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
. . . 326 Appendix 331 Al Renewal theory . . . . . . . . . . . .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . . .. The twobarrier ruin problem . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . 340 A4 Some linear algebra . . . .. . .. . . . . . .. . . . . . . 290 5 Regenerative simulation . . . 287 4 Importance sampling for the finite horizon case . . .. . 316 5 Principles for premium calculation . . . . . . . .. . . . . . .. . . . 259 3 The renewal model . . .. . . . . . . . . . . . . . 331 A2 WienerHopf factorization . . . . . .. . . . . . . . 323 6 Reinsurance . 297 2 Further applications of martingales . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . .. . . . . . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . ... . . 336 A3 Matrixexponentials . . . . . .. . . . . . . . . 350 Bibliography Index 363 383 . .. . . . . . 271 6 Reservedependent premiums . . . . . . . . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . .. 344 AS Complements on phasetype distributions . . . .. . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . . . . . . . . . . . . . 261 4 Models with dependent input . . . . . . . . .. . . 251 2 The compound Poisson model . . . . . . . . . . . . . .. 292 6 Sensitivity analysis . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 3 Large deviations .. . .
This page is intentionally left blank .
Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. and the series editor Ole BarndorffNielsen for their patience. the idea was close to expand these to a short book on the subject. But the pace was much slower than expected. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. this applies to longrange dependence which is intensely studied in the neighboring ix . In particular. and has been an active area of research from the days of Lundberg all the way up to today.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. and other projects absorbed my interest. The course was never realized. and my belief was that this could be done rather quickly. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. and the result is now that the book is much more related to my own research than the initial outline. 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. Apart from these remarks. Since I was to produce some handouts for the students anyway. 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!). Thus. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. which has in particular removed one of the standard criticisms of the area. if the formulations occasionally give a different impression. it is not by intention. However. As an excuse: many of these projects were related to the book. that it can only say something about very simple models and questions. I have deliberately stayed away from discussing the practical relevance of the theory. University of Copenhagen. It has obviously not been possible to cover all subareas. the book is basically mathematical in its flavour.
2.g. VII. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157].13. VII. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].g. an area which is becoming increasingly important. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties.6 (to understand the PollaczeckKhinchine formula in 111. another by method.x PREFACE field of queueing theory. some papers not cited in the text but judged to be of interest are included in the Bibliography.3. IX. 111. see also Schmidli [325] and the references in Asmussen & Taksar [52]. VI.lth.4a. The rest is up to your specific interests. In the classical setting of CramerLundberg models.15. Another interesting area which is not covered is dynamic control.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . read Chapter I. For a second reading. Finally. for the effects on tail probabilities. A book like this can be organized in many ways. Good luck! I have tried to be fairly exhaustive in citing references close to the text. the first part of 11. The main motivation comes from statistical data for network traffic (e.g.maths . Resnick & Samorodnitsky [303] and references therein. IV. Concerning ruin probabilities.13. 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). I intend to keep a list of misprints and remarks posted on my web page. [381]).13 and IX. see in particular Michna [259].2 more properly). e. One is by model.2. it has not been possible to incorporate more numerical examples than the few there are. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. More recently. In addition. IV. Hojgaard & Taksar [206]. For a brief orientation. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. The present book is in between these two possibilities. 111.14. http:// www. the standard stochastic control setting of diffusion models has been considered.13 and XI.1.5. see e. IV. incorporate 11.89. X. Asmussen. It is obvious that such a system involves a number of inconsistencies and omissions. I regret that due to time constraints.lth. Willinger et al.45. VIII.se Lund February 2000 Soren Asmussen . Here is a suggestion on how to get started with the book.
1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.5 from Asmussen [21] with permission from CRC Press. A number of other figures were supplied by Christian Geisler Asmussen .8 . not least the more complicated ones. of which there are not many at this stage . as well as some additional references continue to be at the web page. 111 . . Parts of X. Fig.1 by Bjarne Hojgaard and the table in Example 111. Section VII . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. More substantial remarks. were produced by Lone Juul Hansen . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . Fig.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII.6 is reprinted from Asmussen & Schmidt [49] and parts of IX. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . many of which were pointed out by Hanspeter Schmidli .2 by Rafal Kulik .1 and X.PREFACE xi The second printing differs from the first only by minor corrections. Section VIII. Parts of II. Aarhus.4 from Asmussen.6. 5. IV.6 by my 1999 simulation class in Lund. supported by Center for Mathematical Physics and Stochastics (MaPhySto).
This page is intentionally left blank .
respectively. (1. T) as ruin probabilities with infinite horizon and finite horizon .Chapter I Introduction 1 The risk process In this chapter . is a model for the time evolution of the reserves of an insurance company.2) (O<t<T Ro=ul. we introduce some general notation and terminology. A risk reserve process { Rt}t>o.T) = P inf Rt < 0 I . They are the main topics of study of the present book.Rt.i(u.1) We also refer to t/) ( u) and 0(u. (1. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. (1. For mathematical purposes. results and topics to be studied in the rest of the book.4) O<t<oo O<t<T 1 . it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . and give a very brief summary of some of the models. M = (1.3) sup St. 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}. as defined in broad terms . MT = sup St. We denote throughout the initial reserve by u = Ro.
E Uk. Figure 1. say. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. T3. the ruin probabilities can then alternatively be written as .7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. we see that Nt Nt Rt = u + pt . (1. However. That is. and T1 is the time of the first claim.2 CHAPTER I. Thus. respectively. 1. We denote the interarrival times of claims by T2.5) i..pt. • Premiums flow in at rate p.6) Sofar we have not imposed any assumptions on the risk reserve process. St = E Uk .1. . the time of arrival of the nth claim is an = T1 + • • • + Tn. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. the number Nt of arrivals in [0. (1. per unit time.T) = F (MT > u) = P(r(u) < T).b(u) = P (r(u) < oo) = P(M > u)..i(u.1 . t] is finite. Putting things together. (1.
then M < oo a.s. It would appear obvious. For the purpose of studying ruin probabilities this distinction is.s. on Fig. We shall discuss Brownian motion somewhat in Chapter IV. one may well argue that Brownian motion in itself could be a reasonable model. and hence O(u) < 1 for all sufficiently large u.8) holds. 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.b(u) = 1 for all u.e. however. but as an approximation to the risk process rather than as a model of intrinsic merit. since any modeling involves some approximative assumptions. We shall not deal with this case either. of course. • Brownian motion or more general diffusions. that the insurance company should try to ensure 77 > 0. and the basic ruin probabilities are derived in XI. 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. however..20%. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. immaterial. We study this case in Ch. (1.1 the slope of {Rt} should depend also on the level). one could well replace Rt by Rtnr(u) or RtA. However.. and hence . 1.) V 0. . VII.1. say 10% .1. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.(. though many results are straightforward to generalize from the compound Poisson model. If 77 > 0. not discuss whether this actually corresponds to practice. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. 1. then M = oo a. allowing a countable infinity of jumps on Fig. Thus. a basic references is Gerber [127].1 Assume that (1. t * oo. for example. and in fact: Proposition 1.8) The interpretation of p is as the average amount of claim per unit time. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure.1. If 77 < 0. we shall.
and that . t t p  p' t ^ oo. are i.. This case is referred to as the mixed Poisson process.i(u. not all models considered in the literature have this feature: Example 1. zP(u .. where {Nt} is a Poisson process with rate .s.T) for {Rt} is given by V)(u) = t/i (u).v. and here (1. _ St __ k =1 Uk pt a4.d. are i. St In concrete models. 0 We shall only encounter a few instances of a Cox process. then similarly limSt/t < 0.10) hold with p constant.6EU (on the average. .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.Q (say) and U1.4 CHAPTER I. if {(3(t)} is nonergodic.. The simplest example is 3(t) = V where V is a r . INTRODUCTION Proof It follows from (1. Here it is easy to see that p = . it is not too difficult to show that p as defined by (1.s.i. M < oo a.i. namely that M = oo a. . However. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). namely. rl > 0. 0(u. k=1 (1. then this limit is > 0 which implies St a$ oo and hence M = oo a.11) . Thus p may well be random for such processes..8) that F N. we obtain typically a somewhat stronger conclusion.s. in connection with risk processes in a Markovian or periodic environment (Chapter VI). Proposition 1.3 Assume p 54 1 and define Rt = Rt1p. U2.b(u) < 1 for all u when rl > 0. . (1. If U1.T) = i. this needs to be verified in each separate case. (1..10) is a property which we will typically encounter.Tp).8). However.10) Again. If u oo. corresponding to the Pdlya process. and independent of {(0(t). Nt)}. U2. tb(u) = 1 for all u holds also when rl = 0.d. (1. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. with the most notable special case being V having a Gamma distribution.oo t 0 J (provided the limit exists). and independent of {Nt}.8) is given by ^t p = EU • lim it (3(s) ds t. If 77 < 0.
Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. another important early Swedish work is Tacklind [373].g. Sundt [354]. in particular. Pentikainen & Pesonen [101]. Some early surveys are given in Cramer [91]. De Vylder [110]. the assumption > 0 is equivalent to p < 1. In the even more general area of nonlife insurance mathematics. we shall be able to identify p with the traffic intensity of an associated queue. the role of the result is to justify to take p = 1.g. in a number of models. but in probability and applied probability as a whole. Insurance: Mathematics and Economics. Note that life insurance (e.. Daykin. lighttailed distributions (sometimes the term . We roughly classify these into two groups . Schmidt & Teugels [307] and Seal [326]. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Cox processes are treated extensively in Grandell [171]. often referred to as collective risk theory or just risk theory.. was largely initiated in Sweden in the first half of the century.2. [101]. Buhlmann [82]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Notes and references The study of ruin probabilities. the claim arrivals are Poisson or renewal at the same time). many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. Grandell [171]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. see e . Gerber [157]. Segerdahl [334] and Philipson [289]. Some of the main general ideas were laid down by Lundberg [250]. the research literature is often published in journals like Astin Bulletin . Schmidli. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. the recent survey by Grandell [173] and references therein. Gerber [159]) has a rather different flavour. Taylor [364].. [330]. Straub [353].. Hipp & Michel [198]. For mixed Poisson processes and Polya processes. Heilmann [191]. Note that when p = 1. The Swedish school was pioneering not only in risk theory. see also Chapter XI. Rolski. Besides in standard journals in probability and applied probability. Daykin et al. [134]. [76]. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. U2. Since { Rt } has premium rate 1. and we do not get near to the topic anywhere in this book. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Embrechts et al. 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.
For example in the compound Poisson model.3) .f. 2a Lighttailed distributions Example 2.O(u) can be found in closed form. but different more restrictive definitions are often used: subexponential.g.g. regularly varying (see below) or even regularly varying with infinite variance.f.8.B(x) satisfies B(x) = O(e8x) for some s > 0. the m.2 and /LB is the mean of B.u at the time of ruin given r(u) is again exponential u with rate 8.e. INTRODUCTION 'Cramertype conditions' is used). (2. In particular.6 CHAPTER I. Example 2 . s<8. P B[s]= (8Is ) . 6 has density r(p)xPleax b(x) P and m. B is heavytailed if b[s] = oo for all s > 0. B[s] is finite for some s > 0. Here lighttailed means that the tail B(x) = 1 . As in a number of other applied probability areas. and can also be interpreted as the (constant) failure rate b(x)/B(x). a fact which turns out to contain considerable information.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. if 1 °O AB Jbos x B(dx) > 0. for the compound Poisson model with exponential claim sizes the ruin probability . On the more heuristical side. 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'. Equivalently. then the conditional distribution of X .1) The parameter 6 is referred to as the rate or the intensity. The crucial feature is the lack of memory: if X is exponential with rate 6.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). i.2) = 0. where B(bo. and heavytailed distributions. In contrast. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) .2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. the exponential distribution is by far the simplest to deal with in risk theory as well.
0..i.. we develop computationally tractable results mainly for the Erlang case (p = 1. P b(x) = r` aibiea. by Grandell & Segerdahl [175] and Thorin [369]. .2. In particular. then X v Xl + • • • + X. .v. p)..) VarX1 (EX )2 p is < 1 for p > 1.2) can be considered as the pth power of the exponential density (2. . The exact form of the tail B(x) is given by the incomplete Gamma function r(x. u .d. X2.. p.. and exponential with rate d. if p is integer and X has the gamma distribution p.y i=1 where >i ai = 1.v. the squared coefficient of variation (s. This special case is referred to as the Erlang distribution with p stages.ate (b2 ): L• i=o In the present text. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p .).1 Poisson events in [0. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. An important property of the hyperexponential distribution is that its s. where X1. p) = J tPletdt. u Example 2 . x] so that B(x) = r` e. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. p) °° where r (x. are i.c. or just the Erlang(p) distribution. In particular.1) (or the 1/pth root if p < 1). B(x) = r(p) Asymptotically.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. one has r(bx. among others. is > 1... 2. > 1 for p < 1 and = 1 for p = 1 (the exponential case). i = 1. the Gamma density (2. Ruin probabilities for the general case has been studied.c. 0 < ai < 1.
but the current trend in applied probability is to restrict attention to the class of phasetype distributions.f. there exists a xo < oo such that B(x) = 0 for x > xo. 1)' is the column vector with 1 at all entries.g. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.e. a. The density and c. 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. Important special cases are the exponential. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. T) or sometimes the triple (E.f. the Erlang and the hyperexponential distributions.d. Example 2 .5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.7) are possibly complexvalued but the parameters in (2.8) are realvalued. equivalently. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained.6.. However.(2.6.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. INTRODUCTION Example 2 . .8) j=1 j=1 j=1 where the parameters in (2. resp. B(x) = aeTxe where t = Te and e = (1 . 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 .1 and defer further details to u Chapter VIII.xo)+ is covered by the reinsurer). We give a more comprehensive treatment in VIII. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. T) is called the representation. (or.. The couple (a. Example 2 . The parameters of a phasetype distribution is the set E of transient states. are b(x) = aeTxt. of which one is absorbing and the rest transient.8 CHAPTER I.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . which is slightly smaller but more amenable to probabilistic reasoning. See XI.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. We give some theory for matrixu exponential distribution in VIII. q2 q3 (2. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.
12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. Here failure rates b(x) = b(x)/B(x) play an important role. the tail is B (x ) 2 x. (2.p a 1 (2.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.2. the mean u is eµ+a /2 and the second moment is e2µ+2o2. In particular.1. u Example 2 . b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . the exponential distribution representing the simplest example since here b(x) is constant. a)/A)a+1' x > a.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. we obtain the Weibull distribution B(x) = eCx'. (2.10) The loinormal distribution has moments of all orders. p is defined as the distribution of ev where V .13) u . However. a2). or equivalently as the distribution of a°U+µ where U .11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. Writing c = d/r. (2. and then b(x) = 0.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.u l b(x) = d dx or J ax lor 1 exp Asymptotically. All moments are finite.N(p.pl = 1 W (logx . b(x) = crx''le`xr.9) which is heavytailed when 0 < r < I. There are various variants of the definition around. It follows that the density is 't (1ogX . CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. in practice one may observe that b(x) is either decreasing or increasing and may try to model smooth (incerasing or decreasing) deviations from constancy by 6(x) = dx''1 (0 < r < oo).1).N(0. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. Example 2 . x < a.
's of the form YX.10 CHAPTER I. where Y is Pareto distributed with a = (p .12) (here L (x) * 1) and ( 2. Choudhury & Whitt [1] as the class of distributions of r.16) 11 Example 2. oo) is slowly varying . In general. A = 1 and X is standard exponential. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. in particular. (2.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. the density is { 3 (1 . satisfies L(xt)/L(x) 4 1.17) where L (x) is slowly varying. The density is 8p(log x)pi b(x) . 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). (2. u . u Example 2 .x6+lr(p) (2.1)/p. in particular.v. 6 is defined as the distribution of et' where V has the gamma density (2.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. examples of distributions with regularly varying tails are the Pareto distribution (2. the loggamma distribution is a Pareto distribution.14) The pth moment is finite if p < 5 and infinite if p > 5. x 4 oo (any L having a limit in (0. Thus.(1 + 2x + 2x2)e2x) p = 2 (2. another standard example is (log x)').(1 + Zx + $ p = 3. INTRODUCTION Example 2. { s () 1s+3s29s3log(1+2s I p=3. B(x) = O(xP).L( x ). x + 00.e.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) . i. For p = 1.13).15) x2 + 16x3 ) a3x/2) 3 (1 . The simplest examples correspond to p small and integervalued.2).
3 The arrival process For the purpose of modeling a risk process ..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. 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. the knowledge of the claim size distribution will typically be based upon statistical data. (2. Namely. Also. However. When studying ruin probabilities. From a practical point of view. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. this phenomenon represents one of the true controversies of the area... and so is the Weibull distribution with 0 < r < 1. for example the lognormal distribution is subexponential (but not regularly varying). By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1.3. Similar discussion applies to the distribution of the accumulated claims (XI. U2.1. one may argue that this difficulty is not resticted to ruin probability theory alone.1) that any distribution with a regularly varying tail is subexponential.18) B(x) It can be proved (see IX. We return to a closer study in IX. which each have a ( timehomogeneous) small rate of experiencing a . and based upon such information it seems questionable to extrapolate to tail behaviour. The reason is in part mathematical since this model is the easiest to analyze. though the proof of this is nontrivial.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed.. Thus. but the model also admits a natural interpretation : a large portfolio of insurance holders . but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. THE ARRIVAL PROCESS 11 Example 2. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. the claim size distribution represents of course only one aspect (though a major one). 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.
so that . found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . we study this case in VI . T2.. where {/3 (t)}too is an arbitrary stochastic process . This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . This model . see 11. In order to prove reasonably substantial and interesting results . 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. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. are i. However . Some of them have concentrated on the marginal distribution of NT (say T = one year ). the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . the negative binomial distribution. with a common term {Nt} is a Markovmodulated Poisson process .6. and also that the ruin problem may be hard to analyze .12 CHAPTER I. Cox processes are. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . the periodic and the Markov modulated models also have attractive features . The compound Poisson model is studied in detail in Chapters III. too general and one neeed to specialize to more concrete assumptions . epidemics in life insurance etc.e. has some mathematically appealing random walk features . A more appealing way to allow for inhomogeneity is by means of an intensity . in particular to allow for certain inhomogeneities. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). but with a general not necessarily exponential distribution ).. INTRODUCTION claim .d. in Chapter VII). which facilitate the analysis. I. Historically.. Another one is Cox processes. however.. This applies also to the case where the claim size distribution depends on the time of the year or . to be studied in Chapter V. Nevertheless . 5. e. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too.i. To the author 's knowledge .g. when Jt = i.8 (t) is a periodic function of t. such that 8(t) = . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. 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 .3(t) fluctuating over time. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. In others. IV (and. gives rise to an arrival process which is very close to a Poisson process.(3. Mathematically. An obvious example is 3(t) depending on the time of the year (the season). with the extension to premiums depending on the reserve. getting away from the simple Poisson process seems a crucial step in making the model more realistic.
others being branching processes.v. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. and here (4. methods or modeling ideas developed in one area often has relevance for the other one as well. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. Thus. 0(u) = P(V > u). reliability. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. More generally. and which seems well motivated from a practical point of view as well.T) = P(VT > u). queueing theory. Markovmodulated or periodic can be related to queues with similar characteristics.0 (u. that quite often the emphasis is on computing expected values like EV.'s like V is available. It should be noted. 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 . A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. dam/storage processes. Similarly.6) . time series and Gaussian processes. The ones which appear most related to risk theory are queueing theory and dam/storage processes. Some of these have a certain resemblance in flavour and methodology. however.1) holds as well provided the risk process has a premium rule depending on the reserve. (4.1) permitting to translate freely between risk theory and the queueing/storage setting. interacting particle systems. stochastic differential equations. others are quite different.1). stochastic geometry. point processes and so on. and a lot of information on steadystate r. The study of the steady state is by far the most dominant topic of queueing and storage theory. extreme value theory. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI.4. In fact. it is desirable to have a set of formulas like (4. genetics models. this amounts to Vo having the stationary distribution of {Vt}). R = p(R) in between jumps. Mathematically. ruin probabilities for risk processes with an input process which is renewal. and the limit t 4 oo is the steadystate limit. In the setting of (4. this gives only f0 O°i (u)du which is of limited . with Poisson arrivals and constant release rule p(x) = 1. 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.1) where V is the limit in distribution of Vt as t + oo.
INTRODUCTION intrinsic interest . B(x) = ebx.6. the functions w x f d 1 exdx () . p = 0/8 and y = 8 .3.1 . The fact that Theorem H. which gives a sample path version of (4. have to some extent a different flavour. 3. • The compound Poisson model with a claim size distribution degenerate at one point.p(y) y^ Jo p(x) can be written in closed form. the two areas.3. as is typically the case. Similarly. 3.1. Vi(u. Thus .2).'s like the environmental process {Jt} in a Markovmodulated setting. A prototype of the duality results in this book is Theorem 11.. Example VIII. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. which can be expanded into a sum of exponential terms by diagonalization (see. see Corollary III. see Boxma & Cohen [74] and Abate & Whitt [3]. • The compound Poisson model with some rather special heavytailed claim size distributions. .8. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). Here ?P(u) is explicit provided that . • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . 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.1). e .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 ). Here O(u) = peryu where 3 is the arrival intensity.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r.T). 4b Exact solutions Of course .3. 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. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.14 CHAPTER I. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B.v. though overlapping.g. see Corollary VII. The infinite horizon (steady state ) case is covered by letting T oo.3.
T). 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.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). We don't discuss Laplace transform inversion much. the formulas ( IV. see VIII. However. Embrechts.S(u) 1S(oo) f °D exp {. but are somewhat out of the mainstream of the area . T) themselves.Lef$er function. For the finite horizon ruin probability 0(u. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .2) is the natural scale. it is easier to find the Laplace transforms = f e8 .1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. 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 . the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. . O(u. A notable fact ( see again XI.b(u)du . (u. Here are some of the main approaches: Laplace transform inversion Often. relevant references are Grubel [179]. Ab(u). • An astable Levy process with drift . 1).4. f {eXp U LX 2.ff 2µ(y)/a2(y) dy} dx . 191).u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. Abate & Whitt [2]. esuTb( u. Given this can be done.f f 2µ(y)/a2(y) dy} dx  (4. a2 (x): Ip (u) = where S(u) = f °O exp {. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. [s. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. T) can then be calculated numerically by some method for transform inversion.7.
as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). and carry out the solution by some standard numerical method.3) where C = (1 . and in particular the naive idea of conditioning upon process behaviour in [0.and integral equations The idea is here to express 'O(u) or '(u. u * oo.or integral equation.g. dt] most often leads to equations involving both differential and integral terms. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. T) as the solution to a differential.) B[s]. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). U is explicit in terms of the model parameters. (4. .3) in the compound Poisson model which is an integral equation of Volterra type. and in quite a few cases (Chapter VIII). However.4) 00['Y]1)'Y = 0. see VIII.1) and y > 0 is the solution of the Lundberg equation (4.p)/(13B'[ry] . Differential.3.7. In the compound Poisson model with p = 1. it states that i/i(u) . most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole).Ce"u. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. 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.f. which can equivalently be written as f3 [7] = 1 +13 . 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.
in such cases the evaluation of C is more cumbersome. but typically not very precise in their first naive implementation.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. incorporating correction terms may change the picture dramatically. However. Large claims approximations In order for the CramerLundberg approximation to be valid. other approaches are thus required.7 and IV. It has generalizations to the models with renewal arrivals.2. See Chapter IX. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. In fact. a Markovian environment or periodically varying parameters. In the case of heavytailed distributions. T) for large u are available in most of the models we discuss. in some cases the results are even more complete than for light tails. some further possibilities are surveyed in 111 . 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.4. Diffusion approximations are easy to calculate. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. 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. for the compound Poisson model ^(u) p pu In fact .6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. For example. T). Approximations for O(u) as well as for 1(u. (4. u > oo. . when the claim size distribution is of phasetype. the claim size distribution should have an exponentially decreasing tail B(x). often for all u > 0 and not just for large u. corrected diffusion approximations (see IV. J B dx. In particular. This list of approximations does by no means exhaust the topic. However. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself.
How do we produce a confidence interval? And.i. . This procedure in itself is fairly straightforward. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. In practice.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. . one expects a model with a deterministic claim size distribution B. they have however to be estimated from data. fitting a parametric model to U1. this is extrapolation from data due to the extreme sensitivity of the ruin probabilities to the tail of the claim size distribution in particular (in contrast. given NT. and to plot the empirical mean residual life 1 N .18 CHAPTER I. though not too many precise mathematical results have been obtained.3). For example. This is proved for the compound Poisson model in 111. The standard suggestion is to observe that the mean residual life E[U .g. For example. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. as a general rule. UNT may be viewed as an interpolation in or smoothing of the histogram). it splits up into the estimation of the Poisson intensity (the estimator is /l3 = NT/T) and of the parameter(s) of the claim size distribution. it has the advantage of not involving approximations and also... more importantly. UNT are i. . . . in the compound Poisson model. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. e.U(k)) i =k+ i . However. the difficulty comes in when drawing inference about the ruin probabilities. When comparing different risk models. empirical evidence shows that the general principle holds in a broad variety of settings. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. lower bounds etc.d. .8. T]. it is a general principle that adding random variation to a model increases the risk. INTRODUCTION Compared to the CramerLundberg approximation (4..) at various places and in various settings. which is a standard statistical problem since the claim sizes Ui. of being somewhat easier to generalize beyond the compound Poisson setting. obtained say by observing the risk process in [0. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. However. say degenerate at m.k (U(`) .
in this book referred to as [APQ]). See further Embrechts. formula (5.5.. .4. in all other chapters than VI where we just write . because it appears to require an infinitely long simulation. < U(N) are the order statistics based upon N i. Thus Proposition 4.v's) which can be generated by simulation. Still.(5.3) and Section VI. UN. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. Klnppelberg & Mikosch [134]. respectively. CONVENTIONS 19 as function of U(k). 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. this is a straightforward way to estimate finite horizon ruin probabilities. Truncation to a finite horizon has been used..3 (or just VI. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget..v.2. However. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. For example.3). (or a functional of the expectation of a set of r. where U(1) < . and of course the method is relevant in risk theory as well. The infinite horizon case presents a difficulty.i... good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. reference [14].d. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. claims U1.e. and look at them to see whether they exhibit the expected behaviour or some surprises come up. to observe whether one or the other limiting behaviour is apparent in the tail. .2.3) or Section 3 of Chapter VI are referred to as Proposition VI. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. We look at a variety of such methods in Chapter X. formula VI. The chapter number is specified only when it is not the current one. and in fact methods from that area can often be used in risk theory as well . A main problem is that ruin is typically a rare event (i. but is not very satisfying.
2. moment generating function.s.3) or Section 3. B is concentrated on [0. B(x) = P(X < x) = fx. as for typical claim size distributions.d.d. B(dy). squared coefficient of variation.h. for a probability distribution IIGII = 1.r.g.g. mation. random variable s. and for a defective probability distribution IIGII < 1. right hand side (of equation) r. E expectation.c. h + 0. w. A different type of asymptotics: less precise.o. or a more precise one like eh . (A.f. i.g. log E[s] where b[s] is the m.s. References like Proposition A.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c.29) refer to the Appendix.h. In particular. infinitely often l. IIGII the total mass (variation ) of a (signed ) measure G .d.t. The Laplace transform is b[s].e. say a heuristic approxi1 + h + h2/2. (moment generating function) fm e82B(dx) of the distribution B. .B(x) = P(X > x) of B.v. i.20 CHAPTER L INTRODUCTION Proposition 4. if B(x) . B(x) the tail 1 . cumulative distribution function P(X < x) c. E.f. then for 1s < 5).v. with respect to w. oo).f.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.f. left hand side (of equation) m.f. with probability Mathematical notation P probability.g.ceax. B[s] the m.i. If. n i oo. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. cumulant generating function. EX2/(EX)2. formula (5.4.g. r.p. Abbreviations c. independent identically distributed i. n! 27r nn+1/2en. see under b[s] below. .f.
an example or a remark. xa. . oo) the space of Rvalued functions which are rightcontionuous and have left limits.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). intensity interpretation. Unless otherwise stated. the ith entry is 1 and all other 0. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). Notation like f3i and 3(t) in Chapter VI has a similar .5. the value just before t. 21 D [0. 0 marks the end of a proof.e. Matrices and vectors are denoted by bold letters. In the Frenchinspired literature. 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. only have finitely many jumps in each finite interval. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. i. (the dimension is usually clear from the context and left unspecified in the notation). and column vectors have lowercase Roman letters like t. N(it. . Thus. i. matrices have uppercase Roman or Greek letters like T. all stochastic processes considered in this book are assumed to have sample paths in this space. Usually. of numbers. 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. 7r. row vectors have lowercase Greek letters like a.A] means E[XI(A)]. the processes we consider are piecewise continuous. Xt_ the left limit limstt X8f i. a2) the normal distribution with mean p and variance oa2. R(s) the real part of a complex number s.e. A.. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. a. E[X.e. the ith unit row vector is e'i.. I(A) the indicator function of the event A. . Usually. though slightly more complicated. F o r a given set x1.
1.g. interpretation. ry The adjustment coefficient.5. cf. or quantities with a similar time average interpretation. p the net amount /3pB of claims per unit time. I. 111. FL.22 CHAPTER L INTRODUCTION B the claim size distribution. e. J the rate parameter of B for the exponential case B(x) = eby. 'q the safety loading . though slightly more complicated. cf.1. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation.5. I. . cf. VI. Notation like BE and B(t) in Chapter VI has a similar.
somewhat more advanced than in the rest of the book. More precisely. the level of the exposition is. Sections 4. The topic is. in part.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. Due to the generality of the theory. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. The likelihood ratio approach in Section 2 is basic for most of the models under study. Sections 4. in most cases via likelihood ratio arguments. however. a parallel selfcontained treatment is given of the facts needed there. used in Chapter VI on risk processes in a Markovian (or periodic) environment. however. 5) are. in particular at a first reading of the book. When encountered for the first time in connection with the compound Poisson model in Chapter III. The general theory is. 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. The duality results in Section 3 (and. however. All results are proved elsewhere . The reader should therefore observe that it is possible to skip most of the chapter. strictly speaking. 23 .
u denote the overshoot. T(u) < T] + E [eryST . We get 1 = Ee7So = E e'Y S(. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). however. The more general Theorem 6. the time to ruin r(u) is inf It > 0 : St > u}.5 can be skipped. and in the limit (1. and the ruin probabilities are ip(u) = P (T(u) < oo). T) = P(T(u) < T).1 Assume that (a) for some ry > 0.QµB < 1.(u). (b) St a$ oo on {T(u) = oo}. V) (u. 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. Thus N. f1 . SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.2) As T > oo. T(u) < oo] + 0 = eryuE [e7Vu).0.2 Consider the compound Poisson model with Poisson arrival rate . T(u) > T] . As usual.24 CHAPTER II.1 is basic for the study of the compound Poisson model in Chapter III. (1. claim size distribution B and p = . StUit. Let e(u) = ST(u) . using r(u) A T invokes no problems because r(u) A T is bounded by T). Proposition 1. 1 Martingales We consider the claim surplus process {St} of a general risk process.2) takes the form 1 = E [e'ys(). Example 1 .. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .)AT = E [e7ST(°). {e'YS° }t>0 is a martingale.
u Corollary 1. O(u ) < e7".i.Q. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .1. B(x) _ edx.1. Thus. the ruin probability is O(u) = pe. u Corollary 1.r" where y = S .u + x is again just exponential with rate S. A simple calculation (see Proposition III. Example 1 . The available information on this jump is that the distribution given r(u) = t and S.5 For the compound Poisson model with B exponential.4 (LUNDBERG ' S INEQUALITY ) tion 1 . and p =.Q(B[a] .a.1 is satisfied.Q and the U.a it is immediately seen that y = S .. Since {St} has stationary independent increments. the martingale property now follows just as in Example 1.1) shows that Eels. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .= e"(') where K(a) = . condition (a) of Proposition 1. Thus.f. MARTINGALES 25 where {Nt} is a Poisson process with rate . and thus Ee7s° = 1. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.g. From this it is readily seen (see III.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.1 are satisfied. and (b) follows from p < 1 and the law of large numbers (see Proposition III. the conditions of Proposition 1.d.a = a . By standard formulas for the m.2(c)). 1.3 Assume that {Rt} is Brownian motion with variance constant o.(„)_ = x is that of a claim size U given U > u . Then {St } is Brownian motion with variance constant o2 and drift p < 0. Under the conditions of Proposi Proof Just note that C(u) > 0.Ft = a(S" : v < t).2 and drift p > 0. with common distribution B (and independent of {Nt}). Since {St} has stationary independent increments. of the normal distribution. the conditional distribution of the overshoot e(u) = U . and thus by the memoryless property of the exponential distribution .1.x.1) .ap.1) . and thus Ee'rs° = 1. Proof Since c(a) = /3 (B[a] . Eeas° = e"(°) where n(a) = a2a2/2 ./3. are i. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .2.3/6 < 1.
F(S) = P(S) = 1.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. cf. and is further exploited in his book [157]. and by the law of large numbers for the Poisson process . Grandell [171].. F). The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. Proof Just note that ^(u) = 0 by continuity of Brownian motion. B. then z/'(u) = e7" where 'y = 21A/a2. [172]. A somewhat similar u argument gives singularity when B $ B. I. . More recent references are Dassios & Embrechts [98]. on (DE. A E Ft. hence so is Nt = limfyo N2`i.6 N S = { lim Nt I t +00 t gJ are both in F. Embrechts.Ft}too and the Borel afield F. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.v. Grandell & Schmidli [131]. as shown by the following example this setup is too restrictive: typically'. and F. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. the parameters of the two processes can be reconstructed from a single infinite path. which we equip with the natural filtration {. oo). However. Example 2 . and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. The number Nt F) of jumps > e before time t is a (measurable) r. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. Thus the sets S = I tlim +oot t =. A]. (2. But if a $ ^ . then S and S are disjoint . Delbaen & Haezendonck [103] and Schmidli [320]. Theorem 2. P are then singular (concentrated on two disjoint measurable sets).3 below). 0 and claim size distributions B. Two such processes may be represented by probability measures F.1 Let F. P on (DE.e. 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).
A] = E[Lt. P) such that LLt = 1. that the restriction of P to (DE.e. Lets < t. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. we have E [ LTIFT]1 = LT on {T < T}. Ft). F the Borel o•field and P a given probability measure on (DE.Y) such that P(A) = Pt(A). if for some probability measure P and some {. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.2(i).r.Pt}adapted process {Lt}t>o (2.. Then Ft (A) = E[Lt.F such that (2. . G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). Hence E [_ . The truth of this for all A E Y. then {Lt} is a nonnegative martingale w.1) and nonnegativity by letting A = {Lt < 0}.Pt)) The following result gives the connection to martingales.1) holds. _. (i) If {Lt}t> o is a nonnegative martingale w. Then P(A) = E[Lt. A E F. A E Ft . ({Ft} . Lt < 0] can only be nonnegative if P(A) = 0. define P by Pt (A) = E[Lt.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A).2 Let {Ft}t>o be the natural filtration on DE. 1 J (2. 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 . the restriction of P to (DE. then there exists a unique probability measure Pon . Conversely.t. then { 1 P(G) = EG .2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. . G ] = E [LT .F). LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. using the martingale property in the fourth step. Proof Under the assumptions of (i). P be as in Proposition 2. (ii) Conversely. If r is a stopping time and G E PT.F8] = L8 and the martingale property. G C {T < oo}.1) holds. F) such that ELt = 1.r. A E F8.t.Ft}. implies that E[LtI. A]. A].2. By the martingale property. Finally. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. Proposition 2.3 Let {Lt}. ({. This proves (i). ELt = 1 follows by taking A = DE in (2.t.r.Tt) is absolutely continuous w.
we have F(G) = V )(u). (2. SOME GENERAL TOOLS AND RESULTS In the general case . r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1.t.h.1).28 CHAPTER II.s.4) compared to (1. of (2. first in the Markov case and next (Sections 4.Gn {r <T} . t.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. St). 5) for processes with some randomwalklike structure.r.r. each F.1. say.2) follows by monotone convergence. in continuous time (the discrete time case is parallel but slightly simpler). u From Theorem 2. the natural filtration {. applying (2.2) by noting that 1 = ersr(„) = e1'ue7Ou). Xt = (Jt.O(u) = eryuE[e'YC(u). . In the context of ruin probabilities. For the definition. is nonnegative and has Ey Lt = 1 for all x. Consider a (timehomogeneous) Markov process {Xt} with state space E. and this problem will now be studied..4) Proof Letting G = {r(u) < oo}. Xt = St.. where {Rt} is the risk reserve process. To this end. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. Rt) or Xt = (Jt. one would typically have Xt = Rt.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. 1 Since everything is non negative. The crucial step is to obtain information on the process evolving according to F. Lr(u) 11 The advantage of (2.t. Now just rewrite the r. (2. {St} = {u .Ft} . A change of measure is performed by finding a process {Lt} which is a martingale w.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . we assume for simplicity that {Xt} has Dpaths. First we ask when the Markov property is preserved. and letting T * oo.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu).1: Corollary 2.1) in Proposition 1.4 Under condition (a) of Proposition 1. T(u) < oo]. we need the concept of a multiplicative functional. both sides are increasing in T. is Markov w.
0 . which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.Ftmeasurable. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. and then L. The precise meaning of this is the following: being .v.Pt+8measurable r.5) Pxa.7). Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.v. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. s. Indeed.6) implies (2. A].s.Y8f t < s. the natural filtration {Ft} on DE.Ft+8measurable.8) which is the same as (2.r. since Zt • (Y8 o Ot ) is . or. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. nonnegative and Lt+8 = Lt•(Lso9t) (2.7) for any Ftmeasurable Zt and any . (2.. oo).2. The converse follows since the class of r. Zt.F8measurable r. t] * [0.Ft]. o 9t = V.'s of the form Zt • (Y8 o 0t) comprises all r..Ft }.'s of the form fl' f. Similarly. the Markov property can be written E. o 9tI.6) for any .YB] for any Ftmeasurable r. (2.v. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. By definition of Px. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .(A) = Ex [Lt. Vt+e. ({Xt+u} 0<u<8) Theorem 2.(Xtitl) with all t(i) < t + s. (2. t.5 Let {Xt} be Markov w. where Ot is the shift operator. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. Proof Since both sides of (2.7).v.[Y.5) are Tt+e measurable. t and let Px be the probability measure given by t. Y8. (2.Ft] = Ex. for all x. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.t.T9measurable Y8.v.
. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. . The storage process {Vt }o<t<T is essentially defined by time reversion. 3 Duality with other applied probability models In this section. Indeed.. The result is a sample path relation.. Thus R = Ro + f p(R8) ds . The formulation has applications to virtually all the risk models studied in this book. Ro = u (say). . 0 < vl < . T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. In between jumps.. . We work on a finite time interval [0..e. . aN where or* = T UN_k+l. More precisely . with a proof somewhat different from the present one. {Vt} .. R = p(R)). and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.At where At = k: vk <t U. t] = LtExt L8 = Lt.6 For {u ... (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. CN..30 CHAPTER H. } E[Lt+B I. the premium rate is p(r) when the reserve is r (i. t. < aN < T. A more elementary version along the lines of Theorem 2. (3. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. reflection at zero and initiar condition Vo = 0. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed..Ft] = LtE[L8 o 9t I. The corresponding claim sizes are Ul. In between jumps. then Remark 2. . SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. see Dynkin [128] and Kunita [239]..1) The initial condition is arbitrary. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x.5 can be found in Kuchler & Sorensen [240]. UN. the arrival epochs are Qi.
2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.___ . DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.3..1 The events {T(u) < T} and {VT > u} coincide.____•_. In particular..x.e.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u..11 4. Then rt°) > rt°) for all t when u > v.1.__..T) = P(VT > u). That is... V)(u. .T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. Theorem 3..AT_t. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous..3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. 3. instead of (3. (3. V = p(V)).1) we have Vt = At  f P(Vs)ds where A= U= AT . :. {Vt} remains at 0 until the next arrival)._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3._. (3.__.. The sample path relation between these two processes is illustrated in Fig.
Historically.3 and being i. Nevertheless. If VaN > 0.2 from Harrison & Resnick [188]. = r(VT) . u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. say of water in a dam though other interpretations like the amount of goods stored are also possible. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. The arrival epochs correspond to rainfalls. and a general premium rule p(r) when the reserve is r.Ul < roil  Ul = RQ„ Va1V_1 < RQ2.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). if nothing else n = N). we have r(u) > T. 3. one may feel that the interaction between the different areas has been surprisingly limited even up to today. and in between rainfalls water is released at rate p(r) when Vt (the content) is r.2 Consider the compound Poisson risk model with a general premium rule p(r). Then the storage process {Vt} has a proper limit in distribution.d. Proof Let T ^ oo in (3.i.1 with Ro = u = u2). u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . this represents a model for storage.1 and its proof is from Asmussen & Schock Petersen [50]. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. if and only if O(u) < 1 for all u.3). The results can be viewed as special cases of Siegmund duality. say V.U1 > roil . Resnick & Tweedie [79]. and then '0 (u) = P(V > u). the distinction between right.T l . Corollary 3. and so on. We get: Corollary 3. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. Theorem 3. Hence RQ„ > 0 for all n < N. Then Vo. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Then similarly VVN = r0. 3. and since ruin can only occur at the times of claims.1 with Ro = u = ul). we can repeat the argument and get VoN_1 > Ra2 and so on. with distribution B. Historically.U1 = Rol. we have RQ„ < 0 so that indeed 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. see Siegmund [344]. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].
Z2 = Y2....... WN be the Lindley process generated by Z1 = YN. generated by Z1.. W1.Y1 according to Wo = 0. . .1. has a proper limit W in distribution as n + oo.. there is an analogue of Theorem 3. .e.4.. N min (Y1 + + Yn). n=0.s.. and is reset to 0 once the r. . R valued sequence Z1....1..... Xo = 0. For a given i. . .. WN = YN . where the Yi are i . if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. {Wn}n=0. i. Theorem 4. can be viewed as the reflected version of the random walk with increments Z1. W1. = Xo + Y1 + • • • + Y. Z2. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 33 4 Random walks in discrete or continuous time A random walk in discrete time is defined as X.2).. ..YNn+1) n=0.... just verify that the r ..1.. as long as the random walk only takes nonnegative values.....1. Z2....2) satisfies the same recursion as in (4.. Proof By (4.d. Here F is a general probability distribution on R (the special case of F being concentrated on {1..w. I.... Most often. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4.h.. N min (Y1 + • • • + YNn) n=0.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.. 0 Corollary 4. (c) The Lindley process {WN} generated by Zl = Y1.1)). with common distribution F (say).Yl min (YN . For discrete time random walks . hits (oo. the Lindley process Wo... Then the events {r(u) < N} and {WN > u} coincide..1. Z2 .1 in terms of Lindley processes .min n=0.N (4. evolves as a random walk with increments Z1i Z2.2) (for a rigorous proof.. Let further N be fixed and let Wo. In particular.. 1} is often referred to as simple random walk or Bernoulli random walk). . ...1) Thus {Wn}n=o. 0). Z2 = YN_1 i . W2.. ZN = .2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. of (4..N From this the result immediately follows.i..d..1. .
. or M < oo a. SOME GENERAL TOOLS AND RESULTS (d) m = inf. .1.g... Proof Since (YN.3 The i. In general.i. N.s. Next consider change of measure via likelihood ratios. then the restrictions of Fx. 176) but appears to be rather intractable. By Kolmogorov's 01 law. + Z.. doubly u infinite (n'= 0.1 is actually not necessary .ooa. For a random walk. W v m and P(W > u) = P (m > u) = 0(u). ZN or.g.o..34 CHAPTER II. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . Combining these facts gives easily the equivalence of (a)(d). a Markovian change of measure as in Theorem 2.s . w.. (Yi + • • • + Yn) > oo a.=o. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.. (e)Yi+•••+Yn 74 . (d) #. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). Remark 4 .l.. the Y1. . there is a more general version of Corollary 4. 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.. a sufficient condition for (e) is that EY is welldefined and > 0. . .1. the Lindley processes in Corollary 4. e.. ±2.. .d. assumption on the Z1...1. YN).1 have the same distribution for n = 0. One then assumes Yn to be a stationary sequence. . . (e). YN in Theorem 4.1 is equivalent to WN D MN = (Z1 + . ±1... equivalently. Similarly. . . . ..1... Clearly.... Thus the assertion of Theorem 4.N so that WN _P4 M = supra=0..5 does not necessarily lead to a random walk: if.2..) sup n=0.l.) and defines Zn = Yn. either M = oo a.2 and Theorem 4.s. In that case .s.the result is a sample path relation as is Theorem 3. Y1) has the same distribution as (Y1. 0 By the law of large numbers.. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo..
The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .3) holds for n = 1. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. 100 ). Y) = h(Y ) a.s..5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . Conversely. Breiman [78] p. the changed increment distribution is F(x ) = E[h(Y).. y ).4) ({Ln} is the familiar Wald martingale .4 Let {Ln} be a multiplicative functional of a random walk with E_L.nrc(a )} (4.4. implying g (x. Y ) f (Y)] = E[g(O. For n = 2.g. of F). We get: Corollary 4. where h (y) = g(0.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. h(Yn) (4. (a) = log F(a] is the c. Proof If (4. In that case. Y1). the random walk property implies Ex f (Y1) = Eo f (Y1 ). cf. . this means E(g(x. = 1 for all n and x.Y2) = h(1'i)h(I'a). and define Ln by (4.3) holds. e.5 corresponds to a new random walk if and only if Ln = h(Y1) .. Y < x]. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. for some function h with Eh(Y) = 1..3) 1Pxa.4)... Then the change of measure in Theorem 2. Then the change of measure in Theorem 2.4. 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. In particular.f. (4. and so onforn =3..s. Y) f (Y)] for all f and x. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. Since L1 has the form g (Xo.g.
say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). In discrete time. (4. or imbedded into continuous time processes .Xt)I. but we omit the details ). e f x:IxJ>e} v(dx) < oo (4. Roughly. this description needs some amendments. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . The simplest case is 3 = JhvMM < oo. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . 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. say the beginning of each month or year . a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.Ft] = Eof (X.. . they arise as models for the reserve or claim surplus at a discrete sequence of instants.t(i . Xt(2)_t(l).. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. however. (4. Xt =Xo+pt+oBt+Mt.7) for all e > 0. However. corresponds to a process with stationary independent increments and u = p.6) More precisely. see Chapter V. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . v2 = 0 and v = 3B). the claim surplus process for the compound Poisson risk model . {Xt} is a random walk.). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < .1). The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). the tradition in the area is to use continuous time models.. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes).5) Note that the structure of such a process admits a complete description.36 CHAPTER II.Xn < x). we are . i. In continuous time.. a positive measure on R with the properties e J x2v(dx) < oo. . given by a the increment distribution F(x) = P(Xn+l .e. with premium rate p. {Xt} can be written as the independent sum of a pure drift {pt}. In risk theory. the pure jump process is given by its Levy measure v(dx).
] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3.2).8) O<t<T (assuming Wo = Xo = 0 for simplicity). jxJ v(dx) < oo. O(u.10) .min Xt (4. Now consider reflected versions of processes with stationary independent increments. (ex . cf. and b(u) = P(V > u).3 and decreases linearly at rate 1 in between jumps. 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). having Poisson arrivals with rate . has upwards jumps governed by B at the epochs of a Poisson process with rate .Q and distribution B of the service times of the arriving customers.s. oo]. WT = XT . i.e. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. Furthermore. Corollary 4. V E [0. is easily seen to be f3pB < 1. T) = P(VT > u).1)v(dx) (4. Proposition 4. VT + V for some r. 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. Then the storage process {Vt} has constant release rate 1.1.6). 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. Chapter III. 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. Here workload refers to the fact that we can interpret Vt as the amount of time the server will have to work until the system is empty provided no new customers arrive.v.6 In the compound Poisson risk model with constant premium rate p(r) . [The condition for V < oo a.7 If {Xt} has stationary independent increments as in (4. and the reflected version is then defined by means of the abstract reflection operator as in (4. then Ee'(xtxo) = Eoeaxt = etx(a). where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".4. defined as a system with a single server working at a unit rate.
f. t.g.Xt)g(s. Then the Markov process given by Theorem 2.6) are µ = µ + Oo2 . e. u Note that (4.Xt)I Ftl = Eof (X8)g(s. 5 has stationary independent increments as well. Proof For the first statement . of an infinitely divisible distribution (see. X8) = Eof (X8)L8 = Eof (X8)• For the second.1)eexv(dx). let e" (a ) = Eoeaxl.10) is the LevyKhinchine representation of the c.g.11) (eax . Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) .1 ) v(dx) .1 that E eaMt = exp fmoo In the general case .xo)tk ( e). (4. Theorem 4 . .1.5) and get E [f(Xt+B . Chung [86]). use the representation as limit of compound Poisson processes. we use the characterization (4. In particular.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. v(dx) = e9xv (dx)..Xt)L8 o 0tIFt] = E [f (Xt+s . if Lt = e9(xt . Xt Xo) with E2Lt = 1 for all x. Eea(µt + QBt) = et{aµ +a2OZ/2}. By explicit calculation .Xt)I'Ftl = E [f(Xt+B . we show in the compound Poisson case ( IlvIl < oo) in Proposition III.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. This is of course no coincidence since the distribution of Xl . 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. then the changed parameters in the representation (4. Xt +B .4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax . Q2 = v2.
0 in the following. . Ei instead of P2. u 5 Markov additive processes A Markov additive processes.9 If X0 = 0..5. is defined as a bivariate Markov process {Xt} = {(Jt.1) For shorthand .11 For an example of a likelihood ratio not covered by Theorem 4. Ei.Ft] = Ejt.0. corresponding to p = 1. where . one reason is that in parts of the applied probability literature. b with b(x) > 0 for all x such that b(x) > 0). 0. Then we can write v(dx) _ /3eOxB(dx) = / (dx).3B[B].l3 and claim u size distribution B. abbreviated as MAP in this section2. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.(3B(dx).o[f (S8)g(J8)].St)g(Jt+s)I. v(dx) _ ..(3 = . Example 4 . MAP stands for the Markovian arrival process discussed below..8. As for processes with stationary independent increments . dB/dB = b/b when B.3 and claim size distribution B # B. the structure of MAP's is completely understood when E is finite: 2and only there .. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. B have densities b.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .2. the corresponding claim sizes . 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. <t whenever the RadonNikodym derivative dB/dB exists (e.g.3 and claim size distribution B. Example 4 . 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 . (5. MARKOV ADDITIVE PROCESSES 39 Remark 4. Recalling that U1. B(dx) = B[9] B(dx). Thus (since µ = p = 1.3 =. it is then easily seen that Lt = H dB(Ui) i:o. U2. we write Pi.4). are the arrival times and U1.. . a = 0. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate .
In continuous time (assuming Dpaths). consider the matrix Ft [a] with ijth element least Ei . a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.1 For a MAP in discrete time and with E finite. vi(dx) in (4.. If all Fij are concentrated on (0.9 EE = (iii&ij[a])i j EE . (That a process with this description is a MAP is obvious.jEE• On an interval [t.f. the converse requires a proof. Jn = j. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.J1=j)= Fij (dx) Pij In simulation language.g. Fn[a] = F[a]n where P[a] = P . Y2..o(Ji = j. In addition. the distribution of which has some distribution Bij.. v.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i.it = A.6) depending on i.. {St} evolves like a process with stationary independent increments and the parameters pi.Sr_1.. by generating Yn according to Hij when J„_1 = i. Proposition 5. SOME GENERAL TOOLS AND RESULTS In discrete time. As an example.[a) = (Ei[easl. . An alternative description is in terms of the transition matrix P = (piA. 1 J1 ='^])iJEE = (Fij[a])i .) If E is infinite a MAP may be much more complicated. let {Jt} be standard Brownian motion on the line. which we omit and refer to Neveu [272] or cinlar [87]. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. a MAP is the same as a semiMarkov or Markov renewal process.40 CHAPTER H. Y1 E dx) where Y„ = S„ . 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. As a generalization of the m. oo). with the Y„ being interpreted as interarrival times. {Jt} is specified by its intensity matrix A = (Aij)i.i. t+s) where Jt .
In matrix formulation .1 )v(dx). this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. a= . up to o( h) terms. Jn+1 = A] = 41 Ei[ e 5„.(')(a)) diag + (). Jt = j] Ejesh'^ + E Ak j hEi [ease . we infer that in the discrete time case the .1) } (recall that qjj = 0). Jt = j] is given by etK[a].ijgij(Bij[a] . \ diag Ft[a] = Ft[a]K. Bij (i. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Then the matrix Pt[a] with ijth element Ei [east. Sn ) yields Ei[easn+ '. Jt = k] { 1 . MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.qkj + k?^j qkj Bkj [a] } = Ei [east. Jt = k] { xk kEE j la] . j E E) and So = 0. assume that the Markov chain/process {Jt} is ergodic. kEE Jn = k]Ek[e"Y" . aSt h = (1 + Ajjh) Ei [east . Then.2 Let E be finite and consider a continuous time Markov additive process with parameters A. 00 r(i) (a) = api + a2ot /2 + f (e° . 013 . vi(dx) (i E E).5. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a].4c).1)) . where K[a] = A+ (r. By PerronFrobenius theory (see A. vi(dx). pi. u Proposition 5. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . qij. u In the following.
and we shall take V(a)h(a) = 1. cf.4.e=e°tk. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . its derivatives are 'asymptotic cumulants'. 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.Eikjt = ttc'(0) + ki .42 CHAPTER II.r. of a random walk. Eie"sth^a) = e'Pt[a]h( a) = e. The corresponding left and right eigenvectors v("). (5. Yrh(a ) = 1.f. Proof For the first assertion. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). and write k = k(°). The function ic(a) plays in many respects the same role as the cumulant g. Proof By PerronFrobenius theory (see A. Proposition 5.c(a) (and h(")).h(a)vva)etw(a). Furthermore. . Corollary 5. we are free to impose two normalizations.4 Eie"sth(a) = h=a)et?("). Jt = j] .7. Since v("). h(") are only given up to a constants. Then h(°) = e. cf.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . In particular.t.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). u Let k(a) denote the derivative of h() w. Jeast.2) where 7r = v(°) is the stationary distribution.4c). We also get an analogue of the Wald martingale for random walks: Proposition 5.tK(a)h(a) J jj it L o is a martingale.3 Ei [east. Corollary 5.5 EiSt = tK'(0) + ki . a. Corollary 5. h(") may be chosen with strictly positive components. as will be seen from the following results.
we differentiate (5.+ k. In the same way. u The argument is slightly heuristic (e.a) + ttc (a)2hia ) Multiplying by v=.4) . Corollary 5.. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . (5. tam E tSt a (0). there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).6 For any stopping time T with finite mean. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. Remark 5 . For the second .g. subtraction yields Vary St = tic"(0) + O(1). the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.. . the distribution of Jo).Eikjr . for a random walk: Corollary 5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). 8 Also for E being infinite (possibly uncountable ).e. [E.") }) . t a oo.7 No matter the initial distribution v of Jo. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. one obtains a generalization of Wald's identity EST = Er • ES. Squaring in Corollary 5.5 yields + W (a)k.2ttc (0)Evkjt + 0(i).5. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h. E=ST = tc'(0)E7.4.5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. (5. More precisely. summing and letting a = 0 yields E„ [St + 2Stkj.
s. and the family {f LEE given by Theorem 2. in particular that f is bounded.6) We shall not exploit this approach systematically. s) = ea8h(i). G is defined as Gf (x) = lim Exf (Xt) . where {Jt} is deterministic period motion on E = [0.1) 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. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. 0) = n(a) h(i). u forsEE). Given a function h on E. From (5. we take the martingale property as our basic condition below (though this is automatic in the finite case). 1) (i. Usually. some extra conditions are imposed. Then {Lt } is a multiplicative functional. see.. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).e.3b and Remark VI. St) } as follows.(9) {Lt}t>o = . St)} be a MAP and let 0 be such that h(Jt) OStt. For t small .f (x) tyo t provided the limit exists.6. 0 Proposition 5. In view of this discussion ..5 defines a new MAP.4 that { h(Jt) easttK(a) L o (5. Jt = (s+t) mod 1 P8a. Remark 5. h(Jo) Lo is a Px martingale for each x E E.5) is a martingale . let ha(i.for the present purposes. gha(i. V. xEE .6.9 The condition that (5.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E.5. however.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. First. An example beyond the finite case occurs for periodic risk processes in VI.10 Let {(Jt. however. 0) = h(i )( 1 + ttc(a)). this is. this leads to h(i) + tcha( i. (5.
0<b<oo.7(dx) Bij [0] Bij(dx) in the continuous time case . (5. one can directly verify that (5.5. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. Ai = µi + 0Q.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. We omit the details.. .(0)j. in the discrete time case.Qi < oo and Bi a probability measure. if vi(dx) is compound Poisson. 0<q<1.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. Bi(dx) = Bi(dx). then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].10 is given by P = eK(e) Oh e) F[e]Oh(')..1) holds for the P. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . Bi [0] Remark 5.ic(0)e = ic(0)Oh e) h(e) .tc(0)e = 0 . That 0 < qij < 1 follows from the inequality qb <1. Then the MAP in Proposition 5.1) . In particular.7) In particular. Bi. 0 < qij < 1 and Bij [0] > 0. u Theorem 5. qij = r.c(0)e = tc(0)e .1) eft ea' f ij (dx) = Hij (dx) Hij [0] . 1 + q(b .11 below in the finite case. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). In the infinite case . vi (dx) = f3 Bi(dx) with . That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . ^i = of qij Bij [0] 1 + qij ( Bij [0] .11 Consider the irreducible case with E finite.
H1. .tc(0)I. (dx). Jt = A. is absolutely continuous w. Letting a = 0 yields the stated expression for A. F:j with a density proportional to eei . Similarly. in continuous time ( 5. Further Fib (dx ) = P=(YI E dx. it follows that indeed the normalizing constant is H1 [0]. Hence the same is true for H=j and H. Yi E dx. Now we can write K[a] =A+A ) ( K[a + 0] . this implies k[a] = A 1 ) (K[a + 0] .. a = 0 in (5. In matrix notation ..8).8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A.r. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5.t.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5.tc (') (0) corresponds to the stated parameters µ. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . v= . .8. are probability measures .e) Consider first the discrete time case .8) h(. Jt = j] = hie) . v.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .11.. Jl = j] :(Yi E dx.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . Jl = j) = Ei[Lt.46 CHAPTER II. since Hij. Here the stated formula for P follows immediately by letting t = 1.13) for matrixexponentials . Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. (dx) of a process with stationary independent increments follows from Theorem 4. This shows that F.Bay [0]) That k(') (a + 0) .
has no mass on (oo. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. however. i.6. < x) = 11 (S. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.e. there is. [226] and Miller [260]. Conditions for analogues of Corollary 5. and is typically defective. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. however.Bij[0]) = hjel)ijgijBij[0](Bij[a] . Though the literature on MAP's is extensive.1). [261]. which.)Ajjgij(Bij[a+0] . 7+ < oo).1) = Aij4ij(Bij[a] . 0]. see also Fuh & Lai [149] and Moustakides [264]. 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). oo). [225]. The closest reference on exponential families of random walks on a Markov chain we know of within the more statistical oriented literature is Hoglund [203]. the literature on the continuous time case tends more to deal with special cases..7). [262] in discrete time. h.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . Note that G+ is concentrated on (0. For the Wald identity in Corollary 5.6. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7.+ < x. an extensive bibliography on aspects of the theory can be found in Asmussen [16].3 for an infinite E are given by Ney & Nummelin [266]. . is slightly less general than the present setting. hardly a single comprehensive treatment..
1 For the compound Poisson model with p = 01LB < 1.2) .1. The main result of this section is Theorem 6.5 below. and the maximum M is the total height of the ladder. oo).1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. i. Here bo(x) _ B(x)/µB. Recall that B(x) = 1 .e.. g(y)R+(dy) = E f g(St)dt. R+ is concentrated on (oo.d. they have a semiMarkov structure (but in complete generality.1. 6. For the proof of Theorem 6. at present we concentrate on the first ladder height. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. Thus. In any case.T+ > t)dt = E f 0T+I(St E A) dt. we shall first consider the compound Poisson model in the notation of Example 1. = ST+(1) Figure 6. there are only finitely many). The first ladder step is precisely ST+. i. On Fig.(3B(x ) = pbo(x) on (0. Theorem 6 . G+ is given by the defective density g + (x) =. by approximation with step functions . SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. 6.ST+(1) and so on. In simple cases like the compound Poisson model. the dependence structure seems too complicated to be useful).2.i.00 ). 0 f T+ (6. where basically only stationarity is assumed. it follows that for g > 0 measurable. To illustrate the ideas. see Fig. has no mass on ( 0.e. o 00 (6..48 CHAPTER K. In other cases like the Markovian environment model. the sum of all the ladder steps (if rl > 0.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. 0].1. the second ladder height (step) is ST+(2) . Also.B(x) denotes the tail of B. a fact which turns out to be extremely useful. the ladder heights are i. which gives an explicit expression for G+ in a very general setting.
49 Proof Let T be fixed and define St = ST .ST_t. 0 < t < T) P(STEA.2(a): T+ > t Figure 6.O<t<T) = P(STEA.0<t<T) = F(ST E A. 6. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.ST<St. see Fig.O<t<T). 0]. 0 < t < T. That is. since the distribution of the Poisson process is invariant under time reversion. St S* t a Figure 6.St<0.T+>T) = P(STEA.2.2(b): r+ < t Thus. has the same distribution as {St}o<t<T. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .2 R+ is the restriction of Lebesgue measure to (00.ST<St. . P(STEA. ST < ST_t.6.
Figure 6. That is. The probability of this given { Su}u<t is B(A . oo). oo). and (6.3 where the bold lines correspond to minimal values.St). T+ > t] 0 _ /3 f E[B( A . and since the jump rate is /3.St _)I(r+ > t).t dT. U + St_ E A.2) in the last).. . for A C (0.3 G+ is the restriction of /3R+*B to (0. this is just the Lebesgue measure of A. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . E A} precisely when r+ > t.50 CHAPTER II. 6.3 Lemma 6 .y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS.St _). Fig. we get G+ (A) = f 00 /3 dt E[B(A .r. cf.y) (here we used the fact that the probability of a jump at u t is zero in the second step. SOME GENERAL TOOLS AND RESULTS Integrating w.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .. s. G+(A) = Q f 0 B(A . But since St 4 oo a.
{St+8 . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. .6.) where ak = Ti + • • • + Tk .4) are (ak. oo) x (0. h]} /h (by stationarity. as a point process on [0. 6 .. 4 (the points in the plane are (ak .4). 6 .s > 0). THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. 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). obviously. Nt St =>Uk k=1 t where Nt = max{k = O.* ) and the second the mark (the claim size Uk ).1 With r+(y) = I(y < 0) the density of R+.. i.. Fig. this does not depend on h).z)B(dz) _ f I(x < z)B(dz) _ f (x).1. The points in the plane (marked by x on Fig. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. this is equivalent to the risk process {St*} being stationary in the sense of (6. The traditional representation of the input sequence {(TT.S8 )t> o = {St }t>o for all s > 0. i. assuming basically stationarity in time and space. Uk) (k = 1.3 yields g+ (x) = . 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. 2. the first component representing time (the arrival time o. Uk) for those k for which ak . U k)} k=1 a is as a marked point process M *. In the stationary case. we define the arrival rate as E# { k : ak E [0 . The marked point process ..Q f r+(x . . cf.:T1 +•••+Tk <t}. oo).e.s.e..M o 08 shifted by s is defined the obvious way. Lemma 6. 0 Generalizing the setup.
Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. Section 5) which has pure jump structure corresponding to pi = a = 0. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.52 CHAPTER II. k: vk E [0. = 0 .. We represent M by the sequence (Tk.4 Given a stationary marked point process M*. vi(dx) = . Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists.2.5) does not depend on h.s. . Uk) k=1. See. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . Example 6 . h] Eco(M*) = 1 E f co(M o Bt)dt. The two fundamental formulas connecting M* and M are Eco(M) = aE E. Oh becomes the approximate probability F(ri < h) of an arrival in [0. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. the r. Sigman [348] for these and further aspects of Palm theory. o.5) represents the conditional distribution of M* given vi = 0. h] and the sum approximately ^o(M*)I(ul < h). h. V(M* o eak ).. 0.e.4 Consider a finite Markov additive process (cf. SOME GENERAL TOOLS AND RESULTS M* U. most often one takes h = 1).QiBi(dx). where TI = 0. of (6. As above . This more or less gives a proof that indeed (6. letting h J. i.g.. e. and let T = T2 denote the first proper interarrival time ..
oo a. 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 . qij when {Jt} jumps from i to j and have mark distribution Bij. After that.. A stationary marked point process M* is obtained by assigning Jo distribution Tr.OF(x).6 Under the assumptions of Theorem 6. . and that p = 0EU0 < 1. the ruin probability . It follows that we can describe the Palm version M as follows . the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .5. Note in particular that the Palm distribution of the mark size (i.*(0) with initial reserve u = 0 is p = /3EU0.O for i = j and iriAijgij/.e. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . let U0 be a r.s. v. Before giving the proof. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. First choose (Jo_. Jt = j is iri(3i /. and by some additional arrivals which occur w.p. aij for (i. 5 Consider a general stationary claim surplus process {St }t>o. we note: Corollary 6. If Jt_ = i. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . Jo) w.p. Assume that St * . j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. an arrival for M* occurs before time t + dt w. Then the ladder height distribution G+ is given by the (defective) density g+(x) = .= i.6./. let the arrivals and their marks be generated by {Jt} starting from Jo = j.p.O for i # j.O fo "o F(x)dx = . the probability aij of Jt .OEU0.
.Su_ <0..1] here the r . 0<u<t) = P(St EA. The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0. A standard argument for stationary processes ([78] p. h. { Su}0<u<t is distributed as a process {Su} .5.$St_ u.St<Su.. the mark at time Qk is denoted by Uk.A..o<u<t where a claim arrives at time t and has size Uo. which makes an upwards jump at time .g. Proof of Theorem 6.0<u<tIAt) = P(St EA.Su<0.0<u<t) = P(St EA.. 6.l...5). It follows that for A C (0.o. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. We then represent M by the mark (claim size ) Uo of the arrival at time 0. The last property is referred to as insensitivity in the applied probability literature. 105) shows that one can assume w. 0).. are point processes on (oo . moves down linearly at a unit rate in between jumps and starts from S0 = U.0<u<t) = P(StEA. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. T+ = t given the event At that an arrival at t occurs .Mt).(left limit) when 0 < it < t and is illustrated on Fig . and the kth preceding claim arrives at time t .0<u<tIAt) = P(St EA.Q_k and has size U_ k. (k = St}t>o 1..). . in (0. that M* and M have doubly infinite time (i.St*_ u.54 By (6.s. . in (oo.Su< 0. the arrival times 0 < 0'1 < Q2 < .5.e.St <. CHAPTER H.o. oo)). oo) x (0 . oo ) and the arrival times 0 > 0_1 > a_2 > . The sample path relation between { Su } and { Su } amounts to S„ = St . has a very simple interpretation as the average amount of claims received per unit time . Now conditionally upon At . Let p(t) be the conditional probability that ST+ E A. 2. oo) p(t) = P(St EA.
time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. 6. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. the support of L has right endpoint U0. Uo]. Mt)dt = i3EL(A) o"o . Thus. 0 < u < t } is the event that { Su } has a relative minimum at t . A sample path inspection just as in the proof of Lemma 6 .. and we let L(dy) be the random measure L(A) = fo°° I(St E A. and since by assumption St * oo a. In Fig. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6.6. cf.s. t a oo. 6. Since So = U0. the left endpoint of the support is oo.5 where it = { St < Su.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. Fig. G' (A) = 3 f P(St E A. NIt)dt .5.
2. A further relevant reference related to Corollary 6.5 is due to Schmidt & coworkers [48].6 is Bjork & Grandell [67]. .1). 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]. [263] (a special case of the result appear in Proposition VI.
• the premium rate is p = 1. i. see Chapter IV. . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . say.4 below . and independent of {Nt}. Panjer's recursion ( Corollary XI. For finite horizon ruin probabilities . exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. 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 . being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 .. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. with common distribution B.6) and simulation methods ( Chapter X). Thus .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.d. St = uRt = EUi t. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.. • the claim sizes U1. 4. and assume that • { Nt}t>o is a Poisson process with rate j3. U2. i. 3).e. are i. Some possibilities are numerical Laplace transform inversion via Corollary 3.
f.Rt.1) = t(p . for (d) just note that the kth cumulant of St is tic(k) (0). Proof It was noted in Chapter I that p . we shall start by giving the basic formulas for moments.g. [324]. e . 1 Introduction For later reference. Write pB^) = EUn' YB = Pali = EU. and Schlegel [316].58 CHAPTER III. cumulants . and that B(k)[0] = Pak). (d) The kth cumulant of St is tf3p(k) for k > 2. say stable Levy motion.1). and this immediately yields (a). m.. P = PAB = 1/(1 + rl) Proposition 1.g. Dufresne & Gerber [126].1) .+Uk)P(Nt = k) k=O e8t k=O B[s]k .. 0 .u . {Bt} a Brownian motion and {Rt} a pure jump process. See e. (b) Var St = t.1 is the expected claim surplus per unit time.t = E[Ntµs] .t = t(p . (c) Ee8St = et" (8) where c(s) = f3(B[s] . The same method yields also the variance as Nt Ne Nt Var St = Var E Uk = Var E ^ Uk Nt +EVar [ k=1 k=1 1 k=1 Uk Nt Var [Ntµs] + E[NtVar U] = 113µs + t13Var U = tf3pB2).)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. we get Ee8st = 00 e8t c` Ee8 (U1+. Furrer [150]. For (c). Schmidli [319].'s etc.t = fltpB . where K(k) (0) is the kth derivative of is at 0. of the claim surplus St . We do not spell out in detail such generalizations. THE COMPOUND POISSON MODEL Poisson risk process.s.6pBa).1 (a) ESt = t(13µ$ .t = E E [ U k k=1 k=1 Nt .1). A more formal proof goes as follows: Nt r Nt ESt = E > U k .
S„+V > S„ .1. where Tk is the time between the kth and the (k . we have Sok . Proof We first note that for u. which is often used in the literature for obtaining information about {St} and the ruin probabilities. if t = nh + v with 0 < v < h. St = oo. we get a discrete time random walk imbedded in the claim surplus process {St}. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim.1 = . (d) If 17 = 0.Sok_l = Uk . cf. and the value is then precisely v.V.Tk are i. then lien inft. Obviously. then St> SnhV>Snhh. so that {Sok } is a random walk with mean EUET = EU.1 is the same as if {St} was a random walk indexed by t = 0.i. The right hand inequality in (1.h < St < S(n+1)h + h. Sn+0 . then Snh . The connections to random walks are in fact fundamental. (a) No matter the value of 77.1)th claim. lim supt. meaning that the increments are stationary and independent.. we need the following lemma: Lemma 1. then St 4 co. Indeed. We return to this approach in Chapter V. II..Tk. the Uk .S„ attains its minimal value when there are no arrivals in (u.. obviously 0(u) = F(maxk Sok > u). 2. In particular. . St = oo. 1. The point of view in the present chapter is. (c) If 77 > 0.3) is proved similarly. Here is one immediate application: Proposition 1.3EU01 = 1µs where rt is the safety loading.3 If nh < t < (n + 1)h. u + v]. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.d.4. In this way.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. however. v > 0. (b) If 77 < 0. then St 00. rather to view {St} directly as a random walk in continuous time. For the proof.. For example.
For any fixed h. it is seen that upcrossing occurs at least twice. Notes and references All material of the present section is standard. THE COMPOUND POISSON MODEL Proof of Proposition 1. 2h..2: Proposition 1. Thus using Lemma 1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1._.. {Snh}n=o.2.1.o. (c) are immediate consequences of (a).4 The ruin probability 0(u) is 1 for all u when 77 < 0. and (b). The general case now follows either by another easy application of Lemma 1.6 Often it is of interest to consider size fluctuations. or by a general result on discrete skeletons ([APQ] p. then {St} upcrosses level 0 a.1. hence by induction i. where the size of the portfolio at time t is M(t).. is a discretetime random walk for any h > 0. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.1). this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt. and < 1 for all u when 77 > 0. p. . However.5 The limiting distribution of St . 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 . If rl > 0. There is also a central limit version of Proposition 1..1(b)) that the assertion holds as t 4 oo through values of the form t = 0. u 307)..3. at least once. and hence by the strong law of large numbers. Snh u = 00 (the lemma is not needed for (d))... is a discrete time random walk.. Considering the next downcrossing (which occurs w. {Snh}n=o. Remark 1 . Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. 169) stating that lim infra.s. lim supn_. if P(M > 0) = 1.2.t . 0 Snh = 00.. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. h.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). Snh/n a4' ESh = h(p .1..60 CHAPTER III.p. it suffices to prove 4'(0) = F(M > 0) < 1. This contradicts u St400. h A similar argument for lim sup proves (a). Corollary 1.3. 1 since St 4 oo) and repeating the argument.
the formula for the distribution of M follows . Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). i.1. d. This follows simply by noting that the process repeats itself after reaching a relative maximum. Theorem 2.. IV. p < 1.e. Combined with i/i(u) = P ( M > u). Summing over n.6. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. Note that this . As a vehicle for computing tIi(u).just before ruin is again B0. equivalently.34 or A. nevertheless. We assume throughout rl > 0 or.IIG +II)EG+ . It is crucial to note that for the compound Poisson model. (2. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. 1e. 1 The distribution of M is (1. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. the ladder heights are i. oo ). The expression for g+ was proved in Theorem 11. 11. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . [APQ] Ch. cf. Fig. Thus .1) is not entirely satisfying because of the infinite sum of convolution powers.1) representing the distribution of M as a geometric compound.2.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 .1 provides a representation formula for 0(u). and we further get information about the joint conditional distribution of the surplus and the deficit.1.6. THE POLLACZECKKHINCHINE FORMULA 61 2 The PollaczeckKhinchine formula The time to ruin r(u) is defined as in Chapter I as inf It > 0: St > u}. but we shall be able to extract substantial information from the formula. 0 Alternatively. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). B(x)/aB. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). which we henceforth refer to as the PollaczeckKhinchine formula. n=0 (2..
Feller [143] or Wolff [384]. Theorem 2. We assume rt > 0 throughout. (a) 11 (ST(o)_ > x. 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.6. cf. see Schmidli [323] and references there. In the risk theory literature. f +b (b) the joint distribution of (ST( o).62 CHAPTER III.6.d. ST(o )) given r (0) < oo is the same as the distribution of (VW. Theorem 2 . cf. W are independent. where it requires slightly more calculation. As shown in Theorem 11.just after ruin. 2 The joint distribution of (ST(o )_. ST(o) > y. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). in this setting there is no decomposition of M as a sum of i. ST(o)) is given by the following four equivalent statements: B(z) dz. 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. 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. Again. (d) the marginal distribution of ST(o)_ is B0. . (1 .1 is traditionally carried out for the imbedded discrete time random walk. there is a general marked point process version. V is uniform on (0. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. see for example [APQ].5. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup.5. The proof of Theorem 11. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0.V)W) where V. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.i. [62].2(a) is from Dufresne & Gerber [125]. Beekman [61]. Asmussen & Schmidt [49]. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. (c) the marginal distribution of ST(o)_ is Bo . However.
1)1 00 ( 1 . and hence this overshoot has the same distribution as the claims themselves .p) E pn S n x n. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . 0 . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. Thus r(x) = S(1 .3.6.p)pSe a ( l v)x = p( S . 1 . Thus . SPECIAL CASES OF POLLACZECKKHINCHINE 3a The ruin probability when the initial reserve is zero 63 The case u = 0 is remarkable by giving a formula for V)(u) which depends only on the claim size distribution through its mean: Corollary 3. the result follows .e.1 e ax = n1 (n . also be seen probabilistically without summing infinite series .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. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .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. Integrating from u to oo. the formula for P(O) holds in a more general setting. But claims are exponential . a further relevant reference is Bjork & Grandell [67].p. Let r ( x) be the failure rate of M at x > 0.1 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. however .p. then.2 If B is exponential with rate S.. 3b Exponential claims Corollary 3. As shown in 11.O)e(b0)x.p) = S . The result can. I. use Laplace transforms. For a failure at x.
h. (Example VIII.4) zu P(M > u . (3.2). u + oo. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).T+ <oo). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.y)f3 (y) dy.4) is similar (equivalently. the survival probability Z(u) = 1 .3).+ >u. Then the first term on the r.3) below.3.i(u) satisfies the defective renewal equation Z(u) = 1 . Corollary 3.y)/3B (y) dy.y)G+(dy) For the last identity in (3.+ <U.T+ <oo)+P(M> u. (b) use stopped martingales . 0 Proof Write o (u) as P(M>u) = P(S. II.S.S.p + f u Z(u . just insert the explicit form of G+. cf. is ?7+ ( u). The case of (3.3. (3. 2 is one of the main classical early results in the area.64 CHAPTER III.1. A variety of proofs are available . then 24 1 V. and weights 1/2 for each.p + G+ * Z(u) = 1 .y)G+(dy ) = f U V(u . we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .2) Notes and references Corollary 3.4) can be derived by elementary algebra from (3.+ = y yields P(M>u. (3. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. E. (3. THE COMPOUND POISSON MODEL In VIII. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. (u) 35eu + 35e6u.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 .g.3)). T+ <00) (3.3) Equivalently.s. and conditioning upon S. u .+ <u.1 p pBo(u).1) For a heavytailed B.
The approach there is to condition upon the first claim occuring at time t and having size x .(3 . g./3B[s] which is the same as (3. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.4) can be derived by elementary but tedious manipulations. Corollary 3.3 . 111112 or Feller [143]. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.7) s +.pBo[s] no (1 .8) Proof This can be shown. In fact.(3B[s] 1 . (3. by analytical manipulations (L'Hospital's rule) from (3. 191). eau B(u) du = f PB 3PB SPB 0 o (3./3B[s] . [APQ] pp.s .P)pB' (3.P)PB 2(1 .3 is standard .1 Bo[s] = f oc. 0 Notes and references Corollary 3.p)s s /3 . Of course.5). Embrechts. 206207).PPB2) EM2 = PPB) + QZPBl 2(1 . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.3.. e.g..4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .s . [APQ] pp.3 . Bo of B0 as m e8u B(du) = B[s] . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.5 The first two moments of M are 2 EM .p)2 3(1 . Some relevant references are Abate & Whitt [2]. We omit the details (see.7). In view of (3.5 can be found in virtually any queueing book.p) E p"Bo[s]" = 1 . Griibel & Pitts [132]. for example. see e.Ps s(.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . .5) Proof We first find the m. Also (3.g.p)s .)3B[s]) (3.p = (1 .5).7) and Corollary 3.f.6) 00 = (I .
.3I( 0<y<1)dy Z(y)/3I(0<u.1 < u < n and let Z(u ) denote the r.u + 1 )]k = QZ(u) .66 CHAPTER III.u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary./32(u . of (3.u)]k d 1 u) _ a) n ( du ( k! (1  .4) for Z( u) means f lhu Z(u) = 1.1).u) a)Qea" + (1 . differentiation yields Z(u) _ /3Z(u) . Z^ =eR(k.6 If B is degenerate at p.Q) k=0 k! E e0( = /32(u) .9) follows for 0 < u < 1.Q (k 1 k= n  [O(k .Q) 3e.u)]k1 ku+1) [/3( k .)3(1 . we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .s.9) shown for n . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3. eO('u) [)3(k .u) [N(k .h.3+ 18+ J0 Z(uy).u) [p(k .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 < 1.1).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./3Z(u .z/'(u) takes the form Z(u) L^J L. then p) 1: ep(k u/.9). Assume (3.1)! k=1 u1 .u)]k k! k0 The renewal equation (3. For n < u < n + 1.
CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.a.3) by t.rc(9) = . 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) . Formalizing this for the purpose of studying the whole process {St}. (4. B9(dx) = B[9] B(dx). 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.3B[9]. 0 Notes and references Corollary 3.1) .g.f.(9).2) shows that the solution is Ox [O]0].f. and thus (4. 4 Change of measure via exponential families If X is a random variable with c.1) .g. but will now be repeated for the sake of selfcontainedness.4. The question then naturally arises whether ie is the c.2). of F9.f.g. and define rce by (4. The answer is yes: inserting in (4. or equivalently BB[a] = B[^+ Repeating for t 54 1. (4.Qe(Bo[a] .a.4) works as well. say t = 1: recall from Proposition 1.4. we set up .1) or equivalently.d.(9) is welldefined.4) . K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. F and c.2) (Here 9 is any such number such that r. We could first tentatively consider the claim surplus X = St for a single t. in terms of the c.r. co(a) = rc(a + 9) . 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.3B = . we just have to multiply (4.1 that c(a) = /3(B[a] .f. (4.
d.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .nr. n) for a given n. then EBZ = E [Ze9ST _T"(9)I .. Ti(a)/n. in particular the expression (4.3 and claim size distribution B. and thus (4. The following result (Proposition 4.2 For any fixed T. But let Xk = SkT/n . (4.f. Let FT = o(St : t < T) denote the o•algebra spanned by the St. it suffices to consider the case where Z is measurable w. Proposition 4.68 CHAPTER III.S(k_1)Tln.10) .FT.Tic (0)} ..g.t. replications from Fe (replace x by xi in (4.r. (4.8) By standard measure theory. Z is measurable w. (4. (4.0e and claim size distribution Be.i. The identity (4. G]. and dP(T) dP^T) That is.t. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . EeeBSt + tk(B) = 1. .4). (4. . Then the Xk are i. G]. BB by (4. the corresponding expectation operator is E9. for G E FT. Xn).r. with common c. . Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.9) Proof We first note that for any fixed t.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . oo) governing a given compound Poisson risk process with arrival intensity. G C {T < oo}.8) follows by discrete exponential family theory. .1 Let P be the probability measure on D[0. and PBT) the restriction of PB to FT. t < T. .(9)} (4..3 Let T be any stopping time and let G E FT.d. and define 09.1) and multiply from 1 to n). = exp {BST .7) Proof We must prove that if Z is FTmeasurable. v(Xi.5) for the density of n i.1. .2.FTn) = Q(SkTIn : k = 0. THE COMPOUND POISSON MODEL Definition 4. the PBT) are mutually equivalent on.5) for the density.i. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .
Ee [exp { BST +Trc(9)} I(G) FT)] = 1. c(a) is a convex function of a. The behaviour at zero is given by the first order Taylor expansion c(a) r. Thus by (4.5. and hence (4.9) holds for G as well. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. Given FT.f. Then G E FT.7) holds.r is deterministic. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . 77 Thus.1 It is seen that typically) a ry > 0 satisfying 0 = r. Letting T t oo and using monotone convergence then shows u that (4.r)rc(9)}I . 5 Lundberg conjugation Being a c.1) ..(Y) = 13(B['Y] . Now consider a general G. t = T . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.g. GT C_ Jr < T}. subject to the basic assumption ij > 0 of a positive safety loading.1) _ 1 + a. according to what has just been proved. 5.10). (0) + rc'(0)a = 0 + ES1 a = a (p .FT]] = EB [exp { BST + Trc(9)} I(G)] .ST) + (T .9) holds with G replaced by GT.1(a). Then GT = G n Jr < T} satisfies GT E FT. the typical shape of rc is as in Fig. (4. Thus.7 1 Some discussion further supporting this statement is given in the next section. .
2 s As support for memory.QL instead of /37 and so on in the following . the Lundberg exponent. e. Example 5 . Fig. G = {T(u) < oo} in Theorem 4. 5.a = i(a + 7). THE COMPOUND POISSON MODEL exists . Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. Thus.3. .3. the claim surplus process has positive drift > 0.3.4) yields /3L = b and that BL is again exponential with rate bL =.1) .2)) is 7 = 5/3.1) is precisely what is needed for one of the terms in the exponent .g. Taking T = r(u). Fig.1) (or (5.70 CHAPTER III.s). Thus B[7] = 6/. (5.1 Consider the case of exponential claims.1(b).1(b). u It is a crucial fact that when governed by FL.3) cf. an equivalent version illustrated in Fig. Note that KL (a) = /L (BL [a] . It is then readily seen that the nonzero solution of (5. b[s] = 5/(b . 5. An established terminology is to call y the adjustment coefficient but there are various alternatives around. (5.2 is B(7) = 1 + ^. we further note that ( 5. Equation (5. we write FL instead of F7.1) is known as the Lundberg equation and plays a fundamental role in risk theory . 5. (5.2) 7 Figure 5.4) ELS1 = #L(0) cf. and (4.
(5. (5.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . Then P(ST+ E A) = EL [exp { 7S?+} .3.(u)} .1p .4).(oo) (in the sense of weak convergence w. V) (u) = P(T(u) < oo) = EL [exp {ryS.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.r. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.5). we can rewrite this as 0(u) = e"ELe7^(u).t.1e. To this end. e(u) has a limit i. PL ) with density 1 . we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . Since a7' is continuous and bounded. (5.3 takes a particular simple form.8) . V)(u) < e7u.e7x)G+(dx).5. Letting e(u) = ST(u) .5) Theorem 5 .G+L)(x)) dx ry^+L) J 00 f 0 (1 .ascending ladder height distribution and µ+ its mean. T = T+.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. Proof Just note that e(u) > 0 in (5. G = {S. T(u) < oo] .Ce7u as u 4 oo. see A .P Y j o' xeryxOB (x) dx /3k [Y] .6 ).1 (5.2 (LUNDBERG'S INEQUALITY) For all u > 0.+ E A} in Theorem 4. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.1. take first 0 = ry.7) is the same as (5. 0 Theorem 5 . ST+ E A] .6) Proof By renewal theory. where C .7) 0 and all that is needed to check is that ( 5.
1 above) and that C = p.3 (this was found already in Example 5. Using (5. we get L where 00 (1 .8) yields +L) J0 xel'B ( x) dx (5.72 CHAPTER III.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.10) VW = JI c* e° (x) dx = a (B[a] .12) Example 5 . or equivalently of how close the safety loading 77 is to zero. Noting that SIG(L)II = 1 because of (5.4 Consider first the exponential case b(x) = Seax. 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.1 . THE COMPOUND POISSON MODEL In principle. Then 0(u) = pe(a_Q)u where p = /3/S.7). this solves the problem of evaluating (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .1) (5. . but some tedious (though elementary) calculations remain to bring the expressions on a final form. (5. that 7 = S . u .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. From this it follows.e.")G + (dx ) = 1  J0 00 3B(x) dx = 1p.4). of course.
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)
.
Proposition 1. and hence the ruin probability approximation is b(u) e(bAln)u.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .1 As /3 f Nmax.2) was suggested by De Vylder [109]. Letting Bo be the stationary excess life distribution. [174]) shows that it may produce surprisingly good results.p + p { 1 1p ti 1 .3 and Corollary 3. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . the premiums exceed only slightly the expected claims.p .(3)2 P PB 2µB 2µB Letting /3* = /3/P. p* _ . Proposition 7./3)] 1 . numerical evidence (e.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . we shall represent this situation with a limit where /3 T fl but B is fixed. Though of course it is based upon purely empirical grounds.1.80 CHAPTER III.PBo [s (/3max . 1924.p = (/3max 0)µB. That is. (/3max . the approximating risk process has ruin probability z. we have according to the PollaczeckKhinchine formula in the form (3.3 )1 } _ 1p 1 .P . Mathematically.3* /S.(bA*)".Ps(/3max . 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.g.b(u) = p*e. but has an obvious interpretation also in risk theory: on the average.s(/3max . cf.)3 )PBo µB  . or equivalently that /3 is only slightly smaller than /3max = 1/µ8./3)PBo PB . Notes and references The approximation (7.1. Grandell [171] pp./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . heavy traffic conditions mean that the safety loading q is positive but small.
ze a2unµB laB (7./3)u * v. That is ./3)u). obviously Corollary 7. Numerical evidence shows that the fit of (7.ryu . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7.p _ 2rl11B PB p. 2 provides the better mathematical foundation. or equivalently that 0 is small compared to µB . This follows since rl = 1/p .B AB ) 6()3max _'3) = However .Q T /3max.3) is reasonable for g being say 1020% and u being small or moderate.1 1 . 7d The light traffic approximation As for heavy traffic . Mathematically. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . the premiums are much larger than the expected claims .0)u.2.p. the term light traffic comes from queueing theory. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . VIII).4) suggested by the Cramer Lundberg approximation and Proposition 6.l3)M > (/3max . light traffic conditions mean that the safety loading rl is positive and large . as well as it is needed for the interpolation approximation to be studied in the next subsection. the first results of heavy traffic type seem to be due to Hadwiger [184]. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. and hence 2µ2B 1 . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic .g. while the approximation may be far off for large u. we shall represent this situation with a limit where 3 10 but B is fixed. In the setting of risk theory. However .2 If . It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. light traffic is of some interest as a complement to heavy traffic . [APQ] Ch. u * oo in such a way that (3max .7. These results suggest the approximation Vi(u) e6(0_. Of course. but has an obvious interpretation also in risk theory: on the average .
For a more comprehensive treatment. z/' (u) convergence P(U . 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69].e. ( 3 J O B dx.= 1 aJ 1 a 0+ 1 = = p. THE COMPOUND POISSON MODEL Proposition 7. cf. 0(u) /3 J B(x)dx = /3E[U .3 is the same which comes out by saying that basically ruin can only occur at the F(U . see Daley & Rolski [96]. Indeed. (7.T > u) = J o" B(x + u)/3eax dx . Sigman [347].Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. Asmussen [19] and references there. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically.T > u). En'=2 • • • = O(/32) so that only the first terms matters.u)+.u. 10 ( u The alternative expressions in (7.82 CHAPTER III. Light traffic does not appear to have been studied in risk theory. . the Poisson case is much easier than the renewal case.5) u Proof According to the PollaczeckKhinchine formula. U > u] = /3iE(U .3 As . Omax max m. [97]. u Note that heuristically the light traffic approximation in Proposition 7. Again. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.5) follow by integration by parts. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . by monotone time T of the first claim .(3 10. i. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined.
_(E) (u). 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). Al ./3)) .O(E)(u) 1 (1 .3 and use similar notation for %(B) (u) = (u). the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. where further references can be found . 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). to get nondegenerate limits . no empirical study of the fit of (7.6) (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.VHT) ( ax QmQ ) h (B) ( . . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say).O0 M. (7. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . with rate 1/µB = /3max. one may hope that some correction of the heavy traffic approximation has been obtained. The adaptation to risk theory is new. we see that the following limits HT) (u'). "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. 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. Thus .6) is .8.x . however. . B(2). even if the safety loading is not very small. Substituting v = u(. (U). ) M. Notes and references In the queueing setting . Another main queueing paper is Whitt [380].3). Let OLT) (u) denote the light traffic approximation given by Proposition 7. available. we combine with our explicit knowledge of ip(u) for the exponential claim size distribution E whith the same mean PB as the given one B.Wmax f(x ) dx + pee6mQ. z/i(E) (u) = pe(QmaxQ)u. [84].3n. that is. Instead.
cf. B(' <. this ordering measures difference in variability. 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).6.1 If B(') <d B(2). this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. we can assume that 1) < St 2l for all t. Proof According to the above characterization of stochastical ordering.s. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. U(2) such that U(l) has distribution B('). most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. Proposition 8. then i. In terms of the time to ruin. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x.' 1)(u) < V)(2) (U) for all u. or the existence of random variables U(l). B(') <i. . Proposition 8. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. then . we have the convex ordering. U(2) distribution B(2) and U(1) < U(2) a. B(2)) if f fdB(1) < f fdB(2) for any convex function f. u Of course. we shall need various ordering properties of distributions. THE COMPOUND POISSON MODEL To this end. XI.84 CHAPTER III.1 is quite weak. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x.2 If B(') <j. the proof is complete. then Bill = B(2). A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. Finally. Here convex ordering is useful: Proposition 8. Rather than measuring difference in size. In the literature on risk theory. Taking probabilities. Bill is said to be convexly smaller than B(2) (in symbols. B(2) and PB(1) = µB(2).ill(u) < V)(2) (U) for all u. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. whereas (consider x2) B(2) has the larger variance. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f.
B. The problem is to specify what 'variation' means. This u implies that D <. The heavy traffic approximation (7.4) certainly supports this view: noting that.5 If '0(1)(u) < p(2) (U) for all u and a. we have by Jensen 's inequality that E f (U) > f ( EU). (D) (u) < O(B) (U ) for all u. with fixed mean.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. then B(1) <. . A partial converse to Proposition 8.2 is the following: Proposition 8.1 and µB at 1 so that the safety loading 11 is 10%.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..6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). from which the result immediately follows.4 Let D refer to the distribution degenerate at 'LB . A first attempt would of course be to identify 'variation' with variance. 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. say to p. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. A general picture that emerges from these results and numerical studies like in Example 8.(1) (. B(2).1. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB.. B(2). then /'(')(u) < 0(2)(u) for all u. Then V. and here is one more result of the same flavor: Corollary 8. Hence by the PollaczeckKhinchine formula .8. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. Proof If f is convex. Example 8.e. Corollary 8. larger variance is paramount to larger second moment. Proof Consider the light traffic approximation in Proposition 7.3 If B(1) <.6 Fix /3 at 1/1. and consider the following claim size distributions: B1: the standard exponential distribution with density ay.
. 0.000.0' U0. A standard example from queueing theory is . all distributions have mean 1. 9 Sensitivity estimates In a broad setting. A2 = 3. in comparison to B2 the effect on the ua does not show before a = 0.001 u0. We return to ordering of ruin probabilities in a special problem in VI.1358.e'\1x + 0. Note to make the figures comparable. For B1i B2. Let ua denote the a fractile of the ruin function.1%. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. However.4.lA. 0. B. [166]. Pellerey [287] and (for the convex ordering) Makowski [ 252]. B4: the Pareto distribution with density 3/(1 + 2x)5/2.4142. and consider a = 5%. i. van Heerwarden [189]. 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). B3 the comparison is as expected from the intutition concerning the variability of these distributions.e. B3: the Erlang distribution with density 4xe2x. = 0. 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. In terms of variances o2.) = a. the behaviour of which is governed by a parameter 9. 1%.01%. 11 Notes and references Further relevant references are Goovaerts et al.01%. Kluppelberg [234].86 CHAPTER III. One then obtains the following table: U005 U0. and this is presumably a consequence of a heavier tail rather than larger variance.9A2e'2r where A. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). 1/)(u.
Proof This is an easy time transformation argument in a similar way as in Proposition 1. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . Then if t is large . For example. Similar conclusions will be found below. if = a e(6A)u. a0 as ao 80 19P . t].(u) for large u.Ap). or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. Thus. we may be interested in a'/ap for assesing the effects of a small change in the premium. u Proposition 9. with 0 the vector of service rates at different nodes and routing probabilities.19P a/ . Then ib = Pe(613)u. In the present setting. Example 9. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively .1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Then the arrival rate /3(P) for { R(P) } is )31p.e.01/2u. the premium rate p and the claim size distribution B. increasing in u. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV.2 Consider a risk process { Rt} with a general premium rate p. Let R(P) = Rtli. where the partial derivatives are evaluated at p = 1. and s(9) the expected sojourn time of a customer in the network.1. i. where Q2 = fl ( l2 1113 / _ Ou2v)2. obtained say in the natural way as the empirical arrival rate Nt/t in [0.3. Assume for example that 8 is known. SENSITIVITY ESTIMATES 87 a queueing network.9. a/3 0 . Then a p ao = 00 Qa/. In particular . a2/t). say estimated from data. s(9) is of course the ruin probability t' = Vi(u) (with u fixed) and 0 a set of parameters determining the arrival rate 0. Thus at p = 1. it follows that ' is approximatively normal N(0.. while /3 = j3 is an estimate. the distribution of %3 0 is approximatively normal N(0„ Q/t).
and the proofs of (9.((dx ) = exp {Ox + (t(x) .r.(/3 + y)we(9 + 7. 4) (9 .5) are similar.()^ 1 . The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . 3) ( 9 . but must look for approximations for the sensitivities 0. Similar notation for partial derivatived are used below. Consider first the adjustment coefficient y as function of 3. and write yp = 8y/8/3 and so on .w(6. (9.3) follows by straightforward algebra. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. 9. we can rewrite the Lundberg equation as w(9+ y. However . e. Differentiating w.3 or/and B. various parametric families of claim size distributions could be considered.g.1 or Proposition 9. (3+'y)PC (0+7.()(0 +'0) ' (9 . () Proof According to (9.^)] 1(/3+y)we(9+'y.3.2) (see Remark 9.6 below for some discussion of this assumption).3.uypCe7u urypO.t.()wC(e. (9.Owe (9. it suffices to fix the premium at p = 1 and consider only the effects of changing . In the case of the claim size distribution B.3 70 = 'Ye = = 7 /3(1 we(e +'y. Viei '0(. Proposition 9. x > 0 (9.10) below. namely that of a twoparameter exponential family of the form Bo.0 = t/'(u) and the CramerLundberg constant C.()YC = 1 +y/ /3 \ Q2 From this (9. /3 yields w e(e + Y. () = log(1 + y//3). ()} p(dx) . (. so that heuristically we obtain '00 50ryu = Coe"u .88 CHAPTER III.w(O. for the ruin probabilities . we cannot expect in general to find explicit expressions like in Example 9. ^) . Consider first the case of 8/8/3: .4). THE COMPOUND POISSON MODEL As a consequence.6) As will be seen below. 5) (Q+'Y)[we(0+7. u Now consider the ruin probability 0 = 0 (u) itself. Of course. this intuition is indeed correct. but we shall concentrate on a special structure covering a number of important cases.
w(9.11) Ee.x). ()} .p)/C'y. 11 For the following. () . z2(u) _ 1 ^ e'ri`i7i( u .QB(x) dx. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. 0(u) = /3 Ju"O B(x) dx + f 0 0(u . we get p(u) = J "O B(x) dx + J U O(u .x).2 of the Appendix ).4t (U)e°`U = which are wellknown and easy to show (see e.9.we(9 . But from the proof of Theorem 5.w(9. and alsoo zl(u) + 0 because of B['y ] < oo. ()} . ()] exp {w (O + y.9) (9. O} (9.8).3(x) dx. () . Further write de = [we (9 +'y.([a] = exp {w(9 + a. the proof is complete.3 (see in particular (5. Z(u)/u a C//3PF where PF is the mean of F. () . u 0 Then Z = z + F * Z and F is a proper probability distribution . PF = (1 .4 As u oo. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.x) F(dx ) f u J C F(dx) = C as u 4 oo. (9.(e"U = = wS(O. w((9 + a.3) for z/'(u).x)B(x) dx + J U W(u . Z= zl + z2 where zl (u) = e7u J m B(x)dx. Hence by a variant of the key renewal theorem (Proposition A1.St (U) Ee.x)B(x)dx.12)). Be. we multiply by e7" and let Z(u) = elt" cp(u). z2(U) = e7" J u b(u . () . we note the formulas Ee. BarndorffNielsen [58]). () exp {w(9 + a.C). F(dx) = e'yy/3B(x)dx. By dominated convergence.w(O.8) Letting cp = e0/e/3 and differentiating (9. SENSITIVITY ESTIMATES Proposition 9.10) (9. Combining these estimates .8) (Section 5).g.
8) that cp(u) .x). 8^ ue7u. F(dx) = e7x. 8 8() 8( (9.w (9. ()](e7v . C)] (1 + 7 ) Proposition 9. )}B(dy)• Letting cp it thus follows from (9.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . ()} 1z(dy) = f [t(y) . 0 x Multiplying by e7" and letting Z(u) = e"uV(u).11). ()]B(dy) dx x 0 0C T ON O .x)f3 f ^[t(y) . C) . THE COMPOUND POISSON MODEL [we(e+7. Then as u > oo.w( (0.90 CHAPTER III. u Z2(U) = e7° f u ^/i(u . ()]B(dy) dx.w(0.2) holds.w( (0. ())B(dy) dx.12) f exp {O y + (t(y) .w(e.wc(9.6e7u f "o f[t(y) .5 Assume that (9.e7x/3 f 00 [t(y) .wc (O. By dominated convergence and (9.we (0. ^)} [wc (0 + 7. oo z2 (u) f C . ()]e7vB(dy) 'fCd 7 c .6C do 89 1p 8( 1p Proof By straightforward differentiation.QB(x)dx.wc(O. 2 z 07P N ue7u (3C de . ^) . z = zl + z2. this implies Z = z + F * Z. 01 (i+) do = +'Y.1) B(dy) 'f '[t(y) .9)(9.
9 = S. and also zj (u) 4 0 because of f Hence. t(x) = logx. that C = a.12) takes the form y) alp a .ry) 5a1 cry (5 . Z(u) /3C 91 o c'o e11(t (y) .a log S = log r(c) .C log(9).rye) S 5rya..2) holds with p(dx) = xldx. SENSITIVITY ESTIMATES as u 3 oo. w(e.9. ())B(dy) < oo. () = log r(a) .(log r(a) a log S)} • r(a) 1. We get w( (0. we (9. < = a.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.13) (9.1 ..yu/3C2do u86 89 1p' az/) = 8z/.QS 1 . a /(S .w((9. Here (9. . Example 9./35' a/i'y + aryl 625ry.Y)a+1 ' (9.18) (05 + 57 _'3_y . ( 9. () = C/9 = a/S.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . ue_Yu 'C2d( 8a 8( 1 p . () ='I'(t.pa+1 .3ary tog('Finally.15) (9. by inserting in the above formulas.16) (9. and the proof of the first one u is similar.12) follows.a/35a&y' ' (9. It follows after some elementary calculus that p = a)3/5 and.6 Consider the gamma density b (x ) = Sa xa.1edz = 1 exp {Sx + a log x .Sry a/32 + a/37 + /37 .17) (9.. U 7µF from which the second assertion of (9.
further yield .3.1 16 +ry c C22ry 2( = + 70 We (e.S[a] = exp {w (9 + a. 9 = .21og 2. THE COMPOUND POISSON MODEL Example 9.2) with µ(dx) = 2x3zrdx. Be.2 log (0. Straightforward but tedious calculations .22.log c = 2 In particular. C) = B = Yc = de = do = .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. t(x) _ . ()} = exp {c (C .2a) } Thus the condition B[a*] > 1 + a* /. which we omit in part . () = Cc .CZ try)} 1 C C2 try .w(9.3Ee. C) .9) (() . C = .92 CHAPTER III."62 .1 = eXP {c(C .2 . for a < a* = z (.([Y] . w(e.
to our knowledge.8 The specific form of (9. we can just fix k .. (9.g. To this end. the main tool is simulation.cue_7u)3C P Remark 9. we have assumed k = 2 and ti (x) = x.7 and references there. thus. Also. the exponent is either Ox.. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x).+UNT) > 1. queueing networks) are typically much more complicated than the one considered here. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. Comparatively less work seems to have been done in risk theory. Finally if k = 1. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. However . . BT [a]= NT ^` e"U. if 1 PT = /3TNT(U1+. Van Wouve et al. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . However.a. then ryT < 0. and hence explicit or asymptotic estimates are in general not possible. B are assumed to be completely unknown. or Ct(x). and we estimate y by means of the empirical solution ryT to the Lundberg equation.2) is motivated as follows. the models there (e. That it is no restriction to assume k < 2 follows since if k > 2. then BT and hence ryT is undefined.3C2de 1p' z a = c . 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. In general. in u which case the extension just described applies.1) . Thus.2 of the parameters. Note that if NT = 0.10.12) takes the form a = a 93 ar. for which we refer to 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). the results presented here are new. let NT 16T = ^T . [379] consider a special problem related to reinsurance. sj=1 and let YT be defined by IKT('ryT) = 0.. kT (a) = /T (BT [a] .oo. in which case we can just let t(x) = 0. ae t 1lEY u S _ . Thus.
b[Yp'V21 T { (E[7] . it is easy to see that we can write \ V1 1 l _ . 7T a4' 7. a2 where a2 = /3r.B[7]2 n Hence ( 10.)vl+ N CO. THE COMPOUND POISSON MODEL Theorem 10 . B [7]2 (10.1)2 + E[27] .If ..i3)(B[7] 1) + (3(BT[7]  .1) 'YT .2) rT(7) N N (0. Lemma 10 .a BT[7] I B[7] I + .1 As T 4 oo. If furthermore B[27] < oo. 1) r.3) Proof Since Var(eryU) = we have B[7]. (10. we need a lemma.B[7]) 0+ Iv/o(b[y].'Y . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . since NT /T . B[2'Y]  /3T ) . N ( n[7]. More generally.2 As T * oo.: N 0. vfoVFB[2y].v.1) .1) .B[7]2 }) ( T 0 . B[27] .94 CHAPTER III.T y . For the proof.B[7]2 V2 .Q and Anscombe 's theorem.7 + (.B[7]) + B [7] . . then (10./^ B[27] . V2 are independent N (0.: N ()3.(27)/K'(7)2.3/ T).3T . 16T where V1.'s.2) follows from NT/T a4' .
OT a 95 u 4 /3. and the truth of this for all e > 0 implies ryT at 'y. Let 0 < E < ry. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.3).'(y). To this end . where ryT is some point between ryT and ry. 7T E (y .c'(7) N (0' T (2(7) / N (0. 6"Y (10.1 can be used to obtain error bounds on the ruin probabilities when the parameters . BT[a] 3 B[a].10.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].Q. I. NT BT [a] Hence r. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).1 By the law of large numbers.4) and Lemma 10. Proof of Theorem 10.e.E) < 4T(7T) < 4T(7 + E).KT(7) kT(7) K'(7) . Theorem 10. NT i =1 n'(a) for all a so that for all sufficiently large T K7 . 0 are estimated from data . If ryT E (7  we have KT(7 .E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. By the law of large numbers. Then r. y + E) eventually. first note that e7TU N (e7U u2e27Uo'2/T) 7 .E) < 0 < kT(7 + E) for all sufficiently large T . it follows that 7T7 KT(7T) . °7IT) .4) + E).e) < 0 < r. Combining ( 10.2..(ry + e) and hence KT(7 .e. lcT(a ) 4 /c(a).(ry .
.g. Griibel & Embrechts [292]. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. [197].g. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94].i.96 if a = 2. Frees [146]. Embrechts & Mikosch [133].f. Mammitzsch [253] and Pitts. 6 < 2.Wn) are i .1 : Vt = 0.Q. Wn).e.5%). This approach in fact applies also for many models more general than the compound Poisson one. i .info< „< t S. and the known fact that the Y„ = max Vt tE[W„1. if B is exponential with rate 8 so that ry = 8 .ueryuU ". For this reason . Herkenrath [192].. THE COMPOUND POISSON MODEL Thus an asymptotic upper a confidence bound for a7' (and hence by Lundberg's inequality for 0(u)) is e"TU + f. various alternatives have been developed. Asmussen [23]) can then be used to produce an estimate of ry.3 or equivalently p > 1/2 or 11 < 100%. Notes and references Theorem 10.0) < 5. For example . with a tail of the form P(Y > y) . Letting Wo = 0. Csorgo & Teugels [95].. wn = inf{t > W.e. it means 2 (8 . Hipp [196]. i.. t]}.d..T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. Deheuvels & Steinebach [102]..96 CHAPTER III. satisfies b(.. ft. . the nth busy cycle is then [Wn1.T VIT where r7ry. Vt = St .) = a (e. U2.. One (see Schmidli [321]) is to let {Vt} be the workload process of an M /G/1 queue with the same arrival epochs as the risk process and service times U1. V.1 is from Grandell [170]. > 0 for some t E [Wn_ 1.C1e"a ( see e. = 1.
Unless otherwise stated.s. B[•] and mean AB. exists. 97 . 'y) where c(a) attains it minimum value. In particular.f. the premium rate is 1.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. The notation is essentially as in Chapter III. the Poisson intensity is 0 and the claim size distribution is B with m. The safety loading is q = 1/p . Further let 'Yo be the unique point in (0.1 where p = 13µB. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.g. 0.1 (the role of ryy will be explained in Section 4b). generalizations to other models are either discussed in the Notes and References or in relevant chapters. See Fig. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . 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.1) .
(U) < 00] = ELT(u)ke'YS.r. FL and independent of T(u). By the likelihood identity III.1) (1.(4.. 7. we have for k = 1.1 The claims surplus is {St}. PL = 6/0 = 1/p > 1). (u) is exponential with rate 0 w.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. 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. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .t.u is the overshoot. 1 Exponential claims Proposition 1.(.9). 2 that E [T(u)k. In particular. using that the overshoot l. 1 FL.2) Proof Let as in Example 111.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. Var[T(u) I T(u) < 00] = VarL T( U) .98 CHAPTER IV.5 . the time of ruin is T(u) and ^(u) = ST(t&) . . E[T(u) I T(u) < 00 ] = ELT (U). PROBABILITY OF RUIN IN FINITE TIME Figure 0.
T(u) < oo] fora > r. which leads to the quadratic 02 + (/3 . is V1rLSr( u) +VarL ((PL . 1). where = eBu I 1 . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).1)T(u)) = VarLe(u) + (PL .1)T(u))2 = UL where = s.(PL . u + ELe(u) _ PL .1)T(u) are independent with QL the same mean .(yo) = 2V .2) is aLELT( u) .h.1) .V/ is as asserted. 0 Proposition 1."(ry) = 26//32. Let 0 > yo be determined by ^c(0 ) = a. This means that /3(6/(6 .B = a. .0) .1. we have by Wald's identity that (note that ELSt = t(pL ./3) .h./3 . the 1. of (1. EXPONENTIAL CLAIMS For (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 .1)) ELST(u) ELT(u) (PL .h.1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.s. Wald's second moment identity yields 2 EL (Sr(u) .12 Thus the l.s.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.6 + a)0 .s.6.I (1.1)ELT(u).6a = 0 with solution 0 (the .2).1)2VarLT(u) + 2 Ca 1I VarLT(u).1//32 (6/)3 1)2 26(/3u + 1)/(6 .1 (6)3)2 which is the same as the r. Since Sr (u) and (PL .
are the lengths of of the ladder segments 2. .Y1 Y2 Figure 1.v. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).1 ..100 CHAPTER IV. Y2. But by the fundamental likelihood ratio identity ( Theorem 111. Y(u) belonging to a convolution semigroup ..3 that we can write EeaT( u) = eeuEe 017(o).. More precisely.3) we have E [e«T(u ).9ST(u) +T(u)!c(0)} . T(u) < oo] = EB [exp {aT(u ) . Fig. (1. Cf.. T2 .0. the result follows.T+ Ti a t U T I 1 a i F. Note that it follows from Proposition 1. Ti. . St Ti F..3..v.. and M(u)+1 is the index of the ladder segment corresponding to T(u). Using 5 = 6 .OuEee 04(u) = ee u be BB+B where we used that PB(T(u) < oo ) = 1 because 0 > ryo and hence E9S1 = K'(0) u > 0. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .1 where Y1.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. 1.4.. T(u) < oo] = e.'s with rate 5. are the ladder heights which form a terminating sequence of exponential r.
1.6.3 sin0 + 29) f3(0) = 1+/32/cos9.T is the residual service time of the customer being currently served and U2 .6(u) = Vfl/j l(Su. 2. For j = 0.i. If QT = N > 0. EXPONENTIAL CLAIMS 101 For numerical purposes . Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.T. . Proof We use the formula . Since U1 .. let (cf. 1). Let {Qt} be the queue length process of the queue (number in system.ST).3 Assume that claims are exponential with rate b = 1.T.v.T are conditionally i. the conditional distribution of VT given QT = N is that of EN where the r. .d. T. Corollary 11.T the service times of the customers awaiting service .1.4. where U1.T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1. i.0..T + • • • + UN.I ex cos B cos j O dB fo " .1 )!.cos (u/. T) to be evaluated by numerical integration: Proposition 1. Then V(u. EN has an Erlang distribution with parameters (N.k + 1). [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . UN. .. density xN lex/(N . U2. including the customer being currently served).T.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . UN. then VT = U1. .. and exponential with rate S = 1.T) 1 I fl(O)h(0) fdO where (1.i (u..1(u. cf. the following formula is convenient by allowing t.e. . 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)] L _112 /(k+1)/2 [. in particular equations (1. 8789) 00 E aj j= 00 = 1.(31/2eie .1 00 ok+lR 00 j=k1 +1)/2e .cos((k + 1)0)] f3(0) 00 flk +1 > j=k1 3j/2 COS(jB) l)/2ei(k+1)e )3j/2eije = R)3(k+ (31 /2eie .1 R [.)3k +1 tj g'(QT >. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = . and define tj = e(1+R)Taj/2Ij(2vT T).(31/2 cos (( k + 2)9) .cos((k + 2)9)] d9. k k2 + $k+1 E bj 00 t j .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . Then (see Prabhu [294] pp. similar formulas are in [APQ] pp. 912.cos (( k + 1)0)] f3(9) Hence the integral expression in (1.112 l 1( k +1)/2 [ 31/ 2 cos(kO) .44).i(k +1)e R [/3( klal/2e:0 (01 /2 e . (1.38).3(k +1)/ 2ei(k + l)6 (.13(k +l)/2ei(k +1)9 R E .31 /2eie L 1)] 1 I/31/2eie .8 ) yields F(QT > k + 1) .k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas. f3(0) .ie . let I _ j (x) = Ij (x).102 CHAPTER IV.
u Notes and references Proposition 1.. Seal [327] gives a different numerical integration fomula for 1 .e = e' COS a cos(uf31/2 sin 0). Ui < x I / (note that P(St < x ) = F(x + t.2. equivalently. The first formula. E Fk. t) = P .0(u. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. F(x. however. k! k=O k0 i/z Co Uk ate" o'/z e . The rest of the proof is easy algebra. k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). there are several misprints in the formula there. 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. from the accumulated claim distribution N. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. expresses V)(0. oo (u)31/2e^e)k = )3k z cos(k9) = R k. T) in terms of F(. T). . 2 The ruin probability with no initial reserve In this section . going back to Cramer. however. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. or. Related formulas are in Takacs [359]. t )).7) that _ [^ au ak+l (30 k L. is numerically unstable for large T. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. the numerical examples in [12] are correct). it follows as in (1.T) which.
(0.b (0. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. we define a new claim surplus process St StM NJ Figure 2. f T lStv)} 0<t<T by a 'cyclic translation'.0<w<t} St+v .i. 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. .(6.1 In formulas. 2.T))dv. T) = P(Tr(0) > T) = P(M(0.T) T F(x.t)= {Stv) < SM.1. 1 1 .T)) does not {Stv)} depend on v. [v. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .T].T)dx. co ).S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.3) with A = (0.T))dv E^T I(M(v. Stv^ _ Define M(v. T T o where the second equality follows from II. ") } is at a minimum at time t.104 CHAPTER IV. resp. Then 1 .(. Proof For any v E [0. T]. See Fig.T)) 1 fT P(M(v. v].
in which case there is a last time o where St downcrosses level u. T Theorem 2 . 0<t <Tv}n{ST<ST Sv+St T+v.t)dt.2. t).T)f(I z /)(0. Indeed. T)) dv. then i fT I(M(v. ST > 0.. T) occurs. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .Sv. Obviously.2 10(u.xdx. We claim that if M(0. Fig 2. It follows that if M(0 . T)).T)) dv f T I(M(0. t) denote the density of F(•. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. T) occurs.T) occurs.ST on M (0. w) for some small E. v<t<T}n{ST<STSv+St. this integral is 0 if STv) . there exist v such that M(v. T. v) = M(0. where the last equality follows from ST < St on M(0.2. For example. v). v).T) and Sv < 0 on M(0. we can write M(v.T) = F(u+T. 0<t<v} = {ST < St .. v).Tt))f(u+t. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. we can take v E (w E. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. then M(v. . v < t < T} n M(0. or it occurs. Hence T TE f I( M(v. If ST < 0. T) as {ST<St+ vS. letting w = inf It > 0 : St_ = mino<w<T Sw}. 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 .v<t<T} = {ST<StSv. T) = M(0. T]. T T o i =1 Let f (•. cf.T) occurs or not as long as ST < 0.
106 CHAPTER IV.T) = {St < 0.T)+ J0 T (1V. The following representation of T(0) will be used in the next section.(0. which is independent of St and has the stationary excess distribution B0. Proof of Theorem 111.T) = .2 . 2. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). Hence P(ST<u) = 1 . The proof is combined with the proof of Theorem 111. O(T . {St > .T) = C(z.t).3 Define r_ (z) = inf It > 0 : St = z}. Let Z be a r. Proposition 2. {S t > z.b(u. 0 < t <T . define St = ST . E [t.ST_ t_ and let A(z. 0 < t <T. ST_ _ z} .z. z > 0.p. ST_ _ z}.2 Here o.Tt))P(StE[u. which occurs w. u which is the same as the assertion of the theorem.v. u + dt] and there is no upcrossing of level u after time t.u+dt]). PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. For a fixed T > 0. 0 < t < T. t + dt] occurs if and only if St E [u.2. ST_ _ z}. C*(z.2.
ST(o)_ E [z. A(z. 7( 0) < oo) = P (C(z)) dT. we therefore have P(A(z. Proof of Proposition 2.3). z + dz]. . {St }o<t<T have the same distribution . and since {St}o<t<T.T + dT]. T(0) < oo) = OR(z) dz in (2. Figure 2. z + dz]. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. (2. r(0) < oo) = 3R(z) dz JP(C(z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. z + dz].3.1) z T .T).1) that P(T(0) E [T.2.ST(o) >y. Thus P(Sr(o)_>x.T)). T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. Fig.T))f3B(z) dz dT.1) yields P(ST(o)_ E [z. It follows by division by P(ST(o)_ E [z.3 But by sample path inspection (cf. 2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. T + dT] I S7(o)_ E [z. Hence integrating (2.2. z + dz]) = P(A(z. z + dz].2.T)) = P(Cx.T) = C*(z. u which is the assertion of Theorem 111.
2.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).6.(yo).r(a).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . 3 Laplace transforms Throughout in this section. Tak'ecs [359]. who instead of the present direct proof gave two arguments.5a). one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.(3(B[r( a)] . r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .T+dT]).1 and the present proof is in the spirit of Ballot theorems.1) .3. Lemma 3. z + dz]. Let T_ (y) be defined as Proposition 2. Note that T_ (y) < oo a. Notes and references For Theorems 2.1) where a > r. a martingale proof is in Delbaen & Haezendonck [103].s.3 was noted by Asmussen & Kl(ippelberg [36]. z + dz]. Theorem 2. Lemma 3 . T(0) < oo) = dTP(T_(Z) E [T. see in addition to Prabhu [293] also Seal [326]. r(0) < oo. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx.5 and one upon excursion theory for Markov processes (see IX. Proof Optional stopping of the martingale I er (a) 9 t. I L Let ga(x) be the density of the measure E[ear(°). cf.108 Hence CHAPTER IV. [329].1 Eear( y) = eyr(a). PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. Proposition 2. In the setting of general Levy processes.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.1.2. (3. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. because of77>0.
y + dy].2. Further by Theorem 111. Then by Proposition 2.r(a) b . (u .(v) = ev''(a).x)ga (x) dx where za(u) = f. time T(u): u u Here is a classical result : the double m. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).3.ST(o)_ just before ruin .f. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. Corollary 3.4 E[eaT (o). Hence eb"du E[eaT(").ic(b)/b x(b) + a eb"E[eaT(" ).3 ga[b] = c(b) Proof + b + a . rr(0) < oo) = 1_ r(a) Proof Let b = 0. u . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.x)(a) B(dy)• Lemma 3 .ga [b] 0 TO Using Lemma 3.3.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)) .1] evr(a)B(dy)[ b . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . the result follows after simple algebra.5 f 00 o a/r(a) . r(u) < oo).ga [b] 1 .g. Z = y] = EeaT.°° ga(x)dx.3.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . (Laplace transform) of the ruin Corollary 3. b . E[ear (o) I T(0) < oo .2 P(Z E [y.r(a) = a [B[b] B[r(a)]] .
110 CHAPTER IV. Then given r(u) < 00. (u) t.s. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .w ) v/. and take basically the form of approximations and inequalities. That is. Then as u * oo.e. This proves the first assertion of (4.1 Assume 77 > 0.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. Proposition A1. mu ) ( 0 m < ML '(u) 1 m > rL.1)Er(u) .6. Theorem 4 .(u ) = o(u) a. St/t 1 1/m. the known results are even less explicit than for the exponential claims case. = (p . The first main result of the present section is that the value umL. and hence a.mu D 2 4 N(0. for any m T(u) u . For the proof.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). note that by Wald's identity u + EC(u) = ES. T(u)/u mL as u + oo.2. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .1.2 Assume ri < 0. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model..r(u) = Er(u) • ES.. T(u) a.00 St = lim . for any c > 0 P( Further. (4. we need the following auxiliary result: Proposition 4. By Proposition 111. Later results then deal with more precise and refined versions of this statement. i. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. uoo u using e.s.mL > E T(u) < 00 ) 40. For the second .1) i. cf. P = /3µB > 1.3).h(u.
For (4 .h. T(u) < oo f / 00) e7uE L [e_7 (t1).r(u)/m T(u) ti µB2) Z.s.1. Notes and references Theorem 4. According to Anscombe' s theorem (e.3). which may be viewed both as a refinement of Theorem 4. this can be rewritten as u + 1(u) . PL (•)+ 0.1 is standard.1 (by considering 0(u.3.mu (2) '• m3/2 µB 7 .5) St . and as a timedependent version of the CramerLundberg approximation. proving (4.1).mu m . cf. apB ) .2 of [86]) and (4. again Proposition A1.7 6  11 Proof of Theorem 4.4. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.1).2. Theorem 7.g.6. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0.1) is T (u)  U mL P( T (u) < I > E. If Z .^ N (o.1). T) for T which are close to the critical value umL).2) follows immediately from u (4.6µB2) Z v m (3µB2) Z. and (4. the same conclusion holds with t replaced by r(u). of (4. implying T(u) .N(0. 1'r(U) . Thus. 4).1 The l. Tu) T( u) . though it is not easy to attribute priority to any particular author.t/m D (2) 111 . the result comes out not only by the present direct proof but also from any of the results in the following subsections. . note first that ( Proposition 111. T (u) < 00 J 0(u) e7'PL U \ I T u) .mL U > E.mL >E By Proposition 4.
with w2 as in (4.VU T.v. Then h(u) 4 h(oo) = E f (6(oo)).3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).um.5) For the proof. Proof Define u' = u .^(T(u')). we get E[ T (u) .ul/4. (oo.L+YWLV'U) .f ( (oo)) .ST( u') = u1/4 .))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . E9(Z) (4.)mu \ h(oo)Eg (r(ul) . we can replace T(u) by r(u'). e'°'/b (u.e(u') oo w . O . and similarly as above we get E[f(^(u)) I Fr(u. and thus in (4.112 CHAPTER IV. Hence Ef (Vu )) 9 (T(u.u1/4)I(S(u') > u1 /4) h(oo) + 0.6). then e(u) and r(u) are asymptotically independent in the sense that. PROBABILITY OF RUIN IN FINITE TIME Corollary 4.4 (SIAM'S LEMMA) If 71 < 0. Then the distribution of T(u) . oo).r.a C4'(y )• ( 4.l:(oo) (recall that rt < 0). resp . using that ul/4 . one has 9 (r(u)_rnu) Ef (^(u)) * E. Using ( 4.(u.t.T ( u')] = E[ T ( ul /4 .6) whenever f.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .w2) r. oo ). P because of ^(u') . Let h(u) = E f (^(u)).T(u') given F. S( u ) < ul/4] < ET(ul / 4) = O(ul/4). g are continuous and bounded on [0. letting Z be a N(0. Then for any y.mul h(oo)Eg(Z). we need the following auxiliary result: Proposition 4.4).3).
The present proof is basically that of Siegmund [342]. PL(T(u ) < umL + ywL) 113 4 C4(y). 0. e7v" y < ^'(7) (4 .T) Ce7"4 (T . that for the fit of (4. where we used Stain's lemma in the third step and (4. Segerdahl 's result suggests the approximation b(u. For refinements of Corollary 4.1.umL wI V"U u (4.5 '(u . see Asmussen [12] and Malinovskii [254].4. Theorem 4. yy by 1 K. .(ay) = 17 7y = ay . oo ) as u * oo. u needs to be very large).7) to be valid is that T varies with u in such a way that y(T) has a limit in (.yK(ay)• (4.dependent version of Lundberg's inequality For y > 0. ELe7E (") . 10) '5(u) . Notes and references Corollary 4 . Thus .4) in the last. CL Fig.3 in terms of Edgeworth expansions . WHEN DOES RUIN OCCUR? Proof of Corollary 4. 3 is due to Segerdahl [333]. however . y u) < . y > k'(7) .z/)(u . T(u) < umL + ywL f. 4b Gerber's time.7) to be good.7) To arrive at this . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). in practice one would trust (4. Cf.5) and solve for y = y(T). y u) < e 7v" .9) ( 4 . For practical purposes . just substitute T = umL + ywL in (4.3 ery"z/i(u .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . define ay. The precise condition for (4. also Hoglund [204]. see also von Bahr [55 ] and Gut [182].oo.
v"U. From the proof it is seen that this amounts to that a should maximize ayic(a). we arrive at the expression in (4.t.8 below . if y > 1/ic'(y).8).6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.7 i.6. dy) Notes and references Theorem 4 . An easy combination with the proof of Theorem 111. which may be understood from Theorem 4. yy is sometimes called the timedependent Lundberg exponent.114 CHAPTER IV.5. 5 is due to Gerber [156 ].yu ) = eayuEav [e . the point is that we want to select an a which produces the largest possible exponent in the inequalities. and hence t.Y' (u.3 yields easily the following sharpening of (4.h(u. For a different proof. which shows that the correct rate of decay of tp(u.b (u. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).9): Proposition 4. yu) is e 'Yyu/ . . yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. Hoglund [203] treats the renewal case.1). f Some urther discussion is given in XI. see MartinLM [257] . Numerical comparisons are in Grandell [172 ]. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. 0.ay4(u)+ T(u)K(ay ). Then ic(ay) > 0 (see Fig . and generalizations to more general models are given in Chapter VI.r. In view of Theorem 4. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). the bound a7y° turns out to be rather crude .2.8). u Differentiating w. who used a martingale argument. However. we have rc(ay) < 0 and get (u) . yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). a.
6 with P replaced by Pay and FL by Pay. the formula 0(u.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . (0) r1 (a) ' I. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . and ii(u) .12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. Ea .i(u.4. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. the choice of ay.'(y ). Using Lemma 111. if we want EaT(u) . and b(u.11) ' If y > 1/ r .ay a.. As a motivation.ay y 'Yay  ay . The traditional application of the saddlepoint method is to derive approximations..e. This idea is precisely what characterizes the saddlepoint method. (4. then the relevant choice is precisely a = ay where y = T/u. i.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.8 If y < 1/ic'(ry).e.c(&) = ic(ay) is < 0. then ay > 0. (4. Proposition 4. not inequalities. [eT(u )K( ay). yu) = e. We thereby obtain that T is 'in the center' of the Padistribution of T(u).^3 ]1/ Bay [lay .ayC() . WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. T(u) < yu] .z. T(u) suggests heuristically that l t/. yu ) eaauEaye . u 4 oo.(u.ayuEay f eay^ ( u)+T(u)K(ay). then the solution &y < ay of .yu) c ay .13) . For any a > yo. it is instructive to reinspect the choice of the change of measure in the proof.5.: T. (4. we have ryas = ay .2 yields EaT(u) u u r..
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 .7ruw2 Inserting these estimates in (4.ay + ayl /BLay] .l'B)y /(Pay . where V is normal(0. i B[7ay . (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 . Then ic(a) = .1)3 = (jB"[ay]l (Pay .13). The difficulties in making the proof precise is in part to show (4.c'(a) _ /3a/(8 .ay ) r.13) rigorously.(j (1 .1) . (4. and the equation ic'(a) = 1/y is easily seen to have .I ay &y a ^c'(ay) a (1 +. Example 4.1.9 Assume that B(x) = eay.(ay) _ y(ay . .a.1)3 = y3/3B"[ay]. a nr=. PROBABILITY OF RUIN IN FINITE TIME ry I i .11) follows. Writing r(u) and W2 = I3ay{.4).ay)K(ay) ay ayI&YI For the second term in (4.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.a)2 .1) under Pay mation (4.B[ay] /ay &y y(ay . it seems tempting to apply the normal approxiyu + ul/2wV.12) is 0 entirely similar.ay) ay +.116 CHAPTER IV. V < 01 Ir 00 er(ay)"1'2"'x eyur.a) . we get heuristically that Eay Ler (u)r(ay).13). The proof of (4.3(5/(S .
A related result appears in BarndorffNielsen & Schmidli [59].i )( v s vc ('3 + s _2 / . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. 2 = Var(Si ) the variance.f. (5.11) gives the expression '31/4 ( . is the drift and o./4 ^y for 1/i (u.p. ..= (s.g.because the c.5...1. 0 Notes and references Theorem 4. 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 =.8 is from Arfwedson [9]. yu) when y < 1/ic'('y) = p/1 .1) . DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is . in discrete time: if p = ES.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.tcp) Lo {Wo ( t)}t>0 . is undefined for a > 5). c a 00. 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.ay)3 0 3/2 and (4. then { __ .
St = EN` U= . cf.e.tp). + {Wo(t ) . n/c < t < (n + 1)/c. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.z } {W_1(t )}t>o (5...1) with S.2) t>o where p = pp = p . Indeed . p. of which a particular case is the claim surplus process (see the proof of Theorem 5. We want an approximation of the claim surplus process itself.1)) is inconvenient.1 below). It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).1 As p J. and this can be obtained under the assumption that the safety loading rt is small and positive.t} _ {W_1(t)} . for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.p/c < St(p) < S((n+l)/ c + Pp/c.1.a = Snp) and the inequalities Sn )C . where p is the critical premium rate APBTheorem 5 .118 CHAPTER IV. (5.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. we have o {i!t s: . and consider the limit p j p. 0 . This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III.7c).p. However. 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. Lemma 111.3) whenever c = cp f oo as p 1 p. such that the claim size distribution B and the Poisson rate a are the same for all p (i.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. oo)). a2 =/3µB2) Proof The first step is to note that { WC (St P) . Mathematically.3. this is an easy consequence of (5. Letting c = a2/pp.
2 As p j p. Corollary 5. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. 1. 119 It is wellknown (Corollary XI. has a continuous distribution. is 1/ip (ua2 /IpI. u) is defective when < 0. u) =PIT( (u) < x) = 1 . (ua2 To2 op \ IPI > IG ( T . and in fact some additional arguments are needed to justify (5. ulpl /a2) = e2"1µl / or2.5). (5.h.Ta2 /p2). 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.r.. 196.2 suggests the approximation u 0(u.s.6) from Theorem 5.t. we omit the details .4) Note that IG(.h.1. 263) that the distribution IG(•.6) This is the same as the heavy traffic approximation derived in III. 199. (. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. w. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.s. Because of the direct argument in Chapter III. any probability measure concentrated on the continuous functions. is IG(T.f I \\\ J \ (5.1.T) IG(Tp2/ a2). and the r. u). Corollary 5 . TS(u)=inf{t>0: WW(t)>u}.5) Note that letting T * oo in ( 5.8 or [APQ] p.e.1 I 7= .h.u)..(u) ti IG(oo. since ti(u) has infinite horizon . the continuity argument above does not generalize immediately. [169] or [APQ] pp. C.( ^ I + e2( \ I . However. For practical purposes .5.1 . (5. . we obtain formally the approximation V.. ulpI/a2). see Grandell [ 168]. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.s. ('.7c.
PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5.1 and Section VIII. on the premium rule involving interest. Assume further that 039µB6 < pe. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. that 00 4090. pe .r. For claims with infinite variance. [169]. the B9.5) for the compound Poisson model which does not require much more computation. Furrer. We conclude this section by giving a more general triangular array version of Theorem 5. .g. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. e.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. such that the Poisson rate Oe.6 of [APQ]. in particular for large u.00µB6 + 0. See for example Billingsley [64]. All material of this section can be found in these references. in Asmussen [12].5) and (5. B0 * Boo.. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. as an example of such a generalization we mention the paper [129] by Emanuel et al. and two further standard references in the area are Grandell [168].Po = 09µB6 .Pe. pt? 4 peo. The proof is a straightforward combination of the proof of Theorem 5. However.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. we have ^A. However.t. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. In contrast. the claim size distribution B9 and the premium rate p9 depends on 0. and which is much more precise. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . (5. Theorem 5. The first application in risk theory is Iglehart [207]. the simplicity of (5. In view of the excellent fit of the CramerLundberg approximation.120 CHAPTER IV. in the next subsection we shall derive a refinement of (5.6) are presented. The picture which emerges is that the approximations are not terribly precise.6) therefore does not appear to of much practical relevance for the compound Poisson model. Then as 0 _+ 90.1. as 0 * 00 and that the U2 are uniformly integrable w.
90] B(dx). The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.6. claim size distribution B . 77 = 1/p . For each 9. and we are studying b(u. In this setup. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. Let PO refer to the risk process with parameters e9oz Qo = QB[90]. 9o T 0. CORRECTED DIFFUSION APPROXIMATIONS 121 6 Corrected diffusion approximations The idea behind the simple diffusion approximation is to replace the risk process by a Brownian motion (by fitting the two first moments ) and use the Brownian first passage probabilities as approximation for the ruin probabilities. PB('r(u ) < oo) < 1 for 9 < 0. . Since Brownian motion is skip free. Then r. risk process with safety loading 77 > 0 correspond to 9 = 0 .Q (B[s] .1 > 0. 2. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). ./c(9 . this idea ignores (among other things) the presence of the overshoot e(u). 3.c(s) = . However . it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift).6.ao (0) _ /c(s + 9 . B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .90) and the given risk process corresponds to Poo where 90 = 'yo. which we have seen to play an important role for example for the CramerLundberg approximation . 0(0) = 0.4.s and p = /3µB < 1. In terms of the given risk process with Poisson intensity .90) . P9(r (u) < oo) = 1 for 9 > 0.T) = Peo(r(u) < T) for 90 < 0. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.1) .'(yo) = 0 and let 90 = 'Yo. 77 is close to zero. this is because in the regime of the diffusion approximation . this means the following: 1.9(s) = Ico ( s + 9) . whereas there we let the given 3B. Bo(dx) = B[eo]B(dx). and we want to consider the limit 77 10 corresponding to Oo f 0. Determine yo > 0 by r.
tu2 ) i IG (t. for brevity. (.3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u. u) 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 > 0.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. One has (6.1) . The corrected diffusion approximation to be derived is (u. write r = T(u)..T) 1+u2 (6. 9otc0" (0) = 0061 = ul. 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.7(u)/u2} eh(A.u. The first step in the derivation is to note that µ = k (0) = r0 (00) . C . (. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. u) = euh(a . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.Varo S1 = f30Eo U2 = S1. . C.. i.122 CHAPTER IV.. means up to o(u1) terms): . Theorem 5..(y) = 0. IGu+u2.S. bl IG(t81. C) = 2A + (2 . u) = IG(x/u2.e. (U.1) IG(x. 1) • Since L eatIG (dt. Vargo S.C.2) .() The idea of the proof is to improve upon this by an O (u1) term (in the following. (6. (01. 0o to.3) this implies (take u = 1) Ego exp { . () where h (A. the solution of r. _ ^(u) = ST . .
f.d. 6 . distributed as Z . The solid line represents the exact value .z .3). however.7. in (3) and (4). that whereas the proof of Proposition 6. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6. .'yu /2)(1 + b2/u)} + Aug 1I J . CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .1 + u2 I Indeed.exp { h(A.s. To arrive at (6. p = 0. we get by formal Laplace transform inversion that C 2 u.5) Once this is established . 9o T 0 in such as way that C = Sou is fixed. just replace t by Tb1/u2. is the c. 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.52/u where Z has distribution IG (•.s.1 As u + oo. of a (defective) r. we have p =. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).3.4. . 1. of (6. the r.3 = 0. which is based upon exponential claims with mean µB = 1. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . calculated using numerical integration and Proposition 1. . But the Laplace transform of such a r.v. A numerical illustration is given in Fig.2) is indeed o(u1). In ( 1) and (2).v. however .5) according to (6. bl I IG I t +2 .2).1 below is exact. Note. (6. and the dotted line the corrected diffusion approximation (6. 1% in (2) and (4).2 ). u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r.h. .6.ry2 .1 + 629.h.
() Lemma 6.00 0.111 W(U.07 0.T) 111 0.2 e..W21 0.08 0. For further numerical illustrations. PROBABILITY OF RUIN IN FINITE TIME 0. . (Inc 0s 0.1 proceeds in several steps.19)2 11 20 20 i0 T 1n0 Figure 6. The proof of Proposition 6..02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.aa1 . BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].0 0.TI CHAPTER IV.01 0. OM 0.OOIi O.124 0. the fit at p = 0.EB 0 p ex p ( 7 S h ^)u .u2 2u3 (e .114 0. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.7 or at values of Vi(u) like 1% is unsatisfying.1 It is seen that the numerical fit is extraordinary for p = 0.08 a. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.^) .05{ 0.(061 0.011 L1 60 T IM 11. A51 7(SAT 3 3 h(X.199 0.4 may not be outstanding but nevertheless.T) 0. see Asmussen [12].7. and all of the numerical studies the author knows of indicate that its fit at p = 0. Similarly.T1 00.1 W IU.
C2 = 2).. () 62 Eeo exp u u2 J . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.+ h (A.h.1) h(A.6.(3) J t _ aa1T l + eh(A. (6. 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.co (e) . (6..6) u U3 Lemma 6 . exp ue } al 1J 3 exP I [2). 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .4 Ea.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .. + a1b2 + . the result follows.s.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .4) that the r.(3)Eea LauT exp i 3J .61a2T (B3 .00)(u +C)  'r (. in Lemma 6.co ((/u)) } Let 8 = (2a + (2)1/2 = h(). () + C and note that 2 KO (0) = 102. 1 = PB(T < oo) = Eo0 exp 125 {(B . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .7) 2 2 .C)C/u ..2u (B3 .T (co (8/u) .
2.h.6  d h(A.2u [2A+ (2 3 . 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 6.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). C) ( 1+ u2 The result follows by combining Lemma 6 .() ..\+ (2 (3 e 2u [ (2. we get h(A.x.h (A. Thus by Taylor expansion around ( = 90u. yu/2). [2+ (2 . yu/2) h(A.(2A + ()1/21 exp S h(A. 2 and (6. l Lemma 6 . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2).4. 5 exp { _h(A) (1 + / y u J)) exp 1.s.2) for O(u) (indeed .7) and using eh(a.e h(aS)h (^^ 262 exp {_h(.2 (^/2 + 3y9o + 390) + O(u3). () .\ + () 1 2 / .1 (y/2 + Oo)u . Thus a2 y = 290 + O (u2). There are two reasons for this : in this way.() I 1 + u2 ) y .126 CHAPTER IV. 2 + 00 = . () by h(\. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6. 0 The last step is to replace h(A. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27. yields +90 62 0 + O(u 3) 2u2 +O(u 3).S) d e 62 .
. () (i+a ) 2A + (2 . ()} 3 h (A.e. HOW DOES RUIN OCCUR? exp { h (x. this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. ()} .1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A.T) has not been carried out and seems nontrivial. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. the 'typical' value (say in sense of the conditional mean) was umL.4. Fuh [148] considers the closely related case of discrete time Markov additive processes. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. and to the Markovmodulated model of Chapter VI in Asmussen [16]. the approach to the finite horizon case is in part different and uses local central limit theorems. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed.7.(i+ 62 exP{ h(A. () I 1 + u 2 ) } S 1 . the same as for the unconditional Lundberg process. i. 'yu/2) 127 ( i+ M pz^ exP { h (A.5 in Lemma 6. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. with the translation to risk processes being carried out by the author [12]. that is. 0 1 Proof of Proposition 6.1: Just insert Lemma 6. the analogous analysis of finite horizon ruin probabilities O(u. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. The adaptation to risk theory has not been carried out. The answer is similar: the process behaved as if it changed its whole distribution to FL. In Siegmund's book [346].
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. Theorem 7 . r(u) < oo) . F(u)c] ti e' ru]PL (F(u)`) > 0.ST(u)}t> o is just an independent copy of {St}t>o).3 to . the Poisson rate changes from . Then also P(u)(F(u)) + 1. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .3L and the claim size distribution from B to BL. In fact.FT(u) = o' (T(u ). . then P(u) and FL coincide on .T. so in the in the proof. we give a typical application of Theorem 7.t.(u)_ and similarly the denominator is exactly equal to Ce7u..128 CHAPTER IV. u * oo. (u) and satifying PL(F(u)) * 1.vi(u) Ce'Yu Corollary 7. Recall that .. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.(u). + TMOO ).r. stating roughly that under F(u).TT(u) _measurable.1 Let {F(u)}u>0 be any family of events with F(u) E F.r.(u)_ and ^(u) are independent . F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. P(u) and rate = aL w. . ^(u) is exponential with rate 8 w.F. .1.2 If B is exponential. In the exponential case. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property.t.(u) is not . Proof P(u) (F(u)c) = F(flu)c.EL[e7S.F. Note that basically the difference between FT(u) and . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .(u)_ is that i. FL As example. {ST(u)+t .
3 M(u) pcu) 1 . 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). the queueing results are of a somewhat different type because of the presence of reflection at 0. From a mathematical point of view.eALx) M(u) k=1 u The proof of the second is similar. This is currently a very active area of research. the subject treated in this section leads into the area of large deviations theory. take I(Tk < x) . Proof For the first assertion. Notes and references The results of the present section are part of a more general study carried out by the author [11].(1 .eaLx. . who also treated the heavytailed case. however. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. HOW DOES RUIN OCCUR? Corollary 7.3. see further XI.7.
This page is intentionally left blank .
the Tn are independent. D'2. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. In the socalled zerodelayed case.d. with common distribution B. U2.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. We use much of the same notation as in Chapter I. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).1. with the same distribution A (say) for T2. are i.. Proposition 1.(1.7). and M is the maximum of {St}.Q. r(u) the time to ruin. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA.. T3.i. AA t*oo lim St = lim ESt t tioo t = p . with Nt = # {n: Un <t} the number of arrivals before t. Thus the premium rate is 1.1). see A.. and the one corresponding to T1 = s by 1/i8 (u). . . the one corresponding to the stationary case by 00)(u). of the risk process form a renewal process: letting Tn = Qn . the distribution Al of T1 is A as well.1 (T1 = a1)..Then no matter the distribution Al of T1i B. . {St} is the claim surplus process given by I.... the claim sizes U1. .1 Define p = !µ.
t. and (1 . 3) follows similarly by Blackwell 's renewal theorem. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently).1) ENt/t + 1/µA. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. Thus. the definition 77 = 1/p .St] = a(p . such that no arrivals occur in the off state. 2). we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. A. Nt ESt = E E UI Nt t = ENt•pB . However .2 (DETERMINISTIC ARRIVALS) If A is degenerate. after E. say at a.1 of the safety loading appears reasonable here as well. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). t 4oo Proof Obviously. by the elementary renewal theorem (cf. For (1 .0 > 0. Here are two special cases of the renewal model with a similar direct interpretation: Example 1.1). If the environment is Markovian with transition rate A from on to off and u from OFF to ON.132 Furthermore for any a > 0. OFF. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. The renewal model is often referedd to as the Sparre Andersen process. RENEWAL ARRIVALS lim E [St+a .1) follows .Nt] * a/PA. but the arrival rate in the ON state is . s + t µA PA 0 Of course. From this ( 1. the .3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. stating that E[Nt+a . Nt + EVar U. CHAPTER V. Proposition 1. Example 1 .a is really the accumulated claims over a period u of length a.1 gives the desired interpretation of the constant p as the expected claims per unit time.
. the first term represents the probability F(U1 . we have From this the result immediately follows. For the stationary case.r. INTRODUCTION 133 interarrival times become i. integrate (1. S o<t<oo n=0.1.4) with phase space {oN.s < u).T between a claim U and an interarrival time T. (an arrival occurs necessarily in the ON state. (1.4) w. However. the relevance of the model has been questioned repeatedly. . A is phasetype (Example 1.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0. in general the mechanism generating a renewal arrival process appears much harder to understand.1. Ao.oFF}. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.. we feel it reasonable to present at least some basic features of the model. U1 . if for nothing else then for the mathematical elegance of the subject.i.4) fo Indeed. the fundamental connections to the theory of queues and random walks. and for historical reasons....1.d. u For later use.. Therefore. and then the whole process repeats itself). 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 .s. The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^. 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 . Proof The essence of the argument is that ruin can only occur at claim times.y)B(dy). and the present author agrees to a large extent to this criticism.2.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.t. More precisely. initial vector (1 0) and phase generator 11 However. arrival times.
At the time of death . are independent of {Nt} and i.3* (say ) and the U. the claims and the premium rate are negative so that the risk reserve process . If . U Figure 2.1. i. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. each of which receive a payment at constant rate during the lifetime . 2. then 0 * (u) = 1 for all u > 0.1) +ry. RENEWAL ARRIVALS 2 Exponential claims. resp . 00).0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . The initial reserve is obtained by prepayments from the policy holders. with common distribution B* (say) concentrated on (0. Theorem 2 . Using Lundberg conjugation . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. the remaining part of the prepayment (if any ) is made available to the company. b=1 !=1 where {Nt } is a Poisson process with rate .1 If.a*PB• > 1. < 1. St = t .Ut.d.1) .3* pB. the claim surplus process are given by Nt Nt Rt = u+^U.1 r* (u) One situation where this model is argued to be relevant is life annuities. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . t. A typical sample path of {Rt } is illustrated in Fig.134 CHAPTER V. (2. That is .
the safety loading of { St} is > 0. Define T_ (u) = inf It > 0 : St = u} . Proof Define 135 St =u . and the Lundberg conjugate of {St} is { St } and vice versa. Fig. 2. Let B(dx) = ^e7x B*(dx).Rt.(a) 7 Figure 2. (a) is*(a) (b) . If I3*pB* < 1. B* [7] and let {St} be a compound Poisson risk process with parameters . Then the c. Hence T_(u) < oo a. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. Hence y exists and is unique. St=Rtu=St. Then the function k* is defined on the whole of (oo.2(a). 2. . Since ic'(0) < 0.f. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).2 Assume now . of {St} is c(a) = is*(a7). B*. cf.2(b). B. then by Proposition 111.3*.UB. > 1 .(u ) < oo) = E {e7sr_ (u).g. T_ (u) = inf { t > 0 : St = u 'r* (u).s.* (a) = log Ee'st I. 0) and has typically the shape on Fig.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. and thus 1 = P(T..0.2. 0 Now return to the renewal model.1.
)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[. respectively. and 5PA > 1.. and (2 .1) + ry = 0 which is easily seen to be the same as (2. then .Tr+.4.2) 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 . 3* = 6.+Tn U1 Un.Ui .Un } = max St = t>0 n=0.• • • . To + max {Ul+•••+UnTI. with rate S (say).7r+ 7r EeTo b/(Sa) + +.Ti = U1.u+ and lr+.. According to Theorem 2. Hence M* max {To + Ti + • • • + Tn . RENEWAL ARRIVALS Theorem 2 .2).1.2 If B is exponential..'s and noting that V)*(u) = P(M* > u) so that Theorem 2.. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.Tn} n=0. However. the failure rate of this process is y.. we get Ee'M(d) = Ee°M* _ Y/(. 2.. alternatively termination occurs at a jump time (having rate 8).•. T2 = U2.1 means that M* is exponentially distributed with rate ry.Y] (2.. A variant of the last part of the proof.a) = 1 . the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 . Then B* = A. with the probability that a particular jump time is not followed by any later maximum values being 1 . u Hence P(M(d) > u) _ 1r+e'r". Taking m..136 CHAPTER V. To + M(d) in the notation of Proposition 1.1 it is seen that ruin is equivalent to one of these values being > u. 1) means that 8(A[ry] . Now the value of {St*} just before the nth claim is To +T1* +...f...e. and hence the failure rate .4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. and from Fig ..g.1.Y a I.1..
resp.4... : S(d) > u} . Bads (dx) = . 111.6. This follow since. the imbedded discrete time random walk and Markov additive processes. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. B^d) where Aad> (dt) = ^[ a] A(dt).. we see that ry = 6(1.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". However.3. The probability that the first ladder step is finite is 7r+.B(dx). which states that for a given a. 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^. a ladder step is the overshoot of a claim size.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] . Furthermore. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +.7r+). 0 3 Change of measure via exponential families We shall discuss two points of view. hence exponential with rate b.2. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. Hence the failure rate of M(') is 6(1 .. letting P(d) refer to the renewal risk model with these changed parameters .2.7r+) and hence r+ = 1. Putting this equal to y. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.5.y/b. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.
r.1). 7µA . E(d)e 1' (u).1 For any a such that k(d)' (a) > 0. VIII. (d) (7) _ 0. (a) '(u) < eryu. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0. 00)(u) . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . Proof Proposition 3. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .t. just note that F7(d) is nonlattice when F is so . 187) and thereby for ^(u) to be nonlattice w. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3.4. we get: Proposition 3.C(°)eryu where C(O) = C0[7] .3 For the delayed case Tl = s.1 implies Cu) = e«uE ( 7d)e«^(u) .p)/($B'[7] . 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 . in the easiest nonexponential case where B is phasetype.1) is explicit given 7.Ce"u where C = limu. For claim (b). We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.e.2 In the zerodelayed case.C8e7u where Cs = Ce78B[7]. (b) V)(u) .(u) .2 p. For the stationary case. Consider now the Lundberg case.. and claim (a) follows immediately from this and e (u) > 0. Corollary 3. In fact. ik. let 7 > 0 be the solution of r. cf..138 CHAPTER V. provided the distribution F of U .u the overshoot ..T is nonlattice. i .
dt ) eadt = h ( s) . 0 0 . 0) = ah (s) .a . y) = e°yh(s). The expressions are slightly more complicated and we omit the details..y) B(dy) 0 For the stationary case./c. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.4). E8h0 (Jdt. h(s) = e(a +x( a))8 (3. To determine boundary 0.0 ) = Eo[ha ( Jdt.9. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. 3b Markov additive processes We take the Markov additive point of view of II.dt(ah ( s) + h'(s)) so that Gha ( s. 0) = tc(a)h(s).1) (normalizing by h(0) = 1). Sdt) = h(s .St)} can be defined by taking Jt as the residual time until the next arrival. (u + s . we invoke the behavior at the 1 = h«(0. Let P8f E8 refer to the case Jo = s. IPA 0 Of course. St)} and h. Here K. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v).(s. (s.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).1) = C(O).5. we look for a function h(s) and a k (both depending on a) such that Gh. B(x) = o(e7x) and dominated convergence. delayed version of Lundberg's inequality can be obtained in a e7u.3. According to Remark 11. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . where G is the infinitesimal generator of {Xt} = {(Jt.h'(s). For s > 0.5.(°) ( Ce8B[7] Ao(ds) similar manner. 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] ..
however. RENEWAL ARRIVALS B[a]A[a .s governing {(Jt. (3..1)a in agreement u with Chapter III.s and P(d). J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk..S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .rc(a)] B[a] A[a .rc(a)] = 1. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. Aa as asserted . An easy extension of the argument shows that U1.e(«+k(a))t esy A(dt).2) As in 11.c(a)] which shows that U1.. .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. . we can now for each a define a new probability measure Pa.4 The probability measure Pa. .s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa. ] = E. resp.. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. since JT. [a + /3] A[b .. T2. T2 are independent with distributions Ba.140 CHAPTER V. A[a .5.s(Jo = s) = 1 follows trivially from Lo = 1. . Further. Ba where Aa (dt) . U.8.a . rc(a) = 0 (B[a] .13]A[b . .rc(a)] = B = Ba[13]Aa[5]. Ba(dx) = B(dx). .c(a)] B[a] Proof Pa.a .. Note that the changed distributions of A and B are in general not the same for Pa. 5 For the compound Poisson case where A is exponential with rate .2) means 1 = B[a]/3/(/3+a+rc (a)).e. i. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. Proposition 3. (3. An important exception is. resp. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.2). Remark 3 . = J8 = T2.tK( a)e.
Proof As in the proof of Theorem IV.4. the amount of .r.t. 2.ay+ray))TM(. [APQ] Ch.4.(u. Let M(u) be the number of claims leading to ruin .)+1 e J j e(ay+w(ay))8 e . and U„ the service time of customer n. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue.s eaysr(")+r(u ) K(ay) h (s) . not after time T. For the approach via Markov additive processes. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. for the zerodelayed case zp8(u. u The approach via the embedded random walk is standard. that is.y yuAa y [ay + K(ay) . Label the customers 1.1 and n. Using the Markov additive approach yields for example the following analogue of Theorem IV. 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. XII. The claim for the zerodelayed case follows by integration w. the time from he arrives to the queue till he starts service.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. The virtual waiting time Vt at time t is the residual amount of work at time t. or FCFS = first come first served) queueing discipline and renewal interarrival times.5 Proposition 3. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. and assume that T„ is the time between the arrivals of customers n .. yu) = F'ay.5.yx(ay). yu).g.yu ) e7vu A[ay . which is the same as the asserted inequality for 0. . see e. Then J(rr(u)) = TM(u)+1 and hence Ws(u. and define yy = ay . say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). yu ) < e7yu. it is easily seen that ic(ay ) > 0. that is. Then "^ e(ay+w(aY))8 Ys(u.. defined as the single server queue with first in first out (FIFO.4. . A(ds).
we have Wn = Van(left limit).3) Proof Part (a) is contained in Theorem 11. equivalently. the waiting time a customer would have if he arrived at time t. . Wn converges i n distribution to a random variable W.1. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. (4. in which case {V} remains at zero until time on+1.1.4: Proposition 4.1 and Corollary 11. the proposition follows.1 Wn+1 = (Wn + U.1) The following result shows that {Wn} is a Lindley process in the sense of II.4. and we have P(W > u) = V. but we shall present a slightly different proof via the duality result given in Theorem II. Let the T there be the random time UN.1). If W1 = 0.4.3. It then jumps to VQ„ . The traffic intensity of the queue is p = EU/ET. Thus. 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.• • • Tk ).. we get: Corollary 4.Tn)+. on + Tn) the residual work decreases linearly until possibly zero is hit. (4.142 CHAPTER V.3 Assume rl > 0 or.4): Proposition 4..4. and obviously z/'(u) = limN. then Wn v M. whereas in [On . since customer n arrives at time on."^ Vi(N) (u). Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. Vt converges in distribution to a random variable V.2.n1 (U1 +• • •+Uk Tl . an+1) = [on. Thus Vos}1 _ = (Wn + Un .Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .+ Un. p < 1. Then: (a) as n + oo. (u). equivalently. 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. (4.2) (b) as t * oo... and combining with (4.2 Let Mnd) = maxk=o. and we have P(V > u) = ?/iiol(u). 0 Applying Theorem 11. but interchanging the set .
Letting n oo in Corollary 4. namely W1 = 0 in (a) and Vo = 0.2). It follows that P(WN > u) =. resp . i.4 The steadystate actual waiting time W has the same distribution as M(d). TN) with (TN. by an obvious reversibility argument this does not affect the distribution .. (4.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. Corollary 4. u Now return to the Poisson case .4.(N)(u) has the limit tp(u) for all u. Ti) and similarly for the U.Ao in (b).T* are independent and distributed as U1.y)F(dy).5) Proof Letting n .5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).e. (4. In fact .. T] form a stationary renewal process with interarrival distribution A.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.T*)+ < x) = P(W + U* .. For part (b). hence (since the residual lifetime at 0 and the age at T have the same distribution . { 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).. Then K(x) = J x00K(x . conditioning upon U* .T* = y yields K(x) = P ((W + U* .le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . Then the arrivals of {Rt} in [0. T1. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. which implies the convergence in distribution and (4. K(x) = P(W < x). T1 . cf.. However. Hence for x > 0.T*)+. we obtain: Corollary 4.. Then the corresponding queue is M/G/1. x > 0..oo in Proposition 4.2. and we get: .. we let T be deterministic . where U*. THE DUALITY WITH QUEUEING THEORY 143 (T1.1. A.. as WN. we get W = (W + U* .. convergence in distribution hold for arbitrary initial conditions .
Cohen [88] or [APQ] Ch. W v V.g.5) to hold for all x E R and not just x > 0). That is. Proof For the Poisson case. Asmussen [24] and references there.g. see e. the actual and the virtual waiting time have the same distribution in the steady state. 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. the zerodelayed and the stationary renewal processes are identical. Hence '(u) = Ali(°)(u). implying P(W > u) = P(V > u) for all u. RENEWAL ARRIVALS Corollary 4.6 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 .5) looks like the convolution equation K = F * K but is not the same (one would need (4. The equation (4. VIII).5) is in fact a homogeneous WienerHopf equation. Some early classical papers are Smith [350] and Lindley [246]. .144 CHAPTER V. Note that (4.
Thus. {St} denotes the claim surplus process. • Claims arriving when Jt = i have distribution Bi.(3i when Jt = i. The intensity matrix governing {Jt} is denoted by A = (A. As in Chapter I. Ire = 1. dv. and can be computed as the positive solution of WA = 0. here it exists whenever A is irreducible which is assumed throughout. N.f pi. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). t St = E Ui . .T) = Pi (T(u) < T). • The premium rate when Jt = i is pi. 145 Oj( u. {Jt} describes the environmental conditions for the risk process.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 . i=1 0 and r(u) = inf It > 0: St > u}. M = supt>o St.)iJEE and its stationary limiting distribution by lr..
which is clearly unrealistic. f3i and claim size distributions Bn. leading to E having two states n.5 below. the operational time argument given in Example 1. one expects that 3i > on and presumably also that Bn # Bi. we shall assume that pi = 1.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. Thus. and assume that weather conditions play a major role for the occurence of accidents.a('). say. cf.. An example of how such a mechanism could be relevant in risk theory follows. T(=)). u . respectively. and we have f3. /3 = Nn when j E E(n).1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. we can approximate A(i) with a phasetype distribution (cf.T(n)). We let p Pi = /ji/AB. = iii when j E E(i). Example 1. r^ = P (1.11 below.2.14. 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. Unless otherwise stated. t(i) = T(')e are the exit rates. Proposition 1. cf. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). in blockpartitioned form. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. Example 1 . with rates Ani and Ain. For example. this is no restriction when studying infinite horizon ruin probabilities. According to Theorem A5.4) with representation (E(i). we could distinguish between normal and icy road conditions. a(i). P = E 7riPi. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. and p is the overall average amount of claims per unit time. i and corresponding arrival intensities Qn. Bi. Then the state space for the environment is the disjoint union of E(n) and E(i).1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.146 CHAPTER VI. say.1 Consider car insurance. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. assume that the sojourn time in the icy state has a more general distribution A(i).
1. where W = (w.. This amounts to a family (A(")) ?CH Of sojourn time distributions.1.j = . we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. i ) : n E H. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). iq (visited in that order) and letfOil >. Qi = ..5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. w. resp.. but assume now that the arrival intensity changes during the icy period. T(9) +wggt(9)0.tEH is a transition matrix. one could for example have H = {i1.3 Consider again the alternating renewal model for car insurance in Example 1. A('^).. Then for example wi.3i/pi. and similarly for the normal period.. and 1/ii(u) = t/ii(u). T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) . i E E(n) }. One way to model this would be to take A(') to be Coxian (cf.. u Example 1.. n8}.. . Approximating each A('?) by a phasetype distribution with representation (E('l).. and .a(n).>. i8f n1. 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. is the probability that a long icy period is followed by a short normal one. it = Jel(t).Q.3. the parameters are ^ij = aid/pi. dt. depending only on 77. u From now on..2. 1 . (9) where q = CHI. say it is larger initially. . The simplest model for the arrival intensity amounts to .T(n)). t(n) = T("i)e.n..3i. St = SB=(t). Example VIII.p. Indeed. MODEL AND EXAMPLES 147 Example 1 . In the car insurance example.J017.. such that a sojourn time of type rt is followed by one of type c w. let T 9(T) = f pi. u Example 1 .4) with states i1. say. the state space E for the environment is { ('q.
MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion.7 The Pidistribution of T1 is phasetype with representation (ei. )3*. o = 0. Pi (Ti E dx. Note also that (as the proof shows) 7ri/3i//3* gives the proportion of the claims which are of type i (arrive in state i). the empirical distribution of the claims is B*. dx.e(A(Oi)d'sg)xe.(Qi)diag)• More precisely. JT1 = j) = Qj • e.5. vi(dx) = . iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.148 CHAPTER VI. B* = 1 /^* Bi. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. t l=1 Note that the last statement of the proposition just means that in the limit. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation.A . one can associate in a natural way a standard Poisson one by averaging over the environment. In particular. we put )3* = E 7fi/3i.(3iBi(dx). Nt Nt a . A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. Next we note a semiMarkov structure of the arrival process: 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. qij = 0 in the notation of Chapter 11. N > 1(Ul < x) a4 B*(x). More precisely. .8 As t oo. . Proof The result immediately follows by noting that T1 is obtained as the lifelength of {Jt} killed at the time of the first arrival and that the exit rate obviu ously is f3j in state j.
^j 7riNi. {St} to the compound Poisson model with parameters 0 *. given {Jt}0<t<0. Conditioning .aA . Bi(x). 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. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. e. Bi.4. However . i. aA.. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. Hence Nt'> a . By Proposition A5.6. In particular. Bi.(/3i)aiag). the limiting distribution of the first claim size U1 is B*... oo) as a 4 oo.1. B*..x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). denoting the sizes of the claims arriving in state i by U(') 1 standard law of large numbers yields U(') the N 1: I(Ukik < x) k=1 N a$. A. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. Proposition 1. this converges to the exponential distribution with rate 0* as a * oo. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i .* (u) for all u.2. Then {St)} + {St*} in D[0. cf.7. Nt a' t t iEE Also. zli( (u) . and furthermore in the limit JT. y Ni) i Nti) t a. In particular. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. Proof According to Proposition 1. Then it is standard that ti lt '4' iri as t > oo.. Example 11. has distribution (7ri)3i //3*)iEE and is independent of Ti.9 Consider a Markovmodulated risk process {St} with param eters Ni. and let {St °i} refer to the one with parameters Pi.
2..s = o in state 1 and with intensity 1 . 132=2.2.. Continuing in this manner shows that the limiting distribution of (T. From this the convergence in distribution follows by general facts on weak convergence in D[0.FT. 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 = .2}. which also yield O(a) (u. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. The fact that indeed 0(a) (U) 3 0* (u) follows. the company even suffers an average loss. those of type E7 with intensity z s = 5 in state 1 and with intensity z .10 Let E_{1..1 with periods with positive drift alternating with periods with negative drift... On Fig. 0 Example 1.l3* and U2 having distribution B*. Claims of type E3 arrive with intensity 2 . from Theorem 3.. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). oo). resp. U. That is.1.1 of [145]. U2) are independent of . 1.. B1=3E3+2E7. e. T) for all u and T. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.2 +2 2 = 3. B2=1E3+4E7. > 1. state 1 appears as more dangerous than state 2. s 5 in state 2. Thus. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. marked by thin. with T2 being exponential with rate . Computing the parameters of the averaged compound Poisson model.. is as in {St }. there are p = 2 background states of {Jt}.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. lines in the path of {St}. thick. shows similarly that in the limit (T2. 1. T) + ?P* (u.g.=1.150 CHAPTER VI. and (at least when a is small such that state changes of the environment are infrequent)..s = 1o in state 2. MARKOVIAN ENVIRONMENT upon FT. 9 . we first get that 3 (3* = 2. the paths of the surplus process will exhibit the type of behaviour in Fig.).
3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. (b) St/t * p . 0 The definition (1.1. the averaged compound Poisson model is the same as in III.1). note first that EN Uk')/N a4' µgi.8. iEE . MODEL AND EXAMPLES 151 Figure 1. t + oo.1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. For (b).s.1. = P. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a)..1 a.(3. That is.1 Thus. 01 /. Proof In the notation of Proposition 1. a fraction r. t * oo.11 (a ) ESt/t * p .
X3... MARKOVIAN ENVIRONMENT Corollary 1. See Meier [258] and Ryden [314]. The case 77 > 0 is similarly easy. and hence M = 00. having the Pidistribution of X.12 If 77 < 0... Then by standard Markov process formulas (e. [302]. s.s. Since the X„ are independent . 136 or A. Proposition 1.Jt=i}. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model.Jt=i}. 38) Eiw1 = 1/ir. If 77 > 0..ld. and . [APQ].1 jEE = (p .1 and the Corollary are standard. w2=inf {t>w1:Jt_#i. Theorem II. n n Thus {SWn l is a discrete time random walk with mean zero. some early studies are in Janssen & Reinhard [211]. Now obviously the w. see the Notes to Section 7.1 of St / t is > 0. [212].1)Eiw = 0. limit p .. . and involves a version of the .4. .a form a renewal process . The proof of Proposition 1. + Xn SWn ](1 a .\ i and EiX1 Ei f 13 J. with X2..152 CHAPTER VI. dt . see [APQ] p.g. and so on. PB. Statistical aspects are not treated here. There seems still to be more to be done in this area. and a more comprehensive treatment in Asmussen [16].Eiw o'o Eiw • E ^ifjµs.s. then M < oo a. EiX = 0. [315]. The mainstream of the present chapter follows [16]. Proof The case 77 < 0 is trivial since then the a. also + . and hence 1/ii(u ) = 1 for all i and u. and hence wn /n a4.1(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. and hence oscillates between 0o and oo so that also here M = oo. X2 =SW2 So..2(a) p. Now let r) = 0. In risk theory.. Eiw. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. X 1 =Sty. let some state i be fixed and define w=wl=inf{t >0:Jt_#i.. then M = 00 a. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.
4) we obtain the following result . (y.EA. see also Example II. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. For measurevalued matrices. Proposition 2. •)• kEE Also. However .i.Jr+ =j. IIG+ II denotes the matrix with ijth element IIG+(i. j.6. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. j.5.. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. 6. by specializing results for general stationary risk processes (Theorem II . . let G+(i. j.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. dx)/jBj(y . Let further R denote the preT+ occupation kernel. Thus. Proposition 2. oo)). •). cf. •) II = JG+(i. only with the product of real numbers replaced by convolution of measures.dx). 00 (2. the definition of . j.2 (a) The distribution of M is given by 00 1 . G+ is the matrix whose ijth element is E G +(i. k.x).j E E. we get the same ladder height distribution as for the averaged compound Poisson model.3*B *(y)dy.A) =ZI(St E. B* in Section 1. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). •) * G +(k.g.j.A) = Pt(ST+ E A. e. j.1) 0 (b) G+ (y. oo) = J ao 0 G+(i.(u) = Pi(M < u) = e' E G+ (u)(I .Jt=j)dt. for i.2. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. That is. we define the convolution operation by the same rule as for multiplication of realvalued matrices.IIG +II)e.x.j. oo)) = f R(i. The form of G+ turns out to be explicit (or at least computable).6*. but is substantially more involved than for the compound Poisson case .a/i.2) R(dx)S((y . n=0 (2. T R(i.1 irG+(dy)e =.
and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.6. see Figure 2. hence uniquely specified by its intensity matrix Q (say). and that the environment is j at the nth when we start from i is e . lines in the path of {St}. III to bring R and G+ on a more explicit form . The u proof of (2.3) We let {St*} be defined as {St}. JJ = j.3. That is. St < S* for u < t. To this end . marked by thin. To make Proposition 2.IIG+II)e.1 The following observation is immediate: Proposition 2.154 CHAPTER VI.1) follows by summing over n and j.2 useful . the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2.1 for an illustration in the case of p = 2 environmental states of {Jt}. thick. mx = j when for some (necessarily unique) t we have St = x. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I .3 When q > 0. From this (2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). we need to invoke the timereversed version {Jt } of {Jt} . MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. 0  x Figure 2. resp. {mx} is a non terminating Markov process on E. . we need as in Chapters II.2) is just the same as the proof of Lemma 11.
(/3i) diag. An excursion of {St*} above level x starts at time t if St = x. corresponding to two subexcursions of depth 0.2.0. Furthermore. and the excursion is said to have depth 1 if each of these subexcursions have depth 0.2 where there are three excursions of depth 1. For example the excursion of depth 2 has one subexcursion which is of depth 1. and the excursion ends at time s = inf {v > t : S. s].and a jump (claim arrival) occurs at time t. ( Q( n)) converges monotonically to Q. In general. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.2 . Figure 2. The definitions are illustrated on Fig. 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). the sequence {Q(n)} A* defined by Q(O) = .*. we say that the excursion has depth 0. If there are no jumps in (t. = x}. 2. Q( n+l) _ ^. Proof The argument relies on an interpretation in terms of excursions. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.. Otherwise each jump at a minimum level during the excursion starts a subexcursion.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .(/3i)diag + T S(dx) eQx. } is a minimum value at v = t. {S. 0 mms1   ^O \ T.2.
Similarly by induction . A) = L' U(j.156 CHAPTER VI.St <S*.5) A (note that we use A = {x : x E Al on the r.6) . Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . A). By considering minimum values within the excursion.5 R(i.4) To show Q = cp(Q).Qi + )%pij) Now just note that t pij and insert (2. mx+dx = j) occurs in two ways . 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) = \!. either due to a jump of {Jt } which occurs with intensity A= j.s. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. (2. or through an arrival starting an excursion terminating with J.. Theorem 2 .T+>t) _ ^iF 7ri (JJ =i. 0)). A) = f Pi(mx = j) dx eie4xej dx A u (2.4). Then a jump to j (i. = j. 7rE Proof We shall show that Fi(Jt=j.u< t). p1^) Define a further kernel U by f U(i. Suppose mx = i. It follows that qij = A. of the definition to make U be concentrated on (co. Q = W(Q) follows. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . StEA .St EA.j +/3ipij. h. j. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). Writing out in matrix notation . e.j.. i.(01)diag = Q. Similarly. we first compute qij for i $ j.
St < St U.1 can be rederived using the more detailed form of G+ in Corollary 2.6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2. Jt = i.t... and get irPi(Jt =j..2. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.Jo=i. (b) for z > 0. and this immediately yields (2.Qi)diag. 0 < u < t) = 7rjPj(Jt =i. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx)..StEA. From Qe = 0. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. St E A. G+((z. 0<u<t).St EA..g. 0 +1) = cp (K( n)) defined by K(o) = A . oo))dx.St <Su.0<u<t) = P. We may then assume Ju=Jtu. x < 0.r. S.4]. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. St EA.6 (a) R(dx) = eKxdx.6).7 It is instructive to see how Proposition 2. the CramerLundberg approximation (Section 3). and to obtain a simple solution in the . dt. consider stationary versions of {Jt}.0<u<t. (Jo = j. To this end.S„<0. K( n (d) the sequence converges monotonically to K. e. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).(Jt=j. we shall see that nevertheless we have enough information to derive. {Jt }.(.z+>t) = P. and we let k be the corresponding right eigenvector normalized by Irk = 1. Remark 2.6 is hardly all that explicit in general.=StSt. oo)) = f o' eIXS((x + z. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.
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
9) For the proof. However. we let G+ * W(u) be the vector with ith component E(G+(i. r(u) yu o)(y)eavuEav.5). Our next objective is to improve upon the constant in front of a7u in Lundberg's inequality as well as to supplement with a lower bound: Theorem 3.11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). oo)) and.7. if y > 1lk'(ry).j) * coj)(u) _ f u ^Pj(u . 1 Similarly. exp {e() + r(u))} . yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .5) will produce the maximal ryy for which the argument works. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). (u.i [eT(u)K(av ). yu) f h(av)e v avuE«v. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. for a vector <p(u) = (cpi (u))iEE of functions .00 su e7( ( 3. hj P .(ay)}. dy)• o iEE jEE . av 'i [h..3.i I (a) exp {aye(u) + r(u)r. (3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.y)G+(z.j.8 ) Then for all i E E and all u > 0.V)i(u. as in the classical case (3. 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. we have ic(ay) < 0 and get 'i(u) .i [e*(u)K(av). Chie ryu < Vi(u ) < C+hie 7u. we shall need the matrices G+ and R of Section 2.
dy).3jhj // f 00 R(i. we have G *(N +1) * ^. j.166 CHAPTER VI. _ To see that the ith component of U * G(u) equals ?Pi (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. dx) 100 C .(0) ] (u) < sup Jt t. j.12 Assume sup1.x) x) jEE 0 E Qj f jEE R(i. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. 00 Thus C+ > hj f"o e7(Yu)G +(i. dx ) Bj (u . j. dy) 00 C+ ijhj f R(i. dx) f e7( vu)Bj (dy .x) jEE 00 u 0 //^^ C+E. dy) = aj f Bj(dy .j. j. 0 G+(i. 00 f C_ hj f e(Y)G+(i. Lemma 3 .7. if r+ (n) is the nth ladder epoch. n > oo. jEE u 0 j. MARKOVIAN ENVIRONMENT Lemma 3 . dy) : 1(u) < C+ > hj u e(1tL)G+(i. dx). = Eo G+ G.13 For all i and u.j. Hence lim cp(n) exists and equals U * G. Proof Write UN = EN G+ .x ) = Gi(u). U = U". Then cpin)(u) sit (u) as n + oo.& (u).ery(&u+x)Bj (dy) Bj(u Bj (u .x ) R(i. °O .u IMP:°) (u) I < oo.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).
12) Proof We first note that just as in the proof of Theorem 3. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. j. letting MT = maxo<t<T St.11) C_e7u 57 O+[i. from which the lower inequality follows by letting n * oo. it follows that Vi(u) < C_(yo) h=70)e7ou. jEE estimating the first term in (3. j. dy) jEE o (3. 9 for the last equality in (3. j. this is obvious if n = 0. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. and assuming it shown for n. (3. Then 0< Vi (u )  0i(u. +i . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. u The proof of the upper inequality is similar .13) Hence. 13 and the second by the induction hypothesis .M>u) = Ei [VGJT (u .MT<u. and let 8 = e'(70). we have Vii (u) .u)G+(i. We claim by induction that then cpin) (u) > C_ hie7u for all n. taking cps°) (u) = 0. j.3. y]hj = C_ e7uhi. and the proof of the lower one is similar. let C+(yo) be as in (3.8) with y replaced by yo and hi by h=7o ). T) = Pi (7. 167 u Proof of Theorem 3.10: Theorem 3 .13 Let first cp=°)(u) = C_ hie"u in Lemma 3. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . T) < C+(')' o)hi7u)e7ou8T .n) ( u .tpi(u. Indeed.Pi(MT > u) = Pi(MT < u. (3. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. dy) (3.11).(u) < T ) to 0i (u) which is different from Theorem 3.ST).10 ) by Lemma 3 .M > u) = Pi(ST<u.13.T) = Pi(M > u) . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.10) C_ 1 f hje7(y. MT < u.11. and using Lemma 3 .y)G+(i.
". 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u).5) Note that whereas (4. but that in general the picture is more diverse. this correponds to the usual stochastic ordering of the maxima M'. we define the stochastic ordering by 0' < s.o.. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.31:5)32 . The results to be presented show that quite often this is so.3) Bl <_s. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . < .o. For the notion of monotone Markov processes.0...3) to B = Bi does not depend on i. B2 <_s.. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). <s.33 or Bi 0 Bj. (4. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).2) (4.o.. Bp.1) Obviously. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. V)" if z/i'(u) <'c"" (u). The conditions which play a role in the following are: . we refer to . this is not the case for (4.2) alone just amounts to an ordering of the states. in part from the folklore principle that any added stochastic variation increases the risk. (4. Occasionally we strengthen (4.. we also assume that there exist i # j such that either /3i <. and finally in part from queueing theory..3p. u > 0.3). [177]. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. The motivation that such a result should be true came in part from numerical studies. It was long conjectured that 0* Vi.o.168 CHAPTER VI. (4. The Markov process {Jt} is stochastically monotone (4.4) To avoid trivialities. Further related discussion is given in Grigelionis [176]..
r(u x)dx.2.1 Assume that conditions (4.* For the proof. < a. Then V. 2 If al < . Proof of Theorem 4. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.2)(4.4) say basically that if i < j . Conditions (4. where 7r2+) = QiµBilri/p.4) is automatic in some simple examples like birthdeath processes or p = 2 . we obtain (cf... note that (4.2)(4. Proposition 2.7) 7ri. also Proposition 2.1 for the first term in (4.r (JT(o) = i.x) dx u o i =1 i=1 (4.r (Sr(o) E dx Jr(o) = i. ^i 7ri = 1.. Lemma 4 . Conditioning upon the first ladder epoch..x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.4..9) (4. = b. T(0) < oo) = Bi(x) dx/tcai .4) hold. = aP or b1 = . b1 < .. dx (4.. 3 (a) P.9) follows by considering the increasing functions 3iBi (x) and Oi (u . Section 4.10) Q*B*(u)+. . then j is the more risky state ..1.3iBi(x)YPi(u .13* J0 u 0*(u .7) and Lemma 4. p). Proschan & Walkup [140].1) which with basically the same proof can be found in Asmussen & Schmidt [49]. it follows by a standard .10) and (4.3 for the second) *(u) _ /3 *B* (u) +. 0 Here (4.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. Lemma 4 . 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = ..5 (cf. < bp and 7ri > 0 (i = 1.. 7(0) < oo) = pirf+).6. .2. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4. (b) P. we need two lemmas.6).. the second follows from an extension of Theorem I1.. then P P P 7rjbj.x) of i and using Lemma 4..x)B*(x) dx. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].9 ) below).3* f uB(x) z/^.8) ^j Tri/iBd(x) . Comparing (4.. (4. E 7r i Wi(u . Theorem 4 .
(0) = V. Using (4. MARKOVIAN ENVIRONMENT argument from renewal theory that tk..r (u ) fails for all sufficiently small e > 0.(0) < b *'(0) for e small enough.h.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . Frey.4) is essential (the present author conjectures it is)./3*. Then the l.s. (u) may fail for some u. 01 = 103.0. u Here is a counterexample showing that the inequality tp* (u) < V).1 and Proposition 4. 0. For u = 0. µB. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *.2. i=1 i=1 7'r(0) _ EFioiwi(0) .8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .. of (4..3i. Proof Since 0. it will hold for all sufficiently large u.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. Notes and references The results are from Asmussen. (4. except possibly for a very special situation . Recall that . dominates the solution 0* to the renewal equation (4.4 is the understanding of whether the stochastic monotonicity condition (4.3µi < 1 for all i. and from this the claim follows. As is seen. Rolski & Schmidt [32].4 is not vacuous. (u) is not in general true: Proposition 4.4 Assume that . that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.6).h. What is missing in relation to Theorem 4. µB2 = 104.s..6). = 102. let = ( 1/2 1/2 ) .11) is of order 104 and the r. they are at present not quite complete. of order 101.* (0). Q2 = 1.170 CHAPTER VI. u To see that Proposition 4.0*• i=1 But it is intuitively clear (see Theorem 3. Bi as e J.. Then i/i*(u) < . it is sufficient to show that 0'. this ruin probability is /3iPBi.
Then {(Jt.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0.ld) with generic cycle w = inf{t>0: Jt_54 k. e.)a. it follows by Proposition A1. (4.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.. 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] .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. It is clear that the distribution of X.(a) > 0 for all a 0 0. (4. with strict inequality unless rci (y*) does not depend on iEE. and by Proposition II..1) .g.a = E irirci(a).14) A„(O) iioo varXt t t By convexity. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.5 y < ry*.a. Asmussen [20]) as discussed in 11.13) implies A(a) > 0 for all a.13) (4. Hence if 5i 54 0 for some i E E. 0 . Xt)} is a Markov additive process (a socalled Markovian fluid model.4(b) that the limit in (4. Lemma 4. is nondegenerate unless bi does not depend on i E E.2 we have (Ei[e"X'.g. Proof Define X= f &ids. This implies that A is strictly convex. in particular . cf. Jt = i])' EE = vA+n(6.5.12) iEE Theorem 4. Further (see Corollary 11.5. with strict inequality unless a = 0 or bi = 0 for all i E E.4. (4.1) .14) is nonzero so that A"(0) > 0. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] .5. Now we can view {Xt} as a cumulative process (see A.
.5. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*.172 CHAPTER VI. a = 1 in Lemma 4. The corresponding adjustment coefficient is denoted by ry(e). Further a(1) = rc(y*) by definition of A(. Hence rc (y*) > 0. improving upon more incomplete results from Asmussen. If rci(y* ) is not a constant function of i E E. Hence letting e = 0 in (4.. Since ic is convex with rc'(0) < 0 . where A. MARKOVIAN ENVIRONMENT Proof of Theorem 4.15) once more and letting e = 0 we get .6. we have 7rh' = 0. note that y(a) + mins=1. the basic equation is (A + (rci(y))diag)h = 0. (4. h'(0) = (Ao . 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'.eir)h'(0).15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. we get rc (y*) > 0 which in a similar manner implies that u y < y*. Let bi = rci(y*).. Rolski & Schmidt [32].) and rc (•). Here we put a = 1/e. h depend on the parameter (e or a). Notes and references Theorem 4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h.12) and rc*(y*) = 0.16) Differentiating (4..15) Normalizing h by 7rh = 0.e7r)1 (Ici(Y*))diage. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. Then > risi = 0 because of (4.5 is from Asmussen & O'Cinneide [40]. (4. y.p yi and compute 8y 8a a=0 In both cases. whereas the . Thus y(e) * y* as e 10. Frey. h(0) = e. In the case of e.Qi and Bi are fixed .
THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . .18). Rolski & Schmidt [32]. The analogue of Proposition 4. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . 5 The Markovian arrival process We shall here briefly survey an extension of the model. . We assume that 0 < y < 7i.8 If (4.5. i = 2.8 when ryi < 0 for some i is open. 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. (4. Inserting (4. and may have some relevance in risk theory as well (though this still remains to be implemented).20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0)...19) holds. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. multiplying (4.17) by 7r to the left to get (4. Frey.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. (4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).18) 0 = 27'(0)p+27r(rs. and we have proved: Proposition 4. We get 0 = (aAo + ( lc&Y))diag)h. p.16) yields Proposition 4. (4.i(7' *))diagh'(0).20) Letting a = 0 in (4.17) (4. then 8a a=o All rci (0) Notes and references The results are from Asmussen. which has recently received much attention in the queueing literature.
d. and the marked transitions are then the ones corresponding to arrivals. A(1'k) A(2 k1).174 CHAPTER VI. the definition of Bij is redundant for i i4 j. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. A(1) = A .6i ) diag. with common distribution B. T). . Jt = (Jtl). A(l) = T. the definition of Bi is redundant because of f3i = 0. Here are some main examples: Example 5 . Jt2)) (2. For i = j. j(2) } be two independent environmental processes and let E(k).1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. Bij = B.2).(13i )diag. let { Jt 1) }. that Bii = Bi .4).2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . we use the convention that a1i = f3i where 3i is the Poisson rate in state i. where qij is the probability that a transition i * j is accompanied by a claim with distribution. Note that the case that 0 < qij < 1. II. Bii = Bi . the claim surplus is a Markov additive process (cf. u Example 5 . In the above setting. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. Thus . and thus 1i = 0. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. Indeed. Again . A ( 2) = A (2`1 ) ® A. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process.2 for details).^) etc. We then let (see the Appendix for the Kronecker E = E(1) x E(2). refer to notation) { Jt k) }. This is the only way in which arrivals can occur. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. A(l) = tv.i.2) A(1) = A(' 1) ® A(1. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. B.
where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired.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). 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 . the kth policy enters a recovering state. 11.}.. assume that there is a finite number N of policies. Similarly. more recently. THE MARKOVIAN ARRIVAL PROCESS Bij. claims occur only at state transitions for the environment so that AN2. i2i . Thus.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. as the Markovian arrival process ( MAP). • upon a claim.. E = { WORKING.iN = a2.5.. DIVORCED... possibly having a general phasetype sojourn time.iN C17 AilO.. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate .. Bilo. RETIRED. In this way we can model...iil. MARRIED. superpositions of renewal processes..iN. and that the policy then expires. The versatility of the setup is even greater than for the Markovmodulated model. Example 5 .iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. iN. However .. iN. E 10... e. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.. The individual pays at rate pi when in state i and receives an amount having distribution Bij when his/her state changes from i to j. u Example 5 . or.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij.iN. with rate ai. This means that the environmental states are of the form i1i2 • • • iN with il. say. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.iil. iN = all BOi2. WIDOWED.g.. all Al i2. after which it starts afresh.. INVALIDIZED.1i2 .iN are zero and all Bi are redundant....kl is redundant)..1i2. In fact . DEAD etc.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313].
Lucantoni et at. For the Markovmodulated model. Without loss of generality. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. Let 1 1 /3* _ f /3(t) dt. . Thus at time t the premium rate is p(s + t). )3 t 1 J (6. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. B* = J f B(t) ((*) dt.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . let the period be 1. [248]. from an application point of view. we may assume that the functions /3(t). we talk of s as the 'time of the year'. continuity would hold in presumably all reasonable examples. one needs to assume also (as a minimum) that they are measurable in t.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. Sengupta [336]. where i f00 xB(°) (dx) _ . We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Some main queueing references using the MAP are Ramaswami [298]. • The premium rate at time t of the year is p(t). a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . Neuts [271] and Asmussen & Perry [42]. for s E E = [0. MARKOVIAN ENVIRONMENT sense any arrival stream to a risk process can be approximated by a model of the type studied in this section : any marked point process is the weak limit of a sequence of such models . The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). 1). 1). Lucantoni [248]. Obviously.p)/p. p(t) and B(t) are defined also for t t [0. p * = 0 p(t) dt. 0 < t < 1.176 CHAPTER VI. By periodic extension. • Claims arriving at time t of the year have distribution B(t).3*µs • p = f /3(v) dv 0 0 (6.
1) and Example 1.1 As an example to be used for numerical illustration throughout this section. p* as an averaged version of the periodic model. u Remark 6 .8.3) Note that A enters just as a scaling factor of the time axis.6. (6. not random. in agreement with the general principle of added variation increasing the risk (cf. let .1. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . since the added variation is deterministic. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. We u assume in the rest of this section that p(t) . St = SeI(t). or.10.3(t) = 3A(1 + sin 27rt). In particular. the discussion in 111. Example 6 .(3. In contrast. In contrast. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. respectively. The behaviour of the periodic model needs not to be seen as a violation of this principle.3* = 3A.2 Define T 6(T) = p(t ) dt. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. It is easily seen that . of the periodic model as arising from the compound Poisson model by adding some extra variability. Section 4b). we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.t.w(t). and thus the averaged standard compound Poisson models have the same risk for all A. it turns out that they have the same adjustment coefficient.w(t)) dt). Thus.9). The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . Thus . equivalently. B*. Many of the results given below indicate that the averaged and the periodic model share a number of main features. the average compound Poisson model is the same as in III. one may think of the standard compound Poisson model with parameters 3*. p(t) = A and let B(t) be a mixture of two exponential distributions with intensities 3 and 7 and weights w(t) _ (1 +cos27rt)/2 and 1 . the conditional distribution . 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.
g.1) a = J8 . we obtain E. The claim surplus process {St} may be seen as a Markov additive process.(3(s + t)dt)e«St adt + /3(s + t)dt .1]) ..al.a. we start by deriving formulas giving the m. t + dt] or not.east B(8+t) [a] east .8).5 (see in particular Remark 11.tc* (a)] dv then h (.3(v)(B(vl [a] . a) etw*(a) h(s+t. with some variants in the proofs. a) is periodic on R.a be the c. of the averaged compound Poisson model (the last expression is independent of s by periodicity).Q(v) (B(„) [a] .f.^8 [. The exposition of the present chapter is basically an extract from [44]. As usual. (6. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).(8) [eaSt+dt I7t] = = (1 .g. Jt = (s + t) mod 1 P(8) . Notes and references The model has been studied in risk theory by. with the underlying Markov process {Jt} being deterministic period motion on E = [0. e.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. To this end. [101] .s .1) dv .g. r(u) _ inf It > 0 : St > u} is the time to ruin . 6b Lundberg conjugation Motivated by the discussion in Chapter II. but it turns out to have obvious benefits in terms of guidelining the analysis of the model as a parallel of the analysis for the Markovian environment risk process.a) = exp { . [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.a) Proof Conditioning upon whether a claim occurs in [t. i.e. 1).4) At a first sight this point of view may appear quite artificial. let f 8+1 tc *(a) _ (B* [a] .5. .adt +.. J Theorem 6 .1) .T) = P(8)(r(u) <T). and define h(s. 0 (5)(u. Daykin et..(1 .a . of the claim surplus process.f. 3 E(8)eaSt = h(s. see the Notes to Section 7).3(s + t)dt[B(8+t)[a] . given that the ith claim occurs at time t is B(8+t). Dassios & Embrechts [98] and Asmussen & Rolski [43].
1)dv l og E(8) et where atetk•(a) h(t.t}t>o = h(s. 0) P(8)a.s. a) . a).0(s + t)dt[B(8+t)[a] . at + f log h(s + t. we can write Lo Jt.1]) . so that obviously {Lo.1]) . B) eoSt t.5..3(s + t)[D(8 +t)[a] .1)dv  o h(t.4). St)} . a) Corollary 6. According to Remark 11. dt log E(8)east a + f3(s + t) [B(8+t) [a] . a) et.adt +. 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.(e) Let = h( h(Jo.t = 1 by Theorem 6. u Remark 6. a) = exp I f t3(v)(kv)[a) . it then suffices to note that E(8)Le.3. 0) exist and are finite.(8)east 179 = = = = = E(8)east (1 . St)} and .4 For each 0 such that the integrals in the definition of h(t . a) = h(s.t.9 as follows. With g the infinitesimal generator of {Xt} = {(Jt. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.log h(s. a) Thus E(8)east = h(s + t.c* (e) {Le. 9) is a P ( 8)martingale with mean one. Proof In the Markov additive sense of (6. + v)(B([a] .6. E (8)east (a +.t} is a multiplicative functional for the Markov process { (Jt.9) eastt.5 The formula for h(s) = h(s.1].6 . a) h(s + t. h(s + t.2.
3(v)( Bi"i [a] . That is.y) = eayh(s).T. Lemma 6 .180 CHAPTER VI. yo is determined by 0 = k* (70) = QB*.(3(s)dt) +.1) . 0) = h(s) + dt {ah(s) . Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .3. cf. That is. ( iii) use approximations with piecewiese constant /3(s). it follows by Theorem II.3(s)B(s) [a]h(s). [70] .2. For each 0 satisfying the conditions of Corollary 6.g. correspond to a new periodic risk model with parameters ex . Proposition 6. y solves n* (y) = 0. as above E (s) ha(Jdt.60(t) = a(t)B(t)[0]. Proof (i) Check that m. .3(s)dt • B(s)[a]h(s) = gha(s. of St is as for the asserted periodic risk model.0) = Kh(s). St)} with governing probability measures Fes).1. Now define 'y as the positive solution of the Lundberg equation for the averaged model. MARKOVIAN ENVIRONMENT ha(s. see [44] for 11 a formal proof. Bet)(dx) = ^ B(t ) (dx).a . Proposition 6.3(s)h(s) + h'(s) +. However. J s [. Sdt) = h(s + dt) eadt (1 . we put for short h(s) = h(s. the requirement is cha(i. A further important constant is the value yo (located in (0. When a = y. P(s) (T(u) < oo) = 1 for all u > 0. ry)) at which n* (a) attains its minimum. such that for any s and T < oo.6 The P(s). B(s).5 that we can define a new Markov process {(Jt.4. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.6 ( s ) exp { 0( s )&s) [a] + tc .f.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. 0 < s < 1.7 When a > yo.tc] dv} (normalizing by h(0) = 1). (ii) use Markovmodulated approximations (Section 6c). say.'y). (iv) finally. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.
9(u))} u>0. T) = h(s.2).8) (6.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. 1).9 Assume that there exist open intervals I C [0.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1.10) Then for each a. Corollary 6. ^(u) = ST(u) .g.1. q) = eryx/h(q)). 9(u)) for any bounded continuous function (e. we get: . considered with governing probability measures { E(8) }E[ . a) a > ry0 (6. have components with densities b(8)(x) satisfying inf sEI. Here and in the following. and no matter what is the initial season s.4. and we refer to [44]. The proof involves machinery from the ergodic theory of Markov chains on a general state space. a) e«uE(8 ) e «^ . The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. xEJ 0 (s)b(8)(x) > 0. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. s E I. we need the following auxiliary result . RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. a) TI h(9(u). (6. which is not used elsewhere in the book. J C R+ such that the B(8). e(cc)) Letting u > oo in (6. the mean number of claims per unit time is p« 181 = Jo 1. f (x.7) h(B(u). has a unique stationary distribution.6. Lemma 6 . 0(u)) * (b(oo). Wu). say s0. the Markov process {(^(u).9) 0(')(u) = h(s.1) the distribution of (l: (oo). T(u) < (6.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. B(oo)).2.9) and noting that weak convergence entails convergence of E f (^(u). u which is > 1 by convexity.
10 shows that certainly ry is the correct Lundberg exponent. (6. where e. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. it does not seem within the range of our methods to compute C explicitly. Theorem 6 . Among other things. illustrating that the effect of seasonality increases with A. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.16. MARKOVIAN ENVIRONMENT Theorem 6. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .1.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.10) of Lemma 3. Noting that ^(u) > 0 in ( 6. 1.182 CHAPTER VI. 6. At this stage .9). this provides an algorithm for computing C as a limit. where C(o) = 1 + info < t<i h(t) . A=1/4 A=1 A=4 0 Figure 6.11) Note that ( 6.Ch(s)ery".1. For our basic Example 6 .W.10 Under the condition (6.ir) } Plots of h for different values of A are given in Fig.6 for the Markovmodulated model: Theorem 6 .) C = E1 h(B(oo)) u + oo. Vi(8) (u) . 11 7/'O (u) < C+°)h(s) ery".1 In contrast to h.
3x + (1 .w ) • 7e u{w • 3e3x + ( 1 .7x j dx _7x } _ 6w + 6(1 . we obtain Co) = 1.13 to our basic example.4.42 so that 183 tp(8) (u) < 1. T) and replace the Lundberg exponent ry by ryy = ay . Consider first the timedependent version of Lundberg's inequality.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. we substitute T = yu in 0(u. C_h(s)e7u < V. whereas ay < y. r. yu) 000 (u) .11 as well as it supplements with a lower bound.12 Let 00)(y) 1 Then info < t<i h(t. in our basic example with A = 1. ay) • (6. e7 ( yx)B(t)(dy) > Then for all $ E [0. #c( ay) < 0 when y > 1/tc'('y). the proofs are basically the same as in Section 3 and we refer to [44] for details.17) (6. (ay). 1 (6.w)e4u dx 9w + 7(1 .167r I Cu. Theorem 6 .7e .42 • exp {J_ cos 27rs .(8) (u..(s)(u) < C+h(s)e7". Lundberg's inequality can be con siderably sharpened and extended. 1 ) and all u > 0. e.g.6. (6.w)e4u .w) . 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. RISK THEORY IN A PERIODIC ENVIRONMENT Thus.16 In order to apply Theorem 6.(ay) > 0 when y < 1/ic' (7). We state the results below. Theorem 6.yr. Just as in IV.13 Let = 1 B(t) C o<tf i h(t) 2no f °O e'r(Yx)B( t) (dy)' (x) x 1 B(t) (x) C+ = sup sup o<t<i h ( t) xo J. we first note that the function fu° exu {w • 3e .47r sin 27rs + 167r cos 47rs .0(8) (u+ yu) (6.14) < C+)(y)h(s) e7yu.12) As for the Markovian environment model. .
where the environment at time t is (s + t) mod 1 E [0.16. Finally.19 I eu. .. yo). completing a cycle .16) with 'y replaced by yo and h(t) by h(t. .. 14 Let C+('yo) be as in (6. 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.013.I eu.1 sin 27rs + 1 cos 47rs . 1/i18 1 s (u) > 0.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e. Then 7oudT .T) < C+('Yo)h( s.20).(8)(u.4^ sin 2irs + 16^ cos 41rs .184 CHAPTER VI.\ 3 C+ = sup 6 exp { A (.20 •exp { 2n cos 27rs . the nth Markovian environmental process {Jt} moves cyclically on {1. but thereby also slightly longer.18) Notes and references The material is from Asmussen & Rolski [44]. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. 1) for the environment). and let 8 = er' (Y0). such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. and in fact.0. Of course.19 } 0 <8<1 8 + cos 21rs Thus e.'Yo)e (6. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. exp 2^ cos 21rs . 0 <'p(8)(u ) . n}.L sin 27rs + 1 I cos 47rs .66.1 sin 2irs + 16_ cos 47rs . Thus C_ = 2 inf ex cos 2irs . Thus.g. for A = 1 (where 3 e0.cos 27rs . C+ = 1. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].181 s(u) < 1. with s the initial season. 1).\ = 0 . we have the following result: Theorem 6 .013. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.66. .
but others are also possible. it is desirable to have formulas permitting freely to translate from one setting into the other.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. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i.jEE of the risk process. Let 0j. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). one simple choice is Oni = 0( i . (6. z/'i (u. Bi.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). To this end. Thus. DUAL QUEUEING MODELS 185 within one unit of time on the average . M(n) = Supt>o Stn). . 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. since the settings are equivalent from a mathematical point of view. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. 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.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. Notes and references See Rolski [306]. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model.7. We let {Stn)} (6.1 ((i 1)/n) ) and Bni = B .
ii (u) = it /3 P(W > u. (VT > u I JT = 2).1) over j. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. (7.. and for (7. JT = i) = 'P. .T(V > u I J* = i). T) = 7ri 1 P. Proposition 7. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. I* = i). Proof Consider stationary versions of {Jt}o<t<T. Jo = j. JT = Z). JT = j) = LjPj (VT > u.0i (u . let T . In particular. {Jt }o<t<T• Then we may assume that Jt = JTt.. • The queueing discipline is FIFO. J* = i).3) 7ri where (V.3). JT = i} coincide. J* = i) = P. JJ = i). Proposition 7. JT = i) = P(V > u.2) and use that limF (VT > u.n(VT > u. J*) is the steadystate limit of (Vt. Then Pi(T(u) < T.2) Oi(u) = 1. I* )3i P(V > u.. The first conclusion of that result then states that the events {T(u) < T. J* = i) for all j. JT = j} and {VT > u.P(V > u.1 Assume V0 = 0. . (7. 2 .3. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.4) where 0* = >jEE 7rj/3j. Jt ).oo in u (7.186 CHAPTER VI. (7. JT = j) = 7rjPj(VT > u. Jo = i.1) 7ri In particular. For (7. and (7. The actual waiting time process 1W1.1) follows.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u.=1 .1). (7. Now let In denote the environment when customer n arrives and I* the steadystate limit.2). just sum (7. Taking probabilities and using the stationarity yields 7riPi(T(u) < T.
Proposition 7. P(W >u. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. and one has PI'>(rr(u) < T) = P('_T)(VT > u). we have 1: I(W. on average 0*T customers arrive in [0.l. on average /32TP(V > u. and (7. I* = i.4).6) (7.5) follows from (7.7) (7. a paper relying heavily on classical complex plane methods. a general formalism allowing this type of conclusion is 'conditional PASTA'. With {Vt} denoting the workload process of the periodic queue. and of these.g. see Regterschot & van Doorn [123].4) can be found in Regterschot & de Smit [301]. (7. P(.1 is from Asmussen [16]. N * oo.T)(T(u) <T) = P(8)(VT > u). the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. n=1 N However.7.I *=i).. Lemoine [242].3). that /3(t). B(t) have been periodically extended to negative t). see in particular Harrison & Lemoine [186].4) and (7. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. P(1')(r(u) < oo) = P(')(00) > u).o.8) For treatments of periodic M/G/1 queue. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. I*) with the timeaverage . [243]. J* = i) see W > u. and Rolski [306]. . with (7. The relation (7. if T is large. Taking the ratio yields (7. >u.I.=i) a4. T]. and further references (to which we add Prabhu & Zhu [296]) can be found there. In the setting of the periodic model of Section 6.3) improving somewhat upon (2. A more probabilistic treatment was given by Asmussen [17].7) of that paper.
This page is intentionally left blank .
1) (other terms are accumulated claims or total claims). and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).At + p(R8) ds. Thus in between jumps. and that the claim sizes U1.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 .i.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. i&(u. Zt As earlier. {Rt} moves according to the differential equation R = p(R). and the evolution of the reserve may be described by the equation Rt = u .2) tk(u.6. the aggregate claims in [0. U2. Thus. are i. with common distribution B and independent of {Nt}.. finite horizon.d.. . z/i(u) = F IinffRt< 0IRo=u 1 (1. 189 . However . resp . t] are Nt At = Ui (1.T) = F(T(u) < T). the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.
If Ro = v < u. Example 1.e. say at interest rate b. say e.190 CHAPTER VII. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. dividends are paid out at rate pi .1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.Vi(v) u > e(1 . P(r) _ p + e(r . oo) is given by i (u + p/S).2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. Proposition 1. we can put Rt = Rt + p/S. there is positive probability. Example 1. but when the reserve comes above v.i(u) = 1 for all u. pi > p2 and p(r) = One reason could be competition. That is.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. where one would try to attract new customers as soon as the business has become reasonably safe. Now return to the general model. but assume now that the company borrows the deficit in the bank when the reserve goes negative. it seems reasonable to assume monotonicity (p(r) is u .4 Either i. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. Assume 0(u) < 1 for some u. and the probability of absolute ruin with initial reserve u E [p/S. or o(u) < 1 for all u.2. However. i. rather than when the reserve itself becomes negative.'(u)) > 0 so that V'(v) < 1. In this situation. Thus at deficit x > 0 (meaning Rt = x). when x > p/S. Proof Obviously '(u) < ilb(v) when u > v. that {Rt} will reach level u before the first claim arrives. 1 .p2. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 .p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. No tractable necessary and sufficient condition is known in complete generality of the model. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. Hence in terms of survival probabilities. we get p(r) = p + er.
3µB for all sufficiently large r. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. V = p(V)). one can couple the risk process and the storage process on [0.5 (a) If p(r) < /. 296297): Theorem 1.2) we have t Vt = At .2 once more.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo.1. then l/i(u) < 1 for all u.uo) and. INTRODUCTION 191 decreasing in Example 1. hence Rt < uo also for a whole sequence of is converging to oo. Hence ik(u) < 1 for all u by Proposition III. . which was proved in 11. then ?(u) = 1 for all u. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. { Vt} remains at 0 until the next arrival).b(u.1 and increasing in Example 1. However. {Vt} decreases at rate p(v) when Vt = v (i.6) .2) for r sufficiently large so that p(oo) = limr.1. if and only if V)(u) < 1 for all u. (1. This is basically covered by the following result (but note that the case p(r) . (1.1. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. appealing to Proposition 111. Then if u > no.2(d).o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.6 For any T < oo. In particular. Then 0(u) = P(V > u).4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. Starting from Ro = uo.uo) < 1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.e. let uo be chosen such that p(r) < p = /3µB for r > uo.T) = P(VT > u). that u zPp(u .f p(Vs) ds. say V. Let Op(u) refer to the compound Poisson model with the same 0.2(d)). Proof This follows by a simple comparison with the compound Poisson model. B and (constant) premium rate p. and hence by a geometric trials argument. In case (a). cf.3. and P(Rt + oo) > 0. That is. In case (b)..5) and the process {Vt} has a proper limit in distribution . (1. obviously infu<uo z/'(u) > 0. [APQ] pp. instead of (1.4. We next recall the following results. Theorem 1. we have z/i(u) <p(u . Proposition I1I.1. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. In between jumps.+ p(r) exists.
this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. Oeax f x e'Yg (y) dy } = p) eaxa(x) . oo).8 Assume that B is exponential with rate b. the flow of mass from [0. t + dt] if and only if Vt E [x. It is intuitively obvious and not too hard to prove that G is a mixture of two components. Jo AX) (1. (1. x] to (x. An attempt of an upcrossing occurs as result of an arrival.y)g(y) dy. and the other being given by a density g(x) on (0. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.Sx} dx. the l. say.s. Corollary 1.s. of (1. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. Considering the cases y = 0 and 0 < y < x separately.9) Proof We may rewrite (1. of (1.8) as the rate of upcrossings. 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. one having an atom at 0 of size 'yo. Then the ruin probability is tp (u) = f' g(y)dy.6w(x) . for the storage process {Vt}. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. B(x) = e.y. and is succesful if the jump size is larger than x .6x and that w(x) < oo for all x > 0.192 CHAPTER VII. In view of the path structure of {V t }. we thus need to look more into the stationary distribution G.Qw(x) .7) Proposition 1.7 p(x)g(x) = tofB (x) + a f (x .8) is the rate of downcrossings (the event of an arrival in [t. oo) must be the same as the flow the other way.8) Proof In stationarity. u Define ^x 1 w(x) Jo p(t) dt.h.8) as g(x) = p 1 {yo13e_6x +.h.6 applicable. Now obviously. yo ^ 1 + oo Q exp {. It follows in particular that 0(u) = fg(Y)dy. say. say when {Vt} is in state y. where g(x) = p( ^ exp {. Note that it may happen that w (x) = oo for all x > 0. x + p(x)dt]). (1.Sx}. we arrive at the desired interpretation of the r.
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
b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.V.a) . it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 . . 13/E). The process {St} corresponds to {Wt} so that c(a) or2a2/2 .pa.2) follows by elementary algebra. and the c. From ic(a) = .V < x)]0 + f P(V > p/E ) + eby fv (p/E . RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .2) follows by elementary algebra.2y +µ ) dy .3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. /^ u Example 2 .3. As an example. where V is Gamma(b.200 CHAPTER VII.x) dx e.13 /E) r (.5 The analysis leading to Theorem 2./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . then {Rt} is the diffusion with drift function p+Ex and constant variance a2.f. RESERVEDEPENDENT PREMIUMS u from which (2.3a/ (5 .01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.g. Proof 2 We use Theorem 2. of Z is IogH[a] = f ytc(y)dy = e fa (0.pa.e.3/E) By the memoryless property of the exponential distribution. assume that {Wt} is Brownian motion with drift µ and variance v2. i. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. r (j3/E) In particular. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.
A r. [129] and Harrison [185]. or to nonlinear premium rules p(•).. Corollary 2.v.3) was derived by Emanuel et at.3. [283].e. The formula (2. that the analysis does not seem to carry over to general phasetype distributions. Paulsen [281].4 is classical. [129]. Emanuel et at. 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).Y*p* W*(u) < ery*u = 0. of the form Ei° p"X" with the X„ i. Q2/2E). see e.d.g. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . [357].p*. Gerber [157] p. Delbaen & Haezendonck [104]. Paulsen & Gjessing [286] and Sundt & Teugels [356]. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. Some of these references also go into a stochastic interest rate. 3 The local adjustment coefficient.i. [282]. se e. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). not even Erlang(3) or H3.1) . it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2.8. write y* for the solution of the Lundberg equation f3(B[ry *] . Z is normal (p/E. and since RT = 0 by the continuity of Brownian motion. however. Goldie & Griibel [167]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1.3 is from Harrison [185]. It must be noted. it is also used as basis for a diffusion approximation by these authors. Further studies of the model with interest can be found in Boogaert & Crijns [71]. write Vi* (u) for the ruin probability etc.2 is a special case of a perpetuity. and recall Lundberg 's inequality . for a martingale proof.g.. Gerber [155]. as in the proof of Proposition 2.
1 ).*(u) .1) . x * oo.log '(u)/u < ry*(1 . The steepness assumption and p(x) + oo ensure 'y(x) * So. i. oo for all E > 0. choose c(.C*ef*".e. (3. Then lim sup u>oo u and e E''p(r) + 0.1. i.ap(x). RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.. B(x) > C(2)e(ao+f)x for all x. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.4) we assume existence of y(x) for all x. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. a) = f3(B[a] . log ?i(u) < < 00 JO . Proof of Theorem 3. Let y* < So. the function y(x) of the reserve x obtained by for a fixed x to define y(x) as the adjustment coefficient of the classical risk model with p* = p(x). then log u (u) In the proof as well as in the remaining part of the section . (x.i)eex. and that p(x) * oo.1. 1) and for a given E > 0.e. (3.w (u) J dt > c(3)eeu v 1 p(u+ t) .2) such that p(x) < c(. and (for simplicity) that inf p(x) > (3µs . If 60 s f 6o. choose uo such that p( x) > p* when x > u0E.1 Assume that for some 0 < 5o < oo. as will hold under the steepness assumption of Theorem 3.E). we will use the local adjustment coefficient 'y(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. 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) . 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 .>o 7(x) > 0. it holds that f3[s] T oo. For the last asssertion .3) When trying to extend these results to the model of this chapter where p(x) depends on x. a first step is the following: Theorem 3 .'y ( x)) = 0 where r. When u > uo.5) which implies inf. as solution of the equation n(x. e(1o+e)2 (x ) u > 00. let p* be a in (3. x>0 (3. c(..202 CHAPTER VII.
13 is a technical condition on the claim size distribution B. Then limelog l/ie (u) = I(u).1 only presents a first step. noting that in many cases the constant C is close to 1.3.0 are the same. The form of the result is superficially similar to the CramerLundberg approximation.e.3 Assume that either (a) p(r) is a non decreasing function of r. u Obviously. {Rte)} defined as in (1.ea°/(ecf1)).3 to be reasonably precise and use e` (u) as approximation to 0 (u). and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. UJU > x cannot have a much heavier tail than the claim U itself. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .6) The second main result to be derived states that the bound in Theorem 3.13 below holds. the result is not very informative if bo = oo. (3. For e > 0. which essentially says that an overshoot r. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. Theorem 3 . one can then assume that e = 1 is small enough for Theorem 3.2 is also an approximation under appropriate conditions. 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. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. The rest of this section deals with tail estimates involving the local adjustment coefficient. 3) = (1 .' (u) < eI("). 2.. If p(x) = pis constant .2). or (b) Condition 3. Then . I. . then Rte) = CRtie for all t so that V). (u) = O(u/e). (3.(u) > so. Bucklew [81]).4)eE" Given such an arrival.7) CIO Remarks: 1. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Theorem 3.v. ruin will occur if the claim is at least u + v. the asymptotics u * oo and c . THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. by cU2. However. and in particular. 3.g. Condition 3.
exp {/33w(u) .4 Consider again the exponential case B(x) = eax as in Corollary 1.6).(x) dx. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.2.bx} dx fo 00 1 + [exp {/(3w(x) .bx} dx 1+0. say. the logaritmic form of (3.8.3. RESERVEDEPENDENT PREMIUMS 4. 5. 3a Examples Before giving the proofs of Theorems 3. rather than eI(u)). However. we show how to rewrite the explicit solution for ti(u) in Corollary 1. we consider some simple examples.3. First.204 CHAPTER VII.bu}. we get = 1+ J" AX) exp {(3w (x) . 3.(3/p(x).bx}]o + b /' oo exp low (x) .bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . As typical in large deviations theory. it is formally needed only for Theorem 3.bx} dx . Then y(x) = b . J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.1 + b f e.8 in terms of I(u) when the claims are exponential: Example 3 .bx} dx = 1 + J0 dodx(x) exp {. u .(iw(x) .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).
70 > 0 such that 7(x) < 7o for y < yo. in the definition of AE converges to 0. Similarly. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0..ev 0 O /E) J0 70 70 Yo This implies lim inf A. (3.9 ) 11000 eI(v)dy f000 e.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.e. IE(u) = I(u)/e. note first that the appropriate slow Markov walk assumption amounts to u.3. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion.3 in the particularly simple case of diffusions: Example 3. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy .BE. and (3.2. and (3. For Theorem 3. u .2. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e..0. oo) with drift µ(x) and variance a2 (x) > 0 at x.10) where AE = e log 000 e.fo 7(x)dx/Edy f . (X) = µ(x).3. 0. It is well known that (see Theorem XI.fory(x+u)dxdy ( 3. Be = e log U000 e. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3.7) follows. (3. ry(x /b I u o e f0 °° e  e.I ( u) fool.f y(x)dxd y If 7(x) is increasing . > lime log e = 0 and AE * 0.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .2(X) = ev2(x) so that 7e(x) = 7(x)/e.5 Assume that {Rt} is a diffusion on [0. 3. (u) = I(u) + AE .1. In particular.10 or Karlin & Taylor [222] pp.5) is infx>o 7(x) > 0 which implies that f °O . BE * 0. Choosing yo.1/8 .fa 7(x+u)dx/Edy o The analogue of (3.I ( v )dy fo +u) dxdy . we get r 00 e.9) yields e log .
1 3.5) and 7* = 5 . I(u ) = G1(u) + . the slow Markov walk assumption means 5E = b/c.206 CHAPTER VII. lim sup Af < lim sup c log(1 . 7(x) is typically not explicit. As in Example 3./3 1 AX dx. > . Nevertheless .. Thus 7e(x) _7(x)/e and (3. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. I. Further. G.6) exactly as in Example 3.+1 (u) = o( 1). we have 5 > 7o and get lim inf AE > lime log e .5. however .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. the results are suggestive in their form and much more explicit than anything else in the literature..5 for risk processes with exponential claims is as follows: Example 3 . RESERVEDEPENDENT PREMIUMS The analogue of Example 3.(u) oo. G. + Gq(u) + o(G9(u))• Gi (u) It should be noted .6/p* so that u 1 I (U) = bu .7o C 15 I I.. . .e. 0 Now (3. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. Ignoring 1/5 in the formula there. that the interchange of the slow Markov walk oo is not justified and in fact.Q/p*.5. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. . .6 Assume that B is exponential with rate S.5. this leads to (3.0. .0) = 0. _ . so our approach is to determine standard functions Gl (u). Then the solution of the Lundberg equation is y* = b .0/e. Of course. G.4..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. E+o e*O By (3. 0.
ry*°p*.3. I(u) Pt. x T 1. . (3.1) leads to (S7T N Ocp a. more generally.C2p* C2 = (3clr( a))11'. 1. .1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 .cs(1 .g. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . and hence (3.11) with a > 0.c2 Su a dx ) Su a<1 Su .Y . 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.8 Assume next that B has bounded support..3cse7*I7(77) . if the phase generator is irreducible ( Proposition VIII. u Example 3 . B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . 2. It follows from (3. y = 2 if B is uniform on (0.x)n1.1) leads to . For example. say 1 is the upper limit and B(x) .4) or gamma distributions. This covers mixtures or convolutions of exponentials or. Here B[s] is defined for all s and B[s] .1/a).s)C' f "o o as s T S. ry* loge*+ g7loglogp*.8). Hence (3.12) with y > 1.7 Assume that B(x) . c4 = c2b 1/'/(1 . fu I(u) Su . 77 = 1 if B is degenerate at 1. in the phasetype case . the typical case is a = 1 which holds . u(logu + r7loglogu).y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc. More precisely.11) that b[s] * co as s f S and hence 7* T S as p* + oo.clxale5x 207 (3.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. phasetype distributions (Example 1. e.1/k).:.
of the increment in a small time interval [0.15) Proposition 3. this is only possible if 7o(v) 2 7o(u)• . 3b Proof of Theorem 3.css 2%rc7eC782/2.12).c8 log . (3.f. one could also have considered the increment ru (T1) .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. RESERVEDEPENDENT PREMIUMS Example 3 . (3.3 (B[s] .r„(Ti).9 As a case intermediate between (3. 1 0 3e.f.B[7o (u)] . x f oo .10 Assume that p(x) is a nondecreasing function of x. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .•. ec78)2/2c7 dx C7 . The assumption implies that ru(t) .208 CHAPTER VII.g.(t))dt.ru(TI)) ..u . assume that B(x) CO x2/2c7. 1 = E. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. h 10.13) We get b[s] .4). Hence for u<V.11) and (3.2 We first remark that the definition (3.Ote7o( u)(u.u is a nondecreasing function of u. . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. I (u) c8u log u 0 where c8 = 2/c7.r^. h]. (3.log p*. By convexity of the m .sp(u). of U1 + v .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .e7o ( u)(ul+u r. g. 7 * .14) for the m .1) .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).(T1)) > Ee7o(u)(ul+vr»(Ti)).4) is the formula h logEues ( Rhu) .
16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).u > tp(u).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.7o (u)p(u)• Since (3. Assume (3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.1) .11 Assume that p(x) is a nondecreasing function of x. Also. (3.2 in terms of 7o. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. THE LOCAL ADJUSTMENT COEFFICIENT For (b). note that the assumption implies that ru(t) . Hence 1 = EeYo(u)(U1+uru(T1)) < E.Fu(u ) + J ^(n)(u .17) shown for n and let Fu(x) = P(U1 + u .10(b): Theorem 3. u We prove Theorem 3. fa 7o(y)dy < u7o(u) < xyo (u) for x > u.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .u[70(u)] fo eyo(x)dx . Then (u) < efo Yo(x)dx. it is easily seen that fu x7o(y)dy < xyo (u). We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.4) considered as function of 7 is convex and 0 for y = 0.ru(T1) < x).17) from which the theorem follows by letting n + oo. Separating after whether ruin occurs at the first claim or not. the case of 7 then follows immediately by Proposition 3. this is only possible if yo(u) > 7(u).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .3. Hence „/.(n+l) (u) 1 . we obtain „I.
10(b ) that the bound provided by Theorem 3..: y(u).12 lim sup4^o f log O.. yo(u) appears more difficult to evaluate than y(u).E (u/ n) Y'E (un .E (u) denote the ruin probability for the classical model with 0 replaced by .2.n inf n uk1. in accordance with the notation i/iE (u).2).11 is sharper than the one given by Theorem 3. The probability of this is at least n n.3).n <Z auk}l. the probability that ruin occurs in the CramerLundberg model with p* = pn.n) > k =1 II v ^k n..n AX).3/e and U.x/n. 0.n (starting from u/n) without that 2u/n is upcrossed before ruin.3 is required.nbe C*. by €U=. P k.210 CHAPTER VII.3).n = sup p(x).11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. For these reasons.2. Also.E (u/n) Now as e .n. Lemma 3.n = ku. 0 It follows from Proposition 3.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .E ( u/n) ^•e. for either of Theorems 3. 0. the value of {R(E)} at the time of downcrossing is < unl. let Op*.E (u/n).n u k}1. given downcrossing occurs.I. ryk. Y*u /E..15). define uk.. However. and. in . and here it is easily seen that yo(u) . {RtE)} (starting from u = un. y* evaluated for p* = Pk. C*e where the first equality follows by an easy scaling argument and the approximation by (3. 3.10(a) for some of the inequalities. we used also Proposition 3. (3.. Proof For ruin to occur. pk n = uk_l. resp. (u).e (u) = v'. W O .n so that n.n. op*. (u) < I(u). (un2. we have chosen to work with y(u) as the fundamental local adjustment coefficient. 3c Proof of Theorem 3. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.n.n) must first downcross unl. To this end. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.n. Further. Let Ck.n) pn niE (u /n) n n_1 n.
13 There exists a r. B(x) (3..a( u)z. Indeed .! (u/n) n n m 7k.n + 0(1). since ry' is an increasing function of p'. also ryk.E (u/n) OP +^p•. 3 now follows easily in case (a). v. THE LOCAL ADJUSTMENT COEFFICIENT particular. .n.F (2u/n).3. In case (b).E (u/n) Op•. (3. It follows that n log V'C (u) k =1 log Ypk.log Ck. for all x . so that Theorem 3.n <X<Uk.7k. *p•. k=1 k=1 n u _ nE7 k.n . we need the following condition: Condition 3. uk_1.nk=1 limsupelogv)..i.12 completes the proof.. 0 with n and u fixed.n = sup ?'(x). ne7k.nu /fn( 1 Ck  e. 11 Theorem 3.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u). (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof..nu/en(1 + where o(1) refers to the limit e . /' (u/n) 'T nk.e.n cE (2u/n) Ck ne7k.19) .E (urn) < \ *I. 211 Clearly.nu /En) o(1)). 40 Combining with the upper bound of Lemma 3. 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. in obvious notation one has tC (x) = y(x)/e. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).18) (ii) the family of claim overshoot distributions is stochastically dominated by V. i.
(u/n .eV) • P (T(E) (u.. .EV) = El + E2. For E2.QEU 1 .n V. (R.n < ery1.of:>2 in n(x). (u/n . P (T(E) (u.212 CHAPTER VII. u/n) < oo] .nu /EnE [e71.EV) = EiI 1 .^'' = E [ . v ) = inf { t > 0 : R(c ) < v R) = u } .E (2u/n .E (u/n .R<) (u v). The probability of ruin in between two downcrossings is bounded by Epp ..18) for the last equality). Then the standard Lundberg inequality yields El < E?. u/n) < oo) EV)..1 n. infx>2u /n P(x) . RESERVEDEPENDENT PREMIUMS To complete the proof.nu/En0(1) .EV) = e.( . T() (u. V < u/En] + P(V > u/En) (u/En . u/n)) .E(E) (u. u/n)) I T(E) (u. Then Y'E (u) ^(E) (u. v ) = v . where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. let v < u and define T(E) (u.v.. u/n) < oo] E [OE (u/n . T(E) (u. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.2y 1 ' . u /n) < oo] l = = < E [OE (u/n . N with EN < 1 = infx>2u/nA(x) = 0(1). ) (u u /n)) .5) and the standard formula for b(0). u/n) < oo) . Write EO. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).^(E) (u.V) = e71 nu/Eno(l) (using (3. Ei + E2 < e71. (3.E (0) cf.
21) (the initial condition is r(0) = u in both cases).3. s).u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.r. . Similarly. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.7(x)) (3. 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. it might be possible to show that the limits e . l o JJJ o . the approximation (3. [89].3EU) (3.) = exp . where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . s) as in (3. 0 and b T 00 are interchangeable in the setting of [89].g. T) is maximized over T by taking T as the time for (3. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . u Notes and references With the exception of Theorem 3. 0 ) (= p(x) . the results are from Asmussen & Nielsen [39]. Bucklew [81]).20) (with ic(x. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Comparing these references with the present work shows that in the slow Markov walk setup.21) to pass from u to 0.)Ui } .7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.T) = P „(info<t <T Rt < 0) via related large deviations techniques.1.=1 J An approximation similar to (3.)ds + Y(R2. Typically.J r(Rs)p(R.7) then comes out (at least heuristically) by analytical manipulations with the action integral.4) and the prime meaning differentiation w.t.J y(Rs)dR. the risk process itself is close to the solution of the differential equation r(x) _ r (x. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. Djehiche [122] gives an approximation for tp(u.13. one can in fact arrive at the optimal path by showing that the approximation for 0(u.
see XI.3. We should like.214 CHAPTER VII. however. to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding .g. e. . For different types of applications of large deviations to ruin probabilities .. the exponential distribution ).
Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability.1) is 'Here as usual . if a problem can be solved explicitly when the relevant distributions are exponentials. Note that since (1. refers to the case Jo = i. A distribution B on (0. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. if v = (vi)iEE is a probability distribution. More precisely. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . on Eo = E U {A} where A is some extra state which is absorbing. P. 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. Typically. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. We often write p for the number of elements of E. that is. and not in other cases. 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.
j. T) (or sometimes just (a. (1. Here are some important special cases: Example 1 . B(t) = Fa(^ < t ). i. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. 0 2this means that tii < 0.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. tij > 0 for i 54 j and EjEE tij < 0 .1 The phase diagram of a phasetype distribution with 3 phases.e. The initial vector a is written as a row vector. T is a subintensity matrix2. k}.1 Suppose that p = 1 and write . We now say that B is of phasetype with representation (E. In particular.0 = t11. C is the lifetime sup It > 0 : Jt E E} of {Jt}. Equivalently. and we have t = Te. t1 = /3. 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.2) The interpretation of the column vector t is as the exit rate vector. that is. Then a = a1 = 1. E = {i.216 CHAPTER VIII. an exponential distribution with rate parameter . A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. and the phasetype distribution is the lifetime of a particle with constant failure rate /3.3.e. a. i.
a = (1 0 0 .. p}. . the EP distribution may be represented by the phase diagram (p = 3) Figure 1.x i=1 Thus E _ Si 0 T 0 S2 0 0 . 0 •..1. 6.... 0 0 0 0 S 6 . PHASETYPE DISTRIBUTIONS 217 Example 1. ..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.... 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.. so that the density is P E ai6ie6. .. . 0 S 6 Example 1.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51..2 corresponding to E = {1. 00)) S s o . . 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p... . 0 SP 0 and the phase diagram is (p = 2) .
T). (c) the m.d. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} .1)"n! aT"e.f is B (x) = 1 . p:. i. see A. the Erlang distribution is a special case of a Coxian distribution.f. the backwards equation for {Jt} (e. Theorem 1 . B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . . dp.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. [APQ ] p. The basic analytical properties of phase type distributions are given by the following result .218 CHAPTER VIII.g. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.g. ds^ = ds' = ttlaj + tikpkj. 36) yields s d. Then for i . MATRIXANALYTIC METHODS Figure 1. a. j E E.3 . 5 Let B be phasetype with representation (E.4 For example. (b) the density is b(x ) = B'(x) = aeTxt.aeTxe. Then: (a) the c.e.3 0 Example 1 .t2 yt bP.1 tP1 1 Figure 1. the restriction of P8 to E.
g.6) . and since obviously P° = I.1. and since b[s] = ah. .tii and have an additional time to absorption either go to state j which has m . Alternatively.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .. of the initial sojourn in state i.5) Indeed .5) as hi(tii + s) = ti  t ij hj. After that.p.s j# tii i (1. h = (T + sI)1t. i.e.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). this means in vector notation that (T + sI)h = t.g. = aPxe.tii tii .f. Part (d) follows by differentiating the m. the solution is P8 = eT8. hj . 1. Since 1 .g. j#i jEE tijhj + his = ti.f. For (c). d8 P8 = TP8. we i w.f.n lt .p. B(n)[0] = _ Alternatively. Rewriting ( 1. we arrive once more at the stated expression for B[s]. Then h tit ti + ti3 h j ..T) 't = (. or w.f. (Jx E E) = this proves (a). tij / . PHASETYPE DISTRIBUTIONS 219 That is.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . d" dsn a (.s I . (1) n+1n!aT .1 ) n +l n ! a (s I + T ) .5) ki = 1 + tii L jj:Ai tii (1. the rule (A. ti/ .g. i. for n = 1 we may put ki = Ei( and get as in (1. and (b) then follows from 1: aipF.s) is the m .tii we go to A.n1t = (1)nn!aTn1Te (1)nn! aTne. define hi = Eie8S. in which case the time to absorption is 0 with m .T) 1t.B(x) = 1'a (( > x) = P.
making the problem trivial. another the case p = 2 where explicit diagonalization formulas are always available.s. Example A3.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.220 CHAPTER VIII.h. T= 2 111 so that 2 2 Then (cf. One obvious instance is the hyperexponential distribution. Consider for example 3 9 a= (2 2).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. 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. 0 Example 1. are idempotent. 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 . there are some examples where it is appealing to write T on diagonal form. see the Appendix.
T) is then defined to be oo on a set of probability 1. and in fact one also most often there allows a to have a component ao at A. hail = E=EE a.T). i.e 11BIJ = 1laDD < 1. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.7) Proof According to (A.T) with weight hall and an atom at zero with weight 1 . B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).f. then the matrix m.11aDD. a random variable U having a defective phasetype distribution with representation (a. 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. 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). where the initial vector a is substochastic. This is the traditional choice in the literature. There are two ways to interpret this: • The phasetype distribution B is defective. < 1.29) and Proposition A4.1.7 If B is phasetype with representation (v.4b for definitions and basic rules): Proposition 1. (1.g. • The phasetype distribution B is zeromodified. or one just lets U be undefined on this additional set. .4. i.hall. 0 Sometimes it is relevant also to consider phasetype distributions.e a mixture of a phasetype distribution with representation (a/llall.
In Proposition A5. In older literature. All material of the present section is standard. cf.F. but the relevant T is not irreducible. the result follows (with C = (ah)(ve)). Other expositions of the basic theory of phasetype distributions can be found in [APQ]. Neuts. but in many practical cases. the conditions of Proposition 1. Schmidt & Teugels [307] and Wolff [384]. Using B(x) = aeTxe . let . h can be chosen with strictly positive component. i is real and positive. not only in the tail but in the whole distribution. we give a criterion for asymptotical exponentiality of a phasetype distribution B. T). 1. Example A5. 77 > 0 and k = 0. The Erlang distribution gives an example where k > 0 (in fact.8 Let B be phasetype with representation (a.4c). but todays interest in the topic was largely initiated by M. one has k = 0. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. Rolski. x * oo.8).f. (1. the text is essentially identical to Section 2 of Asmussen [26]. Notes and references The idea behind using phasetype distributions goes back to Erlang. 0 Of course. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. see his book [269] (a historical important intermediate step is Jensen [214]). v.1 of the Appendix.q be the eigenvalue of largest real part of T. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . the Erlang case). cf. 2.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. assume that T is irreducible . distributions with a rational m. B(x) . No satisfying . . O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m.g.8) Proof By PerronFrobenius theory (A..f. let v. See in particular the notes to Section 6. here k = p1). and we have eTx . Here is a sufficient condition: Proposition 1.hve7x.g. 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.Ce7'. Lipsky [247]..222 CHAPTER VIII. Schmidli. Then the tail B(x) is asymptotically exponential. where C. (or Laplace transform) are often used where one would now work instead with phasetype distributions.
.. n=O We may think of the U.r.. as the lifetimes of items (say electrical bulbs) which are replaced upon failure.1.i. however. (2.t.. For this reason.. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. . A related important unsolved problem deals with minimal representations: given a phasetype distribution . and U(A) is then the expected number of replacements (renewals) in A. 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. .+UnEA).g...f.1 of the Appendix.. the renewals form a Poisson process and we have u(x) = 0. is Markov and has two types of jumps .. Then the renewal density exists and is given by u(x) = ae(T+ta)xt.d. be i. + U0 is 0 . but nevertheless. Lebesgue measure. 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 +.. If B is exponential with rate 0..1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together..: U1 + .1 Consider a renewal process with interarrivals which are phasetype with representation (cr. but is in part repeated below. we refer to U as the renewal measure. oo) w. Let U1. JtJt1) Then { 0<t<U1 .T)... or the density is available ) is. known. with common distribution B and define4 U(A) = E# {n = 0. U1<t < U1+U2. 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} 00 = EEI(U1 +. Jt={Jt?ul}. U2.2. if U is absolutely continuous on (0. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. we denote the density by u(x) and refer to u as the renewal density.
which is ti in state i. see Fig. Proof Just note that { it } is a governing phase process for the lifetime. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. that is.T) where vt = ae (T+ta)t . Hence the intensity matrix is T + ta. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter .T + ta). B is defective . and the distribution of Jx is ae ( T+t«)x. which is phase type with representation (v. + Uit_1 where s. The renewal density at x is now just the rate of jumps of the second type. the lifetime of the renewal process.e.224 CHAPTER VIII.T). fi(t) U2 U1 . the density is veTxt = B(x)/µB. This is defined as U1 + . T). However. i. this is welldefined..IIBII which is > 0 in the defective case.. . is the first k with Uk = 00. since Uk = oo with probability 1 . and hence ( 2.T) where v = aT1 /µB. Hence ( 2.3 Consider a renewal process with interarrivals which are phasetype with representation (a. Equivalently. Corollary 2.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a.e. and let µB = aTle be the mean of B. 2. . (b) £(t) has a limiting distribution as t * oo. as the time of the last renewal.1. IIafl < 1. u The argument goes through without change if the renewal process is terminating.1) follows by the law of total probability.1 Corollary 2. u Returning to nonterminating renewal processes . i.U1 U3 U2 U3 U4 Figure 2. and the jumps of the first type are governed by T. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1) remains valid for that case. 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.
The time of the next renewal after t is the time of the next jump of the second type. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. the unique positive solution of ve = 1. (ii) First check the asserted identity for the density: since T. T1 and eTx commute.2.2) v(T + ta) = 0.4 Consider a nonterminating renewal process with two phases. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. cf.e. The renewal density is then aeQtt = (al a2) ( 7i 7"2. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. i. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.1.) ( t2 ) .T) where vt is the distribution of it which is obviously given by the expression in (a).e. hence e(t) is phasetype with representation (vt. According to Example A3. we first compute the stationary distribution of Q.q2. u Example 2 .6. 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). (2. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. Al.
MATRIXANALYTIC METHODS e. Hence 7r = (1/2 1/2). Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .a27r1) (t1 .6 Let B be hyperexponential.4 yields the renewal density as u(t) = 5152 e. A = 25. The present treatment is somewhat more probabilistic. . Then Q= 0 55 )+(1o)=( j ad ). and Example 2.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . t1B 0 Example 2 .t2) 1 + eat (a17r2 .4 yields the renewal density as u(t) = 2 (1 .e2bt) 13 Example 2 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].t2) .5 Let B be Erlang(2).52a1. and Example 2.226 CHAPTER VIII.a27rl) (tl .tl) 7r2t2 + eat (a17r2 . )t (51 .(biaz + aza.
i.1 on the next page. T + to+). Here we have taken the terminating Markov process underlying B with two states. the Markov processes representing ladder steps can be pieced together to one {my}. which occurs at rate ti. 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. however. and rewriting in matrix form yields the phase generator of {my} as T + ta+. with 0 denoting the Poisson intensity. Next. Considering the first. r(u) the time of ruin with initial reserve u. T). . we shall. 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. (b) V.2. For (b). a+j.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. Corollary 3. and M is zeromodified phasetype with representation (a+. Then: (a) G+ is defective phasetype with representation (a+.p. cf. Corollary 2.3. The stars represent the ladder points ST+(k). T). Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. use the phasetype representation of Bo.3. B the claim size distribution.f3aT1. and if there is a subsequent ladder step starting in j whic occurs w. We asssume that B is phasetype with representation (a.) = F(ST(o) E •. Since the results is so basic.i. marked by thin and thick lines on the figure. T(0) < oo) the ladder height distribution and M = supt>o St.e. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). Then each claim (jump) corresponds to one (finite) sample path of the Markov process. we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0. 3. {St} the claim surplus process. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_.1 Assume that the claim size distribution B is phasetype with representation (a. T) where a+ is given by a+ = . represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. add a more selfcontained explanation of why of the phasetype structure is preserved. Thus the total rate is tip + tia+. G+(. Within ladder steps. The essence is contained in Fig.
.228 CHAPTER VIII..1 .t t d kkt S.1 This derivation is a complete proof except for the identification of a+ with . see Corollary 2.QaT1. 0 Example 3.2 Assume that . 3e3x + . 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 .Q = 3 and b(x) = .. This is in fact a simple consequence of the form of the excess distribution B0.3. T = (3 . MATRIXANALYTIC METHODS t . Figure 3.M {mx} ST+(2)  S .
T). we encounter similar expressions for the ruin probabilities in the renewal.2 are taken from Gerber [157]. see Stanford & Stroinski [351] .1): Proposition 4.6).4. For further more or less explicit computations of ruin probabilities. and the argument for the renewal case starts in just the same way (cf.j).1 In the zerodelayed case. We shall derive phasetype representations of the ruin probabilities V) (u). this was obtained in Section 3. T). For the compound Poisson model. The parameters of Example 3. (a) G+ is of phasetype with representation (a+. 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 . The result carries over to B being matrixexponential. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).1 can be found in Neuts [269] (in the setting of M/G/1 queues. That is. his derivation of +'(u) is different. the duality result given in Corollary 11. where a+ is the (defective) .^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.6. see Shin [340]. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). if we define {mz} just as for the Poisson case (cf. 3. T) for some vector a+ = (a+.4. cf.1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+. THE RENEWAL MODEL This is the same matrix as is Example 1. Fig. 4 The renewal model We consider the renewal model in the notation of Chapter V. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. see Section 6. 3. with A denoting the interarrival distribution and B the service time distribution. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. For an attempt. In the next sections. the discussion around Fig.and Markovmodulated models. We assume p = PB/µA < 1 and that B is phasetype with representation (a.
230 distribution of mo. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. Since the conditional distribution of my given T1 = y is ae4y. which for numerical purposes can be solved by iteration. Also. the calculation of the first ladder height is simple in the stationary case: Proposition 4. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). obviously mo = m. 4. where B0 is the stationary excess life distribution corresponding to B. Then {m. But by Corollary 2. In fact. where a(8) = aT1/PA. (c) {mx } is a (terminating) Markov process on E.1) Proof We condition upon T1 = y and define {m. with intensity matrix Q given by Q = T + to+.*'} is Markov with the same transition intensities as {mx}. Proof Obviously. Hence by Theorem 11.*} from {St+y .3. G(') = pBo. 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.Sy} in the same way as {mx} is defined from {St}.T). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. the Palm distribution of the claim size is just B. CHAPTER VIII. but with initial distribution a rather than a+.3 a+ satisfies a+ = V(a+). where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). (4.5.1.T)• Proposition 4.2 The distribution G(s) of the first ladder height of the claim surplus process {Ste) } for the stationary case is phase type with representation (a(8). cf. Nevertheless. The key difference from the Poisson case is that it is more difficult to evaluate a+. Fig. We have now almost collected all pieces of the main result of this section: Theorem 4 .T).1). Then .6.
•. THE RENEWAL MODEL 231 . a+2) = ^p (a+l)) . .1).2. thus . I {mx} . The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. a+ can be computed by iteration of (4. (4.^(8)(u) = a ( 8)e(T+ta +) xe.1/pA. only with initial distribution a(*) for mo.2) where a+ satisfies (4. a+l ) = cp (a+°)) . (4. It remains to prove convergence of the iteration scheme (4.1(b).^(u) = a+e ( T+ta+)xe.4. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. . i y ^ T1= y `•r Figure 4. In particular ..0.1) and a(8) _ aT.. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . i.3).3) (defined on the domain of subprobability vectors .3) Proof The first expression in (4. and that this is given by Proposition 4.2 ) follows from Proposition 4.0) is an increasing function of /3. by a+ = lim a +n) where a+°) .M.. Hence ^p(. Furthermore .1 .1 by noting that the distribution of mo is a+. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4..e.
. n) &+n) T a+. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . 0 0 We next give an alternative algorithm. this implies that ahA[s] # 0. Fn ).1.4) whenever EeR(S)U < oo.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). Assume the assertion shown for n . Then each subexcursion of {St+Tl . . However. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.T1. the normalization is equivalent to F(s) = 1. a+ ) exists .2. 7+ ]. with B[s]. To this end. (4. 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). so to complete the proof it suffices to show that &+ < a+) for all n. we use an argument similar to the proof of Proposition VI. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1.T)'t • A[s] (4.ST. the corresponding right eigenvector may be taken as (sI . For n = 0.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). and let &+". limn4oo a ) < a+.4. Thus by (4.T)1t.5 Let s be some complex number with k(s) > 0.5) Since s $ sp(T).5) yields h = (sI . Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah.g. To prove the converse inequality. both quantities are just 0 . which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). Theorem 4.4). Obviously.) = P(mTl = i. and hence we may assume that h has been normalized such that ahA[s] = 1.P[s] = A[s]B[s]. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. s ¢ sp(T). Then F[s] = a(sI .232 CHAPTER VIII. F[s] being interpreted in the sense of the analytical continuation of the m. Similarly. Then (4. In that case. Proof Suppose first Qh = sh.T)It.} can contain at most n . let F be the distribution of U1 .f. (4. Thus .
. 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 ).T)It. and the topic is classic both in risk theory and queueing theory (recall that we can identify 0(u) with the tail P(W > u) of the GI/PH /1 waiting time W..6) i=1 i=1 Proof Appealing to Theorem 4..p1i . As in Corollary 4. hd. and the solution is . pdhd.T)lt = sh. Pd in the domain ER(s) > 0 . t(ry) > 0. . Given T has been computed.. yd satisfying R(ryi) > 0. we get at a(Q .T) = 1 ata+ = a+..lt we get Qh = (T + to+)h = T(sI . . 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. hd. and hence by the WienerHopf factorization identity (A.9) we have G+[s] = 1 which according to Theorem 1. Let d denote the number of phases. in turn.6. .. This gives d roots 'y. Q has diagonal form d d Q = dpivi®hi = dpihivi... Further.6 Suppose u < 0..1 has the d distinct eigenvalues . 0). Then G+ is phase.' that the equation F(s) = 1 has d distinct roots p1. D that with columns p1 hl. . This immediately implies that Q has the form CD1 and the last assertion on the diagonal form ..5(c) means that a+(sI T)1t = 1. (4. the matrix Q in Theorem 2. W v M(d) in the notation of Chapter V).. THE RENEWAL MODEL 233 Suppose next F(s) = 1.. Corollary 4. In older literature . Since R(s) > 0 and G _ is concentrated on (oo.5. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1..T)lt + t = s(sI . T) with a+ = a(QT)/at. and define hi = (piI .4.... Hence with h = (sI T).. Notes and references Results like those of the present section have a long history. .type with representation (a+. Pd with corresponding eigenvectors hl. . letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . . Q = CD1 where C is the matrix with columns hl. we have IG_ [s] I < 1 .
whereas the approach was introduced in queueing theory by Smith [350]. where R is an unknown matrix. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). The exposition here is based upon [18]. It turns out that subject to the phase. Neuts and his students. For surveys . In queueing theory. e. the background Markov process with p states is {Jt}. Here phase.type assumptions are basic.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18].F. The arrival rate in background state i is a. We assume that each B.type assumption . and appears already in some early work by Wallace [377].) d (see. Asmussen & O'Cinneide [ 41] for a short self. with representation say (a(' ). Numerical examples appear in Asmussen & Rolski [43]. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.234 then in transform terms CHAPTER VIII. see Neuts [269]. [270] and Latouche & Ramaswami [241]. involving . E(t)). For further explicit computations of ruin probabilities in the phasetype renewal case . is phasetype. The distribution of W comes out from the approach but in a rather complicated form . the intensity matrix is A and the stationary row vector is ir . and the distribution of an arrival claim is B. T('). starting around in 1975. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t.contained derivation). That is .. the ruin probability can be found in matrixexponential form just as for the renewal model.. The number of elements of El=> is denoted by q. an alternative approach (the matrixgeometric method ) has been developed largely by M. see Dickson & Hipp [118].g. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. [119]..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. In risk theory. The solutions are based upon iterations schemes like in Theorem 4. The matrix. the fixpoint problems look like R=Ao+RAI+R2A2+ . a pioneering paper in this direction is Tacklind [373].
6. p = ql = Q2 = 2. Section 5b then gives a representation along the lines of Theorem 4. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.4. The version of the process obtained by imposing reflection on the V component is denoted a Markovian fluid and is of considerable interest in telecommunications engineering as model for an ATM (Asynchronuous Transfer Mode) switch. •. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details). The key unknown is the matrix K. 5. has states o.1. and the one E(•) for B. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. states . Diagonalization Consider a process {(It. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It.1 In Fig. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig.2. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms.Vt)} obtained by time reversing the I component. The two environmental states are denoted o. say with slope r(i) on intervals where It = i. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. the phase space E(°) for B.1. 5. O. 5a Calculations via fluid models.5. However. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.
< oo for all s. of E into components indexed by E. resp.1 A complex number s satisfies 'A+ (f3i(Bi[s] . o. a) of {It}. F = E U { (i. 4. •. Recall that in the phasetype case. Eli) + Proposition 5. j = 1.1) if and only if s is an eigenvalue of E.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. If s is such a number. we have more martingales at our disposal. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . a) = 1. i E E. The fluid model on Fig .(Ni)diag r(i.236 CHAPTER VIII. MATRIXANALYTIC METHODS 4. F is the disjoint union of E and the Eli).31a(l) (/3i)diag . In the general formulation . the probability in the Markovmodulated model of upcrossing level u in state i of {Jt} and phase a E Eli) is the same as the probability that the fluid model upcrosses level u in state (i.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . t. in the fluid model Eel'. First. V.1))diag + sII = 0 (5. r(i) _ 1. 5. a E E(i) } . a) : i E E. A claim in state i can then be represented by an E()valued Markov process as on Fig. This implies that in the fluid context. Second.1(a). Let E denote the matrix .Vt)} is then obtained by changing the vertical jumps to segments with slope 1. corresponding to the partitioning + Epp). Bi[s] = a(i)(T(i) + sI)it('). 5. consider the vector a satisfying (A + (13i(Bi[ s] . 4. 2. 4}. '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. Thus F = {o.1))diag ) a = sa and the eigenvector b = . whereas Ee8s' = oo for all t and all s > so where so < oo.1(b) {(It .
A .5. Then E21c+E22d = E21a . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI + E12 (sI . it follows that if Qla(1) 0 0 . and let d = (sI . t(1) 0 0 then also 0 t(2) 0 .E12E22 E21 I . Then (up to a constant) c = a.E22)1 E21a.(sI .sI) (sI . E(1) + + E(P). Noting that E11c + E12d = se by definition.sI 0 0 0 T(3) .Nla(1) 0 0 T 1. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .sI+ ((3ia(i)(T(i) .sI 0 0 t(3) 0 0 = 0. c = a.A . with Eii replaced by Eii .E21a + sd = sd.1).32a(2) (/3i)diag .sI)1t)) iag I = 0 which is the same as (5. assume that a is chosen as asserted which means (Ell .E22 . resp . d = (sI E22)1E21a = E ai(sI . MARKOVMODULATED INPUT 237 indexed by E.sI 0 0 0 T(2) . For the assertions on the eigenvectors. 0 .E22)1 E21) a = 0. iEE (a> of 0* 1 AI.sI ()3i)diag .sI.E22)1 E21a E21a . where c.T('))1t(i) .
To determine 0 (u). u) Iw(u. j. .a)d^ ).. v. the result u follows..( u. Thus 0(u) = esu/d = pe7 ° as u should be. sq with $2s. c j.j)c v . Example 5 .Q. a) = (j./' u = e' (esiuc ( 1) .v) = v) I. Proof Writing Or'Alb( v) = svb( v) as (AI . w(u)=inf{t >O:Vtu}. Then we get V)i (u) as sum of two exponential terms where the rates s1. it follows by Proposition II.. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. v) yields C{V) = e8 .. j) pi( u .5. B2 are both exponential with rates 51 i b2. a) = Pi (Vw(u.4 that {e"1b(v) is a martingale . e89uc(e)) (d(1) . j.. . v > 0. . a )d(a + e8 °vpi (u . v.v}.j. a) and noting that i1 (u) = >I j.upi(u.v) = (j. .v) = Optional stopping at time w (u.v) = = p i( u .2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. a)).4 Assume that E has two states and that B1. q. Then . pi(u. a). d("))1 e. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.a Solving for the pi(u. I' i( V P2 (w (u) < oo..O.sv)b(v) = 0. Example 5 . . a)).3 Consider the Poisson model with exponential claims with rate 5. define w(u.. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.. Iw(u.v) = j)..pi(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . v.v..v)=inf{t >0:Vtu orVt=. j. Here E has one state only.. v = 1. . w(u.238 CHAPTER VIII. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . j.. MATRIXANALYTIC METHODS Theorem 5. j... For u.
In terms of K. the Pidistribution of M is phasetype with representation (E(1) + + E(P).33(e = 0 a(j))(K ®T ( j))(ej (9 I).2. (5. •) is phasetype with representation (E(i). dx)Bj(y . (') a T( However . j.h. i. 0 Theorem 5 .x) 00 f ° (') (j) eT (yy)edx . 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 5.3) .y = to B k7 j # k In particular.5 G+(i.b (u) = Pi(M > u) = 9(i)euue. Proof We must show that G+ (i.k.s. 8^')IT(j)) where e 3^') =. U) where t(j) + t(j)O(j j = k uja. (y.5. according to VI. j. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.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.( 2.xxej • a 00 oo el .3j eye.6 For i E E. j. 9(').2) the l. is 0 /3 f R(i . oo)) j)ye.
that the density b(x) can be written as aeTxt for some row vector a. a m.. say.. For a transition from (j. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. which occurs w. see Example 1.. intensity matrix U. the initial value of (i. equivalently. . Starting from Jo = i. we have the additional possibility of a phase change from a to ry within the ladder step. t) is the representation of the matrixexponential distribution/density): Proposition 6. An alternative characterization is that such a distribution is matrixexponential. bn1 bn). T. and a new ladder step of type k must start in phase y. Furthermore. However.y) to occur when j # k.g. t = (0 0 .2. Numerical illustrations are given in Asmussen & Rolski [43]. which occurs at rate t(i). Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. Bk7 . if b* [0] = b1 +b20+b302 +. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 .p.) which is rational. some square matrix T and some column vector t (the triple (a. i.k y.e.1 Let b(x) be an integrable function on [0. Associated with each ladder step is a phase process. This yields the asserted form of uja.. a) to (k.5). with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). (6. and lifelength M. the current ladder step of type j must terminate. we have sofar concentrated on a claim size distribution B of phasetype.. Then b*[0] is rational if and only b(x) is matrixexponential. which occurs at rate t^^7. i.. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . 0 1)'.f.2) . 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. For j = k. the ratio between two polynomials (for the form of the density. +bn0i1 0n +a10n1 +.240 CHAPTER VIII.. and it just remains to check that U has the asserted form..e. a) is obviously chosen according to e(`). oo) and b* [0] = f °O eBxb(x) dx the Laplace transform.
. (6.e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. u Remark 6.. t). cannot be phasetype. The converse follows from the last statement of the theorem. we can always obtain a real one (a. bn of the denominator to be distinct and expand the r. . One of his elementary criteria. 0 0 .2). ..4) 0 0 1 c This representation is complex. (6. personal communication)..2).(6. Example 6 . . T= 0 0 1 . of (6. S.1) as E 1 c. b(x) > 0 for x > 0.3) 0 0 0 0 0 .47x2 3 1 0 .. then b*[0] = a(0I T)1t which is rational since each element of (01 ..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. where c = 1 + 1/47r 2.2 A remarkable feature of Proposition 6.1 0 0 )3 = (111). since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6..3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). but as follows from Proposition 6.1. Thus. . shows that the distribution B with density b(x) = c(1 cos(21r x))ex. Writing b(x) = c(e( 2ni1 ) y/2 . a2 a1 Proof If b(x) = aeTxt.3) was suggested by Colm O'Cinneide. For a proof.. s) is given by 27r i . see Asmussen & Bladt [29] (the representation (6.an_3 an _ 4 . t= 0 . S = f c/2 0 21ri . T.T)1 is so. (6.47r2 3 . s = c/ 2 . 0 1 an an1 an _2 . MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . Namely./(0 + bi).6...1 0 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. 1 .h. u giving b(x) = E 1 ciebiz/bY.s. 0 0 0 0 1 0 0 . namely to asssume the roots 6l. matrixexponentiality implies a rational transform.
242 CHAPTER VIII.3. t) of b(x). (6. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.5) Thus. For the first. leading to matrix calculus in high dimensions when b is small. 0. t) a phasetype representation with a the initial vector. (6. We recall (see Section 3. and that the minimal number of phases in a phasetype representation increases to 0o as 5 .6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. Consider the distribution with density = 15 ((2e2x . recall that t = Te) that if B is phasetype and (a. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. Then (cf. that despite that the proof of (6. we use a representation (a. T. Corollary 111.4) 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)) . then: Proposition 6. For the second algorithm. q are polynomials without common roots. .6) The remarkable fact is. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . T.5 (6.6) holds true also in the matrixexponential case. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). and can use this to invert by the method of Proposition 6.1)2 + 6).1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. we take as starting point a representation of b* [0] as p( O)/q(9) where p. MATRIXANALYTIC METHODS Example 6 . T the phase generator and t = Te.1 to get i (u) = f3esus. As for the role of matrixexponential distributions in ruin probability calculations. and present two algorithms for calculating '(u) in that setting.
6b* .to+)1T .T . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .7) 9( cf.1t du = .T)1 J0 00 b(x) dx = f aT1t. we get (91. U =. b+ = a+(9I .A .T .to+)1T .T)1T 2 = and 1 = AB IT2 + 82T . .a+(9I .T)1 (91.T .6. since (91T)1T .1t = f3a (0I T)1T1t . we get b+ = 0aT1(9I T).1 = A1 .1UB(B + BVA1UB).T)1t ( l .1 + b+ = b++ 1 . (6. From the general matrix identity ([331] p.to+)1 = (BI .1 = ^(T1 + ( 91T)1).b* (6. 519) (A + UBV ). this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain . with A = 91T.1BVA1. Presumably. b+ = a +(BI .T)1ta+(OI . (6.'t. the assertion is equivalent to a+(BI .6).T)1 so that b* b** b** a+(9I .T)1t.T).5 ).1 + 82 (9I .T)1 + 1 ib* (91. (91.t.1t = b* . but we shall give an algebraic proof. Then in Laplace transform formulation .B=land V=a+.T)1T1t. xb(x) dx = aT2t. Now.T)1t)1a +(9I .T)1 + (6I .
T)1T. see Asmussen & Bladt [29]. .T)1T2t . /3aT1(0I . 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. but the argument of [286] does not apply in any reasonable generality). some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). T). Much of the flavor of this classical approach and many examples are in Cohen [88].la.1. premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E.1. cf. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.8 a(T1 + (01.T)1)t = 8 (1 .5 is similar to arguments used in [29] for formulas in renewal theory.1t = /3a (9I .3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. MATRIXANALYTIC METHODS .1) is the same as the r. See Fig. the ruin probability(u) was found in explicit form for the case of B being exponential. 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. see the Notes to VII.3 is taken). For expositions on the general theory of matrixexponential distributions. of (6. VII. a key early paper is Cox [90] (from where the distribution in Example 6.1.7).h.244 CHAPTER VIII.82b*. From this it is straightforward to check that b**/(b+ . 3. a.s. In Corollary VII. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. which is selfexplanatory given Fig.1.8. 0 Notes and references As noted in the references to section 4. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].b*). 7. The proof of Proposition 6. Lipsky [247] and Asmussen & O'Cinneide [41].
P(tl.u)e = A(u)e (7. O<. Let P(tl. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i. 0 < t < u. though still Markov.I] I 80 { tq f Q(v) dvl t1 1 .1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . t) is the vector of state probabilities for mt. t2) = exp where Q(t) = ds [P(t. Proof The first statement is clear by definition. in contrast to Section 3. t + s) . Figure 7. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. i. By general results on timeinhomogeneous Markov processes.7. Given the v(t) have been computed. the definition of {m8} depends on the initial reserve u = Ro. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.1z I.1) where A(t) = v(u)P(0.t)). Also. Ai(t) = P(mt = i). we obtain V)(u) = P(m„ E E) = v(u)P(0. is no longer timehomogeneous.tl < t2 < u. Note that in general >iEE Vi (U) < 1.e.1 The difference from the case p(r) = p is that {m2}.t2) be the matrix with ijth element P (mt2 =j I mtl = i). In fact.
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. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. 0 < t < u. or it stops between level u + p(u)dt and u. A'(t) = A(t)Q(t) = A(t)(T + tv(u . vi. 0 Thus.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). In the first case. whereas in the second case the probability is p(u)dt • tjvi(u).3dt. Given this occurs.(u) p ( u) = . The intensity of a jump from i to j is tij for jumps of the first type and tivj(u .t and being followed by a downcrossing. Proposition 7. the probability of which is 1 .Sj i)p(u)dt • tji = Sji + p(u)tji dt. we get vi(u) = aidt + (1 .4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. Given A'. {mx} has jumps of two types.246 CHAPTER VIII. Given A. from a computational point of view the remaining problem is to evaluate the v(t). Hence Q(t) _ T + tv(u . given A.t) for the second. dt]. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). (7.t).t)). jEE . MATRIXANALYTIC METHODS However. Thus. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .2 For i E E. 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. the probability that level u is downcrossed for the first time in phase i is ai. two things can happen: either the current jump continues from u + p(u)dt to u.
(u) for any values of u and v such that u < v. supRt>v l t<7 I where o..5). This yields v.. we can first for a given v solve (7. To deal with this.4) backwards for {va (t)}v>t>o.) + 0 as v + oo. Rt . say. refer to the modified process.^ 0. Let p" (t).2.i7rT1/p.s. consider a modification of the original process {Rt} by linearizing the process with some rate p. we have p(r) = p = vi (u) 0aTe.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. after a certain level v. V . RESERVEDEPENDENT PREMIUMS 247 Subtracting v. <oo. (u) on both side and dividing by dt yields the asserted differential u equation.. F" etc.3 For any fixed u > 0.)P"(AnB. Since the processes Rt and Rt coincide under level B. say. When solving the differential equation in Proposition 7.h. vi (U) = lim v= (u). denotes the time of downcrossing level u . then P(A n Bv) _ P"(A n BV'). of (7.7. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.) P"(AnBv) = P(AnB. Now since both P(A n Bv) 3 0 and P"(A n Bv) . Thus. starting from v"(v) = . (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. .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. From Section 3. Then P(B. and similarly P"(Bv) .
. say.10 that in addition to the O'(•). which is available since the z/i'(.) are so.z51(v)). i. where v = inf It > 0 : Rt < v}. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. the evaluation of Vi(u) requires q(u) = 1 . T). > 0} and the last term the contribu tion from {R. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). (7. Therefore u pl(vvueTa t 1. p(r) P.1. The f iin in (7. (u)}. where.1. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. v = u. cf.7) equals 01 (v . To evaluate p1(u).1a.RQ (defined for or < oo only) is defective phasetype with representation (v(u). for u > v the distribution of v . Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u.9. the probability of ruin between a and the next upcrossing of v.1.zp1(u)/(1 . The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. p2. 2/n. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. Corollary 3. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations.V" M 0 . The precision depends on the particular quadrature rule being employed. such that Rt coincide with RI under level v and with Rt above level v.248 CHAPTER VIII. 1/n. assuming u > v for the moment. as well p1(u)..7) f o (the integral is the contribution from {R..e.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.x) dx f v(u)eT xt dx . < 0}). MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.v v(u)eTat 1 1 .. r<v r > v. Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations. However. 2u. let v(u) = a+2ieiT +ta+>)(uv). 0 < u < v. The trapezoidal rule used in [288] gives a precision of 0(n 3).. 3u etc.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. We recall from Propositon VII.
2.(7 The arrival rate is (i = 3. (7. so we consider the nontrivial case example p2 = 4 and p1 = 1.and A2 = 3 .v(u)eTve).to+))11 {e{T®(Ttoy+ ))}„ .u .1 from which we see that pl (u) = 1 + 1 249  1 ..x) dx} V 1 1(v) f V v(u) eTxt.e6u 35 .24e.x) dx 1 ^(v) ( 1 .4 Let {Rt } be as in Example 3.be the eigenvalues of T + to( 2 ).2.v(u ) eTV e J v(u)eTxtz/)l (v . p2 < 3. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e . RESERVEDEPENDENT PREMIUMS 1 .2V"2.v(u)eTVe . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .01 (v .24ev . Example 7. From Example 3.8) The integral in (7.e6v Let Al = 3 + 2V'2.jl (t ®e) Thus.^1(v) 1 .7.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. I. all quantities involved in the computation of b(u) have been found in matrix form.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.21 = ? yields 0(u) = 1.v) + (2^ + 3v2 ea'(u " . Since µB = 5/21.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .e.
/2) ea 1(u .24e5v . 192esv + 8 P1 .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived..24es" .7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.v)esv + 7 4_ 2.1)' ?. ) e sv + ( 2v/2. and one gets 12e5" .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.21(35e6v . 21 3 In particular. Notes and references [30]. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . The analysis and the example are from Asmussen & Bladt . MATRIXANALYTIC METHODS From (7.1 V2 = 4e5"+6 35e6v .24e5v ./2 ea1(u") .1.1 Thus.24es" .250 CHAPTER VIII.2 35e6v .
(b) the lognormal distribution (the distribution of eu where U . B(x) = ex0 with 0<0<1. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail.4.N(µ. x 4 oo. A rough distinction between light and heavy tails is that the m.g.2b. For example. and instead we shall work within the class S of subexponential distributions . For the definition . a2)) with density 1 e(logyFh) 2/2az . we require that B is concentrated on (0. x 2iror2 (c) the Weibull distribution with decreasing failure rate .f. 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 all t > 0. The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. III. oo ) and say then that B is subexponential (B E S) if 251 . For further examples.46 and at numerous later occasions require a light tail.B(x). see I.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. L(tx)/L(x) 4 1. the exponential change of measure techniques discussed in II.
(1. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. 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. one can check that x x where U is uniform on (0. X2 with distribution B. that is.v. note first the following fact: Proposition 1.p. given X1 + X2 > x. we have {max(Xi. That is. (b) liminf BB(() ) > 2. X1 is w. We later show: Proposition 1.v. B(x) ax. P(Xi <yI Xi+X2>x) 1B(y). X2) > x) ^' 2B(x). As contrast to Proposition 1.2 If B E S. In contrast. proving (a). oo). B(x) Here B*2 is the convolution square.3 Consider the standard exponential distribution. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. the distribution of independent r. if X1 + X2 is large .p. Then: (a) P(max(Xi. the r. X2 but none of them exceeds x. P(max(Xi. Proof By the inclusionexclusion formula. X2) > x} C {X1 + X2 > x}. That is. X2) > u x)/B(x) = 2. X2 > x) = 2B(x) . and thus the lim inf in (b) is at least lim inf P(max(Xi. In terms of r. Thus the liminf in Proposition 1. To capture the intuition behind this definition. . oo). x 3 00. then (with high probau bility) so are both of X1. the behaviour in the lighttailed case is illustrated in the following example: Example 1.2B(x). 1/2 'typical' (with distribution B) and w.v.1 Let B be any distribution on (0.1(b) is oo.'s. Since B is concentrated on (0.2. 1/2 it has the distribution of X1I X1 > x. X2) > x}. X2) > x) is P(X1 > x) + P(X2 > x) .1) then means P(X1 +X2 > x) 2P(Xi > x).'s X1. then P(X1>xI X1+X2>x)* 2. 1).F(X1 > x. Thus .252 CHAPTER IX. HEAVY TAILS B*2\ 2.B(x)2 .
S)x. We now turn to the mathematical theory of subexponential distributions. then the overshoot X .S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . If X1 + X2 > x.'s: if X . and combining with Proposition u 1.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. [In terms of r. y] and (y.y)/B(x) > 1. Proposition 1. 1 < B(x ) B( x) Y) < B( 0).4 Any B with a regularly varying tail is subexponential. we get limsupB*2(x)/B(x) < 2.B*(n ) B(dz) (1.1(b) we get B*2(x)/B(x) * 2.y)/B(x) > 1 since y > 0. or they both exceed Sx.5 If B E S. then B(B(x)y) * 1 uniformly in y E [0.B E S. yo] as X + 00.z B(x) . Let 0 < 5 < 1/2.v. Finally lim inf B(x .xIX > x converges in distribution tooo.B*n(x .1. B( 0 .B(y)) .yo]. If lim sup B(x . Hence lim sup a+oo B*2(x) 2B((1 .6)x)/((1 . x]. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . then either one of the Xi exceeds (1 .B(y) = 2. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. a contradiction. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1.] Proof Consider first a fixed y. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .
then for any n B*n(x)/B(x) * n. then e"R(x) * oo. Given e > 0.2. Proposition 1.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. Then by (1. 0 Proof of Proposition 1.B*2 (x) B(x) (x .6 If B E 8.B(x .z) B(dz). and this immediately yields the desired conclusions.254 CHAPTER IX. Proof We use induction. x oo. Proof For 0 < 5 < e. This implies B(x) > c2e5x for all x. its intuitive content is the same as discussed in the case n = 2 above. HEAVY TAILS Corollary 1. B*(n+1) (x I xy + Jxx y) W.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . The first integral is y B(x . so assume the proposition has been shown for n. choose y such that IB*n(x)/B(x) .5 and the induction hypothesis.1) for all large n so that B(n) > cle6n for all n.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) .5 and dominated convergence. 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 . The case n = 2 is just the definition. B(x) \Jo _ B(x .z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.nI < e for x > y.. we have by Proposition 1.z ) (x ) = 1 + (^ B(x .5 that B(n) > e6B(n .z) B(dz) _y B(x) 111 Lx B . O The following result is extremely important and is often taken as definition of the class S.7 If B E S.2). b[c] = oo for all e > 0.
z) B(x) < 1 + A + an sup f x B(x .0 completes the proof.'s such that Xi has distribution Ai.y)Ai(dy) = (x)o(1) (1.X2 > xv) < A1(xv)A2(x v) . Combining these estimates and letting a 4. Proposition 1.2). Proof Let X1. 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.4) . it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Proof Define 5 > 0 by (1+5)2 = 1+e. an = supx>o B*n(x)/B(x). For any fixed v. e > 0. A2 be distributions on (0.1. Xi <= v Ai (x .5 easily yields P(X1 + X2 > x. an+1 fX B*n( *n(x . 0 Lemma 1.z) B(dz) < 1 + A + an(1 + d) . Then Al * A2(x) = P(X1 + X2 > x).ala2B(x)2 which can be neglected.9 Let A1. Then Al * A2 (x) .v. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.y))/B(x).i).z) B(x .ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .y)Ai(dy) v) f o . choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). 0 Proposition 1.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .z) B(dz) x .y)B(dy) = B(x)ov (1)• v (1. Since P(X1+X2 > x.B(x .(al + a2)B(x). X2 be independent r. Then by (1. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. a2 with a1 + a2 > 0.8 If B E S. 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) .X1 > xv.
However.11 Let B E S and let A be any distribution with a ligther tail.y)Ai(dy) = B(x)o„(1).v)B(v) + _'U Aq(x .256 Now (1.9). Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. if q(x) aB(x) for some B E S and some constant a > 0. .aiB(v)) = B(x)o„(1). Hence Corollary 1.5) Here approximately the last term is B(x)o„(1) by ( 1.2A(x).v)Ai(v) .13 Let B have density b and failure rate A(x) such that . HEAVY TAILS 'VV B(x . L2 slowly varying. i = 1. f " By a change of variables.(x) is decreasing for x > x0 with limit 0 at oo. A(x) = o(B(x)). Then B E S provided fo "O exA(x) b(x) dx < oo. it should hold that B1 * B2 E S and B1 * B2 (x) . then so is L = L1 + L2. of (1.4). B2 E S. That is. a2 = 1. B1 * B2 (x) .2. the l. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. it is easy to see that if L1.2aB(x) .Bl (x) + B2 (x) when B1. V (1.B(x) Proof Take Al = A.12 Assume that Bi(x) = Li(x)lxa. That is. whereas the two first yield B(x)(Ai(v) . then A E S. A2 = B so that a1 = 0. Then A * B E S and A * B(x) .Ai(x .Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.h.s.3) follows if CHAPTER LX. In the regularly varying case. u Corollary 1. L2 are slowly varying. u It is tempting to conjecture that S is closed under convolution. with a > 0 and L1. a1 = a2 = a yields A*2(x) .y)B(dy).10 The class S is closed under tailequivalence.5) becomes x B(x . u Corollary 1. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. Proof Taking Al = A2 = A.
Define A(x) = fo . proving B E S. subexponentiality has alrady been proved in Corollary 1. L(x) y° (a .1.. 0 A(x) . B*2(x) . we can use the same domination for the second integral but now the integrand has limit 0 . By (1. the first integral has limit 1 .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . Example 1. The middle bound shows that it converges to b(y) for any fixed y since \ (x .y) < yA(x . Goldie & Teugels [66]): Proposition 1. Since ) (x . a(x) = ax01. x .A(x .y) dy.U) /or) v 2x This yields easily that ex.12.A(y)\(y) dy + fox/ 2 eA(x ).13 works in this setting. Thus by dominated convergence . an integrable function by assumption.1 has limit 1 + 0. Further.A(y)a(y ) = ev'(y) b(y). we first quote Karamata's theorem (Bingham. f ' L(y) dy . the DFR Weibull distriu bution is subexponential. Thus B*2(x )/ B(x) .1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x .3 < 1.y) < A (y) for y < x/2. Thus. Example 1.2).`(x)b(x) is integrable. Thus A(x) is everywhere decreasing.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).15 In the lognormal distribution.1)xcl1 .14 Consider the DFR Weibull case B(x) = ex0 with 0 <.. In the regularly varying case.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).(x . and exa(x)b(x) = (3x01e(10)x9 is integrable.16 For L(x) slowly varying and a > 1. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. Jo For y < x/2. Then b(x) = Ox0lexp. To illustrate how Proposition 1.A(xy)A ( y). replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. Then B(x) = eA(x).(y) dy. the u lognormal distribution is subexponential. Thus.y) * 0.
L(x)/x" and )t(x) . the monotonicity condition in Proposition 1.13 may present a problem in some cases so that the direct proof in Proposition 1.ea b(x) is integrable.1)] I X > x) L(x[1 + y/(a . More precisely.1)])a 1 1 .a/x. we get (1. (1 + y/(a .258 From this we get CHAPTER IX. Proof ( a): Using Karamata's theorem.y(x)B(x).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 . yo] .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)X(x)/x > y) = P(X > x[1 + y/(a . HEAVY TAILS Proposition 1.1))a .)/Y(x) > y) (1 + y/(a .6) EX(x) . 'y(x) = EXix>.1/A(x) and P(X ixil'Y (x) > y) * e'.4 is necessary in full generality. f O B(y) dy . then 7(x) x/(a . then B(x) . Then: Proposition 1. (c) Under the assumptions of either ( a) or (b). the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.1)]) xa L(x) (x[1 + y/(a . Thus exa(x)b(x) .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. let X W = X .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. Then 7(x) .1) and P(X (.xjX > x. However.E(X .
Notes and references A good general reference for subexponential distribution is Embrechts. u a oo. i. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). with EzK < oo for some z > 1. nG(u). then Vi(u) P Bo(u).v. . 0 G(u) L G(u) .1 If Bo E S.EK G(u).+YK> u) = ^•P(K = n)G* n(u ) . be i. Let St = Ei ` Ui . . r(u) = inf it > 0.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .n0 1•P(K= n)•n = EK.15. Bo(x) = f0 B(y) dy / µB.2. Y2.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. 1..14. Then P(Y1 + • • • + YK > u) . THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . St > u}. with common distribution G E S and let K be an independent integervalued r.and lognormal distributions . Kliippelberg & Mikosch [134]. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u.. Examples 1.. Recall that B0 denotes the stationary excess distribution. P The proof is based upon the following lemma (stated slightly more generally than needed at present).1/. d.. Theorem 2 .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/.t be the claim surplus at time t and M = sups>0 St. We get p(yl+.nn. Lemma 2. The remaining statement (1. cf.7) is referred to as 1/A(x) being selfneglecting.2 Let Y1. It is trivially verified to hold for the Weibull.A(x) I X > x) = exp {A(x) .(x). 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B.
we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . then Bo(x)/B(x) + 00. r(1/Q) xlQexp B(x) = ex' From this .. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.18.260 CHAPTER IX. Proof of Theorem 2. Bo E S. However. In general: Proposition 2. (2.p) and EzK < oo whenever pz < 1.1 is essentially due to von Bahr [56].1 is notoriously not very accurate. The problem is a very slow rate of convergence as u ^ oo. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. Weibull) one has Bo(x ( B(x) .x400 PBB(x) PB Leta+oo. See also Embrechts & Veraverbeeke [136]. as well as examples where B ¢ S. x 4 00.2.1) In particular . Bo is more heavytailed than B .p)p'. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. .3 If B E S.1. in our three main examples (regular variation .. a]. u The condition Bo E S is for all practical purposes equivalent to B E S.1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) . Note that in these examples . Borovkov [73] and Pakes [280]. Bo E S is immediate in the regularly varying case.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p.µ J ) .?(xµ 8 (x). and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. The approximation in Theorem 2.x^ ) B(x) _ f or ( lox . u x+a Notes and references Theorem 2. see Abate.13). For some numerical studies. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.µB(01 . P(K = k) = (1. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . Since EK = p/(1. lognormal . the result follows immediately from Lemma 2. mathematically one must note that there exist (quite intricate) examples where B E S. Bo ¢ S.
T+ < oo) where r+ = T1 + • • • + T. there is a representation of M similar to the PollaczeckKhinchine formula. in [219] p..+ E A. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. This shows that even the approximation is asymptotically correct in the tail.] The proof is based upon the observation that also in the renewal setting. M = sup s$ . i.Ti. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). p = iB /µA < 1. t9(u) = inf {n : Snd> > u} . T1 the ith interarrival time and Xi = U.. [279].y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. We assume positive safety loading.. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. 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. Snd) = Xl +. Based upon ideas of Hogan [200]. + Xn. Somewhat related work is in Omey & Willekens [278]. . E. The main result is: Theorem 3 . 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .9+ < oo) = P(S. In [1]. also a second order term is introduced but unfortunately it does not present a great improvement.y)/B (x) > 1 uniformly on compact y internals. {n= 0. G+ (A) = P(Sq+ E A...1 gives 1010. (3. i=1 .1. Let U= be the ith claim . To Bo(u) u + 00.3..1) this end .1 when u is small or moderately large. Kalashnikov [219] and Asmussen & Binswanger [27].g. Define further 0 = IIG+II = P(r9+ < oo). THE RENEWAL MODEL 261 Choudhury & Whitt [1]. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2.} Then ik(u) = F ( M > u) = P(i9 (u) < oo). Then K M=EY.e.
(b) and does not rely on the structure Xi = Ui .oo..N. 0] to the integral is O(F(x)) = o(FI(x)).Ti).i. cf. U_ (dy) is close to Lebesgue measure on (.d. FI (x) _ fz ° F(y) dy.3 G+ (x) . oo) = F(S. (3. The heuristics is now that because of (b). Lemma 3 .Y2. x + oo.1) is equivalent to P(M > u) " .FI(u).B(x). with distribution G+/9 (the distribution of S. we will use the fact that the proof of (3. As for the compound Poisson model. x > 0. and hence FI(x) .PBBo(x). B(x) _ J O° B(B(x)y) A(dy) f 1 .9)9'' and Y1. 0] normalized by IPG_ I so that we should have to G+(x) . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.1) holds for a general random walk satisfying the analogues of (a).3) and we will prove it in this form (in the next Section.2 F(x) . P(K = k) = (1 . u a 00.(. d+ < oo). Proof By dominated convergence and (b). G_(A) = P(S.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_. Proof Let R+(A) = E E'+ ' I(S.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_. Write G+( x) = G+ ( x.262 CHAPTER IX. are independent of K and i.. HEAVY TAILS where K is geometric with parameter 9. A.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. x * oo. Let F denote the distribution of the Xi and F1 the integrated tail.y+ given r+ < oo).y) dy = 1 Pi (X) oo IPG_ I . Lemma 3 .2). 0 The lemma implies that (3.FI(x) /IPG_I. the contribution from the interval (..1 IPG_ I / F(x . whereas for large y . 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.g+ > x. Then 0 0 F( x . A(dy) = 1.
u Proof of Theorem 3.3. 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 .3.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. THE RENEWAL MODEL 263 We now make this precise.1. In the lattice case. n] < (1 + e)/1µc_ I for n > N.2).1. (3. Given e.(dy) fN FI ( x) + lim sup ZY00 N F(x .e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.G+[s]) . then by Blackwell 's renewal theorem U_ (n .1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. If G_ is nonlattice.2) yields 00 F F I (u) P(M > u) _ E(1 . n] is just the probability of a renewal at n.UG_ I x. By Lemma 3.y) U.O[s])(1 . the proof is complete.oo Fj(x) N J (1 +6)2 I {IC_ I lim sup X400 FI(x + N) _ (1 + e)z (x) I Pi µ G_ I Here in the third step we used that (b) implies B(x)/Bo(x) + 0 and hence F(x)/FI(x) 4 0. Similarly.1. and that U_(n . > (1 .9)IpG_ I Differentiating the WienerHopf factorization identity (A.1. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.1. we can assume that the span is 1 and then the same conclusion holds since then U(n .I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .9) 1 . Hence using dominated convergence precisely as for the compound Poisson model.1 I . choose N such that F(n .=1 BIp G_ I (1. F(Y= > x) FI(x)/(OIp _ 1).F[s] = (1 . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .0)0k k I(u) A. n] + 1/I µG_ I.
on the set {M > u.4 on the joint distribution of (S. Note that substantially sharper statements than Lemma 3. u)). allowing also for possible dependence between the arrival process and the claim sizes. Mn < u}.a. HEAVY TAILS µF = (1 .Sty(u)_I < a} we have w(u) < oo.Se(u)_1 < a) = o(Fj(u)).(1 .So(u)) are available. . Therefore by Lemma 3. Notes and references Theorem 3. Sty(u) .a. must attain a maximum > 0 so that P(M > u.0)ua_ .a) N P(M > u). we have P(M E (u . with roots in von Bahr [56] and Pakes [280].1)6+[0] .. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .SS(u)}n=o. 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)) > P(w(u) < oo)(i lp (0))• On the other hand.a. S+9(u) . Then P(M E (u . S+q(u) . 10(0) But since P(M > u .AB iP We conclude by a lemma needed in the next section: Lemma 3 .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . Proof Let w(u) = inf {n : Sid) E (u . P(M > u.(u)+n . see Asmussen & Kliippelberg [36]. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.264 and letting s = 0 yields CHAPTER IX.u)) = o(P (M > u)) = o(FI(u)).1 is due to Embrechts & Veraverbeke [136].IIG+II)µc_ = (1 .yiui_1.2.4 For any a < oo. and {Su..l. u)..a.
. . Assume that the claim surplus process {St}t>o has a regenerative structure in the sense that there exists a renewal process Xo = 0 < Xl <.1 Note that no specific sample path structure of {St} (like in Fig.n n=0..1) . (corresponding to the filled circles on Fig.. < 0 and EoX < oo where X = X2 ..... +1. 4.. M... M* = max S.4.1. The idea is now to observe that in the zerodelayed case. see [47]. {SX1+t . E0.i. Theorem 4.Sxi}0<t<x2Xl .1 = max k=0. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. We give here one of them. {Sn}n=o. Figure 4.Sxk}o<t<xk+1xk is the same for all k = 1. 2.F*(X) = P0(Si > x) . Schmidli & Schmidt [47]. 4.1 except for the first one) is a random walk. Thus the assumption . and apply it to the Markovmodulated model of Chapter VI. M = sup St.. We return to this point in Example 4.. such that {SXo+t  SXo}0<t< X 1Xo ... examples and counterexamples. See Fig.d.X2 < ..X1 is the generic cycle.. We let F* denote the Podistribution of Si. Define S.4 below.1 where the filled circles symbolize a regeneration in the path... (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.1 based upon a regenerative assumption. t>0 S.1. 0o(u) etc.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. = Sx. assume pp.1) is assumed. and the distribution of {Sxk+t . G(x) (4. 4. For further approaches..
2) to show F(M* > u) > 1. N N Xi=0 N Figure 4. The one we focus on is Fo (Mix) > x) .. Fo(Si > X).3) hold. Proof Since M > M*.2..4) liminf u>oo F(M > u) .* i o<t<xn+1x. u p 00. jF11 F* (U). Sxn +t . it suffices by (4. (4.1 Assume that (4. See Fig. HEAVY TAILS for some G such that both G E S and Go E S makes (3. the assumption means that Mix) and Sl are not too far away. Since clearly M(x) > Sl .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) applicable so that F(M* > u) 141 F*(u).2 Theorem 4.S. Then '00 (u) = Fo(M > u) .266 CHAPTER IX.1) and (4.3) where Mnx) = sup o<t<xn +1 X. (4. (4.. 4..Sxn = sup Sxn+t .
assume the path structure Nt St = EUit+Zt i=1 .Mn +1 >aV(u n=1 00 S..5) which follows since Po (M > u. To this end.4. E (u .E) Po ( n max St u. u))/P(M* = 0) = o(Po(M* > u)).S.e)Po (MMX> > x).: Sn > u} .(u) . 2. 0 yields (4. )) > (1 . Under suitable conditions .. Mn+l > a V (u .. choose a such that Po(Si > x ) > (1 .4.Po (M* > u. Let a > 0 be fixed. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . Theorem 4. /3(u) = inf{n=1.a.. S. u)} < P(M* E (u . Then by Lemma 3. MW O(u)+1 < a) IN ( U n=1 A1. Po(M* > u) .. .2.1 can be rewritten as 00 (U) (4.. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4.1 = limti00 St/t.: S.(1 .. Letting first u + oo and next e .e)Po (M > u.Sn 0<t<x„+j ( 1 .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.( u)1 > a) 00 1: Po(Mn<u.a. M^xu)+l > a) .+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). x > a.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .4).e)Po (M > U). Given e > 0.
X and N. independent of {> CHAPTER IX.} and satisfying Zt/t N. we get 00 (u) 1 IPF. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .Q = EoNx/EoX. the proof of Lemma 4. Mix) < > UE + i=1 o<t<x Thus Theorem 4.6 below.X both have tails of sup Zt.4).8) x Write . cf. since the tail of Zx is lighter than B(x) by (iv).3PB. and ENX Ui . HEAVY TAILS N` U. Assume further that (i) both B and Bo are subexponential.7).I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .1 is in force. (ii) EozNX < oo for some z > 1.268 with {Zt} continuous. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.'s order EoNx • B(x).p) Ju P Bo(u) 1p 0 . i=1 (4.6) holds with p = . The same is true for Sl. a4' 0. Proof It is easily seen that the r.v. oX (see Proposition A1. Corollary 4. (iii) For some o field Y.2 Assume that {St} is regenerative and satisfies (4. and also for Mix) since Nx FNX U.
More precisely..4.e. and for some constants ci < oo such that cl + • • • + c.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. . Again . 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. X3 = 2. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. MODELS WITH DEPENDENT INPUT 269 Example 4 .. Bo E S. we conclude just as in Example 4. we assume that B E S. Bo E S. Zt . (iii) is obvious.6) u holds. X2 = 1.9). Assume that B E S. Then (4. i. 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) holds.6) holds.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). Theorem 4. > 0. and taking F = o.3 that (4. of claims arriving in [0. . The key step of the proof is the following lemma. In particular. < 1. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. The number N.t + EN'I Ui where {>N`1 Ui ...0 (thus (iv) is trivial). We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. . The arrival rate is /3i and the claim size distribution Bi when Jt = i.6) holds. Thus we conclude that (4.3 As a first quick application. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. The regenerative assumption is satisfied if we take Xo = Xi = 0. . Taking again Xo = Xi = 0. in particular lighttailed. X2 = 1. Example 4 .(NX). consider the periodic model of VI. We consider the case where one or more of the claim size distributions Bi are heavytailed.4 Assume that St = Zt . then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1..
.."+Np . u Proof of Theorem 4.. 2 . . we can define the regenerations points as the times of returns to i.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . oo) and define p Ni Yx = EEX'i . Thus dominated convergence yields ( P(Yo>x P(Yo>x .v. .5. 1. .. HEAVY TAILS Lemma 4 .. Let {Fi}t=1 P be a family of distributions on [0. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . . i=1 Proof Consider first the case X = 0..270 CHAPTER IX.c'(x) where c = ciENi . X > 0 a r. i1 = E\ G(x) In the general case.. 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. and F a aalgebra such that (N1. cp with cl + > 0 it holds that Fi(x) . oo) such that G E S and some c1.G( x ) > ciNi . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. P P P(YX and > x I. as x a oo. . are independent with distribution Fi for Xij.}P. i=1 P(Yx > x ^) < P(Y0 > x I.F) < CG(x)zn'1+. NP ) be a random vector in {0. If Jo = i.2. For lighttailed distributions.X i=1 j=1 where conditionally upon F the Xi. . and that for some + cp distribution G on [0. Then P P(Yx > x) .Fmeasurable. NP ) and X are . An easy conditioning argument then yields the result when Jo is u random. The same dominated convergence argument completes the proof. and the rest of the argument is then just as the proof of Corollary 4. 6 Let (N1. ..ciG(x).^•) G(x) P ^ E ciNi = C.
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). Theorem 2. this is applied for example to risk processes with Poisson cluster arrivals.4. and the final reduction by Jelenkovic & Lazar [213]. this should be compared with the normal limit for the lighttailed case. The main result of this section. r(u) is the time of ruin and as in IV. cf. i. For further studies of perturbations like in Corollary 4. Combined with the approximation for O(u). cf.. Notes and references Theorem 4.4. states that under mild additional conditions. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential.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.T2. Theorem 4..i. the discussion provides an alternative point of view to some results in Chapter IV.pl(1 .5 shows that basically only the tail dominant claim size distributions (those with c. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form.4. As usual. 5a Excursion theory for Markov processes Let until further notice {St} be an arbitrary Markov process with state space E (we write Px when So = x) and m a stationary measure. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).7). We start by reviewing some general facts which are fundamental for the analysis.T2. It follows from Theorem 4. VI. I T(u) < oo). Within the class of risk processes in a Markovian environment.1 is from Asmussen.. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. ) form a general stationary sequence and the U.4. see Schlegel [316]. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. Then O(u) .e. Theorem 5.7.1. An improvement was given in Asmussen & Hojgaard [33].2 and Example 4. The present approach via Theorem 4. 5 was first proved by Asmussen. we let PN"N = P(. m is a (orfinite) . cf.5.. In contrast. IV.. in particular Proposition 2..3. > 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. and independent of (T1. i. Essentially. as well as a condition for (4.d. ). Schmidli & Schmidt [47].6) to hold in a situation where the interclaim times (T1.p)Bo(u).
a familiar case is time reversion (here m is the stationary distribution). 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} . k on E.t.272 CHAPTER IX. and starting from So = y. Rt is distributed as x + t . Thus.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). to consider only the case Px(w(F`) = 0) 0. (note that we allow x. x = 0+ and F = (0. follows by the substitution y = x . the whole of R and not as usual impose the restrictions x > 0. . in the terminology of general Markov process theory. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. Then (5.rij = mjsji where r13.z.>N` Ui. The equality of the l. {St} and {Rt} are in classical duality w. where we can take h.)k(x .t. Sw(F. k as indicator functions. w(Fc) < oo ) 'In general Markov process theory. j. The simplest example is a discrete time discrete state space chain. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. We let QS be the corresponding distribution and Qx.y = Qx (.h. Proof Starting from Ro = x. m.= y. Lebesgue measure.s. y to vary in.h. For the present purposes it suffices .2) means ffh(a. say.r. .2) with t = 1 means m. Say {St} is reflected Brownian motion on [0. t. Let G denote the distribution of ENt U. to the r.00). r.). but the example of relevance for us is the following: Proposition 5. St is distributed as y . however .t + EI U. y = 0). and (5.s=j are the transition probabilities for {St}.2) for all bounded measurable functions h. for states i. oo). u For F C E.s. resp. {Rt}.
1 for the case F = (oo.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. . in = y..y(2p21 .5. 5.2 Qy.y = Qy Q. and we let Qy y refer to the time reversed excursion .(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. 0]. But in the risk theory example (corresponding to which the sample paths are drawn). this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. in E F. QR and QRy are defined similarly. [note that w(z) < oo a. . i1. Sw(F)1 = y) .y() = P ({SW(F`)t} 0<t<w(F °) E So = x. In particular: Corollary 5.SS(F. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.13AB < 1] Proof of Theorem 5. io = x.y as a measure on all strings of the form i0i1 . .. That is. Sw(Fo)_ should be interpreted as Sw(F^)_1). 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). .3 The distribution of r(0) given r(0) < oo..2. Thus.. oo) = r(0) x= St y (a) Figure 5.. Sn+1 E Fc) nx. The theorem is illustrated in Fig ..s. 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. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. z > 0. Qx. We can then view Qy.). in with i0.. Sn = in = y. S. when p = . /^s x (S1 = Z1.= y) Theorem 5 .itt) = P Px(w(Fc) < 00. w(0. x = 0.. the one in (b) is the time reversed path.
i„_iEF Similarly.. Rn = in = x.. i0 = y. ... . .(F<)1 = Y) S S and Qx y( ipil . R .TI( 2n2n _1 ....gilt' k=1 ii .in1 .. 20 = y.. .. MY Thus Qx(ioii .. S.rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 . = in = y.y(inin _ E SYj jEF` 00 Sxik _1 .. i0) Q x.. HEAVY TAILS E E Px (Si = 21i ....... in = x. . Si11 S 1 .. Rn+1 E FC) TioilTili2.. Sn+1 E Fc) n=1 i1...... in) = oo jEF^ Sxin1 . 2p)...... in with 20... 2p) when 20. 21 . ..ik_1EF ..274 note that Fx(w(Fc) < 00.in E F.. Rn = in = x.. .. S. note first that Pt' (R l = il.. To show Q y x (i0 i 1 ...... R Qy x(2p21 .. Si1y k=1 i1 . Rn+1 E F`) F (w(Fc) < 00..... Silt' E SO k=1 i1. 2n) = Qx.. in) = Qx. t' y and Qy x are measures on all strings of the form ipi l ... Silt' E Sxik_1 .. Si1y 00 jEF° E E 5xik_ 1 . in = x. in E F.ik_1EF Sxin_1 .ii ... in)  Pt' (R1 = ii.J (i.. . Si l io E mjSjx..ik1EF Similarly but easier Sxin_1 . .. (Fc)1 = y) 00 CHAPTER IX..
Y > y} .B(a) +a PBBo(u) .'s are defined w.p.5. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.r. Y > u). FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments .')density of Y is B(y)/[.UBBo(u)]. the case r (O) < oo. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . P(o) ).2.t. that is.2.2 The distribution of (Y. Bo") is also the P(u. 5. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. P(") = P(.r.t. the distribution w. U T(O) = T (u) Y Figure 5. S. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. y > u. Z) is described in Theorem 111.v. 7(0) < oo.')distribution of Yu is Bo"). that is. That is.(o) > y} = {T(0) < oo. Now the P(u. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . the P(u. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). see Fig.2. ST(o) > y.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . 1 w .
1. in particular of Z. . We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .276 CHAPTER IX. 5.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. Since w(z)/z a$. i. The idea is now to observe that if K(u) = n. and YI. let r+(1) = T(0).. . . P(Z < a I Y > u) 3 0. Fig. then by the subexponential property Yn must be large. must be large and Z1. are i..: r+ (n) < oo. r(u)/Z 4 1/(1 . Z = ZI but relative to the kth ladder segment. Zk be defined similarly as Y = Y1.T+(2). Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.i.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). Z1).. Then 7(u)/y(u) ^ W/(1 . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. .p). Then.. Then Corollary 5. more precisely. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. i..3) holds. K(u) = n). we get the same asymptotics as when n = 1..3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z).r+ (n . Z/'y(u) * W in Pi "' ')distribution .. > u with high probability. We let K(u) = inf In = 1. Zn). HEAVY TAILS Let {w(z)}Z^.. Recall the definition of the auxiliary function y(x) in Section 1. 1/(1 . the random vectors (YI. cf.. It is straightforward that under the conditions of Proposition 1. z ^ oo. (Y. denote the ladder epochs and let Yk. Since the conditional distribution of Z is known (viz.. a slight rewriting may be more appealing. Hence Z. That is . However.p) in Pi"'')distribution. Yn_1 'typical'. the duration T+ (n) .. .18(c) Bo")(yY (u)) + P(W > y) ( 5.3.e. Bo") ).d.. 4 Assume that Bo E S and that (5. and since its dominates the first n .1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. conditionally upon r+ (n) < oo.e. this in principle determines the asymptotic behaviour of r(u). . and distributed as (Y. In the proof. .. Z).p) then yields the final result T(u)/y(u) + W/(1 .p) in F(u) distribution.p).. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .. 2.
5. A"(u) _ {K(u)=n} = {Y1+ P(. Further.. then IIP( I A'(u)) Taking A'(u) = {Y..u) E • I A'(u)) = Bo (n1) ®Bou) . Proof We shall use the easily proved fact that if A'(u).n).. > u}.n) (y1. Y„1.3 In the following. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.u) II 0. II ' II denotes the total variation norm between probability measures and ® product measure. I A'(u)) = P(u.Bo (ri1) ®B( . I A"(u ))II + 0. Yn .. P(.Yn1iYn . .u) E •) .2. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).. P (Yj. +Yn1<u. suitably adapted). .5 Ilp(u. Lemma 5.Yl+ +Yf1>u}... . .
.. then 11P(Z E •) .' = y is BM....6. wk(Zk) has a proper limit distribution as u + oo for k < n.. in particular his Proposition (2.. ... the marginal distribution of Zk is Bo for k < n.4.i... By Lemma 5..+y 1 p"F(Yn > u) P)Pn1 P/(1 ..6 IIPIu'n ) CHAPTER IX.1 P PBo(u) • P(W/(1 .. HEAVY TAILS ((Z1'.P(Y' E •)II * 0.. P(u) since by Theorem 2. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 ..7 O (u. ..). in our example Y = (Y1. Let {wl(z)}.p) < y). n_1 < u. Proof of Theorem 5..P) > y) Corollary 5. the density of Yn is B(y)/[IBBO(u)].. . The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). For Theorem 5. Z11). the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn).. .1). Z.. Proof Let (Y11.4).. . k = 1. (Y. . the discussion just before the statement of Theorem 5. y > u. Y") u etc. Zn are independent.. . .. .y(u)T) .Bo (n1) ®Bo' 0. .P) Bo(u) for n = 1. Similarly (replace u by 0). Then according to Section 5a. . The first step is to observe that K(u) has a proper limit distribution w. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . and that Yk has marginal distribution B0 for k = 1. 2. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) . Y'. n.u has distribution Bout That is. Zn). .1.r. Z' are arbitrary random vectors. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. see Fitzsimmons [144])..t.. n .. Y1 +. and clearly Zi..2. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. copies of {w(z)}. Zn) E •) .1 and Y„ . {wn(z)} be i..d.P(Z' E •)II > 0 (here Y.278 Lemma 5 . be independent random vectors such that the conditional distribution of Zk given Y.
y) . Theorem 6 .(3 u u J B(y) dy .1 Assume that B is subexponential and that p(x) > 00.6. Then 0 (u) Qf "O ^) dy. 3. Assume for simplicity that {Vt} regenerates in state 0 . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. claim size distribution B. and premium rate p(x) at level x of the reserve. nontrivial and we refer to Asmussen [22]. the probability that is exceeds u is then B(u . cf. Then P(MT > u) .2.(u) = P(V > u) = f f (y) dy .e. Corollary II. the results only cover the regularly varying case. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . and define the cycle as a = inf{t>0: Vt=0. one expects the level y form which the big jump occurs to be 0(1). = supo<t<0. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. that fo p(x)1 dx < oo. p(Y) and the result follows. The heuristic motivation is the usual in the heavytailed area.2 Define M. however. x > oo./3Ea B(u).1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.1.. 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). Proof of Theorem 6. More precisely. u (6. V. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. that MQ becomes large as consequence of one big jump. . The rigorous proof is. 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 . T) when T + oo with u fixed.B(u). i. Asmussen & Teugels [53] studied approximations of i (u.
Hence f (u)r(u) = D(u) = Do(u) . by regenerative process theory. u Notes and references The results are from Asmussen [22].280 CHAPTER IX. Then D(u) = f(u)p(u) and. 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. Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u).q(u) Now just use that p(x) * oo implies q (x) + 0. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. D(u) = DQ(u)/µ.q ( u)) 1 . It is also shown in that paper that typically.P(MT > u) $B(u) Ft µ(1 . . where also the (easier) case of p(x) having a finite limit is treated .
. 281 . We shall be brief concerning general aspects and refer to standard textbooks like Bratley.. z) 2 = Zit NE ii ii According to standard central limit theory . 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.. 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 ). The crude Monte Carlo ( CMC) method then amounts to simulating i.. and this is the form in which the result of the simulation experiment is commonly reported. replicates Zl. Fox & Schrage [77]. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. topics of direct relevance for the study of ruin probabilities are treated in more depth.d.i. Ripley [304].z) 4 N(0. 4Z). vrN(z .96s z f (1. Hence 1. ZN.2) is an asymptotic 95% confidence interval .
Letting Z' = E[Z I Y]. 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. Therefore. and in most cases this modest increase of N is totally unproblematic. generated at the same time as Z. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. there are others which are widely used in other areas and potentially useful also for ruin probabilities. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. We survey two methods which are used below to study ruin probabilities. conditional Monte Carlo and importance sampling. one can argue that unless Var(Z') is considerable smaller than Var(Z). variance reduction is hardly worthwhile. we then have EZ = EZ = z.b(u. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. and a longer CPU time to produce one replication. However. Further. SIMULATION METHODOLOGY In the setting of ruin probabilities.282 CHAPTER X. and many sophisticated ideas have been developed. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). This is a classical area of the simulation literature. 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. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . Say that Var(Z') = Var(Z)/2. v. an added programming effort. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). so that Z' is a candidate for a Monte Carlo estimator of z. We mention in particular ( regression adjusted) control variates and common random numbers. The situation is more intricate for the infinite horizon ruin probability 0(u). Then replacing the number of replications N by 2N will give the same precision for the CMC method as when simulating N' = N replications of Z'. Z = I inf Rt < 0 (0<t<T = I('r(u) < T).
3) Thus. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. Thus we cannot compute L = Z/z (further. even if the optimal change of measure is not practical.[E(LZ)] = E Z2 Zz . This may also be difficult to assess . .v. L1). the argument cheats because we are simulating since z is not avaliable analytically. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.1.96 sis v^ N 2 1 where srs = N j(LiZi . In order to achieve (1. Thus. However.z2 = 0.E [Z Z]2 = z2 . it appears that we have produced an estimator with variance zero. (1. To this end.e..3). 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. 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. (ZN. but tentatively.. . L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0).zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . one would try to choose P to make large values of Z more likely. i. using the CMC method one generates (Z1. . and the problem is to make an efficient choice. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).zrs. L such that z = EZ = E[LZ]. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio.
N .e. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. However. 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.96 2 z2 z increases like z1 as z .96oz /(zV) = 0.100 .. and let Z(u) be a Monte Carlo estimator of z(u).5 or even much smaller . i. as is the case of typical interest. An example where this works out nicely is given in Section 3. it does not help telling whether z is of the magnitude 104. The CMC method leads to a variance of oZ = z(1 . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . We then . 10 .z) 1 > 00. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.e. Z z V5 In other words . u + oo. let z(u) = P(A(u)). To introduce these.96 2Z ( 1 . the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 .e. Thus.0. say of the order 103 or less. we may try to make P look as much like P(•IA) as possible. Again. if z is small. I. However.z) which tends to zero as z ^ 0. For each u. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large.. In ruin probability theory.z) 1001.1. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .284 CHAPTER X. the optimal P is the conditional distribution given A. and further it is usually not practicable to simulate from P(•IA). in terms of the halfwidth of the confidence interval.B = iP(AB) = P(BIA). The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. Z = I(A) and A is a rare event. say 10%. z I. large sample sizes are required. a confidence interval of width 10 4 may look small. This leads to the equation 1. assume that the A(u) are rare in the sense that z(u) * 0. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). but if the point estimate z is of the order 105.1.
. so that NE (u) may go to infinity. Generate X1. XK from the density bo(x)..log z(u) of (1. 2.1) may be written as V) (u) = P(M > u). Thus. Let M .(2. i. the PollaczeckKhinchine formula III. let Z +. Therefore .d. The term logarithmic comes from the equivalent form . the mathematical definition puts certain restrictions on this growth rate.0. P(K = k) = (1 .4).e. .2. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.log Var(Z(u)) lim inf > 2 u+oo . where X1.1. . but as a CMC method . F(K = k) = (1 . it is appealing to combine with some variance reduction method .p)pk.4) for any e > 0. Notes and references For surveys on rare events simulation. Generate K as geometric. Otherwise. logarithmic efficiency is almost as good as bounded relative error. X2. and in practice. 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 M = X1 + • • • + XK. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded.i. The algorithm gives a solution to the infinite horizon problem . are i. 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.. According to the above discussion. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. 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. 3. it is not efficient for large u .p)pk. where Z = I(M > u) may be generated as follows: 1. which gives a logarithmically efficient estimator . However. see Asmussen & Rubinstein [45] and Heidelberger [190]. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. . let Z +. If M > u.X1 + + XK. O (u) = z = EZ.
+XK > uIXl... Xl > x. . < X(K) throw away the largest one X(K). For the simulation.. conditional probability.. Z(1) (u) has a smaller variance than Zl (x). Then (cf.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). and that Bo(y) = 1.286 CHAPTER X. and let Z(2)(u) = _ P (SK B0((u > u I X(l).2. Theorem IX... assume in the following that Bo(x) .b(u) = P (Xl +•••+XK>u) = EF[Xl + ..X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . and considering only the remaining ones.p/(l . SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.XK_1).... As a conditional Monte Carlo estimator .. XK1. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). So. 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. note first that To check the formula for the P(X(n) > x I X(1)..Xl .... .XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0...X(2).p)Bo(x). Thus. Z(1)(u) is defined as 0).X(2). K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.. X1 + + XK_ 1 > x when X1 > x. just note that EZ(1)(u ) 2 > E[Bo (x .SK1)2. we thus generate K and X1i . 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..... form the order statistics X(1) < X(2) < .L(x)/x`' with a > 0 and L(x) slowly varying. .XK1] = EBo(uX1 . we generate only X1. The idea of [27] is to avoid this problem by discarding the largest X.X(K1)) .. y < 0). To see this.... However..X1 .1) V)(u) . . compute Y = u . A first obvious idea is to use conditional Monte Carlo: write i. XK..
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. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. l)) BO(X(n1)) Theorem 2 . use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . and define )3L. The algorithm is sofar the only efficient one which has been developed for the heavytailed case.. the algorithm for generating Z = Z(u) is: 1. X(n1)) Bo((x . in the renewal or Markov. that is. BL by I3L = /3B[y]. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. .S(n_1) I X(1). IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). 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.. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck._1) > P(X(n) > _ X X(1). X . and simulate from FL. ..y. . Notes and references The proof of Theorem 2. using 13L. B. For practical purposes. and we refer to [27]. X(2). Asmussen . X(2).S (n1)) V X (. However . .. . 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 .modulated model P(r+ < oo) and G+ are not explicit ).1 is elementary but lengty. X(n1)) P(X(TZ) + S(. BL(dx) = e7sB(dx)/B[y].Khinchine formula and importance sampling .u is the representation 0(u) = e7sr(u) overshoot (cf.3. BL instead of 0.1) . Thus. 111. for the purpose of recording Z(u) = erysr(u). Then the algorithm given by { Z (2) (u) } is logarithmically efficient. X (. X(2). Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] .. . l)) .5)..
r(u) < oo) and FL (both measures restricted to. and avoid simulating the known part e7". If S > u. Let Sf0 CHAPTER X. 4. cf. Let FL (dx) = 'For the renewal model. We formulate this in a slightly more general random walk setting '. to deal with the infinite horizon problem ... and assume that µF < 0 and that F[y] = 1. In fact: Theorem 3. B) is not logarithmically efficient when (/3. More precisely. A > AL as in Chapter V.S+U . The estimator is then M(u) /3eQT' dB Z(u) (Ui) j=1 )3 e $Ti dB where M(u) is the number of claims leading to ruin. Let X1. M(u) = inf {n : S„ > u}. = X1 + . Generate T as being exponential with parameter ..2 The estimator (3. BL).288 2. and the change of measure F r FL corresponds to B > BL. b) # (/3L. SIMULATION METHODOLOGY 3. . We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.Q. + X.3. .7 tell that P(. so that changing the measure to FL is close to the optimal scheme for importance sampling .T. b different from . Xi = U. namely ELery£("). Let S . be i. BL could improve the variance of the estimator .d. Proof Just note that EZ(u)2 < e .F. The algorithm generalizes easily to the renewal model . There are various intuitive reasons that this should be a good algorithm.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.(u)) are asymptotically coincide on {r(u)} < oo.i.r(u) < oo) = 1. one must restrict attention to the case 4µB > 1.. with distribution F.1) (simulated with parameters ^3. let Z F e_'s.l3 and U from B.. return to 3.. It resolves the infinite horizon problem since FL(. and we have: Theorem 3. Ti. P'[y] < oo for some ry > 0. In detail . The proof is given below as a corollary to Theorem 3. Otherwise. let S.QL.. the results of IV. X2.2ryu _ z (u)2/C2. u It is tempting to ask whether choosing importance sampling parameters . The answer is no. the discussion at the end of Section 1b.
Proof The first statement is proved exactly as Theorem 3 .. .yu = G. = c'. where e' = EL Iog dFL (Xi) > 0 by the information inequality. (3.2'X1 .P = FL. The importance sampling estimator is then Z( u) = e'rSM( ). where Kl og (X) (j) 2 ) = log dFL (Xi) . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). When F # FL. . Since K1.2ryELXi. 1. is not logarithmically efficient. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . By the chain rule for RadonNikodym derivatives. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . Here ELK..+KM(u)}.d. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2...i.2 > 0. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. More generally. are i.3 The estimator (3.2) dF Theorem 3.2ryELXi)} ... Since ELM(u)/u + 1/ELXi.3.. + KM(u))} = exp {ELM(u)(E . K2.yu+elu u +oo etry' 1 > lim up C2e2. it thus follows that for 0 < e < e'/ELXi. write W(F IF) _ F(XI). F(XM(u))..2) (simulated with distribution F of the X3 has bounded relative error when .. For the second.
The easy case is y > 1/k'(y) where O(u. T) with T < oo.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". U' . then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.3'eO'x f f P (U" . Next. with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. claim size distributions B'. The results of IV. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. then /3' B' = B".i. This immediately yields / = 3". the references in Asmussen & Rubinstein [45] and Heidelberger [190]. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz.4. 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 .e. Then according to Theorem 3.'(y).3. /3'eQ'YR'( x + y) dy = .4 indicate that we can expect a major difference according to whether y < 1/r. so that one would expect the change of measure F 4 FL to produce close to optimal results.1 If y > 1/ic'('y). we write T = yu. U' . generic interarrival times T' . see e. B" and generic claim sizes U'. First by the memoryless distribution of the exponential distribution . optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. Further discussion is in Lehtonen & Nyrhinen [245]. SIMULATION METHODOLOGY u Proof of Theorem 3.B'=B". . In fact: Proposition 4.290 which completes the proof.T".T' D U" . BL) has bounded relative error. CHAPTER X.'(y) or y > 1/r.T". Consider compound Poisson risk process with intensities /3'.2. all that needs to be shown is that if U' .T" has a left exponential tail with rate /3'.g. /3".T" > x) J /3"e0 yB (x + y) dy = .T' has a left exponential tail with rate /3' and U" . 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. T". In [13]. The optimality result Theorem 3.1 is from Lehtonen & Nyrhinen [244]. yu) is close to zk(u).T' = U" . As in IV. U". we conclude by differentiation that Bo(x)=B' (x)forallx > 0. The queueing literature on related algorithms is extensive .
yu) in the sense that .8 has a stronger conclusion than (4.5) follows. 7y (4. .4. and that ryy > ry.4.1) so that z(u) = zP(u. that ryy = ay . We next consider the case y < 1/r. T( u) < yu] e2ryyuEay le 2ay^(u). IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. the result follows as in the proof of Theorem 3.1) which is all that is needed here can be showed much easier . Proof Since ryy > y.yu. one would expect that the change of measure F Pay is in some sense optimal. T(u) < yu] e Hence by (4.'(7).(u) * 1 (Theorem IV.log Var(Z(u)) _ .(ay). Remark 4 .yy> 2 . Let Qy2 = . yu) = eayu Eay Leay^(u)+r(u)K(ay). Bay) is logarithmically efficient. lim inf uoo 27yu .8).1). yu)/z.4.3) and we have: Theorem 4.1.yk(ay) determines the order of magnitude of z'(u. (4.log 4')u) 4 u (Theorem IV.4.log Var(Z(u)) l im of . We recall that ay is defined as the solution of a'(a) = 1/y. 3 Theorem IV.O(u. and in fact. Further . yu) is of order of magnitude a71.to g x ( u ) u u so that (1. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). (4. Bounding u ELZ(u)2 above by a7u.3) (simulated with parameters /gay.1).2) Since the definition of ay is equivalent to Eay r(u) .2 The estimator (4. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. T(u) < yu] .
a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . there exists a dual process { V t} such that i..yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.7y x(u) > hm inf u+Oo U . the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).1) is used to study Voo by simulating {Rt} (for example.a vt(u). Hence lira inf log ryyu + vyu 1/2 tc(ay) . N(0. zi(u) = INV.yu)/(uyu1/2) . we believe that there are examples also in risk theory where (5.o.u1/2 < r(u) < yu Le ] l = e.2). Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). However. related discussion is given in a heavy traffic limit q J. In Asmussen [13].T) = P O<t<T inf Rt < 0 = P(VT > u).Qyu1/2 < T(u) C yu e.3. > u) = E f I(VV > u) dt 0 (5.4.1) (see Proposition IV.1) where the identity for Vi(u) requires that Vt has a limit in distribution V.1) may be useful.4). and (5. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) . yu . yu .ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. In most of the simulation literature (say in queueing applications).1): then by Proposition A1.4. the object of interest is {Vt} rather than {Rt}.292 CHAPTER X.b(u. 0 Notes and references The results of the present section are new. 0 rather than when u 3 oo. (5.3: for many risk processes {Rt }.uaoo U That lim sup < follows similarly but easier as when estimating En. Z (u)2 above.2) . One main example is {Vt} being regenerative (see A. 5 Regenerative simulation Our starting point is the duality representations in 11.(av)Eav l e. '%(u) = P I info Rt < 0) = P(VV > u).o .
E). consider first the case of independent cycles . record Zi'i = (Z1').. 02) (5. which we survey below . i. EZ2'i = z2 = E Thus. Then Z(1). z2) z2/z1 yields Vh = (z2/z2 1/zl). Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Vh = (8h/8z1 8h/ 8z2). )). Z2 . .d. letting J0 'o I (Vt > u) dt . z2)) > N(O. Zl the LLN yields Z1 a$' Z(1) +.5. Z(N) are i . the regenerative estimator z%(u) is consistent... REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. + Z1N>) .. oh) for h : R2 ^ R and Ch = VhEVh. The method of regenerative simulation.. For details .. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. Therefore . Z2 a4* z2. and Z2'>) where Zi'i = w. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.. For the ith cycle. . 2.h (zl. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. is the cycle length. + Z2N)) . Taking h(zl. let E denote the 2 x 2 covariance matrix of Z(').+Z(N) z 1. Thus.. i (^(u) . Then (Z1z1i Z2z2 ) 4 N2(0. Z1 = (Zl1i +. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). EZ1'i = z1 = Ew. Z2 = N (X21' + . provides estimates for F ( V.3) ... > u) (and more general expectations Eg(V. To derive confidence intervals . j = 1..t(u)) 4 N(0.
() dx so that differentiation yields zS d( fco(x)f(x.v.0 . 6 Sensitivity analysis We return to the problem of 111 . There is potential also for combining with some variance reduction method. () depending on C. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . 9.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . However .9. the expectation z = EZ of a single r. Before going into the complications of ruin probabilities . In 111.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1.294 where 01 2 CHAPTER X.g S12 (5. Notes and references The literature on regenerative simulation is extensive.Z) ^Z(=) . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . () dx f Ax) (dl d()f (x' () f ( z.2. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . Rubinstein [310] and Rubinstein & Melamed [311].96s/v"N.C)dx = f w(x) d( f ( x. consider an extremely simple example . SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . () dx = E[SZ] f(X. Let X have a density f (x. asymptotic estimates were derived using the renewal equation for z /i(u). v. The regenerative method is not likely to be efficient for large u but rather a brute force one.g. Z of the form Z = ^p(X) where X is a r . see e. with distribution depending on a parameter (.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . in some situations it may be the only one resolving the infinite horizon problem . Then z(() = f cp(x) f (x. We here consider simulation algorithms which have the potential of applying to substantially more complex situations.
() where U is uniform(0. C). Then z(() = Ecp(h(U. Then . for some Co = (o(U). So assume that a r. The following example demonstrates how the SF method handles this situation. with density f (x. () = . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. For the SF method. ( where h( (u. () is an unbiased Monte Carlo estimator of zS. ()). () is increasing in C. IPA will estimate zS by 0 which is obviously not correct. if f (x. one. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. zc = E [d co(h(U. this is usually unproblematic and involves some application of dominated convergence . SZ is an unbiased Monte Carlo estimator of z(. the Poisson rate /3 in the compound Poisson model. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. () = d log f (X.log U/(. () = log U/(2. however . this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators .v. ()) is 0 w . For IPA there are. just take cp as an indicator function . ()) h((U. Example 6 .r. C) f(X. Let M(u) be the number of claims up to the time r(u) of ruin (thus. () Thus. In the setting of ruin probabilities . The likelihood ratio up to r(u) for two Poisson processes with rates /3. () can be generated as h(U. () = (8/8()h (u. Thus . C)). one can take h (U. = E [`d (h(U. To see this. giving h( (U. Thus. ()) d( hc(U.1). For example . /3o is M(u) Oe (3T: < oo) . 11 /3oeOoT.1 Consider the sensitivity tka(u) w.t.6. cp(h(U. cp' (h(U.() d( is the score function familiar from statistics .r.t. I(r(u) . () _ (eSx. /3 is 0. say W(x) = I(x > xo) and assume that h(U. p. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. r(u) = Tl + • • • +TM(u)).
3 (u) is of the order of magnitude ue7u.3 (u) (to generate Zp (u). Thus. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . BL). SIMULATION METHODOLOGY Taking expectation. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). for different models and for the sensitivities w. j3 and letting flo = 0.3L. a relevant reference is VazquezAbad [374].1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. whereas for the SF method we refer to Rubinstein & Shapiro [312]. 4) that V5. In the setting of ruin probabilities. differentiating w. different parameters.296 CHAPTER X.t. To resolve the infinite horizon problem . change the measure to FL as when simulating tp(u).r. in part for different measures of risk than ruin probabilities.r.t. Example 6.9 . ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .T(u)) I(T(u) < co) ] . However. We then arrive at the estimator ZZ(u) = (M(u) . the risk process should be simulated with parameters .T(u)) e7uerVu) for ?P. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . .0(1) so that in fact the estimator Zf(u) has bounded relative error. There have been much work on resolving the difficulties associated with IPA pointed out above. We recall (Proposition 111.
(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u... T(u. defined as Ro = u (with u E {0. X2.1}valued . '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 ). either this makes no difference (P(R. 297 . R„ = u+X.i. are i. with P(Xk = 1) = 9.. }). That is. a) = r(u) A T+(a). .. Consider first a Bernoulli random walk.. as e. where X1. Y'a(U) = P(T (u) = r+(a)) = 1 . a) = r(u)).d. and {1. Besides its intrinsic interest ...(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo..+• • •+X.1..g.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.P(•r(u. wherel T(u) = inf {t > 0 : Rt < 0} . in most cases . in the Bernoulli random walk example below. The twobarrier ruin problem The twobarrier ruin probability 0. T+(a) = inf It > 0 : Rt > al..
1.e.1) = (19)4/'0(a3)+9ba(a1). u + 1.o)T/la (1) + 8z/'u(3). Wald's exponential martingale is defined as in 11. one elementary but difficult to generalize to other models.1) is solution.a) = 0) + zap ( R.0)/0.r(u. C1_0\a.(u) I\ e = 1 oa ' ()i a = u. and in view of the discrete nature of a Bernoulli random walk we write z = e7.2). We choose a = ry where ry is the Lundberg exponent.. z and the solution is z = (1 .. u Proof 2. = (1 .a(u)). then 'Oa(u) _ au a We give two proofs . and insertion shows that ( 1. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).o» = z°P (RT ( u. By optional stopping. zu = EzRO = EzRT(u.(1B)u oJ 0.. Proof 1.298 CHAPTER XI. the solution of F[.1) o If 0 = 1/ 2.. MISCELLANEOUS TOPICS Proposition 1.4) by ea(u+Xl+.1 For a Bernoulli random walk with 0 0 1/2..2) Oa(a .y] = 1. In a general random walk setting . where a is any number such that Ee°X = F[a] <oo. tba(2) _ (1 . 7/la(a . i.o)'t/1a(a ..a) Y.+Xn) F[ a]n n=0.. and the other more advanced but applicable also in some other settings. = z°Va(u) + za(1  .. (1. The martingale is then {zuzXl+•••+X„ } = {zR° }. The Lundberg equation becomes 1=F[ry]=(19)+9z..(4.
1) for p # 0.1). pa( u) _ u Corollary 1.e7u)/(e7° . thenz1 (u)=1.ba(u) = e2µa . RANDOM WALK.• a2µa e2µu . Corollary 1. (1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. Proof Let a+ oo in (1. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance .1).1 If p = 0. then Vi(u) = 1. Applying optional stopping to the exponential martingale {e7R. If p = 0.4 For a Brownian motion with drift u > 0. Then for p 0 0.2) is trivial (z = 1).} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)). u Proposition 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 . If p<0.2 For a Bernoulli random walk with 9 > 1/2. However. If 9 = 1/2. . {R.0a(u)). (1.u)/u. 1h (u) = a el u \1 If 9 < 1/ 2.1 yields 't/la(u) = (a . and solving for 9/la(u) yields Z/)a(u) = (e 76 .1.zu)/(za . i1(u) = e211 . then Proof Since 'Oa (U)  au a Eea(R°. BROWNIAN MOTION.5) .
however. . (1.0 (u) (where u p =.. a) I R(u a ) < 0] P (R(u . 5).300 Proof Let a * oo in (1.e7a Again . letting a * oo yields the standard expression pe7u for .+^a(u))^(a) If 7k(a) < 1. the paths are upwards skipfree but not downwards.a) = a on {r (u. valid if p < 1 (otherwise .616).a) < 0) + e7°P (R(u. 1 . (u) _ O(u) . Ic 5ry 'pa(u) Using y = 6 . and hence e7u = Ee7Ro E [e7R(.a) = a) = 5 y = P (R (u. we obtain 'Oa a7u . Here is one more case where this is feasible: Example 1. implying R(u.4).5a). CHAPTER XI.a) = r+ (a)} and similarly for the boundary 0.e7a (u) = 6 /0 .0(a) 0 < u < a.7/la(u)). this immediately yields (1. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. 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) = a ) + e ' ° ( 1 .5 Consider the compound Poisson model with exponential claims (with rate. 0.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. VIII.a ) < 0) + e 7aF ( R (u. 7/'(u) = 1).7). For most standard risk processes . Here the undershoot under 0 is exponential with rate 5. and thus one encounters the problem of controlling the undershoot under level 0. passing to even more general cases the method quickly becomes unfeasible (see.vi(a) Proof By the upwards skipfree property. say. 7O(u) = 7/la(u) + (1 .3. However.7) .
T) P(MT > u) where MT = maxo<t<T St. (1. 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 .9) = 2P(ST > u).ST>U). Hence P(MT>u. RANDOM WALK.. T(u) E dT.11) VIT ) Proof For p = 0. = 1 .8 )..r(u). TWO BARRIERS 301 Note thas this argument has already been used in VII.8) Proof In terms of the claim surplus process { St} = {u .1. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .10) Pµ (T(u) < T) !. the density dPµ / dP0 of St is eµsttµ2/2.. For µ # 0. MT > U) = P(ST > u) + P(ST > u) (1. Then the density and c.. + µ2T) } .1a for computing ruin probabilities for a twostep premium function.7 For Brownian motion with drift 0.f. 10) follows then by straightforward differentiation. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .. we have ili(u.)_ _( u)µ2 /2. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. and hence Pµ('r(u) E dT) = Eo [e µsr(. 0(u.2 .d. P(MT > u) = P(ST > u) + P(ST < u. Here {St } is Brownian motion with drift 0 (starting from 0).µ so that {St} is Brownian motion with drift µ . (i). MT > u) = P (ST > u) + P (ST > u. BROWNIAN MOTION. Corollary 1. ( 1.Rt}.ST<u) = P(MT>u.µ%T (1.11 ) is the same as (1.4) I = .µ T I + e2µ"4) ( . and (1 . T ) = P(T(u) < T ) = 241.8 Let {Rt} be Brownian motion with drift . For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. (1.
12) is the same as ( 1.g. oo) with drift µ(x) and variance a2 (x) at x. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. but we omit the details. Breiman [78] or Karlin & Taylor [222] p. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .9). oo). Thus.9 For Bernoulli random walk with 9 = 1/2.9) goes through unchanged. (1. as defined in (1.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.s. T are integervalued and nonnegative. whenever u. the behaviour at the boundary 0 is more complicated and it may happen. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. is finite for all x > 0.10). e.. Vi(u.. see e..8 also applies to the case 9 54 1/2. S(oo) = f c s(y)dy. as defined above as the probability of actually hitting 0. Proof The argument leading to ( 1. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions.10 Consider a diffusion process {Rt} on [0. u Small modifications also apply to Bernoulli random walks: Proposition 1. If this assumption fails.T+2. and (1. Let s(y) = ef0 ry(.T (1.12) P(ST = v) = 0 otherwise. The same argument as used for Corollary 1.11) then follows by checking that the derivative of the r. and in a similar spirit as in VII.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). Here {2T( (v}TT)/2) v=T.h.302 CHAPTER XI.. is (1. 226).T) = P(ST = u) + 2P (ST > u). that 0(u).10) and that the value at 0 is 0.g. We finally consider a general diffusion {Rt} on [0.T)dx. MISCELLANEOUS TOPICS which is the same as (1. Theorem 1. 0 0 (1. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). S(x) = f x s(y)dy.T2.13) with 0 as lower limit of integration.
14) S(oo) < 00. we can ignore the possibility of ruin or hitting the upper barrier a before dt. For generalizations of Proposition 1. Lemma 1.16). Then YIa. A classical reference for further aspects of Bernoulli random walks is Feller [142]. b = 0. Using s'/ s = 2p/a2. and we get Wo.b(Rdt). [117]. A good introduction to diffusions is in Karlin & Taylor [222].1.e. (1 . E„ q(Rdt) = q(u)+Lq(u)dt. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.11 Let 0 < b < u < a and let t&0. the function S(x) is .b(u)dt.ba. Letting a T oo and considering the cases S(oo) = oo. TWO BARRIERS 303 for x > 0. In view of (1.17) Hence L. Letting b J. (1. Notes and references All material of the present section is standard. S(oo) < oo separately u completes the proof.10. 15) i.14) fails. RANDOM WALK.b(u) be the probability that {Rt} hits b before a starting from u.S(b) Proof Recall that under mild conditions on q.b('u) = Eu &0. 191195 for material related to Theorem 1. see Asmussen & Perry [42]. .16) yields 4b (u) = 1 .e LVa.S(u) (1. BROWNIAN MOTION.b('u) = Eu . If b < u < a.S(u)/S(a). Assume further that S (x) as defined in (1.6 to Markovmodulated models . so that Y)n.10.b(u) + L. O. If (1.b(b) = 1. then 0 < 2l.b = 0 implies that VQ b/s is constant.b (Rdt) = Oa. Further references on twobarrier ruin problems include Dickson & Gray [116]. i.13) is finite for all x > 0. see in particular pp.0(u) = 1 for all u > 0.16) S(a) . The obvious boundary conditions '0a.b = a+/3S. 0 Proof of Theorem 1. if (1. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . 0 in (1. Wa.b(a) = 0 then yield the result. then.b(u) = S(a) . 1'.ba.(u) < 1 for all u > 0 and ^ S^ Conversely.
(2. 1 y < k (y). Lo I. defined by the density 1/va(u)s(u) showing up in (1.(a) (2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. Remark 11.9 ) and optional stopping applied to the stopping time r(u) A T.(7) . y > .3. variance 0.3) < e 7yu.6.1 ) was given already in II. information on ruin probabilities can be obtained . IV. (2. Markovmodulated Brownian models . one works instead with a lower limit 5 > 0 of integration in (1.17).2) C_e7u < t(u) < C+e _7u. (2. The emphasis is often on stationary distributions . much of the literature dels with the pure drift case. Lo is a martingale (cf.(.1) (2.4. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.5) A martingale proof of (2. yu) where W (ay) = y. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.6) .(T(u)AT) r.304 CHAPTER XI.13)). but by duality. Another basic quantity is the speed measure M .5. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. with the drift and the variance depending on an underlying Markov process . equivalently.1. correponding to piecewise linear paths or . 7y = ay .ytc (ay). yielding eau = Ee. and here are alternative martingale proofs of the rest .aR.t&(u..4) I. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.4.2. is currently an extremely active area of research.aRo . (2.5): _ z/'(u) < e 7u. yu) '+/1(u) .)AT .o•K(a) = Ee . 111 .
1. (B(y) . . Proof of ( 2.B(r))/B(r).(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).2.yu))• b(u. FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem).3).6) with = 'y that eyu .T)  V. (2. Rt has distribution B(r + dy)/B(r). so that i/1(uL yu) < eayu .f. u Proof of (2.(u.2): As noted in Proposition II.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.4): We take a = ay in (2.7R.( u ) I T(U) < 00] .3).1. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.E [e. Let H(dt.6). y > r. A claim leading to ruin at time t has c.yu) Y Similarly for (2. eyuk (ay) = e7yu e > eyu"(ay ) ij(u. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt. we have tc(ay) > 0 and we can bound (2. it follows easily from (2.1 .d. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.)r(u)r. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.6) below by 1 E Le7Rr(. when Rt_ = r.yuk (ay)(u&(u. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.. and the proof of the lower inequality is similar.4). For (2. Equivalently. RT(u)_) given r(u) < oo. dr) denote the conditional distribution of (T(u). we have ic(ay ) < 0 and use the lower bound E [e7Rr („).
not quite so much in insurance risk. large deviations results been. if x > EX1.^ e nn 1 > x n 0o 2xn (3. MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results. For example. + X. logarithmic asymptotics .g. its generality. and gave sharp asymptotics for probabilities of the form P (S. . gn 4 0. we will write fn 1. However ..1) where we return to the values of 0.306 CHAPTER XI.gn if nioo lim 109 fn = 1 log gn (later in this section. Thus. The advantage of the large deviations approach is./n E I) for intervals I C R.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.1) is an example of sharp asymptotics : . the parameter will be u rather than n).. nroo n n /// Note in particular that (3. Cramer considered a random walk Sn = X1 + .1). which in the setting of (3. . v2 later..nn or C2e. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. The classical result in the area is Cramer's theorem. (3.. The limit result (3.the correct sharp asymptotics might as well have +.1) does not capture the \ in (3. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . however . then P C S. gn with fn + 0 . e.3na with a < 1. 1) but only the dominant exponential term . and that a considerable body of theory has been developed.means (as at other places in the book) that the ratio is one in the limit (here n * oo). ri. cle .2) can be rewritten as F (Sn/n > x) 1g a'fin.1 We will go into some more detail concerning (3. Example 3. Accordingly.2).?n typically only give the dominant term in an asymptotic expression . (3.(B) = log EeOX 1 is defined for sufficiently many 0. such that the cumulant generating function r. large deviations results have usually a weaker form. in being capable of treating many models beyond simple random walks which are not easily treated by other models . Thus . og For sequences fn.
3) is put equal to x.4 enn +1.1). V > 0 e.q = rc* (x). 2 where o2 = o2(x) = rc"(0). LARGE DEVIATIONS Define rc* as the convex conjugate of rc. since Sn is asymptotically normal w. Most often. we get P(Sn/n > x) E [e9nx +nK(9)9" '. rc*(x) = sup(Ox . replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.(0)) e 307 (other names are the entropy. Define . and hence for large n P(Sn/n > x) > E [e.rc(0) where 0 = 0(x) is the solution of x = rc'(0). (3.e.r.9S„+n' ( 9).t. In fact. i.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . S rtn > x 1. nx < Sn < nx + 1. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .r.4) immediately yields (3.(e)i XI E dx]. More precisely.960/) * 0. which is a saddlepoint equation . of P(X1 E dx) = E[e9X1K.425.3.96o /] > 0. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). exponential change of measure is a key tool in large deviations methods. Since P nn > x) = E {e_8 ' ( 9). P with mean nx and variance no.4) n Next.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.sseo f which in conjunction with (3. if we replace Sn by nx + o / V where V is N(0. we have P(nx < Sn < nx + 1.2).
(iv) tc(ry) = 0 and r.e < 8 < y + e.dxn) = 05nKn(7)Fn(dx1. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .1)... however.v.. X2. see Jensen u [215] or [APQ] p.3 For each i > 0. Further main results in large deviations theory are the GartnerEllis theorem... asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0./^ >7 < zn n for n n0.. 1) and no such that Sn . integrates to 1 by the definition of Icn). Pn Sn1 .'s. and write Sn = X1 + • • • + Xn. and the WentzellFreidlin theory of slow Markov walks. Mogulskii's theorem which gives path asymptotics. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.. we introduce a change of measure for X1. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). there exists z E (0. Assume that there exists 'y. We shall need: Lemma 3 . . .308 CHAPTER XI.2 (GLYNN & WHITT [163]) Let X1. Xn) and sn = x1 + • • • + xn (note that the r.. 1].. MISCELLANEOUS TOPICS which is the same as (3.dxn) where Fn is the distribution of (X1i . 260 for details. (ii) lim supn. We further write µ = tc'(ry). is differentiable at ry with 0 < K'(y) < 00. . commonly denoted as is the saddlepoint approximation.p > 7 < zn. n Icn(0) exists and is finite for ry . which is of similar spirit as the dicussion in VII...'(u) )Ng a"u. The substitution by V needs.. .s. (iii) #c (8) = limn. Ee9X n < oo for e < 0 < e.. be a sequence of r. Xn given by Fn(dxl. For the proof.o log Ee9Sn /n. Sanov's theorem which give rare events asymptotics for empirical distributions. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 ...h. to be made rigorous. . In the application of large deviations to ruin probabilities.e < 8 < y + e.. that is. Then i/.3..
y) .ne(p+ 17).+r7) < zn for n > no.m(7). for Sn. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). Since I EeqOX „ ] 1/q is bounded for large n by (ii). can be chosen strictly negative by taking 9 small enough.]1/q = e.n > u ) = [ Em [em Em 1e. 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. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. S. the r .2.077 n^oo n and by Taylor expansion and (iv ).n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.2.71 < e and jq9j < e... mµ Sm > u] km e7Sm+n. it is easy to see that the r. 0.r (7) n = e. Clearly. > 1 +17] m(7).n m µ 1 + rl .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 +.91) + o(O ) as 0 J.> . h. This establishes the first claim of the lemma .YS.+r.s. The corresponding claim for Pn(Sn/n < µ .77) follows by symmetry (note that the argument did not use µ > 0). is of order .Bµ .W. We first show that lim inf„_. ( U) P(S. S.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . The rest of the argument is as before. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.. u Proof of Theorem 3. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.s.h. h. in particular the r.. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.3.. Then V. 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. log zl'(u)/u > 'y.µ?7 .s.
this is possible by (iii). n=1 . Obviously. logO(u)/u > ry..(•) goes to 1 by Lemma 3. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.I < µl1 1+77 I M 1_ 1+277 S.+wn(7). I2 = F(T(u) = n).log z) /2 and Sn Fn\ n >lb+S) <Zn. MISCELLANEOUS TOPICS (7).310 ]Em I e..3.n(ry)/u 4 0andm/u* (1 + r7)/µ.YS +^c CHAPTER XI. (iv) and Lemma 3. For lim supu... 0 yields liminfu __. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. I > IL exp `S.. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).. P(T(u) = n) < P(Sn > u) = En [e7S.0 log i'(u )/u < 'y.6) for some z < 1 and all n > n(E). we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n).7) so that n(b) I1 < e'Yu E en. and since Ic.(Y).n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. 3. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. Pn \ > la+ 8 I < zn (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 .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.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).10) 00 I4 < E F(Sn_1 < u. C 26u `p / +1 I e6u(1+6)/µ (3. Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. Sn1 C U. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .3. u . S.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. eryu en logz/2p n nt n. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.
it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 . we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). I2.b)/i(7).('+'Y). ryue«iu . u . Then for n large. Corollary 3.u(1+b)/rc'(7)). > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. 2. Letting c11 = maxn<n.4. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. IV.xl/2 to give the desired conclusion.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . For I. we need to redefine n(b) as L.Q is so small that w = 1 . the typical time is u/rc'(7) just as for the compound Poisson model.3ui where . this is straightforward since the last inequality in (3. (7 + a) < 2arc'(7). 13 = P(T (u) E (u(1 b)l^ (7). For 14.8) by P(S. the last steps of (3. u(1 + b)/i(7)) Proof Since V.. e'.4/3rc'(y) > 0.z 1/z For I1. it suffices to show that for j = 1.. 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.7' a"ju.4 Under the assumptions of Theorem 3. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . cf.312 CHAPTER XI. For 12. we replace the bound P(Sn > u ) < 1 used in (3.(u) = I1+I2+I3+I4'^ ery( u). say n n1.2.
t] is Rt = E V (Un) n: o„ <t .5 Assume the Xn form a stationary Gaussian sequence with mean p < 0.f. criteria are given in Duffield & O'Connell [124].. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .1. Assuming that the further regularity conditions can be verified. and we conclude that Theorem 3 .. but nevertheless. the key condition similar to (iii).v. V(s) with m. Xk+l) k=1 00 naoo n provided the sum converges absolutely. If {St}t> 0 is the claims surplus process.12) k=0.13) One would expect this to hold in considerable generality. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.3(s) at time s. Obviously many of the most interesting examples have a continuous time scale. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.. 09(9)... 11 Inspection of the proof of Theorem 3.e. Theorem 3.(O) = 9µ+02 for all 9 E R. i. Thus the total reward in the interval [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. 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.3..1. An event occuring at time s is rewarded by a r.2 shows that the discrete time structure is used in an essential way. whether P ( sup St > u ltg a ^" 0<t<oo // (3.LARGE DEVIATIONS 313 Example 3 . we shall give two continuous time examples and tacitly assume that this can be done.. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. and in fact.'(y) > 0. 2 is in force with y = 2p/wz.14) is needed in both examples .g. r. To verify these in concrete examples may well present considerable difficulties.
(3. we conclude that Cu) log e7 u (cf. Thus by (3.Q„) . derive .'`1 U. <t which is a shotnoise process. we have S.. At = . 0 Example 3 . the best estimator of /3µB based upon Ft. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .14) (to see this . one would take p(t) = (1 + rt)At/ t. if the nth claim arrives at time a. but that a claim is not settled immediately. nondecreasing and with finite limits Un as s T oo ( thus. 7 Given the safety loading 77.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . a differential equation in t). (9) < oo for 9 < 'y + C.g. assuming a continuous premium inflow at unit rate. where Ft = a(A8 : 0 < s < t).14). it contributes to St by the amount Un(t . 0 and since EeOUn(8) + Ee°U^ as s * oo.1) ds rt (3..9t = /3 J t (Ee8U° i8l . this is not realistic . Example 3. Un represents the total payment for the nth claim). the above discussion of discrete skeletons).s).15) . Most obviously.9t. 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) ds . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . the Cramer.1) ds .2 are trivial to verify. MISCELLANEOUS TOPICS are the event times. Of course .t.1) . We further assume that the processes {U1(s)}8>0 are i..Lundberg model has the larger ruin probability. More precisely. we have rct (9)/t 4 ic (9). then the payments from the company in [on.0 and assume there are y.v. If the nth claim arrives at time Qn = s.314 where the an CHAPTER XI.It. n: o. O'n +S] is a r . e.d. Un(s). i.. . We let ic (9) = 3(EeWU° . is At . e > 0 such that ic('y) = 0 and that r.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. Kt (0) t (Ee9U"it8i J0 . = U„ ( t . Of course. Thus. Thus. leading to St = At(1+77) Joo t S8 ds. Then logEeOR° = J0 /3(s)(^8(9) . Since the remaining conditions of Theorem 3.
and since the remaining conditions are trivial to verify.(1 +i) f > i= 1 s ds = E Ui 1 . i./3. Thus. (3.e. we conclude that t. the solution of /3(Eelu .log Oi are i.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.b(u) IN a'Yu (cf.d. the Vi = .1) . rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (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.16) i=1 o i=1 Let ict (a) = log Eeast .(1 + 17)0µB = 0. Indeed. LARGE DEVIATIONS With the Qi the arrival times. (3.20) (3. uniform (0.i.3.i. one has y > y' (3.21) This follows from the probabilistic interpretation Si EN '1 Yi where Yi = Ui( 1+(1 +r7)log ©i) = Ui(1(1 +17)Vi) where the Oi are i . Ui Nt / t 01i 315 St = Ui . 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 . equivalently. standard exponential .(1 + r7) log t (3. we have Nt t N.2 hold.1) or . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .18) Thus (iii) of Theorem 3. It then follows from (3. To see this .19) with equality if and only if U is degenerate.d.
see also Nyrhinen [275] for Theorem 3. Further. we then take t = 1 so that p.316 CHAPTER XI. with common distribution B and independent of Nt. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. this in turn yields y > y*. a* (s) are convex with tc'(0) < 0 . and k(x) < 0. 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.2 expressing the finite horizon ruin probabilities in terms of the distribution of A.1 .(1 + ri)y*x is convex with k(oo) = 00.1 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. are i. For notational simplicity. The main example is Nt being Poisson with rate fit. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. Therefore e7'U _ k(U) E [1+(1+77)y*U] . Further applications of large deviations idea in risk theory occur in Djehiche [122]. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. k(0) = 0. we are interested in estimating P(A > x) for large x. [257] and Nyrhinen [275].19). the function k(x) = e7*x . k'(0) < 0.i.2. the study is motivated from the formulas in IV.xo. and since tc(s). [245]. MartinL6f [256]. In addition to Glynn & Whitt [163].20) is due to Tatyana Turova. though we do not always spell this out. x > x0. say one year. For Example 3. In particular. at time t. see Nyrhinen [275] and Asmussen [25]. much of the analysis carries over to more general cases. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. Further. using that Ek(U) = 0 because of (3. This implies n(y*) < 0. Lehtonen & Nyrhinen [244]. y = y* can only occur if U . . 0 < x < x0.7.d. MISCELLANEOUS TOPICS Next. rc*' (0 ) < 0. = P(N = n) = e(3an However.. the proof of (3. assuming that the U.
For a given x. Proposition 4.ic(9) = . K'(0) _ ic'(9) = x. Hence P(A > x) = E e [e9A+ ic(9).9(Ax). only with 0 replaced by a9 and B by B9.1). (4. A E dx] . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. The analysis largely follows Example 3.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). Then Ee"A = e'(") where x(a) _ 0(B[a] . 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 .1.4. B"' [s] lim (B".e.2) implies that the limiting Pedistribution of (A . B"[s] = oo. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). A > x)] = eex+K( e)E9 [e . no(a) = logE9e'A = rc(a + 9) . 818' where s' = sup{s : B[s] < oo}.[s])3/2 = 0.1).3B"[9].x)//3B"[9] is standard normal.1) where )30 = . This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. In particular. Vare(A) = s.1 Assume that lim8T8. we define the saddlepoint 9 = 9(x) by EBA = x. i. Then as x * oo.3e(bo[a] ."(0) = .
i.1). some regularity of the density b(x) of B is required.2) is often referred to as the Esscher approximation. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. just the same dominated convergence argument as in the proof of Theorem 2. more generally. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. leading to P(A > x) :. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. Remark 4 .(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. 4b The NP approximation In many cases . under the Poisson assumption (4.1 yields: Proposition 4. 2). either of the following is sufficient: A.(D X . In fact. For example.4) .1 goes all the way back to Esscher [141]. B covers distributions with finite support or with a density not too far from ax° with a > 1. The present proof is somewhat heuristical in the CLT steps.l3pB. or.2i and that (A . For example.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. 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 . b(x) = q(x)eh(z).x') where x' = sup {x : b(x) > 0}. In particular.e. then P(A > x) . and (4. Thus . [138]. b is logconcave. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4.318 CHAPTER XI. Notes and references Proposition 4. bounded with b(x) . the distribution of A is approximately normal . 1 . it holds that EA = .3) and related results u for the case of main interest . large x. Jensen [215] and references therein.ycix °ie6x B. it is quite questionable to use (4. For a rigorous proof. Furthermore 00 b(x)Sdx < oo for some ( E (1.v. A covers the exponential distribution and phasetype distributions.Q{AB (4. Var(A) _ ^3p.2 If B is subexponential and EzN < oo for some z > 1. see Embrechts et al. b is gammalike. For details.EN B(x).
are the cumulants . f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . defined as the the solution of P(A < yle) = 1 . then P(Y < y) 4(y) . Remark 4. K3 = E(Y . . . In concrete examples .1). so that 1(u) 3 exp { . the density of Y is 1 °° _ eiuy f(u) du 2x _. u5. are small. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter.c2i.5) follows by integration. A particular case is a. K2 = Var (Y). the NP (normal power) approximation deals with the quantile al_E.3!).2X2 ..h.e.l = EY. one needs to show that 163..EY)3. (4. as u2 u3 u4 9(u) = Ee'uY = exp {iuci .6) .4.5) is obtained by noting that by Fourier inversion. where Kl .6(1 . s.equantile in the distribution of Y.. If this holds ..2 2 .EA)/ Var(A) and let yl_E.5 (y3 . which is often denoted VaR (the Value at Risk). resp. and so as a first approximation we obtain a1_E = EA + yle Var(A) .99. Let Y = (A . Var(Y) = 1 as above . K4 . ylE should be close to zl_E (cf.3& (y).. Thus if EY = 0..i 3 K3 } Pt^ exp ..y2)^P(y)• 319 Note as a further warning that the r. If the distribution of Y is close to N(0.: EA + zl_E Var(A) . . of (4.5).s. the standard normal distribution. and from this (4. however. zl_e be the 1 . Heuristically..f. in particular. the CLT for Y = Y6 is usually derived via expanding the ch. Rather than with the tail probabilities F(A > x). (4.2 ^ \1 .2K3 + 4i 64 + .i 6 r 1 3 so that we should take b = ic3/6 in (4.. one expects the u3 term to dominate the terms of order u4.5) may be negative and is not necessarily an increasing function of y for jyj large.
EA ) / Var(A). k3 is small for large /3 but dominates 1c4. let pn Pn = (a+ = P(N = n). Another main reference is Daykin et at.5) by noting that the 4. For example.. We can rewrite (4. that [101] distinguishes between the NP and Edgeworth approximations..yi. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . the kth cumulant of A is /3PBk' and so s. [101].1) E (A . This leads to t( yl E) .S(1 . n = 1.E .E)A1 l E) 1 E 4)(yl E) ^' . MISCELLANEOUS TOPICS A correction term may be computed from (4. and assume that there exist n ) Pn_i .5(1 . K5 .1).E )Azl E) 4(z1E) + ( ylE .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.. as required . this holds with a = 0.E(/3PB^1 )1^2 + s(z1E .6pBki) d/2.k = /3µB^1 / (. In particular . ...320 CHAPTER XI.zlE)W(zlE) 1 .3ni /3 .7) as 1 (3) a1E = Qµa +z1 .E = z1E + S(zi_E .1)! n ^eQ .(y) terms dominate the S(1 . 4c Panjer 's recursion Consider A = constants a.zi.zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .1)EY3.y2)cp( y) term. Using Y = (A . Note.E + (yl.S(1 . however.zl E)V(zl_E) .6 (1 .. 21 . this yields the NP approximation 6(Z1 _E .zl E)^o(zl E) .1)^ 2) µ'E Notes and references We have followed largely Sundt [354].. b such that EN 1 U%.
. then j (a + b!) 1ag k_1 3 gkfj.4.} and write gj = 2 . .14) is therefore a + b/n. n = k=n1 9k(n1 )9j k • (4. n. .k . By symmetry. . The expression for fo is obvious. fj = E (a+ b k =1 )9kfi_k . j1 g.10) f o = po...12). j = 1. . .11) Remark 4. (4. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. . Hence by (4..4 Assume that B is concentrated on {0. u Proof of Proposition 4.14) is independent of i = 1.. E[a +bU=I >Ui =j l i=1 J (4. if go = 0.. 1.13) Namely. 2.13) but only O(j2) for Proposition 4... 2. j = 0. .1. (4. (4. and calculating the gj*n recursively by 9*1 = 9j. (4.5 The crux of Proposition 4.4. j = 1.12) where g*n is the nth convolution power of g.4 is that the algorithm is much faster than the naive method. 2... .12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 .4.. Since the sum over i is na + b. Then fo = >20 9onpn and fi = 1 E In particular. the value of (4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.9). the complexity (number of arithmetic operations required) is O(j3) for (4.. fj = P(A = j)..
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
then (6..6. ' ii (6.h * (X) . Local reinsurance with h(u) = (u .4) and u g(x) = e7x in Ohlin's lemma. Eexp 7 [E [Ui . that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .5) holds trivially. expectations like those in (6.P I = EC [7]N.3).h(Ui)P JJJ l:='l {ry ] or. ry* > 0 because of (6.6) u where C[ry] = Ee'r(u4(u)). and so on. as often local as global. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .i. Proof As in the proof of Proposition 6.d. this implies 7* > 7.h(u) and any P satisfying (6.h( UU) = EN • E[U .4).h(U)].5) reduce quite a lot.h(U) (as in the proof of Proposition 6. u But since ry > 0.4).6).4.d. N E X . This follows by taking Xl = U A b.b)+ = Eh(U) (and the same P) satisfies 71 > ry. appealing to (6.6 Assume the Ui are i. it suffices to show that Eexp {ry ii 'UiAb.h(Ui)] .5 Because of the independence assumptions . we get EX = EN • EU. the excess ofloss rule hl (u) = (u . however. i.h(Ui)] .P.4).b)+ with b determined by E(U . Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. Then for any local retention function u . we get N 1 = Eexp ry E[Ui ii .h(Ui)] . The arrangement used in practice is. Remark 6. X2 = U .P } < 1 = Eexp E[Ui.P > EexP{7[X . y = Ei [Ui . .P]}. (6.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. (6. Assuming for simplicity that the Ui are i.
. Heilman [191] and Sundt [354].g.330 CHAPTER XI. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. The original reference for Ohlin's lemma is Ohlin [277]. [76]. see also Sundt [354].many texts on insurance mathematics. The present proof is from van Dawen [99]. Bowers et at. See further Hesselager [194] and Dickson & Waters [120]. e.
T„_1). .. . If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Then Blackwell 's renewal theorem holds. note in particular that U({0}) = 1..e.. are independent and Y1. Y. Y2.} for any h > 0. t +a]). . not concentrated on {h. The renewal theorem asserts that U(dt) is close to dt/µ... The number max k : Tk_j < t of renewals in [0. t 00 (A.. then Stone 's decomposition holds : U = U. the renewal process is called zerodelayed. the distribution of Yo is called the delay distribution. Y1. t]) so that U(t + a) . The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. t] is denoted by Nt. some condition is needed: that F is nonlattice.r. Lebesgue measure dt normalized by the mean to of F. Y2. The mathematical representation is either the ordered set 0 < To < T1 < . i.. .Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. denoted by F in the following and referred to as the interarrival distribution. Lebesgue measure for some n > 1). + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . all have the same distribution. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. = T„ .. of epochs or the set Y1. 2h. If Yo = 0. stating that U(t+a)U (t) ^ a.. when t is large.t. The point process is called a renewal process if Yo. of interarrival times and the time Yo = To of the first arrival (that is.. Technically. That is.U(t) is the expected number of renewals in (t..1) (here U(t) = U([0.
. in convolution notation Z = z + F * Z. and F(dx) a known probability measure . z(x) = 0. Then Z(u) 4 z(oo).i".9.4) that z is Lebesgue integrable with limZ.2) has the unique solution Z = U * z.2) Z(u) = J0 u z(x)U(dx).R.out. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. Under weak regularity conditions (see [APQJ Ch. z(u) a known function. Equivalently.i.2 Assume that Z solves the renewal equation (A.3) Further.e. oo).5) 2This condition can be weakened considerably . u u PF 4 00. Both result are valid for delayed renewal processes. µF (A. U Z(u . then Z(u) i f0 z(x)dx . and that F has a bounded density2. IV). IV). then it suffices for (A.a.2). wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. i. but suffices for the present purposes . Note in particular that F is spreadout if F has a density f. (A. (A. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . (A.332 APPENDIX u(t) has limit 1/µ as t 4 oo.EN(t) . In 111.4) If F is spread.x)F(dx).1 if F is nonlattice and z (u) is directly Riemann integrable (d. stating that U(t)/t > 1/p. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. that z(u) has a limit z(oo) (say) as u 4 oo. the statements being EN(t + a) . see [APQ] Ch. resp. (A.
.. 1c Regenerative processes Let {T. Assuming that y can be chosen such that f °° Ox F(dx) = 1. Hence by dominated convergence.t. of Yo. a basic reason that renewal theory is relevant is the renewal equation II. • . results from the case fo F(dx) = 1 can then be used to study Z and thereby Z... or many queueing processes. The simplest case is when {Xt} has i.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. that F is a probability measure.. and its distribution does not depend on k.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. Yk ).t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO).(3. To this end. The kth cycle is defined as {XTk+t}o<t<Yk . Y2. equivalently.k+t }t>o is independent of To.5a.. Tk (or. the present more general definition is needed to deal with say Harris recurrent Markov chains.3) satisfied by the ruin probability for the compound Poisson model. Tk and {Xt }o<t<Tk • For example. This program has been carried out in III. T1. However. Note.. . F(dx) = e7xF(dx). multiply (A.} be a renewal process. however. .e. that the existence of y may fail for heavytailed F. The distribution F of Y1. z(x) = e7xz(x). {Tn} if for any k. Z(u) U = 1 u 1 u f z(u . The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To.i. Eo etc. is called the cycle length distribution and as before. T1. We let FO. the postTk process {XT. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. A stochastic process {Xt}t>0 with a general state space E is called regenerative w.x)u(x) dx = z(u( 1 . where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). refer to the zerodelayed case. A regenerative process converges in distribution under very mild conditions: .. . i. Here the relevant F does not have mass one (F is defective). cycles. 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. Y1 . 0 PF µF 11 In risk theory. . ..r.d.. we let µ denote its mean.
assume that p < 00 and define Un = ZT}1 .334 APPENDIX Proposition A1. e(t )) . [0.. Then it (ii. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. under the condition of Blackwell's renewal theorem. (b) If in addition Var(Ul ) < oo.'s by e..tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.Tk : t < Tk} as the age.. but in fact.ZTOI < 00.3. where the distribution of X.i. {Tn}. and q(t) = sup It .0 is called cumulative w. fi (t) = inf {Tk .4 Let {Zt}t^. Then {Zt}t^. in total variation.e.r. Then Xt Di X. resp ..+ X.. Otherwise . then e (t) . (A.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo.e.r.t : t < Tk}.t. then (Zt . oo).. {Tn}. µ 0 If F is spreadout..t.d. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. are i. and we have: holds more generally that (rl(t). {Tn} if the processes {ZT +t . i. 2.6) id Cumulative processes Let {Tn} be a renewal process with i. We denote the limiting r. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.oo (i.t..d. cycles (we allow a different distribution of the first cycle). is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. then Xt .ZT Then: (a) If E sup I ZTo+t .v.i. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.0 be cumulative w.. oo). C). r.. {i7(t)} are Markov with state spaces (0..ZT }0<t<Y„+. This is the case considered in [APQ] V.r. 0<t<Yi then Zt /t a$• EU1/µ.. Then {e(t)}. If p = oo. just the same proof as there carries over to show: Proposition A1. for n = 1.
.t.5 Under the condition of Blackwell's renewal theorem. (c) the marginal distribution of q is FO.4. if in addition EYo < oo. but governed by a Markov chain {Jn} (we . (b) the joint distribution of (ri.^(t))} as a regenerative process. Since the maximum Mn of n i. assume first the renewal process is zerodelayed..i.d. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. . the joint distribution of (rl. Yl > t]. U(x + 1) .U(x) (c < oo because it is easily seen that U(x + 1) .'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). Yo > 0] + f Eo^ (t .y)P(Yo E dy) . ^ > y) = 1 f +Y (z)dz. In IV.y) = f U(t . the sum is o(t) so that Eo£(t)/t + 0 .v. EC(t)/t + 0. and the equivalence of (a) with (b)(d) is an easy exercise. Proof The number Nt of renewal before t satisfies Nt/t a4' p. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).d. r. the first statement follows.6 Consider a renewal process with µ < oo. we can bound e(t) by M(t) = max {Yk : k < 2t/p}. V is uniform on (0. l:) is the same as the distribution of (VW. Hence t t lt ) = f U(dy)z(t . = z is Foz) The proof of (a) is straightforward by viewing {(r. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . we used: Proposition A1.(t). Then fi(t)/t a4' 0 and. and the conditional distribution of ri given l. In the general case. (d) the marginal distribution of ^ is FO. Since z ( k) < E[Yi . use t E^(t)/t = E[Yo . W are independent. (1 V)W) where V. Y1i Y2..t. ^) is given by the following four equivalent statements: (a) P (77 > x. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .U(x) < U( 1)). are not i.i. Hence for t large enough.APPENDIX 335 Theorem A1. For the second. Y1 > t] 4 0.
.. . Jo. Yn..t.g. T_=inf{n>0: Sn<0}. These facts allow many definitions and results to be reduced to ordinary renewal.. Y1.} is nonlattice (it is easily seen that this definition does not depend on i).. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj.. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. IT. oo).+ < x. e. the semiregenerative process is called nonlattice if {T.r. .. with common distribution F. = io for some arbitrary but fixed reference state io E E. Then Xt 4 Xo. in [APQ]. Sn = X1 + • • • + Xn the associated random walk. Jn = i is the same as the P.. J1 i . . . the conditional distribution of {XT„+t}t>o given Yo. be i. . Notes and references Renewal theory and regenerative processes are treated.) and (Fij )i.jEE is a family of distributions on (0.d. For example..and regenerative processes.. A stochastic process {Xt}t>o is called semiregenerative w. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.}. . A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. . Alsmeyer [5] and Thorisson [372]. X2. A Markov renewal process {Tn} contains an imbedded renewal process. and define r+=inf{n>0: Sn>0}..r.t. . oo).. G_(x) = P(ST_ < x. namely {Twk } where {Wk } is the sequence of instants w where Jo. . Let X1. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. .7 Consider a nonlattice semiregenerative process. Jn_1.T_ < oo). 0] or (0 .. where the distribution of X.. r+ < oo). The semiregenerative process is then regenerative w. Further: Proposition A1.. G+(x) = P(S. < yIJ) = Fij( y) on {Jn= i.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. Jn +1=j} where J = a(JO.i . the Markov renewal process if for any n.
oo) (A. G+. F(A . F(A) + (G+ * G_)(A).. the renewal measures U+=>G+.S. n=0 n=0 00 00 and the T+. 0]). A C (0. (d) R+ = U_. A C (oo. define w as the time where the preT_ path S1. More rigorously.S. A C (0. 0] and (0.7). oo).T_=n} = {S.x)R+(dx). oo). 0]. (e) R_ = U+. 0<j<m.7) (A.1 . A C (oo. (A. n 0 R_(A) = E I(Sn E A).8) (e.7) follows since G+(A) = 0 when A C (oo.r.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).G+ * G_: (b) G_ (A) = f °° F(A . G_.=n w=m i Figure A.>0. S.. On {T_ > 2}. Proof Considering the restrictions of measures to (oc. Sr_ _1 is at its minimum .APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F.x)R_ (dx). we consider the last such time (to make w unique) so that {w=m. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. 0).=EGn. In (A. U. m<j<n}. u . . .. (c) G+(A) = f °. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).g. n=0 The basic identities are the following: Theorem A2.1 (a) F = G+ + G_ . >0.
m. E du) = P(T_=nm.1.XnEAx) 00 f 0 f 0 00 00 1: F(A .0<k<ri . SmEdu) = P(T+=m. A.1) that P(Sj Sn.. SnEAIS.8) is similar. m it follows (see Fig..x)P(Sk < 0.3 8 APPENDIX Reversing the time points 0. S. ..>0. (A. 0<j<m.du) (G+ * G)(A)• C llecting terms._ E A . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0..+ E du)P(S. It follows that for n > 2 F (7. .3._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). and reversing the order of summation yields P(T_ > 2.= n.F(r_n_mSrEA_u).u) f0m m=1 n=m+1 00 J0 OO P(S. ST_ E A) P(T+ = m. 0 < k < n. and the proof of (A. Aso. .+ E du) E P(S. (b) follows from 00 G+ (A) _ E F(Sn E A. Sn1 E dx) n=1  F(A . S. m=1 f S mming over n = 2.Sn_1Edx.1). ST+Edu). clearly (Sj Sm>0.. ST_ E A ._ = n . A.7) follows..x)R+(dx).. Sr_ E Adu) (s ee again Fig . m < j <n..
Sn_k. the analogue of a random walk is a process with stationary independent increments (a Levy process. and G+.G_[s]) (A.0+[s])(1 .6.1.T+> n) = P(Sk < O.SnEA) = P(Sn<Sk. this holds always on the line its = 0. E. 6+ [s].2 In terms of m.4). However.1). cf.O<k<n. oo). P(SnEA . u Remark A2. it serves as model and motivation for a number of results and arguments in continuous time.0<k<n. see e.g. Another main extension of the theory deals with Markov dependence. In discrete time. 0]. Then for A C (oo.9) whenever F[s].P as a product H+H_ of functions with such properties.1. is based upon representing G+ as in (b). Nevertheless. 11. u Notes and references In its above discrete time version.F[s] = (1 . WienerHopf theory is only used at a few places in this book. if {St} is Brownian motion.APPENDIX 339 and the proof of (c) is similar. see for example Bingham [65].g. In this generality of. Sk = X1 + • • • + Xk = Sn . and sometimes in a larger strip. the survey [15] by the author and the extensive list of references there. such developments motivate the approach in Chapter VI on the Markovian environment model.SnEA) = P(Sn<Sk. The present proof of Theorem A2.SnEA) is the probability that n is a weak descending ladder point with Sn E A. being concentrated at 0.0<k<n. we can rewrite (a) as 1 .'s. and the proof of (e) is similar.O<k<n. The classical analytical form of the WienerHopf problem is to write 1 . there is no direct analogue of Theorem A2. For example. and similarly H_ (s) = 1 .G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. the derivation of the form of G+ for the compound Poisson model (Theorem 11.. Summing over n yields R+ (A) = U_ (A).. For (d).f.SnEA) = P(SnSn_ k. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). . G_ [s] are defined at the same time.1(a) is from Kennedy [228]. Again.s. there are direct analogues of Theorem A2. G_ are trivial. which is basic for the PollaczeckKhinchine formula. consider a fixed n and let Xk = Xn_k+l. In continuous time. a number of related identities can be derived. Since G+ is concentrated on (0. then T+ = inf It > 0 : St = 0} is 0 a. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).g.
_I 0 (A.12) eA'AO = Ale AA (A.10) d dteAt = AeAt = eAtA (A. ere A is the eigenvalue of largest absolute value.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. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). 0 .13) henever A is a diagonal matrix with all diagonal elements nonzero.11) A f eAtdt = eA. and eQ can then be computed as the mth power (by squaring if = 2). whereas there is no similar single established a proach in the case of matrix exponentials. Thus. Here are.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. three of the c rrently most widely used ones: xample A3.5 that when handling phase type distributi ons. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). 1. Eo Kn/n! converges rapidly and can be evaluated without p oblems. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . 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). Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. however . To circumvent this. if m is s fficiently large. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). It is seen from Theorem VIII.
e. However .APPENDIX 341 Example A3. 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 . letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(..4 (DIAGONALIZATION) Assume that Q has diagonal form. The probabilistic reason that (A..3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.. p different eigenvalues Aj i . the procedure consists in choosing some suitable i > 0.3 i (A. 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. some jumps are dummy in the sense that no state transition occurs ). what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.. Zo = a (Z = QZ. i. i. One then can reduce to p linear differential equations by noting that k = ZQ. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.14) 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}. Ap. To this end. Let vi.e.]t)n (A. vp be the corresponding left . assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.7t) n=0 n! u °O n Pn (to see this. ..15) Then it is easily checked that P is a transition matrix . In practice. . The approach is in particular convenient if one wants eQt for many different u values of t.2 (UNIFORMIZATION) Formally. Here is a further method which appears quite appealing at a first sight: Example A3 . 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. condition upon the number n of Poisson events in [Olt])  Example A3. Zo = h).
and we may adapt some normalization convention ensuring vihi = 1. i= 1 i=1 P P (A.g H1. D = ) 2 2 .. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. and writing eQt as eQt = He°tH1 = H (e\it)di. hp the corresponding right (column) eigenvectors. i=1 i=1 Thus. There are.. and vihi ¢ 0. hi have been computed. of largest real part is often real (say. we have an explicit formula for eQt once the A j.17) eQt = E e\`thivi = E ea:thi ® vi.. hp. say Al. vi. i # j. some cases remain where diagonalization may still be appealing. say A = (Ai)diag. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. under the conditions of the PerronFrobenius theorem). however. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421.16) (A.342 APPENDIX (row) eigenvectors and hl.. The phenomenon occurs not least when the dimension p is large...18) contains terms which almost cancel and the loss of digits may be disasterous. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. we can take H as the matrix with columns hl.. Complex calculus : Typically. Then vihj = 0. v5Q = Aivi. Example A3. (A. Nevertheless. Then P P Q = > Aihivi = E Aihi (9 vi.18) Namely.. (A. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. not all ai are real. the eigenvalue. and hence A2 is so because of A2 = tr(Q). Qhi = vihi.
The other eigenvalue is A = A2 = q1 .6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. u Example A3. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.k1).21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. Then 7r = (ir1 7r2 ) = a (q21 Al .7 Let 3 9 2 14 7 11 2 2 . i. where (A.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. v2 and h2 can be computed in just the same way. Of course. l ab (g12g21 + (A1  411) 2) = 1. 1) . k  C k2 ) =b ( A1 q 1 Q11 / where a . b are any constants ensuring//Irk = 1. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e.20) ir = q2 ql qi +q 2 9l +q2 (A. replacing ai by A2.e.19) Example A3 . h2 = Thus.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.q. However. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 .
23) . and a generalized inverse may not unique. 2 2 1=ab(142+(1+2)2 ) = tab.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 . They are most often constructed by imposing some additional properties . (A. but only that dimensions match .satisfying AAA = A. ir =a(2 9 9 14 2 1 3 2 2)' k=b 14 =b 1+ 2 ir1 k1 ir2 k1 _ 9 2 10 5 7 9 70 1 ' 7r1 k2 7r2 k2 10 9 9 10 10 + 7 1 10 10 10 1 10 7 10 9 70 9 10 0 e4" = e_.11/2 + 5 1. A+AA+ = A+. A2 = 3/2 .344 Then D= 2+ 11)' 7 T4 2 =52. Generalized inverses play an important role in statistics.11/2 . (AA+)' = AA+.22) Note that in this generality it is not assumed that A is necessarily square.. for example AA+A = A.. APPENDIX x1 3/2 .
.e ® 7r)1.23) is called the MoorePenrose inverse of A. Then for some b > 0.g. lt o eAx dx = te7r + D(eAt .P + e7r)1 (here ( I . are ordered such that Al > 0.P + e7r ).. and exists and is unique (see for example Rao [300]). Rather than with generalized inverses .P). 0 01 In applied probability. one is also faced with singular matrices .I) (A. (I . . (A. if A is a possibly singular covariance matrix (nonnegative definite).ew.1 goes under the name fundamental matrix of the Markov chain).e.= (I .24) = te7r . Here is a typical result on the role of such matrices in applied probability: Proposition A4. one then works with Q = (Q . E. and define D = (A .APPENDIX 345 A matrix A+ satisfying (A.D + O(ebt). and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . = 0 where m < p is the rank of A... Am > 0.eir)1 = I .1 Let A be an irreducible intensity matrix with stationary row vector it.eir ).g.1Q = Q(Q .. Assume that a unique stationary distribution w exists . then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . ( Q . _ A. Am+1 = . ..25) . These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix.eir )1.
Note that h ® it has rank 1. see below.e.26) 2 = 2 e7r + tD . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. .27) Proof Let A(t).2 Let it be a row vector with m components and h a column vector with k components. and the columns to h. Equivalently. Then A(O) _ B(O) = 0.h. the r. 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. it follows that h ® it is the k x m matrix with ijth element hi7rj .I)}.D + D2 + O(ebt).3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ).91a(2) .h. For example. Interpreting 7r.I) (A.eir)eAt = eAt = A'(t).J {xe^r + D(e . the formulas involving O(e6t) follow by PerronFrobenius theory.s. respectively.I)} dx. (A. B'(t) = e7r + DAeAt = eir + (I ..s.I) . the rows are proportional to it. resp. I. o Finally.24). (A. 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.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx . . of (A.DZ(ent . and in fact any rank 1 matrix can be written on this form. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . h as 1 x m and k x 1 matrices.2e7r . h ® it reduces to hit in standard matrix notation. B(t) denote the l.
50 6 7 6 4f 4. (A.29). C2 = h2 are column vectors. (A B)' = eA®B e! L 1=0 0 . if A ® I occurs k times. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).31) Indeed. then v1B1h1 and v2B2h2 are real numbers. each of which is A ® I or I ® B.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.3vV/72f 20. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .4vf.4 eA® B = eA ®eB. Proof We shall use the binomial formula A crucial property is the fact that the functional equation for the exponential t / l (A ®B)t = I k Ak 0 B1k k=0 (A. (A.A9.5v/.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).31).3v'6. Using (A.3V8.APPENDIX 347 Then A®B = 2 f 20. and the number of such factors is precisely given by the relevant binomial coefficient. such a factor is Ak (&B 1k according to (A.29) If A and B are both square (k1 = ml and k2 = m2).(A. if Al = vi. A2 = v2 are row vectors and C1 = h1. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.3f 4v/. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .28) In particular.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.
in the definition (A. h. the same time. P(2). where transition matrix of the bivariate Markov chain {X n1). k any column vectors. {Yt(1). From what has been said about matrices of {Yt( 1). { 1't(1) }.6 Suppose that A and of B.32).33) . Thus . and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . p = P(1) ® {X }. Let P8f P(Sl). Yt(2 ) }. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. P(t) Yt(2) }. represents ces Q( 1).I)(h ® k).348 APPENDIX Remark A4. { On the other hand. Let further it. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .3 < 0 Lemma A4 . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. independent Markov chains. we have P8 = Pal) ® p(2). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . A special case of Proposition A4.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . (A. first term on the r . n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1).s. 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 +. X ) }. resp . Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . Q(2).
the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). . We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. . f o r each i. h such that vh = 1. we have AO = 1. so that by asssumption A ® B is u invertible... we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. .34) Note that for a transition matrix. ao). and if we normalize v.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. and appeal to (A.. That is.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. . n.. . 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements.7 Let A be a p x pmatrix with nonnegative elements. see e. E (0. = j and atk_li. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . [APQ] X. and the corresponding left and right eigenvectors v. and the corresponding left and right eigenvectors v. . Here is the PerronFrobenius theorem. h can be chosen with strictly positive elements. in such that io = i. il..1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. (b) if in addition A is aperiodic.. which can be found in a great number of books. i.. > 0 for k = 1. then IN < Ao for all A E sp(A).g. Similarly.29). h can be chosen with 3By this. p there should exist io. Now note that the eigenvalues of A ® B are of the form a +.12).APPENDIX 349 Proof According to (A. h = e and v = 7r (the stationary row vector). Then the eigenvalue Ao with largest real part is simple and real. . A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. j = 1. (A. . then An = Aohv+O(µ") = Aoh®v+O(µ") for some u.
Then for any (3. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. 10) and use the formula me at e Bt = e 00 Antn = e . not only in the tail but in the whole distribution..(ti)ding. let {Yti°i } be a Markov process with initial distribution a and intensity .35) for some p E (oo. I. h = e and v = 7r (the stationary row vector). if we normalize v. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A.8 is most often not stated explicitly in textbooks. To this end.1 Let Q be a proper irreducible intensity matrix with stationary distribution a.(3. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. Note that for an intensity matrix.n t AL n=0 n! (cf.8. the analogy of this procedure with unformization. Corollary A4. Furthermore. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . we have A0 = 0. h such that vh = 1.e.350 APPENDIX strictly positive elements. The next result gives a condition for asymptotical exponentiality.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). the condition is that t is small compared to Q. relate the eigenvalues of B to those of B via (A. For example. the phasetype distribution B(a) with representation (. Ao). one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic). note that we can write the phase generator T as Q . T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. Proposition A5.1.2). but is an easy consequence of the PerronFrobenius theorem. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. Example A3.
Then {Ix} is a Markov process with to = Yo. a'/a + 1. 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. 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}. = YQ(x). and this easily yields a(x)/x a' 1/t*.1.bij) Hence the intensity matrix of { Ix} is (qij/ti)i. a' = a . We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .2 Pi (c(a) > x.Yj(av) = j f . Proof Assume first ti > 0 for all i and let I. and write Yt = Yt(1).(a) > x .9. prove a somewhat more general result which was used in the proof of Proposition VI.YQ(av) = j) Pi ( ci(a'V) > x. Hence O ((a) aa. By the law of large numbers for Markov processes . and that Yt(a) = Yat for all t. We can assume that Jta) = Yt(°). in fact .APPENDIX 351 ((1) etc. Then a(a'V)/a (aV) a' 1. from which the phase process is terminated . Conditioning upon whether { Yt} changes state in [0.aE where 0 < e < 1). a . Let further V be exponential with intensity V and independent of everything else. v/ t.x (1 . has a limit distribution: Proposition A5. Since JJ(.g.a' + oo (e. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. J(()) _ = i) + at•x t tt' . In addition to the asymptotic exponentiality.jEE. dx/ti] or not. fo tY dv/t a$' t*. t < (a). {t Y( a) } v>0 . We shall . it states that the state. J^O)_ = j) Pi (v(aaV) > x.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((.
P. 2. A distribution B on {1. (c) the nth moment k 1 k"bkis 1)"n!aP"p.. the simplest discrete phasetype distribution: here E has only one element.. Et II I a(a^V) > x) at' . an easy modification of the argument yields finally the result for the case where t. zkbk is za(I . Penev & Turbin [238].. ' pk 0 k>1 11 Theorem A5.. these results are in the spirit of rare events theory for regenerative processes (e... Example A5. 1 k=1 1 0 otherwise.} is said to be discrete phasetype with representation (E..j) and initial distribution a. K}.. > 0}. = 0 for one or more i.. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).3 As the exponential distribution is the simplest continuous phasetype distribution. See also Korolyuk. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. k = 1. u Notes and references Propositions A5. let E and Pkj j=k1.g. . is discrete phasetype. say bk = 0. . (b) the generating function b[z] _ E' . Example A5..1 and A5. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. However.5 Let B be discrete phasetype with representation (P.zP)'p.2 do not appear to be in the literature.Pe. so we shall be brief. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. k>1. Then P is substochastic and the vector of exit probabilities is p = e .x k > K. Indeed. 2. so is the geometric distribution. a = b = (bk)k=1. with point probabilities bk = (1 .p)k1 p. a). Then: (a) The point probabilities are bk = aPklp.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.4 Any discrete distribution B with finite support.. Keilson [223]. .
r + 1. a.T(2)). r . T= ( 0 T(2) ) (A. 11 Example A5.. a' .. { Jt 2) } with lifetimes U1 . Jt t > U1 + U2. as is seen by minor modifications of Example A5. initial distribution a and phase generator T.a(1). resp..a(2). Then the convolution B = B1 * B2 is phasetype with representation (E. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).6 is the Erlang distribution Er which is the convolution of r exponential distributions. U2. . The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. _ i E E(1) T(1) t(1)a(2) i E E(2) .2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. and a=1).T(1)). and hence the negative binomial distribution is discrete phaseu type..APPENDIX 353 5c Closure properties Example A5. A.6 (CONVOLUTIONS) Let B1. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. resp.6.1 This corresponds to a convolution of r geometric distributions with the same parameter p.{ 0.36) in blockpartitioned notation (where we could also write a as (a (1) 0)).7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. B2 be phasetype with representations (E(1). (E(2). Then {Jt} has lifetime U1 + U2 .
0)a(2))). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). Then the mixture B = 9B1 + (1 .T(1)).d.a(2). U2.p)pn1B*n. T) and C = EO°_1(1 . if U1.'). Thus.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.E) where a(°) = fAa(a)v(da).8 (FINITE MIXTURES) Let B1. Example A5. then C is the distribution of Ul + • • • + UN. A reduced phase diagram is 0a(1) E(1) A . a.T. and o'i Oa. In risk theory. Let B(") be the corresponding phasetype distribution.4 . i E E(2) 0 T(2) =IT (in blockpartitioned notation.. To obtain a phase process for C. B2 be phasetype with representations (E(1).0)a(2) E(2) Figure A. are i.i. a mixture of more than two phasetype distributions is seen to be phasetype.T(2)).3 In exactly the same way. and consider B(") = fA B(a) v(da) where v is a probability measure on A. p at each termination.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. with common distribution and N is independent of the Uk and geometrically distributed with parameter p. this means that a = (Oa(1) (1 . a reduced phase diagram is f a E t Figure A.0)ai2)..a(1). Example A5.. Then it is trivial to see that B(") is u phasetype with representation (a("). P(N = n) = (1 . resp. we need to restart the phase process for B w. a. Equivalently.354 APPENDIX Example A5. (E(2). one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.O)B2 (0 < 0 < 1) is phasetype with representation (E.p)pn1.p. i E E(1) T 0 I (A.37) (1) (1 .
aF[T]. say with distribution F. let B be a continuous phasetype distribution with representation (F. be the point probabilities of a discrete phasetype distribution with representation (E. P).11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U .2. Minor modifications of the argument show that 1.f. . let the phase space be E x F = {i j : i E E. then Jy has distribution aeTx.X)+ is zeromodified phasetype with representation (E. then U1 +• is phasetype with representation (E. resp. U2. U2 be random variables with distributions B1. a(1) ® a(2 ). +UN 2. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta).2. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. a.7.. i.x)+. if U1. T + ta.v. T(2) ). 12 (PHASETYPE COMPOUNDS ) Let fl. X independent of U. are i..T + pta). E).1. resp.. T) and C = F. let {Jtl)}. cf.T) where F[T] = J0 "o eTx F(dx) u is the matrix m.. If we replace x by a r. v.d. of F. Example A5 . Example A5.9) that (U . a(2). then C is the distribution of U1 + • • • + UN. a. then U1 + • • + UN is zeromodified phasetype with representation (a. B2 of phasetype with representations (E('). T). it follows by mixing (Example A5.. Example A5 . if B is defective and N + 1 is the first n with U„ = oo. but the same T.. v. Indeed. Note that this was exactly the structure of the lifetime of a terminating renewal u process. { Jt2) } be independent with lifetimes U1.T) if U is phasetype with representation (E. 13 (MINIMA AND MAXIMA ) Let U1. T(1) ® T(2)). Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. Thus the representation is (E(1) x E(2). Corollary VIII. Proposition VIII. j E F}. To obtain a phase representation for C . U2.°_1 f„ B*?l. . T + pta). Equivalently. { 4 } as exit of {Jt}. It is zeromodified phasetype with representation (E. say v..aeTx.°. cf.. a. If U1 has a different initial vector. For U1 A U2.g.a(1). with common distribution B and N is independent of the Uk with P(N = n) = f. f2.APPENDIX 355 and C is phasetype with representation (E. if {Jt} is a phase process for U. (E(2). To see this.TWWW).
Example A5. Proof Assume first that B is a onepoint distribution. see Neuts [269] (where the proof. By the diagonal argument (subsequent thinnings). elementary) Let {bk} be any dense sequence of continuity points for B(x). cf... Thus the state space is E(1 ) x E(2) U E(1) U E( 2). i= 1 C.14 To a given distribution B on (0. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. we can assume that ID.B(bk) I < 1/n for n > k... The general case now follows easily from this. with weight pi(n) for xi(n).(bk) + B(bk) for all k. Let the support of Dn be {xl(n). Now we can find first a sequence {Dm} of distributions with finite support such that D. That is. say degenerate at b.xq(n)(n)}.8. Hence it is immediate that Bn 4 B. r # oo. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. Then we must find phasetype distributions Bn with B.. oo). there is a sequence {B. relies more on matrix algebra than the probabilistic interpretation exploited here). Here are the details at two somewhat different levels of abstraction: (diagonal argument ..n = I:pi(n)Er v ( __ ) n) ) a= 1 ..2) } to go on (on E(2)) when { i 1) } exits.(Sn) with Sn = n/b. we need to allow { Jt. and vice versa. any distribution B on (0.} of phasetype distributions such that Bn 3 B as n + oo.. Then from above. however. 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 .(n) = D. and the closedness of the class of phasetype distributions under the formation of finite mixtures.(bk) + B(bk) for all k as n * oo.356 APPENDIX For U1 V U2. the initial vector is (a(1) (& a (2) 0 0). q(n) q(n) pi(n)a .(bk)'. and let Bn be the Erlang distribution E. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .
For a general Bo. then it is immediate that WI(B) = p2(B) for all distributions B on [0. x 4 oo. replications). oo) approximation Assume that we can compute a functional W(B) when B is phasetype.i. E E.D(bk)I < n. u 2 (abstract topological ) The essence of the argument above is that the closure (w.n( b k ) . . 2.n (bk) .. i = 1. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. u Theorem A5. Let E be the class of functions f : [0. the class CO of all discrete distributions.14 is fundamental and can motivate phasetype assumptions. In particular.d. we can then approximate Bo by a phasetype B... Since PET is closed under the continuous operation of formation of finite mixtures. oo) * [0. and we can take Bn = Cr(n). It should be noted.. But To is the class G of all distributions on [0. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f.. i. 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. and that cp is known to be continuous. k < n. for some a < oo.. Then ICr( n ). oo) and any fl.(x)Bf. k < n.t.r.e. oo) such that f (x) = O(e«x).B(bk )I < . Corollary A5. f2. say on the claim size distribution B in risk theory.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). one would use the B given by some statistical fitting procedure (see below).n.. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions.15 To a given distribution B on (0 . if information on Bo is given in terms of observations (i. If Cpl (B) and ^02(B) are weakly continuous. Hence G C PET and L = PIT. compute W(B) and use this quantity as an approximation to cp(B0). oo)..( dx) * f r f{(x)B(dx). however. PIT contains all finite mixtures of onepoint distributions.
then cc f (x)Bn ( dx) = (?!c ) e'= . .38) We first show that for each f E E. . n B=az.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).  APPENDIX B implies that 00 o o 00 n. and the case of a general f then follows from the definition of the class E and a uniform integrability argument.39) Indeed..f (x)B(dx).2 . i = 1. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). .. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. .. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge.(dx) > J fi(x)B(dx). TO (A....n(dx) < 1+.358 Proof By Fatou' s lemma. if f (x ) = e°x. 2.. for each i.. By (A. oo). i = 1.38 ). . n. n. \\ 0 Corollary A5.n(dx) + f 0 fi(x)Dn(dx).oo J fi(x)B. liminf B.. we may assume that in the proof of Theorem A5.16 To a given distribution B on (0 . and hence we may choose r(n) such that L 9l) f (x)Cr(n). f° xtBn(dx ) * f °° x`B( dx).f ' f (x)B(dx).f (z) = f = 1 1 1 1n/ o . f00 fi(x)Cr. Now returning to the proof of (A.. i=1..39). i = 1...
/3) is defined as the unique solution > 0 of B[y] = l+y/j3. 0 as i * oo. (N or a given distribution Bo. For practical purposes. . . 5e Phasetype fitting As has been mentioned a number of times already.17 To a given /3 > 0 and a given distribution B on (0.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal.16. The adjustment coefficient is a fundamental quantity. from a more conceptual . /3) = ry for all n./3). e ) and ei J.> y for some sequence {ei} with ei E (0. The present section is a survey of some of the available approaches and software for inplementing this. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. I. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. lim inf > is proved similarly. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . . the adjustment coefficient 'y = 7(B. one can obtain 7(Bn. However. Proof Let fi(x) = el'r+E. there is a sequence {B. . then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . .l3µb < 1. but are certainly not unexpected.18 In the setting of Corollary A5. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . oo) with B[y +e] < oo for some e > y = 7(B. lim sup ryn < 7. . the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. O We state without proof the following result: Corollary A5.. (N. Notes and references Theorem A5.e.14 is classical.3).. .. If ei > 0. the remaining results may be slightly stronger than those given in the literature.
The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa .'s). and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. It seems therefore a key issue to develop methods allowing for a more general phase diagram.. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be).g. and in practice this sets a limitation to the usefulness (the curse of dimensionality .g.g . e . The characteristics of all of these methods is that even the number of parameters may be low (e. [70]) restrict attention to acyclic phase type distributions . 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 L1 distance between the c . giving mass 1 /N to each S=. we do not not want to perform matrix calculus in hundreds or thousands dimensions). A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . the number of phases required for a good fit will typically be much larger. and used a nonlinear programming approach .. we have constructed a sequence { B.f. at a a number of selected points .f. . 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. 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. for some suitable large n. A method developed by Bobbio and coworkers (see e. risk theory. defined by the absence of loops in the phase diagram . [202]. d. [317] ) has considered an extension of this setup.. g.d. In a series of papers (e.} of phasetype distribution such that Bo. . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . cf. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. Of course. three for a mixture of two Erlangs ). The likelihood function is maximized by a local linearization method allowing to use linear programming techniques.. a program package written in C for the SUN workstation or the PC is available as shareware. and as fitted distribution we may take B. Schmickler (the MEDA package.g.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . reliability or queueing theory. [216] ). Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . (N is the empirical distribution Be.
.T(n) (Nik IC1..(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). eieT(n)((k. it is easy to see that N (k Ea(n). Thus.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .T (n)(TiI(1. .APPENDIX 361 tion Sk was available. . jEEA.g. .. (N) tJk Ea ( n).T(n) (Ti ^^ 1... one is lead to an iterative scheme. the methods of [70] and [38] appear to produce almost identical results.. it seems open whether the restriction to the acyclic case is a severe loss of generality. EN where ai = N 1 I ((k) = i) tii=i iEE.. then the estimators would be of simple occurenceexposure type. In practice. e. In fact.. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. (n+1) _ Ea (n). Nii = = .x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. ..g.. E. = j) f k=1 k =1 tE[0. N Ti = I(J= i) dt. since this is parameterdependent.(N) = E Ea(n). The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.
This page is intentionally left blank .
New York. New York. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. pp. Stoch. Asmussen (1987) Applied Probability and Queues. Act. New York. Skand. Chichester New York. ed. 213229. Arfwedson ( 1955) Research in collective risk theory. J. In: Limit Theorems and Related Problems (A. Queueing Systems 5 . Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . 311338. 191223. 1985. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Arfwedson ( 1954) Research in collective risk theory. Abate & W. 38. Abate. Asmussen (1985) Conjugate processes and the simulation of ruin problems. 1984 . [3] J. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Appl. [2] J. The case of equal risk sums. [5] G. 57. 14. Teubner. [4] M. [9] G. [12] S. B. AT&T.Bibliography [1] J. Probab. The Mathematical Scientist 14.G. Stuttgart. 587. 20. John Wiley & Sons. Aktuar Tidskr. 53100. Adv. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . [11) S. 253267. Optimizations Software. [6] V. [7] E. 3157 .A. 37. [14] S. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Preprint. [13] S. Queueing Systems 16. [10] K. G. II. 143170. 345368. Skand. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Aktuar Tidskr. Dover. [8] G. Borokov. with applications to risk reserve processes and the GI /G/1 queue. 101116.L.). Anantharam (1988 ) How large delays build up in a GI/GI11 queue. 219229. Scand. Queueing Systems 10. Choudhury & W.). Asmussen (1984) Approximations for the probability of ruin within finite time . Alsmeyer (1991) Erneuerungstheorie. 363 . Proc. Transactions XVth International Congress of Actuaries. Abramowitz & I. ibid. Appl. Abate & W. [15] S.
S. Dshalalow ed. Alfa & S. Asmussen & K. Asmussen (1995) Stationary distributions via first passage times. 106119.K. 37. Statist. [19] S. [28] S. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. 313341. J. A. Proc.). Asmussen & M. 4966. Appl.). [21] S. 297318. Bernoulli 6. [23] S. Ann. Chakravarty. J. Marcel Dekker. Scand. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [24] S. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue.). 193226. Act. [20] S. Asmussen & M. T. Probab. 2943. New York. [18] S. Asmussen (1999) On the ruin problem for some adapted premium rules. Riga Aviation University. Asmussen (1992b) Light traffic equivalence in single server queues. 189201. Floe Henriksen & C. Act. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour.364 BIBLIOGRAPHY [16] S. Scand. 20. SIAM Review 40. [33] S. 1996. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. 69100. 315. L. Ann. 1989 . Proc. Asmussen & B. . 54. 96107. 1936. [31] S. 79102. K. Appl. Astin Bulletin 27. Th. Asmussen. [30] S. 137168. 27. Asmussen. 313326. Binswanger & B. [27] S. 354374. CRC Press. Scand. Probab. 303322. 25. Stoch. Rolski & V. Random Processes 2. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Appl. Florida. Probab. 1999 . stationary distributions and first passage probabilities. J. Andronov. [26] S. Appl. Scand. [29] S. Methods V Problems (J. 2. Act. Probabilistic Analysis of Rare Events (V.M. Boca Raton. [22] S. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. MatrixAnalytic Methods in Stochastic Models (A. Kalashnikov & A. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Asmussen (2000) Matrixanalytic models and their analysis. Asmussen. [34] S. Asmussen & B. eds. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. J. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Asmussen (1989b) Risk theory in a Markovian environment. Stoch. Hojgaard (2000) Rare events simulation for heavytailed distributions. 2149. eds. Advances in Queueing: Models. [32] S. Extremes 1 . 8. Stochastic Models 11. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Frey. 772789. [25] S. [17] S. Ann. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 555574.
64. Insurance: Mathematics and Economics 20. Res. Asmussen & R. Marked point processes as limits of Markovian arrival streams.).Y. Asmussen & V. Asmussen & D. 30. Example of excessoffloss reinsurance for an insurance corporation. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Nielsen (1995) Ruin probabilities via local adjustment coefficients . 410433. Asmussen & G. J. B. Probab. Asmussen. Sigman (1996) Monotone stochastic recursions and their duals. [51] S. 299324. (to appear). with applications to insurance risk Stoch. Finance and Stochastics 4. 31. Schmidli & V. Wiley. Advances in Queueing: Models. Appl. [48] S.Y. [41] S. Asmussen & C. . Oper. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Prob 32 . Probab.A. Math. Kliippelberg (1996) Large deviations results for subexponential tails. Proc. Management Science 45. Asmussen & C. 19. 103125. [42] S.BIBLIOGRAPHY 365 [35] S. Taksar (2000) Optimal risk control and dividend distribution policies . Asmussen. [40] S. Schmidt (1995) Ladder height distributions with marks. 115. Dshalalow ed. Asmussen & R. Opns. Th. Hojgaard & M. Appl. [39] S. Statist. 20. 429466. Proc. Appl. 10. O.A. Adv.). 259274. Eng. Scand. [36] S. [47] S. Fitting phasetype distributions via the EM algorithm. Asmussen . 736755. 23. Sc. Appl. Asmussen & M. 913916. 365372. first passage problems and extreme value theory for queues. Appl.M. Asmussen & V. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Inf. Adv. Koole (1993). Olsson (1996). Rubinstein (1999) Sensitivity analysis of insurance risk models. H. Asmussen & T. Asmussen & K. [46] S. Asmussen & S. Stoch. 422447. 120. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. 58. Asmussen & T. Appl. Insurance: Mathematics and Economics 10. Probab. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. [50] S. [52] S. [45] S. Read eds. Asmussen & H. Res. CRC Press. 11251141. [37] S. Probab. 105119. Stochastic Models 8. Methods 8 Problems (J. Perry (1992) On cycle maxima. [49] S. Johnson. [43] S. [38] S. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Nerman & M. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & C. J. 419441. [44] S. Florida. 421458. J. Supplementary Volume (Kotz. Boca Raton. Statistica Neerlandica 47.
H. Scand. Soc. 129134. [72] P. [60] J. Act. J.A. Bingham (1975) Fluctuation theory in continuous time. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails .H. Bobbio & M. Springer . [65] N. 77111. 1974. Appl. Insurance: Mathematics and Economics 4. [54] K. Beekman (1985) A series for infinite time ruin probabilities. Cambridge University Press. Probab. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Act. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Borovkov (1976) Asymptotic Methods in Queueing Theory. J. Cox (1972) A low traffic approximation for queues. SIAM. [68] T. Probab. 221232. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. 1975. Scand. 9. 661667. 275279. [58] O.M. BarndorffNielsen & H. [59] O. Boogaert & V. Boogaert & A. Asmussen & J. de Waegenaere (1990) Simulation of ruin probabilities. Oxford. 169186. Ney ( 1972) Branching Processes . New York. Scand. [66] N. 832840. J. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. [55] B. Halsted Press. Athreya & P. Beekman (1974) Two Stochastic Processes. . 190204. Grandell (1985) An insensitivity property of the ruin probability. 9599. [73] A. J. New York. Beekman (1969) A ruin function approximation. [56] B. Act. Baker (1977) The Numerical Solution of Integral Equations. [62] J. 148156. 1995. [61] J. Wiley.B .L. C. New York. J. [57] C. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Clarendon Press. SpringerVerlag. [64] P. Actuaries 21. Teugels (1987) Regular Variation. Goldie & J.366 BIBLIOGRAPHY [53] S.T. Insurance: Mathematics and Economics 6.R. Bjork & J. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Billingsley (1968) Convergence of Probability Measures. Scand. 610. J. Berlin. Appl.A. [63] A. Adv. Bloomfield & D.H. [69] P. 1985 . Insurance: Mathematics and Economics 9. Wiley. Trans.L. 4148. Scand. 1988 . Cambridge. Berman & R. Bingham. 33. J. 705766. Bjork & J. Appl. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Probab. Stochastic Models 10. [71] P. Act. Act. [70] A. 11811190. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. 7.J. [67] T.
392433. Syst. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Tech. Oper. Chandy & M. Skandia Jubilee Volume. Burman & D. Philos. IEEE Trans. J. Reiser eds. II. Brockwell . SpringerVerlag. Cramer (1955) Collective risk theory. B. [93] K. [82] H. Nonparametric estimators for the probability of ruin. [76] N. Burman & D . verve. Insurance : Mathematics and Economics 9. Aut. [81] J. New York. Jr. [79] P.Y . Qinlar (1972) Markov additive processes.L. Fox & L. 93121. 907920. Resnick & R. Academic Press. 16. [80] F. New York San Francisco London. H. Springer.C. 24.). Hickman. Amsterdam.M. Croux & N. [77] P. Bratley.L.) NorthHolland.I.BIBLIOGRAPHY 367 [74] O. Gerber. Wiley. Proc. Itasca. S. 62. IEEE J. Amsterdam. [84] D. Bucklew ( 1990) Large Deviation Techniques in Decision. 34. [92] H. Fort & G. Wahrscheinlichkeitsth. Berlin. Schrage (1987) A Guide to Simulation. [85] W.W. Cohen (1982) The Single Server Queue (2nd ed. NorthHolland. AddisonWesley.). Reading. Computer Performance (K. Res. Cottrell.J. 105119.A.A. Appl.R. Veraverbeke (1990). J. De Vylder (1986) Ordering of risks and ruin probabilities. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Illinois.W. The Society of Actuaries . 749763. Cramer (1930) On the Mathematical Theory of Risk. The Jubilee volume of Forsakringsbolaget Skandia. Probab. [88] J.C. Bell. 2338. Bux & U. Cox (1955) Use of complex probabilities in the theory of stochastic processes. [87] E. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . 14331453. 14. Soc. Boxma & J.R. [75] O. Jones & C. D.U. M. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Tweedie (1982) Storage processes with general release rule and additive inputs . Stockholm. [89] M. Herzog (1977) The phase concept : approximations of measured data and performance analysis.J. Cambr. Broeck .Y.L. Simulation and Estimation. . [86] K. 51. J. [91] H. Biihlmann ( 1970) Mathematical Methods in Risk Theory. [90] D. Breiman ( 1968) Probability.R. Z. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Boxma & J. Bowers . Area Commun. Insurance : Mathematics and Economics 5. 313319. Sel. New York. [78] L. Stockholm. Goovaerts & F. Control AC28. 3539. Chung (1974) A Course in Probability Theory (2nd ed.L. 177204. Nesbitt (1986) Actuarial Mathematics ..W. [83] D. Queueing Systems 33. Geb. Adv. 127130.
Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Suppl. 85116. Rolski (1991) Light traffic approximations in queues. 27. [107] L. Zeitouni (1992) Large Deviations Techniques and Applications. Deheuvels & J.B. J. Act. [98] A. 16. Haezendonck (1987) Classical risk theory in an economic environment. Csorgo & J. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. J. Daley & T. English version published in Skand. Res. 3. T. Rolski (1984) A light traffic approximation for a single server queue. De Vylder & M. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . 121131. Insurance: Mathematics and Economics 10. [105] A. Jones and Bartlett. Scand. Insurance: Mathematics and Economics 7. Insurance: Mathematics and Economics 3. 9 . 201206.D. 7083 (1969). 583602. Csorgo & J. Tidskr. . [111] F. [101] C. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Dempster. N. Statist. Roy. Aktuar. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Scand. Math. 3750. Stochastic Models 5. [110] F. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Comp. [102] P. Embrechts (1989). 17. 435436. 22. [112] F. New York. Dembo & O. Insurance: Mathematics and Economics 4. Dassios & P. Boston. Appl. Delbaen & J. Oper. [100] A. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. [99] R. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. [104] F. Math. 5771.P.. Math. 88101.L. Appl. Springer. Probab. Actuarial J. Goovaerts (1984) Bounds for classical ruin probabilities. Laird & D. [108] F. De Vylder (1977) A new proof of a known result in risk theory. Stockholm. Martingales and insurance risk. [109] F. 227229. Daykin. Devroye (1986) Nonuniform Random Variate Generation. Haezendonck (1985) Inversed martingales in risk theory. Soc. J. Daley & T. 624628. 114119.M. 1990 . Pentikainen & E.368 BIBLIOGRAPHY [94] M. De Vylder (1996) Advanced Risk Theory. De Vylder & M. Delbaen & J. Chapman & Hall. [97] D. 135159.. 181217. Editions de 1'Universite de Bruxelles. J. [106] A. Pesonen (1994) Practical Risk Theory for Actuaries. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Oper. [96] D. Insurance: Mathematics and Economics 6. Res. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . [95] M. [103] F.
363374. 118. [131] P. J. British Actuarial J. [120] D.U.W.M. Scand. Philos. Act. 5159.C. Dickson (1995) A review of Panjer's recursion formula and its applications.S. 1993 . Insurance: Mathematics and Economics 7. Scand. J.C. Oper. 105115. Insurance : Mathematics and Economics 22. Act.R. Dufresne & H.C. Emanuel . Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Hipp (1999) Ruin problems for phasetype(2) risk processes. Insurance : Mathematics and Economics 11.K.BIBLIOGRAPHY 369 [113] D.M. Act. [118] D. . 193199. 1984 . Insurance: Mathematics and Economics 25.M. Letters 7. [124] N.C. [117] D. J. [128] E.M. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. 229232. with applications. Scand. 3745. Gerber (1988) The surpluses immediately before and at ruin. Embrechts .M.J. Berlin Gottingen Heidelberg. 6180. 1975 . Act. 1741. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. [123] E. Insurance : Mathematics and Economics 7. Dynkin (1965) Markov Processes I. 1. 4259. Scand. Embrechts ( 1988 ). Scand. Act. Dufresne & H. Insurance : Mathematics and Economics 10.J. Djehiche (1993) A large deviation estimate for ruin probabilities. 191207. [126] F. [119] D. [116] D. Dickson & J. Proc. Dickson & H. 1994 . Grandell & H. 1993 . Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Astin Bulletin 21. Dickson & H. Dufresne. Regterschot (1988) Conditional PASTA.M. Waters (1999) Ruin probabilities wih compounding. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 4962. J. 107124.C. Dickson & C.R. Scand. Camb. Insurance: Mathematics and Economics 19. 251262. Scand. Act. van Doorn & G. Gerber & E. Act.C.M. [115] D. 174186. [122] B. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue.U.C. 177192. H.M. 131138.M.G. [130] P. Waters (1996 ) Reinsurance and ruin . Harrison & A. [121] D. Duffield & N. J. 1984 .C. Soc.M. [125] F. Ruin estimates for large claims . J. Dickson (1994) An upper bound for the probability of ultimate ruin. Math. Springer . 2000 . and the amount of claim causing ruin. Res. Shiu (1991) Risk theory with the Gamma process. 269274.B.C. [114] D.R. Dickson & C. J. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. J.R. Dickson (1992) On the distribution of the surplus prior to ruin . 147167.C. Dickson & J. [127] F.U. J. [129] D .
[139] A. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics .M. Wiley. Esary. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Adv.K. Astin Bulletin 16. Embrechts. Embrechts & J. Math. 14661474. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [138] P. 17. Lai (1998) Wald's equations. Probab. Konig. Probab.D.370 BIBLIOGRAPHY [132] P. 8190. 131137. Erlang 2. Approximations for compound Poisson and Polya processes. Trans. Maejima & J. Probab. Jensen. Embrechts.A. Appl.). Nyt Tidsskrift for Matematik B20. [143] W. [133] P. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Act. A bootstrap procedure for estimating the adjustment coefficient. J. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . [147] M. Appl. Insurance: Mathematics and Economics 7.J. Probab. J. Embrechts & T. Statistica Neerlandica 47. 3339. D. C. [148] C. [136] P. Akt. Proschan & D.L. Schmidt (1992) An insensitivity property of ladder height distributions. Arndt & V. Walkup (1967) Association of random variables. Adv. Embrechts & N. with applications. [137] P.K. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A.). Frees (1986) Nonparametric estimation of the probability of ruin. Ann. Fuh & T. 35. Danish Academy Tech. 26. Scand.D. Frenz & V. 1998 . Appl. Sci. 695712. Fitzsimmons (1987). New York. Wiley. Skand. U. Insurance: Mathematics and Economics 10. [149] C. Appl. M. 566580. Kliippelberg & T. Teugels (1985). 616624. Insurance : Mathematics and Economics 1. [141] F. 269274. [144] P. 623637.W. [145] P. [150] H. Embrechts. . [146] E. 29. J. Probab. Villasenor (1988) Ruin estimates for large claims. Griibel & S. Mikosch (1991). Schmidli (1994) Ruin estimation for a general insurance risk model. Statist. [135] P. 159178. 38. On the excursions of Markov processes in classical duality. New York. 181190. Springer. [140] J. 5975. Probab. Fields 75. Adv. Tidsskr. Appl. [134] P. Esscher (1932) On the probability function in the collective theory of risk. 175195. F. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Th. Rel. 29. R.W. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. [142] W. Heidelberg. first passage times and moments of ladder variables in Markov random walks. 5572. Wiley. Embrechts & H. Schmidt (1982) Queues and Point Processes. Furrer (1998) Risk processes perturbed by astable Levy motion.L. Franken. 404422. 5974. J.D.
ETH Zurich. Glasserman & S. Amsterdam.G. Galambos & J. Scand. [152] H. van Heerwarden & T. [155] H. [157] H. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. London. Aktuarietidskrift 1970 . J. 424445. [160] H.N.U. [167] C. 1978 . [158] H. contained in the authors PhD thesis.J. 6370. Birkhauser . 5. Gerber (1986).S. Vers.U. 7778. Gerber (1979) An Introduction to Mathematical Risk Theory. Probab. 31A. Appl. Kou (1995) Limits of first passage times to rare sets in regenerative processes. 1523. Schweiz. Ann. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Springer. Goldie & R. Ver. A. [164] B. Act. 28.. Gani.). 463480. Math. Unpublished. [163] P. Adv. Vers. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'.U. de Vylder & J.U. Kaas. Suppl.V. Gerber (1988) Mathematical fun with ruin theory. University of Pennsylvania. Glynn & W. 131156. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Basel. [169] J. Insurance : Mathematics and Economics 7. Schweiz. Goovaerts. Kluwer. In Studies in Applied Probability (J. Insurance: Mathematics and Economics 20.BIBLIOGRAPHY 371 [151] H. [154] H. [162] P. J. 1981 . .U. Insurance: Mathematics and Economics 15. Z. R. 73. S. Life Insurance Mathematics.M. Goovaerts. [168] J. Mitt.71. Boston. eds. Skand. Griibel (1996) Perpetuities with thin tails. Amsterdam. [165] M. Appl.U. Dordrecht. 205216.. Appl. Michna & A. 97114. F. Bauwelinckx (1990) Effective Actuarial Methods. J. Scand. Scand. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. 2336. 1977.U. NorthHolland.W.). Mitt. Probab.J. Weron (1997) Stable Levy motion approximation in collective risk theory. [156] H. Gerber (1973) Martingales in risk theory. Ver. [159] H. 3752. Haezendonck (1984) Insurance Premiums. NorthHolland. Act. Gnedenko & I. Act. J. Math. [166] M. 105115. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Grandell (1977) A class of approximations of ruin probabilities.E. [161] P. 205210. Furrer. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Furrer & H. [153] H. Huebner Foundation Monographs. Probab.
Winter School on Stochastic Analysis and Appl . Segerdahl (1971) A comparison of some approximations of ruin probabilities. Harrison & S. Grandell & C. 566576.and Sozialforschung 6. M. [189] A. Grandell (1990) Aspects of Risk Theory. 1. Ibragimov. [175] J. Grnbel (1991) G/G/1 via FFT. Amsterdam. 3041. .I. 167176. [174] J. Chapman & Hall. 355365. SpringerVerlag. Grandell (1997) Mixed Poisson Processes.V.372 BIBLIOGRAPHY [170] J. Gyllenberg & D. Math. Griibel & R. Insurance: Mathematics and Economics 26.J. Matem. Opns . Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Lemoine (1977) Limit theorems for periodic queues. Appl.. Appl. Res. SpringerVerlag. 33. [180] R. [182] A. 54. M. Theory and Applications. 347358. ed. [171] J. [172] J. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Appl. New York. Proc . Probab. 8. 14. Resnick (1977) The recurrence classification of risk and storage processes .A. Th. M. Applied Statistics 40. [178] B. Informatica 2. Math. Rink. [176] B. AkademieVerlag. Probab.). Stoch. Astin Bulletin 29. [186] J. Berlin.S.I. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . [173] J. 5 . KTH. [181] D. Gusak & V. Probability Theory and Mathematical Statistics (I. Proc. 197214. Res. [187] J. Liet. [184] H .O. 14. Appl. Grandell (1999) Simple approximations of ruin functions. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Proc. [185] J. Archiv v. Preprint. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. mathematische Wirtschaft . Opns . Tinbergen Institute Research Series 20. 400409. [177] B. [188] J. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes .243255. Aktuar Tidskr. M. Grandell ( 1992) Finite time ruin probabilities and martingales . Grandell (1979) Empirical bounds for ruin probabilities. 7590. J. Grigelionis (1996) Lundbergtype stochastic processes. Harrison & S. 3 . Statistical Algorithm 265. 332. Harrison (1977) Ruin problems with compounding assets . Skand. E. Harrison & A. [179] R. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Gut (1988 ) Stopped Random Walks . 5766. 6779. Stoch. [183] M.131135. 143158. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting.
Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Hoglund (1991). Ann. Huskova. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Proc. [208] V. Appl.L. Asmussen & 0. Studies in Statistical Quality Control and Reliability 1992 :4. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Math. [202] 0. Horvath & E. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Wahrscheinlichkeitsth. Heilmann (1987) Grundbegriffe der Risikotheorie. ACM TOMACS 6. [200] M.L. 1990 . Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Verlag Versicherungswirtschaft e. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. 305313. Insurance: Mathematics and Economics 5. Insurance: Mathematics and Economics 5. [193] U. New York. . 166180. 259268. Karlsruhe. Probab. Probab.M. 8996. [191] W. Geb. Probab. 377381. Electronic Letters 35. [204] T. Mathematical Statistics. 8095. Z. 123151. J. 1998 . 11631174. Wiley. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. No. Nerman (1992) EMPHT . [195] B. [203] T.V. Versicherungswirtschaft. verve. [192] U. Hipp & R. The ruin problem for finite Markov chains.. [206] B. J. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Chalmers University of Technology and the University of Goteborg. 25. 3.A. Klugman (1984) Loss Distributions. Nachrichten 30. 23. 4385.BIBLIOGRAPHY 373 [190] P. Karlsruhe. on Asymptotic Statistics (P. Hojgaard & M. Iversen & L. Astin Bulletin 19. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. J. Haggstrom. Iglehart (1969) Diffusion approximations in collective risk theory. Hipp (1989a) Efficient estimators for ruin probabilities. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. 285293. [197] C. Ann.B. 29. Hogg & S.R. 12981310. Scand.V. [194] 0. [205] T. Ann. Foruth Prague Symp. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. J.a program for fitting phasetype distributions.). Act. eds. Statist. 285292. 19. [207] D. Prob. Mandl & M. [196] C. 18. [201] L. S. 378389. 5770. Scand. Appl. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. [199] R. 6. Hill (1975) A simple general approach to inference about the tail of a distribution. [198] C. Act.
3. 35. Math. Astin Bull. Insurance: Mathematics and Economics 5. 173190.C. J. Soc. Jelenkovic & A. Proc.M. [217] R. N. J. Kluwer. 325247.A. ibid. 866870. [211] J.G. 11. Taaffe (1989/90) Matching moments to phase distributions. Probab. Jensen (1953) A Distribution Model Applicable to Economics. [227] J.B. 711743. Keilson & D. Princeton. [223] J. 6. Goovaerts (1986) General bound on ruin probabilities. [218] V. Kaas & M. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. ibid. Amer. 31. Copenhagen. Statist. Jensen (1995) Saddle Point Approximations.374 BIBLIOGRAPHY [209] D. Stochastic Models 1. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. 8793. NEC Research Institute. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison.J. Johnson & M. Oxford. Technometrics 30. 259281. 547567. 283306.G.M. Act. Clarendon Press. Jagerman (1985) Certain Volterra integral equations arising in queueing.M. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. [216] M. [221] E. Academic Press. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain.R. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. 116129. 561563. Keilson & D. 15. Addenda to for processes defined on a finite Markov chain. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Astin Bull. University of Chicago Press.J. 61.L. . [226] J. 165167. Ann.G.M. Proc. [214] A. Kalashnikov (1996) Twosided bounds of ruin probabilities. 123133. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Manuscript. 1996. Soc. 60. 118.M. Cambridge Philos. Janssen & J. 187193. Keilson & D. Kao (1988) Computing the phasetype renewal and related functions. Karlin & H. [210] D. [225] J. J. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Soc. Munksgaard. [212] J. Taylor (1981) A Second Course in Stochastic Processes. Appl. Act. Chicago. [219] V. Wishart (1964). 4151. Probab. 63. 6. 37. Cambridge Philos. New York. Proc. Appl. Stochastic Models 5. [215] J. [213] P. Scand. N.P. [228] J. Kemperman (1961) The Passage Problem for a Markov Chain. [224] J. [222] S. 239256.H. J. [220] V. Cambridge Philos. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes.
[247] L. Lipsky (1992) Queueing Theory .J. Appl.P. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. [236] C. Statist. 1995 . [241] G. Kunita (1976) Absolute continuity of Markov processes. Penev & A. [243] A.BIBLIOGRAPHY 375 [229] J. J. Probab. 26. SIAM. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. [248] D. 146. 48. Kingman (1964) A martingale inequality in the theory of queues. SpringerVerlag. Insurance: Mathematics and Economics 8. Lemoine (1981) On queues with periodic Poisson input. [233] C. . 383392. [239] H. Probab. Latouche & V. 60 . Lehtonen & H. Kingman (1962) On queues in heavy traffic.C. 277289. Macmillan. Math. 24. 125147. 359361. 132141. Sorensen (1997) Exponential Families of Stochastic Processes.a Linear Algebraic Approach. 279285. 4477. Philos. 6075. 4958. [245] T. Lecture Notes in Mathematics 511. 154168. Soc. Camb. Oxford 12. Klnppelberg & T. J. 858874. [246] D. Roy. Appl. [237] C. Act. (235] C. 889900. J. I.F.. Klnppelberg & T. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Proc. Bernoulli 1 . Soc. Kingman (1961) A convexity property of positive matrices . J. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities.J. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . J. [231] J. Philos. [230] J. 283284. Appl. 18.C. Proc. Kliippelberg & U.C. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Scand. Actuarial J. Cambr. Probab. B24. Seminaire de Probabilties X . Kliippelberg (1988) Subexponential distributions and integrated tails. Quart. [232] C. 25.F. Lehtonen & H. Kuchler & M. Quart. 133135 (in Russian). Soc. Korolyuk. 1998 . Cybernetica 4. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Adv.F. [244] T. [234] C.F. J. SpringerVerlag. Insurance : Mathematics and Economics 12.S. Scand. 390397. Stochastic Models 7. Act. Scand. 259264. New York. Lindley (1952) The theory of queues with a single server. J. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Mikosch (1995) Modelling delay in claim settlement. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. [238] V. [242] A. Prob. Appl. [240] U.
K.S. J. Ann. Appl. J. Hoist eds. F. Appl. Insurance : Mathematics and Economics 5. . Ann. [255] V. [254] V . 1994. [253] V. J. Appl. Uppsala. 32. [261] H. 58. Lundberg (1926) Forsdkringsteknisk Riskutjdmning.376 BIBLIOGRAPHY [249] D. 370377.D. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Scand. [260] H. 1261 (0?)1270. 124147. MartinL6f (1983) Entropy estimates for ruin probabilities . Soc. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . 1996 . Statist. [251] F. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. . 676705.D. Stoch. 2939. [259] Z . Probability and Mathematical Statistics (A. [266] P.K. Probab. 429448. Math. Miyazawa & V. Proc. 4859.V. J. Anal.M.V. Act. Appl. 763776. Proc. with an application . Soc. 36. Nagaev (1957) Some limit theorems for stationary Markov chains. Ney & E.D. Michna ( 1998) Selfsimilar processes in collective risk theory. Lucantoni . [257] A. Englunds Boktryckeri AB. Makowski ( 1994) On an elementary characterization of the increasing convex order . Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Scand. Probab. a useful concept in risk theory. MeierHellstern & M . Cambridge Philos. 58 . Act. [263] M . 834841. J. 147149. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Opns. Scand. [252] A. Schmidt ( 1993) On ladder height distributions of general risk processes . Stockholm. 286298.F.). [262] H . 268285. Probab. 11 . 378406. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . 31. Nummelin (1987) Markov additive processes I. 161174. 2. Act. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Cambridge Philos. Th. Math. Europ. Gut & J. 15. Eigenvalue properties and limit theorems . 3. 22. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Probab. J. [250] F. Ann. 1986 . Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. K. Meier ( 1984). [258] K . Appl. Probab.S. Probab . [256] A. 29 . Res. Appl. [264] G. II Aterforsdkring av Kollektivrisker . [265] S. Adv. MartinL3f (1986) Entropy. 223235. 561592. Almqvist & Wiksell. J.
Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Norberg ( 1990) Risk theory and its statistics environment . Omey & E.a survey. Omey & E. Paulsen & H. [270] M . [275] H. Neuts (1977) A versatile Markovian point process. 123144. 327361. J. 71. 733746. 36. Abh. Stoch. [272] J . Hamburg. [280] A. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. IEICE Trans. Neuts ( 1978) Renewal processes of phase type. New York. 46. Appl. Proc.F. Probab. [271] M. [276] C. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Paulsen (1998) Ruin theory with compounding assets . Johns Hopkins University Press. 215223. [278] E. [284] J. Naval Research Logistics Quarterly 25. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . J. 21. Appl. Sem. 555564. Baltimore . [274] H . Probab. Probab. Models 6. Appl. 16.F. J. Stoch. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 249266. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. [273] R.F. [285] J. [279] E. Stoch. 12551265. 12. 339353. 10081026. Pakes (1975) On the tail of waiting time distributions. 445454. 75. Proc. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Statistics 21. Astin Bulletin 5 . Stoch. . 30. Proc. Appl. [277] J. Appl. O'Cinneide (1990) Characterization of phasetype distributions.G. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Paulsen & H. [282] J .F. Commun. 764779. Adv. Math. 316. Neuts (1992) Models based on the Markovian arrival process. 135148. E75B . 273299. Appl.K. London. [283] J. Paulsen (1993) Risk theory in a stochastic economic environment. Stochastic Models 3 . 311342.K.F. 157. [268] M . Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Marcel Dekker.A. Appl. Appl. Insurance: Mathematics and Economics 20. [269] M .BIBLIOGRAPHY 377 [267] M . Stoch. Probab. [281] J. Nyrhinen (1999) Large deviations for the time of ruin. Proc. Insurance: Mathematics and Economics 22.
[297] R. Queues. Pitman (1980) Subexponential distribution functions.. 757764. R. Act. Regterschot & J.A.M. Heidelberg. 235243. Rolski (1987) Approximation of periodic queues. [293] N. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. 4568. Prabhu (1980) Stochastic Storage Processes. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Res. Res. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Adv. Opns. 2343. [306] T. and Dams. Soc. 4.U. Probab. 32. New York. Ann. 568576. Berlin. J. [288] S. 12. [289] C. Math. Ann. Pyke (1959) The supremum and infimum of the Poisson process. Math. Springer. Astin Bulletin XIV.K.U. 139143. [305] C.L. Pellerey (1995) On the preservation of some orderings of risks under convolution. Aktuar. . New York. Rao (1965 ) Linear Statistical Inference and Its Applications. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. 1989 . Philipson (1968) A review of the collective theory of risk. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Queueing Systems 5. 965985. Math. Probab. Statist. Scand. J. Embrechts (1996) Confidence bounds for the adjustment coefficient. Prabhu & Zhu (1989) Markovmodulated queueing systems. Math. 465483. Ripley (1987) Stochastic Simulation . Appl.J. Math. Tidskr. [303] S. [304] B. Statist. [295] N. J. 61. Insurance: Mathematics and Economics 3. Wiley. 29.H. Griibel & P. 820827. [300] C. 19. 222261. Paulsen & H. 29A. Gjessing (1997c) Ruin theory with stochastic return on investments.G.M. 691707. Inst.G. [291] S. 2330. [298] V. Adv. Ann. Wiley. [296] N. Probab . Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. [299] C. Appl. Ann. Probab.U. Austr.M. 537555. Insurance: Mathematics and Economics 16. Prabhu (1961). Appl. On the ruin problem of collective risk theory. 147159. [287] F. Insurance Risk. Prabhu (1965) Queues and Inventories. 215246. [294] N. Resnick & G.K. [301] G. 11.J. Pitts. Appl. Probab. 46. Oper. 28. 337347.U. Adv.R. 390413. Appl. 117133. [292] S. [290] E. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Wiley. Skand. [302] J. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance.378 BIBLIOGRAPHY [286] J. 30. Stat.
Wiley. J.J. Rubinstein (1981) Simulation and the Monte Carlo Method. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. 431447 [316] S. Appl. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Appi. 5. [319] H. Act. Rolski. 1997. [326] H. Ann. Wiley. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). [320] H. Schmidli. 795829. 262286. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. 135149. 145165. 93104. 11. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Data Anal. H. Operationsforsch.M. 155188. [308] S. 121133. Wiley. 21. Statist. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Insurance: Mathematics and Economics 16. Probab . Science Culture Technology Publishing . Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 5. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. [325] H . Astin Bulletin 29. 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. Stoch. Rubinstein & B. 365388. 687699. Stochastic Models 10 .Y. V. Statist. Proc. Rubinstein & A. Singapore. [315] T. [313] M. . 4857.Y. J. [322] H. Insurance: Mathematics and Economics 22. Scand. [310] R. 131156. [318] H. 2001 . [314] T. Probab. 227244. [311] R. Wiley. Teugels (1999) Stochastic Processes for Insurance and Finance. Schmidli ( 1999b) Perturbed risk processes: a review . Math. Schmidli ( 1996) Martingales and insurance risk. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Th. Stochastic Models 6. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method.Y. 4068. 5. Appl. [312] R. Adv. 7. Stochastic Models 10.L. [309] H. Wiley. Melamed (1998) Modern Simulation and Modelling. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [323] H. Rudemo (1973) Point processes generated by transitions of a Markov chain. J. Act. Probab.BIBLIOGRAPHY 379 [307] T. [324] H . Scand. Seal ( 1969) The Stochastic Theory of a Risk Business. [317] L. 417421. Schmidt & J. [321] H. Comp. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model .
1974. 171178. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Math. Shwartz & A. 429442. 61. Act. L. Verein Schweiz Versich. [333] C: O. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Chapman & Hall.380 BIBLIOGRAPHY [327] H. Math. 72.G.L. Stochastic Models 6. Ann. Seal (1974) The numerical calculation of U(w. t). Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. [331] G. Sengupta (1990) The semiMarkov queue: theory and applications. Insurance: Mathematics and Economics 8. Shin (1987) Convolution of uniform distributions and ruin probability. Probab. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. [330] H. Math. [338] E. Siegmund ( 1985) Sequential Analysis . Probab. 1955. 11. 701719. the probability of nonruin in an interval (0. 75. . t). [341] E. Versich. SpringerVerlag. In Probability and statistics . J.S. 11. Seber (1984) Multivariate Observations. Aktuar Tidsskr.W. Almqvist & Wiksell. Shaked & J. Segerdahl (1959) A survey of results in the collective theory of risk. pp. 121139. 2236. Mitt.F. New York. Probab . Ann. Scand. Academic Press. Grenander).L. [334] C: O. Verein Schweiz. Siegmund (1975) The time until ruin in collective risk theory. Stockholm. [332] C : O. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. [347] K. Adv. Seal ( 1972). Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Mitt. 157166. 72. 4383.S. [344] D.the Harald Cramer volume (ed. 191197. 159180. Skand. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . Risk teory and the single server queue . 243249. Appl. [343] D. Probab. Verein Schweiz. 4. Sigman (1992) Light traffic for workload in queues. 383413.A.L. Seal (1978) Survival Probabilities. Act. 279299. [339] A.S. Weiss (1995) Large Deviations for Performance Analysis. [346) D . [337] M. 1987. 21. [342] D. Appl. Statist. Wiley. J. Versich. [335] B . [348] K .W. Wiley. . Shantikumar (1993) Stochastic Orders and Their Applications. Appl. [345] D. Mitteil. 483487. 77100. [336] B. Th. 4. 673684. Scand. Adv. [328] H. [340] E. Shin (1989) Ruin probability by operational calculus. Aktuar Tidsskr. Queueing Systems 11. [329] H. 914924. Chapman and Hall. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case.
[366] O. Scand. J. Soc 49. Insurance: Mathematics and Economics 16. Thorin (1977). Unpublished manuscript. 163175.L. 65102. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes.S. 5776. Sundt & W. Insurance: Mathematics and Economics 1. 118. Act. Cambridge Philos. 65102. Adv.L. [359] L. 2739. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . J. New York. Astin Bulletin 12. 114137.). Hoesman ( 1989) Moderate. Verlag Versicherungswirtschaft e. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Straub (1988) NonLife Insurance Mathematics. Actuarial J. [353] E. Appl. 21. Slud & C. 1980 . Wiley. Taylor (1986) Claims reserving in NonLife Insurance. 722. [350] W. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . [360] G. Scand. 8199. [369] O. [354] B. Ruin probabilities prepared for numerical calculations.). Sundt & J. John Wiley & Sons. 235254. Scand. 197208. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed.L. J.. Astin Bulletin 24. Scand. [356] B . Act.. 149162. [352] D. 449461. Daley ed. NorthHolland. 1976.J. 1982 . [361] G. J. Springer. [357] B.C. [362] G.J. J. Thorin (1982) Probabilities of ruin. Act.and largedeviations probabilities in actuarial risk theory. [363] G. Proceeedings of the IEEE 77. Astin Bulletin 16 . Scand. Scand.V. Teugels (1982) Estimation of ruin probabilities . Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Teugels ( 1995 ) Ruin estimates under interest force .C.C. Smith ( 1953) Distribution of queueing times. [364] G. Act. 725741. . New York. [358] R. [351] D . Proc.BIBLIOGRAPHY 381 [349] E. Sundt & J. Act.A. 1978. [355] B . Insurance: Mathematics and Economics 19. Taylor (1979) Probability of ruin with variable premium rate. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Stanford & K. 1974.C.L. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Probab. [365] J. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. [368] O. Karlsruhe.C. [367] O. 8594. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes.
Van Wouve. 4963. M. [374] F. Willinger. De Vylder & M. 743756. 148164. Wolff (1990) Stochastic Modeling and the Theory of Queues.R. Skand. [377] V. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. [381] W. thesis. De Vylder. Thorisson (2000) Coupling. 142. Statistical Science 10. 37. . D. F. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Taqqu. Submitted.E. Probab. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. [376] A. Tacklind (1942) Sur le risque dans les jeux inequitables. University of Michigan. 31. [378] H. Dordrecht Boston Lancaster. Reidel. [373] S. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Scand. [382] G. M. 5762. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion.382 BIBLIOGRAPHY [370] 0. [371] 0.E. Unpublished Ph. R. Appl. Wald (1947) Sequential Analysis. 231246. 6785. [384] R. [372] H. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. Tidskr. Aktuar. 137153.W. Wiley. Stationarity and Regeneration. [383] G. Goovaerts. Sherman & D. J. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. [375] N. [380] W. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Astin Bulletin IX. Res.). Oper. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. PrenticeHall. 1942. Willmot & X. J. Insurance: Mathematics and Economics 15. Thorin & N. 1983 . Act. Lin (1994) Lundberg bounds on the tails of compound distributions. J. Astin Bulletin VII. Insurance : Mathematics and Economics 13. SpringerVerlag. Haezendonck eds. Thorin & N. In: Premium Calculation in Insurance (F. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. 936952. [379] M.
328330. 138139. 4851. 80 81.259261. 7079. 218 Cox process 4. 217.6779. 15.242. 360 excursion 155156.281.314316. 122. 278 gamma distribution 67. 17.318320 change of measure 2630. 162164. 239. 89. 196201 inverse Gaussian distribution 76. 17.249250 integral equation 16 Lindley 143 renewal 64. 3436. 7879. 318319 Erlang distribution 7. 323 Coxian distribution 147.100. 301 Kronecker product. 1415.307312 compound Poisson model 4.178184. 1819. 5. 341. 141144.185187. 91. 7179.135. 271274.and sum 221. 9396. 117127 corrected 121127 duality 1314.228229. 7475. 248 WienerHopf 144 interest rate 190.292293 Edgeworth expansion 113. 117128. 12 CramerLundberg approximation 1617.Index adjustment coefficient 17.226.160167. 3839. 308. 119.203. 2526.359 aggregate claims 103106. 79.251280 heavy traffic 76. 9496.9899.200201. 111117. 39. 226. 227229.293294. 302303 diffusion approximation 17.301 central limit theorem 60 . 14.249.272. 1112. 2425.285292. 205. 40. 3032. 207 heavytailed distribution 6.287292. 37.269. 332333 Volterra 192194. 316323 Bessel function 102. 5796. 170173.182. 8283 hyperexponential distribution 7. 9293.299. 283. 189. 97129. 361 diffusion 3.4447. 245248.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 97.150.203. 180182. 86. 135.137141. 110113.346349 383 . 201 Brownian motion 3 .86.121129. 201214. 3334.217.
234240. see also sensitivity analysis phasetype distribution 8.148. 38. 257.201.234.350361 Poisson process Markovmodulated 12 periodic 12.287291 INDEX matrix equation . 234 matrixexponential distribution 240244 matrixexponentials 14. 100.302. 52 53.160161.298299.349 350 perturbation 172173.288290.287. 9899.161.304 process 2830.178182. 44. 132133. 149. 108109. 261264. 35. 59. 42.340350 multiplicative functional 2830.161164. 227228. 16.128129.339 large deviations 129. 2730.134135. 37. 179 NP approximation 318320 Palm distribution 5253.336339 Laplace transform 15. 96. 3947. 267269 Panjer's recursion 320323 Pareto distribution 910. 203 Markov additive process 12. 44. 71. 112113. 7179. 14. 133. 3947. 269 PerronFrobenius theory 4142. nonlinear 155. 213214. 6162. 178 modulation 12. 4446. 16. 162.227230. 5758.348 terminating 215216. 3639. 38.146148. 99.285287 queue 14 . 25. 108 life insurance 5. 295. 32.336339 . 7576. 141144. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 271274.161. 176185.315 inequality 1718. 145187.261264.218221. 133. 86 periodicity 12.123. 142 likelihood ratio : see change of measure lognormal distribution 9. 134135. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 35. 137139. 41. 251. 138.180. 304305 random walk 3336. 113114. 171.259261.238.174.297299.152160. 15. 39.240244.275278. 203204.215250. 154.384 ladder heights 4756.139141. 6970. 25. 144. 65. 157.108. 245 M/G/1 13. 175 light traffic 8183 Lindley integral equation 143 process 3334. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 306316 Levy process 3. 80. 230. 134. 106108. 260 Lundberg conjugation 6979 .269271.
87. 7475. see also matrixexponential distribution regenerative process 264 268. 335336 sensitivity analysis 8693.273274. 1819. 146. 141144. 240. 332333 model 12.186. 327 . 281296 stable process 15. 177 timereversion 14. 3032. 292294. 162. 9693. 123. 280. 107. 338 utility 324. 222. 131144. 152. 120 statistics x. 160. 317318 semiMarkov 147. 172173. 251. 244. 233. 213. 331336 equation 64.262263. 261264 reservedependent premiums 14. 186187 renewal process 131. 251. 260 reinsurance 8. 257. 191192.154157. 5455.314. 186187 virtual: see workload rational Laplace transform 8. 307308. 223226.359361 stochastic control x stochastic ordering 18. 60. 8386. 229234. 238 saddlepoint method 115117. 174. 233234. 89.279280 Rouche roots 158. 260 WienerHopf theory 144. 168172 storage process 13. 12. 294296 shotnoise process 314 simulation 19. 31. 147. 333334 regular variation 10.INDEX 385 waiting time 141. 189214. 37.336339 workload 13. 253. 256258. 279280 subexponential distribution 11.244250. 11. 251280 time change 4. 4950. 326330 Weibull distribution 9.
Some i (l I JL I J r of the topics are Lundberg's inequality. Markovmodulation or periodicity.. P'i yfliother approximations (e.Advanced Series on Statistical Science & Applied Probability . y finite horizon ruin probabilities. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. the ^W A l \ i l ' ''' CramerLundberg approximation. for heavytailed claim size distributions).Vol." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. I 1! Ruin Probabilities .g. "This book is a must for anybody working in applied probability. It is a comprehensive treatment of the known results on ruin probabilities. 2 A I 11 JjVb l' i  i Yj .com 2779 he 9 "789810ll22293211 . worldscientific. extensions of the classical compound Poisson model to allow f o r reservedependent premiums..T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. exact solutions. Special features of the book are the emphasis on change of measure techniques.
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.