This action might not be possible to undo. Are you sure you want to continue?
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. . ... . . . . . . . . . . . . . . . .. . . 290 5 Regenerative simulation . . . . . . .. 292 6 Sensitivity analysis . . . . . . . . . . . .. . . . . . . . . . 297 2 Further applications of martingales . . . . .. . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . 340 A4 Some linear algebra . . . . . . .. . . . . . . . . . . . . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . 271 6 Reservedependent premiums . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . The twobarrier ruin problem . . . . . . 251 2 The compound Poisson model . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . 264 5 Finitehorizon ruin probabilities .. . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . 323 6 Reinsurance . . . . . . . . . . . . .. .. . . ... . . . . . . . . . . . . 306 4 The distribution of the aggregate claims . 336 A3 Matrixexponentials . . .. . . . . . . . . . . 331 A2 WienerHopf factorization . . . . . . . . . . . . . . . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . 316 5 Principles for premium calculation . . 259 3 The renewal model . . . . . 326 Appendix 331 Al Renewal theory . . . . . . . . . . . . . . . . . . . .. . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . 261 4 Models with dependent input . . . . . . 287 4 Importance sampling for the finite horizon case . . .. .. 304 3 Large deviations . . . . . . . .. . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . . . . . . . . . . . 350 Bibliography Index 363 383 . .
This page is intentionally left blank .
and the series editor Ole BarndorffNielsen for their patience. The course was never realized. 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. and has been an active area of research from the days of Lundberg all the way up to today. University of Copenhagen. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. In particular. but the handouts were written and the book was started (even a contract was signed with a deadline I do not dare to write here!). A similar thank goes to all colleagues who encouraged me to finish the project and continued to refer to the book by Asmussen which was to appear in a year which continued to be postponed. which has in particular removed one of the standard criticisms of the area. Since I was to produce some handouts for the students anyway. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. Thus. the book is basically mathematical in its flavour. But the pace was much slower than expected. if the formulations occasionally give a different impression. it is not by intention. that it can only say something about very simple models and questions.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. However. and the result is now that the book is much more related to my own research than the initial outline. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. the idea was close to expand these to a short book on the subject. and other projects absorbed my interest. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. Apart from these remarks. this applies to longrange dependence which is intensely studied in the neighboring ix . It has obviously not been possible to cover all subareas. and my belief was that this could be done rather quickly. I have deliberately stayed away from discussing the practical relevance of the theory.
the first part of 11.13. Asmussen. Finally.g. 111.maths . it has not been possible to incorporate more numerical examples than the few there are. VI. VIII. the standard stochastic control setting of diffusion models has been considered. It is obvious that such a system involves a number of inconsistencies and omissions.g. For a second reading. Willinger et al. 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). Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties.2 more properly). see e.45.13 and XI. read Chapter I.2. In the classical setting of CramerLundberg models.g. see also Schmidli [325] and the references in Asmussen & Taksar [52]. I intend to keep a list of misprints and remarks posted on my web page.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. One is by model. [381]).lth. A book like this can be organized in many ways.1.3.6 (to understand the PollaczeckKhinchine formula in 111. see in particular Michna [259]. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.15. some papers not cited in the text but judged to be of interest are included in the Bibliography. X. I regret that due to time constraints. More recently. Here is a suggestion on how to get started with the book. for the effects on tail probabilities. Hojgaard & Taksar [206].se Lund February 2000 Soren Asmussen . In addition. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. incorporate 11.4a. The rest is up to your specific interests.13. IV. VII. 111. The main motivation comes from statistical data for network traffic (e.5.2. another by method. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. Concerning ruin probabilities.14. For a brief orientation.x PREFACE field of queueing theory. The present book is in between these two possibilities. IV. e. Resnick & Samorodnitsky [303] and references therein.89.lth. http:// www. Another interesting area which is not covered is dynamic control.13 and IX. IV. VII. an area which is becoming increasingly important. Good luck! I have tried to be fairly exhaustive in citing references close to the text. IX.
Fig.6 by my 1999 simulation class in Lund. were produced by Lone Juul Hansen . 5.2 by Rafal Kulik .PREFACE xi The second printing differs from the first only by minor corrections.4 from Asmussen.6.8 . IV. as well as some additional references continue to be at the web page. not least the more complicated ones.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Parts of X. 111 . More substantial remarks.1 and X. A number of other figures were supplied by Christian Geisler Asmussen . supported by Center for Mathematical Physics and Stochastics (MaPhySto). of which there are not many at this stage . Fig. Section VIII. . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Section VII . Aarhus. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX.5 from Asmussen [21] with permission from CRC Press. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. Parts of II. Schmidli & Schmidt [47] with the permission from Applied Probability Trust .1 by Bjarne Hojgaard and the table in Example 111.6 is reprinted from Asmussen & Schmidt [49] and parts of IX. many of which were pointed out by Hanspeter Schmidli .
This page is intentionally left blank .
(1. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. A risk reserve process { Rt}t>o.Chapter I Introduction 1 The risk process In this chapter . respectively. M = (1.4) O<t<oo O<t<T 1 .3) sup St. MT = sup St. as defined in broad terms . We denote throughout the initial reserve by u = Ro.i(u. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . is a model for the time evolution of the reserves of an insurance company. we introduce some general notation and terminology.2) (O<t<T Ro=ul.T) = P inf Rt < 0 I . results and topics to be studied in the rest of the book.1) We also refer to t/) ( u) and 0(u. and give a very brief summary of some of the models. For mathematical purposes. (1. They are the main topics of study of the present book. T) as ruin probabilities with infinite horizon and finite horizon . (1.Rt.
we see that Nt Nt Rt = u + pt .5) i. St = E Uk .1 . INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. per unit time. Figure 1. 1.E Uk. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. say. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. Thus.b(u) = P (r(u) < oo) = P(M > u). (1. respectively. t] is finite.. .i(u.1.6) Sofar we have not imposed any assumptions on the risk reserve process. the number Nt of arrivals in [0. (1. and T1 is the time of the first claim.2 CHAPTER I.. We denote the interarrival times of claims by T2. the time of arrival of the nth claim is an = T1 + • • • + Tn. • Premiums flow in at rate p.T) = F (MT > u) = P(r(u) < T). Putting things together. T3. However.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. (1. That is.pt. the ruin probabilities can then alternatively be written as .
20%.s.1. We shall not deal with this case either.. though many results are straightforward to generalize from the compound Poisson model. and in fact: Proposition 1. that the insurance company should try to ensure 77 > 0. allowing a countable infinity of jumps on Fig. of course. It would appear obvious.1..e. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. one could well replace Rt by Rtnr(u) or RtA. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. Thus. If 77 > 0. on Fig.1 Assume that (1. for example. then M = oo a. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.s.8) The interpretation of p is as the average amount of claim per unit time. We study this case in Ch. • Brownian motion or more general diffusions. we shall. immaterial. However. and hence O(u) < 1 for all sufficiently large u.8) holds.) V 0. t * oo. 1. For the purpose of studying ruin probabilities this distinction is. however. 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. one may well argue that Brownian motion in itself could be a reasonable model. however. . VII. We shall discuss Brownian motion somewhat in Chapter IV. a basic references is Gerber [127].1 the slope of {Rt} should depend also on the level). and hence . then M < oo a.b(u) = 1 for all u. since any modeling involves some approximative assumptions. say 10% . not discuss whether this actually corresponds to practice. 1. and the basic ruin probabilities are derived in XI. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. (1. A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p.1. If 77 < 0.(. but as an approximation to the risk process rather than as a model of intrinsic merit.
The simplest example is 3(t) = V where V is a r . _ St __ k =1 Uk pt a4. St In concrete models. k=1 (1.. If u oo.Q (say) and U1.d.6EU (on the average. t t p  p' t ^ oo. 0 We shall only encounter a few instances of a Cox process. Proposition 1.10) hold with p constant.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. namely.Tp). it is not too difficult to show that p as defined by (1. tb(u) = 1 for all u holds also when rl = 0. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. Nt)}. . . However. and independent of {Nt}. and independent of {(0(t).8). are i.. This case is referred to as the mixed Poisson process. Here it is easy to see that p = .10) is a property which we will typically encounter.i.oo t 0 J (provided the limit exists). (1. (1. are i.8) is given by ^t p = EU • lim it (3(s) ds t. then similarly limSt/t < 0..3 Assume p 54 1 and define Rt = Rt1p.. with the most notable special case being V having a Gamma distribution.b(u) < 1 for all u when rl > 0..v.T) for {Rt} is given by V)(u) = t/i (u). rl > 0. INTRODUCTION Proof It follows from (1.11) . zP(u .s. if {(3(t)} is nonergodic. . Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). this needs to be verified in each separate case. then this limit is > 0 which implies St a$ oo and hence M = oo a. However. corresponding to the Pdlya process. and that . not all models considered in the literature have this feature: Example 1. U2. (1.Q claims arrive per unit time and the mean of a single claim is EU) and that also Nt t aoo t lira EEUk = p.s. U2.T) = i. If 77 < 0.10) Again. in connection with risk processes in a Markovian or periodic environment (Chapter VI).8) that F N.i. Thus p may well be random for such processes. namely that M = oo a. M < oo a. and here (1. 0(u. If U1.d. we obtain typically a somewhat stronger conclusion.s.i(u.4 CHAPTER I. where {Nt} is a Poisson process with rate .
but in probability and applied probability as a whole. Daykin et al. Insurance: Mathematics and Economics. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. in a number of models. For mixed Poisson processes and Polya processes. Gerber [157]. Taylor [364]. Cox processes are treated extensively in Grandell [171]. Some of the main general ideas were laid down by Lundberg [250]. see e . Some early surveys are given in Cramer [91].g. Sundt [354]. Notes and references The study of ruin probabilities.. we shall be able to identify p with the traffic intensity of an associated queue. Since { Rt } has premium rate 1.g.. [76]. another important early Swedish work is Tacklind [373]. Besides in standard journals in probability and applied probability. the research literature is often published in journals like Astin Bulletin . was largely initiated in Sweden in the first half of the century. Straub [353]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Embrechts et al. Note that life insurance (e. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. often referred to as collective risk theory or just risk theory. In the even more general area of nonlife insurance mathematics. [330]. Grandell [171]. U2. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example... The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Schmidt & Teugels [307] and Seal [326]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Pentikainen & Pesonen [101]. We roughly classify these into two groups . Some main later textbooks are (in alphabetical order) Buhlmann [82]. De Vylder [110]. [101]. and we do not get near to the topic anywhere in this book.2. the assumption > 0 is equivalent to p < 1. Gerber [159]) has a rather different flavour. in particular. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. the role of the result is to justify to take p = 1. The Swedish school was pioneering not only in risk theory. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Note that when p = 1. Hipp & Michel [198]. Schmidli. Buhlmann [82]. the claim arrivals are Poisson or renewal at the same time). Heilmann [191]. the recent survey by Grandell [173] and references therein. 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. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. see also Chapter XI. lighttailed distributions (sometimes the term . Rolski. Daykin. Segerdahl [334] and Philipson [289]. [134].
3) .g.e.2 and /LB is the mean of B.B(x) satisfies B(x) = O(e8x) for some s > 0. if 1 °O AB Jbos x B(dx) > 0. As in a number of other applied probability areas. s<8.6 CHAPTER I.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. B[s] is finite for some s > 0. In contrast. INTRODUCTION 'Cramertype conditions' is used).x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). The crucial feature is the lack of memory: if X is exponential with rate 6.2) = 0. for the compound Poisson model with exponential claim sizes the ruin probability . 6 has density r(p)xPleax b(x) P and m. the m. where B(bo.1) The parameter 6 is referred to as the rate or the intensity. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . and heavytailed distributions. regularly varying (see below) or even regularly varying with infinite variance. (2.u at the time of ruin given r(u) is again exponential u with rate 8. For example in the compound Poisson model.f. then the conditional distribution of X .f.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. and can also be interpreted as the (constant) failure rate b(x)/B(x). 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'. Example 2 .8. a fact which turns out to contain considerable information. B is heavytailed if b[s] = oo for all s > 0. 2a Lighttailed distributions Example 2. i.g.O(u) can be found in closed form. but different more restrictive definitions are often used: subexponential. Equivalently. the exponential distribution is by far the simplest to deal with in risk theory as well. In particular. Here lighttailed means that the tail B(x) = 1 . P B[s]= (8Is ) . On the more heuristical side.
. the Gamma density (2.1) (or the 1/pth root if p < 1). P b(x) = r` aibiea. p) = J tPletdt. X2.. 0 < ai < 1. An important property of the hyperexponential distribution is that its s. p) °° where r (x. we develop computationally tractable results mainly for the Erlang case (p = 1..c. 0.v. among others.y i=1 where >i ai = 1. are i. p). 2. by Grandell & Segerdahl [175] and Thorin [369]. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2..v.d..ate (b2 ): L• i=o In the present text. the squared coefficient of variation (s. B(x) = r(p) Asymptotically. one has r(bx. This special case is referred to as the Erlang distribution with p stages.2) can be considered as the pth power of the exponential density (2.2. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . The exact form of the tail B(x) is given by the incomplete Gamma function r(x. In particular. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. i = 1.. is > 1. > 1 for p < 1 and = 1 for p = 1 (the exponential case). and exponential with rate d.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. p. . Ruin probabilities for the general case has been studied.c. In particular.. then X v Xl + • • • + X. where X1. if p is integer and X has the gamma distribution p. .1 Poisson events in [0. or just the Erlang(p) distribution.i.).. u Example 2 . x] so that B(x) = r` e. u .) VarX1 (EX )2 p is < 1 for p > 1.
The couple (a.8) j=1 j=1 j=1 where the parameters in (2. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . are b(x) = aeTxt. 1)' is the column vector with 1 at all entries. We give some theory for matrixu exponential distribution in VIII.d. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.g. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .6.xo)+ is covered by the reinsurer).4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. B(x) = aeTxe where t = Te and e = (1 . Important special cases are the exponential.6.. resp. equivalently. there exists a xo < oo such that B(x) = 0 for x > xo.(2. See XI. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. 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. However. (or. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. which is slightly smaller but more amenable to probabilistic reasoning. q2 q3 (2. INTRODUCTION Example 2 . The density and c. of which one is absorbing and the rest transient.8) are realvalued.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. Example 2 .f. a. 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.e. We give a more comprehensive treatment in VIII.8 CHAPTER I.7) are possibly complexvalued but the parameters in (2. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. the Erlang and the hyperexponential distributions. T) or sometimes the triple (E..f. The parameters of a phasetype distribution is the set E of transient states. . T) is called the representation. Example 2 .1 and defer further details to u Chapter VIII.
p a 1 (2.9) which is heavytailed when 0 < r < I.10) The loinormal distribution has moments of all orders.13) u .N(0. a)/A)a+1' x > a. p is defined as the distribution of ev where V . x < a.N(p. (2.1. In particular. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x.2.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. the exponential distribution representing the simplest example since here b(x) is constant. and then b(x) = 0. It follows that the density is 't (1ogX . a2). (2. or equivalently as the distribution of a°U+µ where U . we obtain the Weibull distribution B(x) = eCx'. Example 2 .7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. However. 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).8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. Here failure rates b(x) = b(x)/B(x) play an important role.pl = 1 W (logx . (2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. the mean u is eµ+a /2 and the second moment is e2µ+2o2. Writing c = d/r. There are various variants of the definition around. b(x) = crx''le`xr.1). All moments are finite. the tail is B (x ) 2 x. u Example 2 .u l b(x) = d dx or J ax lor 1 exp Asymptotically.
16) 11 Example 2.14) The pth moment is finite if p < 5 and infinite if p > 5.(1 + 2x + 2x2)e2x) p = 2 (2. Choudhury & Whitt [1] as the class of distributions of r. For p = 1. B(x) = O(xP). the density is { 3 (1 .15) x2 + 16x3 ) a3x/2) 3 (1 . in particular.'s of the form YX. x + 00.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. oo) is slowly varying .v.10 CHAPTER I.13).e. INTRODUCTION Example 2.L( x ). another standard example is (log x)'). A = 1 and X is standard exponential. satisfies L(xt)/L(x) 4 1. Thus. The motivation for this class is the fact that the Laplace transform is explicit (which is not the case for the Pareto or other standard heavytailed distributions). In general.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. { s () 1s+3s29s3log(1+2s I p=3. 6 is defined as the distribution of et' where V has the gamma density (2. in particular. i. (2. The simplest examples correspond to p small and integervalued. examples of distributions with regularly varying tails are the Pareto distribution (2. the loggamma distribution is a Pareto distribution. The density is 8p(log x)pi b(x) .12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) .x6+lr(p) (2. (2.17) where L (x) is slowly varying. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. u Example 2 .12) (here L (x) * 1) and ( 2.2).(1 + Zx + $ p = 3. where Y is Pareto distributed with a = (p .1)/p. x 4 oo (any L having a limit in (0. u .
. this phenomenon represents one of the true controversies of the area.1. one may argue that this difficulty is not resticted to ruin probability theory alone. for example the lognormal distribution is subexponential (but not regularly varying).3. The reason is in part mathematical since this model is the easiest to analyze. From a practical point of view. but the model also admits a natural interpretation : a large portfolio of insurance holders .13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2.4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. Also. and so is the Weibull distribution with 0 < r < 1.18) B(x) It can be proved (see IX. However. 3 The arrival process For the purpose of modeling a risk process . and based upon such information it seems questionable to extrapolate to tail behaviour. (2. At least as important is the specification of the structure of the point process {Nt } of claim arrivals and its possible dependence with the claims. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed.1) that any distribution with a regularly varying tail is subexponential.. the claim size distribution represents of course only one aspect (though a major one). Thus. the knowledge of the claim size distribution will typically be based upon statistical data. 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... THE ARRIVAL PROCESS 11 Example 2. Namely. Similar discussion applies to the distribution of the accumulated claims (XI. which each have a ( timehomogeneous) small rate of experiencing a . though the proof of this is nontrivial. When studying ruin probabilities. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. We return to a closer study in IX. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. U2.
such that 8(t) = . T2.d. but with a general not necessarily exponential distribution ).12 CHAPTER I. This model . we study this case in VI . in particular to allow for certain inhomogeneities. so that . The compound Poisson model is studied in detail in Chapters III. 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 .6. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications.g. when Jt = i.(3. the periodic and the Markov modulated models also have attractive features . the negative binomial distribution.8 (t) is a periodic function of t.3(t) fluctuating over time. In order to prove reasonably substantial and interesting results . Historically. This applies also to the case where the claim size distribution depends on the time of the year or . I. In others. to be studied in Chapter V. Cox processes are. Some of them have concentrated on the marginal distribution of NT (say T = one year ). However . and also that the ruin problem may be hard to analyze . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. are i. 5. with the extension to premiums depending on the reserve. which facilitate the analysis.e. with a common term {Nt} is a Markovmodulated Poisson process . found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . Another one is Cox processes. in Chapter VII). An obvious example is 3(t) depending on the time of the year (the season).i. too general and one neeed to specialize to more concrete assumptions . e. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . A more appealing way to allow for inhomogeneity is by means of an intensity . IV (and. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available .. where {/3 (t)}too is an arbitrary stochastic process . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. INTRODUCTION claim . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). To the author 's knowledge . getting away from the simple Poisson process seems a crucial step in making the model more realistic. see 11. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. epidemics in life insurance etc. Nevertheless . however. has some mathematically appealing random walk features .. Mathematically. gives rise to an arrival process which is very close to a Poisson process. 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. 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)..
that quite often the emphasis is on computing expected values like EV. In the setting of (4.v. interacting particle systems. with Poisson arrivals and constant release rule p(x) = 1. Thus.1). dam/storage processes. extreme value theory.1) permitting to translate freely between risk theory and the queueing/storage setting. Similarly. methods or modeling ideas developed in one area often has relevance for the other one as well. It should be noted.1) where V is the limit in distribution of Vt as t + oo. 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. ruin probabilities for risk processes with an input process which is renewal. this amounts to Vo having the stationary distribution of {Vt}). point processes and so on.T) = P(VT > u). The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. genetics models. stochastic geometry. Mathematically. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. others being branching processes.0 (u. however. The study of the steady state is by far the most dominant topic of queueing and storage theory.'s like V is available. others are quite different. Markovmodulated or periodic can be related to queues with similar characteristics. reliability. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. and a lot of information on steadystate r. 0(u) = P(V > u). it is desirable to have a set of formulas like (4. 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 . and the limit t 4 oo is the steadystate limit. this gives only f0 O°i (u)du which is of limited . The ones which appear most related to risk theory are queueing theory and dam/storage processes. stochastic differential equations. time series and Gaussian processes.1) holds as well provided the risk process has a premium rule depending on the reserve. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas. and which seems well motivated from a practical point of view as well. More generally. queueing theory. In fact. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. Some of these have a certain resemblance in flavour and methodology.6) . (4.4. and here (4.
2).1).g. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .1 . 4b Exact solutions Of course ..8. 3. • The compound Poisson model with some rather special heavytailed claim size distributions.3. as is typically the case. which can be expanded into a sum of exponential terms by diagonalization (see.1. The infinite horizon (steady state ) case is covered by letting T oo. have to some extent a different flavour. Vi(u. Similarly. • The compound Poisson model with a claim size distribution degenerate at one point. B(x) = ebx. much of the study of finite horizon problems (often referred to as transient behaviour) in queueing theory deals with busy period analysis which has no interpretation in risk theory at all.14 CHAPTER I.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. e .6. .3. 3.T).v. see Corollary VII. the functions w x f d 1 exdx () .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 ).p(y) y^ Jo p(x) can be written in closed form. A prototype of the duality results in this book is Theorem 11. Example VIII. Here ?P(u) is explicit provided that . The fact that Theorem H. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. the two areas. INTRODUCTION intrinsic interest . Thus . though overlapping. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u).3. see Boxma & Cohen [74] and Abate & Whitt [3].'s like the environmental process {Jt} in a Markovmodulated setting. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B.3. Here O(u) = peryu where 3 is the arrival intensity. which gives a sample path version of (4. p = 0/8 and y = 8 . see Corollary III.
Given this can be done. For the finite horizon ruin probability 0(u. [s. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. Embrechts.b(u)du . A notable fact ( see again XI. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. T) can then be calculated numerically by some method for transform inversion. (u. 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. Abate & Whitt [2]. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p.2) is the natural scale.ff 2µ(y)/a2(y) dy} dx .7.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). esuTb( u. a2 (x): Ip (u) = where S(u) = f °O exp {.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions.f f 2µ(y)/a2(y) dy} dx  (4. . it is easier to find the Laplace transforms = f e8 .4. but are somewhat out of the mainstream of the area . 191). • An astable Levy process with drift . T). relevant references are Grubel [179]. T) themselves.Lef$er function. 1). the formulas ( IV. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. We don't discuss Laplace transform inversion much. the only example of something like an explicit expression is the compound Poisson model with constant premium rate p = 1 and exponential claim size distribution . T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). O(u. Here are some of the main approaches: Laplace transform inversion Often. However.S(u) 1S(oo) f °D exp {. see VIII. f {eXp U LX 2. Ab(u). Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution.
U is explicit in terms of the model parameters.1) and y > 0 is the solution of the Lundberg equation (4. and in particular the naive idea of conditioning upon process behaviour in [0. which can equivalently be written as f3 [7] = 1 +13 .3) in the compound Poisson model which is an integral equation of Volterra type. Differential.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). see VIII.3.p)/(13B'[ry] .7. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.3) where C = (1 .4) 00['Y]1)'Y = 0.g. 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. T) as the solution to a differential.f. . (4. One example where this is feasible is the renewal equation for tl'(u) (Corollary III.) B[s]. However. An example where this idea can be carried through by means of a suitable choice of supplementary variables is the case of statedependent premium p(x) and phasetype claims.and integral equations The idea is here to express 'O(u) or '(u. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). 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. In the compound Poisson model with p = 1. and in quite a few cases (Chapter VIII). For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.or integral equation. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. u * oo. and carry out the solution by some standard numerical method. dt] most often leads to equations involving both differential and integral terms.Ce"u. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).
when the claim size distribution is of phasetype. However. in such cases the evaluation of C is more cumbersome. T) for large u are available in most of the models we discuss.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. the claim size distribution should have an exponentially decreasing tail B(x). J B dx. but typically not very precise in their first naive implementation. It has generalizations to the models with renewal arrivals. T).2. for the compound Poisson model ^(u) p pu In fact . a Markovian environment or periodically varying parameters. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. See Chapter IX. Approximations for O(u) as well as for 1(u. Large claims approximations In order for the CramerLundberg approximation to be valid. For example. However. Diffusion approximations are easy to calculate. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. in 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.4.7 and IV. other approaches are thus required. (4. corrected diffusion approximations (see IV. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. In the case of heavytailed distributions. often for all u > 0 and not just for large u. In particular. incorporating correction terms may change the picture dramatically. In fact. u > oo. This list of approximations does by no means exhaust the topic. some further possibilities are surveyed in 111 .6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. 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. .
which is a standard statistical problem since the claim sizes Ui. it has the advantage of not involving approximations and also.g. . . more importantly. and to plot the empirical mean residual life 1 N . lower bounds etc. empirical evidence shows that the general principle holds in a broad variety of settings. as a general rule. the difficulty comes in when drawing inference about the ruin probabilities. For example.3). . of being somewhat easier to generalize beyond the compound Poisson setting. T]. For example. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. In practice. . We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. UNT are i.x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case.k (U(`) . This procedure in itself is fairly straightforward. 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. . The standard suggestion is to observe that the mean residual life E[U . . When comparing different risk models.18 CHAPTER I.U(k)) i =k+ i . though not too many precise mathematical results have been obtained. obtained say by observing the risk process in [0. given NT. UNT may be viewed as an interpolation in or smoothing of the histogram).) at various places and in various settings. to have smaller ruin probabilities than when B is nondegenerate with the same mean m.. This is proved for the compound Poisson model in 111. fitting a parametric model to U1. INTRODUCTION Compared to the CramerLundberg approximation (4. e. they have however to be estimated from data. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. in the compound Poisson model. it is a general principle that adding random variation to a model increases the risk. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. one expects a model with a deterministic claim size distribution B. However..d. However. How do we produce a confidence interval? And.8. 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.i.. say degenerate at m.
3) or Section 3 of Chapter VI are referred to as Proposition VI.2.(5. because it appears to require an infinitely long simulation. CONVENTIONS 19 as function of U(k).5. (or a functional of the expectation of a set of r. and of course the method is relevant in risk theory as well. < U(N) are the order statistics based upon N i. claims U1. where U(1) < . 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics.3) and Section VI. The infinite horizon case presents a difficulty. respectively.. . formula VI. UN. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths.i. but is not very satisfying. in this book referred to as [APQ]). good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget..4.. Still. reference [14].v's) which can be generated by simulation. in all other chapters than VI where we just write . to observe whether one or the other limiting behaviour is apparent in the tail. this is a straightforward way to estimate finite horizon ruin probabilities. However. Thus Proposition 4. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. See further Embrechts. and in fact methods from that area can often be used in risk theory as well .. Truncation to a finite horizon has been used. .. We look at a variety of such methods in Chapter X. For example.e.3 (or just 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.d.3). A main problem is that ruin is typically a rare event (i.2. formula (5. The chapter number is specified only when it is not the current one. Klnppelberg & Mikosch [134]. and look at them to see whether they exhibit the expected behaviour or some surprises come up. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time.v.
Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. (A. right hand side (of equation) r.v. with respect to w. A different type of asymptotics: less precise. formula (5.g.f. .3) or Section 3.c.h. i.4.v. E. log E[s] where b[s] is the m.f.29) refer to the Appendix. oo).d.e. In particular.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c.B(x) = P(X > x) of B. (moment generating function) fm e82B(dx) of the distribution B. infinitely often l. as for typical claim size distributions. and for a defective probability distribution IIGII < 1. i.ceax. B(x) the tail 1 .d. see under b[s] below.d.f. Abbreviations c.t. B[s] the m.p.20 CHAPTER L INTRODUCTION Proposition 4. r.g.f. B is concentrated on [0. mation.s.2. E expectation. moment generating function. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. left hand side (of equation) m.g. for a probability distribution IIGII = 1. B(dy).g. squared coefficient of variation. h + 0. then for 1s < 5). IIGII the total mass (variation ) of a (signed ) measure G .i. w. if B(x) .f. B(x) = P(X < x) = fx. cumulative distribution function P(X < x) c. or a more precise one like eh .g. . n i oo. independent identically distributed i. random variable s.o. with probability Mathematical notation P probability.r. say a heuristic approxi1 + h + h2/2.h. References like Proposition A. n! 27r nn+1/2en. cumulant generating function. The Laplace transform is b[s]. EX2/(EX)2.f.s. If.
a2) the normal distribution with mean p and variance oa2. N(it. intensity interpretation.e. E[X.5. Usually. 0 marks the end of a proof. A.e. 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. all stochastic processes considered in this book are assumed to have sample paths in this space. . In the Frenchinspired literature. row vectors have lowercase Greek letters like a. the processes we consider are piecewise continuous. Thus. the ith entry is 1 and all other 0. an example or a remark. 21 D [0. 7r.. F o r a given set x1. though slightly more complicated. Matrices and vectors are denoted by bold letters. . i. Notation like f3i and 3(t) in Chapter VI has a similar . Unless otherwise stated. i. Xt_ the left limit limstt X8f i. (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(A) the indicator function of the event A. oo) the space of Rvalued functions which are rightcontionuous and have left limits. only have finitely many jumps in each finite interval. R(s) the real part of a complex number s. the value just before t. matrices have uppercase Roman or Greek letters like T. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty.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). . (the dimension is usually clear from the context and left unspecified in the notation). CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B.e. of numbers.A] means E[XI(A)]. xa.. Usually. the ith unit row vector is e'i. 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. and column vectors have lowercase Roman letters like t. a.
.5. ry The adjustment coefficient. 111. J the rate parameter of B for the exponential case B(x) = eby. I. interpretation. VI. or quantities with a similar time average interpretation. FL. e.1. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation.g.1.5. p the net amount /3pB of claims per unit time.22 CHAPTER L INTRODUCTION B the claim size distribution. cf. cf. though slightly more complicated. 'q the safety loading . cf. I. Notation like BE and B(t) in Chapter VI has a similar.
however. The reader should therefore observe that it is possible to skip most of the chapter. 23 . More precisely. in particular at a first reading of the book. The duality results in Section 3 (and. strictly speaking. Sections 4. used in Chapter VI on risk processes in a Markovian (or periodic) environment.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. When encountered for the first time in connection with the compound Poisson model in Chapter III. in most cases via likelihood ratio arguments. the level of the exposition is. The topic is. however. Sections 4. somewhat more advanced than in the rest of the book. Due to the generality of the theory. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. however. All results are proved elsewhere . a parallel selfcontained treatment is given of the facts needed there. in part. The general theory is. 5) are. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. The likelihood ratio approach in Section 2 is basic for most of the models under study. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. not crucial for the rest of the book.
f1 .24 CHAPTER II. StUit.u denote the overshoot.QµB < 1.2) takes the form 1 = E [e'ys(). Thus N.1 Assume that (a) for some ry > 0. {e'YS° }t>0 is a martingale.1 is basic for the study of the compound Poisson model in Chapter III. The more general Theorem 6.0. and in the limit (1. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}).(u)..2) As T > oo. V) (u. T) = P(T(u) < T). and the ruin probabilities are ip(u) = P (T(u) < oo).2 Consider the compound Poisson model with Poisson arrival rate . Proposition 1. 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. 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}. using r(u) A T invokes no problems because r(u) A T is bounded by T). T(u) < T] + E [eryST . We get 1 = Ee7So = E e'Y S(. Example 1 . T(u) > T] . T(u) < oo] + 0 = eryuE [e7Vu).5 can be skipped. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . claim size distribution B and p = . (b) St a$ oo on {T(u) = oo}. Let e(u) = ST(u) .)AT = E [e7ST(°). As usual. however.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. (1.
and thus Ee'rs° = 1.a it is immediately seen that y = S . Thus. Then {St } is Brownian motion with variance constant o2 and drift p < 0. and thus by the memoryless property of the exponential distribution ./3.a = a . Eeas° = e"(°) where n(a) = a2a2/2 .1) shows that Eels.2(c)). Example 1 .5 For the compound Poisson model with B exponential. the ruin probability is O(u) = pe. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .a.Q and the U. u Corollary 1.2 and drift p > 0. Since {St} has stationary independent increments. B(x) _ edx.. with common distribution B (and independent of {Nt}). O(u ) < e7".1.= e"(') where K(a) = . A simple calculation (see Proposition III. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where . Thus.1) . By standard formulas for the m.d.3/6 < 1. of the normal distribution.4 (LUNDBERG ' S INEQUALITY ) tion 1 . The available information on this jump is that the distribution given r(u) = t and S.1) .2. Since {St} has stationary independent increments. the conditions of Proposition 1.1. Under the conditions of Proposi Proof Just note that C(u) > 0. 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 .f.1 are satisfied. are i. MARTINGALES 25 where {Nt} is a Poisson process with rate .3 Assume that {Rt} is Brownian motion with variance constant o. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . Proof Since c(a) = /3 (B[a] . and p =. and thus Ee7s° = 1.1 is satisfied.ap. u Corollary 1.Q(B[a] .u + x is again just exponential with rate S.Q. condition (a) of Proposition 1. the conditional distribution of the overshoot e(u) = U . From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2. From this it is readily seen (see III.1. the martingale property now follows just as in Example 1.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. 1.Ft = a(S" : v < t).g.r" where y = S .i.x.
. Grandell [171]. oo).1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. [172]. More recent references are Dassios & Embrechts [98]. B. and is further exploited in his book [157]. However. on (DE. 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. F(S) = P(S) = 1. Proof Just note that ^(u) = 0 by continuity of Brownian motion.v.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. Two such processes may be represented by probability measures F. which we equip with the natural filtration {. Grandell & Schmidli [131]. .F). cf. A]. A somewhat similar u argument gives singularity when B $ B. as shown by the following example this setup is too restrictive: typically'. the parameters of the two processes can be reconstructed from a single infinite path. The number Nt F) of jumps > e before time t is a (measurable) r. then z/'(u) = e7" where 'y = 21A/a2. (2. and F.3 below).Ft}too and the Borel afield F. P on (DE. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. Embrechts. P are then singular (concentrated on two disjoint measurable sets). Delbaen & Haezendonck [103] and Schmidli [320]. A E Ft. But if a $ ^ .6 N S = { lim Nt I t +00 t gJ are both in F. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. Thus the sets S = I tlim +oot t =. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. Theorem 2. and by the law of large numbers for the Poisson process .e. I.1 Let F. Example 2 . F). then S and S are disjoint . and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. 0 and claim size distributions B. hence so is Nt = limfyo N2`i.
By the martingale property.1) and nonnegativity by letting A = {Lt < 0}. ({. the restriction of P to (DE.2.F). . Then P(A) = E[Lt. we have E [ LTIFT]1 = LT on {T < T}. 1 J (2.Pt}adapted process {Lt}t>o (2. Finally. F) such that ELt = 1. Lets < t. A E F. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.1) holds. implies that E[LtI. A].F such that (2.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. F the Borel o•field and P a given probability measure on (DE.3 Let {Lt}.Tt) is absolutely continuous w.F8] = L8 and the martingale property.e. Hence E [_ . _.1) holds.r. If r is a stopping time and G E PT. G ] = E [LT . P) such that LLt = 1..t. define P by Pt (A) = E[Lt. then there exists a unique probability measure Pon . The truth of this for all A E Y. using the martingale property in the fourth step. then { 1 P(G) = EG . . (i) If {Lt}t> o is a nonnegative martingale w.r. Then Ft (A) = E[Lt. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. A] = E[Lt.t.Pt)) The following result gives the connection to martingales. if for some probability measure P and some {. ({Ft} .2 Let {Ft}t>o be the natural filtration on DE. (ii) Conversely. Lt < 0] can only be nonnegative if P(A) = 0. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). then {Lt} is a nonnegative martingale w. This proves (i).Y) such that P(A) = Pt(A).t.Ft}. Proposition 2. Ft). Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. P be as in Proposition 2.2(i). ELt = 1 follows by taking A = DE in (2. that the restriction of P to (DE. A]. G C {T < oo}. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.r. A E F8. Conversely. 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 . Proof Under the assumptions of (i). A E Ft .
A change of measure is performed by finding a process {Lt} which is a martingale w. For the definition.2) by noting that 1 = ersr(„) = e1'ue7Ou). the natural filtration {. is nonnegative and has Ey Lt = 1 for all x.h. Xt = St.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . each F. first in the Markov case and next (Sections 4.1: Corollary 2.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu).28 CHAPTER II. in continuous time (the discrete time case is parallel but slightly simpler). u From Theorem 2. . of (2. Lr(u) 11 The advantage of (2.1). The crucial step is to obtain information on the process evolving according to F. we need the concept of a multiplicative functional. Consider a (timehomogeneous) Markov process {Xt} with state space E. {St} = {u .4) Proof Letting G = {r(u) < oo}. and this problem will now be studied. St). (2. First we ask when the Markov property is preserved.t. and letting T * oo. In the context of ruin probabilities. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. is Markov w.O(u) = eryuE[e'YC(u). Rt) or Xt = (Jt.. t. applying (2. SOME GENERAL TOOLS AND RESULTS In the general case .4) compared to (1. Now just rewrite the r. where {Rt} is the risk reserve process.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian.2) follows by monotone convergence.1. say. To this end.r. one would typically have Xt = Rt. we have F(G) = V )(u). 1 Since everything is non negative.r.Ft} . T(u) < oo].3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. we assume for simplicity that {Xt} has Dpaths. (2.1) in Proposition 1.Gn {r <T} . 5) for processes with some randomwalklike structure.t. both sides are increasing in T.. Xt = (Jt. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.s.4 Under condition (a) of Proposition 1.
this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. and then L. Zt. The precise meaning of this is the following: being .r. 0 . which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.'s of the form Zt • (Y8 o 0t) comprises all r. oo). Similarly.s. t.2. (2. ({Xt+u} 0<u<8) Theorem 2.Y8f t < s. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. o 9t = V.Ft+8measurable. nonnegative and Lt+8 = Lt•(Lso9t) (2.5) Pxa. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. s. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. A]. (2. By definition of Px. t] * [0..v.t. or. for all x.F8measurable r. the natural filtration {Ft} on DE.T9measurable Y8.Pt+8measurable r.5 Let {Xt} be Markov w. (2.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . Proof Since both sides of (2.Ft }. Vt+e. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].v.. the Markov property can be written E.v.6) for any . since Zt • (Y8 o Ot ) is .7) for any Ftmeasurable Zt and any . Indeed.8) which is the same as (2.7).6) implies (2. The converse follows since the class of r.7).Ft].YB] for any Ftmeasurable r. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.v.Ftmeasurable. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.v. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. (2.[Y.(Xtitl) with all t(i) < t + s.'s of the form fl' f. t and let Px be the probability measure given by t. Y8. o 9tI.Ft] = Ex.5) are Tt+e measurable. where Ot is the shift operator.(A) = Ex [Lt.
(3. Indeed. the arrival epochs are Qi. 3 Duality with other applied probability models In this section. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. reflection at zero and initiar condition Vo = 0. < aN < T.. . . . The formulation has applications to virtually all the risk models studied in this book. see Dynkin [128] and Kunita [239]. (using the Markov property in the second step) so that the martingale property is automatic.. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. In between jumps.Ft] = LtE[L8 o 9t I.. R = p(R)). We work on a finite time interval [0. In between jumps.. aN where or* = T UN_k+l. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. More precisely .5 can be found in Kuchler & Sorensen [240]. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. Thus R = Ro + f p(R8) ds .. The storage process {Vt }o<t<T is essentially defined by time reversion.At where At = k: vk <t U. The result is a sample path relation. . the premium rate is p(r) when the reserve is r (i... UN.. Ro = u (say). } E[Lt+B I. . and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.. t. 0 < vl < . The corresponding claim sizes are Ul.e.1) The initial condition is arbitrary. {Vt} . t] = LtExt L8 = Lt.. CN. with a proof somewhat different from the present one.. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage.6 For {u . then Remark 2.30 CHAPTER H. A more elementary version along the lines of Theorem 2. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or.
V = p(V))..AT_t. ._.e. {Vt} remains at 0 until the next arrival)._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3..1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. (3.11 4. :... DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.____•_. Theorem 3. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous. (3.x. V)(u.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.1 The events {T(u) < T} and {VT > u} coincide.__.. In particular..T) = P(VT > u).2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. The sample path relation between these two processes is illustrated in Fig.___ .3.1. instead of (3.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. 3.__.1) we have Vt = At  f P(Vs)ds where A= U= AT .... Then rt°) > rt°) for all t when u > v. That is.
see Siegmund [344]. Corollary 3.1 with Ro = u = u2). If VaN > 0. 3. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. Proof Let T ^ oo in (3. and then '0 (u) = P(V > u). the distinction between right. Then similarly VVN = r0. we have RQ„ < 0 so that indeed r(u) < T. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. and since ruin can only occur at the times of claims.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. Theorem 3.3). Historically.1 and its proof is from Asmussen & Schock Petersen [50]. = r(VT) .2 from Harrison & Resnick [188]. say V. Hence RQ„ > 0 for all n < N. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. . with distribution B. Then the storage process {Vt} has a proper limit in distribution. this represents a model for storage. 3.i.1 with Ro = u = ul). and a general premium rule p(r) when the reserve is r. Hence if n satisfies VVN_n+1 = 0 (such a n exists.T l . We get: Corollary 3. Historically. we have r(u) > T.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig.3 and being i. we can repeat the argument and get VoN_1 > Ra2 and so on. The arrival epochs correspond to rainfalls.d. Nevertheless. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. if nothing else n = N).Ul < roil  Ul = RQ„ Va1V_1 < RQ2. say of water in a dam though other interpretations like the amount of goods stored are also possible. if and only if O(u) < 1 for all u. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.U1 = Rol. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate .2 Consider the compound Poisson risk model with a general premium rule p(r).U1 > roil . Then Vo. one may feel that the interaction between the different areas has been surprisingly limited even up to today. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. and so on. Resnick & Tweedie [79]. The results can be viewed as special cases of Siegmund duality.
.... W1. where the Yi are i .1. = Xo + Y1 + • • • + Y... . can be viewed as the reflected version of the random walk with increments Z1. WN be the Lindley process generated by Z1 = YN. . .1) Thus {Wn}n=o.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. W1...2) satisfies the same recursion as in (4.. N min (Y1 + • • • + YNn) n=0.. and is reset to 0 once the r.....Y1 according to Wo = 0. Here F is a general probability distribution on R (the special case of F being concentrated on {1.. evolves as a random walk with increments Z1i Z2. Then the events {r(u) < N} and {WN > u} coincide. Proof By (4. Z2 = Y2..h. Xo = 0. . N min (Y1 + + Yn). (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. ZN = . For discrete time random walks ... R valued sequence Z1..w.1)).2) (for a rigorous proof..e. Z2 .. . with common distribution F (say).YNn+1) n=0. the Lindley process Wo.1. hits (oo.....d.. .. n=0.. there is an analogue of Theorem 3.Yl min (YN .. 0).2)... WN = YN . just verify that the r . Z2.i.. W2. I. ...4. 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.. (c) The Lindley process {WN} generated by Zl = Y1.1.. Let further N be fixed and let Wo. Most often. i.. In particular. Z2. For a given i. {Wn}n=0.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0..1.s... is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4...N From this the result immediately follows. generated by Z1...1...min n=0. as long as the random walk only takes nonnegative values.d. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. Z2 = YN_1 i .. 0 Corollary 4. has a proper limit W in distribution as n + oo. ... 1} is often referred to as simple random walk or Bernoulli random walk).. of (4.1.N (4. Theorem 4. . .1 in terms of Lindley processes .
. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . ±2.1 is equivalent to WN D MN = (Z1 + . . Remark 4 .N so that WN _P4 M = supra=0.s.1....o.=o. then the restrictions of Fx. Next consider change of measure via likelihood ratios. Proof Since (YN.. a Markovian change of measure as in Theorem 2. ±1. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. the Lindley processes in Corollary 4. N.. either M = oo a. In general.. ..s..g. (e).. Y1) has the same distribution as (Y1.s ..) sup n=0.1 is actually not necessary ...l. .d. Thus the assertion of Theorem 4.5 does not necessarily lead to a random walk: if.. .. Clearly.. F has a strictly positive density and the Px corresponds to a Markov chain such that the density of X1 given Xo = x is also strictly positive.i. By Kolmogorov's 01 law. (Yi + • • • + Yn) > oo a. . .ooa.. For a random walk.) and defines Zn = Yn. doubly u infinite (n'= 0.the result is a sample path relation as is Theorem 3..g. a sufficient condition for (e) is that EY is welldefined and > 0.. . + Z.. e.1. 0 By the law of large numbers. .2 and Theorem 4. equivalently. or M < oo a.. . (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ).34 CHAPTER II. Similarly.. Combining these facts gives easily the equivalence of (a)(d).1 have the same distribution for n = 0. 176) but appears to be rather intractable. SOME GENERAL TOOLS AND RESULTS (d) m = inf. (e)Yi+•••+Yn 74 .. . w.l. ..2. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo..3 The i. In that case . One then assumes Yn to be a stationary sequence.. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.1.1.s. ZN or. W v m and P(W > u) = P (m > u) = 0(u). there is a more general version of Corollary 4. YN). (d) #. the Y1. YN in Theorem 4. assumption on the Z1.
for some function h with Eh(Y) = 1. Proof If (4.3) holds for n = 1. Y < x]. h(Yn) (4.4).. then n n Ex [f f = Ex H fi a( YY) i=1 i_1 ( Y=) h(YY) H Ef=(Y=)h(Y=) = II J fi(Y )P( d) from which the random walk property is immediate with the asserted form of F. 100 ).nrc(a )} (4. Then the change of measure in Theorem 2.3) holds.4 Let {Ln} be a multiplicative functional of a random walk with E_L.f. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .. Y1). Conversely. of F).. Then the change of measure in Theorem 2. Since L1 has the form g (Xo.g. Y) f (Y)] for all f and x. Breiman [78] p..s. cf. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.4. implying g (x. = 1 for all n and x. this means E(g(x. We get: Corollary 4.. Y) = h(Y ) a. e. (4.Y2) = h(1'i)h(I'a). RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .4. In particular.s.5 corresponds to a new random walk if and only if Ln = h(Y1) . the random walk property implies Ex f (Y1) = Eo f (Y1 ). and so onforn =3. and define Ln by (4.g... Y ) f (Y)] = E[g(O.3) 1Pxa. In that case. . For n = 2. where h (y) = g(0. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. (a) = log F(a] is the c. the changed increment distribution is F(x ) = E[h(Y).4) ({Ln} is the familiar Wald martingale . y ).
{Xt} can be written as the independent sum of a pure drift {pt}. .). this description needs some amendments. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. Roughly. the claim surplus process for the compound Poisson risk model . The simplest case is 3 = JhvMM < oo. 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. given by a the increment distribution F(x) = P(Xn+l .. i.e. the tradition in the area is to use continuous time models. they arise as models for the reserve or claim surplus at a discrete sequence of instants. . e f x:IxJ>e} v(dx) < oo (4. we are ..7) for all e > 0. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . In risk theory. Xt =Xo+pt+oBt+Mt. {Xt} is a random walk.. with premium rate p. but we omit the details ). In discrete time. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o).36 CHAPTER II.1). which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .Xn < x). the pure jump process is given by its Levy measure v(dx).. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . However.5) Note that the structure of such a process admits a complete description. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.Ft] = Eof (X. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < .Xt)I. Xt(2)_t(l). corresponds to a process with stationary independent increments and u = p. or imbedded into continuous time processes . (4. say the beginning of each month or year . v2 = 0 and v = 3B). (4. see Chapter V.6) More precisely. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). a positive measure on R with the properties e J x2v(dx) < oo. however.t(i . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). In continuous time.
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. VT + V for some r. jxJ v(dx) < oo. Then the storage process {Vt} has constant release rate 1.v. Furthermore.Q and distribution B of the service times of the arriving customers.3 and decreases linearly at rate 1 in between jumps. and the reflected version is then defined by means of the abstract reflection operator as in (4.1)v(dx) (4. defined as a system with a single server working at a unit rate.7 If {Xt} has stationary independent increments as in (4. having Poisson arrivals with rate . is easily seen to be f3pB < 1. where VT is the virtual waiting time at time T in an initially empty M/G/1 queue with the same arrival rate /3 and the service times having the same distribution B as the claims in the risk process.1. Now consider reflected versions of processes with stationary independent increments. 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. WT = XT . where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". Chapter III. V E [0.6 In the compound Poisson risk model with constant premium rate p(r) .s. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.2). (ex .] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3.6).min Xt (4. Corollary 4. Proposition 4. i. oo]. and b(u) = P(V > u). then Ee'(xtxo) = Eoeaxt = etx(a).8) O<t<T (assuming Wo = Xo = 0 for simplicity). has upwards jumps governed by B at the epochs of a Poisson process with rate . [The condition for V < oo a.10) .4. O(u. First assume in the setting of Section 3 that {Rt} is the risk reserve process for the compound Poisson risk model with constant premium rate p(r) = 1. cf.e. T) = P(VT > u). 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).
f. (4. t. then the changed parameters in the representation (4. In particular. 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. we use the characterization (4.xo)tk ( e).1 that E eaMt = exp fmoo In the general case . u Note that (4. if Lt = e9(xt . .6) are µ = µ + Oo2 . Chung [86]). 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.Xt)I Ftl = Eof (X8)g(s. This is of course no coincidence since the distribution of Xl . of an infinitely divisible distribution (see. use the representation as limit of compound Poisson processes. X8) = Eof (X8)L8 = Eof (X8)• For the second. By explicit calculation . Theorem 4 .Xt)L8 o 0tIFt] = E [f (Xt+s . Then the Markov process given by Theorem 2.Xt)g(s..38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. Q2 = v2. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. let e" (a ) = Eoeaxl. Xt +B . Xt Xo) with E2Lt = 1 for all x.Xt)I'Ftl = E [f(Xt+B .1 ) v(dx) .g. Proof For the first statement . 5 has stationary independent increments as well.g. Eea(µt + QBt) = et{aµ +a2OZ/2}.10) is the LevyKhinchine representation of the c.1)eexv(dx).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 .11) (eax .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. v(dx) = e9xv (dx).5) and get E [f(Xt+B . e.
Example 4 . As for processes with stationary independent increments .l3 and claim u size distribution B. u 5 Markov additive processes A Markov additive processes.1) For shorthand ... U2. corresponding to p = 1. <t whenever the RadonNikodym derivative dB/dB exists (e.(3 = . (5.Ft] = Ejt. we write Pi. B(dx) = B[9] B(dx).. a = 0. v(dx) _ . and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.8. let the given Markov process (specified by the Px) be the claim surplus process of a compound Poisson risk process with Poisson rate 0 and claim size distribution B. .9 If X0 = 0. Thus (since µ = p = 1.0. Ei instead of P2. Example 4 .11 For an example of a likelihood ratio not covered by Theorem 4.. the structure of MAP's is completely understood when E is finite: 2and only there .St)g(Jt+s)I. Then we can write v(dx) _ /3eOxB(dx) = / (dx). is defined as a bivariate Markov process {Xt} = {(Jt.2.5. 0..3B[B]. B have densities b.3 =. MARKOV ADDITIVE PROCESSES 39 Remark 4. b with b(x) > 0 for all x such that b(x) > 0).o[f (S8)g(J8)]. are the arrival times and U1. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.3 and claim size distribution B # B. abbreviated as MAP in this section2. Ei. one reason is that in parts of the applied probability literature. the corresponding claim sizes .(3B(dx). MAP stands for the Markovian arrival process discussed below.3 and claim size distribution B. where .10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . .0 in the following. 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 .g. it is then easily seen that Lt = H dB(Ui) i:o. Recalling that U1. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate .4). dB/dB = b/b when B.
. 1 J1 ='^])iJEE = (Fij[a])i . In addition.f.9 EE = (iii&ij[a])i j EE . oo). 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. v. {St} evolves like a process with stationary independent increments and the parameters pi.[a) = (Ei[easl.Sr_1.i. As a generalization of the m. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.. Fn[a] = F[a]n where P[a] = P .40 CHAPTER H. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. {Jt} is specified by its intensity matrix A = (Aij)i. If all Fij are concentrated on (0. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. Y1 E dx) where Y„ = S„ .jEE• On an interval [t. Y2. Jn = j. the converse requires a proof. (That a process with this description is a MAP is obvious. a MAP is the same as a semiMarkov or Markov renewal process. let {Jt} be standard Brownian motion on the line. the distribution of which has some distribution Bij. t+s) where Jt .. SOME GENERAL TOOLS AND RESULTS In discrete time.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i.J1=j)= Fij (dx) Pij In simulation language. vi(dx) in (4. with the Y„ being interpreted as interarrival times. As an example.) If E is infinite a MAP may be much more complicated. In continuous time (assuming Dpaths). . An alternative description is in terms of the transition matrix P = (piA.. consider the matrix Ft [a] with ijth element least Ei . which we omit and refer to Neveu [272] or cinlar [87].it = A. by generating Yn according to Hij when J„_1 = i.1 For a MAP in discrete time and with E finite.o(Ji = j.. Proposition 5.6) depending on i.g.
Jt = k] { 1 . \ diag Ft[a] = Ft[a]K. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. Then.2 Let E be finite and consider a continuous time Markov additive process with parameters A. aSt h = (1 + Ajjh) Ei [east .qkj + k?^j qkj Bkj [a] } = Ei [east. Jn+1 = A] = 41 Ei[ e 5„. j E E) and So = 0. Jt = j] is given by etK[a]. assume that the Markov chain/process {Jt} is ergodic. vi(dx). Jt = k] { xk kEE j la] .1) } (recall that qjj = 0). where K[a] = A+ (r. 013 . Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .(')(a)) diag + (). this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. By PerronFrobenius theory (see A. Sn ) yields Ei[easn+ '. u Proposition 5. up to o( h) terms. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . qij. a= . u In the following.4c). vi(dx) (i E E).1)) . we infer that in the discrete time case the .5. pi. Then the matrix Pt[a] with ijth element Ei [east.ijgij(Bij[a] . 00 r(i) (a) = api + a2ot /2 + f (e° . Bij (i. Jt = j] Ejesh'^ + E Ak j hEi [ease . which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. In matrix formulation . kEE Jn = k]Ek[e"Y" .1 )v(dx). MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.
t. Proposition 5.c(a) (and h(")).tK(a)h(a) J jj it L o is a martingale. cf.42 CHAPTER II.h(a)vva)etw(a).4c). (5. Jeast.3 Ei [east. .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). we are free to impose two normalizations. h(") may be chosen with strictly positive components. its derivatives are 'asymptotic cumulants'. Corollary 5. as will be seen from the following results. Yrh(a ) = 1. and we shall take V(a)h(a) = 1.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l .7. Jt = j] .r.4. and write k = k(°). Proof By PerronFrobenius theory (see A. Since v("). a.Eikjt = ttc'(0) + ki . In particular. Corollary 5. h(") are only given up to a constants. cf. We also get an analogue of the Wald martingale for random walks: Proposition 5. The function ic(a) plays in many respects the same role as the cumulant g.e=e°tk. Furthermore. Proof For the first assertion.4 Eie"sth(a) = h=a)et?(").5 EiSt = tK'(0) + ki . The corresponding left and right eigenvectors v("). and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . of a random walk. Eie"sth^a) = e'Pt[a]h( a) = e. Then h(°) = e. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). SOME GENERAL TOOLS AND RESULTS matrix F[a] has a real eigenvalue ic(a) with maximal absolute value and that in the continuous time case K[a] has a real eigenvalue K(a) with maximal real part.2) where 7r = v(°) is the stationary distribution. u Let k(a) denote the derivative of h() w.f. Corollary 5.
for a random walk: Corollary 5. summing and letting a = 0 yields E„ [St + 2Stkj.Eikjr . Remark 5 .7 No matter the initial distribution v of Jo. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.e. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) ..5.+ k.4. u The argument is slightly heuristic (e.") }) . Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.5 yields + W (a)k. E=ST = tc'(0)E7. (5. Squaring in Corollary 5. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.5. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. More precisely.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h. one obtains a generalization of Wald's identity EST = Er • ES. 8 Also for E being infinite (possibly uncountable ).2ttc (0)Evkjt + 0(i).g. (5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .6 For any stopping time T with finite mean. .a) + ttc (a)2hia ) Multiplying by v=. t a oo. subtraction yields Vary St = tic"(0) + O(1). Corollary 5.4) .3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. For the second . we differentiate (5. tam E tSt a (0). In the same way.. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . the distribution of Jo). [E. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).
inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. From (5. and the family {f LEE given by Theorem 2. where {Jt} is deterministic period motion on E = [0. see. this leads to h(i) + tcha( i.s.5) is a martingale . An example beyond the finite case occurs for periodic risk processes in VI. Usually. 0) = n(a) h(i). xEE . Then {Lt } is a multiplicative functional. in particular that f is bounded.10 Let {(Jt. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i)..(9) {Lt}t>o = . u forsEE).6. some extra conditions are imposed.3b and Remark VI. s) = ea8h(i).e. St) } as follows. however.9 The condition that (5. (5. however. 0 Proposition 5. St)} be a MAP and let 0 be such that h(Jt) OStt.4 that { h(Jt) easttK(a) L o (5.6) We shall not exploit this approach systematically. G is defined as Gf (x) = lim Exf (Xt) . 0) = h(i )( 1 + ttc(a)).for the present purposes. Remark 5. First.5.f (x) tyo t provided the limit exists. we take the martingale property as our basic condition below (though this is automatic in the finite case). Jt = (s+t) mod 1 P8a. this is.5 defines a new MAP. 1) (i. V.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. For t small . h(Jo) Lo is a Px martingale for each x E E.6. In view of this discussion . gha(i.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E.. let ha(i. Given a function h on E.
That 0 < qij < 1 follows from the inequality qb <1.. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. Ai = µi + 0Q.c(0)e = tc(0)e .(0)j. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.tc(0)e = 0 .Qi < oo and Bi a probability measure. Bi(dx) = Bi(dx). Bi. qij = r.1) holds for the P. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). Then the MAP in Proposition 5. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . if vi(dx) is compound Poisson. (5. in the discrete time case. 1 + q(b . 0<b<oo. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .7) In particular.10 is given by P = eK(e) Oh e) F[e]Oh('). then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].ic(0)e = ic(0)Oh e) h(e) .7(dx) Bij [0] Bij(dx) in the continuous time case .St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. ^i = of qij Bij [0] 1 + qij ( Bij [0] . In particular. 0<q<1. We omit the details.11 below in the finite case. 0 < qij < 1 and Bij [0] > 0.5.11 Consider the irreducible case with E finite. u Theorem 5. In the infinite case .1) . .1) eft ea' f ij (dx) = Hij (dx) Hij [0] . Bi [0] Remark 5. vi (dx) = f3 Bi(dx) with .. one can directly verify that (5.
Hence the same is true for H=j and H. H1.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A.13) for matrixexponentials . First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . Now we can write K[a] =A+A ) ( K[a + 0] .11.t.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() .8. it follows that indeed the normalizing constant is H1 [0]. a = 0 in (5. Here the stated formula for P follows immediately by letting t = 1... this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Yi E dx.8) h(. Jt = A. Similarly.tc (') (0) corresponds to the stated parameters µ. this implies k[a] = A 1 ) (K[a + 0] . Further Fib (dx ) = P=(YI E dx.8). since Hij. (dx) of a process with stationary independent increments follows from Theorem 4. v. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. .. Letting a = 0 yields the stated expression for A.46 CHAPTER II.r.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . (dx). Jl = j) = Ei[Lt. In matrix notation . This shows that F.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .tc(0)I. is absolutely continuous w.Bay [0]) That k(') (a + 0) . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. . are probability measures . v= .e) Consider first the discrete time case . Jl = j] :(Yi E dx. F:j with a density proportional to eei . in continuous time ( 5. Jt = j] = hie) .
Though the literature on MAP's is extensive. see also Fuh & Lai [149] and Moustakides [264]. .6.. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. h. and is typically defective.7). 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]. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7.6.3 for an infinite E are given by Ney & Nummelin [266]. For the Wald identity in Corollary 5. < x) = 11 (S. oo). is slightly less general than the present setting. Note that G+ is concentrated on (0. the literature on the continuous time case tends more to deal with special cases. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. [226] and Miller [260]. has no mass on (oo.Bij[0]) = hjel)ijgijBij[0](Bij[a] . [262] in discrete time. Conditions for analogues of Corollary 5. i.e.1) = Aij4ij(Bij[a] . which. 0]. there is.1). [225].+ < x.)Ajjgij(Bij[a+0] . hardly a single comprehensive treatment. 7+ < oo). 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). however. however. [261]. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5..(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 .
SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. see Fig. the ladder heights are i. it follows that for g > 0 measurable. To illustrate the ideas. they have a semiMarkov structure (but in complete generality. On Fig.(3B(x ) = pbo(x) on (0.T+ > t)dt = E f 0T+I(St E A) dt. In any case. Theorem 6 .e. oo).d. and the maximum M is the total height of the ladder. i. by approximation with step functions . 6.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. 0].48 CHAPTER K.e. the dependence structure seems too complicated to be useful). g(y)R+(dy) = E f g(St)dt.B(x) denotes the tail of B. the sum of all the ladder steps (if rl > 0. Thus. The main result of this section is Theorem 6.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. there are only finitely many). Also.i. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. The first ladder step is precisely ST+.1. 0 f T+ (6.. a fact which turns out to be extremely useful. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+.5 below..ST+(1) and so on. Recall that B(x) = 1 . we shall first consider the compound Poisson model in the notation of Example 1.1. 6. Here bo(x) _ B(x)/µB. In simple cases like the compound Poisson model. G+ is given by the defective density g + (x) =. the second ladder height (step) is ST+(2) . = ST+(1) Figure 6.1. o 00 (6. For the proof of Theorem 6. at present we concentrate on the first ladder height.00 ). which gives an explicit expression for G+ in a very general setting. 1 For the compound Poisson model with p = 01LB < 1. i. R+ is concentrated on (oo. has no mass on ( 0.2) .2. In other cases like the Markovian environment model. where basically only stationarity is assumed.
49 Proof Let T be fixed and define St = ST . 0 < t < T. P(STEA.ST_t.ST<St.2(a): T+ > t Figure 6.2(b): r+ < t Thus. . has the same distribution as {St}o<t<T. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. 0 < t < T) P(STEA. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . St S* t a Figure 6.2 R+ is the restriction of Lebesgue measure to (00. 6. ST < ST_t.O<t<T) = P(STEA.ST<St.0<t<T) = F(ST E A. That is.2.O<t<T).T+>T) = P(STEA. since the distribution of the Poisson process is invariant under time reversion. see Fig.6. 0].St<0.
s.St _). The probability of this given { Su}u<t is B(A . SOME GENERAL TOOLS AND RESULTS Integrating w. U + St_ E A.St). it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .St _)I(r+ > t). for A C (0..T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . . cf. we get G+ (A) = f 00 /3 dt E[B(A . oo). this is just the Lebesgue measure of A.r.3 Lemma 6 . Fig.2) in the last). T+ > t] 0 _ /3 f E[B( A . Figure 6.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS.3 where the bold lines correspond to minimal values. E A} precisely when r+ > t. and (6. That is. G+(A) = Q f 0 B(A .. and since the jump rate is /3.50 CHAPTER II. 6. But since St 4 oo a.t dT.y) (here we used the fact that the probability of a jump at u t is zero in the second step. oo).3 G+ is the restriction of /3R+*B to (0.
i. 2. this does not depend on h).1. the first component representing time (the arrival time o. h]} /h (by stationarity.e.z)B(dz) _ f I(x < z)B(dz) _ f (x).6. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6..4) are (ak. Uk) for those k for which ak .. The traditional representation of the input sequence {(TT. {St+8 . In the stationary case. Fig.4). Nt St =>Uk k=1 t where Nt = max{k = O. i. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. oo) x (0.M o 08 shifted by s is defined the obvious way. cf. . Uk) (k = 1.S8 )t> o = {St }t>o for all s > 0.. 6 .. 6 . oo).s > 0). we define the arrival rate as E# { k : ak E [0 . this is equivalent to the risk process {St*} being stationary in the sense of (6.Q f r+(x . Lemma 6.1 With r+(y) = I(y < 0) the density of R+.s.* ) and the second the mark (the claim size Uk ). 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.T+ < oo).e. as a point process on [0.) where ak = Ti + • • • + Tk . 4 (the points in the plane are (ak . . 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.. U k)} k=1 a is as a marked point process M *. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0.3 yields g+ (x) = . 0 Generalizing the setup. The marked point process .:T1 +•••+Tk <t}. assuming basically stationarity in time and space. obviously. The points in the plane (marked by x on Fig.
i. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. and let T = T2 denote the first proper interarrival time . The two fundamental formulas connecting M* and M are Eco(M) = aE E. most often one takes h = 1). We represent M by the sequence (Tk.5) does not depend on h. V(M* o eak ).52 CHAPTER II. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . h. vi(dx) = . h] and the sum approximately ^o(M*)I(ul < h). e.e. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. letting h J.QiBi(dx). k: vk E [0.s. Example 6 . Oh becomes the approximate probability F(ri < h) of an arrival in [0.g. h] Eco(M*) = 1 E f co(M o Bt)dt.2. Sigman [348] for these and further aspects of Palm theory.. the r.5) represents the conditional distribution of M* given vi = 0.. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.. where TI = 0.4 Given a stationary marked point process M*.. SOME GENERAL TOOLS AND RESULTS M* U. Uk) k=1. .4 Consider a finite Markov additive process (cf. Section 5) which has pure jump structure corresponding to pi = a = 0. o. See. = 0 . where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. As above . This more or less gives a proof that indeed (6. 0. of (6.
THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. First choose (Jo_. If Jt_ = i. and by some additional arrivals which occur w.*(0) with initial reserve u = 0 is p = /3EU0.p. . the probability aij of Jt . 5 Consider a general stationary claim surplus process {St }t>o. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.5. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. Jo) w.s./.= i. A stationary marked point process M* is obtained by assigning Jo distribution Tr. an arrival for M* occurs before time t + dt w. After that. we note: Corollary 6. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . Before giving the proof. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = .OF(x).p.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . and that p = 0EU0 < 1. Assume that St * . let the arrivals and their marks be generated by {Jt} starting from Jo = j. 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 .e.oo a.OEU0. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. v.O for i # j.O for i = j and iriAijgij/.O fo "o F(x)dx = . qij when {Jt} jumps from i to j and have mark distribution Bij.p. Jt = j is iri(3i /.. aij for (i. let U0 be a r. the ruin probability . Note in particular that the Palm distribution of the mark size (i. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .6. It follows that we can describe the Palm version M as follows .6 Under the assumptions of Theorem 6.
g.Su_ <0. Proof of Theorem 6. .. Now conditionally upon At . the mark at time Qk is denoted by Uk. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. 2.o.1] here the r . T+ = t given the event At that an arrival at t occurs . It follows that for A C (0.$St_ u. oo ) and the arrival times 0 > 0_1 > a_2 > . Let p(t) be the conditional probability that ST+ E A. in (0. are point processes on (oo . and the kth preceding claim arrives at time t ..St<Su. that M* and M have doubly infinite time (i.. has a very simple interpretation as the average amount of claims received per unit time . { Su}0<u<t is distributed as a process {Su} . ..e.l. moves down linearly at a unit rate in between jumps and starts from S0 = U.o. oo) x (0 . The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0.0<u<tIAt) = P(St EA.5. CHAPTER H. A standard argument for stationary processes ([78] p. We then represent M by the mark (claim size ) Uo of the arrival at time 0. which makes an upwards jump at time . . oo) p(t) = P(St EA. 0). 6.5.0<u<t) = P(StEA.Mt).Su< 0. (k = St}t>o 1. in (oo.St <.0<u<tIAt) = P(St EA..5).Q_k and has size U_ k. h..A. the arrival times 0 < 0'1 < Q2 < . The sample path relation between { Su } and { Su } amounts to S„ = St . 105) shows that one can assume w.).s. The last property is referred to as insensitivity in the applied probability literature..54 By (6. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. oo)). 0<u<t) = P(St EA.St*_ u.o<u<t where a claim arrives at time t and has size Uo..0<u<t) = P(St EA.Su<0.(left limit) when 0 < it < t and is illustrated on Fig .
2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. Mt)dt = i3EL(A) o"o . and since by assumption St * oo a. 6. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6.6.s. t a oo. Since So = U0. NIt)dt . and we let L(dy) be the random measure L(A) = fo°° I(St E A. Fig. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. Uo].. A sample path inspection just as in the proof of Lemma 6 . G' (A) = 3 f P(St E A. Thus. In Fig. 0 < u < t } is the event that { Su } has a relative minimum at t . cf. the support of L has right endpoint U0.5 where it = { St < Su. 6.5.5 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. the left endpoint of the support is oo.
5 is due to Schmidt & coworkers [48]. A further relevant reference related to Corollary 6. . [263] (a special case of the result appear in Proposition VI.1).6 is Bjork & Grandell [67].2. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6.56 CHAPTER II. [147].
.6) and simulation methods ( Chapter X).Chapter III The compound Poisson model We consider throughout this chapter a risk reserve process {Rt } t>o in the terminology and notation of Chapter I.. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. 3). 4. Some possibilities are numerical Laplace transform inversion via Corollary 3. i. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . • the premium rate is p = 1. i. see Chapter IV. Panjer's recursion ( Corollary XI. and assume that • { Nt}t>o is a Poisson process with rate j3. are i. For finite horizon ruin probabilities . say. with common distribution B. Thus . St = uRt = EUi t. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. • the claim sizes U1.4 below . U2.. and independent of {Nt}.e. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. 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 .d. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 .
6pBa).)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. and Schlegel [316]. 0 . Proof It was noted in Chapter I that p . for (d) just note that the kth cumulant of St is tic(k) (0). of the claim surplus St . For (c). say stable Levy motion.g. Schmidli [319]. See e.1) = t(p . The same method yields also the variance as Nt Ne Nt Var St = Var E Uk = Var E ^ Uk Nt +EVar [ k=1 k=1 1 k=1 Uk Nt Var [Ntµs] + E[NtVar U] = 113µs + t13Var U = tf3pB2). (d) The kth cumulant of St is tf3p(k) for k > 2.1) . Furrer [150]. (c) Ee8St = et" (8) where c(s) = f3(B[s] . THE COMPOUND POISSON MODEL Poisson risk process. cumulants . and this immediately yields (a).1 (a) ESt = t(13µ$ .t = fltpB .1 is the expected claim surplus per unit time. A more formal proof goes as follows: Nt r Nt ESt = E > U k .Rt.u .+Uk)P(Nt = k) k=O e8t k=O B[s]k .1). 1 Introduction For later reference.t = t(p . {Bt} a Brownian motion and {Rt} a pure jump process. we shall start by giving the basic formulas for moments.1)..f. e . We do not spell out in detail such generalizations. and that B(k)[0] = Pak). Write pB^) = EUn' YB = Pali = EU. [324]. (b) Var St = t.t = E E [ U k k=1 k=1 Nt . where K(k) (0) is the kth derivative of is at 0.'s etc. Dufresne & Gerber [126].t = E[Ntµs] .g. we get Ee8st = 00 e8t c` Ee8 (U1+. P = PAB = 1/(1 + rl) Proposition 1.. m.s.58 CHAPTER III.
we have Sok ... u + v]. and the value is then precisely v.. Proof We first note that for u. (d) If 17 = 0. For example. In this way.3 If nh < t < (n + 1)h.Tk are i. meaning that the increments are stationary and independent.1)th claim. cf. we get a discrete time random walk imbedded in the claim surplus process {St}.Sok_l = Uk . where Tk is the time between the kth and the (k .1 is the same as if {St} was a random walk indexed by t = 0. S„+V > S„ .1.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. The right hand inequality in (1.Tk. so that {Sok } is a random walk with mean EUET = EU. lim supt. which is often used in the literature for obtaining information about {St} and the ruin probabilities. however. The point of view in the present chapter is. then St> SnhV>Snhh. rather to view {St} directly as a random walk in continuous time. We return to this approach in Chapter V. 1. obviously 0(u) = F(maxk Sok > u).. (b) If 77 < 0. In particular.d.1 = . For the proof.3EU01 = 1µs where rt is the safety loading.h < St < S(n+1)h + h. St = oo.4. St = oo.S„ attains its minimal value when there are no arrivals in (u. (a) No matter the value of 77. if t = nh + v with 0 < v < h.i. Here is one immediate application: Proposition 1. then Snh . the Uk . then lien inft. 2.3) is proved similarly. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. The connections to random walks are in fact fundamental. we need the following lemma: Lemma 1.V. Sn+0 . Obviously. then St 00. (c) If 77 > 0. . II. v > 0. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. then St 4 co. Indeed.
However. For any fixed h. The general case now follows either by another easy application of Lemma 1.. Considering the next downcrossing (which occurs w.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). where the size of the portfolio at time t is M(t). and hence by the strong law of large numbers.. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. 169) stating that lim infra.2: Proposition 1. hence by induction i..1.6 Often it is of interest to consider size fluctuations. at least once. Snh/n a4' ESh = h(p . {Snh}n=o. Thus using Lemma 1. or by a general result on discrete skeletons ([APQ] p. and < 1 for all u when 77 > 0. it is seen that upcrossing occurs at least twice. 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 . is a discretetime random walk for any h > 0.60 CHAPTER III. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].5 The limiting distribution of St .p.3. ._... h A similar argument for lim sup proves (a).1(b)) that the assertion holds as t 4 oo through values of the form t = 0.3. Notes and references All material of the present section is standard.4 The ruin probability 0(u) is 1 for all u when 77 < 0.t . if P(M > 0) = 1. This contradicts u St400.1). this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt. u 307).1. h.2. lim supn_. 1 since St 4 oo) and repeating the argument. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.2. There is also a central limit version of Proposition 1.o. THE COMPOUND POISSON MODEL Proof of Proposition 1.1.s. {Snh}n=o. If rl > 0. Corollary 1. is a discrete time random walk. and (b).... p. then {St} upcrosses level 0 a.. 0 Snh = 00. (c) are immediate consequences of (a). Snh u = 00 (the lemma is not needed for (d)). Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. it suffices to prove 4'(0) = F(M > 0) < 1.. 2h. Remark 1 .
we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .2. oo ). n=0 (2. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. cf.1. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ).6. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). the ladder heights are i.1.34 or A.1) is not entirely satisfying because of the infinite sum of convolution powers. p < 1. but we shall be able to extract substantial information from the formula. B(x)/aB. 0 Alternatively. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. The expression for g+ was proved in Theorem 11.1) representing the distribution of M as a geometric compound. 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}. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. 11. nevertheless. We assume throughout rl > 0 or. Note that this . The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . equivalently. i. As a vehicle for computing tIi(u). and we further get information about the joint conditional distribution of the surplus and the deficit. 1e. Theorem 2. (2.e.IIG +II)EG+ . which we henceforth refer to as the PollaczeckKhinchine formula.just before ruin is again B0. IV. Combined with i/i(u) = P ( M > u)..1 provides a representation formula for 0(u). Summing over n.P) E PnBon(u) . Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . Thus . 1 The distribution of M is (1.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). [APQ] Ch.. d. Fig.6. the formula for the distribution of M follows . This follows simply by noting that the process repeats itself after reaching a relative maximum. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. It is crucial to note that for the compound Poisson model.
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). cf. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections.2(a) is from Dufresne & Gerber [125]. However.V)W) where V. where it requires slightly more calculation. cf. [62]. .d.5. Again. Theorem 2. cf.5. In the risk theory literature. ST(o)) is given by the following four equivalent statements: B(z) dz. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. V is uniform on (0. 2 The joint distribution of (ST(o )_.just after ruin. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). Theorem 2 . Theorem A1.2 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. We assume rt > 0 throughout. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.62 CHAPTER III. see for example [APQ]. As shown in Theorem 11. Beekman [61]. (d) the marginal distribution of ST(o)_ is B0. there is a general marked point process version. ST(o )) given r (0) < oo is the same as the distribution of (VW.6. 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. (1 . W are independent. Asmussen & Schmidt [49]. ST(o) > y. (c) the marginal distribution of ST(o)_ is Bo . The proof of Theorem 11. see Schmidli [323] and references there.6.1 is traditionally carried out for the imbedded discrete time random walk. Feller [143] or Wolff [384]. f +b (b) the joint distribution of (ST( o). ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. (a) 11 (ST(o)_ > x. in this setting there is no decomposition of M as a sum of i.i. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup.
use Laplace transforms.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..3 so that the conditional distribution of M given M > 0 is exponential with rate S '3 and 0(u) = P(M > u) = P(M > 0)P(M > uIM > 0) = pe(6Mu. Alternatively.1)1 00 ( 1 .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. hence without memory. As shown in 11.p)pSe a ( l v)x = p( S . Thus .3. Thus r(x) = S(1 . then. The result can. the formula for P(O) holds in a more general setting.e. the result follows . however . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . But claims are exponential .O)e(b0)x.1 e ax = n1 (n . a further relevant reference is Bjork & Grandell [67].p) = S . B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . 0 .p.6.p) E pn S n x n. 3b Exponential claims Corollary 3. 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. I. Integrating from u to oo. 1 . also be seen probabilistically without summing infinite series .2 If B is exponential with rate S.p. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. and hence this overshoot has the same distribution as the claims themselves . For a failure at x. Let r ( x) be the failure rate of M at x > 0.
then 24 1 V.+ <U.4) can be derived by elementary algebra from (3. and weights 1/2 for each.y)/3B (y) dy.y)G+(dy ) = f U V(u .T+ <oo). (3.p + G+ * Z(u) = 1 . (Example VIII.T+ <oo)+P(M> u. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. II.y)G+(dy) For the last identity in (3. u + oo. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). just insert the explicit form of G+. (3.3) below.i(u) satisfies the defective renewal equation Z(u) = 1 . u .s. Then the first term on the r.64 CHAPTER III.1 p pBo(u). (u) 35eu + 35e6u. Corollary 3.3).2) Notes and references Corollary 3. (3.h. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. 0 Proof Write o (u) as P(M>u) = P(S. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .3) Equivalently. E.+ <u. A variety of proofs are available .4) zu P(M > u .1) For a heavytailed B.g. The case of (3. 2 is one of the main classical early results in the area.3)).1. THE COMPOUND POISSON MODEL In VIII.3. T+ <00) (3. cf. (b) use stopped martingales .4) is similar (equivalently.y)f3 (y) dy.p + f u Z(u .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 . is ?7+ ( u). (3. the survival probability Z(u) = 1 . and conditioning upon S.S.2).S. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.3.+ = y yields P(M>u.+ >u.
The approach there is to condition upon the first claim occuring at time t and having size x .5 can be found in virtually any queueing book.f.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . Also (3. 206207).PPB2) EM2 = PPB) + QZPBl 2(1 . Of course.(3B[s] 1 . for example. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. eau B(u) du = f PB 3PB SPB 0 o (3.5 The first two moments of M are 2 EM .p) E p"Bo[s]" = 1 . We omit the details (see. In view of (3. g.P)pB' (3.p)s s /3 .5).pBo[s] no (1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.3 is standard . . Embrechts.5) Proof We first find the m. (3.3 .4) can be derived by elementary but tedious manipulations. 0 Notes and references Corollary 3.3 . [APQ] pp. 191). e./3B[s] which is the same as (3.7) and Corollary 3.p = (1 . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.Ps s(..8) Proof This can be shown. In fact. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.7).p)s . Corollary 3.5)./3B[s] . see e. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. [APQ] pp.)3B[s]) (3. Bo of B0 as m e8u B(du) = B[s] .g.1 Bo[s] = f oc.7) s +.(3 .Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .P)PB 2(1 . by analytical manipulations (L'Hospital's rule) from (3.p)2 3(1 .s .g..s .3. 111112 or Feller [143].6) 00 = (I . Griibel & Pitts [132]. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. Some relevant references are Abate & Whitt [2].
For n < u < n + 1. Assume (3. then p) 1: ep(k u/.u) a)Qea" + (1 .1).)3(1 .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.9) shown for n .4) for Z( u) means f lhu Z(u) = 1.1)! k=1 u1 .Q) 3e.u)]k k! k0 The renewal equation (3.3+ 18+ J0 Z(uy).1).u)]k1 ku+1) [/3( k . Z^ =eR(k.u)]k d 1 u) _ a) n ( du ( k! (1  .Q (k 1 k= n  [O(k . differentiation yields Z(u) _ /3Z(u) . we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1.9) follows for 0 < u < 1.h.9).u) [p(k .Q) k=0 k! E e0( = /32(u) .z/'(u) takes the form Z(u) L^J L. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.6 If B is degenerate at p.s. of (3.1 < u < n and let Z(u ) denote the r.3I( 0<y<1)dy Z(y)/3I(0<u.66 CHAPTER III.u) [N(k . ./32(u .u + 1 )]k = QZ(u) .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3. eO('u) [)3(k ./3Z(u .
or equivalently BB[a] = B[^+ Repeating for t 54 1. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.f. The answer is yes: inserting in (4.4.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. co(a) = rc(a + 9) . of F9. B9(dx) = B[9] B(dx).f. it follows that Z(u) = 2(u) for n<u<n+1. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.1 that c(a) = /3(B[a] .r. F and c. and thus (4.d.a.2) (Here 9 is any such number such that r.3) by t. in terms of the c. Formalizing this for the purpose of studying the whole process {St}.g. say t = 1: recall from Proposition 1. 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) . 0 Notes and references Corollary 3.(9) is welldefined.4. 4 Change of measure via exponential families If X is a random variable with c. We could first tentatively consider the claim surplus X = St for a single t.g.2) shows that the solution is Ox [O]0].4) works as well.1) . (4.rc(9) = . we just have to multiply (4.(9). K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.g. (4. and define rce by (4.3B = .2).Qe(Bo[a] .3B[9].f.1) . (4. but will now be repeated for the sake of selfcontainedness.1) or equivalently. we set up .f. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). The question then naturally arises whether ie is the c.4) .a.
(4.Tic (0)} . Let FT = o(St : t < T) denote the o•algebra spanned by the St. v(Xi.. for G E FT. (4.FTn) = Q(SkTIn : k = 0.1) and multiply from 1 to n).i.8) By standard measure theory.3 and claim size distribution B. THE COMPOUND POISSON MODEL Definition 4. BB by (4. But let Xk = SkT/n . EeeBSt + tk(B) = 1.5) for the density.S(k_1)Tln. The following result (Proposition 4.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .0e and claim size distribution Be.2.(9)} (4. Xn). Then the Xk are i.f.1. the PBT) are mutually equivalent on. replications from Fe (replace x by xi in (4. . G].t. . and dP(T) dP^T) That is. and thus (4. (4. with common c.r. t < T. .68 CHAPTER III. then EBZ = E [Ze9ST _T"(9)I . Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. Ti(a)/n.i.1 Let P be the probability measure on D[0.r.5) for the density of n i. and define 09. (4.d. G]..9) Proof We first note that for any fixed t.2 For any fixed T. Proposition 4. The identity (4. oo) governing a given compound Poisson risk process with arrival intensity.4). n) for a given n. the corresponding expectation operator is E9.7) Proof We must prove that if Z is FTmeasurable. .. = exp {BST .nr. and PBT) the restriction of PB to FT. G C {T < oo}. (4.FT.10) . Z is measurable w. in particular the expression (4. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .t.8) follows by discrete exponential family theory.d.g. . .3 Let T be any stopping time and let G E FT. it suffices to consider the case where Z is measurable w.
.1 It is seen that typically) a ry > 0 satisfying 0 = r.FT]] = EB [exp { BST + Trc(9)} I(G)] .f.1) . c(a) is a convex function of a. t = T .(Y) = 13(B['Y] .r)rc(9)}I . so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. Given FT.7 1 Some discussion further supporting this statement is given in the next section. Then GT = G n Jr < T} satisfies GT E FT. Thus by (4. 5. The behaviour at zero is given by the first order Taylor expansion c(a) r. 77 Thus.5. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.g. GT C_ Jr < T}. (4. (0) + rc'(0)a = 0 + ES1 a = a (p .10).1(a).. Thus. Letting T t oo and using monotone convergence then shows u that (4.7) holds.r is deterministic. the typical shape of rc is as in Fig.1) _ 1 + a. according to what has just been proved. 5 Lundberg conjugation Being a c.ST) + (T . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T. Now consider a general G. Then G E FT. and hence (4.9) holds with G replaced by GT.9) holds for G as well. subject to the basic assumption ij > 0 of a positive safety loading.
1(b).3. e.a = i(a + 7).1) (or (5. (5. the Lundberg exponent. It is then readily seen that the nonzero solution of (5.2)) is 7 = 5/3.2 s As support for memory. an equivalent version illustrated in Fig. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. Thus.s).3.3. u It is a crucial fact that when governed by FL.1 Consider the case of exponential claims. . we write FL instead of F7. Thus B[7] = 6/.1) is precisely what is needed for one of the terms in the exponent . Note that KL (a) = /L (BL [a] . 5. THE COMPOUND POISSON MODEL exists . (5.g.2 is B(7) = 1 + ^.4) yields /3L = b and that BL is again exponential with rate bL =. An established terminology is to call y the adjustment coefficient but there are various alternatives around. Fig.1) .4) ELS1 = #L(0) cf.QL instead of /37 and so on in the following . (5. 5.1(b). Example 5 . we further note that ( 5. G = {T(u) < oo} in Theorem 4. Equation (5.70 CHAPTER III. and (4. 5. Fig. b[s] = 5/(b . Taking T = r(u).2) 7 Figure 5. the claim surplus process has positive drift > 0.3) cf.1) is known as the Lundberg equation and plays a fundamental role in risk theory .
PL ) with density 1 . Proof Just note that e(u) > 0 in (5. V)(u) < e7u. (5.5).3.G+L)(x)) dx ry^+L) J 00 f 0 (1 . e(u) has a limit i.1.4).t.1e.6 ).6) Proof By renewal theory. 0 Theorem 5 .ascending ladder height distribution and µ+ its mean. take first 0 = ry.+ E A} in Theorem 4. T = T+.r. ST+ E A] .7) is the same as (5.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.5) Theorem 5 .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. To this end. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. Since a7' is continuous and bounded.5.8) . (5.7) 0 and all that is needed to check is that ( 5.Ce7u as u 4 oo.3 takes a particular simple form.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . where C . we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . we can rewrite this as 0(u) = e"ELe7^(u). Then P(ST+ E A) = EL [exp { 7S?+} . V) (u) = P(T(u) < oo) = EL [exp {ryS.e7x)G+(dx). T(u) < oo] . see A .(u)} .1p . G = {S. (5.P Y j o' xeryxOB (x) dx /3k [Y] .(oo) (in the sense of weak convergence w.2 (LUNDBERG'S INEQUALITY) For all u > 0. Letting e(u) = ST(u) .1 (5.
(5.12) Example 5 . From this it follows.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. 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.72 CHAPTER III.4 Consider first the exponential case b(x) = Seax. THE COMPOUND POISSON MODEL In principle.4).10) VW = JI c* e° (x) dx = a (B[a] .")G + (dx ) = 1  J0 00 3B(x) dx = 1p.1) (5. Noting that SIG(L)II = 1 because of (5. but some tedious (though elementary) calculations remain to bring the expressions on a final form.8) yields +L) J0 xel'B ( x) dx (5. that 7 = S . Then 0(u) = pe(a_Q)u where p = /3/S.1 above) and that C = p.7). this solves the problem of evaluating (5.3 (this was found already in Example 5.1 . of course. we get L where 00 (1 .11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.e. Using (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . . or equivalently of how close the safety loading 77 is to zero. u .
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)
.
Mathematically. Proposition 1. but has an obvious interpretation also in risk theory: on the average.3 )1 } _ 1p 1 . we shall represent this situation with a limit where /3 T fl but B is fixed. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .s(/3max .1.p = (/3max 0)µB. That is. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.g.3* /S. Letting Bo be the stationary excess life distribution./3)PBo PB . p* _ . Though of course it is based upon purely empirical grounds./3)] 1 .(bA*)".P .80 CHAPTER III. [174]) shows that it may produce surprisingly good results. we have according to the PollaczeckKhinchine formula in the form (3. Notes and references The approximation (7.1 As /3 f Nmax.1. numerical evidence (e.(3)2 P PB 2µB 2µB Letting /3* = /3/P./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . or equivalently that /3 is only slightly smaller than /3max = 1/µ8. the approximating risk process has ruin probability z.PBo [s (/3max . cf. Grandell [171] pp.2) was suggested by De Vylder [109].3 and Corollary 3.)3 )PBo µB  . (/3max .8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . and hence the ruin probability approximation is b(u) e(bAln)u.Ps(/3max . heavy traffic conditions mean that the safety loading q is positive but small.p + p { 1 1p ti 1 .p . Proposition 7.b(u) = p*e. the premiums exceed only slightly the expected claims. 1924.
2.2 If . the first results of heavy traffic type seem to be due to Hadwiger [184].ze a2unµB laB (7. In the setting of risk theory.3) is reasonable for g being say 1020% and u being small or moderate. Numerical evidence shows that the fit of (7./3)u). VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. u * oo in such a way that (3max . Of course. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce.p _ 2rl11B PB p. but has an obvious interpretation also in risk theory: on the average . 2 provides the better mathematical foundation. as well as it is needed for the interpolation approximation to be studied in the next subsection. 7d The light traffic approximation As for heavy traffic . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e .ryu . However . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic .4) suggested by the Cramer Lundberg approximation and Proposition 6. light traffic conditions mean that the safety loading rl is positive and large . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230].B AB ) 6()3max _'3) = However . [APQ] Ch.7.0)u. . we shall represent this situation with a limit where 3 10 but B is fixed. These results suggest the approximation Vi(u) e6(0_. obviously Corollary 7. the premiums are much larger than the expected claims . the term light traffic comes from queueing theory. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . This follows since rl = 1/p . VIII).1 1 . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . or equivalently that 0 is small compared to µB .p.g. and hence 2µ2B 1 . Mathematically./3)u * v. while the approximation may be far off for large u. That is .Q T /3max.l3)M > (/3max . light traffic is of some interest as a complement to heavy traffic .
u. i. En'=2 • • • = O(/32) so that only the first terms matters. 0(u) /3 J B(x)dx = /3E[U . For a more comprehensive treatment.3 is the same which comes out by saying that basically ruin can only occur at the F(U . Asmussen [19] and references there. Light traffic does not appear to have been studied in risk theory.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. by monotone time T of the first claim . 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. THE COMPOUND POISSON MODEL Proposition 7.T > u). Omax max m. (7.(3 10. cf.u)+.T > u) = J o" B(x + u)/3eax dx . Indeed. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. the Poisson case is much easier than the renewal case.e. u Note that heuristically the light traffic approximation in Proposition 7. 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 . The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 .= 1 aJ 1 a 0+ 1 = = p. . U > u] = /3iE(U . and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. 10 ( u The alternative expressions in (7. [97]. Sigman [347]. see Daley & Rolski [96]. ( 3 J O B dx. z/' (u) convergence P(U .3 As .82 CHAPTER III.5) follow by integration by parts. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. Again.5) u Proof According to the PollaczeckKhinchine formula.
Thus . . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. Al . we see that the following limits HT) (u').VHT) ( ax QmQ ) h (B) ( .O(E)(u) 1 (1 . no empirical study of the fit of (7.3n. one may hope that some correction of the heavy traffic approximation has been obtained. Notes and references In the queueing setting . 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. however. Substituting v = u(. available.O0 M. _(E) (u). ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. (U). with rate 1/µB = /3max. [84]. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless .6) is . Another main queueing paper is Whitt [380]. B(2). 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). 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). z/i(E) (u) = pe(QmaxQ)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. to get nondegenerate limits ./3)) .Wmax f(x ) dx + pee6mQ. even if the safety loading is not very small. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . 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. where further references can be found .3 and use similar notation for %(B) (u) = (u). . The adaptation to risk theory is new.x . ) M.3). that is.8. Instead. Let OLT) (u) denote the light traffic approximation given by Proposition 7.
Bill is said to be convexly smaller than B(2) (in symbols. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. B(') <i. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. we have the convex ordering. B(' <. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. cf.6. Proposition 8. then . equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. or the existence of random variables U(l). B(2)) if f fdB(1) < f fdB(2) for any convex function f. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.ill(u) < V)(2) (U) for all u. u Of course. Finally. B(2) and PB(1) = µB(2). In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. In terms of the time to ruin. THE COMPOUND POISSON MODEL To this end. we shall need various ordering properties of distributions. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. then Bill = B(2).84 CHAPTER III. we can assume that 1) < St 2l for all t. U(2) distribution B(2) and U(1) < U(2) a. 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). .1 If B(') <d B(2). whereas (consider x2) B(2) has the larger variance. then i. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x.s. U(2) such that U(l) has distribution B('). this ordering measures difference in variability. Rather than measuring difference in size.1 is quite weak.2 If B(') <j. In the literature on risk theory. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. Here convex ordering is useful: Proposition 8. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B.' 1)(u) < V)(2) (U) for all u. Taking probabilities. XI. the proof is complete. Proposition 8. Proof According to the above characterization of stochastical ordering.
B(2). Corollary 8.6 Fix /3 at 1/1. The problem is to specify what 'variation' means.8. larger variance is paramount to larger second moment. from which the result immediately follows. The heavy traffic approximation (7. . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. A partial converse to Proposition 8.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).(1) (. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB.. B(2). (D) (u) < O(B) (U ) for all u. A first attempt would of course be to identify 'variation' with variance.e.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. Example 8.4 Let D refer to the distribution degenerate at 'LB . Bo1) <_d Bo2) which implies the same order relation for all convolution powers.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. we have by Jensen 's inequality that E f (U) > f ( EU). and here is one more result of the same flavor: Corollary 8. then /'(')(u) < 0(2)(u) for all u. This u implies that D <. with fixed mean.2 is the following: Proposition 8. A general picture that emerges from these results and numerical studies like in Example 8. Proof Consider the light traffic approximation in Proposition 7. Proof If f is convex. then B(1) <. Then V.3 If B(1) <.3 provides another instance of this. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I.1. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.1 and µB at 1 so that the safety loading 11 is 10%. say to p. B.. Hence by the PollaczeckKhinchine formula .4) certainly supports this view: noting that.5 If '0(1)(u) < p(2) (U) for all u and a.
A2 = 3. B3: the Erlang distribution with density 4xe2x. For B1i B2. Kluppelberg [234]. i. In terms of variances o2.1%. all distributions have mean 1. Note to make the figures comparable.. [166]. 0.0' U0. One then obtains the following table: U005 U0. Pellerey [287] and (for the convex ordering) Makowski [ 252]. and this is presumably a consequence of a heavier tail rather than larger variance.9A2e'2r where A.4142. 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. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).e.) = a. We return to ordering of ruin probabilities in a special problem in VI.001 u0. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.e'\1x + 0.4. 11 Notes and references Further relevant references are Goovaerts et al. 1%. van Heerwarden [189]. 9 Sensitivity estimates In a broad setting. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.01%. B4: the Pareto distribution with density 3/(1 + 2x)5/2. B3 the comparison is as expected from the intutition concerning the variability of these distributions. Let ua denote the a fractile of the ruin function. in comparison to B2 the effect on the ua does not show before a = 0. the behaviour of which is governed by a parameter 9.86 CHAPTER III.000.1358. B. = 0.01%. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. 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).lA. A standard example from queueing theory is . 0. and consider a = 5%. 1/)(u. However.
Then if t is large . Then ib = Pe(613)u.1. and s(9) the expected sojourn time of a customer in the network. obtained say in the natural way as the empirical arrival rate Nt/t in [0. t]. a/3 0 .1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Let R(P) = Rtli.19P a/ . with 0 the vector of service rates at different nodes and routing probabilities. In the present setting. the premium rate p and the claim size distribution B.9. For example. the distribution of %3 0 is approximatively normal N(0„ Q/t).3. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. Then a p ao = 00 Qa/.01/2u..Ap). where the partial derivatives are evaluated at p = 1. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. say estimated from data. if = a e(6A)u. Similar conclusions will be found below. Then the arrival rate /3(P) for { R(P) } is )31p. Assume for example that 8 is known. Example 9.2 Consider a risk process { Rt} with a general premium rate p. i. while /3 = j3 is an estimate. In particular . Thus. Proof This is an easy time transformation argument in a similar way as in Proposition 1. where Q2 = fl ( l2 1113 / _ Ou2v)2. 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.(u) for large u.e. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . SENSITIVITY ESTIMATES 87 a queueing network. u Proposition 9. increasing in u. it follows that ' is approximatively normal N(0. a0 as ao 80 19P . a2/t). we may be interested in a'/ap for assesing the effects of a small change in the premium. Thus at p = 1.
g. (3+'y)PC (0+7. 4) (9 .3.88 CHAPTER III. but must look for approximations for the sensitivities 0. Differentiating w. /3 yields w e(e + Y.()^ 1 .6 below for some discussion of this assumption).w(O. In the case of the claim size distribution B. THE COMPOUND POISSON MODEL As a consequence. Viei '0(.5) are similar. 5) (Q+'Y)[we(0+7. Proposition 9.6) As will be seen below. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . Consider first the adjustment coefficient y as function of 3. and the proofs of (9. (9. However .Owe (9.0 = t/'(u) and the CramerLundberg constant C. Of course. 3) ( 9 . (9. Similar notation for partial derivatived are used below. 9. Consider first the case of 8/8/3: . () = log(1 + y//3).4).^)] 1(/3+y)we(9+'y.t.3 70 = 'Ye = = 7 /3(1 we(e +'y. but we shall concentrate on a special structure covering a number of important cases.2) (see Remark 9. x > 0 (9.3) follows by straightforward algebra. u Now consider the ruin probability 0 = 0 (u) itself. so that heuristically we obtain '00 50ryu = Coe"u .()wC(e. various parametric families of claim size distributions could be considered. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable.3. ()} p(dx) . this intuition is indeed correct. and write yp = 8y/8/3 and so on . for the ruin probabilities .(/3 + y)we(9 + 7.r. we can rewrite the Lundberg equation as w(9+ y. we cannot expect in general to find explicit expressions like in Example 9.10) below.()(0 +'0) ' (9 .((dx ) = exp {Ox + (t(x) .()YC = 1 +y/ /3 \ Q2 From this (9.3 or/and B.1 or Proposition 9. e. it suffices to fix the premium at p = 1 and consider only the effects of changing .w(6.uypCe7u urypO. (. namely that of a twoparameter exponential family of the form Bo. ^) . () Proof According to (9.
O} (9. Z= zl + z2 where zl (u) = e7u J m B(x)dx.4 As u oo. PF = (1 . () .St (U) Ee. ()} . F(dx) = e'yy/3B(x)dx. Be.([a] = exp {w(9 + a.3 (see in particular (5. ()} . () . the proof is complete.9) (9.w(9.8).2 of the Appendix ). it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.11) Ee. () .(e"U = = wS(O. By dominated convergence.3(x) dx. () exp {w(9 + a. we note the formulas Ee. SENSITIVITY ESTIMATES Proposition 9.4t (U)e°`U = which are wellknown and easy to show (see e.8) (Section 5). Further write de = [we (9 +'y.x)B(x) dx + J U W(u . But from the proof of Theorem 5.x) F(dx ) f u J C F(dx) = C as u 4 oo. Hence by a variant of the key renewal theorem (Proposition A1. z2(u) _ 1 ^ e'ri`i7i( u . Combining these estimates . z2(U) = e7" J u b(u . 0(u) = /3 Ju"O B(x) dx + f 0 0(u .x).we(9 .10) (9. and alsoo zl(u) + 0 because of B['y ] < oo. ()] exp {w (O + y. 11 For the following.9.3) for z/'(u).x). BarndorffNielsen [58]). w((9 + a. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. we multiply by e7" and let Z(u) = elt" cp(u).p)/C'y.w(9. () .8) Letting cp = e0/e/3 and differentiating (9. (9. u 0 Then Z = z + F * Z and F is a proper probability distribution . Z(u)/u a C//3PF where PF is the mean of F.12)). we get p(u) = J "O B(x) dx + J U O(u .x)B(x)dx.QB(x) dx.C).g.w(O.
())B(dy) dx.9)(9. u Z2(U) = e7° f u ^/i(u .wc (O.1) B(dy) 'f '[t(y) . oo z2 (u) f C .w (9. F(dx) = e7x. 0 x Multiplying by e7" and letting Z(u) = e"uV(u).11).90 CHAPTER III. this implies Z = z + F * Z.w(0. ^) .12) f exp {O y + (t(y) . 8^ ue7u. ()} 1z(dy) = f [t(y) .we (0. C)] (1 + 7 ) Proposition 9.QB(x)dx.x)f3 f ^[t(y) . )}B(dy)• Letting cp it thus follows from (9.5 Assume that (9.wc(O.6C do 89 1p 8( 1p Proof By straightforward differentiation. ()](e7v .6e7u f "o f[t(y) . 01 (i+) do = +'Y.w( (0.w(e.w( (0. ^)} [wc (0 + 7. ()]B(dy) dx.e7x/3 f 00 [t(y) . C) .8) that cp(u) . By dominated convergence and (9. 2 z 07P N ue7u (3C de . ()]B(dy) dx x 0 0C T ON O . z = zl + z2.wc(9. ()]e7vB(dy) 'fCd 7 c . THE COMPOUND POISSON MODEL [we(e+7. 8 8() 8( (9. Then as u > oo.2) holds.x).lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .
It follows after some elementary calculus that p = a)3/5 and.(log r(a) a log S)} • r(a) 1. () ='I'(t.QS 1 ./35' a/i'y + aryl 625ry.yu/3C2do u86 89 1p' az/) = 8z/.17) (9. ue_Yu 'C2d( 8a 8( 1 p . < = a.ry) 5a1 cry (5 .1 .2) holds with p(dx) = xldx. . Example 9. we (9.18) (05 + 57 _'3_y . SENSITIVITY ESTIMATES as u 3 oo.16) (9.pa+1 . Here (9.. U 7µF from which the second assertion of (9. w(e. ())B(dy) < oo.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.6 Consider the gamma density b (x ) = Sa xa.a/35a&y' ' (9. Z(u) /3C 91 o c'o e11(t (y) . ( 9..14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .rye) S 5rya.9.Sry a/32 + a/37 + /37 .12) follows. a /(S . We get w( (0. by inserting in the above formulas. and the proof of the first one u is similar..a log S = log r(c) . 9 = S.3ary tog('Finally.w((9. () = log r(a) .15) (9.Y)a+1 ' (9. () = C/9 = a/S.C log(9). that C = a.13) (9. t(x) = logx.12) takes the form y) alp a .1edz = 1 exp {Sx + a log x . and also zj (u) 4 0 because of f Hence.
which we omit in part .2 .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. w(e.CZ try)} 1 C C2 try .w(9.9) (() . further yield . Be.2a) } Thus the condition B[a*] > 1 + a* /. ()} = exp {c (C .1 = eXP {c(C .21og 2.92 CHAPTER III. THE COMPOUND POISSON MODEL Example 9. t(x) _ .S[a] = exp {w (9 + a. C = . C) = B = Yc = de = do = .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.([Y] .2) with µ(dx) = 2x3zrdx.1 16 +ry c C22ry 2( = + 70 We (e. C) .3.3Ee. Straightforward but tedious calculations . () = Cc .2 log (0.22."62 . 9 = . for a < a* = z (.log c = 2 In particular.
[379] consider a special problem related to reinsurance. sj=1 and let YT be defined by IKT('ryT) = 0..a. BT [a]= NT ^` e"U. Note that if NT = 0. Van Wouve et al. Thus. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). 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). ae t 1lEY u S _ .2 of the parameters. . and we estimate y by means of the empirical solution ryT to the Lundberg equation. for which we refer to X. let NT 16T = ^T . the main tool is simulation. and hence explicit or asymptotic estimates are in general not possible. kT (a) = /T (BT [a] . In general.oo.7 and references there. or Ct(x). to our knowledge. To this end.g. if 1 PT = /3TNT(U1+. thus. B are assumed to be completely unknown. we have assumed k = 2 and ti (x) = x. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.+UNT) > 1. However. Also. That it is no restriction to assume k < 2 follows since if k > 2.cue_7u)3C P Remark 9. Finally if k = 1. the results presented here are new. (9. we can just fix k . However .3C2de 1p' z a = c .. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T .12) takes the form a = a 93 ar.2) is motivated as follows. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. in which case we can just let t(x) = 0. then ryT < 0. queueing networks) are typically much more complicated than the one considered here.1) . the exponent is either Ox. Thus.10.8 The specific form of (9. Comparatively less work seems to have been done in risk theory. then BT and hence ryT is undefined. the models there (e.. in u which case the extension just described applies.
If .B[7]2 n Hence ( 10.B[7]2 }) ( T 0 . (10. B [7]2 (10. If furthermore B[27] < oo.2 As T * oo.1)2 + E[27] .b[Yp'V21 T { (E[7] .'s.: N 0.: N ()3. vfoVFB[2y].'Y .1) .B[7]2 V2 .(27)/K'(7)2.v.T y . 16T where V1. V2 are independent N (0. 7T a4' 7. THE COMPOUND POISSON MODEL Theorem 10 . a2 where a2 = /3r. since NT /T . N ( n[7]. 1) r.94 CHAPTER III. then (10.3/ T).B[7]) + B [7] .B[7]) 0+ Iv/o(b[y].a BT[7] I B[7] I + . More generally.2) follows from NT/T a4' . .1) .i3)(B[7] 1) + (3(BT[7]  .2) rT(7) N N (0.7 + (.1 As T 4 oo. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . B[2'Y]  /3T ) .3) Proof Since Var(eryU) = we have B[7]. Lemma 10 .1) 'YT ./^ B[27] . it is easy to see that we can write \ V1 1 l _ .3T ..Q and Anscombe 's theorem. B[27] .)vl+ N CO. For the proof. we need a lemma.
NT i =1 n'(a) for all a so that for all sufficiently large T K7 .3).1 can be used to obtain error bounds on the ruin probabilities when the parameters .'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. it follows that 7T7 KT(7T) .(ry + e) and hence KT(7 . first note that e7TU N (e7U u2e27Uo'2/T) 7 . °7IT) .E) < 0 < kT(7 + E) for all sufficiently large T . To this end .e. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.e) < 0 < r. y + E) eventually. where ryT is some point between ryT and ry. lcT(a ) 4 /c(a). BT[a] 3 B[a].1 By the law of large numbers.E) < 4T(7T) < 4T(7 + E). If ryT E (7  we have KT(7 .4) and Lemma 10. 7T E (y .'(y). and the truth of this for all e > 0 implies ryT at 'y. 0 are estimated from data .Q.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.e. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).10..c'(7) N (0' T (2(7) / N (0. Then r.2. Proof of Theorem 10.(ry . 6"Y (10. Theorem 10.KT(7) kT(7) K'(7) . NT BT [a] Hence r. By the law of large numbers. OT a 95 u 4 /3.4) + E). Combining ( 10. Let 0 < E < ry. I.
C1e"a ( see e. and the known fact that the Y„ = max Vt tE[W„1.. Wn).) = a (e. Csorgo & Teugels [95].info< „< t S.e. [197]. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. it means 2 (8 .1 is from Grandell [170]. = 1. wn = inf{t > W.. ft.T VIT where r7ry.f.d.3 or equivalently p > 1/2 or 11 < 100%. Frees [146].Q. > 0 for some t E [Wn_ 1.. t]}. 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. Notes and references Theorem 10.1 : Vt = 0.. Hipp [196]. U2.. i. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. For example . Herkenrath [192]. 6 < 2.e. For this reason . 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. the nth busy cycle is then [Wn1. i . This approach in fact applies also for many models more general than the compound Poisson one. .g. Letting Wo = 0.0) < 5. Deheuvels & Steinebach [102].ueryuU ".i. if B is exponential with rate 8 so that ry = 8 . Mammitzsch [253] and Pitts. Embrechts & Mikosch [133].. various alternatives have been developed. satisfies b(.5%).96 CHAPTER III.Wn) are i ...96 if a = 2. with a tail of the form P(Y > y) . Griibel & Embrechts [292].g. Vt = St . Asmussen [23]) can then be used to produce an estimate of ry. V.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc.
g.1 (the role of ryy will be explained in Section 4b). Further let 'Yo be the unique point in (0. See Fig. Unless otherwise stated. exists. The notation is essentially as in Chapter III.s. In particular. B[•] and mean AB. the premium rate is 1.f. 97 .1 where p = 13µB. 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.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .1) . 0. generalizations to other models are either discussed in the Notes and References or in relevant chapters. The safety loading is q = 1/p . 'y) where c(a) attains it minimum value. the Poisson intensity is 0 and the claim size distribution is B with m.
using that the overshoot l. we have for k = 1. 1 FL. 2 that E [T(u)k. the time of ruin is T(u) and ^(u) = ST(t&) .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. FL and independent of T(u).r.u is the overshoot. . In particular. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .5 .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.(U) < 00] = ELT(u)ke'YS.9).2) Proof Let as in Example 111. PROBABILITY OF RUIN IN FINITE TIME Figure 0.98 CHAPTER IV.(.1) (1. Var[T(u) I T(u) < 00] = VarL T( U) . PL = 6/0 = 1/p > 1). 7.1 The claims surplus is {St}.t. By the likelihood identity III.(4. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1. E[T(u) I T(u) < 00 ] = ELT (U).. (u) is exponential with rate 0 w. 1 Exponential claims Proposition 1.
2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.V/ is as asserted. 1).1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.2) is aLELT( u) ./3) .6. T(u) < oo] fora > r.B = a.1 (6)3)2 which is the same as the r.h.1//32 (6/)3 1)2 26(/3u + 1)/(6 .s.6a = 0 with solution 0 (the .12 Thus the l.1)) ELST(u) ELT(u) (PL . we have by Wald's identity that (note that ELSt = t(pL .I (1.1.1)T(u))2 = UL where = s.1)2VarLT(u) + 2 Ca 1I VarLT(u).1)T(u)) = VarLe(u) + (PL . where = eBu I 1 .s.(PL . .1)ELT(u).h. is V1rLSr( u) +VarL ((PL .(yo) = 2V .0) .h. Let 0 > yo be determined by ^c(0 ) = a.s. which leads to the quadratic 02 + (/3 ."(ry) = 26//32.2). EXPONENTIAL CLAIMS For (1 . This means that /3(6/(6 . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).1) .1)T(u) are independent with QL the same mean . Wald's second moment identity yields 2 EL (Sr(u) . of (1.3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 . Since Sr (u) and (PL .6 + a)0 . u + ELe(u) _ PL . the 1. 0 Proposition 1./3 .
M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . ... and M(u)+1 is the index of the ladder segment corresponding to T(u)..3) we have E [e«T(u ).Y1 Y2 Figure 1. T2 .1 . are the lengths of of the ladder segments 2. Ti..3 that we can write EeaT( u) = eeuEe 017(o). Y(u) belonging to a convolution semigroup .9ST(u) +T(u)!c(0)} . the result follows. Cf.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 ) .. Note that it follows from Proposition 1. (1.OuEee 04(u) = ee u be BB+B where we used that PB(T(u) < oo ) = 1 because 0 > ryo and hence E9S1 = K'(0) u > 0. More precisely..T+ Ti a t U T I 1 a i F. 1.1 where Y1. Using 5 = 6 . St Ti F.v. Y2.. T(u) < oo] = e.v.0. are the ladder heights which form a terminating sequence of exponential r.'s with rate 5. . Fig. But by the fundamental likelihood ratio identity ( Theorem 111.3.100 CHAPTER IV. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).4.
. the conditional distribution of VT given QT = N is that of EN where the r.i. .T. Since U1 .1. including the customer being currently served). the following formula is convenient by allowing t.3 Assume that claims are exponential with rate b = 1.4.e.T + • • • + UN. UN... let (cf. . Proof We use the formula .v.T) 1 I fl(O)h(0) fdO where (1. Corollary 11.T is the residual service time of the customer being currently served and U2 . Let {Qt} be the queue length process of the queue (number in system. and exponential with rate S = 1.T.d..1(u. i. 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 .1 )!..6(u) = Vfl/j l(Su. density xN lex/(N . U2. 1). cf. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . EXPONENTIAL CLAIMS 101 For numerical purposes .cos (u/.3 sin0 + 29) f3(0) = 1+/32/cos9. Then V(u.T are conditionally i.T the service times of the customers awaiting service . where U1.. If QT = N > 0.ST). Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. T) to be evaluated by numerical integration: Proposition 1. .T. 2. . T.I ex cos B cos j O dB fo " . .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. For j = 0.k + 1). EN has an Erlang distribution with parameters (N.1. UN.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .6.i (u.0. then VT = U1..
(1.44). PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.38).3(k +1)/ 2ei(k + l)6 (. in particular equations (1.102 CHAPTER IV. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .1)] L _112 /(k+1)/2 [.13(k +l)/2ei(k +1)9 R E .1 R [.)3k +1 tj g'(QT >.1 00 ok+lR 00 j=k1 +1)/2e . similar formulas are in [APQ] pp.(31/2 cos (( k + 2)9) . 8789) 00 E aj j= 00 = 1.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .31 /2eie L 1)] 1 I/31/2eie .8 ) yields F(QT > k + 1) . k k2 + $k+1 E bj 00 t j . Then (see Prabhu [294] pp.(31/2eie .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .cos (( k + 1)0)] f3(9) Hence the integral expression in (1. f3(0) .ie . let I _ j (x) = Ij (x).i(k +1)e R [/3( klal/2e:0 (01 /2 e . 912. and define tj = e(1+R)Taj/2Ij(2vT T).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 .cos((k + 2)9)] d9.
0(u.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. there are several misprints in the formula there. u Notes and references Proposition 1. k! k=O k0 i/z Co Uk ate" o'/z e . E Fk. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. going back to Cramer. however. it follows as in (1. Related formulas are in Takacs [359]. oo (u)31/2e^e)k = )3k z cos(k9) = R k. expresses V)(0. T). k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). t) = P . We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p.e = e' COS a cos(uf31/2 sin 0). 2 The ruin probability with no initial reserve In this section . is numerically unstable for large T. and the next one (often called Seal's formula but originating from Prabhu [293]) shows how to reduce the case u 54 0 to this.7) that _ [^ au ak+l (30 k L. Seal [327] gives a different numerical integration fomula for 1 . the numerical examples in [12] are correct)..2. t )). however. F(x. or. Ui < x I / (note that P(St < x ) = F(x + t. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. The rest of the proof is easy algebra. . The first formula. equivalently. T) in terms of F(.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. from the accumulated claim distribution N.
v]. Stv^ _ Define M(v.(.(6. 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. See Fig. [v.(0. 1 1 .0<w<t} St+v .3) with A = (0.T))dv E^T I(M(v.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.T) T F(x.T)) 1 fT P(M(v.1. ") } is at a minimum at time t. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . co ).T].T)) does not {Stv)} depend on v.104 CHAPTER IV.1 In formulas. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.T))dv. . Proof For any v E [0.t)= {Stv) < SM. Then 1 .i.b (0. 2. T) = P(Tr(0) > T) = P(M(0.T)dx. T]. f T lStv)} 0<t<T by a 'cyclic translation'. resp.
Fig 2. 0<t<v} = {ST < St . ST > 0.xdx.T)) dv f T I(M(0. T) occurs. It follows that if M(0 . t).. T) as {ST<St+ vS. T) = M(0. this integral is 0 if STv) . Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. 0<t <Tv}n{ST<ST Sv+St T+v. then M(v. T Theorem 2 . T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . in which case there is a last time o where St downcrosses level u.Tt))f(u+t. T].T) = F(u+T.v<t<T} = {ST<StSv.ST on M (0. T) occurs. we can take v E (w E. T)) dv.. Indeed. where the last equality follows from ST < St on M(0. v) = M(0. there exist v such that M(v. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0.T) and Sv < 0 on M(0.Sv.2 10(u. t) denote the density of F(•. T)). v). w) for some small E. letting w = inf It > 0 : St_ = mino<w<T Sw}. Obviously.T)f(I z /)(0. For example. Hence T TE f I( M(v. We claim that if M(0.2. T T o i =1 Let f (•. T. .T) occurs or not as long as ST < 0.2. v<t<T}n{ST<STSv+St. If ST < 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 . or it occurs. v).T) occurs. v < t < T} n M(0. v). then i fT I(M(v. we can write M(v. cf.t)dt.
b(u.T) = C(z. ST_ _ z}. u + dt] and there is no upcrossing of level u after time t. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).T)+ J0 T (1V.Tt))P(StE[u. 0 < t <T . Proposition 2. z > 0. The proof is combined with the proof of Theorem 111. ST_ _ z} . 0 < t < T. The following representation of T(0) will be used in the next section.z.t).2.p.T) = .3 Define r_ (z) = inf It > 0 : St = z}. Hence P(ST<u) = 1 .ST_ t_ and let A(z. define St = ST .2 . O(T . E [t. Proof of Theorem 111. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. Let Z be a r. {S t > z.2 Here o.u+dt]). ST_ _ z}. For a fixed T > 0.v. C*(z. {St > .106 CHAPTER IV. 0 < t <T.(0. 2.T) = {St < 0. which occurs w.2. t + dt] occurs if and only if St E [u. u which is the same as the assertion of the theorem. which is independent of St and has the stationary excess distribution B0.
THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.3 But by sample path inspection (cf. r(0) < oo) = 3R(z) dz JP(C(z. z + dz].T))f3B(z) dz dT. z + dz].3). z + dz].2.T + dT]. 2. Hence integrating (2. z + dz]. It follows by division by P(ST(o)_ E [z. T(0) < oo) = OR(z) dz in (2.3. Fig. (2.2.T). 7( 0) < oo) = P (C(z)) dT. {St }o<t<T have the same distribution . T + dT] I S7(o)_ E [z. z + dz]) = P(A(z.1) z T . Figure 2.1) yields P(ST(o)_ E [z. ST(o)_ E [z. . A(z. u which is the assertion of Theorem 111. we therefore have P(A(z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.T)).1) that P(T(0) E [T.ST(o) >y.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. 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.T)) = P(Cx.2. and since {St}o<t<T.T) = C*(z. Proof of Proposition 2.
I L Let ga(x) be the density of the measure E[ear(°). 2. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. Note that T_ (y) < oo a.T+dT]). T(0) < oo) = dTP(T_(Z) E [T.3 was noted by Asmussen & Kl(ippelberg [36]. [329]. 3 Laplace transforms Throughout in this section.108 Hence CHAPTER IV.1 Eear( y) = eyr(a). Lemma 3.1) where a > r. Lemma 3 . cf. see in addition to Prabhu [293] also Seal [326]. z + dz]. Proof Optional stopping of the martingale I er (a) 9 t.5 and one upon excursion theory for Markov processes (see IX. Tak'ecs [359].6.s. z + dz]. r(0) < oo. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. a martingale proof is in Delbaen & Haezendonck [103]. who instead of the present direct proof gave two arguments.r(a).T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.(yo).2. (3. some relevant references are Shtatland [338] and Gusak & Korolyuk [181].1.1 and the present proof is in the spirit of Ballot theorems.5a).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x .3. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. Let T_ (y) be defined as Proposition 2. In the setting of general Levy processes.(3(B[r( a)] . Notes and references For Theorems 2. Proposition 2. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . because of77>0. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. Theorem 2.1) .c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).
r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. Corollary 3.ga [b] 1 .3. (Laplace transform) of the ruin Corollary 3.1] evr(a)B(dy)[ b .ST(o)_ just before ruin .x)ga (x) dx where za(u) = f. the result follows after simple algebra.5 f 00 o a/r(a) .2 P(Z E [y. LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .3.ic(b)/b x(b) + a eb"E[eaT(" ). T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). rr(0) < oo) = 1_ r(a) Proof Let b = 0.°° ga(x)dx. Hence eb"du E[eaT(").ga [b] 0 TO Using Lemma 3. b .4 E[eaT (o).2.(v) = ev''(a).f. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z. r(u) < oo). £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. time T(u): u u Here is a classical result : the double m. y + dy]. u . Z = y] = EeaT.x)(a) B(dy)• Lemma 3 .T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .3.r(a) = a [B[b] B[r(a)]] .3 ga[b] = c(b) Proof + b + a .r(a) b . (u .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)) . Then by Proposition 2. Further by Theorem 111.g. E[ear (o) I T(0) < oo .
1)Er(u) . i. uoo u using e.2 Assume ri < 0. we need the following auxiliary result: Proposition 4. Proposition A1.(u ) = o(u) a. cf. For the proof. St/t 1 1/m.1) i. mu ) ( 0 m < ML '(u) 1 m > rL. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .s.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant).mu D 2 4 N(0. Then given r(u) < 00. and take basically the form of approximations and inequalities. Later results then deal with more precise and refined versions of this statement.h(u. Then as u * oo. That is.3).. T(u)/u mL as u + oo.mL > E T(u) < 00 ) 40. T(u) a. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . For the second . This proves the first assertion of (4. Theorem 4 .s. and hence a.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo.w ) v/. P = /3µB > 1.110 CHAPTER IV.e.1 Assume 77 > 0.r(u) = Er(u) • ES. for any m T(u) u . t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.6. (u) t.00 St = lim . the known results are even less explicit than for the exponential claims case. The first main result of the present section is that the value umL.1.2. (4. note that by Wald's identity u + EC(u) = ES. = (p . for any c > 0 P( Further. By Proposition 111.
For (4 .h.1).5) St . PL (•)+ 0.7 6  11 Proof of Theorem 4. cf.1 The l.6µB2) Z v m (3µB2) Z. Tu) T( u) . and (4. again Proposition A1.mL U > E. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0.t/m D (2) 111 .3). 1'r(U) .r(u)/m T(u) ti µB2) Z. proving (4. though it is not easy to attribute priority to any particular author. T(u) < oo f / 00) e7uE L [e_7 (t1). T) for T which are close to the critical value umL).2 of [86]) and (4.4.2) follows immediately from u (4.g.3. Notes and references Theorem 4. this can be rewritten as u + 1(u) . Thus. T (u) < 00 J 0(u) e7'PL U \ I T u) . According to Anscombe' s theorem (e. the same conclusion holds with t replaced by r(u).1).mL >E By Proposition 4. Theorem 7. If Z .1) is T (u)  U mL P( T (u) < I > E. 4). and as a timedependent version of the CramerLundberg approximation.s. the result comes out not only by the present direct proof but also from any of the results in the following subsections.1 is standard.N(0.2. implying T(u) .1 (by considering 0(u. apB ) . 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.^ N (o.mu (2) '• m3/2 µB 7 . which may be viewed both as a refinement of Theorem 4. .1).6.mu m .1. of (4. note first that ( Proposition 111.
ul/4. oo).a C4'(y )• ( 4. one has 9 (r(u)_rnu) Ef (^(u)) * E. we need the following auxiliary result: Proposition 4.6) whenever f.(u.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .3). we get E[ T (u) .4 (SIAM'S LEMMA) If 71 < 0. E9(Z) (4.ST( u') = u1/4 . S( u ) < ul/4] < ET(ul / 4) = O(ul/4). Proof Define u' = u .3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).t. e'°'/b (u. with w2 as in (4. oo ).v. Let h(u) = E f (^(u)). g are continuous and bounded on [0.u1/4)I(S(u') > u1 /4) h(oo) + 0.T(u') given F. and thus in (4. Using ( 4. and similarly as above we get E[f(^(u)) I Fr(u. then e(u) and r(u) are asymptotically independent in the sense that. letting Z be a N(0.mul h(oo)Eg(Z). P because of ^(u') . resp . O .4).f ( (oo)) . we can replace T(u) by r(u'). (oo.L+YWLV'U) .l:(oo) (recall that rt < 0). Then the distribution of T(u) .^(T(u')). Then for any y.w2) r.T ( u')] = E[ T ( ul /4 .6).)mu \ h(oo)Eg (r(ul) . Then h(u) 4 h(oo) = E f (6(oo)).112 CHAPTER IV.e(u') oo w .um.r.5) For the proof. Hence Ef (Vu )) 9 (T(u.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . PROBABILITY OF RUIN IN FINITE TIME Corollary 4.VU T. using that ul/4 .
4b Gerber's time.yK(ay)• (4.5 '(u .9) ( 4 .1. Notes and references Corollary 4 .4) in the last.z/)(u .(ay) = 17 7y = ay .4. just substitute T = umL + ywL in (4. oo ) as u * oo.3 in terms of Edgeworth expansions . u needs to be very large). CL Fig. 10) '5(u) .5) and solve for y = y(T). Thus . WHEN DOES RUIN OCCUR? Proof of Corollary 4. where we used Stain's lemma in the third step and (4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). see also von Bahr [55 ] and Gut [182]. T(u) < umL + ywL f. yy by 1 K. 3 is due to Segerdahl [333]. define ay. also Hoglund [204].7) to be valid is that T varies with u in such a way that y(T) has a limit in (.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . y u) < . in practice one would trust (4. The precise condition for (4. see Asmussen [12] and Malinovskii [254]. Segerdahl 's result suggests the approximation b(u. For practical purposes . PL(T(u ) < umL + ywL) 113 4 C4(y). y > k'(7) .7) to be good. y u) < e 7v" . e7v" y < ^'(7) (4 .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). however . 0. . The present proof is basically that of Siegmund [342]. ELe7E (") . Theorem 4. For refinements of Corollary 4.3 ery"z/i(u .oo. that for the fit of (4.7) To arrive at this .umL wI V"U u (4.dependent version of Lundberg's inequality For y > 0. Cf.T) Ce7"4 (T .
5 is due to Gerber [156 ]. which may be understood from Theorem 4.ay4(u)+ T(u)K(ay ).r. Then ic(ay) > 0 (see Fig .6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.5. a. if y > 1/ic'(y). we have rc(ay) < 0 and get (u) .8).2.h(u. Hoglund [203] treats the renewal case. see MartinLM [257] . For a different proof.yu ) = eayuEav [e . PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). An easy combination with the proof of Theorem 111.7 i. yu) is e 'Yyu/ . yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). and hence t. which shows that the correct rate of decay of tp(u. who used a martingale argument. f Some urther discussion is given in XI. 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. 0.b (u. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .3 yields easily the following sharpening of (4.8 below . dy) Notes and references Theorem 4 .1).114 CHAPTER IV. Numerical comparisons are in Grandell [172 ]. . we arrive at the expression in (4. In view of Theorem 4. From the proof it is seen that this amounts to that a should maximize ayic(a).t. and generalizations to more general models are given in Chapter VI. u Differentiating w.v"U.Y' (u. However.9): Proposition 4.6. 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.8). the bound a7y° turns out to be rather crude .
The traditional application of the saddlepoint method is to derive approximations. if we want EaT(u) . we have ryas = ay . As a motivation.11) ' If y > 1/ r . T(u) < yu] . i. Proposition 4. (4. (0) r1 (a) ' I.ay y 'Yay  ay .. (4. yu) = e.13) .z.c(&) = ic(ay) is < 0. then the solution &y < ay of .12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III.ay a.e.: T. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . the choice of ay. We thereby obtain that T is 'in the center' of the Padistribution of T(u).(u. (4.ayuEay f eay^ ( u)+T(u)K(ay). then the relevant choice is precisely a = ay where y = T/u.i(u.ayC() . not inequalities.'(y ). u 4 oo. This idea is precisely what characterizes the saddlepoint method.e.4. then ay > 0.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. Ea . T(u) suggests heuristically that l t/. the formula 0(u. For any a > yo. and b(u. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. and ii(u) . it is instructive to reinspect the choice of the change of measure in the proof.. yu ) eaauEaye .8 If y < 1/ic'(ry).5.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . [eT(u )K( ay).2 yields EaT(u) u u r.6 with P replaced by Pay and FL by Pay. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.. Using Lemma 111.yu) c ay .^3 ]1/ Bay [lay .
1)3 = (jB"[ay]l (Pay .9 Assume that B(x) = eay. i B[7ay .13).13) rigorously.a)2 . Then ic(a) = . where V is normal(0.(ay) _ y(ay . 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).1B[ay]1 ) y(ay .3(5/(S . The difficulties in making the proof precise is in part to show (4.1) .11) follows. The proof of (4. (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 . Writing r(u) and W2 = I3ay{.1) under Pay mation (4.(j (1 . (4.ay ) r. .a. we get heuristically that Eay Ler (u)r(ay).4).ay)K(ay) ay ayI&YI For the second term in (4.c'(a) _ /3a/(8 .12) is 0 entirely similar. PROBABILITY OF RUIN IN FINITE TIME ry I i . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.13).c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.1)3 = y3/3B"[ay]. a nr=.l'B)y /(Pay .ay + ayl /BLay] . it seems tempting to apply the normal approxiyu + ul/2wV. Example 4.116 CHAPTER IV. and the equation ic'(a) = 1/y is easily seen to have .1.B[ay] /ay &y y(ay .I ay &y a ^c'(ay) a (1 +.a) .7ruw2 Inserting these estimates in (4. V < 01 Ir 00 er(ay)"1'2"'x eyur.ay) ay +.
.1) . 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. then { __ .ay)3 0 3/2 and (4. (5.i )( v s vc ('3 + s _2 / .5.1.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . in discrete time: if p = ES.11) gives the expression '31/4 ( ./4 ^y for 1/i (u. A related result appears in BarndorffNielsen & Schmidli [59].= (s.. is undefined for a > 5).f. 0 Notes and references Theorem 4. is the drift and o. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .p. c a 00. and next to note that such an approximation in particular implies that the first passage probabilities are close.. 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 =.. yu) when y < 1/ic'('y) = p/1 . 2 = Var(Si ) the variance.g.because the c. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.8 is from Arfwedson [9].tcp) Lo {Wo ( t)}t>0 .
for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.118 CHAPTER IV. However. such that the claim size distribution B and the Poisson rate a are the same for all p (i. 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. Mathematically.tp). of which a particular case is the claim surplus process (see the proof of Theorem 5.3.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. We want an approximation of the claim surplus process itself. this is an easy consequence of (5.7c).p. (5.1 below). we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. and consider the limit p j p.2) t>o where p = pp = p .a = Snp) and the inequalities Sn )C . cf.1 As p J. Indeed . St = EN` U= .1) with S. Lemma 111..z } {W_1(t )}t>o (5.1. where p is the critical premium rate APBTheorem 5 . 0 . PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0.1)) is inconvenient. + {Wo(t ) .t} _ {W_1(t)} . and this can be obtained under the assumption that the safety loading rt is small and positive.p/c < St(p) < S((n+l)/ c + Pp/c. p. a2 =/3µB2) Proof The first step is to note that { WC (St P) . oo)).tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.. Letting c = a2/pp. n/c < t < (n + 1)/c.3) whenever c = cp f oo as p 1 p. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). we have o {i!t s: .e.
and in fact some additional arguments are needed to justify (5. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.8 or [APQ] p. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.u). u) =PIT( (u) < x) = 1 .6) from Theorem 5. Corollary 5.(u) ti IG(oo.6) This is the same as the heavy traffic approximation derived in III.1 I 7= .h. (.h. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. u).Ta2 /p2). u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. w.s.1.( ^ I + e2( \ I . u) is defective when < 0. has a continuous distribution. see Grandell [ 168]. . we omit the details . 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. we obtain formally the approximation V. For practical purposes . ('.f I \\\ J \ (5.4) Note that IG(. However. ulpl /a2) = e2"1µl / or2. 199. is IG(T.1. 1. ulpI/a2). and the r. (ua2 To2 op \ IPI > IG ( T .5. TS(u)=inf{t>0: WW(t)>u}.T) IG(Tp2/ a2). the continuity argument above does not generalize immediately. 263) that the distribution IG(•.1 .7c. is 1/ip (ua2 /IpI..2 As p j p.. 196. 119 It is wellknown (Corollary XI.r.5).t. C. [169] or [APQ] pp.s. any probability measure concentrated on the continuous functions.e. (5..5) Note that letting T * oo in ( 5. Because of the direct argument in Chapter III.s.h. since ti(u) has infinite horizon . (5. Corollary 5 .2 suggests the approximation u 0(u.
pe .g. in Asmussen [12]. and which is much more precise. However. . 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe .6) are presented.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. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. the B9. In view of the excellent fit of the CramerLundberg approximation. For claims with infinite variance. the simplicity of (5.5) and (5. (5. The proof is a straightforward combination of the proof of Theorem 5. in the next subsection we shall derive a refinement of (5. B0 * Boo.Po = 09µB6 . Assume further that 039µB6 < pe. the claim size distribution B9 and the premium rate p9 depends on 0.6) therefore does not appear to of much practical relevance for the compound Poisson model.00µB6 + 0. pt? 4 peo. 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. Then as 0 _+ 90.1. See for example Billingsley [64].Pe. that 00 4090. on the premium rule involving interest. The first application in risk theory is Iglehart [207]. The picture which emerges is that the approximations are not terribly precise. such that the Poisson rate Oe.6 of [APQ]. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. we have ^A. Theorem 5. as 0 * 00 and that the U2 are uniformly integrable w.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. In contrast. However.r. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.1 and Section VIII. We conclude this section by giving a more general triangular array version of Theorem 5. All material of this section can be found in these references. Furrer. in particular for large u. as an example of such a generalization we mention the paper [129] by Emanuel et al.. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].t. e.120 CHAPTER IV. [169]. and two further standard references in the area are Grandell [168].
In this setup. 3.c(s) = .ao (0) _ /c(s + 9 . Since Brownian motion is skip free. . Determine yo > 0 by r. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].s and p = /3µB < 1. 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. this means the following: 1.1 > 0. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.90) and the given risk process corresponds to Poo where 90 = 'yo.6. this is because in the regime of the diffusion approximation . .4.Q (B[s] . However . and we are studying b(u. PB('r(u ) < oo) < 1 for 9 < 0. 77 = 1/p . 77 is close to zero. For each 9.'(yo) = 0 and let 90 = 'Yo. claim size distribution B . Let PO refer to the risk process with parameters e9oz Qo = QB[90]. which we have seen to play an important role for example for the CramerLundberg approximation .90] B(dx). P9(r (u) < oo) = 1 for 9 > 0. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o).1) . In terms of the given risk process with Poisson intensity .T) = Peo(r(u) < T) for 90 < 0. this idea ignores (among other things) the presence of the overshoot e(u).9(s) = Ico ( s + 9) . Bo(dx) = B[eo]B(dx).6. Then r. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . 2./c(9 . The objective of the corrected diffusion approximation is to take this and other deficits into consideration. risk process with safety loading 77 > 0 correspond to 9 = 0 . 0(0) = 0. 9o T 0. and we want to consider the limit 77 10 corresponding to Oo f 0. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). whereas there we let the given 3B.90) .
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.3) this implies (take u = 1) Ego exp { .() 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.T) 1+u2 (6. .u. u) = IG(x/u2. 9otc0" (0) = 0061 = ul. 1) • Since L eatIG (dt. () where h (A. u) = euh(a . .. tu2 ) i IG (t.C. (.. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. means up to o(u1) terms): . _ ^(u) = ST . the solution of r.e. One has (6. u) denotes the distribution function of the passage time of Brownian motion {W((t)} with unit variance and drift C from level 0 to level u > 0. C . for brevity. (6.122 CHAPTER IV. Vargo S.S. and Si = QoEoU2 = Q B"'['Yo Eo U3 ].Varo S1 = f30Eo U2 = S1. 0o to. IGu+u2.7(u)/u2} eh(A.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. (01. The first step in the derivation is to note that µ = k (0) = r0 (00) . S2 = 3E0U2 Bier [Yo] 3B"[Yo] Write the initial reserve u for the given risk process as u = C/Oo ( note that C < 0) and.1) ...1) IG(x.(y) = 0. Theorem 5. C) = 2A + (2 . C. i. (. (U. write r = T(u).2) . bl IG(t81.
we get by formal Laplace transform inversion that C 2 u.s. p = 0. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r.2 ). 9o T 0 in such as way that C = Sou is fixed.s. that whereas the proof of Proposition 6. of a (defective) r.1 + 629.h.2) is indeed o(u1). CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.f.5) Once this is established .1 below is exact. in (3) and (4). however. Note.v.6.v.ry2 . 1% in (2) and (4). To arrive at (6. The solid line represents the exact value .3. and the dotted line the corrected diffusion approximation (6.2). (6. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .exp { h(A. calculated using numerical integration and Proposition 1.7.4.52/u where Z has distribution IG (•.3). which is based upon exponential claims with mean µB = 1. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .3 = 0. 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. just replace t by Tb1/u2.5) according to (6. But the Laplace transform of such a r.d. . 6 . is the c. . A numerical illustration is given in Fig. of (6. bl I IG I t +2 . distributed as Z . . The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). In ( 1) and (2).'yu /2)(1 + b2/u)} + Aug 1I J . .1 + u2 I Indeed. the r.1 As u + oo.h. however . the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6. we have p =.z . 1.
.111 W(U.01 0.1 It is seen that the numerical fit is extraordinary for p = 0.1 W IU.EB 0 p ex p ( 7 S h ^)u .T) 0. Similarly.07 0.(061 0. (Inc 0s 0.114 0.199 0.08 a.05{ 0.u2 2u3 (e .19)2 11 20 20 i0 T 1n0 Figure 6. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34]. see Asmussen [12].T) 111 0.08 0.00 0.() Lemma 6.011 L1 60 T IM 11.0 0. the fit at p = 0. The proof of Proposition 6.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.OOIi O. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.aa1 . OM 0. A51 7(SAT 3 3 h(X.7 or at values of Vi(u) like 1% is unsatisfying.4 may not be outstanding but nevertheless.W21 0.124 0. . For further numerical illustrations.^) . and all of the numerical studies the author knows of indicate that its fit at p = 0.1 proceeds in several steps. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p. PROBABILITY OF RUIN IN FINITE TIME 0..TI CHAPTER IV.T1 00.7.2 e.
+ h (A. () + C and note that 2 KO (0) = 102.61a2T (B3 .. 1 = PB(T < oo) = Eo0 exp 125 {(B .6) u U3 Lemma 6 . in Lemma 6. exp ue } al 1J 3 exP I [2).h.. + a1b2 + .4) that the r.7) 2 2 .C)C/u .r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .1) h(A. 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. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.00)(u +C)  'r (. 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 .s.T (co (8/u) ..4 Ea. () 62 Eeo exp u u2 J . the result follows.co (e) .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .(3)Eea LauT exp i 3J .6. 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .2u (B3 .(3) J t _ aa1T l + eh(A. (6.co ((/u)) } Let 8 = (2a + (2)1/2 = h()..C2 = 2). (6.
Thus by Taylor expansion around ( = 90u. [2+ (2 . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). we get h(A. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. yu/2). C) ( 1+ u2 The result follows by combining Lemma 6 .s.2. () by h(\.h.\ + () 1 2 / . and inserting this and 9o 2 = S/u on the r. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. 0 The last step is to replace h(A. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.7) and using eh(a. () .h (A.2 (^/2 + 3y9o + 390) + O(u3).4. 2 + 00 = . we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.2) for O(u) (indeed . Thus a2 y = 290 + O (u2). l Lemma 6 .x.e h(aS)h (^^ 262 exp {_h(. There are two reasons for this : in this way.. yields +90 62 0 + O(u 3) 2u2 +O(u 3).(2A + ()1/21 exp S h(A.S) d e 62 .2u [2A+ (2 3 .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).() . 5 exp { _h(A) (1 + / y u J)) exp 1.6  d h(A.() I 1 + u2 ) y .\+ (2 (3 e 2u [ (2. 2 and (6.126 CHAPTER IV. yu/2) h(A.1 (y/2 + Oo)u .
i. the same as for the unconditional Lundberg process. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed.4. () (i+a ) 2A + (2 . ()} 3 h (A.(i+ 62 exP{ h(A. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. that is. Fuh [148] considers the closely related case of discrete time Markov additive processes. the 'typical' value (say in sense of the conditional mean) was umL. ()} . The adaptation to risk theory has not been carried out. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. 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. the approach to the finite horizon case is in part different and uses local central limit theorems. In Siegmund's book [346]. HOW DOES RUIN OCCUR? exp { h (x. the analogous analysis of finite horizon ruin probabilities O(u.e.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. () I 1 + u 2 ) } S 1 . . 0 1 Proof of Proposition 6.5 in Lemma 6. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs.1: Just insert Lemma 6. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. 'yu/2) 127 ( i+ M pz^ exP { h (A. The answer is similar: the process behaved as if it changed its whole distribution to FL. with the translation to risk processes being carried out by the author [12].7. and to the Markovmodulated model of Chapter VI in Asmussen [16].T) has not been carried out and seems nontrivial. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin.
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.1 Let {F(u)}u>0 be any family of events with F(u) E F.(u).1. stating roughly that under F(u).ST(u)}t> o is just an independent copy of {St}t>o).t.2 If B is exponential.(u) is not . u * oo. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property. .. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .TT(u) _measurable. Theorem 7 .3 to . + TMOO ). (u) and satifying PL(F(u)) * 1. the Poisson rate changes from . ^(u) is exponential with rate 8 w. In the exponential case.F. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . then P(u) and FL coincide on . . Recall that .EL[e7S. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. we give a typical application of Theorem 7.128 CHAPTER IV. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7].vi(u) Ce'Yu Corollary 7.F.t. Proof P(u) (F(u)c) = F(flu)c.r.(u)_ and similarly the denominator is exactly equal to Ce7u. FL As example. .(u)_ and ^(u) are independent .FT(u) = o' (T(u ). F(u)c] ti e' ru]PL (F(u)`) > 0.(u)_ is that i. {ST(u)+t ..r. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. Note that basically the difference between FT(u) and . r(u) < oo) . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). P(u) and rate = aL w. Then also P(u)(F(u)) + 1.3L and the claim size distribution from B to BL.T. In fact. so in the in the proof.
the queueing results are of a somewhat different type because of the presence of reflection at 0. From a mathematical point of view. Notes and references The results of the present section are part of a more general study carried out by the author [11].7.3 M(u) pcu) 1 . the subject treated in this section leads into the area of large deviations theory. Proof For the first assertion.(1 . take I(Tk < x) .eALx) M(u) k=1 u The proof of the second is similar. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. HOW DOES RUIN OCCUR? Corollary 7. who also treated the heavytailed case. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). This is currently a very active area of research. however.eaLx. see further XI. .3.
This page is intentionally left blank .
7). A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA.1 (T1 = a1).d. . the one corresponding to the stationary case by 00)(u). . Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. Thus the premium rate is 1.Then no matter the distribution Al of T1i B. of the risk process form a renewal process: letting Tn = Qn .. In the socalled zerodelayed case. .1. and M is the maximum of {St}.Q. . Proposition 1.(1. with the same distribution A (say) for T2.. with Nt = # {n: Un <t} the number of arrivals before t.. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).. U2. and the one corresponding to T1 = s by 1/i8 (u). We use much of the same notation as in Chapter I.i. the claim sizes U1. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .. with common distribution B. AA t*oo lim St = lim ESt t tioo t = p . the distribution Al of T1 is A as well. see A..1).. the Tn are independent. are i. D'2.1 Define p = !µ. T3. r(u) the time to ruin. {St} is the claim surplus process given by I.
a is really the accumulated claims over a period u of length a. s + t µA PA 0 Of course. The renewal model is often referedd to as the Sparre Andersen process. the .0 > 0. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Proposition 1. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). Here are two special cases of the renewal model with a similar direct interpretation: Example 1. From this ( 1.1 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. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. Nt + EVar U. Example 1 . RENEWAL ARRIVALS lim E [St+a .1). OFF. the definition 77 = 1/p . 3) follows similarly by Blackwell 's renewal theorem.132 Furthermore for any a > 0. If the environment is Markovian with transition rate A from on to off and u from OFF to ON. t 4oo Proof Obviously. but the arrival rate in the ON state is . after E.2 (DETERMINISTIC ARRIVALS) If A is degenerate.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. and (1 . However . 2). The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. For (1 . stating that E[Nt+a . Nt ESt = E E UI Nt t = ENt•pB . CHAPTER V. Thus. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.Nt] * a/PA. such that no arrivals occur in the off state. say at a.St] = a(p . by the elementary renewal theorem (cf.1) follows .t. A.1) ENt/t + 1/µA.
oFF}. For the stationary case. the fundamental connections to the theory of queues and random walks. Proof The essence of the argument is that ruin can only occur at claim times. integrate (1. S o<t<oo n=0. The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^.s. initial vector (1 0) and phase generator 11 However.1. u For later use. (1. in general the mechanism generating a renewal arrival process appears much harder to understand. and the present author agrees to a large extent to this criticism.4) fo Indeed. we feel it reasonable to present at least some basic features of the model.d. INTRODUCTION 133 interarrival times become i.t. and then the whole process repeats itself).4) with phase space {oN. the first term represents the probability F(U1 . A is phasetype (Example 1. However.1.r. arrival times. 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 . if for nothing else then for the mathematical elegance of the subject. U1 ... More precisely.. (an arrival occurs necessarily in the ON state. as follows easily by noting that the evolution of the risk process after time s is that of a renewal risk model with initial reserve U1 ..T between a claim U and an interarrival time T.y)B(dy). Ao. .i.s < u).} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . Therefore. the relevance of the model has been questioned repeatedly. we have From this the result immediately follows.4) w.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo..1. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1..2. and for historical reasons.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0.
The initial reserve is obtained by prepayments from the policy holders.a*PB• > 1. U Figure 2.1 If. 00). Theorem 2 . are independent of {Nt} and i. i. RENEWAL ARRIVALS 2 Exponential claims. resp . That is . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. with common distribution B* (say) concentrated on (0. t. the remaining part of the prepayment (if any ) is made available to the company. Using Lundberg conjugation . 2.1. A typical sample path of {Rt } is illustrated in Fig. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion .134 CHAPTER V. the claim surplus process are given by Nt Nt Rt = u+^U. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.1) .1) +ry. St = t .Ut.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 . < 1. then 0 * (u) = 1 for all u > 0.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .1 r* (u) One situation where this model is argued to be relevant is life annuities.3* (say ) and the U.3* pB. (2. each of which receive a payment at constant rate during the lifetime . If .
* (a) = log Ee'st I. Proof Define 135 St =u . Let B(dx) = ^e7x B*(dx). If I3*pB* < 1. St=Rtu=St. B* [7] and let {St} be a compound Poisson risk process with parameters .(a) 7 Figure 2. 2. . Define T_ (u) = inf It > 0 : St = u} . EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. and the Lundberg conjugate of {St} is { St } and vice versa.UB.2 Assume now .2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. of {St} is c(a) = is*(a7). Since ic'(0) < 0. Hence T_(u) < oo a.s.f. and thus 1 = P(T. Hence y exists and is unique. B. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). B*. then by Proposition 111.Rt. 0 Now return to the renewal model.(u ) < oo) = E {e7sr_ (u).. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. cf.1. T_ (u) = inf { t > 0 : St = u 'r* (u). the safety loading of { St} is > 0.g.2. 0) and has typically the shape on Fig. > 1 .0. Then the function k* is defined on the whole of (oo. Then the c.2(a).3*.2(b). (a) is*(a) (b) . Fig. 2.
7r+ 7r EeTo b/(Sa) + +.Un } = max St = t>0 n=0.1.+Tn U1 Un. Taking m. the failure rate of this process is y..1 it is seen that ruin is equivalent to one of these values being > u. and (2 .Tr+.Tn} n=0.a) = 1 . u Hence P(M(d) > u) _ 1r+e'r"...Y] (2.• • • . RENEWAL ARRIVALS Theorem 2 .e.1.4.2 If B is exponential.g. To + max {Ul+•••+UnTI. with the probability that a particular jump time is not followed by any later maximum values being 1 .f. 3* = 6. T2 = U2. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII. According to Theorem 2.. alternatively termination occurs at a jump time (having rate 8)... we get Ee'M(d) = Ee°M* _ Y/(.Y a I.2) 7 and7r+=1Proof We can couple the renewal model { St} and the compound Poisson model {St*} with negative claims in such a way the interarrival times of { St*} are To ... and hence the failure rate .136 CHAPTER V. and 5PA > 1.Ui .1.'s and noting that V)*(u) = P(M* > u) so that Theorem 2.1) + ry = 0 which is easily seen to be the same as (2. Now the value of {St*} just before the nth claim is To +T1* +....2).Ti = U1.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.. To + M(d) in the notation of Proposition 1. and from Fig .. with rate S (say). then .•. However. A variant of the last part of the proof. 2. Then B* = A. 1) means that 8(A[ry] .u+ and lr+.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 . respectively. Hence M* max {To + Ti + • • • + Tn .1 means that M* is exponentially distributed with rate ry..
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^.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r".3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.. Bads (dx) = ..4.. The probability that the first ladder step is finite is 7r+. However. Hence the failure rate of M(') is 6(1 .2. B^d) where Aad> (dt) = ^[ a] A(dt). 0 3 Change of measure via exponential families We shall discuss two points of view. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . letting P(d) refer to the renewal risk model with these changed parameters . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. resp. This follow since.y/b. Putting this equal to y. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II.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] .5. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. we see that ry = 6(1.3.6.7r+) and hence r+ = 1. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. a ladder step is the overshoot of a claim size. hence exponential with rate b. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.2. the imbedded discrete time random walk and Markov additive processes. Furthermore.B(dx).7r+).. : S(d) > u} . which states that for a given a. 111.
It should be noted that the computation of the CramerLundberg constant C is much more complicated for the renewal case than for the compound Poisson case where C = (1 .3 For the delayed case Tl = s.p)/($B'[7] .C(°)eryu where C(O) = C0[7] .. In fact.(u) . provided the distribution F of U . (b) V)(u) .t.Ce"u where C = limu.C8e7u where Cs = Ce78B[7].2 p.. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. we get: Proposition 3. (a) '(u) < eryu.1).1 implies Cu) = e«uE ( 7d)e«^(u) . the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.2 In the zerodelayed case. let 7 > 0 be the solution of r.. ik.1 For any a such that k(d)' (a) > 0.r. 187) and thereby for ^(u) to be nonlattice w. (d) (7) _ 0. Consider now the Lundberg case. i .e. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .u the overshoot .138 CHAPTER V.T is nonlattice. 7µA . just note that F7(d) is nonlattice when F is so . VIII. 00)(u) .4. Proof Proposition 3. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0. Corollary 3. For the stationary case. E(d)e 1' (u). O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . and claim (a) follows immediately from this and e (u) > 0. For claim (b). We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . cf. in the easiest nonexponential case where B is phasetype.1) is explicit given 7.
To determine boundary 0.0 ) = Eo[ha ( Jdt. we look for a function h(s) and a k (both depending on a) such that Gh. h(s) = e(a +x( a))8 (3.1) (normalizing by h(0) = 1). delayed version of Lundberg's inequality can be obtained in a e7u.(s. 0) = tc(a)h(s). Sdt) = h(s .y) B(dy) 0 For the stationary case.9.5.3. where G is the infinitesimal generator of {Xt} = {(Jt. (s. St)} and h. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . y) = e°yh(s).Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). 0) = ah (s) .h'(s). 3b Markov additive processes We take the Markov additive point of view of II. IPA 0 Of course.a . The expressions are slightly more complicated and we omit the details.4). we invoke the behavior at the 1 = h«(0.5. 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] .dt(ah ( s) + h'(s)) so that Gha ( s. E8h0 (Jdt. Let P8f E8 refer to the case Jo = s. Here K.dt ) eadt = h ( s) .St)} can be defined by taking Jt as the residual time until the next arrival.. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. (u + s . According to Remark 11./c.1) = C(O).. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. For s > 0. B(x) = o(e7x) and dominated convergence.(°) ( Ce8B[7] Ao(ds) similar manner. 0 0 .
Further. An easy extension of the argument shows that U1.140 CHAPTER V. Remark 3 . (3. T2 are independent with distributions Ba.rc(a)] B[a] A[a .4 The probability measure Pa..8.. Aa as asserted . ] = E.5.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .. . Note that the changed distributions of A and B are in general not the same for Pa.s(Jo = s) = 1 follows trivially from Lo = 1.2) As in 11. rc(a) = 0 (B[a] . Proposition 3.c(a)] B[a] Proof Pa. . RENEWAL ARRIVALS B[a]A[a .a . An important exception is. resp.1)a in agreement u with Chapter III.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.s governing {(Jt.a .rc(a)] = 1. = J8 = T2.. i. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.s and P(d). [a + /3] A[b . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.2).rc(a)] = B = Ba[13]Aa[5]. . 5 For the compound Poisson case where A is exponential with rate . T2.e(«+k(a))t esy A(dt). Ease AU1+6T2 [ AU1+6T2 = Ea a LT. (3. ...2) means 1 = B[a]/3/(/3+a+rc (a)).13]A[b . . U. Ba where Aa (dt) . Ba(dx) = B(dx).c(a)] which shows that U1. resp. we can now for each a define a new probability measure Pa. however.e.tK( a)e. since JT. A[a .
not after time T. yu ) < e7yu. For the approach via Markov additive processes. 2. 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. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.4. and define yy = ay . and assume that T„ is the time between the arrivals of customers n . Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. see e.r.g. XII.4. Proof As in the proof of Theorem IV. [APQ] Ch. which is the same as the asserted inequality for 0. Then "^ e(ay+w(aY))8 Ys(u.5.yx(ay). T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.4. .t. yu) = F'ay. Using the Markov additive approach yields for example the following analogue of Theorem IV.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. A(ds). that is. or FCFS = first come first served) queueing discipline and renewal interarrival times. yu). .s eaysr(")+r(u ) K(ay) h (s) . that is.yu ) e7vu A[ay . it is easily seen that ic(ay ) > 0. The claim for the zerodelayed case follows by integration w.. and U„ the service time of customer n. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. Label the customers 1. Let M(u) be the number of claims leading to ruin .5 Proposition 3. the time from he arrives to the queue till he starts service. for the zerodelayed case zp8(u.y yuAa y [ay + K(ay) .. Then J(rr(u)) = TM(u)+1 and hence Ws(u.1 and n.(u.)+1 e J j e(ay+w(ay))8 e . The virtual waiting time Vt at time t is the residual amount of work at time t. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).ay+ray))TM(. the amount of . defined as the single server queue with first in first out (FIFO. u The approach via the embedded random walk is standard. .
(4.4): Proposition 4. 0 Applying Theorem 11. in which case {V} remains at zero until time on+1.• • • Tk ). Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. then Wn v M. (4..1) The following result shows that {Wn} is a Lindley process in the sense of II.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . since customer n arrives at time on. Let the T there be the random time UN.1 and Corollary 11. whereas in [On . we have Wn = Van(left limit). and we have P(W > u) = V.4: Proposition 4. The traffic intensity of the queue is p = EU/ET. 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.4.. Thus.+ Un.2) (b) as t * oo.1.4.1 Wn+1 = (Wn + U. (4. Wn converges i n distribution to a random variable W. and combining with (4."^ Vi(N) (u). but we shall present a slightly different proof via the duality result given in Theorem II. on + Tn) the residual work decreases linearly until possibly zero is hit.142 CHAPTER V. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. the waiting time a customer would have if he arrived at time t. the proposition follows.3 Assume rl > 0 or. Then: (a) as n + oo.2 Let Mnd) = maxk=o. equivalently. equivalently.1).3. It then jumps to VQ„ ..2.3) Proof Part (a) is contained in Theorem 11. .4.n1 (U1 +• • •+Uk Tl . an+1) = [on. Vt converges in distribution to a random variable V.1. we get: Corollary 4.. p < 1. The next result summarizes the fundamental duality relations between the steadystate behaviour of the queue and the ruin probabilities (part (a) was essentially derived already in 11. but interchanging the set . and we have P(V > u) = ?/iiol(u). and obviously z/'(u) = limN. (u). If W1 = 0.Tn)+. Thus Vos}1 _ = (Wn + Un .
2. T] form a stationary renewal process with interarrival distribution A. u Now return to the Poisson case . i. but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. Then K(x) = J x00K(x .4 The steadystate actual waiting time W has the same distribution as M(d). (4. T1.. K(x) = P(W < x). THE DUALITY WITH QUEUEING THEORY 143 (T1.Ao in (b). T1 . x > 0. TN) with (TN. where U*. Ti) and similarly for the U.e.y)F(dy)..2).1..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. Then the corresponding queue is M/G/1.le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . It follows that P(WN > u) =. we get W = (W + U* .T*)+ < x) = P(W + U* . and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. as WN. Hence for x > 0. conditioning upon U* . resp . by an obvious reversibility argument this does not affect the distribution ..T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). However. which implies the convergence in distribution and (4.. (4.(N)(u) has the limit tp(u) for all u... hence (since the residual lifetime at 0 and the age at T have the same distribution .5) Proof Letting n .T* are independent and distributed as U1...T* = y yields K(x) = P ((W + U* .4. In fact . Corollary 4. { 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). Letting n oo in Corollary 4.. For part (b).5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). namely W1 = 0 in (a) and Vo = 0. convergence in distribution hold for arbitrary initial conditions .oo in Proposition 4.T*)+. and we get: . cf. A. Then the arrivals of {Rt} in [0. we obtain: Corollary 4. we let T be deterministic .
implying P(W > u) = P(V > u) for all u. Cohen [88] or [APQ] Ch. see e. Some early classical papers are Smith [350] and Lindley [246].6 For the M/G/1 queue with p < 1. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . VIII). the zerodelayed and the stationary renewal processes are identical. Note that (4.5) to hold for all x E R and not just x > 0).144 CHAPTER V.g. RENEWAL ARRIVALS Corollary 4. . That is. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. Asmussen [24] and references there. the actual and the virtual waiting time have the same distribution in the steady state.g. The equation (4.5) is in fact a homogeneous WienerHopf equation. Hence '(u) = Ali(°)(u).5) looks like the convolution equation K = F * K but is not the same (one would need (4. Proof For the Poisson case. W v V.
Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . • Claims arriving when Jt = i have distribution Bi. As in Chapter I. {St} denotes the claim surplus process. and can be computed as the positive solution of WA = 0. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). here it exists whenever A is irreducible which is assumed throughout. ..(3i when Jt = i. M = supt>o St.f pi. dv. N. Thus. 145 Oj( u. i=1 0 and r(u) = inf It > 0: St > u}. {Jt} describes the environmental conditions for the risk process. The intensity matrix governing {Jt} is denoted by A = (A.T) = Pi (T(u) < T). Ire = 1.)iJEE and its stationary limiting distribution by lr. t St = E Ui . • The premium rate when Jt = i is pi.
2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.5 below. cf. Bi. T(=)). and we have f3. the operational time argument given in Example 1. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. we shall assume that pi = 1. = iii when j E E(i). and p is the overall average amount of claims per unit time. /3 = Nn when j E E(n). one expects that 3i > on and presumably also that Bn # Bi. a(i). According to Theorem A5. Proposition 1. with rates Ani and Ain.a('). Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). we can approximate A(i) with a phasetype distribution (cf.. in blockpartitioned form. leading to E having two states n. An example of how such a mechanism could be relevant in risk theory follows.14.T(n)).146 CHAPTER VI. we could distinguish between normal and icy road conditions. i and corresponding arrival intensities Qn. Unless otherwise stated. cf. which is clearly unrealistic. u . the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. Example 1. Then the state space for the environment is the disjoint union of E(n) and E(i). For example. say. respectively. and assume that weather conditions play a major role for the occurence of accidents. t(i) = T(')e are the exit rates. Thus.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential.11 below. 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. this is no restriction when studying infinite horizon ruin probabilities. assume that the sojourn time in the icy state has a more general distribution A(i). Example 1 .1 Consider car insurance. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. r^ = P (1. f3i and claim size distributions Bn. We let p Pi = /ji/AB.4) with representation (E(i).2. say. P = E 7riPi.
one could for example have H = {i1. n8}. 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.a(n).Q.3i/pi.1. the state space E for the environment is { ('q. 1 . and .. the parameters are ^ij = aid/pi.j = . i8f n1.2.J017. resp.. (9) where q = CHI.tEH is a transition matrix. it = Jel(t). Indeed.. u Example 1. say. iq (visited in that order) and letfOil >..1. Then for example wi. depending only on 77. u Example 1 .p.. . .. Example VIII. and 1/ii(u) = t/ii(u). One way to model this would be to take A(') to be Coxian (cf.. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). In the car insurance example..>.3 Consider again the alternating renewal model for car insurance in Example 1. A('^). T(9) +wggt(9)0. let T 9(T) = f pi.. w.. MODEL AND EXAMPLES 147 Example 1 . The simplest model for the arrival intensity amounts to .n. T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) .T(n)). say it is larger initially. i E E(n) }. Qi = . and similarly for the normal period. is the probability that a long icy period is followed by a short normal one. t(n) = T("i)e.. Approximating each A('?) by a phasetype distribution with representation (E('l). 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. such that a sojourn time of type rt is followed by one of type c w. St = SB=(t).. This amounts to a family (A(")) ?CH Of sojourn time distributions.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t.3. but assume now that the arrival intensity changes during the icy period. u From now on. dt. i ) : n E H.3i.4) with states i1. where W = (w.
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. Next we note a semiMarkov structure of the arrival process: Proposition 1. )3*.148 CHAPTER VI.A . dx. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. 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. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. . o = 0. More precisely. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi.(Qi)diag)• More precisely. Pi (Ti E dx. In particular. Nt Nt a .8 As t oo. we put )3* = E 7fi/3i. B* = 1 /^* Bi. qij = 0 in the notation of Chapter 11.7 The Pidistribution of T1 is phasetype with representation (ei. one can associate in a natural way a standard Poisson one by averaging over the environment. vi(dx) = . .e(A(Oi)d'sg)xe.(3iBi(dx). N > 1(Ul < x) a4 B*(x). t l=1 Note that the last statement of the proposition just means that in the limit. 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). The key property for much of the analysis presented below is the following immediate observation: Proposition 1.5. the empirical distribution of the claims is B*.
{St} to the compound Poisson model with parameters 0 *. oo) as a 4 oo. and let {St °i} refer to the one with parameters Pi. In particular. cf. i. 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. By Proposition A5. Hence Nt'> a . has distribution (7ri)3i //3*)iEE and is independent of Ti. zli( (u) . N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. and furthermore in the limit JT. Bi. Bi(x). Proposition 1. this converges to the exponential distribution with rate 0* as a * oo.aA . e. Nt a' t t iEE Also.6... Proof According to Proposition 1. Conditioning .7. the limiting distribution of the first claim size U1 is B*. In particular.. Bi. A.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). the Fidistribution of T1 in {St(a ) } is phase type with representation (E. B*. However .* (u) for all u.1. Example 11.. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. ^j 7riNi. aA.4. 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 . we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti.2. y Ni) i Nti) t a. 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$.9 Consider a Markovmodulated risk process {St} with param eters Ni. Then it is standard that ti lt '4' iri as t > oo. Then {St)} + {St*} in D[0. given {Jt}0<t<0.(/3i)aiag)..
g. shows similarly that in the limit (T2... Thus..2. T) for all u and T. with T2 being exponential with rate .s = o in state 1 and with intensity 1 . e. and (at least when a is small such that state changes of the environment are infrequent).2. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7.2 +2 2 = 3.. there are p = 2 background states of {Jt}. s 5 in state 2. which also yield O(a) (u. oo). state 1 appears as more dangerous than state 2. The fact that indeed 0(a) (U) 3 0* (u) follows. 0 Example 1. thick. the paths of the surplus process will exhibit the type of behaviour in Fig. marked by thin.l3* and U2 having distribution B*.10 Let E_{1. Continuing in this manner shows that the limiting distribution of (T. B1=3E3+2E7. T) + ?P* (u. B2=1E3+4E7. from Theorem 3. Claims of type E3 arrive with intensity 2 . MARKOVIAN ENVIRONMENT upon FT. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).. is as in {St }. 1. On Fig. U. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. we first get that 3 (3* = 2.s = 1o in state 2.. U2) are independent of . those of type E7 with intensity z s = 5 in state 1 and with intensity z . 1.150 CHAPTER VI. 9 .. resp. 132=2.=1.1 with periods with positive drift alternating with periods with negative drift.1 of [145].1. That is. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.2}. Computing the parameters of the averaged compound Poisson model. From this the convergence in distribution follows by general facts on weak convergence in D[0. 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 = . lines in the path of {St}.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. > 1...FT.). the company even suffers an average loss.
= P.11 (a ) ESt/t * p .1).1. 01 /. t * oo. MODEL AND EXAMPLES 151 Figure 1. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a).3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. iEE . (b) St/t * p .1 Thus. note first that EN Uk')/N a4' µgi. Proof In the notation of Proposition 1. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs.1.. a fraction r. 0 The definition (1. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.s.8.1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. For (b).1 a. t + oo.(3. That is. the averaged compound Poisson model is the same as in III.
[212]. and hence 1/ii(u ) = 1 for all i and u. PB. Now let r) = 0... [302]. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. 136 or A. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. [315]. dt . Now obviously the w. See Meier [258] and Ryden [314]. then M = 00 a.4. X3. The case 77 > 0 is similarly easy.1)Eiw = 0. Proof The case 77 < 0 is trivial since then the a..s..1 and the Corollary are standard.152 CHAPTER VI. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.2(a) p.g. MARKOVIAN ENVIRONMENT Corollary 1. If 77 > 0.s.Jt=i}. The mainstream of the present chapter follows [16]. Then by standard Markov process formulas (e. Eiw.ld. There seems still to be more to be done in this area. In risk theory. Since the X„ are independent .. Statistical aspects are not treated here. with X2. . having the Pidistribution of X. see the Notes to Section 7. also + . The proof of Proposition 1. and . and hence M = 00. [APQ].\ i and EiX1 Ei f 13 J. Theorem II..1 jEE = (p . X 1 =Sty. s. EiX = 0. then M < oo a.Jt=i}.a form a renewal process . and a more comprehensive treatment in Asmussen [16].12 If 77 < 0. and involves a version of the . and hence oscillates between 0o and oo so that also here M = oo. limit p .. some early studies are in Janssen & Reinhard [211].1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.0i(u) < 1 for all i and u.Eiw o'o Eiw • E ^ifjµs.. X2 =SW2 So.1 of St / t is > 0. and hence wn /n a4. Proposition 1. 38) Eiw1 = 1/ir. + Xn SWn ](1 a . see [APQ] p.. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. n n Thus {SWn l is a discrete time random walk with mean zero. and so on.. . w2=inf {t>w1:Jt_#i.
see also Example II. n=0 (2. the definition of . which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary.i. Thus. The form of G+ turns out to be explicit (or at least computable).EA. e. j. for i. oo)). Let further R denote the preT+ occupation kernel. but is substantially more involved than for the compound Poisson case . For measurevalued matrices. 00 (2. j.j E E.1) 0 (b) G+ (y.6*.6. j.3*B *(y)dy.j. However . 6. (y.a/i.2) R(dx)S((y . let G+(i.4) we obtain the following result . k.g. •)• kEE Also. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element.A) = Pt(ST+ E A.1 irG+(dy)e =. j. cf. we define the convolution operation by the same rule as for multiplication of realvalued matrices. we get the same ladder height distribution as for the averaged compound Poisson model. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). G+ is the matrix whose ijth element is E G +(i.IIG +II)e.5. Proposition 2.Jt=j)dt.2.A) =ZI(St E. by specializing results for general stationary risk processes (Theorem II . j. •). T R(i. •) II = JG+(i. .. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.(u) = Pi(M < u) = e' E G+ (u)(I .2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. only with the product of real numbers replaced by convolution of measures.dx). Proposition 2. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.Jr+ =j.x). IIG+ II denotes the matrix with ijth element IIG+(i. •) * G +(k. oo)) = f R(i.x. That is. dx)/jBj(y .j.2 (a) The distribution of M is given by 00 1 . oo) = J ao 0 G+(i. B* in Section 1.
see Figure 2.1 The following observation is immediate: Proposition 2.154 CHAPTER VI. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . 0  x Figure 2. {mx} is a non terminating Markov process on E. To make Proposition 2. mx = j when for some (necessarily unique) t we have St = x. thick.3. The u proof of (2. JJ = j.IIG+II)e. From this (2. hence uniquely specified by its intensity matrix Q (say).3 When q > 0.3) We let {St*} be defined as {St}.1) follows by summing over n and j. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. . and that the environment is j at the nth when we start from i is e . MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. we need as in Chapters II. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. resp. we need to invoke the timereversed version {Jt } of {Jt} . lines in the path of {St}. St < S* for u < t. III to bring R and G+ on a more explicit form .2 useful .6. To this end . That is. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ).1 for an illustration in the case of p = 2 environmental states of {Jt}.2) is just the same as the proof of Lemma 11. marked by thin.
Figure 2. } is a minimum value at v = t. An excursion of {St*} above level x starts at time t if St = x. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. the sequence {Q(n)} A* defined by Q(O) = .(/3i) diag. we say that the excursion has depth 0. ( Q( n)) converges monotonically to Q. and the excursion ends at time s = inf {v > t : S.2 where there are three excursions of depth 1. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx). Proof The argument relies on an interpretation in terms of excursions. Furthermore.2. In general. If there are no jumps in (t. Q( n+l) _ ^. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. Otherwise each jump at a minimum level during the excursion starts a subexcursion.2 . {S. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. The definitions are illustrated on Fig. 0 mms1   ^O \ T. corresponding to two subexcursions of depth 0.and a jump (claim arrival) occurs at time t.(/3i)diag + T S(dx) eQx.0. = x}. For example the excursion of depth 2 has one subexcursion which is of depth 1.2. s].*.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .. 2.
= j.T+>t) _ ^iF 7ri (JJ =i.Qi + )%pij) Now just note that t pij and insert (2. i. Then a jump to j (i.s. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs .4) To show Q = cp(Q).5) A (note that we use A = {x : x E Al on the r. It follows that qij = A..u< t). of the definition to make U be concentrated on (co.6) . 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. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. 7rE Proof We shall show that Fi(Jt=j. or through an arrival starting an excursion terminating with J.4).156 CHAPTER VI. Suppose mx = i. Similarly. StEA . either due to a jump of {Jt } which occurs with intensity A= j. e. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .j +/3ipij. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). mx+dx = j) occurs in two ways .(01)diag = Q. (2. Q = W(Q) follows. Writing out in matrix notation . Theorem 2 . h. we first compute qij for i $ j.. A) = f Pi(mx = j) dx eie4xej dx A u (2.St EA. By considering minimum values within the excursion. Similarly by induction . 0)).5 R(i. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii.St <S*. p1^) Define a further kernel U by f U(i. j. A) = L' U(j. A).j.
0 +1) = cp (K( n)) defined by K(o) = A . (Jo = j.7 It is instructive to see how Proposition 2..g. we shall see that nevertheless we have enough information to derive.=StSt. K( n (d) the sequence converges monotonically to K.1 can be rederived using the more detailed form of G+ in Corollary 2. the CramerLundberg approximation (Section 3).S„<0. St EA. consider stationary versions of {Jt}. We may then assume Ju=Jtu.r.6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2. St < St U.. Jt = i. where A is the diagonal matrix with 7r on the diagonal: Corollary 2..z+>t) = P.6 (a) R(dx) = eKxdx. oo))dx. e.St <Su. x < 0. St E A. From Qe = 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). and we let k be the corresponding right eigenvector normalized by Irk = 1.4]. To this end. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. 0 < u < t) = 7rjPj(Jt =i. S.(Jt=j.t.Jo=i.Qi)diag. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). Remark 2. dt. (b) for z > 0. {Jt }.StEA.. and get irPi(Jt =j.0<u<t) = P. 0<u<t).St EA. and this immediately yields (2. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.6). and to obtain a simple solution in the .2.0<u<t.6 is hardly all that explicit in general.. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. oo)) = f o' eIXS((x + z. G+((z.(.
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
Chie ryu < Vi(u ) < C+hie 7u.00 su e7( ( 3.11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy).y)G+(z. exp {e() + r(u))} . yu) f h(av)e v avuE«v. av 'i [h. oo)) and. r(u) yu o)(y)eavuEav. we have ic(ay) < 0 and get 'i(u) . 1 Similarly. (u. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . we shall need the matrices G+ and R of Section 2.5) will produce the maximal ryy for which the argument works. if y > 1lk'(ry).9) For the proof.i [eT(u)K(av ).3. dy)• o iEE jEE . as in the classical case (3. hj P .7. we let G+ * W(u) be the vector with ith component E(G+(i.j) * coj)(u) _ f u ^Pj(u . 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.i I (a) exp {aye(u) + r(u)r. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).5).. (3. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i.(ay)}.j. However. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.8 ) Then for all i E E and all u > 0. for a vector <p(u) = (cpi (u))iEE of functions . 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.i [e*(u)K(av).V)i(u.
U = U".x ) R(i.x) x) jEE 0 E Qj f jEE R(i.ery(&u+x)Bj (dy) Bj(u Bj (u .3jhj // f 00 R(i. j. 0 G+(i.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. dx). dy) = aj f Bj(dy .u IMP:°) (u) I < oo.x) jEE 00 u 0 //^^ C+E.12 Assume sup1.& (u). n > oo. MARKOVIAN ENVIRONMENT Lemma 3 . dx) f e7( vu)Bj (dy . Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. °O .13 For all i and u. dy) : 1(u) < C+ > hj u e(1tL)G+(i. dx) 100 C . dy) 00 C+ ijhj f R(i. Hence lim cp(n) exists and equals U * G. 00 Thus C+ > hj f"o e7(Yu)G +(i. if r+ (n) is the nth ladder epoch. j.j. = Eo G+ G. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).(0) ] (u) < sup Jt t.166 CHAPTER VI.x ) = Gi(u). 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. Proof Write UN = EN G+ . dx ) Bj (u . we have G *(N +1) * ^. jEE u 0 j. j. Lemma 3 .7.j. 00 f C_ hj f e(Y)G+(i. j. Then cpin)(u) sit (u) as n + oo.dy). _ To see that the ith component of U * G(u) equals ?Pi (u).
letting MT = maxo<t<T St. jEE estimating the first term in (3.13) Hence.ST).10 ) by Lemma 3 .11) C_e7u 57 O+[i. Then 0< Vi (u )  0i(u. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. T) = Pi (7. T) < C+(')' o)hi7u)e7ou8T . 13 and the second by the induction hypothesis . u The proof of the upper inequality is similar . j. from which the lower inequality follows by letting n * oo. +i .12) Proof We first note that just as in the proof of Theorem 3.10: Theorem 3 .n) ( u .13. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.Pi(MT > u) = Pi(MT < u. j. taking cps°) (u) = 0. 9 for the last equality in (3.11. and the proof of the lower one is similar. y]hj = C_ e7uhi.T) = Pi(M > u) . (3.(u) < T ) to 0i (u) which is different from Theorem 3. and let 8 = e'(70).10) C_ 1 f hje7(y. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. j.u)G+(i. j. MT < u. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .8) with y replaced by yo and hi by h=7o ). dy) jEE u U +C_ hje7( yu)G jEE"" +(i. and using Lemma 3 .tpi(u. it follows that Vi(u) < C_(yo) h=70)e7ou.3. this is obvious if n = 0.y)G+(i. we have Vii (u) . (3. and assuming it shown for n. dy) (3. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp.11). 167 u Proof of Theorem 3. let C+(yo) be as in (3.M > u) = Pi(ST<u. Indeed.MT<u.M>u) = Ei [VGJT (u . We claim by induction that then cpin) (u) > C_ hie7u for all n.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. dy) jEE o (3.
we also assume that there exist i # j such that either /3i <. this is not the case for (4. The Markov process {Jt} is stochastically monotone (4..0.1) Obviously.. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). V)" if z/i'(u) <'c"" (u). we refer to . in part from the folklore principle that any added stochastic variation increases the risk.o.o.2) (4.4) To avoid trivialities. u > 0.3). this correponds to the usual stochastic ordering of the maxima M'.3) Bl <_s.5) Note that whereas (4. we define the stochastic ordering by 0' < s. The results to be presented show that quite often this is so.2) alone just amounts to an ordering of the states. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained.3) to B = Bi does not depend on i.31:5)32 . < . The conditions which play a role in the following are: . B2 <_s. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). It was long conjectured that 0* Vi. Bp. ".. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. and finally in part from queueing theory. (4.168 CHAPTER VI.. The motivation that such a result should be true came in part from numerical studies. [177].. but that in general the picture is more diverse.o. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).33 or Bi 0 Bj.3p... (4. (4. <s. For the notion of monotone Markov processes. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .o.. Further related discussion is given in Grigelionis [176]. Occasionally we strengthen (4.
b1 < . .6.1) which with basically the same proof can be found in Asmussen & Schmidt [49].x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.9 ) below)...x) dx u o i =1 i=1 (4.2)(4..2... Lemma 4 . and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.4) say basically that if i < j . also Proposition 2. Conditioning upon the first ladder epoch.2.8) ^j Tri/iBd(x) . 0 Here (4.9) follows by considering the increasing functions 3iBi (x) and Oi (u . where 7r2+) = QiµBilri/p. Proof of Theorem 4.2)(4..x)B*(x) dx.r(u x)dx. T(0) < oo) = Bi(x) dx/tcai . we need two lemmas. = aP or b1 = .10) and (4. ..13* J0 u 0*(u . Theorem 4 . < bp and 7ri > 0 (i = 1.7) and Lemma 4.. Proposition 2.r (Sr(o) E dx Jr(o) = i. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].. Then V. Lemma 4 . it follows by a standard . Comparing (4.7) 7ri. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. we obtain (cf.1 Assume that conditions (4. dx (4. then P P P 7rjbj.9) (4. E 7r i Wi(u ...* For the proof.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB.10) Q*B*(u)+. 3 (a) P. 7(0) < oo) = pirf+).5 (cf. Proschan & Walkup [140]. the second follows from an extension of Theorem I1.6). note that (4.4.4) hold.3* f uB(x) z/^. (4.. ^i 7ri = 1.r (JT(o) = i. Section 4.3 for the second) *(u) _ /3 *B* (u) +.x) of i and using Lemma 4.1 for the first term in (4. 2 If al < .1. p). = b. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .3iBi(x)YPi(u . then j is the more risky state . (b) P..4) is automatic in some simple examples like birthdeath processes or p = 2 .. < a. Conditions (4.
let = ( 1/2 1/2 ) . that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. u To see that Proposition 4.3i.* (0).0*• i=1 But it is intuitively clear (see Theorem 3. µB. they are at present not quite complete. What is missing in relation to Theorem 4. Proof Since 0.r (u ) fails for all sufficiently small e > 0. Recall that .(0) = V..1 and Proposition 4. µB2 = 104. i=1 i=1 7'r(0) _ EFioiwi(0) .6).../3*. (4.170 CHAPTER VI. (u) may fail for some u. dominates the solution 0* to the renewal equation (4.6).11) i=1` and that A has the form eAo for some fixed intensity matrix A0.(0) < b *'(0) for e small enough. Using (4.11) is of order 104 and the r. For u = 0. of (4. Q2 = 1. Rolski & Schmidt [32].4) is essential (the present author conjectures it is). Bi as e J.3µi < 1 for all i.s.s. except possibly for a very special situation .8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .4 is the understanding of whether the stochastic monotonicity condition (4.0. Frey. of order 101.h.4 is not vacuous.. u Here is a counterexample showing that the inequality tp* (u) < V).h. Then the l. 01 = 103. 0. and from this the claim follows. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. it will hold for all sufficiently large u.4 Assume that . (u) is not in general true: Proposition 4. it is sufficient to show that 0'. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. As is seen.2. Notes and references The results are from Asmussen.. = 102.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . this ruin probability is /3iPBi. Then i/i*(u) < .
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) > 0.2 we have (Ei[e"X'. it follows by Proposition A1.5 y < ry*. Proof Define X= f &ids.1) . with strict inequality unless a = 0 or bi = 0 for all i E E. Hence if 5i 54 0 for some i E E.12) iEE Theorem 4.ld) with generic cycle w = inf{t>0: Jt_54 k.4(b) that the limit in (4. e.5. (4. Lemma 4. with strict inequality unless rci (y*) does not depend on iEE.(a) > 0 for all a 0 0. Jt = i])' EE = vA+n(6. (4. It is clear that the distribution of X.14) is nonzero so that A"(0) > 0. Then {(Jt.)a. Asmussen [20]) as discussed in 11.5.5.a.13) implies A(a) > 0 for all a.13) (4.g. and by Proposition II. is nondegenerate unless bi does not depend on i E E. 0 .4.a = E irirci(a).7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. This implies that A is strictly convex..Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state..g.1) . in particular . which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] . cf.14) A„(O) iioo varXt t t By convexity. Further (see Corollary 11. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . Xt)} is a Markov additive process (a socalled Markovian fluid model. Now we can view {Xt} as a cumulative process (see A. (4.
16) Differentiating (4. Since ic is convex with rc'(0) < 0 . this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. Let bi = rci(y*). multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. Then > risi = 0 because of (4. whereas the . h'(0) = (Ao . If rci(y* ) is not a constant function of i E E. y.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .12) and rc*(y*) = 0.) and rc (•). and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. Hence rc (y*) > 0. (4. note that y(a) + mins=1. where A.p yi and compute 8y 8a a=0 In both cases.e7r)1 (Ici(Y*))diage. improving upon more incomplete results from Asmussen. we have 7rh' = 0. (4. Thus y(e) * y* as e 10. Notes and references Theorem 4.. Rolski & Schmidt [32].6. Further a(1) = rc(y*) by definition of A(. h depend on the parameter (e or a).Qi and Bi are fixed . the basic equation is (A + (rci(y))diag)h = 0.eir)h'(0). Here we put a = 1/e..15) Normalizing h by 7rh = 0.15) once more and letting e = 0 we get .172 CHAPTER VI.5.. we get rc (y*) > 0 which in a similar manner implies that u y < y*. Hence letting e = 0 in (4. In the case of e. MARKOVIAN ENVIRONMENT Proof of Theorem 4.. h(0) = e. a = 1 in Lemma 4. Frey.5 is from Asmussen & O'Cinneide [40]. The corresponding adjustment coefficient is denoted by ry(e).
8 If (4.5. (4.17) 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). We assume that 0 < y < 7i. then 8a a=o All rci (0) Notes and references The results are from Asmussen. . Rolski & Schmidt [32].20) Letting a = 0 in (4.18).. (4. i = 2.8 when ryi < 0 for some i is open. and we have proved: Proposition 4.16) yields Proposition 4. (4.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0). The analogue of Proposition 4.i(7' *))diagh'(0).18) 0 = 27'(0)p+27r(rs. Frey. and may have some relevance in risk theory as well (though this still remains to be implemented). THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . multiplying (4.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. p. 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. Inserting (4. which has recently received much attention in the queueing literature.17) (4. We get 0 = (aAo + ( lc&Y))diag)h. 5 The Markovian arrival process We shall here briefly survey an extension of the model.19) holds. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .
We then let (see the Appendix for the Kronecker E = E(1) x E(2). A(1'k) A(2 k1).6i ) diag. Again . MARKOVIAN ENVIRONMENT f o r a transition i + j by A . In the above setting. the definition of Bi is redundant because of f3i = 0. For i = j.2). we use the convention that a1i = f3i where 3i is the Poisson rate in state i. the claim surplus is a Markov additive process (cf. Bij = B.2) A(1) = A(' 1) ® A(1.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . Jt = (Jtl).4). The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process.d. A ( 2) = A (2`1 ) ® A. A(l) = tv. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. j(2) } be two independent environmental processes and let E(k). . that Bii = Bi . A(1) = A . the definition of Bij is redundant for i i4 j. where qij is the probability that a transition i * j is accompanied by a claim with distribution. Thus .i. u Example 5 . the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. T). refer to notation) { Jt k) }. let { Jt 1) }.(13i )diag. A(l) = T. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. II. Jt2)) (2. Here are some main examples: Example 5 . with common distribution B. B.2 for details).174 CHAPTER VI.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. and thus 1i = 0. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. Indeed. Bii = Bi . This is the only way in which arrivals can occur. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. and the marked transitions are then the ones corresponding to arrivals. Note that the case that 0 < qij < 1.^) etc. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.
...iN.. as the Markovian arrival process ( MAP). WIDOWED...g.. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. INVALIDIZED.iN = a2. E 10.}. Thus.iN C17 AilO. E = { WORKING...5.kl is redundant). The versatility of the setup is even greater than for the Markovmodulated model. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. • upon a claim.. iN = all BOi2. more recently. i2i .. and that the policy then expires. However . Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . claims occur only at state transitions for the environment so that AN2. or. RETIRED.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. In this way we can model..iil. after which it starts afresh.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. all Al i2.1i2 . MARRIED. u Example 5 .... with rate ai. DIVORCED.. This means that the environmental states are of the form i1i2 • • • iN with il..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.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. iN. Example 5 .iil. Similarly. superpositions of renewal processes. 11. 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.. In fact . DEAD etc.iN are zero and all Bi are redundant. 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 . where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. say. the kth policy enters a recovering state. THE MARKOVIAN ARRIVAL PROCESS Bij. iN.. Bilo... e. 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. assume that there is a finite number N of policies.1i2.
Some main queueing references using the MAP are Ramaswami [298].p)/p. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. Obviously.3*µs • p = f /3(v) dv 0 0 (6. Sengupta [336]. Thus at time t the premium rate is p(s + t). By periodic extension. [248]. 0 < t < 1. • The premium rate at time t of the year is p(t). Neuts [271] and Asmussen & Perry [42]. Lucantoni [248]. 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. let the period be 1. we may assume that the functions /3(t). p(t) and B(t) are defined also for t t [0. Let 1 1 /3* _ f /3(t) dt. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). 1). )3 t 1 J (6. . Without loss of generality. continuity would hold in presumably all reasonable examples. for s E E = [0. • Claims arriving at time t of the year have distribution B(t). a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. For the Markovmodulated model. where i f00 xB(°) (dx) _ .176 CHAPTER VI. 1). We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. we talk of s as the 'time of the year'. Lucantoni et at. one needs to assume also (as a minimum) that they are measurable in t. p * = 0 p(t) dt. B* = J f B(t) ((*) dt. 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 . from an application point of view.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* .2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size.
Example 6 .w(t). St = SeI(t). Thus . It is easily seen that . or. u Remark 6 .2 Define T 6(T) = p(t ) dt.9).t.6. Section 4b).8. the discussion in 111.3* = 3A. In contrast. in agreement with the general principle of added variation increasing the risk (cf. In contrast. one may think of the standard compound Poisson model with parameters 3*. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . p* = A whereas B* is a mixture of exponential distributions with intensities 3 and 7 and weights 1/2 for each (1/2 = ff w(t)dt = f o (1. the average compound Poisson model is the same as in III. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u.3(t) = 3A(1 + sin 27rt). and thus the averaged standard compound Poisson models have the same risk for all A.3) Note that A enters just as a scaling factor of the time axis. respectively. it turns out that they have the same adjustment coefficient.1) and Example 1.1. the conditional distribution .w(t)) dt).1 As an example to be used for numerical illustration throughout this section. 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 . since the added variation is deterministic. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1.(3. The behaviour of the periodic model needs not to be seen as a violation of this principle.10. Many of the results given below indicate that the averaged and the periodic model share a number of main features. B*. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. let . In particular. equivalently. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . (6. of the periodic model as arising from the compound Poisson model by adding some extra variability. not random. Thus. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. We u assume in the rest of this section that p(t) . p* as an averaged version of the periodic model.
a.east B(8+t) [a] east .(1 . r(u) _ inf It > 0 : St > u} is the time to ruin . t + dt] or not.5 (see in particular Remark 11. with the underlying Markov process {Jt} being deterministic period motion on E = [0. 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. [101] .4) At a first sight this point of view may appear quite artificial.a) = exp { . with some variants in the proofs.1) . J Theorem 6 .^8 [. 1). a) is periodic on R.8). we obtain E. i.tc* (a)] dv then h (.(8) [eaSt+dt I7t] = = (1 . 0 (5)(u. Jt = (s + t) mod 1 P(8) .1) a = J8 . 6b Lundberg conjugation Motivated by the discussion in Chapter II.3(v)(B(vl [a] .g..adt +.g. Notes and references The model has been studied in risk theory by.Q(v) (B(„) [a] . see the Notes to Section 7).f.a be the c.s . Dassios & Embrechts [98] and Asmussen & Rolski [43].5. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). (6. Daykin et. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. .178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. 3 E(8)eaSt = h(s.al.e. The exposition of the present chapter is basically an extract from [44].a) Proof Conditioning upon whether a claim occurs in [t. and define h(s.1]) ..(3(s + t)dt)e«St adt + /3(s + t)dt ..1) dv .g. of the claim surplus process. a) etw*(a) h(s+t. we start by deriving formulas giving the m. As usual.f. of the averaged compound Poisson model (the last expression is independent of s by periodicity).3(s + t)dt[B(8+t)[a] . The claim surplus process {St} may be seen as a Markov additive process. let f 8+1 tc *(a) _ (B* [a] . To this end.T) = P(8)(r(u) <T).a . e. given that the ith claim occurs at time t is B(8+t).
h(s + t.t}t>o = h(s. + v)(B([a] . Proof In the Markov additive sense of (6.6 . a) . a) = exp I f t3(v)(kv)[a) . dt log E(8)east a + f3(s + t) [B(8+t) [a] .1)dv  o h(t. at + f log h(s + t. a).t} is a multiplicative functional for the Markov process { (Jt..t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.adt +.1)dv l og E(8) et where atetk•(a) h(t. a) et. With g the infinitesimal generator of {Xt} = {(Jt.2.s. a) Thus E(8)east = h(s + t.1]) .9 as follows.(e) Let = h( h(Jo. u Remark 6.t = 1 by Theorem 6. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11. a) = h(s.3(s + t)[D(8 +t)[a] .5. it then suffices to note that E(8)Le. a) Corollary 6.3. E (8)east (a +.4).1].(8)east 179 = = = = = E(8)east (1 .0(s + t)dt[B(8+t)[a] . a) h(s + t. 9) is a P ( 8)martingale with mean one. 0) exist and are finite.c* (e) {Le. 0) P(8)a. St)} .log h(s.6.4 For each 0 such that the integrals in the definition of h(t . B) eoSt t. so that obviously {Lo.5 The formula for h(s) = h(s. St)} and .1]) .9) eastt.* (a) h(s. we can write Lo Jt. According to Remark 11.
such that for any s and T < oo.(3(s)dt) +.g. correspond to a new periodic risk model with parameters ex . Proposition 6. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. 0) = h(s) + dt {ah(s) . However.2. Sdt) = h(s + dt) eadt (1 . A further important constant is the value yo (located in (0. yo is determined by 0 = k* (70) = QB*. Bet)(dx) = ^ B(t ) (dx).T.'y).3(s)h(s) + h'(s) +. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.6 The P(s). P(s) (T(u) < oo) = 1 for all u > 0. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . as above E (s) ha(Jdt. of St is as for the asserted periodic risk model. it follows by Theorem II.f. When a = y. cf.0) = Kh(s). we put for short h(s) = h(s.3. ry)) at which n* (a) attains its minimum.7 When a > yo. see [44] for 11 a formal proof. Now define 'y as the positive solution of the Lundberg equation for the averaged model.60(t) = a(t)B(t)[0].4. . J s [. Lemma 6 .y) = eayh(s). MARKOVIAN ENVIRONMENT ha(s. 0 < s < 1. the requirement is cha(i.1. B(s). That is.a . [70] .3(s)B(s) [a]h(s).180 CHAPTER VI.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. ( iii) use approximations with piecewiese constant /3(s). say. (iv) finally.5 that we can define a new Markov process {(Jt. That is. (ii) use Markovmodulated approximations (Section 6c). St)} with governing probability measures Fes). For each 0 satisfying the conditions of Corollary 6. Proposition 6.1) . Proof (i) Check that m.6 ( s ) exp { 0( s )&s) [a] + tc .3(v)( Bi"i [a] .3(s)dt • B(s)[a]h(s) = gha(s. y solves n* (y) = 0.tc] dv} (normalizing by h(0) = 1).
and no matter what is the initial season s.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. Here and in the following. s E I.1) the distribution of (l: (oo). q) = eryx/h(q)). which is not used elsewhere in the book.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.9(u))} u>0.4.1. Wu). xEJ 0 (s)b(8)(x) > 0. the mean number of claims per unit time is p« 181 = Jo 1. J C R+ such that the B(8). T(u) < (6. a) e«uE(8 ) e «^ .9) 0(')(u) = h(s.7) h(B(u). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. 0(u)) * (b(oo). The proof involves machinery from the ergodic theory of Markov chains on a general state space.9 Assume that there exist open intervals I C [0. Lemma 6 . The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. u which is > 1 by convexity. and we refer to [44]. has a unique stationary distribution. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. say s0. 9(u)) for any bounded continuous function (e.8) (6. we need the following auxiliary result . a) a > ry0 (6.2). T) = h(s.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. the Markov process {(^(u). ^(u) = ST(u) .g. e(cc)) Letting u > oo in (6.2.6. considered with governing probability measures { E(8) }E[ .10) Then for each a. 1). (6. have components with densities b(8)(x) satisfying inf sEI.9) and noting that weak convergence entails convergence of E f (^(u). f (x. Corollary 6. we get: . B(oo)). a) TI h(9(u).
A=1/4 A=1 A=4 0 Figure 6.1.1. illustrating that the effect of seasonality increases with A. (6.W.) C = E1 h(B(oo)) u + oo.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.16. Theorem 6 .ir) } Plots of h for different values of A are given in Fig. 6. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .10 Under the condition (6. where e.9).10) of Lemma 3. it does not seem within the range of our methods to compute C explicitly. 11 7/'O (u) < C+°)h(s) ery". Noting that ^(u) > 0 in ( 6.1 In contrast to h. 10 shows that certainly ry is the correct Lundberg exponent.Ch(s)ery". we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. At this stage .6 for the Markovmodulated model: Theorem 6 . Vi(8) (u) . For our basic Example 6 .11) Note that ( 6. where C(o) = 1 + info < t<i h(t) . Among other things.182 CHAPTER VI. MARKOVIAN ENVIRONMENT Theorem 6. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. 1. this provides an algorithm for computing C as a limit.
7e .42 so that 183 tp(8) (u) < 1. Theorem 6.11 as well as it supplements with a lower bound. T) and replace the Lundberg exponent ry by ryy = ay . r..(s)(u) < C+h(s)e7".4.w)e4u . in our basic example with A = 1. (6.w) . the proofs are basically the same as in Section 3 and we refer to [44] for details.12) As for the Markovian environment model. e. e7 ( yx)B(t)(dy) > Then for all $ E [0. 1 (6.6. we first note that the function fu° exu {w • 3e .g. We state the results below.17) (6.(ay) > 0 when y < 1/ic' (7). we substitute T = yu in 0(u.16 In order to apply Theorem 6. (ay).yr. Consider first the timedependent version of Lundberg's inequality. C_h(s)e7u < V.47r sin 27rs + 167r cos 47rs .w ) • 7e u{w • 3e3x + ( 1 . Just as in IV. yu) 000 (u) . .15) The next result improves upon the constant C+) in front of eryu in Theorem 6.42 • exp {J_ cos 27rs .(8) (u.3x + (1 .7x j dx _7x } _ 6w + 6(1 . RISK THEORY IN A PERIODIC ENVIRONMENT Thus. Theorem 6 .14) < C+)(y)h(s) e7yu. we obtain Co) = 1.13 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.167r I Cu. where ay is the unique solution of W(ay) =y• (6.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.12 Let 00)(y) 1 Then info < t<i h(t.13 to our basic example.0(8) (u+ yu) (6. ay) • (6. whereas ay < y.w)e4u dx 9w + 7(1 . Lundberg's inequality can be con siderably sharpened and extended. #c( ay) < 0 when y > 1/tc'('y). 1 ) and all u > 0.
C+ = 1. 14 Let C+('yo) be as in (6. completing a cycle .\ 3 C+ = sup 6 exp { A (. 1). 1/i18 1 s (u) > 0.16) with 'y replaced by yo and h(t) by h(t.181 s(u) < 1. n}.20 •exp { 2n cos 27rs . we have the following result: Theorem 6 . Of course.\ = 0 .9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.I eu. where the environment at time t is (s + t) mod 1 E [0. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.66.L sin 27rs + 1 I cos 47rs . and in fact. . and let 8 = er' (Y0).'Yo)e (6. Then 7oudT .cos 27rs . but thereby also slightly longer. with s the initial season.16. exp 2^ cos 21rs . Thus.1 sin 2irs + 16_ cos 47rs .184 CHAPTER VI. for A = 1 (where 3 e0. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.18) Notes and references The material is from Asmussen & Rolski [44].19 } 0 <8<1 8 + cos 21rs Thus e.(8)(u.0. 0 <'p(8)(u ) .1 sin 27rs + 1 cos 47rs .66.g. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. yo). .T) < C+('Yo)h( s.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.19 I eu. the nth Markovian environmental process {Jt} moves cyclically on {1.. .20). Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. Finally. Thus C_ = 2 inf ex cos 2irs . such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. 1) for the environment).4^ sin 2irs + 16^ cos 41rs . 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.013.
(6. Notes and references See Rolski [306].21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. one simple choice is Oni = 0( i . Let 0j. since the settings are equivalent from a mathematical point of view. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. M(n) = Supt>o Stn). AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. Bi. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. To this end. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). 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. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. We let {Stn)} (6. z/'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.1 ((i 1)/n) ) and Bni = B .7. . but others are also possible. Thus. 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.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. DUAL QUEUEING MODELS 185 within one unit of time on the average .jEE of the risk process. it is desirable to have formulas permitting freely to translate from one setting into the other. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t).
JT = j) = 7rjPj(VT > u. (7. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. I* )3i P(V > u.186 CHAPTER VI. JT = j) = LjPj (VT > u.3) 7ri where (V.1 Assume V0 = 0. .n(VT > u.. JT = i) = 'P.1) 7ri In particular. JT = i} coincide. Then Pi(T(u) < T. Proof Consider stationary versions of {Jt}o<t<T.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. Jo = i. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. In particular. T) = 7ri 1 P.2) and use that limF (VT > u. ii (u) = it /3 P(W > u..4) where 0* = >jEE 7rj/3j. Jt ). JT = j} and {VT > u. {Jt }o<t<T• Then we may assume that Jt = JTt. and (7.1) over j.1). Proposition 7.2). The actual waiting time process 1W1.P(V > u. (7. • The queueing discipline is FIFO.. J*) is the steadystate limit of (Vt.2) Oi(u) = 1.0i (u . just sum (7. Jo = j. let T . MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.1) follows.3). Proposition 7. J* = i). The first conclusion of that result then states that the events {T(u) < T.T(V > u I J* = i). JJ = i).oo in u (7.3. 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. (VT > u I JT = 2). and for (7. JT = Z). Now let In denote the environment when customer n arrives and I* the steadystate limit. 2 . For (7. I* = i). .=1 . (7. (7. JT = i) = P(V > u. J* = i) for all j. J* = i) = P.
7) of that paper. I* = i.4) can be found in Regterschot & de Smit [301]. I*) with the timeaverage . Proposition 7. on average 0*T customers arrive in [0. (7. P(W >u. a paper relying heavily on classical complex plane methods. B(t) have been periodically extended to negative t).g.=i) a4. if T is large.l.3) improving somewhat upon (2. [243].3). and further references (to which we add Prabhu & Zhu [296]) can be found there.1 is from Asmussen [16]. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84].8) For treatments of periodic M/G/1 queue. >u. and Rolski [306]. A more probabilistic treatment was given by Asmussen [17].6) (7. that /3(t). a general formalism allowing this type of conclusion is 'conditional PASTA'. J* = i) see W > u. see Regterschot & van Doorn [123].4).o. T]. P(. 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. we have 1: I(W. P(1')(r(u) < oo) = P(')(00) > u). With {Vt} denoting the workload process of the periodic queue. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution.7. and (7. with (7.5) follows from (7. see in particular Harrison & Lemoine [186]. on average /32TP(V > u. The relation (7.. and of these. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].7) (7. Lemoine [242].T)(T(u) <T) = P(8)(VT > u). n=1 N However. In the setting of the periodic model of Section 6. . Taking the ratio yields (7.I. N * oo.4) and (7. and one has PI'>(rr(u) < T) = P('_T)(VT > u).I *=i).
This page is intentionally left blank .
Zt As earlier. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). {Rt} moves according to the differential equation R = p(R). 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 . and the evolution of the reserve may be described by the equation Rt = u ..At + p(R8) ds. i&(u. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.T) = F(T(u) < T). the aggregate claims in [0. resp . with common distribution B and independent of {Nt}. are i. 189 . and that the claim sizes U1. . U2. finite horizon. t] are Nt At = Ui (1.2) tk(u.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.d. Thus in between jumps.6. z/i(u) = F IinffRt< 0IRo=u 1 (1.i. Thus.1) (other terms are accumulated claims or total claims).
A basic question is thus which premium rules p(r) ensure that 'O(u) < 1.4 Either i. and the probability of absolute ruin with initial reserve u E [p/S. it seems reasonable to assume monotonicity (p(r) is u . No tractable necessary and sufficient condition is known in complete generality of the model. Hence in terms of survival probabilities. that {Rt} will reach level u before the first claim arrives. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p.'(u)) > 0 so that V'(v) < 1. oo) is given by i (u + p/S).p2.2. rather than when the reserve itself becomes negative. Example 1.190 CHAPTER VII. we get p(r) = p + er. Example 1. i. when x > p/S. P(r) _ p + e(r . there is positive probability. Proposition 1.e.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. Proof Obviously '(u) < ilb(v) when u > v. say e.i(u) = 1 for all u.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Now return to the general model. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. In this situation. dividends are paid out at rate pi . we can put Rt = Rt + p/S. That is. 1 . say at interest rate b.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. where one would try to attract new customers as soon as the business has become reasonably safe. If Ro = v < u.Vi(v) u > e(1 . pi > p2 and p(r) = One reason could be competition. but when the reserve comes above v. Assume 0(u) < 1 for some u. Thus at deficit x > 0 (meaning Rt = x). However. or o(u) < 1 for all 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. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 .
and P(Rt + oo) > 0. which was proved in 11.2 once more. Then 0(u) = P(V > u). let uo be chosen such that p(r) > p = 0ILB + e for r > uo. we have z/i(u) <p(u . Then if u > no. appealing to Proposition 111. that u zPp(u . In particular.3µB for all sufficiently large r.3. In case (b). We next recall the following results. hence Rt < uo also for a whole sequence of is converging to oo.uo) and. This is basically covered by the following result (but note that the case p(r) .2) for r sufficiently large so that p(oo) = limr.1.6) . That is.e.f p(Vs) ds. Starting from Ro = uo.2(d).1 and increasing in Example 1. and hence by a geometric trials argument. INTRODUCTION 191 decreasing in Example 1. say V.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.T) = P(VT > u). Hence ik(u) < 1 for all u by Proposition III.2) we have t Vt = At .uo) < 1. { Vt} remains at 0 until the next arrival). instead of (1.5 (a) If p(r) < /.1. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. T] i n such a way that the events {r(u) <T} and {VT > u} coincide.. {Vt} decreases at rate p(v) when Vt = v (i. one can couple the risk process and the storage process on [0. . the probability that Rt < uo for some t is at least tp(0) = 1 (cf.1.4. then ?(u) = 1 for all u.b(u. [APQ] pp. then l/i(u) < 1 for all u. (1. obviously infu<uo z/'(u) > 0. 296297): Theorem 1. if and only if V)(u) < 1 for all u. let uo be chosen such that p(r) < p = /3µB for r > uo. Proof This follows by a simple comparison with the compound Poisson model. In between jumps. Proposition I1I. cf.5) and the process {Vt} has a proper limit in distribution .4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. (1. Let Op(u) refer to the compound Poisson model with the same 0. In case (a).+ p(r) exists. Theorem 1. V = p(V)).1. (1.2(d)).6 For any T < oo. However. B and (constant) premium rate p.
x] to (x.8) Proof In stationarity. the flow of mass from [0. of (1. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. yo ^ 1 + oo Q exp {.8) as g(x) = p 1 {yo13e_6x +. It is intuitively obvious and not too hard to prove that G is a mixture of two components.Sx} dx.9) Proof We may rewrite (1. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. we thus need to look more into the stationary distribution G. the l. Note that it may happen that w (x) = oo for all x > 0. Considering the cases y = 0 and 0 < y < x separately.8) as the rate of upcrossings. t + dt] if and only if Vt E [x. we arrive at the desired interpretation of the r. In view of the path structure of {V t }. where g(x) = p( ^ exp {. u Define ^x 1 w(x) Jo p(t) dt. (1.6x and that w(x) < oo for all x > 0. say when {Vt} is in state y. and is succesful if the jump size is larger than x . and the other being given by a density g(x) on (0. of (1. say.h. Oeax f x e'Yg (y) dy } = p) eaxa(x) .7) Proposition 1. for the storage process {Vt}. An attempt of an upcrossing occurs as result of an arrival. x + p(x)dt]). say.192 CHAPTER VII.y)g(y) dy. Jo AX) (1.s. oo) must be the same as the flow the other way. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. (1.s. Then the ruin probability is tp (u) = f' g(y)dy.h.8 Assume that B is exponential with rate b. B(x) = e.7 p(x)g(x) = tofB (x) + a f (x .Sx}. one having an atom at 0 of size 'yo.y.6w(x) . Corollary 1. oo). 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.8) is the rate of downcrossings (the event of an arrival in [t.6 applicable. Now obviously. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t.Qw(x) . It follows in particular that 0(u) = fg(Y)dy.
1. INTRODUCTION
where c(x) = 1o + fo elyg(y) dy so that (x) = eaxg(x) _
193
1
p(x)
nkx).
Thus log rc(x) = log rc(0) + Jo X L dt = log rc(0) + /3w(x), p(t) c(x) = rc (0)em"lxl = Yoes"lxl, g(x) = eaxK' (x) = e6x ,Yo)3w'(x)e'6"lxl which is the same as the expression in (1.9). That 'Yo has the asserted value is u a consequence of 1 = I I G I I = yo + f g• Remark 1.9 The exponential case in Corollary 1.8 is the only one in which explicit formulas are known (or almost so; see further the notes to Section 2), and thus it becomes important to develop algorithms for computing the ruin probabilities. We next outline one possible approach based upon the integral equation (1.8) (another one is based upon numerical solution of a system of differential equations which can be derived under phasetype assumptions, see further VIII.7). A Volterra integral equation has the general form x g(x) = h(x) + f K(x, y)9(y) dy, 0 (1.10)
where g(x) is an unknown function (x > 0), h(x) is known and K(x,y) is a suitable kernel. Dividing (1.8) by p(x) and letting K(x, y) _ ,QB(x  y) _ 'YoIB(x) p(x) , h(x) p(x) we see that for fixed to, the function g(x) in (1.8) satisfies (1.10). For the purpose of explicit computation of g(x) (and thereby %(u)), the general theory of Volterra equations does not seem to lead beyond the exponential case already treated in Corollary 1.8. However, one might try instead a numerical solution. We consider the simplest possible approach based upon the most basic numerical integration procedure, the trapezoidal rule hfxN() dx = 2 [f ( xo) + 2f (xi) + 2f ( x2) + ... + 2f (XN1) + f (xN)1
p
194
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
where xk = x0 + kh. Fixing h > 0, letting x0 = 0 (i.e. xk = kh) and writing 9k = 9(xk ), Kk,e = K(xk, xe), this leads to h 9N = hN + 2 {KN,09o+KN,N9N}+h{KN,191+'''+KN,N19N1},
i.e. 9 N=
hN+ ZKN ,ogo +h{KN,lgl+•••+KN,N19N1} 1  ZKNN
(
1.11
)
In the case of (1.8), the unknown yo is involved. However, (1.11) is easily seen to be linear in yo. One therefore first makes a trial solution g*(x) corresponding to yo = 1, i.e. h(x) = h*(x) = (3B(x)/p(x), and computes f o' g*(x)dx numerically (by truncation and using the gk). Then g(x) = yog*(x), and IIGII = 1 then yields f 00 g*(x)dx (1.12) 1= 1+ 'Yo from which yo and hence g(x) and z/'(u) can be computed. u
la Twostep premium functions
We now assume the premium function to be constant in two levels as in Example 1.1, p(r) _ J 1'1 r < v P2 r > v. (1.13)
We may think of the risk reserve process Rt as pieced together of two risk reserve processes R' and Rt with constant premiums p1, P2, such that Rt coincide with Rt under level v and with above level v. For an example of a sample path, Rt see Fig. 1.1.
Rt
V
Figure 1.1
1. INTRODUCTION
195
Proposition 1.10 Let V)' (u) denote the ruin probability of {Rt}, define a = inf It > 0 : Rt < v}, let pi ( u) be the probability of ruin between a and the next upcrossing of v (including ruin possibly at a), and let q(u) = 1  V" (u) Then
1  q(u) + q ( u)z,b(v) p1(v) u = 0<u<v v
0 < u < v. (1.14)
1 + pi (v )  '02 (0) pi (u) + (0, (u  v)  pi (u)) z/i(v ) v < u < oo.
Proof Let w = inf{ t > 0 1 Rt= v or Rt < 0} and let Q1 (u) = Pu(RC,, = v) be the probability of upcrossing level v before ruin given the process starts at u < v. If we for a moment consider the process under level v, Rt , only, we get Vil (u ) = 1  q, (u ) + g1(u),O1( v). Solving for ql (u), it follows that q1 (u) = q(u). With this interpretation of q(u) is follows that if u < v then the probability of ruin will be the sum of the probability of being ruined before upcrossing v, 1  q(u), and the probability of ruin given we hit v first , q(u)z'(v). Similarly, if u > v then the probability of ruin is the sum of being ruined between a and the next upcrossing of v which is pl (u), and the probability of ruin given the process hits v before ( oo, 0) again after a, (Pu(a < oo )  p1(u))''(v) = (Vi2(u  v)  p1 (u))''(v)• This yields the expression for u > v, and the one for u = v then immediately follows. u Example 1 .11 Assume that B is exponential, B(x) = e62. Then
01 (u)
_
0 e .yiu ,,2 (u) = )3 e 72u p1S P2S
1  ~ ery1u p1S 1  Q eryly P1S
where ry; = S  ,Q/p;, so that
q

Furthermore , for u > v P(a < oo ) = 02(u  v) and the conditional distribution of v  Ro given a < oo is exponential with rate S . If v  Ro < 0, ruin occurs at time a . If v  R, = x E [0, v], the probability of ruin before the next upcrossing of v is 1  q(v  x). Hence
196
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
( pi(u) _ 02 ( u  v){ aav + J (1  q(v  x))bedxdx 0 I
1 a e 7i(v x)
eP2,e 7z(uv)
1
_
P16 0 1  a e7iv P16
Se6xdx
1  e 6V Qbe72(uv)
P2 1 
a
e 71v (e(71 6)v  1)
1  p1(71  b)
Ie71v P16
p2be 7z(uv) 1 _
1  e71v a
1  e 7iv P '6
0
Also for general phasetype distributions, all quantities in Proposition 1.10 can be found explicitly, see VIII.7.
Notes and references Some early references drawing attention to the model are Dawidson [100] and Segerdahl [332]. For the absolute ruin problem, see Gerber [155] and Dassios & Embrechts [98]. Equation (1.6) was derived by Harrison & Resnick [186] by a different approach, whereas (1.5) is from Asmussen & Schock Petersen [50]; see further the notes to II.3. One would think that it should be possible to derive the representations (1.7), (1.8) of the ruin probabilities without reference to storage processes. No such direct derivation is, however, known to the author. For some explicit solutions beyond Corollary 1.8, see the notes to Section 2 Remark 1.9 is based upon Schock Petersen [288]; for complexity and accuracy aspects, see the Notes to VIII.7. Extensive discussion of the numerical solution of Volterra equations can be found in Baker [57]; see also Jagerman [209], [210].
2 The model with interest
In this section, we assume that p(x) = p + Ex. This example is of particular application relevance because of the interpretation of f as interest rate. However, it also turns out to have nice mathematical features.
2. THE MODEL WITH INTEREST
197
A basic tool is a representation of the ruin probability in terms of a discounted stochastic integral Z =  f eEtdSt 0 (2.1)
w.r.t. the claim surplus process St = At  pt = EN` U;  pt of the associated compound Poisson model without interest . Write Rt") when Ro = u. We first note that: Proposition 2.1 Rt") = eetu + Rt°) Proof The result is obvious if one thinks in economic terms and represents the reserve at time t as the initial reserve u with added interest plus the gains/deficit from the claims and incoming premiums. For a more formal mathematical proof, note that
dR(u) = p + eR(u)  dAt,
d [R(")  eetu] = p + e [R(u)  eEtu]  dAt . Since R( ;u)  eE'0u = 0 for all u, Rt")  eEtu must therefore be independent of u which yields the result. 0 Let
Zt = eetR(0) = eet (ft (p + eR(°)) ds  At I
Then dZt = e Et (_edt
f t (p + eR°) ds + (p + eR°)) dt + e dt A dA
v Z,, =  eetdSt,
= e_et (pdt  dAt) = eEtdSt. / Thus 0 where the last integral exists pathwise because {St} is of locally bounded variation. Proposition 2.2 The r.v. Z in (2.1) is welldefined and finite, with distribution H(z) = P(Z < z) given by the m.g.f.
H[a] = Ee" = exp
where k(a) _
(aeEt) dt} = exp {f °° k
k
{fa
(y) dy}
13(B[a]  1)  pa. Further Zt a ' Z
as t + oo.
198
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
Proof Let Mt =At tAUB. Then St = Mt+t(/3pBp) and {M„} is a martingale. eEtdMt} From this it follows immediately that {fo is again a martingale. The mean is 0 and (since Var(dMt) = /3PB2)dt)
Var (
Z
'
e'tdMt )
J e eft/3p(B)dt = a2B (1  e2ev). o
/' v
(2)
Hence the limit as v 3 oo exists by the convergence theorem for L2bounded martingales, and we have v
Zv =
v
eEtdSt = f et(dMt + (,3pB  p)dt)
o o

0  f0"
J
a'
0  f 0 oo
eEt
(dMt + (3p$ 
p)dt)
eEtdSt = Z.
Now if X1i X2, ... are i.i.d. with c.g.f. 0 and p < 1, we obtain the c .g.f. of E0° p'Xn at c as
00
00
00
log E fl ea°n X„
n=1
= log 11 e0(av ") _
n=1
E 0(apn). n=1
Letting p = eEh, Xn = Snh  S( n+1)h, we have q5(a) = hic( a), and obtain the c.g.f. of Z =  f0,30 e'tdSt as 00 00 00 lim E 0(apn ) = li h E rc(ae Fnh) = f tc (aet) dt;
n=1 1 n=1 0
the last expression for H[a] follows by the substitution y = aeEt Theorem 2.3 z/'(u) = H(u) E [H(RT(u)) I r(u) < oo] .
u
Proof Write r = r(u) for brevity. On {r < oo }, we have

u + Z =
(u + Zr ) + ( Z  Zr) = e
ET {e
(u + Zr)  f '* eE(tT )dSt] T
e
ET [
R( u)
+ Z`],
2. THE MODEL WITH INTEREST
199
where Z* =  K* eE(tT)dSt is independent of F, and distributed as Z. The last equality followed from Rt") = eEt(Zt + u), cf. Proposition 2.1, which also yields r < oo on {Z < u}. Hence H(u) = P(u + Z < 0) = P(RT + Z* < 0; r < oo) zb(u)E [P(RT + Z* < 0 I)7T, r < oo)] _ O(u)E [H(RT(")) I r(u) < oo] .
Corollary 2.4 Assume that B is exponential, B(x) = e6', and that p(x) _ p + Ex with p > 0. Then
. o€Q/E Ir, (8(p + cu);
V) (u)
aA/Epal Ee 6n1 E +^3E1 / E
1\ E E
1r
Cbp;
E El al
where 1'(x; i) = f 2°° tnletdt is the incomplete Gamma function. Proof 1 We use Corollary 1.8 and get
w(x) fo P + Etdt = g(x) = p +0x
e log(p + Ex)  e loge,
exp {  log(p + Ex)   log p  6x }
pal(p + ex)plE1e6^ J ryo)3 70 = 1 + J p) exp {Ow(x)  Sx} dx x r^ = 1+ ' /E (p + Ex)01'leax dx + 0
f J
= 1+
a
Epo/ E
f yI/ E 1e 6(Y P)/E dy
P (
1+ OEA/E 1e6 P /Er
60/e po/ e
,;,3 )
E E
lp(u) = to foo a exp {w(x)  bx} AX)
acO/E" 1 ePE l
Yo
50 1epolE
(
+ cu); 0)
5(p
E E
Proof 2 We use Theorem 2. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .pa. and the c.2) follows by elementary algebra. . i.pa.13 /E) r (. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.5 The analysis leading to Theorem 2. As an example. /^ u Example 2 .f.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.e.3a/ (5 . H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.2y +µ ) dy . RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .V < x)]0 + f P(V > p/E ) + eby fv (p/E ./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . 13/E).x) dx e. then {Rt} is the diffusion with drift function p+Ex and constant variance a2. 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 . RESERVEDEPENDENT PREMIUMS u from which (2.2) follows by elementary algebra.g.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.3/E) By the memoryless property of the exponential distribution.a) . r (j3/E) In particular.200 CHAPTER VII.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. From ic(a) = . where V is Gamma(b. assume that {Wt} is Brownian motion with drift µ and variance v2.3.V. of Z is IogH[a] = f ytc(y)dy = e fa (0.
it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. Corollary 2. not even Erlang(3) or H3. Z is normal (p/E. or to nonlinear premium rules p(•).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).3) was derived by Emanuel et at.d.3 is from Harrison [185]. it is also used as basis for a diffusion approximation by these authors. Some of these references also go into a stochastic interest rate. The formula (2.3.p*.v. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. and since RT = 0 by the continuity of Brownian motion. Paulsen & Gjessing [286] and Sundt & Teugels [356]. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. for a martingale proof. [282]. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). write y* for the solution of the Lundberg equation f3(B[ry *] .. Emanuel et at. [357]. Further studies of the model with interest can be found in Boogaert & Crijns [71].g. write Vi* (u) for the ruin probability etc. however. se e.i. see e.1) . [129]. and recall Lundberg 's inequality . A r. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) .8. Gerber [155]. as in the proof of Proposition 2. Gerber [157] p.. Goldie & Griibel [167].2 is a special case of a perpetuity. [283]. that the analysis does not seem to carry over to general phasetype distributions.Y*p* W*(u) < ery*u = 0. [129] and Harrison [185].g. Delbaen & Haezendonck [104]. Paulsen [281]. 3 The local adjustment coefficient. Q2/2E). It must be noted. of the form Ei° p"X" with the X„ i.e.
e.4) we assume existence of y(x) for all x. 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. we will use the local adjustment coefficient 'y(x). let p* be a in (3. The steepness assumption and p(x) + oo ensure 'y(x) * So. choose uo such that p( x) > p* when x > u0E.log '(u)/u < ry*(1 . choose c(. log ?i(u) < < 00 JO . (x.. i. Let y* < So. i. 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) . If 60 s f 6o.C*ef*".w (u) J dt > c(3)eeu v 1 p(u+ t) .1) .i)eex. c(. as will hold under the steepness assumption of Theorem 3.*(u) .1 ). a) = f3(B[a] .2) such that p(x) < c(.'y ( x)) = 0 where r. When u > uo. then log u (u) In the proof as well as in the remaining part of the section . oo for all E > 0. it holds that f3[s] T oo. obviously O(u) can be bounded with the probability that the Cramer Lundberg compound Poisson model with premium rate p* downcrosses level uE starting from u .1 Assume that for some 0 < 5o < oo. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. e(1o+e)2 (x ) u > 00. as solution of the equation n(x. and that p(x) * oo.1.. x * oo.e.202 CHAPTER VII.E). 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).>o 7(x) > 0.3) When trying to extend these results to the model of this chapter where p(x) depends on x. x>0 (3. (3. 1) and for a given E > 0. (3. Then lim sup u>oo u and e E''p(r) + 0. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.ap(x).5) which implies inf. B(x) > C(2)e(ao+f)x for all x. Proof of Theorem 3.1. For the last asssertion . a first step is the following: Theorem 3 . and (for simplicity) that inf p(x) > (3µs .
Then .e.13 is a technical condition on the claim size distribution B.6) The second main result to be derived states that the bound in Theorem 3. the asymptotics u * oo and c . Theorem 3 . or (b) Condition 3.g. Condition 3. Then limelog l/ie (u) = I(u). However. I. For e > 0. (u) = O(u/e).(u) > so. {Rte)} defined as in (1.v. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. 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. . u Obviously.4)eE" Given such an arrival. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. If p(x) = pis constant . (3. 3) = (1 .2). the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. 2. Theorem 3.0 are the same.1 only presents a first step. (3.13 below holds.2 is also an approximation under appropriate conditions. The form of the result is superficially similar to the CramerLundberg approximation.3.3 Assume that either (a) p(r) is a non decreasing function of r. ruin will occur if the claim is at least u + v. one can then assume that e = 1 is small enough for Theorem 3. UJU > x cannot have a much heavier tail than the claim U itself.ea°/(ecf1)). the result is not very informative if bo = oo. and in particular. noting that in many cases the constant C is close to 1.7) CIO Remarks: 1. then Rte) = CRtie for all t so that V).3 to be reasonably precise and use e` (u) as approximation to 0 (u).' (u) < eI("). which essentially says that an overshoot r. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. by cU2.. 3. Bucklew [81]). The rest of this section deals with tail estimates involving the local adjustment coefficient.
say. 3. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. we show how to rewrite the explicit solution for ti(u) in Corollary 1.bx} dx = 1 + J0 dodx(x) exp {. we get = 1+ J" AX) exp {(3w (x) .(3/p(x). Then y(x) = b . it is formally needed only for Theorem 3. the logaritmic form of (3. However.6). and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.bx} dx fo 00 1 + [exp {/(3w(x) .bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .3.2.1 + b f e.8 in terms of I(u) when the claims are exponential: Example 3 . 5.(iw(x) .4 Consider again the exponential case B(x) = eax as in Corollary 1. 3a Examples Before giving the proofs of Theorems 3. u .7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u). J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .bx} dx 1+0.bx} dx .(x) dx.204 CHAPTER VII. RESERVEDEPENDENT PREMIUMS 4. First. rather than eI(u)).bx}]o + b /' oo exp low (x) .3.bu}. As typical in large deviations theory.exp {/33w(u) .8. we consider some simple examples.
In particular.2.10 or Karlin & Taylor [222] pp.I ( v )dy fo +u) dxdy . 70 > 0 such that 7(x) < 7o for y < yo.e. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.fory(x+u)dxdy ( 3. and (3.5 Assume that {Rt} is a diffusion on [0. Be = e log U000 e.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. IE(u) = I(u)/e. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3.10) where AE = e log 000 e.1/8 .BE..I ( u) fool.f y(x)dxd y If 7(x) is increasing . BE * 0. Choosing yo.5) is infx>o 7(x) > 0 which implies that f °O .7) follows. 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. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . in the definition of AE converges to 0. Similarly.2(X) = ev2(x) so that 7e(x) = 7(x)/e. (u) = I(u) + AE .3.9 ) 11000 eI(v)dy f000 e. we get r 00 e.1. For Theorem 3.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .9) yields e log . u . (3.3 in the particularly simple case of diffusions: Example 3. > lime log e = 0 and AE * 0. and (3. (3.fa 7(x+u)dx/Edy o The analogue of (3.2. oo) with drift µ(x) and variance a2 (x) > 0 at x. 0.ev 0 O /E) J0 70 70 Yo This implies lim inf A.0. ry(x /b I u o e f0 °° e  e.fo 7(x)dx/Edy f . (X) = µ(x).3. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0. It is well known that (see Theorem XI.. note first that the appropriate slow Markov walk assumption amounts to u. 3.
however .+1 (u) = o( 1). . that the interchange of the slow Markov walk oo is not justified and in fact. lim sup Af < lim sup c log(1 .5. G. Then the solution of the Lundberg equation is y* = b . > .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. Further. so our approach is to determine standard functions Gl (u). I./3 1 AX dx..7o C 15 I I. 0.5.4. I(u ) = G1(u) + .7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B. . As in Example 3. this leads to (3.6 Assume that B is exponential with rate S.(u) oo. Nevertheless .6/p* so that u 1 I (U) = bu . 7(x) is typically not explicit. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo.. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. G. _ . the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. we have 5 > 7o and get lim inf AE > lime log e .5) and 7* = 5 . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.5 for risk processes with exponential claims is as follows: Example 3 . Of course. .206 CHAPTER VII.0.0/e. ..5. Thus 7e(x) _7(x)/e and (3. Ignoring 1/5 in the formula there.0) = 0. 0 Now (3..6) exactly as in Example 3. + Gq(u) + o(G9(u))• Gi (u) It should be noted . the results are suggestive in their form and much more explicit than anything else in the literature.1 3. the slow Markov walk assumption means 5E = b/c. E+o e*O By (3.e. G. .Q/p*.
B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .1/a).x)n1. 2. More precisely. ry* loge*+ g7loglogp*.3.4) or gamma distributions.3cse7*I7(77) . (3. y = 2 if B is uniform on (0. It follows from (3. This covers mixtures or convolutions of exponentials or. phasetype distributions (Example 1.1) leads to (S7T N Ocp a. Here B[s] is defined for all s and B[s] . and hence (3.g. x T 1. I(u) Pt.7 Assume that B(x) .c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. fu I(u) Su .c2 Su a dx ) Su a<1 Su .. For example.12) with y > 1. u(logu + r7loglogu). 1. in the phasetype case .1) leads to . . Hence (3.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . 77 = 1 if B is degenerate at 1.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.s)C' f "o o as s T S. more generally.clxale5x 207 (3. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. e. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . if the phase generator is irreducible ( Proposition VIII.:. say 1 is the upper limit and B(x) .cs(1 .11) with a > 0. c4 = c2b 1/'/(1 . the typical case is a = 1 which holds . .11) that b[s] * co as s f S and hence 7* T S as p* + oo. u Example 3 .1/k).ry*°p*.8).Y .8 Assume next that B has bounded support.C2p* C2 = (3clr( a))11'.
r^.f.1) . RESERVEDEPENDENT PREMIUMS Example 3 .c8 log . Hence for u<V.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .12).9 As a case intermediate between (3.14) for the m .f.4).4) is the formula h logEues ( Rhu) . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .ru(TI)) . ec78)2/2c7 dx C7 .11) and (3. .g. assume that B(x) CO x2/2c7. I (u) c8u log u 0 where c8 = 2/c7. h].2 We first remark that the definition (3.u .3 (B[s] . 1 = E. (3. one could also have considered the increment ru (T1) .Ote7o( u)(u.u is a nondecreasing function of u.10 Assume that p(x) is a nondecreasing function of x. (3.13) We get b[s] . this is only possible if 7o(v) 2 7o(u)• . (3.r„(Ti). By convexity of the m . g. of the increment in a small time interval [0.B[7o (u)] .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.. 1 0 3e.(t))dt. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.•.css 2%rc7eC782/2. of U1 + v .log p*.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).15) Proposition 3.e7o ( u)(ul+u r. x f oo . 3b Proof of Theorem 3. 7 * .(T1)) > Ee7o(u)(ul+vr»(Ti)).208 CHAPTER VII.sp(u). h 10. The assumption implies that ru(t) .
Then (u) < efo Yo(x)dx.7o (u)p(u)• Since (3. (3. Also. Hence „/.Fu(u ) + J ^(n)(u . The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. fa 7o(y)dy < u7o(u) < xyo (u) for x > u.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).u[70(u)] fo eyo(x)dx .17) shown for n and let Fu(x) = P(U1 + u .(n+l) (u) 1 .1) .es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.u > tp(u).2 in terms of 7o. u We prove Theorem 3.10(b): Theorem 3.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . Assume (3.4) considered as function of 7 is convex and 0 for y = 0. THE LOCAL ADJUSTMENT COEFFICIENT For (b). it is easily seen that fu x7o(y)dy < xyo (u). this is only possible if yo(u) > 7(u). We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. note that the assumption implies that ru(t) .11 Assume that p(x) is a nondecreasing function of x. the case of 7 then follows immediately by Proposition 3.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. Separating after whether ruin occurs at the first claim or not.ru(T1) < x). Hence 1 = EeYo(u)(U1+uru(T1)) < E. we obtain „I.17) from which the theorem follows by letting n + oo.3.
11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. (u) < I(u).e (u) = v'.E ( u/n) ^•e. However. W O .I. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. let Op*.E (u/n).nbe C*. P k.n <Z auk}l.n) > k =1 II v ^k n.2.n) pn niE (u /n) n n_1 n. (un2. the probability that ruin occurs in the CramerLundberg model with p* = pn. and here it is easily seen that yo(u) .n. in accordance with the notation i/iE (u).3/e and U. {RtE)} (starting from u = un.E (u/n) Now as e .n.n.. in .E (u/ n) Y'E (un . resp.2.10(a) for some of the inequalities. for either of Theorems 3. pk n = uk_l.n so that n.n = ku. For these reasons. To this end..10(b ) that the bound provided by Theorem 3..3). Also.: y(u).n = sup p(x). 0.n inf n uk1.E (u) denote the ruin probability for the classical model with 0 replaced by .n AX). by €U=. The probability of this is at least n n. Let Ck.12 lim sup4^o f log O. 0 It follows from Proposition 3. 3. (u).3).11 is sharper than the one given by Theorem 3. the value of {R(E)} at the time of downcrossing is < unl.n (starting from u/n) without that 2u/n is upcrossed before ruin.n u k}1. we used also Proposition 3. Proof For ruin to occur. and. op*.15). Y*u /E. define uk. ryk.2).x/n.3 is required..210 CHAPTER VII. 3c Proof of Theorem 3. Lemma 3. 0. y* evaluated for p* = Pk. C*e where the first equality follows by an easy scaling argument and the approximation by (3. (3. Further.. yo(u) appears more difficult to evaluate than y(u)..3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.n. given downcrossing occurs. we have chosen to work with y(u) as the fundamental local adjustment coefficient.n) must first downcross unl.
E (u/n) Op•. we need the following condition: Condition 3. .n = sup ?'(x). 3 now follows easily in case (a).nu/en(1 + where o(1) refers to the limit e . 11 Theorem 3. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). so that Theorem 3. uk_1.nu /fn( 1 Ck  e.19) .n cE (2u/n) Ck ne7k. since ry' is an increasing function of p'.E (u/n) OP +^p•.F (2u/n). i.i. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.n + 0(1). 211 Clearly. In case (b). in obvious notation one has tC (x) = y(x)/e.18) (ii) the family of claim overshoot distributions is stochastically dominated by V. for all x .13 There exists a r.n .log Ck. THE LOCAL ADJUSTMENT COEFFICIENT particular. *p•. Indeed .12 completes the proof. v. /' (u/n) 'T nk. 0 with n and u fixed. k=1 k=1 n u _ nE7 k..nk=1 limsupelogv). (3. It follows that n log V'C (u) k =1 log Ypk.n. 40 Combining with the upper bound of Lemma 3.nu /En) o(1)). also ryk.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).a( u)z. B(x) (3. ne7k.3...7k.e.! (u/n) n n m 7k..n <X<Uk. 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.E (urn) < \ *I.
R<) (u v). P (T(E) (u.v.nu/En0(1) . (R. u /n) < oo] l = = < E [OE (u/n . Then the standard Lundberg inequality yields El < E?. . (3. infx>2u /n P(x) .V) = e71 nu/Eno(l) (using (3. V < u/En] + P(V > u/En) (u/En . let v < u and define T(E) (u. Ei + E2 < e71. Then Y'E (u) ^(E) (u.nu /EnE [e71.n < ery1.EV) = El + E2. (u/n . we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.EV) = EiI 1 .n V. T(E) (u.. (u/n . v ) = inf { t > 0 : R(c ) < v R) = u } ..E(E) (u.E (0) cf. RESERVEDEPENDENT PREMIUMS To complete the proof. T() (u.5) and the standard formula for b(0).E (u/n . u/n) < oo) EV).1 n. N with EN < 1 = infx>2u/nA(x) = 0(1). where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. u/n)) .212 CHAPTER VII. u /en 0(i) _n so that E2 < e2ryl nu/En0(1). The probability of ruin in between two downcrossings is bounded by Epp . Write EO.^(E) (u.eV) • P (T(E) (u.QEU 1 .. v ) = v .of:>2 in n(x). u/n) < oo) .E (2u/n .( . u/n) < oo] E [OE (u/n .^'' = E [ ..2y 1 ' . u/n)) I T(E) (u.EV) = e.18) for the last equality). ) (u u /n)) . For E2. u/n) < oo] .
7) then comes out (at least heuristically) by analytical manipulations with the action integral. s) as in (3. Similarly.) = exp . (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.7(x)) (3. the results are from Asmussen & Nielsen [39].r.1.3EU) (3.J y(Rs)dR. Djehiche [122] gives an approximation for tp(u.J r(Rs)p(R. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. l o JJJ o .13. one can in fact arrive at the optimal path by showing that the approximation for 0(u. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . the approximation (3. s).21) (the initial condition is r(0) = u in both cases). u Notes and references With the exception of Theorem 3.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. T) is maximized over T by taking T as the time for (3.t. 0 and b T 00 are interchangeable in the setting of [89].u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.)Ui } . 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. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.4) and the prime meaning differentiation w.20) (with ic(x. . Typically. 0 ) (= p(x) .=1 J An approximation similar to (3.21) to pass from u to 0. Comparing these references with the present work shows that in the slow Markov walk setup. [89].g. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. the risk process itself is close to the solution of the differential equation r(x) _ r (x. Bucklew [81]). whereas the most probable path leading to ruin is the solution of r(x) _ k (x.T) = P „(info<t <T Rt < 0) via related large deviations techniques. it might be possible to show that the limits e .)ds + Y(R2.3. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .
214 CHAPTER VII.3. For different types of applications of large deviations to ruin probabilities . see XI.g. We should like. . the exponential distribution ). e. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding .. to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. however.
if a problem can be solved explicitly when the relevant distributions are exponentials. A distribution B on (0. Typically. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. We often write p for the number of elements of E. 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<. Note that since (1. More precisely. refers to the case Jo = i. on Eo = E U {A} where A is some extra state which is absorbing. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory.1) is 'Here as usual . F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 .Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. if v = (vi)iEE is a probability distribution. P. and not in other cases. that is.
e. i. Here are some important special cases: Example 1 .0 = t11. T) (or sometimes just (a. k}. Then a = a1 = 1. 0 2this means that tii < 0.3. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. We now say that B is of phasetype with representation (E. a. that is. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. tij > 0 for i 54 j and EjEE tij < 0 .T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. j. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. E = {i. C is the lifetime sup It > 0 : Jt E E} of {Jt}. and we have t = Te.e. T is a subintensity matrix2. an exponential distribution with rate parameter . In particular. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions.1 The phase diagram of a phasetype distribution with 3 phases. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process.1 Suppose that p = 1 and write . B(t) = Fa(^ < t ).216 CHAPTER VIII. Equivalently. (1. t1 = /3.2) The interpretation of the column vector t is as the exit rate vector. The initial vector a is written as a row vector. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one. i.
. p}. 0 •.1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. so that the density is P E ai6ie6.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.. 0 S 6 Example 1. 00)) S s o .......1. 0 SP 0 and the phase diagram is (p = 2) . the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. . . a = (1 0 0 . ..2 corresponding to E = {1.x i=1 Thus E _ Si 0 T 0 S2 0 0 .. . 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 ..... 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •. PHASETYPE DISTRIBUTIONS 217 Example 1. 6. 0 0 0 0 S 6 .
Then: (a) the c. ds^ = ds' = ttlaj + tikpkj.1 tP1 1 Figure 1.aeTxe.3 . Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.1)"n! aT"e. 5 Let B be phasetype with representation (E.4 For example. Then for i . the Erlang distribution is a special case of a Coxian distribution.t2 yt bP. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . MATRIXANALYTIC METHODS Figure 1.g. p:. [APQ ] p. The basic analytical properties of phase type distributions are given by the following result . T). Theorem 1 . a. (b) the density is b(x ) = B'(x) = aeTxt. (c) the m. dp. . 36) yields s d.f.g. i. the backwards equation for {Jt} (e.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. see A. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} .3 0 Example 1 .218 CHAPTER VIII. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. the restriction of P8 to E.e.d. j E E.f is B (x) = 1 .
jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). Part (d) follows by differentiating the m.g.g..T) 1t. or w.5) as hi(tii + s) = ti  t ij hj.n lt . 1.5) Indeed .f.6) . Since 1 . and (b) then follows from 1: aipF. in which case the time to absorption is 0 with m . Then h tit ti + ti3 h j .. ti/ .p.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . for n = 1 we may put ki = Ei( and get as in (1. . d" dsn a (. Alternatively.tii and have an additional time to absorption either go to state j which has m . hj . the rule (A. and since b[s] = ah. i.1 ) n +l n ! a (s I + T ) .e. d8 P8 = TP8. the solution is P8 = eT8. i. tij / .g. PHASETYPE DISTRIBUTIONS 219 That is.tii tii . of the initial sojourn in state i.1.tii is the rate of the exponential holding time of state i and hence (tii)/(tii . we arrive once more at the stated expression for B[s]. h = (T + sI)1t.tii we go to A. = aPxe. For (c).g.5) ki = 1 + tii L jj:Ai tii (1.f.n1t = (1)nn!aTn1Te (1)nn! aTne. After that.p. Rewriting ( 1. (1) n+1n!aT .s j# tii i (1. we i w.s I .f.T) 't = (. and since obviously P° = I. this means in vector notation that (T + sI)h = t.s) is the m . (Jx E E) = this proves (a).B(x) = 1'a (( > x) = P. j#i jEE tijhj + his = ti. define hi = Eie8S. B(n)[0] = _ Alternatively.f.
another the case p = 2 where explicit diagonalization formulas are always available. see the Appendix.s."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. Consider for example 3 9 a= (2 2). 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. 0 Example 1. Example A3. One obvious instance is the hyperexponential distribution.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.h. T= 2 111 so that 2 2 Then (cf.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. are idempotent. there are some examples where it is appealing to write T on diagonal form. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .220 CHAPTER VIII.
5 and serves at this stage to introduce Kronecker notation and calculus (see A. (1.e a mixture of a phasetype distribution with representation (a/llall. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).T) with weight hall and an atom at zero with weight 1 . • The phasetype distribution B is zeromodified.e 11BIJ = 1laDD < 1.4b for definitions and basic rules): Proposition 1. a random variable U having a defective phasetype distribution with representation (a.f.hall. .1. 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 4.4. < 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). or one just lets U be undefined on this additional set. where the initial vector a is substochastic.29) and Proposition A4. and in fact one also most often there allows a to have a component ao at A.11aDD.T).7 If B is phasetype with representation (v.7) Proof According to (A. There are two ways to interpret this: • The phasetype distribution B is defective.g. i. This is the traditional choice in the literature. T) is then defined to be oo on a set of probability 1. 0 Sometimes it is relevant also to consider phasetype distributions. then the matrix m. hail = E=EE a. i.
not only in the tail but in the whole distribution.222 CHAPTER VIII. (1. 2. Neuts. See in particular the notes to Section 6. 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. No satisfying .q be the eigenvalue of largest real part of T. one has k = 0. All material of the present section is standard.. the result follows (with C = (ah)(ve)). but the relevant T is not irreducible. Rolski.g. Notes and references The idea behind using phasetype distributions goes back to Erlang.8 Let B be phasetype with representation (a. Example A5. Using B(x) = aeTxe . O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m.. where C. cf. distributions with a rational m. Then the tail B(x) is asymptotically exponential.1 of the Appendix. i is real and positive.8) Proof By PerronFrobenius theory (A. .f. (or Laplace transform) are often used where one would now work instead with phasetype distributions. let . Schmidt & Teugels [307] and Wolff [384]. 0 Of course.f. but in many practical cases. here k = p1).4c). In Proposition A5. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. see his book [269] (a historical important intermediate step is Jensen [214]). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form.Ce7'. v.hve7x.8). B(x) .g. Here is a sufficient condition: Proposition 1. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . 1. The Erlang distribution gives an example where k > 0 (in fact. Lipsky [247]. the text is essentially identical to Section 2 of Asmussen [26]. Schmidli. x * oo. h can be chosen with strictly positive component. assume that T is irreducible . cf. the conditions of Proposition 1. we give a criterion for asymptotical exponentiality of a phasetype distribution B. 77 > 0 and k = 0.F.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. In older literature. the Erlang case). and we have eTx . T). Other expositions of the basic theory of phasetype distributions can be found in [APQ]. but todays interest in the topic was largely initiated by M. let v.
. JtJt1) Then { 0<t<U1 . Then the renewal density exists and is given by u(x) = ae(T+ta)xt. + U0 is 0 . what is the smallest possible dimension of the phase space E? 2 Renewal theory A summary of the renewal theory in general is given in A. Jt={Jt?ul}.. we refer to U as the renewal measure. A related important unsolved problem deals with minimal representations: given a phasetype distribution .. known.. with common distribution B and define4 U(A) = E# {n = 0.. the jumps of the j(k) and the it } k) to the next J( k+l) A jump jumps corresponding to a transition from one Jt 4Here the empty sum U1 +.. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. the renewals form a Poisson process and we have u(x) = 0.r. oo) w.d.. and U(A) is then the expected number of replacements (renewals) in A..i.t.. +UnEA} 00 = EEI(U1 +. n=O We may think of the U. U1<t < U1+U2.2. be i.. (2...1 of the Appendix. if U is absolutely continuous on (0.1 Consider a renewal process with interarrivals which are phasetype with representation (cr.: U1 + .1. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m.+UnEA). as the lifetimes of items (say electrical bulbs) which are replaced upon failure.T). or the density is available ) is. U2.f. .. we denote the density by u(x) and refer to u as the renewal density. but is in part repeated below. however.. Lebesgue measure. Let U1. For this reason. .1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. is Markov and has two types of jumps . the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.g. but nevertheless. If B is exponential with rate 0.
which is ti in state i. (b) £(t) has a limiting distribution as t * oo. and the distribution of Jx is ae ( T+t«)x. the lifetime of the renewal process. as the time of the last renewal. which is phase type with representation (v. see Fig. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . Hence ( 2. i.1) follows by the law of total probability. u Returning to nonterminating renewal processes . IIafl < 1. Corollary 2. However.T) where v = aT1 /µB. this is welldefined.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Proof Just note that { it } is a governing phase process for the lifetime. is the first k with Uk = 00. + Uit_1 where s. .T). and let µB = aTle be the mean of B.1) remains valid for that case. the density is veTxt = B(x)/µB. Hence the intensity matrix is T + ta. and the jumps of the first type are governed by T.T + ta).T) where vt = ae (T+ta)t . define the excess life e(t) at time t as the time until the next renewal following t. Then the lifetime is zeromodified phase type with representation (a. T)..3 Consider a renewal process with interarrivals which are phasetype with representation (a.U1 U3 U2 U3 U4 Figure 2.e. fi(t) U2 U1 . and hence ( 2. 2.IIBII which is > 0 in the defective case. Equivalently.e. . Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.1 Corollary 2.1. B is defective . the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter .224 CHAPTER VIII. since Uk = oo with probability 1 . u The argument goes through without change if the renewal process is terminating. i.. This is defined as U1 + . The renewal density at x is now just the rate of jumps of the second type. that is.
2) v(T + ta) = 0.e. hence e(t) is phasetype with representation (vt.4 Consider a nonterminating renewal process with two phases. T1 and eTx commute. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.) ( t2 ) . Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. i. u Example 2 . (ii) First check the asserted identity for the density: since T.T) where vt is the distribution of it which is obviously given by the expression in (a).q2.1. we first compute the stationary distribution of Q. cf. (2.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.2. Al. The renewal density is then aeQtt = (al a2) ( 7i 7"2. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say). Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.6. The time of the next renewal after t is the time of the next jump of the second type.e. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . According to Example A3. the unique positive solution of ve = 1. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.
)t (51 .4 yields the renewal density as u(t) = 2 (1 .4 yields the renewal density as u(t) = 5152 e. Hence 7r = (1/2 1/2). The present treatment is somewhat more probabilistic. and Example 2.e2bt) 13 Example 2 .a27r1) (t1 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. and Example 2.6 Let B be hyperexponential.52a1. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . MATRIXANALYTIC METHODS e.tl) 7r2t2 + eat (a17r2 .t2) .226 CHAPTER VIII.(biaz + aza. . A = 25. Then Q= 0 55 )+(1o)=( j ad ).`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .t2) 1 + eat (a17r2 .a27rl) (tl .5 Let B be Erlang(2). t1B 0 Example 2 .
) = F(ST(o) E •. with 0 denoting the Poisson intensity. and if there is a subsequent ladder step starting in j whic occurs w. G+(.e. cf.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. however. a+j. Considering the first. T(0) < oo) the ladder height distribution and M = supt>o St. Thus the total rate is tip + tia+. and rewriting in matrix form yields the phase generator of {my} as T + ta+. (b) V.3. the Markov processes representing ladder steps can be pieced together to one {my}. Corollary 2. {St} the claim surplus process. Then each claim (jump) corresponds to one (finite) sample path of the Markov process.1 Assume that the claim size distribution B is phasetype with representation (a. Here we have taken the terminating Markov process underlying B with two states. T) where a+ is given by a+ = . T). Corollary 3.2.3.i. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. 3. 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. B the claim size distribution. and M is zeromodified phasetype with representation (a+. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). which occurs at rate ti. r(u) the time of ruin with initial reserve u. add a more selfcontained explanation of why of the phasetype structure is preserved. Then: (a) G+ is defective phasetype with representation (a+. The stars represent the ladder points ST+(k). The essence is contained in Fig. We asssume that B is phasetype with representation (a.p. 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. we shall. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. i. marked by thin and thick lines on the figure. Next. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. Since the results is so basic. use the phasetype representation of Bo. T). . For (b). 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. Within ladder steps.1 on the next page.f3aT1. T + to+).
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 . MATRIXANALYTIC METHODS t .3. 0 Example 3. 3e3x + .. Figure 3.QaT1.2 Assume that . see Corollary 2..1 This derivation is a complete proof except for the identification of a+ with .Q = 3 and b(x) = .1 .M {mx} ST+(2)  S .t t d kkt S. T = (3 .228 CHAPTER VIII.. This is in fact a simple consequence of the form of the excess distribution B0.
6). but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. (a) G+ is of phasetype with representation (a+.2 are taken from Gerber [157]. We assume p = PB/µA < 1 and that B is phasetype with representation (a. 3. T). his derivation of +'(u) is different. 4 The renewal model We consider the renewal model in the notation of Chapter V.1 In the zerodelayed case.6.1): Proposition 4. For further more or less explicit computations of ruin probabilities. see Section 6. so that as there 229 9 9 e(T+ta+)u 1 9 e_u 10 70 10 70 7 10 Thus 1 7 9 10 ) + e6'4 ( 10 10 . It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially.4. the discussion around Fig. where a+ is the (defective) . T) for some vector a+ = (a+. The result carries over to B being matrixexponential. see Shin [340].j). we encounter similar expressions for the ruin probabilities in the renewal.and Markovmodulated models. That is. 3. the duality result given in Corollary 11. see Stanford & Stroinski [351] .T). if we define {mz} just as for the Poisson case (cf. with A denoting the interarrival distribution and B the service time distribution. THE RENEWAL MODEL This is the same matrix as is Example 1. For an attempt.1 can be found in Neuts [269] (in the setting of M/G/1 queues. For the compound Poisson model.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+.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.4. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). this was obtained in Section 3. cf. In the next sections. Fig. We shall derive phasetype representations of the ruin probabilities V) (u). The parameters of Example 3. and the argument for the renewal case starts in just the same way (cf. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).
6. Since the conditional distribution of my given T1 = y is ae4y.T)• Proposition 4. G(') = pBo. where B0 is the stationary excess life distribution corresponding to B.1) Proof We condition upon T1 = y and define {m. But by Corollary 2. In fact.*} from {St+y . with intensity matrix Q given by Q = T + to+. CHAPTER VIII.5.Sy} in the same way as {mx} is defined from {St}. where a(8) = aT1/PA. Then . 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+). (c) {mx } is a (terminating) Markov process on E. B0 is phasetype with representation (aT1/µa.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. (4. We have now almost collected all pieces of the main result of this section: Theorem 4 .230 distribution of mo. 4. the calculation of the first ladder height is simple in the stationary case: Proposition 4. which for numerical purposes can be solved by iteration.3. Hence by Theorem 11.1.T). Then {m. Fig.*'} is Markov with the same transition intensities as {mx}. Nevertheless.1).3 a+ satisfies a+ = V(a+). where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). the Palm distribution of the claim size is just B. cf. Proof Obviously. obviously mo = m.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). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4.T). but with initial distribution a rather than a+. Also. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.
•.2. a+ can be computed by iteration of (4. THE RENEWAL MODEL 231 . a+l ) = cp (a+°)) . .3) (defined on the domain of subprobability vectors . . The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3.2 ) follows from Proposition 4.^(u) = a+e ( T+ta+)xe. i.e.0.1 by noting that the distribution of mo is a+.0) is an increasing function of /3.1/pA. (4.1 .^(8)(u) = a ( 8)e(T+ta +) xe.4. I {mx} .. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. and that this is given by Proposition 4.1). a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .3) Proof The first expression in (4. i y ^ T1= y `•r Figure 4. Furthermore .1) and a(8) _ aT.M. only with initial distribution a(*) for mo.2) where a+ satisfies (4.. by a+ = lim a +n) where a+°) . The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. (4.3).. a+2) = ^p (a+l)) .1(b). thus .. It remains to prove convergence of the iteration scheme (4. Hence ^p(. In particular .
Theorem 4.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). For n = 0. .232 CHAPTER VIII. In that case. Similarly. we use an argument similar to the proof of Proposition VI. Fn ).1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). Obviously. (4.) = P(mTl = i.5 Let s be some complex number with k(s) > 0. so to complete the proof it suffices to show that &+ < a+) for all n. Thus . Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1.P[s] = A[s]B[s]. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. However. n) &+n) T a+. To prove the converse inequality.g. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). Then (4. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah.. the normalization is equivalent to F(s) = 1. Assume the assertion shown for n .5) yields h = (sI . F[s] being interpreted in the sense of the analytical continuation of the m. s ¢ sp(T). (4.ST. the corresponding right eigenvector may be taken as (sI . 0 0 We next give an alternative algorithm. 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). Thus by (4.f.4).} can contain at most n . Proof Suppose first Qh = sh. Then F[s] = a(sI .T)'t • A[s] (4.T)It. 7+ ]. To this end.5) Since s $ sp(T). and let &+".T)1t. both quantities are just 0 . Then each subexcursion of {St+Tl .4. let F be the distribution of U1 .4) whenever EeR(S)U < oo.1. and hence we may assume that h has been normalized such that ahA[s] = 1. with B[s]. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.2.T1. this implies that ahA[s] # 0. limn4oo a ) < a+. a+ ) exists .
we have IG_ [s] I < 1 . Then G+ is phase. hd.6... Q has diagonal form d d Q = dpivi®hi = dpihivi.T) = 1 ata+ = a+. 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. Q = CD1 where C is the matrix with columns hl.. This gives d roots 'y. yd satisfying R(ryi) > 0. Hence with h = (sI T). T) with a+ = a(QT)/at. Further. 0). (4.. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1.4.. Corollary 4. THE RENEWAL MODEL 233 Suppose next F(s) = 1.. .lt we get Qh = (T + to+)h = T(sI .. . .5.. .T)lt = sh. D that with columns p1 hl..... we get at a(Q . and the solution is .T)It.9) we have G+[s] = 1 which according to Theorem 1.T)lt + t = s(sI ..1 has the d distinct eigenvalues .6 Suppose u < 0. Pd with corresponding eigenvectors hl.. In older literature .type with representation (a+. 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. As in Corollary 4..6) i=1 i=1 Proof Appealing to Theorem 4. and hence by the WienerHopf factorization identity (A. W v M(d) in the notation of Chapter V). Given T has been computed.p1i . in turn. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . Notes and references Results like those of the present section have a long history. . and define hi = (piI . Pd in the domain ER(s) > 0 .' that the equation F(s) = 1 has d distinct roots p1.. the matrix Q in Theorem 2.5(c) means that a+(sI T)1t = 1. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . pdhd. t(ry) > 0. . The roots are counted and located by Rouche' s theorem (a classical result from complex analysis giving a criterion for two complex functions to have the same number of zeros within the unit circle ). hd.. Since R(s) > 0 and G _ is concentrated on (oo. . Let d denote the number of phases.
the ruin probability can be found in matrixexponential form just as for the renewal model. [119]. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ).type assumptions are basic. whereas the approach was introduced in queueing theory by Smith [350]. In risk theory. For surveys . and appears already in some early work by Wallace [377]. involving .234 then in transform terms CHAPTER VIII. see Dickson & Hipp [118]. and the distribution of an arrival claim is B. The solutions are based upon iterations schemes like in Theorem 4. 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. starting around in 1975. is phasetype.F. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. The arrival rate in background state i is a. T('). 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. see Neuts [269].exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. We assume that each B. The distribution of W comes out from the approach but in a rather complicated form .. The exposition here is based upon [18]. with representation say (a(' ). e. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. Here phase. the fixpoint problems look like R=Ao+RAI+R2A2+ .contained derivation). the intensity matrix is A and the stationary row vector is ir . [270] and Latouche & Ramaswami [241]. In queueing theory. where R is an unknown matrix.4. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. Numerical examples appear in Asmussen & Rolski [43].type assumption . a pioneering paper in this direction is Tacklind [373].. the background Markov process with p states is {Jt}. For further explicit computations of ruin probabilities in the phasetype renewal case . The matrix. That is . Neuts and his students.) d (see.. It turns out that subject to the phase. Asmussen & O'Cinneide [ 41] for a short self. E(t)). The number of elements of El=> is denoted by q. an alternative approach (the matrixgeometric method ) has been developed largely by M.g.
MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. The key unknown is the matrix K. •. Diagonalization Consider a process {(It. states . 5a Calculations via fluid models. 5.1.1. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. has states o. However. O. p = ql = Q2 = 2.6. 5. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. The two environmental states are denoted o.1 In Fig. 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).4. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. 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. the phase space E(°) for B. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. say with slope r(i) on intervals where It = i. and the one E(•) for B. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. Section 5b then gives a representation along the lines of Theorem 4.Vt)} obtained by time reversing the I component.5.2.
The intensity matrix for { It} is (taking p = 3 for simplicity) I A . of E into components indexed by E. t. 4. a) = 1. •. '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. we have more martingales at our disposal. V. a) : i E E. 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. i E E.1(a).1) if and only if s is an eigenvalue of E.236 CHAPTER VIII. This implies that in the fluid context. in the fluid model Eel'. 4}. corresponding to the partitioning + Epp). A claim in state i can then be represented by an E()valued Markov process as on Fig. The fluid model on Fig . j = 1. < oo for all s. If s is such a number. 2. F = E U { (i. a E E(i) } . Second. MATRIXANALYTIC METHODS 4. consider the vector a satisfying (A + (13i(Bi[ s] . F is the disjoint union of E and the Eli).1))diag + sII = 0 (5. Recall that in the phasetype case. whereas Ee8s' = oo for all t and all s > so where so < oo.1(b) {(It . resp. o. Thus F = {o. r(i) _ 1. Bi[s] = a(i)(T(i) + sI)it(').1))diag ) a = sa and the eigenvector b = . 5. 5.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. Let E denote the matrix . First. a) of {It}. Eli) + Proposition 5. 4.1 A complex number s satisfies 'A+ (f3i(Bi[s] .31a(l) (/3i)diag .(Ni)diag r(i.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. In the general formulation .
E21a + sd = sd. d = (sI E22)1E21a = E ai(sI . c = a.A . Then E21c+E22d = E21a . Then (up to a constant) c = a. it follows that if Qla(1) 0 0 . and let d = (sI .32a(2) (/3i)diag . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .5.sI+ ((3ia(i)(T(i) . E(1) + + E(P). Noting that E11c + E12d = se by definition.E22)1 E21a E21a .sI 0 0 0 T(2) . 0 .sI)1t)) iag I = 0 which is the same as (5.E22)1 E21) a = 0.sI ()3i)diag . MARKOVMODULATED INPUT 237 indexed by E. with Eii replaced by Eii .sI 0 0 t(3) 0 0 = 0. For the assertions on the eigenvectors.sI) (sI . resp .E22)1 E21a. t(1) 0 0 then also 0 t(2) 0 . where c.Nla(1) 0 0 T 1.1).sI.E12E22 E21 I . iEE (a> of 0* 1 AI.(sI .T('))1t(i) . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell . assume that a is chosen as asserted which means (Ell .A .sI 0 0 0 T(3) .E22 .sI + E12 (sI .
B2 are both exponential with rates 51 i b2. v. a). Iw(u. j. a) and noting that i1 (u) = >I j.( u. Here E has one state only. Example 5 .Q.4 that {e"1b(v) is a martingale ...4 Assume that E has two states and that B1.. .. j) pi( u .v)=inf{t >0:Vtu orVt=. Then . .v) = (j. j. v > 0. To determine 0 (u). sq with $2s.upi(u.j. v. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . w(u)=inf{t >O:Vtu}. c j.v) = Optional stopping at time w (u.. . s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 .. a )d(a + e8 °vpi (u . a)).. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p.pi(u. u) Iw(u. j.3 Consider the Poisson model with exponential claims with rate 5.sv)b(v) = 0. a)).j)c v . the result u follows. pi(u. Thus 0(u) = esu/d = pe7 ° as u should be. define w(u.. v) yields C{V) = e8 . q.v. Then we get V)i (u) as sum of two exponential terms where the rates s1.. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. a) = (j.v) = v) I.5. d("))1 e. Proof Writing Or'Alb( v) = svb( v) as (AI . For u. a) = Pi (Vw(u. .v) = j). Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. ./' u = e' (esiuc ( 1) .. j.v) = = p i( u . it follows by Proposition II. I' i( V P2 (w (u) < oo.. v.a)d^ ). j.v}..238 CHAPTER VIII.. MATRIXANALYTIC METHODS Theorem 5. e89uc(e)) (d(1) . w(u. Example 5 .a Solving for the pi(u.O.. v = 1.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. .
5 G+(i.3j eye. 0 Theorem 5 . j. the Pidistribution of M is phasetype with representation (E(1) + + E(P). U) where t(j) + t(j)O(j j = k uja.6 For i E E. 8^')IT(j)) where e 3^') =. •) is phasetype with representation (E(i). (y.x) 00 f ° (') (j) eT (yy)edx . i.5.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. (5.s.33(e = 0 a(j))(K ®T ( j))(ej (9 I). we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5.xxej • a 00 oo el .h.y = to B k7 j # k In particular.2) the l. In terms of K. (') a T( However . Proof We must show that G+ (i. is 0 /3 f R(i . dx)Bj(y .( 2.b (u) = Pi(M > u) = 9(i)euue. j. oo)) j)ye. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.2.k. according to VI. j.3) . 9(').
Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. the ratio between two polynomials (for the form of the density. if b* [0] = b1 +b20+b302 +.2) . For a transition from (j. . and lifelength M. and a new ladder step of type k must start in phase y.k y.) which is rational..240 CHAPTER VIII. i.e.p. Starting from Jo = i. Bk7 . Then b*[0] is rational if and only b(x) is matrixexponential.e. we have the additional possibility of a phase change from a to ry within the ladder step. Furthermore. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). say. some square matrix T and some column vector t (the triple (a. intensity matrix U. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . +bn0i1 0n +a10n1 +. However. we have sofar concentrated on a claim size distribution B of phasetype. (6.g. bn1 bn)... which occurs w. the current ladder step of type j must terminate. a) is obviously chosen according to e(`). t) is the representation of the matrixexponential distribution/density): Proposition 6. a) to (k.. T. which occurs at rate t^^7. that the density b(x) can be written as aeTxt for some row vector a. t = (0 0 . and it just remains to check that U has the asserted form. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. Associated with each ladder step is a phase process. For j = k. see Example 1.. a m. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. 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. Numerical illustrations are given in Asmussen & Rolski [43]. i. which occurs at rate t(i).y) to occur when j # k.1 Let b(x) be an integrable function on [0. An alternative characterization is that such a distribution is matrixexponential.5)... equivalently. the initial value of (i..f. 0 1)'. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 .2. This yields the asserted form of uja.
One of his elementary criteria. .2 A remarkable feature of Proposition 6. Thus. 1 .an_3 an _ 4 . S = f c/2 0 21ri . (6. t= 0 .3) was suggested by Colm O'Cinneide.3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]./(0 + bi).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. T. 0 0 0 0 1 0 0 . .s. namely to asssume the roots 6l. we can always obtain a real one (a.1 0 0 )3 = (111). Writing b(x) = c(e( 2ni1 ) y/2 . cannot be phasetype. t).3) 0 0 0 0 0 . shows that the distribution B with density b(x) = c(1 cos(21r x))ex. Example 6 . since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6. s = c/ 2 . where c = 1 + 1/47r 2. u Remark 6..2). personal communication). bn of the denominator to be distinct and expand the r..T)1 is so.1 0 . matrixexponentiality implies a rational transform.47x2 3 1 0 . s) is given by 27r i . a2 a1 Proof If b(x) = aeTxt. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . u giving b(x) = E 1 ciebiz/bY..3) that we can take 0 1 0 0 a= (1 + 47r2 0 0)...1.h. b(x) > 0 for x > 0. Namely. For a proof. (6..2).4) 0 0 1 c This representation is complex.6. S.47r2 3 . .. 0 0 . . see Asmussen & Bladt [29] (the representation (6.e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3.1) as E 1 c. of (6.. T= 0 0 1 . . but as follows from Proposition 6.(6. The converse follows from the last statement of the theorem. 0 1 an an1 an _2 . then b*[0] = a(0I T)1t which is rational since each element of (01 . (6..
6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. T.6) holds true also in the matrixexponential case.242 CHAPTER VIII. (6. that despite that the proof of (6. and can use this to invert by the method of Proposition 6. leading to matrix calculus in high dimensions when b is small. (6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . q are polynomials without common roots. T the phase generator and t = Te.1)2 + 6). . MATRIXANALYTIC METHODS Example 6 . we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. For the second algorithm.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)) .5 (6.6) The remarkable fact is. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). recall that t = Te) that if B is phasetype and (a. T. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. then: Proposition 6. Then (cf.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . For the first. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.1 to get i (u) = f3esus. As for the role of matrixexponential distributions in ruin probability calculations. Corollary 111.5) Thus. We recall (see Section 3. and present two algorithms for calculating '(u) in that setting. t) a phasetype representation with a the initial vector. t) of b(x). we take as starting point a representation of b* [0] as p( O)/q(9) where p. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. Consider the distribution with density = 15 ((2e2x . 0.3. we use a representation (a.
T)1T1t. Now.T .1BVA1.T .7) 9( cf.T)1t.'t.A .T)1t)1a +(9I .a+(9I . U =. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain . From the general matrix identity ([331] p. but we shall give an algebraic proof.T)1 + 1 ib* (91.1UB(B + BVA1UB).to+)1 = (BI . (91.to+)1T .5 ).1t = f3a (0I T)1T1t .6b* . b+ = a+(9I .1t = b* . 519) (A + UBV ). with A = 91T. b+ = a +(BI .T .1 + b+ = b++ 1 . we get (91. (6.1t du = .T)1 J0 00 b(x) dx = f aT1t. we get b+ = 0aT1(9I T).B=land V=a+.6).T)1t ( l .1 = ^(T1 + ( 91T)1). xb(x) dx = aT2t.T)1 + (6I . (6.b* (6. Presumably.6.1 + 82 (9I .T)1ta+(OI .to+)1T . since (91T)1T .T)1 so that b* b** b** a+(9I . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .t. the assertion is equivalent to a+(BI .T)1T 2 = and 1 = AB IT2 + 82T .T)1 (91.1 = A1 . . Then in Laplace transform formulation .T).
From this it is straightforward to check that b**/(b+ . 0 Notes and references As noted in the references to section 4. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. For expositions on the general theory of matrixexponential distributions.1. the ruin probability(u) was found in explicit form for the case of B being exponential. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. 3.T)1T. T).8. 7. a. VII. Lipsky [247] and Asmussen & O'Cinneide [41].T)1T2t .5 is similar to arguments used in [29] for formulas in renewal theory. which is selfexplanatory given Fig. . see the Notes to VII.1. The proof of Proposition 6.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E.T)1)t = 8 (1 .8 a(T1 + (01.1t = /3a (9I .1.1) is the same as the r.7). In Corollary VII. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. but the argument of [286] does not apply in any reasonable generality). see Asmussen & Bladt [29]. 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. /3aT1(0I .s.82b*. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4.1. a key early paper is Cox [90] (from where the distribution in Example 6.244 CHAPTER VIII. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. of (6. See Fig.b*).3 is taken).h.la. cf. Much of the flavor of this classical approach and many examples are in Cohen [88]. MATRIXANALYTIC METHODS .
7. t2) = exp where Q(t) = ds [P(t. In fact. t) is the vector of state probabilities for mt.e. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. Proof The first statement is clear by definition. Also. is no longer timehomogeneous.1) where A(t) = v(u)P(0. Note that in general >iEE Vi (U) < 1.u)e = A(u)e (7. O<.tl < t2 < u. though still Markov.1z I.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i. i. Ai(t) = P(mt = i). t + s) . the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. Given the v(t) have been computed. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .I] I 80 { tq f Q(v) dvl t1 1 .t2) be the matrix with ijth element P (mt2 =j I mtl = i). in contrast to Section 3. Let P(tl. we obtain V)(u) = P(m„ E E) = v(u)P(0. 0 < t < u. the definition of {m8} depends on the initial reserve u = Ro. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •).t)). P(tl. By general results on timeinhomogeneous Markov processes. Figure 7.1 The difference from the case p(r) = p is that {m2}.
jEE . Hence Q(t) _ T + tv(u . Thus. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). Given A.Sj i)p(u)dt • tji = Sji + p(u)tji dt. 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. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . (7. the probability of which is 1 . Given A'. 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. we get vi(u) = aidt + (1 . two things can happen: either the current jump continues from u + p(u)dt to u. whereas in the second case the probability is p(u)dt • tjvi(u).246 CHAPTER VIII.t)).t).t) for the second. or it stops between level u + p(u)dt and u. Given this occurs. 0 < t < u. dt]. 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. Proposition 7. A'(t) = A(t)Q(t) = A(t)(T + tv(u . 0 Thus. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . {mx} has jumps of two types. vi.t and being followed by a downcrossing.2 For i E E. given A. the probability that level u is downcrossed for the first time in phase i is ai. from a computational point of view the remaining problem is to evaluate the v(t).(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). MATRIXANALYTIC METHODS However.3dt. In the first case.(u) p ( u) = .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.
5)..7. Rt . we have p(r) = p = vi (u) 0aTe.i7rT1/p. we can first for a given v solve (7.2. When solving the differential equation in Proposition 7. Let p" (t).) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. then P(A n Bv) _ P"(A n BV').s. say. Since the processes Rt and Rt coincide under level B. . Thus. starting from v"(v) = . of (7.)P"(AnB. Then P(B. F" etc. denotes the time of downcrossing level u . and similarly P"(Bv) .^ 0.) P"(AnBv) = P(AnB. (u) on both side and dividing by dt yields the asserted differential u equation.4) backwards for {va (t)}v>t>o..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. we face the difficulty that no boundary conditions is immediately available. To deal with this. vi (U) = lim v= (u). after a certain level v. (u) for any values of u and v such that u < v.3 For any fixed u > 0. <oo. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. RESERVEDEPENDENT PREMIUMS 247 Subtracting v.. From Section 3. V . supRt>v l t<7 I where o. say. refer to the modified process. This yields v. Now since both P(A n Bv) 3 0 and P"(A n Bv) .) + 0 as v + oo. consider a modification of the original process {Rt} by linearizing the process with some rate p.. (v) is given by the r. P u which implies that v.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB.h.
1a. To evaluate p1(u). while the fourthorder RungeKutta method implemented in [30] gives 0(n5).. p(r) P.e.. 1/n. 0 < u < v. v = u. < 0}).7) equals 01 (v . Therefore u pl(vvueTa t 1. say. 2/n. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.RQ (defined for or < oo only) is defective phasetype with representation (v(u). i. Corollary 3. 3u etc. such that Rt coincide with RI under level v and with Rt above level v. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. However.v v(u)eTat 1 1 . Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations.) are so. 2u. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. cf.1. p2.9. T).zp1(u)/(1 . the probability of ruin between a and the next upcrossing of v. The precision depends on the particular quadrature rule being employed.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. for u > v the distribution of v .. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. assuming u > v for the moment. The trapezoidal rule used in [288] gives a precision of 0(n 3).z51(v)). (u)}.10 that in addition to the O'(•). MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.248 CHAPTER VIII. the evaluation of Vi(u) requires q(u) = 1 . (7. let v(u) = a+2ieiT +ta+>)(uv). where v = inf It > 0 : Rt < v}.V" M 0 . The algorithm based upon numerical solution of a Volterra integral equation (Remark VII...1.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.7) f o (the integral is the contribution from {R. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. which is available since the z/i'(. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. > 0} and the last term the contribu tion from {R. as well p1(u).x) dx f v(u)eT xt dx . r<v r > v. where. The f iin in (7. We recall from Propositon VII.1.
1 from which we see that pl (u) = 1 + 1 249  1 . I. RESERVEDEPENDENT PREMIUMS 1 . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .x) dx} V 1 1(v) f V v(u) eTxt.7.01 (v . From Example 3.2.^1(v) 1 .24e.and A2 = 3 ..jl (t ®e) Thus.be the eigenvalues of T + to( 2 ).e.v) + (2^ + 3v2 ea'(u " .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. Example 7. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. Since µB = 5/21.21 = ? yields 0(u) = 1.2.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . all quantities involved in the computation of b(u) have been found in matrix form.u .8) The integral in (7.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. (7.(7 The arrival rate is (i = 3. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .v(u)eTVe .4 Let {Rt } be as in Example 3.v(u)eTve).x) dx 1 ^(v) ( 1 .e6u 35 .to+))11 {e{T®(Ttoy+ ))}„ .24ev .2V"2.v(u ) eTV e J v(u)eTxtz/)l (v .e6v Let Al = 3 + 2V'2. p2 < 3. so we consider the nontrivial case example p2 = 4 and p1 = 1.
1./2) ea 1(u ./2 ea1(u") .1 Thus. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .. and one gets 12e5" .250 CHAPTER VIII. 21 3 In particular.24es" . 192esv + 8 P1 .24e5v .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4. The analysis and the example are from Asmussen & Bladt .2 35e6v .24e5v .1)' ?.1 V2 = 4e5"+6 35e6v .24es" .21(35e6v .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. ) e sv + ( 2v/2. MATRIXANALYTIC METHODS From (7.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v. Notes and references [30].v)esv + 7 4_ 2.
Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail.f. for all t > 0.B(x).46 and at numerous later occasions require a light tail. A rough distinction between light and heavy tails is that the m. x 4 oo.N(µ. we require that B is concentrated on (0. a2)) with density 1 e(logyFh) 2/2az . For further examples.g. the exponential change of measure techniques discussed in II. L(tx)/L(x) 4 1. 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. For example. III.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) = L(x)/x" where a > 0 and L(x) is slowly varying.2b. (b) the lognormal distribution (the distribution of eu where U . x 2iror2 (c) the Weibull distribution with decreasing failure rate . B(x) = ex0 with 0<0<1. oo ) and say then that B is subexponential (B E S) if 251 . and instead we shall work within the class S of subexponential distributions . The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. For the definition .4. see I.
1 Let B be any distribution on (0.'s X1. That is.1(b) is oo. then (with high probau bility) so are both of X1. P(max(Xi. In terms of r.v. X1 is w. note first the following fact: Proposition 1. the behaviour in the lighttailed case is illustrated in the following example: Example 1. 1). that is. 1/2 it has the distribution of X1I X1 > x. the r. X2) > x}. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. (b) liminf BB(() ) > 2.p.v. Thus the liminf in Proposition 1.3 Consider the standard exponential distribution. the distribution of independent r. B(x) Here B*2 is the convolution square. we have {max(Xi. We later show: Proposition 1. B(x) ax. . Then: (a) P(max(Xi. X2 with distribution B. X2 but none of them exceeds x. (1. x 3 00. In contrast. if X1 + X2 is large . Since B is concentrated on (0. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. P(Xi <yI Xi+X2>x) 1B(y). Thus .B(x)2 . proving (a). X2) > x) is P(X1 > x) + P(X2 > x) .F(X1 > x. That is.252 CHAPTER IX. oo). X2) > x} C {X1 + X2 > x}. X2) > x) ^' 2B(x). X2 > x) = 2B(x) .2B(x).2.v. and thus the lim inf in (b) is at least lim inf P(max(Xi. To capture the intuition behind this definition. HEAVY TAILS B*2\ 2. one can check that x x where U is uniform on (0.1) then means P(X1 +X2 > x) 2P(Xi > x).p. Proof By the inclusionexclusion formula. given X1 + X2 > x. The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi. X2) > u x)/B(x) = 2. oo).'s.2 If B E S. As contrast to Proposition 1. 1/2 'typical' (with distribution B) and w. then P(X1>xI X1+X2>x)* 2.
[In terms of r. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.xIX > x converges in distribution tooo. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. and combining with Proposition u 1.'s: if X . we get limsupB*2(x)/B(x) < 2.6)x)/((1 .] Proof Consider first a fixed y.B*(n ) B(dz) (1. a contradiction.B(y) = 2.v. If X1 + X2 > x.1.y)/B(x) > 1. 1 < B(x ) B( x) Y) < B( 0). then B(B(x)y) * 1 uniformly in y E [0. B( 0 .yo].S)x. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. Hence lim sup a+oo B*2(x) 2B((1 .B(y)) . Finally lim inf B(x . Proposition 1.B E S.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.4 Any B with a regularly varying tail is subexponential. or they both exceed Sx. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . x].1(b) we get B*2(x)/B(x) * 2. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . Let 0 < 5 < 1/2. then the overshoot X .y)/B(x) > 1 since y > 0. We now turn to the mathematical theory of subexponential distributions. If lim sup B(x .S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. y] and (y. then either one of the Xi exceeds (1 .z B(x) .B*n(x . yo] as X + 00.5 If B E S.
O The following result is extremely important and is often taken as definition of the class S. so assume the proposition has been shown for n.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1. Then by (1. b[c] = oo for all e > 0. then for any n B*n(x)/B(x) * n.nI < e for x > y. Proof For 0 < 5 < e.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .z ) (x ) = 1 + (^ B(x .z) B(dz).5 and dominated convergence. This implies B(x) > c2e5x for all x. Proposition 1.254 CHAPTER IX. x oo. and this immediately yields the desired conclusions. Given e > 0.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . The case n = 2 is just the definition. its intuitive content is the same as discussed in the case n = 2 above.2. Proof We use induction.7 If B E S. 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 .1) for all large n so that B(n) > cle6n for all n.5 that B(n) > e6B(n .y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. HEAVY TAILS Corollary 1. 0 Proof of Proposition 1.6 If B E 8.5 and the induction hypothesis. The first integral is y B(x . we have by Proposition 1. choose y such that IB*n(x)/B(x) . then e"R(x) * oo..z) B(dz) _y B(x) 111 Lx B .2).B(x .B*2 (x) B(x) (x . B*(n+1) (x I xy + Jxx y) W. B(x) \Jo _ B(x .
X2 > xv) < A1(xv)A2(x v) . A2 be distributions on (0. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. Proof Let X1. Combining these estimates and letting a 4. a2 with a1 + a2 > 0.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .(al + a2)B(x).4) . X2 be independent r.3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Then Al * A2 (x) .v.z) B(x .B(x . Proposition 1.z) B(dz) x . x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e. Since P(X1+X2 > x. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.z) B(dz) < 1 + A + an(1 + d) . an = supx>o B*n(x)/B(x). Xi <= v Ai (x . Then by (1. an+1 fX B*n( *n(x . Then Al * A2(x) = P(X1 + X2 > x).y)Ai(dy) = (x)o(1) (1. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .8 If B E S.X1 > xv. Proof Define 5 > 0 by (1+5)2 = 1+e.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .5 easily yields P(X1 + X2 > x.i).z) B(x) < 1 + A + an sup f x B(x . e > 0. 0 Proposition 1. For any fixed v.'s such that Xi has distribution Ai. 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) .0 completes the proof. 0 Lemma 1.y)Ai(dy) v) f o .2).y)B(dy) = B(x)ov (1)• v (1.9 Let A1.ala2B(x)2 which can be neglected. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T).1.y))/B(x).
u It is tempting to conjecture that S is closed under convolution.(x) is decreasing for x > x0 with limit 0 at oo. if q(x) aB(x) for some B E S and some constant a > 0.11 Let B E S and let A be any distribution with a ligther tail. of (1. it is easy to see that if L1. A2 = B so that a1 = 0. B1 * B2 (x) . L2 are slowly varying.2.256 Now (1.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.aiB(v)) = B(x)o„(1).3) follows if CHAPTER LX. u Corollary 1.B(x) Proof Take Al = A. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.s.h. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«.5) becomes x B(x . A(x) = o(B(x)). whereas the two first yield B(x)(Ai(v) . i = 1. a2 = 1. u Corollary 1.2aB(x) .y)B(dy). That is. HEAVY TAILS 'VV B(x . That is. a1 = a2 = a yields A*2(x) . then so is L = L1 + L2. B2 E S.5) Here approximately the last term is B(x)o„(1) by ( 1. In the regularly varying case. the l.Ai(x .12 Assume that Bi(x) = Li(x)lxa.2A(x). However.y)Ai(dy) = B(x)o„(1).v)Ai(v) . f " By a change of variables. with a > 0 and L1.v)B(v) + _'U Aq(x . then A E S. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.9). L2 slowly varying. Then A * B E S and A * B(x) . V (1. Then B E S provided fo "O exA(x) b(x) dx < oo. Proof Taking Al = A2 = A.13 Let B have density b and failure rate A(x) such that . Hence Corollary 1. . We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. it should hold that B1 * B2 E S and B1 * B2 (x) .4).Bl (x) + B2 (x) when B1.10 The class S is closed under tailequivalence.
Thus by dominated convergence . Jo For y < x/2.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).A(xy)A ( y). elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. an integrable function by assumption. Goldie & Teugels [66]): Proposition 1. To illustrate how Proposition 1.1 has limit 1 + 0. Thus.U) /or) v 2x This yields easily that ex. Thus A(x) is everywhere decreasing. the first integral has limit 1 . In the regularly varying case. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Since ) (x .3 < 1.2). f ' L(y) dy .y) < yA(x . The middle bound shows that it converges to b(y) for any fixed y since \ (x .12. Thus B*2(x )/ B(x) . By (1.. we can use the same domination for the second integral but now the integrand has limit 0 .(x . the DFR Weibull distriu bution is subexponential. Example 1. Then b(x) = Ox0lexp.A(y)a(y ) = ev'(y) b(y).14 Consider the DFR Weibull case B(x) = ex0 with 0 <. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. Define A(x) = fo . proving B E S. L(x) y° (a .(y) dy.15 In the lognormal distribution.`(x)b(x) is integrable. subexponentiality has alrady been proved in Corollary 1. and exa(x)b(x) = (3x01e(10)x9 is integrable.1)xcl1 . Example 1.16 For L(x) slowly varying and a > 1.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx ..y) < A (y) for y < x/2.1. Thus. x .y) * 0. Further.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).13 works in this setting.y) dy. we first quote Karamata's theorem (Bingham.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . 0 A(x) .A(x . a(x) = ax01. Then B(x) = eA(x). the u lognormal distribution is subexponential. B*2(x) .A(y)\(y) dy + fox/ 2 eA(x ).
1)]) xa L(x) (x[1 + y/(a . Then 7(x) .y(x)B(x).13 may present a problem in some cases so that the direct proof in Proposition 1.E(X . HEAVY TAILS Proposition 1.1)] I X > x) L(x[1 + y/(a . let X W = X .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1. 'y(x) = EXix>.258 From this we get CHAPTER IX. More precisely. Then: Proposition 1. the monotonicity condition in Proposition 1.x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a .ea b(x) is integrable.6) EX(x) . (1 + y/(a . the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. Proof ( a): Using Karamata's theorem. f O B(y) dy . However.xjX > x.L(x)/x" and )t(x) . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.1) and P(X (.4 is necessary in full generality. Thus exa(x)b(x) .1)X(x)/x > y) = P(X > x[1 + y/(a . we get (1.1/A(x) and P(X ixil'Y (x) > y) * e'.1))a .a/x. (c) Under the assumptions of either ( a) or (b). . then 7(x) x/(a .1)])a 1 1 . yo] . then B(x) .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.)/Y(x) > y) (1 + y/(a .
Theorem 2 .2 Let Y1. 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. cf. Lemma 2... .7) is referred to as 1/A(x) being selfneglecting. Kliippelberg & Mikosch [134]. P The proof is based upon the following lemma (stated slightly more generally than needed at present).A(x) I X > x) = exp {A(x) . Bo(x) = f0 B(y) dy / µB.and lognormal distributions . r(u) = inf it > 0.14. with EzK < oo for some z > 1. 1. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).t be the claim surplus at time t and M = sups>0 St.. We get p(yl+. The remaining statement (1. u a oo.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . be i. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Let St = Ei ` Ui . Examples 1.(x).2. It is trivially verified to hold for the Weibull. then Vi(u) P Bo(u).EK G(u). nG(u).1/.8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. 0 G(u) L G(u) .1 If Bo E S.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. Recall that B0 denotes the stationary excess distribution. St > u}. . Y2.+YK> u) = ^•P(K = n)G* n(u ) .n0 1•P(K= n)•n = EK. Then P(Y1 + • • • + YK > u) . i. d. Notes and references A good general reference for subexponential distribution is Embrechts.15.. with common distribution G E S and let K be an independent integervalued r.nn.v.
Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. Proof of Theorem 2.260 CHAPTER IX.µB(01 . then Bo(x)/B(x) + 00.13). The approximation in Theorem 2. lognormal . Since EK = p/(1.1.x^ ) B(x) _ f or ( lox .18. . x 4 00. Bo E S. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. The problem is a very slow rate of convergence as u ^ oo.µ J ) . mathematically one must note that there exist (quite intricate) examples where B E S. Borovkov [73] and Pakes [280]. For some numerical studies.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. u The condition Bo E S is for all practical purposes equivalent to B E S. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant..1 is essentially due to von Bahr [56].1) In particular . r(1/Q) xlQexp B(x) = ex' From this . Weibull) one has Bo(x ( B(x) . see Abate. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) .1 is notoriously not very accurate.p) and EzK < oo whenever pz < 1. the result follows immediately from Lemma 2.. Bo ¢ S.1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) . a].x400 PBB(x) PB Leta+oo.2. u x+a Notes and references Theorem 2. See also Embrechts & Veraverbeeke [136]. (2. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . The PollaczeckKhinchine formula states that (in the setup of Lemma 2.3 If B E S. Note that in these examples . In general: Proposition 2. Bo is more heavytailed than B . P(K = k) = (1. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. Bo E S is immediate in the regularly varying case. as well as examples where B ¢ S.p)p'. in our three main examples (regular variation .?(xµ 8 (x). However.
in [219] p. + Xn. To Bo(u) u + 00. i=1 . ..1 when u is small or moderately large. (3.e.. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. i.9+ < oo) = P(S. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V.+ E A. Kalashnikov [219] and Asmussen & Binswanger [27].1 gives 1010.1) this end . let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. We assume positive safety loading. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Then l/i(u) 1 P P [Note that (b) in particular holds if B E S.Ti....1.] The proof is based upon the observation that also in the renewal setting. E. G+ (A) = P(Sq+ E A. In [1]. T1 the ith interarrival time and Xi = U. also a second order term is introduced but unfortunately it does not present a great improvement. This shows that even the approximation is asymptotically correct in the tail.} Then ik(u) = F ( M > u) = P(i9 (u) < oo).3. Then K M=EY.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. Snd) = Xl +. Define further 0 = IIG+II = P(r9+ < oo). M = sup s$ .. The main result is: Theorem 3 . {n= 0.. Let U= be the ith claim . THE RENEWAL MODEL 261 Choudhury & Whitt [1]. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).y)/B (x) > 1 uniformly on compact y internals. p = iB /µA < 1. Based upon ideas of Hogan [200]. Somewhat related work is in Omey & Willekens [278]. there is a representation of M similar to the PollaczeckKhinchine formula. [279]. T+ < oo) where r+ = T1 + • • • + T. t9(u) = inf {n : Snd> > u} .g.
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.1) is equivalent to P(M > u) " .2).d.Ti).B(x).. the contribution from the interval (. x * oo. FI (x) _ fz ° F(y) dy. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. The heuristics is now that because of (b).FI(u). and hence FI(x) .3) and we will prove it in this form (in the next Section.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.. Write G+( x) = G+ ( x. P(K = k) = (1 . x + oo.FI(x) /IPG_I.Y2. are independent of K and i. 0] normalized by IPG_ I so that we should have to G+(x) . U_ (dy) is close to Lebesgue measure on (.9)9'' and Y1. cf. we will use the fact that the proof of (3.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.(. B(x) _ J O° B(B(x)y) A(dy) f 1 . oo) = F(S.PBBo(x).3 G+ (x) . Proof Let R+(A) = E E'+ ' I(S. A(dy) = 1. (3.. 0] to the integral is O(F(x)) = o(FI(x)).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_. Proof By dominated convergence and (b). whereas for large y . A.262 CHAPTER IX.1) holds for a general random walk satisfying the analogues of (a).g+ > x. d+ < oo).i. Let F denote the distribution of the Xi and F1 the integrated tail. x > 0. HEAVY TAILS where K is geometric with parameter 9. with distribution G+/9 (the distribution of S.y+ given r+ < oo). Lemma 3 . G_(A) = P(S. As for the compound Poisson model. 0 The lemma implies that (3. Then 0 0 F( x . (b) and does not rely on the structure Xi = Ui . u a 00.N.2 F(x) .1 IPG_ I / F(x . Lemma 3 .y) dy = 1 Pi (X) oo IPG_ I .oo.
2). n] + 1/I µG_ I.G+[s]) . In the lattice case.9)IpG_ I Differentiating the WienerHopf factorization identity (A. Hence using dominated convergence precisely as for the compound Poisson model.3.UG_ I x. Similarly. we can assume that the span is 1 and then the same conclusion holds since then U(n .1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. and in the last that FI is asymptotically proportional to Bo E S. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . then by Blackwell 's renewal theorem U_ (n . F(Y= > x) FI(x)/(OIp _ 1). the proof is complete. > (1 . n] is just the probability of a renewal at n.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .1.F[s] = (1 .O[s])(1 .3. THE RENEWAL MODEL 263 We now make this precise.1.1. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .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. Given e. n] < (1 + e)/1µc_ I for n > N.1 I . By Lemma 3.1. If G_ is nonlattice.(dy) fN FI ( x) + lim sup ZY00 N F(x . (3.=1 BIp G_ I (1.9) 1 . u Proof of Theorem 3.2) yields 00 F F I (u) P(M > u) _ E(1 .1. and that U_(n .y) U.0)0k k I(u) A.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. choose N such that F(n .y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.
S+q(u) .a.a. on the set {M > u. u)).(1 .. Proof Let w(u) = inf {n : Sid) E (u . with roots in von Bahr [56] and Pakes [280]. u)..So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u .1 is due to Embrechts & Veraverbeke [136]. must attain a maximum > 0 so that P(M > u. S+9(u) . Note that substantially sharper statements than Lemma 3.u)) = o(P (M > u)) = o(FI(u)).AB iP We conclude by a lemma needed in the next section: Lemma 3 . Notes and references Theorem 3. 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.(u)+n . see Asmussen & Kliippelberg [36].a) N P(M > u). Therefore by Lemma 3.4 on the joint distribution of (S.264 and letting s = 0 yields CHAPTER IX. .l.IIG+II)µc_ = (1 . FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .a.1)6+[0] . allowing also for possible dependence between the arrival process and the claim sizes. Sty(u) .0)ua_ . u)) > P(w(u) < oo)(i lp (0))• On the other hand. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. P(M > u. we have P(M E (u .2.4 For any a < oo.SS(u)}n=o. Mn < u}..Sty(u)_I < a} we have w(u) < oo.Se(u)_1 < a) = o(Fj(u)).So(u)) are available. Then P(M E (u .yiui_1. and {Su. HEAVY TAILS µF = (1 .a. 10(0) But since P(M > u .
We give here one of them. E0. M* = max S. M = sup St.1) . {SX1+t ... assume pp.X1 is the generic cycle. 4. Theorem 4... (corresponding to the filled circles on Fig.4 below.Sxi}0<t<x2Xl .1 Note that no specific sample path structure of {St} (like in Fig. see [47]. 4.F*(X) = P0(Si > x) . (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. such that {SXo+t  SXo}0<t< X 1Xo . G(x) (4.4. Figure 4. See Fig. < 0 and EoX < oo where X = X2 .. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. Schmidli & Schmidt [47]. The idea is now to observe that in the zerodelayed case.. examples and counterexamples.1.. M. 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 <. Define S. For further approaches. We return to this point in Example 4..d. 0o(u) etc. and the distribution of {Sxk+t .Sxk}o<t<xk+1xk is the same for all k = 1. t>0 S. Thus the assumption . . = Sx. and apply it to the Markovmodulated model of Chapter VI...1) is assumed... {Sn}n=o.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen..X2 < .1 except for the first one) is a random walk. We let F* denote the Podistribution of Si.. +1.1 based upon a regenerative assumption. 4.1 where the filled circles symbolize a regeneration in the path...1 = max k=0... 2..i.1..n n=0.
3) hold.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.3) where Mnx) = sup o<t<xn +1 X.S.2 Theorem 4. (4.1) and (4.. (4. the assumption means that Mix) and Sl are not too far away.266 CHAPTER IX. it suffices by (4. N N Xi=0 N Figure 4. The one we focus on is Fo (Mix) > x) . Then '00 (u) = Fo(M > u) . Fo(Si > X). Proof Since M > M*.Sxn = sup Sxn+t . 4.1 Assume that (4.* i o<t<xn+1x. See Fig.2. (4..3) applicable so that F(M* > u) 141 F*(u). jF11 F* (U). HEAVY TAILS for some G such that both G E S and Go E S makes (3. Sxn +t .. u p 00.2) to show F(M* > u) > 1. Since clearly M(x) > Sl .4) liminf u>oo F(M > u) ..
Let a > 0 be fixed. To this end.a. E (u .e)Po (MMX> > x).+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). Given e > 0. x > a. u)} < P(M* E (u . .1 = limti00 St/t. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. )) > (1 . Mn+l > a V (u .. M^xu)+l > a) . 2.(u) . S.( u)1 > a) 00 1: Po(Mn<u. MW O(u)+1 < a) IN ( U n=1 A1. Po(M* > u) .Sn 0<t<x„+j ( 1 . Under suitable conditions . Theorem 4. u))/P(M* = 0) = o(Po(M* > u)).(1 .e)Po (M > U).5) which follows since Po (M > u...4.a.: Sn > u} ..6) 1 p pBo(u) u where B is the Palm distribution of claims and p .4).Po (M* > u..: S.Mn +1 >aV(u n=1 00 S. assume the path structure Nt St = EUit+Zt i=1 .e)Po (M > u.2. /3(u) = inf{n=1.4.E) Po ( n max St u. Letting first u + oo and next e ..1 can be rewritten as 00 (U) (4. 0 yields (4.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. choose a such that Po(Si > x ) > (1 . Then by Lemma 3.S. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 ..
and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . (iii) For some o field Y.3PB. Assume further that (i) both B and Bo are subexponential.'s order EoNx • B(x). a4' 0. Corollary 4.1 is in force. the proof of Lemma 4.6) holds with p = . since the tail of Zx is lighter than B(x) by (iv).v.268 with {Zt} continuous. Proof It is easily seen that the r.p) Ju P Bo(u) 1p 0 .8) x Write . i=1 (4. X and N.Q = EoNx/EoX. cf.6 below.7). and also for Mix) since Nx FNX U. Mix) < > UE + i=1 o<t<x Thus Theorem 4.2 Assume that {St} is regenerative and satisfies (4. The same is true for Sl. we get 00 (u) 1 IPF.4). Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . independent of {> CHAPTER IX.X both have tails of sup Zt.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . oX (see Proposition A1.} and satisfying Zt/t N. (ii) EozNX < oo for some z > 1. HEAVY TAILS N` U. and ENX Ui . are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.
We consider the case where one or more of the claim size distributions Bi are heavytailed. and for some constants ci < oo such that cl + • • • + c. MODELS WITH DEPENDENT INPUT 269 Example 4 . X3 = 2. Then (4.. consider the periodic model of VI. Again .. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. Assume that B E S. .4 Assume that St = Zt .. < 1. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.6) u holds.t + EN'I Ui where {>N`1 Ui .5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.4. More precisely. The number N.. and taking F = o. X2 = 1.3 As a first quick application. we conclude just as in Example 4. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. Taking again Xo = Xi = 0. we assume that B E S.0 (thus (iv) is trivial). The regenerative assumption is satisfied if we take Xo = Xi = 0.6) holds. . Zt .9). Theorem 4. (i) holds. . Example 4 . 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. i.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive). In particular. The arrival rate is /3i and the claim size distribution Bi when Jt = i.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. .(NX). of claims arriving in [0. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. > 0. 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 .. X3 = 2.e. in particular lighttailed. (iii) is obvious. Bo E S. The key step of the proof is the following lemma. Bo E S.3 that (4. Thus we conclude that (4.6) holds. X2 = 1.
5. An easy conditioning argument then yields the result when Jo is u random... .c'(x) where c = ciENi . X > 0 a r. Assume EzN1+"'+Np < oo for some z > 1 and all i. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.X i=1 j=1 where conditionally upon F the Xi. NP ) and X are . and that for some + cp distribution G on [0. .^•) G(x) P ^ E ciNi = C. 1. . The same dominated convergence argument completes the proof. Let {Fi}t=1 P be a family of distributions on [0. and the rest of the argument is then just as the proof of Corollary 4. i=1 Proof Consider first the case X = 0.. oo) and define p Ni Yx = EEX'i . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. NP ) be a random vector in {0. .ciG(x). 2 . u Proof of Theorem 4. ... Then P P(Yx > x) .F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . . . 6 Let (N1. are independent with distribution Fi for Xij.270 CHAPTER IX.Fmeasurable. Thus dominated convergence yields ( P(Yo>x P(Yo>x . Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .. If Jo = i.v. cp with cl + > 0 it holds that Fi(x) .}P. For lighttailed distributions. HEAVY TAILS Lemma 4 .G( x ) > ciNi . and F a aalgebra such that (N1. i1 = E\ G(x) In the general case.. as x a oo.F) < CG(x)zn'1+."+Np .2. we can define the regenerations points as the times of returns to i. oo) such that G E S and some c1. P P P(YX and > x I. i=1 P(Yx > x ^) < P(Y0 > x I..
6) to hold in a situation where the interclaim times (T1. ).2 and Example 4. An improvement was given in Asmussen & Hojgaard [33]. Theorem 2. 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). The main result of this section. we let PN"N = P(. Schmidli & Schmidt [47]. m is a (orfinite) ..7).T2. as well as a condition for (4.i. r(u) is the time of ruin and as in IV.7. ) form a general stationary sequence and the U.4. Combined with the approximation for O(u). IV.5. For further studies of perturbations like in Corollary 4. the discussion provides an alternative point of view to some results in Chapter IV. and the final reduction by Jelenkovic & Lazar [213]. Within the class of risk processes in a Markovian environment. this should be compared with the normal limit for the lighttailed case. I T(u) < oo). As usual. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case.. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential... Theorem 4.T2. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. cf. this is applied for example to risk processes with Poisson cluster arrivals.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.1.. In contrast. Essentially.pl(1 .1 is from Asmussen.5 shows that basically only the tail dominant claim size distributions (those with c.4. 5 was first proved by Asmussen.e.4. see Schlegel [316]. We start by reviewing some general facts which are fundamental for the analysis..d. It follows from Theorem 4.p)Bo(u). Theorem 5. in particular Proposition 2. cf. and independent of (T1.4. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. VI.3. Then O(u) . 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. Notes and references Theorem 4. states that under mild additional conditions. The present approach via Theorem 4. cf. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). i. i.
t. oo). y to vary in.t. y = 0). m. but the example of relevance for us is the following: Proposition 5. Rt is distributed as x + t .1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . We let QS be the corresponding distribution and Qx.= y.00). 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} . Thus. the whole of R and not as usual impose the restrictions x > 0.2) means ffh(a. {Rt}.s. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. in the terminology of general Markov process theory. (note that we allow x.)k(x . u For F C E.s=j are the transition probabilities for {St}. Say {St} is reflected Brownian motion on [0. Lebesgue measure.y = Qx (. Sw(F. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. where we can take h.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). to the r. and (5. The equality of the l. Let G denote the distribution of ENt U. k on E. x = 0+ and F = (0. For the present purposes it suffices . follows by the substitution y = x .h. t. j.). say. and starting from So = y.s. .rij = mjsji where r13. {St} and {Rt} are in classical duality w. for states i.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. to consider only the case Px(w(F`) = 0) 0. Proof Starting from Ro = x.272 CHAPTER IX. St is distributed as y . .t + EI U. k as indicator functions.z. a familiar case is time reversion (here m is the stationary distribution). r.2) with t = 1 means m. however . resp. The simplest example is a discrete time discrete state space chain.h.r. w(Fc) < oo ) 'In general Markov process theory. Then (5.>N` Ui.
1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. 5. oo) = r(0) x= St y (a) Figure 5.3 The distribution of r(0) given r(0) < oo. The theorem is illustrated in Fig . Sn = in = y. Sw(F)1 = y) . 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). 0]. S.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. QR and QRy are defined similarly.= y) Theorem 5 . the one in (b) is the time reversed path. Sw(Fo)_ should be interpreted as Sw(F^)_1).5. z > 0.s. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. /^s x (S1 = Z1. . Thus.y = Qy Q. w(0.13AB < 1] Proof of Theorem 5.1 for the case F = (oo. in = y. But in the risk theory example (corresponding to which the sample paths are drawn). io = x... 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...SS(F.itt) = P Px(w(Fc) < 00.. . That is.y(2p21 . We can then view Qy.2 Qy. [note that w(z) < oo a.. Qx. . and we let Qy y refer to the time reversed excursion .).. Sn+1 E Fc) nx. .y as a measure on all strings of the form i0i1 .y() = P ({SW(F`)t} 0<t<w(F °) E So = x. in E F. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. in with i0.. x = 0.2. when p = . In particular: Corollary 5. i1. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit.
S... 21 .ik_1EF Sxin_1 .... Si1y 00 jEF° E E 5xik_ 1 ...... . Silt' E SO k=1 i1..TI( 2n2n _1 ...J (i. Rn+1 E FC) TioilTili2. Si l io E mjSjx. note first that Pt' (R l = il..... Si11 S 1 .. in = x. HEAVY TAILS E E Px (Si = 21i . R Qy x(2p21 ..274 note that Fx(w(Fc) < 00...... 2p). in) = Qx. MY Thus Qx(ioii . in = x.. i0) Q x..... 2n) = Qx.rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .(F<)1 = Y) S S and Qx y( ipil . ... S....i„_iEF Similarly.in1 .. (Fc)1 = y) 00 CHAPTER IX. Rn = in = x. R .ik_1EF .. ...in E F. .. in) = oo jEF^ Sxin1 ....ik1EF Similarly but easier Sxin_1 . Si1y k=1 i1 .. To show Q y x (i0 i 1 .. Sn+1 E Fc) n=1 i1.. Silt' E Sxik_1 . i0 = y. t' y and Qy x are measures on all strings of the form ipi l ... in E F..ii . = in = y.. .. . 20 = y..y(inin _ E SYj jEF` 00 Sxik _1 . Rn+1 E F`) F (w(Fc) < 00. ..gilt' k=1 ii . in)  Pt' (R1 = ii.. .. .. 2p) when 20... in with 20... Rn = in = x.
Now the P(u. Z) is described in Theorem 111.')density of Y is B(y)/[.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . 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 .2. that is.t. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.t. that is. Y > u).r.p. the distribution w. P(") = P(. Y > y} . To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment .'s are defined w. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r. 5.2 The distribution of (Y. Bo") is also the P(u.r. the case r (O) < oo.2. P(o) ).UBBo(u)].B(a) +a PBBo(u) . We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. the P(u.')distribution of Yu is Bo"). see Fig.5. That is. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . ST(o) > y. 1 w . S. y > u.2. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y).v. 7(0) < oo. U T(O) = T (u) Y Figure 5.(o) > y} = {T(0) < oo.
1/(1 ... we get the same asymptotics as when n = 1. are i. and YI... 2. K(u) = n). . z ^ oo. in particular of Z. Z = ZI but relative to the kth ladder segment. 5. Then Corollary 5. . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. Zn).. It is straightforward that under the conditions of Proposition 1. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.. i. Since the conditional distribution of Z is known (viz. HEAVY TAILS Let {w(z)}Z^. conditionally upon r+ (n) < oo. must be large and Z1. Then 7(u)/y(u) ^ W/(1 .T+(2).. r(u)/Z 4 1/(1 . it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . Z).p) in Pi"'')distribution.. .r+ (n .p). . Yn_1 'typical'.. Zk be defined similarly as Y = Y1.. The idea is now to observe that if K(u) = n. Then. a slight rewriting may be more appealing.i.1. Since w(z)/z a$. the duration T+ (n) .e. i. > u with high probability. P(Z < a I Y > u) 3 0.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1..p) then yields the final result T(u)/y(u) + W/(1 . denote the ladder epochs and let Yk. . let r+(1) = T(0). this in principle determines the asymptotic behaviour of r(u). That is .p). (Y. Hence Z.. Bo") ).18(c) Bo")(yY (u)) + P(W > y) ( 5. .3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). .e. However. and since its dominates the first n .p) in F(u) distribution. Recall the definition of the auxiliary function y(x) in Section 1. Z1).276 CHAPTER IX. and distributed as (Y. then by the subexponential property Yn must be large. the random vectors (YI. Fig.3.d.. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. more precisely. We let K(u) = inf In = 1. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .3) holds.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). cf. In the proof.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. 4 Assume that Bo E S and that (5... Z/'y(u) * W in Pi "' ')distribution .: r+ (n) < oo.
A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).u) E •) . > u}.5... Proof We shall use the easily proved fact that if A'(u). Yn ... P(. ... Further. .u) E • I A'(u)) = Bo (n1) ®Bou) .u) II 0.n).2.Bo (ri1) ®B( . II ' II denotes the total variation norm between probability measures and ® product measure. then IIP( I A'(u)) Taking A'(u) = {Y. +Yn1<u. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. A"(u) _ {K(u)=n} = {Y1+ P(. Lemma 5.Yl+ +Yf1>u}. I A"(u ))II + 0.Yn1iYn .. . the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. . P (Yj.n) (y1. Y„1.3 In the following. suitably adapted).5 Ilp(u. I A'(u)) = P(u.
1).u has distribution Bout That is. y > u.. HEAVY TAILS ((Z1'. . and that Yk has marginal distribution B0 for k = 1.y(u)T) . k = 1.+y 1 p"F(Yn > u) P)Pn1 P/(1 .. Zn). Zn are independent. .1 and Y„ .t..1. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. see Fitzsimmons [144]).P(Z' E •)II > 0 (here Y. then 11P(Z E •) . n_1 < u. The first step is to observe that K(u) has a proper limit distribution w. Y") u etc. . Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. in our example Y = (Y1. Let {wl(z)}.. the discussion just before the statement of Theorem 5.i.P(Y' E •)II * 0..P) Bo(u) for n = 1.. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). Y'. Y1 +.P) > y) Corollary 5.p) < y). Then according to Section 5a.. By Lemma 5.. Proof Let (Y11..6 IIPIu'n ) CHAPTER IX. the density of Yn is B(y)/[IBBO(u)]. .. .. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .... the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). n .d..4). 2. wk(Zk) has a proper limit distribution as u + oo for k < n. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.4...6. For Theorem 5.).. . . P(u) since by Theorem 2.... Z11). n.2. (Y.1 P PBo(u) • P(W/(1 .. .. {wn(z)} be i. 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 •) . Zn) E •) .. Z' are arbitrary random vectors.' = y is BM.. and clearly Zi. be independent random vectors such that the conditional distribution of Zk given Y. . in particular his Proposition (2.Bo (n1) ®Bo' 0. . Z. . It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.. Similarly (replace u by 0)..7 O (u.r. copies of {w(z)}. Proof of Theorem 5.. the marginal distribution of Zk is Bo for k < n.278 Lemma 5 . .
however. More precisely. T) when T + oo with u fixed. Theorem 6 .2 Define M. and premium rate p(x) at level x of the reserve.B(u). Corollary II. claim size distribution B. one expects the level y form which the big jump occurs to be 0(1).2. i. p(Y) and the result follows. that MQ becomes large as consequence of one big jump. . Assume for simplicity that {Vt} regenerates in state 0 .e.1 Assume that B is subexponential and that p(x) > 00./3Ea B(u).1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. the results only cover the regularly varying case. Proof of Theorem 6. the probability that is exceeds u is then B(u . Then P(MT > u) .1.. V. The rigorous proof is. u (6.6. = supo<t<0. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . x > oo. 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). Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33].(3 u u J B(y) dy . nontrivial and we refer to Asmussen [22]. 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 . that fo p(x)1 dx < oo. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.y) . Asmussen & Teugels [53] studied approximations of i (u. The heuristic motivation is the usual in the heavytailed area. 3. Then 0 (u) Qf "O ^) dy.(u) = P(V > u) = f f (y) dy . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. cf. and define the cycle as a = inf{t>0: Vt=0.
280 CHAPTER IX. D(u) = DQ(u)/µ. Hence f (u)r(u) = D(u) = Do(u) . u Notes and references The results are from Asmussen [22]. by regenerative process theory. Then D(u) = f(u)p(u) and.P(MT > u) $B(u) Ft µ(1 . where also the (easier) case of p(x) having a finite limit is treated . there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.q(u) Now just use that p(x) * oo implies q (x) + 0. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. . It is also shown in that paper that typically.q ( u)) 1 . 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).
The crude Monte Carlo ( CMC) method then amounts to simulating i. .z) 4 N(0. z) 2 = Zit NE ii ii According to standard central limit theory ... and this is the form in which the result of the simulation experiment is commonly reported. vrN(z . Hence 1. Ripley [304].d. 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.. 4Z).i. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. ZN. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. Rubinstein [310] or Rubinstein & Melamed [311] for more detail .Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . where a2 = Var(Z ).2) is an asymptotic 95% confidence interval .96s z f (1. topics of direct relevance for the study of ruin probabilities are treated in more depth. Fox & Schrage [77]. 281 . replicates Zl.
We survey two methods which are used below to study ruin probabilities. Letting Z' = E[Z I Y]. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. one can argue that unless Var(Z') is considerable smaller than Var(Z). Further. v. and many sophisticated ideas have been developed. and a longer CPU time to produce one replication. However. generated at the same time as Z. an added programming effort. T): just simulate the risk process {Rt} up to time T (or T n 7(u)) and let Z be the indicator that ruin has occurred. we then have EZ = EZ = z. and in most cases this modest increase of N is totally unproblematic. This is a classical area of the simulation literature. We mention in particular ( regression adjusted) control variates and common random numbers.b(u. conditional Monte Carlo and importance sampling. Therefore.282 CHAPTER X. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). variance reduction is hardly worthwhile. 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'. Say that Var(Z') = Var(Z)/2. The situation is more intricate for the infinite horizon ruin probability 0(u). so that Z' is a candidate for a Monte Carlo estimator of z. SIMULATION METHODOLOGY In the setting of ruin probabilities. 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. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) .
z2 = 0. (1.3). GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. the argument cheats because we are simulating since z is not avaliable analytically. Importance sampling The idea is to compute z = EZ by simulating from a probability measure P different from the given probability measure F and having the property that there exists a r. . a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. This may also be difficult to assess . L such that z = EZ = E[LZ]. even if the optimal change of measure is not practical. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. In order to achieve (1. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. Then z Var(LZ) = E(LZ)2 . Nevertheless.3) Thus.v.. and the problem is to make an efficient choice.[E(LZ)] = E Z2 Zz . To this end.e. Thus.E [Z Z]2 = z2 . it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.96 sis v^ N 2 1 where srs = N j(LiZi .zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi .1. . it appears that we have produced an estimator with variance zero. However. . L1). i. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).. Thus we cannot compute L = Z/z (further. but tentatively. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. (ZN. one would try to choose P to make large values of Z more likely. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). using the CMC method one generates (Z1.zrs.
N . For each u. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . let z(u) = P(A(u)).0.B = iP(AB) = P(BIA).284 CHAPTER X.e. u + oo. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small .100 .96oz /(zV) = 0. and further it is usually not practicable to simulate from P(•IA). However. However. The CMC method leads to a variance of oZ = z(1 . We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. We then . assume that the A(u) are rare in the sense that z(u) * 0. in terms of the halfwidth of the confidence interval.z) 1 > 00. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .z) 1001. say 10%. it does not help telling whether z is of the magnitude 104.5 or even much smaller . A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. 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. if z is small. we may try to make P look as much like P(•IA) as possible.. but if the point estimate z is of the order 105.96 2 z2 z increases like z1 as z .z) which tends to zero as z ^ 0. In ruin probability theory. Z z V5 In other words .1. i.96 2Z ( 1 .e. Thus. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. This leads to the equation 1. An example where this works out nicely is given in Section 3.e. a confidence interval of width 10 4 may look small. z I. 10 . say of the order 103 or less. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). Again. the optimal P is the conditional distribution given A. I. as is the case of typical interest.1. and let Z(u) be a Monte Carlo estimator of z(u). To introduce these.. Z = I(A) and A is a rare event. large sample sizes are required.
i.e.4). but as a CMC method . where Z = I(M > u) may be generated as follows: 1.p)pk. XK from the density bo(x).. Thus.4) for any e > 0. . The term logarithmic comes from the equivalent form . 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. it is appealing to combine with some variance reduction method . According to the above discussion. O (u) = z = EZ.2. and in practice.log z(u) of (1. with common density bo(x) = B(x)/µB and K is geometric with parameter p. This allows Var(Z(u)) to decrease slightly slower than z(u)2. where X1. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. The algorithm gives a solution to the infinite horizon problem . i. where M = X1 + • • • + XK. Let M . which gives a logarithmically efficient estimator .p)pk. are i. F(K = k) = (1 . Generate K as geometric.X1 + + XK. Otherwise.. 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.. SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo. it is not efficient for large u . let Z +.0. If M > u. the mathematical definition puts certain restrictions on this growth rate. 3. see Asmussen & Rubinstein [45] and Heidelberger [190]. .. Generate X1. .1) may be written as V) (u) = P(M > u). let Z +. X2. logarithmic efficiency is almost as good as bounded relative error. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. However. P(K = k) = (1 . the PollaczeckKhinchine formula III. 2.log Var(Z(u)) lim inf > 2 u+oo .d. Therefore . so that NE (u) may go to infinity.1. Notes and references For surveys on rare events simulation.(2.
XK_1)...Xl ... XK. just note that EZ(1)(u ) 2 > E[Bo (x . For the simulation. Z(1) (u) has a smaller variance than Zl (x)..p)Bo(x)..b(u) = P (Xl +•••+XK>u) = EF[Xl + .1) V)(u) . form the order statistics X(1) < X(2) < .. XK1.. The idea of [27] is to avoid this problem by discarding the largest X. and considering only the remaining ones. .SK1)2... X1 + + XK_ 1 > x when X1 > x. conditional probability. Thus. As a conditional Monte Carlo estimator .. Theorem IX.2..... K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. Z(1)(u) is defined as 0). < X(K) throw away the largest one X(K).X(2). y < 0). .p/(l . we generate only X1. assume in the following that Bo(x) .XK1] = EBo(uX1 .X1 ..X(K1)) . Xl > x.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. and the problem is to produce an estimator Z(u) with a variance going to zero not slower (in the logarithmic sense ) than Bo(u)2.L(x)/x`' with a > 0 and L(x) slowly varying. However. and let Z(2)(u) = _ P (SK B0((u > u I X(l).S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1).. . asymptotically it presents no improvement : the variance is of the same order of magnitude F(x)..X(2).. So.286 CHAPTER X...... note first that To check the formula for the P(X(n) > x I X(1).. we thus generate K and X1i . 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. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .. Then (cf.. A first obvious idea is to use conditional Monte Carlo: write i. compute Y = u . To see this.. .+XK > uIXl. and that Bo(y) = 1.
using 13L. . X(n1)) Bo((x . and simulate from FL. X(2). use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . and define )3L.Khinchine formula and importance sampling . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). X .S(n_1) I X(1). Then the algorithm given by { Z (2) (u) } is logarithmically efficient. Notes and references The proof of Theorem 2.u is the representation 0(u) = e7sr(u) overshoot (cf. X(2). 111. X(2). 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying.S (n1)) V X (. BL by I3L = /3B[y].y. 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. that is. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". . Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . X(n1)) P(X(TZ) + S(. .modulated model P(r+ < oo) and G+ are not explicit ).1) . Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] ..1 is elementary but lengty.3. .. BL(dx) = e7sB(dx)/B[y]._1) > P(X(n) > _ X X(1). For practical purposes. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck.... However . B. X (. l)) BO(X(n1)) Theorem 2 . The algorithm is sofar the only efficient one which has been developed for the heavytailed case. BL instead of 0. l)) . Thus. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. for the purpose of recording Z(u) = erysr(u). the algorithm for generating Z = Z(u) is: 1. in the renewal or Markov.. Asmussen .5). . . and we refer to [27]. 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.
. = X1 + .i.2 The estimator (3. . b different from . b) # (/3L.3.7 tell that P(.. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.1) (simulated with parameters ^3. namely ELery£("). Let X1. If S > u. + X. Let Sf0 CHAPTER X.Q. In fact: Theorem 3...F. 4.. to deal with the infinite horizon problem . cf. SIMULATION METHODOLOGY 3. u It is tempting to ask whether choosing importance sampling parameters . Let S . More precisely. and we have: Theorem 3.288 2. P'[y] < oo for some ry > 0. let Z F e_'s. the discussion at the end of Section 1b. The proof is given below as a corollary to Theorem 3. BL). X2. the results of IV. so that changing the measure to FL is close to the optimal scheme for importance sampling .1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. The algorithm generalizes easily to the renewal model . with distribution F. Proof Just note that EZ(u)2 < e . It resolves the infinite horizon problem since FL(.S+U . r(u) < oo) and FL (both measures restricted to.. and assume that µF < 0 and that F[y] = 1. return to 3.r(u) < oo) = 1. Let FL (dx) = 'For the renewal model.QL. and avoid simulating the known part e7". BL could improve the variance of the estimator .2ryu _ z (u)2/C2. M(u) = inf {n : S„ > u}. Generate T as being exponential with parameter . Xi = 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..l3 and U from B.T.. B) is not logarithmically efficient when (/3. Otherwise. In detail . There are various intuitive reasons that this should be a good algorithm. and the change of measure F r FL corresponds to B > BL. let S. be i. We formulate this in a slightly more general random walk setting '. Ti. one must restrict attention to the case 4µB > 1. The answer is no.(u)) are asymptotically coincide on {r(u)} < oo.d. A > AL as in Chapter V.
3 The estimator (3. (3... .2ryELXi. + KM(u))} = exp {ELM(u)(E ..yu+elu u +oo etry' 1 > lim up C2e2. The importance sampling estimator is then Z( u) = e'rSM( ). it thus follows that for 0 < e < e'/ELXi. is not logarithmically efficient. are i.. More generally.i. K2.3. Since K1.+KM(u)}. By the chain rule for RadonNikodym derivatives.P = FL. . When F # FL. For the second.2 > 0. = c'. where e' = EL Iog dFL (Xi) > 0 by the information inequality. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.. write W(F IF) _ F(XI)..2) dF Theorem 3.2ryELXi)} . Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . F(XM(u)).yu = G.d.2'X1 . let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.2) (simulated with distribution F of the X3 has bounded relative error when .. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). Here ELK.. Proof The first statement is proved exactly as Theorem 3 . 1. where Kl og (X) (j) 2 ) = log dFL (Xi) .. Since ELM(u)/u + 1/ELXi.
T' has a left exponential tail with rate /3' and U" . CHAPTER X.1 If y > 1/ic'('y). T) with T < oo.g. In [13].3. 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 . As in IV. The queueing literature on related algorithms is extensive . U". BL) has bounded relative error. . The optimality result Theorem 3. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16].4 indicate that we can expect a major difference according to whether y < 1/r.e.'(y) or y > 1/r. The results of IV. Next. U' . optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.290 which completes the proof. /3'eQ'YR'( x + y) dy = . In fact: Proposition 4.T' D U" .T' = U" . then /3' B' = B". Consider compound Poisson risk process with intensities /3'. so that one would expect the change of measure F 4 FL to produce close to optimal results.T".i. we conclude by differentiation that Bo(x)=B' (x)forallx > 0. generic interarrival times T' .T". T".4. Further discussion is in Lehtonen & Nyrhinen [245]. claim size distributions B'. Then according to Theorem 3. yu) is close to zk(u).T" has a left exponential tail with rate /3'.2. /3". SIMULATION METHODOLOGY u Proof of Theorem 3.'(y). we write T = yu. the references in Asmussen & Rubinstein [45] and Heidelberger [190]. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. The easy case is y > 1/k'(y) where O(u. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz.B'=B".3'eO'x f f P (U" . see e.T" > x) J /3"e0 yB (x + y) dy = . First by the memoryless distribution of the exponential distribution .1 is from Lehtonen & Nyrhinen [244].3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". This immediately yields / = 3". B" and generic claim sizes U'. U' . all that needs to be shown is that if U' . then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].
8 has a stronger conclusion than (4.(u) * 1 (Theorem IV. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. . and in fact. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).1. T(u) < yu] e Hence by (4. Further .log Var(Z(u)) l im of .4. 3 Theorem IV.1).yu.to g x ( u ) u u so that (1.'(7).2 The estimator (4.log 4')u) 4 u (Theorem IV.yk(ay) determines the order of magnitude of z'(u.4. yu) = eayu Eay Leay^(u)+r(u)K(ay). one would expect that the change of measure F Pay is in some sense optimal. T(u) < yu] . T( u) < yu] e2ryyuEay le 2ay^(u). Let Qy2 = .yy> 2 .2) Since the definition of ay is equivalent to Eay r(u) . the result follows as in the proof of Theorem 3.5) follows. yu) in the sense that .3) (simulated with parameters /gay. Remark 4 .(ay).O(u. Bay) is logarithmically efficient.3) and we have: Theorem 4. 7y (4. lim inf uoo 27yu . (4. yu)/z. We next consider the case y < 1/r. We recall that ay is defined as the solution of a'(a) = 1/y. (4. yu) is of order of magnitude a71.4.1). we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. Proof Since ryy > y. Bounding u ELZ(u)2 above by a7u. and that ryy > ry.4.1) which is all that is needed here can be showed much easier .log Var(Z(u)) _ . that ryy = ay .1) so that z(u) = zP(u.8).
4.1) (see Proposition IV.(av)Eav l e.o.b(u.292 CHAPTER X.2).7y x(u) > hm inf u+Oo U .2) .1) is used to study Voo by simulating {Rt} (for example. yu .. zi(u) = INV. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .4. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). N(0.yu)/(uyu1/2) .1) may be useful.uaoo U That lim sup < follows similarly but easier as when estimating En. '%(u) = P I info Rt < 0) = P(VV > u).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). Z (u)2 above. we believe that there are examples also in risk theory where (5.3: for many risk processes {Rt }. However. and (5. Hence lira inf log ryyu + vyu 1/2 tc(ay) . In Asmussen [13]. yu . > u) = E f I(VV > u) dt 0 (5. related discussion is given in a heavy traffic limit q J.3.4).u1/2 < r(u) < yu Le ] l = e.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . the object of interest is {Vt} rather than {Rt}.1) where the identity for Vi(u) requires that Vt has a limit in distribution V. 0 Notes and references The results of the present section are new. (5.1): then by Proposition A1. there exists a dual process { V t} such that i. 0 rather than when u 3 oo.o . In most of the simulation literature (say in queueing applications). One main example is {Vt} being regenerative (see A.a vt(u).T) = P O<t<T inf Rt < 0 = P(VT > u). 5 Regenerative simulation Our starting point is the duality representations in 11.Qyu1/2 < T(u) C yu e.
(u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. EZ1'i = z1 = Ew. Thus.. z2)) > N(O.E).d.. i. For details . + Z2N)) . REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Then Z(1). Therefore . For the ith cycle.5. j = 1. which we survey below .3) .. . To derive confidence intervals . EZ2'i = z2 = E Thus. . Z(N) are i . Z2 . let E denote the 2 x 2 covariance matrix of Z('). Then (Z1z1i Z2z2 ) 4 N2(0. Z1 = (Zl1i +.t(u)) 4 N(0. The method of regenerative simulation. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. Z2 a4* z2. Zl the LLN yields Z1 a$' Z(1) +. 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'). the regenerative estimator z%(u) is consistent. 02) (5. consider first the case of independent cycles . a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi.h (zl. Vh = (8h/8z1 8h/ 8z2). Taking h(zl.. + Z1N>) . 2. oh) for h : R2 ^ R and Ch = VhEVh.... letting J0 'o I (Vt > u) dt ..+Z(N) z 1. > u) (and more general expectations Eg(V. z2) z2/z1 yields Vh = (z2/z2 1/zl). record Zi'i = (Z1'). and Z2'>) where Zi'i = w. ))... Z2 = N (X21' + . is the cycle length.. provides estimates for F ( V. i (^(u) .
() dx f Ax) (dl d()f (x' () f ( z. However .g. see e.Z) ^Z(=) . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . Then z(() = f cp(x) f (x.v. Let X have a density f (x.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . the expectation z = EZ of a single r. in some situations it may be the only one resolving the infinite horizon problem . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . Z of the form Z = ^p(X) where X is a r . SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . () dx = E[SZ] f(X. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. Notes and references The literature on regenerative simulation is extensive.96s/v"N.294 where 01 2 CHAPTER X. 6 Sensitivity analysis We return to the problem of 111 . asymptotic estimates were derived using the renewal equation for z /i(u). There is potential also for combining with some variance reduction method. The regenerative method is not likely to be efficient for large u but rather a brute force one. () dx so that differentiation yields zS d( fco(x)f(x. Before going into the complications of ruin probabilities . v.C)dx = f w(x) d( f ( x. consider an extremely simple example . Rubinstein [310] and Rubinstein & Melamed [311].5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. In 111. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims .2. 9. with distribution depending on a parameter (.0 . () depending on C.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .g S12 (5.9.
one. SZ is an unbiased Monte Carlo estimator of z(. Example 6 . ()). A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w.t. Thus. /3o is M(u) Oe (3T: < oo) . ()) d( hc(U. Then . if f (x. say W(x) = I(x > xo) and assume that h(U. one can take h (U. () = . cp' (h(U. () where U is uniform(0. () can be generated as h(U. () is an unbiased Monte Carlo estimator of zS. however . with density f (x. So assume that a r. ()) h((U. () _ (eSx. For the SF method. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. cp(h(U. C) f(X. p. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. () = (8/8()h (u. C)). = E [`d (h(U. /3 is 0. this is usually unproblematic and involves some application of dominated convergence . In the setting of ruin probabilities . () = d log f (X.r. The likelihood ratio up to r(u) for two Poisson processes with rates /3. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. ( where h( (u. 11 /3oeOoT. giving h( (U.t. () = log U/(2.1). IPA will estimate zS by 0 which is obviously not correct.v. just take cp as an indicator function .() d( is the score function familiar from statistics . this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . To see this. () Thus. () is increasing in C.log U/(. for some Co = (o(U). For example .6. zc = E [d co(h(U. r(u) = Tl + • • • +TM(u)).1 Consider the sensitivity tka(u) w. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. Thus . Then z(() = Ecp(h(U. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. The following example demonstrates how the SF method handles this situation. Let M(u) be the number of claims up to the time r(u) of ruin (thus. C). ()) is 0 w .r. For IPA there are. the Poisson rate /3 in the compound Poisson model. I(r(u) .
T(u)) I(T(u) < co) ] .r. There have been much work on resolving the difficulties associated with IPA pointed out above. j3 and letting flo = 0. whereas for the SF method we refer to Rubinstein & Shapiro [312].296 CHAPTER X. Thus. for different models and for the sensitivities w. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). SIMULATION METHODOLOGY Taking expectation.r. a relevant reference is VazquezAbad [374]. change the measure to FL as when simulating tp(u).1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .9 . in part for different measures of risk than ruin probabilities. However.0(1) so that in fact the estimator Zf(u) has bounded relative error.3L. Example 6. . BL). we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . 4) that V5. In the setting of ruin probabilities. To resolve the infinite horizon problem .t.T(u)) e7uerVu) for ?P.3 (u) is of the order of magnitude ue7u. We then arrive at the estimator ZZ(u) = (M(u) .t. differentiating w. different parameters.3 (u) (to generate Zp (u). the risk process should be simulated with parameters . the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. We recall (Proposition 111.
in the Bernoulli random walk example below. with P(Xk = 1) = 9. either this makes no difference (P(R. as e.i. are i. defined as Ro = u (with u E {0. Consider first a Bernoulli random walk.1.1}valued .+• • •+X.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other... 297 .. 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ). X2.d. in most cases .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. where X1. That is. The twobarrier ruin problem The twobarrier ruin probability 0..Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.. T+(a) = inf It > 0 : Rt > al. Y'a(U) = P(T (u) = r+(a)) = 1 . a) = r(u) A T+(a).P(•r(u. Besides its intrinsic interest .. T(u. wherel T(u) = inf {t > 0 : Rt < 0} .g... . Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.. a) = r(u)). }). and {1. R„ = u+X..
The Lundberg equation becomes 1=F[ry]=(19)+9z... and the other more advanced but applicable also in some other settings.(1B)u oJ 0. Proof 1.1) is solution.o)T/la (1) + 8z/'u(3).298 CHAPTER XI.. zu = EzRO = EzRT(u.+Xn) F[ a]n n=0.. u + 1..0)/0.. then 'Oa(u) _ au a We give two proofs .. = z°Va(u) + za(1  .1.e. i. C1_0\a. u Proof 2.o» = z°P (RT ( u.1) o If 0 = 1/ 2. the solution of F[.4) by ea(u+Xl+. where a is any number such that Ee°X = F[a] <oo.(u) I\ e = 1 oa ' ()i a = u.r(u.(4. Wald's exponential martingale is defined as in 11.. By optional stopping. one elementary but difficult to generalize to other models. = (1 .2) Oa(a .a) = 0) + zap ( R.a(u)).o)'t/1a(a . z and the solution is z = (1 .y] = 1. MISCELLANEOUS TOPICS Proposition 1.1) = (19)4/'0(a3)+9ba(a1). Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). 7/la(a . In a general random walk setting .1 For a Bernoulli random walk with 0 0 1/2. We choose a = ry where ry is the Lundberg exponent. and in view of the discrete nature of a Bernoulli random walk we write z = e7.a) Y. The martingale is then {zuzXl+•••+X„ } = {zR° }. (1.2).. and insertion shows that ( 1. tba(2) _ (1 .
{Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. {R. BROWNIAN MOTION.ba(u) = e2µa . i1(u) = e211 . However. RANDOM WALK.u)/u. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .1 yields 't/la(u) = (a . . then Vi(u) = 1. 1h (u) = a el u \1 If 9 < 1/ 2.• a2µa e2µu . If 9 = 1/2.zu)/(za .1). Then for p 0 0. If p<0.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance .4 For a Brownian motion with drift u > 0.} 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)). then Proof Since 'Oa (U)  au a Eea(R°. and solving for 9/la(u) yields Z/)a(u) = (e 76 .5) . Corollary 1.2) is trivial (z = 1).1) for p # 0. (1.1 If p = 0.. pa( u) _ u Corollary 1. Applying optional stopping to the exponential martingale {e7R.1). Proof Let a+ oo in (1. (1. If p = 0. thenz1 (u)=1.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za .1.e7u)/(e7° .2 For a Bernoulli random walk with 9 > 1/2. u Proposition 1.0a(u)).
and thus one encounters the problem of controlling the undershoot under level 0. 0. this immediately yields (1.a ) < 0) + e 7aF ( R (u. VIII.7) .+^a(u))^(a) If 7k(a) < 1. (u) _ O(u) . 1 . (1.e7a Again . Here the undershoot under 0 is exponential with rate 5.e7a (u) = 6 /0 .0 (u) (where u p =.vi(a) Proof By the upwards skipfree property. Ic 5ry 'pa(u) Using y = 6 . . It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. 7/'(u) = 1).5a).a) = a on {r (u. say.5 Consider the compound Poisson model with exponential claims (with rate. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.a) = r+ (a)} and similarly for the boundary 0. passing to even more general cases the method quickly becomes unfeasible (see.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. However. valid if p < 1 (otherwise .616).a) < 0) + e7°P (R(u.300 Proof Let a * oo in (1. a) I R(u a ) < 0] P (R(u .3.7/la(u))..a) = a) = 5 y = P (R (u. and hence e7u = Ee7Ro E [e7R(. however. CHAPTER XI. the paths are upwards skipfree but not downwards. we obtain 'Oa a7u .4).a) = a ) + e ' ° ( 1 . letting a * oo yields the standard expression pe7u for . 7O(u) = 7/la(u) + (1 .7). 5). For most standard risk processes .0(a) 0 < u < a. implying R(u. Here is one more case where this is feasible: Example 1.
(i). Then the density and c. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. Corollary 1.)_ _( u)µ2 /2. + µ2T) } .9) = 2P(ST > u).8 Let {Rt} be Brownian motion with drift . P(MT > u) = P(ST > u) + P(ST < u.1a for computing ruin probabilities for a twostep premium function.d. TWO BARRIERS 301 Note thas this argument has already been used in VII.ST>U).10) Pµ (T(u) < T) !. MT > U) = P(ST > u) + P(ST > u) (1. we have ili(u.. BROWNIAN MOTION.Rt}.µ%T (1.T) P(MT > u) where MT = maxo<t<T St. (1.. = 1 .8) Proof In terms of the claim surplus process { St} = {u . Here {St } is Brownian motion with drift 0 (starting from 0). RANDOM WALK.. 0(u. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T . T(u) E dT. Hence P(MT>u.. and (1 . 10) follows then by straightforward differentiation.ST<u) = P(MT>u.1. T ) = P(T(u) < T ) = 241.2 .f.. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .11) VIT ) Proof For p = 0. For µ # 0. MT > u) = P (ST > u) + P (ST > u.µ T I + e2µ"4) ( .µ so that {St} is Brownian motion with drift µ .r(u). For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.8 ). and hence Pµ('r(u) E dT) = Eo [e µsr(. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . ( 1.7 For Brownian motion with drift 0. the density dPµ / dP0 of St is eµsttµ2/2.4) I = .11 ) is the same as (1. (1.
T+2. Breiman [78] or Karlin & Taylor [222] p.9 For Bernoulli random walk with 9 = 1/2. Let s(y) = ef0 ry(.10 Consider a diffusion process {Rt} on [0.11) then follows by checking that the derivative of the r. Thus. Vi(u.13) with 0 as lower limit of integration. Proof The argument leading to ( 1. is finite for all x > 0. oo) with drift µ(x) and variance a2 (x) at x..10) and that the value at 0 is 0. whenever u. If this assumption fails. e. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x).10). close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). The expression for F ( ST = v) is just a standard formula for the u binomial distribution.T)dx. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .T2.h.302 CHAPTER XI.9). is (1. u Small modifications also apply to Bernoulli random walks: Proposition 1.g. see e. The same argument as used for Corollary 1.. 226). T are integervalued and nonnegative. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. and in a similar spirit as in VII. Theorem 1.T (1.12) is the same as ( 1. MISCELLANEOUS TOPICS which is the same as (1.T) = P(ST = u) + 2P (ST > u). and (1. 0 0 (1. We finally consider a general diffusion {Rt} on [0.8 also applies to the case 9 54 1/2..g. S(x) = f x s(y)dy. Here {2T( (v}TT)/2) v=T. S(oo) = f c s(y)dy. (1.9) goes through unchanged. oo).3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.12) P(ST = v) = 0 otherwise. as defined in (1. but we omit the details.. as defined above as the probability of actually hitting 0. that 0(u). the behaviour at the boundary 0 is more complicated and it may happen.s.
and we get Wo. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.b = 0 implies that VQ b/s is constant. 191195 for material related to Theorem 1. see in particular pp.b(u)dt.13) is finite for all x > 0. see Asmussen & Perry [42].b(Rdt).1. 0 in (1.(u) < 1 for all u > 0 and ^ S^ Conversely. The obvious boundary conditions '0a. the function S(x) is .11 Let 0 < b < u < a and let t&0.10.10.16) yields 4b (u) = 1 .S(u)/S(a). then 0 < 2l. Wa. Lemma 1. Letting a T oo and considering the cases S(oo) = oo. RANDOM WALK. BROWNIAN MOTION.b(a) = 0 then yield the result.b = a+/3S. If (1.e LVa. Using s'/ s = 2p/a2.16) S(a) .S(u) (1. Notes and references All material of the present section is standard.14) S(oo) < 00. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .b (Rdt) = Oa.S(b) Proof Recall that under mild conditions on q. b = 0.0(u) = 1 for all u > 0. A classical reference for further aspects of Bernoulli random walks is Feller [142]. we can ignore the possibility of ruin or hitting the upper barrier a before dt. S(oo) < oo separately u completes the proof. Letting b J. E„ q(Rdt) = q(u)+Lq(u)dt. Then YIa. 1'. 15) i.b(u) be the probability that {Rt} hits b before a starting from u.e.ba.6 to Markovmodulated models . then. A good introduction to diffusions is in Karlin & Taylor [222]. Further references on twobarrier ruin problems include Dickson & Gray [116].14) fails. In view of (1. If b < u < a. so that Y)n. Assume further that S (x) as defined in (1. i.b('u) = Eu &0.17) Hence L. if (1.b(u) = S(a) .16). For generalizations of Proposition 1. O. TWO BARRIERS 303 for x > 0. (1. (1 . 0 Proof of Theorem 1. [117].b('u) = Eu .ba.b(b) = 1. .b(u) + L.
one works instead with a lower limit 5 > 0 of integration in (1.4.o•K(a) = Ee . Remark 11.5. Markovmodulated Brownian models . (2.1.6) . Lo is a martingale (cf.1) (2. correponding to piecewise linear paths or .9 ) and optional stopping applied to the stopping time r(u) A T.(.aR. 7y = ay . variance 0.3) < e 7yu.17).(a) (2.1 ) was given already in II.(7) . yu) where W (ay) = y. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.6. yielding eau = Ee. and here are alternative martingale proofs of the rest .t&(u. (2. 1 y < k (y). (2. IV..(T(u)AT) r.13)). yu) '+/1(u) . (2.2) C_e7u < t(u) < C+e _7u. y > . with the drift and the variance depending on an underlying Markov process . which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. much of the literature dels with the pure drift case. The emphasis is often on stationary distributions . where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. is currently an extremely active area of research.4) I. Lo I. information on ruin probabilities can be obtained . equivalently.5) A martingale proof of (2.3. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0.304 CHAPTER XI. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.4. Another basic quantity is the speed measure M .5): _ z/'(u) < e 7u.)AT . defined by the density 1/va(u)s(u) showing up in (1.ytc (ay). but by duality.aRo . 111 . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.2.
yu))• b(u. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. it follows easily from (2.3). dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt. y > r. dr) denote the conditional distribution of (T(u).4): We take a = ay in (2.4). so that i/1(uL yu) < eayu .B(r))/B(r).1. when Rt_ = r.f. eyuk (ay) = e7yu e > eyu"(ay ) ij(u.3). For (2.yu) Y Similarly for (2.6).E [e.6) with = 'y that eyu . Proof of ( 2. and the proof of the lower inequality is similar. .d.2): As noted in Proposition II.( u ) I T(U) < 00] . A claim leading to ruin at time t has c.6) below by 1 E Le7Rr(. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. Rt has distribution B(r + dy)/B(r). we have tc(ay) > 0 and we can bound (2.(u. (2. RT(u)_) given r(u) < oo.7R. u Proof of (2.1.)r(u)r. Let H(dt. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).1 .T)  V..2.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). (B(y) . dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. Equivalently. 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).T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.yuk (ay)(u&(u.
and gave sharp asymptotics for probabilities of the form P (S. e.306 CHAPTER XI.1). The last decades have seen a boom in the area and a considerable body of applications in queueing theory. nroo n n /// Note in particular that (3. Accordingly. the parameter will be u rather than n).(B) = log EeOX 1 is defined for sufficiently many 0.means (as at other places in the book) that the ratio is one in the limit (here n * oo). its generality. gn 4 0. such that the cumulant generating function r. cle .the correct sharp asymptotics might as well have +. Cramer considered a random walk Sn = X1 + .2). og For sequences fn. The limit result (3. The classical result in the area is Cramer's theorem. in being capable of treating many models beyond simple random walks which are not easily treated by other models . large deviations results have usually a weaker form. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . ri. .1) where we return to the values of 0. The advantage of the large deviations approach is..?n typically only give the dominant term in an asymptotic expression . Example 3.1) is an example of sharp asymptotics : .3na with a < 1. however . large deviations results been. if x > EX1.. However . 1) but only the dominant exponential term . (3.1) does not capture the \ in (3.nn or C2e. MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results.2) can be rewritten as F (Sn/n > x) 1g a'fin. (3. .1 We will go into some more detail concerning (3. Thus . and that a considerable body of theory has been developed. v2 later.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.g. we will write fn 1. then P C S. + X../n E I) for intervals I C R. logarithmic asymptotics .gn if nioo lim 109 fn = 1 log gn (later in this section. Thus.. not quite so much in insurance risk.^ e nn 1 > x n 0o 2xn (3. which in the setting of (3. For example. gn with fn + 0 .
and hence for large n P(Sn/n > x) > E [e. rc*(x) = sup(Ox .9S„+n' ( 9). we have P(nx < Sn < nx + 1. More precisely. LARGE DEVIATIONS Define rc* as the convex conjugate of rc.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.t. of P(X1 E dx) = E[e9X1K. since Sn is asymptotically normal w.425. (3. i. Define . exponential change of measure is a key tool in large deviations methods.sseo f which in conjunction with (3.q = rc* (x). we get P(Sn/n > x) E [e9nx +nK(9)9" '.rc(0) where 0 = 0(x) is the solution of x = rc'(0). P with mean nx and variance no. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). S rtn > x 1. Since P nn > x) = E {e_8 ' ( 9).3. V > 0 e. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.1).3) is put equal to x.4) n Next.r.96o /] > 0.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . which is a saddlepoint equation . In fact.e.960/) * 0. if we replace Sn by nx + o / V where V is N(0. 2 where o2 = o2(x) = rc"(0).4 enn +1.r. nx < Sn < nx + 1.(e)i XI E dx].(0)) e 307 (other names are the entropy.2). Most often. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .4) immediately yields (3.
.3. n Icn(0) exists and is finite for ry .o log Ee9Sn /n... In the application of large deviations to ruin probabilities. there exists z E (0. ..e < 8 < y + e. however..1)..e < 8 < y + e. Xn) and sn = x1 + • • • + xn (note that the r.. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. to be made rigorous.s.p > 7 < zn. Ee9X n < oo for e < 0 < e. Pn Sn1 .'(u) )Ng a"u... X2.h. we introduce a change of measure for X1.3 For each i > 0.. integrates to 1 by the definition of Icn)./^ >7 < zn n for n n0. and the WentzellFreidlin theory of slow Markov walks. Further main results in large deviations theory are the GartnerEllis theorem. We further write µ = tc'(ry). commonly denoted as is the saddlepoint approximation.. Sanov's theorem which give rare events asymptotics for empirical distributions.dxn) = 05nKn(7)Fn(dx1. ... we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . see Jensen u [215] or [APQ] p. .v. MISCELLANEOUS TOPICS which is the same as (3. that is. be a sequence of r. Mogulskii's theorem which gives path asymptotics. 1]. which is of similar spirit as the dicussion in VII.. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. We shall need: Lemma 3 . is differentiable at ry with 0 < K'(y) < 00...'s. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . For the proof. and write Sn = X1 + • • • + Xn.dxn) where Fn is the distribution of (X1i . Then i/. (iii) #c (8) = limn. 1) and no such that Sn . (iv) tc(ry) = 0 and r. Xn given by Fn(dxl. Assume that there exists 'y. (ii) lim supn. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). . .308 CHAPTER XI..2 (GLYNN & WHITT [163]) Let X1. The substitution by V needs. 260 for details.
we get lim sup 1 log Pn (Sn1 /n > µ + r7) < 0(1i + r7) + i(p(0 +'Y))/p n+oo n and by Taylor expansion.> . h.3. log zl'(u)/u > 'y.. ( U) P(S..n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .Bµ . h. We first show that lim inf„_. Clearly.m(7). can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. Then V.Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +.n > u ) = [ Em [em Em 1e. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.ne(p+ 17).n m µ 1 + rl .2. The rest of the argument is as before.. can be chosen strictly negative by taking 9 small enough. Since I EeqOX „ ] 1/q is bounded for large n by (ii).2. S. 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. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.W. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).]1/q = e. mµ Sm > u] km e7Sm+n.µ?7 .s.r (7) n = e. The corresponding claim for Pn(Sn/n < µ .077 n^oo n and by Taylor expansion and (iv ). in particular the r.h. u Proof of Theorem 3. S.+r7) < zn for n > no. it is easy to see that the r. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.s. is of order . 0.91) + o(O ) as 0 J.. This establishes the first claim of the lemma . the r . > 1 +17] m(7).y) .77) follows by symmetry (note that the argument did not use µ > 0).+r.YS.s.71 < e and jq9j < e.. for Sn.
and since Ic.YS +^c CHAPTER XI.(•) goes to 1 by Lemma 3. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).6) for some z < 1 and all n > n(E). Pn \ > la+ 8 I < zn (3.3. For lim supu... MISCELLANEOUS TOPICS (7).(Y). I > IL exp `S.7) so that n(b) I1 < e'Yu E en... n=1 ... P(T(u) = n) < P(Sn > u) = En [e7S. Sn > u] < eYu+Kn(7)pn(Sn > u) (3.I < µl1 1+77 I M 1_ 1+277 S. Obviously. 0 yields liminfu __. this is possible by (iii). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.310 ]Em I e. (iv) and Lemma 3.0 log i'(u )/u < 'y.+wn(7). 3. 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)..n(ry)/u 4 0andm/u* (1 + r7)/µ. 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 (.log z) /2 and Sn Fn\ n >lb+S) <Zn.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. logO(u)/u > ry.
eryu en logz/2p n nt n.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) . Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . Sn1 C U.10) 00 I4 < E F(Sn_1 < u.3. S.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. C 26u `p / +1 I e6u(1+6)/µ (3. 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 . LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. u .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).
we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). (7 + a) < 2arc'(7). Corollary 3.8) by P(S.u(1+b)/rc'(7)). we need to redefine n(b) as L. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.b)/i(7). it suffices to show that for j = 1. ryue«iu .z 1/z For I1. IV. the typical time is u/rc'(7) just as for the compound Poisson model.Q is so small that w = 1 .2.. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. the last steps of (3.7' a"ju.(u) = I1+I2+I3+I4'^ ery( u). we replace the bound P(Sn > u ) < 1 used in (3. For 14. e'.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 .3ui where . cf. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .312 CHAPTER XI.. For 12.4/3rc'(y) > 0. u . 13 = P(T (u) E (u(1 b)l^ (7). 2.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . u(1 + b)/i(7)) Proof Since V. I2. Letting c11 = maxn<n. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r.xl/2 to give the desired conclusion.('+'Y).4 Under the assumptions of Theorem 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. this is straightforward since the last inequality in (3. Then for n large. say n n1. For I.4.
1. we shall give two continuous time examples and tacitly assume that this can be done. 2 is in force with y = 2p/wz.. i..12) k=0.2 shows that the discrete time structure is used in an essential way.1.v. Obviously many of the most interesting examples have a continuous time scale. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. 09(9)... but nevertheless.3.'(y) > 0. 11 Inspection of the proof of Theorem 3.LARGE DEVIATIONS 313 Example 3 .g. To verify these in concrete examples may well present considerable difficulties. r. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * ..14) is needed in both examples .. An event occuring at time s is rewarded by a r. Assuming that the further regularity conditions can be verified. the key condition similar to (iii). V(s) with m. 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. Thus the total reward in the interval [0. If {St}t> 0 is the claims surplus process.f. t] is Rt = E V (Un) n: o„ <t . and we conclude that Theorem 3 . Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. Theorem 3. whether P ( sup St > u ltg a ^" 0<t<oo // (3..2 then immediately yields the estimate log F( sup Skh > u) a7u (3. and in fact. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.3(s) at time s.(O) = 9µ+02 for all 9 E R. criteria are given in Duffield & O'Connell [124].. Xk+l) k=1 00 naoo n provided the sum converges absolutely.e.13) One would expect this to hold in considerable generality.
this is not realistic .Q„) . Thus by (3. Kt (0) t (Ee9U"it8i J0 .. More precisely. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known.'`1 U. 7 Given the safety loading 77..0 and assume there are y.9t = /3 J t (Ee8U° i8l . It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .1) ds .14). a differential equation in t). is At . if the nth claim arrives at time a. assuming a continuous premium inflow at unit rate.1) .314 where the an CHAPTER XI. . Of course . (3. Then logEeOR° = J0 /3(s)(^8(9) .v. Of course. = U„ ( t . Since the remaining conditions of Theorem 3. we conclude that Cu) log e7 u (cf. Un(s).9t. derive .. but that a claim is not settled immediately.s). the best estimator of /3µB based upon Ft.g.t. i. nondecreasing and with finite limits Un as s T oo ( thus. we have S. We further assume that the processes {U1(s)}8>0 are i. O'n +S] is a r .. We let ic (9) = 3(EeWU° . n: o. MISCELLANEOUS TOPICS are the event times. where Ft = a(A8 : 0 < s < t). 0 and since EeOUn(8) + Ee°U^ as s * oo.Lundberg model has the larger ruin probability. If the nth claim arrives at time Qn = s. Example 3. it contributes to St by the amount Un(t . leading to St = At(1+77) Joo t S8 ds. one would take p(t) = (1 + rt)At/ t. Thus. the Cramer.d.2 are trivial to verify. (9) < oo for 9 < 'y + C. the above discussion of discrete skeletons).6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .15) . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .It. 0 Example 3 . then the payments from the company in [on. we have rct (9)/t 4 ic (9). <t which is a shotnoise process. Most obviously. Un represents the total payment for the nth claim). Thus.1) ds . e > 0 such that ic('y) = 0 and that r.14) (to see this . At = . e.1) ds rt (3.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.
18) Thus (iii) of Theorem 3. we have Nt t N. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .20) (3. again the above discussion of discrete skeletons) where y solves ic('y) = 0 It is interesting to compare the adjustment coefficient y with the one y* of the CramerLundberg model. we conclude that t. the solution of /3(Eelu .i. the Vi = . which yields eau f 1 t(1+n )audtl = E r Ee°Y = E [O(1+n)aueaul = E [eau J L Jo J L1+(l+r))aUJ .(1 + r7) log t (3.i.d. (3. i. Ui Nt / t 01i 315 St = Ui .1) . and since the remaining conditions are trivial to verify. (3. uniform (0. one has y > y' (3.log Oi are i. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .e./3.(1 +i) f > i= 1 s ds = E Ui 1 . Indeed.16) i=1 o i=1 Let ict (a) = log Eeast .3. equivalently.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 .19) with equality if and only if U is degenerate.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.1) or . standard exponential . Thus. LARGE DEVIATIONS With the Qi the arrival times.2 hold.d.b(u) IN a'Yu (cf.(1 + 17)0µB = 0.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. It then follows from (3. To see this .
(1 + ri)y*x is convex with k(oo) = 00. MartinL6f [256]. and k(x) < 0. [257] and Nyrhinen [275]. Therefore e7'U _ k(U) E [1+(1+77)y*U] . the proof of (3. assuming that the U. = P(N = n) = e(3an However. Further. [245]. y = y* can only occur if U . the function k(x) = e7*x .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. x > x0. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. . For Example 3.20) is due to Tatyana Turova. we then take t = 1 so that p. using that Ek(U) = 0 because of (3.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. MISCELLANEOUS TOPICS Next. much of the analysis carries over to more general cases.316 CHAPTER XI. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U.1 . Dembo & Zeitouni [105] and Shwartz & Weiss [339].19). this in turn yields y > y*.d. see also Nyrhinen [275] for Theorem 3. In addition to Glynn & Whitt [163]. For notational simplicity. k(0) = 0. 0 < x < x0. Further.. Lehtonen & Nyrhinen [244]. 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. and since tc(s). rc*' (0 ) < 0. see Nyrhinen [275] and Asmussen [25]. we are interested in estimating P(A > x) for large x. the study is motivated from the formulas in IV. though we do not always spell this out. Further applications of large deviations idea in risk theory occur in Djehiche [122]. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. at time t. a* (s) are convex with tc'(0) < 0 . with common distribution B and independent of Nt.7. The main example is Nt being Poisson with rate fit. This implies n(y*) < 0.xo. are i. In particular.i.2. say one year. k'(0) < 0.
9(Ax).4.3e(bo[a] .[s])3/2 = 0.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).1) where )30 = ."(0) = . B"' [s] lim (B".1). A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz . Vare(A) = s. For a given x. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.1 Assume that lim8T8. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. Then as x * oo. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).x)//3B"[9] is standard normal. Then Ee"A = e'(") where x(a) _ 0(B[a] .3B"[9]. K'(0) _ ic'(9) = x. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. B"[s] = oo. Hence P(A > x) = E e [e9A+ ic(9). In particular. The analysis largely follows Example 3.1.1).2) implies that the limiting Pedistribution of (A . i.ic(9) = . no(a) = logE9e'A = rc(a + 9) . we define the saddlepoint 9 = 9(x) by EBA = x. A > x)] = eex+K( e)E9 [e .e. only with 0 replaced by a9 and B by B9. Proposition 4. A E dx] . 818' where s' = sup{s : B[s] < oo}. (4.
1). Y satisfies 9(u) ti eu2/2(1 + ibu3) (4.e.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. see Embrechts et al.EN B(x). i. some regularity of the density b(x) of B is required.l3pB.1 goes all the way back to Esscher [141]. For example. more generally. it is quite questionable to use (4. leading to P(A > x) :. 2). Remark 4 . just the same dominated convergence argument as in the proof of Theorem 2. For details. [138]. then P(A > x) .2 If B is subexponential and EzN < oo for some z > 1.x') where x' = sup {x : b(x) > 0}. For a rigorous proof. bounded with b(x) . large x. Notes and references Proposition 4. it holds that EA = . In particular. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. b(x) = q(x)eh(z).3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. either of the following is sufficient: A. b is logconcave.1 yields: Proposition 4. The present proof is somewhat heuristical in the CLT steps. under the Poisson assumption (4. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. 1 . For example.2) is often referred to as the Esscher approximation. 3 A word of warning should be said right away : the CLT (and the Edgeworth expansion) can only be expected to provide a good fit in the center of the distribution .2i and that (A . the distribution of A is approximately normal . and (4. Jensen [215] and references therein. 4b The NP approximation In many cases .Q{AB (4.3) and related results u for the case of main interest . or.318 CHAPTER XI. Furthermore 00 b(x)Sdx < oo for some ( E (1. A covers the exponential distribution and phasetype distributions. b is gammalike. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. B covers distributions with finite support or with a density not too far from ax° with a > 1.(D X .v.ycix °ie6x B. Var(A) _ ^3p.4) . Thus . In fact.
5) follows by integration.6) .4.6(1 .f. In concrete examples .. the CLT for Y = Y6 is usually derived via expanding the ch. K2 = Var (Y).. the NP (normal power) approximation deals with the quantile al_E.5) may be negative and is not necessarily an increasing function of y for jyj large.3& (y).i 3 K3 } Pt^ exp .i 6 r 1 3 so that we should take b = ic3/6 in (4.1). Thus if EY = 0. . .5). If this holds . (4..l = EY.c2i. then P(Y < y) 4(y) .5) is obtained by noting that by Fourier inversion.3!). If the distribution of Y is close to N(0. Var(Y) = 1 as above . so that 1(u) 3 exp { .. and from this (4.equantile in the distribution of Y.: EA + zl_E Var(A) .EY)3.. however. Remark 4.y2)^P(y)• 319 Note as a further warning that the r. defined as the the solution of P(A < yle) = 1 . K3 = E(Y .2K3 + 4i 64 + . Rather than with the tail probabilities F(A > x). as u2 u3 u4 9(u) = Ee'uY = exp {iuci . in particular.2X2 .s.e. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. A particular case is a.. are small. which is often denoted VaR (the Value at Risk). . the density of Y is 1 °° _ eiuy f(u) du 2x _. ylE should be close to zl_E (cf. and so as a first approximation we obtain a1_E = EA + yle Var(A) . s..99.h.5 (y3 . where Kl . one expects the u3 term to dominate the terms of order u4. K4 . the standard normal distribution. one needs to show that 163. Heuristically. zl_e be the 1 .2 2 . (4. of (4.EA)/ Var(A) and let yl_E. are the cumulants . Let Y = (A . resp. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) ..2 ^ \1 . u5..
MISCELLANEOUS TOPICS A correction term may be computed from (4. this yields the NP approximation 6(Z1 _E .zi. Note.1) E (A .S(1 . Using Y = (A . that [101] distinguishes between the NP and Edgeworth approximations..y2)cp( y) term. 21 ..320 CHAPTER XI.5) by noting that the 4.E ..1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. k3 is small for large /3 but dominates 1c4.1)! n ^eQ .3ni /3 .E = z1E + S(zi_E . K5 . as required . 4c Panjer 's recursion Consider A = constants a.zl E)V(zl_E) .zlE)W(zlE) 1 .E)A1 l E) 1 E 4)(yl E) ^' .7) as 1 (3) a1E = Qµa +z1 . and assume that there exist n ) Pn_i . For example.. let pn Pn = (a+ = P(N = n). however.yi.. the kth cumulant of A is /3PBk' and so s.k = /3µB^1 / (. [101].S(1 . b such that EN 1 U%. n = 1. In particular . this holds with a = 0.E(/3PB^1 )1^2 + s(z1E .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .. .6pBki) d/2.5(1 .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4. We can rewrite (4.1).EA ) / Var(A).E + (yl. This leads to t( yl E) .zl E)^o(zl E) .E )Azl E) 4(z1E) + ( ylE .6 (1 .1)EY3. Another main reference is Daykin et at.(y) terms dominate the S(1 .
11) Remark 4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4... then j (a + b!) 1ag k_1 3 gkfj. . 1... j = 0. . if go = 0.k . the complexity (number of arithmetic operations required) is O(j3) for (4. j = 1. 2. .5 The crux of Proposition 4. Since the sum over i is na + b. .. Then fo = >20 9onpn and fi = 1 E In particular. The expression for fo is obvious. . and calculating the gj*n recursively by 9*1 = 9j. (4. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4..12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 .} and write gj = 2 . the value of (4.12). .4. .. (4. fj = P(A = j).14) is independent of i = 1. . .13) Namely. j1 g.1. 2. fj = E (a+ b k =1 )9kfi_k . (4.4 Assume that B is concentrated on {0.10) f o = po. u Proof of Proposition 4. By symmetry. n = k=n1 9k(n1 )9j k • (4.. Hence by (4.4.14) is therefore a + b/n.. E[a +bU=I >Ui =j l i=1 J (4.13) but only O(j2) for Proposition 4.. 2. (4..12) where g*n is the nth convolution power of g.4.4 is that the algorithm is much faster than the naive method. n. j = 1.9)..
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).
6).5) reduce quite a lot.d. however. (6.h(Ui)] . and so on. appealing to (6.4).6 Assume the Ui are i. The arrangement used in practice is. Local reinsurance with h(u) = (u .P I = EC [7]N. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h(U)]. this implies 7* > 7.d.b)+ = Eh(U) (and the same P) satisfies 71 > ry.P.h(U) (as in the proof of Proposition 6. ' ii (6.h(Ui)] . Proof As in the proof of Proposition 6. expectations like those in (6.P } < 1 = Eexp E[Ui.3).5 Because of the independence assumptions . i. we get N 1 = Eexp ry E[Ui ii . the excess ofloss rule hl (u) = (u . Eexp 7 [E [Ui .4). (6.b)+ with b determined by E(U . as often local as global. y = Ei [Ui .h(Ui)] . ry* > 0 because of (6.h(Ui)P JJJ l:='l {ry ] or.6. then (6. it suffices to show that Eexp {ry ii 'UiAb.4).h(u) and any P satisfying (6. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . we get EX = EN • EU. This follows by taking Xl = U A b. N E X .P]}. X2 = U . Assuming for simplicity that the Ui are i.4) and u g(x) = e7x in Ohlin's lemma.5) holds trivially. Remark 6..h( UU) = EN • E[U .h * (X) . .6) u where C[ry] = Ee'r(u4(u)).4. Then for any local retention function u .P > EexP{7[X . u But since ry > 0.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .i.
see also Sundt [354].330 CHAPTER XI. Bowers et at. See further Hesselager [194] and Dickson & Waters [120].g. The present proof is from van Dawen [99]. [76]. The original reference for Ohlin's lemma is Ohlin [277]. e. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. Heilman [191] and Sundt [354]. .many texts on insurance mathematics.
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.. .. are independent and Y1. 2h.. Y. then Stone 's decomposition holds : U = U. Then Blackwell 's renewal theorem holds. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. i. not concentrated on {h. t]) so that U(t + a) .U(t) is the expected number of renewals in (t. = T„ ..T„_1)..} for any h > 0. Y2.t. t +a]).. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 .e. of interarrival times and the time Yo = To of the first arrival (that is.r. the renewal process is called zerodelayed.. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Y1. some condition is needed: that F is nonlattice. t 00 (A.. t] is denoted by Nt. stating that U(t+a)U (t) ^ a. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. the distribution of Yo is called the delay distribution. note in particular that U({0}) = 1. The point process is called a renewal process if Yo. If Yo = 0. when t is large.. all have the same distribution. . Lebesgue measure dt normalized by the mean to of F. denoted by F in the following and referred to as the interarrival distribution.1) (here U(t) = U([0. The renewal theorem asserts that U(dt) is close to dt/µ. of epochs or the set Y1.. Y2. The mathematical representation is either the ordered set 0 < To < T1 < . Lebesgue measure for some n > 1). . . The number max k : Tk_j < t of renewals in [0.. Technically. That is.
µF (A. Equivalently.out. and F(dx) a known probability measure .332 APPENDIX u(t) has limit 1/µ as t 4 oo.i. z(u) a known function. (A.3) Further. IV).9.2). (A. see [APQ] Ch. (A. stating that U(t)/t > 1/p.4) that z is Lebesgue integrable with limZ. i. Both result are valid for delayed renewal processes.5) 2This condition can be weakened considerably . then Z(u) i f0 z(x)dx .x)F(dx). u u PF 4 00. U Z(u . (A.2) has the unique solution Z = U * z. then it suffices for (A. but suffices for the present purposes . the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.2) Z(u) = J0 u z(x)U(dx). A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. In 111.2 Assume that Z solves the renewal equation (A.R.. resp. Under weak regularity conditions (see [APQJ Ch. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 .a. and that F has a bounded density2. that z(u) has a limit z(oo) (say) as u 4 oo. z(x) = 0.4) If F is spread. Note in particular that F is spreadout if F has a density f. in convolution notation Z = z + F * Z.e.1 if F is nonlattice and z (u) is directly Riemann integrable (d. the statements being EN(t + a) . wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. oo).i". IV). Then Z(u) 4 z(oo).EN(t) .
APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. . F(dx) = e7xF(dx). or many queueing processes. To this end. that the existence of y may fail for heavytailed F. The simplest case is when {Xt} has i. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). i.t. multiply (A.. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i.. . however. T1.. Y2.. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.5a. . This program has been carried out in III.3) satisfied by the ruin probability for the compound Poisson model. Tk (or. the present more general definition is needed to deal with say Harris recurrent Markov chains.... Here the relevant F does not have mass one (F is defective). we let µ denote its mean. Assuming that y can be chosen such that f °° Ox F(dx) = 1. T1.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO).} be a renewal process. • . The kth cycle is defined as {XTk+t}o<t<Yk . and its distribution does not depend on k. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. A regenerative process converges in distribution under very mild conditions: . the postTk process {XT. . 1c Regenerative processes Let {T. The distribution F of Y1. is called the cycle length distribution and as before. Note. equivalently. of Yo.. a basic reason that renewal theory is relevant is the renewal equation II. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. {Tn} if for any k. . asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. .i. Z(u) U = 1 u 1 u f z(u . We let FO.e. Tk and {Xt }o<t<Tk • For example.(3.x)u(x) dx = z(u( 1 .r. z(x) = e7xz(x). cycles. However. Y1 . this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. Yk ).d.. Eo etc. However. Hence by dominated convergence.k+t }t>o is independent of To. that F is a probability measure.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). refer to the zerodelayed case. 0 PF µF 11 In risk theory.
.. then (Zt . i..ZTOI < 00. Then it (ii. oo). {Tn} if the processes {ZT +t . This is the case considered in [APQ] V. r..ZT }0<t<Y„+.oo (i..0 is called cumulative w.. If p = oo.. {Tn}. and q(t) = sup It . assume that p < 00 and define Un = ZT}1 .334 APPENDIX Proposition A1.r.ZT Then: (a) If E sup I ZTo+t . µ 0 If F is spreadout.3..'s by e. for n = 1.r.d. Then {Zt}t^. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. are i.r..t.e. Then Xt Di X. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.v. {i7(t)} are Markov with state spaces (0. Otherwise .tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.t. {Tn}. under the condition of Blackwell's renewal theorem.t : t < Tk}. 0<t<Yi then Zt /t a$• EU1/µ. just the same proof as there carries over to show: Proposition A1.. 2.t.i. and we have: holds more generally that (rl(t).3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo..Tk : t < Tk} as the age. e(t )) .+ X. resp . (b) If in addition Var(Ul ) < oo. but in fact. in total variation.d. C). where the distribution of X.. Then {e(t)}.4 Let {Zt}t^. (A.e.i. fi (t) = inf {Tk . is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. cycles (we allow a different distribution of the first cycle).0 be cumulative w. then e (t) . [0. oo). An example is Zt = fo f (X8) ds where {Xt} is regenerative w. then Xt . We denote the limiting r.6) id Cumulative processes Let {Tn} be a renewal process with i.
Y1 > t] 4 0. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.5 Under the condition of Blackwell's renewal theorem. = z is Foz) The proof of (a) is straightforward by viewing {(r. r. Y1i Y2. (b) the joint distribution of (ri. In IV. EC(t)/t + 0. Then fi(t)/t a4' 0 and. (1 V)W) where V. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo .. if in addition EYo < oo.i. we used: Proposition A1.4.. Hence t t lt ) = f U(dy)z(t . U(x + 1) . Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). (c) the marginal distribution of q is FO. the first statement follows. Yo > 0] + f Eo^ (t . Proof The number Nt of renewal before t satisfies Nt/t a4' p.y)P(Yo E dy) . Hence for t large enough. are not i. the joint distribution of (rl. ^) is given by the following four equivalent statements: (a) P (77 > x.U(x) (c < oo because it is easily seen that U(x + 1) .i.d. use t E^(t)/t = E[Yo . Since z ( k) < E[Yi . Since the maximum Mn of n i.y) = f U(t .(t). W are independent. For the second. ^ > y) = 1 f +Y (z)dz.APPENDIX 335 Theorem A1. and the equivalence of (a) with (b)(d) is an easy exercise.U(x) < U( 1)).v. Yl > t].t. assume first the renewal process is zerodelayed. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y)..t.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup.d. V is uniform on (0. (d) the marginal distribution of ^ is FO. In the general case. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. and the conditional distribution of ri given l.6 Consider a renewal process with µ < oo. the sum is o(t) so that Eo£(t)/t + 0 . .^(t))} as a regenerative process. l:) is the same as the distribution of (VW. but governed by a Markov chain {Jn} (we .
the conditional distribution of {XT„+t}t>o given Yo.T_ < oo). For example. Yn.. where the distribution of X. Jn +1=j} where J = a(JO. Jo.) and (Fij )i.. J1 i .t. namely {Twk } where {Wk } is the sequence of instants w where Jo. . Notes and references Renewal theory and regenerative processes are treated. e. < yIJ) = Fij( y) on {Jn= i. . X2. .. in [APQ]. G+(x) = P(S.i . A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo..} is nonlattice (it is easily seen that this definition does not depend on i). distribution ofjXt}t>o itself where Pi refers to the case Jo = i. Let X1.. These facts allow many definitions and results to be reduced to ordinary renewal. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. Jn_1. . = io for some arbitrary but fixed reference state io E E..g...jEE is a family of distributions on (0. Further: Proposition A1. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. the semiregenerative process is called nonlattice if {T. Sn = X1 + • • • + Xn the associated random walk. Alsmeyer [5] and Thorisson [372]. Y1.d. oo). .and regenerative processes. and define r+=inf{n>0: Sn>0}. Jn = i is the same as the P. oo). IT.. . Then Xt 4 Xo. the Markov renewal process if for any n. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj..r.r. 0] or (0 . . The semiregenerative process is then regenerative w..7 Consider a nonlattice semiregenerative process..336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.+ < x.t. with common distribution F. T_=inf{n>0: Sn<0}. . . .. A stochastic process {Xt}t>o is called semiregenerative w. A Markov renewal process {Tn} contains an imbedded renewal process.. be i. r+ < oo).. G_(x) = P(ST_ < x.}.
r. S. the renewal measures U+=>G+. Proof Considering the restrictions of measures to (oc. F(A .S.. 0]. . >0. A C (oo.7) (A.x)R+(dx).. A C (0.g.>0. A C (oo. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. Sr_ _1 is at its minimum . 0]). (e) R_ = U+.T_=n} = {S. (c) G+(A) = f °.=EGn. G_. 0). In (A. oo). define w as the time where the preT_ path S1. (A. G+. More rigorously.8) (e. (d) R+ = U_. F(A) + (G+ * G_)(A). n 0 R_(A) = E I(Sn E A).APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). oo) (A.x)R_ (dx).1 .1 (a) F = G+ + G_ . n=0 The basic identities are the following: Theorem A2.S. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). .=n w=m i Figure A. u . oo). we consider the last such time (to make w unique) so that {w=m. 0<j<m. A C (0. On {T_ > 2}.7).. U. m<j<n}.7) follows since G+(A) = 0 when A C (oo. n=0 n=0 00 00 and the T+. 0] and (0.G+ * G_: (b) G_ (A) = f °° F(A .
and the proof of (A..7) follows. ST+Edu). SmEdu) = P(T+=m. m it follows (see Fig.XnEAx) 00 f 0 f 0 00 00 1: F(A . E du) = P(T_=nm. 0<j<m.3.x)P(Sk < 0. It follows that for n > 2 F (7. A. Sr_ E Adu) (s ee again Fig . S._ = n .8) is similar. A. . ._ E A . (b) follows from 00 G+ (A) _ E F(Sn E A. (A.1) that P(Sj Sn. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. 0 < k < n. clearly (Sj Sm>0._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). SnEAIS. S. .x)R+(dx). m=1 f S mming over n = 2. and reversing the order of summation yields P(T_ > 2.. ST_ E A) P(T+ = m..m. Aso.Sn_1Edx.1. m < j <n.. Sn1 E dx) n=1  F(A .0<k<ri .= n.F(r_n_mSrEA_u)..>0.3 8 APPENDIX Reversing the time points 0. ST_ E A .+ E du)P(S.u) f0m m=1 n=m+1 00 J0 OO P(S.+ E du) E P(S.1)..du) (G+ * G)(A)• C llecting terms...
's. 11.. E.0<k<n. For example. it serves as model and motivation for a number of results and arguments in continuous time.9) whenever F[s]. Summing over n yields R+ (A) = U_ (A). is based upon representing G+ as in (b). Nevertheless.2 In terms of m. then T+ = inf It > 0 : St = 0} is 0 a. which is basic for the PollaczeckKhinchine formula.g.SnEA) is the probability that n is a weak descending ladder point with Sn E A. 6+ [s]. u Remark A2.P as a product H+H_ of functions with such properties. we can rewrite (a) as 1 .O<k<n.SnEA) = P(SnSn_ k.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.1).1. and similarly H_ (s) = 1 . The classical analytical form of the WienerHopf problem is to write 1 . being concentrated at 0. P(SnEA . In continuous time.0<k<n.s. the analogue of a random walk is a process with stationary independent increments (a Levy process.4). Sk = X1 + • • • + Xk = Sn .g.SnEA) = P(Sn<Sk. For (d). G_ are trivial. Since G+ is concentrated on (0. and the proof of (e) is similar. the derivation of the form of G+ for the compound Poisson model (Theorem 11..APPENDIX 339 and the proof of (c) is similar. there are direct analogues of Theorem A2. there is no direct analogue of Theorem A2.G_[s]) (A.g. The present proof of Theorem A2.1(a) is from Kennedy [228]. consider a fixed n and let Xk = Xn_k+l. and sometimes in a larger strip. this holds always on the line its = 0.F[s] = (1 . WienerHopf theory is only used at a few places in this book.f. see for example Bingham [65].6. . Another main extension of the theory deals with Markov dependence. and G+.0+[s])(1 . and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). such developments motivate the approach in Chapter VI on the Markovian environment model. a number of related identities can be derived. see e. In this generality of. if {St} is Brownian motion.SnEA) = P(Sn<Sk. In discrete time. 0]. However.O<k<n. u Notes and references In its above discrete time version. oo). G_ [s] are defined at the same time. the survey [15] by the author and the extensive list of references there. Again.Sn_k.1. Then for A C (oo.T+> n) = P(Sk < O. cf.
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. Eo Kn/n! converges rapidly and can be evaluated without p oblems. three of the c rrently most widely used ones: xample A3. however . and eQ can then be computed as the mth power (by squaring if = 2). ere A is the eigenvalue of largest absolute value.11) A f eAtdt = eA. _I 0 (A.13) henever A is a diagonal matrix with all diagonal elements nonzero. 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). JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).12) eA'AO = Ale AA (A.10) d dteAt = AeAt = eAtA (A. if m is s fficiently large. 0 . It is seen from Theorem VIII.5 that when handling phase type distributi ons.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. To circumvent this. Thus. 1. Here are. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. 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. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting .
some jumps are dummy in the sense that no state transition occurs ). the procedure consists in choosing some suitable i > 0. Zo = h). p different eigenvalues Aj i . i. Ap.e. 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 .. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. The approach is in particular convenient if one wants eQt for many different u values of t.4 (DIAGONALIZATION) Assume that Q has diagonal form.e. The probabilistic reason that (A. we have k = QK (or KQ) which is a system of p2 linear differential equations which can be solved numerically by standard algorithms (say the RungeKutta method) subject to the boundary condition Ko = I. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. .]t)n (A. condition upon the number n of Poisson events in [Olt])  Example A3. In practice.15) Then it is easily checked that P is a transition matrix . construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. 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.. However .APPENDIX 341 Example A3.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.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}.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. One then can reduce to p linear differential equations by noting that k = ZQ.. Here is a further method which appears quite appealing at a first sight: Example A3 . Let vi.. i. .3 i (A. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. Zo = a (Z = QZ.. To this end.2 (UNIFORMIZATION) Formally.7t) n=0 n! u °O n Pn (to see this. vp be the corresponding left .
18) contains terms which almost cancel and the loss of digits may be disasterous. (A. we have an explicit formula for eQt once the A j. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. of largest real part is often real (say. hp the corresponding right (column) eigenvectors.17) eQt = E e\`thivi = E ea:thi ® vi. (A.. say A = (Ai)diag. Complex calculus : Typically. under the conditions of the PerronFrobenius theorem). vi. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.. Qhi = vihi. hp. i # j. i= 1 i=1 P P (A..5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. the eigenvalue. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.. and hence A2 is so because of A2 = tr(Q).16) (A. hi have been computed. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. D = ) 2 2 . i=1 i=1 Thus. Nevertheless.g H1. The phenomenon occurs not least when the dimension p is large. Then P P Q = > Aihivi = E Aihi (9 vi...342 APPENDIX (row) eigenvectors and hl. Example A3.18) Namely.. not all ai are real. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. v5Q = Aivi.. say Al. some cases remain where diagonalization may still be appealing. we can take H as the matrix with columns hl. and vihi ¢ 0. Then vihj = 0. and we may adapt some normalization convention ensuring vihi = 1. however. and writing eQt as eQt = He°tH1 = H (e\it)di. There are.
l ab (g12g21 + (A1  411) 2) = 1. 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. i. k  C k2 ) =b ( A1 q 1 Q11 / where a .7 Let 3 9 2 14 7 11 2 2 . v2 and h2 can be computed in just the same way. Then 7r = (ir1 7r2 ) = a (q21 Al .APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.20) ir = q2 ql qi +q 2 9l +q2 (A. The other eigenvalue is A = A2 = q1 .e. replacing ai by A2. u Example A3. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. 1) .19) Example A3 . where (A.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. Of course.q. However. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. h2 = Thus.k1).
22) Note that in this generality it is not assumed that A is necessarily square. A+AA+ = A+. APPENDIX x1 3/2 .23) .satisfying AAA = A.6.. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. (A. (A+A)' = A+A.5 .11/2 . Generalized inverses play an important role in statistics.344 Then D= 2+ 11)' 7 T4 2 =52. 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_. but only that dimensions match . They are most often constructed by imposing some additional properties . for example AA+A = A. (A..11/2 + 5 1. (AA+)' = AA+. 2 2 1=ab(142+(1+2)2 ) = tab. and a generalized inverse may not unique. A2 = 3/2 .
one is also faced with singular matrices . = 0 where m < p is the rank of A.g.P).APPENDIX 345 A matrix A+ satisfying (A.23) is called the MoorePenrose inverse of A. .g. Then for some b > 0. Am > 0.I) (A.D + O(ebt).24) = te7r . Am+1 = .eir )1.P + e7r)1 (here ( I .= (I . 0 01 In applied probability. Here is a typical result on the role of such matrices in applied probability: Proposition A4..P + e7r ).1 goes under the name fundamental matrix of the Markov chain).e.eir)1 = I . one then works with Q = (Q .. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . lt o eAx dx = te7r + D(eAt .1 Let A be an irreducible intensity matrix with stationary row vector it. (A. if A is a possibly singular covariance matrix (nonnegative definite). _ A. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . .25) .. Rather than with generalized inverses . E. and exists and is unique (see for example Rao [300])..eir ).ew..e ® 7r)1. are ordered such that Al > 0. (I . then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. ( Q .1Q = Q(Q . Assume that a unique stationary distribution w exists . and define D = (A ..
B(t) denote the l. B'(t) = e7r + DAeAt = eir + (I .I) .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt .27) Proof Let A(t).D + D2 + O(ebt).24). For example. and the columns to h. .2 Let it be a row vector with m components and h a column vector with k components. the formulas involving O(e6t) follow by PerronFrobenius theory.91a(2) .I)}. 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. and in fact any rank 1 matrix can be written on this form.eir)eAt = eAt = A'(t).2e7r .26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .J {xe^r + D(e . Interpreting 7r.DZ(ent . see below. Then A(O) _ B(O) = 0. .e. of (A. Equivalently.26) 2 = 2 e7r + tD . the r.I)} dx.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). (A.h. resp.s.. it follows that h ® it is the k x m matrix with ijth element hi7rj .I) (A.h. the rows are proportional to it. respectively. I. h ® it reduces to hit in standard matrix notation. (A. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. h as 1 x m and k x 1 matrices. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.s. 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. o Finally. Note that h ® it has rank 1.
4vf.(A.31) Indeed. C2 = h2 are column vectors.29).29) If A and B are both square (k1 = ml and k2 = m2). then v1B1h1 and v2B2h2 are real numbers.3f 4v/.3V8.A9.APPENDIX 347 Then A®B = 2 f 20. and the number of such factors is precisely given by the relevant binomial coefficient.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .50 6 7 6 4f 4.3vV/72f 20. 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.3v'6. such a factor is Ak (&B 1k according to (A.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.5v/.31). (A B)' = eA®B e! L 1=0 0 . then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). if Al = vi. if A ® I occurs k times. each of which is A ® I or I ® B. Using (A.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). (A.4 eA® B = eA ®eB. (A. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.28) In particular. A2 = v2 are row vectors and C1 = h1.
s. P(2). Let P8f P(Sl). resp . P(t) Yt(2) }. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). Thus .348 APPENDIX Remark A4.33) . k any column vectors. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). X ) }. first term on the r .5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. p = P(1) ® {X }.I)(h ® k). represents ces Q( 1). Let further it. we have P8 = Pal) ® p(2). {Yt(1). { 1't(1) }. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . in the definition (A. { On the other hand.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . A special case of Proposition A4. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }.3 < 0 Lemma A4 . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. where transition matrix of the bivariate Markov chain {X n1). 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 +.6 Suppose that A and of B. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. independent Markov chains. h. (A. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . the same time. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . Q(2).32). From what has been said about matrices of {Yt( 1). Yt(2 ) }.
. il... p there should exist io. h can be chosen with 3By this.APPENDIX 349 Proof According to (A. and if we normalize v..3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. (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. h can be chosen with strictly positive elements. . then IN < Ao for all A E sp(A). in such that io = i..1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. . .g. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k).The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. [APQ] X.34) Note that for a transition matrix. Here is the PerronFrobenius theorem. .. f o r each i. we have AO = 1. . A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix.29). > 0 for k = 1.. see e. and the corresponding left and right eigenvectors v. i. ao). .12). and the corresponding left and right eigenvectors v. E (0. h = e and v = 7r (the stationary row vector). We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. . = j and atk_li.7 Let A be a p x pmatrix with nonnegative elements. n.. Similarly. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. j = 1. That is. h such that vh = 1. Then the eigenvalue Ao with largest real part is simple and real. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. (A. Now note that the eigenvalues of A ® B are of the form a +. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . which can be found in a great number of books. and appeal to (A.
not only in the tail but in the whole distribution. the phasetype distribution B(a) with representation (. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). The next result gives a condition for asymptotical exponentiality. I.(ti)ding.35) for some p E (oo.350 APPENDIX strictly positive elements.1.e. Furthermore. h = e and v = 7r (the stationary row vector). Proposition A5. but is an easy consequence of the PerronFrobenius theorem.. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . if we normalize v. the analogy of this procedure with unformization.(3. note that we can write the phase generator T as Q .(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0).1 Let Q be a proper irreducible intensity matrix with stationary distribution a. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. the condition is that t is small compared to Q. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Ao). T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. Corollary A4. Example A3. To this end. 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). A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.8. we have A0 = 0. let {Yti°i } be a Markov process with initial distribution a and intensity . Then for any (3. 10) and use the formula me at e Bt = e 00 Antn = e .n t AL n=0 n! (cf. relate the eigenvalues of B to those of B via (A. h such that vh = 1.2). then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. Note that for an intensity matrix. For example.8 is most often not stated explicitly in textbooks.
Conditioning upon whether { Yt} changes state in [0. Let further V be exponential with intensity V and independent of everything else. t < (a). 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.aE where 0 < e < 1). has a limit distribution: Proposition A5. a'/a + 1. 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}. from which the phase process is terminated .jEE.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. By the law of large numbers for Markov processes . and that Yt(a) = Yat for all t. a' = a . dx/ti] or not. {t Y( a) } v>0 . Proof Assume first ti > 0 for all i and let I. Then {Ix} is a Markov process with to = Yo.9. v/ t. in fact . We shall .APPENDIX 351 ((1) etc.a' + oo (e. prove a somewhat more general result which was used in the proof of Proposition VI. Hence O ((a) aa. We can assume that Jta) = Yt(°). and this easily yields a(x)/x a' 1/t*.1. Then a(a'V)/a (aV) a' 1. J^O)_ = j) Pi (v(aaV) > x.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.YQ(av) = j) Pi ( ci(a'V) > x. it states that the state. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. fo tY dv/t a$' t*.g.2 Pi (c(a) > x.x (1 . Since JJ(. a . and write Yt = Yt(1).Yj(av) = j f . In addition to the asymptotic exponentiality. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .(a) > x . J(()) _ = i) + at•x t tt' . = YQ(x).
1 and A5. See also Korolyuk. > 0}. so is the geometric distribution. Et II I a(a^V) > x) at' . Then P is substochastic and the vector of exit probabilities is p = e .x k > K.. Then: (a) The point probabilities are bk = aPklp..zP)'p.. . .+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. is discrete phasetype. with point probabilities bk = (1 . 1 k=1 1 0 otherwise. Penev & Turbin [238]. let E and Pkj j=k1. 2.j) and initial distribution a.. k = 1...} is said to be discrete phasetype with representation (E. a). ' pk 0 k>1 11 Theorem A5. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case.5 Let B be discrete phasetype with representation (P.2 do not appear to be in the literature. u Notes and references Propositions A5.. k>1. (c) the nth moment k 1 k"bkis 1)"n!aP"p. an easy modification of the argument yields finally the result for the case where t.. Indeed. zkbk is za(I . the simplest discrete phasetype distribution: here E has only one element.3 As the exponential distribution is the simplest continuous phasetype distribution. say bk = 0. = 0 for one or more i. these results are in the spirit of rare events theory for regenerative processes (e. so we shall be brief.. P. . A distribution B on {1. (b) the generating function b[z] _ E' . and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.. Example A5. However. Keilson [223]..Pe. 2. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).4 Any discrete distribution B with finite support.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) . Example A5. a = b = (bk)k=1.g. K}. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p.p)k1 p.
_ i E E(1) T(1) t(1)a(2) i E E(2) . and a=1).a(1).7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.{ 0.r + 1. 11 Example A5. Then {Jt} has lifetime U1 + U2 .a(2). { Jt 2) } with lifetimes U1 . a' .APPENDIX 353 5c Closure properties Example A5.1 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) }. . A... U2. resp. as is seen by minor modifications of Example A5.6 is the Erlang distribution Er which is the convolution of r exponential distributions.T(1)). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). Jt t > U1 + U2. and hence the negative binomial distribution is discrete phaseu type. initial distribution a and phase generator T.. (E(2). Then the convolution B = B1 * B2 is phasetype with representation (E.. r .T(2)).36) in blockpartitioned notation (where we could also write a as (a (1) 0)). resp. T= ( 0 T(2) ) (A.6 (CONVOLUTIONS) Let B1. B2 be phasetype with representations (E(1).6. a. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.
p)pn1. Example A5. A reduced phase diagram is 0a(1) E(1) A .10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. we need to restart the phase process for B w. resp. In risk theory.0)a(2) E(2) Figure A. To obtain a phase process for C. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). Thus.T. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.0)a(2))).p)pn1B*n.. are i. a reduced phase diagram is f a E t Figure A. p at each termination. P(N = n) = (1 . then C is the distribution of Ul + • • • + UN..O)B2 (0 < 0 < 1) is phasetype with representation (E.a(2). if U1.'). U2.T(2)). and consider B(") = fA B(a) v(da) where v is a probability measure on A. a. Example A5. Then the mixture B = 9B1 + (1 . this means that a = (Oa(1) (1 ..354 APPENDIX Example A5.E) where a(°) = fAa(a)v(da). with common distribution and N is independent of the Uk and geometrically distributed with parameter p.T(1)). Equivalently. a.37) (1) (1 .0)ai2).p.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a.8 (FINITE MIXTURES) Let B1. and o'i Oa. T) and C = EO°_1(1 . i E E(1) T 0 I (A. (E(2). B2 be phasetype with representations (E(1).3 In exactly the same way. a mixture of more than two phasetype distributions is seen to be phasetype. i E E(2) 0 T(2) =IT (in blockpartitioned notation.i.a(1).4 .d. Let B(") be the corresponding phasetype distribution. Then it is trivial to see that B(") is u phasetype with representation (a(").
a(1). a(1) ® a(2 ).aF[T]. let the phase space be E x F = {i j : i E E. T + ta. To obtain a phase representation for C . T) and C = F. U2 be random variables with distributions B1.d. { Jt2) } be independent with lifetimes U1. E).. say v. are i. T(2) ). of F. if {Jt} is a phase process for U. It is zeromodified phasetype with representation (E.1. v..2. resp.aeTx. If U1 has a different initial vector.v.APPENDIX 355 and C is phasetype with representation (E.T + pta). Minor modifications of the argument show that 1. Example A5. cf. then Jy has distribution aeTx. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype.°. i.x)+. with common distribution B and N is independent of the Uk with P(N = n) = f. then C is the distribution of U1 + • • • + UN. let {Jtl)}.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . then U1 +• is phasetype with representation (E. let B be a continuous phasetype distribution with representation (F. Example A5 . if B is defective and N + 1 is the first n with U„ = oo. Note that this was exactly the structure of the lifetime of a terminating renewal u process. a(2). Indeed. T(1) ® T(2)). be the point probabilities of a discrete phasetype distribution with representation (E.. T). 13 (MINIMA AND MAXIMA ) Let U1. { 4 } as exit of {Jt}. but the same T.TWWW). if U1. then U1 + • • + UN is zeromodified phasetype with representation (a..T) if U is phasetype with representation (E. +UN 2. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }.. X independent of U.X)+ is zeromodified phasetype with representation (E. T + pta)..g. it follows by mixing (Example A5.°_1 f„ B*?l. To see this. For U1 A U2. Proposition VIII.. B2 of phasetype with representations (E('). . resp. a.. cf. P). Example A5 . a.7.2. 12 (PHASETYPE COMPOUNDS ) Let fl. (E(2).9) that (U . U2. a. j E F}.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). say with distribution F. v. If we replace x by a r. . f2. Equivalently.f. U2. Corollary VIII. Thus the representation is (E(1) x E(2).
.. see Neuts [269] (where the proof.(bk) + B(bk) for all k as n * oo. we need to allow { Jt. The general case now follows easily from this. Here are the details at two somewhat different levels of abstraction: (diagonal argument .. Then from above.. and let Bn be the Erlang distribution E. however. cf.(Sn) with Sn = n/b.xq(n)(n)}. Let the support of Dn be {xl(n). i= 1 C. and vice versa.n = I:pi(n)Er v ( __ ) n) ) a= 1 . Now we can find first a sequence {Dm} of distributions with finite support such that D.. any distribution B on (0.2) } to go on (on E(2)) when { i 1) } exits. oo). oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.(bk)'. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n..356 APPENDIX For U1 V U2. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . there is a sequence {B. Example A5. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). Then we must find phasetype distributions Bn with B.} of phasetype distributions such that Bn 3 B as n + oo. 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 . Proof Assume first that B is a onepoint distribution. That is.B(bk) I < 1/n for n > k. and the closedness of the class of phasetype distributions under the formation of finite mixtures.(n) = D. relies more on matrix algebra than the probabilistic interpretation exploited here). By the diagonal argument (subsequent thinnings).(bk) + B(bk) for all k. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. r # oo. elementary) Let {bk} be any dense sequence of continuity points for B(x). Hence it is immediate that Bn 4 B. with weight pi(n) for xi(n)..14 To a given distribution B on (0..8. say degenerate at b. the initial vector is (a(1) (& a (2) 0 0). q(n) q(n) pi(n)a . we can assume that ID.
15 To a given distribution B on (0 .. the class CO of all discrete distributions.(x)Bf. u 2 (abstract topological ) The essence of the argument above is that the closure (w. replications). 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.i. Since PET is closed under the continuous operation of formation of finite mixtures. oo) * [0. Corollary A5. Let E be the class of functions f : [0.. oo). oo) such that f (x) = O(e«x).14 is fundamental and can motivate phasetype assumptions. i. then it is immediate that WI(B) = p2(B) for all distributions B on [0.t. Then ICr( n ). But To is the class G of all distributions on [0. x 4 oo. In particular.. however. If Cpl (B) and ^02(B) are weakly continuous.n...e.r. E E. and that cp is known to be continuous. f2. one would use the B given by some statistical fitting procedure (see below). 2. PIT contains all finite mixtures of onepoint distributions. oo) approximation Assume that we can compute a functional W(B) when B is phasetype.d. say on the claim size distribution B in risk theory.. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. For a general Bo. . k < n. i = 1. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. Hence G C PET and L = PIT.n( b k ) . u Theorem A5.n (bk) . if information on Bo is given in terms of observations (i. we can then approximate Bo by a phasetype B. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. oo) and any fl. compute W(B) and use this quantity as an approximation to cp(B0).B(bk )I < .. and we can take Bn = Cr(n). for some a < oo..D(bk)I < n.( dx) * f r f{(x)B(dx). It should be noted.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). k < n.
\\ 0 Corollary A5.16 To a given distribution B on (0 ... 2. n. then cc f (x)Bn ( dx) = (?!c ) e'= . .f (x)B(dx)..38 ).2 .. f00 fi(x)Cr. . and hence we may choose r(n) such that L 9l) f (x)Cr(n). . oo). Now returning to the proof of (A.oo J fi(x)B. TO (A.n(dx) < 1+.. liminf B. i=1... n.. .39) Indeed. for each i.. f° xtBn(dx ) * f °° x`B( dx)..(dx) > J fi(x)B(dx).358 Proof By Fatou' s lemma. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. i = 1.. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.39).f (z) = f = 1 1 1 1n/ o . .38) We first show that for each f E E.. i = 1.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). and the case of a general f then follows from the definition of the class E and a uniform integrability argument. we may assume that in the proof of Theorem A5. if f (x ) = e°x.  APPENDIX B implies that 00 o o 00 n. i = 1.f ' f (x)B(dx).n(dx) + f 0 fi(x)Dn(dx). n B=az. By (A.
0 as i * oo. there is a sequence {B. The present section is a survey of some of the available approaches and software for inplementing this. ./3) is defined as the unique solution > 0 of B[y] = l+y/j3.. . For practical purposes. oo) with B[y +e] < oo for some e > y = 7(B. the remaining results may be slightly stronger than those given in the literature. O We state without proof the following result: Corollary A5. /3) = ry for all n.14 is classical.17 To a given /3 > 0 and a given distribution B on (0. . lim inf > is proved similarly. Notes and references Theorem A5. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .e. e ) and ei J. The adjustment coefficient is a fundamental quantity. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. If ei > 0. . . (N or a given distribution Bo. (N. . and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. 5e Phasetype fitting As has been mentioned a number of times already..l3µb < 1.3).16. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i .. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. I./3). However. the adjustment coefficient 'y = 7(B. from a more conceptual .18 In the setting of Corollary A5. Proof Let fi(x) = el'r+E.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.> y for some sequence {ei} with ei E (0. but are certainly not unexpected. one can obtain 7(Bn. lim sup ryn < 7.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . .
three for a mixture of two Erlangs ). and as fitted distribution we may take B.. .g. Schmickler (the MEDA package.'s). The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. and in practice this sets a limitation to the usefulness (the curse of dimensionality . Of course. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . at a a number of selected points . we have constructed a sequence { B. In a series of papers (e.d. [317] ) has considered an extension of this setup.g . d. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. . we do not not want to perform matrix calculus in hundreds or thousands dimensions). B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. the L1 distance between the c .360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . cf. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be).g. risk theory. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm .. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . The characteristics of all of these methods is that even the number of parameters may be low (e. e . [216] ). g. where more than two Erlangs are allowed and in addition to the exact matching of the first three moments a more general deviation measure is minimized (e.} of phasetype distribution such that Bo. giving mass 1 /N to each S=. a program package written in C for the SUN workstation or the PC is available as shareware. [70]) restrict attention to acyclic phase type distributions .. (N is the empirical distribution Be. A method developed by Bobbio and coworkers (see e.f. It seems therefore a key issue to develop methods allowing for a more general phase diagram. [202]. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments .f.g. reliability or queueing theory. and used a nonlinear programming approach . for some suitable large n. defined by the absence of loops in the phase diagram . 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. the number of phases required for a good fit will typically be much larger..
eieT(n)((k. since this is parameterdependent.APPENDIX 361 tion Sk was available...T (n)(TiI(1. E. the methods of [70] and [38] appear to produce almost identical results. . then the estimators would be of simple occurenceexposure type. In practice. = j) f k=1 k =1 tE[0. (N) tJk Ea ( n). .. EN where ai = N 1 I ((k) = i) tii=i iEE... one is lead to an iterative scheme. (n+1) _ Ea (n). e.g...(N) = E Ea(n).. it seems open whether the restriction to the acyclic case is a severe loss of generality. In fact. N Ti = I(J= i) dt.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . Thus. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.T(n) (Ti ^^ 1. Nii = = . .. it is easy to see that N (k Ea(n). (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). . jEEA. .g.T(n) (Nik IC1.
This page is intentionally left blank .
[10] K. Alsmeyer (1991) Erneuerungstheorie. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Probab. Dover.G. [13] S. 20. Queueing Systems 5 . [3] J. Abate & W. B. [11) S. Scand. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. [7] E. Arfwedson ( 1954) Research in collective risk theory. Adv. Preprint. 219229. Asmussen (1984) Approximations for the probability of ruin within finite time . Asmussen (1985) Conjugate processes and the simulation of ruin problems. The case of equal risk sums. [14] S.). AT&T. 3157 . [5] G. with applications to risk reserve processes and the GI /G/1 queue. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 101116. New York.L. Aktuar Tidskr. New York. Borokov. Abate. John Wiley & Sons. ed. pp. G. Abramowitz & I. 143170. 38. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Proc. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Transactions XVth International Congress of Actuaries. 1985. Teubner.Bibliography [1] J. [2] J. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . 311338. ibid. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. 587. Stoch. 57. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . New York. Appl.). Optimizations Software. Asmussen (1987) Applied Probability and Queues. [15] S. [6] V. Abate & W. Skand. Queueing Systems 10. Queueing Systems 16. 253267. J. Choudhury & W. Skand. The Mathematical Scientist 14. [9] G. Aktuar Tidskr. Stuttgart. [12] S. II. Arfwedson ( 1955) Research in collective risk theory. In: Limit Theorems and Related Problems (A. [4] M. 213229. 1984 . [8] G. 53100. Chichester New York.A. Act. 363 . 345368. 191223. 14. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Appl. 37.
Asmussen & M. 2. [19] S. Act. Asmussen (1995) Stationary distributions via first passage times. Probab. Astin Bulletin 27. Proc. Boca Raton. Ann. Extremes 1 . Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. 137168. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Probabilistic Analysis of Rare Events (V. Andronov. Stoch. 555574. [23] S. Act. Dshalalow ed.M. 1996.K. 1936. Alfa & S. New York. Hojgaard (2000) Rare events simulation for heavytailed distributions. Ann. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. 2943. [18] S. 8. Asmussen & B.). Asmussen. eds. Methods V Problems (J. 315. T. Appl. Asmussen (1999) On the ruin problem for some adapted premium rules. 193226. [26] S. 4966. Asmussen. Riga Aviation University. Scand. 37. Asmussen & M. Statist. 69100. [21] S. J. Asmussen. 20. [33] S. 106119. J. [17] S. 96107. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Marcel Dekker. [31] S. 297318.). 27.). 1989 . Asmussen (1992b) Light traffic equivalence in single server queues. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Random Processes 2. 79102. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. 313326. Scand. . 2149. Appl. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Asmussen & K. Frey. Stoch. Chakravarty. J. [32] S. Bernoulli 6. Asmussen (1998c) A probabilistic look at the WienerHopf equation. Th. Asmussen & B. 1999 . [28] S. Scand.S. L. K.364 BIBLIOGRAPHY [16] S. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Advances in Queueing: Models. 313341. [27] S. 303322. [25] S. 54. 189201. [20] S. [29] S. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Probab. [34] S. A. Florida. MatrixAnalytic Methods in Stochastic Models (A. 354374. Act. eds. Kalashnikov & A. SIAM Review 40. Stochastic Models 11. Appl. Binswanger & B. Proc. 25. Ann. Asmussen (2000) Matrixanalytic models and their analysis. [24] S. J. 772789. Rolski & V. Probab. stationary distributions and first passage probabilities. Appl. Floe Henriksen & C. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Asmussen (1989b) Risk theory in a Markovian environment. Scand. [30] S. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. CRC Press. [22] S.
Prob 32 . [44] S. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. J. Asmussen & C. Asmussen . [43] S. Probab. Taksar (2000) Optimal risk control and dividend distribution policies . 11251141. Asmussen. 913916. Methods 8 Problems (J.). 105119. Probab. 19. (to appear). [49] S.Y. Asmussen. Example of excessoffloss reinsurance for an insurance corporation. 20. Opns. Schmidli & V. 120. 64. Florida. Marked point processes as limits of Markovian arrival streams. Math. Asmussen & G. Appl. 31. 259274. Asmussen & R. Asmussen & H. [51] S. Olsson (1996). Stochastic Models 8. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Management Science 45.M. Read eds. 58. Johnson. Schmidt (1995) Ladder height distributions with marks. Stoch. Asmussen & R. 410433. Proc. [38] S. Nerman & M. Kliippelberg (1996) Large deviations results for subexponential tails.). Probab. Statist. [50] S. Adv.Y. 429466. 103125. Res. Sc. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Hojgaard & M. Rubinstein (1999) Sensitivity analysis of insurance risk models. J. 422447.BIBLIOGRAPHY 365 [35] S. 421458. Asmussen & M. 299324. Advances in Queueing: Models. B. Asmussen & S. Th. CRC Press. Oper. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Wiley. Asmussen & T. [52] S. Perry (1992) On cycle maxima. Appl. Asmussen & V.A. Asmussen & T. Asmussen & C. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Res. Appl. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. H. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. 365372. . [42] S. Insurance: Mathematics and Economics 20. Sigman (1996) Monotone stochastic recursions and their duals. first passage problems and extreme value theory for queues. Asmussen & D. 10. 23. [39] S. 419441. Inf. Finance and Stochastics 4.A. Appl. Boca Raton. Appl. Adv. Scand. [46] S. [40] S. Dshalalow ed. Koole (1993). [37] S. [48] S. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & C. Proc. 736755. Insurance: Mathematics and Economics 10. O. with applications to insurance risk Stoch. Statistica Neerlandica 47. Eng. [47] S. Fitting phasetype distributions via the EM algorithm. Asmussen & V. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. [45] S. 115. 30. [41] S. Supplementary Volume (Kotz. Appl. Probab. J. [36] S. Asmussen & K.
Billingsley (1968) Convergence of Probability Measures. Berlin. Trans. New York. 190204. J. 221232.L. Teugels (1987) Regular Variation. J. [66] N. 1985 . 705766.H. 9599. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. [60] J.H. [72] P. Insurance: Mathematics and Economics 9. [64] P. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. [56] B. 4148. [57] C. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist.H. [65] N. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. New York. C. Bingham (1975) Fluctuation theory in continuous time. [59] O. Scand.R.J. [55] B. Act. Wiley. 33. [61] J. [62] J. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Probab. Berman & R. Bobbio & M. Appl. Boogaert & A. Scand. SpringerVerlag. de Waegenaere (1990) Simulation of ruin probabilities. Stochastic Models 10. Scand. [73] A. Asmussen & J. Bjork & J. Beekman (1969) A ruin function approximation. Cambridge University Press. [71] P. Goldie & J. 7.L. 129134. Halsted Press. Bjork & J. J. Soc. Scand. New York. J. Bingham. Grandell (1985) An insensitivity property of the ruin probability. Borovkov (1976) Asymptotic Methods in Queueing Theory. Bloomfield & D. [54] K. [67] T. Wiley. Act. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Ney ( 1972) Branching Processes . Springer . Probab. 169186. . 1975. 9.T. Probab. Cambridge.A. Athreya & P. 1988 . Cox (1972) A low traffic approximation for queues.B . [70] A. Boogaert & V.366 BIBLIOGRAPHY [53] S. Adv. SIAM. J. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. [58] O. BarndorffNielsen & H. Act. [69] P. [68] T. 11811190. 610. Beekman (1974) Two Stochastic Processes. Act. [63] A. Oxford. 275279. Beekman (1985) A series for infinite time ruin probabilities. 1995. 1974. J. Scand. Baker (1977) The Numerical Solution of Integral Equations. Actuaries 21.A. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Appl. Clarendon Press. J. Insurance: Mathematics and Economics 4. 148156. 661667. Appl. 832840. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. Act. 77111. Insurance: Mathematics and Economics 6.M.
Burman & D . Queueing Systems 33. Cramer (1930) On the Mathematical Theory of Risk.C. 177204. Croux & N. [75] O. Cambr. Jones & C. Boxma & J. Burman & D. Area Commun. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic.M.Y . [76] N. New York San Francisco London. Tech. H. . [89] M. [91] H. 16. Cohen (1982) The Single Server Queue (2nd ed. [93] K.). [86] K. Skandia Jubilee Volume. [82] H. The Jubilee volume of Forsakringsbolaget Skandia. 14.. Cox (1955) Use of complex probabilities in the theory of stochastic processes. 24. Stockholm. Breiman ( 1968) Probability. Nesbitt (1986) Actuarial Mathematics . [80] F. verve. [81] J.L. [90] D. Resnick & R. AddisonWesley. Z.). Qinlar (1972) Markov additive processes. Academic Press. S. J. Jr. 93121. Cramer (1955) Collective risk theory. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Philos. Boxma & J. Oper. Cottrell. Bux & U. [77] P. 51. Syst. Aut. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . [88] J. Goovaerts & F.A. Reiser eds. Res. Brockwell . 907920. Chandy & M. 14331453. Bucklew ( 1990) Large Deviation Techniques in Decision. Proc. Gerber. [87] E.R.) NorthHolland. Tweedie (1982) Storage processes with general release rule and additive inputs . [78] L. 127130. Bowers .L. 392433. Adv. Probab. J.I.A. Geb. Amsterdam. [79] P. [92] H. Fox & L. Appl. Insurance : Mathematics and Economics 5. Chung (1974) A Course in Probability Theory (2nd ed.R.L. [85] W. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Berlin. Wahrscheinlichkeitsth. Bell.C. M. 62.U.J. New York.L. NorthHolland. Computer Performance (K. 2338.W. J. Nonparametric estimators for the probability of ruin. 313319. Sel. 105119. Bratley. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . B. Reading. Simulation and Estimation. Stockholm. Hickman. II. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Broeck . Soc. Fort & G. Control AC28.Y. IEEE J.R.J.BIBLIOGRAPHY 367 [74] O. Amsterdam. Herzog (1977) The phase concept : approximations of measured data and performance analysis. D. Veraverbeke (1990). The Society of Actuaries . Springer.W. De Vylder (1986) Ordering of risks and ruin probabilities. [83] D. 34. [84] D. Wiley. IEEE Trans. Insurance : Mathematics and Economics 9. Itasca. New York. SpringerVerlag. Illinois. 749763. Schrage (1987) A Guide to Simulation.W. 3539.
Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Delbaen & J.. Roy. Insurance: Mathematics and Economics 10. 5771.D. Insurance: Mathematics and Economics 3. Res. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Springer. 85116. Rolski (1984) A light traffic approximation for a single server queue. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. 114119. 1990 . Csorgo & J. J.P.. Suppl. Daykin. 3750. 201206. [100] A. Scand. Haezendonck (1985) Inversed martingales in risk theory. 121131. 27.M. Laird & D. De Vylder (1996) Advanced Risk Theory. [111] F. Deheuvels & J. Stockholm. Zeitouni (1992) Large Deviations Techniques and Applications. English version published in Skand. 9 . 624628. Math. N. [112] F. Haezendonck (1987) Classical risk theory in an economic environment. 181217. [102] P. [96] D. [107] L. [108] F. Oper. [105] A. Insurance: Mathematics and Economics 6. Probab. [95] M. 135159. Embrechts (1989). van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. [104] F. [106] A. De Vylder & M. Stochastic Models 5. New York. Appl. Act. Daley & T. [103] F. Csorgo & J. Martingales and insurance risk. 16. Tidskr. J. [99] R. De Vylder (1977) A new proof of a known result in risk theory. Dembo & O. . J. Appl. Goovaerts (1984) Bounds for classical ruin probabilities. [98] A. Dempster. [109] F. 17. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Pentikainen & E. Delbaen & J. Soc. De Vylder & M. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . [101] C. J. [97] D. Math. 7083 (1969). Editions de 1'Universite de Bruxelles. Aktuar. 3. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. T. Actuarial J. Jones and Bartlett. 88101. 435436. Boston.B. Res. Insurance: Mathematics and Economics 7. Insurance: Mathematics and Economics 4. 583602. Dassios & P. Statist. Chapman & Hall. Oper. Math. [110] F. Devroye (1986) Nonuniform Random Variate Generation. Pesonen (1994) Practical Risk Theory for Actuaries.368 BIBLIOGRAPHY [94] M. Rolski (1991) Light traffic approximations in queues. Comp. Scand. Daley & T. 22. 227229.L.
M. 1993 . Scand. Dickson (1992) On the distribution of the surplus prior to ruin . Shiu (1991) Risk theory with the Gamma process. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Dufresne & H.U. and the amount of claim causing ruin. J.R. Act. Scand. Dynkin (1965) Markov Processes I. 6180. with applications. Scand. Act. [129] D . Insurance: Mathematics and Economics 25. Act.R. Dickson & J. 1. Hipp (1999) Ruin problems for phasetype(2) risk processes. J. Gerber & E. [120] D. Dufresne & H. 363374. Insurance : Mathematics and Economics 11. [127] F. Act. 5159. Harrison & A. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Waters (1996 ) Reinsurance and ruin . Regterschot (1988) Conditional PASTA. 1993 . 251262. J. Philos. Astin Bulletin 21. [130] P.C.W. [121] D.M. [114] D. 1975 .U. [131] P.C.C. [124] N. Letters 7.C.M. Dickson (1995) A review of Panjer's recursion formula and its applications. Ruin estimates for large claims . 118. J. Insurance: Mathematics and Economics 7. 191207. Dickson (1994) An upper bound for the probability of ultimate ruin. Duffield & N. H. [128] E.M.C. 4962. 174186. Embrechts ( 1988 ).C. 105115. Embrechts . J. [116] D. [123] E.M. [125] F. 1994 . Dufresne. Springer . Act. [119] D.J.M. 2000 .M. Berlin Gottingen Heidelberg. Waters (1999) Ruin probabilities wih compounding.C.M. Dickson & C. 147167. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. .M. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . Emanuel . 131138.S. Dickson & J.G. Insurance : Mathematics and Economics 7.C. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. van Doorn & G. 1741.U. Math. [118] D. Dickson & H. Camb. 4259. Djehiche (1993) A large deviation estimate for ruin probabilities. [122] B. Insurance: Mathematics and Economics 19.K. 193199. J. Soc. Act.J. Insurance : Mathematics and Economics 22.C. Oper. Proc. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. 1984 . 107124. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. 177192. Scand. J. Dickson & H. 1984 . 3745.R. J.B. Res. Scand. Act.R. J. Insurance : Mathematics and Economics 10. [115] D. [126] F.BIBLIOGRAPHY 369 [113] D. Scand. Dickson & C. 269274. British Actuarial J. Scand. Gerber (1988) The surpluses immediately before and at ruin. 229232. Grandell & H. [117] D.C.M.
Statistica Neerlandica 47. Villasenor (1988) Ruin estimates for large claims. Insurance: Mathematics and Economics 7. Jensen. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks.K. Appl. Arndt & V. [139] A. Fuh & T. Act. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Schmidli (1994) Ruin estimation for a general insurance risk model. [144] P. Ann. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. 8190. Embrechts & T. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Frenz & V. Embrechts. R. A bootstrap procedure for estimating the adjustment coefficient. Probab. [134] P. Fields 75. [150] H. 181190. Embrechts.K. Nyt Tidsskrift for Matematik B20. Approximations for compound Poisson and Polya processes.M. Probab. Maejima & J. Wiley. Trans.370 BIBLIOGRAPHY [132] P. [146] E. J. [145] P.W.L. Insurance: Mathematics and Economics 10. [133] P.D. 17. Mikosch (1991). Probab.A. Proschan & D. New York. Embrechts & J. Konig. first passage times and moments of ladder variables in Markov random walks. Skand. [137] P. [147] M. Heidelberg.D. Embrechts & H. Tidsskr. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A.W. 695712. [141] F. Astin Bulletin 16. J. Adv. Adv. 269274. 38. C. Frees (1986) Nonparametric estimation of the probability of ruin.J. 131137. 29. 3339. Th. J. [148] C. Appl. New York. Insurance : Mathematics and Economics 1. Embrechts & N. 159178. On the excursions of Markov processes in classical duality. Teugels (1985). Math. Statist. 616624. [138] P. Kliippelberg & T. Wiley. 175195. 5974. Furrer (1998) Risk processes perturbed by astable Levy motion. 1998 .L. 35. Franken. [136] P. Appl. D. Embrechts. Erlang 2. Lai (1998) Wald's equations. Appl. Sci. 566580. Probab.). [140] J. Fitzsimmons (1987). 26. Springer. Danish Academy Tech. Griibel & S. with applications. Adv. 623637. 14661474. 404422. Rel.). Scand. Esary. Schmidt (1982) Queues and Point Processes. 5975. Esscher (1932) On the probability function in the collective theory of risk. J. [149] C. [142] W. [143] W. Wiley. 5572. F.D. Walkup (1967) Association of random variables. Probab. M. 29. . Probab. U. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [135] P. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Akt. Appl. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Schmidt (1992) An insensitivity property of ladder height distributions.
[164] B. 3752.J. Act. Vers. Adv.U. Grandell (1977) A class of approximations of ruin probabilities.V. NorthHolland. London.U. Skand. Scand. Schweiz. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Boston. Birkhauser .71. Probab. F. Math. J. Basel. Z. [152] H. Scand. Appl. 31A. 105115. [163] P. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. [165] M. 2336.E. J. 7778.N. Kaas. J. Gerber (1988) Mathematical fun with ruin theory. Furrer & H.BIBLIOGRAPHY 371 [151] H. Schweiz. [156] H. Springer..S. 28. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. [157] H. Goldie & R. Michna & A. 1977. S. Gerber (1979) An Introduction to Mathematical Risk Theory. 6370. [166] M.U. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. In Studies in Applied Probability (J. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Math. Aktuarietidskrift 1970 . Ann. 97114. . Haezendonck (1984) Insurance Premiums. Goovaerts. 1978 .G.U. University of Pennsylvania.).M. 5. [158] H. 463480. Huebner Foundation Monographs. J. Amsterdam. 1981 . Ver. Probab. Furrer. Kou (1995) Limits of first passage times to rare sets in regenerative processes. NorthHolland. Mitt. Dordrecht. Insurance: Mathematics and Economics 20. [162] P. Life Insurance Mathematics. R. 205216. Gerber (1973) Martingales in risk theory. Galambos & J. [155] H. Ver. Appl. Probab. [168] J. Insurance : Mathematics and Economics 7. [169] J. Gani. contained in the authors PhD thesis. Scand. Vers. ETH Zurich. [159] H. Unpublished. [154] H.U. Bauwelinckx (1990) Effective Actuarial Methods. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Suppl. 1523. 73. Kluwer. Mitt. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier.. 424445. [153] H.). Griibel (1996) Perpetuities with thin tails. [160] H.J. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Gnedenko & I. Amsterdam. de Vylder & J. van Heerwarden & T. A. Glasserman & S. Act. 205210. Gerber (1986). Weron (1997) Stable Levy motion approximation in collective risk theory. Goovaerts.W.U.U. Act. Insurance: Mathematics and Economics 15. Glynn & W. Appl. 131156. [161] P. eds. [167] C. Glasserman (1991) Gradient Estimation via Perturbation Analysis.
Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . M. 347358. 7590. Ibragimov. 566576. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. 6779. 332. Proc . Statistical Algorithm 265. [174] J. Opns .J. [181] D. Insurance: Mathematics and Economics 26. Applied Statistics 40. Griibel & R. Probability Theory and Mathematical Statistics (I. Harrison & A. Harrison (1977) Ruin problems with compounding assets . Stoch. Chapman & Hall. Amsterdam. Grandell (1979) Empirical bounds for ruin probabilities. [179] R. [178] B. [176] B.S.. Gyllenberg & D. M. [172] J. Informatica 2. Appl. 33. Res. 3041. Lemoine (1977) Limit theorems for periodic queues. Math. Grandell (1990) Aspects of Risk Theory. . Proc. KTH.and Sozialforschung 6. [185] J. Grandell (1999) Simple approximations of ruin functions. Proc. SpringerVerlag. Gusak & V.I. Aktuar Tidskr. 8.O.I. [173] J. Probab.243255.A. Theory and Applications. Grnbel (1991) G/G/1 via FFT.V. New York. 14. [184] H . Rink. [183] M. [177] B. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Liet. 5 . Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . [186] J. Archiv v. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Grigelionis (1996) Lundbergtype stochastic processes. ed. Math. Th. [189] A. 355365. Segerdahl (1971) A comparison of some approximations of ruin probabilities. AkademieVerlag.131135. Resnick (1977) The recurrence classification of risk and storage processes . Tinbergen Institute Research Series 20. 5766. Grandell (1997) Mixed Poisson Processes. Appl. Appl.372 BIBLIOGRAPHY [170] J. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Probab. Appl. M. 3 . Matem. [180] R. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . 14. [171] J. [182] A. Grandell & C. Gut (1988 ) Stopped Random Walks .). 400409. M. Harrison & S. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. E. 167176. 143158. 1. Skand. 54. [188] J. Grandell ( 1992) Finite time ruin probabilities and martingales . Res. Opns . Astin Bulletin 29. [187] J. [175] J. mathematische Wirtschaft . Berlin. 197214. Preprint. Winter School on Stochastic Analysis and Appl . Stoch. J. SpringerVerlag. Harrison & S.
Klugman (1984) Loss Distributions. Z. Foruth Prague Symp. 305313. [200] M. [198] C. J. 1998 . Taksar (2000) Optimal proportional reinsurance policies for diffusion models. on Asymptotic Statistics (P. Hogg & S. Horvath & E. Proc. J. Scand. [196] C. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities.. Nerman (1992) EMPHT . J.L. [201] L.A. [207] D. Geb. Insurance: Mathematics and Economics 5. [195] B. Math. Verlag Versicherungswirtschaft e. 19. [208] V. 25.L. Ann. 8996. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Insurance: Mathematics and Economics 5. .). 11631174. [194] 0. 259268. 8095. [197] C. ACM TOMACS 6. Electronic Letters 35. Hoglund (1991). 1990 . Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Mandl & M. [192] U.B. S. 377381. 6. Statist. [204] T. 123151. [193] U. Karlsruhe. 378389. [191] W. Probab. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Hipp (1989a) Efficient estimators for ruin probabilities. Huskova. Probab. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Studies in Statistical Quality Control and Reliability 1992 :4.M. Asmussen & 0.BIBLIOGRAPHY 373 [190] P. Iglehart (1969) Diffusion approximations in collective risk theory. 23. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. [199] R. [202] 0. The ruin problem for finite Markov chains. 29. Hipp & R. [205] T. [203] T. Ann. Astin Bulletin 19. [206] B. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. 285292.V. 285293. 5770. No.V. Haggstrom. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Act. 12981310. eds. 166180. verve. Karlsruhe. Iversen & L. Hill (1975) A simple general approach to inference about the tail of a distribution. Versicherungswirtschaft. Act. Heilmann (1987) Grundbegriffe der Risikotheorie. Scand. Ann. Appl. Mathematical Statistics. Chalmers University of Technology and the University of Goteborg. Nachrichten 30. Wahrscheinlichkeitsth.a program for fitting phasetype distributions. Hojgaard & M. J. Probab. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. 18. Wiley. New York. Appl. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. 3. Prob.R. 4385.
M. J. Astin Bull. 31. Statist. 15.G. J. J. Goovaerts (1986) General bound on ruin probabilities.L. New York. Kemperman (1961) The Passage Problem for a Markov Chain. Act. Proc. 63.374 BIBLIOGRAPHY [209] D. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues.H. [221] E. Probab. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [218] V. Jensen (1953) A Distribution Model Applicable to Economics. 6. 37. 11. Stochastic Models 5. Appl. Taaffe (1989/90) Matching moments to phase distributions. Soc. 325247. Soc.M. Appl. [216] M.C. Chicago. Cambridge Philos. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. 118. Cambridge Philos. [227] J.G.G. Proc. Amer. Act. [213] P.R. 173190. Taylor (1981) A Second Course in Stochastic Processes.A. [224] J. 123133. ibid. Kalashnikov (1996) Twosided bounds of ruin probabilities. [214] A.B. Addenda to for processes defined on a finite Markov chain.J. Princeton. [212] J. University of Chicago Press. Technometrics 30. 116129. Math. [210] D. 3. Wishart (1964). 283306. Astin Bull. [223] J. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Keilson & D. NEC Research Institute. Munksgaard. Soc. Ann.M. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. 561563. Proc. [215] J. Kao (1988) Computing the phasetype renewal and related functions.M. [220] V. Clarendon Press. [217] R. N. Insurance: Mathematics and Economics 5. Oxford. [219] V. ibid. Keilson & D. 259281. Jensen (1995) Saddle Point Approximations. 711743. [228] J. Keilson & D. [222] S. Academic Press. Scand. Stochastic Models 1. 60. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. . 6. 547567. 8793. 35. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Manuscript. Copenhagen. Karlin & H. 165167. [211] J. 1996. N. Kluwer. J. Kaas & M. Janssen & J. 61. 239256. Probab. [226] J. Johnson & M. Cambridge Philos.M. [225] J. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. 866870. Jagerman (1985) Certain Volterra integral equations arising in queueing. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Jelenkovic & A.J. 187193.P. 4151.
I.C. Klnppelberg & T. Kingman (1964) A martingale inequality in the theory of queues. Lecture Notes in Mathematics 511. [236] C. Soc. Insurance : Mathematics and Economics 12. Kunita (1976) Absolute continuity of Markov processes. 24. Klnppelberg & T. Oxford 12. 359361. 146. 133135 (in Russian). SpringerVerlag. B24. SpringerVerlag. Macmillan. Probab. Lemoine (1981) On queues with periodic Poisson input. [234] C. Scand. Latouche & V. Seminaire de Probabilties X . 858874.F. [240] U. [246] D. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. . Appl. Act. [245] T. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. 60 . Probab. Kingman (1961) A convexity property of positive matrices .J. Insurance: Mathematics and Economics 8. J. J. 26. [238] V. Camb. J. [237] C. Lehtonen & H. Prob.P. 18. New York. Stochastic Models 7. 132141. 283284. J. Bernoulli 1 . 154168. 889900. Lehtonen & H. Kingman (1962) On queues in heavy traffic. 4477. Philos. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . [241] G. 25. Soc. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process.J. Soc. Kuchler & M. Kliippelberg & U. Scand. Statist. 48.C. Korolyuk.S. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment.F. Proc. [248] D. 390397. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. 277289. Lindley (1952) The theory of queues with a single server.C. (235] C.a Linear Algebraic Approach. [242] A. [243] A. Act. [247] L. Appl. Scand. Proc. 6075. Roy. Cambr. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Cybernetica 4. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. [232] C. 125147. J.BIBLIOGRAPHY 375 [229] J. Kliippelberg (1988) Subexponential distributions and integrated tails. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. Sorensen (1997) Exponential Families of Stochastic Processes. Adv. Actuarial J. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. [239] H. Quart. [230] J. 259264. [231] J. J. 1998 . Appl. 279285. 383392. [233] C.F. 4958. [244] T. Philos. Quart. 1995 . SIAM. Mikosch (1995) Modelling delay in claim settlement. Math. J.F. Probab. Penev & A.. Lipsky (1992) Queueing Theory . Appl.
[259] Z . Appl. Ney & E. J. . Europ. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . [255] V. 58 .D. 36. [256] A.K. Appl. Math. 834841. 22. Michna ( 1998) Selfsimilar processes in collective risk theory. 2.F. J. 124147.V. 15. 763776.V. 1996 . Stoch. Scand. MartinL6f (1983) Entropy estimates for ruin probabilities . Makowski ( 1994) On an elementary characterization of the increasing convex order . [257] A. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . 1986 . Act. 2939. Soc. [253] V. with an application . Miyazawa & V. 676705. Ann. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. [260] H. 147149. Appl. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes .K. 429448. Statist. Ann. J. 286298. Anal. Schmidt ( 1993) On ladder height distributions of general risk processes . Res. 561592. [250] F. 3. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Eigenvalue properties and limit theorems . 58. Stockholm. [251] F. 161174. [252] A. Nagaev (1957) Some limit theorems for stationary Markov chains. 32. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Gut & J. Insurance : Mathematics and Economics 5. Opns.). Soc. [265] S. J. .S. 223235. Hoist eds. F. MeierHellstern & M . Proc. Almqvist & Wiksell. Act. K. Probab. Cambridge Philos. J. Appl.M. [254] V . Appl. Cambridge Philos.376 BIBLIOGRAPHY [249] D. [258] K . [264] G.D. Ann. 1994. Th. 4859. Lucantoni . 370377. 1261 (0?)1270. 31. II Aterforsdkring av Kollektivrisker . Probability and Mathematical Statistics (A. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Probab. J. 11 . J. Probab. a useful concept in risk theory. [266] P. 29 . Englunds Boktryckeri AB. Scand. Nummelin (1987) Markov additive processes I. Proc. Act. Math. Adv. 268285. Appl. [261] H. [262] H . Probab. Meier ( 1984). Probab . [263] M . Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . MartinL3f (1986) Entropy.S. Probab. 378406. Scand. Lundberg (1926) Forsdkringsteknisk Riskutjdmning.D. Uppsala.
[282] J . 12551265. London. J. Hamburg. Proc.A. [280] A.G.K. 339353. 316. E75B . [278] E. Appl. Statistics 21. [271] M. Models 6.BIBLIOGRAPHY 377 [267] M . Appl. Marcel Dekker. 71. [275] H. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. 30. 157. Appl. Omey & E. [273] R. New York. Probab. Appl. Appl. [277] J. Probab. Neuts (1977) A versatile Markovian point process. Stoch. Paulsen & H. 36. Astin Bulletin 5 . 764779. [274] H . 311342. Proc.F. . Paulsen (1993) Risk theory in a stochastic economic environment. Abh. Nyrhinen (1999) Large deviations for the time of ruin. Appl. [283] J.a survey. [281] J. Pakes (1975) On the tail of waiting time distributions. Stoch. 327361. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Math. Paulsen & H. Omey & E. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . [268] M . Stoch. 733746. Probab. J.F. Insurance: Mathematics and Economics 22. Appl. Appl. Commun. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. 75. 12. [269] M . 10081026. 249266. 273299. Johns Hopkins University Press. Stoch.F.F. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. [270] M . [279] E. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. [276] C. J. Proc. Adv. [284] J. Insurance: Mathematics and Economics 20.K. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. 215223. Stochastic Models 3 . Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. 16. Stoch. 123144. Baltimore . 135148. Neuts (1992) Models based on the Markovian arrival process. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Sem. 445454. [285] J. 555564. Norberg ( 1990) Risk theory and its statistics environment . 21. Paulsen (1998) Ruin theory with compounding assets . IEICE Trans. [272] J . Neuts ( 1978) Renewal processes of phase type.F. Proc. Probab. Naval Research Logistics Quarterly 25. O'Cinneide (1990) Characterization of phasetype distributions. 46.
Philipson (1968) A review of the collective theory of risk.H. 147159.U.378 BIBLIOGRAPHY [286] J. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. New York. 390413.U.J. and Dams. Prabhu (1961). Pyke (1959) The supremum and infimum of the Poisson process. 757764. Appl. Prabhu (1980) Stochastic Storage Processes. . [301] G. 12. Pitts. Aktuar. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. 235243. Stat. Insurance: Mathematics and Economics 3. [297] R. Scand. Math. [289] C. Austr. [294] N. 19. New York. 28. [306] T.K. Regterschot & J. Ann. 29. J. J. Ramaswami (1980) The N/G/1 queue and its detailed analysis. 11.M. Inst.K.G. Springer. 4568. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Res. Math. Math. Adv. Probab. 2343. [292] S. 61. Pellerey (1995) On the preservation of some orderings of risks under convolution. Astin Bulletin XIV. On the ruin problem of collective risk theory. J.U. [304] B. Probab . Pitman (1980) Subexponential distribution functions. [300] C. [295] N. 32. Oper. [293] N. Queues. 4. [305] C. 139143. Rolski (1987) Approximation of periodic queues. Statist. 820827. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. 29A. Probab. 465483. [303] S. 537555. 222261. [302] J. 117133.R. 965985. Statist.G. Tidskr. Wiley. Prabhu (1965) Queues and Inventories. Prabhu & Zhu (1989) Markovmodulated queueing systems. [298] V. Rao (1965 ) Linear Statistical Inference and Its Applications. Adv. Res. Paulsen & H. Heidelberg. Adv. Skand. Insurance: Mathematics and Economics 16. 691707. Opns. Act. Appl. Appl. Probab. Resnick & G. [291] S. Wiley. Probab. 568576. [296] N. Ann. 46. Gjessing (1997c) Ruin theory with stochastic return on investments. Ann. Appl. 2330. Ripley (1987) Stochastic Simulation . Soc. Berlin. 337347. 1989 . Math. [287] F. Appl. Ann. [299] C.M.A. Embrechts (1996) Confidence bounds for the adjustment coefficient. 30. [288] S. Griibel & P. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services.L. Wiley. R.. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains.U. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. [290] E. Queueing Systems 5.J. Math. 215246. Insurance Risk.M.
Act. Th.Y. Wiley. Stoch. Rubinstein & A. [322] H. Wiley. [309] H. Rubinstein (1981) Simulation and the Monte Carlo Method. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. 4068. Appl. Schmidli ( 1999b) Perturbed risk processes: a review . Probab . Data Anal. 227244. Scand. Rolski. [321] H. [319] H. Seal ( 1969) The Stochastic Theory of a Risk Business. Statist. 365388. Stochastic Models 10. [326] H. 5. J. Schmidli. [310] R. [324] H .J. 5. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. [313] M. Wiley. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion .Y. J. . Stochastic Models 6.BIBLIOGRAPHY 379 [307] T. 795829. Proc. 417421. 93104. Probab. [323] H. [314] T. [325] H . 7. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Melamed (1998) Modern Simulation and Modelling. [317] L. 11. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Appl. Schmidli ( 1996) Martingales and insurance risk. 431447 [316] S. 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. V. Scand. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Wiley. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest.M. Insurance: Mathematics and Economics 22. 155188. 131156. Wiley. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. 21.L. [315] T. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Rudemo (1973) Point processes generated by transitions of a Markov chain. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Comp. Ann. Appi. 121133. 1997. Math. [308] S.Y. 5. Insurance: Mathematics and Economics 16. [312] R. Statist. Singapore. Operationsforsch. Act. H. Teugels (1999) Stochastic Processes for Insurance and Finance. Rubinstein & B. 135149. Schmidt & J. Probab. 4857. 145165. J. 687699. [320] H. 262286. Astin Bulletin 29. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . [318] H. Stochastic Models 10 . [311] R. Science Culture Technology Publishing . 2001 . Adv.
Th. Mitt.S. . the probability of nonruin in an interval (0. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. Act. Shin (1987) Convolution of uniform distributions and ruin probability. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case.W. 11. Wiley. Appl. Risk teory and the single server queue . Weiss (1995) Large Deviations for Performance Analysis. 673684. [346) D .380 BIBLIOGRAPHY [327] H. [331] G. 159180. [333] C: O. 11. 383413. Seber (1984) Multivariate Observations. [343] D. 701719. Insurance: Mathematics and Economics 8. 21. Seal ( 1972). Probab. Verein Schweiz Versich. Shin (1989) Ruin probability by operational calculus. 483487. Scand. 171178. Shantikumar (1993) Stochastic Orders and Their Applications. . Almqvist & Wiksell. 914924. Probab. Verein Schweiz.L. Appl. Academic Press. Math.the Harald Cramer volume (ed. 279299. 191197.S. Scand. In Probability and statistics .L. 72. Grenander). [339] A. [329] H. 4. Math.A. Queueing Systems 11. [330] H. Seal (1978) Survival Probabilities. Adv. J. Sigman (1992) Light traffic for workload in queues. 157166. New York. [332] C : O. L. Adv. [347] K. Chapman and Hall. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Stochastic Models 6. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. 72. pp. Probab. Chapman & Hall. 2236. [345] D. 61. Ann. t). Sengupta (1990) The semiMarkov queue: theory and applications. Siegmund (1975) The time until ruin in collective risk theory. [340] E.S. SpringerVerlag. 121139. 243249. Versich. Probab . Versich. Verein Schweiz. Stockholm. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Wiley. Shaked & J. J. [337] M. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. 1987. 77100. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . Math. [335] B . Statist. [344] D. [341] E. [336] B. Shwartz & A. Mitteil. 4383. [342] D. Siegmund ( 1985) Sequential Analysis . Skand. 75. Aktuar Tidsskr. Seal (1974) The numerical calculation of U(w. 1955. Segerdahl (1959) A survey of results in the collective theory of risk.W. [334] C: O.F.G. 429442. [348] K .L. [328] H. Aktuar Tidsskr. Act. Mitt. Ann. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . 4. 1974. t). Appl. [338] E.
J.. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Act. Teugels ( 1995 ) Ruin estimates under interest force . Insurance: Mathematics and Economics 16. Act. Daley ed. [359] L. [365] J. Stanford & K. 8594. 2739. 1974. Probab. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D.L. Sundt & J.J. 235254. Adv.C. Proc. Actuarial J. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. [357] B. Act. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . Astin Bulletin 16 . [367] O. Straub (1988) NonLife Insurance Mathematics. Ruin probabilities prepared for numerical calculations. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. 197208.J. [361] G. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Scand. Springer. Scand. [358] R. Insurance: Mathematics and Economics 19. New York. Act.L. Thorin (1982) Probabilities of ruin. Verlag Versicherungswirtschaft e.L. Insurance: Mathematics and Economics 1. Cambridge Philos. Teugels (1982) Estimation of ruin probabilities . Unpublished manuscript. 725741. 149162. Appl. 118. Sundt & W.). 5776.C. Hoesman ( 1989) Moderate. Wiley. Soc 49. Proceeedings of the IEEE 77. [355] B .BIBLIOGRAPHY 381 [349] E.C. 65102. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . [353] E. [362] G. Thorin (1977). Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . .V. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Scand. [356] B . [352] D. 722. J.S.C. 8199. J. Scand. 65102. 1980 . [363] G. Slud & C. [354] B. Taylor (1979) Probability of ruin with variable premium rate. NorthHolland. Sundt & J. 1982 . Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. [369] O. 21. [364] G. Smith ( 1953) Distribution of queueing times. Astin Bulletin 24. 1978.). [366] O.and largedeviations probabilities in actuarial risk theory. Karlsruhe.. [350] W. Astin Bulletin 12. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Scand. Act. 163175.C. [351] D . Scand. 449461. John Wiley & Sons. 114137. J.L. New York. [368] O. J.A. 1976. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Taylor (1986) Claims reserving in NonLife Insurance. [360] G.
1983 . J. [374] F. Wiley. [373] S. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. 142.E. De Vylder & M. De Vylder. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. F. Submitted. [375] N. Insurance : Mathematics and Economics 13. Thorin & N. Aktuar. Goovaerts. Taqqu. Scand.). [378] H. Tacklind (1942) Sur le risque dans les jeux inequitables. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. R. Haezendonck eds. [371] 0. Willmot & X. M. Skand. J. Res. . [379] M. J. [372] H. Reidel. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. [381] W. Stationarity and Regeneration. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Astin Bulletin VII. Appl. Wolff (1990) Stochastic Modeling and the Theory of Queues. 148164. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. University of Michigan. Act. Tidskr. 31. [376] A. 6785. PrenticeHall. Thorin & N. 936952. Lin (1994) Lundberg bounds on the tails of compound distributions. [383] G.W. Probab. 1942. Wald (1947) Sequential Analysis. SpringerVerlag. Oper. D. Sherman & D. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Astin Bulletin IX.E.382 BIBLIOGRAPHY [370] 0. Waters (1983) Probability of ruin for a risk process with claims cost inflation. 5762. 4963. Dordrecht Boston Lancaster. [377] V.R. 137153. [380] W. Unpublished Ph. M. 37. Van Wouve. thesis. 231246. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Thorisson (2000) Coupling. [382] G. Insurance: Mathematics and Economics 15. 743756. Willinger. Statistical Science 10. [384] R. In: Premium Calculation in Insurance (F.
323 Coxian distribution 147.4447.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.226.307312 compound Poisson model 4.217. 245248. 91. 119. 97. 217. 283. 196201 inverse Gaussian distribution 76. 17.259261.318320 change of measure 2630.293294. 138139. 110113.346349 383 .292293 Edgeworth expansion 113.178184. 135. 7475. 8283 hyperexponential distribution 7.150. 180182. 89. 17. 97129. 7879. 117128. 111117.200201. 9496. 332333 Volterra 192194.100. 1819. 201 Brownian motion 3 .299. 189. 2425.121129. 201214.6779. 1415. 7179.281. 37. 318319 Erlang distribution 7.160167. 207 heavytailed distribution 6.272. 302303 diffusion approximation 17. 170173. 2526.203.328330. 15. 40. 278 gamma distribution 67.314316. 79. 162164. 117127 corrected 121127 duality 1314.359 aggregate claims 103106.242. 3032. 141144. 3436.203.228229.9899.301 central limit theorem 60 .269. 239. 360 excursion 155156. 5796. 9396. 3839. 12 CramerLundberg approximation 1617.137141.182.285292. 39. 80 81. 248 WienerHopf 144 interest rate 190. 4851. 1112.and sum 221.Index adjustment coefficient 17. 205. 3334. 122. 14.249250 integral equation 16 Lindley 143 renewal 64. 9293. 316323 Bessel function 102. 271274. 86. 7079. 226. 361 diffusion 3.86.287292. 5.249. 308. 218 Cox process 4. 341.251280 heavy traffic 76. 227229. 301 Kronecker product.135.185187.
349 350 perturbation 172173. 142 likelihood ratio : see change of measure lognormal distribution 9.134135. 14.304 process 2830. 261264. 44.285287 queue 14 . 37. 267269 Panjer's recursion 320323 Pareto distribution 910. 39. 6162. 171. 100. 52 53. 179 NP approximation 318320 Palm distribution 5253. 71. 145187. 260 Lundberg conjugation 6979 . 162. 3639. 157. 4446.215250. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 65.234240.298299.348 terminating 215216. 7576.123.160161. 295.336339 . see also sensitivity analysis phasetype distribution 8.384 ladder heights 4756. 138. 149. 178 modulation 12.240244.128129. 5758. 96. 141144. 16.339 large deviations 129. 44.218221.275278. 25.180. 112113.227230. 9899. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 251. 38. 137139. 132133. 2730.139141. 41.148.174. 108109.161. 154. 25. 3947. 245 M/G/1 13. 7179. 133.161.238.161164.269271.336339 Laplace transform 15. 59.201.315 inequality 1718. 99.302. 3947. 133. 175 light traffic 8183 Lindley integral equation 143 process 3334. 6970.340350 multiplicative functional 2830.108. 306316 Levy process 3.178182.288290. 144. 234 matrixexponential distribution 240244 matrixexponentials 14. 176185. 134135. 203 Markov additive process 12.350361 Poisson process Markovmodulated 12 periodic 12. 134. 15. 16. 257. 203204.297299. 113114. 269 PerronFrobenius theory 4142. 42.152160.287291 INDEX matrix equation . 230.259261. 86 periodicity 12. 227228.146148. 32. 35. 80.261264. 35. nonlinear 155. 304305 random walk 3336.234. 213214. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 271274.287. 38. 106108. 108 life insurance 5.
294296 shotnoise process 314 simulation 19. 189214. 141144. 9693.336339 workload 13. 260 reinsurance 8.262263. 317318 semiMarkov 147. 233234. 5455. 240. 213. 37. 191192.279280 Rouche roots 158. 338 utility 324. 326330 Weibull distribution 9. 12. 238 saddlepoint method 115117. 160. 327 . see also matrixexponential distribution regenerative process 264 268. 331336 equation 64.273274. 11. 279280 subexponential distribution 11. 168172 storage process 13. 229234. 292294. 251. 261264 reservedependent premiums 14. 31. 177 timereversion 14. 280. 281296 stable process 15. 260 WienerHopf theory 144.359361 stochastic control x stochastic ordering 18. 1819. 131144. 244. 7475. 257. 107. 146. 251. 256258. 147.186. 4950. 251280 time change 4. 123.314. 152.244250. 335336 sensitivity analysis 8693. 253. 332333 model 12. 60. 222.INDEX 385 waiting time 141. 186187 renewal process 131. 186187 virtual: see workload rational Laplace transform 8. 233. 172173. 162. 174. 307308. 120 statistics x. 89. 223226. 3032. 87.154157. 333334 regular variation 10. 8386.
"This book is a must for anybody working in applied probability. Markovmodulation or periodicity.Vol. It is a comprehensive treatment of the known results on ruin probabilities. Special features of the book are the emphasis on change of measure techniques. 2 A I 11 JjVb l' i  i Yj . phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. I 1! Ruin Probabilities . the ^W A l \ i l ' ''' CramerLundberg approximation.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. for heavytailed claim size distributions)." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.g..Advanced Series on Statistical Science & Applied Probability .com 2779 he 9 "789810ll22293211 . exact solutions. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. y finite horizon ruin probabilities. worldscientific. P'i yfliother approximations (e. Some i (l I JL I J r of the topics are Lundberg's inequality..
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.