CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics
LARGE DEVIATIONS AND IDEMPOTENT PROBABILITY
...

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics

LARGE DEVIATIONS AND IDEMPOTENT PROBABILITY

© 2001 by Chapman & Hall/CRC

119

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics Main Editors H. Brezis, Université de Paris R.G. Douglas, Texas A&M University A. Jeffrey, University of Newcastle upon Tyne (Founding Editor)

Editorial Board H. Amann, University of Zürich R. Aris, University of Minnesota G.I. Barenblatt, University of Cambridge H. Begehr, Freie Universität Berlin P. Bullen, University of British Columbia R.J. Elliott, University of Alberta R.P. Gilbert, University of Delaware R. Glowinski, University of Houston D. Jerison, Massachusetts Institute of Technology K. Kirchgässner, Universität Stuttgart B. Lawson, State University of New York B. Moodie, University of Alberta S. Mori, Kyoto University L.E. Payne, Cornell University D.B. Pearson, University of Hull I. Raeburn, University of Newcastle G.F. Roach, University of Strathclyde I. Stakgold, University of Delaware W.A. Strauss, Brown University J. van der Hoek, University of Adelaide

© 2001 by Chapman & Hall/CRC

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics

LARGE DEVIATIONS AND IDEMPOTENT PROBABILITY

ANATOLII PUHALSKII

CHAPMAN & HALL/CRC Boca Raton London New York Washington, D.C.

© 2001 by Chapman & Hall/CRC

119

C0198_Disclaimer Page 1 Friday, March 30, 2001 2:22 PM

Library of Congress Cataloging-in-Publication Data Puhalskii, Anatolii. Large deviations and idempotent probability / Anatolii Puhalskii. p. cm. -- (Chapman & Hall/CRC monographs and surveys in pure and applied mathematices ; 119) Includes bibliographical references and index. ISBN 1-58488-198-4 (alk. paper) 1. Large deviations. 2. Probability measures. 3. Idempotents. I. Title. II. Series. QA273.67 .P83 2000 519.5′34--dc21

00-065883

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage or retrieval system, without prior permission in writing from the publisher. The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC for such copying. Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation, without intent to infringe.

Visit the CRC Press Web site at www.crcpress.com © 2001 by Chapman & Hall/CRC No claim to original U.S. Government works International Standard Book Number 1-58488-198-4 Library of Congress Card Number 00-065883 Printed in the United States of America 1 2 3 4 5 6 7 8 9 0 Printed on acid-free paper

© 2001 by Chapman & Hall/CRC

To my parents

© 2001 by Chapman & Hall/CRC

Contents Preface Basic notation

xi 1

I Idempotent Probability Theory

3

1 Idempotent probability measures 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9

Idempotent measures . . . . . . . . . . . . . Measurable functions . . . . . . . . . . . . . Modes of convergence . . . . . . . . . . . . Idempotent integration . . . . . . . . . . . . Product spaces . . . . . . . . . . . . . . . . Independence and conditioning . . . . . . . Idempotent measures on topological spaces Idempotent measures on projective limits . Topological spaces of idempotent probabilities . . . . . . . . . . . . . . . . . . 1.10 Derived weak convergence . . . . . . . . . . 1.11 Laplace-Fenchel transform . . . . . . . . . .

2 Maxingales 2.1 2.2 2.3 2.4

Idempotent stopping times . . . . . Idempotent processes . . . . . . . . . Exponential maxingales . . . . . . . Wiener and Poisson idempotent processes . . . . . . . . . . . . . . . 2.5 Idempotent stochastic integrals . . . 2.6 Idempotent Ito dierential equations vii

© 2001 by Chapman & Hall/CRC

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

5

5 14 17 20 32 36 51 58

. . . . . . 64 . . . . . . 79 . . . . . . 84

91

. . . . . . . . . . 91 . . . . . . . . . . 95 . . . . . . . . . . 104 . . . . . . . . . . 114 . . . . . . . . . . 124 . . . . . . . . . . 151

viii 2.7 Semimaxingales . . . . . . . . . . . . . . . . . . . . . 170 2.8 Maxingale problems . . . . . . . . . . . . . . . . . . . 202

II Large Deviation Convergence of Semimartingales

251

3 Large deviation convergence

253

4 The method of nite-dimensional distributions

289

3.1 Large deviation convergence in Tihonov spaces . . . . 253 3.2 Large deviation convergence in the Skorohod space . . . . . . . . . . . . . . . . . . . . 276 4.1 Convergence of stochastic exponentials 4.2 Convergence of characteristics . . . . . 4.2.1 The case of small jumps . . . . 4.2.2 The general case . . . . . . . . 4.3 Corollaries . . . . . . . . . . . . . . . . 4.4 Applications to partial-sum processes .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

5 The method of the maxingale problem

5.1 Convergence of stochastic exponentials . . . . . . . . . 5.1.1 Proofs . . . . . . . . . . . . . . . . . . . . . . . 5.2 Convergence of characteristics . . . . . . . . . . . . . . 5.2.1 Exponential tightness results . . . . . . . . . . 5.2.2 LD accumulation points as solutions to maxingale problems . . . . . . . . . . . . . . . . . . . 5.2.3 Proofs of the main results . . . . . . . . . . . . 5.3 Large deviation convergence results . . . . . . . . . . . 5.4 Large deviation convergence of Markov processes . . .

290 305 316 325 332 342

355

356 360 373 380 391 404 406 414

6 Large deviation convergence of queueing processes 433 6.1 Moderate deviations in queueing networks . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 Idempotent diusion approximation for single server queues . . . . . . . . . . . . . . . . . . . 6.1.2 Idempotent diusion approximation for queueing networks . . . . . . . . . . . . . . . . 6.2 Very large and moderate deviations for many server queues . . . . . . . . . . . . . . . . . . . . . . . . . . .

© 2001 by Chapman & Hall/CRC

433 433 442 450

ix

Appendix A Auxiliary lemmas Appendix B Notes and remarks Bibliography

© 2001 by Chapman & Hall/CRC

457 467 483

Preface This book has grown out of an approach to establishing the large deviation principle (LDP) for probability measures that originates from viewing the LDP as an analogue of weak convergence of probability measures and develops tools for proving it along the lines of the ones used in weak convergence theory. Let us recall that, given a Hausdor topological space E equipped with Borel -algebra B(E ), a function I : E ! [0; 1] such that the sets fz 2 E : I (z ) ag are compact for a 2 R+ , a net fP ; 2 g of probability measures on (E; B(E )), and a net of non-negative numbers fr ; 2 g such that r ! 1 as 2 , the net fP ; 2 g is said to obey the LDP with rate function I for scale r if 1 lim sup ln P (F ) inf I (z ) for F being a closed subset of E; z 2F 2 r 1 lim inf ln P (G) inf I (z ) for G being an open subset of E: z 2G 2 r The de nition being modelled after the de nition of weak convergence of probability measures, it is not surprising that there are similarities between methods of deriving the LDP and weak convergence, e.g., both theories make use of characteristic functionals, projective limit arguments, continuous mappings, characterisation of relative compactness in terms of certain tightness conditions, and others. Our purpose is to explore this analogy in more depth and systematically build on it for studying properties of the LDP. The rst important step is to recognise and treat the rate function as a limit case of the probability measure rather than merely as an asymptotic value. More precisely, we consider the set function (A) on E de ned by (A) = supz2A exp( I (z )) as an analogue and a limit of probabilities, so we call it \a deviability". We next look for properxi © 2001 by Chapman & Hall/CRC

xii ties of that are inherited from probabilities in the hope that this will help us to identify it, e.g., we are interested in an analogue of the martingale property. A distinctive feature of deviability is that it is \maxitive" in that (A [ B ) = (A) _ (B ). Maxitive set functions have been known as possibility measures in possibility theory and idempotent measures in idempotent measure theory (also referred to by the names \max-plus calculus" and \min-plus calculus"). We adopt the name \idempotent measure"; on the other hand, the notation , which is used not only for deviabilities but also for general idempotent measures \of mass 1", is borrowed from possibility theory. Developing \a stochastic calculus" for idempotent measures is the subject of part I of the book. We start with basic axioms, consider extension theorems, measurability issues, idempotent expectations and conditional idempotent expectations, topologies on spaces of idempotent measures, and other analogues of the constructions of probability theory. The axioms for an idempotent measure are mostly the same as the ones used in possibility theory and idempotent measure theory so we recover some of the results of these theories. Besides, we extensively analyse the -smoothness property of idempotent measures that requires a certain type of \continuity from above" and has been prompted by the fact that deviabilities are -smooth with respect to decreasing nets of closed sets. We also undertake a study in the spirit of the general theory of stochastic processes of idempotent analogues of stopping times, ltrations, stochastic processes, Ito dierential equations, martingales and semimartingales (which we call maxingales and semimaxingales, respectively), and martingale problems (referred to as maxingale problems). Being motivated by applications to large deviation theory, by no means do we consider analogues of all standard probability topics, the most notable omissions being analogues of the theory of limit theorems and theory of Markov processes. Our focus is on developing weak convergence theory for idempotent measures and those parts of \maxingale theory" that are instrumental in deriving large deviation limit theorems. Part II studies the large deviation setting. In order to emphasise the view of a deviability as a limit of probabilities, we refer to \the LDP for the P with rate function I " as \large deviation (LD) convergence of the P to ". Thus, the P are said to LD converge to

© 2001 by Chapman & Hall/CRC

xiii at rate r if lim sup P (F )1=r (F ) for F closed; (0.0.1) 2 lim inf P (G)1=r (G) for G open: (0.0.2) 2 In our study we actually use a dierent form of the de nition of LD convergence that states that the P LD converge to (at rate r ) if Z

lim 2

h(z )r dP (z )

E

1=r

= sup h(z )(z ) z 2E

(0.0.3)

for all R+ -valued bounded and continuous functions h on E . For general Hausdor spaces properties (0.0.1) and (0.0.2) are stronger than (0.0.3). One could draw the line by referring to (0.0.1) and (0.0.2) as narrow large deviation convergence and (0.0.3) as weak large deviation convergence. However, for Tihonov spaces (i.e., completely regular T1 -spaces), which we mostly consider and which seem to suÆce for concrete large deviation settings, the two forms are equivalent; so we refer to the property speci ed by (0.0.3) as large deviation convergence. The advantage of using de nition (0.0.3) is that many proofs can be signi cantly shortened (which fact is explained to some extent by both the limit and pre-limit objects being certain norms). We explore general properties of LD convergence in the form (0.0.3) in the rst section of part II, where our methods are similar to those of studying weak convergence of measures and idempotent measures. The rest of part II considers LD convergence of the distributions of semimartingales for the Skorohod topology on the space of right-continuous with left-hand limits Rd -valued functions on R+ . Here we are able to implement the approaches used for deriving convergence in distribution for semimartingales such as characterisations of limits in terms of their nite-dimensional distributions and as solutions to martingale problems. We interpret the limit deviabilities as distributions of idempotent processes and state the results in the form of LD convergence in distribution of semimartingales to semimaxingales. For example, we formulate the LDP for diusion processes with small diusion terms as LD convergence in distribution to an idempotent diusion. We give applications to LD convergence of Markov processes and processes arising in queueing systems. Our

© 2001 by Chapman & Hall/CRC

xiv results for queues are in the same theme as corresponding weak convergence results. As a byproduct, the results of part II show that possibility theory can be viewed as a large deviation limit of probability theory. The book concludes with two appendices. Appendix A proves certain auxiliary results invoked in the main body of the book. Appendix B contains additional comments on the results and bibliographical notes; the latter re ect the author's view of the related work and are by their very nature subjective, nor do we make any claim to completeness of the list of references.

© 2001 by Chapman & Hall/CRC

1

Basic notation R+ R+

a_b a+ a^b bac f Æg N Z+

xy jxj T kk 1=2 Ac AB P (A) Q(A) 1(A); 1A int A cl A B(E ) B(R+) B [0; t]

= [0; 1) = [0; 1] the maximum of a and b = a_0 the minimum of a and b the integer part of a the composition of functions f and g the set of natural numbers the set of non-negative integers the inner product of vectors x and y the Euclidean norm of a vector x the transpose of a matrix the operator norm of a matrix the pseudo inverse of a matrix the square root of a positive semi-de nite symmetric matrix the complement of a set A the collection of functions from a set B to a set A the power set of a set A the collection of nite subsets of a set A the indicator function of a set A the interior of a subset A of a topological space the closure of a subset A of a topological space the Borel -algebra on a topological space E the Lebesgue -algebra on R+ the Lebesgue -algebra on [0; t]

© 2001 by Chapman & Hall/CRC

Part I

Idempotent Probability Theory

3 © 2001 by Chapman & Hall/CRC

Chapter 1

Idempotent probability measures In this chapter we introduce idempotent analogues of basic objects of probability theory such as probability measures, random variables, expectations, conditional probabilities and expectations, and others, and study their properties.

1.1 Idempotent measures In this section we de ne the notion of an idempotent measure and obtain an extension theorem. We also introduce idempotent analogues of a measure space and probability space. Let be a set and E be a collection of subsets of , which contains ;. Let P ( ) denote the power set of . We reserve symbols and to denote directed sets, and J to denote arbitrary index sets.

De nition 1.1.1. A set function : P ( ) ! R + is an idempotent measure on if the following conditions hold:

(0) (;) = 0; (1) (A [ B ) = (A) _ (B ); (2) ([ A ) = sup (A ) for every increasing net fA ; 2 g of subsets of . If, in addition,

5 © 2001 by Chapman & Hall/CRC

6

Idempotent probability measures

() ( ) = 1, the idempotent measure is called an idempotent probability measure or idempotent probability, for short, and denoted by . If, in addition to (0), (1) and (2), (3) (\ F ) = inf (F ) for every decreasing net fF ; 2 g of elements of E , then we say that the idempotent measure is -smooth relative to E , or, for short, is an E -idempotent measure. Remark 1.1.2. Throughout, we use the terms \increasing" and \decreasing" as synonyms of \non-decreasing" and \non-increasing", respectively. Remark 1.1.3. Property (1) shows that is an increasing and subadditive set function in that (A) (B ) if A B and (A [ B ) (A) + (B ). The following characterisation of idempotent measures is a straightforward consequence of the de nition. Lemma 1.1.4. Conditions (1) and (2) are equivalent to the condition [ Aj = sup (Aj ) (1.1.1) j j 2J for every collection fAj ; j 2 J g of subsets of , which in turn is equivalent to the representation

(A) = sup (f!g); A : !2A

(1.1.2)

The function (f!g) is called the density of . We also refer to property (2) as -smoothness along increasing nets (it should not be confused with -smoothness, which concerns decreasing nets of elements of E ) and to property (1.1.1) as -maxitivity. For set functions that are only de ned on subsets of we use a similar terminology introduced by the following de nition. De nition 1.1.5. A set function : E ! R + is maxitive (respectively, -maxitive) on E if (A [ B ) = (A) _ (B ) for every A 2 E and B 2 E such that A [ B 2 E (respectively, [j 2J Aj = supj 2J (Aj ) for every collection of sets Aj 2 E ; j 2 J; such that [j2J Aj 2 E ).

© 2001 by Chapman & Hall/CRC

7

Idempotent measures

Given a collection E , we denote by Eu (respectively, Ei ) the collection of arbitrary unions (respectively, intersections) of elements of E . If nite unions (respectively, intersections) of sets from E belong to E , then we may and often do assume that the sets in an in nite union (respectively, intersection) of elements of E form an increasing (respectively, decreasing) net relative to a directed set. We also denote Eiu = (Ei )u and observe that it coincides with Eui = (Eu )i . The collection Eiu is clearly closed under the formation of arbitrary unions and intersections. We recall the following de nitions.

De nition 1.1.6. A collection E of subsets of is called a paving on

if it contains ; and is closed under the formation of nite unions and intersections. A collection E of subsets of is called a -system if it is closed under the formation of nite intersections. The next lemma shows that if is an idempotent measure that is -smooth relative to a paving E , then the values of on Eiu are uniquely speci ed by the values on E so that there is at most one extension of from E to Eiu .

Theorem 1.1.7. Let E be a -system containing ; and be an E idempotent measure. Then

(A) = sup (B ); A 2 Eiu ; B 2Ei : B A

(B ) = inf (F ); B 2 Ei : F 2E : F B

The proof follows by -maxitivity and -smoothness of , and the fact that each set in Ei is an intersection of a decreasing net of elements of E . The following simple fact is useful for extension theorems (see Theorem 1.1.9 below). We give the proof to show a typical argument.

Lemma 1.1.8. If an idempotent measure is -smooth relative to

a -system E containing ;, then it is -smooth relative to Ei . Proof. T Let fA ; 2 g be a decreasing net of elements of Ei , i.e., A = 2 F , where F 2 E . Let be the collection of nite sequences Æ = f( i1 i1 ); ( i2 i2 ); : : : ; ( ik ik )g, where ij 2 , i1 i2 : : : ik and l 2 l for l = i1 ; i2 ; : : : ; ik . We say

© 2001 by Chapman & Hall/CRC

8

Idempotent probability measures

that Æ Æ0 if all the pairs ( ) that appear in Æ are also conT 0 tained in Æ . For Æ 2 , let BÆ = ( )2Æ F . Then is a directed set and fBÆ ; Æ 2 g is a decreasing net. Also BÆ 2 E and T T A = 2 Æ2 BÆ ; so, since is an E -idempotent measure, \

2

A

= inf (BÆ ): Æ2

(1.1.3)

For arbitrary Æ = f( i1 i1 ); ( i2 i2 ); : : : ; ( ik ik )g, let Æ ij ; j = 1; : : : ; k. Then, since fA ; 2 g is a decreasing net, it follows that BÆ A Æ ; hence,

(BÆ ) A

Æ

inf (A ) = inf (A ): 2 Æ

Thus, in view of (1.1.3),

\

2

A

inf (A ): 2

We consider now the issue of extending set functions to idempotent measures. Theorem 1.1.9. Let E be a paving on . Let be an R + -valued maxitive function on E such that (;) = 0. 1. The set function can be extended to an idempotent measure on if and only if it is -smooth along increasing nets, i.e., for every increasing net fF g of elements of E whose union belongs to E we have

[

F = sup (F ):

The extension is uniquely speci ed on Eu .

2. The set function can be extended to an E -idempotent measure if and only if the following condition holds.

(S ) If fF1; g and fF2; g are respective increasing and decreasing nets of elements of E such that [

F1;

© 2001 by Chapman & Hall/CRC

\

F2; ;

Idempotent measures

9

then

sup (F1; ) inf (F2; ):

The idempotent measure is then also -smooth relative to Ei and is uniquely speci ed on Eiu . Proof. We rst consider part 1. Necessity of the condition is obvious. We prove suÆciency. We rst note that in view of maxitivity of and the fact that E is closed under the formation of nite unions the condition of -smoothness along increasing nets implies -maxitivity of on E . For ! 2 , let (f!g) = inf (F ); (1.1.4) F 2E : !2F

and for A let (A) = sup (f!g): (1.1.5) !2A Clearly, (A) is an idempotent measure. We prove that agrees with on E . Let F 2 E . By (1.1.4) (f!g) (F ) if ! 2 F , so by (1.1.5) (F ) (F ). Conversely, given " > 0, let F ! 2 E for ! 2 F be such that ! 2 F ! and (f!g) S(F ! ) ". Then by (1.1.5) (F ) sup!2F (F ! ) . Since F = !2F (F \ F ! ), where F \ F ! 2 E by the fact that E is closed under the formation of nite intersections, and is -maxitive and increasing on E , (F ) = sup (F \ F ! ) sup (F ! ) (F )+ ": !2F !2F Part 1 is proved. We prove part 2. It is obvious that if there exists an idempotent measure , which is -smooth relative to E and coincides with on E , then condition (S ) holds. For the converse, we note that condition (S ) implies the condition of -smoothness of relative to increasing nets of elements of E in part 1. Therefore, by part 1 the set function de ned by (1.1.4) and (1.1.5) is an idempotent measure, which extends . We prove that is an Ei -idempotent measure. Note that since is maxitive on E and E is closed under the formation of nite unions, condition (S ) extends to the case where fF1; g is an

© 2001 by Chapman & Hall/CRC

10

Idempotent probability measures

arbitrary collection of elements of E . Next, by Lemma 1.1.8 it suÆces to check (3) for decreasing nets of elements of E . Let F # F , where F 2 E . Given " > 0, we choose for every ! 2 F sets F ! 2 E as in the proof of suÆciency in part 1. Since [!2F F ! \ F and condition (S ) extends to arbitrary collections fF1;j g of elements of E , we conclude that sup!2F ! (F ! ) inf (F ) so that (F ) sup!2F (F ! ) " inf (F ) ", which completes the proof. The fact that is unique on Eiu follows by Theorem 1.1.7.

Remark 1.1.10. If E is a ring, i.e., closed under the formation of dierences, then condition (S ) is equivalent to continuity of at 0: if F # ;, then (F ) # 0. Theorem 1.1.9 is then an analogue of Caratheodory's theorem, see, e.g., Halmos [58].

Remark 1.1.11. Wang and Klir [133, Theorem 4.9] prove an ex-

tension theorem in the theme of part 1 for the case where the collection E is not necessarily closed under the formation of nite unions and intersections. Then the requirements on of maxitivity and -smoothness along increasing nets are replaced by the following P consistency condition: if a collection fFj g of elements of E and F 2 E are such that F [j Fj , then (F ) supj (Fj ). Similarly, part 2 admits a version for collections E that have the only property of including the empty set. Condition (S ) then has to be replaced by the following condition: (S 0 ) If fF1;j g is a collection of elements of E and fF2; g is a decreasing net of elements of E such that [

j

F1;j

\

F2; ;

then

sup (F1;j ) inf (F2; ): j

Condition (S 0 ) is necessary and suÆcient for to be extended to an E -idempotent measure. If E is a -system, then by Lemma 1.1.8 the extension is -smooth relative to Ei . We note also that condition (S 0 ) implies the P-consistency condition.

© 2001 by Chapman & Hall/CRC

11

Idempotent measures

Finally, if in part 2 we only omit the requirement that E be a -system, then the extension also exists and is an E -idempotent measure.

We will be interested in more special collections of subsets of

than pavings and -systems. The following notion plays a central part in our analysis below.

De nition 1.1.12. A collection A of subsets of is a -algebra if it contains ; and is closed under the formation of complements and arbitrary unions. The elements of A are referred to as A-measurable subsets of .

The power set P ( ) is obviously a -algebra, we refer to it as the discrete -algebra.

De nition 1.1.13. A collection E of subsets of is said to be atomic 0

if it has a subcollection E = fA g, consisting of non-empty subsets of , such that either A \ A0 = ; or A = A0 for every and 0 , and F 2 E if and only if F = [A , where the union is taken over A F; A 2 E 0 . The elements of E 0 are called the atoms of E .

The structure of -algebras is revealed by the next theorem, which follows from the de nition.

Theorem 1.1.14.

A collection A, which contains -algebra if and only if it is atomic.

;

and , is a

We denote as [!]A the atom of a -algebra A that contains ! 2 . We note that A 2 A if and only if A = [!2A [!]A , where an empty union is assumed to be empty. Remark 1.1.15. The relation R on de ned by (!; !0 ) 2 R if !0 and ! belong to the same atom of a -algebra A is obviously an A ! if we want equivalence relation. We denote this by !0 ! (or !0 to emphasise the -algebra to which the equivalence relation refers). Note that !0 ! if and only if !0 2 [!]A if and only if ! 2 [!0 ]A if and only if [!]A = [!0 ]A . The following simple observation is frequently used below.

Corollary 1.1.16. A set A is an element of a -algebra A on

if and only if [!]A A for every ! 2 A.

© 2001 by Chapman & Hall/CRC

12

Idempotent probability measures

De nition 1.1.17.0 We say that a -algebra A0 is a sub- -algebra of a -algebra A if A

A.

Lemma 1.1.18. A -algebra0 A0 is a sub- -algebra of a -algebra A if and only if the atoms of A are unions of the atoms of A.

We refer to the smallest -algebra containing a collection E as the -algebra generated by E and denote it as (E ). It is obviously unambiguously de ned.

De nition 1.1.19. We say that a collection E of subsets of is a c semi- -algebra if it includes ;, is a -system, and F F 2 E.

2 Eiu for every

The structure of semi- -algebras is similar to that of -algebras.

Lemma 1.1.20. A -system E , which includes ;, is a semi- -algebra

if and only if Ei is atomic and the union of the atoms of Ei equals . Proof. It is obvious that if Ei is atomic and its atoms make up , then E is a semi- -algebra. For the converse, given ! 2 , we take as the atom about ! the intersection of all elements of E that contain !.

The preceding proof also proves the following lemma.

Lemma 1.1.21. If E is a semi- -algebra, then (E ) = Eiu. Theorem 1.1.9 and Remark 1.1.11 yield the following fact.

Corollary 1.1.22. Let be a set function on a semi- -algebra E 0

such that (;) = 0. If condition (S ) holds, then has a unique extension to an Ei -idempotent measure on the -algebra generated by E. Proof. We only need to extend to a maxitive set function on the collection of nite unions of elements of E by setting

(

k [ i=1

Fi ) = max (Fi ) i=1;:::;k

(1.1.6)

and apply part 2 of Theorem 1.1.9. The fact that the extension (1.1.6) is unambiguously de ned follows from condition (S 0 ).

© 2001 by Chapman & Hall/CRC

Idempotent measures

13

The notion of a -algebra is obviously an analogue of the notion of a -algebra. The next de nition paraphrases the de nition of complete -algebras.

De nition 1.1.23. We say that a -algebra A is complete with re-

spect to an idempotent measure on (or -complete, for short) if every ! 2 such that (f!g) = 0 is an atom of A.

De nition 1.1.24. We call the completion of a -algebra A with respect to idempotent measure the -algebra that has as its atoms all the elements of idempotent measure 0 and the atoms of A without the elements of idempotent measure zero. We denote the completion by A .

Remark 1.1.25. Clearly, A is the smallest complete -algebra containing A.

De nition 1.1.26. A set with a -algebra of subsets of is called a -measurable space and is denoted as ( ; A).

We now de ne an analogue of a measure space.

De nition 1.1.27. A triplet ( ; A; ), where is a set, A is a -

algebra of subsets of and is an idempotent measure on , is called an idempotent measure space. We denote the idempotent measure space ( ; P ( ); ) as ( ; ). If is an idempotent probability, we refer to ( ; A; ) as an idempotent probability space.

De nition 1.1.28. Given an idempotent measure and a -algebra

A on the set function A de ned by A(A) = (A); A 2 A; is called the restriction of to A.

Remark 1.1.29. As we will see, it is often the case that an idempo-

tent measure is originally speci ed on a -algebra. Though by Theorem 1.1.9 it can always be extended to an idempotent measure on P ( ), this extension might not be unique, which justi es restricting our consideration to the elements of A. To emphasise this we refer to as an idempotent measure on ( ; A). We note, however, that ambiguity in extending does not necessarily lead to ambiguity in the end results. In the sequel, we use the same symbol to denote some extension of from A to P ( ). On the other hand, given an idempotent measure space ( ; A; ), where is uniquely speci ed on A, we could reduce it to a space

© 2001 by Chapman & Hall/CRC

14

Idempotent probability measures

with the discrete -algebra by introducing the factor-space of with respect to the equivalence relation speci ed by the atoms of A. In this sense considering arbitrary -algebras does not give anything new. However, it comes in useful if we need to deal with a collection of algebras on the same set as in Chapter 2, where -algebras are used to keep track of \the history of a process".

1.2 Measurable functions In this section we introduce measurable maps of spaces with idempotent measures. Let and 0 be sets, and E and E 0 be respective collections of subsets of and 0 , both containing ;. De nition 1.2.1. For a function f : ! 0, we de ne 1the 0 collection of subsets of generated by f as the collection f (E ) = ff 1(B ); B E 0g. We also refer to functions de ned on idempotent probability spaces as idempotent variables. The following lemma is a consequence of the de nition. Lemma 1.2.2. If E 0 is a -algebra (respectively, a -system, a 0 paving, a semi- -algebra) on , then f 1(E 0 ) is a -algebra (respectively, a -system, a paving, a semi- -algebra). The collection of the atoms of the -algebra f 1 (E 0 ) is the collection ff 1 (A0 )g, where A0 are the atoms of the -algebra E 0 . De nition 1.2.3. A function f : ! 0 is said to be E =E 0 1 0 measurable if f (E ) E . The following result is obvious. Lemma 1.2.4. A0 function f : ! 0 is Eiu=Eiu0 -measurable if and only if it is Eiu =E -measurable. As a consequence, a function f : ! 0 is A=A0 -measurable, where A and A0 are -algebras on respective sets and 0 , if and only if the inverse images of the atoms of A0 belong to A. Thus, we have the following. Corollary 1.2.5. A function f : ! 0 is A=A0-measurable if and0 only if every atom of A is mapped into a subset of some atom of A . In particular, f is A=P ( 0 )-measurable if and only if it is constant on the atoms of A.

© 2001 by Chapman & Hall/CRC

Measurable functions

15

In the sequel, we refer to A=P ( 0 )-measurable functions as Ameasurable functions or idempotent variables on ( ; A). Note that A is A-measurable if and only if 1(A) : ! R+ is Ameasurable.

Lemma 1.2.6. Let ( ; A; ) be an idempotent measure space and 0 A

be the completion of A with respect to . If f : ! is A measurable, then there exists an A-measurable idempotent variable f 0 such that f 0 = f -a.e. Proof. Let [!]A be an atom of A. We de ne f 0(!0 ) = f (~!) for all !0 2 [!]A , where !~ 2 [!]A is such that (~!) > 0 if ([!]A ) > 0 and !~ is an arbitrary element of [!]A otherwise. Then f 0 is A-measurable and f 0 = f -a.e. by the construction of A (see De nition 1.1.24).

The next lemma is a version of Doob's result.

Lemma 1.2.7. Let a -algebra A on be generated by a function 0 0 0 f : ! , where is equipped with a -algebra A . A function g :

! 00 is A-measurable if and only if there exists an A0 -measurable function h : 0 ! 00 such that g = h Æ f .

Proof. SuÆciency of the condition is obvious. We prove the necessity. Since A is generated by f and g is A-measurable, for arbitrary !00 2

00 there exists A0!00 2 A0 such that g 1 (!00 ) = f 1 (A0!00 ). Since the sets g 1 (!00 ) are disjoint, the sets A0!00 ; !00 2 00 ; are also disjoint. Therefore, letting h(!0 ) = !00 for !00 2 00 and !0 2 A0!00 , and h(!0 ) = !^ for !0 2 [!00 2 00 A0!00 c , where !^ is a xed element of 00 , de nes h unambiguously. Clearly, h Æ f (!) = g(!); ! 2 .

We have the following corollary for functions assuming values on the real line.

Corollary 1.2.8. Let A be a -algebra on

. If functions fj : ! J j 2 J; are A-measurable and F : R ! R, then F ((fj )j 2J ) is A-measurable. In particular, supj fj and inf j fj are A-measurable, R;

and if is a directed set, then lim sup2 f and lim inf 2 f are A-measurable.

We now consider images of idempotent measures under mappings. Let be an idempotent measure on . The next lemma is straightforward.

© 2001 by Chapman & Hall/CRC

16

Idempotent probability measures

Lemma 1.2.9. Let f : ! 0 . 0 0 1 0

Then the set function 0 on 0 de ned by (A ) = f (A ) for A0 0 is an idempotent measure on 0 . De nition 1.2.10. The set function 0 as de ned 1in the lemma is called the image of under f and denoted by Æ f .

For the image of a -smooth idempotent measure to be a smooth idempotent measure, we need to impose conditions on the mapping. By Luzin's theorem in measure theory a real-valued function of a real argument is Borel-measurable if and only if it is continuous on \large" sets (closed or compact). We turn the theorem into the de nition of a measurability concept. The rst step is to introduce an abstract analogue of the concept of a tight measure.

De nition 1.2.11. Let be an E -idempotent measure. We say that

a collection T of subsets of is tightening for if T \ F 2 E for T 2 T and F 2 E , and for arbitrary " > 0 there exists T 2 T such that (T c ) ". We then also say that is tight relative to T , or T -tight, for short.

We next de ne \Luzin measurability".

De nition 1.2.12.

Let T be a tightening collection for an E idempotent measure . A function f : ! 0 is called Luzin (E ; T )=E 0 -measurable if the restriction of f to an arbitrary T 2 T T is ET =E 0 {measurable, where ET = fT F; F 2 Eg. Remark 1.2.13.1 Equivalently, f : ! 0 is Luzin (E ; T )=E 0 { T 0 measurable if f (F ) T 2 E for every F 0 2 E 0 and T 2 T . Note also that E =E 0 -measurability implies Luzin (E ; T )=E 0 -measurability. On the other hand, the collection T = f g is trivially a tightening collection for so that E =E 0 -measurability is a speci c case of Luzin (E ; T )=E 0 -measurability. We refer to Luzin (E ; T )=E 0 -measurable functions as Luzin measurable functions if the collections E , T and E 0 are understood. The purpose of introducing the concept of Luzin measurability is seen from the following theorem.

Theorem 1.2.14. Let be an E -idempotent measure and T 0 tightening collection for . If a function f :

© 2001 by Chapman & Hall/CRC

!

be a

is Luzin

17

Modes of convergence

(E ; T )=E 0 -measurable, then the image 0 = Æ f 1 of under f is a -smooth relative to E 0 idempotent measure on 0. If f (T ) \ F 0 2 E 0 whenever T 2 T and F 0 2 E 0 , then 0 is f (T )-tight. Proof. By Lemma 1.2.9 0 is an idempotent measure. We check property (3). Let F0 2 E 0 be a decreasing net. Given " > 0, let T 2 T be such that (T c) < ". Then

0 (F0 ) = (f 1 (F0 )) f 1(F0 )

\

T + ":

1 (F 0 ) T T decrease as well. Therefore, Since the F0 decrease, the f T since f 1(F0 ) T 2 E , by -smoothness of

f 1(F0 )

\

T

\

!

f 1 (F0 )

\

T

0

\

F0

proving (3). The fact that the idempotent measure 0 is f (T )-tight provided f (T ) \ F 0 2 E0 whenever T 2 T and F 0 2 E 0 follows since is T -tight and 0 f (T )c = (f 1 (f (T )))c (T c): We x some more terminology.

De nition 1.2.15. Let ( ; ) be an idempotent probability space 0 0

and f : ! be an -valued idempotent variable. The idempotent probability Æ f 1 is called the idempotent distribution (or idempotent law) of f under . If 0 is a metric space and limr!1 (f 62 Br (z )) = 0, where z is a xed element of E and Br (z ) denotes the closed r-ball about z , then f is called a proper idempotent variable.

1.3 Modes of convergence We consider idempotent analogues of convergence in measure and convergence almost everywhere. Let ( ; ) be an idempotent measure space. Let f and f denote idempotent variables on with values in a metric space E with metric .

De nition 1.3.1. We say that a net ff; 2 g converges -a.e. to f if

! 2 : f (!) 6! f (!) = 0:

© 2001 by Chapman & Hall/CRC

18

Idempotent probability measures

More generally, we say that a property concerning elements of

holds -a.e. (or a.e. if is understood) if the idempotent measure of the set where the property does not hold equals 0.

De nition 1.3.2. We say that a net ff; 2 g converges to f in idempotent measure (or in idempotent measure if is understood) if for every " > 0

lim (! 2 : (f (!); f (!)) > ") = 0:

2

Note that since ((f; g) > ") ((f; f ) > "=2) + ((g; f ) > "=2), the limit in idempotent measure is unique -a.e. The same fact is of course true for convergence -a.e. We denote convergence in idempotent measure by f ! f.

Lemma 1.3.3. (\Borel-Cantelli") Let fA; 2 g be a net of subsets of . If (A ) ! 0, then lim sup2 A = 0. Proof. The claim follows since

\ [

lim sup A = 2

0 2 0

A

inf 0 2 sup0 (A ):

Theorem 1.3.4.

(\Egorov") f ! f if and only if for every " > 0 there exists a set A" such that (Ac" ) " and sup!2A" (f (!); f (!)) ! 0.

Proof. If ff ; 2 g is a net such that f ! f , then for every Æ > 0 there exists Æ 2 such that ((f ; f ) > Æ) < " for all Æ ; hence, (f ; f ) Æ on the set A" = f! 2 : (f!g) "g. The converse is obvious.

Theorem 1.3.5. If f ! f , then f ! f -a.e.

Proof. By Lemma 1.3.3 lim sup (f ; f ) > " = 0.

The next result gives a partial converse.

Theorem 1.3.6. If f ! f; 2 ; -a.e., then there exists a net fh ; 2 g,

which converges to f in idempotent measure and is

© 2001 by Chapman & Hall/CRC

19

Modes of convergence

such that fh (!); 2 g is a subnet of ff (!); 2 g for every ! 2 . If ffn ; n 2 N g is a sequence such that fn ! f -a.e., then for every ! 2 there exists a subsequence kn (!) such that kn (!) n and fkn ! f . Proof. Let = f(; ") : 2 ; 2 R+ g. We turn into a directed set by de ning that (; ") (0 ; "0 ) if 0 ; " "0 . For = (; ), we de ne (!) as 0 such that (f~(!); f (!)) for all ~ 0 if such a 0 exists and (!) = otherwise. Let h (!) = f (!) (!): Clearly, fh (!); 2 g is a subnet of ff (!); 2 g for every ! 2 . Since f ! f -a.e., 0 exists for almost every ! 2 ; hence, (h ~ (!); f (!)) for almost every ! 2 if ~ . In the case of sequences, we de ne kn (!) = minfl n : (fl (!); f (!)) 1=ng and kn (!) = n if no such l exists.

Remark 1.3.7.

Generally speaking, convergence -a.e. does not imply convergence in idempotent measure. Consider the following example. Let = [0; 1] and (f!g) = 1 for ! 2 [0; 1]. Let fn(!) = n!; ! 2 [0; 1=n], fn (!) = 2 n!; ! 2 [1=n; 2=n] and fn (!) = 0 elsewhere. Then fn(!) ! 0 as n ! 1 for every ! 2 [0; 1]. However, (! : fn(!) = 1) = 1.

We now consider Cauchy nets.

Lemma 1.3.8. If ((f; f0 ) > ") ! 0 as ; 0 2 for every " > 0 and (E; ) is complete, then ff g converges -a.e.

Proof. In analogy with the proof of Lemma 1.3.3, for " > 0, \ [

0

(f ; f0 ) > "

inf sup (f ; f0 ) > " = 0: 0

Thus, the net ff g is Cauchy -a.e., so it converges -a.e. by completeness of (E; ).

Theorem 1.3.9. If ((f; f0 ) > ") ! 0 as ; 0 2 and (E; ) is complete, then ff g converges in idempotent measure.

© 2001 by Chapman & Hall/CRC

20

Idempotent probability measures

Proof. By Lemma 1.3.8 f ! f -a.e. Let us choose as in the proof of Theorem 1.3.6. Then, for > 0, taking = (; =2) and using the inequalities (f ; f ) "=2 -a.e. and (!) ,

((f ; f ) > ") ((f ; f ) > "=2) _ ((f ; f ) > "=2) ( sup (f; f0 ) > "=2) = sup ((f; f0 ) > "=2): 0

0

The latter term goes to 0 as 2 by hypotheses. The last result of the section shows that for -smooth idempotent measures and decreasing nets of R+ -valued Luzin measurable functions convergence -almost everywhere implies convergence in idempotent measure. Let U denote the paving on R+ consisting of intervals [a; 1), where a 2 R+ , and ;. Theorem 1.3.10. Let be -smooth relative to a collection E and let T be a tightening collection for . If f; 2 g is a decreasing net of Luzin (E ; T )=U -measurable R+ -valued functions, which converges -a.e. to 0, then f ; 2 g converges to 0 in idempotent measure. Proof. Let Æ > 0, > 0 and T 2 T such that (T c) ". Then Æ f Æg \ T + ": Since f Æg \ T 2 E, by -smoothness of relative to E we have that f Æg \ T ! \ f Æg \ T = 0.

Remark 1.3.11. The result is a counterpart of the fact in probability

theory stating that a monotonic sequence of non-negative random variables that converges to zero in probability also converges to zero almost surely.

1.4 Idempotent integration In this section we develop an idempotent analogue of integration theory. Let ( ; ) be an idempotent measure space such that ( ) < 1. We adopt the convention that 1 0 = 0. De nition 1.4.1. For a function f on with values in R + we de ne the idempotent integral of f with respect to by _

f d = sup a (f a): a2R+

W

For A , we let A f d =

© 2001 by Chapman & Hall/CRC

W

f

1(A) d.

21

Idempotent integration

Idempotent integral is called idempotent expectation if is an idempotentWprobability. In the sequel we also denote idempotent integrals as f (!) d(!) and, if is an idempotent probability, as Sf , Sf (!), or S f , the latter notation is used to emphasise the idempotent probability for which the idempotent integral is evaluated. The next lemma follows by de nition.

Lemma 1.4.2. Let f : ! R +. The following equivalent representations hold. _

f d = sup a(f = a) = sup f (!)(f!g) !2

a2R+

= sup f (!)([!]f !2

1 (P (R+ )) ):

Remark 1.4.3. If is originally de ned on a -algebra A, then the

value of the idempotent integral of a function f depends generally speaking on what extensionW of to P ( ) we consider. However, if f is A-measurable, then f d is de ned unambiguously, which follows by the last equality in Lemma 1.4.2.WTo emphasise this we W sometimes denote the integral as f dA , f (!) dA (!) and, if is an idempotent probability, as SA f and SA f (!), where A and A denote the respective restrictions of and to A. A careful examination of the proofs below shows that if we require that the functions and sets considered in the statements be A-measurable, then the results are insensitive to the particular extension of to P ( ).

The results below whose proofs are omitted directly follow from Lemma 1.4.2. We consider only R+ -valued integrands, the corresponding properties for R + -valued integrands are derived similarly.

Theorem 1.4.4. Let f; g be R+ -valued functions on . The following properties hold. (JS0) (JS1)

_

_

0 d = 0

f d

_

g d if f

© 2001 by Chapman & Hall/CRC

g

22

Idempotent probability measures

(JS2) (JS3) (JS4) (JS5)

_

(cf ) d = c

_

f d; c 2 R+

_

_

_

_

(f _ g) d =

(f + g) d

_

_

j f d

(JS6)

_

f d +

g dj

well de ned

sup fj d = sup j 2J j 2J

_

f d _

_

_

g d

_

g d

jf gj d provided the left-hand side is

fj d, where fj : ! R+ ; j 2 J

The following Chebyshev-type inequality plays as important a role below as its counterpart does in probability theory.

Lemma 1.4.5. If f : ! R+ , then 1_ (f a) f 1(f a) d; a

a > 0:

We also have an analogue of the change-of-variables formula in the Lebesgue integral. Theorem 1.4.6. 1Let 0 be an idempotent measure on a set 0 such 0 0 that = Æ f for some f : ! . Then, for a function g : 0 ! R+ , _

0

g d0 =

_

g Æf d:

The following \Holder" inequalities are also useful. For f : ! 1=p W p R + and p > 0 we de ne kf k p = and kf k 1 =

f d sup!: (f!g)>0 f (!).

Lemma 1.4.7. Let f; g : ! R+ .

1. Let p 2 [1; 1] and q 2 [1; 1] be such that 1=p + 1=q = 1. Then W fg d kf k p kgk q .

2. If ( ) = 1, then, for 0 < p < q, kf k p kf k q .

© 2001 by Chapman & Hall/CRC

23

Idempotent integration

We are interested in convergence properties of idempotent integrals, so we study an analogue of the concept of uniform integrability.

De nition 1.4.8. We say that a function f : ! R+ is maximable (orW-maximable if the idempotent measure needs to be emphasised) W if f d < 1 and, moreover, f 1(f > a) d ! 0 as a ! 1. The following version of La Vallee-Poussin's theorem holds.

Theorem 1.4.9. A function f : ! R+ is maximable if and only if there exists a monotonically increasing function F : W that F (x)=x ! 1 as x ! 1 and F Æ f d < 1.

R+

! R+

such

Proof. We prove that the condition is suÆcient for f to be maximable. Given " > 0, let a > 0 be such that x=F (x) " for x a. Then _

f 1(f > a) d "

_

F Æf 1(f > a) d "

_

F Æf d:

Conversely, let f be maximable. Since there is no loss of generality in assuming that kf k 1 = 1, we de ne x F (x) = W :

f 1(f x) d Then F is monotonic and F (x)=x ! 1 as x ! 1 by maximability W of f . Also, F Æ f d 1.

De nition 1.4.10. A collection ffj ; j 2 J g of R+ -valued functions on is said to be uniformly maximable (or -uniformly maximable) if

sup j 2J

_

fj 1(fj > a) d ! 0 as a ! 1:

Theorem 1.4.11. A collection ffj ; j 2 J g is uniformly maximable if and only if the following conditions hold: (i) sup j 2J

_

fj d < 1, W

(ii) for every " > 0 there exists > 0 such that supj 2J A fj d < " for every set A such that (A) < .

© 2001 by Chapman & Hall/CRC

24

Idempotent probability measures

Proof. Let ffj ; j 2 J g be uniformly maximable. Then (i) and (ii) follow by the inequality _

A

fj d

_

fj 1(fj > a) d + a(A); a > 0:

The converse follows by the fact that since for all j 2 J and a large enough

(fj > a)

W

fj d

< ; a W where is chosen as in condition (ii), we have fj 1(fj > a) d < "; j 2 J:

Corollary 1.4.12. Let f : ! R+ be maximable. Then the set W function A ! f 1(A) d; A ; is absolutely continuous with respect to in the sense that for every " > 0 there exists Æ > 0 such W that f 1(A) d < " for all A such that (A) < Æ. Theorem 1.4.13. A collection ffj ; j 2 J g of R+ -valued functions on is uniformly maximable if and only if supj 2J fj is maximable.

Proof. SuÆciency of the condition is obvious. Conversely, let ffj ; j 2 J g be uniformly maximable. For > 0, let > 0 be chosen as in part (ii) of Theorem 1.4.11. By \the Chebyshev inequality" (supj fj W a) supj fj d=a, the latter supremum being less than if a is large enough in view of condition (i) of Theorem 1.4.11. Then by W Theorem 1.4.4 and the choice of we have sup j fj 1(supj fj

W 0 a)d = supj fj 1(supj 0 fj a)d .

The following analogue of La Vallee-Poussin's theorem is a simple consequence of Theorem 1.4.9 and Theorem 1.4.13.

Corollary 1.4.14. A collection ffj ; j 2 J g of R+ -valued functions on is uniformly maximable if and only if there exists a monotonically increasing function F : R+ ! R+ such that F (x)=x ! 1 as W x ! 1 and supj 2J F Æ fj d < 1.

The next easy corollary gives simple conditions for uniform maximability.

Corollary 1.4.15. A collection ffj ; j 2 J g is uniformly maximable if either one of the following conditions holds:

© 2001 by Chapman & Hall/CRC

25

Idempotent integration

1. fj f; j 2 J; where f is maximable, 2. sup j 2J 3. sup j 2J

_

_

fj1+" d < 1 for some " > 0, exp(fj ) d < 1 for some > 0.

We now consider uniformly maximable nets. De nition 1.4.16. A net ff; 2 g of R+ -valued functions on

is said to be uniformly maximable if lim sup 2

_

f 1(f > a) d ! 0 as a ! 1:

We have the following analogues of the properties of uniformly maximable collections of functions. Similar proofs apply.

Theorem 1.4.17. A net ff; 2 g is uniformly maximable if and only if the following conditions hold: (i) lim sup 2

_

f d < 1,

(ii) for every W" > 0 there exists > 0 such that lim sup2 A f d < " for every net of sets fA ; 2 g such that lim sup2 (A ) < .

Corollary 1.4.18.

A net ff ; 2 g is uniformly maximable if either one of the following conditions holds: 1. lim sup 2 2. lim sup 2

_

_

f1+" d < 1 for some " > 0, exp(f ) d < 1 for some > 0.

We now study convergence of idempotent integrals.

Theorem 1.4.19. Let ff; 2 g be a net of R+ -valued functions on and f be an R+ -valued function on .

1. (\the Fatou lemma"). If lim inf 2 f f -a.e., then

lim inf

_

f d

© 2001 by Chapman & Hall/CRC

_

f d:

26

Idempotent probability measures

2. (\the Lebesgue dominated convergence theorem"). If f and the net ff g is uniformly maximable, then

lim

_

f d =

_

f d:

3. (\the Lebesgue monotone convergence theorem"). If f -a.e., then

lim

_

f d =

_

! f

"f

f d:

4. Let be -smooth relative to a collection E of subsets of and T be a tightening collection for . Let f be Luzin (E ; T )=U measurable and maximable functions. If f # f -a.e., then

lim

_

f d =

_

f d:

Proof. We only prove part 4. Since ff ; 2 g is uniformly maximable and is T -tight, we can and do assume that the f are bounded by some N and E =U -measurable. Let for m 2 N

i i i i 1 f ; gm = max 1 f : i=1;:::;mN m i=1;:::;mN m m m We have that jgm; f j 1=m and jgm f j 1=m so that gm; = max _

_

j gm; d

f dj

( ) ; m

_

_

j gm d

f dj

( ) : m

On the other hand, using properties of and the facts that the f decrease and ff i=mg 2 E , as 2 , _

i i i f # max f i=1;:::;mN m m i=1;:::;mN m

gm; d = max

We have the following useful consequence.

© 2001 by Chapman & Hall/CRC

=

_

mi

gm d:

27

Idempotent integration

Corollary 1.4.20. Let be an E -idempotent measure with a tight-

ening collection T . Let f : ! R+ be Luzin (E ; T )=U -measurable and -maximable. Then the set function 0 de ned by 0 (A) = W

f 1(A) d; A ; is an E -idempotent measure, which has T as a tightening collection. Proof. It is obvious that 0 is a nite idempotent measure. It is -smooth relative to E by Theorem 1.4.19. It is T -tight since for a 2 R+ and T 2 T _ 0 (T c ) a(T c)+ f 1(f > a) d:

The following lemma establishes connection between convergence in idempotent measure and in \L1 ()". Lemma 1.4.21. Let ff; 2 g be a net of R+ -valued functions on

and f : ! R+ . W 1. If jf f j d ! 0, then f ! f .

2. If f ! f and the net ff g is uniformly maximable, then W

jf f j d ! 0. Proof. Part 1 follows by \the Chebyshev inequality". For part 2, note that if f ! f and the net ff g is uniformly maximable, then f is maximable by Theorem 1.4.19. Hence, by the inequality _

_

jf fj 1(jf fj > a) d [f 1(f > a)+f 1(f > a)] d

the net fjf f jg is uniformly maximable, so Theorem 1.4.19.

W

jf fj d ! 0 by

We now prove an analogue of Daniell's representation theorem, see, e.g., Meyer [88], stating that idempotent integral is speci ed by properties (JS2) and (JS3). Theorem 1.4.22. Let H be a set of R+ -valued functions on , which contains the zero function and is closed under multiplication by nonnegative scalars and the formation of maximums and minimums. Let E denote the paving on consisting of the sets ff ag; f 2 H; a 2 R + , and ;. Let V : H ! R + be a non-negative homogeneous maxitive functional, i.e., V has the properties

© 2001 by Chapman & Hall/CRC

28

Idempotent probability measures

(V 1) V (cf ) = c V (f ); c 2 R+ ; f

2 H; (V 2) V (f _ g) = V (f ) _ V (g); f; g 2 H. Then the following holds. 1. There exists an idempotent measure on such that

V (f ) =

_

f d; f

2 H;

if and only if V is -smooth along increasing nets in the sense that for every increasing net ff g of bounded functions from H such that sup f 2 H we have

V (sup f ) = sup V (f ):

The idempotent measure is uniquely speci ed on Eu . It is an idempotent probability if H contains the function identically equal to 1 and

(V 0) V (1) = 1. 2. Let H, in addition, have either one of the following properties: (a) if f 2 H, then (f 1) _ 0 2 H, (b) if f 2 H, then f ^ 1 2 H, and H is closed under multiplication. Then there exists an E -idempotent measure on such that

V (f ) =

_

f d; f

2 H;

if and only if the following condition holds: (VC) for every nets ff g and fg g of bounded functions from H, which are increasing and decreasing, respectively, and such that sup f inf g , we have

sup V (f ) inf V (g ):

© 2001 by Chapman & Hall/CRC

29

Idempotent integration

The idempotent measure is also -smooth relative to Ei and is uniquely speci ed on Eiu . If, in addition, for every f; g 2 H, we have (f g) _ 0 2 H and V (f + g) V (f ) + V (g) if f + g 2 H, then condition (VC) is equivalent to Daniell's condition: (VD) if f # 0, where the f are bounded functions from H, then V (f) # 0. Proof. We rst deal with the necessity parts. The condition of smoothness of V along increasing nets is necessary for existence of in part 1 by Theorem 1.4.19. Let us show necessity of (V C ) for being a -smooth idempotent measure. The condition follows by Theorem 1.4.19 (with T = f g) if we note that every f 2 H is E =U -measurable so that

sup

_

f d =

_

sup f d; inf

_

f d =

_

inf f d:

We prove suÆciency in part 1. Let 1 be the set whose elements are sets [0; a) f!g; a 2 R+ ; ! 2 . For R+ -valued functions f on , let Wf = f[0; a) f!g : a f (!)g. We de ne E1 = fWf ; f 2 Hg: By the assumptions on H the collection E1 is a paving on 1 . We set U (Wf ) = V (f ). Then the set function U is maxitive on E1 and satis es the hypotheses of part 1 of Theorem 1.1.9 (in particular, since V (0) = 0 by (V 1), it follows that U (;) = U (W0 ) = V (0) = 0). We de ne an extension of U to a set function on P ( 1 ) as in the proof of Theorem 1.1.9 by

U ([0; a) f!g) = inf U (Wf ); U (A1 ) =

f 2H: f (!)a

(1.4.1)

sup U [0; a) f!g ; A1 1 : [0;a)f!g2A1 By part 1 of Theorem 1.1.9 U is an idempotent measure on 1 , which extends U . Since by (V 1), for c 2 R+ and f 2 H, U (Wcf ) = V (cf ) = c V (f ) = c U (Wf ); equality (1.4.1) implies that

U ([0; c a) f!g) = c U ([0; a) f!g):

© 2001 by Chapman & Hall/CRC

(1.4.2)

30

Idempotent probability measures

Now, for A , we de ne (A) = U ([0; 1) A): Clearly, is an idempotent measure on . Also, for arbitrary f 2 H, by -maxitivity of U and (1.4.2)

V (f ) = U (Wf ) = U

[

[0; a) f!g

!2 : af (!)

= sup U [0; a) f!g = sup aU [0; 1) f!g !2 : !2 : af (!) af (!) _ = sup f (!)(f!g) = f d: !2

Part 1 is proved. We prove part 2. Note that under (V C ) the set function U satis es condition (S ) of part 2 of Theorem 1.1.9. Therefore, U is an E1-idempotent measure. We check the -smoothness property for . Let condition (a) of part 2 hold. Since for f : ! R+ and a > 0

f 1(f a) = xinf x 2R+ a

x +1

+

(1.4.3)

and the functions in the in mum belong to H, the sets W1(f a) belong to E1;i for f 2 H and a > 0. Now, let ff a g; f 2 H; 2 ; be a decreasing net of sets from E . We prove that \

inf (f a ) =

ff ag :

(1.4.4)

It is suÆcientTto consider the case a > 0. Then by (1.4.3) W1(f a ) = 2 F; for some F; 2 E1 ; where = R+ . For nite Tsubsets 0 and 0 of and , respectively, we de ne G0 ; 0 = 20 F; . Then 2 0

G0 ; 0

\ \

20

2

F; =

\

20

W1(f a ) W1(f0 a0 ) ;

(1.4.5) T 0 0 where is such that for all 2 0 . Also 0 ; 0 G0 ; 0 = T W 1(f a ) : Since fG0 ; 0 ; (0 ; 0 ) 2 Q() Q( )g is a decreasing net of elements of E1 with respect to the partial order on the

© 2001 by Chapman & Hall/CRC

31

Idempotent integration

pairs (0 ; 0 ) by inclusion and U is -smooth relative to E1 , inf U (G0 ; 0 ) = U

0 ; 0

= U W1

T

ff a g

\

= sup aU [0; 1) a2[0;1)

0 ; 0

= U \

G0 ; 0 = U

\

[

a2[0;1)

[0; a)

W1(f a )

\

ff ag

\

ff ag =

ff ag :

Also by (1.4.5) inf 0 ; 0 U (G0 ; 0 ) inf (f a ). Therefore, \

ff ag inf (f a );

and (1.4.4) follows. Now, is an Ei {idempotent measure by Lemma 1.1.8. If the condition that (f 1) _ 0 2 H for f 2 H is replaced by the conditions that f ^ 1 2 H for f 2 H and H is closed under multiplication, the same proof applies except that (1.4.3) is replaced by the equality 1(f a) = inf x2N (f=a)x ^ 1. To end the proof, let us assume that (f g) _ 0 2 H if f; g 2 H, V is subadditive, i.e., V (f + g) V (f ) + V (g) if f; g; f + g 2 H, and (VD) holds. We check that (VC) holds. Let f ", g # and sup f inf g , where the f and g are bounded functions from H. Then

V (g ) V (g

_ f) V (f)+ V ((g f) _ 0): f ) _ 0 2 H and tends monotonically to zero, an appli-

Since (g cation of (VD) yields (VC).

Remark 1.4.23. Under the hypotheses of part 1 (respectively, part

2) of Theorem 1.4.22 the functional V has a unique extension to a non-negative homogeneous maxitive functional on the set of Eu measurable (respectively, Eiu -measurable) R+ -valued functions on .

Remark 1.4.24.

If H, in addition to the hypotheses of Theorem 1.4.22, is such that (1 f ) _ 0 2 H for every f 2 H, then E is a semi- -algebra, hence, Eiu is a -algebra, which is also generated by sets ff = ag; f 2 H; a 2 R+ . In particular, by the preceding

© 2001 by Chapman & Hall/CRC

32

Idempotent probability measures

remark under the hypotheses of part 2 V has a unique extension on the set of R+ -valued functions that are measurable with respect to the -algebra generated by the elements of H.

Remark 1.4.25. Theorem 1.4.22 implies Theorem 1.1.9 if we take H = fc1(F ); F 2 E ; c 2 R+ g and V (c 1(F )) = c (F ) (note that c 1(F ) ^ 1 = (c ^ 1)1(F ) and (c 1(F ) 1) _ 0 = (c 1) _ 0 1(F ) so that f ^ 1 2 H and (f

1) _ 0 2 H if f

2 H).

Remark 1.4.26. Similarly to Theorem 1.1.9 (see Remark 1.1.11),

Theorem 1.4.22 admits a version where the collection H need not be closed under the formation of maximums and minimums. Then the maxitivity condition (V 2) and the -smoothness condition along increasing nets should be replaced in part 1 by the following condition: if a collection ffj g of elements of H and f 2 H are such that f supj fj , then V (f ) supj V (fj ). In part 2, condition (VC) would have to be replaced by the following: (V C 0) If ffj g is a collection of bounded functions from H and fg g is a decreasing net of bounded functions from H such that supj fj inf g ; then supj V (fj ) inf V (g ): Condition (V C 0 ) is necessary and suÆcient for to be -smooth relative to E . If H is closed under the formation of minimums, then the extension is -smooth relative to Ei .

1.5 Product spaces This section considers products of idempotent measure spaces. De nition 1.5.1. Let ( ; A) and ( 0 ; A0) be -measurable0 spaces. We de ne the product -algebra as the -algebra on with the atoms [!]A [!0 ]A0 . It is denoted by A A0 . The -measurable space ( 0 ; A A0 ) is called the product of ( ; A) and ( 0 ; A0 ). Remark 1.5.2. The product -algebra A A0 0is generated by the semi- -algebra consisting of the rectangles A A , where A 2 A and

© 2001 by Chapman & Hall/CRC

33

Product spaces

A0 2 A0. Since A A0 is also generated by the collection of sets A 0 and A0 , where A 2 A and A0 2 A0 , the -algebra A A0 is the smallest -algebra A~ on 0 such that the projections (!; !0 ) ! ! and (!; !0 ) ! !0 are A~=A-measurable and A~=A0 -measurable, respectively. Lemma 1.5.3. Let A 2 A A0 . Then the projection pr 0 A = f!0 2

0 : (!; !0 ) 2 A for some ! 2 g and cross-sections A! = f!0 2 0 : (!; !0 ) 2 Ag, where ! 2 , are elements of A0 . Also, if f : 0 ! R + is A A0 -measurable, then the cross-section f! (! 0 ) = f (!; ! 0 ) is A0-measurable as a function of !0. Conversely, if A! 2 A0 for every ! 2 and A!0 2 A for every !0 2 0 , then A 2 A A0 . Proof. We start with cross-sections. Let !0 2 A! . Then (!; !0 ) 2 A and by Corollary 1.1.16 [!]A [!0 ]A0 = [(!; !0 )]A A0 2 A. Thus, [!0 ]A0 2 A! and by Corollary 1.1.16 A! 2 A0 . The case of the projection is considered similarly. One could also use the representation pr 0 A = [!2 A! and the de nition of a -algebra. Measurability of f! follows since f! 1 (x) = f 1(x) ! for x 2 R+ . For the nal statement, we rst note that the cross-sections A! depend only on the atom to which ! belongs for otherwise there would exist !1 2 , A ! , !0 2 A and !0 62 A , !2 2 and !0 2 0 such that !1 2 !1 !2 which would imply that !1 2 A!0 but !2 62 A!0 so that A!0 62 A. The required now follows by the equality A = [!2 [!]A A! and the de nition of A A0 .

De nition 1.5.4. Let (

; A) and ( 0 ; A0 ) be -measurable spaces. 0

A function k : P ( ) ! [0; 1] is called an idempotent transition kernel from ( ; A) to ( 0 ; A0 ) if the following holds: (a) for every ! 2 , the function k(!; A0 ); A0 0 ; is an idempotent probability measure on 0 , (b) for every A0 2 A0 , the function k(!; A0 ); ! 2 ; is Ameasurable. Lemma 1.5.5. Let ( ; A) and ( 0; A0) be -measurable spaces. 1. Let be an idempotent measure on ( ; A) and k be an idempotent transition kernel from ( ; A) to ( 0 ; A0 ). Then there is a unique idempotent measure ~ on ( 0 ; A A0 ) such that _ ~(AA0 ) = k(!; A0 ) d(!); A 2 A; A0 2 A0: (1.5.1) A

© 2001 by Chapman & Hall/CRC

34

Idempotent probability measures

In particular, ~(A 0 ) = (A). If a function f : 0 ! R+ is A A0 -measurable, then the function _ g(!) = f! (!0 ) k(!; d!0 ); ! 2 ;

0

is A-measurable and _

0

f d~ =

_

g d:

2. Conversely, if ~ and are idempotent measures on respective -measurable spaces ( 0 ; A A0 ) and ( ; A) such that ~(A 0 ) = (A) for A 2 A, then there exists an idempotent transition kernel k from ( ; A) to ( 0 ; A0 ) such that (1.5.1) holds. The idempotent probability (k(!; A0 ); A0 2 A0 ) is uniquely speci ed on ( ; A0 ) for -almost all !. Remark 1.5.6. As we mentioned in Remark 1.1.29, saying \ is an idempotent measure on ( ; A)" is to mean that is uniquely speci ed only for elements of A. Likewise, uniqueness of ~ is claimed on A A0. Proof of Lemma 1.5.5. We begin with part 1. Since necessarily ~([!]A [!0 ]A0 ) = k(!; [!0 ]A0 )([!]A ), ~ is uniquely speci ed on A A0. We can de ne ~ on P ( 0) by ~(f(!; !0 )g) = k(!; f!0 g)(f!g); ~(A A0 ) = sup ~ f(!; !0 )g : !2A; !0 2A0 The only thing that requires proof is that g is A-measurable. This follows by the fact that f! (!0 )k(!; f!0 g) = f!0 (!)k(!; f!0 g) is Ameasurable in ! for every !0 2 A0 by Lemma 1.5.3 and Corollary 1.2.8. In part 2 necessarily ~([!]A [!0 ]A0 ) k(!; f!0 g) = ([!]A ) if ([!]A ) > 0, which proves uniqueness. For existence, we de ne k(!; f!0 g) by the latter formula if ([!]A ) > 0 and let k(!; A) be arbitrary idempotent probability measures on ( 0 ; A0 ) that are constant on the atoms of A on the rest of .

© 2001 by Chapman & Hall/CRC

35

Product spaces

Remark 1.5.7.0 If k(!; f!0 g) =0 0(f!0 g), where0 00 is an idempotent

measure on , then ~ f(!; ! )g = (f!g) (f! g) and we obtain an analogue of Fubini's theorem. The idempotent measure ~ is called the product of idempotent measures and 0 and denoted as 0 . The idempotent measure space ( 0 ; A A0 ; 0 ) is called the product of the idempotent measure spaces ( ; A; ) and ( 0 ; A0 ; 0 ). We consider now -smoothness and tightness properties of idempotent measures on product spaces. Theorem 1.5.8. Let us assume that an idempotent measure on

is -smooth relative to a collection E P ( ) such that ; 2 E and has a tightening collection T P ( ). Let an idempotent transition kernel k(!; A0 ) from ( ; P ( )) to ( 0 ; P ( 0 )) and collections E 0 P ( 0 ) and T 0 P ( 0 ), such that ; 2 E 0 and T 0 \ F 0 2 E 0 for every T 0 2 T 0 and F 0 2 E 0 , satisfy the following conditions: 1. k(!; A0 ) is Luzin (E ; T )=U -measurable in ! for every A0 2 E 0 , 2. k(!; A0 ) is a -smooth idempotent measure in A0 relative to E 0 for every ! 2 , 3. for every > 0 and T 2 T there exists T 0 2 T 0 such that sup!2T k(!; 0 n T 0 ) . Then the idempotent measure ~ on 0 de ned by ~ f(!; !0 )g = k(!; f!0 g)(f!g) is -smooth relative to E E 0 = fF F 0 ; F 2 E ; F 0 2 E 0 g and has the tightening collection T T 0 = fT T 0; T 2 T ; T 0 2 T 0g. Proof. We rst check the -smoothness. Let fF F 0 ; 2 g be a decreasing net of elements of E E 0 . Let F F 0 = \ 2 (F F 0 ). We have that _ ~(F F 0 ) = k(!; F 0 ) 1(! 2 F ) d(!): (1.5.2)

The functions k(!; F 0 ) 1(! 2 F ) are bounded, Luzin (E ; T )=U measurable and monotonically converge as 2 to k(!; F 0 ) 1(! 2 F ). Therefore by Theorem 1.4.19 ~(F F 0 ) ! ~(F F 0 ) checking -smoothness of ~. The fact that T T 0 is a tightening collection for ~ follows since by (1.5.2) ~ (T T 0 )c = ~ ( T 0 c )[(T c 0) (T c)_ sup k !; T 0 c : !2T

© 2001 by Chapman & Hall/CRC

36

Idempotent probability measures

For future use we also introduce the following notion.

De nition 1.5.9. Let A be a -algebra on a set and B be a -

algebra on a set . We de ne the product B A as the collection of subsets of that are expressed as unions of sets B [z ]A , where B 2 B and [z ]A are atoms of A, such that each atom of A appears in a union only once.

Remark 1.5.10. Clearly, B A is a -algebra but not a -algebra. 1.6 Independence and conditioning Let ( ; ) be an idempotent probability space.

De nition 1.6.1. A nite collection fAi ; i = 1; : : : ; kg of subsets of

is independent if k \

i=1

Ai =

k Y i=1

(Ai ):

A collection fA g of subsets of is independent if every nite subcollection is independent. A collection fE g of families of subsets of is independent if every collection of sets fA g, where A 2 E , is independent.

Lemma 1.6.2. A collection of families fE g is independent if and

only if the collection of families f(E )u g is independent. De nition 1.6.3. Let a set 0 be equipped with a -algebra 0 A0. A0 collection ff g of idempotent variables on with values in is A independent (or independent if A0 = P ( 0 )) if the collection of the algebras f 1 (A0 ) is independent. An idempotent variable f : ! 0 and a collection E of subsets of are A0-independent (or independent if A0 = P ( 0 )) if the -algebra f 1(A0 ) and E form an independent collection.

Remark 1.6.4. Loosely, we will often say that sets, -algebras or idempotent variables, respectively, are independent if the associated collections are independent.

© 2001 by Chapman & Hall/CRC

Independence and conditioning

37

Remark 1.6.5. Clearly, sets are independent if and only if their indicator functions are independent.

Lemma 1.6.6. Let idempotent variables f : ! 0 , where 0 is 0

equipped with a -algebra A , be measurable relative to respective algebras A on . If the collection fA g is independent, then the collection ff g is A0 -independent.

We have the following consequences. Lemma 1.6.7. 1. Let f : ! 0 and 0 be equipped with the discrete -algebra. The collection ff g is independent if and only if the collection of sets ff 1 (! )g is independent for all ! 2 0 . 2. Let ( ; A), ( 0 ; A0 ) and ( 00 ; A00 ) be -measurable spaces. Let f : ! 0 and F : 0 ! 00 be A=A0 - and A0 =A00 -measurable, respectively. If f is independent of a -algebra B on , then F Æ f is also independent of B. In the rest of the section we consider R+ -valued functions on

unless speci ed otherwise. We assume that R+ is equipped with the discrete -algebra P (R + ) and, according to the convention adopted in Section 1.2, refer to A=P (R + )-measurable functions f : ! R+ as A-measurable. We recall that if A is a -algebra on , then f is A-measurable if and only if the inverse images of one-element subsets of R+ belong to A.

Lemma 1.6.8. If f : ! R+ and g : ! R+ are independent, then S (fg) = S (f )S (g).

Proof. The result follows by the representations f (!) = supx2R+ x 1(f (!) = x), g(!) = supx2R+ x 1(g(!) = x) and (fg)(!) = supx;y2R+ (xy) 1(f (!) = x) 1(g(!) = y), and properties of idempotent expectations.

We now de ne conditional idempotent probabilities and conditional idempotent expectations. Let A be a -algebra on , E be a -system of subsets of containing ; and T be a collection of subsets of such that T \ F 2 E for T 2 T and F 2 E . For economy of notation we denote (f!g) as (!).

© 2001 by Chapman & Hall/CRC

38

De nition 1.6.9. 0 0

Idempotent probability measures

The conditional idempotent probability (! jA)(!) of ! given A is de ned by 8 0 A !); if ([!] ) > 0; < (! ) 1(! 0 A (!0 jA)(!) = ([!]A ) :~ 0 (! ); if ([!]A ) = 0; ~ is some idempotent probability on . where For B , we de ne (!0 jB ) = (!0 jAB )(!), where ! 2 B and AB is a -algebra, which has B as an atom (note that the right-hand side does not depend on the particular choice of ! 2 B and AB ). If A , then the conditional idempotent probability of A given A is de ned by (AjA)(!) = sup (!0 jA)(!); ! 2 : !0 2A Similarly, (AjB ) = sup!0 2A (!0 jAB )(!); ! 2 B . If is an E -idempotent probability, then the conditional E idempotent probability given A is de ned in an analogous manner except that ~ is required to be an E -idempotent probability on . ~ to be T -tight. Likewise, if is in addition T -tight, we require 0 0 If f : ! , where is equipped with a -algebra A0, we de ne (Ajf ) = Ajf 1 (A0 ) . Remark 1.6.10. According to the de nition, conditional idempotent probability is uniquely speci ed -a.e. More precisely, if N = f! 2

: ([!]A ) = 0g, then N 2 A, (N ) = 0 and (A \ [!]A ) (AjA)(!) = for all A and ! 2 N c: ([!]A ) Also, if B is such that (B ) > 0, then our de nition of (AjB ) agrees with the \standard" one in that (AjB ) = (A \ B )=(B ). Remark 1.6.11. Let us assume that is uniquely speci ed only on 0 a -algebra A A. Then, recalling that the atoms of A are unions of the atoms of A0 , we have for A0 2 A0 by the above de nition that 8 0 A !); if ([!] ) > 0; < ([! ]A0 ) 1(! 0 A 0 ([! ]A0 jA)(!) = ([!]A ) :~ ([!0 ]A0 ); if ([!]A ) = 0; so the values of the conditional idempotent probability on the elements of A0 do not depend on the extension of to P ( ).

© 2001 by Chapman & Hall/CRC

Independence and conditioning

39

We now list properties of conditional idempotent probabilities. Theorem 1.6.12. The function ((AjA)(!); A ; ! 2 ) has the following properties: 1. it is A{measurable in ! for all A , 2. it is an idempotent probability in A for every ! 2 , 3. for all A and B 2 A,

A \ B = S [(AjA)(!) 1(! 2 B )]:

4. If is an E -idempotent probability and [!]A 2 E (respectively, B 2 E ), then ((AjA)(!); A ) (respectively, ((AjB ); A

)) is an E -idempotent probability in A.

5. If is T -tight, then ((AjA)(!); A ) is T -tight for all ! 2 and ((AjB ); A ) is T -tight for all B . Proof. We begin with property 1. The function ! ! ([!]A ) is A-measurable since it is constant on the atoms of A. Obviously, 1(!0 !) is also0 A-measurable in !. By Lemma 1.2.8 we conclude that ! ! (! jA)(!) is A-measurable; hence, (AjA)(!) is also A-measurable. Property 2 follows by the de nition. We prove property 3. Assume, rst, that B = [^!]A for some !^ 2 . Then, adopting the convention that (!)=([!]A ) = 0 if ([!]A ) = 0, S (AjA)(!) 1(! !^ ) = sup sup (!0 jA)(!) 1(! !^ )(!) !2 !0 2A 0 (! ) = sup sup 1 (!0 !) 1(! !^ )(!) 0 ([ ! ] ) A !2 ! 2A (!0 ) = sup 1 (!0 !^ )([^! ]A ) = sup (!0 ) 1(!0 !^ ) 0 ([^ ! ] ) A ! 2A !0 2A = (A \ [^! ]A ): S In general, if B 2 A, then B = !2B [!]A , and this case reduces to the preceding one. Part 4 of the lemma concerning ! is a consequence of the de nition when ([!]A ) = 0. If ([!]A ) > 0, then by Remark 1.6.10 (A \ [!]A ) (AjA)(!) = ; ([!]A )

© 2001 by Chapman & Hall/CRC

40

Idempotent probability measures

which is a -smooth idempotent probability relative to E provided so is since [!]A 2 E and E is a -system. The proof for B is similar. In part 5 we also can assume that (!) > 0. Since is T -tight, given > 0, there exists T 2 T such that (T c ) ([!]A ), which implies that (T c jA)(!) . The proof for B is similar.

Remark 1.6.13. According to the lemma (AjA)(!) is an analogue

of regular conditional probability (cf., e.g., Ikeda and Watanabe [66, de nition 3.2]). Remark 1.6.14. Let

and 0 be -measurable sets with respective -algebras A and A0, and let be an idempotent probability on 0. Then k(!; A0 ) = ([!]A A0 j[!]A 0 ) is an idempotent transition kernel from ( ; A) to ( 0 ; A0 ).

We now de ne conditional idempotent expectations.

De nition 1.6.15. If f

is an R+ -valued function on , then the conditional idempotent expectation of f given A is de ned as S (f jA)(!) = sup f (!0)(!0 jA)(!): !0 2

If g : ! ^ , where ^ is equipped with a -algebra A^, we de ne S (f jg) = S f jg 1 (A^) . We also let S (f jg = !^ ) = sup f (!0 )(!0 jg = !^ ): !0 2

Remark 1.6.16. Note that S (f jA)(!) < 1 -a.e. if Sf < 1 and (AjA)(!) = S (1(A)jA)(!) -a.e. Remark 1.6.17. We will often use the following form of the de nition of conditional idempotent expectation: sup!0 2 f (!0) 1 !0 2 [!]A (!0 ) -a.e. S (f jA)(!) = [!]A

Remark 1.6.18. If A A0 and f is A0-measurable, then S (f jA)(!) = sup f (!0)([!0 ]A0 jA)(!) !0 2

so that S (f jA) depends only on the values of on A0 .

© 2001 by Chapman & Hall/CRC

(1.6.1)

Independence and conditioning

41

Remark 1.6.19.

In order to refer explicitly to the idempotent probability , we denote the conditional idempotent expectation as S (f jA). Since (!0 jA)(!) is speci ed {a.e. in !, S (f jA)(!) is also speci ed up to a set of zero idempotent probability. We call any such function a version of the conditional idempotent expectation. The following result is a consequence of the de nitions. Lemma 1.6.20. Let f : ! R+ and g : ! 0, where 0 is equipped with the discrete -algebra. Then -a.e. S (f jg)(!) = sup S (f jg = !0 ) 1(g(!) = !0 ); !0 2 0 in particular, -a.e.

(Ajg)(!) = sup (Ajg = !0 ) 1(g(!) = !0 ); A : !0 2 0 Conditional idempotent expectations have properties similar to the properties of conditional expectations in probability theory. They are summarised in the next lemma. All the equalities and inequalities involving conditional idempotent expectations are understood to hold -a.e. Following a convention of probability theory, we routinely omit argument ! in conditional idempotent probabilities and idempotent expectations.

Lemma 1.6.21. Let f , fj , and g denote R+ -valued functions on , A and B denote -algebras on . S (f jA) is A{measurable. If f = g -a.e., then S (f jA) = S (gjA). S (0jA) = 0, S (1jA) = 1: S (c f jA) = c S (f jA); c 2 R+ : S supj 2J fj jA = supj 2J S (fj jA): jS (f jA) S (gjA)j S (jf gjjA); if Sf < 1 and Sg < 1.

and 1. 2. 3. 4. 5. 6.

© 2001 by Chapman & Hall/CRC

42

Idempotent probability measures

7. S S (f jA) = Sf: 8 If f is independent of A, then S (f jA) = Sf . 9. Let ( 0 ; A0 ) be a -measurable space, F : 0 ! R+ be such that the cross-sections F! : 0 ! R+ are A0 -measurable for every ! 2 , and h : ! 0 be A=A0 -measurable. Then S F (; h)jA = S F (; x)jA jx=h. In particular, if g : ! R+ is A-measurable, then S (fgjA) = gS (f jA) and S (gjA) = g. 10. If B A, then S S (f jA)jB = S (f jB): 11. If A and B are independent, and f is independent of B, then S (f j (A; B)) = S (f jA). 12. If f is maximable, then the family fS (f jB); B Ag is uniformly maximable. 13. If 0 < p q, then S (f pjA) 1=p S (f q jA) 1=q . 14. If p 1 and q 1 are such that 1=p + 1=q = 1, then S (fgjA) S (f pjA) 1=p S (f q jA) 1=q . 15. (f ajA) S (f jA)=a; a > 0: Proof. Properties 1{6 follow by de nition. Consider property 7. By the de nition of conditional idempotent expectation, properties of idempotent integrals, and part 3 of Theorem 1.6.12,

S S (f jA) = S sup f (!0 )(!0 jA)(!) !0 2

= sup f (!0 )S [(!0 jA)(!)] = sup f (!0 )(!0 ) = Sf: !0 2

!0 2

Consider property 8. According to the de nitions -a.e. in ! 1(!0 !) (!0 ): S (f jA)(!) = sup f (!0) ([!]A ) !0 2

Since 1(!0 !)=([!]A ) is A{measurable in !0 and f (!0 ) is independent of A, 1(!0 !) (!0 ) sup f (!0 ) ([!]A ) !0 2

1(!0 !) (!0) = Sf: = sup f (!0 )(!0 ) sup !0 2

!0 2 ([! ]A )

© 2001 by Chapman & Hall/CRC

Independence and conditioning

43

Property 8 is proved. We prove property 9. We have that -a.e.

S (F (; h)jA)(!) = sup F (!0 ; h(!0 ))(!0 jA)(!) !0 2

(!0 ) = sup F (!0 ; h(!0 )) 1(!0 !): ([!]A ) !0 2

By A=A0{measurability of h and A0 -measurability of F!0 we have that F (!0 ; h(!0 )) = F (!0 ; h(!)) if !0 !, so we conclude that (!0 ) S (F (; h)jA)(!) = sup F (!0 ; h(!)) 1 (!0 !) 0 ([ ! ] ) A ! 2

= S (F (; x)jA)(!)jx=h(!) : Property 9 is proved. Proof of property 10. It is easy to derive from the de nition of conditional idempotent expectation that for {almost all ! S S (f jA)jB (!) = sup f (!00 ) sup (!0 jB)(!)(!00 jA)(!0 ): !00 2

!0 2

Hence, it suÆces to prove that for {almost all ! and all !00 sup (!0 jB)(!)(!00 jA)(!0 ) = (!00 jB)(!): (1.6.2) !0 2

If (!00 ) = 0, then (!00 jB)(!) = 0 {a.e. in !, so the right-hand side of (1.6.2) is equal to 0 -a.e. As for the left-hand side, if (!00 ) = 0 and (!00 jA)(!0 ) > 0, then (!0 ) = 0. The latter implies that if, in addition (!0 jB)(!) > 0, then (!) = 0. Thus, we conclude that if (!00 ) = 0 and (!0 jB)(!)(!00 jA)(!0 ) > 0, then (!) = 0. Hence, the left-hand side of (1.6.2) also is {a.e. equal to 0 when (!00 ) = 0. This ends the proof of (1.6.2) when (!00 ) = 0. Let (!00 ) > 0. We can assume that (!) > 0 so that ([!]A ) > 0 and ([!]B ) > 0. Then by the de nition of conditional idempotent probability the right-hand side of (1.6.2) takes the form (!00 ) (!00 jB)(!) = 1(!00 B !): (1.6.3) ([!]B ) On the other hand, (!00 ) > 0 implies that ([!00 ]A ) > 0, hence, A !00 , and by the de nition of conditional ([!0 ]A ) > 0 when !0

© 2001 by Chapman & Hall/CRC

44

Idempotent probability measures

idempotent probability we have for the left-hand side of (1.6.2) sup (!0 jB)(!)(!00 jA)(!0 ) !0 2

(!0 ) B !) (!00 ) : = sup 1 (!0 ([!0 ]A ) A !00 ([! ]B ) !0 : !0 A !00 implies that [!0 ] = [!00 ] . It also implies, The equivalence !0 A A B B B !) if 0 00 0 since B A, that ! ! , and hence 1(! !) = 1(!00 A !00 . Thus, on replacing in the latter supremum [!0 ] with [!00 ] !0 A A B !) with 1(!00 B !), we obtain that the supremum coinand 1(!0 cides with the right-hand side of (1.6.3). Equality (1.6.2) is proved. Property 10 is proved. We prove property 11. Let C = (A; B). Obviously, [!]C = [!]A \ [!]B . Then by de nition -a.e. 1(!0 2 [!]C ) (!0 ) S (f jC )(!) = sup f (!0 ) ([!]C ) !0 0 1(! 2 [!]A \ [!]B ) (!0 ) = sup f (!0) ([!]A \ [!]B ) !0 h 1(!0 2 [!]A) (!0)ih 1(!0 2 [!]B ) (!0 )i = sup f (!0 ) ([!]A ) ([!]B ) !0 = S (f jA)(!); where the equality before the last one is obtained by using the fact that f and A are independent of B. We prove property 12. By property 1 S (f jB) is B{measurable, so, successively applying properties 9 and 7, we have for a > 0

S S (f jB) 1 S (f jB) > a = S S (f 1 S (f jB) > a) jB = S f 1(S (f jB) > a) :

Now, by properties of idempotent expectations, for b > 0,

S f 1(S (f jB) > a) = S f 1(f > b) 1(S (f jB) > a) _ S f 1(f S f 1(f > b) _ b (S (f jB) > a)

b) 1(S (f jB) > a)

S f 1(f > b) _ ab Sf ;

© 2001 by Chapman & Hall/CRC

45

Independence and conditioning

where the last inequality is by \the Chebyshev inequality" and property 7. Hence, lim sup sup S S (f jB)1(S (f jB) > a) a!1 BA

S f 1(f > b) ;

which goes to 0 as b ! 1 by maximability of f . Inequalities 13 and 14 follow from the de nitions. Property 15, similarly to the probability theory counterpart, follows by the inequality 1(f a) f=a. The following lemma, which contains facts on convergence of conditional idempotent expectations analogous to facts from probability theory, is a consequence of the de nition of conditional idempotent expectation, Theorem 1.4.19, Theorem 1.6.12, and Lemma 1.6.21.

Lemma 1.6.22. Let A be a -algebra on .

Let ff ; 2 g be a net of R+ -valued functions and f be an R+ -valued function on .

f {a.e., then lim inf S (f jA) S (f jA) {a.e. 2

1. If lim inf 2 f

2. If S jf

2 ; then S jS (f jA) S (f jA)j ! 0:

3. If f

f j ! 0;

! f; 2 ; and ff g is uniformly maximable, then

S (f 4. If f

jA) ! S (f jA):

" f; 2 ; {a.e., then lim S (f jA) = S (f jA) {a.e. 2

5. Let be an E -idempotent probability, T be a tightening collection for , and the f be Luzin (E ; T )=U -measurable and maximable. Let the atoms of A belong to E . If f # f -a.e., then

S (f

jA) # S (f jA) -a.e.

© 2001 by Chapman & Hall/CRC

46

Idempotent probability measures

We now give versions of Levy's upward and downward theorems.

Lemma 1.6.23. Let A be a -algebra on and f : ! R+ . either one of the following conditions hold: 1.

fA ; T

2 g

2.

fA ; 2 g

2 A ,

is a decreasing net of -algebras and

Let

A =

is an increasing net of -algebras, E includes S the atoms of the A , A = 2 A , is a T -tight E idempotent probability, and f is Luzin (E ; T )=U -measurable and maximable.

Then

S (f jA) = lim S (f jA ) -a.e. 2

Proof. We prove part 1. Note that the net f[!]A ; 2 g is increasing for every ! 2 and [!]A = [ 2 [!]A . Equality (1.6.1) implies the claim by -maxitivity of and part 3 of Theorem 1.4.19. The proof of part 2 is similar: we note that the net f[!]A ; 2 g is decreasing for every ! 2 and [!]A = \ 2 [!]A , invoke (1.6.1), part 4 of Theorem 1.4.19, and the -smoothness property of .

In the sequel we will need the following characterisation of conditional idempotent expectations, which is a straightforward consequence of (1.6.1).

Lemma 1.6.24. Let f : ! R+ , g : ! R+ and A be a -algebra on . Then S (f jA) S (gjA) -a.e. if and only if S (f 1(A)) S (g 1(A)) for every A 2 A. The next implication of the lemma characterises conditional idempotent expectation in a manner similar to the de nition of conditional expectation in probability theory. We say that functions f and g on ( ; ) are indistinguishable if (f 6= g) = 0.

Lemma 1.6.25.

Let A be a -algebra and f : ! R+ . Then S (f jA) is the only up to indistinguishability function g : ! R+ , which is A-measurable and satis es the equality S (fh) = S (gh) for all A-measurable functions h : ! R+ .

© 2001 by Chapman & Hall/CRC

47

Independence and conditioning

Remark 1.6.26. If we de ned conditional idempotent expectation

by the property in the lemma, then De nition 1.6.9 would prove existence. In general, one cannot replace in Lemma 1.6.25 algebras with -algebras. Let us consider the following example. Let = [0; 1], (!) = 1 for every ! 2 [0; 1], B [0; 1] be the Borel -algebra on [0; 1], and f be an R+ -valued function on that is not Borel measurable. Suppose there exists conditional idempo tent expectation g of f given B [0; 1] satisfying the requirements of Lemma 1.6.25. Since, given x 2 [0; 1], the function 1(! = x) is Borel, we have sup!2[0;1] 1(! = x)f (!) = sup!2[0;1] 1(! = x)g(!), so that g(x) = f (x) for every x 2 [0; 1], which contradicts the requirement that g be Borel measurable.

We give \a transitivity law" for conditional idempotent expectations.

Lemma 1.6.27. Let f : ! 0 , g : 0 ! R+ and A0 be a -algebra 0 on . Then

S (g Æf jf 1 (A0 ))(!) = SÆf 1 (gjA0 )(f (!))

-a.e.

Proof. Let h(!0 ) = SÆf 1 (gjA0 )(!0 ); !0 2 0 . By Lemma 1.6.25 for every A0 -measurable function v : 0 ! R+ we have SÆf 1 (gv) = SÆf 1 (hv). By Theorem 1.4.6 this implies the equality S (g Æ f v Æ f ) = S(h Æ f v Æ f ). Since by Lemma 1.2.7 an arbitrary f 1 (A0 )measurable R+ -valued function on is of the form v Æ f for a suitable A0-measurable function v and h Æ f is f 1(A0)-measurable, we conclude by Lemma 1.6.25 again that h Æf = S(g Æf jf 1 (A0 )) -a.e.

The next lemma concerns evaluating conditional idempotent expectations for product idempotent probabilities.

Lemma 1.6.28. Let ( ; A0 ; ) and ( 0; A0 ; 0) be idempotent 0prob-

ability spaces and be equipped with -algebra A A and idempotent probability 0 . Let f : 0 ! R+ . Then S0 (f jA A0) = S (f~jA); where f~(!) = S0 (f! (!0 )jA0 ); ! 2

: In particular, if g : ! R+ and g0 : 0 ! R+ , then S0 (gg0 jA A0 ) = S (gjA)S (g0 jA0 ).

© 2001 by Chapman & Hall/CRC

48

Idempotent probability measures

Proof. The required follows since S0 (f jA A0 ) = sup f (!; !0 ) 0 (!; !0 )jA A0 (!;!0 )2 0 = sup sup f (!; !0 )0 (!0 jA0 ) (!jA): !2 !0 2 0

In probability theory existence of conditional expectation is proved by means of the Radon-Nikodym theorem. We can do without an analogue of the latter; however, it comes in useful below, so we state and prove it. De nition 1.6.29. Let A be a -algebra on0 . Let and 0 be idempotent probabilities on . We say that is absolutely continuous with respect to on A if for every " > 0 there exists Æ > 0 such that (A) < Æ implies that 0 (A) < for all A 2 A. Remark 1.6.30. Equivalently, we may require the above condition to hold only for the atoms of A. Note also that our de nition implies that if 0 is absolutely continuous with respect to , then 0 (A) = 0 whenever A 2 A and (A) = 0; by contrast with the situation in measure theory, the converse is not true. We have chosen the \strong" version as a de nition since it implies maximability of the Radon-Nikodym derivative (see Theorem 1.6.34 below). Remark 1.6.31. If 0 and are restricted to A, we simply say that 0 is absolutely continuous with respect to . De nition 1.6.32. We say that a function f : !0 R+ is a RadonNikodym derivative of an idempotent probability with respect to an idempotent probability Won a -algebra A if f is A-measurable, -maximable and 0 (A) = A f d for all A 2 A. We then denote f = d=d0 . We also write d = f d0 . Lemma 1.6.33. If f = d0 =d, then f (!) = 0 ([!]A )=([!]A ) a.e. Proof. Let (!) > 0. By A-measurability of f it is constant on [!]A so that _ 0 ([!]A ) = f d = f (!)([!]A ): [!]A

© 2001 by Chapman & Hall/CRC

49

Independence and conditioning

Thus, a Radon-Nikodym derivative is unique -a.e. Theorem 1.6.34. An idempotent probability 0 is absolutely continuous with respect to an idempotent probability on A if and only if there exists a Radon-Nikodym derivative of with respect to 0 on A. Proof. Existence of the derivative implies the absolute continuity by Corollary 1.4.12. For the converse, we de ne the derivative as in W 0 Lemma 1.6.33. Then f is A-measurable and (A) = A f d for all A 2 A. To show f is -maximable, we write for a > 0 _

f 1(f > a) d = 0 (f > a) 0 (! : (!) 1=a):

Since 0 is absolutely continuous with respect to , the latter idempotent probability can be made less than arbitrary > 0 by choosing a large enough. Lemma 1.6.35. Let an idempotent probability 0 on be absolutely continuous with respect to on a -algebra A. Let f = d0 =d > 0 -a.e. Then, for a function g : ! R+ and -algebra B A, -a.e. S (fgjB) S0 (gjB) = : S (f jB) Proof. Since 0 = S S (f jB) 1(S (f jB) = 0) = S f 1(S (f jB) = 0) and f > 0 -a.e., it follows that S (f jB) > 0 -a.e., so the right-hand side in the above equality is well de ned -a.e. We next have by properties of conditional idempotent expectations that for B2B

S S0 (gjB)S (f jB) 1(B ) = S S S0 (gjB)f 1(B )jB = S S0 (gjB)f 1(B ) = S0 S0 (gjB) 1(B ) = S0 (g 1(B )) = S(fg 1(B )):

Thus, by Lemma 1.6.25 S0 (gjB)S (f jB) = S (fgjB). We end the section by giving versions of Lemma 1.6.24 and Lemma 1.6.25 for -smooth tight idempotent probabilities.

© 2001 by Chapman & Hall/CRC

50

Idempotent probability measures

Lemma 1.6.36. Let E be a semi- -algebra, be -smooth relative

to E and T be a tightening collection for . Let f : ! R+ and g : ! R+ be Luzin (E ; T )=U -measurable and maximable functions. Then the following holds. 1. S (f j (E )) S (gj (E )) -a.e. if and only if S (f 1(A)) S (g 1(A)) for every A 2 E .

2. S (f j (E )) is the only up to indistinguishability Luzin (E ; T )=U measurable maximable function f 0 : ! R+ that is (E )measurable and satis es the equality S (fh) = S (f 0 h) for all E -measurable functions h : ! R+ . Proof. Necessity of the condition in part 1 is obvious. We prove suÆciency. Let (A) = S (f 1(A)) and (A) = S (g 1(A)). By Corollary 1.4.20 and are E -idempotent measures on such that on E . Theorem 1.1.7 implies that on Eiu . Since E is a semi- algebra, Eiu = (E ), completing the proof by Lemma 1.6.24. Part 2 is a consequence of part 1.

Lemma 1.6.37. Let be -smooth relative to a collection E and has

a tightening collection T . Let H be a collection of Luzin (E ; T )=U measurable R+ -valued functions on that contains the zero function, is closed under multiplication by non-negative scalars and the formation of maximums and minimums, and is such that if h 2 H, then (h 1) _ 0 2 H and (1 h) _ 0 2 H. Let f : ! R+ be maximable and Luzin (E ; T )=U -measurable and let A denote the -algebra generated by the elements of H. If a Luzin (E ; T )=U -measurable maximable function g : ! R+ is A-measurable and such that S (fh) = S (gh) for all h 2 H, then g = S (f jA) -a.e. Proof. Let V (h) = S (fh); h 2 H. By Theorem 1.4.19 and the hypotheses the functional V satis es condition (V C ) of Theorem 1.4.22. Therefore, V has a unique extension to a non-negative homogeneous maxitive functional on the set of A-measurable R+ -valued functions on (see Remark 1.4.24). The same fact is true for the functional V 0 (h) = S (gh). Hence, S (fh) = S (gh) for every Ameasurable R+ -valued function h : ! R+ and the required follows by Lemma 1.6.25.

© 2001 by Chapman & Hall/CRC

Topological spaces

51

1.7 Idempotent measures on topological spaces In the next three sections we consider -smooth idempotent measures on topological spaces. Let E be a Hausdor topological space. It will play the part of the set . The part of the collection E will be played either by the collection F of closed subsets of E or the collection K of compact subsets of E . We consider only nite idempotent measures throughout.

De nition 1.7.1. We say that an idempotent measure on E is tight if it has K as a tightening collection, i.e., for every " > 0 there exists compact K E such that (K c ) ".

Note that since E is Hausdor, F -idempotent measures are Kidempotent measures; also the classes of tight F -idempotent measures and tight K-idempotent measures coincide. Therefore, we occasionally refer to tight F -idempotent measures on Hausdor spaces as tight -smooth idempotent measures. We denote (z ) = (fz g) so that symbol will alternatingly be used for an idempotent measure and its density. To avoid confusion we will denote the density by (z ) and the idempotent measure by . We recall that according to Lemma 1.1.4

(A) = sup (z ); A E: z 2A

(1.7.1)

The next lemma relates properties of (z ) to properties of . Recall that a function f : E ! R+ is upper semi-continuous if the sets fz 2 E : f (z) ag are closed.

De nition 1.7.2. A function f : E ! R+ is said to be upper compact if the sets fz 2 E : f (z ) ag are compact for a > 0.

Remark 1.7.3. Upper compact functions attain suprema on closed sets.

Lemma 1.7.4. Let a function (z) :

E ! R+ and set function : P (E ) ! R+ be related by (1.7.1). Then the following holds. 1. If the set function is an F {idempotent measure on E , then the function (z ) is upper semi-continuous.

© 2001 by Chapman & Hall/CRC

52

Idempotent probability measures

2. If the function (z ) is upper semi-continuous, then the set function is a K{idempotent measure. If E is either rst countable or locally compact, then the converse is also true. 3. The set function is a tight F {idempotent measure on E if and only if the function (z ) is upper compact. Proof. Let be an F {idempotent measure and let z ! z; 2 . Then applying the -smoothness property of to the sets F = clfz0 ; 0 g and noting that \ F = fz g, we conclude that lim sup (z ) (z ), i.e., is upper semi-continuous. The rst assertion in part 2 follows from Lemma 1.7.6 below with f (z ) = (z ) 1(z 2 K ), where fK g is a decreasing net of compacts. The second assertion in the case E is rst countable is proved by the argument of the proof of part 1 since we can assume that fz g is countable so that the F are compact. If E is locally compact, then, given z ! z and > 0, by -smoothness there exists a compact K such that z 2 K , z 2 K for \large" and (K ) (z ) + . If is a tight F -idempotent measure, then (z ) is upper semicontinuous by part 1; besides, if K is a compact such that (K c) < a, then fz : (z ) ag K , so (z ) is upper compact. If (z ) is upper compact, then is a K-idempotent measure by part 2. It is tight since, given > 0, one can take K = fz : (z ) "g.

In the sequel, we denote

K (a) = fz 2 E : (z ) ag; a > 0:

(1.7.2)

Remark 1.7.5. According to the lemma, if is a tight F -idempotent

measure, then the sets K (a); a > 0; make up a tightening collection for . We give the lemma used in the above proof. Lemma 1.7.6. Let f be a net of R+ -valued upper compact functions on E monotonically decreasing and converging pointwise to function f . Then

sup f(z ) # sup f (z ): z 2E Proof. For " > 0, let B = fz 2 E : f (z ) supz0 2E f (z 0 ) + "g. The sets B are compact, decreasing and \B = ;. Hence, B0 = ; for some 0 . z 2E

© 2001 by Chapman & Hall/CRC

53

Topological spaces

The following useful fact is in the same theme.

Theorem 1.7.7.

Let be an F -idempotent measure on E . Let be a collection of R+ -valued bounded upper semicontinuous functions closed under the formation of minimums. Then

ffj ; j 2 J g inf j 2J

_

E

fj d =

_

E

inf fj d:

j 2J

Proof. The claim follows by the last assertion of Theorem 1.4.19 if we observe that T = fE g is a tightening collection for and upper semi-continuous functions are Luzin (F ; T )=U -measurable.

The next result is an analogue of Ulam's theorem, see, e.g., Billingsley [11].

Theorem 1.7.8. Let E be either homeomorphic to a complete metric space or locally compact. If is an F -idempotent measure on E , then is tight.

Proof. Let E be metrised by a complete metric. Given Æ > 0, let OÆ denote the collection of nite unions of open Æ-balls in E ordered by inclusion. It follows by -smoothness of relative to F that limO2OÆ (E n O) = 0: Therefore, for arbitrary > 0 and n 2 Nthere n A exist open 1=n-balls An;1 ; : : : ; An;kn such that E n [ki=1 n;i < : k 1 n The set A = \n=1 [i=1 An;i is totally bounded so that, since E is complete, the set K = cl A is compact. Also (E nK ) supn2N E n n A [ki=1 n;i : If E is locally compact, then the -smoothness property of implies that for every > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that (E n [ki=1 Ai ) < :

Let E 0 be a Hausdor topological space and F 0 denote the collection of closed subsets of E 0 . We introduce a class of maps f : E ! E 0 that preserve the property of an idempotent measure being tight and -smooth relative to the collection of closed sets.

De nition 1.7.9. Let0 be a tight F -idempotent measure on E . A

function f : E ! E is said to be -Luzin measurable (or simply Luzin measurable if is understood) if it is continuous when restricted to the compacts K (a); a > 0.

© 2001 by Chapman & Hall/CRC

54

Idempotent probability measures

Remark0 1.7.10. The de nition adapts the abstract notion of Luzin

(E ; T )=E -measurable functions in that f is -Luzin measurable if and only if it is (F ; K )=F 0 -measurable, where K = fK (a); a > 0g.

The equality f K (a) = KÆf 1 (a), which is valid for arbitrary f : E ! E 0 , and Theorem 1.2.14 yield the following result.

Theorem 1.7.11. If is a tight F -idempotent 0measure 0 on E and1 f is a -Luzin measurable mapping from E to E , then = Æ f is a tight F 0 -idempotent measure on E 0 .

An application to f being an embedding provides the following extension result. Corollary 1.7.12. Let E E 0 and the topology on E be ner than the topology induced by the topology on E 0 . Let be a tight F idempotent measure on E . Then the set function 0 on E 0 de ned by 0 (A0 ) = (A0 \ E ) is a tight F 0 -idempotent measure on E 0 .

In view of an important role played by tight F -idempotent probabilities in large deviation theory, we give them a special name.

De nition 1.7.13. A tight F -idempotent probability is called a deviability. If the idempotent distribution of an idempotent variable is a deviability, we also refer to it as a deviability distribution.

Remark 1.7.14. By Theorem 1.7.8 on complete metric spaces and

on locally compact spaces F -idempotent probabilities are deviabilities.

The following is a corollary of Lemma 1.7.4 that will be used frequently below.

Corollary 1.7.15.

An idempotent measure is a deviability if and only if the density (z ) is an upper compact function and supz2E (z ) = 1.

Remark 1.7.16. Since upper compact functions attain suprema on closed sets, for every deviability there exists z (z ) = 1.

2E

such that

Remark 1.7.17. We note that is a deviability if and only if the function I (z ) = ln (z ) is a tight probability rate function in the sense that the sets fz 2 E : I (z ) ag are compact for all a 2 R+ and inf z2E I (z ) = 0.

© 2001 by Chapman & Hall/CRC

55

Topological spaces

De nition 1.7.18.

1. Let ( ; ) be an idempotent probability space. An idempotent variable f : ! E is called a Luzin idempotent variable on ( ; ) if its idempotent distribution Æ f 1 is a deviability on E .

2. Let, in addition, be a Hausdor topological space and be a deviability. An idempotent variable f : ! E is called a strictly Luzin variable on ( ; ) if it is -Luzin measurable.

Remark 1.7.19. By Theorem 1.7.11 strictly Luzin idempotent variables are Luzin idempotent variables.

Strictly Luzin idempotent variables have another useful property. It is a topological version of Corollary 1.4.20. Lemma 1.7.20. Let be a Hausdor topological space and be a deviability on . Let f be a maximable R+ -valued strictly Luzin idempotent variable on ( ; ) such that S f = 1. Then the set function 0 (A) = S f 1(A); A E; is a deviability on E . We now consider topological versions of Theorem 1.4.22. They are analogues of Riesz' representation theorem. Let CK+ (E ) denote the set of R+ -valued continuous functions on E with compact support. Theorem 1.7.21. Let E be a locally compact Hausdor topological + space and V : CK (E ) ! R+ be a functional with properties (V 1) and (V 2) from Theorem 1.4.22, i.e., (V 1) V (c f ) = c V (f ); c 2 R+ ; (V 2) V (f _ g) = V (f ) _ V (g): Then there exists a K{idempotent measure on E such that

V (f ) =

_

E

f d; f

2 CK+(E ):

The idempotent measure is uniquely speci ed on P (E ). Proof. By Theorem 1.4.22 in order to prove existence of we need to check that if ff' ; ' 2 g and fg ; 2 g are, respectively, increasing and decreasing nets of elements of CK+ (E ) such that

sup f' (z ) inf g (z ); z 2 E; 2

'2

© 2001 by Chapman & Hall/CRC

56

Idempotent probability measures

then sup V (f' ) inf V (g ): 2

(1.7.3)

'2

Replacing if necessary the net ff' g by the net ff' ^ g ^g, where ^ 2 is picked arbitrarily, we can assume that all the above functions are supported by a compact K . Let hK 2 CK+ (E ) be such that hK = 1 on K . Given " > 0, the net f(g =(f' _ ("hK )) 1)+ g indexed by (; ) monotonically converges to 0 on K . By Dini's theorem the convergence is uniform, so there exist '0 and 0 such that

g (z ) (1+")(f' (z )_("hK (z ))); z 2 E;

0 ; ' '0 :

Therefore, by the properties of V

V (g ) (1+")(V (f' )_("V (hK )));

0 ; ' '0 ;

which implies (1.7.3) since " > 0 is arbitrary. Since Kiu = P (E ), by Theorem 1.4.22 is uniquely speci ed on P (E ). Corollary 1.7.22. Let and 0 be K-idempotent measures on a locally compact Hausdor topological space E . If _

E

f d =

_

E

f d0 ; f 2 CK+ (E );

then = 0 on P (E ).

The following version for compact spaces has an analogous proof.

Theorem 1.7.23. Let E be a compact Hausdor topological space

and H be a set of R+ -valued continuous functions on E that contains the zero function, is closed under the multiplication by non-negative scalars and formation of maximums and minimums, and is such that if f 2 H, then (f 1) _ 0 2 H. If V : H ! R+ is a functional with properties (V 1) and (V 2), then there exists a (KH )i {idempotent measure on E such that

V (f ) =

_

E

f d; f

© 2001 by Chapman & Hall/CRC

2 H;

57

Topological spaces

where KH is the collection of compacts fz 2 E : f (z ) ag; a 2 R + ; f 2 H. The idempotent measure is uniquely speci ed on (KH )iu . If, in addition, H contains constants and

(V 0) V (1) = 1,

then is a (KH )i -idempotent probability.

Remark 1.7.24.

Theorem 1.7.21 can be derived from Theorem 1.7.23 if one recalls that a locally compact Hausdor space is homeomorphic to an open subset of a compact Hausdor space.

For an R+ -valued function f on E , let kf k = supz2E f (z ). Let Cb+ (E ) denote the set of R+ -valued bounded continuous functions on E . We recall that Tihonov spaces are completely regular T1 -spaces, see, e.g., Kelley [71]. Theorem 1.7.25. Let E be a Tihonov topological space and V : + Cb (E ) ! R+ be a functional with properties (V 1) and (V 2), which is, in addition, tight in the sense that for arbitrary " > 0 there exists a compact K E such that V (f ) " kf k for every f 2 Cb+(E ) that equals 0 on K . Then there exists a tight F {idempotent measure on E such that

V (f ) =

_

E

f d; f

2 Cb+(E ):

The idempotent measure is speci ed uniquely on P (E ). If, in addition, condition (V 0) holds, then is a tight F -idempotent probability. Proof. Let E be the Stone-Czech compacti cation of E , see, e.g., Engelking [47]. We de ne a functional V on Cb+(E ) by V (f ) = V (f ), where f denotes the restriction of f 2 Cb+(E ) to E . It is obvious that V satis es the conditions of Theorem 1.7.21. By Theorem 1.7.21 there exists a K{idempotent measure on E , where K is the collection of compact subsets of E , such that

V (f ) =

_

E

f d; f

2 Cb+(E ):

(1.7.4)

We show that is K-tight, i.e., for every " > 0 there exists a compact K E such that (E n K ) ". Let K be as in the hypotheses.

© 2001 by Chapman & Hall/CRC

58

Idempotent probability measures

The set E n K is open in E so, since E is Tihonov, 1(E n K ) = sup f over f 2 Cb+(E ) such that f 1(E n K ). Therefore, since all these f are equal to 0 on K and is an idempotent measure, (E n K ) = supf V (f ) = supf V (f ) ": Since the embedding E ! E is continuous, the restriction of to E , de ned by (A) = (A) for A E , is a K-idempotent measure. It is tight since W is K-tight. Moreover, K -tightness of implies that ( E n E ) = 0, so E f d = W + E f d: Thus, by (1.7.4) we have for f 2 Cb (E ), denoting by f the continuous extension of f to E , that

V (f ) = V (f ) =

_

E

f d =

_

E

f d:

Remark 1.7.26. According to the+ theorem and part (JS 5) of The-

orem 1.4.4, the functional V : Cb (E ) ! R+ admits a continuous extension to a functional on the space of bounded R+ -valued functions on E with sup-norm. Theorem 1.7.27. Let and 0 be F -idempotent measures on a Tihonov topological space E . If _

E

f d =

_

E

f d0 ; f 2 Cb+ (E );

then = 0 on P (E ).

Proof. Given z 2 E , we have, since E is Tihonov, that 1(fz g) = inf ff 2 Cb+ (E ) : f (z ) = 1g. Therefore, by Theorem 1.7.7

(z ) =

inf +

_

f 2Cb (E ): E f (z )=1

f d;

0 (z ) =

inf +

_

f 2Cb (E ): E f (z )=1

f d0 :

1.8 Idempotent measures on projective limits Our purpose here is to prove analogues of extension theorems for projective systems in measure theory. We are only able to get nice results for projective systems of tight -smooth idempotent measures.

© 2001 by Chapman & Hall/CRC

Projective limits

59

We actually formulate the results for idempotent probabilities, which are our main concern in this book. Let (E ) 2 be a net of Hausdor topological spaces indexed by a directed set . We assume that for all , 2 , 2 , there are maps : E ! E such that = Æ for . We denote by F the collections of closed subsets of the E . Let the E be equipped with deviabilities (i.e., tight F idempotent probabilities). For " > 0, we denote K"; = fz 2 E : (z ) "g. We assume that the maps are Luzin measurable, i.e., their restrictions to the K"; are continuous. The deviabilities are assumed to form a projective system in that = Æ 1 for . We note that this implies that K; = K; . Let E be a Hausdor topological space and maps : E ! E be such that = Æ for . Let F denote the collection of closed subsets of E . Theorem 1.8.1. Let the maps ; 2 ; separate points in E . Let for every " > 0 there exist a compact K" E such that the restrictions of the maps to K" are continuous and (; K ) ( K")c ": Then there exists a deviability on E such that = Æ 1 . Deviability is uniquely speci ed by (z ) = inf 2 ( z ): Remark 1.8.2. Condition (; K ), being an analogue of the ("; K )condition for Radon measures (Schwartz [118]) is referred to below as such. We rst consider the following special case. Lemma 1.8.3. Let E be the projective limit of the system (E ) and the be the canonical projections from E to E . Then the assertion of Theorem 1.8.1 holds. Proof. We de ne (z ) = inf ( z ); z 2 E; 2 and (A) = supz2A (z ); A E: Let K" = fz Then

K =

\

1 K; ;

© 2001 by Chapman & Hall/CRC

(1.8.1)

2 E : (z) "g: (1.8.2)

60

Idempotent probability measures

so K is the projective limit of the (K; ; 2 ). Therefore, K is compact for 2 (0; 1]. It is also empty for > 1. Thus, is a deviability on E by Corollary 1.7.15. Since the \bonding" maps : K; ! K; are onto and continuous, the maps : K ! K; are also onto, see Engelking [47, Corollary 3.2.15]. Therefore,

K; = K ;

(1.8.3)

which is equivalent to the equality (z ) = sup (z ); z 2 1 z

(1.8.4)

which proves that = Æ 1 . Conversely, if is a deviability on E such that = Æ 1 ; 2 ; then (1.8.4) holds, which is equivalent to (1.8.3), which implies, since K is compact, that K is the projective limit of the (K; ; 2 ) so that (1.8.2) holds, Engelking [47, Proposition 2.5.6]. The latter is equivalent to (1.8.1).

Remark 1.8.4. Note that the (; K )-condition is satis ed in this setting in view of (1.8.3). We call the projective limit of the .

For a proof of the general case we need a lemma. Lemma 1.8.5. Let E and E 0 be Hausdor topological spaces. Let E 0 be endowed with an F 0 -idempotent probability 0 , where F 0 is the collection of closed subsets of E 0 . Let an injective mapping h : E ! E 0 and a collection K^ of compact subsets of E be such that the restrictions of h to the elements of K^ are continuous and h(K^ ) is a tightening collection for 0 . Then there exists a unique deviability on E such that 0 = Æ h 1 . It is speci ed by the equality (A) = 0 (h(A)); A E . Proof. We de ne as in the statement of the lemma. In order to check that is -smooth relative to F , let us consider a decreasing net fF ; 2 g of closed subsets of E . For > 0, we choose a compact K 2 K^ such that 0 (E 0 n h(K )) < . Then by the fact that h is injective (F ) = 0 (h(F )) 0 h(F )\h(K ) + = 0 h(F \K ) +:

© 2001 by Chapman & Hall/CRC

61

Projective limits

Since h is continuous when restricted to K and 0 is -smooth relative to F 0 , we have that \

inf 0 h(F \ K ) = 0 h(F \ K ) 2 2

\

0 h

=

2 \

F

2

F :

Thus, \

inf (F ) F + ; 2 2 implying that is -smooth relative to F . Finally, is tight since by h being injective (K c ) = 0 (h(K )c ). Proof of Theorem 1.8.1. Let E 0 be the projective limit of the (E ), let K"0 ; " > 0; be the respective projective limits of the (K"; ), and let 0 : E 0 ! E be the canonical projections. By the part of the theorem already proved there exists a unique deviability 0 on E 0 such that = 0 Æ 0 1 . On the other hand, it is easy to see that there exists a map h : E ! E 0 , which is continuous when restricted to the sets K" , and is such that 0 Æh = . By the ("; K )-condition K K; , which implies that h(K" ) K"0 . Therefore, fh(K" )g is a tightening collection for 0 . Since the family separates points in E , h is injective. Thus, by Lemma 1.8.5 there exists a unique deviability on E such that 0 = Æh 1 , which implies that = Æ 1 . Also (z ) = 0 (h(z )) = inf 0 Æh(z ) = inf ( z ); z 2 E: 2 2

We now consider an application to product spaces. Let fEj ; j 2 J g be a family of Hausdor topological spaces with collections Fj of closed sets. Let be the set of nite subsetsQof elements of J . For 2 let E denote the Cartesian product j2 Ej with product topology; E is endowed with the collection F of closed sets. Let the sets E be equipped with F -idempotent probabilities . As above we denote K"; = fz 2 E : (z ) "g. For , let

© 2001 by Chapman & Hall/CRC

62

Idempotent probability measures

Q

denote the canonical projection E ! E . Let E j 2J Ej be equipped with a Hausdor topology, which is ner than the relative product topology, and E denote the restriction to E of the canonical Q projection j 2J Ej ! E . The next result follows by Theorem 1.8.1.

Theorem 1.8.6. Let the idempotent probabilities 1

form a projective system, i.e., = Æ if . If for every > 0 there exists a compact subset K of E such that (E K )c for all 2 , then there exists a tight F -idempotent probability on E such that = Æ E 1 ; 2 . It is uniquely speci ed by (z ) = inf 2 (E z ). In particular, exists if E is equipped with product topology and the idempotent probabilities j ; j 2 J; are tight. Proof. Only the last claim requires proof. We check that Q the are ^ tight if the j ; j 2 J; are tight. For " > 0, let K"; = j 2 K";j , c ) ". Then K ^ "; is compact and where j (K";j

K^ ";c

=

[

j2

c j 1 K";j

c = sup j 1 K";j j2 c ": = sup j K";j j2

For the (; K )-condition, we can take K = \j jE 1 K;j . The following consequence of Lemma 1.8.5 complements Corollary 1.7.12. Its weaker version has been used in the proof of Theorem 1.7.25. Corollary 01.8.7. Let E and E 0 be Hausdor topological spaces such that E E and the topology of E is ner than the topology induced by the topology of E 0 . Let 0 be a deviability on E 0 . If the collection of compact subsets of E is a tightening collection for 0 , then the set function on E de ned by (A) = 0 (A) is a deviability on E . We now give topological versions of the results of Section 1.5. De nition 1.8.8. Let E and E 0 be Hausdor topological spaces. A function k(z; A0 ) : E P (E 0 ) ! [0; 1] is called a deviability transition kernel from E into E 0 if the following conditions hold: 1. k(z; A0 ) is upper semi-continuous in z for every closed set A0 E0,

© 2001 by Chapman & Hall/CRC

Projective limits

63

2. k(z; A0 ) is a deviability in A0 for every z 2 E and is uniformly tight on compact subsets of E in the sense that for every compact K E and > 0 there exists a compact K 0 E 0 such that supz2K k(z; E 0 n K 0 ) . Theorem 1.5.8 yields the following result.

Theorem 1.8.9.

Let be a tight F -idempotent measure on a Hausdor topological space E . Let E 0 be another Hausdor topological space and E E 0 be equipped with product topology. Let k(z; A) : E P (E 0 ) ! [0; 1] be a deviability transition kernel from E into E 0 . Then the idempotent measure ~ on E E 0 de ned by ~(z; z 0 ) = k(z; z 0 )(z ) is tight and -smooth relative to the collection of closed subsets of E E 0 . Proof. By Theorem 1.5.8 ~ is -smooth relative to the collection fF F 0 g, where F and F 0 are closed in E and E 0 , respectively, and has the tightening collection fK K 0 g, where K and K 0 are compact in E and E 0 , respectively. The -smoothness property implies that ~(z; z 0 ) is an upper semi-continuous function on E E 0 , and the tightness property then allows us to deduce that the function is actually upper compact. Corollary 1.8.10. If and 0 0are deviabilities on respective Hausdor topological spaces E and E , then the product idempotent measure ~ = 0 is a deviability on E E 0 equipped with product topology.

Combining the latter with Theorem 1.8.6 yields the following existence result.

Corollary 1.8.11.

Let j ; j 2 J; be a collection of deviabilities on respective Hausdor spaces Ej . Then there exists an idempotent probability space ( ; ) and independent Luzin idempotent variables fj ; j 2 J; on ( ; ) whose respective deviability distributions are the j .

The following condition will be used for checking that a function is a deviability transition kernel. Lemma 1.8.12. Let E and E 0 be Hausdor0 topological0 spaces and let E be rst countable. If a function k(z; z ) : E E ! [0; 1] is

© 2001 by Chapman & Hall/CRC

64

Idempotent probability measures

upper semi-continuous in (z; z 0 ), the sets fz 0 2 E 0 : supz2K k(z; z 0 ) ag, where a 2 (0; 1], are relatively compact subsets of E 0 for every compact K E , and supz0 2E 0 k(z; z 0 ) = 1 for every z 2 E , then k(z; A0 ) = supz0 2A0 k(z; z 0 ) is a deviability transition kernel from E into E 0 . Proof. We rst note that the hypotheses imply that the function supz2K k(z; z 0 ) is upper compact for every compact K E so that condition 2 in the de nition of a deviability transition kernel holds. We check condition 1. Let A0 be a closed subset of E 0 and zn ! z as n ! 1. The set K = [n2N fzn g [ fz g is a compact subset of E . Therefore, given an arbitrary > 0, there exists compact K 0 E 0 such that k(zn ; E 0 n K 0 ) ; n 2 N . Hence,

lim sup k(zn ; A0 ) lim sup k(zn ; K 0 \ A0) + k(z; K 0 \ A0 ) + n!1 n!1 k(z; A0 ) + ; where the second inequality follows by upper semi-continuity of k(z; z 0 ) and the fact that K 0 \ A0 is compact.

1.9 Topological spaces of idempotent probabilities In this section we consider topologies on the space of idempotent probability measures on a topological space, our main concern being the weak topology. Let E be a topological space and let IM(E ) denote the set of F -idempotent probabilities on E , where F is the collection of closed subsets of E . As above we denote by Cb+ (E ) the set of all R+ -valued bounded continuous functions on E . De nition 1.9.1. The weak topology on IM(E ) is the coarsest W topology for which the maps ! E h d are continuous for all h 2 Cb+ (E ). According to the de nition, a base for the weak topology consists W W 0 0 of sets f 2 IM(E ) : j E hi d E hi dj < "; i = 1; : : : ; k g, where 2 IM(E ); hi 2 Cb+ (E ); " > 0: For most of the section we assume that E is a Tihonov topological space, in which case IM(E ) is a Hausdor topological space by Theorem 1.7.27. We denote coniw vergence in the weak topology as ! . Our purpose is to show that

© 2001 by Chapman & Hall/CRC

65

Spaces of idempotent probabilities

the weak convergence of idempotent probabilities has many of the properties of the weak convergence of probability measures. Let us denote by C +b (E ) the set of all upper semi-continuous bounded R+ -valued functions on E , and by C +b (E ) the set of all lower semi-continuous bounded R+ -valued functions on E . For a function h : E ! R+ , let h and h denote the respective upper semicontinuous and lower semi-continuous envelopes of h de ned by

h=

inf

f and h =

f 2C + b (E ): f h

sup g:

g2C + b (E ): gh W

W

We say that h is continuous relative to if E h d = E h d. We say that h is upper semi-continuous (respectively, lower semiW W continuous) relative to if E h d = E h d (respectively, W W h d = E E h d). We adopt a similar terminology for sets. We call a set H E continuous relative to if (int H ) = (cl H ). We call a set H E closed (respectively, open) relative to if (H ) = (cl H ) (respectively, (H ) = (int H )). If E is a Tihonov space, a set is continuous (closed or open, respectively) relative to if and only if its indicator function is continuous (upper semi-continuous or lower semi-continuous, respectively) relative to .

Theorem 1.9.2. (Portmanteau theorem) Let E be a Tihonov topological space. Let 2 IM(E ) and 2 IM(E ); The following conditions are equivalent. iw !

1:

_

2:

E

3: (i)

h d !

lim inf

(ii) lim sup

_

E _ E

_

E

2 ; be a net.

h d for all h 2 Cb+(E )

g d f d

_

E _ E

g d

for all g 2 C +b (E )

f d for all f

2 C +b (E )

30 : The inequalities of part 3 hold for all lower semi-continuous relative to , bounded functions g : E ! R+ and all upper semicontinuous relative to , bounded functions f : E ! R+ , respectively

© 2001 by Chapman & Hall/CRC

66

Idempotent probability measures

4: (i) lim inf (G) (G) (ii) lim sup (F ) (F )

for all open G E for all closed F E

40 : The inequalities of part 4 hold for all open relative to sets G and closed relative to sets F , respectively 5: lim (H ) = (H ) for all continuous relative to sets HE 6:

lim

7:

lim

_

E

_

E

h d = h d =

_

E

_

E

h d

for all continuous relative to bounded functions h : E ! R+

h d

for all bounded functions h : E ! R+ that are uniformly continuous with respect to a given uniformity on E

Proof. Conditions 1 and 2 are equivalent by the de nition of the weak topology. Clearly, 2 ) 7, 3 , 30 , 3 ) 2, 3 ) 4, 30 ) 6, 4 , 40 , 40 ) 5, and 6 ) 2. We prove the implication 2 ) 4. To prove 2 ) 4(i), we note that, since E is Tihonov and G is open, 1(G) = sup h over h 2 Cb+W(E ) such that h 1(G). Therefore, by Theorem 1.4.4 (G) = suph E h d, W so that if h 1(G) is such that (G) E h" d + ", then

lim inf (G) lim

_

E

h" d =

_

E

h" d (G) ":

The proof of 4(ii) is analogous if we note that 1F = inf h over h 2W Cb+(E ) such that h 1(F ) so that by Theorem 1.7.7 (F ) = inf h E h d . We prove that 4(i) ) 3(i) and 4(ii) ) 3(ii). For g 2 C +b (E ) such that kgk = 1 let

gk (z ) = max

hi

i=0;:::;k 1

1 k

g(z ) >

i i ; k 2 N: k

W

Since E gk d = maxi=0;:::;k 1 i=k (g(z ) > i=k) and the sets fz : g(z) > xg are open by the lower semi-continuity of g, 4(i) yields lim inf

_

E

gk d

© 2001 by Chapman & Hall/CRC

_

E

gk d:

67

Spaces of idempotent probabilities

As gk (z ) < g(z ) gk (z ) + 1=k, by Theorem 1.4.4 lim inf

_

E

gd lim inf

_

E

gk d

_

E

gk d

_

E

gd

1 ; k

which yields 3(i). The proof of 4(ii) ) 3(ii) is similar if we consider the functions fk (z ) = maxi=0;:::;k 1 (i + 1)=k 1 f (z ) i=k : Now we prove 5 ) 4. Let G be open andWÆ > 0. Let h be a function from Cb+ (E ) such that h 1(G) and E h d (G) Æ. Let Hu = fz 2 E : h(z ) ug; u 2 [0; 1]: Then the function (Hu ) increases as u # 0 so it has at most countably many jumps. Also W (Hu ) E h d u, so (Hu ) (G) 2Æ for u small enough. Thus, there exists " > 0 such that (H" ) (G) 2Æ and (Hu ) is continuous at ". By -maxitivity of the latter is equivalent to H" being continuous relative to . Thus, we conclude that lim inf (G) lim (H") = (H" ) (G) 2Æ:

The proof of 4(ii) is similar. We prove that 7 ) 4(ii). Let V be a uniformity on E and F be a closed subset of E . Let f g be a collection of pseudo-metrics on E , uniformly continuous with respect to V , which is closed under the formation of maximums and such that 1(F ) = inf ">0 inf (1 (z; F )=")+ , where (z; F ) = inf z0 2F (z; z 0 ). The functions (1 (z; F )=")+ are bounded and uniformly continuous with respect to V so that by Theorem 1.7.7 _

lim sup (F ) inf inf lim (1 (z; F )=")+ d

= inf inf ">0

">0

_

E

(1

E (z; F )=")+ d

=

_

E

inf inf (1 (z; F )=")+ d = (F ):

">0

The implication 7 ) 4(i) is proved in an analogous manner.

Remark 1.9.3. As the proof shows, in part 7 it is enough to require

that the convergences hold for functions h that are Lipshitz continuous with respect to the pseudo-metrics specifying the uniformity.

© 2001 by Chapman & Hall/CRC

68

Idempotent probability measures

Remark 1.9.4. In particular, Theorem 1.9.2 implies that the weak 0

topology on IM(E ) is also generated by the subbase f 2 IM(E ) : 0 (G) > (G) "g; f0 2 IM(E ) : 0 (F ) < (F ) + "g; where the G are open, F are closed, " > 0, 2 IM(E ), as well as by the subbase f0 2 IM(E ) : j0(H ) (H )j < "g, where the H are continuous relative to , " > 0; 2 IM(E ). Remark 1.9.5. The de nition of the weak topology also applies to arbitrary nite F -idempotent measures. Theorem 1.9.2 is retained. Remark 1.9.6. For general Hausdor topological spaces the convergences in part 4 of Theorem 1.9.2, which specify the narrow topology, see, e.g., O'Brien and Vervaat [97], imply convergence in the weak topology. On the other hand, if we de ned the weak topology in analogy with Topse [125] by requiring that it be the weakest topology W such that the evaluations ! E g d areW lower semi-continuous for all g 2 C +b (E ) and the evaluations ! E f d are upper semicontinuous for all f 2 C +b (E ), then the weak topology would be equivalent to the narrow topology. Also for this topology the requirement of E being Tihonov in Theorem 1.9.17 below can be relaxed. Corollary 1.9.7. Let E be a Tihonov topological space. Let E0 be a subset of E equipped with relative topology. Let 2 IM(E ) and 2 IM(E ) be such that (E n E0 ) = (E n E0 ) = 0 and the ~ and ~, restrictions of and to E0 , which are denoted by respectively, are -smooth relative to the collection of closed subsets iw iw ~ ~ ! of E0 . Then ! if and only if . Remark 1.9.8. The -smoothness property in the hypotheses holds if either E0 is a closed subset of E or the and are tight F idempotent probabilities on E . We next give suÆcient conditions for continuity relative to and the other related notions. Let E0 E . De nition 1.9.9. We say that a set H E is E0 {closed if it contains all its accumulation points that are in E0 , i.e., cl H \ E0 H . We say that a set H E is E0 {open if every point of H \ E0 is an interior point of H , i.e., H \ E0 int H . Remark 1.9.10. Note that both the interior and closure are taken in E . Also, H is E0 {open if and only if its complement in E is E0 {closed.

© 2001 by Chapman & Hall/CRC

69

Spaces of idempotent probabilities

De nition 1.9.11.

A function h : E ! R+ is called E0 {upper (respectively, E0 {lower) semi-continuous if the sets fz 2 E : f (z ) ag (respectively, fz 2 E : f (z ) ag), a 2 R+ , are E0 {closed. A function h : E ! R+ is said to be E0 {continuous if h 1 (G) is E0 { open for each open G R+ .

Remark 1.9.12. An indicator function 1(A), A E , is E0{upper (respectively, E0 {lower) semi-continuous if and only if A is E0 {closed (respectively, E0 {open).

Remark 1.9.13.

If E is Hausdor, then a function h is E0 upper (respectively, E0 -lower) semi-continuous if and only if lim sup h(z ) h(z ) (respectively, lim inf h(z ) h(z )) for every 2 2 net z ! z 2 E0 . Similarly, h : E ! R+ is E0 {continuous if and only if lim h(z ) = h(z ) for every net z ! z 2 E0 . 2 Let us say that is supported by E0 if (E n E0 ) = 0.

Lemma 1.9.14. Let E be Hausdor and be supported by E0.

1. If a function h : E ! R+ is E0 {continuous (E0 {uppersemi-continuous or E0 {lower-semi-continuous, respectively), then it is continuous (upper semi-continuous or lower semicontinuous, respectively) relative to . 2. If a set H E is E0 {continuous (E0 {closed or E0 {open, respectively), then it is continuous (closed or open, respectively) relative to .

Proof. Part 1 follows by the fact that on Hausdor spaces h(z ) = lim sup h(z 0 ) and h(z ) = lim inf h(z 0 ); U 2Uz z 0 2U U 2Uz z 0 2U

where Uz is the collection of open neighbourhoods of z ordered by inclusion. Part 2 is a consequence of the de nitions. Our next goal is to prove a Prohorov criterion of weak relative compactness. We denote by IMt (E ) the set of tight F -idempotent probabilities on E .

De nition 1.9.15.

1. A subset inf K 2K sup2A (K c ) = 0.

© 2001 by Chapman & Hall/CRC

A of IMt(E )

is called tight if

70

Idempotent probability measures

2. A net f ; 2 g in inf K 2K lim sup2 (K c ) = 0.

IM(E )

is called tight if

De nition 1.9.16. A net f ; 2 g in IM(E ) is called relatively compact if every subnet of f ; 2 g contains a weakly convergent subsubnet.

We also use the standard de nition that a subset of IM(E ) is relatively compact for the weak topology if its closure is compact.

Theorem 1.9.17.

1. Let E be a Tihonov topological space. If a subset A of IMt (E ) (respectively, a net f ; 2 g in IM(E )) is tight, then A (respectively, f ; 2 g) is relatively compact, the accumulation points being elements of IMt(E ).

2. Let E be homeomorphic to a complete metric space. If a subset A of IM(E ) is relatively compact, then A is tight. 3. Let E be locally compact and Hausdor. If a subset A of IM(E ) (respectively, a net f ; 2 g in IM(E ) ) is relatively compact, then A (respectively, f ; 2 g) is tight. Proof. We prove part 1 by proving that every tight net f ; 2 g in IM(E ) contains a subnet converging to an element of IMt (E ). Let Cb;+1 (E ) = ff 2 Cb+ (E ) : kf k 1g. The mapping ! W ( E f d; f 2 Cb;+1 (E )) de nes a homeomorphism between space + IM(E ) and a subspace of space [0; 1]Cb;1 (E) with product topology. The latter space being compact by Tihonov's theorem and Hausdor, there exists a subnet f0 ; 0 2 0 g of f ; 2 g that converges to + + an element of [0; 1]Cb;1 (E ) . By the de nition of topology on [0W; 1]Cb;1 (E ) and properties of idempotent integral this implies that E f d0 converges for every f 2 Cb+ (E ). Denoting the limits by V (f ) we conclude in view of Theorem 1.4.4 that the functional f ! V (f ) has properties (V 0){(V 2). Tightness of f ; 2 g implies that the functional is tight in the sense of Theorem 1.7.25. Thus, the functional V (f ) satis es all theWconditions of Theorem 1.7.25; according to the theorem V (f ) = E f d; f 2 Cb+ (E ); for some tight iw F -idempotent probability , which implies that 0 ! . This completes the proof of part 1.

© 2001 by Chapman & Hall/CRC

71

Spaces of idempotent probabilities

For part 2, we may assume that E is a complete metric space. Also replacing A by its closure, we may assume that A is a compact subset of IM(E ). We rst show that for all " > 0 and Æ > 0 there exist open Æ-balls A1 ; : : : ; Ak such that

En

k [ i=1

Ai < "; 2 A:

(1.9.1)

Since each 2 A is tight by Theorem 1.7.8, we can choose compacts K in E such that (E nK ) < "=2. Let B;1 ; : : : ; B;l be open Æ-balls that cover K so that

En

l [ i=1

B;i < "=2; 2 A:

(1.9.2)

Let n

l [

i=1 l [

G = 0 2 IM(E ) : 0 E n < En

i=1

B;i

B;i +

"o ; 2 A: (1.9.3) 2

Since by Remark 1.9.4 fG ; 2 Ag is an open cover of the compact set A, there exist G1 ; : : : ; G p that also cover A so S P 2 pj=1 G j ; 2 A: Then denoting k = pj=1 l j and Cj = Slj i=1 Bj ;i ; j = 1; : : : ; p; and taking A1 = B1 ;1 ; : : : ; Al1 = B1 ;l1 ; Al1 +1 = B2 ;1 ; : : : ; Ak = Bp ;lp we have by (1.9.3) and (1.9.2)

En

k [ j =1

Aj

j=1min (E nCj ) ;:::;p j=1 max [j (E nCj ) + "=2] < "; 2 A; ;:::;p

which is the required property. Therefore for arbitrary " > 0 and k = 1; 2; : : : ; there exist open 1=k-balls Ak1 ; : : : ; Aknk such that

En

nk [

i=1

Aki < "; 2 A:

© 2001 by Chapman & Hall/CRC

(1.9.4)

72

Idempotent probability measures

T Snk The set A = 1 k=1 i=1 Aki is totally bounded and hence relatively compact by completeness of E . At the same time, by (1.9.4) and -maxitivity of

nk [

(E nA) = sup E n Aki k2N i=1

"; 2 A;

i.e., the set A is tight. Now let E be locally compact and Hausdor. The case of a relatively compact set A is tackled similarly to part 2 in that one can show that there exist open sets Ai with compact closures such that (1.9.1) holds. Now, let a net f ; 2 g be relatively compact in IM(E ). It is suÆcient to prove that for every " > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that

k [

lim sup E n Ai 2 i=1

":

(1.9.5)

We introduce a partial order on the collection O of nite unions O of open sets with compact closures so that O O0 if O O0 . Let denote the set of pairs (; O). We turn into a directed set by de ning that (; A) (0 ; A0 ) if 0 and A A0 . We denote (E n O) = x . Let fx 0 ; 0 2 0 g be a subnet of fx ; 2 g such that lim supO2O lim sup2 x = lim 0 2 0 x 0 (cf. Kelley [71]). The mapping from 0 to in the de nition of a subnet induces a mapping from 0 to by associating with elements of 0 the rst components of the corresponding elements of . This de nes a subnet f 0 g of . Since for every O 2 O the second component in ( 0 ) contains O for all 0 large enough, we have that lim 0 2 0 x 0 lim sup 0 2 0 0 (E n O): Let f 00 ; 00 2 00 g be a subnet of f 0 ; 0 2 0 g that weakly converges to a deviability . De ning x 00 = x 0 , where 0 is the image of 00 under the mapping from 00 to 0 in the de nition of a subnet, we conclude that fx 00 g is a subnet of fx 0 g. Therefore, for arbitrary O 2 O by Theorem 1.9.2 (E nO) lim sup 00 (E nO) 00lim 00 x 00 = lim sup lim sup x : 00 2 00 2 O2O 2 Since is -smooth relative to the collection of closed subsets of E and the union of the sets O equals E by local compactness of E , it follows that limO2O (E n O) = (;) = 0; so we conclude that lim supO2O lim sup2 (E nO) = 0; as claimed.

© 2001 by Chapman & Hall/CRC

Spaces of idempotent probabilities

73

We now assume that E is a metric space with metric and prove that the weak topology on IM(E ) is metrisable. The next lemma has been proved in the proof of Theorem 1.7.8. Lemma 1.9.18. If 2 IM(E ), then is separable in the sense that for every > 0 and Æ > 0 there exists a nite collection of open k Æ-balls A1 ; A2 ; : : : ; Ak such that E n [i=1Ai < . We now de ne an idempotent analogue of the Prohorov metric. We denote by B (z ) the closed ball of radius about z , by A the closed {neighbourhood of a set A, and let A = E n (E n A) . De nition 1.9.19. Given ; 0 2 IM(E ), we de ne p(; 0 ) = inf > 0 : (z ) 0 B(z ) + ; 0 (z ) B (z ) + for all z 2 E :

It is not diÆcult to check that p is a metric on IM(E ). The next lemma follows by -maxitivity of an idempotent measure. Lemma 1.9.20. Let ; 0 2 IM(E ). Then p(; 0 ) = inf > 0 : (A) 0 A + ; 0 (A) A + for all A E :

Remark 1.9.21. One obtains the same metric as p if, as in the standard de nition of the Prohorov metric, one considers open rather than closed -neighbourhoods of z and A, respectively. Theorem 1.9.22. The metric p is compatible with the weak topology on IM(E ). Proof. We rst prove that the topology induced by p is ner than the weak topology. By Remark 1.9.4 it is suÆcient to prove that, given 2 IM(E ), a closed set F , an open set G, and > 0, there exists Æ > 0 such that f0 2 IM(E ) : p(; 0 ) < Æg f0 2 IM(E ) : 0(F ) < (F ) + g

and

f0 2 IM(E ) : p(; 0 ) < Æg f0 2 IM(E ) : 0 (G) > (G) g:

© 2001 by Chapman & Hall/CRC

74

Idempotent probability measures

Since is -smooth relative to F , there exists Æ 2 (0; =2) such that (F ) (F Æ ) =2. Therefore, if p(; 0 ) < Æ, then 0 (F ) < (F Æ ) + Æ (F ) + proving the rst inclusion. For the second inclusion, using -maxitivity of , we choose Æ 2 (0; =2) such that (G) (G Æ )+ =2. Then, if p(; 0 ) < Æ, then (G) (G Æ )+ =2 < 0 (G) + . Conversely, we show using again Remark 1.9.4 that given and > 0 there exists a collection Hi ; i = 1; : : : ; k of sets, which are continuous relative to , and Æ > 0 such that

f0 2 IM(E ) : j0(Hi) (Hi)j < Æ; i = 1; : : : ; kg f0 2 IM(E ) : p(; 0 ) < g: Let Æ < =3. By separability of there exist closed Æ=2-balls B1 ; : : : ;Bk 1 centred at z1 ; : : : ; zk 1 , respectively, such that E n [ki=11Bi < Æ. By -maxitivity of for each i = 1; 2; : : : ; k 1 there exists a closed ball Hi centred at zi of radius not less than Æ=2 and not greater than Æ, which is a continuous set relative to . We also Æ0 k 1 take Hk = E n[i=1 Bi , where Æ0 > 0 is chosen such that Hk is continuous relative to and (Hk ) 2Æ. Let 0 2 IM(E ) be such that j0 (Hi) (Hi)j < Æ; i = 1; : : : ; k. If z 2 Hi for some i = 1; : : : ; k 1, then (z ) (Hi ) < 0 (Hi ) + Æ 0 (B2Æ (z )) + Æ. Similarly, 0 (z ) < (B2Æ (z )) + Æ. If z 62 [ki=11Hi , then (z ) < Æ; also, since E n [ki=11 Hi Hk , we have that 0 (z ) 0 (Hk ) < (Hk ) + Æ 3Æ. We thus conclude that p(; 0 ) < . We next show that the weak topology on IM(E ) is also metrised by a Kantorovich-Wasserstein metric. For f 2 Cb+ (E ) let 0 kf k BL = sup f (z) _ sup0 jf (z)(z; zf0()z )j : z 2E z 6=z Clearly, if kf k BL < 1, then f is bounded and Lipshitz-continuous. De nition 1.9.23. For ; 0 2 IM(E ), we let

BL (; 0 ) =

sup

_ _ f d0 : f d

f 2Cb+ (E ): E kf k BL 1

© 2001 by Chapman & Hall/CRC

E

75

Spaces of idempotent probabilities

It is not diÆcult to check that BL is a metric on IM(E ). Lemma0 1.9.24. Let and 0 be in IM(E ). Then BL (; 0 ) 2p(; ).

Proof. We have for f such that kf k BL 1 and Æ > 0 _ f d

E

_

_

f d0 sup

_

f d

f d0

z 2E B (z ) E BÆ (z ) Æ _ _ sup f d f (z )0 (z ) f (z )(z ) f d0 z 2E B (z ) BÆ (z ) Æ 0 Æ + sup (z ) BÆ (z ) BÆ (z ) 0 (z ) ;

_

z 2E

_

yielding the required. As a consequence, we have the following result.

Theorem 1.9.25. The metric BL metrises the weak topology on IM(E ).

Proof. By Theorem 1.9.2 and Remark 1.9.3 the convergence BL ( ; ) ! 0 implies the convergence ! . Thus, the topology induced by the metric BL is ner than the weak topology. The converse follows by Theorem 1.9.22 and Lemma 1.9.24.

Since on a metric space the notions of sequential compactness and compactness are identical, metrisability of IM(E ) allows us to give criteria for sequential compactness. We rst recall relevant de nitions.

De nition 1.9.26. A subset A of IM(E ) is called relatively sequentially compact (for the weak topology) if every sequence fn ; n 2 N g of elements of A contains a weakly convergent subsubsequence.

Combining Theorem 1.9.17 and Theorem 1.9.22 yields the following result.

Theorem 1.9.27. Let E be a metric space.

1. If a subset A of IM(E ) is tight, then A is relatively sequentially compact, the accumulation points being elements of IMt (E ).

© 2001 by Chapman & Hall/CRC

76

Idempotent probability measures

2. Let E be homeomorphic to a complete metric space. If a subset A of IM(E ) is relatively sequentially compact, then A is tight.

We give, however, a proof of part 1 that does not use Theorem 1.9.17. Proof. Let n 2 IM(E ); n 2 N . Let us assume rst that E is a compact metric space. The set Cb;+1 (E ) of R+ -valued continuous functions on E that are bounded by 1, endowed with the topology of uniform convergence, is a separable metric space. Let Cb;+1;d (E ) + denote its countable dense subset. The set [0; 1]Cb;1;d (E ) with product topology is sequentially compact, so by a diagonal argument there W exists a subsequence nk such that the sequences f E f dnk ; k 2 N g converge for all f 2 Cb;+1;d (E ). Since Cb;+1;d (E ) is dense in Cb;+1 (E ), it follows by properties of idempotent integral (more speci cally, by W part (JS 5) of Theorem 1.4.4) that the sequences f E f dnk ; k 2 N g converge for all f 2 Cb;+1 (E ), which implies in analogy with the proof of Theorem 1.9.17 that there exists an F -idempotent probability iw on E such that nk ! as k ! 1. Let us now assume that E is a separable metric space. Then it is embedded into a compact metric space E 0 . We extend idempotent probabilities on E to idempotent probabilities on E 0 by letting 0 (A) = A0 \ E ; A0 E 0 : Let f0nk g be a subsequence of f0n ; n 2 N g which weakly converges to a deviability 0 on E 0 . Tightness of fn ; n 2 N g implies that the collection of compact subsets of E is a tightening collection for 0 so that the set function de ned by (A) = 0 (A); A E; is a tight F idempotent probability on E . Also 0 (E 0 n E ) = 0. We check that iw nk ! . Let f be a uniformly continuous function from Cb+ (E ) and f 0 be an element of Cb+(E 0 ) that extends f . By Theorem 1.9.2 W W 0 0 0 0 f d0 . SinceWthe 0nk and 0 are supported by E , E 0 f dnk ! EW we conclude that E f dnk ! E f d, which proves the required by Theorem 1.9.2. Now, if E is an arbitrary metric space, then by the tightness condition there exists a -compact metric space E 0 E such that n (E n E 0 ) = 0 for all n. Since E 0 is separable in relative topology, applying the part just proved to the restrictions 0n of the n to E 0 , we deduce existence of a subsequence f0nk g that weakly converges to a deviability 0 on E 0 . Let (A) = 0 (A \ E 0 ); A E . Then

© 2001 by Chapman & Hall/CRC

77

Spaces of idempotent probabilities

is a tight F -idempotent probability on E . Finally, if f 2 Cb+ (E ), then its restriction f 0 to E 0 belongsWto Cb+ (E 0 ). SinceW 0n (E n E 0 ) = 0 (E n E 0 ) = 0, we have that 0 0 = dnk and nk EfW E 0 f dW W W 0 0 0 0 0 0 f d = f d . Thus, convergence f d ! nk E0 E W E0 E 0 f d W yields convergence E f dnk ! E f d. A modi cation of the argument used in the proof of Lemma 1.9.24 allows us to obtain the following result. Theorem 1.9.28. Let f ; 2 g be a net of F -idempotent probabilities on a Tihonov space E that weakly converges to 2 IM(E ). Let G be a subset of Cb+ (E ) consisting of uniformly bounded and pointwise equicontinuous functions, i.e., supf 2G supz2E f (z ) < 1; and for every > 0 and z 2 E there exists an open neighbourhood Uz of z such that supf 2G supz0 2U jf (z ) f (z 0 )j : Then _ lim sup f d

f 2G E

_

E

z f d = 0:

Proof. We x > 0. For each z 2 E let Uz be as in the statement of the theorem. We show that the Uz can be assumed to be continuous relative to . Let fz be continuous functions with values in [0; 1] such thatfz (z ) = 1 and fz (z 0 ) = 0 on E n Uz . The function fz 1 ((x; 1]) ; x 2 [0; 1); is monotonically decreasing, so it has 1 (x; 1] is an open -continuous set if continuity points. Since f z fz 1((x; 1]) is continuous at x by -smoothness of , the claim has been proved. Let Uz1 ; : : : ; Uzk be such that E n[ki=1 Uzi < . Then, denoting a = supf 2G supz2E f (z ), we have _ f d

E

+ max

i=1;:::;k

_

E

f d

E n[ki=1 Uzi _ f d

_ f d

Uzi

_

f d +

_

E n[ki=1 Uzi

f d

Uzi

a E n [ki=1Uzi + a + 2 i=1 max sup jf (z ) f (zi )j ;:::;k z 2Uzi

+ max f (zi )j (Uzi ) i=1;:::;k

(Uzi )j:

Since by \the Portmanteau theorem" the latter maximum converges to 0 as 2 and lim sup2 E n[ki=1 Uzi E n[ki=1 Uzi < ; the proof is complete.

© 2001 by Chapman & Hall/CRC

78

Idempotent probability measures

If we replace space Cb+(E ) in the de nition of the weak topology by space CK+ (E ) of R+ -valued continuous functions with compact support, then we arrive at the de nition of the vague topology. However, to obtain nice properties, we have to consider the space of K{idempotent measures.

De nition 1.9.29. The vague topology on the set of K{idempotent

measures on a W topological space E is the coarsest topology for which the maps ! E h d are continuous for all h 2 CK+(E ).

If E is locally compact and Hausdor the vague topology has properties similar to the above properties of the weak topology. For instance, the space of K{idempotent measures is a Hausdor topological space and there is an easy analogue of Theorem 1.9.2. A distinguishing feature of the vague topology is that the space of K{ idempotent measures is compact.

Theorem 1.9.30. Let E be a locally compact Hausdor topologi-

cal space. Then the space of K{idempotent measures with the vague topology is compact.

The proof is similar to the proof of part 1 of Theorem 1.9.17, the main distinction being the use of Theorem 1.7.21 in place of Theorem 1.7.25. We end the section by indicating a connection between the vague and weak topologies, on the one hand, and Mosco convergence, on the other hand. The proof is straightforward.

Theorem 1.9.31.

I. Let E be a locally compact Hausdor topological space. Let be a K-idempotent probability and f ; 2 g be a net of K-idempotent probabilities on E . The following pairs of conditions are equivalent:

(M ) 1. for every z 2 E and net z ! z , lim sup2 (z ) (z ); 2. for every z 2 E there exists a net z ! z such that lim2 (z ) = (z ); (V ) 1. for every compact set K E , lim sup2 (K ) (K ); 2. for every open set G E , lim inf 2 (G) (G):

© 2001 by Chapman & Hall/CRC

Derived weak convergence

79

II. Let E be a Hausdor topological space. Let be a tight F idempotent probability and f ; 2 g be a tight net of F idempotent probabilities on E . Then the next pair of conditions is equivalent to both of the above:

(N ) 1. for every closed set F E , lim sup2 (F ) (F ); 2. for every open set G E , lim inf 2 (G) (G):

1.10 Derived weak convergence The results of this section give conditions for weak convergence of idempotent probabilities that are derived from weakly convergent idempotent probabilities. We also introduce convergence in idempotent distribution as an alternative way of viewing weak convergence of idempotent probabilities.

De nition 1.10.1. Let fX ; 2 g be a net of idempotent variables

de ned on respective idempotent probability spaces ( ; ) and assuming values in a topological space E and X be an idempotent variable de ned on an idempotent probability space ( ; ) and assuming values in E . Let the idempotent distributions of the X and X be F -idempotent probabilities on E . We say that the net fX ; 2 g iw converges in idempotent distribution to X if Æ X 1 ! Æ X 1. id We denote convergence in idempotent distribution by ! . Since convergence in idempotent distribution is the weak convergence of induced idempotent laws, the theory of Section 1.9 applies, e.g., there is a version of the Portmanteau theorem. Depending on the concrete situation it can be more convenient to formulate results on weak convergence of idempotent laws as convergence in idempotent distribution or vice versa.

Lemma 1.10.2. Let E be a Tihonov topological space, and and iw be F -idempotent probabilities on E such that ! . Let functions h : E ! R+ be uniformly bounded and a function h : E ! R+ be such that

lim h (z ) = h(z )

2

© 2001 by Chapman & Hall/CRC

80

Idempotent probability measures

for -almost every z 2 E and every net z ! z as 2 . Then

lim 2

_

E

h (z ) d (z ) =

_

E

h(z ) d(z ):

Proof. Let

h (z ) = inf sup sup h (z 0 );

(1.10.1)

U 2Uz z 0 2U 0

where Uz denotes the set of open neighbourhoods of z . Since the convergence condition in the hypotheses equivalently requires that for every z 2 E such that (z ) > 0 and > 0 there exist an open neighbourhood U of z and such that jh0 (z 0 ) h(z )j < for all z 0 2 U and 0 , we conclude that, given > 0 and z 2 E such that (z ) > 0, there exists such that h0 (z ) h(z ) + for all 0 . Therefore, introducing h(z ) = inf 2 h (z ), we have that h is upper semi-continuous and h(z ) h(z ) -a.e. Also, since the net fh g consists of upper semi-continuous bounded functions, is monotonically decreasing and converges to h, by Theorem 1.7.7 W W lim2 E h dW = E h dWso that for arbitrary > 0 there exists 0 such that E h0 d E h d + : Using the fact that h h0 ; 0 ; by (1.10.1) and Theorem 1.9.2 applied to h0 we obtain _

_

lim sup h (z ) d (z ) lim sup h0 (z ) d (z ) 2 E 2 E _ _ _ h0 (z) d(z) h(z) d(z) + h(z) d(z) + : E

E

E

The complementary inequality lim inf 2

_

E

h (z ) d (z )

_

E

h(z ) d(z )

is proved by a symmetric argument. Speci cally, we de ne

h (z ) = sup inf inf h (z 0 ); h(z ) = sup h (z ) U 2Uz z 0 2U 0 2 semi-continuous, h h0 if 0 , and note that the h are lower W W h h -a.e., and lim2 E h d = E h d. Therefore, for an

© 2001 by Chapman & Hall/CRC

81

Derived weak convergence

arbitrary > 0 and suitable 1 lim inf 2 _

_

E

h (z ) d (z ) lim inf _

2

_

E

h1 (z ) d (z )

h1 (z) d(z) h(z) d(z) E

E

_

E

h(z ) d(z ) :

As a consequence, we obtain the following version of the continuous mapping theorem on preservation of weak convergence of probability measures under mappings. We formulate it in terms of convergence in idempotent distribution. Theorem 1.10.3. Let E and E 0 be Tihonov topological spaces, and X and X be Luzin idempotent variables with values in E such that id X ! X . Let functions f : E ! E 0 , 2 , be Luzin measurable relative to the respective idempotent distributions of the X and f : E ! E 0 be Luzin measurable relative to the idempotent distribution of X . If for almost every z 2 E with respect to the idempotent distribution of X and every net z ! z we have that f(z ) ! f (z ), id then f Æ X ! f ÆX . Proof. Let and denote the respective idempotent distributions of X and X on E . Since the idempotent distribution of f Æ X is Æ f 1 and the idempotent distribution of f Æ X is Æ f 1 , for an R+ -valued bounded continuous function h on E 0 by a change of variables and Lemma 1.10.2 _ _ lim h(z 0 ) d Æ f 1 (z 0 ) = lim h Æ f (z ) d (z ) 2 0 2 E E _ _ = h Æ f (z ) d(z ) = h(z 0 ) d Æ f 1(z 0 ): E

E0

We thus have the following \continuous mapping theorem". Corollary 1.10.4. Let E and E 0 be Tihonov topological spaces, and id X and X be Luzin variables with values in E . If X ! X as 2 and f : E ! E 0 is Luzin measurable with respect to the distribution of X for every 2 and continuous a.e. with respect id to the distribution of X , then f Æ X ! f Æ X.

© 2001 by Chapman & Hall/CRC

82

Idempotent probability measures

We denote by Li (X ) the idempotent distribution of an idempotent variable X .

Lemma 1.10.5. Let E be a metric space with metric , and let X

and Y ; where 2 , 2 , and are directed sets, be nets of idempotent variables with values in E de ned on respective idempotent probability spaces ( ; ), whose idempotent distributions are F -idempotent probabilities on E . Let

lim lim sup (X ; Y ) " = 0; " > 0; 2 2 and

iw ~ Li X !

as 2 ,

~ ; 2 ; are F -idempotent probabilities on E . where Then, for an F -idempotent probability on E , we have that iw Li(Y ) ! as 2

if and only if

~

iw ! as 2 :

iw ~ ! Proof. We prove suÆciency of the condition so we assume that . Let F be a closed subset of E . Since

(Y 2 F ) (X 2 F )+ (X ; Y ) " ; by hypotheses ~ (F ) lim sup (Y 2 F ) lim sup 2 2 + lim sup lim sup ((X ; Y ) ") (F ); 2 2 and hence by the -smoothness property of lim sup (Y 2 F ) (F ): 2

© 2001 by Chapman & Hall/CRC

(1.10.2)

Derived weak convergence

Let G be an open subset of E . Then fX 2 G " g fY f(X ; Y ) "g; hence, since the G " are open as well,

83

2 Gg [

lim inf (Y 2 G) lim inf lim inf (X 2 G " ) 2 2 2 lim sup lim sup ((X ; Y ) ") (G " ): 2 2 Observing that [">0 G " = G so that by -maxitivity (G " ) ! (G) as " ! 0, we conclude that lim inf (Y 2 G) (G); 2 which together with (1.10.2) ends the proof of the suÆciency part. The converse is proved in an analogous manner. The following special case is useful. Given a net fZ ; 2 g of idempotent variables de ned on ( ; ) and assuming values in a metric space E with metric , we write that Z ! z 2 E , if lim2 (Z ; z ) > = 0 for every > 0. Lemma 1.10.6. Let E be a metric space with metric , and let X and Y ; where 2 , be nets of idempotent variables on ( ; ) with values in E , whose idempotent distributions are F -idempotent iw probabilities on E . If Li (X ) ! , where is an F -idempotent iw probability on E , and (X ; Y ) ! 0 as 2 , then Li (Y ) ! . Lemma 1.10.7. Let fX ; 2 g be a net of idempotent variables and X be an idempotent variable. Let all the variables be de ned on ( ; ), assume values in a metric space E and have F -idempotent id probabilities on E as distributions. If X ! X , then X ! X . If id X ! z 2 E , then X ! z . Proof. The rst property follows by convergence properties of idempotent integrals (Theorem 1.4.19). For the second one, let denote the metric on E , X denote the idempotent distribution of X , and 1z denote the unit mass at z. Then the convergence X !id z implies that _ _ lim 1^(z 0 ; z ) dX (z 0 ) = 1^(z 0 ; z ) d 1z (z 0 ) = 0: 2 E E

© 2001 by Chapman & Hall/CRC

84

Idempotent probability measures

We now consider joint convergence. Lemma 1.10.8. Let X and Y ; 2 , be nets of idempotent variables on respective idempotent probability spaces ( ; ) with values in Tihonov spaces E and E 0 , respectively. Let X and Y be idempotent variables on an idempotent probability space ( ; ) with values in E and E 0 , respectively. Let the idempotent distributions of the X , Y ; X , and Y be -smooth relative to the associated collections of closed sets. Let E E 0 be equipped with product topology. id id 1. If X ! X, Y ! Y , X and Y are independent, and X and id Y are independent, then (X ; Y ) ! (X; Y ).

id 2. Let E and E 0 be metric spaces. If X ! X and Y ! z , then id (X ; Y ) ! (X; z ). Proof. We prove part 1. Let h(z; z 0 ); (z; z 0 ) 2 E E 0 ; be an R+ valued bounded function that is uniformly continuous with respect to a uniformity on E E 0 . By Theorem 1.9.28 lim sup sup h(z; z 0 ) ÆY 1 (z 0 ) sup h(z; z 0 )ÆY 1 (z 0 ) = 0: 2 z 2E z 0 2E 0 z 0 2E 0 (1.10.3) Since supz0 2E 0 h(z; z 0 ) Æ Y 1 (z 0 ) is continuous in z 2 E , 1 lim sup sup h(z; z 0 ) Æ Y 1 (z 0 ) Æ X (z ) 2 z 2E z 0 2E 0 = sup sup h(z; z 0 ) Æ Y 1 (z 0 ) Æ X 1 (z ): (1.10.4) z 2E z 0 2E 0 Equations (1.10.3) and (1.10.4) imply the required. For part 2 we observe that ( 0 )((X ; Y ); (X ; z )) ! 0, where 0 is a product metric on E E 0 , so that the required follows by part 1 and Lemma 1.10.6.

1.11 Laplace-Fenchel transform This section introduces idempotent analogues of the characteristic function and standard probability distributions, and develops related techniques. Let ( ; ) be an idempotent probability space. We denote the idempotent distribution Æf 1 of an idempotent variable f : ! Rd by f .

© 2001 by Chapman & Hall/CRC

85

Laplace-Fenchel transform

Remark 1.11.1. Clearly, given an idempotent probability ^ on Rd ,

we can always construct an idempotent variable whose idempotent ^ . This is the \canonical" idempotent variable f (x) = distribution is x. We refer to representations like this as \canonical settings".

We recall that for d-dimensional vectors x and y we denote as x y the inner product.

De nition 1.11.2. Given f : ! Rd , the Laplace-Fenchel transform of f is the R + -valued function Lf () = Sef (!) ; 2 Rd :

By Theorem 1.4.6 we can also write _ Lf () = ex df (x): Rd

Remark 1.11.3. Note that ln Lf () isf the convex conjugate, or the Legendre-Fenchel transform, of

ln (x) in that

ln Lf () = sup ( x +ln f (x)): x2Rd

Lemma 1.11.4.

An R + -valued function L(); 2 Rd ; is the Laplace-Fenchel transform of an Rd -valued idempotent variable if and only if ln L() is convex and lower semi-continuous, and L(0) = 1. Proof. Necessity of the conditions follows from the de nition of the Laplace-Fenchel transform and properties of convex conjugates. Conversely, let ln L() be convex and lower semi-continuous, and L(0) = 1. We de ne (x) by

ln (x) = sup ( x ln L()); 2 Rd : 2Rd Then is an idempotent probability and L is the Laplace-Fenchel transform of the canonical variable on (Rd ; ) by properties of convex conjugates, Rockafellar [117, x26]. According to the above proof, we can recover f from Lf by the equality f (x) = infd e x Lf () (1.11.1) 2R

© 2001 by Chapman & Hall/CRC

86

Idempotent probability measures

if we know, in addition, that ln f (x) is lower semi-continuous and convex. We refer to (1.11.1) as the inversion formula. It is useful, however, to have conditions for the inversion formula to hold that are expressed only in terms of the properties of Lf . We recall some notions of convex analysis, Rockafellar [117]. Let a function g(); 2 Rd ; assume values in ( 1; 1]. The domain of g as de ned as dom g = f 2 Rd : g() < 1g and the function g is said to be essentially smooth if the following conditions hold, Rockafellar [117, x26], (a) int(dom g) is not empty, (b) g is dierentiable on int(dom g), (c) limk!1 jrg(k )j = 1 whenever fk g is a sequence of elements of int(dom g) converging to a boundary point of int(dom g). We also denote as ri A the relative interior of a set A.

Lemma 1.11.5. Let Lf (); 2 Rd ; be essentially smooth. Then the inversion formula holds.

Proof. Let (x) denote the right-hand side of (1.11.1). Since ln (x); x 2 Rd ; is the convex conjugate of ln Lf () and the latter is convex and lower semi-continuous, it follows that ln Lf () is the convex conjugate of ln (x), Rockafellar [117, x26]. Since ln Lf () is essentially smooth, we conclude that ln (x) is essentially strictly convex, Rockafellar [117, Theorem 26.3], hence, strictly convex on ri(dom ln ). Since also ln (x) is the bipolar of ln f (x), the two functions coincide by Lemma A.1 in Appendix A.

Remark 1.11.6. As the proof shows, one can weaken the requirement of essential smoothness of Lf to the requirement that the convex conjugate of ln Lf () be strictly convex on the relative interior of its domain.

We have a simple corollary, which shows that the Laplace-Fenchel transform can help us to identify Luzin idempotent variables.

Lemma 1.11.7. If Lf (); 2 Rd ; is essentially smooth and 0 2 int(dom Lf ), then f is a Luzin idempotent variable.

© 2001 by Chapman & Hall/CRC

87

Laplace-Fenchel transform

Proof. By the inversion formula f (x) is upper semi-continuous being the in mum of continuous functions of x so that by Lemma 1.7.4 f is a K-idempotent probability. By \the Chebyshev inequality" for a > 0 and > 0 such that the -ball about the origin in Rd belongs to int(dom Lf )

f (jxj a) = sup f (x a) e 2Rd : jj=

a

sup Lf ():

2Rd : jj=

The latter supremum being nite by continuity of Lf () on int(dom Lf ), the right-most side tends to 0 as a ! 1. Thus, f is tight. The following property provides us with a means of proving independence of idempotent variables on . For idempotent variables f1 : ! Rd1 and f2 : ! Rd2 , we denote by Lf1 ;f2 (1 ; 2 ); 1 2 R d1 ; 2 2 Rd2 ; the Laplace-Fenchel transform of (f1 ; f2 ) : ! R d1 +d2 .

Lemma 1.11.8.

1. If f1 and f2 are independent, then Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ):

2. Let Lf1 () and Lf2 () be essentially smooth. If Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ), then f1 and f2 are independent. Proof. The rst part follows by de nition. The second part follows by Lemma 1.11.5 for denoting by f1 ;f2 (x1 ; x2 ) the joint idempotent distribution of f1 and f2 , we have by the lemma since Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ) is essentially smooth as well

f1 ;f2 (x1 ; x2 ) = infd e 1 x1 2 x2 Lf1 ;f2 (1 ; 2 ) 1 2R 1 ; 2 2Rd2 x = infd e 1 1 Lf1 (1 ) infd e 2 x2 Lf2 (2 ) 1 2R 1 2 2R 2 = f1 (x1 )f2 (x2 ):

Corollary 1.11.9. Let A be a -algebra on and f : ! Rd .

If S exp( f )jA (!) is constant for -almost ! and is an essentially smooth function of 2 Rd , then f is independent of A.

© 2001 by Chapman & Hall/CRC

88

Idempotent probability measures

Proof. From the hypotheses and properties of conditional idempotent expectations, for 1 2 Rd ; 2 2 R and A 2 A,

S exp(1 f ) exp(2 1(A)) = S exp(1 f )jA S exp(2 1(A)) = S exp(1 f )jA exp(2 )(A) _ (Ac ) :

The required follows by Lemma 1.11.8. We now introduce idempotent analogues of standard probability distributions by requiring that their Laplace-Fenchel transforms be identical to the Laplace transforms of their probabilistic counterparts.

De nition 1.11.10. We say that f : !d Rd is idempotent Gaussian with parameters (m; ), where m 2 R and is a d d positive semi-de nite symmetric matrix, if Lf () = exp m + =2 .

Remark 1.11.11. Below we occasionally refer to m as the idempotent mean and as the idempotent covariance of an idempotent Gaussian variable f .

The next lemma follows by Lemma 1.11.5.

Lemma 1.11.12. An idempotent variable f : !f Rd is idempotent

Gaussian with parameters (m; ) if and only if (x) = exp (x m) (x m)=2 if x m is in the range of and f (x) = 0 otherwise.

The following is a consequence of the de nition.

Lemma 1.11.13. An idempotent variable f : ! Rd is idempotent Gaussian if and only if f :

every 2 Rd .

!R

is idempotent Gaussian for

De nition 1.11.14. We say that f : ! R + is idempotent Poisson with a parameter > 0 if Lf () = exp (e

1) ; 2 R.

An application of Lemma 1.11.5 yields the idempotent distribution of the Poisson idempotent variable.

Lemma 1.11.15.

An idempotent variable f : ! R+ is idempotent Poisson with a parameter > 0 if and only if f (x) = exp x ln(x=) + x ; x 2 R+ , where 0 ln 0 = 0.

© 2001 by Chapman & Hall/CRC

89

Laplace-Fenchel transform

Remark 1.11.16.

By Lemma 1.11.7 both Gaussian and Poisson idempotent variables are Luzin idempotent variables. We now apply the Laplace-Fenchel transform to limit theorems. The following result is \an idempotent law of large numbers". Lemma 1.11.17. Let ffi; i 2 N g be a sequence of independent identically distributed Rd -valued idempotent variables on an idempotent probability space ( ; ) such that S exp( f1 ) < 1 for from a neighbourhood of the origin. If (jf1 j > ) < 1 for every > 0, then n 1X f ! 0 as n ! 1: n i=1 i

Proof. It is suÆcient to check that for arbitrary Æ > 0 and arbitrary 2 Rd such that S exp( f1 ) < 1 n 1 X lim f > j j Æ = 0: n!1 n i=1 i

(1.11.2)

Since by \the Chebyshev inequality" for 2 (0; 1] n 1 X

n i=1

fi > jjÆ

P

Sef1 n S exp ni=1 fi = jjÆ ; exp(njjÆ) e

the required would follow if there exists 2 (0; 1] such that Sef1 < 1: (1.11.3) ejjÆ We have for > 0 Sef1 = Sef1 1(j f1 j > jj) _ Sef1 1(j f1 j jj) Sef1 1(j f1j > jj) _ ejj: (1.11.4) Let < Æ. Then the second term on the right-most side of (1.11.4) is less than ejjÆ . For the rst term we have in view of the hypotheses that Sef1 1(j f1 j > jj) lim = (jf1 j > jj) < 1: !0 ejjÆ Thus, (1.11.3) holds for > 0 small enough, which concludes the proof.

© 2001 by Chapman & Hall/CRC

90

Idempotent probability measures

Remark 1.11.18. Thus, an analogue of the expectation of a random

variable for an idempotent variable f in the law of large numbers is an element a such that (jf aj > ) < 1 for every > 0. The following lemma is an analogue of the method of characteristic functions in weak convergence theory. Lemma 1.11.19. Let f ; 2 g be a net of deviabilities on Rd . If W x d (x) ! L() as 2 for all 2 Rd , where L() is essen Rd e tially smooth, lower semi-continuous and such that 0 2 int(dom L), iw then ! , where deviability is given by the inversion formula (1.11.1). W Proof. Let us denote L () = Rd ex d (x). We show that the net f ; 2 g is tight. Indeed, since L () ! L() and L() is nite in a neighbourhood of the origin by the fact that 0 2 W int (dom L), there exists r > 0 such that lim sup Rd erjxj d (x) < 1; and then \the Chebyshev inequality" (fx : jxj > Ag) W e rA Rd erjxj d (x) yields the claim. Therefore, by Theorem 1.9.17 f ; 2 g has accumulation points in IMt (Rd ). Let f0 ; 0 2 0 g be a subnet of f ; 2 g that weakly converges to ~ 2 IMt (Rd ). Then the convergence L0 () ! L() implies that if 2 int(dom L), then for suitable W " > 0 lim sup0 Rd e(1+")x d0 (x) < 1: Hence, the function (ex ; x 2WRd ) is uniformly maximable with respect to f0 ; 0 2 0 g, W x x ~ x); implying that so lim0 Rd e d0 (x) = Rd e d( _ ~ x) = L() ex d( (1.11.5) Rd

for all 2 int(dom L). We prove that (1.11.5) actually holds for all 2 Rd . Let L~ () denote the left-hand side of (1.11.5). Since L() is essentially smooth, it follows that jrL~ (n )j ! 1 for every sequence n in int(dom L) that converges to a boundary point of dom L. Since L~ () is convex, it follows that L~ () = 1 for all 2= cl(dom L), which implies that L~ () = L() for 2= cl(dom L). Finally, if is a boundary point of dom L, then lower semi-continuity and convexity of L() imply that L() = limn L(n ), where n is a sequence of points from int(dom L) converging to . For the same reason this holds for L~ (), so we conclude that L~ () = L() for all 2 Rd . Then ~ = , which ends the proof. Lemma 1.11.5 implies that

© 2001 by Chapman & Hall/CRC

Chapter 2

Maxingales In this chapter we develop elements of idempotent stochastic calculus. We are mostly interested in studying idempotent analogues of martingales and martingale problems (which we call maxingales and maxingale problems, respectively).

2.1 Idempotent stopping times In this section we de ne stopping times with respect to -algebras and study their properties. The concepts, results and proofs are analogous to those in the general theory of stochastic processes, see Dellacherie [34] or Meyer [88]. Therefore, we omit proofs that are analogous to proofs in these books. Let be a set.

De nition 2.1.1. An indexed collection A = fAt; t 2 R+ g of -

algebras on is called a ow of -algebras if As At for s t: We also refer to a ow of -algebras as a - ow. We say that the - ow T A is right-continuous if As = t>s At for all s 2 R+ .

Remark 2.1.2. Recall that As At if and only if the atoms of As

are unions of the atoms of At . Remark 2.1.3. Given a - ow A, there is a natural rightcontinuous - ow associated with A that is de ned by A+ = (At+ ; t 2 R+ ), where At+ = \>0 At+.

We assume as given a ow of -algebras A and a -algebra A1 such that As A1 for s 2 R+ . 91 © 2001 by Chapman & Hall/CRC

92

Maxingales

De nition 2.1.4. An R + -valued function on is said to be an idempotent stopping time relative to A, or, for short, an A-stopping

time, if f sg 2 As for all s 2 R+ .

Lemma 2.1.5.

1. A function : ! R + is an A-stopping time if and only if f = sg 2 As for all s 2 R+ (in particular, constants are stopping times); if A is right-continuous, an equivalent condition is that f < sg 2 As for all s 2 R+ .

2. If and are A-stopping times, then _ , ^ and + are A-stopping times.

3. If ; 2 , are A-stopping times, then sup 2 is an A-stopping time; if A is right-continuous, then inf 2 , lim inf 2 and lim sup 2 (in the latter two cases is a directed set) are A-stopping times.

We now introduce -algebras associated with stopping times. Let us rst prove two simple facts. Recall that [!]A denotes the atom of a -algebra A about !. Lemma 2.1.6. Let be an A-stopping time. If !00 2 [!0]A(!0) , then (!00 ) = (!0 ) and [!0 ]A(!0 ) = [!00 ]A(!00 ) . Proof. Since !0 2 f! : (!) (!0 )g 2 A(!0 ) , [!0 ]A(!0 ) is an atom of A(!0 ) , and !00 2 [!0 ]A(!0 ) , it follows that !00 2 f! : (!) (!0 )g so that

(!00 ) (!0 ):

(2.1.1)

Therefore, [!00 ]A(!00 ) [!00 ]A(!0 ) = [!0 ]A(!0 ) . Thus, !0 2 [!00 ]A(!00 ) . The argument of the proof of (2.1.1) with roles of !0 and !00 switched then shows that we actually have equality in (2.1.1). Equality of the atoms in the statement is now self-evident. The following lemma is an easy consequence.

Lemma 2.1.7. Let be an A-stopping time.

Then the collection

f[!]A(!) ; ! 2 g is a partition of in that two arbitrary sets from

the collection are either disjoint or coincide and the union of the sets from the collection equals .

© 2001 by Chapman & Hall/CRC

93

Stopping times

De nition 2.1.8. Let be an A-stopping time. We denote by A the -algebra that has atoms [!]A(!) .

The next lemma shows that our de nition is consistent with the corresponding de nition in the general theory of stochastic processes.

Lemma 2.1.9. The -algebra A is the collection of subsets A of

such that A 2 A1 and A \ f sg 2 As for all s 2 R+ .

Proof. Clearly, A A1 . Let us show that [!]A(!) \f sg 2 As . Let !0 2 [!]A(!) \ f sg. Then by Lemma 2.1.6 [!0 ]As [!0 ]A(!0 ) = [!]A(!) . Also [!0 ]As f sg. Therefore, by Corollary 1.1.16 [!]A(!) \ f sg 2 As . Conversely, let A 2 A1 be such that A \ f sg 2 As for all s 2 R+ and let ! 2 A. We prove that A 2 A by proving that [!]A (!) A. If (!) < 1, then the required follows since ! 2 A \ f (!)g 2 A(!) so that [!]A (!) A \ f (!)g. If (!) = 1, then [!]A(!) is an atom of A1 and the required follows since A 2 A1.

Remark 2.1.10. The -algebra A can also be de ned as the col-

lection of subsets A of such that A \ f = sg 2 As for all s 2 R + .

Lemma 2.1.11.

measurable.

1. Let be an A-stopping time. Then is A -

2. Let be an A-stopping time and Then is an A-stopping time.

be

A -measurable.

3. Let and be A-stopping times such that . Then is a stopping time relative to the - ow fA+t ; t 2 R+ g.

4. Let and be A-stopping times such that . Then A A . 5. Let and be A-stopping times. Then the sets f < gT, f = g and fT > g belong both to A and A . Also A f = g = A f = g.

Let, given t > 0, At denote the -algebra generated by the algebras As for s < t. It is obvious that the atoms of At are of the form [!]At = \s t, we have that [!]A(!) [!]At . Therefore, [!](!) A. Since ! 2 f > tg 2 At , we have that [!]A(!) [!]At f > tg. The claim is proved. Thus, A^ A . The reverse inclusion follows since [!]A(!) = \t 0, and X0 (!) = 0. Then Xt is continuous at t = 0 for every ! 2 but it is not continuous in idempotent probability.

In the sequel we refer to idempotent processes with continuous (right-continuous, respectively) paths as continuous (rightcontinuous, respectively) idempotent processes. The following simple fact is useful.

Lemma 2.2.11. If X0 is a proper idempotent variable and X

is continuous in idempotent probability, then X is a proper idempotent process.

© 2001 by Chapman & Hall/CRC

99

Idempotent processes

We now introduce the class of Luzin-continuous processes, which is smaller than the class of processes continuous in idempotent probability. Let C (R + ; Rd ) denote the space of Rd -valued continuous functions on R+ equipped with the metric supst jxs ys j ^ 1 ; 1+t t2R+

dC (x; y) = sup

where x = (xt ; t 2 R+ ) and y = (yt ; t 2 R+ ). We recall that with this metric C (R + ; Rd ) is a complete separable metric space.

De nition 2.2.12. We say that an idempotent process X with Xcon-

tinuous trajectories is Luzin-continuous if the restriction of to C (R + ; R d ) is a deviability on C (R + ; R d ). Let, in addition, be a Hausdor topological space and be a deviability on . An idempotent process X is called a strictly Luzin-continuous idempotent process on ( ; ) if the mapping ! ! (Xt (!); t 2 R+ ) from to C (R + ; R d ) is a strictly Luzin idempotent variable.

Theorem 2.2.13.

A Luzin (respectively, strictly Luzin) idempotent process X is Luzin-continuous (respectively, strictly Luzincontinuous) if and only if for arbitrary T > 0 and > 0

lim sup jXt Xs j > = 0: Æ!0 s;t2[0;T ]: js tjÆ Proof. According to Corollary 1.8.7 a Luzin idempotent process X is Luzin-continuous if and only if inf K 2K (X 62 K ) = 0, where K is the class of compact subsets of C (R + ; Rd ). By a standard argument based on Arzela-Ascoli's theorem this is equivalent to the convergences

lim jX0 j > A = 0; lim A!1 Æ!0

sup jXt Xsj > = 0; s;t2[0;T ]: js tjÆ

where T > 0 and > 0 are arbitrary. The rst condition is a consequence of X0 being Luzin. The second condition is equivalent to the condition in the statement by -maxitivity of . The proof for strictly Luzin idempotent processes is similar. We now consider measurability issues in the spirit of the general theory of stochastic processes. We assume the discrete -algebra on

© 2001 by Chapman & Hall/CRC

100

Maxingales

unless otherwise speci ed. Let A = (At ; t 2 R+ ) be a ow of -algebras on . De nition 2.2.14. We say that an idempotent process X is Aadapted if Xt is At -measurable for every t 2 R+ . Let B [0; t] At be the product of the Lebesgue -algebra on [0; t] and the -algebra At in the sense of De nition 1.5.9. We refer to elements of B [0; t] At as progressively measurable sets. De nition 2.2.15. An idempotent process X is said to be progressively measurable (or A-progressively measurable) if the mappings (s; !) ! X (s; !) from [0; t] to Rd are B [0; t] At =B(Rd )measurable for all t 2 R+ . Lemma 2.2.16. An idempotent process X = (Xt (!); t 2 R+ ; ! 2 ) is A-progressively measurable if and only if it is A-adapted and the functions (Xt (!); t 2 R+ ) are B(R+ )=B(Rd )-measurable in t for all ! 2 . Proof. Let X be A-progressively measurable. Then, given t2 R+ and x 2 Rd , we have f(s; !) : Xs (!) = x;s 2 [0; t]g 2 B [0; t] At so that ftg f! : Xt (!) = xg 2 B [0; t] At , which implies that f! : Xt(!) = xg 2 At . Thus, X is A-adapted. Next, given !^ and a Borel subset of Rd , we have that f(s; !) : Xs (!) 2 g \ [0; t] [^!]At 2 B [0; t] At : Since Xs (!) = Xs (^!) for s 2 [0; t] if ! 2 [^!]At by A-adaptedness of X and Corollary 1.2.5, we conclude that fs 2 [0; t] : Xs (^!) 2 g [^!]At 2 B [0; t] At so that fs 2 [0; t] : Xs (^!) 2 g 2 B [0; t] . Necessity is proved. Conversely, since Rd

f(s; !) : Xs(!) 2 g =

\

[0; t]

[

!2

\

fs : Xs (!) 2 g [0; t] f!g ;

fs : Xs(!) 2 g \ [0; t] 2 B [0; t] for ! 2 , and Xs(!) = Xs(!0 ) if s 2 [0; t] and ! At !0 , the de nition of B [0; t] At implies suÆciency.

Lemma 2.2.17. Let X beR an A-progressively measurable idempotent t process. Let the integrals 0 XsR(!) ds, t 2 R+ ; ! 2 ; be well de ned. Then the idempotent process 0t Xs (!) ds; t 2 R+ is A-adapted. © 2001 by Chapman & Hall/CRC

101

Idempotent processes

Proof. Since X is A-adapted, 0t Xs (!) ds is constant on the atoms of At , so it is At -measurable by Corollary 1.2.5. R

The next result is a version of Lemma 2.1.18.

Lemma 2.2.18. Let X be A-adapted and D Rd . Let D = inf ft 2 R+

: Xt 2 Dg. Let either one of the conditions hold:

1. X is right-continuous and D is closed,

A is right-continuous. Then D is an A-stopping time. 2.

Proof. Under the rst condition, we have by right-continuity of X and closedness of D [

f! : D (!) tg = f! : Xs(!) 2 Dg 2 At: st

The second part follows by Lemma 2.1.18. We can also adapt the proof of the lemma by writing [

f! : D (!) < tg = f! : Xs(!) 2 Dg 2 At: s 0, and t 2 R+

(t X = xjpt X = x0 ) = (t X = x):

2. An A-adapted Luzin idempotent process X on ( ; ) has Aindependent increments if and only if for x 2 (Rd )R+ and t 2 R + -a.e.

(t X = xjAt ) = (t X = x):

De nition 2.2.25. An idempotent process X = (Xt ; t 2 R+ ) is said

to be idempotent Gaussian if its nite-dimensional distributions are idempotent Gaussian.

The next two theorems consider convergence in idempotent distribution for Luzin-continuous idempotent processes. We state the results in the form of weak convergence of associated deviabilities. The following tightness theorem is an obvious consequence of ArzelaAscoli's theorem.

Theorem 2.2.26. A net f ; 2 g of deviabilities on C (R + ; Rd )

is tight if and only if 1:Æ lim lim sup (fx : jx0 j > Ag) = 0, A!1 2:Æ lim lim sup (fx : sup jxt xs j > g) = 0; > 0; T > 0: Æ!0 s;t2[0;T ]: js tj 0. In particular, x^ is absolutely continuous and x^ 0 = 0. By the de nition of conditional idempotent expectation

SW (Yt jCs )(^x) = sup Yt (x)W (xjCs )(^x):

x2C

(2.4.4)

The conditional idempotent probability W (xjCs )(^x) is not equal to zero only if x and x^ belong to the same atom of Cs , i.e., ps x = ps x^ . For these x, W (xjCs )(^x) =

W (x) W (x) = : W ([^x]Cs ) supx0 : ps x0 =psx^ W (x0 )

(2.4.5)

Easy calculationsusing (2.4.1) show that the latter supremum equals Rs 2 ^_ u du=2 ; so, assuming that W (x) > 0 and x is thus exp 0x absolutely continuous and x0 = 0, by (2.4.5) and the equality xu = x^ u ; u s; 1 Z1 W _ 2u du : (xjCs )(^x) = exp x 2 s

Recalling that s x = (xs+u ps x = psx^ , then W (xjCs )(^x) = W (sx)

© 2001 by Chapman & Hall/CRC

x^ s; u 2 R+ ), we thus have that if -a.e.

117

Wiener and Poisson idempotent processes

Therefore, by the de nition of Yt (x), (2.4.4) and (2.4.3)

SW (Yt jCs )(^x) = sup Yt (x)W (xjCs )(^x)

x2C : ps x=ps x^

i 2 s h 2 (t s) W sup exp (xt x^ s) (s x) 2 x2C 2 W = Ys (^x) sup Yt s (sx) (s x) = Ys(^x)SW Yt s (x) = Ys(^x):

= exp x^ s

x2C

x) follows by Corollary 1.4.15 and the equal-

W -maximability of Yt (

ities

SW (Yt2 ) = exp(2 t)SW (2xt 22 t) = exp(2 t); where the latter equality follows by (2.4.2) and (2.4.3) with replaced by 2. Also, x0 = 0 W -a.e. by the de nition of W . This ends the proof of the implication 1 ! 3. We prove that part 3 implies part 2. The maxingale property of Y easily implies that 1 SW (exp((xt xs))jCs ) = exp 2 (t s) ; 2 which yields the required by De nition 1.11.10, Corollary 1.11.9 and Lemma 1.11.12. We prove that part 2 implies part 1. By independence of increments the nite-dimensional idempotent distributions of W are Gaussian so by Remark 1.11.16 they are deviabilities on the associated spaces. Therefore, by Corollary 2.2.6 the idempotent distribution of W is -smooth relative to the collection of closed subsets of R R+ . Thus, for x 2 R R+ such that x0 = 0 by Theorem 2.2.2 and independence of increments of W (! : W (!) = x) = t inf (! : Wti (!) = xti ; i = 1; : : : ; k) ;:::;t 1

k

k Y

= t inf (! : Wt1 (!) = xt1 ) ;:::;t 1

k

= xti

i=2

xti 1 ) = exp

(! : Wti (!) Wti 1 (!) k X (xti 1 sup 2 t1 ;:::;tk i=1 ti

xti 1 )2 : ti

1

The Rlatter supremum equals +1 if x is not absolutely continuous and 01 x_ 2t dt if x is absolutely continuous.

© 2001 by Chapman & Hall/CRC

118

Maxingales

The next result shows that the Wiener idempotent process is Luzin-continuous.

Lemma 2.4.3. The Wiener idempotent probability is a deviability, i.e., it is tight and -smooth relative to the collection of closed subsets of C (R + ; R). Proof. By Corollary 1.7.15 it would be suÆcient to prove that W (x) is an upper compact function of x 2 C (R + ; R), which is an easy exercise. We give, however, a dierent \probabilistic" proof based on Theorem 2.4.2. Let X be the canonical idempotent process on (C (R + ; R); W ). By Theorem 2.2.13 it suÆces to prove that for arbitrary T > 0 and > 0

lim sup W jXt Xsj > = 0: Æ!0 s;t2[0;T ]: js tjÆ By Theorem 2.4.2 and the Chebyshev inequality for > 0

W jXt Xs j > = W Xt Xs > _ W Xs Xt > SW exp (Xt Xs ) SW exp (Xs Xt ) _ exp() exp() exp 2 jt sj=2 = ; exp()

which implies the required since is arbitrary.

De nition 2.4.4. We say that a continuous idempotent process W is Wiener relative to a - ow A (or an A-Wiener idempotent process for short) if the idempotent process M () de ned in the statement of Theorem 2.4.2 is an A-exponential maxingale starting at 1 for every 2 R.

Lemma 2.4.5. A continuous idempotent process W is A-Wiener if and only if the idempotent process M () de ned in the statement of Theorem 2.4.2 is an A-local exponential maxingale starting at 1 for every 2 R.

Proof. Let M () be an A-local exponential maxingale starting at 1 for every 2 R. By Lemma 2.3.13 we only need to prove that for

© 2001 by Chapman & Hall/CRC

Wiener and Poisson idempotent processes

119

every s 2 R+ the process (Mt^s (); t 2 R+ ) is uniformly maximable. As in the proof of Theorem 2.4.2

S Mt^s ()2

S Mt^s (2) exp(2s) exp(2 s);

where the latter inequality follows by Lemma 2.3.13. The uniform maximability follows by Corollary 1.4.15. Obviously, if M () is an A-exponential maxingale, then it is an AW -exponential maxingale. Thus, we have the following consequence of Theorem 2.4.2.

Corollary 2.4.6. If W is an A-Wiener idempotent process, then it has properties described in parts 1 and 2 of Theorem 2.4.2.

Corollary 2.4.7. If W Wt

is an A-Wiener idempotent process, then Ws is independent of As for t s.

We now consider the multi-dimensional case.

De nition 2.4.8. 1 d

An Rd -valued idempotent process W = (W ; : : : ; W ) on ( ; ) is called a d-dimensional Wiener idempotent process if the processes W 1 ; : : : ; W d are independent Wiener idempotent processes.

The next theorem is proved similarly to Theorem 2.4.2. Let, as above, AW denote the ow of -algebras generated by W and W denote the idempotent distribution of W . Let Ed denote the d d identity matrix.

Theorem 2.4.9. Let W

be an Rd -valued idempotent process. The following conditions are equivalent: 1. W is a d-dimensional Wiener idempotent process,

2. the density of W is given by (x 2 RR+ ) 8 1 Z1 > > 2 ds ; > > _ j x j if x is absolutely exp s > < 2 continuous and 0 W (x) = > > > x0 = 0; > > : 0; otherwise,

© 2001 by Chapman & Hall/CRC

120

Maxingales

3. W is an idempotent process with independent increments, W0 = 0 -a.e., and increments Wt Ws are idempotent Gaussian with parameters (0; (t s)Ed ) so that

(Wt Ws = x) = exp

jxj2

2(t s)

; x 2 Rd ;

4. the idempotent process M () = (Mt (); t 2 R+ ) de ned by 1 Mt () = exp Wt jj2 t ; 2 is an AW -exponential maxingale such that M0 () = 1 -a.e. for every 2 Rd .

Clearly, a d-dimensional idempotent Wiener process is Luzincontinuous. Also, since W has independent increments, which are idempotent Gaussian variables, we have the following corollary.

Corollary 2.4.10. A d-dimensional idempotent Wiener process is an idempotent Gaussian process.

Part 4 of Theorem 2.4.9 implies the following analogues of the properties of the Wiener process. Let e = (t; t 2 R+ ). We recall that Æ denotes the composition map.

Corollary 2.4.11. 1. Let W be a one-dimensional idempotent Wiener process and 2 R+ . Then the idempotent process W Æ (e) has the same idempotent distribution as 1=2 W .

2. Let W1 ; : : : ; Wk be independent d-dimensional idempotent Wiener processes and 1 ; :P : : ; k be l d matrices. Then the idempotent distribution of ki=1 i Wi coincides with the idemPk T 1=2 W , where W is an lpotent distribution of i=1 i i dimensional idempotent Wiener process.

De nition 2.4.12. An Rd -valued continuous idempotent process W is called a d-dimensional Wiener idempotent process with respect to a - ow A (or A-Wiener, for short) if the idempotent process M () de ned in the statement of Theorem 2.4.9 is an A-exponential maxind gale starting at 1 for every 2 R .

The proof of the following lemma is similar to the proof of Lemma 2.4.5.

© 2001 by Chapman & Hall/CRC

Wiener and Poisson idempotent processes

121

Lemma 2.4.13. An Rd -valued idempotent process W is a ddimensional A-Wiener idempotent process if and only if the idempotent process M () de ned in the statement of Theorem 2.4.9 is an A-local exponential maxingale starting at 1 for every 2 Rd .

We de ne now the Poisson idempotent process. As above, we assume that 0 ln 0 = 0. De nition 2.4.14. We say that an idempotent probability N on C (R + ; R ) is the Poisson idempotent probability if it has density de ned by 8 Z1 > > > > _ _ _ exp ( x ln x x + 1) ds ; if x is absolutely > s s s > > > < continuous, 0 N (x) = x_ s 2 R+ a.e. > > > > > and x0 = 0, > > > :0; otherwise. (2.4.6) A Poisson idempotent process N = (Nt ; t 2 R+ ) on ( ; ) is an idempotent process with paths from C (R + ; R) and idempotent distribution N , i.e., (N = x) = N (x); x 2 C (R + ; R). We call N a canonical Poisson idempotent process if it is the canonical process on (C (R + ; R); N ). Remark 2.4.15. According to the de nition, the Poisson idempotent process has increasing paths -a.e. We now give a characterisation of the Poisson idempotent process in the spirit of Watanabe's characterisation of the Poisson process and analogous to that of the Wiener idempotent process. Let AN = N (AN t ; t 2 R+ ), where At denotes the -algebra on generated by the maps ! ! Ns(!) for s t. Theorem 2.4.16. Let N be an R-valued idempotent process. The following statements are equivalent: 1. N is a Poisson idempotent process, 2. N is an idempotent process with independent increments, N0 = 0 -a.e., and increments Nt Ns for s < t are idempotent Poisson with parameters t s, i.e., x (Nt Ns = x) = exp x ln +x (t s) ; x 2 R+ ; t s

© 2001 by Chapman & Hall/CRC

122

Maxingales

3. the idempotent process M () = (Mt (); t 2 R+ ) de ned by

Mt () = exp Nt (e 1)t

is an AN -exponential maxingale such that M0 () = 1 -a.e. for every 2 R. Proof. The proof is analogous to the proof of Theorem 2.4.2. We prove that part 1 implies part 3. Let N be a Poisson idempotent process. As in the proof of Theorem 2.4.2 it is suÆcient to prove that the idempotent process Y = (Yt (x); t 2 R+ ; x 2 C (R + ; R)) de ned by

Yt (x) = exp xt (e 1)t

(2.4.7)

SN Yt (x) = 1:

(2.4.8)

is a C(R+ ; R)-exponential maxingale on (C (R + ; R); N ). We rst note that by (2.4.6)

Let x^ 2 C (R + ; R) be such that N (^x) > 0. In particular, x^ is absolutely continuous, increasing and x^ 0 = 0. By the de nition of conditional idempotent expectation SN (Yt jCs )(^x) = sup Yt (x)N (xjCs )(^x): (2.4.9) x2C By the reasoning used in the proof of Theorem 2.4.2 we have that N (xjCs )(^x) = N (s x):

Therefore, by the de nition of Yt (x), again repeating the argument of the proof of Theorem 2.4.2, SN (Yt jCs )(^x) = sup Yt (x)N (xjCs )(^x) x2C : ps x=ps x^ = Ys (^x) sup Yt s (s x)N (s x) = Ys (^x)SN Yt s (x) = Ys (^x): x2C N -maximability of Yt (x) follows by Corollary 1.4.15 and the equality

SN (Yt2 ) = exp (e2

© 2001 by Chapman & Hall/CRC

2e + 1)t SN 2xt (e2 1)t = exp (e2 2e + 1)t ;

123

Wiener and Poisson idempotent processes

where the latter equality follows by (2.4.7) and (2.4.8) with replaced by 2. Finally, Y0 (x) = 1 N -a.e. since x0 = 0 -a.e. This ends the proof of the implication 1 ! 3. We prove that part 3 implies part 2. The maxingale property of Y yields

SN (exp((xt

xs))jCs ) = exp (e

1)(t s) ;

which implies the required by Corollary 1.11.9 and Lemma 1.11.15. To prove that part 2 implies part 1 we write for an increasing function x 2 RR+ such that x0 = 0 by Theorem 2.2.2 and independence of increments of N (! :

(! : Nti (!) = xti ; i = 1; : : : ; k) N (!) = x) = t1inf ;:::;tk

= t inf (! : ;:::;t 1

= xti

k

Nt1 (!) = xt1 )

xti 1 ) = exp

sup

k Y

(! :

i=2 k X

t1 ;:::;tk i=1

(xti

(xti

Nti (!) Nti 1 (!)

xti 1 ) ln xttii txiti1 1

xti 1 ) + (ti

ti 1 ) :

The Rlatter supremum equals +1 if x is not absolutely continuous and 01 (x_ t ln x_ t x_ t + 1) dt if x is absolutely continuous. The next result shows that N is a Luzin-continuous idempotent process.

Lemma 2.4.17. The Poisson idempotent probability is a deviability, i.e., it is tight and -smooth relative to the collection of closed subsets of C (R + ; R). Proof. We give again a \probabilistic" proof. Let X be the canonical process on (C (R + ; R); N ). By Theorem 2.2.13 it suÆces to prove that for arbitrary T > 0 and > 0

lim sup N Xt Xs > = 0: Æ!0 s;t2[0;T ]: 0t sÆ

© 2001 by Chapman & Hall/CRC

124

Maxingales

By Theorem 2.4.16 and the Chebyshev inequality for > 0

S N exp (Xt Xs ) Xs > exp() exp (e 1)(t s) ; = exp() which implies the required since is arbitrary. N Xt

De nition 2.4.18. We say that a continuous idempotent process N is Poisson relative to a - ow A (or A-Poisson idempotent process for short) if the idempotent process M () de ned in the statement of Theorem 2.4.16 is an A-exponential maxingale for every 2 R such that M0 () = 1 -a.e.

Corollary 2.4.19. If N is an A-Poisson idempotent process, then

it has properties described in parts 1 and 2 of Theorem 2.4.16. Also, Nt Ns is independent of As for s t.

Lemma 2.4.20. An R-valued continuous idempotent process N is A-Poisson if and only if the idempotent process M () de ned in the statement of Theorem 2.4.16 is an A-local exponential maxingale starting at 1 for every 2 R.

2.5 Idempotent stochastic integrals Let ( ; ) be an idempotent probability space with a ow of algebras A = (At ; t 2 R+ ).

De nition 2.5.1. An Rd -valued continuous A-adapted idempotent

process M = (Mt ; t 2 R+ ) such that M0 = 0 is called a local maxingale (maxingale or uniformly maximable maxingale, respectively) with a quadratic characteristic hM i relative to A (or an A-local maxingale, A-maxingale, or uniformly maximable Amaxingale, respectively, for short) if there exists an Rdd -valued continuous A-adapted idempotent process hM i = (hM it ; t 2 R+ ) such that hM i0 = 0, hM it hM is for 0 s t are positive semi-de nite symmetric d d matrices and the idempotent process (exp( Mt hM it =2); t 2 R+ ) is an A-local exponential maxingale (respectively, A-exponential maxingale, uniformly maximable Aexponential maxingale) for every 2 Rd .

© 2001 by Chapman & Hall/CRC

125

Idempotent stochastic integrals

By Theorem 2.4.9 an Rd -valued Wiener idempotent process is a local maxingale with a quadratic characteristic Ed t. Lemma 2.4.13 yields the following converse.

Corollary 2.5.2. Let a continuous idempotent process M be an Rd valued A-local maxingale with a quadratic characteristic (Ed t; t 2 R + ). Then M is a d-dimensional A-Wiener idempotent process. The following consequence of Lemma 2.3.14 is also useful.

Lemma 2.5.3. Let a continuous idempotent process M be an Rd valued A-local maxingale with a quadratic characteristic hM i. Then for a > 0, b > 0, c > 0, and nite A-stopping times and such

that

( sup jMt M j a) ec(b t

a) _( khM i

hM i k > 2b=c):

Proof. By \the Doob stopping theorem" the idempotent process exp (Mt+ M ) (hM it+ hM i )=2 ; t 2 R+ ; 2 Rd ; is a supermaxingale relative to the - ow (At+ ; t 2 R+ ). Also, is a stopping time relative to (At+ ; t 2 R+ ) by Lemma 2.1.11. Hence, by Lemma 2.3.14 and -maxitivity of

( sup jMt M j a) = sup ( sup (Mt M ) a) t jj=1 t ec(b a) _ sup ( (hM i hM i ) > 2b=c) jj=1 = ec(b a) _ ( khM i hM i k > 2b=c): Properties of the trajectories of hM i are often translated into the corresponding properties of M . The next result follows by Lemma 2.5.3 and Theorem 2.2.13.

Lemma 2.5.4. Let a continuous idempotent process M be a local maxingale relative to the ow A with a quadratic characteristic hM i. 1. If hM i is a proper idempotent process, then M is a proper idempotent process.

© 2001 by Chapman & Hall/CRC

126

Maxingales

2. If hM i is continuous (respectively, stopping-time-rightcontinuous) in idempotent probability, then M is continuous (respectively, stopping-time-right-continuous) in idempotent probability. 3. Let M be Luzin. If hM i is Luzin-continuous, then M is Luzincontinuous.

We assume in the rest of the section that the - ow A is complete in the sense of the following de nition.

De nition 2.5.5.

We say that a ow of -algebras on ( ; ) is complete if the -algebras in the ow are complete with respect to .

Remark 2.5.6. Clearly, if A = (At ; t 2 R+ ) is a - ow, then A = (At ; t 2 R+ ), where the At are the completions of the At with respect to , is a complete - ow. We refer to it as the completion of A with respect to (or the -completion of A). The next lemma extends Lemma 1.2.6.

Lemma 2.5.7. Let A be the completion of A with respect to . If is an A -stopping time, then there exists an A-stopping time 0 0

such that = -a.e. Proof. We de ne 0 as follows: if (!) > 0, then 0 (!) = (!); if (!) = 0 and there exists !~ such that (~!) > 0 and ! 2 [~!]A(~!) , then 0 (!) = (~! ); if (!) = 0 and no such !~ exists, then 0 (!) = 1. We rst check that 0 is well de ned. Indeed, suppose for some ! such that (!) = 0 there exist !~ and !^ such that (~!) > 0, (^! ) > 0, ! 2 [~!]A(~!) , ! 2 [^!]A(^!) , and (~!) (^! ); then [~!]A(~!) = [!]A(~!) [!]A(^!) = [^!]A(^!) ; hence, !~ 2 [^!]A(^!) since (~!) > 0 and (^!) > 0, so (~! ) = (^! ) by Lemma 2.1.6 proving the claim. We check that 0 is an A-stopping time. Let 0 (!0 ) = t and 00 ! 2 [!0 ]At , where t 2 R+ . We have to check that 0 (!00 ) = t. If (!0 ) > 0 and (!00 ) > 0, then 0 (!0 ) = (!0 ), 0 (!00 ) = (!00 ) and !00 2 [!0 ]At ; hence, (!0 ) = t so !00 2 [!0 ]A(!0 ) and (!0 ) = (!00 ) by Lemma 2.1.6. If (!0 ) > 0 and (!00 ) = 0, then as above (!0 ) = t, so !00 2 [!0 ]A(!0 ) and by de nition 0 (!00 ) = (!0 ). If (!0 ) = 0, then, since 0 (!0 ) is nite, there exists !~ such that (~!) > 0, !0 2 [~!]A(~!) and 0 (!0 ) = (~!); hence, !00 2 [~!]A(~!) . If, in addition,

© 2001 by Chapman & Hall/CRC

127

Idempotent stochastic integrals

(!00 ) > 0, then !00 2 [~!]A(~!) so by Lemma 2.1.6 0 (!00 ) = (~!); if (!00 ) = 0, then the latter equality holds by de nition. We say that an Rmd -valued continuous idempotent process X is absolutely continuous if the entry processes have -a.e. absolutely continuous with respect to Lebesgue measure trajectories; if X is, in addition, A-adapted, then we denote by X_ an A-progressively measurable idempotent process such that X_ (!) is a version of the RadonNikodym derivative of X (!) with respect to Lebesgue measure for almost all !. (For instance, we could de ne X_ t (!) as the left derivative of X (!) at t if the latter exists and let X_ t (!) = 0 otherwise.) For a quadratic characteristic hM i to be absolutely continuous, it is actually suÆcient that the diagonal Rentries are absolutely continuous. Note that if this is the case, then 0t khM_ isk ds < 1; t 2 R+ : The following lemma shows, in particular, that absolute continuity of the quadratic characteristic of a local maxingale implies absolute continuity of the local maxingale itself. It also lays the groundwork for the de nition of idempotent integrals with respect to local maxingales. We recall that denotes the pseudo-inverse of a matrix .

Lemma 2.5.8. Let an Rd -valued continuous idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i. Then the following holds.

1. M is absolutely continuous. 2. M_ s (!) belongs to the range of hM_ is (!) for almost all s and -almost all !, and h

Z1

S exp

0

i 1 hM_ is ) ds 2 h 1 Z1 i = S exp M_ s hM_ is M_ s ds 1; 2

sup ( M_ s 2Rd

0

R in particular, 01 M_ s idempotent variable.

© 2001 by Chapman & Hall/CRC

hM_ is M_ s ds

is a -a.e. nite proper

128

Maxingales

Let, in addition, (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued progressively measurable idempotent process such that Zt

0

A-

ks hM_ issT k ds < 1; t 2 R+ ; -a.e.

Then Zt

jsM_ sj ds < 1; t 2 R+ ; -a.e.

0

Proof. Let, for 0 s1 t1 : : : sk tk and i 2 Rd ; i = 1; : : : ; k, k hX

Z = exp

i=1

i Mti Msi

k 1X (hM iti 2 i=1 i

We show that

SZ 1:

i

hM isi )i : (2.5.1) (2.5.2)

Let

n = inf ft 2 R+ :

khM it k ng:

(2.5.3)

By Lemma 2.2.18 the n are A-stopping times. Also by Lemma 2.3.13

S exp( Mt^n hM it^n =2)2 S exp(2 Mt^n (2) hM it^n (2)=2) exp(njj2 ) exp(njj2 ) so that by Lemma 2.3.13 fn g is a localising sequence of stopping times for every local maxingale exp( Mt hM it =2); t 2 R+ , 2 Rd . Let 1 Yni = exp i Mti ^n Msi ^n i (hM iti ^n hM isi ^n )i : 2 (2.5.4)

© 2001 by Chapman & Hall/CRC

129

Idempotent stochastic integrals

Since (exp(i Mt^n i exponential maxingale, S (Yni jAsi ) = 1: By (2.5.4) and (2.5.5) we have

S

k Y i=1

Yni = S

kY1

i=1

hM it^n i=2); t 2

R+ )

is an

A-

(2.5.5)

Yni S (Ynk jAsk ) = S Q

kY2

i=1

Yni S (Ynk

1

jAsk 1 )

= : : : = SYn1 = 1:

Since n ! 1, by (2.5.1) Z = limn!1 ki=1 Yni , and \the Fatou lemma" (see Theorem 1.4.19) yields (2.5.2). This inequality implies in view of the de nition of idempotent expectation that k X

k 1X sup i Mti Msi i (hM iti hM isi )i 2 i=1 f(si ;ti )g;fi g i=1 < 1 -a.e. (2.5.6) Now let us suppose that there exists ! 2 such that (!) > 0 and T > 0 such that Mt (!) is not absolutely continuous on [0; T ]. Then there exists " > 0 such that for every Æ > 0 there exist nonoverlapping subintervals f(sÆi ; tÆi )g of [0; T ] such that X

i

(tÆi sÆi ) < Æ and

X

i

jMtÆi (!) MsÆi (!)j > ":

(2.5.7)

Let Æi be such that jÆi j = 1 and Æi (MtÆi (!) MsÆi (!)) = jMtÆi (!) MsÆi (!)j. Given A > 0 we choose Æ > 0 such that X khM itÆi (!) hM isÆi (!)k < A12 ; i

which is possible in view of absolute continuity of hM i. Then by (2.5.7) k X i=1

(AÆi ) MtÆi (!) MsÆi (!)

k 1X (AÆ ) (hM itÆi (!) 2 i=1 i

© 2001 by Chapman & Hall/CRC

hM isÆi (!))(AÆi ) A" 12 ;

130

Maxingales

which contradicts (2.5.6) since A is arbitrary. Part 1 is proved. For part 2, we note that by properties of idempotent expectation and (2.5.2)

sup Z = sup SZ 1: f(si ;ti )g;fi g f(si ;ti )g;fi g By Lemma A.2 in Appendix A the supremum on the left-most side R equals 01 sup2Rd ( M_ s hM_ is =2) ds. We thus have i h Z1 1 S exp sup ( M_ s hM_ is ) ds 1: 2 2Rd

S

0

The supremum in the integral is nite if and only if M_ s is orthogonal to the nullspace of hM_ is , which is equivalent to M_ s being in the range of hM_ is. It is then equal to M_ s hM_ is M_ s =2, which proves the rst claim of part 2. The second one is an obvious consequence of the rst. For the nal assertion, we note that since M_ s (!) belongs to the range of hM_ is(!) for almost all s and -almost all !, we have that M_ s (!) = hM_ is(!)hM_ is (!)M_ s (!) for almost all s and -almost all !. Therefore, by the Cauchy-Schwarz inequality -a.e. Zt

jsM_ sj ds

0

Zt

0

kshM_ issT k ds

1=2 Zt

0

1=2

M_ s hM_ is M_ s ds

;

where the right-hand side is nite -a.e. by hypotheses and the part of the lemma already proved.

De nition 2.5.9. Let idempotent processes M and be as in the statement of Lemma 2.5.8. An idempotent process X = (Xt (!); t 2 R + ; ! 2 ) de ned by

Xt (!) =

8 t Z > >

0;

> 0 > :~

Xt (!);

if (!) = 0;

where X~ t (!) is a continuous idempotent process, is called an idempotent stochastic integral of with respect to M and denoted by Rt _ M = 0 s Ms ds; t 2 R+ .

© 2001 by Chapman & Hall/CRC

131

Idempotent stochastic integrals

In particular, if M is a d-dimensional Wiener idempotent process Rt 2 W and 0 ks k ds < 1; t 2 R+ ; -a.e., the integral W is called an idempotent Ito integral.

Clearly, an idempotent stochastic integral is a continuous Aadapted process and is speci ed uniquely -a.e. We show that under certain conditions on the integrands idempotent stochastic integrals are local maxingales with quadratic characteristics. We begin with an approximation lemma.

Lemma 2.5.10. Let an Rd -valued continuous A-adapted idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i. Let (sk (!); s 2 R+ ; ! 2 ); k 2 N ; and (s (!); s 2 R+ ; ! 2 ) be A-progressively measurable Rmd -valued idempotent processes such that for t 2 R+ Zt

0

kshM_ issT k ds < 1; Zt

0

Zt

0

ksk hM_ is(sk )T k ds < 1;

k(sk s)hM_ is(sk s)T k ds ! 0 as k ! 1: (2.5.8a)

If the idempotent processes Rk M are A-local maxingales with the quadratic characteristics 0t sk hM_ is (sk )T ds; t 2 R+ , then the idempotent process R M is an A-local maxingale with the quadratic characteristic 0t shM_ is sT ds; t 2 R+ .

Proof. The idempotent processes M and k M are well de ned by Lemma 2.5.8. The idempotent process M is A-adapted by Lemma 2.2.17 and completeness of A. We introduce, for 2 Rm ,

1 2

Zt () = exp ( M )t

Ztk () = exp (k M )t

1 2

Zt

0

Zt

0

s hM_ is sT ds ; (2.5.9a)

sk hM_ is(sk )T ds : (2.5.9b)

The idempotent processes (Ztk (); t 2 R+ ) are A-local exponential maxingales by hypotheses. We have to prove that the idempotent

© 2001 by Chapman & Hall/CRC

132

Maxingales

process (Zt (); t 2 R+ ) is an A-local exponential maxingale. Let us note that by Lemma 2.5.8

SZt () 1; SZtk () 1; 2 Rm :

(2.5.10)

Let us introduce for n 2 N

n = inf ft 2 R+ : Zt

nk = inf ft 2 R+ :

0

Zt

0

kshM_ issT k ds ng; (2.5.11a)

ksk hM_ is(sk )T k ds n + 1g ^ n:(2.5.11b)

We show that the idempotent processes (Zt^n (); t 2 R+ ) are uniformly maximable. By (2.5.9a), (2.5.11a) and (2.5.10)

S Zt^n

()2 = S

h

Zt^n (2) exp

tZ^n

0

Thus,

sup S Zt^n ()2 < 1; t2R+

shM_ is sT ds

i

exp jj2n : (2.5.12)

proving uniform maximability of (Zt^n (); t 2 R+ ) by Corollary 1.4.15. A similar argument shows that S Ztk^nk ()2 exp jj2 (n + 1) : Therefore, the collection fZtk^nk (); k 2 N ; t 2 R+ g is uniformly maximable; in particular, the (Ztk^nk (); t 2 R+ ); k 2 N ; are uniformly maximable exponential maxingales. By Lemma 1.6.22 it thus suÆces to prove that Ztk^nk () ! Zt^n () as k ! 1. Since (2.5.12) implies that Zt^n () is a proper idempotent variable, it follows by the de nitions (2.5.9a) and (2.5.9b) that the required is a consequence of the convergences as k ! 1 tZ^nk 0

(k M )t^nk sk hM_ is (sk )T ds !

© 2001 by Chapman & Hall/CRC

! ( M )t^n ; (2.5.13a)

tZ^n 0

s hM_ issT ds: (2.5.13b)

133

Idempotent stochastic integrals

Let us rst note that in view of (2.5.11a), (2.5.11b) and (2.5.8a) we have that

lim t ^ nk 6= t ^ n = 0: k!1 Limit (2.5.13a) follows now by the inequalities j (k M )t^nk

( M )t^n j > Z t

+

0 Zt

0

j (sk

s)M_ s j ds

Zt

0

0

t ^ nk 6= t ^ n

1=2

(sk s)hM_ is(sk s)T ds Zt

M_ s hM_ is M_ s ds A

j (sk s)M_ s(!)j ds > " ;

0 Z t

(2.5.14)

1=2

M_ s hM_ is M_ s ds

;

exp( A=2);

and (2.5.8a). Similarly, limit (2.5.13b) follows by (2.5.14), (2.5.8a), (2.5.11a), and the inequality tZ^n

0

sk hM_ is(sk )T ds

tZ^n

2

0

tZ^n

0

(sk

0

s)hM_ is(sk

shM_ is sT ds

s hM_ is sT ds 1=2

s)T ds

1=2

+

tZ^n 0

© 2001 by Chapman & Hall/CRC

tZ^n

(sk

s)hM_ is (sk

s )T ds:

134

Maxingales

Theorem 2.5.11. Let an Rd -valued continuous A-adapted idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i, which is a proper idempotent process. Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively measurable idempotent process such that for t 2 R+ Zt

Zt

0

0

kshM_ is sT k ds < 1;

ks hM_ issT k 1( ks k > A) ds ! 0

as A ! 1:

Let there exist functions nA : R+ ! [0; 1], where A 2 R+ , such that nA (x) = 1 if x A, nA(x) A=x if x A, and for all A large enough

lim Æ!0

Z t

0

k snA( ks k ) s Æ nA( ks Æ k ) hM_ is

s nA( ks k ) s Æ nA ( ks

T

Æk )

k ds > = 0; t 2 R+ ; > 0;

where s (!) = 0 for s < 0. Then the idempotent stochastic integral M is an A-local maxingale with the quadratic characteristic Zt

h M it = shM_ issT ds: 0

The latter is a proper idempotent process. Proof. We have to prove that Z () = (Zt (); t 2 R+ ), 2 Rm , de ned as in (2.5.9a) is an A-local exponential maxingale. It is Aadapted by Lemma 2.2.17 and completeness of the - ow A. Let us rst consider the case that

s (!) =

k X i=1

fi (!) 1(ti

© 2001 by Chapman & Hall/CRC

1 ;ti ] (s);

135

Idempotent stochastic integrals

where 0 = t0 < t1 < : : : < tk and the fi are Ati 1 -measurable and bounded Rmd -valued idempotent variables. Then k hX

Zt () = exp

i=1

fi Mti ^t

Mti 1 ^t

i 1 fi (hM iti ^t hM iti 1 ^t )fiT : (2.5.15) 2 Let n be de ned by (2.5.3). By Lemma 2.5.8 S [exp( Mt^n hM it^n =2)] 1, which implies as in the proof of Lemma 2.5.10 that the idempotent process (exp( Mt^n hM it^n =2); t 2 R+ ) is uniformly maximable. Hence, by Lemma 2.3.13 (exp( Mt^n hM it^n =2); t 2 R+ ) is a uniformly maximable A-exponential maxingale. Let

Yti = exp fi Mti ^t^n Mti 1 ^t^n 1 T fi (hM iti ^t^n hM iti 1 ^t^n )fi : (2.5.16) 2 Since fi is Ati 1 -measurable, Lemma 1.6.21 implies that S (Yti jAti 1 ) = 1:

(2.5.17)

By (2.5.15), (2.5.16), and (2.5.17) we have

SZt^n () = S

k Y i=1

=S

Yti = S

kY2

i=1

kY1

i=1

Yti S (Ytk

Yti S (Ytk jAtk 1 ) 1

jAtk 2 ) = : : : = SYt1 = 1:

Since the latter holds for all 2 Rm , we have by (2.5.15) and (2.5.3) that

S Zt^n (2) exp(jj2 Bnk) = exp(jj2 Bnk); where B is an upper bound for the kfi k 2 . By Corollary 1.4.15 we conclude that the idempotent process (Zt^n (); t 2 R+ ) is uniformly S Zt^n ()2

maximable.

© 2001 by Chapman & Hall/CRC

136

Maxingales

The maxingale property of (Zt^n (); t 2 R+ ) is checked similarly. Let s t. If s tk , then Zs^n () = Zt^n () and the maxingale property trivially holds. Let ti 1 < s ti for some i. Since by the argument used above S (Zt^n ()jAti ) = Zti ^n (), it follows that

S (Zt^n ()jAs ) = S (S (Zt^n ()jAti )jAs ) = S (Zti ^n ()jAs ) = Zti 1 ^n ()S (Ytii jAs): By (2.5.16), since fi is As-measurable, S (Ytii jAs ) = exp fi (Ms^n Mti 1 ^n ) 1 fi (hM is^n hM iti 1 ^n )fiT 2 S exp fi(Mti ^n Ms^n ) 1 fi (hM iti ^n hM is^n )fiT jAs 2 = exp fi (Ms^n Mti 1 ^n ) 1 fi (hM is^n hM iti 1 ^n )fiT ; 2 where the latter equality follows by the maxingale property of exp( Mt^n hM it^n =2) and Lemma 1.6.21. Putting everything together, we conclude that S (Zt^n ()jAs ) = Zs^n (): Let us now assume that s is bounded and locally continuous in s uniformly on , i.e., wT (Æ) = sup sup ks (!) t (!)k ! 0 as Æ ! 0, T > 0: !2 s;tT js tjÆ

R Let us rst note that 0t khM_ is k ds is a proper idempotent variable since the integral is not greater than the sum of the diagonal entries of hM it and the latter is a proper idempotent variable. Let

sk

=

2

k X i=1

(i

1)=k

1(s 2 ((i

© 2001 by Chapman & Hall/CRC

1)=k; i=k]):

137

Idempotent stochastic integrals

Then by the part just proved the Z k (), de ned as Z () with s changed to sk , are A-local exponential maxingales. We also have that for k t Zt

0

Zt

k(s sk )hM_ is(s sk )T k ds wt (1=k)2 khM_ isk ds 0

Rt 0

so that by Lemma 2.5.10 and the fact that khM_ is k ds is a proper idempotent variable we conclude that Z () is an A-local exponential maxingale. Let us assume now that s in the statement of the theorem is bounded. We introduce the Steklov functions

sk

=k

Zs

u du:

s 1=k

Since ksk (!) tk (!)k 2k supu;! ku (!)k jt sj; the functions sk are continuous in s uniformly over !. They are also bounded and properly measurable so that by the part just proved the associated idempotent processes Z k () are A-local exponential maxingales. We again apply Lemma 2.5.10 to deduce that Z () is an A-local exponential maxingale. We have Zt

0

=

k(s sk )hM_ is(s sk )T k ds Zt

0

Z1=k

Z1=k

k k (s s u) du hM_ is k (s s u)T du k ds

Zt

0 Z1=k

0

0

k

0

k(s s u)hM_ is(s s u)T k du ds sup

0u1=k

Zt

0

k(s s u)hM_ is(s s u)T k ds:

The latter supremum converges in idempotent probability to 0 as k ! 1 by hypotheses.

© 2001 by Chapman & Hall/CRC

138

Maxingales

Finally, if s in the statement of the theorem is not bounded, we de ne sk = nk ( ks k )s : Then by hypotheses Zt

0

k(s sk )hM_ is(s sk )T k ds Zt

0 kshM_ issT k 1( ks k k) ds ! 0

as k ! 1, and since associated with the k idempotent processes Z k () are A-local exponential maxingales by the part already proved, Lemma 2.5.10 implies that Z () is an A-local exponential maxingale. The fact that h M i is a proper idempotent process is obvious. For idempotent Ito integrals we have the following corollary.

Theorem 2.5.12.

Let W be an Rd -valued A-Wiener idempotent process. Let (s (!); s 2 R+ ; ! 2 ) be an RmR d -valued Aprogressively measurable idempotent process such that 0t ks k 2 ds < 1 and for a function nA as in Theorem 2.5.11 Zt

ksnA( ks k ) s+Æ nA( ks+Æ k )k 2 ds ! 0

0 Zt

0

as Æ ! 0; t 2 R+ ;

for all A large enough, ksk 2 1( ks k > A) ds ! 0

as A ! 1; t 2 R+ :

Then W is an A-local maxingale with the quadratic characteristic Zt

h W it = ssT ds; 0

which is a proper process.

Remark 2.5.13. R imply that

t 0 ks

The convergence conditions in Theorem 2.5.12 s+Æ k 2 ds ! 0 as Æ ! 0. Therefore, by M.

© 2001 by Chapman & Hall/CRC

139

Idempotent stochastic integrals

Riesz's criterion for relative compactness in L2 , see, e.g., Kantorovich and Akilov [70], under the hypotheses the idempotent distribution of ! ! (s (!); s t) is a deviability in L2 ([0; t]; R md ) for all t 2 R+ .

We now prove that in analogy with stochastic calculus under certain regularity conditions local maxingales with quadratic characteristics are idempotent Ito integrals. We adapt notation (1.7.2) to denote

K (a) = f! 2 : (!) ag; a 2 (0; 1]:

Theorem 2.5.14. Let an Rd -valued continuous A-adapted idempotent process M be an A-local Rmaxingale with an absolutely continut T ous quadratic characteristic ( 0 s s ds; t 2 R+ ), which is a proper idempotent process, where the s are d d matrices. If for t 2 R+ and a 2 (0; 1] Zt

ks s+Æ k 2 ds ! 0

as Æ ! 0;

0

inf inf inf s (!)sT (!) > 0;

!2K (a) st 2Rd : jj=1

then there exists a d-dimensional such that M = W .

A-Wiener idempotent process W

Proof. We de ne a continuous idempotent process W = 1 M; where the right-hand side is well de ned by Lemma 2.5.8. Clearly, M = W . Let nA (x) = 1(x A), where x 2 R+ and A 2 R+ . Then, denoting s 1 = s = 0 for s < 0, we have Zt

0

k s 1Æ nA( ks 1Æ k ) s 1nA( ks 1 k ) ssT T

s 1Æ nA ( ks 1Æ k ) s 1 nA( ks 1 k )

=

Zt

0

k ds

ks 1Æ snA( ks 1Æ k ) s Æ nA( ks 1 k )

s nA( ks 1Æ k ) s Æ nA ( ks 1 k ) T (sT Æ )

© 2001 by Chapman & Hall/CRC

1

k ds

140

Maxingales

Zt

sup kssT k 1 st sup kssT k 1 st

0

ksnA( ks 1Æ k ) s Æ nA( ks 1 k )k 2 ds

15

Zt

ks s Æ k 2 ds

0

+ 10

Zt

0

ksk 2 1( ks 1 k > A) ds :

By hypotheses the latter converges in idempotent probability to 0 as Æ ! 0 for all large A. Hence, by Theorem 2.5.11 W is an A-local maxingale with the quadratic characteristic (Ed t; t 2 R+ ); hence, W is a Wiener idempotent process by Corollary 2.5.2.

Remark 2.5.15. One can replace the in mum over s above by the essential in mum with respect to Lebesgue measure. This remark also concerns other conditions of a similar sort. We now consider versions for strictly Luzin idempotent processes de ned on Hausdor topological spaces with deviabilities. We start with a lemma on properties of trajectories. Lemma 2.5.16. Let be a dHausdor topological space and be a deviability on . Let an R -valued continuous A-adapted strictly Luzin idempotent process M be an A-local maxingale on ( ; ) with an absolutely continuous quadratic characteristic hM i such that (hM_ is ; s 2 R+ ) is a strictly Luzin idempotent process and for t 2 R+ Zt

0

khM_ isk 1( khM_ isk > A) ds ! 0

as A ! 1:

Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively meaRt T _ surable strictly Luzin idempotent process such that 0 ks hM is s k ds < 1 and Zt

0

ks hM_ issT k 1( ks k > A) ds ! 0

R Then both M and ( 0t s hM_ is sT ds; t continuous idempotent processes.

© 2001 by Chapman & Hall/CRC

as A ! 1; t 2 R+ :

2 R+ ) are strictly Luzin-

141

Idempotent stochastic integrals

Proof. We rst check that M is strictly Luzin. Let ! 2 K (a), where a 2 (0; 1], and ! ! !0 as 2 . We write for t 2 R + and A 2 R+ , denoting kA ( ) = (A + 1 k k )+ ^ 1 for a matrix ,

k Mt (!) Mt (!0 )k 2 sup

!2K (a)

Zt

s (!) kA (s (!)) M_ s (!) ds

0 t Z Zt

+ kA (s (! ))M_ s (! ) ds kA (s (!0 ))M_ s (! ) ds

0 0 t Z Zt

0 _ + kA (s (! ))Ms (! ) ds kA (s (!0 ))M_ s (!0 ) ds

: 0 0

(2.5.18)

We estimate the rst term on the right-hand side of (2.5.18) as

Zt

0

s (!) kA (s (!)) M_ s (!) ds

Z t

0

1=2

ks(!)hM_ is(!)s(!)T k 1( ks (!)k > A) ds Zt

1=2

M_ s (!) hM_ is (!) M_ s (!) ds

ds;

0

which converges to 0 as A ! 1 uniformly over ! 2 K (a) by Lemma 2.5.8 and the hypotheses. For the second term on the right-hand side of (2.5.18) we write

Zt

_ s (! ) ds

kA (s (! ))M 0

© 2001 by Chapman & Hall/CRC

Zt

0

kA (s (!0 ))M_ s (! ) ds

142

Maxingales

Z t

kA (s (! )) kA (s (!0 )) hM_ is (! )

0

T 1=2

ds

kA (s (! )) kA (s (!0 )) Zt

1=2

M_ s (! ) hM_ is (! ) M_ s (! ) ds

:

0

The second multiplier on the right is bounded in by Lemma 2.5.8. The integrand in the rst one can be estimated for B 2 R+ as

kA (s (! )) kA (s (!0 )) hM_ is (! ) kA (s (! )) kA (s (!0 )) T

B kA(s(!)) kA(s(!0 ))k2 + 4(A + 1)2 khM_ is (! )k 1( khM_ is (! )k > B );

which implies by hypotheses and Lebesgue's dominated convergence theorem that the second term on the right-hand side of (2.5.18) converges to 0 as 2 . Finally, the third term on the right-hand side of (2.5.18) converges to 0 as 2 by the following argument: since hM is (! ) ! Rt Rt 0 _ hM is (! ), it follows that f M (! ) ds ! 0 fsM_ s (!0 ) ds for step Rt 0 s s functions fs; since the 0 M_ s (! )hM_ is (! ) M_ s (! ) ds are uniformly Rt bounded and 0 M_ s (!0 ) hM_ is (!0 ) M_ s (!0 ) ds is nite, by the CauchyShwarz inequality the class of functions fs for which the latter convergence holds is closed under bounded pointwise convergence; so a monotone class argument shows that it contains all bounded Borelmeasurable functions fs (see the proof of (2.7.28) in the proof of Lemma 2.7.17 below for a more detailed argument of this sort). Now, in order to check that M is strictly Luzin-continuous it is suÆcient to show that uniformly over ! 2 K (a), where a 2 (0; 1], the functions ( Mt (!); t 2 R+ ) are uniformly continuous in t 2

© 2001 by Chapman & Hall/CRC

143

Idempotent stochastic integrals

2 R+ . We have for s; t 2 [0; T ] k Mt (!) Ms(!)k 2

[0; T ] for every T Zt

ku(!)hM_ iu(!)u (!)T k du s

Zt

s

M_ u (!) hM_ iu (!) M_ u (!) du:

The second term on the right is bounded on K (a) by Lemma 2.5.8. The rst term is not greater for A 2 R+ and B 2 R+ than Zt

s

ku (!)hM_ iu(!)u (!)T k 1( ku (!)k > A) du + A2

Zt

s

khM_ iu(!)k 1( khM_ iu(!)k > B ) du + A2 B (t s);

which implies the required in view of the hypotheses. R The proof of ( 0t s hM_ is sT ds; t 2 R+ ) being strictly Luzincontinuous uses similar ideas. Let ! ! !0 , where ! 2 K (a). Then by hypotheses and Lebesgue's dominated convergence theorem for t 2 R+ , A 2 R+ , B 2 R+ , and 2 Rd lim

Zt

0

kA (s (! ))kB (hM_ is (! ))kA (sT (! )) ds =

Zt

0

Since Zt

0

kA (s (!0 ))kB (hM_ is (!0 ))kA (sT (!0 )) ds:

s (! )hM_ is (! )sT (! ) ds Zt

kA (s(!))kB (hM_ is(!))kA (sT (!)) ds 0

© 2001 by Chapman & Hall/CRC

144

Maxingales

+j

Zt

j

2

0

ks(!)hM_ is(!)sT (!)k 1( ks(! )k > A) ds +j

j

2 A2

Zt

0

khM_ is(!)k 1( khM_ is(!)k > B ) ds

and by hypotheses lim sup

A!1

Zt

0

ks(!)hM_ is(!)sT (! )k 1( ks (!)k > A) ds = 0; lim sup

B !1

Zt

0

khM_ is(!)k 1( khM_ is(!)k > B ) ds = 0;

we conclude that lim sup

Zt

0

s(! )hM_ is (! )sT (! ) ds Zt

s(!0 )hM_ is(!0)sT (!0) ds: 0

Fatou's lemma provides the reverse inequality. Thus, R ( 0t shM_ is sT ds; t 2 R+ ) is strictly Luzin. Next, for s; t 2 [0; T ]

Zt

u 0 ZT

hM_ iu uT du

Zs

0

uhM_ iu uT du

kuhM_ iu uT k 1( ku k > A) du 0

+ A2

ZT

0

khM_ iuk 1( khM_ iuk > B ) du + A2B jt sj; R

which implies uniform continuity of ( 0t s hM_ is sT ds; t [0; T ] uniformly over ! 2 K (a).

© 2001 by Chapman & Hall/CRC

2

R+ )

on

145

Idempotent stochastic integrals

Remark 2.5.17. The hypotheses imply that both M

and hM i are

also strictly Luzin-continuous. Theorem 2.5.18. Let be da Hausdor topological space and be a deviability on . Let an R -valued, continuous, strictly Luzin Aadapted idempotent process M be an A-local maxingale on ( ; ) with an absolutely continuous quadratic characteristic hM i such that (hM_ is ; s 2 R+ ) is a strictly Luzin idempotent process and sup sup khM_ is (!)k < 1; t 2 R+ ; a 2 (0; 1]: st !2K (a) Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively meaR surable strictly Luzin idempotent process such that 0t ks k 2 ds < 1; t 2 R+ ; and Zt

0

ks k 2 1( ks k > A) ds ! 0

as A ! 1; t 2 R+ :

Then the idempotent process M is an A-local maxingale with the quadratic characteristic Zt

h M it = shM_ issT ds: 0

Both M and h M i are strictly Luzin-continuous idempotent processes. Proof. We check the local maxingale property for M . Taking in the hypotheses of Theorem 2.5.11 nA(x) = (A + 1 x)+ ^ 1, we have by hypotheses that ! ! (s (!)nA ks (!)k ; s 2 [0; t]) is a continuous mapping from K (a) to L2 ([0; t]; R md ). Since Zt

0

k snA( ks k ) s Æ nA( ks Æ k ) hM_ is

s nA( ks k ) s Æ nA ( ks Zt

T

Æk )

k ds

sup khM_ isk ksnA( ks k ) s Æ nA( ks Æ k )k 2 ds st

© 2001 by Chapman & Hall/CRC

0

(2.5.19)

146

Maxingales

and by M. Riesz's criterion for relative compactness in L2 the righthand side of (2.5.19) converges to 0 as Æ ! 0 uniformly over ! 2 K (a), we conclude by Theorem 2.5.11 that M is an A-local maxingale with the quadratic characteristic in the statement of the theorem. Both M and h M i are strictly Luzin-continuous by Lemma 2.5.16. The following version of Theorem 2.5.12 is proved along the lines of the proof of Theorem 2.5.18.

Theorem 2.5.19. Let dW

be an Rm -valued A-Wiener idempotent process and X be an R -valued A-adapted Luzin-continuous idempotent process with idempotent distribution X . Let (s (x); s 2 R + ; x 2 C (R + ; R d )) be an R km -valued C(R+ ; R d )-progressively measurableR strictly Luzin idempotent process on (C (R + ; Rd ); X ) such that 0t ks (x)k 2 ds < 1; t 2 R+ ; x 2 C (R + ; Rd ); and Zt

0

ks(x)k 2 1( ks (x)k > A) ds ! 0 X

as A ! 1; t 2 R+ :

Then the idempotent process (X ) W is an A-local maxingale with the quadratic characteristic Zt

h(X ) W it = s(X )s (X )T ds; 0

which is a Luzin-continuous idempotent processes.

Remark 2.5.20. For the convergence condition in the hypotheses to hold it is suÆcient that for every compact K C and t 2 R+ Zt

0

sup ks (x)k 2 ds < 1; x2K

in particular, it is suÆcient that (s (x); s satis es the linear-growth condition

2

R+ ;

x 2 C (R + ; Rd ))

kt (x)k 2 lt(1+supjxsj2 ); x 2 C (R + ; Rd ); t 2 R+ ; st

R where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ .

© 2001 by Chapman & Hall/CRC

Idempotent stochastic integrals

147

Remark 2.5.21.

Note that the maxingale property in Theorem 2.5.19 Rdoes not generally hold for discontinuous s (x). For example, 0t sign (Ws )W_ s ds; t 2 R+ , where W is an R-valued idempotent Wiener process, is not a local maxingale, since -a.e. Rt _ 0 sign (Ws )Ws ds = jWt j:

The next consequence of Theorem 2.5.14 considers strictly Luzin processes on C (R + ; Rd ).

Theorem 2.5.22. Let be a deviability on C (R + ; Rd ) and d C (R+ ; R ) denote the -completion of dthe - ow C(R+ ; Rd ). Let M = (Mt (x); t 2 R+ ; x 2 C (R + ; R )) be a strictly Luzin idempotent process, which is a C(R+ ; RdR)-local maxingale on (C (R + ; Rd ); ) with quadratic characteristic ( 0t s(x)s (x)T ds; t 2 R + ), where (s (x); s 2 R + ; x 2 C (R + ; R d )) is an R dd -valued C(R+ ; Rd )-progressively measurable R strictly Luzin idempotent process on (C (R + ; Rd ); ) such that 0t ks (x)k 2 ds < 1; t 2 R+ ; x 2 d C (R + ; R Zt

0

and

);

ks (x)k 2 1( ks (x)k > A) ds ! 0 as A ! 1, t 2 R+ ;

inf inf inf s (x)s (x)T > 0; a 2 (0; 1]: jj=1

x2K (a) st 2Rd :

Then there exists a strictly Luzin-continuous d-dimensional C(R+ ; Rd )-Wiener idempotent process W on (C (R + ; Rd ); ) such that M = W . Proof. The hypotheses and M. Riesz's criterion for compactness in L2 imply that Zt

ks s+Æ k 2 ds ! 0

0

as Æ ! 0. Thus, the hypotheses of Theorem 2.5.14 hold so that, according to the proof of the theorem, W = 1 M is a Wiener idempotent process. It is strictly Luzin-continuous by Lemma 2.5.16.

© 2001 by Chapman & Hall/CRC

148

Maxingales

Theorem 2.5.23. Let the - ow A be right-continuous. Let an Rvalued continuous A-adapted idempotent process M be an A-local

maxingale with a continuous quadratic characteristic hM i such that limt!1 hM it = 1 in idempotent probability. Then there exists a Wiener idempotent process W such that Ms = WhM is -a.e. Proof. Let s = inf ft 2 R+ : hM it > sg: The function s is an A-stopping time by Lemma 2.2.18 and is a proper idempotent variable. Since hM is = s, Lemma 2.3.13 implies that (exp(Mt^s 2 hM it^s =2); t 2 R+ ) is a uniformly maximable Aexponential maxingale so that by Theorem 2.3.8 the idempotent process (exp(Ms 2 s=2); s 2 R+ ) is an exponential maxingale rela~ = (A~s; s 2 R+ ), where A~s = As and M1 = 0 tive to the - ow A for de niteness. Since the idempotent process Ws = Ms also is ~ -adapted by Lemma 2.2.19, a.e. continuous by Lemma 2.5.4 and A ~ -Wiener idempotent process by Theorem 2.4.2. The it is an A proof is complete if we show that MhM is = Ms -a.e. Since fhM is < tg = fhM it > hM isg, hM is s and A is right-continuous, it follows that hM is is an A-stopping time. Therefore, Lemma 2.5.3 implies that for a > 0, b > 0 and c > 0

( sup jMt Ms j a) ec(b a) _(hM ihM is hM is > 2b=c): sthM is Since hM ihM is = hM is -a.e., we conclude that (supsthM is jMt Ms j > 0) = 0.

Remark 2.5.24. Note that the - ow A~ is also right-continuous by Lemma 2.1.11 and the fact that s is right-continuous.

We now consider analogues of Girsanov's theorem.

Theorem 2.5.25. Let W

be an Rm -valued A-Wiener idempotent process on ( ; ). Let (bs (!); s 2 R+ ; ! 2 ) be an Rm -valued Aprogressively measurable idempotent process such that the idempotent R R processes exp 0t ( bs + ) W_ s ds 0t j bs + j2 ds=2 ; t 2 R+ are well-de ned A-exponential maxingales under for all 2 Rm . Let

Mt = exp

Zt

0

1 bs W_ s ds 2

© 2001 by Chapman & Hall/CRC

Zt

0

jbsj2 ds :

149

Idempotent stochastic integrals

If there exists an idempotent probability 0 on such that its restrictions 0t to the -algebras At are expressed as d0t = Mt d; t 2 R+ , then the idempotent process X = (Xt ; t 2 R+ ), de ned by

Xt =

Zt

bs ds + Wt ;

0

is an A-Wiener idempotent process under 0 .

Proof. We rst note that since Mt > 0, the sets of 0 -idempotent probability 0 have -idempotent probability 0; hence, the ow A is 0 -complete. For s t and 2 Rm in view of Lemma 1.6.35

S0 exp Xt

1 2 jj t jA 2 s S Mt exp Xt 12 jj2 t jAs = : (2.5.20) S (Mt jAs )

By the property of R de nition of X and R M , and the maxingale exp 0t ( bs + ) W_ s ds 0t j bs + j2 ds=2 ; t 2 R+

S Mt exp Xt

= S exp Zs

= exp

Zt

1 2 jj t jAs 2 1 2

( br + ) W_ r dr

0

( br + ) W_ r dr

0

1 2

Zs

Zt

j br + j2 dr jAs

0

j br + j2 dr

0

1 2 jj s : 2 Also S (Mt jAs ) = Ms by the maxingale property of Mt . Thus, by (2.5.20) = Ms exp Xs

1 2 1 j j t jAs = exp Xs jj2 s : 2 2 The 0 -maximability property is obvious.

S0 exp Xt

© 2001 by Chapman & Hall/CRC

150

Maxingales

We now give a version for the canonical setting. We denote C (R + ; R d ) by C .

Theorem 2.5.26. Let space C C (R + ; Rm ) be endowed with a deviability such that the idempotent process W de ned by Wt (x; w) = wt is an Rm -valued Wiener idempotent process. Let Y be de ned by Yt (x; w) = xt and Y denote the idempotent distribution of Y . Let (bs (x); s 2 R+ ; x 2 C ) be an Rm -valued C -progressively measurable Y

bounded strictly Luzin idempotent process on (C ; ). Let M and X be de ned as in Theorem 2.5.25. Then there exists an idempotent probability 0 on C such that its restrictions 0t to the -algebras Ct Ct(R+ ; Rm ) are expressed as d0t = Mt d; t 2 R+ . The idempotent process X is a Wiener idempotent process under 0 . Proof. Let A denote the natural - ow on C C (R + ; Rm ) completed with respect theidempotent proR t to . Theorem 2.5.19 R timplies that 2 _ cess exp 0 ( bs + ) Ws ds 0 j bs + j ds=2 ; t 2 R+ is an A-local exponential maxingale under for arbitrary 2 Rm ; it is actually maxingale, which is derived from the fact R t an exponential 2 that 0 j bs + j ds is a bounded idempotent variable by a standard argument (cf. the proof of Lemma 2.5.10). We now check existence of 0 . We rst show that the idempotent probabilities 0t de ned by d0t = Mt d are deviabilities on C C (R + ; Rm ). Since M is an exponential maxingale on (C C (R + ; Rm ); ), by Lemma 1.7.20 it is enough to check that M (x; w) is strictly Luzin on (C C (R + ; Rm ); ), which follows by Lemma 2.5.16 and the fact that bs (x) is strictly Luzin and bounded. Let 00t denote the deviabilities on the spaces C ([0; t]; R d Rm ) of continuous Rd Rm -valued functions on [0; t] that are the images of 0t under the mappings x ! (xs ; s 2 [0; t]). The maxingale property 00 d m of M implies that t ; C ([0; t]; R R ) is a projective system so that by Lemma 1.8.3 and the fact that the projective limit of the C ([0; t]; R d R m ); t 2 R + ; is homeomorphic to C C (R + ; R m ) we conclude that there exists a deviability 0 on C C (R + ; Rm ) that extends the 0t . The fact that X is idempotent Wiener follows by Theorem 2.5.25.

© 2001 by Chapman & Hall/CRC

151

Idempotent Ito equations

2.6 Idempotent Ito dierential equations This section studies idempotent analogues of Ito dierential equations. We x space C (R + ; Rd ), which we denote throughout the section by C ; the associated - ow C(R+ ; R d ) = (Ct (R+ ; Rd ); t 2 R+ ) is denoted as C = (Ct ; t 2 R+ ). Given space C (R + ; Rm ) we denote m m by C W t (R + ; R ) the completion of Ct (R + ; R ) with respect to the Wiener idempotent probability W on C (R + ; Rm ). Let bt (x); t 2 R+ ; x 2 C ; and t (x); t 2 R+ ; x 2 C ; be respective d R -valued and Rdm -valued functions, which are continuous in x for every t, C-progressively measurable in (t; x), and Zt

0

jbs(x)j ds < 1;

Zt

0

ks(x)k 2 ds < 1; t 2 R+ :

We introduce the equation Zt

Zt

0

0

Xt = X0 + bs(X ) ds + s (X )W_ s ds;

(2.6.1)

where W = (Ws ; s 2 R+ ) is an Rm -valued Wiener idempotent process, X = (Xs ; s 2 R+ ) is an Rd -valued continuous idempotent process, and the second integral on the right is an Ito idempotent integral. The ow A in the following de nitions is assumed to be complete with respect to the associated idempotent probability.

De nition 2.6.1. We sayd that equation (2.6.1) with an initial idempotent distribution on R has a solution if there exist an idempotent probability space ( ; ) with a - ow A, an Rd -valued continuous A-adapted idempotent process X = (Xs; s 2 R+ ) and an Rm -valued A-Wiener idempotent process W = (Ws; s 2 R+ ) on ( ; ) such that X0 has the idempotent distribution , and (2.6.1) holds for all t 2 R+ -a.e. in ! 2 . The pair (X; W ) is then called a solution to the equation. We say that existence holds for (2.6.1) if a solution exists for every .

De nition 2.6.2. We say that equation (2.6.1) with an initial deviability distribution on Rd has a Luzin solution if there exist an idempotent probability space ( ; ) with a - ow A, an Rd -valued © 2001 by Chapman & Hall/CRC

152

Maxingales

continuous A-adapted idempotent process X = (Xs ; s 2 R+ ) and an Rm -valued A-Wiener idempotent process W = (Ws ; s 2 R+ ) on ( ; ) such that (X; W ) is a Luzin-continuous idempotent process on C C (R + ; R m ), X0 has the idempotent distribution , and (2.6.1) holds for all t 2 R+ -a.e. in ! 2 . The pair (X; W ) is then called a Luzin solution to the equation with initial deviability distribution . We say that Luzin existence holds for (2.6.1) if a Luzin solution exists for every deviability .

Remark 2.6.3. We sometimes loosely refer to X alone as a solution (respectively, Luzin solution).

Remark 2.6.4.

We recall that by de nition (X; W ) is a Luzincontinuous idempotent process if the idempotent distribution of (X; W ) is a deviability on C C (R + ; Rm ).

Remark 2.6.5. Let (X; W ) be a solution (respectively, Luzin soX;W

lution) on some idempotent probability space and denote the idempotent distribution of (X; W ). Then by the transitivity property of conditional idempotent expectations the canonical idempotent process (x; w) on C C (R + ; Rm ) with the natural - ow completed with respect to X;W is a solution (respectively, Luzin solution) under X;W . Thus, a solution (respectively, Luzin solution) can always be implemented on the canonical space. Therefore, we occasionally refer to the idempotent distribution of (X; W ) as a solution (respectively, Luzin solution) as well. We denote by (respectively, x ) the idempotent distribution of (X; W ) associated with an initial idempotent distribution (respectively, with an initial condition X0 = x 2 Rd ).

De nition 2.6.6.

We say that uniqueness (respectively, Luzin uniqueness) holds for (2.6.1) if for every two solutions (respectively, Luzin solutions) (X; W ) and (X 0 ; W 0 ) such that the idempotent distributions of X0 and X00 coincide the idempotent distributions of (X; W ) and (X 0 ; W 0 ) also coincide.

De nition 2.6.7. We say that strong existence holds for equation d m

(2.6.1) if there exists a function F : R C (R + ; R ) ! C such m that the function w ! F (x; w) is C W t (R + ; R )=Ct -measurable for every x 2 Rd and t 2 R+ and, given an A0 -measurable idempotent variable f 2 Rd and an Rm -valued A-Wiener idempotent process W = (Ws ; s 2 R+ ) both de ned on an idempotent probability space

© 2001 by Chapman & Hall/CRC

Idempotent Ito equations

153

( ; ) with a - ow A, the idempotent process X = F (f; W ) satis es (2.6.1) for all t 2 R+ -a.e., and X0 = f -a.e. The idempotent process X is then called a strong solution to the equation with an initial condition f . If the function F (x; w) is, in addition, continuous on Rd KW (a), a 2 (0; 1], then we say that Luzin strong existence holds, and X is called a Luzin strong solution. Remark 2.6.8. We recall that KW (a) = fw 2 C (R + ; Rm ) : W (w) ag. Remark 2.6.9. We note that a strong solution X is AW-adapted since W is A-adapted, A is complete and w ! F (x; w) is C t (R+ ; Rm )=Ct measurable. De nition 2.6.10. We say that (2.6.1) has a unique strong solution (respectively, Luzin strong solution) if strong existence (respectively, Luzin strong existence) holds and if, given a solution (respectively, Luzin solution) (X; W ) with an initial condition X0 on an idempotent probability space ( ; ) with a - ow A, we have that X = F (X0 ; W ) -a.e., where F is the function in the de nition of a strong solution (respectively, Luzin strong solution). De nition 2.6.11. We say that pathwise uniqueness holds for (2.6.1) if for every two solutions (X; W ) and (X 0 ; W 0 ), which are de ned on the same idempotent probability space ( ; ) with the same - ow A, we have X = X 0 -a.e. provided X0 = X00 and W = W 0 -a.e. We refer to (2.6.1) as an idempotent Ito dierential equation and to X as an idempotent diusion. The functions bt (x) and t (x)t (x)T are occasionally referred to as in nitesimal drift and diusion coef cients, respectively. We also use the following short-hand notation for (2.6.1) X_ t = bt (X )+ t (X )W_ t : It is clear that uniqueness implies Luzin uniqueness, Luzin strong existence implies strong existence, and strong existence implies existence. We study some other relationships between the introduced concepts. We rst discuss the role of initial conditions in the above de nitions. The next lemma follows by Lemma 1.5.5 and Theorem 1.8.9. It implies, in particular, that if Luzin existence holds, then existence holds.

© 2001 by Chapman & Hall/CRC

154

Maxingales

Lemma 2.6.12.

1. If for every initial condition x 2 Rd there exists a solution x , then existence holds. Speci cally, given an initial idempotent distribution , the idempotent distribution de ned by (x; w) = supx2Rd x (x; w)(x) is a solution to (2.6.1).

2. If for every initial condition x 2 Rd there exists a Luzin solution x , which is a deviability transition kernel from Rd into C C (R + ; R m ), then Luzin existence holds. More speci cally, given an initial deviability , the idempotent probability de ned as in part 1 is a Luzin solution with initial deviability distribution .

Lemma 2.6.13. If the idempotent distributions of0 every two solu0 tions (respectively, Luzin solutions) (X; W ) and (X ; W ) with initial condition x coincide for every x 2 Rd , then uniqueness (respectively, Luzin uniqueness) holds.

Proof. Let X be a solution with an initial idempotent distribution de ned on an idempotent probability space ( ; ) with a - ow A. For x 2 Rd such that (x) > 0, let x(A) = (AjX0 = x); A

. Then it follows from the de nition of conditional idempotent probability that X0 = x x -a.e.; since W is independent of A0 by the de nition of an A-Wiener idempotent process, it is independent of X0 , so that W is an A-Wiener idempotent process on ( ; x ); (2.6.1) holds x -a.e. since it holds -a.e. Thus, X is a solution of (2.6.1) with initial condition x on the space ( ; x ) for the AWiener idempotent process W ; hence, the idempotent distribution of (X; W ) under x is speci ed uniquely. By Theorem 1.6.12 we have for A C C (R + ; Rm ) that

((X; W ) 2 A) = sup x ((X; W ) 2 A)(X0 = x) x2Rd = sup x ((X; W ) 2 A)(x); (2.6.2) x2Rd which implies that the idempotent distribution of (X; W ) under is speci ed uniquely. We now turn to Luzin uniqueness. We assume the canonical setting so that = C C (R + ; Rm ), is a deviability on , and (X; W ) is the canonical idempotent process. Let be an initial

© 2001 by Chapman & Hall/CRC

Idempotent Ito equations

155

deviability distribution on Rd . Then x de ned as in the preceding proof is a deviability on by Theorem 1.6.12 and the fact that the set fX0 = xg belongs to the collection of closed subsets of . Since (X; W ) is a Luzin solution such that X0 = x x -a.e., deviability x is speci ed uniquely. Since (2.6.2) holds for this case as well, is speci ed uniquely. The next lemma follows by similar arguments and the de nitions. Lemma 2.6.14. 1. If there exists a function F (x; w) in the de nition of a strong solution (respectively, Luzin strong solution) such that the idempotent process X = F (x; W ) is a strong solution (respectively, Luzin strong solution) for an initial condition x, then strong existence (respectively, Luzin strong existence) holds. 2. If pathwise uniqueness holds for solutions with initial condition x for every x 2 Rd , then pathwise uniqueness holds. Let us associate with (2.6.1) a collection of ordinary dierential equations depending on absolutely R 1 continuous functions w = (wt ; t 2 R + ) 2 C (R + ; R m ) such that 0 jw_ t j2 dt < 1:

x_ t = bt (x)+t (x)w_ t a.e. in t; x0 = x 2 Rd ; (2.6.3) where the x = (xt ; t 2 R+ ) 2 C are absolutely continuous functions. De nition 2.6.15. We say that the extension condition holds for d equation (2.6.3) if for every x 2 R and w 2 C (R + ; Rm ) such that R1 2 0 jw_ t j dt < 1 the following holds: if a function (xt ), de ned on an interval [0; T ], satis es (2.6.3) for t 2 [0; T ], then it can be extended to a solution of (2.6.3) on R+ . Remark 2.6.16. The extension condition implies existence of solutions for every equation (2.6.3). Lemma 2.6.17. 1. If the extension condition holds for (2.6.3), then existence holds for the idempotent Ito dierential equation (2.6.1). If existence holds for (2.6.1), then every ordinary dierential equation (2.6.3) has a solution. 2. If every ordinary dierential equation (2.6.3) has at most one solution, then pathwise uniqueness and uniqueness hold for (2.6.1).

© 2001 by Chapman & Hall/CRC

156

Maxingales

3. Luzin strong existence implies Luzin existence. 4. If strong existence (respectively, Luzin strong existence) and pathwise uniqueness hold for (2.6.1), then there is a unique strong solution (respectively, Luzin strong solution). Proof. Let the extension condition hold for (2.6.3). We de ne an idempotent distribution x on C C (R + ; Rm ) by x (x; w) = W (w) if x and w satisfy (2.6.3) and x0 = x, and x (x; w) = 0 otherwise. Let (X; W ) denote the canonical idempotent process on C C (R + ; R m ) and A = fAt ; t 2 R + g be the natural - ow completed with respect to x. Then (X; W ) satis es (2.6.1) for the initial condition x x-a.e. We show that W is an A-Wiener idempotent process. It is suÆcient to check that W has A-independent increments, i.e., x (t w0 jAt ) = x (t w0 ) x -a.e., where t 2 R+ and w0 2 C (R + ; Rm ) is such that W (w0 ) > 0. Let (x00 ; w00 ) be such that x pt 1 Æ pt (x00; w00 ) > 0. Then

x(t w0jAt )(x00 ; w00 ) 0 00 00 x (x; w) : t w = t w ; pt (x; w) = pt (x ; w ) = x (x; w) : pt (x; w) = pt(x00 ; w00 ) : Since x pt 1 Æ pt (x00 ; w00 ) > 0, the pair (x00 ; w00 ) satis es (2.6.3) on [0; t]. The extension condition implies that for every w such that pt w = pt w00 and t w = t w0 there exists a solution to (2.6.3) on R+ that coincides with x00 on [0; t]. Therefore, by the de nition of x we have that

x (x; w) : t w = t w0 ; pt(x; w) = pt (x00 ; w00 )

= W (w : t w = t w0 ; pt w = pt w00 ):

By a similar reasoning

x (x; w) : pt (x; w) = pt(x00 ; w00 ) = W (w : pt w = pt w00 ): Thus, by independence of increments of W

x(t w0jAt )(x00 ; w00 ) =

© 2001 by Chapman & Hall/CRC

: t w = t w0 ; pt w = pt w00 ) W (w : pt w = pt w00 ) = W (t w0 )

W (w

157

Idempotent Ito equations

as required. Thus, (X; W ) is a solution of (2.6.1) on (C C (R + ; R m ); x ) with - ow A for every x 2 Rd . By Lemma 2.6.12 existence holds. Conversely, let existence hold. Let x be an idempotent distribution of (X; W ) on C C (R + ; Rm ) for an initial condition x. Since supx x (x; w) = W (w), it follows that if W (w) > 0, then there exists x such that x (x; w) > 0 so that (x; w) satisfy (2.6.3) and x0 = x. This ends the proof of part 1. We now prove part 2. The fact that pathwise uniqueness holds if every dierential equation (2.6.3) has at most one solution is obvious. Let us assume that x is a solution on C C (R + ; R m ) with an initial condition x. By de nition if x (x; w) > 0, then x solves (2.6.3) with x0 = x. By uniqueness for (2.6.3) x is a unique solution for given w W and x. Since we must have that supx x (x; w) = (w) it follows that x (x; w) = W (w) so that x coincides with the solution x de ned in the proof of part 1. Luzin strong existence implies Luzin existence by Theorem 1.7.11. Part 4 is obvious.

Remark 2.6.18. Note that under the hypotheses of part 20 we have

pathwise uniqueness even for two solutions (X; W ) and (X ; W ) that are not necessarily associated with the same - ow. The latter is also true if there is a unique strong solution.

The following lemma takes advantage of the proof of Lemma 2.6.17 to indicate a candidate for a solution of (2.6.1). The proof also relies on Lemma 2.6.12.

Lemma 2.6.19. Let the extension condition hold for (2.6.3). Then the idempotent distribution

(x; w) = supd x(x; w)(x); x2R

where

8 W > :

0;

if x_ t = bt (x) + t (x)w_ t a.e. and x0 = x; otherwise;

is a solution for an initial idempotent distribution . If x (x; w) is a deviability transition kernel from Rd into C C (R + ; R m ), then is a Luzin solution for an initial deviability .

© 2001 by Chapman & Hall/CRC

158

Maxingales

Remark 2.6.20. Easy supw x(x; w) is given by

Xx (x) = exp

1 2

calculations show that

Xx (x)

=

Z1

0

(x_ t bt (x)) t (x)t (x)T (x_ t bt (x)) dt

if x0 = x, x is absolutely continuous and x_ t bt (x) is in the range of t (x) a.e., and X x (x) = 0 otherwise. We also note that the range of t (x) coincides with the range of t (x)t (x)T . Theorem 2.6.21. 1. If pathwise uniqueness holds, then uniqueness holds. 2. Let pathwise uniqueness hold. If existence (respectively, Luzin existence) holds, then strong existence (respectively, Luzin strong existence) holds so that there exists a unique strong solution (respectively, Luzin strong solution). Proof. Let (X; W ) and (X 0 ; W 0 ) be two solutions of (2.6.1) with an initial condition x 2 Rd on respective idempotent probability spaces ( ; ) and ( 0 ; 0 ) with respective - ows A and A0 . Let us introduce the conditional idempotent distributions w (A) = (X 2 AjW = w) and 0 w (A) = 0 (X 0 2 AjW 0 = w). We show that for x 2 C and t 2 R+

w pt 1 (pt x) = (pt X = pt xjpt W = pt w) for W {almost all w; (2.6.4)

i.e., the left-hand side depends only on the piece of w up to t. (Of course, a similar relation holds for 0 .) Recalling the notation t ws = wt+s wt ; s 2 R+ ; w 2 C (R + ; Rm ), we have, for w such that W (w) > 0, in view of the fact that pt X and pt W are At -measurable, and t W is independent of At , that (pt X = pt x; W = w w pt t x) = (W = w) (pt X = pt x; pt W = pt w; t W = t w) = (pt W = pt w; t W = t w) (pt X = pt x; pt W = pt w)(t W = t w) = (pt W = pt w)(t W = t w) = (pt X = pt xjpt W = pt w): 1 (p

© 2001 by Chapman & Hall/CRC

159

Idempotent Ito equations

The claim has been proved. ~ on ~ = C C C (R + ; Rm ) We de ne an idempotent probability by ~ x; x0 ; w) = w (x)0 w (x0 )W (w): ( (2.6.5)

Clearly, ~ fxg C fwg = (X = x; W = w) and ~ C fx0 g fwg = 0(X 0 = x0 ; W 0 = w). Let C~t be the completion ~ and C ~ = (C~t ; t 2 R+ ). of Ct Ct Ct (R+ ; Rm ) with respect to ~We check that the canonical idempotent process (wt ; t 2 R+ ) is a C ~ ~ Wiener idempotent process on ( ; ). It obviously has idempotent distribution W . By Theorem 2.4.9 it is suÆcient to check that t w is ~ -independent of C~t . We have, for x^ ; x^ 0 2 C and w; ^ w~ 2 C (R + ; Rm ), in view of (2.6.5), ~ (x; x0 ; w) : pt (x; x0 ; w) = pt (^x; x^ 0 ; w^ ); t w = w~ = sup 1 pt (x; x0 ; w) = pt (^x; x^ 0 ; w^); t w = w~ (x;x0 ;w)2 ~ w (x)0 w (x0 )W (w) = sup 1 pt w = pt w; ^ t w = w~ w (pt 1 (pt x^ )) w 0 w (pt 1 (pt x^ 0 ))W (w) = sup 1(pt w = pt w^ )w (p 1 (pt x^ ))0 w (p 1 (pt x^ 0 ))W (w)

w

sup 1(t w = w~ )W (w)

w

t

t

~ pt (x; x0 ; w) = pt (^x; x^ 0 ; w^ )( ~ t w = w~ ); =

where the equality before the last one follows by (2.6.4) and the fact that t w is independent of pt w under W . Thus, (x; w) and (x0 ; w) are two solutions to (2.6.1) on the same idempotent probability space and adapted to the same - ow. By ~ pathwise uniqueness we conclude that x = x0 -a.e. so ~ x; x0 ; w) = 0: sup 1(x 6= x0 )( (2.6.6) 0 ~ (x;x ;w)2

Therefore, ~ fxg C fwg) = (( ~ x; x; w)) ((X; W ) = (x; w)) = ( ~ C fxg fwg) = 0 ((X 0 ; W 0 ) = (x; w)) = (

© 2001 by Chapman & Hall/CRC

160

Maxingales

so that uniqueness holds. Next, by (2.6.6) and (2.6.5) supx;x0 1(x 6= x0 )w (x)0 w (x0 ) = 0 for W -almost all w. Fixing w such that W (w) > 0 and picking x~ 0 such that 0w (~x0 ) > 0 we de ne F (x; w) = x~ 0. Since x = x~ 0 for w -almost all x, we have that x = F (x; w) whenever (X = x; W = w) > 0. If W (w) = 0, we de ne F (x; w) arbitrarily. By construction, X = F (x; W ) -a.e. m We prove that F (x; w) is C W t (R + ; R )=Ct -measurable in w for every x 2 Rd and t 2 R+ . Since X = F (x; W ) -a.e., we have that (X = xjW = w) = 1(F (x; w) = x) if (W = w) > 0. Therefore, by (2.6.4) for W -almost all w

1(F (x; w) 2 pt 1(pt x)) = (pt X = pt xjW = w) = (pt X = pt xjpt W = pt w): Since the right-most side, for xed x, is a function of pt w, we conm clude that fw : F (x; w) 2 pt 1 (pt x)g 2 C W t (R + ; R ). Thus, strong

existence holds, and by part 4 of Lemma 2.6.17 there exists a unique strong solution. For the part concerned with Luzin solutions, we need to check, in addition, that if Luzin existence holds, then F (x; w) is continuous in (x; w) on Rd KW (a); a 2 (0; 1]. Let (xn ; wn ) 2 Rd KW (a); a 2 ^ W ^ ) be a Luzin (0; 1]; converge to (~x; w~ ) and xn = F (xn ; wn ). Let (X; ^ ^ solution on an idempotent probability space ( ; ) with an initial ^ X^ 0 = xn) = 1; n 2 N ; and ( ^ X^0 = x~) = condition X^ 0 such that ( ^ ^ ^ ^ ^ ^ ^ ^ 1. Since X = F (X0 ; W ) -a.e., ((X; X0 ; W ) = (xn ; xn ; wn )) = ^ X^0 = xn )W (wn ) a. Since (X; ^ W ^ ) is a Luzin solution, the ( ^ ^ ^ ^ set f(x; x; w) : ((X; X0 ; W ) = (x; x; w)) ag is compact, which implies that the sequence f(xn ; xn ; wn ); n 2 N g is relatively compact and every x; x~; w~) is such that ^ (X;^ X^0 ; W^ ) = accumulation point (~ (~x; x~; w~ ) a > 0. Hence, x~ = F (~x; w~ ). According to Lemma 2.6.17 existence and uniqueness issues for (2.6.1) and (2.6.3) are closely related. Thus, the methods of the theory of ordinary dierential equations apply to the study of existence and pathwise uniqueness. We recall that bt (x) and t (x) are assumed to be continuous in x.

© 2001 by Chapman & Hall/CRC

161

Idempotent Ito equations

Theorem 2.6.22.

1. Let bt (x) and t (x) satisfy the lineargrowth conditions

jbt(x)j lt (1 + supjxs j); kt (x)k 2 lt (1 + supjxsj2 ); st st t 2 R+ ; x 2 C ; R where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ : Then existence holds for (2.6.1).

2. Let bt (x) and t (x) be locally Lipshitz continuous, i.e., for every a 2 R+ , t 2 R+ and x; y 2 C ; such that supst jxs j a and supst jys j a we have

jbt (x) bt (y)j kta supjxs ysj; st 2 a kt (x) t (y)k kt supjxs ysj2 ; st

R

where kta is Lebesgue measurable and 0t ksa ds < 1; t 2 R+ : Then pathwise uniqueness holds for (2.6.1). Proof. For the existence part we have to check the extension condition for (2.6.3). We use the method of successive approximations. Standard are omitted. Let x 2 Rd , w 2 C (R + ; Rm ) be such R 1 details 2 that 0 jw_ s j ds < 1, and a function x^ t ; t 2 [0; T ]; satisfy (2.6.3) on [0; T ]. Let unt be successive approximations de ned by

unt +1

Zt

= x+ bs

(un ) ds+

0

Zt

s (un )w_ s ds; t 2 R+ ;

(2.6.7)

0

where u0t = x^ t for t 2 [0; T ] and u0t = x^ T for t T . Then by the Cauchy-Schwarz inequality and the linear-growth conditions

junt +1 j2

© 2001 by Chapman & Hall/CRC

Z t 2 3x +3

jj

0

2

jbs(un )j ds

162

Maxingales

+3

Zt

ks

(un )k 2 ds

0

jj

3x2 +3

Zt

2

Zt

jw_ sj2 ds

0

ls ds +

0

Zt

0 t Z

+3 2

Zt

jw_ sj

2 ds

ls ds +

0

Zt

0

ls ds

0

jw_ sj

2 ds

Zt

0

ls supjunr j2 ds: rs

This string of inequalities shows in particular that the right-hand side of (2.6.7) is well de ned. Denoting ftn = supst juns j2 , we conclude that, given T > 0, there exist constants A1 and A2 such that for all tT

ftn+1

Zt

A1 + A2 lsfsn ds; 0

which implies by a version of Gronwall's inequality that

ftn A1 exp A2

Zt

ls ds ; t T:

0

Therefore, supn supst juns j < 1, which easily implies by (2.6.7) and the linear-growth conditions that the sequence f(unt ; t 2 R+ ); n 2 N g is locally equicontinuous, so that by Arzela-Ascoli's theorem it is relatively compact in C . In a standard way, by using continuity of bt (x) and t (x) in x and the linear-growth conditions, it follows that every accumulation point of f(unt ; t 2 R+ )g solves (2.6.3). It coincides with x^ on [0; T ] since unt = x^ t for t 2 [0; T ]. The uniqueness part is also proved by a standard argument. Let u and v be two solutions of (2.6.3) such that u0 = v0 = x. Let a (x) = inf ft 2 R+ : jxt j ag; x 2 C : Then denoting uat = ut^ a (u)^ a (v) and vta = vt^ a (v)^ a (u) , by the Cauchy-Schwarz inequality and Lipshitz

© 2001 by Chapman & Hall/CRC

163

Idempotent Ito equations

continuity conditions

juat

vta j2

2

Zt

0

ksa ds

Zt

0

ksa supjuar

vra j2 ds

rs

+2

Zt

0

Zt

jw_ sj2 ds ksa supjuar vra j2 ds; 0

rs

so that ua = va by Gronwall's inequality. Hence, a (u) = a (v) and ut = vt for t ua . Letting a ! 1 completes the proof.

Remark 2.6.23. Under the hypotheses of part 2 the stronger version of pathwise uniqueness of Remark 2.6.18 holds.

We now strengthen part 1 of Theorem 2.6.22.

Theorem 2.6.24. Under the hypotheses of part 1 of Theorem 2.6.22 Luzin existence holds for equation (2.6.1).

Proof. Since according to the proof of Theorem 2.6.22 the extension condition holds for (2.6.3) under the hypotheses, by Lemma 2.6.19 it suÆces to check that x(x; w) de ned in the statement of the lemma is a deviability transition kernel. By Lemma 1.8.12 this can be done by proving that x (x; w) is upper semi-continuous in (x; x; w) and the sets f(x; w) : supjxjA x(x; w) ag are relatively compact for every A 2 R+ and a 2 (0; 1]. We consider the upper semi-continuity rst. Let (xn ; xn ; wn ) ! (~x; x~ ; w~ ) as n ! 1. We can obviously assume that the (xn ; xn ; wn ) satisfy equation (2.6.3) for otherwise xn (xn ; wn ) = 0. In addition, since by de nition xn (xn ; wn ) = W (wn ), we may assume that W (wn ) a > 0. We check that x~ is a solution to (2.6.3) associated with x~ and w~ . The linear-growth conditions, continuity of bs (x) and s(x) in x, Lebesgue's dominated convergence theorem, and Lemma 2.5.16 yield

lim

Zt

n!1

lim

n!1

Zt

0

0

bs(x

n ) ds

=

s (xn )w_ sn ds =

© 2001 by Chapman & Hall/CRC

Zt

0 Zt 0

bs (~x) ds; s (~x)w~_ s ds;

164

Maxingales

proving the claim. Thus, x~ (~x; w~ ) = W (w~ ). Since by upper semi-continuity of W (w) we have that lim supn!1 xn (xn ; wn ) = lim supn!1 W (wn ) W (w~ ), the required follows. Let us check that the set f(x; w) : supjxjA x(x; w) ag is relatively compact. Let a sequence f(xn ; xn ; wn ); n 2 N g be such that jxn j A and xn (xn ; wn ) a(1 1=n). Since xn (xn ; wn ) = W (wn ) and W (w) is upper compact, we may assume that the xn and wn converge to some x~ and w~ , respectively. Since the xn are solutions of (2.6.3),

jx

n j2 t

Z t

3(xn )2 + 3

jbs(x

0 Z t

jxnt xnsj2 2

s

n )j ds 2 + 3 2

Zt

jw_ sn j2 ds

Zt

0 Zt

0 Zt

s

s

ks(xn )k 2 ds;

jbr (xn)j dr + 2 jw_ rn j2 dr kr (xn )k 2 dr:

Since also inf n W (wn ) > 0, we conclude in analogy with the proof of Theorem 2.6.22 that the sequence fxn g is relatively compact in C.

Remark 2.6.25. One can also show that in the hypotheses of part 1 of Theorem 2.6.22 the idempotent process X de ned for (x; w) 2 C Rt C (R + ; R m ) by Xt (x; w) = xt x b ( x ) ds is a local maxingale on 0 s m m C C (R + ; R ); C C(R+ ; R ); x with quadratic characteristic Rt T (x) ds; t 2 R . ( x ) s + s 0

Combining Lemma 2.6.19, Theorem 2.6.21, Theorem 2.6.22, and Theorem 2.6.24 we obtain the following existence and uniqueness result.

Theorem 2.6.26. Let bt (x) and t (x) be locally Lipshitz-continuous and satisfy the linear-growth conditions. Then the equation X_ t = bt (X )+ t (X )W_ t ; X0 = x;

has a unique Luzin solution, which is also a strong Luzin solution. The deviability distribution of X is given by 1 Z1 X x (x) = exp (x_ t bt (x)) t (x)t (x)T (x_ t bt (x)) dt 2 0

© 2001 by Chapman & Hall/CRC

165

Idempotent Ito equations

if x0 = x, x is absolutely continuous and x_ t of t (x) a.e., and X x (x) = 0 otherwise.

bt (x) is in the range

Remark 2.6.27. Existence of a strong solution under the hypothe-

ses can also be proved directly by using a version of the method of successive approximations.

As a consequence of \the Girsanov theorem" (Theorem 2.5.26) we have the following existence and uniqueness result.

Theorem 2.6.28. Let (s (x); s 2 R+ ; x 2 C ) be an Rm valued C(R+ ; Rm )-progressively measurable bounded function such that s (x) is continuous in x for s 2 R+ . Then Luzin existence and uniqueness hold for the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x;

(2.6.8)

if and only if Luzin existence and uniqueness hold for the equation X_ t = bt (X ) t (X )t (X ) +t (X )W_ t ; X0 = x;

(2.6.9)

where W is an Rm -valued A-Wiener idempotent process.

Corollary 2.6.29. Let (bs (x); s 2 R+ ; x 2 C ) be bounded. Luzin existence and uniqueness hold for the equation

Then

X_ t = bt (X )+ W_ t ; X0 = x; where W is an Rd -valued A-Wiener idempotent process. Let X denote the idempotent distribution of X . Then X (x) = 0 unless x0 = 0 and x is absolutely continuous. For these x

x) = exp

X (

1 2

Z1

0

jx_t bt(x)j2 dt :

We now outline another approach, which is analogous to the martingale problem approach and which we will explore in detail later in the text. We state the result for Luzin solutions, which are our main concern below. Given a deviability on C , we denote by C the -completion of the - ow C.

© 2001 by Chapman & Hall/CRC

166

Maxingales

Theorem 2.6.30. Let the matrix s(x) have size d d and for every compact K C and t 2 R+

lim sup

a!1 x2K

Zt

0

ks(x)k 2 1( ks (x)k > a) ds = 0;

inf inf s (x)s (x)T > 0; xinf 2K st 2Rd :

lim sup

a!1 x2K

Zt

0

jj=1

jbs(x)j 1(jbs (x)j > a) ds = 0:

Then the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x; has a Luzin solution (X; W ) if and only if there exists a deviability on C such that R x0 = x -a.e., and the idempotent process Mt (x) = xt x 0t bs (x) ds is a C-localR maxingale on (C ; ) with the quadratic characteristic hM it (x) = 0t s (x)s (x)T ds and is strictly Luzin. The solution (X; W ) is unique if and only if the deviability is unique. The idempotent distribution of X then coincides with . Proof. Let (X; W ) be a Luzin solution on an idempotent prob~ with a - ow A. Then by Theorem 2.5.19 ability space ( ; ) Rt _ ( 0 s(X )Ws ds;R t 2 R+ ) is an A-local maxingale with the quadratic characteristic ( 0t s(X )s (X )T ds; t 2 R+ ). The idempotent distribution of X is the required deviability . R t Conversely, let Mt (x) = xt x 0 bs (x) ds be a strictly Luzin idempotent process, which is a RC-local maxingale with the quadratic characteristic hM it (x) = 0t s (x)s (x)T ds on (C ; ). Then by Theorem 2.5.22 there exists a strictly Luzin-continuous CRt _ Wiener idempotent process W on (C ; ) such that Mt = 0 s Ws ds, which implies that the canonical process on (C ; ) and W make up a Luzin solution to the equation.

Remark 2.6.31. Under the hypotheses, both M and hM i are also strictly Luzin-continuous on (C ; ).

We conclude the section with an existence result for an equation with respect to Poisson idempotent processes, which will be used

© 2001 by Chapman & Hall/CRC

167

Idempotent Ito equations

in a queueing application later on. We con ne ourselves to Luzin solutions. Let (us (x); s 2 R+ ; x 2 C (R + ; R)) and (vs (x); s 2 R+ ; x 2 C (R + ; R )) be C(R + ; R )-progressively measurable R + -valued functions, which are continuous in x and such that Zt

0

us (x) ds < 1;

Zt

0

us (x) ds < 1; t 2 R+ ; x 2 C (R + ; R):

Let us consider the equation

Xt = x+N1

Z t

us (X ) ds

N2

0

Zt

vs (X ) ds ;

(2.6.10)

0

where N1 and N2 are independent Poisson idempotent processes and x 2 R.

De nition 2.6.32. We say that equation (2.6.10) has a Luzin solution if there exist an idempotent probability space ( ; ) equipped with a - ow A and R-valued continuous idempotent processes X = (Xs ; s 2 R+ ), N1 = (N1 (s); s 2 R+ ) and ( ; ) such that the following holds 1. X , are

R

N1 ( 0t us(X ) ds); t 2 R+

A-adapted,

and

N2 = (N2(s); s 2 R+ ) on R

N2( 0t vs(X ) ds); t 2 R+

R

2. the R idempotent processes N1(r + 0Rt us(X ) ds) N1(R0t us(X ) ds); r 2 R+ and N2(r + 0t vs(X ) ds) N1( 0t vs(X ) ds); r 2 R+ , when conditioned on At, where t 2 R+ , are independent Poisson idempotent processes, 3. (X; N1 ; N2 ) is a Luzin-continuous idempotent process, 4. (2.6.10) holds for t 2 R+ -a.e. in ! 2 . The triplet (X; N1 ; N2 ) is then called a Luzin solution to the equation with initial condition x.

The next existence result is an analogue of Theorem 2.6.24 and is proved along the same lines.

© 2001 by Chapman & Hall/CRC

168

Maxingales

Theorem 2.6.33. Let us(x) and vs(x), in addition to the above conditions, satisfy the linear-growth condition us (x) + vs (x) ls (1 + supts jxt j), where ls is locally integrable. Then equation (2.6.10) has a Luzin solution (X; N1 ; N2 ) on an idempotent probability space ( ; ) with a - ow A such that the idempotent distribution of X has density

x) = exp

X x(

Z1

0

sup x_ t (e 1)ut (x) (e 2R

if x is absolutely continuous and

1)vt (x) dt

x0 = x, and Xx (x) = 0 otherwise.

Proof. We rst prove an analogue of the extension property for the equation

xt = x+n1

Zt

0

us(x) ds

n2

Zt

0

vs (x) ds ; t 2 R+ :

(2.6.11)

More speci cally, we prove that given n1 2 C (R + ; R) and n2 2 C (R + ; R ) such that N (n1 )N (n2 ) > 0, where N is the Poisson idempotent probability, every solution of (2.6.11) on an interval [0; T ] can be extended to a solution on R+ . Since by properties of the idempotent Poisson process for n 2 C (R + ; R), A 2 R+ and t 2 R+ N

n(t) S N en(t) e(e 1)t > A A(1+t) = A(1+t) 1+t e e

and the latter ratio is less than N (n1 ) for all A large enough, we conclude that n1 (t) A(1 + t) for all t 2 R+ if A is large. The same fact holds for n2 . Since also n1 and n2 are continuous, the claim follows by a successive approximation argument as in the proof of Theorem 2.6.22. We de ne idempotent probability X;N1 ;N2 on RC (R + ; R 3 ) by X; RN1 ;N2 (x; n1; n2 ) = N (n1 )N (n2 ) if xt = x + n1 0t us (x) ds n2 0t vs (x) ds ; t 2 R+ , and X;N1 ;N2 (x; n1 ; n2 ) = 0 otherwise. Let (X; N1 ; N2 ) be the canonical process on C (R + ; R3 ). It clearly satis es (2.6.10) X;N1 ;N2 -a.e. We check that (X; N1 ; N2 ) is Luzincontinuous, i.e., that X;N1 ;N2 is a deviability by showing that the sets K (a) = f(x; n1 ; n2 ) 2 C (R + ; R3 ) : X;N1 ;N2 (x; n1 ; n2 ) ag are

© 2001 by Chapman & Hall/CRC

169

Idempotent Ito equations

compact for all a 2 (0; 1], which is carried out as in the proof of Theorem 2.6.24. In some more detail, let (xk ; n1;k ; n2;k ) 2 K (a); k 2 N . Since N is a Luzin-continuous idempotent process by Lemma 2.4.17, we have by Theorem 2.2.13 that for T 2 R+ and > 0 lim N Æ!0

sup jn(t) n(s)j > = 0: s;t2[0;T ]: jt sjÆ

Since N (n1;k ) a > 0 and N (n2;k ) a > 0, it follows that the functions n1;k and n2;k , k 2 N , are locally uniformly equicontinuous. Besides in analogy with the above argument there exists B > 0 such that n1;k (t) + n2;k (t) B (1 + t) for t 2 R+ and k 2 N . Since also the (xk ; n1;k ; n2;k ) satisfy (2.6.11), a standard argument shows that the xk are uniformly bounded on bounded intervals and locally uniformly equicontinuous as well. Arzela-Ascoli's theorem implies that the sequence f(xk ; n1;k ; n2;k ); k 2 N g is relatively compact in C (R + ; R 3 ). Let (~ x; n~ 1; n~ 2 ) beX;Nan;Naccumulation point. It clearly satN 1 2 is es (2.6.11). Therefore, (~x; n~ 1 ; n~ 2 ) = (~n1 )N (~n2 ) a; where the latter inequality follows since (~n1 ; n~ 2 ) is an accumulation point of (n1;k ; n2;k ), N (n1 )N (n2 ) a, and the function N (n) is upper semi-continuous. Let At be the -algebra generated by the atoms pt x, pR0t us (x) ds n1 and pR t vs (x) ds n2 , where (x; n1 ; n2 ) 2 C (R + ; R3 ). De ning 0 A =R tfAt ; t 2 R+ g, we see that the X, R idempotent processes N1( 0 us(X ) ds); t 2 R+ , and N2( 0t vs(X ) ds); t 2 R+ are Aadapted. We nowR show that the idempotent processes N1(r + Rt Rt t 0 uRs (X ) ds) N1 ( 0 us (X ) ds); r 2 R + and N2 (r + 0 vs (X ) ds) N1( 0t vs(X ) ds); r 2 R+ , when conditioned on At, are independent Poisson idempotent processes. Equivalently, we have to prove that for n01 ; n02 2 C (R + ; R) the following holds X;N1 ;N2 -a.e. X;N1 ;N2 R0t us (X ) ds N1 = n01 ; R0t vs (X ) ds N2 = n02 jAt = N (n01 )N (n02 ): (2.6.12)

Let n001 ; n002 ; x00 2 C (R + ; R) be such that X;N1 ;N2 (x00 ; n001 ; n002 ) > 0. Then by the de nition of X;N1 ;N2 , the extension property and the

© 2001 by Chapman & Hall/CRC

170

Maxingales

properties of the Poisson idempotent probability X;N1 ;N2 R0t us (X ) ds N1 = n01 ; R0t vs (X ) ds N2 = n02 ; pt X = pt x00 ; pR0t us (X ) ds N1 = pR0t us (x00 ) ds n001 ; pR0t vs (X ) ds N2 = pR0t vs (x00 ) ds n002 = sup 1 R0t us(x) ds n1 = n01; R0t vs(x) dsn2 = n02; (x;n1 ;n2 )2C (R+ ;R3 ) pt x = pt x00 ; pR0t us (x) ds n1 = pR0t us (x00 ) ds n001 ; pR0t vs (x) ds n2 = pR0t vs (x00 ) ds n002 X;N1 ;N2 (x; n1 ; n2 ) = sup 1 R0t us(x00 ) ds n1 = n01; R0t vs(x00 ) ds n2 = n02; (n1 ;n2 )2C (R+ ;R2 ) pR0t us (x00 ) ds n1 = pR0t us (x00 ) ds n001 ; pR0t vs (x00 ) ds n2 = pR0t vs (x00 ) ds n002 N (n1 )N (n2 ) = sup 1 R0t us (x00 ) ds n1 = n01 ; n1 2C (R+ ;R) pR0t us (x00 ) ds n1 = pR0t us (x00 ) ds n001 sup 1 R0t vs (x00 ) ds n2 = n02 ; n2 2C (R+ ;R) N 00 R R p 0t vs (x00 ) ds n2 = p 0t vs (x00 ) ds n2 (n1 )N (n2 ) = N n1 : pR t 00 n1 = pR t 00 n00 N (n0 ) 0 us (x ) ds

0 us (x ) ds 1

1

N n2 : pR0t vs (x00 ) ds n2 = pR0t vs (x00 ) ds n002 N (n02 ) = X;N1 ;N2 pt X = pt x00 ; pR t us (X ) ds N1 = pR t us (x00 ) ds n001 ; 0 0 N 0 N 0 00 R R p t vs (X ) ds N2 = p t vs (x00 ) ds n2 (n1 ) (n2 ):

0

0

Equality (2.6.12) follows. X;N1 ;N2 (x; n ; n ) coinFinally, the fact that X 1 2 x (x) = supn1 ;n2 cides with the expression for X (x) in the statement of the theorem follows by routine calculations.

Remark 2.6.34. Rt

Rt We refer to N u ( X ) ds ; t 2 R+ and 1 s 0 N2 0 vs(X ) ds ; t 2 R+ as Poisson idempotent processes of rates us (X ) and vs (X ), respectively.

2.7 Semimaxingales In this section we consider idempotent analogues of semimartingales and associated integrals. Let ( ; ) be an idempotent probability space with a - ow A. Let (Gt (; !); t 2 R+ ; ! 2 ); 2 Rd ;

© 2001 by Chapman & Hall/CRC

Semimaxingales

171

be R-valued A-adapted continuous idempotent processes such that G0 (; !) = Gt (0; !) = 0. We refer to G() = (Gt (; !); t 2 R+ ; ! 2

); 2 Rd ; as a cumulant.

De nition 2.7.1. We say that an Rd -valued A-adapted continuous idempotent process X on ( ; ) is an A-semimaxingale with cumulant G() if the idempotent process Y () = (Yt (; !); t 2 R+ ; ! 2 ) de ned by

Yt () = exp( (Xt X0 ) Gt ())

(2.7.1)

is an A-local exponential maxingale for every 2 Rd . If, in addition, G() is an increasing function of t for all and !, X is called an A-local maxingale.

Remark 2.7.2. We occasionally say that X is a semimaxingale on ( ; A; ) rather than that it is an A-semimaxingale. Examples of semimaxingales are the idempotent Wiener process and the idempotent Poisson process. Also local maxingales with quadratic characteristics are semimaxingales. Another example is provided by idempotent processes with independent increments. Recall that AX denotes the - ow generated by an idempotent process X.

Theorem 2.7.3. Let X be a continuous A-adapted idempotent process with independent increments such that the function G~ t () = ln S exp( (Xt X0 )) is nite for all 2 Rd . Then X is an AX -

semimaxingale with cumulant G~ (). If, in addition, X0 is Luzin and G~ t () is dierentiable in , then X is Luzin-continuous. Proof. We only prove the Luzin-continuity. Since G~ t () is dierentiable in , by Lemma 1.11.7 the increments Xt Xs are Luzin idempotent variables, so by independence of increments, Corollary 1.8.10 and Theorem 1.7.11 X is Luzin. It is Luzin-continuous by Theorem 2.2.13 and the fact that G~ t () is continuous in t.

Our primary goal is to show that integrals with respect to semimaxingales also give rise to local exponential maxingales. The methods are analogous to those we used in Section 2.5. In view of applications to large deviation theory we are interested in studying semimaxingales on spaces of trajectories that are also strictly Luzin

© 2001 by Chapman & Hall/CRC

172

Maxingales

idempotent processes with respect to a deviability. Therefore, both in this section and the next one we assume that is the space C = C (R + ; R d ) and is a deviability on C . We equip C with the

ow C = (Cs ; s 2 R+ ) that is the completion of the natural - ow C = (Cs; s 2 R+ ) with respect to , where the Cs are the -algebras generated by the mappings x ! xt ; t 2 [0; s]; for x 2 C . We start with some simple properties.

De nition 2.7.4. We say that an R +-valued function on C is a strictly Luzin stopping time on (C ; C ; ) if it is a C{stopping time and is nite and continuous when restricted to K (a) for a 2 (0; 1]. The following result is standard.

Lemma 2.7.5. Let (Ht (x); t 2 R+ ; x 2 C ) be an R+ -valued increasing continuous C {adapted strictly Luzin idempotent process. Let, for c 2 R+ ,

(x) = inf ft 2 R+ : Ht (x)+ t cg:

Then (x); x 2 C ; is a strictly Luzin stopping time on (C ; C ; ).

Proof. The fact that is a C {stopping time follows by Lemma 2.2.18. Since Ht (x) is increasing, continuous in t and continuous in x on K (a); a 2 (0; 1]; it is also continuous on K (a) as a map from C to C (R + ; R) by Polya's theorem, see Liptser and Shiryaev [79, Problem 5.3.2]. Continuity of (x) on K (a) follows now by Whitt [135, Theorem 7.2] since Ht (x) + t is strictly increasing.

De nition 2.7.6. We say that a semimaxingale (respectively, local maxingale) X with cumulant G() is a strictly Luzin semimaxingale (respectively, local maxingale) on (C ; C ; ) if X and G() for every d

2 R are strictly Luzin idempotent processes. We say that X is a strictly Luzin-continuous semimaxingale (respectively, local maxingale) if X and G() for every 2 Rd are strictly Luzin-continuous idempotent processes.

Lemma 2.7.7. If X is either a strictly Luzin local maxingale or a strictly Luzin-continuous semimaxingale on (C ; C ; ), then the local exponential maxingales Y () admit localising sequences of strictly Luzin stopping times.

© 2001 by Chapman & Hall/CRC

173

Semimaxingales

Proof. Let X be a strictly Luzin local maxingale with cumulant G(). Let n = inf ft 2 R+ : Gt (2) + t ng, where n 2 N . Since G(2) is a strictly Luzin idempotent process and is increasing, n is a strictly Luzin stopping time by Lemma 2.7.5. The idempotent process (Yt^n (); t 2 R+ ) is uniformly maximable because

S Yt^n ()2

S Yt^n (2) exp(Gt^n (2)) en;

where the rst inequality holds since Gt () is non-negative and the second follows by the de nition of n and the fact that (Yt^n (); t 2 R + ) is a supermaxingale starting at 1. If X is a strictly Luzin-continuous semimaxingale, then the above argument applies with n = inf ft 2 R+ : supst jGs (2)j_jGs ()j + t ng. In connection with the lemma we introduce the following.

De nition 2.7.8. A local exponential maxingale M on (C ; C ; ) is

called a strictly Luzin-continuous local exponential maxingale if it is a strictly Luzin-continuous idempotent process and admits a localising sequence of strictly Luzin stopping times.

Remark 2.7.9. Note that if M is a strictly Luzin-continuous local

exponential maxingale and is a strictly Luzin stopping time, then Mt^ is a strictly Luzin idempotent variable. In the rest of the section X is the canonical idempotent process on C , i.e., Xt (x) = xt , and the following is assumed to hold:

X is a semimaxingale on (C ; C ; ) with cumulant G().

Lemma 2.7.10. The idempotent process X is a semimaxingale with 1 cumulant G() under the deviability (jx0 = x) for Æ 0 -almost d all x 2 R .

We omit a simple proof and only note that (jx0 = x) is a deviability by Lemma 1.6.12. We introduce an idempotent measure that is to play an important part in the sequel. Let 0 be the set of all Rd {valued, piecewise constant functions ((t); t 2 R+ ) of the form

(t) =

k X i=1

i 1(t 2 (ti 1 ; ti ]);

© 2001 by Chapman & Hall/CRC

174

Maxingales

where 0 t0 < t1 < : : : < tk ; i 2 Rd ; i = 0; : : : ; k; k de ne for x 2 C and x 2 Rd Z1 I(x) = sup (t)dxt dGt ((t); x); ((t))20 0

2 N : We (2.7.2)

where the integral is understood as a nite sum so that Z1

0

((t) dxt

=

dGt ((t); x))

k X i=1

i (xti

xti 1 )

(Gti (i ; x) Gti 1 (i ; x)) ; (2.7.3)

and let

((x) = exp( I(x)); (2.7.4) Ix(x) = I(x); if x0 = x; (2.7.5) 1;

otherwise; x(x) = exp( Ix(x)); x 2 C ; x( ) = sup x(x); C : (2.7.6) x2 We have our rst property of x .

Lemma 2.7.11. For Æ 0 1-almost all x 2 Rd , (xjx0 = x) x(x); x 2 C . In particular, (x) (x) and is an idempotent probability.

Proof. By Lemma 2.7.10 it is enough to check that (x) x (x) assuming that x0 = x -a.e. We follow the argument of the proof of Lemma 2.5.8. Let, for 0 s1 t1 : : : sk tk and i 2 Rd ; i = 1; : : : ; k, k hX

Z~ = exp

i=1

i Xti Xsi

k X i=1

i

Gti (i ) Gsi (i ) :

The de nition of x and argument of the proof of Lemma 2.5.8 imply that it suÆces to show that S Z~ 1. Let

n = inf t 2 R+ : max jGt (i )j_jGt (2i )j i=1;:::;k

© 2001 by Chapman & Hall/CRC

n

175

Semimaxingales

be a common localising sequence for the Y (i ); i = 1; : : : ; k, and

Yni = exp i Xti ^n Xsi ^n

Gti ^n (i ) Gsi ^n (i ) :

Then S (Yni jCsi ) = 1 so that

S

k Y i=1

Yni = S = S

kY1

i=1 kY2 i=1

Yni S (Ynk jCsk )

Yni S (Ynk

jC

1 ) sk 1

= : : : = S Yn1 = 1:

Q Since n ! 1, it follows that Z~ = limn!1 ki=1 Yni so that \the Fatou lemma" (see Theorem 1.4.19) yields the required. Finally, is an idempotent probability since is an idempotent probability and (x) 1 by de nition.

Below, we are mostly concerned with the case where the cumulant G() = (Gt (; x); t 2 R+ ; x 2 C ); 2 Rd , is absolutely continuous so that it has the form

Gt (; x) =

Zt

0

gs (; x) ds; 2 Rd ; t 2 R+ ; x 2 C ;

(2.7.7)

where gs (; x) is Lebesgue integrable in s. The next lemma gives the form of x for absolutely continuous cumulants. It also shows that our usage of the notation is consistent with the one in Section 2.6 (see Remark 2.6.20). Let

hs (y; x) = sup ( y gs (; x)) 2Rd

(2.7.8)

be the convex conjugate, or the Legendre{Fenchel transform, of gs (; x). It is non-negative provided gs (0; x) = 0.

Lemma 2.7.12. Let an R-valued function gs(; x) be Lebesgue measurable in s and continuous in , gs (0; x) = 0, and for A > 0, t 2 R+ and x 2 C Zt

0

sup jgs (; x)j ds < 1: jj=A

© 2001 by Chapman & Hall/CRC

176

Maxingales

If G() has the form (2.7.7), then 8 1 Z > > < I(x) = > 0 hs (x_ s; x) ds; if x is absolutely continuous, > : +1; otherwise. In particular, X is absolutely continuous under the hypotheses.

Proof. If x is absolutely continuous, then the desired representation follows by Lemma A.2 in Appendix A with f (t; ) = x_ t gt (; x); (2.7.3) and (2.7.5). Let x not be absolutely continuous on an interval [0; T ]. Then we can choose " > 0 such that for every Æ > 0 there exist 0 t1 < : : : < t2l T satisfying l X i=1

(t2i t2i 1 ) < Æ;

l X i=1

jxt2i xt2i 1 j > ":

(2.7.9)

For N > 0, we take

xt2i xt2i 1 1 jx xt2i 1 j (t2i 1 ;t2i ](t) i=1 t2i (of course we may assume that jxt2i xt2i 1 j > 0). N (t) = N

l X

Then by (2.7.3), (2.7.5), and (2.7.9)

I(x) =N

Z1

0 l X

i=1

[N (t) dxt

jxt2i xt2i 1 j > N"

ZT

0

dGt (N (t); x)] l Zt2i X i=1 t2i 1

gt (N (t); x) dt

l [

sup jgt (; x)j 1 t 2 (t2i 1 ; t2i ] dt: jj=N i=1

By (2.7.9) the latter integrand goes to 0 in measure as Æ ! 0 so that by Lebesgue's dominated convergence theorem the integral converges to 0 as Æ ! 0. Thus, I(x) > N" for arbitrary N .

© 2001 by Chapman & Hall/CRC

177

Semimaxingales

Finally, since by Lemma 2.7.11 (X = xjX0 = x) x(x) for Æ0 1 -almost all x, X is absolutely continuous under (jX0 = x) for these x. Since (X = x) = supx2Rd (X = xjX0 = x)(X0 = x), it follows that X is absolutely continuous under . We assume in the rest of the section that G() is given by (2.7.7). Let us further assume that Gt (; x) = Bt0 (x)+ G^ t (; x); (2.7.10)

where B 0 = (Bt0 (x); t 2 R+ ; x 2 C ) is an Rd -valued C -adapted idempotent process such that B00 (x) = 0 and G^ () = (G^ t (; x); t 2 R + ; x 2 C ), 2 R d , are R+ -valued C -adapted idempotent processes such that G^ 0 (; x) = G^ t (0; x) = 0. Since G() is absolutely continuous in t, we assume that both B 0 and G^ () are absolutely continuous so that

Bt0 (x) = G^ t (; x) =

Zt

0 Zt 0

bs (x) ds;

(2.7.11)

g^s (; x) ds;

(2.7.12)

where (bs (x)) is C-progressively measurable, 0t jbs (x)jds < 1, (^gs (; x)) is R+ -valued, B([0; t]) B(R d ) Ct =B(RR+ ){measurable as a map from [0; t] Rd C to R+ , g^s (0; x) = 0, and 0t g^s (; x) ds < 1 for t 2 R+ ; 2 Rd and x 2 C . (The product of a -algebra and a -algebra has been introduced in De nition 1.5.9, the product of two -algebras has a standard meaning.) Thus, gs (; x) from (2.7.7) has the form

gs (; x) = bs (x)+^gs (; x):

R

We now introduce more conditions on bs (x) and g^s (; x).

(2.7.13)

(I ) The idempotent process (bs (x)) is strictly Luzin on (C ; ) and Zt

0

sup

x2K (a)

jbs(x)j ds < 1

for all a 2 (0; 1] and t 2 R+ .

© 2001 by Chapman & Hall/CRC

178

Maxingales

(II ) The function (^gs (; x)) is continuous in (; x) when restricted to Rd K (a) for a 2 (0; 1], convex in 2 Rd , and sup sup sup g^s (; x) < 1; lim sup sup g^s (; x) = 0 !0 st x2K (a) jjA st x2K(a) for all a 2 (0; 1], t 2 R+ and A 2 R+ . Let Mt = Xt X0 Bt0 : Then M = (Mt (x); t 2 R+ ; x 2 C ) is a C { local maxingale with cumulant G^ () and the following \canonical decomposition" holds X = X0 + B 0 + M: (2.7.14) 0 Under (I ) and (II ) the idempotent processes B and M are strictly Luzin-continuous. As above, we denote by M_ a C -progressively measurable idempotent process that coincides with the RadonNikodym derivative of M with respect to Lebesgue measure -a.e. We note that for absolutely continuous x x_ s gs (; x) = M_ s (x) g^s (; x); (2.7.15) so by (2.7.8), Lemma 2.7.12, (2.7.10), (2.7.14), and Lemma 2.7.11 Z1 I(x) = h^ t (M_ t (x); x) dt < 1 -a.e.; (2.7.16) 0

where h^ t (y; x) = sup (y g^t (; x)); y 2 Rd ; t 2 R+ ; x 2 C : 2Rd

(2.7.17) De nition 2.7.13. Let ^ denote the set of all C{ progressively measurable strictly Luzin idempotent processes = ((t; x); t 2 R+ ; x 2 C ) on (C ; ) such that for 2 R, t 2 R+ and x 2 C R d {valued

Zt

0

g^s ((s; x); x) ds < 1

(2.7.18)

and, moreover, Zt

0

g^s ((s; x); x) 1(j(s; x)j > A) ds ! 0 as A ! 1:

© 2001 by Chapman & Hall/CRC

(2.7.19)

179

Semimaxingales

Remark 2.7.14.

If condition (II ) holds, then bounded C { progressively measurable strictly Luzin idempotent processes belong to ^ .

Lemma 2.7.15. Let conditions (RI ) and (II ) hold. Let = ((t; x); t 2 R+ ; x 2 C ) 2 ^ . Then 0t (s; x) M_ s ds, where t 2 R+ , is well de ned and nite -a.e. Proof. Since M is absolutely continuous, the integrand in the statement is well de ned -a.e. We show that -a.e. Zt

0

j(s; x) M_ s j ds < 1:

Since h^ s (y; x) is the convex conjugate of g^s (; x) and g^s (; x) is nonnegative, by Young's inequality (see, e.g., Krasnosel'skii and Rutickii [75])

j(s; x) M_ sj = (s; x) sign (s; x) M_ s M_ s g^s (s; x) sign ((s; x) M_ s); x + h^ s (M_ s; x) g^s((s; x); x) + g^s( (s; x); x) + h^ s(M_ s ; x):

The integral from 0 to t of the right-most side is nite -a.e. by the de nition of ^ and (2.7.16). Given an Rd -valued Lebesgue measurable in s function = ((s; x)), we introduce an idempotent process Z () = (Zt (; x); t 2 R + ; x 2 C ) by Z t

Zt (; x) = exp

0

(s; x)x_ s gs ((s; x); x) ds

(2.7.20)

if the integral on the right-hand side is well de ned and nite, and let Zt (; x) = 0 otherwise. If 2 ^ and conditions (I ) and (II ) hold, then by Lemma 2.7.15, the de nition of ^ and (2.7.15) equality (2.7.20) holds for t 2 R+ -a.e. The main result of this section is the following theorem. Theorem 2.7.16. Let conditions (I ) and (II ) hold. If 2 ^ , then the idempotent process Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ).

© 2001 by Chapman & Hall/CRC

180

Maxingales

The proof proceeds through a string of lemmas. Let us rst note that by Lemma 2.7.11, Lemma 2.7.12 and condition (II )

S Z () 1 (2.7.21) for an arbitrary R+ -valued function on C . In the lemmas below we assume that conditions (I ) and (II ) hold. Lemma 2.7.17. Let 2 ^ . Then the idempotent process Z () is strictly Luzin-continuous. If (x); x 2 C ; is a strictly Luzin stopping time, then Zt^ () is a Ct^ {measurable strictly Luzin variable. Proof. The argument is similar to the one we used in the proof of Lemma 2.5.16. We begin the proof of Z () being strictly Luzincontinuous by proving that Zt () is a strictly Luzin variable. Since by (2.7.21) and \the Chebyshev inequality" Zt () is a proper idempotent variable, by (2.7.20) it is suÆcient to check that the maps x ! Rt Rt _ 0 (s; x) Ms (x) ds and x ! 0 g^s ((s; x); x) ds are continuous when restricted to sets K (a); a 2 (0; 1]: Let xk ! x^ as k ! 1, where xk ; x^ 2 K(a). We rst check the convergence lim

k!1

Denoting

Zt

0

(s; x

k )M_

x

s(

k ) ds =

Zt

0

(s; x^ )M_ s (^x) ds:

A (t; x) = (t; x)iA (j(t; x)j); where iA (x) = (A +1 x)+ ^ 1; x 2 R+ ; we have that, for A > 0, Z t (s; k ) 0 Zt

x M_ s(xk ) ds

Zt

0

(s; x^ )

(2.7.22)

(2.7.23) (2.7.24) _ Ms (^ ) ds

x

j(s; xk ) M_ s(xk )j 1(j(s; xk )j > A) ds 0

+

Zt

0

© 2001 by Chapman & Hall/CRC

j(s; x^ ) M_ s(^x)j 1(j(s; x^ )j > A) ds

181

Semimaxingales

+

Zt

j(A (s; xk ) A(s; x^ )) M_ s(xk )j ds

0 Zt + A (s; ^ ) 0

x M_ s(xk ) ds

Zt

0

A (s; x^ )

M_ s (^ ) ds :

x

(2.7.25)

Note that all the terms in (2.7.22) and (2.7.25) are well de ned by Lemma 2.7.15 and the fact that (A (t; x); t 2 R+ ; x 2 C ) 2 ^ . We prove that each term on the right of (2.7.25) tends to 0 as k ! 1 and A ! 1. Since by Young's inequality for > 0 1 1 y g^t (; x)+ h^ t (y; x); we have for x 2 C and > 0 by (2.7.16) Zt

0

j(s; x) 1(j(s; x)j > A) M_ s(x)j ds

1 +

1

Zt

0 Zt 0 Zt

g^s (s; x) sign ((s; x) M_ s (x)); x 1(j(s; x)j > A) ds

h^ s (M_ s (x); x) ds

1 [^gs((s; x); x) + g^s( (s; x); x)] 1(j(s; x)j > A) ds 0

so that by (2.7.19) and the inequality I(x) holds since (x) (x)) lim sup sup A!1 x2K (a)

Zt

0

1 I(x); (2.7.26) ln a on K (a) (which +

j(s; x) M_ s(x)j 1(j(s; x)j > A) ds lna ; > 0:

Since is arbitrary, we have thereby proved that the rst term on the right of (2.7.25) tends to 0 as k ! 1 and A ! 1, and the second one tends to 0 as A ! 1.

© 2001 by Chapman & Hall/CRC

182

Maxingales

For the third term, we write analogously to (2.7.26) for > 0 Zt

0

j(A (s; xk ) A(s; x^ )) M_ s(xk )j ds

1

Zt

0

g^s ((A (s; xk ) A (s; x^ )); xk ) ds

1 +

Zt

0

g^s ( (A (s; xk ) A (s; x^ )); xk ) ds +

1 k I(x ): (2.7.27)

Since is strictly Luzin, (s; xk ) ! (s; x^ ); hence, by (2.7.23) and (2.7.24) A (s; xk ) A (s; x^ ) ! 0. Therefore, by condition (II ) g^s ((A (s; xk ) A (s; x^ )); xk ) ! g^s (0; x^ ) = 0 as k ! 1. Thus, by Lebesgue's dominated convergence theorem and condition (II ) (recall that by (2.7.23) and (2.7.24) jA (s; xk ) A (s; x^ )j 2(A +1)), the rst two terms on the right of (2.7.27) tend to 0 as k ! 1. Since is arbitrary, the inequality I(xk ) ln a implies that the third term on the right of (2.7.25) tends to 0 as k ! 1. Thus, we are left to prove that lim

k!1

Zt

0

x

A (s; ^ )M_

x

s(

k ) ds =

Zt

0

A (s; x^ )M_ s (^x) ds:

(2.7.28)

Let ~ be the set of bounded Rd {valued Borel functions ((t); t 2 R+ ) such that lim

k!1

Zt

0

(s) M_ s (xk ) ds =

Zt

0

(s) M_ s (^x) ds:

We prove that ~ consists, in fact, of all Rd {valued bounded Borel functions, which will imply (2.7.28) since A (s; x^ ), being Lebesguemeasurable in s, coincides a.e. with some Borel-measurable function. Since Ms (x) is continuous on K (a) by (I ) and (II ), the convergence xk ! x^ implies that ~ contains all piecewise constant functions ((t)). Now, by a standard monotone class argument, see, e.g.,

© 2001 by Chapman & Hall/CRC

183

Semimaxingales

Meyer [88], it is suÆcient to prove that ~ is closed under bounded pointwise convergence. We prove this by an argument similar to the one we used above: let n (s) ! (s) as n ! 1, where jn (s)j A and j(s)j A. Then, as in (2.7.26), we have with the use of Young's inequality for > 0 Zt

0

j(n (s) (s)) M_ s(x)j ds

1

Zt

0

[^gs ((n (s) (s)); x) + g^s ( (n (s) (s)); x)] ds

1 I(x); so that condition (II ) and Lebesgue's dominated convergence theorem yield +

lim

sup

Zt

n!1 x2K (a)

j(n (s) (s)) M_ s(x)j ds = 0:

0

Hence, lim lim sup

n!1 k!1

Zt

0

lim

n!1

j(n(s) (s)) M_ s(xk )j ds = 0;

Zt

0

j(n (s) (s)) M_ s(^x)j ds = 0;

which proves the claim. Thus, R t (2.7.28) and with it (2.7.22) have been proved. Continuity of x ! 0 (s; x) M_ s (x) ds on K (a) has been proved. In order to prove that Zt

0

g^s ((s; x

x

k ); k ) ds !

Zt

0

g^s ((s; x^ ); x^ ) ds;

(2.7.29)

we note that since (s; x) is continuous in x on K (a) by the de nition of ^ and g^s (; x) is continuous in (; x) by condition (II ),

© 2001 by Chapman & Hall/CRC

184

Maxingales

g^s ((s; xk ); xk ) ! g^s ((s; x^ ); x^ ) as k ! 1: Therefore, condition (II ) and Lebesgue's dominated convergence theorem yield for A > 0 lim

Zt

k!1

0

g^s ((s; xk ); xk )iA (j(s; xk )j) ds =

Zt

0

g^s ((s; x^ ); x^ )iA (j(s; x^ )j) ds

so that lim sup k!1

Zt

g^s ((s; xk ); xk ) 1(j(s; xk )j A) ds

0

Zt

g^s((s; x^ ); x^ ) ds: 0

The latter inequality, (2.7.19) and Fatou's lemma imply the convergence (2.7.29). To complete the proof of Z () being strictly Luzin-continuous, by Theorem 2.2.13 it is suÆcient to show that for T 2 R+ and > 0

lim sup jZt () Zs()j > = 0: Æ!0 s;t2[0;T ]: js tjÆ Since for A > 0

jZt () Zs ()j > (Zs() > A) _ jZt ()=Zs() 1j > =A and (Zs () > A) 1=A by (2.7.21), we deduce that the required would follow by lim Æ!0

Zt sup s;t2[0;T ]:

js tjÆ s

Z t sup g^u ((u; s;t2[0;T ]:

lim Æ!0

js tjÆ s

© 2001 by Chapman & Hall/CRC

(u; x) M_ u (x) du > = 0;

x); x) du >

= 0:

185

Semimaxingales

The rst convergence follows by the inequality Zt

s

j(u; x) M_ u(x)j du

1

ZT

0

[^gu ((u; x); x) + g^u ( (u; x); x)] 1(j(u; x)j > A) du 1 +

Zt

s

1 sup g^u (; x) du + I(x); A > 0; > 0; jjA

derived in analogy with (2.7.26) and condition (II ). The second convergence follows by condition (II ) and the inequality Zt

s

g^u ((u; x); x) du

ZT

0

g^u ((u; x); x) 1(j(u; x)j > A) du +

Zt

sup g^u (; x) du: j jA s

Finally, since and (bs (x)) are C -progressively measurable, g^s (; x) is B([0; t]) B(Rd ) Ct =B(R+ ){measurable as a map from [0; t] _ is C -progressively measurable, and C is comR d C to R + , M plete, Ct^ {measurability of Zt^ () follows by Lemma 2.2.17 and Lemma 2.2.19. The fact that Zt^ () is strictly Luzin measurable follows by the rst part of the lemma. We now address the uniform maximability issue. The next lemma is in the theme of Lemmas 2.7.5 and 2.7.7. Lemma 2.7.18. Let 2 ^ and (Gt ; t 2 R+ ), G0 = 0, be an increasing continuous C {adapted strictly Luzin idempotent process such that for -almost all x Zt

0

g^s (2(s; x); x) ds Gt (x); t 2 R+ :

© 2001 by Chapman & Hall/CRC

186

Maxingales

Let for N

2N

N (x) = inf t 2 R+ : Gt (x)+ t N :

Then N (x) is a strictly Luzin stopping time, the idempotent process fZt^N (); t 2 R+ g is uniformly maximable, and, moreover, S Zt^N ()2 eN : Lemma 2.7.19. Let a function = ((t; x); t 2 R+ ; x 2 C ) 2 ^ be of the form:

(t; x) =

k X i=1

i (x)1(ti

1 ;ti ] (t);

where k 2 N ; 0 t0 < t1 < : : : < tk , and the i (x) are Rd {valued, bounded and Cti 1 {measurable strictly Luzin variables on (C ; C ; ). Then Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ).

Proof. Since by Lemma 2.7.17 Z () is a C {adapted strictly Luzincontinuous idempotent process, we have to check that there exists an increasing to in nity sequence of strictly Luzin stopping times N (x); N 2 N , such that the (Zt^ N (x) ((x); x); t 2 R+ ) are uniformly maximable exponential maxingales. Let A be a bound for i (x), i.e., ji (x)j A; i = 1; : : : ; k; x 2 C . We introduce for t 2 R+ , x 2 C and N 2 N

N (x) =

8 > >

> :

N;

Zt

0

sup g^s (; x) ds + t N ; if (x) > 0; jj2A if (x) = 0:

By condition (II ) and Lemma 2.7.18 the N (x) are strictly Luzin stopping times and the idempotent processes (Zt^N (); t 2 R+ ) are uniformly maximable. Let us check that the (Zt^N (); t 2 R+ ); N 2 N ; are exponential maxingales. Let 0 s < t. We have to prove that

S (Zt^N ()jCs ) = Zs^N ():

(2.7.31)

(As above relations involving conditional idempotent expectations are understood to hold -a.e.) We begin with a proof of

S (Zt^N ()jCti ^t ) = Zti ^t^N (); i = 1; : : : ; k:

© 2001 by Chapman & Hall/CRC

(2.7.32)

187

Semimaxingales

We note that by (2.7.1) and (2.7.20) -a.e.

Zt ((x); x) =

Yti ^t (i (x); x) : Y ( ( x ) ; x ) t ^ t i i 1 i=1

k Y

(2.7.33)

Since the idempotent processes Y () = (Yt (; x); t 2 R+ ; x 2 C ), 2 Rd , are local exponential maxingales, Lemma 2.7.18 and Lemma 2.3.13 imply that the (Yt^N (); t 2 R+ ); jj A; are uniformly maximable exponential maxingales. We prove (2.7.32) by inverse induction in i: by (2.7.33) Ztk ^t^N () = Zt^N () so that (2.7.32) holds for i = k. Suppose that (2.7.32) holds for some i 2 f2; : : : ; kg: We prove it for (i 1). By properties of conditional idempotent expectations

S (Zt^N ()jCti 1 ^t ) = S [S (Zt^N ()jCti ^t )jCti 1 ^t ] = S (Zti ^t^N ()jCti 1 ^t ): By (2.7.33) and properties of conditional idempotent expectations

S (Zti ^t^N ()jCti 1 ^t ) =

Ytj ^t^N (x) (j (x); x) Y ( (x); x) j =1 tj 1 ^t^N (x) j i 1 Y

S(Yti ^t^N (x) (i (x); x)jCti 1 ^t ) : (2.7.34) Yti 1 ^t^N (x) (i (x); x) Let t ti 1 . Since i (x) is Cti 1 {measurable, by properties of conditional idempotent expectations

S (Yti ^t^N (x) (i (x); x)jCti 1 ^t ) = S (Yti ^t^N (x) (i (x); x)jCti 1 ) = S (Yti ^t^N (x) (; x)jCti 1 )j=i (x) = Yti 1 ^t^N (x) (i (x); x); where in the latter equality we used that (Yt^N (x) (; x); t 2 R+ ) is an exponential maxingale. So, if t ti 1 ,

S (Yti ^t^N (x) (i (x); x)jCti 1 ^t ) = Yti 1 ^t^N (x) (i (x); x):

© 2001 by Chapman & Hall/CRC

188

Maxingales

This also is true if t ti 1 . Substituting the equality into (2.7.34) obtains

S (Zti ^t^N ()jCti 1 ^t ) =

Ytj ^t^N (x) (j (x); x) Y ( (x); x) j =1 tj 1 ^t^N (x) j i 1 Y

= Zti 1 ^t^N ():

Equality (2.7.32) is proved. Now, (2.7.31) is obvious if s tk since in this case Zs^N () = Zt^N () = Ztk ^N (). Let s < tk and i0 2 f1; : : : ; kg be such that ti0 1 s < ti0 . By properties of conditional idempotent expectations, (2.7.32) and (2.7.33)

S (Zt^N ()jCs ) = S [S (Zt^N ()jCti0 ^t )jCs ] = S (Zti0 ^t^N ()jCs ) S (Yti0 ^t^N (x) (i0 (x); x)jCs ) = Zti0 1 ^s^N () : (2.7.35) Yti0 1 ^s^N (x) (i0 (x); x)

Now, as above, since i0 (x) is Cti0 1 {measurable,

S (Yti0 ^t^N (x) (i0 (x); x)jCs ) = S (Yti0 ^t^N (x) (; x)jCs )j=i0 (x) = Yti0 ^s^N (x) (i0 (x); x): Substituting this into (2.7.35), we deduce by (2.7.33) that (2.7.31) holds. Lemma 2.7.19 proves the assertion of Theorem 2.7.16 for piecewise constant and bounded functions . To handle general 2 ^ , we will use the following approximation result, which extends Lemma 2.5.10. Lemma 2.7.20. Let k = (k (t; x); t 2 R+ ; x 2 C ); kk 2 N ; be uniformly bounded functions from ^ such that the Z ( ) are strictly Luzin-continuous local exponential maxingales. If = ((t; x); t 2 ^ is bounded and is a limit of the k in the sense that R+ ; x 2 C ) 2 Zt

0

g^s ((k (s; x) (s; x)); x)ds ! 0 as k ! 1; 2 R; t 2 R+ ;

© 2001 by Chapman & Hall/CRC

189

Semimaxingales

then Z () is a strictly Luzin-continuous local exponential maxingale. Proof. The proof uses the ideas of the proof of Lemma 2.5.10. Let for N 2 N and k 2 N

N (x) = inf t 2 R+ :

x) = inf t 2 R+ :

Nk (

Zt

0

Zt

0

g^s (2(s; x); x)ds + t

N ;

(2.7.36)

g^s (2k (s; x); x) ds + t

N + 1 ^ N (x):

(2.7.37)

By Lemma 2.7.18 and condition (II ) the N and Nk are strictly Luzin stopping times, and (Zt^N (); t 2 R+ ) and (Zt^Nk (k ); t 2 R + ); k 2 N ; are uniformly maximable idempotent processes. In particular, by Lemma 2.3.13 the (Zt^Nk (k ); t 2 R+ ); k 2 N ; are uniformly maximable exponential maxingales. Suppose we have proved that for every R+ -valued bounded and continuous function f on C and t 2 R+ lim S Zt^Nk (x) (k (x); x)f (x) = SZt^N (x) ((x); x)f (x): (2.7.38) k Since the (Zt^Nk ( ); t 2 R+ ); k 2 N ; are exponential maxingales, this would imply that (Zt^N (); t 2 R+ ) is an exponential maxingale as well; hence, since (Zt^N (); t 2 R+ ) is uniformly maximable and N (x) ! 1 as N ! 1, this would prove in view of Lemma 2.7.17 that Z () is a strictly Luzin-continuous local exponential maxingale. Therefore, we prove next (2.7.38). By Lemma 2.7.18 and (2.7.37) S Zt^Nk (x) (k (x); x)2 exp(N + 1), so the family fZt^Nk (k ); k 2 N g is uniformly maximable, and by \the Lebesgue dominated convergence theorem" (see Theorem 1.4.19) (2.7.38) would follow by k!1

Zt^Nk (k ) ! Zt^N () as k ! 1:

© 2001 by Chapman & Hall/CRC

(2.7.39)

190

Maxingales

As a rst step, we prove that, for every 2 R and t 2 R+ , Zt

0

jg^s (k (s; x); x) g^s ((s; x); x)jds ! 0 as k ! 1:

(2.7.40) By convexity of g^s (; x) in , for " 2 (0; 1=2], g^s ((s; x); x) (1 2")^gs (k (s; x); x) + "g^s (2k (s; x); x) +"g^s ((s; x) k (s; x)); x ; " hence, since g^s (; x) 0,

g^s ((s; x); x) g^s (k (s; x); x) " sup g^s (; x) jj2Ajj

+ "g^s ((s; x) k (s; x)); x ; " where A is an upper bound for j(s; x)j; jk (s; x)j; k 2 N ; s 2 R+ , x 2 C . Interchanging k (s; x) and (s; x) and integrating, we arrive at the inequality for x such that (x) > 0 Zt

0

jg^s (k (s; x); x) g^s ((s; x); x)jds

"

Zt

0

Zt

sup g^s (; x) ds + " g^s ((s; x) k (s; x)); x ds " jj2Ajj 0

+"

Zt

g^s

0

k ( (s; x) (s; x)); x ds; "

where the right-hand side is nite by condition (II ). By hypotheses, the latter two integrals on the right tend in deviability to 0 as k ! 1, so, for > 0, Zt

lim sup (x : jg^s (k (s; x); x) g^s ((s; x); x)jds > ) k!1 0

(x : "

© 2001 by Chapman & Hall/CRC

Zt

0

sup g^s (; x)ds > =3); jj2Ajj

191

Semimaxingales

where the right-hand side is not greater than arbitrary a > 0 if " > 0 is chosen small enough to satisfy Zt

" sup

x2K (a) 0

sup g^s (; x) ds 3 jj2Ajj

(use condition (II )). Limit (2.7.40) is proved. It implies, since by (2.7.36) and (2.7.37)

fx : N (x) ^ t 6= Nk (x) ^ tg fx :

Zt

0

jg^s(2k (s; x); x) g^s(2(s; x); x)j ds > 1g;

that limk!1 (N (x) ^ t 6= Nk (x) ^ t) = 0: Hence, by the inequality (jZt^Nk (x) (k (x); x) Zt^N (x) ((x); x)j > ) (N (x) ^ t 6= Nk (x) ^ t) + (jZt^N (x) (k (x); x) Zt^N (x) ((x); x)j > ); (2.7.20), and the fact that Zt^N (x) ((x); x) is a proper idempotent variable, limit (2.7.39) would follow by t^Z N (x)

k (s;

0

x) M_ s(x)ds !

t^Z N (x)

(s; x) M_ s (x) ds

0

as k ! 1; (2.7.41)

and t^Z N (x)

g^s

0

(k (s;

x); x)ds !

t^Z N (x) 0

g^s ((s; x); x) ds as k ! 1:

The latter limit obviously follows by (2.7.40). Limit (2.7.41) is proved with a tool we have already used: for > 0, by Young's inequality,

© 2001 by Chapman & Hall/CRC

192

Maxingales

in view of (2.7.17) and (2.7.16), t^ (x) ZN k (s; x) M_ s (x) ds 0

1

Zt

0

t^Z N (x)

(s; x) M_ s (x) ds

0

g^s ((k (s; x) (s; x)); x) + g^s (((s; x)

k (s; x)); x) ds +

which implies (2.7.41) in view of the hypotheses.

1 I(x);

The next lemma and its proof are prompted by Theorem 2.5.11. Lemma 2.7.21. Let 2 ^ and be bounded. Then Z () is a strictly Luzin-continuous local exponential maxingale. Proof. Let us rst assume that is, in addition, locally uniformly continuous in t uniformly in x 2 K (a) for all a 2 (0; 1], i.e., for all T >0 wT;a (Æ) = sup sup j(t; x) (s; x)j ! 0 as Æ ! 0: x2K (a) s;t2[0;T ]: js tjÆ (2.7.42) Let for k 2 N

k (t;

x) =

k2 X i

i=1 (k (t;

k

1

;x

1( i k1 ; ki ](t):

Then the k = x); t 2 R+ ; x 2 C ) are bounded and piecewise constant functions from ^ , which implies by Lemma 2.7.19 that the Z (k ); k 2 N ; are strictly Luzin-continuous local exponential maxingales. Also, since jk (s; x) (s; x)j wt;a (1=k) for s 2 [0; t] and x 2 K (a) if k t, Zt

sup g^s ((k (s; x) (s; x)); x) ds x2K (a) 0

sup

Zt

sup

x2K(a) 0 jjwt;a(1=k)

© 2001 by Chapman & Hall/CRC

g^s (; x) ds:

193

Semimaxingales

Since the right-hand side converges to 0 as k ! 1 by (2.7.42) and condition (II ), we conclude that the k and meet the conditions of Lemma 2.7.20; hence, Z () is a strictly-Luzin local exponential maxingale. Let be an arbitrary bounded function from ^ . We introduce the Steklov functions

k (t;

x) = k

Zt

(s; x)ds; t 2 R+ ; x 2 C ; k

2 N;

(2.7.43)

t 1=k

where (s; x) = 0 if s 0. Then the functions k = (k (t; x); t 2 ^ are bounded and jk (t; x) k (s; x)j R+ ; x 2 C ) 2 2k supv;x j(v; x)jjt sj: Hence, by the part just proved the Z (k ) are strictly Luzin-continuous local exponential maxingales. By Lemma 2.7.20 Z () is a strictly Luzin-continuous local exponential maxingale provided Zt

0

g^s ((k (s; x) (s; x)); x)ds ! 0 as k ! 1;

2 R; t

2 R+ : (2.7.44)

By convexity of g^s (; x) in and (2.7.43) Zt

0

=

g^s ((k (s; x) (s; x)); x)ds Zt

g^s k

0

Zt

Z1=k

0

0

k

Z1=k

0

((s v; x) (s; x))dv; x ds

g^s (((s v; x) (s; x)); x) dv ds

sup

0v1=k

© 2001 by Chapman & Hall/CRC

Zt

0

g^s (((s v; x) (s; x)); x) ds;

194

Maxingales

so we prove (2.7.44) by proving that Zt

0

g^s (((s v; x) (s; x)); x)ds ! 0 as v ! 0:

(2.7.45)

Let a 2 (0; 1]. Firstly, we prove that for every x^ 2 C lim

v!0

Zt

0

sup g^s (((s v; x^ ) (s; x^ )); x)ds = 0:

x2K(a)

(2.7.46)

The argument is standard. If ((s; x^ ); s 2 R+ ) is continuous in s, then (s v; x^ ) (s; x^ ) ! 0 as v ! 0 and the required follows by condition (II ) and boundedness of ((s; x^ ); s 2 R+ ). If ((s; x^ ); s 2 R+ ) is an arbitrary bounded Lebesgue-measurable function, then, given arbitrary " > 0, we can choose by Luzin's R theorem a bounded continuous function (^ (s); s 2 R+ ) such that 0t 1(^ (s) 6= (s; x^ )) ds < "; and (2.7.46) then holds for ((s; x^ ); s 2 R+ ) since it holds for (^ (s); s 2 R+ ) and ((s; x^ ); s 2 R+ ) is bounded. Limit (2.7.46) is proved. We denote

g t;a () = sup sup g^s (; x): st

x2K(a)

By continuity of (s; y) in y on K (a), boundedness of (s; y) and condition (II ) we have that Zt

0

g t;a (3((s; x) (s; y))) ds +

Zt

0

g t;a ( 3((s; x) (s; y))) ds ! 0

as y ! x; y 2 K (a): Hence, for x 2 K (a) and " > 0, there exists

© 2001 by Chapman & Hall/CRC

195

Semimaxingales

an open subset U" (x) of K (a) such that

U" (x) fy 2 K (a) : +

Zt

0 Zt 0

g t;a (3((s; x) (s; y))) ds

g t;a ( 3((s; x) (s; y))) ds < "g:

By compactness of K (a), there exist x1 ; : : : ; xk 2 K (a) such that K (a) [ki=1 U" (xi ), which means that for every x 2 K (a) there exists i 2 f1; : : : ; kg such that Zt

0

g t;a (3((s; xi ) (s; x))) ds +

Zt

0

gt;a ( 3((s; xi ) (s; x))) ds < ": (2.7.47)

Next, by convexity of g^s (; x) in , for x 2 K (a), Zt

0

g^s (((s v; x) (s; x)); x) ds Zt

13 gt;a (3((s v; x) (s v; xi ))) ds

+

Zt

0

0

g t;a (3((s; xi ) (s; x))) ds +

Zt

0

g^s (3((s v; xi ) (s; xi )); x) ds ; (2.7.48)

where i is chosen so that (2.7.47) holds. By (2.7.46), if v is small enough, then for all i = 1; : : : ; k Zt

0

sup g^s (3((s v; xi ) (s; xi )); x) ds < ";

x2K (a)

© 2001 by Chapman & Hall/CRC

196

Maxingales

so, by (2.7.47) and (2.7.48) sup

Zt

x2K (a) 0

g^s (((s v; x) (s; x)); x) ds < ":

Limit (2.7.45) has been proved. Limit (2.7.44) has been proved. Proof of Theorem 2.7.16. Let 2 ^ and A = (rA (t; x); t 2 R + ; x 2 C ), where for 2 R d and A > 0 8 jj A; < ; rA = : jj A; jj > A: Since the map ! rA is continuous, it follows that A 2 ^ . Also A is bounded so that by Lemma 2.7.21 Z (A ) is a strictly Luzincontinuous local exponential maxingale. Since g^t (; x) is non-negative, convex in and g^t (0; x) = 0, we have that g^t (rA ; x) is increasing in A, so g^t (rA ; x) " g^t (; x) as A ! 1: (2.7.49) Let N be de ned by (2.7.36). By (2.7.49) and Lemma 2.7.18 the collections of strictly Luzin variables fZt^N (x) ((x); x); t 2 R+ g and fZt^N (x) (A(x); x); t 2 R+ ; A 2 R+ g are uniformly maximable. In particular, (Zt^N (x) (A (x); x); t 2 R+ ; x 2 C ) is a uniformly maximable exponential maxingale, so by Lemma 2.7.17 the theorem is proved if lim S Zt^N (x) (A (x); x)f (x) = S Zt^N (x) ((x); x)f (x) A!1 for every bounded, continuous and non-negative f , which by The orem 1.4.19 is implied by the convergence Zt^N (x) (A (x); x) ! Zt^N (x) ((x); x). Using the fact that Zt^N (x) ((x); x) is a proper idempotent variable, we prove the latter convergence by proving that as A ! 1 Zt

0

g^s ((s; x); x) g^s (rA (s; x); x) ds ! 0;

Zt

0

j((s; x) rA(s; x)) M_ s(x)jds ! 0:

© 2001 by Chapman & Hall/CRC

(2.7.50a) (2.7.50b)

197

Semimaxingales

The rst convergence follows by (2.7.49), the fact that the integral in (2.7.50a) is a strictly Luzin variable (use condition (II ) and (2.7.19)) and Dini's theorem. To prove (2.7.50b) we write as in the proof of Lemma 2.7.17 for > 0 Zt

0

j((s; x) rA(s; x)) M_ s(x)jds

1

Zt

0

g^s (((s; x) rA (s; x)); x)

1 + g^s ( ((s; x) rA (s; x)); x) ds + I (x): (2.7.51) Next, by the de nition of rA , convexity of g^s (; x) in and the fact that g^s (0; x) = 0 g^s (((s; x) rA (s; x)); x) g^s ((s; x); x) 1(j(s; x)j > A); g^s ( ((s; x) rA (s; x)); x) g^s ( (s; x); x) 1(j(s; x)j > A): Hence, as 2 ^ , (2.7.19) implies that Zt

0

g^s (((s; x) rA (s; x)); x) + g^s ( ((s; x) rA (s; x)); x) ds

! 0 as A ! 1;

so, since is arbitrary, (2.7.51) yields (2.7.50b). As a byproduct, we can prove that certain integrals with respect to X are semimaxingales. De nition 2.7.22. Let be the subset of ^ consisting of functions such that for t 2 R+ Zt

and

0 Zt

0

j(s; x) bs (x)j ds < 1; x 2 C ; j(s; x)bs (x)j 1(j(s; x)j > A) ds ! 0

© 2001 by Chapman & Hall/CRC

as A ! 1.

198

Maxingales

For 2 , we de ne the idempotent process X = ( Xt ; t 2 R + ) by

Xt =

Zt

0

Zt

(s; x)bs (x) ds+ (s; x)M_ s (x) ds;

(2.7.52)

0

if the integrals are well de ned and nite, and Xt = X^ t otherwise, where X^ is a continuous idempotent process. By Lemma 2.7.15 and the de nition of we have that (2.7.52) holds -a.e.

Theorem 2.7.23. Let gs(; x) be given by (2.7.13), conditions (I ) and (II ) hold, and 2 . Then the idempotent process X is a strictly Luzin-continuous semimaxingale on (C ; C ; ) with cumulant G () = (Gt (; x); t 2 R+ ; x 2 C ) given by Gt (; x) =

Zt

0

gs ((s; x); x) ds; t 2 R+ ; 2 R; x 2 C :

Proof. By Theorem 2.7.16 (exp( Xt Gt ()); t 2 R+ ) is a strictly Luzin-continuous local exponential maxingale. The fact that both X and G () are strictly Luzin-continuous C -adapted idempotent processes follows by the proof of Lemma 2.7.17, condition (I ), C progressive measurability of (bs (x)), and the de nition of .

In large deviation limit theorems G() is often more speci c than in (2.7.10) and de ned in terms of \characteristics", which we now introduce. Let cs (x); s 2 R+ ; x 2 C be a C -progressively measurable idempotent process with values in theR space of symmetric, positive semi-de nite d d-matrices such that 0t kcs (x)k ds < 1 for t 2 R+ and x 2 C ; s( ; x); s 2 R+ ; 2 B(Rd ); x 2 C be a transition kernel (for each x) from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that

© 2001 by Chapman & Hall/CRC

199

Semimaxingales

for t 2 R+ ; x 2 C and 2 R+

t (f0g; x) = 0; Zt Z

0

Z

Zt Z

0

Rd

Rd

Z

Rd

jxj2 ^ 1 t (dx; x) < 1;

jxj2 ^ 1 s(dx; x) ds < 1;

ejxj 1(jxj > 1) t (dx; x) < 1;

ejxj 1(jxj > 1) s (dx; x) ds < 1;

Rd

(2.7.53a)

(2.7.53b)

and the functions ( Rd f (x) s(dx; x); s 2 R+ ) are C -progressively measurable for Borel functions f such that the integrals are well de ned; ^s( ; x); s 2 R+ ; 2 B(Rd ); x 2 C be a transition kernel (for each x) from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for s 2 R + ; x 2 C and 2 B (R d ) R

^s ( ; x) s ( ; x); ^s(Rd ; x) 1;

(2.7.54)

and the functions ( Rd f (x) ^s(dx; x); s 2 R+ ) are C -progressively measurable for Borel functions f such that the integrals are well de ned. We say that the semimaxingale X has local characteristics (b; c; ; ^), where (bs (x)) is as above, if the associated cumulant is given by (2.7.7), where R

1 gs (; x) = bs (x) + cs (x) + 2

+ ln 1 +

Z

(ex

Rd

1)^s (dx; x)

Z

Rd

(ex 1 x)s (dx; x) Z Rd

(ex

1)^s (dx; x) : (2.7.55)

Remark 2.7.24. The right-hand sidedis well de ned since RRd (ex 1)^s (dx; x) > R 1 by the fact that ^s(R ; x) 1. It is also not diÆcult to check that 0t supjjAjgs (; x)jds < 1 for t 2 R+ and A 2 R+ . © 2001 by Chapman & Hall/CRC

200

Maxingales

Let

Ct (x) =

Zt

0

cs (x) ds:

(2.7.56)

We call the idempotent process B 0 = (Bt0 (x); t 2 R+ ; x 2 C ) (de ned by (2.7.11)) the rst characteristic of X \without truncation", C = (Ct (x); t 2 R+ ; x 2 C ) the second characteristic, s(dx; x) the density of the measure of jumps, and ^s(dx; x) the density of the discontinuous measure of jumps. The quadruplet (B 0 ; C; ; ^) is referred to as the characteristics of the semimaxingale X \without truncation". In large deviation limit theorems we will also need characteristics \associated with limiters".

De nition 2.7.25. A Borel function h :

Rd

! Rd

is said to be a limiter if it is bounded and h(x) = x in a neighbourhood of the origin.

For Borel functions f , for which the integrals below are well de ned, we introduce the notation

f (x) t (x) =

Zt Z

0

Rd

f (x) s(dx; x)ds; f (x) ^s (x) =

Z Rd

f (x)^s (dx; x):

The rst characteristic of X associated with a limiter h is an idempotent process B = (Bt (x); t 2 R+ ; x 2 C ) de ned by

Bt (x) = Bt0 (x)+(h(x) x) t (x):

(2.7.57)

The modi ed second characteristic associated with h is an idempotent process C~ = (C~t (x); t 2 R+ ; x 2 C ) such that the C~t (x) are symmetric positive semi-de nite d d-matrices, speci ed by the equalities

C~t (x) = Ct (x) + ( h(x))2 t (x) Zt

0

© 2001 by Chapman & Hall/CRC

( h(x) ^s (x))2 ds; 2 Rd : (2.7.58)

201

Semimaxingales

Analogously, the modi ed second characteristic \without truncation" C~ 0 = (C~t0 (x); t 2 R+ ; x 2 C ) is speci ed by the equalities

C~t0 (x) = Ct (x)+(x)2 t (x)

Zt

0

(x^s (x))2 ds:

(2.7.59) Note that (2.7.57) implies that if B = (Bt (x); t 2 R+ ; x 2 C ) is the rst characteristic associated with a limiter h(x), then

Bt (x) = Bt (x)+(h (x) h(x)) t (x): (2.7.60) Obviously, Bt (x); Ct (x); C~t (x); C~t0 (x); and f (x) t (x) (when well de ned) are continuous in t and Ct {measurable in x. We note that the cumulant assumes the form 1 Gt (; x) = Bt (x) + Ct (x) + (ex 1 h(x)) t (x) 2 +

Zt

ln 1 + (ex

0

Remark 2.7.26.

1) ^s (x) (ex

1) ^s(x) ds:

(2.7.61)

The de nition of the characteristics of a semimaxingale is motivated by the de nition of the characteristics of a semimartingale and the fact that semimaxingales are \large deviation limits" of semimartingales as is shown in part II. Note that the expression (2.7.61) for the cumulant is analogous to the logarithm of the stochastic exponential of a semimartingale, Liptser and Shiryaev [79], Jacod and Shiryaev [67]. Thus, Bt is \the drift term", Ct is \the diusion term", s ds is \the predictable measure of jumps", and ^s ds is \the discontinuous part of the predictable measure of jumps". As we will see in part II, this analogy is not only in form, but it also helps us to formulate conditions under which large deviation limit theorems for semimartingales can be proved. The terminology \truncated" and \nontruncated" characteristics is also inherited from semimartingales. The analogy with semimartingales would be more complete if we required in De nition 2.7.1 that Gt (; !) be, in addition, of locally bounded variation in t. In fact, this property holds if a semimaxingale admits characteristics as is the case in all our examples of semimaxingales. However, since a number of

© 2001 by Chapman & Hall/CRC

202

Maxingales

properties of semimaxingales do not depend on G() being of locally bounded variation, we have decided not to include this requirement in the de nition. We now state conditions on the characteristics that imply conditions (I ) and (II ). Lemma 2.7.27. Let the canonical idempotent process X be a semimaxingale on (C ; C ; ) with local characteristics (Rb; c; ; ^), where b and c are strictly Luzin R idempotent processes, and Rd (exp( x) 1 x) s (dx; x) and Rd exp( x) ^s (dx; x) are continuous in (; x) when restricted to Rd K (a) for a 2 (0; 1] and s 2 R+ . If for all a 2 (0; 1], A 2 R+ and t 2 R+ Zt

sup jbs (x)j ds < 1; sup sup kcs (x)k < 1; st x2K (a)

0

x2K (a)

sup sup sup st x2K (a) jjA

Z

1 x)s (dx; x) < 1;

(ex

Rd

lim sup sup

Z

!0 st x2K (a)

1 x)s (dx; x) = 0;

(ex

Rd

sup sup sup st x2K (a) jjA

Z lim sup sup !0 st x2K (a) Rd

Z

Rd

ex ^s (dx; x) ds < 1;

ex 1 ^s (dx; x) = 0;

then the associated cumulant satis es conditions (I ) and (II ).

2.8 Maxingale problems In this section we are concerned with identifying deviabilities for which the canonical idempotent process on C (R + ; R d ) is a semimaxingale with a given cumulant. As in the preceding section we denote C = C (R + ; Rd ), Ct = Ct (R+ ; Rd ), and C = (Ct ; t 2 R+ ). We also assume as given a cumulant G() on C . Let x 2 Rd . De nition 2.8.1. We say that a deviability on C is a solution to the maxingale problem (x; G) if the canonical process X is a semimaxingale with cumulant G() on (C ; C; ) such that X0 = x -a.e.

© 2001 by Chapman & Hall/CRC

203

Maxingale problems

Examples are provided by the Wiener and Poisson idempotent processes: by Theorem 2.4.2 the Wiener idempotent probability on C (R + ; R) solves the maxingale problem (0; G) with Gt (; x) = 2 t=2 and by Theorem 2.4.16 the Poisson idempotent probability on C (R + ; R) solves the maxingale problem (0; G) with Gt (; x) = (exp 1)t. In view of applications to large deviation theory, we are interested in nding conditions when a maxingale problem has a unique solution. Existence issues are of less importance to us. Besides, large deviation convergence theorems of part II also imply that under their hypotheses the associated maxingale problems have solutions.

De nition 2.8.2. We say that uniqueness holds for the maxingale problem (x; G) if it has at most one solution.

Our candidate for a solution of (x; G) is the idempotent measure

x de ned in (2.7.6). We begin with the case where the cumulant does not depend on x. Lemma 2.8.3. Let Gt (; x) not depend on x 2 C , be dierentiable in for all t 2 R+ and the dierences Gt () Gs () be convex in for all 0 s < t. Then the idempotent measures t1 ;:::;tk on (Rd )k , where 0 t1 < t2 < : : : < tk , speci ed by the densities

t1 ;:::;tk (x1 ; : : : ; xk ) =

k Y

inf e (xi

d i=1 2R

xi 1 ) eGti () Gti 1 () ;

where t0 = 0 and x0 = x, form a projective system of deviabilities, which has x as the projective limit. Proof. By dierentiability of Gt (), Lemma 1.11.4 and Lemma 1.11.7 the t1 ;:::;tk are deviabilities. In order to prove that they form a projective system it suÆces to check that, given ti 1 < ti < ti+1 , xi 1 and xi+1 , we have

inf sup 1 (xi xi 1 )+2 (xi+1 xi) (Gti (1 ) Gti 1 (1 ))

xi 2Rd 1 2Rd ; 2 2Rd (Gti+1 (2 )

Gti (2 )) = sup (xi+1 2Rd

© 2001 by Chapman & Hall/CRC

xi 1 ) (Gti+1 () Gti 1 ()) :

204

Maxingales

The latter equality follows by a minimax argument, see, e.g., Aubin and Ekeland [6]. By Theorem 2.2.4 the projective limit of the t1 ;:::;tk coincides with x .

Remark 2.8.4. hypotheses.

As a consequence,

x is a deviability under the

The following existence and uniqueness result, which is a converse to Theorem 2.7.3, shows that the distributions of certain idempotent processes with independent increments are uniquely speci ed by the associated cumulants. We denote Gt () = sup0st jGs ()j.

Theorem 2.8.5. Let Gt (; x) not depend on x 2 C . Let, in addition, Gt () be dierentiable in for all t 2 R+ and the dierences Gt () Gs () be convex in for all 0 s < t. Then x is a unique solution to problem (x; G). The canonical idempotent process X has independent increments and starts at x under x . Proof. We de ne t1 ;:::;tk as in Lemma 2.8.3. The construction of the t1 ;:::;tk and the fact that they induce deviability x on C imply that the canonical idempotent process X has independent increments and starts at x under x , and Sx exp((Xt Xs)) = exp(Gt () Gs ())

so that X is a semimaxingale with cumulant Gt (). We prove uniqueness. Let be a solution of (x; G) and X be the canonical idempotent process on C . We note that for every n 2 N and 2 Rd the family fYt^n (); t 2 R+ g is {uniformly maximable. (Recall that Y () is de ned by (2.7.1).) To see this, we write by (2.7.1)

S Yt^n ()2 = sup Yt^n (2; x) exp( 2Gt^n ()) exp(Gt^n (2))(x)

x2C

SYt^n(2) exp(2Gn () + Gn (2)):

Since S Yt^n (2) SY0 (2) = 1 by Lemma 2.3.13, the uniform maximability is proved. Then by \the Doob stopping theorem" (Theorem 2.3.8) the sequence fn; n 2 N g is a localising sequence for Y () and, in particular, Y () is a C{exponential maxingale under . Therefore, for 0 s < t,

S exp((Xt Xs ))jCs = exp(Gt () Gs ());

© 2001 by Chapman & Hall/CRC

(2.8.1)

205

Maxingale problems

so by Lemma 1.11.9 X has independent increments under . Let X t0 ;t1 ;:::;tk , 0 = t0 < t1 < t1 < : : : < tk , denote nite-dimensional idempotent distributions of X so that X t0 ;t1 ;:::;tk (x0 ; x1 ; : : : ; xk ) = (Xt0 = x0 ; Xt1 = x1 ; : : : ; Xtk = xk ). By independence of increments of X we have that X t0 ;t1 ;:::;tk (x0 ; x1 ; : : : ; xk ) = Qk X X 0 (x0 ) i=1 ti 1 ;ti (xi 1 ; xi ). Also by (2.8.1) and Lemma 1.11.5 X ti

e 1 ;ti (xi 1 ; xi ) = inf 2Rd

(xi xi 1 ) exp(G t

i ()

Gti 1 ()):

Thus, X t0 ;t1 ;:::;tk = t0 ;t1 ;:::;tk Since by Theorem 2.2.2 and Corollary 1.7.12 X (x) = inf t1 ;:::;tk X t0 ;t1 ;:::;tk (xt0 ; xt1 ; : : : ; xtk ), the required follows by Lemma 2.8.3.

Remark 2.8.6. Note also that if X is an idempotent process with

independent increments and cumulant G() on an idempotent prob- ability space ( ; ), then Gt () Gs () = ln S exp (Xt Xs ) so that Gt () Gs () is convex for s < t.

If Gt () is given in terms of characteristics by (2.7.7) and (2.7.55), we have the following consequence.

Corollary 2.8.7.

Let Gt () have the form (2.7.7) and (2.7.55), where bs , cs , s , and ^s do not depend on x. If, in addition, (L0 ) 1 + inf e jxj 1 ^s > 0; t 2 R+ ; 2 R+ ; st then x is a unique deviability on C such that X is a semimaxingale starting at x with independent increments and local characteristics (b; c; ; ^) under x . Proof. By Theorem 2.8.5 we only need to prove that Gt () is dierentiable in , which follows by condition (L0 ).

The next lemma gives suÆcient conditions for condition (L0 ) to hold.

Lemma 2.8.8. Condition (L0) holds if at least one of the following conditions holds

1. alim ^ (jxj a) > 0; t 2 R+ ; !1 sinf t s 2. sup ^s (Rd ) < 1; t 2 R+ ; st

© 2001 by Chapman & Hall/CRC

206

Maxingales

3. for every t 2 R+ there exists > 0 such that sup ejxj ^s < 1: st Proof. It is straightforward to see that conditions 1 and 2 imply (L0 ). Let condition 3 hold. We have for 2 R+ , s 2 R+ and b 2 R+

1 + (e jxj 1) ^s 1 + e jxj 1(jxj < b) 1 ^s 1 + e b=(1 e b ejxj) 1 ^s = 1 + (e b= 1)^s (Rd ) e b= e b ejxj ^s : (2.8.2)

Since ^s (Rd ) 2 [0; 1], it follows that 1 + (e b= 1)^s (Rd ) e b= , and, hence picking b such that eb 2 supst ejxj ^s ; which is possible by hypotheses, we conclude that the left-most side of (2.8.2) is bounded from below by e b= =2 for s t. We now consider the case where the cumulant depends on x. The following is an existence result for a diusion maxingale problem. It is a consequence of Theorem 2.5.19 and Theorem 2.6.24. We denote xt = sup0stjxsj. Theorem 2.8.9. Let the maxingale problem (x; G) be speci ed by the cumulant

Gt (; x) =

Zt

0

1 bs(x) ds + 2

Zt

0

cs (x) ds;

where functions (bs (x)) and (cs (x)), assuming values in Rd and the space of symmetric positive semi-de nite d d-matrices, respectively, are C-progressively measurable and continuous in x, and Rt Rt j b ( x ) j ds < 1 and k c ( x ) 0 s 0 s k ds < 1. Let also bt (x) and ct (x) satisfy the linear-growth conditions jbt (x)j lt(1+x ); kct (x)k lt(1+x 2 ); t 2 R+ ; x 2 C ; t

t

Rt 0 ls ds

0. Then for s t by the de nition of X;N1 ;N2

S (Zt jAs^N )(x0 ; n01 ; n02 ) =

sup

(x;n1 ;n2 )2C (R+ ;R3 ) t^ZN (x) t^ZN (x) exp n1 us( ) ds (e 1) us( ) ds 0 0 t^ZN (x) t^ZN (x) + n2 vs ( ) ds (e 1) vs ( ) ds 0 0 X; N ; N 1 2 ( ; n1 ; n2 ) s^N ( 0 ; n01 ; n02 ): (2.8.3)

x

x

x

x

x

x

jA

Since X;N1 ;N2 (x; n1 ; n2 )jAs^N (x0 ; n01 ; n02 ) N (R s^N (x0 ) ur (x0) dr n1)N (R s^N (x0 ) vr (x0 ) dr n2); 0 0

it follows that the right-hand side of (2.8.3) is not greater than Zs^N (x0 ) (x0 ). For the reverse inequality, let x^ be de ned on [0; s ^ N (x0 )] by x^ (q) = x0 (q), on [s ^ N (x0 ); t ^ N (^x)] as a solution of the equation

x^ q = x0

s^N (x0 )

Zq

+ e

© 2001 by Chapman & Hall/CRC

s^N (x0 )

ur (^x) dr e

Zq

s^N (x0 )

vr (^x) dr;

209

Maxingale problems

and on [t ^ N (^x); 1) by Zq

x^ q = x^ t^N (^x) +

t^N (^x)

ur (^x) dr

Zq

t^N (^x)

vr (^x) dr:

The latter two equations have solutions by a standard R 0 argument. We de ne n^ 1 on 0; 0s^N (x ) ur (x0 ) dr by R R 0 n^ 1 (q) = n01 (q), on 0s^N (x ) ur (x0 ) dr; 0t^N (^x) ur (^x) dr by R R s^N (x0 ) 0 0 ) dr and n^ 1 (q) = n01 0s^N (x ) ur (x0 ) dr + e q u ( x r 0 R R on 0t^N (^x) ur (^x) dr; 1) by n^ 1 (q) = n^ 1 0t^N (^x) ur (^x) dr + R t^N (^x) q ur (^x) dr . Similarly, n^ 2 (q) = n02 (q) on 0 R s^N (x0 ) R 0 0; 0 vr (x0 ) dr , n^ 2 (q) = n02 0s^N (x ) vr (x0 ) dr + e q R s^N (x0 ) R s^N (x0 ) R vr (x0 ) dr on vr (x0 ) dr; 0t^N (^x) vr (^x) dr , 0 0 R R t^N (^x) and n^ 2 (q) = n^ 2 0t^N (^x) vr (^x) dr + q vr (^x) dr on 0 R t^N (^x) vr (^x) dr; 1). Then (^x; n^ 1 ; n^ 2 ) satis es equation (2.6.11), 0 X; N ; N 1 2 (^ so x; n^ 1 ; n^ 2) = N (^n1 )N (^n2). Since

exp n^ 1

t^ZN (^x)

ur (^x) dr

0

= exp n01

(e

ur (x0 ) dr

0

(e

1)

us (^x) ds N (^n1 )

0

s^N (x0 ) Z

1)

t^ZN (^x)

s^Z N (x0 ) 0

ur (x0 ) dr N [n01 ]As^N (x0 )

and

exp

n^ 2

t^ZN (^x) 0

© 2001 by Chapman & Hall/CRC

vr (^x) dr (e

1)

t^ZN (^x) 0

vs (^x) ds N (^n2 )

210

Maxingales

= exp

n02

s^Z N (x0 )

vr (x0 ) dr

0

(e

1)

s^Z N (x0 ) 0

vr (x0 ) dr N [n02 ]As^N (x0 ) ;

we conclude that the expression in the supremum on the right-hand side of (2.8.3) evaluated at (^x; n^ 1 ; n^ 2 ) equals Zs^N (x0 ) (x0 ). Thus, S (Zt^N jAs^N ) = Zs^N . Therefore,

S (Zt^N jAs ) = S (Zt^N jAs^N ) 1(s N ) _ ZN 1(s > N ) = Zs^N 1(s N ) _ ZN 1(s > N ) = Zs^N :

We now study the uniqueness issue. We introduce for future use for t 2 R+

It (x) = x;t(x) = exp(

Zt

sup

((s))20

0

(s) dxs dGs ((s); x)); (2.8.4) (

x0 = x; Ix;t(x) = I1t(;x); ifotherwise ; Ix;t(x)); x;t( ) = sup x;t(x); C : x2C

(2.8.5) (2.8.6)

As we have seen, x is a natural candidate for a solution to (x; G). We rst show that it is a tight -smooth idempotent measure on C under fairly general assumptions.

De nition 2.8.11. We say that G() satis es the linear-growth con-

dition if there exist R+ {valued, increasing and continuous in t functions F l () = (Ftl (); t 2 R+ ); 2 Rd ; such that F0l () = Ftl (0) = 0 and for some increasing function kt 2 R+ we have for all 0 s < t , x 2 C and 2 Rd Gt (; x) Gs (; x) F l ((1+kt x )) F l ((1+kt x )): t

t

s

t

Lemma 2.8.12. Let G() satisfy the linear-growth condition. Then x is a tight -smooth idempotent measure on C .

© 2001 by Chapman & Hall/CRC

211

Maxingale problems

Proof. With no loss of generality we assume that x = 0 and check that I0 (x) = ln 0 (x) de ned by (2.7.2), (2.7.3), and (2.7.5), where x = 0, is a tight rate function on C in the sense of Remark 1.7.17, i.e., the sets LI0 (a) = fx 2 C : I0 (x) ag are compact for all a 2 R+ . Let x 2 LI0 (a). Then x0 = 0 by (2.7.5). By (2.7.2), (2.7.3), (2.7.5), and the linear-growth condition we have for 0 s < t, denoting by ei ; i = 1; : : : ; 2d; the d-vector, whose b(i + 1)=2cth entry is 1 if i is odd, 1 if i is even, and the rest of the entries are equal to 0, jxt xsj d max ei (xt xs) i=1;:::;2d 1 + kt xt 1 + kt xt ei ei d i=1max Gt G +a s ;:::;2d 1 + kt xt 1 + kt xt d max Ftl (ei ) Fsl (ei ) + da: (2.8.7) i=1;:::;2d

Since Ftl () is increasing in t, x has bounded variation over bounded intervals; therefore, since kt is increasing, for T > 0 ZT

0

d Vart x 1 + kT xt

2d X

d

i=1

FTl (ei )+ da

(recall that F0l () = 0), which implies that ZT

0

dxt 1 + cT xt

where cT = kT that

d

2d X

i=1

FTl (ei )+ da;

_ 1. By (2.8.8) and the fact that x0 = 0 we deduce 2d

X ln(1+ cT xT ) cT d( FTl (ei )+ a):

i=1

Hence, sup

x2L 0 (a) I

(2.8.8)

xT < 1:

(2.8.9) (2.8.10)

In analogy with (2.8.7) we can write for T > 0 and b > 0 jx x j b ei (xt xs ) b t s d max i=1;:::;2d 1 + kt xt 1 + kt xt d max Ftl (bei ) Fsl (bei ) + da: i=1;:::;2d

© 2001 by Chapman & Hall/CRC

212

Maxingales

Therefore, for Æ > 0 sup

x2L 0 (a) I

jxt xsj

sup

s;t2[0;T ]: js tjÆ

db i=1max sup ;:::;2d

s;t2[0;T ]: js tjÆ

l F (bei )

t

Fsl (bei ) +a (1+kT sup

x2L 0 (a) I

so by continuity of Ftl () in t lim sup sup Æ!0 x2LI0 (a)

sup

s;t2[0;T ]: js tjÆ

jxt xsj

xT ); !

da 1 + kT sup xT : b x2LI0 (a)

Letting b ! 1 and using (2.8.10), we conclude that the left-hand side of the latter inequality is 0. An application of Arzela{Ascoli's theorem ends the proof.

Remark 2.8.13. The above proof shows that if Ix;T (x) a, then the bound (2.8.9) holds, which implies that the sets [st fxs : x;s(x)

g are bounded under the hypotheses for t 2 R+ and 2 (0; 1]. By Lemma 2.7.11 if solves the maxingale problem (x; G), then (x) x (x); x 2 C : (2.8.11) Our goal is to establish conditions when we actually have equality above. Clearly, we need only to be concerned with the case x (x) > 0. We assume in the sequel that the cumulant is given by R t (2.7.7) and (2.7.13), where (bs (x)) is C-progressively measurable, 0 jbs (x)jds < 1, (^gs(; x)) is non-negative, continuous in and B([0; t]) B(Rd )

Ct =B(R+R ){measurable as a map from [0; t] R d C to R+ , g^s (0; x) = t 0, and 0 supjj=A g^s (; x) ds < 1 for t 2 R+ ; A 2 R+ and x 2 C . By Lemma 2.7.12 if x (x) > 0; then x is absolutely continuous. We note that by (2.7.2), (2.7.5), (2.7.6), (2.8.4), (2.8.5), and (2.8.6) x;t(x) # x(x); x 2 C ; as t ! 1: (2.8.12) The argument of the proof of Lemma 2.7.12 also shows that 8 t Z > >

0 > :

hs (x_ s ; x) ds

1

© 2001 by Chapman & Hall/CRC

if x is absolutely continuous on [0; t]; otherwise.

(2.8.13)

213

Maxingale problems

For the following theorem we recall that Z () = (Zt (; x); t 2 R + ; x 2 C ) is given by (2.7.20) if the integral on the right-hand side is well de ned and nite, and Zt (; x) = 0 otherwise.

Theorem 2.8.14. Let deviability solve dthe maxingale problem (x; G) and x^ 2 C . Let there exist an R {valued function ^ = (^ (s; x); s 2 R+ ; x 2 C ) with the following properties. a) Z (^ ) is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ) and admits a localising sequence of strictly Luzin C-stopping times; b) if x~ is such that x (~x) > 0 and a.e. in s 2 [0; t]

^ (s; x~ )x~_ s gs (^ (s; x~ ); x~ ) = sup (x~_ s gs (; x~ )); 2Rd

then x~ s = x^ s ; s 2 [0; t]; where t 2 R+ .

(2.8.14)

Then (ps 1 Æ ps x^ ) = x;s(^x); s 2 R+ ; and (^x) = x (^x).

Proof. Let fN ; N 2 N g be a localising sequence of strictly Luzin Cstopping times for Z (^ ) so that the (Zt^N (^ ); t 2 R+ ) are strictly Luzin uniformly maximable exponential maxingales under . Since S Zt^N (^ ) = 1 and limA!1 S Zt^N (^ ) 1(Zt^N (^ ) > A) = 0; for A large enough S Zt^N (^ ) 1(Zt^N (^ ) A) = 1 so that the inequality

S Zt^N (^ ) 1(Zt^N (^ ) A) sup Zt^N (x) (^(x); x)(x) _ (aA);

x2K(a)

where a 2 (0; 1], implies that for a small enough supx2K(a) Zt^N (x) (^ (x); x)(x) 1: On the other hand, by the de nitions of Z (^ ) and x , and (2.8.11)

Zt^N (x) ((x); x)(x) Zt^N (x) ((x); x)x (x) 1:

Thus, if a > 0 is small enough, then sup Zt^N (x) (^ (x); x)(x) = 1:

x2K (a)

© 2001 by Chapman & Hall/CRC

(2.8.15) (2.8.16)

214

Maxingales

Being a strictly Luzin idempotent variable on (C ; ), Zt^N (x) (^ (x); x) is continuous in x when restricted to K (a). Also (x), being a deviability density, is upper semi-continuous. Therefore, the product Zt^N (x) (^ (x); x)(x) is upper semicontinuous when restricted to K (a). As the latter set is compact, the supremum in (2.8.16) is attained, so for some xN 2 C

Zt^N (xN ) (^ (xN ); xN )(xN ) = 1

(2.8.17)

(we suppress in xN dependence on t). Then by (2.8.15)

Zt^N (xN ) (^ (xN ); xN )x (xN ) = 1:

(2.8.18)

In particular, x (xN ) > 0 so that by (2.7.6) and Lemma 2.7.12 xN is absolutely continuous and xN0 = x. Thus, by the de nitions of Z and x as well as Lemma 2.7.12 we have that almost everywhere on [0; t ^ N (xN )]

^ (s; xN )x_ Ns gs (^ (s; xN ); xN ) = sup (x_ Ns gs (; xN )) 2Rd

and Z1

t^N (xN )

sup ( x_ Ns gs (; xN )) ds = 0: 2Rd

(2.8.19) (2.8.20)

Hence, by the requirements on ^ we conclude that xNs = x^ s ; 0 s t ^ N (xN ). Since t ^ N (x) is a C{stopping time, by Lemma 2.2.21 t ^ N (xN ) = t ^ N (^x). Therefore, by (2.7.5), (2.7.6), Lemma 2.7.12, (2.8.5), (2.8.6), (2.8.20), and (2.8.13)

x;t^N (^x) (^x) = x;t^N (xN )(xN ) = x(xN ): (2.8.21) Thus, by (2.8.17), (2.8.18) and (2.8.21), (xN ) = x(xN ) = x;t^N (^x) (^x); which implies, since xN 2 pt^1N (^x) Æ pt^N (^x) x^ , that (pt^1N (^x) Æpt^N (^x) x^ ) (xN ) = x;t^N (^x) (^x): On the other hand, by (2.8.12)

© 2001 by Chapman & Hall/CRC

(2.8.22)

x;t(x) x(x) for x 2 C

and

Maxingale problems

215

t 2 R+ , so by (2.8.11), (2.8.4), (2.8.5), (2.8.6), and Lemma 2.2.21

(pt^1N (^x) Æ pt^N (^x) x^ ) = supf(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g supfx(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g supfx;t^N (x)(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g = x;t^N (^x) (^x): Comparing this with (2.8.22) yields (pt^1N (^x) Æ pt^N (^x) x^ ) = x;t^N (^x) (^x): (2.8.23) 1 ^ = pt 1 Æpt x^ ; Since N (^x) ! 1, it follows that \1 N =1 pt^N (^x) Æpt^N (^x) x so since the sets pt^1N (^x) Æpt^N (^x) x^ are closed, by the -smoothness property of deviability (pt 1 Æpt x^ ) = lim (pt^1N (^x) Æpt^N (^x) x^ ): (2.8.24) N !1 Also the convergence N (^x) ! 1, (2.8.4), (2.8.5), and (2.8.6) yield x;t(^x) = Nlim (^x): (2.8.25) !1 x;t^N (^x) Putting together (2.8.23), (2.8.24) and (2.8.25) results in the equality (pt 1 Æ pt x^ ) = x;t (^x). The nal assertion follows by taking in both sides of the latter equality the limit as t ! 1, using -smoothness of and (2.8.12).

Remark 2.8.15. In the sequel we routinely omit indications that cer-

tain relations hold almost everywhere with respect to Lebesgue measure when this is understood. Conditions for Z () to be a strictly Luzin-continuous exponential maxingale on (C ; C ; ) are given in Theorem 2.7.16. However, the deviability is not known to us, so it would be diÆcult to verify the hypotheses of the theorem. Therefore, we introduce somewhat cruder conditions, which have the advantage of not involving . The following conditions replace conditions (I ) and (II ). I The function bs (x) is continuous in x and Zt

0

sup jbs (x)j ds < 1

x2K

for all compacts K C and t 2 R+ .

© 2001 by Chapman & Hall/CRC

216

Maxingales

II The function g^s (; x) is convex in 2 Rd , continuous in (; x) R d C , and sup sup sup g^s (; x) < 1; jjA st x2K

2

lim sup sup g^s (; x) = 0 x2K

!0 st

for all compacts K C , t 2 R+ and A 2 R+ .

If gs (; x) has the form (2.7.55), Lemma 2.7.27 provides suÆcient conditions for conditions I and II to hold in terms of characteristics.

Lemma 2.8.16. Let gs (; x) be given by (2.7.55) , where cs (x); s 2 d R + ; x 2 C , s ( ; x); s 2 R + ; 2 B (R ); x 2 C , and ^s ( ; x); s 2 R+ ; 2 B(Rd ); x 2 C are as de ned in Section 2.7 with Cprogressive measurability replaced by C-progressive measurability. R Let bs (x) and cs (x)R be continuous in x, and Rd (exp( x) 1 x) s (dx; x) and Rd exp( x) ^s (dx; x) be continuous in (; x). If for all compacts K C , A 2 R+ and t 2 R+ Zt

0

sup jbs (x)j ds < 1; sup sup kcs (x)k < 1; st x2K

x2K

sup sup st x2K

Z

Rd

lim sup sup x2K

!0 st

1 Ajxj) s (dx; x) < 1;

(eAjxj Z

1 x) s (dx; x) = 0;

(ex

Rd

sup sup st x2K

Z Rd

Z lim sup sup !0 st x2K Rd

eAjxj ^s(dx; x) ds < 1; ex

)

1 ^s (dx; x = 0;

then conditions I and II are satis ed.

We now de ne the class of integrands.

De nition 2.8.17. Let ^ denote the set of all Rd {valued C{ progressively measurable idempotent processes = ((t; x); t 2 R + ; x 2 C ) such that the (t; x) are continuous in x, for 2 R , © 2001 by Chapman & Hall/CRC

217

Maxingale problems

t 2 R+ and x 2 C Zt

0

g^s ((s; x); x) ds < 1

and for every compact K C

sup

Zt

x2K 0

g^s ((s; x); x) 1(j(s; x)j > A) ds ! 0 as A ! 1:

Obviously, ^ ^ for every deviability on C .

Remark 2.8.18. If x is a deviability, then both in conditions I and II and the de nition of ^ we could consider only compacts Kx (a), where a 2 (0; 1] and Kx (a) = fx : x (x) ag, and require that x be such that x (x) > 0. The following consequence of Theorem 2.7.16 allows us to check that Z () satis es condition a) of Theorem 2.8.14. Theorem 2.8.19. Let conditions I and II hold. If 2 ^ and solves the maxingale problem (x; G), then the idempotent process Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ), which admits a localising sequence of strictly Luzin Cstopping times. Proof. By Theorem 2.7.16 Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ). It is straightforward to check that the sequence fN ; N 2 N g de ned by

N (x) = inf ft 2 R+ :

Zt

0

g^s (2(s; x); x) ds+t N g

is a localising sequence of strictly Luzin-continuous C-stopping times. We now consider the issue of choosing the function ^ to satisfy condition b) of Theorem 2.8.14. There are two ways of doing this as is shown by the following lemma. We denote by rgs (; x) the gradient of gs (; x) with respect to if it is well de ned.

© 2001 by Chapman & Hall/CRC

218

Maxingales

Lemma 2.8.20. Let gs(; x) be dierentiable in .

Let x^

absolutely continuous. 1. If a function ^ = (^ (s; x); s 2 R+ ; x 2 C ) is such that x^_ s = rgs(^(s; x); x)

2C

be

for x -almost all x and almost all s 2 R+ , then condition b) of Theorem 2.8.14 is satis ed. 2. If a function ^ = (^ (s); s 2 R+ ) is such that the equation x_ s = rgs(^(s); x); x0 = x;

has the only solution x = x^ , then condition b) of Theorem 2.8.14 is satis ed. Proof. By the requirements on x~ in part b) of Theorem 2.8.14 and the necessary condition for attaining supremum x~_ s = rgs (^ (s; x~ ); x~ ). Hence, if ^ is from part 1 of the lemma, then x^_ s = x~_ s so that since x^ 0 = x~ 0 = x we conclude that x^ = x~ . If ^ is from part 2 of the lemma, then x^ = x~ by de nition.

We are able to obtain somewhat general uniqueness results only for the case where ^ is chosen as in part 1 of Lemma 2.8.20, so we concentrate on that case. However, we give examples that show an application of the approach in part 2. We next state a uniqueness result for a \diusion" maxingale problems. Theorem 2.8.21. Let the canonical process X be a semimaxingale under with local characteristics (b; c; 0; 0) starting at x. Let the following conditions hold: 1. the functions bs (x) and cs (x) are continuous in x 2 C , 2. for every t 2 R+ and compact K C Zt

0

sup jbs (x)j2 ds < 1; sup sup kcs (x)k < 1; st x2K

x2K

3. for every t 2 R+ and compact K C

inf inf inf cs (x) > 0:

st x2K 2Rd : jj=1

© 2001 by Chapman & Hall/CRC

219

Maxingale problems

Then = x . Proof. We rst note that g^t (; x) = ct (x)=2. Let x^ 2 C be such that x (^x) > 0. We prove that (^x) = x (^x). We apply Theorem 2.8.14. By Lemma 2.8.16 conditions I and II are satis ed. Since x^ is absolutely continuous and ct (x) is positive de nite for all x 2 C , we can de ne

^ (t; x) = ct (x)

1

(x^_ t bt(x)):

(2.8.26)

Since x^_ t = rgt (^ (t; x); x) for all x, by Lemma 2.8.20 ^ satis es the condition of part b) of Theorem 2.8.14, so by part a) of The^ orem 2.8.14 and Theorem 2.8.19 it suÆces to check that ^ 2 . Firstly, we note that for x 2 K , where K is compact, Zt

0

sup jx^_ s bs(x)j2 ds < 1:

(2.8.27)

x2K

Indeed, we have Zt

0

sup jx^_ s

bs (x)j2 ds 2

x2K

Zt

jx^_ s bs(^x)j2 ds

0 Zt

+2

0

sup jbs (^x) bs (x)j2 ds: (2.8.28)

x2K

Since by Lemma 2.7.12

I(^x) = 12

Z1

0

(x^_ s

bs (^x)) cs (^x) 1 (x^_ s bs (^x)) ds

1 sup kcs (^x)k 2 st

1

Zt

0

jx^_ s bs(^x)j2 ds;

we conclude that the rst term on the right of (2.8.28) is nite. The second one is nite by hypotheses. Inequality (2.8.27) is proved.

© 2001 by Chapman & Hall/CRC

220

Maxingales

Next, we have by (2.8.26) for x 2 K (a) and 2 R+ , denoting c = inf st inf x2K inf 2Rd : cs (x), that jj=1 Zt

0

g^s (^ (s; x); x) 1(j^ (s; x)j > A)ds

2 = 2

Zt

0

(x^_ s

bs (x)) cs (x) 1 (x^_ s bs (x)) 1(j^ (s; x)j > A)ds

2 c 2

1

Zt

0

jx^_ s bs(x)j2 1(j^ (s; x)j > A)ds

so that by (2.8.27) and absolute continuity of the Lebesgue integral the required limit lim sup

A!1 x2K

Zt

0

g^s (^ (s; x); x) 1(j^ (s; x)j > A)ds = 0

would follow by lim sup

A!1 x2K

Zt

0

1(j^ (s; x)j > A)ds = 0:

The latter limit follows since by (2.8.26), (2.8.27) and hypotheses sup

Zt

x2K 0

j^(s; x)j2 ds c

2

Zt

0

sup jx^_ s bs(x)j2 ds < 1:

x2K

As a consequence of this result, Theorem 2.6.24 and Theorem 2.6.30, we have the following existence and uniqueness result for idempotent Ito equations.

Theorem 2.8.22. Let (bds(x); s 2 R+ ; dxd2 C ) and (s(x); s 2 R + ; x 2 C ) be respective R -valued and R -valued C-progressively measurable idempotent processes. Let the following conditions hold:

© 2001 by Chapman & Hall/CRC

221

Maxingale problems

1. bs (x) and s (x) are continuous in x 2 C ,

2. linear growth: for every t 2 R+ and x 2 C Zt

0

jbs(x)j2 ds +sup ks(x)k 2 < 1; 1 + xs 2

st

1 + xs 2

3. for every t 2 R+ and compact K C

inf inf inf s (x)s (x)T > 0:

st x2K 2Rd : jj=1

Then the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x; has a unique Luzin solution. The idempotent distribution of X is given by 1 Z1 X (x_ t bt (x)) t (x)t (x)T 1 (x_ t bt (x)) dt (x) = exp 2 0

if x is absolutely continuous and

x0 = x, and X (x) = 0 otherwise.

Proof. By Theorem 2.6.24 the equation has a Luzin solution X . By Theorem 2.6.30 and Theorem 2.8.21 uniqueness holds. The form of X is given in Lemma 2.6.19.

Remark 2.8.23. One can weaken the conditions of Theorem 2.8.22 by requiring that the non-degeneracy condition 3 in the hypotheses hold for compacts Kx (a), where a 2 (0; 1], and x in conditions 1 and 2 is such that x(x) > 0. If the function gs (; x) is more general than in Theorem 2.8.21, we cannot apply Theorem 2.8.14 to all x 2 C such that x (x) > 0, so we have to introduce additional regularity conditions, e.g., require that the derivative x_ s be locally bounded. As for \nonregular" x, a way to establish for them the equality (x) = x (x) is to see to it that each such x can properly be approximated by \regular" x. Below we assume without further mentioning that is a solution to the maxingale problem (x; G).

© 2001 by Chapman & Hall/CRC

222

Maxingales

De nition 2.8.24. Let D C : We de ne the x{closure of D as the set of all x 2 C such that x (x) > 0 for which there exists a sequence xk 2 D such that xk ! x and x;t (xk ) ! x;t (x) as k ! 1 for all t 2 R+ . We say that D is x-dense in C if its x{closure coincides with the set fx 2 C : x(x) > 0g. Remark 2.8.25. If conditions I and II hold, then by (2.8.4), (2.8.5) and (2.8.6) x;t (x) is an upper semi-continuous function of x. In this case in the above de nition of the x {closure of D it is suÆcient to require that xk 2 D be such that xk ! x and lim inf k!1 x;t (xk ) x;t (x); t 2 R+ . Lemma 2.8.26. If1 (pt 1 Æ pt x) = x;t(x); t 2 R+ ; for all x 2 D C , then (pt Æ pt x) = x;t (x); t 2 R+ ; for all x from the x{closure of D. If, in addition, the set D is x{dense in C , then uniqueness holds for the maxingale problem (x; G) with = x . Proof. Let x belong to the x {closure of D. Let xk 2 D be such that xk ! x and x;t (xk ) ! x;t (x); t 2 R+ . Since pt xk ! pt x, it follows that, for arbitrary " > 0, pt 1 Æ pt xk pt 1 B" (pt x) for all k large enough, where B" (pt x) is the closed "-ball about pt x. Sinceby the -smoothness property of deviability lim"!0 pt 1 B"(pt x) = (pt 1 Æ pt x); we conclude that lim supk!1 (pt 1 Æ pt xk ) (pt 1 Æ pt x): Since also (pt 1 Æ pt xk ) = x;t (xk ) ! x;t (x) as k ! 1, we have that x;t (x) (pt 1 Æ pt x): On the other hand, x;t (x) x(pt 1 Æpt x) (pt 1 Æpt x1) by the de nitions of x;t and x, and (2.8.11), so x;t (x) = (pt Æ pt x); t 2 R+ . Finally, if D is x -dense in C , then by the part just proved, the -smoothness property of deviability and (2.8.12) (x) = limt!1 (pt 1 Æpt x) = limt!1 x;t (x) = x (x) when x (x) > 0. If x(x) = 0, then (x) = 0 = x(x) by (2.8.11). Theorem 2.8.27. Let gs(; x) meet conditions I and II and be differentiable in . Let there exist a family fGm ; m 2 N g of subsets of R d and an R d -valued function ((s; x; y ); s 2 R + ; x 2 C ; y 2 Rd ), which is B[0; t] Ct B(Rd )=B(Rd ) {measurable when restricted to [0; t] C Rd for t 2 R+ , continuous in x, bounded on the sets [0; t] K Gm , where t 2 R+ , K C and is compact, and m 2 N , and such that

y = rgs ((s; x; y); x)

© 2001 by Chapman & Hall/CRC

(2.8.29)

Maxingale problems

223

for y 2 [1 m=1 Gm , (almost all) s 2 R+ and x {almost all x. Let 1 [ D= fx 2 C : x is absolutely continuous, x0 = x, and m=1

x_ s 2 Gm ; s 2 R+ g: (2.8.30) Then (pt 1 Æ pt x) = x;t (x); t 2 R+ ; whenever x is in the x { closure of D. If the set D is, in addition, x {dense in C , then uniqueness holds for the maxingale problem (x; G) with = x . Proof. For x^ 2 D we de ne ^ (s; x) = (s; x; x^_ s ): The function ^ (s; x) is C-progressively measurable, bounded on the sets [0; T ] K for T 2 R+ and continuous in x. Therefore (^ (s; x); s 2 R+ ; x 2 C ) 2

^ so that by Theorem 2.8.19 Z (^ ) is a strictly Luzin-continuous local exponential maxingale under and admits a localising sequence of C-stopping times. Since also by (2.8.29) x^_ s = rgs (^ (s; x); x) for (almost all) s 2 R+ and x -almost all x, Theorem 2.8.14 and Lemma 2.8.20 imply that (pt 1 Æ pt x^ ) = x;t (^x); t 2 R+ , and an application of Lemma 2.8.26 ends the proof. We give an application. Theorem 2.8.28. Let d = 1. Let the canonical process X on C be a semimaxingale starting at x 2 R+ under with local characteristics (bq; 0; ; 0), where s( ; x) = 1(qs (x) 2 )bs (x); and bs (x) and qs(x) are R+ {valued functions, which are C-progressively measurable in (s; x) and continuous in x 2 C . Let for every t 2 R+ and compact KC inf inf b (x) > 0; sup sup bs (x) < 1; st x2K s st x2K inf inf q (x) > 0; sup sup qs(x) < 1: st x2K s st x2K Then = x . Proof. We apply Theorem 2.8.27. Conditions I and II are met by Lemma 2.8.16. The function gs (; x) = eqs (x) 1 bs (x) is differentiable in . We take in the hypotheses of Theorem 2.8.27 Gm = [1=m; m] and 8 1 y < ln ; y > 0; (s; x; y) = qs(x) bs (x)qs (x) : 0; y 0:

© 2001 by Chapman & Hall/CRC

224

Maxingales

We check that D is x {dense in C . Let x 2 C be such that x (x) > 0. By Lemma 2.7.12 x is absolutely continuous; also

I(x) =

Z1

0

sup(x_ t (eqt(x) 1)bt (x))dt 2R

so that x_ t 0 a.e. We de ne xk by xk0 = x and x_ ks = (x_ s 1(x_ s k)) _ k1 . Convergence xk ! x is obvious. We prove that for t 2 R+ lim

k!1

Zt

0

sup(x_ ks 2R =

(eqs (x Zt

k)

1)bs (xk ))ds

sup(x_ s 2R

0

(eqs (x)

1)bs (x))ds: (2.8.31)

We have Zt

0

sup(x_ ks 2R

(eqs (x

k)

1)bs (xk )) ds

Zt

x_ s ln x_ s qs(xk ) bs(xk )qs(xk ) 0 1 x_ s k; x_ s k1 ds =

+

Zt

0

1 1 ln k k kqs(x ) kbs (x )qs (xk )

x_ s + bs(xk ) qs(xk ) 1 k) + b ( x kqs (xk ) s

1 x_ s < k1 + 1(x_ s > k)ds:

Since 0t x_ s ln x_ s 1(x_ s > 1)ds < 1 by the fact that I(x) < 1 and hypotheses, bs (xk ) ! bs (x) and qs (xk ) ! qs(x) as k ! 1, Lebesgue's dominated convergence theorem implies that the right-hand side conR

© 2001 by Chapman & Hall/CRC

225

Maxingale problems

verges to Zt

0

x_ s ln x_ s qs (x) bs(x)qs (x)

x_ s + bs(x) 1(x_ s > 0) ds qs(x) +

Zt

0

bs(x) 1(x_ s = 0) ds

ending the proof of (2.8.31). We now consider a version of the above result, which will be used in an application to the analysis of a many-server queue in part II. This result also shows the use of the other method of choosing the function ^ .

Theorem 2.8.29. Let d = 1. Let deviability on C

be such that the canonical process X is a semimaxingale on (C ; ) starting at x 2 R+ with local characteristics (b; 0; ; 0), where

s( ; x) = 1(1 2 )vs (x) + 1( 1 2 )us (x)(xs ^ ms (x)); bs(x) = vs (x) us (x)(xs ^ ms (x));

and vs (x), us (x) and ms(x) are R+ {valued functions, which are Cprogressively measurable in (s; x) and locally Lipshitz-continuous in x 2 C . Let also for every t 2 R+ and compact K C

inf inf vs (x) > 0; sup sup vs (x) < 1; st x2K inf inf u (x) > 0; sup sup us (x) < 1; st x2K s st x2K inf inf m (x) > 0: st x2K s st x2K

Then = x .

Proof. We rst consider the case where x > 0. Let x^ be such that x(^x) > 0, sups2[0;t]jx^_ sj < 1 and inf s2[0;t] x^ s > 0 for t 2 R+ . We de ne a function ^ (s) by the equality

x^_ s = e^(s) vs(^x)

e

© 2001 by Chapman & Hall/CRC

^ (s) u

x x^ s ^ms(^x):

s (^ )

(2.8.32)

226

Maxingales

The function ^ (s) is well de ned, satis es the conditions of part 2 of Lemma 2.8.20 and, being locally bounded and Lebesgue measurable, ^ also by Lemma 2.8.16 conditions I and II are met; is an element of ; so Z (^ ) is a strictly Luzin exponential maxingale by Theorem 2.8.19 and admits a localising sequence of strictly Luzin C-stopping times. Therefore, by Theorem 2.8.14 (^x) = x (^x) and (pt 1 Æ pt x^ ) = x;t(^x); t 2 R+ . For more general functions x^ such that x (^x) > 0 we apply Lemma 2.8.26. Let us consider the instance where inf s2[0;t] x^ s > 0 for t 2 R+ . We de ne Zs

x^ ks = x + x^_ p 1 jx^_ pj k dp:

(2.8.33)

0

Then sups2[0;t] jx^ ks x^ sj ! 0 as k ! 1; in particular, k lim inf k!1 inf s2[0;t] x^ s > 0 so by the part just proved (pt 1 Æ pt x^ k ) = x;t(^xk ); t 2 R+ . Since for absolutely continuous x

It(x) =

Zt

0

sup x_ s (e 1)vs (x) (e 2R

1)us (x)

xs^ms(x) ds;

to prove that (^x) = x (^x) it is suÆcient to show by Lemma 2.8.26 and Remark 2.8.25 that for all t 2 R+ lim sup k!1 (e

Zt

0

Zt

sup x^_ ks 2R

(e

1)vs (^xk )

1)us (^xk ) x^ ks ^ ms (^xk ) ds

sup x^_ s (e 1)vs (^x) (e 0

2R

1)us (^x) x^ s ^ms(^x) ds:

(2.8.34)

We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R

© 2001 by Chapman & Hall/CRC

1)us (^xk ) x^ ks ^ms (^xk ) ds

227

Maxingale problems

Zt

sup x^_ s (e 1)vs(^xk ) (e +

0 Zt 0

2R

1)vs (^xk ) (e

sup (e 2R

1)us (^xk ) x^ ks ^ms(^xk ) ds

1)us (^xk ) x^ ks ^ ms(^xk )

1 jx^_ sj > k ds:

(2.8.35)

The second integral on the right converges to 0 as k ! 1. We work with the rst integral. Let C1 > 0 and C2 > 0 be respective upper and lower bounds over large values of k and s 2 [0; t] for the k k x^ s ^ ms(^x ) , C3 > 0 and C4 > 0 { upper and lower bounds for the vs (^xk ), and C5 > 0 and C6 > 0 { upper and lower bounds for the us (^xk ). Let k (s) be the points where the supremums in the rst integral on the right-hand side of (2.8.35) are attained so that

x^_ s = ek (s) vs(^xk )

e

k (s) u

xk ) x^ ks ^ms(^xk ):

s (^

(2.8.36)

k (s) u (^ k ^k ^ Since for k (s) positive ek (s) vs (^xk ) e s x ) x s k ms (^xk ) C4 e (s) C5 C1 and for k (s) negative ek (s) vs (^xk ) e k (s) us (^xk ) x^ ks ^ ms(^xk ) C3 e k (s) C6 C2 ; we conclude that k (s)

e

jx^ sj +C C5 C1 _1; e _

4

jx^Cs j +C C3 _1; s 2 [0; t]: _

k (s)

6 2

We thus write for the rst integral on the right of (2.8.35) Zt

0

=

sup x^_ s (e 1)vs (^xk ) (e 2R Zt

0

k (s)x^_ s (e

© 2001 by Chapman & Hall/CRC

(2.8.37)

1)us (^xk ) x^ ks ^ ms(^xk ) ds

1)vs (^xk )

k (s)

(e

k (s)

1)us (^xk ) x^ ks ^ ms (^xk ) ds

228

Maxingales

Zt

0 Zt

+

0 Zt

+

k k (s)x^_ s (e (s) 1)vs (^x) (e

je

k (s)

1jjus (^xk ) x^ ks ^ ms (^xk )

sup x^_ s (e 1)vs(^x) (e + +

0 Zt 0

1)us (^x) x^ s ^ms(^x) ds

jek (s) 1jjvs (^xk ) vs(^x)j ds

0 Zt

0 Zt

k (s)

2R

us (^x) x^ s ^ ms(^x) j ds

1)us (^x) x^ s ^ ms (^x) ds

jek (s) 1jjvs (^xk ) vs(^x)j ds

je

k (s)

1jjus (^xk ) x^ ks ^ ms (^xk )

us (^x) x^ s ^ ms (^x) j ds:

By (2.8.37) and the facts that x^ is absolutely continuous, us (x), vs (x) and ms(x) are continuous in x and bounded, and the x^ k converge uniformly on [0; t] to x^ as k ! 1, the latter two integrals converge to 0 as k ! 1, so (2.8.34) follows. Let us now assume that x^ is an arbitrary function such that x(^x) > 0. Clearly, x^ is R+ -valued, absolutely continuous, and x^ 0 = x > 0. Let x^ ks = x + R0s x^_ u 1(^xu k 1=k) du. Since x^_ s = 0 on the set fx^ s = 0g (a.e.), we have that x^ ! x^ uniformly on bounded intervals. Since x^ ks = x^ s _ (1=k) for k large, by the part proved (pt 1 Æ pt x^ k ) = x;t (^xk ); t 2 R+ , so by Lemma 2.8.26 in order to prove that (^x) = x (^x) it is suÆcient to show that (2.8.34) holds. We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R

© 2001 by Chapman & Hall/CRC

1)us (^xk ) x^ ks ^ms (^xk ) ds

229

Maxingale problems

=

Zt

0

sup x^_ s (e 1)vs (^xk ) (e 2R

1)us (^xk ) x^ s ^ ms(^xk )

1(^xs 1=k) ds +

Zt

0

sup (e 2R

1)vs (^xk ) (e

1)us (^xk ) x^ ks ^ ms(^xk )

1(^xs < 1=k) ds:

(2.8.38) Routine calculations show that R t the second integral on the right-hand side of (2.8.38) converges to 0 vs (^x) 1(^xs = 0) ds as k ! 1. De ning k (s) on the set fs : x^ s 1=kg by x^_ s = ek (s) vs(^xk ) e k (s) us(^xk ) x^ s ^ms(^xk ); we have estimates analogous to (2.8.37) with the right inequality replaced by _ x^ s ^ms(^xk )e k (s) jx^ sjC+6 C3 _ x^ s ^ms(^xk ) so that Zt

0

sup x^_ s 2R

1)vs (^xk ) (e

(e

1)us (^xk ) x^ s ^ ms (^xk )

1(^xs 1=k) ds Zt

sup x^_ s (e 1)vs (^x) (e 0

2R

1)us (^x) x^ s ^ ms (^x)

1(^xs > 0) ds + +

Zt

0 Zt 0

jek (s) 1jjvs (^xk ) vs(^x)j ds

x^ s ^ ms(^xk )je +

Zt

0

© 2001 by Chapman & Hall/CRC

k (s)

je

k (s)

1jjus (^xk ) us (^x)j ds 1jjms (^xk ) ms (^x)jus (^x) ds:

230

Maxingales

The required convergence follows by the fact that the last three integrals on the right-hand side converge to 0 by the same argument as above. We now consider the case x = 0. Let x^ be such that x (^x) > 0, sups2[0;t]jx^_ s j < 1, inf s2[0;] x^_ s > 0, and inf s2[;t] x^ s > 0 for t 2 R+ and some > 0. We again de ne ^ (s) by (2.8.32). It is evidently bounded on [; t]. Next, since by (2.8.32) x^_ s e^(s) vs (^x), we have that e^(s) x^_ s =C30 ; hence, x^_ s = e^ (s) vs(^x) e ^(s) us (^x) x^ s ^ ms (^x) e^ (s) C40 C30 C50 x^ s ^ ms(^x) =x^_ s ; where C30 ; C40 and C50 have a similar meaning as above. We thus conclude that there exist A1 , A2 and A3 , which depend on x^ s ; s 2 [0; ]; but do not explicitly depend on x^_ s , such that for s 2 [0; ] A A1 x^_ s e^ (s) A2 x^_ s + _ 3 : (2.8.39) x^ s Since x^_ s is bounded both from below and above on [0; ], so is ^ (s). Hence, by Theorem 2.8.14 (^x) = x (^x) and (pt 1 Æ pt x^ ) = x;t(^x); t 2 R+ . Next, let x^ be such that x (^x) > 0, inf s2[0;] x^_ s > 0, and inf s2[;t] x^ s > 0 for t 2 R+ and some > 0. We de ne x^ k in analogy with the case x > 0 by (2.8.33). Then following the same line of reasoning we need bounds on k (s) de ned by (2.8.36) in order to prove that Zt

jek (s) 1jjvs (^xk ) vs(^x)j ds ! 0;

Zt

Z

0

je

k (s)

1jjus (^xk ) x^ ks ^ ms(^xk )

(2.8.40)

us(^x) x^ s ^ ms(^x) j ds ! 0;

(2.8.41)

jek (s) 1jjvs (^xk ) vs(^x)j ds ! 0;

(2.8.42)

Z

je

k (s)

0

© 2001 by Chapman & Hall/CRC

1jjus (^xk ) x^ ks ^ ms(^xk )

us (^x) x^ s ^ ms (^x) j ds ! 0:

(2.8.43)

231

Maxingale problems

On the interval [; t] bounds (2.8.37) apply yielding convergences (2.8.40) and (2.8.41). Bounds on [0; ] that imply limits (2.8.42) and (2.8.43) are given by (2.8.39) with suitable A1 , A2 and A3 . Next, let x^ be such that x (^x) > 0 and not identically equal to 0. Let k = inf fs 2 R+ : x^ s = 1=kg. We de ne x^ k by x^ ks = sx^ k =k for s 2 [0; k ] and x^ ks = x^ s _ (1=k) for s k . Then (pt 1 Æ pt x^ k ) = x;t(^xk ); t 2 R+ , so we need to prove that (2.8.34) holds. We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R Zk

sup x^_ ks (e 1)vs(^xk ) (e +

0 Zt 0

1)us (^xk ) x^ ks ^ ms(^xk ) ds

2R

sup x^_ s (e 1)vs (^xk ) (e 2R

1)us (^xk ) x^ ks ^ms(^xk ) ds

1)us (^xk ) x^ s ^ ms(^xk )

1(^xs 1=k) ds +

Zt

0

sup (e 2R

1)vs (^xk ) (e

1)us (^xk ) x^ ks ^ ms(^xk )

1(^xs < 1=k) ds:

(2.8.44)

By an argument similar to the one used for deriving an asymptotic bound for the right-hand side of (2.8.38) the limit superior as k ! 1 of the sum of the latter two integrals is not greater than It (^x). Let k (s) be the points, where the supremums in the rst integral on the right of (2.8.44) are attained. Then the integrand takes the form k k k (s)x^_ ks (e (s) 1)vs (^xk ) (e (s) 1)us(^xk ) x^ ks ^ ms(^xk ) k (s)x^_ ks + vs(^xk ) + us(^xk ) x^ ks ^ ms(^xk ) :

Since the derivatives x^_ ks equal x^ k =k and k ! 0 as k ! 1, the estimates (2.8.39) applied to x^ ks show that the rst integral on the right of (2.8.44) converges to 0 as k ! 1 provided limk!1 x^ Rk ln(^xk =k ) = 0: The latter limit follows since x^ k ln(^xk =k ) 0 k jx^_ s lnjx^_ s jj ds Rt and 0 jx^_ s lnjx^_ s jj ds < 1 since It (^x) < 1.

© 2001 by Chapman & Hall/CRC

232

Maxingales

Finally, if x^ s = 0 for all s 2 R+ , then we let x^ ks = s=k for s 2 [0; 1=k] and x^ ks = 1=k for s 1=k.

Remark 2.8.30. This theorem illustrates the general feature that

if the function ^ is chosen as in part 2 of Lemma 2.8.20, then one needs to impose Lipshitz continuity conditions on the coeÆcients.

Our purpose now is to state uniqueness results for more general functions gs (; x). In the next lemma, given a closed convex set F Rm , we denote as projF the projection of 2 Rm onto F ; for a closed convex cone N 2 Rd , we denote as a N the aÆne hull of N ; N ? = f 2 Rd : y 0 for all y 2 N g denotes the polar cone of N . As above, ri N denotes the relative interior of N .

Lemma 2.8.31. Let conditions I and II hold, and gs(; x) be dif-d ferentiable in . Let there exist a closed convex cone N R such that gs (; x) is strictly convex in 2 a N , gs (; x) gs (proja N ; x) ; 2 Rd ; s 2 R+ ; x 2 C ; and the following holds: 1. for every t 2 R+ and compact K C

g (; x) = 1; lim inf inf s 2 a N : st x2K jprojN j jprojN j!1

2. for every t 2 R+ , compact K C and A 2 R+

inf

inf inf gs (; x) >

2 a N : st x2K jprojN jA

1;

3. for every t 2 R+ , A 2 R+ , x 2 C , and sequence xk ! x

lim

k!1

Zt

0

sup

2 a N : jprojN jA

jgs (; xk ) gs(; x)j ds = 0;

4. if N ? 6= 0, then for every t 2 R+ and x 2 C

lim sup sup 2N ? : st jj!1

© 2001 by Chapman & Hall/CRC

gs (; x) jj

0:

233

Maxingale problems

Then (pt 1 Æ pt x) = x;t(x); t 2 R+ ; for every x such that x0 = x, x_ s 2 N (a.e.) and sups2R+ jx_ s j < 1. Also, if x(x) > 0, then x_ s 2 N (a.e.) Proof. We apply Theorem 2.8.27. Denote as N 0 the polar cone of N relative to a N , i.e., N 0 = N ? \ a N and let for m 2 N

Nm = Gm =

fy 2 a N : y m1 jjjyj for all 2 N 0g; fy 2 Nm : jyj mg:

We next de ne (s; x; y) in the statement of Theorem 2.8.27. If y 2= ri N , we set (s; x; y) = 0. Let y 2 ri N (the latter set is nonempty since N is convex, von Leichtweiss [131]). Since ri N = [1 m=1 Gm , we have that y 2 Gm for some m. We now prove that for every compact K C and t 2 R+ there exists C0 > 0 such that the inequality y gs (; x) 0 holds for x 2 K , s t and 2 a N such that jj > C0 . If jyj = 0, then a N = N , so = projN and in view of condition 1 y gs (; x) is negative for x 2 K and s t if jj is large enough. Let us assume now that jyj > 0. We rst consider the case when 1 2 a N is such that y jjjyj. Then for s t 2m 1 y gs (; x) jjjyj inf uinft gu(; x) 2m 2 a N

which by conditions 1 and 2 of the lemma is negative for all x 2 K and s t if jj is large enough. Now, let 2 a N be such that 1 y jjjyj. Then 2m

jprojN j 1 : jj 2(m + 1)

(2.8.45)

To see this, write = 1 + 2 , where 1 = projN 0 and 2 = projN . 1 Since y 2 Gm Nm and 1 2 N 0 , it follows that 1 y j jjyj; m 1 1 on the other hand, y jjjyj, so 2m j2 jjyj 2 y = y 1 y jmyj j1 j j2j jmyj j2j j2 j :

© 2001 by Chapman & Hall/CRC

234

Maxingales

1 jj Hence, j2 j 1 + 2m (recall that jyj > 0) which is equivalent m to (2.8.45) by the de nition of 2 . Inequality (2.8.45) and condition 1 imply that, given L > 0, we have, if jj is large enough, that for x 2 K and s t

y gs (; x) jjjyj

L jj; 2(m + 1)

which is negative if L has been chosen large enough. Thus, in all the cases we have that y gs (; x) 0 for all x 2 K and s t if 2 a N is such that jj is large enough. The claim is proved. We thus conclude, since y = proja N y and by hypotheses gs (; x) gs (proja N ; x); that for x 2 K and s t sup ( y

2Rd

gs (; x)) = sup ( y gs (; x)) 2 a N = sup ( y 2 a N : jjC0

gs (; x)); s t: (2.8.46)

Since the function y gs (; x) is strictly concave in 2 a N by the hypotheses, it attains supremum on f 2 a N : jj C0 g at a unique point, which we take as (s; x; y). Thus, we have de ned (s; x; y) for all y 2 Rd . We check that it satis es the conditions of Theorem 2.8.27. By de nition (s; x; y)y gs ((s; x; y); x) = sup (y gs (; x)) 2Rd for s 2 R+ ; y 2 [1 m=1 Gm ; x 2 C , which implies (2.8.29). Equality (2.8.46) shows that (s; x; y) is bounded on the sets [0; t] K Gm , where t 2 R+ , K is compact, and m 2 N . Also it is B[0; t] Ct

B(Rd )=B(Rd ) {measurable when restricted to [0; t]C R d for t 2 R+ since gs (; x) is C-progressively measurable in (s; x) and continuous in , and for 2 B(Rd )

f(s; x; y) : (s; x; y) 2 g = f(s; x; y) : sup (y gs (; x)) = sup (y gs (; x))g; d 2

© 2001 by Chapman & Hall/CRC

\ a N

2R

235

Maxingale problems

if

63 0, and f(s; x; y) : (s; x; y) 2 g = f(s; x; y) : sup ( y gs (; x)) = sup ( y gs (; x))g d 2

\ a N

[

2R

f(s; x; y) : y 2= ri N g;

if 3 0, where sup; = 1. (Note that by continuity of gs (; x) in the supremums may be taken over the rationals.) Finally, (s; x; y) is continuous in x. Indeed, let xk ! x as k ! 1. If y 2= ri N , then (s; xk ; y) = (s; x; y) = 0. Let y 2 ri N . Hence, y 2 Gm for some m and, since the set fxk ; k 2 N g is relatively compact, the set f(s; xk ; y); k 2 N g is bounded; so there exist 0 2 a N and subsequence k0 such that (s; xk0 ; y) ! 0 as k0 ! 1. Since (s; xk ; y) y gs ((s; xk ; y); xk ) y gs (; xk ) for 2 Rd and gs (; x) is continuous in (; x), we conclude that 0 y gs (0 ; x) y gs (; x) so that 0 y gs (0 ; x) = sup2 a N (y gs (; x)): Since the point where the supremum is attained is unique, 0 = (s; x; y) proving that (s; xk ; y) ! (s; x; y) as k ! 1. Thus, existence of (s; x; y) in the statement of Theorem 2.8.27 is proved. Let x^ be such that

x^ 0 = x; x^_ s 2 N;

A = sup jx^_ s j < 1: s2R+

(2.8.47)

We prove that (pt 1Æpt x^ ) = x;t (^x); t 2 R+ . Let D be de ned as in Theorem 2.8.27. According to Theorem 2.8.27, the required equality will follow if we nd a sequence xk 2 D, which converges to x^ as k ! 1, and is such that lim

k!1

x;t(xk ) = x;t(^x); t 2 R+ :

(2.8.48)

Since ri N is nonempty, there exist y^ 2 N and r 2 (0; 1) such that jy^j = 1 and y^ rjj for all 2 N 0 . We observe that if y 2 N and jyj A, then, given k 2 N ; there exist k > 0 such that k ! 0 as k ! 1 and y + k y^ 2 Gk if k k0 = b(A + 1)=rc + 1. Indeed, since y^ rjj and y 0 if 2 N 0 , jy^j = 1 and jyj A, we have, for 2 N 0 and 0 < k < 1, that (y + k y^) k rjj k rjjjy + k y^j=(1+ A); so if k = (1+ A)=(rk), then y + k y^ 2 Gk for k k0 .

© 2001 by Chapman & Hall/CRC

236

Maxingales

We de ne next xks = x^ s + k y^s; s 2 R+ : Since x^_ s 2 N; we have by (2.8.47) that x_ ks 2 Gk , so xk 2 D by (2.8.30). By Lemma 2.7.12, (2.8.6), and Remark 2.8.25, for (2.8.48) we need to prove that lim sup k!1

Zt

0

x

sup ( _ ks 2Rd

gs(; x

k ))ds

Zt

0

sup (x^_ s gs(; x^ ))ds: 2Rd

Since for y 2 ri N the left equality in (2.8.46) holds and x_ ks 2 Gk ri N , it suÆces to prove that lim sup k!1

Zt

0

sup ( x_ ks 2 a N

Zt

0

gs (; xk )) ds sup ( x^_ s 2 a N

gs (; x^ )) ds: (2.8.49)

We denote ks = (s; xk ; x_ ks ) and observe that ks 2 a N . Since x_ ks 2 N , it follows that ks x_ ks projN ks x_ ks , so, by the de nition of (s; x; y) 0 sup ( x_ ks 2 a N

gs (; xk )) jprojN ks jjx_ ks j gs (ks ; xk );

which implies by the facts that x_ ks is bounded (see (2.8.47)), ks a N and condition 1 holds that

B = sup sup jprojN ks j < 1: kk0 st

2

(2.8.50)

Next, by the de nitions of ks and xks sup ( x_ ks gs (; xk )) 2 a N = (ks x^_ s gs (ks ; x^ )) + k ks y^ (gs (ks ; xk ) gs (ks ; x^ )): (2.8.51) Since jy^j = 1 and y^ 2 N , we have that ks y^ jprojN ks j B: Also, by (2.8.50) for k k0 and s t

jgs (ks ; xk ) gs(ks ; x^ )j

© 2001 by Chapman & Hall/CRC

sup jgs (; xk ) gs (; x^ )j: 2 a N : jprojN jB

237

Maxingale problems

Therefore, by (2.8.51) and (2.8.50) for k k0 and s t sup ( x_ ks

2 a N

gs (; xk )) sup ( x^_ s gs (; x^ )) + Bk 2 a N

+

sup jgs (; xk ) 2 a N : jprojN jB

gs (; x^ )j

which implies (2.8.49) by the convergence k ! 0 and condition 3, nishing the proof of (2.8.48). We consider now the second assertion of the lemma. There is something to prove only if N 6= Rd . If y 2 Rd nN , then y "jjjyj for some 2 N ? and " > 0, so by condition 4 of the lemma, for every x 2 C , lim sup 2N ? : jj!1

y gs (; x) jj

"jyj;

and therefore sup2Rd (y gs (; x)) = 1. By Lemma 2.7.12, (2.7.6) and (2.7.8) this implies that x_ s 2 N if x (x) > 0. We now apply Theorem 2.8.27 and Lemma 2.8.31 to a proof of the following uniqueness result.

Theorem 2.8.32. Let conditions I anddII hold, and gs(; x) be differentiable and strictly convex in lowing hold:

2R

. Let, in addition, the fol-

1. for every t 2 R+ and compact K C

lim inf inf jj!1 st x2K

gs (; x) jj = 1;

2. for every t 2 R+ , compact K C and A 2 R+

inf inf inf g (; x) > jjA st x2K s

1;

x 2 C there exists l > 1 such that g (l; x) lim inf inf s > 1; jj!1 st lgs (; x)

3. for every t 2 R+ and

© 2001 by Chapman & Hall/CRC

238

Maxingales

4. for every t 2 R+ , compact K C and x 2 K there exist > 0 and > 0 such that

lim inf inf jj!1 st

gs (; x0 ) inf > 0: x0 2K : 0 gs (; x) sup jxr xr j rt

Then uniqueness holds for the maxingale problem (x; G) with = x. Proof. We rst observe that conditions 1 and 2 of the theorem imply that for every t 2 R+ and compact K C

inf inf inf gs (; x) >

2Rd st x2K

1:

(2.8.52)

It is easy to see that all the conditions of Lemma 2.8.31 hold with N = Rd . (In particular, condition 3 follows by conditions I and II.) Hence, by Lemma 2.8.31 (pt 1 Æpt x) = x;t (x); t 2 R+ ; if, in addition, x0 = x and sups2R+ jx_ s j < 1. The proof of Lemma 2.8.31 also shows that there exists function (s; x; y) satisfying the conditions of Theorem 2.8.27 with Gm = fy 2 Rd : jyj mg so that by Theorem 2.8.27 it suÆces to prove that the set D = fx : sups2R+ jx_ s j < 1; x0 = xg is x{dense in C . Assuming with no loss of generality that x = 0, we x x^ 2 C such that 0 (^x) > 0 and look for xk 2 D such that xk ! x^ as k ! 1 and lim It (xk ) = It (^x); t 2 R+ :

k!1

(2.8.53)

We de ne

x

k s

=

Zs

0

x^_ u 1(jx^_ u j k) du:

(2.8.54)

R The convergence xk ! x^ follows by the fact that 0t jx^_ u j du < 1 and Lebesgue's dominated convergence theorem. By Remark 2.8.25 for (2.8.53) it suÆces to show that

lim sup It (xk ) It (^x); t 2 R+ ; k!1

© 2001 by Chapman & Hall/CRC

(2.8.55)

239

Maxingale problems

where by Lemma 2.7.12 and (2.8.54)

It(^x)

=

It (xk )

=

Zt

0 Zt 0

sup ( x^_ s

2Rd

gs (; x^ )) ds;

sup ( x^_ s 1(jx^_ s j k) gs (; xk )) ds:

2Rd

Noting that by (2.8.54)

It (x

k) =

Zt

0

1(jx^_ sj > k) supd ( 2R

+

Zt

0

gs (; xk )) ds

1(jx^_ s j k) supd ( x^_ s 2R

gs (; xk )) ds

and that condition 1 of the lemma and conditions I and II easily yield the convergence lim

k!1

Zt

0

1(jx^_ s j > k) supd ( 2R

gs (; xk )) ds = 0;

we have that (2.8.55) would follow by Zt

Zt

lim sup sup (x^_ s gs (; xk )) ds sup (x^_ s gs (; x^ )) ds: k!1 2Rd 2Rd 0 0 (2.8.56) By the de nition of (s; x; y) (s; x^ ; x^_ s ) x^_ s gs ((s; x^ ; x^_ s ); x^ ) = sup ( x^_ s gs (; x^ )); (2.8.57a) 2Rd (s; xk ; x^_ s ) x^_ s gs ((s; xk ; x^_ s ); xk ) = sup ( x^_ s gs (; xk )): (2.8.57b) 2Rd We denote ^ s = (s; x^ ; x^_ s ); ^ ks = (s; xk ; x^_ s ) 1(j(s; xk ; x^_ s )j ak ); (2.8.58)

© 2001 by Chapman & Hall/CRC

240

Maxingales

where ak " 1 are chosen so that Zt

lim sup sup ( x^_ s k!1 2Rd 0

gs (; xk )) ds Zt

= lim sup (^ ks x^_ s k!1 0

gs (^ ks ; xk )) ds:

Then (2.8.56) and the lemma would be proved if Zt

lim sup (^ ks x^_ s k!1

gs (^ ks ; xk )) ds

0

Zt

(^s x^_ s gs(^s; x^ )) ds: (2.8.59) 0

We rst note that by continuity of (s; x; y) in x, the convergence xk ! x^ and (2.8.58) lim ^ k k!1 s

= ^ s ; s 2 R+ :

(2.8.60)

We now show that for some > 0 sup k

Zt

0

jgs (^ks ; x^ )j ds < 1:

(2.8.61)

By Young's inequality for every > 0 Zt

0

1 ^ ks x^_ s ds

Zt

0

1 gs (^ ks ; x^ ) ds + It (^x)

(the integrals are well de ned since ^ ks is bounded). Thus, since ^ ks x^_ s gs (^ ks ; xk ) 0 (see (2.8.57a) and (2.8.58)), we have that Zt

0

gs (^ ks ; xk ) ds

© 2001 by Chapman & Hall/CRC

1

Zt

0

1 gs (^ ks ; x^ ) ds+ It (^x):

(2.8.62)

241

Maxingale problems

Condition 3 of the lemma implies that, given t 2 R+ and x 2 C , there exist l > 1 and " > 0 such that gs (l; x) (1 + ")lgs (; x) for all jj large enough and s t (by condition 1 of the lemma gs (; x) is nonnegative if jj is large enough). Hence, gs (lp; x) (1 + ")p lp gs (; x) for arbitrary p 2 N so that for arbitrary M > 1 there exists L > 1 such that for all jj large enough

gs (; x) ML gs ; x ; s t: L

(2.8.63)

Combining this with condition 4 and recalling condition 1, we conclude that there exist Æ > 0; > 0 and k0 2 N such that for arbitrary M > 1 there exist L > 1 and A > 0 for which gs (; xk ) Ægs (; x^ ) ÆML gs ; x^ ; s t; L for all k k0 and jj A. Choosing now = =L and M = 2=(Æ), we have that gs (; xk ) 2gs (; x^ )=; s t; when k k0 and jj A. Hence, for k k0 Zt

0

x

gs (^ ks ; k ) ds Zt

0

Zt

0

inf g (; x jjA s

2 inf gs (; xk ) ds + jjA

k ) ds +

Zt

0

Zt

0

gs (^ ks ; x^ ) ds

gs (^ ks ; xk ) 1(j^ ks j > A) ds 2

Zt

0

sup gs (; x^ ) ds; jjA

and (2.8.62) yields for k k0 after a simple algebra Zt

0

gs (^ ks ; x^ ) ds It (^x) + 2

Zt

0

sup gs (; x^ ) ds jjA

Zt

0

inf g (; xk ) ds; jjA s

which yields, since It (^x) < 1 and conditions I and II hold, sup k

Zt

0

gs (^ ks ; x^ ) ds < 1:

© 2001 by Chapman & Hall/CRC

242

Maxingales

Inequality (2.8.61) now follows since the functions gs (^ ks ; x^ ) are bounded from below uniformly in s 2 [0; t] and k in view of (2.8.52). Inequalities (2.8.61) and (2.8.62) yield, since by (2.8.52) the functions gs (^ ks ; xk ) are bounded from below uniformly in s t and k, sup k

Zt

0

jgs (^ks ; xk )j ds < 1:

(2.8.64)

Since condition I, the convergences xk ! x^ and (2.8.60) imply that gs (^ ks ; xkR) ! gs (^ s ; x^ ); s t; as k ! 1, by Fatou's lemma and (2.8.64) 0t jgs (^ s ; x^ )j ds < 1; hence, by Lemma 2.7.12 Zt

0

(^ s x^_ s )_0 ds It (^x)+

Zt

0

jgs(^s ; x^ )j ds < 1:

(2.8.65)

On the other hand, since by (2.8.57a) and (2.8.58) ^ s x^ s gs (^ s ; x^ ) 0 and by (2.8.52) the function gs (^ s ; x^ ) is bounded from below on [0; t], we conclude that ^ s x^ s is bounded from below on [0; t]. This fact and (2.8.65) show that Zt

0

j^s x^_ sj ds < 1;

(2.8.66)

R in particular, 0t ^ s x^_ s ds is well de ned and nite. Since the convergence gs (^ ks ; xk ) ! gs (^ s ; x^ ); s t; and uniform boundedness from below of the functions gs (^ ks ; xk ); s 2 [0; t]; k 2 N ; also imply by Fatou's lemma that

lim inf k!1

Zt

0

gs (^ ks ; xk ) ds

Zt

0

gs (^ s ; x^ ) ds;

we conclude that (2.8.59) would follow by lim sup k!1

Zt

0

x

^ ks ^_ s ds

Zt

0

^ s x^_ s ds:

To prove the latter, note that by La Vallee Poussin's theorem the sequence f(^ ks ; s 2 [0; t]); k 2 N g is uniformly integrable with respect

© 2001 by Chapman & Hall/CRC

243

Maxingale problems

to Lebesgue measure in view of (2.8.61) and condition 1 of the lemma. This fact and convergence (2.8.60) yield for arbitrary m > 0 lim

k!1

Zt

x 1(jx^_ s j m) ds =

^ ks ^_ s

0

Zt

0

^ s x^_ s 1(jx^_ s j m) ds:

In view of (2.8.66), we thus complete the proof by showing that lim sup lim sup m!1 k!1

Zt

^ ks x^_ s 1(jx^_ s j > m) ds 0:

0

(2.8.67)

Given arbitrary " > 0, by (2.8.61) we can choose M1 > 0 such that 1 sup M1 k

Zt

0

jgs(^ks ; x^ )j ds ":

(2.8.68)

By (2.8.63) we can choose A1 > 0 and L1 > 0 such that gs (; x^ ) M1 L1 gs (=L1 ; x^ ) for s 2 [0; t] and jj > A1 . Young's inequality then yields Zt

0

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds

L1 +

Zt

0

hZt

0

gs

^k

sup ( x^_ s 2Rd

M1 1

Zt

0

x 1(jx^_ s j > m) 1(j^ks j > A1 ) ds

s;^ L1

i

gs (; x^ )) 1(jx^_ s j > m) ds

jgs (^ks ; x^ )j ds + L1

Zt

0

sup ( x^_ s

2Rd

gs (; x^ )) 1(jx^_ s j > m) ds:

Inequality (2.8.68) and niteness of It (^x) then imply that lim sup lim sup m!1 k!1

Zt

0

© 2001 by Chapman & Hall/CRC

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds ":

244

Maxingales

Since Zt

0

Zt

x 1(jx^_ s j > m) ds A1 jx^_ sj 1(jx^_ s j > m) ds

^ ks ^_ s

+

Zt

0

and

Rt 0

0

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds;

jx^_ sj ds < 1, the proof of (2.8.67) is over.

The following uniqueness result, which is stated in terms of bs (x) and g^s (; x), is a direct consequence of Theorem 2.8.32.

Theorem 2.8.33. Let conditions I and II hold, and g^s(; x) be differentiable and strictly convex in . Let the following hold: 1. for every t 2 R+ and compact K C

lim inf inf jj!1 st x2K

g^s (; x) jj = 1;

2. for every t 2 R+ and compact K C

sup sup jbs (x)j < 1; st x2K

x 2 C there exists l > 1 such that g^ (l; x) lim inf inf s > 1; jj!1 st lg^s (; x) 4. for every t 2 R+ , compact K C and x 2 K there exist > 0 3. for every t 2 R+ and

and > 0 such that

lim inf inf jj!1 st

g^s (; x0 ) inf > 0: g^s (; x) x0 2K : 0 sup jxr xr j rt

Then = x .

As another consequence, we have the following.

© 2001 by Chapman & Hall/CRC

245

Maxingale problems

Theorem 2.8.34.

Let be a deviability on C under which the canonical idempotent process X is a semimaxingale starting at x with local characteristics (b; c; ; 0). Let the following conditions hold: 1. the functions Rbs (x) and cs (x) are continuous in x 2 C and the function Rd (ex 1 x) s (dx; x) is continuous in (; x) 2 Rd C ; 2. for every t 2 R+ , A > 0 and compact K C

sup sup jbs (x)j < 1; sup sup kcs (x)k < 1; st x2K st x2K Z sup sup (eAjxj 1 Ajxj) s (dx; x) < 1 st x2K d R

and

lim sup sup !0 st x2K

Z Rd

(ex 1 x) s (dx; x) = 0;

3. for every t 2 R+ and compact K C there exists B 2 R+ such that a)

Z

1( x > B ) s(dx; x) > 0;

inf inf inf 2Rd : st x2K Rd jj=1

b) for every x 2 K there exist > 0 and > 0 such that

lim inf inf inf v!1 2Rd : st jj=1 Z

Rd

Z

Rd

Then = x .

© 2001 by Chapman & Hall/CRC

inf

x0 2K : sup jxr x0r j rt

exp(v x) 1( x > B ) s (dx; x0 ) exp(v x) 1( x > B ) s(dx; x)

> 0:

246

Maxingales

Proof. We check the conditions of Theorem 2.8.33. Conditions I and II are met by Lemma 2.8.16. Condition 2 of Theorem 2.8.33 holds by hypotheses. Condition 1 of Theorem 2.8.33 follows since g^s (; x) grows at least exponentially fast in , which results from the inequalities

g^s (; x)

Z

1 x) s (dx; x)

(ex

Rd

(ejjB 1 jjB )

Z

Rd

1( x > B jj) s(dx; x)

and condition 3a). Let us consider condition 3 of Theorem 2.8.33. We have that Z

1 g^s (; x) = cs (x) + 2

Rd

Z

(ex

1 x) s (dx; x)

12 cs(x) + 13 (e2x 1 2 x) 1( x > 0) s(dx; x) 1 + 2

Z Rd

Rd

( x)2 1( x 0) s (dx; x)

1 1 g^ (2; x) + 3 s 2

Z Rd

( x)2 s (dx; x):

Therefore, by the fact that g^s (; x) grows at least exponentially fast in and conditions 2 and 3a) of the theorem

g^ (2; x) lim inf inf s jj!1 st g^s (; x)

3;

verifying condition 3 of Theorem 2.8.33. Finally, condition 4 of Theorem 2.8.33 follows by the inequalities

g^s (; x)

Z Rd

ex=2 1( x > B jj) s (dx; x)

© 2001 by Chapman & Hall/CRC

247

Maxingale problems

when jj is large enough, and 1 1 () cs (x)() + 2 2

g^s (; x) 1 + eBjj 2

Z

(x)2 s(dx; x)+

Rd

Z

Rd

Z Rd

( x)2 s (dx; x)

ex 1(x > B jj) s (dx; x);

and conditions 2 and 3.

Remark 2.8.35. Condition 3b) holds if condition 3a) holds and R ln Rd eujxj 1(jxj > B )s (dx; x) lim sup sup sup < 1: u

u!1 st x2K

Theorems 2.8.21, 2.8.28, 2.8.32, 2.8.33, and 2.8.34 are concerned with \the nondegenerate case" singled out by condition 1 of Theorem 2.8.33. We now consider another degenerate case along with Theorem 2.8.29, which takes advantage of the generality of Lemma 2.8.31.

Theorem 2.8.36.

Let be a deviability on C under which the canonical process X is a semimaxingale starting at x with local characteristics (b; 0; ; 0) such that for some l 2 N and vi 2 Rd

bs(x) =

l X i=1

b(si) (x)vi ;

where R+ -valued functions progressively measurable. Let also for every t 2 R+

x) > 0;

inf inf b(i) ( st x2K s

Then = x .

s ( ; x) =

l X i=1

1(vi 2

)b(si) (x);

b(si) (x) are continuous in

x

and

C-

and compact K C

sup sup b(si) (x) < 1; 1 i l: st x2K

Proof. Let N denote the smallest closed convex cone containing P v1 ; : : : ; vl . Noting that gs (; x) = li=1 exp( vi ) 1 b(si) (x), one can see that the hypotheses of Lemma 2.8.31 are satis ed. Speci cally, the following stronger versions of conditions 1 and 2 of Lemma 2.8.31 hold:

© 2001 by Chapman & Hall/CRC

248

Maxingales

10 : for every t 2 R+ and compact K C

g (; x) limd inf inf s = 1; s t x 2 K j projN j 2R : jprojN j!1

20 : for every t 2 R+ and compact K C inf inf inf gs (; x) >

2Rd st x2K

1:

Property 20 is obvious, property 10 follows by the inequality

jprojN j c i=1 max ( vi ) _ 0; ;:::;l

(2.8.69)

where c is a constant depending only on v1 ; v2 ; : : : ; vl . By Lemma 2.8.31 (pt 1 Æ pt x) = x;t (x) when x0 = x, x_ s 2 N and sups2R+ jx_ s j < 1. Therefore, in analogy with the proof of Theorem 2.8.32 it suÆces to show that the set D = fx 2 C : x_ s 2 N; sups2R+ jx_ s j < 1; x0 = xg is x {dense in C . Let x^ 2 C be such that x (^x) > 0 and xk be de ned by (2.8.54). By Lemma 2.8.31 x^_ s 2 N , so xk 2 D. The0 argument of the proof of Theorem 2.8.32 with the use of property 2 implies that it suÆces to establish (2.8.56) for t 2 R+ . Since x^_ s 2 N , we have that x^_ s projN x^_ s for 2 Rd , which implies by properties 10 and 20 that sup2Rd ( x^_ s gs (; xk )) sup2Rd (projN x^_ s gs (; xk )) < 1, k 2 N . Therefore, by a measurable-selection theorem there exist Lebesgue measurable functions (~ ks ; s 2 R+ ) such that "

~ ks x^_ s gs (~ ks ; xk ) sup ( x^_ s gs (; xk )) 2Rd

1 k

#

_0:

Then, for suitable ak > 0, the functions ^ ks = ~ ks 1(j~ ks j bounded and satisfy the equality Zt

lim sup sup ( x^_ s k!1 2Rd 0

gs (; xk )) ds Zt

= lim sup (^ ks x^_ s k!1 0

© 2001 by Chapman & Hall/CRC

ak ) are

gs (^ ks ; xk )) ds

249

Maxingale problems

so (2.8.56) and the theorem are proved if Zt

lim sup (^ ks x^_ s gs (^ ks ; xk )) ds k!1 0

Zt

0

sup (x^_ s gs (; x^ )) ds:

2Rd

Next, the hypotheses on b(si) (x) imply that, given arbitrary " 2 (0; 1), we have, for k large enough, (1 ")

Zt

0

x

b(si) ( k ) ds

Zt

0

x

b(si) (^ ) ds

Zt

(1+") b(si) (xk ) ds: 0

Hence, for arbitrary 2 Rd Zt

(1 + ")

0

gs (; x

k ) ds

l X i=1

evi =

Zt

Zt

0

0

x

b(si) (^ ) ds

gs (; x^ ) ds

l 1+"X 1 " i=1

2"

Zt

0 Z l t X

1 " i=1

0

b(si) (^x) ds b(si) (^x) ds:

Therefore, Zt

0

(^ ks x^_ s

gs (^ ks ; xk )) ds +"

Zt

0

Zt

0

sup ( x^_ s 2Rd

gs (; x^ )) ds

l 2" X gs (^ ks ; xk ) ds + 1 " i=1

Zt

0

b(si) (^x) ds;

and, since " can be taken arbitrarily small, the proof is complete if sup k

Zt

0

jgs (^ks ; xk )j ds < 1:

(2.8.70)

The proof of the latter inequality is similar to that of (2.8.64) in the proof of Theorem 2.8.32. More speci cally, it is not diÆcult to check that under the assumptions of the theorem the following holds:

© 2001 by Chapman & Hall/CRC

250

Maxingales

1. for every t 2 R+ , compact K C and A 2 R+ , Zt

0

sup

x2K

sup

2Rd : jprojN jA

jgs(; x)j ds < 1;

2. for every t 2 R+ and x 2 C , there exists l > 1 such that

g (l; x) lim inf inf s > 1; d s t lgs (; x) 2R : jprojN j!1

3. for every t 2 R+ , compact K C and x 2 K , there exist > 0 and > 0 such that gs (; x0 ) > 0: lim inf inf inf st x0 2K : 0 gs (; x) 2Rd : sup jxr xr j jprojN j!1 rt Note that part 2 follows from (2.8.69), the other two properties being obvious. These conditions, along with properties 10 and 20 above, imply (2.8.70) in the same way as in the proof of Theorem 2.8.32 conditions I and II together with conditions 1{4 of Theorem 2.8.32 implied (2.8.64) with jprojN j playing the role of jj.

© 2001 by Chapman & Hall/CRC

Part II

Large Deviation Convergence of Semimartingales

251 © 2001 by Chapman & Hall/CRC

Chapter 3

Large deviation convergence This chapter contains basic facts on large deviation convergence in Tihonov spaces and their adaptation to the setting of the Skorohod space.

3.1 Large deviation convergence in Tihonov spaces In this section we develop the theory of large deviation convergence in Tihonov spaces. Our exposition is along the lines of the content of Section 1.9. Let E be a topological space with Borel -algebra B(E ). Let be a directed set, fP ; 2 g be a net of probability measures on (E; B(E )), and fr ; 2 g be a net of real numbers greater than 1 converging to 1 as 2 . We recall that Cb+(E ), C +b (E ), and C +b (E ) denote the respective sets of R+ -valued bounded continuous functions on E , R+ -valued bounded upper semi-continuous functions on E , and R+ -valued bounded lower semi-continuous functions on E . Let be an F -idempotent probability on E , where F denotes the collection of closed subsets of E .

De nition 3.1.1. We say that the net fP ; 2 g large deviation 253 © 2001 by Chapman & Hall/CRC

254

Large deviation convergence

(LD) converges at rate r to if for every h 2 Cb+(E ) Z

lim 2

1=r

h(z )r dP (z )

=

E

_

E

h(z ) d(z ):

(3.1.1)

Remark 3.1.2. One could also consider the version of the above

de nition where h ranges in the set of R+ -valued bounded continuous functions on E of compact support. Then the de nition we have given would refer to \weak large deviation convergence", while the case of compactly supported h would specify \vague large deviation convergence". Since our focus is on \weak large deviation convergence", we simply call it \large deviation convergence".

Note that if E is a Tihonov topological space, then according to Theorem 1.7.27 the F -idempotent probability is uniquely speci ed by the right-hand sides of (3.1.1). We generally denote the large deviation convergence by P rld! . Since the net r is xed in the

rest of the chapter, we simplify the notation by writing P ld ! . We R 1=r 1=r 1=r denote P (A) = P (A) and kf k = E f (z )r dP (z ) , where f : E ! R+ . We state a Portmanteau theorem for large deviation convergence.

Theorem 3.1.3. Let E be a Tihonov topological space. The following conditions are equivalent.

ld ! : _ 2: (i) lim inf kgk g d

1:

P

(ii) lim sup kf k

E _ E

for all g 2 C +b (E );

f d for all f 2 C +b (E ):

20 : The inequalities of part 2 hold for all lower semi-continuous relative to , bounded Borel-measurable functions g : E ! R+ and all upper semi-continuous relative to , bounded Borel-measurable functions f : E ! R+ , respectively. 3: (i)

lim inf P1=r (G) (G)

(ii) lim sup P1=r (F ) (F )

© 2001 by Chapman & Hall/CRC

for all open sets G E; for all closed sets F

E:

255

LD convergence in Tihonov spaces

30 : The inequalities of part 3 hold for all open relative to Borelmeasurable sets G and closed relative to Borel-measurable sets F , respectively. 4: lim P1=r (H ) = (H ) for all continuous relative to Borel-measurable sets H E: 5:

lim khk =

_

6:

lim khk =

_

E

h d

for all continuous relative to bounded Borel-measurable functions h : E ! R+ :

h d

for all bounded Borel-measurable E functions h : E ! R+ that are uniformly continuous with respect to a given uniformity on E . Proof. The proof is almost identical to the one of Theorem 1.9.2. We give it here to make the reading easier. Clearly, 1 ) 6, 2 , 20 , 2 ) 1, 2 ) 3, 2 ) 5, 3 , 30 , 30 ) 4, and 5 ) 1. We prove the implication 1 ) 3. To prove 1 ) 3(i), we note that, since E is Tihonov and G is open, 1(G) = sup h over h 2 Cb+W(E ) such that h 1(G). Therefore, by Theorem 1.4.4 (G) = suph E h d, W so that if h 1(G) is such that (G) E h" d + ", then

lim inf P1=r (G) lim kh" k =

_

E

h" d (G) ":

The proof of 3(ii) is analogous if we note that 1F = inf h over h2W Cb+ (E ) such that h 1(F ) so that by Theorem 1.4.19 (F ) = inf h E h d. We prove that 3(i) ) 2(i) and 3(ii) ) 2(ii). For g 2 C +b (E ) such that kgk = 1 let hi i i gk (z ) = max 1 g(z ) > ; k 2 N: i=0;:::;k 1 k k Since the sets fz : g(z ) > xg are open by the lower semi-continuity of g, 3(i) yields lim inf kgk k max lim inf

i=0max ;:::;k

P1=r g(z ) >

i i k

k _ i i _ g(z ) > = gk d g d k k E E

i=0;:::;k 1 hi

1

hi

© 2001 by Chapman & Hall/CRC

1 : k

256

Large deviation convergence

The proof of 3(ii) ) 2(ii) is similar if we consider fk (z ) = maxi=0;:::;k 1 (i + 1)=k 1(f (z ) i=k) : Now we prove 4 ) 3. Let G be open andWÆ > 0. Let h be a function from Cb+ (E ) such that h 1(G) and E h d (G) Æ. Let Hu = fz 2 E : h(z ) ug; u 2 [0; 1]: Then the function (Hu ) increases as u #W0. Therefore, it has at most countably many jumps. Also (Hu ) E h d u, so (Hu ) (G) 2Æ for u small enough. Thus, there exists " > 0 such that (H" ) (G) 2Æ and (Hu ) is continuous at ". By -maxitivity of the latter is equivalent to H" being continuous relative to , so we conclude that lim inf P1=r (G) lim P1=r (H") = (H" ) (G) 2Æ:

The proof of 3(ii) is similar. We prove that 6 ) 3(ii). Let V be a uniformity on E and F be a closed subset of E . Let f g be a collection of uniformly continuous with respect to V pseudo-metrics on E , which is closed under the formation of maximums and such that 1(F ) = inf ">0 inf (1 (z; F )=")+ . (As above, (z; F ) = inf z0 2F (z; z 0 ).) The functions (1 (z; F )=")+ are bounded and uniformly continuous with respect to V so that by Theorem 1.7.7 +

lim sup P1=r (F ) inf inf lim k 1 (z; F )=" ">0

= inf inf ">0

_

E

k

(1 (z; F )=")+ d(z ) = (F ):

The implication 6 ) 3(i) is proved in an analogous manner.

Remark 3.1.4. As the proof shows, in part 6 it is enough to require

that the convergences hold for functions h that are Lipshitz continuous with respect to the pseudo-metrics specifying the uniformity.

Remark 3.1.5. Part 3 of the theorem can be used to de ne \narrow

large deviation convergence", which is identical to the de nition of the large deviation principle, see, e.g., Varadhan [128]. Thus, on Tihonov spaces large deviation convergence is equivalent to the large deviation principle.

We recall that Br (z ) denotes the closed r-ball about an element z of a metric space.

© 2001 by Chapman & Hall/CRC

LD convergence in Tihonov spaces

257

Corollary 3.1.6. Let E be a Tihonov topological space. If P !ld , then

(z ) = lim0 lim inf P1=r (U ) = lim0 lim sup P1=r (cl U ); U 2Uz 2 U 2Uz 2 where Uz0 is a collection of open neighbourhoods of z whose closures decrease to z . In particular, if E is a metric space, then

(z ) = lim lim sup P1=r (Br (z )): r!0 2 Proof. The claim follows by the inequalities

(z ) (U ) lim inf P1=r (U ) lim sup P1=r (cl U ) 2 2 (cl U); and the fact that limU 2Uz0 (cl U ) = (z ). The following fact follows by Theorem 3.1.3.

Corollary 3.1.7. Let E be a Tihonov topological space. Let a Borel

subset E0 of E be equipped with relative topology. Let P (E n E0 ) = ~, (E n E0 ) = 0 and the restriction of to E0 , which is denoted by be -smooth relative to the collection of closed subsets of E0 . Then ld ld ~ P ! if and only if P~ ! , where P~ denotes the restriction of P to E0 .

Remark 3.1.8. The -smoothness property of ~ holds if either E0 is a closed subset of E or is a deviability on E .

Lemma 1.9.14 provides us with the following corollary.

Corollary 3.1.9. Let E be a Tihonov topological space and P !ld , where is supported by E0 E . Then the following holds. 1.

lim khk =

_

E

h d

for all E0 -continuous bounded Borel-measurable functions h : E ! R+ ;

© 2001 by Chapman & Hall/CRC

258

Large deviation convergence

2. (i)

lim inf kgk

_

E

g d

for all E0 -lower-semi-continuous bounded Borel-measurable functions g : E ! R+ ; _ (ii) lim sup kf k f d

E

for all E0 -upper-semi-continuous bounded Borel-measurable functions f : E ! R+ ;

lim inf P1=r (G) (G) for all E0 -open Borel-measurable sets G E; (ii) lim sup P1=r (F ) (F )

3. (i)

for all E0 -closed Borel-measurable sets F 4.

E;

lim P1=r (H ) = (H ) for all E0 -continuous Borel-measurable sets H E:

The next corollary allows one to strengthen topology for which LD convergence can be proved. We say that two topologies on a topological space are locally equivalent at a given point if they have equivalent local bases at the point.

Corollary 3.1.10. Let O1 and O2 be Tihonov topologies on E , and

let O2 be ner than O1 . Let E0 E be such that O1 and O2 are ld locally equivalent at every z 2 E0 . If P ! for topology O1 , the P can be extended to probabilities on the Borel -algebra of E generated by O2 , is supported by E0 and the restriction of to E0 is -smooth relative to the collection of closed subsets of E0 for the topology induced on E0 by O1 , then is a -smooth idempotent probability relative to the collection of sets closed in topology O2 and ld P ! for topology O2 . Proof. Since the topologies induced on E0 by O1 and O2 coincide, (E n E0 ) = 0 and the restriction of to E0 is -smooth relative to the collection of closed subsets of E0 for the topology induced on E0 by O1 , is a -smooth idempotent probability relative to the collection of sets closed in topology O2 . The required LD convergence follows by the fact that if h : E ! R+ is continuous for topology O2 , then it is E0 -continuous for topology O1 .

© 2001 by Chapman & Hall/CRC

259

LD convergence in Tihonov spaces

Below, we will need convergence of integrals of not necessarily bounded functions. This requires an analogue of uniform integrability. De nition 3.1.11. A Borel-measurable function f : E ! R+ is said to be uniformly exponentially integrable (of order r ) with respect to the net fP g if Z

lim lim sup a!1 2

E

1=r

f (z )r 1(f (z ) > a) dP (z )

= 0:

In analogy to uniform integrability by Chebyshev's inequality the uniform exponential integrability holds if for some > 0 Z

lim sup

1=r

f (z )r (1+") dP (z )

< 1:

E

Lemma 3.1.12. Let E be Tihonov.

ld Let P ! as 2 and be supported by E0 E . Then the following holds. 1: For all E0 {continuous and uniformly exponentially integrable with respect to fP g Borel-measurable functions h : E ! R+ Z

lim 2

1=r

h(z )r dP (z )

=

E

_

h(z ) d(z ):

E

2. For all E0 {lower-semi-continuous Borel-measurable functions g : E ! R+

lim inf 2

Z

g(z )r dP (z )

1=r

_

g(z) d(z): E

E

Proof. The second part, being \a Fatou lemma for LD convergence", is proved by a similar means: for a 2 R+ by Corollary 3.1.9

lim inf 2

Z

E

lim inf 2

1=r

g(z )r dP (z ) Z

1=r

(g(z ) ^ a)r dP (z )

E

_

(g(z) ^ a) d(z): E

© 2001 by Chapman & Hall/CRC

260

Large deviation convergence

W

Since the latter converges to E g(z ) d(z ) as a ! 1, the proof of part 2 is over. Part 1 follows by part 2 and the inequalities Z

lim sup 2

E

1=r r h(z ) dP (z ) Z

lim sup 2

+ lim sup 2 _

E

Z

E

1=r

h(z )r 1(h(z ) a)dP (z )

1=r

h(z )r 1(h(z ) > a)dP (z )

h(z) 1(h(z) a) d(z) E

Z

+ lim sup 2

E

1=r

h(z )r 1(h(z ) > a)dP (z )

;

where the latter inequality holds by Corollary 3.1.9. The following lemma gives an extension in a dierent direction.

Lemma 3.1.13. Let E be a Tihonov topological space.

ld Let P ! . Let h : E ! R+ be uniformly bounded and Borel-measurable functions such that for a function h : E ! R+

lim h (z ) = h(z )

2

for -almost every z 2 E and every net z ! z as 2 . Then Z

lim 2

1=r

h (z )r dP (z )

E

=

_

E

h(z ) d(z ):

Proof. The proof is similar to the one of Lemma 1.10.2 so we only sketch it. De ning h (z ) = inf U 2Uz supz0 2U sup0 h (z 0 ), where as we recall Uz denotes the collection of open neighbourhoods of z , and h(z ) = inf 2 h (z ), we can write for arbitrary > 0 and suitable 0

© 2001 by Chapman & Hall/CRC

261

LD convergence in Tihonov spaces

by Theorem 3.1.3 applied to h0 that Z

lim sup 2

E

1=r

h (z )r dP (z ) Z

lim sup 2

E

1=r

h0 (z )r dP (z )

_

_

E

h0 (z ) d(z )

h(z ) d(z ) +

E

_

E

h(z ) d(z ) + :

The complementary inequality lim inf 2

Z

1=r

h (z )r dP (z )

_

h(z) d(z) E

E

is proved by a mirror argument. Speci cally, de ning h (z ) = supU 2Uz inf z0 2U inf 0 h (z 0 ) and h(z ) = sup2 h (z ), we have for arbitrary > 0 and suitable 1 lim inf 2

Z

E

lim inf 2

1=r

h (z )r dP (z ) Z

E

1=r

h1 (z )r dP (z )

_

E

_

h1 (z) d(z) E

h(z ) d(z )

_

E

h(z ) d(z )

:

As a consequence of Lemma 3.1.13, we obtain the following version of the contraction principle on preservation of LD convergence under mappings. Theorem 3.1.14. Let E be a Hausdor topological space and E 0 be a Tihonov topological space. Let be a deviability on E . Let Borelmeasurable functions f : E ! E 0 , 2 ; and a -Luzin-measurable function f : E ! E 0 be such that for -almost every z 2 E and ld every net z ! z we have that f (z ) ! f (z ). If P ! , then ld 1 1 P Æ f ! Æ f .

© 2001 by Chapman & Hall/CRC

262

Large deviation convergence

Proof. The proof is similar to the one of Theorem 1.10.3. We rst note that Æf 1 is a deviability on E 0 by Theorem 1.7.11. Next, for an R+ -valued bounded continuous function h on E 0 by a change of variables and Lemma 3.1.13

lim

2

Z

E0

1=r 1 0 r 0 h(z ) dP Æ f (z )

Z

= lim 2

E

h Æ f (z )r dP (z ) =

_

E

1=r

h Æ f (z ) d(z ) =

_

E0

h(z 0 ) d Æ f 1(z 0 ):

The following consequence is often used below.

Corollary 3.1.15. Let E be a Hausdor topological space, E 0 ldbe a

Tihonov topological space, and be a deviability on E . If P ! as 2 and f : E ! E 0 is Borel measurable and continuous -a.e., ld then P Æ f 1 ! Æf 1.

We now derive a criterion of \large deviation relative compactness" in the theme of Prohorov's one for weak convergence.

De nition 3.1.16. An F {idempotent probability on E is called

a large deviation (LD) accumulation point of fP ; 2 g (for rate r ) if there exists a subnet fP0 ; 0 2 0 g of fP ; 2 g that LD converges (at rate r0 ) to .

De nition 3.1.17. The net fP ; 2 g is called large deviation 0 0 (LD) relatively compact (for rate r ) if every subnet fP0 ; of fP ; 2 g has an LD accumulation point (for rate r0 ).

2 g

We recall that K denotes the collection of compact subsets of E .

De nition 3.1.18. The net fP ; 2 g is called exponentially tight (of order r ) if inf K 2K lim sup2 P1=r (K c ) = 0.

Theorem 3.1.19. Let E be a Tihonov topological space.

1. If the net fP ; 2 g is exponentially tight, then it is LD relatively compact, the accumulation points being deviabilities.

© 2001 by Chapman & Hall/CRC

LD convergence in Tihonov spaces

263

2. Let E be, in addition, a locally compact Hausdor topological space. If the net fP ; 2 g is LD relatively compact, then it is exponentially tight. Proof. The proof is analogous to the proof of Theorem 1.9.17. We start with part 1. Let Cb;+1 (E ) = ff 2 Cb+(E ) : kf k 1g. For a given 2 , the mapping V : f ! kf k ; f 2 Cb;+1 (E ); is an element + of the space [0; 1]Cb;1 (E ) . The latter space, endowed with product topology, is compact and Hausdor. Therefore, the net fV ; 2 g + is relatively compact on [0; 1]Cb;1 (E ) so that there exists a subnet + fV0 ; 0 2 0g that converges to an element V of [0; 1]Cb;1 (E) . We extend V to a functional on Cb+(E ) by letting V (c f ) = cV (f ); c 2 C + (E ) R + . By the de nition of topology on [0; 1] b;1

lim kf k 0 = V (f ); f 2 Cb+ (E ):

0 20

(3.1.2)

The latter implies that V satis es conditions (V 0), (V 1) and (V 2) of Theorem 1.7.25, i.e., (V 0) V (1) = 1, (V 1) V (c f ) = cV (f ); c 2 R+ ,

(V 2) V (f _ g) = V (f ) _ V (g). The rst two properties directly follow by (3.1.2). The third property is valid in view of the inequalities kf k _ kgk kf _ gk 21=r kf k _ kgk and (3.1.2). Also, exponential tightness of fP ; 2 g and (3.1.2) imply that V is tight in the sense of Theorem 1.7.25. Thus, the functional V satis es all the conditions of Theorem 1.7.25, so according to the W theorem V (f ) = E f d; f 2 Cb+(E ); for some deviability , which ld implies that P0 ! (at rate r0 ). This completes the proof of part 1. Part 2 follows since by the argument of the proof of (1.9.5), where we use large deviation relative compactness of fP g in place of weak relative compactness of f g, for arbitrary " > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that

k [

lim sup P1=r E n Ai 2 i=1

© 2001 by Chapman & Hall/CRC

":

264

Large deviation convergence

(Cf. also the proof of part 2 of Theorem 3.1.28 below.)

Corollary 3.1.20.

Let fP; ; 2 g, 2 , be nets of Borel measures on respective Tihonov topological spaces E . If the nets fP; ; 2 g are exponentially tight (of order r) for every 2 , then there exists a subnet f(P0 ; ; 2 ); 0 2 0 g of f(P; ; 2 ); 2 g such that the nets fP0 ; ; 0 2 0 g LD converge (at rate r0 ) to deviabilities on the E for every 2 . R

Proof. For f 2 Cb+(E ), let V; (f ) = E fr dP; 1=r . By Tihonov's theorem the set f(V; (f ); f 2 Cb;+1 (E ); 2 ); 2 + Q g is a relatively compact subset of 2 [0; 1]Cb;1 (E ) , where the latter set is equipped with product topology. Thus, there exists a convergent subnet f(V0 ; (f ); f 2 Cb;+1 (E ); 2 ); 0 2 0 g. Now the required follows by the argument of the proof of Theorem 3.1.19.

Theorem 3.1.19 allows us to introduce the following useful concept.

De nition 3.1.21. Let E be a Tihonov topological space and E0 E: We say that the net fP ; 2 g is E0 -exponentially tight if it is exponentially tight and every LD accumulation point is supported by E0 . The following is a version of the contraction principle.

Corollary 3.1.22. Let E be a Tihonov topological space, E0 E , 0

and E be a locally compact Hausdor topological space. If the net fP ; 2 g is E0 {exponentially tight and a function f : E ! E 0 is Borel measurable and E0 -continuous, then the net fP Æf 1 ; 2 g is exponentially tight. Proof. By part 1 of Theorem 3.1.19 the net fP ; 2 g is LD relatively compact. Since f is continuous a.e. with respect to every LD accumulation point of fP ; 2 g, by Corollary 3.1.15 the net fP Æf 1 ; 2 g is LD relatively compact as well; hence, it is exponentially tight by part 2 of Theorem 3.1.19.

We now assume that E is a metric space and introduce \metrics" for large deviation convergence. We rst de ne an analogue of the

© 2001 by Chapman & Hall/CRC

265

LD convergence in Tihonov spaces

Prohorov metric. We again assume as given a net fP ; 2 g of Borel measures on E , a net of real numbers fr ; 2 g greater than 1 converging to 1, and an F -idempotent probability on E . The analogue of the Prohorov metric is de ned by

pld (P ; ) = inf > 0 : P1=r (F ) F + ;

(F ) P1=r F + for all closed F

E : (As above, for A E , we denote A = fz 2 E : (z; A) g). The

next lemma follows by regularity of Borel measures and -maxitivity of idempotent measures.

Lemma 3.1.23. We can equivalently write

pld (P ; ) = inf > 0 : P1=r (A) A + ;

(A) P1=r A + for all A 2 B(E ) :

Remark 3.1.24. The distance pld could equivalently be de ned in terms of open -neighbourhoods.

Theorem 3.1.25. The net fP g LD converges to if and only if ld p (P ; ) ! 0 as 2 .

Proof. The proof is analogous to the proof of Theorem 1.9.22. We rst prove that if pld (P ; ) ! 0, then the P LD converge to . By Theorem 3.1.3 it is suÆcient to prove that, given a closed set F , an open set G, and > 0, there exists Æ > 0 such that if pld (P ; ) < 1=r 1=r Æ, then P (F ) < (F ) + and P (G) > (G) . Since is -smooth relative to F , there exists Æ1 2 (0; =2) such that 1=r (F ) (F Æ1 ) =2. Therefore, if pld (P ; ) < Æ1 , then P (F ) < (F Æ1 ) + Æ1 (F ) + . Next, using -maxitivity of , we choose Æ2 2 (0; =2) such that (G) (G Æ2 ) + =2. Then, if pld (P ; ) < Æ2 , then (G) (G Æ2 ) + =2 < P1=r (G) + . Taking Æ = Æ1 ^ Æ2 proves the claim. ld Conversely, let P ! . We show using again Theorem 3.1.3 that given > 0 there exists a collection of sets Hi; i = 1; : : : ; k and Æ > 0 such that the Hi are continuous relative to and the fact that jP1=r (Hi) (Hi)j < Æ; i = 1; : : : ; k, implies that pld(P ; ) < .

© 2001 by Chapman & Hall/CRC

266

Large deviation convergence

Let Æ < =4. Let closed Æ=2-balls B1 ; :: : ; Bl centred at z1 ; : : : ; zl , respectively, be such that E n [li=1 Bi < Æ. By -maxitivity of for each i = 1; 2; : : : ; l there exists a closed ball Bi0 centred at zi of radius not less than Æ=2 and not greater than Æ, which is a continuous set relative to . Observing that a nite union of sets continuous relative to is also continuous relative to , we take as H1 ; : : : ; Hk 1 the collection of arbitrary unions of the balls B10 ; : : : ; Bl0 . We also take 0 Hk = E n [li=1 Bi0 Æ , where Æ0 > 0 and is chosen such that Hk is continuous relative to and (Hk ) 2Æ. Let jP1=r (Hi ) (Hi)j < Æ; i = 1; : : : ; k. Let F be a closed subset of E and let H 0 be the largest set out of H1 ; : : : ; Hk 1 such that F has non-empty intersection with each of the sets Bi0 that make up H 0 . Then H 0 F 2Æ so that (F ) (F \ H 0 ) + (F \ Hk ) (H 0 ) + (Hk ) < P1=r (H 0 ) + 3Æ P1=r (F 2Æ ) + 3Æ. By a symmetric argument P1=r (F ) P1=r (F \ H 0 )+P1=r (F \Hk ) P1=r (H 0 )+P1=r (Hk ) < (H 0 )+(Hk )+2Æ (F 2Æ ) + 4Æ. Thus, pld (P ; ) < . by

We now de ne an analogue of the Kantorovich-Wasserstein metric

ldBL (P ; ) =

Z + f 2Cb (E ): E kf k BL 1

sup

f (z )r dP (z )

1=r _

E

f (z ) d(z ) :

Theorem 3.1.26. The net fP g LD converges to if and only if ld BL (P ; ) ! 0 as 2 .

Proof. The fact that the convergence ld BL (P ; ) ! 0 implies LD convergence of the P to follows from Theorem 3.1.3 and Remark 3.1.4. For the converse, by Theorem 3.1.25 it suÆces to prove ld that if P ! , then

ld lim sup ld BL (P ; ) 2p (P ; ) 2

0:

Let kf k BL 1. Given Æ > 0, we choose open Æ-balls AÆ (zk ); k = 1; 2; : : : ; l; such that E n [lk=1AÆ (zk ) < Æ. Since lim sup2 P1=r E n [lk=1 AÆ (zk ) E n [lk=1AÆ (zk ) ;we may as sume that P1=r E n[lk=1AÆ (zk ) < Æ. Abbreviating p = pld (P ; )

© 2001 by Chapman & Hall/CRC

267

LD convergence in Tihonov spaces

and recalling that AÆ (z )p denotes the closed p -neighbourhood of AÆ (z ), we have Z

f (z )r dP

(z )

1=r

E

Z

l k=1 max ;:::;l

1=r

f (z )r dP (z )

+Æ

AÆ (zk )

1=r

l1=r k=1 max f (zk ) + Æ P (AÆ (zk )) + Æ ;:::;l 1 =r l k=1 max f (zk ) (AÆ (zk )p ) + p + l1=r Æ + Æ ;:::;l _ 1 =r l f (z) + Æ + p d(z) + l1=r p + l1=r Æ + Æ E

_

l1=r f (z) d(z) + 2l1=r p + 2l1=r Æ + Æ: (3.1.3) E

Similarly, _

E

f (z ) d(z ) max f (zk )(AÆ (zk )) + Æ k=1;:::;l

k=1 max f (zk ) P1=r (AÆ (zk )p ) + p + Æ ;:::;l k=1 max ;:::;l

Z

Z

AÆ (zk )p

f (z ) + Æ + p

f (z ) + Æ + p

E

r

Z

r

dP

1=r

dP (z )

1=r

+ p + Æ

+ p + Æ

f (z )r dP (z )

1=r

+ 2p + 2Æ: (3.1.4)

E

Inequalities (3.1.3) and (3.1.4) imply that

ldBL (P ; ) 2p l1=r 1 (1+2p )+2l1=r Æ +Æ Since l1=r ! 1 as 2 , p 1, and Æ > 0 is arbitrary, the proof is complete. We now consider sequential compactness for metric spaces. We thus assume that = N and replace general nets fr g by sequences frng. Probability measures are denoted by Pn .

© 2001 by Chapman & Hall/CRC

268

Large deviation convergence

De nition 3.1.27.

A sequence fPn ; n 2 N g is LD relatively sequentially compact (for rate rn ) if every subsequence fPn0 g of fPn g contains a further subsequence fPn00 g that LD converges (at rate rn00 ) to an F -idempotent probability on E .

Theorem 3.1.28.

1. Let E be a metric space. If a sequence

fPn ; n 2 N g of probabilities on (E; B(E )) is exponentially tight, then it is LD relatively sequentially compact, the LD accumulation points being deviabilities.

2. Let E be homeomorphic to a complete separable metric space. If a sequence fPn ; n 2 N g is LD relatively sequentially compact, then it is exponentially tight. Proof. We prove part 1. Let us assume rst that E is a separable metric space. Then it is embedded as a dense subspace into a compact metric space E 0 . We extend probabilities on (E; B(E )) to probabilities on (E 0 ; B(E 0 )) by letting P 0 (A0 ) = P A0 \ E ; A0 2 B(E 0 ): The set Cb;+1 (E 0 ) of R+ -valued continuous functions on E 0 that are bounded by 1, endowed with the topology of uniform convergence, is a separable metric space. Let Cb;+1;d (E 0 ) denote a countable dense + 0 subset. The set [0; 1]Cb;1;d (E ) with product topology is sequentially compact, so the diagonal argument yields existence of a subsequence nk such that the sequences f kf k 0nk ; k 2 N g converge for all f 2 Cb;+1;d (E 0 ), where kf k 0nk refers to the norm relative to Pn0 k . The inequality j kgk 0n kgk 0mj j kf k 0n kf k 0m j+2 sup jg(z0 ) f (z0 )j z 0 2E 0 and the fact that Cb;+1;d (E 0 ) is dense in Cb;+1 (E 0 ) show that the sequences f kf k 0nk ; k 2 N g converge for all f 2 Cb;+1 (E 0 ), which implies in analogy with the proof of Theorem 3.1.19 that there exld ists a deviability 0 on E 0 such that Pn0 k ! 0 as k ! 1 at rate rnk . Exponential tightness of fPn ; n 2 N g implies that inf K 2K 0 (E 0 n K ) = 0 (where K is the collection of compact subsets of E ) so that 0 (E 0 n E ) = 0 and the set function de ned by (A) = 0 (A); A E; is a deviability on E by Corollary 1.7.12. W It is left to check that kf k nk ! E f d for all f 2 Cb+(E ). By Theorem 3.1.3 we may assume that f is uniformly continuous on E so that it can be extended to f 0 2 Cb+ (E 0 ), see, e.g., Engelking [47].

© 2001 by Chapman & Hall/CRC

269

LD convergence in Tihonov spaces

0 k 0 ! W 0 f 0 d0 , kf 0 k 0 = kf k n The W required follows since k f nk nk k E W and E 0 f 0 d0 = E f d. Now, if E is an arbitrary metric space, then by the exponential tightness condition there exists a -compact space E 0 E such that limn!1 Pn1=rn (E n E 0 ) = 0. Applying the part just proved to the probabilities Pn0 on the separable metric space E 0 de ned by Pn0 (A) = Pn (A)=Pn (E 0 ); A 2 B(E 0 ), we deduce existence of an LD convergent subsequence for the Pn0 . This provides us with an LD convergent subsequence for the Pn . Part 1 is proved. For part 2 we rst check that for every Æ > 0 and > 0 there exist open Æ-balls A1 ; : : : ; Ak such that

lim sup Pn1=rn n!1

En

k [ i=1

Ai

:

(3.1.5)

Let open Æ-balls Ai be such that [1 i=1 Ai = E . Let subsequences kl and nl be such that lim sup lim sup Pn1=rn k!1 n!1 ld and Pnl ! 0 for some arbitrary k, k

[ 0 E n Ai

i=1

En

k [ i=1

Ai = lim

l!1

Pn1l=rnl

En

kl [ i=1

Ai

F -idempotent probability 0. Then, for

1=r lim sup Pnl nl E l!1

n

k [ i=1

Ai

llim !1

1=rnl Pnl E

n

kl [ i=1

Ai :

The required inequality (3.1.5) follows since limk!1 0 E n k [i=1Ai = 0. Since each Pn is tight by Ulam's theorem, (3.1.5) implies that for arbitrary " > 0 and k 2 N ; there exist open 1=k-balls Ak1 ; : : : ; Aknk such that for all n 2 N

Pn1=rn E n

nk [ i=1

Aki

© 2001 by Chapman & Hall/CRC

2"k :

270

Large deviation convergence

T Snk The set B = 1 k=1 i=1 Aki is totally bounded and hence relatively compact by completeness of E . Also for every n 2 N

Pn1=rn (E nB )

1 X k=1

Pn1=rn E n

nk [ i=1

Aki

":

Remark 3.1.29. Part 1 also follows by Theorem 3.1.19 and Theorem 3.1.25 (or Theorem 3.1.26).

As a consequence, we have the following version of Corollary 3.1.22. The proof is similar.

Corollary 3.1.30. Let E be a metric space and E 0 be homeomorphic

to a complete separable metric space. Let E0 E . If a sequence fPn ; n 2 N g of probabilities on (E; B(E )) is E0 {exponentially tight and a function f : E ! E 0 is Borel measurable and E0 -continuous, then the sequence fPn Æ f 1 ; n 2 N g is exponentially tight.

As an illustration of the use of LD relative compactness arguments, we prove Gartner's theorem. Let L(X ) denote the distribution of a random variable X and E denote expectation with respect to a probability measure P .

Theorem 3.1.31. Let fX ; 2 g be a net of Rk {valued random

variables de ned on respective probability spaces ( ; F ; P ) such that for each 2 Rk 1 lim ln E exp r X = G(); 2 r

where G() is an R -valued lower semi-continuous and essentially ld smooth convex function such that 0 2 int (dom G). Then L(X ) ! ; at rate r , where is the deviability speci ed by the density (x) = exp sup2Rk ( x G()) : Proof. We act as in the proof of Lemma 1.11.19. We rst note that the net fL(X ); 2 g is exponentially tight. To see this, we write by Chebyshev's inequality, for A > 0 and > 0, denoting by ei ,

© 2001 by Chapman & Hall/CRC

271

LD convergence in Tihonov spaces

i = 1; : : : ; 2k, the 2k-vector, whose b(k + 1)=2cth entry is 1 if k is odd, 1 if k is even, and the rest of the entries are equal to 0, P1=r (jX j > A) max P1=r (ei X > A=k) i=1;:::;2k

exp( A=k) i=1max E exp(r ei X ) 1=r : ;:::;2k The exponential tightness follows since by hypotheses

lim E exp(r ei X ) 1=r = exp(G(ei )); 2 where the right-hand side is nite if is small enough by the fact that G() is nite in a neighbourhood of the origin. Therefore, by Theorem 3.1.19 there exists a subnet fX 0 ; 0 2 0 g ld ~ of fX ; 2 g and a deviability ~ on Rk such that L(X 0 ) ! : Next, it follows from Chebyshev's inequality that if 2 int(dom G), then the function exp( x); x 2 Rk is uniformly exponentially integrable with respect to fL(X 0 ); 0 2 0 g, so by Lemma 3.1.12 0 1=r0 _ ~ x); = exp( x) d(

lim0 E exp(r0 X )

Rk

2 int(dom G):

W

~ x) = exp(G()) for all 2 int(dom G), Thus, Rk exp( x) d( ~ = . which as in the proof of Lemma 1.11.19 implies that The following result is proved similarly to Theorem 1.9.28. Theorem 3.1.32. Let E be a Tihonov space. Let G be a subset of + Cb (E ) consisting of uniformly bounded and pointwise equicontinuous functions, i.e., supf 2G supz2E f (z ) < 1 and for every > 0 and z 2 E there exists an open neighbourhood Uz of z such that ld supf 2G supz0 2Uz jf (z ) f (z 0 )j : If P ! , then Z

lim sup f 2G

f r dP

E

1=r

_

E

f d = 0:

As we have mentioned, if we replace space Cb+ (E ) in the de nition of weak LD convergence by space CK+(E ) of R+ -valued continuous functions with compact support, then we obtain the notion of vague LD convergence.

© 2001 by Chapman & Hall/CRC

272

Large deviation convergence

De nition 3.1.33. We say that a net fP ; 2 g of probabilities on (E; B(E )) vaguely LD converges at rate r to a probability on E if for every f 2 CK+(E ) Z

lim 2

f (z )r dP (z )

E

1=r

=

_

E

K-idempotent

f (z ) d(z ):

If E is locally compact and Hausdor, the vague LD convergence has properties similar to the properties of the weak LD convergence. For instance, there is an easy analogue of Theorem 1.9.2. A distinctive feature of this type of LD convergence is that nets of probability measures are LD relatively compact.

Theorem 3.1.34. Let E be a locally compact Hausdor topologi-

cal space. Then a net fP ; 2 g of probabilities on (E; B(E )) is vaguely LD relatively compact.

The proof is similar to the proof of part 1 of Theorem 3.1.19, the main distinction being the use of Theorem 1.7.21 in place of Theorem 1.7.25. At times it is more intuitive to formulate large deviation convergence of probability measures as large deviation convergence in distribution of the associated random variables.

De nition 3.1.35. Let fX ; 2 g be a net of random variables

de ned on respective probability spaces ( ; F ; P ) and assuming values in a topological space E and X be an idempotent variable de ned on an idempotent probability space ( ; ) and assuming values in E , whose idempotent distribution is -smooth relative to the collection of closed subsets of E . We say that the net fX ; 2 g large ld deviation converges in distribution to X if P Æ X 1 ! Æ X 1. ld We denote large deviation convergence in distribution by ! as well. Whether this notation refers to large deviation convergence of probability measures or large deviation convergence in distribution of random variables should be clear from the context. We will also occasionally say that a net of random variables is LD relatively compact if the associated net of laws is LD relatively compact. We have the following version of Lemma 3.1.12. Let us say that a net f ; 2 g of R+ -valued random variables on respective prob-

© 2001 by Chapman & Hall/CRC

273

LD convergence in Tihonov spaces

ability spaces ( ; F ; P ) is uniformly exponentially integrable relative to fP ; 2 g (with rate r ) if lim lim sup E r 1( > A) A!1 2

1=r

= 0:

Lemma 3.1.36. Let !ld

, where is an R+ -valued idempotent variable on an idempotent probability space ( ; ). If the net f ; 2 g is uniformly exponentially integrable relative to fP ; 2 g, then lim2 E r 1=r = S: We now prove a number of technical lemmas. Let us recall that if X and Y are random variables with values in respective separable metric spaces E and E 0 with Borel -algebras, then (X; Y ) is a random variable in E E 0 with product topology and Borel algebra; in particular, if E = E 0 and denotes the metric on E , then (X; Y ) is a random variable. The following result is an analogue of Lemma 1.10.5 and admits a proof along the same lines. We give another proof that illustrates the use of metrics.

Lemma 3.1.37. Let E be a separable metric space with metric ,

and let X and Y ; where 2 , 2 , and are directed sets, be nets of random variables with values in E , de ned on respective probability spaces ( ; F ; P ). Let

lim lim sup P1=r (X ; Y ) " = 0; " > 0; 2 2 and

ld L X ! as 2 , where ; 2 ; are F -idempotent probabilities on E . Then, for an F -idempotent probability on E , we have that ld L(Y ) ! as 2

if and only if

iw ! as 2 :

© 2001 by Chapman & Hall/CRC

274

Large deviation convergence

Proof. The claims follow by Theorem 1.9.25 and Theorem 3.1.26 since in view of the de nitions of BL and ld BL BL (

; ) ld BL L(Y );

Z + sup f 2Cb+ (E ):

kf k BL 1

f (X )r dP

ldBL L(X ); +

ldBL L(X );

1=r Z

Z

1=r r f (Y ) dP

r

1 ^ (X ; Y )

dP

1=r

:

We will often use the case where the X do not depend on . Lemma 3.1.38. Let E be a separable metric space with metric , and let X and Y ; where 2 , be nets of random variables de ned on respective probability spaces ( ; F ; P ) with values in E . If ld L(X ) ! , where is an F -idempotent probability on E , and 1=r P X ;Y !

ld 0 as 2 , then L(Y ) ! . We give another application of metrics. De nition 3.1.39. We say that a net fX ; 2 g of random variables on ( ; F ; P ) assuming values in a metric space E with metric converges to z 2 E super-exponentially in probability at rate r (or simply super-exponentially in probability if the rate is under-

stood) and write X every > 0.

1=r P

! z if lim2 P1=r ((X ; z) > ) = 0 for 1=r

Remark 3.1.40. Note that X P! Z

lim 2

r

1 ^ (X ; z )

dP

1=r

z if and only if = 0:

Lemma 3.1.41. Let fX ; 2 g be a net of random variables on

respective probability spaces ( ; F ; P ) assuming values in a metric 1=r P X !

ld space E with metric . Then z if and only if L(X ) ! 1z , where 1z denotes the unit mass at z .

© 2001 by Chapman & Hall/CRC

275

LD convergence in Tihonov spaces

1z , then by the de nition of LD convergence 1=r _ 1^(X ; z ) r dP = 1^(z 0 ; z ) d 1z (z 0 ) = 0:

ld Proof. If L(X ) !

lim

2

Z

E

The converse follows since as in the proof of Lemma 3.1.37

1

ldBL (L(X ); z )

Z

r

1 ^ (X ; z )

dP

1=r

:

The next lemma considers joint LD convergence. We formulate the results in the language of LD convergence in distribution. Lemma 3.1.42. Let E and E 0 be separable metric spaces, and let X and Y ; where 2 , be nets of random variables on ( ; F ; P ) with values in E and E 0 , respectively. Let X and Y be idempotent variables on an idempotent probability space ( ; ) with values in E and E 0 , respectively, whose idempotent distributions are -smooth relative to the associated collections of closed sets. Let E E 0 be equipped with product topology. ld ld 1. If X ! X, Y ! Y , X and Y are independent, and X ld and Y are independent, then (X ; Y ) ! (X; Y ).

2. If

ld X !

X and

1=r P Y !

z , then (X ; Y )

ld ! (X; z ).

Proof. The proof of part 1 uses Theorem 3.1.32 and is analogous to the proof of part 1 of Lemma 1.10.8. In more detail, let PX and PY denote the respective distributions of X and Y , and X and Y denote the respective idempotent distributions of X and Y ; Theorem 3.1.32 implies that for an R+ -valued bounded uniformly continuous function h(z; z 0 ) on E E 0 Z Z lim 2 E E0 Z

E

1=r

h(z; z 0 )r dPY (z 0 ) dPX (z )

sup h(z; z 0 ) Y (z 0 ) r dPX (z ) z 0 2E 0

© 2001 by Chapman & Hall/CRC

1=r

= 0: (3.1.6)

276

Large deviation convergence

Also, since supz0 2E 0 h(z; z 0 ) Y (z 0 ) is continuous in z 2 E , lim 2

Z

E

1=r sup h(z; z 0 ) Y (z 0 ) r dPX (z ) z 0 2E 0 = sup sup h(z; z 0 ) Y (z 0 ) X (z ): (3.1.7) z 2E z 0 2E 0

The required follows by (3.1.6) and (3.1.7). The proof of part 2 is also analogous: we observe that ( 1=r

P 0 )((X ; Y ); (X ; z )) ! 0, where 0 is a product metric on E E 0 , so that the required follows by part 1 and Lemma 3.1.38.

3.2 Large deviation convergence in the Skorohod space The purpose of this section is to lay groundwork for deriving large deviation convergence results for semimartingales. We begin by introducing basic notation for the Skorohod space. We denote by D = D (R + ; Rd ); d 2 N ; the space of Rd {valued, right-continuous with left-hand limits functions x = (xt ; t 2 R+ ). We equip it with the Skorohod J1 topology and metrise it by the Skorohod{Prohorov{Lindvall metric denoted by S , under which it is a complete separable metric space. Let D denote the Borel { algebra on D , Dt , for t 2 R+ , denote the sub{{algebra generated by the coordinate maps x ! xs ; s t; and D = (Dt ; t 2 R+ ). (Note that the ow D is not right-continuous.) Given x 2 D , we denote xt = supst jxsj; xt = sups 0 and Æ > 0, we de ne the modulus of continuity w0 (x; Æ) = inf max wx [tj 1 ; tj ) ; T

(tj ) j =1;:::;k

where wx [s; t) = supu;v2[s;t) jxu xv j; s < t; and the in mum is taken over all collections (tj ) such that 0 = t0 < t1 < : : : < tk = T and tj tj 1 > Æ for j < k. The next theorem routinely follows by characterisation of compacts in D , see, e.g., Jacod and Shiryaev [67], and is analogous to tightness conditions for sequences of probabilities in D , cf. Ethier and Kurtz [48].

Theorem 3.2.1. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ). The net fL(X ); 2 g is exponentially tight if and only if for all T > 0 and > 0 (i) lim lim sup P1=r sup jXt j > A = 0; A!1 2 tT (ii) lim lim sup P1=r wT0 (X ; Æ) > = 0: Æ!0 2 The concept of E0 -exponential tightness with E0 = C plays an important role in the developments below, so we repeat it here.

© 2001 by Chapman & Hall/CRC

278

Large deviation convergence

De nition 3.2.2.

We say that a net fP ; 2 g of probability measures on D is C -exponentially tight if it is exponentially tight and every LD accumulation point is supported by C . The next result gives conditions for C -exponential tightness. Theorem 3.2.3. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be a net of stochastic processes with paths in D , de ned on respective stochastic bases ( ; F ; F ; P ). The net fL(X ); 2 g is C { exponentially tight if and only if either one of the following equivalent conditions I or II holds. I (i) lim lim sup P1=r jX0 j > A = 0; A!1 2 (ii) lim lim sup P1=r sup jXt Xs j > = 0; Æ!0 2 s;t2[0;T ]: js tjÆ

lim lim sup P1=r sup jXt j > A = 0; A!1 2 tT 1=r (ii) lim lim sup sup P sup jX+t X j > = 0; Æ!0 2 2ST (F ) 0tÆ where T > 0 and > 0 are arbitrary, and ST (F ) denotes the set of all F {stopping times not greater than T . Remark 3.2.4. A similar result holds if the X assume values in a complete separable metric space E . Then instead of conditions I(i) and II(i) one should require that the nets fL(Xt ); 2 g be exponentially tight in E for all t 2 R+ , and in I(ii) and II(ii) replace the moduli of the increments of the X by the distances between the values of the X , cf. Ethier and Kurtz [48]. We precede the proof with a lemma. In the rest of the book we use E to denote expectation and F to denote -algebras. Lemma 3.2.5. Let i; i 2 N ; be positive random variables on a probability space ( ; F ; P ) and let

II (i)

(

max k 2 N : At = 0; If Æ > 0, t 2 R+ and N

P (At N ) 1

© 2001 by Chapman & Hall/CRC

Pk i=1 i

2N

t NÆ

t ;

if 1 t; otherwise:

are such that NÆ=t > 1, then

1

max P (k Æ; At k):

k=1;:::;N

279

LD convergence in the Skorohod space

Proof. In view of Chebyshev's inequality for n N

P (At N ) P (At N; n Æ) + P (At N; n Æ) 1 k=1max P (At N; k Æ) + E n 1(At N ) : (3.2.1) ;:::;N Æ Choosing n such that E n 1(At N ) = minkN E k 1(At N ) , we estimate the expectation on the right of (3.2.1) as follows

N1

E n 1(At N )

=

1 E N

N X

E k 1(At N )

k=1 N X k=1

k 1(At N )

Nt P (At N ):

Substituting the estimate into the right-hand side of (3.2.1) gives the required inequality. Proof of Theorem 3.2.3. It is obvious that part I implies part II. Let C -exponential tightness of fL(X ); 2 g hold. We prove that the assertion of part I holds. The argument is fairly standard. We derive part I(ii). Let us denote x;Æ = P1=r sup s;t2[0;T ]: jXt Xs j > . js tjÆ In analogy with the diagonal argument in the proof of Theorem 1.9.17 there exists a subnet fP0 ; x0 ; 0 2 0g of fP ; x;Æ ; 2 ; Æ > ld 0g such that P0 ! for some deviability and lim0 20 x0 = lim supÆ!0 lim sup2 x;Æ . Then Corollary 3.1.9 and the fact that (D n C ) = 0 yield for arbitrary Æ0 > 0

lim sup lim sup P1=r Æ!0 2

sup jXt s;t2[0;T ]: js tjÆ

Xs j >

0 0 1=r lim0 sup P0 0 sup jXt Xs j > 0 2

s;t2[0;T ]: js tjÆ0

x 2 C : sup jxt xsj :

s;t2[0;T ]: js tjÆ0

The claim follows since the right-hand side can be made arbitrarily small by choosing Æ0 in view of the -smoothness property of a deviability with respect to decreasing nets of closed sets. Part I(i) is derived similarly.

© 2001 by Chapman & Hall/CRC

280

Large deviation convergence

Thus, it remains to prove that part II implies the C -exponential tightness. We rst prove that under the hypotheses the net fL(X ); 2 g is exponentially tight in D . We apply Theorem 3.2.1. The rst condition of the theorem holds by hypotheses. We check the second. We de ne stopping times

0 = 0; k = inf ft 2 R+ : jXt X

k 1

j =2g; k 2 N :

Introducing AT = maxfk 2 N : k T g and k = k have for Æ < T by the fact that wX [k 1 ; k ) < P wT0 (X ; Æ) P AT

1; min k Æ

k 1 , we

kAT N ) + P min k Æ; 1 AT < N kAT N 1 i XX P (AT N ) + P k Æ; AT = i i=1 k=1 P(AT N ) + N 2 k=1max P k Æ; ;:::;N

P (AT

Since by Lemma 3.2.5, for N

P (AT

AT

k :

2 N,

2T N ) 2 k=1max P k ; AT k ; ;:::;N N

we obtain the estimate 2T P wT0 (X ; Æ) 2 max P k ; AT k k=1;:::;N N 2 + N max P k Æ; AT k : (3.2.2)

k=1;:::;N

Next, by right-continuity of X

P k Æ; AT

k = P sup jX

k 1 +t tÆ P sup jX ^T +t X ^T j 2 k 1 k 1 tÆ sup P sup jX+t tÆ 2ST (F )

© 2001 by Chapman & Hall/CRC

X

k 1

j 2 ; k T

X j

: (3.2.3) 2

281

LD convergence in the Skorohod space

Similarly,

P k

2T ;A N T

k sup P sup jX+t X j 2 : 2ST (F )

t2T=N

(3.2.4)

Substituting (3.2.3) and (3.2.4) into (3.2.2) yields

P1=r wT0 (X ; Æ)

21=r sup P1=r sup jX+t X j 2 2ST (F ) + N 2=r

tÆ

sup P1=r

2ST (F )

sup jX+t t2T=N

X j

: 2

Taking on the left-hand side the limits, rstly, as 2 , then as N ! 1, and nally as Æ ! 0, checks the required condition. We now prove the C -exponential tightness. Let be deviability on D , which is an LD accumulation point for fL(X )g, and let x 2 D be a discontinuous function. We show that (x) = 0. Let t 2 R+ and > 0 be such that jxt j . Using Proposition VI.2.1 of Jacod and Shiryaev [67] (see also Liptser and Pukhalskii [78] for details), we have that if Æ > 0 is small enough, then n

fy 2 D : S (x; y) Æg y 2 D : sup jysj 2 n

y 2 D : sup jyt n

y2D

jt sjÆ : sup jys t<st+Æ

o

jt sjÆ

o ysj 4 o [n yt j 4 y2D :

jyt yt Æ j 8 o [n y 2 D : sup jys yt Æ j 8 : t Æ<st

Then by Corollary 3.1.6 (x) lim sup lim sup P1=r S (x; X ) Æ Æ!0 2 lim sup lim sup 31=r sup P1=r sup jXu+s ut Æ!0 2 0<sÆ

The latter limit equals 0 by hypotheses.

© 2001 by Chapman & Hall/CRC

Xu j

: 8

o

282

Large deviation convergence

The following analogue of the Lenglart{Rebolledo inequality will allow us to estimate the probabilities in part II(ii) \in predictable terms".

Lemma 3.2.6. Let X = (Xt ; t 2 R+ ) and Y = (Yt; t 2 R+ ) be positive processes on a stochastic basis ( ; F ; F; P ). If E (X =Y ) 1 for every F-stopping time < 1, then for every F-stopping time 1, a > 0 and b > 0,

P sup Xt a t

ab + P sup Yt > b

t

(here supt1 = supt2R+ ). Proof. We de ne the stopping time = inf ft 2 R+ : Xt ag 1: If P ( < 1) = 1, then

P (sup Xt > a) P (X^

a) P (Y^ > b) +P (X^ a; Y^ b) P (Y^ > b)+P (X^ =Y^ a=b): t

By Chebyshev's inequality

P sup Xt > a t

P (Y^ > b)+ ab P sup Yt > b + ab :

t

To obtain the required, note that

P sup Xt a = lim P sup Xt > a N !1

t

t

1 : N

If P ( < 1) < 1, then by the part just proved we have for N > 0

P sup Xt a t

P sup Xt > a N1

t

1 b = lim P sup Xt > a P sup Yt > b + : M !1 N a 1=N t t^M

Since N is arbitrary, the proof is over. The following useful fact is a direct consequence of Theorem 3.2.3.

© 2001 by Chapman & Hall/CRC

LD convergence in the Skorohod space

283

Corollary 3.2.7.

Let fX ; 2 g and fY ; 2 g be nets of stochastic processes with paths in respective Skorohod spaces D (R + ; R d ) and D (R + ; R k ) such that X and Y are de ned on a common probability space ( ; F ; P ) for every 2 . If the net fL(X ); 2 g is C (R + ; Rd )-exponentially tight and the net fL(Y ); 2 g is C (R + ; Rk )-exponentially tight, then the net fL(X ; Y ); 2 g of distributions on D (R + ; Rd Rk ) is C (R + ; R d R k )-exponentially tight.

The next two results concern methods of identifying LD limits. The following theorem presents the method of nite-dimensional distributions.

Theorem 3.2.8. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be

a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ). Let the net fL(X ); 2 g of the distributions of the X on D be C {exponentially tight. Let for all k 2 N ; and t1 < : : : < tk 2 U , where U is a dense subset of R+ , as 2 , ld L(Xt1 ; : : : ; Xtk ) ! t1 ;:::;tk ;

where t1 ;:::;tk are deviabilities on (Rd )k . Then idempotent probability on D with density (x) = inf t1 ;:::;tk 2U t1 ;:::;tk xt1 ; : : : ; xtk if x = (xt ; t 2 R+ ) 2 C ; and (x) = 0 if x = (xt ; t 2 R+ ) 2 D n C , is a deviability on D , and ld L(X ) ! : Proof. By Theorem 3.1.19 fL(X ); 2 g is LD relatively compact. Let 0 be an LD accumulation point. It suÆces to prove that 0 = . By C -exponential tightness of fL(X ); 2 g this is true on D n C . Let x 2 C . By the contraction principle (Corollary 3.1.15) we have that 0 Æ t11;:::;tk = t1 ;:::;tk . By Theorem 2.2.2 and Remark 2.2.3 0 (x) = inf t1 ;:::;tk 2U t1 ;:::;tk (xt1 ; : : : ; xtk ) = (x).

The following result, which roughly shows that LD limits in distribution of non-negative martingales are exponential maxingales, lays a foundation for the maxingale problem method of proving LD convergence. We denote E1=r = (E )1=r , where, as above, E denotes expectation with respect to P .

© 2001 by Chapman & Hall/CRC

284

Large deviation convergence

Theorem 3.2.9. Let X ; 2 ; be processes with paths in D ned on respective stochastic bases ( ; F ; F ; P ). Let the

denet fL(X ); 2 g be C {exponentially tight and be a deviability on D supported by C , which is an LD accumulation point of fL(X ); 2 g. Let M = (Mt; t 2 R+ ); 2 ; be R+ {valued martingales on ( ; F ; F ; P ) such that the net f(Mt )1=r ; 2 g is uniformly exponentially integrable relative to the net fP ; 2 g for each t 2 R+ , and let Mt (x); t 2 R+ ; x 2 D ; be an R+ {valued function, which, for every t 2 R+ , is C {continuous, Borel measurable and, if restricted to C , Ct {measurable in x. If, for every t 2 R+ , as 2 ,

(Mt )1=r Mt (X )

1=r P

! 0;

then (Mt (x); t 2 R+ ; x 2 C ) is a C-exponential maxingale on (C ; ).

Proof. Since fL(X ); 2 g is C {exponentially tight, we can asld sume, by taking a subnet if necessary, that L(X ) ! . Consider a function f (x) = g(xt1 ; : : : ; xtk ), where 0 t1 < : : : < tk and g : (Rd )k ! R+ is continuous and bounded. Since f and Mt are Borel measurable, C {continuous and (D nC ) = 0, by the contraction principle for every t 2 R+

ld L Mt(X )f (X ) ! Æh 1 ; where h : D ! R+ is de ned by h(x) = Mt (x)f (x). Since by

hypotheses and boundedness of f (Mt )1=r f (X )

Mt

1=r P (X )f (X ) !

0;

it follows by Lemma 3.1.38 that

ld L (Mt)1=r f (X ) ! Æ h 1:

(3.2.5)

By uniform exponential integrability of f(Mt )1=r ; 2 g relative to fP ; 2 g, boundedness of f and Lemma 3.1.36 we then have

lim E 1=r Mt f (X )r = sup x Æ h 1 (x) 2 x2R+ = sup Mt (x)f (x)(x): (3.2.6) x2C

© 2001 by Chapman & Hall/CRC

285

LD convergence in the Skorohod space

(The last equality is the change-of-variables formula from Theorem 1.4.6.) Now let 0 s < t and ti s; i = 1; : : : ; k. By the martingale property

E Mt f (X )r = E Ms f (X )r ; so (3.2.6) yields the maxingale property sup f (x)Mt (x)(x) = sup f (x)Ms (x)(x): x2C

x2C

The collection of functions g(xt1 ; : : : ; xtk ) such that 0 t1 < : : : < tk s and g : (Rd )k ! R+ are continuous and bounded satis es the requirements on Hs in part 2 of Lemma 2.3.5 and also generates Cs. An application of Lemma 2.3.5 shows that M is an exponential maxingale on (C ; C; ). Also taking f = 1 in (3.2.5), we obtain by Lemma 3.1.12 for a > 0 lim inf E1=r Mt 1 (Mt )1=r > a 2 sup Mt (x) 1(Mt (x) > a)(x): x2C

The latter implies, since f(Mt )1=r ; 2 g is uniformly exponentially integrable relative to fP ; 2 g, that (Mt (x); x 2 C ) is {maximable. In certain cases C -exponential tightness allows one to establish LD convergence for the locally uniform topology on D . The following result is an adaptation of Theorem 3.1.10.

Theorem 3.2.10. Let X !ld

X for the Skorohod topology, where X is a Luzin-continuous idempotent process. If the X are random variables on D relative to the locally uniform topology on D , then ld X ! X for the locally uniform topology. Proof. Since convergence in the Skorohod topology to a continuous function is equivalent to locally uniform convergence, the Skorohod and locally uniform topologies are locally equivalent at every x 2 C so Theorem 3.1.10 applies.

© 2001 by Chapman & Hall/CRC

286

Large deviation convergence

We now discuss composition and rst-passage-time mappings. If x = (x1 ; : : : ; xd ) 2 D and the component functions of y = (y1 ; : : : ; yd ) 2 D are increasing and R+ -valued, we de ne the composition x Æ y by x Æ y = ((x1yt1 ; : : : ; xdyd ); t 2 R+ ).

Lemma 3.2.11.

t

ld Let X ! X , where X is a Luzin-continuous idempotent process. Let Y be stochastic processes with paths in D , whose component processes are R+ -valued and increasing, such that 1=r P Y !

y^ 2 C . Then X Æ Y !ld X Æ y^ . Proof. Clearly, y^ is component-wise R+ -valued and increasing so ld that X Æ y^ is well de ned. By Lemma 3.1.42 (X ; Y ) ! (X; y^ ). The claim now follows by Corollary 3.1.15 and the fact that the composition map (x; y) ! x Æ y is continuous at (x; y) such that x and y are continuous, Billingsley [11], Whitt [135, Theorem 3.1]. De nition 3.2.12. Given an R+ -valued function x = (xt ; t 2 R+ ) from D (R + ; R) with sample paths that are unbounded above, the associated rst-passage-time function x( 1) = (x(t 1) ; t 2 R+ ) 2 D (R + ; R )

x(t

is de ned by

xs > tg; t 2 R+ : If x = (x1 ; : : : ; xd ) 2 D is such that the xi ; i = 1; : : : ; d; are R+ valued and unbounded above, we de ne x( 1) = (x1 ( 1) ; : : : ; xd ( 1) ). 1)

= inf fs 2 R+ :

In the next lemma c ! 1 as 2 . We also denote = (t; t 2 R+ ) and, for a vector = (1 ; : : : ; d ), we let e = (t; t 2 R+ ) and 1 = (1=1 ; : : : ; 1=d ) if has positive entries. For vectors = (1 ; : : : ; d ) and = ( 1 ; : : : ; d ), we denote = (1 1 ; : : : ; d d ). Lemma 3.2.13. Let fX ; 2 g be a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ) such that the component processes are R+ -valued and unbounded above. Let X be an Rd -valued Luzin-continuous idempotent process de ned on an idempotent probability space ( ; ). 1. Let, in addition, the component idempotent processes of X be R + -valued, unbounded above and strictly increasing -a.e. If ld ld X ! X , then X ( 1) ! X ( 1) .

e

© 2001 by Chapman & Hall/CRC

287

LD convergence in the Skorohod space

2. Let Y = (Yt ; t 2 R+ ) be stochastic processes with paths 0 0 from D (R + ; Rd ) and Y be an Rd -valued Luzin-continuous 0 idempotent process. Let 2 Rd be such that ! , where is entrywise positive. Let, in addition, X0 = 0. ld If (c (X e); Y ) ! (X; Y ) in D (R + ; Rd+d0 ), then ld (c (X ( 1) 1 e); c (X 0 e); Y ) ! ( 1 XÆ 1 2 d + d ( e); X; Y ) in D (R + ; R ). Proof. Part 1 follows by Corollary 3.1.15 and the fact that the map x ! x( 1) is continuous at strictly increasing x 2 D , Whitt [135, Theorem 7.2]. We prove part 2. We rst consider the case = 1d , where 1d is ld a d-vector with unity entries. Since c (X 1d e) ! X as 2 , P

1=r

! 1, and X is Luzin-continuous, it follows that X ! 1d e 1=r ( 1) P in D so by part 1 and Lemma 3.1.41 X ! 1d e in D . By ( 1) ld Lemma 3.2.11 (c (X 1d e)ÆX ; c (X 1d e); Y ) ! (X; X; Y ). c

Since

c (X (

1)

1d e) = c (1d e

X ) Æ X (

1)

by Lemma 3.1.38 it suÆces to prove that c in D , which would follow by sup

t2[0;T ]

c jX ÆX (t 1)

+ c (X Æ X (

1)

1=r P

1d e) !

(X ÆX ( 1)

1=r P

tj

1d e) ;

! 0 ; T 2 R+ :

0

(3.2.7)

Since ( 1)

0 sup (X Æ X t t2[0;T ]

t) (X0 )+ _

sup

t2[0;X (T 1) ]

(Xt )+ ;

we have for A 2 R+ and > 0 that

P1=r ( sup c jX Æ X (t t2[0;T ]

1)

tj > ) P1=r X (T

1)

>A

+ P1=r sup c ((X0 )+ _ (Xt )+ ) > : (3.2.8) t2[0;A]

© 2001 by Chapman & Hall/CRC

288

Large deviation convergence

Since the function x ! supt2[0;A] (xt )+ is continuous at continuous x, Liptser and Shiryaev [79], and x ! x0 is continuous, by Corolld lary 3.1.15 and the LD convergence c (X 1d e) ! X lim sup P1=r sup c ((X0 )+ _ (Xt )+ ) ) 2 t2[0;A] sup X0+ _ (Xt )+ = 0 ; t2[0;A] proving that the second term on the right of (3.2.8) tends to 0 as 2 P

1=r

. The rst term goes to 0 as 2 and A ! 1 since X ( 1) ! 1d e. The limit (3.2.7) has been proved so that the claim for the case = 1d has been proved. For general the hypotheses imply by Lemma 3.2.11 that ld c (X Æ ( 1 e) 1d e); Y ) ! (X Æ ( 1 e); Y ) so by the part ld proved c (X Æ ( 1 e))( 1) 1d e); c (X Æ ( 1 e) 1d e); Y ) ! X Æ ( 1 e); X Æ ( 1 e); Y . Since (X Æ ( 1 e))( 1) 1d e =

X ( 1) 1 e , by Corollary 3.1.15 and Lemma 3.2.11 c X ( 1) ld 1 e); c (X e); Y ) ! 1 X Æ ( 1 e); X; Y .

© 2001 by Chapman & Hall/CRC

Chapter 4

The method of nite-dimensional distributions In this chapter we consider the method of nite-dimensional distributions of identifying LD accumulation points. It is best suited for studying LD convergence to idempotent processes with independent increments and is based on Theorem 3.2.8. As in the preceding section, we consider a net fX ; 2 g of stochastic processes, which for the most part are semimartingales de ned on respective stochastic bases ( ; F ; F ; P ) and having paths in D = D (R + ; Rd ). The ltrations F = (Ft ; t 2 R+ ) are assumed to be complete and rightcontinuous; E denotes expectation with respect to P . We assume as xed a net fr ; 2 g of real numbers greater than 1 converging to 1 as 2 , which is used as a rate for LD convergences below; the latter refer to the Skorohod topology. We retain the rest of the notation of Section 3.2, e.g., we write E1=r for (E )1=r . Section 4.1 formulates conditions for LD convergence in distribution in terms of convergence of the stochastic exponentials of the semimartingales, Section 4.2 gives conditions on convergence of the predictable characteristics, Sections 4.3 and 4.4 consider implications of the general results. 289 © 2001 by Chapman & Hall/CRC

290

Finite-dimensional LD convergence

4.1 Convergence of stochastic exponentials In this section we use the method of nite dimensional distributions to derive conditions for LD convergence of semimartingales in a Skorohod space in terms of convergence of the associated stochastic exponentials. We start by introducing the general setting for both this chapter and the next one. For the notions and facts from stochastic calculus used below we refer the reader to Jacod and Shiryaev [67] and Liptser and Shiryaev [79]. Let X = (Xt ; t 2 R+ ); 2 ; be Rd {valued semimartingales de ned on stochastic bases ( ; F ; F ; P ), where F = (Ft ; t 2 R + ). All the X , as well as all the processes we consider below, have paths in an appropriate Skorohod space (which is D for the X ). We recall that a Borel function h : Rd ! Rd is said to be a limiter if it is bounded and h(x) = x in a neighbourhood of the origin. Every truncation function as de ned in Jacod and Shiryaev [67] is a limiter. In the same way as it is done for truncation functions one can de ne the triplet of the predictable characteristics of a semimartingale associated with a limiter. This slight extension of the class of truncation functions is convenient for technical reasons in that it allows us to consider characteristics associated with limiters that do not vanish at in nity by contrast with truncation functions. Let h(x) be a limiter. Then 1 h (x) = h(r x) (4.1.1) r is also a limiter, and we denote by (B ; C ; ) the triplet of the predictable characteristics of X associated with h (x) (i.e., de ned as if h (x) were a truncation function). We also say that (B ; C ; ) corresponds to h(x). We recall that this is equivalent to X having the following canonical representation:

Xt = X0 +Bt +Xt;c +h (x)( )t +(x h (x))t ;

(4.1.2)

where B = (Bt ; t 2 R+ ); B0 = 0; is an Rd {valued F {predictable process with bounded variation over bounded intervals; X ;c = (Xt;c ; t 2 R+ ); X0;c = 0; is an Rd {valued continuous local martingale with respect to F that is the continuous martingale part of X ;

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

291

is the measure associated with jumps of X , i.e., ([0; t]; ) =

X

0<st

1(Xs 2 nf0g); 2 B(Rd );

is an F {predictable random measure on (R+ Rd ; B(R+ ) B(Rd )) that is the F {compensator of . We use to denote integration so that h (x) (

)

t

=

(x h (x)) t =

f (x) t =

Zt Z

0 Rd Zt Z 0 Rd Zt Z 0

h (x)( (ds; dx) (ds; dx)); (x h (x)) (ds; dx);

f (x) (ds; dx):

Rd

In analogy with earlier notation for semimaxingales we also denote

f (x) t

=

Z

f (x) (ftg; dx):

Rd

An Rdd {valued continuous process C = (Ct ; t 2 R+ ); C0 = 0; is de ned to be the F {predictable quadratic-variation process of X ;c . We also de ne the continuous part of the predictable measure of jumps by

;c(ds; dx) = 1( (fsg; R d ) = 0) (ds; dx): We consider the version (B ; C ; ) of the characteristics for which identically: Ct Cs; 0 s < t; is a symmetric positive semi-de nite d d{ matrix,

(f0g; R d ) = 0; (R+ ; f0g) = 0; (ftg; R d ) 1; (4.1.3a) (jxj2 ^ 1) t < 1; (4.1.3b) Bt = h (x) t : (4.1.3c)

© 2001 by Chapman & Hall/CRC

292

Finite-dimensional LD convergence

We recall that C and do not depend on the choice of h, while if B = (B t ; t 2 R+ ) is the rst characteristic corresponding to another limiter h(x), then

B t Bt = (h (x) h (x)) t :

(4.1.4)

Mt;Æ = Xt;c +x 1(r jxj Æ)( )t ; t 2 R+ ;

(4.1.5)

M~ t = Xt;c + h (x) ( )t ; t 2 R+ ;

(4.1.6)

Along with C , we introduce Rdd {valued processes C ;Æ = (Ct;Æ ; t 2 ~ = (C~t ; t 2 R+ ) that are the respective F { R + ); Æ > 0; and C predictable quadratic-variation processes of the locally squareintegrable martingales and and are speci ed by the equalities

Ct;Æ = Ct + x 1(r jxj Æ) 2 t X x 1(r jxj Æ) s 2 (4.1.7)

st

and

C~t = Ct + ( h (x))2 t

X

st

h (x) s 2 ; (4.1.8)

where 2 Rd . The processes C ;Æ and C~ are referred to as modi ed second characteristics. We recall that X is a special semimartingale if

jxj 1(jxj > 1) t < 1; t 2 R+ :

(4.1.9)

Then one can consider the predictable triplet of X \without truncation", i.e., assume that in (4.1.2) h (x) = x. We denote the process B corresponding to this \nontruncation" by B 0 = (B 0 t ; t 2 R+ ), so the predictable triplet without truncation is (B 0 ; C ; ). As it follows by (4.1.3c) and (4.1.4), B 0t = x t ; (4.1.10) 0 B = B + (x h (x)) : (4.1.11) t

© 2001 by Chapman & Hall/CRC

t

t

293

Convergence of stochastic exponentials

If, moreover,

jxj2 1(jxj > 1) t < 1; t 2 R+ ;

(4.1.12)

then X is said to be a locally square integrable semimartingale. In that case one can de ne \nontruncated modi ed second characteristics" C~t0 = (C~t0 ; t 2 R+ ) by X C~t0 = Ct+(x)2 t (xs)2 : (4.1.13) st The following stronger condition on plays an important role below and is further referred to as the Cramer condition: (Cr) ejxj 1(jxj > 1) t < 1; for all t 2 R+ ; 2 R+ : Under (Cr), we can de ne the stochastic cumulant 1 Gt () = B 0t + Ct+(ex 1 x)t; 2 Rd ; t 2 R+ : 2 (4.1.14) The process G () = (Gt (); t 2 R+ ) is a real-valued F {predictable process with bounded variation over bounded intervals, in particular, a semimartingale, so that we can de ne the associated stochastic (or Doleans{Dade) exponential E () = (Et (); t 2 R+ ); 2 Rd ; by

Et() = eGt ()

Y

(1+Gs ())e

0<st

Gs () :

(4.1.15)

(The right-hand side is well de ned and is a semimartingale with paths in D (R + ; R), see Liptser and Shiryaev [79, Theorem 2.4.1], Jacod and Shiryaev [67, Theorem I.4.61].) By (4.1.14), (4.1.10) and (4.1.3a) Gs () =

Z

(ex 1) (fsg; dx) > 1;

(4.1.16)

Rd

so, as we are going to see,

Et() > 0; t 2 R+ :

(4.1.17)

Therefore, we can de ne the processes Y () = (Yt (); t 2 R+ ); 2 R d ; by Y () = e(Xt X0 ) E () 1 : (4.1.18) t

© 2001 by Chapman & Hall/CRC

t

294

Finite-dimensional LD convergence

A fundamental property of the stochastic exponential is expressed by the following lemma. Lemma 4.1.1. Under the Cramer condition the process Y () = (Yt (); t 2 R+ ); 2 Rd ; is a well-de ned local martingale on ( ; F ; F ; P ). Proof. We rst check Y () is well de ned by showing that (4.1.17) holds. Since G () has bounded variation over bounded intervals, X

Next,

jGs ()j < 1:

(4.1.19)

0<st Y

0<st

(1 + Gs ())

exp 2

X

0<st

Y

jGs()j

(1 + Gs ()):

0<st: jGs()j>1=2 By (4.1.19) the product on the right has nitely many terms, which are positive, and is itself positive. We now check the local martingale property of Y (). Let for k = 2; 3; : : :

k = inf t 2 R+ : 1+Gt () < 1=k : Since by (4.1.19) G has a nite number of jumps less than 1=k 1 on a bounded interval, it follows that 1 + Gk () < 1=k. Thus, the k are F -predictable stopping times as the debuts of predictable sets whose graphs belong to the sets, see Dellacherie [34, IV-T.16], Jacod and Shiryaev [67, I.2.13]. Also k ! 1 as k ! 1. Hence, there exist F -stopping times k < k such that k ! 1 as k ! 1. It is suÆcient to show that the (Yt^k (); t 2 R+ ) are local martingales relative to F. Since k < k , we have that 1 (4.1.20) inf (1+Gt ()) : tk k P By (4.1.19) 0<st jln(1 + Gs ()j < 1, so the process ln Et () = Gt ()

© 2001 by Chapman & Hall/CRC

X

0<st

Gs ()+

X

0<st

ln(1+Gs ()) (4.1.21)

295

Convergence of stochastic exponentials

is well de ned and is a semimartingale. Let

Ut = (Xt X0 ) ln Et (): Then Yt () = exp Ut so that by the Ito formula

Yt^k () = 1 +

tZ^k

eUs

0

+

1 dUs + 2

X

0<st^k

eUs

tZ^k 0

eUs dhU c is

eUs

eUs Us ; (4.1.22)

where hU c i denotes the predictable quadratic-variation process of the continuous martingale part of U . Noting that hU c i = C and invoking the canonical decomposition of the special semimartingale X with no truncation Xt = X0 + Bt0 + Mt ; where M is a local martingale relative to F , we derive from (4.1.22) after some algebra that

Yt^k () = 1 + +

tZ^kZ 0

tZ^k 0

eUs dMs

eUs ex 1 x ( )(ds; dx)

Rd

X

0<st^k

eUs

eXs 1 + Gs ()

1 Gs (): (4.1.23)

The rst integral on the right of (4.1.23) is a stochastic integral of a locally bounded predictable process with respect to a local martingale, hence, it is a local martingale. The second integral is an integral with to a martingale measure and is a local martingale since R t R respect Us ex 1 x (ds; dx) < 1, Jacod and Shiryaev [67, e 0 Rd II.1.28]. Let us consider the sum on the right of (4.1.23). Firstly, it is

© 2001 by Chapman & Hall/CRC

296

Finite-dimensional LD convergence

absolutely convergent by (4.1.19) and (4.1.20). Secondly, by (4.1.16) X

0<st^k X

=

eUs

=

eXs 1 + Gs ()

eUs eXs

tZ^kZ 0

1

eUs

1 Gs ()

0<st^k X 0<st^k

Gs () 2 1 + Gs ()

ex

eUs

Rd

Gs () 1 + Gs ()

Gs () 1 ( )(ds; dx); 1 + Gs ()

which is a local martingale by Jacod and Shiryaev [67, II.1.28] and the fact that X

0<st^k

eUs

Gs () 2 < 1: 1 + Gs ()

Let (Gt (); t 2 R+ ; 2 Rd ) be an R-valued function, which is continuous in t, dierentiable in , and such that the increments Gt () Gs () are convex functions of 2 Rd for 0 s < t and G0 () = Gt (0) = 0. Let x0 , where x0 2 Rd , be de ned as in Section 2.7 (see (2.7.6)). We recall that it is a deviability on C by Lemma 2.8.3. We denote its extension to a deviability on D with support in C as x0 as well. Let X be the canonical idempotent process on (D ; x0 ). We state the central result of the section.

Theorem 4.1.2. Let the X satisfy the Cramer condition. 2 ,

(0) and for all T > 0 and 2 Rd (sup E )

1=r P X0 ! x0 ; 1=r

P 1 ln Et (r ) Gt () ! 0 ; tT r

sup

© 2001 by Chapman & Hall/CRC

If, as

297

Convergence of stochastic exponentials

ld then X ! X:

We begin the proof with preliminary results. The hypotheses of Theorem 4.1.2 are assumed to hold. We also assume with no loss of generality that x0 = 0. We introduce the F -stopping times ( 2 Rd )

() = inf t 2 R+ : Et (r )1=r _ Et (r ) 1=r 2eGt () or Et (2r )1=r _ Et (2r ) 1=r 2eGt (2) : (4.1.24) By condition (sup E )

lim P 1=r () t = 0; t 2 R+ ; 2 Rd : 2

(4.1.25)

Being the debut of a predictable set whose graph belongs to the set, () is an F {predictable stopping time. Thus, () is P -a.s. announced by an increasing sequence of F {stopping times. Since () > 0, there exist F {stopping times () < () such that

P1=r () +

1 r

() < 1

+ P1=r () r ; () = 1

r1 : (4.1.26)

In view of the inequality

P ( () t) P ( () t + 1)

+ P () +

1 r

(t + 1) ^ () ;

we have from (4.1.25) and (4.1.26) that

lim P1=r () t = 0; t 2 R+ : 2

(4.1.27)

Note also that by (4.1.24) and the inequality () < ()

Et^ () (r)1=r _ Et^ () (r ) 1=r < 2eGt () ; (4.1.28) Et^ () (2r )1=r _ Et^ () (2r) 1=r < 2eGt (2) : (4.1.29)

© 2001 by Chapman & Hall/CRC

298

Finite-dimensional LD convergence

Lemma 4.1.3. For all 2 Rd , the process Y () is a positive supermartingale relative to F . The process (Yt^ () (); t 2 [0; T ]) is a square integrable martingale relative to F for every T > 0 and E Yt^ () (r )2

23r er[Gt (2)+2Gt ()] ; 2 Rd ; t 2 R+ :

(4.1.30)

Proof. By Lemma 4.1.1 Y () is a positive local martingale relative to F. Hence it is a supermartingale. So we have to prove only (4.1.30). By the supermartingale property of Y () and the fact that Y0 () = 1, for all nite F -stopping times ,

E Y () 1; 2 Rd :

(4.1.31)

In view of (4.1.29) and the de nition of Y (), (4.1.31) with 2r implies that E exp 2r (Xt^ () X0 ) 2r exp r Gt (2) ; 2 Rd ; which by (4.1.28) yields

E Yt^ () (r )2

2r exp rGt (2) 22r exp 2rGt () ;

proving the lemma.

Lemma 4.1.4. Let for 0 = t0 < t1 < : : : < tk Zi () = Xti ^ () Xti 1 ^ () ; 2 Rd : n

P

o

Then, for all 1 ; : : : ; k 2 Rd , the net exp ki=1 i Zi (i ) ; 2 is uniformly exponentially integrable with respect to the P and k X 1=r lim E exp r i 2 i=1

k X Zi (i) = exp i=1

Gti (i ) Gti 1 (i ) :

Proof. Let

i =

i X j =1

j Zj (j ); i = 1; : : : ; k; 0 = 0:

© 2001 by Chapman & Hall/CRC

(4.1.32)

299

Convergence of stochastic exponentials

We rst prove that for i = 1; : : : ; k

E exp(2r i ) 22r i

i Y j =1

exp r Gtj (2j )+Gtj 1 (2j ) :

The proof of Lemma 4.1.3 implies that E exp(2r i) < view of (4.1.32) and the de nitions of Zi () and Y (), we Lemma 4.1.3 and (4.1.29) for i = 1; : : : ; k h

E exp(2r i ) j Fti

(4.1.33) 1. In have by

i

1 exp(2r i 1) exp 2r i (Xti 1 ^ (i ) X0 ) Yti 1 ^ (i ) (2r i )2r er Gti (2i ) = exp(2r i 1 )Eti 1 ^ (i ) (2r i ) 1 2r er Gti (2i ) :

Applying to the latter (4.1.29) again we deduce h

E exp(2r i ) j Fti

i

1 exp(2r i 1)22r exp r

Gti (2i ) + Gti 1 (2i ) :

This P proves (4.1.33). Uniform exponential integrability of k exp i=1 i Zi (i ) ; 2 is implied by (4.1.33) with i = k. We prove the convergence required in the lemma by proving that for i = 1; : : : ; k lim E 1=r exp(r i ) = egi ; 2

(4.1.34)

where

gi =

i X j =1

Gtj (j ) Gtj 1 (j ) ; i = 1; : : : ; k; g0 = 0;

provided (4.1.34) holds for (i 1). For Æ 2 (0; 1=2) we de ne the sets n

BÆ = ! 2 : Eti 1 ^ (i ) (r i )1=r e Gti 1 (i ) 1 Æ o or Eti ^ (i ) (r i ) 1=r eGti (i ) 1 Æ

© 2001 by Chapman & Hall/CRC

300

Finite-dimensional LD convergence

and AÆ = n BÆ. By (sup E ) and (4.1.27), as 2 ,

P1=r BÆ

! 0; P1=r AÆ ! 1 :

(4.1.35)

Applying the Cauchy-Schwarz inequality we have by (4.1.33) and (4.1.35) lim E1=r exp(r i )1(BÆ ) = 0; 2

which implies that (4.1.34) would follow from lim inf lim inf E1=r exp(r i)1(AÆ ) Æ!0 2 = lim sup lim sup E1=r exp(r i )1(AÆ ) = egi : (4.1.36) Æ!0 2

Let

RÆ = exp(r i 1 )Yti ^ (i ) (r i )Yti 1 ^ (i ) (r i ) 1 : By Lemma 4.1.3

E RÆ

= E exp(r i 1 );

(4.1.37) so by our assumption

lim E1=r RÆ = egi 1 :

(4.1.38)

2

On the other hand, (4.1.32), (4.1.37), and the de nitions of Y () and Zi() yield

RÆ = exp(r i)Eti ^ (i ) (r i )

E

1 ti 1 ^ (i ) (r i ):

(4.1.39)

Therefore, applying the Cauchy-Schwarz inequality we have in view of (4.1.28), (4.1.35) and (4.1.33) that lim2 E1=r RÆ 1(BÆ ) = 0; which by (4.1.38) obtains lim E 1=r RÆ 1(AÆ ) = egi 1 : 2

(4.1.40)

By de nition, we have that on AÆ

Eti 1 ^ (i )(r i) (1 + Æ)r exp rGti 1 (i) ; Eti ^ (i ) (ri ) 1 (1 + Æ)r exp rGti (i) :

© 2001 by Chapman & Hall/CRC

301

Convergence of stochastic exponentials

Therefore, by (4.1.39)

RÆ 1(AÆ ) exp(r i)(1 + Æ)2r exp r(Gti (i) Gti 1 (i)) 1(AÆ ): The latter implies by (4.1.40) that lim inf lim inf E1=r exp(r i )1(AÆ ) Æ!0 2 egi 1 exp Gti (i)

Gti 1 (i ) = egi :

In an analogous manner, the inequalities

Eti 1 ^ (i )(r i) (1 Æ)r exp rGti 1 (i) ; Eti^ (i )(r i) 1 (1 Æ)r exp rGti (i) yield the limit lim sup lim sup E1=r exp(r i )1(AÆ ) Æ!0 2

egi :

Limits (4.1.36) are proved. The last preliminary result needed for the proof of Theorem 4.1.2 is the following version of Theorem 3.1.31, which is proved in a similar manner.

Lemma 4.1.5. Let fK ; 2 g be a net of Rm {valued random vari-

ables de ned on respective probability spaces ( ; F ; P ) and K be an Rm -valued idempotent variable de ned on an idempotent probability space ( ; ). Let S exp( K ) be nite for and dierentiable in 2 Rm . Let there exist nets fZ (); 2 g of Rm {valued ran1=r P Z () !

dom variables de ned on ( ; F ; P ) such that 0 1=r and E exp r Z () ! S exp( K ) as 2 , and the net fexp Z () ; 2 g is uniformly exponentially integrable for ld 2 Rm . Then K ! K:

K

Now we proceed with the proof of Theorem 4.1.2 itself. Recall that x0 = 0.

© 2001 by Chapman & Hall/CRC

302

Finite-dimensional LD convergence

Proof of Theorem 4.1.2. We apply Theorem 3.2.8 to the net fL(X ); 2 g. We verify the C -exponential tightness by checking the conditions of part II of Theorem 3.2.3. We begin with condition (i). In view of (4.1.31) and the de nition of Y (), we can apply to exp r (Xt X0 ) ; t 2 R+ and E (r ) Lemma 3.2.6 to obtain for all A > 0; B > 0 and L > 0

P sup exp r (Xt tL

X0 )

erA er(B

A)

+ P sup Et (r ) er B ; 2 Rd : (4.1.41) tL Taking B > GL () + 1, we have by (sup E )

lim P 1=r sup Et (r ) er B = 0; 2 tT and then (4.1.41) yields, for 2 Rd , lim sup P1=r sup (Xt X0 ) > A tT 2 Since is arbitrary, this implies that

eB A ! 0 as A ! 1:

lim lim sup P1=r sup jXt X0 j > A = 0: A!1 2 tT 1=r P X0 !

As by hypotheses 0, we obtain (i). Turning to (ii) it is again suÆcient to prove that for all 2 R d ; 6= 0; > 0; and T > 0, lim lim sup sup P1=r sup (Xt+ X ) > = 0: Æ!0 2 2ST (F ) tÆ jj (4.1.42) By Lemma 4.1.3 and Doob's stopping theorem we have for every nite F { stopping times and such that "

#

Y (r ) E 1: Y (r ) Fixing 2 ST (F ), let for t 2 R+

Xt; = Xt+ X ; Et; () = Et+ () E ()

© 2001 by Chapman & Hall/CRC

(4.1.43) (4.1.44a) (4.1.44b)

303

Convergence of stochastic exponentials

and introduce the ltration F; = (Ft+ ; t 2 R+ ). Let be a nite F; {stopping time. Then ( + ) is an F{stopping time so that by (4.1.43) (with = + ), (4.1.44a), (4.1.44b), and the de nition of Y () we have "

E

exp(r X; ) E; (r)

#

1:

As is an arbitrary nite F; {stopping time, by Lemma 3.2.6 we conclude that for all 2 Rd ; > 0; Æ > 0, and > 0

P sup Xt; tÆ

jj er (

)jj +

P sup Et; (r )1=r tÆ

ejj : (4.1.45)

By (4.1.44b) 1 1 sup ln Et; (r ) ln E (r ) G () r tÆ r 1 + sup ln Et+ (r ) Gt+ () + sup jGt+ () G ()j: tÆ r tÆ (4.1.46)

Since Gt () is continuous in t, 1 sup jGt () Gs ()j jj 2 jt sjÆ 0t;sT +Æ for all suÆciently small Æ. Thus, (4.1.46) yields for these Æ by the fact that T

P sup Et; (r )1=r ejj tÆ

P sup r1 ln Et(r ) Gt () j4j : tT +Æ Substituting the right-hand side into (4.1.45) and using (sup E ) we obtain for = 6 0

lim sup lim sup P1=r sup Xt; Æ!0 2 tÆ jj

© 2001 by Chapman & Hall/CRC

e(

)jj :

304

Finite-dimensional LD convergence

Taking = =2 and jj ! 1, and using (4.1.44a) we arrive at (4.1.42). C {exponential tightness of fL(X )g follows. Let us check LD convergence of nite-dimensional distributions. Let 0 = t0 < t1 < : : : < tk , 1 ; : : : ; k 2 Rd , and X denote the canonical idempotent process on D . We also denote m = d k; = (1 ; : : : ; k ) 2 Rm ; K = (Xt1 X0 ; : : : ; Xtk Xtk 1 ); K = (Xt1 X0 ; : : : ; Xtk Xtk 1 ); and Z () = (Z1 (1 ); : : : ; Zk (k )); where the Zi (i ) are de ned in Lemma 4.1.4. Then by (4.1.27)

P1=r jK Z ()j > "

P1=r K 6= Z ()

k X i=1

P1=r (i ) ti

! 0:

By Theorem 2.8.5 if D is equipped with 0 , then X is an idempotent process with independent increments starting at 0 and such that S0 exp (Xt Xs ) = exp Gt () Gs () : By Lemma 4.1.4 fK ; 2 g and fZ (); 2 g; 2 Rm ; satisfy the condild tions of Lemma 4.1.5 so that K ! K: The contraction principle then implies by the de nition of K that Xt1 X0 ; : : : ; Xtk

1=r P ld X0 ! Xt1 X0 ; : : : ; Xtk X0 : Since X0 ! 0 and X0 = 0 0 -a.e., by Lemma 3.1.42 and the contraction principle ld Xt0 ; Xt1 ; : : : ; Xtk ! Xt0 ; Xt1 ; : : : ; Xtk proving the nite dimen-

sional LD convergence.

Remark 4.1.6. By Theorem 2.8.5 under x0 X is an idempotent

process with independent increments starting at x0 and having cumulant G(). In particular, X satis es \the Cramer condition" S exp jXt j < 1 for 2 R+ and t 2 R+ .

We now give two more versions of the theorem in which the X do not have to be semimartingales. Inspection of the above proof shows that the critical property of E () is the one stated in Lemma 4.1.1 that the process Y () = (Yt (); t 2 R+ ); 2 Rd ; is a local martingale. Therefore, the following extension of Theorem 4.1.2 holds. The function Gt () satis es the same conditions as above.

Theorem 4.1.7. Let X ; 2 ; be stochastic processes with paths in D de ned on respective stochastic bases ( ; F ; F ; P ). Let for © 2001 by Chapman & Hall/CRC

305

Convergence of characteristics

every 2 Rd and 2 there exist F {predictable positive pro cesses E () = Et (); t 2 R+ , E0 () = 1, such that the processes Y () = Yt (); t 2 R+ de ned by

Yt () = exp (Xt X0 )

are F {local martingales. If

1=r P X0 ! x0

1 sup r

tT

Et ()

1

and, for all T > 0 and 2 Rd ,

ln Et (r )

ld as 2 , then L(X ) !

1=r P Gt ()

! 0;

x0 :

As a consequence, we obtain the following result for processes with independent increments that are not necessarily semimartingales.

Theorem 4.1.8. Let X be processes with independent increments with paths in D such that E exp (Xt and 2 Rd . If

1=r P X0 ! x0

1 sup r

tT

X0 ) < 1 for all t 2 R+

and, for all T > 0 and 2 Rd ,

ln E exp

(Xt

X0 )

Gt ()

! 0;

x0 : Remark 4.1.9. If the X ld then L(X ) !

are semimartingales with independent increments, then the assertions of Theorems 4.1.2 and 4.1.8 coincide since the triplets(B 0 ; C ; ) are deterministic and Et () = E exp (Xt X0 ) (cf. Jacod and Shiryaev [67, II.4.15]).

4.2 Convergence of characteristics In this section we give results on LD convergence in terms of the characteristics of the semimartingales. This allows us to do without the Cramer condition; we require instead that the measure of \big

© 2001 by Chapman & Hall/CRC

306

Finite-dimensional LD convergence

jumps" be small. As in the preceding section, we consider a net fX ; 2 g of semimartingales on ( ; F ; F; P ) with predictable triplets (B ; C ; ) corresponding to a limiter h(x). We also consider as given a cumulant Gt (), which does not depend on x and is de ned as in (2.7.7) and (2.7.55) by

Gt () =

Zt

gs () ds; 2 Rd ; t 2 R+ ;

(4.2.1)

0

where 1 gs () = bs + cs + 2

+ ln 1 +

Z

(ex

Rd

Z

(ex

Rd

1)^s (dx)

1 x)s (dx)

Z

(ex

Rd

1)^s (dx) ; (4.2.2)

(Rbs ; s 2 R+ ) is an Rd -valued Lebesgue-measurable function such that t 0 jbs j ds < 1 for t 2 R + , (cs ; s 2 R+ ) is a Lebesgue-measurable function with values in the space of symmetric, positive semi-de nite d d-matrices such that Rt 0 kcs k ds < 1 for t 2 R + , (s ( ); s 2 R+ ; 2 B(Rd )) is a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for t 2 R+ and 2 R+

t (f0g) = 0; Z Rd

Z

jxj2 ^ 1 t (dx) < 1;

Rd

ejxj 1(jxj > 1) t (dx) < 1;

jxj2 ^ 1 t < 1; ejxj 1(jxj > 1) t < 1; (4.2.3) ^s( ); s 2 R+ ; 2 B(Rd ) is a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for s 2 R+ and 2 B(Rd ) ^s ( ) s( ); ^s(Rd ) 1: (4.2.4) We also assume the following condition to hold

© 2001 by Chapman & Hall/CRC

307

Convergence of characteristics

sup ejxj ^s < 1; t 2 R+ ; 2 R+ : st By Lemma 2.8.8 condition (L1 ) implies condition (L0 ) of Corollary 2.8.7. We recall that Gt () is dierentiable in under (L0 ). Let x0 2 Rd . By Corollary 2.8.7 x0 is a deviability on D and the canonical idempotent process X is a Luzin-continuous semimaxingale with independent increments on (D ; x0 ) starting at x0 and having local characteristics (b; c; ; ^). As in Section 2.7, we denote as (B; C; ; ^) the characteristics of X associated with a limiter h(x); B 0 denotes the nontruncated rst characteristic and C~ denotes the modi ed second characteristic, which are de ned by (2.7.11), (2.7.57), and (2.7.58). To recall, (L1 )

B0 = t

Zt

0

bs ds; Bt = Bt0 +(h(x) x)t ; Ct =

Zt

cs ds;

0

and for 2 Rd

C~t = Ct + ( h(x))2 t

Zt

( h(x) ^s )2 ds:

0

Let U be a dense subset of R+ and let Cb denote the set of functions f : Rd ! Rd that are bounded, continuous and equal to 0 in a neighbourhood of the origin. For f 2 Cb , we denote f (x) = f (r x)=r . We consider the following conditions (0) (A) (a)

1=r P X0 ! x0 as 2 ; 1=r lim lim sup P ([0; t]; jxj > A)1=r > " = 0; t > 0; A!1 2 1=r 1 r jxj lim lim sup P e (r jxj > a) (jxj A) t a!1 2 r

(sup B )

1

sup jBt tT

© 2001 by Chapman & Hall/CRC

Bt j

1=r P

1

" > 0; >"

= 0; t > 0; > 0; A > 0; " > 0:

! 0 as 2 ; T > 0;

308

Finite-dimensional LD convergence

(C )

lim lim sup P1=r Æ!0 2

(C~ )

kr C~t

( ) (^ )

kr Ct;Æ Ct k > " = 0; t 2 U; " > 0; C~t k

f (x) t f (x) t 1 X f (r x) s k r 0<st

1=r P

! 0 as 2 ; t 2 U;

1=r P

! 0 as 2 ; t 2 U; f 2 Cb;

Zt

f (x) ^s k ds

1=r P

! 0 as 2 ;

0

k = 2; 3; : : : ; t 2 U; f 2 Cb : The following theorem is the main result of the section. Theorem 4.2.1. Let the limiter h(x) be continuous. If conditions ld (0), (A) + (a), (sup B ), (C ) (or (C~ )), ( ), and (^) hold, then X ! X. Remark 4.2.2. Conditions (C ) and (C~ ) equivalently require convergence of the entries of the respective matrices r Ct;Æ and r C~t to the corresponding entries of Ct . Remark 4.2.3. The ldassertion of the theorem is equivalent to the LD convergence L(X ) ! x0 . We also recall that by Lemma 2.7.12 under the hypotheses

x0 (x) = exp

Z1

0

sup x_ t gt () dt 2Rd

if x is absolutely continuous and x0 = x0 , and x0 (x) = 0 otherwise.

Remark 4.2.4. The theorem also holds if condition (a) is replaced by the following weaker condition

(a0 )

where

1=r 1 lim lim sup P j (x) ;c a!1 2 r a;A; t 1 X (x) > " = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0;

(x) = (er jxj 1) 1(r jxj > a) 1(jxj A): ja;A;

© 2001 by Chapman & Hall/CRC

309

Convergence of characteristics

Before proceeding with a proof we give a version for processes with independent increments (PII) that are not necessarily semimartingales. Let X be PII. Then, given a limiter h(x), the X admit decomposition (4.1.2), see Jacod and Shiryaev [67, II.5], where B = (Bt ; t 2 R+ ); B0 = 0; is an Rd {valued right-continuous with left limits (deterministic) function; X ;c = (Xt;c ; t 2 R+ ); X0;c = 0; is an Rd {valued continuous local martingale with respect to F that is the continuous martingale part of X ; is the measure associated with jumps of X ; is a (deterministic) measure on (R+ Rd ; B(R+ ) B(Rd )) that is the F{compensator of . Also, relations (4.1.3a) and (4.1.3c) hold. Let C = (Ct ; t 2 R+ ); C0 = 0; be the F {predictable quadraticvariation process of X ;c. Then C is a deterministic Rdd {valued continuous function such that the matrices Ct Cs are symmetric and positive semi-de nite for s t. As above, we denote by C ;Æ and C~ the F -predictable quadratic-variation processes of the respective local martingales M ;Æ and M from the respective equalities (4.1.5) and (4.1.6). As with C , the processes C ;Æ and C~ are actually deterministic matrix-valued functions. Since X is not necessarily a semimartingale, the function B might no longer have bounded variation over bounded intervals and condition (4.1.3b) is not in general satis ed; so, one cannot specify C ;Æ and C~ by the respective equalities (4.1.7) and (4.1.8). Instead, we have for 2 Rd and t 2 R+

Ct;Æ = Ct + x 1(r jxj Æ) x 1(r jxj Æ)s 2 t X + x 1(r jxj Æ) s 2 1 (fsg; R d ) (4.2.5)

0<st

and C~t = Ct + h (x) h (x) s 2 t X + h (x) s 2 1 (fsg; R d ) : (4.2.6)

0<st

© 2001 by Chapman & Hall/CRC

310

Finite-dimensional LD convergence

The right-hand sides are well de ned since by Jacod and Shiryaev [67, II.5.6] h (x)

Bs 2 t +

X

jBsj2 1 (fsg; Rd ) < 1:

0<st

(4.2.7) Formulas (4.2.5) and (4.2.6) reduce to (4.1.7) and (4.1.8) if (4.1.3b) holds. Since the characteristics of a PII are deterministic, conditions (A), (a), (a0 ), (sup B ), (C ), (C~ ), ( ), and (^ ) take the form (A)I (a)I (a0 )I

(sup B )I (C )I (C~ )I ( )I (^ )I

lim lim sup ([0; t]; jxj > A)1=r = 0; t > 0;

A!1 2

1 r jxj e 1(rjxj > a) 1(jxj A) t = 0; r t > 0; > 0; A > 0; " > 0: 1 lim lim sup j (x) t;c a!1 2 r a;A; 1 X (x) = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0;

lim lim sup

a!1 2

sup jBt Bt j ! 0 as 2 ; T > 0; tT lim lim sup kr Ct;Æ Ct k = 0; t 2 U; Æ!0 2 lim kr C~t C~t k = 0; t 2 U; 2 f (x) t f (x) t ! 0 as 2 ; t 2 U; f 2 Cb ; 1 X f (r x) s k r 0<st

Theorem 4.2.5.

Zt

0

f (x) ^s k ds ! 0 as 2 ; k = 2; 3; : : : ; t 2 U; f

2 Cb :

Let X be PII with predictable characteristics (B ; C ; ) corresponding to a continuous limiter h(x). Let conditions (0), (A)I , (a)I (or (a0 )I ), (sup B )I , (C )I (or (C~I )), ( )I , ld and (^ )I hold. Then X ! X as 2 .

© 2001 by Chapman & Hall/CRC

311

Convergence of characteristics

We prove Theorems 4.2.1 and 4.2.5 in parallel. The argument actually refers to the X being semimartingales. The modi cations needed when the X are PII are either self-evident (e.g., replacing super-exponential convergence in probability by deterministic convergence) or explicitly mentioned. The proof proceeds through a number of steps: we rst establish interconnections between the conditions of the theorem, then derive the assertions of the theorems for the case of jumps of order r 1 and nally consider the general setting.

Lemma 4.2.6. Let Z ;Æ

= (Zt;Æ ; t 2 R+ ); Z0;Æ = 0; 2 ; Æ > 0; be R+ -valued increasing processes on respective probability spaces ( ; F ; P ) and Z = (Zt ; t 2 R+ ); Z0 = 0, be a (deterministic) R+ valued increasing continuous function. If, for all t 2 U and " > 0,

lim lim sup P1=r jZt;Æ Zt j > " = 0; Æ!0 2 then this convergence is uniform so that

lim lim sup P1=r sup jZt;Æ Zt j > " = 0; T > 0; " > 0: Æ!0 2 tT Proof. The argument is standard. Let wt (Æ) denote the modulus of continuity of Z on [0; t], i.e., wt (Æ) = sup u;vt: jZu Zv j: For N 2 N , ju vjÆ we choose tNi 2 U; i = 0; : : : ; kN , such that 0 = tN0 < tN1 < : : : < tNkN 1 < T tNkN < T +1 and jtNi tNi 1 j 1=N; i = 1; : : : ; kN . Then, since the Z ;Æ and Z are increasing, for t 2 [tNi 1 ; tNi ]; i = 1; : : : ; kN , we have that

jZt;Æ Ztj jZt;Æ ZtNi 1 j _ jZt;Æ ZtNi j N N i i 1 ZtNi 1 j + jZtNi ZtNi 1 j; jZt;Æ ZtNi j _ jZt;Æ N N i 1 i and hence using that Z0;Æ = Z0 = 0

1 sup jZt;Æ Zt j max N jZt;Æ Z j + w : N T +1 N ti i N i=1;:::;k tT Since wT +1 (1=N ) ! 0 as N ! 1 by continuity of Z , for arbitrary " > 0 we have for all N large enough

P1=r

sup jZt;Æ tT

© 2001 by Chapman & Hall/CRC

Zt j > "

kN X i=1

P1=r jZt;Æ j > "=2 : N ZtN i i

312

Finite-dimensional LD convergence

The latter goes to 0 as choice of the tNi .

2 and Æ ! 0 by hypotheses and the

As a consequence, we have the following. Corollary 4.2.7. Conditions (C ), (C~ ), and ( ) are equivalent to the following respective conditions (sup C ), (sup C~ ), and (sup ). (sup C )

lim lim sup P1=r sup kr Ct;Æ Æ!0 2 tT

(sup C~ )

lim P 1=r sup kr C~t 2 tT

(sup )

sup f (x)

tT

t

Ct k > " = 0;

" > 0; T > 0;

C~t k > " = 0; " > 0; T > 0;

f (x)

P 1=r t

! 0 as 2 ; T > 0; f 2 Cb :

The second preliminary lemma shows that if condition ( ) holds, then condition (sup B ) is invariant with respect to the choice of a continuous limiter h.

Lemma 4.2.8. If condition ( ) holds, then condition (sup B ) does not depend on the choice of a continuous limiter h(x).

Proof. Let B = (Bt ; t 2 R+ ) and B = (Bt ; t 2 R+ ) be the rst characteristics of X corresponding to continuous limiters h(x) and h (x), respectively. Let Bt and Bt be the rst characteristics of X associated with h(x) and h (x), respectively. By (4.1.4), up to a P null set,

Bt = Bt +(h (x) h (x)) t ; so that by the de nitions of Bt and Bt , h (x) and h (x), up to a P -null set,

Bt Bt = (Bt Bt )+(h (x) h (x))t (h (x) h(x))t ; and the equivalence of (sup B ) and (sup B ) under ( ) follows by the equivalence of ( ) and (sup ), and the fact that h h 2 Cb . We now consider implications of conditions ( ) and (^ ).

© 2001 by Chapman & Hall/CRC

313

Convergence of characteristics

Lemma 4.2.9. Let ( ) and (^ ) hold. Then for " > 0 and t 2 U 1 1. lim lim sup P1=r f (r x) 1(r jxj > Æ) t Æ!0 2 r

f (x) t > " = 0; for all R+ -valued bounded and continuous functions f (x); x 2 R d ; such that f (x) cjxj2 ; c > 0; in a neighbourhood of the origin; Æ 2. lim lim sup P1=r j g(r x)j 1(r jxj > Æ) t > " = 0 Æ!0 2 r and 1 X 3. lim lim sup P1=r g(r x) 1(r jxj > Æ) s k Æ!0 2 r 0<st Zt

g(x) ^s k ds > " = 0;

0

k = 2; 3; : : : ; for all R-valued bounded and continuous functions g(x); x 2 Rd ; such that jg(x)j cjxj; c > 0; in a neighbourhood of the origin. Proof. Let

fr (x) =

jxj 1+ ^ 1; r > 0; x 2 Rd : r

Then, since f (x) 0 and fÆ=2 (x) 1(jxj > Æ) fÆ (x), 1 r

f (r x) 1

(r jxj > Æ) t

f (x)

t

f (x) 1(jxj Æ) t + f (x)(fÆ=2 (x) fÆ (x)) t 1 + max f (r x)fÆ=i (r x) t f (x)fÆ=i (x) t : i=1;2 r

The last term on the right goes to 0 super-exponentially in probability as 2 by ( ) and the inclusion f (x)fr (x) 2 Cb . The sum of the two other terms does not exceed, for Æ small enough, 2cjxj2 1(jxj 2Æ) t ; which goes to 0 as Æ ! 0 by (4.2.3) and Lebesgue's dominated convergence theorem. Part 1 is proved.

© 2001 by Chapman & Hall/CRC

314

Finite-dimensional LD convergence

Now we prove part 2. By ( ), using that 1(jxj > Æ) and g(x)fÆ=2 (x) 2 Cb ,

fÆ=2 (x)

Æ lim sup P1=r jg(r x)j 1(r jxj > Æ) t > " r 2 lim sup P1=r rÆ jg(r x)jfÆ=2 (rx) t > " 2 1 Æjg(x)jfÆ=2 (x) t > "=2 : (4.2.8)

Now, for > Æ=2, since fÆ=2 (x) 1(jxj > Æ=2),

Æjg(x)jfÆ=2 (x) t Æjg(x)j 1(jxj > ) t + Æjg(x)j 1(Æ=2 < jxj ) t : The rst term on the right, obviously, goes to 0 as Æ ! 0. The second one, by the assumptions on g(x) and with the use of Chebyshev's inequality, is not greater than (take small enough) 2jg(x)jjxj 1(jxj ) t 2cjxj2 1(jxj ) t ; and goes to 0 as ! 0, as above. Thus, the right-hand side of (4.2.8) is zero for Æ small enough and part 2 is proved. We prove part 3. By (^ ) and the inclusion g(x)fÆ (x) 2 Cb , the required would follow by lim lim sup P1=r Æ!0 2

k 1 X g (r x) 1(r jxj > Æ ) s r 0<st k " g(r x)fÆ (r x) s > = 0 (4.2.9) 2

and Zt lim Æ!0 0

g(x) ^s

k

g(x)fÆ (x) ^s

k ds

= 0:

The validity of the latter limit is obvious since g(x) and fÆ (x) are uniformly bounded, fÆ (x) ! 1 as Æ ! 0 for x 6= 0, and (4.2.4) holds. For (4.2.9), we write, by the inequalities jxk yk j k(x _ y)k 1 jx

© 2001 by Chapman & Hall/CRC

315

Convergence of characteristics

yj; x; y > 0, and 1(jxj > 2Æ) fÆ (x) 1(jxj > Æ), k k 1 X g ( r x ) 1 ( r j x j > Æ ) g ( r x ) f ( r x ) Æ s s r 0<st X jg(r x)j 1(r jxj > Æ) s k 1 rk 0<st jg(r x)j 1(r jxj 2Æ) s : (4.2.10)

Applying to the rst integral on the right of (4.2.10) Jensen's inequality and recalling that (fsg; R d ) 1; we conclude that, for Æ small enough, the right-hand side of (4.2.10) is not greater than 2Æc

k X jg(r x)jk r 0<st

1

1(r jxj > Æ) s ;

and an application of the assertion of part 2 yields (4.2.9). Part 3 is proved.

Lemma 4.2.10. Under ( ) and (^), conditions (C ) and (C~ ) are equivalent.

Proof. By the de nitions of Ct;Æ , C~t, C~t , and h (x), it suÆces to prove that for t 2 U; " > 0, 1 lim lim sup P1=r ( h(r x))2 1(r jxj > Æ) t Æ!0 2 r

( h(x))2 t > " = 0;

Æ X jh(r x)j 1(r jxj > Æ) s > " = 0; r 0<st 1 X h(r x) 1(r jxj > Æ) s 2 lim lim sup P1=r Æ!0 2 r 0<st

lim lim sup P1=r Æ!0 2

Zt

h(x) ^s 2 ds > " = 0:

0

The limits follow by the respective parts 1, 2 and 3 of Lemma 4.2.9.

© 2001 by Chapman & Hall/CRC

316

Finite-dimensional LD convergence

4.2.1 The case of small jumps In this subsection we prove Theorems 4.2.1 and 4.2.5 for the case of jump size of order 1=r .

Theorem 4.2.11. Let the X be semimartingales (respectively, PII). Let conditions (0), (sup B ), (C ) (or (C~ )), ( ), and (^) (respectively, (0), (sup B )I , (C )I (or (C~ )I ), ( )I , and (^ )I ) hold, and, in addition, for some a > 0,

([0; t]; fr jxj > ag) = 0; t > 0; 2 :

( F )

ld Then X ! X as 2 .

Proof. We prove the theorem by checking the hypotheses of Theorem 4.1.2 in the semimartingale case, respectively, Theorem 4.1.8 in the PII case. By condition ( F ) the Cramer condition (Cr) is met by the X so that the associated stochastic exponentials E () if the X are semimartingales, respectively, the E exp (Xt X0 ) if the X are PII, are well de ned. For economy of notation we denote the latter expectation by Et () in the PII case as well. Since also the cumulant Gt () satis es the conditions of Theorem 4.1.2, by Theorem 4.1.2 (respectively, Theorem 4.1.8) in order to prove Theorem 4.2.11 it is suÆcient to check that as 2 1 ln Et (r ) Gt () tT r

sup

(sup E )

1=r P

! 0:

(In the PII case the convergence is deterministic.) Let us rstly note that by conditions ( F ) and ( )

t (jxj > a) = 0 (a.e.)

(4.2.11)

We choose h(x) = x for jxj a so that by ( F ) we have B = B 0 . In view of (4.1.14), (4.1.15) and (4.1.16) we can write

Et() = exp Bt + 21 Ct + (ex 1 x) t;c Y e Bs 1 + (ex 1) s : (4.2.12) st

© 2001 by Chapman & Hall/CRC

317

Convergence of characteristics

(In the PII case one needs to use the argument of the proofs of Jacod and Shiryaev [67, Theorems II.4.15 and II.5.2].) We show that the right-hand side of (4.2.12) is well de ned. Since ejxj 1 jxj t;c < 1; 2 R+ ; (4.2.13) all the individual terms are well de ned. To show that the product is convergent, note that Bs = x s so that e Bs 1+(ex 1) = 1+ e(x Bs ) (x B ) 1 s Bs

+ Bs

s s d (fsg; R ) :

+ e 1 1 Thus, the expression on the left-hand side is not less than 1. Also, by the inequalities 0 exp(u) 1 u exp(juj)juj2 =2; u 2 R; and jBsj a=r , condition ( F ) and the choice of h(x) X

ln e Bs 1 + (ex

0<st X 0<st

1) s

j2j e2jjar 2

jh (x) Bsj2 s + jBsj2 1 (fsg; Rd ) ;

the latter sum being convergent by (4.1.3b) (respectively, by (4.2.7)). Let us denote for s 2 R+ as = (fsg; R d ); (4.2.14) and for Æ > 0, 2 Rd , x;Æ = x 1(r jxj Æ) s ; (4.2.15a) s ;Æ x Ds () = (e 1) 1(r jxj > Æ) s ; (4.2.15b) F (well de ned by ( )), ;Æ Rs () = exp( (x 1(r jxj Æ) x;Æ s )) 1 (x 1(r jxj Æ) x;Æ (4.2.15c) s ) s ; ;Æ ;Æ ;Æ Qs () = (exp( xs ) 1 + xs )(1 as );(4.2.15d) ;Æ ;Æ ;Æ G;Æ (4.2.15e) s () = exp( xs )Ds () + Rs () ;Æ +Qs (); (4.2.15f) ;Æ ;c x Ut () = (e 1 x) 1(r jxjÆ) t (4.2.15g) (well de ned by (4.2.13)); ;Æ Vt () = (ex 1 x) 1(r jxj>Æ) t (4.2.15h) F (well de ned by ( )):

© 2001 by Chapman & Hall/CRC

318

Finite-dimensional LD convergence

For the sequel, we put down the following obvious relations: 0 as 1; jx;Æ s j

Æ r

(4.2.16)

and

exp

jja 1 D;Æ () exp jja 1 s r r

(4.2.17)

(use ( F ) and (4.1.3a)).

Lemma 4.2.12. The following representation holds (LS )

ln Et () = Bt + Vt;Æ () + Yt;Æ () + Zt;Æ ();

where

Yt;Æ () =

Zt;Æ () =

X

0<st

ln(1+ Ds;Æ ()) Ds;Æ () ;

1 ;Æ ln(1 + G;Æ s ()) + Ut () + 2 Ct 0<st X

X

0<st

ln(1 + Ds;Æ ()):

Proof. The key is to observe that ;Æ 1+(ex 1)s = exp(x;Æ s )(1+ Gs ());

(4.2.18)

which follows by routine calculations using (4.2.14){(4.2.15e). Substituting the right-hand side into (4.2.12) and taking into account (4.1.10), (4.2.15g) and (4.2.15h) yields

Et() = exp Bt + 12 Ct + Ut;Æ () + Vt;Æ () X

0<st

© 2001 by Chapman & Hall/CRC

Y

Ds;Æ ()

st

(1 + G;Æ s ()); (4.2.19)

319

Convergence of characteristics

which is equivalent to (LS ) provided the right-hand sides of (4.2.19) and (LS ) are well de ned, i.e., X

jDs;Æ ()j < 1;

(4.2.20a)

1 + G;Æ s () > 0; ;Æ j ln(1 + Gs ())j < 1;

(4.2.20b) (4.2.20c)

1 + Ds;Æ () > 0; j ln(1 + Ds;Æ ())j < 1:

(4.2.20d) (4.2.20e)

0<st

X

0<st X

0<st

Inequality (4.2.20a) follows by (4.2.15b), ( F ) and the fact that by (4.1.3b) ([0; t]; fjxj > "g) < 1; " > 0. Inequality (4.2.20b) follows since by (4.2.18), ( F ) and (4.2.16) 1 + G;Æ s () exp

jjÆ=r 1 + exp( jja=r ) 1 as exp jj(Æ + a)=r : (4.2.21)

Next, by (4.2.15e) and (4.2.16) X

0<st

jjÆ=r jG;Æ s ()j e

X

jDs;Æ ()j

0<st X

+

(Rs;Æ () + Q;Æ s ()) (4.2.22)

0<st

(note that Rs;Æ () 0 and Q;Æ s () 0). By (4.2.15c) and (4.2.15d), using (4.2.16) and the inequality exp(u) 1 u (juj2 =2) exp(juj); u 2 R, we have X

(Rs;Æ () + Q;Æ s ())

0<st e2jjÆ=r X

2

0<st

(x 1(r jxj Æ) x;Æ s )

2

s

2 + x;Æ (1 s

as ) ;

whichPis nite by (4.1.3b). In view of (4.2.20a) we thus have that 0<st jG;Æ s ()j < 1, which implies (4.2.20c) by (4.2.21),

© 2001 by Chapman & Hall/CRC

320

Finite-dimensional LD convergence

(4.2.15e), (4.2.17), (4.2.16), and the fact that Rs;Æ () and Q;Æ s () are non-negative. Inequality (4.2.20d) follows by (4.2.15b). Finally, inequality (4.2.20e) follows by (4.2.20a) and the left-hand side of (4.2.17). Now we give a similar representation for Gt (). Let Vt () = (ex 1 x) t ; (4.2.23)

Yt () =

Zt

ln 1 + (ex

1) ^s

(ex

1) ^s ds; (4.2.24)

0

1 Zt () = Ct : (4.2.25) 2 By (4.2.11) and the choice of h(x) (recall we take h(x) = x; jxj a) (LS )

Gt () = Bt + Vt () + Yt () + Zt ():

Decompositions (LS ) and (LS ) show that (sup E ) would follow if for every T > 0 and " > 0

)

sup jBt tT

Bt j

1=r P

! 0 as 2 ;

1 lim lim sup P1=r sup j Vt;Æ (r ) Vt ()j > " = 0; Æ!0 2 tT r 1

) lim lim sup P1=r sup j Yt;Æ (r ) Yt ()j > " = 0; Æ!0 2 tT r 1 Æ) lim lim sup P1=r sup j Zt;Æ (r ) Zt ()j > " = 0: Æ!0 2 tT r Part ) is just condition (sup B ). By (4.2.15h), (4.2.23), ( F ), and (4.2.11), part 1 of Lemma 4.2.9 yields

)

1 lim lim sup P1=r j Vt;Æ (r ) Vt ()j > " = 0; t 2 U; " > 0; Æ!0 2 r and an application of Lemma 4.2.6 proves part ). We now prove part ). Let

(x) = x ln(1+ x); x > 1:

© 2001 by Chapman & Hall/CRC

(4.2.26)

321

Convergence of characteristics

By the de nitions of Yt;Æ () and Yt () (see Lemma 4.2.12 and (4.2.24))

Yt;Æ () = Zt

Yt () =

X

0<st

(Ds;Æ ());

(4.2.27)

(ex

1) ^s ds:

(4.2.28)

0

Since (x) > 0, an application of Lemma 4.2.6 implies that ) would follow by 1 lim lim sup P1=r j Yt;Æ (r ) Yt ()j > " = 0; t 2 U; " > 0: Æ!0 2 r (4.2.29) Let u = exp( jja) 1 and v = exp(jja) 1. Since the function (x)=x2 is continuous on [u; v], by Weierstrass' theorem it can uniformly be approximated on [u; v] by polynomials, so, given arbitrary > 0, there exists a polynomial q (x) with powers not less than 2 such that j (x) q (x)j < x2 ; x 2 [u; v]: Now, by (4.2.17) Ds;Æ (r ) 2 [u; v], and by (4.2.11) and (4.2.4) (ex 1) ^s 2 [u; v] (a.e.) Thus, recalling (4.2.27) and (4.2.28),

P1=r

j r1 Yt;Æ (r) Yt()j > "

1=r 1 X P q Ds;Æ (r ) r 0<st Zt " q (ex 1) ^s ds > 3 0 1 X " + P1=r Ds;Æ (r )2 > r 0<st 3

+

1

Zt

0

(ex

1) ^s 2 ds >

" ; 3

and since can be taken arbitrarily small and the smallest power in

© 2001 by Chapman & Hall/CRC

322

Finite-dimensional LD convergence

q is not less than 2, (4.2.29) is implied by 1 X lim lim sup P1=r D;Æ (r )k Æ!0 2 r 0<st s Zt

(ex

0

lim lim sup lim sup P1=r 2

A!1 Æ!0

1) ^s k ds > = 0;

> 0; t 2 U; k = 2; 3; : : : ; (4.2.30a) 1 X ;Æ Ds (r )2 > A r 0<st = 0: (4.2.30b)

Limit (4.2.30a) follows by part 3 of Lemma 4.2.9 in view of (4.2.15b), ( F ) and (4.2.11). The lim sup2 in (4.2.30b) being by (^ ) not R t x 2 greater than 1 0 (je 1j ^s ) ds > A=2 for all Æ > 0, equals 0 for all large A. Limit (4.2.29) is proved. Part ) is proved. We prove part Æ). Let us denote 1 ;Æ 2 L;Æ s () = 2 (x 1(r jxj Æ ) xs ) s ; (4.2.31a) 1 2 (4.2.31b) Ks;Æ () = ( x;Æ s ) (1 as ); 2 ;Æ Hs;Æ () = L;Æ (4.2.31c) s () + Ks (); 1 Wt;Æ () = ( x)2 1(r jxj Æ) t;c: (4.2.31d) 2 Then by (4.1.7), (4.2.15a) and (4.2.14) X 1 ;Æ 1 Ct = Ct +Wt;Æ ()+ Hs;Æ (): (4.2.32) 2 2 0<st Hence, in view of the de nitions of Zt;Æ () (see Lemma 4.2.12) and Zt () (see (4.2.25)), and the fact that by Corollary 4.2.7 condition (sup C ) holds, Æ) would follow from 1 Æ0 ) lim lim sup P1=r sup j (Ut;Æ (r ) Wt;Æ (r ))j > " = 0; Æ!0 2 tT r 1 X Æ00 ) lim lim sup P1=r ln(1 + G;Æ s (r )) Æ!0 2 r 0<st Hs;Æ (r ) + ln(1 + Ds;Æ (r )) > " = 0:

© 2001 by Chapman & Hall/CRC

323

Convergence of characteristics

Let us note that condition (C ) implies that

lim lim sup lim sup P1=r r Ct;Æ > A = 0: A!1 Æ!0 2

(4.2.33)

For limit Æ0 ), we note that by the inequality jeu 1 u (ejuj =6)juj3 ; (4.2.15g) and (4.2.31d) ejjÆ sup jUt;Æ (r ) Wt;Æ (r )j jjÆ WT;Æ (r ): 3 tT

u2 =2j (4.2.34)

Next, since Hs;Æ () and Ct are non-negative, by (4.2.32) WT;Æ () Ct;Æ =2, so that by (4.2.33) 1 lim lim sup lim sup P1=r WT;Æ (r ) > A = 0; A!1 Æ!0 r 2

which together with (4.2.34) implies Æ0 ). We prove Æ00 ). Let us rst note that for t 2 R+ ; " > 0 and 2 Rd by part 2 of Lemma 4.2.9, ( F ) and (4.2.15b) Æ X lim lim sup P1=r j Ds;Æ (r )j > " = 0: Æ!0 2 r 0<st

(4.2.35)

Next, (4.2.15e) implies by Taylor's formula ;Æ ln(1+ G;Æ s (r )) = ln(1+ Ds (r ))+

Ts ; Fs

where ;Æ Ts = (exp( r x;Æ s ) 1)Ds (r ) +Rs;Æ (r ) + Q;Æ s (r ); ;Æ Fs = 1 + Ds (r ) + Ts ; 0 1;

(4.2.36) (4.2.37)

and thus 1 X ln(1 + G;Æ s (r )) r 0<st

© 2001 by Chapman & Hall/CRC

Hs;Æ (r ) + ln(1 + Ds;Æ (r ))

A1 + A2 + A3 ;

324

Finite-dimensional LD convergence

where

A1 =

1 X jDs;Æ (r )j j exp( r x;Æ s ) 1j r 0<st Fs

A2 A3

+Hs;Æ (r ) ; 1 X jTs j ;Æ = H (r ); r 0<st Fs s 1 X 1 = (jR;Æ (r ) L;Æ s (r )j r 0<st Fs s ;Æ +jQ;Æ s (r ) Ks (r )j)

(4.2.38a)

(4.2.38b)

00

(4.2.38c)

(for the latter equality recall (4.2.31c)). We thus prove Æ ) by proving that

lim lim sup P1=r Ai > " = 0; " > 0; i = 1; 2; 3: Æ!0 2 We begin with some estimates. Since by (4.2.16) j exp( r x;Æ ) 1j jjÆejjÆ ; s

(4.2.39)

(4.2.40)

we have, using (4.2.17), ;Æ jjÆ jja j exp( rx;Æ s ) 1jjDs (r )j jjÆe e :

(4.2.41)

From (4.2.15c) and (4.2.15d) using (4.2.16) and the inequality exp(u) 1 u (juj2 =2) exp(juj); we have 1 2 2 jjÆ jRs;Æ (r)j 2jj2 Æ2 e2jjÆ ; jQ;Æ s (r )j 2 jj Æ e ; whereafter in view of (4.2.36) and (4.2.41) jT j 4jjÆe2jja provided jjÆ 1; Æ a: (4.2.42) s

Further, the left inequality in (4.2.17) and (4.2.37) yield, in view of (4.2.42), 1 1 Fs e jja provided Æjj e 3jja ; Æ a: (4.2.43) 2 8 Besides, (4.2.31a){(4.2.31c) and (4.2.16) imply

Hs;Æ (r ) 3jj2 Æ2 :

© 2001 by Chapman & Hall/CRC

(4.2.44)

Convergence of characteristics

325

Now, (4.2.39) for i = 1 follows by (4.2.38a), (4.2.35), (4.2.43), (4.2.40), and (4.2.44). Next, since by (4.2.32) 1 X ;Æ 1 Hs (r ) r Ct;Æ ; r 0<st 2

(4.2.45)

by (4.2.42), (4.2.43) and (4.2.38b) for Æ small enough

A2 4jjÆe3jja r Ct;Æ ; and (4.2.39) for i = 2 follows by (4.2.33). Finally, let i = 3. From (4.2.15c), (4.2.31a), (4.2.15d), and (4.2.31b) analogously to (4.2.34) 2Æjj 2Æjj ;Æ jRs;Æ (r) L;Æ e Ls (r ); s (r )j 3 Æjj Æjj ;Æ ;Æ jQ;Æ e Ks (r ); s (r ) Ks (r )j 3 and hence in view of (4.2.38c), (4.2.43) and (4.2.31c) for Æ small enough

A3 2Æjje2Æjj ejja

1 X ;Æ H (r ); r 0<st s

so that (4.2.45) and (4.2.33) yield (4.2.39) for i = 3. Part Æ00 ) is proved. Limit (sup E ) and with it Theorem 4.2.11 are proved.

4.2.2 The general case In this subsection we prove Theorems 4.2.1 and 4.2.5. The proof relies heavily on the theory of weak convergence for deviabilities. Since by Lemma 4.2.10 conditions (C ) and (C~ ) are equivalent under conditions ( ) and (^ ), we assume that conditions (0), (A)+(a), (sup B), (C ), ( ), and (^ ) hold. The proof below also applies to the case where condition (a0 ) is assumed instead of condition (a). The X are either semimartingales or PII. For a > 0, we de ne the limiters

ha (x) =

a 1 ^ 1 x; ha (x) = ha (r x); x 2 Rd ; jxj r

© 2001 by Chapman & Hall/CRC

(4.2.46)

326

Finite-dimensional LD convergence

and introduce processes X ;a = (X t;a ; t 2 R+ ) and X^ ;a = (X^ t;a ; t 2 R + ) by

X t;a = X^ t;a =

X

0<st Xt

(Xs

ha (Xs ));

(4.2.47)

X t;a :

(4.2.48)

Let Pa denote the distribution of X^ ;a and let = x0 , which is the idempotent distribution of X . Let (B ;a ; C ; ) be the triplet of X corresponding to ha . Since the jumps of X^ ;a are ha (Xs ), the triplet of X^ ;a corresponding to ha is (B ;a ; C ; ;a ), where

;a ([0; t]; ) = ([0; t]; (ha ) 1 ( )); t 2 R+ ;

2 B(Rd ):

(4.2.49) The semimartingales (respectively, PII) X^ ;a will LD converge in distribution to a certain semimaxingale (respectively, a semimaxingale with independent increments) X a . We de ne the latter as having characteristics (B a ; C; a ; ^a ) relative to ha , which are de ned in analogy with the characteristics of X^ ;a in that B a is the rst characteristic of X associated with ha (x),

a ([0; t]; ) = ([0; t]; ha 1 ( )); ^ta ( ) = ^t (ha 1 ( ));

2 B(Rd ):

(4.2.50) Corollary 2.8.7 implies that exists and is a Luzin-continuous idempotent process. We denote its idempotent distribution by a and the associated cumulant by Ga () = (Gat (); t 2 R+ ); 2 Rd . The proof of Theorem 4.2.1 (respectively, Theorem 4.2.5) consists in proving that

Xa

ld a (i) Pa ! as 2 ; a > 0; 1=r ;a j > " = 0; t > 0; " > 0; (ii) alim lim sup P sup j X s !1 2 st

(iii) a

iw ! as a ! 1.

Since part (ii) implies that

lim lim sup P1=r S (X ; X^ ;a ) > " = 0; " > 0; a!1 2 ld by Lemma 3.1.37 (i) through (iii) would yield P ! as required.

© 2001 by Chapman & Hall/CRC

327

Convergence of characteristics

Lemma 4.2.13. Part (i) holds.

Proof. We show that the X^ ;a and X a satisfy conditions (0), (sup B), (C ), ( ), and (^ ). Condition (0) holds by the hypotheses of Theorem 4.2.1. Since the rst characteristics of the X^ ;a and X a associated with ha coincide with the respective rst characteristics of X and X , condition (sup B ) for X^ ;a and X a is identical to condition (sup B ) for X and X . We now check that conditions (C ) are identical. By (4.2.49) and since (ha ) 1 (x) = fxg if r jxj < a, for Æ < a

;a [0; t];

\ fr jxj Æg = [0; t]; \ frjxj Æg ; 2 B(Rd );

and hence (4.1.7) (respectively, (4.2.5)) yields Ct;a;Æ = Ct;Æ for Æ < a (with obvious notation). The claim follows. The fact that ( ) for the X^ ;a and X a is implied by ( ) for X and X follows by the equalities 1 1 f (r x) t;a = f (ha (r x)) t ; r r

f (x) ta = f (ha (x)) t ;

and the inclusion f Æ ha 2 Cb if f 2 Cb . Condition (^ ) for the X^ ;a and X a is checked similarly. Since also X^ ;a satis es ( F ) by (4.2.49) and (4.2.46), Theorem 4.2.11 yields the assertion of the lemma.

Remark 4.2.14. Note that it is here, while checking (C ), that we used the property that the limiters in (4.2.46), by contrast with truncation functions, do not vanish at in nity.

Now we proceed with a proof of (ii).

Lemma 4.2.15.

If f (x); x 2 Rd ; is an R+ -valued bounded Borel function equal to 0 in a neighbourhood of the origin, then for all > 0 and > 0

P f (x) t > e + P (ef (x) 1) t;c +

© 2001 by Chapman & Hall/CRC

X

ln 1+(ef (x) 1) s >

0<st e +P

(ef (x)

1) t > :

328

Finite-dimensional LD convergence

Proof. Let Yt = f (x) t: Then Y = (Yt ; t 2 R+ ) has bounded variation over bounded intervals, and is, therefore, a semimartingale. The associated stochastic exponential E ;Y () = (Et;Y (); t 2 R + ); 2 R ; is of the form Y

t () Et;Y () = eG;Y (1+G;Y s ())e

st

G;Y s () ;

f (x) 1) : Lemma 4.1.1 implies that the where G;Y t () = (e t process exp(Yt )=Et;Y (); t 2 R+ is a supermartingale relative to F; hence, E exp(Y)=E;Y () 1 for every nite F-stopping time . Since

ln Et;Y () = (ef (x) 1)t;c +

X

0<st

ln 1+(ef (x) 1)s ;

an application of Lemma 3.2.6 yields the left inequality. The right inequality follows since ln(1 + x) x. The next lemma proves part (ii).

Lemma 4.2.16. Both under conditions (A) + (a) and (A) + (a0 ) for every " > 0

1=r lim lim sup P sup jXs;a j > " = 0; t 2 R+ : a!1 2 st

Proof. Since by (4.2.46) and (4.2.47) X

sup jXs;a j jXs j 1(r jXsj > a); st 0<st we have, for A > 0; " > 0, P sup jXs;a j > ") P (sup jXs j > A

st

+ P

X

0<st

st

jXs j 1(r jXs j > a) 1(jXs j A) > " :

© 2001 by Chapman & Hall/CRC

(4.2.51)

329

Convergence of characteristics

By the Lenglart-Rebolledo inequality, see, e.g., Liptser and Shiryaev [79, Theorem 1.9.3], for > 0

P sup jXs j > A st

e

X

P

1(jXs j > A) 1

0<st r + P ([0; t]; fjxj

> Ag) > e

r ;

and, hence, by (A)

lim sup lim sup P1=r sup jXs j > A st A!1 2

e ! 0 as ! 1:

(4.2.52) For the second term on the right-hand side of (4.2.51), we have by Lemma 4.2.15 for > 0 and > 0

P

X

0<st

jXs j 1(r jXs j > a) 1(jXs j A) > "

exp r( ") 1 (x) ;c + 1 + P ja;A; t r r

X

0<st

(x) > ; ln 1+ ja;A; s

which yields under condition (a0 ) 1=r lim lim sup P a!1 2

X

0<st

jXsj 1(r jXsj > a)

1(jXs j A) > "

= 0: (4.2.53)

Limits (4.2.52) and (4.2.53) in view of (4.2.51) prove the claim under conditions (A) + (a0 ). Since (a) is stronger than (a0 ), the required also holds under (A) + (a). It is left to prove (iii). We use the method of nite-dimensional distributions for idempotent processes, so we prove that nitedimensional idempotent distributions of the X a converge to nite dimensional idempotent distributions of X as a ! 1 and that the net fLi (X a ); a 2 R+ g is tight. We begin the proof of the convergence of nite-dimensional distributions by checking that Gat () ! Gt () as a ! 1.

© 2001 by Chapman & Hall/CRC

330

Finite-dimensional LD convergence

Lemma 4.2.17. For t 2 R+ and 2 Rd , as a ! 1, supjGas () Gs ()j ! 0: st Proof. Let as in (4.2.26) (x) = x ln(1 + x); x > 1. By (4.2.50) and (2.7.61)

1 Gat () = Bta + Ct + (ex 2 Zt

1 ha (x)) ta

1 1) ^sa ds = Bta + Ct 2

(ex

0

+ (eha (x)

1 ha (x)) t

Zt

(ex

0

1) ^sa ds:

Also by (2.7.61) 1 Gt () = Bta + Ct + (ex 2

1 ha (x)) t Zt

(ex

1) ^s ds:

(ex

1) ^s ds :

0

Thus,

jGat () Gt ()j ejjjxj 1(jxj > a) t Z t +

(eha (x)

0

1) ^s ds

Zt

0

(4.2.54)

The right inequality in (4.2.3) implies that the rst term on the righthand side of (4.2.54) tends to 0 as a ! 1. By the fact that (x) is positive, in order to prove that the second term tends to 0 uniformly over bounded intervals by Polya's theorem it is suÆcient to check convergence to 0 for every t 2 R+ . Since ha (x) ! h(x) as a ! 1, by Lebesgue's convergence theorem, (4.2.3) and (4.2.4) dominated h ( x ) x a e 1 ^s ! e 1 ^s : Therefore, the required convergence

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

331

would follow by Lebesgue's dominated convergence theorem, (4.2.3) and (4.2.4) provided lim sup ex 1(jxj > a) ^s = 0; a!1 st ha (x) 1 ^ > 0; lim inf 1 + inf e s a!1 st which hold by condition (L1 ). We now prove (iii).

iw Lemma 4.2.18. a !

as a ! 1: Proof. Since the a and are supported by C and the topology on C coincides with the one induced by the Skorohod topology, by Corollary 1.9.7 and Remark 1.9.8 we may apply the method of nite dimensional distributions of Theorem 2.2.27. Since both the X a and X are Luzin idempotent processes with independent increments, by Lemma 1.10.8 weak convergence of nite-dimensional idempotent distributions would follow from weak convergence of one-dimensional idempotent distributions. The latter follows by Lemma 1.11.19 and Lemma 4.2.17 if we recall that Gt () is dierentiable in . We check tightness of the net fa ; a 2 R+ g by verifying the conditions of Theorem 2.2.26. Condition 1Æ is obvious since a (x) = 0 if x0 6= x0 . Let us consider condition 2Æ . We denote, for Æ > 0; T > 0 and > 0; ATÆ; = fx 2 C : sup jxt xs j > g s;t2[0;T ]: js tjÆ Let ei ; 1 i 2d; denote the vector, whose b(i + 1)=2cth entry equals 1 if i is odd and -1 if i is even, the rest of the entries being equal to 0. Denoting by I a the rate function associated with a , we have for > 0; s < t, that if ei (xt xs ) > > 0; then I a (x) (Gat (ei ) Gas (ei )): Therefore, ATÆ; x 2 C : sup max ei (xt xs ) > d s;t2[0;T ]: i=1;:::;2d js tjÆ

2d n [

i=1

x 2 C : I a (x) d

© 2001 by Chapman & Hall/CRC

o

sup (Gat (ei ) Gas (ei )) ; s;t2[0;T ]: js tjÆ

332

Finite-dimensional LD convergence

and hence a (ATÆ; ) = sup exp I a (x) x2ATÆ; i=1max exp + sup Gat (ei ) Gas (ei ) ;:::;2d d s;t2[0;T ]: js tjÆ =d e max exp sup Gat (ei ) Gt (ei ) i=1;:::;2d t2[0;T ] + sup Gt (ei ) Gs (ei ) : s;t2[0;T ]: js tjÆ

By Lemma 4.2.17 and continuity of G() we conclude that lim sup lim sup a (ATÆ; ) e =d ! 0 as ! 1: Æ!0 a!1 Tightness of fa g is proved. Part (iii) is proved. Thus all the assertions (i), (ii) and (iii) are proved and by ld Lemma 3.1.37 P ! as 2 . Theorems 4.2.1 and 4.2.5 have been proved.

4.3 Corollaries In this section we discuss conditions and implications of Theorem 4.2.1. The PII case of Theorem 4.2.5 can be considered similarly. Thus, the X are semimartingales in what follows. We start with \integrable" versions when the convergence conditions can be checked for nontruncated characteristics. Let us recall that the nontruncated modi ed second characteristic C~ 0 of X is de ned by (see (2.7.59))

C~t0 = Ct +( x)2 t

Zt

( x ^s )2 ds:

0

If the X are special semimartingales, we introduce the conditions (sup B 0 )

sup jB 0 t tT

© 2001 by Chapman & Hall/CRC

1=r

P Bt0 j ! 0 as 2 ; T > 0;

333

Corollaries

and (I1 )

lim lim sup P1=r jxj 1(r jxj > a) t > " = 0; a!1 2 t > 0; " > 0:

If the X are also locally square integrable semimartingales, we introduce the conditions

kr C~t0

(C~ 0 ) and (I2 )

C~t0 k

1=r P

! 0 as 2 ; t 2 U;

lim lim sup P1=r r jxj2 1(r jxj > a) t > " = 0; a!1 2 t > 0; " > 0:

Note that (I2 ) implies (I1 ).

Lemma 4.3.1.

1. Let the X be special semimartingales. If conditions ( ) and (I1 ) hold, then conditions (sup B ) and (sup B 0 ) are equivalent.

2. Let the X be locally square integrable semimartingales. If conditions ( ), (^ ) and (I2 ) hold, then conditions (C ) and (C~ 0 ) are equivalent. Proof. The proofs are analogous to the proofs of Lemmas 4.2.8 and 4.2.10, respectively. For part 1 we write B 0 B 0 = B Bt t

t

+ (ha (x) h (x)) t (ha (x) h(x)) t + (x ha (x)) t (x ha (x)) t ;

where ha is from (4.2.46). Since condition ( ) implies condition (sup ), the expression in the rst parentheses on the right converges as 2 super-exponentially in probability to 0 locally uniformly in t. Since

jx ha(x)j t jxj 1(r jxj a) t; jx ha (x)j t jxj 1(jxj a) t ; (x ha (x)) t converges super-exponentially in probability to 0 as 2 and a ! 1 locally uniformly in t by (I1 ) and (x ha (x)) t

© 2001 by Chapman & Hall/CRC

334

Finite-dimensional LD convergence

converges to 0 as a ! 1 locally uniformly in t by (4.2.3). Part 1 is proved. In order to prove part 2, it suÆces to show as in the proof of Lemma 4.2.10 that for t 2 U; " > 0,

lim lim sup P1=r r ( x)2 1(r jxj > Æ) t Æ!0 2 lim lim sup P1=r Æ!0 2

Æ

( x)2 t > "

= 0; jxj 1(r jxj > Æ) s > "

X

0<st

X lim lim sup P1=r r Æ!0 2 0<st Zt

= 0; x 1(r jxj > Æ) s 2

x ^s 2 ds > " = 0:

0

The rst convergence follows by the inequality, where a > Æ, r ( 1

x)2 1(rjxj > Æ) t ( x)2 t r ( ha (rx))2 1(rjxj > Æ) t ( ha (x))2 t

+ jj2 r jxj2 1(r jxj > a) t + jj2 jxj2 1(jxj > a) t ;

part 1 of Lemma 4.2.9, condition (I2 ), and (4.2.3). The second convergence follows by the inequality

Æ

X

0<st

jxj 1(r jxj > Æ) s rÆ

X

jha (r x)j 1(r jxj > Æ) s

0<st

+Æ

X

jxj 1(r jxj > a) s;

0<st

part 2 of Lemma 4.2.9 and condition (I1 ). For the third convergence, we write X r 0<st

x 1

(r jxj > Æ)s 2

© 2001 by Chapman & Hall/CRC

Zt

0

x^s 2 ds Q1 +Q2 +Q3 ;

335

Corollaries

where

1 X ha (r x) 1(r jxj > Æ) s 2 r 0<st

Q1 =

Zt

X Q2 = r 0<st X 1

r 0<st

Zt Q3 = 0

ha (x) ^s 2 ds ;

0

x 1(r jxj > Æ) s

2

ha (r x) 1(r jxj > Æ) s 2 ; 2

ha (x) ^s ds

Zt

x ^s 2 ds :

0

Quantity Q1 converges super-exponentially in probability to 0 as 2 by part 3 of Lemma 4.2.9. For Q2 we have

Q2 2jj2 r

jj

jj

2 2r

+ 2 2r

X

0<st X

0<st X 0<st

jxj 1(r jxj > a) s jxj 1(r jxj > Æ) s

jxj 1(r jxj > a) s jxj 1(a rjxj > Æ) s

jxj 1(r jxj > a)s 4jj2 r

2 X

jxj2 1(r jxj > a) s:

0<st

The latter sum goes to 0 as 2 super-exponentially in probability by (I2 ). By a similar argument,

Q3

jj

42

Zt

0

( x)2 1(jxj > a) ^s ds;

and converges to 0 as a ! 1 by (4.2.3). Part 2 is proved. The next result is a direct consequence of Theorem 4.2.1 and Lemma 4.3.1.

© 2001 by Chapman & Hall/CRC

336

Finite-dimensional LD convergence

Theorem 4.3.2.

I. Let the X be special semimartingales and condition (I1 ) hold. If conditions (0), (A) + (a), (sup B 0 ), (C ) (or (C~ ) associated with a continuous limiter), ( ), and (^) ld hold, then X ! X as 2 .

II. Let the X be locally square integrable semimartingales and condition (I2 ) hold. If conditions (0), (A)+(a), (sup B 0), (C~ 0 ), ld ( ), and (^ ) hold, then X ! X as 2 .

Remark 4.3.3. Similarly, in the statements below we can replace 0

condition (sup B ) by condition (sup B ) each time condition (I1 ) holds and replace condition (C~ ) by condition (C~ 0 ) each time condition (I2 ) holds.

We next consider the \quasi-continuous" case ^s (Rd ) = 0. It is singled out by the condition (QC )

1=r

P 1 X (fsg; fr jxj > g)2 ! 0; t > 0; > 0: r 0<st

Since (QC ) implies (^ ) with ^s (Rd ) = 0, condition (L1 ) trivially holds. We thus obtain the following corollary of Theorem 4.2.1.

Corollary 4.3.4. Let condition (QC ) hold and the limiter h(x) be continuous. If conditions (0), (A) + (a), (sup B ), (C ) (or (C~ )), and ld ( ) hold, then X ! X as 2 .

As a consequence, we derive a result on LD convergence to the Poisson idempotent process. Since the latter by Theorem 2.4.16 has characteristics Bt0 = t, Ct = 0, t ( ) = 1(1 2 ), and ^t ( ) = 0, we have the following result.

Corollary 4.3.5. Let the semimartingales X be one-dimensional

and the limiter h(x) be continuous at x = 1. Let N be a Poisson idempotent process. Let conditions (0) for x0 = 0, (A) + (a) and (QC ) hold. If

sup jBt tT

h(1)tj

lim lim sup P1=r Æ!0 2

© 2001 by Chapman & Hall/CRC

1=r P

! 0

as 2 ; T > 0;

kr Ct;Æ k > " = 0; t 2 U; " > 0;

337

Corollaries

and for all " 2 (0; 1=2), as 2 , 1=r

P 1 ([0; t]; fjr x 1j "g) ! t; t 2 U; r

1=r

P 1 ([0; t]; fr jxj > "g \ fjr x 1j > "g) ! 0; t > 0; r ld then X ! N as 2 . Proof. Let h(x) be continuous. The rst characteristic of X associated with h(x) equals h(1)t. Therefore, the rst two convergences in the statement check conditions (sup B ) and (C ). We check condition ( ). Let f 2 Cb and > 0 be arbitrary. Let > 0 be such that jf (x) f (1)j if jx 1j and f (x) = 0 if jxj . Then by the fact that f t = f (1)t jf t f j krf k [0; t]; fr jxj > "g \ fjrx 1j > "g

1 + [0; t]; fjr x 1j "g r 1 + jf (1)j [0; t]; fjr x 1j "g t ; r which implies condition ( ) by hypotheses and arbitrariness of . The stated LD convergence follows now by Corollary 4.3.4. Now let h(x) be continuous at x = 1 and h(x) be a continuous limiter such that h(1) = h(1). We denote by B the rst characteristic of X corresponding to h(x). Given > 0, we choose > 0 such that h(x) = h(x) = 0 if jxj and jh(x) h(x)j if jx 1j : Then by (4.1.4) denoting khk = supx2Rd jh(x)j and khk = supx2Rd jh(x)j

jB t h(1)tj jBt h(1)tj + h (x) h (x) t jBt h(1)tj + r1 [0; t]; fjr x 1j "g k hk + khk + [0; t]; fr jxj > "g \ fjr x 1j > "g : r

Since is arbitrary, the hypotheses imply that The claim follows by the part already proved.

© 2001 by Chapman & Hall/CRC

B t

h(1)t

1=r P

! 0.

338

Finite-dimensional LD convergence

Let us now consider the case t ( ) = 0. It is implied by the condition 1=r P 1 (MD) ([0; t]; fr jxj > g) ! 0 as 2 ; t > 0; > 0: r In the large deviation theory terminology this is the case of \moderate deviations". We recall that if = 0, then the idempotent deviability distribution of X has density 1 Z1 x0 (x) = exp 2 (x_ t bt ) ct (x_ t bt ) dt 0

if x is absolutely continuous, x0 = x0 and x_ t bt is in the range of ct (a.e.), and x0 (x) = 0 otherwise. By Theorem 2.6.26 X is the Luzin-continuous idempotent Gaussian diusion Zt

Zt

0

0

Xt = x0 + bs ds + c1s=2 W_ s ds;

(4.3.1)

where W is a Wiener idempotent process. Hence, this is a \central limit theorem" setting. Lemma 4.3.6. Let B and C~ , B and Ce be the rst and modi ed second characteristics of X corresponding to respective limiters h(x) and h(x), not necessarily continuous. If condition (MD) holds, then, as 2 , sup jBt tT

r kC~t

1=r P Bt j !

Ce

tk

0; T > 0;

1=r P

! 0; t > 0:

Proof. For B and B the claim is a direct consequence of (4.1.4), the de nition of a limiter and (MD). For C~ and Ce , we have by the de nition of modi ed second characteristics, choosing > 0 such that h(x) = h(x) = x; jxj ; that 1 r kC~t () Ce t ()k jj2 (khk2 + khk2 ) ([0; t]; fr jxj > g) r 1 X + jj2 (khk + khk)2 (fsg; fr jxj > g): r 0<st

© 2001 by Chapman & Hall/CRC

339

Corollaries

Thus, the claim follows by (MD). We introduce the conditions (sup B 0 )

sup jBt tT

0

(C0 )

kr C~t

1=r

P Bt0 j ! 0 as 2 ; T > 0;

Ct k

1=r P

! 0 as 2 ; t 2 U:

Theorem 4.2.1 and Lemma 4.3.6 yield the following.

Corollary 4.3.7. (\the LD central limit theorem") Let X be given 0 by (4.3.1). Let conditions (0), (A) + (a), (sup B0 ), (C0 ), and (MD) ld hold with some limiter h(x). Then X ! X as 2 .

If we would like to use nontruncated characteristics of the X , we could require the following Lindeberg condition (L2 )

r j

1=r P (r jxj > ) t !

j1

x2

0; t > 0; > 0;

which implies both conditions (I2 ) and (MD). Let us introduce the condition (C 0 ) 0

krC~t0

Ct k

1=r P

! 0 as 2 ; t 2 U:

We thus have the following version.

Corollary 4.3.8. Let X be given by (4.3.1).

Let the X be locally square integrable semimartingales and condition (L2 ) hold. If conld ditions (0), (A) + (a), (sup B 0 ), and (C00 ) hold, then X ! X as 2 .

Now we consider simpler versions of conditions (A) + (a) on the jumps of the X . We rst note that condition (A) can be checked by checking the condition (A0 )

([0; t]; jxj > A)1=r

1=r P

! 0 as 2 ; t > 0; 9A > 0;

and condition (a) can be checked by checking the condition

© 2001 by Chapman & Hall/CRC

340

Finite-dimensional LD convergence

1=r

P 1 r jxj e 1(r jxj > a) 1(jxj A) t ! 0 r as 2 ; t > 0; > 0; A > 0; 9a > 0:

(a0 )

Let us also note that if condition (A0 ) holds and the convergence in (a0 ) holds for every a > 0, then condition (MD) holds. The following observation comes in useful below.

Lemma 4.3.9. Condition (a) is implied by the conditions (a1 )

(a2 )

1=r 1 [0 ; t ] ; f r j x j > a g > = 0; lim lim sup P a!1 2 r t > 0; > 0;

lim lim sup P1=r a!1 2

1 r

r ZA

eu [0; t]; fr jxj > ug du > = 0;

a

t > 0; > 0; > 0; A > 0;

Proof. The claim follows since

1 r jxj e 1 1(r jxj > a) 1(jxj A) t r Z1 1 eu 1(r jxj > u)du 1(r jxj > a) 1(jxj A) t = r

r1

0 r Z A

eu [0; t]; fr jxj > ug \ fr jxj > ag du

0

r1 [0; t]; fr jxj > ag

+

1 r

ZR

eu du

0 r Z A

eu [0; t]; fr jxj > ug du;

R

where R > 0 is arbitrary. The following conditions can also be used for checking conditions (A) + (a).

© 2001 by Chapman & Hall/CRC

341

Corollaries

(V S )

lim lim sup P1=r ([0; t]; fr jxj > ag)1=r > " = 0; a!1 2 t > 0; " > 0; 1=r P 1 =r )

! 0 as 2 ; t > 0; > 0: Clearly, (V S0 ) ) (V S ) ) (A) + (a). Since also (V S0 ) ) (MD), by

(V S0 )

([0; t]; fr

jxj > g

Corollary 4.3.7 we have the following.

Corollary 4.3.10. Let X be given by (4.3.1).

Let conditions (0), ld 0 (V S0 ), (sup B0 ), and (C0 ) hold with some limiter h(x). Then X ! X as 2 .

Now we consider the case of the \classical" large deviation setting when the Cramer condition holds: ejxj 1(jxj > 1) < 1; 2 ; t > 0; > 0: t

We introduce the conditions 1=r 1 r jxj > " = 0; (Ie ) lim lim sup P e 1 ( r j x j > a ) t a!1 2 r t > 0; " > 0; > 0; (Le )

1=r

P 1 r jxj e 1(rjxj > ) t ! 0 as 2 ; r t > 0; > 0; > 0:

Condition (Le ) can be called an exponential Lindeberg condition. Obviously, (Le ) ) (Ie ) ) (A)+(a), (Ie ) ) (I2 ), and (Le ) ) (L2 ) ) (MD). By Theorem 4.3.2 the implication (Ie ) ) (I2 ) allows us to consider nontruncated characteristics under (Ie ). We thus have the following result.

Corollary 4.3.11.

Let the Cramer condition and condition (Ie ) hold. If conditions (0), (sup B 0 ), (C~ 0 ), ( ), and (^) hold, then ld X ! X as 2 .

The following is an application to point processes.

Corollary 4.3.12.

Let Xt = Nt =r , where N = (Nt ; t 2 R+ ) are one-dimensional point processes with compensators A = (At ; t 2 R+ ).

© 2001 by Chapman & Hall/CRC

342

Finite-dimensional LD convergence

a) If, as [0; 1],

2 ,

for some Lebesgue measurable function s

1 A r t

1=r P

1 X (As )k r 0<st

1=r

2

Zt

! t + s ds ; t 2 U;

P

!

Zt

0

0

ks ds ; t 2 U; k = 2; 3; : : : ;

ld then X ! X as 2 , where X is the idempotent process with independent increments with local characteristics bt = 1+t , ct = 0, t ( ) = (1 + t ) 1(1 2 ), and ^t ( ) = t 1(1 2 ). b) In particular, if

1 A r t 1 X (As )2 r 0<st ld then X !

1=r P

! t; t 2 U;

1=r P

! 0; t > 0;

N as 2 , where N is an idempotent Poisson process.

Proof. In part a) the nontruncated characteristics of X are of the form B 0 t = At =r , Ct = 0, ([0; t]; ) = 1(r 1 2 )At so that R conditions (sup B 0 ), (C~ 0 ), ( ), and (^ ) hold with Bt = t + 0t s ds, Ct = 0, and t and ^t as indicated in the statement. Part b) is a consequence of part a).

The implications (Le ) ) (L2 ) and (Le ) following version of Corollary 4.3.8.

) (A) + (a) give the

Corollary 4.3.13. Let X be given by (4.3.1). Let the Cram er con0 0 dition and condition (Le ) hold. If conditions (0), (sup B ), and (C0 ) ld hold, then X ! X as 2 .

4.4 Applications to partial-sum processes In this section we consider applications of the above results to the setting of the processes of partial sums of random variables. Let

© 2001 by Chapman & Hall/CRC

343

Applications

fin; i 2 N g; n 2 N ; be sequences of

R d -valued

random variables de ned on respective probability spaces ( n ; Fn ; Pn ) and adapted to discrete-time ltrations Fn = fFin ; i 2 N g. Let rn ! 1 as n ! 1 and bX ntc n Xt = in ; t 2 R+ ; (4.4.1) i=1 P0 where i=1 = 0. The predictable triplet (B n ; C n ; n ) (Xtn ; t R+ ) corresponding to a limiter h(x) is given by

2

bntc 1 X = En h(rn in )jFin 1 ; rn i=1 bntc X n [0; t]; = Pn (in 2 n f0gjFin 1 );

Btn

i=1

of X n =

Ctn = 0;

2 B(Rd );

where En denotes expectation with respect to Pn . We consider large deviation convergence of the X n with rate rn . Let X be a semimaxingale with characteristics (B; C; ; ^) and modi ed second characteristic C~ associated with h(x) as de ned in Section 4.2. The conditions of Theorem 4.2.1 assume the form. ntc bX 1=rn 1 =r n (A) lim lim sup Pn Pn (jin j > AjFin 1 ) > = 0; A!1 n!1 i=1 t > 0; " > 0; b nt c X 1=rn 1 rn jin j E e (a) lim lim sup P n n a!1 n!1 rn i=1 1(rnjin j > a) 1(jin j A jFin 1 > " = 0; t > 0; > 0; A > 0; " > 0; ntc 1 bX n 1 =r 0 n (a ) alim sup Pn ln 1 + En (ern ji j 1) 1(rn jin j > a) !1 lim r n!1 n i=1 1(jin j AjFin 1 > " = 0; t > 0; > 0; A > 0; " > 0; ntc bX 1=rn 1 (sup B ) sup En h(rn in )jFin 1 Bt Pn! 0 as n ! 1; tT rn i=1 T > 0;

© 2001 by Chapman & Hall/CRC

344 (C )

Finite-dimensional LD convergence

ntc bX 1 =r n lim lim sup Pn En ( in )2 rn Æ!0 n!1 i=1 n n En i (rn i Æ) in 1 2

1 j j jF

1(rnjinj Æ)jFin 1 Ct >

= 0; t 2 U; " > 0; 2 Rd ;

ntc bX 1=rn 1 n ))2 jF n lim P E ( h ( r n n n i i 1 n!1 rn i=1 En h(rn in )jFin 1 2 C~t > = 0; t 2 U; " > 0; 2 Rd ; bntc 1=rn 1 X ( ) En f (rnin )jFin 1 Pn! f (x) t as n ! 1; rn i=1 t 2 U; f 2 Cb ; Zt bntc k Pn1=rn 1 X n n (^ ) En (f (rn i )jFi 1 ) ! f (x) ^s k ds as n ! 1; rn i=1 0 k = 2; 3; : : : ; t 2 U; f 2 Cb : The following theorem is a triangular array version of Theorem 4.2.1. Theorem 4.4.1. Let X n be de ned by (4.4.1) and h(x) be a continuous limiter. If conditions (A) + (a) , (sup B ) , (C ) (or (C~ )), ld ( ) , and (^ ) hold, then X n ! X. The integrable and square integrable versions look as follows. As above, B 0 and C~ 0 denote the nontruncated rst and modi ed second characteristics of X , respectively. Theorem 4.4.2. I. Let Enjinj < 1; i 2 N . Let

(C~ )

(I1 )

bntc X 1 =r n lim lim sup P En a!1 n!1 n i=1

jinj 1(rn jinj > a)jFin 1

> " = 0; t > 0; " > 0

and

(sup B 0 )

bntc X sup En (in tT i=1

© 2001 by Chapman & Hall/CRC

1=rn

jFin 1 ) Bt0 Pn! 0

as n ! 1;

T > 0:

345

Applications

If, in addition, conditions (A) + (a) , (C ) (or (C~ ) with ld a continuous limiter), ( ) , and (^ ) hold, then X n ! X as n ! 1.

II. Let En jin j2 < 1; i 2 N . Let

(I2 )

lim lim sup P 1=rn a!1 n!1 n

rn

bX ntc i=1

En jin j2 1(rn jin j > a)jFin 1

> " = 0; t > 0; " > 0

and

(C~ 0 )

rn

bX ntc i=1

En (( in )2 jFin 1 )

En ( in jFin 1 )

2

Pn1=rn

! C~t0 as n ! 1; t 2 U; 2 Rd :

If, in addition, conditions (A) + (a) , (sup B 0) , ( ) , and ld (^ ) hold, then X n ! X as n ! 1. Conditions (QC ) and (MD) look in the triangular array setting as follows. bntc 1=rn 1 X (QC ) Pn (rn jin j > jFin 1 ) 2 Pn! 0 as n ! 1; rn i=1 t > 0; > 0; bntc 1=rn 1 X (MD) Pn (rn jin j > jFin 1 ) Pn! 0 as n ! 1; rn i=1 t > 0; > 0: The other conditions take the form bX ntc 1=rn (L2 ) rn En jin j2 1(rn jin j > )jFin 1 Pn! 0 as n ! 1; i=1

(A0 )

ntc bX i=1

t > 0; > 0; 1=rn

Pn (jin j > AjFin 1 )

© 2001 by Chapman & Hall/CRC

Pn1=rn

! 0 as n ! 1; t > 0; " > 0; 9A > 0;

346

Finite-dimensional LD convergence

(a0 )

bntc 1=rn 1 X n En ern ji j 1(rn jin j > a) 1(jin j A jFin 1 Pn! 0 rn i=1 as n ! 1; t > 0; > 0; A > 0; 9a > 0:

(a1 )

lim lim sup P 1=rn a!1 n!1 n

(a2 )

lim lim sup P 1=rn a!1 n!1 n

lim lim sup P 1=rn a!1 n!1 n

(V S )

ntc bX

(V S0 )

(Ie )

i=1

bntc 1 X Pn (rn jin j > ajFin 1 ) > = 0; rn i=1 t > 0; > 0;

bntc 1 X rn i=1

ntc bX i=1

r ZnA a

eu Pn (rn jin j > ujFin 1 )du >

= 0;

t > 0; > 0; > 0; A > 0; 1=rn

Pn (rn jin j > ajFin 1 )

> = 0;

t > 0; > 0; 1=rn

Pn (rn jin j > jFin 1 )

Pn1=rn

! 0 as n ! 1; t > 0; > 0;

lim lim sup P 1=rn a!1 n!1 n

bntc n 1 X En ern ji j 1(rn jin j > a)jFin 1 > rn i=1 = 0; t > 0;

bntc 1=rn 1 X n En ern ji j 1(rn jin j > )jFin 1 Pn! 0 as n ! 1; rn i=1 t > 0; > 0;

(Le )

(sup B00 )

(C0 )

bntc X 1=rn 1 sup En h(rn in )jFin 1 ) Bt0 Pn! 0 tT rn i=1 as n ! 1; T > 0;

bntc 1 X E (( h(rn in ))2 jFin 1 ) rn i=1 n Pn1=rn

2

En ( h(rn in )jFin 1 )

! Ct as n ! 1; t > 0; 2 Rd ;

© 2001 by Chapman & Hall/CRC

347

Applications

bntc X 0 (C0 ) rn En (( in )2 jFin 1 ) i=1

2 Pn1=rn

En ( in jFin 1 )

! Ct as n ! 1; t > 0; 2 Rd :

We now consider versions of the results of the preceding section on LD convergence in distribution to idempotent diusions. Let X be the idempotent diusion given by (4.3.1), where x0 = 0. Corollary 4.4.3. Let Enjinj2 < 1; i 20 N ; and condition (L2 ) hold. If conditions (A) + (a) , (sup B ) , and (C00 ) hold, then ld Xn ! X as n ! 1. Corollary 4.4.4. Let conditions (V S0) , (sup B00 ) and (C0 ) hold ld with some limiter h(x). Then X n ! X as n ! 1. Corollary 4.4.5. Let E exp(jin j) < 1; 2 R+ ; and condition ld (Le ) hold. If conditions (sup B 0 ) and (C00 ) hold, then X n ! X as n ! 1. We next consider an application to a typical moderate deviation setting. Theorem 4.4.6. Let i; i 2 N ; be i.i.d. Rd -valued random variables on a probability space ( ; F ; P ) such that E j1 j2 < 1 and E1 = 0, and let bntc 1 X n Xt = ; bn i=1 i where bn =n ! 0 and b2n =n ! 1 as n ! 1. If, for some v > 0,

2

lim nP (j1 j > vbn ) n=bn = 0; (4.4.2) n!1 then the X n LD converge in distribution at rate b2n =n to the Luzincontinuous idempotent diusion X = (E1 1T )1=2 W , where W is an R d -valued Wiener idempotent process. The deviability distribution of X is idempotent Gaussian and given by 1 Z1 X _ t (E1 1T ) x_ t dt ; (x) = exp x 2 0

if x is absolutely continuous, x0 = 0 and x_ t belongs to the range of E1 1T (a.e.), and X (x) = 0 otherwise.

© 2001 by Chapman & Hall/CRC

348

Finite-dimensional LD convergence

Proof. We take rn = b2n =n. It is easy to check that condition (L2 ) holds so we can apply Corollary 4.4.3. Since E1 = 0, condition (sup B 0 ) holds with Bt0 = 0. Condition (C00 ) holds with Ct = (E1 1T )t. We thus need to check conditions (A) + (a) . In view of Lemma 4.3.9 it is suÆcient to check conditions (A0 ) , (a1 ) and (a2 ) . We assume with no loss of generality that bn , n=bn and b2n =n are monotonically increasing. Condition (A0 ) has the form n=b2n

lim sup bntcP (j1 j > Abn ) n!1

=0

for some A 2 R+ and follows from (4.4.2) with A = v. Condition (a1 ) assumes the form

nbntc n lim lim sup P j j > a = 0: 1 a!1 n!1 b2n bn We actually check that

nbntc n lim P j j > v = 0: 1 n!1 b2n bn

(4.4.3)

Let integer N = N (n) be such that n bN < bN +1 : bn Then, noting that bN +1 bn =n , we have

bN (N + 1)=N 2bN by monotonicity of

nbntc n P j j > v tb2N +1 P (j1 j > vbN ) 4tb2N P (j1j > vbN ): 1 b2n bn Since N ! 1 as n ! 1, condition (4.4.2) implies that

lim 4tb2N P (j1 j > vbN ) n!1 so that

N=b2 N

lim 4tb2N P (j1 j > vbN ) = 0:

n!1

Limit (4.4.3) follows.

© 2001 by Chapman & Hall/CRC

=0

349

Applications

Condition (a2 ) assumes the form 2 Ab Z n =n

nbntc lim lim sup 2 a!1 n!1 bn

eu P j1 j >

a

u n du = 0: bn

We actually prove that for all A and a large enough

nbntc lim n!1 b2n

2 Ab Z n =n a

eu P j1 j >

u n du = 0: bn

(4.4.4)

Let A v and de ne integer L = L(n; u) by

bL

u n A, we have that u A and, hence, bL+1 > n=bn . Therefore, by the fact that as above bL+1 2bL , for u 2 [a; Ab2n =n]

nbntc u u n e P j j > teu b2L+1P (j1 j > AbL) 1 b2n bn 4teu b2LP (j1 j > AbL ): (4.4.6) Since bL+1 n=bn for u a, we have that L ! 1 as n uniformly over u a, so by (4.4.2) and the fact that A v

lim sup LP (j1 j > AbL ) n!1 ua

L=b2 L

!1

= 0:

Therefore, given arbitrary > 0, for all n large enough and u a

LP (j1 j > AbL ) e

b2L =L :

(4.4.7)

Since in the integral in (4.4.4) u Ab2n =n, it follows by (4.4.5) that bL bn , so by monotonicity L n and bL =L bn =n, which implies by (4.4.5) that

b2L L

u bL bnn bL2+1 bnn 2A :

© 2001 by Chapman & Hall/CRC

(4.4.8)

350

Finite-dimensional LD convergence

Using (4.4.7) and (4.4.8) we obtain by (4.4.6) that for n large enough and u 2 [a; Ab2n =n]

b2 nbntc u u n e P j j > 4teu L e 1 2 bn bn L

b2L =L

eu e e

b2L =(2L)

=(4A) 1 u ;

hence,

nbntc lim sup 2 bn n!1

2 Ab Z n =n

eu P

a

j1j > u bn du n

Since the latter integral converges to 0 as proved.

Z1

e

=(4A) 1 u du:

a

! 1, limit (4.4.4) is

Remark 4.4.7. As the proof shows, under the hypotheses condition

(a0 ) holds. The next result is in the same theme but considers triangular arrays of row-wise i.i.d.r.v. Theorem 4.4.8.d Let in; i 2 N ; n 2 N ; be a triangular array of row-wise i.i.d. R -valued random variables on respective probability spaces ( n ; Fn ; Pn ) such that En 1n = 0 and En j1n j2 < 1, and let bntc 1 X n n Xt = ; bn i=1 i where bn =n ! 0 and b2n =n ! 1 as n ! 1. If En 1n 1nT ! as n ! 1, where is a positive semi-de nite symmetric matrix and either

sup En j1nj2+Æ < 1 for some Æ > 0 and n

b2n n ln n

or

© 2001 by Chapman & Hall/CRC

as n ! 1; (4.4.9)

b2n ! 0 as n ! 1 n for some > 0 and 2 (0; 1], (4.4.10)

sup En exp( j1n j ) < 1 and n

!0

351

Applications

then the X n LD converge in distribution at rate b2n =n to the Luzincontinuous idempotent Gaussian diusion X = (T )1=2 W , where W is an Rd -valued Wiener idempotent process. Proof. We take rn = b2n =n and apply Corollary 4.4.3. Since the moment conditions imply (L2 ) , and conditions (sup B 0) and (C00 ) hold, we have to check conditions (A) + (a). Under (4.4.9) this is done by checking condition (V S0 ) , which takes the form n n=b2n 2 nn=bn Pn j1nj > ! 0 as n ! 1 bn

and follows by (4.4.9). Under condition (4.4.10) we check that conditions (A0 ) and (a0 ) hold. Condition (A0 ) holds since by (4.4.10)

nPn

j j > An=b2n b

n 1

n

2

nn=b2n En exp( j1nj ) n=bn exp A n=b2n ; which converges to 0 as n ! 1.

Veri cation of condition (a0 ) is a bit more intricate. Let us assume that < 1. We have for > 0, > 0, A > 0, and > 0 j n j n2 bn n bn n 1 A E exp j j 1 j j > 1 b2n n n 1 n 1 bn 2 nb2 En exp bnn j1nj 1 bnn j1nj1 < 1 bnn j1nj > n b j n j n2 b + 2 En exp n j1n j 1 n j1n j1 1 1 A : bn n n bn (4.4.11)

The rst term on the right of (4.4.11) is not greater than n2 n j 1 bn j n j > E exp j n 1 1 b2n n 2 nb2 En exp j1n j + 12 j1nj exp 21 bn n n (4.4.12)

© 2001 by Chapman & Hall/CRC

352

Finite-dimensional LD convergence

and converges to 0 as n ! 1 if < =2 by the moment condition in (4.4.10) and the assumption n=bn ! 1. We estimate the second term on the right of (4.4.11) as j n j n2 bn n bn n 1 1 A j j 1 j j 1 E exp n 1 1 b2n n n bn n =(1 ) n2 b2n b2 exp A n =(1 ) b En exp( j1n j ) ; n n which goes to 0 by (4.4.10). Thus, the right-hand side of (4.4.11) goes to 0 as n ! 1 so that condition (a0 ) is checked for < 1. If = 1, the required follows by (4.4.12).

Remark 4.4.9. If the distributions of the 1n do not depend on n,

the moment conditions above imply condition (4.4.2). Remark 4.4.10. We have actually checked that under (4.4.10) condition (a0 ) holds for every a > 0. We now consider examples on LD convergence to dierent kinds of idempotent processes. Example 4.4.11. "Very large deviations" Let X n be given by (4.4.1). Let rn = n, which speci es the set-up of \very large deviations". We assume that in = g(i=n; i )=n, where i ; i 2 N ; are Rd -valued i.i.d.r.v. on a probability space ( ; F ; P ), and g : R+ Rd ! Rd is continuous in the rst variable and such that E exp(jg(t; 1 )j) < 1 for all > 0; t > 0. It is easy to see that all the conditions of Corollary 4.3.11 hold with

B0 = t

Zt

Eg(s; 1 ) ds; Ct = 0; t ( ) = ^t ( ) = P g(t; 1 ) 2

0

It is instructive to note that the Cramer condition is not indispensable in this sort of result. Indeed, let in = in=n, where in; i 2 N ; are R + -valued random variables, i.i.d. for each n with the distribution function 1(x > n2) : P (1n x) = 1 exp( x2 ) 1(x n2 ) n2 exp( n4 ) x

© 2001 by Chapman & Hall/CRC

nf0g :

353

Applications

Then conditions (A) + (a) are easily seen to hold while neither condition (V S ) nor (Ie ) is satis ed, and even E1n = 1. The other conditions of Theorem 4.4.1 are satis ed as well with Z1 Bt = t h(x)d(1 exp( x2 )); Ct = 0; 0

t ( ) = ^t ( ) =

Z

1(x > 0)d(1

exp( x2 ));

2 B(Rd ):

Example 4.4.12. LD convergence to Poisson idempotent processes.

Let X n be given by (4.4.1). Let in = in=rn , where rn ! 1, rn =n ! 0 as n ! 1, and fin ; i 2 N g are independent r.v. assuming values 1 and 0 with respective probabilities rn =n and (1 rn =n). ld N as n ! 1 Then part b) of Corollary 4.3.12 implies that X n ! at rate rn , where N is a Poisson idempotent process. Note that here n ([0; t]; frn jxj > g)=rn ! t for < 1, so condition (MD) does not hold, while condition (QC ) does. Example 4.4.13. LD convergence of empirical processes. Let n 1 X rn n Xt = 1 i n t ; rn i=1 where i are i.i.d.r.v. with values in R+ , whose distribution admits density g(x), which is continuous and positive at 0. Also rn ! 1 and rn =n ! 0 as n ! 1. We denote by G(x) the distribution function of 1 and introduce the point process Ntn = rn Xtn . Then the compensator of N n = (Ntn ; t 2 R+ ) relative to the natural ltration is, Jacod and Shiryaev [67, II.3.32], r n Zt n g s X r r n n n ds: Ant = rn 1 i s r n i=1 n 1 G ns 0 n It is not diÆcult to check that n 1X r 1=rn 1 i n s Pn! 0 n i=1 n

© 2001 by Chapman & Hall/CRC

354

Finite-dimensional LD convergence

1=rn

as n ! 1 and hence Ant =rn Pn! tg(0): Part b) of Corollary 4.3.12 ld implies that the X n ! X as n ! 1 at rate rn , where Xt = Ng(0)t , N being an idempotent Poisson process.

© 2001 by Chapman & Hall/CRC

Chapter 5

The method of the maxingale problem The method of nite-dimensional distributions considered in Chapter 4 does not allow us to prove LD convergence in distribution to idempotent processes other than idempotent processes with independent increments. In this chapter we consider a dierent approach, which is an analogue of the martingale problem approach in weak convergence theory and consists in identifying the limit deviability as a solution to a maxingale problem. As in Chapter 4, we consider a net of semimartingales fX ; 2 g de ned on respective stochastic bases ( ; F ; F ; P ) with paths in D = D (R + ; Rd ). We assume as xed a net fr ; 2 g of real numbers greater than 1 converging to 1 as 2 . It is used as a rate for LD convergences below, which refer to the Skorohod topology. The limit semimaxingale X is assumed to be \canonical" in that it is de ned on D by Xt (x) = xt ; x 2 D ; t 2 R+ . It will actually be Luzin-continuous so that we can equivalently consider it as the canonical idempotent process on C = C (R + ; Rd ). The next two sections are concerned with identifying maxingale problems whose solutions are LD accumulution points of fL(X ); 2 g: Section 5.1 speci es the maxingale problem in terms of convergence of stochastic exponentials and assumes the Cramer condition for the X , while Section 5.2 considers convergence of the characteristics of the semimaxingales and does without the Cramer condition. Section 5.3 is devoted to speci c LD convergence results. Section 5.4 considers applications to large deviation convergence of Markov processes. 355 © 2001 by Chapman & Hall/CRC

356

Maxingale problem

5.1 Convergence of stochastic exponentials This section contains results on LD convergence of semimartingales stated in terms of convergence of the associated stochastic exponentials. Let G() = Gt (; x); t 2 R+ ; x 2 D ; 2 Rd ; be an R-valued function such that G0 (; x) = Gt (0; x) = 0, which is continuous in t and D-adapted in x. As above, we refer to G() as a cumulant. We introduce a number of conditions.

De nition 5.1.1. The function G() is said to satisfy the uniform continuity condition if the map x ! (Gt (; x); t 2 R+ ) is a C { continuous map from

D

into C (R + ; R).

De nition 5.1.2. Let F = (Ft (x); t 2 R+ ; x 2 D ); F0 (x) = 0; be an continuous function. We say that F satis es the majoration condition if there exists an R-valued, increasing and continuous function F = (F t ; t 2 R+ ); F 0 = 0; such that for all 0 s < t, R -valued

sup(Ft (x) Fs (x)) F t F s :

x2D

(5.1.1)

The function F is said to satisfy the local majoration condition if, for each b > 0, there exists an R-valued, increasing and continuous in t function F b = (F bt ; t 2 R+ ); F b0 = 0; such that, for all 0 s < t,

sup (Ft (x) Fs (x)) F bt F bs : xx12D:b

(5.1.2)

Remark 5.1.3. If F is D-adapted, then, being continuous, it is Dpredictable, so the preceding supremum may be taken over x 2 D such that xt b. More generally, if is a nite D{stopping time on D , then, for every nite D{stopping time , sup (F (x) F (x)) = sup (F (x) F (x)) x2D : x2D : x1 b

x b

(See, e.g., Jacod and Shiryaev [67, III.2.43].)

At times we require that the restriction of G() to C satisfy the linear-growth condition of De nition 2.8.11, which we recall here.

© 2001 by Chapman & Hall/CRC

357

Convergence of stochastic exponentials

De nition 5.1.4. We say that G() satis es the linear-growth con-

dition if there exist R+ {valued, increasing and continuous in t functions F l () = (Ftl (); t 2 R+ ); 2 Rd ; such that F0l () = Ftl (0) = 0 and for some R+ -valued increasing function kt we have for all 0 s < t , x 2 C and 2 Rd

Gt (; x) Gs (; x) Ftl ((1+kt xt )) Fsl ((1+kt xt )):

We also recall that a deviability on C is a solution to the maxingale problem (x0 ; G), where x0 2 Rd , if the canonical process X on C is a semimaxingale with cumulant G() on (C ; C; ) such that X0 = x0 -a.e. (De nition 2.8.1). Let the X satisfy the Cramer condition (Cr) and E () = Et (); t 2 R+ ; 2 Rd ; be the associated stochastic exponentials. The following conditions on the X are similar to those used in Section 4.1: 1=r P X0 ! x0

(0) (sup E )

as 2 ;

1=r

P 1 sup j ln Et (r ) Gt (; X )j ! 0 as 2 ; tT r 2 Rd ; T > 0:

Theorem 5.1.5. d

Let the X satisfy (Cr), and G(), for each 2 R , satisfy the uniform continuity and majoration conditions. If conditions (0) and (sup E ) hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.1.6. By the fact that a cumulant G() that does not depend on x satis es the uniform continuity and majoration conditions, Theorems 5.1.5 and 2.8.5 imply Theorem 4.1.2.

The majoration condition on G() is too restrictive in applications. We replace it next by the local majoration condition and another condition. Recall that x0 is de ned by (2.7.6) and x;t by (2.8.6). The following is the condition we will require. (NE ) The function x (x); x 2 C ; is upper compact and the sets [s2[0;t]fxs : x;s(x) ag are bounded for a 2 (0; 1] and t 2 R+ .

© 2001 by Chapman & Hall/CRC

358

Maxingale problem

Remark 5.1.7.

Condition (NE ) implies that smooth idempotent measure on C .

x0

is a tight -

Remark 5.1.8.

By Lemma 2.8.12 and Remark 2.8.13 condition (NE ) is met when G() satis es the linear-growth condition.

Let us de ne for x 2 D

N (x) = inf ft 2 R+ : xt +t N g; N

2 N:

(5.1.3)

The next version of Lemma 2.7.5, which is proved by a similar argument, implies that N is a D{stopping time and is C {continuous.

Lemma 5.1.9. Let (Ht (x); t 2 R+ ; x 2 D ) be an R+ -valued D{ adapted function, which is continuous and increasing in t and C { continuous in x. Let for c 2 R+ (x) = inf ft 2 R+ : Ht (x)+ t cg: Then (x); x 2 D ; is a D{stopping time and is C -continuous. The following condition is a localised version of (sup E ).

(sup E )loc

1=r

P 1 sup j ln Et^N (X ) (r ) Gt^N (X ) (; X )j ! 0 tT r as 2 ; 2 Rd ; T > 0; N 2 N :

Theorem 5.1.10. Let the X satisfy (Cr), G(), for each 2 Rd ,

satisfy the uniform continuity and local majoration conditions, and (NE ) hold. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.1.11. The uniform continuity and majoration conditions used above can be somewhat modi ed. Let us say that G() satis es the continuity condition for a given 2 Rd if Gt (; x) is C { continuous in x for all t from a dense subset of R+ . Let us say that an R-valued function F = (Ft (x); t 2 R+ ; x 2 D ) obeys the strict majoration condition if (5.1.1) holds with the increments on the left-hand side replaced by their absolute values. Similarly, we can de ne the local strict majoration condition by taking absolute values on the left of (5.1.2). Since the local strict majoration condition

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

359

and the continuity condition imply the uniform continuity condition for G(), we could in Theorem 5.1.5 (respectively, Theorem 5.1.10) require only the continuity condition if we strengthened the majoration condition (respectively, local majoration condition) to the strict majoration condition (respectively, local strict majoration condition).

Since the linear-growth condition on G() implies both the local majoration condition and (NE ), we obtain the following important consequence of Theorem 5.1.10.

Theorem 5.1.12. Let the X satisfy (Cr) and the cumulant G(), d

for each 2 R , satisfy the uniform continuity and linear-growth conditions. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

The proofs of Theorems 5.1.5, 5.1.10, and 5.1.12 below show that the only property of the processes E that matters, besides being positive and predictable, is that they satisfy the assertion of Lemma 4.1.1. This observation allows us, as in Chapter 4, to extend the theorems to the case when the processes X are not necessarily semimartingales if we postulate the property stated in Lemma 4.1.1. More speci cally, let us consider the following condition on processes X with paths in D de ned on stochastic bases ( ; F ; F ; P ) .

(E )

For each 2 , there exist F {predictable positive processes E () = (Et (); t 2 R+ ); 2 Rd , such that E0 () = 1 and the processes exp( (Xt X0 ))Et () 1 ; t 2 R+ are F {local martingales.

Then we have the following extension of Theorem 5.1.12.

Theorem 5.1.13. Let X ( ; F ; F ; P ) , which satisfy d

be stochastic processes on condition (E ) , and let G(), for each 2 R , satisfy the uniform continuity and lineargrowth conditions. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Theorems 5.1.5 and 5.1.10 admit similar versions.

© 2001 by Chapman & Hall/CRC

360

Maxingale problem

Remark 5.1.14. In Theorems 5.1.5, 5.1.10, 5.1.12, and 5.1.13 we

can equivalently describe the accumulation points by saying that if is an accumulation point of fL(X ); 2 g, then the canonical process X on (C ; C; ) is a Luzin-continuous semimaxingale with cumulant G() starting at x0 . Also if G() and x0 uniquely specify ld , then X ! X.

5.1.1 Proofs In the proofs below we assume with no loss of generality that x0 = 0. We start with two preliminary lemmas. We assume the conditions imposed on the X at the beginning of the section. As above we denote by ei ; i = 1; : : : ; 2d; the d-vector, whose b(i + 1)=2cth entry equals 1 if i is odd and -1 if i is even, the rest of the entries being equal to 0.

Lemma 5.1.15. For every nite F{stopping time , a > 0, b > 0, c > 0, and u 2 R+ the following inequalities hold P sup jXt+ tu

X j a

2d exp dc (b a)

1 b + 2d max P sup (ln Et+ (cei ) ln E (cei )) > i=1;:::;2d d tu c c 2d exp d (b a) 1 b + 2d max P sup (Gt+ (cei ) G (cei )) > : i=1;:::;2d d tu c

Proof. The second inequality is implied by the rst since by (4.1.15)

ln Et+ (cei ) ln E (cei ) Gt+ (cei ) G (cei ): The rst inequality results from Lemma 3.2.6. Speci cally, let

Zt; () = Yt+ ()=Y (); t 2 R+ ; where Yt () is de ned by (4.1.18). By Lemma 4.1.1 and Doob's stopping theorem Z ; () = (Zt; (); t 2 R+ ) is an R+ -valued local martingale with respect to the ltration F; = (Ft+ ; t 2 R+ );

© 2001 by Chapman & Hall/CRC

361

Convergence of stochastic exponentials

hence, EZ; () 1 for every F; {stopping time . Lemma 3.2.6 and the de nition of Y () then yield for i = 1; : : : ; 2d c a P sup ei (Xt+ X ) exp (b a) d d tu 1 b + P sup (ln Et+ (cei ) ln E (cei )) > ; d tu c hence,

P sup jXt+ tu

X j a

a 2d i=1max P sup ei (Xt+ X ) ;:::;2d d c

tu

2d exp d (b a)

1 b + 2d max P sup (ln Et+ (cei ) ln E (cei )) > : i=1;:::;2d d tu c

Next comes one of the most technically important results of the chapter. It is more general than is required at the moment for the proofs of Theorems 5.1.5, 5.1.10 and 5.1.12, but this generality will be exploited while proving Theorems 5.2.9, 5.2.12, and 5.2.15 below. Let fX 0 ; 2 g, where X 0 = (X 0 t ; t 2 R+ ), be along with X a net of Rd {valued semimartingales de ned on ( ; F ; F ; P ). We consider the pair (X ; X 0 ) as a process with paths in the Skorohod space D 0 = D (R + ; Rd Rd ). The space D 0 is equipped with the natural ow of -algebras D0 = (Dt0 ; t 2 R+ ), de ned in analogy with D, and elements of D 0 are denoted by (x; x0 ). We denote by C 0 = C (R + ; R d Rd ) the subspace of D 0 of continuous functions equipped with the - ow C0 = (Ct0 ; t 2 R+ ) as de ned in Section 3.2. For (x; x0 ) 2 D 0 and 2 Rd , we introduce Y 0 (; (x; x0 )) = exp(xt Gt (; x0 )); t 2 R+ ; (5.1.4) t

and let

Y 0 () = (Yt0 (; (x; x0 )); t 2 R+ ; (x; x0 ) 2 C 0 ):

(5.1.5)

Deviability 0 on C 0 is said to be a solution of maxingale problem (M 0 ) if

© 2001 by Chapman & Hall/CRC

362

Maxingale problem

x00 = 0

(M 0 )

Y

(); 2 Rd ;

0 { a.e.; is a C0 {local exponential maxingale on (C 0 ; 0 ):

Theorem 5.1.16. Let G(0 ) satisfy the uniform continuity condi0 tion. If the net fL((X ; X )); 2 g is C {exponentially tight, and conditions (0) and 1=r

1 P 0 sup j ln Et (r ) Gt (; X )j ! 0 as 2 ; tT r T > 0; 2 Rd ; hold, then every LD accumulation point of fL((X ; X 0 )); 2 g (when restricted to C 0 ) is a solution to (M 0 ). Proof. Let 0 be an LD accumulation point of fL((X ; X 0 )); 2 ld g. To simplify notation, we assume that L((X ; X 0 )) ! 0 as 2 : By C 0 {exponential tightness of fL((X ; X 0 )); 2 g the deviability 0 is supported by C 0 , so it can be considered as a deviability on C 0 . We show that 0 ((x; x0 ) : x0 6= 0) = 0. Since the map 0 : (x; x0 ) ! x0 from D 0 into Rd is continuous, by the contraction principle

(sup E )0

ld 0 L(X0 ) ! Æ0 1 as 2 ;

and then by (0) and the de nition of LD convergence 0 Æ 1 (x) = 1(x = 0); x 2 Rd ; 0

which is equivalent to the required. Now we prove that the Y 0 (); 2 Rd ; are C0 {local exponential maxingales on (C 0 ; 0 ). We do that by reduction to Theorem 3.2.9. As above we denote Gt (; x0 ) = supst jGs (; x0 )j: By the uniform continuity condition on G() the function Gt (; x0 ) is C {continuous in x0 2 D for each t 2 R+ . For N 2 N and x0 2 D we introduce (x0 ) = inf ft 2 R+ : G (; x0 )_G (2; x0 )+t N g: (5.1.6) N

t

t

By Lemma 5.1.9 N (x0 ); x0 2 D ; is a nite D{stopping time and is C {continuous. Therefore, N , as a function on D 0 , is a D0 {stopping time and is C 0 {continuous.

© 2001 by Chapman & Hall/CRC

363

Convergence of stochastic exponentials

Let also, for N

2 N and 2 ;

N = inf t 2 R+ : Et (r )1=r _ Et (r ) 1=r _ Et (2r)1=r _ Et(2r ) 1=r 2eN : (5.1.7) Then by F {predictability and right continuity of Et (), N is an F{predictable stopping time (see Dellacherie [34, IV.T.16]), and by (5.1.6) and (sup E )0 lim P 1=r N 2

N (X 0 ) = 0:

The facts that N is F {predictable and N > 0 P {a.s. (since E0() = 1) imply as in the proof of Theorem 4.1.2 that there exist nite F {stopping times N such that

N < N P {a.s.

(5.1.8)

and lim P 1=r N 2

N (X 0 ) = 0:

(5.1.9)

Note that by (5.1.7) and (5.1.8) for t 2 R+ P -a.s.

Et^ (r) _ Et^ (r) 1 _ Et^ (2r ) _ Et^ (2r ) N

N

N

1

N

< 2r er N : (5.1.10)

Now, since Lemma 4.1.1 implies that Y () is a supermartingale so EY () 1 for every nite F {stopping time , we have by (5.1.10) and the de nition of Y () in (4.1.18) that

E Yt^ (r )2 = E Yt^ (2r )Et^ (2r )Et^ (r ) N N N N 23r e3Nr :

2

Thus, in view of Doob's stopping theorem, (Yt^ (r ); t 2 R+ ) is a N square-integrable martingale and for every F {stopping time

E1=r Y^ (r )2 N

© 2001 by Chapman & Hall/CRC

8e3N :

(5.1.11)

364

Maxingale problem

Next, by the respective de nitions (4.1.18) and (5.1.4) of Yt () and Yt0 (; (x; x0 )), and the inequality jeu 1j jujejuj ; u 2 R; we have that for A > 0; " > 0; > 0, and T > 0

P sup Yt0^ (; (X ; X 0 )) Yt^ (r )1=r > " N N tT P sup jjjXt j > A + P (jX0 j > ) tT

" + P sup Et^ (r ) 1=r > e A (jj) 1 e jj 2 N tT " +P sup exp( Gt^ (; X 0 )) Et^ (r ) 1=r > e A : N 2 N tT (5.1.12)

We prove that the right-hand side converges super-exponentially to 0 as 2 . By C 0 {exponential tightness of fL((X ; X 0 )); 2 g and Theorem 3.2.3 lim lim sup P1=r (sup jjjXt j > A) = 0; tT

(5.1.13)

lim P 1=r (jX0 j > ) = 0; 2

(5.1.14)

A!1 2

by (0)

and by (5.1.10)

lim lim sup P1=r sup Et^ (r ) !0 2 N tT

1=r

" > e A (jj) 1 ejj 2 = 0: (5.1.15)

Finally, (5.1.10) and (sup E )0 are easily seen to imply that

lim P1=r sup exp( Gt^ (; X 0 )) N 2 tT

Et^ (r)

" > e 2

N

A

1=r

= 0: (5.1.16)

By (5.1.13){(5.1.16) the right-hand side of (5.1.12) raised to the power of 1=r goes to 0 in the limit limA!1 lim sup!0 lim sup2 .

© 2001 by Chapman & Hall/CRC

365

Convergence of stochastic exponentials

Thus, we have proved that as 2 sup Yt0^ (; (X ; X 0 )) N tT

1=r P 1 =r Yt^ (r ) ! N

0; T > 0:

As a consequence, introducing

N = N ^ N (X 0 );

(5.1.17)

we have that for t 2 R+ 1=r P 0 0 1 =r Yt^ (; (X ; X )) Yt^ (r ) ! 0 as 2 ; N N

hence, since by (5.1.9) and (5.1.17)

lim P1=r N (X 0 ) 6= N = 0; 2 we arrive at the convergence

Yt0^ (X 0 ) (; (X ; X 0 )) Yt^ (r )1=r N N

1=r P

! 0 as 2 :

(5.1.18) Now we check the conditions of Theorem 3.2.9 with D 0 as D , (X ; X 0 ) as X , YN () = (Yt^ (r ); t 2 R+ ) as M , (x; x0 ) as x, N and Yt0^ (x0 ) (; (x; x0 )) as Mt (x). N The net fL((X ; X 0 )); 2 g is C 0 {exponentially tight by hypotheses. Next, since N (x0 ) is a D{stopping time and X 0 is F { adapted, N (X 0 ) is an F {stopping time; since N also is an F { stopping time, we conclude in view of (5.1.17) that N is an F { stopping time. Therefore, recalling that (Yt^ (r ); t 2 R+ ) is N

a square-integrable martingale with respect to F and N N , we have that YN () is a square-integrable martingale too. Moreover, in view of (5.1.11), the net fYt^ (r )1=r ; 2 g is uniN formly exponentially integrable relative to fP g for all t 2 R+ . Also Yt0^ (x0 ) (; (x; x0 )) is C 0 {continuous by the fact that, as we remarked earlier, N (x0 ) is C 0 {continuous, and Yt0 (; (x; x0 )) is continuous in

© 2001 by Chapman & Hall/CRC

366

Maxingale problem

(t; (x; x0 )) at (x; x0 ) 2 C 0 by the uniform continuity condition on G() and (5.1.4). Finally, since N (x0 ) is a D0 {stopping time, it is a C0 {stopping time if restricted to C 0 . Therefore, Yt0^ (x0 ) (; (x; x0 )) is Ct0 { measurable by Lemma 2.2.19. Since also (5.1.18) holds, we conclude that the conditions of Theorem 3.2.9 are met with the above choice of M , X and Mt (x). The theorem implies that the function (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) is a C0 {exponential maxingale on N (C 0 ; 0 ). 0 {uniform maximability of (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) is N proved in analogy with (5.1.11). Since by (5.1.6) and continuity of Gt (; x) in t we have that jGt^ (x0 ) (; x0 )j N and N jGt^N (x0 ) (2; x0 )j N; it follows by (5.1.4) that

xx

xx

sup Y 0 0 (; ( ; 0 ))2 0 (( ; 0 )) (x;x0 )2C 0 t^N (x ) e3N sup Yt0^ (x0 ) (2; ( ; N (x;x0 )2C 0

x x0 ))0 ((x; x0 )) = e3N ;

where the latter equality holds by the maxingale property of (Yt0^ (x0 ) (2; (x; x0 )); t 2 R+ ). Thus, (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) N N is 0 {uniformly maximable by Corollary 1.4.15. We are now in a position to prove Theorem 5.1.5. Proof of Theorem 5.1.5. We apply Theorem 5.1.16 with X 0 = X . All we need to prove is that under the conditions of Theorem 5.1.5 the net fL((X ; X )); 2 g is C 0 {exponentially tight in D 0 or, equivalently, the net fL(X ); 2 g is C {exponentially tight in D . We check the C {exponential tightness by verifying the conditions of part II of Theorem 3.2.3. This is carried out similarly to the argument used in the proof of Theorem 4.2.11 with the use of Lemma 5.1.15. We consider only condition II(ii) of Theorem 3.2.3, because II(i) is checked in an analogous manner. By Lemma 5.1.15 for T > 0; > 0; c > 0; 0 < Æ < 1, and 2

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

367

ST (F) P sup jXt+

X j >

tÆ

2d exp cr 2d

1 ln Et+ (r cei ) ln E (r cei ) + 2d max P sup i=1;:::;2d 2d tÆ cr c 1 ln Et (r cei ) ln Es (r cei ) 2d i=1max P sup ;:::;2d 2d s;tT +1 r sts+Æ cr + 2d exp : (5.1.19) 2d

Applying successively (sup E ) and the majoration condition on G(), we have for i = 1; : : : ; 2d 1 c sup ln Et (r cei ) ln Es (r cei ) 2d s;tT +1 r sts+Æ lim sup P1=r sup Gt (cei ; X ) Gs (cei ; X ) 3cd 2 s;tT +1 sts+Æ 1 sup Git Gis 3cd ; s;tT +1 sts+Æ

lim sup P1=r 2

where Gi = (Git ; t 2 R+ ) is a function majorising G(cei ). By continuity of Git in t the latter indicator is 0 for all small Æ > 0. Hence, by (5.1.19) c lim sup lim sup sup P1=r sup jXt+ X j > exp : 2d Æ!0 2 2ST (F ) tÆ Since c is arbitrarily large, condition II(ii) of Theorem 3.2.3 has been checked.

For a proof of Theorem 5.1.10, we need another auxiliary result which will also be used in the proof of Theorem 5.2.12. Let the maps p~N : D ! D ; N 2 N ; be de ned by (~pN x)t = xt^N (x) ; x 2 D ; t 2 R+ ;

© 2001 by Chapman & Hall/CRC

(5.1.20)

368

Maxingale problem

where the N are from (5.1.3). The maps p~N are C {continuous since the N are C {continuous and Skorohod convergence to continuous functions is equivalent to locally uniform convergence. Let also

X ;N = p~N X

(5.1.21)

and Y N () = (YtN (; x); t 2 R+ ; x 2 C ) be de ned by

YtN (; x) = Yt^N (x) (; x):

Let for N (M N )

(5.1.22)

2 N maxingale problems (M N ) on C be de ned by

x0N= 0

Y ();

2 Rd ;

N {a.e.; is a C-local exponential maxingale on (C ; N ):

Lemma 5.1.17.

Let the nets fL(X ;N ); 2 g; N 2 N ; be C {exponentially tight and every LD accumulation point of fL(X ;N ); 2 g solve (M N ). If, in addition, (NE ) holds, then

lim lim sup P1=r N (X ) t = 0; t 2 R+ : N !1 2 Proof. We rst note that since N solves (M N ), an argument similar to the one used in the proof of Lemma 2.7.11 shows that

N (x) 0;N (x) (x):

(5.1.23)

Let P1Æ=rÆ N Æ (X Æ ) t ; 2 be a subnet of 1=r P N (X ) t ; (; N ) 2 N such that lim P1Æ=rÆ N Æ (X Æ ) t 2 = lim sup lim sup P1=r (N (X ) t): (5.1.24) N !1 2 By Corollary 3.1.20 there exists a subnet f(L(X Æ Æ;N ); N 2 N ); 2 g of f(L(X Æ ;N ); N 2 N ); 2 g such that ld Æ Æ ;N N L(X ) ! ; N 2 N ; as 2 at rate rÆ Æ , where N are deviabilities on D with support in C . Since N (D n C ) = 0, we

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

369

identify N with its restriction to C . By (5.1.3) and Lemma 5.1.9 N (x); x 2 C ; is a nite C{stopping time so that the -algebra CN is well de ned. We prove that there exists a deviability on C such that (A) = N (A); A 2 CN ; N

2 N:

(5.1.25)

This is done by applying Theorem 1.8.1. Let us check that fN ; N 2 N g is a projective system of deviabilities on the same space C with the p~N as \bonding maps". In other words, we have to check that 0 ld N = 0N Æ p~N1 for N 0 > N . Since L(X Æ Æ;N 0 ) ! N 0 , X ;N = p~N X ;N (see (5.1.20) and (5.1.21)), and p~N is C {continuous, by ld the contraction principle L(X Æ Æ;N ) ! N 0 Æ p~N1 . Since also ld N L(X Æ Æ;N ) ! , by uniqueness of an LD limit N = N 0 Æp~N1 . In order to apply Theorem 1.8.1 we have to check the (; K )condition. By the rst part of condition (NE ) it is suÆcient to check that KN 0 () p~N 0 K0 () for arbitrary 2 (0; 1] and N 0 2 N . 0 Let xN 2 KN 0 (). The fact that fN ; N 2 N g is a projective system of deviabilities allows us to construct functions xN 2 C0 ; N 0= N 0 ; N 0 + 1; : : : ; such that p~N xN +1 = xN and N (xN ) = N (xN ). Since the sequence fN (xN )g is increasing, it converges to a limit L. Since by (5.1.23) 0;N (xN ) (xN ) N (xN ) and the sequence fxNN (xN ) + N (xN ); N = N 0 ; N 0 + 1; : : :g is unbounded, the second part of condition (NE ) implies that L0 = 1. Hence, there exists a 0 function x^ 2 C that coincides with xN on [0; N 0 (xN )] and coincides with the xN on [N 1 (xN ); N (xN )] for N = N 0 +1; N 0 +2; : : :. Since 0;N (^x) (^x) = 0;N (xN )(xN ) and 0;N (^x)(^x) ! 0 (^x) as N ! 1, we conclude that 0(^x) as required. Hence, by Theorem 1.8.1 there exists a deviability on C such that N = Æ p~N1 , which is equivalent to (5.1.25) by Lemma 2.2.21. Since N (X ) = N (X ;N ), N Æ Æ ! 1 as 2 , the set fx 2 D : N (x) tg is C {closed, N (D nC ) = 0, and fx 2 C : N (x) tg 2 CN , we have by (5.1.24), Corollary 3.1.9, and (5.1.25) that for arbitrary N 0 2 N lim sup lim sup P1=r (N (X ) t) N !1 2 = lim P1Æ=rÆÆ Æ N Æ Æ (X Æ Æ ) t 2

© 2001 by Chapman & Hall/CRC

370

Maxingale problem

lim sup P1Æ=rÆÆ Æ (N 0 (X Æ 2

Æ;N 0 ) t) N 0 (N 0 (x) t)

= (N 0 (x) t): (5.1.26) By the -smoothness property of deviability lim (N (x) t) = N !1

\

N 2N

fx 2 C : N (x) tg = 0;

which together with (5.1.26) proves the lemma. Proof of Theorem 5.1.10. We begin by showing that the nets fL(X ;N ); 2 g; N 2 N ; de ned in (5.1.21), are C {exponentially tight and their respective LD accumulation points solve (M N ). We rst check that fX ;N ; 2 g satis es the conditions of Theorem 5.1.5 with GN () = (Gt^N (x) (; x); t 2 R+ ; x 2 D ) as G(). Condition (0) is obvious. Next, X ;N has as its stochastic exponential the process E ;N () = (Et^N (X ) (); t 2 R+ ). Hence, (sup E )loc implies (sup E ) for X ;N with GN () as G(). Now we check that GN () satis es the conditions imposed in Theorem 5.1.5 on G(). Let us consider the majoration condition. Since by the de nition of N , x(t^N (x)) N; x 2 D , we have that, for 0 s t,

sup(Gt^N (x) (; x) Gs^N (x) (; x))

x2D

=

sup

x2D : (t^N ( )) N

x

(Gt^N (x) (; x) Gs^N (x) (; x)): (5.1.27)

x

By Remark 5.1.3 and the facts that G() is D{adapted and t ^ N (x) is a D{stopping time, the right-hand side of (5.1.27) equals sup(Gt^N (x) (; x) Gs^N (x) (; x)) over x 2 D such that x1 N ; so by the local majoration condition on G() (say, with GN for given ) if x1 N , then

Gt^N (x) (; x) Gs^N (x) (; x) GNt^N (x)

GNs^N (x) GNt GNs ;

where for the last inequality we used that GNt is increasing in t. Hence, sup(Gt^N (x) (; x) Gs^N (x) (; x)) GNt GNs ; x2D

© 2001 by Chapman & Hall/CRC

(5.1.28)

371

Convergence of stochastic exponentials

proving the majoration condition for GN (). Next, obviously, Gt^N (x) (; x) is Dt {measurable in x 2 D and continuous in t. We check that it is C {continuous in x uniformly over t 2 [0; T ] for arbitrary T > 0. Let xn ! x 2 C . We again x and denote by GN the associated local majorant for G(). For arbitrary " > 0, by continuity of GNt and Gt (; x) in t we can choose Æ > 0, Æ < N (x) ^ 1, such that sup jGNu u;vT ju vjÆ

GNv j "; sup jGt^(N (x) tT

Since N (xn ) ! N (x) as n large enough to have

Æ) (;

x)

Gt^N (x) (; x)j ":

! 1 by Lemma 2.7.5, we can take n

jN (xn) N (x)j Æ; and then, for t T ,

(5.1.29)

Gt^N (xn ) (; xn ) Gt^N (x) (; x) (Gt^N (xn )(; xn ) Gt^(N (x) Æ) (; xn )) + jGt^(N (x) Æ) (; xn ) Gt^(N (x) Æ) (; x)j + jGt^(N (x) Æ) (; x) Gt^N (x) (; x)j (Gt^N (xn )(; xn ) Gt^(N (x) Æ) (; xn )) + jGt^(N (x) Æ) (; xn ) Gt^(N (x) Æ) (; x)j + ": (5.1.30) By (5.1.29), (5.1.28) and the choice of Æ

Gt^N (xn ) (; xn ) Gt^(N (x) Æ) (; xn ) = Gt^N (xn ) (; xn ) Gt^(N (x) Æ)^N (xn ) (; xn ) GNt^N (xn ) GNt^(N (x)

": Therefore, (5.1.30) yields by C {continuity of the mapping x ! (Gt (; x); t 2 R+ ) lim sup sup(Gt^N (xn ) (; xn ) Gt^N (x) (; x)) 2": n!1 tT

The complementary inequality

lim sup sup(Gt^N (x) (; x) Gt^N (xn ) (; xn )) 2" n!1 tT

© 2001 by Chapman & Hall/CRC

Æ)

372

Maxingale problem

is proved similarly if we choose Æ > 0 such that 2Æ N +1 (x) N (x), sup jGNu +1 GNv +1 j "; u;vT ju vjÆ sup jGt^(N (x)+Æ) (; x) Gt^N (x) (; x)j "; tT and consider n for which, in addition to (5.1.29), N (x) + Æ N +1 (xn ). Thus, fX ;N ; 2 g and GN (); 2 Rd ; satisfy all the conditions of Theorem 5.1.5. Hence, the net fL(X ;N ); 2 g is C { exponentially tight, and if N is an LD accumulation point, then x0 = 0 N { a.e.Nand the function Y N () = (Y Nt (; x); t 2 R+ ; x 2 C ) de ned by Y t (; x) = exp( xt Gt^N (x) (; x)) is a C{local exponential maxingale on (C ; N ). Therefore, to prove that N solves (M N ), it is left to show that N Y t (; x) = YtN (; x) N {a.e., which in view of (5.1.22) and the de nition of Y follows by the equality

xt^N (x) = xt

N {a.e.

(5.1.31)

To see the latter, let fX 0 ;N ; 0 2 0 g be a subnet of fX ;N ; 2 g that LD converges to . Since Xt;N = Xt;N ^N (X ;N ) by (5.1.20) and N (5.1.21), is supported by C , N (x) is C {continuous, and the set fx 2 D : xt^N (x) 6= xt g is C {open, by Corollary 3.1.9 0

0

N 0 = lim sup P0 0 Xt ;N 6= Xt^;N 0 ;N ) (xt 6= xt^N (x) ); ( X N 0 0 2 which proves (5.1.31). Thus, the nets fX ;N ; 2 g; N 2 N ; satisfy the conditions of Lemma 5.1.17. By the lemma for T > 0 1=r

lim lim sup P1=r (N (X ) T ) = 0;

(5.1.32)

N !1 2

and using (5.1.20) and (5.1.21) we have that

lim lim sup P1=r sup jXt Xt;N j > 0 = 0; N !1 2 tT

© 2001 by Chapman & Hall/CRC

373

Convergence of characteristics

which implies by C {exponential tightness of fL(X ;N ); 2 g for every N 2 N and Theorem 3.2.3 that fL(X ); 2 g is C { exponentially tight. Also (5.1.32) and (sup E )loc imply (sup E ). Thus, all the conditions of Theorem 5.1.16 with X 0 = X hold. An application of that theorem ends the proof. Theorem 5.1.12 follows by Theorem 5.1.10 and Remark 5.1.8.

5.2 Convergence of characteristics This section formulates conditions on convergence of the characteristics of the X in order for the net fL(X ); 2 g to be exponentially tight with all the LD accumulation points being solutions of a maxingale problem. We retain the above notation. As in Section 4.2 the cumulant in the limiting maxingale problem will have the semimaxingale representation (2.7.7) and (2.7.55), however, the characteristics can depend on x, on the one hand, and are de ned for x 2 D , on the other hand.

De nition 5.2.1. Let us say that a function f : R+ D ! Rk is D{progressively measurable if its restriction to [0; t] D is B([0; t])

k Dt =B(R )-measurable.

We assume as given the following objects: (bs (x); s 2 R+ ; x 2R D ) is an Rd -valued D-progressively measurable function such that 0tjbs (x)jds < 1 for t 2 R+ and x 2 D , cs (x); s 2 R+ ; x 2 D is a D-progressively measurable function with values in the R t space of symmetric, positive semi-de nite d d-matrices such that 0 kcs (x)k ds < 1 for t 2 R+ and x 2 D , s( ; x); s 2 R+ ; 2 B(Rd ); x 2 D is a transition kernel from ([0; t] D ; B ([0; t]) Dt ) into (Rd ; B(Rd )) for every t 2 R+ such that for t 2 R+ ; x 2 D and 2 R+ Z

Rd

j j ^ 1 t(dx; x) < 1; x2

Z

Rd

ejxj 1(jxj > 1) t (dx; x) < 1;

t (f0g; x) = 0; jxj2 ^ 1 t (x) < 1; ejxj 1(jxj > 1) t (x) < 1;

^s( ; x); s 2 R+ ; 2 B(Rd ); x 2 D is a transition kernel from ([0; t] D ; B (R+ ) Dt ) into (Rd ; B(Rd )) for every t 2 R+ such that

© 2001 by Chapman & Hall/CRC

374

Maxingale problem

for s 2 R+ ; x 2 D and

2 B(Rd )

^s ( ; x) s ( ; x); ^s(Rd ; x) 1:

(5.2.2)

Since D-progressively measurable functions are C-progressively measurable, the restrictions of (bs (x)); (cs (x)); (s ( ; x)), and (^s ( ; x)) to C satisfy the conditions on the local characteristics of a semimaxingale as de ned in Section 4.2. Also, given a limiter h : Rd ! Rd we de ne extensions of the characteristics of a semimaxingale to D by

B 0 (x) =

Zt

bs (x) ds;

(5.2.3)

Bt (x) = Bt0 (x) + (h(x) x) t (x);

(5.2.4)

t

Ct (x) =

0

Zt

0

cs (x) ds;

(5.2.5)

and refer to B 0 = (Bt0 (x); t 2 R+ ; x 2 D ) as the rst characteristic \without truncation" of the limiting semimaxingale, to B = (Bt (x); t 2 R+ ; x 2 D ) as the rst characteristic associated with limiter h(x), to C = (Ct (x); t 2 R+ ; x 2 D ) as the second characteristic, to s ( ; x) as the density of the measure of jumps, and to ^s( ; x) as the density of the discontinuous measure of jumps. If B = (Bt (x); t 2 R+ ; x 2 D ) is the rst characteristic associated with a limiter h (x), then

Bt (x) = Bt (x)+(h (x) h(x)) t (x):

The modi ed second characteristic C~ = (C~t (x); t speci ed by the equality

(5.2.6)

2 R+ ; x 2 D ) is

C~t (x) = Ct (x) + ( h(x))2 t (x) Zt

0

( h(x) ^s (x))2 ds; 2 Rd : (5.2.7)

We introduce a number of conditions on the characteristics, which are analogues of the continuity and majoration conditions on G() in Section 5.1. As above, we denote by U a dense subset of R+ .

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

375

De nition 5.2.2. We say that B (respectively, C ; C~ ; ; or ^) satis es the continuity condition if Bt (x) (respectively, Ct (x); C~t (x); f (x) t (x) for fR : Rd ! R Borel measurable and such that jf (x)j 1 ^ jxj2 ; 0t (g(x) ^s(x))k ds for g : Rd ! R Borel measurable and bounded, and k = 2; 3; : : :) is C {continuous in x for all t 2 U.

Remark 5.2.3.

If the continuity condition on holds, then by (5.2.6) the continuity condition on B does not depend on a limiter. If, in addition, the continuity condition on ^ holds, then by (5.2.7) the continuity conditions on C and C~ are equivalent.

Occasionally, we will need a stronger form of the continuity condition on B which is an analogue of the uniform continuity condition for G().

De nition 5.2.4. We say that B satis es the uniform continuity condition if the map x ! (Bt (x); t 2 R+ ) is C {continuous as a map from

D

into C .

Remark 5.2.5. Since Ct (x); C~t (x) and f (x) t (x), if f 0,

are increasing, continuous in t and equal to 0 at 0, the continuity conditions on C , C~ and are equivalent to C {continuity of the respective maps x ! (Ct (x); t 2 R+ ) from D into C (R + ; Rdd ), x ! (C~t (x); t 2 R+ ) from D into C (R + ; Rdd ) and x ! (f (x) t (x); t 2 R + ), where jf (x)j 1 ^ jxj2 , from D into C (R + ; R ). Thus, the continuity conditions on C , C~ and imply the associated uniform continuity conditions. Therefore, we will sometimes also be referring to the continuity conditions for C , C~ and as uniform continuity conditions. De nition 5.2.6. We say that B (respectively, C , C~ , or ) satis es the majoration condition (respectively, the local majoration condition) if the functions ( Bt (x); t 2 R+ ; x 2 D ) for all 2 Rd ; ( Ct (x); t 2 R+ ; x 2 D ) for all 2 Rd ; ( C~t (x); t 2 R+ ; x 2 D ) for all 2 Rd ; (f (x) t (x); t 2 R+ ; x 2 D ) for all f 2 Cb satisfy the majoration condition (respectively, the local majoration condition).

Remark 5.2.7. The majoration condition (respectively, the local majoration condition) on B equivalently requires that the function (Vart B (x); t 2 R+ ; x 2 D ) of total variation of B obey the majoration condition (respectively, the local majoration condition). The

© 2001 by Chapman & Hall/CRC

376

Maxingale problem

majoration conditions (respectively, the local majoration conditions) on C and C~ are equivalent to the majoration conditions (respectively, local majoration conditions) on the respective functions of the sums of the diagonal entries of Ct and C~t .

De nition 5.2.8. We say that satis es the C {local boundedness condition if for every compact K C and every > 0; t > 0; sup ejxj 1(jxj > 1) t (x) < 1: (5.2.8) x2K

We say that ^ satis es the C {local boundedness condition if for every compact K C and every > 0; t > 0; sup sup ejxj ^s (x) < 1: (5.2.9) x2K st

We now state conditions on the triplets of the X . They are similar to the conditions of Section 4.2. Actually, the conditions on the X0 and big jumps are the same. We repeat them here for completeness. Let (B ; C ; ) be the predictable characteristics of X corresponding to a limiter h(x). As above, x0 2 Rd . (0) (A)

1=r P X0 ! x0 as 2 ; lim lim sup P1=r ([0; t]; jxj > A) 1=r A!1 2

> " = 0;

t > 0; " > 0;

(a)

(sup (C ) (C~ ) ( )

lim lim sup P1=r

a!1 2

1 r jxj e 1 (r jxj > a) 1(jxj A) t > " r = 0; t > 0; > 0; A > 0; " > 0;

1=r P B) sup jBt Bt (X )j ! 0 as 2 ; T > 0; tT lim lim sup P1=r ( kr Ct;Æ Ct (X )k > ") = 0; t 2 U; " > 0, Æ!0 2 1=r P ~ ~ kr Ct Ct(X )k ! 0 as 2 ; t 2 U; 1=r P f (x) t f (x) t (X ) ! 0 as 2 ; t 2 U; f 2 Cb ;

© 2001 by Chapman & Hall/CRC

377

Convergence of characteristics

1 X (^ ) f (r x) s k r 0<st

Zt

f (x) ^s

0

1=r k P (X ) ds !

0 as 2 ;

t 2 U; k = 2; 3; : : : ; f

2 Cb :

(We recall that f (x) = f (r x)=r .) Theorem 5.2.9. Let h(x) be continuous, B , C (respectively, C~ ), , and ^ satisfy the continuity conditions, and and ^ satisfy the C {local boundedness conditions. Let also the majoration conditions on B , C (respectively, C~ ) and hold. If conditions (0); (A) + (a); (sup B ); (C ) (respectively, (C~ )), ( ), and (^ ) hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point is a solution to the maxingale problem (x0 ; G). Remark 5.2.10. Condition (a) can be replaced with the condition (a0 )

where ja;A;

1=r 1 j (x) t;c lim lim sup P a!1 2 r a;A; 1 X (x) > " = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0; (x) = (er jxj 1) 1(r jxj > a) 1(jxj A):

Remark 5.2.11. Theorems 5.2.9 and 2.8.5 imply Theorem 4.2.1. Next comes a locally bounded version. Let us de ne N (x) by (5.1.3) and introduce the conditions (A)loc (a)loc

lim lim sup P1=r ([0; t ^ N (X )]; jxj > A)1=r > " = 0; A!1 2 t > 0; N 2 N ; " > 0; 1=r er jxj 1(r jxj > a) 1(jxj A) t^N (X ) lim lim sup P a!1 2 r > " = 0; t > 0; N 2 N ; > 0; A > 0; " > 0;

(sup B )loc

sup jBt^N (X ) tT

© 2001 by Chapman & Hall/CRC

1=r P Bt^N (X ) (X )j !

0 as 2 ;

T > 0; N

2 N;

378

Maxingale problem

(C )loc lim lim sup P1=r kr Ct;Æ ^N (X ) Æ!0 2 (C~ )loc

kr C~t^N (X )

( )loc

f (x) t^N (X )

(^ )loc

1 r st^

X N (X )

Ct^N (X ) (X )k > " = 0; t 2 U; N

C~t^N (X ) (X )k

1=r P

2 N ; " > 0;

! 0 as 2 ; t 2 U; N 2 N ; P

1=r

f (x) t^N (X ) (X ) ! 0 as 2 ; t 2 U; N 2 N ; f 2 Cb ; k

f (r x) s

t^ZN (X ) 0

f (x) ^s (X ) k ds 1=r P

! 0 as 2 ; t 2 U; N 2 N ; k = 2; 3; : : : ; f 2 Cb :

Theorem 5.2.12. Let h(x) be continuous.

Let B , C (respectively, ~ C ), , and ^ satisfy the continuity conditions, and and ^ satisfy the C {local boundedness conditions. Let the local majoration conditions on B , C (respectively, C~ ) and hold. Let condition (NE ) hold. If conditions (0); (A)loc + (a)loc ; (sup B )loc ; (C )loc (respectively, (C~ )loc ), ( )loc , and (^ )loc hold, then the net fL(X ); 2 g is C { exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.2.13. Condition (a)loc can be replaced with the condition (a0 )loc

1=r 1 lim lim sup P j (x) t;c ^N (X ) a!1 2 r a;A; X 1 (x) > " = 0; + ln 1 + ja;A; s r st^ (X ) N t > 0; N 2 N ; > 0; A > 0; " > 0:

De nition 5.2.14. We say that bs(x), respectively, cs(x), meets the linear-growth condition if there R exists an R + -valued Lebesgue measurable function ls such that 0t ls ds < 1; t 2 R+ ; and

jbs(x)j (1+ xs )ls ;

© 2001 by Chapman & Hall/CRC

(5.2.10)

379

Convergence of characteristics

respectively,

kcs (x)k (1+(xs )2 )ls:

(5.2.11)

We say that meets the linear-growth condition if Z

1 x)s (dx; x)

(ex

Rd

Z

(e(1+xs )x

Rd

1 (1 + xs ) x)ms (dx); 2 Rd ; (5.2.12)

where ms (dx) is aR transition kernel from (R+ ; B(R+ )) into tR d d (R ; B(R )) such that 0 Rd (exp(jxj) 1 jxj)ms (dx)ds < 1; t > 0; > 0. Theorem 5.2.15. Let h(x) be continuous, B , C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let the linear-growth conditions on bs (x), cs (x), and s( ; x) hold. Let ^ satisfy the C {local boundedness condition. If conditions (0); (A)loc + (a)loc ; (sup B )loc ; (C )loc (respectively, ~ (C )loc ), ( )loc , and (^ )loc hold, then the net fL(X ); 2 g is C { exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.2.16. 0 (a )loc .

Condition (a)loc can be replaced with condition

Remark 5.2.17. In the above theorems we can equivalently describe the accumulation points by saying that if is an LD accumulation point of fL(X ); 2 g, then the canonical process X on (C ; C; )

is a Luzin-continuous semimaxingale with characteristics (B; C; ; ^) ld starting at x0 . If (B; C; ; ^) and x0 uniquely specify , then X ! X.

The proofs use the ideas of the proofs of Theorems 4.2.1, 5.1.5, and 5.1.10. An outline is as follows: we introduce truncated processes X^ ;a as in the proof of Theorem 4.2.1, establish that the pairs (X^ ;a ; X ) as random elements of D 0 (= D (R + ; Rd Rd )) satisfy the conditions of Theorem 5.1.16, then observe in view of Lemma 4.2.16

© 2001 by Chapman & Hall/CRC

380

Maxingale problem

that condition (A) + (a) implies that the nets f(X ; X ); 2 g and f(X^ ;a ; X ); 2 g asymptotically (as a ! 1) have the same LD limit, and derive the statements of Theorems 5.2.9, 5.2.12 and 5.2.15 by taking the limit as a ! 1 in the maxingale problems associated with the nets f(X^ ;a ; X ); 2 g. This turns out to be quite a long way. The next subsection studies required exponential tightness properties. After, LD accumulation points are identi ed as solutions to certain maxingale problems, and nally the proofs of the above results are given. We assume in the proofs that x0 = 0.

5.2.1 Exponential tightness results

We develop exponential tightness results for (X^ ;a ; X ). The next lemma extends Lemma 4.2.6. Lemma 5.2.18. Let Z ;Æ = (Zt;Æ ; t 2 R+ ); Z0;Æ = 0; Æ > 0; 2 ; and Z = (Zt ; t 2 R+ ); Z0 = 0; 2 ; be Rd {valued, componentwise increasing processes with paths in D de ned on respective probability spaces ( ; F ; P ) such that for all t 2 U; " > 0; lim lim sup P1=r (jZt;Æ Ztj > ") = 0:

Æ!0 2

If the net fL(Z ); 2 g is C {exponentially tight, the latter convergence is uniform on bounded intervals, i.e.,

lim lim sup P1=r (sup jZt;Æ Zt j > ") = 0; T > 0; " > 0: Æ!0 2 tT Proof. Acting as in the proof of Lemma 4.2.6 for N 2 N we choose tNi 2 U; i = 0; : : : ; kN , such that 0 = tN0 < tN1 < : : : < tNkN 1 < T tNkN < T + 1 and jtNi tNi 1j 1=N; i = 1; : : : ; kN . Then by the same argument sup jZt;Æ Zt j max N jZt;Æ N ZtN j+ sup jZt Zs j: i i i=1;:::;k tT s;tT +1: js tj1=N Therefore,

P1=r

sup jZt;Æ tT

© 2001 by Chapman & Hall/CRC

Zt j > "

kN X

i=1 1=r + P

P1=r jZt;Æ N i

sup jZt s;tT +1: js tj1=N

ZtNi j > "=2

Zs j > =2

381

Convergence of characteristics

so that by hypotheses lim lim sup P1=r (sup jZt;Æ Æ!0 2 tT

Zt j > ")

lim sup P1=r 2

The right-hand side tends to 0 as N

sup jZt s;tT +1: js tj1=N

Zs j > =2

! 1 by Theorem 3.2.3.

We now study, as in the proof of Theorem 4.2.1, when one can replace conditions (C ), (C~ ) and ( ) with the associated uniform versions. Let us introduce the conditions: (sup C )

lim lim sup P1=r (sup kr Ct;Æ tT

Æ!0 2

(sup C~ )

sup kr C~t tT

(sup )

sup jf (x) t tT

C~t (X )k

Ct (X )k > ") = 0;

1=r

" > 0; T > 0;

P

! 0 as 2 ; T > 0;

f (x) t

1=r P (X )j !

0 as 2 ;

f 2 Cb ; T > 0: In some of the statements below we say, with a slight abuse of terminology, that nets of laws of Rm {valued random processes are C { exponentially tight if they are C (R + ; Rm ){exponentially tight. The meaning should be clear from the context. Lemma 5.2.19. 1. Let the net fL(C (X )); 2 g (respectively, fL(C~ (X )); 2 g; fL(f (X )); 2 g; f 2 Cb) be C { exponentially tight. Then condition (C ) (respectively, (C~ ); ( )) is equivalent to condition (sup C ) (respectively, (sup C~ ); (sup )). 2. If the nets fL(f (x) (X )); 2 g; f 2 Cb ; are C { exponentially tight and condition ( ) holds, then condition (sup B ) does not depend on the particular choice of a continuous limiter h. Proof. The rst part follows by Lemma 5.2.18. The second part follows from the rst and is proved in the same way as Lemma 4.2.8.

© 2001 by Chapman & Hall/CRC

382

Maxingale problem

Now, we state and prove an exponential tightness result. We recall the de nition of X^ ;a = (X^t;a ; t 2 R+ ); a 2 R+ ; from Subsection 4.2.2:

X^ t;a = Xt X t;a ;

(5.2.13)

where

Xt;a = ha (x) =

X

ha (Xs ));

(5.2.14)

a 1 ^ 1 x; ha (x) = ha (r x): jxj r

(5.2.15)

st

(Xs

We also recall that C 0 = C (R + ; Rd Rd ).

Theorem 5.2.20.

1. Let the nets fL(B (X )); 2 g, fL(C (X )); 2 g (respectively fL(C~ (X )); 2 g ), and fL(f (x) (X )); 2 g; f 2 Cb; be C {exponentially tight. If conditions (0), (sup B ), (C ) (respectively, (C~ )), and ( ) hold, then the net fL(X^ ;a ); 2 g is C {exponentially tight. If, in addition, conditions (A) + (a) hold, then the net fL(X ); 2 g is C {exponentially tight, so, the net fL((X^ ;a ; X )); 2 g is C 0 {exponentially tight.

2. Let the net fL(X ); 2 g be C {exponentially tight. If the function B (respectively, C ; C~ ; ) satis es the uniform continuity condition, then the net fL(B (X )); 2 g (respectively, fL(C (X )); 2 g; fL(C~ (X )); 2 g; fL(f (x) (X )); 2 g; f 2 Cb ) is C {exponentially tight. Proof. Part 2 follows from Theorem 3.2.3 via a diagonal argument. Let, given T > 0 and > 0, fL(X 0 ); r0 ; p0 ; 0 2 0 g be a subnet of fL(X ); r ; P1=r sups;t2[0;T ]:jBt (X ) Bs (X )j > ; (; Æ) 2 js tjÆ ld 0 (0; 1)g such that L(X 0 ) ! at rate r0 , where 0 is supported by C , and

lim p0 = lim lim sup P1=r Æ!0 2 0 20

© 2001 by Chapman & Hall/CRC

sup jBt (X ) Bs (X )j > : s;t2[0;T ]: js tjÆ

383

Convergence of characteristics

Then for arbitrary Æ^ > 0 by the contraction principle (Corollary 3.1.15) and uniform continuity condition for B

lim lim sup P1=r Æ!0 2

sup jBt (X ) Bs (X )j > = lim p0 0 20 s;t2[0;T ]: js tjÆ 0 0 1=r0 lim0 sup P0 sup jBt (X ) Bs(X )j > 20 s;t2[0;T ]: js tjÆ^ 0 x 2 C : sup jBt (x) Bs(x)j : s;t2[0;T ]: js tjÆ^

By -smoothness of 0 the latter deviability tends to 0 as Æ^ ! 0. This checks condition I(ii) of Theorem 3.2.3. Condition I(i) holds since B0 (x) = 0 completing the proof of C -exponential tightness of fL(B (X )); 2 g. Proofs for the other processes are similar. We prove part 1. Let us assume, rst, that the nets fL(B (X )); 2 g, fL(C (X )); 2 g and fL(f (x) (X )); 2 g; f 2 Cb ; are C {exponentially tight. We begin with a proof of ^ ;a ); 2 g. Let C {exponential tightness for fL(X

Xt = X0 + Bt;Æ + Mt;Æ + x 1(r jxj > Æ) t be the canonical representation of X associated with the truncation function x 1(r jxj Æ), where Æ < a, so that B ;Æ = (Bt;Æ ; t 2 R+ ); B0;Æ = 0, is an F {predictable process with bounded variation over bounded intervals; M ;Æ = (Mt;Æ ; t 2 R+ ); M0;Æ = 0, is the F {locally squareintegrable martingale de ned by

Mt;Æ = Xt;c + x 1(r jxj Æ) ( )t : Since by (5.2.14) and (5.2.13) X^ t;a = Xt Æ < a, we have

(5.2.16) (x

ha (x)) t and

X^ t;a = X0 + Bt;Æ + Mt;Æ + ha (x) 1(r jxj > Æ) t ; so, by Theorem 3.2.3 and condition (0) in order to prove C { exponential tightness of fL(X^ ;a ); 2 g it suÆces to prove that,

© 2001 by Chapman & Hall/CRC

384

Maxingale problem

for all T > 0; > 0, lim lim sup P1=r (sup jBt;Æ j > A) = 0; A!1 2 tT

(5.2.17a)

= 0;

(5.2.17b) (5.2.17c)

lim lim sup sup P 1=r (sup jBt;Æ + !0 2 2ST (F ) t

Bt;Æ j > )

lim lim lim sup P1=r (sup Mt;Æ > A) = 0; Æ!0 A!1 2 tT 1=r ;Æ lim lim sup lim sup sup P (sup Mt;Æ + Mt Æ!0 !0 t 2 2ST (F )

k

k

k

= 0;

j1

lim lim sup P1=r ( ha (x) A!1 2

j

(r jxj > Æ) T

= 0;

lim lim sup sup P1=r !0 2 2ST (F )

Z+Z

Rd

k > ) (5.2.17d)

> A) (5.2.17e)

jha (x)j 1(r jxj > Æ)

(ds; dx) > = 0:

(5.2.17f)

We begin with B ;Æ . Let B~ ;Æ be the rst characteristic of X associated with hÆ (hÆ is de ned as ha with Æ = a). Then by (4.1.4)

Bt;Æ = B~t;Æ +(x 1(r jxj Æ) hÆ (x)) t ; so, for 0 s < t, using (5.2.15),

jBt;Æ Bs;Æ j jB~t;Æ B~s;Æ j Z t Z + s Rd

Æ x 1 (r jxj > Æ) (ds; dx) r jxj

jB~t;Æ

B~s;Æ j +

Zt Z

s

Rd

f~Æ(x) (ds; dx); (5.2.18)

where f~Æ (x) = (2jxj=Æ 1)+ ^ 1 (Recall that, by the notation introduced in Section 4.2, f~Æ(x) = f~Æ (r x)=r .)

© 2001 by Chapman & Hall/CRC

385

Convergence of characteristics

Let B~tÆ (x) be the rst characteristic of X associated with the limiter hÆ (x) so that it is de ned by (5.2.4) with hÆ (x) as h(x). Then in view of (5.2.6) fL(B~ Æ (X )); 2 g is C {exponentially tight since fL(B (X )); 2 g and fL((hÆ (x) h(x)) (X ))); 2 g are both C {exponentially tight, hence, by (sup B ) with hÆ in place of h and Theorem 3.2.3 the net fL(B~ ;Æ ); 2 g is C {exponentially tight. Since f~Æ 2 Cb , the net fL(f~Æ (X )); 2 g is C {exponentially tight by hypotheses so that by Lemma 5.2.18 condition (sup ) holds with f = f~Æ . Theorem 3.2.3 implies that the net fL(f~Æ ); 2 g is C {exponentially tight. Inequality (5.2.18), C {exponential tightness of fL(B~ ;Æ ); 2 g and fL(f~Æ ); 2 g imply by Theorem 3.2.3 that the B ;Æ satisfy (5.2.17a) and (5.2.17b). Now we prove (5.2.17e) and (5.2.17f). The process (jha (x)j 1(r jxj > Æ) t ; t 2 R+ ) has as its stochastic cumulant the process (exp(jha (x)j) 1) 1(r jxj > Æ) t ; t 2 R+ ; 2 R. Then for 2 ST (F ) and c > 0 by the second inequality in Lemma 5.1.15 with d = 1

P

Z+Z

Rd

jha (x)j 1(r jxj > Æ) (ds; dx) >

2 exp r c2

1 + 2 max P i=1;2 r

Z+Z

Rd

exp(( 1)i jha (r x)jc) 1 1(r jxj > Æ)

(ds; dx) >

c ; 2

hence,

P1=r

Z+Z

Rd

jha (x)j 1(r jxj > Æ) (ds; dx) >

© 2001 by Chapman & Hall/CRC

21=r exp c2

386

Maxingale problem

+ 21=r P1=r

Z+Z

f (x) (ds; dx) >

Rd

c ; (5.2.19) 2

where f (x) = (exp(cjha (x)j) 1)f~Æ (x). Since f (x) belongs to Cb , the net fL(f (x) (X )); 2 g is C {exponentially tight by hypotheses. From (sup ) and Theorem 3.2.3 we derive that the net fL(f (x) ); 2 g is C {exponentially tight. Then (5.2.19) and Theorem 3.2.3 imply (5.2.17f). Limit (5.2.17e) is proved similarly. Now we prove (5.2.17c) and (5.2.17d). Denoting the stochastic cumulant associated with M ;Æ by G~ ;Æ () = (G~ ;Æ t (); t 2 R + ) and applying Lemma 5.1.15, we reduce the proof of (5.2.17c) and (5.2.17d) to the proof of the respective limits (i = 1; : : : ; 2d; T > 0; > 0) 1 lim lim lim sup P1=r sup G~ ;Æ ( r e ) > A = 0; (5.2.20a) i t Æ!0 A!1 2 tT r 1 lim lim sup lim sup sup P1=r sup (G~ ;Æ t+ (r ei ) Æ!0 !0 r t 2 2ST (F ) G~ ;Æ ( r e )) >

= 0: (5.2.20b) i

Now, by (5.2.16) the measure of jumps of M ;Æ is

~;Æ ([0; t];

)=

X

0<st

1

Z Rd

x 1(r jxj Æ)( (fsg; dx)

(fsg; dx)) 2

nf0g ; 2 B(Rd );

and then by (4.1.14) and the fact that the F {predictable quadraticvariation process of X ;c is C 1 ;Æ 1 G~ ;Æ t () = 2 Ct + exp (x 1(r jxj Æ ) xs ) (x 1(r jxj Æ) x;Æ s ) t ; (5.2.21) where

x;Æ s =

Z Rd

x 1(r jxj Æ) (fsg; dx):

© 2001 by Chapman & Hall/CRC

(5.2.22)

387

Convergence of characteristics

We note that, since by (4.1.3a) (fsg; R d ) 1, we have Æ jx;Æ (5.2.23) s j r : Applying Taylor's formula to the integrand in (5.2.21), we obtain 1 ~ ;Æ G (r ) = Tt;Æ ()+ rt;Æ (); (5.2.24) r t where 1 1 2 Tt;Æ () = (r Ct)+ r ((x 1(r jxj Æ) x;Æ s )) t 2 2 (5.2.25) and jje2jjÆ r ((x 1(r jxj Æ) x;Æ ))2 sup jrs;Æ ()j Æ s t 3 st (5.2.26) (for the last inequality we also used (5.2.23)). Since by (5.2.22) 2 2 ( (x 1(r jxj Æ) x;Æ s )) t = ( x) 1(r jxj Æ ) t X 2 d ( x;Æ s ) (2 (fsg; R )); 0<st

where the sum is over s such that (fsg; R d ) > 0, and (fsg; R d ) 1, we have by (4.1.7), (5.2.22) and (5.2.25) that 0 Tt;Æ () Ts;Æ () r (Ct;Æ Cs;Æ ); s < t; and by (4.1.7) and (5.2.26) that jje2jjÆ r C ;Æ : sup jrs;Æ ()j Æ t 3 st

(5.2.27) (5.2.28)

Since fL(C (X )); 2 g is C {exponentially tight, an application of (sup C ) yields in view of Theorem 3.2.3

lim lim sup P1=r kr CT;Æ+1 k > A = 0; Æ!0 2 A!1 ;Æ lim lim sup sup P1=r sup r kCt;Æ + C k > = 0; Æ!0 2 2ST (F ) t !0

© 2001 by Chapman & Hall/CRC

388

Maxingale problem

for every > 0. The rst of these relations together with (5.2.28) implies that lim lim sup P1=r ( sup jrt;Æ ()j > ") = 0; " > 0; tT +1

Æ!0 2

(5.2.29)

and together with (5.2.27) that lim lim lim sup P1=r ( sup Tt;Æ () > A) = 0; Æ!0 A!1 2 tT +1

(5.2.30)

while the second one and (5.2.27) yield ;Æ lim lim sup lim sup sup P1=r (sup jTt;Æ + () T ()j > ") t 2 2ST (F ) = 0; " > 0: (5.2.31)

Æ!0 !0

In view of (5.2.24), limits (5.2.29) and (5.2.30) prove (5.2.20a), while limits (5.2.29) and (5.2.31) prove (5.2.20b). C {exponential tightness of the net fL(X^ ;a ); 2 g has been proved. C {exponential tightness of the net fL(X ); 2 g under (A) + (a) follows by Theorem 3.2.3 and Lemma 4.2.16. Finally, by Corollary 3.2.7 C 0 { exponential tightness of the net fL((X^ ;a ; X )); 2 g is implied by C {exponential tightness of both fL(X^ ;a ); 2 g and fL(X ); 2 g. We now assume that instead of condition (C ) and C {exponential tightness of fL(C (X )); 2 g we have condition (C~ ) and C { exponential tightness of fL(C~ (X )); 2 g. As above, it is suf cient to prove C {exponential tightness of fL(X^ ;a ); 2 g; a > 0: We again consider a canonical representation of X , but this time with respect to hÆ (x) from (5.2.15) with Æ < a so we replace a by Æ in (5.2.15) and substitute throughout in the preceding argument hÆ (x) for x 1(r jxj Æ). Retaining the above notation for the components of the representation, we again reduce the task to proving (5.2.17a){(5.2.17f). Limits (5.2.17a) and (5.2.17b) follow by (sup B ) and C { exponential tightness of fL(B (X )); 2 g and fL(f (X )); 2 g; f 2 Cb . The proofs of (5.2.17e) and (5.2.17f) do not change. The proofs of (5.2.17c) and (5.2.17d) also proceed along the same lines, (5.2.27) and (5.2.28) being replaced by (with hÆ (x) in place of

© 2001 by Chapman & Hall/CRC

389

Convergence of characteristics

x 1(r jxj Æ) in (5.2.22), (5.2.25) and (5.2.26)) 0 Tt;Æ () Ts;Æ ()

r(C~t;Æ C~s;Æ ); (5.2.32) jj sup jrs;Æ ()j Æ e2jjÆ r C~t;Æ ; (5.2.33) 3 st

where C~t;Æ is de ned by (4.1.8) with hÆ (x) as h(x). By (4.1.8) for 0 s t

r (C~t;Æ Z Z

+ r

(s;t] Rd

+ r

C~s;Æ ) r (C~t C~s)

(

X

s 0 be such that h(x) = x; jxj c: Then we obtain for Æ < c, recalling the notation jjhjj = sup jh(x)j, x2Rd Z Z

(s;t] Rd

j( h(x))2 ( hÆ (x))2 j (du; dx)

Z j j2 2 2 (jjhjj + Æ )

r2

Z

1(r jxj > c) (du; dx)

(s;t] Rd Z Z 2

+j

j

(s;t] Rd

jxj2 1(Æ < rjxj c) (du; dx)

and X

s 0; > 0; > 0; and Æ < 1 ^ c ^ jjhjj, by (sup C~ ) and (sup ) lim sup P1=r sup j r (C~t;Æ C~s;Æ )j > 2 js tj s;tT lim sup P1=r sup j r(C~t (X ) C~s(X ))j > 6 2 js tj s;tT

Z Z

+ lim sup P1=r 10jj2 jjhjj2 sup 1(jxj > c=2) 2 js tj (s;t] d R s;tT (du; dx)(X ) > 6 Z Z 1=r 2 + lim sup P 4jj sup jxj2 1(Æ=2 < jxj 2c) 2 js tj (s;t] d R s;tT (du; dx)(X ) > ; 6 where the right-hand side goes to 0 as ! 0 by C {exponential tightness of fL(C~ (X )); 2 g and fL(f (X )); 2 g; f 2 Cb ; and Theorem 3.2.3. By (5.2.32) this proves (5.2.31). Limits (5.2.29) and (5.2.30) are proved similarly. Thus, C {exponential tightness of fL(X^ ;a ); 2 g under the new set of assumptions has been proved. The theorem has been proved.

© 2001 by Chapman & Hall/CRC

391

Convergence of characteristics

Remark 5.2.21. According to the theorem, under (0), (a) + (A),

(sup B ), (C ) (or (C~ )), and ( ), and the uniform continuity conditions on B , C (or C~ ), and , C {exponential tightness of the net fL(X ); 2 g is equivalent to C {exponential tightness of the nets fL(B (X )); 2 g, fL(C (X )); 2 g (or fL(C~ (X )); 2 g ), and fL(f (x) (X )); 2 g; f 2 Cb .

5.2.2 LD accumulation points as solutions to maxingale problems In this subsection, assuming that either the net fL((X^ ;a ; X )); 2 g is C 0 {exponentially tight or the net fL(X ); 2 g is C { exponentially tight, we characterise their LD accumulation points as solutions to maxingale problems. The rst step is to consider small jumps as in Theorem 4.2.11. As in Subsection 4.2.2 the triplet of X^ ;a without truncation is (B ;a ; C ; ;a ), where B ;a is the rst characteristic of B corresponding to ha and

;a ([0; t]; ) = ([0; t]; (ha ) 1 ( ));

2 B(Rd ); t 2 R+ :

We also note that by (5.2.15) and (5.2.35)

;a ([0; t]; fr jxj > ag) = 0; t 2 R+ ; 2 :

(5.2.35) (5.2.36)

Since the jumps of X^ ;a are bounded in modulus by a=r , its stochastic exponential is well de ned. We denote it by E^;a () = (E^t;a (); t 2 R+ ); 2 Rd . The \limiting" semimaxingale X a is de ned in analogy with Subsection 4.2.2 as having characteristics (B a ; C; a ; ^a ) relative to ha , which are de ned for x 2 D and 2 B(Rd ) by

Bta (x) = Bt0 (x) + (ha (x) x) t (x); ta ( ; x) = t (ha 1 ( ); x); ^ta (x) = ^t (ha 1 ( ); x):

(5.2.37) (5.2.38)

We also note that

ta (fjxj > ag; x) = 0:

© 2001 by Chapman & Hall/CRC

(5.2.39)

392

Maxingale problem

The associated cumulant is given by 1 G^ at (; x) = Bta (x)+ Ct (x) +(ex 1 ha (x)) ta (x) 2 +

Zt

ln 1 + (ex

0

1) ^sa (x)

ex

1

^sa (x) ds: (5.2.40)

Obviously G^ at (; x) is continuous in t and Dt {measurable in x. Let (M^ a ) denote the maxingale problem (M 0 ) introduced in Subsection 5.1.1 with G() replaced with G^ a () = (G^ at (; x); t 2 R+ ; x 2 ^ a on C 0 solves (M^ a ) if D ); 2 Rd , i.e., deviability x0a = 0 d ^ a {a.e. (M^ a ) Y^ (); 2 R ; is a C 0 {local exponential maxingale on (C 0 ; ^ a ); where Y^ a () = (Y^ta (; (x; x0 )); t 2 R+ ; (x; x0 ) 2 C 0 ) is de ned by

Y^ta (; (x; x0 )) = exp( xt G^ at (; x0 )): (5.2.41) Theorem 5.2.22. Let the net fL((X^ ;a ; X )); 2 g be C 0 { exponentially tight, let B satisfy the uniform continuity condition, and let C (respectively C~ ), and ^ satisfy the continuity conditions. If conditions (0), (sup B ), (C ) (respectively (C~ )), ( ), and (^ ) hold, then every LD accumulation point of fL((X^ ;a ; X )); 2 g solves (M^ a ).

The idea of the proof is to apply Theorem 5.1.16 to the pair (X^ ;a ; X ) (note that X^ ;a plays the role of X in Theorem 5.1.16, and X plays the role of X 0 ). We again need auxiliary results. The second part of the next lemma extends Lemma 4.2.9.

Lemma 5.2.23. I. Let the net fL(X ); 2 g be C {exponentially

tight and satisfy the continuity condition. Then, for c > 0; " > 0 and t > 0 1.

lim lim sup P1=r jxj2 1(jxj Æ) t (X ) " = 0; Æ!0 2

2.

lim lim sup P1=r Æ(jxj ^ c) 1(jxj Æ) t (X ) " = 0: Æ!0 2

© 2001 by Chapman & Hall/CRC

393

Convergence of characteristics

II. If, in addition, conditions ( ) and (^ ) hold, then for " > 0 and t2U 1 1. lim lim sup P1=r f (r x) 1(r jxj > Æ) t f (x) t (X ) Æ!0 2 r >" =0 for all R+ -valued bounded continuous functions f (x); x 2 Rd ; such that f (x) cjxj2 in a neighbourhood of 0 for some c > 0; Æ 2. lim lim sup P1=r jg(r x)j 1(r jxj > Æ) s > " = 0; Æ!0 2 r and 1 X g(r x) 1(r jxj > Æ) s k 3. lim lim sup P1=r Æ!0 2 r 0<st Zt

g(x) ^s (X ) k ds > " = 0; k = 2; 3; : : :

0

for all R-valued, bounded and continuous functions g(x); x 2 Rd ; such that jg(x)j cjxj in a neighbourhood of 0 for some c > 0. Proof. We begin with part I. We denote

HÆ (x) = jxj2 1(jxj Æ) t (x);

x 2 D:

(5.2.42)

Let a subnet f(L(X 0 ); r0 ; p0 ); 0 2 0 g of f(L(X ); r ; P1=r HÆ (X ) " ); (; Æ) 2 (0; 1)g be such that

lim p0 = lim sup lim sup P1=r HÆ (X ) " 0 0 2 Æ!0 2

ld 0 and L(X 0 ) ! at rate r0 , where deviability 0 is supported by C . Then for arbitrary > 0 and t 2 U

lim sup lim sup P1=r HÆ (X ) " = lim p0 0 20 Æ!0 2 1=r 0 lim0 sup P0 0 H (X ) " 0 x 2 C : H (x) " ; 20 where the latter inequality follows by Corollary 3.1.9 since the continuity condition on implies that H (x) is C {upper-semi-continuous. Upper semi-continuity of H (x) on C implies that the sets fx 2 C :

© 2001 by Chapman & Hall/CRC

394

Maxingale problem

H (x) "g are closed. They converge to ; as ! 0 by Lebesgue's bounded convergence theorem. The -smoothness property of devia bility then implies that lim!0 0 x 2 C : H (x) " = 0: The rst assertion of part I is proved. For the second, picking > 0, we have for Æ small enough and x 2 D, Æ(jxj ^ c) 1(jxj Æ) t (x) Æ(jxj ^ c) 1(jxj ) t (x) + jxj2 1(jxj ) t (x); (5.2.43) where for the second summand, which is well de ned by (2.7.53a) and (2.7.54), we used Chebyshev's inequality. Let k(x) = (2jxj= 1)+ ^ 1; x 2 Rd . Since the function (jxj^ c)k(x) belongs to Cb and we can assume that t 2 U , C {exponential tightness of fL(X ); 2 g and the continuity condition on imply by Corollary 3.1.22 that the net fL((jxj ^ c)k(x) t (X )); 2 g is exponentially tight in R; hence, by Theorem 3.2.3, since 1(jxj ) k(x), " lim lim sup P1=r Æ(jxj^c) 1(jxj )t (X ) = 0: Æ!0 2 2 Inequality (5.2.43) and the rst assertion of part I yield the required. Part I is proved. Part II is proved in analogy with Lemma 4.2.9, necessary modi cations make use of part I and are obvious. The next lemma follows by Lemma 5.2.23 in the same way as Lemma 4.2.10 follows by Lemma 4.2.9.

Lemma 5.2.24. Let the net fL(X ); 2 g be C -exponentially tight

and satisfy the continuity condition. Then, under conditions ( ) and (^ ), condition (C ) is equivalent to condition (C~ ), which, hence, does not depend on the choice of h.

Lemma 5.2.25. Under the uniform continuity condition on B aand

the continuity conditions on C (or C~ ), and ^, the function G^ () satis es the uniform continuity condition. Proof. Since under the continuity conditions on and ^ the continuity conditions for C and C~ are equivalent, we can assume that the

© 2001 by Chapman & Hall/CRC

395

Convergence of characteristics

continuity condition on C holds. We prove that each of the functions on the right of (5.2.40) is C {continuous as a map from D into C (R + ; R ). The function B a has the required property by hypotheses and the fact that, in view of the continuity condition on , the uniform continuity condition for B does not depend on a limiter. The same fact is clearly also true for the next two terms on the right-hand side of (5.2.40) (use (5.2.38) for the third term). Let us consider the last term. Recalling the notation (x) = x ln(1 + x); x > 1 (seeR (4.2.26)), we have, since (x) 0, that it t x a is suÆcient to show that 0 (e 1) ^s (x) ds is C {continuous for each t 2 U . Let xn ! x^ 2 C and " > 0. By (5.2.39)

e jja 1

Z

(ex 1)^sa (dx; x) ejja 1;

Rd

x 2 D:

Also by Weierstrass' theorem there exists a polynomial q(u) = Pl k q(u)j " for k=2 dk x ; l 2; u 2 R+ ; such that j (u) x 2 [exp( jja) 1; exp(jja) 1]. Since by the continuity condition on ^ and (5.2.38) lim n!1

Zt

q

(ex

0

x

1)^sa ( n ) ds =

Zt

0

q (ex 1)^sa (^x) ds;

we obtain that Zt lim sup

n!1

(ex

0

1) ^sa (xn ) ds

Zt

(ex

0

1) ^sa (^x) ds 2":

Since " is arbitrary, the lemma is proved. Proof of Theorem 5.2.22. Since by Lemma 5.2.25 the function G^ a () satis es the uniform continuity condition, it is suÆcient to prove in view of Theorem 5.1.16 that 1=r

P 1 sup j ln E^t;a (r ) G^ at (; X )j ! 0 as 2 ; T > 0: tT r (5.2.44)

© 2001 by Chapman & Hall/CRC

396

Maxingale problem

The idea is to follow the proof of Theorem 4.2.1 in order to derive (5.2.44) from convergence of the characteristics of the X^ ;a ; 2 : Let C^t;a;Æ denote the modi ed second characteristic of X^ ;a corresponding to the truncation function x 1(jxj Æ) so that

C^t;a;Æ = Ct + ( x)2 1(r jxjÆ) t;a X x 1(r jxjÆ) s;a 2 ; 2 Rd : (5.2.45) 0<st

We note that the following conditions hold (sup (C a ) ( a ) (^ a )

1=r P ;a a sup jBt Bt (X )j ! 0 as 2 ; T > 0; tT lim lim sup P1=r ( kr C^t;a;Æ Ct (X )k > ") = 0; t 2 U; " > 0, Æ!0 2 1=r P ;a a f (x) t f (x) t (X ) ! 0 as 2 ; t 2 U; f 2 Cb ; 1=r Zt k k P 1 X ;a a f (r x) s f (x) ^s (X ) ds ! 0 r 0<st 0 as 2 ; t 2 U; k = 2; 3; : : : ; f 2 Cb :

Ba)

Proof is by the argument of the proof of Lemma 4.2.13. Condition (sup B a ) is actually condition (sup B ) with ha (x) as h(x) and holds since (sup B ) does not depend on the choice of a limiter by Lemma 5.2.19 and Theorem 5.2.20. Since by (5.2.35) for Æ < a

;a ([0; t];

\fr jxj Æg) = ([0; t]; \fr jxj Æg);

(5.2.45) and (4.1.7) imply that C^t;a;Æ = Ct;Æ when Æ < a, so that conditions (C a ) and (C ) coincide. Similarly, the argument of the proof of Lemma 4.2.13 shows in view of (5.2.38) and (5.2.35) that ( a ) and (^ a ) are implied by ( ) and (^ ). We now prove that conditions (sup B a ), (C a ), ( a ), and (^ a ) imply (5.2.44). This is carried out analogously to the proof of Theorem 4.2.11. Therefore, we do not give all the details but only indicate modi cations that have to be made in that proof. For this reason, we extensively use the notation of the proof of Theorem 4.2.11. Let us rst note that by (5.2.37) and (5.2.38) the functions Bta (x) and ta ( ; x) satisfy the same continuity conditions as imposed on

© 2001 by Chapman & Hall/CRC

397

Convergence of characteristics

Bt (x) and t ( ; x) in the statement of Theorem 5.2.22. Also by Lemma 5.2.19 conditions (sup C a ) and (sup a ) hold (with obvious notation). We now de ne in analogy with (4.2.14), (4.2.15a){ (4.2.15h), substituting ;a for , as = ;a (fsg; Rd ); and, for Æ > 0 and 2 Rd ,

x;Æ = x 1(r jxj Æ) s;a; s Ds;Æ () = (ex 1) 1(r jxj > Æ) s;a ; Rs;Æ () = exp( (x 1(r jxj Æ) x;Æ s )) 1 ;Æ (x 1(r jxj Æ) xs ) s;a ; ;Æ ;Æ Q;Æ s () = (exp( xs ) 1 + xs )(1 as ); ;Æ ;Æ ;Æ ;Æ G;Æ s () = exp( xs )Ds () + Rs () + Qs (); Ut;Æ () = (ex 1 x) 1(r jxjÆ) s;a;c; Vt;Æ () = (ex 1 x) 1(r jxj>Æ) s;a ; where t 2 R+ ; s 2 R+ , and ;a;c(ds; dx) is the continuous part of ;a (ds; dx). Let, as in Lemma 4.2.12,

Yt;Æ () = Zt;Æ ()

=

X

0<st X

(Ds;Æ ());

(5.2.46)

1 ;Æ ln(1 + G;Æ s ()) + Ut () + 2 Ct 0<st X

0<st

ln(1 + Ds;Æ ());

and, as in (4.2.23){(4.2.25),

Vt (; x) = (ex 1 x) ta (x); Yt (; x) = Zt (; x) =

© 2001 by Chapman & Hall/CRC

Zt

0

(ex

1 Ct (x) 2

1) ^sa (x) ds;

398

Maxingale problem

(we \bar" here Yt (; x) not to confuse it with earlier notation). All the quantities above are well de ned by the same argument as in Subsection 4.2.1. Then exactly as in the proof of Theorem 4.2.11 convergence (5.2.44) would hold provided for every T > 0 and " > 0

) )

) Æ)

sup jBt;a tT

1=r P a Bt (X )j !

0 as 2 ;

1 lim lim sup P1=r sup j Vt;Æ (r ) Vt (; X )j > " = 0; Æ!0 2 tT r 1 lim lim sup P1=r sup j Yt;Æ (r ) Yt (; X )j > " = 0; Æ!0 2 tT r 1 lim lim sup P1=r sup j Zt;Æ (r ) Zt (; X )j > " = 0: Æ!0 2 tT r

Limit ) is just (sup B a ) which we have already proved. For part ), we rst note that by part II.1 of Lemma 5.2.23 applied to fL(X^ ;a ); 2 g and a, in which hypotheses boundedness of the associated function f follows by (5.2.36) and (5.2.39), we have that 1 lim lim sup P1=r Vt;Æ (r ) Vt (; X ) > " = 0: Æ!0 2 r

Since by Theorem 3.2.3 and (5.2.39) the net fL((Vt (; X ); t 2 R + )); 2 g is C {exponentially tight, Lemma 5.2.18 implies ). We prove ) by the argument of the proof of part ) in Theorem 4.2.11 (a similar argument we have already used in the proof of Lemma 5.2.25). Theorem 3.2.3 implies in view of (5.2.2) and (5.2.39) that the net fL(Yt (; X ); t 2 R+ ); 2 g is C {exponentially tight. Therefore, by (5.2.46) we have, in view of Lemma 5.2.18, that ) would follow from 1 lim lim sup P1=r j Yt;Æ (r ) Yt (; X )j > " = 0; Æ!0 2 r " > 0; t 2 U: (5.2.47)

Next, noting that by (5.2.36) e jja 1 (er x 1) s;a ejja 1 and in view of (5.2.39), we have as in the proof of ) while proving

© 2001 by Chapman & Hall/CRC

399

Convergence of characteristics

Theorem 4.2.11 that (5.2.47) is implied by 1 X lim lim sup P1=r D;Æ (r )k Æ!0 2 r 0<st s Zt

0

and

(ex

1) ^sa (X ) k ds > = 0;

> 0; k = 2; 3; : : : ; (5.2.48)

1 X lim lim sup lim sup P1=r Ds;Æ (r )2 > A = 0; A!1 Æ!0 r 0<st 2 (5.2.49)

lim lim sup P1=r A!1 2

Zt

0

jex 1j^sa (X ) 2ds > A = 0;

(5.2.50) where t 2 U . Limit (5.2.50) follows by (5.2.39). Limit (5.2.49) is easily deduced from (^ a ) and (5.2.50). Limit (5.2.48) follows by part II.2 of Lemma 5.2.23, (5.2.36) and (5.2.39). Part ) is proved. To prove Æ), we introduce as in the proof of Theorem 4.2.11 1 2 ;a j (x 1(r jxj Æ) x;Æ s )j s 2 1 2 + j x;Æ s j (1 as ); 2 1 Wt;Æ () = j xj2 1(r jxj Æ) t;a;c: 2

Hs;Æ () =

Then by (5.2.45) and the de nitions of x;Æ s and as X 1 ^ ;a;Æ 1 Ct = Ct + Wt;Æ ()+ Hs;Æ () 2 2 0<st

so that by the de nitions of Zt;Æ () and Zt (; x) convergence (sup C a ) implies that Æ) would follow by 1 Æ0 ) lim lim sup P1=r sup j (Ut;Æ (r ) Wt;Æ (r ))j > " = 0; Æ!0 2 tT r " > 0; T > 0;

© 2001 by Chapman & Hall/CRC

400

Æ00 )

Maxingale problem

1 X ;Æ lim lim sup P1=r j ln(1 + G;Æ s (r )) (Hs (r ) Æ!0 2 r 0<st

+ ln(1 + Ds;Æ (r )))j > " = 0; " > 0; t > 0:

0

0

Limit Æ ) is proved as Æ ) in the proof of Theorem 4.2.11 if we note that by Theorem 5.2.20 the net fL(C (X )); 2 g is C { exponentially tight so that by (C a ) and Theorem 3.2.3 lim lim sup lim sup P1=r (jr C^t;a;Æ j > A) = 0: (5.2.51) A!1 Æ!0 2 As for Æ00 ), the argument is again as in the proof of Theorem 4.2.11. We rst note that by part II.3 of Lemma 5.2.23 and (5.2.36) Æ X j Ds;Æ (r )j > " = 0; " > 0: lim lim sup P1=r Æ!0 2 r 0<st Now the rest of the proof is the same as the proof of Æ00 ) in the proof of Theorem 4.2.11 (with the use of (5.2.51) in due place). Thus, ), ), ), and Æ) have been proved, and (5.2.44) has been proved. By Theorem 5.1.16, we have thus proved Theorem 5.2.22 under conditions (sup B ), (C ), ( ), and (^ ). The fact that (C ) can be replaced by (C~ ) follows by Lemma 5.2.24. Theorem 5.2.22 has been proved.

Theorem 5.2.26. Let the net fL(X ); 2 g be C {exponentially

tight, let B satisfy the uniform continuity condition, and let C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let also and ^ satisfy the C {local boundedness conditions (5.2.8) and (5.2.9). If conditions (0), (A) + (a), (sup B ), (C ) (respectively, (C~ )), ( ), and (^ ) hold, then every LD accumulation point of fL(X ); 2 g is a solution of (0; G). The idea of the proof is to \take the limit as a ! 1" in Theorem 5.2.22. The next lemma proves that limits in the weak topology of solutions to (M^ a ) solve (x0 ; G). Lemma 5.2.27. Let and ^ satisfy the C {local boundedness conditions (5.2.8) and (5.2.9), B satisfy the uniform continuity condition, and C , and ^ satisfy the continuity conditions. Let deviabilities ^ a ; a > 0; on C 0 solve (M^ a ). Let be a deviability on C and devia ^ on C 0 be de ned as ( ^ x; x0 ) = (x) 1(x = x0 ); (x; x0 ) 2 C 0 : bility iw ^ as a ! 1, then solves (0; G). If ^ a !

© 2001 by Chapman & Hall/CRC

401

Convergence of characteristics

Proof. We begin by proving that for every compact K C

lim sup sup jG^ as (; x) Gs (; x)j = 0; t > 0:

(5.2.52)

a!1 x2K st

By the de nitions of G^ a () and G()

jG^ as (; x) Gs(; x)j ejxj 1(jxj > a) s(x) Z t +

1) ^s (x) ds

1 + (eha (x)

0

Zt

1) ^s (x) ds (5.2.53)

1 + (ex

0

The rst term on the right tends to 0 as a ! 1 uniformly over x 2 K by (5.2.8). Since by (5.2.9) as in the proof of Lemma 4.2.17 lim sup sup ex 1(jxj > a) ^s(x) = 0; a!1 x2K st Z ha (x) 1) ^ (x) > 0; lim inf inf inf 1 + ( e s a!1 x2K st Rd

the second term on the right of (5.2.53) also tends to 0. Limit (5.2.52) has been proved. Now, for r 2 R+ and (x; x0 ) 2 C 0 we introduce the C0 {stopping times

r (x; x0 ) = inf ft 2 R+ : G (; x0 )_x +t rg (5.2.54) t

t

and

r;a (x; x0 ) = inf ft 2 R+ : G^ at (; x0 )_xt +t rg; a > 0: (5.2.55) By Lemma 5.2.25 G^ a () satis es the uniform continuity condition, and (5.2.52) then implies that the map x ! (Gt (; x); t 2 R+ ) is a continuous map from C into C (R + ; R). Therefore, by Lemma 5.1.9

r and r;a are continuous in (x; x0 ) 2 C 0 , also by (5.2.52) for every compact K 0 C 0 lim

sup

a!1 (x;x0 )2K 0

j r;a(x; x0 ) r (x; x0 )j = 0; r 2 R+ :

© 2001 by Chapman & Hall/CRC

(5.2.56)

402

Maxingale problem

^ a is a solution to (M^ a ), Y^ a () de ned by (5.2.41) is a C0 { Since local exponential maxingale on (C 0 ; ^ a ). It is also continuous in the time variable, hence, by part 2 of Lemma 2.3.13 and continuity of

r;a (x; x0 ) the function Y^ a;N () = Y^ta^ r;a (x;x0 ) (; (x; x0 )); t 2 R+ is a C0{local exponential maxingale on (C 0 ; ^ a ). Since by (5.2.41) and (5.2.55) Y^ a;N () is bounded, we conclude that Y^ a;N () is a C0 { uniformly maximable exponential maxingale on (C 0 ; ^ a ). Hence, for all 0 s < t and every R+ -valued continuous and bounded Cs0 { measurable function f (x; x0 ) ^ a (x; x0 ) sup Y^ta^ r;a (x;x0 ) (; (x; x0 ))f (x; x0 ) (x;x0 )2C 0 ^ a (x; x0 ): (5.2.57) = sup Y^sa^ r;a(x;x0 ) (; (x; x0 ))f (x; x0 ) (x;x0 )2C 0 We now prove that the equality is preserved on taking in both sides the limits as a ! 1. More precisely, we prove that for every R+ valued bounded continuous function f (x; x0 ) on C 0 and t > 0 lim sup Y^ a r;a 0 (; (x; x0 ))f (x; x0 )^ a (x; x0 ) a!1 (x;x0 )2C 0 t^ (x;x ) ^ x; x0 ); (5.2.58) = sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( 0 0 (x;x )2C where Y 0 () is de ned by (5.1.4), i.e., Yt0 (; (x; x0 )) = exp( xt Gt (; x0 )): As a rst step, we prove that lim sup jY^ a r;a 0 (; (x; x0 )) Yt0^ r (x;x0 ) (; (x; x0 ))j a!1 (x;x0 )2C 0 t^ (x;x ) ^ a (x; x0 ) = 0: (5.2.59) Let K 0 be a compact in C 0 . By (5.2.56) and Arzela{Ascoli's theorem lim sup jx r;a 0 xt^ r (x;x0 ) j = 0; a!1 (x;x0 )2K 0 t^ (x;x ) and by (5.2.52), (5.2.56) and continuity of Gt (; x) in (t; x) 2 R+ C lim sup jG^ a r;a 0 (; x0 ) Gt^ r (x;x0 ) (; x0 )j = 0; a!1 (x;x0 )2K 0 t^ (x;x ) which imply that lim sup jY^ a r;a 0 (; (x; x0 )) Yt0^ r (x;x0 ) (; (x; x0 ))j = 0: a!1 (x;x0 )2K 0 t^ (x;x ) (5.2.60)

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

403

Let fak ; k 2 N g be a subsequence, along which lim supa!1 of the iw ^ supremums in (5.2.59) is attained. Since ^ ak ! as k ! 1, by a ^ k Theorem 1.9.27 the sequence f ; k 2 N g is tight. Given " > 0, we choose K 0 such that lim supk!1 ^ ak (C 0 nK 0 ) < ": Then, since Y^ta^ r;a (x;x0 ) (; (x; x0 )) and Yt0^ r (x;x0 ) (; (x; x0 )) are bounded above by e(1+jj)r , lim sup sup Y^ta^k r;ak (x;x0 ) (; (x; x0 ))^ ak (x; x0 ) < "e(1+jj)r ; k!1 (x;x0 )2C 0 nK 0 ^ ak (x; x0 ) < "e(1+jj)r : lim sup sup Yt0^ r;ak (x;x0 ) (; (x; x0 )) k!1 (x;x0 )2C 0 nK 0 These inequalities and (5.2.60) imply (5.2.59) (recall that ^ a (x; x0 ) 1). iw ^ Next, using the convergence ^ a ! and the fact that 0 0 Yt^ r (x;x0 ) (; (x; x )) is bounded and continuous in (x; x0 ) 2 C 0 , we have by the de nition of idempotent weak convergence lim sup Y 0 r 0 (; (x; x0 ))f (x; x0 )^ a (x; x0 ) a!1 (x;x0 )2C 0 t^ (x;x ) ^ x; x0 ); = sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( (x;x0 )2C 0 which by (5.2.59) concludes the proof of (5.2.58). Equalities (5.2.57) and (5.2.58) imply that ^ x; x0 ) sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( 0 0 (x;x )2C ^ x; x0 ); = sup Ys0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( (x;x0 )2C 0 ^ x; x0 ) = (x) 1(x = x0 ), we conclude that and, since ( (Yt^ r (x;x) (; x); x 2 C ; t 2 R+ ) satis es the maxingale property with respect to . Being bounded, it is a C{uniformly maximable exponential maxingale on (C ; ). Since also (Yt (; x); x 2 C ; t 2 R+ ) is C{adapted and r (x; x); x 2 C ; is a continuous C{stopping time, it follows that Y () is a C{local exponential maxingale on (C ; ). The equality (x0 6= 0) = 0 holds since ^ a ((x; x0 ) : x0 6= 0) (( ^ x; x0 ) : x0 6= 0) 0 = lim inf a!1 = (x : x0 6= 0):

© 2001 by Chapman & Hall/CRC

404

Maxingale problem

Proof of Theorem 5.2.26. By (5.2.7) and Lemma 5.2.24 it is suÆcient to prove the part of the statement that concerns C . Let be a deviability on D supported by C that is an LD accumulation point of fL(X ); 2 g so that along a subnet ld L(X 0 ) ! :

(5.2.61) By Theorem 5.2.20 the nets fL(X ; X ); 2 g are C 0 { exponentially tight for all a > 000. Therefore, by Corollary 3.1.20 00 ;a ^ there exists a subnet f(L(X ; X ); a > 0); 00 2 00 g of 0 ;a 0 ^ f(L(X ; X ); a > 0); 0 2 0g such that for every a > 0 ^ 0 ;a

0

ld ^ a L(X^ 00 ;a ; X 00 ) ! ;

(5.2.62)

ld ^ L(X 00 ; X 00 ) ! ;

(5.2.63)

where ^ a are deviabilities on D 0 supported by C 0 . Convergence (5.2.61) implies that in D 0 ^ is the deviability on D 0 de ned by where ^ x; x0 ) = (x) 1(x = x0 ); (x; x0 ) 2 D 0 : ( Also by Lemma 4.2.16 we have that for every > 0 1=r 0 ^ ;a ; X )) > = 0; lim lim sup P (( X ; X ) ; ( X S a!1 2

(5.2.64) where 0S is the Skorohod{Prohorov{Lindvall metric on D 0 . Limits (5.2.62), (5.2.63) and (5.2.64) imply by Lemma 3.1.37 that iw ^ ^a ! ^ a solves problem (M^ a ). as a ! 1: By Theorem 5.2.22 An application of Lemma 5.2.27 ends the proof.

5.2.3 Proofs of the main results Proof of Theorem 5.2.9. By the majoration conditions in the theorem and Theorem 3.2.3 the nets fL(B (X )); 2 g, fL(C (X )); 2 g (respectively, fL(C~ (X )); 2 g), and fL(f (X )); 2 g; f 2 Cb ; are C {exponentially tight. Then by Theorem 5.2.20 the net fL(X ); 2 g is C {exponentially tight. Also the continuity and majoration conditions on B imply the uniform continuity condition. An application of Theorem 5.2.26 concludes the proof.

© 2001 by Chapman & Hall/CRC

405

Convergence of characteristics

Proof of Theorem 5.2.12. The argument is similar to the one in the proof of Theorem 5.1.10. We give only the main points. Let X ;N ; 2 ; N 2 N ; be de ned by (5.1.21). Then it is veri ed analogously to the proof of Theorem 5.1.10 that the net fX ;N ; 2 g satis es the conditions of Theorem 5.2.9 for every N 2 N with Bt (x); Ct (x); C~t (x); t (dx; x), and ^t (dx; x) replaced, respectively, by BtN (x); CtN (x); C~tN (x); tN (dx; x), and ^tN (dx; x) de ned as

BtN (x) = Bt^N (x) (x); CtN (x) = Ct^N (x) (x); C~tN (x) = C~t^N (x) (x); tN (dx; x) = t (dx; x) 1(t N (x)); ^tN (dx; x) = ^t (dx; x) 1(t N (x)):

In particular, the majoration conditions are checked in a manner similar to the proof of Theorem 5.1.10; checking the continuity conditions is simple (note, however, that the proof of the continuity condition for ^N uses the inequality ^s(Rd ; x) 1). By Theorem 5.2.9 the nets fL(X ;N ); 2 g; N 2 N ; are C { exponentially tight. Let N ; N 2 N ; be their respective LD accumulation points. It follows as in the proof of Theorem 5.1.10 that N solves (M N ). Condition (NE ) yields by Lemma 5.1.17 lim lim sup P1=r (N (X ) t) = 0; t 2 R+ ;

N !1 2

which implies, again as in the proof of Theorem 5.1.10, that fL(X ); 2 g is C {exponentially tight. The uniform continuity condition for B follows by the continuity and local majoration conditions. An application of Theorem 5.2.26 ends the proof. Proof of Theorem 5.2.15. By the de nition of the cumulant for 0 < s 0, we introduce

Xt;A = Xt

X

0<st

Xs 1(jXs j > A):

Since X ;A = (Xt;A ; t 2 R+ ) has bounded jumps, it satis es the Cramer condition. Let G;A () = (G;A t (); t 2 R + ) be the asso;A ciated stochastic cumulant and E () = (Et;A (); t 2 R+ ) be the stochastic exponential of G;A (). Then the following holds.

Theorem 5.3.1. Let the cumulant G() satisfy the uniform continuity condition and the linear-growth condition. Let condition (0) hold and, for some A > 0,

© 2001 by Chapman & Hall/CRC

407

LD convergence results

([0; t]; jxj

(A0 ) (sup

E A)

1=r P 1 =r > A)

! 0

as 2 ; t > 0;

1=r

1 )j P! 0 sup j ln Et;A ( r ) G ( ; X t^N (X ) ^N (X ) tT r as 2 ; T > 0; N 2 N ; 2 Rd :

Then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves the maxingale problem (x0 ; G). If the ld latter problem has the unique solution x0 , then L(X ) ! x0 . Proof. According to the proof of Lemma 4.2.16 condition (A0 ) implies that

P sup jXt Xt;A j > 0 tT

Therefore, condition (sup

! 0 as 2 ; T > 0:

(5.3.1)

E A) implies that 1=r

P 1 ;A ;A sup j ln Et^N (X ;A ) (r ) Gt^N (X ;A ) (; X )j ! 0 tT r as 2 ; T > 0; N 2 N ; 2 Rd :

By Theorem 5.1.12 the net fL(X ;A ); 2 g is C {exponentially tight, and its every LD accumulation point solves the maxingale problem (x0 ; G). By Theorem 3.2.3 and (5.3.1) the net fL(X ); 2 g is C {exponentially tight. By Lemma 3.1.38 and (5.3.1) every LD accumulation point of fL(X ); 2 g is an LD accumulation point of fL(X ;A ); 2 g. Theorems 5.1.5 and 5.1.10 admit similar versions. We next concentrate on consequences of Theorem 5.2.15 as the most useful one for applications. Since considerations below are along the lines of the content of Section 4.3 and the proofs use similar ideas, we omit details. As in Section 4.3, we begin with integrable versions when one can consider nontruncated characteristics. We introduce the following localised versions of conditions (I1 ) and (I2 ). (I1 )loc

lim lim sup P1=r jxj 1(r jxj > a) t^N (X ) > " = 0; a!1 2 " > 0; t > 0; N 2 N ;

© 2001 by Chapman & Hall/CRC

408

Maxingale problem

lim lim sup P1=r r jxj2 1(r jxj > a) t^N (X ) > " = 0; a!1 2 " > 0; t > 0; N 2 N :

(I2 )loc

Clearly, (I2 )loc implies (I1 )loc . We recall that the modi ed second characteristic without truncation C~ 0 = (C~t0 (x); t 2 R+ ; x 2 C ) of X is speci ed by the equalities

C~t0 (x) = Ct (x)+(x)2 t (x)

Zt

0

(x^s (x))2 ds; 2 Rd :

Lemma 5.3.2.

1. Let the X be special semimartingales. If, in addition, condition (I1 )loc holds, then condition (sup B )loc is equivalent to the condition

(sup B 0 )loc

sup jB 0 t^N (X ) tT

B0

t^N (X )

1=r P (X )j !

as 2 ; T > 0; N

0

2 N:

2. Let the X be locally square-integrable semimartingales. If, in addition, condition (I2 )loc holds, then condition (C~ )loc is equivalent to the condition

(C~ 0 )loc

kr C~t0^N (X )

C~t0^N (X ) (X )k

1=r P

! 0 as 2 ; t 2 U; N 2 N :

The proof is similar to the proof of Theorem 4.3.2. As in Section 4.3, in view of the lemma Remark 4.3.3 applies to the setting of this section as well. We now introduce simpli ed versions of the other conditions. For (^ )loc and ( )loc , we consider the conditions (QC )loc

(MD)loc

1=r

X P 1 2 (fsg; fr jxj > g) ! 0 as 2 ; r 0<st^ (X ) N t > 0; > 0; N 2 N ;

1=r

P 1 ([0; t ^ N (X )]; fr jxj > g) ! 0 as 2 ; r t > 0; > 0; N 2 N :

© 2001 by Chapman & Hall/CRC

LD convergence results

409

Obviously, condition (QC )loc implies condition (^ )loc with ^( ; x) = 0. Condition (MD)loc , which is stronger than (QC )loc, implies both (^ )loc and ( )loc with ( ; x) = ^( ; x) = 0 . It thus de nes the case of moderate deviations considered in more detail below for the Markov setting. Then by Theorem 5.2.15 we have the following generalisation of Corollary 4.3.4. Theorem 5.3.3. Let the limiter h(x) be continuous, and B , C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let the linear-growth conditions (5.2.10), (5.2.11) and (5.2.12) hold. If conditions (0); (A)loc + (a)loc ; (sup B )loc; (C )loc (respectively, (C~ )loc ), ( )loc , and (QC )loc hold, then the net fL(X ); 2 g is C { exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant 1 Gt (; x) = Bt0 (x)+ Ct (x)+(ex 1 x)t (x): 2 If the latter problem has the unique solution x0 (e.g., either Theold rem 2.8.33 or Theorem 2.8.34 applies), then L(X ) ! x0 as 2 : By Remark 4.3.3 the theorem also has locally integrable and locally square-integrable versions. R t The next result follows by Theorem 5.2.15 with Bt (x) d= 0 us (x) ds, Ct (x) = 0, and s ( ; x) = us (x)1(1 2 ); 2 B (R ), and Theorem 2.8.10. It extends Corollary 4.3.5. Theorem 5.3.4. Let d = 1 and conditions (0), (A)loc + (a)loc and (QC )loc hold. Let the limiter h(x) be continuous at x = 1. Let (us (x); x 2 D ; s 2 R+ ) be a D-progressively measurable R+ -valued function, which is C {continuous in xR and satis es the linear-growth condition us (x) (1 + xs )ls , where 0t ls ds < 1; t 2 R+ . If t^ZN (X ) 1=r P sup jBt^N (X ) h(1) us(X )dsj ! 0 as 2 ; tT 0 T > 0; N 2 N ; 1=r ;Æ lim lim sup P (r kCt^N (X ) k > ") = 0; Æ!0 2 t 2 U; " > 0; N 2 N ;

© 2001 by Chapman & Hall/CRC

410

Maxingale problem

and, for all " 2 (0; 1=2); t 2 R+ and N

2 N ; as 2 ,

1 ([0; t ^ N (X )]; fjr x 1j < "g) r 1 ([0; t ^ N (X )]; fjr xj > "g r

t^ZN (X )

us

1=r P (X )ds !

0 \

0;

1=r P

fjr x 1j > "g) ! 0;

then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumuR lant Gt (; x) = (e 1) 0t us (x) ds: If the latter problem has a unique solution (e.g., by Theorem 2.8.28 inf inf us(x) > 0 and st x2K sup sup us (x) < 1 for every compact K C and t 2 R+ ), then st x2K ld X ! X as 2 ; where X is the Luzin-continuous idempotent Poisson process of rate us(X ) starting at x0 with idempotent distribution x0 , whose density is given by Z1 x0 (x) = exp sup x_ t (e 1)ut (x) dt 2R 0

if x is absolutely continuous and x0 = x0 , and x0 (x) = 0 otherwise.

Let us introduce the following conditions. (sup B00 )loc

sup jBt^N (X ) tT

Bt0^N (X ) (X )j

1=r P

! 0 as 2 ; T > 0; N 2 N ;

(C0 )loc

kr C~t^N (X )

1=r P Ct^N (X ) (X )k !

(C 0 )loc

kr C~t0^N (X )

1=r P Ct^N (X ) (X )k !

0

(L2 )loc

r j

1=r P (r jxj > ) t^N (X ) !

j1

x2

0 as 2 ; t 2 U; N 2 N ;

© 2001 by Chapman & Hall/CRC

0 as 2 ; t 2 U; N 2 N ;

0 as 2 ; t > 0; N 2 N ; > 0:

411

LD convergence results

Note that the latter condition is a localised version of the Lindeberg condition. As above, it implies both (I2 )loc and (MD)loc and allows us to do without truncation. The following result extends Corollaries 4.3.7 and 4.3.8. The proof is similar and also uses Theorem 2.8.9.

Theorem 5.3.5. Let the functions (bs(x); s 2 R+ ; x 2 D ) and (cs (x); s 2 R+ ; x 2 D ) be C {continuous and satisfy the linear-growth conditions

jbt (x)j lt(1+ xt ); kct (x)k lt (1+ xt 2); R

where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ : Let conditions (0) and (A)loc + (a)loc hold. If, in addition, either conditions (MD)loc , (sup B00 )loc and (C0 )loc hold for some limiter h(x) or conditions (L2 )loc , (sup B 0 )loc , and (C00 )loc hold, then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant 1 Gt (; x) = Bt0 (x)+ Ct (x): 2 If, in addition, uniqueness holds for problem (x0 ; G) (e.g., according to Theorem 2.8.21, inf inf inf cs (x) > 0 and sup sup kcs (x)k < jj=1 st x2K st x2K ld 1 for every t 2 R+ and compact K C ), then X ! X as 2 , where X = (Xt ; t 2 R+ ) is the Luzin-continuous idempotent diusion

X_ t = bt (X )+ c1t =2 (X )W_ t ; X0 = x0 ; whose deviability distribution is given by 1 Z1 x0 (x) = exp 2 (x_ t bt (x))ct (x) (x_ t bt (x)) dt 0

if x0 = x0 , x is absolutely continuous and x_ t of ct (x) a.e., and x0 (x) = 0 otherwise.

bt (x) is in the range

Now, we turn our attention to conditions (A)loc + (a)loc . Let us introduce the condition (V S0 )loc

([0; t ^

N

© 2001 by Chapman & Hall/CRC

(X )]; fr

jxj > g

1=r P 1 =r )

! 0 as 2 ; t > 0; > 0; N 2 N :

412

Maxingale problem

Condition (V S0 )loc implies both (A)loc + (a)loc and (MD)loc so that by Lemma 5.3.2 and Theorem 5.3.5 we have the following result.

Theorem 5.3.6. Let the X be locally square-integrable.

Then the assertion of Theorem 5.3.5 holds if instead of conditions (0), (A)loc + (a)loc , (MD)loc , (sup B00 )loc , and (C0 )loc one requires conditions (0), (I2 )loc , (V S0 )loc , (sup B 0 )loc , and (C00 )loc .

Conditions (A)loc + (a)loc are also implied by the following condition (V S )loc , which is weaker than (V S0 )loc : (V S )loc

lim lim sup P1=r ([0; t ^ N (X )]; fr jxj > ag)1=r > " a!1 2 = 0; t > 0; " > 0; N 2 N :

For the sequel, we note that conditions (A)loc + (a)loc are implied by the conditions (A0 )loc (a0 )loc

P

1=r

([0; t ^ N (X )]; jxj > A)1=r ! 0 as 2 ; t > 0; N 2 N ; 9A > 0; 1=r

P 1 r jxj e 1(r jxj > a) 1(jxj A) t^N (X ) ! 0 r as 2 ; t > 0; > 0; A > 0; N 2 N ; 9a > 0:

If, in addition, the convergence in (a0 )loc holds for every a > 0, then (MD)loc holds. Let us assume that the Cramer condition (Cr) holds, i.e., j x e j 1(jxj > 1) t < 1; t > 0; > 0: Then moment conditions can be used to check (A)loc + (a)loc . More speci cally, let us introduce the conditions 1=r 1 r jxj e 1 ( r j x j > a ) > " (Ie )loc alim lim sup P t ^ ( X ) N !1 2 r = 0; t > 0; " > 0; > 0; N 2 N ; (Le )loc

1=r

P 1 r jxj e 1 (r jxj > ) t^N (X ) ! 0 as 2 ; r t > 0; " > 0; N 2 N :

© 2001 by Chapman & Hall/CRC

413

LD convergence results

Note that (Le )loc is an exponential analogue of the Lindeberg condition. Then (Le )loc ) (Ie )loc ) (A)loc + (a)loc ; (Ie )loc ) (I2 )loc and (Le )loc ) (L2 )loc ) (MD)loc . In particular, we can check (A)loc +(a)loc by checking (Ie )loc (e.g., in Theorem 5.3.3). As another illustration, Theorem 5.3.4 allows us to state the following extension of part b) of Corollary 4.3.12.

Theorem 5.3.7.

Let Xt = Nt =r , where N = (Nt ; t 2 R+ ) are one-dimensional point processes with respective compensators A = (At ; t 2 R+ ). Let (us (x); x 2 D ; s 2 R+ ) be a Dprogressively measurable R+ -valued function, which is C {continuous in x and R t satis es the linear-growth condition us (x) (1 + xs )ls , where 0 ls ds < 1; t 2 R+ . Let the maxingale problem (0; G) with R cumulant Gt (; x) = (e 1) 0t us (x) ds has the unique solution 0 (e.g., by Theorem 2.8.28 inf inf us (x) > 0 and sup sup us (x) < 1 st x2K st x2K for every compact K C and t 2 R+ ). Let X be the Poisson idempotent process of rate us(X ) with idempotent distribution 0 . If, as 2 ,

1 A r t^N (X )

t^ZN (X )

us

1=r P (X )ds !

0

0

and 1=r

X P 1 (As )2 ! 0; r 0<st^ (X ) N

ld then X ! X as 2 :

Since (Le )loc implies (A)loc + (a)loc , (I2 )loc and (MD)loc , Lemma 5.3.2 and Theorem 5.3.5 result in the following version of Theorem 5.3.6.

Theorem 5.3.8.

Let the Cramer condition hold. Then in Theorem 5.3.5 one can replace conditions (A)loc + (a)loc , (MD)loc , (sup B00 )loc , and (C0 )loc with conditions (Le )loc , (sup B 0 )loc , and (C00 )loc .

© 2001 by Chapman & Hall/CRC

414

Maxingale problem

5.4 Large deviation convergence of Markov processes We now consider implications of the above results for the Markov setting. In the next theorem we assume that X are generally speaking non-time homogeneous \continuous-time Markov processes with generators At " in the sense that the At map the functions (ex ; x 2 R d ); 2 R d ; into B(R + ) B (R d )=B (R )-measurable functions of (t; x) R and the processes (exp( Xt ) exp( X0 ) 0t As exp( Xs ) ds; t 2 R + ) are well-de ned local martingales on ( ; F ; F ; P ). Let S denote the state space of X .

Theorem 5.4.1. d

Let gt (; x); t 2 R+ ; 2 Rd ; x 2 Rd , be a B(R+ ) B(R ) B(Rd )=B(R)-measurable R-valued function such that gt (0; x) = 0 . Let us assume that gt (; x) is continuous in x and meets the linear-growth condition gt (; x) kt (jj(1 + jxj)) , where t 2 R+ , 2 Rd , and the function kRt () is R+ -valued, Lebesgue measurable in t, increasing in , and 0t ks () ds < 1; t 2 R+ ; 2 R+ . If, as 2 ,

1=r P X0 ! x0

and, for T > 0; N

2 N,

1 sup sup j exp( r x)At exp(r x) gt (; x)j ! 0; tT x2S :jxjN r then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves maxingale problem (x0 ; G) associated with the cumulant

Gt (; x) =

Zt

0

gs (; xs ) ds: R

Proof. Since the processes (exp(r (Xt X0 )) exp( 0t exp( r Xs )As exp(r Xs )ds); t 2 R+ ) are local martingales on ( ; F ; F ; P ) , Liptser and Shiryaev [79, Theorem 2.5.1] (see also Ethier and Kurtz [48, Lemma 3.2]), the claim follows by Theorem 5.1.13.

We next consider applications of Theorem 5.2.15. The semimartingales X are assumed to be Markov processes. To simplify

© 2001 by Chapman & Hall/CRC

415

LD convergence of Markov processes

the conditions, we distinguish between the continuous- and discretetime cases. Let f ; 2 g and f ; 2 g be nets of real numbers tending to 1 as 2 . In the continuous-time case, we assume that the predictable triplets of X corresponding to a limiter h(x) are given by:

Bt

=

Ct =

1 r

([0; t]; dx) =

Zt

0

Zt

bs (Xs )ds;

(5.4.1)

cs (Xs )ds;

(5.4.2)

s ( dx; Xs )ds;

(5.4.3)

0 Zt 0

where bs (u) is an R d {valued B (R + ) B (R d )=B (R d ){measurable funcRt tion such that 0 jbs (xs )j ds < 1 for t 2 R+ and x 2 D , cs (u) is a function with values in the space of symmetric positive semi-de nite d d {matrices, which is B(R+ ) B(Rd )=B(R dd ){measurable and Rt such that 0 kcs (xs )k ds < 1 for t 2 R+ and x 2 D , s (dx; u) is a transition kernel from (R+ Rd ; B(R+ ) B(Rd )) into (Rd ; B(Rd )) such that

t (f0g; u) Zt Z

0

Rd

= 0;

Z

Rd

1 ^ jxj2 t (dx; u) < 1;

1 ^ jxj2 s (dx; xs ) ds < 1; t 2 R+ ; x 2 D ; u 2 Rd :

In the discrete-time case, we assume that X is a pure jump process with predictable measure of jumps bX tc ( dx; X ([0; t]; dx) = ^i= (5.4.4) (i 1)= ); i=1 where ^i= (dx; u), for every i 2 N ; (Rd ; B(Rd )) into (Rd ; B(Rd )) such that

is a transition kernel from

(f0g; u) = 0; ^ (R d ; u) 1; i 2 N ; u 2 R d : ^i= i=

© 2001 by Chapman & Hall/CRC

416

Maxingale problem

The parameter can be interpreted as the frequency of jumps of X and the parameter 1= as the size of the jumps. Depending on the relative speed at which and go to 1, there are two dierent asymptotics, which are referred to as \very large deviations", when and are of the same order and one takes r = = , and \moderate deviations", when = ! 1 but = 2 ! 0; and one takes r = 2 = . Let us recall the de nition of the essential supremum of a collection of measurable functions, see, e.g., Neveu [94, II.4]. Let fj = (fj (x); x 2 R+ ); j 2 J; be a collection of B(R+ )=B(R){measurable R -valued functions on R + ; for a B(R + )=B (R ){measurable R -valued function f = (f (x); x 2 R+ ), we say that f = ess sup j 2J fj if, for every j 2 J , f (x) fj (x) for almost all x 2 R+ and f (x) g(x) for almost all x 2 R+ for every B(R+ )=B(R){measurable R-valued function g = (g(x); x 2 R+ ) such that, for every j 2 J , g(x) fj (x) for almost all x 2 R+ . Note that this usage is dierent from the interpretation above. We consider, rst, the case of very large deviations: = = r . Let us introduce the following versions of the Cramer condition: ess sup jujv Zt

0

Z

(ex

Rd

ess sup jujv

Z

1 x) t (dx; u) < 1;

(ex

Rd

1 x) s (dx; u) ds < 1;

2 Rd ; t 2 R+ ; v 2 R+ ; (5.4.5)

in the continuous-time case, and Z

Rd

(dx; u) < 1; 2 R d ; i 2 N ; u 2 R d ; ex ^i=r

(5.4.6)

in the discrete-time case. Under these conditions, the X are locally square integrable semimartingales, so, we may and will take h(x) = x so that B is the rst characteristic \without truncation" (B = B 0 ), and de ne Z 1 g s (; u) = bs (u)+ cs (u)+ (ex 1 x) s (dx; u) 2 Rd

© 2001 by Chapman & Hall/CRC

417

LD convergence of Markov processes

in the continuous-time case and Z g s (; u) = ln 1+ (ex 1) ^(br sc+1)=r (dx; u) Rd

in the discrete-time case. Let gs (; u) be a B(R+ ) B(Rd ) B(Rd )=B(R)-measurable function, which is continuous in u and satis es the following lineargrowth condition

jgs(; u)j g~s(jj(1+ juj)); whereR g~s (y) is R+ -valued, B(R+ )=B(R)-measurable in s, increasing in y, 0t g~s (y) ds < 1; t 2 R+ ; y 2 R+ ; and gs (0; u) = 0 . We then have the following version of Theorem 5.4.1. Theorem 5.4.2. Let = = r and the Cramer condition (5.4.5) in the continuous-time case, respectively, the Cramer con1=r P X0 !

dition (5.4.6) in the discrete-time case, hold. Let 2 . If, for all 2 Rd , t 2 R+ and v 2 R+ , as 2 , Zt

0

x0 as

ess sup jg s (; u) g s(; u)j ds ! 0; jujv

then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant

Gt (; x) =

Zt

0

g s(; xs ) ds:

If the latter problem has the unique solution x0 (e.g., the conditions ld of Theorem 2.8.32 hold for gs (; x) = g s (; xs )), then L(X ) ! x0 as 2 : Proof. We have the following representations for the stochastic exponential E () = (Et (); t 2 R+ ); 2 Rd ; associated with X . In the continuous-time case,

1 ln Et (r ) = r

Zt

0

© 2001 by Chapman & Hall/CRC

gs (; Xs ) ds:

418

Maxingale problem

In the discrete-time case, by the equalities g s = g br sc=r and Xs = Xbr sc=r , 1 1 ln Et (r ) = r r

brX tc 1 i=0

Therefore, in both cases for N 1 r

ln Et^N (X ) (r ) Zt ess sup jgs(; u) jujN 0

)= g i=r (; Xi=r

2N

brZtc=r 0

gs (; Xs ) ds:

Gt^N (X ) (; X )

g s (; u)j ds +

Zt

br tc=r

sup jgs (; u)j ds; jujN

and the claim follows by Theorem 5.1.12. We next state a \very large deviation" result in terms of characteristics, which does not require the Cramer condition. We con ne ourselves to the continuous-time case. We assume all the above conditions on bs (u), cs (u) and s(dx; u) to hold except the Cramer condition (5.4.5). Instead, we assume that the rst characteristic B corresponds to a continuous limiter h(x). We de ne positive semide nite symmetric matrices c~s (u) by

c~s (u)

=

cs (u) +

Z Rd

( h (x))2 s (dx; u); 2 Rd :

We next introduce the limit idempotent process. Let bs(u) be an Rd {valued B(R+ ) B(Rd )=B(R d ){measurable function, cs (u) be a B(R+ ) B(Rd )=B(R dd ){measurable function with values in the space of positive semi-de nite symmetric d d {matrices and s (dx; u) be a transition kernel from (R+ Rd ; B(R+ ) B(Rd )) into

© 2001 by Chapman & Hall/CRC

419

LD convergence of Markov processes

(Rd ; B(Rd )) such that

s (f0g; u) = 0;

Z

(ex

1 x) s (dx; u) < 1;

Rd

2 Rd ; s 2 R+ ; u 2 Rd :

Let the following linear-growth conditions be satis ed:

jbs(u)j ls(1+juj); kcs(u)k ls(1+juj2 ); Z

(ejxj

Rd

(5.4.7)

1 jxj)s (dx; u) Z

(ejxj(1+juj)

1 jxj(1 + juj))ms (dx); 2 R+ ;

Rd

whereR ls is an R+ -valued B(R+ )=B(R+ ){measurable function such that 0t ls ds < 1 and ms(dxR) isR a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that 0t Rd (exp(jxj) 1 jxj)ms (dx)ds < 1; t 2 R+ ; 2 R+ . In addition, we assume that the functions

u ! bs (u); u ! cs (u); u !

Z

f (x)s (dx; u);

Rd

for f continuous and such that jf (x)j 1 ^ jxj2 ; are continuous in u 2 Rd . We also de ne positive semi-de nite symmetric matrices c~s (u) by

c~s (u) = cs (u) +

Z

( h(x))2 s (dx; u); 2 Rd :

Rd

Theorem 5.4.3. Let = = r and the above conditions hold. © 2001 by Chapman & Hall/CRC

420 Let Zt

0

Maxingale problem

1=r P X0 ! x0

ess sup jujv

and, for all t 2 R+ and v 2 R+ ,

jbs(u)

bs (u)j ds ! 0;

Zt

0

Zt

Z ess sup jujv d 0 R

f (x)s (dx; u)

ess sup kc~s (u) c~s (u)k ds ! 0; jujv

Z

f (x)s(dx; u) ds

! 0; f 2 Cb:

Rd

Let also Z t

lim lim sup a!1 2

0

ess sup s (fjxj > ag; u) ds jujv

1=r

= 0;

v 2 R+ ; t 2 R+ : (5.4.8)

Then the net fL(X ); 2 g is C {exponentially tight. If is an LD accumulation point of fL(X ); 2 g , then the canonical idempotent process X is a Luzin-continuous semimaxingale with local characteristics (b; c; ; 0) on (C ; C; ). If the idempotent distribution Li(X ) of X is speci ed uniquely (e.g., Theorem 2.8.34 applies), then ld Li(X ) = x0 and X ! X as 2 : If, in addition, the Cramer condition (5.4.5) holds, then condition (5.4.8) can be replaced with the condition

lim lim sup a!1 2

Zt

Z

0

Rd

ess sup jujv

ejxj 1(jxj > a)s (dx; u) ds = 0; t 2 R+ ; 2 R+ ; v 2 R+ : (5.4.9)

In the latter case we can take h(x) = x, i.e., consider nontruncated characteristics. Proof. Condition (5.4.8) implies condition (V S )loc , and condition (5.4.9) implies condition (Ie )loc . Either one of these conditions implies conditions (a)loc + (A)loc . Therefore, the claim follows by Theorem 5.2.15. We can consider nontruncated characteristics under (5.4.9) by Lemma 5.3.2 and the fact that (Ie )loc implies (I2 )loc .

© 2001 by Chapman & Hall/CRC

LD convergence of Markov processes

421

We consider now moderate deviations so ! 1 and ! 1 in such a way that = ! 1 and = 2 ! 0. Let r = 2 = . We assume the locally square-integrable case, i.e., ess sup jujv Zt

0

Z

jxj2 t(dx; u) < 1;

Rd

ess sup jujv

Z Rd

jxj2 s(dx; u) ds < 1; t 2 R+ ; v 2 R+ ; 2 ;

in the continuous-time case, and Z

(dx; u) < 1; i 2 N ; u 2 R d ; 2 ; jxj2^i=

Rd

in the discrete-time case. In the discrete-time case we also assume that the X are martingales, i.e., Z

(dx; u) = 0; i 2 N ; u 2 R d ; 2 : x^i=

Rd

Then the X are locally square integrable semimartingales, so we choose nontruncated predictable characteristics. According to (5.4.2), (5.4.3), (5.4.4), and the equality r = 2 , the (nontruncated) modi ed predictable second characteristics of the X are of the form: in the continuous-time case

C~ 0 = t

2

Zt

0

c~0s (Xs ) ds;

where c~0s (u) are positive semi-de nite symmetric matrices de ned by

c~0s (u) = cs (u) +

Z

Rd

© 2001 by Chapman & Hall/CRC

( x)2 s (dx; u);

422

Maxingale problem

in the discrete-time case bZtc= 0 C~t = 2 c~0s (Xs ) ds; 0

where c~0s (u) are positive semi-de nite symmetric matrices de ned by Z 0 c~s (u) = ( x)2 ^(b sc+1)= (dx; u): Rd

Let us assume that bs(u) and cs (u) satisfy the conditions stated before Theorem 5.4.3 (i.e., measurability, linear growth and continuity in u). We introduce the following conditions on the predictable measures of jumps and rates of convergence. (P) For some Æ > 0 lim sup

Zt

0

ess sup jujv

Z

jxj2+Æ s(dx; u) ds < 1; t 2 R+ ; v 2 R+ ;

Rd

in the continuous-time case, respectively, b tc Z 1 X (dx; u) < 1; t 2 R ; v 2 R ; sup jxj2+Æ ^i= lim sup + + i=1 jujv d R

in the discrete-time case, and 2 =( ln ) ! 0 as 2 . (SE) For some 2 (0; 1] and > 0 lim sup

Zt

0

ess sup jujv

Z

Rd

exp( jxj )s (dx; u) ds < 1;

t 2 R+ ; v 2 R + ;

in the continuous-time case, respectively, b tc Z 1 X (dx; u) < 1; lim sup sup exp( jxj )^i= i=1 jujv d R

© 2001 by Chapman & Hall/CRC

t 2 R+ ; v 2 R + ;

423

LD convergence of Markov processes

in the discrete-time case, and 2 = ! 0 as 2 . The next theorem extends Theorem 4.4.8.

Theorem 5.4.4. Let = ! 1 and = 2 ! 0 as 2 .

Let either condition (P) or condition (SE) hold. Let the law of a Luzincontinuous semimaxingale X with local characteristics (b; c; 0; 0) starting at x0 be speci ed uniquely (e.g., according to Theorem 2.8.21, inf inf inf cs (u) > 0 and sup inf kcs (u)k < 1; t 2 R+ ; v 2 jj=1 st jujv st jujv R + ). If, as 2 , and v 2 R+ Zt

0

1=r P X0 ! x0 ,

where r = 2 = , and for all t 2 R+

ess sup jbs (u) bs (u)j ds ! 0; jujv

Zt

0

ess sup kc~0s (u) cs (u)k ds ! 0; jujv

ld then X ! X as 2 at rate r . Proof. The proof is almost the same as for Theorem 4.4.8. In some more detail, either one of conditions (P) or (SE) implies (L2 )loc . Since by hypotheses conditions (sup B 0)loc and (C00 )loc hold, according to Theorem 5.3.5 one needs to check conditions (A)loc + (a)loc . If condition (P) is satis ed, then condition (V S0 )loc holds, which implies (A)loc + (a)loc . If condition (SE) is satis ed, then conditions (A0 )loc + (a0 )loc can be veri ed as in the proof of Theorem 4.4.8.

Remark 5.4.5. We recall that by Theorem 2.8.9 under the hypothe-

ses X is a Luzin-continuous idempotent process satisfying the equation

X_ t = bt (Xt )+ t (Xt )W_ t ; X0 = x0 ; and the deviability distribution of X has density given by 1 Z1 X (x) = exp (x_ s bs (xs ))cs (xs ) (x_ s bs (xs )) ds 2 0

if x is absolutely continuous, x0 = x0 and x_ s bs(xs ) is in the range of cs (xs ) a.e., and X (x) = 0 otherwise.

© 2001 by Chapman & Hall/CRC

424

Maxingale problem

We conclude the section with some illustrative examples. To simplify notation, we consider one-dimensional settings.

Example 5.4.6.

Let R-valued processes X ";Æ; = (Xt";Æ; ; t 2 R+ ), indexed by " > 0; Æ > 0 and > 0, be de ned on respective stochastic bases ( ";Æ; ; F";Æ; ; F";Æ; ; P";Æ; ) and satisfy the equations

Xt";Æ;

= x0 + +Æ

Zt

p bs(Xs";Æ; ) ds + "

0 Zt Z

0 G

fs(Xs";Æ; ; y)

Zt

0

s (Xs";Æ; ) dWs"

N (ds; dy) 1 ds m(dy) ;

where (G; G ) is a measurable space, bs(u); s (u) and fs (u; y) are respective B(R+ ) B(R)=B(R ), B(R+ ) B(R)=B(R ) and B(R+ )

B(R) G =B(R)-measurable functions, W " = (Ws"; s 2 R+ ) are Rvalued Wiener processes, m(dy) is a non-negative - nite measure on (G; G ), and N = (N (ds; dy)) are Poisson random measures on R + G with intensity measures 1 ds m(dy ). We also assume that bs (u) and s(u) are continuous in u, lim

u!v

Z

jfs(u; y) fs (v; y)j2 m(dy) = 0;

G

the following linear-growth conditions are met:

bs (u)2 +s(u)2 ls (1+u2 ); jfs(u; y)j hs (y)(1+juj); where ls and hs (y) are R+ -valued andR increasing in s, hs (y) is G =B(R+ ){measurable for every s 2 R+ , G hs(y)2 m(dy) < 1, and the non-degeneracy condition holds: inf inf (s

st jujv

(u)2 +

Z

fs(u; y)2 m(dy)) > 0; t 2 R+ ; v 2 R+ :

G

For existence of X ";Æ; see, e.g., Gihman and Skorohod [54, Chapter 5]. Let us consider the following moment conditions on the jumps of the X ";Æ; :

© 2001 by Chapman & Hall/CRC

425

LD convergence of Markov processes

~ for some Æ > 0 (P) Zt

0

ess sup jujv

Z

jfs(u; y)j2+Æ m(dy) ds < 1; t 2 R+ ; v 2 R+ ;

G

f for some 2 (0; 1] and > 0 (SE) Zt

0

ess sup jujv

Z

exp jfs (u; y)j m(dy) ds < 1; t 2 R+ ; v 2 R+ :

G

Let an idempotent Luzin-continuous process X satisfy the equation Z

X_ t = bt (Xt )+ t (Xt )2 + ft (Xt ; y)2 m(dy) 1=2 W_ t ; X0 = x0 ;

G

where W is an R-valued idempotent Wiener process. The process X is well de ned by Theorems 2.6.24 and 2.8.21.

Theorem 5.4.7. Let ! 0, ! 20, and Æ ! 0 in such a way that 2 1 1

= Æ = . If, in addition, either Æ ln( ) ! 1 and condition ld g) holds, then X ";Æ; ! (P~ ) holds, or Æ2 1 ! 1 and condition (SE X at rate 1=. Proof. The predictable characteristics without truncation of X ";Æ; are of the form

B 0";Æ; = t

Zt

0

bs(Xs";Æ; ) ds;

g(x) t";Æ; =

1

Ct";Æ;

Zt Z

0 G

="

Zt

0

s(Xs";Æ; )2 ds;

g Æfs (Xs";Æ; ; y) m(dy) ds for g Borel and bounded:

It is straightforward to see that the convergence hypotheses of Theorem 5.4.4 hold for = (; Æ; ), R = 1 , = Æ 1 , b;Æ; s (u) = bs (u), ;Æ; 2 2 and c~s (u) = cs (u) = s (u) + G fs(u; y) m(dy): The moment conditions on the jumps are the same as in Theorem 5.4.4.

© 2001 by Chapman & Hall/CRC

426

Maxingale problem

Let us assume, in addition, that the function bs (u) is dierentiable in u (for almost all s) and the derivative b0s (u) is bounded on bounded domains. We denote as (xt ; t 2 R+ ) the solution of the equation

xt = x0 +

Zt

bs (xs ) ds

0

(for existence and uniqueness of (xt ) see, e.g., Coddington and Levinson [26, Chapter II]). We introduce the processes X~ ; = (X~t; ; t 2 R+ ) by

X~ t; =

r

;; Xt xt ;

where > 0 and de ne \a non-time-homogeneous idempotent Ornstein-Uhlenbeck process" X~ = (X~t ; t 2 R+ ) by Z

~_ t ; X~ 0 = 0; X~_ t = b0t (xt )X~ t + t (xt )2 + ft (xt ; y)2 m(dy) 1=2 W

G

~ is an R-valued idempotent Wiener process. By Theowhere W rem 2.6.26 the latter equation has a unique Luzin strong solution with idempotent distribution speci ed by the density

x) = exp

X~ (

1 2

Z1

0

x_ t R

b0t (xt )xt 2 dt t (xt )2 + G ft (xt ; y)2 m(dy)

if x0 = 0 and x is absolutely continuous, and X~ (x) = 0 otherwise.

Theorem 5.4.8. Let ! 0 and ! 0 in such a way that1 = ! 1. If, in addition, either condition (P~ ) holds and ln( ) ! 1, or ld ~ g) holds and 2 ! 1, then X ~ ; ! condition (SE X at rate 1=. Proof. We again invoke Theorem 5.4.4. Since X~ ; satis es the equa-

© 2001 by Chapman & Hall/CRC

427

LD convergence of Markov processes

tion

X~ t;

=

Z t r r

bs

0 Zt

X~ s; + xs

bs (xs ) ds

r p + s X~ s; + xs dWs

p +

0 Zt Z

fs

r

0 G

X~ s; + xs; y

N (ds; dy) 1ds m(dy) ;

it follows that the predictable characteristics of X~ ; without truncation are of the form

B 0 ; = t

Zt r r

0

bs

Ct; g(x) t;

=

1

Zt Z

=

X~ s; + xs

Zt

0

s

r

bs (xs ) ds;

2 X~ s; + xs ds;

r p g fs X~ s; + xs; y m(dy) ds;

0 G

for g Borel and bounded: Therefore, letting = (; ), = 1 and = () 1=2 , in the notation of Theorem 5.4.4 r r bs (u) = bs u + xs bs (xs ) ; 2 r c~0s (u) = s u + xs Z r 2 + fs u + xs ; y m(dy) G

so that we have the convergences Zt

0

ess sup jujv

jbs (u)

bs(u)j ds ! 0;

© 2001 by Chapman & Hall/CRC

Zt

0

ess sup jc~0s (u) cs j ds ! 0; jujv

428

Maxingale problem

where

Z

bs (u) = b0s (xs )u; cs = s (xs )2 + fs(xs ; y)2 m(dy): G

Now the claimed LD convergence follows by Theorem 5.4.4.

Example 5.4.9.

Let R-valued processes X n = (Xtn ; t 2 R+ ), where n 2 N , be de ned on respective stochastic bases ( n ; Fn ; Fn = (Ftn ; t 2 R+ ); Pn ) and have the form 1 Xtn = N n n n

Zt

0

f Xsn ; Ynsn ds ;

where f (x; y) is an R+ -valued Borel function, N n = (Ntn ; t 2 R+ ) are Poisson processes on ( n ; Fn ; Fn ; Pn ), and Y n = (Ytn ; t 2 R+ ) are n ; t 2 R ); P ): Ornstein-Uhlenbeck processes on ( n ; Fn ; Gn = (Ft=n + n

Ytn

=

Zt

0

1 Ysn ds + p Wtn; n

W n = (Wtn ; t 2 R+ ) being Wiener processes on ( n ; Fn ; Gn ; Pn ). The processes X n are well de ned since the N n are piecewise constant. We assume that f (x; y) is continuous at points (x; 0) for x 2 R+ and is such that sup0xa;y2R f (x; y) < 1 for a > 0, f (x; 0) > 0 for x 2 R+ , and the function f (x; 0) grows at most linearly as x ! 1. We prove that the X n LD converge at rate n to the Luzincontinuous idempotent process X satisfying the equation Xt = N

Z t

f (Xs ; 0) ds ;

0

where N is a Poisson idempotent process, and having the idempotent distribution with density Z1 X (x) = exp sup x_ t (e 1)f (xt ; 0) dt 2R 0

© 2001 by Chapman & Hall/CRC

429

LD convergence of Markov processes

if x is absolutely continuous, increasing and x0 = 0, and X (x) = 0 otherwise. The idempotent process X is well de ned by Theorem 2.6.33. Let us denote tn = Yntn . For tn we have the equation

tn

= n

Zt

0

^ tn; snds + W

(5.4.10)

^ n = (W ^ tn; t 2 R+ ) is a Wiener process on ( n ; Fn ; Fn ; Pn ). where W One that the nX n have Fn -compensators Ant = R t cann show n n 0 f (Xs ; s ) ds so that by Theorems 2.8.10, 2.8.28, and 5.3.7 the claim would follow by t^Z (X n ) N sup f (Xsn ; sn)ds tT 0

1=n f (Xsn ; 0)ds Pn

t^ZN (X n ) 0

! 0 as n ! 1; T > 0; N

2 N:

By continuity of f at points (x; 0), where x 2 R+ ; and the boundedness condition supx2[0;a];y2R f (x; y) < 1, for arbitrary " > 0 there exists Æ > 0 such that 8 >

: tT 0

f (Xsn ; 0)ds

t^ZN (X n ) 0 8 "

9 > = > ; 9 =

: 1(jsnj > Æ)ds > 2" ; ; 0

so by \the Chebyshev inequality" it is suÆcient to show that lim P 1=n n!1 n

ZT

0

jsnj2 ^1 ds > = 0; > 0:

(5.4.11)

Let g(x); x 2 R; be a twice dierentiable non-negative function with bounded rst and second derivatives, and such that g(x) = x2 =2; jxj 1; and xg0 (x) 1 if jxj 1 (e.g., g(x) = 1=2 + ln jxj +

© 2001 by Chapman & Hall/CRC

430

Maxingale problem

(ln jxj)2 ; jxj 1). By Ito's formula and (5.4.10)

g(tn ) = g(0)

n

Zt

g0 (n)n ds + s

0

Zt

s

0

1 g0 (sn) dW^ sn + 2

Zt

0

g00 (sn )ds:

Since g0 is bounded, for all 2 R

En exp

ZT

0

2 ^ sn g0 (sn )dW 2

ZT

0

g0 (sn)2 ds = 1

and hence

En exp

g(Tn )

g(0) + n

ZT

g0 (n)n ds

0

s

s

ZT 2

2

0

2

ZT

0

g00 (sn )ds

g0 (sn )2 ds = 1;

which implies, since g is non-negative, g0 and g00 are bounded, g0 (x) = x; jxj 1; and g0 (x)x 1 for jxj 1, that for some function F ()

En exp n

ZT

0

jsnj2 ^ 1 ds F ():

Hence, lim sup Pn1=n n!1

ZT

0

jsnj2 ^ 1 ds >

ZT

exp( ) lim sup En exp(n jsnj2 ^ 1 ds) 1=n n!1

0

Since is arbitrary, (5.4.11) is proved.

© 2001 by Chapman & Hall/CRC

exp( ):

431

LD convergence of Markov processes

Example 5.4.10. This example considers a discrete-time case and builds on Example 4.4.12. Let n ! 1, ! 1 and ! 1 in such a way that n= ! 1, and i ; i = 1; 2; : : :, where = (n; ; ), be i.i.d. indicator random variables, which equal 1 with probability =(n ) and 0 with probability 1 =(n ). We de ne random variables Yk by

Yk = Yk 1 +

b f (YX k 1 =)c

i=1

i ; Y0 = 0;

where f (x) is a continuous positive function, growing no faster than linearly as x ! 1. Let the process X = (Xt ; t 2 R+ ) be de ned by Xt = Ybntc =. Then X is a point process whose compensator A = (At ; t 2 R+ ) relative to the ltration generated by X is given by

At

bntX c 1 )c: = E1 b f (Xi=n i=0

We have At

Zt

0

f (Xs ) ds

bnt Z c=n 0

b f (Xs )c f (X ) ds s +

Zt

bntc=n

f (Xs ) ds

and

1 X 2 As 2 2 bntc sup ( f (x))2 : st^ (X ) n 0xN N

Hence, by Theorems 5.3.7 and 2.8.28 the net fX ; 2 g LD converges at rate to the semimaxingale X with local characteristics (b; 0; ; 0), where bs (x) = f (xs) and ( ; x) = 1(1 2 )f (xs); equiva R lently X is the Luzin solution of the equation Xt = N 0t f (Xs ) ds ;

© 2001 by Chapman & Hall/CRC

432

Maxingale problem

where N is an idempotent Poisson process, whose idempotent distribution has density Z1 X sup x_ t (e 1)f (xt ) dt (x) = exp 2R 0

if x is absolutely continuous, increasing and x0 = 0, and X (x) = 0 otherwise.

Remark 5.4.11.

It is straightforward to extend Examples 5.4.6, 5.4.9 and 5.4.10 to the case where the coeÆcients depend on the past.

© 2001 by Chapman & Hall/CRC

Chapter 6

Large deviation convergence of queueing processes In this chapter we apply the results on large deviation convergence of semimartingales for deriving large deviation asymptotics in queueing systems.

6.1 Moderate deviations in queueing networks In this section we prove LD convergence of queueing processes in single server queues and networks of single server queues to idempotent diusions.

6.1.1 Idempotent diusion approximation for single server queues We consider a sequence of FIFO single server queues indexed by n . For the nth system, we denote by Ant the number of arrivals by time t , by Stn the number of customers served for the rst t units of the server's busy time, by Dtn the number of departures by time t , by Qnt the queue length at time t , by Wtn the un nished work at time t , by Ctn the completed work at time t , by Hkn the waiting time of the k th customer, and by Lnk the departure time of the k th 433 © 2001 by Chapman & Hall/CRC

434

LD convergence for queues

customer. We also introduce

Vkn = minft 2 R+ : Stn kg; k 2 Z+;

(6.1.1)

which, for k 2 N , is the cumulative service time of the rst k customers. All the objects referring to the nth system are assumed to be de ned on a complete probability space ( n ; Fn ; Pn ). Also all the processes are assumed to have trajectories from the associated Skorohod space. The above processes are connected by the following equalities

Ctn =

Zt

0

Wtn = W0n + V n Æ Ant Ctn; Zt

1(Wsn > 0) ds = 1(Qns > 0) ds;

(6.1.2)

Qnt = Qn0 + Ant Dtn; Dtn = S n Æ Ctn;

(6.1.3) (6.1.4)

0

p

Let bn ! 1 and bn = n ! 0 as n ! 1 , and n and n be positive numbers. We de ne the associated normalized and timescaled processes by 1 An = (Ant ; t 2 R+ ); Ant = p (Annt n nt); (6.1.5) bn n 1 n nt); S n = (S nt ; t 2 R+ ); S nt = p (Snt (6.1.6) n bn n 1 (6.1.7) Qn = (Qnt ; t 2 R+ ); Qnt = p Qnnt ; bn n 1 n nt) ; C n = (C nt ; t 2 R+ ); C nt = p (Cnt (6.1.8) bn n 1 V n = (V nt ; t 2 R+ ); V nt = p (Vbnntc n 1 nt); bn n 1 n n n n D = (Dt ; t 2 R+ ); Dt = p (Dnt n nt); bn n 1 W n = (W nt ; t 2 R+ ); W nt = p Wntn ; bn n 1 H n = (H nt ; t 2 R+ ); H nt = p Hbnntc+1 ; bn n 1 Ln = (Lnt ; t 2 R+ ); Lnt = p (Lnbntc+1 n 1 nt): bn n

© 2001 by Chapman & Hall/CRC

435

Moderate deviations for networks

We assume that n ! > 0 and n ! > 0 as n ! 1 , and \the near-heavy-traÆc condition" holds: 1p n(n n) ! c; c 2 R: bn

(6.1.9)

Note that (6.1.9) implies that = . We recall that the one-dimensional Skorohod re ection map x ! R(x) is characterised by the property that z = R(x) is an only R+ valued R 1function such that z = x + y, where y is increasing, y0 = 0 and 0 1(zt > 0) dyt = 0. It is a continuous map from D (R + ; R) to D (R + ; R ) and can explicitly be written as

R(x)t = xt 0infst xs ^ 0; t 2 R+ ;

(6.1.10)

where x 2 D (R + ; R) and x0 2 R+ , see, e.g., Ikeda and Watanabe [66]. Let WA = (WA;t ; t 2 R+ ) and WS = (WS;t ; t 2 R+ ) be independent idempotent Wiener processes on an idempotent probability space ( ; ). Let A and S be real numbers. As above we denote e = (t; t 2 R+ ). In the theorems below LD convergence refers to the 2 Skorohod topology and rate rn = bn .

Theorem 6.1.1. Let (A ; S ) ! (A WA; S WS ) and n

n

ld

ld Then Qn ! Q, where Q = (Qt ; t 2 R+ ) is an continuous idempotent process de ned by

2 n Pn1=bn Q0 ! q0 .

R + -valued

Luzin-

Q = R q0 + A WA S WS + ce :

Proof. Let us denote A = A WA and S = S WS . By (6.1.3), (6.1.4), (6.1.2), (6.1.7), (6.1.5), (6.1.6), and (6.1.8)

Qnt = Qn0 + Ant

S n Æ C 0tn + +

pn

C nt =

© 2001 by Chapman & Hall/CRC

bn

n

Zt

0

pn Zt bn

0

pn bn

(n

n )t

1(Qns = 0) ds;

(6.1.11)

1(Qns = 0) ds;

(6.1.12)

436

LD convergence for queues

where 1 n C 0tn = Cnt =

n

Zt

0

1(Qns > 0) ds:

(6.1.13)

Since Qnt is non-negative and 0t 1(Qns = 0) ds increases only when Qnt = 0 , (6.1.11) allows us to conclude that n

Q =R

R

n

Since by (6.1.12) and (6.1.11)

n C

n

p

n S Æ C 0n + (n

Qn0 + An

bn

n C

n )e :

(6.1.14)

p

n S Æ C 0n + (n n )e Qn;

= Qn0 + An

n

bn

it follows by (6.1.14) that n

p

n S Æ C 0n + (n b

= Qn0 + An

n

R

Qn0 + An

n n

p

n S Æ C 0n + (n

Therefore, by (6.1.10) n jC nt j 2 sup Qn0 + Ans st 1=b2n The convergences Qn0 Pn! q0

n )e bn

p

n S Æ C 0sn + (n n

bn

n)e :

n )s ;

t 2 R+ :

ld and (An ; S n ) ! (A; S ) , the fact that A and S are proper idempotent processes, the inequality C 0tn t , and (6.1.9) imply that 2

1=bn ( jC n j > a) = 0: lim lim sup P n t n a!1 n!1

Hence, by (6.1.12) and the facts that Zt

0

1

(Qns

= 0) ds

pn=b ! 1 and ! > 0 n n

2 Pn1=bn

! 0 as n ! 1; t 2 R+ ; 2

(6.1.15)

1=bn which implies by (6.1.13) that C 0n Pn! e. Then \the time-change theorem" (Lemma 3.2.11) implies by the LD convergence of (An ; S n )

© 2001 by Chapman & Hall/CRC

437

Moderate deviations for networks

to (A; S ) that the sequence f(An ; S n Æ C 0n ); n 2 N g LD converges to 2 n Pn1=bn ld (A; S ) as well. Since Q0 ! q0 , we have that (Qn0 ; An ; S n Æ C 0n ) ! (q0 ; A; S ) by Lemma 3.1.42. By (6.1.14) and continuity of re ection n n n n 0 n p Q is a continuous function of (Q0 ; A ; S Æ C ; ( n=bn )(n n)e). Therefore, the LD convergence of (Qn0 ; An ; S n Æ C 0n ), the near-heavy traÆc condition (6.1.9), and the contraction principle yield the required LD convergence of fQn ; n 2 N g. The idempotent process Q is Luzin-continuous since A and S are Luzin-continuous and R is continuous.

Remark 6.1.2. If we assume, in addition to the hypotheses of The2 1=bn ld orem 6.1.1, that W n0 Pn! q0 =, then (Qn ; Dn ; W n ; C n ; H n ; Ln ) ! (Q; D; W; C; H; L), where

Dt = AWA;t Qt + ct; Wt = Ct =

A WA;t S WS;t + ct Ht = W (t=); Lt =

Qt ; Wt ;

Dt= :

The following lemma gives an explicit expression for the idempotent distribution of Q. Let Q denote the idempotent distribution of Q and I Q (q) = ln Q (q) be the associated rate function.

Lemma 6.1.3. Let A2 + S2 > 0. The rate function I Q is given by q) = 2(2 1+ 2 ) A S

I Q(

Z1

0

1(qt > 0)(q_ t +

c)2 dt

1(c > 0)c2 Z1 1(qt = 0) dt;

2(A2 + S2 )

0

q is a non-negative and absolutely continuous function such that q0 = q0, and IQ(q) = 1 otherwise. if

For a proof, we need the following result.

© 2001 by Chapman & Hall/CRC

438

LD convergence for queues

Lemma 6.1.4. Let z 2 C (R + ; R) be non-negative and x 2 C (R + ; R) be absolutely continuous. Then z = R(x) if and only if z is absolutely continuous and there exists an absolutely continuous function y 2 C (R + ; R )

such that

z_ t = x_ t + y_ t

a.e.

and

y0 = 0; y_ t 2 R+ a.e.; zt y_ t = 0 a.e. Also z_ t = 0 a.e. on the set ft : zt = 0g. Proof. SuÆciency of the condition follows by the de nition of the re ection mapping. Conversely, if y = R(x) x, then yt ys Rt _ j x j du for 0 s t, so y is absolutely continuous. The other s u conditions on y follow from the de nition of re ection. For the nal part, note that a.e. z_ t = limh!0(zt+h zt )=h. The numerator in the latter fraction being non-negative since zt = 0 implies that the fraction is non-negative for h positive and non-positive for h negative. Hence, the limit is zero. Proof of Lemma 6.1.3. Let 2 = A2 + S2 . By Corollary 2.4.11 we may assume that A WA + S WS = Ws , where W is an idempotent Wiener process. By Theorem 6.1.1 and the de nition of the image idempotent measure

Q (q) = supfW (w);

q = R(q0 + w + ce)g:

Therefore, q0 = q0 and q is absolutely continuous Q -a.e. For these q by the de nition of an idempotent Wiener process, Lemma 6.1.4 and Lemma A.2 in Appendix A

q) = y: y0=0inf;y_ t2R+; 21

I Q(

1(q >0)y_ =0; q_ t=t w_ t +ct+y_ t

Z1

0

w_ t2 dt

1 = 2 2

© 2001 by Chapman & Hall/CRC

Z1

0

q_ t _ t 2R+ ; y_ t : yinf

1(qt >0)y_ t =0

c

y_ t 2 dt

439

Moderate deviations for networks

1 = 2 2 +

1 22

1 = 2 2

Z1

0 Z1 0 Z1 0

1(qt > 0) q_ t

c 2 dt

1(qt = 0) y2infR+ q_ t 1(qt > 0) q_ t

c

2

c y 2 dt Z1 1 (c > 0)c2 dt + 1(qt = 0) dt:

22

0

The next lemma formulates the LD convergence conditions in the hypotheses of Theorem 6.1.1 in terms of interarrival and service times. Let Ukn = inf ft 2 R+ : Ant kg ; k 2 Z+; and U nt = Ubnntc p n 1 nt = bn n : Recalling also (6.1.1) we have by Lemma 3.2.13 the following.

Lemma 6.1.5. The LD convergencen (Ann; S n) !ld

(A WA ; S WS ) holds if and only if the sequence f(U ; V ); n 2 N g LD converges to (A 3=2 WA ; S 3=2 WS ). We now specify the results to the case of GI=GI=1 queues, i.e., we assume that the An and S n are renewal processes. Let us denote by uni ; i 2 N ; the time between the i th and (i + 1) th arrivals and by vin ; i 2 N ; the service time of the i th customer in the n th system. By hypothesis the sequences funi ; i 2 N g and fvin ; i 2 N g are independent i.i.d. Theorem 4.4.8 provides us with the following way of checking the convergence requirements of Lemma 6.1.5.

Lemma 6.1.6. Let either one of the following conditions hold: (i) supn En (un1 )2+ < 1; supn En (v1n )2+ < 1 for some > 0, and b2n = ln n ! 0; (ii) supn En exp((un1 ) ) < 1; supn En exp((v1n ) ) < 1 for some > 0; 0 < 1, and b2n =n =2 ! 0. If En un1 ! 1 ; En v1n ! 1 ; Varn un1 ! A2 =3 ; and Varn v1n ld S2 =3 ; then (U n ; V n ) ! (A 3=2 WA; S 3=2 WS ).

© 2001 by Chapman & Hall/CRC

!

440

LD convergence for queues

We now establish LD convergence for stationary waiting times. Let partial-sum processes U 0 n = (U 0 nk; k 2 Z+) and V n = (Vkn ; k 2 Z+) be given by k

k

i=1

i=1

X X U 0 nk = uni ; U 0 n0 = 0; Vkn = vin ; V0n = 0;

(6.1.16)

so that, as above, Vkn , for k 2 N , is the cumulative service time of the rst k customers. The equation for waiting times is

Hkn+1 = H1n +Vkn U 0 nk min (H1n +Vin U 0 ni )^0:

(6.1.17)

1ik

We recall that if En v1n < Eun1 , then the waiting times Hkn converge in distribution as k ! 1 to the proper random variable supk2Z+(Vkn U 0 nk ) (see, e.g.,pBorovkov [15]). We denote the latter by H0n and let H n0 = H0n =(bn n) .

Theorem 6.1.7.

Let either p one of conditions (i) or (ii) of Lemma 6.1.6 hold. Let ( n=bn )(Eun1 Ev1n ) ! c0 > 0, Varn un1 ! U2 ; Varn v1n ! V2 , where U2 + V2 > 0, as n ! 1. Then the sequence fH n0 ; n 2 N g LD converges in distribution to an exponentially distributed R+ -valued idempotent variable with density 0 2 2 (x) = exp 2c x=(U + V ) ; x 2 R+ . Proof. Since H0n is distributed as supk2Z+(Vkn a Borel subset A of R+ , n Pn (H 0

U 0 nk ) , we have, for ! A

1

2 A) Pn b pn sup (Vkn U 0nk) 2 n 0kbntc Pn( sup (Vkn U 0nk) 0): k>bntc

Let

U~ 0n = (U~t0n ; t 2 R+ ); U~t0n =

1

p (U 0 n Eun1 nt); bn n bntc 1 V~ n = (V~tn ; t 2 R+ ); V~tn = p (Vbnntc Ev1n nt): bn n

© 2001 by Chapman & Hall/CRC

441

Moderate deviations for networks

ld Since p by Theorem 4.4.8 (U~ 0n ; V~ n ) ! (U WA ; V WS ) , (En un1 n 0 En v1 ) n=bn ! c by hypotheses, and

1

p sup (V n U 0nk) bn n 0kbntc k = sup 0st

V ns

U~s0n

(En un1

En v1n )

pn bn

s ;

by the contraction principle and the fact that U WA + V WS = W , where 2 = A2 + S2 and W is an idempotent Wiener process, we have that 1

ld p sup (V n U 0nk) ! sup (Ws c0 s): bn n 0kbntc k 0st

Let t denote the idempotent variable on the right-hand side and = sups2R+ (Ws c0 s). It is an easy exercise to check that has idempotent distribution in the statement of the theorem; in particular, it is a Luzin idempotent variable. We show that t converges to as t ! 1 in idempotent distribution. The map w ! sup0st (ws c0 s) from C (R + ; R) to R is continuous, so t is a Luzin idempotent variable as well. Assuming that both and t are de ned on (C (R + ; R); W ), we have that t monotonically converges to zero W -almost everywhere, so by Theorem 1.3.10 the convergence is actually in deviability W and by Lemma 1.10.7 id t ! . Thus, by Lemma 3.1.37 the required would follow by

2

lim lim sup P 1=bn sup (Vkn U 0 nk ) 0 = 0: t!1 n!1 n k>bntc Denoting Æn = En (un1 Æn > 0 ,

(6.1.18)

v1n ) and in = vin uni + Æn , we have, since

Pn sup (Vkn U 0 nk ) 0 k>bntc

1 X l=blog2 (nt)c

© 2001 by Chapman & Hall/CRC

Pn

max

k=2l +1;:::;2l+1

k X i=1

!

in

kÆn

0

!

442

LD convergence for queues

+

1 X l=blog2 (nt)c

1 X

l=blog2 (nt)c

0 2l X @ Pn n

i=1

Pn

i

max

1

2l 1Æn A k X

k=1;:::;2l i=1

2

in

1 X

l=blog2 (nt)c

2l 1 Æn

Pn

max

!

k X

k=1;:::;2l i=1

in

!

2l 1Æ

n

:

Limit (6.1.18) now follows bypLemma A.3 in Appendix A and the near-heavy traÆc condition ( n=bn )Æn ! c0 > 0 as n ! 1 .

6.1.2 Idempotent diusion approximation for queueing networks We now extend some of the results of the preceding subsection to the queueing-network set-up. We consider a sequence of networks indexed by n. The nth network has a homogeneous customer population and consists of K FIFO single server stations. The network is open so customers arrive from outside and eventually leave. For the nth network, let An;k t ; k = 1; : : : ; K; denote the cumulative number of customers who arrive at station k from outside the network during the interval [0; t], and let Stn;k ; k = 1; : : : ; K; denote the cumulative number of customers who complete service at station k during the rst t units of busy time of that station. We call An = (An;k ; k = 1; : : : ; K ), where An;k = (An;k t ; t 2 R + ), and n;k n n;k n;k S = (S ; k = 1; : : : ; K ), where S = (St ; t 2 R+ ), the arrival process and service process, respectively (note that some of the entries in An may equal zero). We associate with the stations of the network the processes Rn;k = (Rn;kl ; l = 1; : : : ; K ); k = 1; : : : ; K , where n;kl ; m 2 N ), and Rn;kl denotes the cumulative number of Rn;kl = (Rm m customers among the rst m customers who depart station k that go directly to station l. The process Rn = (Rn;kl ; k; l = 1; : : : ; K ) is referred to as the routing process. We consider the processes An;k , S n;k and Rn;k as random elements of the respective Skorohod spaces D (R + ; R ), D (R + ; R ) and D (R + ; R K ); accordingly, An , S n and Rn are considered as random elements of D (R + ; RK ), D (R + ; RK ) and D (R + ; R K K ), respectively. We assume that the data associated with the nth network is de ned on a probability space ( n ; Fn ; Pn ).

© 2001 by Chapman & Hall/CRC

443

Moderate deviations for networks

We next introduce normalized and time-scaled versions of the arrival process, service process and routing process. Let n;k 2 R+ ; n;k 2 R+ ; and pkl 2 [0; 1], k = 1; : : : ; K; l = 1; : : : ; K . We de ne

An;k t =

n;k An;k nt pn;k nt ; S n;k = Snt pn;k nt ; t bn n bn n n;kl Rbntc pkl bntc p Rn;kl ; t = bn n

(6.1.19) (6.1.20)

p

where, as above, bn ! 1 and bn = n ! 0, and let An = (An;k ; k = 1; : : : ; K ), S n = (S n;k ; k = 1; : : : ; K ), Rn;k = (Rn;kl ; l = 1; : : : ; K ); k = 1; : : : ; K , and Rn = (Rn;kl ; k; l = 1; : : : ; K ). Again the latter processes are considered as random elements of D (R + ; RK ), D (R + ; R K ), D (R + ; R K ), and D (R + ; R K K ), respectively. Also we denote n = (n;k ; k = 1; : : : ; K ), n = (n;k ; k = 1; : : : ; K ) and P = (pkl ; k = 1; : : : ; K; l = 1; : : : ; K ). Elements of RK are regarded as column-vectors. Our main concern here is the queue-length process Qn = n;k n;k (Q ; k = 1; : : : ; K ), where Qn;k = (Qn;k t ; t 2 R + ), Qt denoting the number of customers at station k at time t. The associated normalized and time-scaled process Qn = (Qn;k ; k = 1; : : : ; K ) is de ned by

Qn;k t =

Qn;k pnt : bn n

(6.1.21)

In analogy with the hypotheses of Subsection 6.1.1 we assume that n ! = (^ 1 ; : : : ; ^ K ) and n ! = (^1 ; : : : ; ^K ) as n ! 1, where is a component-wise positive vector, and that \the nearheavy traÆc condition" holds: for some c 2 RK

pn bn

(n (EK P T )n ) ! c as n ! 1;

(6.1.22)

in particular,

= (EK P T ):

(6.1.23)

(Recall that EK denotes the identity K K matrix.) We also assume that the spectral radius of the matrix P is less than unity.

© 2001 by Chapman & Hall/CRC

444

LD convergence for queues

We recall that the skew re ection mapping RP , Harrison and Reiman [59], Reiman [115], is de ned as the map from D (R + ; RK ) into D (R + ; RK ) associating to each x = (xt ; t 2 R+ ) 2 D (R + ; RK ) such that xk0 2 R+ ; k = 1; : : : ; K; a function z = (zt ; t 2 R+ ) 2 D (R + ; R K ) such that

z = x + (EK P T )y, 2. y is componentwise increasing and y0k = 0; k = 1; : : : ; K , 1.

3.

Z1

z 2 R+ and zkt dytk = 0, k = 1; : : : ; K . k t

0

The map RP is well de ned and Lipshitz continuous for the locally uniform and Skorohod topologies on D (R + ; RK ) , Harrison and Reiman [59], Reiman [115], Chen and Whitt [23]. As in Subsection 6.1.1 all LD convergences below refer to the rate rn = b2n and the Skorohod topology. We recall the notation introduced at the end of Section 3.2. If x 2 D (R + ; RK ) has componentwise increasing R+ -valued paths, then, for y 2 D (R + ; R K ), we denote y Æ x = ((yxk kt ; k = 1; : : : ; K ); t 2 R+ ); analogously, if rt = (rklt ; k; l = 1; : : : ; K ) 2 RK K , then r Æ xt = (rkl xkt ; k; l = 1; : : : ; K ) . For vectors = (1 ; : : : ; K ) 2 RK and = ( 1 ; : : : ; K ) 2 RK , we denote = (1 1 ; : : : ; K K ) 2 RK . Let 1 denote the K -vector with all the entries equal to 1.

Theorem 26.1.8. Let the near-heavy-traÆc condition (6.1.22) hold. 1=bn

Let Qn0 Pn! q0 . Let the sequence f(An ; S n ; Rn ); n 2 N g LD converge in D (R + ; RK RK RK K ) to an idempotent process (A; S; R), where A = (A1 ; : : : ; AK ), S = (S 1 ; : : : ; S K ), and R = (R1 T ; : : : ; RK T ) are de ned by

A = AWA ; S = S WS ; Rk = kR WRk ; k = 1; : : : ; K; WA ; WS ; WRk ; k = 1; : : : ; K , being mutually independent K { dimensional idempotent Wiener processes and A; S ; kR ; k = ld 1; : : : ; K; being K K matrices. Then Qn ! Q, where Q is a Luzin-continuous idempotent process given by Q = RP (q0 + A +(R Æ e)T 1 (EK P T )S + ce):

© 2001 by Chapman & Hall/CRC

445

Moderate deviations for networks

Proof. The proof is a straightforward extension of the proof of Theorem 6.1.1. In analogy with (6.1.3), (6.1.4) and (6.1.2), we have that for k = 1; : : : ; K

Qn;k t

K X n;k n;k = Q0 + At + Rn;lk Æ Dtn;l l=1

Dtn;k ;

where Dtn;k = SRn;k t 1(Qn;k >0)ds : Introducing s

0

n;k C0 = t

Zt

0

1

(Qn;k s

n;k n;k Dnt 0 ; > 0)ds; D t =

n

we then have by (6.1.19), (6.1.20) and (6.1.21) that n;k n;k Qn;k t = Q0 + At +

+ +

K X l=1

pn bn

K X l=1

Rn;lk Æ D0 t

n;l

p

K X n;l n;k n 0 n;k 0 plk S Æ C t S Æ C t + (n;k + plk n;l n;k )t n;l

0 @n;k

bn

Zt

0

1

(Qn;k s

K X

= 0)ds

l=1

l=1

plk n;l

Zt

0

1

1

(Qn;l s

= 0)dsA ;

or in vector form

Qn = Qn0 + An + (Rn Æ D0 )T 1 (EK P T )S n Æ C 0 pn + (n (EK P T )n )e (EK P T )n C n ; (6.1.24) bn n

n

p

n;k where C nt = (C n;k n=bn 0t 1(Qn;k t ; k = 1; : : : ; K ), C t = s = n n;k n n;k 0 0 0 0 0)ds; C t = (C t ; k = 1; : : : ; K ), and D t = (D t ; k = 1; : : : ; K ). Hence, by the de nition of the re ection map RP R

Qn = RP Qn0 + An + (Rn Æ D0 )T 1 (EK P T ) S n Æ C 0 pn + (n (EK P T )n )e ; (6.1.25) bn n

© 2001 by Chapman & Hall/CRC

n

446

LD convergence for queues

so that from (6.1.24) (EK P T ) n C n n n = Qn0 + An + (Rn Æ D0 )T 1 (EK P T ) S n Æ C 0 pn + ( (EK P T )n )e bn n RP Qn0 + An + (Rn Æ D0n)T 1 (EK P T ) S n Æ C 0 n pn + (n (EK P T )n )e : bn

Since RP is a bounded map (in the sense that if z = RP (x), then sup0st zs K (t) sup0st jxs j, where K (t) depends only on t), the

p

1=b2n

ld convergences Qn0 Pn! q0 , (An ; S n ; Rn ) ! (A; S; R), and n=bn ! 1, the near-heavy-traÆc condition (6.1.22), and the facts that EK P T is nonsingular and is component-wise positive yield by the argument of the proof of (6.1.15) Zt

0

1

(Qn;k s

= 0)ds

2 Pn1=bn

! 0 as n ! 1; k = 1; : : : ; K; t 2 R+ ; 2

n;k 1=bn n;k implying that C 0 Pn! e as n ! 1. Then, since D0 t = 2 2 n;k n Pn1=bn 0 Pn1=bn 0 n;k n;k S Æ C =n and S =n ! e, we have that D ! e, so by n n ld (q0 ; A; S; R Æ 1 e) Lemma 3.2.11 (Qn0 ; An ; S n Æ C 0 ; Rn Æ D0 ) ! K K K K in D (R + ; R R R ). The claim now follows by (6.1.22), (6.1.25), continuity of the re ection and the contraction principle. The idempotent process Q is Luzin-continuous since the idempotent processes A, S and R are Luzin-continuous and RP is continuous.

Remark 6.1.9. One cann;kalso prove LD convergences for waiting and

sojourn times. Let Wt ; k = 1; : : : ; K; denote the virtual waiting p n;k time at station k at time t . We de ne W n;k t = Wnt =(bn n) and let W n = ((W n;k t ; k = 1; : : : ; K ); t 2 R + ) . For a vector k = k (k1 ; : : : ; kl ) , where ki 2 f1; 2; : : : ; K g , let An; t denote the number of customers with the routing (k1 ; k2 ; : : : ; kl ) who exogenously arrive by t , Ymn;k denote the sojourn time of the m th exogenous customer with

© 2001 by Chapman & Hall/CRC

447

Moderate deviations for networks

p

k n;k n;k the routing (k1 ; k2 ; : : : ; kl ) , and Y n; = t = Ybntc+1 =(bn n) , Y k 0n;k = (An;k =n; t 2 R ) . If, in addition to the (Y n; + t ; t 2 R+ ) , A nt hypotheses of the theorem,

2 n Pn1=bn W !

w0 , where q0 = w0 , then 2

1=bn ld (Qn ; W n ) ! (Q; W ) , where Q = W . If, in addition, A0n;k Pn! ld (W; Y ); where k e as n !P 1 ; where k > 0 , then (W n; Y n;k) ! Y Æ (k e) = li=1 Wki .

We now give an explicit expression for the idempotent distribution of Q. We de ne some more notation. For a subset J of f1; 2; : : : ; K g, we set FJ = f = (1 ; : : : ; K ) 2 RK+ : k = 0; k 2 k J; k > 0; k 62 J g and F J = f = (1 ; : : : ; K ) 2 RK + : = 0; k 2 J g ; 1J denotes the K -vector with entries from J equal to 1 and the rest of the entries equal to 0 ; J c denotes the complement of 0 K J . Let also RK; + denote the interior of R + , and K the set of all the subsets of f1; 2; : : : ; K g except the empty set. We introduce the positive semi-de nite symmetric matrix = ATA + (EK

P T )S TS (EK

Lemma 6.1.10.

P) +

K X k=1

^k R;k TR;k :

Let be positive de nite. Then the idempotent distribution of Q has rate function Z1 1 1(qt 2 RK;+ 0 )(q_ t r) 1(q_ t r) dt IQ (q) = 2 0 1 Z X1 + 1(qt 2 FJ ) y2infF c (q_ t 1J c r (EK P T )y) 1 2 J J 2K 0 (q_ t 1J c r (EK P T )y) dt when q 2 C (R + ; Rk ) is absolutely continuous and IQ (q) = 1 otherwise.

q0

= q0 , and

For a proof we need the following lemma which extends Lemma 6.1.4 and has a similar proof.

Lemma 6.1.11. Let z 2 C (R + ; RK ) have R+ -valued entry functions and x 2 C (R + ; RK ) be absolutely continuous. Then z = RP (x) if © 2001 by Chapman & Hall/CRC

448

LD convergence for queues

and only if z is absolutely continuous and there exists an absolutely continuous function y 2 C (R + ; RK ) such that

z_ t = x_ t +(EK

P T )y_ t a.e.

and

y0k = 0; y_ tk 2 R+ a.e.; zkt y_ tk = 0 a.e.; k = 1; 2; : : : ; K: Also z_ kt = 0 a.e. on the set ft 2 R+ : zkt = 0g, k = 1; 2; : : : ; K . Proof of Lemma 6.1.10. By the de nition of Q and Lemma 6.1.11 I Q (q) = 1 unless q0 = q0 and q is absolutely continuous. Since the idempotent Wiener processes WA ; WS and WRk ; k = 1; : : : ; K; are mutually independent, we have by Corollary 2.4.11 that

A + (R Æ e

)T

1 (EK

P T )S A WA +

K X

kR WRk Æ (^k 1e)

k=1 P T )

(EK

S WS

=

1=2 W;

where W is a K -dimensional idempotent Wiener process. Therefore,

q) =

I Q( =

1 2

Z1

inf

w2C (R+ ;RK ): q=RP (q0 + 1=2 w+ce)

q_ t : q yk =0

inf K k

y2R+ t 0

1 2

Z1

jw_ t j2 dt

0

c (EK

P T )y

q_ t

1

c (EK

P T )y dt;

which obviously coincides with the expression for I Q in the statement of the lemma. Let us now consider the i.i.d. case. Let, for some K 0 , ^ k > 0 0 when k = 1; : : : ; K 0 , and An;k t = 0 when k = K + 1; : : : ; K . n;k n;k Let the processes A ; k = 1; : : : ; K 0 , S ; k = 1; : : : ; K; and Rn;k ; k = 1; : : : ; K; be mutually independent for each n. Let the processes An;k ; k = 1; : : : ; K 0 ; and S n;k ; k = 1; : : : ; K , be renewal processes with times between renewals having nite second moments.

© 2001 by Chapman & Hall/CRC

449

Moderate deviations for networks

Let u^n;k ; k = 1; : : : ; K 0 ; denote the generic exogenous interarrival time and v^n;k ; k = 1; : : : ; K , the generic service time, for station k . Let, in addition, the routing mechanism not depend on n and be i.i.d. at each station with pkl being the probability of going directly from station k to station l. Lemma 6.1.12. Let under the above hypotheses, as n ! 1, En u^n;k ! 1=^ k ; Var u^n;k ! 2 ; k = 1; : : : ; K 0 ;

En

v^n;k

! 1=^k ; Var

v^n;k

u;k 2 ; ! v;k

k = 1; : : : ; K;

and either one of the following conditions be met: (i) supn En (^un;k )2+ < 1; k = 1; : : : ; K 0 ; and supn En (^vn;k )2+ < 1; k = 1; : : : ; K; for some > 0, and b2n= ln n ! 0; (ii) supn En exp((^un;k ) ) < 1; k = 1; : : : ; K 0 ; and supn En exp((^vn;k ) ) < 1; k = 1; : : : ; K; for some > 0 and 0 < 1, and b2n =n =2 ! 0. ld Then the LD convergence (An ; S n ; Rn ) ! (A; S; R) in the hypotheses of Theorem 6.1.8 holds for 2 ; : : : ; 2 ); ATA = diag(A; 1 A;K T 2 2 ); S S = diag(S;1 ; : : : ; S;K T pkl ); if m = l; k k R R = pkl (1 p p if m 6= l; l;m kl km ; k; l; m = 1; : : : ; K:

where

2 = 2 0 2 0 ^3 A;k u;k k ; k = 1; : : : ; K ; A;k = 0; k = K + 1; : : : ; K; 2 = 2 3 S;k v;k ^ k ; k = 1; : : : ; K:

Proof. Since the An , S n and Rn;k ; k = 1; : : : ; K; are mutually independent, by Lemma 3.1.42 it is suÆcient to prove the entry-wise conld ld ld vergence, i.e., An ! AWA , S n ! S WS and Rnk ! R;k WR;k ; k = 1; : : : ; K: All of them follow by Theorem 4.4.8. In more detail, for the LD convergence of the Rn;k we write bX ntc n;k i pnpk ; Rn;k = t b n i=1

© 2001 by Chapman & Hall/CRC

450

LD convergence for queues

where pk = (pkl ; l = 1; : : : ; K ) and in;k ; i 2 N ; are i.i.d. K -vectors, which have one entry equal to 1 and the rest equal to 0, the probability of the lth entry being equal to 1pbeing pkl . Clearly, the conditions of Theorem 4.4.8 are met with bn n as bn and (

()l;m =

pkl (1 pkl ) pkl pkm

if m = l; if m 6= l:

6.2 Very large and moderate deviations for many server queues In this section we derive results on LD convergence for many server queues. We consider a sequence of many server queues with exponential service times and Poisson arrival processes, which may be non-time-homogeneous. Arriving customers who nd no available servers form a queue and are served in the order of arrival. At time t the nth queueing system has Ktn homogeneous servers in parallel, arrival rate nt and service rate nt. We assume that the following expansions hold

p

p

nt = n0;t + nbn 1;t + O( n); p b nt = 0;t + pn 1;t + O(1= n); n p p n Kt = n0;t + nbn 1;t + O( n);

p

(6.2.1a) (6.2.1b) (6.2.1c)

where bn ! 1, bn = n ! 0, the functions 0;t , 1;t , 0;t , 1;t , 0;t , and 1;t are Lebesgue measurable, the functions 0;t , 1;t , 0;t , 1;t , and 1;t are bounded on bounded intervals, and the O's are uniform in t over bounded intervals. We do not rule out the case 0;t = 1, which corresponds to an in nite server queue. Let An = (Ant ; t 2 R+ ) and B n;k = (Btn;k ; t 2 R+ ); k 2 N ; be independent Poisson processes of respective rates nt and nt at time t. We assume that the objects associated with the nth system are de ned on a complete probability space ( n ; Fn ; Pn ). All the processes are considered as random elements of D (R + ; R). Denoting by Qnt the number of customers in the nth system at time t, we have

© 2001 by Chapman & Hall/CRC

451

Many server queues

that distributionally the process Qn = (Qnt ; t equation

Qnt

= Qn0 + Ant

Ktn Zt X

2

1(Qns k) dBsn;k :

k=1 0

R+ )

satis es the (6.2.2)

Let Zt

Mtn = Ant

0

ns ds+

Ktn Zt X k=1 0

1(Qns k)

dBsn;k n;k s ds :

(6.2.3) Then = t 2 R+ ) is a local martingale with respect to the ltration (Ftn ; t 2 R+ ), where Ftn = \>0Gtn+ and Gtn is the sub--algebra of F n generated by Qn0 , Ans ; Bsn;k ; s 2 [0; t]; k = 1; : : : ; K; and sets of Pn -measure zero. The predictable quadraticvariation process of M n has the form

Mn

hM n i

(Mtn ; Fn =

t

=

Zt

0

ns ds +

Zt

0

Qns ^ Ksn ns ds:

(6.2.4)

We write equation (6.2.2) in the form

Qnt

= Qn0 +

Zt

0

ns ds

Zt

0

Qns ^Ksn ns ds+Mtn :

(6.2.5)

Pn The following auxiliary result is standard. We denote by ! convergence in probability. Pn Lemma 6.2.1. Let Qn0 =n ! q0 2 R+ as n ! 1. Then, for T 2 R+ , Qn Pn sup t qt ! 0; t2[0;T ] n

where q = (qt ; t 2 R+ ) is the solution to the dierential equation

q_t = 0;t 0;t (qt ^ 0;t ):

Remark 6.2.2.

(6.2.6)

Equation (6.2.6) has a unique solution by Caratheodory's theorem, see, e.g., Coddington and Levinson [26].

© 2001 by Chapman & Hall/CRC

452

LD convergence for queues

Pn Proof of Lemma 6.2.1. It is easyR to check that RAnt =n2 ! 0. By R (6.2.4) hM n it 0t ns ds + 0t Qns ns ds 0t ns ds + Qn0 t + Rt n n n 2 Pn 0 As s ds so by (6.2.1a), (6.2.1b) and (6.2.1c) hM it =n ! 0 as n ! 1, which implies by the Lenglart-Rebolledo inequality that Pn supt2[0;T ] jMtn j=n ! 0 as n ! 1. Applying a standard tightness argument to (6.2.5) and using the fact the functions 0;t , 1;t , 0;t , and 1;t are bounded on bounded intervals, we conclude that the sequence of laws of the processes Qn =n is relatively compact in distribution, all the limit points satisfying equation (6.2.6) with probability 1. The solution of the latter equation being unique completes the proof.

The next result gives an idempotent diusion approximation for Let pn Qn t q n (6.2.7) Xt = bn n t and X n = (Xtn ; t 2 R+ ).

Qn .

2 Pn1=bn n Let X0 ! x0 2 R as n ! 1. If, in addition, inf s2[0;t] 0;s + 0;s(qs ^ 0;s ) > 0; t 2 R+ ; then X n ld! X as n ! 1 b2n for the Skorohod topology, where X = (Xt ; t 2 R+ ) is the idempotent

Theorem 6.2.3.

diusion speci ed by the equation

X_ t = 1;t 1;t (qt ^ 0;t ) 0;t 1(qt < 1;t )Xt + 1(qt = 0;t )(Xt ^1;t )+ 1(qt > 0;t )1;t q + 0;t + 0;t (qt ^ 0;t )W_ t ; X0 = x0 ; with W = (Wt ; t 2 R+ ) being an idempotent Wiener process. Proof. By (6.2.5), (6.2.6) and (6.2.7) we can write

Xtn

= X0n +

Zt

0

pn n s 0;s ds b n n

Zt

0

© 2001 by Chapman & Hall/CRC

b qs + Xsn pn ^ n

p

Ksn n n ( n bn s

0;s ) ds

453

Many server queues

pn Zt bn

b Kn 1 qs + Xsn pn ^ s qs ^ 0;s 0;s ds + p Mtn : n n nbn

0

Therefore, the rst characteristic of X n without truncation is

B 0n = t

Zt

bns (Xsn ) ds;

0

where

pn n s 0;s bn n pn

1 C~t0n = 2 bn

Zt

0

p

b Kn n n qs + u pn ^ s ( 0;s ) n bn s n b Kn qs + u pn ^ s qs ^ 0;s 0;s : bn n n The modi ed second characteristic without truncation C~ 0n = (C~t0n ; t 2 R+ ) p coincides with the predictable quadratic-variation process of (Mtn =( nbn ); t 2 R+ ) and by (6.2.4) has the form bns (u) =

c~0sn (Xsn ) ds;

where

ns n b Kn + s qs + pn u ^ s : n n n Easy calculations show that for t 2 R+ and v 2 R+ c~0sn (u) =

lim

n!1

Zt

0

lim

n!1

where

ess sup jbns (u) bs (u)j ds = 0; jujv Zt

0

ess sup jc~0sn (u) cs j ds = 0; jujv

bs(u) = 1;s 1;s (qs ^ 0;s ) 0;s 1(qs < 0;s )u + 1(qs = 0;s )(u ^ 1;s) + 1(qs > 0;s )1;s ; cs = 0;s + 0;s(qs ^ 0;s ):

© 2001 by Chapman & Hall/CRC

(6.2.8) (6.2.9)

454

LD convergence for queues

Thus, the convergence conditions of Theorem 5.4.4 are satis ed. The moment conditions are satis ed since the jumps of X n are bounded p above by 1=(bn n). Also the law of the semimaxingale with local characteristics (b; c; 0; 0) is uniquely speci ed by Theorem 2.8.21. ld Therefore, by Theorem 5.4.4 X n ! X at rate b2n as n ! 1.

Remark 6.2.4. We note that the idempotent distribution of X has density

x) = exp

X (

1 2

Z1

x_ t

bt (xt ) 2 dt ct

0

if x is absolutely continuous and x0 = x0 , and X (x) = 0 otherwise, where bt and ct are de ned by the respective equalities (6.2.8) and (6.2.9).

We consider now very large deviations. Let us assume, in addition, that inf s2[0;t] 0;s > 0; t 2 R+ . Let a process Y n = (Ytn ; t 2 R + ) be de ned by Ytn = Qn t =n. Let N1 = (N1 (t); t 2 R + ) and N2 = (N2(t); t 2 R+ ) be independent Poisson idempotent processes on an idempotent probability space ( ; ). Let Y = (Yt ; t 2 R+ ) be a Luzin solution of the equation

Yt = y0 +N1

Z t

0;s ds

N2

0

Zt

(Ys ^0;s)0;s ds ; y0 2 R+ ;

0

such that the idempotent distribution of Y has density

x) = exp

Y (

Z1

0

sup x_ t 2R

(e (e

1)0;t

1)(xt ^ 0;t )0;t dt

if x0 = y0 and x is absolutely continuous, and Y (x) = 0 otherwise. It is well de ned by Theorems 2.6.33, 2.8.10 and 2.8.29. 1=n

Theorem 6.2.5. If Y0n P!n Skorohod topology.

© 2001 by Chapman & Hall/CRC

y0 , then Y n

ld ! n

Y as n

! 1 for the

455

Many server queues

Proof. By (6.2.5)

Ytn

= Y0n +

Zt

0

Zt

ns ds n

0

Ysn ^

Ksn n 1 ds + Mtn : n s n

Therefore, Y n is a squarely integrable semimartingale. Its rst characteristic without truncation is given by

B 0n = t

Zt n

s

0

n

Ksn n n Ys ^ ds; n s

the predictable measure of jumps in view of (6.2.2) is given by

n([0; t];

)=n

Zt n 1

1 2 n n s

0

+ns Ysn ^

Ksn 1 1 n2 n

ds;

and the modi ed second characteristic without truncation in view of (6.2.4) is given by 1 C~t0n = n

Zt n

s

0

n

+ Ysn ^

Ksn n ds: n s

Then the convergence conditions of Theorem 5.4.3 hold with h(x) = x for

bs (u) = 0;s (u ^ 0;s )0;s ; c~s (u) = 0;s + (u ^ 0;s )0;s; s( ) = 0;s 1(1 2 ) + (u ^ 0;s)0;s 1( 1 2 ): The Cramer condition holds since the jumps of Y n are not greater than 1=n. Thus, by Theorem 5.4.3 the sequence of laws of the Y n is C (R + ; R )-exponentially tight of order n and every LD accumulation point for LD convergence of rate n is the law of a semimaxingale with local characteristics (b; 0; ; 0). By Theorem 2.8.29 the latter law is unique, hence, it is the LD limit of the laws of the Y n .

© 2001 by Chapman & Hall/CRC

Appendix A

Auxiliary lemmas This Appendix contains lemmas we referred to in the main body of the book. We rst prove the fact from convex analysis used in the proof of Lemma 1.11.5. For it we adopt the usual de nitions and notation from convex analysis, Rockafellar [117]. For a subset A of a Euclidean space, cl A denotes its closure, ri A the relative interior, rb A = cl Anri A the relative boundary, and conv A the convex hull of A. Let f be a function from Rd , d 2 N , into ] 1; 1]. Its conjugate (or the Legendre{Fenchel transform) f is de ned by f () = sup x f (x) ; 2 Rd ; x2Rd and the bipolar f of f is de ned as the conjugate of f :

f (x) = sup x f () ; x 2 Rd : 2Rd

Obviously, f is convex and lower semi-continuous. We denote by epi f the epigraph of f : epi f = f(x; y) 2 Rd R : y f (x)g; and let dom f = fx 2 Rd : f (x) < 1g denote the eective domain of f . The convex hull conv f of f is de ned by epi(conv f ) = conv(epi f ) and the lower semi-continuous hull cl f by epi(cl f ) = cl(epi f ): If f is convex, then @f (x) denotes the subdierential of f at x. We say that f is essentially strictly convex if it is strictly convex on every convex subset of the set of those x for which the 457 © 2001 by Chapman & Hall/CRC

458

Auxiliary lemmas

set @f (x) is nonempty. If f is essentially strictly convex, then it is strictly convex on ri(dom f ), Rockafellar [117].

Lemma A.1. If f

: Rd !] 1; 1] is a lower semi-continuous function and its bipolar f is strictly convex on ri(dom f ), then f = f . Proof. It is obvious that f f . So we prove the opposite inequality. By Rockafellar [117, Corollary 12.1.1 and the argument below] we have f = cl(conv f ): (A.1) We rst prove that f (x) f (x); x 2 ri(dom f ):

(A.2)

Assume the contrary, i.e., that for some x0 2 ri(dom f ) and > 0 we have f (x0 ) > f (x0 )+ : (A.3) Since x0 2 ri(dom f ), by Rockafellar [117, Theorem 23.4] the set @f (x0 ) is nonempty. Let 0 2 @f (x0 ). Then by the de nition of f f (x) 0 x f (0 ) (A.4) and by Rockafellar [117, Theorem 23.5] f (x0 ) = 0 x0 f (0 ):

(A.5)

It is easy to see that strict convexity of f implies that f (x) > 0 x f (0 ); x 6= x0 :

(A.6)

Indeed, if for some x 6= x0 we had equality in (A.4), then by convexity of f and (A.5) f (z ) 0 z f (0 ); for every z 2 [x0 ; x[, which together with (A.4) would yield f (z ) = 0 z f (0 ); z 2 [x0 ; x[: On the other hand, [x0 ; x[ ri(dom f ) (by Rockafellar [117, Theorem 6.1] and since x 2 dom f if there is equality in (A.4)). Thus, f would fail to be strictly convex on ri(dom f ), and (A.6) is proved.

© 2001 by Chapman & Hall/CRC

Appendix A

459

By lower semi-continuity of f we can choose " > 0 such that "j0 j < =3 and

f (x) > f (x0 ) ; jx x0 j < ": (A.7) 3 For this ", we choose Æ > 0; Æ < =3, satisfying the inclusion fx : 0 x f (0 )+Æ f (x)g fx : jx x0 j < "g: (A.8) In order to show that such a Æ exists, let us denote by AÆ the set on the left of (A.8). Then by (A.5) and (A.6) \

Æ>0

AÆ = fx0 g:

(A.9)

The hyperplane in Rd R de ned by the equation y = 0 x f (0 )+ Æ; x 2 Rd ; y 2 R; is parallel to the hyperplane y = 0 x f (0 ). The latter in view of (A.5) and (A.6) has with epi f the only point x0 in common. Then by Rockafellar [117, Corollary 8.4.1] the sets AÆ are bounded. They are closed since f is lower semi-continuous. Thus, the AÆ are compacts and (A.9) easily implies that for all Æ > 0 small enough AÆ fx : jx x0 j < "g proving (A.8). For the chosen Æ and ", we de ne fÆ;"(x) = max f (x); 0 x f (0 )+ Æ : (A.10) Obviously, fÆ;" is convex, lower semi-continuous and fÆ;"(x0 ) > f (x0 ) by (A.5). If we show that

fÆ;"(x) f (x); x 2 Rd ;

(A.11)

this would contradict (A.1), and (A.2) would be proved. It is clear that (A.11) holds on the set fx : jx x0 j "g since fÆ;"(x) = f (x) for these x by (A.8) and (A.10). If jx x0 j < ", then using (A.5), (A.3) and (A.7) we have

0 x f (0 ) + Æ = 0 (x x0 ) + f (x0 ) + Æ 2 + Æ < f (x) < "j0 j + f (x0 ) + Æ "j0 j + f (x) 3 (the latter inequality holds by the choice of " and Æ). Since, as we noted, f f , this proves (A.11) on fx : jx x0 j < "g. Thus (A.2) is proved.

© 2001 by Chapman & Hall/CRC

460

Auxiliary lemmas

Now if x 2 rb(dom f ) we have by Rockafellar [117, Theorem 7.5] in view of lower semi-continuity of f that for arbitrary z 2 ri(dom f ) f (x) = lim f (1 )z + x : (A.12) "1 By Rockafellar [117, Theorem 6.1] [z; x[ ri(dom f ), and then by the part just proved f (1 )z + x = f (1 )z + x ; 0 < 1; so that by lower semi-continuity of f and (A.12) we have that f (x) f (x) proving the assertion of the lemma for x 2 cl(dom f ). Finally, for x 62 cl(dom f ) we obviously have f (x) = f (x) = 1. For the next lemma we recall that 0 denotes the set of all Rd { valued piecewise constant functions ((t); t 2 R+ ) of the form

(t) =

k X i=1

i 1(t 2 (ti 1 ; ti ]);

where 0 t0 < t1 < : : : < tk ; i 2 Rd ; i = 1; : : : ; k; k 2 N :

Lemma A.2. Let f (t; ); t 2 R+ ; 2 Rd , be an R-valued function, which is Lebesgue R T measurable in t, continuous in , and is such that f (t; 0) = 0 and 0 f (t; )dt is well de ned for T 2 R+ and 2 Rd . Then for T 2 R+ ZT

0

ZT

sup f (t; ) dt = sup f (t; (t)) dt: ((t))20 0

2Rd

(A.13)

Proof. We denote F (t) = sup2Rd f (t; ): Since the supremum may be taken over the rational in view of continuity of f (t; ) in , the function F (t) is Lebesgue measurable and non-negative, so that the integral on the left-hand side of (A.13) is well de ned. Given arbitrary " > 0, we introduce the set

A" = f(t; ) 2 [0; T ]R d :

© 2001 by Chapman & Hall/CRC

1

f (t; ) (F (t) ")+ ^ 1" g:

461

Appendix A

By a measurable selection theorem, see, e.g., Clarke [25], Ethier and Kurtz [48], there exists an Rd {valued Lebesgue measurable function ~ " (t) such that 1 f (t; ~"(t)) (F (t) ")+ ^ 1" ; t 2 [0; T ]: By Luzin's theoremthere exists a continuous function " (t) such that RT 2 ~ 0 1 " (t) 6= " (t) dt < . Then ZT

f (t; " (t)) _ 0 dt

0

ZT

f (t; ~ "(t)) dt

0

ZT

0

1 (F (t) ")+ ^ dt : "

Since ( (t)) is continuous, it can be approximated by functions from 0 . Since f (t; ) is continuous in and f (t; 0) = 0, by Fatou's lemma there exists a function 0 2 0 such that ZT

f (t; 0 (t)) dt

0

ZT

f (t; " (t)) _ 0 dt :

0

Thus, since > 0 is arbitrary, ZT

0

ZT

sup f (t; ) dt sup f (t; (t)) dt: ((t))20 0

2Rd

The reverse inequality is obvious. We state and prove the lemma used in the proof of Theorem 6.1.7. Lemma A.3. Let fin; i 2 N g; n 2 N , be a triangular array of rowwise i.i.d. real-valued r.v. with zero mean on respective probability spaces ( n ; Fn ; Pn ). Let bn ! 1 as n ! 1, and > 0. p (i) If bn = n ! 0 as n ! 1 and, for some " > 0, we have supn En j1n j2+" < 1, then there exist n0 , t0 > 0, C1 > 0, and C2 > 0 such that, for all t t0 and n n0 ,

Pn

1

k X

p max n > t k=1;:::;bntc bn n i=1 i

!

p exp( C1b2n t) + C2

© 2001 by Chapman & Hall/CRC

" 1 b2+ n : (A.14) n"=2 t"=2

462

Auxiliary lemmas

(ii) If, for some > 0 and 2 (0; 1], we have supn En exp( j1n j ) < 1 and b2n =n =2 ! 0 as n ! 1, then there exist n00, t00 > 0, C10 > 0 and C20 > 0 such that, for all t t00 and n n00 ,

Pn

1

k X

p max in > t k=1;:::;bntc bn n i=1

!

p exp( C10 b2n t) 0 p

+ exp( C2 (bn nt) ): (A.15) Proof. The argument uses the ideas of the proof of Theorem 4.4.8. Let the conditions of (i) hold. We rst prove that there exist C1 > 0 and t0 such that for t t0

Pn

max

k=1;:::;bntc bn

1

pn

k X i=1

in 1

p pbnn jinj t > t

!

p exp( C1 b2n t): (A.16)

We denote B = supn En j1n j2+" + 1 and bn n p n ^in = pbn in 1 pbn jin j pt En i 1 p ji j t : n n n (A.17) By Doob's inequality (see, e.g., Liptser and Shiryaev [79, Theorem 1.9.1]), for > 0, bntc ! 2^1n k E e X n 1 t Pn max 2 ^in > eb2n t : (A.18) 2 k=1;:::;bntc bn i=1

p Since En ^1n = 0, j^1n j 2 t and En (^1n )2 En (1n )2 b2n =n, it follows that p p b2 n En e2^1 1+22 e4 t En (^1n )2 1+22 e4 t n B; n so p n bntc En e2^1 exp(22 e4 t Btb2n): p

Choosing in (A.18) = 1= t, we obtain for t t0 = (4e4 B=)2 and C1 = =2

Pn

k 1 X t max 2 ^n > 2 k=1;:::;bntc bn i=1 i

© 2001 by Chapman & Hall/CRC

!

p exp( C1b2n t):

(A.19)

463

Appendix A

Now note that, since En 1n = 0, En n 1

1

p pbnn j1n j t

= En n 1

1

p pbnn j1nj > t " B b1+ n n(1+ ")=2 t(1+")=2 ;

hence,

bntpc E n 1 pbn jnj a b"n Bt(1 ")=2; bn n n 1 n 1 n"=2 p so, by the fact that bn = n ! 0 as n ! 1 and (A.17), for all n large enough and t t0 Pn

max

k=1;:::;bntc bn

1

pn

k X i=1

in

1

p pbnn jinj t > t

!

k 1 X t Pn k=1max ^in > ; 2 2 ;:::;bntc bn i=1

which together with (A.19) proves (A.16). The estimate (A.14) now follows by (A.16) and the inequalities

Pn

max

k=1;:::;bntc bn

Pn

1

pn

k X i=1

!

in

> t !

p b p max in 1 pn jin j t > t n k=1;:::;bntc bn n i=1 p b + Pn max pn jin j > t (A.20) k=1;:::;bntc n 1

k X

and

Pn

p b max pn jin j > t k=1;:::;bntc n

Part (i) is proved.

© 2001 by Chapman & Hall/CRC

p bntcPn pbnn j1nj > t 2+" bntc nb1+n "=2 t1+B"=2 :

464

Auxiliary lemmas

For part (ii), we write

Pn

max

k=1;:::;bntc bn

Pn

1

pn

max

k=1;:::;bntc bn 0

+ Pn @

1

p

k X i=1

1

pn

bX ntc

bn n i=1

!

in > t

k X i=1

in 1

p pbnn jin j t > t2

!

1

bn n p n ji j 1 pn ji j > t > t2 A : (A.21)

Noting that the conditions of part (ii) imply the conditions of part (i), we estimate the rst term on the right of (A.21) with the help of (A.16). For the second, we use the inequality 0

Pn @

bX ntc

1

p

bn n i=1

Pn b n + Pn

1

p t A b n n n ji j 1 pn ji j > t > 2

p pn k=1max j in j > t ;:::;bntc

1

1

p

bntc X

bn n i=1

jinj

1

p pbnn jinj > t 1 n p t p 1 bn n ji j t > 2 : (A.22)

We rst work with the second probability on the right. We have, for > 0 by Chebyshev's inequality

Pn

bX ntc

b p p j in j 1 pn jin j > t bn n i=1 n

1

1 bn1pn jinj pt

© 2001 by Chapman & Hall/CRC

>

t 2

465

Appendix A

p En exp 2 pbnn j1nj 1 pbnn j1nj > t 1 p bntc n p 1 b n j1 j t exp( b2n t)

n

p exp nt log En exp 2 pbnn j1nj 1 pbnn j1nj > t 1 b 1pn j1nj pt b2nt : (A.23)

n

Next, for 0 < < 1; c > 0 and c =4;

p b b 1 n p p j j t En exp 2 pn j1n j 1 pn j1n j > t 1 n n bn n 1 bn n bn n 1 En exp 2 pn j1 j 1 pn j1 j > c 1 n p 1 bnpn j1 j t p b b b n n n n n 1 n + En exp 2 p j1 j 1 p j1 j c 1 p j1 j > t n n n p ! 1 p En exp j1n j exp 2b2n t cb n n

+ En exp

p

2c + j1n j exp 2

p

2

p

nt bn

!

: (A.24)

Taking = 1=(2 t) and c = t=2, and using the condition n =2 =b2n ! 1 as n ! 1, we conclude that the rightmost side p of (A.24) is not greater than exp C~ nt=bn for some C~ > 0. Substituting the estimate into (A.23) and again using the convergence n =2 =b2n ! 1 implies that, for all n and t large enough, bX ntc

b p p j in j 1 pn jin j > t bn n i=1 n 1 1 b pn jinj pt > t2

Pn

1

n

© 2001 by Chapman & Hall/CRC

p exp C100b2n t :

466

Auxiliary lemmas

By a similar argument, this bound is seen to hold for = 1 as well. Finally, the rst term on the right of (A.22) is estimated as

p En e pj1n j n p Pn max j j > t nt (b nt) bn n i=1;:::;bntc i e n p exp( C20 (bn nt) ):

1

Substituting the estimates into (A.21) nishes the proof of (ii).

© 2001 by Chapman & Hall/CRC

Appendix B

Notes and remarks Part I This part considers idempotent analogues of the constructions of probability theory. They also belong to the realm of possibility theory so one can replace the adjective \idempotent" with \possibilistic" (or, perhaps, \fuzzy"). The observation of the analogy between certain probabilistic and \max-plus" constructions seems to have rst been made in Baccelli et al. [8].

Section 1.1 It appears that maxitive measures were rst introduced by Shilkret [119], who also studied properties such as convergence, Egorov's theorem, and others. Idempotent measures are known as possibility measures in fuzzy measure theory, see, e.g., Dubois and Prade [40], Wang and Klir [133], de Cooman, Kerre and Vanmassenhove [33], Pap [101] (who also uses the term \maxitive measure"), Mesiar [86], and references therein; and as A-measures in idempotent measure theory, see Kolokoltsov and Maslov [73] (del Moral in Kolokoltsov and Maslov [73] uses the name \performance measure"). Another name is \cost measure", see Akian, Quadrat and Viot [2, 3]. Both Wang and Klir [133] and Kolokoltsov and Maslov [73] use the requirement of -maxitivity as a de nition and call the property fuzzy additivity and complete additivity, respectively. Pap [101] uses the name \complete maxitivity". Some authors replace the -maxitivity 467 © 2001 by Chapman & Hall/CRC

468

Notes and remarks

property by -maxitivity, see, e.g., Akian [1], Pap [101]. In the topological setting similar objects have been studied by Norberg [95], and O'Brien and Vervaat [97]. The latter authors use the name \supmeasure", which is explained by the \sup-representation" (1.1.2), and require certain inner and outer regularity properties rather than -smoothness. Possibility measures with inner and outer regularity properties on topological spaces have been considered by Janssen, de Cooman and Kerre [68]. Our usage of the concept of -smoothness is consistent with the one adopted in measure theory, see Topse [125], Vakhania, Tarieladze and Chobanyan [126]. A -smooth idempotent measure is a speci c case of a Choquet capacity, see, e.g., Meyer [88] or Neveu [94]. Our study uses some of the ideas as well as the terminology of the theory of Choquet capacities. The de nition of a maxitive set function is due to Norberg [95]. Since the collection Eiu contains the collection of E -analytic (or Suslin) subsets of , see, e.g., Kuratowski and Mostowski [76], Meyer [88] or Neveu [94] for the de nition, Theorem 1.1.7 is a (very simple) analogue of Choquet's theorem, Meyer [88, T19], Neveu [94]. The de nition of a paving is borrowed from Meyer [88]. Theorem 1.1.9 is in the theme of Meyer [88, Theorem IIIT23] and Wang and Klir [133, Theorem 4.9], and runs parallel to the result on the extension of a measure from a ring to a -ring, see, e.g., Halmos [58]. The proof uses the construction of Wang and Klir. One can also obtain the same extension of by applying the construction used by Meyer [88, IIIT23]: de ne, for A 2 Eu , (A) = sup (F ); F 2E F A and, for arbitrary B , let (B ) = inf (A): A2Eu AB However, checking the necessary properties is more complicated. The latter approach is better suited to -maxitive measures, cf. Akian [1]. Objects that we call -algebras have been known in possibility theory as ample elds or complete elds, see Wang [132], De Cooman and Kerre [31, 32], and Wang and Klir [133]. Both our de nition

© 2001 by Chapman & Hall/CRC

Appendix B

469

of and notation for atoms are consistent with Wang [132], and De Cooman and Kerre [31, 32]. Most of the properties of -algebras stated in this section can be found in these papers. Corollary 1.1.22 has been prompted by Neveu [94, Proposition I.6.1].

Section 1.2 Functions measurable with respect to ample elds are called fuzzy variables in fuzzy measure theory, see Wang [132], De Cooman and Kerre [31, 32], and Wang and Klir [133]; Janssen, de Cooman and Kerre [68] use the name \possibilistic variables". Measurability properties for more general set functions are considered in Pap [101]. Images of possibility measures are considered in Wang [132], de Cooman, Kerre and Vanmassenhove [33]. Lemma 1.2.7 is an analogue of Doob's theorem on representation of measurable functions, see, e.g., Meyer [88, IT18]. The proof is also along the lines of the proof given in Meyer [88, IT18]. O'Brien and Vervaat [97] distinguish between tightness and classical tightness. We need only the latter concept for which we reserve the name \tightness". The notion of Luzin measurability with respect to idempotent measures is an analogue of Luzin measurability in measure theory, see Schwartz [118], Vakhania, Tarieladze and Chobanyan [126]. Theorem 1.2.14 is an abstract version of a result in large deviation theory (cf., e.g., Deuschel and Stroock [36]).

Section 1.3 Modes of convergence have also been studied by Shilkret [119] and del Moral in Kolokoltsov and Maslov [73]. Extensions of many of the results of the section to more general set functions are given in Pap [101]. For analogues in probability theory see, e.g., Shiryaev [120].

Section 1.4 The notion of idempotent integral has been introduced by Shilkret [119], who also studied its basic properties, but the name seems to

© 2001 by Chapman & Hall/CRC

470

Notes and remarks

be due to Maslov [84, 85]. Similar constructions appear in Norberg [95], see also Vervaat [130]. More general integrals are studied in fuzzy measure theory, see Dubois and Prade [40], de Cooman, Kerre and Vanmassenhove [33], Wang and Klir [133], Wu, Wang, and Ma [137], Pap [101], de Cooman and Kerre [31], Mesiar [86], Guo, Zhang, and Wu [56], Mesiar and Pap [87], and references therein; integrals of lattice-valued functions have been considered in Akian [1], de Cooman and Kerre [31], Mesiar and Pap [87], Pap [101]. Lemma 1.4.5 also holds for so called pan integrals, Wang and Klir [133]. Part 1 of Lemma 1.4.7 appears in Kolokoltsov and Maslov [73], Theorem 1.4.11 is stated by del Moral in Kolokoltsov and Maslov [73], it also appears in Puhalskii [111]. SuÆciency of the existence of the function F for uniform maximability in Corollary 1.4.14 is stated by del Moral in Kolokoltsov and Maslov [73], who also studies convergence of idempotent integrals. Our analysis of the convergence properties is based on Puhalskii [111]. For convergence properties in a more general setting see Pap [101]. The proof of Theorem 1.4.22 uses the ideas of the proof of Daniell's theorem in Meyer [88, III.2.24]. For another form of the Daniell property see Pap [101].

Section 1.5 Products of idempotent measures have been studied in more generality in fuzzy set theory, see, e.g., de Cooman, Kerre and Vanmassenhove [33], Janssen, de Cooman and Kerre [68], and references therein; they have also been analysed by Kolokoltsov and Maslov [73]. Products of ample elds have been considered by Wang [132].

Section 1.6 Exposition is based on Puhalskii [111], who however conditions on collections of analytic sets rather than -algebras. The notions of independence and conditioning for idempotent variables in the fuzzy set theory context have been studied in Wang [132], de Cooman, Kerre and Vanmassenhove [33]. Similar properties as well as a de nition of conditional idempotent expectation with respect to -algebras are considered by del Moral in Kolokoltsov and Maslov [73] (see Re-

© 2001 by Chapman & Hall/CRC

Appendix B

471

mark 1.6.26). For other approaches see Akian, Quadrat and Viot [2, 3]. Absolute continuity has been studied in the fuzzy set theory context, see, e.g., Mesiar [86] and references therein; a general treatment appears in Pap [101]; however, the de nitions are stated for -algebras and the weaker notion of absolute continuity (cf. Remark 1.6.30).

Section 1.7 The concept of Luzin measurability for idempotent variables on topological spaces has been introduced in Puhalskii [111], for the measuretheoretic analogue see Schwartz [118], and Vakhania, Tarieladze and Chobanyan [126]. The rst result in the theme of Theorems 1.7.21, 1.7.23, and 1.7.25 seems to be due to Choquet [24]. Theorems 1.7.21 and 1.7.25 appear in Breyer and Gulinski [17]; our proof of Theorem 1.7.25 uses their idea of invoking the Stone-Czech compacti cation. Puhalskii [107, 108] proves a similar result for metric spaces under the additional condition of sub-additivity of the functional V . Kolokoltsov and Maslov [73, Theorem 1.5, ch.1] prove the result of Theorem 1.7.21 for a locally compact normal space and functions with values in an idempotent metric semiring. They also prove the stated representation for the case where V is a continuous homeomorphism from Cb+(E ) equipped with the topology of pointwise convergence to an idempotent metric semiring and E is Tihonov. Akian [1] considers the same characterisation in terms of continuity for integrals of lattice-valued functions. Algebraic versions appear in Litvinov, Maslov and Shpiz [80, 81].

Section 1.8 In this section we use some of the ideas of Schwartz [118]. The result in Lemma 1.8.3 is a special case of a result in large deviation theory due to Dawson and Gartner [28], see also Dembo and Zeitouni [35]. The setting of Theorem 1.8.6 for regular possibility measures has been considered by Janssen, de Cooman and Kerre [68].

© 2001 by Chapman & Hall/CRC

472

Notes and remarks

Sections 1.9 and 1.10 The results are modelled after weak convergence theory of probability measures, see Billingsley [11], Parthasarathy [102], Topse [125, 124], Vakhania, Tarieladze and Chobanyan [126]. For a prototype see Vervaat [129]. The setting of metric spaces is studied in Puhalskii [108]. For facts about uniform spaces used in the proof of Theorem 1.9.2 see, e.g., Engelking [47]. Theorem 1.9.28 is an analogue of Ranga Rao's result, see Vakhania, Tarieladze and Chobanyan [126]. More general compactness results and other properties of the vague topology are in O'Brien and Vervaat [97] and O'Brien and Watson [99]. For the de nitions of the Prohorov and Kantorovich-Wasserstein metrics for probability measures see, e.g., Dudley [41]. Jiang and O'Brien [69] de ne the Prohorov metric on a space of set functions that includes idempotent probability measures and probability measures and show, in particular, that it metrises convergence of sequences in the narrow topology; they also show that the KantorovichWasserstein metric has this property for sequences of the exponentials of rate functions and address the issue of characterising tight collections as totally bounded sets. For the de nition and properties of Mosco convergence see Mosco [92], Zabell [138] and references therein.

Section 1.11 Kolokoltsov and Maslov [73] refer to the Laplace-Fenchel transform as the Fourier-Legendre transform. The inversion formula appears in Puhalskii [108]. Lemma 1.11.19 is also taken from the latter paper. For required facts from convex analysis see Rockafellar [117] and Appendix A.

Sections 2.1 { 2.6 The results and approaches are analogous to those in stochastic calculus, see Dellacherie [34], Elliott [45], Ikeda and Watanabe [66], Jacod and Shiryaev [67], Liptser and Shiryaev [79], Meyer [88], Neveu [94], ksendal [100], and Stroock and Varadhan [123]. Idempotent martingales have been considered by Del Moral in Kolokoltsov and Maslov [73] (for conditional expectations with respect to -algebras).

© 2001 by Chapman & Hall/CRC

473

Appendix B

For other approaches see Akian, Quadrat and Viot [2, 3]. \Possibilistic" processes have been studied in Janssen, de Cooman and Kerre [68]. Theorems 2.2.26 and 2.2.27 are taken from Puhalskii [108]. Section 2.3 is based on Puhalskii [111]. The de nitions of the idempotent Wiener and Poisson processes in Section 2.4 are motivated by the fact that the associated rate functions appear in the large deviation principles for Wiener and Poisson processes, respectively, see, e.g., Borovkov [13], Freidlin and Wentzell [51]. For the properties of the pseudo-inverses of matrices see, e.g., Campbell and Meyer [20]. Theorem 2.6.22 is in essence the Picard-Lindelof-Caratheodory theorem, see, e.g., Coddington and Levinson [26], Hartman [60].

Sections 2.7 The results are based on Puhalskii [109, 111]. We follow the ideas of stochastic calculus, see, e.g., Liptser and Shiryaev [79], Jacod and Shiryaev [67]; in particular, the de nition of a semimaxingale is analogous to the exponential characterisation of semimartingales. Del Moral in Kolokoltsov and Maslov [73] has considered idempotent semimartingales for conditional expectations with respect to algebras. Lemma 2.7.5 is in essence due to Liptser and Shiryaev [79, Theorem 6.2.3], whose argument also applies to the proof. Theorem 2.7.16 admits a revealing interpretation in terms of Orlicz spaces, Krasnosel'skii and Rutickii [75]. Speci cally, for x 2 C and t 2 R+ , let Lg^(x) (0; t) denote the set of all functions f (s); s t; such that Zt

0

g^s

1 f (s); x ds < 1;

for some > 0. Lg^(x) (0; t) is easily seen to be a vector space. Let for f 2 Lg^(x) (0; t)

Zt

kf k Lg^( )(0;t) = inf > 0 : g^s x

© 2001 by Chapman & Hall/CRC

0

1 f (s); x ds 1 :

474

Notes and remarks

This can be shown to de ne a seminorm on Lg^(x) (0; t), which is a norm if g^s (; x) 6= 0 for 6= 0 (cf. Krasnosel'skii and Rutickii [75]). Let us assume, for the moment, that g^s (; x) does not depend on s: g^s (; x) = g^(; x). Then the above norm is called the Luxembourg norm and Lg^(x) (0; t) is called an Orlicz space, Krasnosel'skii and Rutickii [75]. Also in this case the set of functions, for which (2.7.18) holds, is the closure of the space of bounded functions in the Luxembourg norm, Krasnosel'skii and Rutickii [75]. In analogy with Krasnosel'skii and Rutickii [75], we denote this set by Eg^(x) (0; t). We then have the following insight into the statement of Theorem 2.7.16. Let Lg^(x) (0; t) be the set of functions f such that Zt

0

g^(f (s); x) ds < 1:

Then, by Krasnosel'skii and Rutickii [75], we have the strict inclusions

Eg^(x) (0; t) Lg^(x) (0; t) Lg^(x) (0; t);

unless g^(; x) satis es the weak growth condition (or the 2 { condition):

g^(2; x) lim sup < 1: !1 g^(; x)

If g(; x) has the semimartingale representation (2.7.55), then the weak growth condition means that Ks (Rd ; x) = Ls (Rd ; x) = 0, i.e., it holds only in \the diusion case". Hence, generally, the class of functions , for which we have proved that Z () is a -local exponential maxingale, is smaller than the class de ned by the condition Zt

0

jgs ((s; x); x)j ds < 1:

We do not know if Theorem 2.7.16 can be extended to a larger set of functions . We also note that the proof of Theorem 2.7.16 has been prompted by the methods of the theory of Orlicz spaces.

© 2001 by Chapman & Hall/CRC

Appendix B

475

Section 2.8 The results are based on Puhalskii [112]. Our conditions for uniqueness of solutions to maxingale problems are similar to conditions required for corresponding martingale problems. It is thus instructive to compare our results with those for martingale problems in Ikeda and Watanabe [66], Jacod and Shiryaev [67], Stroock and Varadhan [123]. The setting of Theorem 2.8.5 corresponds to the situation where a martingale problem is speci ed by a deterministic triplet of predictable characteristics so that the associated process is a process with independent increments; the problem then has a unique solution, see, e.g., Jacod and Shiryaev [67, Theorem III.2.16]. The function (s; x; y) in the hypotheses of Theorem 2.8.27 exists and equals the gradient rhs (y; x) if the latter exists for (almost all) s 2 R+ ; y 2 G and x -almost all x, and is bounded on the sets [0; t] K Gm , where t 2 R+ ; m 2 N and K is compact in C . The role of conditions I and II and the conditions in Theorem 2.8.33 is analogous to the role of conditions A{E in Wentzell [134]. A distinctive feature of our conditions is that they are stated only in terms of the cumulant and do not invoke its Fenchel{Legendre transform (as in Wentzell [134]). We believe that this makes the conditions easier to check. Also we relax the requirements on boundedness and continuity of the cumulant. Condition (2.8.14) is analogous to condition III in Liptser and Puhalskii [78]. The regularisation approach of Lemma 2.8.26 applied later in the section has earlier been used in Wentzell [134], and Liptser and Puhalskii [78] in the large deviation setting. For background on the notions used in Lemma 2.8.31 see Aubin and Cellina [5], von Leichtweiss [131], and Rockafellar [117]. For measurable-selection theorems see Clarke [25, Theorem 4.1.1], Ethier and Kurtz [48], or Dellacherie [34, IT37].

Part II Exposition is based on Puhalskii [106] { [114]. Standard manuals on large deviation theory are Freidlin and Wentzell [51], Varadhan [128], Deuschel and Stroock [36], Dembo and Zeitouni [35].

© 2001 by Chapman & Hall/CRC

476

Notes and remarks

Section 3.1 Some of the results of this section are large deviation convergence versions of the results formulated in the setting of the large deviation principle, see Varadhan [128], Stroock [122], Deuschel and Stroock [36], Dembo and Zeitouni [35], Bryc [18], Dinwoodie [37]. For an approach from the point of view of convergence of capacities see O'Brien and Vervaat [97], O'Brien [96], O'Brien and Watson [99]. Since this section considers similar issues as Section 1.9, most of the comments to that section apply here. In particular, there are many analogies with results in weak convergence theory, see, Billingsley [11], Parthasarathy [102], Topse [125, 124], Vakhania, Tarieladze and Chobanyan [126]. The de nition of the large deviation convergence and the term itself have been introduced in Puhalskii [108, 109]. Properties of a more general type of convergence have been considered by Pap [101], Mesiar and Pap [87]. Theorem 3.1.3 for the setting of metric spaces has appeared in Puhalskii [107]. It combines a number of earlier results. The fact that part 3 implies part 2 is \Varadhan's lemma", Varadhan [127], who also proves Lemma 3.1.12, the converse under the additional condition of exponential tightness is due to Bryc, see Varadhan [128], Bryc [18], Dembo and Zeitouni [35]. Instead of the de nition we have adopted for large deviation convergence one could use part 2 of Theorem 3.1.3 in order to de ne \weak large deviation convergence". It would then be equivalent to the large deviation principle, or \narrow large deviation convergence", cf. Remark 1.9.6. The name \contraction principle" as given by Varadhan [128] refers to the case of continuous f in Corollary 3.1.15. Corollary 3.1.15 for convergence of sequences has appeared in Puhalskii [106], it is referred to in Puhalskii and Whitt [113] as the extended contraction principle. Theorem 3.1.14 for metric spaces and sequences of probability measures has been proved in Puhalskii [110]; Chaganty [22] proves the statement in the setting of Polish spaces under the assumption that the convergence fn(zn ) ! f (z ) holds for every z 2 E , a similar type of condition has been used by Dinwoodie and Zabell [38]. Other versions and generalisations have been considered by Deuschel and Stroock [36] and O'Brien [96]. Theorems 3.1.19 and 3.1.28 are analogues of Prohorov's criterion for weak convergence, Prohorov [104], see also Billingsley [11],

© 2001 by Chapman & Hall/CRC

Appendix B

477

Vakhania, Tarieladze and Chobanyan [126]. Theorem 3.1.28 has appeared in Puhalskii [106]. The proofs here are along the lines of the one in Puhalskii [107], who also mentions the extension to Tihonov spaces. A more general setting has been considered by O'Brien and Vervaat [97] so that Theorems 3.1.19 and 3.1.28 for separable metric spaces follow from Theorem 3.11 and Lemma 5.2 there; an announcement had been made in Vervaat [129]. Lynch and Sethuraman [82] have proved that LD convergence implies exponential tightness for sequences of probability measures on complete separable metric spaces; an extension appears in Jiang and O'Brien [69]. De Acosta [30] has generalised and simpli ed the proof of Puhalskii [106] to give weaker conditions for part 1 of Theorem 3.1.19 to hold, in particular, extending the result to Hausdor topological spaces (similarly to Prohorov's tightness criterion in weak convergence theory, see Topse [125]) and probabilities on non-Borel -algebras. For other versions and extensions see Vervaat [129], O'Brien and Vervaat [98]. The vague large deviation convergence has extensively been studied by O'Brien and Vervaat [97, 98] who prove Theorem 3.1.34. Jiang and O'Brien [69], considering a more general setting, prove that the Prohorov and the Kantorovich-Wasserstein metrics metrise LD convergence of sequences. They also address the issue of characterising tight collections as totally bounded sets. See also Dembo and Zeitouni [35] for some of the results. Theorem 3.1.31, which is usually referred to as the GartnerEllis theorem, has been proved by Gartner [53] for the case where G() is smooth and nite everywhere. The extension to the essentially smooth case has been obtained by Freidlin and Wentzell [51] and, later and apparently independently, by Ellis [46]. The name \Gartner-Ellis" appears to have been rst used in Bucklew [19]. Our method of proof follows Puhalskii [109] and O'Brien and Vervaat [98]. Theorem 3.1.32 is an analogue of Ranga Rao's result in weak convergence, see Vakhania, Tarieladze and Chobanyan [126]. In a somewhat dierent form it appears in Jiang and O'Brien [69]. Part 1 of Lemma 3.1.42 in the form of the LDP extends to regular spaces, see Cegla and Klimek [21].

© 2001 by Chapman & Hall/CRC

478

Notes and remarks

Section 3.2 For more background on the concepts and properties concerning stochastic processes on Skorohod spaces the reader is referred to Ethier and Kurtz [48], Jacod and Shiryaev [67], Lindvall [77], Liptser and Shiryaev [79], Skorohod [121], and Whitt [135]. The notion of C -exponential tightness has been introduced in Liptser and Puhalskii [78]. Theorem 3.2.3 is an analogue of Aldous' tightness condition, Aldous [4], and has appeared in Puhalskii [106]. Lemma 3.2.5 can also be used in order to give a somewhat dierent proof of Aldous' original result. Feng and Kurtz [50] obtain dierent exponential tightness conditions. For the Lenglart-Rebolledo inequality, see, e.g., Liptser and Shiryaev [79, Theorem 1.9.3]. Theorems 3.2.8 and 3.2.9, taken from Puhalskii [106, 107, 109, 111], are analogues of the methods of nite-dimensional distributions and the martingale problem in weak convergence theory, cf. Jacod and Shiryaev [67], Liptser and Shiryaev [79], Ethier and Kurtz [48]. Lemmas 3.2.11 and 3.2.13 are LD convergence versions of results in Puhalskii and Whitt [113], which are also more general. See also Lemma 4.2 there for more detail about the proof of Lemma 3.2.13. Prototypes for weak convergence are in Whitt [135].

Chapter 4 The method of nite-dimensional distributions is essentially an adaptation of projective limit arguments, see Dawson and Gartner [28], to the setting of stochastic processes. It was rst used by Varadhan [127], see also Dembo and Zeitouni [35]. Our exposition follows Puhalskii [108, 109]. The main results, both in content and form, are analogous to results on convergence in distribution of a sequence of semimartingales to a process with independent increments in Liptser and Shiryaev [79] and Jacod and Shiryaev [67]. For results on the LDP for processes with independent increments see Varadhan [127], Borovkov [13], Mogulskii [90, 91], Lynch and Sethuraman [82], de Acosta [29], Puhalskii [106]. Lemma 4.1.1 is an analogue of results in Liptser and Shiryaev [79, Chapter 2, x3], Jacod and Shiryaev [67, Chapter 2, x2d] for complex-valued stochastic exponentials. Theorem 4.1.2 is an ana-

© 2001 by Chapman & Hall/CRC

479

Appendix B

logue of Jacod and Shiryaev [67, Theorem VIII.2.30]. Lemma 4.2.6 is a variation on the theme of Polya's theorem, see, e.g., Liptser and Shiryaev [79, Problem 5.3.2]. The proof of Theorem 4.2.11 uses the method of the proof of Theorem 5.4.1 in Liptser and Shiryaev [79]. In particular the fundamental decomposition (LS ) originates in Lemma 5.4.1 there. In the proof of Theorem 4.2.11 condition (^ ) has been used only while proving convergence ). Since ) clearly holds under the condition

(log ^)

1 X f (r x) s r 0<st

Zt

ln 1 + f (rx) s

f (x) Ls ln 1 + f (x) Ls ) ds

0

1=r P

! 0

as 2 ; t 2 U; f 2 Cb ;

and, as the proof shows, condition (^ ) implies condition (log ^), it follows that the theorem holds when condition (^ ) is replaced by condition (log ^). One can show that these two conditions are actually equivalent. Corollary 4.3.5 and its proof are analogous to Proposition VIII.3.40 in Jacod and Shiryaev [67]. Condition (L2 ) and Lemma 4.3.9 originate from Djellout [39]. The argument of the proof of Theorem 4.4.6 follows Djellout [39]. LDPs for partial-sum processes have been studied in Varadhan [127], Borovkov [13], and Mogulskii [90, 91]. Conditions (4.4.10) have been found by Ibragimov and Linnik [61, Theorem 13.1.1] who studied exact asymptotics in the nonfunctional case and showed that the moment condition in (4.4.10) is in a certain sense necessary, see [61, Theorem 13.1.2]. Mogulskii [90, Theorem 1] has established a functional LDP under (4.4.10) for the setting of Theorem 4.4.6. Logarithmic asymptotics for sums of i.i.d.r.v. under (4.4.9) can also be derived from the estimates of the convergence rate in the CLT in Ibragimov and Linnik [61, Chapter 3] and Petrov [103, Chapter 5]. Example 4.4.13 is motivated by Theorem VIII.3.43 in Jacod and Shiryaev [67].

© 2001 by Chapman & Hall/CRC

480

Notes and remarks

Chapter 5 The results are based on Puhalskii [111, 112] (note, however, that our condition (NE ) is somewhat stronger than in Puhalskii [111], so a correction needs to be made in that paper). The majoration and continuity conditions are similar to those used in weak convergence theory, see, e.g., Jacod and Shiryaev [67, VI.3.34]. Theorems 5.3.3, 5.3.4 and 5.3.5 are analogues of respective Theorems 8.2.1, 8.4.2 and 8.3.1 in Liptser and Shiryaev [79]. For the case where At is continuous Theorem 5.3.7 is an analogue of Theorem 8.4.1 in Liptser and Shiryaev [79]. Theorem 5.4.3 is an analogue of Theorems IX.4.8 and IX.4.15 in Jacod and Shiryaev [67]. The setting of Liptser and Puhalskii [78], who have considered large deviations for quasi-continuous processes with the Cramer condition on the jumps and linearly growing coeÆcients, is a special case of the setting of Theorem 5.3.3 for the case where conditions (A)loc + (a)loc are checked by checking (Ie )loc . In particular, Theorem 2.2 there follows by Theorem 5.3.7. As for Theorem 2.1 of Liptser and Puhalskii [78], it is not quite clear if it follows from our results since both results require some implicit conditions on the rate functions, which are diÆcult to compare. However, for the case of explicit suÆcient conditions given by Theorem 9.1 in Liptser and Puhalskii [78], one can obtain Theorem 2.1 there as a corollary of Theorems 5.3.3 and 2.8.33. The Markov setting has been analysed in Freidlin and Wentzell [51], Wentzell [134], Dupuis and Ellis [43], Feng [49], Feng and Kurtz [50], see also Azencott [7], Baldi [9], Baldi and Chaleyat-Maurel [10], Friedman [52], Makhno [83], Micami [89], Narita [93], Remillard and Dawson [116]. Diusions with dependence on the past have been considered by Cutland [27]. Markov processes with discontinuous statistics have been studied by Blinovskii and Dobrushin [12], Dupuis, Ellis and Weiss [44], Dupuis and Ellis [42], Korostelev and Leonov [74], Ignatyuk, Malyshev and Scherbakov [65]. Theorem 5.4.1, taken from Puhalskii [111], is a large-deviation analogue of a result that derives convergence of Markov semigroups from convergence of associated generators, see, e.g., Ethier and Kurtz [48]. Feng [49] has obtained a similar result for Markov processes with values in general metric spaces by the methods of nonlinear semigroup convergence, see Feng and Kurtz [50] for further develop-

© 2001 by Chapman & Hall/CRC

Appendix B

481

ments in this direction. Theorem 5.4.4 relaxes the assumptions of Theorems 4.4.1, 4.4.2, 4:4:20 , and 4.4.3 in Wentzell [134], and Theorem 5.4.2 in combination with Theorem 2.8.33 relaxes the assumptions of Wentzell's Theorems 4.3.1, 3.2.3 and 3.2.3'. The main improvements are that we do not require that either bs (u), or cs (u), or s(dx; u), or gs (; u) be either continuous in the time variable or bounded, and the associated convergences may take place only locally uniformly, not necessarily uniformly on the entire space. Theorem 5.4.8 has been motivated by Theorem 4.5.2 in Wentzell [134]. A similar result can be proved for discrete-time processes. Example 2 is prompted by Problem 8.4.1 in Liptser and Shiryaev [79]. The proof of (5.4.11) is based on Liptser's idea (private communication).

Section 6.1 The results are based on Puhalskii [114], where more detail is given. They complement results on diusion approximation for queues in Kingman [72], Prohorov [105], Iglehart and Whitt [64], Borovkov [15], and Reiman [115]. Theorem 6.1.7 is an analogue of diusion approximation results in Prohorov [105] and Borovkov [15]; the proof borrows ideas used in these proofs. The results of Subsection 6.1.2 are inspired by Reiman [115].

Section 6.2 Theorem 6.2.3 complements diusion approximation results by Iglehart [62, 63], Borovkov [14, 16], Hal n and Whitt [57], and Whitt [136]. For other results on large deviation asymptotics for many server queues see Glynn [55] and Zajic [139].

© 2001 by Chapman & Hall/CRC

Bibliography [1] M. Akian. Densities of idempotent measures and large deviations. Trans. Am. Math. Soc., 351(11):4515{4543, 1999. [2] M. Akian, J.-P. Quadrat, and M. Viot. Bellman processes. In 11th Conference on Analysis and Optimization of Systems: Discrete Event Systems, volume 199 of Lecture Notes in Control and Information Sciences. Springer, 1994. [3] M. Akian, J.-P. Quadrat, and M. Viot. Duality between probability and optimization. In J. Gunawerdena, editor, Idempotency. Cambridge University Press, 1998. [4] D. Aldous. Stopping times and tightness. Ann. Prob., 6:335{ 340, 1978. [5] J.-P. Aubin and A. Cellina. Dierential Inclusions. Springer, 1984. [6] J.-P. Aubin and I. Ekeland. Applied Nonlinear Analysis. Wiley, 1984. [7] R. Azencott. Grandes deviations et applications. In Lecture Notes Math., volume 774, pages 1{176. Springer, 1980. [8] F. Baccelli, G. Cohen, G.J. Olsder, and J.P. Quadrat. Synchronisation and Linearity: an Algebra for Discrete Event Systems. Wiley, 1992. [9] P. Baldi. Large deviations for diusion processes with homogenization and applications. Ann. Prob., 19(2):509{524, 1991. [10] P. Baldi and M. Chaleyat-Maurel. An extension of VentcelFreidlin estimates. In Lecture Notes Math., volume 1316, pages 305{327. Springer, 1988. 483 © 2001 by Chapman & Hall/CRC

484

Bibliography

[11] P. Billingsley. Convergence of Probability Measures. Wiley, 1968. [12] V.M. Blinovskii and R.L. Dobrushin. Process level large deviations for a class of piecewise homogeneous random walks. In The Dynkin Festschrift: Markov Processes and their Applications, pages 1{59. Birkhauser, 1994. [13] A.A. Borovkov. Boundary-value problems for random walks and large deviations in function spaces. Th. Prob. Appl., 12(4):575{595, 1967. [14] A.A. Borovkov. On limit laws for service processes in multi{ channel systems. Siberian Math. J., 8:746{762, 1967 (in Russian). [15] A.A. Borovkov. Stochastic Processes in Queueing Theory. Nauka, 1972 (in Russian, English translation: Springer, 1976). [16] A.A. Borovkov. Asymptotic Methods in Queueing Theory. Nauka, 1980 (in Russian, English translation: Wiley, 1984). [17] V.V. Breyer and O.V. Gulinsky. Large deviations in in nite dimensional vector spaces. Preprint MIPT 96-5, Moscow Institute of Physics and Technology, 1996 (in Russian). [18] W. Bryc. Large deviations by the asymptotic value method. In M. Pinsky, editor, Diusion processes and related problems in analysis, pages 447{472. Birkhauser, 1990. [19] J.A. Bucklew. Large Deviations Techniques in Decision, Simulation, and Estimation. Wiley, 1990. [20] S.L. Campbell and C.D. Meyer, Jr. Generalized Inverses of Linear Transformations. Pitman, 1979. [21] W. Cegla and M. Klimek. Criterion for the large deviation principle. Proc. Roy. Irish Acad., 90A(1):5{10, 1990. [22] N.R. Chaganty. Large deviations for joint distributions and statistical applications. Technical Report TR93-2, Department of Mathematics and Statistics, Old Dominion University, Norfolk, Va, 1993.

© 2001 by Chapman & Hall/CRC

Bibliography

485

[23] H. Chen and W. Whitt. Diusion approximations for open queueing networks with service interruptions. Queueing Systems, 13:335{359, 1993. [24] G. Choquet. Theory of capacities. Ann. Inst. Fourier, 5:131{ 295, 1955. [25] F.H. Clarke. Optimization and Nonsmooth Analysis. Wiley, 1983. [26] E.A. Coddington and N. Levinson. Theory of Ordinary Dierential Equations. McGraw-Hill, 1955. [27] N.J. Cutland. An extension of the Ventcel-Freidlin large deviation principle. Stochastics, 24:121{149, 1988. [28] D.A. Dawson and J. Gartner. Large deviations from the McKean-Vlasov limit for weakly interacting diusions. Stochastics, 20:247{308, 1987. [29] A. de Acosta. Large deviations for vector-valued Levy processes. Stoch. Proc. Appl., 51:75{115, 1994. [30] A. de Acosta. Exponential tightness and projective systems in large deviation theory. In Festschrift for Lucien Le Cam, pages 143{156. Springer, 1997. [31] G. de Cooman and E. Kerre. Possibility and necessity integrals. Fuzzy Sets and Systems, 77:207{227, 1996. [32] G. de Cooman and E.E. Kerre. Ample elds. Simon Stevin, 67:235{244, 1993. [33] G. de Cooman, E.E. Kerre, and F.R. Vanmassenhove. Possibility theory: an integral theoretic approach. Fuzzy Sets and Systems, 46:287{299, 1992. [34] C. Dellacherie. Capacites et Processus Stochastiques. Springer, 1972. [35] A. Dembo and O. Zeitouni. Large Deviations Techniques and Applications. Springer, second edition, 1998.

© 2001 by Chapman & Hall/CRC

486

Bibliography

[36] J.D. Deuschel and D.W. Stroock. Large Deviations. Academic Press, 1989. [37] I.H. Dinwoodie. Identifying a large deviation rate function. Ann. Prob., 21(1):216{231, 1993. [38] I.H. Dinwoodie and S.L. Zabell. Large deviations for sequences of mixtures. In J.K. Ghosh et al., editor, Statistics and Probability. A Bahadur Festschrift. Wiley, 1993. [39] H. Djellout. Moderate deviations for martingale dierences, 2000 (submitted for publication). [40] D. Dubois and H. Prade. Possibility Theory. Plenum Press, 1988. [41] R.M. Dudley. Real Analysis and Probability. Wadsworth & Brooks/Cole, 1989. [42] P. Dupuis and R. Ellis. Large deviations for Markov processes with discontinuous statistics. II. Prob. Th. Rel. Fields, 91:153{ 194, 1992. [43] P. Dupuis and R. Ellis. A Weak Convergence Approach to the Theory of Large Deviations. Wiley, 1997. [44] P. Dupuis, R.S. Ellis, and A. Weiss. Large deviations for Markov processes with discontinuous statistics. I. Ann. Prob., 19:1280{1297, 1991. [45] R.J. Elliott. Stochastic Calculus and Applications. Springer, 1982. [46] R.S. Ellis. Large deviations for a general class of random vectors. Ann. Prob., 12(1):1{12, 1984. [47] R. Engelking. General Topology. PWN, 1977. [48] S.N. Ethier and T.G. Kurtz. Markov Processes. Characterization and Convergence. Wiley, 1986. [49] J. Feng. Martingale problems for large deviations of Markov processes. Stoch. Proc. Appl., 81(2):165{216, 1999.

© 2001 by Chapman & Hall/CRC

Bibliography

487

[50] J. Feng and T.G. Kurtz. Large deviations for stochastic processes (preliminary manuscript), 2000. [51] M.I. Freidlin and A.D. Wentzell. Random Perturbations of Dynamical Systems. Nauka, 1979 (in Russian, English translation: Springer, 1984). [52] A. Friedman. Stochastic Dierential Equations and Applications, volume 2. Academic Press, 1976. [53] J. Gartner. On large deviations from the invariant measure. Th. Prob. Appl., 22(1):24{39, 1977. [54] I.I. Gihman and A.V. Skorohod. Stochastic Dierential Equations and their Applications. Naukova Dumka, 1982 (in Russian). [55] P.W. Glynn. Large deviations for the in nite server queue in heavy traÆc. In Stochastic networks, volume 71 of IMA Vol. Math. Appl., pages 387{394. Springer, 1995. [56] C. Guo, D. Zhang, and C. Wu. Generalized fuzzy integrals of fuzzy-valued functions. Fuzzy Sets and Systems, 97:123{128, 1998. [57] S. Hal n and W. Whitt. Heavy-traÆc limits for queues with many exponential servers. Oper. Res., 29:567{588, 1981. [58] P.R. Halmos. Measure Theory. Springer, 1974. [59] J.M. Harrison and M.I. Reiman. Re ected Brownian motion on an orthant. Ann. Prob., 9:302{308, 1981. [60] P. Hartman. Ordinary Dierential Equations. Wiley, 1964. [61] I.A. Ibragimov and Yu.V. Linnik. Independent and Stationary Related Random Variables. Nauka, 1965 (in Russian). [62] D.L. Iglehart. Limit diusion approximations for the many server queue and the repairman problem. J. Appl. Prob., 2:429{ 441, 1965. [63] D.L. Iglehart. Weak convergence of compound stochastic processes. Stoch. Proc. Appl, 1:11{31, 1973.

© 2001 by Chapman & Hall/CRC

488

Bibliography

[64] D.L. Iglehart and W. Whitt. Multiple channel queues in heavy traÆc, I and II. Adv. Appl. Prob., 2:150{177 and 355{369, 1970. [65] I.A. Ignatyuk, V. Malyshev, and V.V. Scherbakov. Boundary eects in large deviation problems. Russ. Math. Surv., 49(2):41{99, 1994. [66] N. Ikeda and S. Watanabe. Stochastic Dierential Equations and Diusion Processes. North Holland, second edition, 1989. [67] J. Jacod and A.N. Shiryaev. Limit Theorems for Stochastic Processes. Springer, 1987. [68] H. Janssen, G. de Cooman, and E.E. Kerre. A DaniellKolmogorov theorem for supremum preserving upper probabilities. Fuzzy Sets and Systems, 102(3):429{444, 1999. [69] T. Jiang and G.L. O'Brien. The metric of large deviation convergence. J. Theoret. Prob., 13(3):805{823, 2000. [70] L.V. Kantorovich and G.P. Akilov. Functional Analysis in Normed Spaces. Pergamon Press, 1964. Original edition: Funktsional'nyi analiz v normirovannikh prostranstvakh, Fizmatgiz (in Russian). [71] J.L. Kelley. General Topology. Springer, 1975. [72] J.F.C. Kingman. On queues in heavy traÆc. J. Roy. Statist. Soc., B24:383{392, 1962. [73] V.N. Kolokoltsov and V.P. Maslov. Idempotent Analysis and Its Applications. Kluwer, 1997. [74] A.P. Korostelev and S.L. Leonov. An action functional for a diusion process with discontinuous drift. Th. Prob. Appl., 37(3):543{550, 1992 (in Russian: Teor. Veroyatn. i Primen., 1992, v. 37, no. 3, pp. 570-576). [75] M.A. Krasnosel'skii and Ya.B. Rutickii. Convex Functions and Orlicz Spaces. Noordho, 1961. [76] K. Kuratowski and A. Mostowski. Set Theory. North-Holland{ PWN, 1967.

© 2001 by Chapman & Hall/CRC

Bibliography

489

[77] T. Lindvall. Weak convergence of probability measures and random functions in the function space D[0; 1). J. Appl. Prob., 10:109{121, 1973. [78] R.Sh. Liptser and A. Puhalskii. Limit theorems on large deviations for semimartingales. Stoch. Stoch. Rep., 38:201{249, 1992. [79] R.Sh. Liptser and A.N. Shiryaev. Theory of Martingales. Kluwer, 1989. [80] G.L. Litvinov, V.P. Maslov, and G.B. Shpiz. Idempotent functional analysis. An algebraic approach. Technical report, International Centre "Sophus Lie", 1998 (in Russian). [81] G.L. Litvinov, V.P. Maslov, and G.B. Shpiz. Linear functionals on idempotent spaces. An algebraic approach. Dokl. Akad. Nauk, 363(3):298{300, 1998 (in Russian). [82] J. Lynch and J. Sethuraman. Large deviations for processes with independent increments. Ann. Prob., 15(2):610{627, 1987. [83] S.Ya. Makhno. A large deviation theorem for a class of diusion processes. Teor. Veroyatnost. i Primenen., 39(3):554{566, 1994 (English translation: Th. Prob. Appl. 39(1994), no. 3, 437-447 (1995)). [84] V. Maslov. Methode Operatorielles. Mir, 1987 (in French). [85] V.P. Maslov. Asymptotic Methods of Solving PseudoDierential Equations. Nauka, 1987 (in Russian). [86] R. Mesiar. Possibility measures, integration and fuzzy possibility measures. Fuzzy Sets and Systems, 92:191{196, 1997. [87] R. Mesiar and E. Pap. Idempotent integral as limit of gintegrals. Fuzzy Sets and Systems, 102:385{392, 1999. [88] P.A. Meyer. Probability and Potentials. Blaisdell, 1966. [89] T. Micami. Some generalizations of Wentzell's lower estimates on large deviations. Stochastics, 24(4):269{284, 1988.

© 2001 by Chapman & Hall/CRC

490

Bibliography

[90] A.A. Mogulskii. Large deviations for trajectories of multidimensional random walks. Theory Prob. Appl., 21(2):300{315, 1976. [91] A.A. Mogulskii. Large deviations for processes with independent increments. Ann. Prob., 21(1):202{215, 1993. [92] U. Mosco. On the continuity of the Young-Fenchel transform. J. Math. Anal. Appl., 35(3):518{535, 1971. [93] K. Narita. Large deviation principle for diusion processes. Tsukuba J. Math., 12(1):211{229, 1988. [94] J. Neveu. Bases Mathematiques du Calcul des Probabilites. Masson et Cie, 1964 (in French). [95] T. Norberg. Random capacities and their distributions. Prob. Th. Rel. Fields, 73(2):281{297, 1986. [96] G.L. O'Brien. Sequences of capacities, with connections to large deviation theory. J. Theoret. Probab., 9(1):19{35, 1995. [97] G.L. O'Brien and W. Vervaat. Capacities, large deviations and loglog laws. In S. Cambanis, G. Samorodnitsky, and M. Taqqu, editors, Stable Processes and Related Topics, volume 25 of Progress in Probability, pages 43{83. Birkhauser, 1991. [98] G.L. O'Brien and W. Vervaat. Compactness in the theory of large deviations. Stoch. Proc. Appl., 57:1{10, 1995. [99] G.L. O'Brien and S. Watson. Relative compactness for capacities, measures, upper semicontinuous functions and closed sets. J. Theoret. Prob., 11(3):577{588, 1998. [100] B. ksendal. Stochastic Dierential Equations. Springer, 1998. [101] E. Pap. Null-Additive Set Functions. Kluwer, 1995. [102] K.R. Parthasarathy. Probability Measures on Metric Spaces. Academic Press, 1967. [103] V.V. Petrov. Limit Theorems for Sums of Independent Random Variables. Nauka, second edition, 1987 (in Russian).

© 2001 by Chapman & Hall/CRC

Bibliography

491

[104] Yu. V. Prohorov. Convergence of stochastic processes and limit theorems in probability theory. Th. Prob. Appl., 1:157{214, 1956. [105] Yu.V. Prohorov. Transient phenomena in queueing processes. Lit. Mat. Rink., 3:199{206, 1963 (in Russian). [106] A. Puhalskii. On functional principle of large deviations. In V. Sazonov and T. Shervashidze, editors, New Trends in Probability and Statistics, volume 1, pages 198{218. VSP/Moks'las, 1991. [107] A. Puhalskii. On the theory of large deviations. Th. Prob. Appl., 38:490{497, 1993. [108] A. Puhalskii. Large deviations of semimartingales via convergence of the predictable characteristics. Stoch. Stoch. Rep., 49:27{85, 1994. [109] A. Puhalskii. The method of stochastic exponentials for large deviations. Stoch. Proc. Appl., 54:45{70, 1994. [110] A. Puhalskii. Large deviation analysis of the single server queue. Queueing Systems, 21:5{66, 1995. [111] A. Puhalskii. Large deviations of semimartingales: a maxingale problem approach. I. Limits as solutions to a maxingale problem. Stoch. Stoch. Rep., 61:141{243, 1997. [112] A. Puhalskii. Large deviations of semimartingales: a maxingale problem approach. II. Uniqueness for the maxingale problem. Applications. Stoch. Stoch. Rep., 68:65{143, 1999. [113] A. Puhalskii and W. Whitt. Functional large deviation principles for rst-passage-time processes. Ann. Appl. Prob., 7(2):362{381, 1997. [114] A.A. Puhalskii. Moderate deviations for queues in critical loading. Queueing Systems, 31:359{392, 1999. [115] M.I. Reiman. Open queueing networks in heavy traÆc. Math. Oper. Res., 9:441{458, 1984.

© 2001 by Chapman & Hall/CRC

492

Bibliography

[116] B. Remillard and D.A. Dawson. Laws of the iterated logarithm and large deviations for a class of diusion processes. Can. J. Statist., 17(4):349{376, 1989. [117] R.T. Rockafellar. Convex Analysis. Princeton University Press, 1970. [118] L. Schwartz. Radon Measures on Arbitrary Topological Spaces and Cylindrical Measures. Oxford University Press, 1973. [119] N. Shilkret. Maxitive measure and integration. Indag. Math., 33:109{116, 1971. [120] A.N. Shiryaev. Probability, volume 95 of Graduate Texts in Mathematics. Springer, second edition, 1996 (Translated from the rst (1980) Russian edition by R. P. Boas). [121] A.V. Skorohod. Limit theorems for stochastic processes. Th. Prob. Appl., 1:261{292, 1956. [122] D.W. Stroock. An Introduction to the Theory of Large Deviations. Springer, 1984. [123] D.W. Stroock and S.R.S. Varadhan. Multidimensional Diusion Processes. Springer, 1979. [124] F. Topse. Compactness in spaces of measures. Studia Mathematica, 36:195{221, 1970. [125] F. Topse. Topology and Measure, volume 133 of Lecture Notes in Mathematics. Springer, 1970. [126] N.N. Vakhania, V.I. Tarieladze, and S.A. Chobanyan. Probability Distributions on Banach Spaces. Nauka, 1985 (in Russian, English translation: Reidel, 1987). [127] S.R.S. Varadhan. Asymptotic probabilities and dierential equations. Comm. Pure Appl. Math., 19(3):261{286, 1966. [128] S.R.S. Varadhan. Large Deviations and Applications. SIAM, 1984. [129] W. Vervaat. Narrow and vague convergence of set functions. Statist. & Prob. Lett., 6(5):295{298, 1988.

© 2001 by Chapman & Hall/CRC

Bibliography

493

[130] W. Vervaat. Random uppersemicontinuous functions and extremal processes. Technical Report MS-8801, Center for Math. and Comp. Sci., Amsterdam, 1988. [131] K. von Leichtweiss. Konvexe Mengen. VEB Deutscher Verlag der Wissenschaften, 1980. [132] P.-Z. Wang. Fuzzy contactability and fuzzy variables. Fuzzy Sets and Systems, 8:81{92, 1982. [133] Z. Wang and G.J. Klir. Fuzzy Measure Theory. Plenum Press, 1992. [134] A.D. Wentzell. Limit Theorems on Large Deviations for Markov Stochastic Processes. Nauka, 1986 (in Russian, English translation: Kluwer, 1990). [135] W. Whitt. Some useful functions for functional limit theorems. Math. Oper. Res., 5(1):67{85, 1980. [136] W. Whitt. On the heavy-traÆc limit theorem for GI=G=1 queues. Adv. Appl. Prob., 14:171{190, 1982. [137] C. Wu, S. Wang, and M. Ma. Generalized fuzzy integrals: Part I. Fundamental concepts. Fuzzy Sets and Systems, 57:219{226, 1993. [138] S.L. Zabell. Mosco convergence in locally convex spaces. J. Function. Anal., 110(1):226{246, 1992. [139] T. Zajic. Rough asymptotics for tandem non-homogeneous M=G=1 queues via Poissonized empirical processes. Queueing Systems, 29(2-4):161{174, 1998.

© 2001 by Chapman & Hall/CRC

LARGE DEVIATIONS AND IDEMPOTENT PROBABILITY

© 2001 by Chapman & Hall/CRC

119

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics Main Editors H. Brezis, Université de Paris R.G. Douglas, Texas A&M University A. Jeffrey, University of Newcastle upon Tyne (Founding Editor)

Editorial Board H. Amann, University of Zürich R. Aris, University of Minnesota G.I. Barenblatt, University of Cambridge H. Begehr, Freie Universität Berlin P. Bullen, University of British Columbia R.J. Elliott, University of Alberta R.P. Gilbert, University of Delaware R. Glowinski, University of Houston D. Jerison, Massachusetts Institute of Technology K. Kirchgässner, Universität Stuttgart B. Lawson, State University of New York B. Moodie, University of Alberta S. Mori, Kyoto University L.E. Payne, Cornell University D.B. Pearson, University of Hull I. Raeburn, University of Newcastle G.F. Roach, University of Strathclyde I. Stakgold, University of Delaware W.A. Strauss, Brown University J. van der Hoek, University of Adelaide

© 2001 by Chapman & Hall/CRC

CHAPMAN & HALL/CRC Monographs and Surveys in Pure and Applied Mathematics

LARGE DEVIATIONS AND IDEMPOTENT PROBABILITY

ANATOLII PUHALSKII

CHAPMAN & HALL/CRC Boca Raton London New York Washington, D.C.

© 2001 by Chapman & Hall/CRC

119

C0198_Disclaimer Page 1 Friday, March 30, 2001 2:22 PM

Library of Congress Cataloging-in-Publication Data Puhalskii, Anatolii. Large deviations and idempotent probability / Anatolii Puhalskii. p. cm. -- (Chapman & Hall/CRC monographs and surveys in pure and applied mathematices ; 119) Includes bibliographical references and index. ISBN 1-58488-198-4 (alk. paper) 1. Large deviations. 2. Probability measures. 3. Idempotents. I. Title. II. Series. QA273.67 .P83 2000 519.5′34--dc21

00-065883

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage or retrieval system, without prior permission in writing from the publisher. The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC for such copying. Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation, without intent to infringe.

Visit the CRC Press Web site at www.crcpress.com © 2001 by Chapman & Hall/CRC No claim to original U.S. Government works International Standard Book Number 1-58488-198-4 Library of Congress Card Number 00-065883 Printed in the United States of America 1 2 3 4 5 6 7 8 9 0 Printed on acid-free paper

© 2001 by Chapman & Hall/CRC

To my parents

© 2001 by Chapman & Hall/CRC

Contents Preface Basic notation

xi 1

I Idempotent Probability Theory

3

1 Idempotent probability measures 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9

Idempotent measures . . . . . . . . . . . . . Measurable functions . . . . . . . . . . . . . Modes of convergence . . . . . . . . . . . . Idempotent integration . . . . . . . . . . . . Product spaces . . . . . . . . . . . . . . . . Independence and conditioning . . . . . . . Idempotent measures on topological spaces Idempotent measures on projective limits . Topological spaces of idempotent probabilities . . . . . . . . . . . . . . . . . . 1.10 Derived weak convergence . . . . . . . . . . 1.11 Laplace-Fenchel transform . . . . . . . . . .

2 Maxingales 2.1 2.2 2.3 2.4

Idempotent stopping times . . . . . Idempotent processes . . . . . . . . . Exponential maxingales . . . . . . . Wiener and Poisson idempotent processes . . . . . . . . . . . . . . . 2.5 Idempotent stochastic integrals . . . 2.6 Idempotent Ito dierential equations vii

© 2001 by Chapman & Hall/CRC

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

5

5 14 17 20 32 36 51 58

. . . . . . 64 . . . . . . 79 . . . . . . 84

91

. . . . . . . . . . 91 . . . . . . . . . . 95 . . . . . . . . . . 104 . . . . . . . . . . 114 . . . . . . . . . . 124 . . . . . . . . . . 151

viii 2.7 Semimaxingales . . . . . . . . . . . . . . . . . . . . . 170 2.8 Maxingale problems . . . . . . . . . . . . . . . . . . . 202

II Large Deviation Convergence of Semimartingales

251

3 Large deviation convergence

253

4 The method of nite-dimensional distributions

289

3.1 Large deviation convergence in Tihonov spaces . . . . 253 3.2 Large deviation convergence in the Skorohod space . . . . . . . . . . . . . . . . . . . . 276 4.1 Convergence of stochastic exponentials 4.2 Convergence of characteristics . . . . . 4.2.1 The case of small jumps . . . . 4.2.2 The general case . . . . . . . . 4.3 Corollaries . . . . . . . . . . . . . . . . 4.4 Applications to partial-sum processes .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

5 The method of the maxingale problem

5.1 Convergence of stochastic exponentials . . . . . . . . . 5.1.1 Proofs . . . . . . . . . . . . . . . . . . . . . . . 5.2 Convergence of characteristics . . . . . . . . . . . . . . 5.2.1 Exponential tightness results . . . . . . . . . . 5.2.2 LD accumulation points as solutions to maxingale problems . . . . . . . . . . . . . . . . . . . 5.2.3 Proofs of the main results . . . . . . . . . . . . 5.3 Large deviation convergence results . . . . . . . . . . . 5.4 Large deviation convergence of Markov processes . . .

290 305 316 325 332 342

355

356 360 373 380 391 404 406 414

6 Large deviation convergence of queueing processes 433 6.1 Moderate deviations in queueing networks . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 Idempotent diusion approximation for single server queues . . . . . . . . . . . . . . . . . . . 6.1.2 Idempotent diusion approximation for queueing networks . . . . . . . . . . . . . . . . 6.2 Very large and moderate deviations for many server queues . . . . . . . . . . . . . . . . . . . . . . . . . . .

© 2001 by Chapman & Hall/CRC

433 433 442 450

ix

Appendix A Auxiliary lemmas Appendix B Notes and remarks Bibliography

© 2001 by Chapman & Hall/CRC

457 467 483

Preface This book has grown out of an approach to establishing the large deviation principle (LDP) for probability measures that originates from viewing the LDP as an analogue of weak convergence of probability measures and develops tools for proving it along the lines of the ones used in weak convergence theory. Let us recall that, given a Hausdor topological space E equipped with Borel -algebra B(E ), a function I : E ! [0; 1] such that the sets fz 2 E : I (z ) ag are compact for a 2 R+ , a net fP ; 2 g of probability measures on (E; B(E )), and a net of non-negative numbers fr ; 2 g such that r ! 1 as 2 , the net fP ; 2 g is said to obey the LDP with rate function I for scale r if 1 lim sup ln P (F ) inf I (z ) for F being a closed subset of E; z 2F 2 r 1 lim inf ln P (G) inf I (z ) for G being an open subset of E: z 2G 2 r The de nition being modelled after the de nition of weak convergence of probability measures, it is not surprising that there are similarities between methods of deriving the LDP and weak convergence, e.g., both theories make use of characteristic functionals, projective limit arguments, continuous mappings, characterisation of relative compactness in terms of certain tightness conditions, and others. Our purpose is to explore this analogy in more depth and systematically build on it for studying properties of the LDP. The rst important step is to recognise and treat the rate function as a limit case of the probability measure rather than merely as an asymptotic value. More precisely, we consider the set function (A) on E de ned by (A) = supz2A exp( I (z )) as an analogue and a limit of probabilities, so we call it \a deviability". We next look for properxi © 2001 by Chapman & Hall/CRC

xii ties of that are inherited from probabilities in the hope that this will help us to identify it, e.g., we are interested in an analogue of the martingale property. A distinctive feature of deviability is that it is \maxitive" in that (A [ B ) = (A) _ (B ). Maxitive set functions have been known as possibility measures in possibility theory and idempotent measures in idempotent measure theory (also referred to by the names \max-plus calculus" and \min-plus calculus"). We adopt the name \idempotent measure"; on the other hand, the notation , which is used not only for deviabilities but also for general idempotent measures \of mass 1", is borrowed from possibility theory. Developing \a stochastic calculus" for idempotent measures is the subject of part I of the book. We start with basic axioms, consider extension theorems, measurability issues, idempotent expectations and conditional idempotent expectations, topologies on spaces of idempotent measures, and other analogues of the constructions of probability theory. The axioms for an idempotent measure are mostly the same as the ones used in possibility theory and idempotent measure theory so we recover some of the results of these theories. Besides, we extensively analyse the -smoothness property of idempotent measures that requires a certain type of \continuity from above" and has been prompted by the fact that deviabilities are -smooth with respect to decreasing nets of closed sets. We also undertake a study in the spirit of the general theory of stochastic processes of idempotent analogues of stopping times, ltrations, stochastic processes, Ito dierential equations, martingales and semimartingales (which we call maxingales and semimaxingales, respectively), and martingale problems (referred to as maxingale problems). Being motivated by applications to large deviation theory, by no means do we consider analogues of all standard probability topics, the most notable omissions being analogues of the theory of limit theorems and theory of Markov processes. Our focus is on developing weak convergence theory for idempotent measures and those parts of \maxingale theory" that are instrumental in deriving large deviation limit theorems. Part II studies the large deviation setting. In order to emphasise the view of a deviability as a limit of probabilities, we refer to \the LDP for the P with rate function I " as \large deviation (LD) convergence of the P to ". Thus, the P are said to LD converge to

© 2001 by Chapman & Hall/CRC

xiii at rate r if lim sup P (F )1=r (F ) for F closed; (0.0.1) 2 lim inf P (G)1=r (G) for G open: (0.0.2) 2 In our study we actually use a dierent form of the de nition of LD convergence that states that the P LD converge to (at rate r ) if Z

lim 2

h(z )r dP (z )

E

1=r

= sup h(z )(z ) z 2E

(0.0.3)

for all R+ -valued bounded and continuous functions h on E . For general Hausdor spaces properties (0.0.1) and (0.0.2) are stronger than (0.0.3). One could draw the line by referring to (0.0.1) and (0.0.2) as narrow large deviation convergence and (0.0.3) as weak large deviation convergence. However, for Tihonov spaces (i.e., completely regular T1 -spaces), which we mostly consider and which seem to suÆce for concrete large deviation settings, the two forms are equivalent; so we refer to the property speci ed by (0.0.3) as large deviation convergence. The advantage of using de nition (0.0.3) is that many proofs can be signi cantly shortened (which fact is explained to some extent by both the limit and pre-limit objects being certain norms). We explore general properties of LD convergence in the form (0.0.3) in the rst section of part II, where our methods are similar to those of studying weak convergence of measures and idempotent measures. The rest of part II considers LD convergence of the distributions of semimartingales for the Skorohod topology on the space of right-continuous with left-hand limits Rd -valued functions on R+ . Here we are able to implement the approaches used for deriving convergence in distribution for semimartingales such as characterisations of limits in terms of their nite-dimensional distributions and as solutions to martingale problems. We interpret the limit deviabilities as distributions of idempotent processes and state the results in the form of LD convergence in distribution of semimartingales to semimaxingales. For example, we formulate the LDP for diusion processes with small diusion terms as LD convergence in distribution to an idempotent diusion. We give applications to LD convergence of Markov processes and processes arising in queueing systems. Our

© 2001 by Chapman & Hall/CRC

xiv results for queues are in the same theme as corresponding weak convergence results. As a byproduct, the results of part II show that possibility theory can be viewed as a large deviation limit of probability theory. The book concludes with two appendices. Appendix A proves certain auxiliary results invoked in the main body of the book. Appendix B contains additional comments on the results and bibliographical notes; the latter re ect the author's view of the related work and are by their very nature subjective, nor do we make any claim to completeness of the list of references.

© 2001 by Chapman & Hall/CRC

1

Basic notation R+ R+

a_b a+ a^b bac f Æg N Z+

xy jxj T kk 1=2 Ac AB P (A) Q(A) 1(A); 1A int A cl A B(E ) B(R+) B [0; t]

= [0; 1) = [0; 1] the maximum of a and b = a_0 the minimum of a and b the integer part of a the composition of functions f and g the set of natural numbers the set of non-negative integers the inner product of vectors x and y the Euclidean norm of a vector x the transpose of a matrix the operator norm of a matrix the pseudo inverse of a matrix the square root of a positive semi-de nite symmetric matrix the complement of a set A the collection of functions from a set B to a set A the power set of a set A the collection of nite subsets of a set A the indicator function of a set A the interior of a subset A of a topological space the closure of a subset A of a topological space the Borel -algebra on a topological space E the Lebesgue -algebra on R+ the Lebesgue -algebra on [0; t]

© 2001 by Chapman & Hall/CRC

Part I

Idempotent Probability Theory

3 © 2001 by Chapman & Hall/CRC

Chapter 1

Idempotent probability measures In this chapter we introduce idempotent analogues of basic objects of probability theory such as probability measures, random variables, expectations, conditional probabilities and expectations, and others, and study their properties.

1.1 Idempotent measures In this section we de ne the notion of an idempotent measure and obtain an extension theorem. We also introduce idempotent analogues of a measure space and probability space. Let be a set and E be a collection of subsets of , which contains ;. Let P ( ) denote the power set of . We reserve symbols and to denote directed sets, and J to denote arbitrary index sets.

De nition 1.1.1. A set function : P ( ) ! R + is an idempotent measure on if the following conditions hold:

(0) (;) = 0; (1) (A [ B ) = (A) _ (B ); (2) ([ A ) = sup (A ) for every increasing net fA ; 2 g of subsets of . If, in addition,

5 © 2001 by Chapman & Hall/CRC

6

Idempotent probability measures

() ( ) = 1, the idempotent measure is called an idempotent probability measure or idempotent probability, for short, and denoted by . If, in addition to (0), (1) and (2), (3) (\ F ) = inf (F ) for every decreasing net fF ; 2 g of elements of E , then we say that the idempotent measure is -smooth relative to E , or, for short, is an E -idempotent measure. Remark 1.1.2. Throughout, we use the terms \increasing" and \decreasing" as synonyms of \non-decreasing" and \non-increasing", respectively. Remark 1.1.3. Property (1) shows that is an increasing and subadditive set function in that (A) (B ) if A B and (A [ B ) (A) + (B ). The following characterisation of idempotent measures is a straightforward consequence of the de nition. Lemma 1.1.4. Conditions (1) and (2) are equivalent to the condition [ Aj = sup (Aj ) (1.1.1) j j 2J for every collection fAj ; j 2 J g of subsets of , which in turn is equivalent to the representation

(A) = sup (f!g); A : !2A

(1.1.2)

The function (f!g) is called the density of . We also refer to property (2) as -smoothness along increasing nets (it should not be confused with -smoothness, which concerns decreasing nets of elements of E ) and to property (1.1.1) as -maxitivity. For set functions that are only de ned on subsets of we use a similar terminology introduced by the following de nition. De nition 1.1.5. A set function : E ! R + is maxitive (respectively, -maxitive) on E if (A [ B ) = (A) _ (B ) for every A 2 E and B 2 E such that A [ B 2 E (respectively, [j 2J Aj = supj 2J (Aj ) for every collection of sets Aj 2 E ; j 2 J; such that [j2J Aj 2 E ).

© 2001 by Chapman & Hall/CRC

7

Idempotent measures

Given a collection E , we denote by Eu (respectively, Ei ) the collection of arbitrary unions (respectively, intersections) of elements of E . If nite unions (respectively, intersections) of sets from E belong to E , then we may and often do assume that the sets in an in nite union (respectively, intersection) of elements of E form an increasing (respectively, decreasing) net relative to a directed set. We also denote Eiu = (Ei )u and observe that it coincides with Eui = (Eu )i . The collection Eiu is clearly closed under the formation of arbitrary unions and intersections. We recall the following de nitions.

De nition 1.1.6. A collection E of subsets of is called a paving on

if it contains ; and is closed under the formation of nite unions and intersections. A collection E of subsets of is called a -system if it is closed under the formation of nite intersections. The next lemma shows that if is an idempotent measure that is -smooth relative to a paving E , then the values of on Eiu are uniquely speci ed by the values on E so that there is at most one extension of from E to Eiu .

Theorem 1.1.7. Let E be a -system containing ; and be an E idempotent measure. Then

(A) = sup (B ); A 2 Eiu ; B 2Ei : B A

(B ) = inf (F ); B 2 Ei : F 2E : F B

The proof follows by -maxitivity and -smoothness of , and the fact that each set in Ei is an intersection of a decreasing net of elements of E . The following simple fact is useful for extension theorems (see Theorem 1.1.9 below). We give the proof to show a typical argument.

Lemma 1.1.8. If an idempotent measure is -smooth relative to

a -system E containing ;, then it is -smooth relative to Ei . Proof. T Let fA ; 2 g be a decreasing net of elements of Ei , i.e., A = 2 F , where F 2 E . Let be the collection of nite sequences Æ = f( i1 i1 ); ( i2 i2 ); : : : ; ( ik ik )g, where ij 2 , i1 i2 : : : ik and l 2 l for l = i1 ; i2 ; : : : ; ik . We say

© 2001 by Chapman & Hall/CRC

8

Idempotent probability measures

that Æ Æ0 if all the pairs ( ) that appear in Æ are also conT 0 tained in Æ . For Æ 2 , let BÆ = ( )2Æ F . Then is a directed set and fBÆ ; Æ 2 g is a decreasing net. Also BÆ 2 E and T T A = 2 Æ2 BÆ ; so, since is an E -idempotent measure, \

2

A

= inf (BÆ ): Æ2

(1.1.3)

For arbitrary Æ = f( i1 i1 ); ( i2 i2 ); : : : ; ( ik ik )g, let Æ ij ; j = 1; : : : ; k. Then, since fA ; 2 g is a decreasing net, it follows that BÆ A Æ ; hence,

(BÆ ) A

Æ

inf (A ) = inf (A ): 2 Æ

Thus, in view of (1.1.3),

\

2

A

inf (A ): 2

We consider now the issue of extending set functions to idempotent measures. Theorem 1.1.9. Let E be a paving on . Let be an R + -valued maxitive function on E such that (;) = 0. 1. The set function can be extended to an idempotent measure on if and only if it is -smooth along increasing nets, i.e., for every increasing net fF g of elements of E whose union belongs to E we have

[

F = sup (F ):

The extension is uniquely speci ed on Eu .

2. The set function can be extended to an E -idempotent measure if and only if the following condition holds.

(S ) If fF1; g and fF2; g are respective increasing and decreasing nets of elements of E such that [

F1;

© 2001 by Chapman & Hall/CRC

\

F2; ;

Idempotent measures

9

then

sup (F1; ) inf (F2; ):

The idempotent measure is then also -smooth relative to Ei and is uniquely speci ed on Eiu . Proof. We rst consider part 1. Necessity of the condition is obvious. We prove suÆciency. We rst note that in view of maxitivity of and the fact that E is closed under the formation of nite unions the condition of -smoothness along increasing nets implies -maxitivity of on E . For ! 2 , let (f!g) = inf (F ); (1.1.4) F 2E : !2F

and for A let (A) = sup (f!g): (1.1.5) !2A Clearly, (A) is an idempotent measure. We prove that agrees with on E . Let F 2 E . By (1.1.4) (f!g) (F ) if ! 2 F , so by (1.1.5) (F ) (F ). Conversely, given " > 0, let F ! 2 E for ! 2 F be such that ! 2 F ! and (f!g) S(F ! ) ". Then by (1.1.5) (F ) sup!2F (F ! ) . Since F = !2F (F \ F ! ), where F \ F ! 2 E by the fact that E is closed under the formation of nite intersections, and is -maxitive and increasing on E , (F ) = sup (F \ F ! ) sup (F ! ) (F )+ ": !2F !2F Part 1 is proved. We prove part 2. It is obvious that if there exists an idempotent measure , which is -smooth relative to E and coincides with on E , then condition (S ) holds. For the converse, we note that condition (S ) implies the condition of -smoothness of relative to increasing nets of elements of E in part 1. Therefore, by part 1 the set function de ned by (1.1.4) and (1.1.5) is an idempotent measure, which extends . We prove that is an Ei -idempotent measure. Note that since is maxitive on E and E is closed under the formation of nite unions, condition (S ) extends to the case where fF1; g is an

© 2001 by Chapman & Hall/CRC

10

Idempotent probability measures

arbitrary collection of elements of E . Next, by Lemma 1.1.8 it suÆces to check (3) for decreasing nets of elements of E . Let F # F , where F 2 E . Given " > 0, we choose for every ! 2 F sets F ! 2 E as in the proof of suÆciency in part 1. Since [!2F F ! \ F and condition (S ) extends to arbitrary collections fF1;j g of elements of E , we conclude that sup!2F ! (F ! ) inf (F ) so that (F ) sup!2F (F ! ) " inf (F ) ", which completes the proof. The fact that is unique on Eiu follows by Theorem 1.1.7.

Remark 1.1.10. If E is a ring, i.e., closed under the formation of dierences, then condition (S ) is equivalent to continuity of at 0: if F # ;, then (F ) # 0. Theorem 1.1.9 is then an analogue of Caratheodory's theorem, see, e.g., Halmos [58].

Remark 1.1.11. Wang and Klir [133, Theorem 4.9] prove an ex-

tension theorem in the theme of part 1 for the case where the collection E is not necessarily closed under the formation of nite unions and intersections. Then the requirements on of maxitivity and -smoothness along increasing nets are replaced by the following P consistency condition: if a collection fFj g of elements of E and F 2 E are such that F [j Fj , then (F ) supj (Fj ). Similarly, part 2 admits a version for collections E that have the only property of including the empty set. Condition (S ) then has to be replaced by the following condition: (S 0 ) If fF1;j g is a collection of elements of E and fF2; g is a decreasing net of elements of E such that [

j

F1;j

\

F2; ;

then

sup (F1;j ) inf (F2; ): j

Condition (S 0 ) is necessary and suÆcient for to be extended to an E -idempotent measure. If E is a -system, then by Lemma 1.1.8 the extension is -smooth relative to Ei . We note also that condition (S 0 ) implies the P-consistency condition.

© 2001 by Chapman & Hall/CRC

11

Idempotent measures

Finally, if in part 2 we only omit the requirement that E be a -system, then the extension also exists and is an E -idempotent measure.

We will be interested in more special collections of subsets of

than pavings and -systems. The following notion plays a central part in our analysis below.

De nition 1.1.12. A collection A of subsets of is a -algebra if it contains ; and is closed under the formation of complements and arbitrary unions. The elements of A are referred to as A-measurable subsets of .

The power set P ( ) is obviously a -algebra, we refer to it as the discrete -algebra.

De nition 1.1.13. A collection E of subsets of is said to be atomic 0

if it has a subcollection E = fA g, consisting of non-empty subsets of , such that either A \ A0 = ; or A = A0 for every and 0 , and F 2 E if and only if F = [A , where the union is taken over A F; A 2 E 0 . The elements of E 0 are called the atoms of E .

The structure of -algebras is revealed by the next theorem, which follows from the de nition.

Theorem 1.1.14.

A collection A, which contains -algebra if and only if it is atomic.

;

and , is a

We denote as [!]A the atom of a -algebra A that contains ! 2 . We note that A 2 A if and only if A = [!2A [!]A , where an empty union is assumed to be empty. Remark 1.1.15. The relation R on de ned by (!; !0 ) 2 R if !0 and ! belong to the same atom of a -algebra A is obviously an A ! if we want equivalence relation. We denote this by !0 ! (or !0 to emphasise the -algebra to which the equivalence relation refers). Note that !0 ! if and only if !0 2 [!]A if and only if ! 2 [!0 ]A if and only if [!]A = [!0 ]A . The following simple observation is frequently used below.

Corollary 1.1.16. A set A is an element of a -algebra A on

if and only if [!]A A for every ! 2 A.

© 2001 by Chapman & Hall/CRC

12

Idempotent probability measures

De nition 1.1.17.0 We say that a -algebra A0 is a sub- -algebra of a -algebra A if A

A.

Lemma 1.1.18. A -algebra0 A0 is a sub- -algebra of a -algebra A if and only if the atoms of A are unions of the atoms of A.

We refer to the smallest -algebra containing a collection E as the -algebra generated by E and denote it as (E ). It is obviously unambiguously de ned.

De nition 1.1.19. We say that a collection E of subsets of is a c semi- -algebra if it includes ;, is a -system, and F F 2 E.

2 Eiu for every

The structure of semi- -algebras is similar to that of -algebras.

Lemma 1.1.20. A -system E , which includes ;, is a semi- -algebra

if and only if Ei is atomic and the union of the atoms of Ei equals . Proof. It is obvious that if Ei is atomic and its atoms make up , then E is a semi- -algebra. For the converse, given ! 2 , we take as the atom about ! the intersection of all elements of E that contain !.

The preceding proof also proves the following lemma.

Lemma 1.1.21. If E is a semi- -algebra, then (E ) = Eiu. Theorem 1.1.9 and Remark 1.1.11 yield the following fact.

Corollary 1.1.22. Let be a set function on a semi- -algebra E 0

such that (;) = 0. If condition (S ) holds, then has a unique extension to an Ei -idempotent measure on the -algebra generated by E. Proof. We only need to extend to a maxitive set function on the collection of nite unions of elements of E by setting

(

k [ i=1

Fi ) = max (Fi ) i=1;:::;k

(1.1.6)

and apply part 2 of Theorem 1.1.9. The fact that the extension (1.1.6) is unambiguously de ned follows from condition (S 0 ).

© 2001 by Chapman & Hall/CRC

Idempotent measures

13

The notion of a -algebra is obviously an analogue of the notion of a -algebra. The next de nition paraphrases the de nition of complete -algebras.

De nition 1.1.23. We say that a -algebra A is complete with re-

spect to an idempotent measure on (or -complete, for short) if every ! 2 such that (f!g) = 0 is an atom of A.

De nition 1.1.24. We call the completion of a -algebra A with respect to idempotent measure the -algebra that has as its atoms all the elements of idempotent measure 0 and the atoms of A without the elements of idempotent measure zero. We denote the completion by A .

Remark 1.1.25. Clearly, A is the smallest complete -algebra containing A.

De nition 1.1.26. A set with a -algebra of subsets of is called a -measurable space and is denoted as ( ; A).

We now de ne an analogue of a measure space.

De nition 1.1.27. A triplet ( ; A; ), where is a set, A is a -

algebra of subsets of and is an idempotent measure on , is called an idempotent measure space. We denote the idempotent measure space ( ; P ( ); ) as ( ; ). If is an idempotent probability, we refer to ( ; A; ) as an idempotent probability space.

De nition 1.1.28. Given an idempotent measure and a -algebra

A on the set function A de ned by A(A) = (A); A 2 A; is called the restriction of to A.

Remark 1.1.29. As we will see, it is often the case that an idempo-

tent measure is originally speci ed on a -algebra. Though by Theorem 1.1.9 it can always be extended to an idempotent measure on P ( ), this extension might not be unique, which justi es restricting our consideration to the elements of A. To emphasise this we refer to as an idempotent measure on ( ; A). We note, however, that ambiguity in extending does not necessarily lead to ambiguity in the end results. In the sequel, we use the same symbol to denote some extension of from A to P ( ). On the other hand, given an idempotent measure space ( ; A; ), where is uniquely speci ed on A, we could reduce it to a space

© 2001 by Chapman & Hall/CRC

14

Idempotent probability measures

with the discrete -algebra by introducing the factor-space of with respect to the equivalence relation speci ed by the atoms of A. In this sense considering arbitrary -algebras does not give anything new. However, it comes in useful if we need to deal with a collection of algebras on the same set as in Chapter 2, where -algebras are used to keep track of \the history of a process".

1.2 Measurable functions In this section we introduce measurable maps of spaces with idempotent measures. Let and 0 be sets, and E and E 0 be respective collections of subsets of and 0 , both containing ;. De nition 1.2.1. For a function f : ! 0, we de ne 1the 0 collection of subsets of generated by f as the collection f (E ) = ff 1(B ); B E 0g. We also refer to functions de ned on idempotent probability spaces as idempotent variables. The following lemma is a consequence of the de nition. Lemma 1.2.2. If E 0 is a -algebra (respectively, a -system, a 0 paving, a semi- -algebra) on , then f 1(E 0 ) is a -algebra (respectively, a -system, a paving, a semi- -algebra). The collection of the atoms of the -algebra f 1 (E 0 ) is the collection ff 1 (A0 )g, where A0 are the atoms of the -algebra E 0 . De nition 1.2.3. A function f : ! 0 is said to be E =E 0 1 0 measurable if f (E ) E . The following result is obvious. Lemma 1.2.4. A0 function f : ! 0 is Eiu=Eiu0 -measurable if and only if it is Eiu =E -measurable. As a consequence, a function f : ! 0 is A=A0 -measurable, where A and A0 are -algebras on respective sets and 0 , if and only if the inverse images of the atoms of A0 belong to A. Thus, we have the following. Corollary 1.2.5. A function f : ! 0 is A=A0-measurable if and0 only if every atom of A is mapped into a subset of some atom of A . In particular, f is A=P ( 0 )-measurable if and only if it is constant on the atoms of A.

© 2001 by Chapman & Hall/CRC

Measurable functions

15

In the sequel, we refer to A=P ( 0 )-measurable functions as Ameasurable functions or idempotent variables on ( ; A). Note that A is A-measurable if and only if 1(A) : ! R+ is Ameasurable.

Lemma 1.2.6. Let ( ; A; ) be an idempotent measure space and 0 A

be the completion of A with respect to . If f : ! is A measurable, then there exists an A-measurable idempotent variable f 0 such that f 0 = f -a.e. Proof. Let [!]A be an atom of A. We de ne f 0(!0 ) = f (~!) for all !0 2 [!]A , where !~ 2 [!]A is such that (~!) > 0 if ([!]A ) > 0 and !~ is an arbitrary element of [!]A otherwise. Then f 0 is A-measurable and f 0 = f -a.e. by the construction of A (see De nition 1.1.24).

The next lemma is a version of Doob's result.

Lemma 1.2.7. Let a -algebra A on be generated by a function 0 0 0 f : ! , where is equipped with a -algebra A . A function g :

! 00 is A-measurable if and only if there exists an A0 -measurable function h : 0 ! 00 such that g = h Æ f .

Proof. SuÆciency of the condition is obvious. We prove the necessity. Since A is generated by f and g is A-measurable, for arbitrary !00 2

00 there exists A0!00 2 A0 such that g 1 (!00 ) = f 1 (A0!00 ). Since the sets g 1 (!00 ) are disjoint, the sets A0!00 ; !00 2 00 ; are also disjoint. Therefore, letting h(!0 ) = !00 for !00 2 00 and !0 2 A0!00 , and h(!0 ) = !^ for !0 2 [!00 2 00 A0!00 c , where !^ is a xed element of 00 , de nes h unambiguously. Clearly, h Æ f (!) = g(!); ! 2 .

We have the following corollary for functions assuming values on the real line.

Corollary 1.2.8. Let A be a -algebra on

. If functions fj : ! J j 2 J; are A-measurable and F : R ! R, then F ((fj )j 2J ) is A-measurable. In particular, supj fj and inf j fj are A-measurable, R;

and if is a directed set, then lim sup2 f and lim inf 2 f are A-measurable.

We now consider images of idempotent measures under mappings. Let be an idempotent measure on . The next lemma is straightforward.

© 2001 by Chapman & Hall/CRC

16

Idempotent probability measures

Lemma 1.2.9. Let f : ! 0 . 0 0 1 0

Then the set function 0 on 0 de ned by (A ) = f (A ) for A0 0 is an idempotent measure on 0 . De nition 1.2.10. The set function 0 as de ned 1in the lemma is called the image of under f and denoted by Æ f .

For the image of a -smooth idempotent measure to be a smooth idempotent measure, we need to impose conditions on the mapping. By Luzin's theorem in measure theory a real-valued function of a real argument is Borel-measurable if and only if it is continuous on \large" sets (closed or compact). We turn the theorem into the de nition of a measurability concept. The rst step is to introduce an abstract analogue of the concept of a tight measure.

De nition 1.2.11. Let be an E -idempotent measure. We say that

a collection T of subsets of is tightening for if T \ F 2 E for T 2 T and F 2 E , and for arbitrary " > 0 there exists T 2 T such that (T c ) ". We then also say that is tight relative to T , or T -tight, for short.

We next de ne \Luzin measurability".

De nition 1.2.12.

Let T be a tightening collection for an E idempotent measure . A function f : ! 0 is called Luzin (E ; T )=E 0 -measurable if the restriction of f to an arbitrary T 2 T T is ET =E 0 {measurable, where ET = fT F; F 2 Eg. Remark 1.2.13.1 Equivalently, f : ! 0 is Luzin (E ; T )=E 0 { T 0 measurable if f (F ) T 2 E for every F 0 2 E 0 and T 2 T . Note also that E =E 0 -measurability implies Luzin (E ; T )=E 0 -measurability. On the other hand, the collection T = f g is trivially a tightening collection for so that E =E 0 -measurability is a speci c case of Luzin (E ; T )=E 0 -measurability. We refer to Luzin (E ; T )=E 0 -measurable functions as Luzin measurable functions if the collections E , T and E 0 are understood. The purpose of introducing the concept of Luzin measurability is seen from the following theorem.

Theorem 1.2.14. Let be an E -idempotent measure and T 0 tightening collection for . If a function f :

© 2001 by Chapman & Hall/CRC

!

be a

is Luzin

17

Modes of convergence

(E ; T )=E 0 -measurable, then the image 0 = Æ f 1 of under f is a -smooth relative to E 0 idempotent measure on 0. If f (T ) \ F 0 2 E 0 whenever T 2 T and F 0 2 E 0 , then 0 is f (T )-tight. Proof. By Lemma 1.2.9 0 is an idempotent measure. We check property (3). Let F0 2 E 0 be a decreasing net. Given " > 0, let T 2 T be such that (T c) < ". Then

0 (F0 ) = (f 1 (F0 )) f 1(F0 )

\

T + ":

1 (F 0 ) T T decrease as well. Therefore, Since the F0 decrease, the f T since f 1(F0 ) T 2 E , by -smoothness of

f 1(F0 )

\

T

\

!

f 1 (F0 )

\

T

0

\

F0

proving (3). The fact that the idempotent measure 0 is f (T )-tight provided f (T ) \ F 0 2 E0 whenever T 2 T and F 0 2 E 0 follows since is T -tight and 0 f (T )c = (f 1 (f (T )))c (T c): We x some more terminology.

De nition 1.2.15. Let ( ; ) be an idempotent probability space 0 0

and f : ! be an -valued idempotent variable. The idempotent probability Æ f 1 is called the idempotent distribution (or idempotent law) of f under . If 0 is a metric space and limr!1 (f 62 Br (z )) = 0, where z is a xed element of E and Br (z ) denotes the closed r-ball about z , then f is called a proper idempotent variable.

1.3 Modes of convergence We consider idempotent analogues of convergence in measure and convergence almost everywhere. Let ( ; ) be an idempotent measure space. Let f and f denote idempotent variables on with values in a metric space E with metric .

De nition 1.3.1. We say that a net ff; 2 g converges -a.e. to f if

! 2 : f (!) 6! f (!) = 0:

© 2001 by Chapman & Hall/CRC

18

Idempotent probability measures

More generally, we say that a property concerning elements of

holds -a.e. (or a.e. if is understood) if the idempotent measure of the set where the property does not hold equals 0.

De nition 1.3.2. We say that a net ff; 2 g converges to f in idempotent measure (or in idempotent measure if is understood) if for every " > 0

lim (! 2 : (f (!); f (!)) > ") = 0:

2

Note that since ((f; g) > ") ((f; f ) > "=2) + ((g; f ) > "=2), the limit in idempotent measure is unique -a.e. The same fact is of course true for convergence -a.e. We denote convergence in idempotent measure by f ! f.

Lemma 1.3.3. (\Borel-Cantelli") Let fA; 2 g be a net of subsets of . If (A ) ! 0, then lim sup2 A = 0. Proof. The claim follows since

\ [

lim sup A = 2

0 2 0

A

inf 0 2 sup0 (A ):

Theorem 1.3.4.

(\Egorov") f ! f if and only if for every " > 0 there exists a set A" such that (Ac" ) " and sup!2A" (f (!); f (!)) ! 0.

Proof. If ff ; 2 g is a net such that f ! f , then for every Æ > 0 there exists Æ 2 such that ((f ; f ) > Æ) < " for all Æ ; hence, (f ; f ) Æ on the set A" = f! 2 : (f!g) "g. The converse is obvious.

Theorem 1.3.5. If f ! f , then f ! f -a.e.

Proof. By Lemma 1.3.3 lim sup (f ; f ) > " = 0.

The next result gives a partial converse.

Theorem 1.3.6. If f ! f; 2 ; -a.e., then there exists a net fh ; 2 g,

which converges to f in idempotent measure and is

© 2001 by Chapman & Hall/CRC

19

Modes of convergence

such that fh (!); 2 g is a subnet of ff (!); 2 g for every ! 2 . If ffn ; n 2 N g is a sequence such that fn ! f -a.e., then for every ! 2 there exists a subsequence kn (!) such that kn (!) n and fkn ! f . Proof. Let = f(; ") : 2 ; 2 R+ g. We turn into a directed set by de ning that (; ") (0 ; "0 ) if 0 ; " "0 . For = (; ), we de ne (!) as 0 such that (f~(!); f (!)) for all ~ 0 if such a 0 exists and (!) = otherwise. Let h (!) = f (!) (!): Clearly, fh (!); 2 g is a subnet of ff (!); 2 g for every ! 2 . Since f ! f -a.e., 0 exists for almost every ! 2 ; hence, (h ~ (!); f (!)) for almost every ! 2 if ~ . In the case of sequences, we de ne kn (!) = minfl n : (fl (!); f (!)) 1=ng and kn (!) = n if no such l exists.

Remark 1.3.7.

Generally speaking, convergence -a.e. does not imply convergence in idempotent measure. Consider the following example. Let = [0; 1] and (f!g) = 1 for ! 2 [0; 1]. Let fn(!) = n!; ! 2 [0; 1=n], fn (!) = 2 n!; ! 2 [1=n; 2=n] and fn (!) = 0 elsewhere. Then fn(!) ! 0 as n ! 1 for every ! 2 [0; 1]. However, (! : fn(!) = 1) = 1.

We now consider Cauchy nets.

Lemma 1.3.8. If ((f; f0 ) > ") ! 0 as ; 0 2 for every " > 0 and (E; ) is complete, then ff g converges -a.e.

Proof. In analogy with the proof of Lemma 1.3.3, for " > 0, \ [

0

(f ; f0 ) > "

inf sup (f ; f0 ) > " = 0: 0

Thus, the net ff g is Cauchy -a.e., so it converges -a.e. by completeness of (E; ).

Theorem 1.3.9. If ((f; f0 ) > ") ! 0 as ; 0 2 and (E; ) is complete, then ff g converges in idempotent measure.

© 2001 by Chapman & Hall/CRC

20

Idempotent probability measures

Proof. By Lemma 1.3.8 f ! f -a.e. Let us choose as in the proof of Theorem 1.3.6. Then, for > 0, taking = (; =2) and using the inequalities (f ; f ) "=2 -a.e. and (!) ,

((f ; f ) > ") ((f ; f ) > "=2) _ ((f ; f ) > "=2) ( sup (f; f0 ) > "=2) = sup ((f; f0 ) > "=2): 0

0

The latter term goes to 0 as 2 by hypotheses. The last result of the section shows that for -smooth idempotent measures and decreasing nets of R+ -valued Luzin measurable functions convergence -almost everywhere implies convergence in idempotent measure. Let U denote the paving on R+ consisting of intervals [a; 1), where a 2 R+ , and ;. Theorem 1.3.10. Let be -smooth relative to a collection E and let T be a tightening collection for . If f; 2 g is a decreasing net of Luzin (E ; T )=U -measurable R+ -valued functions, which converges -a.e. to 0, then f ; 2 g converges to 0 in idempotent measure. Proof. Let Æ > 0, > 0 and T 2 T such that (T c) ". Then Æ f Æg \ T + ": Since f Æg \ T 2 E, by -smoothness of relative to E we have that f Æg \ T ! \ f Æg \ T = 0.

Remark 1.3.11. The result is a counterpart of the fact in probability

theory stating that a monotonic sequence of non-negative random variables that converges to zero in probability also converges to zero almost surely.

1.4 Idempotent integration In this section we develop an idempotent analogue of integration theory. Let ( ; ) be an idempotent measure space such that ( ) < 1. We adopt the convention that 1 0 = 0. De nition 1.4.1. For a function f on with values in R + we de ne the idempotent integral of f with respect to by _

f d = sup a (f a): a2R+

W

For A , we let A f d =

© 2001 by Chapman & Hall/CRC

W

f

1(A) d.

21

Idempotent integration

Idempotent integral is called idempotent expectation if is an idempotentWprobability. In the sequel we also denote idempotent integrals as f (!) d(!) and, if is an idempotent probability, as Sf , Sf (!), or S f , the latter notation is used to emphasise the idempotent probability for which the idempotent integral is evaluated. The next lemma follows by de nition.

Lemma 1.4.2. Let f : ! R +. The following equivalent representations hold. _

f d = sup a(f = a) = sup f (!)(f!g) !2

a2R+

= sup f (!)([!]f !2

1 (P (R+ )) ):

Remark 1.4.3. If is originally de ned on a -algebra A, then the

value of the idempotent integral of a function f depends generally speaking on what extensionW of to P ( ) we consider. However, if f is A-measurable, then f d is de ned unambiguously, which follows by the last equality in Lemma 1.4.2.WTo emphasise this we W sometimes denote the integral as f dA , f (!) dA (!) and, if is an idempotent probability, as SA f and SA f (!), where A and A denote the respective restrictions of and to A. A careful examination of the proofs below shows that if we require that the functions and sets considered in the statements be A-measurable, then the results are insensitive to the particular extension of to P ( ).

The results below whose proofs are omitted directly follow from Lemma 1.4.2. We consider only R+ -valued integrands, the corresponding properties for R + -valued integrands are derived similarly.

Theorem 1.4.4. Let f; g be R+ -valued functions on . The following properties hold. (JS0) (JS1)

_

_

0 d = 0

f d

_

g d if f

© 2001 by Chapman & Hall/CRC

g

22

Idempotent probability measures

(JS2) (JS3) (JS4) (JS5)

_

(cf ) d = c

_

f d; c 2 R+

_

_

_

_

(f _ g) d =

(f + g) d

_

_

j f d

(JS6)

_

f d +

g dj

well de ned

sup fj d = sup j 2J j 2J

_

f d _

_

_

g d

_

g d

jf gj d provided the left-hand side is

fj d, where fj : ! R+ ; j 2 J

The following Chebyshev-type inequality plays as important a role below as its counterpart does in probability theory.

Lemma 1.4.5. If f : ! R+ , then 1_ (f a) f 1(f a) d; a

a > 0:

We also have an analogue of the change-of-variables formula in the Lebesgue integral. Theorem 1.4.6. 1Let 0 be an idempotent measure on a set 0 such 0 0 that = Æ f for some f : ! . Then, for a function g : 0 ! R+ , _

0

g d0 =

_

g Æf d:

The following \Holder" inequalities are also useful. For f : ! 1=p W p R + and p > 0 we de ne kf k p = and kf k 1 =

f d sup!: (f!g)>0 f (!).

Lemma 1.4.7. Let f; g : ! R+ .

1. Let p 2 [1; 1] and q 2 [1; 1] be such that 1=p + 1=q = 1. Then W fg d kf k p kgk q .

2. If ( ) = 1, then, for 0 < p < q, kf k p kf k q .

© 2001 by Chapman & Hall/CRC

23

Idempotent integration

We are interested in convergence properties of idempotent integrals, so we study an analogue of the concept of uniform integrability.

De nition 1.4.8. We say that a function f : ! R+ is maximable (orW-maximable if the idempotent measure needs to be emphasised) W if f d < 1 and, moreover, f 1(f > a) d ! 0 as a ! 1. The following version of La Vallee-Poussin's theorem holds.

Theorem 1.4.9. A function f : ! R+ is maximable if and only if there exists a monotonically increasing function F : W that F (x)=x ! 1 as x ! 1 and F Æ f d < 1.

R+

! R+

such

Proof. We prove that the condition is suÆcient for f to be maximable. Given " > 0, let a > 0 be such that x=F (x) " for x a. Then _

f 1(f > a) d "

_

F Æf 1(f > a) d "

_

F Æf d:

Conversely, let f be maximable. Since there is no loss of generality in assuming that kf k 1 = 1, we de ne x F (x) = W :

f 1(f x) d Then F is monotonic and F (x)=x ! 1 as x ! 1 by maximability W of f . Also, F Æ f d 1.

De nition 1.4.10. A collection ffj ; j 2 J g of R+ -valued functions on is said to be uniformly maximable (or -uniformly maximable) if

sup j 2J

_

fj 1(fj > a) d ! 0 as a ! 1:

Theorem 1.4.11. A collection ffj ; j 2 J g is uniformly maximable if and only if the following conditions hold: (i) sup j 2J

_

fj d < 1, W

(ii) for every " > 0 there exists > 0 such that supj 2J A fj d < " for every set A such that (A) < .

© 2001 by Chapman & Hall/CRC

24

Idempotent probability measures

Proof. Let ffj ; j 2 J g be uniformly maximable. Then (i) and (ii) follow by the inequality _

A

fj d

_

fj 1(fj > a) d + a(A); a > 0:

The converse follows by the fact that since for all j 2 J and a large enough

(fj > a)

W

fj d

< ; a W where is chosen as in condition (ii), we have fj 1(fj > a) d < "; j 2 J:

Corollary 1.4.12. Let f : ! R+ be maximable. Then the set W function A ! f 1(A) d; A ; is absolutely continuous with respect to in the sense that for every " > 0 there exists Æ > 0 such W that f 1(A) d < " for all A such that (A) < Æ. Theorem 1.4.13. A collection ffj ; j 2 J g of R+ -valued functions on is uniformly maximable if and only if supj 2J fj is maximable.

Proof. SuÆciency of the condition is obvious. Conversely, let ffj ; j 2 J g be uniformly maximable. For > 0, let > 0 be chosen as in part (ii) of Theorem 1.4.11. By \the Chebyshev inequality" (supj fj W a) supj fj d=a, the latter supremum being less than if a is large enough in view of condition (i) of Theorem 1.4.11. Then by W Theorem 1.4.4 and the choice of we have sup j fj 1(supj fj

W 0 a)d = supj fj 1(supj 0 fj a)d .

The following analogue of La Vallee-Poussin's theorem is a simple consequence of Theorem 1.4.9 and Theorem 1.4.13.

Corollary 1.4.14. A collection ffj ; j 2 J g of R+ -valued functions on is uniformly maximable if and only if there exists a monotonically increasing function F : R+ ! R+ such that F (x)=x ! 1 as W x ! 1 and supj 2J F Æ fj d < 1.

The next easy corollary gives simple conditions for uniform maximability.

Corollary 1.4.15. A collection ffj ; j 2 J g is uniformly maximable if either one of the following conditions holds:

© 2001 by Chapman & Hall/CRC

25

Idempotent integration

1. fj f; j 2 J; where f is maximable, 2. sup j 2J 3. sup j 2J

_

_

fj1+" d < 1 for some " > 0, exp(fj ) d < 1 for some > 0.

We now consider uniformly maximable nets. De nition 1.4.16. A net ff; 2 g of R+ -valued functions on

is said to be uniformly maximable if lim sup 2

_

f 1(f > a) d ! 0 as a ! 1:

We have the following analogues of the properties of uniformly maximable collections of functions. Similar proofs apply.

Theorem 1.4.17. A net ff; 2 g is uniformly maximable if and only if the following conditions hold: (i) lim sup 2

_

f d < 1,

(ii) for every W" > 0 there exists > 0 such that lim sup2 A f d < " for every net of sets fA ; 2 g such that lim sup2 (A ) < .

Corollary 1.4.18.

A net ff ; 2 g is uniformly maximable if either one of the following conditions holds: 1. lim sup 2 2. lim sup 2

_

_

f1+" d < 1 for some " > 0, exp(f ) d < 1 for some > 0.

We now study convergence of idempotent integrals.

Theorem 1.4.19. Let ff; 2 g be a net of R+ -valued functions on and f be an R+ -valued function on .

1. (\the Fatou lemma"). If lim inf 2 f f -a.e., then

lim inf

_

f d

© 2001 by Chapman & Hall/CRC

_

f d:

26

Idempotent probability measures

2. (\the Lebesgue dominated convergence theorem"). If f and the net ff g is uniformly maximable, then

lim

_

f d =

_

f d:

3. (\the Lebesgue monotone convergence theorem"). If f -a.e., then

lim

_

f d =

_

! f

"f

f d:

4. Let be -smooth relative to a collection E of subsets of and T be a tightening collection for . Let f be Luzin (E ; T )=U measurable and maximable functions. If f # f -a.e., then

lim

_

f d =

_

f d:

Proof. We only prove part 4. Since ff ; 2 g is uniformly maximable and is T -tight, we can and do assume that the f are bounded by some N and E =U -measurable. Let for m 2 N

i i i i 1 f ; gm = max 1 f : i=1;:::;mN m i=1;:::;mN m m m We have that jgm; f j 1=m and jgm f j 1=m so that gm; = max _

_

j gm; d

f dj

( ) ; m

_

_

j gm d

f dj

( ) : m

On the other hand, using properties of and the facts that the f decrease and ff i=mg 2 E , as 2 , _

i i i f # max f i=1;:::;mN m m i=1;:::;mN m

gm; d = max

We have the following useful consequence.

© 2001 by Chapman & Hall/CRC

=

_

mi

gm d:

27

Idempotent integration

Corollary 1.4.20. Let be an E -idempotent measure with a tight-

ening collection T . Let f : ! R+ be Luzin (E ; T )=U -measurable and -maximable. Then the set function 0 de ned by 0 (A) = W

f 1(A) d; A ; is an E -idempotent measure, which has T as a tightening collection. Proof. It is obvious that 0 is a nite idempotent measure. It is -smooth relative to E by Theorem 1.4.19. It is T -tight since for a 2 R+ and T 2 T _ 0 (T c ) a(T c)+ f 1(f > a) d:

The following lemma establishes connection between convergence in idempotent measure and in \L1 ()". Lemma 1.4.21. Let ff; 2 g be a net of R+ -valued functions on

and f : ! R+ . W 1. If jf f j d ! 0, then f ! f .

2. If f ! f and the net ff g is uniformly maximable, then W

jf f j d ! 0. Proof. Part 1 follows by \the Chebyshev inequality". For part 2, note that if f ! f and the net ff g is uniformly maximable, then f is maximable by Theorem 1.4.19. Hence, by the inequality _

_

jf fj 1(jf fj > a) d [f 1(f > a)+f 1(f > a)] d

the net fjf f jg is uniformly maximable, so Theorem 1.4.19.

W

jf fj d ! 0 by

We now prove an analogue of Daniell's representation theorem, see, e.g., Meyer [88], stating that idempotent integral is speci ed by properties (JS2) and (JS3). Theorem 1.4.22. Let H be a set of R+ -valued functions on , which contains the zero function and is closed under multiplication by nonnegative scalars and the formation of maximums and minimums. Let E denote the paving on consisting of the sets ff ag; f 2 H; a 2 R + , and ;. Let V : H ! R + be a non-negative homogeneous maxitive functional, i.e., V has the properties

© 2001 by Chapman & Hall/CRC

28

Idempotent probability measures

(V 1) V (cf ) = c V (f ); c 2 R+ ; f

2 H; (V 2) V (f _ g) = V (f ) _ V (g); f; g 2 H. Then the following holds. 1. There exists an idempotent measure on such that

V (f ) =

_

f d; f

2 H;

if and only if V is -smooth along increasing nets in the sense that for every increasing net ff g of bounded functions from H such that sup f 2 H we have

V (sup f ) = sup V (f ):

The idempotent measure is uniquely speci ed on Eu . It is an idempotent probability if H contains the function identically equal to 1 and

(V 0) V (1) = 1. 2. Let H, in addition, have either one of the following properties: (a) if f 2 H, then (f 1) _ 0 2 H, (b) if f 2 H, then f ^ 1 2 H, and H is closed under multiplication. Then there exists an E -idempotent measure on such that

V (f ) =

_

f d; f

2 H;

if and only if the following condition holds: (VC) for every nets ff g and fg g of bounded functions from H, which are increasing and decreasing, respectively, and such that sup f inf g , we have

sup V (f ) inf V (g ):

© 2001 by Chapman & Hall/CRC

29

Idempotent integration

The idempotent measure is also -smooth relative to Ei and is uniquely speci ed on Eiu . If, in addition, for every f; g 2 H, we have (f g) _ 0 2 H and V (f + g) V (f ) + V (g) if f + g 2 H, then condition (VC) is equivalent to Daniell's condition: (VD) if f # 0, where the f are bounded functions from H, then V (f) # 0. Proof. We rst deal with the necessity parts. The condition of smoothness of V along increasing nets is necessary for existence of in part 1 by Theorem 1.4.19. Let us show necessity of (V C ) for being a -smooth idempotent measure. The condition follows by Theorem 1.4.19 (with T = f g) if we note that every f 2 H is E =U -measurable so that

sup

_

f d =

_

sup f d; inf

_

f d =

_

inf f d:

We prove suÆciency in part 1. Let 1 be the set whose elements are sets [0; a) f!g; a 2 R+ ; ! 2 . For R+ -valued functions f on , let Wf = f[0; a) f!g : a f (!)g. We de ne E1 = fWf ; f 2 Hg: By the assumptions on H the collection E1 is a paving on 1 . We set U (Wf ) = V (f ). Then the set function U is maxitive on E1 and satis es the hypotheses of part 1 of Theorem 1.1.9 (in particular, since V (0) = 0 by (V 1), it follows that U (;) = U (W0 ) = V (0) = 0). We de ne an extension of U to a set function on P ( 1 ) as in the proof of Theorem 1.1.9 by

U ([0; a) f!g) = inf U (Wf ); U (A1 ) =

f 2H: f (!)a

(1.4.1)

sup U [0; a) f!g ; A1 1 : [0;a)f!g2A1 By part 1 of Theorem 1.1.9 U is an idempotent measure on 1 , which extends U . Since by (V 1), for c 2 R+ and f 2 H, U (Wcf ) = V (cf ) = c V (f ) = c U (Wf ); equality (1.4.1) implies that

U ([0; c a) f!g) = c U ([0; a) f!g):

© 2001 by Chapman & Hall/CRC

(1.4.2)

30

Idempotent probability measures

Now, for A , we de ne (A) = U ([0; 1) A): Clearly, is an idempotent measure on . Also, for arbitrary f 2 H, by -maxitivity of U and (1.4.2)

V (f ) = U (Wf ) = U

[

[0; a) f!g

!2 : af (!)

= sup U [0; a) f!g = sup aU [0; 1) f!g !2 : !2 : af (!) af (!) _ = sup f (!)(f!g) = f d: !2

Part 1 is proved. We prove part 2. Note that under (V C ) the set function U satis es condition (S ) of part 2 of Theorem 1.1.9. Therefore, U is an E1-idempotent measure. We check the -smoothness property for . Let condition (a) of part 2 hold. Since for f : ! R+ and a > 0

f 1(f a) = xinf x 2R+ a

x +1

+

(1.4.3)

and the functions in the in mum belong to H, the sets W1(f a) belong to E1;i for f 2 H and a > 0. Now, let ff a g; f 2 H; 2 ; be a decreasing net of sets from E . We prove that \

inf (f a ) =

ff ag :

(1.4.4)

It is suÆcientTto consider the case a > 0. Then by (1.4.3) W1(f a ) = 2 F; for some F; 2 E1 ; where = R+ . For nite Tsubsets 0 and 0 of and , respectively, we de ne G0 ; 0 = 20 F; . Then 2 0

G0 ; 0

\ \

20

2

F; =

\

20

W1(f a ) W1(f0 a0 ) ;

(1.4.5) T 0 0 where is such that for all 2 0 . Also 0 ; 0 G0 ; 0 = T W 1(f a ) : Since fG0 ; 0 ; (0 ; 0 ) 2 Q() Q( )g is a decreasing net of elements of E1 with respect to the partial order on the

© 2001 by Chapman & Hall/CRC

31

Idempotent integration

pairs (0 ; 0 ) by inclusion and U is -smooth relative to E1 , inf U (G0 ; 0 ) = U

0 ; 0

= U W1

T

ff a g

\

= sup aU [0; 1) a2[0;1)

0 ; 0

= U \

G0 ; 0 = U

\

[

a2[0;1)

[0; a)

W1(f a )

\

ff ag

\

ff ag =

ff ag :

Also by (1.4.5) inf 0 ; 0 U (G0 ; 0 ) inf (f a ). Therefore, \

ff ag inf (f a );

and (1.4.4) follows. Now, is an Ei {idempotent measure by Lemma 1.1.8. If the condition that (f 1) _ 0 2 H for f 2 H is replaced by the conditions that f ^ 1 2 H for f 2 H and H is closed under multiplication, the same proof applies except that (1.4.3) is replaced by the equality 1(f a) = inf x2N (f=a)x ^ 1. To end the proof, let us assume that (f g) _ 0 2 H if f; g 2 H, V is subadditive, i.e., V (f + g) V (f ) + V (g) if f; g; f + g 2 H, and (VD) holds. We check that (VC) holds. Let f ", g # and sup f inf g , where the f and g are bounded functions from H. Then

V (g ) V (g

_ f) V (f)+ V ((g f) _ 0): f ) _ 0 2 H and tends monotonically to zero, an appli-

Since (g cation of (VD) yields (VC).

Remark 1.4.23. Under the hypotheses of part 1 (respectively, part

2) of Theorem 1.4.22 the functional V has a unique extension to a non-negative homogeneous maxitive functional on the set of Eu measurable (respectively, Eiu -measurable) R+ -valued functions on .

Remark 1.4.24.

If H, in addition to the hypotheses of Theorem 1.4.22, is such that (1 f ) _ 0 2 H for every f 2 H, then E is a semi- -algebra, hence, Eiu is a -algebra, which is also generated by sets ff = ag; f 2 H; a 2 R+ . In particular, by the preceding

© 2001 by Chapman & Hall/CRC

32

Idempotent probability measures

remark under the hypotheses of part 2 V has a unique extension on the set of R+ -valued functions that are measurable with respect to the -algebra generated by the elements of H.

Remark 1.4.25. Theorem 1.4.22 implies Theorem 1.1.9 if we take H = fc1(F ); F 2 E ; c 2 R+ g and V (c 1(F )) = c (F ) (note that c 1(F ) ^ 1 = (c ^ 1)1(F ) and (c 1(F ) 1) _ 0 = (c 1) _ 0 1(F ) so that f ^ 1 2 H and (f

1) _ 0 2 H if f

2 H).

Remark 1.4.26. Similarly to Theorem 1.1.9 (see Remark 1.1.11),

Theorem 1.4.22 admits a version where the collection H need not be closed under the formation of maximums and minimums. Then the maxitivity condition (V 2) and the -smoothness condition along increasing nets should be replaced in part 1 by the following condition: if a collection ffj g of elements of H and f 2 H are such that f supj fj , then V (f ) supj V (fj ). In part 2, condition (VC) would have to be replaced by the following: (V C 0) If ffj g is a collection of bounded functions from H and fg g is a decreasing net of bounded functions from H such that supj fj inf g ; then supj V (fj ) inf V (g ): Condition (V C 0 ) is necessary and suÆcient for to be -smooth relative to E . If H is closed under the formation of minimums, then the extension is -smooth relative to Ei .

1.5 Product spaces This section considers products of idempotent measure spaces. De nition 1.5.1. Let ( ; A) and ( 0 ; A0) be -measurable0 spaces. We de ne the product -algebra as the -algebra on with the atoms [!]A [!0 ]A0 . It is denoted by A A0 . The -measurable space ( 0 ; A A0 ) is called the product of ( ; A) and ( 0 ; A0 ). Remark 1.5.2. The product -algebra A A0 0is generated by the semi- -algebra consisting of the rectangles A A , where A 2 A and

© 2001 by Chapman & Hall/CRC

33

Product spaces

A0 2 A0. Since A A0 is also generated by the collection of sets A 0 and A0 , where A 2 A and A0 2 A0 , the -algebra A A0 is the smallest -algebra A~ on 0 such that the projections (!; !0 ) ! ! and (!; !0 ) ! !0 are A~=A-measurable and A~=A0 -measurable, respectively. Lemma 1.5.3. Let A 2 A A0 . Then the projection pr 0 A = f!0 2

0 : (!; !0 ) 2 A for some ! 2 g and cross-sections A! = f!0 2 0 : (!; !0 ) 2 Ag, where ! 2 , are elements of A0 . Also, if f : 0 ! R + is A A0 -measurable, then the cross-section f! (! 0 ) = f (!; ! 0 ) is A0-measurable as a function of !0. Conversely, if A! 2 A0 for every ! 2 and A!0 2 A for every !0 2 0 , then A 2 A A0 . Proof. We start with cross-sections. Let !0 2 A! . Then (!; !0 ) 2 A and by Corollary 1.1.16 [!]A [!0 ]A0 = [(!; !0 )]A A0 2 A. Thus, [!0 ]A0 2 A! and by Corollary 1.1.16 A! 2 A0 . The case of the projection is considered similarly. One could also use the representation pr 0 A = [!2 A! and the de nition of a -algebra. Measurability of f! follows since f! 1 (x) = f 1(x) ! for x 2 R+ . For the nal statement, we rst note that the cross-sections A! depend only on the atom to which ! belongs for otherwise there would exist !1 2 , A ! , !0 2 A and !0 62 A , !2 2 and !0 2 0 such that !1 2 !1 !2 which would imply that !1 2 A!0 but !2 62 A!0 so that A!0 62 A. The required now follows by the equality A = [!2 [!]A A! and the de nition of A A0 .

De nition 1.5.4. Let (

; A) and ( 0 ; A0 ) be -measurable spaces. 0

A function k : P ( ) ! [0; 1] is called an idempotent transition kernel from ( ; A) to ( 0 ; A0 ) if the following holds: (a) for every ! 2 , the function k(!; A0 ); A0 0 ; is an idempotent probability measure on 0 , (b) for every A0 2 A0 , the function k(!; A0 ); ! 2 ; is Ameasurable. Lemma 1.5.5. Let ( ; A) and ( 0; A0) be -measurable spaces. 1. Let be an idempotent measure on ( ; A) and k be an idempotent transition kernel from ( ; A) to ( 0 ; A0 ). Then there is a unique idempotent measure ~ on ( 0 ; A A0 ) such that _ ~(AA0 ) = k(!; A0 ) d(!); A 2 A; A0 2 A0: (1.5.1) A

© 2001 by Chapman & Hall/CRC

34

Idempotent probability measures

In particular, ~(A 0 ) = (A). If a function f : 0 ! R+ is A A0 -measurable, then the function _ g(!) = f! (!0 ) k(!; d!0 ); ! 2 ;

0

is A-measurable and _

0

f d~ =

_

g d:

2. Conversely, if ~ and are idempotent measures on respective -measurable spaces ( 0 ; A A0 ) and ( ; A) such that ~(A 0 ) = (A) for A 2 A, then there exists an idempotent transition kernel k from ( ; A) to ( 0 ; A0 ) such that (1.5.1) holds. The idempotent probability (k(!; A0 ); A0 2 A0 ) is uniquely speci ed on ( ; A0 ) for -almost all !. Remark 1.5.6. As we mentioned in Remark 1.1.29, saying \ is an idempotent measure on ( ; A)" is to mean that is uniquely speci ed only for elements of A. Likewise, uniqueness of ~ is claimed on A A0. Proof of Lemma 1.5.5. We begin with part 1. Since necessarily ~([!]A [!0 ]A0 ) = k(!; [!0 ]A0 )([!]A ), ~ is uniquely speci ed on A A0. We can de ne ~ on P ( 0) by ~(f(!; !0 )g) = k(!; f!0 g)(f!g); ~(A A0 ) = sup ~ f(!; !0 )g : !2A; !0 2A0 The only thing that requires proof is that g is A-measurable. This follows by the fact that f! (!0 )k(!; f!0 g) = f!0 (!)k(!; f!0 g) is Ameasurable in ! for every !0 2 A0 by Lemma 1.5.3 and Corollary 1.2.8. In part 2 necessarily ~([!]A [!0 ]A0 ) k(!; f!0 g) = ([!]A ) if ([!]A ) > 0, which proves uniqueness. For existence, we de ne k(!; f!0 g) by the latter formula if ([!]A ) > 0 and let k(!; A) be arbitrary idempotent probability measures on ( 0 ; A0 ) that are constant on the atoms of A on the rest of .

© 2001 by Chapman & Hall/CRC

35

Product spaces

Remark 1.5.7.0 If k(!; f!0 g) =0 0(f!0 g), where0 00 is an idempotent

measure on , then ~ f(!; ! )g = (f!g) (f! g) and we obtain an analogue of Fubini's theorem. The idempotent measure ~ is called the product of idempotent measures and 0 and denoted as 0 . The idempotent measure space ( 0 ; A A0 ; 0 ) is called the product of the idempotent measure spaces ( ; A; ) and ( 0 ; A0 ; 0 ). We consider now -smoothness and tightness properties of idempotent measures on product spaces. Theorem 1.5.8. Let us assume that an idempotent measure on

is -smooth relative to a collection E P ( ) such that ; 2 E and has a tightening collection T P ( ). Let an idempotent transition kernel k(!; A0 ) from ( ; P ( )) to ( 0 ; P ( 0 )) and collections E 0 P ( 0 ) and T 0 P ( 0 ), such that ; 2 E 0 and T 0 \ F 0 2 E 0 for every T 0 2 T 0 and F 0 2 E 0 , satisfy the following conditions: 1. k(!; A0 ) is Luzin (E ; T )=U -measurable in ! for every A0 2 E 0 , 2. k(!; A0 ) is a -smooth idempotent measure in A0 relative to E 0 for every ! 2 , 3. for every > 0 and T 2 T there exists T 0 2 T 0 such that sup!2T k(!; 0 n T 0 ) . Then the idempotent measure ~ on 0 de ned by ~ f(!; !0 )g = k(!; f!0 g)(f!g) is -smooth relative to E E 0 = fF F 0 ; F 2 E ; F 0 2 E 0 g and has the tightening collection T T 0 = fT T 0; T 2 T ; T 0 2 T 0g. Proof. We rst check the -smoothness. Let fF F 0 ; 2 g be a decreasing net of elements of E E 0 . Let F F 0 = \ 2 (F F 0 ). We have that _ ~(F F 0 ) = k(!; F 0 ) 1(! 2 F ) d(!): (1.5.2)

The functions k(!; F 0 ) 1(! 2 F ) are bounded, Luzin (E ; T )=U measurable and monotonically converge as 2 to k(!; F 0 ) 1(! 2 F ). Therefore by Theorem 1.4.19 ~(F F 0 ) ! ~(F F 0 ) checking -smoothness of ~. The fact that T T 0 is a tightening collection for ~ follows since by (1.5.2) ~ (T T 0 )c = ~ ( T 0 c )[(T c 0) (T c)_ sup k !; T 0 c : !2T

© 2001 by Chapman & Hall/CRC

36

Idempotent probability measures

For future use we also introduce the following notion.

De nition 1.5.9. Let A be a -algebra on a set and B be a -

algebra on a set . We de ne the product B A as the collection of subsets of that are expressed as unions of sets B [z ]A , where B 2 B and [z ]A are atoms of A, such that each atom of A appears in a union only once.

Remark 1.5.10. Clearly, B A is a -algebra but not a -algebra. 1.6 Independence and conditioning Let ( ; ) be an idempotent probability space.

De nition 1.6.1. A nite collection fAi ; i = 1; : : : ; kg of subsets of

is independent if k \

i=1

Ai =

k Y i=1

(Ai ):

A collection fA g of subsets of is independent if every nite subcollection is independent. A collection fE g of families of subsets of is independent if every collection of sets fA g, where A 2 E , is independent.

Lemma 1.6.2. A collection of families fE g is independent if and

only if the collection of families f(E )u g is independent. De nition 1.6.3. Let a set 0 be equipped with a -algebra 0 A0. A0 collection ff g of idempotent variables on with values in is A independent (or independent if A0 = P ( 0 )) if the collection of the algebras f 1 (A0 ) is independent. An idempotent variable f : ! 0 and a collection E of subsets of are A0-independent (or independent if A0 = P ( 0 )) if the -algebra f 1(A0 ) and E form an independent collection.

Remark 1.6.4. Loosely, we will often say that sets, -algebras or idempotent variables, respectively, are independent if the associated collections are independent.

© 2001 by Chapman & Hall/CRC

Independence and conditioning

37

Remark 1.6.5. Clearly, sets are independent if and only if their indicator functions are independent.

Lemma 1.6.6. Let idempotent variables f : ! 0 , where 0 is 0

equipped with a -algebra A , be measurable relative to respective algebras A on . If the collection fA g is independent, then the collection ff g is A0 -independent.

We have the following consequences. Lemma 1.6.7. 1. Let f : ! 0 and 0 be equipped with the discrete -algebra. The collection ff g is independent if and only if the collection of sets ff 1 (! )g is independent for all ! 2 0 . 2. Let ( ; A), ( 0 ; A0 ) and ( 00 ; A00 ) be -measurable spaces. Let f : ! 0 and F : 0 ! 00 be A=A0 - and A0 =A00 -measurable, respectively. If f is independent of a -algebra B on , then F Æ f is also independent of B. In the rest of the section we consider R+ -valued functions on

unless speci ed otherwise. We assume that R+ is equipped with the discrete -algebra P (R + ) and, according to the convention adopted in Section 1.2, refer to A=P (R + )-measurable functions f : ! R+ as A-measurable. We recall that if A is a -algebra on , then f is A-measurable if and only if the inverse images of one-element subsets of R+ belong to A.

Lemma 1.6.8. If f : ! R+ and g : ! R+ are independent, then S (fg) = S (f )S (g).

Proof. The result follows by the representations f (!) = supx2R+ x 1(f (!) = x), g(!) = supx2R+ x 1(g(!) = x) and (fg)(!) = supx;y2R+ (xy) 1(f (!) = x) 1(g(!) = y), and properties of idempotent expectations.

We now de ne conditional idempotent probabilities and conditional idempotent expectations. Let A be a -algebra on , E be a -system of subsets of containing ; and T be a collection of subsets of such that T \ F 2 E for T 2 T and F 2 E . For economy of notation we denote (f!g) as (!).

© 2001 by Chapman & Hall/CRC

38

De nition 1.6.9. 0 0

Idempotent probability measures

The conditional idempotent probability (! jA)(!) of ! given A is de ned by 8 0 A !); if ([!] ) > 0; < (! ) 1(! 0 A (!0 jA)(!) = ([!]A ) :~ 0 (! ); if ([!]A ) = 0; ~ is some idempotent probability on . where For B , we de ne (!0 jB ) = (!0 jAB )(!), where ! 2 B and AB is a -algebra, which has B as an atom (note that the right-hand side does not depend on the particular choice of ! 2 B and AB ). If A , then the conditional idempotent probability of A given A is de ned by (AjA)(!) = sup (!0 jA)(!); ! 2 : !0 2A Similarly, (AjB ) = sup!0 2A (!0 jAB )(!); ! 2 B . If is an E -idempotent probability, then the conditional E idempotent probability given A is de ned in an analogous manner except that ~ is required to be an E -idempotent probability on . ~ to be T -tight. Likewise, if is in addition T -tight, we require 0 0 If f : ! , where is equipped with a -algebra A0, we de ne (Ajf ) = Ajf 1 (A0 ) . Remark 1.6.10. According to the de nition, conditional idempotent probability is uniquely speci ed -a.e. More precisely, if N = f! 2

: ([!]A ) = 0g, then N 2 A, (N ) = 0 and (A \ [!]A ) (AjA)(!) = for all A and ! 2 N c: ([!]A ) Also, if B is such that (B ) > 0, then our de nition of (AjB ) agrees with the \standard" one in that (AjB ) = (A \ B )=(B ). Remark 1.6.11. Let us assume that is uniquely speci ed only on 0 a -algebra A A. Then, recalling that the atoms of A are unions of the atoms of A0 , we have for A0 2 A0 by the above de nition that 8 0 A !); if ([!] ) > 0; < ([! ]A0 ) 1(! 0 A 0 ([! ]A0 jA)(!) = ([!]A ) :~ ([!0 ]A0 ); if ([!]A ) = 0; so the values of the conditional idempotent probability on the elements of A0 do not depend on the extension of to P ( ).

© 2001 by Chapman & Hall/CRC

Independence and conditioning

39

We now list properties of conditional idempotent probabilities. Theorem 1.6.12. The function ((AjA)(!); A ; ! 2 ) has the following properties: 1. it is A{measurable in ! for all A , 2. it is an idempotent probability in A for every ! 2 , 3. for all A and B 2 A,

A \ B = S [(AjA)(!) 1(! 2 B )]:

4. If is an E -idempotent probability and [!]A 2 E (respectively, B 2 E ), then ((AjA)(!); A ) (respectively, ((AjB ); A

)) is an E -idempotent probability in A.

5. If is T -tight, then ((AjA)(!); A ) is T -tight for all ! 2 and ((AjB ); A ) is T -tight for all B . Proof. We begin with property 1. The function ! ! ([!]A ) is A-measurable since it is constant on the atoms of A. Obviously, 1(!0 !) is also0 A-measurable in !. By Lemma 1.2.8 we conclude that ! ! (! jA)(!) is A-measurable; hence, (AjA)(!) is also A-measurable. Property 2 follows by the de nition. We prove property 3. Assume, rst, that B = [^!]A for some !^ 2 . Then, adopting the convention that (!)=([!]A ) = 0 if ([!]A ) = 0, S (AjA)(!) 1(! !^ ) = sup sup (!0 jA)(!) 1(! !^ )(!) !2 !0 2A 0 (! ) = sup sup 1 (!0 !) 1(! !^ )(!) 0 ([ ! ] ) A !2 ! 2A (!0 ) = sup 1 (!0 !^ )([^! ]A ) = sup (!0 ) 1(!0 !^ ) 0 ([^ ! ] ) A ! 2A !0 2A = (A \ [^! ]A ): S In general, if B 2 A, then B = !2B [!]A , and this case reduces to the preceding one. Part 4 of the lemma concerning ! is a consequence of the de nition when ([!]A ) = 0. If ([!]A ) > 0, then by Remark 1.6.10 (A \ [!]A ) (AjA)(!) = ; ([!]A )

© 2001 by Chapman & Hall/CRC

40

Idempotent probability measures

which is a -smooth idempotent probability relative to E provided so is since [!]A 2 E and E is a -system. The proof for B is similar. In part 5 we also can assume that (!) > 0. Since is T -tight, given > 0, there exists T 2 T such that (T c ) ([!]A ), which implies that (T c jA)(!) . The proof for B is similar.

Remark 1.6.13. According to the lemma (AjA)(!) is an analogue

of regular conditional probability (cf., e.g., Ikeda and Watanabe [66, de nition 3.2]). Remark 1.6.14. Let

and 0 be -measurable sets with respective -algebras A and A0, and let be an idempotent probability on 0. Then k(!; A0 ) = ([!]A A0 j[!]A 0 ) is an idempotent transition kernel from ( ; A) to ( 0 ; A0 ).

We now de ne conditional idempotent expectations.

De nition 1.6.15. If f

is an R+ -valued function on , then the conditional idempotent expectation of f given A is de ned as S (f jA)(!) = sup f (!0)(!0 jA)(!): !0 2

If g : ! ^ , where ^ is equipped with a -algebra A^, we de ne S (f jg) = S f jg 1 (A^) . We also let S (f jg = !^ ) = sup f (!0 )(!0 jg = !^ ): !0 2

Remark 1.6.16. Note that S (f jA)(!) < 1 -a.e. if Sf < 1 and (AjA)(!) = S (1(A)jA)(!) -a.e. Remark 1.6.17. We will often use the following form of the de nition of conditional idempotent expectation: sup!0 2 f (!0) 1 !0 2 [!]A (!0 ) -a.e. S (f jA)(!) = [!]A

Remark 1.6.18. If A A0 and f is A0-measurable, then S (f jA)(!) = sup f (!0)([!0 ]A0 jA)(!) !0 2

so that S (f jA) depends only on the values of on A0 .

© 2001 by Chapman & Hall/CRC

(1.6.1)

Independence and conditioning

41

Remark 1.6.19.

In order to refer explicitly to the idempotent probability , we denote the conditional idempotent expectation as S (f jA). Since (!0 jA)(!) is speci ed {a.e. in !, S (f jA)(!) is also speci ed up to a set of zero idempotent probability. We call any such function a version of the conditional idempotent expectation. The following result is a consequence of the de nitions. Lemma 1.6.20. Let f : ! R+ and g : ! 0, where 0 is equipped with the discrete -algebra. Then -a.e. S (f jg)(!) = sup S (f jg = !0 ) 1(g(!) = !0 ); !0 2 0 in particular, -a.e.

(Ajg)(!) = sup (Ajg = !0 ) 1(g(!) = !0 ); A : !0 2 0 Conditional idempotent expectations have properties similar to the properties of conditional expectations in probability theory. They are summarised in the next lemma. All the equalities and inequalities involving conditional idempotent expectations are understood to hold -a.e. Following a convention of probability theory, we routinely omit argument ! in conditional idempotent probabilities and idempotent expectations.

Lemma 1.6.21. Let f , fj , and g denote R+ -valued functions on , A and B denote -algebras on . S (f jA) is A{measurable. If f = g -a.e., then S (f jA) = S (gjA). S (0jA) = 0, S (1jA) = 1: S (c f jA) = c S (f jA); c 2 R+ : S supj 2J fj jA = supj 2J S (fj jA): jS (f jA) S (gjA)j S (jf gjjA); if Sf < 1 and Sg < 1.

and 1. 2. 3. 4. 5. 6.

© 2001 by Chapman & Hall/CRC

42

Idempotent probability measures

7. S S (f jA) = Sf: 8 If f is independent of A, then S (f jA) = Sf . 9. Let ( 0 ; A0 ) be a -measurable space, F : 0 ! R+ be such that the cross-sections F! : 0 ! R+ are A0 -measurable for every ! 2 , and h : ! 0 be A=A0 -measurable. Then S F (; h)jA = S F (; x)jA jx=h. In particular, if g : ! R+ is A-measurable, then S (fgjA) = gS (f jA) and S (gjA) = g. 10. If B A, then S S (f jA)jB = S (f jB): 11. If A and B are independent, and f is independent of B, then S (f j (A; B)) = S (f jA). 12. If f is maximable, then the family fS (f jB); B Ag is uniformly maximable. 13. If 0 < p q, then S (f pjA) 1=p S (f q jA) 1=q . 14. If p 1 and q 1 are such that 1=p + 1=q = 1, then S (fgjA) S (f pjA) 1=p S (f q jA) 1=q . 15. (f ajA) S (f jA)=a; a > 0: Proof. Properties 1{6 follow by de nition. Consider property 7. By the de nition of conditional idempotent expectation, properties of idempotent integrals, and part 3 of Theorem 1.6.12,

S S (f jA) = S sup f (!0 )(!0 jA)(!) !0 2

= sup f (!0 )S [(!0 jA)(!)] = sup f (!0 )(!0 ) = Sf: !0 2

!0 2

Consider property 8. According to the de nitions -a.e. in ! 1(!0 !) (!0 ): S (f jA)(!) = sup f (!0) ([!]A ) !0 2

Since 1(!0 !)=([!]A ) is A{measurable in !0 and f (!0 ) is independent of A, 1(!0 !) (!0 ) sup f (!0 ) ([!]A ) !0 2

1(!0 !) (!0) = Sf: = sup f (!0 )(!0 ) sup !0 2

!0 2 ([! ]A )

© 2001 by Chapman & Hall/CRC

Independence and conditioning

43

Property 8 is proved. We prove property 9. We have that -a.e.

S (F (; h)jA)(!) = sup F (!0 ; h(!0 ))(!0 jA)(!) !0 2

(!0 ) = sup F (!0 ; h(!0 )) 1(!0 !): ([!]A ) !0 2

By A=A0{measurability of h and A0 -measurability of F!0 we have that F (!0 ; h(!0 )) = F (!0 ; h(!)) if !0 !, so we conclude that (!0 ) S (F (; h)jA)(!) = sup F (!0 ; h(!)) 1 (!0 !) 0 ([ ! ] ) A ! 2

= S (F (; x)jA)(!)jx=h(!) : Property 9 is proved. Proof of property 10. It is easy to derive from the de nition of conditional idempotent expectation that for {almost all ! S S (f jA)jB (!) = sup f (!00 ) sup (!0 jB)(!)(!00 jA)(!0 ): !00 2

!0 2

Hence, it suÆces to prove that for {almost all ! and all !00 sup (!0 jB)(!)(!00 jA)(!0 ) = (!00 jB)(!): (1.6.2) !0 2

If (!00 ) = 0, then (!00 jB)(!) = 0 {a.e. in !, so the right-hand side of (1.6.2) is equal to 0 -a.e. As for the left-hand side, if (!00 ) = 0 and (!00 jA)(!0 ) > 0, then (!0 ) = 0. The latter implies that if, in addition (!0 jB)(!) > 0, then (!) = 0. Thus, we conclude that if (!00 ) = 0 and (!0 jB)(!)(!00 jA)(!0 ) > 0, then (!) = 0. Hence, the left-hand side of (1.6.2) also is {a.e. equal to 0 when (!00 ) = 0. This ends the proof of (1.6.2) when (!00 ) = 0. Let (!00 ) > 0. We can assume that (!) > 0 so that ([!]A ) > 0 and ([!]B ) > 0. Then by the de nition of conditional idempotent probability the right-hand side of (1.6.2) takes the form (!00 ) (!00 jB)(!) = 1(!00 B !): (1.6.3) ([!]B ) On the other hand, (!00 ) > 0 implies that ([!00 ]A ) > 0, hence, A !00 , and by the de nition of conditional ([!0 ]A ) > 0 when !0

© 2001 by Chapman & Hall/CRC

44

Idempotent probability measures

idempotent probability we have for the left-hand side of (1.6.2) sup (!0 jB)(!)(!00 jA)(!0 ) !0 2

(!0 ) B !) (!00 ) : = sup 1 (!0 ([!0 ]A ) A !00 ([! ]B ) !0 : !0 A !00 implies that [!0 ] = [!00 ] . It also implies, The equivalence !0 A A B B B !) if 0 00 0 since B A, that ! ! , and hence 1(! !) = 1(!00 A !00 . Thus, on replacing in the latter supremum [!0 ] with [!00 ] !0 A A B !) with 1(!00 B !), we obtain that the supremum coinand 1(!0 cides with the right-hand side of (1.6.3). Equality (1.6.2) is proved. Property 10 is proved. We prove property 11. Let C = (A; B). Obviously, [!]C = [!]A \ [!]B . Then by de nition -a.e. 1(!0 2 [!]C ) (!0 ) S (f jC )(!) = sup f (!0 ) ([!]C ) !0 0 1(! 2 [!]A \ [!]B ) (!0 ) = sup f (!0) ([!]A \ [!]B ) !0 h 1(!0 2 [!]A) (!0)ih 1(!0 2 [!]B ) (!0 )i = sup f (!0 ) ([!]A ) ([!]B ) !0 = S (f jA)(!); where the equality before the last one is obtained by using the fact that f and A are independent of B. We prove property 12. By property 1 S (f jB) is B{measurable, so, successively applying properties 9 and 7, we have for a > 0

S S (f jB) 1 S (f jB) > a = S S (f 1 S (f jB) > a) jB = S f 1(S (f jB) > a) :

Now, by properties of idempotent expectations, for b > 0,

S f 1(S (f jB) > a) = S f 1(f > b) 1(S (f jB) > a) _ S f 1(f S f 1(f > b) _ b (S (f jB) > a)

b) 1(S (f jB) > a)

S f 1(f > b) _ ab Sf ;

© 2001 by Chapman & Hall/CRC

45

Independence and conditioning

where the last inequality is by \the Chebyshev inequality" and property 7. Hence, lim sup sup S S (f jB)1(S (f jB) > a) a!1 BA

S f 1(f > b) ;

which goes to 0 as b ! 1 by maximability of f . Inequalities 13 and 14 follow from the de nitions. Property 15, similarly to the probability theory counterpart, follows by the inequality 1(f a) f=a. The following lemma, which contains facts on convergence of conditional idempotent expectations analogous to facts from probability theory, is a consequence of the de nition of conditional idempotent expectation, Theorem 1.4.19, Theorem 1.6.12, and Lemma 1.6.21.

Lemma 1.6.22. Let A be a -algebra on .

Let ff ; 2 g be a net of R+ -valued functions and f be an R+ -valued function on .

f {a.e., then lim inf S (f jA) S (f jA) {a.e. 2

1. If lim inf 2 f

2. If S jf

2 ; then S jS (f jA) S (f jA)j ! 0:

3. If f

f j ! 0;

! f; 2 ; and ff g is uniformly maximable, then

S (f 4. If f

jA) ! S (f jA):

" f; 2 ; {a.e., then lim S (f jA) = S (f jA) {a.e. 2

5. Let be an E -idempotent probability, T be a tightening collection for , and the f be Luzin (E ; T )=U -measurable and maximable. Let the atoms of A belong to E . If f # f -a.e., then

S (f

jA) # S (f jA) -a.e.

© 2001 by Chapman & Hall/CRC

46

Idempotent probability measures

We now give versions of Levy's upward and downward theorems.

Lemma 1.6.23. Let A be a -algebra on and f : ! R+ . either one of the following conditions hold: 1.

fA ; T

2 g

2.

fA ; 2 g

2 A ,

is a decreasing net of -algebras and

Let

A =

is an increasing net of -algebras, E includes S the atoms of the A , A = 2 A , is a T -tight E idempotent probability, and f is Luzin (E ; T )=U -measurable and maximable.

Then

S (f jA) = lim S (f jA ) -a.e. 2

Proof. We prove part 1. Note that the net f[!]A ; 2 g is increasing for every ! 2 and [!]A = [ 2 [!]A . Equality (1.6.1) implies the claim by -maxitivity of and part 3 of Theorem 1.4.19. The proof of part 2 is similar: we note that the net f[!]A ; 2 g is decreasing for every ! 2 and [!]A = \ 2 [!]A , invoke (1.6.1), part 4 of Theorem 1.4.19, and the -smoothness property of .

In the sequel we will need the following characterisation of conditional idempotent expectations, which is a straightforward consequence of (1.6.1).

Lemma 1.6.24. Let f : ! R+ , g : ! R+ and A be a -algebra on . Then S (f jA) S (gjA) -a.e. if and only if S (f 1(A)) S (g 1(A)) for every A 2 A. The next implication of the lemma characterises conditional idempotent expectation in a manner similar to the de nition of conditional expectation in probability theory. We say that functions f and g on ( ; ) are indistinguishable if (f 6= g) = 0.

Lemma 1.6.25.

Let A be a -algebra and f : ! R+ . Then S (f jA) is the only up to indistinguishability function g : ! R+ , which is A-measurable and satis es the equality S (fh) = S (gh) for all A-measurable functions h : ! R+ .

© 2001 by Chapman & Hall/CRC

47

Independence and conditioning

Remark 1.6.26. If we de ned conditional idempotent expectation

by the property in the lemma, then De nition 1.6.9 would prove existence. In general, one cannot replace in Lemma 1.6.25 algebras with -algebras. Let us consider the following example. Let = [0; 1], (!) = 1 for every ! 2 [0; 1], B [0; 1] be the Borel -algebra on [0; 1], and f be an R+ -valued function on that is not Borel measurable. Suppose there exists conditional idempo tent expectation g of f given B [0; 1] satisfying the requirements of Lemma 1.6.25. Since, given x 2 [0; 1], the function 1(! = x) is Borel, we have sup!2[0;1] 1(! = x)f (!) = sup!2[0;1] 1(! = x)g(!), so that g(x) = f (x) for every x 2 [0; 1], which contradicts the requirement that g be Borel measurable.

We give \a transitivity law" for conditional idempotent expectations.

Lemma 1.6.27. Let f : ! 0 , g : 0 ! R+ and A0 be a -algebra 0 on . Then

S (g Æf jf 1 (A0 ))(!) = SÆf 1 (gjA0 )(f (!))

-a.e.

Proof. Let h(!0 ) = SÆf 1 (gjA0 )(!0 ); !0 2 0 . By Lemma 1.6.25 for every A0 -measurable function v : 0 ! R+ we have SÆf 1 (gv) = SÆf 1 (hv). By Theorem 1.4.6 this implies the equality S (g Æ f v Æ f ) = S(h Æ f v Æ f ). Since by Lemma 1.2.7 an arbitrary f 1 (A0 )measurable R+ -valued function on is of the form v Æ f for a suitable A0-measurable function v and h Æ f is f 1(A0)-measurable, we conclude by Lemma 1.6.25 again that h Æf = S(g Æf jf 1 (A0 )) -a.e.

The next lemma concerns evaluating conditional idempotent expectations for product idempotent probabilities.

Lemma 1.6.28. Let ( ; A0 ; ) and ( 0; A0 ; 0) be idempotent 0prob-

ability spaces and be equipped with -algebra A A and idempotent probability 0 . Let f : 0 ! R+ . Then S0 (f jA A0) = S (f~jA); where f~(!) = S0 (f! (!0 )jA0 ); ! 2

: In particular, if g : ! R+ and g0 : 0 ! R+ , then S0 (gg0 jA A0 ) = S (gjA)S (g0 jA0 ).

© 2001 by Chapman & Hall/CRC

48

Idempotent probability measures

Proof. The required follows since S0 (f jA A0 ) = sup f (!; !0 ) 0 (!; !0 )jA A0 (!;!0 )2 0 = sup sup f (!; !0 )0 (!0 jA0 ) (!jA): !2 !0 2 0

In probability theory existence of conditional expectation is proved by means of the Radon-Nikodym theorem. We can do without an analogue of the latter; however, it comes in useful below, so we state and prove it. De nition 1.6.29. Let A be a -algebra on0 . Let and 0 be idempotent probabilities on . We say that is absolutely continuous with respect to on A if for every " > 0 there exists Æ > 0 such that (A) < Æ implies that 0 (A) < for all A 2 A. Remark 1.6.30. Equivalently, we may require the above condition to hold only for the atoms of A. Note also that our de nition implies that if 0 is absolutely continuous with respect to , then 0 (A) = 0 whenever A 2 A and (A) = 0; by contrast with the situation in measure theory, the converse is not true. We have chosen the \strong" version as a de nition since it implies maximability of the Radon-Nikodym derivative (see Theorem 1.6.34 below). Remark 1.6.31. If 0 and are restricted to A, we simply say that 0 is absolutely continuous with respect to . De nition 1.6.32. We say that a function f : !0 R+ is a RadonNikodym derivative of an idempotent probability with respect to an idempotent probability Won a -algebra A if f is A-measurable, -maximable and 0 (A) = A f d for all A 2 A. We then denote f = d=d0 . We also write d = f d0 . Lemma 1.6.33. If f = d0 =d, then f (!) = 0 ([!]A )=([!]A ) a.e. Proof. Let (!) > 0. By A-measurability of f it is constant on [!]A so that _ 0 ([!]A ) = f d = f (!)([!]A ): [!]A

© 2001 by Chapman & Hall/CRC

49

Independence and conditioning

Thus, a Radon-Nikodym derivative is unique -a.e. Theorem 1.6.34. An idempotent probability 0 is absolutely continuous with respect to an idempotent probability on A if and only if there exists a Radon-Nikodym derivative of with respect to 0 on A. Proof. Existence of the derivative implies the absolute continuity by Corollary 1.4.12. For the converse, we de ne the derivative as in W 0 Lemma 1.6.33. Then f is A-measurable and (A) = A f d for all A 2 A. To show f is -maximable, we write for a > 0 _

f 1(f > a) d = 0 (f > a) 0 (! : (!) 1=a):

Since 0 is absolutely continuous with respect to , the latter idempotent probability can be made less than arbitrary > 0 by choosing a large enough. Lemma 1.6.35. Let an idempotent probability 0 on be absolutely continuous with respect to on a -algebra A. Let f = d0 =d > 0 -a.e. Then, for a function g : ! R+ and -algebra B A, -a.e. S (fgjB) S0 (gjB) = : S (f jB) Proof. Since 0 = S S (f jB) 1(S (f jB) = 0) = S f 1(S (f jB) = 0) and f > 0 -a.e., it follows that S (f jB) > 0 -a.e., so the right-hand side in the above equality is well de ned -a.e. We next have by properties of conditional idempotent expectations that for B2B

S S0 (gjB)S (f jB) 1(B ) = S S S0 (gjB)f 1(B )jB = S S0 (gjB)f 1(B ) = S0 S0 (gjB) 1(B ) = S0 (g 1(B )) = S(fg 1(B )):

Thus, by Lemma 1.6.25 S0 (gjB)S (f jB) = S (fgjB). We end the section by giving versions of Lemma 1.6.24 and Lemma 1.6.25 for -smooth tight idempotent probabilities.

© 2001 by Chapman & Hall/CRC

50

Idempotent probability measures

Lemma 1.6.36. Let E be a semi- -algebra, be -smooth relative

to E and T be a tightening collection for . Let f : ! R+ and g : ! R+ be Luzin (E ; T )=U -measurable and maximable functions. Then the following holds. 1. S (f j (E )) S (gj (E )) -a.e. if and only if S (f 1(A)) S (g 1(A)) for every A 2 E .

2. S (f j (E )) is the only up to indistinguishability Luzin (E ; T )=U measurable maximable function f 0 : ! R+ that is (E )measurable and satis es the equality S (fh) = S (f 0 h) for all E -measurable functions h : ! R+ . Proof. Necessity of the condition in part 1 is obvious. We prove suÆciency. Let (A) = S (f 1(A)) and (A) = S (g 1(A)). By Corollary 1.4.20 and are E -idempotent measures on such that on E . Theorem 1.1.7 implies that on Eiu . Since E is a semi- algebra, Eiu = (E ), completing the proof by Lemma 1.6.24. Part 2 is a consequence of part 1.

Lemma 1.6.37. Let be -smooth relative to a collection E and has

a tightening collection T . Let H be a collection of Luzin (E ; T )=U measurable R+ -valued functions on that contains the zero function, is closed under multiplication by non-negative scalars and the formation of maximums and minimums, and is such that if h 2 H, then (h 1) _ 0 2 H and (1 h) _ 0 2 H. Let f : ! R+ be maximable and Luzin (E ; T )=U -measurable and let A denote the -algebra generated by the elements of H. If a Luzin (E ; T )=U -measurable maximable function g : ! R+ is A-measurable and such that S (fh) = S (gh) for all h 2 H, then g = S (f jA) -a.e. Proof. Let V (h) = S (fh); h 2 H. By Theorem 1.4.19 and the hypotheses the functional V satis es condition (V C ) of Theorem 1.4.22. Therefore, V has a unique extension to a non-negative homogeneous maxitive functional on the set of A-measurable R+ -valued functions on (see Remark 1.4.24). The same fact is true for the functional V 0 (h) = S (gh). Hence, S (fh) = S (gh) for every Ameasurable R+ -valued function h : ! R+ and the required follows by Lemma 1.6.25.

© 2001 by Chapman & Hall/CRC

Topological spaces

51

1.7 Idempotent measures on topological spaces In the next three sections we consider -smooth idempotent measures on topological spaces. Let E be a Hausdor topological space. It will play the part of the set . The part of the collection E will be played either by the collection F of closed subsets of E or the collection K of compact subsets of E . We consider only nite idempotent measures throughout.

De nition 1.7.1. We say that an idempotent measure on E is tight if it has K as a tightening collection, i.e., for every " > 0 there exists compact K E such that (K c ) ".

Note that since E is Hausdor, F -idempotent measures are Kidempotent measures; also the classes of tight F -idempotent measures and tight K-idempotent measures coincide. Therefore, we occasionally refer to tight F -idempotent measures on Hausdor spaces as tight -smooth idempotent measures. We denote (z ) = (fz g) so that symbol will alternatingly be used for an idempotent measure and its density. To avoid confusion we will denote the density by (z ) and the idempotent measure by . We recall that according to Lemma 1.1.4

(A) = sup (z ); A E: z 2A

(1.7.1)

The next lemma relates properties of (z ) to properties of . Recall that a function f : E ! R+ is upper semi-continuous if the sets fz 2 E : f (z) ag are closed.

De nition 1.7.2. A function f : E ! R+ is said to be upper compact if the sets fz 2 E : f (z ) ag are compact for a > 0.

Remark 1.7.3. Upper compact functions attain suprema on closed sets.

Lemma 1.7.4. Let a function (z) :

E ! R+ and set function : P (E ) ! R+ be related by (1.7.1). Then the following holds. 1. If the set function is an F {idempotent measure on E , then the function (z ) is upper semi-continuous.

© 2001 by Chapman & Hall/CRC

52

Idempotent probability measures

2. If the function (z ) is upper semi-continuous, then the set function is a K{idempotent measure. If E is either rst countable or locally compact, then the converse is also true. 3. The set function is a tight F {idempotent measure on E if and only if the function (z ) is upper compact. Proof. Let be an F {idempotent measure and let z ! z; 2 . Then applying the -smoothness property of to the sets F = clfz0 ; 0 g and noting that \ F = fz g, we conclude that lim sup (z ) (z ), i.e., is upper semi-continuous. The rst assertion in part 2 follows from Lemma 1.7.6 below with f (z ) = (z ) 1(z 2 K ), where fK g is a decreasing net of compacts. The second assertion in the case E is rst countable is proved by the argument of the proof of part 1 since we can assume that fz g is countable so that the F are compact. If E is locally compact, then, given z ! z and > 0, by -smoothness there exists a compact K such that z 2 K , z 2 K for \large" and (K ) (z ) + . If is a tight F -idempotent measure, then (z ) is upper semicontinuous by part 1; besides, if K is a compact such that (K c) < a, then fz : (z ) ag K , so (z ) is upper compact. If (z ) is upper compact, then is a K-idempotent measure by part 2. It is tight since, given > 0, one can take K = fz : (z ) "g.

In the sequel, we denote

K (a) = fz 2 E : (z ) ag; a > 0:

(1.7.2)

Remark 1.7.5. According to the lemma, if is a tight F -idempotent

measure, then the sets K (a); a > 0; make up a tightening collection for . We give the lemma used in the above proof. Lemma 1.7.6. Let f be a net of R+ -valued upper compact functions on E monotonically decreasing and converging pointwise to function f . Then

sup f(z ) # sup f (z ): z 2E Proof. For " > 0, let B = fz 2 E : f (z ) supz0 2E f (z 0 ) + "g. The sets B are compact, decreasing and \B = ;. Hence, B0 = ; for some 0 . z 2E

© 2001 by Chapman & Hall/CRC

53

Topological spaces

The following useful fact is in the same theme.

Theorem 1.7.7.

Let be an F -idempotent measure on E . Let be a collection of R+ -valued bounded upper semicontinuous functions closed under the formation of minimums. Then

ffj ; j 2 J g inf j 2J

_

E

fj d =

_

E

inf fj d:

j 2J

Proof. The claim follows by the last assertion of Theorem 1.4.19 if we observe that T = fE g is a tightening collection for and upper semi-continuous functions are Luzin (F ; T )=U -measurable.

The next result is an analogue of Ulam's theorem, see, e.g., Billingsley [11].

Theorem 1.7.8. Let E be either homeomorphic to a complete metric space or locally compact. If is an F -idempotent measure on E , then is tight.

Proof. Let E be metrised by a complete metric. Given Æ > 0, let OÆ denote the collection of nite unions of open Æ-balls in E ordered by inclusion. It follows by -smoothness of relative to F that limO2OÆ (E n O) = 0: Therefore, for arbitrary > 0 and n 2 Nthere n A exist open 1=n-balls An;1 ; : : : ; An;kn such that E n [ki=1 n;i < : k 1 n The set A = \n=1 [i=1 An;i is totally bounded so that, since E is complete, the set K = cl A is compact. Also (E nK ) supn2N E n n A [ki=1 n;i : If E is locally compact, then the -smoothness property of implies that for every > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that (E n [ki=1 Ai ) < :

Let E 0 be a Hausdor topological space and F 0 denote the collection of closed subsets of E 0 . We introduce a class of maps f : E ! E 0 that preserve the property of an idempotent measure being tight and -smooth relative to the collection of closed sets.

De nition 1.7.9. Let0 be a tight F -idempotent measure on E . A

function f : E ! E is said to be -Luzin measurable (or simply Luzin measurable if is understood) if it is continuous when restricted to the compacts K (a); a > 0.

© 2001 by Chapman & Hall/CRC

54

Idempotent probability measures

Remark0 1.7.10. The de nition adapts the abstract notion of Luzin

(E ; T )=E -measurable functions in that f is -Luzin measurable if and only if it is (F ; K )=F 0 -measurable, where K = fK (a); a > 0g.

The equality f K (a) = KÆf 1 (a), which is valid for arbitrary f : E ! E 0 , and Theorem 1.2.14 yield the following result.

Theorem 1.7.11. If is a tight F -idempotent 0measure 0 on E and1 f is a -Luzin measurable mapping from E to E , then = Æ f is a tight F 0 -idempotent measure on E 0 .

An application to f being an embedding provides the following extension result. Corollary 1.7.12. Let E E 0 and the topology on E be ner than the topology induced by the topology on E 0 . Let be a tight F idempotent measure on E . Then the set function 0 on E 0 de ned by 0 (A0 ) = (A0 \ E ) is a tight F 0 -idempotent measure on E 0 .

In view of an important role played by tight F -idempotent probabilities in large deviation theory, we give them a special name.

De nition 1.7.13. A tight F -idempotent probability is called a deviability. If the idempotent distribution of an idempotent variable is a deviability, we also refer to it as a deviability distribution.

Remark 1.7.14. By Theorem 1.7.8 on complete metric spaces and

on locally compact spaces F -idempotent probabilities are deviabilities.

The following is a corollary of Lemma 1.7.4 that will be used frequently below.

Corollary 1.7.15.

An idempotent measure is a deviability if and only if the density (z ) is an upper compact function and supz2E (z ) = 1.

Remark 1.7.16. Since upper compact functions attain suprema on closed sets, for every deviability there exists z (z ) = 1.

2E

such that

Remark 1.7.17. We note that is a deviability if and only if the function I (z ) = ln (z ) is a tight probability rate function in the sense that the sets fz 2 E : I (z ) ag are compact for all a 2 R+ and inf z2E I (z ) = 0.

© 2001 by Chapman & Hall/CRC

55

Topological spaces

De nition 1.7.18.

1. Let ( ; ) be an idempotent probability space. An idempotent variable f : ! E is called a Luzin idempotent variable on ( ; ) if its idempotent distribution Æ f 1 is a deviability on E .

2. Let, in addition, be a Hausdor topological space and be a deviability. An idempotent variable f : ! E is called a strictly Luzin variable on ( ; ) if it is -Luzin measurable.

Remark 1.7.19. By Theorem 1.7.11 strictly Luzin idempotent variables are Luzin idempotent variables.

Strictly Luzin idempotent variables have another useful property. It is a topological version of Corollary 1.4.20. Lemma 1.7.20. Let be a Hausdor topological space and be a deviability on . Let f be a maximable R+ -valued strictly Luzin idempotent variable on ( ; ) such that S f = 1. Then the set function 0 (A) = S f 1(A); A E; is a deviability on E . We now consider topological versions of Theorem 1.4.22. They are analogues of Riesz' representation theorem. Let CK+ (E ) denote the set of R+ -valued continuous functions on E with compact support. Theorem 1.7.21. Let E be a locally compact Hausdor topological + space and V : CK (E ) ! R+ be a functional with properties (V 1) and (V 2) from Theorem 1.4.22, i.e., (V 1) V (c f ) = c V (f ); c 2 R+ ; (V 2) V (f _ g) = V (f ) _ V (g): Then there exists a K{idempotent measure on E such that

V (f ) =

_

E

f d; f

2 CK+(E ):

The idempotent measure is uniquely speci ed on P (E ). Proof. By Theorem 1.4.22 in order to prove existence of we need to check that if ff' ; ' 2 g and fg ; 2 g are, respectively, increasing and decreasing nets of elements of CK+ (E ) such that

sup f' (z ) inf g (z ); z 2 E; 2

'2

© 2001 by Chapman & Hall/CRC

56

Idempotent probability measures

then sup V (f' ) inf V (g ): 2

(1.7.3)

'2

Replacing if necessary the net ff' g by the net ff' ^ g ^g, where ^ 2 is picked arbitrarily, we can assume that all the above functions are supported by a compact K . Let hK 2 CK+ (E ) be such that hK = 1 on K . Given " > 0, the net f(g =(f' _ ("hK )) 1)+ g indexed by (; ) monotonically converges to 0 on K . By Dini's theorem the convergence is uniform, so there exist '0 and 0 such that

g (z ) (1+")(f' (z )_("hK (z ))); z 2 E;

0 ; ' '0 :

Therefore, by the properties of V

V (g ) (1+")(V (f' )_("V (hK )));

0 ; ' '0 ;

which implies (1.7.3) since " > 0 is arbitrary. Since Kiu = P (E ), by Theorem 1.4.22 is uniquely speci ed on P (E ). Corollary 1.7.22. Let and 0 be K-idempotent measures on a locally compact Hausdor topological space E . If _

E

f d =

_

E

f d0 ; f 2 CK+ (E );

then = 0 on P (E ).

The following version for compact spaces has an analogous proof.

Theorem 1.7.23. Let E be a compact Hausdor topological space

and H be a set of R+ -valued continuous functions on E that contains the zero function, is closed under the multiplication by non-negative scalars and formation of maximums and minimums, and is such that if f 2 H, then (f 1) _ 0 2 H. If V : H ! R+ is a functional with properties (V 1) and (V 2), then there exists a (KH )i {idempotent measure on E such that

V (f ) =

_

E

f d; f

© 2001 by Chapman & Hall/CRC

2 H;

57

Topological spaces

where KH is the collection of compacts fz 2 E : f (z ) ag; a 2 R + ; f 2 H. The idempotent measure is uniquely speci ed on (KH )iu . If, in addition, H contains constants and

(V 0) V (1) = 1,

then is a (KH )i -idempotent probability.

Remark 1.7.24.

Theorem 1.7.21 can be derived from Theorem 1.7.23 if one recalls that a locally compact Hausdor space is homeomorphic to an open subset of a compact Hausdor space.

For an R+ -valued function f on E , let kf k = supz2E f (z ). Let Cb+ (E ) denote the set of R+ -valued bounded continuous functions on E . We recall that Tihonov spaces are completely regular T1 -spaces, see, e.g., Kelley [71]. Theorem 1.7.25. Let E be a Tihonov topological space and V : + Cb (E ) ! R+ be a functional with properties (V 1) and (V 2), which is, in addition, tight in the sense that for arbitrary " > 0 there exists a compact K E such that V (f ) " kf k for every f 2 Cb+(E ) that equals 0 on K . Then there exists a tight F {idempotent measure on E such that

V (f ) =

_

E

f d; f

2 Cb+(E ):

The idempotent measure is speci ed uniquely on P (E ). If, in addition, condition (V 0) holds, then is a tight F -idempotent probability. Proof. Let E be the Stone-Czech compacti cation of E , see, e.g., Engelking [47]. We de ne a functional V on Cb+(E ) by V (f ) = V (f ), where f denotes the restriction of f 2 Cb+(E ) to E . It is obvious that V satis es the conditions of Theorem 1.7.21. By Theorem 1.7.21 there exists a K{idempotent measure on E , where K is the collection of compact subsets of E , such that

V (f ) =

_

E

f d; f

2 Cb+(E ):

(1.7.4)

We show that is K-tight, i.e., for every " > 0 there exists a compact K E such that (E n K ) ". Let K be as in the hypotheses.

© 2001 by Chapman & Hall/CRC

58

Idempotent probability measures

The set E n K is open in E so, since E is Tihonov, 1(E n K ) = sup f over f 2 Cb+(E ) such that f 1(E n K ). Therefore, since all these f are equal to 0 on K and is an idempotent measure, (E n K ) = supf V (f ) = supf V (f ) ": Since the embedding E ! E is continuous, the restriction of to E , de ned by (A) = (A) for A E , is a K-idempotent measure. It is tight since W is K-tight. Moreover, K -tightness of implies that ( E n E ) = 0, so E f d = W + E f d: Thus, by (1.7.4) we have for f 2 Cb (E ), denoting by f the continuous extension of f to E , that

V (f ) = V (f ) =

_

E

f d =

_

E

f d:

Remark 1.7.26. According to the+ theorem and part (JS 5) of The-

orem 1.4.4, the functional V : Cb (E ) ! R+ admits a continuous extension to a functional on the space of bounded R+ -valued functions on E with sup-norm. Theorem 1.7.27. Let and 0 be F -idempotent measures on a Tihonov topological space E . If _

E

f d =

_

E

f d0 ; f 2 Cb+ (E );

then = 0 on P (E ).

Proof. Given z 2 E , we have, since E is Tihonov, that 1(fz g) = inf ff 2 Cb+ (E ) : f (z ) = 1g. Therefore, by Theorem 1.7.7

(z ) =

inf +

_

f 2Cb (E ): E f (z )=1

f d;

0 (z ) =

inf +

_

f 2Cb (E ): E f (z )=1

f d0 :

1.8 Idempotent measures on projective limits Our purpose here is to prove analogues of extension theorems for projective systems in measure theory. We are only able to get nice results for projective systems of tight -smooth idempotent measures.

© 2001 by Chapman & Hall/CRC

Projective limits

59

We actually formulate the results for idempotent probabilities, which are our main concern in this book. Let (E ) 2 be a net of Hausdor topological spaces indexed by a directed set . We assume that for all , 2 , 2 , there are maps : E ! E such that = Æ for . We denote by F the collections of closed subsets of the E . Let the E be equipped with deviabilities (i.e., tight F idempotent probabilities). For " > 0, we denote K"; = fz 2 E : (z ) "g. We assume that the maps are Luzin measurable, i.e., their restrictions to the K"; are continuous. The deviabilities are assumed to form a projective system in that = Æ 1 for . We note that this implies that K; = K; . Let E be a Hausdor topological space and maps : E ! E be such that = Æ for . Let F denote the collection of closed subsets of E . Theorem 1.8.1. Let the maps ; 2 ; separate points in E . Let for every " > 0 there exist a compact K" E such that the restrictions of the maps to K" are continuous and (; K ) ( K")c ": Then there exists a deviability on E such that = Æ 1 . Deviability is uniquely speci ed by (z ) = inf 2 ( z ): Remark 1.8.2. Condition (; K ), being an analogue of the ("; K )condition for Radon measures (Schwartz [118]) is referred to below as such. We rst consider the following special case. Lemma 1.8.3. Let E be the projective limit of the system (E ) and the be the canonical projections from E to E . Then the assertion of Theorem 1.8.1 holds. Proof. We de ne (z ) = inf ( z ); z 2 E; 2 and (A) = supz2A (z ); A E: Let K" = fz Then

K =

\

1 K; ;

© 2001 by Chapman & Hall/CRC

(1.8.1)

2 E : (z) "g: (1.8.2)

60

Idempotent probability measures

so K is the projective limit of the (K; ; 2 ). Therefore, K is compact for 2 (0; 1]. It is also empty for > 1. Thus, is a deviability on E by Corollary 1.7.15. Since the \bonding" maps : K; ! K; are onto and continuous, the maps : K ! K; are also onto, see Engelking [47, Corollary 3.2.15]. Therefore,

K; = K ;

(1.8.3)

which is equivalent to the equality (z ) = sup (z ); z 2 1 z

(1.8.4)

which proves that = Æ 1 . Conversely, if is a deviability on E such that = Æ 1 ; 2 ; then (1.8.4) holds, which is equivalent to (1.8.3), which implies, since K is compact, that K is the projective limit of the (K; ; 2 ) so that (1.8.2) holds, Engelking [47, Proposition 2.5.6]. The latter is equivalent to (1.8.1).

Remark 1.8.4. Note that the (; K )-condition is satis ed in this setting in view of (1.8.3). We call the projective limit of the .

For a proof of the general case we need a lemma. Lemma 1.8.5. Let E and E 0 be Hausdor topological spaces. Let E 0 be endowed with an F 0 -idempotent probability 0 , where F 0 is the collection of closed subsets of E 0 . Let an injective mapping h : E ! E 0 and a collection K^ of compact subsets of E be such that the restrictions of h to the elements of K^ are continuous and h(K^ ) is a tightening collection for 0 . Then there exists a unique deviability on E such that 0 = Æ h 1 . It is speci ed by the equality (A) = 0 (h(A)); A E . Proof. We de ne as in the statement of the lemma. In order to check that is -smooth relative to F , let us consider a decreasing net fF ; 2 g of closed subsets of E . For > 0, we choose a compact K 2 K^ such that 0 (E 0 n h(K )) < . Then by the fact that h is injective (F ) = 0 (h(F )) 0 h(F )\h(K ) + = 0 h(F \K ) +:

© 2001 by Chapman & Hall/CRC

61

Projective limits

Since h is continuous when restricted to K and 0 is -smooth relative to F 0 , we have that \

inf 0 h(F \ K ) = 0 h(F \ K ) 2 2

\

0 h

=

2 \

F

2

F :

Thus, \

inf (F ) F + ; 2 2 implying that is -smooth relative to F . Finally, is tight since by h being injective (K c ) = 0 (h(K )c ). Proof of Theorem 1.8.1. Let E 0 be the projective limit of the (E ), let K"0 ; " > 0; be the respective projective limits of the (K"; ), and let 0 : E 0 ! E be the canonical projections. By the part of the theorem already proved there exists a unique deviability 0 on E 0 such that = 0 Æ 0 1 . On the other hand, it is easy to see that there exists a map h : E ! E 0 , which is continuous when restricted to the sets K" , and is such that 0 Æh = . By the ("; K )-condition K K; , which implies that h(K" ) K"0 . Therefore, fh(K" )g is a tightening collection for 0 . Since the family separates points in E , h is injective. Thus, by Lemma 1.8.5 there exists a unique deviability on E such that 0 = Æh 1 , which implies that = Æ 1 . Also (z ) = 0 (h(z )) = inf 0 Æh(z ) = inf ( z ); z 2 E: 2 2

We now consider an application to product spaces. Let fEj ; j 2 J g be a family of Hausdor topological spaces with collections Fj of closed sets. Let be the set of nite subsetsQof elements of J . For 2 let E denote the Cartesian product j2 Ej with product topology; E is endowed with the collection F of closed sets. Let the sets E be equipped with F -idempotent probabilities . As above we denote K"; = fz 2 E : (z ) "g. For , let

© 2001 by Chapman & Hall/CRC

62

Idempotent probability measures

Q

denote the canonical projection E ! E . Let E j 2J Ej be equipped with a Hausdor topology, which is ner than the relative product topology, and E denote the restriction to E of the canonical Q projection j 2J Ej ! E . The next result follows by Theorem 1.8.1.

Theorem 1.8.6. Let the idempotent probabilities 1

form a projective system, i.e., = Æ if . If for every > 0 there exists a compact subset K of E such that (E K )c for all 2 , then there exists a tight F -idempotent probability on E such that = Æ E 1 ; 2 . It is uniquely speci ed by (z ) = inf 2 (E z ). In particular, exists if E is equipped with product topology and the idempotent probabilities j ; j 2 J; are tight. Proof. Only the last claim requires proof. We check that Q the are ^ tight if the j ; j 2 J; are tight. For " > 0, let K"; = j 2 K";j , c ) ". Then K ^ "; is compact and where j (K";j

K^ ";c

=

[

j2

c j 1 K";j

c = sup j 1 K";j j2 c ": = sup j K";j j2

For the (; K )-condition, we can take K = \j jE 1 K;j . The following consequence of Lemma 1.8.5 complements Corollary 1.7.12. Its weaker version has been used in the proof of Theorem 1.7.25. Corollary 01.8.7. Let E and E 0 be Hausdor topological spaces such that E E and the topology of E is ner than the topology induced by the topology of E 0 . Let 0 be a deviability on E 0 . If the collection of compact subsets of E is a tightening collection for 0 , then the set function on E de ned by (A) = 0 (A) is a deviability on E . We now give topological versions of the results of Section 1.5. De nition 1.8.8. Let E and E 0 be Hausdor topological spaces. A function k(z; A0 ) : E P (E 0 ) ! [0; 1] is called a deviability transition kernel from E into E 0 if the following conditions hold: 1. k(z; A0 ) is upper semi-continuous in z for every closed set A0 E0,

© 2001 by Chapman & Hall/CRC

Projective limits

63

2. k(z; A0 ) is a deviability in A0 for every z 2 E and is uniformly tight on compact subsets of E in the sense that for every compact K E and > 0 there exists a compact K 0 E 0 such that supz2K k(z; E 0 n K 0 ) . Theorem 1.5.8 yields the following result.

Theorem 1.8.9.

Let be a tight F -idempotent measure on a Hausdor topological space E . Let E 0 be another Hausdor topological space and E E 0 be equipped with product topology. Let k(z; A) : E P (E 0 ) ! [0; 1] be a deviability transition kernel from E into E 0 . Then the idempotent measure ~ on E E 0 de ned by ~(z; z 0 ) = k(z; z 0 )(z ) is tight and -smooth relative to the collection of closed subsets of E E 0 . Proof. By Theorem 1.5.8 ~ is -smooth relative to the collection fF F 0 g, where F and F 0 are closed in E and E 0 , respectively, and has the tightening collection fK K 0 g, where K and K 0 are compact in E and E 0 , respectively. The -smoothness property implies that ~(z; z 0 ) is an upper semi-continuous function on E E 0 , and the tightness property then allows us to deduce that the function is actually upper compact. Corollary 1.8.10. If and 0 0are deviabilities on respective Hausdor topological spaces E and E , then the product idempotent measure ~ = 0 is a deviability on E E 0 equipped with product topology.

Combining the latter with Theorem 1.8.6 yields the following existence result.

Corollary 1.8.11.

Let j ; j 2 J; be a collection of deviabilities on respective Hausdor spaces Ej . Then there exists an idempotent probability space ( ; ) and independent Luzin idempotent variables fj ; j 2 J; on ( ; ) whose respective deviability distributions are the j .

The following condition will be used for checking that a function is a deviability transition kernel. Lemma 1.8.12. Let E and E 0 be Hausdor0 topological0 spaces and let E be rst countable. If a function k(z; z ) : E E ! [0; 1] is

© 2001 by Chapman & Hall/CRC

64

Idempotent probability measures

upper semi-continuous in (z; z 0 ), the sets fz 0 2 E 0 : supz2K k(z; z 0 ) ag, where a 2 (0; 1], are relatively compact subsets of E 0 for every compact K E , and supz0 2E 0 k(z; z 0 ) = 1 for every z 2 E , then k(z; A0 ) = supz0 2A0 k(z; z 0 ) is a deviability transition kernel from E into E 0 . Proof. We rst note that the hypotheses imply that the function supz2K k(z; z 0 ) is upper compact for every compact K E so that condition 2 in the de nition of a deviability transition kernel holds. We check condition 1. Let A0 be a closed subset of E 0 and zn ! z as n ! 1. The set K = [n2N fzn g [ fz g is a compact subset of E . Therefore, given an arbitrary > 0, there exists compact K 0 E 0 such that k(zn ; E 0 n K 0 ) ; n 2 N . Hence,

lim sup k(zn ; A0 ) lim sup k(zn ; K 0 \ A0) + k(z; K 0 \ A0 ) + n!1 n!1 k(z; A0 ) + ; where the second inequality follows by upper semi-continuity of k(z; z 0 ) and the fact that K 0 \ A0 is compact.

1.9 Topological spaces of idempotent probabilities In this section we consider topologies on the space of idempotent probability measures on a topological space, our main concern being the weak topology. Let E be a topological space and let IM(E ) denote the set of F -idempotent probabilities on E , where F is the collection of closed subsets of E . As above we denote by Cb+ (E ) the set of all R+ -valued bounded continuous functions on E . De nition 1.9.1. The weak topology on IM(E ) is the coarsest W topology for which the maps ! E h d are continuous for all h 2 Cb+ (E ). According to the de nition, a base for the weak topology consists W W 0 0 of sets f 2 IM(E ) : j E hi d E hi dj < "; i = 1; : : : ; k g, where 2 IM(E ); hi 2 Cb+ (E ); " > 0: For most of the section we assume that E is a Tihonov topological space, in which case IM(E ) is a Hausdor topological space by Theorem 1.7.27. We denote coniw vergence in the weak topology as ! . Our purpose is to show that

© 2001 by Chapman & Hall/CRC

65

Spaces of idempotent probabilities

the weak convergence of idempotent probabilities has many of the properties of the weak convergence of probability measures. Let us denote by C +b (E ) the set of all upper semi-continuous bounded R+ -valued functions on E , and by C +b (E ) the set of all lower semi-continuous bounded R+ -valued functions on E . For a function h : E ! R+ , let h and h denote the respective upper semicontinuous and lower semi-continuous envelopes of h de ned by

h=

inf

f and h =

f 2C + b (E ): f h

sup g:

g2C + b (E ): gh W

W

We say that h is continuous relative to if E h d = E h d. We say that h is upper semi-continuous (respectively, lower semiW W continuous) relative to if E h d = E h d (respectively, W W h d = E E h d). We adopt a similar terminology for sets. We call a set H E continuous relative to if (int H ) = (cl H ). We call a set H E closed (respectively, open) relative to if (H ) = (cl H ) (respectively, (H ) = (int H )). If E is a Tihonov space, a set is continuous (closed or open, respectively) relative to if and only if its indicator function is continuous (upper semi-continuous or lower semi-continuous, respectively) relative to .

Theorem 1.9.2. (Portmanteau theorem) Let E be a Tihonov topological space. Let 2 IM(E ) and 2 IM(E ); The following conditions are equivalent. iw !

1:

_

2:

E

3: (i)

h d !

lim inf

(ii) lim sup

_

E _ E

_

E

2 ; be a net.

h d for all h 2 Cb+(E )

g d f d

_

E _ E

g d

for all g 2 C +b (E )

f d for all f

2 C +b (E )

30 : The inequalities of part 3 hold for all lower semi-continuous relative to , bounded functions g : E ! R+ and all upper semicontinuous relative to , bounded functions f : E ! R+ , respectively

© 2001 by Chapman & Hall/CRC

66

Idempotent probability measures

4: (i) lim inf (G) (G) (ii) lim sup (F ) (F )

for all open G E for all closed F E

40 : The inequalities of part 4 hold for all open relative to sets G and closed relative to sets F , respectively 5: lim (H ) = (H ) for all continuous relative to sets HE 6:

lim

7:

lim

_

E

_

E

h d = h d =

_

E

_

E

h d

for all continuous relative to bounded functions h : E ! R+

h d

for all bounded functions h : E ! R+ that are uniformly continuous with respect to a given uniformity on E

Proof. Conditions 1 and 2 are equivalent by the de nition of the weak topology. Clearly, 2 ) 7, 3 , 30 , 3 ) 2, 3 ) 4, 30 ) 6, 4 , 40 , 40 ) 5, and 6 ) 2. We prove the implication 2 ) 4. To prove 2 ) 4(i), we note that, since E is Tihonov and G is open, 1(G) = sup h over h 2 Cb+W(E ) such that h 1(G). Therefore, by Theorem 1.4.4 (G) = suph E h d, W so that if h 1(G) is such that (G) E h" d + ", then

lim inf (G) lim

_

E

h" d =

_

E

h" d (G) ":

The proof of 4(ii) is analogous if we note that 1F = inf h over h 2W Cb+(E ) such that h 1(F ) so that by Theorem 1.7.7 (F ) = inf h E h d . We prove that 4(i) ) 3(i) and 4(ii) ) 3(ii). For g 2 C +b (E ) such that kgk = 1 let

gk (z ) = max

hi

i=0;:::;k 1

1 k

g(z ) >

i i ; k 2 N: k

W

Since E gk d = maxi=0;:::;k 1 i=k (g(z ) > i=k) and the sets fz : g(z) > xg are open by the lower semi-continuity of g, 4(i) yields lim inf

_

E

gk d

© 2001 by Chapman & Hall/CRC

_

E

gk d:

67

Spaces of idempotent probabilities

As gk (z ) < g(z ) gk (z ) + 1=k, by Theorem 1.4.4 lim inf

_

E

gd lim inf

_

E

gk d

_

E

gk d

_

E

gd

1 ; k

which yields 3(i). The proof of 4(ii) ) 3(ii) is similar if we consider the functions fk (z ) = maxi=0;:::;k 1 (i + 1)=k 1 f (z ) i=k : Now we prove 5 ) 4. Let G be open andWÆ > 0. Let h be a function from Cb+ (E ) such that h 1(G) and E h d (G) Æ. Let Hu = fz 2 E : h(z ) ug; u 2 [0; 1]: Then the function (Hu ) increases as u # 0 so it has at most countably many jumps. Also W (Hu ) E h d u, so (Hu ) (G) 2Æ for u small enough. Thus, there exists " > 0 such that (H" ) (G) 2Æ and (Hu ) is continuous at ". By -maxitivity of the latter is equivalent to H" being continuous relative to . Thus, we conclude that lim inf (G) lim (H") = (H" ) (G) 2Æ:

The proof of 4(ii) is similar. We prove that 7 ) 4(ii). Let V be a uniformity on E and F be a closed subset of E . Let f g be a collection of pseudo-metrics on E , uniformly continuous with respect to V , which is closed under the formation of maximums and such that 1(F ) = inf ">0 inf (1 (z; F )=")+ , where (z; F ) = inf z0 2F (z; z 0 ). The functions (1 (z; F )=")+ are bounded and uniformly continuous with respect to V so that by Theorem 1.7.7 _

lim sup (F ) inf inf lim (1 (z; F )=")+ d

= inf inf ">0

">0

_

E

(1

E (z; F )=")+ d

=

_

E

inf inf (1 (z; F )=")+ d = (F ):

">0

The implication 7 ) 4(i) is proved in an analogous manner.

Remark 1.9.3. As the proof shows, in part 7 it is enough to require

that the convergences hold for functions h that are Lipshitz continuous with respect to the pseudo-metrics specifying the uniformity.

© 2001 by Chapman & Hall/CRC

68

Idempotent probability measures

Remark 1.9.4. In particular, Theorem 1.9.2 implies that the weak 0

topology on IM(E ) is also generated by the subbase f 2 IM(E ) : 0 (G) > (G) "g; f0 2 IM(E ) : 0 (F ) < (F ) + "g; where the G are open, F are closed, " > 0, 2 IM(E ), as well as by the subbase f0 2 IM(E ) : j0(H ) (H )j < "g, where the H are continuous relative to , " > 0; 2 IM(E ). Remark 1.9.5. The de nition of the weak topology also applies to arbitrary nite F -idempotent measures. Theorem 1.9.2 is retained. Remark 1.9.6. For general Hausdor topological spaces the convergences in part 4 of Theorem 1.9.2, which specify the narrow topology, see, e.g., O'Brien and Vervaat [97], imply convergence in the weak topology. On the other hand, if we de ned the weak topology in analogy with Topse [125] by requiring that it be the weakest topology W such that the evaluations ! E g d areW lower semi-continuous for all g 2 C +b (E ) and the evaluations ! E f d are upper semicontinuous for all f 2 C +b (E ), then the weak topology would be equivalent to the narrow topology. Also for this topology the requirement of E being Tihonov in Theorem 1.9.17 below can be relaxed. Corollary 1.9.7. Let E be a Tihonov topological space. Let E0 be a subset of E equipped with relative topology. Let 2 IM(E ) and 2 IM(E ) be such that (E n E0 ) = (E n E0 ) = 0 and the ~ and ~, restrictions of and to E0 , which are denoted by respectively, are -smooth relative to the collection of closed subsets iw iw ~ ~ ! of E0 . Then ! if and only if . Remark 1.9.8. The -smoothness property in the hypotheses holds if either E0 is a closed subset of E or the and are tight F idempotent probabilities on E . We next give suÆcient conditions for continuity relative to and the other related notions. Let E0 E . De nition 1.9.9. We say that a set H E is E0 {closed if it contains all its accumulation points that are in E0 , i.e., cl H \ E0 H . We say that a set H E is E0 {open if every point of H \ E0 is an interior point of H , i.e., H \ E0 int H . Remark 1.9.10. Note that both the interior and closure are taken in E . Also, H is E0 {open if and only if its complement in E is E0 {closed.

© 2001 by Chapman & Hall/CRC

69

Spaces of idempotent probabilities

De nition 1.9.11.

A function h : E ! R+ is called E0 {upper (respectively, E0 {lower) semi-continuous if the sets fz 2 E : f (z ) ag (respectively, fz 2 E : f (z ) ag), a 2 R+ , are E0 {closed. A function h : E ! R+ is said to be E0 {continuous if h 1 (G) is E0 { open for each open G R+ .

Remark 1.9.12. An indicator function 1(A), A E , is E0{upper (respectively, E0 {lower) semi-continuous if and only if A is E0 {closed (respectively, E0 {open).

Remark 1.9.13.

If E is Hausdor, then a function h is E0 upper (respectively, E0 -lower) semi-continuous if and only if lim sup h(z ) h(z ) (respectively, lim inf h(z ) h(z )) for every 2 2 net z ! z 2 E0 . Similarly, h : E ! R+ is E0 {continuous if and only if lim h(z ) = h(z ) for every net z ! z 2 E0 . 2 Let us say that is supported by E0 if (E n E0 ) = 0.

Lemma 1.9.14. Let E be Hausdor and be supported by E0.

1. If a function h : E ! R+ is E0 {continuous (E0 {uppersemi-continuous or E0 {lower-semi-continuous, respectively), then it is continuous (upper semi-continuous or lower semicontinuous, respectively) relative to . 2. If a set H E is E0 {continuous (E0 {closed or E0 {open, respectively), then it is continuous (closed or open, respectively) relative to .

Proof. Part 1 follows by the fact that on Hausdor spaces h(z ) = lim sup h(z 0 ) and h(z ) = lim inf h(z 0 ); U 2Uz z 0 2U U 2Uz z 0 2U

where Uz is the collection of open neighbourhoods of z ordered by inclusion. Part 2 is a consequence of the de nitions. Our next goal is to prove a Prohorov criterion of weak relative compactness. We denote by IMt (E ) the set of tight F -idempotent probabilities on E .

De nition 1.9.15.

1. A subset inf K 2K sup2A (K c ) = 0.

© 2001 by Chapman & Hall/CRC

A of IMt(E )

is called tight if

70

Idempotent probability measures

2. A net f ; 2 g in inf K 2K lim sup2 (K c ) = 0.

IM(E )

is called tight if

De nition 1.9.16. A net f ; 2 g in IM(E ) is called relatively compact if every subnet of f ; 2 g contains a weakly convergent subsubnet.

We also use the standard de nition that a subset of IM(E ) is relatively compact for the weak topology if its closure is compact.

Theorem 1.9.17.

1. Let E be a Tihonov topological space. If a subset A of IMt (E ) (respectively, a net f ; 2 g in IM(E )) is tight, then A (respectively, f ; 2 g) is relatively compact, the accumulation points being elements of IMt(E ).

2. Let E be homeomorphic to a complete metric space. If a subset A of IM(E ) is relatively compact, then A is tight. 3. Let E be locally compact and Hausdor. If a subset A of IM(E ) (respectively, a net f ; 2 g in IM(E ) ) is relatively compact, then A (respectively, f ; 2 g) is tight. Proof. We prove part 1 by proving that every tight net f ; 2 g in IM(E ) contains a subnet converging to an element of IMt (E ). Let Cb;+1 (E ) = ff 2 Cb+ (E ) : kf k 1g. The mapping ! W ( E f d; f 2 Cb;+1 (E )) de nes a homeomorphism between space + IM(E ) and a subspace of space [0; 1]Cb;1 (E) with product topology. The latter space being compact by Tihonov's theorem and Hausdor, there exists a subnet f0 ; 0 2 0 g of f ; 2 g that converges to + + an element of [0; 1]Cb;1 (E ) . By the de nition of topology on [0W; 1]Cb;1 (E ) and properties of idempotent integral this implies that E f d0 converges for every f 2 Cb+ (E ). Denoting the limits by V (f ) we conclude in view of Theorem 1.4.4 that the functional f ! V (f ) has properties (V 0){(V 2). Tightness of f ; 2 g implies that the functional is tight in the sense of Theorem 1.7.25. Thus, the functional V (f ) satis es all theWconditions of Theorem 1.7.25; according to the theorem V (f ) = E f d; f 2 Cb+ (E ); for some tight iw F -idempotent probability , which implies that 0 ! . This completes the proof of part 1.

© 2001 by Chapman & Hall/CRC

71

Spaces of idempotent probabilities

For part 2, we may assume that E is a complete metric space. Also replacing A by its closure, we may assume that A is a compact subset of IM(E ). We rst show that for all " > 0 and Æ > 0 there exist open Æ-balls A1 ; : : : ; Ak such that

En

k [ i=1

Ai < "; 2 A:

(1.9.1)

Since each 2 A is tight by Theorem 1.7.8, we can choose compacts K in E such that (E nK ) < "=2. Let B;1 ; : : : ; B;l be open Æ-balls that cover K so that

En

l [ i=1

B;i < "=2; 2 A:

(1.9.2)

Let n

l [

i=1 l [

G = 0 2 IM(E ) : 0 E n < En

i=1

B;i

B;i +

"o ; 2 A: (1.9.3) 2

Since by Remark 1.9.4 fG ; 2 Ag is an open cover of the compact set A, there exist G1 ; : : : ; G p that also cover A so S P 2 pj=1 G j ; 2 A: Then denoting k = pj=1 l j and Cj = Slj i=1 Bj ;i ; j = 1; : : : ; p; and taking A1 = B1 ;1 ; : : : ; Al1 = B1 ;l1 ; Al1 +1 = B2 ;1 ; : : : ; Ak = Bp ;lp we have by (1.9.3) and (1.9.2)

En

k [ j =1

Aj

j=1min (E nCj ) ;:::;p j=1 max [j (E nCj ) + "=2] < "; 2 A; ;:::;p

which is the required property. Therefore for arbitrary " > 0 and k = 1; 2; : : : ; there exist open 1=k-balls Ak1 ; : : : ; Aknk such that

En

nk [

i=1

Aki < "; 2 A:

© 2001 by Chapman & Hall/CRC

(1.9.4)

72

Idempotent probability measures

T Snk The set A = 1 k=1 i=1 Aki is totally bounded and hence relatively compact by completeness of E . At the same time, by (1.9.4) and -maxitivity of

nk [

(E nA) = sup E n Aki k2N i=1

"; 2 A;

i.e., the set A is tight. Now let E be locally compact and Hausdor. The case of a relatively compact set A is tackled similarly to part 2 in that one can show that there exist open sets Ai with compact closures such that (1.9.1) holds. Now, let a net f ; 2 g be relatively compact in IM(E ). It is suÆcient to prove that for every " > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that

k [

lim sup E n Ai 2 i=1

":

(1.9.5)

We introduce a partial order on the collection O of nite unions O of open sets with compact closures so that O O0 if O O0 . Let denote the set of pairs (; O). We turn into a directed set by de ning that (; A) (0 ; A0 ) if 0 and A A0 . We denote (E n O) = x . Let fx 0 ; 0 2 0 g be a subnet of fx ; 2 g such that lim supO2O lim sup2 x = lim 0 2 0 x 0 (cf. Kelley [71]). The mapping from 0 to in the de nition of a subnet induces a mapping from 0 to by associating with elements of 0 the rst components of the corresponding elements of . This de nes a subnet f 0 g of . Since for every O 2 O the second component in ( 0 ) contains O for all 0 large enough, we have that lim 0 2 0 x 0 lim sup 0 2 0 0 (E n O): Let f 00 ; 00 2 00 g be a subnet of f 0 ; 0 2 0 g that weakly converges to a deviability . De ning x 00 = x 0 , where 0 is the image of 00 under the mapping from 00 to 0 in the de nition of a subnet, we conclude that fx 00 g is a subnet of fx 0 g. Therefore, for arbitrary O 2 O by Theorem 1.9.2 (E nO) lim sup 00 (E nO) 00lim 00 x 00 = lim sup lim sup x : 00 2 00 2 O2O 2 Since is -smooth relative to the collection of closed subsets of E and the union of the sets O equals E by local compactness of E , it follows that limO2O (E n O) = (;) = 0; so we conclude that lim supO2O lim sup2 (E nO) = 0; as claimed.

© 2001 by Chapman & Hall/CRC

Spaces of idempotent probabilities

73

We now assume that E is a metric space with metric and prove that the weak topology on IM(E ) is metrisable. The next lemma has been proved in the proof of Theorem 1.7.8. Lemma 1.9.18. If 2 IM(E ), then is separable in the sense that for every > 0 and Æ > 0 there exists a nite collection of open k Æ-balls A1 ; A2 ; : : : ; Ak such that E n [i=1Ai < . We now de ne an idempotent analogue of the Prohorov metric. We denote by B (z ) the closed ball of radius about z , by A the closed {neighbourhood of a set A, and let A = E n (E n A) . De nition 1.9.19. Given ; 0 2 IM(E ), we de ne p(; 0 ) = inf > 0 : (z ) 0 B(z ) + ; 0 (z ) B (z ) + for all z 2 E :

It is not diÆcult to check that p is a metric on IM(E ). The next lemma follows by -maxitivity of an idempotent measure. Lemma 1.9.20. Let ; 0 2 IM(E ). Then p(; 0 ) = inf > 0 : (A) 0 A + ; 0 (A) A + for all A E :

Remark 1.9.21. One obtains the same metric as p if, as in the standard de nition of the Prohorov metric, one considers open rather than closed -neighbourhoods of z and A, respectively. Theorem 1.9.22. The metric p is compatible with the weak topology on IM(E ). Proof. We rst prove that the topology induced by p is ner than the weak topology. By Remark 1.9.4 it is suÆcient to prove that, given 2 IM(E ), a closed set F , an open set G, and > 0, there exists Æ > 0 such that f0 2 IM(E ) : p(; 0 ) < Æg f0 2 IM(E ) : 0(F ) < (F ) + g

and

f0 2 IM(E ) : p(; 0 ) < Æg f0 2 IM(E ) : 0 (G) > (G) g:

© 2001 by Chapman & Hall/CRC

74

Idempotent probability measures

Since is -smooth relative to F , there exists Æ 2 (0; =2) such that (F ) (F Æ ) =2. Therefore, if p(; 0 ) < Æ, then 0 (F ) < (F Æ ) + Æ (F ) + proving the rst inclusion. For the second inclusion, using -maxitivity of , we choose Æ 2 (0; =2) such that (G) (G Æ )+ =2. Then, if p(; 0 ) < Æ, then (G) (G Æ )+ =2 < 0 (G) + . Conversely, we show using again Remark 1.9.4 that given and > 0 there exists a collection Hi ; i = 1; : : : ; k of sets, which are continuous relative to , and Æ > 0 such that

f0 2 IM(E ) : j0(Hi) (Hi)j < Æ; i = 1; : : : ; kg f0 2 IM(E ) : p(; 0 ) < g: Let Æ < =3. By separability of there exist closed Æ=2-balls B1 ; : : : ;Bk 1 centred at z1 ; : : : ; zk 1 , respectively, such that E n [ki=11Bi < Æ. By -maxitivity of for each i = 1; 2; : : : ; k 1 there exists a closed ball Hi centred at zi of radius not less than Æ=2 and not greater than Æ, which is a continuous set relative to . We also Æ0 k 1 take Hk = E n[i=1 Bi , where Æ0 > 0 is chosen such that Hk is continuous relative to and (Hk ) 2Æ. Let 0 2 IM(E ) be such that j0 (Hi) (Hi)j < Æ; i = 1; : : : ; k. If z 2 Hi for some i = 1; : : : ; k 1, then (z ) (Hi ) < 0 (Hi ) + Æ 0 (B2Æ (z )) + Æ. Similarly, 0 (z ) < (B2Æ (z )) + Æ. If z 62 [ki=11Hi , then (z ) < Æ; also, since E n [ki=11 Hi Hk , we have that 0 (z ) 0 (Hk ) < (Hk ) + Æ 3Æ. We thus conclude that p(; 0 ) < . We next show that the weak topology on IM(E ) is also metrised by a Kantorovich-Wasserstein metric. For f 2 Cb+ (E ) let 0 kf k BL = sup f (z) _ sup0 jf (z)(z; zf0()z )j : z 2E z 6=z Clearly, if kf k BL < 1, then f is bounded and Lipshitz-continuous. De nition 1.9.23. For ; 0 2 IM(E ), we let

BL (; 0 ) =

sup

_ _ f d0 : f d

f 2Cb+ (E ): E kf k BL 1

© 2001 by Chapman & Hall/CRC

E

75

Spaces of idempotent probabilities

It is not diÆcult to check that BL is a metric on IM(E ). Lemma0 1.9.24. Let and 0 be in IM(E ). Then BL (; 0 ) 2p(; ).

Proof. We have for f such that kf k BL 1 and Æ > 0 _ f d

E

_

_

f d0 sup

_

f d

f d0

z 2E B (z ) E BÆ (z ) Æ _ _ sup f d f (z )0 (z ) f (z )(z ) f d0 z 2E B (z ) BÆ (z ) Æ 0 Æ + sup (z ) BÆ (z ) BÆ (z ) 0 (z ) ;

_

z 2E

_

yielding the required. As a consequence, we have the following result.

Theorem 1.9.25. The metric BL metrises the weak topology on IM(E ).

Proof. By Theorem 1.9.2 and Remark 1.9.3 the convergence BL ( ; ) ! 0 implies the convergence ! . Thus, the topology induced by the metric BL is ner than the weak topology. The converse follows by Theorem 1.9.22 and Lemma 1.9.24.

Since on a metric space the notions of sequential compactness and compactness are identical, metrisability of IM(E ) allows us to give criteria for sequential compactness. We rst recall relevant de nitions.

De nition 1.9.26. A subset A of IM(E ) is called relatively sequentially compact (for the weak topology) if every sequence fn ; n 2 N g of elements of A contains a weakly convergent subsubsequence.

Combining Theorem 1.9.17 and Theorem 1.9.22 yields the following result.

Theorem 1.9.27. Let E be a metric space.

1. If a subset A of IM(E ) is tight, then A is relatively sequentially compact, the accumulation points being elements of IMt (E ).

© 2001 by Chapman & Hall/CRC

76

Idempotent probability measures

2. Let E be homeomorphic to a complete metric space. If a subset A of IM(E ) is relatively sequentially compact, then A is tight.

We give, however, a proof of part 1 that does not use Theorem 1.9.17. Proof. Let n 2 IM(E ); n 2 N . Let us assume rst that E is a compact metric space. The set Cb;+1 (E ) of R+ -valued continuous functions on E that are bounded by 1, endowed with the topology of uniform convergence, is a separable metric space. Let Cb;+1;d (E ) + denote its countable dense subset. The set [0; 1]Cb;1;d (E ) with product topology is sequentially compact, so by a diagonal argument there W exists a subsequence nk such that the sequences f E f dnk ; k 2 N g converge for all f 2 Cb;+1;d (E ). Since Cb;+1;d (E ) is dense in Cb;+1 (E ), it follows by properties of idempotent integral (more speci cally, by W part (JS 5) of Theorem 1.4.4) that the sequences f E f dnk ; k 2 N g converge for all f 2 Cb;+1 (E ), which implies in analogy with the proof of Theorem 1.9.17 that there exists an F -idempotent probability iw on E such that nk ! as k ! 1. Let us now assume that E is a separable metric space. Then it is embedded into a compact metric space E 0 . We extend idempotent probabilities on E to idempotent probabilities on E 0 by letting 0 (A) = A0 \ E ; A0 E 0 : Let f0nk g be a subsequence of f0n ; n 2 N g which weakly converges to a deviability 0 on E 0 . Tightness of fn ; n 2 N g implies that the collection of compact subsets of E is a tightening collection for 0 so that the set function de ned by (A) = 0 (A); A E; is a tight F idempotent probability on E . Also 0 (E 0 n E ) = 0. We check that iw nk ! . Let f be a uniformly continuous function from Cb+ (E ) and f 0 be an element of Cb+(E 0 ) that extends f . By Theorem 1.9.2 W W 0 0 0 0 f d0 . SinceWthe 0nk and 0 are supported by E , E 0 f dnk ! EW we conclude that E f dnk ! E f d, which proves the required by Theorem 1.9.2. Now, if E is an arbitrary metric space, then by the tightness condition there exists a -compact metric space E 0 E such that n (E n E 0 ) = 0 for all n. Since E 0 is separable in relative topology, applying the part just proved to the restrictions 0n of the n to E 0 , we deduce existence of a subsequence f0nk g that weakly converges to a deviability 0 on E 0 . Let (A) = 0 (A \ E 0 ); A E . Then

© 2001 by Chapman & Hall/CRC

77

Spaces of idempotent probabilities

is a tight F -idempotent probability on E . Finally, if f 2 Cb+ (E ), then its restriction f 0 to E 0 belongsWto Cb+ (E 0 ). SinceW 0n (E n E 0 ) = 0 (E n E 0 ) = 0, we have that 0 0 = dnk and nk EfW E 0 f dW W W 0 0 0 0 0 0 f d = f d . Thus, convergence f d ! nk E0 E W E0 E 0 f d W yields convergence E f dnk ! E f d. A modi cation of the argument used in the proof of Lemma 1.9.24 allows us to obtain the following result. Theorem 1.9.28. Let f ; 2 g be a net of F -idempotent probabilities on a Tihonov space E that weakly converges to 2 IM(E ). Let G be a subset of Cb+ (E ) consisting of uniformly bounded and pointwise equicontinuous functions, i.e., supf 2G supz2E f (z ) < 1; and for every > 0 and z 2 E there exists an open neighbourhood Uz of z such that supf 2G supz0 2U jf (z ) f (z 0 )j : Then _ lim sup f d

f 2G E

_

E

z f d = 0:

Proof. We x > 0. For each z 2 E let Uz be as in the statement of the theorem. We show that the Uz can be assumed to be continuous relative to . Let fz be continuous functions with values in [0; 1] such thatfz (z ) = 1 and fz (z 0 ) = 0 on E n Uz . The function fz 1 ((x; 1]) ; x 2 [0; 1); is monotonically decreasing, so it has 1 (x; 1] is an open -continuous set if continuity points. Since f z fz 1((x; 1]) is continuous at x by -smoothness of , the claim has been proved. Let Uz1 ; : : : ; Uzk be such that E n[ki=1 Uzi < . Then, denoting a = supf 2G supz2E f (z ), we have _ f d

E

+ max

i=1;:::;k

_

E

f d

E n[ki=1 Uzi _ f d

_ f d

Uzi

_

f d +

_

E n[ki=1 Uzi

f d

Uzi

a E n [ki=1Uzi + a + 2 i=1 max sup jf (z ) f (zi )j ;:::;k z 2Uzi

+ max f (zi )j (Uzi ) i=1;:::;k

(Uzi )j:

Since by \the Portmanteau theorem" the latter maximum converges to 0 as 2 and lim sup2 E n[ki=1 Uzi E n[ki=1 Uzi < ; the proof is complete.

© 2001 by Chapman & Hall/CRC

78

Idempotent probability measures

If we replace space Cb+(E ) in the de nition of the weak topology by space CK+ (E ) of R+ -valued continuous functions with compact support, then we arrive at the de nition of the vague topology. However, to obtain nice properties, we have to consider the space of K{idempotent measures.

De nition 1.9.29. The vague topology on the set of K{idempotent

measures on a W topological space E is the coarsest topology for which the maps ! E h d are continuous for all h 2 CK+(E ).

If E is locally compact and Hausdor the vague topology has properties similar to the above properties of the weak topology. For instance, the space of K{idempotent measures is a Hausdor topological space and there is an easy analogue of Theorem 1.9.2. A distinguishing feature of the vague topology is that the space of K{ idempotent measures is compact.

Theorem 1.9.30. Let E be a locally compact Hausdor topologi-

cal space. Then the space of K{idempotent measures with the vague topology is compact.

The proof is similar to the proof of part 1 of Theorem 1.9.17, the main distinction being the use of Theorem 1.7.21 in place of Theorem 1.7.25. We end the section by indicating a connection between the vague and weak topologies, on the one hand, and Mosco convergence, on the other hand. The proof is straightforward.

Theorem 1.9.31.

I. Let E be a locally compact Hausdor topological space. Let be a K-idempotent probability and f ; 2 g be a net of K-idempotent probabilities on E . The following pairs of conditions are equivalent:

(M ) 1. for every z 2 E and net z ! z , lim sup2 (z ) (z ); 2. for every z 2 E there exists a net z ! z such that lim2 (z ) = (z ); (V ) 1. for every compact set K E , lim sup2 (K ) (K ); 2. for every open set G E , lim inf 2 (G) (G):

© 2001 by Chapman & Hall/CRC

Derived weak convergence

79

II. Let E be a Hausdor topological space. Let be a tight F idempotent probability and f ; 2 g be a tight net of F idempotent probabilities on E . Then the next pair of conditions is equivalent to both of the above:

(N ) 1. for every closed set F E , lim sup2 (F ) (F ); 2. for every open set G E , lim inf 2 (G) (G):

1.10 Derived weak convergence The results of this section give conditions for weak convergence of idempotent probabilities that are derived from weakly convergent idempotent probabilities. We also introduce convergence in idempotent distribution as an alternative way of viewing weak convergence of idempotent probabilities.

De nition 1.10.1. Let fX ; 2 g be a net of idempotent variables

de ned on respective idempotent probability spaces ( ; ) and assuming values in a topological space E and X be an idempotent variable de ned on an idempotent probability space ( ; ) and assuming values in E . Let the idempotent distributions of the X and X be F -idempotent probabilities on E . We say that the net fX ; 2 g iw converges in idempotent distribution to X if Æ X 1 ! Æ X 1. id We denote convergence in idempotent distribution by ! . Since convergence in idempotent distribution is the weak convergence of induced idempotent laws, the theory of Section 1.9 applies, e.g., there is a version of the Portmanteau theorem. Depending on the concrete situation it can be more convenient to formulate results on weak convergence of idempotent laws as convergence in idempotent distribution or vice versa.

Lemma 1.10.2. Let E be a Tihonov topological space, and and iw be F -idempotent probabilities on E such that ! . Let functions h : E ! R+ be uniformly bounded and a function h : E ! R+ be such that

lim h (z ) = h(z )

2

© 2001 by Chapman & Hall/CRC

80

Idempotent probability measures

for -almost every z 2 E and every net z ! z as 2 . Then

lim 2

_

E

h (z ) d (z ) =

_

E

h(z ) d(z ):

Proof. Let

h (z ) = inf sup sup h (z 0 );

(1.10.1)

U 2Uz z 0 2U 0

where Uz denotes the set of open neighbourhoods of z . Since the convergence condition in the hypotheses equivalently requires that for every z 2 E such that (z ) > 0 and > 0 there exist an open neighbourhood U of z and such that jh0 (z 0 ) h(z )j < for all z 0 2 U and 0 , we conclude that, given > 0 and z 2 E such that (z ) > 0, there exists such that h0 (z ) h(z ) + for all 0 . Therefore, introducing h(z ) = inf 2 h (z ), we have that h is upper semi-continuous and h(z ) h(z ) -a.e. Also, since the net fh g consists of upper semi-continuous bounded functions, is monotonically decreasing and converges to h, by Theorem 1.7.7 W W lim2 E h dW = E h dWso that for arbitrary > 0 there exists 0 such that E h0 d E h d + : Using the fact that h h0 ; 0 ; by (1.10.1) and Theorem 1.9.2 applied to h0 we obtain _

_

lim sup h (z ) d (z ) lim sup h0 (z ) d (z ) 2 E 2 E _ _ _ h0 (z) d(z) h(z) d(z) + h(z) d(z) + : E

E

E

The complementary inequality lim inf 2

_

E

h (z ) d (z )

_

E

h(z ) d(z )

is proved by a symmetric argument. Speci cally, we de ne

h (z ) = sup inf inf h (z 0 ); h(z ) = sup h (z ) U 2Uz z 0 2U 0 2 semi-continuous, h h0 if 0 , and note that the h are lower W W h h -a.e., and lim2 E h d = E h d. Therefore, for an

© 2001 by Chapman & Hall/CRC

81

Derived weak convergence

arbitrary > 0 and suitable 1 lim inf 2 _

_

E

h (z ) d (z ) lim inf _

2

_

E

h1 (z ) d (z )

h1 (z) d(z) h(z) d(z) E

E

_

E

h(z ) d(z ) :

As a consequence, we obtain the following version of the continuous mapping theorem on preservation of weak convergence of probability measures under mappings. We formulate it in terms of convergence in idempotent distribution. Theorem 1.10.3. Let E and E 0 be Tihonov topological spaces, and X and X be Luzin idempotent variables with values in E such that id X ! X . Let functions f : E ! E 0 , 2 , be Luzin measurable relative to the respective idempotent distributions of the X and f : E ! E 0 be Luzin measurable relative to the idempotent distribution of X . If for almost every z 2 E with respect to the idempotent distribution of X and every net z ! z we have that f(z ) ! f (z ), id then f Æ X ! f ÆX . Proof. Let and denote the respective idempotent distributions of X and X on E . Since the idempotent distribution of f Æ X is Æ f 1 and the idempotent distribution of f Æ X is Æ f 1 , for an R+ -valued bounded continuous function h on E 0 by a change of variables and Lemma 1.10.2 _ _ lim h(z 0 ) d Æ f 1 (z 0 ) = lim h Æ f (z ) d (z ) 2 0 2 E E _ _ = h Æ f (z ) d(z ) = h(z 0 ) d Æ f 1(z 0 ): E

E0

We thus have the following \continuous mapping theorem". Corollary 1.10.4. Let E and E 0 be Tihonov topological spaces, and id X and X be Luzin variables with values in E . If X ! X as 2 and f : E ! E 0 is Luzin measurable with respect to the distribution of X for every 2 and continuous a.e. with respect id to the distribution of X , then f Æ X ! f Æ X.

© 2001 by Chapman & Hall/CRC

82

Idempotent probability measures

We denote by Li (X ) the idempotent distribution of an idempotent variable X .

Lemma 1.10.5. Let E be a metric space with metric , and let X

and Y ; where 2 , 2 , and are directed sets, be nets of idempotent variables with values in E de ned on respective idempotent probability spaces ( ; ), whose idempotent distributions are F -idempotent probabilities on E . Let

lim lim sup (X ; Y ) " = 0; " > 0; 2 2 and

iw ~ Li X !

as 2 ,

~ ; 2 ; are F -idempotent probabilities on E . where Then, for an F -idempotent probability on E , we have that iw Li(Y ) ! as 2

if and only if

~

iw ! as 2 :

iw ~ ! Proof. We prove suÆciency of the condition so we assume that . Let F be a closed subset of E . Since

(Y 2 F ) (X 2 F )+ (X ; Y ) " ; by hypotheses ~ (F ) lim sup (Y 2 F ) lim sup 2 2 + lim sup lim sup ((X ; Y ) ") (F ); 2 2 and hence by the -smoothness property of lim sup (Y 2 F ) (F ): 2

© 2001 by Chapman & Hall/CRC

(1.10.2)

Derived weak convergence

Let G be an open subset of E . Then fX 2 G " g fY f(X ; Y ) "g; hence, since the G " are open as well,

83

2 Gg [

lim inf (Y 2 G) lim inf lim inf (X 2 G " ) 2 2 2 lim sup lim sup ((X ; Y ) ") (G " ): 2 2 Observing that [">0 G " = G so that by -maxitivity (G " ) ! (G) as " ! 0, we conclude that lim inf (Y 2 G) (G); 2 which together with (1.10.2) ends the proof of the suÆciency part. The converse is proved in an analogous manner. The following special case is useful. Given a net fZ ; 2 g of idempotent variables de ned on ( ; ) and assuming values in a metric space E with metric , we write that Z ! z 2 E , if lim2 (Z ; z ) > = 0 for every > 0. Lemma 1.10.6. Let E be a metric space with metric , and let X and Y ; where 2 , be nets of idempotent variables on ( ; ) with values in E , whose idempotent distributions are F -idempotent iw probabilities on E . If Li (X ) ! , where is an F -idempotent iw probability on E , and (X ; Y ) ! 0 as 2 , then Li (Y ) ! . Lemma 1.10.7. Let fX ; 2 g be a net of idempotent variables and X be an idempotent variable. Let all the variables be de ned on ( ; ), assume values in a metric space E and have F -idempotent id probabilities on E as distributions. If X ! X , then X ! X . If id X ! z 2 E , then X ! z . Proof. The rst property follows by convergence properties of idempotent integrals (Theorem 1.4.19). For the second one, let denote the metric on E , X denote the idempotent distribution of X , and 1z denote the unit mass at z. Then the convergence X !id z implies that _ _ lim 1^(z 0 ; z ) dX (z 0 ) = 1^(z 0 ; z ) d 1z (z 0 ) = 0: 2 E E

© 2001 by Chapman & Hall/CRC

84

Idempotent probability measures

We now consider joint convergence. Lemma 1.10.8. Let X and Y ; 2 , be nets of idempotent variables on respective idempotent probability spaces ( ; ) with values in Tihonov spaces E and E 0 , respectively. Let X and Y be idempotent variables on an idempotent probability space ( ; ) with values in E and E 0 , respectively. Let the idempotent distributions of the X , Y ; X , and Y be -smooth relative to the associated collections of closed sets. Let E E 0 be equipped with product topology. id id 1. If X ! X, Y ! Y , X and Y are independent, and X and id Y are independent, then (X ; Y ) ! (X; Y ).

id 2. Let E and E 0 be metric spaces. If X ! X and Y ! z , then id (X ; Y ) ! (X; z ). Proof. We prove part 1. Let h(z; z 0 ); (z; z 0 ) 2 E E 0 ; be an R+ valued bounded function that is uniformly continuous with respect to a uniformity on E E 0 . By Theorem 1.9.28 lim sup sup h(z; z 0 ) ÆY 1 (z 0 ) sup h(z; z 0 )ÆY 1 (z 0 ) = 0: 2 z 2E z 0 2E 0 z 0 2E 0 (1.10.3) Since supz0 2E 0 h(z; z 0 ) Æ Y 1 (z 0 ) is continuous in z 2 E , 1 lim sup sup h(z; z 0 ) Æ Y 1 (z 0 ) Æ X (z ) 2 z 2E z 0 2E 0 = sup sup h(z; z 0 ) Æ Y 1 (z 0 ) Æ X 1 (z ): (1.10.4) z 2E z 0 2E 0 Equations (1.10.3) and (1.10.4) imply the required. For part 2 we observe that ( 0 )((X ; Y ); (X ; z )) ! 0, where 0 is a product metric on E E 0 , so that the required follows by part 1 and Lemma 1.10.6.

1.11 Laplace-Fenchel transform This section introduces idempotent analogues of the characteristic function and standard probability distributions, and develops related techniques. Let ( ; ) be an idempotent probability space. We denote the idempotent distribution Æf 1 of an idempotent variable f : ! Rd by f .

© 2001 by Chapman & Hall/CRC

85

Laplace-Fenchel transform

Remark 1.11.1. Clearly, given an idempotent probability ^ on Rd ,

we can always construct an idempotent variable whose idempotent ^ . This is the \canonical" idempotent variable f (x) = distribution is x. We refer to representations like this as \canonical settings".

We recall that for d-dimensional vectors x and y we denote as x y the inner product.

De nition 1.11.2. Given f : ! Rd , the Laplace-Fenchel transform of f is the R + -valued function Lf () = Sef (!) ; 2 Rd :

By Theorem 1.4.6 we can also write _ Lf () = ex df (x): Rd

Remark 1.11.3. Note that ln Lf () isf the convex conjugate, or the Legendre-Fenchel transform, of

ln (x) in that

ln Lf () = sup ( x +ln f (x)): x2Rd

Lemma 1.11.4.

An R + -valued function L(); 2 Rd ; is the Laplace-Fenchel transform of an Rd -valued idempotent variable if and only if ln L() is convex and lower semi-continuous, and L(0) = 1. Proof. Necessity of the conditions follows from the de nition of the Laplace-Fenchel transform and properties of convex conjugates. Conversely, let ln L() be convex and lower semi-continuous, and L(0) = 1. We de ne (x) by

ln (x) = sup ( x ln L()); 2 Rd : 2Rd Then is an idempotent probability and L is the Laplace-Fenchel transform of the canonical variable on (Rd ; ) by properties of convex conjugates, Rockafellar [117, x26]. According to the above proof, we can recover f from Lf by the equality f (x) = infd e x Lf () (1.11.1) 2R

© 2001 by Chapman & Hall/CRC

86

Idempotent probability measures

if we know, in addition, that ln f (x) is lower semi-continuous and convex. We refer to (1.11.1) as the inversion formula. It is useful, however, to have conditions for the inversion formula to hold that are expressed only in terms of the properties of Lf . We recall some notions of convex analysis, Rockafellar [117]. Let a function g(); 2 Rd ; assume values in ( 1; 1]. The domain of g as de ned as dom g = f 2 Rd : g() < 1g and the function g is said to be essentially smooth if the following conditions hold, Rockafellar [117, x26], (a) int(dom g) is not empty, (b) g is dierentiable on int(dom g), (c) limk!1 jrg(k )j = 1 whenever fk g is a sequence of elements of int(dom g) converging to a boundary point of int(dom g). We also denote as ri A the relative interior of a set A.

Lemma 1.11.5. Let Lf (); 2 Rd ; be essentially smooth. Then the inversion formula holds.

Proof. Let (x) denote the right-hand side of (1.11.1). Since ln (x); x 2 Rd ; is the convex conjugate of ln Lf () and the latter is convex and lower semi-continuous, it follows that ln Lf () is the convex conjugate of ln (x), Rockafellar [117, x26]. Since ln Lf () is essentially smooth, we conclude that ln (x) is essentially strictly convex, Rockafellar [117, Theorem 26.3], hence, strictly convex on ri(dom ln ). Since also ln (x) is the bipolar of ln f (x), the two functions coincide by Lemma A.1 in Appendix A.

Remark 1.11.6. As the proof shows, one can weaken the requirement of essential smoothness of Lf to the requirement that the convex conjugate of ln Lf () be strictly convex on the relative interior of its domain.

We have a simple corollary, which shows that the Laplace-Fenchel transform can help us to identify Luzin idempotent variables.

Lemma 1.11.7. If Lf (); 2 Rd ; is essentially smooth and 0 2 int(dom Lf ), then f is a Luzin idempotent variable.

© 2001 by Chapman & Hall/CRC

87

Laplace-Fenchel transform

Proof. By the inversion formula f (x) is upper semi-continuous being the in mum of continuous functions of x so that by Lemma 1.7.4 f is a K-idempotent probability. By \the Chebyshev inequality" for a > 0 and > 0 such that the -ball about the origin in Rd belongs to int(dom Lf )

f (jxj a) = sup f (x a) e 2Rd : jj=

a

sup Lf ():

2Rd : jj=

The latter supremum being nite by continuity of Lf () on int(dom Lf ), the right-most side tends to 0 as a ! 1. Thus, f is tight. The following property provides us with a means of proving independence of idempotent variables on . For idempotent variables f1 : ! Rd1 and f2 : ! Rd2 , we denote by Lf1 ;f2 (1 ; 2 ); 1 2 R d1 ; 2 2 Rd2 ; the Laplace-Fenchel transform of (f1 ; f2 ) : ! R d1 +d2 .

Lemma 1.11.8.

1. If f1 and f2 are independent, then Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ):

2. Let Lf1 () and Lf2 () be essentially smooth. If Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ), then f1 and f2 are independent. Proof. The rst part follows by de nition. The second part follows by Lemma 1.11.5 for denoting by f1 ;f2 (x1 ; x2 ) the joint idempotent distribution of f1 and f2 , we have by the lemma since Lf1 ;f2 (1 ; 2 ) = Lf1 (1 )Lf2 (2 ) is essentially smooth as well

f1 ;f2 (x1 ; x2 ) = infd e 1 x1 2 x2 Lf1 ;f2 (1 ; 2 ) 1 2R 1 ; 2 2Rd2 x = infd e 1 1 Lf1 (1 ) infd e 2 x2 Lf2 (2 ) 1 2R 1 2 2R 2 = f1 (x1 )f2 (x2 ):

Corollary 1.11.9. Let A be a -algebra on and f : ! Rd .

If S exp( f )jA (!) is constant for -almost ! and is an essentially smooth function of 2 Rd , then f is independent of A.

© 2001 by Chapman & Hall/CRC

88

Idempotent probability measures

Proof. From the hypotheses and properties of conditional idempotent expectations, for 1 2 Rd ; 2 2 R and A 2 A,

S exp(1 f ) exp(2 1(A)) = S exp(1 f )jA S exp(2 1(A)) = S exp(1 f )jA exp(2 )(A) _ (Ac ) :

The required follows by Lemma 1.11.8. We now introduce idempotent analogues of standard probability distributions by requiring that their Laplace-Fenchel transforms be identical to the Laplace transforms of their probabilistic counterparts.

De nition 1.11.10. We say that f : !d Rd is idempotent Gaussian with parameters (m; ), where m 2 R and is a d d positive semi-de nite symmetric matrix, if Lf () = exp m + =2 .

Remark 1.11.11. Below we occasionally refer to m as the idempotent mean and as the idempotent covariance of an idempotent Gaussian variable f .

The next lemma follows by Lemma 1.11.5.

Lemma 1.11.12. An idempotent variable f : !f Rd is idempotent

Gaussian with parameters (m; ) if and only if (x) = exp (x m) (x m)=2 if x m is in the range of and f (x) = 0 otherwise.

The following is a consequence of the de nition.

Lemma 1.11.13. An idempotent variable f : ! Rd is idempotent Gaussian if and only if f :

every 2 Rd .

!R

is idempotent Gaussian for

De nition 1.11.14. We say that f : ! R + is idempotent Poisson with a parameter > 0 if Lf () = exp (e

1) ; 2 R.

An application of Lemma 1.11.5 yields the idempotent distribution of the Poisson idempotent variable.

Lemma 1.11.15.

An idempotent variable f : ! R+ is idempotent Poisson with a parameter > 0 if and only if f (x) = exp x ln(x=) + x ; x 2 R+ , where 0 ln 0 = 0.

© 2001 by Chapman & Hall/CRC

89

Laplace-Fenchel transform

Remark 1.11.16.

By Lemma 1.11.7 both Gaussian and Poisson idempotent variables are Luzin idempotent variables. We now apply the Laplace-Fenchel transform to limit theorems. The following result is \an idempotent law of large numbers". Lemma 1.11.17. Let ffi; i 2 N g be a sequence of independent identically distributed Rd -valued idempotent variables on an idempotent probability space ( ; ) such that S exp( f1 ) < 1 for from a neighbourhood of the origin. If (jf1 j > ) < 1 for every > 0, then n 1X f ! 0 as n ! 1: n i=1 i

Proof. It is suÆcient to check that for arbitrary Æ > 0 and arbitrary 2 Rd such that S exp( f1 ) < 1 n 1 X lim f > j j Æ = 0: n!1 n i=1 i

(1.11.2)

Since by \the Chebyshev inequality" for 2 (0; 1] n 1 X

n i=1

fi > jjÆ

P

Sef1 n S exp ni=1 fi = jjÆ ; exp(njjÆ) e

the required would follow if there exists 2 (0; 1] such that Sef1 < 1: (1.11.3) ejjÆ We have for > 0 Sef1 = Sef1 1(j f1 j > jj) _ Sef1 1(j f1 j jj) Sef1 1(j f1j > jj) _ ejj: (1.11.4) Let < Æ. Then the second term on the right-most side of (1.11.4) is less than ejjÆ . For the rst term we have in view of the hypotheses that Sef1 1(j f1 j > jj) lim = (jf1 j > jj) < 1: !0 ejjÆ Thus, (1.11.3) holds for > 0 small enough, which concludes the proof.

© 2001 by Chapman & Hall/CRC

90

Idempotent probability measures

Remark 1.11.18. Thus, an analogue of the expectation of a random

variable for an idempotent variable f in the law of large numbers is an element a such that (jf aj > ) < 1 for every > 0. The following lemma is an analogue of the method of characteristic functions in weak convergence theory. Lemma 1.11.19. Let f ; 2 g be a net of deviabilities on Rd . If W x d (x) ! L() as 2 for all 2 Rd , where L() is essen Rd e tially smooth, lower semi-continuous and such that 0 2 int(dom L), iw then ! , where deviability is given by the inversion formula (1.11.1). W Proof. Let us denote L () = Rd ex d (x). We show that the net f ; 2 g is tight. Indeed, since L () ! L() and L() is nite in a neighbourhood of the origin by the fact that 0 2 W int (dom L), there exists r > 0 such that lim sup Rd erjxj d (x) < 1; and then \the Chebyshev inequality" (fx : jxj > Ag) W e rA Rd erjxj d (x) yields the claim. Therefore, by Theorem 1.9.17 f ; 2 g has accumulation points in IMt (Rd ). Let f0 ; 0 2 0 g be a subnet of f ; 2 g that weakly converges to ~ 2 IMt (Rd ). Then the convergence L0 () ! L() implies that if 2 int(dom L), then for suitable W " > 0 lim sup0 Rd e(1+")x d0 (x) < 1: Hence, the function (ex ; x 2WRd ) is uniformly maximable with respect to f0 ; 0 2 0 g, W x x ~ x); implying that so lim0 Rd e d0 (x) = Rd e d( _ ~ x) = L() ex d( (1.11.5) Rd

for all 2 int(dom L). We prove that (1.11.5) actually holds for all 2 Rd . Let L~ () denote the left-hand side of (1.11.5). Since L() is essentially smooth, it follows that jrL~ (n )j ! 1 for every sequence n in int(dom L) that converges to a boundary point of dom L. Since L~ () is convex, it follows that L~ () = 1 for all 2= cl(dom L), which implies that L~ () = L() for 2= cl(dom L). Finally, if is a boundary point of dom L, then lower semi-continuity and convexity of L() imply that L() = limn L(n ), where n is a sequence of points from int(dom L) converging to . For the same reason this holds for L~ (), so we conclude that L~ () = L() for all 2 Rd . Then ~ = , which ends the proof. Lemma 1.11.5 implies that

© 2001 by Chapman & Hall/CRC

Chapter 2

Maxingales In this chapter we develop elements of idempotent stochastic calculus. We are mostly interested in studying idempotent analogues of martingales and martingale problems (which we call maxingales and maxingale problems, respectively).

2.1 Idempotent stopping times In this section we de ne stopping times with respect to -algebras and study their properties. The concepts, results and proofs are analogous to those in the general theory of stochastic processes, see Dellacherie [34] or Meyer [88]. Therefore, we omit proofs that are analogous to proofs in these books. Let be a set.

De nition 2.1.1. An indexed collection A = fAt; t 2 R+ g of -

algebras on is called a ow of -algebras if As At for s t: We also refer to a ow of -algebras as a - ow. We say that the - ow T A is right-continuous if As = t>s At for all s 2 R+ .

Remark 2.1.2. Recall that As At if and only if the atoms of As

are unions of the atoms of At . Remark 2.1.3. Given a - ow A, there is a natural rightcontinuous - ow associated with A that is de ned by A+ = (At+ ; t 2 R+ ), where At+ = \>0 At+.

We assume as given a ow of -algebras A and a -algebra A1 such that As A1 for s 2 R+ . 91 © 2001 by Chapman & Hall/CRC

92

Maxingales

De nition 2.1.4. An R + -valued function on is said to be an idempotent stopping time relative to A, or, for short, an A-stopping

time, if f sg 2 As for all s 2 R+ .

Lemma 2.1.5.

1. A function : ! R + is an A-stopping time if and only if f = sg 2 As for all s 2 R+ (in particular, constants are stopping times); if A is right-continuous, an equivalent condition is that f < sg 2 As for all s 2 R+ .

2. If and are A-stopping times, then _ , ^ and + are A-stopping times.

3. If ; 2 , are A-stopping times, then sup 2 is an A-stopping time; if A is right-continuous, then inf 2 , lim inf 2 and lim sup 2 (in the latter two cases is a directed set) are A-stopping times.

We now introduce -algebras associated with stopping times. Let us rst prove two simple facts. Recall that [!]A denotes the atom of a -algebra A about !. Lemma 2.1.6. Let be an A-stopping time. If !00 2 [!0]A(!0) , then (!00 ) = (!0 ) and [!0 ]A(!0 ) = [!00 ]A(!00 ) . Proof. Since !0 2 f! : (!) (!0 )g 2 A(!0 ) , [!0 ]A(!0 ) is an atom of A(!0 ) , and !00 2 [!0 ]A(!0 ) , it follows that !00 2 f! : (!) (!0 )g so that

(!00 ) (!0 ):

(2.1.1)

Therefore, [!00 ]A(!00 ) [!00 ]A(!0 ) = [!0 ]A(!0 ) . Thus, !0 2 [!00 ]A(!00 ) . The argument of the proof of (2.1.1) with roles of !0 and !00 switched then shows that we actually have equality in (2.1.1). Equality of the atoms in the statement is now self-evident. The following lemma is an easy consequence.

Lemma 2.1.7. Let be an A-stopping time.

Then the collection

f[!]A(!) ; ! 2 g is a partition of in that two arbitrary sets from

the collection are either disjoint or coincide and the union of the sets from the collection equals .

© 2001 by Chapman & Hall/CRC

93

Stopping times

De nition 2.1.8. Let be an A-stopping time. We denote by A the -algebra that has atoms [!]A(!) .

The next lemma shows that our de nition is consistent with the corresponding de nition in the general theory of stochastic processes.

Lemma 2.1.9. The -algebra A is the collection of subsets A of

such that A 2 A1 and A \ f sg 2 As for all s 2 R+ .

Proof. Clearly, A A1 . Let us show that [!]A(!) \f sg 2 As . Let !0 2 [!]A(!) \ f sg. Then by Lemma 2.1.6 [!0 ]As [!0 ]A(!0 ) = [!]A(!) . Also [!0 ]As f sg. Therefore, by Corollary 1.1.16 [!]A(!) \ f sg 2 As . Conversely, let A 2 A1 be such that A \ f sg 2 As for all s 2 R+ and let ! 2 A. We prove that A 2 A by proving that [!]A (!) A. If (!) < 1, then the required follows since ! 2 A \ f (!)g 2 A(!) so that [!]A (!) A \ f (!)g. If (!) = 1, then [!]A(!) is an atom of A1 and the required follows since A 2 A1.

Remark 2.1.10. The -algebra A can also be de ned as the col-

lection of subsets A of such that A \ f = sg 2 As for all s 2 R + .

Lemma 2.1.11.

measurable.

1. Let be an A-stopping time. Then is A -

2. Let be an A-stopping time and Then is an A-stopping time.

be

A -measurable.

3. Let and be A-stopping times such that . Then is a stopping time relative to the - ow fA+t ; t 2 R+ g.

4. Let and be A-stopping times such that . Then A A . 5. Let and be A-stopping times. Then the sets f < gT, f = g and fT > g belong both to A and A . Also A f = g = A f = g.

Let, given t > 0, At denote the -algebra generated by the algebras As for s < t. It is obvious that the atoms of At are of the form [!]At = \s t, we have that [!]A(!) [!]At . Therefore, [!](!) A. Since ! 2 f > tg 2 At , we have that [!]A(!) [!]At f > tg. The claim is proved. Thus, A^ A . The reverse inclusion follows since [!]A(!) = \t 0, and X0 (!) = 0. Then Xt is continuous at t = 0 for every ! 2 but it is not continuous in idempotent probability.

In the sequel we refer to idempotent processes with continuous (right-continuous, respectively) paths as continuous (rightcontinuous, respectively) idempotent processes. The following simple fact is useful.

Lemma 2.2.11. If X0 is a proper idempotent variable and X

is continuous in idempotent probability, then X is a proper idempotent process.

© 2001 by Chapman & Hall/CRC

99

Idempotent processes

We now introduce the class of Luzin-continuous processes, which is smaller than the class of processes continuous in idempotent probability. Let C (R + ; Rd ) denote the space of Rd -valued continuous functions on R+ equipped with the metric supst jxs ys j ^ 1 ; 1+t t2R+

dC (x; y) = sup

where x = (xt ; t 2 R+ ) and y = (yt ; t 2 R+ ). We recall that with this metric C (R + ; Rd ) is a complete separable metric space.

De nition 2.2.12. We say that an idempotent process X with Xcon-

tinuous trajectories is Luzin-continuous if the restriction of to C (R + ; R d ) is a deviability on C (R + ; R d ). Let, in addition, be a Hausdor topological space and be a deviability on . An idempotent process X is called a strictly Luzin-continuous idempotent process on ( ; ) if the mapping ! ! (Xt (!); t 2 R+ ) from to C (R + ; R d ) is a strictly Luzin idempotent variable.

Theorem 2.2.13.

A Luzin (respectively, strictly Luzin) idempotent process X is Luzin-continuous (respectively, strictly Luzincontinuous) if and only if for arbitrary T > 0 and > 0

lim sup jXt Xs j > = 0: Æ!0 s;t2[0;T ]: js tjÆ Proof. According to Corollary 1.8.7 a Luzin idempotent process X is Luzin-continuous if and only if inf K 2K (X 62 K ) = 0, where K is the class of compact subsets of C (R + ; Rd ). By a standard argument based on Arzela-Ascoli's theorem this is equivalent to the convergences

lim jX0 j > A = 0; lim A!1 Æ!0

sup jXt Xsj > = 0; s;t2[0;T ]: js tjÆ

where T > 0 and > 0 are arbitrary. The rst condition is a consequence of X0 being Luzin. The second condition is equivalent to the condition in the statement by -maxitivity of . The proof for strictly Luzin idempotent processes is similar. We now consider measurability issues in the spirit of the general theory of stochastic processes. We assume the discrete -algebra on

© 2001 by Chapman & Hall/CRC

100

Maxingales

unless otherwise speci ed. Let A = (At ; t 2 R+ ) be a ow of -algebras on . De nition 2.2.14. We say that an idempotent process X is Aadapted if Xt is At -measurable for every t 2 R+ . Let B [0; t] At be the product of the Lebesgue -algebra on [0; t] and the -algebra At in the sense of De nition 1.5.9. We refer to elements of B [0; t] At as progressively measurable sets. De nition 2.2.15. An idempotent process X is said to be progressively measurable (or A-progressively measurable) if the mappings (s; !) ! X (s; !) from [0; t] to Rd are B [0; t] At =B(Rd )measurable for all t 2 R+ . Lemma 2.2.16. An idempotent process X = (Xt (!); t 2 R+ ; ! 2 ) is A-progressively measurable if and only if it is A-adapted and the functions (Xt (!); t 2 R+ ) are B(R+ )=B(Rd )-measurable in t for all ! 2 . Proof. Let X be A-progressively measurable. Then, given t2 R+ and x 2 Rd , we have f(s; !) : Xs (!) = x;s 2 [0; t]g 2 B [0; t] At so that ftg f! : Xt (!) = xg 2 B [0; t] At , which implies that f! : Xt(!) = xg 2 At . Thus, X is A-adapted. Next, given !^ and a Borel subset of Rd , we have that f(s; !) : Xs (!) 2 g \ [0; t] [^!]At 2 B [0; t] At : Since Xs (!) = Xs (^!) for s 2 [0; t] if ! 2 [^!]At by A-adaptedness of X and Corollary 1.2.5, we conclude that fs 2 [0; t] : Xs (^!) 2 g [^!]At 2 B [0; t] At so that fs 2 [0; t] : Xs (^!) 2 g 2 B [0; t] . Necessity is proved. Conversely, since Rd

f(s; !) : Xs(!) 2 g =

\

[0; t]

[

!2

\

fs : Xs (!) 2 g [0; t] f!g ;

fs : Xs(!) 2 g \ [0; t] 2 B [0; t] for ! 2 , and Xs(!) = Xs(!0 ) if s 2 [0; t] and ! At !0 , the de nition of B [0; t] At implies suÆciency.

Lemma 2.2.17. Let X beR an A-progressively measurable idempotent t process. Let the integrals 0 XsR(!) ds, t 2 R+ ; ! 2 ; be well de ned. Then the idempotent process 0t Xs (!) ds; t 2 R+ is A-adapted. © 2001 by Chapman & Hall/CRC

101

Idempotent processes

Proof. Since X is A-adapted, 0t Xs (!) ds is constant on the atoms of At , so it is At -measurable by Corollary 1.2.5. R

The next result is a version of Lemma 2.1.18.

Lemma 2.2.18. Let X be A-adapted and D Rd . Let D = inf ft 2 R+

: Xt 2 Dg. Let either one of the conditions hold:

1. X is right-continuous and D is closed,

A is right-continuous. Then D is an A-stopping time. 2.

Proof. Under the rst condition, we have by right-continuity of X and closedness of D [

f! : D (!) tg = f! : Xs(!) 2 Dg 2 At: st

The second part follows by Lemma 2.1.18. We can also adapt the proof of the lemma by writing [

f! : D (!) < tg = f! : Xs(!) 2 Dg 2 At: s 0, and t 2 R+

(t X = xjpt X = x0 ) = (t X = x):

2. An A-adapted Luzin idempotent process X on ( ; ) has Aindependent increments if and only if for x 2 (Rd )R+ and t 2 R + -a.e.

(t X = xjAt ) = (t X = x):

De nition 2.2.25. An idempotent process X = (Xt ; t 2 R+ ) is said

to be idempotent Gaussian if its nite-dimensional distributions are idempotent Gaussian.

The next two theorems consider convergence in idempotent distribution for Luzin-continuous idempotent processes. We state the results in the form of weak convergence of associated deviabilities. The following tightness theorem is an obvious consequence of ArzelaAscoli's theorem.

Theorem 2.2.26. A net f ; 2 g of deviabilities on C (R + ; Rd )

is tight if and only if 1:Æ lim lim sup (fx : jx0 j > Ag) = 0, A!1 2:Æ lim lim sup (fx : sup jxt xs j > g) = 0; > 0; T > 0: Æ!0 s;t2[0;T ]: js tj 0. In particular, x^ is absolutely continuous and x^ 0 = 0. By the de nition of conditional idempotent expectation

SW (Yt jCs )(^x) = sup Yt (x)W (xjCs )(^x):

x2C

(2.4.4)

The conditional idempotent probability W (xjCs )(^x) is not equal to zero only if x and x^ belong to the same atom of Cs , i.e., ps x = ps x^ . For these x, W (xjCs )(^x) =

W (x) W (x) = : W ([^x]Cs ) supx0 : ps x0 =psx^ W (x0 )

(2.4.5)

Easy calculationsusing (2.4.1) show that the latter supremum equals Rs 2 ^_ u du=2 ; so, assuming that W (x) > 0 and x is thus exp 0x absolutely continuous and x0 = 0, by (2.4.5) and the equality xu = x^ u ; u s; 1 Z1 W _ 2u du : (xjCs )(^x) = exp x 2 s

Recalling that s x = (xs+u ps x = psx^ , then W (xjCs )(^x) = W (sx)

© 2001 by Chapman & Hall/CRC

x^ s; u 2 R+ ), we thus have that if -a.e.

117

Wiener and Poisson idempotent processes

Therefore, by the de nition of Yt (x), (2.4.4) and (2.4.3)

SW (Yt jCs )(^x) = sup Yt (x)W (xjCs )(^x)

x2C : ps x=ps x^

i 2 s h 2 (t s) W sup exp (xt x^ s) (s x) 2 x2C 2 W = Ys (^x) sup Yt s (sx) (s x) = Ys(^x)SW Yt s (x) = Ys(^x):

= exp x^ s

x2C

x) follows by Corollary 1.4.15 and the equal-

W -maximability of Yt (

ities

SW (Yt2 ) = exp(2 t)SW (2xt 22 t) = exp(2 t); where the latter equality follows by (2.4.2) and (2.4.3) with replaced by 2. Also, x0 = 0 W -a.e. by the de nition of W . This ends the proof of the implication 1 ! 3. We prove that part 3 implies part 2. The maxingale property of Y easily implies that 1 SW (exp((xt xs))jCs ) = exp 2 (t s) ; 2 which yields the required by De nition 1.11.10, Corollary 1.11.9 and Lemma 1.11.12. We prove that part 2 implies part 1. By independence of increments the nite-dimensional idempotent distributions of W are Gaussian so by Remark 1.11.16 they are deviabilities on the associated spaces. Therefore, by Corollary 2.2.6 the idempotent distribution of W is -smooth relative to the collection of closed subsets of R R+ . Thus, for x 2 R R+ such that x0 = 0 by Theorem 2.2.2 and independence of increments of W (! : W (!) = x) = t inf (! : Wti (!) = xti ; i = 1; : : : ; k) ;:::;t 1

k

k Y

= t inf (! : Wt1 (!) = xt1 ) ;:::;t 1

k

= xti

i=2

xti 1 ) = exp

(! : Wti (!) Wti 1 (!) k X (xti 1 sup 2 t1 ;:::;tk i=1 ti

xti 1 )2 : ti

1

The Rlatter supremum equals +1 if x is not absolutely continuous and 01 x_ 2t dt if x is absolutely continuous.

© 2001 by Chapman & Hall/CRC

118

Maxingales

The next result shows that the Wiener idempotent process is Luzin-continuous.

Lemma 2.4.3. The Wiener idempotent probability is a deviability, i.e., it is tight and -smooth relative to the collection of closed subsets of C (R + ; R). Proof. By Corollary 1.7.15 it would be suÆcient to prove that W (x) is an upper compact function of x 2 C (R + ; R), which is an easy exercise. We give, however, a dierent \probabilistic" proof based on Theorem 2.4.2. Let X be the canonical idempotent process on (C (R + ; R); W ). By Theorem 2.2.13 it suÆces to prove that for arbitrary T > 0 and > 0

lim sup W jXt Xsj > = 0: Æ!0 s;t2[0;T ]: js tjÆ By Theorem 2.4.2 and the Chebyshev inequality for > 0

W jXt Xs j > = W Xt Xs > _ W Xs Xt > SW exp (Xt Xs ) SW exp (Xs Xt ) _ exp() exp() exp 2 jt sj=2 = ; exp()

which implies the required since is arbitrary.

De nition 2.4.4. We say that a continuous idempotent process W is Wiener relative to a - ow A (or an A-Wiener idempotent process for short) if the idempotent process M () de ned in the statement of Theorem 2.4.2 is an A-exponential maxingale starting at 1 for every 2 R.

Lemma 2.4.5. A continuous idempotent process W is A-Wiener if and only if the idempotent process M () de ned in the statement of Theorem 2.4.2 is an A-local exponential maxingale starting at 1 for every 2 R.

Proof. Let M () be an A-local exponential maxingale starting at 1 for every 2 R. By Lemma 2.3.13 we only need to prove that for

© 2001 by Chapman & Hall/CRC

Wiener and Poisson idempotent processes

119

every s 2 R+ the process (Mt^s (); t 2 R+ ) is uniformly maximable. As in the proof of Theorem 2.4.2

S Mt^s ()2

S Mt^s (2) exp(2s) exp(2 s);

where the latter inequality follows by Lemma 2.3.13. The uniform maximability follows by Corollary 1.4.15. Obviously, if M () is an A-exponential maxingale, then it is an AW -exponential maxingale. Thus, we have the following consequence of Theorem 2.4.2.

Corollary 2.4.6. If W is an A-Wiener idempotent process, then it has properties described in parts 1 and 2 of Theorem 2.4.2.

Corollary 2.4.7. If W Wt

is an A-Wiener idempotent process, then Ws is independent of As for t s.

We now consider the multi-dimensional case.

De nition 2.4.8. 1 d

An Rd -valued idempotent process W = (W ; : : : ; W ) on ( ; ) is called a d-dimensional Wiener idempotent process if the processes W 1 ; : : : ; W d are independent Wiener idempotent processes.

The next theorem is proved similarly to Theorem 2.4.2. Let, as above, AW denote the ow of -algebras generated by W and W denote the idempotent distribution of W . Let Ed denote the d d identity matrix.

Theorem 2.4.9. Let W

be an Rd -valued idempotent process. The following conditions are equivalent: 1. W is a d-dimensional Wiener idempotent process,

2. the density of W is given by (x 2 RR+ ) 8 1 Z1 > > 2 ds ; > > _ j x j if x is absolutely exp s > < 2 continuous and 0 W (x) = > > > x0 = 0; > > : 0; otherwise,

© 2001 by Chapman & Hall/CRC

120

Maxingales

3. W is an idempotent process with independent increments, W0 = 0 -a.e., and increments Wt Ws are idempotent Gaussian with parameters (0; (t s)Ed ) so that

(Wt Ws = x) = exp

jxj2

2(t s)

; x 2 Rd ;

4. the idempotent process M () = (Mt (); t 2 R+ ) de ned by 1 Mt () = exp Wt jj2 t ; 2 is an AW -exponential maxingale such that M0 () = 1 -a.e. for every 2 Rd .

Clearly, a d-dimensional idempotent Wiener process is Luzincontinuous. Also, since W has independent increments, which are idempotent Gaussian variables, we have the following corollary.

Corollary 2.4.10. A d-dimensional idempotent Wiener process is an idempotent Gaussian process.

Part 4 of Theorem 2.4.9 implies the following analogues of the properties of the Wiener process. Let e = (t; t 2 R+ ). We recall that Æ denotes the composition map.

Corollary 2.4.11. 1. Let W be a one-dimensional idempotent Wiener process and 2 R+ . Then the idempotent process W Æ (e) has the same idempotent distribution as 1=2 W .

2. Let W1 ; : : : ; Wk be independent d-dimensional idempotent Wiener processes and 1 ; :P : : ; k be l d matrices. Then the idempotent distribution of ki=1 i Wi coincides with the idemPk T 1=2 W , where W is an lpotent distribution of i=1 i i dimensional idempotent Wiener process.

De nition 2.4.12. An Rd -valued continuous idempotent process W is called a d-dimensional Wiener idempotent process with respect to a - ow A (or A-Wiener, for short) if the idempotent process M () de ned in the statement of Theorem 2.4.9 is an A-exponential maxind gale starting at 1 for every 2 R .

The proof of the following lemma is similar to the proof of Lemma 2.4.5.

© 2001 by Chapman & Hall/CRC

Wiener and Poisson idempotent processes

121

Lemma 2.4.13. An Rd -valued idempotent process W is a ddimensional A-Wiener idempotent process if and only if the idempotent process M () de ned in the statement of Theorem 2.4.9 is an A-local exponential maxingale starting at 1 for every 2 Rd .

We de ne now the Poisson idempotent process. As above, we assume that 0 ln 0 = 0. De nition 2.4.14. We say that an idempotent probability N on C (R + ; R ) is the Poisson idempotent probability if it has density de ned by 8 Z1 > > > > _ _ _ exp ( x ln x x + 1) ds ; if x is absolutely > s s s > > > < continuous, 0 N (x) = x_ s 2 R+ a.e. > > > > > and x0 = 0, > > > :0; otherwise. (2.4.6) A Poisson idempotent process N = (Nt ; t 2 R+ ) on ( ; ) is an idempotent process with paths from C (R + ; R) and idempotent distribution N , i.e., (N = x) = N (x); x 2 C (R + ; R). We call N a canonical Poisson idempotent process if it is the canonical process on (C (R + ; R); N ). Remark 2.4.15. According to the de nition, the Poisson idempotent process has increasing paths -a.e. We now give a characterisation of the Poisson idempotent process in the spirit of Watanabe's characterisation of the Poisson process and analogous to that of the Wiener idempotent process. Let AN = N (AN t ; t 2 R+ ), where At denotes the -algebra on generated by the maps ! ! Ns(!) for s t. Theorem 2.4.16. Let N be an R-valued idempotent process. The following statements are equivalent: 1. N is a Poisson idempotent process, 2. N is an idempotent process with independent increments, N0 = 0 -a.e., and increments Nt Ns for s < t are idempotent Poisson with parameters t s, i.e., x (Nt Ns = x) = exp x ln +x (t s) ; x 2 R+ ; t s

© 2001 by Chapman & Hall/CRC

122

Maxingales

3. the idempotent process M () = (Mt (); t 2 R+ ) de ned by

Mt () = exp Nt (e 1)t

is an AN -exponential maxingale such that M0 () = 1 -a.e. for every 2 R. Proof. The proof is analogous to the proof of Theorem 2.4.2. We prove that part 1 implies part 3. Let N be a Poisson idempotent process. As in the proof of Theorem 2.4.2 it is suÆcient to prove that the idempotent process Y = (Yt (x); t 2 R+ ; x 2 C (R + ; R)) de ned by

Yt (x) = exp xt (e 1)t

(2.4.7)

SN Yt (x) = 1:

(2.4.8)

is a C(R+ ; R)-exponential maxingale on (C (R + ; R); N ). We rst note that by (2.4.6)

Let x^ 2 C (R + ; R) be such that N (^x) > 0. In particular, x^ is absolutely continuous, increasing and x^ 0 = 0. By the de nition of conditional idempotent expectation SN (Yt jCs )(^x) = sup Yt (x)N (xjCs )(^x): (2.4.9) x2C By the reasoning used in the proof of Theorem 2.4.2 we have that N (xjCs )(^x) = N (s x):

Therefore, by the de nition of Yt (x), again repeating the argument of the proof of Theorem 2.4.2, SN (Yt jCs )(^x) = sup Yt (x)N (xjCs )(^x) x2C : ps x=ps x^ = Ys (^x) sup Yt s (s x)N (s x) = Ys (^x)SN Yt s (x) = Ys (^x): x2C N -maximability of Yt (x) follows by Corollary 1.4.15 and the equality

SN (Yt2 ) = exp (e2

© 2001 by Chapman & Hall/CRC

2e + 1)t SN 2xt (e2 1)t = exp (e2 2e + 1)t ;

123

Wiener and Poisson idempotent processes

where the latter equality follows by (2.4.7) and (2.4.8) with replaced by 2. Finally, Y0 (x) = 1 N -a.e. since x0 = 0 -a.e. This ends the proof of the implication 1 ! 3. We prove that part 3 implies part 2. The maxingale property of Y yields

SN (exp((xt

xs))jCs ) = exp (e

1)(t s) ;

which implies the required by Corollary 1.11.9 and Lemma 1.11.15. To prove that part 2 implies part 1 we write for an increasing function x 2 RR+ such that x0 = 0 by Theorem 2.2.2 and independence of increments of N (! :

(! : Nti (!) = xti ; i = 1; : : : ; k) N (!) = x) = t1inf ;:::;tk

= t inf (! : ;:::;t 1

= xti

k

Nt1 (!) = xt1 )

xti 1 ) = exp

sup

k Y

(! :

i=2 k X

t1 ;:::;tk i=1

(xti

(xti

Nti (!) Nti 1 (!)

xti 1 ) ln xttii txiti1 1

xti 1 ) + (ti

ti 1 ) :

The Rlatter supremum equals +1 if x is not absolutely continuous and 01 (x_ t ln x_ t x_ t + 1) dt if x is absolutely continuous. The next result shows that N is a Luzin-continuous idempotent process.

Lemma 2.4.17. The Poisson idempotent probability is a deviability, i.e., it is tight and -smooth relative to the collection of closed subsets of C (R + ; R). Proof. We give again a \probabilistic" proof. Let X be the canonical process on (C (R + ; R); N ). By Theorem 2.2.13 it suÆces to prove that for arbitrary T > 0 and > 0

lim sup N Xt Xs > = 0: Æ!0 s;t2[0;T ]: 0t sÆ

© 2001 by Chapman & Hall/CRC

124

Maxingales

By Theorem 2.4.16 and the Chebyshev inequality for > 0

S N exp (Xt Xs ) Xs > exp() exp (e 1)(t s) ; = exp() which implies the required since is arbitrary. N Xt

De nition 2.4.18. We say that a continuous idempotent process N is Poisson relative to a - ow A (or A-Poisson idempotent process for short) if the idempotent process M () de ned in the statement of Theorem 2.4.16 is an A-exponential maxingale for every 2 R such that M0 () = 1 -a.e.

Corollary 2.4.19. If N is an A-Poisson idempotent process, then

it has properties described in parts 1 and 2 of Theorem 2.4.16. Also, Nt Ns is independent of As for s t.

Lemma 2.4.20. An R-valued continuous idempotent process N is A-Poisson if and only if the idempotent process M () de ned in the statement of Theorem 2.4.16 is an A-local exponential maxingale starting at 1 for every 2 R.

2.5 Idempotent stochastic integrals Let ( ; ) be an idempotent probability space with a ow of algebras A = (At ; t 2 R+ ).

De nition 2.5.1. An Rd -valued continuous A-adapted idempotent

process M = (Mt ; t 2 R+ ) such that M0 = 0 is called a local maxingale (maxingale or uniformly maximable maxingale, respectively) with a quadratic characteristic hM i relative to A (or an A-local maxingale, A-maxingale, or uniformly maximable Amaxingale, respectively, for short) if there exists an Rdd -valued continuous A-adapted idempotent process hM i = (hM it ; t 2 R+ ) such that hM i0 = 0, hM it hM is for 0 s t are positive semi-de nite symmetric d d matrices and the idempotent process (exp( Mt hM it =2); t 2 R+ ) is an A-local exponential maxingale (respectively, A-exponential maxingale, uniformly maximable Aexponential maxingale) for every 2 Rd .

© 2001 by Chapman & Hall/CRC

125

Idempotent stochastic integrals

By Theorem 2.4.9 an Rd -valued Wiener idempotent process is a local maxingale with a quadratic characteristic Ed t. Lemma 2.4.13 yields the following converse.

Corollary 2.5.2. Let a continuous idempotent process M be an Rd valued A-local maxingale with a quadratic characteristic (Ed t; t 2 R + ). Then M is a d-dimensional A-Wiener idempotent process. The following consequence of Lemma 2.3.14 is also useful.

Lemma 2.5.3. Let a continuous idempotent process M be an Rd valued A-local maxingale with a quadratic characteristic hM i. Then for a > 0, b > 0, c > 0, and nite A-stopping times and such

that

( sup jMt M j a) ec(b t

a) _( khM i

hM i k > 2b=c):

Proof. By \the Doob stopping theorem" the idempotent process exp (Mt+ M ) (hM it+ hM i )=2 ; t 2 R+ ; 2 Rd ; is a supermaxingale relative to the - ow (At+ ; t 2 R+ ). Also, is a stopping time relative to (At+ ; t 2 R+ ) by Lemma 2.1.11. Hence, by Lemma 2.3.14 and -maxitivity of

( sup jMt M j a) = sup ( sup (Mt M ) a) t jj=1 t ec(b a) _ sup ( (hM i hM i ) > 2b=c) jj=1 = ec(b a) _ ( khM i hM i k > 2b=c): Properties of the trajectories of hM i are often translated into the corresponding properties of M . The next result follows by Lemma 2.5.3 and Theorem 2.2.13.

Lemma 2.5.4. Let a continuous idempotent process M be a local maxingale relative to the ow A with a quadratic characteristic hM i. 1. If hM i is a proper idempotent process, then M is a proper idempotent process.

© 2001 by Chapman & Hall/CRC

126

Maxingales

2. If hM i is continuous (respectively, stopping-time-rightcontinuous) in idempotent probability, then M is continuous (respectively, stopping-time-right-continuous) in idempotent probability. 3. Let M be Luzin. If hM i is Luzin-continuous, then M is Luzincontinuous.

We assume in the rest of the section that the - ow A is complete in the sense of the following de nition.

De nition 2.5.5.

We say that a ow of -algebras on ( ; ) is complete if the -algebras in the ow are complete with respect to .

Remark 2.5.6. Clearly, if A = (At ; t 2 R+ ) is a - ow, then A = (At ; t 2 R+ ), where the At are the completions of the At with respect to , is a complete - ow. We refer to it as the completion of A with respect to (or the -completion of A). The next lemma extends Lemma 1.2.6.

Lemma 2.5.7. Let A be the completion of A with respect to . If is an A -stopping time, then there exists an A-stopping time 0 0

such that = -a.e. Proof. We de ne 0 as follows: if (!) > 0, then 0 (!) = (!); if (!) = 0 and there exists !~ such that (~!) > 0 and ! 2 [~!]A(~!) , then 0 (!) = (~! ); if (!) = 0 and no such !~ exists, then 0 (!) = 1. We rst check that 0 is well de ned. Indeed, suppose for some ! such that (!) = 0 there exist !~ and !^ such that (~!) > 0, (^! ) > 0, ! 2 [~!]A(~!) , ! 2 [^!]A(^!) , and (~!) (^! ); then [~!]A(~!) = [!]A(~!) [!]A(^!) = [^!]A(^!) ; hence, !~ 2 [^!]A(^!) since (~!) > 0 and (^!) > 0, so (~! ) = (^! ) by Lemma 2.1.6 proving the claim. We check that 0 is an A-stopping time. Let 0 (!0 ) = t and 00 ! 2 [!0 ]At , where t 2 R+ . We have to check that 0 (!00 ) = t. If (!0 ) > 0 and (!00 ) > 0, then 0 (!0 ) = (!0 ), 0 (!00 ) = (!00 ) and !00 2 [!0 ]At ; hence, (!0 ) = t so !00 2 [!0 ]A(!0 ) and (!0 ) = (!00 ) by Lemma 2.1.6. If (!0 ) > 0 and (!00 ) = 0, then as above (!0 ) = t, so !00 2 [!0 ]A(!0 ) and by de nition 0 (!00 ) = (!0 ). If (!0 ) = 0, then, since 0 (!0 ) is nite, there exists !~ such that (~!) > 0, !0 2 [~!]A(~!) and 0 (!0 ) = (~!); hence, !00 2 [~!]A(~!) . If, in addition,

© 2001 by Chapman & Hall/CRC

127

Idempotent stochastic integrals

(!00 ) > 0, then !00 2 [~!]A(~!) so by Lemma 2.1.6 0 (!00 ) = (~!); if (!00 ) = 0, then the latter equality holds by de nition. We say that an Rmd -valued continuous idempotent process X is absolutely continuous if the entry processes have -a.e. absolutely continuous with respect to Lebesgue measure trajectories; if X is, in addition, A-adapted, then we denote by X_ an A-progressively measurable idempotent process such that X_ (!) is a version of the RadonNikodym derivative of X (!) with respect to Lebesgue measure for almost all !. (For instance, we could de ne X_ t (!) as the left derivative of X (!) at t if the latter exists and let X_ t (!) = 0 otherwise.) For a quadratic characteristic hM i to be absolutely continuous, it is actually suÆcient that the diagonal Rentries are absolutely continuous. Note that if this is the case, then 0t khM_ isk ds < 1; t 2 R+ : The following lemma shows, in particular, that absolute continuity of the quadratic characteristic of a local maxingale implies absolute continuity of the local maxingale itself. It also lays the groundwork for the de nition of idempotent integrals with respect to local maxingales. We recall that denotes the pseudo-inverse of a matrix .

Lemma 2.5.8. Let an Rd -valued continuous idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i. Then the following holds.

1. M is absolutely continuous. 2. M_ s (!) belongs to the range of hM_ is (!) for almost all s and -almost all !, and h

Z1

S exp

0

i 1 hM_ is ) ds 2 h 1 Z1 i = S exp M_ s hM_ is M_ s ds 1; 2

sup ( M_ s 2Rd

0

R in particular, 01 M_ s idempotent variable.

© 2001 by Chapman & Hall/CRC

hM_ is M_ s ds

is a -a.e. nite proper

128

Maxingales

Let, in addition, (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued progressively measurable idempotent process such that Zt

0

A-

ks hM_ issT k ds < 1; t 2 R+ ; -a.e.

Then Zt

jsM_ sj ds < 1; t 2 R+ ; -a.e.

0

Proof. Let, for 0 s1 t1 : : : sk tk and i 2 Rd ; i = 1; : : : ; k, k hX

Z = exp

i=1

i Mti Msi

k 1X (hM iti 2 i=1 i

We show that

SZ 1:

i

hM isi )i : (2.5.1) (2.5.2)

Let

n = inf ft 2 R+ :

khM it k ng:

(2.5.3)

By Lemma 2.2.18 the n are A-stopping times. Also by Lemma 2.3.13

S exp( Mt^n hM it^n =2)2 S exp(2 Mt^n (2) hM it^n (2)=2) exp(njj2 ) exp(njj2 ) so that by Lemma 2.3.13 fn g is a localising sequence of stopping times for every local maxingale exp( Mt hM it =2); t 2 R+ , 2 Rd . Let 1 Yni = exp i Mti ^n Msi ^n i (hM iti ^n hM isi ^n )i : 2 (2.5.4)

© 2001 by Chapman & Hall/CRC

129

Idempotent stochastic integrals

Since (exp(i Mt^n i exponential maxingale, S (Yni jAsi ) = 1: By (2.5.4) and (2.5.5) we have

S

k Y i=1

Yni = S

kY1

i=1

hM it^n i=2); t 2

R+ )

is an

A-

(2.5.5)

Yni S (Ynk jAsk ) = S Q

kY2

i=1

Yni S (Ynk

1

jAsk 1 )

= : : : = SYn1 = 1:

Since n ! 1, by (2.5.1) Z = limn!1 ki=1 Yni , and \the Fatou lemma" (see Theorem 1.4.19) yields (2.5.2). This inequality implies in view of the de nition of idempotent expectation that k X

k 1X sup i Mti Msi i (hM iti hM isi )i 2 i=1 f(si ;ti )g;fi g i=1 < 1 -a.e. (2.5.6) Now let us suppose that there exists ! 2 such that (!) > 0 and T > 0 such that Mt (!) is not absolutely continuous on [0; T ]. Then there exists " > 0 such that for every Æ > 0 there exist nonoverlapping subintervals f(sÆi ; tÆi )g of [0; T ] such that X

i

(tÆi sÆi ) < Æ and

X

i

jMtÆi (!) MsÆi (!)j > ":

(2.5.7)

Let Æi be such that jÆi j = 1 and Æi (MtÆi (!) MsÆi (!)) = jMtÆi (!) MsÆi (!)j. Given A > 0 we choose Æ > 0 such that X khM itÆi (!) hM isÆi (!)k < A12 ; i

which is possible in view of absolute continuity of hM i. Then by (2.5.7) k X i=1

(AÆi ) MtÆi (!) MsÆi (!)

k 1X (AÆ ) (hM itÆi (!) 2 i=1 i

© 2001 by Chapman & Hall/CRC

hM isÆi (!))(AÆi ) A" 12 ;

130

Maxingales

which contradicts (2.5.6) since A is arbitrary. Part 1 is proved. For part 2, we note that by properties of idempotent expectation and (2.5.2)

sup Z = sup SZ 1: f(si ;ti )g;fi g f(si ;ti )g;fi g By Lemma A.2 in Appendix A the supremum on the left-most side R equals 01 sup2Rd ( M_ s hM_ is =2) ds. We thus have i h Z1 1 S exp sup ( M_ s hM_ is ) ds 1: 2 2Rd

S

0

The supremum in the integral is nite if and only if M_ s is orthogonal to the nullspace of hM_ is , which is equivalent to M_ s being in the range of hM_ is. It is then equal to M_ s hM_ is M_ s =2, which proves the rst claim of part 2. The second one is an obvious consequence of the rst. For the nal assertion, we note that since M_ s (!) belongs to the range of hM_ is(!) for almost all s and -almost all !, we have that M_ s (!) = hM_ is(!)hM_ is (!)M_ s (!) for almost all s and -almost all !. Therefore, by the Cauchy-Schwarz inequality -a.e. Zt

jsM_ sj ds

0

Zt

0

kshM_ issT k ds

1=2 Zt

0

1=2

M_ s hM_ is M_ s ds

;

where the right-hand side is nite -a.e. by hypotheses and the part of the lemma already proved.

De nition 2.5.9. Let idempotent processes M and be as in the statement of Lemma 2.5.8. An idempotent process X = (Xt (!); t 2 R + ; ! 2 ) de ned by

Xt (!) =

8 t Z > >

0;

> 0 > :~

Xt (!);

if (!) = 0;

where X~ t (!) is a continuous idempotent process, is called an idempotent stochastic integral of with respect to M and denoted by Rt _ M = 0 s Ms ds; t 2 R+ .

© 2001 by Chapman & Hall/CRC

131

Idempotent stochastic integrals

In particular, if M is a d-dimensional Wiener idempotent process Rt 2 W and 0 ks k ds < 1; t 2 R+ ; -a.e., the integral W is called an idempotent Ito integral.

Clearly, an idempotent stochastic integral is a continuous Aadapted process and is speci ed uniquely -a.e. We show that under certain conditions on the integrands idempotent stochastic integrals are local maxingales with quadratic characteristics. We begin with an approximation lemma.

Lemma 2.5.10. Let an Rd -valued continuous A-adapted idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i. Let (sk (!); s 2 R+ ; ! 2 ); k 2 N ; and (s (!); s 2 R+ ; ! 2 ) be A-progressively measurable Rmd -valued idempotent processes such that for t 2 R+ Zt

0

kshM_ issT k ds < 1; Zt

0

Zt

0

ksk hM_ is(sk )T k ds < 1;

k(sk s)hM_ is(sk s)T k ds ! 0 as k ! 1: (2.5.8a)

If the idempotent processes Rk M are A-local maxingales with the quadratic characteristics 0t sk hM_ is (sk )T ds; t 2 R+ , then the idempotent process R M is an A-local maxingale with the quadratic characteristic 0t shM_ is sT ds; t 2 R+ .

Proof. The idempotent processes M and k M are well de ned by Lemma 2.5.8. The idempotent process M is A-adapted by Lemma 2.2.17 and completeness of A. We introduce, for 2 Rm ,

1 2

Zt () = exp ( M )t

Ztk () = exp (k M )t

1 2

Zt

0

Zt

0

s hM_ is sT ds ; (2.5.9a)

sk hM_ is(sk )T ds : (2.5.9b)

The idempotent processes (Ztk (); t 2 R+ ) are A-local exponential maxingales by hypotheses. We have to prove that the idempotent

© 2001 by Chapman & Hall/CRC

132

Maxingales

process (Zt (); t 2 R+ ) is an A-local exponential maxingale. Let us note that by Lemma 2.5.8

SZt () 1; SZtk () 1; 2 Rm :

(2.5.10)

Let us introduce for n 2 N

n = inf ft 2 R+ : Zt

nk = inf ft 2 R+ :

0

Zt

0

kshM_ issT k ds ng; (2.5.11a)

ksk hM_ is(sk )T k ds n + 1g ^ n:(2.5.11b)

We show that the idempotent processes (Zt^n (); t 2 R+ ) are uniformly maximable. By (2.5.9a), (2.5.11a) and (2.5.10)

S Zt^n

()2 = S

h

Zt^n (2) exp

tZ^n

0

Thus,

sup S Zt^n ()2 < 1; t2R+

shM_ is sT ds

i

exp jj2n : (2.5.12)

proving uniform maximability of (Zt^n (); t 2 R+ ) by Corollary 1.4.15. A similar argument shows that S Ztk^nk ()2 exp jj2 (n + 1) : Therefore, the collection fZtk^nk (); k 2 N ; t 2 R+ g is uniformly maximable; in particular, the (Ztk^nk (); t 2 R+ ); k 2 N ; are uniformly maximable exponential maxingales. By Lemma 1.6.22 it thus suÆces to prove that Ztk^nk () ! Zt^n () as k ! 1. Since (2.5.12) implies that Zt^n () is a proper idempotent variable, it follows by the de nitions (2.5.9a) and (2.5.9b) that the required is a consequence of the convergences as k ! 1 tZ^nk 0

(k M )t^nk sk hM_ is (sk )T ds !

© 2001 by Chapman & Hall/CRC

! ( M )t^n ; (2.5.13a)

tZ^n 0

s hM_ issT ds: (2.5.13b)

133

Idempotent stochastic integrals

Let us rst note that in view of (2.5.11a), (2.5.11b) and (2.5.8a) we have that

lim t ^ nk 6= t ^ n = 0: k!1 Limit (2.5.13a) follows now by the inequalities j (k M )t^nk

( M )t^n j > Z t

+

0 Zt

0

j (sk

s)M_ s j ds

Zt

0

0

t ^ nk 6= t ^ n

1=2

(sk s)hM_ is(sk s)T ds Zt

M_ s hM_ is M_ s ds A

j (sk s)M_ s(!)j ds > " ;

0 Z t

(2.5.14)

1=2

M_ s hM_ is M_ s ds

;

exp( A=2);

and (2.5.8a). Similarly, limit (2.5.13b) follows by (2.5.14), (2.5.8a), (2.5.11a), and the inequality tZ^n

0

sk hM_ is(sk )T ds

tZ^n

2

0

tZ^n

0

(sk

0

s)hM_ is(sk

shM_ is sT ds

s hM_ is sT ds 1=2

s)T ds

1=2

+

tZ^n 0

© 2001 by Chapman & Hall/CRC

tZ^n

(sk

s)hM_ is (sk

s )T ds:

134

Maxingales

Theorem 2.5.11. Let an Rd -valued continuous A-adapted idempotent process M be an A-local maxingale with an absolutely continuous quadratic characteristic hM i, which is a proper idempotent process. Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively measurable idempotent process such that for t 2 R+ Zt

Zt

0

0

kshM_ is sT k ds < 1;

ks hM_ issT k 1( ks k > A) ds ! 0

as A ! 1:

Let there exist functions nA : R+ ! [0; 1], where A 2 R+ , such that nA (x) = 1 if x A, nA(x) A=x if x A, and for all A large enough

lim Æ!0

Z t

0

k snA( ks k ) s Æ nA( ks Æ k ) hM_ is

s nA( ks k ) s Æ nA ( ks

T

Æk )

k ds > = 0; t 2 R+ ; > 0;

where s (!) = 0 for s < 0. Then the idempotent stochastic integral M is an A-local maxingale with the quadratic characteristic Zt

h M it = shM_ issT ds: 0

The latter is a proper idempotent process. Proof. We have to prove that Z () = (Zt (); t 2 R+ ), 2 Rm , de ned as in (2.5.9a) is an A-local exponential maxingale. It is Aadapted by Lemma 2.2.17 and completeness of the - ow A. Let us rst consider the case that

s (!) =

k X i=1

fi (!) 1(ti

© 2001 by Chapman & Hall/CRC

1 ;ti ] (s);

135

Idempotent stochastic integrals

where 0 = t0 < t1 < : : : < tk and the fi are Ati 1 -measurable and bounded Rmd -valued idempotent variables. Then k hX

Zt () = exp

i=1

fi Mti ^t

Mti 1 ^t

i 1 fi (hM iti ^t hM iti 1 ^t )fiT : (2.5.15) 2 Let n be de ned by (2.5.3). By Lemma 2.5.8 S [exp( Mt^n hM it^n =2)] 1, which implies as in the proof of Lemma 2.5.10 that the idempotent process (exp( Mt^n hM it^n =2); t 2 R+ ) is uniformly maximable. Hence, by Lemma 2.3.13 (exp( Mt^n hM it^n =2); t 2 R+ ) is a uniformly maximable A-exponential maxingale. Let

Yti = exp fi Mti ^t^n Mti 1 ^t^n 1 T fi (hM iti ^t^n hM iti 1 ^t^n )fi : (2.5.16) 2 Since fi is Ati 1 -measurable, Lemma 1.6.21 implies that S (Yti jAti 1 ) = 1:

(2.5.17)

By (2.5.15), (2.5.16), and (2.5.17) we have

SZt^n () = S

k Y i=1

=S

Yti = S

kY2

i=1

kY1

i=1

Yti S (Ytk

Yti S (Ytk jAtk 1 ) 1

jAtk 2 ) = : : : = SYt1 = 1:

Since the latter holds for all 2 Rm , we have by (2.5.15) and (2.5.3) that

S Zt^n (2) exp(jj2 Bnk) = exp(jj2 Bnk); where B is an upper bound for the kfi k 2 . By Corollary 1.4.15 we conclude that the idempotent process (Zt^n (); t 2 R+ ) is uniformly S Zt^n ()2

maximable.

© 2001 by Chapman & Hall/CRC

136

Maxingales

The maxingale property of (Zt^n (); t 2 R+ ) is checked similarly. Let s t. If s tk , then Zs^n () = Zt^n () and the maxingale property trivially holds. Let ti 1 < s ti for some i. Since by the argument used above S (Zt^n ()jAti ) = Zti ^n (), it follows that

S (Zt^n ()jAs ) = S (S (Zt^n ()jAti )jAs ) = S (Zti ^n ()jAs ) = Zti 1 ^n ()S (Ytii jAs): By (2.5.16), since fi is As-measurable, S (Ytii jAs ) = exp fi (Ms^n Mti 1 ^n ) 1 fi (hM is^n hM iti 1 ^n )fiT 2 S exp fi(Mti ^n Ms^n ) 1 fi (hM iti ^n hM is^n )fiT jAs 2 = exp fi (Ms^n Mti 1 ^n ) 1 fi (hM is^n hM iti 1 ^n )fiT ; 2 where the latter equality follows by the maxingale property of exp( Mt^n hM it^n =2) and Lemma 1.6.21. Putting everything together, we conclude that S (Zt^n ()jAs ) = Zs^n (): Let us now assume that s is bounded and locally continuous in s uniformly on , i.e., wT (Æ) = sup sup ks (!) t (!)k ! 0 as Æ ! 0, T > 0: !2 s;tT js tjÆ

R Let us rst note that 0t khM_ is k ds is a proper idempotent variable since the integral is not greater than the sum of the diagonal entries of hM it and the latter is a proper idempotent variable. Let

sk

=

2

k X i=1

(i

1)=k

1(s 2 ((i

© 2001 by Chapman & Hall/CRC

1)=k; i=k]):

137

Idempotent stochastic integrals

Then by the part just proved the Z k (), de ned as Z () with s changed to sk , are A-local exponential maxingales. We also have that for k t Zt

0

Zt

k(s sk )hM_ is(s sk )T k ds wt (1=k)2 khM_ isk ds 0

Rt 0

so that by Lemma 2.5.10 and the fact that khM_ is k ds is a proper idempotent variable we conclude that Z () is an A-local exponential maxingale. Let us assume now that s in the statement of the theorem is bounded. We introduce the Steklov functions

sk

=k

Zs

u du:

s 1=k

Since ksk (!) tk (!)k 2k supu;! ku (!)k jt sj; the functions sk are continuous in s uniformly over !. They are also bounded and properly measurable so that by the part just proved the associated idempotent processes Z k () are A-local exponential maxingales. We again apply Lemma 2.5.10 to deduce that Z () is an A-local exponential maxingale. We have Zt

0

=

k(s sk )hM_ is(s sk )T k ds Zt

0

Z1=k

Z1=k

k k (s s u) du hM_ is k (s s u)T du k ds

Zt

0 Z1=k

0

0

k

0

k(s s u)hM_ is(s s u)T k du ds sup

0u1=k

Zt

0

k(s s u)hM_ is(s s u)T k ds:

The latter supremum converges in idempotent probability to 0 as k ! 1 by hypotheses.

© 2001 by Chapman & Hall/CRC

138

Maxingales

Finally, if s in the statement of the theorem is not bounded, we de ne sk = nk ( ks k )s : Then by hypotheses Zt

0

k(s sk )hM_ is(s sk )T k ds Zt

0 kshM_ issT k 1( ks k k) ds ! 0

as k ! 1, and since associated with the k idempotent processes Z k () are A-local exponential maxingales by the part already proved, Lemma 2.5.10 implies that Z () is an A-local exponential maxingale. The fact that h M i is a proper idempotent process is obvious. For idempotent Ito integrals we have the following corollary.

Theorem 2.5.12.

Let W be an Rd -valued A-Wiener idempotent process. Let (s (!); s 2 R+ ; ! 2 ) be an RmR d -valued Aprogressively measurable idempotent process such that 0t ks k 2 ds < 1 and for a function nA as in Theorem 2.5.11 Zt

ksnA( ks k ) s+Æ nA( ks+Æ k )k 2 ds ! 0

0 Zt

0

as Æ ! 0; t 2 R+ ;

for all A large enough, ksk 2 1( ks k > A) ds ! 0

as A ! 1; t 2 R+ :

Then W is an A-local maxingale with the quadratic characteristic Zt

h W it = ssT ds; 0

which is a proper process.

Remark 2.5.13. R imply that

t 0 ks

The convergence conditions in Theorem 2.5.12 s+Æ k 2 ds ! 0 as Æ ! 0. Therefore, by M.

© 2001 by Chapman & Hall/CRC

139

Idempotent stochastic integrals

Riesz's criterion for relative compactness in L2 , see, e.g., Kantorovich and Akilov [70], under the hypotheses the idempotent distribution of ! ! (s (!); s t) is a deviability in L2 ([0; t]; R md ) for all t 2 R+ .

We now prove that in analogy with stochastic calculus under certain regularity conditions local maxingales with quadratic characteristics are idempotent Ito integrals. We adapt notation (1.7.2) to denote

K (a) = f! 2 : (!) ag; a 2 (0; 1]:

Theorem 2.5.14. Let an Rd -valued continuous A-adapted idempotent process M be an A-local Rmaxingale with an absolutely continut T ous quadratic characteristic ( 0 s s ds; t 2 R+ ), which is a proper idempotent process, where the s are d d matrices. If for t 2 R+ and a 2 (0; 1] Zt

ks s+Æ k 2 ds ! 0

as Æ ! 0;

0

inf inf inf s (!)sT (!) > 0;

!2K (a) st 2Rd : jj=1

then there exists a d-dimensional such that M = W .

A-Wiener idempotent process W

Proof. We de ne a continuous idempotent process W = 1 M; where the right-hand side is well de ned by Lemma 2.5.8. Clearly, M = W . Let nA (x) = 1(x A), where x 2 R+ and A 2 R+ . Then, denoting s 1 = s = 0 for s < 0, we have Zt

0

k s 1Æ nA( ks 1Æ k ) s 1nA( ks 1 k ) ssT T

s 1Æ nA ( ks 1Æ k ) s 1 nA( ks 1 k )

=

Zt

0

k ds

ks 1Æ snA( ks 1Æ k ) s Æ nA( ks 1 k )

s nA( ks 1Æ k ) s Æ nA ( ks 1 k ) T (sT Æ )

© 2001 by Chapman & Hall/CRC

1

k ds

140

Maxingales

Zt

sup kssT k 1 st sup kssT k 1 st

0

ksnA( ks 1Æ k ) s Æ nA( ks 1 k )k 2 ds

15

Zt

ks s Æ k 2 ds

0

+ 10

Zt

0

ksk 2 1( ks 1 k > A) ds :

By hypotheses the latter converges in idempotent probability to 0 as Æ ! 0 for all large A. Hence, by Theorem 2.5.11 W is an A-local maxingale with the quadratic characteristic (Ed t; t 2 R+ ); hence, W is a Wiener idempotent process by Corollary 2.5.2.

Remark 2.5.15. One can replace the in mum over s above by the essential in mum with respect to Lebesgue measure. This remark also concerns other conditions of a similar sort. We now consider versions for strictly Luzin idempotent processes de ned on Hausdor topological spaces with deviabilities. We start with a lemma on properties of trajectories. Lemma 2.5.16. Let be a dHausdor topological space and be a deviability on . Let an R -valued continuous A-adapted strictly Luzin idempotent process M be an A-local maxingale on ( ; ) with an absolutely continuous quadratic characteristic hM i such that (hM_ is ; s 2 R+ ) is a strictly Luzin idempotent process and for t 2 R+ Zt

0

khM_ isk 1( khM_ isk > A) ds ! 0

as A ! 1:

Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively meaRt T _ surable strictly Luzin idempotent process such that 0 ks hM is s k ds < 1 and Zt

0

ks hM_ issT k 1( ks k > A) ds ! 0

R Then both M and ( 0t s hM_ is sT ds; t continuous idempotent processes.

© 2001 by Chapman & Hall/CRC

as A ! 1; t 2 R+ :

2 R+ ) are strictly Luzin-

141

Idempotent stochastic integrals

Proof. We rst check that M is strictly Luzin. Let ! 2 K (a), where a 2 (0; 1], and ! ! !0 as 2 . We write for t 2 R + and A 2 R+ , denoting kA ( ) = (A + 1 k k )+ ^ 1 for a matrix ,

k Mt (!) Mt (!0 )k 2 sup

!2K (a)

Zt

s (!) kA (s (!)) M_ s (!) ds

0 t Z Zt

+ kA (s (! ))M_ s (! ) ds kA (s (!0 ))M_ s (! ) ds

0 0 t Z Zt

0 _ + kA (s (! ))Ms (! ) ds kA (s (!0 ))M_ s (!0 ) ds

: 0 0

(2.5.18)

We estimate the rst term on the right-hand side of (2.5.18) as

Zt

0

s (!) kA (s (!)) M_ s (!) ds

Z t

0

1=2

ks(!)hM_ is(!)s(!)T k 1( ks (!)k > A) ds Zt

1=2

M_ s (!) hM_ is (!) M_ s (!) ds

ds;

0

which converges to 0 as A ! 1 uniformly over ! 2 K (a) by Lemma 2.5.8 and the hypotheses. For the second term on the right-hand side of (2.5.18) we write

Zt

_ s (! ) ds

kA (s (! ))M 0

© 2001 by Chapman & Hall/CRC

Zt

0

kA (s (!0 ))M_ s (! ) ds

142

Maxingales

Z t

kA (s (! )) kA (s (!0 )) hM_ is (! )

0

T 1=2

ds

kA (s (! )) kA (s (!0 )) Zt

1=2

M_ s (! ) hM_ is (! ) M_ s (! ) ds

:

0

The second multiplier on the right is bounded in by Lemma 2.5.8. The integrand in the rst one can be estimated for B 2 R+ as

kA (s (! )) kA (s (!0 )) hM_ is (! ) kA (s (! )) kA (s (!0 )) T

B kA(s(!)) kA(s(!0 ))k2 + 4(A + 1)2 khM_ is (! )k 1( khM_ is (! )k > B );

which implies by hypotheses and Lebesgue's dominated convergence theorem that the second term on the right-hand side of (2.5.18) converges to 0 as 2 . Finally, the third term on the right-hand side of (2.5.18) converges to 0 as 2 by the following argument: since hM is (! ) ! Rt Rt 0 _ hM is (! ), it follows that f M (! ) ds ! 0 fsM_ s (!0 ) ds for step Rt 0 s s functions fs; since the 0 M_ s (! )hM_ is (! ) M_ s (! ) ds are uniformly Rt bounded and 0 M_ s (!0 ) hM_ is (!0 ) M_ s (!0 ) ds is nite, by the CauchyShwarz inequality the class of functions fs for which the latter convergence holds is closed under bounded pointwise convergence; so a monotone class argument shows that it contains all bounded Borelmeasurable functions fs (see the proof of (2.7.28) in the proof of Lemma 2.7.17 below for a more detailed argument of this sort). Now, in order to check that M is strictly Luzin-continuous it is suÆcient to show that uniformly over ! 2 K (a), where a 2 (0; 1], the functions ( Mt (!); t 2 R+ ) are uniformly continuous in t 2

© 2001 by Chapman & Hall/CRC

143

Idempotent stochastic integrals

2 R+ . We have for s; t 2 [0; T ] k Mt (!) Ms(!)k 2

[0; T ] for every T Zt

ku(!)hM_ iu(!)u (!)T k du s

Zt

s

M_ u (!) hM_ iu (!) M_ u (!) du:

The second term on the right is bounded on K (a) by Lemma 2.5.8. The rst term is not greater for A 2 R+ and B 2 R+ than Zt

s

ku (!)hM_ iu(!)u (!)T k 1( ku (!)k > A) du + A2

Zt

s

khM_ iu(!)k 1( khM_ iu(!)k > B ) du + A2 B (t s);

which implies the required in view of the hypotheses. R The proof of ( 0t s hM_ is sT ds; t 2 R+ ) being strictly Luzincontinuous uses similar ideas. Let ! ! !0 , where ! 2 K (a). Then by hypotheses and Lebesgue's dominated convergence theorem for t 2 R+ , A 2 R+ , B 2 R+ , and 2 Rd lim

Zt

0

kA (s (! ))kB (hM_ is (! ))kA (sT (! )) ds =

Zt

0

Since Zt

0

kA (s (!0 ))kB (hM_ is (!0 ))kA (sT (!0 )) ds:

s (! )hM_ is (! )sT (! ) ds Zt

kA (s(!))kB (hM_ is(!))kA (sT (!)) ds 0

© 2001 by Chapman & Hall/CRC

144

Maxingales

+j

Zt

j

2

0

ks(!)hM_ is(!)sT (!)k 1( ks(! )k > A) ds +j

j

2 A2

Zt

0

khM_ is(!)k 1( khM_ is(!)k > B ) ds

and by hypotheses lim sup

A!1

Zt

0

ks(!)hM_ is(!)sT (! )k 1( ks (!)k > A) ds = 0; lim sup

B !1

Zt

0

khM_ is(!)k 1( khM_ is(!)k > B ) ds = 0;

we conclude that lim sup

Zt

0

s(! )hM_ is (! )sT (! ) ds Zt

s(!0 )hM_ is(!0)sT (!0) ds: 0

Fatou's lemma provides the reverse inequality. Thus, R ( 0t shM_ is sT ds; t 2 R+ ) is strictly Luzin. Next, for s; t 2 [0; T ]

Zt

u 0 ZT

hM_ iu uT du

Zs

0

uhM_ iu uT du

kuhM_ iu uT k 1( ku k > A) du 0

+ A2

ZT

0

khM_ iuk 1( khM_ iuk > B ) du + A2B jt sj; R

which implies uniform continuity of ( 0t s hM_ is sT ds; t [0; T ] uniformly over ! 2 K (a).

© 2001 by Chapman & Hall/CRC

2

R+ )

on

145

Idempotent stochastic integrals

Remark 2.5.17. The hypotheses imply that both M

and hM i are

also strictly Luzin-continuous. Theorem 2.5.18. Let be da Hausdor topological space and be a deviability on . Let an R -valued, continuous, strictly Luzin Aadapted idempotent process M be an A-local maxingale on ( ; ) with an absolutely continuous quadratic characteristic hM i such that (hM_ is ; s 2 R+ ) is a strictly Luzin idempotent process and sup sup khM_ is (!)k < 1; t 2 R+ ; a 2 (0; 1]: st !2K (a) Let (s (!); s 2 R+ ; ! 2 ) be an Rmd -valued A-progressively meaR surable strictly Luzin idempotent process such that 0t ks k 2 ds < 1; t 2 R+ ; and Zt

0

ks k 2 1( ks k > A) ds ! 0

as A ! 1; t 2 R+ :

Then the idempotent process M is an A-local maxingale with the quadratic characteristic Zt

h M it = shM_ issT ds: 0

Both M and h M i are strictly Luzin-continuous idempotent processes. Proof. We check the local maxingale property for M . Taking in the hypotheses of Theorem 2.5.11 nA(x) = (A + 1 x)+ ^ 1, we have by hypotheses that ! ! (s (!)nA ks (!)k ; s 2 [0; t]) is a continuous mapping from K (a) to L2 ([0; t]; R md ). Since Zt

0

k snA( ks k ) s Æ nA( ks Æ k ) hM_ is

s nA( ks k ) s Æ nA ( ks Zt

T

Æk )

k ds

sup khM_ isk ksnA( ks k ) s Æ nA( ks Æ k )k 2 ds st

© 2001 by Chapman & Hall/CRC

0

(2.5.19)

146

Maxingales

and by M. Riesz's criterion for relative compactness in L2 the righthand side of (2.5.19) converges to 0 as Æ ! 0 uniformly over ! 2 K (a), we conclude by Theorem 2.5.11 that M is an A-local maxingale with the quadratic characteristic in the statement of the theorem. Both M and h M i are strictly Luzin-continuous by Lemma 2.5.16. The following version of Theorem 2.5.12 is proved along the lines of the proof of Theorem 2.5.18.

Theorem 2.5.19. Let dW

be an Rm -valued A-Wiener idempotent process and X be an R -valued A-adapted Luzin-continuous idempotent process with idempotent distribution X . Let (s (x); s 2 R + ; x 2 C (R + ; R d )) be an R km -valued C(R+ ; R d )-progressively measurableR strictly Luzin idempotent process on (C (R + ; Rd ); X ) such that 0t ks (x)k 2 ds < 1; t 2 R+ ; x 2 C (R + ; Rd ); and Zt

0

ks(x)k 2 1( ks (x)k > A) ds ! 0 X

as A ! 1; t 2 R+ :

Then the idempotent process (X ) W is an A-local maxingale with the quadratic characteristic Zt

h(X ) W it = s(X )s (X )T ds; 0

which is a Luzin-continuous idempotent processes.

Remark 2.5.20. For the convergence condition in the hypotheses to hold it is suÆcient that for every compact K C and t 2 R+ Zt

0

sup ks (x)k 2 ds < 1; x2K

in particular, it is suÆcient that (s (x); s satis es the linear-growth condition

2

R+ ;

x 2 C (R + ; Rd ))

kt (x)k 2 lt(1+supjxsj2 ); x 2 C (R + ; Rd ); t 2 R+ ; st

R where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ .

© 2001 by Chapman & Hall/CRC

Idempotent stochastic integrals

147

Remark 2.5.21.

Note that the maxingale property in Theorem 2.5.19 Rdoes not generally hold for discontinuous s (x). For example, 0t sign (Ws )W_ s ds; t 2 R+ , where W is an R-valued idempotent Wiener process, is not a local maxingale, since -a.e. Rt _ 0 sign (Ws )Ws ds = jWt j:

The next consequence of Theorem 2.5.14 considers strictly Luzin processes on C (R + ; Rd ).

Theorem 2.5.22. Let be a deviability on C (R + ; Rd ) and d C (R+ ; R ) denote the -completion of dthe - ow C(R+ ; Rd ). Let M = (Mt (x); t 2 R+ ; x 2 C (R + ; R )) be a strictly Luzin idempotent process, which is a C(R+ ; RdR)-local maxingale on (C (R + ; Rd ); ) with quadratic characteristic ( 0t s(x)s (x)T ds; t 2 R + ), where (s (x); s 2 R + ; x 2 C (R + ; R d )) is an R dd -valued C(R+ ; Rd )-progressively measurable R strictly Luzin idempotent process on (C (R + ; Rd ); ) such that 0t ks (x)k 2 ds < 1; t 2 R+ ; x 2 d C (R + ; R Zt

0

and

);

ks (x)k 2 1( ks (x)k > A) ds ! 0 as A ! 1, t 2 R+ ;

inf inf inf s (x)s (x)T > 0; a 2 (0; 1]: jj=1

x2K (a) st 2Rd :

Then there exists a strictly Luzin-continuous d-dimensional C(R+ ; Rd )-Wiener idempotent process W on (C (R + ; Rd ); ) such that M = W . Proof. The hypotheses and M. Riesz's criterion for compactness in L2 imply that Zt

ks s+Æ k 2 ds ! 0

0

as Æ ! 0. Thus, the hypotheses of Theorem 2.5.14 hold so that, according to the proof of the theorem, W = 1 M is a Wiener idempotent process. It is strictly Luzin-continuous by Lemma 2.5.16.

© 2001 by Chapman & Hall/CRC

148

Maxingales

Theorem 2.5.23. Let the - ow A be right-continuous. Let an Rvalued continuous A-adapted idempotent process M be an A-local

maxingale with a continuous quadratic characteristic hM i such that limt!1 hM it = 1 in idempotent probability. Then there exists a Wiener idempotent process W such that Ms = WhM is -a.e. Proof. Let s = inf ft 2 R+ : hM it > sg: The function s is an A-stopping time by Lemma 2.2.18 and is a proper idempotent variable. Since hM is = s, Lemma 2.3.13 implies that (exp(Mt^s 2 hM it^s =2); t 2 R+ ) is a uniformly maximable Aexponential maxingale so that by Theorem 2.3.8 the idempotent process (exp(Ms 2 s=2); s 2 R+ ) is an exponential maxingale rela~ = (A~s; s 2 R+ ), where A~s = As and M1 = 0 tive to the - ow A for de niteness. Since the idempotent process Ws = Ms also is ~ -adapted by Lemma 2.2.19, a.e. continuous by Lemma 2.5.4 and A ~ -Wiener idempotent process by Theorem 2.4.2. The it is an A proof is complete if we show that MhM is = Ms -a.e. Since fhM is < tg = fhM it > hM isg, hM is s and A is right-continuous, it follows that hM is is an A-stopping time. Therefore, Lemma 2.5.3 implies that for a > 0, b > 0 and c > 0

( sup jMt Ms j a) ec(b a) _(hM ihM is hM is > 2b=c): sthM is Since hM ihM is = hM is -a.e., we conclude that (supsthM is jMt Ms j > 0) = 0.

Remark 2.5.24. Note that the - ow A~ is also right-continuous by Lemma 2.1.11 and the fact that s is right-continuous.

We now consider analogues of Girsanov's theorem.

Theorem 2.5.25. Let W

be an Rm -valued A-Wiener idempotent process on ( ; ). Let (bs (!); s 2 R+ ; ! 2 ) be an Rm -valued Aprogressively measurable idempotent process such that the idempotent R R processes exp 0t ( bs + ) W_ s ds 0t j bs + j2 ds=2 ; t 2 R+ are well-de ned A-exponential maxingales under for all 2 Rm . Let

Mt = exp

Zt

0

1 bs W_ s ds 2

© 2001 by Chapman & Hall/CRC

Zt

0

jbsj2 ds :

149

Idempotent stochastic integrals

If there exists an idempotent probability 0 on such that its restrictions 0t to the -algebras At are expressed as d0t = Mt d; t 2 R+ , then the idempotent process X = (Xt ; t 2 R+ ), de ned by

Xt =

Zt

bs ds + Wt ;

0

is an A-Wiener idempotent process under 0 .

Proof. We rst note that since Mt > 0, the sets of 0 -idempotent probability 0 have -idempotent probability 0; hence, the ow A is 0 -complete. For s t and 2 Rm in view of Lemma 1.6.35

S0 exp Xt

1 2 jj t jA 2 s S Mt exp Xt 12 jj2 t jAs = : (2.5.20) S (Mt jAs )

By the property of R de nition of X and R M , and the maxingale exp 0t ( bs + ) W_ s ds 0t j bs + j2 ds=2 ; t 2 R+

S Mt exp Xt

= S exp Zs

= exp

Zt

1 2 jj t jAs 2 1 2

( br + ) W_ r dr

0

( br + ) W_ r dr

0

1 2

Zs

Zt

j br + j2 dr jAs

0

j br + j2 dr

0

1 2 jj s : 2 Also S (Mt jAs ) = Ms by the maxingale property of Mt . Thus, by (2.5.20) = Ms exp Xs

1 2 1 j j t jAs = exp Xs jj2 s : 2 2 The 0 -maximability property is obvious.

S0 exp Xt

© 2001 by Chapman & Hall/CRC

150

Maxingales

We now give a version for the canonical setting. We denote C (R + ; R d ) by C .

Theorem 2.5.26. Let space C C (R + ; Rm ) be endowed with a deviability such that the idempotent process W de ned by Wt (x; w) = wt is an Rm -valued Wiener idempotent process. Let Y be de ned by Yt (x; w) = xt and Y denote the idempotent distribution of Y . Let (bs (x); s 2 R+ ; x 2 C ) be an Rm -valued C -progressively measurable Y

bounded strictly Luzin idempotent process on (C ; ). Let M and X be de ned as in Theorem 2.5.25. Then there exists an idempotent probability 0 on C such that its restrictions 0t to the -algebras Ct Ct(R+ ; Rm ) are expressed as d0t = Mt d; t 2 R+ . The idempotent process X is a Wiener idempotent process under 0 . Proof. Let A denote the natural - ow on C C (R + ; Rm ) completed with respect theidempotent proR t to . Theorem 2.5.19 R timplies that 2 _ cess exp 0 ( bs + ) Ws ds 0 j bs + j ds=2 ; t 2 R+ is an A-local exponential maxingale under for arbitrary 2 Rm ; it is actually maxingale, which is derived from the fact R t an exponential 2 that 0 j bs + j ds is a bounded idempotent variable by a standard argument (cf. the proof of Lemma 2.5.10). We now check existence of 0 . We rst show that the idempotent probabilities 0t de ned by d0t = Mt d are deviabilities on C C (R + ; Rm ). Since M is an exponential maxingale on (C C (R + ; Rm ); ), by Lemma 1.7.20 it is enough to check that M (x; w) is strictly Luzin on (C C (R + ; Rm ); ), which follows by Lemma 2.5.16 and the fact that bs (x) is strictly Luzin and bounded. Let 00t denote the deviabilities on the spaces C ([0; t]; R d Rm ) of continuous Rd Rm -valued functions on [0; t] that are the images of 0t under the mappings x ! (xs ; s 2 [0; t]). The maxingale property 00 d m of M implies that t ; C ([0; t]; R R ) is a projective system so that by Lemma 1.8.3 and the fact that the projective limit of the C ([0; t]; R d R m ); t 2 R + ; is homeomorphic to C C (R + ; R m ) we conclude that there exists a deviability 0 on C C (R + ; Rm ) that extends the 0t . The fact that X is idempotent Wiener follows by Theorem 2.5.25.

© 2001 by Chapman & Hall/CRC

151

Idempotent Ito equations

2.6 Idempotent Ito dierential equations This section studies idempotent analogues of Ito dierential equations. We x space C (R + ; Rd ), which we denote throughout the section by C ; the associated - ow C(R+ ; R d ) = (Ct (R+ ; Rd ); t 2 R+ ) is denoted as C = (Ct ; t 2 R+ ). Given space C (R + ; Rm ) we denote m m by C W t (R + ; R ) the completion of Ct (R + ; R ) with respect to the Wiener idempotent probability W on C (R + ; Rm ). Let bt (x); t 2 R+ ; x 2 C ; and t (x); t 2 R+ ; x 2 C ; be respective d R -valued and Rdm -valued functions, which are continuous in x for every t, C-progressively measurable in (t; x), and Zt

0

jbs(x)j ds < 1;

Zt

0

ks(x)k 2 ds < 1; t 2 R+ :

We introduce the equation Zt

Zt

0

0

Xt = X0 + bs(X ) ds + s (X )W_ s ds;

(2.6.1)

where W = (Ws ; s 2 R+ ) is an Rm -valued Wiener idempotent process, X = (Xs ; s 2 R+ ) is an Rd -valued continuous idempotent process, and the second integral on the right is an Ito idempotent integral. The ow A in the following de nitions is assumed to be complete with respect to the associated idempotent probability.

De nition 2.6.1. We sayd that equation (2.6.1) with an initial idempotent distribution on R has a solution if there exist an idempotent probability space ( ; ) with a - ow A, an Rd -valued continuous A-adapted idempotent process X = (Xs; s 2 R+ ) and an Rm -valued A-Wiener idempotent process W = (Ws; s 2 R+ ) on ( ; ) such that X0 has the idempotent distribution , and (2.6.1) holds for all t 2 R+ -a.e. in ! 2 . The pair (X; W ) is then called a solution to the equation. We say that existence holds for (2.6.1) if a solution exists for every .

De nition 2.6.2. We say that equation (2.6.1) with an initial deviability distribution on Rd has a Luzin solution if there exist an idempotent probability space ( ; ) with a - ow A, an Rd -valued © 2001 by Chapman & Hall/CRC

152

Maxingales

continuous A-adapted idempotent process X = (Xs ; s 2 R+ ) and an Rm -valued A-Wiener idempotent process W = (Ws ; s 2 R+ ) on ( ; ) such that (X; W ) is a Luzin-continuous idempotent process on C C (R + ; R m ), X0 has the idempotent distribution , and (2.6.1) holds for all t 2 R+ -a.e. in ! 2 . The pair (X; W ) is then called a Luzin solution to the equation with initial deviability distribution . We say that Luzin existence holds for (2.6.1) if a Luzin solution exists for every deviability .

Remark 2.6.3. We sometimes loosely refer to X alone as a solution (respectively, Luzin solution).

Remark 2.6.4.

We recall that by de nition (X; W ) is a Luzincontinuous idempotent process if the idempotent distribution of (X; W ) is a deviability on C C (R + ; Rm ).

Remark 2.6.5. Let (X; W ) be a solution (respectively, Luzin soX;W

lution) on some idempotent probability space and denote the idempotent distribution of (X; W ). Then by the transitivity property of conditional idempotent expectations the canonical idempotent process (x; w) on C C (R + ; Rm ) with the natural - ow completed with respect to X;W is a solution (respectively, Luzin solution) under X;W . Thus, a solution (respectively, Luzin solution) can always be implemented on the canonical space. Therefore, we occasionally refer to the idempotent distribution of (X; W ) as a solution (respectively, Luzin solution) as well. We denote by (respectively, x ) the idempotent distribution of (X; W ) associated with an initial idempotent distribution (respectively, with an initial condition X0 = x 2 Rd ).

De nition 2.6.6.

We say that uniqueness (respectively, Luzin uniqueness) holds for (2.6.1) if for every two solutions (respectively, Luzin solutions) (X; W ) and (X 0 ; W 0 ) such that the idempotent distributions of X0 and X00 coincide the idempotent distributions of (X; W ) and (X 0 ; W 0 ) also coincide.

De nition 2.6.7. We say that strong existence holds for equation d m

(2.6.1) if there exists a function F : R C (R + ; R ) ! C such m that the function w ! F (x; w) is C W t (R + ; R )=Ct -measurable for every x 2 Rd and t 2 R+ and, given an A0 -measurable idempotent variable f 2 Rd and an Rm -valued A-Wiener idempotent process W = (Ws ; s 2 R+ ) both de ned on an idempotent probability space

© 2001 by Chapman & Hall/CRC

Idempotent Ito equations

153

( ; ) with a - ow A, the idempotent process X = F (f; W ) satis es (2.6.1) for all t 2 R+ -a.e., and X0 = f -a.e. The idempotent process X is then called a strong solution to the equation with an initial condition f . If the function F (x; w) is, in addition, continuous on Rd KW (a), a 2 (0; 1], then we say that Luzin strong existence holds, and X is called a Luzin strong solution. Remark 2.6.8. We recall that KW (a) = fw 2 C (R + ; Rm ) : W (w) ag. Remark 2.6.9. We note that a strong solution X is AW-adapted since W is A-adapted, A is complete and w ! F (x; w) is C t (R+ ; Rm )=Ct measurable. De nition 2.6.10. We say that (2.6.1) has a unique strong solution (respectively, Luzin strong solution) if strong existence (respectively, Luzin strong existence) holds and if, given a solution (respectively, Luzin solution) (X; W ) with an initial condition X0 on an idempotent probability space ( ; ) with a - ow A, we have that X = F (X0 ; W ) -a.e., where F is the function in the de nition of a strong solution (respectively, Luzin strong solution). De nition 2.6.11. We say that pathwise uniqueness holds for (2.6.1) if for every two solutions (X; W ) and (X 0 ; W 0 ), which are de ned on the same idempotent probability space ( ; ) with the same - ow A, we have X = X 0 -a.e. provided X0 = X00 and W = W 0 -a.e. We refer to (2.6.1) as an idempotent Ito dierential equation and to X as an idempotent diusion. The functions bt (x) and t (x)t (x)T are occasionally referred to as in nitesimal drift and diusion coef cients, respectively. We also use the following short-hand notation for (2.6.1) X_ t = bt (X )+ t (X )W_ t : It is clear that uniqueness implies Luzin uniqueness, Luzin strong existence implies strong existence, and strong existence implies existence. We study some other relationships between the introduced concepts. We rst discuss the role of initial conditions in the above de nitions. The next lemma follows by Lemma 1.5.5 and Theorem 1.8.9. It implies, in particular, that if Luzin existence holds, then existence holds.

© 2001 by Chapman & Hall/CRC

154

Maxingales

Lemma 2.6.12.

1. If for every initial condition x 2 Rd there exists a solution x , then existence holds. Speci cally, given an initial idempotent distribution , the idempotent distribution de ned by (x; w) = supx2Rd x (x; w)(x) is a solution to (2.6.1).

2. If for every initial condition x 2 Rd there exists a Luzin solution x , which is a deviability transition kernel from Rd into C C (R + ; R m ), then Luzin existence holds. More speci cally, given an initial deviability , the idempotent probability de ned as in part 1 is a Luzin solution with initial deviability distribution .

Lemma 2.6.13. If the idempotent distributions of0 every two solu0 tions (respectively, Luzin solutions) (X; W ) and (X ; W ) with initial condition x coincide for every x 2 Rd , then uniqueness (respectively, Luzin uniqueness) holds.

Proof. Let X be a solution with an initial idempotent distribution de ned on an idempotent probability space ( ; ) with a - ow A. For x 2 Rd such that (x) > 0, let x(A) = (AjX0 = x); A

. Then it follows from the de nition of conditional idempotent probability that X0 = x x -a.e.; since W is independent of A0 by the de nition of an A-Wiener idempotent process, it is independent of X0 , so that W is an A-Wiener idempotent process on ( ; x ); (2.6.1) holds x -a.e. since it holds -a.e. Thus, X is a solution of (2.6.1) with initial condition x on the space ( ; x ) for the AWiener idempotent process W ; hence, the idempotent distribution of (X; W ) under x is speci ed uniquely. By Theorem 1.6.12 we have for A C C (R + ; Rm ) that

((X; W ) 2 A) = sup x ((X; W ) 2 A)(X0 = x) x2Rd = sup x ((X; W ) 2 A)(x); (2.6.2) x2Rd which implies that the idempotent distribution of (X; W ) under is speci ed uniquely. We now turn to Luzin uniqueness. We assume the canonical setting so that = C C (R + ; Rm ), is a deviability on , and (X; W ) is the canonical idempotent process. Let be an initial

© 2001 by Chapman & Hall/CRC

Idempotent Ito equations

155

deviability distribution on Rd . Then x de ned as in the preceding proof is a deviability on by Theorem 1.6.12 and the fact that the set fX0 = xg belongs to the collection of closed subsets of . Since (X; W ) is a Luzin solution such that X0 = x x -a.e., deviability x is speci ed uniquely. Since (2.6.2) holds for this case as well, is speci ed uniquely. The next lemma follows by similar arguments and the de nitions. Lemma 2.6.14. 1. If there exists a function F (x; w) in the de nition of a strong solution (respectively, Luzin strong solution) such that the idempotent process X = F (x; W ) is a strong solution (respectively, Luzin strong solution) for an initial condition x, then strong existence (respectively, Luzin strong existence) holds. 2. If pathwise uniqueness holds for solutions with initial condition x for every x 2 Rd , then pathwise uniqueness holds. Let us associate with (2.6.1) a collection of ordinary dierential equations depending on absolutely R 1 continuous functions w = (wt ; t 2 R + ) 2 C (R + ; R m ) such that 0 jw_ t j2 dt < 1:

x_ t = bt (x)+t (x)w_ t a.e. in t; x0 = x 2 Rd ; (2.6.3) where the x = (xt ; t 2 R+ ) 2 C are absolutely continuous functions. De nition 2.6.15. We say that the extension condition holds for d equation (2.6.3) if for every x 2 R and w 2 C (R + ; Rm ) such that R1 2 0 jw_ t j dt < 1 the following holds: if a function (xt ), de ned on an interval [0; T ], satis es (2.6.3) for t 2 [0; T ], then it can be extended to a solution of (2.6.3) on R+ . Remark 2.6.16. The extension condition implies existence of solutions for every equation (2.6.3). Lemma 2.6.17. 1. If the extension condition holds for (2.6.3), then existence holds for the idempotent Ito dierential equation (2.6.1). If existence holds for (2.6.1), then every ordinary dierential equation (2.6.3) has a solution. 2. If every ordinary dierential equation (2.6.3) has at most one solution, then pathwise uniqueness and uniqueness hold for (2.6.1).

© 2001 by Chapman & Hall/CRC

156

Maxingales

3. Luzin strong existence implies Luzin existence. 4. If strong existence (respectively, Luzin strong existence) and pathwise uniqueness hold for (2.6.1), then there is a unique strong solution (respectively, Luzin strong solution). Proof. Let the extension condition hold for (2.6.3). We de ne an idempotent distribution x on C C (R + ; Rm ) by x (x; w) = W (w) if x and w satisfy (2.6.3) and x0 = x, and x (x; w) = 0 otherwise. Let (X; W ) denote the canonical idempotent process on C C (R + ; R m ) and A = fAt ; t 2 R + g be the natural - ow completed with respect to x. Then (X; W ) satis es (2.6.1) for the initial condition x x-a.e. We show that W is an A-Wiener idempotent process. It is suÆcient to check that W has A-independent increments, i.e., x (t w0 jAt ) = x (t w0 ) x -a.e., where t 2 R+ and w0 2 C (R + ; Rm ) is such that W (w0 ) > 0. Let (x00 ; w00 ) be such that x pt 1 Æ pt (x00; w00 ) > 0. Then

x(t w0jAt )(x00 ; w00 ) 0 00 00 x (x; w) : t w = t w ; pt (x; w) = pt (x ; w ) = x (x; w) : pt (x; w) = pt(x00 ; w00 ) : Since x pt 1 Æ pt (x00 ; w00 ) > 0, the pair (x00 ; w00 ) satis es (2.6.3) on [0; t]. The extension condition implies that for every w such that pt w = pt w00 and t w = t w0 there exists a solution to (2.6.3) on R+ that coincides with x00 on [0; t]. Therefore, by the de nition of x we have that

x (x; w) : t w = t w0 ; pt(x; w) = pt (x00 ; w00 )

= W (w : t w = t w0 ; pt w = pt w00 ):

By a similar reasoning

x (x; w) : pt (x; w) = pt(x00 ; w00 ) = W (w : pt w = pt w00 ): Thus, by independence of increments of W

x(t w0jAt )(x00 ; w00 ) =

© 2001 by Chapman & Hall/CRC

: t w = t w0 ; pt w = pt w00 ) W (w : pt w = pt w00 ) = W (t w0 )

W (w

157

Idempotent Ito equations

as required. Thus, (X; W ) is a solution of (2.6.1) on (C C (R + ; R m ); x ) with - ow A for every x 2 Rd . By Lemma 2.6.12 existence holds. Conversely, let existence hold. Let x be an idempotent distribution of (X; W ) on C C (R + ; Rm ) for an initial condition x. Since supx x (x; w) = W (w), it follows that if W (w) > 0, then there exists x such that x (x; w) > 0 so that (x; w) satisfy (2.6.3) and x0 = x. This ends the proof of part 1. We now prove part 2. The fact that pathwise uniqueness holds if every dierential equation (2.6.3) has at most one solution is obvious. Let us assume that x is a solution on C C (R + ; R m ) with an initial condition x. By de nition if x (x; w) > 0, then x solves (2.6.3) with x0 = x. By uniqueness for (2.6.3) x is a unique solution for given w W and x. Since we must have that supx x (x; w) = (w) it follows that x (x; w) = W (w) so that x coincides with the solution x de ned in the proof of part 1. Luzin strong existence implies Luzin existence by Theorem 1.7.11. Part 4 is obvious.

Remark 2.6.18. Note that under the hypotheses of part 20 we have

pathwise uniqueness even for two solutions (X; W ) and (X ; W ) that are not necessarily associated with the same - ow. The latter is also true if there is a unique strong solution.

The following lemma takes advantage of the proof of Lemma 2.6.17 to indicate a candidate for a solution of (2.6.1). The proof also relies on Lemma 2.6.12.

Lemma 2.6.19. Let the extension condition hold for (2.6.3). Then the idempotent distribution

(x; w) = supd x(x; w)(x); x2R

where

8 W > :

0;

if x_ t = bt (x) + t (x)w_ t a.e. and x0 = x; otherwise;

is a solution for an initial idempotent distribution . If x (x; w) is a deviability transition kernel from Rd into C C (R + ; R m ), then is a Luzin solution for an initial deviability .

© 2001 by Chapman & Hall/CRC

158

Maxingales

Remark 2.6.20. Easy supw x(x; w) is given by

Xx (x) = exp

1 2

calculations show that

Xx (x)

=

Z1

0

(x_ t bt (x)) t (x)t (x)T (x_ t bt (x)) dt

if x0 = x, x is absolutely continuous and x_ t bt (x) is in the range of t (x) a.e., and X x (x) = 0 otherwise. We also note that the range of t (x) coincides with the range of t (x)t (x)T . Theorem 2.6.21. 1. If pathwise uniqueness holds, then uniqueness holds. 2. Let pathwise uniqueness hold. If existence (respectively, Luzin existence) holds, then strong existence (respectively, Luzin strong existence) holds so that there exists a unique strong solution (respectively, Luzin strong solution). Proof. Let (X; W ) and (X 0 ; W 0 ) be two solutions of (2.6.1) with an initial condition x 2 Rd on respective idempotent probability spaces ( ; ) and ( 0 ; 0 ) with respective - ows A and A0 . Let us introduce the conditional idempotent distributions w (A) = (X 2 AjW = w) and 0 w (A) = 0 (X 0 2 AjW 0 = w). We show that for x 2 C and t 2 R+

w pt 1 (pt x) = (pt X = pt xjpt W = pt w) for W {almost all w; (2.6.4)

i.e., the left-hand side depends only on the piece of w up to t. (Of course, a similar relation holds for 0 .) Recalling the notation t ws = wt+s wt ; s 2 R+ ; w 2 C (R + ; Rm ), we have, for w such that W (w) > 0, in view of the fact that pt X and pt W are At -measurable, and t W is independent of At , that (pt X = pt x; W = w w pt t x) = (W = w) (pt X = pt x; pt W = pt w; t W = t w) = (pt W = pt w; t W = t w) (pt X = pt x; pt W = pt w)(t W = t w) = (pt W = pt w)(t W = t w) = (pt X = pt xjpt W = pt w): 1 (p

© 2001 by Chapman & Hall/CRC

159

Idempotent Ito equations

The claim has been proved. ~ on ~ = C C C (R + ; Rm ) We de ne an idempotent probability by ~ x; x0 ; w) = w (x)0 w (x0 )W (w): ( (2.6.5)

Clearly, ~ fxg C fwg = (X = x; W = w) and ~ C fx0 g fwg = 0(X 0 = x0 ; W 0 = w). Let C~t be the completion ~ and C ~ = (C~t ; t 2 R+ ). of Ct Ct Ct (R+ ; Rm ) with respect to ~We check that the canonical idempotent process (wt ; t 2 R+ ) is a C ~ ~ Wiener idempotent process on ( ; ). It obviously has idempotent distribution W . By Theorem 2.4.9 it is suÆcient to check that t w is ~ -independent of C~t . We have, for x^ ; x^ 0 2 C and w; ^ w~ 2 C (R + ; Rm ), in view of (2.6.5), ~ (x; x0 ; w) : pt (x; x0 ; w) = pt (^x; x^ 0 ; w^ ); t w = w~ = sup 1 pt (x; x0 ; w) = pt (^x; x^ 0 ; w^); t w = w~ (x;x0 ;w)2 ~ w (x)0 w (x0 )W (w) = sup 1 pt w = pt w; ^ t w = w~ w (pt 1 (pt x^ )) w 0 w (pt 1 (pt x^ 0 ))W (w) = sup 1(pt w = pt w^ )w (p 1 (pt x^ ))0 w (p 1 (pt x^ 0 ))W (w)

w

sup 1(t w = w~ )W (w)

w

t

t

~ pt (x; x0 ; w) = pt (^x; x^ 0 ; w^ )( ~ t w = w~ ); =

where the equality before the last one follows by (2.6.4) and the fact that t w is independent of pt w under W . Thus, (x; w) and (x0 ; w) are two solutions to (2.6.1) on the same idempotent probability space and adapted to the same - ow. By ~ pathwise uniqueness we conclude that x = x0 -a.e. so ~ x; x0 ; w) = 0: sup 1(x 6= x0 )( (2.6.6) 0 ~ (x;x ;w)2

Therefore, ~ fxg C fwg) = (( ~ x; x; w)) ((X; W ) = (x; w)) = ( ~ C fxg fwg) = 0 ((X 0 ; W 0 ) = (x; w)) = (

© 2001 by Chapman & Hall/CRC

160

Maxingales

so that uniqueness holds. Next, by (2.6.6) and (2.6.5) supx;x0 1(x 6= x0 )w (x)0 w (x0 ) = 0 for W -almost all w. Fixing w such that W (w) > 0 and picking x~ 0 such that 0w (~x0 ) > 0 we de ne F (x; w) = x~ 0. Since x = x~ 0 for w -almost all x, we have that x = F (x; w) whenever (X = x; W = w) > 0. If W (w) = 0, we de ne F (x; w) arbitrarily. By construction, X = F (x; W ) -a.e. m We prove that F (x; w) is C W t (R + ; R )=Ct -measurable in w for every x 2 Rd and t 2 R+ . Since X = F (x; W ) -a.e., we have that (X = xjW = w) = 1(F (x; w) = x) if (W = w) > 0. Therefore, by (2.6.4) for W -almost all w

1(F (x; w) 2 pt 1(pt x)) = (pt X = pt xjW = w) = (pt X = pt xjpt W = pt w): Since the right-most side, for xed x, is a function of pt w, we conm clude that fw : F (x; w) 2 pt 1 (pt x)g 2 C W t (R + ; R ). Thus, strong

existence holds, and by part 4 of Lemma 2.6.17 there exists a unique strong solution. For the part concerned with Luzin solutions, we need to check, in addition, that if Luzin existence holds, then F (x; w) is continuous in (x; w) on Rd KW (a); a 2 (0; 1]. Let (xn ; wn ) 2 Rd KW (a); a 2 ^ W ^ ) be a Luzin (0; 1]; converge to (~x; w~ ) and xn = F (xn ; wn ). Let (X; ^ ^ solution on an idempotent probability space ( ; ) with an initial ^ X^ 0 = xn) = 1; n 2 N ; and ( ^ X^0 = x~) = condition X^ 0 such that ( ^ ^ ^ ^ ^ ^ ^ ^ 1. Since X = F (X0 ; W ) -a.e., ((X; X0 ; W ) = (xn ; xn ; wn )) = ^ X^0 = xn )W (wn ) a. Since (X; ^ W ^ ) is a Luzin solution, the ( ^ ^ ^ ^ set f(x; x; w) : ((X; X0 ; W ) = (x; x; w)) ag is compact, which implies that the sequence f(xn ; xn ; wn ); n 2 N g is relatively compact and every x; x~; w~) is such that ^ (X;^ X^0 ; W^ ) = accumulation point (~ (~x; x~; w~ ) a > 0. Hence, x~ = F (~x; w~ ). According to Lemma 2.6.17 existence and uniqueness issues for (2.6.1) and (2.6.3) are closely related. Thus, the methods of the theory of ordinary dierential equations apply to the study of existence and pathwise uniqueness. We recall that bt (x) and t (x) are assumed to be continuous in x.

© 2001 by Chapman & Hall/CRC

161

Idempotent Ito equations

Theorem 2.6.22.

1. Let bt (x) and t (x) satisfy the lineargrowth conditions

jbt(x)j lt (1 + supjxs j); kt (x)k 2 lt (1 + supjxsj2 ); st st t 2 R+ ; x 2 C ; R where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ : Then existence holds for (2.6.1).

2. Let bt (x) and t (x) be locally Lipshitz continuous, i.e., for every a 2 R+ , t 2 R+ and x; y 2 C ; such that supst jxs j a and supst jys j a we have

jbt (x) bt (y)j kta supjxs ysj; st 2 a kt (x) t (y)k kt supjxs ysj2 ; st

R

where kta is Lebesgue measurable and 0t ksa ds < 1; t 2 R+ : Then pathwise uniqueness holds for (2.6.1). Proof. For the existence part we have to check the extension condition for (2.6.3). We use the method of successive approximations. Standard are omitted. Let x 2 Rd , w 2 C (R + ; Rm ) be such R 1 details 2 that 0 jw_ s j ds < 1, and a function x^ t ; t 2 [0; T ]; satisfy (2.6.3) on [0; T ]. Let unt be successive approximations de ned by

unt +1

Zt

= x+ bs

(un ) ds+

0

Zt

s (un )w_ s ds; t 2 R+ ;

(2.6.7)

0

where u0t = x^ t for t 2 [0; T ] and u0t = x^ T for t T . Then by the Cauchy-Schwarz inequality and the linear-growth conditions

junt +1 j2

© 2001 by Chapman & Hall/CRC

Z t 2 3x +3

jj

0

2

jbs(un )j ds

162

Maxingales

+3

Zt

ks

(un )k 2 ds

0

jj

3x2 +3

Zt

2

Zt

jw_ sj2 ds

0

ls ds +

0

Zt

0 t Z

+3 2

Zt

jw_ sj

2 ds

ls ds +

0

Zt

0

ls ds

0

jw_ sj

2 ds

Zt

0

ls supjunr j2 ds: rs

This string of inequalities shows in particular that the right-hand side of (2.6.7) is well de ned. Denoting ftn = supst juns j2 , we conclude that, given T > 0, there exist constants A1 and A2 such that for all tT

ftn+1

Zt

A1 + A2 lsfsn ds; 0

which implies by a version of Gronwall's inequality that

ftn A1 exp A2

Zt

ls ds ; t T:

0

Therefore, supn supst juns j < 1, which easily implies by (2.6.7) and the linear-growth conditions that the sequence f(unt ; t 2 R+ ); n 2 N g is locally equicontinuous, so that by Arzela-Ascoli's theorem it is relatively compact in C . In a standard way, by using continuity of bt (x) and t (x) in x and the linear-growth conditions, it follows that every accumulation point of f(unt ; t 2 R+ )g solves (2.6.3). It coincides with x^ on [0; T ] since unt = x^ t for t 2 [0; T ]. The uniqueness part is also proved by a standard argument. Let u and v be two solutions of (2.6.3) such that u0 = v0 = x. Let a (x) = inf ft 2 R+ : jxt j ag; x 2 C : Then denoting uat = ut^ a (u)^ a (v) and vta = vt^ a (v)^ a (u) , by the Cauchy-Schwarz inequality and Lipshitz

© 2001 by Chapman & Hall/CRC

163

Idempotent Ito equations

continuity conditions

juat

vta j2

2

Zt

0

ksa ds

Zt

0

ksa supjuar

vra j2 ds

rs

+2

Zt

0

Zt

jw_ sj2 ds ksa supjuar vra j2 ds; 0

rs

so that ua = va by Gronwall's inequality. Hence, a (u) = a (v) and ut = vt for t ua . Letting a ! 1 completes the proof.

Remark 2.6.23. Under the hypotheses of part 2 the stronger version of pathwise uniqueness of Remark 2.6.18 holds.

We now strengthen part 1 of Theorem 2.6.22.

Theorem 2.6.24. Under the hypotheses of part 1 of Theorem 2.6.22 Luzin existence holds for equation (2.6.1).

Proof. Since according to the proof of Theorem 2.6.22 the extension condition holds for (2.6.3) under the hypotheses, by Lemma 2.6.19 it suÆces to check that x(x; w) de ned in the statement of the lemma is a deviability transition kernel. By Lemma 1.8.12 this can be done by proving that x (x; w) is upper semi-continuous in (x; x; w) and the sets f(x; w) : supjxjA x(x; w) ag are relatively compact for every A 2 R+ and a 2 (0; 1]. We consider the upper semi-continuity rst. Let (xn ; xn ; wn ) ! (~x; x~ ; w~ ) as n ! 1. We can obviously assume that the (xn ; xn ; wn ) satisfy equation (2.6.3) for otherwise xn (xn ; wn ) = 0. In addition, since by de nition xn (xn ; wn ) = W (wn ), we may assume that W (wn ) a > 0. We check that x~ is a solution to (2.6.3) associated with x~ and w~ . The linear-growth conditions, continuity of bs (x) and s(x) in x, Lebesgue's dominated convergence theorem, and Lemma 2.5.16 yield

lim

Zt

n!1

lim

n!1

Zt

0

0

bs(x

n ) ds

=

s (xn )w_ sn ds =

© 2001 by Chapman & Hall/CRC

Zt

0 Zt 0

bs (~x) ds; s (~x)w~_ s ds;

164

Maxingales

proving the claim. Thus, x~ (~x; w~ ) = W (w~ ). Since by upper semi-continuity of W (w) we have that lim supn!1 xn (xn ; wn ) = lim supn!1 W (wn ) W (w~ ), the required follows. Let us check that the set f(x; w) : supjxjA x(x; w) ag is relatively compact. Let a sequence f(xn ; xn ; wn ); n 2 N g be such that jxn j A and xn (xn ; wn ) a(1 1=n). Since xn (xn ; wn ) = W (wn ) and W (w) is upper compact, we may assume that the xn and wn converge to some x~ and w~ , respectively. Since the xn are solutions of (2.6.3),

jx

n j2 t

Z t

3(xn )2 + 3

jbs(x

0 Z t

jxnt xnsj2 2

s

n )j ds 2 + 3 2

Zt

jw_ sn j2 ds

Zt

0 Zt

0 Zt

s

s

ks(xn )k 2 ds;

jbr (xn)j dr + 2 jw_ rn j2 dr kr (xn )k 2 dr:

Since also inf n W (wn ) > 0, we conclude in analogy with the proof of Theorem 2.6.22 that the sequence fxn g is relatively compact in C.

Remark 2.6.25. One can also show that in the hypotheses of part 1 of Theorem 2.6.22 the idempotent process X de ned for (x; w) 2 C Rt C (R + ; R m ) by Xt (x; w) = xt x b ( x ) ds is a local maxingale on 0 s m m C C (R + ; R ); C C(R+ ; R ); x with quadratic characteristic Rt T (x) ds; t 2 R . ( x ) s + s 0

Combining Lemma 2.6.19, Theorem 2.6.21, Theorem 2.6.22, and Theorem 2.6.24 we obtain the following existence and uniqueness result.

Theorem 2.6.26. Let bt (x) and t (x) be locally Lipshitz-continuous and satisfy the linear-growth conditions. Then the equation X_ t = bt (X )+ t (X )W_ t ; X0 = x;

has a unique Luzin solution, which is also a strong Luzin solution. The deviability distribution of X is given by 1 Z1 X x (x) = exp (x_ t bt (x)) t (x)t (x)T (x_ t bt (x)) dt 2 0

© 2001 by Chapman & Hall/CRC

165

Idempotent Ito equations

if x0 = x, x is absolutely continuous and x_ t of t (x) a.e., and X x (x) = 0 otherwise.

bt (x) is in the range

Remark 2.6.27. Existence of a strong solution under the hypothe-

ses can also be proved directly by using a version of the method of successive approximations.

As a consequence of \the Girsanov theorem" (Theorem 2.5.26) we have the following existence and uniqueness result.

Theorem 2.6.28. Let (s (x); s 2 R+ ; x 2 C ) be an Rm valued C(R+ ; Rm )-progressively measurable bounded function such that s (x) is continuous in x for s 2 R+ . Then Luzin existence and uniqueness hold for the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x;

(2.6.8)

if and only if Luzin existence and uniqueness hold for the equation X_ t = bt (X ) t (X )t (X ) +t (X )W_ t ; X0 = x;

(2.6.9)

where W is an Rm -valued A-Wiener idempotent process.

Corollary 2.6.29. Let (bs (x); s 2 R+ ; x 2 C ) be bounded. Luzin existence and uniqueness hold for the equation

Then

X_ t = bt (X )+ W_ t ; X0 = x; where W is an Rd -valued A-Wiener idempotent process. Let X denote the idempotent distribution of X . Then X (x) = 0 unless x0 = 0 and x is absolutely continuous. For these x

x) = exp

X (

1 2

Z1

0

jx_t bt(x)j2 dt :

We now outline another approach, which is analogous to the martingale problem approach and which we will explore in detail later in the text. We state the result for Luzin solutions, which are our main concern below. Given a deviability on C , we denote by C the -completion of the - ow C.

© 2001 by Chapman & Hall/CRC

166

Maxingales

Theorem 2.6.30. Let the matrix s(x) have size d d and for every compact K C and t 2 R+

lim sup

a!1 x2K

Zt

0

ks(x)k 2 1( ks (x)k > a) ds = 0;

inf inf s (x)s (x)T > 0; xinf 2K st 2Rd :

lim sup

a!1 x2K

Zt

0

jj=1

jbs(x)j 1(jbs (x)j > a) ds = 0:

Then the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x; has a Luzin solution (X; W ) if and only if there exists a deviability on C such that R x0 = x -a.e., and the idempotent process Mt (x) = xt x 0t bs (x) ds is a C-localR maxingale on (C ; ) with the quadratic characteristic hM it (x) = 0t s (x)s (x)T ds and is strictly Luzin. The solution (X; W ) is unique if and only if the deviability is unique. The idempotent distribution of X then coincides with . Proof. Let (X; W ) be a Luzin solution on an idempotent prob~ with a - ow A. Then by Theorem 2.5.19 ability space ( ; ) Rt _ ( 0 s(X )Ws ds;R t 2 R+ ) is an A-local maxingale with the quadratic characteristic ( 0t s(X )s (X )T ds; t 2 R+ ). The idempotent distribution of X is the required deviability . R t Conversely, let Mt (x) = xt x 0 bs (x) ds be a strictly Luzin idempotent process, which is a RC-local maxingale with the quadratic characteristic hM it (x) = 0t s (x)s (x)T ds on (C ; ). Then by Theorem 2.5.22 there exists a strictly Luzin-continuous CRt _ Wiener idempotent process W on (C ; ) such that Mt = 0 s Ws ds, which implies that the canonical process on (C ; ) and W make up a Luzin solution to the equation.

Remark 2.6.31. Under the hypotheses, both M and hM i are also strictly Luzin-continuous on (C ; ).

We conclude the section with an existence result for an equation with respect to Poisson idempotent processes, which will be used

© 2001 by Chapman & Hall/CRC

167

Idempotent Ito equations

in a queueing application later on. We con ne ourselves to Luzin solutions. Let (us (x); s 2 R+ ; x 2 C (R + ; R)) and (vs (x); s 2 R+ ; x 2 C (R + ; R )) be C(R + ; R )-progressively measurable R + -valued functions, which are continuous in x and such that Zt

0

us (x) ds < 1;

Zt

0

us (x) ds < 1; t 2 R+ ; x 2 C (R + ; R):

Let us consider the equation

Xt = x+N1

Z t

us (X ) ds

N2

0

Zt

vs (X ) ds ;

(2.6.10)

0

where N1 and N2 are independent Poisson idempotent processes and x 2 R.

De nition 2.6.32. We say that equation (2.6.10) has a Luzin solution if there exist an idempotent probability space ( ; ) equipped with a - ow A and R-valued continuous idempotent processes X = (Xs ; s 2 R+ ), N1 = (N1 (s); s 2 R+ ) and ( ; ) such that the following holds 1. X , are

R

N1 ( 0t us(X ) ds); t 2 R+

A-adapted,

and

N2 = (N2(s); s 2 R+ ) on R

N2( 0t vs(X ) ds); t 2 R+

R

2. the R idempotent processes N1(r + 0Rt us(X ) ds) N1(R0t us(X ) ds); r 2 R+ and N2(r + 0t vs(X ) ds) N1( 0t vs(X ) ds); r 2 R+ , when conditioned on At, where t 2 R+ , are independent Poisson idempotent processes, 3. (X; N1 ; N2 ) is a Luzin-continuous idempotent process, 4. (2.6.10) holds for t 2 R+ -a.e. in ! 2 . The triplet (X; N1 ; N2 ) is then called a Luzin solution to the equation with initial condition x.

The next existence result is an analogue of Theorem 2.6.24 and is proved along the same lines.

© 2001 by Chapman & Hall/CRC

168

Maxingales

Theorem 2.6.33. Let us(x) and vs(x), in addition to the above conditions, satisfy the linear-growth condition us (x) + vs (x) ls (1 + supts jxt j), where ls is locally integrable. Then equation (2.6.10) has a Luzin solution (X; N1 ; N2 ) on an idempotent probability space ( ; ) with a - ow A such that the idempotent distribution of X has density

x) = exp

X x(

Z1

0

sup x_ t (e 1)ut (x) (e 2R

if x is absolutely continuous and

1)vt (x) dt

x0 = x, and Xx (x) = 0 otherwise.

Proof. We rst prove an analogue of the extension property for the equation

xt = x+n1

Zt

0

us(x) ds

n2

Zt

0

vs (x) ds ; t 2 R+ :

(2.6.11)

More speci cally, we prove that given n1 2 C (R + ; R) and n2 2 C (R + ; R ) such that N (n1 )N (n2 ) > 0, where N is the Poisson idempotent probability, every solution of (2.6.11) on an interval [0; T ] can be extended to a solution on R+ . Since by properties of the idempotent Poisson process for n 2 C (R + ; R), A 2 R+ and t 2 R+ N

n(t) S N en(t) e(e 1)t > A A(1+t) = A(1+t) 1+t e e

and the latter ratio is less than N (n1 ) for all A large enough, we conclude that n1 (t) A(1 + t) for all t 2 R+ if A is large. The same fact holds for n2 . Since also n1 and n2 are continuous, the claim follows by a successive approximation argument as in the proof of Theorem 2.6.22. We de ne idempotent probability X;N1 ;N2 on RC (R + ; R 3 ) by X; RN1 ;N2 (x; n1; n2 ) = N (n1 )N (n2 ) if xt = x + n1 0t us (x) ds n2 0t vs (x) ds ; t 2 R+ , and X;N1 ;N2 (x; n1 ; n2 ) = 0 otherwise. Let (X; N1 ; N2 ) be the canonical process on C (R + ; R3 ). It clearly satis es (2.6.10) X;N1 ;N2 -a.e. We check that (X; N1 ; N2 ) is Luzincontinuous, i.e., that X;N1 ;N2 is a deviability by showing that the sets K (a) = f(x; n1 ; n2 ) 2 C (R + ; R3 ) : X;N1 ;N2 (x; n1 ; n2 ) ag are

© 2001 by Chapman & Hall/CRC

169

Idempotent Ito equations

compact for all a 2 (0; 1], which is carried out as in the proof of Theorem 2.6.24. In some more detail, let (xk ; n1;k ; n2;k ) 2 K (a); k 2 N . Since N is a Luzin-continuous idempotent process by Lemma 2.4.17, we have by Theorem 2.2.13 that for T 2 R+ and > 0 lim N Æ!0

sup jn(t) n(s)j > = 0: s;t2[0;T ]: jt sjÆ

Since N (n1;k ) a > 0 and N (n2;k ) a > 0, it follows that the functions n1;k and n2;k , k 2 N , are locally uniformly equicontinuous. Besides in analogy with the above argument there exists B > 0 such that n1;k (t) + n2;k (t) B (1 + t) for t 2 R+ and k 2 N . Since also the (xk ; n1;k ; n2;k ) satisfy (2.6.11), a standard argument shows that the xk are uniformly bounded on bounded intervals and locally uniformly equicontinuous as well. Arzela-Ascoli's theorem implies that the sequence f(xk ; n1;k ; n2;k ); k 2 N g is relatively compact in C (R + ; R 3 ). Let (~ x; n~ 1; n~ 2 ) beX;Nan;Naccumulation point. It clearly satN 1 2 is es (2.6.11). Therefore, (~x; n~ 1 ; n~ 2 ) = (~n1 )N (~n2 ) a; where the latter inequality follows since (~n1 ; n~ 2 ) is an accumulation point of (n1;k ; n2;k ), N (n1 )N (n2 ) a, and the function N (n) is upper semi-continuous. Let At be the -algebra generated by the atoms pt x, pR0t us (x) ds n1 and pR t vs (x) ds n2 , where (x; n1 ; n2 ) 2 C (R + ; R3 ). De ning 0 A =R tfAt ; t 2 R+ g, we see that the X, R idempotent processes N1( 0 us(X ) ds); t 2 R+ , and N2( 0t vs(X ) ds); t 2 R+ are Aadapted. We nowR show that the idempotent processes N1(r + Rt Rt t 0 uRs (X ) ds) N1 ( 0 us (X ) ds); r 2 R + and N2 (r + 0 vs (X ) ds) N1( 0t vs(X ) ds); r 2 R+ , when conditioned on At, are independent Poisson idempotent processes. Equivalently, we have to prove that for n01 ; n02 2 C (R + ; R) the following holds X;N1 ;N2 -a.e. X;N1 ;N2 R0t us (X ) ds N1 = n01 ; R0t vs (X ) ds N2 = n02 jAt = N (n01 )N (n02 ): (2.6.12)

Let n001 ; n002 ; x00 2 C (R + ; R) be such that X;N1 ;N2 (x00 ; n001 ; n002 ) > 0. Then by the de nition of X;N1 ;N2 , the extension property and the

© 2001 by Chapman & Hall/CRC

170

Maxingales

properties of the Poisson idempotent probability X;N1 ;N2 R0t us (X ) ds N1 = n01 ; R0t vs (X ) ds N2 = n02 ; pt X = pt x00 ; pR0t us (X ) ds N1 = pR0t us (x00 ) ds n001 ; pR0t vs (X ) ds N2 = pR0t vs (x00 ) ds n002 = sup 1 R0t us(x) ds n1 = n01; R0t vs(x) dsn2 = n02; (x;n1 ;n2 )2C (R+ ;R3 ) pt x = pt x00 ; pR0t us (x) ds n1 = pR0t us (x00 ) ds n001 ; pR0t vs (x) ds n2 = pR0t vs (x00 ) ds n002 X;N1 ;N2 (x; n1 ; n2 ) = sup 1 R0t us(x00 ) ds n1 = n01; R0t vs(x00 ) ds n2 = n02; (n1 ;n2 )2C (R+ ;R2 ) pR0t us (x00 ) ds n1 = pR0t us (x00 ) ds n001 ; pR0t vs (x00 ) ds n2 = pR0t vs (x00 ) ds n002 N (n1 )N (n2 ) = sup 1 R0t us (x00 ) ds n1 = n01 ; n1 2C (R+ ;R) pR0t us (x00 ) ds n1 = pR0t us (x00 ) ds n001 sup 1 R0t vs (x00 ) ds n2 = n02 ; n2 2C (R+ ;R) N 00 R R p 0t vs (x00 ) ds n2 = p 0t vs (x00 ) ds n2 (n1 )N (n2 ) = N n1 : pR t 00 n1 = pR t 00 n00 N (n0 ) 0 us (x ) ds

0 us (x ) ds 1

1

N n2 : pR0t vs (x00 ) ds n2 = pR0t vs (x00 ) ds n002 N (n02 ) = X;N1 ;N2 pt X = pt x00 ; pR t us (X ) ds N1 = pR t us (x00 ) ds n001 ; 0 0 N 0 N 0 00 R R p t vs (X ) ds N2 = p t vs (x00 ) ds n2 (n1 ) (n2 ):

0

0

Equality (2.6.12) follows. X;N1 ;N2 (x; n ; n ) coinFinally, the fact that X 1 2 x (x) = supn1 ;n2 cides with the expression for X (x) in the statement of the theorem follows by routine calculations.

Remark 2.6.34. Rt

Rt We refer to N u ( X ) ds ; t 2 R+ and 1 s 0 N2 0 vs(X ) ds ; t 2 R+ as Poisson idempotent processes of rates us (X ) and vs (X ), respectively.

2.7 Semimaxingales In this section we consider idempotent analogues of semimartingales and associated integrals. Let ( ; ) be an idempotent probability space with a - ow A. Let (Gt (; !); t 2 R+ ; ! 2 ); 2 Rd ;

© 2001 by Chapman & Hall/CRC

Semimaxingales

171

be R-valued A-adapted continuous idempotent processes such that G0 (; !) = Gt (0; !) = 0. We refer to G() = (Gt (; !); t 2 R+ ; ! 2

); 2 Rd ; as a cumulant.

De nition 2.7.1. We say that an Rd -valued A-adapted continuous idempotent process X on ( ; ) is an A-semimaxingale with cumulant G() if the idempotent process Y () = (Yt (; !); t 2 R+ ; ! 2 ) de ned by

Yt () = exp( (Xt X0 ) Gt ())

(2.7.1)

is an A-local exponential maxingale for every 2 Rd . If, in addition, G() is an increasing function of t for all and !, X is called an A-local maxingale.

Remark 2.7.2. We occasionally say that X is a semimaxingale on ( ; A; ) rather than that it is an A-semimaxingale. Examples of semimaxingales are the idempotent Wiener process and the idempotent Poisson process. Also local maxingales with quadratic characteristics are semimaxingales. Another example is provided by idempotent processes with independent increments. Recall that AX denotes the - ow generated by an idempotent process X.

Theorem 2.7.3. Let X be a continuous A-adapted idempotent process with independent increments such that the function G~ t () = ln S exp( (Xt X0 )) is nite for all 2 Rd . Then X is an AX -

semimaxingale with cumulant G~ (). If, in addition, X0 is Luzin and G~ t () is dierentiable in , then X is Luzin-continuous. Proof. We only prove the Luzin-continuity. Since G~ t () is dierentiable in , by Lemma 1.11.7 the increments Xt Xs are Luzin idempotent variables, so by independence of increments, Corollary 1.8.10 and Theorem 1.7.11 X is Luzin. It is Luzin-continuous by Theorem 2.2.13 and the fact that G~ t () is continuous in t.

Our primary goal is to show that integrals with respect to semimaxingales also give rise to local exponential maxingales. The methods are analogous to those we used in Section 2.5. In view of applications to large deviation theory we are interested in studying semimaxingales on spaces of trajectories that are also strictly Luzin

© 2001 by Chapman & Hall/CRC

172

Maxingales

idempotent processes with respect to a deviability. Therefore, both in this section and the next one we assume that is the space C = C (R + ; R d ) and is a deviability on C . We equip C with the

ow C = (Cs ; s 2 R+ ) that is the completion of the natural - ow C = (Cs; s 2 R+ ) with respect to , where the Cs are the -algebras generated by the mappings x ! xt ; t 2 [0; s]; for x 2 C . We start with some simple properties.

De nition 2.7.4. We say that an R +-valued function on C is a strictly Luzin stopping time on (C ; C ; ) if it is a C{stopping time and is nite and continuous when restricted to K (a) for a 2 (0; 1]. The following result is standard.

Lemma 2.7.5. Let (Ht (x); t 2 R+ ; x 2 C ) be an R+ -valued increasing continuous C {adapted strictly Luzin idempotent process. Let, for c 2 R+ ,

(x) = inf ft 2 R+ : Ht (x)+ t cg:

Then (x); x 2 C ; is a strictly Luzin stopping time on (C ; C ; ).

Proof. The fact that is a C {stopping time follows by Lemma 2.2.18. Since Ht (x) is increasing, continuous in t and continuous in x on K (a); a 2 (0; 1]; it is also continuous on K (a) as a map from C to C (R + ; R) by Polya's theorem, see Liptser and Shiryaev [79, Problem 5.3.2]. Continuity of (x) on K (a) follows now by Whitt [135, Theorem 7.2] since Ht (x) + t is strictly increasing.

De nition 2.7.6. We say that a semimaxingale (respectively, local maxingale) X with cumulant G() is a strictly Luzin semimaxingale (respectively, local maxingale) on (C ; C ; ) if X and G() for every d

2 R are strictly Luzin idempotent processes. We say that X is a strictly Luzin-continuous semimaxingale (respectively, local maxingale) if X and G() for every 2 Rd are strictly Luzin-continuous idempotent processes.

Lemma 2.7.7. If X is either a strictly Luzin local maxingale or a strictly Luzin-continuous semimaxingale on (C ; C ; ), then the local exponential maxingales Y () admit localising sequences of strictly Luzin stopping times.

© 2001 by Chapman & Hall/CRC

173

Semimaxingales

Proof. Let X be a strictly Luzin local maxingale with cumulant G(). Let n = inf ft 2 R+ : Gt (2) + t ng, where n 2 N . Since G(2) is a strictly Luzin idempotent process and is increasing, n is a strictly Luzin stopping time by Lemma 2.7.5. The idempotent process (Yt^n (); t 2 R+ ) is uniformly maximable because

S Yt^n ()2

S Yt^n (2) exp(Gt^n (2)) en;

where the rst inequality holds since Gt () is non-negative and the second follows by the de nition of n and the fact that (Yt^n (); t 2 R + ) is a supermaxingale starting at 1. If X is a strictly Luzin-continuous semimaxingale, then the above argument applies with n = inf ft 2 R+ : supst jGs (2)j_jGs ()j + t ng. In connection with the lemma we introduce the following.

De nition 2.7.8. A local exponential maxingale M on (C ; C ; ) is

called a strictly Luzin-continuous local exponential maxingale if it is a strictly Luzin-continuous idempotent process and admits a localising sequence of strictly Luzin stopping times.

Remark 2.7.9. Note that if M is a strictly Luzin-continuous local

exponential maxingale and is a strictly Luzin stopping time, then Mt^ is a strictly Luzin idempotent variable. In the rest of the section X is the canonical idempotent process on C , i.e., Xt (x) = xt , and the following is assumed to hold:

X is a semimaxingale on (C ; C ; ) with cumulant G().

Lemma 2.7.10. The idempotent process X is a semimaxingale with 1 cumulant G() under the deviability (jx0 = x) for Æ 0 -almost d all x 2 R .

We omit a simple proof and only note that (jx0 = x) is a deviability by Lemma 1.6.12. We introduce an idempotent measure that is to play an important part in the sequel. Let 0 be the set of all Rd {valued, piecewise constant functions ((t); t 2 R+ ) of the form

(t) =

k X i=1

i 1(t 2 (ti 1 ; ti ]);

© 2001 by Chapman & Hall/CRC

174

Maxingales

where 0 t0 < t1 < : : : < tk ; i 2 Rd ; i = 0; : : : ; k; k de ne for x 2 C and x 2 Rd Z1 I(x) = sup (t)dxt dGt ((t); x); ((t))20 0

2 N : We (2.7.2)

where the integral is understood as a nite sum so that Z1

0

((t) dxt

=

dGt ((t); x))

k X i=1

i (xti

xti 1 )

(Gti (i ; x) Gti 1 (i ; x)) ; (2.7.3)

and let

((x) = exp( I(x)); (2.7.4) Ix(x) = I(x); if x0 = x; (2.7.5) 1;

otherwise; x(x) = exp( Ix(x)); x 2 C ; x( ) = sup x(x); C : (2.7.6) x2 We have our rst property of x .

Lemma 2.7.11. For Æ 0 1-almost all x 2 Rd , (xjx0 = x) x(x); x 2 C . In particular, (x) (x) and is an idempotent probability.

Proof. By Lemma 2.7.10 it is enough to check that (x) x (x) assuming that x0 = x -a.e. We follow the argument of the proof of Lemma 2.5.8. Let, for 0 s1 t1 : : : sk tk and i 2 Rd ; i = 1; : : : ; k, k hX

Z~ = exp

i=1

i Xti Xsi

k X i=1

i

Gti (i ) Gsi (i ) :

The de nition of x and argument of the proof of Lemma 2.5.8 imply that it suÆces to show that S Z~ 1. Let

n = inf t 2 R+ : max jGt (i )j_jGt (2i )j i=1;:::;k

© 2001 by Chapman & Hall/CRC

n

175

Semimaxingales

be a common localising sequence for the Y (i ); i = 1; : : : ; k, and

Yni = exp i Xti ^n Xsi ^n

Gti ^n (i ) Gsi ^n (i ) :

Then S (Yni jCsi ) = 1 so that

S

k Y i=1

Yni = S = S

kY1

i=1 kY2 i=1

Yni S (Ynk jCsk )

Yni S (Ynk

jC

1 ) sk 1

= : : : = S Yn1 = 1:

Q Since n ! 1, it follows that Z~ = limn!1 ki=1 Yni so that \the Fatou lemma" (see Theorem 1.4.19) yields the required. Finally, is an idempotent probability since is an idempotent probability and (x) 1 by de nition.

Below, we are mostly concerned with the case where the cumulant G() = (Gt (; x); t 2 R+ ; x 2 C ); 2 Rd , is absolutely continuous so that it has the form

Gt (; x) =

Zt

0

gs (; x) ds; 2 Rd ; t 2 R+ ; x 2 C ;

(2.7.7)

where gs (; x) is Lebesgue integrable in s. The next lemma gives the form of x for absolutely continuous cumulants. It also shows that our usage of the notation is consistent with the one in Section 2.6 (see Remark 2.6.20). Let

hs (y; x) = sup ( y gs (; x)) 2Rd

(2.7.8)

be the convex conjugate, or the Legendre{Fenchel transform, of gs (; x). It is non-negative provided gs (0; x) = 0.

Lemma 2.7.12. Let an R-valued function gs(; x) be Lebesgue measurable in s and continuous in , gs (0; x) = 0, and for A > 0, t 2 R+ and x 2 C Zt

0

sup jgs (; x)j ds < 1: jj=A

© 2001 by Chapman & Hall/CRC

176

Maxingales

If G() has the form (2.7.7), then 8 1 Z > > < I(x) = > 0 hs (x_ s; x) ds; if x is absolutely continuous, > : +1; otherwise. In particular, X is absolutely continuous under the hypotheses.

Proof. If x is absolutely continuous, then the desired representation follows by Lemma A.2 in Appendix A with f (t; ) = x_ t gt (; x); (2.7.3) and (2.7.5). Let x not be absolutely continuous on an interval [0; T ]. Then we can choose " > 0 such that for every Æ > 0 there exist 0 t1 < : : : < t2l T satisfying l X i=1

(t2i t2i 1 ) < Æ;

l X i=1

jxt2i xt2i 1 j > ":

(2.7.9)

For N > 0, we take

xt2i xt2i 1 1 jx xt2i 1 j (t2i 1 ;t2i ](t) i=1 t2i (of course we may assume that jxt2i xt2i 1 j > 0). N (t) = N

l X

Then by (2.7.3), (2.7.5), and (2.7.9)

I(x) =N

Z1

0 l X

i=1

[N (t) dxt

jxt2i xt2i 1 j > N"

ZT

0

dGt (N (t); x)] l Zt2i X i=1 t2i 1

gt (N (t); x) dt

l [

sup jgt (; x)j 1 t 2 (t2i 1 ; t2i ] dt: jj=N i=1

By (2.7.9) the latter integrand goes to 0 in measure as Æ ! 0 so that by Lebesgue's dominated convergence theorem the integral converges to 0 as Æ ! 0. Thus, I(x) > N" for arbitrary N .

© 2001 by Chapman & Hall/CRC

177

Semimaxingales

Finally, since by Lemma 2.7.11 (X = xjX0 = x) x(x) for Æ0 1 -almost all x, X is absolutely continuous under (jX0 = x) for these x. Since (X = x) = supx2Rd (X = xjX0 = x)(X0 = x), it follows that X is absolutely continuous under . We assume in the rest of the section that G() is given by (2.7.7). Let us further assume that Gt (; x) = Bt0 (x)+ G^ t (; x); (2.7.10)

where B 0 = (Bt0 (x); t 2 R+ ; x 2 C ) is an Rd -valued C -adapted idempotent process such that B00 (x) = 0 and G^ () = (G^ t (; x); t 2 R + ; x 2 C ), 2 R d , are R+ -valued C -adapted idempotent processes such that G^ 0 (; x) = G^ t (0; x) = 0. Since G() is absolutely continuous in t, we assume that both B 0 and G^ () are absolutely continuous so that

Bt0 (x) = G^ t (; x) =

Zt

0 Zt 0

bs (x) ds;

(2.7.11)

g^s (; x) ds;

(2.7.12)

where (bs (x)) is C-progressively measurable, 0t jbs (x)jds < 1, (^gs (; x)) is R+ -valued, B([0; t]) B(R d ) Ct =B(RR+ ){measurable as a map from [0; t] Rd C to R+ , g^s (0; x) = 0, and 0t g^s (; x) ds < 1 for t 2 R+ ; 2 Rd and x 2 C . (The product of a -algebra and a -algebra has been introduced in De nition 1.5.9, the product of two -algebras has a standard meaning.) Thus, gs (; x) from (2.7.7) has the form

gs (; x) = bs (x)+^gs (; x):

R

We now introduce more conditions on bs (x) and g^s (; x).

(2.7.13)

(I ) The idempotent process (bs (x)) is strictly Luzin on (C ; ) and Zt

0

sup

x2K (a)

jbs(x)j ds < 1

for all a 2 (0; 1] and t 2 R+ .

© 2001 by Chapman & Hall/CRC

178

Maxingales

(II ) The function (^gs (; x)) is continuous in (; x) when restricted to Rd K (a) for a 2 (0; 1], convex in 2 Rd , and sup sup sup g^s (; x) < 1; lim sup sup g^s (; x) = 0 !0 st x2K (a) jjA st x2K(a) for all a 2 (0; 1], t 2 R+ and A 2 R+ . Let Mt = Xt X0 Bt0 : Then M = (Mt (x); t 2 R+ ; x 2 C ) is a C { local maxingale with cumulant G^ () and the following \canonical decomposition" holds X = X0 + B 0 + M: (2.7.14) 0 Under (I ) and (II ) the idempotent processes B and M are strictly Luzin-continuous. As above, we denote by M_ a C -progressively measurable idempotent process that coincides with the RadonNikodym derivative of M with respect to Lebesgue measure -a.e. We note that for absolutely continuous x x_ s gs (; x) = M_ s (x) g^s (; x); (2.7.15) so by (2.7.8), Lemma 2.7.12, (2.7.10), (2.7.14), and Lemma 2.7.11 Z1 I(x) = h^ t (M_ t (x); x) dt < 1 -a.e.; (2.7.16) 0

where h^ t (y; x) = sup (y g^t (; x)); y 2 Rd ; t 2 R+ ; x 2 C : 2Rd

(2.7.17) De nition 2.7.13. Let ^ denote the set of all C{ progressively measurable strictly Luzin idempotent processes = ((t; x); t 2 R+ ; x 2 C ) on (C ; ) such that for 2 R, t 2 R+ and x 2 C R d {valued

Zt

0

g^s ((s; x); x) ds < 1

(2.7.18)

and, moreover, Zt

0

g^s ((s; x); x) 1(j(s; x)j > A) ds ! 0 as A ! 1:

© 2001 by Chapman & Hall/CRC

(2.7.19)

179

Semimaxingales

Remark 2.7.14.

If condition (II ) holds, then bounded C { progressively measurable strictly Luzin idempotent processes belong to ^ .

Lemma 2.7.15. Let conditions (RI ) and (II ) hold. Let = ((t; x); t 2 R+ ; x 2 C ) 2 ^ . Then 0t (s; x) M_ s ds, where t 2 R+ , is well de ned and nite -a.e. Proof. Since M is absolutely continuous, the integrand in the statement is well de ned -a.e. We show that -a.e. Zt

0

j(s; x) M_ s j ds < 1:

Since h^ s (y; x) is the convex conjugate of g^s (; x) and g^s (; x) is nonnegative, by Young's inequality (see, e.g., Krasnosel'skii and Rutickii [75])

j(s; x) M_ sj = (s; x) sign (s; x) M_ s M_ s g^s (s; x) sign ((s; x) M_ s); x + h^ s (M_ s; x) g^s((s; x); x) + g^s( (s; x); x) + h^ s(M_ s ; x):

The integral from 0 to t of the right-most side is nite -a.e. by the de nition of ^ and (2.7.16). Given an Rd -valued Lebesgue measurable in s function = ((s; x)), we introduce an idempotent process Z () = (Zt (; x); t 2 R + ; x 2 C ) by Z t

Zt (; x) = exp

0

(s; x)x_ s gs ((s; x); x) ds

(2.7.20)

if the integral on the right-hand side is well de ned and nite, and let Zt (; x) = 0 otherwise. If 2 ^ and conditions (I ) and (II ) hold, then by Lemma 2.7.15, the de nition of ^ and (2.7.15) equality (2.7.20) holds for t 2 R+ -a.e. The main result of this section is the following theorem. Theorem 2.7.16. Let conditions (I ) and (II ) hold. If 2 ^ , then the idempotent process Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ).

© 2001 by Chapman & Hall/CRC

180

Maxingales

The proof proceeds through a string of lemmas. Let us rst note that by Lemma 2.7.11, Lemma 2.7.12 and condition (II )

S Z () 1 (2.7.21) for an arbitrary R+ -valued function on C . In the lemmas below we assume that conditions (I ) and (II ) hold. Lemma 2.7.17. Let 2 ^ . Then the idempotent process Z () is strictly Luzin-continuous. If (x); x 2 C ; is a strictly Luzin stopping time, then Zt^ () is a Ct^ {measurable strictly Luzin variable. Proof. The argument is similar to the one we used in the proof of Lemma 2.5.16. We begin the proof of Z () being strictly Luzincontinuous by proving that Zt () is a strictly Luzin variable. Since by (2.7.21) and \the Chebyshev inequality" Zt () is a proper idempotent variable, by (2.7.20) it is suÆcient to check that the maps x ! Rt Rt _ 0 (s; x) Ms (x) ds and x ! 0 g^s ((s; x); x) ds are continuous when restricted to sets K (a); a 2 (0; 1]: Let xk ! x^ as k ! 1, where xk ; x^ 2 K(a). We rst check the convergence lim

k!1

Denoting

Zt

0

(s; x

k )M_

x

s(

k ) ds =

Zt

0

(s; x^ )M_ s (^x) ds:

A (t; x) = (t; x)iA (j(t; x)j); where iA (x) = (A +1 x)+ ^ 1; x 2 R+ ; we have that, for A > 0, Z t (s; k ) 0 Zt

x M_ s(xk ) ds

Zt

0

(s; x^ )

(2.7.22)

(2.7.23) (2.7.24) _ Ms (^ ) ds

x

j(s; xk ) M_ s(xk )j 1(j(s; xk )j > A) ds 0

+

Zt

0

© 2001 by Chapman & Hall/CRC

j(s; x^ ) M_ s(^x)j 1(j(s; x^ )j > A) ds

181

Semimaxingales

+

Zt

j(A (s; xk ) A(s; x^ )) M_ s(xk )j ds

0 Zt + A (s; ^ ) 0

x M_ s(xk ) ds

Zt

0

A (s; x^ )

M_ s (^ ) ds :

x

(2.7.25)

Note that all the terms in (2.7.22) and (2.7.25) are well de ned by Lemma 2.7.15 and the fact that (A (t; x); t 2 R+ ; x 2 C ) 2 ^ . We prove that each term on the right of (2.7.25) tends to 0 as k ! 1 and A ! 1. Since by Young's inequality for > 0 1 1 y g^t (; x)+ h^ t (y; x); we have for x 2 C and > 0 by (2.7.16) Zt

0

j(s; x) 1(j(s; x)j > A) M_ s(x)j ds

1 +

1

Zt

0 Zt 0 Zt

g^s (s; x) sign ((s; x) M_ s (x)); x 1(j(s; x)j > A) ds

h^ s (M_ s (x); x) ds

1 [^gs((s; x); x) + g^s( (s; x); x)] 1(j(s; x)j > A) ds 0

so that by (2.7.19) and the inequality I(x) holds since (x) (x)) lim sup sup A!1 x2K (a)

Zt

0

1 I(x); (2.7.26) ln a on K (a) (which +

j(s; x) M_ s(x)j 1(j(s; x)j > A) ds lna ; > 0:

Since is arbitrary, we have thereby proved that the rst term on the right of (2.7.25) tends to 0 as k ! 1 and A ! 1, and the second one tends to 0 as A ! 1.

© 2001 by Chapman & Hall/CRC

182

Maxingales

For the third term, we write analogously to (2.7.26) for > 0 Zt

0

j(A (s; xk ) A(s; x^ )) M_ s(xk )j ds

1

Zt

0

g^s ((A (s; xk ) A (s; x^ )); xk ) ds

1 +

Zt

0

g^s ( (A (s; xk ) A (s; x^ )); xk ) ds +

1 k I(x ): (2.7.27)

Since is strictly Luzin, (s; xk ) ! (s; x^ ); hence, by (2.7.23) and (2.7.24) A (s; xk ) A (s; x^ ) ! 0. Therefore, by condition (II ) g^s ((A (s; xk ) A (s; x^ )); xk ) ! g^s (0; x^ ) = 0 as k ! 1. Thus, by Lebesgue's dominated convergence theorem and condition (II ) (recall that by (2.7.23) and (2.7.24) jA (s; xk ) A (s; x^ )j 2(A +1)), the rst two terms on the right of (2.7.27) tend to 0 as k ! 1. Since is arbitrary, the inequality I(xk ) ln a implies that the third term on the right of (2.7.25) tends to 0 as k ! 1. Thus, we are left to prove that lim

k!1

Zt

0

x

A (s; ^ )M_

x

s(

k ) ds =

Zt

0

A (s; x^ )M_ s (^x) ds:

(2.7.28)

Let ~ be the set of bounded Rd {valued Borel functions ((t); t 2 R+ ) such that lim

k!1

Zt

0

(s) M_ s (xk ) ds =

Zt

0

(s) M_ s (^x) ds:

We prove that ~ consists, in fact, of all Rd {valued bounded Borel functions, which will imply (2.7.28) since A (s; x^ ), being Lebesguemeasurable in s, coincides a.e. with some Borel-measurable function. Since Ms (x) is continuous on K (a) by (I ) and (II ), the convergence xk ! x^ implies that ~ contains all piecewise constant functions ((t)). Now, by a standard monotone class argument, see, e.g.,

© 2001 by Chapman & Hall/CRC

183

Semimaxingales

Meyer [88], it is suÆcient to prove that ~ is closed under bounded pointwise convergence. We prove this by an argument similar to the one we used above: let n (s) ! (s) as n ! 1, where jn (s)j A and j(s)j A. Then, as in (2.7.26), we have with the use of Young's inequality for > 0 Zt

0

j(n (s) (s)) M_ s(x)j ds

1

Zt

0

[^gs ((n (s) (s)); x) + g^s ( (n (s) (s)); x)] ds

1 I(x); so that condition (II ) and Lebesgue's dominated convergence theorem yield +

lim

sup

Zt

n!1 x2K (a)

j(n (s) (s)) M_ s(x)j ds = 0:

0

Hence, lim lim sup

n!1 k!1

Zt

0

lim

n!1

j(n(s) (s)) M_ s(xk )j ds = 0;

Zt

0

j(n (s) (s)) M_ s(^x)j ds = 0;

which proves the claim. Thus, R t (2.7.28) and with it (2.7.22) have been proved. Continuity of x ! 0 (s; x) M_ s (x) ds on K (a) has been proved. In order to prove that Zt

0

g^s ((s; x

x

k ); k ) ds !

Zt

0

g^s ((s; x^ ); x^ ) ds;

(2.7.29)

we note that since (s; x) is continuous in x on K (a) by the de nition of ^ and g^s (; x) is continuous in (; x) by condition (II ),

© 2001 by Chapman & Hall/CRC

184

Maxingales

g^s ((s; xk ); xk ) ! g^s ((s; x^ ); x^ ) as k ! 1: Therefore, condition (II ) and Lebesgue's dominated convergence theorem yield for A > 0 lim

Zt

k!1

0

g^s ((s; xk ); xk )iA (j(s; xk )j) ds =

Zt

0

g^s ((s; x^ ); x^ )iA (j(s; x^ )j) ds

so that lim sup k!1

Zt

g^s ((s; xk ); xk ) 1(j(s; xk )j A) ds

0

Zt

g^s((s; x^ ); x^ ) ds: 0

The latter inequality, (2.7.19) and Fatou's lemma imply the convergence (2.7.29). To complete the proof of Z () being strictly Luzin-continuous, by Theorem 2.2.13 it is suÆcient to show that for T 2 R+ and > 0

lim sup jZt () Zs()j > = 0: Æ!0 s;t2[0;T ]: js tjÆ Since for A > 0

jZt () Zs ()j > (Zs() > A) _ jZt ()=Zs() 1j > =A and (Zs () > A) 1=A by (2.7.21), we deduce that the required would follow by lim Æ!0

Zt sup s;t2[0;T ]:

js tjÆ s

Z t sup g^u ((u; s;t2[0;T ]:

lim Æ!0

js tjÆ s

© 2001 by Chapman & Hall/CRC

(u; x) M_ u (x) du > = 0;

x); x) du >

= 0:

185

Semimaxingales

The rst convergence follows by the inequality Zt

s

j(u; x) M_ u(x)j du

1

ZT

0

[^gu ((u; x); x) + g^u ( (u; x); x)] 1(j(u; x)j > A) du 1 +

Zt

s

1 sup g^u (; x) du + I(x); A > 0; > 0; jjA

derived in analogy with (2.7.26) and condition (II ). The second convergence follows by condition (II ) and the inequality Zt

s

g^u ((u; x); x) du

ZT

0

g^u ((u; x); x) 1(j(u; x)j > A) du +

Zt

sup g^u (; x) du: j jA s

Finally, since and (bs (x)) are C -progressively measurable, g^s (; x) is B([0; t]) B(Rd ) Ct =B(R+ ){measurable as a map from [0; t] _ is C -progressively measurable, and C is comR d C to R + , M plete, Ct^ {measurability of Zt^ () follows by Lemma 2.2.17 and Lemma 2.2.19. The fact that Zt^ () is strictly Luzin measurable follows by the rst part of the lemma. We now address the uniform maximability issue. The next lemma is in the theme of Lemmas 2.7.5 and 2.7.7. Lemma 2.7.18. Let 2 ^ and (Gt ; t 2 R+ ), G0 = 0, be an increasing continuous C {adapted strictly Luzin idempotent process such that for -almost all x Zt

0

g^s (2(s; x); x) ds Gt (x); t 2 R+ :

© 2001 by Chapman & Hall/CRC

186

Maxingales

Let for N

2N

N (x) = inf t 2 R+ : Gt (x)+ t N :

Then N (x) is a strictly Luzin stopping time, the idempotent process fZt^N (); t 2 R+ g is uniformly maximable, and, moreover, S Zt^N ()2 eN : Lemma 2.7.19. Let a function = ((t; x); t 2 R+ ; x 2 C ) 2 ^ be of the form:

(t; x) =

k X i=1

i (x)1(ti

1 ;ti ] (t);

where k 2 N ; 0 t0 < t1 < : : : < tk , and the i (x) are Rd {valued, bounded and Cti 1 {measurable strictly Luzin variables on (C ; C ; ). Then Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ).

Proof. Since by Lemma 2.7.17 Z () is a C {adapted strictly Luzincontinuous idempotent process, we have to check that there exists an increasing to in nity sequence of strictly Luzin stopping times N (x); N 2 N , such that the (Zt^ N (x) ((x); x); t 2 R+ ) are uniformly maximable exponential maxingales. Let A be a bound for i (x), i.e., ji (x)j A; i = 1; : : : ; k; x 2 C . We introduce for t 2 R+ , x 2 C and N 2 N

N (x) =

8 > >

> :

N;

Zt

0

sup g^s (; x) ds + t N ; if (x) > 0; jj2A if (x) = 0:

By condition (II ) and Lemma 2.7.18 the N (x) are strictly Luzin stopping times and the idempotent processes (Zt^N (); t 2 R+ ) are uniformly maximable. Let us check that the (Zt^N (); t 2 R+ ); N 2 N ; are exponential maxingales. Let 0 s < t. We have to prove that

S (Zt^N ()jCs ) = Zs^N ():

(2.7.31)

(As above relations involving conditional idempotent expectations are understood to hold -a.e.) We begin with a proof of

S (Zt^N ()jCti ^t ) = Zti ^t^N (); i = 1; : : : ; k:

© 2001 by Chapman & Hall/CRC

(2.7.32)

187

Semimaxingales

We note that by (2.7.1) and (2.7.20) -a.e.

Zt ((x); x) =

Yti ^t (i (x); x) : Y ( ( x ) ; x ) t ^ t i i 1 i=1

k Y

(2.7.33)

Since the idempotent processes Y () = (Yt (; x); t 2 R+ ; x 2 C ), 2 Rd , are local exponential maxingales, Lemma 2.7.18 and Lemma 2.3.13 imply that the (Yt^N (); t 2 R+ ); jj A; are uniformly maximable exponential maxingales. We prove (2.7.32) by inverse induction in i: by (2.7.33) Ztk ^t^N () = Zt^N () so that (2.7.32) holds for i = k. Suppose that (2.7.32) holds for some i 2 f2; : : : ; kg: We prove it for (i 1). By properties of conditional idempotent expectations

S (Zt^N ()jCti 1 ^t ) = S [S (Zt^N ()jCti ^t )jCti 1 ^t ] = S (Zti ^t^N ()jCti 1 ^t ): By (2.7.33) and properties of conditional idempotent expectations

S (Zti ^t^N ()jCti 1 ^t ) =

Ytj ^t^N (x) (j (x); x) Y ( (x); x) j =1 tj 1 ^t^N (x) j i 1 Y

S(Yti ^t^N (x) (i (x); x)jCti 1 ^t ) : (2.7.34) Yti 1 ^t^N (x) (i (x); x) Let t ti 1 . Since i (x) is Cti 1 {measurable, by properties of conditional idempotent expectations

S (Yti ^t^N (x) (i (x); x)jCti 1 ^t ) = S (Yti ^t^N (x) (i (x); x)jCti 1 ) = S (Yti ^t^N (x) (; x)jCti 1 )j=i (x) = Yti 1 ^t^N (x) (i (x); x); where in the latter equality we used that (Yt^N (x) (; x); t 2 R+ ) is an exponential maxingale. So, if t ti 1 ,

S (Yti ^t^N (x) (i (x); x)jCti 1 ^t ) = Yti 1 ^t^N (x) (i (x); x):

© 2001 by Chapman & Hall/CRC

188

Maxingales

This also is true if t ti 1 . Substituting the equality into (2.7.34) obtains

S (Zti ^t^N ()jCti 1 ^t ) =

Ytj ^t^N (x) (j (x); x) Y ( (x); x) j =1 tj 1 ^t^N (x) j i 1 Y

= Zti 1 ^t^N ():

Equality (2.7.32) is proved. Now, (2.7.31) is obvious if s tk since in this case Zs^N () = Zt^N () = Ztk ^N (). Let s < tk and i0 2 f1; : : : ; kg be such that ti0 1 s < ti0 . By properties of conditional idempotent expectations, (2.7.32) and (2.7.33)

S (Zt^N ()jCs ) = S [S (Zt^N ()jCti0 ^t )jCs ] = S (Zti0 ^t^N ()jCs ) S (Yti0 ^t^N (x) (i0 (x); x)jCs ) = Zti0 1 ^s^N () : (2.7.35) Yti0 1 ^s^N (x) (i0 (x); x)

Now, as above, since i0 (x) is Cti0 1 {measurable,

S (Yti0 ^t^N (x) (i0 (x); x)jCs ) = S (Yti0 ^t^N (x) (; x)jCs )j=i0 (x) = Yti0 ^s^N (x) (i0 (x); x): Substituting this into (2.7.35), we deduce by (2.7.33) that (2.7.31) holds. Lemma 2.7.19 proves the assertion of Theorem 2.7.16 for piecewise constant and bounded functions . To handle general 2 ^ , we will use the following approximation result, which extends Lemma 2.5.10. Lemma 2.7.20. Let k = (k (t; x); t 2 R+ ; x 2 C ); kk 2 N ; be uniformly bounded functions from ^ such that the Z ( ) are strictly Luzin-continuous local exponential maxingales. If = ((t; x); t 2 ^ is bounded and is a limit of the k in the sense that R+ ; x 2 C ) 2 Zt

0

g^s ((k (s; x) (s; x)); x)ds ! 0 as k ! 1; 2 R; t 2 R+ ;

© 2001 by Chapman & Hall/CRC

189

Semimaxingales

then Z () is a strictly Luzin-continuous local exponential maxingale. Proof. The proof uses the ideas of the proof of Lemma 2.5.10. Let for N 2 N and k 2 N

N (x) = inf t 2 R+ :

x) = inf t 2 R+ :

Nk (

Zt

0

Zt

0

g^s (2(s; x); x)ds + t

N ;

(2.7.36)

g^s (2k (s; x); x) ds + t

N + 1 ^ N (x):

(2.7.37)

By Lemma 2.7.18 and condition (II ) the N and Nk are strictly Luzin stopping times, and (Zt^N (); t 2 R+ ) and (Zt^Nk (k ); t 2 R + ); k 2 N ; are uniformly maximable idempotent processes. In particular, by Lemma 2.3.13 the (Zt^Nk (k ); t 2 R+ ); k 2 N ; are uniformly maximable exponential maxingales. Suppose we have proved that for every R+ -valued bounded and continuous function f on C and t 2 R+ lim S Zt^Nk (x) (k (x); x)f (x) = SZt^N (x) ((x); x)f (x): (2.7.38) k Since the (Zt^Nk ( ); t 2 R+ ); k 2 N ; are exponential maxingales, this would imply that (Zt^N (); t 2 R+ ) is an exponential maxingale as well; hence, since (Zt^N (); t 2 R+ ) is uniformly maximable and N (x) ! 1 as N ! 1, this would prove in view of Lemma 2.7.17 that Z () is a strictly Luzin-continuous local exponential maxingale. Therefore, we prove next (2.7.38). By Lemma 2.7.18 and (2.7.37) S Zt^Nk (x) (k (x); x)2 exp(N + 1), so the family fZt^Nk (k ); k 2 N g is uniformly maximable, and by \the Lebesgue dominated convergence theorem" (see Theorem 1.4.19) (2.7.38) would follow by k!1

Zt^Nk (k ) ! Zt^N () as k ! 1:

© 2001 by Chapman & Hall/CRC

(2.7.39)

190

Maxingales

As a rst step, we prove that, for every 2 R and t 2 R+ , Zt

0

jg^s (k (s; x); x) g^s ((s; x); x)jds ! 0 as k ! 1:

(2.7.40) By convexity of g^s (; x) in , for " 2 (0; 1=2], g^s ((s; x); x) (1 2")^gs (k (s; x); x) + "g^s (2k (s; x); x) +"g^s ((s; x) k (s; x)); x ; " hence, since g^s (; x) 0,

g^s ((s; x); x) g^s (k (s; x); x) " sup g^s (; x) jj2Ajj

+ "g^s ((s; x) k (s; x)); x ; " where A is an upper bound for j(s; x)j; jk (s; x)j; k 2 N ; s 2 R+ , x 2 C . Interchanging k (s; x) and (s; x) and integrating, we arrive at the inequality for x such that (x) > 0 Zt

0

jg^s (k (s; x); x) g^s ((s; x); x)jds

"

Zt

0

Zt

sup g^s (; x) ds + " g^s ((s; x) k (s; x)); x ds " jj2Ajj 0

+"

Zt

g^s

0

k ( (s; x) (s; x)); x ds; "

where the right-hand side is nite by condition (II ). By hypotheses, the latter two integrals on the right tend in deviability to 0 as k ! 1, so, for > 0, Zt

lim sup (x : jg^s (k (s; x); x) g^s ((s; x); x)jds > ) k!1 0

(x : "

© 2001 by Chapman & Hall/CRC

Zt

0

sup g^s (; x)ds > =3); jj2Ajj

191

Semimaxingales

where the right-hand side is not greater than arbitrary a > 0 if " > 0 is chosen small enough to satisfy Zt

" sup

x2K (a) 0

sup g^s (; x) ds 3 jj2Ajj

(use condition (II )). Limit (2.7.40) is proved. It implies, since by (2.7.36) and (2.7.37)

fx : N (x) ^ t 6= Nk (x) ^ tg fx :

Zt

0

jg^s(2k (s; x); x) g^s(2(s; x); x)j ds > 1g;

that limk!1 (N (x) ^ t 6= Nk (x) ^ t) = 0: Hence, by the inequality (jZt^Nk (x) (k (x); x) Zt^N (x) ((x); x)j > ) (N (x) ^ t 6= Nk (x) ^ t) + (jZt^N (x) (k (x); x) Zt^N (x) ((x); x)j > ); (2.7.20), and the fact that Zt^N (x) ((x); x) is a proper idempotent variable, limit (2.7.39) would follow by t^Z N (x)

k (s;

0

x) M_ s(x)ds !

t^Z N (x)

(s; x) M_ s (x) ds

0

as k ! 1; (2.7.41)

and t^Z N (x)

g^s

0

(k (s;

x); x)ds !

t^Z N (x) 0

g^s ((s; x); x) ds as k ! 1:

The latter limit obviously follows by (2.7.40). Limit (2.7.41) is proved with a tool we have already used: for > 0, by Young's inequality,

© 2001 by Chapman & Hall/CRC

192

Maxingales

in view of (2.7.17) and (2.7.16), t^ (x) ZN k (s; x) M_ s (x) ds 0

1

Zt

0

t^Z N (x)

(s; x) M_ s (x) ds

0

g^s ((k (s; x) (s; x)); x) + g^s (((s; x)

k (s; x)); x) ds +

which implies (2.7.41) in view of the hypotheses.

1 I(x);

The next lemma and its proof are prompted by Theorem 2.5.11. Lemma 2.7.21. Let 2 ^ and be bounded. Then Z () is a strictly Luzin-continuous local exponential maxingale. Proof. Let us rst assume that is, in addition, locally uniformly continuous in t uniformly in x 2 K (a) for all a 2 (0; 1], i.e., for all T >0 wT;a (Æ) = sup sup j(t; x) (s; x)j ! 0 as Æ ! 0: x2K (a) s;t2[0;T ]: js tjÆ (2.7.42) Let for k 2 N

k (t;

x) =

k2 X i

i=1 (k (t;

k

1

;x

1( i k1 ; ki ](t):

Then the k = x); t 2 R+ ; x 2 C ) are bounded and piecewise constant functions from ^ , which implies by Lemma 2.7.19 that the Z (k ); k 2 N ; are strictly Luzin-continuous local exponential maxingales. Also, since jk (s; x) (s; x)j wt;a (1=k) for s 2 [0; t] and x 2 K (a) if k t, Zt

sup g^s ((k (s; x) (s; x)); x) ds x2K (a) 0

sup

Zt

sup

x2K(a) 0 jjwt;a(1=k)

© 2001 by Chapman & Hall/CRC

g^s (; x) ds:

193

Semimaxingales

Since the right-hand side converges to 0 as k ! 1 by (2.7.42) and condition (II ), we conclude that the k and meet the conditions of Lemma 2.7.20; hence, Z () is a strictly-Luzin local exponential maxingale. Let be an arbitrary bounded function from ^ . We introduce the Steklov functions

k (t;

x) = k

Zt

(s; x)ds; t 2 R+ ; x 2 C ; k

2 N;

(2.7.43)

t 1=k

where (s; x) = 0 if s 0. Then the functions k = (k (t; x); t 2 ^ are bounded and jk (t; x) k (s; x)j R+ ; x 2 C ) 2 2k supv;x j(v; x)jjt sj: Hence, by the part just proved the Z (k ) are strictly Luzin-continuous local exponential maxingales. By Lemma 2.7.20 Z () is a strictly Luzin-continuous local exponential maxingale provided Zt

0

g^s ((k (s; x) (s; x)); x)ds ! 0 as k ! 1;

2 R; t

2 R+ : (2.7.44)

By convexity of g^s (; x) in and (2.7.43) Zt

0

=

g^s ((k (s; x) (s; x)); x)ds Zt

g^s k

0

Zt

Z1=k

0

0

k

Z1=k

0

((s v; x) (s; x))dv; x ds

g^s (((s v; x) (s; x)); x) dv ds

sup

0v1=k

© 2001 by Chapman & Hall/CRC

Zt

0

g^s (((s v; x) (s; x)); x) ds;

194

Maxingales

so we prove (2.7.44) by proving that Zt

0

g^s (((s v; x) (s; x)); x)ds ! 0 as v ! 0:

(2.7.45)

Let a 2 (0; 1]. Firstly, we prove that for every x^ 2 C lim

v!0

Zt

0

sup g^s (((s v; x^ ) (s; x^ )); x)ds = 0:

x2K(a)

(2.7.46)

The argument is standard. If ((s; x^ ); s 2 R+ ) is continuous in s, then (s v; x^ ) (s; x^ ) ! 0 as v ! 0 and the required follows by condition (II ) and boundedness of ((s; x^ ); s 2 R+ ). If ((s; x^ ); s 2 R+ ) is an arbitrary bounded Lebesgue-measurable function, then, given arbitrary " > 0, we can choose by Luzin's R theorem a bounded continuous function (^ (s); s 2 R+ ) such that 0t 1(^ (s) 6= (s; x^ )) ds < "; and (2.7.46) then holds for ((s; x^ ); s 2 R+ ) since it holds for (^ (s); s 2 R+ ) and ((s; x^ ); s 2 R+ ) is bounded. Limit (2.7.46) is proved. We denote

g t;a () = sup sup g^s (; x): st

x2K(a)

By continuity of (s; y) in y on K (a), boundedness of (s; y) and condition (II ) we have that Zt

0

g t;a (3((s; x) (s; y))) ds +

Zt

0

g t;a ( 3((s; x) (s; y))) ds ! 0

as y ! x; y 2 K (a): Hence, for x 2 K (a) and " > 0, there exists

© 2001 by Chapman & Hall/CRC

195

Semimaxingales

an open subset U" (x) of K (a) such that

U" (x) fy 2 K (a) : +

Zt

0 Zt 0

g t;a (3((s; x) (s; y))) ds

g t;a ( 3((s; x) (s; y))) ds < "g:

By compactness of K (a), there exist x1 ; : : : ; xk 2 K (a) such that K (a) [ki=1 U" (xi ), which means that for every x 2 K (a) there exists i 2 f1; : : : ; kg such that Zt

0

g t;a (3((s; xi ) (s; x))) ds +

Zt

0

gt;a ( 3((s; xi ) (s; x))) ds < ": (2.7.47)

Next, by convexity of g^s (; x) in , for x 2 K (a), Zt

0

g^s (((s v; x) (s; x)); x) ds Zt

13 gt;a (3((s v; x) (s v; xi ))) ds

+

Zt

0

0

g t;a (3((s; xi ) (s; x))) ds +

Zt

0

g^s (3((s v; xi ) (s; xi )); x) ds ; (2.7.48)

where i is chosen so that (2.7.47) holds. By (2.7.46), if v is small enough, then for all i = 1; : : : ; k Zt

0

sup g^s (3((s v; xi ) (s; xi )); x) ds < ";

x2K (a)

© 2001 by Chapman & Hall/CRC

196

Maxingales

so, by (2.7.47) and (2.7.48) sup

Zt

x2K (a) 0

g^s (((s v; x) (s; x)); x) ds < ":

Limit (2.7.45) has been proved. Limit (2.7.44) has been proved. Proof of Theorem 2.7.16. Let 2 ^ and A = (rA (t; x); t 2 R + ; x 2 C ), where for 2 R d and A > 0 8 jj A; < ; rA = : jj A; jj > A: Since the map ! rA is continuous, it follows that A 2 ^ . Also A is bounded so that by Lemma 2.7.21 Z (A ) is a strictly Luzincontinuous local exponential maxingale. Since g^t (; x) is non-negative, convex in and g^t (0; x) = 0, we have that g^t (rA ; x) is increasing in A, so g^t (rA ; x) " g^t (; x) as A ! 1: (2.7.49) Let N be de ned by (2.7.36). By (2.7.49) and Lemma 2.7.18 the collections of strictly Luzin variables fZt^N (x) ((x); x); t 2 R+ g and fZt^N (x) (A(x); x); t 2 R+ ; A 2 R+ g are uniformly maximable. In particular, (Zt^N (x) (A (x); x); t 2 R+ ; x 2 C ) is a uniformly maximable exponential maxingale, so by Lemma 2.7.17 the theorem is proved if lim S Zt^N (x) (A (x); x)f (x) = S Zt^N (x) ((x); x)f (x) A!1 for every bounded, continuous and non-negative f , which by The orem 1.4.19 is implied by the convergence Zt^N (x) (A (x); x) ! Zt^N (x) ((x); x). Using the fact that Zt^N (x) ((x); x) is a proper idempotent variable, we prove the latter convergence by proving that as A ! 1 Zt

0

g^s ((s; x); x) g^s (rA (s; x); x) ds ! 0;

Zt

0

j((s; x) rA(s; x)) M_ s(x)jds ! 0:

© 2001 by Chapman & Hall/CRC

(2.7.50a) (2.7.50b)

197

Semimaxingales

The rst convergence follows by (2.7.49), the fact that the integral in (2.7.50a) is a strictly Luzin variable (use condition (II ) and (2.7.19)) and Dini's theorem. To prove (2.7.50b) we write as in the proof of Lemma 2.7.17 for > 0 Zt

0

j((s; x) rA(s; x)) M_ s(x)jds

1

Zt

0

g^s (((s; x) rA (s; x)); x)

1 + g^s ( ((s; x) rA (s; x)); x) ds + I (x): (2.7.51) Next, by the de nition of rA , convexity of g^s (; x) in and the fact that g^s (0; x) = 0 g^s (((s; x) rA (s; x)); x) g^s ((s; x); x) 1(j(s; x)j > A); g^s ( ((s; x) rA (s; x)); x) g^s ( (s; x); x) 1(j(s; x)j > A): Hence, as 2 ^ , (2.7.19) implies that Zt

0

g^s (((s; x) rA (s; x)); x) + g^s ( ((s; x) rA (s; x)); x) ds

! 0 as A ! 1;

so, since is arbitrary, (2.7.51) yields (2.7.50b). As a byproduct, we can prove that certain integrals with respect to X are semimaxingales. De nition 2.7.22. Let be the subset of ^ consisting of functions such that for t 2 R+ Zt

and

0 Zt

0

j(s; x) bs (x)j ds < 1; x 2 C ; j(s; x)bs (x)j 1(j(s; x)j > A) ds ! 0

© 2001 by Chapman & Hall/CRC

as A ! 1.

198

Maxingales

For 2 , we de ne the idempotent process X = ( Xt ; t 2 R + ) by

Xt =

Zt

0

Zt

(s; x)bs (x) ds+ (s; x)M_ s (x) ds;

(2.7.52)

0

if the integrals are well de ned and nite, and Xt = X^ t otherwise, where X^ is a continuous idempotent process. By Lemma 2.7.15 and the de nition of we have that (2.7.52) holds -a.e.

Theorem 2.7.23. Let gs(; x) be given by (2.7.13), conditions (I ) and (II ) hold, and 2 . Then the idempotent process X is a strictly Luzin-continuous semimaxingale on (C ; C ; ) with cumulant G () = (Gt (; x); t 2 R+ ; x 2 C ) given by Gt (; x) =

Zt

0

gs ((s; x); x) ds; t 2 R+ ; 2 R; x 2 C :

Proof. By Theorem 2.7.16 (exp( Xt Gt ()); t 2 R+ ) is a strictly Luzin-continuous local exponential maxingale. The fact that both X and G () are strictly Luzin-continuous C -adapted idempotent processes follows by the proof of Lemma 2.7.17, condition (I ), C progressive measurability of (bs (x)), and the de nition of .

In large deviation limit theorems G() is often more speci c than in (2.7.10) and de ned in terms of \characteristics", which we now introduce. Let cs (x); s 2 R+ ; x 2 C be a C -progressively measurable idempotent process with values in theR space of symmetric, positive semi-de nite d d-matrices such that 0t kcs (x)k ds < 1 for t 2 R+ and x 2 C ; s( ; x); s 2 R+ ; 2 B(Rd ); x 2 C be a transition kernel (for each x) from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that

© 2001 by Chapman & Hall/CRC

199

Semimaxingales

for t 2 R+ ; x 2 C and 2 R+

t (f0g; x) = 0; Zt Z

0

Z

Zt Z

0

Rd

Rd

Z

Rd

jxj2 ^ 1 t (dx; x) < 1;

jxj2 ^ 1 s(dx; x) ds < 1;

ejxj 1(jxj > 1) t (dx; x) < 1;

ejxj 1(jxj > 1) s (dx; x) ds < 1;

Rd

(2.7.53a)

(2.7.53b)

and the functions ( Rd f (x) s(dx; x); s 2 R+ ) are C -progressively measurable for Borel functions f such that the integrals are well de ned; ^s( ; x); s 2 R+ ; 2 B(Rd ); x 2 C be a transition kernel (for each x) from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for s 2 R + ; x 2 C and 2 B (R d ) R

^s ( ; x) s ( ; x); ^s(Rd ; x) 1;

(2.7.54)

and the functions ( Rd f (x) ^s(dx; x); s 2 R+ ) are C -progressively measurable for Borel functions f such that the integrals are well de ned. We say that the semimaxingale X has local characteristics (b; c; ; ^), where (bs (x)) is as above, if the associated cumulant is given by (2.7.7), where R

1 gs (; x) = bs (x) + cs (x) + 2

+ ln 1 +

Z

(ex

Rd

1)^s (dx; x)

Z

Rd

(ex 1 x)s (dx; x) Z Rd

(ex

1)^s (dx; x) : (2.7.55)

Remark 2.7.24. The right-hand sidedis well de ned since RRd (ex 1)^s (dx; x) > R 1 by the fact that ^s(R ; x) 1. It is also not diÆcult to check that 0t supjjAjgs (; x)jds < 1 for t 2 R+ and A 2 R+ . © 2001 by Chapman & Hall/CRC

200

Maxingales

Let

Ct (x) =

Zt

0

cs (x) ds:

(2.7.56)

We call the idempotent process B 0 = (Bt0 (x); t 2 R+ ; x 2 C ) (de ned by (2.7.11)) the rst characteristic of X \without truncation", C = (Ct (x); t 2 R+ ; x 2 C ) the second characteristic, s(dx; x) the density of the measure of jumps, and ^s(dx; x) the density of the discontinuous measure of jumps. The quadruplet (B 0 ; C; ; ^) is referred to as the characteristics of the semimaxingale X \without truncation". In large deviation limit theorems we will also need characteristics \associated with limiters".

De nition 2.7.25. A Borel function h :

Rd

! Rd

is said to be a limiter if it is bounded and h(x) = x in a neighbourhood of the origin.

For Borel functions f , for which the integrals below are well de ned, we introduce the notation

f (x) t (x) =

Zt Z

0

Rd

f (x) s(dx; x)ds; f (x) ^s (x) =

Z Rd

f (x)^s (dx; x):

The rst characteristic of X associated with a limiter h is an idempotent process B = (Bt (x); t 2 R+ ; x 2 C ) de ned by

Bt (x) = Bt0 (x)+(h(x) x) t (x):

(2.7.57)

The modi ed second characteristic associated with h is an idempotent process C~ = (C~t (x); t 2 R+ ; x 2 C ) such that the C~t (x) are symmetric positive semi-de nite d d-matrices, speci ed by the equalities

C~t (x) = Ct (x) + ( h(x))2 t (x) Zt

0

© 2001 by Chapman & Hall/CRC

( h(x) ^s (x))2 ds; 2 Rd : (2.7.58)

201

Semimaxingales

Analogously, the modi ed second characteristic \without truncation" C~ 0 = (C~t0 (x); t 2 R+ ; x 2 C ) is speci ed by the equalities

C~t0 (x) = Ct (x)+(x)2 t (x)

Zt

0

(x^s (x))2 ds:

(2.7.59) Note that (2.7.57) implies that if B = (Bt (x); t 2 R+ ; x 2 C ) is the rst characteristic associated with a limiter h(x), then

Bt (x) = Bt (x)+(h (x) h(x)) t (x): (2.7.60) Obviously, Bt (x); Ct (x); C~t (x); C~t0 (x); and f (x) t (x) (when well de ned) are continuous in t and Ct {measurable in x. We note that the cumulant assumes the form 1 Gt (; x) = Bt (x) + Ct (x) + (ex 1 h(x)) t (x) 2 +

Zt

ln 1 + (ex

0

Remark 2.7.26.

1) ^s (x) (ex

1) ^s(x) ds:

(2.7.61)

The de nition of the characteristics of a semimaxingale is motivated by the de nition of the characteristics of a semimartingale and the fact that semimaxingales are \large deviation limits" of semimartingales as is shown in part II. Note that the expression (2.7.61) for the cumulant is analogous to the logarithm of the stochastic exponential of a semimartingale, Liptser and Shiryaev [79], Jacod and Shiryaev [67]. Thus, Bt is \the drift term", Ct is \the diusion term", s ds is \the predictable measure of jumps", and ^s ds is \the discontinuous part of the predictable measure of jumps". As we will see in part II, this analogy is not only in form, but it also helps us to formulate conditions under which large deviation limit theorems for semimartingales can be proved. The terminology \truncated" and \nontruncated" characteristics is also inherited from semimartingales. The analogy with semimartingales would be more complete if we required in De nition 2.7.1 that Gt (; !) be, in addition, of locally bounded variation in t. In fact, this property holds if a semimaxingale admits characteristics as is the case in all our examples of semimaxingales. However, since a number of

© 2001 by Chapman & Hall/CRC

202

Maxingales

properties of semimaxingales do not depend on G() being of locally bounded variation, we have decided not to include this requirement in the de nition. We now state conditions on the characteristics that imply conditions (I ) and (II ). Lemma 2.7.27. Let the canonical idempotent process X be a semimaxingale on (C ; C ; ) with local characteristics (Rb; c; ; ^), where b and c are strictly Luzin R idempotent processes, and Rd (exp( x) 1 x) s (dx; x) and Rd exp( x) ^s (dx; x) are continuous in (; x) when restricted to Rd K (a) for a 2 (0; 1] and s 2 R+ . If for all a 2 (0; 1], A 2 R+ and t 2 R+ Zt

sup jbs (x)j ds < 1; sup sup kcs (x)k < 1; st x2K (a)

0

x2K (a)

sup sup sup st x2K (a) jjA

Z

1 x)s (dx; x) < 1;

(ex

Rd

lim sup sup

Z

!0 st x2K (a)

1 x)s (dx; x) = 0;

(ex

Rd

sup sup sup st x2K (a) jjA

Z lim sup sup !0 st x2K (a) Rd

Z

Rd

ex ^s (dx; x) ds < 1;

ex 1 ^s (dx; x) = 0;

then the associated cumulant satis es conditions (I ) and (II ).

2.8 Maxingale problems In this section we are concerned with identifying deviabilities for which the canonical idempotent process on C (R + ; R d ) is a semimaxingale with a given cumulant. As in the preceding section we denote C = C (R + ; Rd ), Ct = Ct (R+ ; Rd ), and C = (Ct ; t 2 R+ ). We also assume as given a cumulant G() on C . Let x 2 Rd . De nition 2.8.1. We say that a deviability on C is a solution to the maxingale problem (x; G) if the canonical process X is a semimaxingale with cumulant G() on (C ; C; ) such that X0 = x -a.e.

© 2001 by Chapman & Hall/CRC

203

Maxingale problems

Examples are provided by the Wiener and Poisson idempotent processes: by Theorem 2.4.2 the Wiener idempotent probability on C (R + ; R) solves the maxingale problem (0; G) with Gt (; x) = 2 t=2 and by Theorem 2.4.16 the Poisson idempotent probability on C (R + ; R) solves the maxingale problem (0; G) with Gt (; x) = (exp 1)t. In view of applications to large deviation theory, we are interested in nding conditions when a maxingale problem has a unique solution. Existence issues are of less importance to us. Besides, large deviation convergence theorems of part II also imply that under their hypotheses the associated maxingale problems have solutions.

De nition 2.8.2. We say that uniqueness holds for the maxingale problem (x; G) if it has at most one solution.

Our candidate for a solution of (x; G) is the idempotent measure

x de ned in (2.7.6). We begin with the case where the cumulant does not depend on x. Lemma 2.8.3. Let Gt (; x) not depend on x 2 C , be dierentiable in for all t 2 R+ and the dierences Gt () Gs () be convex in for all 0 s < t. Then the idempotent measures t1 ;:::;tk on (Rd )k , where 0 t1 < t2 < : : : < tk , speci ed by the densities

t1 ;:::;tk (x1 ; : : : ; xk ) =

k Y

inf e (xi

d i=1 2R

xi 1 ) eGti () Gti 1 () ;

where t0 = 0 and x0 = x, form a projective system of deviabilities, which has x as the projective limit. Proof. By dierentiability of Gt (), Lemma 1.11.4 and Lemma 1.11.7 the t1 ;:::;tk are deviabilities. In order to prove that they form a projective system it suÆces to check that, given ti 1 < ti < ti+1 , xi 1 and xi+1 , we have

inf sup 1 (xi xi 1 )+2 (xi+1 xi) (Gti (1 ) Gti 1 (1 ))

xi 2Rd 1 2Rd ; 2 2Rd (Gti+1 (2 )

Gti (2 )) = sup (xi+1 2Rd

© 2001 by Chapman & Hall/CRC

xi 1 ) (Gti+1 () Gti 1 ()) :

204

Maxingales

The latter equality follows by a minimax argument, see, e.g., Aubin and Ekeland [6]. By Theorem 2.2.4 the projective limit of the t1 ;:::;tk coincides with x .

Remark 2.8.4. hypotheses.

As a consequence,

x is a deviability under the

The following existence and uniqueness result, which is a converse to Theorem 2.7.3, shows that the distributions of certain idempotent processes with independent increments are uniquely speci ed by the associated cumulants. We denote Gt () = sup0st jGs ()j.

Theorem 2.8.5. Let Gt (; x) not depend on x 2 C . Let, in addition, Gt () be dierentiable in for all t 2 R+ and the dierences Gt () Gs () be convex in for all 0 s < t. Then x is a unique solution to problem (x; G). The canonical idempotent process X has independent increments and starts at x under x . Proof. We de ne t1 ;:::;tk as in Lemma 2.8.3. The construction of the t1 ;:::;tk and the fact that they induce deviability x on C imply that the canonical idempotent process X has independent increments and starts at x under x , and Sx exp((Xt Xs)) = exp(Gt () Gs ())

so that X is a semimaxingale with cumulant Gt (). We prove uniqueness. Let be a solution of (x; G) and X be the canonical idempotent process on C . We note that for every n 2 N and 2 Rd the family fYt^n (); t 2 R+ g is {uniformly maximable. (Recall that Y () is de ned by (2.7.1).) To see this, we write by (2.7.1)

S Yt^n ()2 = sup Yt^n (2; x) exp( 2Gt^n ()) exp(Gt^n (2))(x)

x2C

SYt^n(2) exp(2Gn () + Gn (2)):

Since S Yt^n (2) SY0 (2) = 1 by Lemma 2.3.13, the uniform maximability is proved. Then by \the Doob stopping theorem" (Theorem 2.3.8) the sequence fn; n 2 N g is a localising sequence for Y () and, in particular, Y () is a C{exponential maxingale under . Therefore, for 0 s < t,

S exp((Xt Xs ))jCs = exp(Gt () Gs ());

© 2001 by Chapman & Hall/CRC

(2.8.1)

205

Maxingale problems

so by Lemma 1.11.9 X has independent increments under . Let X t0 ;t1 ;:::;tk , 0 = t0 < t1 < t1 < : : : < tk , denote nite-dimensional idempotent distributions of X so that X t0 ;t1 ;:::;tk (x0 ; x1 ; : : : ; xk ) = (Xt0 = x0 ; Xt1 = x1 ; : : : ; Xtk = xk ). By independence of increments of X we have that X t0 ;t1 ;:::;tk (x0 ; x1 ; : : : ; xk ) = Qk X X 0 (x0 ) i=1 ti 1 ;ti (xi 1 ; xi ). Also by (2.8.1) and Lemma 1.11.5 X ti

e 1 ;ti (xi 1 ; xi ) = inf 2Rd

(xi xi 1 ) exp(G t

i ()

Gti 1 ()):

Thus, X t0 ;t1 ;:::;tk = t0 ;t1 ;:::;tk Since by Theorem 2.2.2 and Corollary 1.7.12 X (x) = inf t1 ;:::;tk X t0 ;t1 ;:::;tk (xt0 ; xt1 ; : : : ; xtk ), the required follows by Lemma 2.8.3.

Remark 2.8.6. Note also that if X is an idempotent process with

independent increments and cumulant G() on an idempotent prob- ability space ( ; ), then Gt () Gs () = ln S exp (Xt Xs ) so that Gt () Gs () is convex for s < t.

If Gt () is given in terms of characteristics by (2.7.7) and (2.7.55), we have the following consequence.

Corollary 2.8.7.

Let Gt () have the form (2.7.7) and (2.7.55), where bs , cs , s , and ^s do not depend on x. If, in addition, (L0 ) 1 + inf e jxj 1 ^s > 0; t 2 R+ ; 2 R+ ; st then x is a unique deviability on C such that X is a semimaxingale starting at x with independent increments and local characteristics (b; c; ; ^) under x . Proof. By Theorem 2.8.5 we only need to prove that Gt () is dierentiable in , which follows by condition (L0 ).

The next lemma gives suÆcient conditions for condition (L0 ) to hold.

Lemma 2.8.8. Condition (L0) holds if at least one of the following conditions holds

1. alim ^ (jxj a) > 0; t 2 R+ ; !1 sinf t s 2. sup ^s (Rd ) < 1; t 2 R+ ; st

© 2001 by Chapman & Hall/CRC

206

Maxingales

3. for every t 2 R+ there exists > 0 such that sup ejxj ^s < 1: st Proof. It is straightforward to see that conditions 1 and 2 imply (L0 ). Let condition 3 hold. We have for 2 R+ , s 2 R+ and b 2 R+

1 + (e jxj 1) ^s 1 + e jxj 1(jxj < b) 1 ^s 1 + e b=(1 e b ejxj) 1 ^s = 1 + (e b= 1)^s (Rd ) e b= e b ejxj ^s : (2.8.2)

Since ^s (Rd ) 2 [0; 1], it follows that 1 + (e b= 1)^s (Rd ) e b= , and, hence picking b such that eb 2 supst ejxj ^s ; which is possible by hypotheses, we conclude that the left-most side of (2.8.2) is bounded from below by e b= =2 for s t. We now consider the case where the cumulant depends on x. The following is an existence result for a diusion maxingale problem. It is a consequence of Theorem 2.5.19 and Theorem 2.6.24. We denote xt = sup0stjxsj. Theorem 2.8.9. Let the maxingale problem (x; G) be speci ed by the cumulant

Gt (; x) =

Zt

0

1 bs(x) ds + 2

Zt

0

cs (x) ds;

where functions (bs (x)) and (cs (x)), assuming values in Rd and the space of symmetric positive semi-de nite d d-matrices, respectively, are C-progressively measurable and continuous in x, and Rt Rt j b ( x ) j ds < 1 and k c ( x ) 0 s 0 s k ds < 1. Let also bt (x) and ct (x) satisfy the linear-growth conditions jbt (x)j lt(1+x ); kct (x)k lt(1+x 2 ); t 2 R+ ; x 2 C ; t

t

Rt 0 ls ds

0. Then for s t by the de nition of X;N1 ;N2

S (Zt jAs^N )(x0 ; n01 ; n02 ) =

sup

(x;n1 ;n2 )2C (R+ ;R3 ) t^ZN (x) t^ZN (x) exp n1 us( ) ds (e 1) us( ) ds 0 0 t^ZN (x) t^ZN (x) + n2 vs ( ) ds (e 1) vs ( ) ds 0 0 X; N ; N 1 2 ( ; n1 ; n2 ) s^N ( 0 ; n01 ; n02 ): (2.8.3)

x

x

x

x

x

x

jA

Since X;N1 ;N2 (x; n1 ; n2 )jAs^N (x0 ; n01 ; n02 ) N (R s^N (x0 ) ur (x0) dr n1)N (R s^N (x0 ) vr (x0 ) dr n2); 0 0

it follows that the right-hand side of (2.8.3) is not greater than Zs^N (x0 ) (x0 ). For the reverse inequality, let x^ be de ned on [0; s ^ N (x0 )] by x^ (q) = x0 (q), on [s ^ N (x0 ); t ^ N (^x)] as a solution of the equation

x^ q = x0

s^N (x0 )

Zq

+ e

© 2001 by Chapman & Hall/CRC

s^N (x0 )

ur (^x) dr e

Zq

s^N (x0 )

vr (^x) dr;

209

Maxingale problems

and on [t ^ N (^x); 1) by Zq

x^ q = x^ t^N (^x) +

t^N (^x)

ur (^x) dr

Zq

t^N (^x)

vr (^x) dr:

The latter two equations have solutions by a standard R 0 argument. We de ne n^ 1 on 0; 0s^N (x ) ur (x0 ) dr by R R 0 n^ 1 (q) = n01 (q), on 0s^N (x ) ur (x0 ) dr; 0t^N (^x) ur (^x) dr by R R s^N (x0 ) 0 0 ) dr and n^ 1 (q) = n01 0s^N (x ) ur (x0 ) dr + e q u ( x r 0 R R on 0t^N (^x) ur (^x) dr; 1) by n^ 1 (q) = n^ 1 0t^N (^x) ur (^x) dr + R t^N (^x) q ur (^x) dr . Similarly, n^ 2 (q) = n02 (q) on 0 R s^N (x0 ) R 0 0; 0 vr (x0 ) dr , n^ 2 (q) = n02 0s^N (x ) vr (x0 ) dr + e q R s^N (x0 ) R s^N (x0 ) R vr (x0 ) dr on vr (x0 ) dr; 0t^N (^x) vr (^x) dr , 0 0 R R t^N (^x) and n^ 2 (q) = n^ 2 0t^N (^x) vr (^x) dr + q vr (^x) dr on 0 R t^N (^x) vr (^x) dr; 1). Then (^x; n^ 1 ; n^ 2 ) satis es equation (2.6.11), 0 X; N ; N 1 2 (^ so x; n^ 1 ; n^ 2) = N (^n1 )N (^n2). Since

exp n^ 1

t^ZN (^x)

ur (^x) dr

0

= exp n01

(e

ur (x0 ) dr

0

(e

1)

us (^x) ds N (^n1 )

0

s^N (x0 ) Z

1)

t^ZN (^x)

s^Z N (x0 ) 0

ur (x0 ) dr N [n01 ]As^N (x0 )

and

exp

n^ 2

t^ZN (^x) 0

© 2001 by Chapman & Hall/CRC

vr (^x) dr (e

1)

t^ZN (^x) 0

vs (^x) ds N (^n2 )

210

Maxingales

= exp

n02

s^Z N (x0 )

vr (x0 ) dr

0

(e

1)

s^Z N (x0 ) 0

vr (x0 ) dr N [n02 ]As^N (x0 ) ;

we conclude that the expression in the supremum on the right-hand side of (2.8.3) evaluated at (^x; n^ 1 ; n^ 2 ) equals Zs^N (x0 ) (x0 ). Thus, S (Zt^N jAs^N ) = Zs^N . Therefore,

S (Zt^N jAs ) = S (Zt^N jAs^N ) 1(s N ) _ ZN 1(s > N ) = Zs^N 1(s N ) _ ZN 1(s > N ) = Zs^N :

We now study the uniqueness issue. We introduce for future use for t 2 R+

It (x) = x;t(x) = exp(

Zt

sup

((s))20

0

(s) dxs dGs ((s); x)); (2.8.4) (

x0 = x; Ix;t(x) = I1t(;x); ifotherwise ; Ix;t(x)); x;t( ) = sup x;t(x); C : x2C

(2.8.5) (2.8.6)

As we have seen, x is a natural candidate for a solution to (x; G). We rst show that it is a tight -smooth idempotent measure on C under fairly general assumptions.

De nition 2.8.11. We say that G() satis es the linear-growth con-

dition if there exist R+ {valued, increasing and continuous in t functions F l () = (Ftl (); t 2 R+ ); 2 Rd ; such that F0l () = Ftl (0) = 0 and for some increasing function kt 2 R+ we have for all 0 s < t , x 2 C and 2 Rd Gt (; x) Gs (; x) F l ((1+kt x )) F l ((1+kt x )): t

t

s

t

Lemma 2.8.12. Let G() satisfy the linear-growth condition. Then x is a tight -smooth idempotent measure on C .

© 2001 by Chapman & Hall/CRC

211

Maxingale problems

Proof. With no loss of generality we assume that x = 0 and check that I0 (x) = ln 0 (x) de ned by (2.7.2), (2.7.3), and (2.7.5), where x = 0, is a tight rate function on C in the sense of Remark 1.7.17, i.e., the sets LI0 (a) = fx 2 C : I0 (x) ag are compact for all a 2 R+ . Let x 2 LI0 (a). Then x0 = 0 by (2.7.5). By (2.7.2), (2.7.3), (2.7.5), and the linear-growth condition we have for 0 s < t, denoting by ei ; i = 1; : : : ; 2d; the d-vector, whose b(i + 1)=2cth entry is 1 if i is odd, 1 if i is even, and the rest of the entries are equal to 0, jxt xsj d max ei (xt xs) i=1;:::;2d 1 + kt xt 1 + kt xt ei ei d i=1max Gt G +a s ;:::;2d 1 + kt xt 1 + kt xt d max Ftl (ei ) Fsl (ei ) + da: (2.8.7) i=1;:::;2d

Since Ftl () is increasing in t, x has bounded variation over bounded intervals; therefore, since kt is increasing, for T > 0 ZT

0

d Vart x 1 + kT xt

2d X

d

i=1

FTl (ei )+ da

(recall that F0l () = 0), which implies that ZT

0

dxt 1 + cT xt

where cT = kT that

d

2d X

i=1

FTl (ei )+ da;

_ 1. By (2.8.8) and the fact that x0 = 0 we deduce 2d

X ln(1+ cT xT ) cT d( FTl (ei )+ a):

i=1

Hence, sup

x2L 0 (a) I

(2.8.8)

xT < 1:

(2.8.9) (2.8.10)

In analogy with (2.8.7) we can write for T > 0 and b > 0 jx x j b ei (xt xs ) b t s d max i=1;:::;2d 1 + kt xt 1 + kt xt d max Ftl (bei ) Fsl (bei ) + da: i=1;:::;2d

© 2001 by Chapman & Hall/CRC

212

Maxingales

Therefore, for Æ > 0 sup

x2L 0 (a) I

jxt xsj

sup

s;t2[0;T ]: js tjÆ

db i=1max sup ;:::;2d

s;t2[0;T ]: js tjÆ

l F (bei )

t

Fsl (bei ) +a (1+kT sup

x2L 0 (a) I

so by continuity of Ftl () in t lim sup sup Æ!0 x2LI0 (a)

sup

s;t2[0;T ]: js tjÆ

jxt xsj

xT ); !

da 1 + kT sup xT : b x2LI0 (a)

Letting b ! 1 and using (2.8.10), we conclude that the left-hand side of the latter inequality is 0. An application of Arzela{Ascoli's theorem ends the proof.

Remark 2.8.13. The above proof shows that if Ix;T (x) a, then the bound (2.8.9) holds, which implies that the sets [st fxs : x;s(x)

g are bounded under the hypotheses for t 2 R+ and 2 (0; 1]. By Lemma 2.7.11 if solves the maxingale problem (x; G), then (x) x (x); x 2 C : (2.8.11) Our goal is to establish conditions when we actually have equality above. Clearly, we need only to be concerned with the case x (x) > 0. We assume in the sequel that the cumulant is given by R t (2.7.7) and (2.7.13), where (bs (x)) is C-progressively measurable, 0 jbs (x)jds < 1, (^gs(; x)) is non-negative, continuous in and B([0; t]) B(Rd )

Ct =B(R+R ){measurable as a map from [0; t] R d C to R+ , g^s (0; x) = t 0, and 0 supjj=A g^s (; x) ds < 1 for t 2 R+ ; A 2 R+ and x 2 C . By Lemma 2.7.12 if x (x) > 0; then x is absolutely continuous. We note that by (2.7.2), (2.7.5), (2.7.6), (2.8.4), (2.8.5), and (2.8.6) x;t(x) # x(x); x 2 C ; as t ! 1: (2.8.12) The argument of the proof of Lemma 2.7.12 also shows that 8 t Z > >

0 > :

hs (x_ s ; x) ds

1

© 2001 by Chapman & Hall/CRC

if x is absolutely continuous on [0; t]; otherwise.

(2.8.13)

213

Maxingale problems

For the following theorem we recall that Z () = (Zt (; x); t 2 R + ; x 2 C ) is given by (2.7.20) if the integral on the right-hand side is well de ned and nite, and Zt (; x) = 0 otherwise.

Theorem 2.8.14. Let deviability solve dthe maxingale problem (x; G) and x^ 2 C . Let there exist an R {valued function ^ = (^ (s; x); s 2 R+ ; x 2 C ) with the following properties. a) Z (^ ) is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ) and admits a localising sequence of strictly Luzin C-stopping times; b) if x~ is such that x (~x) > 0 and a.e. in s 2 [0; t]

^ (s; x~ )x~_ s gs (^ (s; x~ ); x~ ) = sup (x~_ s gs (; x~ )); 2Rd

then x~ s = x^ s ; s 2 [0; t]; where t 2 R+ .

(2.8.14)

Then (ps 1 Æ ps x^ ) = x;s(^x); s 2 R+ ; and (^x) = x (^x).

Proof. Let fN ; N 2 N g be a localising sequence of strictly Luzin Cstopping times for Z (^ ) so that the (Zt^N (^ ); t 2 R+ ) are strictly Luzin uniformly maximable exponential maxingales under . Since S Zt^N (^ ) = 1 and limA!1 S Zt^N (^ ) 1(Zt^N (^ ) > A) = 0; for A large enough S Zt^N (^ ) 1(Zt^N (^ ) A) = 1 so that the inequality

S Zt^N (^ ) 1(Zt^N (^ ) A) sup Zt^N (x) (^(x); x)(x) _ (aA);

x2K(a)

where a 2 (0; 1], implies that for a small enough supx2K(a) Zt^N (x) (^ (x); x)(x) 1: On the other hand, by the de nitions of Z (^ ) and x , and (2.8.11)

Zt^N (x) ((x); x)(x) Zt^N (x) ((x); x)x (x) 1:

Thus, if a > 0 is small enough, then sup Zt^N (x) (^ (x); x)(x) = 1:

x2K (a)

© 2001 by Chapman & Hall/CRC

(2.8.15) (2.8.16)

214

Maxingales

Being a strictly Luzin idempotent variable on (C ; ), Zt^N (x) (^ (x); x) is continuous in x when restricted to K (a). Also (x), being a deviability density, is upper semi-continuous. Therefore, the product Zt^N (x) (^ (x); x)(x) is upper semicontinuous when restricted to K (a). As the latter set is compact, the supremum in (2.8.16) is attained, so for some xN 2 C

Zt^N (xN ) (^ (xN ); xN )(xN ) = 1

(2.8.17)

(we suppress in xN dependence on t). Then by (2.8.15)

Zt^N (xN ) (^ (xN ); xN )x (xN ) = 1:

(2.8.18)

In particular, x (xN ) > 0 so that by (2.7.6) and Lemma 2.7.12 xN is absolutely continuous and xN0 = x. Thus, by the de nitions of Z and x as well as Lemma 2.7.12 we have that almost everywhere on [0; t ^ N (xN )]

^ (s; xN )x_ Ns gs (^ (s; xN ); xN ) = sup (x_ Ns gs (; xN )) 2Rd

and Z1

t^N (xN )

sup ( x_ Ns gs (; xN )) ds = 0: 2Rd

(2.8.19) (2.8.20)

Hence, by the requirements on ^ we conclude that xNs = x^ s ; 0 s t ^ N (xN ). Since t ^ N (x) is a C{stopping time, by Lemma 2.2.21 t ^ N (xN ) = t ^ N (^x). Therefore, by (2.7.5), (2.7.6), Lemma 2.7.12, (2.8.5), (2.8.6), (2.8.20), and (2.8.13)

x;t^N (^x) (^x) = x;t^N (xN )(xN ) = x(xN ): (2.8.21) Thus, by (2.8.17), (2.8.18) and (2.8.21), (xN ) = x(xN ) = x;t^N (^x) (^x); which implies, since xN 2 pt^1N (^x) Æ pt^N (^x) x^ , that (pt^1N (^x) Æpt^N (^x) x^ ) (xN ) = x;t^N (^x) (^x): On the other hand, by (2.8.12)

© 2001 by Chapman & Hall/CRC

(2.8.22)

x;t(x) x(x) for x 2 C

and

Maxingale problems

215

t 2 R+ , so by (2.8.11), (2.8.4), (2.8.5), (2.8.6), and Lemma 2.2.21

(pt^1N (^x) Æ pt^N (^x) x^ ) = supf(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g supfx(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g supfx;t^N (x)(x); x 2 pt^1N (^x) Æ pt^N (^x) x^ g = x;t^N (^x) (^x): Comparing this with (2.8.22) yields (pt^1N (^x) Æ pt^N (^x) x^ ) = x;t^N (^x) (^x): (2.8.23) 1 ^ = pt 1 Æpt x^ ; Since N (^x) ! 1, it follows that \1 N =1 pt^N (^x) Æpt^N (^x) x so since the sets pt^1N (^x) Æpt^N (^x) x^ are closed, by the -smoothness property of deviability (pt 1 Æpt x^ ) = lim (pt^1N (^x) Æpt^N (^x) x^ ): (2.8.24) N !1 Also the convergence N (^x) ! 1, (2.8.4), (2.8.5), and (2.8.6) yield x;t(^x) = Nlim (^x): (2.8.25) !1 x;t^N (^x) Putting together (2.8.23), (2.8.24) and (2.8.25) results in the equality (pt 1 Æ pt x^ ) = x;t (^x). The nal assertion follows by taking in both sides of the latter equality the limit as t ! 1, using -smoothness of and (2.8.12).

Remark 2.8.15. In the sequel we routinely omit indications that cer-

tain relations hold almost everywhere with respect to Lebesgue measure when this is understood. Conditions for Z () to be a strictly Luzin-continuous exponential maxingale on (C ; C ; ) are given in Theorem 2.7.16. However, the deviability is not known to us, so it would be diÆcult to verify the hypotheses of the theorem. Therefore, we introduce somewhat cruder conditions, which have the advantage of not involving . The following conditions replace conditions (I ) and (II ). I The function bs (x) is continuous in x and Zt

0

sup jbs (x)j ds < 1

x2K

for all compacts K C and t 2 R+ .

© 2001 by Chapman & Hall/CRC

216

Maxingales

II The function g^s (; x) is convex in 2 Rd , continuous in (; x) R d C , and sup sup sup g^s (; x) < 1; jjA st x2K

2

lim sup sup g^s (; x) = 0 x2K

!0 st

for all compacts K C , t 2 R+ and A 2 R+ .

If gs (; x) has the form (2.7.55), Lemma 2.7.27 provides suÆcient conditions for conditions I and II to hold in terms of characteristics.

Lemma 2.8.16. Let gs (; x) be given by (2.7.55) , where cs (x); s 2 d R + ; x 2 C , s ( ; x); s 2 R + ; 2 B (R ); x 2 C , and ^s ( ; x); s 2 R+ ; 2 B(Rd ); x 2 C are as de ned in Section 2.7 with Cprogressive measurability replaced by C-progressive measurability. R Let bs (x) and cs (x)R be continuous in x, and Rd (exp( x) 1 x) s (dx; x) and Rd exp( x) ^s (dx; x) be continuous in (; x). If for all compacts K C , A 2 R+ and t 2 R+ Zt

0

sup jbs (x)j ds < 1; sup sup kcs (x)k < 1; st x2K

x2K

sup sup st x2K

Z

Rd

lim sup sup x2K

!0 st

1 Ajxj) s (dx; x) < 1;

(eAjxj Z

1 x) s (dx; x) = 0;

(ex

Rd

sup sup st x2K

Z Rd

Z lim sup sup !0 st x2K Rd

eAjxj ^s(dx; x) ds < 1; ex

)

1 ^s (dx; x = 0;

then conditions I and II are satis ed.

We now de ne the class of integrands.

De nition 2.8.17. Let ^ denote the set of all Rd {valued C{ progressively measurable idempotent processes = ((t; x); t 2 R + ; x 2 C ) such that the (t; x) are continuous in x, for 2 R , © 2001 by Chapman & Hall/CRC

217

Maxingale problems

t 2 R+ and x 2 C Zt

0

g^s ((s; x); x) ds < 1

and for every compact K C

sup

Zt

x2K 0

g^s ((s; x); x) 1(j(s; x)j > A) ds ! 0 as A ! 1:

Obviously, ^ ^ for every deviability on C .

Remark 2.8.18. If x is a deviability, then both in conditions I and II and the de nition of ^ we could consider only compacts Kx (a), where a 2 (0; 1] and Kx (a) = fx : x (x) ag, and require that x be such that x (x) > 0. The following consequence of Theorem 2.7.16 allows us to check that Z () satis es condition a) of Theorem 2.8.14. Theorem 2.8.19. Let conditions I and II hold. If 2 ^ and solves the maxingale problem (x; G), then the idempotent process Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ), which admits a localising sequence of strictly Luzin Cstopping times. Proof. By Theorem 2.7.16 Z () is a strictly Luzin-continuous local exponential maxingale on (C ; C ; ). It is straightforward to check that the sequence fN ; N 2 N g de ned by

N (x) = inf ft 2 R+ :

Zt

0

g^s (2(s; x); x) ds+t N g

is a localising sequence of strictly Luzin-continuous C-stopping times. We now consider the issue of choosing the function ^ to satisfy condition b) of Theorem 2.8.14. There are two ways of doing this as is shown by the following lemma. We denote by rgs (; x) the gradient of gs (; x) with respect to if it is well de ned.

© 2001 by Chapman & Hall/CRC

218

Maxingales

Lemma 2.8.20. Let gs(; x) be dierentiable in .

Let x^

absolutely continuous. 1. If a function ^ = (^ (s; x); s 2 R+ ; x 2 C ) is such that x^_ s = rgs(^(s; x); x)

2C

be

for x -almost all x and almost all s 2 R+ , then condition b) of Theorem 2.8.14 is satis ed. 2. If a function ^ = (^ (s); s 2 R+ ) is such that the equation x_ s = rgs(^(s); x); x0 = x;

has the only solution x = x^ , then condition b) of Theorem 2.8.14 is satis ed. Proof. By the requirements on x~ in part b) of Theorem 2.8.14 and the necessary condition for attaining supremum x~_ s = rgs (^ (s; x~ ); x~ ). Hence, if ^ is from part 1 of the lemma, then x^_ s = x~_ s so that since x^ 0 = x~ 0 = x we conclude that x^ = x~ . If ^ is from part 2 of the lemma, then x^ = x~ by de nition.

We are able to obtain somewhat general uniqueness results only for the case where ^ is chosen as in part 1 of Lemma 2.8.20, so we concentrate on that case. However, we give examples that show an application of the approach in part 2. We next state a uniqueness result for a \diusion" maxingale problems. Theorem 2.8.21. Let the canonical process X be a semimaxingale under with local characteristics (b; c; 0; 0) starting at x. Let the following conditions hold: 1. the functions bs (x) and cs (x) are continuous in x 2 C , 2. for every t 2 R+ and compact K C Zt

0

sup jbs (x)j2 ds < 1; sup sup kcs (x)k < 1; st x2K

x2K

3. for every t 2 R+ and compact K C

inf inf inf cs (x) > 0:

st x2K 2Rd : jj=1

© 2001 by Chapman & Hall/CRC

219

Maxingale problems

Then = x . Proof. We rst note that g^t (; x) = ct (x)=2. Let x^ 2 C be such that x (^x) > 0. We prove that (^x) = x (^x). We apply Theorem 2.8.14. By Lemma 2.8.16 conditions I and II are satis ed. Since x^ is absolutely continuous and ct (x) is positive de nite for all x 2 C , we can de ne

^ (t; x) = ct (x)

1

(x^_ t bt(x)):

(2.8.26)

Since x^_ t = rgt (^ (t; x); x) for all x, by Lemma 2.8.20 ^ satis es the condition of part b) of Theorem 2.8.14, so by part a) of The^ orem 2.8.14 and Theorem 2.8.19 it suÆces to check that ^ 2 . Firstly, we note that for x 2 K , where K is compact, Zt

0

sup jx^_ s bs(x)j2 ds < 1:

(2.8.27)

x2K

Indeed, we have Zt

0

sup jx^_ s

bs (x)j2 ds 2

x2K

Zt

jx^_ s bs(^x)j2 ds

0 Zt

+2

0

sup jbs (^x) bs (x)j2 ds: (2.8.28)

x2K

Since by Lemma 2.7.12

I(^x) = 12

Z1

0

(x^_ s

bs (^x)) cs (^x) 1 (x^_ s bs (^x)) ds

1 sup kcs (^x)k 2 st

1

Zt

0

jx^_ s bs(^x)j2 ds;

we conclude that the rst term on the right of (2.8.28) is nite. The second one is nite by hypotheses. Inequality (2.8.27) is proved.

© 2001 by Chapman & Hall/CRC

220

Maxingales

Next, we have by (2.8.26) for x 2 K (a) and 2 R+ , denoting c = inf st inf x2K inf 2Rd : cs (x), that jj=1 Zt

0

g^s (^ (s; x); x) 1(j^ (s; x)j > A)ds

2 = 2

Zt

0

(x^_ s

bs (x)) cs (x) 1 (x^_ s bs (x)) 1(j^ (s; x)j > A)ds

2 c 2

1

Zt

0

jx^_ s bs(x)j2 1(j^ (s; x)j > A)ds

so that by (2.8.27) and absolute continuity of the Lebesgue integral the required limit lim sup

A!1 x2K

Zt

0

g^s (^ (s; x); x) 1(j^ (s; x)j > A)ds = 0

would follow by lim sup

A!1 x2K

Zt

0

1(j^ (s; x)j > A)ds = 0:

The latter limit follows since by (2.8.26), (2.8.27) and hypotheses sup

Zt

x2K 0

j^(s; x)j2 ds c

2

Zt

0

sup jx^_ s bs(x)j2 ds < 1:

x2K

As a consequence of this result, Theorem 2.6.24 and Theorem 2.6.30, we have the following existence and uniqueness result for idempotent Ito equations.

Theorem 2.8.22. Let (bds(x); s 2 R+ ; dxd2 C ) and (s(x); s 2 R + ; x 2 C ) be respective R -valued and R -valued C-progressively measurable idempotent processes. Let the following conditions hold:

© 2001 by Chapman & Hall/CRC

221

Maxingale problems

1. bs (x) and s (x) are continuous in x 2 C ,

2. linear growth: for every t 2 R+ and x 2 C Zt

0

jbs(x)j2 ds +sup ks(x)k 2 < 1; 1 + xs 2

st

1 + xs 2

3. for every t 2 R+ and compact K C

inf inf inf s (x)s (x)T > 0:

st x2K 2Rd : jj=1

Then the equation

X_ t = bt (X )+ t (X )W_ t ; X0 = x; has a unique Luzin solution. The idempotent distribution of X is given by 1 Z1 X (x_ t bt (x)) t (x)t (x)T 1 (x_ t bt (x)) dt (x) = exp 2 0

if x is absolutely continuous and

x0 = x, and X (x) = 0 otherwise.

Proof. By Theorem 2.6.24 the equation has a Luzin solution X . By Theorem 2.6.30 and Theorem 2.8.21 uniqueness holds. The form of X is given in Lemma 2.6.19.

Remark 2.8.23. One can weaken the conditions of Theorem 2.8.22 by requiring that the non-degeneracy condition 3 in the hypotheses hold for compacts Kx (a), where a 2 (0; 1], and x in conditions 1 and 2 is such that x(x) > 0. If the function gs (; x) is more general than in Theorem 2.8.21, we cannot apply Theorem 2.8.14 to all x 2 C such that x (x) > 0, so we have to introduce additional regularity conditions, e.g., require that the derivative x_ s be locally bounded. As for \nonregular" x, a way to establish for them the equality (x) = x (x) is to see to it that each such x can properly be approximated by \regular" x. Below we assume without further mentioning that is a solution to the maxingale problem (x; G).

© 2001 by Chapman & Hall/CRC

222

Maxingales

De nition 2.8.24. Let D C : We de ne the x{closure of D as the set of all x 2 C such that x (x) > 0 for which there exists a sequence xk 2 D such that xk ! x and x;t (xk ) ! x;t (x) as k ! 1 for all t 2 R+ . We say that D is x-dense in C if its x{closure coincides with the set fx 2 C : x(x) > 0g. Remark 2.8.25. If conditions I and II hold, then by (2.8.4), (2.8.5) and (2.8.6) x;t (x) is an upper semi-continuous function of x. In this case in the above de nition of the x {closure of D it is suÆcient to require that xk 2 D be such that xk ! x and lim inf k!1 x;t (xk ) x;t (x); t 2 R+ . Lemma 2.8.26. If1 (pt 1 Æ pt x) = x;t(x); t 2 R+ ; for all x 2 D C , then (pt Æ pt x) = x;t (x); t 2 R+ ; for all x from the x{closure of D. If, in addition, the set D is x{dense in C , then uniqueness holds for the maxingale problem (x; G) with = x . Proof. Let x belong to the x {closure of D. Let xk 2 D be such that xk ! x and x;t (xk ) ! x;t (x); t 2 R+ . Since pt xk ! pt x, it follows that, for arbitrary " > 0, pt 1 Æ pt xk pt 1 B" (pt x) for all k large enough, where B" (pt x) is the closed "-ball about pt x. Sinceby the -smoothness property of deviability lim"!0 pt 1 B"(pt x) = (pt 1 Æ pt x); we conclude that lim supk!1 (pt 1 Æ pt xk ) (pt 1 Æ pt x): Since also (pt 1 Æ pt xk ) = x;t (xk ) ! x;t (x) as k ! 1, we have that x;t (x) (pt 1 Æ pt x): On the other hand, x;t (x) x(pt 1 Æpt x) (pt 1 Æpt x1) by the de nitions of x;t and x, and (2.8.11), so x;t (x) = (pt Æ pt x); t 2 R+ . Finally, if D is x -dense in C , then by the part just proved, the -smoothness property of deviability and (2.8.12) (x) = limt!1 (pt 1 Æpt x) = limt!1 x;t (x) = x (x) when x (x) > 0. If x(x) = 0, then (x) = 0 = x(x) by (2.8.11). Theorem 2.8.27. Let gs(; x) meet conditions I and II and be differentiable in . Let there exist a family fGm ; m 2 N g of subsets of R d and an R d -valued function ((s; x; y ); s 2 R + ; x 2 C ; y 2 Rd ), which is B[0; t] Ct B(Rd )=B(Rd ) {measurable when restricted to [0; t] C Rd for t 2 R+ , continuous in x, bounded on the sets [0; t] K Gm , where t 2 R+ , K C and is compact, and m 2 N , and such that

y = rgs ((s; x; y); x)

© 2001 by Chapman & Hall/CRC

(2.8.29)

Maxingale problems

223

for y 2 [1 m=1 Gm , (almost all) s 2 R+ and x {almost all x. Let 1 [ D= fx 2 C : x is absolutely continuous, x0 = x, and m=1

x_ s 2 Gm ; s 2 R+ g: (2.8.30) Then (pt 1 Æ pt x) = x;t (x); t 2 R+ ; whenever x is in the x { closure of D. If the set D is, in addition, x {dense in C , then uniqueness holds for the maxingale problem (x; G) with = x . Proof. For x^ 2 D we de ne ^ (s; x) = (s; x; x^_ s ): The function ^ (s; x) is C-progressively measurable, bounded on the sets [0; T ] K for T 2 R+ and continuous in x. Therefore (^ (s; x); s 2 R+ ; x 2 C ) 2

^ so that by Theorem 2.8.19 Z (^ ) is a strictly Luzin-continuous local exponential maxingale under and admits a localising sequence of C-stopping times. Since also by (2.8.29) x^_ s = rgs (^ (s; x); x) for (almost all) s 2 R+ and x -almost all x, Theorem 2.8.14 and Lemma 2.8.20 imply that (pt 1 Æ pt x^ ) = x;t (^x); t 2 R+ , and an application of Lemma 2.8.26 ends the proof. We give an application. Theorem 2.8.28. Let d = 1. Let the canonical process X on C be a semimaxingale starting at x 2 R+ under with local characteristics (bq; 0; ; 0), where s( ; x) = 1(qs (x) 2 )bs (x); and bs (x) and qs(x) are R+ {valued functions, which are C-progressively measurable in (s; x) and continuous in x 2 C . Let for every t 2 R+ and compact KC inf inf b (x) > 0; sup sup bs (x) < 1; st x2K s st x2K inf inf q (x) > 0; sup sup qs(x) < 1: st x2K s st x2K Then = x . Proof. We apply Theorem 2.8.27. Conditions I and II are met by Lemma 2.8.16. The function gs (; x) = eqs (x) 1 bs (x) is differentiable in . We take in the hypotheses of Theorem 2.8.27 Gm = [1=m; m] and 8 1 y < ln ; y > 0; (s; x; y) = qs(x) bs (x)qs (x) : 0; y 0:

© 2001 by Chapman & Hall/CRC

224

Maxingales

We check that D is x {dense in C . Let x 2 C be such that x (x) > 0. By Lemma 2.7.12 x is absolutely continuous; also

I(x) =

Z1

0

sup(x_ t (eqt(x) 1)bt (x))dt 2R

so that x_ t 0 a.e. We de ne xk by xk0 = x and x_ ks = (x_ s 1(x_ s k)) _ k1 . Convergence xk ! x is obvious. We prove that for t 2 R+ lim

k!1

Zt

0

sup(x_ ks 2R =

(eqs (x Zt

k)

1)bs (xk ))ds

sup(x_ s 2R

0

(eqs (x)

1)bs (x))ds: (2.8.31)

We have Zt

0

sup(x_ ks 2R

(eqs (x

k)

1)bs (xk )) ds

Zt

x_ s ln x_ s qs(xk ) bs(xk )qs(xk ) 0 1 x_ s k; x_ s k1 ds =

+

Zt

0

1 1 ln k k kqs(x ) kbs (x )qs (xk )

x_ s + bs(xk ) qs(xk ) 1 k) + b ( x kqs (xk ) s

1 x_ s < k1 + 1(x_ s > k)ds:

Since 0t x_ s ln x_ s 1(x_ s > 1)ds < 1 by the fact that I(x) < 1 and hypotheses, bs (xk ) ! bs (x) and qs (xk ) ! qs(x) as k ! 1, Lebesgue's dominated convergence theorem implies that the right-hand side conR

© 2001 by Chapman & Hall/CRC

225

Maxingale problems

verges to Zt

0

x_ s ln x_ s qs (x) bs(x)qs (x)

x_ s + bs(x) 1(x_ s > 0) ds qs(x) +

Zt

0

bs(x) 1(x_ s = 0) ds

ending the proof of (2.8.31). We now consider a version of the above result, which will be used in an application to the analysis of a many-server queue in part II. This result also shows the use of the other method of choosing the function ^ .

Theorem 2.8.29. Let d = 1. Let deviability on C

be such that the canonical process X is a semimaxingale on (C ; ) starting at x 2 R+ with local characteristics (b; 0; ; 0), where

s( ; x) = 1(1 2 )vs (x) + 1( 1 2 )us (x)(xs ^ ms (x)); bs(x) = vs (x) us (x)(xs ^ ms (x));

and vs (x), us (x) and ms(x) are R+ {valued functions, which are Cprogressively measurable in (s; x) and locally Lipshitz-continuous in x 2 C . Let also for every t 2 R+ and compact K C

inf inf vs (x) > 0; sup sup vs (x) < 1; st x2K inf inf u (x) > 0; sup sup us (x) < 1; st x2K s st x2K inf inf m (x) > 0: st x2K s st x2K

Then = x .

Proof. We rst consider the case where x > 0. Let x^ be such that x(^x) > 0, sups2[0;t]jx^_ sj < 1 and inf s2[0;t] x^ s > 0 for t 2 R+ . We de ne a function ^ (s) by the equality

x^_ s = e^(s) vs(^x)

e

© 2001 by Chapman & Hall/CRC

^ (s) u

x x^ s ^ms(^x):

s (^ )

(2.8.32)

226

Maxingales

The function ^ (s) is well de ned, satis es the conditions of part 2 of Lemma 2.8.20 and, being locally bounded and Lebesgue measurable, ^ also by Lemma 2.8.16 conditions I and II are met; is an element of ; so Z (^ ) is a strictly Luzin exponential maxingale by Theorem 2.8.19 and admits a localising sequence of strictly Luzin C-stopping times. Therefore, by Theorem 2.8.14 (^x) = x (^x) and (pt 1 Æ pt x^ ) = x;t(^x); t 2 R+ . For more general functions x^ such that x (^x) > 0 we apply Lemma 2.8.26. Let us consider the instance where inf s2[0;t] x^ s > 0 for t 2 R+ . We de ne Zs

x^ ks = x + x^_ p 1 jx^_ pj k dp:

(2.8.33)

0

Then sups2[0;t] jx^ ks x^ sj ! 0 as k ! 1; in particular, k lim inf k!1 inf s2[0;t] x^ s > 0 so by the part just proved (pt 1 Æ pt x^ k ) = x;t(^xk ); t 2 R+ . Since for absolutely continuous x

It(x) =

Zt

0

sup x_ s (e 1)vs (x) (e 2R

1)us (x)

xs^ms(x) ds;

to prove that (^x) = x (^x) it is suÆcient to show by Lemma 2.8.26 and Remark 2.8.25 that for all t 2 R+ lim sup k!1 (e

Zt

0

Zt

sup x^_ ks 2R

(e

1)vs (^xk )

1)us (^xk ) x^ ks ^ ms (^xk ) ds

sup x^_ s (e 1)vs (^x) (e 0

2R

1)us (^x) x^ s ^ms(^x) ds:

(2.8.34)

We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R

© 2001 by Chapman & Hall/CRC

1)us (^xk ) x^ ks ^ms (^xk ) ds

227

Maxingale problems

Zt

sup x^_ s (e 1)vs(^xk ) (e +

0 Zt 0

2R

1)vs (^xk ) (e

sup (e 2R

1)us (^xk ) x^ ks ^ms(^xk ) ds

1)us (^xk ) x^ ks ^ ms(^xk )

1 jx^_ sj > k ds:

(2.8.35)

The second integral on the right converges to 0 as k ! 1. We work with the rst integral. Let C1 > 0 and C2 > 0 be respective upper and lower bounds over large values of k and s 2 [0; t] for the k k x^ s ^ ms(^x ) , C3 > 0 and C4 > 0 { upper and lower bounds for the vs (^xk ), and C5 > 0 and C6 > 0 { upper and lower bounds for the us (^xk ). Let k (s) be the points where the supremums in the rst integral on the right-hand side of (2.8.35) are attained so that

x^_ s = ek (s) vs(^xk )

e

k (s) u

xk ) x^ ks ^ms(^xk ):

s (^

(2.8.36)

k (s) u (^ k ^k ^ Since for k (s) positive ek (s) vs (^xk ) e s x ) x s k ms (^xk ) C4 e (s) C5 C1 and for k (s) negative ek (s) vs (^xk ) e k (s) us (^xk ) x^ ks ^ ms(^xk ) C3 e k (s) C6 C2 ; we conclude that k (s)

e

jx^ sj +C C5 C1 _1; e _

4

jx^Cs j +C C3 _1; s 2 [0; t]: _

k (s)

6 2

We thus write for the rst integral on the right of (2.8.35) Zt

0

=

sup x^_ s (e 1)vs (^xk ) (e 2R Zt

0

k (s)x^_ s (e

© 2001 by Chapman & Hall/CRC

(2.8.37)

1)us (^xk ) x^ ks ^ ms(^xk ) ds

1)vs (^xk )

k (s)

(e

k (s)

1)us (^xk ) x^ ks ^ ms (^xk ) ds

228

Maxingales

Zt

0 Zt

+

0 Zt

+

k k (s)x^_ s (e (s) 1)vs (^x) (e

je

k (s)

1jjus (^xk ) x^ ks ^ ms (^xk )

sup x^_ s (e 1)vs(^x) (e + +

0 Zt 0

1)us (^x) x^ s ^ms(^x) ds

jek (s) 1jjvs (^xk ) vs(^x)j ds

0 Zt

0 Zt

k (s)

2R

us (^x) x^ s ^ ms(^x) j ds

1)us (^x) x^ s ^ ms (^x) ds

jek (s) 1jjvs (^xk ) vs(^x)j ds

je

k (s)

1jjus (^xk ) x^ ks ^ ms (^xk )

us (^x) x^ s ^ ms (^x) j ds:

By (2.8.37) and the facts that x^ is absolutely continuous, us (x), vs (x) and ms(x) are continuous in x and bounded, and the x^ k converge uniformly on [0; t] to x^ as k ! 1, the latter two integrals converge to 0 as k ! 1, so (2.8.34) follows. Let us now assume that x^ is an arbitrary function such that x(^x) > 0. Clearly, x^ is R+ -valued, absolutely continuous, and x^ 0 = x > 0. Let x^ ks = x + R0s x^_ u 1(^xu k 1=k) du. Since x^_ s = 0 on the set fx^ s = 0g (a.e.), we have that x^ ! x^ uniformly on bounded intervals. Since x^ ks = x^ s _ (1=k) for k large, by the part proved (pt 1 Æ pt x^ k ) = x;t (^xk ); t 2 R+ , so by Lemma 2.8.26 in order to prove that (^x) = x (^x) it is suÆcient to show that (2.8.34) holds. We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R

© 2001 by Chapman & Hall/CRC

1)us (^xk ) x^ ks ^ms (^xk ) ds

229

Maxingale problems

=

Zt

0

sup x^_ s (e 1)vs (^xk ) (e 2R

1)us (^xk ) x^ s ^ ms(^xk )

1(^xs 1=k) ds +

Zt

0

sup (e 2R

1)vs (^xk ) (e

1)us (^xk ) x^ ks ^ ms(^xk )

1(^xs < 1=k) ds:

(2.8.38) Routine calculations show that R t the second integral on the right-hand side of (2.8.38) converges to 0 vs (^x) 1(^xs = 0) ds as k ! 1. De ning k (s) on the set fs : x^ s 1=kg by x^_ s = ek (s) vs(^xk ) e k (s) us(^xk ) x^ s ^ms(^xk ); we have estimates analogous to (2.8.37) with the right inequality replaced by _ x^ s ^ms(^xk )e k (s) jx^ sjC+6 C3 _ x^ s ^ms(^xk ) so that Zt

0

sup x^_ s 2R

1)vs (^xk ) (e

(e

1)us (^xk ) x^ s ^ ms (^xk )

1(^xs 1=k) ds Zt

sup x^_ s (e 1)vs (^x) (e 0

2R

1)us (^x) x^ s ^ ms (^x)

1(^xs > 0) ds + +

Zt

0 Zt 0

jek (s) 1jjvs (^xk ) vs(^x)j ds

x^ s ^ ms(^xk )je +

Zt

0

© 2001 by Chapman & Hall/CRC

k (s)

je

k (s)

1jjus (^xk ) us (^x)j ds 1jjms (^xk ) ms (^x)jus (^x) ds:

230

Maxingales

The required convergence follows by the fact that the last three integrals on the right-hand side converge to 0 by the same argument as above. We now consider the case x = 0. Let x^ be such that x (^x) > 0, sups2[0;t]jx^_ s j < 1, inf s2[0;] x^_ s > 0, and inf s2[;t] x^ s > 0 for t 2 R+ and some > 0. We again de ne ^ (s) by (2.8.32). It is evidently bounded on [; t]. Next, since by (2.8.32) x^_ s e^(s) vs (^x), we have that e^(s) x^_ s =C30 ; hence, x^_ s = e^ (s) vs(^x) e ^(s) us (^x) x^ s ^ ms (^x) e^ (s) C40 C30 C50 x^ s ^ ms(^x) =x^_ s ; where C30 ; C40 and C50 have a similar meaning as above. We thus conclude that there exist A1 , A2 and A3 , which depend on x^ s ; s 2 [0; ]; but do not explicitly depend on x^_ s , such that for s 2 [0; ] A A1 x^_ s e^ (s) A2 x^_ s + _ 3 : (2.8.39) x^ s Since x^_ s is bounded both from below and above on [0; ], so is ^ (s). Hence, by Theorem 2.8.14 (^x) = x (^x) and (pt 1 Æ pt x^ ) = x;t(^x); t 2 R+ . Next, let x^ be such that x (^x) > 0, inf s2[0;] x^_ s > 0, and inf s2[;t] x^ s > 0 for t 2 R+ and some > 0. We de ne x^ k in analogy with the case x > 0 by (2.8.33). Then following the same line of reasoning we need bounds on k (s) de ned by (2.8.36) in order to prove that Zt

jek (s) 1jjvs (^xk ) vs(^x)j ds ! 0;

Zt

Z

0

je

k (s)

1jjus (^xk ) x^ ks ^ ms(^xk )

(2.8.40)

us(^x) x^ s ^ ms(^x) j ds ! 0;

(2.8.41)

jek (s) 1jjvs (^xk ) vs(^x)j ds ! 0;

(2.8.42)

Z

je

k (s)

0

© 2001 by Chapman & Hall/CRC

1jjus (^xk ) x^ ks ^ ms(^xk )

us (^x) x^ s ^ ms (^x) j ds ! 0:

(2.8.43)

231

Maxingale problems

On the interval [; t] bounds (2.8.37) apply yielding convergences (2.8.40) and (2.8.41). Bounds on [0; ] that imply limits (2.8.42) and (2.8.43) are given by (2.8.39) with suitable A1 , A2 and A3 . Next, let x^ be such that x (^x) > 0 and not identically equal to 0. Let k = inf fs 2 R+ : x^ s = 1=kg. We de ne x^ k by x^ ks = sx^ k =k for s 2 [0; k ] and x^ ks = x^ s _ (1=k) for s k . Then (pt 1 Æ pt x^ k ) = x;t(^xk ); t 2 R+ , so we need to prove that (2.8.34) holds. We have Zt

0

sup x^_ ks (e 1)vs (^xk ) (e 2R Zk

sup x^_ ks (e 1)vs(^xk ) (e +

0 Zt 0

1)us (^xk ) x^ ks ^ ms(^xk ) ds

2R

sup x^_ s (e 1)vs (^xk ) (e 2R

1)us (^xk ) x^ ks ^ms(^xk ) ds

1)us (^xk ) x^ s ^ ms(^xk )

1(^xs 1=k) ds +

Zt

0

sup (e 2R

1)vs (^xk ) (e

1)us (^xk ) x^ ks ^ ms(^xk )

1(^xs < 1=k) ds:

(2.8.44)

By an argument similar to the one used for deriving an asymptotic bound for the right-hand side of (2.8.38) the limit superior as k ! 1 of the sum of the latter two integrals is not greater than It (^x). Let k (s) be the points, where the supremums in the rst integral on the right of (2.8.44) are attained. Then the integrand takes the form k k k (s)x^_ ks (e (s) 1)vs (^xk ) (e (s) 1)us(^xk ) x^ ks ^ ms(^xk ) k (s)x^_ ks + vs(^xk ) + us(^xk ) x^ ks ^ ms(^xk ) :

Since the derivatives x^_ ks equal x^ k =k and k ! 0 as k ! 1, the estimates (2.8.39) applied to x^ ks show that the rst integral on the right of (2.8.44) converges to 0 as k ! 1 provided limk!1 x^ Rk ln(^xk =k ) = 0: The latter limit follows since x^ k ln(^xk =k ) 0 k jx^_ s lnjx^_ s jj ds Rt and 0 jx^_ s lnjx^_ s jj ds < 1 since It (^x) < 1.

© 2001 by Chapman & Hall/CRC

232

Maxingales

Finally, if x^ s = 0 for all s 2 R+ , then we let x^ ks = s=k for s 2 [0; 1=k] and x^ ks = 1=k for s 1=k.

Remark 2.8.30. This theorem illustrates the general feature that

if the function ^ is chosen as in part 2 of Lemma 2.8.20, then one needs to impose Lipshitz continuity conditions on the coeÆcients.

Our purpose now is to state uniqueness results for more general functions gs (; x). In the next lemma, given a closed convex set F Rm , we denote as projF the projection of 2 Rm onto F ; for a closed convex cone N 2 Rd , we denote as a N the aÆne hull of N ; N ? = f 2 Rd : y 0 for all y 2 N g denotes the polar cone of N . As above, ri N denotes the relative interior of N .

Lemma 2.8.31. Let conditions I and II hold, and gs(; x) be dif-d ferentiable in . Let there exist a closed convex cone N R such that gs (; x) is strictly convex in 2 a N , gs (; x) gs (proja N ; x) ; 2 Rd ; s 2 R+ ; x 2 C ; and the following holds: 1. for every t 2 R+ and compact K C

g (; x) = 1; lim inf inf s 2 a N : st x2K jprojN j jprojN j!1

2. for every t 2 R+ , compact K C and A 2 R+

inf

inf inf gs (; x) >

2 a N : st x2K jprojN jA

1;

3. for every t 2 R+ , A 2 R+ , x 2 C , and sequence xk ! x

lim

k!1

Zt

0

sup

2 a N : jprojN jA

jgs (; xk ) gs(; x)j ds = 0;

4. if N ? 6= 0, then for every t 2 R+ and x 2 C

lim sup sup 2N ? : st jj!1

© 2001 by Chapman & Hall/CRC

gs (; x) jj

0:

233

Maxingale problems

Then (pt 1 Æ pt x) = x;t(x); t 2 R+ ; for every x such that x0 = x, x_ s 2 N (a.e.) and sups2R+ jx_ s j < 1. Also, if x(x) > 0, then x_ s 2 N (a.e.) Proof. We apply Theorem 2.8.27. Denote as N 0 the polar cone of N relative to a N , i.e., N 0 = N ? \ a N and let for m 2 N

Nm = Gm =

fy 2 a N : y m1 jjjyj for all 2 N 0g; fy 2 Nm : jyj mg:

We next de ne (s; x; y) in the statement of Theorem 2.8.27. If y 2= ri N , we set (s; x; y) = 0. Let y 2 ri N (the latter set is nonempty since N is convex, von Leichtweiss [131]). Since ri N = [1 m=1 Gm , we have that y 2 Gm for some m. We now prove that for every compact K C and t 2 R+ there exists C0 > 0 such that the inequality y gs (; x) 0 holds for x 2 K , s t and 2 a N such that jj > C0 . If jyj = 0, then a N = N , so = projN and in view of condition 1 y gs (; x) is negative for x 2 K and s t if jj is large enough. Let us assume now that jyj > 0. We rst consider the case when 1 2 a N is such that y jjjyj. Then for s t 2m 1 y gs (; x) jjjyj inf uinft gu(; x) 2m 2 a N

which by conditions 1 and 2 of the lemma is negative for all x 2 K and s t if jj is large enough. Now, let 2 a N be such that 1 y jjjyj. Then 2m

jprojN j 1 : jj 2(m + 1)

(2.8.45)

To see this, write = 1 + 2 , where 1 = projN 0 and 2 = projN . 1 Since y 2 Gm Nm and 1 2 N 0 , it follows that 1 y j jjyj; m 1 1 on the other hand, y jjjyj, so 2m j2 jjyj 2 y = y 1 y jmyj j1 j j2j jmyj j2j j2 j :

© 2001 by Chapman & Hall/CRC

234

Maxingales

1 jj Hence, j2 j 1 + 2m (recall that jyj > 0) which is equivalent m to (2.8.45) by the de nition of 2 . Inequality (2.8.45) and condition 1 imply that, given L > 0, we have, if jj is large enough, that for x 2 K and s t

y gs (; x) jjjyj

L jj; 2(m + 1)

which is negative if L has been chosen large enough. Thus, in all the cases we have that y gs (; x) 0 for all x 2 K and s t if 2 a N is such that jj is large enough. The claim is proved. We thus conclude, since y = proja N y and by hypotheses gs (; x) gs (proja N ; x); that for x 2 K and s t sup ( y

2Rd

gs (; x)) = sup ( y gs (; x)) 2 a N = sup ( y 2 a N : jjC0

gs (; x)); s t: (2.8.46)

Since the function y gs (; x) is strictly concave in 2 a N by the hypotheses, it attains supremum on f 2 a N : jj C0 g at a unique point, which we take as (s; x; y). Thus, we have de ned (s; x; y) for all y 2 Rd . We check that it satis es the conditions of Theorem 2.8.27. By de nition (s; x; y)y gs ((s; x; y); x) = sup (y gs (; x)) 2Rd for s 2 R+ ; y 2 [1 m=1 Gm ; x 2 C , which implies (2.8.29). Equality (2.8.46) shows that (s; x; y) is bounded on the sets [0; t] K Gm , where t 2 R+ , K is compact, and m 2 N . Also it is B[0; t] Ct

B(Rd )=B(Rd ) {measurable when restricted to [0; t]C R d for t 2 R+ since gs (; x) is C-progressively measurable in (s; x) and continuous in , and for 2 B(Rd )

f(s; x; y) : (s; x; y) 2 g = f(s; x; y) : sup (y gs (; x)) = sup (y gs (; x))g; d 2

© 2001 by Chapman & Hall/CRC

\ a N

2R

235

Maxingale problems

if

63 0, and f(s; x; y) : (s; x; y) 2 g = f(s; x; y) : sup ( y gs (; x)) = sup ( y gs (; x))g d 2

\ a N

[

2R

f(s; x; y) : y 2= ri N g;

if 3 0, where sup; = 1. (Note that by continuity of gs (; x) in the supremums may be taken over the rationals.) Finally, (s; x; y) is continuous in x. Indeed, let xk ! x as k ! 1. If y 2= ri N , then (s; xk ; y) = (s; x; y) = 0. Let y 2 ri N . Hence, y 2 Gm for some m and, since the set fxk ; k 2 N g is relatively compact, the set f(s; xk ; y); k 2 N g is bounded; so there exist 0 2 a N and subsequence k0 such that (s; xk0 ; y) ! 0 as k0 ! 1. Since (s; xk ; y) y gs ((s; xk ; y); xk ) y gs (; xk ) for 2 Rd and gs (; x) is continuous in (; x), we conclude that 0 y gs (0 ; x) y gs (; x) so that 0 y gs (0 ; x) = sup2 a N (y gs (; x)): Since the point where the supremum is attained is unique, 0 = (s; x; y) proving that (s; xk ; y) ! (s; x; y) as k ! 1. Thus, existence of (s; x; y) in the statement of Theorem 2.8.27 is proved. Let x^ be such that

x^ 0 = x; x^_ s 2 N;

A = sup jx^_ s j < 1: s2R+

(2.8.47)

We prove that (pt 1Æpt x^ ) = x;t (^x); t 2 R+ . Let D be de ned as in Theorem 2.8.27. According to Theorem 2.8.27, the required equality will follow if we nd a sequence xk 2 D, which converges to x^ as k ! 1, and is such that lim

k!1

x;t(xk ) = x;t(^x); t 2 R+ :

(2.8.48)

Since ri N is nonempty, there exist y^ 2 N and r 2 (0; 1) such that jy^j = 1 and y^ rjj for all 2 N 0 . We observe that if y 2 N and jyj A, then, given k 2 N ; there exist k > 0 such that k ! 0 as k ! 1 and y + k y^ 2 Gk if k k0 = b(A + 1)=rc + 1. Indeed, since y^ rjj and y 0 if 2 N 0 , jy^j = 1 and jyj A, we have, for 2 N 0 and 0 < k < 1, that (y + k y^) k rjj k rjjjy + k y^j=(1+ A); so if k = (1+ A)=(rk), then y + k y^ 2 Gk for k k0 .

© 2001 by Chapman & Hall/CRC

236

Maxingales

We de ne next xks = x^ s + k y^s; s 2 R+ : Since x^_ s 2 N; we have by (2.8.47) that x_ ks 2 Gk , so xk 2 D by (2.8.30). By Lemma 2.7.12, (2.8.6), and Remark 2.8.25, for (2.8.48) we need to prove that lim sup k!1

Zt

0

x

sup ( _ ks 2Rd

gs(; x

k ))ds

Zt

0

sup (x^_ s gs(; x^ ))ds: 2Rd

Since for y 2 ri N the left equality in (2.8.46) holds and x_ ks 2 Gk ri N , it suÆces to prove that lim sup k!1

Zt

0

sup ( x_ ks 2 a N

Zt

0

gs (; xk )) ds sup ( x^_ s 2 a N

gs (; x^ )) ds: (2.8.49)

We denote ks = (s; xk ; x_ ks ) and observe that ks 2 a N . Since x_ ks 2 N , it follows that ks x_ ks projN ks x_ ks , so, by the de nition of (s; x; y) 0 sup ( x_ ks 2 a N

gs (; xk )) jprojN ks jjx_ ks j gs (ks ; xk );

which implies by the facts that x_ ks is bounded (see (2.8.47)), ks a N and condition 1 holds that

B = sup sup jprojN ks j < 1: kk0 st

2

(2.8.50)

Next, by the de nitions of ks and xks sup ( x_ ks gs (; xk )) 2 a N = (ks x^_ s gs (ks ; x^ )) + k ks y^ (gs (ks ; xk ) gs (ks ; x^ )): (2.8.51) Since jy^j = 1 and y^ 2 N , we have that ks y^ jprojN ks j B: Also, by (2.8.50) for k k0 and s t

jgs (ks ; xk ) gs(ks ; x^ )j

© 2001 by Chapman & Hall/CRC

sup jgs (; xk ) gs (; x^ )j: 2 a N : jprojN jB

237

Maxingale problems

Therefore, by (2.8.51) and (2.8.50) for k k0 and s t sup ( x_ ks

2 a N

gs (; xk )) sup ( x^_ s gs (; x^ )) + Bk 2 a N

+

sup jgs (; xk ) 2 a N : jprojN jB

gs (; x^ )j

which implies (2.8.49) by the convergence k ! 0 and condition 3, nishing the proof of (2.8.48). We consider now the second assertion of the lemma. There is something to prove only if N 6= Rd . If y 2 Rd nN , then y "jjjyj for some 2 N ? and " > 0, so by condition 4 of the lemma, for every x 2 C , lim sup 2N ? : jj!1

y gs (; x) jj

"jyj;

and therefore sup2Rd (y gs (; x)) = 1. By Lemma 2.7.12, (2.7.6) and (2.7.8) this implies that x_ s 2 N if x (x) > 0. We now apply Theorem 2.8.27 and Lemma 2.8.31 to a proof of the following uniqueness result.

Theorem 2.8.32. Let conditions I anddII hold, and gs(; x) be differentiable and strictly convex in lowing hold:

2R

. Let, in addition, the fol-

1. for every t 2 R+ and compact K C

lim inf inf jj!1 st x2K

gs (; x) jj = 1;

2. for every t 2 R+ , compact K C and A 2 R+

inf inf inf g (; x) > jjA st x2K s

1;

x 2 C there exists l > 1 such that g (l; x) lim inf inf s > 1; jj!1 st lgs (; x)

3. for every t 2 R+ and

© 2001 by Chapman & Hall/CRC

238

Maxingales

4. for every t 2 R+ , compact K C and x 2 K there exist > 0 and > 0 such that

lim inf inf jj!1 st

gs (; x0 ) inf > 0: x0 2K : 0 gs (; x) sup jxr xr j rt

Then uniqueness holds for the maxingale problem (x; G) with = x. Proof. We rst observe that conditions 1 and 2 of the theorem imply that for every t 2 R+ and compact K C

inf inf inf gs (; x) >

2Rd st x2K

1:

(2.8.52)

It is easy to see that all the conditions of Lemma 2.8.31 hold with N = Rd . (In particular, condition 3 follows by conditions I and II.) Hence, by Lemma 2.8.31 (pt 1 Æpt x) = x;t (x); t 2 R+ ; if, in addition, x0 = x and sups2R+ jx_ s j < 1. The proof of Lemma 2.8.31 also shows that there exists function (s; x; y) satisfying the conditions of Theorem 2.8.27 with Gm = fy 2 Rd : jyj mg so that by Theorem 2.8.27 it suÆces to prove that the set D = fx : sups2R+ jx_ s j < 1; x0 = xg is x{dense in C . Assuming with no loss of generality that x = 0, we x x^ 2 C such that 0 (^x) > 0 and look for xk 2 D such that xk ! x^ as k ! 1 and lim It (xk ) = It (^x); t 2 R+ :

k!1

(2.8.53)

We de ne

x

k s

=

Zs

0

x^_ u 1(jx^_ u j k) du:

(2.8.54)

R The convergence xk ! x^ follows by the fact that 0t jx^_ u j du < 1 and Lebesgue's dominated convergence theorem. By Remark 2.8.25 for (2.8.53) it suÆces to show that

lim sup It (xk ) It (^x); t 2 R+ ; k!1

© 2001 by Chapman & Hall/CRC

(2.8.55)

239

Maxingale problems

where by Lemma 2.7.12 and (2.8.54)

It(^x)

=

It (xk )

=

Zt

0 Zt 0

sup ( x^_ s

2Rd

gs (; x^ )) ds;

sup ( x^_ s 1(jx^_ s j k) gs (; xk )) ds:

2Rd

Noting that by (2.8.54)

It (x

k) =

Zt

0

1(jx^_ sj > k) supd ( 2R

+

Zt

0

gs (; xk )) ds

1(jx^_ s j k) supd ( x^_ s 2R

gs (; xk )) ds

and that condition 1 of the lemma and conditions I and II easily yield the convergence lim

k!1

Zt

0

1(jx^_ s j > k) supd ( 2R

gs (; xk )) ds = 0;

we have that (2.8.55) would follow by Zt

Zt

lim sup sup (x^_ s gs (; xk )) ds sup (x^_ s gs (; x^ )) ds: k!1 2Rd 2Rd 0 0 (2.8.56) By the de nition of (s; x; y) (s; x^ ; x^_ s ) x^_ s gs ((s; x^ ; x^_ s ); x^ ) = sup ( x^_ s gs (; x^ )); (2.8.57a) 2Rd (s; xk ; x^_ s ) x^_ s gs ((s; xk ; x^_ s ); xk ) = sup ( x^_ s gs (; xk )): (2.8.57b) 2Rd We denote ^ s = (s; x^ ; x^_ s ); ^ ks = (s; xk ; x^_ s ) 1(j(s; xk ; x^_ s )j ak ); (2.8.58)

© 2001 by Chapman & Hall/CRC

240

Maxingales

where ak " 1 are chosen so that Zt

lim sup sup ( x^_ s k!1 2Rd 0

gs (; xk )) ds Zt

= lim sup (^ ks x^_ s k!1 0

gs (^ ks ; xk )) ds:

Then (2.8.56) and the lemma would be proved if Zt

lim sup (^ ks x^_ s k!1

gs (^ ks ; xk )) ds

0

Zt

(^s x^_ s gs(^s; x^ )) ds: (2.8.59) 0

We rst note that by continuity of (s; x; y) in x, the convergence xk ! x^ and (2.8.58) lim ^ k k!1 s

= ^ s ; s 2 R+ :

(2.8.60)

We now show that for some > 0 sup k

Zt

0

jgs (^ks ; x^ )j ds < 1:

(2.8.61)

By Young's inequality for every > 0 Zt

0

1 ^ ks x^_ s ds

Zt

0

1 gs (^ ks ; x^ ) ds + It (^x)

(the integrals are well de ned since ^ ks is bounded). Thus, since ^ ks x^_ s gs (^ ks ; xk ) 0 (see (2.8.57a) and (2.8.58)), we have that Zt

0

gs (^ ks ; xk ) ds

© 2001 by Chapman & Hall/CRC

1

Zt

0

1 gs (^ ks ; x^ ) ds+ It (^x):

(2.8.62)

241

Maxingale problems

Condition 3 of the lemma implies that, given t 2 R+ and x 2 C , there exist l > 1 and " > 0 such that gs (l; x) (1 + ")lgs (; x) for all jj large enough and s t (by condition 1 of the lemma gs (; x) is nonnegative if jj is large enough). Hence, gs (lp; x) (1 + ")p lp gs (; x) for arbitrary p 2 N so that for arbitrary M > 1 there exists L > 1 such that for all jj large enough

gs (; x) ML gs ; x ; s t: L

(2.8.63)

Combining this with condition 4 and recalling condition 1, we conclude that there exist Æ > 0; > 0 and k0 2 N such that for arbitrary M > 1 there exist L > 1 and A > 0 for which gs (; xk ) Ægs (; x^ ) ÆML gs ; x^ ; s t; L for all k k0 and jj A. Choosing now = =L and M = 2=(Æ), we have that gs (; xk ) 2gs (; x^ )=; s t; when k k0 and jj A. Hence, for k k0 Zt

0

x

gs (^ ks ; k ) ds Zt

0

Zt

0

inf g (; x jjA s

2 inf gs (; xk ) ds + jjA

k ) ds +

Zt

0

Zt

0

gs (^ ks ; x^ ) ds

gs (^ ks ; xk ) 1(j^ ks j > A) ds 2

Zt

0

sup gs (; x^ ) ds; jjA

and (2.8.62) yields for k k0 after a simple algebra Zt

0

gs (^ ks ; x^ ) ds It (^x) + 2

Zt

0

sup gs (; x^ ) ds jjA

Zt

0

inf g (; xk ) ds; jjA s

which yields, since It (^x) < 1 and conditions I and II hold, sup k

Zt

0

gs (^ ks ; x^ ) ds < 1:

© 2001 by Chapman & Hall/CRC

242

Maxingales

Inequality (2.8.61) now follows since the functions gs (^ ks ; x^ ) are bounded from below uniformly in s 2 [0; t] and k in view of (2.8.52). Inequalities (2.8.61) and (2.8.62) yield, since by (2.8.52) the functions gs (^ ks ; xk ) are bounded from below uniformly in s t and k, sup k

Zt

0

jgs (^ks ; xk )j ds < 1:

(2.8.64)

Since condition I, the convergences xk ! x^ and (2.8.60) imply that gs (^ ks ; xkR) ! gs (^ s ; x^ ); s t; as k ! 1, by Fatou's lemma and (2.8.64) 0t jgs (^ s ; x^ )j ds < 1; hence, by Lemma 2.7.12 Zt

0

(^ s x^_ s )_0 ds It (^x)+

Zt

0

jgs(^s ; x^ )j ds < 1:

(2.8.65)

On the other hand, since by (2.8.57a) and (2.8.58) ^ s x^ s gs (^ s ; x^ ) 0 and by (2.8.52) the function gs (^ s ; x^ ) is bounded from below on [0; t], we conclude that ^ s x^ s is bounded from below on [0; t]. This fact and (2.8.65) show that Zt

0

j^s x^_ sj ds < 1;

(2.8.66)

R in particular, 0t ^ s x^_ s ds is well de ned and nite. Since the convergence gs (^ ks ; xk ) ! gs (^ s ; x^ ); s t; and uniform boundedness from below of the functions gs (^ ks ; xk ); s 2 [0; t]; k 2 N ; also imply by Fatou's lemma that

lim inf k!1

Zt

0

gs (^ ks ; xk ) ds

Zt

0

gs (^ s ; x^ ) ds;

we conclude that (2.8.59) would follow by lim sup k!1

Zt

0

x

^ ks ^_ s ds

Zt

0

^ s x^_ s ds:

To prove the latter, note that by La Vallee Poussin's theorem the sequence f(^ ks ; s 2 [0; t]); k 2 N g is uniformly integrable with respect

© 2001 by Chapman & Hall/CRC

243

Maxingale problems

to Lebesgue measure in view of (2.8.61) and condition 1 of the lemma. This fact and convergence (2.8.60) yield for arbitrary m > 0 lim

k!1

Zt

x 1(jx^_ s j m) ds =

^ ks ^_ s

0

Zt

0

^ s x^_ s 1(jx^_ s j m) ds:

In view of (2.8.66), we thus complete the proof by showing that lim sup lim sup m!1 k!1

Zt

^ ks x^_ s 1(jx^_ s j > m) ds 0:

0

(2.8.67)

Given arbitrary " > 0, by (2.8.61) we can choose M1 > 0 such that 1 sup M1 k

Zt

0

jgs(^ks ; x^ )j ds ":

(2.8.68)

By (2.8.63) we can choose A1 > 0 and L1 > 0 such that gs (; x^ ) M1 L1 gs (=L1 ; x^ ) for s 2 [0; t] and jj > A1 . Young's inequality then yields Zt

0

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds

L1 +

Zt

0

hZt

0

gs

^k

sup ( x^_ s 2Rd

M1 1

Zt

0

x 1(jx^_ s j > m) 1(j^ks j > A1 ) ds

s;^ L1

i

gs (; x^ )) 1(jx^_ s j > m) ds

jgs (^ks ; x^ )j ds + L1

Zt

0

sup ( x^_ s

2Rd

gs (; x^ )) 1(jx^_ s j > m) ds:

Inequality (2.8.68) and niteness of It (^x) then imply that lim sup lim sup m!1 k!1

Zt

0

© 2001 by Chapman & Hall/CRC

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds ":

244

Maxingales

Since Zt

0

Zt

x 1(jx^_ s j > m) ds A1 jx^_ sj 1(jx^_ s j > m) ds

^ ks ^_ s

+

Zt

0

and

Rt 0

0

^ ks x^_ s 1(jx^_ s j > m) 1(j^ ks j > A1 ) ds;

jx^_ sj ds < 1, the proof of (2.8.67) is over.

The following uniqueness result, which is stated in terms of bs (x) and g^s (; x), is a direct consequence of Theorem 2.8.32.

Theorem 2.8.33. Let conditions I and II hold, and g^s(; x) be differentiable and strictly convex in . Let the following hold: 1. for every t 2 R+ and compact K C

lim inf inf jj!1 st x2K

g^s (; x) jj = 1;

2. for every t 2 R+ and compact K C

sup sup jbs (x)j < 1; st x2K

x 2 C there exists l > 1 such that g^ (l; x) lim inf inf s > 1; jj!1 st lg^s (; x) 4. for every t 2 R+ , compact K C and x 2 K there exist > 0 3. for every t 2 R+ and

and > 0 such that

lim inf inf jj!1 st

g^s (; x0 ) inf > 0: g^s (; x) x0 2K : 0 sup jxr xr j rt

Then = x .

As another consequence, we have the following.

© 2001 by Chapman & Hall/CRC

245

Maxingale problems

Theorem 2.8.34.

Let be a deviability on C under which the canonical idempotent process X is a semimaxingale starting at x with local characteristics (b; c; ; 0). Let the following conditions hold: 1. the functions Rbs (x) and cs (x) are continuous in x 2 C and the function Rd (ex 1 x) s (dx; x) is continuous in (; x) 2 Rd C ; 2. for every t 2 R+ , A > 0 and compact K C

sup sup jbs (x)j < 1; sup sup kcs (x)k < 1; st x2K st x2K Z sup sup (eAjxj 1 Ajxj) s (dx; x) < 1 st x2K d R

and

lim sup sup !0 st x2K

Z Rd

(ex 1 x) s (dx; x) = 0;

3. for every t 2 R+ and compact K C there exists B 2 R+ such that a)

Z

1( x > B ) s(dx; x) > 0;

inf inf inf 2Rd : st x2K Rd jj=1

b) for every x 2 K there exist > 0 and > 0 such that

lim inf inf inf v!1 2Rd : st jj=1 Z

Rd

Z

Rd

Then = x .

© 2001 by Chapman & Hall/CRC

inf

x0 2K : sup jxr x0r j rt

exp(v x) 1( x > B ) s (dx; x0 ) exp(v x) 1( x > B ) s(dx; x)

> 0:

246

Maxingales

Proof. We check the conditions of Theorem 2.8.33. Conditions I and II are met by Lemma 2.8.16. Condition 2 of Theorem 2.8.33 holds by hypotheses. Condition 1 of Theorem 2.8.33 follows since g^s (; x) grows at least exponentially fast in , which results from the inequalities

g^s (; x)

Z

1 x) s (dx; x)

(ex

Rd

(ejjB 1 jjB )

Z

Rd

1( x > B jj) s(dx; x)

and condition 3a). Let us consider condition 3 of Theorem 2.8.33. We have that Z

1 g^s (; x) = cs (x) + 2

Rd

Z

(ex

1 x) s (dx; x)

12 cs(x) + 13 (e2x 1 2 x) 1( x > 0) s(dx; x) 1 + 2

Z Rd

Rd

( x)2 1( x 0) s (dx; x)

1 1 g^ (2; x) + 3 s 2

Z Rd

( x)2 s (dx; x):

Therefore, by the fact that g^s (; x) grows at least exponentially fast in and conditions 2 and 3a) of the theorem

g^ (2; x) lim inf inf s jj!1 st g^s (; x)

3;

verifying condition 3 of Theorem 2.8.33. Finally, condition 4 of Theorem 2.8.33 follows by the inequalities

g^s (; x)

Z Rd

ex=2 1( x > B jj) s (dx; x)

© 2001 by Chapman & Hall/CRC

247

Maxingale problems

when jj is large enough, and 1 1 () cs (x)() + 2 2

g^s (; x) 1 + eBjj 2

Z

(x)2 s(dx; x)+

Rd

Z

Rd

Z Rd

( x)2 s (dx; x)

ex 1(x > B jj) s (dx; x);

and conditions 2 and 3.

Remark 2.8.35. Condition 3b) holds if condition 3a) holds and R ln Rd eujxj 1(jxj > B )s (dx; x) lim sup sup sup < 1: u

u!1 st x2K

Theorems 2.8.21, 2.8.28, 2.8.32, 2.8.33, and 2.8.34 are concerned with \the nondegenerate case" singled out by condition 1 of Theorem 2.8.33. We now consider another degenerate case along with Theorem 2.8.29, which takes advantage of the generality of Lemma 2.8.31.

Theorem 2.8.36.

Let be a deviability on C under which the canonical process X is a semimaxingale starting at x with local characteristics (b; 0; ; 0) such that for some l 2 N and vi 2 Rd

bs(x) =

l X i=1

b(si) (x)vi ;

where R+ -valued functions progressively measurable. Let also for every t 2 R+

x) > 0;

inf inf b(i) ( st x2K s

Then = x .

s ( ; x) =

l X i=1

1(vi 2

)b(si) (x);

b(si) (x) are continuous in

x

and

C-

and compact K C

sup sup b(si) (x) < 1; 1 i l: st x2K

Proof. Let N denote the smallest closed convex cone containing P v1 ; : : : ; vl . Noting that gs (; x) = li=1 exp( vi ) 1 b(si) (x), one can see that the hypotheses of Lemma 2.8.31 are satis ed. Speci cally, the following stronger versions of conditions 1 and 2 of Lemma 2.8.31 hold:

© 2001 by Chapman & Hall/CRC

248

Maxingales

10 : for every t 2 R+ and compact K C

g (; x) limd inf inf s = 1; s t x 2 K j projN j 2R : jprojN j!1

20 : for every t 2 R+ and compact K C inf inf inf gs (; x) >

2Rd st x2K

1:

Property 20 is obvious, property 10 follows by the inequality

jprojN j c i=1 max ( vi ) _ 0; ;:::;l

(2.8.69)

where c is a constant depending only on v1 ; v2 ; : : : ; vl . By Lemma 2.8.31 (pt 1 Æ pt x) = x;t (x) when x0 = x, x_ s 2 N and sups2R+ jx_ s j < 1. Therefore, in analogy with the proof of Theorem 2.8.32 it suÆces to show that the set D = fx 2 C : x_ s 2 N; sups2R+ jx_ s j < 1; x0 = xg is x {dense in C . Let x^ 2 C be such that x (^x) > 0 and xk be de ned by (2.8.54). By Lemma 2.8.31 x^_ s 2 N , so xk 2 D. The0 argument of the proof of Theorem 2.8.32 with the use of property 2 implies that it suÆces to establish (2.8.56) for t 2 R+ . Since x^_ s 2 N , we have that x^_ s projN x^_ s for 2 Rd , which implies by properties 10 and 20 that sup2Rd ( x^_ s gs (; xk )) sup2Rd (projN x^_ s gs (; xk )) < 1, k 2 N . Therefore, by a measurable-selection theorem there exist Lebesgue measurable functions (~ ks ; s 2 R+ ) such that "

~ ks x^_ s gs (~ ks ; xk ) sup ( x^_ s gs (; xk )) 2Rd

1 k

#

_0:

Then, for suitable ak > 0, the functions ^ ks = ~ ks 1(j~ ks j bounded and satisfy the equality Zt

lim sup sup ( x^_ s k!1 2Rd 0

gs (; xk )) ds Zt

= lim sup (^ ks x^_ s k!1 0

© 2001 by Chapman & Hall/CRC

ak ) are

gs (^ ks ; xk )) ds

249

Maxingale problems

so (2.8.56) and the theorem are proved if Zt

lim sup (^ ks x^_ s gs (^ ks ; xk )) ds k!1 0

Zt

0

sup (x^_ s gs (; x^ )) ds:

2Rd

Next, the hypotheses on b(si) (x) imply that, given arbitrary " 2 (0; 1), we have, for k large enough, (1 ")

Zt

0

x

b(si) ( k ) ds

Zt

0

x

b(si) (^ ) ds

Zt

(1+") b(si) (xk ) ds: 0

Hence, for arbitrary 2 Rd Zt

(1 + ")

0

gs (; x

k ) ds

l X i=1

evi =

Zt

Zt

0

0

x

b(si) (^ ) ds

gs (; x^ ) ds

l 1+"X 1 " i=1

2"

Zt

0 Z l t X

1 " i=1

0

b(si) (^x) ds b(si) (^x) ds:

Therefore, Zt

0

(^ ks x^_ s

gs (^ ks ; xk )) ds +"

Zt

0

Zt

0

sup ( x^_ s 2Rd

gs (; x^ )) ds

l 2" X gs (^ ks ; xk ) ds + 1 " i=1

Zt

0

b(si) (^x) ds;

and, since " can be taken arbitrarily small, the proof is complete if sup k

Zt

0

jgs (^ks ; xk )j ds < 1:

(2.8.70)

The proof of the latter inequality is similar to that of (2.8.64) in the proof of Theorem 2.8.32. More speci cally, it is not diÆcult to check that under the assumptions of the theorem the following holds:

© 2001 by Chapman & Hall/CRC

250

Maxingales

1. for every t 2 R+ , compact K C and A 2 R+ , Zt

0

sup

x2K

sup

2Rd : jprojN jA

jgs(; x)j ds < 1;

2. for every t 2 R+ and x 2 C , there exists l > 1 such that

g (l; x) lim inf inf s > 1; d s t lgs (; x) 2R : jprojN j!1

3. for every t 2 R+ , compact K C and x 2 K , there exist > 0 and > 0 such that gs (; x0 ) > 0: lim inf inf inf st x0 2K : 0 gs (; x) 2Rd : sup jxr xr j jprojN j!1 rt Note that part 2 follows from (2.8.69), the other two properties being obvious. These conditions, along with properties 10 and 20 above, imply (2.8.70) in the same way as in the proof of Theorem 2.8.32 conditions I and II together with conditions 1{4 of Theorem 2.8.32 implied (2.8.64) with jprojN j playing the role of jj.

© 2001 by Chapman & Hall/CRC

Part II

Large Deviation Convergence of Semimartingales

251 © 2001 by Chapman & Hall/CRC

Chapter 3

Large deviation convergence This chapter contains basic facts on large deviation convergence in Tihonov spaces and their adaptation to the setting of the Skorohod space.

3.1 Large deviation convergence in Tihonov spaces In this section we develop the theory of large deviation convergence in Tihonov spaces. Our exposition is along the lines of the content of Section 1.9. Let E be a topological space with Borel -algebra B(E ). Let be a directed set, fP ; 2 g be a net of probability measures on (E; B(E )), and fr ; 2 g be a net of real numbers greater than 1 converging to 1 as 2 . We recall that Cb+(E ), C +b (E ), and C +b (E ) denote the respective sets of R+ -valued bounded continuous functions on E , R+ -valued bounded upper semi-continuous functions on E , and R+ -valued bounded lower semi-continuous functions on E . Let be an F -idempotent probability on E , where F denotes the collection of closed subsets of E .

De nition 3.1.1. We say that the net fP ; 2 g large deviation 253 © 2001 by Chapman & Hall/CRC

254

Large deviation convergence

(LD) converges at rate r to if for every h 2 Cb+(E ) Z

lim 2

1=r

h(z )r dP (z )

=

E

_

E

h(z ) d(z ):

(3.1.1)

Remark 3.1.2. One could also consider the version of the above

de nition where h ranges in the set of R+ -valued bounded continuous functions on E of compact support. Then the de nition we have given would refer to \weak large deviation convergence", while the case of compactly supported h would specify \vague large deviation convergence". Since our focus is on \weak large deviation convergence", we simply call it \large deviation convergence".

Note that if E is a Tihonov topological space, then according to Theorem 1.7.27 the F -idempotent probability is uniquely speci ed by the right-hand sides of (3.1.1). We generally denote the large deviation convergence by P rld! . Since the net r is xed in the

rest of the chapter, we simplify the notation by writing P ld ! . We R 1=r 1=r 1=r denote P (A) = P (A) and kf k = E f (z )r dP (z ) , where f : E ! R+ . We state a Portmanteau theorem for large deviation convergence.

Theorem 3.1.3. Let E be a Tihonov topological space. The following conditions are equivalent.

ld ! : _ 2: (i) lim inf kgk g d

1:

P

(ii) lim sup kf k

E _ E

for all g 2 C +b (E );

f d for all f 2 C +b (E ):

20 : The inequalities of part 2 hold for all lower semi-continuous relative to , bounded Borel-measurable functions g : E ! R+ and all upper semi-continuous relative to , bounded Borel-measurable functions f : E ! R+ , respectively. 3: (i)

lim inf P1=r (G) (G)

(ii) lim sup P1=r (F ) (F )

© 2001 by Chapman & Hall/CRC

for all open sets G E; for all closed sets F

E:

255

LD convergence in Tihonov spaces

30 : The inequalities of part 3 hold for all open relative to Borelmeasurable sets G and closed relative to Borel-measurable sets F , respectively. 4: lim P1=r (H ) = (H ) for all continuous relative to Borel-measurable sets H E: 5:

lim khk =

_

6:

lim khk =

_

E

h d

for all continuous relative to bounded Borel-measurable functions h : E ! R+ :

h d

for all bounded Borel-measurable E functions h : E ! R+ that are uniformly continuous with respect to a given uniformity on E . Proof. The proof is almost identical to the one of Theorem 1.9.2. We give it here to make the reading easier. Clearly, 1 ) 6, 2 , 20 , 2 ) 1, 2 ) 3, 2 ) 5, 3 , 30 , 30 ) 4, and 5 ) 1. We prove the implication 1 ) 3. To prove 1 ) 3(i), we note that, since E is Tihonov and G is open, 1(G) = sup h over h 2 Cb+W(E ) such that h 1(G). Therefore, by Theorem 1.4.4 (G) = suph E h d, W so that if h 1(G) is such that (G) E h" d + ", then

lim inf P1=r (G) lim kh" k =

_

E

h" d (G) ":

The proof of 3(ii) is analogous if we note that 1F = inf h over h2W Cb+ (E ) such that h 1(F ) so that by Theorem 1.4.19 (F ) = inf h E h d. We prove that 3(i) ) 2(i) and 3(ii) ) 2(ii). For g 2 C +b (E ) such that kgk = 1 let hi i i gk (z ) = max 1 g(z ) > ; k 2 N: i=0;:::;k 1 k k Since the sets fz : g(z ) > xg are open by the lower semi-continuity of g, 3(i) yields lim inf kgk k max lim inf

i=0max ;:::;k

P1=r g(z ) >

i i k

k _ i i _ g(z ) > = gk d g d k k E E

i=0;:::;k 1 hi

1

hi

© 2001 by Chapman & Hall/CRC

1 : k

256

Large deviation convergence

The proof of 3(ii) ) 2(ii) is similar if we consider fk (z ) = maxi=0;:::;k 1 (i + 1)=k 1(f (z ) i=k) : Now we prove 4 ) 3. Let G be open andWÆ > 0. Let h be a function from Cb+ (E ) such that h 1(G) and E h d (G) Æ. Let Hu = fz 2 E : h(z ) ug; u 2 [0; 1]: Then the function (Hu ) increases as u #W0. Therefore, it has at most countably many jumps. Also (Hu ) E h d u, so (Hu ) (G) 2Æ for u small enough. Thus, there exists " > 0 such that (H" ) (G) 2Æ and (Hu ) is continuous at ". By -maxitivity of the latter is equivalent to H" being continuous relative to , so we conclude that lim inf P1=r (G) lim P1=r (H") = (H" ) (G) 2Æ:

The proof of 3(ii) is similar. We prove that 6 ) 3(ii). Let V be a uniformity on E and F be a closed subset of E . Let f g be a collection of uniformly continuous with respect to V pseudo-metrics on E , which is closed under the formation of maximums and such that 1(F ) = inf ">0 inf (1 (z; F )=")+ . (As above, (z; F ) = inf z0 2F (z; z 0 ).) The functions (1 (z; F )=")+ are bounded and uniformly continuous with respect to V so that by Theorem 1.7.7 +

lim sup P1=r (F ) inf inf lim k 1 (z; F )=" ">0

= inf inf ">0

_

E

k

(1 (z; F )=")+ d(z ) = (F ):

The implication 6 ) 3(i) is proved in an analogous manner.

Remark 3.1.4. As the proof shows, in part 6 it is enough to require

that the convergences hold for functions h that are Lipshitz continuous with respect to the pseudo-metrics specifying the uniformity.

Remark 3.1.5. Part 3 of the theorem can be used to de ne \narrow

large deviation convergence", which is identical to the de nition of the large deviation principle, see, e.g., Varadhan [128]. Thus, on Tihonov spaces large deviation convergence is equivalent to the large deviation principle.

We recall that Br (z ) denotes the closed r-ball about an element z of a metric space.

© 2001 by Chapman & Hall/CRC

LD convergence in Tihonov spaces

257

Corollary 3.1.6. Let E be a Tihonov topological space. If P !ld , then

(z ) = lim0 lim inf P1=r (U ) = lim0 lim sup P1=r (cl U ); U 2Uz 2 U 2Uz 2 where Uz0 is a collection of open neighbourhoods of z whose closures decrease to z . In particular, if E is a metric space, then

(z ) = lim lim sup P1=r (Br (z )): r!0 2 Proof. The claim follows by the inequalities

(z ) (U ) lim inf P1=r (U ) lim sup P1=r (cl U ) 2 2 (cl U); and the fact that limU 2Uz0 (cl U ) = (z ). The following fact follows by Theorem 3.1.3.

Corollary 3.1.7. Let E be a Tihonov topological space. Let a Borel

subset E0 of E be equipped with relative topology. Let P (E n E0 ) = ~, (E n E0 ) = 0 and the restriction of to E0 , which is denoted by be -smooth relative to the collection of closed subsets of E0 . Then ld ld ~ P ! if and only if P~ ! , where P~ denotes the restriction of P to E0 .

Remark 3.1.8. The -smoothness property of ~ holds if either E0 is a closed subset of E or is a deviability on E .

Lemma 1.9.14 provides us with the following corollary.

Corollary 3.1.9. Let E be a Tihonov topological space and P !ld , where is supported by E0 E . Then the following holds. 1.

lim khk =

_

E

h d

for all E0 -continuous bounded Borel-measurable functions h : E ! R+ ;

© 2001 by Chapman & Hall/CRC

258

Large deviation convergence

2. (i)

lim inf kgk

_

E

g d

for all E0 -lower-semi-continuous bounded Borel-measurable functions g : E ! R+ ; _ (ii) lim sup kf k f d

E

for all E0 -upper-semi-continuous bounded Borel-measurable functions f : E ! R+ ;

lim inf P1=r (G) (G) for all E0 -open Borel-measurable sets G E; (ii) lim sup P1=r (F ) (F )

3. (i)

for all E0 -closed Borel-measurable sets F 4.

E;

lim P1=r (H ) = (H ) for all E0 -continuous Borel-measurable sets H E:

The next corollary allows one to strengthen topology for which LD convergence can be proved. We say that two topologies on a topological space are locally equivalent at a given point if they have equivalent local bases at the point.

Corollary 3.1.10. Let O1 and O2 be Tihonov topologies on E , and

let O2 be ner than O1 . Let E0 E be such that O1 and O2 are ld locally equivalent at every z 2 E0 . If P ! for topology O1 , the P can be extended to probabilities on the Borel -algebra of E generated by O2 , is supported by E0 and the restriction of to E0 is -smooth relative to the collection of closed subsets of E0 for the topology induced on E0 by O1 , then is a -smooth idempotent probability relative to the collection of sets closed in topology O2 and ld P ! for topology O2 . Proof. Since the topologies induced on E0 by O1 and O2 coincide, (E n E0 ) = 0 and the restriction of to E0 is -smooth relative to the collection of closed subsets of E0 for the topology induced on E0 by O1 , is a -smooth idempotent probability relative to the collection of sets closed in topology O2 . The required LD convergence follows by the fact that if h : E ! R+ is continuous for topology O2 , then it is E0 -continuous for topology O1 .

© 2001 by Chapman & Hall/CRC

259

LD convergence in Tihonov spaces

Below, we will need convergence of integrals of not necessarily bounded functions. This requires an analogue of uniform integrability. De nition 3.1.11. A Borel-measurable function f : E ! R+ is said to be uniformly exponentially integrable (of order r ) with respect to the net fP g if Z

lim lim sup a!1 2

E

1=r

f (z )r 1(f (z ) > a) dP (z )

= 0:

In analogy to uniform integrability by Chebyshev's inequality the uniform exponential integrability holds if for some > 0 Z

lim sup

1=r

f (z )r (1+") dP (z )

< 1:

E

Lemma 3.1.12. Let E be Tihonov.

ld Let P ! as 2 and be supported by E0 E . Then the following holds. 1: For all E0 {continuous and uniformly exponentially integrable with respect to fP g Borel-measurable functions h : E ! R+ Z

lim 2

1=r

h(z )r dP (z )

=

E

_

h(z ) d(z ):

E

2. For all E0 {lower-semi-continuous Borel-measurable functions g : E ! R+

lim inf 2

Z

g(z )r dP (z )

1=r

_

g(z) d(z): E

E

Proof. The second part, being \a Fatou lemma for LD convergence", is proved by a similar means: for a 2 R+ by Corollary 3.1.9

lim inf 2

Z

E

lim inf 2

1=r

g(z )r dP (z ) Z

1=r

(g(z ) ^ a)r dP (z )

E

_

(g(z) ^ a) d(z): E

© 2001 by Chapman & Hall/CRC

260

Large deviation convergence

W

Since the latter converges to E g(z ) d(z ) as a ! 1, the proof of part 2 is over. Part 1 follows by part 2 and the inequalities Z

lim sup 2

E

1=r r h(z ) dP (z ) Z

lim sup 2

+ lim sup 2 _

E

Z

E

1=r

h(z )r 1(h(z ) a)dP (z )

1=r

h(z )r 1(h(z ) > a)dP (z )

h(z) 1(h(z) a) d(z) E

Z

+ lim sup 2

E

1=r

h(z )r 1(h(z ) > a)dP (z )

;

where the latter inequality holds by Corollary 3.1.9. The following lemma gives an extension in a dierent direction.

Lemma 3.1.13. Let E be a Tihonov topological space.

ld Let P ! . Let h : E ! R+ be uniformly bounded and Borel-measurable functions such that for a function h : E ! R+

lim h (z ) = h(z )

2

for -almost every z 2 E and every net z ! z as 2 . Then Z

lim 2

1=r

h (z )r dP (z )

E

=

_

E

h(z ) d(z ):

Proof. The proof is similar to the one of Lemma 1.10.2 so we only sketch it. De ning h (z ) = inf U 2Uz supz0 2U sup0 h (z 0 ), where as we recall Uz denotes the collection of open neighbourhoods of z , and h(z ) = inf 2 h (z ), we can write for arbitrary > 0 and suitable 0

© 2001 by Chapman & Hall/CRC

261

LD convergence in Tihonov spaces

by Theorem 3.1.3 applied to h0 that Z

lim sup 2

E

1=r

h (z )r dP (z ) Z

lim sup 2

E

1=r

h0 (z )r dP (z )

_

_

E

h0 (z ) d(z )

h(z ) d(z ) +

E

_

E

h(z ) d(z ) + :

The complementary inequality lim inf 2

Z

1=r

h (z )r dP (z )

_

h(z) d(z) E

E

is proved by a mirror argument. Speci cally, de ning h (z ) = supU 2Uz inf z0 2U inf 0 h (z 0 ) and h(z ) = sup2 h (z ), we have for arbitrary > 0 and suitable 1 lim inf 2

Z

E

lim inf 2

1=r

h (z )r dP (z ) Z

E

1=r

h1 (z )r dP (z )

_

E

_

h1 (z) d(z) E

h(z ) d(z )

_

E

h(z ) d(z )

:

As a consequence of Lemma 3.1.13, we obtain the following version of the contraction principle on preservation of LD convergence under mappings. Theorem 3.1.14. Let E be a Hausdor topological space and E 0 be a Tihonov topological space. Let be a deviability on E . Let Borelmeasurable functions f : E ! E 0 , 2 ; and a -Luzin-measurable function f : E ! E 0 be such that for -almost every z 2 E and ld every net z ! z we have that f (z ) ! f (z ). If P ! , then ld 1 1 P Æ f ! Æ f .

© 2001 by Chapman & Hall/CRC

262

Large deviation convergence

Proof. The proof is similar to the one of Theorem 1.10.3. We rst note that Æf 1 is a deviability on E 0 by Theorem 1.7.11. Next, for an R+ -valued bounded continuous function h on E 0 by a change of variables and Lemma 3.1.13

lim

2

Z

E0

1=r 1 0 r 0 h(z ) dP Æ f (z )

Z

= lim 2

E

h Æ f (z )r dP (z ) =

_

E

1=r

h Æ f (z ) d(z ) =

_

E0

h(z 0 ) d Æ f 1(z 0 ):

The following consequence is often used below.

Corollary 3.1.15. Let E be a Hausdor topological space, E 0 ldbe a

Tihonov topological space, and be a deviability on E . If P ! as 2 and f : E ! E 0 is Borel measurable and continuous -a.e., ld then P Æ f 1 ! Æf 1.

We now derive a criterion of \large deviation relative compactness" in the theme of Prohorov's one for weak convergence.

De nition 3.1.16. An F {idempotent probability on E is called

a large deviation (LD) accumulation point of fP ; 2 g (for rate r ) if there exists a subnet fP0 ; 0 2 0 g of fP ; 2 g that LD converges (at rate r0 ) to .

De nition 3.1.17. The net fP ; 2 g is called large deviation 0 0 (LD) relatively compact (for rate r ) if every subnet fP0 ; of fP ; 2 g has an LD accumulation point (for rate r0 ).

2 g

We recall that K denotes the collection of compact subsets of E .

De nition 3.1.18. The net fP ; 2 g is called exponentially tight (of order r ) if inf K 2K lim sup2 P1=r (K c ) = 0.

Theorem 3.1.19. Let E be a Tihonov topological space.

1. If the net fP ; 2 g is exponentially tight, then it is LD relatively compact, the accumulation points being deviabilities.

© 2001 by Chapman & Hall/CRC

LD convergence in Tihonov spaces

263

2. Let E be, in addition, a locally compact Hausdor topological space. If the net fP ; 2 g is LD relatively compact, then it is exponentially tight. Proof. The proof is analogous to the proof of Theorem 1.9.17. We start with part 1. Let Cb;+1 (E ) = ff 2 Cb+(E ) : kf k 1g. For a given 2 , the mapping V : f ! kf k ; f 2 Cb;+1 (E ); is an element + of the space [0; 1]Cb;1 (E ) . The latter space, endowed with product topology, is compact and Hausdor. Therefore, the net fV ; 2 g + is relatively compact on [0; 1]Cb;1 (E ) so that there exists a subnet + fV0 ; 0 2 0g that converges to an element V of [0; 1]Cb;1 (E) . We extend V to a functional on Cb+(E ) by letting V (c f ) = cV (f ); c 2 C + (E ) R + . By the de nition of topology on [0; 1] b;1

lim kf k 0 = V (f ); f 2 Cb+ (E ):

0 20

(3.1.2)

The latter implies that V satis es conditions (V 0), (V 1) and (V 2) of Theorem 1.7.25, i.e., (V 0) V (1) = 1, (V 1) V (c f ) = cV (f ); c 2 R+ ,

(V 2) V (f _ g) = V (f ) _ V (g). The rst two properties directly follow by (3.1.2). The third property is valid in view of the inequalities kf k _ kgk kf _ gk 21=r kf k _ kgk and (3.1.2). Also, exponential tightness of fP ; 2 g and (3.1.2) imply that V is tight in the sense of Theorem 1.7.25. Thus, the functional V satis es all the conditions of Theorem 1.7.25, so according to the W theorem V (f ) = E f d; f 2 Cb+(E ); for some deviability , which ld implies that P0 ! (at rate r0 ). This completes the proof of part 1. Part 2 follows since by the argument of the proof of (1.9.5), where we use large deviation relative compactness of fP g in place of weak relative compactness of f g, for arbitrary " > 0 there exist open sets A1 ; : : : ; Ak with compact closures such that

k [

lim sup P1=r E n Ai 2 i=1

© 2001 by Chapman & Hall/CRC

":

264

Large deviation convergence

(Cf. also the proof of part 2 of Theorem 3.1.28 below.)

Corollary 3.1.20.

Let fP; ; 2 g, 2 , be nets of Borel measures on respective Tihonov topological spaces E . If the nets fP; ; 2 g are exponentially tight (of order r) for every 2 , then there exists a subnet f(P0 ; ; 2 ); 0 2 0 g of f(P; ; 2 ); 2 g such that the nets fP0 ; ; 0 2 0 g LD converge (at rate r0 ) to deviabilities on the E for every 2 . R

Proof. For f 2 Cb+(E ), let V; (f ) = E fr dP; 1=r . By Tihonov's theorem the set f(V; (f ); f 2 Cb;+1 (E ); 2 ); 2 + Q g is a relatively compact subset of 2 [0; 1]Cb;1 (E ) , where the latter set is equipped with product topology. Thus, there exists a convergent subnet f(V0 ; (f ); f 2 Cb;+1 (E ); 2 ); 0 2 0 g. Now the required follows by the argument of the proof of Theorem 3.1.19.

Theorem 3.1.19 allows us to introduce the following useful concept.

De nition 3.1.21. Let E be a Tihonov topological space and E0 E: We say that the net fP ; 2 g is E0 -exponentially tight if it is exponentially tight and every LD accumulation point is supported by E0 . The following is a version of the contraction principle.

Corollary 3.1.22. Let E be a Tihonov topological space, E0 E , 0

and E be a locally compact Hausdor topological space. If the net fP ; 2 g is E0 {exponentially tight and a function f : E ! E 0 is Borel measurable and E0 -continuous, then the net fP Æf 1 ; 2 g is exponentially tight. Proof. By part 1 of Theorem 3.1.19 the net fP ; 2 g is LD relatively compact. Since f is continuous a.e. with respect to every LD accumulation point of fP ; 2 g, by Corollary 3.1.15 the net fP Æf 1 ; 2 g is LD relatively compact as well; hence, it is exponentially tight by part 2 of Theorem 3.1.19.

We now assume that E is a metric space and introduce \metrics" for large deviation convergence. We rst de ne an analogue of the

© 2001 by Chapman & Hall/CRC

265

LD convergence in Tihonov spaces

Prohorov metric. We again assume as given a net fP ; 2 g of Borel measures on E , a net of real numbers fr ; 2 g greater than 1 converging to 1, and an F -idempotent probability on E . The analogue of the Prohorov metric is de ned by

pld (P ; ) = inf > 0 : P1=r (F ) F + ;

(F ) P1=r F + for all closed F

E : (As above, for A E , we denote A = fz 2 E : (z; A) g). The

next lemma follows by regularity of Borel measures and -maxitivity of idempotent measures.

Lemma 3.1.23. We can equivalently write

pld (P ; ) = inf > 0 : P1=r (A) A + ;

(A) P1=r A + for all A 2 B(E ) :

Remark 3.1.24. The distance pld could equivalently be de ned in terms of open -neighbourhoods.

Theorem 3.1.25. The net fP g LD converges to if and only if ld p (P ; ) ! 0 as 2 .

Proof. The proof is analogous to the proof of Theorem 1.9.22. We rst prove that if pld (P ; ) ! 0, then the P LD converge to . By Theorem 3.1.3 it is suÆcient to prove that, given a closed set F , an open set G, and > 0, there exists Æ > 0 such that if pld (P ; ) < 1=r 1=r Æ, then P (F ) < (F ) + and P (G) > (G) . Since is -smooth relative to F , there exists Æ1 2 (0; =2) such that 1=r (F ) (F Æ1 ) =2. Therefore, if pld (P ; ) < Æ1 , then P (F ) < (F Æ1 ) + Æ1 (F ) + . Next, using -maxitivity of , we choose Æ2 2 (0; =2) such that (G) (G Æ2 ) + =2. Then, if pld (P ; ) < Æ2 , then (G) (G Æ2 ) + =2 < P1=r (G) + . Taking Æ = Æ1 ^ Æ2 proves the claim. ld Conversely, let P ! . We show using again Theorem 3.1.3 that given > 0 there exists a collection of sets Hi; i = 1; : : : ; k and Æ > 0 such that the Hi are continuous relative to and the fact that jP1=r (Hi) (Hi)j < Æ; i = 1; : : : ; k, implies that pld(P ; ) < .

© 2001 by Chapman & Hall/CRC

266

Large deviation convergence

Let Æ < =4. Let closed Æ=2-balls B1 ; :: : ; Bl centred at z1 ; : : : ; zl , respectively, be such that E n [li=1 Bi < Æ. By -maxitivity of for each i = 1; 2; : : : ; l there exists a closed ball Bi0 centred at zi of radius not less than Æ=2 and not greater than Æ, which is a continuous set relative to . Observing that a nite union of sets continuous relative to is also continuous relative to , we take as H1 ; : : : ; Hk 1 the collection of arbitrary unions of the balls B10 ; : : : ; Bl0 . We also take 0 Hk = E n [li=1 Bi0 Æ , where Æ0 > 0 and is chosen such that Hk is continuous relative to and (Hk ) 2Æ. Let jP1=r (Hi ) (Hi)j < Æ; i = 1; : : : ; k. Let F be a closed subset of E and let H 0 be the largest set out of H1 ; : : : ; Hk 1 such that F has non-empty intersection with each of the sets Bi0 that make up H 0 . Then H 0 F 2Æ so that (F ) (F \ H 0 ) + (F \ Hk ) (H 0 ) + (Hk ) < P1=r (H 0 ) + 3Æ P1=r (F 2Æ ) + 3Æ. By a symmetric argument P1=r (F ) P1=r (F \ H 0 )+P1=r (F \Hk ) P1=r (H 0 )+P1=r (Hk ) < (H 0 )+(Hk )+2Æ (F 2Æ ) + 4Æ. Thus, pld (P ; ) < . by

We now de ne an analogue of the Kantorovich-Wasserstein metric

ldBL (P ; ) =

Z + f 2Cb (E ): E kf k BL 1

sup

f (z )r dP (z )

1=r _

E

f (z ) d(z ) :

Theorem 3.1.26. The net fP g LD converges to if and only if ld BL (P ; ) ! 0 as 2 .

Proof. The fact that the convergence ld BL (P ; ) ! 0 implies LD convergence of the P to follows from Theorem 3.1.3 and Remark 3.1.4. For the converse, by Theorem 3.1.25 it suÆces to prove ld that if P ! , then

ld lim sup ld BL (P ; ) 2p (P ; ) 2

0:

Let kf k BL 1. Given Æ > 0, we choose open Æ-balls AÆ (zk ); k = 1; 2; : : : ; l; such that E n [lk=1AÆ (zk ) < Æ. Since lim sup2 P1=r E n [lk=1 AÆ (zk ) E n [lk=1AÆ (zk ) ;we may as sume that P1=r E n[lk=1AÆ (zk ) < Æ. Abbreviating p = pld (P ; )

© 2001 by Chapman & Hall/CRC

267

LD convergence in Tihonov spaces

and recalling that AÆ (z )p denotes the closed p -neighbourhood of AÆ (z ), we have Z

f (z )r dP

(z )

1=r

E

Z

l k=1 max ;:::;l

1=r

f (z )r dP (z )

+Æ

AÆ (zk )

1=r

l1=r k=1 max f (zk ) + Æ P (AÆ (zk )) + Æ ;:::;l 1 =r l k=1 max f (zk ) (AÆ (zk )p ) + p + l1=r Æ + Æ ;:::;l _ 1 =r l f (z) + Æ + p d(z) + l1=r p + l1=r Æ + Æ E

_

l1=r f (z) d(z) + 2l1=r p + 2l1=r Æ + Æ: (3.1.3) E

Similarly, _

E

f (z ) d(z ) max f (zk )(AÆ (zk )) + Æ k=1;:::;l

k=1 max f (zk ) P1=r (AÆ (zk )p ) + p + Æ ;:::;l k=1 max ;:::;l

Z

Z

AÆ (zk )p

f (z ) + Æ + p

f (z ) + Æ + p

E

r

Z

r

dP

1=r

dP (z )

1=r

+ p + Æ

+ p + Æ

f (z )r dP (z )

1=r

+ 2p + 2Æ: (3.1.4)

E

Inequalities (3.1.3) and (3.1.4) imply that

ldBL (P ; ) 2p l1=r 1 (1+2p )+2l1=r Æ +Æ Since l1=r ! 1 as 2 , p 1, and Æ > 0 is arbitrary, the proof is complete. We now consider sequential compactness for metric spaces. We thus assume that = N and replace general nets fr g by sequences frng. Probability measures are denoted by Pn .

© 2001 by Chapman & Hall/CRC

268

Large deviation convergence

De nition 3.1.27.

A sequence fPn ; n 2 N g is LD relatively sequentially compact (for rate rn ) if every subsequence fPn0 g of fPn g contains a further subsequence fPn00 g that LD converges (at rate rn00 ) to an F -idempotent probability on E .

Theorem 3.1.28.

1. Let E be a metric space. If a sequence

fPn ; n 2 N g of probabilities on (E; B(E )) is exponentially tight, then it is LD relatively sequentially compact, the LD accumulation points being deviabilities.

2. Let E be homeomorphic to a complete separable metric space. If a sequence fPn ; n 2 N g is LD relatively sequentially compact, then it is exponentially tight. Proof. We prove part 1. Let us assume rst that E is a separable metric space. Then it is embedded as a dense subspace into a compact metric space E 0 . We extend probabilities on (E; B(E )) to probabilities on (E 0 ; B(E 0 )) by letting P 0 (A0 ) = P A0 \ E ; A0 2 B(E 0 ): The set Cb;+1 (E 0 ) of R+ -valued continuous functions on E 0 that are bounded by 1, endowed with the topology of uniform convergence, is a separable metric space. Let Cb;+1;d (E 0 ) denote a countable dense + 0 subset. The set [0; 1]Cb;1;d (E ) with product topology is sequentially compact, so the diagonal argument yields existence of a subsequence nk such that the sequences f kf k 0nk ; k 2 N g converge for all f 2 Cb;+1;d (E 0 ), where kf k 0nk refers to the norm relative to Pn0 k . The inequality j kgk 0n kgk 0mj j kf k 0n kf k 0m j+2 sup jg(z0 ) f (z0 )j z 0 2E 0 and the fact that Cb;+1;d (E 0 ) is dense in Cb;+1 (E 0 ) show that the sequences f kf k 0nk ; k 2 N g converge for all f 2 Cb;+1 (E 0 ), which implies in analogy with the proof of Theorem 3.1.19 that there exld ists a deviability 0 on E 0 such that Pn0 k ! 0 as k ! 1 at rate rnk . Exponential tightness of fPn ; n 2 N g implies that inf K 2K 0 (E 0 n K ) = 0 (where K is the collection of compact subsets of E ) so that 0 (E 0 n E ) = 0 and the set function de ned by (A) = 0 (A); A E; is a deviability on E by Corollary 1.7.12. W It is left to check that kf k nk ! E f d for all f 2 Cb+(E ). By Theorem 3.1.3 we may assume that f is uniformly continuous on E so that it can be extended to f 0 2 Cb+ (E 0 ), see, e.g., Engelking [47].

© 2001 by Chapman & Hall/CRC

269

LD convergence in Tihonov spaces

0 k 0 ! W 0 f 0 d0 , kf 0 k 0 = kf k n The W required follows since k f nk nk k E W and E 0 f 0 d0 = E f d. Now, if E is an arbitrary metric space, then by the exponential tightness condition there exists a -compact space E 0 E such that limn!1 Pn1=rn (E n E 0 ) = 0. Applying the part just proved to the probabilities Pn0 on the separable metric space E 0 de ned by Pn0 (A) = Pn (A)=Pn (E 0 ); A 2 B(E 0 ), we deduce existence of an LD convergent subsequence for the Pn0 . This provides us with an LD convergent subsequence for the Pn . Part 1 is proved. For part 2 we rst check that for every Æ > 0 and > 0 there exist open Æ-balls A1 ; : : : ; Ak such that

lim sup Pn1=rn n!1

En

k [ i=1

Ai

:

(3.1.5)

Let open Æ-balls Ai be such that [1 i=1 Ai = E . Let subsequences kl and nl be such that lim sup lim sup Pn1=rn k!1 n!1 ld and Pnl ! 0 for some arbitrary k, k

[ 0 E n Ai

i=1

En

k [ i=1

Ai = lim

l!1

Pn1l=rnl

En

kl [ i=1

Ai

F -idempotent probability 0. Then, for

1=r lim sup Pnl nl E l!1

n

k [ i=1

Ai

llim !1

1=rnl Pnl E

n

kl [ i=1

Ai :

The required inequality (3.1.5) follows since limk!1 0 E n k [i=1Ai = 0. Since each Pn is tight by Ulam's theorem, (3.1.5) implies that for arbitrary " > 0 and k 2 N ; there exist open 1=k-balls Ak1 ; : : : ; Aknk such that for all n 2 N

Pn1=rn E n

nk [ i=1

Aki

© 2001 by Chapman & Hall/CRC

2"k :

270

Large deviation convergence

T Snk The set B = 1 k=1 i=1 Aki is totally bounded and hence relatively compact by completeness of E . Also for every n 2 N

Pn1=rn (E nB )

1 X k=1

Pn1=rn E n

nk [ i=1

Aki

":

Remark 3.1.29. Part 1 also follows by Theorem 3.1.19 and Theorem 3.1.25 (or Theorem 3.1.26).

As a consequence, we have the following version of Corollary 3.1.22. The proof is similar.

Corollary 3.1.30. Let E be a metric space and E 0 be homeomorphic

to a complete separable metric space. Let E0 E . If a sequence fPn ; n 2 N g of probabilities on (E; B(E )) is E0 {exponentially tight and a function f : E ! E 0 is Borel measurable and E0 -continuous, then the sequence fPn Æ f 1 ; n 2 N g is exponentially tight.

As an illustration of the use of LD relative compactness arguments, we prove Gartner's theorem. Let L(X ) denote the distribution of a random variable X and E denote expectation with respect to a probability measure P .

Theorem 3.1.31. Let fX ; 2 g be a net of Rk {valued random

variables de ned on respective probability spaces ( ; F ; P ) such that for each 2 Rk 1 lim ln E exp r X = G(); 2 r

where G() is an R -valued lower semi-continuous and essentially ld smooth convex function such that 0 2 int (dom G). Then L(X ) ! ; at rate r , where is the deviability speci ed by the density (x) = exp sup2Rk ( x G()) : Proof. We act as in the proof of Lemma 1.11.19. We rst note that the net fL(X ); 2 g is exponentially tight. To see this, we write by Chebyshev's inequality, for A > 0 and > 0, denoting by ei ,

© 2001 by Chapman & Hall/CRC

271

LD convergence in Tihonov spaces

i = 1; : : : ; 2k, the 2k-vector, whose b(k + 1)=2cth entry is 1 if k is odd, 1 if k is even, and the rest of the entries are equal to 0, P1=r (jX j > A) max P1=r (ei X > A=k) i=1;:::;2k

exp( A=k) i=1max E exp(r ei X ) 1=r : ;:::;2k The exponential tightness follows since by hypotheses

lim E exp(r ei X ) 1=r = exp(G(ei )); 2 where the right-hand side is nite if is small enough by the fact that G() is nite in a neighbourhood of the origin. Therefore, by Theorem 3.1.19 there exists a subnet fX 0 ; 0 2 0 g ld ~ of fX ; 2 g and a deviability ~ on Rk such that L(X 0 ) ! : Next, it follows from Chebyshev's inequality that if 2 int(dom G), then the function exp( x); x 2 Rk is uniformly exponentially integrable with respect to fL(X 0 ); 0 2 0 g, so by Lemma 3.1.12 0 1=r0 _ ~ x); = exp( x) d(

lim0 E exp(r0 X )

Rk

2 int(dom G):

W

~ x) = exp(G()) for all 2 int(dom G), Thus, Rk exp( x) d( ~ = . which as in the proof of Lemma 1.11.19 implies that The following result is proved similarly to Theorem 1.9.28. Theorem 3.1.32. Let E be a Tihonov space. Let G be a subset of + Cb (E ) consisting of uniformly bounded and pointwise equicontinuous functions, i.e., supf 2G supz2E f (z ) < 1 and for every > 0 and z 2 E there exists an open neighbourhood Uz of z such that ld supf 2G supz0 2Uz jf (z ) f (z 0 )j : If P ! , then Z

lim sup f 2G

f r dP

E

1=r

_

E

f d = 0:

As we have mentioned, if we replace space Cb+ (E ) in the de nition of weak LD convergence by space CK+(E ) of R+ -valued continuous functions with compact support, then we obtain the notion of vague LD convergence.

© 2001 by Chapman & Hall/CRC

272

Large deviation convergence

De nition 3.1.33. We say that a net fP ; 2 g of probabilities on (E; B(E )) vaguely LD converges at rate r to a probability on E if for every f 2 CK+(E ) Z

lim 2

f (z )r dP (z )

E

1=r

=

_

E

K-idempotent

f (z ) d(z ):

If E is locally compact and Hausdor, the vague LD convergence has properties similar to the properties of the weak LD convergence. For instance, there is an easy analogue of Theorem 1.9.2. A distinctive feature of this type of LD convergence is that nets of probability measures are LD relatively compact.

Theorem 3.1.34. Let E be a locally compact Hausdor topologi-

cal space. Then a net fP ; 2 g of probabilities on (E; B(E )) is vaguely LD relatively compact.

The proof is similar to the proof of part 1 of Theorem 3.1.19, the main distinction being the use of Theorem 1.7.21 in place of Theorem 1.7.25. At times it is more intuitive to formulate large deviation convergence of probability measures as large deviation convergence in distribution of the associated random variables.

De nition 3.1.35. Let fX ; 2 g be a net of random variables

de ned on respective probability spaces ( ; F ; P ) and assuming values in a topological space E and X be an idempotent variable de ned on an idempotent probability space ( ; ) and assuming values in E , whose idempotent distribution is -smooth relative to the collection of closed subsets of E . We say that the net fX ; 2 g large ld deviation converges in distribution to X if P Æ X 1 ! Æ X 1. ld We denote large deviation convergence in distribution by ! as well. Whether this notation refers to large deviation convergence of probability measures or large deviation convergence in distribution of random variables should be clear from the context. We will also occasionally say that a net of random variables is LD relatively compact if the associated net of laws is LD relatively compact. We have the following version of Lemma 3.1.12. Let us say that a net f ; 2 g of R+ -valued random variables on respective prob-

© 2001 by Chapman & Hall/CRC

273

LD convergence in Tihonov spaces

ability spaces ( ; F ; P ) is uniformly exponentially integrable relative to fP ; 2 g (with rate r ) if lim lim sup E r 1( > A) A!1 2

1=r

= 0:

Lemma 3.1.36. Let !ld

, where is an R+ -valued idempotent variable on an idempotent probability space ( ; ). If the net f ; 2 g is uniformly exponentially integrable relative to fP ; 2 g, then lim2 E r 1=r = S: We now prove a number of technical lemmas. Let us recall that if X and Y are random variables with values in respective separable metric spaces E and E 0 with Borel -algebras, then (X; Y ) is a random variable in E E 0 with product topology and Borel algebra; in particular, if E = E 0 and denotes the metric on E , then (X; Y ) is a random variable. The following result is an analogue of Lemma 1.10.5 and admits a proof along the same lines. We give another proof that illustrates the use of metrics.

Lemma 3.1.37. Let E be a separable metric space with metric ,

and let X and Y ; where 2 , 2 , and are directed sets, be nets of random variables with values in E , de ned on respective probability spaces ( ; F ; P ). Let

lim lim sup P1=r (X ; Y ) " = 0; " > 0; 2 2 and

ld L X ! as 2 , where ; 2 ; are F -idempotent probabilities on E . Then, for an F -idempotent probability on E , we have that ld L(Y ) ! as 2

if and only if

iw ! as 2 :

© 2001 by Chapman & Hall/CRC

274

Large deviation convergence

Proof. The claims follow by Theorem 1.9.25 and Theorem 3.1.26 since in view of the de nitions of BL and ld BL BL (

; ) ld BL L(Y );

Z + sup f 2Cb+ (E ):

kf k BL 1

f (X )r dP

ldBL L(X ); +

ldBL L(X );

1=r Z

Z

1=r r f (Y ) dP

r

1 ^ (X ; Y )

dP

1=r

:

We will often use the case where the X do not depend on . Lemma 3.1.38. Let E be a separable metric space with metric , and let X and Y ; where 2 , be nets of random variables de ned on respective probability spaces ( ; F ; P ) with values in E . If ld L(X ) ! , where is an F -idempotent probability on E , and 1=r P X ;Y !

ld 0 as 2 , then L(Y ) ! . We give another application of metrics. De nition 3.1.39. We say that a net fX ; 2 g of random variables on ( ; F ; P ) assuming values in a metric space E with metric converges to z 2 E super-exponentially in probability at rate r (or simply super-exponentially in probability if the rate is under-

stood) and write X every > 0.

1=r P

! z if lim2 P1=r ((X ; z) > ) = 0 for 1=r

Remark 3.1.40. Note that X P! Z

lim 2

r

1 ^ (X ; z )

dP

1=r

z if and only if = 0:

Lemma 3.1.41. Let fX ; 2 g be a net of random variables on

respective probability spaces ( ; F ; P ) assuming values in a metric 1=r P X !

ld space E with metric . Then z if and only if L(X ) ! 1z , where 1z denotes the unit mass at z .

© 2001 by Chapman & Hall/CRC

275

LD convergence in Tihonov spaces

1z , then by the de nition of LD convergence 1=r _ 1^(X ; z ) r dP = 1^(z 0 ; z ) d 1z (z 0 ) = 0:

ld Proof. If L(X ) !

lim

2

Z

E

The converse follows since as in the proof of Lemma 3.1.37

1

ldBL (L(X ); z )

Z

r

1 ^ (X ; z )

dP

1=r

:

The next lemma considers joint LD convergence. We formulate the results in the language of LD convergence in distribution. Lemma 3.1.42. Let E and E 0 be separable metric spaces, and let X and Y ; where 2 , be nets of random variables on ( ; F ; P ) with values in E and E 0 , respectively. Let X and Y be idempotent variables on an idempotent probability space ( ; ) with values in E and E 0 , respectively, whose idempotent distributions are -smooth relative to the associated collections of closed sets. Let E E 0 be equipped with product topology. ld ld 1. If X ! X, Y ! Y , X and Y are independent, and X ld and Y are independent, then (X ; Y ) ! (X; Y ).

2. If

ld X !

X and

1=r P Y !

z , then (X ; Y )

ld ! (X; z ).

Proof. The proof of part 1 uses Theorem 3.1.32 and is analogous to the proof of part 1 of Lemma 1.10.8. In more detail, let PX and PY denote the respective distributions of X and Y , and X and Y denote the respective idempotent distributions of X and Y ; Theorem 3.1.32 implies that for an R+ -valued bounded uniformly continuous function h(z; z 0 ) on E E 0 Z Z lim 2 E E0 Z

E

1=r

h(z; z 0 )r dPY (z 0 ) dPX (z )

sup h(z; z 0 ) Y (z 0 ) r dPX (z ) z 0 2E 0

© 2001 by Chapman & Hall/CRC

1=r

= 0: (3.1.6)

276

Large deviation convergence

Also, since supz0 2E 0 h(z; z 0 ) Y (z 0 ) is continuous in z 2 E , lim 2

Z

E

1=r sup h(z; z 0 ) Y (z 0 ) r dPX (z ) z 0 2E 0 = sup sup h(z; z 0 ) Y (z 0 ) X (z ): (3.1.7) z 2E z 0 2E 0

The required follows by (3.1.6) and (3.1.7). The proof of part 2 is also analogous: we observe that ( 1=r

P 0 )((X ; Y ); (X ; z )) ! 0, where 0 is a product metric on E E 0 , so that the required follows by part 1 and Lemma 3.1.38.

3.2 Large deviation convergence in the Skorohod space The purpose of this section is to lay groundwork for deriving large deviation convergence results for semimartingales. We begin by introducing basic notation for the Skorohod space. We denote by D = D (R + ; Rd ); d 2 N ; the space of Rd {valued, right-continuous with left-hand limits functions x = (xt ; t 2 R+ ). We equip it with the Skorohod J1 topology and metrise it by the Skorohod{Prohorov{Lindvall metric denoted by S , under which it is a complete separable metric space. Let D denote the Borel { algebra on D , Dt , for t 2 R+ , denote the sub{{algebra generated by the coordinate maps x ! xs ; s t; and D = (Dt ; t 2 R+ ). (Note that the ow D is not right-continuous.) Given x 2 D , we denote xt = supst jxsj; xt = sups 0 and Æ > 0, we de ne the modulus of continuity w0 (x; Æ) = inf max wx [tj 1 ; tj ) ; T

(tj ) j =1;:::;k

where wx [s; t) = supu;v2[s;t) jxu xv j; s < t; and the in mum is taken over all collections (tj ) such that 0 = t0 < t1 < : : : < tk = T and tj tj 1 > Æ for j < k. The next theorem routinely follows by characterisation of compacts in D , see, e.g., Jacod and Shiryaev [67], and is analogous to tightness conditions for sequences of probabilities in D , cf. Ethier and Kurtz [48].

Theorem 3.2.1. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ). The net fL(X ); 2 g is exponentially tight if and only if for all T > 0 and > 0 (i) lim lim sup P1=r sup jXt j > A = 0; A!1 2 tT (ii) lim lim sup P1=r wT0 (X ; Æ) > = 0: Æ!0 2 The concept of E0 -exponential tightness with E0 = C plays an important role in the developments below, so we repeat it here.

© 2001 by Chapman & Hall/CRC

278

Large deviation convergence

De nition 3.2.2.

We say that a net fP ; 2 g of probability measures on D is C -exponentially tight if it is exponentially tight and every LD accumulation point is supported by C . The next result gives conditions for C -exponential tightness. Theorem 3.2.3. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be a net of stochastic processes with paths in D , de ned on respective stochastic bases ( ; F ; F ; P ). The net fL(X ); 2 g is C { exponentially tight if and only if either one of the following equivalent conditions I or II holds. I (i) lim lim sup P1=r jX0 j > A = 0; A!1 2 (ii) lim lim sup P1=r sup jXt Xs j > = 0; Æ!0 2 s;t2[0;T ]: js tjÆ

lim lim sup P1=r sup jXt j > A = 0; A!1 2 tT 1=r (ii) lim lim sup sup P sup jX+t X j > = 0; Æ!0 2 2ST (F ) 0tÆ where T > 0 and > 0 are arbitrary, and ST (F ) denotes the set of all F {stopping times not greater than T . Remark 3.2.4. A similar result holds if the X assume values in a complete separable metric space E . Then instead of conditions I(i) and II(i) one should require that the nets fL(Xt ); 2 g be exponentially tight in E for all t 2 R+ , and in I(ii) and II(ii) replace the moduli of the increments of the X by the distances between the values of the X , cf. Ethier and Kurtz [48]. We precede the proof with a lemma. In the rest of the book we use E to denote expectation and F to denote -algebras. Lemma 3.2.5. Let i; i 2 N ; be positive random variables on a probability space ( ; F ; P ) and let

II (i)

(

max k 2 N : At = 0; If Æ > 0, t 2 R+ and N

P (At N ) 1

© 2001 by Chapman & Hall/CRC

Pk i=1 i

2N

t NÆ

t ;

if 1 t; otherwise:

are such that NÆ=t > 1, then

1

max P (k Æ; At k):

k=1;:::;N

279

LD convergence in the Skorohod space

Proof. In view of Chebyshev's inequality for n N

P (At N ) P (At N; n Æ) + P (At N; n Æ) 1 k=1max P (At N; k Æ) + E n 1(At N ) : (3.2.1) ;:::;N Æ Choosing n such that E n 1(At N ) = minkN E k 1(At N ) , we estimate the expectation on the right of (3.2.1) as follows

N1

E n 1(At N )

=

1 E N

N X

E k 1(At N )

k=1 N X k=1

k 1(At N )

Nt P (At N ):

Substituting the estimate into the right-hand side of (3.2.1) gives the required inequality. Proof of Theorem 3.2.3. It is obvious that part I implies part II. Let C -exponential tightness of fL(X ); 2 g hold. We prove that the assertion of part I holds. The argument is fairly standard. We derive part I(ii). Let us denote x;Æ = P1=r sup s;t2[0;T ]: jXt Xs j > . js tjÆ In analogy with the diagonal argument in the proof of Theorem 1.9.17 there exists a subnet fP0 ; x0 ; 0 2 0g of fP ; x;Æ ; 2 ; Æ > ld 0g such that P0 ! for some deviability and lim0 20 x0 = lim supÆ!0 lim sup2 x;Æ . Then Corollary 3.1.9 and the fact that (D n C ) = 0 yield for arbitrary Æ0 > 0

lim sup lim sup P1=r Æ!0 2

sup jXt s;t2[0;T ]: js tjÆ

Xs j >

0 0 1=r lim0 sup P0 0 sup jXt Xs j > 0 2

s;t2[0;T ]: js tjÆ0

x 2 C : sup jxt xsj :

s;t2[0;T ]: js tjÆ0

The claim follows since the right-hand side can be made arbitrarily small by choosing Æ0 in view of the -smoothness property of a deviability with respect to decreasing nets of closed sets. Part I(i) is derived similarly.

© 2001 by Chapman & Hall/CRC

280

Large deviation convergence

Thus, it remains to prove that part II implies the C -exponential tightness. We rst prove that under the hypotheses the net fL(X ); 2 g is exponentially tight in D . We apply Theorem 3.2.1. The rst condition of the theorem holds by hypotheses. We check the second. We de ne stopping times

0 = 0; k = inf ft 2 R+ : jXt X

k 1

j =2g; k 2 N :

Introducing AT = maxfk 2 N : k T g and k = k have for Æ < T by the fact that wX [k 1 ; k ) < P wT0 (X ; Æ) P AT

1; min k Æ

k 1 , we

kAT N ) + P min k Æ; 1 AT < N kAT N 1 i XX P (AT N ) + P k Æ; AT = i i=1 k=1 P(AT N ) + N 2 k=1max P k Æ; ;:::;N

P (AT

Since by Lemma 3.2.5, for N

P (AT

AT

k :

2 N,

2T N ) 2 k=1max P k ; AT k ; ;:::;N N

we obtain the estimate 2T P wT0 (X ; Æ) 2 max P k ; AT k k=1;:::;N N 2 + N max P k Æ; AT k : (3.2.2)

k=1;:::;N

Next, by right-continuity of X

P k Æ; AT

k = P sup jX

k 1 +t tÆ P sup jX ^T +t X ^T j 2 k 1 k 1 tÆ sup P sup jX+t tÆ 2ST (F )

© 2001 by Chapman & Hall/CRC

X

k 1

j 2 ; k T

X j

: (3.2.3) 2

281

LD convergence in the Skorohod space

Similarly,

P k

2T ;A N T

k sup P sup jX+t X j 2 : 2ST (F )

t2T=N

(3.2.4)

Substituting (3.2.3) and (3.2.4) into (3.2.2) yields

P1=r wT0 (X ; Æ)

21=r sup P1=r sup jX+t X j 2 2ST (F ) + N 2=r

tÆ

sup P1=r

2ST (F )

sup jX+t t2T=N

X j

: 2

Taking on the left-hand side the limits, rstly, as 2 , then as N ! 1, and nally as Æ ! 0, checks the required condition. We now prove the C -exponential tightness. Let be deviability on D , which is an LD accumulation point for fL(X )g, and let x 2 D be a discontinuous function. We show that (x) = 0. Let t 2 R+ and > 0 be such that jxt j . Using Proposition VI.2.1 of Jacod and Shiryaev [67] (see also Liptser and Pukhalskii [78] for details), we have that if Æ > 0 is small enough, then n

fy 2 D : S (x; y) Æg y 2 D : sup jysj 2 n

y 2 D : sup jyt n

y2D

jt sjÆ : sup jys t<st+Æ

o

jt sjÆ

o ysj 4 o [n yt j 4 y2D :

jyt yt Æ j 8 o [n y 2 D : sup jys yt Æ j 8 : t Æ<st

Then by Corollary 3.1.6 (x) lim sup lim sup P1=r S (x; X ) Æ Æ!0 2 lim sup lim sup 31=r sup P1=r sup jXu+s ut Æ!0 2 0<sÆ

The latter limit equals 0 by hypotheses.

© 2001 by Chapman & Hall/CRC

Xu j

: 8

o

282

Large deviation convergence

The following analogue of the Lenglart{Rebolledo inequality will allow us to estimate the probabilities in part II(ii) \in predictable terms".

Lemma 3.2.6. Let X = (Xt ; t 2 R+ ) and Y = (Yt; t 2 R+ ) be positive processes on a stochastic basis ( ; F ; F; P ). If E (X =Y ) 1 for every F-stopping time < 1, then for every F-stopping time 1, a > 0 and b > 0,

P sup Xt a t

ab + P sup Yt > b

t

(here supt1 = supt2R+ ). Proof. We de ne the stopping time = inf ft 2 R+ : Xt ag 1: If P ( < 1) = 1, then

P (sup Xt > a) P (X^

a) P (Y^ > b) +P (X^ a; Y^ b) P (Y^ > b)+P (X^ =Y^ a=b): t

By Chebyshev's inequality

P sup Xt > a t

P (Y^ > b)+ ab P sup Yt > b + ab :

t

To obtain the required, note that

P sup Xt a = lim P sup Xt > a N !1

t

t

1 : N

If P ( < 1) < 1, then by the part just proved we have for N > 0

P sup Xt a t

P sup Xt > a N1

t

1 b = lim P sup Xt > a P sup Yt > b + : M !1 N a 1=N t t^M

Since N is arbitrary, the proof is over. The following useful fact is a direct consequence of Theorem 3.2.3.

© 2001 by Chapman & Hall/CRC

LD convergence in the Skorohod space

283

Corollary 3.2.7.

Let fX ; 2 g and fY ; 2 g be nets of stochastic processes with paths in respective Skorohod spaces D (R + ; R d ) and D (R + ; R k ) such that X and Y are de ned on a common probability space ( ; F ; P ) for every 2 . If the net fL(X ); 2 g is C (R + ; Rd )-exponentially tight and the net fL(Y ); 2 g is C (R + ; Rk )-exponentially tight, then the net fL(X ; Y ); 2 g of distributions on D (R + ; Rd Rk ) is C (R + ; R d R k )-exponentially tight.

The next two results concern methods of identifying LD limits. The following theorem presents the method of nite-dimensional distributions.

Theorem 3.2.8. Let fX ; 2 g, where X = (Xt ; t 2 R+ ), be

a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ). Let the net fL(X ); 2 g of the distributions of the X on D be C {exponentially tight. Let for all k 2 N ; and t1 < : : : < tk 2 U , where U is a dense subset of R+ , as 2 , ld L(Xt1 ; : : : ; Xtk ) ! t1 ;:::;tk ;

where t1 ;:::;tk are deviabilities on (Rd )k . Then idempotent probability on D with density (x) = inf t1 ;:::;tk 2U t1 ;:::;tk xt1 ; : : : ; xtk if x = (xt ; t 2 R+ ) 2 C ; and (x) = 0 if x = (xt ; t 2 R+ ) 2 D n C , is a deviability on D , and ld L(X ) ! : Proof. By Theorem 3.1.19 fL(X ); 2 g is LD relatively compact. Let 0 be an LD accumulation point. It suÆces to prove that 0 = . By C -exponential tightness of fL(X ); 2 g this is true on D n C . Let x 2 C . By the contraction principle (Corollary 3.1.15) we have that 0 Æ t11;:::;tk = t1 ;:::;tk . By Theorem 2.2.2 and Remark 2.2.3 0 (x) = inf t1 ;:::;tk 2U t1 ;:::;tk (xt1 ; : : : ; xtk ) = (x).

The following result, which roughly shows that LD limits in distribution of non-negative martingales are exponential maxingales, lays a foundation for the maxingale problem method of proving LD convergence. We denote E1=r = (E )1=r , where, as above, E denotes expectation with respect to P .

© 2001 by Chapman & Hall/CRC

284

Large deviation convergence

Theorem 3.2.9. Let X ; 2 ; be processes with paths in D ned on respective stochastic bases ( ; F ; F ; P ). Let the

denet fL(X ); 2 g be C {exponentially tight and be a deviability on D supported by C , which is an LD accumulation point of fL(X ); 2 g. Let M = (Mt; t 2 R+ ); 2 ; be R+ {valued martingales on ( ; F ; F ; P ) such that the net f(Mt )1=r ; 2 g is uniformly exponentially integrable relative to the net fP ; 2 g for each t 2 R+ , and let Mt (x); t 2 R+ ; x 2 D ; be an R+ {valued function, which, for every t 2 R+ , is C {continuous, Borel measurable and, if restricted to C , Ct {measurable in x. If, for every t 2 R+ , as 2 ,

(Mt )1=r Mt (X )

1=r P

! 0;

then (Mt (x); t 2 R+ ; x 2 C ) is a C-exponential maxingale on (C ; ).

Proof. Since fL(X ); 2 g is C {exponentially tight, we can asld sume, by taking a subnet if necessary, that L(X ) ! . Consider a function f (x) = g(xt1 ; : : : ; xtk ), where 0 t1 < : : : < tk and g : (Rd )k ! R+ is continuous and bounded. Since f and Mt are Borel measurable, C {continuous and (D nC ) = 0, by the contraction principle for every t 2 R+

ld L Mt(X )f (X ) ! Æh 1 ; where h : D ! R+ is de ned by h(x) = Mt (x)f (x). Since by

hypotheses and boundedness of f (Mt )1=r f (X )

Mt

1=r P (X )f (X ) !

0;

it follows by Lemma 3.1.38 that

ld L (Mt)1=r f (X ) ! Æ h 1:

(3.2.5)

By uniform exponential integrability of f(Mt )1=r ; 2 g relative to fP ; 2 g, boundedness of f and Lemma 3.1.36 we then have

lim E 1=r Mt f (X )r = sup x Æ h 1 (x) 2 x2R+ = sup Mt (x)f (x)(x): (3.2.6) x2C

© 2001 by Chapman & Hall/CRC

285

LD convergence in the Skorohod space

(The last equality is the change-of-variables formula from Theorem 1.4.6.) Now let 0 s < t and ti s; i = 1; : : : ; k. By the martingale property

E Mt f (X )r = E Ms f (X )r ; so (3.2.6) yields the maxingale property sup f (x)Mt (x)(x) = sup f (x)Ms (x)(x): x2C

x2C

The collection of functions g(xt1 ; : : : ; xtk ) such that 0 t1 < : : : < tk s and g : (Rd )k ! R+ are continuous and bounded satis es the requirements on Hs in part 2 of Lemma 2.3.5 and also generates Cs. An application of Lemma 2.3.5 shows that M is an exponential maxingale on (C ; C; ). Also taking f = 1 in (3.2.5), we obtain by Lemma 3.1.12 for a > 0 lim inf E1=r Mt 1 (Mt )1=r > a 2 sup Mt (x) 1(Mt (x) > a)(x): x2C

The latter implies, since f(Mt )1=r ; 2 g is uniformly exponentially integrable relative to fP ; 2 g, that (Mt (x); x 2 C ) is {maximable. In certain cases C -exponential tightness allows one to establish LD convergence for the locally uniform topology on D . The following result is an adaptation of Theorem 3.1.10.

Theorem 3.2.10. Let X !ld

X for the Skorohod topology, where X is a Luzin-continuous idempotent process. If the X are random variables on D relative to the locally uniform topology on D , then ld X ! X for the locally uniform topology. Proof. Since convergence in the Skorohod topology to a continuous function is equivalent to locally uniform convergence, the Skorohod and locally uniform topologies are locally equivalent at every x 2 C so Theorem 3.1.10 applies.

© 2001 by Chapman & Hall/CRC

286

Large deviation convergence

We now discuss composition and rst-passage-time mappings. If x = (x1 ; : : : ; xd ) 2 D and the component functions of y = (y1 ; : : : ; yd ) 2 D are increasing and R+ -valued, we de ne the composition x Æ y by x Æ y = ((x1yt1 ; : : : ; xdyd ); t 2 R+ ).

Lemma 3.2.11.

t

ld Let X ! X , where X is a Luzin-continuous idempotent process. Let Y be stochastic processes with paths in D , whose component processes are R+ -valued and increasing, such that 1=r P Y !

y^ 2 C . Then X Æ Y !ld X Æ y^ . Proof. Clearly, y^ is component-wise R+ -valued and increasing so ld that X Æ y^ is well de ned. By Lemma 3.1.42 (X ; Y ) ! (X; y^ ). The claim now follows by Corollary 3.1.15 and the fact that the composition map (x; y) ! x Æ y is continuous at (x; y) such that x and y are continuous, Billingsley [11], Whitt [135, Theorem 3.1]. De nition 3.2.12. Given an R+ -valued function x = (xt ; t 2 R+ ) from D (R + ; R) with sample paths that are unbounded above, the associated rst-passage-time function x( 1) = (x(t 1) ; t 2 R+ ) 2 D (R + ; R )

x(t

is de ned by

xs > tg; t 2 R+ : If x = (x1 ; : : : ; xd ) 2 D is such that the xi ; i = 1; : : : ; d; are R+ valued and unbounded above, we de ne x( 1) = (x1 ( 1) ; : : : ; xd ( 1) ). 1)

= inf fs 2 R+ :

In the next lemma c ! 1 as 2 . We also denote = (t; t 2 R+ ) and, for a vector = (1 ; : : : ; d ), we let e = (t; t 2 R+ ) and 1 = (1=1 ; : : : ; 1=d ) if has positive entries. For vectors = (1 ; : : : ; d ) and = ( 1 ; : : : ; d ), we denote = (1 1 ; : : : ; d d ). Lemma 3.2.13. Let fX ; 2 g be a net of stochastic processes with paths in D de ned on respective probability spaces ( ; F ; P ) such that the component processes are R+ -valued and unbounded above. Let X be an Rd -valued Luzin-continuous idempotent process de ned on an idempotent probability space ( ; ). 1. Let, in addition, the component idempotent processes of X be R + -valued, unbounded above and strictly increasing -a.e. If ld ld X ! X , then X ( 1) ! X ( 1) .

e

© 2001 by Chapman & Hall/CRC

287

LD convergence in the Skorohod space

2. Let Y = (Yt ; t 2 R+ ) be stochastic processes with paths 0 0 from D (R + ; Rd ) and Y be an Rd -valued Luzin-continuous 0 idempotent process. Let 2 Rd be such that ! , where is entrywise positive. Let, in addition, X0 = 0. ld If (c (X e); Y ) ! (X; Y ) in D (R + ; Rd+d0 ), then ld (c (X ( 1) 1 e); c (X 0 e); Y ) ! ( 1 XÆ 1 2 d + d ( e); X; Y ) in D (R + ; R ). Proof. Part 1 follows by Corollary 3.1.15 and the fact that the map x ! x( 1) is continuous at strictly increasing x 2 D , Whitt [135, Theorem 7.2]. We prove part 2. We rst consider the case = 1d , where 1d is ld a d-vector with unity entries. Since c (X 1d e) ! X as 2 , P

1=r

! 1, and X is Luzin-continuous, it follows that X ! 1d e 1=r ( 1) P in D so by part 1 and Lemma 3.1.41 X ! 1d e in D . By ( 1) ld Lemma 3.2.11 (c (X 1d e)ÆX ; c (X 1d e); Y ) ! (X; X; Y ). c

Since

c (X (

1)

1d e) = c (1d e

X ) Æ X (

1)

by Lemma 3.1.38 it suÆces to prove that c in D , which would follow by sup

t2[0;T ]

c jX ÆX (t 1)

+ c (X Æ X (

1)

1=r P

1d e) !

(X ÆX ( 1)

1=r P

tj

1d e) ;

! 0 ; T 2 R+ :

0

(3.2.7)

Since ( 1)

0 sup (X Æ X t t2[0;T ]

t) (X0 )+ _

sup

t2[0;X (T 1) ]

(Xt )+ ;

we have for A 2 R+ and > 0 that

P1=r ( sup c jX Æ X (t t2[0;T ]

1)

tj > ) P1=r X (T

1)

>A

+ P1=r sup c ((X0 )+ _ (Xt )+ ) > : (3.2.8) t2[0;A]

© 2001 by Chapman & Hall/CRC

288

Large deviation convergence

Since the function x ! supt2[0;A] (xt )+ is continuous at continuous x, Liptser and Shiryaev [79], and x ! x0 is continuous, by Corolld lary 3.1.15 and the LD convergence c (X 1d e) ! X lim sup P1=r sup c ((X0 )+ _ (Xt )+ ) ) 2 t2[0;A] sup X0+ _ (Xt )+ = 0 ; t2[0;A] proving that the second term on the right of (3.2.8) tends to 0 as 2 P

1=r

. The rst term goes to 0 as 2 and A ! 1 since X ( 1) ! 1d e. The limit (3.2.7) has been proved so that the claim for the case = 1d has been proved. For general the hypotheses imply by Lemma 3.2.11 that ld c (X Æ ( 1 e) 1d e); Y ) ! (X Æ ( 1 e); Y ) so by the part ld proved c (X Æ ( 1 e))( 1) 1d e); c (X Æ ( 1 e) 1d e); Y ) ! X Æ ( 1 e); X Æ ( 1 e); Y . Since (X Æ ( 1 e))( 1) 1d e =

X ( 1) 1 e , by Corollary 3.1.15 and Lemma 3.2.11 c X ( 1) ld 1 e); c (X e); Y ) ! 1 X Æ ( 1 e); X; Y .

© 2001 by Chapman & Hall/CRC

Chapter 4

The method of nite-dimensional distributions In this chapter we consider the method of nite-dimensional distributions of identifying LD accumulation points. It is best suited for studying LD convergence to idempotent processes with independent increments and is based on Theorem 3.2.8. As in the preceding section, we consider a net fX ; 2 g of stochastic processes, which for the most part are semimartingales de ned on respective stochastic bases ( ; F ; F ; P ) and having paths in D = D (R + ; Rd ). The ltrations F = (Ft ; t 2 R+ ) are assumed to be complete and rightcontinuous; E denotes expectation with respect to P . We assume as xed a net fr ; 2 g of real numbers greater than 1 converging to 1 as 2 , which is used as a rate for LD convergences below; the latter refer to the Skorohod topology. We retain the rest of the notation of Section 3.2, e.g., we write E1=r for (E )1=r . Section 4.1 formulates conditions for LD convergence in distribution in terms of convergence of the stochastic exponentials of the semimartingales, Section 4.2 gives conditions on convergence of the predictable characteristics, Sections 4.3 and 4.4 consider implications of the general results. 289 © 2001 by Chapman & Hall/CRC

290

Finite-dimensional LD convergence

4.1 Convergence of stochastic exponentials In this section we use the method of nite dimensional distributions to derive conditions for LD convergence of semimartingales in a Skorohod space in terms of convergence of the associated stochastic exponentials. We start by introducing the general setting for both this chapter and the next one. For the notions and facts from stochastic calculus used below we refer the reader to Jacod and Shiryaev [67] and Liptser and Shiryaev [79]. Let X = (Xt ; t 2 R+ ); 2 ; be Rd {valued semimartingales de ned on stochastic bases ( ; F ; F ; P ), where F = (Ft ; t 2 R + ). All the X , as well as all the processes we consider below, have paths in an appropriate Skorohod space (which is D for the X ). We recall that a Borel function h : Rd ! Rd is said to be a limiter if it is bounded and h(x) = x in a neighbourhood of the origin. Every truncation function as de ned in Jacod and Shiryaev [67] is a limiter. In the same way as it is done for truncation functions one can de ne the triplet of the predictable characteristics of a semimartingale associated with a limiter. This slight extension of the class of truncation functions is convenient for technical reasons in that it allows us to consider characteristics associated with limiters that do not vanish at in nity by contrast with truncation functions. Let h(x) be a limiter. Then 1 h (x) = h(r x) (4.1.1) r is also a limiter, and we denote by (B ; C ; ) the triplet of the predictable characteristics of X associated with h (x) (i.e., de ned as if h (x) were a truncation function). We also say that (B ; C ; ) corresponds to h(x). We recall that this is equivalent to X having the following canonical representation:

Xt = X0 +Bt +Xt;c +h (x)( )t +(x h (x))t ;

(4.1.2)

where B = (Bt ; t 2 R+ ); B0 = 0; is an Rd {valued F {predictable process with bounded variation over bounded intervals; X ;c = (Xt;c ; t 2 R+ ); X0;c = 0; is an Rd {valued continuous local martingale with respect to F that is the continuous martingale part of X ;

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

291

is the measure associated with jumps of X , i.e., ([0; t]; ) =

X

0<st

1(Xs 2 nf0g); 2 B(Rd );

is an F {predictable random measure on (R+ Rd ; B(R+ ) B(Rd )) that is the F {compensator of . We use to denote integration so that h (x) (

)

t

=

(x h (x)) t =

f (x) t =

Zt Z

0 Rd Zt Z 0 Rd Zt Z 0

h (x)( (ds; dx) (ds; dx)); (x h (x)) (ds; dx);

f (x) (ds; dx):

Rd

In analogy with earlier notation for semimaxingales we also denote

f (x) t

=

Z

f (x) (ftg; dx):

Rd

An Rdd {valued continuous process C = (Ct ; t 2 R+ ); C0 = 0; is de ned to be the F {predictable quadratic-variation process of X ;c . We also de ne the continuous part of the predictable measure of jumps by

;c(ds; dx) = 1( (fsg; R d ) = 0) (ds; dx): We consider the version (B ; C ; ) of the characteristics for which identically: Ct Cs; 0 s < t; is a symmetric positive semi-de nite d d{ matrix,

(f0g; R d ) = 0; (R+ ; f0g) = 0; (ftg; R d ) 1; (4.1.3a) (jxj2 ^ 1) t < 1; (4.1.3b) Bt = h (x) t : (4.1.3c)

© 2001 by Chapman & Hall/CRC

292

Finite-dimensional LD convergence

We recall that C and do not depend on the choice of h, while if B = (B t ; t 2 R+ ) is the rst characteristic corresponding to another limiter h(x), then

B t Bt = (h (x) h (x)) t :

(4.1.4)

Mt;Æ = Xt;c +x 1(r jxj Æ)( )t ; t 2 R+ ;

(4.1.5)

M~ t = Xt;c + h (x) ( )t ; t 2 R+ ;

(4.1.6)

Along with C , we introduce Rdd {valued processes C ;Æ = (Ct;Æ ; t 2 ~ = (C~t ; t 2 R+ ) that are the respective F { R + ); Æ > 0; and C predictable quadratic-variation processes of the locally squareintegrable martingales and and are speci ed by the equalities

Ct;Æ = Ct + x 1(r jxj Æ) 2 t X x 1(r jxj Æ) s 2 (4.1.7)

st

and

C~t = Ct + ( h (x))2 t

X

st

h (x) s 2 ; (4.1.8)

where 2 Rd . The processes C ;Æ and C~ are referred to as modi ed second characteristics. We recall that X is a special semimartingale if

jxj 1(jxj > 1) t < 1; t 2 R+ :

(4.1.9)

Then one can consider the predictable triplet of X \without truncation", i.e., assume that in (4.1.2) h (x) = x. We denote the process B corresponding to this \nontruncation" by B 0 = (B 0 t ; t 2 R+ ), so the predictable triplet without truncation is (B 0 ; C ; ). As it follows by (4.1.3c) and (4.1.4), B 0t = x t ; (4.1.10) 0 B = B + (x h (x)) : (4.1.11) t

© 2001 by Chapman & Hall/CRC

t

t

293

Convergence of stochastic exponentials

If, moreover,

jxj2 1(jxj > 1) t < 1; t 2 R+ ;

(4.1.12)

then X is said to be a locally square integrable semimartingale. In that case one can de ne \nontruncated modi ed second characteristics" C~t0 = (C~t0 ; t 2 R+ ) by X C~t0 = Ct+(x)2 t (xs)2 : (4.1.13) st The following stronger condition on plays an important role below and is further referred to as the Cramer condition: (Cr) ejxj 1(jxj > 1) t < 1; for all t 2 R+ ; 2 R+ : Under (Cr), we can de ne the stochastic cumulant 1 Gt () = B 0t + Ct+(ex 1 x)t; 2 Rd ; t 2 R+ : 2 (4.1.14) The process G () = (Gt (); t 2 R+ ) is a real-valued F {predictable process with bounded variation over bounded intervals, in particular, a semimartingale, so that we can de ne the associated stochastic (or Doleans{Dade) exponential E () = (Et (); t 2 R+ ); 2 Rd ; by

Et() = eGt ()

Y

(1+Gs ())e

0<st

Gs () :

(4.1.15)

(The right-hand side is well de ned and is a semimartingale with paths in D (R + ; R), see Liptser and Shiryaev [79, Theorem 2.4.1], Jacod and Shiryaev [67, Theorem I.4.61].) By (4.1.14), (4.1.10) and (4.1.3a) Gs () =

Z

(ex 1) (fsg; dx) > 1;

(4.1.16)

Rd

so, as we are going to see,

Et() > 0; t 2 R+ :

(4.1.17)

Therefore, we can de ne the processes Y () = (Yt (); t 2 R+ ); 2 R d ; by Y () = e(Xt X0 ) E () 1 : (4.1.18) t

© 2001 by Chapman & Hall/CRC

t

294

Finite-dimensional LD convergence

A fundamental property of the stochastic exponential is expressed by the following lemma. Lemma 4.1.1. Under the Cramer condition the process Y () = (Yt (); t 2 R+ ); 2 Rd ; is a well-de ned local martingale on ( ; F ; F ; P ). Proof. We rst check Y () is well de ned by showing that (4.1.17) holds. Since G () has bounded variation over bounded intervals, X

Next,

jGs ()j < 1:

(4.1.19)

0<st Y

0<st

(1 + Gs ())

exp 2

X

0<st

Y

jGs()j

(1 + Gs ()):

0<st: jGs()j>1=2 By (4.1.19) the product on the right has nitely many terms, which are positive, and is itself positive. We now check the local martingale property of Y (). Let for k = 2; 3; : : :

k = inf t 2 R+ : 1+Gt () < 1=k : Since by (4.1.19) G has a nite number of jumps less than 1=k 1 on a bounded interval, it follows that 1 + Gk () < 1=k. Thus, the k are F -predictable stopping times as the debuts of predictable sets whose graphs belong to the sets, see Dellacherie [34, IV-T.16], Jacod and Shiryaev [67, I.2.13]. Also k ! 1 as k ! 1. Hence, there exist F -stopping times k < k such that k ! 1 as k ! 1. It is suÆcient to show that the (Yt^k (); t 2 R+ ) are local martingales relative to F. Since k < k , we have that 1 (4.1.20) inf (1+Gt ()) : tk k P By (4.1.19) 0<st jln(1 + Gs ()j < 1, so the process ln Et () = Gt ()

© 2001 by Chapman & Hall/CRC

X

0<st

Gs ()+

X

0<st

ln(1+Gs ()) (4.1.21)

295

Convergence of stochastic exponentials

is well de ned and is a semimartingale. Let

Ut = (Xt X0 ) ln Et (): Then Yt () = exp Ut so that by the Ito formula

Yt^k () = 1 +

tZ^k

eUs

0

+

1 dUs + 2

X

0<st^k

eUs

tZ^k 0

eUs dhU c is

eUs

eUs Us ; (4.1.22)

where hU c i denotes the predictable quadratic-variation process of the continuous martingale part of U . Noting that hU c i = C and invoking the canonical decomposition of the special semimartingale X with no truncation Xt = X0 + Bt0 + Mt ; where M is a local martingale relative to F , we derive from (4.1.22) after some algebra that

Yt^k () = 1 + +

tZ^kZ 0

tZ^k 0

eUs dMs

eUs ex 1 x ( )(ds; dx)

Rd

X

0<st^k

eUs

eXs 1 + Gs ()

1 Gs (): (4.1.23)

The rst integral on the right of (4.1.23) is a stochastic integral of a locally bounded predictable process with respect to a local martingale, hence, it is a local martingale. The second integral is an integral with to a martingale measure and is a local martingale since R t R respect Us ex 1 x (ds; dx) < 1, Jacod and Shiryaev [67, e 0 Rd II.1.28]. Let us consider the sum on the right of (4.1.23). Firstly, it is

© 2001 by Chapman & Hall/CRC

296

Finite-dimensional LD convergence

absolutely convergent by (4.1.19) and (4.1.20). Secondly, by (4.1.16) X

0<st^k X

=

eUs

=

eXs 1 + Gs ()

eUs eXs

tZ^kZ 0

1

eUs

1 Gs ()

0<st^k X 0<st^k

Gs () 2 1 + Gs ()

ex

eUs

Rd

Gs () 1 + Gs ()

Gs () 1 ( )(ds; dx); 1 + Gs ()

which is a local martingale by Jacod and Shiryaev [67, II.1.28] and the fact that X

0<st^k

eUs

Gs () 2 < 1: 1 + Gs ()

Let (Gt (); t 2 R+ ; 2 Rd ) be an R-valued function, which is continuous in t, dierentiable in , and such that the increments Gt () Gs () are convex functions of 2 Rd for 0 s < t and G0 () = Gt (0) = 0. Let x0 , where x0 2 Rd , be de ned as in Section 2.7 (see (2.7.6)). We recall that it is a deviability on C by Lemma 2.8.3. We denote its extension to a deviability on D with support in C as x0 as well. Let X be the canonical idempotent process on (D ; x0 ). We state the central result of the section.

Theorem 4.1.2. Let the X satisfy the Cramer condition. 2 ,

(0) and for all T > 0 and 2 Rd (sup E )

1=r P X0 ! x0 ; 1=r

P 1 ln Et (r ) Gt () ! 0 ; tT r

sup

© 2001 by Chapman & Hall/CRC

If, as

297

Convergence of stochastic exponentials

ld then X ! X:

We begin the proof with preliminary results. The hypotheses of Theorem 4.1.2 are assumed to hold. We also assume with no loss of generality that x0 = 0. We introduce the F -stopping times ( 2 Rd )

() = inf t 2 R+ : Et (r )1=r _ Et (r ) 1=r 2eGt () or Et (2r )1=r _ Et (2r ) 1=r 2eGt (2) : (4.1.24) By condition (sup E )

lim P 1=r () t = 0; t 2 R+ ; 2 Rd : 2

(4.1.25)

Being the debut of a predictable set whose graph belongs to the set, () is an F {predictable stopping time. Thus, () is P -a.s. announced by an increasing sequence of F {stopping times. Since () > 0, there exist F {stopping times () < () such that

P1=r () +

1 r

() < 1

+ P1=r () r ; () = 1

r1 : (4.1.26)

In view of the inequality

P ( () t) P ( () t + 1)

+ P () +

1 r

(t + 1) ^ () ;

we have from (4.1.25) and (4.1.26) that

lim P1=r () t = 0; t 2 R+ : 2

(4.1.27)

Note also that by (4.1.24) and the inequality () < ()

Et^ () (r)1=r _ Et^ () (r ) 1=r < 2eGt () ; (4.1.28) Et^ () (2r )1=r _ Et^ () (2r) 1=r < 2eGt (2) : (4.1.29)

© 2001 by Chapman & Hall/CRC

298

Finite-dimensional LD convergence

Lemma 4.1.3. For all 2 Rd , the process Y () is a positive supermartingale relative to F . The process (Yt^ () (); t 2 [0; T ]) is a square integrable martingale relative to F for every T > 0 and E Yt^ () (r )2

23r er[Gt (2)+2Gt ()] ; 2 Rd ; t 2 R+ :

(4.1.30)

Proof. By Lemma 4.1.1 Y () is a positive local martingale relative to F. Hence it is a supermartingale. So we have to prove only (4.1.30). By the supermartingale property of Y () and the fact that Y0 () = 1, for all nite F -stopping times ,

E Y () 1; 2 Rd :

(4.1.31)

In view of (4.1.29) and the de nition of Y (), (4.1.31) with 2r implies that E exp 2r (Xt^ () X0 ) 2r exp r Gt (2) ; 2 Rd ; which by (4.1.28) yields

E Yt^ () (r )2

2r exp rGt (2) 22r exp 2rGt () ;

proving the lemma.

Lemma 4.1.4. Let for 0 = t0 < t1 < : : : < tk Zi () = Xti ^ () Xti 1 ^ () ; 2 Rd : n

P

o

Then, for all 1 ; : : : ; k 2 Rd , the net exp ki=1 i Zi (i ) ; 2 is uniformly exponentially integrable with respect to the P and k X 1=r lim E exp r i 2 i=1

k X Zi (i) = exp i=1

Gti (i ) Gti 1 (i ) :

Proof. Let

i =

i X j =1

j Zj (j ); i = 1; : : : ; k; 0 = 0:

© 2001 by Chapman & Hall/CRC

(4.1.32)

299

Convergence of stochastic exponentials

We rst prove that for i = 1; : : : ; k

E exp(2r i ) 22r i

i Y j =1

exp r Gtj (2j )+Gtj 1 (2j ) :

The proof of Lemma 4.1.3 implies that E exp(2r i) < view of (4.1.32) and the de nitions of Zi () and Y (), we Lemma 4.1.3 and (4.1.29) for i = 1; : : : ; k h

E exp(2r i ) j Fti

(4.1.33) 1. In have by

i

1 exp(2r i 1) exp 2r i (Xti 1 ^ (i ) X0 ) Yti 1 ^ (i ) (2r i )2r er Gti (2i ) = exp(2r i 1 )Eti 1 ^ (i ) (2r i ) 1 2r er Gti (2i ) :

Applying to the latter (4.1.29) again we deduce h

E exp(2r i ) j Fti

i

1 exp(2r i 1)22r exp r

Gti (2i ) + Gti 1 (2i ) :

This P proves (4.1.33). Uniform exponential integrability of k exp i=1 i Zi (i ) ; 2 is implied by (4.1.33) with i = k. We prove the convergence required in the lemma by proving that for i = 1; : : : ; k lim E 1=r exp(r i ) = egi ; 2

(4.1.34)

where

gi =

i X j =1

Gtj (j ) Gtj 1 (j ) ; i = 1; : : : ; k; g0 = 0;

provided (4.1.34) holds for (i 1). For Æ 2 (0; 1=2) we de ne the sets n

BÆ = ! 2 : Eti 1 ^ (i ) (r i )1=r e Gti 1 (i ) 1 Æ o or Eti ^ (i ) (r i ) 1=r eGti (i ) 1 Æ

© 2001 by Chapman & Hall/CRC

300

Finite-dimensional LD convergence

and AÆ = n BÆ. By (sup E ) and (4.1.27), as 2 ,

P1=r BÆ

! 0; P1=r AÆ ! 1 :

(4.1.35)

Applying the Cauchy-Schwarz inequality we have by (4.1.33) and (4.1.35) lim E1=r exp(r i )1(BÆ ) = 0; 2

which implies that (4.1.34) would follow from lim inf lim inf E1=r exp(r i)1(AÆ ) Æ!0 2 = lim sup lim sup E1=r exp(r i )1(AÆ ) = egi : (4.1.36) Æ!0 2

Let

RÆ = exp(r i 1 )Yti ^ (i ) (r i )Yti 1 ^ (i ) (r i ) 1 : By Lemma 4.1.3

E RÆ

= E exp(r i 1 );

(4.1.37) so by our assumption

lim E1=r RÆ = egi 1 :

(4.1.38)

2

On the other hand, (4.1.32), (4.1.37), and the de nitions of Y () and Zi() yield

RÆ = exp(r i)Eti ^ (i ) (r i )

E

1 ti 1 ^ (i ) (r i ):

(4.1.39)

Therefore, applying the Cauchy-Schwarz inequality we have in view of (4.1.28), (4.1.35) and (4.1.33) that lim2 E1=r RÆ 1(BÆ ) = 0; which by (4.1.38) obtains lim E 1=r RÆ 1(AÆ ) = egi 1 : 2

(4.1.40)

By de nition, we have that on AÆ

Eti 1 ^ (i )(r i) (1 + Æ)r exp rGti 1 (i) ; Eti ^ (i ) (ri ) 1 (1 + Æ)r exp rGti (i) :

© 2001 by Chapman & Hall/CRC

301

Convergence of stochastic exponentials

Therefore, by (4.1.39)

RÆ 1(AÆ ) exp(r i)(1 + Æ)2r exp r(Gti (i) Gti 1 (i)) 1(AÆ ): The latter implies by (4.1.40) that lim inf lim inf E1=r exp(r i )1(AÆ ) Æ!0 2 egi 1 exp Gti (i)

Gti 1 (i ) = egi :

In an analogous manner, the inequalities

Eti 1 ^ (i )(r i) (1 Æ)r exp rGti 1 (i) ; Eti^ (i )(r i) 1 (1 Æ)r exp rGti (i) yield the limit lim sup lim sup E1=r exp(r i )1(AÆ ) Æ!0 2

egi :

Limits (4.1.36) are proved. The last preliminary result needed for the proof of Theorem 4.1.2 is the following version of Theorem 3.1.31, which is proved in a similar manner.

Lemma 4.1.5. Let fK ; 2 g be a net of Rm {valued random vari-

ables de ned on respective probability spaces ( ; F ; P ) and K be an Rm -valued idempotent variable de ned on an idempotent probability space ( ; ). Let S exp( K ) be nite for and dierentiable in 2 Rm . Let there exist nets fZ (); 2 g of Rm {valued ran1=r P Z () !

dom variables de ned on ( ; F ; P ) such that 0 1=r and E exp r Z () ! S exp( K ) as 2 , and the net fexp Z () ; 2 g is uniformly exponentially integrable for ld 2 Rm . Then K ! K:

K

Now we proceed with the proof of Theorem 4.1.2 itself. Recall that x0 = 0.

© 2001 by Chapman & Hall/CRC

302

Finite-dimensional LD convergence

Proof of Theorem 4.1.2. We apply Theorem 3.2.8 to the net fL(X ); 2 g. We verify the C -exponential tightness by checking the conditions of part II of Theorem 3.2.3. We begin with condition (i). In view of (4.1.31) and the de nition of Y (), we can apply to exp r (Xt X0 ) ; t 2 R+ and E (r ) Lemma 3.2.6 to obtain for all A > 0; B > 0 and L > 0

P sup exp r (Xt tL

X0 )

erA er(B

A)

+ P sup Et (r ) er B ; 2 Rd : (4.1.41) tL Taking B > GL () + 1, we have by (sup E )

lim P 1=r sup Et (r ) er B = 0; 2 tT and then (4.1.41) yields, for 2 Rd , lim sup P1=r sup (Xt X0 ) > A tT 2 Since is arbitrary, this implies that

eB A ! 0 as A ! 1:

lim lim sup P1=r sup jXt X0 j > A = 0: A!1 2 tT 1=r P X0 !

As by hypotheses 0, we obtain (i). Turning to (ii) it is again suÆcient to prove that for all 2 R d ; 6= 0; > 0; and T > 0, lim lim sup sup P1=r sup (Xt+ X ) > = 0: Æ!0 2 2ST (F ) tÆ jj (4.1.42) By Lemma 4.1.3 and Doob's stopping theorem we have for every nite F { stopping times and such that "

#

Y (r ) E 1: Y (r ) Fixing 2 ST (F ), let for t 2 R+

Xt; = Xt+ X ; Et; () = Et+ () E ()

© 2001 by Chapman & Hall/CRC

(4.1.43) (4.1.44a) (4.1.44b)

303

Convergence of stochastic exponentials

and introduce the ltration F; = (Ft+ ; t 2 R+ ). Let be a nite F; {stopping time. Then ( + ) is an F{stopping time so that by (4.1.43) (with = + ), (4.1.44a), (4.1.44b), and the de nition of Y () we have "

E

exp(r X; ) E; (r)

#

1:

As is an arbitrary nite F; {stopping time, by Lemma 3.2.6 we conclude that for all 2 Rd ; > 0; Æ > 0, and > 0

P sup Xt; tÆ

jj er (

)jj +

P sup Et; (r )1=r tÆ

ejj : (4.1.45)

By (4.1.44b) 1 1 sup ln Et; (r ) ln E (r ) G () r tÆ r 1 + sup ln Et+ (r ) Gt+ () + sup jGt+ () G ()j: tÆ r tÆ (4.1.46)

Since Gt () is continuous in t, 1 sup jGt () Gs ()j jj 2 jt sjÆ 0t;sT +Æ for all suÆciently small Æ. Thus, (4.1.46) yields for these Æ by the fact that T

P sup Et; (r )1=r ejj tÆ

P sup r1 ln Et(r ) Gt () j4j : tT +Æ Substituting the right-hand side into (4.1.45) and using (sup E ) we obtain for = 6 0

lim sup lim sup P1=r sup Xt; Æ!0 2 tÆ jj

© 2001 by Chapman & Hall/CRC

e(

)jj :

304

Finite-dimensional LD convergence

Taking = =2 and jj ! 1, and using (4.1.44a) we arrive at (4.1.42). C {exponential tightness of fL(X )g follows. Let us check LD convergence of nite-dimensional distributions. Let 0 = t0 < t1 < : : : < tk , 1 ; : : : ; k 2 Rd , and X denote the canonical idempotent process on D . We also denote m = d k; = (1 ; : : : ; k ) 2 Rm ; K = (Xt1 X0 ; : : : ; Xtk Xtk 1 ); K = (Xt1 X0 ; : : : ; Xtk Xtk 1 ); and Z () = (Z1 (1 ); : : : ; Zk (k )); where the Zi (i ) are de ned in Lemma 4.1.4. Then by (4.1.27)

P1=r jK Z ()j > "

P1=r K 6= Z ()

k X i=1

P1=r (i ) ti

! 0:

By Theorem 2.8.5 if D is equipped with 0 , then X is an idempotent process with independent increments starting at 0 and such that S0 exp (Xt Xs ) = exp Gt () Gs () : By Lemma 4.1.4 fK ; 2 g and fZ (); 2 g; 2 Rm ; satisfy the condild tions of Lemma 4.1.5 so that K ! K: The contraction principle then implies by the de nition of K that Xt1 X0 ; : : : ; Xtk

1=r P ld X0 ! Xt1 X0 ; : : : ; Xtk X0 : Since X0 ! 0 and X0 = 0 0 -a.e., by Lemma 3.1.42 and the contraction principle ld Xt0 ; Xt1 ; : : : ; Xtk ! Xt0 ; Xt1 ; : : : ; Xtk proving the nite dimen-

sional LD convergence.

Remark 4.1.6. By Theorem 2.8.5 under x0 X is an idempotent

process with independent increments starting at x0 and having cumulant G(). In particular, X satis es \the Cramer condition" S exp jXt j < 1 for 2 R+ and t 2 R+ .

We now give two more versions of the theorem in which the X do not have to be semimartingales. Inspection of the above proof shows that the critical property of E () is the one stated in Lemma 4.1.1 that the process Y () = (Yt (); t 2 R+ ); 2 Rd ; is a local martingale. Therefore, the following extension of Theorem 4.1.2 holds. The function Gt () satis es the same conditions as above.

Theorem 4.1.7. Let X ; 2 ; be stochastic processes with paths in D de ned on respective stochastic bases ( ; F ; F ; P ). Let for © 2001 by Chapman & Hall/CRC

305

Convergence of characteristics

every 2 Rd and 2 there exist F {predictable positive pro cesses E () = Et (); t 2 R+ , E0 () = 1, such that the processes Y () = Yt (); t 2 R+ de ned by

Yt () = exp (Xt X0 )

are F {local martingales. If

1=r P X0 ! x0

1 sup r

tT

Et ()

1

and, for all T > 0 and 2 Rd ,

ln Et (r )

ld as 2 , then L(X ) !

1=r P Gt ()

! 0;

x0 :

As a consequence, we obtain the following result for processes with independent increments that are not necessarily semimartingales.

Theorem 4.1.8. Let X be processes with independent increments with paths in D such that E exp (Xt and 2 Rd . If

1=r P X0 ! x0

1 sup r

tT

X0 ) < 1 for all t 2 R+

and, for all T > 0 and 2 Rd ,

ln E exp

(Xt

X0 )

Gt ()

! 0;

x0 : Remark 4.1.9. If the X ld then L(X ) !

are semimartingales with independent increments, then the assertions of Theorems 4.1.2 and 4.1.8 coincide since the triplets(B 0 ; C ; ) are deterministic and Et () = E exp (Xt X0 ) (cf. Jacod and Shiryaev [67, II.4.15]).

4.2 Convergence of characteristics In this section we give results on LD convergence in terms of the characteristics of the semimartingales. This allows us to do without the Cramer condition; we require instead that the measure of \big

© 2001 by Chapman & Hall/CRC

306

Finite-dimensional LD convergence

jumps" be small. As in the preceding section, we consider a net fX ; 2 g of semimartingales on ( ; F ; F; P ) with predictable triplets (B ; C ; ) corresponding to a limiter h(x). We also consider as given a cumulant Gt (), which does not depend on x and is de ned as in (2.7.7) and (2.7.55) by

Gt () =

Zt

gs () ds; 2 Rd ; t 2 R+ ;

(4.2.1)

0

where 1 gs () = bs + cs + 2

+ ln 1 +

Z

(ex

Rd

Z

(ex

Rd

1)^s (dx)

1 x)s (dx)

Z

(ex

Rd

1)^s (dx) ; (4.2.2)

(Rbs ; s 2 R+ ) is an Rd -valued Lebesgue-measurable function such that t 0 jbs j ds < 1 for t 2 R + , (cs ; s 2 R+ ) is a Lebesgue-measurable function with values in the space of symmetric, positive semi-de nite d d-matrices such that Rt 0 kcs k ds < 1 for t 2 R + , (s ( ); s 2 R+ ; 2 B(Rd )) is a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for t 2 R+ and 2 R+

t (f0g) = 0; Z Rd

Z

jxj2 ^ 1 t (dx) < 1;

Rd

ejxj 1(jxj > 1) t (dx) < 1;

jxj2 ^ 1 t < 1; ejxj 1(jxj > 1) t < 1; (4.2.3) ^s( ); s 2 R+ ; 2 B(Rd ) is a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that for s 2 R+ and 2 B(Rd ) ^s ( ) s( ); ^s(Rd ) 1: (4.2.4) We also assume the following condition to hold

© 2001 by Chapman & Hall/CRC

307

Convergence of characteristics

sup ejxj ^s < 1; t 2 R+ ; 2 R+ : st By Lemma 2.8.8 condition (L1 ) implies condition (L0 ) of Corollary 2.8.7. We recall that Gt () is dierentiable in under (L0 ). Let x0 2 Rd . By Corollary 2.8.7 x0 is a deviability on D and the canonical idempotent process X is a Luzin-continuous semimaxingale with independent increments on (D ; x0 ) starting at x0 and having local characteristics (b; c; ; ^). As in Section 2.7, we denote as (B; C; ; ^) the characteristics of X associated with a limiter h(x); B 0 denotes the nontruncated rst characteristic and C~ denotes the modi ed second characteristic, which are de ned by (2.7.11), (2.7.57), and (2.7.58). To recall, (L1 )

B0 = t

Zt

0

bs ds; Bt = Bt0 +(h(x) x)t ; Ct =

Zt

cs ds;

0

and for 2 Rd

C~t = Ct + ( h(x))2 t

Zt

( h(x) ^s )2 ds:

0

Let U be a dense subset of R+ and let Cb denote the set of functions f : Rd ! Rd that are bounded, continuous and equal to 0 in a neighbourhood of the origin. For f 2 Cb , we denote f (x) = f (r x)=r . We consider the following conditions (0) (A) (a)

1=r P X0 ! x0 as 2 ; 1=r lim lim sup P ([0; t]; jxj > A)1=r > " = 0; t > 0; A!1 2 1=r 1 r jxj lim lim sup P e (r jxj > a) (jxj A) t a!1 2 r

(sup B )

1

sup jBt tT

© 2001 by Chapman & Hall/CRC

Bt j

1=r P

1

" > 0; >"

= 0; t > 0; > 0; A > 0; " > 0:

! 0 as 2 ; T > 0;

308

Finite-dimensional LD convergence

(C )

lim lim sup P1=r Æ!0 2

(C~ )

kr C~t

( ) (^ )

kr Ct;Æ Ct k > " = 0; t 2 U; " > 0; C~t k

f (x) t f (x) t 1 X f (r x) s k r 0<st

1=r P

! 0 as 2 ; t 2 U;

1=r P

! 0 as 2 ; t 2 U; f 2 Cb;

Zt

f (x) ^s k ds

1=r P

! 0 as 2 ;

0

k = 2; 3; : : : ; t 2 U; f 2 Cb : The following theorem is the main result of the section. Theorem 4.2.1. Let the limiter h(x) be continuous. If conditions ld (0), (A) + (a), (sup B ), (C ) (or (C~ )), ( ), and (^) hold, then X ! X. Remark 4.2.2. Conditions (C ) and (C~ ) equivalently require convergence of the entries of the respective matrices r Ct;Æ and r C~t to the corresponding entries of Ct . Remark 4.2.3. The ldassertion of the theorem is equivalent to the LD convergence L(X ) ! x0 . We also recall that by Lemma 2.7.12 under the hypotheses

x0 (x) = exp

Z1

0

sup x_ t gt () dt 2Rd

if x is absolutely continuous and x0 = x0 , and x0 (x) = 0 otherwise.

Remark 4.2.4. The theorem also holds if condition (a) is replaced by the following weaker condition

(a0 )

where

1=r 1 lim lim sup P j (x) ;c a!1 2 r a;A; t 1 X (x) > " = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0;

(x) = (er jxj 1) 1(r jxj > a) 1(jxj A): ja;A;

© 2001 by Chapman & Hall/CRC

309

Convergence of characteristics

Before proceeding with a proof we give a version for processes with independent increments (PII) that are not necessarily semimartingales. Let X be PII. Then, given a limiter h(x), the X admit decomposition (4.1.2), see Jacod and Shiryaev [67, II.5], where B = (Bt ; t 2 R+ ); B0 = 0; is an Rd {valued right-continuous with left limits (deterministic) function; X ;c = (Xt;c ; t 2 R+ ); X0;c = 0; is an Rd {valued continuous local martingale with respect to F that is the continuous martingale part of X ; is the measure associated with jumps of X ; is a (deterministic) measure on (R+ Rd ; B(R+ ) B(Rd )) that is the F{compensator of . Also, relations (4.1.3a) and (4.1.3c) hold. Let C = (Ct ; t 2 R+ ); C0 = 0; be the F {predictable quadraticvariation process of X ;c. Then C is a deterministic Rdd {valued continuous function such that the matrices Ct Cs are symmetric and positive semi-de nite for s t. As above, we denote by C ;Æ and C~ the F -predictable quadratic-variation processes of the respective local martingales M ;Æ and M from the respective equalities (4.1.5) and (4.1.6). As with C , the processes C ;Æ and C~ are actually deterministic matrix-valued functions. Since X is not necessarily a semimartingale, the function B might no longer have bounded variation over bounded intervals and condition (4.1.3b) is not in general satis ed; so, one cannot specify C ;Æ and C~ by the respective equalities (4.1.7) and (4.1.8). Instead, we have for 2 Rd and t 2 R+

Ct;Æ = Ct + x 1(r jxj Æ) x 1(r jxj Æ)s 2 t X + x 1(r jxj Æ) s 2 1 (fsg; R d ) (4.2.5)

0<st

and C~t = Ct + h (x) h (x) s 2 t X + h (x) s 2 1 (fsg; R d ) : (4.2.6)

0<st

© 2001 by Chapman & Hall/CRC

310

Finite-dimensional LD convergence

The right-hand sides are well de ned since by Jacod and Shiryaev [67, II.5.6] h (x)

Bs 2 t +

X

jBsj2 1 (fsg; Rd ) < 1:

0<st

(4.2.7) Formulas (4.2.5) and (4.2.6) reduce to (4.1.7) and (4.1.8) if (4.1.3b) holds. Since the characteristics of a PII are deterministic, conditions (A), (a), (a0 ), (sup B ), (C ), (C~ ), ( ), and (^ ) take the form (A)I (a)I (a0 )I

(sup B )I (C )I (C~ )I ( )I (^ )I

lim lim sup ([0; t]; jxj > A)1=r = 0; t > 0;

A!1 2

1 r jxj e 1(rjxj > a) 1(jxj A) t = 0; r t > 0; > 0; A > 0; " > 0: 1 lim lim sup j (x) t;c a!1 2 r a;A; 1 X (x) = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0;

lim lim sup

a!1 2

sup jBt Bt j ! 0 as 2 ; T > 0; tT lim lim sup kr Ct;Æ Ct k = 0; t 2 U; Æ!0 2 lim kr C~t C~t k = 0; t 2 U; 2 f (x) t f (x) t ! 0 as 2 ; t 2 U; f 2 Cb ; 1 X f (r x) s k r 0<st

Theorem 4.2.5.

Zt

0

f (x) ^s k ds ! 0 as 2 ; k = 2; 3; : : : ; t 2 U; f

2 Cb :

Let X be PII with predictable characteristics (B ; C ; ) corresponding to a continuous limiter h(x). Let conditions (0), (A)I , (a)I (or (a0 )I ), (sup B )I , (C )I (or (C~I )), ( )I , ld and (^ )I hold. Then X ! X as 2 .

© 2001 by Chapman & Hall/CRC

311

Convergence of characteristics

We prove Theorems 4.2.1 and 4.2.5 in parallel. The argument actually refers to the X being semimartingales. The modi cations needed when the X are PII are either self-evident (e.g., replacing super-exponential convergence in probability by deterministic convergence) or explicitly mentioned. The proof proceeds through a number of steps: we rst establish interconnections between the conditions of the theorem, then derive the assertions of the theorems for the case of jumps of order r 1 and nally consider the general setting.

Lemma 4.2.6. Let Z ;Æ

= (Zt;Æ ; t 2 R+ ); Z0;Æ = 0; 2 ; Æ > 0; be R+ -valued increasing processes on respective probability spaces ( ; F ; P ) and Z = (Zt ; t 2 R+ ); Z0 = 0, be a (deterministic) R+ valued increasing continuous function. If, for all t 2 U and " > 0,

lim lim sup P1=r jZt;Æ Zt j > " = 0; Æ!0 2 then this convergence is uniform so that

lim lim sup P1=r sup jZt;Æ Zt j > " = 0; T > 0; " > 0: Æ!0 2 tT Proof. The argument is standard. Let wt (Æ) denote the modulus of continuity of Z on [0; t], i.e., wt (Æ) = sup u;vt: jZu Zv j: For N 2 N , ju vjÆ we choose tNi 2 U; i = 0; : : : ; kN , such that 0 = tN0 < tN1 < : : : < tNkN 1 < T tNkN < T +1 and jtNi tNi 1 j 1=N; i = 1; : : : ; kN . Then, since the Z ;Æ and Z are increasing, for t 2 [tNi 1 ; tNi ]; i = 1; : : : ; kN , we have that

jZt;Æ Ztj jZt;Æ ZtNi 1 j _ jZt;Æ ZtNi j N N i i 1 ZtNi 1 j + jZtNi ZtNi 1 j; jZt;Æ ZtNi j _ jZt;Æ N N i 1 i and hence using that Z0;Æ = Z0 = 0

1 sup jZt;Æ Zt j max N jZt;Æ Z j + w : N T +1 N ti i N i=1;:::;k tT Since wT +1 (1=N ) ! 0 as N ! 1 by continuity of Z , for arbitrary " > 0 we have for all N large enough

P1=r

sup jZt;Æ tT

© 2001 by Chapman & Hall/CRC

Zt j > "

kN X i=1

P1=r jZt;Æ j > "=2 : N ZtN i i

312

Finite-dimensional LD convergence

The latter goes to 0 as choice of the tNi .

2 and Æ ! 0 by hypotheses and the

As a consequence, we have the following. Corollary 4.2.7. Conditions (C ), (C~ ), and ( ) are equivalent to the following respective conditions (sup C ), (sup C~ ), and (sup ). (sup C )

lim lim sup P1=r sup kr Ct;Æ Æ!0 2 tT

(sup C~ )

lim P 1=r sup kr C~t 2 tT

(sup )

sup f (x)

tT

t

Ct k > " = 0;

" > 0; T > 0;

C~t k > " = 0; " > 0; T > 0;

f (x)

P 1=r t

! 0 as 2 ; T > 0; f 2 Cb :

The second preliminary lemma shows that if condition ( ) holds, then condition (sup B ) is invariant with respect to the choice of a continuous limiter h.

Lemma 4.2.8. If condition ( ) holds, then condition (sup B ) does not depend on the choice of a continuous limiter h(x).

Proof. Let B = (Bt ; t 2 R+ ) and B = (Bt ; t 2 R+ ) be the rst characteristics of X corresponding to continuous limiters h(x) and h (x), respectively. Let Bt and Bt be the rst characteristics of X associated with h(x) and h (x), respectively. By (4.1.4), up to a P null set,

Bt = Bt +(h (x) h (x)) t ; so that by the de nitions of Bt and Bt , h (x) and h (x), up to a P -null set,

Bt Bt = (Bt Bt )+(h (x) h (x))t (h (x) h(x))t ; and the equivalence of (sup B ) and (sup B ) under ( ) follows by the equivalence of ( ) and (sup ), and the fact that h h 2 Cb . We now consider implications of conditions ( ) and (^ ).

© 2001 by Chapman & Hall/CRC

313

Convergence of characteristics

Lemma 4.2.9. Let ( ) and (^ ) hold. Then for " > 0 and t 2 U 1 1. lim lim sup P1=r f (r x) 1(r jxj > Æ) t Æ!0 2 r

f (x) t > " = 0; for all R+ -valued bounded and continuous functions f (x); x 2 R d ; such that f (x) cjxj2 ; c > 0; in a neighbourhood of the origin; Æ 2. lim lim sup P1=r j g(r x)j 1(r jxj > Æ) t > " = 0 Æ!0 2 r and 1 X 3. lim lim sup P1=r g(r x) 1(r jxj > Æ) s k Æ!0 2 r 0<st Zt

g(x) ^s k ds > " = 0;

0

k = 2; 3; : : : ; for all R-valued bounded and continuous functions g(x); x 2 Rd ; such that jg(x)j cjxj; c > 0; in a neighbourhood of the origin. Proof. Let

fr (x) =

jxj 1+ ^ 1; r > 0; x 2 Rd : r

Then, since f (x) 0 and fÆ=2 (x) 1(jxj > Æ) fÆ (x), 1 r

f (r x) 1

(r jxj > Æ) t

f (x)

t

f (x) 1(jxj Æ) t + f (x)(fÆ=2 (x) fÆ (x)) t 1 + max f (r x)fÆ=i (r x) t f (x)fÆ=i (x) t : i=1;2 r

The last term on the right goes to 0 super-exponentially in probability as 2 by ( ) and the inclusion f (x)fr (x) 2 Cb . The sum of the two other terms does not exceed, for Æ small enough, 2cjxj2 1(jxj 2Æ) t ; which goes to 0 as Æ ! 0 by (4.2.3) and Lebesgue's dominated convergence theorem. Part 1 is proved.

© 2001 by Chapman & Hall/CRC

314

Finite-dimensional LD convergence

Now we prove part 2. By ( ), using that 1(jxj > Æ) and g(x)fÆ=2 (x) 2 Cb ,

fÆ=2 (x)

Æ lim sup P1=r jg(r x)j 1(r jxj > Æ) t > " r 2 lim sup P1=r rÆ jg(r x)jfÆ=2 (rx) t > " 2 1 Æjg(x)jfÆ=2 (x) t > "=2 : (4.2.8)

Now, for > Æ=2, since fÆ=2 (x) 1(jxj > Æ=2),

Æjg(x)jfÆ=2 (x) t Æjg(x)j 1(jxj > ) t + Æjg(x)j 1(Æ=2 < jxj ) t : The rst term on the right, obviously, goes to 0 as Æ ! 0. The second one, by the assumptions on g(x) and with the use of Chebyshev's inequality, is not greater than (take small enough) 2jg(x)jjxj 1(jxj ) t 2cjxj2 1(jxj ) t ; and goes to 0 as ! 0, as above. Thus, the right-hand side of (4.2.8) is zero for Æ small enough and part 2 is proved. We prove part 3. By (^ ) and the inclusion g(x)fÆ (x) 2 Cb , the required would follow by lim lim sup P1=r Æ!0 2

k 1 X g (r x) 1(r jxj > Æ ) s r 0<st k " g(r x)fÆ (r x) s > = 0 (4.2.9) 2

and Zt lim Æ!0 0

g(x) ^s

k

g(x)fÆ (x) ^s

k ds

= 0:

The validity of the latter limit is obvious since g(x) and fÆ (x) are uniformly bounded, fÆ (x) ! 1 as Æ ! 0 for x 6= 0, and (4.2.4) holds. For (4.2.9), we write, by the inequalities jxk yk j k(x _ y)k 1 jx

© 2001 by Chapman & Hall/CRC

315

Convergence of characteristics

yj; x; y > 0, and 1(jxj > 2Æ) fÆ (x) 1(jxj > Æ), k k 1 X g ( r x ) 1 ( r j x j > Æ ) g ( r x ) f ( r x ) Æ s s r 0<st X jg(r x)j 1(r jxj > Æ) s k 1 rk 0<st jg(r x)j 1(r jxj 2Æ) s : (4.2.10)

Applying to the rst integral on the right of (4.2.10) Jensen's inequality and recalling that (fsg; R d ) 1; we conclude that, for Æ small enough, the right-hand side of (4.2.10) is not greater than 2Æc

k X jg(r x)jk r 0<st

1

1(r jxj > Æ) s ;

and an application of the assertion of part 2 yields (4.2.9). Part 3 is proved.

Lemma 4.2.10. Under ( ) and (^), conditions (C ) and (C~ ) are equivalent.

Proof. By the de nitions of Ct;Æ , C~t, C~t , and h (x), it suÆces to prove that for t 2 U; " > 0, 1 lim lim sup P1=r ( h(r x))2 1(r jxj > Æ) t Æ!0 2 r

( h(x))2 t > " = 0;

Æ X jh(r x)j 1(r jxj > Æ) s > " = 0; r 0<st 1 X h(r x) 1(r jxj > Æ) s 2 lim lim sup P1=r Æ!0 2 r 0<st

lim lim sup P1=r Æ!0 2

Zt

h(x) ^s 2 ds > " = 0:

0

The limits follow by the respective parts 1, 2 and 3 of Lemma 4.2.9.

© 2001 by Chapman & Hall/CRC

316

Finite-dimensional LD convergence

4.2.1 The case of small jumps In this subsection we prove Theorems 4.2.1 and 4.2.5 for the case of jump size of order 1=r .

Theorem 4.2.11. Let the X be semimartingales (respectively, PII). Let conditions (0), (sup B ), (C ) (or (C~ )), ( ), and (^) (respectively, (0), (sup B )I , (C )I (or (C~ )I ), ( )I , and (^ )I ) hold, and, in addition, for some a > 0,

([0; t]; fr jxj > ag) = 0; t > 0; 2 :

( F )

ld Then X ! X as 2 .

Proof. We prove the theorem by checking the hypotheses of Theorem 4.1.2 in the semimartingale case, respectively, Theorem 4.1.8 in the PII case. By condition ( F ) the Cramer condition (Cr) is met by the X so that the associated stochastic exponentials E () if the X are semimartingales, respectively, the E exp (Xt X0 ) if the X are PII, are well de ned. For economy of notation we denote the latter expectation by Et () in the PII case as well. Since also the cumulant Gt () satis es the conditions of Theorem 4.1.2, by Theorem 4.1.2 (respectively, Theorem 4.1.8) in order to prove Theorem 4.2.11 it is suÆcient to check that as 2 1 ln Et (r ) Gt () tT r

sup

(sup E )

1=r P

! 0:

(In the PII case the convergence is deterministic.) Let us rstly note that by conditions ( F ) and ( )

t (jxj > a) = 0 (a.e.)

(4.2.11)

We choose h(x) = x for jxj a so that by ( F ) we have B = B 0 . In view of (4.1.14), (4.1.15) and (4.1.16) we can write

Et() = exp Bt + 21 Ct + (ex 1 x) t;c Y e Bs 1 + (ex 1) s : (4.2.12) st

© 2001 by Chapman & Hall/CRC

317

Convergence of characteristics

(In the PII case one needs to use the argument of the proofs of Jacod and Shiryaev [67, Theorems II.4.15 and II.5.2].) We show that the right-hand side of (4.2.12) is well de ned. Since ejxj 1 jxj t;c < 1; 2 R+ ; (4.2.13) all the individual terms are well de ned. To show that the product is convergent, note that Bs = x s so that e Bs 1+(ex 1) = 1+ e(x Bs ) (x B ) 1 s Bs

+ Bs

s s d (fsg; R ) :

+ e 1 1 Thus, the expression on the left-hand side is not less than 1. Also, by the inequalities 0 exp(u) 1 u exp(juj)juj2 =2; u 2 R; and jBsj a=r , condition ( F ) and the choice of h(x) X

ln e Bs 1 + (ex

0<st X 0<st

1) s

j2j e2jjar 2

jh (x) Bsj2 s + jBsj2 1 (fsg; Rd ) ;

the latter sum being convergent by (4.1.3b) (respectively, by (4.2.7)). Let us denote for s 2 R+ as = (fsg; R d ); (4.2.14) and for Æ > 0, 2 Rd , x;Æ = x 1(r jxj Æ) s ; (4.2.15a) s ;Æ x Ds () = (e 1) 1(r jxj > Æ) s ; (4.2.15b) F (well de ned by ( )), ;Æ Rs () = exp( (x 1(r jxj Æ) x;Æ s )) 1 (x 1(r jxj Æ) x;Æ (4.2.15c) s ) s ; ;Æ ;Æ ;Æ Qs () = (exp( xs ) 1 + xs )(1 as );(4.2.15d) ;Æ ;Æ ;Æ G;Æ (4.2.15e) s () = exp( xs )Ds () + Rs () ;Æ +Qs (); (4.2.15f) ;Æ ;c x Ut () = (e 1 x) 1(r jxjÆ) t (4.2.15g) (well de ned by (4.2.13)); ;Æ Vt () = (ex 1 x) 1(r jxj>Æ) t (4.2.15h) F (well de ned by ( )):

© 2001 by Chapman & Hall/CRC

318

Finite-dimensional LD convergence

For the sequel, we put down the following obvious relations: 0 as 1; jx;Æ s j

Æ r

(4.2.16)

and

exp

jja 1 D;Æ () exp jja 1 s r r

(4.2.17)

(use ( F ) and (4.1.3a)).

Lemma 4.2.12. The following representation holds (LS )

ln Et () = Bt + Vt;Æ () + Yt;Æ () + Zt;Æ ();

where

Yt;Æ () =

Zt;Æ () =

X

0<st

ln(1+ Ds;Æ ()) Ds;Æ () ;

1 ;Æ ln(1 + G;Æ s ()) + Ut () + 2 Ct 0<st X

X

0<st

ln(1 + Ds;Æ ()):

Proof. The key is to observe that ;Æ 1+(ex 1)s = exp(x;Æ s )(1+ Gs ());

(4.2.18)

which follows by routine calculations using (4.2.14){(4.2.15e). Substituting the right-hand side into (4.2.12) and taking into account (4.1.10), (4.2.15g) and (4.2.15h) yields

Et() = exp Bt + 12 Ct + Ut;Æ () + Vt;Æ () X

0<st

© 2001 by Chapman & Hall/CRC

Y

Ds;Æ ()

st

(1 + G;Æ s ()); (4.2.19)

319

Convergence of characteristics

which is equivalent to (LS ) provided the right-hand sides of (4.2.19) and (LS ) are well de ned, i.e., X

jDs;Æ ()j < 1;

(4.2.20a)

1 + G;Æ s () > 0; ;Æ j ln(1 + Gs ())j < 1;

(4.2.20b) (4.2.20c)

1 + Ds;Æ () > 0; j ln(1 + Ds;Æ ())j < 1:

(4.2.20d) (4.2.20e)

0<st

X

0<st X

0<st

Inequality (4.2.20a) follows by (4.2.15b), ( F ) and the fact that by (4.1.3b) ([0; t]; fjxj > "g) < 1; " > 0. Inequality (4.2.20b) follows since by (4.2.18), ( F ) and (4.2.16) 1 + G;Æ s () exp

jjÆ=r 1 + exp( jja=r ) 1 as exp jj(Æ + a)=r : (4.2.21)

Next, by (4.2.15e) and (4.2.16) X

0<st

jjÆ=r jG;Æ s ()j e

X

jDs;Æ ()j

0<st X

+

(Rs;Æ () + Q;Æ s ()) (4.2.22)

0<st

(note that Rs;Æ () 0 and Q;Æ s () 0). By (4.2.15c) and (4.2.15d), using (4.2.16) and the inequality exp(u) 1 u (juj2 =2) exp(juj); u 2 R, we have X

(Rs;Æ () + Q;Æ s ())

0<st e2jjÆ=r X

2

0<st

(x 1(r jxj Æ) x;Æ s )

2

s

2 + x;Æ (1 s

as ) ;

whichPis nite by (4.1.3b). In view of (4.2.20a) we thus have that 0<st jG;Æ s ()j < 1, which implies (4.2.20c) by (4.2.21),

© 2001 by Chapman & Hall/CRC

320

Finite-dimensional LD convergence

(4.2.15e), (4.2.17), (4.2.16), and the fact that Rs;Æ () and Q;Æ s () are non-negative. Inequality (4.2.20d) follows by (4.2.15b). Finally, inequality (4.2.20e) follows by (4.2.20a) and the left-hand side of (4.2.17). Now we give a similar representation for Gt (). Let Vt () = (ex 1 x) t ; (4.2.23)

Yt () =

Zt

ln 1 + (ex

1) ^s

(ex

1) ^s ds; (4.2.24)

0

1 Zt () = Ct : (4.2.25) 2 By (4.2.11) and the choice of h(x) (recall we take h(x) = x; jxj a) (LS )

Gt () = Bt + Vt () + Yt () + Zt ():

Decompositions (LS ) and (LS ) show that (sup E ) would follow if for every T > 0 and " > 0

)

sup jBt tT

Bt j

1=r P

! 0 as 2 ;

1 lim lim sup P1=r sup j Vt;Æ (r ) Vt ()j > " = 0; Æ!0 2 tT r 1

) lim lim sup P1=r sup j Yt;Æ (r ) Yt ()j > " = 0; Æ!0 2 tT r 1 Æ) lim lim sup P1=r sup j Zt;Æ (r ) Zt ()j > " = 0: Æ!0 2 tT r Part ) is just condition (sup B ). By (4.2.15h), (4.2.23), ( F ), and (4.2.11), part 1 of Lemma 4.2.9 yields

)

1 lim lim sup P1=r j Vt;Æ (r ) Vt ()j > " = 0; t 2 U; " > 0; Æ!0 2 r and an application of Lemma 4.2.6 proves part ). We now prove part ). Let

(x) = x ln(1+ x); x > 1:

© 2001 by Chapman & Hall/CRC

(4.2.26)

321

Convergence of characteristics

By the de nitions of Yt;Æ () and Yt () (see Lemma 4.2.12 and (4.2.24))

Yt;Æ () = Zt

Yt () =

X

0<st

(Ds;Æ ());

(4.2.27)

(ex

1) ^s ds:

(4.2.28)

0

Since (x) > 0, an application of Lemma 4.2.6 implies that ) would follow by 1 lim lim sup P1=r j Yt;Æ (r ) Yt ()j > " = 0; t 2 U; " > 0: Æ!0 2 r (4.2.29) Let u = exp( jja) 1 and v = exp(jja) 1. Since the function (x)=x2 is continuous on [u; v], by Weierstrass' theorem it can uniformly be approximated on [u; v] by polynomials, so, given arbitrary > 0, there exists a polynomial q (x) with powers not less than 2 such that j (x) q (x)j < x2 ; x 2 [u; v]: Now, by (4.2.17) Ds;Æ (r ) 2 [u; v], and by (4.2.11) and (4.2.4) (ex 1) ^s 2 [u; v] (a.e.) Thus, recalling (4.2.27) and (4.2.28),

P1=r

j r1 Yt;Æ (r) Yt()j > "

1=r 1 X P q Ds;Æ (r ) r 0<st Zt " q (ex 1) ^s ds > 3 0 1 X " + P1=r Ds;Æ (r )2 > r 0<st 3

+

1

Zt

0

(ex

1) ^s 2 ds >

" ; 3

and since can be taken arbitrarily small and the smallest power in

© 2001 by Chapman & Hall/CRC

322

Finite-dimensional LD convergence

q is not less than 2, (4.2.29) is implied by 1 X lim lim sup P1=r D;Æ (r )k Æ!0 2 r 0<st s Zt

(ex

0

lim lim sup lim sup P1=r 2

A!1 Æ!0

1) ^s k ds > = 0;

> 0; t 2 U; k = 2; 3; : : : ; (4.2.30a) 1 X ;Æ Ds (r )2 > A r 0<st = 0: (4.2.30b)

Limit (4.2.30a) follows by part 3 of Lemma 4.2.9 in view of (4.2.15b), ( F ) and (4.2.11). The lim sup2 in (4.2.30b) being by (^ ) not R t x 2 greater than 1 0 (je 1j ^s ) ds > A=2 for all Æ > 0, equals 0 for all large A. Limit (4.2.29) is proved. Part ) is proved. We prove part Æ). Let us denote 1 ;Æ 2 L;Æ s () = 2 (x 1(r jxj Æ ) xs ) s ; (4.2.31a) 1 2 (4.2.31b) Ks;Æ () = ( x;Æ s ) (1 as ); 2 ;Æ Hs;Æ () = L;Æ (4.2.31c) s () + Ks (); 1 Wt;Æ () = ( x)2 1(r jxj Æ) t;c: (4.2.31d) 2 Then by (4.1.7), (4.2.15a) and (4.2.14) X 1 ;Æ 1 Ct = Ct +Wt;Æ ()+ Hs;Æ (): (4.2.32) 2 2 0<st Hence, in view of the de nitions of Zt;Æ () (see Lemma 4.2.12) and Zt () (see (4.2.25)), and the fact that by Corollary 4.2.7 condition (sup C ) holds, Æ) would follow from 1 Æ0 ) lim lim sup P1=r sup j (Ut;Æ (r ) Wt;Æ (r ))j > " = 0; Æ!0 2 tT r 1 X Æ00 ) lim lim sup P1=r ln(1 + G;Æ s (r )) Æ!0 2 r 0<st Hs;Æ (r ) + ln(1 + Ds;Æ (r )) > " = 0:

© 2001 by Chapman & Hall/CRC

323

Convergence of characteristics

Let us note that condition (C ) implies that

lim lim sup lim sup P1=r r Ct;Æ > A = 0: A!1 Æ!0 2

(4.2.33)

For limit Æ0 ), we note that by the inequality jeu 1 u (ejuj =6)juj3 ; (4.2.15g) and (4.2.31d) ejjÆ sup jUt;Æ (r ) Wt;Æ (r )j jjÆ WT;Æ (r ): 3 tT

u2 =2j (4.2.34)

Next, since Hs;Æ () and Ct are non-negative, by (4.2.32) WT;Æ () Ct;Æ =2, so that by (4.2.33) 1 lim lim sup lim sup P1=r WT;Æ (r ) > A = 0; A!1 Æ!0 r 2

which together with (4.2.34) implies Æ0 ). We prove Æ00 ). Let us rst note that for t 2 R+ ; " > 0 and 2 Rd by part 2 of Lemma 4.2.9, ( F ) and (4.2.15b) Æ X lim lim sup P1=r j Ds;Æ (r )j > " = 0: Æ!0 2 r 0<st

(4.2.35)

Next, (4.2.15e) implies by Taylor's formula ;Æ ln(1+ G;Æ s (r )) = ln(1+ Ds (r ))+

Ts ; Fs

where ;Æ Ts = (exp( r x;Æ s ) 1)Ds (r ) +Rs;Æ (r ) + Q;Æ s (r ); ;Æ Fs = 1 + Ds (r ) + Ts ; 0 1;

(4.2.36) (4.2.37)

and thus 1 X ln(1 + G;Æ s (r )) r 0<st

© 2001 by Chapman & Hall/CRC

Hs;Æ (r ) + ln(1 + Ds;Æ (r ))

A1 + A2 + A3 ;

324

Finite-dimensional LD convergence

where

A1 =

1 X jDs;Æ (r )j j exp( r x;Æ s ) 1j r 0<st Fs

A2 A3

+Hs;Æ (r ) ; 1 X jTs j ;Æ = H (r ); r 0<st Fs s 1 X 1 = (jR;Æ (r ) L;Æ s (r )j r 0<st Fs s ;Æ +jQ;Æ s (r ) Ks (r )j)

(4.2.38a)

(4.2.38b)

00

(4.2.38c)

(for the latter equality recall (4.2.31c)). We thus prove Æ ) by proving that

lim lim sup P1=r Ai > " = 0; " > 0; i = 1; 2; 3: Æ!0 2 We begin with some estimates. Since by (4.2.16) j exp( r x;Æ ) 1j jjÆejjÆ ; s

(4.2.39)

(4.2.40)

we have, using (4.2.17), ;Æ jjÆ jja j exp( rx;Æ s ) 1jjDs (r )j jjÆe e :

(4.2.41)

From (4.2.15c) and (4.2.15d) using (4.2.16) and the inequality exp(u) 1 u (juj2 =2) exp(juj); we have 1 2 2 jjÆ jRs;Æ (r)j 2jj2 Æ2 e2jjÆ ; jQ;Æ s (r )j 2 jj Æ e ; whereafter in view of (4.2.36) and (4.2.41) jT j 4jjÆe2jja provided jjÆ 1; Æ a: (4.2.42) s

Further, the left inequality in (4.2.17) and (4.2.37) yield, in view of (4.2.42), 1 1 Fs e jja provided Æjj e 3jja ; Æ a: (4.2.43) 2 8 Besides, (4.2.31a){(4.2.31c) and (4.2.16) imply

Hs;Æ (r ) 3jj2 Æ2 :

© 2001 by Chapman & Hall/CRC

(4.2.44)

Convergence of characteristics

325

Now, (4.2.39) for i = 1 follows by (4.2.38a), (4.2.35), (4.2.43), (4.2.40), and (4.2.44). Next, since by (4.2.32) 1 X ;Æ 1 Hs (r ) r Ct;Æ ; r 0<st 2

(4.2.45)

by (4.2.42), (4.2.43) and (4.2.38b) for Æ small enough

A2 4jjÆe3jja r Ct;Æ ; and (4.2.39) for i = 2 follows by (4.2.33). Finally, let i = 3. From (4.2.15c), (4.2.31a), (4.2.15d), and (4.2.31b) analogously to (4.2.34) 2Æjj 2Æjj ;Æ jRs;Æ (r) L;Æ e Ls (r ); s (r )j 3 Æjj Æjj ;Æ ;Æ jQ;Æ e Ks (r ); s (r ) Ks (r )j 3 and hence in view of (4.2.38c), (4.2.43) and (4.2.31c) for Æ small enough

A3 2Æjje2Æjj ejja

1 X ;Æ H (r ); r 0<st s

so that (4.2.45) and (4.2.33) yield (4.2.39) for i = 3. Part Æ00 ) is proved. Limit (sup E ) and with it Theorem 4.2.11 are proved.

4.2.2 The general case In this subsection we prove Theorems 4.2.1 and 4.2.5. The proof relies heavily on the theory of weak convergence for deviabilities. Since by Lemma 4.2.10 conditions (C ) and (C~ ) are equivalent under conditions ( ) and (^ ), we assume that conditions (0), (A)+(a), (sup B), (C ), ( ), and (^ ) hold. The proof below also applies to the case where condition (a0 ) is assumed instead of condition (a). The X are either semimartingales or PII. For a > 0, we de ne the limiters

ha (x) =

a 1 ^ 1 x; ha (x) = ha (r x); x 2 Rd ; jxj r

© 2001 by Chapman & Hall/CRC

(4.2.46)

326

Finite-dimensional LD convergence

and introduce processes X ;a = (X t;a ; t 2 R+ ) and X^ ;a = (X^ t;a ; t 2 R + ) by

X t;a = X^ t;a =

X

0<st Xt

(Xs

ha (Xs ));

(4.2.47)

X t;a :

(4.2.48)

Let Pa denote the distribution of X^ ;a and let = x0 , which is the idempotent distribution of X . Let (B ;a ; C ; ) be the triplet of X corresponding to ha . Since the jumps of X^ ;a are ha (Xs ), the triplet of X^ ;a corresponding to ha is (B ;a ; C ; ;a ), where

;a ([0; t]; ) = ([0; t]; (ha ) 1 ( )); t 2 R+ ;

2 B(Rd ):

(4.2.49) The semimartingales (respectively, PII) X^ ;a will LD converge in distribution to a certain semimaxingale (respectively, a semimaxingale with independent increments) X a . We de ne the latter as having characteristics (B a ; C; a ; ^a ) relative to ha , which are de ned in analogy with the characteristics of X^ ;a in that B a is the rst characteristic of X associated with ha (x),

a ([0; t]; ) = ([0; t]; ha 1 ( )); ^ta ( ) = ^t (ha 1 ( ));

2 B(Rd ):

(4.2.50) Corollary 2.8.7 implies that exists and is a Luzin-continuous idempotent process. We denote its idempotent distribution by a and the associated cumulant by Ga () = (Gat (); t 2 R+ ); 2 Rd . The proof of Theorem 4.2.1 (respectively, Theorem 4.2.5) consists in proving that

Xa

ld a (i) Pa ! as 2 ; a > 0; 1=r ;a j > " = 0; t > 0; " > 0; (ii) alim lim sup P sup j X s !1 2 st

(iii) a

iw ! as a ! 1.

Since part (ii) implies that

lim lim sup P1=r S (X ; X^ ;a ) > " = 0; " > 0; a!1 2 ld by Lemma 3.1.37 (i) through (iii) would yield P ! as required.

© 2001 by Chapman & Hall/CRC

327

Convergence of characteristics

Lemma 4.2.13. Part (i) holds.

Proof. We show that the X^ ;a and X a satisfy conditions (0), (sup B), (C ), ( ), and (^ ). Condition (0) holds by the hypotheses of Theorem 4.2.1. Since the rst characteristics of the X^ ;a and X a associated with ha coincide with the respective rst characteristics of X and X , condition (sup B ) for X^ ;a and X a is identical to condition (sup B ) for X and X . We now check that conditions (C ) are identical. By (4.2.49) and since (ha ) 1 (x) = fxg if r jxj < a, for Æ < a

;a [0; t];

\ fr jxj Æg = [0; t]; \ frjxj Æg ; 2 B(Rd );

and hence (4.1.7) (respectively, (4.2.5)) yields Ct;a;Æ = Ct;Æ for Æ < a (with obvious notation). The claim follows. The fact that ( ) for the X^ ;a and X a is implied by ( ) for X and X follows by the equalities 1 1 f (r x) t;a = f (ha (r x)) t ; r r

f (x) ta = f (ha (x)) t ;

and the inclusion f Æ ha 2 Cb if f 2 Cb . Condition (^ ) for the X^ ;a and X a is checked similarly. Since also X^ ;a satis es ( F ) by (4.2.49) and (4.2.46), Theorem 4.2.11 yields the assertion of the lemma.

Remark 4.2.14. Note that it is here, while checking (C ), that we used the property that the limiters in (4.2.46), by contrast with truncation functions, do not vanish at in nity.

Now we proceed with a proof of (ii).

Lemma 4.2.15.

If f (x); x 2 Rd ; is an R+ -valued bounded Borel function equal to 0 in a neighbourhood of the origin, then for all > 0 and > 0

P f (x) t > e + P (ef (x) 1) t;c +

© 2001 by Chapman & Hall/CRC

X

ln 1+(ef (x) 1) s >

0<st e +P

(ef (x)

1) t > :

328

Finite-dimensional LD convergence

Proof. Let Yt = f (x) t: Then Y = (Yt ; t 2 R+ ) has bounded variation over bounded intervals, and is, therefore, a semimartingale. The associated stochastic exponential E ;Y () = (Et;Y (); t 2 R + ); 2 R ; is of the form Y

t () Et;Y () = eG;Y (1+G;Y s ())e

st

G;Y s () ;

f (x) 1) : Lemma 4.1.1 implies that the where G;Y t () = (e t process exp(Yt )=Et;Y (); t 2 R+ is a supermartingale relative to F; hence, E exp(Y)=E;Y () 1 for every nite F-stopping time . Since

ln Et;Y () = (ef (x) 1)t;c +

X

0<st

ln 1+(ef (x) 1)s ;

an application of Lemma 3.2.6 yields the left inequality. The right inequality follows since ln(1 + x) x. The next lemma proves part (ii).

Lemma 4.2.16. Both under conditions (A) + (a) and (A) + (a0 ) for every " > 0

1=r lim lim sup P sup jXs;a j > " = 0; t 2 R+ : a!1 2 st

Proof. Since by (4.2.46) and (4.2.47) X

sup jXs;a j jXs j 1(r jXsj > a); st 0<st we have, for A > 0; " > 0, P sup jXs;a j > ") P (sup jXs j > A

st

+ P

X

0<st

st

jXs j 1(r jXs j > a) 1(jXs j A) > " :

© 2001 by Chapman & Hall/CRC

(4.2.51)

329

Convergence of characteristics

By the Lenglart-Rebolledo inequality, see, e.g., Liptser and Shiryaev [79, Theorem 1.9.3], for > 0

P sup jXs j > A st

e

X

P

1(jXs j > A) 1

0<st r + P ([0; t]; fjxj

> Ag) > e

r ;

and, hence, by (A)

lim sup lim sup P1=r sup jXs j > A st A!1 2

e ! 0 as ! 1:

(4.2.52) For the second term on the right-hand side of (4.2.51), we have by Lemma 4.2.15 for > 0 and > 0

P

X

0<st

jXs j 1(r jXs j > a) 1(jXs j A) > "

exp r( ") 1 (x) ;c + 1 + P ja;A; t r r

X

0<st

(x) > ; ln 1+ ja;A; s

which yields under condition (a0 ) 1=r lim lim sup P a!1 2

X

0<st

jXsj 1(r jXsj > a)

1(jXs j A) > "

= 0: (4.2.53)

Limits (4.2.52) and (4.2.53) in view of (4.2.51) prove the claim under conditions (A) + (a0 ). Since (a) is stronger than (a0 ), the required also holds under (A) + (a). It is left to prove (iii). We use the method of nite-dimensional distributions for idempotent processes, so we prove that nitedimensional idempotent distributions of the X a converge to nite dimensional idempotent distributions of X as a ! 1 and that the net fLi (X a ); a 2 R+ g is tight. We begin the proof of the convergence of nite-dimensional distributions by checking that Gat () ! Gt () as a ! 1.

© 2001 by Chapman & Hall/CRC

330

Finite-dimensional LD convergence

Lemma 4.2.17. For t 2 R+ and 2 Rd , as a ! 1, supjGas () Gs ()j ! 0: st Proof. Let as in (4.2.26) (x) = x ln(1 + x); x > 1. By (4.2.50) and (2.7.61)

1 Gat () = Bta + Ct + (ex 2 Zt

1 ha (x)) ta

1 1) ^sa ds = Bta + Ct 2

(ex

0

+ (eha (x)

1 ha (x)) t

Zt

(ex

0

1) ^sa ds:

Also by (2.7.61) 1 Gt () = Bta + Ct + (ex 2

1 ha (x)) t Zt

(ex

1) ^s ds:

(ex

1) ^s ds :

0

Thus,

jGat () Gt ()j ejjjxj 1(jxj > a) t Z t +

(eha (x)

0

1) ^s ds

Zt

0

(4.2.54)

The right inequality in (4.2.3) implies that the rst term on the righthand side of (4.2.54) tends to 0 as a ! 1. By the fact that (x) is positive, in order to prove that the second term tends to 0 uniformly over bounded intervals by Polya's theorem it is suÆcient to check convergence to 0 for every t 2 R+ . Since ha (x) ! h(x) as a ! 1, by Lebesgue's convergence theorem, (4.2.3) and (4.2.4) dominated h ( x ) x a e 1 ^s ! e 1 ^s : Therefore, the required convergence

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

331

would follow by Lebesgue's dominated convergence theorem, (4.2.3) and (4.2.4) provided lim sup ex 1(jxj > a) ^s = 0; a!1 st ha (x) 1 ^ > 0; lim inf 1 + inf e s a!1 st which hold by condition (L1 ). We now prove (iii).

iw Lemma 4.2.18. a !

as a ! 1: Proof. Since the a and are supported by C and the topology on C coincides with the one induced by the Skorohod topology, by Corollary 1.9.7 and Remark 1.9.8 we may apply the method of nite dimensional distributions of Theorem 2.2.27. Since both the X a and X are Luzin idempotent processes with independent increments, by Lemma 1.10.8 weak convergence of nite-dimensional idempotent distributions would follow from weak convergence of one-dimensional idempotent distributions. The latter follows by Lemma 1.11.19 and Lemma 4.2.17 if we recall that Gt () is dierentiable in . We check tightness of the net fa ; a 2 R+ g by verifying the conditions of Theorem 2.2.26. Condition 1Æ is obvious since a (x) = 0 if x0 6= x0 . Let us consider condition 2Æ . We denote, for Æ > 0; T > 0 and > 0; ATÆ; = fx 2 C : sup jxt xs j > g s;t2[0;T ]: js tjÆ Let ei ; 1 i 2d; denote the vector, whose b(i + 1)=2cth entry equals 1 if i is odd and -1 if i is even, the rest of the entries being equal to 0. Denoting by I a the rate function associated with a , we have for > 0; s < t, that if ei (xt xs ) > > 0; then I a (x) (Gat (ei ) Gas (ei )): Therefore, ATÆ; x 2 C : sup max ei (xt xs ) > d s;t2[0;T ]: i=1;:::;2d js tjÆ

2d n [

i=1

x 2 C : I a (x) d

© 2001 by Chapman & Hall/CRC

o

sup (Gat (ei ) Gas (ei )) ; s;t2[0;T ]: js tjÆ

332

Finite-dimensional LD convergence

and hence a (ATÆ; ) = sup exp I a (x) x2ATÆ; i=1max exp + sup Gat (ei ) Gas (ei ) ;:::;2d d s;t2[0;T ]: js tjÆ =d e max exp sup Gat (ei ) Gt (ei ) i=1;:::;2d t2[0;T ] + sup Gt (ei ) Gs (ei ) : s;t2[0;T ]: js tjÆ

By Lemma 4.2.17 and continuity of G() we conclude that lim sup lim sup a (ATÆ; ) e =d ! 0 as ! 1: Æ!0 a!1 Tightness of fa g is proved. Part (iii) is proved. Thus all the assertions (i), (ii) and (iii) are proved and by ld Lemma 3.1.37 P ! as 2 . Theorems 4.2.1 and 4.2.5 have been proved.

4.3 Corollaries In this section we discuss conditions and implications of Theorem 4.2.1. The PII case of Theorem 4.2.5 can be considered similarly. Thus, the X are semimartingales in what follows. We start with \integrable" versions when the convergence conditions can be checked for nontruncated characteristics. Let us recall that the nontruncated modi ed second characteristic C~ 0 of X is de ned by (see (2.7.59))

C~t0 = Ct +( x)2 t

Zt

( x ^s )2 ds:

0

If the X are special semimartingales, we introduce the conditions (sup B 0 )

sup jB 0 t tT

© 2001 by Chapman & Hall/CRC

1=r

P Bt0 j ! 0 as 2 ; T > 0;

333

Corollaries

and (I1 )

lim lim sup P1=r jxj 1(r jxj > a) t > " = 0; a!1 2 t > 0; " > 0:

If the X are also locally square integrable semimartingales, we introduce the conditions

kr C~t0

(C~ 0 ) and (I2 )

C~t0 k

1=r P

! 0 as 2 ; t 2 U;

lim lim sup P1=r r jxj2 1(r jxj > a) t > " = 0; a!1 2 t > 0; " > 0:

Note that (I2 ) implies (I1 ).

Lemma 4.3.1.

1. Let the X be special semimartingales. If conditions ( ) and (I1 ) hold, then conditions (sup B ) and (sup B 0 ) are equivalent.

2. Let the X be locally square integrable semimartingales. If conditions ( ), (^ ) and (I2 ) hold, then conditions (C ) and (C~ 0 ) are equivalent. Proof. The proofs are analogous to the proofs of Lemmas 4.2.8 and 4.2.10, respectively. For part 1 we write B 0 B 0 = B Bt t

t

+ (ha (x) h (x)) t (ha (x) h(x)) t + (x ha (x)) t (x ha (x)) t ;

where ha is from (4.2.46). Since condition ( ) implies condition (sup ), the expression in the rst parentheses on the right converges as 2 super-exponentially in probability to 0 locally uniformly in t. Since

jx ha(x)j t jxj 1(r jxj a) t; jx ha (x)j t jxj 1(jxj a) t ; (x ha (x)) t converges super-exponentially in probability to 0 as 2 and a ! 1 locally uniformly in t by (I1 ) and (x ha (x)) t

© 2001 by Chapman & Hall/CRC

334

Finite-dimensional LD convergence

converges to 0 as a ! 1 locally uniformly in t by (4.2.3). Part 1 is proved. In order to prove part 2, it suÆces to show as in the proof of Lemma 4.2.10 that for t 2 U; " > 0,

lim lim sup P1=r r ( x)2 1(r jxj > Æ) t Æ!0 2 lim lim sup P1=r Æ!0 2

Æ

( x)2 t > "

= 0; jxj 1(r jxj > Æ) s > "

X

0<st

X lim lim sup P1=r r Æ!0 2 0<st Zt

= 0; x 1(r jxj > Æ) s 2

x ^s 2 ds > " = 0:

0

The rst convergence follows by the inequality, where a > Æ, r ( 1

x)2 1(rjxj > Æ) t ( x)2 t r ( ha (rx))2 1(rjxj > Æ) t ( ha (x))2 t

+ jj2 r jxj2 1(r jxj > a) t + jj2 jxj2 1(jxj > a) t ;

part 1 of Lemma 4.2.9, condition (I2 ), and (4.2.3). The second convergence follows by the inequality

Æ

X

0<st

jxj 1(r jxj > Æ) s rÆ

X

jha (r x)j 1(r jxj > Æ) s

0<st

+Æ

X

jxj 1(r jxj > a) s;

0<st

part 2 of Lemma 4.2.9 and condition (I1 ). For the third convergence, we write X r 0<st

x 1

(r jxj > Æ)s 2

© 2001 by Chapman & Hall/CRC

Zt

0

x^s 2 ds Q1 +Q2 +Q3 ;

335

Corollaries

where

1 X ha (r x) 1(r jxj > Æ) s 2 r 0<st

Q1 =

Zt

X Q2 = r 0<st X 1

r 0<st

Zt Q3 = 0

ha (x) ^s 2 ds ;

0

x 1(r jxj > Æ) s

2

ha (r x) 1(r jxj > Æ) s 2 ; 2

ha (x) ^s ds

Zt

x ^s 2 ds :

0

Quantity Q1 converges super-exponentially in probability to 0 as 2 by part 3 of Lemma 4.2.9. For Q2 we have

Q2 2jj2 r

jj

jj

2 2r

+ 2 2r

X

0<st X

0<st X 0<st

jxj 1(r jxj > a) s jxj 1(r jxj > Æ) s

jxj 1(r jxj > a) s jxj 1(a rjxj > Æ) s

jxj 1(r jxj > a)s 4jj2 r

2 X

jxj2 1(r jxj > a) s:

0<st

The latter sum goes to 0 as 2 super-exponentially in probability by (I2 ). By a similar argument,

Q3

jj

42

Zt

0

( x)2 1(jxj > a) ^s ds;

and converges to 0 as a ! 1 by (4.2.3). Part 2 is proved. The next result is a direct consequence of Theorem 4.2.1 and Lemma 4.3.1.

© 2001 by Chapman & Hall/CRC

336

Finite-dimensional LD convergence

Theorem 4.3.2.

I. Let the X be special semimartingales and condition (I1 ) hold. If conditions (0), (A) + (a), (sup B 0 ), (C ) (or (C~ ) associated with a continuous limiter), ( ), and (^) ld hold, then X ! X as 2 .

II. Let the X be locally square integrable semimartingales and condition (I2 ) hold. If conditions (0), (A)+(a), (sup B 0), (C~ 0 ), ld ( ), and (^ ) hold, then X ! X as 2 .

Remark 4.3.3. Similarly, in the statements below we can replace 0

condition (sup B ) by condition (sup B ) each time condition (I1 ) holds and replace condition (C~ ) by condition (C~ 0 ) each time condition (I2 ) holds.

We next consider the \quasi-continuous" case ^s (Rd ) = 0. It is singled out by the condition (QC )

1=r

P 1 X (fsg; fr jxj > g)2 ! 0; t > 0; > 0: r 0<st

Since (QC ) implies (^ ) with ^s (Rd ) = 0, condition (L1 ) trivially holds. We thus obtain the following corollary of Theorem 4.2.1.

Corollary 4.3.4. Let condition (QC ) hold and the limiter h(x) be continuous. If conditions (0), (A) + (a), (sup B ), (C ) (or (C~ )), and ld ( ) hold, then X ! X as 2 .

As a consequence, we derive a result on LD convergence to the Poisson idempotent process. Since the latter by Theorem 2.4.16 has characteristics Bt0 = t, Ct = 0, t ( ) = 1(1 2 ), and ^t ( ) = 0, we have the following result.

Corollary 4.3.5. Let the semimartingales X be one-dimensional

and the limiter h(x) be continuous at x = 1. Let N be a Poisson idempotent process. Let conditions (0) for x0 = 0, (A) + (a) and (QC ) hold. If

sup jBt tT

h(1)tj

lim lim sup P1=r Æ!0 2

© 2001 by Chapman & Hall/CRC

1=r P

! 0

as 2 ; T > 0;

kr Ct;Æ k > " = 0; t 2 U; " > 0;

337

Corollaries

and for all " 2 (0; 1=2), as 2 , 1=r

P 1 ([0; t]; fjr x 1j "g) ! t; t 2 U; r

1=r

P 1 ([0; t]; fr jxj > "g \ fjr x 1j > "g) ! 0; t > 0; r ld then X ! N as 2 . Proof. Let h(x) be continuous. The rst characteristic of X associated with h(x) equals h(1)t. Therefore, the rst two convergences in the statement check conditions (sup B ) and (C ). We check condition ( ). Let f 2 Cb and > 0 be arbitrary. Let > 0 be such that jf (x) f (1)j if jx 1j and f (x) = 0 if jxj . Then by the fact that f t = f (1)t jf t f j krf k [0; t]; fr jxj > "g \ fjrx 1j > "g

1 + [0; t]; fjr x 1j "g r 1 + jf (1)j [0; t]; fjr x 1j "g t ; r which implies condition ( ) by hypotheses and arbitrariness of . The stated LD convergence follows now by Corollary 4.3.4. Now let h(x) be continuous at x = 1 and h(x) be a continuous limiter such that h(1) = h(1). We denote by B the rst characteristic of X corresponding to h(x). Given > 0, we choose > 0 such that h(x) = h(x) = 0 if jxj and jh(x) h(x)j if jx 1j : Then by (4.1.4) denoting khk = supx2Rd jh(x)j and khk = supx2Rd jh(x)j

jB t h(1)tj jBt h(1)tj + h (x) h (x) t jBt h(1)tj + r1 [0; t]; fjr x 1j "g k hk + khk + [0; t]; fr jxj > "g \ fjr x 1j > "g : r

Since is arbitrary, the hypotheses imply that The claim follows by the part already proved.

© 2001 by Chapman & Hall/CRC

B t

h(1)t

1=r P

! 0.

338

Finite-dimensional LD convergence

Let us now consider the case t ( ) = 0. It is implied by the condition 1=r P 1 (MD) ([0; t]; fr jxj > g) ! 0 as 2 ; t > 0; > 0: r In the large deviation theory terminology this is the case of \moderate deviations". We recall that if = 0, then the idempotent deviability distribution of X has density 1 Z1 x0 (x) = exp 2 (x_ t bt ) ct (x_ t bt ) dt 0

if x is absolutely continuous, x0 = x0 and x_ t bt is in the range of ct (a.e.), and x0 (x) = 0 otherwise. By Theorem 2.6.26 X is the Luzin-continuous idempotent Gaussian diusion Zt

Zt

0

0

Xt = x0 + bs ds + c1s=2 W_ s ds;

(4.3.1)

where W is a Wiener idempotent process. Hence, this is a \central limit theorem" setting. Lemma 4.3.6. Let B and C~ , B and Ce be the rst and modi ed second characteristics of X corresponding to respective limiters h(x) and h(x), not necessarily continuous. If condition (MD) holds, then, as 2 , sup jBt tT

r kC~t

1=r P Bt j !

Ce

tk

0; T > 0;

1=r P

! 0; t > 0:

Proof. For B and B the claim is a direct consequence of (4.1.4), the de nition of a limiter and (MD). For C~ and Ce , we have by the de nition of modi ed second characteristics, choosing > 0 such that h(x) = h(x) = x; jxj ; that 1 r kC~t () Ce t ()k jj2 (khk2 + khk2 ) ([0; t]; fr jxj > g) r 1 X + jj2 (khk + khk)2 (fsg; fr jxj > g): r 0<st

© 2001 by Chapman & Hall/CRC

339

Corollaries

Thus, the claim follows by (MD). We introduce the conditions (sup B 0 )

sup jBt tT

0

(C0 )

kr C~t

1=r

P Bt0 j ! 0 as 2 ; T > 0;

Ct k

1=r P

! 0 as 2 ; t 2 U:

Theorem 4.2.1 and Lemma 4.3.6 yield the following.

Corollary 4.3.7. (\the LD central limit theorem") Let X be given 0 by (4.3.1). Let conditions (0), (A) + (a), (sup B0 ), (C0 ), and (MD) ld hold with some limiter h(x). Then X ! X as 2 .

If we would like to use nontruncated characteristics of the X , we could require the following Lindeberg condition (L2 )

r j

1=r P (r jxj > ) t !

j1

x2

0; t > 0; > 0;

which implies both conditions (I2 ) and (MD). Let us introduce the condition (C 0 ) 0

krC~t0

Ct k

1=r P

! 0 as 2 ; t 2 U:

We thus have the following version.

Corollary 4.3.8. Let X be given by (4.3.1).

Let the X be locally square integrable semimartingales and condition (L2 ) hold. If conld ditions (0), (A) + (a), (sup B 0 ), and (C00 ) hold, then X ! X as 2 .

Now we consider simpler versions of conditions (A) + (a) on the jumps of the X . We rst note that condition (A) can be checked by checking the condition (A0 )

([0; t]; jxj > A)1=r

1=r P

! 0 as 2 ; t > 0; 9A > 0;

and condition (a) can be checked by checking the condition

© 2001 by Chapman & Hall/CRC

340

Finite-dimensional LD convergence

1=r

P 1 r jxj e 1(r jxj > a) 1(jxj A) t ! 0 r as 2 ; t > 0; > 0; A > 0; 9a > 0:

(a0 )

Let us also note that if condition (A0 ) holds and the convergence in (a0 ) holds for every a > 0, then condition (MD) holds. The following observation comes in useful below.

Lemma 4.3.9. Condition (a) is implied by the conditions (a1 )

(a2 )

1=r 1 [0 ; t ] ; f r j x j > a g > = 0; lim lim sup P a!1 2 r t > 0; > 0;

lim lim sup P1=r a!1 2

1 r

r ZA

eu [0; t]; fr jxj > ug du > = 0;

a

t > 0; > 0; > 0; A > 0;

Proof. The claim follows since

1 r jxj e 1 1(r jxj > a) 1(jxj A) t r Z1 1 eu 1(r jxj > u)du 1(r jxj > a) 1(jxj A) t = r

r1

0 r Z A

eu [0; t]; fr jxj > ug \ fr jxj > ag du

0

r1 [0; t]; fr jxj > ag

+

1 r

ZR

eu du

0 r Z A

eu [0; t]; fr jxj > ug du;

R

where R > 0 is arbitrary. The following conditions can also be used for checking conditions (A) + (a).

© 2001 by Chapman & Hall/CRC

341

Corollaries

(V S )

lim lim sup P1=r ([0; t]; fr jxj > ag)1=r > " = 0; a!1 2 t > 0; " > 0; 1=r P 1 =r )

! 0 as 2 ; t > 0; > 0: Clearly, (V S0 ) ) (V S ) ) (A) + (a). Since also (V S0 ) ) (MD), by

(V S0 )

([0; t]; fr

jxj > g

Corollary 4.3.7 we have the following.

Corollary 4.3.10. Let X be given by (4.3.1).

Let conditions (0), ld 0 (V S0 ), (sup B0 ), and (C0 ) hold with some limiter h(x). Then X ! X as 2 .

Now we consider the case of the \classical" large deviation setting when the Cramer condition holds: ejxj 1(jxj > 1) < 1; 2 ; t > 0; > 0: t

We introduce the conditions 1=r 1 r jxj > " = 0; (Ie ) lim lim sup P e 1 ( r j x j > a ) t a!1 2 r t > 0; " > 0; > 0; (Le )

1=r

P 1 r jxj e 1(rjxj > ) t ! 0 as 2 ; r t > 0; > 0; > 0:

Condition (Le ) can be called an exponential Lindeberg condition. Obviously, (Le ) ) (Ie ) ) (A)+(a), (Ie ) ) (I2 ), and (Le ) ) (L2 ) ) (MD). By Theorem 4.3.2 the implication (Ie ) ) (I2 ) allows us to consider nontruncated characteristics under (Ie ). We thus have the following result.

Corollary 4.3.11.

Let the Cramer condition and condition (Ie ) hold. If conditions (0), (sup B 0 ), (C~ 0 ), ( ), and (^) hold, then ld X ! X as 2 .

The following is an application to point processes.

Corollary 4.3.12.

Let Xt = Nt =r , where N = (Nt ; t 2 R+ ) are one-dimensional point processes with compensators A = (At ; t 2 R+ ).

© 2001 by Chapman & Hall/CRC

342

Finite-dimensional LD convergence

a) If, as [0; 1],

2 ,

for some Lebesgue measurable function s

1 A r t

1=r P

1 X (As )k r 0<st

1=r

2

Zt

! t + s ds ; t 2 U;

P

!

Zt

0

0

ks ds ; t 2 U; k = 2; 3; : : : ;

ld then X ! X as 2 , where X is the idempotent process with independent increments with local characteristics bt = 1+t , ct = 0, t ( ) = (1 + t ) 1(1 2 ), and ^t ( ) = t 1(1 2 ). b) In particular, if

1 A r t 1 X (As )2 r 0<st ld then X !

1=r P

! t; t 2 U;

1=r P

! 0; t > 0;

N as 2 , where N is an idempotent Poisson process.

Proof. In part a) the nontruncated characteristics of X are of the form B 0 t = At =r , Ct = 0, ([0; t]; ) = 1(r 1 2 )At so that R conditions (sup B 0 ), (C~ 0 ), ( ), and (^ ) hold with Bt = t + 0t s ds, Ct = 0, and t and ^t as indicated in the statement. Part b) is a consequence of part a).

The implications (Le ) ) (L2 ) and (Le ) following version of Corollary 4.3.8.

) (A) + (a) give the

Corollary 4.3.13. Let X be given by (4.3.1). Let the Cram er con0 0 dition and condition (Le ) hold. If conditions (0), (sup B ), and (C0 ) ld hold, then X ! X as 2 .

4.4 Applications to partial-sum processes In this section we consider applications of the above results to the setting of the processes of partial sums of random variables. Let

© 2001 by Chapman & Hall/CRC

343

Applications

fin; i 2 N g; n 2 N ; be sequences of

R d -valued

random variables de ned on respective probability spaces ( n ; Fn ; Pn ) and adapted to discrete-time ltrations Fn = fFin ; i 2 N g. Let rn ! 1 as n ! 1 and bX ntc n Xt = in ; t 2 R+ ; (4.4.1) i=1 P0 where i=1 = 0. The predictable triplet (B n ; C n ; n ) (Xtn ; t R+ ) corresponding to a limiter h(x) is given by

2

bntc 1 X = En h(rn in )jFin 1 ; rn i=1 bntc X n [0; t]; = Pn (in 2 n f0gjFin 1 );

Btn

i=1

of X n =

Ctn = 0;

2 B(Rd );

where En denotes expectation with respect to Pn . We consider large deviation convergence of the X n with rate rn . Let X be a semimaxingale with characteristics (B; C; ; ^) and modi ed second characteristic C~ associated with h(x) as de ned in Section 4.2. The conditions of Theorem 4.2.1 assume the form. ntc bX 1=rn 1 =r n (A) lim lim sup Pn Pn (jin j > AjFin 1 ) > = 0; A!1 n!1 i=1 t > 0; " > 0; b nt c X 1=rn 1 rn jin j E e (a) lim lim sup P n n a!1 n!1 rn i=1 1(rnjin j > a) 1(jin j A jFin 1 > " = 0; t > 0; > 0; A > 0; " > 0; ntc 1 bX n 1 =r 0 n (a ) alim sup Pn ln 1 + En (ern ji j 1) 1(rn jin j > a) !1 lim r n!1 n i=1 1(jin j AjFin 1 > " = 0; t > 0; > 0; A > 0; " > 0; ntc bX 1=rn 1 (sup B ) sup En h(rn in )jFin 1 Bt Pn! 0 as n ! 1; tT rn i=1 T > 0;

© 2001 by Chapman & Hall/CRC

344 (C )

Finite-dimensional LD convergence

ntc bX 1 =r n lim lim sup Pn En ( in )2 rn Æ!0 n!1 i=1 n n En i (rn i Æ) in 1 2

1 j j jF

1(rnjinj Æ)jFin 1 Ct >

= 0; t 2 U; " > 0; 2 Rd ;

ntc bX 1=rn 1 n ))2 jF n lim P E ( h ( r n n n i i 1 n!1 rn i=1 En h(rn in )jFin 1 2 C~t > = 0; t 2 U; " > 0; 2 Rd ; bntc 1=rn 1 X ( ) En f (rnin )jFin 1 Pn! f (x) t as n ! 1; rn i=1 t 2 U; f 2 Cb ; Zt bntc k Pn1=rn 1 X n n (^ ) En (f (rn i )jFi 1 ) ! f (x) ^s k ds as n ! 1; rn i=1 0 k = 2; 3; : : : ; t 2 U; f 2 Cb : The following theorem is a triangular array version of Theorem 4.2.1. Theorem 4.4.1. Let X n be de ned by (4.4.1) and h(x) be a continuous limiter. If conditions (A) + (a) , (sup B ) , (C ) (or (C~ )), ld ( ) , and (^ ) hold, then X n ! X. The integrable and square integrable versions look as follows. As above, B 0 and C~ 0 denote the nontruncated rst and modi ed second characteristics of X , respectively. Theorem 4.4.2. I. Let Enjinj < 1; i 2 N . Let

(C~ )

(I1 )

bntc X 1 =r n lim lim sup P En a!1 n!1 n i=1

jinj 1(rn jinj > a)jFin 1

> " = 0; t > 0; " > 0

and

(sup B 0 )

bntc X sup En (in tT i=1

© 2001 by Chapman & Hall/CRC

1=rn

jFin 1 ) Bt0 Pn! 0

as n ! 1;

T > 0:

345

Applications

If, in addition, conditions (A) + (a) , (C ) (or (C~ ) with ld a continuous limiter), ( ) , and (^ ) hold, then X n ! X as n ! 1.

II. Let En jin j2 < 1; i 2 N . Let

(I2 )

lim lim sup P 1=rn a!1 n!1 n

rn

bX ntc i=1

En jin j2 1(rn jin j > a)jFin 1

> " = 0; t > 0; " > 0

and

(C~ 0 )

rn

bX ntc i=1

En (( in )2 jFin 1 )

En ( in jFin 1 )

2

Pn1=rn

! C~t0 as n ! 1; t 2 U; 2 Rd :

If, in addition, conditions (A) + (a) , (sup B 0) , ( ) , and ld (^ ) hold, then X n ! X as n ! 1. Conditions (QC ) and (MD) look in the triangular array setting as follows. bntc 1=rn 1 X (QC ) Pn (rn jin j > jFin 1 ) 2 Pn! 0 as n ! 1; rn i=1 t > 0; > 0; bntc 1=rn 1 X (MD) Pn (rn jin j > jFin 1 ) Pn! 0 as n ! 1; rn i=1 t > 0; > 0: The other conditions take the form bX ntc 1=rn (L2 ) rn En jin j2 1(rn jin j > )jFin 1 Pn! 0 as n ! 1; i=1

(A0 )

ntc bX i=1

t > 0; > 0; 1=rn

Pn (jin j > AjFin 1 )

© 2001 by Chapman & Hall/CRC

Pn1=rn

! 0 as n ! 1; t > 0; " > 0; 9A > 0;

346

Finite-dimensional LD convergence

(a0 )

bntc 1=rn 1 X n En ern ji j 1(rn jin j > a) 1(jin j A jFin 1 Pn! 0 rn i=1 as n ! 1; t > 0; > 0; A > 0; 9a > 0:

(a1 )

lim lim sup P 1=rn a!1 n!1 n

(a2 )

lim lim sup P 1=rn a!1 n!1 n

lim lim sup P 1=rn a!1 n!1 n

(V S )

ntc bX

(V S0 )

(Ie )

i=1

bntc 1 X Pn (rn jin j > ajFin 1 ) > = 0; rn i=1 t > 0; > 0;

bntc 1 X rn i=1

ntc bX i=1

r ZnA a

eu Pn (rn jin j > ujFin 1 )du >

= 0;

t > 0; > 0; > 0; A > 0; 1=rn

Pn (rn jin j > ajFin 1 )

> = 0;

t > 0; > 0; 1=rn

Pn (rn jin j > jFin 1 )

Pn1=rn

! 0 as n ! 1; t > 0; > 0;

lim lim sup P 1=rn a!1 n!1 n

bntc n 1 X En ern ji j 1(rn jin j > a)jFin 1 > rn i=1 = 0; t > 0;

bntc 1=rn 1 X n En ern ji j 1(rn jin j > )jFin 1 Pn! 0 as n ! 1; rn i=1 t > 0; > 0;

(Le )

(sup B00 )

(C0 )

bntc X 1=rn 1 sup En h(rn in )jFin 1 ) Bt0 Pn! 0 tT rn i=1 as n ! 1; T > 0;

bntc 1 X E (( h(rn in ))2 jFin 1 ) rn i=1 n Pn1=rn

2

En ( h(rn in )jFin 1 )

! Ct as n ! 1; t > 0; 2 Rd ;

© 2001 by Chapman & Hall/CRC

347

Applications

bntc X 0 (C0 ) rn En (( in )2 jFin 1 ) i=1

2 Pn1=rn

En ( in jFin 1 )

! Ct as n ! 1; t > 0; 2 Rd :

We now consider versions of the results of the preceding section on LD convergence in distribution to idempotent diusions. Let X be the idempotent diusion given by (4.3.1), where x0 = 0. Corollary 4.4.3. Let Enjinj2 < 1; i 20 N ; and condition (L2 ) hold. If conditions (A) + (a) , (sup B ) , and (C00 ) hold, then ld Xn ! X as n ! 1. Corollary 4.4.4. Let conditions (V S0) , (sup B00 ) and (C0 ) hold ld with some limiter h(x). Then X n ! X as n ! 1. Corollary 4.4.5. Let E exp(jin j) < 1; 2 R+ ; and condition ld (Le ) hold. If conditions (sup B 0 ) and (C00 ) hold, then X n ! X as n ! 1. We next consider an application to a typical moderate deviation setting. Theorem 4.4.6. Let i; i 2 N ; be i.i.d. Rd -valued random variables on a probability space ( ; F ; P ) such that E j1 j2 < 1 and E1 = 0, and let bntc 1 X n Xt = ; bn i=1 i where bn =n ! 0 and b2n =n ! 1 as n ! 1. If, for some v > 0,

2

lim nP (j1 j > vbn ) n=bn = 0; (4.4.2) n!1 then the X n LD converge in distribution at rate b2n =n to the Luzincontinuous idempotent diusion X = (E1 1T )1=2 W , where W is an R d -valued Wiener idempotent process. The deviability distribution of X is idempotent Gaussian and given by 1 Z1 X _ t (E1 1T ) x_ t dt ; (x) = exp x 2 0

if x is absolutely continuous, x0 = 0 and x_ t belongs to the range of E1 1T (a.e.), and X (x) = 0 otherwise.

© 2001 by Chapman & Hall/CRC

348

Finite-dimensional LD convergence

Proof. We take rn = b2n =n. It is easy to check that condition (L2 ) holds so we can apply Corollary 4.4.3. Since E1 = 0, condition (sup B 0 ) holds with Bt0 = 0. Condition (C00 ) holds with Ct = (E1 1T )t. We thus need to check conditions (A) + (a) . In view of Lemma 4.3.9 it is suÆcient to check conditions (A0 ) , (a1 ) and (a2 ) . We assume with no loss of generality that bn , n=bn and b2n =n are monotonically increasing. Condition (A0 ) has the form n=b2n

lim sup bntcP (j1 j > Abn ) n!1

=0

for some A 2 R+ and follows from (4.4.2) with A = v. Condition (a1 ) assumes the form

nbntc n lim lim sup P j j > a = 0: 1 a!1 n!1 b2n bn We actually check that

nbntc n lim P j j > v = 0: 1 n!1 b2n bn

(4.4.3)

Let integer N = N (n) be such that n bN < bN +1 : bn Then, noting that bN +1 bn =n , we have

bN (N + 1)=N 2bN by monotonicity of

nbntc n P j j > v tb2N +1 P (j1 j > vbN ) 4tb2N P (j1j > vbN ): 1 b2n bn Since N ! 1 as n ! 1, condition (4.4.2) implies that

lim 4tb2N P (j1 j > vbN ) n!1 so that

N=b2 N

lim 4tb2N P (j1 j > vbN ) = 0:

n!1

Limit (4.4.3) follows.

© 2001 by Chapman & Hall/CRC

=0

349

Applications

Condition (a2 ) assumes the form 2 Ab Z n =n

nbntc lim lim sup 2 a!1 n!1 bn

eu P j1 j >

a

u n du = 0: bn

We actually prove that for all A and a large enough

nbntc lim n!1 b2n

2 Ab Z n =n a

eu P j1 j >

u n du = 0: bn

(4.4.4)

Let A v and de ne integer L = L(n; u) by

bL

u n A, we have that u A and, hence, bL+1 > n=bn . Therefore, by the fact that as above bL+1 2bL , for u 2 [a; Ab2n =n]

nbntc u u n e P j j > teu b2L+1P (j1 j > AbL) 1 b2n bn 4teu b2LP (j1 j > AbL ): (4.4.6) Since bL+1 n=bn for u a, we have that L ! 1 as n uniformly over u a, so by (4.4.2) and the fact that A v

lim sup LP (j1 j > AbL ) n!1 ua

L=b2 L

!1

= 0:

Therefore, given arbitrary > 0, for all n large enough and u a

LP (j1 j > AbL ) e

b2L =L :

(4.4.7)

Since in the integral in (4.4.4) u Ab2n =n, it follows by (4.4.5) that bL bn , so by monotonicity L n and bL =L bn =n, which implies by (4.4.5) that

b2L L

u bL bnn bL2+1 bnn 2A :

© 2001 by Chapman & Hall/CRC

(4.4.8)

350

Finite-dimensional LD convergence

Using (4.4.7) and (4.4.8) we obtain by (4.4.6) that for n large enough and u 2 [a; Ab2n =n]

b2 nbntc u u n e P j j > 4teu L e 1 2 bn bn L

b2L =L

eu e e

b2L =(2L)

=(4A) 1 u ;

hence,

nbntc lim sup 2 bn n!1

2 Ab Z n =n

eu P

a

j1j > u bn du n

Since the latter integral converges to 0 as proved.

Z1

e

=(4A) 1 u du:

a

! 1, limit (4.4.4) is

Remark 4.4.7. As the proof shows, under the hypotheses condition

(a0 ) holds. The next result is in the same theme but considers triangular arrays of row-wise i.i.d.r.v. Theorem 4.4.8.d Let in; i 2 N ; n 2 N ; be a triangular array of row-wise i.i.d. R -valued random variables on respective probability spaces ( n ; Fn ; Pn ) such that En 1n = 0 and En j1n j2 < 1, and let bntc 1 X n n Xt = ; bn i=1 i where bn =n ! 0 and b2n =n ! 1 as n ! 1. If En 1n 1nT ! as n ! 1, where is a positive semi-de nite symmetric matrix and either

sup En j1nj2+Æ < 1 for some Æ > 0 and n

b2n n ln n

or

© 2001 by Chapman & Hall/CRC

as n ! 1; (4.4.9)

b2n ! 0 as n ! 1 n for some > 0 and 2 (0; 1], (4.4.10)

sup En exp( j1n j ) < 1 and n

!0

351

Applications

then the X n LD converge in distribution at rate b2n =n to the Luzincontinuous idempotent Gaussian diusion X = (T )1=2 W , where W is an Rd -valued Wiener idempotent process. Proof. We take rn = b2n =n and apply Corollary 4.4.3. Since the moment conditions imply (L2 ) , and conditions (sup B 0) and (C00 ) hold, we have to check conditions (A) + (a). Under (4.4.9) this is done by checking condition (V S0 ) , which takes the form n n=b2n 2 nn=bn Pn j1nj > ! 0 as n ! 1 bn

and follows by (4.4.9). Under condition (4.4.10) we check that conditions (A0 ) and (a0 ) hold. Condition (A0 ) holds since by (4.4.10)

nPn

j j > An=b2n b

n 1

n

2

nn=b2n En exp( j1nj ) n=bn exp A n=b2n ; which converges to 0 as n ! 1.

Veri cation of condition (a0 ) is a bit more intricate. Let us assume that < 1. We have for > 0, > 0, A > 0, and > 0 j n j n2 bn n bn n 1 A E exp j j 1 j j > 1 b2n n n 1 n 1 bn 2 nb2 En exp bnn j1nj 1 bnn j1nj1 < 1 bnn j1nj > n b j n j n2 b + 2 En exp n j1n j 1 n j1n j1 1 1 A : bn n n bn (4.4.11)

The rst term on the right of (4.4.11) is not greater than n2 n j 1 bn j n j > E exp j n 1 1 b2n n 2 nb2 En exp j1n j + 12 j1nj exp 21 bn n n (4.4.12)

© 2001 by Chapman & Hall/CRC

352

Finite-dimensional LD convergence

and converges to 0 as n ! 1 if < =2 by the moment condition in (4.4.10) and the assumption n=bn ! 1. We estimate the second term on the right of (4.4.11) as j n j n2 bn n bn n 1 1 A j j 1 j j 1 E exp n 1 1 b2n n n bn n =(1 ) n2 b2n b2 exp A n =(1 ) b En exp( j1n j ) ; n n which goes to 0 by (4.4.10). Thus, the right-hand side of (4.4.11) goes to 0 as n ! 1 so that condition (a0 ) is checked for < 1. If = 1, the required follows by (4.4.12).

Remark 4.4.9. If the distributions of the 1n do not depend on n,

the moment conditions above imply condition (4.4.2). Remark 4.4.10. We have actually checked that under (4.4.10) condition (a0 ) holds for every a > 0. We now consider examples on LD convergence to dierent kinds of idempotent processes. Example 4.4.11. "Very large deviations" Let X n be given by (4.4.1). Let rn = n, which speci es the set-up of \very large deviations". We assume that in = g(i=n; i )=n, where i ; i 2 N ; are Rd -valued i.i.d.r.v. on a probability space ( ; F ; P ), and g : R+ Rd ! Rd is continuous in the rst variable and such that E exp(jg(t; 1 )j) < 1 for all > 0; t > 0. It is easy to see that all the conditions of Corollary 4.3.11 hold with

B0 = t

Zt

Eg(s; 1 ) ds; Ct = 0; t ( ) = ^t ( ) = P g(t; 1 ) 2

0

It is instructive to note that the Cramer condition is not indispensable in this sort of result. Indeed, let in = in=n, where in; i 2 N ; are R + -valued random variables, i.i.d. for each n with the distribution function 1(x > n2) : P (1n x) = 1 exp( x2 ) 1(x n2 ) n2 exp( n4 ) x

© 2001 by Chapman & Hall/CRC

nf0g :

353

Applications

Then conditions (A) + (a) are easily seen to hold while neither condition (V S ) nor (Ie ) is satis ed, and even E1n = 1. The other conditions of Theorem 4.4.1 are satis ed as well with Z1 Bt = t h(x)d(1 exp( x2 )); Ct = 0; 0

t ( ) = ^t ( ) =

Z

1(x > 0)d(1

exp( x2 ));

2 B(Rd ):

Example 4.4.12. LD convergence to Poisson idempotent processes.

Let X n be given by (4.4.1). Let in = in=rn , where rn ! 1, rn =n ! 0 as n ! 1, and fin ; i 2 N g are independent r.v. assuming values 1 and 0 with respective probabilities rn =n and (1 rn =n). ld N as n ! 1 Then part b) of Corollary 4.3.12 implies that X n ! at rate rn , where N is a Poisson idempotent process. Note that here n ([0; t]; frn jxj > g)=rn ! t for < 1, so condition (MD) does not hold, while condition (QC ) does. Example 4.4.13. LD convergence of empirical processes. Let n 1 X rn n Xt = 1 i n t ; rn i=1 where i are i.i.d.r.v. with values in R+ , whose distribution admits density g(x), which is continuous and positive at 0. Also rn ! 1 and rn =n ! 0 as n ! 1. We denote by G(x) the distribution function of 1 and introduce the point process Ntn = rn Xtn . Then the compensator of N n = (Ntn ; t 2 R+ ) relative to the natural ltration is, Jacod and Shiryaev [67, II.3.32], r n Zt n g s X r r n n n ds: Ant = rn 1 i s r n i=1 n 1 G ns 0 n It is not diÆcult to check that n 1X r 1=rn 1 i n s Pn! 0 n i=1 n

© 2001 by Chapman & Hall/CRC

354

Finite-dimensional LD convergence

1=rn

as n ! 1 and hence Ant =rn Pn! tg(0): Part b) of Corollary 4.3.12 ld implies that the X n ! X as n ! 1 at rate rn , where Xt = Ng(0)t , N being an idempotent Poisson process.

© 2001 by Chapman & Hall/CRC

Chapter 5

The method of the maxingale problem The method of nite-dimensional distributions considered in Chapter 4 does not allow us to prove LD convergence in distribution to idempotent processes other than idempotent processes with independent increments. In this chapter we consider a dierent approach, which is an analogue of the martingale problem approach in weak convergence theory and consists in identifying the limit deviability as a solution to a maxingale problem. As in Chapter 4, we consider a net of semimartingales fX ; 2 g de ned on respective stochastic bases ( ; F ; F ; P ) with paths in D = D (R + ; Rd ). We assume as xed a net fr ; 2 g of real numbers greater than 1 converging to 1 as 2 . It is used as a rate for LD convergences below, which refer to the Skorohod topology. The limit semimaxingale X is assumed to be \canonical" in that it is de ned on D by Xt (x) = xt ; x 2 D ; t 2 R+ . It will actually be Luzin-continuous so that we can equivalently consider it as the canonical idempotent process on C = C (R + ; Rd ). The next two sections are concerned with identifying maxingale problems whose solutions are LD accumulution points of fL(X ); 2 g: Section 5.1 speci es the maxingale problem in terms of convergence of stochastic exponentials and assumes the Cramer condition for the X , while Section 5.2 considers convergence of the characteristics of the semimaxingales and does without the Cramer condition. Section 5.3 is devoted to speci c LD convergence results. Section 5.4 considers applications to large deviation convergence of Markov processes. 355 © 2001 by Chapman & Hall/CRC

356

Maxingale problem

5.1 Convergence of stochastic exponentials This section contains results on LD convergence of semimartingales stated in terms of convergence of the associated stochastic exponentials. Let G() = Gt (; x); t 2 R+ ; x 2 D ; 2 Rd ; be an R-valued function such that G0 (; x) = Gt (0; x) = 0, which is continuous in t and D-adapted in x. As above, we refer to G() as a cumulant. We introduce a number of conditions.

De nition 5.1.1. The function G() is said to satisfy the uniform continuity condition if the map x ! (Gt (; x); t 2 R+ ) is a C { continuous map from

D

into C (R + ; R).

De nition 5.1.2. Let F = (Ft (x); t 2 R+ ; x 2 D ); F0 (x) = 0; be an continuous function. We say that F satis es the majoration condition if there exists an R-valued, increasing and continuous function F = (F t ; t 2 R+ ); F 0 = 0; such that for all 0 s < t, R -valued

sup(Ft (x) Fs (x)) F t F s :

x2D

(5.1.1)

The function F is said to satisfy the local majoration condition if, for each b > 0, there exists an R-valued, increasing and continuous in t function F b = (F bt ; t 2 R+ ); F b0 = 0; such that, for all 0 s < t,

sup (Ft (x) Fs (x)) F bt F bs : xx12D:b

(5.1.2)

Remark 5.1.3. If F is D-adapted, then, being continuous, it is Dpredictable, so the preceding supremum may be taken over x 2 D such that xt b. More generally, if is a nite D{stopping time on D , then, for every nite D{stopping time , sup (F (x) F (x)) = sup (F (x) F (x)) x2D : x2D : x1 b

x b

(See, e.g., Jacod and Shiryaev [67, III.2.43].)

At times we require that the restriction of G() to C satisfy the linear-growth condition of De nition 2.8.11, which we recall here.

© 2001 by Chapman & Hall/CRC

357

Convergence of stochastic exponentials

De nition 5.1.4. We say that G() satis es the linear-growth con-

dition if there exist R+ {valued, increasing and continuous in t functions F l () = (Ftl (); t 2 R+ ); 2 Rd ; such that F0l () = Ftl (0) = 0 and for some R+ -valued increasing function kt we have for all 0 s < t , x 2 C and 2 Rd

Gt (; x) Gs (; x) Ftl ((1+kt xt )) Fsl ((1+kt xt )):

We also recall that a deviability on C is a solution to the maxingale problem (x0 ; G), where x0 2 Rd , if the canonical process X on C is a semimaxingale with cumulant G() on (C ; C; ) such that X0 = x0 -a.e. (De nition 2.8.1). Let the X satisfy the Cramer condition (Cr) and E () = Et (); t 2 R+ ; 2 Rd ; be the associated stochastic exponentials. The following conditions on the X are similar to those used in Section 4.1: 1=r P X0 ! x0

(0) (sup E )

as 2 ;

1=r

P 1 sup j ln Et (r ) Gt (; X )j ! 0 as 2 ; tT r 2 Rd ; T > 0:

Theorem 5.1.5. d

Let the X satisfy (Cr), and G(), for each 2 R , satisfy the uniform continuity and majoration conditions. If conditions (0) and (sup E ) hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.1.6. By the fact that a cumulant G() that does not depend on x satis es the uniform continuity and majoration conditions, Theorems 5.1.5 and 2.8.5 imply Theorem 4.1.2.

The majoration condition on G() is too restrictive in applications. We replace it next by the local majoration condition and another condition. Recall that x0 is de ned by (2.7.6) and x;t by (2.8.6). The following is the condition we will require. (NE ) The function x (x); x 2 C ; is upper compact and the sets [s2[0;t]fxs : x;s(x) ag are bounded for a 2 (0; 1] and t 2 R+ .

© 2001 by Chapman & Hall/CRC

358

Maxingale problem

Remark 5.1.7.

Condition (NE ) implies that smooth idempotent measure on C .

x0

is a tight -

Remark 5.1.8.

By Lemma 2.8.12 and Remark 2.8.13 condition (NE ) is met when G() satis es the linear-growth condition.

Let us de ne for x 2 D

N (x) = inf ft 2 R+ : xt +t N g; N

2 N:

(5.1.3)

The next version of Lemma 2.7.5, which is proved by a similar argument, implies that N is a D{stopping time and is C {continuous.

Lemma 5.1.9. Let (Ht (x); t 2 R+ ; x 2 D ) be an R+ -valued D{ adapted function, which is continuous and increasing in t and C { continuous in x. Let for c 2 R+ (x) = inf ft 2 R+ : Ht (x)+ t cg: Then (x); x 2 D ; is a D{stopping time and is C -continuous. The following condition is a localised version of (sup E ).

(sup E )loc

1=r

P 1 sup j ln Et^N (X ) (r ) Gt^N (X ) (; X )j ! 0 tT r as 2 ; 2 Rd ; T > 0; N 2 N :

Theorem 5.1.10. Let the X satisfy (Cr), G(), for each 2 Rd ,

satisfy the uniform continuity and local majoration conditions, and (NE ) hold. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.1.11. The uniform continuity and majoration conditions used above can be somewhat modi ed. Let us say that G() satis es the continuity condition for a given 2 Rd if Gt (; x) is C { continuous in x for all t from a dense subset of R+ . Let us say that an R-valued function F = (Ft (x); t 2 R+ ; x 2 D ) obeys the strict majoration condition if (5.1.1) holds with the increments on the left-hand side replaced by their absolute values. Similarly, we can de ne the local strict majoration condition by taking absolute values on the left of (5.1.2). Since the local strict majoration condition

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

359

and the continuity condition imply the uniform continuity condition for G(), we could in Theorem 5.1.5 (respectively, Theorem 5.1.10) require only the continuity condition if we strengthened the majoration condition (respectively, local majoration condition) to the strict majoration condition (respectively, local strict majoration condition).

Since the linear-growth condition on G() implies both the local majoration condition and (NE ), we obtain the following important consequence of Theorem 5.1.10.

Theorem 5.1.12. Let the X satisfy (Cr) and the cumulant G(), d

for each 2 R , satisfy the uniform continuity and linear-growth conditions. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

The proofs of Theorems 5.1.5, 5.1.10, and 5.1.12 below show that the only property of the processes E that matters, besides being positive and predictable, is that they satisfy the assertion of Lemma 4.1.1. This observation allows us, as in Chapter 4, to extend the theorems to the case when the processes X are not necessarily semimartingales if we postulate the property stated in Lemma 4.1.1. More speci cally, let us consider the following condition on processes X with paths in D de ned on stochastic bases ( ; F ; F ; P ) .

(E )

For each 2 , there exist F {predictable positive processes E () = (Et (); t 2 R+ ); 2 Rd , such that E0 () = 1 and the processes exp( (Xt X0 ))Et () 1 ; t 2 R+ are F {local martingales.

Then we have the following extension of Theorem 5.1.12.

Theorem 5.1.13. Let X ( ; F ; F ; P ) , which satisfy d

be stochastic processes on condition (E ) , and let G(), for each 2 R , satisfy the uniform continuity and lineargrowth conditions. If conditions (0) and (sup E )loc hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Theorems 5.1.5 and 5.1.10 admit similar versions.

© 2001 by Chapman & Hall/CRC

360

Maxingale problem

Remark 5.1.14. In Theorems 5.1.5, 5.1.10, 5.1.12, and 5.1.13 we

can equivalently describe the accumulation points by saying that if is an accumulation point of fL(X ); 2 g, then the canonical process X on (C ; C; ) is a Luzin-continuous semimaxingale with cumulant G() starting at x0 . Also if G() and x0 uniquely specify ld , then X ! X.

5.1.1 Proofs In the proofs below we assume with no loss of generality that x0 = 0. We start with two preliminary lemmas. We assume the conditions imposed on the X at the beginning of the section. As above we denote by ei ; i = 1; : : : ; 2d; the d-vector, whose b(i + 1)=2cth entry equals 1 if i is odd and -1 if i is even, the rest of the entries being equal to 0.

Lemma 5.1.15. For every nite F{stopping time , a > 0, b > 0, c > 0, and u 2 R+ the following inequalities hold P sup jXt+ tu

X j a

2d exp dc (b a)

1 b + 2d max P sup (ln Et+ (cei ) ln E (cei )) > i=1;:::;2d d tu c c 2d exp d (b a) 1 b + 2d max P sup (Gt+ (cei ) G (cei )) > : i=1;:::;2d d tu c

Proof. The second inequality is implied by the rst since by (4.1.15)

ln Et+ (cei ) ln E (cei ) Gt+ (cei ) G (cei ): The rst inequality results from Lemma 3.2.6. Speci cally, let

Zt; () = Yt+ ()=Y (); t 2 R+ ; where Yt () is de ned by (4.1.18). By Lemma 4.1.1 and Doob's stopping theorem Z ; () = (Zt; (); t 2 R+ ) is an R+ -valued local martingale with respect to the ltration F; = (Ft+ ; t 2 R+ );

© 2001 by Chapman & Hall/CRC

361

Convergence of stochastic exponentials

hence, EZ; () 1 for every F; {stopping time . Lemma 3.2.6 and the de nition of Y () then yield for i = 1; : : : ; 2d c a P sup ei (Xt+ X ) exp (b a) d d tu 1 b + P sup (ln Et+ (cei ) ln E (cei )) > ; d tu c hence,

P sup jXt+ tu

X j a

a 2d i=1max P sup ei (Xt+ X ) ;:::;2d d c

tu

2d exp d (b a)

1 b + 2d max P sup (ln Et+ (cei ) ln E (cei )) > : i=1;:::;2d d tu c

Next comes one of the most technically important results of the chapter. It is more general than is required at the moment for the proofs of Theorems 5.1.5, 5.1.10 and 5.1.12, but this generality will be exploited while proving Theorems 5.2.9, 5.2.12, and 5.2.15 below. Let fX 0 ; 2 g, where X 0 = (X 0 t ; t 2 R+ ), be along with X a net of Rd {valued semimartingales de ned on ( ; F ; F ; P ). We consider the pair (X ; X 0 ) as a process with paths in the Skorohod space D 0 = D (R + ; Rd Rd ). The space D 0 is equipped with the natural ow of -algebras D0 = (Dt0 ; t 2 R+ ), de ned in analogy with D, and elements of D 0 are denoted by (x; x0 ). We denote by C 0 = C (R + ; R d Rd ) the subspace of D 0 of continuous functions equipped with the - ow C0 = (Ct0 ; t 2 R+ ) as de ned in Section 3.2. For (x; x0 ) 2 D 0 and 2 Rd , we introduce Y 0 (; (x; x0 )) = exp(xt Gt (; x0 )); t 2 R+ ; (5.1.4) t

and let

Y 0 () = (Yt0 (; (x; x0 )); t 2 R+ ; (x; x0 ) 2 C 0 ):

(5.1.5)

Deviability 0 on C 0 is said to be a solution of maxingale problem (M 0 ) if

© 2001 by Chapman & Hall/CRC

362

Maxingale problem

x00 = 0

(M 0 )

Y

(); 2 Rd ;

0 { a.e.; is a C0 {local exponential maxingale on (C 0 ; 0 ):

Theorem 5.1.16. Let G(0 ) satisfy the uniform continuity condi0 tion. If the net fL((X ; X )); 2 g is C {exponentially tight, and conditions (0) and 1=r

1 P 0 sup j ln Et (r ) Gt (; X )j ! 0 as 2 ; tT r T > 0; 2 Rd ; hold, then every LD accumulation point of fL((X ; X 0 )); 2 g (when restricted to C 0 ) is a solution to (M 0 ). Proof. Let 0 be an LD accumulation point of fL((X ; X 0 )); 2 ld g. To simplify notation, we assume that L((X ; X 0 )) ! 0 as 2 : By C 0 {exponential tightness of fL((X ; X 0 )); 2 g the deviability 0 is supported by C 0 , so it can be considered as a deviability on C 0 . We show that 0 ((x; x0 ) : x0 6= 0) = 0. Since the map 0 : (x; x0 ) ! x0 from D 0 into Rd is continuous, by the contraction principle

(sup E )0

ld 0 L(X0 ) ! Æ0 1 as 2 ;

and then by (0) and the de nition of LD convergence 0 Æ 1 (x) = 1(x = 0); x 2 Rd ; 0

which is equivalent to the required. Now we prove that the Y 0 (); 2 Rd ; are C0 {local exponential maxingales on (C 0 ; 0 ). We do that by reduction to Theorem 3.2.9. As above we denote Gt (; x0 ) = supst jGs (; x0 )j: By the uniform continuity condition on G() the function Gt (; x0 ) is C {continuous in x0 2 D for each t 2 R+ . For N 2 N and x0 2 D we introduce (x0 ) = inf ft 2 R+ : G (; x0 )_G (2; x0 )+t N g: (5.1.6) N

t

t

By Lemma 5.1.9 N (x0 ); x0 2 D ; is a nite D{stopping time and is C {continuous. Therefore, N , as a function on D 0 , is a D0 {stopping time and is C 0 {continuous.

© 2001 by Chapman & Hall/CRC

363

Convergence of stochastic exponentials

Let also, for N

2 N and 2 ;

N = inf t 2 R+ : Et (r )1=r _ Et (r ) 1=r _ Et (2r)1=r _ Et(2r ) 1=r 2eN : (5.1.7) Then by F {predictability and right continuity of Et (), N is an F{predictable stopping time (see Dellacherie [34, IV.T.16]), and by (5.1.6) and (sup E )0 lim P 1=r N 2

N (X 0 ) = 0:

The facts that N is F {predictable and N > 0 P {a.s. (since E0() = 1) imply as in the proof of Theorem 4.1.2 that there exist nite F {stopping times N such that

N < N P {a.s.

(5.1.8)

and lim P 1=r N 2

N (X 0 ) = 0:

(5.1.9)

Note that by (5.1.7) and (5.1.8) for t 2 R+ P -a.s.

Et^ (r) _ Et^ (r) 1 _ Et^ (2r ) _ Et^ (2r ) N

N

N

1

N

< 2r er N : (5.1.10)

Now, since Lemma 4.1.1 implies that Y () is a supermartingale so EY () 1 for every nite F {stopping time , we have by (5.1.10) and the de nition of Y () in (4.1.18) that

E Yt^ (r )2 = E Yt^ (2r )Et^ (2r )Et^ (r ) N N N N 23r e3Nr :

2

Thus, in view of Doob's stopping theorem, (Yt^ (r ); t 2 R+ ) is a N square-integrable martingale and for every F {stopping time

E1=r Y^ (r )2 N

© 2001 by Chapman & Hall/CRC

8e3N :

(5.1.11)

364

Maxingale problem

Next, by the respective de nitions (4.1.18) and (5.1.4) of Yt () and Yt0 (; (x; x0 )), and the inequality jeu 1j jujejuj ; u 2 R; we have that for A > 0; " > 0; > 0, and T > 0

P sup Yt0^ (; (X ; X 0 )) Yt^ (r )1=r > " N N tT P sup jjjXt j > A + P (jX0 j > ) tT

" + P sup Et^ (r ) 1=r > e A (jj) 1 e jj 2 N tT " +P sup exp( Gt^ (; X 0 )) Et^ (r ) 1=r > e A : N 2 N tT (5.1.12)

We prove that the right-hand side converges super-exponentially to 0 as 2 . By C 0 {exponential tightness of fL((X ; X 0 )); 2 g and Theorem 3.2.3 lim lim sup P1=r (sup jjjXt j > A) = 0; tT

(5.1.13)

lim P 1=r (jX0 j > ) = 0; 2

(5.1.14)

A!1 2

by (0)

and by (5.1.10)

lim lim sup P1=r sup Et^ (r ) !0 2 N tT

1=r

" > e A (jj) 1 ejj 2 = 0: (5.1.15)

Finally, (5.1.10) and (sup E )0 are easily seen to imply that

lim P1=r sup exp( Gt^ (; X 0 )) N 2 tT

Et^ (r)

" > e 2

N

A

1=r

= 0: (5.1.16)

By (5.1.13){(5.1.16) the right-hand side of (5.1.12) raised to the power of 1=r goes to 0 in the limit limA!1 lim sup!0 lim sup2 .

© 2001 by Chapman & Hall/CRC

365

Convergence of stochastic exponentials

Thus, we have proved that as 2 sup Yt0^ (; (X ; X 0 )) N tT

1=r P 1 =r Yt^ (r ) ! N

0; T > 0:

As a consequence, introducing

N = N ^ N (X 0 );

(5.1.17)

we have that for t 2 R+ 1=r P 0 0 1 =r Yt^ (; (X ; X )) Yt^ (r ) ! 0 as 2 ; N N

hence, since by (5.1.9) and (5.1.17)

lim P1=r N (X 0 ) 6= N = 0; 2 we arrive at the convergence

Yt0^ (X 0 ) (; (X ; X 0 )) Yt^ (r )1=r N N

1=r P

! 0 as 2 :

(5.1.18) Now we check the conditions of Theorem 3.2.9 with D 0 as D , (X ; X 0 ) as X , YN () = (Yt^ (r ); t 2 R+ ) as M , (x; x0 ) as x, N and Yt0^ (x0 ) (; (x; x0 )) as Mt (x). N The net fL((X ; X 0 )); 2 g is C 0 {exponentially tight by hypotheses. Next, since N (x0 ) is a D{stopping time and X 0 is F { adapted, N (X 0 ) is an F {stopping time; since N also is an F { stopping time, we conclude in view of (5.1.17) that N is an F { stopping time. Therefore, recalling that (Yt^ (r ); t 2 R+ ) is N

a square-integrable martingale with respect to F and N N , we have that YN () is a square-integrable martingale too. Moreover, in view of (5.1.11), the net fYt^ (r )1=r ; 2 g is uniN formly exponentially integrable relative to fP g for all t 2 R+ . Also Yt0^ (x0 ) (; (x; x0 )) is C 0 {continuous by the fact that, as we remarked earlier, N (x0 ) is C 0 {continuous, and Yt0 (; (x; x0 )) is continuous in

© 2001 by Chapman & Hall/CRC

366

Maxingale problem

(t; (x; x0 )) at (x; x0 ) 2 C 0 by the uniform continuity condition on G() and (5.1.4). Finally, since N (x0 ) is a D0 {stopping time, it is a C0 {stopping time if restricted to C 0 . Therefore, Yt0^ (x0 ) (; (x; x0 )) is Ct0 { measurable by Lemma 2.2.19. Since also (5.1.18) holds, we conclude that the conditions of Theorem 3.2.9 are met with the above choice of M , X and Mt (x). The theorem implies that the function (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) is a C0 {exponential maxingale on N (C 0 ; 0 ). 0 {uniform maximability of (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) is N proved in analogy with (5.1.11). Since by (5.1.6) and continuity of Gt (; x) in t we have that jGt^ (x0 ) (; x0 )j N and N jGt^N (x0 ) (2; x0 )j N; it follows by (5.1.4) that

xx

xx

sup Y 0 0 (; ( ; 0 ))2 0 (( ; 0 )) (x;x0 )2C 0 t^N (x ) e3N sup Yt0^ (x0 ) (2; ( ; N (x;x0 )2C 0

x x0 ))0 ((x; x0 )) = e3N ;

where the latter equality holds by the maxingale property of (Yt0^ (x0 ) (2; (x; x0 )); t 2 R+ ). Thus, (Yt0^ (x0 ) (; (x; x0 )); t 2 R+ ) N N is 0 {uniformly maximable by Corollary 1.4.15. We are now in a position to prove Theorem 5.1.5. Proof of Theorem 5.1.5. We apply Theorem 5.1.16 with X 0 = X . All we need to prove is that under the conditions of Theorem 5.1.5 the net fL((X ; X )); 2 g is C 0 {exponentially tight in D 0 or, equivalently, the net fL(X ); 2 g is C {exponentially tight in D . We check the C {exponential tightness by verifying the conditions of part II of Theorem 3.2.3. This is carried out similarly to the argument used in the proof of Theorem 4.2.11 with the use of Lemma 5.1.15. We consider only condition II(ii) of Theorem 3.2.3, because II(i) is checked in an analogous manner. By Lemma 5.1.15 for T > 0; > 0; c > 0; 0 < Æ < 1, and 2

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

367

ST (F) P sup jXt+

X j >

tÆ

2d exp cr 2d

1 ln Et+ (r cei ) ln E (r cei ) + 2d max P sup i=1;:::;2d 2d tÆ cr c 1 ln Et (r cei ) ln Es (r cei ) 2d i=1max P sup ;:::;2d 2d s;tT +1 r sts+Æ cr + 2d exp : (5.1.19) 2d

Applying successively (sup E ) and the majoration condition on G(), we have for i = 1; : : : ; 2d 1 c sup ln Et (r cei ) ln Es (r cei ) 2d s;tT +1 r sts+Æ lim sup P1=r sup Gt (cei ; X ) Gs (cei ; X ) 3cd 2 s;tT +1 sts+Æ 1 sup Git Gis 3cd ; s;tT +1 sts+Æ

lim sup P1=r 2

where Gi = (Git ; t 2 R+ ) is a function majorising G(cei ). By continuity of Git in t the latter indicator is 0 for all small Æ > 0. Hence, by (5.1.19) c lim sup lim sup sup P1=r sup jXt+ X j > exp : 2d Æ!0 2 2ST (F ) tÆ Since c is arbitrarily large, condition II(ii) of Theorem 3.2.3 has been checked.

For a proof of Theorem 5.1.10, we need another auxiliary result which will also be used in the proof of Theorem 5.2.12. Let the maps p~N : D ! D ; N 2 N ; be de ned by (~pN x)t = xt^N (x) ; x 2 D ; t 2 R+ ;

© 2001 by Chapman & Hall/CRC

(5.1.20)

368

Maxingale problem

where the N are from (5.1.3). The maps p~N are C {continuous since the N are C {continuous and Skorohod convergence to continuous functions is equivalent to locally uniform convergence. Let also

X ;N = p~N X

(5.1.21)

and Y N () = (YtN (; x); t 2 R+ ; x 2 C ) be de ned by

YtN (; x) = Yt^N (x) (; x):

Let for N (M N )

(5.1.22)

2 N maxingale problems (M N ) on C be de ned by

x0N= 0

Y ();

2 Rd ;

N {a.e.; is a C-local exponential maxingale on (C ; N ):

Lemma 5.1.17.

Let the nets fL(X ;N ); 2 g; N 2 N ; be C {exponentially tight and every LD accumulation point of fL(X ;N ); 2 g solve (M N ). If, in addition, (NE ) holds, then

lim lim sup P1=r N (X ) t = 0; t 2 R+ : N !1 2 Proof. We rst note that since N solves (M N ), an argument similar to the one used in the proof of Lemma 2.7.11 shows that

N (x) 0;N (x) (x):

(5.1.23)

Let P1Æ=rÆ N Æ (X Æ ) t ; 2 be a subnet of 1=r P N (X ) t ; (; N ) 2 N such that lim P1Æ=rÆ N Æ (X Æ ) t 2 = lim sup lim sup P1=r (N (X ) t): (5.1.24) N !1 2 By Corollary 3.1.20 there exists a subnet f(L(X Æ Æ;N ); N 2 N ); 2 g of f(L(X Æ ;N ); N 2 N ); 2 g such that ld Æ Æ ;N N L(X ) ! ; N 2 N ; as 2 at rate rÆ Æ , where N are deviabilities on D with support in C . Since N (D n C ) = 0, we

© 2001 by Chapman & Hall/CRC

Convergence of stochastic exponentials

369

identify N with its restriction to C . By (5.1.3) and Lemma 5.1.9 N (x); x 2 C ; is a nite C{stopping time so that the -algebra CN is well de ned. We prove that there exists a deviability on C such that (A) = N (A); A 2 CN ; N

2 N:

(5.1.25)

This is done by applying Theorem 1.8.1. Let us check that fN ; N 2 N g is a projective system of deviabilities on the same space C with the p~N as \bonding maps". In other words, we have to check that 0 ld N = 0N Æ p~N1 for N 0 > N . Since L(X Æ Æ;N 0 ) ! N 0 , X ;N = p~N X ;N (see (5.1.20) and (5.1.21)), and p~N is C {continuous, by ld the contraction principle L(X Æ Æ;N ) ! N 0 Æ p~N1 . Since also ld N L(X Æ Æ;N ) ! , by uniqueness of an LD limit N = N 0 Æp~N1 . In order to apply Theorem 1.8.1 we have to check the (; K )condition. By the rst part of condition (NE ) it is suÆcient to check that KN 0 () p~N 0 K0 () for arbitrary 2 (0; 1] and N 0 2 N . 0 Let xN 2 KN 0 (). The fact that fN ; N 2 N g is a projective system of deviabilities allows us to construct functions xN 2 C0 ; N 0= N 0 ; N 0 + 1; : : : ; such that p~N xN +1 = xN and N (xN ) = N (xN ). Since the sequence fN (xN )g is increasing, it converges to a limit L. Since by (5.1.23) 0;N (xN ) (xN ) N (xN ) and the sequence fxNN (xN ) + N (xN ); N = N 0 ; N 0 + 1; : : :g is unbounded, the second part of condition (NE ) implies that L0 = 1. Hence, there exists a 0 function x^ 2 C that coincides with xN on [0; N 0 (xN )] and coincides with the xN on [N 1 (xN ); N (xN )] for N = N 0 +1; N 0 +2; : : :. Since 0;N (^x) (^x) = 0;N (xN )(xN ) and 0;N (^x)(^x) ! 0 (^x) as N ! 1, we conclude that 0(^x) as required. Hence, by Theorem 1.8.1 there exists a deviability on C such that N = Æ p~N1 , which is equivalent to (5.1.25) by Lemma 2.2.21. Since N (X ) = N (X ;N ), N Æ Æ ! 1 as 2 , the set fx 2 D : N (x) tg is C {closed, N (D nC ) = 0, and fx 2 C : N (x) tg 2 CN , we have by (5.1.24), Corollary 3.1.9, and (5.1.25) that for arbitrary N 0 2 N lim sup lim sup P1=r (N (X ) t) N !1 2 = lim P1Æ=rÆÆ Æ N Æ Æ (X Æ Æ ) t 2

© 2001 by Chapman & Hall/CRC

370

Maxingale problem

lim sup P1Æ=rÆÆ Æ (N 0 (X Æ 2

Æ;N 0 ) t) N 0 (N 0 (x) t)

= (N 0 (x) t): (5.1.26) By the -smoothness property of deviability lim (N (x) t) = N !1

\

N 2N

fx 2 C : N (x) tg = 0;

which together with (5.1.26) proves the lemma. Proof of Theorem 5.1.10. We begin by showing that the nets fL(X ;N ); 2 g; N 2 N ; de ned in (5.1.21), are C {exponentially tight and their respective LD accumulation points solve (M N ). We rst check that fX ;N ; 2 g satis es the conditions of Theorem 5.1.5 with GN () = (Gt^N (x) (; x); t 2 R+ ; x 2 D ) as G(). Condition (0) is obvious. Next, X ;N has as its stochastic exponential the process E ;N () = (Et^N (X ) (); t 2 R+ ). Hence, (sup E )loc implies (sup E ) for X ;N with GN () as G(). Now we check that GN () satis es the conditions imposed in Theorem 5.1.5 on G(). Let us consider the majoration condition. Since by the de nition of N , x(t^N (x)) N; x 2 D , we have that, for 0 s t,

sup(Gt^N (x) (; x) Gs^N (x) (; x))

x2D

=

sup

x2D : (t^N ( )) N

x

(Gt^N (x) (; x) Gs^N (x) (; x)): (5.1.27)

x

By Remark 5.1.3 and the facts that G() is D{adapted and t ^ N (x) is a D{stopping time, the right-hand side of (5.1.27) equals sup(Gt^N (x) (; x) Gs^N (x) (; x)) over x 2 D such that x1 N ; so by the local majoration condition on G() (say, with GN for given ) if x1 N , then

Gt^N (x) (; x) Gs^N (x) (; x) GNt^N (x)

GNs^N (x) GNt GNs ;

where for the last inequality we used that GNt is increasing in t. Hence, sup(Gt^N (x) (; x) Gs^N (x) (; x)) GNt GNs ; x2D

© 2001 by Chapman & Hall/CRC

(5.1.28)

371

Convergence of stochastic exponentials

proving the majoration condition for GN (). Next, obviously, Gt^N (x) (; x) is Dt {measurable in x 2 D and continuous in t. We check that it is C {continuous in x uniformly over t 2 [0; T ] for arbitrary T > 0. Let xn ! x 2 C . We again x and denote by GN the associated local majorant for G(). For arbitrary " > 0, by continuity of GNt and Gt (; x) in t we can choose Æ > 0, Æ < N (x) ^ 1, such that sup jGNu u;vT ju vjÆ

GNv j "; sup jGt^(N (x) tT

Since N (xn ) ! N (x) as n large enough to have

Æ) (;

x)

Gt^N (x) (; x)j ":

! 1 by Lemma 2.7.5, we can take n

jN (xn) N (x)j Æ; and then, for t T ,

(5.1.29)

Gt^N (xn ) (; xn ) Gt^N (x) (; x) (Gt^N (xn )(; xn ) Gt^(N (x) Æ) (; xn )) + jGt^(N (x) Æ) (; xn ) Gt^(N (x) Æ) (; x)j + jGt^(N (x) Æ) (; x) Gt^N (x) (; x)j (Gt^N (xn )(; xn ) Gt^(N (x) Æ) (; xn )) + jGt^(N (x) Æ) (; xn ) Gt^(N (x) Æ) (; x)j + ": (5.1.30) By (5.1.29), (5.1.28) and the choice of Æ

Gt^N (xn ) (; xn ) Gt^(N (x) Æ) (; xn ) = Gt^N (xn ) (; xn ) Gt^(N (x) Æ)^N (xn ) (; xn ) GNt^N (xn ) GNt^(N (x)

": Therefore, (5.1.30) yields by C {continuity of the mapping x ! (Gt (; x); t 2 R+ ) lim sup sup(Gt^N (xn ) (; xn ) Gt^N (x) (; x)) 2": n!1 tT

The complementary inequality

lim sup sup(Gt^N (x) (; x) Gt^N (xn ) (; xn )) 2" n!1 tT

© 2001 by Chapman & Hall/CRC

Æ)

372

Maxingale problem

is proved similarly if we choose Æ > 0 such that 2Æ N +1 (x) N (x), sup jGNu +1 GNv +1 j "; u;vT ju vjÆ sup jGt^(N (x)+Æ) (; x) Gt^N (x) (; x)j "; tT and consider n for which, in addition to (5.1.29), N (x) + Æ N +1 (xn ). Thus, fX ;N ; 2 g and GN (); 2 Rd ; satisfy all the conditions of Theorem 5.1.5. Hence, the net fL(X ;N ); 2 g is C { exponentially tight, and if N is an LD accumulation point, then x0 = 0 N { a.e.Nand the function Y N () = (Y Nt (; x); t 2 R+ ; x 2 C ) de ned by Y t (; x) = exp( xt Gt^N (x) (; x)) is a C{local exponential maxingale on (C ; N ). Therefore, to prove that N solves (M N ), it is left to show that N Y t (; x) = YtN (; x) N {a.e., which in view of (5.1.22) and the de nition of Y follows by the equality

xt^N (x) = xt

N {a.e.

(5.1.31)

To see the latter, let fX 0 ;N ; 0 2 0 g be a subnet of fX ;N ; 2 g that LD converges to . Since Xt;N = Xt;N ^N (X ;N ) by (5.1.20) and N (5.1.21), is supported by C , N (x) is C {continuous, and the set fx 2 D : xt^N (x) 6= xt g is C {open, by Corollary 3.1.9 0

0

N 0 = lim sup P0 0 Xt ;N 6= Xt^;N 0 ;N ) (xt 6= xt^N (x) ); ( X N 0 0 2 which proves (5.1.31). Thus, the nets fX ;N ; 2 g; N 2 N ; satisfy the conditions of Lemma 5.1.17. By the lemma for T > 0 1=r

lim lim sup P1=r (N (X ) T ) = 0;

(5.1.32)

N !1 2

and using (5.1.20) and (5.1.21) we have that

lim lim sup P1=r sup jXt Xt;N j > 0 = 0; N !1 2 tT

© 2001 by Chapman & Hall/CRC

373

Convergence of characteristics

which implies by C {exponential tightness of fL(X ;N ); 2 g for every N 2 N and Theorem 3.2.3 that fL(X ); 2 g is C { exponentially tight. Also (5.1.32) and (sup E )loc imply (sup E ). Thus, all the conditions of Theorem 5.1.16 with X 0 = X hold. An application of that theorem ends the proof. Theorem 5.1.12 follows by Theorem 5.1.10 and Remark 5.1.8.

5.2 Convergence of characteristics This section formulates conditions on convergence of the characteristics of the X in order for the net fL(X ); 2 g to be exponentially tight with all the LD accumulation points being solutions of a maxingale problem. We retain the above notation. As in Section 4.2 the cumulant in the limiting maxingale problem will have the semimaxingale representation (2.7.7) and (2.7.55), however, the characteristics can depend on x, on the one hand, and are de ned for x 2 D , on the other hand.

De nition 5.2.1. Let us say that a function f : R+ D ! Rk is D{progressively measurable if its restriction to [0; t] D is B([0; t])

k Dt =B(R )-measurable.

We assume as given the following objects: (bs (x); s 2 R+ ; x 2R D ) is an Rd -valued D-progressively measurable function such that 0tjbs (x)jds < 1 for t 2 R+ and x 2 D , cs (x); s 2 R+ ; x 2 D is a D-progressively measurable function with values in the R t space of symmetric, positive semi-de nite d d-matrices such that 0 kcs (x)k ds < 1 for t 2 R+ and x 2 D , s( ; x); s 2 R+ ; 2 B(Rd ); x 2 D is a transition kernel from ([0; t] D ; B ([0; t]) Dt ) into (Rd ; B(Rd )) for every t 2 R+ such that for t 2 R+ ; x 2 D and 2 R+ Z

Rd

j j ^ 1 t(dx; x) < 1; x2

Z

Rd

ejxj 1(jxj > 1) t (dx; x) < 1;

t (f0g; x) = 0; jxj2 ^ 1 t (x) < 1; ejxj 1(jxj > 1) t (x) < 1;

^s( ; x); s 2 R+ ; 2 B(Rd ); x 2 D is a transition kernel from ([0; t] D ; B (R+ ) Dt ) into (Rd ; B(Rd )) for every t 2 R+ such that

© 2001 by Chapman & Hall/CRC

374

Maxingale problem

for s 2 R+ ; x 2 D and

2 B(Rd )

^s ( ; x) s ( ; x); ^s(Rd ; x) 1:

(5.2.2)

Since D-progressively measurable functions are C-progressively measurable, the restrictions of (bs (x)); (cs (x)); (s ( ; x)), and (^s ( ; x)) to C satisfy the conditions on the local characteristics of a semimaxingale as de ned in Section 4.2. Also, given a limiter h : Rd ! Rd we de ne extensions of the characteristics of a semimaxingale to D by

B 0 (x) =

Zt

bs (x) ds;

(5.2.3)

Bt (x) = Bt0 (x) + (h(x) x) t (x);

(5.2.4)

t

Ct (x) =

0

Zt

0

cs (x) ds;

(5.2.5)

and refer to B 0 = (Bt0 (x); t 2 R+ ; x 2 D ) as the rst characteristic \without truncation" of the limiting semimaxingale, to B = (Bt (x); t 2 R+ ; x 2 D ) as the rst characteristic associated with limiter h(x), to C = (Ct (x); t 2 R+ ; x 2 D ) as the second characteristic, to s ( ; x) as the density of the measure of jumps, and to ^s( ; x) as the density of the discontinuous measure of jumps. If B = (Bt (x); t 2 R+ ; x 2 D ) is the rst characteristic associated with a limiter h (x), then

Bt (x) = Bt (x)+(h (x) h(x)) t (x):

The modi ed second characteristic C~ = (C~t (x); t speci ed by the equality

(5.2.6)

2 R+ ; x 2 D ) is

C~t (x) = Ct (x) + ( h(x))2 t (x) Zt

0

( h(x) ^s (x))2 ds; 2 Rd : (5.2.7)

We introduce a number of conditions on the characteristics, which are analogues of the continuity and majoration conditions on G() in Section 5.1. As above, we denote by U a dense subset of R+ .

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

375

De nition 5.2.2. We say that B (respectively, C ; C~ ; ; or ^) satis es the continuity condition if Bt (x) (respectively, Ct (x); C~t (x); f (x) t (x) for fR : Rd ! R Borel measurable and such that jf (x)j 1 ^ jxj2 ; 0t (g(x) ^s(x))k ds for g : Rd ! R Borel measurable and bounded, and k = 2; 3; : : :) is C {continuous in x for all t 2 U.

Remark 5.2.3.

If the continuity condition on holds, then by (5.2.6) the continuity condition on B does not depend on a limiter. If, in addition, the continuity condition on ^ holds, then by (5.2.7) the continuity conditions on C and C~ are equivalent.

Occasionally, we will need a stronger form of the continuity condition on B which is an analogue of the uniform continuity condition for G().

De nition 5.2.4. We say that B satis es the uniform continuity condition if the map x ! (Bt (x); t 2 R+ ) is C {continuous as a map from

D

into C .

Remark 5.2.5. Since Ct (x); C~t (x) and f (x) t (x), if f 0,

are increasing, continuous in t and equal to 0 at 0, the continuity conditions on C , C~ and are equivalent to C {continuity of the respective maps x ! (Ct (x); t 2 R+ ) from D into C (R + ; Rdd ), x ! (C~t (x); t 2 R+ ) from D into C (R + ; Rdd ) and x ! (f (x) t (x); t 2 R + ), where jf (x)j 1 ^ jxj2 , from D into C (R + ; R ). Thus, the continuity conditions on C , C~ and imply the associated uniform continuity conditions. Therefore, we will sometimes also be referring to the continuity conditions for C , C~ and as uniform continuity conditions. De nition 5.2.6. We say that B (respectively, C , C~ , or ) satis es the majoration condition (respectively, the local majoration condition) if the functions ( Bt (x); t 2 R+ ; x 2 D ) for all 2 Rd ; ( Ct (x); t 2 R+ ; x 2 D ) for all 2 Rd ; ( C~t (x); t 2 R+ ; x 2 D ) for all 2 Rd ; (f (x) t (x); t 2 R+ ; x 2 D ) for all f 2 Cb satisfy the majoration condition (respectively, the local majoration condition).

Remark 5.2.7. The majoration condition (respectively, the local majoration condition) on B equivalently requires that the function (Vart B (x); t 2 R+ ; x 2 D ) of total variation of B obey the majoration condition (respectively, the local majoration condition). The

© 2001 by Chapman & Hall/CRC

376

Maxingale problem

majoration conditions (respectively, the local majoration conditions) on C and C~ are equivalent to the majoration conditions (respectively, local majoration conditions) on the respective functions of the sums of the diagonal entries of Ct and C~t .

De nition 5.2.8. We say that satis es the C {local boundedness condition if for every compact K C and every > 0; t > 0; sup ejxj 1(jxj > 1) t (x) < 1: (5.2.8) x2K

We say that ^ satis es the C {local boundedness condition if for every compact K C and every > 0; t > 0; sup sup ejxj ^s (x) < 1: (5.2.9) x2K st

We now state conditions on the triplets of the X . They are similar to the conditions of Section 4.2. Actually, the conditions on the X0 and big jumps are the same. We repeat them here for completeness. Let (B ; C ; ) be the predictable characteristics of X corresponding to a limiter h(x). As above, x0 2 Rd . (0) (A)

1=r P X0 ! x0 as 2 ; lim lim sup P1=r ([0; t]; jxj > A) 1=r A!1 2

> " = 0;

t > 0; " > 0;

(a)

(sup (C ) (C~ ) ( )

lim lim sup P1=r

a!1 2

1 r jxj e 1 (r jxj > a) 1(jxj A) t > " r = 0; t > 0; > 0; A > 0; " > 0;

1=r P B) sup jBt Bt (X )j ! 0 as 2 ; T > 0; tT lim lim sup P1=r ( kr Ct;Æ Ct (X )k > ") = 0; t 2 U; " > 0, Æ!0 2 1=r P ~ ~ kr Ct Ct(X )k ! 0 as 2 ; t 2 U; 1=r P f (x) t f (x) t (X ) ! 0 as 2 ; t 2 U; f 2 Cb ;

© 2001 by Chapman & Hall/CRC

377

Convergence of characteristics

1 X (^ ) f (r x) s k r 0<st

Zt

f (x) ^s

0

1=r k P (X ) ds !

0 as 2 ;

t 2 U; k = 2; 3; : : : ; f

2 Cb :

(We recall that f (x) = f (r x)=r .) Theorem 5.2.9. Let h(x) be continuous, B , C (respectively, C~ ), , and ^ satisfy the continuity conditions, and and ^ satisfy the C {local boundedness conditions. Let also the majoration conditions on B , C (respectively, C~ ) and hold. If conditions (0); (A) + (a); (sup B ); (C ) (respectively, (C~ )), ( ), and (^ ) hold, then the net fL(X ); 2 g is C {exponentially tight and its every LD accumulation point is a solution to the maxingale problem (x0 ; G). Remark 5.2.10. Condition (a) can be replaced with the condition (a0 )

where ja;A;

1=r 1 j (x) t;c lim lim sup P a!1 2 r a;A; 1 X (x) > " = 0; + ln 1 + ja;A; s r 0<st t > 0; > 0; A > 0; " > 0; (x) = (er jxj 1) 1(r jxj > a) 1(jxj A):

Remark 5.2.11. Theorems 5.2.9 and 2.8.5 imply Theorem 4.2.1. Next comes a locally bounded version. Let us de ne N (x) by (5.1.3) and introduce the conditions (A)loc (a)loc

lim lim sup P1=r ([0; t ^ N (X )]; jxj > A)1=r > " = 0; A!1 2 t > 0; N 2 N ; " > 0; 1=r er jxj 1(r jxj > a) 1(jxj A) t^N (X ) lim lim sup P a!1 2 r > " = 0; t > 0; N 2 N ; > 0; A > 0; " > 0;

(sup B )loc

sup jBt^N (X ) tT

© 2001 by Chapman & Hall/CRC

1=r P Bt^N (X ) (X )j !

0 as 2 ;

T > 0; N

2 N;

378

Maxingale problem

(C )loc lim lim sup P1=r kr Ct;Æ ^N (X ) Æ!0 2 (C~ )loc

kr C~t^N (X )

( )loc

f (x) t^N (X )

(^ )loc

1 r st^

X N (X )

Ct^N (X ) (X )k > " = 0; t 2 U; N

C~t^N (X ) (X )k

1=r P

2 N ; " > 0;

! 0 as 2 ; t 2 U; N 2 N ; P

1=r

f (x) t^N (X ) (X ) ! 0 as 2 ; t 2 U; N 2 N ; f 2 Cb ; k

f (r x) s

t^ZN (X ) 0

f (x) ^s (X ) k ds 1=r P

! 0 as 2 ; t 2 U; N 2 N ; k = 2; 3; : : : ; f 2 Cb :

Theorem 5.2.12. Let h(x) be continuous.

Let B , C (respectively, ~ C ), , and ^ satisfy the continuity conditions, and and ^ satisfy the C {local boundedness conditions. Let the local majoration conditions on B , C (respectively, C~ ) and hold. Let condition (NE ) hold. If conditions (0); (A)loc + (a)loc ; (sup B )loc ; (C )loc (respectively, (C~ )loc ), ( )loc , and (^ )loc hold, then the net fL(X ); 2 g is C { exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.2.13. Condition (a)loc can be replaced with the condition (a0 )loc

1=r 1 lim lim sup P j (x) t;c ^N (X ) a!1 2 r a;A; X 1 (x) > " = 0; + ln 1 + ja;A; s r st^ (X ) N t > 0; N 2 N ; > 0; A > 0; " > 0:

De nition 5.2.14. We say that bs(x), respectively, cs(x), meets the linear-growth condition if there R exists an R + -valued Lebesgue measurable function ls such that 0t ls ds < 1; t 2 R+ ; and

jbs(x)j (1+ xs )ls ;

© 2001 by Chapman & Hall/CRC

(5.2.10)

379

Convergence of characteristics

respectively,

kcs (x)k (1+(xs )2 )ls:

(5.2.11)

We say that meets the linear-growth condition if Z

1 x)s (dx; x)

(ex

Rd

Z

(e(1+xs )x

Rd

1 (1 + xs ) x)ms (dx); 2 Rd ; (5.2.12)

where ms (dx) is aR transition kernel from (R+ ; B(R+ )) into tR d d (R ; B(R )) such that 0 Rd (exp(jxj) 1 jxj)ms (dx)ds < 1; t > 0; > 0. Theorem 5.2.15. Let h(x) be continuous, B , C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let the linear-growth conditions on bs (x), cs (x), and s( ; x) hold. Let ^ satisfy the C {local boundedness condition. If conditions (0); (A)loc + (a)loc ; (sup B )loc ; (C )loc (respectively, ~ (C )loc ), ( )loc , and (^ )loc hold, then the net fL(X ); 2 g is C { exponentially tight and its every LD accumulation point solves the maxingale problem (x0 ; G).

Remark 5.2.16. 0 (a )loc .

Condition (a)loc can be replaced with condition

Remark 5.2.17. In the above theorems we can equivalently describe the accumulation points by saying that if is an LD accumulation point of fL(X ); 2 g, then the canonical process X on (C ; C; )

is a Luzin-continuous semimaxingale with characteristics (B; C; ; ^) ld starting at x0 . If (B; C; ; ^) and x0 uniquely specify , then X ! X.

The proofs use the ideas of the proofs of Theorems 4.2.1, 5.1.5, and 5.1.10. An outline is as follows: we introduce truncated processes X^ ;a as in the proof of Theorem 4.2.1, establish that the pairs (X^ ;a ; X ) as random elements of D 0 (= D (R + ; Rd Rd )) satisfy the conditions of Theorem 5.1.16, then observe in view of Lemma 4.2.16

© 2001 by Chapman & Hall/CRC

380

Maxingale problem

that condition (A) + (a) implies that the nets f(X ; X ); 2 g and f(X^ ;a ; X ); 2 g asymptotically (as a ! 1) have the same LD limit, and derive the statements of Theorems 5.2.9, 5.2.12 and 5.2.15 by taking the limit as a ! 1 in the maxingale problems associated with the nets f(X^ ;a ; X ); 2 g. This turns out to be quite a long way. The next subsection studies required exponential tightness properties. After, LD accumulation points are identi ed as solutions to certain maxingale problems, and nally the proofs of the above results are given. We assume in the proofs that x0 = 0.

5.2.1 Exponential tightness results

We develop exponential tightness results for (X^ ;a ; X ). The next lemma extends Lemma 4.2.6. Lemma 5.2.18. Let Z ;Æ = (Zt;Æ ; t 2 R+ ); Z0;Æ = 0; Æ > 0; 2 ; and Z = (Zt ; t 2 R+ ); Z0 = 0; 2 ; be Rd {valued, componentwise increasing processes with paths in D de ned on respective probability spaces ( ; F ; P ) such that for all t 2 U; " > 0; lim lim sup P1=r (jZt;Æ Ztj > ") = 0:

Æ!0 2

If the net fL(Z ); 2 g is C {exponentially tight, the latter convergence is uniform on bounded intervals, i.e.,

lim lim sup P1=r (sup jZt;Æ Zt j > ") = 0; T > 0; " > 0: Æ!0 2 tT Proof. Acting as in the proof of Lemma 4.2.6 for N 2 N we choose tNi 2 U; i = 0; : : : ; kN , such that 0 = tN0 < tN1 < : : : < tNkN 1 < T tNkN < T + 1 and jtNi tNi 1j 1=N; i = 1; : : : ; kN . Then by the same argument sup jZt;Æ Zt j max N jZt;Æ N ZtN j+ sup jZt Zs j: i i i=1;:::;k tT s;tT +1: js tj1=N Therefore,

P1=r

sup jZt;Æ tT

© 2001 by Chapman & Hall/CRC

Zt j > "

kN X

i=1 1=r + P

P1=r jZt;Æ N i

sup jZt s;tT +1: js tj1=N

ZtNi j > "=2

Zs j > =2

381

Convergence of characteristics

so that by hypotheses lim lim sup P1=r (sup jZt;Æ Æ!0 2 tT

Zt j > ")

lim sup P1=r 2

The right-hand side tends to 0 as N

sup jZt s;tT +1: js tj1=N

Zs j > =2

! 1 by Theorem 3.2.3.

We now study, as in the proof of Theorem 4.2.1, when one can replace conditions (C ), (C~ ) and ( ) with the associated uniform versions. Let us introduce the conditions: (sup C )

lim lim sup P1=r (sup kr Ct;Æ tT

Æ!0 2

(sup C~ )

sup kr C~t tT

(sup )

sup jf (x) t tT

C~t (X )k

Ct (X )k > ") = 0;

1=r

" > 0; T > 0;

P

! 0 as 2 ; T > 0;

f (x) t

1=r P (X )j !

0 as 2 ;

f 2 Cb ; T > 0: In some of the statements below we say, with a slight abuse of terminology, that nets of laws of Rm {valued random processes are C { exponentially tight if they are C (R + ; Rm ){exponentially tight. The meaning should be clear from the context. Lemma 5.2.19. 1. Let the net fL(C (X )); 2 g (respectively, fL(C~ (X )); 2 g; fL(f (X )); 2 g; f 2 Cb) be C { exponentially tight. Then condition (C ) (respectively, (C~ ); ( )) is equivalent to condition (sup C ) (respectively, (sup C~ ); (sup )). 2. If the nets fL(f (x) (X )); 2 g; f 2 Cb ; are C { exponentially tight and condition ( ) holds, then condition (sup B ) does not depend on the particular choice of a continuous limiter h. Proof. The rst part follows by Lemma 5.2.18. The second part follows from the rst and is proved in the same way as Lemma 4.2.8.

© 2001 by Chapman & Hall/CRC

382

Maxingale problem

Now, we state and prove an exponential tightness result. We recall the de nition of X^ ;a = (X^t;a ; t 2 R+ ); a 2 R+ ; from Subsection 4.2.2:

X^ t;a = Xt X t;a ;

(5.2.13)

where

Xt;a = ha (x) =

X

ha (Xs ));

(5.2.14)

a 1 ^ 1 x; ha (x) = ha (r x): jxj r

(5.2.15)

st

(Xs

We also recall that C 0 = C (R + ; Rd Rd ).

Theorem 5.2.20.

1. Let the nets fL(B (X )); 2 g, fL(C (X )); 2 g (respectively fL(C~ (X )); 2 g ), and fL(f (x) (X )); 2 g; f 2 Cb; be C {exponentially tight. If conditions (0), (sup B ), (C ) (respectively, (C~ )), and ( ) hold, then the net fL(X^ ;a ); 2 g is C {exponentially tight. If, in addition, conditions (A) + (a) hold, then the net fL(X ); 2 g is C {exponentially tight, so, the net fL((X^ ;a ; X )); 2 g is C 0 {exponentially tight.

2. Let the net fL(X ); 2 g be C {exponentially tight. If the function B (respectively, C ; C~ ; ) satis es the uniform continuity condition, then the net fL(B (X )); 2 g (respectively, fL(C (X )); 2 g; fL(C~ (X )); 2 g; fL(f (x) (X )); 2 g; f 2 Cb ) is C {exponentially tight. Proof. Part 2 follows from Theorem 3.2.3 via a diagonal argument. Let, given T > 0 and > 0, fL(X 0 ); r0 ; p0 ; 0 2 0 g be a subnet of fL(X ); r ; P1=r sups;t2[0;T ]:jBt (X ) Bs (X )j > ; (; Æ) 2 js tjÆ ld 0 (0; 1)g such that L(X 0 ) ! at rate r0 , where 0 is supported by C , and

lim p0 = lim lim sup P1=r Æ!0 2 0 20

© 2001 by Chapman & Hall/CRC

sup jBt (X ) Bs (X )j > : s;t2[0;T ]: js tjÆ

383

Convergence of characteristics

Then for arbitrary Æ^ > 0 by the contraction principle (Corollary 3.1.15) and uniform continuity condition for B

lim lim sup P1=r Æ!0 2

sup jBt (X ) Bs (X )j > = lim p0 0 20 s;t2[0;T ]: js tjÆ 0 0 1=r0 lim0 sup P0 sup jBt (X ) Bs(X )j > 20 s;t2[0;T ]: js tjÆ^ 0 x 2 C : sup jBt (x) Bs(x)j : s;t2[0;T ]: js tjÆ^

By -smoothness of 0 the latter deviability tends to 0 as Æ^ ! 0. This checks condition I(ii) of Theorem 3.2.3. Condition I(i) holds since B0 (x) = 0 completing the proof of C -exponential tightness of fL(B (X )); 2 g. Proofs for the other processes are similar. We prove part 1. Let us assume, rst, that the nets fL(B (X )); 2 g, fL(C (X )); 2 g and fL(f (x) (X )); 2 g; f 2 Cb ; are C {exponentially tight. We begin with a proof of ^ ;a ); 2 g. Let C {exponential tightness for fL(X

Xt = X0 + Bt;Æ + Mt;Æ + x 1(r jxj > Æ) t be the canonical representation of X associated with the truncation function x 1(r jxj Æ), where Æ < a, so that B ;Æ = (Bt;Æ ; t 2 R+ ); B0;Æ = 0, is an F {predictable process with bounded variation over bounded intervals; M ;Æ = (Mt;Æ ; t 2 R+ ); M0;Æ = 0, is the F {locally squareintegrable martingale de ned by

Mt;Æ = Xt;c + x 1(r jxj Æ) ( )t : Since by (5.2.14) and (5.2.13) X^ t;a = Xt Æ < a, we have

(5.2.16) (x

ha (x)) t and

X^ t;a = X0 + Bt;Æ + Mt;Æ + ha (x) 1(r jxj > Æ) t ; so, by Theorem 3.2.3 and condition (0) in order to prove C { exponential tightness of fL(X^ ;a ); 2 g it suÆces to prove that,

© 2001 by Chapman & Hall/CRC

384

Maxingale problem

for all T > 0; > 0, lim lim sup P1=r (sup jBt;Æ j > A) = 0; A!1 2 tT

(5.2.17a)

= 0;

(5.2.17b) (5.2.17c)

lim lim sup sup P 1=r (sup jBt;Æ + !0 2 2ST (F ) t

Bt;Æ j > )

lim lim lim sup P1=r (sup Mt;Æ > A) = 0; Æ!0 A!1 2 tT 1=r ;Æ lim lim sup lim sup sup P (sup Mt;Æ + Mt Æ!0 !0 t 2 2ST (F )

k

k

k

= 0;

j1

lim lim sup P1=r ( ha (x) A!1 2

j

(r jxj > Æ) T

= 0;

lim lim sup sup P1=r !0 2 2ST (F )

Z+Z

Rd

k > ) (5.2.17d)

> A) (5.2.17e)

jha (x)j 1(r jxj > Æ)

(ds; dx) > = 0:

(5.2.17f)

We begin with B ;Æ . Let B~ ;Æ be the rst characteristic of X associated with hÆ (hÆ is de ned as ha with Æ = a). Then by (4.1.4)

Bt;Æ = B~t;Æ +(x 1(r jxj Æ) hÆ (x)) t ; so, for 0 s < t, using (5.2.15),

jBt;Æ Bs;Æ j jB~t;Æ B~s;Æ j Z t Z + s Rd

Æ x 1 (r jxj > Æ) (ds; dx) r jxj

jB~t;Æ

B~s;Æ j +

Zt Z

s

Rd

f~Æ(x) (ds; dx); (5.2.18)

where f~Æ (x) = (2jxj=Æ 1)+ ^ 1 (Recall that, by the notation introduced in Section 4.2, f~Æ(x) = f~Æ (r x)=r .)

© 2001 by Chapman & Hall/CRC

385

Convergence of characteristics

Let B~tÆ (x) be the rst characteristic of X associated with the limiter hÆ (x) so that it is de ned by (5.2.4) with hÆ (x) as h(x). Then in view of (5.2.6) fL(B~ Æ (X )); 2 g is C {exponentially tight since fL(B (X )); 2 g and fL((hÆ (x) h(x)) (X ))); 2 g are both C {exponentially tight, hence, by (sup B ) with hÆ in place of h and Theorem 3.2.3 the net fL(B~ ;Æ ); 2 g is C {exponentially tight. Since f~Æ 2 Cb , the net fL(f~Æ (X )); 2 g is C {exponentially tight by hypotheses so that by Lemma 5.2.18 condition (sup ) holds with f = f~Æ . Theorem 3.2.3 implies that the net fL(f~Æ ); 2 g is C {exponentially tight. Inequality (5.2.18), C {exponential tightness of fL(B~ ;Æ ); 2 g and fL(f~Æ ); 2 g imply by Theorem 3.2.3 that the B ;Æ satisfy (5.2.17a) and (5.2.17b). Now we prove (5.2.17e) and (5.2.17f). The process (jha (x)j 1(r jxj > Æ) t ; t 2 R+ ) has as its stochastic cumulant the process (exp(jha (x)j) 1) 1(r jxj > Æ) t ; t 2 R+ ; 2 R. Then for 2 ST (F ) and c > 0 by the second inequality in Lemma 5.1.15 with d = 1

P

Z+Z

Rd

jha (x)j 1(r jxj > Æ) (ds; dx) >

2 exp r c2

1 + 2 max P i=1;2 r

Z+Z

Rd

exp(( 1)i jha (r x)jc) 1 1(r jxj > Æ)

(ds; dx) >

c ; 2

hence,

P1=r

Z+Z

Rd

jha (x)j 1(r jxj > Æ) (ds; dx) >

© 2001 by Chapman & Hall/CRC

21=r exp c2

386

Maxingale problem

+ 21=r P1=r

Z+Z

f (x) (ds; dx) >

Rd

c ; (5.2.19) 2

where f (x) = (exp(cjha (x)j) 1)f~Æ (x). Since f (x) belongs to Cb , the net fL(f (x) (X )); 2 g is C {exponentially tight by hypotheses. From (sup ) and Theorem 3.2.3 we derive that the net fL(f (x) ); 2 g is C {exponentially tight. Then (5.2.19) and Theorem 3.2.3 imply (5.2.17f). Limit (5.2.17e) is proved similarly. Now we prove (5.2.17c) and (5.2.17d). Denoting the stochastic cumulant associated with M ;Æ by G~ ;Æ () = (G~ ;Æ t (); t 2 R + ) and applying Lemma 5.1.15, we reduce the proof of (5.2.17c) and (5.2.17d) to the proof of the respective limits (i = 1; : : : ; 2d; T > 0; > 0) 1 lim lim lim sup P1=r sup G~ ;Æ ( r e ) > A = 0; (5.2.20a) i t Æ!0 A!1 2 tT r 1 lim lim sup lim sup sup P1=r sup (G~ ;Æ t+ (r ei ) Æ!0 !0 r t 2 2ST (F ) G~ ;Æ ( r e )) >

= 0: (5.2.20b) i

Now, by (5.2.16) the measure of jumps of M ;Æ is

~;Æ ([0; t];

)=

X

0<st

1

Z Rd

x 1(r jxj Æ)( (fsg; dx)

(fsg; dx)) 2

nf0g ; 2 B(Rd );

and then by (4.1.14) and the fact that the F {predictable quadraticvariation process of X ;c is C 1 ;Æ 1 G~ ;Æ t () = 2 Ct + exp (x 1(r jxj Æ ) xs ) (x 1(r jxj Æ) x;Æ s ) t ; (5.2.21) where

x;Æ s =

Z Rd

x 1(r jxj Æ) (fsg; dx):

© 2001 by Chapman & Hall/CRC

(5.2.22)

387

Convergence of characteristics

We note that, since by (4.1.3a) (fsg; R d ) 1, we have Æ jx;Æ (5.2.23) s j r : Applying Taylor's formula to the integrand in (5.2.21), we obtain 1 ~ ;Æ G (r ) = Tt;Æ ()+ rt;Æ (); (5.2.24) r t where 1 1 2 Tt;Æ () = (r Ct)+ r ((x 1(r jxj Æ) x;Æ s )) t 2 2 (5.2.25) and jje2jjÆ r ((x 1(r jxj Æ) x;Æ ))2 sup jrs;Æ ()j Æ s t 3 st (5.2.26) (for the last inequality we also used (5.2.23)). Since by (5.2.22) 2 2 ( (x 1(r jxj Æ) x;Æ s )) t = ( x) 1(r jxj Æ ) t X 2 d ( x;Æ s ) (2 (fsg; R )); 0<st

where the sum is over s such that (fsg; R d ) > 0, and (fsg; R d ) 1, we have by (4.1.7), (5.2.22) and (5.2.25) that 0 Tt;Æ () Ts;Æ () r (Ct;Æ Cs;Æ ); s < t; and by (4.1.7) and (5.2.26) that jje2jjÆ r C ;Æ : sup jrs;Æ ()j Æ t 3 st

(5.2.27) (5.2.28)

Since fL(C (X )); 2 g is C {exponentially tight, an application of (sup C ) yields in view of Theorem 3.2.3

lim lim sup P1=r kr CT;Æ+1 k > A = 0; Æ!0 2 A!1 ;Æ lim lim sup sup P1=r sup r kCt;Æ + C k > = 0; Æ!0 2 2ST (F ) t !0

© 2001 by Chapman & Hall/CRC

388

Maxingale problem

for every > 0. The rst of these relations together with (5.2.28) implies that lim lim sup P1=r ( sup jrt;Æ ()j > ") = 0; " > 0; tT +1

Æ!0 2

(5.2.29)

and together with (5.2.27) that lim lim lim sup P1=r ( sup Tt;Æ () > A) = 0; Æ!0 A!1 2 tT +1

(5.2.30)

while the second one and (5.2.27) yield ;Æ lim lim sup lim sup sup P1=r (sup jTt;Æ + () T ()j > ") t 2 2ST (F ) = 0; " > 0: (5.2.31)

Æ!0 !0

In view of (5.2.24), limits (5.2.29) and (5.2.30) prove (5.2.20a), while limits (5.2.29) and (5.2.31) prove (5.2.20b). C {exponential tightness of the net fL(X^ ;a ); 2 g has been proved. C {exponential tightness of the net fL(X ); 2 g under (A) + (a) follows by Theorem 3.2.3 and Lemma 4.2.16. Finally, by Corollary 3.2.7 C 0 { exponential tightness of the net fL((X^ ;a ; X )); 2 g is implied by C {exponential tightness of both fL(X^ ;a ); 2 g and fL(X ); 2 g. We now assume that instead of condition (C ) and C {exponential tightness of fL(C (X )); 2 g we have condition (C~ ) and C { exponential tightness of fL(C~ (X )); 2 g. As above, it is suf cient to prove C {exponential tightness of fL(X^ ;a ); 2 g; a > 0: We again consider a canonical representation of X , but this time with respect to hÆ (x) from (5.2.15) with Æ < a so we replace a by Æ in (5.2.15) and substitute throughout in the preceding argument hÆ (x) for x 1(r jxj Æ). Retaining the above notation for the components of the representation, we again reduce the task to proving (5.2.17a){(5.2.17f). Limits (5.2.17a) and (5.2.17b) follow by (sup B ) and C { exponential tightness of fL(B (X )); 2 g and fL(f (X )); 2 g; f 2 Cb . The proofs of (5.2.17e) and (5.2.17f) do not change. The proofs of (5.2.17c) and (5.2.17d) also proceed along the same lines, (5.2.27) and (5.2.28) being replaced by (with hÆ (x) in place of

© 2001 by Chapman & Hall/CRC

389

Convergence of characteristics

x 1(r jxj Æ) in (5.2.22), (5.2.25) and (5.2.26)) 0 Tt;Æ () Ts;Æ ()

r(C~t;Æ C~s;Æ ); (5.2.32) jj sup jrs;Æ ()j Æ e2jjÆ r C~t;Æ ; (5.2.33) 3 st

where C~t;Æ is de ned by (4.1.8) with hÆ (x) as h(x). By (4.1.8) for 0 s t

r (C~t;Æ Z Z

+ r

(s;t] Rd

+ r

C~s;Æ ) r (C~t C~s)

(

X

s 0 be such that h(x) = x; jxj c: Then we obtain for Æ < c, recalling the notation jjhjj = sup jh(x)j, x2Rd Z Z

(s;t] Rd

j( h(x))2 ( hÆ (x))2 j (du; dx)

Z j j2 2 2 (jjhjj + Æ )

r2

Z

1(r jxj > c) (du; dx)

(s;t] Rd Z Z 2

+j

j

(s;t] Rd

jxj2 1(Æ < rjxj c) (du; dx)

and X

s 0; > 0; > 0; and Æ < 1 ^ c ^ jjhjj, by (sup C~ ) and (sup ) lim sup P1=r sup j r (C~t;Æ C~s;Æ )j > 2 js tj s;tT lim sup P1=r sup j r(C~t (X ) C~s(X ))j > 6 2 js tj s;tT

Z Z

+ lim sup P1=r 10jj2 jjhjj2 sup 1(jxj > c=2) 2 js tj (s;t] d R s;tT (du; dx)(X ) > 6 Z Z 1=r 2 + lim sup P 4jj sup jxj2 1(Æ=2 < jxj 2c) 2 js tj (s;t] d R s;tT (du; dx)(X ) > ; 6 where the right-hand side goes to 0 as ! 0 by C {exponential tightness of fL(C~ (X )); 2 g and fL(f (X )); 2 g; f 2 Cb ; and Theorem 3.2.3. By (5.2.32) this proves (5.2.31). Limits (5.2.29) and (5.2.30) are proved similarly. Thus, C {exponential tightness of fL(X^ ;a ); 2 g under the new set of assumptions has been proved. The theorem has been proved.

© 2001 by Chapman & Hall/CRC

391

Convergence of characteristics

Remark 5.2.21. According to the theorem, under (0), (a) + (A),

(sup B ), (C ) (or (C~ )), and ( ), and the uniform continuity conditions on B , C (or C~ ), and , C {exponential tightness of the net fL(X ); 2 g is equivalent to C {exponential tightness of the nets fL(B (X )); 2 g, fL(C (X )); 2 g (or fL(C~ (X )); 2 g ), and fL(f (x) (X )); 2 g; f 2 Cb .

5.2.2 LD accumulation points as solutions to maxingale problems In this subsection, assuming that either the net fL((X^ ;a ; X )); 2 g is C 0 {exponentially tight or the net fL(X ); 2 g is C { exponentially tight, we characterise their LD accumulation points as solutions to maxingale problems. The rst step is to consider small jumps as in Theorem 4.2.11. As in Subsection 4.2.2 the triplet of X^ ;a without truncation is (B ;a ; C ; ;a ), where B ;a is the rst characteristic of B corresponding to ha and

;a ([0; t]; ) = ([0; t]; (ha ) 1 ( ));

2 B(Rd ); t 2 R+ :

We also note that by (5.2.15) and (5.2.35)

;a ([0; t]; fr jxj > ag) = 0; t 2 R+ ; 2 :

(5.2.35) (5.2.36)

Since the jumps of X^ ;a are bounded in modulus by a=r , its stochastic exponential is well de ned. We denote it by E^;a () = (E^t;a (); t 2 R+ ); 2 Rd . The \limiting" semimaxingale X a is de ned in analogy with Subsection 4.2.2 as having characteristics (B a ; C; a ; ^a ) relative to ha , which are de ned for x 2 D and 2 B(Rd ) by

Bta (x) = Bt0 (x) + (ha (x) x) t (x); ta ( ; x) = t (ha 1 ( ); x); ^ta (x) = ^t (ha 1 ( ); x):

(5.2.37) (5.2.38)

We also note that

ta (fjxj > ag; x) = 0:

© 2001 by Chapman & Hall/CRC

(5.2.39)

392

Maxingale problem

The associated cumulant is given by 1 G^ at (; x) = Bta (x)+ Ct (x) +(ex 1 ha (x)) ta (x) 2 +

Zt

ln 1 + (ex

0

1) ^sa (x)

ex

1

^sa (x) ds: (5.2.40)

Obviously G^ at (; x) is continuous in t and Dt {measurable in x. Let (M^ a ) denote the maxingale problem (M 0 ) introduced in Subsection 5.1.1 with G() replaced with G^ a () = (G^ at (; x); t 2 R+ ; x 2 ^ a on C 0 solves (M^ a ) if D ); 2 Rd , i.e., deviability x0a = 0 d ^ a {a.e. (M^ a ) Y^ (); 2 R ; is a C 0 {local exponential maxingale on (C 0 ; ^ a ); where Y^ a () = (Y^ta (; (x; x0 )); t 2 R+ ; (x; x0 ) 2 C 0 ) is de ned by

Y^ta (; (x; x0 )) = exp( xt G^ at (; x0 )): (5.2.41) Theorem 5.2.22. Let the net fL((X^ ;a ; X )); 2 g be C 0 { exponentially tight, let B satisfy the uniform continuity condition, and let C (respectively C~ ), and ^ satisfy the continuity conditions. If conditions (0), (sup B ), (C ) (respectively (C~ )), ( ), and (^ ) hold, then every LD accumulation point of fL((X^ ;a ; X )); 2 g solves (M^ a ).

The idea of the proof is to apply Theorem 5.1.16 to the pair (X^ ;a ; X ) (note that X^ ;a plays the role of X in Theorem 5.1.16, and X plays the role of X 0 ). We again need auxiliary results. The second part of the next lemma extends Lemma 4.2.9.

Lemma 5.2.23. I. Let the net fL(X ); 2 g be C {exponentially

tight and satisfy the continuity condition. Then, for c > 0; " > 0 and t > 0 1.

lim lim sup P1=r jxj2 1(jxj Æ) t (X ) " = 0; Æ!0 2

2.

lim lim sup P1=r Æ(jxj ^ c) 1(jxj Æ) t (X ) " = 0: Æ!0 2

© 2001 by Chapman & Hall/CRC

393

Convergence of characteristics

II. If, in addition, conditions ( ) and (^ ) hold, then for " > 0 and t2U 1 1. lim lim sup P1=r f (r x) 1(r jxj > Æ) t f (x) t (X ) Æ!0 2 r >" =0 for all R+ -valued bounded continuous functions f (x); x 2 Rd ; such that f (x) cjxj2 in a neighbourhood of 0 for some c > 0; Æ 2. lim lim sup P1=r jg(r x)j 1(r jxj > Æ) s > " = 0; Æ!0 2 r and 1 X g(r x) 1(r jxj > Æ) s k 3. lim lim sup P1=r Æ!0 2 r 0<st Zt

g(x) ^s (X ) k ds > " = 0; k = 2; 3; : : :

0

for all R-valued, bounded and continuous functions g(x); x 2 Rd ; such that jg(x)j cjxj in a neighbourhood of 0 for some c > 0. Proof. We begin with part I. We denote

HÆ (x) = jxj2 1(jxj Æ) t (x);

x 2 D:

(5.2.42)

Let a subnet f(L(X 0 ); r0 ; p0 ); 0 2 0 g of f(L(X ); r ; P1=r HÆ (X ) " ); (; Æ) 2 (0; 1)g be such that

lim p0 = lim sup lim sup P1=r HÆ (X ) " 0 0 2 Æ!0 2

ld 0 and L(X 0 ) ! at rate r0 , where deviability 0 is supported by C . Then for arbitrary > 0 and t 2 U

lim sup lim sup P1=r HÆ (X ) " = lim p0 0 20 Æ!0 2 1=r 0 lim0 sup P0 0 H (X ) " 0 x 2 C : H (x) " ; 20 where the latter inequality follows by Corollary 3.1.9 since the continuity condition on implies that H (x) is C {upper-semi-continuous. Upper semi-continuity of H (x) on C implies that the sets fx 2 C :

© 2001 by Chapman & Hall/CRC

394

Maxingale problem

H (x) "g are closed. They converge to ; as ! 0 by Lebesgue's bounded convergence theorem. The -smoothness property of devia bility then implies that lim!0 0 x 2 C : H (x) " = 0: The rst assertion of part I is proved. For the second, picking > 0, we have for Æ small enough and x 2 D, Æ(jxj ^ c) 1(jxj Æ) t (x) Æ(jxj ^ c) 1(jxj ) t (x) + jxj2 1(jxj ) t (x); (5.2.43) where for the second summand, which is well de ned by (2.7.53a) and (2.7.54), we used Chebyshev's inequality. Let k(x) = (2jxj= 1)+ ^ 1; x 2 Rd . Since the function (jxj^ c)k(x) belongs to Cb and we can assume that t 2 U , C {exponential tightness of fL(X ); 2 g and the continuity condition on imply by Corollary 3.1.22 that the net fL((jxj ^ c)k(x) t (X )); 2 g is exponentially tight in R; hence, by Theorem 3.2.3, since 1(jxj ) k(x), " lim lim sup P1=r Æ(jxj^c) 1(jxj )t (X ) = 0: Æ!0 2 2 Inequality (5.2.43) and the rst assertion of part I yield the required. Part I is proved. Part II is proved in analogy with Lemma 4.2.9, necessary modi cations make use of part I and are obvious. The next lemma follows by Lemma 5.2.23 in the same way as Lemma 4.2.10 follows by Lemma 4.2.9.

Lemma 5.2.24. Let the net fL(X ); 2 g be C -exponentially tight

and satisfy the continuity condition. Then, under conditions ( ) and (^ ), condition (C ) is equivalent to condition (C~ ), which, hence, does not depend on the choice of h.

Lemma 5.2.25. Under the uniform continuity condition on B aand

the continuity conditions on C (or C~ ), and ^, the function G^ () satis es the uniform continuity condition. Proof. Since under the continuity conditions on and ^ the continuity conditions for C and C~ are equivalent, we can assume that the

© 2001 by Chapman & Hall/CRC

395

Convergence of characteristics

continuity condition on C holds. We prove that each of the functions on the right of (5.2.40) is C {continuous as a map from D into C (R + ; R ). The function B a has the required property by hypotheses and the fact that, in view of the continuity condition on , the uniform continuity condition for B does not depend on a limiter. The same fact is clearly also true for the next two terms on the right-hand side of (5.2.40) (use (5.2.38) for the third term). Let us consider the last term. Recalling the notation (x) = x ln(1 + x); x > 1 (seeR (4.2.26)), we have, since (x) 0, that it t x a is suÆcient to show that 0 (e 1) ^s (x) ds is C {continuous for each t 2 U . Let xn ! x^ 2 C and " > 0. By (5.2.39)

e jja 1

Z

(ex 1)^sa (dx; x) ejja 1;

Rd

x 2 D:

Also by Weierstrass' theorem there exists a polynomial q(u) = Pl k q(u)j " for k=2 dk x ; l 2; u 2 R+ ; such that j (u) x 2 [exp( jja) 1; exp(jja) 1]. Since by the continuity condition on ^ and (5.2.38) lim n!1

Zt

q

(ex

0

x

1)^sa ( n ) ds =

Zt

0

q (ex 1)^sa (^x) ds;

we obtain that Zt lim sup

n!1

(ex

0

1) ^sa (xn ) ds

Zt

(ex

0

1) ^sa (^x) ds 2":

Since " is arbitrary, the lemma is proved. Proof of Theorem 5.2.22. Since by Lemma 5.2.25 the function G^ a () satis es the uniform continuity condition, it is suÆcient to prove in view of Theorem 5.1.16 that 1=r

P 1 sup j ln E^t;a (r ) G^ at (; X )j ! 0 as 2 ; T > 0: tT r (5.2.44)

© 2001 by Chapman & Hall/CRC

396

Maxingale problem

The idea is to follow the proof of Theorem 4.2.1 in order to derive (5.2.44) from convergence of the characteristics of the X^ ;a ; 2 : Let C^t;a;Æ denote the modi ed second characteristic of X^ ;a corresponding to the truncation function x 1(jxj Æ) so that

C^t;a;Æ = Ct + ( x)2 1(r jxjÆ) t;a X x 1(r jxjÆ) s;a 2 ; 2 Rd : (5.2.45) 0<st

We note that the following conditions hold (sup (C a ) ( a ) (^ a )

1=r P ;a a sup jBt Bt (X )j ! 0 as 2 ; T > 0; tT lim lim sup P1=r ( kr C^t;a;Æ Ct (X )k > ") = 0; t 2 U; " > 0, Æ!0 2 1=r P ;a a f (x) t f (x) t (X ) ! 0 as 2 ; t 2 U; f 2 Cb ; 1=r Zt k k P 1 X ;a a f (r x) s f (x) ^s (X ) ds ! 0 r 0<st 0 as 2 ; t 2 U; k = 2; 3; : : : ; f 2 Cb :

Ba)

Proof is by the argument of the proof of Lemma 4.2.13. Condition (sup B a ) is actually condition (sup B ) with ha (x) as h(x) and holds since (sup B ) does not depend on the choice of a limiter by Lemma 5.2.19 and Theorem 5.2.20. Since by (5.2.35) for Æ < a

;a ([0; t];

\fr jxj Æg) = ([0; t]; \fr jxj Æg);

(5.2.45) and (4.1.7) imply that C^t;a;Æ = Ct;Æ when Æ < a, so that conditions (C a ) and (C ) coincide. Similarly, the argument of the proof of Lemma 4.2.13 shows in view of (5.2.38) and (5.2.35) that ( a ) and (^ a ) are implied by ( ) and (^ ). We now prove that conditions (sup B a ), (C a ), ( a ), and (^ a ) imply (5.2.44). This is carried out analogously to the proof of Theorem 4.2.11. Therefore, we do not give all the details but only indicate modi cations that have to be made in that proof. For this reason, we extensively use the notation of the proof of Theorem 4.2.11. Let us rst note that by (5.2.37) and (5.2.38) the functions Bta (x) and ta ( ; x) satisfy the same continuity conditions as imposed on

© 2001 by Chapman & Hall/CRC

397

Convergence of characteristics

Bt (x) and t ( ; x) in the statement of Theorem 5.2.22. Also by Lemma 5.2.19 conditions (sup C a ) and (sup a ) hold (with obvious notation). We now de ne in analogy with (4.2.14), (4.2.15a){ (4.2.15h), substituting ;a for , as = ;a (fsg; Rd ); and, for Æ > 0 and 2 Rd ,

x;Æ = x 1(r jxj Æ) s;a; s Ds;Æ () = (ex 1) 1(r jxj > Æ) s;a ; Rs;Æ () = exp( (x 1(r jxj Æ) x;Æ s )) 1 ;Æ (x 1(r jxj Æ) xs ) s;a ; ;Æ ;Æ Q;Æ s () = (exp( xs ) 1 + xs )(1 as ); ;Æ ;Æ ;Æ ;Æ G;Æ s () = exp( xs )Ds () + Rs () + Qs (); Ut;Æ () = (ex 1 x) 1(r jxjÆ) s;a;c; Vt;Æ () = (ex 1 x) 1(r jxj>Æ) s;a ; where t 2 R+ ; s 2 R+ , and ;a;c(ds; dx) is the continuous part of ;a (ds; dx). Let, as in Lemma 4.2.12,

Yt;Æ () = Zt;Æ ()

=

X

0<st X

(Ds;Æ ());

(5.2.46)

1 ;Æ ln(1 + G;Æ s ()) + Ut () + 2 Ct 0<st X

0<st

ln(1 + Ds;Æ ());

and, as in (4.2.23){(4.2.25),

Vt (; x) = (ex 1 x) ta (x); Yt (; x) = Zt (; x) =

© 2001 by Chapman & Hall/CRC

Zt

0

(ex

1 Ct (x) 2

1) ^sa (x) ds;

398

Maxingale problem

(we \bar" here Yt (; x) not to confuse it with earlier notation). All the quantities above are well de ned by the same argument as in Subsection 4.2.1. Then exactly as in the proof of Theorem 4.2.11 convergence (5.2.44) would hold provided for every T > 0 and " > 0

) )

) Æ)

sup jBt;a tT

1=r P a Bt (X )j !

0 as 2 ;

1 lim lim sup P1=r sup j Vt;Æ (r ) Vt (; X )j > " = 0; Æ!0 2 tT r 1 lim lim sup P1=r sup j Yt;Æ (r ) Yt (; X )j > " = 0; Æ!0 2 tT r 1 lim lim sup P1=r sup j Zt;Æ (r ) Zt (; X )j > " = 0: Æ!0 2 tT r

Limit ) is just (sup B a ) which we have already proved. For part ), we rst note that by part II.1 of Lemma 5.2.23 applied to fL(X^ ;a ); 2 g and a, in which hypotheses boundedness of the associated function f follows by (5.2.36) and (5.2.39), we have that 1 lim lim sup P1=r Vt;Æ (r ) Vt (; X ) > " = 0: Æ!0 2 r

Since by Theorem 3.2.3 and (5.2.39) the net fL((Vt (; X ); t 2 R + )); 2 g is C {exponentially tight, Lemma 5.2.18 implies ). We prove ) by the argument of the proof of part ) in Theorem 4.2.11 (a similar argument we have already used in the proof of Lemma 5.2.25). Theorem 3.2.3 implies in view of (5.2.2) and (5.2.39) that the net fL(Yt (; X ); t 2 R+ ); 2 g is C {exponentially tight. Therefore, by (5.2.46) we have, in view of Lemma 5.2.18, that ) would follow from 1 lim lim sup P1=r j Yt;Æ (r ) Yt (; X )j > " = 0; Æ!0 2 r " > 0; t 2 U: (5.2.47)

Next, noting that by (5.2.36) e jja 1 (er x 1) s;a ejja 1 and in view of (5.2.39), we have as in the proof of ) while proving

© 2001 by Chapman & Hall/CRC

399

Convergence of characteristics

Theorem 4.2.11 that (5.2.47) is implied by 1 X lim lim sup P1=r D;Æ (r )k Æ!0 2 r 0<st s Zt

0

and

(ex

1) ^sa (X ) k ds > = 0;

> 0; k = 2; 3; : : : ; (5.2.48)

1 X lim lim sup lim sup P1=r Ds;Æ (r )2 > A = 0; A!1 Æ!0 r 0<st 2 (5.2.49)

lim lim sup P1=r A!1 2

Zt

0

jex 1j^sa (X ) 2ds > A = 0;

(5.2.50) where t 2 U . Limit (5.2.50) follows by (5.2.39). Limit (5.2.49) is easily deduced from (^ a ) and (5.2.50). Limit (5.2.48) follows by part II.2 of Lemma 5.2.23, (5.2.36) and (5.2.39). Part ) is proved. To prove Æ), we introduce as in the proof of Theorem 4.2.11 1 2 ;a j (x 1(r jxj Æ) x;Æ s )j s 2 1 2 + j x;Æ s j (1 as ); 2 1 Wt;Æ () = j xj2 1(r jxj Æ) t;a;c: 2

Hs;Æ () =

Then by (5.2.45) and the de nitions of x;Æ s and as X 1 ^ ;a;Æ 1 Ct = Ct + Wt;Æ ()+ Hs;Æ () 2 2 0<st

so that by the de nitions of Zt;Æ () and Zt (; x) convergence (sup C a ) implies that Æ) would follow by 1 Æ0 ) lim lim sup P1=r sup j (Ut;Æ (r ) Wt;Æ (r ))j > " = 0; Æ!0 2 tT r " > 0; T > 0;

© 2001 by Chapman & Hall/CRC

400

Æ00 )

Maxingale problem

1 X ;Æ lim lim sup P1=r j ln(1 + G;Æ s (r )) (Hs (r ) Æ!0 2 r 0<st

+ ln(1 + Ds;Æ (r )))j > " = 0; " > 0; t > 0:

0

0

Limit Æ ) is proved as Æ ) in the proof of Theorem 4.2.11 if we note that by Theorem 5.2.20 the net fL(C (X )); 2 g is C { exponentially tight so that by (C a ) and Theorem 3.2.3 lim lim sup lim sup P1=r (jr C^t;a;Æ j > A) = 0: (5.2.51) A!1 Æ!0 2 As for Æ00 ), the argument is again as in the proof of Theorem 4.2.11. We rst note that by part II.3 of Lemma 5.2.23 and (5.2.36) Æ X j Ds;Æ (r )j > " = 0; " > 0: lim lim sup P1=r Æ!0 2 r 0<st Now the rest of the proof is the same as the proof of Æ00 ) in the proof of Theorem 4.2.11 (with the use of (5.2.51) in due place). Thus, ), ), ), and Æ) have been proved, and (5.2.44) has been proved. By Theorem 5.1.16, we have thus proved Theorem 5.2.22 under conditions (sup B ), (C ), ( ), and (^ ). The fact that (C ) can be replaced by (C~ ) follows by Lemma 5.2.24. Theorem 5.2.22 has been proved.

Theorem 5.2.26. Let the net fL(X ); 2 g be C {exponentially

tight, let B satisfy the uniform continuity condition, and let C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let also and ^ satisfy the C {local boundedness conditions (5.2.8) and (5.2.9). If conditions (0), (A) + (a), (sup B ), (C ) (respectively, (C~ )), ( ), and (^ ) hold, then every LD accumulation point of fL(X ); 2 g is a solution of (0; G). The idea of the proof is to \take the limit as a ! 1" in Theorem 5.2.22. The next lemma proves that limits in the weak topology of solutions to (M^ a ) solve (x0 ; G). Lemma 5.2.27. Let and ^ satisfy the C {local boundedness conditions (5.2.8) and (5.2.9), B satisfy the uniform continuity condition, and C , and ^ satisfy the continuity conditions. Let deviabilities ^ a ; a > 0; on C 0 solve (M^ a ). Let be a deviability on C and devia ^ on C 0 be de ned as ( ^ x; x0 ) = (x) 1(x = x0 ); (x; x0 ) 2 C 0 : bility iw ^ as a ! 1, then solves (0; G). If ^ a !

© 2001 by Chapman & Hall/CRC

401

Convergence of characteristics

Proof. We begin by proving that for every compact K C

lim sup sup jG^ as (; x) Gs (; x)j = 0; t > 0:

(5.2.52)

a!1 x2K st

By the de nitions of G^ a () and G()

jG^ as (; x) Gs(; x)j ejxj 1(jxj > a) s(x) Z t +

1) ^s (x) ds

1 + (eha (x)

0

Zt

1) ^s (x) ds (5.2.53)

1 + (ex

0

The rst term on the right tends to 0 as a ! 1 uniformly over x 2 K by (5.2.8). Since by (5.2.9) as in the proof of Lemma 4.2.17 lim sup sup ex 1(jxj > a) ^s(x) = 0; a!1 x2K st Z ha (x) 1) ^ (x) > 0; lim inf inf inf 1 + ( e s a!1 x2K st Rd

the second term on the right of (5.2.53) also tends to 0. Limit (5.2.52) has been proved. Now, for r 2 R+ and (x; x0 ) 2 C 0 we introduce the C0 {stopping times

r (x; x0 ) = inf ft 2 R+ : G (; x0 )_x +t rg (5.2.54) t

t

and

r;a (x; x0 ) = inf ft 2 R+ : G^ at (; x0 )_xt +t rg; a > 0: (5.2.55) By Lemma 5.2.25 G^ a () satis es the uniform continuity condition, and (5.2.52) then implies that the map x ! (Gt (; x); t 2 R+ ) is a continuous map from C into C (R + ; R). Therefore, by Lemma 5.1.9

r and r;a are continuous in (x; x0 ) 2 C 0 , also by (5.2.52) for every compact K 0 C 0 lim

sup

a!1 (x;x0 )2K 0

j r;a(x; x0 ) r (x; x0 )j = 0; r 2 R+ :

© 2001 by Chapman & Hall/CRC

(5.2.56)

402

Maxingale problem

^ a is a solution to (M^ a ), Y^ a () de ned by (5.2.41) is a C0 { Since local exponential maxingale on (C 0 ; ^ a ). It is also continuous in the time variable, hence, by part 2 of Lemma 2.3.13 and continuity of

r;a (x; x0 ) the function Y^ a;N () = Y^ta^ r;a (x;x0 ) (; (x; x0 )); t 2 R+ is a C0{local exponential maxingale on (C 0 ; ^ a ). Since by (5.2.41) and (5.2.55) Y^ a;N () is bounded, we conclude that Y^ a;N () is a C0 { uniformly maximable exponential maxingale on (C 0 ; ^ a ). Hence, for all 0 s < t and every R+ -valued continuous and bounded Cs0 { measurable function f (x; x0 ) ^ a (x; x0 ) sup Y^ta^ r;a (x;x0 ) (; (x; x0 ))f (x; x0 ) (x;x0 )2C 0 ^ a (x; x0 ): (5.2.57) = sup Y^sa^ r;a(x;x0 ) (; (x; x0 ))f (x; x0 ) (x;x0 )2C 0 We now prove that the equality is preserved on taking in both sides the limits as a ! 1. More precisely, we prove that for every R+ valued bounded continuous function f (x; x0 ) on C 0 and t > 0 lim sup Y^ a r;a 0 (; (x; x0 ))f (x; x0 )^ a (x; x0 ) a!1 (x;x0 )2C 0 t^ (x;x ) ^ x; x0 ); (5.2.58) = sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( 0 0 (x;x )2C where Y 0 () is de ned by (5.1.4), i.e., Yt0 (; (x; x0 )) = exp( xt Gt (; x0 )): As a rst step, we prove that lim sup jY^ a r;a 0 (; (x; x0 )) Yt0^ r (x;x0 ) (; (x; x0 ))j a!1 (x;x0 )2C 0 t^ (x;x ) ^ a (x; x0 ) = 0: (5.2.59) Let K 0 be a compact in C 0 . By (5.2.56) and Arzela{Ascoli's theorem lim sup jx r;a 0 xt^ r (x;x0 ) j = 0; a!1 (x;x0 )2K 0 t^ (x;x ) and by (5.2.52), (5.2.56) and continuity of Gt (; x) in (t; x) 2 R+ C lim sup jG^ a r;a 0 (; x0 ) Gt^ r (x;x0 ) (; x0 )j = 0; a!1 (x;x0 )2K 0 t^ (x;x ) which imply that lim sup jY^ a r;a 0 (; (x; x0 )) Yt0^ r (x;x0 ) (; (x; x0 ))j = 0: a!1 (x;x0 )2K 0 t^ (x;x ) (5.2.60)

© 2001 by Chapman & Hall/CRC

Convergence of characteristics

403

Let fak ; k 2 N g be a subsequence, along which lim supa!1 of the iw ^ supremums in (5.2.59) is attained. Since ^ ak ! as k ! 1, by a ^ k Theorem 1.9.27 the sequence f ; k 2 N g is tight. Given " > 0, we choose K 0 such that lim supk!1 ^ ak (C 0 nK 0 ) < ": Then, since Y^ta^ r;a (x;x0 ) (; (x; x0 )) and Yt0^ r (x;x0 ) (; (x; x0 )) are bounded above by e(1+jj)r , lim sup sup Y^ta^k r;ak (x;x0 ) (; (x; x0 ))^ ak (x; x0 ) < "e(1+jj)r ; k!1 (x;x0 )2C 0 nK 0 ^ ak (x; x0 ) < "e(1+jj)r : lim sup sup Yt0^ r;ak (x;x0 ) (; (x; x0 )) k!1 (x;x0 )2C 0 nK 0 These inequalities and (5.2.60) imply (5.2.59) (recall that ^ a (x; x0 ) 1). iw ^ Next, using the convergence ^ a ! and the fact that 0 0 Yt^ r (x;x0 ) (; (x; x )) is bounded and continuous in (x; x0 ) 2 C 0 , we have by the de nition of idempotent weak convergence lim sup Y 0 r 0 (; (x; x0 ))f (x; x0 )^ a (x; x0 ) a!1 (x;x0 )2C 0 t^ (x;x ) ^ x; x0 ); = sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( (x;x0 )2C 0 which by (5.2.59) concludes the proof of (5.2.58). Equalities (5.2.57) and (5.2.58) imply that ^ x; x0 ) sup Yt0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( 0 0 (x;x )2C ^ x; x0 ); = sup Ys0^ r (x;x0 ) (; (x; x0 ))f (x; x0 )( (x;x0 )2C 0 ^ x; x0 ) = (x) 1(x = x0 ), we conclude that and, since ( (Yt^ r (x;x) (; x); x 2 C ; t 2 R+ ) satis es the maxingale property with respect to . Being bounded, it is a C{uniformly maximable exponential maxingale on (C ; ). Since also (Yt (; x); x 2 C ; t 2 R+ ) is C{adapted and r (x; x); x 2 C ; is a continuous C{stopping time, it follows that Y () is a C{local exponential maxingale on (C ; ). The equality (x0 6= 0) = 0 holds since ^ a ((x; x0 ) : x0 6= 0) (( ^ x; x0 ) : x0 6= 0) 0 = lim inf a!1 = (x : x0 6= 0):

© 2001 by Chapman & Hall/CRC

404

Maxingale problem

Proof of Theorem 5.2.26. By (5.2.7) and Lemma 5.2.24 it is suÆcient to prove the part of the statement that concerns C . Let be a deviability on D supported by C that is an LD accumulation point of fL(X ); 2 g so that along a subnet ld L(X 0 ) ! :

(5.2.61) By Theorem 5.2.20 the nets fL(X ; X ); 2 g are C 0 { exponentially tight for all a > 000. Therefore, by Corollary 3.1.20 00 ;a ^ there exists a subnet f(L(X ; X ); a > 0); 00 2 00 g of 0 ;a 0 ^ f(L(X ; X ); a > 0); 0 2 0g such that for every a > 0 ^ 0 ;a

0

ld ^ a L(X^ 00 ;a ; X 00 ) ! ;

(5.2.62)

ld ^ L(X 00 ; X 00 ) ! ;

(5.2.63)

where ^ a are deviabilities on D 0 supported by C 0 . Convergence (5.2.61) implies that in D 0 ^ is the deviability on D 0 de ned by where ^ x; x0 ) = (x) 1(x = x0 ); (x; x0 ) 2 D 0 : ( Also by Lemma 4.2.16 we have that for every > 0 1=r 0 ^ ;a ; X )) > = 0; lim lim sup P (( X ; X ) ; ( X S a!1 2

(5.2.64) where 0S is the Skorohod{Prohorov{Lindvall metric on D 0 . Limits (5.2.62), (5.2.63) and (5.2.64) imply by Lemma 3.1.37 that iw ^ ^a ! ^ a solves problem (M^ a ). as a ! 1: By Theorem 5.2.22 An application of Lemma 5.2.27 ends the proof.

5.2.3 Proofs of the main results Proof of Theorem 5.2.9. By the majoration conditions in the theorem and Theorem 3.2.3 the nets fL(B (X )); 2 g, fL(C (X )); 2 g (respectively, fL(C~ (X )); 2 g), and fL(f (X )); 2 g; f 2 Cb ; are C {exponentially tight. Then by Theorem 5.2.20 the net fL(X ); 2 g is C {exponentially tight. Also the continuity and majoration conditions on B imply the uniform continuity condition. An application of Theorem 5.2.26 concludes the proof.

© 2001 by Chapman & Hall/CRC

405

Convergence of characteristics

Proof of Theorem 5.2.12. The argument is similar to the one in the proof of Theorem 5.1.10. We give only the main points. Let X ;N ; 2 ; N 2 N ; be de ned by (5.1.21). Then it is veri ed analogously to the proof of Theorem 5.1.10 that the net fX ;N ; 2 g satis es the conditions of Theorem 5.2.9 for every N 2 N with Bt (x); Ct (x); C~t (x); t (dx; x), and ^t (dx; x) replaced, respectively, by BtN (x); CtN (x); C~tN (x); tN (dx; x), and ^tN (dx; x) de ned as

BtN (x) = Bt^N (x) (x); CtN (x) = Ct^N (x) (x); C~tN (x) = C~t^N (x) (x); tN (dx; x) = t (dx; x) 1(t N (x)); ^tN (dx; x) = ^t (dx; x) 1(t N (x)):

In particular, the majoration conditions are checked in a manner similar to the proof of Theorem 5.1.10; checking the continuity conditions is simple (note, however, that the proof of the continuity condition for ^N uses the inequality ^s(Rd ; x) 1). By Theorem 5.2.9 the nets fL(X ;N ); 2 g; N 2 N ; are C { exponentially tight. Let N ; N 2 N ; be their respective LD accumulation points. It follows as in the proof of Theorem 5.1.10 that N solves (M N ). Condition (NE ) yields by Lemma 5.1.17 lim lim sup P1=r (N (X ) t) = 0; t 2 R+ ;

N !1 2

which implies, again as in the proof of Theorem 5.1.10, that fL(X ); 2 g is C {exponentially tight. The uniform continuity condition for B follows by the continuity and local majoration conditions. An application of Theorem 5.2.26 ends the proof. Proof of Theorem 5.2.15. By the de nition of the cumulant for 0 < s 0, we introduce

Xt;A = Xt

X

0<st

Xs 1(jXs j > A):

Since X ;A = (Xt;A ; t 2 R+ ) has bounded jumps, it satis es the Cramer condition. Let G;A () = (G;A t (); t 2 R + ) be the asso;A ciated stochastic cumulant and E () = (Et;A (); t 2 R+ ) be the stochastic exponential of G;A (). Then the following holds.

Theorem 5.3.1. Let the cumulant G() satisfy the uniform continuity condition and the linear-growth condition. Let condition (0) hold and, for some A > 0,

© 2001 by Chapman & Hall/CRC

407

LD convergence results

([0; t]; jxj

(A0 ) (sup

E A)

1=r P 1 =r > A)

! 0

as 2 ; t > 0;

1=r

1 )j P! 0 sup j ln Et;A ( r ) G ( ; X t^N (X ) ^N (X ) tT r as 2 ; T > 0; N 2 N ; 2 Rd :

Then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves the maxingale problem (x0 ; G). If the ld latter problem has the unique solution x0 , then L(X ) ! x0 . Proof. According to the proof of Lemma 4.2.16 condition (A0 ) implies that

P sup jXt Xt;A j > 0 tT

Therefore, condition (sup

! 0 as 2 ; T > 0:

(5.3.1)

E A) implies that 1=r

P 1 ;A ;A sup j ln Et^N (X ;A ) (r ) Gt^N (X ;A ) (; X )j ! 0 tT r as 2 ; T > 0; N 2 N ; 2 Rd :

By Theorem 5.1.12 the net fL(X ;A ); 2 g is C {exponentially tight, and its every LD accumulation point solves the maxingale problem (x0 ; G). By Theorem 3.2.3 and (5.3.1) the net fL(X ); 2 g is C {exponentially tight. By Lemma 3.1.38 and (5.3.1) every LD accumulation point of fL(X ); 2 g is an LD accumulation point of fL(X ;A ); 2 g. Theorems 5.1.5 and 5.1.10 admit similar versions. We next concentrate on consequences of Theorem 5.2.15 as the most useful one for applications. Since considerations below are along the lines of the content of Section 4.3 and the proofs use similar ideas, we omit details. As in Section 4.3, we begin with integrable versions when one can consider nontruncated characteristics. We introduce the following localised versions of conditions (I1 ) and (I2 ). (I1 )loc

lim lim sup P1=r jxj 1(r jxj > a) t^N (X ) > " = 0; a!1 2 " > 0; t > 0; N 2 N ;

© 2001 by Chapman & Hall/CRC

408

Maxingale problem

lim lim sup P1=r r jxj2 1(r jxj > a) t^N (X ) > " = 0; a!1 2 " > 0; t > 0; N 2 N :

(I2 )loc

Clearly, (I2 )loc implies (I1 )loc . We recall that the modi ed second characteristic without truncation C~ 0 = (C~t0 (x); t 2 R+ ; x 2 C ) of X is speci ed by the equalities

C~t0 (x) = Ct (x)+(x)2 t (x)

Zt

0

(x^s (x))2 ds; 2 Rd :

Lemma 5.3.2.

1. Let the X be special semimartingales. If, in addition, condition (I1 )loc holds, then condition (sup B )loc is equivalent to the condition

(sup B 0 )loc

sup jB 0 t^N (X ) tT

B0

t^N (X )

1=r P (X )j !

as 2 ; T > 0; N

0

2 N:

2. Let the X be locally square-integrable semimartingales. If, in addition, condition (I2 )loc holds, then condition (C~ )loc is equivalent to the condition

(C~ 0 )loc

kr C~t0^N (X )

C~t0^N (X ) (X )k

1=r P

! 0 as 2 ; t 2 U; N 2 N :

The proof is similar to the proof of Theorem 4.3.2. As in Section 4.3, in view of the lemma Remark 4.3.3 applies to the setting of this section as well. We now introduce simpli ed versions of the other conditions. For (^ )loc and ( )loc , we consider the conditions (QC )loc

(MD)loc

1=r

X P 1 2 (fsg; fr jxj > g) ! 0 as 2 ; r 0<st^ (X ) N t > 0; > 0; N 2 N ;

1=r

P 1 ([0; t ^ N (X )]; fr jxj > g) ! 0 as 2 ; r t > 0; > 0; N 2 N :

© 2001 by Chapman & Hall/CRC

LD convergence results

409

Obviously, condition (QC )loc implies condition (^ )loc with ^( ; x) = 0. Condition (MD)loc , which is stronger than (QC )loc, implies both (^ )loc and ( )loc with ( ; x) = ^( ; x) = 0 . It thus de nes the case of moderate deviations considered in more detail below for the Markov setting. Then by Theorem 5.2.15 we have the following generalisation of Corollary 4.3.4. Theorem 5.3.3. Let the limiter h(x) be continuous, and B , C (respectively, C~ ), , and ^ satisfy the continuity conditions. Let the linear-growth conditions (5.2.10), (5.2.11) and (5.2.12) hold. If conditions (0); (A)loc + (a)loc ; (sup B )loc; (C )loc (respectively, (C~ )loc ), ( )loc , and (QC )loc hold, then the net fL(X ); 2 g is C { exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant 1 Gt (; x) = Bt0 (x)+ Ct (x)+(ex 1 x)t (x): 2 If the latter problem has the unique solution x0 (e.g., either Theold rem 2.8.33 or Theorem 2.8.34 applies), then L(X ) ! x0 as 2 : By Remark 4.3.3 the theorem also has locally integrable and locally square-integrable versions. R t The next result follows by Theorem 5.2.15 with Bt (x) d= 0 us (x) ds, Ct (x) = 0, and s ( ; x) = us (x)1(1 2 ); 2 B (R ), and Theorem 2.8.10. It extends Corollary 4.3.5. Theorem 5.3.4. Let d = 1 and conditions (0), (A)loc + (a)loc and (QC )loc hold. Let the limiter h(x) be continuous at x = 1. Let (us (x); x 2 D ; s 2 R+ ) be a D-progressively measurable R+ -valued function, which is C {continuous in xR and satis es the linear-growth condition us (x) (1 + xs )ls , where 0t ls ds < 1; t 2 R+ . If t^ZN (X ) 1=r P sup jBt^N (X ) h(1) us(X )dsj ! 0 as 2 ; tT 0 T > 0; N 2 N ; 1=r ;Æ lim lim sup P (r kCt^N (X ) k > ") = 0; Æ!0 2 t 2 U; " > 0; N 2 N ;

© 2001 by Chapman & Hall/CRC

410

Maxingale problem

and, for all " 2 (0; 1=2); t 2 R+ and N

2 N ; as 2 ,

1 ([0; t ^ N (X )]; fjr x 1j < "g) r 1 ([0; t ^ N (X )]; fjr xj > "g r

t^ZN (X )

us

1=r P (X )ds !

0 \

0;

1=r P

fjr x 1j > "g) ! 0;

then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumuR lant Gt (; x) = (e 1) 0t us (x) ds: If the latter problem has a unique solution (e.g., by Theorem 2.8.28 inf inf us(x) > 0 and st x2K sup sup us (x) < 1 for every compact K C and t 2 R+ ), then st x2K ld X ! X as 2 ; where X is the Luzin-continuous idempotent Poisson process of rate us(X ) starting at x0 with idempotent distribution x0 , whose density is given by Z1 x0 (x) = exp sup x_ t (e 1)ut (x) dt 2R 0

if x is absolutely continuous and x0 = x0 , and x0 (x) = 0 otherwise.

Let us introduce the following conditions. (sup B00 )loc

sup jBt^N (X ) tT

Bt0^N (X ) (X )j

1=r P

! 0 as 2 ; T > 0; N 2 N ;

(C0 )loc

kr C~t^N (X )

1=r P Ct^N (X ) (X )k !

(C 0 )loc

kr C~t0^N (X )

1=r P Ct^N (X ) (X )k !

0

(L2 )loc

r j

1=r P (r jxj > ) t^N (X ) !

j1

x2

0 as 2 ; t 2 U; N 2 N ;

© 2001 by Chapman & Hall/CRC

0 as 2 ; t 2 U; N 2 N ;

0 as 2 ; t > 0; N 2 N ; > 0:

411

LD convergence results

Note that the latter condition is a localised version of the Lindeberg condition. As above, it implies both (I2 )loc and (MD)loc and allows us to do without truncation. The following result extends Corollaries 4.3.7 and 4.3.8. The proof is similar and also uses Theorem 2.8.9.

Theorem 5.3.5. Let the functions (bs(x); s 2 R+ ; x 2 D ) and (cs (x); s 2 R+ ; x 2 D ) be C {continuous and satisfy the linear-growth conditions

jbt (x)j lt(1+ xt ); kct (x)k lt (1+ xt 2); R

where lt is Lebesgue measurable and 0t ls ds < 1; t 2 R+ : Let conditions (0) and (A)loc + (a)loc hold. If, in addition, either conditions (MD)loc , (sup B00 )loc and (C0 )loc hold for some limiter h(x) or conditions (L2 )loc , (sup B 0 )loc , and (C00 )loc hold, then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant 1 Gt (; x) = Bt0 (x)+ Ct (x): 2 If, in addition, uniqueness holds for problem (x0 ; G) (e.g., according to Theorem 2.8.21, inf inf inf cs (x) > 0 and sup sup kcs (x)k < jj=1 st x2K st x2K ld 1 for every t 2 R+ and compact K C ), then X ! X as 2 , where X = (Xt ; t 2 R+ ) is the Luzin-continuous idempotent diusion

X_ t = bt (X )+ c1t =2 (X )W_ t ; X0 = x0 ; whose deviability distribution is given by 1 Z1 x0 (x) = exp 2 (x_ t bt (x))ct (x) (x_ t bt (x)) dt 0

if x0 = x0 , x is absolutely continuous and x_ t of ct (x) a.e., and x0 (x) = 0 otherwise.

bt (x) is in the range

Now, we turn our attention to conditions (A)loc + (a)loc . Let us introduce the condition (V S0 )loc

([0; t ^

N

© 2001 by Chapman & Hall/CRC

(X )]; fr

jxj > g

1=r P 1 =r )

! 0 as 2 ; t > 0; > 0; N 2 N :

412

Maxingale problem

Condition (V S0 )loc implies both (A)loc + (a)loc and (MD)loc so that by Lemma 5.3.2 and Theorem 5.3.5 we have the following result.

Theorem 5.3.6. Let the X be locally square-integrable.

Then the assertion of Theorem 5.3.5 holds if instead of conditions (0), (A)loc + (a)loc , (MD)loc , (sup B00 )loc , and (C0 )loc one requires conditions (0), (I2 )loc , (V S0 )loc , (sup B 0 )loc , and (C00 )loc .

Conditions (A)loc + (a)loc are also implied by the following condition (V S )loc , which is weaker than (V S0 )loc : (V S )loc

lim lim sup P1=r ([0; t ^ N (X )]; fr jxj > ag)1=r > " a!1 2 = 0; t > 0; " > 0; N 2 N :

For the sequel, we note that conditions (A)loc + (a)loc are implied by the conditions (A0 )loc (a0 )loc

P

1=r

([0; t ^ N (X )]; jxj > A)1=r ! 0 as 2 ; t > 0; N 2 N ; 9A > 0; 1=r

P 1 r jxj e 1(r jxj > a) 1(jxj A) t^N (X ) ! 0 r as 2 ; t > 0; > 0; A > 0; N 2 N ; 9a > 0:

If, in addition, the convergence in (a0 )loc holds for every a > 0, then (MD)loc holds. Let us assume that the Cramer condition (Cr) holds, i.e., j x e j 1(jxj > 1) t < 1; t > 0; > 0: Then moment conditions can be used to check (A)loc + (a)loc . More speci cally, let us introduce the conditions 1=r 1 r jxj e 1 ( r j x j > a ) > " (Ie )loc alim lim sup P t ^ ( X ) N !1 2 r = 0; t > 0; " > 0; > 0; N 2 N ; (Le )loc

1=r

P 1 r jxj e 1 (r jxj > ) t^N (X ) ! 0 as 2 ; r t > 0; " > 0; N 2 N :

© 2001 by Chapman & Hall/CRC

413

LD convergence results

Note that (Le )loc is an exponential analogue of the Lindeberg condition. Then (Le )loc ) (Ie )loc ) (A)loc + (a)loc ; (Ie )loc ) (I2 )loc and (Le )loc ) (L2 )loc ) (MD)loc . In particular, we can check (A)loc +(a)loc by checking (Ie )loc (e.g., in Theorem 5.3.3). As another illustration, Theorem 5.3.4 allows us to state the following extension of part b) of Corollary 4.3.12.

Theorem 5.3.7.

Let Xt = Nt =r , where N = (Nt ; t 2 R+ ) are one-dimensional point processes with respective compensators A = (At ; t 2 R+ ). Let (us (x); x 2 D ; s 2 R+ ) be a Dprogressively measurable R+ -valued function, which is C {continuous in x and R t satis es the linear-growth condition us (x) (1 + xs )ls , where 0 ls ds < 1; t 2 R+ . Let the maxingale problem (0; G) with R cumulant Gt (; x) = (e 1) 0t us (x) ds has the unique solution 0 (e.g., by Theorem 2.8.28 inf inf us (x) > 0 and sup sup us (x) < 1 st x2K st x2K for every compact K C and t 2 R+ ). Let X be the Poisson idempotent process of rate us(X ) with idempotent distribution 0 . If, as 2 ,

1 A r t^N (X )

t^ZN (X )

us

1=r P (X )ds !

0

0

and 1=r

X P 1 (As )2 ! 0; r 0<st^ (X ) N

ld then X ! X as 2 :

Since (Le )loc implies (A)loc + (a)loc , (I2 )loc and (MD)loc , Lemma 5.3.2 and Theorem 5.3.5 result in the following version of Theorem 5.3.6.

Theorem 5.3.8.

Let the Cramer condition hold. Then in Theorem 5.3.5 one can replace conditions (A)loc + (a)loc , (MD)loc , (sup B00 )loc , and (C0 )loc with conditions (Le )loc , (sup B 0 )loc , and (C00 )loc .

© 2001 by Chapman & Hall/CRC

414

Maxingale problem

5.4 Large deviation convergence of Markov processes We now consider implications of the above results for the Markov setting. In the next theorem we assume that X are generally speaking non-time homogeneous \continuous-time Markov processes with generators At " in the sense that the At map the functions (ex ; x 2 R d ); 2 R d ; into B(R + ) B (R d )=B (R )-measurable functions of (t; x) R and the processes (exp( Xt ) exp( X0 ) 0t As exp( Xs ) ds; t 2 R + ) are well-de ned local martingales on ( ; F ; F ; P ). Let S denote the state space of X .

Theorem 5.4.1. d

Let gt (; x); t 2 R+ ; 2 Rd ; x 2 Rd , be a B(R+ ) B(R ) B(Rd )=B(R)-measurable R-valued function such that gt (0; x) = 0 . Let us assume that gt (; x) is continuous in x and meets the linear-growth condition gt (; x) kt (jj(1 + jxj)) , where t 2 R+ , 2 Rd , and the function kRt () is R+ -valued, Lebesgue measurable in t, increasing in , and 0t ks () ds < 1; t 2 R+ ; 2 R+ . If, as 2 ,

1=r P X0 ! x0

and, for T > 0; N

2 N,

1 sup sup j exp( r x)At exp(r x) gt (; x)j ! 0; tT x2S :jxjN r then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves maxingale problem (x0 ; G) associated with the cumulant

Gt (; x) =

Zt

0

gs (; xs ) ds: R

Proof. Since the processes (exp(r (Xt X0 )) exp( 0t exp( r Xs )As exp(r Xs )ds); t 2 R+ ) are local martingales on ( ; F ; F ; P ) , Liptser and Shiryaev [79, Theorem 2.5.1] (see also Ethier and Kurtz [48, Lemma 3.2]), the claim follows by Theorem 5.1.13.

We next consider applications of Theorem 5.2.15. The semimartingales X are assumed to be Markov processes. To simplify

© 2001 by Chapman & Hall/CRC

415

LD convergence of Markov processes

the conditions, we distinguish between the continuous- and discretetime cases. Let f ; 2 g and f ; 2 g be nets of real numbers tending to 1 as 2 . In the continuous-time case, we assume that the predictable triplets of X corresponding to a limiter h(x) are given by:

Bt

=

Ct =

1 r

([0; t]; dx) =

Zt

0

Zt

bs (Xs )ds;

(5.4.1)

cs (Xs )ds;

(5.4.2)

s ( dx; Xs )ds;

(5.4.3)

0 Zt 0

where bs (u) is an R d {valued B (R + ) B (R d )=B (R d ){measurable funcRt tion such that 0 jbs (xs )j ds < 1 for t 2 R+ and x 2 D , cs (u) is a function with values in the space of symmetric positive semi-de nite d d {matrices, which is B(R+ ) B(Rd )=B(R dd ){measurable and Rt such that 0 kcs (xs )k ds < 1 for t 2 R+ and x 2 D , s (dx; u) is a transition kernel from (R+ Rd ; B(R+ ) B(Rd )) into (Rd ; B(Rd )) such that

t (f0g; u) Zt Z

0

Rd

= 0;

Z

Rd

1 ^ jxj2 t (dx; u) < 1;

1 ^ jxj2 s (dx; xs ) ds < 1; t 2 R+ ; x 2 D ; u 2 Rd :

In the discrete-time case, we assume that X is a pure jump process with predictable measure of jumps bX tc ( dx; X ([0; t]; dx) = ^i= (5.4.4) (i 1)= ); i=1 where ^i= (dx; u), for every i 2 N ; (Rd ; B(Rd )) into (Rd ; B(Rd )) such that

is a transition kernel from

(f0g; u) = 0; ^ (R d ; u) 1; i 2 N ; u 2 R d : ^i= i=

© 2001 by Chapman & Hall/CRC

416

Maxingale problem

The parameter can be interpreted as the frequency of jumps of X and the parameter 1= as the size of the jumps. Depending on the relative speed at which and go to 1, there are two dierent asymptotics, which are referred to as \very large deviations", when and are of the same order and one takes r = = , and \moderate deviations", when = ! 1 but = 2 ! 0; and one takes r = 2 = . Let us recall the de nition of the essential supremum of a collection of measurable functions, see, e.g., Neveu [94, II.4]. Let fj = (fj (x); x 2 R+ ); j 2 J; be a collection of B(R+ )=B(R){measurable R -valued functions on R + ; for a B(R + )=B (R ){measurable R -valued function f = (f (x); x 2 R+ ), we say that f = ess sup j 2J fj if, for every j 2 J , f (x) fj (x) for almost all x 2 R+ and f (x) g(x) for almost all x 2 R+ for every B(R+ )=B(R){measurable R-valued function g = (g(x); x 2 R+ ) such that, for every j 2 J , g(x) fj (x) for almost all x 2 R+ . Note that this usage is dierent from the interpretation above. We consider, rst, the case of very large deviations: = = r . Let us introduce the following versions of the Cramer condition: ess sup jujv Zt

0

Z

(ex

Rd

ess sup jujv

Z

1 x) t (dx; u) < 1;

(ex

Rd

1 x) s (dx; u) ds < 1;

2 Rd ; t 2 R+ ; v 2 R+ ; (5.4.5)

in the continuous-time case, and Z

Rd

(dx; u) < 1; 2 R d ; i 2 N ; u 2 R d ; ex ^i=r

(5.4.6)

in the discrete-time case. Under these conditions, the X are locally square integrable semimartingales, so, we may and will take h(x) = x so that B is the rst characteristic \without truncation" (B = B 0 ), and de ne Z 1 g s (; u) = bs (u)+ cs (u)+ (ex 1 x) s (dx; u) 2 Rd

© 2001 by Chapman & Hall/CRC

417

LD convergence of Markov processes

in the continuous-time case and Z g s (; u) = ln 1+ (ex 1) ^(br sc+1)=r (dx; u) Rd

in the discrete-time case. Let gs (; u) be a B(R+ ) B(Rd ) B(Rd )=B(R)-measurable function, which is continuous in u and satis es the following lineargrowth condition

jgs(; u)j g~s(jj(1+ juj)); whereR g~s (y) is R+ -valued, B(R+ )=B(R)-measurable in s, increasing in y, 0t g~s (y) ds < 1; t 2 R+ ; y 2 R+ ; and gs (0; u) = 0 . We then have the following version of Theorem 5.4.1. Theorem 5.4.2. Let = = r and the Cramer condition (5.4.5) in the continuous-time case, respectively, the Cramer con1=r P X0 !

dition (5.4.6) in the discrete-time case, hold. Let 2 . If, for all 2 Rd , t 2 R+ and v 2 R+ , as 2 , Zt

0

x0 as

ess sup jg s (; u) g s(; u)j ds ! 0; jujv

then the net fL(X ); 2 g is C {exponentially tight, and its every LD accumulation point solves problem (x0 ; G) with cumulant

Gt (; x) =

Zt

0

g s(; xs ) ds:

If the latter problem has the unique solution x0 (e.g., the conditions ld of Theorem 2.8.32 hold for gs (; x) = g s (; xs )), then L(X ) ! x0 as 2 : Proof. We have the following representations for the stochastic exponential E () = (Et (); t 2 R+ ); 2 Rd ; associated with X . In the continuous-time case,

1 ln Et (r ) = r

Zt

0

© 2001 by Chapman & Hall/CRC

gs (; Xs ) ds:

418

Maxingale problem

In the discrete-time case, by the equalities g s = g br sc=r and Xs = Xbr sc=r , 1 1 ln Et (r ) = r r

brX tc 1 i=0

Therefore, in both cases for N 1 r

ln Et^N (X ) (r ) Zt ess sup jgs(; u) jujN 0

)= g i=r (; Xi=r

2N

brZtc=r 0

gs (; Xs ) ds:

Gt^N (X ) (; X )

g s (; u)j ds +

Zt

br tc=r

sup jgs (; u)j ds; jujN

and the claim follows by Theorem 5.1.12. We next state a \very large deviation" result in terms of characteristics, which does not require the Cramer condition. We con ne ourselves to the continuous-time case. We assume all the above conditions on bs (u), cs (u) and s(dx; u) to hold except the Cramer condition (5.4.5). Instead, we assume that the rst characteristic B corresponds to a continuous limiter h(x). We de ne positive semide nite symmetric matrices c~s (u) by

c~s (u)

=

cs (u) +

Z Rd

( h (x))2 s (dx; u); 2 Rd :

We next introduce the limit idempotent process. Let bs(u) be an Rd {valued B(R+ ) B(Rd )=B(R d ){measurable function, cs (u) be a B(R+ ) B(Rd )=B(R dd ){measurable function with values in the space of positive semi-de nite symmetric d d {matrices and s (dx; u) be a transition kernel from (R+ Rd ; B(R+ ) B(Rd )) into

© 2001 by Chapman & Hall/CRC

419

LD convergence of Markov processes

(Rd ; B(Rd )) such that

s (f0g; u) = 0;

Z

(ex

1 x) s (dx; u) < 1;

Rd

2 Rd ; s 2 R+ ; u 2 Rd :

Let the following linear-growth conditions be satis ed:

jbs(u)j ls(1+juj); kcs(u)k ls(1+juj2 ); Z

(ejxj

Rd

(5.4.7)

1 jxj)s (dx; u) Z

(ejxj(1+juj)

1 jxj(1 + juj))ms (dx); 2 R+ ;

Rd

whereR ls is an R+ -valued B(R+ )=B(R+ ){measurable function such that 0t ls ds < 1 and ms(dxR) isR a transition kernel from (R+ ; B(R+ )) into (Rd ; B(Rd )) such that 0t Rd (exp(jxj) 1 jxj)ms (dx)ds < 1; t 2 R+ ; 2 R+ . In addition, we assume that the functions

u ! bs (u); u ! cs (u); u !

Z

f (x)s (dx; u);

Rd

for f continuous and such that jf (x)j 1 ^ jxj2 ; are continuous in u 2 Rd . We also de ne positive semi-de nite symmetric matrices c~s (u) by

c~s (u) = cs (u) +

Z

( h(x))2 s (dx; u); 2 Rd :

Rd

Theorem 5.4.3. Let = = r and the above conditions hold. © 2001 by Chapman & Hall/CRC

420 Let Zt

0

Maxingale problem

1=r P X0 ! x0

ess sup jujv

and, for all t 2 R+ and v 2 R+ ,

jbs(u)

bs (u)j ds ! 0;

Zt

0

Zt

Z ess sup jujv d 0 R

f (x)s (dx; u)

ess sup kc~s (u) c~s (u)k ds ! 0; jujv

Z

f (x)s(dx; u) ds

! 0; f 2 Cb:

Rd

Let also Z t

lim lim sup a!1 2

0

ess sup s (fjxj > ag; u) ds jujv

1=r

= 0;

v 2 R+ ; t 2 R+ : (5.4.8)

Then the net fL(X ); 2 g is C {exponentially tight. If is an LD accumulation point of fL(X ); 2 g , then the canonical idempotent process X is a Luzin-continuous semimaxingale with local characteristics (b; c; ; 0) on (C ; C; ). If the idempotent distribution Li(X ) of X is speci ed uniquely (e.g., Theorem 2.8.34 applies), then ld Li(X ) = x0 and X ! X as 2 : If, in addition, the Cramer condition (5.4.5) holds, then condition (5.4.8) can be replaced with the condition

lim lim sup a!1 2

Zt

Z

0

Rd

ess sup jujv

ejxj 1(jxj > a)s (dx; u) ds = 0; t 2 R+ ; 2 R+ ; v 2 R+ : (5.4.9)

In the latter case we can take h(x) = x, i.e., consider nontruncated characteristics. Proof. Condition (5.4.8) implies condition (V S )loc , and condition (5.4.9) implies condition (Ie )loc . Either one of these conditions implies conditions (a)loc + (A)loc . Therefore, the claim follows by Theorem 5.2.15. We can consider nontruncated characteristics under (5.4.9) by Lemma 5.3.2 and the fact that (Ie )loc implies (I2 )loc .

© 2001 by Chapman & Hall/CRC

LD convergence of Markov processes

421

We consider now moderate deviations so ! 1 and ! 1 in such a way that = ! 1 and = 2 ! 0. Let r = 2 = . We assume the locally square-integrable case, i.e., ess sup jujv Zt

0

Z

jxj2 t(dx; u) < 1;

Rd

ess sup jujv

Z Rd

jxj2 s(dx; u) ds < 1; t 2 R+ ; v 2 R+ ; 2 ;

in the continuous-time case, and Z

(dx; u) < 1; i 2 N ; u 2 R d ; 2 ; jxj2^i=

Rd

in the discrete-time case. In the discrete-time case we also assume that the X are martingales, i.e., Z

(dx; u) = 0; i 2 N ; u 2 R d ; 2 : x^i=

Rd

Then the X are locally square integrable semimartingales, so we choose nontruncated predictable characteristics. According to (5.4.2), (5.4.3), (5.4.4), and the equality r = 2 , the (nontruncated) modi ed predictable second characteristics of the X are of the form: in the continuous-time case

C~ 0 = t

2

Zt

0

c~0s (Xs ) ds;

where c~0s (u) are positive semi-de nite symmetric matrices de ned by

c~0s (u) = cs (u) +

Z

Rd

© 2001 by Chapman & Hall/CRC

( x)2 s (dx; u);

422

Maxingale problem

in the discrete-time case bZtc= 0 C~t = 2 c~0s (Xs ) ds; 0

where c~0s (u) are positive semi-de nite symmetric matrices de ned by Z 0 c~s (u) = ( x)2 ^(b sc+1)= (dx; u): Rd

Let us assume that bs(u) and cs (u) satisfy the conditions stated before Theorem 5.4.3 (i.e., measurability, linear growth and continuity in u). We introduce the following conditions on the predictable measures of jumps and rates of convergence. (P) For some Æ > 0 lim sup

Zt

0

ess sup jujv

Z

jxj2+Æ s(dx; u) ds < 1; t 2 R+ ; v 2 R+ ;

Rd

in the continuous-time case, respectively, b tc Z 1 X (dx; u) < 1; t 2 R ; v 2 R ; sup jxj2+Æ ^i= lim sup + + i=1 jujv d R

in the discrete-time case, and 2 =( ln ) ! 0 as 2 . (SE) For some 2 (0; 1] and > 0 lim sup

Zt

0

ess sup jujv

Z

Rd

exp( jxj )s (dx; u) ds < 1;

t 2 R+ ; v 2 R + ;

in the continuous-time case, respectively, b tc Z 1 X (dx; u) < 1; lim sup sup exp( jxj )^i= i=1 jujv d R

© 2001 by Chapman & Hall/CRC

t 2 R+ ; v 2 R + ;

423

LD convergence of Markov processes

in the discrete-time case, and 2 = ! 0 as 2 . The next theorem extends Theorem 4.4.8.

Theorem 5.4.4. Let = ! 1 and = 2 ! 0 as 2 .

Let either condition (P) or condition (SE) hold. Let the law of a Luzincontinuous semimaxingale X with local characteristics (b; c; 0; 0) starting at x0 be speci ed uniquely (e.g., according to Theorem 2.8.21, inf inf inf cs (u) > 0 and sup inf kcs (u)k < 1; t 2 R+ ; v 2 jj=1 st jujv st jujv R + ). If, as 2 , and v 2 R+ Zt

0

1=r P X0 ! x0 ,

where r = 2 = , and for all t 2 R+

ess sup jbs (u) bs (u)j ds ! 0; jujv

Zt

0

ess sup kc~0s (u) cs (u)k ds ! 0; jujv

ld then X ! X as 2 at rate r . Proof. The proof is almost the same as for Theorem 4.4.8. In some more detail, either one of conditions (P) or (SE) implies (L2 )loc . Since by hypotheses conditions (sup B 0)loc and (C00 )loc hold, according to Theorem 5.3.5 one needs to check conditions (A)loc + (a)loc . If condition (P) is satis ed, then condition (V S0 )loc holds, which implies (A)loc + (a)loc . If condition (SE) is satis ed, then conditions (A0 )loc + (a0 )loc can be veri ed as in the proof of Theorem 4.4.8.

Remark 5.4.5. We recall that by Theorem 2.8.9 under the hypothe-

ses X is a Luzin-continuous idempotent process satisfying the equation

X_ t = bt (Xt )+ t (Xt )W_ t ; X0 = x0 ; and the deviability distribution of X has density given by 1 Z1 X (x) = exp (x_ s bs (xs ))cs (xs ) (x_ s bs (xs )) ds 2 0

if x is absolutely continuous, x0 = x0 and x_ s bs(xs ) is in the range of cs (xs ) a.e., and X (x) = 0 otherwise.

© 2001 by Chapman & Hall/CRC

424

Maxingale problem

We conclude the section with some illustrative examples. To simplify notation, we consider one-dimensional settings.

Example 5.4.6.

Let R-valued processes X ";Æ; = (Xt";Æ; ; t 2 R+ ), indexed by " > 0; Æ > 0 and > 0, be de ned on respective stochastic bases ( ";Æ; ; F";Æ; ; F";Æ; ; P";Æ; ) and satisfy the equations

Xt";Æ;

= x0 + +Æ

Zt

p bs(Xs";Æ; ) ds + "

0 Zt Z

0 G

fs(Xs";Æ; ; y)

Zt

0

s (Xs";Æ; ) dWs"

N (ds; dy) 1 ds m(dy) ;

where (G; G ) is a measurable space, bs(u); s (u) and fs (u; y) are respective B(R+ ) B(R)=B(R ), B(R+ ) B(R)=B(R ) and B(R+ )

B(R) G =B(R)-measurable functions, W " = (Ws"; s 2 R+ ) are Rvalued Wiener processes, m(dy) is a non-negative - nite measure on (G; G ), and N = (N (ds; dy)) are Poisson random measures on R + G with intensity measures 1 ds m(dy ). We also assume that bs (u) and s(u) are continuous in u, lim

u!v

Z

jfs(u; y) fs (v; y)j2 m(dy) = 0;

G

the following linear-growth conditions are met:

bs (u)2 +s(u)2 ls (1+u2 ); jfs(u; y)j hs (y)(1+juj); where ls and hs (y) are R+ -valued andR increasing in s, hs (y) is G =B(R+ ){measurable for every s 2 R+ , G hs(y)2 m(dy) < 1, and the non-degeneracy condition holds: inf inf (s

st jujv

(u)2 +

Z

fs(u; y)2 m(dy)) > 0; t 2 R+ ; v 2 R+ :

G

For existence of X ";Æ; see, e.g., Gihman and Skorohod [54, Chapter 5]. Let us consider the following moment conditions on the jumps of the X ";Æ; :

© 2001 by Chapman & Hall/CRC

425

LD convergence of Markov processes

~ for some Æ > 0 (P) Zt

0

ess sup jujv

Z

jfs(u; y)j2+Æ m(dy) ds < 1; t 2 R+ ; v 2 R+ ;

G

f for some 2 (0; 1] and > 0 (SE) Zt

0

ess sup jujv

Z

exp jfs (u; y)j m(dy) ds < 1; t 2 R+ ; v 2 R+ :

G

Let an idempotent Luzin-continuous process X satisfy the equation Z

X_ t = bt (Xt )+ t (Xt )2 + ft (Xt ; y)2 m(dy) 1=2 W_ t ; X0 = x0 ;

G

where W is an R-valued idempotent Wiener process. The process X is well de ned by Theorems 2.6.24 and 2.8.21.

Theorem 5.4.7. Let ! 0, ! 20, and Æ ! 0 in such a way that 2 1 1

= Æ = . If, in addition, either Æ ln( ) ! 1 and condition ld g) holds, then X ";Æ; ! (P~ ) holds, or Æ2 1 ! 1 and condition (SE X at rate 1=. Proof. The predictable characteristics without truncation of X ";Æ; are of the form

B 0";Æ; = t

Zt

0

bs(Xs";Æ; ) ds;

g(x) t";Æ; =

1

Ct";Æ;

Zt Z

0 G

="

Zt

0

s(Xs";Æ; )2 ds;

g Æfs (Xs";Æ; ; y) m(dy) ds for g Borel and bounded:

It is straightforward to see that the convergence hypotheses of Theorem 5.4.4 hold for = (; Æ; ), R = 1 , = Æ 1 , b;Æ; s (u) = bs (u), ;Æ; 2 2 and c~s (u) = cs (u) = s (u) + G fs(u; y) m(dy): The moment conditions on the jumps are the same as in Theorem 5.4.4.

© 2001 by Chapman & Hall/CRC

426

Maxingale problem

Let us assume, in addition, that the function bs (u) is dierentiable in u (for almost all s) and the derivative b0s (u) is bounded on bounded domains. We denote as (xt ; t 2 R+ ) the solution of the equation

xt = x0 +

Zt

bs (xs ) ds

0

(for existence and uniqueness of (xt ) see, e.g., Coddington and Levinson [26, Chapter II]). We introduce the processes X~ ; = (X~t; ; t 2 R+ ) by

X~ t; =

r

;; Xt xt ;

where > 0 and de ne \a non-time-homogeneous idempotent Ornstein-Uhlenbeck process" X~ = (X~t ; t 2 R+ ) by Z

~_ t ; X~ 0 = 0; X~_ t = b0t (xt )X~ t + t (xt )2 + ft (xt ; y)2 m(dy) 1=2 W

G

~ is an R-valued idempotent Wiener process. By Theowhere W rem 2.6.26 the latter equation has a unique Luzin strong solution with idempotent distribution speci ed by the density

x) = exp

X~ (

1 2

Z1

0

x_ t R

b0t (xt )xt 2 dt t (xt )2 + G ft (xt ; y)2 m(dy)

if x0 = 0 and x is absolutely continuous, and X~ (x) = 0 otherwise.

Theorem 5.4.8. Let ! 0 and ! 0 in such a way that1 = ! 1. If, in addition, either condition (P~ ) holds and ln( ) ! 1, or ld ~ g) holds and 2 ! 1, then X ~ ; ! condition (SE X at rate 1=. Proof. We again invoke Theorem 5.4.4. Since X~ ; satis es the equa-

© 2001 by Chapman & Hall/CRC

427

LD convergence of Markov processes

tion

X~ t;

=

Z t r r

bs

0 Zt

X~ s; + xs

bs (xs ) ds

r p + s X~ s; + xs dWs

p +

0 Zt Z

fs

r

0 G

X~ s; + xs; y

N (ds; dy) 1ds m(dy) ;

it follows that the predictable characteristics of X~ ; without truncation are of the form

B 0 ; = t

Zt r r

0

bs

Ct; g(x) t;

=

1

Zt Z

=

X~ s; + xs

Zt

0

s

r

bs (xs ) ds;

2 X~ s; + xs ds;

r p g fs X~ s; + xs; y m(dy) ds;

0 G

for g Borel and bounded: Therefore, letting = (; ), = 1 and = () 1=2 , in the notation of Theorem 5.4.4 r r bs (u) = bs u + xs bs (xs ) ; 2 r c~0s (u) = s u + xs Z r 2 + fs u + xs ; y m(dy) G

so that we have the convergences Zt

0

ess sup jujv

jbs (u)

bs(u)j ds ! 0;

© 2001 by Chapman & Hall/CRC

Zt

0

ess sup jc~0s (u) cs j ds ! 0; jujv

428

Maxingale problem

where

Z

bs (u) = b0s (xs )u; cs = s (xs )2 + fs(xs ; y)2 m(dy): G

Now the claimed LD convergence follows by Theorem 5.4.4.

Example 5.4.9.

Let R-valued processes X n = (Xtn ; t 2 R+ ), where n 2 N , be de ned on respective stochastic bases ( n ; Fn ; Fn = (Ftn ; t 2 R+ ); Pn ) and have the form 1 Xtn = N n n n

Zt

0

f Xsn ; Ynsn ds ;

where f (x; y) is an R+ -valued Borel function, N n = (Ntn ; t 2 R+ ) are Poisson processes on ( n ; Fn ; Fn ; Pn ), and Y n = (Ytn ; t 2 R+ ) are n ; t 2 R ); P ): Ornstein-Uhlenbeck processes on ( n ; Fn ; Gn = (Ft=n + n

Ytn

=

Zt

0

1 Ysn ds + p Wtn; n

W n = (Wtn ; t 2 R+ ) being Wiener processes on ( n ; Fn ; Gn ; Pn ). The processes X n are well de ned since the N n are piecewise constant. We assume that f (x; y) is continuous at points (x; 0) for x 2 R+ and is such that sup0xa;y2R f (x; y) < 1 for a > 0, f (x; 0) > 0 for x 2 R+ , and the function f (x; 0) grows at most linearly as x ! 1. We prove that the X n LD converge at rate n to the Luzincontinuous idempotent process X satisfying the equation Xt = N

Z t

f (Xs ; 0) ds ;

0

where N is a Poisson idempotent process, and having the idempotent distribution with density Z1 X (x) = exp sup x_ t (e 1)f (xt ; 0) dt 2R 0

© 2001 by Chapman & Hall/CRC

429

LD convergence of Markov processes

if x is absolutely continuous, increasing and x0 = 0, and X (x) = 0 otherwise. The idempotent process X is well de ned by Theorem 2.6.33. Let us denote tn = Yntn . For tn we have the equation

tn

= n

Zt

0

^ tn; snds + W

(5.4.10)

^ n = (W ^ tn; t 2 R+ ) is a Wiener process on ( n ; Fn ; Fn ; Pn ). where W One that the nX n have Fn -compensators Ant = R t cann show n n 0 f (Xs ; s ) ds so that by Theorems 2.8.10, 2.8.28, and 5.3.7 the claim would follow by t^Z (X n ) N sup f (Xsn ; sn)ds tT 0

1=n f (Xsn ; 0)ds Pn

t^ZN (X n ) 0

! 0 as n ! 1; T > 0; N

2 N:

By continuity of f at points (x; 0), where x 2 R+ ; and the boundedness condition supx2[0;a];y2R f (x; y) < 1, for arbitrary " > 0 there exists Æ > 0 such that 8 >

: tT 0

f (Xsn ; 0)ds

t^ZN (X n ) 0 8 "

9 > = > ; 9 =

: 1(jsnj > Æ)ds > 2" ; ; 0

so by \the Chebyshev inequality" it is suÆcient to show that lim P 1=n n!1 n

ZT

0

jsnj2 ^1 ds > = 0; > 0:

(5.4.11)

Let g(x); x 2 R; be a twice dierentiable non-negative function with bounded rst and second derivatives, and such that g(x) = x2 =2; jxj 1; and xg0 (x) 1 if jxj 1 (e.g., g(x) = 1=2 + ln jxj +

© 2001 by Chapman & Hall/CRC

430

Maxingale problem

(ln jxj)2 ; jxj 1). By Ito's formula and (5.4.10)

g(tn ) = g(0)

n

Zt

g0 (n)n ds + s

0

Zt

s

0

1 g0 (sn) dW^ sn + 2

Zt

0

g00 (sn )ds:

Since g0 is bounded, for all 2 R

En exp

ZT

0

2 ^ sn g0 (sn )dW 2

ZT

0

g0 (sn)2 ds = 1

and hence

En exp

g(Tn )

g(0) + n

ZT

g0 (n)n ds

0

s

s

ZT 2

2

0

2

ZT

0

g00 (sn )ds

g0 (sn )2 ds = 1;

which implies, since g is non-negative, g0 and g00 are bounded, g0 (x) = x; jxj 1; and g0 (x)x 1 for jxj 1, that for some function F ()

En exp n

ZT

0

jsnj2 ^ 1 ds F ():

Hence, lim sup Pn1=n n!1

ZT

0

jsnj2 ^ 1 ds >

ZT

exp( ) lim sup En exp(n jsnj2 ^ 1 ds) 1=n n!1

0

Since is arbitrary, (5.4.11) is proved.

© 2001 by Chapman & Hall/CRC

exp( ):

431

LD convergence of Markov processes

Example 5.4.10. This example considers a discrete-time case and builds on Example 4.4.12. Let n ! 1, ! 1 and ! 1 in such a way that n= ! 1, and i ; i = 1; 2; : : :, where = (n; ; ), be i.i.d. indicator random variables, which equal 1 with probability =(n ) and 0 with probability 1 =(n ). We de ne random variables Yk by

Yk = Yk 1 +

b f (YX k 1 =)c

i=1

i ; Y0 = 0;

where f (x) is a continuous positive function, growing no faster than linearly as x ! 1. Let the process X = (Xt ; t 2 R+ ) be de ned by Xt = Ybntc =. Then X is a point process whose compensator A = (At ; t 2 R+ ) relative to the ltration generated by X is given by

At

bntX c 1 )c: = E1 b f (Xi=n i=0

We have At

Zt

0

f (Xs ) ds

bnt Z c=n 0

b f (Xs )c f (X ) ds s +

Zt

bntc=n

f (Xs ) ds

and

1 X 2 As 2 2 bntc sup ( f (x))2 : st^ (X ) n 0xN N

Hence, by Theorems 5.3.7 and 2.8.28 the net fX ; 2 g LD converges at rate to the semimaxingale X with local characteristics (b; 0; ; 0), where bs (x) = f (xs) and ( ; x) = 1(1 2 )f (xs); equiva R lently X is the Luzin solution of the equation Xt = N 0t f (Xs ) ds ;

© 2001 by Chapman & Hall/CRC

432

Maxingale problem

where N is an idempotent Poisson process, whose idempotent distribution has density Z1 X sup x_ t (e 1)f (xt ) dt (x) = exp 2R 0

if x is absolutely continuous, increasing and x0 = 0, and X (x) = 0 otherwise.

Remark 5.4.11.

It is straightforward to extend Examples 5.4.6, 5.4.9 and 5.4.10 to the case where the coeÆcients depend on the past.

© 2001 by Chapman & Hall/CRC

Chapter 6

Large deviation convergence of queueing processes In this chapter we apply the results on large deviation convergence of semimartingales for deriving large deviation asymptotics in queueing systems.

6.1 Moderate deviations in queueing networks In this section we prove LD convergence of queueing processes in single server queues and networks of single server queues to idempotent diusions.

6.1.1 Idempotent diusion approximation for single server queues We consider a sequence of FIFO single server queues indexed by n . For the nth system, we denote by Ant the number of arrivals by time t , by Stn the number of customers served for the rst t units of the server's busy time, by Dtn the number of departures by time t , by Qnt the queue length at time t , by Wtn the un nished work at time t , by Ctn the completed work at time t , by Hkn the waiting time of the k th customer, and by Lnk the departure time of the k th 433 © 2001 by Chapman & Hall/CRC

434

LD convergence for queues

customer. We also introduce

Vkn = minft 2 R+ : Stn kg; k 2 Z+;

(6.1.1)

which, for k 2 N , is the cumulative service time of the rst k customers. All the objects referring to the nth system are assumed to be de ned on a complete probability space ( n ; Fn ; Pn ). Also all the processes are assumed to have trajectories from the associated Skorohod space. The above processes are connected by the following equalities

Ctn =

Zt

0

Wtn = W0n + V n Æ Ant Ctn; Zt

1(Wsn > 0) ds = 1(Qns > 0) ds;

(6.1.2)

Qnt = Qn0 + Ant Dtn; Dtn = S n Æ Ctn;

(6.1.3) (6.1.4)

0

p

Let bn ! 1 and bn = n ! 0 as n ! 1 , and n and n be positive numbers. We de ne the associated normalized and timescaled processes by 1 An = (Ant ; t 2 R+ ); Ant = p (Annt n nt); (6.1.5) bn n 1 n nt); S n = (S nt ; t 2 R+ ); S nt = p (Snt (6.1.6) n bn n 1 (6.1.7) Qn = (Qnt ; t 2 R+ ); Qnt = p Qnnt ; bn n 1 n nt) ; C n = (C nt ; t 2 R+ ); C nt = p (Cnt (6.1.8) bn n 1 V n = (V nt ; t 2 R+ ); V nt = p (Vbnntc n 1 nt); bn n 1 n n n n D = (Dt ; t 2 R+ ); Dt = p (Dnt n nt); bn n 1 W n = (W nt ; t 2 R+ ); W nt = p Wntn ; bn n 1 H n = (H nt ; t 2 R+ ); H nt = p Hbnntc+1 ; bn n 1 Ln = (Lnt ; t 2 R+ ); Lnt = p (Lnbntc+1 n 1 nt): bn n

© 2001 by Chapman & Hall/CRC

435

Moderate deviations for networks

We assume that n ! > 0 and n ! > 0 as n ! 1 , and \the near-heavy-traÆc condition" holds: 1p n(n n) ! c; c 2 R: bn

(6.1.9)

Note that (6.1.9) implies that = . We recall that the one-dimensional Skorohod re ection map x ! R(x) is characterised by the property that z = R(x) is an only R+ valued R 1function such that z = x + y, where y is increasing, y0 = 0 and 0 1(zt > 0) dyt = 0. It is a continuous map from D (R + ; R) to D (R + ; R ) and can explicitly be written as

R(x)t = xt 0infst xs ^ 0; t 2 R+ ;

(6.1.10)

where x 2 D (R + ; R) and x0 2 R+ , see, e.g., Ikeda and Watanabe [66]. Let WA = (WA;t ; t 2 R+ ) and WS = (WS;t ; t 2 R+ ) be independent idempotent Wiener processes on an idempotent probability space ( ; ). Let A and S be real numbers. As above we denote e = (t; t 2 R+ ). In the theorems below LD convergence refers to the 2 Skorohod topology and rate rn = bn .

Theorem 6.1.1. Let (A ; S ) ! (A WA; S WS ) and n

n

ld

ld Then Qn ! Q, where Q = (Qt ; t 2 R+ ) is an continuous idempotent process de ned by

2 n Pn1=bn Q0 ! q0 .

R + -valued

Luzin-

Q = R q0 + A WA S WS + ce :

Proof. Let us denote A = A WA and S = S WS . By (6.1.3), (6.1.4), (6.1.2), (6.1.7), (6.1.5), (6.1.6), and (6.1.8)

Qnt = Qn0 + Ant

S n Æ C 0tn + +

pn

C nt =

© 2001 by Chapman & Hall/CRC

bn

n

Zt

0

pn Zt bn

0

pn bn

(n

n )t

1(Qns = 0) ds;

(6.1.11)

1(Qns = 0) ds;

(6.1.12)

436

LD convergence for queues

where 1 n C 0tn = Cnt =

n

Zt

0

1(Qns > 0) ds:

(6.1.13)

Since Qnt is non-negative and 0t 1(Qns = 0) ds increases only when Qnt = 0 , (6.1.11) allows us to conclude that n

Q =R

R

n

Since by (6.1.12) and (6.1.11)

n C

n

p

n S Æ C 0n + (n

Qn0 + An

bn

n C

n )e :

(6.1.14)

p

n S Æ C 0n + (n n )e Qn;

= Qn0 + An

n

bn

it follows by (6.1.14) that n

p

n S Æ C 0n + (n b

= Qn0 + An

n

R

Qn0 + An

n n

p

n S Æ C 0n + (n

Therefore, by (6.1.10) n jC nt j 2 sup Qn0 + Ans st 1=b2n The convergences Qn0 Pn! q0

n )e bn

p

n S Æ C 0sn + (n n

bn

n)e :

n )s ;

t 2 R+ :

ld and (An ; S n ) ! (A; S ) , the fact that A and S are proper idempotent processes, the inequality C 0tn t , and (6.1.9) imply that 2

1=bn ( jC n j > a) = 0: lim lim sup P n t n a!1 n!1

Hence, by (6.1.12) and the facts that Zt

0

1

(Qns

= 0) ds

pn=b ! 1 and ! > 0 n n

2 Pn1=bn

! 0 as n ! 1; t 2 R+ ; 2

(6.1.15)

1=bn which implies by (6.1.13) that C 0n Pn! e. Then \the time-change theorem" (Lemma 3.2.11) implies by the LD convergence of (An ; S n )

© 2001 by Chapman & Hall/CRC

437

Moderate deviations for networks

to (A; S ) that the sequence f(An ; S n Æ C 0n ); n 2 N g LD converges to 2 n Pn1=bn ld (A; S ) as well. Since Q0 ! q0 , we have that (Qn0 ; An ; S n Æ C 0n ) ! (q0 ; A; S ) by Lemma 3.1.42. By (6.1.14) and continuity of re ection n n n n 0 n p Q is a continuous function of (Q0 ; A ; S Æ C ; ( n=bn )(n n)e). Therefore, the LD convergence of (Qn0 ; An ; S n Æ C 0n ), the near-heavy traÆc condition (6.1.9), and the contraction principle yield the required LD convergence of fQn ; n 2 N g. The idempotent process Q is Luzin-continuous since A and S are Luzin-continuous and R is continuous.

Remark 6.1.2. If we assume, in addition to the hypotheses of The2 1=bn ld orem 6.1.1, that W n0 Pn! q0 =, then (Qn ; Dn ; W n ; C n ; H n ; Ln ) ! (Q; D; W; C; H; L), where

Dt = AWA;t Qt + ct; Wt = Ct =

A WA;t S WS;t + ct Ht = W (t=); Lt =

Qt ; Wt ;

Dt= :

The following lemma gives an explicit expression for the idempotent distribution of Q. Let Q denote the idempotent distribution of Q and I Q (q) = ln Q (q) be the associated rate function.

Lemma 6.1.3. Let A2 + S2 > 0. The rate function I Q is given by q) = 2(2 1+ 2 ) A S

I Q(

Z1

0

1(qt > 0)(q_ t +

c)2 dt

1(c > 0)c2 Z1 1(qt = 0) dt;

2(A2 + S2 )

0

q is a non-negative and absolutely continuous function such that q0 = q0, and IQ(q) = 1 otherwise. if

For a proof, we need the following result.

© 2001 by Chapman & Hall/CRC

438

LD convergence for queues

Lemma 6.1.4. Let z 2 C (R + ; R) be non-negative and x 2 C (R + ; R) be absolutely continuous. Then z = R(x) if and only if z is absolutely continuous and there exists an absolutely continuous function y 2 C (R + ; R )

such that

z_ t = x_ t + y_ t

a.e.

and

y0 = 0; y_ t 2 R+ a.e.; zt y_ t = 0 a.e. Also z_ t = 0 a.e. on the set ft : zt = 0g. Proof. SuÆciency of the condition follows by the de nition of the re ection mapping. Conversely, if y = R(x) x, then yt ys Rt _ j x j du for 0 s t, so y is absolutely continuous. The other s u conditions on y follow from the de nition of re ection. For the nal part, note that a.e. z_ t = limh!0(zt+h zt )=h. The numerator in the latter fraction being non-negative since zt = 0 implies that the fraction is non-negative for h positive and non-positive for h negative. Hence, the limit is zero. Proof of Lemma 6.1.3. Let 2 = A2 + S2 . By Corollary 2.4.11 we may assume that A WA + S WS = Ws , where W is an idempotent Wiener process. By Theorem 6.1.1 and the de nition of the image idempotent measure

Q (q) = supfW (w);

q = R(q0 + w + ce)g:

Therefore, q0 = q0 and q is absolutely continuous Q -a.e. For these q by the de nition of an idempotent Wiener process, Lemma 6.1.4 and Lemma A.2 in Appendix A

q) = y: y0=0inf;y_ t2R+; 21

I Q(

1(q >0)y_ =0; q_ t=t w_ t +ct+y_ t

Z1

0

w_ t2 dt

1 = 2 2

© 2001 by Chapman & Hall/CRC

Z1

0

q_ t _ t 2R+ ; y_ t : yinf

1(qt >0)y_ t =0

c

y_ t 2 dt

439

Moderate deviations for networks

1 = 2 2 +

1 22

1 = 2 2

Z1

0 Z1 0 Z1 0

1(qt > 0) q_ t

c 2 dt

1(qt = 0) y2infR+ q_ t 1(qt > 0) q_ t

c

2

c y 2 dt Z1 1 (c > 0)c2 dt + 1(qt = 0) dt:

22

0

The next lemma formulates the LD convergence conditions in the hypotheses of Theorem 6.1.1 in terms of interarrival and service times. Let Ukn = inf ft 2 R+ : Ant kg ; k 2 Z+; and U nt = Ubnntc p n 1 nt = bn n : Recalling also (6.1.1) we have by Lemma 3.2.13 the following.

Lemma 6.1.5. The LD convergencen (Ann; S n) !ld

(A WA ; S WS ) holds if and only if the sequence f(U ; V ); n 2 N g LD converges to (A 3=2 WA ; S 3=2 WS ). We now specify the results to the case of GI=GI=1 queues, i.e., we assume that the An and S n are renewal processes. Let us denote by uni ; i 2 N ; the time between the i th and (i + 1) th arrivals and by vin ; i 2 N ; the service time of the i th customer in the n th system. By hypothesis the sequences funi ; i 2 N g and fvin ; i 2 N g are independent i.i.d. Theorem 4.4.8 provides us with the following way of checking the convergence requirements of Lemma 6.1.5.

Lemma 6.1.6. Let either one of the following conditions hold: (i) supn En (un1 )2+ < 1; supn En (v1n )2+ < 1 for some > 0, and b2n = ln n ! 0; (ii) supn En exp((un1 ) ) < 1; supn En exp((v1n ) ) < 1 for some > 0; 0 < 1, and b2n =n =2 ! 0. If En un1 ! 1 ; En v1n ! 1 ; Varn un1 ! A2 =3 ; and Varn v1n ld S2 =3 ; then (U n ; V n ) ! (A 3=2 WA; S 3=2 WS ).

© 2001 by Chapman & Hall/CRC

!

440

LD convergence for queues

We now establish LD convergence for stationary waiting times. Let partial-sum processes U 0 n = (U 0 nk; k 2 Z+) and V n = (Vkn ; k 2 Z+) be given by k

k

i=1

i=1

X X U 0 nk = uni ; U 0 n0 = 0; Vkn = vin ; V0n = 0;

(6.1.16)

so that, as above, Vkn , for k 2 N , is the cumulative service time of the rst k customers. The equation for waiting times is

Hkn+1 = H1n +Vkn U 0 nk min (H1n +Vin U 0 ni )^0:

(6.1.17)

1ik

We recall that if En v1n < Eun1 , then the waiting times Hkn converge in distribution as k ! 1 to the proper random variable supk2Z+(Vkn U 0 nk ) (see, e.g.,pBorovkov [15]). We denote the latter by H0n and let H n0 = H0n =(bn n) .

Theorem 6.1.7.

Let either p one of conditions (i) or (ii) of Lemma 6.1.6 hold. Let ( n=bn )(Eun1 Ev1n ) ! c0 > 0, Varn un1 ! U2 ; Varn v1n ! V2 , where U2 + V2 > 0, as n ! 1. Then the sequence fH n0 ; n 2 N g LD converges in distribution to an exponentially distributed R+ -valued idempotent variable with density 0 2 2 (x) = exp 2c x=(U + V ) ; x 2 R+ . Proof. Since H0n is distributed as supk2Z+(Vkn a Borel subset A of R+ , n Pn (H 0

U 0 nk ) , we have, for ! A

1

2 A) Pn b pn sup (Vkn U 0nk) 2 n 0kbntc Pn( sup (Vkn U 0nk) 0): k>bntc

Let

U~ 0n = (U~t0n ; t 2 R+ ); U~t0n =

1

p (U 0 n Eun1 nt); bn n bntc 1 V~ n = (V~tn ; t 2 R+ ); V~tn = p (Vbnntc Ev1n nt): bn n

© 2001 by Chapman & Hall/CRC

441

Moderate deviations for networks

ld Since p by Theorem 4.4.8 (U~ 0n ; V~ n ) ! (U WA ; V WS ) , (En un1 n 0 En v1 ) n=bn ! c by hypotheses, and

1

p sup (V n U 0nk) bn n 0kbntc k = sup 0st

V ns

U~s0n

(En un1

En v1n )

pn bn

s ;

by the contraction principle and the fact that U WA + V WS = W , where 2 = A2 + S2 and W is an idempotent Wiener process, we have that 1

ld p sup (V n U 0nk) ! sup (Ws c0 s): bn n 0kbntc k 0st

Let t denote the idempotent variable on the right-hand side and = sups2R+ (Ws c0 s). It is an easy exercise to check that has idempotent distribution in the statement of the theorem; in particular, it is a Luzin idempotent variable. We show that t converges to as t ! 1 in idempotent distribution. The map w ! sup0st (ws c0 s) from C (R + ; R) to R is continuous, so t is a Luzin idempotent variable as well. Assuming that both and t are de ned on (C (R + ; R); W ), we have that t monotonically converges to zero W -almost everywhere, so by Theorem 1.3.10 the convergence is actually in deviability W and by Lemma 1.10.7 id t ! . Thus, by Lemma 3.1.37 the required would follow by

2

lim lim sup P 1=bn sup (Vkn U 0 nk ) 0 = 0: t!1 n!1 n k>bntc Denoting Æn = En (un1 Æn > 0 ,

(6.1.18)

v1n ) and in = vin uni + Æn , we have, since

Pn sup (Vkn U 0 nk ) 0 k>bntc

1 X l=blog2 (nt)c

© 2001 by Chapman & Hall/CRC

Pn

max

k=2l +1;:::;2l+1

k X i=1

!

in

kÆn

0

!

442

LD convergence for queues

+

1 X l=blog2 (nt)c

1 X

l=blog2 (nt)c

0 2l X @ Pn n

i=1

Pn

i

max

1

2l 1Æn A k X

k=1;:::;2l i=1

2

in

1 X

l=blog2 (nt)c

2l 1 Æn

Pn

max

!

k X

k=1;:::;2l i=1

in

!

2l 1Æ

n

:

Limit (6.1.18) now follows bypLemma A.3 in Appendix A and the near-heavy traÆc condition ( n=bn )Æn ! c0 > 0 as n ! 1 .

6.1.2 Idempotent diusion approximation for queueing networks We now extend some of the results of the preceding subsection to the queueing-network set-up. We consider a sequence of networks indexed by n. The nth network has a homogeneous customer population and consists of K FIFO single server stations. The network is open so customers arrive from outside and eventually leave. For the nth network, let An;k t ; k = 1; : : : ; K; denote the cumulative number of customers who arrive at station k from outside the network during the interval [0; t], and let Stn;k ; k = 1; : : : ; K; denote the cumulative number of customers who complete service at station k during the rst t units of busy time of that station. We call An = (An;k ; k = 1; : : : ; K ), where An;k = (An;k t ; t 2 R + ), and n;k n n;k n;k S = (S ; k = 1; : : : ; K ), where S = (St ; t 2 R+ ), the arrival process and service process, respectively (note that some of the entries in An may equal zero). We associate with the stations of the network the processes Rn;k = (Rn;kl ; l = 1; : : : ; K ); k = 1; : : : ; K , where n;kl ; m 2 N ), and Rn;kl denotes the cumulative number of Rn;kl = (Rm m customers among the rst m customers who depart station k that go directly to station l. The process Rn = (Rn;kl ; k; l = 1; : : : ; K ) is referred to as the routing process. We consider the processes An;k , S n;k and Rn;k as random elements of the respective Skorohod spaces D (R + ; R ), D (R + ; R ) and D (R + ; R K ); accordingly, An , S n and Rn are considered as random elements of D (R + ; RK ), D (R + ; RK ) and D (R + ; R K K ), respectively. We assume that the data associated with the nth network is de ned on a probability space ( n ; Fn ; Pn ).

© 2001 by Chapman & Hall/CRC

443

Moderate deviations for networks

We next introduce normalized and time-scaled versions of the arrival process, service process and routing process. Let n;k 2 R+ ; n;k 2 R+ ; and pkl 2 [0; 1], k = 1; : : : ; K; l = 1; : : : ; K . We de ne

An;k t =

n;k An;k nt pn;k nt ; S n;k = Snt pn;k nt ; t bn n bn n n;kl Rbntc pkl bntc p Rn;kl ; t = bn n

(6.1.19) (6.1.20)

p

where, as above, bn ! 1 and bn = n ! 0, and let An = (An;k ; k = 1; : : : ; K ), S n = (S n;k ; k = 1; : : : ; K ), Rn;k = (Rn;kl ; l = 1; : : : ; K ); k = 1; : : : ; K , and Rn = (Rn;kl ; k; l = 1; : : : ; K ). Again the latter processes are considered as random elements of D (R + ; RK ), D (R + ; R K ), D (R + ; R K ), and D (R + ; R K K ), respectively. Also we denote n = (n;k ; k = 1; : : : ; K ), n = (n;k ; k = 1; : : : ; K ) and P = (pkl ; k = 1; : : : ; K; l = 1; : : : ; K ). Elements of RK are regarded as column-vectors. Our main concern here is the queue-length process Qn = n;k n;k (Q ; k = 1; : : : ; K ), where Qn;k = (Qn;k t ; t 2 R + ), Qt denoting the number of customers at station k at time t. The associated normalized and time-scaled process Qn = (Qn;k ; k = 1; : : : ; K ) is de ned by

Qn;k t =

Qn;k pnt : bn n

(6.1.21)

In analogy with the hypotheses of Subsection 6.1.1 we assume that n ! = (^ 1 ; : : : ; ^ K ) and n ! = (^1 ; : : : ; ^K ) as n ! 1, where is a component-wise positive vector, and that \the nearheavy traÆc condition" holds: for some c 2 RK

pn bn

(n (EK P T )n ) ! c as n ! 1;

(6.1.22)

in particular,

= (EK P T ):

(6.1.23)

(Recall that EK denotes the identity K K matrix.) We also assume that the spectral radius of the matrix P is less than unity.

© 2001 by Chapman & Hall/CRC

444

LD convergence for queues

We recall that the skew re ection mapping RP , Harrison and Reiman [59], Reiman [115], is de ned as the map from D (R + ; RK ) into D (R + ; RK ) associating to each x = (xt ; t 2 R+ ) 2 D (R + ; RK ) such that xk0 2 R+ ; k = 1; : : : ; K; a function z = (zt ; t 2 R+ ) 2 D (R + ; R K ) such that

z = x + (EK P T )y, 2. y is componentwise increasing and y0k = 0; k = 1; : : : ; K , 1.

3.

Z1

z 2 R+ and zkt dytk = 0, k = 1; : : : ; K . k t

0

The map RP is well de ned and Lipshitz continuous for the locally uniform and Skorohod topologies on D (R + ; RK ) , Harrison and Reiman [59], Reiman [115], Chen and Whitt [23]. As in Subsection 6.1.1 all LD convergences below refer to the rate rn = b2n and the Skorohod topology. We recall the notation introduced at the end of Section 3.2. If x 2 D (R + ; RK ) has componentwise increasing R+ -valued paths, then, for y 2 D (R + ; R K ), we denote y Æ x = ((yxk kt ; k = 1; : : : ; K ); t 2 R+ ); analogously, if rt = (rklt ; k; l = 1; : : : ; K ) 2 RK K , then r Æ xt = (rkl xkt ; k; l = 1; : : : ; K ) . For vectors = (1 ; : : : ; K ) 2 RK and = ( 1 ; : : : ; K ) 2 RK , we denote = (1 1 ; : : : ; K K ) 2 RK . Let 1 denote the K -vector with all the entries equal to 1.

Theorem 26.1.8. Let the near-heavy-traÆc condition (6.1.22) hold. 1=bn

Let Qn0 Pn! q0 . Let the sequence f(An ; S n ; Rn ); n 2 N g LD converge in D (R + ; RK RK RK K ) to an idempotent process (A; S; R), where A = (A1 ; : : : ; AK ), S = (S 1 ; : : : ; S K ), and R = (R1 T ; : : : ; RK T ) are de ned by

A = AWA ; S = S WS ; Rk = kR WRk ; k = 1; : : : ; K; WA ; WS ; WRk ; k = 1; : : : ; K , being mutually independent K { dimensional idempotent Wiener processes and A; S ; kR ; k = ld 1; : : : ; K; being K K matrices. Then Qn ! Q, where Q is a Luzin-continuous idempotent process given by Q = RP (q0 + A +(R Æ e)T 1 (EK P T )S + ce):

© 2001 by Chapman & Hall/CRC

445

Moderate deviations for networks

Proof. The proof is a straightforward extension of the proof of Theorem 6.1.1. In analogy with (6.1.3), (6.1.4) and (6.1.2), we have that for k = 1; : : : ; K

Qn;k t

K X n;k n;k = Q0 + At + Rn;lk Æ Dtn;l l=1

Dtn;k ;

where Dtn;k = SRn;k t 1(Qn;k >0)ds : Introducing s

0

n;k C0 = t

Zt

0

1

(Qn;k s

n;k n;k Dnt 0 ; > 0)ds; D t =

n

we then have by (6.1.19), (6.1.20) and (6.1.21) that n;k n;k Qn;k t = Q0 + At +

+ +

K X l=1

pn bn

K X l=1

Rn;lk Æ D0 t

n;l

p

K X n;l n;k n 0 n;k 0 plk S Æ C t S Æ C t + (n;k + plk n;l n;k )t n;l

0 @n;k

bn

Zt

0

1

(Qn;k s

K X

= 0)ds

l=1

l=1

plk n;l

Zt

0

1

1

(Qn;l s

= 0)dsA ;

or in vector form

Qn = Qn0 + An + (Rn Æ D0 )T 1 (EK P T )S n Æ C 0 pn + (n (EK P T )n )e (EK P T )n C n ; (6.1.24) bn n

n

p

n;k where C nt = (C n;k n=bn 0t 1(Qn;k t ; k = 1; : : : ; K ), C t = s = n n;k n n;k 0 0 0 0 0)ds; C t = (C t ; k = 1; : : : ; K ), and D t = (D t ; k = 1; : : : ; K ). Hence, by the de nition of the re ection map RP R

Qn = RP Qn0 + An + (Rn Æ D0 )T 1 (EK P T ) S n Æ C 0 pn + (n (EK P T )n )e ; (6.1.25) bn n

© 2001 by Chapman & Hall/CRC

n

446

LD convergence for queues

so that from (6.1.24) (EK P T ) n C n n n = Qn0 + An + (Rn Æ D0 )T 1 (EK P T ) S n Æ C 0 pn + ( (EK P T )n )e bn n RP Qn0 + An + (Rn Æ D0n)T 1 (EK P T ) S n Æ C 0 n pn + (n (EK P T )n )e : bn

Since RP is a bounded map (in the sense that if z = RP (x), then sup0st zs K (t) sup0st jxs j, where K (t) depends only on t), the

p

1=b2n

ld convergences Qn0 Pn! q0 , (An ; S n ; Rn ) ! (A; S; R), and n=bn ! 1, the near-heavy-traÆc condition (6.1.22), and the facts that EK P T is nonsingular and is component-wise positive yield by the argument of the proof of (6.1.15) Zt

0

1

(Qn;k s

= 0)ds

2 Pn1=bn

! 0 as n ! 1; k = 1; : : : ; K; t 2 R+ ; 2

n;k 1=bn n;k implying that C 0 Pn! e as n ! 1. Then, since D0 t = 2 2 n;k n Pn1=bn 0 Pn1=bn 0 n;k n;k S Æ C =n and S =n ! e, we have that D ! e, so by n n ld (q0 ; A; S; R Æ 1 e) Lemma 3.2.11 (Qn0 ; An ; S n Æ C 0 ; Rn Æ D0 ) ! K K K K in D (R + ; R R R ). The claim now follows by (6.1.22), (6.1.25), continuity of the re ection and the contraction principle. The idempotent process Q is Luzin-continuous since the idempotent processes A, S and R are Luzin-continuous and RP is continuous.

Remark 6.1.9. One cann;kalso prove LD convergences for waiting and

sojourn times. Let Wt ; k = 1; : : : ; K; denote the virtual waiting p n;k time at station k at time t . We de ne W n;k t = Wnt =(bn n) and let W n = ((W n;k t ; k = 1; : : : ; K ); t 2 R + ) . For a vector k = k (k1 ; : : : ; kl ) , where ki 2 f1; 2; : : : ; K g , let An; t denote the number of customers with the routing (k1 ; k2 ; : : : ; kl ) who exogenously arrive by t , Ymn;k denote the sojourn time of the m th exogenous customer with

© 2001 by Chapman & Hall/CRC

447

Moderate deviations for networks

p

k n;k n;k the routing (k1 ; k2 ; : : : ; kl ) , and Y n; = t = Ybntc+1 =(bn n) , Y k 0n;k = (An;k =n; t 2 R ) . If, in addition to the (Y n; + t ; t 2 R+ ) , A nt hypotheses of the theorem,

2 n Pn1=bn W !

w0 , where q0 = w0 , then 2

1=bn ld (Qn ; W n ) ! (Q; W ) , where Q = W . If, in addition, A0n;k Pn! ld (W; Y ); where k e as n !P 1 ; where k > 0 , then (W n; Y n;k) ! Y Æ (k e) = li=1 Wki .

We now give an explicit expression for the idempotent distribution of Q. We de ne some more notation. For a subset J of f1; 2; : : : ; K g, we set FJ = f = (1 ; : : : ; K ) 2 RK+ : k = 0; k 2 k J; k > 0; k 62 J g and F J = f = (1 ; : : : ; K ) 2 RK + : = 0; k 2 J g ; 1J denotes the K -vector with entries from J equal to 1 and the rest of the entries equal to 0 ; J c denotes the complement of 0 K J . Let also RK; + denote the interior of R + , and K the set of all the subsets of f1; 2; : : : ; K g except the empty set. We introduce the positive semi-de nite symmetric matrix = ATA + (EK

P T )S TS (EK

Lemma 6.1.10.

P) +

K X k=1

^k R;k TR;k :

Let be positive de nite. Then the idempotent distribution of Q has rate function Z1 1 1(qt 2 RK;+ 0 )(q_ t r) 1(q_ t r) dt IQ (q) = 2 0 1 Z X1 + 1(qt 2 FJ ) y2infF c (q_ t 1J c r (EK P T )y) 1 2 J J 2K 0 (q_ t 1J c r (EK P T )y) dt when q 2 C (R + ; Rk ) is absolutely continuous and IQ (q) = 1 otherwise.

q0

= q0 , and

For a proof we need the following lemma which extends Lemma 6.1.4 and has a similar proof.

Lemma 6.1.11. Let z 2 C (R + ; RK ) have R+ -valued entry functions and x 2 C (R + ; RK ) be absolutely continuous. Then z = RP (x) if © 2001 by Chapman & Hall/CRC

448

LD convergence for queues

and only if z is absolutely continuous and there exists an absolutely continuous function y 2 C (R + ; RK ) such that

z_ t = x_ t +(EK

P T )y_ t a.e.

and

y0k = 0; y_ tk 2 R+ a.e.; zkt y_ tk = 0 a.e.; k = 1; 2; : : : ; K: Also z_ kt = 0 a.e. on the set ft 2 R+ : zkt = 0g, k = 1; 2; : : : ; K . Proof of Lemma 6.1.10. By the de nition of Q and Lemma 6.1.11 I Q (q) = 1 unless q0 = q0 and q is absolutely continuous. Since the idempotent Wiener processes WA ; WS and WRk ; k = 1; : : : ; K; are mutually independent, we have by Corollary 2.4.11 that

A + (R Æ e

)T

1 (EK

P T )S A WA +

K X

kR WRk Æ (^k 1e)

k=1 P T )

(EK

S WS

=

1=2 W;

where W is a K -dimensional idempotent Wiener process. Therefore,

q) =

I Q( =

1 2

Z1

inf

w2C (R+ ;RK ): q=RP (q0 + 1=2 w+ce)

q_ t : q yk =0

inf K k

y2R+ t 0

1 2

Z1

jw_ t j2 dt

0

c (EK

P T )y

q_ t

1

c (EK

P T )y dt;

which obviously coincides with the expression for I Q in the statement of the lemma. Let us now consider the i.i.d. case. Let, for some K 0 , ^ k > 0 0 when k = 1; : : : ; K 0 , and An;k t = 0 when k = K + 1; : : : ; K . n;k n;k Let the processes A ; k = 1; : : : ; K 0 , S ; k = 1; : : : ; K; and Rn;k ; k = 1; : : : ; K; be mutually independent for each n. Let the processes An;k ; k = 1; : : : ; K 0 ; and S n;k ; k = 1; : : : ; K , be renewal processes with times between renewals having nite second moments.

© 2001 by Chapman & Hall/CRC

449

Moderate deviations for networks

Let u^n;k ; k = 1; : : : ; K 0 ; denote the generic exogenous interarrival time and v^n;k ; k = 1; : : : ; K , the generic service time, for station k . Let, in addition, the routing mechanism not depend on n and be i.i.d. at each station with pkl being the probability of going directly from station k to station l. Lemma 6.1.12. Let under the above hypotheses, as n ! 1, En u^n;k ! 1=^ k ; Var u^n;k ! 2 ; k = 1; : : : ; K 0 ;

En

v^n;k

! 1=^k ; Var

v^n;k

u;k 2 ; ! v;k

k = 1; : : : ; K;

and either one of the following conditions be met: (i) supn En (^un;k )2+ < 1; k = 1; : : : ; K 0 ; and supn En (^vn;k )2+ < 1; k = 1; : : : ; K; for some > 0, and b2n= ln n ! 0; (ii) supn En exp((^un;k ) ) < 1; k = 1; : : : ; K 0 ; and supn En exp((^vn;k ) ) < 1; k = 1; : : : ; K; for some > 0 and 0 < 1, and b2n =n =2 ! 0. ld Then the LD convergence (An ; S n ; Rn ) ! (A; S; R) in the hypotheses of Theorem 6.1.8 holds for 2 ; : : : ; 2 ); ATA = diag(A; 1 A;K T 2 2 ); S S = diag(S;1 ; : : : ; S;K T pkl ); if m = l; k k R R = pkl (1 p p if m 6= l; l;m kl km ; k; l; m = 1; : : : ; K:

where

2 = 2 0 2 0 ^3 A;k u;k k ; k = 1; : : : ; K ; A;k = 0; k = K + 1; : : : ; K; 2 = 2 3 S;k v;k ^ k ; k = 1; : : : ; K:

Proof. Since the An , S n and Rn;k ; k = 1; : : : ; K; are mutually independent, by Lemma 3.1.42 it is suÆcient to prove the entry-wise conld ld ld vergence, i.e., An ! AWA , S n ! S WS and Rnk ! R;k WR;k ; k = 1; : : : ; K: All of them follow by Theorem 4.4.8. In more detail, for the LD convergence of the Rn;k we write bX ntc n;k i pnpk ; Rn;k = t b n i=1

© 2001 by Chapman & Hall/CRC

450

LD convergence for queues

where pk = (pkl ; l = 1; : : : ; K ) and in;k ; i 2 N ; are i.i.d. K -vectors, which have one entry equal to 1 and the rest equal to 0, the probability of the lth entry being equal to 1pbeing pkl . Clearly, the conditions of Theorem 4.4.8 are met with bn n as bn and (

()l;m =

pkl (1 pkl ) pkl pkm

if m = l; if m 6= l:

6.2 Very large and moderate deviations for many server queues In this section we derive results on LD convergence for many server queues. We consider a sequence of many server queues with exponential service times and Poisson arrival processes, which may be non-time-homogeneous. Arriving customers who nd no available servers form a queue and are served in the order of arrival. At time t the nth queueing system has Ktn homogeneous servers in parallel, arrival rate nt and service rate nt. We assume that the following expansions hold

p

p

nt = n0;t + nbn 1;t + O( n); p b nt = 0;t + pn 1;t + O(1= n); n p p n Kt = n0;t + nbn 1;t + O( n);

p

(6.2.1a) (6.2.1b) (6.2.1c)

where bn ! 1, bn = n ! 0, the functions 0;t , 1;t , 0;t , 1;t , 0;t , and 1;t are Lebesgue measurable, the functions 0;t , 1;t , 0;t , 1;t , and 1;t are bounded on bounded intervals, and the O's are uniform in t over bounded intervals. We do not rule out the case 0;t = 1, which corresponds to an in nite server queue. Let An = (Ant ; t 2 R+ ) and B n;k = (Btn;k ; t 2 R+ ); k 2 N ; be independent Poisson processes of respective rates nt and nt at time t. We assume that the objects associated with the nth system are de ned on a complete probability space ( n ; Fn ; Pn ). All the processes are considered as random elements of D (R + ; R). Denoting by Qnt the number of customers in the nth system at time t, we have

© 2001 by Chapman & Hall/CRC

451

Many server queues

that distributionally the process Qn = (Qnt ; t equation

Qnt

= Qn0 + Ant

Ktn Zt X

2

1(Qns k) dBsn;k :

k=1 0

R+ )

satis es the (6.2.2)

Let Zt

Mtn = Ant

0

ns ds+

Ktn Zt X k=1 0

1(Qns k)

dBsn;k n;k s ds :

(6.2.3) Then = t 2 R+ ) is a local martingale with respect to the ltration (Ftn ; t 2 R+ ), where Ftn = \>0Gtn+ and Gtn is the sub--algebra of F n generated by Qn0 , Ans ; Bsn;k ; s 2 [0; t]; k = 1; : : : ; K; and sets of Pn -measure zero. The predictable quadraticvariation process of M n has the form

Mn

hM n i

(Mtn ; Fn =

t

=

Zt

0

ns ds +

Zt

0

Qns ^ Ksn ns ds:

(6.2.4)

We write equation (6.2.2) in the form

Qnt

= Qn0 +

Zt

0

ns ds

Zt

0

Qns ^Ksn ns ds+Mtn :

(6.2.5)

Pn The following auxiliary result is standard. We denote by ! convergence in probability. Pn Lemma 6.2.1. Let Qn0 =n ! q0 2 R+ as n ! 1. Then, for T 2 R+ , Qn Pn sup t qt ! 0; t2[0;T ] n

where q = (qt ; t 2 R+ ) is the solution to the dierential equation

q_t = 0;t 0;t (qt ^ 0;t ):

Remark 6.2.2.

(6.2.6)

Equation (6.2.6) has a unique solution by Caratheodory's theorem, see, e.g., Coddington and Levinson [26].

© 2001 by Chapman & Hall/CRC

452

LD convergence for queues

Pn Proof of Lemma 6.2.1. It is easyR to check that RAnt =n2 ! 0. By R (6.2.4) hM n it 0t ns ds + 0t Qns ns ds 0t ns ds + Qn0 t + Rt n n n 2 Pn 0 As s ds so by (6.2.1a), (6.2.1b) and (6.2.1c) hM it =n ! 0 as n ! 1, which implies by the Lenglart-Rebolledo inequality that Pn supt2[0;T ] jMtn j=n ! 0 as n ! 1. Applying a standard tightness argument to (6.2.5) and using the fact the functions 0;t , 1;t , 0;t , and 1;t are bounded on bounded intervals, we conclude that the sequence of laws of the processes Qn =n is relatively compact in distribution, all the limit points satisfying equation (6.2.6) with probability 1. The solution of the latter equation being unique completes the proof.

The next result gives an idempotent diusion approximation for Let pn Qn t q n (6.2.7) Xt = bn n t and X n = (Xtn ; t 2 R+ ).

Qn .

2 Pn1=bn n Let X0 ! x0 2 R as n ! 1. If, in addition, inf s2[0;t] 0;s + 0;s(qs ^ 0;s ) > 0; t 2 R+ ; then X n ld! X as n ! 1 b2n for the Skorohod topology, where X = (Xt ; t 2 R+ ) is the idempotent

Theorem 6.2.3.

diusion speci ed by the equation

X_ t = 1;t 1;t (qt ^ 0;t ) 0;t 1(qt < 1;t )Xt + 1(qt = 0;t )(Xt ^1;t )+ 1(qt > 0;t )1;t q + 0;t + 0;t (qt ^ 0;t )W_ t ; X0 = x0 ; with W = (Wt ; t 2 R+ ) being an idempotent Wiener process. Proof. By (6.2.5), (6.2.6) and (6.2.7) we can write

Xtn

= X0n +

Zt

0

pn n s 0;s ds b n n

Zt

0

© 2001 by Chapman & Hall/CRC

b qs + Xsn pn ^ n

p

Ksn n n ( n bn s

0;s ) ds

453

Many server queues

pn Zt bn

b Kn 1 qs + Xsn pn ^ s qs ^ 0;s 0;s ds + p Mtn : n n nbn

0

Therefore, the rst characteristic of X n without truncation is

B 0n = t

Zt

bns (Xsn ) ds;

0

where

pn n s 0;s bn n pn

1 C~t0n = 2 bn

Zt

0

p

b Kn n n qs + u pn ^ s ( 0;s ) n bn s n b Kn qs + u pn ^ s qs ^ 0;s 0;s : bn n n The modi ed second characteristic without truncation C~ 0n = (C~t0n ; t 2 R+ ) p coincides with the predictable quadratic-variation process of (Mtn =( nbn ); t 2 R+ ) and by (6.2.4) has the form bns (u) =

c~0sn (Xsn ) ds;

where

ns n b Kn + s qs + pn u ^ s : n n n Easy calculations show that for t 2 R+ and v 2 R+ c~0sn (u) =

lim

n!1

Zt

0

lim

n!1

where

ess sup jbns (u) bs (u)j ds = 0; jujv Zt

0

ess sup jc~0sn (u) cs j ds = 0; jujv

bs(u) = 1;s 1;s (qs ^ 0;s ) 0;s 1(qs < 0;s )u + 1(qs = 0;s )(u ^ 1;s) + 1(qs > 0;s )1;s ; cs = 0;s + 0;s(qs ^ 0;s ):

© 2001 by Chapman & Hall/CRC

(6.2.8) (6.2.9)

454

LD convergence for queues

Thus, the convergence conditions of Theorem 5.4.4 are satis ed. The moment conditions are satis ed since the jumps of X n are bounded p above by 1=(bn n). Also the law of the semimaxingale with local characteristics (b; c; 0; 0) is uniquely speci ed by Theorem 2.8.21. ld Therefore, by Theorem 5.4.4 X n ! X at rate b2n as n ! 1.

Remark 6.2.4. We note that the idempotent distribution of X has density

x) = exp

X (

1 2

Z1

x_ t

bt (xt ) 2 dt ct

0

if x is absolutely continuous and x0 = x0 , and X (x) = 0 otherwise, where bt and ct are de ned by the respective equalities (6.2.8) and (6.2.9).

We consider now very large deviations. Let us assume, in addition, that inf s2[0;t] 0;s > 0; t 2 R+ . Let a process Y n = (Ytn ; t 2 R + ) be de ned by Ytn = Qn t =n. Let N1 = (N1 (t); t 2 R + ) and N2 = (N2(t); t 2 R+ ) be independent Poisson idempotent processes on an idempotent probability space ( ; ). Let Y = (Yt ; t 2 R+ ) be a Luzin solution of the equation

Yt = y0 +N1

Z t

0;s ds

N2

0

Zt

(Ys ^0;s)0;s ds ; y0 2 R+ ;

0

such that the idempotent distribution of Y has density

x) = exp

Y (

Z1

0

sup x_ t 2R

(e (e

1)0;t

1)(xt ^ 0;t )0;t dt

if x0 = y0 and x is absolutely continuous, and Y (x) = 0 otherwise. It is well de ned by Theorems 2.6.33, 2.8.10 and 2.8.29. 1=n

Theorem 6.2.5. If Y0n P!n Skorohod topology.

© 2001 by Chapman & Hall/CRC

y0 , then Y n

ld ! n

Y as n

! 1 for the

455

Many server queues

Proof. By (6.2.5)

Ytn

= Y0n +

Zt

0

Zt

ns ds n

0

Ysn ^

Ksn n 1 ds + Mtn : n s n

Therefore, Y n is a squarely integrable semimartingale. Its rst characteristic without truncation is given by

B 0n = t

Zt n

s

0

n

Ksn n n Ys ^ ds; n s

the predictable measure of jumps in view of (6.2.2) is given by

n([0; t];

)=n

Zt n 1

1 2 n n s

0

+ns Ysn ^

Ksn 1 1 n2 n

ds;

and the modi ed second characteristic without truncation in view of (6.2.4) is given by 1 C~t0n = n

Zt n

s

0

n

+ Ysn ^

Ksn n ds: n s

Then the convergence conditions of Theorem 5.4.3 hold with h(x) = x for

bs (u) = 0;s (u ^ 0;s )0;s ; c~s (u) = 0;s + (u ^ 0;s )0;s; s( ) = 0;s 1(1 2 ) + (u ^ 0;s)0;s 1( 1 2 ): The Cramer condition holds since the jumps of Y n are not greater than 1=n. Thus, by Theorem 5.4.3 the sequence of laws of the Y n is C (R + ; R )-exponentially tight of order n and every LD accumulation point for LD convergence of rate n is the law of a semimaxingale with local characteristics (b; 0; ; 0). By Theorem 2.8.29 the latter law is unique, hence, it is the LD limit of the laws of the Y n .

© 2001 by Chapman & Hall/CRC

Appendix A

Auxiliary lemmas This Appendix contains lemmas we referred to in the main body of the book. We rst prove the fact from convex analysis used in the proof of Lemma 1.11.5. For it we adopt the usual de nitions and notation from convex analysis, Rockafellar [117]. For a subset A of a Euclidean space, cl A denotes its closure, ri A the relative interior, rb A = cl Anri A the relative boundary, and conv A the convex hull of A. Let f be a function from Rd , d 2 N , into ] 1; 1]. Its conjugate (or the Legendre{Fenchel transform) f is de ned by f () = sup x f (x) ; 2 Rd ; x2Rd and the bipolar f of f is de ned as the conjugate of f :

f (x) = sup x f () ; x 2 Rd : 2Rd

Obviously, f is convex and lower semi-continuous. We denote by epi f the epigraph of f : epi f = f(x; y) 2 Rd R : y f (x)g; and let dom f = fx 2 Rd : f (x) < 1g denote the eective domain of f . The convex hull conv f of f is de ned by epi(conv f ) = conv(epi f ) and the lower semi-continuous hull cl f by epi(cl f ) = cl(epi f ): If f is convex, then @f (x) denotes the subdierential of f at x. We say that f is essentially strictly convex if it is strictly convex on every convex subset of the set of those x for which the 457 © 2001 by Chapman & Hall/CRC

458

Auxiliary lemmas

set @f (x) is nonempty. If f is essentially strictly convex, then it is strictly convex on ri(dom f ), Rockafellar [117].

Lemma A.1. If f

: Rd !] 1; 1] is a lower semi-continuous function and its bipolar f is strictly convex on ri(dom f ), then f = f . Proof. It is obvious that f f . So we prove the opposite inequality. By Rockafellar [117, Corollary 12.1.1 and the argument below] we have f = cl(conv f ): (A.1) We rst prove that f (x) f (x); x 2 ri(dom f ):

(A.2)

Assume the contrary, i.e., that for some x0 2 ri(dom f ) and > 0 we have f (x0 ) > f (x0 )+ : (A.3) Since x0 2 ri(dom f ), by Rockafellar [117, Theorem 23.4] the set @f (x0 ) is nonempty. Let 0 2 @f (x0 ). Then by the de nition of f f (x) 0 x f (0 ) (A.4) and by Rockafellar [117, Theorem 23.5] f (x0 ) = 0 x0 f (0 ):

(A.5)

It is easy to see that strict convexity of f implies that f (x) > 0 x f (0 ); x 6= x0 :

(A.6)

Indeed, if for some x 6= x0 we had equality in (A.4), then by convexity of f and (A.5) f (z ) 0 z f (0 ); for every z 2 [x0 ; x[, which together with (A.4) would yield f (z ) = 0 z f (0 ); z 2 [x0 ; x[: On the other hand, [x0 ; x[ ri(dom f ) (by Rockafellar [117, Theorem 6.1] and since x 2 dom f if there is equality in (A.4)). Thus, f would fail to be strictly convex on ri(dom f ), and (A.6) is proved.

© 2001 by Chapman & Hall/CRC

Appendix A

459

By lower semi-continuity of f we can choose " > 0 such that "j0 j < =3 and

f (x) > f (x0 ) ; jx x0 j < ": (A.7) 3 For this ", we choose Æ > 0; Æ < =3, satisfying the inclusion fx : 0 x f (0 )+Æ f (x)g fx : jx x0 j < "g: (A.8) In order to show that such a Æ exists, let us denote by AÆ the set on the left of (A.8). Then by (A.5) and (A.6) \

Æ>0

AÆ = fx0 g:

(A.9)

The hyperplane in Rd R de ned by the equation y = 0 x f (0 )+ Æ; x 2 Rd ; y 2 R; is parallel to the hyperplane y = 0 x f (0 ). The latter in view of (A.5) and (A.6) has with epi f the only point x0 in common. Then by Rockafellar [117, Corollary 8.4.1] the sets AÆ are bounded. They are closed since f is lower semi-continuous. Thus, the AÆ are compacts and (A.9) easily implies that for all Æ > 0 small enough AÆ fx : jx x0 j < "g proving (A.8). For the chosen Æ and ", we de ne fÆ;"(x) = max f (x); 0 x f (0 )+ Æ : (A.10) Obviously, fÆ;" is convex, lower semi-continuous and fÆ;"(x0 ) > f (x0 ) by (A.5). If we show that

fÆ;"(x) f (x); x 2 Rd ;

(A.11)

this would contradict (A.1), and (A.2) would be proved. It is clear that (A.11) holds on the set fx : jx x0 j "g since fÆ;"(x) = f (x) for these x by (A.8) and (A.10). If jx x0 j < ", then using (A.5), (A.3) and (A.7) we have

0 x f (0 ) + Æ = 0 (x x0 ) + f (x0 ) + Æ 2 + Æ < f (x) < "j0 j + f (x0 ) + Æ "j0 j + f (x) 3 (the latter inequality holds by the choice of " and Æ). Since, as we noted, f f , this proves (A.11) on fx : jx x0 j < "g. Thus (A.2) is proved.

© 2001 by Chapman & Hall/CRC

460

Auxiliary lemmas

Now if x 2 rb(dom f ) we have by Rockafellar [117, Theorem 7.5] in view of lower semi-continuity of f that for arbitrary z 2 ri(dom f ) f (x) = lim f (1 )z + x : (A.12) "1 By Rockafellar [117, Theorem 6.1] [z; x[ ri(dom f ), and then by the part just proved f (1 )z + x = f (1 )z + x ; 0 < 1; so that by lower semi-continuity of f and (A.12) we have that f (x) f (x) proving the assertion of the lemma for x 2 cl(dom f ). Finally, for x 62 cl(dom f ) we obviously have f (x) = f (x) = 1. For the next lemma we recall that 0 denotes the set of all Rd { valued piecewise constant functions ((t); t 2 R+ ) of the form

(t) =

k X i=1

i 1(t 2 (ti 1 ; ti ]);

where 0 t0 < t1 < : : : < tk ; i 2 Rd ; i = 1; : : : ; k; k 2 N :

Lemma A.2. Let f (t; ); t 2 R+ ; 2 Rd , be an R-valued function, which is Lebesgue R T measurable in t, continuous in , and is such that f (t; 0) = 0 and 0 f (t; )dt is well de ned for T 2 R+ and 2 Rd . Then for T 2 R+ ZT

0

ZT

sup f (t; ) dt = sup f (t; (t)) dt: ((t))20 0

2Rd

(A.13)

Proof. We denote F (t) = sup2Rd f (t; ): Since the supremum may be taken over the rational in view of continuity of f (t; ) in , the function F (t) is Lebesgue measurable and non-negative, so that the integral on the left-hand side of (A.13) is well de ned. Given arbitrary " > 0, we introduce the set

A" = f(t; ) 2 [0; T ]R d :

© 2001 by Chapman & Hall/CRC

1

f (t; ) (F (t) ")+ ^ 1" g:

461

Appendix A

By a measurable selection theorem, see, e.g., Clarke [25], Ethier and Kurtz [48], there exists an Rd {valued Lebesgue measurable function ~ " (t) such that 1 f (t; ~"(t)) (F (t) ")+ ^ 1" ; t 2 [0; T ]: By Luzin's theoremthere exists a continuous function " (t) such that RT 2 ~ 0 1 " (t) 6= " (t) dt < . Then ZT

f (t; " (t)) _ 0 dt

0

ZT

f (t; ~ "(t)) dt

0

ZT

0

1 (F (t) ")+ ^ dt : "

Since ( (t)) is continuous, it can be approximated by functions from 0 . Since f (t; ) is continuous in and f (t; 0) = 0, by Fatou's lemma there exists a function 0 2 0 such that ZT

f (t; 0 (t)) dt

0

ZT

f (t; " (t)) _ 0 dt :

0

Thus, since > 0 is arbitrary, ZT

0

ZT

sup f (t; ) dt sup f (t; (t)) dt: ((t))20 0

2Rd

The reverse inequality is obvious. We state and prove the lemma used in the proof of Theorem 6.1.7. Lemma A.3. Let fin; i 2 N g; n 2 N , be a triangular array of rowwise i.i.d. real-valued r.v. with zero mean on respective probability spaces ( n ; Fn ; Pn ). Let bn ! 1 as n ! 1, and > 0. p (i) If bn = n ! 0 as n ! 1 and, for some " > 0, we have supn En j1n j2+" < 1, then there exist n0 , t0 > 0, C1 > 0, and C2 > 0 such that, for all t t0 and n n0 ,

Pn

1

k X

p max n > t k=1;:::;bntc bn n i=1 i

!

p exp( C1b2n t) + C2

© 2001 by Chapman & Hall/CRC

" 1 b2+ n : (A.14) n"=2 t"=2

462

Auxiliary lemmas

(ii) If, for some > 0 and 2 (0; 1], we have supn En exp( j1n j ) < 1 and b2n =n =2 ! 0 as n ! 1, then there exist n00, t00 > 0, C10 > 0 and C20 > 0 such that, for all t t00 and n n00 ,

Pn

1

k X

p max in > t k=1;:::;bntc bn n i=1

!

p exp( C10 b2n t) 0 p

+ exp( C2 (bn nt) ): (A.15) Proof. The argument uses the ideas of the proof of Theorem 4.4.8. Let the conditions of (i) hold. We rst prove that there exist C1 > 0 and t0 such that for t t0

Pn

max

k=1;:::;bntc bn

1

pn

k X i=1

in 1

p pbnn jinj t > t

!

p exp( C1 b2n t): (A.16)

We denote B = supn En j1n j2+" + 1 and bn n p n ^in = pbn in 1 pbn jin j pt En i 1 p ji j t : n n n (A.17) By Doob's inequality (see, e.g., Liptser and Shiryaev [79, Theorem 1.9.1]), for > 0, bntc ! 2^1n k E e X n 1 t Pn max 2 ^in > eb2n t : (A.18) 2 k=1;:::;bntc bn i=1

p Since En ^1n = 0, j^1n j 2 t and En (^1n )2 En (1n )2 b2n =n, it follows that p p b2 n En e2^1 1+22 e4 t En (^1n )2 1+22 e4 t n B; n so p n bntc En e2^1 exp(22 e4 t Btb2n): p

Choosing in (A.18) = 1= t, we obtain for t t0 = (4e4 B=)2 and C1 = =2

Pn

k 1 X t max 2 ^n > 2 k=1;:::;bntc bn i=1 i

© 2001 by Chapman & Hall/CRC

!

p exp( C1b2n t):

(A.19)

463

Appendix A

Now note that, since En 1n = 0, En n 1

1

p pbnn j1n j t

= En n 1

1

p pbnn j1nj > t " B b1+ n n(1+ ")=2 t(1+")=2 ;

hence,

bntpc E n 1 pbn jnj a b"n Bt(1 ")=2; bn n n 1 n 1 n"=2 p so, by the fact that bn = n ! 0 as n ! 1 and (A.17), for all n large enough and t t0 Pn

max

k=1;:::;bntc bn

1

pn

k X i=1

in

1

p pbnn jinj t > t

!

k 1 X t Pn k=1max ^in > ; 2 2 ;:::;bntc bn i=1

which together with (A.19) proves (A.16). The estimate (A.14) now follows by (A.16) and the inequalities

Pn

max

k=1;:::;bntc bn

Pn

1

pn

k X i=1

!

in

> t !

p b p max in 1 pn jin j t > t n k=1;:::;bntc bn n i=1 p b + Pn max pn jin j > t (A.20) k=1;:::;bntc n 1

k X

and

Pn

p b max pn jin j > t k=1;:::;bntc n

Part (i) is proved.

© 2001 by Chapman & Hall/CRC

p bntcPn pbnn j1nj > t 2+" bntc nb1+n "=2 t1+B"=2 :

464

Auxiliary lemmas

For part (ii), we write

Pn

max

k=1;:::;bntc bn

Pn

1

pn

max

k=1;:::;bntc bn 0

+ Pn @

1

p

k X i=1

1

pn

bX ntc

bn n i=1

!

in > t

k X i=1

in 1

p pbnn jin j t > t2

!

1

bn n p n ji j 1 pn ji j > t > t2 A : (A.21)

Noting that the conditions of part (ii) imply the conditions of part (i), we estimate the rst term on the right of (A.21) with the help of (A.16). For the second, we use the inequality 0

Pn @

bX ntc

1

p

bn n i=1

Pn b n + Pn

1

p t A b n n n ji j 1 pn ji j > t > 2

p pn k=1max j in j > t ;:::;bntc

1

1

p

bntc X

bn n i=1

jinj

1

p pbnn jinj > t 1 n p t p 1 bn n ji j t > 2 : (A.22)

We rst work with the second probability on the right. We have, for > 0 by Chebyshev's inequality

Pn

bX ntc

b p p j in j 1 pn jin j > t bn n i=1 n

1

1 bn1pn jinj pt

© 2001 by Chapman & Hall/CRC

>

t 2

465

Appendix A

p En exp 2 pbnn j1nj 1 pbnn j1nj > t 1 p bntc n p 1 b n j1 j t exp( b2n t)

n

p exp nt log En exp 2 pbnn j1nj 1 pbnn j1nj > t 1 b 1pn j1nj pt b2nt : (A.23)

n

Next, for 0 < < 1; c > 0 and c =4;

p b b 1 n p p j j t En exp 2 pn j1n j 1 pn j1n j > t 1 n n bn n 1 bn n bn n 1 En exp 2 pn j1 j 1 pn j1 j > c 1 n p 1 bnpn j1 j t p b b b n n n n n 1 n + En exp 2 p j1 j 1 p j1 j c 1 p j1 j > t n n n p ! 1 p En exp j1n j exp 2b2n t cb n n

+ En exp

p

2c + j1n j exp 2

p

2

p

nt bn

!

: (A.24)

Taking = 1=(2 t) and c = t=2, and using the condition n =2 =b2n ! 1 as n ! 1, we conclude that the rightmost side p of (A.24) is not greater than exp C~ nt=bn for some C~ > 0. Substituting the estimate into (A.23) and again using the convergence n =2 =b2n ! 1 implies that, for all n and t large enough, bX ntc

b p p j in j 1 pn jin j > t bn n i=1 n 1 1 b pn jinj pt > t2

Pn

1

n

© 2001 by Chapman & Hall/CRC

p exp C100b2n t :

466

Auxiliary lemmas

By a similar argument, this bound is seen to hold for = 1 as well. Finally, the rst term on the right of (A.22) is estimated as

p En e pj1n j n p Pn max j j > t nt (b nt) bn n i=1;:::;bntc i e n p exp( C20 (bn nt) ):

1

Substituting the estimates into (A.21) nishes the proof of (ii).

© 2001 by Chapman & Hall/CRC

Appendix B

Notes and remarks Part I This part considers idempotent analogues of the constructions of probability theory. They also belong to the realm of possibility theory so one can replace the adjective \idempotent" with \possibilistic" (or, perhaps, \fuzzy"). The observation of the analogy between certain probabilistic and \max-plus" constructions seems to have rst been made in Baccelli et al. [8].

Section 1.1 It appears that maxitive measures were rst introduced by Shilkret [119], who also studied properties such as convergence, Egorov's theorem, and others. Idempotent measures are known as possibility measures in fuzzy measure theory, see, e.g., Dubois and Prade [40], Wang and Klir [133], de Cooman, Kerre and Vanmassenhove [33], Pap [101] (who also uses the term \maxitive measure"), Mesiar [86], and references therein; and as A-measures in idempotent measure theory, see Kolokoltsov and Maslov [73] (del Moral in Kolokoltsov and Maslov [73] uses the name \performance measure"). Another name is \cost measure", see Akian, Quadrat and Viot [2, 3]. Both Wang and Klir [133] and Kolokoltsov and Maslov [73] use the requirement of -maxitivity as a de nition and call the property fuzzy additivity and complete additivity, respectively. Pap [101] uses the name \complete maxitivity". Some authors replace the -maxitivity 467 © 2001 by Chapman & Hall/CRC

468

Notes and remarks

property by -maxitivity, see, e.g., Akian [1], Pap [101]. In the topological setting similar objects have been studied by Norberg [95], and O'Brien and Vervaat [97]. The latter authors use the name \supmeasure", which is explained by the \sup-representation" (1.1.2), and require certain inner and outer regularity properties rather than -smoothness. Possibility measures with inner and outer regularity properties on topological spaces have been considered by Janssen, de Cooman and Kerre [68]. Our usage of the concept of -smoothness is consistent with the one adopted in measure theory, see Topse [125], Vakhania, Tarieladze and Chobanyan [126]. A -smooth idempotent measure is a speci c case of a Choquet capacity, see, e.g., Meyer [88] or Neveu [94]. Our study uses some of the ideas as well as the terminology of the theory of Choquet capacities. The de nition of a maxitive set function is due to Norberg [95]. Since the collection Eiu contains the collection of E -analytic (or Suslin) subsets of , see, e.g., Kuratowski and Mostowski [76], Meyer [88] or Neveu [94] for the de nition, Theorem 1.1.7 is a (very simple) analogue of Choquet's theorem, Meyer [88, T19], Neveu [94]. The de nition of a paving is borrowed from Meyer [88]. Theorem 1.1.9 is in the theme of Meyer [88, Theorem IIIT23] and Wang and Klir [133, Theorem 4.9], and runs parallel to the result on the extension of a measure from a ring to a -ring, see, e.g., Halmos [58]. The proof uses the construction of Wang and Klir. One can also obtain the same extension of by applying the construction used by Meyer [88, IIIT23]: de ne, for A 2 Eu , (A) = sup (F ); F 2E F A and, for arbitrary B , let (B ) = inf (A): A2Eu AB However, checking the necessary properties is more complicated. The latter approach is better suited to -maxitive measures, cf. Akian [1]. Objects that we call -algebras have been known in possibility theory as ample elds or complete elds, see Wang [132], De Cooman and Kerre [31, 32], and Wang and Klir [133]. Both our de nition

© 2001 by Chapman & Hall/CRC

Appendix B

469

of and notation for atoms are consistent with Wang [132], and De Cooman and Kerre [31, 32]. Most of the properties of -algebras stated in this section can be found in these papers. Corollary 1.1.22 has been prompted by Neveu [94, Proposition I.6.1].

Section 1.2 Functions measurable with respect to ample elds are called fuzzy variables in fuzzy measure theory, see Wang [132], De Cooman and Kerre [31, 32], and Wang and Klir [133]; Janssen, de Cooman and Kerre [68] use the name \possibilistic variables". Measurability properties for more general set functions are considered in Pap [101]. Images of possibility measures are considered in Wang [132], de Cooman, Kerre and Vanmassenhove [33]. Lemma 1.2.7 is an analogue of Doob's theorem on representation of measurable functions, see, e.g., Meyer [88, IT18]. The proof is also along the lines of the proof given in Meyer [88, IT18]. O'Brien and Vervaat [97] distinguish between tightness and classical tightness. We need only the latter concept for which we reserve the name \tightness". The notion of Luzin measurability with respect to idempotent measures is an analogue of Luzin measurability in measure theory, see Schwartz [118], Vakhania, Tarieladze and Chobanyan [126]. Theorem 1.2.14 is an abstract version of a result in large deviation theory (cf., e.g., Deuschel and Stroock [36]).

Section 1.3 Modes of convergence have also been studied by Shilkret [119] and del Moral in Kolokoltsov and Maslov [73]. Extensions of many of the results of the section to more general set functions are given in Pap [101]. For analogues in probability theory see, e.g., Shiryaev [120].

Section 1.4 The notion of idempotent integral has been introduced by Shilkret [119], who also studied its basic properties, but the name seems to

© 2001 by Chapman & Hall/CRC

470

Notes and remarks

be due to Maslov [84, 85]. Similar constructions appear in Norberg [95], see also Vervaat [130]. More general integrals are studied in fuzzy measure theory, see Dubois and Prade [40], de Cooman, Kerre and Vanmassenhove [33], Wang and Klir [133], Wu, Wang, and Ma [137], Pap [101], de Cooman and Kerre [31], Mesiar [86], Guo, Zhang, and Wu [56], Mesiar and Pap [87], and references therein; integrals of lattice-valued functions have been considered in Akian [1], de Cooman and Kerre [31], Mesiar and Pap [87], Pap [101]. Lemma 1.4.5 also holds for so called pan integrals, Wang and Klir [133]. Part 1 of Lemma 1.4.7 appears in Kolokoltsov and Maslov [73], Theorem 1.4.11 is stated by del Moral in Kolokoltsov and Maslov [73], it also appears in Puhalskii [111]. SuÆciency of the existence of the function F for uniform maximability in Corollary 1.4.14 is stated by del Moral in Kolokoltsov and Maslov [73], who also studies convergence of idempotent integrals. Our analysis of the convergence properties is based on Puhalskii [111]. For convergence properties in a more general setting see Pap [101]. The proof of Theorem 1.4.22 uses the ideas of the proof of Daniell's theorem in Meyer [88, III.2.24]. For another form of the Daniell property see Pap [101].

Section 1.5 Products of idempotent measures have been studied in more generality in fuzzy set theory, see, e.g., de Cooman, Kerre and Vanmassenhove [33], Janssen, de Cooman and Kerre [68], and references therein; they have also been analysed by Kolokoltsov and Maslov [73]. Products of ample elds have been considered by Wang [132].

Section 1.6 Exposition is based on Puhalskii [111], who however conditions on collections of analytic sets rather than -algebras. The notions of independence and conditioning for idempotent variables in the fuzzy set theory context have been studied in Wang [132], de Cooman, Kerre and Vanmassenhove [33]. Similar properties as well as a de nition of conditional idempotent expectation with respect to -algebras are considered by del Moral in Kolokoltsov and Maslov [73] (see Re-

© 2001 by Chapman & Hall/CRC

Appendix B

471

mark 1.6.26). For other approaches see Akian, Quadrat and Viot [2, 3]. Absolute continuity has been studied in the fuzzy set theory context, see, e.g., Mesiar [86] and references therein; a general treatment appears in Pap [101]; however, the de nitions are stated for -algebras and the weaker notion of absolute continuity (cf. Remark 1.6.30).

Section 1.7 The concept of Luzin measurability for idempotent variables on topological spaces has been introduced in Puhalskii [111], for the measuretheoretic analogue see Schwartz [118], and Vakhania, Tarieladze and Chobanyan [126]. The rst result in the theme of Theorems 1.7.21, 1.7.23, and 1.7.25 seems to be due to Choquet [24]. Theorems 1.7.21 and 1.7.25 appear in Breyer and Gulinski [17]; our proof of Theorem 1.7.25 uses their idea of invoking the Stone-Czech compacti cation. Puhalskii [107, 108] proves a similar result for metric spaces under the additional condition of sub-additivity of the functional V . Kolokoltsov and Maslov [73, Theorem 1.5, ch.1] prove the result of Theorem 1.7.21 for a locally compact normal space and functions with values in an idempotent metric semiring. They also prove the stated representation for the case where V is a continuous homeomorphism from Cb+(E ) equipped with the topology of pointwise convergence to an idempotent metric semiring and E is Tihonov. Akian [1] considers the same characterisation in terms of continuity for integrals of lattice-valued functions. Algebraic versions appear in Litvinov, Maslov and Shpiz [80, 81].

Section 1.8 In this section we use some of the ideas of Schwartz [118]. The result in Lemma 1.8.3 is a special case of a result in large deviation theory due to Dawson and Gartner [28], see also Dembo and Zeitouni [35]. The setting of Theorem 1.8.6 for regular possibility measures has been considered by Janssen, de Cooman and Kerre [68].

© 2001 by Chapman & Hall/CRC

472

Notes and remarks

Sections 1.9 and 1.10 The results are modelled after weak convergence theory of probability measures, see Billingsley [11], Parthasarathy [102], Topse [125, 124], Vakhania, Tarieladze and Chobanyan [126]. For a prototype see Vervaat [129]. The setting of metric spaces is studied in Puhalskii [108]. For facts about uniform spaces used in the proof of Theorem 1.9.2 see, e.g., Engelking [47]. Theorem 1.9.28 is an analogue of Ranga Rao's result, see Vakhania, Tarieladze and Chobanyan [126]. More general compactness results and other properties of the vague topology are in O'Brien and Vervaat [97] and O'Brien and Watson [99]. For the de nitions of the Prohorov and Kantorovich-Wasserstein metrics for probability measures see, e.g., Dudley [41]. Jiang and O'Brien [69] de ne the Prohorov metric on a space of set functions that includes idempotent probability measures and probability measures and show, in particular, that it metrises convergence of sequences in the narrow topology; they also show that the KantorovichWasserstein metric has this property for sequences of the exponentials of rate functions and address the issue of characterising tight collections as totally bounded sets. For the de nition and properties of Mosco convergence see Mosco [92], Zabell [138] and references therein.

Section 1.11 Kolokoltsov and Maslov [73] refer to the Laplace-Fenchel transform as the Fourier-Legendre transform. The inversion formula appears in Puhalskii [108]. Lemma 1.11.19 is also taken from the latter paper. For required facts from convex analysis see Rockafellar [117] and Appendix A.

Sections 2.1 { 2.6 The results and approaches are analogous to those in stochastic calculus, see Dellacherie [34], Elliott [45], Ikeda and Watanabe [66], Jacod and Shiryaev [67], Liptser and Shiryaev [79], Meyer [88], Neveu [94], ksendal [100], and Stroock and Varadhan [123]. Idempotent martingales have been considered by Del Moral in Kolokoltsov and Maslov [73] (for conditional expectations with respect to -algebras).

© 2001 by Chapman & Hall/CRC

473

Appendix B

For other approaches see Akian, Quadrat and Viot [2, 3]. \Possibilistic" processes have been studied in Janssen, de Cooman and Kerre [68]. Theorems 2.2.26 and 2.2.27 are taken from Puhalskii [108]. Section 2.3 is based on Puhalskii [111]. The de nitions of the idempotent Wiener and Poisson processes in Section 2.4 are motivated by the fact that the associated rate functions appear in the large deviation principles for Wiener and Poisson processes, respectively, see, e.g., Borovkov [13], Freidlin and Wentzell [51]. For the properties of the pseudo-inverses of matrices see, e.g., Campbell and Meyer [20]. Theorem 2.6.22 is in essence the Picard-Lindelof-Caratheodory theorem, see, e.g., Coddington and Levinson [26], Hartman [60].

Sections 2.7 The results are based on Puhalskii [109, 111]. We follow the ideas of stochastic calculus, see, e.g., Liptser and Shiryaev [79], Jacod and Shiryaev [67]; in particular, the de nition of a semimaxingale is analogous to the exponential characterisation of semimartingales. Del Moral in Kolokoltsov and Maslov [73] has considered idempotent semimartingales for conditional expectations with respect to algebras. Lemma 2.7.5 is in essence due to Liptser and Shiryaev [79, Theorem 6.2.3], whose argument also applies to the proof. Theorem 2.7.16 admits a revealing interpretation in terms of Orlicz spaces, Krasnosel'skii and Rutickii [75]. Speci cally, for x 2 C and t 2 R+ , let Lg^(x) (0; t) denote the set of all functions f (s); s t; such that Zt

0

g^s

1 f (s); x ds < 1;

for some > 0. Lg^(x) (0; t) is easily seen to be a vector space. Let for f 2 Lg^(x) (0; t)

Zt

kf k Lg^( )(0;t) = inf > 0 : g^s x

© 2001 by Chapman & Hall/CRC

0

1 f (s); x ds 1 :

474

Notes and remarks

This can be shown to de ne a seminorm on Lg^(x) (0; t), which is a norm if g^s (; x) 6= 0 for 6= 0 (cf. Krasnosel'skii and Rutickii [75]). Let us assume, for the moment, that g^s (; x) does not depend on s: g^s (; x) = g^(; x). Then the above norm is called the Luxembourg norm and Lg^(x) (0; t) is called an Orlicz space, Krasnosel'skii and Rutickii [75]. Also in this case the set of functions, for which (2.7.18) holds, is the closure of the space of bounded functions in the Luxembourg norm, Krasnosel'skii and Rutickii [75]. In analogy with Krasnosel'skii and Rutickii [75], we denote this set by Eg^(x) (0; t). We then have the following insight into the statement of Theorem 2.7.16. Let Lg^(x) (0; t) be the set of functions f such that Zt

0

g^(f (s); x) ds < 1:

Then, by Krasnosel'skii and Rutickii [75], we have the strict inclusions

Eg^(x) (0; t) Lg^(x) (0; t) Lg^(x) (0; t);

unless g^(; x) satis es the weak growth condition (or the 2 { condition):

g^(2; x) lim sup < 1: !1 g^(; x)

If g(; x) has the semimartingale representation (2.7.55), then the weak growth condition means that Ks (Rd ; x) = Ls (Rd ; x) = 0, i.e., it holds only in \the diusion case". Hence, generally, the class of functions , for which we have proved that Z () is a -local exponential maxingale, is smaller than the class de ned by the condition Zt

0

jgs ((s; x); x)j ds < 1:

We do not know if Theorem 2.7.16 can be extended to a larger set of functions . We also note that the proof of Theorem 2.7.16 has been prompted by the methods of the theory of Orlicz spaces.

© 2001 by Chapman & Hall/CRC

Appendix B

475

Section 2.8 The results are based on Puhalskii [112]. Our conditions for uniqueness of solutions to maxingale problems are similar to conditions required for corresponding martingale problems. It is thus instructive to compare our results with those for martingale problems in Ikeda and Watanabe [66], Jacod and Shiryaev [67], Stroock and Varadhan [123]. The setting of Theorem 2.8.5 corresponds to the situation where a martingale problem is speci ed by a deterministic triplet of predictable characteristics so that the associated process is a process with independent increments; the problem then has a unique solution, see, e.g., Jacod and Shiryaev [67, Theorem III.2.16]. The function (s; x; y) in the hypotheses of Theorem 2.8.27 exists and equals the gradient rhs (y; x) if the latter exists for (almost all) s 2 R+ ; y 2 G and x -almost all x, and is bounded on the sets [0; t] K Gm , where t 2 R+ ; m 2 N and K is compact in C . The role of conditions I and II and the conditions in Theorem 2.8.33 is analogous to the role of conditions A{E in Wentzell [134]. A distinctive feature of our conditions is that they are stated only in terms of the cumulant and do not invoke its Fenchel{Legendre transform (as in Wentzell [134]). We believe that this makes the conditions easier to check. Also we relax the requirements on boundedness and continuity of the cumulant. Condition (2.8.14) is analogous to condition III in Liptser and Puhalskii [78]. The regularisation approach of Lemma 2.8.26 applied later in the section has earlier been used in Wentzell [134], and Liptser and Puhalskii [78] in the large deviation setting. For background on the notions used in Lemma 2.8.31 see Aubin and Cellina [5], von Leichtweiss [131], and Rockafellar [117]. For measurable-selection theorems see Clarke [25, Theorem 4.1.1], Ethier and Kurtz [48], or Dellacherie [34, IT37].

Part II Exposition is based on Puhalskii [106] { [114]. Standard manuals on large deviation theory are Freidlin and Wentzell [51], Varadhan [128], Deuschel and Stroock [36], Dembo and Zeitouni [35].

© 2001 by Chapman & Hall/CRC

476

Notes and remarks

Section 3.1 Some of the results of this section are large deviation convergence versions of the results formulated in the setting of the large deviation principle, see Varadhan [128], Stroock [122], Deuschel and Stroock [36], Dembo and Zeitouni [35], Bryc [18], Dinwoodie [37]. For an approach from the point of view of convergence of capacities see O'Brien and Vervaat [97], O'Brien [96], O'Brien and Watson [99]. Since this section considers similar issues as Section 1.9, most of the comments to that section apply here. In particular, there are many analogies with results in weak convergence theory, see, Billingsley [11], Parthasarathy [102], Topse [125, 124], Vakhania, Tarieladze and Chobanyan [126]. The de nition of the large deviation convergence and the term itself have been introduced in Puhalskii [108, 109]. Properties of a more general type of convergence have been considered by Pap [101], Mesiar and Pap [87]. Theorem 3.1.3 for the setting of metric spaces has appeared in Puhalskii [107]. It combines a number of earlier results. The fact that part 3 implies part 2 is \Varadhan's lemma", Varadhan [127], who also proves Lemma 3.1.12, the converse under the additional condition of exponential tightness is due to Bryc, see Varadhan [128], Bryc [18], Dembo and Zeitouni [35]. Instead of the de nition we have adopted for large deviation convergence one could use part 2 of Theorem 3.1.3 in order to de ne \weak large deviation convergence". It would then be equivalent to the large deviation principle, or \narrow large deviation convergence", cf. Remark 1.9.6. The name \contraction principle" as given by Varadhan [128] refers to the case of continuous f in Corollary 3.1.15. Corollary 3.1.15 for convergence of sequences has appeared in Puhalskii [106], it is referred to in Puhalskii and Whitt [113] as the extended contraction principle. Theorem 3.1.14 for metric spaces and sequences of probability measures has been proved in Puhalskii [110]; Chaganty [22] proves the statement in the setting of Polish spaces under the assumption that the convergence fn(zn ) ! f (z ) holds for every z 2 E , a similar type of condition has been used by Dinwoodie and Zabell [38]. Other versions and generalisations have been considered by Deuschel and Stroock [36] and O'Brien [96]. Theorems 3.1.19 and 3.1.28 are analogues of Prohorov's criterion for weak convergence, Prohorov [104], see also Billingsley [11],

© 2001 by Chapman & Hall/CRC

Appendix B

477

Vakhania, Tarieladze and Chobanyan [126]. Theorem 3.1.28 has appeared in Puhalskii [106]. The proofs here are along the lines of the one in Puhalskii [107], who also mentions the extension to Tihonov spaces. A more general setting has been considered by O'Brien and Vervaat [97] so that Theorems 3.1.19 and 3.1.28 for separable metric spaces follow from Theorem 3.11 and Lemma 5.2 there; an announcement had been made in Vervaat [129]. Lynch and Sethuraman [82] have proved that LD convergence implies exponential tightness for sequences of probability measures on complete separable metric spaces; an extension appears in Jiang and O'Brien [69]. De Acosta [30] has generalised and simpli ed the proof of Puhalskii [106] to give weaker conditions for part 1 of Theorem 3.1.19 to hold, in particular, extending the result to Hausdor topological spaces (similarly to Prohorov's tightness criterion in weak convergence theory, see Topse [125]) and probabilities on non-Borel -algebras. For other versions and extensions see Vervaat [129], O'Brien and Vervaat [98]. The vague large deviation convergence has extensively been studied by O'Brien and Vervaat [97, 98] who prove Theorem 3.1.34. Jiang and O'Brien [69], considering a more general setting, prove that the Prohorov and the Kantorovich-Wasserstein metrics metrise LD convergence of sequences. They also address the issue of characterising tight collections as totally bounded sets. See also Dembo and Zeitouni [35] for some of the results. Theorem 3.1.31, which is usually referred to as the GartnerEllis theorem, has been proved by Gartner [53] for the case where G() is smooth and nite everywhere. The extension to the essentially smooth case has been obtained by Freidlin and Wentzell [51] and, later and apparently independently, by Ellis [46]. The name \Gartner-Ellis" appears to have been rst used in Bucklew [19]. Our method of proof follows Puhalskii [109] and O'Brien and Vervaat [98]. Theorem 3.1.32 is an analogue of Ranga Rao's result in weak convergence, see Vakhania, Tarieladze and Chobanyan [126]. In a somewhat dierent form it appears in Jiang and O'Brien [69]. Part 1 of Lemma 3.1.42 in the form of the LDP extends to regular spaces, see Cegla and Klimek [21].

© 2001 by Chapman & Hall/CRC

478

Notes and remarks

Section 3.2 For more background on the concepts and properties concerning stochastic processes on Skorohod spaces the reader is referred to Ethier and Kurtz [48], Jacod and Shiryaev [67], Lindvall [77], Liptser and Shiryaev [79], Skorohod [121], and Whitt [135]. The notion of C -exponential tightness has been introduced in Liptser and Puhalskii [78]. Theorem 3.2.3 is an analogue of Aldous' tightness condition, Aldous [4], and has appeared in Puhalskii [106]. Lemma 3.2.5 can also be used in order to give a somewhat dierent proof of Aldous' original result. Feng and Kurtz [50] obtain dierent exponential tightness conditions. For the Lenglart-Rebolledo inequality, see, e.g., Liptser and Shiryaev [79, Theorem 1.9.3]. Theorems 3.2.8 and 3.2.9, taken from Puhalskii [106, 107, 109, 111], are analogues of the methods of nite-dimensional distributions and the martingale problem in weak convergence theory, cf. Jacod and Shiryaev [67], Liptser and Shiryaev [79], Ethier and Kurtz [48]. Lemmas 3.2.11 and 3.2.13 are LD convergence versions of results in Puhalskii and Whitt [113], which are also more general. See also Lemma 4.2 there for more detail about the proof of Lemma 3.2.13. Prototypes for weak convergence are in Whitt [135].

Chapter 4 The method of nite-dimensional distributions is essentially an adaptation of projective limit arguments, see Dawson and Gartner [28], to the setting of stochastic processes. It was rst used by Varadhan [127], see also Dembo and Zeitouni [35]. Our exposition follows Puhalskii [108, 109]. The main results, both in content and form, are analogous to results on convergence in distribution of a sequence of semimartingales to a process with independent increments in Liptser and Shiryaev [79] and Jacod and Shiryaev [67]. For results on the LDP for processes with independent increments see Varadhan [127], Borovkov [13], Mogulskii [90, 91], Lynch and Sethuraman [82], de Acosta [29], Puhalskii [106]. Lemma 4.1.1 is an analogue of results in Liptser and Shiryaev [79, Chapter 2, x3], Jacod and Shiryaev [67, Chapter 2, x2d] for complex-valued stochastic exponentials. Theorem 4.1.2 is an ana-

© 2001 by Chapman & Hall/CRC

479

Appendix B

logue of Jacod and Shiryaev [67, Theorem VIII.2.30]. Lemma 4.2.6 is a variation on the theme of Polya's theorem, see, e.g., Liptser and Shiryaev [79, Problem 5.3.2]. The proof of Theorem 4.2.11 uses the method of the proof of Theorem 5.4.1 in Liptser and Shiryaev [79]. In particular the fundamental decomposition (LS ) originates in Lemma 5.4.1 there. In the proof of Theorem 4.2.11 condition (^ ) has been used only while proving convergence ). Since ) clearly holds under the condition

(log ^)

1 X f (r x) s r 0<st

Zt

ln 1 + f (rx) s

f (x) Ls ln 1 + f (x) Ls ) ds

0

1=r P

! 0

as 2 ; t 2 U; f 2 Cb ;

and, as the proof shows, condition (^ ) implies condition (log ^), it follows that the theorem holds when condition (^ ) is replaced by condition (log ^). One can show that these two conditions are actually equivalent. Corollary 4.3.5 and its proof are analogous to Proposition VIII.3.40 in Jacod and Shiryaev [67]. Condition (L2 ) and Lemma 4.3.9 originate from Djellout [39]. The argument of the proof of Theorem 4.4.6 follows Djellout [39]. LDPs for partial-sum processes have been studied in Varadhan [127], Borovkov [13], and Mogulskii [90, 91]. Conditions (4.4.10) have been found by Ibragimov and Linnik [61, Theorem 13.1.1] who studied exact asymptotics in the nonfunctional case and showed that the moment condition in (4.4.10) is in a certain sense necessary, see [61, Theorem 13.1.2]. Mogulskii [90, Theorem 1] has established a functional LDP under (4.4.10) for the setting of Theorem 4.4.6. Logarithmic asymptotics for sums of i.i.d.r.v. under (4.4.9) can also be derived from the estimates of the convergence rate in the CLT in Ibragimov and Linnik [61, Chapter 3] and Petrov [103, Chapter 5]. Example 4.4.13 is motivated by Theorem VIII.3.43 in Jacod and Shiryaev [67].

© 2001 by Chapman & Hall/CRC

480

Notes and remarks

Chapter 5 The results are based on Puhalskii [111, 112] (note, however, that our condition (NE ) is somewhat stronger than in Puhalskii [111], so a correction needs to be made in that paper). The majoration and continuity conditions are similar to those used in weak convergence theory, see, e.g., Jacod and Shiryaev [67, VI.3.34]. Theorems 5.3.3, 5.3.4 and 5.3.5 are analogues of respective Theorems 8.2.1, 8.4.2 and 8.3.1 in Liptser and Shiryaev [79]. For the case where At is continuous Theorem 5.3.7 is an analogue of Theorem 8.4.1 in Liptser and Shiryaev [79]. Theorem 5.4.3 is an analogue of Theorems IX.4.8 and IX.4.15 in Jacod and Shiryaev [67]. The setting of Liptser and Puhalskii [78], who have considered large deviations for quasi-continuous processes with the Cramer condition on the jumps and linearly growing coeÆcients, is a special case of the setting of Theorem 5.3.3 for the case where conditions (A)loc + (a)loc are checked by checking (Ie )loc . In particular, Theorem 2.2 there follows by Theorem 5.3.7. As for Theorem 2.1 of Liptser and Puhalskii [78], it is not quite clear if it follows from our results since both results require some implicit conditions on the rate functions, which are diÆcult to compare. However, for the case of explicit suÆcient conditions given by Theorem 9.1 in Liptser and Puhalskii [78], one can obtain Theorem 2.1 there as a corollary of Theorems 5.3.3 and 2.8.33. The Markov setting has been analysed in Freidlin and Wentzell [51], Wentzell [134], Dupuis and Ellis [43], Feng [49], Feng and Kurtz [50], see also Azencott [7], Baldi [9], Baldi and Chaleyat-Maurel [10], Friedman [52], Makhno [83], Micami [89], Narita [93], Remillard and Dawson [116]. Diusions with dependence on the past have been considered by Cutland [27]. Markov processes with discontinuous statistics have been studied by Blinovskii and Dobrushin [12], Dupuis, Ellis and Weiss [44], Dupuis and Ellis [42], Korostelev and Leonov [74], Ignatyuk, Malyshev and Scherbakov [65]. Theorem 5.4.1, taken from Puhalskii [111], is a large-deviation analogue of a result that derives convergence of Markov semigroups from convergence of associated generators, see, e.g., Ethier and Kurtz [48]. Feng [49] has obtained a similar result for Markov processes with values in general metric spaces by the methods of nonlinear semigroup convergence, see Feng and Kurtz [50] for further develop-

© 2001 by Chapman & Hall/CRC

Appendix B

481

ments in this direction. Theorem 5.4.4 relaxes the assumptions of Theorems 4.4.1, 4.4.2, 4:4:20 , and 4.4.3 in Wentzell [134], and Theorem 5.4.2 in combination with Theorem 2.8.33 relaxes the assumptions of Wentzell's Theorems 4.3.1, 3.2.3 and 3.2.3'. The main improvements are that we do not require that either bs (u), or cs (u), or s(dx; u), or gs (; u) be either continuous in the time variable or bounded, and the associated convergences may take place only locally uniformly, not necessarily uniformly on the entire space. Theorem 5.4.8 has been motivated by Theorem 4.5.2 in Wentzell [134]. A similar result can be proved for discrete-time processes. Example 2 is prompted by Problem 8.4.1 in Liptser and Shiryaev [79]. The proof of (5.4.11) is based on Liptser's idea (private communication).

Section 6.1 The results are based on Puhalskii [114], where more detail is given. They complement results on diusion approximation for queues in Kingman [72], Prohorov [105], Iglehart and Whitt [64], Borovkov [15], and Reiman [115]. Theorem 6.1.7 is an analogue of diusion approximation results in Prohorov [105] and Borovkov [15]; the proof borrows ideas used in these proofs. The results of Subsection 6.1.2 are inspired by Reiman [115].

Section 6.2 Theorem 6.2.3 complements diusion approximation results by Iglehart [62, 63], Borovkov [14, 16], Hal n and Whitt [57], and Whitt [136]. For other results on large deviation asymptotics for many server queues see Glynn [55] and Zajic [139].

© 2001 by Chapman & Hall/CRC

Bibliography [1] M. Akian. Densities of idempotent measures and large deviations. Trans. Am. Math. Soc., 351(11):4515{4543, 1999. [2] M. Akian, J.-P. Quadrat, and M. Viot. Bellman processes. In 11th Conference on Analysis and Optimization of Systems: Discrete Event Systems, volume 199 of Lecture Notes in Control and Information Sciences. Springer, 1994. [3] M. Akian, J.-P. Quadrat, and M. Viot. Duality between probability and optimization. In J. Gunawerdena, editor, Idempotency. Cambridge University Press, 1998. [4] D. Aldous. Stopping times and tightness. Ann. Prob., 6:335{ 340, 1978. [5] J.-P. Aubin and A. Cellina. Dierential Inclusions. Springer, 1984. [6] J.-P. Aubin and I. Ekeland. Applied Nonlinear Analysis. Wiley, 1984. [7] R. Azencott. Grandes deviations et applications. In Lecture Notes Math., volume 774, pages 1{176. Springer, 1980. [8] F. Baccelli, G. Cohen, G.J. Olsder, and J.P. Quadrat. Synchronisation and Linearity: an Algebra for Discrete Event Systems. Wiley, 1992. [9] P. Baldi. Large deviations for diusion processes with homogenization and applications. Ann. Prob., 19(2):509{524, 1991. [10] P. Baldi and M. Chaleyat-Maurel. An extension of VentcelFreidlin estimates. In Lecture Notes Math., volume 1316, pages 305{327. Springer, 1988. 483 © 2001 by Chapman & Hall/CRC

484

Bibliography

[11] P. Billingsley. Convergence of Probability Measures. Wiley, 1968. [12] V.M. Blinovskii and R.L. Dobrushin. Process level large deviations for a class of piecewise homogeneous random walks. In The Dynkin Festschrift: Markov Processes and their Applications, pages 1{59. Birkhauser, 1994. [13] A.A. Borovkov. Boundary-value problems for random walks and large deviations in function spaces. Th. Prob. Appl., 12(4):575{595, 1967. [14] A.A. Borovkov. On limit laws for service processes in multi{ channel systems. Siberian Math. J., 8:746{762, 1967 (in Russian). [15] A.A. Borovkov. Stochastic Processes in Queueing Theory. Nauka, 1972 (in Russian, English translation: Springer, 1976). [16] A.A. Borovkov. Asymptotic Methods in Queueing Theory. Nauka, 1980 (in Russian, English translation: Wiley, 1984). [17] V.V. Breyer and O.V. Gulinsky. Large deviations in in nite dimensional vector spaces. Preprint MIPT 96-5, Moscow Institute of Physics and Technology, 1996 (in Russian). [18] W. Bryc. Large deviations by the asymptotic value method. In M. Pinsky, editor, Diusion processes and related problems in analysis, pages 447{472. Birkhauser, 1990. [19] J.A. Bucklew. Large Deviations Techniques in Decision, Simulation, and Estimation. Wiley, 1990. [20] S.L. Campbell and C.D. Meyer, Jr. Generalized Inverses of Linear Transformations. Pitman, 1979. [21] W. Cegla and M. Klimek. Criterion for the large deviation principle. Proc. Roy. Irish Acad., 90A(1):5{10, 1990. [22] N.R. Chaganty. Large deviations for joint distributions and statistical applications. Technical Report TR93-2, Department of Mathematics and Statistics, Old Dominion University, Norfolk, Va, 1993.

© 2001 by Chapman & Hall/CRC

Bibliography

485

[23] H. Chen and W. Whitt. Diusion approximations for open queueing networks with service interruptions. Queueing Systems, 13:335{359, 1993. [24] G. Choquet. Theory of capacities. Ann. Inst. Fourier, 5:131{ 295, 1955. [25] F.H. Clarke. Optimization and Nonsmooth Analysis. Wiley, 1983. [26] E.A. Coddington and N. Levinson. Theory of Ordinary Dierential Equations. McGraw-Hill, 1955. [27] N.J. Cutland. An extension of the Ventcel-Freidlin large deviation principle. Stochastics, 24:121{149, 1988. [28] D.A. Dawson and J. Gartner. Large deviations from the McKean-Vlasov limit for weakly interacting diusions. Stochastics, 20:247{308, 1987. [29] A. de Acosta. Large deviations for vector-valued Levy processes. Stoch. Proc. Appl., 51:75{115, 1994. [30] A. de Acosta. Exponential tightness and projective systems in large deviation theory. In Festschrift for Lucien Le Cam, pages 143{156. Springer, 1997. [31] G. de Cooman and E. Kerre. Possibility and necessity integrals. Fuzzy Sets and Systems, 77:207{227, 1996. [32] G. de Cooman and E.E. Kerre. Ample elds. Simon Stevin, 67:235{244, 1993. [33] G. de Cooman, E.E. Kerre, and F.R. Vanmassenhove. Possibility theory: an integral theoretic approach. Fuzzy Sets and Systems, 46:287{299, 1992. [34] C. Dellacherie. Capacites et Processus Stochastiques. Springer, 1972. [35] A. Dembo and O. Zeitouni. Large Deviations Techniques and Applications. Springer, second edition, 1998.

© 2001 by Chapman & Hall/CRC

486

Bibliography

[36] J.D. Deuschel and D.W. Stroock. Large Deviations. Academic Press, 1989. [37] I.H. Dinwoodie. Identifying a large deviation rate function. Ann. Prob., 21(1):216{231, 1993. [38] I.H. Dinwoodie and S.L. Zabell. Large deviations for sequences of mixtures. In J.K. Ghosh et al., editor, Statistics and Probability. A Bahadur Festschrift. Wiley, 1993. [39] H. Djellout. Moderate deviations for martingale dierences, 2000 (submitted for publication). [40] D. Dubois and H. Prade. Possibility Theory. Plenum Press, 1988. [41] R.M. Dudley. Real Analysis and Probability. Wadsworth & Brooks/Cole, 1989. [42] P. Dupuis and R. Ellis. Large deviations for Markov processes with discontinuous statistics. II. Prob. Th. Rel. Fields, 91:153{ 194, 1992. [43] P. Dupuis and R. Ellis. A Weak Convergence Approach to the Theory of Large Deviations. Wiley, 1997. [44] P. Dupuis, R.S. Ellis, and A. Weiss. Large deviations for Markov processes with discontinuous statistics. I. Ann. Prob., 19:1280{1297, 1991. [45] R.J. Elliott. Stochastic Calculus and Applications. Springer, 1982. [46] R.S. Ellis. Large deviations for a general class of random vectors. Ann. Prob., 12(1):1{12, 1984. [47] R. Engelking. General Topology. PWN, 1977. [48] S.N. Ethier and T.G. Kurtz. Markov Processes. Characterization and Convergence. Wiley, 1986. [49] J. Feng. Martingale problems for large deviations of Markov processes. Stoch. Proc. Appl., 81(2):165{216, 1999.

© 2001 by Chapman & Hall/CRC

Bibliography

487

[50] J. Feng and T.G. Kurtz. Large deviations for stochastic processes (preliminary manuscript), 2000. [51] M.I. Freidlin and A.D. Wentzell. Random Perturbations of Dynamical Systems. Nauka, 1979 (in Russian, English translation: Springer, 1984). [52] A. Friedman. Stochastic Dierential Equations and Applications, volume 2. Academic Press, 1976. [53] J. Gartner. On large deviations from the invariant measure. Th. Prob. Appl., 22(1):24{39, 1977. [54] I.I. Gihman and A.V. Skorohod. Stochastic Dierential Equations and their Applications. Naukova Dumka, 1982 (in Russian). [55] P.W. Glynn. Large deviations for the in nite server queue in heavy traÆc. In Stochastic networks, volume 71 of IMA Vol. Math. Appl., pages 387{394. Springer, 1995. [56] C. Guo, D. Zhang, and C. Wu. Generalized fuzzy integrals of fuzzy-valued functions. Fuzzy Sets and Systems, 97:123{128, 1998. [57] S. Hal n and W. Whitt. Heavy-traÆc limits for queues with many exponential servers. Oper. Res., 29:567{588, 1981. [58] P.R. Halmos. Measure Theory. Springer, 1974. [59] J.M. Harrison and M.I. Reiman. Re ected Brownian motion on an orthant. Ann. Prob., 9:302{308, 1981. [60] P. Hartman. Ordinary Dierential Equations. Wiley, 1964. [61] I.A. Ibragimov and Yu.V. Linnik. Independent and Stationary Related Random Variables. Nauka, 1965 (in Russian). [62] D.L. Iglehart. Limit diusion approximations for the many server queue and the repairman problem. J. Appl. Prob., 2:429{ 441, 1965. [63] D.L. Iglehart. Weak convergence of compound stochastic processes. Stoch. Proc. Appl, 1:11{31, 1973.

© 2001 by Chapman & Hall/CRC

488

Bibliography

[64] D.L. Iglehart and W. Whitt. Multiple channel queues in heavy traÆc, I and II. Adv. Appl. Prob., 2:150{177 and 355{369, 1970. [65] I.A. Ignatyuk, V. Malyshev, and V.V. Scherbakov. Boundary eects in large deviation problems. Russ. Math. Surv., 49(2):41{99, 1994. [66] N. Ikeda and S. Watanabe. Stochastic Dierential Equations and Diusion Processes. North Holland, second edition, 1989. [67] J. Jacod and A.N. Shiryaev. Limit Theorems for Stochastic Processes. Springer, 1987. [68] H. Janssen, G. de Cooman, and E.E. Kerre. A DaniellKolmogorov theorem for supremum preserving upper probabilities. Fuzzy Sets and Systems, 102(3):429{444, 1999. [69] T. Jiang and G.L. O'Brien. The metric of large deviation convergence. J. Theoret. Prob., 13(3):805{823, 2000. [70] L.V. Kantorovich and G.P. Akilov. Functional Analysis in Normed Spaces. Pergamon Press, 1964. Original edition: Funktsional'nyi analiz v normirovannikh prostranstvakh, Fizmatgiz (in Russian). [71] J.L. Kelley. General Topology. Springer, 1975. [72] J.F.C. Kingman. On queues in heavy traÆc. J. Roy. Statist. Soc., B24:383{392, 1962. [73] V.N. Kolokoltsov and V.P. Maslov. Idempotent Analysis and Its Applications. Kluwer, 1997. [74] A.P. Korostelev and S.L. Leonov. An action functional for a diusion process with discontinuous drift. Th. Prob. Appl., 37(3):543{550, 1992 (in Russian: Teor. Veroyatn. i Primen., 1992, v. 37, no. 3, pp. 570-576). [75] M.A. Krasnosel'skii and Ya.B. Rutickii. Convex Functions and Orlicz Spaces. Noordho, 1961. [76] K. Kuratowski and A. Mostowski. Set Theory. North-Holland{ PWN, 1967.

© 2001 by Chapman & Hall/CRC

Bibliography

489

[77] T. Lindvall. Weak convergence of probability measures and random functions in the function space D[0; 1). J. Appl. Prob., 10:109{121, 1973. [78] R.Sh. Liptser and A. Puhalskii. Limit theorems on large deviations for semimartingales. Stoch. Stoch. Rep., 38:201{249, 1992. [79] R.Sh. Liptser and A.N. Shiryaev. Theory of Martingales. Kluwer, 1989. [80] G.L. Litvinov, V.P. Maslov, and G.B. Shpiz. Idempotent functional analysis. An algebraic approach. Technical report, International Centre "Sophus Lie", 1998 (in Russian). [81] G.L. Litvinov, V.P. Maslov, and G.B. Shpiz. Linear functionals on idempotent spaces. An algebraic approach. Dokl. Akad. Nauk, 363(3):298{300, 1998 (in Russian). [82] J. Lynch and J. Sethuraman. Large deviations for processes with independent increments. Ann. Prob., 15(2):610{627, 1987. [83] S.Ya. Makhno. A large deviation theorem for a class of diusion processes. Teor. Veroyatnost. i Primenen., 39(3):554{566, 1994 (English translation: Th. Prob. Appl. 39(1994), no. 3, 437-447 (1995)). [84] V. Maslov. Methode Operatorielles. Mir, 1987 (in French). [85] V.P. Maslov. Asymptotic Methods of Solving PseudoDierential Equations. Nauka, 1987 (in Russian). [86] R. Mesiar. Possibility measures, integration and fuzzy possibility measures. Fuzzy Sets and Systems, 92:191{196, 1997. [87] R. Mesiar and E. Pap. Idempotent integral as limit of gintegrals. Fuzzy Sets and Systems, 102:385{392, 1999. [88] P.A. Meyer. Probability and Potentials. Blaisdell, 1966. [89] T. Micami. Some generalizations of Wentzell's lower estimates on large deviations. Stochastics, 24(4):269{284, 1988.

© 2001 by Chapman & Hall/CRC

490

Bibliography

[90] A.A. Mogulskii. Large deviations for trajectories of multidimensional random walks. Theory Prob. Appl., 21(2):300{315, 1976. [91] A.A. Mogulskii. Large deviations for processes with independent increments. Ann. Prob., 21(1):202{215, 1993. [92] U. Mosco. On the continuity of the Young-Fenchel transform. J. Math. Anal. Appl., 35(3):518{535, 1971. [93] K. Narita. Large deviation principle for diusion processes. Tsukuba J. Math., 12(1):211{229, 1988. [94] J. Neveu. Bases Mathematiques du Calcul des Probabilites. Masson et Cie, 1964 (in French). [95] T. Norberg. Random capacities and their distributions. Prob. Th. Rel. Fields, 73(2):281{297, 1986. [96] G.L. O'Brien. Sequences of capacities, with connections to large deviation theory. J. Theoret. Probab., 9(1):19{35, 1995. [97] G.L. O'Brien and W. Vervaat. Capacities, large deviations and loglog laws. In S. Cambanis, G. Samorodnitsky, and M. Taqqu, editors, Stable Processes and Related Topics, volume 25 of Progress in Probability, pages 43{83. Birkhauser, 1991. [98] G.L. O'Brien and W. Vervaat. Compactness in the theory of large deviations. Stoch. Proc. Appl., 57:1{10, 1995. [99] G.L. O'Brien and S. Watson. Relative compactness for capacities, measures, upper semicontinuous functions and closed sets. J. Theoret. Prob., 11(3):577{588, 1998. [100] B. ksendal. Stochastic Dierential Equations. Springer, 1998. [101] E. Pap. Null-Additive Set Functions. Kluwer, 1995. [102] K.R. Parthasarathy. Probability Measures on Metric Spaces. Academic Press, 1967. [103] V.V. Petrov. Limit Theorems for Sums of Independent Random Variables. Nauka, second edition, 1987 (in Russian).

© 2001 by Chapman & Hall/CRC

Bibliography

491

[104] Yu. V. Prohorov. Convergence of stochastic processes and limit theorems in probability theory. Th. Prob. Appl., 1:157{214, 1956. [105] Yu.V. Prohorov. Transient phenomena in queueing processes. Lit. Mat. Rink., 3:199{206, 1963 (in Russian). [106] A. Puhalskii. On functional principle of large deviations. In V. Sazonov and T. Shervashidze, editors, New Trends in Probability and Statistics, volume 1, pages 198{218. VSP/Moks'las, 1991. [107] A. Puhalskii. On the theory of large deviations. Th. Prob. Appl., 38:490{497, 1993. [108] A. Puhalskii. Large deviations of semimartingales via convergence of the predictable characteristics. Stoch. Stoch. Rep., 49:27{85, 1994. [109] A. Puhalskii. The method of stochastic exponentials for large deviations. Stoch. Proc. Appl., 54:45{70, 1994. [110] A. Puhalskii. Large deviation analysis of the single server queue. Queueing Systems, 21:5{66, 1995. [111] A. Puhalskii. Large deviations of semimartingales: a maxingale problem approach. I. Limits as solutions to a maxingale problem. Stoch. Stoch. Rep., 61:141{243, 1997. [112] A. Puhalskii. Large deviations of semimartingales: a maxingale problem approach. II. Uniqueness for the maxingale problem. Applications. Stoch. Stoch. Rep., 68:65{143, 1999. [113] A. Puhalskii and W. Whitt. Functional large deviation principles for rst-passage-time processes. Ann. Appl. Prob., 7(2):362{381, 1997. [114] A.A. Puhalskii. Moderate deviations for queues in critical loading. Queueing Systems, 31:359{392, 1999. [115] M.I. Reiman. Open queueing networks in heavy traÆc. Math. Oper. Res., 9:441{458, 1984.

© 2001 by Chapman & Hall/CRC

492

Bibliography

[116] B. Remillard and D.A. Dawson. Laws of the iterated logarithm and large deviations for a class of diusion processes. Can. J. Statist., 17(4):349{376, 1989. [117] R.T. Rockafellar. Convex Analysis. Princeton University Press, 1970. [118] L. Schwartz. Radon Measures on Arbitrary Topological Spaces and Cylindrical Measures. Oxford University Press, 1973. [119] N. Shilkret. Maxitive measure and integration. Indag. Math., 33:109{116, 1971. [120] A.N. Shiryaev. Probability, volume 95 of Graduate Texts in Mathematics. Springer, second edition, 1996 (Translated from the rst (1980) Russian edition by R. P. Boas). [121] A.V. Skorohod. Limit theorems for stochastic processes. Th. Prob. Appl., 1:261{292, 1956. [122] D.W. Stroock. An Introduction to the Theory of Large Deviations. Springer, 1984. [123] D.W. Stroock and S.R.S. Varadhan. Multidimensional Diusion Processes. Springer, 1979. [124] F. Topse. Compactness in spaces of measures. Studia Mathematica, 36:195{221, 1970. [125] F. Topse. Topology and Measure, volume 133 of Lecture Notes in Mathematics. Springer, 1970. [126] N.N. Vakhania, V.I. Tarieladze, and S.A. Chobanyan. Probability Distributions on Banach Spaces. Nauka, 1985 (in Russian, English translation: Reidel, 1987). [127] S.R.S. Varadhan. Asymptotic probabilities and dierential equations. Comm. Pure Appl. Math., 19(3):261{286, 1966. [128] S.R.S. Varadhan. Large Deviations and Applications. SIAM, 1984. [129] W. Vervaat. Narrow and vague convergence of set functions. Statist. & Prob. Lett., 6(5):295{298, 1988.

© 2001 by Chapman & Hall/CRC

Bibliography

493

[130] W. Vervaat. Random uppersemicontinuous functions and extremal processes. Technical Report MS-8801, Center for Math. and Comp. Sci., Amsterdam, 1988. [131] K. von Leichtweiss. Konvexe Mengen. VEB Deutscher Verlag der Wissenschaften, 1980. [132] P.-Z. Wang. Fuzzy contactability and fuzzy variables. Fuzzy Sets and Systems, 8:81{92, 1982. [133] Z. Wang and G.J. Klir. Fuzzy Measure Theory. Plenum Press, 1992. [134] A.D. Wentzell. Limit Theorems on Large Deviations for Markov Stochastic Processes. Nauka, 1986 (in Russian, English translation: Kluwer, 1990). [135] W. Whitt. Some useful functions for functional limit theorems. Math. Oper. Res., 5(1):67{85, 1980. [136] W. Whitt. On the heavy-traÆc limit theorem for GI=G=1 queues. Adv. Appl. Prob., 14:171{190, 1982. [137] C. Wu, S. Wang, and M. Ma. Generalized fuzzy integrals: Part I. Fundamental concepts. Fuzzy Sets and Systems, 57:219{226, 1993. [138] S.L. Zabell. Mosco convergence in locally convex spaces. J. Function. Anal., 110(1):226{246, 1992. [139] T. Zajic. Rough asymptotics for tandem non-homogeneous M=G=1 queues via Poissonized empirical processes. Queueing Systems, 29(2-4):161{174, 1998.

© 2001 by Chapman & Hall/CRC

Our partners will collect data and use cookies for ad personalization and measurement. Learn how we and our ad partner Google, collect and use data. Agree & close