No

Universitext
Paolo Baldi
Stochastic
Calculus
An Introduction Through Theory and
Exercises
Universitext
Universitext
Series editors
Sheldon Axler
San Francisco State University
Carles Casacuberta
Universitat de Barcelona
Angus MacIntyre
Queen Mary, University of London
Kenneth Ribet
University of California, Berkeley
Claude Sabbah
École polytechnique, CNRS, Université Paris-Saclay, Palaiseau
Endre Süli
University of Oxford
Wojbor A. Woyczyński
Case Western Reserve University
Universitext is a series of textbooks that presents material from a wide variety of

mathematical disciplines at master’s level and beyond. The books, often well class-
tested by their author, may have an informal, personal even experimental approach
to their subject matter. Some of the most successful and established books in the
series have evolved through several editions, always following the evolution of
teaching curricula, to very polished texts.
Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.
More information about this series at http://www.springer.com/series/223

Paolo Baldi
Stochastic Calculus
An Introduction Through Theory
and Exercises
123
Paolo Baldi
Dipartimento di Matematica
UniversitJa di Roma “Tor Vergata”
Roma, Italy
ISSN 0172-5939 ISSN 2191-6675 (electronic)

Universitext
ISBN 978-3-319-62225-5 ISBN 978-3-319-62226-2 (eBook)
DOI 10.1007/978-3-319-62226-2
Library of Congress Control Number: 2017948389
Mathematics Subject Classification (2010): 60H10, 60H05, 60H30, 60G42, 60G44
© Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To the memory of Gianni Di Masi
Preface
Courses in Stochastic Calculus have in the last two decades changed their target
audience. Once this advanced part of mathematics was of interest mainly to
postgraduates intending to pursue an academic research career, but now many
professionals cannot do without the ability to manipulate stochastic models.
The aim of this book is to provide a tool in this direction, starting from a basic
probability background (with measure theory, however). The intended audience
should, moreover, have serious mathematical bases.
The entire content of this book should provide material for a two-semester class.
My experience is that Chaps. 2–9 provide the material for a course of 72 h, including
the time devoted to the exercises.
To be able to manipulate these notions requires the reader to acquire not only the
elements of the theory but also the ability to work with them and to understand their
applications and their connections with other branches of mathematics.
The first of these objectives is taken care of (or at least I have tried to. . . ) by the
development of a large set of exercises which are provided together with extensive
solutions. Exercises are hardwired with the theory and are intended to acquaint the
reader with the full meaning of the theoretical results. This set of exercises with
their solution is possibly the most original part of this work.
As for the applications, this book develops two kinds.
The first is given by modeling applications. Actually there are very many
situations (in finance, telecommunications, control, . . . ) where stochastic processes,
and in particular diffusions, are a natural model. In Chap. 13 we develop financial
applications, currently a rapidly growing area.
Stochastic processes are also connected with other fields in pure mathematics
and in particular with PDEs. Knowledge of diffusion processes contributes to a
better understanding of some aspects of PDE problems and, conversely, the solution
of PDE problems can lead to the computation of quantities of interest related to
diffusion processes. This two-way tight connection between processes and PDEs is
developed in Chap. 10. Further interesting connections between diffusion processes
vii
viii Preface
and other branches of mathematics (algebraic structures, differential geometry, . . . )

are unfortunately not present in this text, primarily for reasons of space.
The first goal is to make the reader familiar with the basic elements of stochastic
processes, such as Brownian motion, martingales, and Markov processes, so that it
is not surprising that stochastic calculus proper begins almost in the middle of the
book.
Chapters 2–3 introduce stochastic processes. After the description of the general
setting of a continuous time stochastic process that is given in Chap. 2, Chap. 3
introduces the prototype of diffusion processes, that is Brownian motion, and
investigates its, sometimes surprising, properties.
Chapters 4 and 5 provide the main elements on conditioning, martingales, and
their applications in the investigation of stochastic processes. Chapter 6 is about
Markov processes.
From Chap. 7 begins stochastic calculus proper. Chapters 7 and 8 are concerned
with stochastic integrals and Ito’s formula. Chapter 9 investigates stochastic dif-
ferential equations, Chap. 10 is about the relationship with PDEs. After the detour
on numerical issues related to diffusion processes of Chap. 11, further notions of
stochastic calculus are investigated in Chap. 12 (Girsanov’s theorem, representation
theorems of martingales) and applications to finance are the object of the last
chapter.
The book is organized in a linear way, almost every section being necessary
for the understanding of the material that follows. The few sections and the single
chapter that can be avoided are marked with an asterisk.
This book is based on courses that I gave first at the University of Pisa, then
at Roma “Tor Vergata” and also at SMI (Scuola Matematica Interuniversitaria) in
Perugia. It has taken advantage of the remarks and suggestions of many cohorts of
students and of colleagues who tried the preliminary notes in other universities.
The list of the people I am indebted to is a long one, starting with the many
students that have suffered under the first versions of this book. G. Letta was very
helpful in clarifying to me quite a few complicated situations. I am also indebted
to C. Costantini, G. Nappo, M. Pratelli, B. Trivellato, and G. Di Masi for useful
remarks on the earlier versions.
I am also grateful for the list of misprints, inconsistencies, and plain mistakes
pointed out to me by M. Gregoratti and G. Guatteri at Milano Politecnico and
B. Pacchiarotti at my University of Roma “Tor Vergata”. And mainly I must mention
L. Caramellino, whose class notes on mathematical finance were the main source of
Chap. 13.
Roma, Italy Paolo Baldi

Contents
1 Elements of Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.1 Probability spaces, random variables .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
1.2 Variance, covariance, law of a r.v. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.3 Independence, product measure . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
1.4 Probabilities on Rm .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 9
1.5 Convergence of probabilities and random variables .. . . . . . . . . . . . . . . 11
1.6 Characteristic functions . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
1.7 Gaussian laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
1.8 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
1.9 Measure-theoretic arguments . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 25
2 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 31
2.1 General facts .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 31
2.2 Kolmogorov’s continuity theorem .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
2.3 Construction of stochastic processes . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 38
2.4 Next. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 42
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 42
3 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
3.1 Definition and general facts . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
3.2 The law of a continuous process, Wiener measure . . . . . . . . . . . . . . . . . 52
3.3 Regularity of the paths . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 53
3.4 Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 57
3.5 Stopping times .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
3.6 The stopping theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 66
3.7 The simulation of Brownian motion.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 70
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 75
4 Conditional Probability .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 85
4.1 Conditioning .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 85
4.2 Conditional expectations .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
ix
x Contents
4.3 Conditional laws. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98

4.4 Conditional laws of Gaussian vectors . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 100
4.5 The augmented Brownian filtration . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 102
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 104
5 Martingales .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
5.1 Definitions and general facts. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
5.2 Discrete time martingales .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 111
5.3 Discrete time martingales: a.s. convergence . . . .. . . . . . . . . . . . . . . . . . . . 114
5.4 Doob’s inequality; Lp convergence, the p > 1 case .. . . . . . . . . . . . . . . . 118
5.5 Uniform integrability and convergence in L1 . . .. . . . . . . . . . . . . . . . . . . . 121
5.6 Continuous time martingales . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 124
5.7 Complements: the Laplace transform . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 133
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
6 Markov Processes .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
6.1 Definitions and general facts. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151
6.2 The Feller and strong Markov properties . . . . . . .. . . . . . . . . . . . . . . . . . . . 160
6.3 Semigroups, generators, diffusions .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 175
7 The Stochastic Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
7.1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
7.2 Elementary processes . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 182
7.3 The stochastic integral . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 185
7.4 The martingale property . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 192
2
7.5 The stochastic integral in Mloc . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 200
7.6 Local martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 205
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
8 Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
8.1 Ito’s formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 215
8.2 Application: exponential martingales . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 224
8.3 Application: Lp estimates . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 229
8.4 The multidimensional stochastic integral . . . . . . .. . . . . . . . . . . . . . . . . . . . 231
8.5 A case study: recurrence of multidimensional
Brownian motion .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 243
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 246
9 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
9.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
9.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 257
9.3 An a priori estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 260
9.4 Existence for Lipschitz continuous coefficients . . . . . . . . . . . . . . . . . . . . 265
9.5 Localization and existence for locally Lipschitz coefficients . . . . . . 270
9.6 Uniqueness in law .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 273
9.7 The Markov property.. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 275
Contents xi
9.8 Lp bounds and dependence on the initial data. . .. . . . . . . . . . . . . . . . . . . . 282

9.9 The square root of a matrix field and the problem
of diffusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 285
9.10 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 290
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 291
10 PDE Problems and Diffusions . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 305
10.1 Representation of the solutions of a PDE problem . . . . . . . . . . . . . . . . . 305
10.2 The Dirichlet problem.. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 307
10.3 Parabolic equations .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 314
10.4 The Feynman–Kac formula .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 317
10.5 The density of the transition function, the backward
equation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 326
10.6 Construction of the solutions of the Dirichlet problem .. . . . . . . . . . . . 328
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 334

11 Simulation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 341
11.1 Numerical approximations of an SDE. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 341
11.2 Strong approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343
11.3 Weak approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 349
11.4 Simulation and PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 350
11.5 Other schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 359
11.6 Practical remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 361
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 363
12 Back to Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 365
12.1 Girsanov’s theorem .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 365
12.2 The Cameron–Martin formula .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 379
12.3 The martingales of the Brownian filtration .. . . . .. . . . . . . . . . . . . . . . . . . . 381
12.4 Equivalent probability measures .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 387
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 389
13 An Application: Finance .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 395
13.1 Stocks and options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 395
13.2 Trading strategies and arbitrage . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 397
13.3 Equivalent martingale measures . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 402
13.4 Replicating strategies and market completeness .. . . . . . . . . . . . . . . . . . . 404
13.5 The generalized Black–Scholes models .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 407
13.6 Pricing and hedging in the generalized Black–Scholes model .. . . . 417
13.7 The standard Black–Scholes model . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 422
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 434
Solutions of the Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 437
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 623
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 625
Common Notations
Real, complex numbers, Rm
x_y D max.x; y/ the largest of the real numbers x and y

x^y D min.x; y/ the smallest of the real numbers x and y
hx; yi The scalar product of x; y 2 Rm
xC , x The positive and negative parts of x 2 R: xC D max.x; 0/, x D
max.x; 0/
jxj According to the context the absolute value of the real number x,
the modulus of the complex number x, or the euclidean norm of
x 2 Rm
<z, =z The real and imaginary parts of z 2 C
BR .x/ fy 2 Rm ; jy xj < Rg, the open ball centered at x with radius R
A , tr A, det A The transpose, trace, determinant of matrix A
Derivatives
@f
, fx Partial derivatives
@xi i
f0 Derivative, gradient
f 00 Second derivative, Hessian
Functional spaces
Mb .E/ Real bounded measurable functions on the topological space E

kf k1 D supx2E jf .x/j if f 2 Mb .E/
xiii
xiv Common Notations
Cb .Rm / The Banach space of real bounded continuous functions on Rm

endowed with the norm k k1
CK .Rm / The subspace of Cb .Rm / of the continuous functions with com-
pact support
Ck .D/ k times differentiable functions on the open set D Rm
C2;1 .D Œ0; TŒ/ Functions twice differentiable in x 2 D and once in t 2 Œ0; TŒ
C .Œ0; T; Rm / The vector space of continuous maps W Œ0; T ! Rm endowed
with the sup norm k k1 . It is a Banach space
C .RC ; Rm / The vector space of continuous maps W RC ! Rm , endowed
with the topology of uniform convergence on compact sets of
RC
Cx .Œ0; T; Rm / The vector space of continuous maps W Œ0; T ! Rm such that
0 D x, endowed with the topology of uniform convergence
Chapter 1
Elements of Probability
In this chapter we recall the basic facts in probability that are required for the
investigation of the stochastic processes that are the object of the subsequent
chapters.
1.1 Probability spaces, random variables
A measurable space is a pair .E; E / where

• E is a set;
• E is a -algebra of subsets of E.
A probability space is a triple .˝; F ; P/ where
• ˝ is a set;
• F is a -algebra of subsets of ˝ (i.e. (˝; F / is a measurable space);
• P is a positive measure on F such that P.˝/ D 1.
The elements of F are called events, P a probability (measure).
If .˝; F ; P/ is a probability space and .E; E / a measurable space, a random
variable (r.v.) is a measurable function X W ˝ ! E, i.e. such that X 1 .A/ 2 F
whenever A 2 E . It is a real random variable if E D R and E D B.R/ (Borel sets
of R).
For a real r.v. X on .˝; F ; P/ weRcan speak of an integral. If X is integrable we
shall usually write EŒX instead of X dP. EŒX is the mathematical expectation.
Sometimes the terms mean or mean value are also used.
If X D .X1 ; : : : ; Xm / is .Rm ; B.Rm //-valued we shall speak of m-dimensional
r.v.’s and, if the components Xi are integrable, we define
EŒX D .EŒX1 ; : : : ; EŒXm / :
We say that two r.v.’s X; Y are equivalent if P.X D Y/ D 1.
© Springer International Publishing AG 2017 1

P. Baldi, Stochastic Calculus, Universitext, DOI 10.1007/978-3-319-62226-2_1
2 1 Elements of Probability
The spaces Lp , 1 p < C1, and L1 are defined as usual as well as the norms
kXkp and kXk1 . Recall in particular that Lp is the set of equivalence classes of r.v.’s
X such that kXkp D EŒjXjp 1=p < C1. It is worth pointing out that Lp is a set of
equivalence classes and not of r.v.’s; this distinction will sometimes be necessary
even if often, in order to simplify the statements, we shall identify a r.v. and its
equivalence class.
For a real r.v. X, let us denote by X D X C X its decomposition into positive
and negative parts. Recall that both X C D X _ 0 and X D .X/ _ 0 are positive
r.v.’s. X is said to be lower semi-integrable (l.s.i.) if X is integrable. In this case it
is possible to define the mathematical expectation
EŒX D EŒX C EŒX ;
which is well defined, even if it can take the value C1. Of course a positive r.v. is
always l.s.i.
The following classical inequalities hold.
Jensen’s inequality. Let X be an m-dimensional integrable r.v. and ˚ W Rm !

R [ fC1g a lower semi-continuous convex function; then ˚.X/ is lower semi-
integrable and
EŒ˚.X/ ˚.EŒX/ (1.1)
(possibly one or both sides in the previous inequality can be equal to C1).
Moreover, if ˚ is strictly convex and ˚.EŒX/ < C1, then the inequality is
strict unless X takes only one value a.s.
Hölder’s inequality. Let Z, W be real positive r.v.’s and ˛, ˇ positive real

numbers such that ˛ C ˇ D 1. Then
EŒZ ˛ W ˇ EŒZ˛ EŒWˇ : (1.2)
From (1.2) it follows that if X, Y are real r.v.’s and p, q positive numbers such that
1
p
C 1q D 1, then, by setting ˛ D 1p , ˇ D 1q , Z D jXjp , W D jYjq , we have
ˇ ˇ
ˇEŒXYˇ EŒjXjp 1=p EŒjYjq 1=q ; (1.3)
which also goes under the name of Hölder’s inequality.

1.2 Variance, covariance, law of a r.v. 3
Minkowski’s inequality. Let X, Y be real r.v.’s and p 1, then
kX C Ykp kXkp C kYkp : (1.4)
Minkowski’s inequality, in particular, implies that Lp , p 1, is a vector space. If

p > q, Jensen’s inequality applied to the continuous convex function ˚.x/ D jxjp=q
gives
kXkpp D EŒjXjp D EŒ˚.jXjq / ˚.EŒjXjq / D EŒjXjq p=q
and therefore, taking the p-th root,
kXkp kXkq : (1.5)
In particular, if p q, Lp Lq .
1.2 Variance, covariance, law of a r.v.
Let X be a real square integrable r.v. (i.e. such that EŒX 2 < C1). Its variance is
the quantity
Var.X/ D EŒ.X EŒX/2 D EŒX 2 EŒX2 :

R
If ˛ > 0, the quantity jxj˛ .dx/
R is the absolute moment of order ˛ of . If ˛ is
a positive integer the quantity x˛ .dx/, if it exists and is finite, is the moment of
order ˛ of .
Markov’s inequality. For every ı > 0; ˇ > 0 we have
EŒjXjˇ
P.jXj ı/ (1.6)
ıˇ
It is apparent from its definition that the variance of a r.v. is so much larger as X takes
values far from its mean EŒX. This intuitive fact is made precise by the following
Chebyshev’s inequality. Let X 2 L2 . Then for every ˛ > 0
Var.X/
P.jX EŒXj ˛/
˛2
This is a very important inequality. It is, of course, a particular case of Markov’s

inequality.
The covariance of two r.v.’s X and Y is defined as
Cov.X; Y/ D EŒXY EŒXEŒY :
If Cov.X; Y/ D 0, X and Y are said to be uncorrelated.

Let X be a r.v. with values in the measurable space .E; E /. It is easy to see that the
set function X defined on E as
X .A/ D P.X 1 .A//
is itself a probability measure (on E ). X is the law of X. It is the image (or pullback)
of P through X. The following proposition provides a formula for the computation
of integrals with respect to an image law. We shall make use of it throughout.
Proposition 1.1 Let X W .˝; F ; P/ ! .E; E / be a r.v., X the law of X. Then

a measurable function f W .E; E / ! .R; B.R// is X -integrable if and only if
f .X/ is P-integrable, and then we have
Z Z
f .x/ X .dx/ D f .X.!// P.d!/ : (1.7)
E ˝
In particular, if X is a real r.v. then X has finite mathematical expectation if and

only if x 7! x is X -integrable, and in this case
Z
EŒX D x X .dx/ :
Also we have, for every ˛,

Z
˛
EŒjXj D jxj˛ X .dx/ :
Therefore X 2 Lp if and only if its law has a finite absolute moment of order p.
By the notations X , X Y we shall mean “X has law ” and “X and Y have
the same law”, respectively.
Given two probabilities P and Q on .˝; F /, we say that Q is absolutely continuous
with respect to P if P.A/ D 0 implies Q.A/ D 0. This relation is denoted P Q.
The Radon–Nikodym Theorem states that if P Q then there exists a r.v. Z 0
1.3 Independence, product measure 5
such that
Z
Q.A/ D Z dP
A
(the converse is obvious). P and Q are said to be equivalent if both P Q and

Q P. In other words, two probabilities are equivalent if and only if they have
the same negligible events. Recall that an event N 2 F is negligible if P.N/ D 0.
Conversely, P and Q are said to be orthogonal if there exists an event A such that
P.A/ D 1 and Q.A/ D 0. If P and Q are orthogonal, of course, it is not possible for
one of them to be absolutely continuous with respect to the other one.
1.3 Independence, product measure
If E , E 0 are -algebras of events of E let us denote by E _ E 0 the smallest -

algebra containing both E and E 0 . This is well defined: it is immediate that the
intersection of any non-empty family of -algebras is again a -algebra. We can
therefore consider the family of all -algebras containing both E and E 0 . This is
non-empty as certainly P.E/ (all subsets of E) is one of them. Then E _ E 0 will be
the intersection of this family of -algebras. This argument will be used from now
on to define -algebras as “the smallest -algebra such that. . . ”.
Let X W .˝; F ; P/ ! .E; E / be a r.v. and denote by .X/ the -algebra
generated by X, i.e. the smallest sub--algebra of F with respect to which X is
measurable. It is easy to see that .X/ is the family of the events of F of the form
X 1 .A0 / with A0 2 E .
The following lemma, characterizing the real .X/-measurable r.v.’s, is very
useful.
Lemma 1.1 (Doob’s measurability criterion) Let X W .˝; F ; P/ !

.E; E / be a r.v. Then every real .X/-measurable r.v. is of the form f .X/,
where f W .E; E / ! .R; B.R// is a measurable map.
The r.v.’s X1 ; : : : ; Xm , taking values respectively in .E1 ; E1 /; : : : ; .Em ; Em /, are

said to be independent if, for every A01 2 E1 ; : : : ; A0m 2 Em ,
P.X1 2 A01 ; : : : ; Xm 2 A0m / D P.X1 2 A01 / : : : P.Xm 2 A0m / :
The events A1 ; : : : ; Am 2 F are said to be independent if and only if
P.Ai1 \ \ Ai` / D P.Ai1 / : : : P.Ai` /

for every choice of 1 ` m and of 1 i1 < i2 < < i` m. A

careful but straightforward computation shows that this definition is equivalent to
the independence of the r.v.’s 1A1 ; : : : ; 1Am .
If F1 ; : : : ; Fm are sub--algebras of F , we say that they are independent if, for
every A1 2 F1 ; : : : ; Am 2 Fm ,
P.A1 \ \ Am / D P.A1 / : : : P.Am / :
It is immediate that the r.v.’s X1 ; : : : ; Xm are independent if and only if so are the
generated -algebras .X1 /; : : : ; .Xm /.
Let .Xi /i2I be a (possibly infinite) family of r.v.’s; they are said to be independent
if and only if the r.v.’s Xi1 ; : : : ; Xim are independent for every choice of a finite
number i1 ; : : : ; im of distinct indices in I. A similar definition is made for an infinite
family of -algebras.
We shall say, finally, that the r.v. X is independent of the -algebra G if and only
if the -algebras .X/ and G are independent. It is easy to see that this happens if
and only if X is independent of every G -measurable r.v. W.
Let us now point out the relation between independence and laws of r.v.’s. If we
denote by i the law of Xi and we define E D E1 Em , E D E1 ˝ ˝ Em , on
the product space .E; E /, we can consider the product measure D 1 ˝ ˝ m .
The following result will be of constant use in the sequel.
Proposition 1.2 In the previous notations the r.v.’s X1 ; : : : ; Xm are indepen-

dent if and only if the law of X D .X1 ; : : : ; Xm / on .E; E / is the product law .
The proof of this proposition makes use of the following theorem.
Theorem 1.1 (Carathéodory’s criterion) Let .E; E / be a measurable

space, 1 , 2 two finite measures on .E; E /; let I E be a family of
subsets of E stable with respect to finite intersections and generating E . Then,
if 1 .E/ D 2 .E/ and 1 and 2 coincide on I , they also coincide on E .
Proof of Proposition 1.2. Let us denote by the law of X and let I be the family
of the sets of E of the form A1 Am with Ai 2 Ei , i D 1; : : : ; m. I
is stable with respect to finite intersections and, by definition, generates E . The
definition of independence states exactly that and coincide on I ; therefore by
Carathéodory’s criterion they coincide on E . t
u
1.3 Independence, product measure 7
Proposition 1.3 If X and Y are real independent integrable r.v.’s, then also
their product is integrable and
EŒXY D EŒXEŒY :
Recalling that Cov.X; Y/ D EŒXY EŒXEŒY, in particular we have that real

integrable and independent r.v.’s are not correlated. Examples show that the converse
is not true.
Remark 1.1 A problem we shall often be confronted with is to prove that

some -algebras are independent. Let us see how, in practice, to perform
this task.
Let F1 and F2 be sub--algebras of F and let C1 F1 , C2 F2
be subclasses generating F1 and F2 , respectively, and stable with respect to
finite intersections. Then F1 and F2 are independent if and only if
P.A1 \ A2 / D P.A1 /P.A2 / (1.8)
for every A1 2 C1 , A2 2 C2 .
Actually, if A2 2 C2 is fixed, the two measures on F1
A1 7! P.A1 \ A2 / and A1 7! P.A1 /P.A2 /
are finite and coincide on C1 ; they also have the same total mass (D P.A2 /).
By Carathéodory’s criterion, Theorem 1.1, they coincide on F1 , hence (1.8)
holds for every A1 2 F1 , A2 2 C2 . By a repetition of this argument with
A1 2 F1 fixed, (1.8) holds for every A1 2 F1 , A2 2 F2 , i.e. F1 and F2 are
independent.
A case of particular interest appears when F1 D .X; X 2 I /, F2 D
.Y; Y 2 J / are -algebras generated by families I and J of r.v.’s
respectively. If we assume that every X 2 I is independent of every Y 2 J ,
is this enough to guarantee the independence of F1 and F2 ?
Thinking about this a bit it is clear that the answer is negative, as even
in elementary classes in probability one deals with examples of r.v.’s that
are independent pairwise but not globally. It might therefore happen that X1
and Y are independent, as well as X2 and Y, whereas the pair .X1 ; X2 / is
not independent of Y. fX1 ; X2 g D I and fYg D J therefore provides a
counterexample.
If, however, we assume that for every choice of X1 ; : : : ; Xn 2 I and
Y1 ; : : : ; Yk 2 J the r.v.’s .X1 ; : : : ; Xn / and .Y1 ; : : : ; Yk / are independent, then
necessarily the two generated -algebras F1 and F2 are independent.
(continued )
Remark 1.1 (continued)

Let us denote by .Ei ; Ei / and .Gj ; Gj / the measurable spaces where the
r.v.’s Xi ; Yj take the their values respectively. Let us consider the class C1 of
the events of the form
fX1 2 A01 ; : : : ; Xn 2 A0n g
for some n 1; X1 ; : : : ; Xn 2 I ; A01 2 E1 ; : : : ; A0n 2 En and similarly define

C2 as the class of the events of the form
fY1 2 B01 ; : : : ; Yk 2 B0k g
for some k 1; Y1 ; : : : ; Yk 2 J ; B01 2 G1 ; : : : ; B0k 2 Gk . One verifies that C1

and C2 are stable with respect to finite intersections and generate .X; X 2
I / and .Y; Y 2 J /, respectively. Moreover, (1.8) is satisfied for every
A1 2 C1 ; A2 2 C2 . Actually, as the r.v.’s .X1 ; : : : ; Xn / and .Y1 ; : : : ; Yk / are
assumed to be independent, the two events
A1 D fX1 2 A01 ; : : : ; Xn 2 A0n g D f.X1 ; : : : ; Xn / 2 A01 A0n g

A2 D fY1 2 B01 ; : : : ; Yk 2 B0k g D f.Y1 ; : : : ; Yk / 2 B01 B0k g
are independent. Therefore, by the previous criterion, F1 and F2 are

independent. In other words, F1 and F2 are independent if and only if the
r.v.’s .X1 ; : : : ; Xn / and .Y1 ; : : : ; Yk / are independent for every choice of n; k
and of X1 ; : : : ; Xn 2 I and Y1 ; : : : ; Yk 2 J .
We shall see later that if the r.v.’s of I and J form together a Gaussian
family, then this criterion can be considerably simplified.
The following result will be of constant use in the sequel.
Theorem 1.2 (Fubini) Let 1 ; 2 be measures respectively on the measur-

able spaces .E1 ; E1 /, .E2 ; E2 / and let f W E1 E2 ! R be an E1 ˝E2 -measurable
function such that at least one of the following is true:
1) f is integrable with respect to 1 ˝ 2
2) f is positive.
Then
a) the functions
Z Z
x1 7! f .x1 ; z/ 2 .z/; x2 7! f .z; x2 / 1 .dz/ (1.9)
are respectively E1 -measurable and E2 -measurable.
(continued)
1.4 Probabilities on Rm 9
Theorem 1.2 (continued)

b)
Z Z Z
f d1 ˝ 2 D 1 .dx1 / f .x1 ; x2 / 2 .dx2 /
E1 E2
Z Z
D 2 .dx2 / f .x1 ; x2 / 1 .dx1 / :
1.4 Probabilities on Rm
Let be a probability on .Rm ; B.Rm // and let us denote by i W Rm ! R its

projection on the i-th coordinate: we call the image of through i the i-th marginal
law of . By the integration rule with respect to an image law, (1.7), the i-th marginal
law is therefore given by
Z
i .A/ D 1A .xi / .dx1 ; : : : ; dxm /; A 2 B.R/ : (1.10)
Rm
If X D .X1 ; : : : ; Xm / is an m-dimensional r.v. with law it is clear that its i-th

marginal i coincides with the law of Xi .
We say that a probability on Rm admits a density (with respect to Lebesgue
measure) if there exists a Borel function f 0 such that for every A 2 B.Rm /
Z
.A/ D f .x/ dx :
A
If admits a density f then its i-th marginal i also admits a density fi , given by
Z
fi .x/ D f .y1 ; : : : ; yi1 ; x; yiC1 ; : : : ; ym / dy1 : : : dyi1 dyiC1 : : : dym
(the existence of the integral and the fact that such an fi is a measurable function are
consequences of Fubini’s Theorem 1.2, see (1.9)).
In any case the previous formulas show that, given , it is possible to determine
its marginal distributions. The converse is not true: it is not possible, in general,
knowing just the laws of the Xi ’s, to deduce the law of X. Unless, of course, the r.v.’s
Xi are independent as, in this case, by Proposition 1.2 the law of X is the product of
its marginals.
Let X be an Rm -valued r.v. with density f , a 2 Rm and A an m m invertible matrix;
it is easy to see that the r.v. Y D AX C a has also a density g given by
g.y/ D j det Aj1 f .A1 .y a// : (1.11)

If X D .X1 ; : : : ; Xm / is an m-dimensional r.v., its covariance matrix is the matrix

C D .cij /i;j defined as
cij D Cov.Xi ; Xj / D EŒ.Xi EŒXi /.Xj EŒXj / D EŒXi Xj EŒXi EŒXj :
It is the matrix having on the diagonal the variances of the components X1 ; : : : ; Xm

and outside the diagonal their covariances. Another more compact way of express-
ing the covariance matrix is the following. Let us assume X centered, then
C D EŒXX ; (1.12)
where we consider X as a column vector; X ( denotes the transpose) is a row vector

and the result of the product XX is an m m matrix whose i; j-th entry is precisely
Xi Xj . If, conversely, X is not centered, we have C D EŒ.X E.X//.X EŒX/ .
This formulation allows us, for instance, to derive immediately how the covariance
matrix transforms with respect to linear transformations: if A is a k m matrix and
Y D AX, then Y is a k-dimensional r.v. and (assume X centered for simplicity)
CY D EŒAX.AX/ D EŒAXX A D AEŒXX A D ACA : (1.13)
If X D .X1 ; : : : ; Xm / is an m-dimensional r.v., the covariance matrix provides

information concerning the correlation among the r.v.’s Xi ; i D 1; : : : ; m. In
particular, if these are pairwise independent the correlation matrix is diagonal. The
converse is not true in general.
An important property of every covariance matrix C is that it is positive

definite, i.e. that for every vector 2 Rm
X
m
hC; i D cij i j 0 : (1.14)
i;jD1
Assuming for simplicity that X is centered, we have cij D EŒXi Xj and
X
m hX
m i h X
m 2 i
hC; i D cij i j D E Xi Xj i j D E Xi i 0; (1.15)
i;jD1 i;jD1 iD1
as the r.v. inside the rightmost expectation is positive.

Let now X and Y be independent m-dimensional r.v.’s with laws and
respectively. We call the law of X C Y, denoted by
, the convolution product of
and . This definition actually only depends on and and not on X and Y:

is actually the image on Rm through the map .x; y/ 7! x C y of the law of .X; Y/ on
Rm Rm that, as we have seen, is the product law ˝ and does not depend on the
particular choice of X and Y.
1.5 Convergence of probabilities and random variables 11
Proposition 1.4 If and are probabilities on Rm having density f and g

respectively with respect to Lebesgue measure, then
also has density h
with respect to Lebesgue measure given by
Z
h.x/ D f .z/g.x z/ dz :
Rm
Proof If A 2 B.Rm /, thanks to the theorem of integration with respect to an image

law, Proposition 1.1,
Z Z

.A/ D 1A .z/
.dz/ D 1A .x C y/ .dx/ .dy/
Z Z Z
D 1A .x C y/f .x/g.y/ dx dy D dx f .z/g.x z/ dz ;
A Rm
which allows us to conclude the proof.

t
u
1.5 Convergence of probabilities and random variables
In this section .E; B.E// denotes a measurable space formed by a topological space
E with its Borel -algebra B.E/.
• Let .n /n be a sequence of finite measures on .E; B.E//. We say that it converges
to weakly if for every continuous bounded function f W E ! R
Z Z
lim fdn D fd : (1.16)
n!1
Note that if n !n!1 weakly, in general we do not have n .A/ !n!1 .A/
for A 2 E , as the indicator function 1A is not, in general, continuous. It can be
proved, however, that n .A/ !n!1 .A/ if .@A/ D 0 and that (1.16) also
holds for functions f such that the set of their points of discontinuity is negligible
with respect to the limit measure .
Let Xn , n 2 N, and X be r.v.’s on .˝; F ; P/ taking values in .E; B.E//.
a:s:
• We say that .Xn /n converges to X almost surely (a.s.), denoted Xn ! X, if there
exists a negligible event N 2 F such that
lim Xn .!/ D X.!/ for every ! … N :

n!1
Lp
• If E D Rm we say that .Xn /n converges to X in Lp , denoted Xn ! X, if X 2 Lp
and
lim EŒjXn Xjp D 0 :

n!1
Thanks to (1.5), Lp -convergence implies Lq -convergence if p > q.

• If the sequence .Xn /n takes its values in a metric space E with a distance denoted
P
by d, we say that it converges to X in probability, denoted Xn ! X, if, for every
ı > 0,
lim P.d.Xn ; X/ > ı/ D 0 :

n!1
L
• We say that .Xn /n converges to X in law, denoted Xn ! X, if n ! weakly,
n , denoting respectively the laws of Xn and X. Note that for this kind of
convergence it is not necessary for the r.v.’s X; Xn ; n D 1; 2; : : : , to be defined on
the same probability space.
The following proposition summarizes the general comparison results between
these notions of convergence.
Lp P a:s: P
Proposition 1.5 If Xn ! X then Xn ! X. If Xn ! X then Xn ! X. If
P L P
Xn ! X then Xn ! X. If Xn ! X then there exists a subsequence .Xnk /k
converging to X a.s.
In particular, the last statement of Proposition 1.5 implies the uniqueness of the
limit in probability.
Proposition 1.6 (Cauchy criterion of convergence in probability) Let

.Xn /n be a sequence of r.v.’s with values in the complete metric space E and
let us assume that for every > 0, " > 0 there exists an n0 2 N such that for
every n; m > n0
P.d.Xn ; Xm / > / < " :
P
Then there exists an E-valued r.v. X such that Xn ! X as n ! 1.
1.6 Characteristic functions 13
Let us consider on the probability space .˝; F ; P/ a sequence .An /n of events.

Let us define an event A (the superior limit of .An /n ) through
1 [
\
lim An D Ak D A :
n!1
nD1 kn
It is easy to see that, equivalently,
1A D lim 1An
n!1
or
A D f!I ! 2 An for infinitely many indices ng :
We shall often be led to the computation of the probability of the superior limit of a
sequence of events. To this end, the key argument is the following.
Proposition 1.7 (Borel–Cantelli lemma) Let .An /n , and A D limn!1 An

as above. Then
P
a) if P1nD1 P.An / < C1, then P.A/ D 0.
1
b) If nD1 P.An / D C1 and, moreover, the events .An /n are pairwise
independent, then P.A/ D 1.
Usually Proposition 1.7 b) is stated under assumption of global independence of

the events .An /n . Actually, pairwise independence is sufficient.
1.6 Characteristic functions
If X is an m-dimensional r.v. and denotes its law, let us define

Z
b
.
/ D eih
;xi .dx/ D EŒeih
;Xi ;
2 Rm :
b
is the characteristic function (very much similar to the Fourier transform) of . It
is defined for every probability on Rm and enjoys the following properties, some
of them being immediate.
1. b
.0/ D 1 and jb.
/j 1, for every
2 Rm .
2. If X and Y are independent r.v.’s with laws and respectively, then we have
b
XCY .
/ D b.
/b .
/.
3. b
is uniformly continuous.
4. If has finite mathematical expectation then b
is differentiable and
Z
@b

.
/ D i xj eih
;xi .dx/ :
@
j
In particular,
Z
@b

.0/ D i xj .dx/ ;
@
j
0 .0/ D iEŒX.
i.e. b
5. If has finite moment of order 2, b
is twice differentiable and
Z
@2 b

.
/ D xk xj eih
;xi .dx/ : (1.17)
@
k @
j
In particular,
Z
@2b
.0/ D xk xj .dx/ : (1.18)
@
k @
j
For m D 1 this relation becomes b 00 .0/ D EŒX 2 . If X is centered, then

the second derivative at 0 is equal to 1 times the variance. If m 2 and X
is centered, then (1.18) states that the Hessian of b at the origin is equal to 1
times the covariance matrix.
Similar statements hold for higher-order moments: if ˛ D .i1 ; : : : ; im / is a
multi-index and j˛j D i1 C C im and if the absolute moment of order j˛j,
EŒjXjj˛j , is finite, then b
is ˛ times differentiable and
Z
@j˛j b

˛
.
/ D ij˛j x˛ d.x/ : (1.19)
@x
These results of differentiability are an immediate consequence of the theorem

of differentiation of integrals depending on a parameter.
Conversely it is not true, in general, that if all the derivatives of order j˛j of
b
are finite, then the moment of order j˛j is finite. It can be shown, however,
that if j˛j is an even number and b is differentiable up to the order j˛j at 0,
then X has a finite moment of order j˛j and therefore (1.19) holds. Thus, in
particular, in dimension 1, if b is twice differentiable at the origin, then X has
finite variance and (1.17) and (1.18) hold.
The previous formulas are very useful as they allow us to obtain the mean,
variance, covariance, moments. . . of a r.v. simply by computing the derivatives
of its characteristic function at 0.
1.7 Gaussian laws 15
6. If b
.
/ D b .
/ for every
2 Rm , then D . This very important property
explains the name “characteristic function”.
7. If X1 ; : : : ; Xm are r.v.’s respectively 1 ; : : : ; n -distributed, then they are inde-
pendent if and only if, denoting by the law of X D .X1 ; : : : ; Xm /,
b
.
1 ; : : : ;
m / D b
1 .
1 / : : : b
m .
m /
(the “only if part” is a consequence of the definitions, the “if” part follows from
6 above).
8. If k is the k-th marginal of then
b
k .
/ D b
.0; : : : ; 0;
; 0; : : : ; 0/ :
"
kth
9. Let b 2 Rk and A a k m matrix. Then Y D AX C b is a Rk -valued r.v.; if

denotes its law, for
2 Rk ,

;Xi
.
/ D EŒeih
;AXCbi D eih
;bi EŒeihA
b .A
/ eih
;bi
Db (1.20)
(again A is the transpose of A).

10. Clearly if n !n!1 weakly then b n .
/ !n!1 b.
/ for every
; indeed,
x 7! eih
;xi is a continuous function having bounded real and imaginary parts.
Conversely, (P. Lévy’s theorem) if
lim b
n .
/ D .
/ for every
2 Rm
n!1
and if is continuous at the origin, then is the characteristic function of

some probability law and n !n!1 weakly.
1.7 Gaussian laws
A probability on R is said to be N.a; 2 / (normal, or Gaussian, with mean a

and variance 2 ), where a 2 R; > 0, if it has density with respect to Lebesgue
measure given by
1 h .x a/2 i
f .x/ D p exp :
2 2 2
Let us compute its characteristic function. We have at first, with the change of
variable x D y a,
Z C1 h .y a/2 i Z C1 x2
1 ei
a i
x 2 2
.
/ D p
b ei
y exp dy D p e e dx
2 1 2 2 2 1
and if we set
Z C1 x2
1
u.
/ D p ei
x e 2 2 dx ;
2 1
we have b
.
/ D u.
/ ei
a . Integrating by parts we have
Z C1 x2
1
u0 .
/ D p ix ei
x e 2 2 dx
2 1
ˇ Z C1
1 x2 ˇC1 2
x2
2 i
x 2 2 ˇ
Dp .i / e e ˇ p ei
x e 2 2 dx
2 1 2 1
D 2
u.
/ :
1 2
2
This is a first order differential equation. Its general solution is u.
/ D c e 2
and, recalling the condition u.0/ D 1, we find
1 2
2 1 2
2
u.
/ D e 2 ; .
/ D ei
a e 2
b :
By points 4. and 5. of the previous section one easily derives, by taking the derivative
at 0, that has mean a and variance 2 . If now N.b; 2 / then
h . 2 C 2 /
2 i
.
/b.
/ D b
.
/b
.
/ D ei
.aCb/ exp :
2

therefore has the same characteristic function as an N.a C b; 2 C 2 / law and,
by 6. of the previous section,
N.a C b; 2 C 2 /. In particular, if X and Y
are independent normal r.v.’s, then X C Y is also normal.
Let X1 ; : : : ; Xm be independent N.0; 1/-distributed r.v.’s and let X D .X1 ; : : : ; Xm /;
then the vector X has density
1 2 1 2 1 1 2
f .x/ D p ex1 =2 : : : p exm =2 D e 2 jxj (1.21)
2 2 .2/ m=2
and, by 7. of the previous section, its characteristic function is given by

2 2 1 2
e
1 =2 : : : e
m =2 D e 2 j
j :
If a 2 Rm and A is an m m matrix then, recalling (1.20), AX C a has characteristic

function
1
;A
i 1
eih
;ai e 2 hA D eih
;ai e 2 h
;
i ; (1.22)
where D AA .
Let a 2 Rm and let be an m m positive semi-definite matrix. A law
on Rm is said to be N.a; / (normal with mean a and covariance matrix ) if
its characteristic function is given by (1.22). This is well defined as we can prove
that (1.22) is certainly the characteristic function of a r.v.
It is actually well-known that if is an m m positive semi-definite matrix,
then there always exists a matrix A such that AA D ; this matrix is unique under
the additional requirement of being symmetric and in this case we will denote it
by 1=2 . Therefore (1.22) is the characteristic function of 1=2 X C a, where X
has density given by (1.21). In particular, (1.21) is the density of an N.0; I/ law (I
denotes the identity matrix).
In particular, we have seen that every X N.a; /-distributed r.v. can be written
as X D a C 1=2 Z, where Z N.0; I/. We shall often take advantage of this
property, which allows us to reduce computations concerning N.a; /-distributed
r.v.’s to N.0; I/-distributed r.v.’s, usually much simpler to deal with. This is also very
useful in dimension 1: a r.v. X N.a; 2 / can always be written as X D a C Z
with Z N.0; 1/.
Throughout the computation of the derivatives at 0, as indicated in 4. and 5. of
the previous section, a is the mean and the covariance matrix. Similarly as in the
one-dimensional case we find that if and are respectively N.a; / and N.b; /,
then
is N.a C b; C / and that the sum of independent normal r.v.’s is also
normal.
If the covariance matrix is invertible and D AA , then it is not difficult
to check that A also must be invertible and therefore, thanks to (1.20), an N.a; /-
distributed r.v. has density g given by (1.11), where f is the N.0; I/ density defined
in (1.21). Developing this relation and noting that det D .det A/2 , we find, more
explicitly, that
1 1 1 .ya/;yai
g.y/ D e 2 h :
.2/m=2 .det /1=2
If, conversely, the covariance matrix is not invertible, it can be shown that the
N.a; / law does not have a density (see Exercise 1.4 for example). In particular, the
N.a; 0/ law is also defined: it is the law having characteristic function
7! eih
;ai ,
and is therefore the Dirac mass at a.
Let X be a Gaussian Rm -valued N.a; /-distributed r.v., A a km matrix and b 2 Rk .
Let us consider the r.v. Y D AX C b, which is Rk -valued. By (1.20)

;ai 1
;A
i 1
;
i
b .A
/ eih
;bi D eih
;bi eihA
.
/ D b e 2 h A D eih
;bCAai e 2 hA A :
Therefore Y is Gaussian with mean b C Aa and variance A A . This fact can be

summarized by saying that
Theorem 1.3 Linear-affine transformations map normal laws into normal

laws.
This is a fundamental property. In particular, it implies that if X and Y are real

r.v.’s with a normal joint law, then X C Y is normal. It suffices to observe that X C Y
is a linear function of the vector .X; Y/.
The same argument gives that if D N.a; / and k is the k-th marginal
distribution of , then k is also normal (the projection on the k-th coordinate is
a linear function). This can also be deduced directly from 8. of the previous section:
1 2
k .
/ D e 2 kk
ei
ak
b (1.23)
and therefore k N.ak ; kk /.

If is diagonal, then using (1.22) and (1.23),
.
1 ; : : : ;
m / D b
b 1 .
1 / : : : b
m .
m / : (1.24)
By 7. of Sect. 1.6 we see then that D 1 ˝ ˝ m . Therefore
if X D .X1 ; : : : ; Xm / is normal and has a diagonal covariance matrix, its

components X1 ; : : : Xm are independent r.v.’s.
Therefore for two real jointly Gaussian r.v.’s X and Y, if they are uncorrelated, they
are also independent. This is a specific property of jointly normal r.v.’s which, as
already remarked, is false in general.
A similar criterion can be stated if X1 ; : : : ; Xm are themselves multidimensional:
if the covariances between the components of Xh and Xk , h 6D k; 1 h; k m,
vanish, then X1 ; : : : ; Xm are independent. Actually, if we denote by h the covariance
matrix of Xh , the covariance matrix of X D .X1 ; : : : ; Xm / turns out to be block
diagonal, with the blocks h on the diagonal. It is not difficult therefore to repeat the
previous argument and show that the relation (1.24) holds between the characteristic
functions b of X and those, b h , of the Xh , which implies the independence of
X1 ; : : : ; Xm .
Definition 1.1 A family I of d-dimensional r.v.’s defined on .˝; F ; P/ is

said to be a Gaussian family if, for every choice of X1 ; : : : ; Xm 2 I , the
dm-dimensional r.v. X D .X1 ; : : : ; Xm / is Gaussian.
Remark 1.2 Let us go back to Remark 1.1: let F1 D .X; X 2 I / and F2 D

.Y; Y 2 J / be -algebras generated respectively by the families I and J
of real r.v.’s. Let us assume, moreover, that I [ J is a Gaussian family. Then
it is immediate that F1 and F2 are independent if and only if
Cov.X; Y/ D 0 for every X 2 I ; Y 2 J : (1.25)
In fact this condition guarantees that the r.v.’s .X1 ; : : : ; Xn / and .Y1 ; : : : ; Yk /
are independent for every choice of X1 ; : : : ; Xn 2 I and Y1 ; : : : ; Yk 2 J
and therefore the criterion of Remark 1.1 is satisfied. Let us recall again
that (1.25) implies the independence of the generated -algebras only under
the assumption that the r.v.’s X 2 I ; Y 2 J are jointly Gaussian.
Proposition 1.8 Let I be a family of d-dimensional r.v.’s. Then it is a

Gaussian family if and only if for every X1 ; : : : ; Xm 2 I and 1 ; : : : ; m 2 Rd ,
the r.v. h1 ; X1 i C C hm ; Xm i is Gaussian.
Proof We shall assume d D 1 in order to make things simple.

Let us assume that every finite linear combination of the r.v.’s of I is Gaussian.
Hence, in particular, every r.v. of I is Gaussian. This implies that if X; Y 2 I then
EŒjXYj EŒX 2 1=2 EŒY 2 1=2 < C1 :
So that Cov.X; Y/ is well defined. Let X1 ; : : : ; Xm 2 I and X D .X1 ; : : : ; Xm / and

let us denote by the law of X. Let us compute its characteristic function b . Let
D .1 ; : : : ; m / 2 Rm . By hypothesis 1 X1 C C m Xm D h; Xi is N.a; 2 /-
distributed for some 2 0, a 2 R. Now we know that X has finite expectation, z,
and covariance matrix , hence
X
m
a D EŒh; Xi D i EŒXi D h; zi
iD1
X
m
Xm
2 D EŒh; Xi2 a2 D i j EŒXi Xj EŒXi EŒXj D ij i j :
i;jD1 i;jD1
Therefore we have
2 =2 1
. / D EŒeih;Xi D eia e
b D eih;zi e 2 h ; i
so X is Gaussian.
Conversely, if X D .X1 ; : : : ; Xm / is Gaussian and 2 Rm , then 1 X1 C C

m Xm D h; Xi is Gaussian, being a linear function of a Gaussian vector (see
Theorem 1.3). t
u
The following property of the Gaussian laws, which we shall use quite often, is,
together with the invariance with respect to affine transformations of Theorem 1.3,
the most important.
Proposition 1.9 Let .Xn /n be a sequence of m-dimensional Gaussian r.v.’s

converging in law to a r.v. X. Then X is also Gaussian.
Proof Thanks to Proposition 1.8 it is sufficient to prove that Y D h; Xi is Gaussian

for every 2 Rm .
The r.v.’s Yn D h; Xn i are Gaussian, being linear functions of Gaussian vectors.
Denoting by mn ; n2 respectively the mean and variance of Yn , the characteristic
function of Yn is
1 2 2
n .
/ D ei
mn e 2 n
:
By hypothesis n .
/ ! .
/ as n ! 1, where is the characteristic function of
Y. Taking the modulus we have, for every
2 R,
1 2 2
e 2 n
! j.
/j :
n!1
This proves that the sequence .n2 /n is bounded. Actually, if there existed a
subsequence converging to C1 we would have j.
/j D 0 for
6D 0 and
j.0/j D 1, which is not possible, being continuous. If 2 denotes the limit of
1 2 2 1 2 2
a subsequence of .n2 /n , then necessarily e 2 n
! e 2
.
Let us prove now that the sequence of the means, .mn /n , is also bounded. Note
that, Yn being Gaussian, we have P.Yn mn / D 12 if n2 > 0 and P.Yn mn / D 1
if n2 D 0, as in this case the law of Yn is the Dirac mass at mn . In any case, P.Yn
mn / 12 .
Let us assume that .mn /n is unbounded. Then there would exist a subsequence
.mnk /k converging to C1 (this argument is easily adapted to the case mnk ! 1).
For every M 2 R we would have, as mnk M for k large,
1,
P.Y M/ lim P.Ynk M/ limn!1 P.Ynk mnk / D
n!1 2
which is not possible, as necessarily limM!C1 P.Y M/ D 0. As .mn /n is

bounded, there exists a convergent subsequence .mnk /k , mnk ! m say. Therefore
1.8 Simulation 21
we have, for every

2 R,
1 2 2 1 2
2
.
/ D lim ei
nk e 2 nk
D ei
m e 2
k!1
hence Y is Gaussian. t
u
1.8 Simulation
Often one needs to compute a probability or an expectation for which there is no

explicit formula. In this case simulation is natural option.
For instance, if X is a r.v. and f a real Borel function, bounded to make things
simple, and we must know the value of the expectation
EŒf .X/ ;
then the Law of Large Numbers states that if .Xn /n is a sequence of independent
identically distributed r.v.’s with the same law as X, then
1X
N
f .Xi / ! EŒf .X/ :
N iD1 N!1
A first question in this direction is to be able to simulate a r.v. with a given

distribution, i.e. to instruct a computer to generate random numbers with a given
distribution.
In this section we shall see how to simulate Gaussian distributed random
numbers.
We will assume that the researcher has access to a programming language with
a random generator uniformly distributed on Œ0; 1, i.e. a command that produces a
random number with this law and such that, moreover, repeated calls to it give rise
to random numbers that are independent.
These are the so-called pseudo-random numbers. We shall not approach the
question of how they are obtained: we shall just be satisfied with the fact that all
the common programming languages (FORTRAN, C, . . . ) comply with our needs.
We shall see later, however, that it is always wise to doubt of the quality of these
generators. . .
The first problem is: how is it possible, starting from a random generator uniform
on Œ0; 1, to obtain a generator producing a r.v. with a Gaussian distribution? Various
algorithms are available (see also Exercise 1.21); curiously most of them produce
simultaneously two independent random numbers with an N.0; 1/ distribution. The
one provided by the following proposition is possibly the simplest.
Proposition 1.10 (The Box–Müller algorithm) Let W; Z be independent

r.v.’s respectively exponential with parameter 12 (see Exercise 1.2) and uniform
p p
on Œ0; 2. Then X D W cos Z and Y D W sin Z are independent and
N.0; 1/-distributed.
p
Proof First, let us compute the density of R D W. The partition function of W is
F.t/ D 1 et=2 (see also Exercise 1.2). Therefore the partition function of R is, for
r > 0,
p 2
FR .r/ D P. W r/ D P.W r2 / D 1 er =2
and FR .r/ D 0 for r 0. Therefore, taking the derivative, its density is fR .r/ D
2 1
rer =2 , r > 0. As the density of Z is equal to 2 on the interval Œ0; 2, the joint
p
density of W and Z is
1 r2 =2
f .r; z/ D re ; for r > 0; 0 z 2;
2
p
and f .r; z/pD 0 otherwise. Let us compute the joint density, g say, of X D W cos Z
and Y D W sin Z: g is characterized by the relation
Z C1 Z C1
EŒ˚.X; Y/ D ˚.x; y/g.x; y/ dx dy (1.26)
1 1
for every bounded Borel function ˚ W R2 ! R. Recalling the formulas of

integration in polar coordinates,
p p Z Z
EŒ˚.X; Y/ D EŒ˚. W cos Z; W sin Z/ D ˚.r cos z; r sin z/f .r; z/ dr dz
Z 2 Z C1
1 2
D dz ˚.r cos z; r sin z/ rer =2 dr
2 0 0
Z C1 Z C1
1 1 2 2
D ˚.x; y/e 2 .x Cy / dx dy
2 1 1
and, comparing with (1.26), we derive that the joint density g of X; Y is
1 1 .x2 Cy2 /
g.x; y/ D e 2 ;
2
which proves simultaneously that X and Y are independent and
N.0; 1/-distributed. t
u
1.8 Simulation 23
Proposition 1.10 suggests the following recipe for the simulation of an N.0; I/-distri-
buted r.v.: let Z be a r.v. uniform on Œ0; 2 and W an exponential r.v. with parameter
1
2
. They can be obtained from a uniform distribution as explained in Exercise 1.2).
Then the r.v.’s
p p
X D W cos Z; Y D W sin Z (1.27)
are Gaussian N.0; I/-distributed and independent. To be more explicit the steps
are:
• Simulate two independent r.v.’s U1 , U2 uniform on Œ0; 1 (these are provided by
the random number generator of the programming language that you are using);
• set Z D 2U1 , W D 2 log.1 U2 /; then Z is uniform on Œ0; 2 and W is
exponential with parameter 12 (see Exercise 1.2);
• then the r.v.’s X, Y as in (1.27) are N.0; 1/-distributed and independent.
This algorithm produces an N.0; 1/-distributed r.v., but of course from this we
can easily obtain an N.m; 2 /-distributed one using the fact that if X is N.0; 1/-
distributed, then m C X is N.m; 2 /-distributed.
Remark 1.3 a) The Box–Müller algorithm introduced in Proposition 1.10 is

necessary when using low level programming languages such as FORTRAN
or C. High level languages such as Matlab, Mathematica or scilab, in fact
already provide routines that directly produce Gaussian-distributed r.v.’s and,
in fact, the simulated paths of the figures of the forthcoming chapters have
been produced with one of them.
However, these high-level languages are interpreted, which means that
they are relatively slow when dealing with a real simulation, requiring the
production of possibly millions of random numbers. In these situations a
compiled language such as FORTRAN or C is necessary.
b) The huge number of simulations that are required by a real life
application introduces another caveat. The random number generator that is
usually available in languages such as FORTRAN or C is based on arithmetic
procedures that have a cycle, i.e. the generator produces numbers a0 ; a1 ; : : :
that can be considered uniformly distributed on Œ0; 1 and independent, but
after N numbers (N Dthe cycle length) it goes back to the beginning of the
cycle, repeating again the same numbers a0 ; a1 ; : : :
This means that it is not fit for your simulation if your application requires
the production of more than N random numbers. In particular, the C command
rnd() has a cycle N 2 109 , which can very well be insufficient (we shall
see that the simulation of a single path may require many random numbers).
Luckily more efficient random number simulators, with a reassuringly
large cycle length, have been developed and are available in all these
languages and should be used instead of the language default.
1.9 Measure-theoretic arguments
Very often in the sequel we shall be confronted with the problem of proving that
a certain statement is true for a large class of functions. Measure theory provides
several tools in order to deal with this kind of question, all based on the same idea:
just prove the statement for a smaller class of functions (for which the check is easy)
and then show that necessarily it must be true for the larger class. In this section we
give, without proof, three results that will be useful in order to produce this kind of
argument.
In general, if E is a set and C a class of parts of E, by .C / we denote the
-algebra generated by C , i.e. the smallest -algebra of parts of E containing C .
Proposition 1.11 Let .E; E / be a measurable space and f a positive real

measurable function on E. Then there exists an increasing sequence . fn /n of
functions of the form
X
m
fn .x/ D ˛i 1Ai .x/ (1.28)
iD1
such that fn % f .
Functions of the form appearing on the right-hand side of (1.28) are called ele-
mentary. Therefore Proposition 1.11 states that every measurable positive function
is the increasing limit of a sequence of elementary functions.
Theorem 1.4 Let E be a set, C a class of parts of E stable with respect to

finite intersections and H a vector space of bounded functions such that:
i) if . fn /n is an increasing sequence of elements of H all bounded above by
the same element of H then supn fn 2 H ; and
ii) H contains the function 1 and the indicator functions of the elements of
C.
Then H contains every bounded .C /-measurable function.
Theorem 1.5 Let H be a vector space of bounded real functions on a set E

such that:
(continued)
1.9 Exercises for Chapter 1 25

i) 1 2 H ;
ii) if . fn /n H and fn !n!1 f uniformly, then f 2 H ; and
iii) if . fn /n H is an increasing sequence of equi-bounded positive functions
and limn!1 fn D f , then f 2 H .
Let L be a subspace of H stable under multiplication. Then H contains
every bounded .L/-measurable function.
Let us prove Theorem 1.1 (Carathéodory’s criterion) as an example of application

of Theorem 1.4. Let H be the set of the real bounded measurable functions f on E
such that
Z Z
f d1 D f d2 :
H is a vector space and satisfies i) of Theorem 1.4 thanks to Lebesgue’s theorem.

Moreover, of course, H contains the function 1 and the indicators of the elements
of I . Therefore H contains every bounded .I /-measurable function. As by
assumption .I / D E , 1 and 2 coincide on E .
This example of application of Theorem 1.4 is typical when one must prove in
general a property that is initially only known for a particular class of events.
Similarly, Theorem 1.5 will be used, typically, in order to extend to every
bounded measurable function a property which is known at first only for a particular
class of functions: for instance, the class of continuous functions on .E; B.E//, if E
is a metric space.
Exercises
1.1 (p. 437) Given a real r.v. X, its partition function (p.f.) is the function
F.t/ D P.X t/ :
Show that two real r.v.’s X and Y have the same p.f. if and only if they have the same
distribution.
Use Carathéodory’s criterion, Theorem 1.1.
1.2 (p. 437) a) A r.v. X has exponential law with parameter if it has density
f .x/ D e x 1Œ0;C1Œ .x/ :

What is the p.f. F of X (see the definition in Exercise 1.1)? Compute the mean and
variance of X.
b) Let U be a r.v. uniformly distributed on Œ0; 1, i.e. having density
f .x/ D 1Œ0;1 .x/:
Compute the mean and variance of U.

c) Let U be as in b).
c1) Compute the law of Z D ˛U, ˛ > 0.
c2) Compute the law of W D 1 log U, > 0.
1.3 (p. 439) a) Let X be a positive r.v. and f W RC ! R a differentiable function

with continuous derivative and such that f .X/ is integrable. Then
Z C1
EŒf .X/ D f .0/ C f 0 .t/P.X t/ dt :
0
b) Let X be a positive integer-valued r.v. Then

1
X
EŒX D P.X k/ :
kD1
R C1 R C1 R C1 Rx
a) If is the law of X, then 0 f 0 .t/ dt t .dx/ D 0 .dx/ 0 f 0 .t/ dt by Fubini’s
theorem.
1.4 (p. 439) Let X be an m-dimensional r.v. and let us denote by C its covariance
matrix.
a) Prove that if X is centered then P.X 2 Im C/ D 1 (Im C is the image of the
matrix C).
b) Deduce that if the covariance matrix C of a r.v. X is not invertible, then the
law of X cannot have a density.
1.5 (p. 439) Let X; Xn ; n D 1; 2; : : : , be Rm -valued r.v.’s. Prove that if from every
subsequence of .Xn /n we can extract a further subsequence convergent to X in Lp
(resp. in probability) then .Xn /n converges to X in Lp (resp. in probability). Is this
also true for a.s. convergence?
1.6 (p. 440) Given an m-dimensional r.v. X, its Laplace transform is the function
Rm 3
! EŒeh
;Xi
(possibly D C1). Prove that, if X N.b; /, then

1
EŒeh
;Xi D eih
;bi e 2 h
;
i :
1.7 (p. 440) Prove points 2 and 8 of Sect. 1.6.

1.8 (p. 440) Let X be a real r.v. having a Laplace law with parameter , i.e. having
density
jxj
f .x/ D e :
2
Compute its characteristic function and its Laplace transform.
1.9 (p. 441) Let X, Y be independent N.0; 1/-distributed r.v.’s. Determine the laws
p
of the two-dimensional r.v.’s .X; X C Y/ and .X; 2X/. Show that these two laws
have the same marginals.
1.10 (p. 442) Let X1 , X2 be independent N.0; 1/-distributed r.v.’s. If Y1 D X1p X2 ,
3
Y2 D X1 CX2 , show that Y1 and Y2 are independent. And if it were Y1 D 12 X1 2 X2 ,
p
3
Y2 D 12 X1 C 2
X2 ?
1.11 (p. 443) a) Let X be an N.; 2 /-distributed r.v. Compute the density of eX
(lognormal law of parameters and 2 ).
b) Show that a lognormal law has finite moments of all orders and compute them.
What are the values of its mean and variance?
2
1.12 (p. 443) Let X be an N.0; 2 /-distributed r.v. Compute, for t 2 R, EŒetX .
1.13 (p. 444) Let X be an N.0; 1/-distributed r.v., ; b real numbers and x; K > 0.
Show that
1 2
EŒ.xebC X K/C D xebC 2 ˚. C / K˚./ ;
where D 1 .log Kx b/ and ˚ denotes the partition function of an N.0; 1/-distribu-
ted r.v. This quantity appears naturally in many questions in mathematical finance,
see Sect. 13.6. (xC denotes the positive part function, xC D x if x 0, xC D 0 if
x < 0.)
1.14 (p. 444) a) Let .Xn /n be a sequence of m-dimensional Gaussian r.v.’s
respectively with mean bn and covariance matrix n . Let us assume that
lim bn WD b; lim n WD :
n!1 n!1
L
Show that Xn ! N.b; / as n ! 1.
b1) Let .Zn /n be a sequence of N.0; 2 /-distributed real independent r.v.’s. Let
.Xn /n be the sequence defined recursively by
X0 D x 2 R; XnC1 D ˛Xn C ZnC1

where j˛j < 1 (i.e., possibly, X1 ; : : : ; Xn ; : : : represent the subsequent positions of

a moving object that at every iteration moves from the actual position Xn to the
position ˛Xn but also undergoes a perturbation ZnC1 ). What is the law of X1 ? And
of X2 ? Prove that, as n ! 1, .Xn /n converges in law and determine the limit law.
b2) Prove that, as n ! 1, the sequence of two-dimensional r.v.’s ..Xn ; XnC1 //n
converges in law and determine the limit law.
1.15 (p. 446) a) Prove that, for every p 0, there exists a constant cp;m > 0 such
that E.jXjp / cp;m E.jXj2 /p=2 for every Gaussian m-dimensional centered r.v. X.
b) Let .Xn /n be a sequence of Gaussian r.v.’s converging to 0 in L2 . Show that
the convergence also takes place in Lp for every p > 0.
Recall the inequality, for x D .x1 ; : : : ; xm / 2 Rm ,
X
m
p2 X
m
jxi jp jxjp m 2 jxi jp :
iD1 iD1
1.16 (p. 446) (Example of a pair of Gaussian r.v.’s whose joint law it is not
Gaussian) Let X; Z be independent r.v.’s with X N.0; 1/ and P.Z D 1/ D P.Z D
1/ D 12 . Let Y D XZ.
a) Prove that Y is itself N.0; 1/.
b) Prove that X C Y is not Gaussian. Does .X; Y/ have a joint Gaussian law?
1.17 (p. 447) a) Let .Xn /n be a sequence of independent N.0; 1/-distributed r.v.’s.
Prove that, for every ˛ > 2,

P Xn > .˛ log n/1=2 for infinitely many indices n D 0 : (1.29)
b) Prove that

P Xn > .2 log n/1=2 for infinitely many indices n D 1 : (1.30)
c) Show that the sequence ..log n/1=2 Xn /n tends to 0 in probability but not a.s.
Use the following inequalities that will be proved later (Lemma 3.2)
Z C1
1 1 x2 =2 2 1 2
xC e ez =2 dz ex =2 :
x x x
1.18 (p. 448) Prove Proposition 1.1 (the integration rule with respect to an image
probability).
1.19 (p. 449) (A very useful measurability criterion) Let X be a map .˝; F / !
.E; E /, D E a family of subsets of E such that .D/ D E and let us assume that
X 1 .A/ 2 F for every A 2 D. Show that X is measurable.
1.20 (p. 449) (A special trick for the L1 convergence of densities) Let Zn ; Z be
positive r.v.’s such that Zn ! Z a.s. as n ! 1 and EŒZn D EŒZ < C1 for every
n. We want to prove that the convergence also takes place in L1 .
a) Let Hn D min.Zn ; Z/. Prove that limn!1 EŒHn D EŒZ.
b) Note that jZn Zj D .Z Hn / C .Zn Hn / and deduce that Zn ! Z also in L1 .
1.21 (p. 449) In the FORTRAN libraries in use in the 70s (but also nowadays . . . ),
in order to generate an N.0; 1/-distributed random number the following procedure
was implemented. If X1 ; : : : ; X12 are independent r.v.’s uniformly distributed on
Œ0; 1, then the number
W D X1 C C X12 6 (1.31)
is (approximately) N.0; 1/-distributed.

a) Can you give a justification of this procedure?
b) Let Z be a N.0; 1/-distributed r.v. What is the value of EŒZ 4 ? What is the
value of EŒW 4 for the r.v. given by (1.31)? What do you think of this procedure?
Chapter 2
Stochastic Processes
2.1 General facts
A stochastic process is a mathematical object that is intended to model the evolution

in time of a random phenomenon. As will become clear in the sequel the appropriate
setting is the following.
A stochastic process is an object of the form
X D .˝; F ; .Ft /t2T ; .Xt /t2T ; P/ ;
where
• .˝; F ; P/ is a probability space;
• T (the times) is a subset of RC ;
• .Ft /t2T is a filtration, i.e. an increasing family of sub--algebras of F :
Fs Ft whenever s t; and
• .Xt /t2T is a family of r.v.’s on .˝; F / taking values in a measurable space
.E; E / such that, for every t, Xt is Ft -measurable. This fact is also expressed
by saying that .Xt /t is adapted to the filtration .Ft /t .
Let us look at some aspects of this definition.

Of course Xt will represent the random position of our object at time t.
At this point the meaning of the -algebras Ft and the role that they are going
to play may not be evident; we will understand them well only later. For now let us
just say that Ft is “the amount of information that is available at time t”, i.e. it is the
family of the events for which, at time t, one knows whether they happened or not.
Note that, if s t, as Xs must be Fs -measurable and Fs Ft , then all the r.v.’s
Xs for s t, must be Ft measurable, so that Ft necessarily contains the -algebra
Gt D .Xs ; s t/. These -algebras form the natural filtration .Gt /t , which is
therefore the smallest possible filtration with respect to which X is adapted. As Xs

32 2 Stochastic Processes
for s t is an Ft measurable r.v., this means intuitively that at time t we know the
positions of the process at the times before timeSt.
In general, if .Ft /t is a filtration, the family t Ft is not necessarily aS-algebra.
By F1 we shall denote the smallest -algebra of parts of ˝ containing W t Ft . This
is a -algebra that can be strictly smaller than F ; it is also denoted t Ft .
Another filtration that we shall often be led to consider is the augmented natural
filtration, .G t /t , where G t is the -algebra obtained by adding to Gt D .Xu ; u t/
the negligible events of F , i.e.
Gt D .Xu ; u t/; G t D .Gt ; N / ;
where N D fAI A 2 F ; P.A/ D 0g.

The space E appearing above (in which the process takes its values) is called the
state space. We shall mainly be interested in the case .E; E / D .Rm ; B.Rm //; in
general, however, it might also be a more involved object (a group of matrices, a
manifold,. . . ).
Via the map ! 7! .t 7! Xt .!// one can always think of ˝ as a subset of ET (ET
is the set of all functions T ! E); for this reason ˝ is also called the space of paths.
This point of view will be developed in Sect. 3.2.
Let .˝; F ; .Ft /t2T ; .Xt /t2T ; P/ and .˝ 0 ; F 0 ; .Ft0 /t2T ; .Xt0 /t2T ; P0 / be two processes.
They are said to be equivalent if for every choice of t1 ; : : : ; tm 2 T the r.v.’s

.Xt1 ; : : : ; Xtm / and .Xt01 ; : : : ; Xt0m / have the same law.
As we shall see in Sect. 2.3, the notion of equivalence of processes is very important:
in a certain sense two equivalent processes “are the same process”, at least in the
sense that they model the same situation.
One process is said to be a modification (or a version) of the other if

.˝; F ; .Ft /t2T ; P/ D .˝ 0 ; F 0 ; .Ft0 /t2T ; P0 / and if, for every t 2 T, Xt D Xt0
P-a.s. In this case, in particular, they are equivalent.
They are said to be indistinguishable if
.˝; F ; .Ft /t2T ; P/ D .˝ 0 ; F 0 ; .Ft0 /t2T ; P0 /
and
P.Xt D Xt0 for every t 2 T/ D 1 :
Two indistinguishable processes are clearly modifications of one another. The

converse is false, as shown in the following example.
2.1 General facts 33
Example 2.1 If ˝ D Œ0; 1; F D B.Œ0; 1/ and P DLebesgue measure, let
Xt .!/ D 1f!g .t/; Xt0 .!/ 0 :
X 0 is a modification of X, because the event fXt 6D Xt0 g is formed by the only

element t, and therefore has probability 0. But the two processes are not indis-
tinguishable, and the event f!I Xt .!/ D Xt0 .!/ for every tg is even empty.
Let us assume from now on that the state space E is a topological space endowed
with its Borel -algebra B.E/ and that T is an interval of RC .
A process is said to be continuous (resp. a.s. continuous) if for every ! (resp.
for almost every !) the map t 7! Xt .!/ is continuous. The definitions of a right-
continuous process, an a.s. right-continuous process, etc., are quite similar.
Note that the processes X and X 0 of Example 2.1 are modifications of each
other but, whereas X 0 is continuous, X is not. Therefore, in general, the
property of being continuous is not preserved when passing from a process
to a modification.
X is said to be measurable if the map .t; !/ 7! Xt .!/ is measurable .T ˝; B.T/˝

F / ! .E; B.E//. It is said to be progressively measurable if for every u 2 T the
map .t; !/ 7! Xt .!/ is measurable .Œ0; u ˝; B.Œ0; u/ ˝ Fu / ! .E; B.E//. A
progressively measurable process is obviously measurable.
It is crucial to assume that the processes we are working with are progressively
measurable (see Examples 2.2 and 2.3 below). Luckily we will be dealing with
continuous processes and the following proposition guarantees that they are pro-
gressively measurable.
Proposition 2.1 Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be a right-continuous

process. Then it is progressively measurable.
Proof For a fixed u 2 T let us consider the piecewise constant process

(
.n/
X kC1 if s 2 Œ 2kn u; kC1
2n
uŒ
Xs D 2n u
Xu if s D u :
.n/
Note that we can write Xs D Xsn where sn > s is a time such that jsn sj u2n
(sn D .k C 1/u2n if s 2 Œ 2kn u; kC12n uŒ). Hence sn & s as n ! 1 and, as X is
.n/
assumed to be right-continuous, Xs ! Xs as n ! 1 for every s u.
Let us prove now that X .n/ is progressively measurable, i.e. that if 2 B.E/
.n/
then the event f.s; !/I s u; Xs .!/ 2 g belongs to B.Œ0; u/ ˝ Fu . This follows
from the relation
f.s; !/I s u; Xs.n/ .!/ 2 g

[ ˚
D k
2n
u; kC1
2n
u X kC1 n u
2 [ fug fXu 2 g
2
k<2n
and the right-hand side is an element of B.Œ0; u/ ˝ Fu .

The map .s; !/ 7! Xs .!/ is now B.Œ0; u/ ˝ Fu -measurable, being the limit of
B.Œ0; u/ ˝ Fu -measurable functions. t
u
Example 2.2 Let us see a situation where the assumption of progressive

measurability is needed. Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ Rbe a process. We shall
t
often be led to the consideration of the process Yt .!/ D 0 Xs .!/ ds (assuming
that t 7! Xt is integrable for almost every !). Is the new process .Yt /t adapted
to the filtration .Ft /t ? That is, is .˝; F ; .Ft /t ; .Yt /t ; P/ a process?
Let us fix t > 0. If X is progressively measurable, then the map .!; s/ 7!
Xs .!/; ! 2 ˝; s t, is Ft ˝ B.Œ0; t/-measurable. Therefore Rt by the first part
of Fubini’s theorem (Theorem 1.2) the map ! 7! 0 Xs .!/ ds D Yt is Ft -
measurable. Therefore Y is adapted and, being continuous, also progressively
measurable. Without the assumption of progressive measurability for X, Y
might not be adapted.
Example 2.3 Let X be a stochastic process and a positive r.v. Sometimes we

shall be led to the consideration of the r.v. X , i.e. the position of X at the random
time . Can we say that X is a r.v.?
The answer is positive if we assume that X is a measurable process: in this
case .!; t/ 7! Xt .!/ is measurable and X is obtained as the composition of the
measurable maps ! 7! .!; .!// and .!; t/ 7! Xt .!/.
In order to be rigorous we will sometimes need to make some assumptions

concerning the filtration. These will not produce difficulties, as we shall soon see
that they are satisfied
T for our processes of interest.
Let FtC D ">0 FtC" . Clearly FtC is a -algebra (the intersection of any
family of -algebras is always a -algebra) and Ft FtC ; we say that the filtration
is right-continuous if FtC D Ft for every t.
A process is said to be standard if

a) the filtration .Ft /t is right-continuous;
b) for every t, Ft contains the negligible events of F .
2.2 Kolmogorov’s continuity theorem 35
Note that the assumption that a process is standard concerns only the filtration and
not, for instance, the r.v.’s Xt .
A situation where this kind of assumption is needed is the following: let X D
.˝; F ; .Ft /t ; .Xt /t ; P/ be a process and .Yt /t a family of r.v.’s such that Xt D Yt
a.s. for every t. In general this does not imply that Y D .˝; F ; .Ft /t ; .Yt /t ; P/ is
a process. Actually, Yt might not be Ft -measurable because the negligible event
Nt D fXt 6D Yt g might not belong to Ft . This problem does not appear if the space
is standard. Moreover, in this case every a.s. continuous process has a continuous
modification.
The fact that the filtration .Ft /t is right-continuous is also a technical assumption
that is often necessary; this explains why we shall prove, as soon as possible, that
we can assume that the processes we are dealing with are standard.
2.2 Kolmogorov’s continuity theorem
We have seen in Example 2.1 that a non-continuous process can have a continuous
modification. The following classical theorem provides a simple criterion in order
to ensure the existence of such a continuous version.
Theorem 2.1 (Kolmogorov’s continuity theorem) Let D Rm be an open

set and .Xy /y2D a family of d-dimensional r.v.’s on .˝; F ; P/ such that there
exist ˛ > 0, ˇ > 0, c > 0 satisfying
EŒjXy Xz jˇ cjy zjmC˛ :
Then there exists a family .e

X y /y of Rd -valued r.v.’s such that
Xy D e
Xy a.s. for every y 2 D
(i.e. e
X is a modification of X) and that, for every ! 2 ˝, the map y 7! e
X y .!/
is continuous and even Hölder continuous with exponent for every < ˇ˛
on every compact subset of D.
In the proof we shall assume D D0; 1Œm . The key technical point is the following
lemma.
Lemma 2.1 Under the assumptions of Theorem 2.1, let D D0; 1Œm and let
us denote by DB the set, which is dense in D, of the dyadic points (i.e. points
having as coordinates fractions with powers of 2 in the denominator). Then,
for every < ˇ˛ , there exists a negligible event N such that the restriction of
X to DB is Hölder continuous with exponent on N c .
Proof For a fixed n let An DB be the set of the points y 2 D whose coordinates
are of the form k2n . Let < ˇ˛ and let
n D fjXy Xz j > 2n for some z; y 2 An with jy zj D 2n g :
If y; z 2 An are such that jy zj D 2n , then, by Markov’s inequality (1.6),
P.jXy Xz j > 2n / 2nˇ E.jXy Xz jˇ / 2n.ˇ ˛m/ :
As the set of the pairs y; z 2 An such that jy zj D 2n has cardinality 2m 2nm ,
P.n / 2m2nmCn.ˇ ˛m/ const 2n
where D ˛ˇ > 0. This is the general term of a convergent series and therefore,
by the Borel–Cantelli lemma, there exists a negligible event N such that, if ! 2 N c ,
we have ! 2 nc eventually. Let us fix now ! 2 N c and let n D n.!/ be such that
! 2 kc for every k > n. Let us assume at first m D 1. Let y 2 DB : if > n and
y 2 Œi 2 ; .i C 1/2 Œ then
X
r
y D i2 C ˛` 2` ;
`DC1
where ˛` D 0 or 1. By the triangle inequality
X
r
ˇ X
kC1
X
k
ˇ
jXy Xi2 j ˇX i2 C ˛` 2` X i2 C ˛` 2` ˇ
kDC1 `DC1 `DC1
X
r
1
2.Ck/ 2 :
kD1
1 2
2.2 Kolmogorov’s continuity theorem 37
Let now y; z 2 DB be such that jy zj 2 ; there are two possibilities: if there
exists an i such that .i 1/2 y < i2 z < .i C 1/2 then
2
jXy Xz j jXz Xi2 jCjXy X.i1/2 jCjXi2 X.i1/2 j 1C 2 :
1 2
Otherwise y, z 2 Œi2 ; .i C 1/2 Œ and then
2
jXy Xz j jXy Xi2 j C jXz Xi2 j 2 :
1 2
Thus we have proved that if y, z 2 DB and jy zj 2 then jXy Xz j k2 ,

where k does not depend on or, equivalently, that if y, z 2 DB and jy zj 2
then
jXy Xz j kjy zj
for every > n. The lemma is therefore proved if m D 1. Let us now consider the
case m > 1. We can repeat the same argument as in dimension 1 and derive that the
previous relation holds as soon as y and z differ at most by one coordinate. Let us
define x.i/ 2 Rm , for i D 0; : : : ; m, by
(
.i/ yi if ji
xj D
zj if j>1:
Therefore x.0/ D z, x.m/ D y and x.i/ and x.iC1/ have all but one of their coordinates
equal, and then
X
m X
m
jXy Xz j jXx.i/ Xx.i1/ j k jx.i/ x.i1/ j mkjy zj ;
iD1 iD1

t
u
˛
Proof of Theorem 2.1 By Lemma 2.1, if < then there exists a negligible event
ˇ
N such that if ! 62 N, the restriction of y 7! Xy .!/ to DB is Hölder continuous
with exponent and hence uniformly continuous. Let us denote by e
X . .!/ its unique
continuous extension to D. It suffices now to prove that e
X y D Xy a.s. This fact is
obvious if y 2 DB . Otherwise let . yn /n DB be a sequence converging to y. The
assumption and Markov’s inequality (1.6) imply that Xyn !n!1 Xy in probability
and therefore, possibly taking a subsequence, Xyn !n!1 Xy a.s. As e X yn D Xyn a.s.
for every n, this implies that Xy D e
X y a.s. t
u
Corollary 2.1 Let X be an Rd -valued process such that there exist ˛ > 0,
ˇ > 0, c > 0 satisfying, for every s; t,
EŒjXt Xs jˇ cjt sj1C˛ :
Then there exists a modification Y of X that is continuous. Moreover, for

every < ˇ˛ the paths of Y are Hölder continuous with exponent on every
bounded time interval.
Example 2.4 In the next chapter we shall see that a Brownian motion is a real-
valued process B D .˝; F ; .Ft /t0 ; .Bt /t0 ; P/ such that
i) B0 D 0 a.s.;
ii) for every 0 s t the r.v. Bt Bs is independent of Fs ;
iii) for every 0 s t Bt Bs is N.0; t s/-distributed.
Let us show that a Brownian motion has a continuous modification. It is
sufficient to check the condition of Corollary 2.1. Let t > s; as Bt Bs
N.0; t s/, we have Bt Bs D .t s/1=2 Z with Z N.0; 1/. Therefore
EŒjBt Bs jˇ D .t s/ˇ=2 EŒjZjˇ :
As EŒjZjˇ < C1 for every ˇ > 0, we can apply Corollary 2.1 with ˛ D
ˇ
2
1. Hence a Brownian motion has a continuous version, which is also Hölder
continuous with exponent for every < 12 ˇ1 ; i.e., ˇ being arbitrary, for
every < 12 .
2.3 Construction of stochastic processes
Let X D .˝; F ; .Ft /t2T ; .Xt /t2T ; P/ be a process taking values in the topological
space .E; B.E// and D .t1 ; : : : ; tn / an n-tuple of elements of T with t1 < < tn .
Then we can consider the r.v.
X D .Xt1 ; : : : ; Xtn / W ˝ ! En D E E
and denote by its distribution. The probabilities are called the finite-dimen-
sional distributions of the process X.
2.3 Construction of stochastic processes 39
Note that two processes have the same finite-dimensional distributions if and
only if they are equivalent.
This family of probabilities is quite important: the finite-dimensional distributions

characterize P in the following sense.
Proposition 2.2 Let X D .˝; F ; .Ft /t2T ; .Xt /t2T ; P/ and X 0 D

.˝; F ; .Ft /t2T ; .Xt /t2T ; P0 / be two processes (defined on the same .˝; F /)
having the same finite-dimensional distributions. Then P and P0 coincide on
the -algebra .Xt ; t 2 T/.
Proof If 1 ; : : : ; n 2 B.E/, as X and X 0 have the same finite-dimensional

distributions,
P.Xt1 2 1 ; : : : ; Xtn 2 n / D .1 n / D P0 .Xt1 2 1 ; : : : ; Xtn 2 n / ;
i.e. P and P0 coincide on the events of the type
fXt1 2 1 ; : : : ; Xtn 2 n g :
As these events form a class that is stable with respect to finite intersections, thanks
to Carathéodory’s criterion, Theorem 1.1, P and P0 coincide on the generated -
algebra, i.e. .Xt ; t 2 T/.
t
u
A very important problem which we are going to be confronted with later is the
converse: given a topological space E, a time span T and a family . /2˘ of finite-
dimensional distributions (˘ D all possible n-tuples of distinct elements of T for
n ranging over the positive integers), does an E-valued stochastic process having
. /2˘ as its family of finite-dimensional distributions exist?
It is clear, however, that the ’s cannot be anything. For instance, if D ft1 ; t2 g
and 0 D ft1 g, then if the . /2˘ were the finite-dimensional distributions of some
process .Xt /t , would be the law of .Xt1 ; Xt2 / and 0 the law of Xt1 . Therefore
0 would necessarily be the first marginal of . This can also be stated by saying
that 0 is the image of through the map p W E E ! E given by p.x1 ; x2 / D x1 .
More generally, in order to be the family of finite-dimensional distributions
of some process X, the family . /2˘ must necessarily satisfy the following
consistency condition.
Condition 2.1 Let D .t1 ; : : : ; tn / and 0 D .t1 ; : : : ; ti1 ; tiC1 ; : : : ; tn /.

Let pi W En ! En1 be the map defined as .x1 ; : : : ; xn / 7!
.x1 ; : : : ; xi1 ; xiC1 ; : : : ; xn /. Then the image of through pi is equal to 0 .
The next theorem states that Condition 2.1 is also sufficient for . /2˘ to be
the system of finite-dimensional distributions of some process X, at least if the
topological space E is sufficiently regular.
Theorem 2.2 (Kolmogorov’s existence theorem) Let E be a complete

separable metric space, . /2˘ a system of finite-dimensional distributions
on E satisfying Condition 2.1. Let ˝ D ET (ET is the set of all paths
T ! E) and let us define Xt .!/ D !.t/, Ft D .Xs ; s t/, F D
.Xt ; t 2 T/. Then there exists on .˝; F / a unique probability P such that
. /2˘ is the family of the finite-dimensional distributions of the process
X D .˝; F ; .Ft /t2T ; .Xt /t2T ; P/.
Uniqueness is of course a consequence of Proposition 2.2.
Example 2.5 (Gaussian processes) An Rm -valued process .Xt /t is said to be

Gaussian if it is a Gaussian family i.e. if its finite-dimensional distributions
are Gaussian. If we define
bt D E.Xt /; its mean function

ij
Ks;t D Cov.Xi .s/; Xj .t//; 1 i; j m; its covariance function ;
then the finite-dimensional distributions of .Xt /t are completely determined

by these two quantities.
Let us, for instance, determine the distribution of the 2m-dimensional r.v.
.Xs ; Xt /, s < t. By assumption it is Gaussian (Proposition 1.8) and has mean
.bs ; bt /. As Kt;t is the m m covariance matrix of Xt , the covariance matrix of
.Xs ; Xt / is the 2m 2m block matrix

Kt;t Ks;t
:
Kt;s Ks;s
The law of .Xs ; Xt / is therefore determined by the functions K and b.

Similarly, if 0 t1 tn , then the r.v. .Xt1 ; : : : ; Xtn / is Gaussian, has
a covariance matrix with blocks Kti ;ti on the diagonal and blocks Kti ;tj outside
the diagonal and mean .bt1 ; : : : ; btn /.
(continued )
2.3 Construction of stochastic processes 41
(continued)
Conversely, given a mean and covariance functions, does an associated
Gaussian process exist?
We must first point out that the covariance function must satisfy an
important property. Let us consider for simplicity the case m D 1, i.e. of a
real-valued process X. A real function .s; t/ 7! C.s; t/ is said to be a positive
definite kernel if, for every choice of t1 ; : : : ; tn 2 RC and 1 ; : : : ; n 2 R,
X
n
C.ti ; tj /i j 0 (2.1)
i;jD1
i.e. if the matrix .C.ti ; tj //i;j is positive definite.

The covariance function K is necessarily a positive definite kernel: actually
.Kti ;tj /ij is the covariance matrix of the random vector .Xt1 ; : : : ; Xtn / and every
covariance matrix is positive definite (as was shown in (1.14)).
Conversely, let us prove that if K is a positive definite kernel, then there
exists a Gaussian process X associated to it. This is a simple application of
Kolmogorov’s existence Theorem 2.2.
If D .t1 ; : : : ; tm /, let us define the finite-dimensional distribution of
X to be Gaussian with covariance matrix
ij D Kti ;tj :
Let us check that this family of finite-dimensional distributions satisfies

the coherence Condition 2.1. If 0 D .t1 ; : : : ; ti1 ; tiC1 ; : : : ; tm / then 0
is Gaussian with covariance matrix obtained by removing the i-th row and
the i-th column from . Let us denote by 0 such a matrix. We must
check that 0 coincides with pi . /, where pi is the map pi .x1 ; : : : ; xn / D
.x1 ; : : : ; xi1 ; xiC1 ; : : : ; xn /.
But this is immediate: pi . / is also a Gaussian law, being the image of
a Gaussian distribution through the linear mapping pi (recall Theorem 1.3).
Moreover, the covariance matrix of pi . / is equal to 0 . To see this simply
observe that if .X1 ; : : : ; Xm / is a r.v. having law and therefore covariance
matrix , then .X1 ; : : : ; Xi1 ; XiC1 ; : : : ; Xm / has distribution
pi . /
and covariance matrix 0 .

Therefore if K is a positive definite kernel, then there exists a stochastic
process X which is Gaussian with covariance function K and mean function
bt 0. Now, for every mean function b the process Yt D Xt C bt is Gaussian
with covariance function K and mean function b.
2.4 Next. . .
In this chapter we have already met some of the relevant problems which arise in
the investigation of stochastic processes:
a) the construction of processes satisfying particular properties (that can be reduced
to finite-dimensional distributions); for instance, in the next chapter we shall
see that it is immediate, from its definition, to determine the finite-dimensional
distributions of a Brownian motion;
b) the regularity of the paths (continuity, . . . );
c) the determination of the probability P of the process, i.e. the computation of the
probability of events connected to it. For instance, for a Brownian motion B, what
is the value of P.sup0st Bs 1/?
Note again that, moving from a process to one of its modifications, the finite-
dimensional distributions do not change, whereas other properties, such as regularity
of the paths, can turn out to be very different, as in Example 2.1.
In the next chapters we investigate a particular class of processes: diffusions.
We shall be led to the development of particular techniques (stochastic integral)
that, together with the two Kolmogorov’s theorems, will allow us first to prove their
existence and then to construct continuous versions. The determination of P, besides
some particular situations, will in general not be so simple. We shall see, however,
that the probability of certain events or the expectations of some functionals of
the process can be obtained by solving suitable PDE problems. Furthermore these
quantities can be computed numerically by methods of simulation.
These processes (i.e. diffusions) are very important
a) first because there are strong links with other areas of mathematics (for example,
the theory of PDEs, but in other fields too)
b) but also because they provide models in many applications (control theory,
filtering, finance, telecommunications, . . . ). Some of these aspects will be
developed in the last chapter.
Exercises
2.1 (p. 451) Let X and Y be two processes that are modifications of one another.
a) Prove that they are equivalent.
b) Prove that if the time set is RC or a subinterval of RC and X and Y are both
a.s. continuous, then they are indistinguishable.
2.2 (p. 451) Let .Xt /0tT be a continuous process and D a dense subset of Œ0; T.
a) Show that .Xt ; t T/ D .Xt ; t 2 D/.
b) What if .Xt /0tT was only right-continuous?
2.3 (p. 452) Let X D .˝; F ; .Ft /t ; .Xt /t be a progressively measurable process
with values in the measurable space .E; E /. Let W E ! G be a measurable
function into the measurable space .G; G /. Prove that the G-valued process t 7!
.Xt / is also progressively measurable.
2.4 (p. 452) a) Let Z1 ; Zn , n D 1; : : : , be r.v.’s on some probability space
.˝; F ; P/. Prove that the event flimn!1 Zn 6D Z1 g is equal to
1 \
[ 1 [
1
fjZn Z1 j mg : (2.2)
mD1 n0 D1 nn0
e ; .F
b) Let .˝; F ; .Ft /t2T ; .Xt /t2T ; P/ and . e̋ ; F ft /t2T ; .e
X t /t2T ; e
P/ be equivalent
processes.
b1) Let .tn /n T. Let us assume that there exists a number ` 2 R such that
lim Xtn D ` a.s.

n!1
Then also
lim e
X tn D ` a.s.
n!1
b2) Let t 2 T. Then if the limit
lim Xs D `
s!t;s2Q
exists for some ` 2 R then also
lim e
Xs D ` : (2.3)
s!t;s2Q
2.5 (p. 453) An example of a process that comes to mind quite naturally is so-
called “white noise”, i.e. a process .Xt /t defined for t 2 Œ0; 1, say, and such that the
r.v.’s Xt are identically distributed centered and square integrable and Xt and Xs are
independent for every s 6D t.
In this exercise we prove that a white noise cannot be a measurable process,
unless it is 0 a.s. Let therefore .Xt /t be a measurable white noise.
a) Prove that, for every a; b 2 Œ0; 1, a b,
Z b Z b
E.Xs Xt / ds dt D 0 : (2.4)
a a
b) Show that for every a; b 2 Œ0; 1

h Z b 2 i
E Xs ds D0:
a
c) Deduce that necessarily Xt .!/ D 0 a.e. in t for almost every !.

2.6 (p. 454) Let .˝; F ; .Ft /t ; .Zt /t ; P/ be an Rm -valued continuous process.
Let Z W ! ! ft ! Zt .!/g be the application associating to every ! 2 ˝
the corresponding path of the process. Z is a map with values in the space
C D C .RC ; Rm /. Let us consider on C the topology of uniform convergence on
compact sets and denote by M the associated Borel -algebra. We want to prove
that Z is measurable.
a) Let, for 2 C , t > 0, " > 0, A;t;" D fw 2 C I jt wt j "g. Prove that
1
Z .A;t;" / 2 F .
b) For 2 C ; T > 0; " > 0, let
U ;T;" D fw 2 C I jt wt j " for every t 2 Œ0; Tg (2.5)
1
(a closed tube of radius " around the path ). Prove that Z .U ;T;" / 2 F.
c) Prove that Z is a .C ; M /-valued r.v.
b) As the paths of C are continuous, U ;T;" D fw 2 C I jr wr j " for every r 2 Œ0; T \ Qg,
which is a countable intersection of events of the form A;t;" . c) Recall Exercise 1.19.
Chapter 3
Brownian Motion
Brownian motion is a particular stochastic process which is the prototype of the

class of processes which will be our main concern. Its investigation is the object of
this chapter.
3.1 Definition and general facts
We already know from the previous chapter the definition of a Brownian motion.
Definition 3.1 A real-valued process B D .˝; F ; .Ft /t0 ; .Bt /t0 ; P/ is a

Brownian motion if
i) B0 D 0 a.s.;
ii) for every 0 s t the r.v. Bt Bs is independent of Fs ;
iii) for every 0 s t Bt Bs has law N.0; t s/.
Remarks 3.1
a) ii) of Definition 3.1 implies that Bt Bs is independent of Bu for every
u s and even from .Bu ; u s/, which is a -algebra that is contained
in Fs . Intuitively this means that the increments of the process after time
s are independent of the path of the process up to time s.
b) A Brownian motion is a Gaussian process, i.e. the joint distributions of
Bt1 ; : : : ; Btm are Gaussian. Let ˛1 ; : : : ; ˛m 2 R, 0 t1 < t2 < < tm : we
must prove that ˛1 Bt1 C C ˛m Btm is a normal r.v., so that we can apply
Proposition 1.8. This is obvious if m D 1, as Definition 3.1 with s D 0
(continued )

46 3 Brownian Motion
Remarks 3.1 (continued)

states that Bt N.0; t/. Let us assume this fact true for m 1 and let us
prove it for m: we can write
˛1 Bt1 C C˛m Btm D Œ˛1 Bt1 C C.˛m1 C˛m /Btm1 C˛m .Btm Btm1 / :
This is a normal r.v., as we have seen in Sect. 1.7, being the sum of
two independent normal r.v.’s (the r.v. between Œ is Ftm1 -measurable
whereas Btm Btm1 is independent of Ftm1 , thanks to ii) of Definition 3.1).
c) For every 0 t0 < < tm the real r.v.’s Btk Btk1 ; k D 1; : : : ; m, are
independent: they are actually jointly Gaussian and pairwise uncorrelated.
d) Sometimes it will be important to specify with respect to which filtration
a Brownian motion is considered. When the probability space is fixed we
shall say that B is an .Ft /t -Brownian motion in order to specify that B D
.˝; F ; .Ft /t ; .Bt /t ; P/ is a Brownian motion. Of course, for every t the
-algebra Ft must necessarily contain the -algebra Gt D .Bs ; s t/
(otherwise .Bt /t would not be adapted to .Ft /t ). It is also clear that if B is
an .Ft /t -Brownian motion it is a fortiori a Brownian motion with respect
to every other filtration .F 0t /t that is smaller than .Ft /t , (i.e. such that
Ft0 Ft for every t 0) provided that B is adapted to .F 0t /t i.e provided
that .F 0t /t contains the natural filtration (see p. 31). Actually if Bt Bs is
independent of Fs , a fortiori it will be independent of Fs0 .
We shall speak of natural Brownian motion when .Ft /t is the natural
filtration.
Proposition 3.1 If B D .˝; F ; .Ft /t ; .Bt /t ; P/ is a Brownian motion then

1) B0 D 0 a.s.;
2) for every 0 t1 < < tm , .Bt1 ; : : : ; Btm / is an m-dimensional centered
normal r.v.;
3) EŒBs Bt D s ^ t.
Conversely, properties 1), 2) and 3) imply that B is a natural Brownian
motion.
Proof If B is a Brownian motion, 1) is obvious; 2) is Remark 3.1 b). As for 3), if

s t, as Bt Bs and Bs are centered and independent
EŒBt Bs D EŒf.Bt Bs / C Bs gBs D EŒ.Bt Bs /Bs C EŒB2s D s D s ^ t :
Conversely, if B satisfies 1), 2) and 3), then i) of Definition 3.1 is obvious. Moreover,
for 0 s < t, Bt Bs is a normal r.v., being a linear function of .Bs ; Bt /, and is
3.1 Definition and general facts 47
centered since both Bt and Bs are so; as, for s t,
EŒ.Bt Bs /2 D EŒB2t C EŒB2s 2EŒBt Bs D t C s 2s D t s ;
Bt Bs has an N.0; t s/ distribution and iii) is satisfied. To complete the proof

we must show that Bt Bs is independent of .Bu ; u s/; Brownian motion being
a Gaussian process, thanks to Remark 1.2 it is sufficient to prove that Bt Bs is
uncorrelated with Bu for every u s. Actually, if u s t,
EŒ.Bt Bs /Bu D t ^ u s ^ u D 0
and Bt Bs is therefore independent of .Bu ; u s/. t

u
We have not yet proved the existence of a process satisfying the conditions of
Definition 3.1, but this is a simple application of Remark 2.5. Actually it is sufficient
to verify that, if t1 t2 tm ,
K.ti ; tj / D ti ^ tj
is a positive definite kernel, i.e. that the matrix with entries ij D ti ^ tj is positive
definite. The simplest way to check this fact is to produce a r.v. having as a
covariance matrix, every covariance matrix being positive definite as pointed out on
p. 10.
Let Z1 ; : : : ; Zm be independent centered Gaussian r.v.’s with Var.Zi / D ti ti1 ,
with the understanding that t0 D 0. Then it is immediate that the r.v. .X1 ; : : : ; Xm /
with Xi D Z1 C C Zi has covariance matrix : as the r.v.’s Zk are independent we
have, for i j,
Cov.Xi ; Xj / D Cov.Z1 C C Zi ; Z1 C C Zj / D Var.Z1 / C C Var.Zi /

D t1 C .t2 t1 / C C .ti ti1 / D ti D ti ^ tj :
The next statement points out that Brownian motion is invariant with respect to
certain transformations.
Proposition 3.2 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion.

Then
Xt D BtCs Bs ; Bt ; cBt=c2 ;

(
tB1=t t>0
Zt D
0 tD0
are also Brownian motions, the first one with respect to the filtration .FtCs /t ,
the second one with respect to .Ft /t , and the third one with respect to .Ft=c2 /t .
.Zt /t is a natural Brownian motion.
Proof Let us prove that X is a Brownian motion. The condition X0 D 0 is immediate

as is the fact that, if v < u, then Xu Xv D BuCs Bs BvCs C Bs D BuCs BvCs
is N.0; u v/-distributed. This also gives that the increment Xu Xv is independent
of FvCs , so that X is an .FtCs /t -Brownian motion.
Let us now prove that Z is a Brownian motion, the other situations being similar
to the case of X. The proof amounts to checking that the three conditions of
Proposition 3.1 are satisfied.
1) of Proposition 3.1 is immediate, 2) follows from the fact that Z is also a
Gaussian family; actually .Zt1 ; : : : ; Ztm / can be written as a linear transformation
of .B1=t1 ; : : : ; B1=tm /, which is Gaussian. Finally, if s < t then 1s ^ 1t D 1t and
EŒZs Zt D st EŒB1=s B1=t D st . 1s ^ 1t / D s D s ^ t
and therefore 3) is satisfied. t

u
We shall speak of the invariance properties of Proposition 3.2 as the scaling
properties of the Brownian motion.
Definition 3.2 An Rm -valued process X D .˝; F ; .Ft /t ; .Xt /t ; P/ is an m-

dimensional Brownian motion if
a) X0 D 0 a.s.;
b) for every 0 s t, the r.v. Xt Xs is independent of Fs ;
c) for every 0 s t, Xt Xs is N.0; .t s/I/-distributed (I is the m m
identity matrix).
Remarks 3.2 Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be an m-dimensional Brownian

motion.
a) The real processes .˝; F ; .Ft /t ; .Xi .t//t ; P/ are Brownian motions for i D
1; : : : ; m. Definition 3.1 is immediately verified.
b) The -algebras .Xi .t/; t 0/, i D 1; : : : ; m, are independent. Keeping
in mind Remark 1.2 c), it is sufficient to check that EŒXi .t/Xj .s/ D 0, for
every s t, i 6D j. But
EŒXi .t/Xj .s/ D EŒ.Xi .t/ Xi .s//Xj .s/ C EŒXi .s/Xj .s/
and the first term on the right-hand side vanishes, Xj .s/ and Xi .t/ Xi .s/
being independent and centered, the second one vanishes too as the
covariance matrix of Xs is diagonal.
Therefore the components .Xi .t//t , i D 1; : : : ; m, of an m-dimensional
Brownian motion are independent real Brownian motions.
Example 3.1 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian

motion. Let us prove that if z 2 Rm and jzj D 1, then the process Xt D hz; Bt i
is a real Brownian motion.
Since, for t s, Bt Bs is independent of Fs , the r.v.
Xt Xs D hz; Bt i hz; Bs i D hz; Bt Bs i
is itself independent of Fs . Moreover, Xt Xs D hz; Bt Bs i is Gaussian,

being a linear combination of B1 .t/ B1 .s/; : : : ; Bm .t/ Bm .s/, which are
jointly Gaussian. The variance of Xt Xs is computed immediately, being a
linear function of Bt Bs : (1.13) gives
Var.Xt Xs / D .t s/z z D .t s/jzj2 D t s :
As a) of Definition 3.1 is obvious, this proves that X is a Brownian motion

with respect to the filtration .Ft /t . Conditions 1), 2), 3) of Proposition 3.1 are
also easily checked. Note. however, that these only ensure that X is a natural
Brownian motion, which is a weaker result.
This, similarly to Proposition 3.2, is a typical exercise: given a Brownian
motion B, prove that some transformation of it gives rise to a Brownian
motion. By now we have the choice of checking that the conditions of
Definition 3.1 or of Proposition 3.1 are satisfied. Note that Proposition 3.1
is useful only if we have to prove that our candidate process is a natural
Brownian motion. Later (starting from Chap. 8) we shall learn more powerful
methods in order to prove that a given process is a Brownian motion.
Of course a) of Remark 3.2 is a particular case of this example: just choose
as z the unit vector along the i-th axis.
We already know (Example 2.4) that a Brownian motion has a continuous mod-
ification. Note that the argument of Example 2.4 also works for an m-dimensional
Brownian motion. From now on, by “Brownian motion” we shall always understand
a Brownian motion that is continuous. Figure 3.1 provides a typical example of a
path of a two-dimensional Brownian motion.
If we go back to Proposition 3.2, if B is a continuous Brownian motion, then
the “new” Brownian motions X, B, .cBt=c2 /t are also obviously continuous. For Z
instead a proof is needed in order to have continuity at 0. In order to do this, note
that the processes .Bt /t and .Zt /t are equivalent. Therefore, as B is assumed to be
continuous,
lim Bt D 0 ;
t!0C
.8
.4
−.4
−1 0
Fig. 3.1 A typical image of a path of a two-dimensional Brownian motion for 0 t 1 (a black
small circle denotes the origin and the position at time 1). For information about the simulation of
Brownian motion see Sect. 3.7
therefore (see Exercise 2.4 b2))
lim Zt D 0
t!0C;t2Q
from which, the process .Zt /t;t>0 being continuous, we derive
lim Zt D 0 :
t!0C
It will be apparent in the sequel that it is sometimes important to specify the filtration
with respect to which a process B is a Brownian motion. The following remark
points out a particularly important typical filtration. Exercise 3.5 deals with a similar
question.
Remark 3.1 (The augmented natural filtration) Let B D .˝; F ; .Ft /t ;

.Bt /t ; P/ be a Brownian motion and let F t be the -algebra that is obtained
by adding to Ft the negligible events of F . Let us prove that B is a Brownian
motion also with respect to .F t /t .
Actually, we must only prove that the increment Bt Bs , s t, is
independent of F s . Let us denote by N the family of the negligible events
of F . Let C be the class of the events of the form A \ G with A 2 Fs
and G 2 N or G D ˝. C is stable with respect to finite intersections and
(continued )

contains Fs (one chooses G D ˝) and N , and therefore generates F s . By
Remark 1.1, in order to prove that Bt Bs and F s are independent we just
have to check that, for every Borel set C R and A and G as above,
P.fBt Bs 2 Cg \ A \ G/ D P.fBt Bs 2 Cg/P.A \ G/ ;
which is immediate. If G D ˝ the relation above is true because then A\G D

A and B is an .Ft /t -Brownian motion, whereas if G 2 N , both members are
equal to 0.
In particular, B is a Brownian motion with respect to the augmented natural
filtration, i.e. with respect to the filtration .G t /t that is obtained by adding the
negligible events of F to the natural filtration .Gt /t .
We shall see in Sect. 4.5 that .G t /t is also right-continuous, so that the
Brownian motion with respect to the natural augmented filtration is a standard
process (see the definition on p. 34).
Remark 3.2 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. We

know that each of the increments Bt Bs , t > s, is independent of Fs . Actually
a stronger statement holds: the -algebra .Bt Bs ; t s/ is independent
of Fs .
This fact is almost obvious if Fs D .Bu ; u s/, i.e. if B is a natural
Brownian motion, as the families of r.v.’s fBu ; u sg and fBv Bs ; v sg
are jointly Gaussian and pairwise non-correlated, so that we can apply the
criterion of Remark 1.2. It requires some care if the filtration is not the natural
one. The proof is left as an exercise (Exercise 3.4).
Note that this is a useful property. For instance, it allows us to state that the
r.v. .Bt1 Bs / C .Bt2 Bs /, t1 ; t2 > s, is independent of Fs . Recall that if X; Y
are r.v.’s that are independent of a -algebra G , it is not true in general that
their sum is also independent of G . Here, however, we can argue that .Bt1
Bs / C .Bt2 Bs / is .Bt Bs ; t s/-measurable and therefore independent of
Fs . Similarly, if b > s, the r.v.
Z b
YD .Bu Bs / du
s
for t > s is independent of Fs : writing the integral as the limit of its Riemann
sums, it is immediate that Y is .Bt Bs ; t s/-measurable and therefore
independent of Fs .
Remark 3.3 Computations concerning Brownian motion repeatedly require a

certain set of formulas typical of Gaussian
p distributions. Let us recall them
(they are all based on the relation Bt t Z with Z N.0; 1/):
1 2
a/ EŒe
Bt D e 2 t
I
(
2 p 1 if t
< 1
2
b/ EŒe
Bt D 12t
1
C1 if t
2
a) (the Laplace transform) is computed in Exercise 1.6, b) in Exercise 1.12.
3.2 The law of a continuous process, Wiener measure
Let D .˝; F ; .Ft /t ; .t /t ; P/ be a continuous Rd -valued process and let W ! 7!

ft 7! t .!/g be the map mapping every ! to its associated path. This map takes its
values in C D C .RC ; Rd /, the space of continuous maps RC ! Rd , endowed
with the topology of uniform convergence on compact sets. Let us denote by M the
corresponding Borel -algebra on C .
Proposition 3.3 W .˝; F / ! .C ; M / is measurable.
The proof of Proposition 3.3 is rather straightforward and we shall skip it (see,
however, Exercise 2.6).
Proposition 3.3 authorizes us to consider on the space .C ; M / the image
probability of P through , called the law of the process . The law of a continuous
process is therefore a probability on the space C of continuous paths.
Let us denote by P the law of a process . If we consider the coordinate r.v.’s
Xt W C ! Rd defined as Xt . / D .t/ (recall that 2 C is a continuous function)
and define Mt D .Xs ; s t/, then
X D .C ; M ; .Mt /t ; .Xt /t ; P /
is itself a stochastic process. By construction this new process has the same finite-
dimensional distributions as . Let A1 ; : : : ; Am 2 B.Rd /, then it is immediate that
1

ft1 2 A1 ; : : : ; tm 2 Am g D Xt1 2 A1 ; : : : ; Xtm 2 Am
3.3 Regularity of the paths 53
so that

P Xt1 2 A1 ; : : : ; Xtm 2 Am D P 1 Xt1 2 A1 ; : : : ; Xtm 2 Am

D P t1 2 A1 ; : : : ; tm 2 Am
and the two processes and X have the same finite-dimensional distributions and are
0
equivalent. This also implies that if and 0 are equivalent processes, then P D P ,
i.e. they have the same law.
In particular, given two (continuous) Brownian motions, they have the same law.
Let us denote this law by PW (recall that this a probability on C ). PW is the Wiener
measure and the process X D .C ; M ; .Mt /t ; .Xt /t ; PW /, having the same finite-
dimensional distributions, is also a Brownian motion: it is the canonical Brownian
motion.
3.3 Regularity of the paths
We have seen that a Brownian motion always admits a continuous version which is,
moreover, -Hölder continuous for every < 12 . It is possible to provide a better
description of the regularity of the paths, in particular showing that, in some sense,
this estimate cannot be improved.
From now on X D .˝; F ; .Ft /t ; .Xt /t ; P/ will denote a (continuous) Brownian
motion. Let us recall that if I R is an interval and f W I ! R is a continuous
function, its modulus of continuity is the function
w.ı/ D sup j f .x/ f . y/j :

x;y2I;jxyjı
The regularity of a function is determined by the behavior of its modulus of

continuity as ı ! 0C. In particular,
• f is uniformly continuous if and only if limı!0C w.ı/ D 0.
• f is Hölder continuous with exponent ˛ if and only if w.ı/ cı ˛ for some c > 0.
• f is Lipschitz continuous if and only if w.ı/ cı for some c > 0.
Theorem 3.1 (P. Lévy’s modulus of continuity) Let X be a real Brownian

motion. For every T > 0
1
P lim sup jXt Xs j D 1 D 1 :
ı!0C 2ı log 1 1=2 0s<tT
ı tsı
We skip the proof of Theorem 3.1, which is somewhat similar to the proof of the
Iterated Logarithm Law, Theorem 3.2, that we shall see soon.
P. Lévy’s theorem asserts that if w.; !/ is the modulus of continuity of Xt .!/ for
t 2 Œ0; T, then P-a.s.
w.ı; !/
lim 1=2 D 1 :
ı!0C 2ı log 1ı
Note that this relation holds for every ! a.s. and does not depend on T.
As w.ı/
ı 1=2
! C1 as ı ! 0C, Theorem 3.1 specifies that the paths of a Brownian
motion cannot be Hölder continuous of exponent 12 on the interval Œ0; T for every T
(Fig. 3.2). More precisely
Corollary 3.1 Outside a set of probability 0 no path is Hölder continuous

with exponent 12 in any time interval I RC having non-empty interior.
Proof It suffices to observe that for q; r 2 QC , 0 q < r,
jXt Xs j
lim sup 1=2 D 1 a.s. (3.1)
ı!0C qs<tr
tsı
2ı log 1ı
thanks to Theorem 3.1 applied to the Brownian motion .XtCq SXq /t . Therefore if
Nq;r is the negligible event on which (3.1) is not satisfied, N D q;r2QC Nq;r is still
negligible. Since an interval I RC having non-empty interior necessarily contains
1.6
1.2
.8
.4
−.4
0 1
Fig. 3.2 Example of the path of a real Brownian motion for 0 t 1 (here the x axis represents
time). As in Fig. 3.1, the lack of regularity is evident as well as the typical oscillatory behavior,
which will be better understood with the help of Theorem 3.2
3.3 Regularity of the paths 55
an interval of the form Œq; r with q < r, q; r 2 QC , no path outside N can be Hölder
continuous with exponent 12 in any time interval I RC having non-empty
interior.
t
u
Let us recall that, given a function f W R ! R, its variation in the interval Œa; b is
the quantity
X
n
ˇ ˇ
Vba f D sup ˇ f .tiC1 / f .ti /ˇ ;

iD1
the supremum being taken among all finite partitions a D t0 < t1 < < tnC1 D b
of the interval Œa; b. f is said to have finite variation if Vba f < C1 for every a,
b 2 R.
Note that a Lipschitz continuous function f is certainly with finite variation: if
we denote by L the Lipschitz constant of f then
X
n
ˇ ˇ X n X
n
ˇ f .tiC1 / f .ti /ˇ L jtiC1 ti j D L .tiC1 ti / D L.b a/ :
iD1 iD1 iD1
Therefore differentiable functions have finite variation on every bounded interval. It

is also immediate that monotone functions have finite variation.
Proposition 3.4 Let X be a real Brownian motion. Let D ft0 ; : : : ; tm g

with s D t0 < t1 < tm D t be a partition of the interval Œs; t, jj D
max0km1 jtkC1 tk j. Then if
X
m1
S D jXtkC1 Xtk j2
kD0
we have
lim S D t s in L2 : (3.2)
jj!0C
In particular, the paths of a Brownian motion do not have finite variation in

any time interval a.s.
Pm1
Proof We have kD0 .tkC1 tk / D .t1 s/ C .t2 t1 / C C .t tm1 / D t s
so we can write
X
m1

S .t s/ D .XtkC1 Xtk /2 .tkC1 tk / :

kD0
We must prove that EŒ.S .t s//2 ! 0 as jj ! 0. Note that .XtkC1 Xtk /2
.tkC1 tk / are independent (the increments of a Brownian motion over disjoint
intervals are independent) and centered; therefore, if h 6D k, the expectation of the
product

.XthC1 Xth /2 .thC1 th / .XtkC1 Xtk /2 .tkC1 tk /
vanishes so that
EŒ.S .t s//2
m1
X X

m1

DE .XtkC1 Xtk /2 .tkC1 tk / .XthC1 Xth /2 .thC1 th /
kD0 hD0
X
m1

2 X
m1 h .Xt Xtk /2 2 i
D E .XtkC1 Xtk /2 .tkC1 tk / D .tkC1 tk /2 E kC1
1 :
kD0 kD0
tkC1 tk
XtkC1 Xtk
But for every k the r.v. p
tkC1 tk is N.0; 1/-distributed and the quantities
h .Xt Xtk /2 2 i
cDE kC1
1
tkC1 tk
are finite and do not depend on k (c D 2, if you really want to compute it. . . ).
Therefore, as jj ! 0,
X
m1 X
m1
EŒ.S .t s//2 D c .tkC1 tk /2 cjj jtkC1 tk j D cjj.t s/ ! 0 ;
kD0 kD0
which proves (3.2). Moreover,
X
m1
ˇ ˇ2 ˇ Xˇ
ˇ m1 ˇ
S D ˇX t Xtk max XtiC1 Xti ˇ
ˇ ˇ ˇX t X t ˇ : (3.3)
kC1 kC1 k
0im1
kD0 kD0
3.4 Asymptotics 57
As the paths are continuous, max0im1 jXtiC1 Xti j ! 0 as jj ! 0C and

therefore if the paths were with finite variation on Œs; t for ! in some event A of
positive probability, then on A we would have
X
m1
ˇ ˇ
lim ˇX t Xtk ˇ < C1
kC1
jj!0C
kD0
and therefore, taking the limit in (3.3), we would have limjj!0C S .!/ D 0 on A,
in contradiction with the first part of the statement. t
u
Let us recall that if f has finite variation, then it is possible to define the integral
Z T
.t/ df .t/
0
for every bounded Borel function . Later we shall need to define an integral of the
type
Z T
.t/ dXt .!/ ;
0
which will be a key tool for the construction of new processes starting from
Brownian motion. Proposition 3.4 states that this cannot be done ! by !, as the
paths of a Brownian motion do not have finite variation. In order to perform this
program we shall construct an ad hoc integral (the stochastic integral).
3.4 Asymptotics
We now present a classical result that gives very useful information concerning the
behavior of the paths of a Brownian motion as t ! 0C and as t ! C1.
Theorem 3.2 (Iterated logarithm law) Let X be a Brownian motion. Then

Xt
P lim D 1 D1: (3.4)
t!0C 2t log log 1 1=2
t
Before giving the proof, let us point out some consequences.

Corollary 3.2
Xt
lim 1=2 D 1 a.s. (3.5)
t!0C 2t log log 1t
Xs
lim 1=2 D 1 a.s. (3.6)
s!C1 2s log log s
Xs
lim 1=2 D 1 a.s. (3.7)
s!C1 2s log log s
Proof Let us prove (3.6). We know from Proposition 3.2 that Zt D tX1=t is a
Brownian motion. Theorem 3.2 applied to this Brownian motion gives
tX1=t
lim 1=2 D 1 a:s:
t!0C 2t log log 1t
Now, with the change of variable s D 1t ,
tX1=t p X1=t Xs
lim 1=2
D lim t 1=2
D lim
t!0C 2t log log 1 t!0C 2 log log 1t s!C1 2s log log s 1=2
t
Similarly, (3.5) and (3.7) follow from Theorem 3.2 applied to the Brownian motions
X, and .tX1=t /t . t
u
Remark 3.4 (3.6) and (3.7) give important information concerning the
asymptotic of the Brownian motion as t ! C1. Indeed they imply
the existence of two sequences of times .tn /n , .sn /n , with limn!1 tn D
limn!1 sn D C1 and such that
p
Xtn .1 "/ 2tn log log tn
p
Xsn .1 "/ 2sn log log sn :
This means that, as t ! C1, the Brownian motion takes arbitrarily large
positive and negative values infinitely many times. It therefore exhibits larger
and larger oscillations. As the paths are continuous, in particular, it visits
every real number infinitely many times.
(3.6) and (3.7) also give a bound on how fast a Brownian motion
moves away from the origin. In particular, (3.6) implies that, for t large,
(continued )
3.4 Asymptotics 59

p
Xt .!/ p.1 C "/ 2t log log t. Similarly, by (3.7), for t large, Xt .!/
.1 C "/ 2t log log t, so that, still for large t,
p
jXt .!/j .1 C "/ 2t log log t : (3.8)
To be precise, there exists a t0 D t0 .!/ such that (3.8) holds for every t t0 .
Similarly, by (3.4) and (3.5), there exist two sequences .tn /n ; .sn /n decreas-
ing to 0 and such that a.s. for every n,
q
Xtn .1 "/ 2tn log log t1n
q
Xsn .1 "/ 2sn log log s1n
In particular, Xsn < 0 < Xtn . By the intermediate value theorem the path
t 7! Xt crosses 0 infinitely many times in the time interval Œ0; " for every
" > 0. This gives a hint concerning the oscillatory behavior of the Brownian
motion.
In order to prove Theorem 3.2 we need some preliminary estimates.
Lemma 3.1 Let X be a continuous Brownian motion. If x > 0, T > 0, then

P sup Xt > x 2P.XT > x/ :
0tT
Proof Let t0 < t1 < < tn D T, I D ft0 ; : : : ; tn g and let D inff jI Xtj > xg. Note
that if XT .!/ > x, then T, i.e. fXT > xg f Tg. Moreover, we have Xtj x
on f D tj g. Hence
X
n X
n
P.XT > x/ D P. T; XT > x/ D P. D tj ; XT > x/ P. D tj ; XT Xtj 0/ :
jD0 jD0
We have f D tj g D fXt1 x; : : : ; Xtj1 x; Xtj > xg. Hence f D tj g is Ftj -

measurable and independent of fXT Xtj 0g; moreover, P.XT Xtj 0/ D 12 , as
XT Xtj is Gaussian and centered; therefore
X 1X 1
n n
P.XT > x/ P. D tj /P.XT Xtj 0/ D P. D tj / D P sup Xt > x :
jD0
2 jD0 2 t2I
Let now In be a sequence of finite subsets of Œ0; T increasing to Œ0; T \ Q. By the

previous inequality and taking the limit on the increasing sequence of events

P sup Xt > x D sup P sup Xt > x 2P.XT > x/ :
tT;t2Q n t2In
As the paths are continuous, suptT;t2Q Xt D suptT Xt , and the statement is proved.
t
u
Lemma 3.2 If x > 0 then

Z C1
1 1 x2 =2 2 1 2
xC e ez =2 dz ex =2 :
x x x
Proof We have
Z C1 Z C1
2 =2 1 2 =2 1 x2 =2
ez dz z ez dz D e
x x x x
and the inequality on the right-hand side is proved. Moreover,
d 1 x2 =2 1 2
e D 1 C 2 ex =2
dx x x
and therefore
Z Z
1 x2 =2 C1
1 2 1 C1 z2 =2
e D 1 C 2 ez =2 dz 1 C 2 e dz :
x x z x x
t
u
Proof of Theorem 3.2 Let us prove first that
Xt
lim 1=2 1 a.s. (3.9)
t!0C 2t log log 1t
1=2
Let .t/ D 2t log log 1t . Let .tn /n be a sequence decreasing to 0, let ı > 0 and
consider the event
˚
An D Xt > .1 C ı/.t/ for some t 2 ŒtnC1 ; tn :
ı being arbitrary, (3.9) follows if we can prove that P.limn!1 An / D 0. Indeed,

recall that limn!1 An is the event that is formed by those !’s that belong to infinitely
3.4 Asymptotics 61
many An . If this set has probability 0, this means that Xt > .1 C ı/.t/ for some t 2
ŒtnC1 ; tn only for finitely many n and therefore that limt!0C .2t log log
Xt
1 1=2 .1 C ı/,
/
t
and ı being arbitrary, this implies (3.9). P
By the Borel–Cantelli lemma it suffices to prove that the series 1 nD1 P.An /
is convergent; we need, therefore, a good upper bound for P.An /. First, as is
increasing,
n o
An sup Xt > .1 C ı/.tnC1 / ;
0ttn
p
and by Lemmas 3.1 and 3.2, as Xtn = tn N.0; 1/,

P.An / D P sup Xt .1 C ı/.tnC1 / 2P Xtn .1 C ı/.tnC1 /
0ttn
X t 1 1=2
t nC1
D 2P p n .1 C ı/ 2 log log
tn tn tnC1
Z C1 r
2 2 2 1 x2n =2
D p ez =2 dz e ;
2 xn xn
t
1 1=2
where xn D .1 C ı/ 2 nC1
tn log log tnC1 . Let us choose now tn D qn with 0 < q <
1, but such that D q.1 C ı/2 > 1. Now if we write ˛ D log 1q > 0, then

1=2
1=2
xn D .1 C ı/ 2q log .n C 1/ log 1q D 2 log.˛.n C 1// :
Therefore
r r
2 1 x2n =2 2 log.˛.nC1// c
P.An / e e D
xn .n C 1/
P1 > 1, the rightmost term is the general term of a convergent series, hence
As
nD1 P.An / < C1 and, by the Borel–Cantelli Lemma, P.limn!1 An / D 0, which
completes the proof of (3.9).
Let us prove now the reverse inequality of (3.9). This will require the use of
the converse part of the Borel–Cantelli Lemma, which holds under the additional
assumption that the events involved are independent. For this reason, we shall first
investigate the behavior of the increments of the Brownian motion. Let again .tn /n
be a sequence decreasing to 0 and let Zn D Xtn XtnC1 . The r.v.’s Zn are independent,
being the increments of X. Then for every x > 1, " > 0, we have
p Xt Xt
P Zn > x tn tnC1 D P p >x
n nC1
Z C1 tn tnC1
(3.10)
1 z2 =2 x 1 2 1 2
Dp e dz 2 p ex =2 p ex =2 ;
2 x x C 1 2 2x 2
where we have taken advantage of the left-hand side inequality of Lemma 3.2. Let
2
tn D qn with 0 < q < 1 and put ˇ D 2.1"/
1q
, ˛ D log 1q . Then
.tn / 1" q
x D .1 "/ p Dp 2 log n log 1q
tn tnC1 1q
s
2.1 "/2 p
D log n log 1q D ˇ log.˛n/
1q
so that (3.10) becomes

c
P.Zn > .1 "/.tn // p
nˇ=2 log n
We can choose q small enough so that ˇ < 2 and the left-hand side becomes the
general term of a divergent series. Moreover, as the r.v.’s Zn are independent, these
events are independent themselves and by the Borel–Cantelli lemma we obtain
Zn > .1 "/.tn / infinitely many times a.s.
On the other hand the upper bound (3.9), which has already been proved, applied to
the Brownian motion X implies that a.s. we have eventually
Xtn > .1 C "/.tn / :
Putting these two relations together we have that a.s. for infinitely many indices n
Xtn D Zn C XtnC1 > .1 "/.tn / .1 C "/.tnC1 /

h .tnC1 / i (3.11)
D .tn / 1 " .1 C "/ :
.tn /
Note that, as log log q1n D log n C log log 1q and limn!1 log.nC1/
log n D 1,
q q
1 1
.tnC1 / 2qnC1 log log qnC1 p log log qnC1 p
lim D lim q D q lim q D q:
n!1 .tn / n!1 1
2qn log log qn n!1 1
log log qn
For every fixed ı > 0 we can choose " > 0 and q > 0 small enough so that
p
1 " .1 C "/ q > 1 ı ;
hence from (3.11)
Xtn > .1 ı/.tn / for infinitely many times a.s.
which completes the proof of the theorem. t

u
3.5 Stopping times 63
3.5 Stopping times
Sometimes, in the investigation of a stochastic process X, we shall consider the value

of X at some random time , i.e. of the r.v. X .!/ .!/. A random time is, in general,
just a r.v. W ˝ ! Œ0; C1Œ. Among random times the stopping times of the next
definition will play a particular role.
Definition 3.3 Let .Ft /t2T be a filtration. A r.v. W ˝ ! T [fC1g is said to

be a stopping time if, for every t 2 T, f tg 2 Ft . Associated to a stopping
time let us define the -algebra
F D fA 2 F1 ; A \ f tg 2 Ft for every t 2 Tg
W
where, as usual, F1 D t Ft .
Note that, in general, a stopping time is allowed to take the value C1. Intuitively
the condition f tg 2 Ft means that at time t we should be able to say whether
t or not. For instance. we shall see that the first time at which a Brownian
motion B comes out of an open set D is a stopping time. Intuitively, at time t we
know the values of Bs for s t and we are therefore able to say whether Bs 62 D for
some s t. Conversely, the last time of visit of B to an open set D is not a stopping
time as in order to say whether some time t is actually the last time of visit we also
need to know the positions of Bs for s > t.
F is, intuitively, the -algebra of the events for which at time we can say
whether they are satisfied or not. The following proposition summarizes some
elementary properties of stopping times. The proof is a straightforward application
of the definitions and it is suggested to do it as an exercise (looking at the actual
proof only later).
Proposition 3.5 Let and be stopping times. Then

a) is F -measurable.
b) _ and ^ are stopping times.
c) If , then F F .
d) F ^ D F \ F .
Proof
a) By Exercise 1.19, we just have to prove that, for every s 0, f sg 2 F . It
is obvious that f sg 2 Fs F1 . We then have to check that, for every t,
f sg\f tg 2 Ft . Now, if t s, we have f sg\f tg D f tg 2 Ft .

If, conversely, t > s, f sg \ f tg D f sg 2 Fs Ft .
b) We have
f ^ tg D f tg [ f tg 2 Ft :
and therefore ^ is a stopping time. The argument for _ is similar:
f _ tg D f tg \ f tg 2 Ft :
c) If A 2 F then, for every t, A \ f tg 2 Ft . As f tg f tg,
A \ f tg D A \ f tg \ f tg 2 Ft :
„ ƒ‚ … „ ƒ‚ …
2Ft 2Ft
d) Thanks to c), F ^ is contained both in F and in F . Let us prove the opposite

inclusion. Let A 2 F \ F ; then A 2 F1 , A \ f tg 2 Ft and A \ f
tg 2 Ft . We have therefore
A \ f ^ tg D A \ .f tg [ f tg/ D .A \ f tg/ [ .A \ f tg/ 2 Ft
and therefore A 2 F ^ . t
u
Note that, in particular, if t 2 RC then t is a (deterministic) stopping time.
Therefore if is a stopping time, by Proposition 3.5 b), ^ t is also a stopping time.
It is actually a bounded stopping time, even if is not. We shall use this fact very
often when dealing with unbounded stopping times.
Proposition 3.6 If is an a.s. finite stopping time and X D .˝; F ;

.Ft /t ; .Xt /t ; P/ is a progressively measurable process, then the r.v. X is F -
measurable.
Proof Let us assume that X takes its values in some measurable space .E; E /. We
must prove that, for every 2 E , fX 2 g 2 F . We know already (Example 2.3)
that X is F1 -measurable, so that fX 2 g 2 F1 .
Recalling the definition of the -algebra F , we must now prove that, for every
t, f tg \fX 2 g 2 Ft . Of course f tg \fX 2 g D f tg \fX ^t 2 g.
The r.v. ^ t is Ft -measurable: the event f ^ t sg is equal to ˝ if s t and
to f sg if s < t and belongs to Ft in both cases.
The r.v. X ^t then turns out to be Ft -measurable as the composition of the maps
! 7! .!; ^t.!//, which is measurable from .˝; Ft / to .˝ Œ0; t; Ft ˝B.Œ0; t//,
and .!; u/ 7! Xu .!/, from .˝ Œ0; t; Ft ˝ B.Œ0; t// to .E; B.E//, which is
3.5 Stopping times 65
Fig. 3.3 The exit time of a two-dimensional Brownian motion B from the unit ball. The three
black small circles denote the origin, the position at time 1 and the exit position B
measurable thanks to the assumption of progressive measurability. Hence f

tg \ fX 2 g 2 Ft as the intersection of Ft -measurable events.
t
u
If A 2 B.E/, let
A D infft 0I Xt … Ag :
A is called the exit time from A (Fig. 3.3). In this definition, as well as in other
similar situations, we shall always understand, unless otherwise indicated, that the
infimum of the empty set is equal to C1. Therefore A D C1 if Xt 2 A for every
t 0. Similarly the r.v.
A D infft 0I Xt 2 Ag
is the entrance time in A. It is clear that it coincides with the exit time from Ac . Are
exit times stopping times? Intuition suggests a positive answer, but we shall see that
some assumptions are required.
Proposition 3.7 Let E be a metric space and X an E-valued continuous

process.
a) If A is an open set, then A is a stopping time.
b) If F is a closed set and the filtration .Ft /t is right continuous, then F is a
stopping time.
Proof
a) If A is an open set, we have
1 \
[
fA > tg D fd.Xr ; Ac / > 1n g 2 Ft :
nD1 r2Q
r<t
Indeed, if ! belongs to the set on the right-hand side, then for some n we have
d.Xr .!/; Ac / > 1n for every r 2 Q \ Œ0; tŒ and, the paths being continuous, we
have d.Xs .!/; Ac / 1n for every s t. Therefore Xs .!/ 2 A for every s t and
A .!/ > t.
The opposite inclusion follows from the fact that if A .!/ > t then Xs .!/ 2 A
for every s t and hence d.Xs .!/; Ac / > 1n for some n and for every s t (the
image of Œ0; t through s 7! Xs .!/ is a compact subset of E and its distance from
the closed set Ac is therefore strictly positive).
Therefore fA tg D fA > tgc 2 Ft for every t.
b) Similarly, if F is closed,
\
fF tg D fXr 2 Fg 2 Ft ;
r2Q;r<t
so that fF < tg 2 Ft . Therefore

\
fF tg D fF < t C "g 2 FtC D Ft
">0
since we have assumed that the filtration is right-continuous.

t
u
3.6 The stopping theorem
Let X be a Brownian motion. In Proposition 3.2 we have seen that, for every
s 0, .XtCs Xs /t is also a Brownian motion. Moreover, it is immediate that it
is independent of Fs . The following result states that these properties remain true if
the deterministic time s is replaced by a stopping time .
Theorem 3.3 Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be an m-dimensional Brow-

nian motion and an a.s. finite stopping time of the filtration .Ft /t . Then
Yt D XtC X is a Brownian motion independent of F .
3.6 The stopping theorem 67
Proof Let us assume first that takes only a discrete set of values: s1 < s2 <
< sk < : : : Let C 2 F ; then, recalling the definition of F , we know that
C \ f sk g 2 Fsk . Actually also C \ f D sk g 2 Fsk , as
C \ f D sk g D .C \ f sk g/ n .C \ f sk1 g/
„ ƒ‚ …
2Fsk1 Fsk
and both events on the right-hand side belong to Fsk . Then, if A1 ; : : : ; An 2 B.Rm /
and C 2 F , we have
P. Yt1 2 A1 ; : : : ; Ytn 2 An ; C/
X
D P.Xt1 C X 2 A1 ; : : : ; Xtn C X 2 An ; D sk ; C/
k
X
D P.Xt1 Csk Xsk 2 A1 ; : : : ; Xtn Csk Xsk 2 An ; D sk ; C /
„ ƒ‚ … „ ƒ‚ …
k Fsk measurable
independent of Fsk
X
D P.Xt1 Csk Xsk 2 A1 ; : : : ; Xtn Csk Xsk 2 An /P. D sk ; C/ :
k
Now recall that .XtCsk Xsk /t is a Brownian motion for every k, so that the quantity
P.Xt1 Csk Xsk 2 A1 ; : : : ; Xtn Csk Xsk 2 An / does not depend on k and is equal to
P.Xt1 2 A1 ; : : : ; Xtn 2 An /. Hence
X
P. Yt1 2 A1 ; : : : ; Ytn 2 An ; C/ D P.Xt1 2 A1 ; : : : ; Xtn 2 An / P. D sk ; C/
k
D P.Xt1 2 A1 ; : : : ; Xtn 2 An /P.C/ :
Letting C D ˝ we have that Y is a Brownian motion (it has the same finite-dimen-
sional distributions as X); letting C 2 F we have that Y is independent of F .
In order to get rid of the assumption that takes a discrete set of values, we use
the following result, which will also be useful later.
Lemma 3.3 Let be a stopping time. Then there exists a non-increasing

sequence .n /n of stopping times each taking a discrete set of values and such
that n & . Moreover, fn D C1g D f D C1g for every n.
Proof Let
8
ˆ
<0
ˆ if .!/ D 0
n .!/ D kC1
if k
< .!/ kC1
ˆ 2n 2n 2n
:̂C1 if .!/ D C1 :
1
As n 2n
n , obviously n & . Moreover,
[ ˚ [ ˚k
fn tg D n D kC1
2n D f.!/ D 0g [ 2n < kC1
2n 2 Ft
k; kC1
2n t k; kC1
2n t
and therefore n is a stopping time. t

u
End of the proof of Theorem 3.3. If is a finite stopping time, let .n /n be a
sequence of stopping times each taking a discrete set of values as in Lemma 3.3.
In the first part of the proof we have proved that, for every choice of t1 ; : : : ; tk ,
the r.v.’s Xt1 Cn Xn ; : : : ; Xtk Cn Xn are independent of Fn and have the
same distribution as Xt1 ; : : : ; Xtk , Therefore for every bounded continuous function
˚ W Rm ! R and C 2 F Fn (recall that, as n then F Fn , see
Proposition 3.5 c)), we have

E ˚.Xt1 Cn Xn ; : : : ; Xtk Cn Xn /1C

D E ˚.Xt1 Cn Xn ; : : : ; Xtk Cn Xn / P.C/ (3.12)

D E ˚.Xt1 ; : : : ; Xtk / P.C/ :
Let us take the limit as n ! 1 in (3.12). As the paths are continuous and by
Lebesgue’s theorem

E ˚.Xt1 Cn Xn ; : : : ; Xtk Cn Xn /1C ! E ˚.Xt1 C X ; : : : ; Xtk C X /1C

n!1
so that

E ˚.Xt1 C X ; : : : ; Xtk C X /1C D E ˚.Xt1 ; : : : ; Xtk / P.C/ : (3.13)
Again by first choosing C D ˝ we find that the joint distributions of .Xt1 C

X ; : : : ; Xtk C X / and .Xt1 ; : : : ; Xtk / coincide. Therefore the process .XtC X /t ,
having the same joint distributions of a Brownian motion, is a Brownian motion
itself. Then taking C 2 F we find that it is independent of F . t
u
Let us describe a first, very useful, application of the stopping theorem. Let a > 0
and a D infftI Xt ag; a is a stopping time, being the exit time from the half-
line 1; aŒ, and is called the passage time at a. By the Iterated Logarithm Law,
Corollary 3.2, the Brownian motion takes arbitrarily large values a.s., hence a is
a.s. finite.
3.6 The stopping theorem 69
It is often of interest to determine the probability that a Brownian motion X

will cross the level a before a time t. This will be a consequence of the following
more general (and interesting) statement. Note that the events fsupst Xs ag and
fa tg coincide.
Theorem 3.4 (The joint distribution of a Brownian motion and of its

running maximum) Let X be a Brownian motion. For every t 0, let
Xt D sup0st Xs . Then, if a 0 and b a,
P.Xt a; Xt b/ D P.Xt 2a b/ : (3.14)
Proof Let a D infft 0I Xt D ag. If we write Wt D XtCa Xa , we have
P.Xt a; Xt b/ D P.a t; Xt b/ D P.a t; Wta b a/ ;
as Wta D Xt Xa D Xt a. We know that W is a Brownian motion independent

of Fa , hence of a . As W has same law as W, the pair .a ; W/, as an RC
C .Œ0; t; R/-valued r.v., has the same law as .a ; W/.
Let A D f.s; / 2 R C .Œ0; t; R/; s t; .t s/ b ag, so that we have
P.a t; Wta b a/ D P..a ; W/ 2 A/. Then, if a b,
P.Xt a; Xt b/ D P.a t; Wta b a/ D P..a ; W/ 2 A/

D P..a ; W/ 2 A/ D P.a t; Xt C Xa b a/
D P.a t; Xt 2a b/ D P.Xt 2a b/
since if Xt 2a b then a fortiori Xt a, so that fXt 2a bg fa tg. t

u
Let us denote by f the joint density of .Xt ; Xt /. From (3.14), for a > 0, b a,
Z b Z C1
dx f .x; y/ dy D P.Xt b; X a/ D P.Xt 2a b/
1 a
Z C1
1 2 =2
Dp p ex dx :
2 .2ab/= t
Taking first the derivative with respect to b we have

Z C1
1 1 2
f .b; y/ dy D p e 2t .2ab/
a 2t
and then taking the derivative with respect to a we find
Corollary 3.3 The joint density of .Xt ; Xt /, i.e. of the running maximum and
the position at time t of a Brownian motion, is
2 1 2
f .b; a/ D p .2a b/ e 2t .2ab/
2t 3
for a > 0; b a and f .b; a/ D 0 otherwise.
Another consequence of Theorem 3.4 is the following, which is an important

improvement of Lemma 3.1.
Corollary 3.4 (The reflection principle)

P.Xt a/ D P.a t/ D 2P.Xt a/ :
Proof We have, from (3.14), P.Xt a; Xt a/ D P.Xt a/. Moreover, fXt

a; Xt ag D fXt ag. Hence
P.Xt a/ D P.Xt a; Xt a/ C P.Xt a; Xt a/ D 2P.Xt a/ :
t
u
Remark 3.5 Because of the symmetry of the centered Gaussian distributions

we have 2P.Xt a/ D P.jXt j a/. Let again Xt D sup0st Xs . As a
consequence of the reflection principle, for every a > 0,
P.Xt a/ D P.a t/ D 2P.Xt a/ D P.jXt j a/ :
This means that, for every t, the two r.v.’s Xt and jXt j have the same
distribution. Of course, the two processes are different: .Xt /t is increasing,
whereas .jXt j/t is not.
3.7 The simulation of Brownian motion
How to simulate the path of a Brownian motion? This is a very simple task, but
it already enables us to investigate a number of interesting situations. Very often a
simulation is the first step toward the understanding of complex phenomena.
3.7 The simulation of Brownian motion 71
Applications may require the computation of the expectation of a functional of

the path of a stochastic process and if closed formulas are not available one must
resort to simulation. This means that one must instruct a computer to simulate many
paths and the corresponding value of the functional: by the law of large numbers
the arithmetic mean of the values obtained in this way is an approximation of the
desired expectation.
In this section we see the first aspects of the problem of simulating a stochastic
process starting from the case of Brownian motion. For more information about
simulation, also concerning more general processes, see Chap. 11. On this subject,
because of today’s ease of access to numerical computation, a vast literature is
available (see, for example, Kloeden and Platen 1992; Fishman 1996; Glasserman
2004).
The first idea (others are possible) for the simulation of a Brownian motion is
very simple: its increments being independent Gaussian r.v.’s, the problem is solved
as soon as we are able to simulate a sequence of independent Gaussian r.v.’s, which
can be done as explained in Sect. 1.8.
If Z1 ; Z2 ; : : : are independent m-dimensional and N.0; I/-distributed r.v.’s on
some probability space .˝; F ; P/, let us choose a grid of times 0 < h < 2h p < :::
where h > 0 is a positive number (typically to be taken small). Then the r.v. h Z1
is N.0; h I/-distributed, i.e. has the same distribution as the increment of a Brownian
motion over a time interval of size h. Hence the r.v.’s
p
Bh .h/ D ph Z1
Bh .2h/ D p h .Z1 C Z2 /
(3.15)
Bh .kh/ D h .Z1 C C Zk /
:::
have the same joint distributions as the positions at times h; 2h; : : : ; kh; : : : of a
Brownian motion.
If, for kh t .k C 1/h, we define Bh .t/ as a linear interpolation of the positions
Bh .kh/ and Bh ..k C 1/h/, this is obviously an approximation of a Brownian motion.
More precisely:
Theorem 3.5 Let T > 0, n > 0 and h D Tn . Let us denote by Ph the law
of the process Bh (Ph is therefore a probability on the canonical space C D
C .Œ0; T; Rm /). Then Ph converges weakly to the Wiener measure PW .
Proof Note first that the law Ph neither depends on the choice of the r.v.’s .Zn /n ,
nor on the probability space on which they are defined, provided they are N.0; I/-
distributed and independent.
Let .C ; M ; PW / be the canonical space and, as usual, denote by Xt the coor-
dinate applications (recall that PW denotes the Wiener measure, as defined p. 53).
The process X D .C ; M ; .Xt /t ; PW / is therefore a Brownian motion. The r.v.’s
1
Zk D p .X.kC1/h Xkh /
h
are therefore independent and N.0; I/-distributed. If we construct the process Bh

starting from these r.v.’s, we have Bh .kh/ D Xkh . The process Bh , in this case, is
nothing else but the piecewise linear interpolation of X. As the paths are continuous
(and therefore uniformly continuous on every compact interval), limh!0 Bh .!; t/ D
Xt .!/ for every ! 2 C uniformly on compact sets. This means that the family of
C -valued r.v.’s .Bh /h converges to X a.s. in the topology of C . Therefore, as a.s.
convergence implies convergence in law, Bh ! X in law as h ! 0.
Hence, as remarked above, Ph ! PW weakly as h ! 0, whatever the choice of
the r.v.’s .Zn /n . t
u
Theorem 3.5 ensures that if W C ! R is a bounded continuous function, then
lim EŒ.Bh / D EŒ.B/ :

h!0
Example 3.2 Let B be a real Brownian motion and let us consider the problem
of computing numerically the quantity
hZ 1
2
i
E eBs ds : (3.16)
0
It is easy to determine the exact value:

Z Z ˇ1 p
1
2
1
1 p ˇ
EŒeBs ds D p ds D 1 C 2sˇ D 3 1 ' 0:7320508 :
0 0 1 C 2s 0
In order to compute by simulation an approximation of this value, note

that, using the same argument as in the proof of Theorem 3.5, the integral
R 1 B2
0 e
s ds is the limit of its Riemann sums, i.e.
X
n Z 1
2 2
lim h eBh .kh/ D eBs ds
h!0 0
kD1
and, by Lebesgue’s theorem,
h Xn
2
i hZ 1
2
i
lim E h eBh .kh/ D E eBs ds :
h!0 0
kD1
(continued )
3.7 The simulation of Brownian motion 73
Example 3.2 (continued)

Hence, if .Zni /n are independent N.0; 1/-distributed r.v.’s then, by the Law of
Large Numbers,
1 X X h.Z i CCZ i /2
N n h X n
2
i
lim e 1 kh DE h eBh .kh/ (3.17)
N!1 N
iD1 kD1 kD1
so that the left-hand side above is close to the quantity (3.16) of interest.
The results for N D 640;000 and various values of the discretization
parameter h are given in Table 3.1. It is apparent that, even for relatively large
values of h, the simulation gives quite accurate results.
Table 3.1 The estimated values of Example 3.2

h Value Error
1
100
0:7295657 0:34%
1
200
0:7311367 0:12%
1
400
0:7315369 0:07%
1
800
0:7320501 105 %
Of course it would be very important to know how close to the true value
EŒ.B/ the approximation EŒ.Bh / is. In other words, it would be very important
to determine the speed of convergence, as h ! 0, of the estimator obtained by
the simulated process. We shall address this question in a more general setting in
Chap. 11.
Note also that in Example 3.2 we obtained an estimator using only the values of
the simulated process at the discretization times kh, k D 0; : : : ; n, and that it was not
important how the simulated process was defined between the discretization times.
This is almost always the case in applications.
Example 3.3 Let B be a real Brownian motion and let us consider the problem
of computing numerically the quantity

P sup Bt 1 : (3.18)
0st
(continued )

We could try to reproduce the simulation algorithm above, i.e. consider the
approximation
1 X
N
X h WD 1 i :
N iD1 fsup0kn Bkh 1g
However, there is a difficulty: the functional C 3 ! 1fsup0sT s 1g

is not continuous because of the presence of the indicator function, so
that Theorem 3.5is not immediately sufficient in order to guarantee that
limh!0 EŒX h D P sup0t1 Bs 1 .
One should, however, recall that the convergence holds if the set A D
fsup0st Bt 1g is such that PW .@A/ D 0, as indicated on p. 11. This is the
case here, as
n o
@A D sup Bt D 1
0st
and the r.v. B D sup0st Bt has a density, which is an immediate

consequence of Remark 3.5. Hence PW .@A/ D P.B D 1/ D 0.
Table 3.2 reproduces the results of the computation of the probabil-
ity (3.18) for different values of h.
Here again N D 640;000. It is apparent that the relative error is much
larger than in Example 3.2 and, more noteworthy, that the error decreases very
slowly as the discretization step h becomes smaller. This is certainly related
to the lack of regularity of the functional to be computed.
Table 3.2 The outcomes of the numerical simulation and their relative errors. Thanks to the
reflection principle, Corollary 3.4, the true value is 2P.B1 1/ D 2.1 ˚.1// D 0:3173, ˚
denoting the partition function of an N.0; 1/-distributed r.v.
h Value Error
1
100
0:2899 8:64%
1
200
0:2975 6:23%
1
400
0:3031 4:45%
1
800
0:3070 3:22%
1
1600
0:3099 2:32%
Exercises
3.1 (p. 455) Let B be a Brownian motion and let s t.

a) Compute EŒBs B2t .
b) Compute EŒB2s B2t .
c) Show that
EŒBs eBs D ses=2 :
d) Compute EŒBs eBt .

Recall: EŒX 4 D 3 if X N.0; 1/.
3.2 (p. 455) Let B be a Brownian motion. Compute

a) limt!C1 EŒ1fBt ag .
b) limt!C1 EŒBt 1fBt ag .
3.3 (p. 456) Let B be a Brownian motion. Compute

p 2
lim t EŒB2t eBt :
t!C1
3.4 (p. 456) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a d-dimensional Brownian motion.
We prove here that the increments .Bt Bs /ts form a family that is independent of
Fs , as announced in Remark 3.2. This fact is almost obvious if Fs D .Bu ; u s/,
i.e. if B is a natural Brownian motion (why?). It requires some dexterity otherwise.
Let s > 0.
a) If s t1 < < tm and 1 ; : : : ; m 2 B.Rd / prove that, for every A 2 Fs ,

P Btm Btm1 2 m ; : : : ; Bt1 Bs 2 1 ; A

D P Btm Btm1 2 m ; : : : ; Bt1 Bs 2 1 P.A/ :
b) Show that the two -algebras
.Btm Bs ; : : : ; Bt1 Bs / and .Btm Btm1 ; Btm1 Btm2 ; : : : ; Bt1 Bs /
are equal.
c) Prove that the -algebra .Bt Bs ; t s/ is independent of Fs .
3.5 (p. 457) Let B be a .Ft /t -Brownian motion, let G a -algebra independent of
W
F1 D t Ft and F e s D Fs _ G . The object of this exercise is to prove that B
e t /t .
remains a Brownian motion with respect to the larger filtration .F
a) Let C be the class of the events of the form A \ G with A 2 Fs and G 2 G .

Prove that C is stable with respect to finite intersections and that it generates the
-algebra Fe s.
e t /t -Brownian motion.
b) Prove that B is also a .F
3.6 (p. 457) Let B be a Brownian motion.

a) Show that if 0 s t, then the joint law of .Bs ; Bt / is
1 1 2 1 1 2
fs;t .x; y/ D p e 2s x p e 2.ts/ . yx/ : (3.19)
2s 2.t s/
b) Show that, for every s > 0,
1
P.Bs < 0; B2s > 0/ D
8
3.7 (p. 459)

a) Let B be a Brownian motion and let Xt D et Be2t .
a1) Show that .Xt /1<t<C1 is a Gaussian process and compute its covariance
function K.s; t/ D Cov.Xs ; Xt /.
a2) Prove that X is a stationary process, i.e. such that for every t1 < t2 < <
tm and h > 0, the r.v.’s .Xt1 ; : : : ; Xtm / and .Xt1 Ch ; : : : ; Xtm Ch / have the same
distribution.
a3) Prove that the kernel K.s; t/ D ejtsj is positive definite (i.e. that it satis-
fies (2.1)).
b) Let .Xt /1<t<C1 be a (not necessarily continuous) centered Gaussian process
with covariance function
p K.s; t/ D ejtsj .
b1) Prove that Wu D u X 1 log u is a natural Brownian motion.
2
b2) Prove that X has a continuous version.
3.8 (p. 459) Sometimes in applications a process appears that is called a -corre-
lated Brownian motion. In dimension 2 it is a Gaussian process .X1 .t/; X2 .t//t such
that X1 and X2 are real Brownian motions, whereas Cov.X1 .t/; X2 .s// D .s ^ t/,
where 1 1 (Figs. 3.4 and 3.5).
a) Let B D .B1 .t/;
pB2 .t// be a two-dimensional Brownian motion. Then if X2 D B2
and X1 .t/ D 1 2 B1 C B2 .t/, then X D .X1 .t/; X2 .t// is a -correlated
Brownian motion.
b) Conversely, if X is a -correlated Brownian motion and jj < 1, let B2 .t/ D X2 .t/
and
1
B1 .t/ D p X1 .t/ p X2 .t/ :
1 2 1 2
Prove that B D .B1 .t/; B2 .t//

p is a two-dimensional Brownian motion such that
X2 .t/ D B2 .t/ and X1 .t/ D 1 2 B1 C B2 .t/.
–.5
–.5 0 1
Fig. 3.4 Example of a path of a two-dimensional -correlated Brownian motion for 0 t 1

and D 0:7 (a black small circle denotes the origin and the position at time 1). Of course, for
D 1 the paths concentrate on the main diagonal
.8
.4
–.4
–.5 0
Fig. 3.5 Example of a path of a two-dimensional -correlated Brownian motion for 0 t 1

and D 0:6. Now the correlation is negative. (a black small circle denotes the origin and the
position at time 1)
3.9 (p. 460) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian

motion. Determine the matrices A 2 M.m/ such that Xt D ABt is also an m-
dimensional Brownian motion.
3.10 (p. 460) Let B be an m-dimensional Brownian motion and A Rm a Borel

set having finite and strictly positive Lebesgue measure. Let
SA .!/ D ft 2 RC I Bt .!/ 2 Ag :
SA .!/ is the set of the times t such that Bt .!/ 2 A. Let us denote by the Lebesgue
measure of R; therefore the quantity EŒ .SA / is the mean time spent by B on the set
A (SA is called the occupation time of A). Prove that
8
<C1 if m 2
Z
EŒ .SA / D 1
: . 2 1/ jxj2m dx
m
if m > 2 :
2 m=2 A
R C1
Recall the definition of the Gamma function: .˛/ D 0 x˛1 ex dx, for ˛ > 0.
R C1
Note that EŒ .SA / D 0 P.Xt 2 A/ dt by Fubini’s theorem.
3.11 (p. 461) Let X be a continuous Gaussian process.

a) Let us denote by G the family of the compactly supported finite measures on
RC (i.e. such that .Œa; bc / D 0 for some a; b 2 RC ). Let
Z
X D Xs d.s/ : (3.20)
Show that .X / 2G is a Gaussian family.

b) Let be a measure on RC such that .Œa; b/ < C1 for every finite interval
Œa; b (i.e. a Borel measure) and let
Z t
Yt D Xs .ds/ :
0
b1) Prove that Y is a Gaussian process.

b2) Let us assume that X is a Brownian motion. Using the relation
Z t
2
.r; t/ D 2 .r; u/ d.u/ (3.21)
r
Rt
prove that Yt has centered Gaussian law with variance t2 D 0 .s; t/2 ds.
Compute Cov. Yt ; Ys /.
P
a) The integral is the limit of the Riemann sums i0 Xi=n .i nt ; .i C 1/ nt / and one can apply
Proposition 1.9;
b2) use the relations (Fubini)
Z t Z s Z t Z s
E Xu d.u/ Xv d.v/ D d.u/ E.Xu Xv / d.v/ ;
0 0 0 0
Z s Z s Z v Z s Z s Z s
v d.v/ D d.v/ du D du d.v/ D .u; s/ du :
0 0 0 0 u 0
3.12 (p. 463) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and let
Z t
Bu
Zt D Bt du :
0 u
a) Note that the integral converges for almost every ! and show that .Zt /t is a
Brownian motion with respect to its natural filtration .Gt /t .
b) Show that .Zt /t is a process adapted to the filtration .Ft /t , but it is not a Brownian
motion with respect to this filtration.
c) Prove that, for every t, Bt is independent of Gt .
a) Use in an appropriate way Exercise 3.11.
3.13 (p. 464) Let B be a Brownian motion. Prove that, for every a > 0 and T > 0,
p
P.Bt a t for every t T/ D 0 :
3.14 (p. 465) Let B be a Brownian motion and b 2 R, > 0. Let Xt D ebtC Bt .
a) Investigate the existence and finiteness of the a.s. limit
lim Xt
t!C1
according to the possible values of b; .

b) Investigate the existence and finiteness of
lim EŒXt (3.22)

t!C1
according to the possible values of b; .
3.15 (p. 465) Let B be a Brownian motion and b 2 R, > 0.

a) For which values of b; , b 6D 0, is the integral
Z C1
ebuC Bu du
0
a.s. finite?
b1) Prove that the r.v.
Z 1
1fBu>0g du
0
is > 0 a.s.
b2) Prove that

Z t
lim 1fBu >0g du D C1 a.s.
t!C1 0
b3) Deduce that

Z C1
e Bu du D C1 a.s.
0
c) For which values of b; is

hZ C1 i
E ebuC Bu du < C1 ‹
0
3.16 (p. 466) (Approximating exit times) Let E be a metric space, D E an open
set, X D .Xt /t a continuous E-valued process, the exit time of X from D. Let .Dn /n
be an increasing sequence of open sets with Dn D for every n > 0, and such that
d.@Dn ; @D/ 1n . Let us denote by n the exit time of X from Dn . Then, as n ! 1,
n ! and, on f < C1g, Xn ! X .
3.17 (p. 467) Let B be an m-dimensional Brownian motion and D an open set
containing the origin. For > 0 let us denote by D the set D homothetic to D and
by and the exit times of B from D and D respectively.
Show that the r.v.’s and 2 have the same law. In particular, EŒ D 2 EŒ,
this relation being true whether the quantities appearing on the two sides of the
equation are finite or not.
3.18 (p. 467) Let S be the unit ball centered at the origin of Rm , m 1, and X a
continuous m-dimensional Brownian motion. Let D infftI Xt … Sg be the exit time
of X from S.
a) Prove that < C1 a.s.; X is therefore a r.v. with values in @S. Show that the
law of X is the .m 1/-dimensional Lebesgue measure of @S, normalized so that
it has total mass 1.
b) Prove that and X are independent.
Recall that an orthogonal matrix transforms a Brownian motion into a Brownian motion
(Exercise 3.9 b)) and use the fact that the only measures on @S that are invariant under the action
of orthogonal matrices are of the form c , where is the .m 1/-dimensional Lebesgue measure
of @S and c some positive constant.
3.19 (p. 468)

a) Let f W Œ0; 1 ! R be a continuous function. Prove that
log ke f k1 D sup f .s/ :

0s1
b) Let B be a Brownian motion.

b1) Prove that, for every t > 0,
Z t Z 1 p
L
eBs ds t e tBs
ds :
0 0
b2) Prove that

Z
1 t
L
p log eBs ds ! sup Bs :
t 0 t!C1 s1
b3) Give an approximation, for t large, of

Z t Z t p
P eBs ds 1 P eBs ds e:77337 t
0 0
3
(0:77337 is, approximately, the quantile of order 4
of the N.0; 1/ distribution).
Recall the relation among the Lp norms: limp!C1 k f kp D k f k1 .
3.20 (p. 469)

a) Let X be a continuous Brownian motion, a > 0 and a D infftI Xt ag the
passage time of X at a. Prove that the law of a has a density, with respect to
Lebesgue measure, given by fa .t/ D 0 for t 0 and (Fig. 3.6)
a 2 =2t
fa .t/ D ea (3.23)
.2/1=2 t3=2
for t > 0. In particular, a does not have finite expectation (and neither does
p
a ). Show that, as t ! C1,
2a 1
P.a > t/ p p
2 t
b) A researcher wants to explore the distribution of 1 by simulation. He therefore

realizes a computer program and simulates N paths of a Brownian motion and
records the passage times at a D 1 of each of them, T1 ; : : : ; TN say.
Assume N D 10;000. What is the probability that Tj 108 for at least one
path? And the probability that Tj 1010 (1 hundred millions) for at least one
path? Why is the program taking so long? This should be done numerically with
a computer.
c) Let us recall that a probability is said to be stable with exponent ˛; 0 < ˛ 2,
if
X1 C C Xn
X1
n1=˛
where X1 ; : : : ; Xn are independent -distributed r.v.’s (for instance, a centered
normal law is stable with exponent 2).
Prove that the law of a is stable with exponent ˛ D 12 .
a) Note that P.a t/ D P.sup0st Bs > a/. c) Use Theorem 3.3 to prove that if X1 ; : : : ; Xn
are independent copies of 1 , then X1 C C Xn has the same distribution as na . This property of
stability of the law of a will also be proved in Exercise 5.30 in quite a different manner. Do not
try to compute the law of X1 C C Xn by convolution!
.5
.25
0 1 2 3 4
Fig. 3.6 The graph of the density f of the passage time at a D 1. Note that, as t ! 0C, f tends
to 0 very fast, whereas its decrease for t ! C1 is much slower, which is also immediate from
(3.23)
3.21 (p. 470)

a) Prove, without explicitly computing the values, that
h i p h i
E sup Bs D t E sup Bs : (3.24)
0st s1
b) Compute
h i
E sup Bs
0st
(recall Exercise 1.3).

3.22 (p. 471)

a) Let B be an m-dimensional Brownian motion and let, for Rm 3 z 6D 0, k > 0,
H D fx 2 Rm ; hz; xi kg be a hyperplane containing the origin in its interior.
Let us denote by the exit time of B from H. The object of this exercise is to
compute the law of and its expectation.
a1) Prove that
1
Xt D hz; Bt i
jzj
is a Brownian motion and that coincides with the passage time of X at a D jzjk .
a2) Compute the density of and EŒ. Is this expectation finite?
b) Let B D .B1 ; B2 / be a two-dimensional -correlated Brownian motion (see
Exercise 3.8) and consider the half-space H D f.x; y/; x C y < 1g.
b1) Find a number ˛ > 0 such that
Xt D ˛.B1 .t/ C B2 .t//
is a natural Brownian motion.

b2) Let be the exit time of B from H. Compute the density of and EŒ. Is this
expectation finite? Compute P. 1/ and determine for which values of this
quantity is maximum.
3.23 (p. 472) (Do Exercise 3.22 first)

a) Let B be a two-dimensional Brownian motion and z 2 R2 ; z 6D 0. Let Xt D
hBt ; zi. Show that there exists a number v > 0 such that Wt D vXt is a Brownian
motion.
b) Let be the first time at which B reaches the straight line x C y D 1.
b1) Let z D .1; 1/. Show that coincides with the passage time at a D 1 of Xt D
hBt ; zi.
b2) Show that the law of has a density and compute it. Compute P. 1/.
c) Let 1 and 2 be the first times at which B reaches the lines 2x C y D 1 and
1
2 x y D 1, respectively.
c1) Compute the densities of 1 and 2 .
c2) Prove that 1 and 2 are independent.
c3) Let D 1 ^ 2 : is the exit time of B out of the shaded infinite region in
Fig. 3.7. With the help of numerical tables compute P. 1/.
c4) Compute P.1 2 /.
........ ......
........ .......
........ .........
........ ...........
........ ...........
........ ............. 2x+y=1
........ ..............
........ ...............
........ ...............
........ .................
........ ..................
........ ...................
........ ...................
........ .....................
•
........ ...................... ...
..
........ ...................... ...
........ ........................ .
...
......
........ ......................... ...
.......
........ .......................... ..............
........ ................................ 1 x−y=1
........ ........................... .... 2
........ ......................... ...
...
........ ................... .
.
........ ...............
........ ............
..
Fig. 3.7 D 1 ^ 2 is the exit time out of the shaded area

Chapter 4
Conditional Probability
4.1 Conditioning
Let .˝; F ; P/ be a probability space and A 2 F an event having strictly positive

probability. Recall that the conditional probability of P with respect to A is the
probability PA on .˝; F /, which is defined as
P.A \ B/
PA .B/ D for every B 2 F : (4.1)
P.A/
Intuitively the situation is the following: initially we know that every event B 2 F
can appear with probability P.B/. If, later, we acquire the information that the event
A has occurred or will certainly occur, we replace the law P with PA , in order to keep
into account the new information.
A similar situation is the following. Let X be a real r.v. and Z another r.v. taking
values in a countable set E and such that P.Z D z/ > 0 for every z 2 E. For every
Borel set A R and for every z 2 E let
P.X 2 A; Z D z/
n.z; A/ D P.X 2 AjZ D z/ D
P.Z D z/
For every z 2 E, A 7! n.z; A/ is a probability on R: it is the conditional law of X

given Z D z. This probability has an intuitive meaning not dissimilar from the one
just pointed out above: A 7! n.z; A/ is the law that it is reasonable to consider for
the r.v. X, should we acquire the information that the event fZ D zg certainly occurs.
The conditional expectation of X given Z D z is defined as the mean of this law, if
it exists:
Z Z
1 EŒX1fZDzg
EŒX jZ D z D x n.z; dx/ D X dP D
P.Z D z/ fZDzg P.Z D z/

86 4 Conditional Probability
This is a very important concept, as we shall see constantly throughout. For this
reason we need to extend it to the case of a general r.v. Z (i.e. without the assumption
that it takes at most only countably many values). This is the object of this chapter.
Let us see first how it is possible to characterize the function h.z/ D EŒX jZ D z
in a way that continues to be meaningful in general. For every B E we have
Z X X
h.Z/ dP D EŒX jZ D zP.Z D z/ D EŒX1fZDzg
fZ2Bg z2B z2B
Z
D EŒX1fZ2Bg D X dP :
fZ2Bg
Therefore the r.v. h.Z/, which is of course .Z/-measurable, is such that its integral
on any event B of .Z/ coincides with the integral of X on the same B. We shall see
that this property characterizes the conditional expectation.
In the following sections we use this property in order to extend the notion of
conditional expectation to a more general (and interesting) situation. We shall come
back to conditional distributions at the end.
4.2 Conditional expectations
Let X be a real r.v. and X D X C X its decomposition into positive and negative
parts and assume X to be lower semi-integrable (l.s.i.) i.e. such that EŒX < C1.
See p. 2 for more explanations.
Definition and Theorem 4.1 Let X be a real l.s.i. r.v. and D F a sub--
algebra. The conditional expectation of X with respect to D, denoted EŒX jD,
is the (equivalence class of) r.v.’s Z which are D-measurable and l.s.i. and
such that for every D 2 D
Z Z
Z dP D X dP : (4.2)
D D
A r.v. Z with these properties always exists and is unique up to P-equivalence.
Proof Existence. Let us assume first that X 0; then let us consider on .˝; D/ the
positive measure
Z
Q.B/ D X dP B2D:
B
4.2 Conditional expectations 87
Obviously Q P (see p. 4); thus by the Radon–Nikodym theorem there exists a

real D-measurable r.v. Z such that
Z
Q.B/ D Z dP
B
and hence Z 2 EŒX jD. Note that if X is integrable then Q is finite and Z is
integrable. For a general r.v. X just decompose X D X C X and check that
EŒX jD D EŒX C jD EŒX jD satisfies the conditions of Definition 4.1. This is
well defined since, X being integrable, EŒX jD is integrable itself and a.s. finite
(so that the form C1 1 cannot appear).
Uniqueness. Let us assume first that X is integrable. If Z1 ; Z2 are D-measurable
and satisfy (4.2) then the event B D fZ1 > Z2 g belongs to D and
Z Z Z
.Z1 Z2 / dP D X dP X dP D 0 :
B B B
As the r.v. Z1 Z2 is strictly positive on B this implies P.B/ D 0 and hence Z2 Z1

a.s. By symmetry, Z1 D Z2 a.s.
In general if X is l.s.i. and not necessarilyR integrable this argument cannot
be applied as is, since it is possible that B X dP D C1. However, B can be
approximated by events Bn on which the integral of X is finite; the details are left as
an exercise. t
u
Remark 4.1 (4.2) states that if Z D EŒX jD, the relation
EŒZW D EŒXW (4.3)
holds for every r.v. W of the form W D 1A , A 2 D. Of course, then, (4.3)

also holds for every linear combination of these indicator functions and,
by standard approximation results (see Proposition 1.11), for every positive
bounded D-measurable r.v. W.
If Y is a r.v. taking values in some measurable space .E; E /, sometimes we shall

write EŒX jY instead of EŒX j.Y/. We know that every real .Y/-measurable r.v.
is of the form g.Y/, where g W E ! R is a measurable function (Doob’s criterion,
Lemma 1.1). Therefore there exists a measurable function g W E ! R such that
EŒX jY D g.Y/ a.s. Sometimes we shall use for the function g, in an evocative
manner, the notation
g.y/ D EŒX jY D y :
Always keeping in mind that every .Y/-measurable r.v. W is of the form .Y/ for
a suitable measurable function , such a g must satisfy the relation
EŒX .Y/ D EŒg.Y/ .Y/ (4.4)
for every bounded measurable function . The next section provides some tools for
the computation of g.y/ D EŒX jY D y.
To be precise (or pedantic. . . ) a conditional expectation is an equivalence class of
r.v.’s but, in order to simplify the arguments, we shall often identify the equivalence
class EŒX jD with one of its elements Z.
Remark 4.2 We shall often be called to verify statements of the form “a

certain r.v. Z is equal to EŒX jD”. On the basis of Definition 4.1 this is
equivalent to show that
a) Z is D-measurable; and
b) EŒZ1D D EŒX1D for every D 2 D.
In fact it will be enough to check b) for every D in a class C D generating
D and stable with respect to finite intersections and containing ˝. This is
actually a simple application of Theorem 1.4: let H be the space of the D-
measurable bounded r.v.’s W such that
EŒWX D EŒWZ : (4.5)
Clearly H contains the function 1 and the indicator functions of the events
of C . Moreover, if .Wn /n is an increasing sequence of functions of H all
bounded above by the same element W 2 H and Wn " W, then we have,
for every n, W1 Wn W . As both W1 and W are bounded (as is every
function of H ), with two applications of Lebesgue’s theorem we have
EŒWX D lim EŒWn X D lim EŒWn Z D EŒWZ:

n!1 n!1
Therefore condition i) of Theorem 1.4 is satisfied and H contains every

.C /-bounded measurable r.v.; therefore (4.5) holds for every D-bounded
measurable r.v. W, which allows us to conclude the proof.
Proposition 4.1 Let X; X1 ; X2 be integrable r.v.’s and ˛; ˇ 2 R. Then

a) EŒ˛X1 C ˇX2 jD D ˛EŒX1 jD C ˇEŒX2 jD a.s.
b) If X 0 a.s., then EŒX jD 0 a.s.
c) EŒEŒX jD D EŒX.
(continued)
Proposition 4.1 (continued)

d) If Z is bounded and D-measurable then EŒZX jD D ZEŒX jD a.s.
e) If D D 0 then EŒEŒX jD 0 jD D EŒX jD a.s.
f) If X is independent of D then EŒX jD D EŒX a.s.
Proof These are immediate applications of the definition; let us look more carefully
at the proofs of the last three statements.
d) As it is immediate that the r.v. Z EŒX jD is D-measurable,

we must only prove
that the r.v. ZEŒX jD satisfies the relation E W ZEŒX jD D EŒWZX for every
r.v. W. This is also immediate, as ZW is itself bounded and D-measurable.
e) The r.v. EŒEŒX jD 0 jD is D-measurable; moreover, if W is bounded and D-
measurable, then it is also D 0 -measurable and, using c) and d),

E WEŒEŒX jD 0 jD D E EŒEŒWX jD 0 jD D E EŒWX jD 0 D EŒWX ;

f) The r.v. ! 7! EŒX is constant, hence D-measurable. If W is D-measurable then
it is independent of X and
EŒWX D EŒWEŒX D EŒWEŒX ;
hence EŒX D EŒX jD a.s. t

u
It is easy to extend Proposition 4.1 to the case of r.v.’s that are only l.s.i. We shall
only need to observe that a) holds only if ˛; ˇ 0 (otherwise ˛X1 C ˇX2 might not
be l.s.i. anymore) and that d) holds only if Z is bounded and positive (otherwise ZX
might not be l.s.i. anymore).
Proposition 4.2 Let X, Xn , n D 1; 2; : : : ; be real l.s.i. r.v.’s. Then

a) (Beppo Levi) if Xn % X a.s. then EŒXn jD % EŒX jD a.s.
b) (Fatou) If limn!1 Xn D X and for every n Xn Z for some integrable r.v.
Z then
EŒX jD lim EŒXn jD a.s.

n!1
c) (Lebesgue) If limn!1 Xn D X and for every n jXn j Z for some integrable

r.v. Z then
lim EŒXn jD D EŒX jD a.s.

n!1
(continued)
Proposition 4.2 (continued)

d) (Jensen’s inequality) If ˚ is a convex lower semi-continuous function R !
R [ fC1g then ˚.Y/ is also l.s.i. and

EŒ˚.Y/jD ˚ EŒY jD a.s.
Proof
a) As .Xn /n is a.s. increasing, the same is true for the sequence .EŒXn jD/n by
Proposition 4.1 b); hence the limit limn!1 EŒXn jD exists a.s. Let us set Z WD
limn!1 EŒXn jD. By Beppo Levi’s theorem applied twice we have, for every
D 2 D,

EŒZ1D D E lim EŒXn jD1D D lim EŒEŒXn jD1D D lim EŒXn 1D D EŒX1D ;
n!1 n!1 n!1
hence Z D EŒX jD a.s.

b) If Yn D infkn Xk then
lim Yn D lim Xn D X :
n!1 n!1
Moreover, .Yn /n is increasing and Yn Xn . We can therefore apply a) and obtain
EŒX jD D lim EŒYn jD lim EŒXn jD :

n!1 n!1
c) Immediate consequence of b), applied first to Xn and then to Xn .

d) The proof is identical to the proof of the usual Jensen’s inequality: recall that a
convex l.s.c. function ˚ is equal to the supremum of all the affine linear functions
that are bounded above by ˚. If f .x/ D ax C b is an affine function minorizing
˚, then

EŒ˚.X/jD EŒ f .X/jD D f EŒX jD :
Then just take the upper bound in f in the previous inequality among all the affine
functions minorizing ˚. t
u
Sometimes we shall write P.AjD/ instead of EŒ1A jD and shall speak of the
conditional probability of A given D.
Remark 4.3 It is immediate that, if X D X 0 a.s., then EŒX jD D EŒX 0 jD a.s.
Actually, if Z is a D-measurable r.v. such that EŒZ1D D EŒX1D for every
D 2 D, then also EŒZ1D D EŒX 0 1D for every D 2 D. The conditional
expectation is therefore defined on equivalence classes of r.v.’s.
Let us investigate the action of the conditional expectation on Lp spaces.
We must not forget that Lp is a space of equivalence classes of r.v.’s, not of
r.v.’s. Taking care of this fact, Jensen’s inequality (Proposition 4.2 d)), applied
to the convex function x 7! jxjp with p 1, gives jEŒX jDjp EŒjXjp jD, so
that
ˇ ˇp

E ˇEŒX jDˇ E EŒjXjp jD/ D EŒjXjp : (4.6)
The conditional expectation is therefore a continuous linear operator Lp !

Lp ; p 1; moreover, it has norm 1, i.e. it is a contraction. The image
of Lp through the operator X 7! EŒX jD is the subspace of Lp , which we
shall denote by Lp .D/, that is formed by the equivalence classes of r.v.’s that
contain at least one D-measurable representative. In particular, since p 1,
by (4.6)
ˇ ˇp
ˇ ˇp
ˇ ˇp
E ˇEŒXn jD EŒX jDˇ D E ˇEŒXn X jDˇ E ˇXn X ˇ ;
Lp Lp
we have that Xn ! X implies EŒXn jD ! EŒX jD.
n!1 n!1
The case L2 deserves particular attention. If X 2 L2 , we have for every
bounded D-measurable r.v. W,

E .X EŒX jD/W D EŒXW E EŒX jDW D 0 : (4.7)
As bounded r.v.’s are dense in L2 , the relation EŒ.X EŒX jD/W D 0 also
holds for every r.v. W 2 L2 .D/. In other words, X EŒX jD is orthogonal to
L2 .D/, i.e. EŒX jD is the orthogonal projection of X on L2 .D/. In particular,
this implies that
EŒX jD is the element of L2 .D/ that minimizes the L2 distance from X.
If for simplicity we set Z D EŒX jD, then, for every W 2 L2 .D/,
kX Wk22 D EŒ.X W/2 D EŒ.X Z C Z W/2

D EŒ.X Z/2 C 2EŒ.X Z/.Z W/ C EŒ.Z W/2 :
(continued )

Note that EŒ.XZ/.ZW/ D 0 thanks to (4.7), since ZW is D-measurable.
Hence the double product above vanishes and
kX Wk22 D EŒ.X Z/2 C EŒ.Z W/2 EŒ.X Z/2 D kX Zk22 ;
where the inequality is even strict, unless Z D W a.s.
Therefore, in the sense of L2 , EŒX jD is the best approximation of X by a
D-measurable r.v. (Fig. 4.1).
Fig. 4.1 The L2 distance between EŒX j D and X is the shortest because the angle between the
segments EŒX j D ! W and EŒX j D ! X is 90ı
Example 4.1 If D D f˝; ;g is the trivial -algebra, then

EŒX jD D EŒX :
Actually, the only D-measurable r.v.’s are the constants and, if c D EŒX jD,
then the constant c is determined by the relationship c D EŒEŒX jD D EŒX.
The notion of expectation thus appears to be a particular case of conditional
expectation.
Example 4.2 Let A 2 F be an event such P.A/ > 0 and let D be the -
algebra fA; Ac ; ˝; ;g. Then EŒX jD, being D-measurable, is constant on A
and on Ac . Its value, c say, on A is determined by the relation
EŒX1A D EŒ1A EŒX jD D cP.A/ :
From this relation and the similar one for Ac we easily derive that
(
1
EŒX1A on A
EŒX jD D P.A/ 1
P.Ac / EŒX1A on A :
c
c
(continued )

R
In particular, EŒX jD is equal toR X dPA on A, where PA is the conditional
probability (4.1), and is equal to X dPAc on Ac .
Example 4.3 Let B D .˝; F ; .Ft /t , .Bt /t ; P/ be a Brownian motion. Then if

st
EŒBt jFs D Bs a.s.
Indeed
EŒBt jFs D EŒBt Bs jFs C EŒBs jFs D EŒBt Bs C Bs D Bs
since Bt Bs is independent of Fs and centered, whereas Bs is already Fs -

measurable.
The computation of a conditional expectation is an operation that we are led to

perform quite often and which, sometimes, is even the objective of the researcher.
Let us now make two remarks that can be of help towards this goal.
Remark 4.4 Sometimes one must compute the conditional expectation of a

r.v. X with respect to a -algebra D that is obtained by adding to a -algebra
D0 the family N of negligible events of a larger -algebra, F for example.
It is useful to observe that
EŒX jD D EŒX jD0 a:s:
If we denote EŒX jD0 by Y, for simplicity, then Y is a fortiori D-measurable.

Moreover, for every event of the form A \ G with A 2 D0 and G 2 N or
G D ˝,
EŒY1A\G D EŒX1A\G :
Actually, both sides vanish if G 2 N whereas they obviously coincide if

G D ˝. As pointed out in Remark 3.1, the events of this form are stable
with respect to finite intersections and generate the -algebra D, therefore
Y D EŒX jD a.s.
Let D be a -algebra. We have seen two situations in which the computation of

EŒjD is easy: if X is D-measurable then
EŒX jD D X
whereas if it is independent of D then
EŒX jD D EŒX : (4.8)
The following lemma, which is quite (very!) useful, combines these two situations.
Lemma 4.1 (The freezing lemma) Let .˝; F ; P/ be a probability space

and G and D independent sub--algebras of F . Let X be a D-measurable
r.v. taking values in the measurable space .E; E / and W E ˝ ! R an
E ˝ G -measurable function such that ! 7! .X.!/; !/ is integrable. Then
EŒ .X; /jD D ˚.X/ ; (4.9)
where ˚.x/ D EŒ .x; /.
Proof Let us assume first that is of the form .x; !/ D f .x/Z.!/, where Z is
G -measurable. In this case, ˚.x/ D f .x/EŒZ and (4.9) becomes
EŒ f .X/Z jD D f .X/EŒZ ;
which is immediately verified. The lemma is therefore proved for all functions of
the type described above and, of course, for their linear combinations. One obtains
the general case with the help of Theorem 1.5. t
u
Lemma 4.1 allows us to easily compute the conditional expectation in many
instances. It says that you can just freeze one of the variables and compute the
expectation of the resulting expression. The following examples show typical
applications.
Example 4.4 Let B be an m-dimensional .Ft /t -Brownian motion and

f W Rm ! R a bounded Borel function. Let s < t. What is the value of
EŒ f .Bt /jFs ?
We can write f .Bt / D f ..Bt Bs / C Bs /: we are then in the situation of
Lemma 4.1, i.e. the computation of the conditional expectation with respect
to Fs of a function of a r.v. that is already Fs -measurable and of a r.v. that is
independent of Fs . If we define .x; !/ D f .x C Bt .!/ Bs .!//, then is
(continued )

B.Rm / ˝ G -measurable with G D .Bt Bs / and we have
EŒ f .Bt /jFs D EŒ f .Bs C .Bt Bs //jFs

(4.10)
D EŒ .Bs ; Bt Bs /jFs D ˚.Bs / ;
where ˚.x/ D EŒ f .x C Bt Bs /. Note that the r.v. EŒ f .Bt /jFs just obtained
turns out to be a function of Bs and is therefore .Bs /-measurable. Hence

EŒ f .Bt /jBs DE EŒ f .Bt /jFs jBs D EŒ˚.Bs /jBs D ˚.Bs / D EŒ f .Bt /jFs ;
i.e. the conditional expectation knowing the position at time s is the same as
the conditional expectation knowing the entire past of the process up to time
s. We shall see later that this means that the Brownian motion is a Markov
process.
It is also possible to make explicit the function ˚: as xCBt Bs N.x; .t
s/I/,
Z h jy xj2 i
1
˚.x/ D EŒ f .x C Bt Bs / D f .y/ exp dy :
Œ2.t s/m=2 2.t s/
Example 4.5 (The position of a Brownian motion at a random time) Let B

be an m-dimensional Brownian motion and a positive r.v. independent of B.
How can we compute the law of the r.v. B ? Lemma 4.1 provides a simple
way of doing this. We assume, for simplicity, > 0 a.s.
Let A Rm be a Borel set. We have

P.B 2 A/ D EŒ1A .B / D E EŒ1A .B /j./ :
This is a very common trick: instead of computing the expectation directly

we first compute a conditional expectation and then its expectation. In the
computation of the conditional expectation above we are in the situation of
Lemma 4.1 as B is a function of the r.v. , which is ./-measurable, and
of the Brownian motion, which is independent of ./. In order to apply the
freezing lemma let A 2 B.Rm / and let us define .t; !/ D 1A .Bt .!//, so that
1A .B / D .; /. Therefore

E EŒ1A .B /j./ D EŒ˚./
(continued )

where
Z jxj2
1
˚.t/ D EŒ .t; / D EŒ1A .Bt / D e 2t dx :
.2t/m=2 A
By Fubini’s theorem and denoting the law of by ,

Z C1
P.B 2 A/ D EŒ1A .B / D EŒ˚./ D ˚.t/ d.t/
0
Z C1 Z jxj2
1 2t
D d.t/ e dx
.2t/m=2 (4.11)
Z0 Z C1 A jxj2
1
D dx e 2t d.t/
0 .2t/m=2
A
„ ƒ‚ …
WDg.x/
so that the density g of B is obtained by computing the integral.

One could also compute the characteristic function of B in quite a similar
way: if .t; !/ D eih
;Bt .!/i , we can write

EŒeih
;B i D E EŒ .; !/j./ D EŒ˚./
t 2
where now ˚./ D EŒ .t; !/ D EŒeih
;Bt i D e 2 j
j . Therefore
Z C1 t
2
EŒeih
;B i D e 2 j
j d.t/ : (4.12)
0
In order to determine the law of B we can therefore choose whether to

compute the density using (4.11) or the characteristic function using (4.12),
according to the difficulty of the computation of the corresponding integral.
Exercises 4.6 and 4.8 give some important examples of the application of this
technique.
Remark 4.5 Sometimes we are confronted with the computation of some-

thing of the kind
hZ T ˇ i
E Xt dt ˇ D ; (4.13)
0
where X is some integrable measurable process and we are tempted to write

hZ T ˇ i
Z T
E Xt dt ˇ D D EŒXt jD dt : (4.14)
0 0
(continued )

Is this correct? It is immediate that if W is a bounded D-measurable r.v. then
h Z T i Z T Z T
E W Xt dt D EŒWXt dt D E WEŒXt jD dt

0 0 0
h Z T i
DE W EŒXt jD dt :
0
There are, however, a couple of issues to be fixed. The first is that the quantity
EŒXt jD is, for every t, defined only a.s. and we do not know whether it has a
version such that t 7! EŒXt jD is integrable.
The second is that we must also prove that the r.v.
Z T
ZD EŒXt jD dt (4.15)
0
is actually D-measurable.
In practice, without looking for a general statement, these question are
easily handled: very often we shall see that t 7! EŒXt jD has a continuous
version so that the integral in (4.15) is well defined. In this case, the r.v. Z
of (4.15) is also D-measurable, as it is the limit of Riemann sums of the form
X
EŒXti jD.tiC1 ti / ;
which are certainly D-measurable.
Example 4.6 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. Com-

pute, for s t,
hZ t ˇ i
ˇ
E Bu du ˇ Fs :
0
We have EŒBu jFs D Bs if u s and EŒBu jFs D Bu if u s (as Bu is then

already Fs -measurable), i.e.
EŒBu jFs D Bsû :
This is a continuous process, so that we can apply formula (4.14), which gives
hZ t ˇ i Z t Z t Z s
ˇ
E Bu du ˇ Fs D EŒBu jFs du D Bsû du D Bu du C .t s/Bs :
0 0 0 0
4.3 Conditional laws
At the beginning of this chapter we spoke of conditional distributions given the value
of some discrete r.v. Z, but then we investigated the conditional expectations, i.e. the
expectations of these conditional distributions. Let us go back now to conditional
distributions.
Let Y; X be r.v.’s with values in the measurable spaces .G; G / and .E; E /,
respectively, and let us denote by Y the law of Y.
A family of probabilities .n.y; dx//y2G on .E; E / is a conditional law of X

given Y if,
i) For every A 2 E , the map y 7! n.y; A/ is G -measurable.
ii) For every A 2 E and B 2 G ,
Z
P.X 2 A; Y 2 B/ D n.y; A/ Y .dy/ : (4.16)
B
Intuitively n.y; dx/, which is a probability on .E; E /, is the law that it is suitable
to consider for the r.v. X, keeping into account the information that Y D y. (4.16)
implies that, if f W E ! R and g W G ! R are functions that are linear combinations
of indicator functions, then
Z Z
EŒ f .X/ .Y/ D f .x/ n.y; dx/ .y/ Y .dy/ : (4.17)
G E
Actually, (4.17) coincides with (4.16) if f D 1A and D 1B . Therefore, by lin-

earity, (4.17) is true if f and are linear combinations of indicator functions. With
the usual approximation arguments of measure theory, such as Proposition 1.11 for
example, we have that (4.17) holds for every choice of positive functions f ; and
then, decomposing into the sum of the positive and negative parts, for every choice
of functions f ; such R that f .X/ .Y/ is integrable.
If we set g.y/ D E f .x/ n.y; dx/, then (4.17) can be written as
Z
EŒ f .X/ .Y/ D g.y/ .y/ Y .dy/ D EŒg.Y/ .Y/ :
G
A comparison with (4.4) shows that this means that .n.y; dx//y2G is a conditional
distribution of X given Y if and only if, for every bounded measurable function f
4.3 Conditional laws 99
such that f .X/ is integrable,

Z
EŒ f .X/jY D g.Y/ D f .x/ n.Y; dx/ a.s. (4.18)
E
i.e. EŒ f .X/jY D y D g.y/. In particular, if X is an integrable real r.v., choosing

f .x/ D x we have
Z
EŒX jY D x n.Y; dx/ a.s. (4.19)
R
R
i.e. EŒX jY D y D R x n.y; dx/. We then find the conditional expectation as the
mean of the conditional law.
The next example highlights a general situation where the computation of the
conditional law is easy.
Example 4.7 Let X; Y be respectively Rd - and Rm -valued r.v.’s having joint

density h.x; y/ with respect to the Lebesgue measure of Rd Rm . Let
Z
hY .y/ D h.x; y/ dx
Rd
be the marginal densityRof Y (see Sect. 1.4) and let Q D fyI hY .y/ D 0g.
Obviously P.Y 2 Q/ D Q hY .y/ dy D 0. If we define
8
< h.x; y/ if y 62 Q
h.xI y/ D hY .y/ (4.20)
:
any arbitrary density if y 2 Q ;
we have immediately that n.y; dx/ D h.xI y/ dx is a conditional law of X

given Y D y: if f and g are bounded measurable functions on Rd and Rm
respectively, then
Z Z
EŒ f .X/g.Y/ D f .x/g.y/h.x; y/ dy dx
Rm Rd
Z Z
D g.y/hY .y/ dy f .x/h.xI y/ dx :
Rm Rd
Therefore, for every measurable function f such that f .X/ is integrable,

Z
EŒ f .X/jY D y D f .x/h.xI y/ dx ;
Rm
(continued )

which allows us to compute explicitly the conditional expectation in many
situations.
Note, however, that we have proved that the conditional expectation EŒX jY
always exists whereas, until now at least, we know nothing about the existence of a
conditional distribution of X given Y D y.
4.4 Conditional laws of Gaussian vectors
In this section we see a particular case of computation of conditional laws (hence

also of conditional expectations) when the r.v. X (whose conditional law we want
to compute) and Y (the conditioning r.v.) are jointly Gaussian. In order to do this
it is also possible to use the procedure of Example 4.7 (taking the quotient of the
joint distribution and the marginal), but the method we are about to present is much
more convenient (besides the fact that a joint density of X and Y does not necessarily
exist).
Let X; Y be Gaussian vectors with values in Rk and Rp respectively. Let
us assume that their joint law on .RkCp ; B.RkCp // is Gaussian with mean and
covariance matrix given respectively by

mX CX CXY
and ;
mY CYX CY
where CX ; CY are the covariance matrices of X and Y, respectively, and
CXY D EŒ.X EŒX/ .Y EŒY/ D CYX

is the k p matrix of the covariances of the components of X and of those of Y; let

us assume, moreover, that CY is strictly positive definite (and therefore invertible).
Let us look first for a k p matrix, A, such that X AY and Y are independent. Let
Z D X AY. The pair .X; Y/ is Gaussian and the same is true for .Z; Y/ which is a
linear function of .X; Y/. Therefore, for the independence of Z and Y, it is sufficient
to check that Cov.Zi ; Yj / D 0 for every i D 1; : : : ; k; j D 1; : : : ; p. In order to keep
the notation simple, let us first assume that the means mX and mY vanish so that the
covariance coincides with the expectation of the product. The condition of absence
of correlation between the components of Z and Y can then be expressed as
0 D EŒZY D EŒ.X AY/Y D EŒXY AEŒYY D CXY ACY :

4.4 Conditional laws of Gaussian vectors 101
Therefore the required property holds with A D CXY CY1 . If we remove the
hypothesis that the means vanish, it suffices to repeat the same computation with
X mX and Y mY instead of X and Y. Thus we can write
X D AY C .X AY/ ;
where X AY and Y are independent. It is rather intuitive now (see Exercise 4.10
for a rigorous, albeit simple, verification) that the conditional law of X given Y D y
is the law of Ay C X AY, since, intuitively, the knowledge of the value of Y does
not give any information on the value of X AY. As X AY is Gaussian, this law is
characterized by its mean
Ay C mX AmY D mX C CXY CY1 .y mY /
and covariance matrix (recall that A D CXY CY1 )
CXAY D EŒ..X mX / A.Y mY //..X mX / A.Y mY //

D CX CXY A ACYX C ACY A
D CX CXY CY1 CXY

CXY CY1 CXY

C CXY CY1 CY CY1 CXY

1
D CX CXY CY CXY ;
where we took advantage of the fact that CY is symmetric and of the relation CYX D

CXY . In conclusion
the conditional distribution of X given Y D y is Gaussian with mean
mX C CXY CY1 .y mY / (4.21)
and covariance matrix
CX CXY CY1 CXY

: (4.22)
In particular, from (4.21), we obtain the value of the conditional expectation,

which is equal to
EŒX jY D mX C CXY CY1 .Y mY / :
If X and Y are both one-dimensional, this formula becomes
Cov.X; Y/
EŒX jY D mX C .Y mY / ; (4.23)
Var.Y/
whereas the variance (4.22) of the conditional distribution of X given Y D y is
Cov.X; Y/2
Var.X/ (4.24)
Var.Y/
and is therefore always smaller that the variance of X. Let us point out two important
things in the previous computation:
• The conditional laws of a Gaussian vector are Gaussian themselves.

• Only the mean of the conditional law depends on the value y of the
conditioning r.v. Y. The covariance matrix of the conditional law does not
depend on the value of Y and can therefore be computed before knowing
the observation Y.
4.5 The augmented Brownian filtration
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional (continuous) Brownian

motion. Let us consider its natural filtration Gt D .Bs ; s t/ and the aug-
mented natural filtration G D .Gt ; N / (see p. 32). We know that B is still a
Brownian motion if considered with respect to .Gt /t and also with respect to .G t /t
(Remark 3.1). In this section we show that this filtration is right-continuous and
therefore .˝; F ; .G t /t ; .Bt /t ; P/ is a standard Brownian motion. More precisely, let
\ _
G sC D G u; G s D Gu :
u>s u<s
It is immediate that Gs D Gs . Indeed, as B is continuous, we have
lim Bu D Bs :
u!s
Hence the r.v. Bs , being the limit of Gs -measurable r.v.’s, is Gs -measurable itself,
which implies Gs Gs . More precisely, we have the following
Proposition 4.3 G s D G s D G sC .
4.5 The augmented Brownian filtration 103
Proof Of course, we only need to prove the rightmost equality. We shall assume for
simplicity m D 1. It is sufficient to show that for every bounded G 1 -measurable
r.v. W
EŒW jG sC D EŒW jG s a.s. (4.25)
This relation, applied to a r.v. W that is already G sC -measurable, will imply that
it is also G s -measurable and therefore that G s G sC . As the reciprocal inclusion
is obvious, the statement will be proved. Note that here we use the fact that the -
algebras G s contain all the negligible events of G 1 : thanks to this fact, if a r.v. is
a.s. equal to a r.v. which is G s -measurable, then it is G s -measurable itself.
We have by Lebesgue’s theorem, for t > s,
h ˇ i
EŒei˛Bt jG sC D ei˛Bs EŒei˛.Bt Bs / jG sC D ei˛Bs E lim ei˛.Bt BsC" / ˇ G sC
"!0
i˛.Bt BsC" /
De i˛Bs
lim EŒe jG sC ;
"!0
as ei˛.Bt BsC" / ! ei˛.Bt Bs / as " ! 0 and these r.v.’s are bounded by 1. As, for " > 0,
Bt BsC" is independent of GsC" and a fortiori of GsC ,
EŒei˛Bt jG sC D ei˛Bs lim EŒei˛.Bt BsC" / jG sC D ei˛Bs lim EŒei˛.Bt BsC" /

"!0 "!0
12 ˛ 2 .ts"/ i˛Bs 12 ˛ 2 .ts/

D ei˛Bs lim e De e :
"!0
Hence, the right-hand side above being Gs -measurable, we have, for t 2 RC , ˛ 2 R,
EŒei˛Bt jG sC D EŒei˛Bt jG s a.s. (4.26)
(if t s this relation is obvious).

We now prove that (4.26) implies that (4.25) holds for every bounded G 1 -
measurable r.v. W. Let us denote by H the vector space of the real r.v.’s Z such that
E.Z jG sC D EŒZ jG s a.s. We have seen that H contains the linear combinations
of r.v.’s of the form <ei˛Bt and =ei˛Bt , t 2 RC , which is a subspace stable under
multiplication; it also contains the r.v.’s that vanish a.s., since for such r.v.’s both
EŒZ jG sC and EŒZ jG s vanish a.s. and are G t -measurable for every t, as the -
algebras G t contain all negligible events.
Hence H satisfies the hypotheses of Theorem 1.5 and therefore contains every
bounded r.v. that is measurable with respect to the smallest -algebra, A , generated
by the r.v.’s ei˛Bt , ˛ 2 R, t 2 RC and by the negligible events of F . But, as
ei˛Bt ei˛Bt
Bt D lim ;
˛!0 2i˛
the r.v.’s Bt are A -measurable and thus A G t for every t 0, hence A G 1 ,

which concludes the proof. u
t
Corollary 4.1 B D .˝; F ; .G t /t ; .Bt /t ; P/ is a standard Brownian motion.
Exercises
4.1 (p. 474) Let X be a real r.v. defined on a probability space .˝; F ; P/ and
G F a sub--algebra. Let D F be another -algebra independent of X and
independent of G .
a) Is it true that
EŒX jG _ D D EŒX jG ‹ (4.27)
b) Show that if D is independent of .X/ _ G , then (4.27) is true.

Use the criterion of Remark 4.2.
4.2 (p. 475) On .˝; F ; P/ let G F be a sub--algebra and Y a r.v. independent

of G . Show that Y cannot be G -measurable unless it is a.s. constant.
4.3 (p. 476) Let P and Q be probabilities on .˝; F / and let us assume that Q
has a density, Z, with respect to P, i.e. Q.A/ D E.1A Z/ for every A 2 F . Let
X W .˝; F / ! .E; E / be a r.v. and let P D X.P/, Q D X.Q/ be the laws of X
d
with respect to P and Q, respectively. Prove that P Q and that dQP D f , where
f .x/ D EŒZ jX D x (E denotes the expectation with respect to P).
4.4 (p. 476) (Conditioning under a change of probability) Let Z be a real positive
r.v. defined on .˝; F ; P/ and G a sub--algebra of F .
a) Prove that, a.s., fZ D 0g fEŒZ jG D 0g (i.e. by conditioning the set of zeros
of a positive r.v. shrinks). Prove that for every r.v. Y such that YZ is integrable,
EŒZY jG D EŒZY jG 1fE.Z jG />0g a.s. (4.28)
b) Let us assume, moreover, that EŒZ D 1. Let Q be the probability on .˝; F /

having density Z with respect to P and let EQ denote the expectation with respect
to Q. Show that EŒZ jG > 0 Q-a.s. (E still denotes the expectation with respect
to P). Show that, if Y is Q-integrable,
EŒYZ jG
EQ ŒY jG D Qa:s: (4.29)
EŒZ jG
a) Try assuming Y 0 first.
4.5 (p. 477) Let D F be a sub--algebra and X an m-dimensional r.v. such that
for every 2 Rm
EŒeih ;Xi jD D EŒeih ;Xi a.s.
Then X is independent of D.
4.6 (p. 477) Let B be an m-dimensional Brownian motion and an exponential r.v.
with parameter and independent of B.
a) Compute the characteristic function of B (the position of the Brownian motion
at the random time /.
b1) Let X be a real r.v. with a Laplace density with parameter , i.e. with density
jxj
fX .x/ D e :
2
Compute the characteristic function of X.
b2) What is the law of B for m D 1?
4.7 (p. 478) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and let W
˝ ! RC be a positive r.v. (not necessarily a stopping time of .Ft /t ) independent
of B.
Prove that Xt D BCt B is a Brownian motion and specify with respect to
which filtration.
4.8 (p. 478) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a two-dimensional Brownian
motion. Let a > 0 and let D infftI B2 .t/ D ag be the passage time of B2 at
a, which is also the entrance time of B in the line y D a. Recall (Sect. 3.6) that
< C1 a.s.
a) Show that the -algebras ./ and G1 D .B1 .u/; u 0/ are independent.
b) Compute the law of B1 ./ (i.e. the law of the abscissa of B at the time it reaches
the line y D a).
b) Recall Example 4.5. The law of is computed in Exercise 3.20. . .
4.9 (p. 479) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. Compute
Z t ˇ Z t
E B2u du ˇ Fs and E B2u dujBs :
s s
4.10 (p. 479) Let .E; E /, .G; G / be measurable spaces, X an E-valued r.v. such
that
X D .Y/ C Z ;
where Y and Z are independent r.v.’s with values respectively in E and G and where
W G ! E is measurable.
Show that the conditional law of X given Y D y is the law of Z C .y/.
4.11 (p. 480)
a) Let X be a signal having a Gaussian N.0; 1/ law. An observer has no access to
the value of X and only knows an observation Y D X C W, where W is a noise,
independent of X and N.0; 2 /-distributed. What is your estimate of the value X
of the signal knowing that Y D y?
b) The same observer, in order to improve its estimate of the signal X, decides to
take two observations Y1 D X C W1 and Y2 D X C W2 , where W1 and W2 are
N.0; 2 /-distributed and the three r.v.’s X; W1 ; W2 are independent. What is the
estimate of X now given Y1 D y1 and Y2 D y2 ? Compare the variance of the
conditional law of X given the observation in situations a) and b).
4.12 (p. 481)

a) Let B be a Brownian motion. What is the conditional law of .Bt1 ; : : : ; Btm / given
B1 D y, 0 t1 < < tm < 1?
b) What is the conditional law of .Bt1 ; : : : ; Btm / given B1 D y; Bv D x, 0 t1 <
< tm < 1 < v?
4.13 (p. 483) Let B be a Brownian motion. What is the joint law of
Z 1
B1 and Bs ds ‹
0
R1
Let us assume we know that 0 Bs ds D x. What is the best estimate of the position
B1 of the Brownian motion at time 1?
The joint law is Gaussian, see Exercise 3.11.
4.14 (p. 483) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and an
N.; 2 /-distributed r.v. independent of .Bt /t . Let
Yt D t C Bt
and Gt D .Ys ; s t/. Intuitively the meaning of this exercise is the following:
starting from the observation of a path Ys .!/; s t, how can the unknown value of
.!/ be estimated? How does this estimate behave as t ! 1? Will it converge to
.!/?
a) Compute Cov.; Ys /, Cov.Ys ; Yt /.
b) Show that .Yt /t is a Gaussian process.
c) Prove that, for every t 0, there exists a number (depending on t) such that
D Yt C Z, with Z independent of Gt .
d) Compute EŒjGt and the variance of EŒjGt . Show that
lim EŒjGt D a.s.

t!C1
4.15 (p. 484) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a natural Brownian motion and, for
0 t 1, let
Xt D Bt tB1 :
The process .Xt /t , defined for 0 t 1, is called the Brownian bridge .

a) Show that .Xt /t is a centered Gaussian process independent of B1 . Compute
E.Xt Xs /.
b) Show that, for s t, the r.v.
1t
Xt Xs
1s
is independent of Xs .
c) Compute EŒXt jXs and show that, with Gs D .Xu ; u s/, for s t
EŒXt jGs D EŒXt jXs :
d) Compute EŒXt jFs . Do the -algebras Fs and Gs coincide?

e) Compute the finite-dimensional distributions of .Xt /t (0 t 1) and show that
they coincide with the finite-dimensional distributions of .Bt /t conditioned given
B1 D 0.
Chapter 5
Martingales
Martingales are stochastic processes that enjoy many important, sometimes surpris-
ing, properties. When studying a process X, it is always a good idea to look for
martingales “associated” to X, in order to take advantage of these properties.
5.1 Definitions and general facts
Let T RC .
Definition 5.1 A real-valued process M D .˝; F ; .Ft /t2T ; .Mt /t2T ; P/ is a

martingale (resp. a supermartingale, a submartingale) if Mt is integrable for
every t 2 T and
E.Mt jFs / D Ms .resp. Ms ; Ms / (5.1)
for every s t.
When the filtration is not specified it is understood to be the natural one.
Examples 5.1
a) If T D N and .Xk /k is a sequence of independent real centered r.v.’s, and
Yn D X1 C C Xn , then .Yn /n is a martingale.
Indeed let Fm D .Y1 ; : : : ; Ym / and also observe that Fm D
.X1 ; : : : ; Xm /. If n > m, as we can write Yn D Ym C XmC1 C C Xn
(continued )

110 5 Martingales
Examples 5.1 (continued)

and, as the r.v.’s XmC1 ; : : : ; Xn are centered and independent of Fm , we
have
E.Yn jFm / D E.Ym jFm / C E.Xn C C XmC1 jFm /

D Ym C E.Xn C C XmC1 / D Ym :
b) Let X be an integrable r.v. and .Fn /n a filtration, then Xn D E.X jFn / is a

martingale. Indeed if n > m, then Fm Fn and
ˇ
E.Xn jFm / D E E.X jFn / ˇ Fm D E.X jFm / D Xm
thanks to Proposition 4.1 e).

c) If B D .˝; F ; .Ft /t ; .Bt /t ; P/ is a Brownian motion, then .Bt /t is a .Ft /t -
martingale, as we saw in Example 4.3.
It is clear that linear combinations of martingales are also martingales and linear
combinations with positive coefficients of supermartingales (resp. submartingales)
are still supermartingales (resp. submartingales). If .Mt /t is a supermartingale, then
.Mt /t is a submartingale and vice versa.
Moreover, if M is a martingale (resp. a submartingale) and ˚ W R ! R is a
convex (resp. increasing convex) function such that ˚.Mt / is also integrable for
every t, then .˚.Mt //t is a submartingale: it is a consequence of Jensen’s inequality,
Proposition 4.2 d). Indeed, if s t,

EŒ˚.Mt /jFs ˚ EŒMt jFs D ˚.Ms / :
In particular, if .Mt /t is a martingale then .jMt j/t is a submartingale (with respect to

the same filtration).
We shall say that a martingale (resp. supermartingale, submartingale) .Mt /t is in
Lp ; p 1, if Mt 2 Lp for every t. We shall speak of square integrable martingales
(resp. supermatingales, submartingales) if p D 2. If .Mt /t is a martingale in Lp ,
p 1, then .jMt jp /t is a submartingale.
Note that it is not true, in general, that if M is a submartingale, then also .jMt j/t
and .Mt2 /t are submartingales (even if M is square integrable): the functions x 7! jxj
and x 7! x2 are convex but not increasing. If M is square integrable .Mt2 /t will,
however, be a submartingale under the additional assumption that M is positive: the
function x 7! x2 is increasing when restricted to RC .
5.2 Discrete time martingales 111
5.2 Discrete time martingales
In this and in the subsequent sections we assume T D N.
A process .An /n is said to be an increasing predictable process for a filtration

.Fn /n if A0 D 0, An AnC1 a.s. and AnC1 is Fn -measurable for every n 0.
As Fn FnC1 , .An /n is, in particular, adapted but, intuitively, it is even a process

such that at time n we know its value at time n C 1 (which accounts for the term
“predictable”). Let .Xn /n be an .Fn /n -submartingale and .An /n the process defined
recursively as
A0 D 0; AnC1 D An C E.XnC1 jFn / Xn : (5.2)

„ ƒ‚ …
0
.An /n is clearly increasing and, as AnC1 is the sum of Fn -measurable r.v.’s, by

construction it is an increasing predictable process. As AnC1 is Fn -measurable, (5.2)
can be rewritten as
E.XnC1 AnC1 jFn / D E.XnC1 jFn / AnC1 D Xn An ;
i.e. Mn D Xn An is an .Fn /n -martingale, that is we have decomposed X into the

sum of a martingale and of an increasing predictable process.
If Xn D Mn0 C A0n were another decomposition of .Xn /n into the sum of a
martingale M 0 and of an increasing predictable process A0 , we would have
A0nC1 A0n D XnC1 Xn .MnC1

0
Mn0 / :
By conditioning with respect to Fn , we have A0nC1 A0n D E.XnC1 jFn / Xn D

AnC1 An ; as A0 D A00 D 0, by recurrence it follows that A0n D An and Mn0 D Mn .
We have therefore proved that
Theorem 5.1 Every submartingale .Xn /n can be decomposed uniquely into

the sum of a martingale .Mn /n and an increasing predictable process .An /n .
This is Doob’s decomposition and the process A is the compensator of .Xn /n .

If .Mn /n is a square integrable martingale, then .Mn2 /n is a submartingale. Its
compensator is called the associated increasing process of .Mn /n .
112 5 Martingales
If .Xn /n is a martingale, the same is true for the stopped process Xn D Xn^ ,

where is a stopping time of the filtration .Fn /n . Indeed as XnC1 D Xn on f ng,

XnC1 Xn D .XnC1

Xn /1f ng C .XnC1

Xn /1f nC1g D .XnC1

Xn /1f nC1g
Hence, as by the definition of a stopping time f n C 1g D f ngc 2 Fn , we

have

E.XnC1 Xn jFn / D EŒ.XnC1 Xn /1f nC1g jFn
(5.3)
D 1f nC1g E.XnC1 Xn jFn / D 0 :
Similarly it is proved that if .Xn /n is a supermartingale (resp. a submartingale) then

the stopped process .Xn /n is again a supermartingale (resp. a submartingale).
The following is the key result from which many properties of martingales
follow.
Theorem 5.2 (Stopping theorem) Let X D .˝; F ; .Fn /n ; .Xn /n ; P/ be a

supermartingale and 1 ; 2 two a.s. bounded stopping times of the filtration
.Fn /n such that 1 2 a.s. Then the r.v.’s X1 ; X2 are integrable and
E.X2 jF1 / X1 :
is immediate, as, for i D 1; 2 and denoting by k a

Proof Integrability of X1 and X2 P
number that majorizes 2 , jXi j kjD1 jXj j. Let us assume, at first, that 2 k 2 N
and let A 2 F1 . As A \ f1 D jg 2 Fj , we have for j k
EŒX1 1A\f1 Djg D EŒXj 1A\f1 Djg EŒXk 1A\f1Djg
and, taking the sum over j; 0 j k,
X
k X
k
EŒX1 1A D EŒXj 1A\f1 Djg EŒXk 1A\f1 Djg D EŒX1 1A :
jD0 jD0
We have therefore proved the statement if 2 is a constant stopping time. Let us now
remove this hypothesis and assume 2 k. If we apply the result proved in the first
part of the proof to the stopped martingale .Xn2 /n and to the stopping times 1 and k
we obtain
EŒX1 1A D EŒX12 1A EŒXk2 1A D EŒX2 1A ;

t
u
5.2 Discrete time martingales 113
Thanks to Theorem 5.2 applied to X and X we have, of course,
Corollary 5.1 If X is a martingale and 1 ; 2 are bounded stopping times such

that 1 2 , then
E.X2 jF1 / D X1 a:s:
Let us point out that the assumption of boundedness of the stopping times in
the two previous statements is essential and that it is easy to find counterexamples
showing that the stopping theorem does not hold under the weaker assumption that
1 and 2 are only finite.
Remark 5.1 In particular, if X is a martingale, Corollary 5.1 applied to 2 D

and 1 D 0 gives
EŒX D EŒE.X jF0 / D EŒX0
for every bounded stopping time . Hence the quantity E.X / is constant as
ranges among the bounded stopping times. In fact this condition is also
sufficient for X to be a martingale (Exercise 5.6).
Theorem 5.3 (Maximal inequalities) Let X be a supermartingale and >

0; then

P sup Xi E.X0 / C E.Xk / ; (5.4)
0ik

P inf Xi EŒXk 1finf0ik Xi g E.jXk j/ : (5.5)
0ik
Proof Let
.!/ D inffnI n k; Xn .!/ g

114 5 Martingales
with the understanding that D k if f g D ;. is a finite stopping time and we have

X on the event fsup0ik Xi g. By Theorem 5.2 applied to 2 D , 1 D 0,
E.X0 / E.X / D EŒX 1fsup0ik Xi g C EŒXk 1fsup0ik Xi < g

P sup Xi E.Xk / ;
1ik
from which we obtain (5.4). If conversely
.!/ D inffnI n k; Xn .!/ g
and .!/ D k if f g D ;, then is still a bounded stopping time and, again by

Theorem 5.2,
E.Xk / E.X / D EŒX 1finf0ik Xi g C EŒXk 1finf0ik Xi > g

P inf Xi C EŒXk 1finf0ik Xn > g
0ik
from which we obtain (5.5).

t
u
5.3 Discrete time martingales: a.s. convergence
One of the reasons why martingales are important is the result of this section; it
guarantees, under rather weak (and easy to check) hypotheses, that a martingale
converges a.s.
Let a b and Xm .!/ b for
m D i C 1; : : : ; j 1. Let
a;b
k
.!/ D number of upcrossings of .Xn .!//nk over the interval Œa; b :
The proof of the theorem of convergence that we have advertised is a bit technical,
but the basic idea is rather simple: in order to prove that a sequence is convergent
one first needs to prove that it does not oscillate too much. For this reason the key
estimate is the following, which states that a supermartingale cannot make too many
upcrossings.
5.3 Discrete time martingales: a.s. convergence 115
Proposition 5.1 If X is a supermartingale, then
.b a/E.a;b
k
/ EŒ.Xk a/ :
Proof Let us define the following sequence of stopping times
1 .!/ D inffiI i k; Xi .!/ < ag

2 .!/ D inffiI 1 .!/ bg
:::
2m1 .!/ D inffiI 2m2 .!/ bg
with the understanding that i D k if f g D ;. Let
˝2m D f2m k; X2m > bg D fa;b

k
mg ;
˝2m1 D fa;b
k
m 1; X2m1 < ag :
It is immediate that ˝i 2 Fi , as i and Xi are Fi -measurable. Moreover, X2m1 <

a on ˝2m1 and X2m > b on ˝2m . By Theorem 5.2, applied to the bounded stopping
times 2m1 and 2m ,
Z Z
0 .X2m1 a/ dP .X2m a/ dP
˝2m1 ˝2m1
Z
.b a/ P.˝2m / C .Xk a/ dP ;
˝2m1 n˝2m
where we take advantage of the fact that 2m D k on ˝2m1 n ˝2m and that X2m > b
on ˝2m . Therefore
Z
.b a/P.˝2m / D .b a/P.a;b m/
k
.Xk a/ dP : (5.6)
˝2m1 n˝2m
Let us recall that (Exercise 1.3)

1
X
k
E.a;b /D k
P.a;b m/ :
mD1
116 5 Martingales
As the events ˝2m1 n ˝2m are pairwise disjoint, taking the sum in m in (5.6) we
have
1
X
.b a/E.a;b
k
/ D .b a/ k
P.a;b m/ EŒ.Xk a/ :
mD1
t
u
Theorem 5.4 Let X be a supermartingale such that
sup E.Xn / < C1 : (5.7)

n0
Then it converges a.s. to a finite limit.
Proof For fixed a < b let us denote by a;b .!/ the number of upcrossings of the
path .Xn .!//n over the interval Œa; b. As .Xn a/ aC C Xn , by Proposition 5.1,
1
E.a;b / D lim E.a;b
k
/ sup EŒ.Xn a/
k!1 b a n0
1 C (5.8)
a C sup E.Xn / < C1 :
ba n0
In particular, a;b < C1 a.s., i.e. a;b < C1 outside a negligible event Na;b ;
considering the union of these negligible events Na;b for every possible a; b 2 Q
with a < b, we can assume that outside a negligible event N we have a;b < C1
for every a, b 2 R.
Let us prove that if ! … N then the sequence .Xn .!//n necessarily has a limit.
Indeed, if this was not the case, let a D limn!1 Xn .!/ < limn!1 Xn .!/ D b. This
implies that .Xn .!//n is close to both a and b infinitely many times. Therefore if
˛; ˇ are such that a < ˛ < ˇ < b we would have necessarily ˛;ˇ .!/ D C1,
which is not possible outside N.
The limit is, moreover, finite. In fact from (5.8)
lim E.a;b / D 0
b!C1
but a;b .!/ is a non-increasing function of b and therefore
lim a;b .!/ D 0 a.s.

b!C1
5.3 Discrete time martingales: a.s. convergence 117
As a;b takes only integer values, a;b .!/ D 0 for large b and .Xn .!//n is therefore
bounded above a.s. Similarly one sees that it is bounded below.
t
u
In particular,
Corollary 5.2 Let X be a positive supermartingale. Then the limit
lim Xn
n!1
exists a.s. and is finite.
Example 5.1 Let .Zn /n be a sequence of i.i.d. r.v.’s taking the values ˙1 with
probability 12 and let X0 D 0 and Xn D Z1 C C Zn for n 1. Let a; b
be positive integers and let D inffnI Xn b or Xn ag, the exit time of
X from the interval a; bŒ. Is < C1 with probability 1? In this case we
can define the r.v. X , which is the position of X when it leaves the interval
a; bŒ. Of course, X can only take the values a or b. What is the value of
P.X D b/?
We know (as in Example 5.1 a)) that X is a martingale. Also .Xn^ /n is
a martingale, which is moreover bounded as it can take only values that are
a and b.
By Theorem 5.4 the limit limn!1 Xn^ exists and is finite. This implies
that < C1 a.s.: as at every iteration X makes steps of size 1 to the right
or to the left, on D C1 we have jX.nC1/^ Xn^ j D 1, so that .Xn^ /n
cannot be a Cauchy sequence.
Therefore necessarily < C1 and the r.v. X is well defined. In order
to compute P.X D b/, let us assume for a moment that we can apply
Theorem 5.2, the stopping theorem, to the stopping times 2 D and 1 D 0
(we cannot because is finite but not bounded), then we would have
0 D EŒX0 D EŒX : (5.9)
From this relation, as P.X D a/ D 1 P.X D b/, we obtain
0 D EŒX D bP.X D b/ aP.X D a/ D bP.X D b/ a.1 P.X D b//
i.e.
a
P.X D b/ D
aCb
(continued )
118 5 Martingales

The problem is therefore solved if (5.9) is satisfied. Let us prove it: for every
n the stopping time ^ n is bounded, therefore the stopping theorem gives
0 D EŒX ^n :
Now observe that limn!1 X ^n D X and that, as a X ^n b, the

sequence .X ^n /n is bounded, so that we can apply Lebesgue’s theorem and
obtain (5.9).
This example shows a typical application of the stopping theorem in order
to obtain the distribution of a process stopped at some stopping time, and also
how to apply the stopping theorem to stopping times that are not bounded:
just apply the stopping theorem to the stopping times ^ n, that are bounded,
and then pass to the limit using Lebesgue’s theorem as in this example or
some other statement, such as Beppo Levi’s theorem, in order to pass to the
limit.
Remark 5.2 Condition (5.7) for a supermartingale is equivalent to requiring

that X is bounded in L1 . Indeed, obviously,
Xk jXk j
so that, if supk EŒjXk j < C1, then also (5.7) holds. Conversely, note that
jXk j D Xk C 2Xk ;
hence
EŒjXk j D EŒXk C 2EŒXk EŒX0 C 2EŒXk ;
so that (5.7) implies boundedness in L1 .
5.4 Doob’s inequality; Lp convergence, the p > 1 case
We say that a martingale M is bounded in Lp if

sup E jMn jp < C1 :

n
5.4 Doob’s inequality; Lp convergence, the p > 1 case 119
Theorem 5.5 (Doob’s inequality) Let M D .˝; F ; .Fn /n ; .Mn /n ; P/ be a

martingale bounded in Lp with p > 1. Then if M D supn jMn j, M 2 Lp and
kM kp q sup kMn kp ; (5.10)

n
p
where q D p1 is the exponent conjugate to p.
Theorem 5.5 is a consequence of the following result, applied to the positive

submartingale .jMn j/n .
Lemma 5.1 If X is a positive submartingale, then for every p > 1 and n 2 N

p p
p
E max Xi E.Xnp / :
0in p1
Proof Let Y D max1in Xi and, for > 0,
.!/ D inffiI 0 i n; Xi .!/ > g

Pn
and .!/ D n C 1 if f g D ;. Then kD1 1f Dkg / D 1fY> g and, for every ˛ > 0,
Z Y Z C1
Yp D p p1 d D p p1 1fY> g d
0 0
Z C1 X
n (5.11)
Dp p1
1f Dkg d :
0 kD1
Moreover, if k n, as Xk on f D kg, we have 1f Dkg Xk 1f Dkg and
X
n X
n
p1 1f Dkg p2 Xk 1f Dkg : (5.12)
kD1 kD1
As X is a submartingale and 1f Dkg is Fk -measurable,
E.Xn 1f Dkg jFk / D 1f Dkg E.Xn jFk / Xk 1f Dkg
and, combining this relation with (5.12), we find
h X
n i h X
n i h X
n i
E p1 1f Dkg E p2 1f Dkg Xk E ˛2 Xn 1f Dkg :
kD1 kD1 kD1
120 5 Martingales
Thanks to this inequality, if we take the expectation of both members in (5.11),
h Z C1 X
n i
1 1
EŒY p E Xn p2 1f Dkg d D EŒXn Y p1 :
p 0 kD1
p1
„ ƒ‚ …
1
D p1 Y p1
Now by Hölder’s inequality (recall that we assume p > 1)

p
EŒY p EŒY p . p1/=p EŒXnp 1=p :
p1
As we know that E.Y p / < C1 (Y is the maximum of a finite number of r.v.’s of

Lp ), we can divide both terms by EŒY p . p1/=p , which gives
p
EŒY p 1=p EŒXnp 1=p ;
p1
from which the statement follows. t

u
Proof of Theorem 5.5 Lemma 5.1 applied to the submartingale Xi D jMi j gives the
inequality

max jMi j qkMn kp D q max kMk kp ;
0in p kn
where we used the fact that k 7! kMk kp is increasing as .jMn jp /n is a submartingale.

Doob’s inequality now follows by taking the limit as n ! 1 and using Beppo
Levi’s theorem. t
u
Thanks to Doob’s inequality (5.10), for p > 1, the behavior of a martingale
bounded in Lp is very nice and simple.
Theorem 5.6 If p > 1 a martingale is bounded in Lp if and only if it

converges a.s. and in Lp .
Proof It is immediate that a martingale M that converges in Lp is also bounded in

Lp .
Conversely, if a martingale M is bounded in Lp with p > 1, then
sup Mn sup jMn j D M :

n0 n0
5.5 Uniform integrability and convergence in L1 121
Since by Doob’s inequality M 2 Lp , we have supn E.Mn / E.M / < C1, so

that the condition (5.7) of Theorem 5.4 is satisfied and M converges a.s. Moreover,
its limit M1 belongs to Lp , as obviously jM1 j M .
Let us prove that the convergence takes place in Lp , i.e. that

lim E jMn M1 jp D 0 :
n!1
We already know that jMn M1 jp !n!1 0 so that we only need to find a bound in
order to apply Lebesgue’s theorem. Thanks to the inequality jx C yjp 2p1 .jxjp C
jyjp /, which holds for every x; y 2 R,
jMn M1 jp 2p1 .jMn jp C jM1 jp / 2p M

p
and as the r.v. M p is integrable by Doob’s maximal inequality, Lebesgue’s theorem

can be applied, giving
lim EŒjMn M1 jp D 0 :
n!1
t
u
We see in the next section that for L1 -convergence of martingales things are very
different.
5.5 Uniform integrability and convergence in L1
The notion of uniform integrability is the key tool for the investigation of L1
convergence of martingales.
Definition 5.2 A family of m-dimensional r.v.’s H is said to be uniformly

integrable if

lim sup E jYj1fjYj>cg D 0 : (5.13)

c!C1 Y2H
The set formed by a single integrable r.v. is the simplest example of a uniformly
integrable family. We have limc!C1 jYj1fjYj>cg D 0 a.s. and, as jYj1fjYj>cg jYj,
by Lebesgue’s theorem,

lim E jYj1fjYj>cg D 0 :
c!C1
122 5 Martingales
Similarly H turns out to be uniformly integrable if there exists an integrable real

r.v. Z such that Z jYj for every Y 2 H . Actually, in this case, for every Y 2 H ,
jYj1fjYj>cg jZj1fjZj>cg so that, for every Y 2 H ,

E jYj1fjYj>cg E Z1fZ>cg ;
from which (5.13) follows. Hence the following theorem is a extension of

Lebesgue’s theorem. For a proof, as well as a proof of the following Proposition 5.3,
see J. Neveu’s book (Neveu 1964, §II-5).
Theorem 5.7 Let .Yn /n be a sequence of r.v.’s converging a.s. to Y. In order

for Y to be integrable and for the convergence to take place in L1 it is necessary
and sufficient that .Yn /n is uniformly integrable.
In any case a uniformly integrable family H is bounded in L1 : let c > 0 be such

that

sup E jYj1fjYj>cg 1 ;
Y2H
then we have, for every Y 2 H ,

E.jYj/ D E jYj1fjYj>cg C E jYj1fjYjcg 1 C c :
The next two characterizations of uniform integrability are going to be useful.
Proposition 5.2 A family H is uniformly integrable if and only if the

following conditions are satisfied.
a) supY2H EŒjYj < C1.
b) For every " > 0 there exists a ı" > 0 such that, for every Y 2 H ,
EŒjYj1A "
for every A 2 F such that P.A/ ı" .
The proof is left as an exercise or, again, see J. Neveu’s book (Neveu 1964,
Proposition II-5-2).
5.5 Uniform integrability and convergence in L1 123
Proposition 5.3 A family H L1 is uniformly integrable if and only if

there exists a positive increasing convex function g W RC ! R such that
limt!C1 g.t/t D C1 and
sup EŒg.jYj/ < C1 :

Y2H
In particular, if H is bounded in Lp for some p > 1 then it is uniformly

integrable: just apply Proposition 5.3 with g.t/ D tp .
Proposition 5.4 Let Y be an integrable r.v. Then the family fE.Y jG /gG , for
G in the class of all sub--algebras of F , is uniformly integrable.
Proof As the family formed by the single r.v. Y is uniformly integrable, by

Proposition 5.3 there exists a positive increasing convex function g W RC ! R
such that limt!C1 g.t/
t
D C1 and EŒg.jYj/ < C1. By Jensen’s inequality,
Proposition 4.2 d),

E g.jE.Y jG /j/ E E.g.jYj/jG / D EŒg.jYj/ < C1
and, again by Proposition 5.3, fE.Y jG /gG is uniformly integrable.

t
u
In particular, if .Fn /n is a filtration on .˝; F ; P/ and Y 2 L1 , .E.Y jFn //n is a
uniformly integrable martingale (recall Example 5.1 b)).
Theorem 5.8 Let .Mn /n be a martingale. Then the following properties are
equivalent
a) .Mn /n converges in L1 ;
b) .Mn /n is uniformly integrable;
c) .Mn /n is of the form Mn D E.Y jFn / for some Y 2 L1 .˝; F ; P/.
If any of these conditions is satisfied then .Mn /n also converges a.s.
Proof If .Mn /n is a uniformly integrable martingale then it is bounded in L1 and

therefore the condition (5.7), supk E.Mk / < C1, is satisfied and M converges a.s.
to some r.v. Y. By Theorem 5.7, Y 2 L1 and the convergence takes place in L1 .
Therefore (Remark 4.3), for every m,
Mm D E.Mn jFm / ! E.Y jFm /

n!1
124 5 Martingales
in L1 . This proves that b))a) and a))c). c))b) is Proposition 5.4.

t
u
The following proposition identifies the limit of a uniformly integrable martingale.
an integrable r.v. on L1 .˝; F ; P/, .Fn /n a filtration

Proposition 5.5 Let Y beW
on .˝; F / and F1 D n Fn , the -algebra generated by the -algebras
.Fn /n . Then
lim E.Y jFn / D E.Y jF1 / a.s. and in L1 :

n!1
Proof Let Z D limn!1 E.Y jFn /. Then Z is F1 -measurable. Let us check that for
every A 2 F1
E.Z1A / D E.Y1A / : (5.14)
By Remark 4.2 it is sufficient to prove this relation for A in a class of events C ,

generating F1 , stable with respect to finite intersections and containing ˝. If A 2
Fm for some m then
E.Z1A / D lim E.E.Y jFn /1A / D E.Y1A /

n!1
because for n m A is Fn -measurable and E.E.Y S jFn /1A / D E.E.1A Y jFn // D

E.Y1A /. (5.14) is therefore proved for A 2 C D n Fn . As this class is stable with
respect to finite intersections, generates F1 and contains ˝, the proof is complete.
t
u
5.6 Continuous time martingales
We now extend the results of the previous sections to the continuous case, i.e. when
the time set is RC or an interval of RC , which is an assumption that we make from
now on.
The main argument is that if .Mt /t is a supermartingale of the filtration .Ft /t ,
then, for every t0 < t1 < < tn , .Mtk /kD0;:::;n is a (discrete time) supermartingale
of the filtration .Ftk /kD0;:::;n to which the results of the previous sections apply.
5.6 Continuous time martingales 125
Theorem 5.9 Let M D .˝; F ; .Ft /t ; .Mt /t ; P/ be a right (or left) continuous
supermartingale. Then for every T > 0, > 0,

P sup Mt E.M0 / C E.MT / ; (5.15)
0tT

P inf Mt E MT 1finf0tT Mt g EŒjMT j : (5.16)
0tT
Proof Let us prove (5.15). Let 0 D t0 < t1 < < tn D T. Then by (5.4) applied
to the supermartingale .Mtk /kD0;:::;n

P sup Mtk E.M0 / C E.MT / :
0kn
Note that the right-hand side does not depend on the choice of t1 ; : : : ; tn . Letting
ft0 ; : : : ; tn g increase to Q\Œ0; T we have that sup0kn Mtk increases to supQ\Œ0;T Mt
and Beppo Levi’s theorem gives

P sup Mt E.M0 / C E.MT / :
Q\Œ0;T
The statement now follows because the paths are right (or left) continuous, so that
sup Mt D sup Mt :
Q\Œ0;T Œ0;T
t
u
The following statements can be proved with similar arguments.
Theorem 5.10 Let X be a right-continuous supermartingale. If Œa; bŒ RC

(b D C1 possibly) and supŒa;bŒ E.Xt / < C1, then limt!b Xt exists a.s.
Theorem 5.11 A right-continuous martingale M is uniformly integrable if

and only if it is of the form Mt D E.Y jFt /, where Y 2 L1 and if and only if it
converges a.s. and in L1 .
126 5 Martingales
Example 5.2 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a real Brownian motion and,
for
2 R, let

2
X t D e
B t 2 t :
Then X is a martingale.
Indeed, with the old trick of separating increment and actual position,

2
2
EŒXt jFs D e 2 t EŒe
.Bt Bs CBs / jFs D e 2 t e
Bs EŒe
.Bt Bs / jFs

2
2
D e 2 t e
B s e 2 .ts/ D Xs :
This is the prototype of an important class of martingales that we shall

investigate later. Now observe that X is a positive martingale and hence, by
Theorem 5.10, the limit limt!C1 Xt exists a.s. and is finite. What is its value?
Is this martingale uniformly integrable? These questions are the object of
Exercise 5.9.
Theorem 5.12 (Doob’s inequality) Let M be a right-continuous martingale

bounded in Lp , p > 1, and let M D supt0 jMt j. Then M 2 Lp and
kM kp q sup kMt kp ;
t0
p
where q D p1 .
Theorem 5.13 (Stopping theorem) Let M be a right-continuous martingale

(resp. supermartingale, submartingale) and 1 ; 2 two stopping times with
1 2 . Then if 2 is bounded a.s.
E.M2 jF1 / D M1 .resp. ; / : (5.17)
Proof Let us assume first that M is a martingale. Let b > 0 and let be a stopping
time taking only finitely many values t1 < t2 < < tm and bounded above
by b. Then by Theorem 5.2, applied to the discrete time martingale .Mtk /kD0;:::;m
with respect to the filtration .Ftk /kD0;:::;m , we have EŒMb jF D M and by
Proposition 5.4 the r.v.’s .M / , for ranging over the set of stopping times taking
only finitely many values bounded above by b, is a uniformly integrable family.
Let now .n1 /n , .n2 /n be sequences of stopping times taking only finitely many

values and decreasing to 1 and 2 , respectively, and such that n2 n1 for every n
as in Lemma 3.3. By Theorem 5.2, applied to the martingale .Mtk /kD0;:::;m , where
t0 < t1 < < tm are the possible values of the two stopping times 1 ; 2 ,
EŒMn2 jFn1 D Mn1 :
Conditioning both sides with respect to F1 , as F1 Fn1 ,
EŒMn2 jF1 D EŒMn1 jF1 : (5.18)
Since n ! 1, Mn1 ! M1 and Mn2 ! M2 (the paths are right-continuous) and

the convergence also takes place in L1 , .Mn1 /n , .Mn2 /n being uniformly integrable.
Therefore one can take the limit in (5.18) leading to (5.17).
Assume now that M is a supermartingale. A repetition of the argument as above
brings us to (5.18) with D replaced by , i.e.
EŒMn2 jF1 EŒMn1 jF1 :
Now the fact that .Mn1 /n and .Mn2 /n are uniformly integrable is trickier and follows
from the next Lemma 5.2 applied to Hn D Fni and Zn D Mni , observing that the
stopping theorem for discrete time supermartingales ensures the relation M i
nC1
EŒMni jF i . The proof for the case of a submartingale is similar.
nC1
t
u
Lemma 5.2 Let .Zn /n be a sequence of integrable r.v.’s, .Hn /n a decreasing

sequence of -algebras, i.e. such that HnC1 Hn for every n D 0; 1; : : : ,
with the property that Zn is Hn -measurable and ZnC1 EŒZn jHnC1 . Then, if
supn EŒZn D ` < C1, .Zn /n is uniformly integrable.
Proof Let " > 0. We must prove that, at least for large n, there exists a c > 0 such
that EŒjZn j1fjZn jcg ". As the sequence of the expectations .EŒZn /n is increasing
to ` there exists a k such that EŒZk ` ". If n k we obtain for every A 2 Hn
EŒZk 1A D EŒE.Zk 1A jHn / D EŒ1A E.Zk jHn / EŒZn 1A ;
hence
EŒZk 1fZn <cg EŒZn 1fZn <cg ;

EŒZk 1fZn cg EŒZn 1fZn cg :
128 5 Martingales
Now, for every n k,
EŒjZn j1fjZn jcg D EŒZn 1fZn cg EŒZn 1fZn cg

D EŒZn EŒZn 1fZn <cg EŒZn 1fZn cg
` EŒZk 1fZn <cg EŒZk 1fZn cg
(5.19)
" C EŒZk EŒZk 1fZn <cg EŒZk 1fZn cg
D " C EŒZk 1fZn cg EŒZk 1fZn cg
D " C EŒjZk j1fjZn jcg :
Moreover, for every n,
1 1 1
P.jZn j c/ EŒjZn j D .2EŒZnC EŒZn / .2` EŒZ0 / : (5.20)
c c c
Let now ı be such that EŒZk 1A " for every A 2 F satisfying P.A/ ı,
as guaranteed by Proposition 5.2 (the single r.v. Zk forms a uniformly integrable
family) and let c be large enough so that P.jZn j c/ ı, thanks to (5.20). Then
by (5.19), for every n k,
EŒjZn j1fjZn jcg " C EŒjZk j1fjZn jcg 2" ;
which is the relation we are looking for.

t
u
Proposition 5.6 Let M D .˝; F ; .Ft /t ; .Mt /t ; P/ be a right-continuous

martingale and a stopping time. Then .Mt^ /t is also an .Ft /t -martingale.
Proof Thanks to Exercise 5.6 we need only prove that, for every bounded stopping
time , E.M ^ / D E.M0 /. As ^ is also a bounded stopping time, this follows
from Theorem 5.13 applied to the two bounded stopping times 1 D 0 and 2 D
^ .
t
u
Example 5.3 Let B be a Brownian motion, a; b > 0 and denote by the exit
time of B from the interval Œa; b. It is immediate that < C1 a.s., thanks
to the Iterated Logarithm Law. By continuity, the r.v. B , i.e. the position of B
at the time at which it exits from Œa; b, can only take the values a and b.
What is the value of P.B D a/?
(continued )

As B is a martingale, by the stopping theorem, Theorem 5.13, applied to
the stopping times 1 D 0, 2 D t ^ , we have for every t > 0
0 D E.B0 / D E.Bt^ / :
Letting t ! C1 Bt^ ! B and, by Lebesgue’s theorem (the inequalities

a Bt^ b are obvious),
0 D E.B / D aP.B D a/ C bP.B D b/

D aP.B D a/ C b.1 P.B D a// D b .a C b/P.B D a/
from which
b a
P.B D a/ D ; P.B D b/ D
aCb aCb
As in Example 5.1, we have solved a problem concerning the distribution of

a process stopped at some stopping time with an application of the stopping
theorem (Theorem 5.13 in this case) to some martingale.
Here, as in Example 5.1, the process B itself is a martingale; in general,
the stopping theorem will be applied to a martingale which is a function of
the process or associated to the process in some way, as developed in the
exercises.
The hypothesis of right continuity in all these statements is not going to be

too constraining, as almost every martingale we are going to deal with will
be continuous. In a couple of situations, however, we shall need the following
result, which gives some simple hypotheses guaranteeing the existence of a right-
continuous modification. See Revuz and Yor (1999) or Karatzas and Shreve (1991)
for a proof.
Theorem 5.14 Let .˝; F ; .Ft /t ; .Mt /t ; P/ be a supermartingale. Let us

assume that the standard hypotheses are satisfied: the filtration .Ft /t is right-
continuous and is augmented with the negligible events of F1 . Then .Mt /t
admits a right-continuous modification if and only if the map t 7! E.Mt / is
continuous.
We have seen in Proposition 3.4 that the paths of a Brownian motion do not have
finite variation a.s. The following two statements show that this is a general property
of every continuous martingale that is square integrable.
130 5 Martingales
Theorem 5.15 Let M be a square integrable continuous martingale with

respect to a filtration .Ft /t . Then if it has finite variation, it is a.s. constant.
Proof Possibly switching to the martingale Mt0 D Mt M0 , we can assume M0 D 0.

Let D 0 < t0 < < tn D t denote a subdivision of the interval Œ0; t and let
jj D max jtiC1 ti j. As was the case when investigating the variation of the paths
of a Brownian motion, we shall first look at the quadratic variation of the paths. Let
us assume first that M has bounded variation on the interval Œ0; T, i.e. that
X
n1
ˇ ˇ
VT .!/ D sup ˇMt .!/ Mti .!/ˇ K
iC1

iD0
for some K > 0 and for every !. Note that EŒMtiC1 Mti D EŒE.MtiC1 jFti /Mti D
EŒMt2i , so that
EŒ.MtiC1 Mti /2 D EŒMt2iC1 2MtiC1 Mti C Mt2i D EŒMt2iC1 Mt2i :
Therefore
hX
n1 i hX
n1 i
EŒMt2 D E .Mt2iC1 Mt2i / D E .MtiC1 Mti /2 :
iD0 iD0
As
ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ
.MtiC1 Mti /2 D ˇMtiC1 Mti ˇ ˇMtiC1 Mti ˇ ˇMtiC1 Mti ˇ max ˇMtiC1 Mti ˇ
in1
we have
h ˇ ˇXn1
ˇ ˇi ˇ ˇ
EŒMt2 E ˇ
max MtiC1 Mti ˇ ˇMt Mt ˇ E VT max ˇMt Mt ˇ
iC1 i iC1 i
0in1 0in1
iD0
ˇ ˇ
KE max ˇMtiC1 Mti ˇ :

The paths of M being continuous, if .n /n is a sequence of subdivisions such that

jn j ! 0 then maxn jMtiC1 Mti j ! 0 as n ! 1 and also maxn jMtiC1 Mti j
Vt K. Therefore by Lebesgue’s theorem
EŒMt2 D 0 ;
which gives Mt D 0 a.s. The time t being arbitrary, we have P.Mq D 0 for every q 2
Q/ D 1, so that, as M is assumed to be continuous, it is equal to 0 for every t a.s.
In order to get rid of the assumption VT K, let Vt be the variation of M up to

time t, i.e.
X
n1
Vt D sup jMtiC1 Mti j

iD0
for ranging among the partitions of the interval Œ0; t. The process .Vt /t is adapted
to the filtration .Ft /t , as Vt is the limit of Ft -measurable r.v.’s. Hence, for every
K > 0, K D inf.s 0I Vs K/ is a stopping time. Note that .Vt /t is continuous,
which is a general property of the variation of a continuous function.
Hence .Vt^K /t is bounded by K and is the variation of the stopped martingale
MtK D MK ^t , which therefore has bounded variation. By the first part of the proof
M K is identically zero and this entails that Mt 0 on fVT Kg. Now just let
K ! C1.
t
u
As a consequence,
a square integrable stochastic process whose paths are differentiable or even

Lipschitz continuous cannot be a martingale.
An important question is whether Doob’s decomposition (Theorem 5.1) also holds

for time continuous submartingales. We shall just mention the following result (see
Revuz and Yor 1999, Chapter IV or Karatzas and Shreve 1991, Chapter 1 for proofs
and more details).
Theorem 5.16 Let M be a square integrable continuous martingale with

respect to a filtration .Ft /t augmented with the negligible events of F1 . Then
there exists a unique increasing continuous process A such that A0 D 0 and
.Mt2 At /t is a .Ft /t -martingale. If D f0 D t0 < t1 < < tm D tg is a
partition of the interval Œ0; t then, in probability,
X
m1
At D lim jMtkC1 Mtk j2 : (5.21)
jj!0
kD0
We shall still call the process .At /t of Theorem 5.16 the associated increasing
process to the square integrable martingale M and sometimes we shall use the
notation hMi. Thus hMi is a process such that t 7! Mt2 hMit is a martingale and
Theorem 5.16 states that, if M is a square integrable continuous martingale, such a
process exists and is unique.
132 5 Martingales
Note that (5.21) implies that the increasing process associated to a square
integrable continuous martingale does not depend on the filtration: if .Mt /t is
a martingale with respect to two filtrations, then the corresponding increasing
associated processes are the same.
Remark 5.3 Under the hypotheses of Theorem 5.16, if is a stopping time

and we consider the stopped martingale Mt D Mt^ (which is also continuous
and square integrable), is it true, as it may appear natural, that its associated
increasing process can be obtained simply by stopping hMi?
The answer is yes: it is clear that .Mt /2 hMit^ D Mt^ 2
hMit^ is a
2
martingale, being obtained by stopping the martingale .Mt hMit /t . By the
uniqueness of the associated increasing process we have therefore, a.s.,
hMit^ D hM it : (5.22)
We are not going to prove Theorem 5.16, because we will be able to explicitly
compute the associated increasing process of the martingales we are going to deal
with.
Example 5.4 If .Bt /t is a Brownian motion of the filtration .Ft /t , then we

know (Example 5.1 c)) that it is also a martingale of that filtration. Therefore
Proposition 3.4, compared with (5.21), can be rephrased by saying that the
increasing process A associated to a Brownian motion is At D t. This is
consistent with the fact that .B2t t/t is a martingale, which is easy to verify
(see also Exercise 5.10).
There are various characterizations of Brownian motion in terms of martingales.

The following one will be very useful later.
Theorem 5.17 Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be an m-dimensional contin-

uous process such that X0 D 0 a.s. and that, for every 2 Rm ,
1 2t
Yt D eih ;Xt iC 2 j j
is an .Ft /t -martingale. Then X is an .Ft /t -Brownian motion.

5.7 Complements: the Laplace transform 133
Proof The martingale relation

1 2 ˇ
1
E eih ;Xt iC 2 j j t ˇ Fs D E.Yt jFs / D Ys D eih ;Xs iC 2 j j s
2
implies that, for every 2 Rm ,

1 2 .ts/
E.eih ;Xt Xs i jFs / D e 2 j j : (5.23)
Taking the expectation in (5.23) we find that the increment Xt Xs is N.0; .t

s/I/-distributed. Moreover, (Exercise 4.5), (5.23) implies that the r.v. Xt Xs is
independent of Fs . Therefore a), b) and c) of Definition 3.2 are satisfied.
t
u
5.7 Complements: the Laplace transform
One of the natural application of martingales appears in the computation of the

Laplace transform of stopping times or of processes stopped at some stopping time
(see Exercises 5.30, 5.31, 5.32, 5.33).
Let be a -finite measure on Rm . Its complex Laplace transform is the function
W C ! C defined as
m m
Z
.z/ D ehz;xi d.x/ (5.24)
Rm
for every z 2 Cm such that

Z
jehz;xi j d.x/ < C1 : (5.25)
Rm
The domain of the Laplace transform, denoted D , is the set of the z 2 Cm for
which (5.25) is satisfied (and therefore for which the Laplace transform is defined).
Recall that ehz;xi D e<hz;xi .cos.=hz; xi/ C i sin.=hz; xi/ so that jehz;xi j D e<hz;xi .
If has density f with respect to Lebesgue measure then, of course,
Z
.z/ D ehz;xi f .x/ dx
Rm
and we shall refer to it indifferently as the Laplace transform of or of f . If X is an

m-dimensional r.v., the Laplace transform of its law X is, of course,
X .z/ D EŒehz;Xi ;
thanks to the integration rule with respect to an image law, Proposition 1.1.
134 5 Martingales
In this section we recall some useful properties of the Laplace transform. We

shall skip the proofs, in order to enhance instead consequences and applications.
We shall assume m D 1, for simplicity. The passage to the case m 2 will be rather
obvious.
1) There exist x1 ; x2 2 R [ fC1g [ f1g; x1 0 x2 , (the convergence
abscissas) such that, if x1 < <z < x2 ,
Z Z
j e j d.t/ D
zt
et<z d.t/ < C1
R R
i.e. z 2 D , whereas
Z
j ezt j d.t/ D C1 (5.26)
R
if <z < x1 or x2 < <z. In other words, the domain D contains an open strip (the
convergence strip) of the form x1 < <z < x2 , which can be empty (if x1 D x2 ).
Let x1 < x2 be real numbers such that
Z
etxi d.t/ < C1; i D 1; 2
R
and let z 2 C such that x1 <z x2 . Then t<z tx2 if t 0 and t<z tx1 if
t 0. In any case, hence et<z etx1 C etx2 and (5.26) holds.
A typical situation appears when the measure has a support that is contained
in RC . In this case jezt j 1 if <z < 0; therefore the Laplace transform is defined at
least on the whole half plane <z < 0 and in this case x1 D 1.
Of course if is a finite measure, D always contains the imaginary axis <z D 0.
Even in this situation it can happen that x1 D x2 D 0. Moreover, if is a probability,
then its characteristic function b is related to its Laplace transform by the relation
b
.t/ D .it/.
Example 5.5
a) Let .dx/ D 10;C1Œ e x be an exponential distribution. We have
Z 1 Z 1
tz t
e e dt D e. z/t dt
0 0
so that the integral converges if and only if <z < . Therefore the
convergence abscissas are x1 D 1, x2 D . For <z < we have
Z 1

.z/ D e. z/t dt D (5.27)
0 z
(continued )

b) If
dt
.dt/ D
.1 C t2 /
(Cauchy law) then the two convergence abscissas are both equal to 0.
Indeed
Z C1
et<z
dt D C1
1 .1 C t2 /
unless <z D 0.
2) If the convergence abscissas are different (and therefore the convergence strip
is non-empty) then is a holomorphic function in the convergence strip.
This property can be easily proved in many ways. For instance, one can prove that
in the convergence strip we can take the derivative under the integral sign in (5.24).
Then if we write z D x C iy we have, using the fact that z 7! ezt is analytic and
satisfies the Cauchy-Riemann equations itself,
@ Z C1
@ @ @
< .z/ = .z/ D <ezt =ezt d.t/ D 0 ;
@x @y 1 @x @y
Z C1
@ @ @ @
< .z/ C = .z/ D <ezt C =ezt d.t/ D 0 ;
@y @x 1 @y @x
so that the Cauchy–Riemann equations are satisfied by . In particular, inside the

convergence strip the Laplace transform is differentiable infinitely many times.
Example 5.6
a) If N.0; 1/, we have already computed its Laplace transform for x 2
2
R in Exercise 1.6, where we found .x/ D ex =2 . As the convergence
abscissas are x1 D 1; x2 D C1, is analytic on the complex plane
and by the uniqueness of the analytic extension this relation gives the value
of for every z 2 C:
2 =2
.z/ D ez : (5.28)
(continued )
136 5 Martingales

The same result might have been derived immediately from the value of
the characteristic function
2 =2
.it/ .t/ D et
Db (5.29)
and again using the argument of analytic extension. Note also that the
analyticity of the Laplace transform might have been used in order to
compute the characteristic function: first compute the Laplace transform
for x 2 R as in Exercise 1.6, then for every z 2 C by analytic extension
which gives the value of the characteristic function by restriction to the
imaginary axis. This is probably the most elegant way of computing the
characteristic function of a Gaussian law.
b) What is the characteristic function of a .˛; / distribution? Recall that,
for ˛; > 0, this is a distribution having density

f .x/ D x˛1 e x
.˛/
for x > 0 and f .x/ D 0 on the negative half-line, where is Euler’s

R C1
Gamma function, .˛/ D 0 x˛1 ex dx. As developed above, we can
compute the Laplace transform on the real axis and derive its value on the
imaginary axis by analytic continuation. If z 2 R and z < . We have
Z Z C1
˛
.z/ D ezx f .x/ dx D x˛1 e x ezx dx
.˛/ 0
Z C1 ˛
˛ ˛ .˛/
D x˛1 e. z/x dx D D ;
.˛/ 0 .˛/ . z/˛ z
where the integral was computed by recognizing, but for the normalizing
constant, the expression of a .˛; z/ distribution. If z the integral
diverges, hence is the convergence abscissa. The characteristic function
of f is then obtained by substituting z D i
and is equal to
˛

7! :
i
The fact that inside the convergence strip is differentiable infinitely many times
and that one can differentiate under the integral sign has another consequence: if 0
belongs to the convergence strip, then
Z
0
.0/ D x d.x/ ;
i.e. the derivative at 0 of the Laplace transform is equal to the expectation. Iterating
this procedure we have that if the origin belongs to the convergence strip then the
r.v. having law has finite moments of all orders.
3) Let us prove that if two -finite measures 1 ; 2 have a Laplace transform
coinciding on an open non-empty interval x1 ; x2 Œ R, then 1 D 2 . Again by the
uniqueness of the analytic extension, the two Laplace transforms have a convergence
strip that contains the strip x1 < <z < x2 and coincide in this strip. Moreover, if
is such that x1 < < x2 and we denote by e j ; j D 1; 2, the measures d ej .t/ D
e t dj .t/, then, denoting by their common Laplace transform,
Z
Q j .z/ D
ezt e t dj .t/ D . C z/ :
R
The measures e
j are finite, as
e
j .R/ D Q j .0/
D . / < C1 :
Up to a division by . /, we can even assume that e

1 and e
2 are probabilities.
Hence, as
b
e
j .t/ D Q j .it/
D . C it/ ;
they have the same characteristic function. Therefore, by 6. of Sect. 1.6, e

1 D e2
and therefore also 1 D 2 .
This proves that the Laplace transform characterizes the probability .
4) As is the case of characteristic functions, it is easy to compute the Laplace
transform of the sum of independent r.v.’s. If X and Y are independent then the
product formula for the expectation of independent r.v.’s gives for the Laplace
transform of their sum
XCY .z/ D EŒez.XCY/ D EŒezX ezY D EŒezX EŒezY D X .z/ Y .z/
provided that both expectations are finite, i.e. the Laplace transform of the sum is
equal to the product of the Laplace transforms and is defined in the intersection of
the two convergence strips.
As for the characteristic functions this formula can be used in order to determine
the distribution of X C Y.
5) A useful property of the Laplace transform: when restricted to Rm is
a convex function. Actually, even log is convex, from which the convexity of
follows, as the exponential function is convex and increasing. Recall Hölder’s
inequality (1.2): if f ; g are positive functions and 0 ˛ 1, then
Z Z ˛ Z 1˛
f .x/˛ g.x/1˛ d.x/ f .x/ d.x/ g.x/ d.x/ :
138 5 Martingales
Therefore
Z Z
˛ x 1˛
.˛
C .1 ˛/ / D e.˛
C.1˛/ /x d.x/ D e
x e d.x/
Z ˛ Z 1˛

x
e d.x/ e x d.x/ D .
/˛ . /1˛ ;
from which we obtain the convexity relation for log :
log .˛
C .1 ˛/ / ˛ log .
/ C .1 ˛/ log . / :
6) Let X be a r.v. and let us assume that its Laplace transform is finite at a
certain value > 0. Then for every x > 0, by Markov’s inequality,
. / D EŒe X e x P.X x/
i.e. we have for the tail of the distribution of X
P.X x/ . /e x :
The finiteness of the Laplace transform therefore provides a useful information

concerning the decrease of the tail of the distribution. This gives to the convergence
abscissa some meaning: if the right abscissa, x2 , is > 0, then, for every " > 0,
P.X x/ const e.x2 "/x
with const possibly depending on ".
Remark 5.4 To fix the ideas, let us consider Exercise 5.33. There we prove
that if
< 0 then
p
2
EŒe
a D ea. 2
/ ; (5.30)
where a is the passage

p time at a > 0 of Xt D Bt C t, > 0. Let us
a. 2 2z /
define g.z/ D e . g defines a holomorphic function also for some
2
values of z with <z > 0, actually on the whole half-plane <z < 2 . Can we
2 2
say that (5.30) also holds for 0
< 2 and that 2 is the convergence
abscissa of the Laplace transform of a ?
The answer is yes, as we shall see now with an argument that can be
reproduced in other similar situations. This is very useful: once the value of
the Laplace transform is computed for some values of its argument, its value
(continued )

is immediately known wherever the argument of analytic continuation can be
applied.
Let x0 < 0. The power series development of the exponential function at
x0 a gives
1
X .x x0 /k
exa D ak ex0 a :
kD0
kŠ
If we choose x real and such that x x0 > 0, the series above has positive
terms and therefore can be integrated by series, which gives
1
X .x x0 /k
EŒe xa
D EŒak ex0 a (5.31)
kD0
kŠ
and this equality remains true also if EŒexa D C1 since the series on the
right-hand side has positive terms. As also g, being holomorphic, has a power
series development
1
X .z x0 /k
g.z/ D ak (5.32)
kD0
kŠ
and the two functions g and z 7! EŒeza coincide on the half-plane <z < 0,
we have ak D EŒak ex0 a and the two series, the one in (5.31) and the other
in (5.32), necessarily have the same radius of convergence. Now a classical
property of holomorphic functions is that their power series expansion
converges in every ball in which they are holomorphic. As g is holomorphic
2 2
on the half-plane <z < 2 , then the series in (5.31) converges for x < 2 ,
2
so that 2 is the convergence abscissa of the Laplace transform of a and the
2
relation (5.30) also holds on <z < 2 .
Exercises
5.1 (p. 485) Let .˝; .Ft /t ; .Xt /t ; P/ be a supermartingale and assume, moreover,
that E.Xt / D const. Then .Xt /t is a martingale. (This is a useful criterion.)
5.2 (p. 486) (Continuation of Example 5.1) Let .Zn /n be a sequence of i.i.d. r.v.’s
taking the values ˙1 with probability 12 and let X0 D 0 and Xn D Z1 C C Zn . Let
140 5 Martingales
a; b be positive integers and let a;b D inffnI Xn b or Xn ag, the exit time of X
from the interval a; bŒ.
a) Compute lima!C1 P.Xa;b D b/.
b) Let b D inffnI Xn bg be the passage time of X at b. Deduce that b < C1 a.s.
5.3 (p. 486) Let .Yn /n be a sequence of i.i.d. r.v.’s such that P.Yi D 1/ D p,
P.Yi D 1/ D q with q > p. Let Xn D Y1 C C Yn .
a) Compute limn!1 1n Xn and show that limn!1 Xn D 1 a.s.
b1) Show that
q Xn
Zn D
p
is a martingale.
b2) As Z is positive, it converges a.s. Determine the value of limn!1 Zn .
c) Let a; b 2 N be positive numbers and let D inffnI Xn D b or Xn D ag. What
is the value of EŒZn^ ? And of EŒZ ?
d) What is the value of P.X D b/ (i.e. what is the probability for the random walk
.Xn /n to exit from the interval a; bŒ at b)?
5.4 (p. 487) Let .Xn /n be a sequence of independent r.v.’s on the probability space
.˝; F ; P/ with mean 0 and variance 2 and let Fn D .Xk ; k n/. Let Mn D
X1 C C Xn and let .Zn /n be a square integrable process predictable with respect
to .Fn /n (i.e. such that ZnC1 is Fn -measurable).
a) Show that
X
n
Yn D Zk Xk
kD1
is a square integrable martingale.

b) Show that EŒYn D 0 and that
X
n
EŒYn2 D 2 EŒZk2 :
kD1
What is the associated increasing process of .Mn /n ? And of .Yn /n ?

c) Let us assume Zk D 1k . Is the martingale .Yn /n uniformly integrable?
5.5 (p. 488) Let .Yn /n1 be a sequence of independent r.v.’s such that
P.Yk D 1/ D 2k
P.Yk D 0/ D 1 2 2k
P.Yk D 1/ D 2k
and let Xn D Y1 C C Yn , Fn D .Xk ; k n/.

a) Show that .Xn /n is a martingale of the filtration .Fn /n .
b) Show that .Xn /n is square integrable and compute its associated increasing
process.
c) Does the martingale .Xn /n converge a.s.? In L1 ? In L2 ?
5.6 (p. 489)

a) Let .Ft /t be a filtration. Let s t and A 2 Fs . Prove that the r.v. defined as
(
s if ! 2 A
.!/ D
t if ! 2 Ac
is a stopping time.
b) Prove that an integrable right-continuous process X D .˝; F ; .Ft /t ; .Xt /t ; P/ is
a martingale if and only if, for every bounded .Ft /t -stopping time , E.X / D
E.X0 /.
5.7 (p. 490) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a real Brownian motion. Prove
that, for every K 2 R,
Mt D .eBt K/C
is an .Ft /t -submartingale.
5.8 (p. 490)
a) Let M be a positive martingale. Prove that, for s < t, fMs D 0g fMt D 0g
a.s.
b) Let M D .Mt /t be a right-continuous martingale.
b1) Prove that if D infftI Mt D 0g, then M D 0 on f < C1g.
b2) Prove that if MT > 0 a.s., then P.Mt > 0 for every t T/ D 1.
b2) Use the stopping theorem with the stopping times T, T > 0, and T ^ .
• Concerning b), let us point out that, in general, it is possible for a continuous
process X to have P.Xt > 0/ D 1 for every t T and P.Xt > 0 for every t
T/ D 0. Even if, at first sight, this seems unlikely because of continuity.
142 5 Martingales
An example is the following, which is an adaptation of Example 2.1: let

˝ D0; 1Œ, F D B.0; 1Œ/ and P DLebesgue measure. Define a process X
as Xt .!/ D 0 if t D !, X0 .!/ D X1 .!/ D 1, Xt .!/ Dlinearly interpolated
otherwise. We have, for every t, fXt > 0g D ˝ n ftg D0; 1Œnftg, which is a set
of probability 1, but every path vanishes for one value of t. This exercise states
that this phenomenon cannot occur if, in addition, X is a martingale (Fig. 5.1).
0 1
Fig. 5.1 A typical graph of t 7! Xt .!/

motion and let, for 2 Rm ,
1 2
Xt D eh ;Bt i 2 j j t :
a) Prove that .Xt /t is an .Ft /t -martingale.

b) As the condition of Theorem 5.12 is satisfied (X is positive) limt!C1 Xt exists
a.s. and is finite. What is its value?
c) Is X uniformly integrable?
5.10 (p. 491) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a real Brownian motion and let
Yt D B2t t.
a) Prove that .Yt /t is an .Ft /t -martingale. Is it uniformly integrable?
b) Let be the exit time of B from the interval a; bŒ. In Example 5.3 we saw
that < C1 a.s. and computed the distribution of X . Can you derive from a)
that EŒB2 D EŒ? What is the value of EŒ? Is EŒ finite?
c) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian motion and let
Yt D jBt j2 mt.
c1) Prove that .Yt /t is an .Ft /t -martingale.
c2) Let us denote by the exit time of B from the ball of radius 1 of Rm . Compute
EŒ.
5.11 (p. 493) Let M D .˝; F ; .Ft /t ; .Mt /t ; P/ be a square integrable martin-
gale.
a) Show that EŒ.Mt Ms /2 D EŒMt2 Ms2 :
b) M is said to have independent increments if, for every t > s, the r.v. Mt Ms is
independent of Fs . Prove that in this case the associated increasing process is
hMit D E.Mt2 / E.M02 / D EŒ.Mt M0 /2 and is therefore deterministic.
c) Show that a Gaussian martingale has necessarily independent increments with
respect to its natural filtration.
d) Let us assume that M D .˝; F ; .Ft /t ; .Mt /t ; P/ has independent increments
and, moreover, is a Gaussian martingale (i.e. simultaneously a martingale and a
Gaussian process). Therefore its associated increasing process is deterministic,
thanks to b) above. Show that, for every
2 R,
1 2
Zt D e
Mt 2
hMi t
(5.33)
is an .Ft /t -martingale.
5.12 (p. 494)

a) Let X D .˝; F ; .Fn /n ; .Xn /n ; P/ be a (discrete time) martingale. Let Y D .Yn /n
be a process equivalent to X (see p. 32 for the definition of equivalent processes).
Show that Y is a martingale with respect to its natural filtration.
b) Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be a (time continuous) martingale. Let Y D .Yt /t
be a process equivalent to X. Show that Y is a martingale with respect to its
natural filtration.
5.13 (p. 495) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion.

a) Show that
Z t
Yt D tBt Bu du
0
is a martingale. Recall (Exercise 3.11) that we know that .Yt /t is a Gaussian

process.
b) Prove that, if t > s, EŒ.Yt Ys /Bu D 0 for every u s.
5.14 (p. 495)

a) Let .Fn /n be a filtration, X an integrable r.v. and an a.s. finite stopping time of
.Fn /n . Let Xn D E.X jFn /; then
E.X jF / D X :
b) Let .Ft /t0 be a filtration, X an integrable r.v. and a a.s. finite stopping time.
Let .Xt /t be a right continuous process such that Xt D E.X jFt / a.s. (thanks
144 5 Martingales
to Theorem 5.14 the process .Xt /t thus defined always has a right-continuous
modification if .Ft /t is the augmented natural filtration). Then
E.X jF / D X :
b) Use Lemma 3.3.
5.15 (p. 496) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and 2 R.
Prove that
Z t
Mt D e t Bt e u Bu du
0
is an .Ft /t -martingale with independent increments.

5.16 (p. 496) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and ; real
numbers with 0. Let
Xt D e Bt t :
a) Compute, for 2 R, > 0, the value of limt!1 E.Xt /.

b) Show that .Xt /t is
8
ˆ
ˆ 2

<a supermartingale if 2
2
a martingale if D
ˆ 2
:̂a submartingale if 2
:
2
c) Assume > 0. Show that there exists an ˛ > 0 such that .Xt˛ /t is a
supermartingale with limt!1 E.Xt˛ / D 0. Deduce that the limit limt!1 Xt exists
a.s. for every 2 R and compute it.
2
d) Prove that, if 2 < , then
Z C1
A1 WD e Bs s ds < C1 a.s.
0
and compute E.A1 /. Prove that the r.v.’s

Z C1 Z C1
Bs s 1 2 s
e ds and eBs ds
0 2 0
have the same law.

5.17 (p. 498) (The law of the supremum of a Brownian motion with a negative
drift)
a) Let .Mt /t be a continuous positive martingale such that limt!C1 Mt D 0 and
M0 D 1 (do you recall any examples of martingales with these properties?). For
x > 1, let x D infftI Mt xg its passage time at x. Show that
lim Mt^x D x1fx <C1g : (5.34)

t!C1
b) Deduce from (5.34) the value of P.x < C1/.

c) Let M D supt0 Mt . Compute P.M x/.
d) Let B be a Brownian motion and, for
> 0, let Xt D Bt
t. Prove that
2
Mt D e2
Bt 2
t is a martingale and show that the r.v. X D supt0 Xt is a.s.
finite and has an exponential law with parameter 2
.
5.18 (p. 499) The aim of this exercise is the computation of the law of the
supremum of a Brownian bridge X. As seen in Exercise 4.15, this is a continuous
Gaussian process, centered and with covariance function Ks;t D s.1 t/ for s t.
a) Show that there exists a Brownian motion B such that, for 0 t < 1,
Xt D .1 t/B t : (5.35)
1t
b) Prove that, for every a > 0,

P sup Xt > a D P sup.Bs as/ > a
0t1 s>0
and deduce the partition function and density of sup0t1 Xt .

b) Use Exercise 5.17 d).
5.19 (p. 499) Let B be a Brownian motion and let D infftI Bt 62 x; 1Œg be the
exit time of B from the interval x; 1Œ.
a) Compute P.B D x/.
b) We want to compute the distribution of the r.v.
Z D min Bt
0t1
where 1 denotes the passage time of B at 1. Z is the minimum level attained by

B before passing at 1. Compute

P.Z x/ D P min Bt x :
0t1
Does the r.v. Z have finite mathematical expectation? If so, what is its value?
146 5 Martingales
5.20 (p. 500) In this exercise we compute the exit distribution from an interval of a
Brownian motion with drift. As in Example 5.3, the problem is very simple as soon
as we find the right martingale. . . Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian
motion and, for > 0, let Xt D Bt C t.
a) Prove that
Mt D e2Xt
b) Let a; b > 0 and let be the exit time of .Xt /t from the interval a; bŒ.
b1) Show that < C1 a.s.
b2) What is the value of P.X D b/? What is the value of the limit of this probability
as ! 1?

a) Show that Xt D B3t 3tBt is an .Ft /t -martingale.
b) Let a; b > 0 and let us denote by the exit time of B from the interval a; bŒ.
Compute the covariance Cov.; B /.
Recall (Exercise 5.10) that we know that is integrable.
5.22 (p. 501) (The product of independent martingales) Let .Mt /t , .Nt /t be
martingales on the same probability space .˝; F ; P/, with respect to the filtrations
.Mt /t , .Nt /t , respectively. Let us assume, moreover, that the filtrations .Mt /t and
.Nt /t are independent. Then the product .Mt Nt /t is a martingale of the filtration
Ht D Mt _ Nt .
5.23 (p. 501) Let .Mt /t be a continuous .Ft /t -martingale.
a) Prove that if .Mt2 /t is also a martingale, then .Mt /t is constant.
0
b1) Prove that if p > 1 and .jMt jp /t is a martingale, then .jMt jp /t is a martingale
for every 1 p0 p.
b2) Prove that if p 2 and .jMt jp /t is a martingale then .Mt /t is constant.

motion.
a) Prove that, if i 6D j, the process .Bi .t/Bj .t//t is a .Ft /t -martingale.
b) Prove that if i1 ; : : : ; id , d m are distinct indices, then t 7! Bi1 .t/ : : : Bid .t/ is a
martingale.
c) Let B be an m m matrix Brownian motion, i.e. Bt D .Bij .t//ij , i; j D 1; : : : ; m,
where the Bij ’s are independent Brownian motions. Prove that Xt D det.Bt / is a
martingale. Prove that Yt D det.Bt Bt / is a submartingale.

a) Compute, for t T,
EŒ1fBT >0g jFt :
b) Let us denote by ˚ the partition function of an N.0; 1/ distribution. Prove that

B
t
Mt D ˚ p
T t
is an .Ft /t -martingale for t < T. Is it uniformly integrable? Compute the a.s.

limit limt!T Mt .

a) Show that the integral
Z t
Bu
Xt D p du
0 u
converges a.s. for every t 0.

b) Show that .Xt /t is a Gaussian process and compute its covariance function Ks;t D
Cov.Xs ; Xt /.
c) Is .Xt /t a martingale?
5.27 (p. 505) Let P, Q be probabilities on .˝; F ; .Ft /t /. Let us assume that, for
every t > 0, the restriction QjFt of Q to Ft is absolutely continuous with respect to
the restriction of P to Ft . Let
dQjFt
Zt D
dPjFt
a) Show that .Zt /t is a martingale.

b) Show that Zt > 0 Q-a.s. and that .Zt1 /t is a Q-supermartingale.
c) Show that if also QjFt PjFt , then .Zt1 /t is a Q-martingale.
Recall (Exercise 5.8) that the set of the zeros t 7! fZt D 0g is increasing.
5.28 (p. 505) Let .Xt /t be a continuous process such that Xt is square integrable
for every t. Let G D .Xu ; u s/ and Gn D .Xsk=2n ; k D 1; : : : ; 2n /. Let t > s.
Does the conditional expectation of Xt given Gn converge to that of Xt given G , as
n ! 1? In other words, if the process is known at the times sk=2n for k D 1; : : : ; 2n ,
is it true that if n is large enough then, in order to predict the future position at time
148 5 Martingales
t, it is almost as if we had the knowledge of the whole path of the process up to time
s?
a) Show that the sequence of -algebras .Gn /n is increasing and that their union
generates G .
b1) Let Zn D E.Xt jGn /. Show that the sequence .Zn /n converges a.s. and in L2 to
E.Xt jG /.
b2) How would the statement of b1) change if we just assumed Xt 2 L1 for every
t?
5.29 (p. 506) Let .Bt /t be an .Ft /t Brownian motion and an integrable stopping
time for the filtration .Ft /t . We want to prove the Wald equalities: EŒB D 0,
EŒB2 D EŒ. The situation is similar to Example 4.5 but the arguments are going
to be different as here is not in general independent of .Bt /t .
a) Prove that, for every t 0, EŒB ^t D 0 and EŒB2^t D EŒ ^ t.
b) Prove that the martingale .B ^t /t is bounded in L2 and that
h i
E sup B2^t 4EŒ : (5.36)
t0
c) Prove that EŒB D 0 and EŒB2 D EŒ.
5.30 (p. 507) (The Laplace transform of the passage time of a Brownian motion)
Let B be a real Brownian motion, a > 0 and a D infftI Bt ag the passage time at
1 2
a. We know from Example 5.2 that Mt D e Bt 2 t is a martingale.
a) Prove that, for 0, EŒMa D 1. Why does the proof not work for < 0?
b) Show that the Laplace transform of a is
p
.
/ D EŒe
a D ea 2
(5.37)
for
0, whereas .
/ D C1 for
> 0.
c) Show that a has a law that is stable with exponent 12 (this was already proved in a
different manner in Exercise 3.20, where the definition of stable law is recalled).
5.31 (p. 507) Let B be a Brownian motion and, for a > 0, let us denote by the
exit time of B from the interval Œa; a. In Example 5.3 we remarked that < C1
a.s. and we computed the distribution of X . Show that, for
> 0,
1
EŒe
D p
cosh.a 2
/
Find the right martingale. . . Recall that B and are independent (Exercise 3.18). Further
properties of exit times from bounded intervals are the object of Exercises 5.32, 8.5 and 10.5.
5.32 (p. 508) Let B be a Brownian motion with respect to a filtration .Ft /t and let
D infftI jBt j ag be the exit time from a; aŒ. In Exercise 5.31 we computed the
Laplace transform
7! EŒe
for
0. Let us investigate this Laplace transform
for
> 0. Is it going to be finite for some values
> 0?
1 2
a) Prove that, for every 2 R, Xt D cos. Bt / e 2 t is an .Ft /t -martingale.

b) Prove that, if j j < 2a ,
1 2 1
EŒe 2 . ^t/
< C1;
cos. a/
1 2
and that the r.v. e 2 is integrable. Prove that
1 2
EŒe
D p ; 0
< (5.38)
cos.a 2
/ 8a2
c) What is the value of E./? Show that 2 Lp for every p 0

d2
d) Determine the spectrum of the Laplace operator 12 dx 2 on the interval a; aŒ,
with the Dirichlet conditions u.a/ D u.a/ D 0, i.e. determine the numbers

2 R such that there exists a function u not identically zero, twice differentiable
in a; aŒ, continuous on Œa; a and such that
(
1 00
2u D
u on a; aŒ
(5.39)
u.a/ D u.a/ D 0:
Which is the largest of these numbers? Do you notice some coincidence with the
results of b)?
Recall that if 0 < < 2 , then cos x cos > 0 for x 2 Œ; . See Exercise 10.5 concerning
the relation between the convergence abscissas of the Laplace transform of the exit time and the
spectrum of the generator of the process.
5.33 (p. 510) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. Let 2 R, a > 0.
We want to investigate the probability of crossing the level a > 0 and the time
needed to do this for the process Xt D Bt C t. For D 0 the reflection principle
already answers this question whereas Exercise 5.17 investigates the case < 0.
We now consider > 0. Let D infft 0I Xt ag.
a) Prove that < C1 a.s.
b) Prove that, for every 2 R,
2
Mt D e Xt . 2 C /t
150 5 Martingales
is an .Ft /t -martingale. What is the value of
2
EŒMt^ D EŒe Xt^ . 2 C /.t^ /
‹
c) Let 0. Show that Mt^ e a and that
2
EŒe. 2 C /
D e a : (5.40)
d) Compute, for
> 0, EŒe
. What is the value of EŒ?
Chapter 6
Markov Processes
In this chapter we introduce an important family of stochastic processes. Diffusions,

which are the object of our investigation in the subsequent chapters, are instances of
Markov processes.
6.1 Definitions and general facts
Let .E; E / be a measurable space.
A Markov transition function on .E; E / is a function p.s; t; x; A/, where s; t; 2

RC , s t, x 2 E, A 2 E , such that
i) for fixed s; t; A, x 7! p.s; t; x; A/ is E -measurable;
ii) for fixed s; t; x, A 7! p.s; t; x; A/ is a probability on .E; E /;
iii) p satisfies the Chapman–Kolmogorov equation
Z
p.s; t; x; A/ D p.u; t; y; A/ p.s; u; x; dy/ (6.1)
E
for every s u t.
Definition 6.1 Given on .E; E / a Markov transition function p and a prob-

ability , an .E; E /-valued process X D .˝; F ; .Ft /tu ; .Xt /tu ; P/ is said
(continued)

152 6 Markov Processes
Definition 6.1 (continued)

to be a Markov process associated to p, starting at time u and with initial
distribution , if
a) Xu has law .
b) (The Markov property) For every bounded measurable function f W E ! R
and t > s u
Z
E. f .Xt /jFs / D f .x/ p.s; t; Xs ; dx/ a.s. (6.2)
E
When the filtration .Ft /t is not specified it is understood, as usual, that it is the
natural filtration.
As we shall see better in the sequel, p.s; t; x; A/ represents the probability that
the process, being at position x at time s, will move to a position in the set A at
time t. The Chapman–Kolmogorov equation intuitively means that if s < u < t, the
probability of moving from position x at time s to a position in A at time t is equal
to the probability of moving to a position y at the intermediate time u and then from
y to A, integrated over all possible values of y.
Remark 6.1
a) Let us have a closer look at the Markov property: if f D 1A , A 2 E , (6.2)
becomes, for s t,
P.Xt 2 AjFs / D p.s; t; Xs ; A/ :
Note that this conditional expectation given Fs only depends on Xs .

Therefore
P.Xt 2 AjXs / D EŒP.Xt 2 AjFs /jXs

(6.3)
D EŒ p.s; t; Xs ; A/jXs D p.s; t; Xs ; A/ D P.Xt 2 AjFs / ;
i.e. the conditional probability of fXt 2 Ag given Xs or given the whole past
Fs coincide. Intuitively, the knowledge of the whole path of the process
up to time s or just of its position at time s give the same information about
the future position of the process at a time t, t s.
(continued )
6.1 Definitions and general facts 153

It is important to understand this aspect of the Markov property when
modeling a random phenomenon with a Markov process. Consider, for
instance, the evolution of financial assets (stocks, bonds, options,. . . ): to
assume that, in view of predicting tomorrow’s price, the knowledge of
the prices of the previous days, knowing today’s price, does not give
any additional valuable information is certainly not correct. Consider the
importance of knowing whether or not we are in a trend of growing prices,
and that to determine this we need to know the price evolution of the last
few days. To model this kind of situation with a Markovian model may
lead to grossly wrong appreciations.
However, we shall see how to derive a non Markovian model from a
Markovian one (see Exercises 6.2 and 6.11, for example).
b) Recalling (4.18), the relation
Z
f .x/ p.s; t; Xs ; dx/ D EŒ f .Xt /jFs D EŒ f .Xt /jXs
E
says that p.s; t; x; / is the conditional distribution of Xt given Xs D x. This

a useful starting point when, given a Markov process, one has to determine
the transition function.
c) For every bounded measurable function f W E ! R and for every s < t,
the map
Z
x 7! f . y/ p.s; t; x; dy/ (6.4)
E
is measurable. Indeed i) of the definition of transition function states

that (6.4) is measurable for every f which is an indicator of a set A 2 E .
Therefore this is true for every linear combination of such indicators and
for every positive measurable function, as these are increasing limits of
linear combinations of indicators. By subtracting the positive and negative
parts, (6.4) is true for every bounded measurable function f .
d) Definition 6.1 allows to determine the finite-dimensional distributions of
a Markov process. Let f0 ; f1 W E ! R be Borel functions and t > u and let
us derive the joint distribution of .Xu ; Xt /. As Xu has law , with the trick
of performing an “intermediate conditioning”,

EŒ f1 .Xt /f0 .Xu / D E f0 .Xu /EŒ f1 .Xt /jFu

h Z i
D E f0 .Xu / f1 .x1 / p.s; t; Xu ; dx1 / (6.5)
Z E Z
D f0 .x0 /.dx0 / f1 .x1 / p.s; t; x0 ; dx1 / :
E E
(continued )

By induction on m we have, for every u t1 < < tm ,
EŒ f0 .Xu /f1 .Xt1 / : : : fm .Xtm /

Z Z
D f0 .x0 /.dx0 / f1 .x1 / p.u; t1 ; x0 ; dx1 / : : :
E Z E
(6.6)
: : : fm1 .xm1 / p.tm2 ; tm1 ; xm2 ; dxm1 /
E Z
fm .xm /p.tm1 ; tm ; xm1 ; dxm / :
E
The meaning of (6.6) may be difficult to decrypt at first, however it is simple:

once the rightmost integral in dxm has been computed, the result is a function
of xm1 , which is integrated in the subsequent integral giving, rise to a
function of xm2 , and so on.
Formula (6.6) states that a priori the finite-dimensional distributions of a
Markov process only depend on the initial distribution , the starting time u
and the transition function p. In particular, two Markov processes associated
to the same transition function and having the same starting time and starting
distribution are equivalent.
Example 6.1 If E D Rm , E D B.Rm /, let

Z h jy xj2 i
1
p.s; t; x; A/ D exp dy : (6.7)
Œ2.t s/m=2 A 2.t s/
Then p is a Markov transition function: p.s; t; x; / is an N.x; .t s/I/ law.

The Chapman–Kolmogorov equation is a consequence of the property of the
normal laws with respect to the convolution product: if A 2 B.Rm /, then
Z
p.u; t; y; A/ p.s; u; x; dy/
Z h jy xj2 i Z h jz yj2 i
1 1
D exp dy exp dz
Œ2.u s/m=2 2.u s/ A Œ2.t u/
m=2 2.t u/
Z Z h jy xj2 i h jz yj2 i
1 1
D dz exp exp dy
A Œ2.u s/m=2 2.u s/ Œ2.t u/m=2 2.t u/
„ ƒ‚ …„ ƒ‚ …
f1 . y/ f2 .zy/
so that the inner integral is the convolution of the densities f1 , which is

N.x; .u s/I/, and f2 N.0; .t u/I/. The result, owing to well-known
(continued )

properties of Gaussian distributions (see Sect. 1.7), is N.x; .t s/I/, i.e.
p.s; t; x; /. Therefore we can conclude that
Z Z h jy xj2 i
1
p.u; t; y; A/ p.s; u; x; dy/ D exp dy D p.s; t; x; A/ :
Œ2.t s/m=2 A 2.t s/
In Example 4.4 we computed

Z h jy B j2 i
1 s
EŒ f .Bt /jFs D f . y/ exp dy
Œ2.t s/m=2 2.t s/
Z
D f .x/ p.s; t; Bs ; dx/ ;
E
i.e. the Markov property, b) of Definition 6.1, is satisfied by Brownian motion,

with respect to the Markov transition function given by (6.7). Therefore
Brownian motion becomes our first example of Markov process. To be
precise, it is a Markov process associated to the transition function (6.7) and
starting at 0 at time 0.
Let us now prove the existence of the Markov process associated to a given
transition function and initial distribution. The idea is simple: thanks to the previous
remark we know what the finite-dimensional distributions are. We must therefore
a) check that these satisfy the coherence Condition 2.1 of Kolmogorov’s existence
Theorem 2.2 so that a stochastic process with these finite-dimensional distribu-
tions does exist;
b) prove that a stochastic process having such finite-dimensional distributions is
actually a Markov process associated to the given transition function, which
amounts to proving that it enjoys the Markov property (6.3).
Assume E to be a complete separable metric space endowed with its Borel -
C
algebra B.E/ and p a Markov transition function on .E; B.E//. Let ˝ D ER ;
an element ! 2 ˝ is therefore a map RC ! E. Let us define Xt W ˝ ! E as
Xt .!/ D !.t/, t 0 and then Ftu D .Xs ; u s t/, F D F1 u
. Let us consider
the system of finite-dimensional distributions defined by (6.6). The Chapman–
Kolmogorov equation (6.1) easily implies that this system of finite-dimensional
distributions satisfies the Condition 2.1 of coherence.
Therefore there exists a unique probability P on .˝; F / such that the probabili-
ties (6.6) are the finite-dimensional distributions of .˝; F ; .Ftu /t ; .Xt /t ; P/. We now
have to check that this is a Markov process associated to p and with initial (i.e. at
time u) distribution . Part a) of Definition 6.1 being immediate, we have to check
the Markov property b), i.e. we must prove that if D 2 Fsu and f W E ! R is
bounded and measurable, then
h Z i
EŒ f .Xt /1D D E 1D f .x/ p.s; t; Xs ; dx/ : (6.8)
E
It will, however, be sufficient (Remark 4.2) to prove this relation for a set D of the
form
D D fXt0 2 B0 ; : : : ; Xtn 2 Bn g ;
where u D t0 < t1 < < tn D s, since by definition the events of this form
generate Fsu , are a class that is stable with respect to finite intersections, and ˝
itself is of this form. For this choice of events the verification of (6.8) is now easy
as both sides are easily expressed in terms of finite-dimensional distributions: as
1fXtk 2Btk g D 1Bk .Xtk /, we have 1D D 1B0 .Xt0 / : : : 1Bn .Xtn / and

EŒ f .Xt /1D D E 1B0 .Xt0 / : : : 1Bn1 .Xtn1 /1Bn .Xs /f .Xt /

Z Z
D 1B0 . y0 /.dy0 / 1B1 . y1 /p.u; t1 ; y0 ; dy1 / : : :
Z Z
::: 1Bn . yn /p.tn1 ; s; yn1 ; dyn / f . y/ p.s; t; yn ; dy/
„ ƒ‚ …
WDe
f . yn /
Z Z Z
D 1B0 . y0 /.dy0 / 1 B 1 . y1 / : : : 1Bn . yn /e
f . yn /p.tn1 ; s; yn1 ; dyn /

D EŒ1B0 .Xt0 / : : : 1Bn1 .Xtn1 /1Bn .Xs /e

f .Xs / D E 1De
f .Xs / :
Hence (6.8) holds for every D 2 Fsu and as e f .Xs / is Fsu -measurable we have proved
that
Z
EŒ f .Xt /jFs D e
f .Xs / D f . y/ p.s; t; Xs ; dy/ :
Therefore .˝; F ; .Ftu /tu ; .Xt /tu ; P/ is a process satisfying conditions a) and b)
of Definition 6.1. The probability P just constructed depends of course, besides p,
on and on u and will be denoted P;u . If D ıx we shall write Px;u instead of Pıx ;u
and we shall denote by Ex;s the expectation computed with respect to Px;s .
Going back to the construction above, we have proved that, if E is a complete
separable metric space, then there exist
a) a measurable space .˝; F / endowed with a family of filtrations .Fts /ts , such
0
that Fts0 Fts if s s0 , t0 t;
b) a family of r.v.’s Xt W ˝ ! E such that ! 7! Xt .!/ is Fts -measurable for every

s t;
c) a family of probabilities .Px;s /x2E;s2RC such that Px;s is a probability on .˝; F1
s
/
with the properties that, for every x 2 E and s; h > 0,
Px;s .Xs D x/ D 1 (6.9)

Z
Ex;s Œ f .XtCh /jFts D f .z/ p.t; t C h; Xt ; dz/ Px;s -a.s. (6.10)
E
Note that in (6.10) the value of the conditional expectation does not depend on s.
A family of processes .˝; F ; .Fts /ts ; .Xt /t ; .Px;s /x;s / satisfying a), b), c) is
called a realization of the Markov process associated to the given transition function
p. In some sense, the realization is a unique space .˝; F / on which we consider
a family of probabilities Px;s that are the laws of the Markov processes associated
to the given transition function, only depending on the starting position and initial
time.
Let us try to familiarize ourselves with these notations. Going back to the
expression of the finite-dimensional distributions (6.6) let us observe that, if D ıx
(i.e. if the starting distribution is concentrated at x), (6.5) with f0 1 gives
Z Z Z
E Œ f .Xt / D
x;s
ıx .dx0 / f .x1 /p.s; t; x0 ; dx1 / D f .x1 /p.s; t; x; dx1 / (6.11)
E E E
so that, with respect to Px;s , p.s; t; x; / is the law of Xt . If f D 1A then the previous
relation becomes
Px;s .Xt 2 A/ D p.s; t; x; A/ : (6.12)
Thanks to (6.11) the Markov property (6.10) can also be written as

Z
Ex;s Œ f .XtCh /jFts D f .z/p.t; tCh; Xt ; dz/ D EXt ;t Œ f .XtCh / Px;s -a.s. ; (6.13)
for every x; s or, for f D 1A ,
Px;s .XtCh 2 AjFts / D p.t; t C h; Xt ; A/ D PXt ;t .XtCh 2 A/ Px;s -a.s. (6.14)
for every x; s. The expression PXt ;t .XtCh 2 A/ can initially create some confusion:
it is simply the composition of the function x 7! Px;t .XtCh 2 A/ with the r.v. Xt .
From now on in this chapter X D .˝; F ; .Fts /ts ; .Xt /t0 ; .Px;s /x;s / will denote the
realization of a Markov process associated to the transition function p.
The Markov property, (6.10) or (6.13), allows us to compute the conditional
expectation with respect to Fts of a r.v., f .XtCh /, that depends on the position of the
process at a fixed time t C h posterior to t. Sometimes, however, it is necessary to
compute the conditional expectation Ex;s .Y jFts / for a r.v. Y that depends, possibly,
on the whole path of the process after time t.
The following proposition states that the Markov property can be extended to
cover this situation. Note that, as intuition suggests, this conditional expectation
also depends only on the position, Xt , at time t.
Let Gts D .Xu ; s u t/. Intuitively, a G1s
-measurable r.v. is r.v. that depends
on the behavior of the paths after time s.
Proposition 6.1 Let Y be a bounded G1

t
-measurable r.v. Then
Ex;s .Y jFts / D EXt ;t .Y/ Px;s -a.s. (6.15)
Proof Let us assume first that Y is of the form f1 .Xt1 / : : : fm .Xtm /, where t t1 <
< tm and f1 ; : : : ; fm are bounded measurable functions. If m D 1 then (6.15) is
the same as (6.13). Let us assume that (6.15) holds for Y as above and for m 1.
Then, conditioning first with respect to Ftsm1 ,
Ex;s Œ f1 .Xt1 / : : : fm .Xtm /jFts

DE f1 .Xt1 / : : : fm1 .Xtm1 /Ex;s Œ fm .Xtm /jFtsm1 jFts

x;s
ˇ
D Ex;s f1 .Xt1 / : : : fm1 .Xtm1 /EXtm1 ;tm1 Œ fm .Xtm / ˇ Fts

D Ex;s Œ f1 .Xt1 / : : :e
f m1 .Xtm1 /jFts ;
where e
f m1 .x/ D fm1 .x/ Ex;tm1 Œ fm .Xtm /; by the induction hypothesis
Ex;s Œ f1 .Xt1 / : : : fm .Xtm /jFts D EXt ;t Œ f1 .Xt1 / : : :e

f m1 .Xtm1 / :
However, by (6.13), EXtm1 ;tm1 Œ fm .Xtm / D Ex;t Œ fm .Xtm /jFttm1 Px;t -a.s. for every
x; t so that
EXt ;t Œ f1 .Xt1 / : : :e
f m1 .Xtm1 / D EXt ;t Œ f1 .Xt1 / : : : fm1 .Xtm1 / EXtm1 ;tm1 Œ fm .Xtm /
D EXt ;t Œ f1 .Xt1 / : : : fm1 .Xtm1 / EXt ;t Œ fm .Xtm /jFttm1
D EXt ;t ŒEXt ;t Œ f1 .Xt1 / : : : fm1 .Xtm1 /fm .Xtm /jFttm1
D EXt ;t Œ f1 .Xt1 / : : : fm1 .Xtm1 /fm .Xtm / :
We obtain the general case using Theorem 1.4.

t
u
Remark 6.2 In some sense the Chapman–Kolmogorov equation follows from

the Markov property. More precisely, let X D .˝; F ; .Fts /ts ; .Xt /t ; .Px;s /x;s /
be a family of stochastic processes such that Xs D x Px;s -a.s. for every x; s and
let p.s; t; x; A/ be a function satisfying the conditions i) and ii) of the definition
of a transition function on p. 151. Then, if for every x 2 E and s u t,
Px;s .Xt 2 AjFus / D p.u; t; Xu ; A/ Px;s -a.s. ;
necessarily p satisfies the Chapman–Kolmogorov equation (6.1) and X is a

realization of the Markov process associated to p.
Indeed first with u D s in the previous relation we obtain, as Xs D x
Px;s -a.s.,
Px;s .Xt 2 AjFss / D p.s; t; Xs ; A/ D p.s; t; x; A/ Px;s -a.s.
i.e. p.s; t; x; / is the law of Xt with respect to Px;s . Then, if s u t,
p.s; t; x; A/ D Px;s .Xt 2 A/ D Ex;s ŒPx;s .Xt 2 AjFus /

Z
D E Œ p.u; t; Xu ; A/ D p.u; t; y; A/p.s; u; x; dy/ ;
x;s
E
which is the required Chapman–Kolmogorov relation.
The probability of making a given transition in a time span h, p.s; s C h; x; A/, in

general depends on the time s. Sometimes the transition function p depends on s and
t only as a function of t s, so that p.s; s C h; x; A/ does not actually depend on s.
In this case, p is said to be time homogeneous and, recalling the expression of the
finite-dimensional distributions,
Ex;s Œ f1 .Xt1 Cs / : : : fm .Xtm Cs / D Ex;0 Œ f1 .Xt1 / : : : fm .Xtm / :
In this case, as the transition p.s; s C h; x; A/ does not depend on s, in some sense the
behavior of the process is the same whatever the initial time. It is now convenient
to fix 0 as the initial time and to set p.t; x; A/ D p.0; t; x; A/, Ft D Ft0 , Px D
Px;0 and consider as a realization X D .˝; F ; .Ft /t ; .Xt /t ; .Px /x /. The Chapman–
Kolmogorov equation becomes, for 0 s < t,
Z
p.t; x; A/ D p.t s; y; A/p.s; x; dy/
and the relations (6.12) and (6.14), for s t, are now
Px .Xt 2 A/ D p.t; x; A/
Px .Xt 2 AjFs / D p.t s; Xs ; A/ D PXs .Xts 2 A/ Px -a.s.
The transition function of Brownian motion, see Example 6.1, is time homogeneous.
Remark 6.3 Transition functions such as that of Brownian motion, Exam-

ple 6.1, satisfy another kind of invariance: by looking back to the expression
for the transition function (6.7), by a change of variable we find that
p.t; x; A/ D p.t; 0; A x/ :
This implies that, if .Bt /t is a Brownian motion, then .Bxt /t , with Bxt D x C Bt ,
is a Markov process associated to the same transition function and starting at
x. Indeed, going back to (4.10),
P.Bxt 2 AjFs / D P.Bt 2 A xjFs / D p.t s; Bs ; A x/

D p.t s; Bs C x; A/ D p.t s; Bxs ; A/ :
Hence the Markov process associated to the transition function of Brownian

motion, (6.7), and starting at x has the same law as t 7! Bt C x.
6.2 The Feller and strong Markov properties
A positive r.v. is said to be an s-stopping time if s and if it is a stopping

time for the filtration .Fts /t , i.e. if f tg 2 Fts for every t s. We define
Fs D fA 2 F1 s
; A \ f tg 2 Fts for every tg.
Definition 6.2 We say that X is strong Markov if, for every x 2 E and A 2
B.E/, for every s 0 and for every finite s-stopping time ,
Px;s .XtC 2 AjFs / D p.; t C ; X ; A/ Px;s -a.s. (6.16)
It is clear that if h s then (6.16) boils down to the usual Markov

property (6.10). (6.16) is, of course, equivalent to
Z
Ex;s Œ f .XtC /jFs D p.; t C ; X ; dy/f . y/ Px;s -a.s. (6.17)
E
for every bounded Borel function f , thanks to the usual arguments of approximating
positive Borel functions with increasing sequences of linear combinations of
indicator functions.
6.2 The Feller and strong Markov properties 161
Example 6.2 It is immediate to derive from Theorem 3.3 (the stopping

theorem for Brownian motion) that a Brownian motion is strong Markov.
Indeed, as the increment XtC X is independent of F we have, by the
freezing Lemma 4.1,
Ex Œ f .XtC /jF D Ex Œ f .XtC X C X /jF D ˚.X / ; (6.18)
where
˚.z/ D Ex Œ f .XtC X C z/ :
Now, as t 7! XtC X is also a Brownian motion, under Px XtC X C z has

law p.t; z; dy/, denoting by p the transition function of a Brownian motion.
Hence
Z
˚.z/ D f . y/ p.t; z; dy/ ;
which, going back to (6.18), allows to conclude the proof.
The strong Markov property allows us to compute the conditional expectation,

given F , of a function of the position of the process at a time subsequent to time .
It is often very useful to have formulas allowing us to compute the conditional
expectation, given F , of a more complicated function of the path of the process
after time . This question resembles the one answered by Proposition 6.1 where
(the simple) Markov property was extended in a similar way.
Here the situation is more complicated, as it does not seem easy to give a precise
meaning to what a “function of the path of the process after time ” is. In order to
do this we are led to the following definitions.
Sometimes the space ˝, on which a Markov process is defined, is a space of
paths, i.e. ! 2 ˝ is a map RC ! E and the process is defined as Xt .!/ D !.t/.
This is the case, for instance, for the process constructed in Kolmogorov’s theorem,
C
where ˝ D ER , or for the canonical space .C ; M ; .Mt /t ; .Xt /t ; .Px /x / introduced
in Sect. 3.2, where C D C .RC ; Rm /. We shall say, in this case, that the Markov
process is canonical. In this case we can define, for every t 2 RC , the map
t W
˝ ! ˝ through

t !.s/ D !.t C s/ :
The maps
t are the translation operators and we have the relations
Xs ı
t .!/ D Xs .
t !/ D XtCs .!/ :
Translation operators are the tool that allows us to give very powerful expressions
of the Markov and strong Markov properties.
If W ˝ ! RC is a r.v., we can define the random translation operator
as
.
!/.t/ D !.t C .!// :
Note that in the time homogeneous case the strong Markov property can be written
as
Z
Ex Œ f .XtC /jF D f . y/p.t; X ; dy/
E
or, with the notations of translation operators,
Ex Œ f .Xt / ı
jF D EX Œ f .Xt / : (6.19)
This leads us to the following extension of the strong Markov property for the time
homogeneous case. The proof is similar to that of Proposition 6.1.
Proposition 6.2 Let X be a canonical time homogeneous strong Markov

process and Y an F1 -measurable r.v., bounded or positive. Then for every
finite stopping time
Ex ŒY ı
jF D EX ŒY : (6.20)
Proof The proof follows the same pattern as Proposition 6.1, first assuming that
Y is of the form Y D f1 .Xt1 / : : : fm .Xtm / where 0 t1 < < tm and f1 ; : : : ; fm
are bounded (or positive) functions. Then (6.20) is immediate if m D 1, thanks
to (6.19). Assume that (6.20) is satisfied for every Y of this form and for m 1.
Then if Y D f1 .Xt1 / : : : fm .Xtm / for every x we have Px -a.s.
Ex ŒY ı
jF D Ex Œ f1 .Xt1 C / : : : fm .Xtm C /jF
D Ex Œ f1 .Xt1 C / : : : fm1 .Xtm1 C /Ex Œ fm .Xtm C /jFtm1 C jF
D Ex Œ f1 .Xt1 C / : : : fm1 .Xtm1 C /EXtm1 C Œ fm .Xtm /jF
D Ex Œ f1 .Xt1 C / : : :e
f m1 .Xtm1 C /jF ;
where e
f m1 .x/ D fm1 .x/Ex Œ fm .Xtm /. By the recurrence hypothesis then, going back
and using the “simple” Markov property,
Ex ŒY ı
jF D EX Œ f1 .Xt1 / : : :e
f m1 .Xtm1 /
D EX Œ f1 .Xt1 / : : : fm1 .Xtm1 /EXtm1 Œ fm .Xtm /
D EX Œ f1 .Xt1 / : : : fm1 .Xtm1 /EX Œ fm .Xtm /jFtm1

D EX ŒEX Œ f1 .Xt1 / : : : fm1 .Xtm1 /fm .Xtm /jFtm1
D EX Œ f1 .Xt1 / : : : fm1 .Xtm1 /fm .Xtm /
D EX ŒY :
Theorem 1.4 allows us to state that if (6.20) is true for every Y of the form
Y D f1 .Xt1 / : : : fm .Xtm / with 0 t1 tm and f1 ; : : : ; fm bounded or positive
functions, then it holds for every bounded (or positive) F1 -measurable r.v.
t
u
Example 6.3 Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical realiza-

tion of an m-dimensional Brownian motion. Let D Rm be a bounded open
set, the exit time from D and W @D ! R a bounded measurable function.
Let
u.x/ D Ex Œ.B / : (6.21)
We see now that u enjoys some important properties.

Let BR .x/ be a ball centered at x with radius R and contained in D, R the
exit time from BR .x/. As > R Px -a.s (starting at x the process must exit
from BR .x/ before leaving D), we have
X D X ı
R
as the paths t 7! Xt .!/ and t 7! XtCR .!/ .!/ exit D at the same position (not
at the same time of course). Hence (6.20) gives
u.x/ D Ex Œ.X / D Ex Œ.X / ı

R
D Ex ŒEx ..X / ı
R jMR / D Ex ŒEXR Œ f .X / :
„ ƒ‚ …
Du.XR /
Hence
Z
u.x/ D Ex Œu.XR / D u. y/ d R . y/ ;
@BR .x/
where R is the law of XR under Px . By Exercise 3.18 R is the normalized

Lebesgue measure of the spherical surface @BR .x/. Therefore the value of u at
(continued )

x coincides with the mean of u on the surface of the sphere @BR .x/ for every
R (small enough so that BR D). It is a well-known result in mathematical
analysis (see Han and Lin 1997, p. 3, for example) that a bounded measurable
function with this property is harmonic, i.e. that it is twice differentiable in D
and such that
4u D 0 ;
where 4 denotes the Laplace operator
Xm
@2
4D
iD1
@x2i
Therefore if one wants to look for a solution of the Dirichlet problem

(
4u D 0 on D
uj@D D
the function u of (6.21) is a promising candidate. We shall discuss this kind of

problem (i.e., the relation between solutions of PDE problems and stochastic
processes) in more detail in Chap. 10.
At this point Brownian motion is our only example of a strong Markov process. We
shall snow see that, in fact, a large class of Markov processes enjoy this property.
Let us assume from now on that E is a metric space (we can therefore speak of
continuous functions) and that E D B.E/.
Definition 6.3 A transition function p is said to enjoy the Feller property if,
for every fixed h 0 and for every bounded continuous function f W E ! R,
the map
Z
.t; z/ 7! f . y/p.t; t C h; z; dy/
E
is continuous. A Markov process X is said to be a Feller process if its transition

function enjoys Feller’s property.
In fact Feller’s property is equivalent to saying that if sn ! s and xn ! x then
p.sn ; sn C h; xn ; / ! p.s; s C h; x; /
n!1
weakly. Hence, in a certain sense, that if x is close to y and s close to t then the
probability p.s; s C h; x; / is close to p.t; t C h; y; /.
Of course if the transition function is time homogeneous then
Z Z
f . y/p.t; t C h; z; dy/ D f . y/p.h; z; dy/
E E
and Feller’s property requires this to be a continuous function of z, for every

bounded continuous function f and for every fixed h > 0.
Remark 6.4 Let p be the transition function of Brownian motion, then

Z Z
1 1 2
f . y/p.h; z; dy/ D f . y/e 2h jzyj dy :
E .2h/ m=2
Rm
It is immediate, taking the derivative under the integral sign, that, if f is a

bounded measurable function, the left-hand side is actually a C1 function of
z. Therefore Brownian motion is a Feller process.
Theorem 6.1 Let X be a right-continuous Feller process. Then it is strong

Markov.
Proof We must prove that, if is a s-stopping time,
Px;s .XtC 2 AjFs / D p.; t C ; X ; A/ Px;s -a.s.
Note first that, as X is right-continuous, then it is progressively measurable and

therefore, by Proposition 3.5, X is Fs -measurable. Therefore the right-hand side
in the previous equation is certainly Fs -measurable. We must then prove that if
2 Fs then for every A 2 B.E/
Px;s .fXtC 2 Ag \ / D Ex;s .p.; t C ; X ; A/1 / :
Let us assume first that takes at most countably many values ftj gj . Then
X
Px;s .fXtC 2 Ag \ / D Px;s .fXtC 2 Ag \ \ f D tj g/
X j
(6.22)
D Px;s .fXtCtj 2 Ag \ \ f D tj g/ :
j
As \ f D tj g 2 Ftsj , by the Markov property (6.10),
Px;s .fXtCtj 2 Ag \ \ f D tj g/ D Ex;s Œ1fXtCtj 2Ag 1 \f Dtj g

D Ex;s Ex;s Œ1fXtCtj 2Ag 1 \f Dtj g jFtsj

D Ex;s 1 \f Dtj g Ex;s Œ1fXtCtj 2Ag jFtsj D Ex;s 1 \f Dtj g p.tj ; t C tj ; Xtj ; A/
so that, substituting into (6.22)

X
Px;s .fXtC 2 Ag \ / D Ex;s 1 \f Dtj g p.tj ; t C tj ; Xtj ; A/

j

DE x;s
p.; t C ; X ; A/1 :
Let now be any finite s-stopping time. By Lemma 3.3 there exists a sequence .n /n
of finite s-stopping times, each taking at most countably many values and decreasing
to . In particular, therefore Fsn Fs . The strong Markov property, already proved
for n ; and the remark leading to (6.17) guarantee that, for every bounded continuous
function f on E,
Z
Ex;s Œ f .XtCn /jFsn D f . y/p.n ; t C n ; Xn ; dy/ :
E
In particular, if 2 Fs Fsn then
h Z i
E Œ f .XtCn /1 D E
x;s x;s
1 f . y/p.n ; t C n ; Xn ; dy/ :
E
By the right continuity of the paths and the Feller property

Z Z
f . y/p.n ; t C n ; Xn ; dy/ ! f . y/p.; t C ; X ; dy/ :
E n!1 E
Since by Lebesgue’s theorem we have Ex;s Œ f .XtCn /1 ! Ex;s Œ f .XtC /1 as

n ! 1 (recall that f is bounded and continuous), we have obtained that, for every
bounded continuous function f ,
h Z i
Ex;s Œ f .XtC /1 D Ex;s 1 f . y/p.; t C ; X ; dy/ :
E
By Theorem 1.5 the previous equation holds for every bounded Borel function f and
we have proved the statement.
t
u
Note that the Feller and right continuity assumptions are not needed in the first part
of the proof of Theorem 6.1. Therefore the strong Markov property holds for every
Markov process when takes at most a countable set of values.
We have already seen (Proposition 4.3) that for a Brownian motion it is always
possible to be in a condition where the filtration is right-continuous.
T In fact, this
holds for every right-continuous Feller process. Let FtC
s
D ">0 FtC" s
.
Theorem 6.2 If X D .˝; F ; .Fts /ts ; .Xt /t ; .Px;s /x;s / is a realization of a

Feller right-continuous Markov process then .˝; F ; .FtC s
/ts ; .Xt /t ; .Px;s /x;s /
is also a realization of a Markov process associated to the same transition
function.
Proof We must prove the Markov property with respect to .FtC

s
/t , i.e. that
Z
EŒ f .XtCh /jFtC
s
D f . y/p.t; t C h; Xt ; dy/ (6.23)
E
where f can be assumed to be bounded continuous on E. A routine application of

Theorem 1.5 allows us to extend this relation to every bounded Borel function f . For
every " > 0 we have, by the Markov property with respect to .Ft /t ,
Z
E x;s
Œ f .XtChC" /jFtC"
s
D f . y/p.t C "; t C h C "; XtC" ; dy/
E
and, by conditioning both sides with respect to FtC

s
, which is contained in FtC"
s
,
hZ ˇ s i
E x;s
Œ f .XtChC" /jFtC
s
DE x;s
f . y/p.t C "; t C h C "; XtC" ; dy/ ˇ FtC :
E
As the paths are right-continuous, f .XtChC" / ! f .XtCh / a.s. as " ! 0C. Hence
the left-hand side converges to Ex;s Œ f .XtCh /jFtC
s
by Lebesgue’s theorem for
conditional expectations (Proposition 4.2 c)). Thanks to the Feller property, the
right-hand side converges to
hZ ˇ s i
Z
Ex;s f . y/p.t; t C h; Xt ; dy/ ˇ FtC D f . y/p.t; t C h; Xt ; dy/ ;
E E
Xt being FtC
s
-measurable, which concludes the proof.
t
u
Recall that a -algebra G is said to be trivial with respect to a probability P if for
every A 2 G the quantity P.A/ can only take the values 0 or 1, i.e. if all events in
G are either negligible or almost sure. Recall
T that we denote by .Gts /t the natural
filtration Gt D .Xu ; s u t/. Let GtC D ">0 GtC" .
s s s
Proposition 6.3 (Blumenthal’s 0-1 law) Let X be a right-continuous Feller

process. Then the -algebra GsC
s
is trivial with respect to the probability Px;s
for every x 2 E, s 0.
Proof By the previous theorem X is a Markov process with respect to the filtration
.FtC
s
/ts . As GsC
s
FsCs
, if A 2 GsC
s
, by Proposition 6.1 for Y D 1A and with
s D t,
1A D Ex;s .1A jFsC

s
/ D Ex;s .1A / a.s.
as Xs D x Px;s -a.s. Therefore Ex;s .1A / D Px;s .A/ can assume only the values 0 or 1.
t
u
Example 6.4 An application clarifying the otherwise mysterious Blumenthal

0-1 law is the following.
Let F E be a closed set and x 2 @F. Let X be a right-continuous Feller
process and let D infftI t s; Xs 2 F c g be the first exit time from F after
time s. X being right-continuous, is a stopping time for the filtration .FtC /t ,
thanks to Proposition 3.7. Then the event f D sg belongs to GsC s
:
f D sg D fthere exists a sequence of times sn & s such that Xsn 2 F c g
and the event on the right-hand side belongs to GsC" s

for every " > 0.
Proposition 6.3 states that the probability P . D s/ (i.e. the probability
x;s
for the process to leave F immediately) can only take the values 0 or 1, i.e.
either all paths exit F immediately Px;s -a.s. or no path does (recall that x 2
@F). Figures 6.1 and 6.2 suggest situations where these two possibilities can
arise.
Fc
F
Fig. 6.1 Typical situation where Px;s . D s/ D 1: the boundary @F near x is smooth and the
oscillations of the path take it immediately outside F
6.3 Semigroups, generators, diffusions 169
Fc
Fig. 6.2 Typical situation where Px;s . D s/ D 0: the set Fc near x is “too thin” to be “caught”
immediately by the path
6.3 Semigroups, generators, diffusions
Let p be a transition function, which for now we assume to be time homogeneous,

on the topological space E. To p we can associate the family of linear operators
.Tt /t0 defined on Mb .E/ (bounded Borel functions on E) by
Z
Tt f .x/ D f . y/ p.t; x; dy/
E
or, if X is a realization of the Markov process associated to p,
Tt f .x/ D Ex Œ f .Xt / : (6.24)

ˇ ˇ
As ˇEx Œ f .Xt /ˇ kf k1 , for every t 0 the operator Tt is a contraction on Mb .E/,
i.e. kTt f k1 kf k1 . Moreover, we have
Ts Tt D TsCt ;
i.e. the family of operators .Tt /t is a semigroup. Indeed, by the Chapman–Kolmogo-

rov equation,
Z Z Z
Ts Tt f .x/ D p.s; x; dy/Tt f . y/ D p.s; x; dy/ p.t; y; dz/f .z/
E E E
Z
D p.t C s; x; dz/f .z/ D TsCt f .x/ :
E
If, moreover, p is Feller, then Tt also operates on Cb , i.e. Tt f 2 Cb if f 2 Cb . If

f 2 Mb .E/ let us define, if the limit exists,
1
Af .x/ D lim ŒTt f .x/ f .x/ : (6.25)
t!0C t
Let us denote by D.A/ the set of functions f 2 Mb .E/ for which the limit in (6.25)
exists for every x. The operator A is defined for f 2 D.A/ and is the infinitesimal
generator of the semigroup .Tt /t or of the Markov process X.
In this section we investigate some properties of the operator A and characterize
an important class of Markov processes by imposing some conditions on A. We
assume from now on that the state space E is an open domain D Rm .
These concepts can be repeated with obvious changes when p is not time homoge-
neous. In this case we can define the family of operators .Ts;t /st through
Z
Ts;t f .x/ D f . y/ p.s; t; x; dy/ D Ex;s Œ f .Xt /
and, for s u t, the Chapman–Kolmogorov equation gives
Ts;t D Ts;u Tu;t :
For a time inhomogeneous Markov process, instead of the operator A we are led to
consider the family of operators .As /s defined, when the expression is meaningful,
by
1
As f .x/ D lim ŒTs;sCh f .x/ f .x/ :
h!0C h
We say that the infinitesimal generator A is local when the value of Af .x/ depends
only on the behavior of f in a neighborhood of x, i.e. if, given two functions f ; g
coinciding in a neighborhood of x, if Af .x/ is defined, then Ag.x/ is also defined and
Af .x/ D Ag.x/.
Proposition 6.4 Let BR .x/ be the sphere of radius R and centered at x and let
us assume that for every x 2 D and R > 0
1
lim p.t; t C h; x; BR .x/c / D 0 : (6.26)
h!0C h
Then At is local.
Proof Let f 2 D.A/ and let R be small enough so that BR .x/ D, then
Z
1 1
ŒTt;tCh f .x/ f .x/ D f . y/ p.t; t C h; x; dy/ f .x/
h h
Z
1
D Œ f . y/ f .x/ p.t; t C h; x; dy/
h
Z Z
1 1
D Œ f . y/ f .x/ p.t; t C h; x; dy/ C Œ f . y/ f .x/ p.t; t C h; x; dy/ :
h BR .x/ h BcR
As
Z ˇ 1Z
1 ˇˇ ˇ ˇ ˇ
ˇ f . y/ f .x/ˇp.t; t C h; x; dy/
ˇ Œ f . y/ f .x/p.t; t C h; x; dy/ˇ
h BR .x/c h BR .x/c
1
2 k f k1 p.t; t C h; x; BR .x/c /
h
we can conclude that the two limits
Z Z
1 1
lim Œ f . y/ f .x/ p.t; t C h; x; dy/; lim Œ f . y/ f .x/ p.t; t C h; x; dy/
h!0C h h!0C h BR .x/
either both exist or neither of them exist; moreover, At f .x/ does not depend on
values of f outside of BR .x/, for every R > 0.
t
u
Note that condition (6.26) simply states that the probability of making a transition
of length R in a time interval of amplitude h goes to 0 as h ! 0 faster than h.
The operator A satisfies the maximum principle in the following form:
Proposition 6.5 (The maximum principle) If f 2 D.A/ and x is a point of

maximum for f , then Af .x/ 0.
Proof
Z Z
Tt f .x/ D f . y/p.t; x; dy/ f .x/p.t; x; dy/ D f .x/
and therefore 1t ŒTt f .x/ f .x/ 0 for every t > 0.

t
u
If A is local, the maximum principle takes the following form. The proof is rather
straightforward.
Proposition 6.6 If A is local, then if f 2 D.A/ and x is a point of relative

maximum for f then Af .x/ 0.
If E D Rm the following proposition provides a way of computing the generator

A, at least for a certain class of functions.
Proposition 6.7 Let us assume that for every R > 0, t 0 the limits
1
lim p.t; t C h; x; BR .x/c / D 0 (6.27)
h!0C h
Z
1
lim . yi xi / p.t; t C h; x; dy/ D bi .x; t/ (6.28)
h!0C h B .x/
R
Z
1
lim . yi xi /. yj xj / p.t; t C h; x; dy/ D aij .x; t/ (6.29)
h!0C h B .x/
R
exist. Then the matrix a.x; t/ is positive semidefinite for every x; t and, if
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ ;
2 i;jD1 @xi @xj iD1
@x i
we have, for every function f 2 C2 .Rm / \ Cb .Rm /,
1
lim ŒTt;tCh f .x/ f .x/ D Lt f .x/ for every x 2 D :
h!0C h
Proof By (6.27), for f 2 C2 .Rm / \ Cb .Rm /

Z
1 1
lim ŒTt;tCh f .x/ f .x/ D lim Œ f . y/ f .x/ p.t; t C h; x; dy/
h!0C h h!0C h
Z
1
D lim Œ f . y/ f .x/ p.t; t C h; x; dy/ :
h!0C h BR .x/
Replacing f with its Taylor development to the second order,
Xm
@f 1 X @2 f
m
f . y/ D f .x/ C .x/ . yi xi / C .x/ . yi xi /. yj xj / C o.jx yj2 /
iD1
@x i 2 i;jD1
@x i @x j
we find
Z
1 1
lim ŒTt;tCh f .x/ f .x/ D Lt f .x/ C lim o.jx yj2 / p.t; t C h; x; dy/ :
h!0C h h!0C h BR .x/
Let us check that the rightmost limit is equal to 0. As the above computation holds
for every R, let R be small enough so that j o.r/
r j " for every 0 < r < R. Then
Z Z
1 1
lim jo.jx yj2/j p.t; t C h; x; dy/ lim " jx yj2 p.t; t C h; x; dy/
h!0C h BR .x/ h!0C h BR .x/
Z X
m X
m
1
D lim " .xi yi /2 p.t; t C h; x; dy/ D " aii .x; t/
h!0C h BR .x/ iD1 iD1
and the conclusion comes from the arbitrariness of ". We still have to prove that the
matrix a.x; t/ is positive semidefinite for every x; t. Let
2 Rm and f 2 C2 .Rm / \
Cb .Rm / be a function such that f . y/ D h
; y xi2 for y in a neighborhood of x.
2
Then, as the first derivatives of f vanish at x, whereas @x@i @xf j .x/ D 2
i
j ,
X
m
1
ha.x; t/
;
i D aij .x; t/
i
j D Lt f .x/ D lim ŒTt;tCh f .x/ f .x/ :
i;jD1
h!0C h
But x is a relative maximum for f , hence, by Proposition 6.6, ha.x; t/

;
i D
Lt f .x/ 0.
t
u
Example 6.5 Let us compute the infinitesimal generator of an m-dimensional

Brownian motion B. Let us first establish that it is local by checking
condition (6.26) of Lemma 6.4. We have
Z
1 1 2
p.h; x; BR .x/c / D e 2h jxyj dy
.2h/m=2 jxyjR
Z Z
1 1
2h jyj2 1 1 2
D e dy D e 2 jzj dz
.2h/m=2 jyjR .2/m=2 jzjR=ph
Z
1 14 jzj2 14 jzj2
D p e e dz
.2/ m=2
jzjR= h
Z
1
4h jRj2 1 1 2 1 2
e e 4 jzj dz const e 4h jRj
.2/ m=2
Rm
so that condition (6.26) is satisfied. Moreover,

Z Z
1 1 yi xi 1 jxyj2
. yi xi / p.h; x; dy/ D e 2h dy
h jyxjR h jyxjR .2h/m=2
(continued )

Z
1 zi 2
D ejzj =2h dz D 0
h jzjR .2h/m=2
as we are integrating an odd function on a symmetric set. Also, with the

change of variable z D yx
p so that dz D dy
h hm=2
,
Z
1
lim . yi xi /. yj xj /p.h; x; dy/
h!0C h jyxjR
Z
1 1 1 2
D lim . yi xi /. yj xj /e 2h jxyj dy
h!0C h jyxjR .2h/m=2
Z Z
1 1
jzj
2 1 1 2
D z z e 2 dw ! zi zj e 2 jzj dw D ıij
.2/m=2 jzjR=ph
i j
h!0 .2/m=2 Rm
as we recognize in the last integral nothing else than the covariance matrix of
an N.0; I/-distributed r.v. Going back to the notations of Proposition 6.7, we
have bi D 0; aij D ıij . Therefore the Brownian motion has an infinitesimal
generator given, for every function f 2 C2 \ Cb , by
1 X @2 f
m
Lf D 12 4f D
2 iD1 @x2i
Definition 6.4 Let
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ (6.30)
2 i;jD1 @xi @xj iD1
@x i
be a differential operator on Rm such that the matrix a.x; t/ is positive

semidefinite. A diffusion associated to the generator Lt is a Markov process
such that
a) it is strongly Markovian,
b) for every function f 2 CK2 .Rm /, if .Ts;t /ts is the semigroup associated to
X,
Z t
Ts;t f .x/ D f .x/ C Ts;u Lu f .x/ du : (6.31)
s
We have, up to now, a candidate example of diffusion, which is Brownian motion.

We know that it enjoys the Feller property (Example 6.4) and, being continuous, it
is also strongly Markovian by Theorem 6.1. Its generator is the Laplace operator
divided by 2 and we shall see soon that b) of Definition 6.4, which is more or less
equivalent to the fact that 12 4 is its generator, also holds.
It is natural now to wonder whether, given a more general differential operator Lt
as in (6.30), an associated diffusion process exists and is unique. In the next chapters
we shall answer this question; the method will be the construction of a new process
starting from Brownian motion, which is itself a diffusion.
Exercises
6.1 (p. 511) (When is a Gaussian process also Markov?) Let X be a centered m-
i;j j
dimensional Gaussian process and let us denote by Ks;t D E.Xsi Xt / its covariance
function. Let us assume that the matrix Ks;s is invertible for every s.
a) Prove that, for every s; t, s t, there exist a matrix Ct;s and a Gaussian r.v. Yt;s
independent of Xs such that
Xt D Ct;s Xs C Yt;s :
What is the conditional law of Xt given Xs D x?

b) Prove that X is a Markov process with respect to its natural filtration .Gt /t if and
only if for every u s t
1
Kt;u D Kt;s Ks;s Ks;u : (6.32)
b) Use the freezing Lemma 4.1; (6.32) is equivalent to requiring that Yt;s is orthogonal to Xu for
every u s.
Z t
Xt D Bu du :
0
a) Is .Xt /t a martingale of the filtration .Ft /t ?

b) Show that .Xt /t is a Gaussian process and compute its covariance function Ks;t D
Cov.Xs ; Xt /. Show that .Xt /t is not a Markov process
c) Compute
Cov.Xt ; Bs /
(be careful to distinguish whether s t or s t).

d) Show that .Bt ; Xt /t is a Markov process.
6.3 (p. 514) Let g; h W RC ! RC be continuous functions with g increasing and

both vanishing at most at 0. Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion
and
Xt D h.t/Bg.t/ :
a1) Prove that X is a Markov process (with respect to which filtration?) and compute
its transition function.pIs it time homogeneous?
t
a2) Assume that h.t/ D pg.t/ . What is the law of Xt for a fixed t? Is X a Brownian
motion?
b) Assume that limt!C1 g.t/ D C1. What can be said about
Xt
lim p ‹
t!C1 2
2g.t/h .t/ log log g.t/
6.4 (p. 515) Let B be a Brownian motion, > 0, and let
Xt D e t Be2 t :
a) Prove that X is a time homogeneous Markov process and determine its

transition function.
b1) Compute the law of Xt . What do you observe?
b2) Prove that for every t1 < t2 < < tm and h > 0 the two r.v.’s .Xt1 ; : : : ; Xtm /
and .Xt1 Ch ; : : : ; Xtm Ch / have the same distribution (i.e. that X is stationary).
c) Let .˝; F ; .Ft /t ; .Zt /t ; .Px /x / be a realization of the Markov process associ-
ated to the transition function computed in a). What is the law of Zt under Px ?
Prove that, whatever the starting point x,
L
Zt ! N.0; 1/ : (6.33)
t!C1
6.5 (p. 516) (Brownian bridge again). We shall make use here of Exercises 4.15
and 6.1. Let B be a real Brownian motion and let Xt D Bt tB1 . We have already
dealt with this process in Exercise 4.15, it is a Brownian bridge.
a) What is its covariance function Ks;t D E.Xs Xt /? Compute the law of Xt and the
conditional law of Xt given Xs D x, 0 s < t 1.
b) Prove that X is a non-homogeneous Markov process and compute its transition
function.
• The generator of the Brownian bridge can be computed using Proposition 6.7. It
will be easier to compute it later using a different approach, see Exercise 9.2.
6.6 (p. 517) (Time reversal of a Brownian motion) Let B be a Brownian motion
and for 0 t 1 let
Xt D B1t :
a) Prove that X is a Gaussian process and compute its transition function.

b) Prove that .Xt /t1 is an inhomogeneous Markov process with respect to its natural
filtration and compute its transition function.
6.7 (p. 517) Let X be an Rm -valued Markov process associated to a transition

function p such that, for some ˇ > 0, " > 0 and c > 0,
Z
jx yjˇ p.s; t; x; dy/ c jt sjmC"
Rm
for every x 2 Rm and s t. Then X has a continuous version and its generator is
local.
6.8 (p. 518) Let X D .˝; F ; .Ft /t ; .Xt /t ; .Px /x / be an m-dimensional time
homogeneous diffusion and let us assume that its transition function p satisfies the
relation
p.t; x; A/ D p.t; 0; A x/
as in Remark 6.3.
a) Prove that, for every bounded Borel function f ,
Z Z
f . y/ p.t; x; dy/ D f .x C y/ p.t; 0; dy/ : (6.34)
Rm
b) Prove that the generator L of X has constant coefficients.
6.9 (p. 519) (A transformation of Markov processes)

a) Let p be a time homogeneous transition function on a measurable space .E; E /,
h W E ! R a strictly positive measurable function and ˛ a number such that for
every x 2 R, t > 0,
Z
h. y/ p.t; x; dy/ D e˛t h.x/ : (6.35)
E
Show that
Z
e˛t
p .t; x; A/ D
h
h. y/ p.t; x; dy/
h.x/ A
is a Markov transition function.

b) Let us assume that (6.35) holds and let us denote by L and Lh the generators of
the semigroups .Tt /t and .Tth /t associated to p and to ph respectively. Show that
if f 2 D.L/, then g D hf belongs to D.Lh / and express Lh g in terms of Lf .
c) Let us assume, moreover, that E D Rm and that p is the Markov transition
function of a diffusion of generator
1X X
m m
@2 @
LD aij .x/ C bi .x/ (6.36)
2 i;jD1 @xi @xj iD1
@x i
Prove that, if h is twice differentiable, CK2 D.Lh / and compute Lh g for g 2 CK2 .
d) Let p be the transition function of an m-dimensional Brownian motion (see
Example 6.1) and let h.x/ D ehv;xi , where v 2 Rm is some fixed vector. Show
that (6.35) holds for some ˛ to be determined and compute Lh g for g 2 CK2 .
6.10 (p. 520) Let E be a topological space. We say that a time homogeneous E-
valued Markov process X associated to the transition function p admits an invariant
(or stationary) measure if is a -finite measure on .E; B.E// such that, for
every compactly supported bounded Borel function f ,
Z Z
Tt f .x/ d.x/ D f .x/ d.x/ (6.37)
E
R
for every t. Recall that Tt f .x/ D E f . y/ p.t; x; dy/. If, moreover, is a probability,
we say that X admits an invariant (or stationary) distribution.
a) Prove that is a stationary distribution if and only if, if X0 has law then Xt also
has law for every t 0.
b) Prove that the Lebesgue measure of Rm is invariant for m-dimensional Brownian
motion.
c) Prove that, if for every x 2 E, the transition function of X is such that
lim p.t; x; A/ D 0 (6.38)

t!C1
for every bounded Borel set A E, then X cannot have an invariant probability.
Deduce that the m-dimensional Brownian motion cannot have an invariant
probability.
d) Prove that if X is a Feller process and there exists a probability on .E; B.E//
such that, for every x 2 E,
lim p.t; x; / D (6.39)

t!C1
in the sense of the weak convergence of probabilities, then is a stationary

distribution.
6.11 (p. 521) (When is a function of a Markov process also Markov?) Let X D
.˝; F ; .Ft /t ; .Xt /t ; P/ be a Markov process associated to the transition function p
and with values in .E; E /. Let .G; G / be a measurable space and ˚ W E ! G a
surjective measurable map. Is it true that Y D .˝; F ; .Ft /t ; .Yt /t ; P/ with Yt D
˚.Xt / is also a Markov process? In this exercise we investigate this question. The
answer is no, in general: the Markov property might be lost for Y. We have seen an
example of this phenomenon in Exercise 6.2.
a) Prove that if the map ˚ is bijective then Y is a Markov process and determine
its transition function.
b) Let us assume that for every A 2 G the transition function p of X satisfies the
relation
p.s; t; x; ˚ 1 .A// D p.s; t; z; ˚ 1 .A// : (6.40)
for every x; z such that ˚.x/ D ˚.z/; let, for 2 G and A 2 G , q.s; t; ; A/ D
p.s; t; x; ˚ 1 .A//, where x is any element of E such that ˚.x/ D . )
b1) Prove that for every bounded measurable function f W G ! R
Z Z
f . y/ q.s; t; ; dy/ D f ı ˚.z/ p.s; t; x; dz/ ; (6.41)
G E
where x 2 E is any element such that ˚.x/ D .

b2) Show that, if Yt D ˚.Xt /, .˝; F ; .Ft /t ; .Yt /t ; P/ is a Markov process
associated to q, which therefore turns out to be a transition function (note that
it is a Markov process with respect to the same filtration).
c) Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be an m-dimensional Brownian motion and
z 2 Rm . Let Yt D jXt zj; Yt is therefore the distance of Xt from z. Show that
.˝; F ; .Ft /t ; .Yt /t ; P/ is a Markov process.
c) Note that if Z N.0; I/, then the transition function of the Brownian motion can be written
as
p
p.s; x; A/ D P. sZ 2 A x/ : (6.42)
Then use the rotational invariance of the distributions N.0; I/.
• Condition (6.40) is called Dynkin’s criterion. It is a simple sufficient condition

for Y to be a Markov process (not necessary however).
• The Markov process introduced in c) is called Bessel process of dimension m. It
is a diffusion on RC which will also be the object of Exercise 8.24.
Chapter 7
The Stochastic Integral
7.1 Introduction
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a (continuous) standard Brownian motion fixed

once and for all: the aim of this chapter is to give a meaning to expressions of the
form
Z T
Xs .!/ dBs .!/ (7.1)
0
where the integrand .Xs /0sT is a process enjoying certain properties to be

specified. As already remarked in Sect. 3.3, this cannot be done path by path as the
function t 7! Bt .!/ does not have finite variation a.s. The r.v. (7.1) is a stochastic
integral and it will be a basic tool for the construction and the investigation of
new processes. For instance, once the stochastic integral is defined, it is possible
to consider a stochastic differential equation: it will be something of the form
dXt D b.Xt / dt C .Xt / dBt ; (7.2)
where b and are suitable functions. To solve it will mean to find a process .Xt /t
such that for every t 0
Z t Z t
Xt D X0 C b.Xs / ds C .Xs / dBs ;
0 0
which is well defined, once the stochastic integral is given a rigorous meaning.
One can view the solution of (7.2) as a model to describe the behavior of objects
following the ordinary differential equation xP t D b.xt /, but whose evolution is also
influenced by random perturbations represented by the term .Xt / dBt .

182 7 The Stochastic Integral
The idea of the construction of the stochastic integral is rather simple: imitating
the definition the Riemann integral, consider first the integral of piecewise constant
processes, i.e. of the form
X
n1
Xt D Xk 1Œtk ;tkC1 Œ .t/ ; (7.3)
kD0
where a D t0 < t1 < < tn D b. For these it is natural to set

Z b X
Xt dBt D Xk .BtkC1 Btk / :
a kD0
Then with processes as in (7.3) we shall approximate more general situations.

Although the idea is simple, technical difficulties will not be in short supply, as
we shall see.
Once the integral is defined
R t we shall be concerned with the investigation of the
newly defined process t 7! 0 Xs dBs . The important feature is that, under suitable
assumptions, it turns out to be a martingale.
7.2 Elementary processes
Let us define first the spaces of processes that will be the integrands of the stochastic
integral.
p
We denote by Mloc .Œa; b/ the space of the equivalence classes of real-valued
progressively measurable processes X D .˝; F ; .Ft /atb ; .Xt /atb ; P/
such that
Z b
jXs jp ds < C1 a.s. (7.4)
a
p
By M p .Œa; b/ we conversely denote the subspace of Mloc .Œa; b/ of the
processes such that
Z b
E jXs jp ds < C1 : (7.5)
a
p
Mloc .Œ0; C1Œ/ (resp. M p .Œ0; C1Œ/) will denote the space of the processes
p
.Xt /t such that .Xt /0tT lies in Mloc .Œ0; T/ (resp. M p .Œ0; T/) for every T > 0.
7.2 Elementary processes 183
Speaking of equivalence classes, we mean that we identify two processes X and X 0

whenever they are indistinguishable, which can also be expressed by saying that
Z b
jXs Xs0 j ds D 0 a.s.
a
Remarks 7.1
a) It is immediate that a continuous and adapted process .Xt /t belongs to
p
Mloc .Œa; b/ for every p 0. Indeed continuity, in addition to the fact
of being adapted, implies progressive measurability (Proposition 2.1).
Moreover, (7.4) is immediate as s 7! jXs .!/jp , being continuous, is
automatically integrable on every bounded interval. By the same argument,
p
multiplying a process in Mloc .Œa; b/ by a bounded progressively measur-
p
able process again gives rise to a process in Mloc .Œa; b/.
p
b) If X 2 Mloc (resp. M ) and is a stopping time of the filtration .Ft /t ,
p
p
then the process t 7! Xt 1ft< g also belongs to Mloc (resp. M p ). Indeed the
process t 7! 1ft< g is itself progressively measurable (it is adapted and
right-continuous) and moreover, as it vanishes for t > ,
Z b Z b^ Z b
jXs j 1fs< g ds D
p
jXs j ds
p
jXs jp ds
a a a
so that condition (7.4) (resp. (7.5)) is also satisfied.

c) It is not difficult to prove that M 2 is a Hilbert space with respect to the
scalar product
Z b Z b
hX; YiM2 D E Xs Ys ds D E.Xs Ys / ds ;
a a
being a closed subspace of L2 .˝ RC ; P ˝ Œa;b /, where Œa;b denotes the

Lebesgue measure of Œa; b.
d) Quite often we shall deal with the problem of checking that a given
progressively measurable process .Xt /t belongs to M 2 .Œa; b/. This is not
difficult. Indeed, by Fubini’s theorem
Z b Z b
2
E jXs j ds D E.jXs j2 / ds
a a
and therefore it is enough to ascertain that s 7! E.jXs j2 / is, for instance,

continuous. The Brownian motion itself is in M 2 , for instance.
p
Among the elements of Mloc there are, in particular, those of the form
X
n1
Xt .!/ D Xi .!/1Œti ;tiC1 Œ .t/ ; (7.6)
iD0
where a D t0 < t1 < < tn D b and, for every i, Xi is a real Fti -measurable r.v.
The condition that Xi is Fti -measurable is needed for the process to be adapted and
ensures progressive measurability (a process as in (7.6) is clearly right-continuous).
We shall call these processes elementary. As
hZ b i hX
n1 i Xn1
E Xt2 dt D E Xi2 .tiC1 ti / D E.Xi2 /.tiC1 ti /
a iD0 iD0
we have that X 2 M 2 .Œa; b/ if and only if the r.v.’s Xi are square integrable.
2
Definition 7.1 Let X 2 Mloc .Œa; b/ be an elementary process as in (7.6). The
Rb
stochastic integral of X (with respect to B), denoted a Xt dBt , is the r.v.
X
n1
Xi .BtiC1 Bti / : (7.7)
iD0
It is easy to see that the stochastic integral is linear in X. Moreover,
Lemma 7.1 If X is an elementary process in M 2 .Œa; b/, then

Z b ˇ
E Xt dBt ˇ Fa D 0 ;
a
h Z b 2 ˇ i hZ b ˇ i
E Xt dBt ˇ Fa D E Xt2 dt ˇ Fa :
a a
In particular, the stochastic integral of an elementary process of M 2 .Œa; b/ is

a centered square integrable r.v. and
h Z b 2 i hZ b i
E Xt dBt DE Xt2 dt : (7.8)
a a
7.3 The stochastic integral 185
Pn1
Proof Let Xt D iD0 Xi 1Œti ;tiC1 Œ .t/; as Xi is square integrable and Fti -measurable
and BtiC1 Bti is independent of Fti , we have EŒXi .BtiC1 Bti /jFti D Xi EŒBtiC1
Bti D 0 and therefore

E.Xi .BtiC1 Bti /jFa / D E EŒXi .BtiC1 Bti /jFti jFa D 0 ;
from which the first relation follows. We also have
h Z b 2 ˇ i hX
n1
ˇ i
E Xt dBt ˇ Fa D E Xi Xj .Bt Bti /.BtjC1 Btj / ˇ Fa : (7.9)
iC1
a i;jD0
Note first that, as Xi is Fti -measurable and therefore independent of BtiC1 Bti , the
r.v.’s Xi2 .BtiC1 Bti /2 are integrable being the product of integrable independent r.v.’s
(Proposition 1.3). Therefore the r.v. Xi Xj .BtiC1 Bti /.BtjC1 Btj / is also integrable,
being the product of the square integrable r.v.’s Xi .BtiC1 Bti / and Xj .BtjC1 Btj /.
We have, for j > i,

E Xi Xj .BtiC1 Bti /.BtjC1 Btj /jFa

ˇ
D E E Xi Xj .BtiC1 Bti /.BtjC1 Btj /jFtj ˇ Fa

D EŒXi Xj .BtiC1 Bti / E.BtjC1 Btj jFtj / jFa D 0 :
„ ƒ‚ …
D0
Therefore in (7.9) the contribution of the terms with i 6D j vanishes. Moreover,

E Xi2 .BtiC1 Bti /2 jFa D E E Xi2 .BtiC1 Bti /2 jFti jFa D EŒXi2 .tiC1 ti /jFa :
Therefore, going back to (7.9),

h Z b 2 ˇ i hX
n1
ˇ i hZ b ˇ i
E Xt dBt ˇ Fa D E 2 ˇ
X .tiC1 ti / Fa D E Xt2 dt ˇ Fa :
i
a iD0 a
t
u
In the proof of Lemma 7.1, note the crucial role played by the assumption that the
r.v.’s Xi are Fti measurable, which is equivalent to requiring that X is progressively
measurable.
7.3 The stochastic integral
In order to define the stochastic integral for processes in M p we need an approxima-

tion result stating that a process in M p can be suitably approximated with elementary
processes.
Given a function f 2 Lp .Œa; b/, p 1, and an equi-spaced grid n D fa D t0 <

t1 < < tn D bg such that tiC1 ti D 1n .b a/, let
X
n1
Gn f D fi 1Œti ;tiC1 Œ ;
iD0
where fi D 0 if i D 0 and
Z Z
1 ti
n ti
fi D f .s/ ds D f .s/ ds
ti ti1 ti1 ba ti1
is the average of f on the interval at the left of ti . As, by Jensen’s inequality,

ˇ Z ˇp Z Z
ˇ 1 ti
ˇ 1 ti
n ti
j fi jp D ˇ f .s/ dsˇ j f .s/jp ds D j f .s/jp ds ;
ti ti1 ti1 ti ti1 ti1 ba ti1
we have
Z X
n1 n Z
X Z
b
j fi jp tiC1 b
jGn f .s/jp ds D .b a/ j f .s/jp ds D j f .s/jp ds
a iD0
n iD1 ti a
(7.10)
and therefore also Gn f 2 Lp .Œa; b/. As G2n f D G2n.G2nC1 f /, the same argument
leading to (7.10) gives
Z b Z b
jG f .s/j ds
2n
p
jG2nC1f .s/jp ds : (7.11)
a a
Gn f is a step function and let us to prove that

Z b
lim jGn f .s/ f .s/jp ds D 0 : (7.12)
n!1 a
Actually, this is immediate if f is continuous, owing to uniform continuity:

Rt let, for a
fixed ", n be such that j f .x/ f .y/j " if jx yj 2n .b a/ and a 1 j f .s/jp ds "p .
For i 1, if s 2 Œti ; tiC1 Œ and u 2 Œti1 ; ti Œ, then ju sj 2n .b a/ so that
j f .s/ f .u/j " and also jGn f .s/ f .s/j ", as Gn f .s/ is the average of f on
Œti1 ; ti Œ. Then we have
Z b Z t1 Z b
jGn f .s/ f .s/j ds D p
j f .s/j ds C p
jGn f .s/ f .s/jp ds "p .1 C b a/
a a t1
which proves (7.12) if f is continuous. In order to prove (7.12) for a general function
f 2 Lp .Œa; b/ one has just to recall that continuous functions are dense in Lp , the
details are left to the reader.
p
Lemma 7.2 Let X 2 Mloc .Œa; b/, then there exists a sequence of elementary
p
processes .Xn /n Mloc .Œa; b/ such that
Z b
lim jXt Xn .t/jp dt D 0 a.s. (7.13)
n!1 a
If X 2 M p .Œa; b/ then there exists a sequence of elementary processes .Xn /n

M p .Œa; b/ such that
Z b
lim E jXt Xn .t/jp dt D 0 : (7.14)
n!1 a
Moreover, the elementary processes Xn can be chosen in such a way that n 7!

Rb
a jXn j ds is increasing.
p
p
Proof If X 2 Mloc then t 7! Xt .!/ is a function in Lp a.s. Let us define
Xn D Gn X. Such R ti a process Xn is adapted as on the interval Œti ; tiC1 Œ it takes
the value ti t1iC1 ti1 Xs ds, which is Fti -measurable, as explained in Remark 2.2.
Finally, (7.13) follows from (7.12).
Let X 2 M p and let n be the equi-spaced grid defined above. We know already
that Xn D Gn X is adapted and moreover, thanks to (7.10),
Z b Z b
jXn .s/j ds
p
jX.s/jp ds
a a
so that Xn 2 M .Œa; b/ and, moreover, by (7.10) we have for every n

p
Z b Z b Z b Z b
jXt Xn .t/jp dt 2p1 jXt jp dt C jXn .t/jp dt 2p jXt jp dt ;
a a a a
hence we can take the expectation in (7.13) and obtain (7.14) using Lebesgue’s
theorem.
Finally the last statement is a consequence of (7.11). t
u
We can now define the stochastic integral for every process X 2 M 2 .Œa; b/: (7.8)
states that the stochastic integral is an isometry between the elementary processes
of M 2 .Œa; b/ and L2 .P/. Lemma 7.2 says that these elementary processes are dense
in M 2 .Œa; b/, so that the isometry can be extended to the whole M 2 .Œa; b/, thus
defining the stochastic integral for every X 2 M 2 .Œa; b/.
In practice, in order to compute the stochastic integral of a process X 2

M 2 .Œa; b/, by now, we must do the following: we must first determine the sequence
Gn X of elementary processes approximating X in the sense of (7.14), and then
compute the limit
Z b
lim Xn .s/ dBs
n!1 a
in the L2 sense. This procedure does not look appealing, but soon we shall see (in the
next chapter) other ways of computing the stochastic integral. This is similar to what
happens with the ordinary Riemann integral: first one defines the integral through
an approximation with step functions and then finds much simpler and satisfactory
ways of making the actual computation with the use of primitives.
Let us investigate the first properties of the stochastic integral. The following
extends to general integrands the properties already known for the stochastic integral
of elementary processes.
Theorem 7.1 If X 2 M 2 .Œa; b/ then

Z b ˇ
E Xt dBt ˇ Fa D 0 ; (7.15)
a
E Xt dBt ˇ Fa D E Xt2 dt ˇ Fa : (7.16)
a a
In particular,
h Z b 2 i hZ b i
E Xt dBt DE Xt2 dt : (7.17)
a a
Proof Let .Xn /n be a sequence of elementary processes approximating X as stated

in Lemma 7.2. Then we know by Lemma 7.1 that (7.15) is true for the stochastic
Rb
integral of processes of Xn and we can take the limit as n ! 1 as a Xn .t/ dBt !
Rb 2
a Xt dBt in L . As for (7.16) we have
E Xn .t/ dBt ˇ Fa D E Xn2 .t/ dt ˇ Fa :
a a
Rb Rb
We can take the limit as n ! 1 at the left-hand side as a Xn .t/ dBt ! a Xt dBt
Rb 2 Rb 2
in L2 hence a Xn .t/ dBt ! a Xt dBt in L1 and we can use Remark 4.3. As
Rb
for the right-hand side we can assume that n 7! a jXn .t/j2 dt is increasing and then
use Beppo Levi’s theorem for the conditional expectation, Proposition 4.2 a). t
u
Remark 7.1 If X; Y 2 M 2 .Œa; b/, then

Z b Z b Z b
E Xs dBs Ys dBs D E.Xs Ys / ds : (7.18)
a a a
Indeed, it is sufficient to apply (7.17) to X C Y and X Y and then develop

and subtract:
Z b 2 Z b 2 Z b 2 Z b Z b
.Xs C Ys / dBs D Xs dBs C Ys dBs C2 Xs dBs Ys dBs
a a a a a
Z b 2 Z b 2 Z b 2 Z b Z b
.Xs Ys / dBs D Xs dBs C Ys dBs 2 Xs dBs Ys dBs
a a a a a
and therefore
Z b Z b
4E Xs dBs Ys dBs
a a
h Z b 2 i h Z b 2 i
DE .Xs C Ys / dBs E .Xs Ys / dBs
a a
hZ b i hZ b i hZ b i
2 2
DE .Xs C Ys / ds E .Xs Ys / ds D 4E Xs Ys ds :
a a a
Examples 7.1
a) Note that the motion Brownian B itself belongs to M 2 . Is it true that
Z 1
1 2
Bs dBs D B ‹ (7.19)
0 2 1
Of course no, as
Z 1
E Bs dBs D 0 whereas E.B21 / D 1 :
0
We shall see, however, that (7.19) becomes true if an extra term is added.
b) Also .B2t /t belongs to M 2 . What is the value of
hZ 1 Z 1 i
E Bs dBs B2s dBs ‹
0 0
(continued )
Examples 7.1 (continued)

Easy
hZ 1 Z 1 i hZ 1 i Z 1
E Bs dBs B2s dBs D E Bs B2s ds D EŒB3s ds D 0 :
0 0 0 0
c) What is the value of

hZ 2 Z 3 i
E Bs dBs Bs dBs ‹
0 1
The two stochastic integrals are over different intervals. If we set

( (
Bs if 0 s 2 0 if 0 s < 1
Xs D Ys D
0 if 2 < s 3 Bs if 1 s 3
then
hZ 2 Z 3 i hZ 3 Z 3 i
E Bs dBs Bs dBs D E Xs dBs Ys dBs
0 1 0 0
Z 3 Z 2
3
D EŒXs Ys ds D s ds D
0 1 2
Of course, this argument can be used in general in order to compute the

expectation of the product of stochastic integrals over different intervals.
In particular, the expectation of the product of two stochastic integrals of
processes of M 2 over disjoint intervals vanishes.
Remark 7.2 As the stochastic integral is an isometry M 2 ! L2 .˝; F ; P/, we

can ask whether every r.v. of L2 .˝; F ; P/ can be obtained as the stochastic
integral of a process of M 2 . One sees immediately that this cannot be as,
by (7.15), every stochastic integral of a process in M 2 has zero mean. We shall
see, however, in Sect. 12.3 that if .Ft /t is the natural augmented filtration of
the Brownian motion B, then every r.v. Z 2 L2 .˝; FT ; P/ can be represented
in the form
Z T
Z D cC Xs dBs
0
with X 2 M 2 .Œ0; T/ and c 2 R. This representation has deep and important
applications that we shall see later.
By the isometry property of the stochastic integral (7.17)

h Z b Z b 2 i h Z b 2 i
E Xn .s/ dBs Xs dBs DE .Xn .s/ Xs / dBs
a a a
Z b
DE jXn .t/ Xt j2 dt
a
so that
Corollary 7.1 If Xn ; X 2 M 2 .Œa; b/ are processes such that

Z b
lim E jXn .t/ Xt j2 dt D 0
n!1 a
then
Z b Z b
L2
Xn .s/ dBs ! Xs dBs :
a n!1 a
Note that we cannot say that the value of the integral at ! depends only on the
paths t 7! Xt .!/ and t 7! Bt .!/, as the integral is not defined pathwise. However,
we have the following
Theorem 7.2 Let X; Y 2 M 2 .Œa; b/ and A 2 F such that Xt D Yt on A for

every t 2 Œa; b. Then on A we have a.s.
Z b Z b
Xt dBt D Yt dBt :
a a
Proof Let .Xn /n and .Yn /n be the sequences of elementary processes that respec-
tively approximate X and Y in M 2 .Œa; b/ constructed on p. 186, i.e. Xn D Gn X,
Yn D Gn Y. A closer look at the definition of Gn shows that Gn X and R b Gn Y also
coincide on A for every t 2 Œa; b and for every n, and therefore also a Xn dBt and
Rb
a Yn dBt . By definition we have
Z b Z b
L2
Xn .t/ dBt ! X.t/ dBt
a n!1 a
Z b Z b
L2
Yn .t/ dBt ! Y.t/ dBt
a n!1 a
and then just recall that L2 -convergence implies a.s. convergence for a subsequence.
t
u
7.4 The martingale property
Let X 2 M 2 .Œ0; T/, then the restriction

Rt of X to Œ0; t, t T also belongs to M 2 .Œ0; t/
and we can consider its integral 0 Xs dBs ; let, for 0 t T, the real-valued process
I be defined as
Z t
I.t/ D Xs dBs :
0
It is clear that
Rt
a) if t > s, I.t/ I.s/ D s Xu dBu ,
b) I.t/ is Ft -measurable for every t. Indeed if .Xn /n is the sequence
Rt of elementary
processes that approximates X in M 2 .Œ0; t/ and In .t/ D 0 Xn .s/ dBs , then it
is immediate that In .t/ is Ft -measurable, given the definition of the stochastic
integral for the elementary processes. Since In .t/ ! I.t/ in L2 as n ! 1, there
exists a subsequence .nk /k such that Ink .t/ ! I.t/ a.s. as k ! 1; therefore
I.t/ is also Ft -measurable (remember that we assume that the Brownian motion
B is standard so that, in particular, Ft contains the negligible events of F and
changing an Ft -measurable r.v. on a negligible event still produces an Ft -measu-
rable r.v.)
Note that if
X
n1
Xt D Xi 1Œti ;tiC1 Œ .t/
iD0
is an elementary process then we can write

Z t X
n1
I.t/ D Xs dBs D Xi .BtiC1 ^t Bti ^t / : (7.20)
0 iD0
7.4 The martingale property 193
In particular, if X is an elementary process then I is a continuous process adapted

to the filtration .Ft /t .
A key property of the process I is the following.
Theorem 7.3 If X 2 M 2 .Œ0; T/, then I is an .Ft /t -square integrable mar-

tingale. Moreover, I has a continuous version and its associated increasing
process is
Z t
At D Xs2 ds : (7.21)
0
Proof If t > s we have, thanks to Theorem 7.1,

Z t ˇ
EŒI.t/ I.s/jFs D E Xu dBu ˇ Fs D 0 a.s.
s
and therefore we have the martingale relation
EŒI.t/jFs D I.s/ C EŒI.t/ I.s/jFs D I.s/ a.s.
Theorem 7.1 also states that I.t/ is square integrable. In order to check that A is
the associated increasing process to the martingale I, we need to verify that Zt D
I.t/2 A.t/ is a martingale. With the decomposition I.t/2 A.t/ D ŒI.s/ C .I.t/
I.s//2 A.s/ .A.t/ A.s// and remarking that I.s/ and A.s/ are Fs -measurable,
EŒI.t/2 A.t/jFs

D I.s/2 A.s/ C E 2I.s/.I.t/ I.s// C .I.t/ I.s//2 .A.t/ A.s//jFs

D I.s/2 A.s/ C 2I.s/ EŒI.t/ I.s/jFs CE .I.t/ I.s//2 .A.t/ A.s//jFs

„ ƒ‚ …
D0
which allows us to conclude that the process defined in (7.21) is the associate
increasing process as, thanks to (7.16), we have a.s.
h Z t 2 ˇ i hZ t ˇ i
EŒ.I.t/ I.s//2 jFs D E Xu dBu ˇ Fs D E Xu2 du ˇ Fs
s s
D EŒA.t/ A.s/jFs :
Let us prove the existence of a continuous version. We already know that this is
true for elementary processes. Let .Xn /n be a sequence of elementary processes in
M 2 .Œ0; T/ such that

Z T
lim E jXn .t/ X.t/j2 dt D 0 :
n!1 0
Rt
Let In .t/ D 0 Xn .s/ dBs . The idea of the existence of a continuous version is simple:
just find a subsequence .nk /k such that the sequence of processes .Ink /k converges
uniformly a.s. The limit, J say, will therefore be a continuous process. As we know
already that
lim Ink .t/ D I.t/ in L2

k!1
we shall have that J.t/ D I.t/ a.s., so that J will be the required continuous version.
In order to prove the uniform convergence of the subsequence .Ink /k we shall write
X
nk
Ink .t/ D In1 .t/ C .Ini .t/ Ini1 .t// (7.22)
iD1
and prove that sup0tT jIni .t/ Ini1 .t/j is the general term of a convergent series.
As .In .t/ Im .t//t is a square integrable continuous martingale, by the maximal
inequality (5.16) applied to the supermartingale .jIn .t/ Im .t/j2 /t and for D "2 ,

P sup jIn .t/ Im .t/j > " D P inf jIn .t/ Im .t/j2 < "2
0tT 0tT
Z
1 2

1 h T 2 i
2 E jIn .T/ Im .T/j D 2 E Xn .s/ Xm .s/ dBs
" " 0
Z T
1
D 2E jXn .s/ Xm .s/j2 ds :
" 0
As .Xn /n is a Cauchy sequence in M 2 , the right-hand side can be made arbitrarily

small for n and m large. If we choose " D 2k then there exists an increasing
sequence .nk /k such that, for every m > nk ,
Z T 1
P sup jInk .t/ Im .t/j > 2k 22k E jXnk .s/ Xm .s/j2 ds 2
0tT 0 k
and therefore
1
P sup jInk .t/ InkC1 .t/j > 2k 2
0tT k
1
But k2
is the general term of a convergent series and, by the Borel–Cantelli lemma,

P sup jInk .t/ InkC1 .t/j > 2k infinitely many times D 0 :
0tT
Therefore a.s. we have eventually
sup jInk .t/ InkC1 .t/j 2k :

0tT
Therefore the series in (7.22) converges uniformly, which concludes the proof. t
u
Rt
From now on, by I.t/ or 0 Xs dBs we shall understand the continuous version.
Theorem 7.3 allows us to apply to the stochastic integral .I.t//t all the nice
properties of square integrable continuous martingales that we have pointed out in
Sect. 5.6. First of all that the paths of I do not have finite variation (Theorem 5.15)
unless they are a.s. constant, which can happen only if Xs D 0 a.s for every s.
The following inequalities also hold
ˇZ t ˇ Z
1 T 2
ˇ ˇ
P sup ˇ Xs dBs ˇ > 2 E Xs ds ; for every > 0
0tT 0 0
h Z
t 2 i Z
h T i (7.23)
E sup Xs dBs 4E Xs2 ds :
0tT 0 0
In fact the first relation follows from the maximal inequality (5.16) applied to
ˇRt ˇ2
the supermartingale Mt D ˇ 0 Xs dBs ˇ : (5.16) states that, for a continuous
supermartingale M,

P inf Mt EŒjMT j
0tT
so that
ˇZ t ˇ ˇZ t ˇ2
ˇ ˇ ˇ ˇ
P sup ˇ Xs dBs ˇ > D P sup ˇ Xs dBs ˇ 2
0tT 0 0tT 0
ˇZ t ˇ2 Z
1 ˇˇ T ˇ2 Z
1 T 2
ˇ ˇ ˇ
D P inf ˇ Xs dBs ˇ 2 2 E ˇ Xs dBs ˇ D 2 E Xs ds :
0tT 0 0 0
As for the second one, we have, from Doob’s inequality (Theorem 5.12),
h Z t 2 i h Z t 2 i
E sup Xs dBs 4 sup E Xs dBs
0tT 0 0tT 0
hZ t i hZ T i
D 4 sup E Xs2 ds D 4E Xs2 ds :
0tT 0 0
As a consequence, we have the following strengthening of Corollary 7.1.
Theorem 7.4 Let Xn ; X 2 M 2 .Œ0; T/ and let us assume that

hZ T i
E jXn .s/ Xs j2 ds ! 0
0 n!1
then
ˇZ t Z t ˇ2
ˇ ˇ
lim E sup ˇ Xn .s/ dBs Xs dBs ˇ D 0 :
n!1 0tT 0 0
Obviously if f 2 L2 .Œ0; T/, then f 2 M 2 .Œ0; T/. In this case the stochastic integral
enjoys an important property.
Proposition 7.1 If f 2 L2 .Œ0; T/ then the process

Z t
It D f .s/ dBs (7.24)
0
is Gaussian.
Proof Let us prove that, for every choice of 0 s1 < < sm T, the r.v. I D
.I.s1 /; :P
: : ; I.sm // is Gaussian. This fact is immediate if f is piecewise constant: if
f .t/ D niD1 i 1Œti1 ;ti Œ , then
X
n
Is D i .Bti ^s Bti1 ^s /
iD1
and the vector I is therefore Gaussian, being a linear function of the r.v.’s Bti ^sj
that are jointly Gaussian. We know that there exists a sequence . fn /n L2 .Œ0; T/
of piecewise constant functions converging to f in L2 .Œ0; T/ and therefore in M 2 .
Rt Rt
Let In .t/ D 0 fn .s/ dBs , I.t/ D 0 f .s/ dBs . Then, by the isometry property of the
stochastic integral (Theorem 7.1), for every t,
Z t
2 L2
E.jIn .t/ It j / D . fn .u/ f .u//2 ds k fn f k22 ! 0
0 n!1
so that In .t/ !n!1 It in L2 and therefore, for every t, It is Gaussian by the properties
of the Gaussian r.v.’s under L2 convergence (Proposition 1.9). Moreover, if 0 s1
sm , then In D .In .s1 /; : : : ; In .sm // is a jointly Gaussian r.v. As it converges,
for n ! 1, to I D .I.s1 /; : : : ; I.sm // in L2 , the random vector I is also jointly
Gaussian, which concludes the proof. t
u
Remark 7.3 Proposition 7.1 implies in particular that

h Z t i 1 Z t
E exp f .s/ dBs D exp f .s/2 ds : (7.25)
0 2 0
Indeed we recognize on the left-hand side the mean of the exponential of a

centered Gaussian r.v. (i.e. its Laplace transform computed at
D 1), which
is equal to the exponential of its variance divided by 2 (see Exercise 1.6).
If is a stopping time of the filtration .Ft /t then, thanks to Corollary 5.6, the process
Z t^
I.t ^ / D Xs dBs
0
R t^
is a .Ft /t -martingale. In particular, E 0 Xs dBs D 0. The following statement is
often useful.
Theorem 7.5 Let be a stopping time of the filtration .Ft /t , with T.

Then if X 2 M 2 .Œ0; T/ also .Xt 1ft< g /t 2 M 2 .Œ0; T/ and
Z Z T
Xs dBs D Xs 1fs< g dBs a.s. (7.26)
0 0
Proof .Xt 1ft< g /t 2 M 2 .Œ0; T/ because the process t ! 1ft< g is bounded, adapted
and right-continuous hence progressively measurable.
(7.26) is immediate if takes at most finitely many values t1 ; : : : ; tn . Actually we

have a.s.
Z X
n Z tk X
n Z tk
#
Xs dBs D 1f Dtk g Xs dBs D 1f Dtk g Xs 1fs< g dBs
0 kD1 0 kD1 0
Z T
D Xs 1fs< g dBs ;
0
where the equality marked with # is a consequence of Theorem 7.2, as on f D tk g

the processes .Xs /s and .Xs 1fs< g/s coincide on the time interval Œ0; tk Œ.
For a general stopping time , let .n /n be a sequence of stopping times taking at
most finitely many values and decreasing to (see Lemma 3.3); possibly replacing
n with n ^ T we can assume n T for every n. Then, the paths of the stochastic
integral being continuous,
Z n Z
lim Xs dBs D Xs dBs a.s. (7.27)
n!1 0 0
Moreover,
Z Z
T ˇ ˇ T
ˇXs 1fs< g Xs 1fs< gˇ2 ds D Xs2 1f s<ng ds :
n
0 0
As limn!1 1f s<n g D 1f Dsg , by Lebesgue’s theorem

Z Z
T ˇ ˇ T
lim ˇXs 1fs< g Xs 1fs< gˇ2 ds D Xs2 1f Dsg ds D 0
n
n!1 0 0
and again by Lebesgue’s theorem

hZ T ˇ ˇ i
lim E ˇXs 1fs< g Xs 1fs< gˇ2 ds D 0 :
n
n!1 0
Therefore, by Corollary 7.1,

Z T Z T
lim Xs 1fs<n g dBs D Xs 1fs< g dBs in L2 :
n!1 0 0
This relation together with (7.27) allows us to conclude the proof of the lemma. u
t
Let X 2 M 2 .Œ0; T/ and let 1 and 2 be stopping times with 1 2 T.
The following properties follow easily from the stopping theorem, Theorem 7.5,
and the martingale property, Theorem 7.3, of the stochastic integral.
Z 1
E Xt dBt D 0 (7.28)
0
h Z 1 2 i Z 1
E Xt dBt DE Xt2 dt (7.29)
0 0
Z 2 ˇ Z 1
E Xt dBt ˇ F1 D Xt dBt ; a.s. (7.30)
0 0
h Z 2 2 ˇ i Z 2 ˇ
E Xt dBt ˇ F1 D E Xt2 dt ˇ F1 ; a.s. (7.31)
1 1
Let Œa; b be an interval such that Xt .!/ D 0 for almost every a t b. Is it true
that t 7! It is constant
˚ onRŒa; b?
Let D inf tI t > a; a Xs2 ds > 0 with the understanding that D b if f g D ;.
t
Then by Lemma 7.5 and (7.23)

ˇZ t^ ˇ2
2 ˇ ˇ
E sup jIt^ Ia j DE sup ˇ Xu dBu ˇ
atb atb a
ˇZ t ˇ2 Z b Z
ˇ ˇ
DE sup ˇ Xu 1fu< g dBu ˇ 4E Xu2 1fu< g du D 4E Xu2 du D 0 ;
atb a a a
where the last equality follows from the fact that Xu .!/ D 0 for almost every u 2
a; .!/Œ. Therefore there exists a negligible event Na;b such that, for ! … Na;b ,
t 7! It .!/ is constant on a; Œ.
Proposition 7.2 Let X 2 M 2 .Œ0; T/. Then there exists a negligible event N
such that, for every ! … N and for every 0 a < b T, if Xt .!/ D 0 for
almost every t 2a; bŒ, then t 7! It .!/ is constant on a; bŒ.
Proof We know already that for every r; q 2 Q\Œ0; T there exists a negligible event
Nr;q such that if Xt .!/ D 0 for almost every t 2r; qŒ, then t 7! It .!/ is constant on
r; qŒ.
Let N be the union of the events Nr;q ; r; q 2 Q \ Œ0; T. N is the negligible event
we were looking for, as if ! … N, then if Xt .!/ D 0 on a; bŒ, t 7! It .!/ is constant
a.s. on every interval r; sŒa; bŒ having rational endpoints and therefore also on
a; bŒ. t
u
2
7.5 The stochastic integral in Mloc
2
We can now define the stochastic integral of a process X 2 Mloc .Œ0; T/. The main
2 2
idea is to approximate X 2 Mloc .Œ0; T/ by processes
Rt in M .Œ0; T/.
Let, for every n > 0, n D infft TI 0 Xs2 ds > ng with the understanding
RT
n D T if 0 Xs2 ds n. Then n is a stopping time and the process Xn .t/ D Xt 1ft<n g
belongs to M 2 .Œ0; T/. Indeed, thanks to Theorem 7.5,
Z T Z T Z n ^T
2 2
Xn .s/ ds D X.s/ 1fs<ng ds D Xs2 ds n
0 0 0
so that Xn 2 M 2 . We can therefore define, for every n 0, the stochastic integral

Z t
In .t/ D Xn .s/ dBs :
0
RT
Let us observe that, as 0 Xs2 ds < C1 by hypothesis, n % T a.s. as n ! C1.
Moreover, if n > m then n > m , so that fn D Tg fm D Tg and the processes
Xn and Xm coincide on fm D Tg. By S Proposition 7.2 In and Im also coincide on
fm D Tg a.s. for every t. As ˝ D n fn D Tg, the sequence .In /n is eventually
constant and therefore converges a.s. to some r.v. I.
2
Definition 7.2 Let X 2 Mloc .Œ0; T/, then its stochastic integral is defined as
Z t Z t
It D Xs dBs D lim Xs 1fs<n g dBs a.s.
0 n!1 0
2
The stochastic integral of a process X 2 Mloc .Œ0; T/ is obviously continuous, as
it coincides with In on fn > Tg and In is continuous as a stochastic integral of a
process in M 2 .
If X 2 M 2 then also X 2 Mloc 2
. Let us verify that in this case Definition 7.2
coincides with the definition given for processes of M 2 in Sect. 7.3, p. 187. Note that
if X 2 M 2 .Œ0; T/, then
Z T Z T
E jXs Xn .s/j2 ds D E jXs j2 ds (7.32)
0 n
since Xs D Xn .s/ if s < n , whereas Xn .s/ D 0 on s n . Therefore the right-hand

side in (7.32) tends to 0 as n ! 1, thanks to Lebesgue’s theorem, so that
Z T
lim E jXs Xn .s/j2 ds D 0 :
n!1 0
2
7.5 The stochastic integral in Mloc 201
Thanks to Corollary 7.1,

ˇZ t Z t ˇ2
ˇ ˇ
lim E sup ˇ Xs dBs Xn .s/ dBs ˇ D 0
n!1 0tT 0 0
and therefore if X 2 M 2 .Œ0; T/ the two definitions coincide.
Remark 7.4 The statement of Theorem 7.2 remains true for stochastic inte-
2
grals of processes of Mloc .Œ0; T. Indeed if Xt D Yt on ˝0 for every t 2 Œa; b,
then this also true for the approximants Xn , Yn . Therefore the stochastic
integrals
Z T Z T
Xn .s/ dBs and Yn .s/ dBs
0 0
also coincide on ˝0 and therefore their a.s. limits also coincide.
We now point out some properties of the stochastic integral when the integrand is
2
a process in X 2 Mloc .Œa; b/. Let us first look for convergence results of processes
2
Xn ; X 2 Mloc .Œa; b/. We shall see that if the processes Xn suitably approximate X
2
in Mloc .Œa; b/, then the stochastic integrals converge in probability. The key tool in
this direction is the following.
2
Lemma 7.3 If X 2 Mloc .Œa; b/, then for every " > 0, > 0
ˇ Z b ˇ Z b
ˇ ˇ
P ˇ Xt dBt ˇ > " P Xt2 dt > C 2
a a "
Rt
Proof Let D infftI t a; a Xs2 ds g. Then we can write
ˇ Z
ˇ b
ˇˇ
Xt dBt ˇ > "
P ˇ
ˇ Z b ˇ a
ˇ Z b ˇ
ˇ ˇ ˇ ˇ
DP ˇ Xt dBt ˇ > "; > T C P ˇ Xt dBt ˇ > "; T (7.33)
a
ˇ Z b ˇ a
ˇ ˇ
P ˇ Xt dBt ˇ > "; > T C P. T/ :
a
As the two processes t 7! Xt and t 7! Xt 1f >t g coincide on ˝0 D f > Tg and by

Remark 7.4,
ˇ Z b ˇ ˇ Z b ˇ
ˇ ˇ ˇ ˇ
P ˇ Xt dBt ˇ > "; > T D P ˇ Xt 1f >tg dBt ˇ > "; > T
a a
ˇ Z b ˇ
ˇ ˇ
P ˇ Xt 1f >tg dBt ˇ > " :
a
Therefore, as
ˇ Z b ˇ2 Z b^
ˇ ˇ
E ˇ Xt 1f >tg dBt ˇ D E Xt2 dt ;
a a
by Chebyshev’s inequality
ˇ Z b ˇ
ˇ ˇ
P ˇ Xt dBt ˇ > "; > T 2
a "
and as
Z b
P. T/ D P Xt2 dt > ;
a
going back to (7.33) we can conclude the proof. t

u
2
Proposition 7.3 Let X; Xn 2 Mloc .Œa; b/, n 1, and let us assume that
Z b
P
jX.t/ Xn .t/j2 dt ! 0:
a n!1
Then
Z b Z b
P
Xn .t/ dBt ! X.t/ dBt :
a n!1 a
Proof By Lemma 7.3

ˇ Z b Z b ˇ ˇ Z b ˇ
ˇ ˇ ˇ ˇ
P ˇ Xn .t/ dBt Xt dBt ˇ > " D P ˇ .Xn .t/ Xt / dBt ˇ > "
a a a
Z b
P jXn .t/ Xt j2 dt > C 2
a "
2
7.5 The stochastic integral in Mloc 203

Let > 0 and let us first choose so that "2
2 and then n0 such that for n > n0
Z b
P jXt Xn .t/j2 dt > <
a 2
Therefore for n > n0

ˇ Z b Z b ˇ
ˇ ˇ
P ˇ Xn .t/ dBt Xt dBt ˇ > " < ;
a a
which allows us to conclude the proof. t

u
The following proposition states, similarly to the Lebesgue integral, that if the
integrand is continuous then the integral is the limit of the Riemann sums. Note,
however, that the limit is only in probability and that the integrand must be computed
at the left end point of every small interval (see Exercise 7.9 to see what happens if
this rule is not respected).
2
Proposition 7.4 If X 2 Mloc .Œa; b/ is a continuous process, then for every
sequence .n /n of partitions a D tn;0 < tn;1 < < tn;mn D b with jn j D
max jtnkC1 tnk j ! 0 we have
n 1
mX Z b
P
X.tn;k /.Btn;kC1 Btn;k / ! X.t/ dBt :
n!1 a
kD0
Proof Let
n 1
mX
Xn .t/ D X.tn;k /1Œtn;k ;tn;kC1 Œ .t/ :
kD0
Xn is an elementary process and
n 1
mX Z b
X.tn;k /.Btn;kC1 Btn;k / D Xn .t/ dBt :
kD0 a
As the paths are continuous,

Z b
jXn .t/ X.t/j2 dt ! 0 a.s.
a n!1
Hence by Proposition 7.3,
n 1
mX Z b Z b
P
X.tn;k /.Btn;kC1 Btn;k / D Xn .t/ dBt ! X.t/ dBt :
a n!1 a
kD0
t
u
Remark 7.5 It is worthwhile to point out that we defined the stochastic

integral under the assumption that the integrand is progressively measurable,
which implies, intuitively, that its value does not depend on the future
increments of the Brownian motion.
This hypothesis is essential in the derivation of some of the most important
properties of the stochastic integral, as already pointed out at the end of the
proof of Lemma 7.1. Exercise 7.9 shows that strange things happen when
one tries to mimic the construction of the stochastic integral without this
assumption.
For the usual integral multiplying constants can be taken in and out of the integral
sign. This is also true for the stochastic integral, but a certain condition is necessary,
which requires attention.
Theorem 7.6 Let Z be a real Fa -measurable r.v.; then for every X 2

2
Mloc .Œa; b/
Z b Z b
ZXt dBt D Z Xt dBt : (7.34)
a a
Proof Let us consider the case of an elementary integrand
X
m
Xt D Xi 1Œti ;tiC1 Œ .t/ :
iD1
P
Then ZX is still an elementary process. Indeed ZXt D niD1 ZXi 1Œti ;tiC1 Œ .t/ and, as Z
is Fti -measurable for every i D 1; : : : ; m, the r.v.’s ZXi remain Fti -measurable (here
the hypothesis that Z is Fa -measurable and therefore Fti -measurable for every i is
crucial). It is therefore immediate that the statement is true for elementary processes.
Once the statement is proved for elementary processes it can be extended first to
processes in M 2 and then in Mloc
2
by straightforward methods of approximation. The
details are left to the reader. t
u
7.6 Local martingales 205
Note also that if the condition “Z Fa -measurable” is not satisfied, the rela-
tion (7.34) is not even meaningful: the process t 7! ZXt might not be adapted and
the stochastic integral in this case has not been defined.
7.6 Local martingales
We have seen that, if X 2 M 2 .Œ0; C1Œ/, .It /t is a martingale and this fact has been
fundamental in order to derive a number of important properties of the stochastic
2
integral as a process. This is not true in general if X 2 Mloc .Œ0; C1Œ/: It in this
case might not be integrable. Nevertheless, let us see in this case how .It /t can be
approximated with martingales.
Definition 7.3 A process M D .˝; F ; .Ft /t ; .Mt /t ; P/ is said to be a local

martingale if there exists an increasing sequence .n /n of stopping times such
that
i) n % C1 as n ! 1 a.s.
ii) .Mt^n /t is a .Ft /t -martingale for every n.
We shall say that a sequence .n /n as above reduces the local martingale M.
It is clear that every martingale is a local martingale (just choose n C1).

The important motivation of this definition is that the stochastic integral of a
2
processR t X2 2 Mloc is a local martingale: going back to Definition 7.2, if n D
infftI 0 Xs ds > ng and
Z t
It D Xs dBs
0
then
Z t^n Z t
It^n D Xs dBs D Xs 1fs<n g dBs
0 0
and the right-hand side is a (square integrable) martingale, being the stochastic
integral of the process s 7! Xs 1fs<n g which belongs to M 2 .
Remark 7.6
a) If M is a local martingale, then Mt might be non-integrable. However, M0
is integrable. In fact M0 D M0^n and .Mt^n /t is a martingale.
b) Every local martingale has a right-continuous modification. Actually, this
is true for all the stopped martingales .Mt^n /t by Theorem 5.14, so that
t 7! Mt has a right continuous modification for t n for every n. Now
just observe that n ! C1 as n ! 1 a.s.
c) In Definition 7.3 we can always assume that .Mt^n /t is a uniformly
integrable martingale for every n. If .n /n reduces M the same is true for
n D n ^ n. Indeed, condition i) is immediate. Also the fact that .Mt^n /t
is a martingale is immediate, being the martingale .Mn ^t /t stopped at time
n. Therefore .n /n also reduces M. Moreover, Mn is integrable because
Mn D Mn^n and .Mt^n /t is a martingale. We also have
Mt^n D EŒMn jFt : (7.35)
In fact for t n we have EŒMn jFt D EŒMn^n jFt D Mt^n ,

thanks to the martingale property of .Mt^n /t , whereas, if t > n, Mn D
Mn^n is already Ft -measurable. (7.35) implies that .Mt^n /t is uniformly
integrable.
d) If, moreover, M is continuous, then we can assume that the stopped
martingale .Mt^n /t is bounded. Let n D infftI jMt j ng and n D n ^n ,
then the sequence .n /n again satisfies the conditions of Definition 7.3 and,
moreover, jMn ^t j n for every t.
We shall always assume that the stopped martingales .Mt^n /t are uniformly
integrable and, if .Mt /t is continuous, that they are also bounded, which is always
possible thanks to the previous remark.
As remarked above, in general a local martingale need not be integrable.
However, it is certainly integrable if it is positive, which is one of the consequences
of the following result.
Proposition 7.5 A positive local martingale is a supermartingale.
Proof Let .n /n be a sequence of stopping times that reduces M. Then Mt^n ! Mt

as n ! C1 and by Fatou’s lemma
E.Mt / lim E.Mt^n / D E.M0 / ;

n!1
7.6 Local martingales 207
therefore Mt is integrable. We must now prove that if A 2 Fs and s < t, then

E.Mt 1A / E.Ms 1A /, i.e. that
EŒ.Mt Ms /1A 0 : (7.36)
We have
EŒ1A\fn >sg .Mt^n Ms / D EŒ1A\fn >sg .Mt^n Ms^n / D 0 : (7.37)
Indeed A \ fn > sg 2 Fs and .Mt^n /t is a martingale, which justifies the last

equality. Also we used the fact that on fn > sg the r.v.’s Ms and Ms^n coincide. As
lim 1A\fn >sg .Mt^n Ms / D 1A .Mt Ms /

n!1
and
1A\fn>sg .Mt^n Ms / Ms ;
we can apply Fatou’s lemma (we proved above that Ms is integrable), which gives
0 D lim EŒ1A\fn >sg .Mt^n Ms / EŒ1A .Mt Ms / ;

n!1
completing the proof. t

u
Remark 7.7 A bounded local martingale M is a martingale. This is almost

obvious as, if the sequence .n /n reduces M, then
E.Mt^n jFs / D Ms^n
and we can take the limit as n ! 1 using Lebesgue’s theorem for conditional
expectations (Proposition 4.2).
Note, however, that other apparently strong assumptions are not sufficient
to guarantee that a local martingale is a martingale. For instance, there
exist uniformly integrable local martingales that are not martingales (see
Example 8.10). A condition for a local martingale to be a true martingale
is provided in Exercise 7.15.
Proposition 7.6 Let M be a continuous local martingale. Then there exists

a unique continuous increasing process .At /t , which we continue to call the
associated increasing process to M, such that Xt D Mt2 At is a continuous
local martingale.
The proof is not difficult and consists in the use of the definition of a local
martingale in order to approximate M with square integrable martingales, for which
Theorem 5.16 holds. Instead of giving the proof, let us see what happens in the
situation that is of greatest interest to us.
2
Rt
Proposition 7.7 If X 2 Mloc .Œ0; T/ and It D 0 Xs dBs , then .It /0tT is a
local martingale whose increasing process is
Z t
At D Xs2 ds :
0
Proof We know already that I is a local martingale. In order to complete

R t the proof
we must prove that .It2 At /t is a local martingale. Let n D infftI 0 Xs2 ds > ng,
then we know already that .n /n reduces I and that
Z t^n Z t
It^n D Xs dBs D 1fs<n g Xs dBs
0 0
is a square integrable .Ft /t -martingale. By Proposition 7.3 and Lemma 7.5,

Z t 2 Z t
2
It^n
At^n D 1fs<ng Xs dBs 1fs<n g Xs2 ds
0 0
is a martingale, which means that .n /n reduces .It2 At /t , which is therefore a local

martingale. t
u
We shall still denote by .hMit /t the associated increasing process of the local
martingale M.
Corollary 7.2 If M and N are continuous local martingales, then there exists
a unique process A with finite variation such that Zt D Mt Nt At is a
continuous local martingale.
The proof of the corollary boils down to the observation that Mt Nt D 14 ..Mt CNt /2
.Mt Nt /2 /, so that At D 14 .hM C Nit hM Nit / satisfies the requirement. .At /t
is a process with finite variation, being the difference of two increasing processes.
We will denote by hM; Ni the process with finite variation of Corollary 7.2.
Exercises
7.1 (p. 523) Let B be a Brownian motion.

a) Compute
Z 1
ZD 1fBt D0g dBt :
0
b) Let
Z 1
ZD 1fBt 0g dBt :
0
Compute EŒZ and Var.Z/.

• Note that the processes t 7! 1fBt D0g and t 7! 1fBt 0g are progressively
measurable, thanks to Exercise 2.3.
7.2 (p. 523) Compute

Z t
E Bs Bu dBu :
0
Rt
7.3 (p. 523) Let B be a standard Brownian motion and Yt D 0 es dBs . If
Z t
Zt D Ys dBs ;
0
compute E.Zt /, E.Zt2 / and E.Zt Zs /.

7.4 (p. 524)
a) Compute, for s t,
h Z t 2 i
E B2s Bu dBu :
s
Rt
b) Prove that, if X 2 M 2 .Œs; t/, the r.v. Z D s Xu dBu is uncorrelated with Bv ; v s,
but, in general, it is not independent of Fs .
7.5 (p. 525) Let

Z t
2
Xt D eBs dBs :
0
Prove that X is a square integrable martingale. Is it bounded in L2 ?

7.6 (p. 525) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a two-dimensional standard

Brownian motion and let
Z t Z s
1 2
Zt D dB1 .s/ eB1 .u/ dB2 .u/ :
0 1 C 4s 0
a) Is Z a martingale? Determine the processes .hZit /t and .hZ; B1 it /t .

b) Prove that the limit Z1 D limt!C1 Zt exists a.s. and in L2 and compute EŒZ1
and Var.Z1 /.
7.7 (p. 526)

Rt
a) Let f 2 L2 .Œs; t/; show that W D s f .u/ dBu is independent of Fs .
b) Let ˚ W RC ! RC be a strictly increasing differentiable function such that
˚.0/ D 0 and let
Z ˚ 1 .t/ p
B˚
t D ˚ 0 .u/ dBu :
0
Then B˚ is a Brownian motion. With respect to which filtration?

Rt
7.8 (p. 527)R Show that the r.v. 0 B2u du is orthogonal in L2 .˝; F ; P/ to every r.v.
s
of the form 0 f .u/ dBu , for every s > 0 and f 2 L2 .Œ0; s/.
7.9 (p. 528) Let B D .˝; F ; .Ft /t ; .Bt /t / be a Brownian motion and let 0 D t0 <
t1 < < tn D t be a partition of the interval Œ0; t and jj D maxiD0;:::;n1 jtiC1 ti j
the amplitude of the partition. Mr. Whynot decides to approximate the stochastic
integral
Z t
Xs dBs
0
for a continuous adapted process X with the Riemann sums
X
n1
XtiC1 .BtiC1 Bti / (7.38)
iD0
instead of the sums
X
n1
Xti .BtiC1 Bti / (7.39)
iD0
as indicated (and recommended!) in Proposition 7.4. He argues that, if this were

an ordinary integral instead of a stochastic integral, then the result would be the
same.
a) Compute the limit of the sums (7.38) as the amplitude of the partition tends to 0
when X D B.
b) Do the same thing assuming that the process X has paths with finite variation.
7.10 (p. 529) Let B be a Brownian motion and f 2 L2 .RC /.

a) Show that the limit
Z t
lim f .s/ dBs
t!C1 0
exists in L2 . Does it also exist a.s.? RT

b) Let g W RC ! RC be a function such that 0 g.s/2 ds < C1 for every T > 0.
b1) Show that
Rt Rt
g.s/ dBs 12 g.s/2 ds
Yt D e 0 0
is a martingale. Compute EŒYt2 .

b2) Prove that if g 2 L2 .RC / then .Yt /t is uniformly integrable.
R C1
b3) Assume, conversely, that 0 g.s/2 ds D C1. Compute EŒYt˛ for ˛ < 1.
What is the value of limt!C1 Yt now? Is .Yt /t uniformly integrable in this
case?
7.11 (p. 530) a) For t < 1 let B a Brownian motion and

Z t
dBs
Yt D .1 t/
0 1s
a1) Show that .Yt /t is a Gaussian process and compute E.Yt Ys /, 0 s < t < 1.
Does this remind you of a process we have already met?
a2) Show that the limit limt!1 Yt exists in L2 and compute it.
b) Let A.s/ D 1Cs
s
and let
Z A.s/
dBu
Ws D
0 1u
Show that .Ws /s is a Brownian motion and deduce that limt!1 Yt D 0 a.s.
7.12 (p. 531) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and > 0. Let
Z t Z t
Yt D e .ts/ dBs ; Zt D e s dBs :
0 0
a) Prove that, for every t > 0, Yt and Zt have the same law and compute it.
b) Show that .Zt /t is a martingale. And .Yt /t ?
c) Show that limt!C1 Zt exists a.s. and in L2 .
d1) Show that limt!C1 Yt exists in law and determine the limit law.
d2) Show that
1
lim EŒ.YtCh Yt /2 D .1 e h /
t!C1
and therefore .Yt /t cannot converge in L2 .
7.13 (p. 532) Let B be a Brownian motion and let e

B0 D 0 and
Z t
e 12u 10u2
Bt D 3 C 2 dBu :
0 t t
a) Show thatR e
B is a natural Brownian motion.
1
b) Let Y D 0 u dBu . Show that Y is independent of eBt for every t 0.
c) Show that the -algebra generated by e Bt , t 0 is strictly smaller than the -
algebra generated by Bt , t 0.
Rt
7.14 (p. 533) Let X 2 M 2 .Œ0; T/. We know that 0 Xs dBs is square integrable. Is
2
the converse true? That is if X 2 Mloc .Œ0; T/ and its stochastic integral is square
integrable, does this imply that X 2 M 2 .Œ0; T/ and that its stochastic integral is
a square integrable martingale? The answer is no (a counterexample is given in
Exercise 8.21). This exercise goes deeper into this question.
2
LetR.˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and let X 2 Mloc .Œ0; T/ and
t
Mt D 0 Xs dBs .
Rt RT
a) Let n D infftI 0 Xs2 ds > ng with the understanding that n D T if 0 Xs2 ds n.
Prove that
h Z n i h i
E Xs2 ds E sup Mt2 :
0 0tT
b) Prove that X 2 M 2 .Œ0; T/ if and only if

h i
E sup Mt2 < C1 : (7.40)
0tT
7.15 (p. 533) Let M D .˝; F ; .Ft /t ; .Mt /t ; P/ be a local martingale and assume
that for every t > 0 the family .Mt^ / is uniformly integrable, with ranging
among all stopping times of .Ft /t . Prove that .Mt /t is a martingale.
Chapter 8
Stochastic Calculus
8.1 Ito’s formula
Let X be a process such that, for every 0 t1 < t2 T,

Z t2 Z t2
Xt2 Xt1 D Ft dt C Gt dBt ;
t1 t1
1 2
where F 2 Mloc .Œ0; T/ and G 2 Mloc .Œ0; T/. We say then that X admits the
stochastic differential
dXt D Ft dt C Gt dBt :
A process admitting a stochastic differential is called an Ito process. An Ito process

is therefore the sum of a process with finite variation and of a local martingale.
Remark 8.1 The stochastic differential is unique: if there existed A1 ; A2 2

1 2
Mloc .Œ0; T/, G1 ; G2 2 Mloc .Œ0; T/ such that
A1 .t/ dt C G1 .t/ dBt D A2 .t/ dt C G2 .t/ dBt ;
then, for every 0 t T, we would have a.s.

Z t Z t
def def
At D .A1 .s/ A2 .s// ds D .G2 .s/ G1 .s// dBs D Gt
0 0
and this is not possible, unless the two integrands are identically zero, as the
left-hand side is a process with finite variation, whereas the right-hand side,
which is a local martingale, is not.

216 8 Stochastic Calculus
We shall write
Z t
hXit D G2s ds :
0
hXi is nothing else than the increasing process (see Proposition 7.6) associated to
the local martingale appearing in the definition of X, and it is well defined thanks to
the previous remark. Similarly, if Y is another Ito process with stochastic differential
dYt D Ht dt C Kt dBt ;
we shall set
Z t
hX; Yit D Gs Ks ds :
0
Example 8.1 Let us compute the stochastic differential of Xt D B2t . The

analogy with standard calculus might suggest that dXt D 2Bt dBRt , but we
have already remarked that this cannot be as it would give B2t D 2 0 Bs dBs ,
t
which is impossible because the stochastic integral is a centered r.v. whereas

B2t is not.
Let 0 t1 < t2 T and n D ft1 D tn;1 < tn;2 < < tn;mn D t2 g be a
partition of Œt1 ; t2 such that jn j ! 0, then by Proposition 7.4, meaning the
limits in probability,
Z t2 n 1
mX
Bt dBt D lim Btn;k ŒBtn;kC1 Btn;k
t1 n!1
kD1
n 1
mX
1 2
D lim ŒBtn;kC1 B2tn;k ŒBtn;kC1 Btn;k 2
2 n!1 kD1
n 1
mX
1 2 1
D ŒBt2 B2t1 lim ŒBt Btn;k 2 :
2 2 n!1 kD1 n;kC1
By Proposition 3.4 the rightmost limit is equal to t2 t1 in L2 ; therefore

Z t2
1 2 1
Bt dBt D .B B2t1 / .t2 t1 / ;
t1 2 t2 2
i.e.
dB2t D dt C 2Bt dBt : (8.1)

8.1 Ito’s formula 217
Example 8.2 Xt D tBt . With the notations of the previous example
Z t2 n 1
mX
t dBt D lim tn;k .Btn;kC1 Btn;k /
t1 n!1
kD1
Z t2 n 1
mX
Bt dt D lim Btn;k .tn;kC1 tn;k / :
t1 n!1
kD1
Let us compute the limit, in probability,
mX
n 1 n 1
mX
lim Btn;k .tn;kC1 tn;k / C tn;k .Btn;kC1 Btn;k /
n!1
kD1 kD1
n mX
n 1
D lim Btn;k .tn;kC1 tn;k / C tn;kC1 .Btn;kC1 Btn;k / (8.2)
n!1
kD1
mXn 1 o
C .tn;k tn;kC1 /.Btn;kC1 Btn;k / :
kD1
But as
ˇ mX
n 1 ˇ
ˇ ˇ
ˇ .tn;kC1 tn;k /.Btn;kC1 Btn;k /ˇ
kD1
n 1
mX
max jBtn;kC1 Btn;k j jtn;kC1 tn;k j ;
kD1;:::;mn 1
kD1
„ ƒ‚ …
Dt2 t1
thanks to the continuity of the paths of B, the quantity above converges to 0

a.s. Therefore the limit in (8.2) is equal to
n 1
mX
lim Btn;k .tn;kC1 tn;k / C tn;kC1 .Btn;kC1 Btn;k /
n!1
kD1
n 1
mX
D lim Œtn;kC1 Btn;kC1 tn;k Btn;k :
n!1
kD1
(continued )

Hence
Z t2 Z t2 n 1
mX
Bt dt C t dBt D lim Œtn;kC1 Btn;kC1 tn;k Btn;k D t2 Bt2 t1 Bt1 ;
t1 t1 n!1
kD1
i.e.
d.tBt / D Bt dt C t dBt : (8.3)
Proposition 8.1 If Xi , i D 1; 2, are processes with stochastic differentials
dXi .t/ D Fi .t/ dt C Gi .t/ dBt
then
d.X1 .t/X2 .t// D X1 .t/ dX2 .t/ C X2 .t/ dX1 .t/ C G1 .t/G2 .t/ dt
(8.4)
D X1 .t/ dX2 .t/ C X2 .t/ dX1 .t/ C dhX1 ; X2 it :
Proof If F1 , G1 , F2 , G2 are constant on Œt1 ; t2 , the statement is a consequence

of (8.1) and (8.3). If, conversely, they are elementary processes, let I1 ; : : : ; Ir be
the subintervals of Œt1 ; t2 on which F1 , F2 , G1 , G2 are constants; the statement
follows now using (8.1) and (8.3) applied to each of the intervals Ik and taking
1
the sum. In general, let Fi;n , Gi;n , i D 1; 2, be elementary processes in Mloc .Œ0; T/
2
and Mloc .Œ0; T/, respectively, and such that
Z T
jFi;n .t/ Fi .t/j dt ! 0 a.s.
0 n!1
Z T
jGi;n .t/ Gi .t/j2 dt ! 0 a.s.
0 n!1
and let
Z t Z t
Xi;n .t/ D Xi .0/ C Fi;n .s/ ds C Gi;n .s/ dBs :
0 0
By Theorem 7.4
P
sup jXi;n .t/ Xi .t/j ! 0
0tT n!1
so that, possibly considering a subsequence, Xi;n .t/ ! Xi .t/ as n ! 1 uniformly

a.s. Therefore
Z t2 Z t2
X1;n .t/F2;n .t/ dt ! X1 .t/F2 .t/ dt a.s.
t1 n!1 t1
Z t2 Z t2
X2;n .t/F1;n .t/ dt ! X2 .t/F1 .t/ dt a.s.
t1 n!1 t1
Z t2 Z t2
G1;n .t/G2;n .t/ dt ! G1 .t/G2 .t/ dt a.s.
t1 n!1 t1
and by Theorem 7.4

Z t2 Z t2
P
X1;n .t/G2;n .t/ dBt ! X1 .t/G2 .t/ dBt
t1 n!1 t1
Z t2 Z t2
P
X2;n .t/G1;n .t/ dBt ! X2 .t/G1 .t/ dBt
t1 n!1 t1
and taking the limit as n ! 1 we obtain (8.4). t

u
Note that the formula for the differential of the product of two processes is a
bit different from the corresponding one for the differential of the product of two
functions. Actually there is an extra term, namely
dhX1 ; X2 it :
Note also that if at least one of G1 and G2 vanishes, then this additional term also
vanishes and we have the usual relation
dX1 .t/X2 .t/ D X1 .t/ dX2 .t/ C X2 .t/ dX1 .t/ :
Let f W R RC be a regular function of .x; t/ and X an Ito process. What is the

stochastic differential of the process t 7! f .Xt ; t/? The answer to this question is
Ito’s formula, which will be a key tool of computation from now on. In the next
theorem note again that the differential df .Xt ; t/ behaves as an ordinary differential
plus an extra term.
Theorem 8.1 (Ito’s formula) Let X be a process with stochastic differential
dXt D F.t/ dt C G.t/ dBt
(continued)

and let f W Rx RCt ! R be a continuous function in .x; t/, once continuously
differentiable in t and twice in x. Then
df .Xt ; t/
@f @f 1 @2 f
D .Xt ; t/ dt C .Xt ; t/ dXt C 2
.Xt ; t/G.t/2 dt
@t
@f @x 2 @x
@f 1 @2 f (8.5)
D .Xt ; t/ C .Xt ; t/F.t/ C 2
.Xt ; t/G.t/2 dt
@t @x 2 @x
@f
C .Xt ; t/G.t/ dBt :
@x
Proof (The main ideas)

1ı Step: let us assume f .x; t/ D xn ; (8.5) becomes
dXtn D nXtn1 dXt C 12 n.n 1/Xtn2 G2t dt :
This relation is obvious for n D 1 and follows easily by induction by

Proposition 8.1.
2ı Step: (8.5) is therefore true if f .x; t/ is a polynomial P.x/. Let us assume now
that f is of the form f .x; t/ D P.x/gt . We have
dP.Xt / D ŒP0 .Xt /Ft C 12 P00 .Xt /G2t dt C P0 .Xt /Gt dBt
dgt D g0t dt
and again (8.5) follows easily for such f , thanks to Proposition 8.1.
3ı Step: (8.5) is therefore true if f is of the form
X̀
f .x; t/ D Pi .x/gi .t/ (8.6)
iD1
where the Pi ’s are polynomials and the gi ’s are differentiable functions. If f is

continuous in .x; t/, once continuously differentiable in t and twice in x, one can
prove that there exists a sequence . fn /n of functions of this form such that
fn .x; t/ ! f .x; t/;

@ @ @ @
fn .x; t/ ! f .x; t/; fn .x; t/ ! f .x; t/;
@x @x @t @t
@2 @2
2
fn .x; t/ ! 2 f .x; t/
@x @x
as n ! 1 uniformly on compact sets of R RC (see Friedman 1975, p. 95 for

a proof). We have, for t1 < t2 T,
fn .Xt2 ; t2 / fn .Xt1 ; t1 /
Z h @f i
t2
n @fn 1 @2 fn 2
D .Xt ; t/ C .Xt ; t/Ft C .X t ; t/G t dt
t1 @t Z t2@x 2 @x2 (8.7)
@fn
C .Xt ; t/Gt dBt :
t1 @x
As n ! 1 the first integral on the right-hand side converges toward

Z h @f i
t2
@f 1 @2 f 2
.Xt ; t/ C .Xt ; t/Ft C .X t ; t/G t dt :
t1 @t @x 2 @x2
Actually, t 7! Xt is continuous and therefore bounded on Œt1 ; t2 for every !;

therefore @@t fn .Xt ; t/ converges uniformly to ft .Xt ; t/ and the same holds for the
other derivatives. Moreover,
Z
ˇ @fn ˇ
ˇ .Xt ; t/ @f .Xt ; t/ˇ2 G2 dt
t1
P
! 0
t1 @x @x t n!1
and by Theorem 7.4

Z Z
t2
@fn P
t2
@f
.Xt ; t/Gt dBt ! .Xt ; t/Gt dBt :
t1 @x n!1 t1 @x
We can therefore take the limit as n ! 1 in (8.7) and the statement is proved.
t
u
Example 8.3 As a particular case, if f W R ! R is twice differentiable and
dXt D Ft dt C Gt dBt
then
1 00 1
df .Xt / D f 0 .Xt / dXt C f .Xt /G2t dt D f 0 .Xt / dXt C f 00 .Xt /dhXit :
2 2
In particular,
1 00
df .Bt / D f 0 .Bt / dBt C f .Bt / dt :
2
Example 8.4 If f W RC ! R is differentiable then

Z T Z T
f .s/ dBs D f .T/BT Bs f 0 .s/ ds :
0 0
Actually Ito’s formula (or more simply (8.4)) gives
df .t/Bt D f 0 .t/Bt dt C f .t/ dBt :
Example 8.5 Let us compute the stochastic differential of

Z t
dBs
Zt D .1 t/
0 1s
In Exercise 7.11 it is shown that Y is a Brownian bridge. There are two

possibilities in order to apply Ito’s formula. The first one is to write Zt D Xt Yt
where
Z t
dBs ,
Xt D 1 t; Yt D
0 1 s
i.e.
1
dXt D dt; dYt D dBt :
1t
Proposition 8.1 gives therefore
dZt D Yt dXt C Xt dYt D Yt dt C dBt : (8.8)
Actually here hX; Yit 0. Observing that Zt D .1t/Yt , the previous relation
becomes
1
dZt D Zt dt C dBt ;
1t
which is our first example of a Stochastic Differential Equation.
A second possibility in order to compute the stochastic differential is to
write Zt D g.Yt ; t/, where g.x; t/ D .1 t/x. Now
@g @g @2 g
.x; t/ D x; .x; t/ D 1 t; .x; t/ D 0
@t @x @x2
(continued )

and the computation gives again
@g @g 1
dZt D .Yt ; t/ dt C .Yt ; t/ dYt D Yt dt C dBt D Zt dt C dBt :
@t @x 1t
Example 8.6 In Exercise 5.13 we proved that

Z t
Yt D tBt Bu du
0
is a martingale. Let us see how Ito’s formula makes all these computations
simpler. We have
dYt D t dBt C Bt dt Bt dt D t dBt
i.e., as Y0 D 0,
Z t
Yt D s dBs
0
so that Y, being the stochastic integral of a (deterministic) process of M 2 , is

a square integrable martingale. This representation also makes it particularly
easy to compute the variance of Y, which was done in Exercise 5.13 with
some effort: we have
Z t
2 t3
E.Yt / D s2 ds D
0 3
Example 8.6 points out how Ito’s formula allows to check that a process is a
martingale: first compute its stochastic differential. If the finite variation term
vanishes (i.e. there is no term in dt) then the process is a stochastic integral and
necessarily a local martingale. If, moreover, it is the stochastic integral of a process
in M 2 , then it is a martingale (we shall see in the examples other ways of proving
that a stochastic integral is actually a martingale). This method of checking that a
process is a martingale also provides immediately the associated increasing process.
Remark 8.2 It is possible to give a heuristic explanation of the extra term

appearing in Ito’s formula. Thinking of differentials in the old way we find
1 00
df .Xt / D f 0 .Xt / dXt C f .Xt /.dXt /2 C : : :
2
1 00
D f 0 .Xt /.At dt C Gt dBt / C f .Xt /.At dt C Gt dBt /2 C : : :
2
1 00
D f 0 .Xt /.At dt C Gt dBt / C f .Xt /.A2t dt2 C 2At Gt dt dBt C G2t .dBt /2 / C : : :
2
At this point in ordinary calculus only the first term is considered, all the terms
in dt2 or of higher order being
p negligible in comparison to dt. But now it turns
out that dBt “behaves as” dt so that the term 12 f 00 .Xt /G2t .dBt /2 is no longer
negligible with respect to dt. . .
8.2 Application: exponential martingales
2
If X 2 Mloc .Œ0; T/ and 2 C let
Z Z
t
2 t
Zt D Xs dBs Xs2 ds
0 2 0 (8.9)
Yt D eZt :
Let us compute the stochastic differential of Y. We must apply Ito’s formula to

t 7! f .Zt / with f .x/ D ex : as f 0 .x/ D f 00 .x/ D ex , we find
1
dYt D eZt dZt C eZt dhZit
2
2 2 1 Zt 2 2 (8.10)
D eZt Xt dBt Xt dt C e Xt dt D Xt Yt dBt ;
2 2
i.e., keeping in mind that Y0 D 1,
Z t
Yt D 1 C Ys Xs dBs : (8.11)
0
Therefore Y is a local martingale. If 2 R it is a positive local martingale and, by

Proposition 7.5, a supermartingale; in particular, if 2 R, then
E.Yt / E.Y0 / D 1
for every t T.
8.2 Application: exponential martingales 225
In the future, we shall often deal with the problem of proving that, if 2 R, Y is
a martingale. Indeed, there are two ways to do this:
1) by showing that E.Yt / D 1 for every t (a supermartingale is a martingale if and
only if it has constant expectation, Exercise 5.1);
2) by proving that YX 2 M 2 .Œ0; T/. In this case Y is even square integrable.
We know (Lemma 3.2) the behavior of the “tail” of the distribution of Bt :
1 2
P.jBt j x/ const e 2t x ;
and the reflection principle allows us to state a similar behavior for the tail of the r.v.
sup0tT jBt j. The next statement allows us to say that a similar behavior is shared
by stochastic integrals, under suitable conditions on the integrand.
Proposition 8.2 (Exponential bound) Let X 2 M 2 .Œ0; T/ be such

RT
that 0 Xs2 ds k a.s. for some constant k > 0. Then, for every x > 0,
ˇZ t ˇ x2
ˇ ˇ
P sup ˇ Xs dBs ˇ x 2e 2k : (8.12)
0tT 0
Rt Rt
Proof Let Mt D 0 Xs dBs and At D hMit D 0 Xs2 ds. Then as At k, for every
> 0,
1 2 1 2
fMt xg D fe
Mt e
x g fe
Mt 2
A t
e
x 2
k g :
1 2
The maximal inequality (5.15) applied to the supermartingale t 7! e
Mt 2
A t gives
1 2 1 2
1 2
P sup Mt x P sup e
Mt 2
At e
x 2
k e
xC 2
k :
0tT 0tT
This inequality holds for every

> 0. The minimum of
7!
x C 12
2 k for
> 0
x2
is attained at
D xk and its value is 2k . Substituting this value into the right-hand
side we get
x2
P sup Mt x e 2k : (8.13)
0tT
The same argument applied to M gives

x2
P inf Mt x e 2k ;
0tT
which allows us to conclude the proof. t

u
P
If the process X appearing in (8.9) is of the form m iD1 i Xi .s/ with X1 ; : : : ; Xm 2
2
Mloc .Œ0; T/, D . 1 ; : : : ; m / 2 Cm , then we obtain that if
Z t X
m
1
Z tX
m
Yt D exp i Xi .s/ dBs i j Xi .s/Xj .s/ ds ;
0 iD1 2 0 i;jD1
we have
X
m Z t
Yt D 1 C i Ys Xi .s/ dBs ; (8.14)
iD1 0
i.e. Y is a local martingale. This is the key tool in order to prove the following
important result.
2
Theorem 8.2 Let X 2 Mloc .Œ0; C1Œ/ be such that
Z C1
Xs2 ds D C1 a.s. (8.15)
0
Ru
Then, if .t/ D inffuI 0 Xs2 ds > tg, the process
Z .t/
Wt D Xs dBs
0
is a (continuous) Brownian motion.
Proof We first compute, using (8.14), the finite-dimensional distributions of W,

verifying that they coincide with those of a Brownian motion. Next we shall prove
that W is continuous. Let 0 t1 < < tm ; as .t/ is a finite stopping time for
every t,
Z .tj / Z .tj /^t Z t
Wtj D Xs dBs D lim Xs dBs D lim Xs 1fs< .tj/g dBs :
0 t!C1 0 t!C1 0
(8.16)
In order to determine the joint distribution of .Wt1 ; : : : ; Wtm / we investigate the

characteristic function
h X
m i h X
m Z .tj / i
E exp i
j Wtj D E exp i
j Xs dBs :
jD1 jD1 0
8.2 Application: exponential martingales 227
Let us write (8.14) setting Xj .s/ D Xs 1fs< .tj/g and j D i

j ,
j 2 R:
hX Z Z t i
1 X
m t m
Yt D exp i
j Xj .s/ dBs C

h
k Xh .s/Xk .s/ ds
jD1 0 2 h;kD1 0
hX Z t^ .tj / Z t^ .th /^ .tk / i (8.17)
1 X
m m
D exp i
j Xs dBs C
h
k Xs2 ds :
jD1 0 2 h;kD1 0
We have, by (8.14),
X
m Z t
Yt D 1 C i
j Ys Xj .s/ dBs : (8.18)
jD1 0
Moreover, Yt is bounded for every t as

Z t Z t^ .th /^ .tk /
Xh .s/Xk .s/ ds D Xs2 ds th ^ tk (8.19)
0 0
and therefore, by (8.17),

h1 X
m i
jYt j exp
h
k .th ^ tk / : (8.20)
2 h;kD1
Y is therefore a bounded local martingale, hence a martingale, from which we

deduce that E.Yt / D E.Y0 / D 1 for every t, i.e.
h X Z Z t^ .th /^ .tk / i
t^ .tj /
1 X
m m
E exp i
j Xs dBs C
h
k Xs2 ds D 1 : (8.21)
jD1 0 2 h;kD1 0
But
Z t^ .th /^ .tk / Z .th /^ .tk /
lim Xs2 ds D Xs2 ds D th ^ tk
t!C1 0 0
and therefore, by (8.16) and Lebesgue’s theorem (the r.v.’s Yt remain bounded thanks
to (8.20)), taking the limit as t ! C1,
h Xm
1 X
m i
1 D lim EŒYt D E exp i
j Wtj C
h
k .th ^ tk / ;
t!C1
jD1
2 h;kD1
i.e.
h Xm i h 1 X m i
E exp i
j Wtj D exp
h
k .th ^ tk / ;
jD1
2 h;kD1
which is the characteristic function of a centered multivariate Gaussian r.v. with

covariance matrix hk D th ^ tk . Therefore W has the same finite-dimensional
distributions as a Brownian motion and is actually a Brownian motion. Rt
We now have to prove that the paths are continuous. Let At D 0 Xs2 ds. If this
process, which is obviously continuous, was strictly increasing then we would have
.t/ D A1 .t/. As the inverse of a continuous function is still continuous, W would
be continuous, being the composition of the two applications
Z s
s 7! Xu dBu and t 7! .t/ :
0
If, conversely, A is not strictly increasing, this simple argument does not work, as
t 7! .t/ might be discontinuous. For instance, if A was constant on an interval
Œa; b; a tg b ;
whereas .u/ < a for every u < t; therefore necessarily limu!t .u/ a < b
.t/. The arguments that follow are required in order to fix this technical point.
R t is simple: if A is constant on Œa; bŒ, then Xt D 0 a.e. on Œa; bŒ, and
However, the idea
therefore t 7! 0 Xu dBu is itself constant on Œa; bŒ.
If a W RC ! RC is a non-increasing right-continuous function, its pseudo-
inverse is defined as
c.t/ D inffsI a.s/ > tg : (8.22)
is therefore the pseudo-inverse of A.
Proposition 8.3 Let a be an increasing right-continuous function. Then

a) its pseudo-inverse c is right-continuous.
b) If limt!t0 c.t/ D L < c.t0 /, then a is constant t0 on ŒL; c.t0 /Œ.
Proof
a) Let .tn /n be a sequence decreasing to t and let us prove that limn!1 c.tn / D c.t/;
c being increasing we have limn!1 c.tn / c.t/; let us assume that c.tn / &
L > c.t/ and let us prove that this is not possible. Let u be a number such that
c.t/ < u < L. As u < c.tn / we have a.u/ tn for every n. As c.t/ < u we have
t < a.u/. These two inequalities are clearly incompatible.
b) If L u < c.t0 /, then a.u/ t for every t < t0 and therefore a.u/ t0 . On the
other hand u < c.t0 / implies a.u/ t0 , and therefore a.u/ D t0 . t
u
8.3 Application: Lp estimates 229
End of the Proof of Theorem 8.2 As is the pseudo-inverse of A and is therefore

right-continuous, W is right-continuous. If was left-continuous at t0 then W would
be continuous at t0 . If, conversely, limt!t0 .t/ D L < .t0 /, then A would
be constant on ŒL; .t0 /Œ and therefore Xs D 0 a.e. on ŒL; .t0 /Œ. Therefore by
Proposition 7.2 WL D W .t0 / a.s. and W is also left-continuous a.s. t
u
Rt
Corollary 8.1 Under the hypotheses of Theorem 8.2, if At D 0 Xs2 ds we
have
Z t
WAt D Xs dBs :
0
Proof By definition, A being increasing, .At / D inffsI As > At g t. If .At / D t

the statement follows immediately from Theorem 8.2. If, conversely, .At / D L > t
then A is constant on ŒL; tŒ and therefore Xs 0 on ŒL; tŒ a.s.; then by Proposition 7.2
and Theorem 8.2
Z .At / Z t
WAt D Xs dBs D Xs dBs :
0 0
t
u
Corollary 8.1 states that a stochastic integral is a “time changed” Brownian motion,
i.e. it is a process that “follows the same paths” of some Brownian motion W
but at a speed that changes as a function of t and !. Moreover, the time change
that determines this “speed” is given by a process A that is nothing else than the
associated increasing process of the stochastic integral.
Corollary 8.1 is obviously useful in order toR obtain results concerning the
t
regularity or the asymptotic of the paths of t 7! 0 Xs dBs , which can be deduced
directly from those of the Brownian motion (Hölder continuity, Lévy’s modulus of
continuity, Iterated Logarithm Law, . . . ).
8.3 Application: Lp estimates
We have seen in Chap. 7 some L2 estimates for the stochastic integral (mostly thanks
to the isometry property M 2 $ L2 ). If, conversely, the integrand belongs to M p ,
p > 2, is it true that the stochastic integral is in Lp ? Ito’s formula allows us to
answer positively and to derive some useful estimates. We shall need them in the
next chapter in order to derive regularity properties of the solutions of Stochastic
Differential Equations with respect to parameters and initial conditions.
2
Proposition 8.4 If p 2 and X 2 Mloc .Œ0; T/, then
ˇZ t ˇp h Z T p2 i Z T
ˇ ˇ p2
E sup ˇ Xs dBs ˇ cp E jXs j2 ds cp T 2 E jXs jp ds
0tT 0 0 0
for some constant cp > 0.
Proof One can of course assume X 2 M p .Œ0; R tT/, otherwise the statement is obvious
(the right-hand side is D C1). Let It D 0 Xs dBs and define It D sup0st jIs j.
.It /t is a square integrable martingale and by Doob’s inequality (Theorem 5.12)
ˇZ t ˇp p p
ˇ ˇ
Xs dBs ˇ D EŒIt
p
E sup ˇ sup E jIt jp
0tT 0 p p p 1 0tT (8.23)
D EŒjIT jp :
p1
Let us apply Ito’s formula to the function f .x/ D jxjp (which is twice differentiable,
as p 2) and to the process I whose stochastic differential is dIt D Xt dBt .
We have f 0 .x/ D p sgn.x/jxjp1 ; f 00 .x/ D p. p 1/jxjp2 , where sgn denotes the
“sign” function (D 1 for x 0 and 1 for x < 0). Then by Ito’s formula
1 00
djIt jp D f 0 .It / dIt C f .It / dhIit
2
1
D jIs jp1 sgn.Is /Xs dBs C p. p 1/jIs jp2 Xs2 ds ;
2
i.e., as I0 D 0,
Z Z
t
1 t
jIt j D p
p
jIs j p1
sgn.Is /Xs dBs C p. p 1/ jIs jp2 Xs2 ds : (8.24)
0 2 0
Let us now first assume jIT j K: this guarantees that jIs jp1 sgn.Is /Xs 2 M 2 .Œ0; T/.
Let us take the expectation in (8.24) recalling that the stochastic integral has zero
mean. By Doob’s inequality, (8.23) and Hölder’s inequality with the exponents p2
p
and p2 , we have
p p 1 p p Z T
EŒIT p EŒjIT j D
p
p. p 1/ E jIs jp2 Xs2 ds
p1 2 p1 0
„ ƒ‚ …
WDc0
Z T h Z T p2 i 2p
2
c0 E IT Xs2 ds c0 EŒIT 1 p E Xs2 ds
p2 p
:
0 0
8.4 The multidimensional stochastic integral 231
As we assume jIT j K, EŒIT p < C1 and in the previous inequality we can divide
2
by EŒIT p 1 p , which gives
h Z t p2 i 2p
p 2
EŒIT p c0 E Xs2 ds ;
0
i.e.
h Z t p2 i
EŒIT c0 Xs2 ds
p p=2
E : (8.25)
0
If, conversely, I is not bounded, then let n D infft TI jIt j ng (.n/ D T if

f g D ;). .n /n is a sequence of stopping times increasing to T, as the paths of I are
continuous and then also bounded. We have therefore In ^t ! It as n ! 1 and
Z t^n Z t
It^n D Xs dBs D Xs 1fs<n g dBs
0 0
so that (8.25) gives

h Z T p2 i h Z T p2 i

jXs j2 1fs<n g ds jXs j2 ds
p p=2 p=2
E.IT^n
/ c0 E c0 E ;
0 0
and we can just apply Fatou’s lemma. Finally, again by Hölder’s inequality,
h Z T p2 i p2
hZ T i
E jXs j2 ds T p E jXs jp ds :
0 0
t
u
8.4 The multidimensional stochastic integral
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a standard d-dimensional continuous Brownian

motion and let Bt D .B1 .t/; : : : ; Bd .t//.
A matrix-valued process .s/ D .ij .s//i;j i D 1; : : : ; m, j D 1; : : : ; d, is said

p
to be in Mloc .Œa; b/, 0 a < b, if .ij .s//s is progressively measurable for
every i; j and
Z b
p
i) for every i; j ij 2 Mloc , i.e. jij .s/jp ds < C1 a.s.
a
We say that 2 M .Œa; b/ if

p
hZ b i
i’) for every i; j ij 2 M , i.e. E
p
jij .s/jp ds < C1.
a
Similarly to the one-dimensional case, for fixed d; m, the space M 2 .Œa; b/ turns to
be a Hilbert space with norm
hZ b i
kk2 D E j.s/j2 ds ;
a
where
X
m X
d
2
j.s/j / D ij .s/2 D tr..s/.s/ / :
iD1 jD1
2
Rb
If 2 Mloc .Œa; b/ then, for every i; j, the stochastic integral a ij .t/ dBj .t/ is already
defined. Let
Z b X d Z b
.t/ dBt D ij .t/ dBj .t/ : (8.26)
a a i
jD1
The stochastic integral in (8.26) is an Rm -valued r.v. Note that the matrix .s/ is,
in general, rectangular and that the process defined by the stochastic integral has a
dimension (m) which may be different from the dimension of the Brownian motion
(d). It is clear, by the properties ofRthe stochastic integral in dimension 1, that, if
2 t
2 Mloc .Œ0; T/, the process It D 0 .s/ dBs has a continuous version and every
component is a local martingale (it is easy to see that a sum of local martingales
with respect to the same filtration is still a local martingale).
In the next statements we determine the associated increasing processes to these
local martingales.
Lemma 8.1 Let X1 ; X2 2 M 2 .Œ0; T/ be real processes and

Z t Z t
I1 .t/ D X1 .u/ dB1 .u/; I2 .t/ D X2 .u/ dB2 .u/ :
0 0
Then, for 0 s t T,
Z t Z t ˇ
E X1 .u/ dB1 .u/ X2 .u/ dB2 .u/ ˇ Fs D 0 : (8.27)
s s
In particular, the process
t 7! I1 .t/I2 .t/
is a martingale and hI1 ; I2 it 0.

Proof Let us first prove the statement for elementary processes. Let
X
n
.1/
X
n
.2/
X1 .t/ D Xi 1Œti ;tiC1 Œ ; X2 .t/ D Xi 1Œti ;tiC1 Œ :
jD1 jD1
Then
hZ t Z t ˇ i
ˇ
E X1 .u/ dB1 .u/ X2 .u/ dB2 .u/ ˇ Fs
s s
hX
n
.1/ .2/ ˇˇ i
E Xi B1 .tiC1 / B1 .ti / Xj B2 .tjC1 / B2 .tj / ˇ Fs
(8.28)
i;jD1
X
n
.1/ .2/ ˇ
D E Xi B1 .tiC1 / B1 .ti / Xj B2 .tjC1 / B2 .tj / ˇ Fs :

i;jD1
Note first that all the terms appearing in the conditional expectation on the right-
.k/
hand side are integrable, as the r.v.’s Xi Bk .tiC1 / Bk .ti / are square integrable,
.k/
being the product of square integrable and independent r.v.’s (recall that Xi is Fti -
.1/ .2/
measurable). Now if ti < tj then, as the r.v.’s Xi ; B1 .tiC1 / B1 .ti /; Xj are already
Ftj -measurable,
.1/ .2/ ˇ
E Xi B1 .tiC1 / B1 .ti / Xj B2 .tjC1 / B2 .tj / ˇ Fs

.1/ .2/ ˇ
ˇ
D E E Xi B1 .tiC1 / B1 .ti / Xj B2 .tjC1 / B2 .tj / ˇ Ftj ˇ Fs

.1/ .2/ ˇ
ˇ
D E Xi B1 .tiC1 / B1 .ti / Xj E B2 .tjC1 / B2 .tj / ˇ Ftj ˇ Fs D 0 :

„ ƒ‚ …
D0
Similarly one proves that the terms with i D j in (8.28) also vanish thanks to the
relation
EŒ.B1 .tiC1 / B1 .ti //.B2 .tiC1 / B2 .ti //jFti

D EŒ.B1 .tiC1 / B1 .ti //.B2 .tiC1 / B2 .ti // D 0 ;
where we use first the independence of .B1 .tiC1 / B1 .ti //.B2 .tiC1 / B2 .ti // with
respect to Fti and then the independence of B1 .tiC1 / B1 .ti / and B2 .tiC1 / B2 .ti /.
(8.27) then follows as the stochastic integrals are the limit in L2 of the stochastic
integrals of approximating elementary processes. t
u
Let now 2 M 2 .Œ0; T/ and
Z t
Xt D .u/ dBu : (8.29)
0
What is the associated increasing process of the square integrable martingale .Xi .t//t
(the i-th component of X)? If we define
Z t
Ik .t/ D ik .u/ dBk .u/
0
then Xi .t/ D I1 .t/ C C Id .t/ and
X
d 2 X
d
Xi .t/2 D Ik .t/ D Ih .t/Ik .t/ :
kD1 h;kD1
By Lemma 8.1 Ih Ik , h 6D k, is a martingale. Moreover, (Proposition 7.3)

Z t
Ik .t/2 ik2 .s/ ds
0
is a martingale. Therefore
d Z
X t
Xi .t/2 ik2 .s/ ds (8.30)
kD1 0
is a martingale, i.e.
d Z
X t
hXi it D ik2 .s/ ds : (8.31)
kD1 0
A repetition of the approximation arguments of Sect. 7.6 gives that if, conversely,
2
2 Mloc .Œ0; T/, then the process (8.30) is a local martingale so that again its
associated increasing process is given by (8.31). From (8.31) we have
d Z
X t
hXi C Xj it D .ik .s/ C jk .s//2 ds
kD1 0
d Z
X t
hXi Xj it D .ik .s/ jk .s//2 ds
kD1 0
1
and, using the formula hXi ; Xj it D 4
.hXi C Xj it hXi Xj it /, we obtain
d Z
X t Z t
hXi ; Xj it D ik .s/jk .s/ ds D aij .s/ ds ; (8.32)
kD1 0 0
where a.s/ D .s/.s/ . As a consequence we have the following, which shows

that also in the multidimensional case the stochastic integral is an isometry between
M 2 and L2 .
Proposition 8.5 If 2 M 2 .Œ0; T/ and X is as in (8.29) then

Z t2
E .Xi .t2 / Xi .t1 //.Xj .t2 / Xj .t1 // D E aij .s/ ds
t1
(recall that a.s/ D .s/.s/ ) and

ˇ Z t2 ˇ2 Z t2 Z t2
ˇ ˇ 2
E ˇ .t/ dBt ˇ D E j.t/j dt D E tr a.t/ dt : (8.33)
t1 t1 t1
Proof As both Xi and Xj are martingales, we have for t1 t2

EŒXi .t2 /Xj .t1 / D E EŒXi .t2 /Xj .t1 /jFt1 D EŒXi .t1 /Xj .t1 / :
Therefore

E .Xi .t2 / Xi .t1 //.Xj .t2 / Xj .t1 //

D E Xi .t2 /Xj .t2 / C Xi .t1 /Xj .t1 / Xi .t1 /Xj .t2 / Xi .t2 /Xj .t1 /

D E Xi .t2 /Xj .t2 / Xi .t1 /Xj .t1 /
but as the process .Xi .t/Xj .t//t is equal to a martingale vanishing at 0 plus the process
hXi ; Xj it we have

Z t2
D E hXi ; Xj it2 hXi ; Xj it1 D E aij .s/ ds :
t1
Now (8.33) follows because
X
m Z t2 X
m
E.jX.t2 / X.t1 /j2 / D EŒ.Xi .t2 / Xi .t1 //2 D E aii .t/ dt :
iD1 t1 iD1
t
u
As is the case for the stochastic integral in dimension 1, the martingale property
allows us to derive some important estimates. The following is a form of Doob’s
maximal inequality.
Proposition 8.6 If 2 M 2 .Œ0; T/ then

h ˇZ t ˇ2 i hZ T i
ˇ ˇ
E sup ˇ .s/ dBs ˇ 4E j.s/j2 ds :
0tT 0 0
Proof
h ˇZ t ˇ2 i h m Z t X
X d 2 i
ˇ ˇ
E sup ˇ .s/ dBs ˇ D E sup ij .s/ dBj .s/
0tT 0 0tT iD1 0 jD1
X
m h Z t X
d 2 i
E sup ij .s/ dBj .s/ :
iD1 0tT 0 jD1
By Doob’s inequality (Theorem 5.12),
h Z t X
d 2 i h Z T X
d 2 i
E sup ij .s/ dBj .s/ 4E ij .s/ dBj .s/
0tT 0 jD1 0 jD1
Z hX i Z
T d T
D4 E ih .s/ih .s/ ds D 4 E . /ii .s/ ds

0 hD1 0
and now just take the sum in i, recalling that j.s/j2 D tr .s/.s/ . t
u
We say that an Rm -valued process X has stochastic differential
dXt D Ft dt C Gt dBt ;
1 2
where F D .F1 ; : : : ; Fm / 2 Mloc .Œ0; T/ and G D .Gij / iD1;:::;m 2 Mloc .Œ0; T/, if for
jD1;:::;d
every 0 t1 < t2 T we have
Z t2 Z t2
Xt2 Xt1 D Ft dt C Gt dBt :
t1 t1
Again do not be confused: the Brownian motion here is d-dimensional whereas the
process X turns out to be m-dimensional. Note that, also in the multidimensional
case, the stochastic differential is unique: just argue coordinate by coordinate, the
In the multidimensional case Ito’s formula takes the following form, which
extends Theorem 8.1. We shall not give the proof (which is however similar).
C
Theorem 8.3 Let u W Rm x Rt ! R be a continuous function with
continuous derivatives ut ; uxi , uxi xj ; i; j D 1; : : : ; m. Let X be a process having
stochastic differential
dXt D Ft dt C Gt dBt :
Then the process .u.Xt ; t//t has stochastic differential
X
m
du.Xt ; t/ D ut .Xt ; t/ dt C uxi .Xt ; t/ dXi .t/
iD1
(8.34)
1X
m
C ux x .Xt ; t/Aij .t/ dt ;
2 i;jD1 i j
where A D GG .
Thanks to (8.32) we have Aij .t/ dt D dhXi ; Xj it , so that (8.34) can also be written as
X
m
1X
m
du.Xt ; t/ D ut .Xt ; t/ dt C uxi .Xt ; t/ dXi .t/ C ux x .Xt ; t/dhXi ; Xj it :
iD1
2 i;jD1 i j
Example 8.7 Let u W Rd ! R be a twice differentiable function. Then,

denoting by u0 the gradient of u,
1
du.Bt / D u0 .Bt / dBt C 4 u.Bt / dt : (8.35)
2
Actually in this case dhBi ; Bj it D 0 for i 6D j and dhBi ; Bi it D dt so that
1X
d
1
uxi xj .Bt /dhBi ; Bj it D 4 u.Bt / dt :
2 i;jD1 2
In particular, if i 6D j, applying the previous remark to the function u.x/ D

xi xj , whose Laplacian is equal to 0,
d.Bi .t/Bj .t// D Bi .t/ dBj .t/ C Bj .t/ dBi .t/ : (8.36)
We find again that Bi Bj is a martingale for i 6D j (see Exercise 5.24). In

addition we obtain immediately the associated increasing process A of this
(continued )

martingale: as
Z t Z t 2
2 2
Bi .t/ Bj .t/ D Bi .t/ dBj .t/ C Bj .t/ dBi .t/
0 0
Z t 2 Z t 2 Z t Z t
D Bj .t/ dBi .t/ C Bi .t/ dBj .t/ C2 Bj .t/ dBi .t/ Bi .t/ dBj .t/
0 0 0 0
and we know that the last term on the right-hand side is already a martingale
(Lemma 8.1), we have that if
Z t
At D .Bi .s/2 C Bj .s/2 / ds
0
then t 7! Bi .t/2 Bj .t/2 At is a martingale. Hence A is the requested associated

increasing process.
Example 8.8 (Bessel square processes) What is the stochastic differential of

Xt D jBt j2 ?
We must compute the stochastic differential of u.Bt / where u W Rd ! R is
given by u.x/ D jxj2 D x21 C C x2d . Obviously
@u @2 u
.x/ D 2xi ; .x/ D 2ıij
@xi @xi @xj
i.e. u0 .x/ D 2x whereas 4u D 2d. Therefore by (8.35)
dXt D d dt C 2Bt dBt (8.37)
and we find again another well-known fact: jBt j2 dt is a martingale. X is a

Bessel square process.
2
If
D .
1 ; : : : ;
m / 2 Rm , .s/ D .ij .s// iD1;:::;m 2 Mloc .Œ0; T/; let a D and
jD1;:::;d
Z t
Xt D .s/ dBs ; (8.38)
0
which is an m-dimensional process. Let

˝ Z t Z
˛ 1 t
Yt
D exp
; .s/ dBs ha.s/
;
i ds ; (8.39)
0 2 0
then, by Ito’s formula applied to the function u.x/ D eh

;xi and to the process
Z Z
t
1 t
Zt
D .s/ dBs a.s/
ds
0 2 0
so that Yt
D u.Zt
/, we obtain
X
m Z t X
m X
d Z t
Yt
D1C
h Ys
dXh .s/ D 1 C
h Ys
hj .s/ dBj .s/ : (8.40)
hD1 0 hD1 jD1 0
Y
is a positive local martingale and therefore a supermartingale. In a way similar
to Proposition 8.2 we get
Proposition 8.7 Let 2 M 2 .Œ0; T/ and

2 Rm with j
j D 1. Then if there
RT
exists a constant k such that 0 ha.s/
;
i ds k we have
x2
P sup jh
; Xt ij x 2e 2k : (8.41)
0tT
RT
Moreover, if 0 ha.s/
;
i ds k for every vector
2 Rm of modulus 1,
x2
P sup jXt j x 2me 2k m : (8.42)
0tT
Proof The first part of the statement comes from a step by step repetition of the
proof of Proposition 8.2: we have for > 0

P sup jh
; Xt ij x D P sup jh
; Xt ij x
0tT 0tT
Z
1 T 2 1
P sup jh
; Xt ij ha.s/
;
i ds x 2 k
0tT 2 0 2
Z
1 T 2 1 2
D P sup exp jh
; Xt ij ha.s/
;
i ds e x 2 k :
0tT 2 0
RT
Now t 7! exp.jh
; Xt ij 12 0 2 ha.s/
;
i ds/ is a continuous supermartingale
whose expectation is smaller than 1. Hence, by the maximal inequality (5.15),
1 2
P sup jh
; Xt ij x e xC 2 k
0tT
and taking the minimum of 7! x C 12 2 k for 2 RC we obtain (8.41).

Moreover, the inequality (8.41) applied to the coordinate vectors gives, for every
i D 1; : : : ; m,
x x2
P sup jXi .t/j p 2e 2k m
0tT m
and now just observe that if sup0tT jXt j x, then for one at least of the coordinates
i; i D 1; : : : ; m, necessarily sup0tT jXi .t/j xm1=2 . Therefore
X m x x2
P sup jXt j x P sup jXi .t/j p 2me 2k m :
0tT iD1 0tT m
t
u
2
Theorem 8.4 Let X D .X1 ; : : : ; Xd / 2 Mloc .Œ0; C1Œ/ such that
Z C1
jXs j2 ds D C1 a.s.
0
Rt
Let At D 0 jXs j2 ds and .t/ D inffsI As > tg; then the process
Z .t/
Wt D Xs dBs
0
Rt
is a (one-dimensional) Brownian motion such that W.At / D 0 Xs dBs :
Proof Identical to the proof of Theorem 8.2. t

u
In particular, if jXs j D 1 a.s. for every s, then At D t and therefore
2
Corollary 8.2 Let X D .X1 ; : : : ; Xd / 2 Mloc .Œ0; C1Œ/ be a process such that
jXs j D 1 for almost every s a.s. Then the process
Z t
Wt D Xs dBs
0
is a Brownian motion.
Sometimes it would be useful to apply Corollary 8.2 to Xt D jB1t j .B1 .t/; : : : ; Bd .t//,
which is obviously such that jXt j D 1. Unfortunately this is not immediately
possible, as X is not defined if jBt j D 0. In the following example we see how

to go round this difficulty.
Example 8.9 (Bessel square processes again) We have seen in (8.37) that
the process Xt D jBt j2 has the stochastic differential
dXt D d dt C 2Bt dBt : (8.43)

p
Now let z0 2 Rd be any vector such that jz0 j D 1 and let (recall that Xt D
jBt j)
Z t
Bs
Wt D p 1fXs6D0g C z0 1fXs D0g dBs : (8.44)
0 Xs
Therefore W is a one-dimensional Brownian motion. Note that

p p
Xt dW.t/ D Bt 1fXt 6D0g dBt C Xt 1fXt D0g z0 dBt D Bt dBt
p
(of course Xt 1fXt D0g D 0) so that
X
d p
Bt dBt D Bi .t/ dBi .t/ D Xt dW.t/ :
iD1
(8.43) can now be written as

p
dXt D d dt C 2 Xt dWt : (8.45)
Therefore the process Xt D jBt j2 is a solution of (8.45), another example

of a Stochastic Differential Equation, which we shall investigate in the next
chapter.
The same arguments developed in the proofs of Theorem 8.4 and Corollary 8.2 give
the following
Proposition 8.8 Let .Os /s be a progressively measurable O.d/-valued (i.e.

orthogonal matrix-valued) process. Then .Os /s 2 M 2 .Œ0; C1Œ/ and
Z t
Xt D Os dBs
0
is itself an .Ft /t -d-dimensional Brownian motion.

Proof .Os /s obviously belongs to M 2 .Œ0; C1Œ/ as all the coefficients of an orthog-
onal matrix are smaller, in modulus, than 1. By the criterion of Theorem 5.17, we
1 2
just need to prove that, for every 2 Rd , Yt D eih ;Xt iC 2 j j t is an .Ft /t -martingale.
By Ito’s formula, or directly by (8.39) and (8.40), we have
Z
˝ t ˛
Yt D 1 C i ; Ys Os dBs :
0
1 2
Since, for s t, jYs j e 2 j j s , the process .Ys Os /s is in M 2 .Œ0; C1Œ/ and therefore
Y is an .Ft /t -martingale, which allows us to conclude the proof. t
u
We conclude this section with an extension to the multidimensional case of the
Lp estimates of Sect. 8.3.
2
Proposition 8.9 If 2 Mloc .Œ0; T/, p 2, then
ˇZ t ˇp Z t
ˇ ˇ p2
E sup ˇ .s/ dBs ˇ c. p; m; d/T 2 E j.s/jp ds :
0tT 0 0
Proof One can repeat the proof of Proposition 8.4, applying Theorem 8.3 to the
function, Rm ! R, u.x/ D jxjp , or using the inequalities (also useful later)
X
m
p2 X
m
jxi jp jxjp m 2 jxi jp x 2 Rm (8.46)
iD1 iD1
X
m
jy1 C C yd jp d p1 jyj jp y 2 Rd : (8.47)
jD1
By Proposition 8.4 and (8.46),
ˇZ t ˇp m ˇX
X d Z t ˇp
ˇ ˇ p2 ˇ ˇ
E sup ˇ .s/ dBs ˇ m 2 E sup ˇ ij .s/ dBj .s/ˇ
0tT 0 0tT iD1 jD1 0
X
m ˇXd Z t ˇp
p2 ˇ ˇ
m 2 E sup ˇ ij .s/ dBj .s/ˇ
iD1 0tT jD1 0

8.5 A case study: recurrence of multidimensional Brownian motion 243
and then by (8.47) and Proposition 8.4
X
m X
d ˇZ t ˇp
p2 ˇ ˇ
m 2 dp1 E sup ˇ ij .s/ dBj .s/ˇ
iD1 jD1 0tT 0
p2 p2
Z T X
m X
d
p3p=2 .2. p 1//p=2 m 2 dp1 T 2 E jij .s/jp ds
0 iD1 jD1
p2
Z T
c. p; m; d/T 2 E j.s/jp ds :
0
t
u

8.5 A case study: recurrence of multidimensional Brownian
motion
Let B be an m-dimensional Brownian motion, with m 2. We know (see

Remark 3.4) that in dimension 1 the Brownian motion visits (infinitely many times)
every real number with probability 1. What can be said in dimension m 2? In
dimension m does Brownian motion also go everywhere?
There are two ways of precisely stating this kind of property for a generic
process X.
p1) X visits every open set D Rm with probability 1.
p2) X visits every point x 2 Rm with probability 1.
A process satisfying p1) is said to be recurrent. A process that is not recurrent is
said to be transient. Of course p2) entails p1).
Let us investigate whether the Brownian motion is transient or recurrent in
dimension 2. We shall treat the case m 3, leaving the case m D 2 as an
exercise (Exercise 8.23). Let x 2 Rm , x 6D 0: we want to investigate the probability
for the Brownian motion to visit a neighborhood of x. It should not be surprising,
as in other questions concerning the visit of a process to some set, that the problem
boils down to finding the right martingale. Let
1
Xt D (8.48)
jBt xjm2
We can write Xt D f .Bt /, where f .z/ D jz xj.m2/ . f has a singularity at x but

is bounded and infinitely many times differentiable outside every ball containing x.
In order to apply Ito’s formula, which in this case is given in (8.35), let us compute
f 0 .z/ and 4f .z/, z 6D x. If g.z/ D jz xj, we have, for z 6D x,
v
uX
@ u
m
@g t .zj xj /2 D zi xi
.z/ D
@zi @zi jD1 jz xj
and therefore
@f .m 2/ zi xi .m 2/.zi xi /

.z/ D D
@zi jz xjm1 jz xj jz xjm
and
@2 f m.m 2/.zi xi /2 m2

2
.z/ D
@zi jz xj mC2 jz xjm
so that
m.m 2/ X
m
m.m 2/
4f .z/ D .zi xi /2 D0:
jz xj mC2
iD1
jz xjm
f is not a C2 function on Rm (it is not even defined everywhere) so that we cannot

directly apply Ito’s formula in order to obtain the stochastic differential of Xt D
f .Bt /. However, let e
f be a function that coincides with f outside a small ball, Vn .x/,
centered at x and with radius 1n with 1n < jxj and extended inside the ball in order to
be C2 . Then by Ito’s formula (recall that ef .0/ D f .0/ D jxj.m2/ )
Z Z
e
t
e 0
t
1 e
f .Bt / D jxj.m2/ C f .Bs / dBs C 4f .Bs / ds :
0 0 2
Let us denote by n the entrance time of B into the small ball. Then, as f coincides
with e
f outside the ball,
Xt^n D f .Bt^n / D ef .Bt^n /

Z t^n Z t^n
.m2/ 0 1 (8.49)
D jxj C f .Bs / dBs C 4f .Bs / ds :
0 2
„0 ƒ‚ …
D0
Therefore .Xt^n /t is a local martingale. As f is bounded outside Vn .x/ . f nm2

outside the ball Vn .x/), .Xt^n /t is also bounded and is therefore a martingale
(a bounded local martingale is a martingale, Remark 7.7). Let us deduce that
P.n < C1/ < 1 if n is large enough. As X0 D jxj.m2/ ,
E.Xt^n / D E.X0 / D jxj.m2/
for every t 0. As Xt^n D nm2 on fn tg, we have E.Xt^n / nm2 P.n t/

and therefore
jxj.m2/ nm2 P.n t/ (8.50)

8.5 A case study: recurrence of multidimensional Brownian motion 245
so that P.n t/ .njxj/.m2/ and as t ! C1 we have P.n < C1/

.njxj/.m2/ and for n large the last quantity is < 1. Therefore with strictly positive
probability the Brownian motion does not visit a small ball around x. In dimension
3 the Brownian motion is therefore transient.
From this estimate we deduce that, with probability 1, B does not visit x: in order
to visit x, B must first enter into the ball of radius 1n centered at x so that for every n
1
P.Bt D x for some t > 0/ P.n < C1/
.njxj/.m2/
and, as n is arbitrary,
P.Bt D x for some t > 0/ D 0 : (8.51)
If m D 2, conversely, the Brownian motion is recurrent (see Exercise 8.23),

but (8.51) is still true: even in dimension 2 the probability for the Brownian motion
to visit a given point x 6D 0 is equal to 0.
The argument proving recurrence in dimension 2 is similar to the one developed
here for dimension 3 (find the right martingale. . . ), but using the function f2 .z/ D
log.jz xj/ instead. What makes the difference with the function f of (8.48) is
that f2 is not bounded outside a neighborhood of x.
Example 8.10 (A uniformly integrable local martingale which is not a mar-

tingale) Since we now know that, with probability 1, B does not visit x, the
process
1
Xt D
jBt xjm2
is well defined and, thanks to (8.49), is a local martingale. We now prove

that it is uniformly integrable but not a martingale, which is a useful
counterexample. In order to prove uniform integrability, let us show that
supt>0 E.Xt / < C1 for every p < m.m 2/1 and therefore that it is a
p
process bounded in Lp for some p > 1. Let r D p.m 2/; then

Z
1 2 =2t
E.Xt / D E.jBt xjr / D jy xjr ejyj
p
dy :
.2t/m=2
We split the integral into two parts: the first one on a ball Sx centered at x and
with radius 12 jxj, the second one on its complement. As jy xjr jxjr 2r
(continued )

outside of Sx ,
Z Z
1 r jyj2 =2t 2r jyj2 =2t 2r
jy xj e dy e dy D
.2t/m=2 Scx jxjr .2t/m=2 jxjr
The remaining term conversely is evaluated by observing that inside Sx we

2 2
have jyj 12 jxj and therefore ejyj =2t ejxj =4t . The integral over Sx is
therefore smaller than
Z Z
1 jxj2 =4t r 1 jxj2 =4t
e jy xj dy D e jzjr dy
.2t/m=2 jyxj 12 jxj .2t/m=2 jzj 12 jxj
Z 1
1 2 2 jxj
D const ejxj =4t m1r d:
.2t/ m=2
0
If p < m.m 2/1 , we have r < m and the last integral is convergent. The
p
quantity E.Xt / is bounded in t and X is uniformly integrable.
If X, which is uniformly integrable, were a martingale, then the limit X1 D
limt!C1 Xt would exist a.s. and in L1 . But by the Iterated Logarithm Law
there exists a.s. a sequence of times .tn /n with limn!1 tn D C1 such that
limn!1 jBtn j D C1, therefore X1 D 0 a.s., in contradiction with the fact
that the convergence must also take place in L1 .
Exercises
8.1 (p. 533) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a real Brownian motion and let
1
Mt D .Bt C t/ e.Bt C 2 t/ :
a) Compute the stochastic differential of .Mt /t .

b) Show that .Mt /t is a martingale.
a; b > 0. Compute the stochastic differential of Xt D .b Bt /.a C Bt / C t and
deduce that it is a martingale.
8.3 (p. 535)
a) Let .Bt /t be a Brownian motion. Determine which of the following processes is
a martingale
Xt D et=2 sin Bt
Yt D et=2 cos Bt :
Compute E.Xt / and E.Yt / for t D 1.

b) Prove that X and Y are Ito’s processes and compute hX; Yit .
8.4 (p. 536)

a) Let > 0. Prove that
Z t
1 2 1 Bt 1 2 t
e Bs 2 s dBs D e 2 1 : (8.52)
0
b)
Z t Rs Rs
Bu dBu 12 B2u du
e 0 0 Bs dBs D‹
0
8.5 (p. 536) Let B be a Brownian motion with respect to the filtration .Ft /t .
a) Show that Xt D B3t 3tBt is an .Ft /t -martingale.
b) Prove that for every n D 1; 2; : : : there exist numbers cn;m ; m Œn=2 such that
the polynomial
Œn=2
X
Pn .x; t/ D cn;m xn2m tm (8.53)
mD0
is such that Xt D Pn .Bt ; t/ is an .Ft /t -martingale. Compute Pn for n D 1; 2; 3; 4.

c) Let a; b > 0 and let be the exit time of B from a; bŒ. In Example 5.3
and Exercise 5.10 we have computed the law of B and E./. Are and B
independent?
8.6 (p. 537) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and 2 R.
a) Prove that
Z t
Mt D e t Bt e u Bu du
0
is a square integrable martingale and compute its associated increasing process.

b) Prove that if < 0 then M is uniformly integrable and determine the
distribution of its limit as t ! C1.
c1) Prove that
1 2 t 1/
Zt D eMt 4 .e
is a martingale.
c2) Let Z1 be the limit of the positive martingale Z as t ! C1 and assume < 0.
What is the law of Z1 ?
p
c3) Assume > 0. Compute EŒZt for p < 1. What is the law of Z1 ?
Z t
Yt D tBt Bu du :
0
In Exercise 5.13 and in Example 8.6 we proved that Y is an .Ft /t -martingale.

Prove that
1 3
Zt D eYt 6 t
8.8 (p. 539) Let B be a Brownian motion and
Z t
dBs
Xt D p
0 2 C sin s
What is the value of

Xt
lim ‹
t!C1 .2t log log t/1=2
For < t < a primitive of .2 C sin t/1 is t 7! 2 31=2 arctan.2 31=2 tan. 2t / C 31=2 /
hence
Z
ds 2
D p
2 C sin s 3
8.9 (p. 540) Let, for > 0,

Z t
Xt D e s Bs ds :
0
Does the limit
lim Xt (8.54)
t!C1
exist? In what sense? Determine the limit and compute its distribution.
Integrate by parts. . .
8.10 (p. 540) Let B be a real Brownian motion and, for " > 0 and t T,
p Z t s
Xt" D 2 sin dBs :
0 "
a) Prove that, for every 0 < t T, Xt" converges in law as " ! 0 to a limiting
distribution to be determined.
b) Prove that, as " ! 0, the law of the continuous process X " for t 2 Œ0; T converges
to the Wiener measure.

Z t
dBs
Xt D p
0 1Cs

Compute P sup0t3 Xt 1 .
8.12 (p. 541) Let, for ˛ > 0,
Z
p t
Xt D ˛ C 1 u˛=2 dBu :
0
a) Compute P.sup0s2 Xs 1/.

b) Let D infft > 0I Xt 1g. Compute the density of . For which values of ˛
does have finite expectation?
8.13 (p. 542) As the Brownian paths are continuous, and therefore locally
bounded, we can consider their derivative in the sense of distributions, i.e. for every
2 CK1 .RC / and for every ! 2 ˝ we can consider the distribution B0 .!/ defined as
Z C1
h; B0 .!/i D t0 Bt .!/ dt :
0
Check that
Z C1
h; B0 i D t dBt a.s.
0
8.14 (p. 542) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and let, for
0 t < 1,
1 B2t
Zt D p exp :
1t 2.1 t/
a) Prove that Z is a continuous .Ft /t -martingale.

Compute, for p > 0, EŒZt . Is Z square integrable? Is Z bounded in L2 ?
p
b)
c) What is the value of limt!1 Zt ?
2
d) Find a process .Xt /t 2 Mloc such that
Rt Rt
Xs dBs 12 Xs2 ds
Zt D e 0 0 :
8.15 (p. 545) Let B D .B1 ; B2 / be a two-dimensional Brownian motion. Which

one of the following is a Brownian motion?
Z t Z t
a) W1 .t/ D sin s dB1 .s/ C cos s dB1 .s/.
Z0 t Z0 t
b) W2 .t/ D sin s dB1 .s/ C cos s dB2 .s/.
Z0 t 0 Z
t
c) W3 .t/ D sin B2 .s/ dB1 .s/ C cos B2 .s/ dB2 .s/.
0 0
8.16 (p. 545) Let .B1 ; B2 ; B3 / be a three-dimensional Brownian motion and

Z t Z t
Xt D sin.B3 .s// dB1 .s/ C cos.B3 .s// dB2 .s/
Z0 t Z0 t (8.55)
Yt D cos.B3 .s// dB1 .s/ C sin.B3 .s// dB2 .s/ :
0 0
a) Prove that .Xt /t and .Yt /t are Brownian motions.

b1) Prove that, for every s; t, Xs and Yt are uncorrelated.
b2) Is .Xt ; Yt /t a two-dimensional Brownian motion?
c) Assume instead that
Z t Z t
Yt D cos.B3 .s// dB1 .s/ C sin.B3 .s// dB2 .s/ :
0 0
Is .Xt ; Yt /t now a two-dimensional Brownian motion?
8.17 (p. 546) Let B D .˝; F ; .Gt /t ; .Bt /t ; P/ be a Brownian motion with respect
et D Gt _ .B1 /.
to its natural filtration .Gt /t and let G
a) Let 0 < s < t < 1 be fixed. Determine a square integrable function ˚ and a
number ˛ (possibly depending on s; t) such that the r.v.
Z s
Bt ˚.u/ dBu ˛B1
0
is orthogonal to B1 and to Bv for every v s.

b) Compute EŒBt j G es . Is B also a Brownian motion with respect to the filtration
e e t /t ?
.G t /t ? Is B a martingale with respect to .G
c) Let, for t 1,
Z
e
t
B1 Bu
Bt D Bt du :
0 1u
c1) Prove that, for 0 < s < t < 1, e Bt eBs is independent of G es .

e e
c2) Prove that, for 0 < t < 1, .Bt /t is a .G t /t -Brownian motion.
c3) Prove that .Bt /t is a Ito process with respect to .Ge t /t .
• This exercise shows that when switching to a larger filtration, in general,

properties such as being a Brownian motion or a martingale are lost, whereas
to be a Ito process is more stable.
8.18 (p. 548) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion and X a
centered square integrable FT -measurable r.v.
a) Let us assume that there exists a process .Yt /t in M 2 .Œ0; T/ such that
Z T
XD Ys dBs : (8.56)
0
Show that such a process is unique.

b1) Let X D B3T . What is the value of Xs D EŒX jFs for 0 s T? What is the
stochastic differential of .Xs /0sT ?
b2) Determine a process .Yt /t such that (8.56) holds in this case.
2
c) Determine a process .Yt /t such that (8.56) holds for X D e BT e 2 T
, for
6D 0.

Z T
ZD Bs ds :
0
a1) Compute Mt D EŒZ jFt .

a2) Compute the stochastic differential of .Mt /t and determine a process X 2 M 2
such that
Z T
ZD Xs dBs :
0
b) Determine a process X 2 M 2 such that

Z Z
T
T2 T
B2s ds D C Xs dBs :
0 2 0
8.20 (p. 550) In Exercise 5.24 we proved that, if B is an m-dimensional Brownian

motion, then Mt D Bi .t/Bj .t/, 1 i; j m, i 6D j, is a square integrable martingale
and in Example 8.7 we have shown that
Z t
hMit D .B2i .s/ C B2j .s// ds :
0
Let
> 0. Determine an increasing process .At /t such that
Zt D e
Bi .t/Bj .t/At
is a local martingale.
8.21 (p. 551)
a) Let Z be an N.0; 1/-distributed r.v. Show that
(
1
˛Z 2 .1 2˛/3=2 if ˛ <
EŒZ 2 e D 2
1
C1 if ˛ 2

b) Let .Bt /t be a Brownian motion and let
Bt B2t
Ht D exp :
.1 t/3=2 2.1 t/
2
Prove that limt!1 Ht D 0 a.s. Prove that H 2 Mloc .Œ0; 1/ but H 62 M 2 .Œ0; 1/.
c) Let, for t < 1,
1 B2t
Xt D p exp :
1t 2.1 t/
Prove that, for t < 1,

Z t
Xt D 1 Hs dBs : (8.57)
0
Prove that limt!1 Xt D 1 a.s.

d) Prove that
Z 1
Hs dBs D 1
0
so that H is an example of an integrand not belonging to M 2 whose stochastic

integral is square integrable.
8.22 (p. 552) (Poincaré’s lemma) Let B1 ; B2 ; : : : be a sequence of one-dimensional

.n/
independent Brownian motions defined on the same probability space and let Bt D
.n/
.B1 .t/; : : : ; Bn .t//; B is therefore an n-dimensional
p Brownian motion. Let n be the
exit time of B.n/ from the ball of radius n.
.n/
a) Show that Mn .t/ D jBt j2 nt is a martingale. Deduce that E.n / D 1 for every n.
.n/
b) If Rn .t/ D 1n jBt j2 , show that there exists a Brownian motion W such that
2 p
dRn .t/ D dt C p Rn .t/ dWt : (8.58)
n
c) Deduce that
h i 16
E sup jRn .t/ tj2
0tn n
and therefore n ! 1 in probability as n ! 1.

d) (Gallardo 1981). Deduce the following classical lemma of H. Poincaré. Let n
p .n 1/-dimensional Lebesgue measure of the spherical
denote the normalized
surface of radius n of Rn . Let d be a fixed positive integer and n;d the
projection Rn ! Rd .
Then as n ! 1 the image of n through n;d converges to the N.0; I/ law of Rd .
c) Zt D Rn .t/ t, suitably stopped, is a martingale.
.n/
d) n is the law of Bn ; n;d .n / is therefore the law of .B1 .n /; : : : ; Bd .n //.
8.23 (p. 553) In this exercise we investigate the properties of recurrence/transience

of Brownian motion in dimension 2, hence completing the analysis of Sect. 8.5.
It turns out that in dimension 2 Brownian motion visits every open set with
probability 1 (which is different from the case of dimension m 3), and is
therefore recurrent, but in contrast to dimension 1, every point x 6D 0 is visited
with probability 0.
Let B be a two-dimensional Brownian motion and x 2 R2 ; x 6D 0. Let n be the
entrance time of B in the ball of radius 1n centered at x, with 1n < jxj, so that this ball
does not contain the origin. Let Xt D log jBt xj, when defined.
a) Prove that .Xt^n /t is a martingale.
b) Let, for 1n < M, n;M be the exit time of B from the annulus fzI 1n < jzxj Mg
with M > jxj so that this annulus contains the origin (Fig. 8.1). Prove that
n;M < C1 a.s. and compute P.jBn;M xj D 1n /.
c) Deduce that P.n < C1/ D 1 for every n and therefore that B is recurrent.
d) Let x 2 R2 , x 6D 0, and let us prove that B does not visit x with probability 1.
Let k > 0 be such that k1k < jxj < k. Let
1
k D kk D infftI jBt xj D kk
g
k D infftI jBt xj D kg :
d1) Deduce from the computation of b) the value of P.k < k /. Compute the limit
of this quantity as k ! 1.
d2) Let D infftI Bt D xg. Note that, for every k, P. < k / P.k < k / and
deduce that P. < C1/ D 0.
Fig. 8.1 n;M is the exit time of the Brownian motion from the shaded annulus. The large circle
has radius M, the small one 1n
8.24 (p. 554) (Bessel processes) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimen-
sional standard Brownian motion with m > 1 and let x 2 Rm with jxj > 0. Then,
if Xt D jBt C xj, X is a Markov process by Exercise 6.11 and Xt > 0 a.s. for every
t 0 a.s. as seen in Sect. 8.5 and Exercise 8.23.
a) Show that Xt D jBt C xj is an Ito process and determine its stochastic differential.
b) Show that if f 2 CK2 .0; C1Œ/ and D jxj, then
1
EŒ f .Xt / f ./ ! Lf ./ ;
t t!C1
where L is a differential operator to be determined.

Chapter 9
Stochastic Differential Equations
In this chapter we introduce the notion of a Stochastic Differential Equation. In

Sects. 9.4, 9.5, 9.6 we investigate existence and uniqueness. In Sect. 9.8 we obtain
some Lp estimates that will allow us to specify the regularity of the paths and
the dependence from the initial conditions. In the last sections we shall see that
the solution of a stochastic differential equation is a Markov process and even a
diffusion associated to a differential operator that we shall specify.
9.1 Definitions
Let b.x; t/ D .bi .x; t//1im and .x; t/ D .ij .x; t// 1im be measurable functions
1jd
defined on Rm Œ0; T and Rm - and M.m; d/-valued respectively (recall that
M.m; d/ D m d real matrices).
Definition 9.1 The process .˝; F ; .Ft /t2Œ0;T ; .t /t2Œu;T ; .Bt /t ; P/ is said to
be a solution of the Stochastic Differential Equation (SDE)
dt D b.t ; t/ dt C .t ; t/ dBt

(9.1)
u D x x 2 Rm
if
a) .˝; F ; .Ft /t ; .Bt /t ; P/ is a d-dimensional standard Brownian motion, and
b) for every t 2 Œu; T we have
Z t Z t
t D x C b.s ; s/ ds C .s ; s/ dBs :
u u

256 9 Stochastic Differential Equations
Of course in Definition 9.1 we require implicitly that s 7! b.s ; s/ and s 7!

1 2
.s ; s/ are processes in Mloc .Œu; T/ and Mloc .Œu; T/, respectively.
Note that the Brownian motion with respect to which the stochastic integrals are
taken and the probability space on which it is defined are not given a priori.
Note also that the solution of a SDE as above is necessarily a continuous process.
In case someone wanted to model some real life phenomenon with such a SDE it is
important to realize this point: (9.1) is not fit to model quantities that make jumps.
Discontinuous behaviors must be modeled using different SDEs (with the Brownian
motion replaced by some more suitable stochastic process).
Some terminology: is the diffusion coefficient, b is the drift.
Definition 9.2 We say that (9.1) has strong solutions if for every standard
Brownian motion .˝; F ; .Ft /t ; .Bt /t ; P/ there exists a process that satisfies
Eq. (9.1).
We shall speak of weak solutions, meaning those in the sense of Definition 9.1.
If is a solution, strong or weak, we can consider the law of the process (see
Sect. 3.2): recall that the map W ˝ ! C .Œ0; T; Rm / defined as
! 7! .t 7! t .!//
is measurable (Proposition 3.3) and the law of the process is the probability on
C .Œ0; T; Rm / that is the image of P through the map .
Definition 9.3 We say that for the SDE (9.1) there is uniqueness in law if,
given two solutions i D .˝ i ; F i ; .Fti /t ; .ti /t ; .Bit /t ; Pi /, i D 1; 2, (possibly
defined on different probability spaces and/or with respect to different Brow-
nian motions) 1 and 2 have the same law.
Definition 9.4 We say that for the SDE (9.1) there is pathwise uniqueness
if, given two solutions .˝; F ; .Ft /t ; .ti /t ; .Bt /t ; P/; i D 1; 2, defined on the
same probability space and with respect to the same Brownian motion, 1 and
2 are indistinguishable, i.e. P.t1 D t2 for every t 2 Œu; T/ D 1.
Note that, whereas the existence of strong solutions immediately implies the
existence of weak solutions, it is less obvious that pathwise uniqueness implies
uniqueness in law, since for the latter we must compare solutions that are defined
9.2 Examples 257
on different probability spaces and with respect to different Brownian motions. It is,
however, possible to prove that pathwise uniqueness implies uniqueness in law.
Remark 9.1 If is a solution of the SDE (9.1) and f W Rm RC ! R is a

C2;1 function, then the stochastic differential of t 7! f .t ; t/ is
@f X @f 1 X @2 f
df .t ; t/ D .t ; t/ dt C .t ; t/ di .t/ C .t ; t/aij .t ; t/ dt ;
@t i
@xi 2 i;j @xi xj
P
where aij D ` i` j` , i.e. a D . Hence we can write
@f X @f
df .t ; t/ D .t ; t/ C Lf .t ; t/ dt C .t ; t/ij .t ; t/ dBj .t/ ;
@t i;j
@xi
where L denotes the second-order differential operator
1X X
m m
@2 @
LD aij .x; t/ C bi .x; t/
2 i;jD1 @xi xj iD1
@xi
In particular, if we look for a function f such that t 7! f .t ; t/ is a martingale,

a first requirement is that
@f
C Lf D 0 ;
@t
which ensures that t 7! f .t ; t/ is a local martingale.

The operator L will play an important role from now on. Note that if dt D
dBt , then L D 12 4, 4 denoting the Laplace operator.
In this chapter we shall generally denote stochastic processes by the symbols , .t /t .
The notations X, .Xt /t , most common in the previous chapters, are now reserved to
denote the canonical process. From now on B D .˝; F ; .Ft /t ; .Bt /t ; P/ will denote
a d-dimensional standard Brownian motion, fixed once and for all.
9.2 Examples
In the next sections we shall investigate the existence and the uniqueness of solutions
of a SDE, but let us first see some particular equations for which it is possible to find
an explicit solution.
Example 9.1 Let us consider the equation, in dimension 1,
dt D t dt C dBt
(9.2)
0 D x
where ; 2 R; i.e. we assume that the drift is linear in and the diffusion
coefficient constant. To solve (9.2) we can use a method which is similar
to the variation of constants for ordinary differential equations: here the
stochastic part of the equation, dBt , plays the role of the constant term. The
“homogeneous” equation would be
dt D t dt
0 D x
whose solution is t D e t x. Let us look for a “particular solution” of (9.2)

of the form t D e t Zt . As, by Ito’s formula,
d.e t Zt / D e t Zt dt C e t dZt
t D e t Zt is a solution if e t dZt D dBt , i.e.

Z t
Zt D e s dBs
0
and, in conclusion,
Z t
t t
t D e xCe e s dBs : (9.3)
0
This is the Ornstein–Uhlenbeck process (in the literature this name is

sometimes reserved for the case > 0). It is a process of particular interest,
being the natural model in many applications, and we shall see in the exercises
some of its important properties (see Exercises 9.1, 9.2, 9.24 and 9.25). Now
let us observe that the computation above can be repeated if we consider an
m-dimensional analog of (9.2), i.e. the equation
dt D t dt C dBt
(9.4)
0 D x
where and are m m and m d matrices, respectively. By a step by step

repetition of the computation above, we have that a solution is
(continued )
9.2 Examples 259

Z t
t D e t x C e t e s dBs ; (9.5)
0
e t denoting the exponential of the matrix t. This is a Gaussian process by

Proposition 7.1, as the stochastic integral above has a deterministic integrand.
Example 9.2 Let us now consider instead the equation in dimension 1
dt D bt dt C t dBt

(9.6)
0 D x :
Now both drift and diffusion coefficient are linear functions of . Dividing
both sides by t we have
dt
D b dt C dBt : (9.7)
t
The term dt t is suggestive of the stochastic differential of log t . It is not quite
this way, as we know that stochastic differentials behave differently from
the usual ones. Anyway, if we compute the stochastic differential of log t ,
assuming that is a solution of (9.6), Ito’s formula gives
dt 1 2
d.log t / D 2 2 t2 dt D b dt C dBt :
t 2t 2
2

Therefore log t D log x C b 2 t C Bt and
2
t D xe.b 2 /tC Bt : (9.8)
In fact this derivation of the solution is not correct: we cannot apply Ito’s
formula to the logarithm function, which is not twice differentiable on R (it
is not even defined on the whole of R). But once the solution (9.8) is derived,
it is easy, always by Ito’s formula, to check that it is actually a solution of
Eq. (9.6). Note that, if x > 0, then the solution remains positive for every
t 0.
This process is geometric Brownian motion and it is one of the processes
to be taken into account as a model for the evolution of quantities that must
(continued )

always stay positive (it is not the only one enjoying this property). Because of
this, it is also used to describe the evolution of prices in financial markets.
Note that, for every t, t appears to be the exponential of a Gaussian r.v. It
has therefore a lognormal distribution (see Exercise 1.11).
This section provides examples of SDE’s for which it is possible to obtain an explicit
solution. This is not a common situation. However developing the arguments of
these two examples it is possible to find a rather explicit solution of a SDE when the
drift and diffusion coefficient are both linear-affine functions. Complete details are
given, in the one-dimensional case, in Exercise 9.11. Other examples of SDE’s for
which an explicit solution can be obtained are developed in Exercises 9.6 and 9.13.
Note that in the two previous examples we have found a solution but we still
know nothing about uniqueness.
9.3 An a priori estimate
In this section we prove some properties of the solution of an SDE before looking
into the question of existence and uniqueness.
Assumption (A) We say that the coefficients b and satisfy Assumption (A)
if they are measurable in .x; t/ and if there exist constants L > 0; M > 0 such
that for every x; y 2 Rm ; t 2 Œ0; T,
jb.x; t/j M.1 C jxj/ j.x; t/j M.1 C jxj/ (9.9)

jb.x; t/ b.y; t/j Ljx yj j.x; t/ .y; t/j Ljx yj : (9.10)
Note that (9.10) requires the coefficients b and to be Lipschitz continuous in x

uniformly in t, whereas (9.9) requires that they have a sublinear growth at infinity.
Note also that if b and do not depend on t then (9.10) implies (9.9). However, it
is useful to keep these two conditions separate in order to clarify their role in what
follows. In the remainder we shall often make use of the inequality, for x1 ; : : : ; xm 2
Rd ,
jx1 C C xm jp mp1 .jx1 jp C C jxm jp / : (9.11)
A key tool in our investigation of SDE’s is the following classical inequality.

9.3 An a priori estimate 261
Lemma 9.1 (Gronwall’s inequality) Let w; v be non-negative real func-

tions on the interval Œa; b with w integrable and v bounded measurable; let
c 0 and let us assume that
Z t
v.t/ c C w.s/v.s/ ds :
a
Then
Rt
v.t/ c e a w.s/ ds
:
Note in particular in the previous statement that if c D 0 then necessarily v 0.

Gronwall’s inequality is used with ordinary differential equations in order to
prove bounds and uniqueness of the solutions. For SDE’s its use is similar.
In particular, Gronwall’s inequality allows to prove the following Theorem 9.1:
it states that if there exists a solution (weak or strong) of the SDE (9.1) with an
initial value that is in Lp with p 2 and independent of the Brownian motion,
then under the condition of sublinear growth (9.10) the solution is necessarily a
process in M p . Note that sublinear growth of the coefficients is required, but no
assumption concerning Lipschitz continuity is made. And still we do not know
whether a solution actually exists.
Theorem 9.1 Let be a solution of

Z t Z t
t D C b.r ; r/ dr C .r ; r/ dBr
u u
where the coefficients b; are measurable in x; t and satisfy (9.9) (sublinear

growth) and is an Fu -measurable r.v. of Lp . Then for every p 2

E sup js jp c. p; T; M/.1 C E.jjp // (9.12)
usT

E sup js jp c. p; T; M/.t u/p=2 .1 C E.jjp // : (9.13)
ust
Proof The idea is to find some inequalities in order to show that the function v.t/ D
EŒsupust js jp satisfies an inequality of the kind
Z t
v.t/ c1 .1 C EŒjjp / C c2 v.s/ ds
u
and then to apply Gronwall’s inequality. There is, however, a difficulty, as in order
to apply Gronwall’s inequality we must know beforehand that such a function v is
bounded. Otherwise it might be v C1. In order to circumvent this difficulty we
shall be obliged stop the process when it takes large values as described below.
Let, for R > 0, R .t/ D t^R where R D infftI u t T; jt j Rg denotes
the exit time of from the open ball of radius R, with the understanding R D T if
jt j < R for every t 2 Œu; T. Then
Z t^R Z t^R
R .t/ D C b.r ; r/ dr C .r ; r/ dBr
Z t u Z tu
DC b.r ; r/1fr<R g dr C .r ; r/1fr<R g dBr (9.14)
Z t u Z ut
DC b.R .r/; r/1fr<R g dr C .R .r/; r/1fr<R g dBr :
u u
th
Taking the modulus, the p power and then the expectation we find, using (9.11),
h i
E sup jR .s/jp
ust
h ˇZ s ˇp i
ˇ ˇ
3 EŒjj C 3 E sup ˇ
p1 p p1
b.R .r/; r/1fr<R g drˇ (9.15)
ˇZ sust u
ˇp
ˇ ˇ
C3p1 E sup ˇ .R .r/; r/1fr<R g dBr ˇ :
ust u
Now, by Hölder’s inequality and (9.9),

h ˇZ s ˇp i hZ t i
ˇ ˇ
E sup ˇ b.R .r/; r/1frR g drˇ T E
p1
jb.R .r/; r/jp 1fr<R g dr
ust u u
hZ t i
T p1 M p E .1 C jR .r/j/p dr :
u
By the Lp inequalities for stochastic integrals (Proposition 8.4) we have

h ˇZ s ˇp i p2
Z t
ˇ ˇ
E sup ˇ .R .r/; r/1fr<R g dBr ˇ cp jt uj 2 E j.R .r/; r/jp dr
ust u u
p2 Z t
cp T 2 Mp E .1 C jR .r/j/p dr :
u
Hence, plugging these two inequalities into (9.15), we find

h i
E sup jR .s/jp
ust
p2
Z
t
3 p1
EŒjj C 3
p p1
M T p1 C cp T 2 E
p
j1 C R .r/jp dr
u
9.3 An a priori estimate 263
Z
p2 t
3 p1
EŒjj C 3
p p1
M T p1 C cp T 2 2p1 E T p C
p
jR .r/jp dr
u
Z
t
c1 . p; T; M/ 1 C Ejj p
C c2 . p; T; M/ E jR .r/jp dr :
u

Let now v.t/ D E supust jR .s/j : from the previous inequality we have
p
Z t
v.t/ c1 . p; T; M/.1 C Ejj / C c2 . p; T; M/p
v.r/ dr :
u
Now jR .t/j D jj if jj R and jR .t/j R otherwise. Hence jR .t/j R _ jj
and v.t/ EŒRp _ jjp < C1. v is therefore bounded and thanks to Gronwall’s
inequality
h i
v.T/ D E sup jR .s/jp c1 . p; T; M/ 1 C EŒjjp eTc2 . p;T;M/
usT

D c. p; T; M/ 1 C EŒjjp :
Note that the right-hand side above does not depend on R. Let us now send R !
1. The first thing to observe is that R !R!1 D T: as is continuous we have
suputR jt jp D Rp on fR < Tg and therefore
h i
E sup jt jp Rp P.R < T/
utR
so that
1 h i k.T; M/.1 C E.jjp //

P.R < T/ E sup jt j p

Rp utR Rp
and therefore P.R < T/ ! 0 as R ! C1. As R ! R is increasing,

limR!C1 R D T a.s. and
sup jR .s/jp ! sup js jp a.s.

usT R!C1 usT
and by Fatou’s lemma (or Beppo Levi’s theorem) we have proved (9.12). As
for (9.13) we have
sup js jp

Z
ˇ s ˇp
ust
ˇZ s ˇp (9.16)
ˇ ˇ ˇ ˇ
2 p1
sup ˇ b.r ; r/ drˇ C 2p1 sup ˇ .r ; r/ dBr ˇ :
ust u ust u
Now, by Hölder’s inequality and (9.12),

h ˇZ s ˇp i hZ t i
ˇ ˇ
E sup ˇ b.r ; r/ drˇ .t u/p1 E jb.r ; r/jp dr
ust u u
hZ t i hZ t i
.t u/p1 M p E .1 C jr j/p dr 2p1 .t u/p1 M p E .1 C jr jp / dr
u u
c1 . p; T; m/.t u/ .1 C EŒjj / :
p p
Similarly, using again Proposition 8.4,

h Z s p i p2
hZ t i
E sup .r ; r/ dBr c.t u/ 2 E j.r ; r/jp dr
ust u u
p2
hZ t i p2
hZ t i
c.t u/ 2 Mp E .1 C jr j/p dr c2p1 .t u/ 2 Mp E .1 C jr jp / dr
u u
c2 . p; T; m/.t u/ p=2
.1 C EŒjj /p
and (9.13) follows from (9.16). t

u
Remark 9.2 Note again that we have proved (9.12) before we knew of the
existence of a solution: (9.12) is an a priori bound of the solutions.
As a first consequence, for p D 2, under assumption (9.9) (sublinearity
of the coefficients) if 2 L2 every solution of the SDE (9.1) is a process
belonging to M 2 . This implies also that the stochastic component
Z t
t 7! .s ; s/ dBs
0
is a square integrable martingale. Actually the process s 7! .s ; s/ itself

belongs to M 2 .Œu; t/: as j.s ; s/j M.1 CR js j/ we have j.s ; s/j2
2M 2 .1Cjs j2 /. Hence, as the stochastic integral u .s ; s/ dBs has expectation
t
equal to 0,
hZ t i
EŒt D EŒ C E b.s ; s/ ds : (9.17)
u
Therefore, intuitively, the coefficient b has the meaning of a trend, i.e. in the
average the process follows the direction of b. In dimension 1, in the average,
the process increases in regions where b is positive and decreases otherwise.
Conversely, determines zero-mean oscillations. If the process is one-
dimensional then, recalling Theorem 8.4, i.e. the fact that a stochastic integral
is a time-changed Brownian motion, regions where is large in absolute value
will be regions where the process undergoes oscillations with high intensity
(continued )
9.4 Existence for Lipschitz continuous coefficients 265

and regions where is small in absolute value will be regions where the
oscillations have a lower intensity.
This is useful information when dealing with the problem of modeling a
given phenomenon with an appropriate SDE.
Remark 9.3 It is useful to point out a by-product of the proof of Theorem 9.1.
If R is the exit time from the sphere of radius R, then, for every 2 Lp ,
k.T; M/.1 C E.jjp //
P.R < T/ D P sup jt j R
utT Rp
Hence limR!C1 P.R < T/ D 0 and this rate of convergence is moreover

uniform for in a bounded set of Lp and does not depend on the Lipschitz
constant of the coefficients.
9.4 Existence for Lipschitz continuous coefficients
In this section we prove existence and uniqueness of the solutions of an SDE under
suitable hypotheses on the coefficients b and .
Let us assume by now that (9.9) (sublinear growth) holds. Then, if Y 2
M 2 .Œ0; T/, as jb.x; s/j2 M 2 .1 C jxj/2 2M 2 .1 C jxj2 /,
hZ t i hZ t i
2 2
E jb.Ys ; s/j ds 2M E .1 C jYs j2 / ds
0
hZ t 0
i (9.18)
2
2M t C E jYs j2 ds
0
whereas, with a similar argument and using Doob’s maximal inequality (the second
of the maximal inequalities (7.23)),
h ˇZ u ˇ2 i hZ t i
ˇ ˇ
E sup ˇ .Ys ; s/ dBs ˇ 4E j.Ys ; s/j2 ds
0ut
hZ t
0 0
i (9.19)
8 M2 t C E jYs j2 ds :
0
We are now able to prove that for an SDE under Assumption (A) there exist strong
solutions and that pathwise uniqueness holds.
Theorem 9.2 Let u 0 and let be an Rm -valued r.v., Fu -measurable and

square integrable. Then under Assumption (A) there exists a 2 M 2 .Œu; T/
such that
Z t Z t
t D C b.s ; s/ ds C .s ; s/ dBs : (9.20)
u u
Moreover, we have pathwise uniqueness, i.e. if 0 is another solution of (9.20)

then
P.t D t0 for every t 2 Œu; T/ D 1 : (9.21)
The proof is very similar to that of comparable theorems for ordinary equations
(the method of successive approximations). Before going into it, let us point out that
to ask for to be Fu -measurable is equivalent to requiring that is independent of
.BtCu Bu ; t 0/, i.e., intuitively, the initial position is assumed to be independent
of the subsequent random evolution.
Actually, if is Fu -measurable, then it is necessarily independent of .BtCu
Bu ; t 0/, as developed in Exercise 3.4.
Conversely, assume u D 0. If is independent of .Bt ; t 0/, then, if Ft0 D
Ft _ ./, B is also an .Ft0 /t -standard Brownian motion (see Exercise 3.5) and now
is F00 -measurable.
Proof We shall assume u D 0 for simplicity. Let us define by recurrence a sequence
of processes by 0 .t/ and
Z t Z t
mC1 .t/ D C b.m .s/; s/ ds C .m .s/; s/ dBs : (9.22)
0 0
The idea of the proof of existence is to show that the processes .m /m converge
uniformly on the time interval Œ0; T to a process that will turn out to be a solution.
Let us first prove, by induction, that
h i .Rt/mC1
E sup jmC1 .u/ m .u/j2 , (9.23)
0ut .m C 1/Š
where R is a positive constant. For m D 0, thanks to Hölder’s inequality,

ˇZ u ˇ2 ˇZ u ˇ2
ˇ ˇ ˇ ˇ
sup j1 .u/ j2 2 sup ˇ b.; s/ dsˇ C 2 sup ˇ .; s/ dBs ˇ
0ut 0ut 0 0ut 0
Z t ˇZ u ˇ2
2 ˇ ˇ
2t jb.; s/j ds C 2 sup ˇ .; s/ dBs ˇ :
0 0ut 0
By (9.18) and (9.19),

E sup j1 .u/ j2 4M 2 t t C tEŒjj2 C 16M 2 t C tEŒjj2 Rt ;
0ut
where R D 16M 2 .T C 1/.1 C E.jj2 //. Let us assume (9.23) true for m 1, then
ˇZ u ˇ2
ˇ 2 ˇ
sup jmC1 .u/ m .u/j 2 sup ˇ b.m .s/; s/ b.m1 .s/; s/ dsˇ
0ut 0ut 0
ˇZ u ˇ2
ˇ ˇ
C 2 sup ˇ .m .s/; s/ .m1 .s/; s/ dBs ˇ :
0ut 0
Now, by Hölder’s inequality,

ˇZ u ˇ2
ˇ ˇ
sup ˇ b.m .s/; s/ b.m1 .s/; s/ dsˇ
0ut 0
Z u Z t
2 2
sup u jb.m .s/; s/ b.m1 .s/; s/j ds tL jm .s/ m1 .s/j2 ds
0ut 0 0
whereas by Doob’s inequality

ˇZ u ˇ2
ˇ ˇ
E sup ˇ .m .s/; s/ .m1 .s/; s/ dBs ˇ
0ut 0
Z t Z t
4E j.m .s/; s/ .m1 .s/; s/j2 ds 4L2 E.jm .s/ m1 .s/j2 / ds
0 0
whence, with a possibly larger value for R,

E sup jmC1 .u/ m .u/j2
0ut
Z t Z t
2 2 2
2L tE jm .s/ m1 .s/j ds C 8L E jm .s/ m1 .s/j2 ds
0 0
Z Z
t t
.Rs/m
R E.jm .s/ m1 .s/j2 / ds R ds
0 0 mŠ
.Rt/ mC1
D
.m C 1/Š
and (9.23) is proved. Markov’s inequality now gives
1 h i
P sup jmC1 .t/ m .t/j > m 22m E sup jmC1 .t/ m .t/j2
0tT 2 0tT
.RT/mC1
22m
.m C 1/Š
As the left-hand side is summable, by the Borel–Cantelli lemma,

1
P sup jmC1 .t/ m .t/j > m for infinitely many indices m D 0 ;
0tT 2
i.e. for almost every ! we have eventually
1
sup jmC1 .t/ m .t/j
0tT 2m
and therefore for fixed ! the series
X
m1
C ŒkC1 .t/ k .t/ D m .t/
kD0
converges uniformly on Œ0; T a.s. Let t D limm!1 m .t/. Then is continuous,

2
being the uniform limit of continuous processes, and therefore 2 Mloc .Œ0; T/. Let
us prove that it is a solution of (9.20). Recall that we have the relation
Z t Z t
mC1 .t/ D C b.m .s/; s/ ds C .m .s/; s/ dBs : (9.24)
0 0
Of course the left-hand side converges uniformly to . Moreover, as b and are

Lipschitz continuous (constant L),
sup jb.m .t/; t/ b.t ; t/j L sup jm .t/ t j :

0tT 0tT
This and the analogous inequality for imply that, uniformly on Œ0; T a.s.,
lim b.m .t/; t/ D b.t ; t/; lim .m .t/; t/ D .t ; t/ :

m!1 m!1
Therefore
Z t Z t
b.m .s/; s/ ds ! b.s ; s/ ds a.s.
0 m!1 0
and
Z T
j.m .t/; t/ .t ; t/j2 dt ! 0 a.s.
0 m!1
and by Theorem 7.3 (a.s. convergence implies convergence in probability)

Z t Z t
P
.m .s/; s/ dBs ! .s ; s/ dBs
0 m!1 0
so that we can take the limit in probability in (9.24) and obtain the relation
Z t Z t
t D C b.s ; s/ ds C .s ; s/ dBs ;
0 0
so that is a solution. Of course, 2 M 2 .Œ0; T/ thanks to Theorem 9.1.

Let us turn to uniqueness. Let 1 ; 2 be two solutions, then
j1 .t/ 2 .t/j

ˇZ t Z
ˇˇ ˇˇ t ˇ
ˇ ˇ
ˇ b.1 .s/; s/ b.2 .s/; s/ dsˇ C ˇ .1 .s/; s/ .2 .s/; s dBs ˇ :
0 0
Using (9.18) and (9.19) and the Lipschitz continuity of the coefficients,
h i
E sup j1 .u/ 2 .u/j2
h ˇZ u 0ut
ˇ2 i
ˇ ˇ
2E sup ˇ Œb.1 .s/; s/ b.2 .s/; s/ dsˇ
0ut 0
h ˇZ u ˇ2 i
ˇ ˇ
C2E sup ˇ Œ.1 .s/; s/ .2 .s/; s/ dBs ˇ
0ut 0
hZ t i (9.25)
2tE jb.1 .s/; s/ b.2 .s/; s/j2 ds
0
hZ t i
C8E j.1 .s/; s/ .2 .s/; s/j2 ds
0Z
t
2
L Œ2T C 8 E j1 .s/ 2 .s/j2 ds :
0
Therefore, if v.t/ D EŒsup0ut j1 .u/ 2 .u/j2 , v is bounded by Theorem 9.1, and
satisfies the relation
Z t
v.t/ c v.s/ ds 0tT
0
with c D L2 .2T C 8/; by Gronwall’s inequality v 0 on Œ0; T, i.e. (9.21).

t
u
Remark 9.4 Let us assume that the initial condition is deterministic, i.e.
x 2 Rm and that the starting time is u D 0. Let us consider the filtration
Ht D .s ; s t/ generated by the solution of (9.1). Then we have Ht G t
(recall that .G t /t denotes the augmented natural filtration of the Brownian
motion).
(continued )

Actually, it is immediate to check by induction that all the approximants
n are such that n .t/ is G t -measurable. Hence the same holds for t , which is
the a.s. limit of .n .t//n .
The opposite inclusion is not always true. Consider, for example, the case
where the last k columns of the diffusion coefficient vanish. In this case the
solution would be independent of the Brownian motions BdkC1 ; : : : ; Bd ,
which cannot be adapted to the filtration .Ht /t .
We shall see later that the question of whether the two filtrations .Ht /t and
.G t /t coincide is of some interest.
9.5 Localization and existence for locally Lipschitz

coefficients
In this section we prove existence and uniqueness under weaker assumptions than
Lipschitz continuity of the coefficients b and . The idea of this extension is
contained in the following result, which is of great importance by itself. It states
that as far as the solution remains inside an open set D, its behavior depends only
on the values of the coefficients inside D.
Theorem 9.3 (Localization) Let bi , i , i D 1; 2, be measurable functions on

Rm Œu; T. Let i ; i D 1; 2, be solutions of the SDE
di .s/ D bi .i .s/; s/ ds C i .i .s/; s/ dBs

i .u/ D
where is a square integrable, Fu -measurable r.v. Let D Rm be an open set

such that, on D Œu; T, b1 D b2 , 1 D 2 and, for every x; y 2 D, u t T,
jbi .x; t/ bi .y; t/j Ljx yj; ji .x; t/ i .y; t/j Ljx yj:
Then, if i denotes the exit time of i from D,

1 ^ T D 2 ^ T q.c. and P 1 .t/ D 2 .t/ for every u t 1 ^ T D 1 :
Proof Let, for t T,
e
b.x; t/ D b1 .x; t/1D .x/ D b2 .x; t/1D .x/; e
.x; t/ D 1 .x; t/1D .x/ D 2 .x; t/1D .x/:
9.5 Localization and existence for locally Lipschitz coefficients 271
Then it is easy to see that, as in the proof of Theorem 9.1,

Z tî Z tî
i .t ^ i / D C b.i .s/; s/ ds C .i .s/; s/ dBs
0 0
Z t Z t
DC e
b.i .s ^ i /; s/ ds C e
.i .s ^ i /; s/ dBs
0 0
and therefore, for t T,
sup j1 .r ^ 1 / 2 .r ^ 2 /j2

0rt
ˇZ r ˇˇ2
ˇ e
2 sup ˇ b.1 .s ^ 1 /; s/ e
b.2 .s ^ 2 /; s/ dsˇ
0rt 0
ˇZ r ˇ2
ˇ ˇ
C2 sup ˇ e
.1 .s ^ 1 /; s/ e
.2 .s ^ 2 /; s/ dBs ˇ :
0rt 0
A repetition of the arguments that led us to (9.25) (Doob’s inequality for the
stochastic integral and Hölder’s inequality for the ordinary one) gives
h i Z t
2 2
E sup j1 .r ^ 1 / 2 .r ^ 2 /j L .2T C 8/E j1 .s ^ 1 / 2 .s ^ 2 /j2 ds
0rt 0
Z t
L2 .2T C 8/E sup j1 .r ^ 1 / 2 .r ^ 2 /j2 dr ;
0 0rs
i.e., if we set v.t/ D E sup0rt j1 .r ^ 1 / 2 .r ^ 2 /j2 , v satisfies the relation

Z t
2
v.t/ L .2T C 8/ v.s/ ds :
0
Now, v is bounded by Theorem 9.1 and Gronwall’s inequality gives v.t/ 0, so

that
h i
v.T/ D E sup j1 .r ^ 1 / 2 .r ^ 2 /j2 D 0 ;
0rT
whence the two statements follow simultaneously.

t
u
Assumption (A’) We say that b and satisfy Assumption (A’) if they are
measurable in .x; t/ and
(continued )
(continued)
i) have sublinear growth (i.e. satisfy (9.9));
ii) are locally Lipschitz continuous in x, i.e. for every N > 0 there exists an
LN > 0 such that, if x; y 2 Rm , jxj N, jyj N, t 2 Œ0; T,
jb.x; t/ b.y; t/j LN jx yj; j.x; t/ .y; t/j LN jx yj ;
i.e. their restriction to every bounded set is Lipschitz continuous.
Theorem 9.4 Under Assumption (A’), let be a square integrable Fu -

measurable r.v. Then there exists a process such that

u D :
Moreover, 2 M 2 .Œ0; T/ and pathwise uniqueness holds.
Proof Existence. We shall assume u D 0. For N > 0 let ˚N 2 CK1 .Rm / such that
0 ˚N 1 and
(
1 if jxj N
˚N .x/ D
0 if jxj N C 1 :
Let
bN .x; t/ D b.x; t/˚N .x/; N .x; t/ D .x; t/˚N .x/ :
N and bN therefore satisfy Assumption (A) and there exists a N 2 M 2 .Œ0; T/ such
that
Z t Z t
N .t/ D C bN .N .s/; s/ ds C N .N .s/; s/ dBs :
0 0
Let N be the exit time of N from the ball of radius N. If N 0 > N, on fjxj Ng
Œ0; T we have bN D bN 0 and N D N 0 . Therefore, by the localization Theorem 9.3,
the two processes N 0 and N coincide for t T until their exit from the ball of
radius N, i.e. on the event fN > Tg a.s. By Remark 9.3, moreover,

P.N > T/ D P sup jN .t/j N ! 1
0tT N!1
and therefore fN > Tg % ˝ a.s. We can therefore define t D N .t/ on fN > Tg:
this is a well defined since if N 0 > N then we know that N 0 .t/ D N .t/ on fN > Tg;
by Theorem 7.2 on the event fN > Tg we have
9.6 Uniqueness in law 273
Z t Z t
t D N .t/ D C bN .N .s/; s/ ds C N .N .s/; s/ dBs
0 0
Z t Z t
DC b.s ; s/ ds C .s ; s/ dBs ;
0 0
hence, by the arbitrariness of N, is a solution. Of course, 2 M 2 .Œ0; T/ thanks to

Theorem 9.1.
Uniqueness. Let 1 ; 2 be two solutions and let, for i D 1; 2, i .N/ D infft
TI ji .t/j > Ng. Then, by Theorem 9.3, 1 .N/ D 2 .N/ a.s. and 1 and 2 coincide
a.s. on f1 .N/ > Tg. As f1 .N/ > Tg % ˝ a.s., 1 and 2 coincide for every
t 2 Œ0; T a.s.
t
u
Remark 9.5 With the notations of the previous proof N ! uniformly for
almost every ! (N .!/ and .!/ even coincide a.s. on Œ0; T for N large
enough). Moreover, if x, the event fN Tg on which N and are
different has a probability that goes to 0 as N ! 1 uniformly for u 2 Œ0; T
and x in a compact set.
Remark 9.6 A careful look at the proofs of Theorems 9.1, 9.2, 9.3, 9.4
shows that, in analogy with ordinary differential equations, hypotheses of
Lipschitz continuity of the coefficients are needed in order to guarantee
local existence and uniqueness (thanks to Gronwall’s inequality) whereas
hypotheses of sublinear growth guarantee global existence. Exercise 9.23
presents an example of an SDE whose coefficients do not satisfy (9.9) and
admits a solution that is defined only on a time interval Œ0; .!/Œ with <
C1 a.s.
9.6 Uniqueness in law
In this section we prove uniqueness in law of the solutions under Assumption (A’).
Theorem 9.5 Let Bi D .˝ i ; F i ; .Fti /t ; .Bit /t ; Pi /, i D 1; 2, be d-dimen-

sional standard Brownian motions, i ; i D 1; 2, m-dimensional r.v.’s in
L2 .˝ i ; Fui ; Pi /, respectively, and having same law. Let us assume that b and
satisfy Assumption (A’) and let i be the solutions of
(continued)

Z t Z t
i .t/ D i C b. i .s/; s/ ds C . i .s/; s/ dBis :
u u
Then the processes . i ; Bi /, i D 1; 2, have the same law.
Proof We assume u D 0. Let us first assume Assumption (A). The idea of the proof
consists in proving by induction that, if ni are the approximants defined in (9.22),
then the processes .ni ; Bi /, i D 1; 2, have the same law for every n.
This is certainly true for n D 0 as 0i i is independent of Bi and 1 and
2 have the same law. Let us assume then that .n1 i
; Bi /, i D 1; 2, have the same
law and let us prove that the same holds for .n ; B /. This means showing that the
i i
finite-dimensional distributions of .ni ; Bi / coincide. This is a consequence of the

following lemma.
Lemma 9.2 Let Bi D .˝ i ; F i ; .Fti /t ; .Bit /t ; Pi /, i D 1; 2, be d-dimen-

sional standard Brownian motions, i ; i D 1; 2, m-dimensional r.v.’s in
L2 .˝ i ; Fui ; Pi /, respectively, and having same law. Let us assume that b
and satisfy Assumption (A) and let 1 ; 2 be processes in Mloc;B 2
1 .Œ0; T/,
2
Mloc;B2 .Œ0; T/ respectively such that .i ; ; B /, i D 1; 2, have the same law.
i i
Then, if
Z t Z t
Yi D i C b.si ; s/ ds C .si ; s/ dBis ;
0 0
the processes .i ; Y i ; Bi / have the same finite-dimensional distributions.
Proof If the i ’s are elementary processes and b and are linear combinations
of functions of the form g.x/1Œu;vŒ .t/ this is immediate as in this case b.si ; s/ and
.si ; s/ are still elementary processes and by Definition 7.15 we have directly that
the finite-dimensional distributions of .1 ; Y 1 ; B1 / and .2 ; Y 2 ; B2 / coincide. The
passage to the general case is done by first approximating the i ’s with elementary
processes and then b and with functions as above and using the fact that a.s.
convergence implies convergence in law. The details are omitted. t
u
End of the Proof of Theorem 9.5 As the equality of the finite-dimensional distri-
butions implies equality of the laws, the processes .ni ; Bi / have the same law.
As .ni ; Bi / converges to . i ; Bi / uniformly in t and therefore in the topology of
C .Œ0; T; RmCd / (see the proof of Theorem 9.2) and the a.s. convergence implies
convergence of the laws, the theorem is proved under Assumption (A).
9.7 The Markov property 275
Let us assume now Assumption (A’). By the first part of this proof, if N1 ; N2 are
the processes as in the proof of Theorem 9.4, .N1 ; B1 / and .N2 ; B2 / have the same
law. As Ni ! i uniformly a.s. (see Remark 9.5), the laws of .Ni ; Bi /, i D 1; 2,
converge as N ! 1 to those of . i ; Bi /, which therefore coincide. t
u
9.7 The Markov property
In this section we prove that, under Assumption (A’), the solution of an SDE is a
Markov process and actually a diffusion associated to a generator that the reader can
already imagine. In Sect. 9.9 we shall consider the converse problem, i.e. whether
for a given differential operator there exists a diffusion process that is associated to
it and whether it is unique in law.
;s
Let, for an Fs -measurable r.v. , t be the solution of

(9.26)
s D
and let us prove first the Markov property, i.e. that, for every 2 B.Rm / and
s u t,
;s
P t 2 jFus D p.u; t; u;s ; / (9.27)
for some transition function p. We know that the transition function p.u; t; x; / is the
law at time t of the process starting at x at time u, i.e. if tx;u is the solution of

(9.28)
u D x
then the candidate transition function is, for 2 B.Rm /,
p.u; t; x; / D P.tx;u 2 / : (9.29)
In order to prove the Markov property (9.27) we note that

Z t Z t
;s
t DC b.v;s ; v/ dv C .v;s ; v/ dBv
Z u s Z u s
DC b.v;s ; v/ dv C .v;s ; v/ dBv

„ s
ƒ‚ s … (9.30)
;s
Du
Z t Z t
C b.v;s ; v/ dv C .v;s ; v/ dBv ;
u u
i.e. the value at time t of the solution starting at at time s is the same as the value
;s
of the solution starting at u at time u, i.e.
;s
;s ;u
t D t u : (9.31)
Let us define
.x; !/ D 1 .tx;u .!//
and observe that
.x/ WD EŒ .x; / D EŒ1 .tx;u / D P.tx;u 2 / D p.u; x; t; / : (9.32)
By (9.31)
;s
;u
.u;s ; !/ D 1 .t u .!//
and, going back to (9.27),

;s ;s ;u
P t 2 jFus D P t u 2 jFus D E .u;s ; /jFus :
;s
The r.v. u is Fus -measurable. If we prove that the r.v. tx;u is independent of Fus ,
then .x; / is independent of Fus and we can apply the freezing lemma (Lemma 4.1)
and obtain
;s
P t 2 jFus D E .u;s ; /jFus D .u;s / D p.u; u;s ; t; /
( is defined above in (9.32)), which proves the Markov property.

Hence in order to complete this proof we still have to
a) prove that tx;u is a r.v. that is independent of Fus , and, unfortunately, also that
b) the map .x; u/ 7! tx;u is measurable for every ! or at least that there is
a measurable version of it. This is the object of the next section, where
(Theorem 9.9) it is shown that, under Assumption (A), there exists a version of
.x; u; t/ 7! tx;u that is continuous in the three arguments x; u; t (it will be another
application of Theorem 2.1, the continuity Kolmogorov theorem).
Fact a) above follows easily from the following lemma. Let Hu0 D .Bv
Bu ; v u/ be the -algebra generated by the increments of B after time u and
Hu the -algebra obtained by adding to Hu0 the negligible sets of F1 . Of course
(Exercise 3.4) Hu is independent of Fu .
Lemma 9.3 Let Y 2 M 2 be a process such that Ys is Hu -measurable for every

s u. Then the r.v.
Z t
Ys dBs (9.33)
u
is also Hu -measurable for every t u and therefore independent of Fu .
P
Proof The statement is immediate if Y D m iD1 Xi 1Œti ;tiC1 Œ is an elementary process.
In this case the r.v.’s Xi are Hu -measurable and in the explicit expression (7.7)
only the increments of B after time u and the values of the Xi appear. It is
also immediate that in the approximation procedure of a general integrand with
elementary processes described at the beginning of Sect. 7.3 if Y is Hu -measurable,
then the approximating elementary processes are also Hu -measurable. Then just
observe that the stochastic integral in (9.33) is the a.s. limit of the integrals of the
approximating elementary processes.
t
u
Let us assume Assumption (A). If we go back to the successive approximations
scheme that is the key argument in the proof of existence in Theorem 9.2, we see
that .tx;u /tu is the limit of the approximating processes m , where 0 .t/ x and
Z t Z t
mC1 .t/ D x C b.m .v/; v/ dv C .m .v/; v/ dBv :
u u
The deterministic process 0 x is obviously Hu -measurable and by recurrence,

using Lemma 9.3, the processes m are such that m .t/ also is Hu -measurable, as
well as .tx;u /tu which is their a.s. limit.
Hence, given as granted b) above, we have proved the Markov property under
Assumption (A).
Theorem 9.6 Under Assumption (A’), p defined in (9.29) is a transition

function and the process .˝; F ; .Ft /t ; .tx;s /t ; P/ is a Markov process starting
at time s, with initial distribution ıx and associated to the transition function p.
Proof Under Assumption (A) the Markov property has already been proved. Under
Assumption (A’) we shall approximate the solution with processes that are solutions
of an SDE satisfying Assumption (A): let N and N be as in the proof of
Theorem 9.4 and let
pN .s; t; x; / WD P.Nx;s .t/ 2 / :

N is the solution of an SDE with coefficients satisfying Assumption (A), and

therefore enjoying the Markov property, i.e.
P.Nx;s .t/ 2 jFus / D pN .u; t; Nx;s .u/; / :
We now have to pass to the limit as N ! 1 in the previous equation. As tx;s and
Nx;s .t/ coincide on fN > tg and P.N t/ ! 0 as N ! 1, we have, using the fact
that probabilities pass to the limit with a monotone sequences of events,
pN .s; t; x; / D P.Nx;s .t/ 2 /

D P.tx;s 2 ; N > t/ C P.Nx;s .t/ 2 ; N t/ ! P.tx;s 2 / (9.34)
N!1
D p.s; t; x; / :
From this relation we derive that
pN .u; t; Nx;s .u/; / ! p.u; t; x;s .u/; /

N!1
as Nx;s .u/ D x;s .u/ for N large enough. Moreover,

P tx;s 2 jFus D E 1 .tx;s /1fN >tg jFus C E 1 .tx;s /1fN tg jFus

D E 1 .Nx;s .t//1fN >tg jFus C E 1 .tx;s /1fN tg jFus

D E 1 .Nx;s .t//jFus C E .1 .tx;s / 1 .Nx;s .t///1fN tg jFus :
The second term on the right-hand side tends to zero a.s., as P.N t/ & 0 as
N ! 1, hence

lim P.Nx;s .t/ 2 jFus / D lim E 1 .Nx;s .t//jFus D P tx;s 2 jFus a:s:
n!1 n!1
and putting things together
P.tx;s 2 jFus / D lim EŒ1 .Nx;s .t//jFus

N!1
D lim pN .u; t; Nx;s .u/; / D p.u; t; x;s .u/; / :

N!1
The Markov property is therefore proved under Assumption (A’). We still have to
prove that p satisfies the Chapman–Kolmogorov equation; this is a consequence of
the argument described in Remark 6.2: as p.s; t; x; / is the law of tx;s
p.s; t; x; / D P.tx;s 2 / D EŒP.tx;s 2 jFus / D EŒp.u; t; ux;s ; /

Z
D p.u; t; y; /p.s; u; x; dy/ :
t
u
We can now construct a realization of the Markov process associated to the solution
of an SDE. In order to do this, let C D C .Œ0; T; Rm / and let Xt W C ! Rm be the
applications defined by Xt . / D .t/. Let M D B.C /, Mts D .Xu ; s u t/,
M1 s
D .Xu ; u s/. On C let Px;s be the law of the process x;s (Sect. 3.2) and
denote by Ex;s the expectation computed with respect to Px;s . Of course, if p is the
transition function defined in (9.29), we have p.s; t; x; / D Px;s .Xt 2 /.
As, by definition, the finite-dimensional distributions of .Xt /ts with respect to
Px;s coincide with those of x;s with respect to P, the following theorem is obvious.
Theorem 9.7 Under Assumption (A’), .C ; M ; .Mts /ts ; .Xt /ts ; .Px;s /x;s / is
a realization of the Markov process associated to the transition function p
defined in (9.29).
Thanks to Theorem 9.5, which guarantees the uniqueness in law, the probability
Px;s does not depend on the Brownian motion that is used to construct the solution
x;s . Therefore the realization .C ; M ; .Mts /ts ; .Xt /ts ; .Px;s /x;s / is well defined. We
shall call this family of processes the canonical Markov process associated to the
SDE (9.1).
Theorem 9.8 Under Assumption (A’) the canonical Markov process associ-
ated to (9.1) is a diffusion with generator
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ ,
2 i;jD1 @xi @xj iD1
@xi
where a.x; t/ D .x; t/.x; t/ .
Proof Let us prove first the Feller property which will imply strong Markovianity
thanks to Theorem 6.1. We must prove that, for every h > 0 and for every bounded
continuous function f , the map
Z
x;t
.x; t/ 7! f .y/p.t; t C h; x; dy/ D Ex;t Œ f .XtCh / D EŒ f .tCh /
is continuous. Under Assumption (A) this is immediate, using the fact that tx;s .!/ is
a continuous function of x; s; t; t s thanks to the forthcoming Theorem 9.9. Under
Assumption (A’) if Nx;s and N are as in the proof of Theorem 9.4 then as x;s and
Nx;s coincide on fN > Tg,
x;t
EŒ f .tCh / D EŒ f .Nx;t .t C h// C E .f .tCh

x;t
/ f .Nx;t .t C h///1fN Tg :
x;t
Now it is easy to deduce that .x; t/ 7! EŒ f .tCh / is continuous, as the first term
on the right-hand side is continuous in .x; t/ whereas the second one is majorized
by 2kf k1 P.N T/ and can be made small uniformly for .x; t/ in a compact set,
thanks to Remark 9.3.
Let us prove that Lt is the generator. By Ito’s formula, as explained in Remark 9.1,
if f 2 CK2 ,
Ts;t f .x/ D EŒ f .tx;s /

hZ t i hZ t Xm X d
@f i
D f .x/ C E Lu f .ux;s / du C E ij .ux;s ; u/ .ux;s / dBj .u/ :
s s iD1 jD1 @xi
The last integrand is a process of M 2 .Œs; t/, as x;s 2 M 2 .Œs; t/ and has a sublinear
growth (the derivatives of f are bounded), therefore the expectation of the stochastic
integral vanishes and we have
Z t Z t
Ts;t f .x/ D f .x/ C EŒLu f .ux;s / du D f .x/ C Ts;u .Lu f /.x/ du :
s s
t
u
Example 9.3 What is the transition function p.t; x; / of the Ornstein–

Uhlenbeck process, i.e. the Markov process that is the solution of
dt D t dt C dBt ‹ (9.35)
By (9.29) p.t; x; / is the law of tx , i.e. of the position at time t of the solution
of (9.35) starting at x. We have found in Sect. 9.2 that
Z t
t D e t x C e t e s dBs :
0
Now simply observe that t has a distribution that is Gaussian with mean e t x
and variance
Z t
2
t2 WD e2 .ts/ 2 ds D .1 e2 t / :
0 2
This observation can be useful if you want to simulate the process : just fix a
discretization step h. Then choose at random a number z with distribution
N.e h x; h2 / and set h D z. Then choose at random a number z with
distribution N.e h h ; h2 / and set 2h D z and so on: the position .kC1/h
will be obtained by sampling a number with distribution N.e h kh ; h2 /.
(continued )

This procedure allows us to simulate subsequent positions of an Ornstein–
Uhlenbeck process exactly, which means that the positions .x; h ; : : : ; mh /
have exactly the same joint distributions as the positions of the Markov
process which is the solution of (9.35) at the times 0; h; : : : ; mh.
This procedure can be easily extended to the case of a multidimensional
Ornstein–Uhlenbeck process, the only difference being that in that case the
transition function is a multivariate Gaussian distribution.
Chapter 11 is devoted to the simulation problems of a more general class
of diffusion processes.
Example 9.4 Let B be a real Brownian motion and X the solution of the two-
dimensional SDE
d1 .t/ D b1 .1 .t// dt C 1 .t/ dBt

d2 .t/ D b2 .2 .t// dt C 2 .t/ dBt :
What is the differential generator of ?

This equation can be written as
dt D b.t / dt C .t / dBt ;

b1 .x1 / x
where b.x1 ; x2 / D and .x/ D 1 : here the Brownian motion B
b2 .x2 / x2
is one-dimensional.
In order to obtain the generator, the only thing to be computed is the matrix
a D of the second-order coefficients. We have
2
x1 x1 x1 x2
a.x/ D .x/ .x/ D x1 x2 D :
x2 x1 x2 x22
Therefore
1 2 @2 @2 @2 @ @
LD x1 2 C x22 2 C 2x1 x2 C b1 .x1 / C b2 .x2 /
2 @x1 @x2 @x1 @x2 @x1 @x2
And if it was
d1 .t/ D b1 .1 .t// dt C 1 .t/ dB1 .t/

d2 .t/ D b2 .2 .t// dt C 2 .t/ dB2 .t/ ;
(continued )

where B D .B1 ; B2 / is a two-dimensional Brownian motion?
Now

x1 0
.x/ D
0 x2
and
2
x1 0
a.x/ D .x/ .x/ D
0 x22
so that
1 2 @2 @2 @ @
LD x1 2 C x22 2 C b1 .x1 / C b2 .x2 /
2 @x1 @x2 @x1 @x2
9.8 Lp bounds and dependence on the initial data
Let us see how the solution depends on the initial value and initial time u. What if
one changes the starting position or the starting time “just a little”? Does the solution
change “just a little” too? Or, to be precise, is it possible to construct solutions so
that there is continuity with respect to initial data?
In some situations, when we are able to construct explicit solutions, the answer to
this question is immediate. For instance, if sx;t is the position at time s of a Brownian
motion starting at x at time t, we can write
sx;t D x C .Bt Bs / ;
where B is a Brownian motion. It is immediate that this is, for every !, a continuous
function of the starting position x, of the starting time t and of the actual time s.
Similar arguments can be developed for other processes for which we have explicit
formulas, such as the Ornstein–Uhlenbeck process or the geometric Brownian
motion of Sect. 9.2.
The aim of this chapter is to give an answer to this question in more general
situations. It should be no surprise that the main tool will be in the end Kolmogorov’s
continuity Theorem 2.1.
9.8 Lp bounds and dependence on the initial data 283
Proposition 9.1 Under Assumption (A), let 1 and 2 be solutions of (9.20)

with initial conditions 1 ; u and 2 ; v, respectively. Let us assume u v.
Then, if p 2,
h i
E sup j1 .t/ 2 .t/jp
vtT (9.36)
c.L; T; M; p/ 1 C E.j1 jp / ju vjp=2 C E.j1 2 jp / :
Proof For t v
1 .t/ 2 .t/
Z Z
t
t
D 1 .v/ 2 C b.1 .r/; r/ b.2 .r/; r/ dr C .1 .r/; r/ .2 .r/; r/ dBr
v v
and therefore, again using Hölder’s inequality on the integral in dt and Doob’s
inequality on the stochastic integral,
h i
E sup j1 .t/ 2 .t/jp
vts

hZ s i
3p1 E j1 .v/ 2 j p
C 3p1 T p1 E jb.1 .r/; r/ b.2 .r/; r/jp dr
v
ˇZ t
h ˇp i
ˇ ˇ
C3p1 E sup ˇ Œ.1 .r/; r/ .2 .r/; r/ dBr ˇ
vts v
Z h i

s
3 p1
E j1 .v/ 2 j p
C .L; T; M; p/ E sup j1 .t/ 2 .t/jp dr :
v vtr
The function s 7! EŒsuputs j1 .t/ 2 .t/jp is bounded thanks to (9.12). One can
therefore apply Gronwall’s inequality and get
h i

E sup j1 .t/ 2 .t/jp c.L; T; M; p/ 1 C E j1 .v/ 2 jp :
vtT
Now, by (9.13),

E j1 .v/ 2 jp 2p1 E j1 .v/ 1 jp C E j1 2 jp

c.T; M; p/ 1 C E j1 jp jv ujp=2 C E j1 2 jp

c0 .T; M; p/ 1 C E j1 jp jv ujp=2 C E j1 2 jp ;

t
u
Let us denote by tx;s the solution of

(9.37)
s D x :
A classical result of ordinary differential equations is that, under suitable conditions,

the solution depends continuously on the initial data x; s. The corresponding result
for SDEs is the following.
Theorem 9.9 Under Assumption (A) there exists a family .Zx;s .t//x;s;t of r.v.’s
such that
a) the map .x; s; t/ 7! Zx;s .t/ is continuous for every ! for x 2 Rm ; s; t 2
RC ; s t.
b) Zx;s .t/ D tx;s a.s. for every x 2 Rm , s; t 2 RC ; s t.
Proof This is an application of Kolmogorov’s continuity Theorem 2.1. Note that,

by (9.30), for every s u; x;u is a solution of

s D sx;u :
Let us fix T > 0 and let u; v; s; t be times with u s; v t. We want to prove, for
x; y with jxj; jyj R the inequality
y;v

E jsx;u t jp c.L; M; T; p; R/ jx yjp C jt sjp=2 C ju vjp=2 : (9.38)
Assume first u v s t. We have

y;v

y;v
E jsx;u t jp 2p1 E jsx;u sy;v jp C 2p1 E jsy;v t jp :
Using Proposition 9.1 in order to bound the first term on the right-hand side, we
have

E jsx;u sy;v jp c.L; T; M; p/ 1 C jxjp ju vjp=2 C jx yjp ;
y;v
whereas for the second one, recalling that t is the solution at time t of the SDE
with initial condition sy;v at time s, (9.12) and (9.13) give
y;v
E jsy;v t jp c. p; T; M/jt sjp=2 .1 C EŒjsy;v jp /

c. p; T; M/jt sjp=2 .1 C jyjp / ;
9.9 The square root of a matrix field and the problemof diffusions 285
which together give (9.38). The same argument proves (9.38) in the case u v
t s.
If, instead, u s v t, we must argue differently (in this case sy;v is not
defined); we have
y;v
E jsx;u t jp

y;u
y;u y;v
3p1 E jsx;u sy;u jp C 3p1 E jsy;u t jp C 3p1 E jt t jp
and again Proposition 9.1, (9.12) and (9.13) give (9.38). In conclusion, for p large
enough,
y;v
mC2C˛
E jsx;u t jp c jx yj C jt sj C ju vj
for some ˛ > 0. Theorem 2.1 guarantees therefore the existence of a continuous
version of tx;s in the three variables x; s; t; s t for jxj R. R being arbitrary, the
statement is proved.
t
u
Note that in the previous proof we can choose ˛ D p2 m 2 and a more precise
application of Theorem 2.1 allows us to state that the paths of x;u are Hölder
1
continuous with exponent < 2p . p 2m/ D 12 .1 mp / and therefore, by the
arbitrariness of p, Hölder continuous with exponent for every < 12 , as for the
Brownian motion.
9.9 The square root of a matrix field and the problem

of diffusions
Given the differential operator
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ , (9.39)
2 i;jD1 @xi @xj iD1
@x i
where the matrix a.x; t/ is positive semidefinite, two questions naturally arise:
• under which conditions does a diffusion associated to Lt exist?
• what about uniqueness?
From the results of the first sections of this chapter it follows that the answer to
the first question is positive provided there exists a matrix field .x; t/ such that
a.x; t/ D .x; t/.x; t/ and that and b satisfy Assumption (A’) (local Lipschitz
continuity and sublinear growth in the x variable, in addition to joint measurability).
In general, as a.x; t/ is positive semidefinite, for fixed x; t there always exists
an m m matrix .x; t/ such that .x; t/.x; t/ D a.x; t/. Moreover, it is unique
under the additional assumption that it is symmetric. We shall denote this symmetric
matrix field by , so that .x; t/2 D a.x; t/.
Let us now investigate the regularity of . Is it possible to take such a square
root in such a way that .t; x/ 7! .x; t/ satisfies Assumptions (A) or (A’)? Note also
that, due to the lack of uniqueness of this square root , one will be led to enquire
whether two different square root fields .x; t/ might produce different diffusions
and therefore if uniqueness is preserved. We shall mention the following results.
See Friedman (1975, p. 128), Priouret (1974, p. 81), or Stroock and Varadhan (1979,
p. 131) for proofs and other details.
Proposition 9.2 Let D Rm be an open set and let us assume a.x; t/ is

positive definite for every .x; t/ 2 D Œ0; T. Then if a is measurable in .x; t/,
the same is true for . If a is locally Lipschitz continuous in x then this is also
true for . If a is Lipschitz continuous in x and uniformly positive definite,
is Lipschitz continuous in x.
To be precise, let us recall that “a.x; t/ positive definite” means that
X
m
ha.x; t/z; zi D aij .x; t/zi zj > 0 (9.40)
i;jD1
for every z 2 Rm , jzj > 0, or, equivalently, that, for every x; t, the smallest eigenvalue
of a.x; t/ is > 0. “a uniformly positive definite” means
X
m
ha.x; t/z; zi D aij .x; t/zi zj > jzj2 (9.41)
i;jD1
for every z 2 Rm and some > 0 or, equivalently, that there exists a > 0 such
that the smallest eigenvalue of a.x; t/ is > for every x; t.
Definition 9.5 We say that the matrix field a is elliptic if a.x; t/ is positive
definite for every x; t and that it is uniformly elliptic if a.x; t/ is uniformly
positive definite.
Remark 9.7 Note that
ha.x; t/z; zi D h.x; t/ .x; t/z; zi D h .x; t/z; .x; t/zi D j .x; t/zj2
(continued )

hence (9.40) is equivalent to requiring that
j .x; t/zj2 > 0
for every z 2 Rm ; z 6D 0, i.e. the matrix .x; t/ W Rm ! Rd must be injective.

As .x; t/ is an md matrix, (9.40) can hold only if d m, otherwise .x; t/
would have a kernel of dimension 1 (recall that m is the dimension of the
diffusion process and d the dimension of the driving Brownian motion).
Of course (9.40) implies that a.x; t/ is invertible for every x; t and (9.41)
that
1
ha.x; t/1 w; wi jwj2 :

If a is not positive definite, the square root may even not be locally Lipschitz
continuous, even if a is Lipschitz pcontinuous. Just consider, for m D 1, the case
a.x/ D jxj and therefore .x/ D jxj. However, we have the following result.
Proposition 9.3 Let D Rm be an open set and assume that a.x; t/ is positive
semidefinite for every .x; t/ 2 D Œ0; T. If a is of class C2 in x on D Œ0; T,
then is locally Lipschitz continuous in x. If, moreover, the derivatives of
order 2 of a are bounded, is Lipschitz continuous in x.
We can now state the main result of this section.
Theorem 9.10 Let us assume that b satisfies Assumption (A’) and that a.x; t/
is measurable in .x; t/, positive semidefinite for every .x; t/ and of class C2 in
x or positive definite for every .x; t/ and locally Lipschitz continuous. Let us
assume, moreover, that, for every t 2 Œ0; T,
ja.x; t/j M 2 .1 C jxj2 / :
Then on the canonical space .C ; M ; .Mts /ts ; .Xt /t0 / there exists a unique
family of probabilities .Px;s /x;s such that .C ; M ; .Mts /ts ; .Xt /ts ; .Px;s /x;s / is
the realization of a diffusion process associated to Lt . Moreover, it is a Feller
process.
Proof If is the symmetric square root of the matrix field a, then the coefficients
b; satisfy Assumption (A’) and if x;s denotes the solution of

(9.42)
s D x
then the probabilities Px;s on .C ; M / introduced p. 279 satisfy the conditions

required.
Uniqueness requires more attention. We must prove that if e
is any matrix field
such that a.x; t/ D e .x; t/ and e
.x; t/e is a solution of
de
t D b.e .e
t ; t/ dt C e t ; t/ dBt
e (9.43)
s D x
then the laws of x;s and e coincide. If a is elliptic, so that a.x; t/ is invertible for
every x; t, and e
is another m m matrix field such that e .x; t/ D a.x; t/ for
.x; t/e
every x; t, then there exists an orthogonal matrix field .x; t/, such that e D (just
set D 1e and use the fact that a and its symmetric square root commute).
Hence e is a solution of
de
t D b.e
t ; t/ dt C .e
t ; t/.e
t ; t/ dBt D b.e
t ; t/ dt C .e
t ; t/ de
Bt ;
where
Z t
e
Bt D .e
s ; s/ dBs
0
is once again a Brownian motion by Proposition 8.8. By Theorem 9.5 (uniqueness

in law under Assumption (A’)) and e have the same law.
This gives the main idea of the proof, but we must take care of other possible
situations. What if a is not elliptic? And if e
was an m d matrix field with d 6D m?
The first situation is taken care of by the following statement.
Lemma 9.4 Let A be an m m semi-positive definite matrix and S1 , S2 be

two m m matrices such that
A D Si Si :
Then there exists an orthogonal matrix Q such that S1 D S2 Q.
Proof There exists an orthogonal matrix O such that A D O DO; where D is

diagonal and with non-negative entries. Then if Ri D OSi , we have
Ri Ri D OSi Si O D OO DOO D D : (9.44)

Suppose A has m0 non-zero eigenvalues, ordered in such a way that

e
D0
DD
0 0
for an m0 m0 diagonal matrix e D with strictly positive entries in its diagonal. Let ej
be the vector having 1 as its j-th coordinate and 0 for the others. Then if j > m0
0 D hDej ; ej i D hRi Ri ej ; ej i D jRi ej j2 ;
which implies that the last m m0 columns of Ri vanish. Hence the last m m0 rows
of Ri vanish and Ri is of the form
0
Ri
Ri D :
0
Clearly R0i R0i D e

D so that if e
Ri D e
D1=2 R0i we have
e 1=2
Ri D e
Rie D1=2 R0i R0i e
D DI:
Hence the rows of e

Ri are orthonormal. Let b
Ri be any .m m0 / m matrix whose
rows complete to an orthonormal basis of R . Then the matrices
m

e
R
RL i D bi
Ri
are orthogonal and there exists an orthogonal matrix Q such that
RL 1 D RL 2 Q :
R1 D e
From this relation, “going backwards” we obtain e R2 Q, then R1 D R2 Q and
finally S1 D S2 Q. t
u
End of the Proof of Theorem 9.10 Let us now consider the case of an m d-matrix
field e such that e 2 D a.x; t/ for every x; t, hence consider in (9.43) a d-
2e
dimensional Brownian motion possibly with d 6D m. Also in this case we have a
diffusion process associated to the generator L and we must prove that it has the
same law as the solution of (9.42). The argument is not really different from those
developed above.
Let us assume first that d > m, i.e. the Brownian motion has a dimension that is
strictly larger than the dimension of the diffusion . Then e .x; t/ has a rank strictly
smaller than d. Let us construct a d d orthogonal matrix .x; t/ by choosing its last
d m columns to be orthogonal unitary vectors in ker e .x; t/ and then completing
the matrix by choosing the other columns so that the columns together form an
orthogonal basis of Rd . Then (9.43) can be written as
t D b.e
de .e
t ; t/ dt C e t ; t/.e
t ; t/.e
t ; t/1 dBt
D b.e
t ; t/ dt C e
.t ; t/ dWt ;
Rt
where e .x; t/.x; t/ and Wt D 0 .e
.x; t/ D e s ; s/1 dBs is a new d-dimensional
Brownian motion. Now with the choice we have made of .x; t/, it is clear that the
last d m columns of .x; t/ vanish. Hence the previous equation can be written
de
t D b.e 2 .e
t ; t/ dt C e et ;
t ; t/ d W
where e 2 .x; t/ is the m m matrix obtained by taking away from .x; t/ the last
d m columns and W e is the m-dimensional Brownian motion formed by the first m
components of W. It is immediate that 2 .x; t/2 .x; t/ D a.x; t/ and by the first part
of the proof we know that the process e has the same law as the solution of (9.42).
It remains to consider the case where 2 is an m d matrix with d < m (i.e. the
driving Brownian motion has a dimension that is smaller than the dimension of ),
but this can be done easily using the same type of arguments developed so far, so
we leave it to the reader.
Note that this proof is actually incomplete, as we should also prove that the
orthogonal matrices .x; t/ above can be chosen in such a way that the matrix field
.x; t/ ! .x; t/ is locally Lipschitz continuous. t
u
9.10 Further reading
The theory of SDEs presented here does not account for other approaches to the
question of existence and uniqueness.
Many deep results are known today concerning SDEs whose coefficients are not
locally Lipschitz continuous. This is important because in the applications examples
of SDEs of this type do arise.
An instance is the so-called square root process (or CIR model, as it is better
known in applications to finance), which is the solution of the SDE
p
dt D .a bt / dt C t dBt :
2
In Exercise 9.19 we see that if a D 4 then a solution exists and can be obtained
as the square of an Ornstein–Uhlenbeck process. But what about uniqueness? And
what about existence for general coefficients a, b and ?
For the problem of existence and uniqueness of the solutions of an SDE there are
two approaches that the interested reader might look at.
The first one is the theory of linear (i.e. real-valued) diffusion as it is developed
(among other places) in Revuz and Yor (1999, p. 300) or Ikeda and Watanabe (1981,
p. 446). The Revuz and Yor (1999) reference is also (very) good for many other
advanced topics in stochastic calculus.
The second approach is the Strook–Varadhan theory that produces existence and
uniqueness results under much weaker assumptions than local Lipschitz continuity
and with a completely different and original approach. Good references are, without
the pretension of being exhaustive, Stroock and Varadhan (1979), Priouret (1974)
(in French), Rogers and Williams (2000) and Ikeda and Watanabe (1981).
Exercises
9.1 (p. 556) Let be the Ornstein–Uhlenbeck process solution of the SDE
dt D t dt C dBt
(9.45)
0 D x
with > 0 (see Example 9.1).

a) Prove that, whatever the starting position x, the law of t converges as t ! C1
to a probability to be determined.
b) Consider now the process but with a starting position 0 D , where is a
square integrable r.v. having law and independent of the Brownian motion B.
Show that t has distribution for every t.
9.2 (p. 557)

a) Let B be a d-dimensional Brownian motion and let us consider the SDE, in
dimension m,
dt D b.t/t dt C .t/ dBt

(9.46)
0 D x
where b and are locally bounded measurable functions of the variable t only,
with values respectively in M.m/ and M.m; d/ (m m and m d matrices
respectively). Find an explicit solution, show that it is a Gaussian process and
compute its mean and covariance functions (see Example 2.5).
b) Let us consider, for t < 1, the SDE
t
dt D dt C dBt
1t (9.47)
0 D x :
Solve this equation explicitly. Compute the mean and covariance functions of
(which is a Gaussian process by a)). Prove that if x D 0 then is a Brownian
bridge.
9.3 (p. 558) (The asymptotic of the Ornstein–Uhlenbeck processes) Let B D

.˝; F ; .Ft /t ; .Bt /t ; P/ be a real Brownian motion and the Ornstein–Uhlenbeck
process solution of the SDE, for 6D 0, > 0,
dt D t dt C dBt
0 D x :
a) Let us assume > 0.

a1) Prove that
t t
lim p D p ; lim p D p
t!C1 log t t!C1 log t
a2) What can be said of limt!C1 t and limt!C1 t ?

b) Let < 0. Prove that there exists an event A 2 F suchpthat limt!C1 t D C1
on A and limt!C1 t D 1 on Ac with P.A/ D ˚. x 2 /, where ˚ denotes
the partition function of an N.0; 1/-distributed r.v. (in particular P.A/ D 12 if
x D 0).
9.4 (p. 560) Let be the solution of the SDE, for t < 1,
1 t p
dt D dt C 1 t dBt
2 1t (9.48)
0 D x :
a) Find the solution of this equation and prove that it is a Gaussian process.
b) Compare the variance of t with the corresponding variance of a Brownian bridge
at time t. Is a Brownian bridge?
9.5 (p. 561) Let B be a real Brownian motion and the solution of the SDE
dt D b.t/t dt C .t/t dBt

(9.49)
0 D x > 0 ;
where b and are measurable and locally bounded functions of the time only.
a) Find an explicit solution of (9.49).
b) Investigate the a.s. convergence of as t ! C1 when
1 2Ct
.1/ .t/ D b.t/ D
1Ct 2.1 C t/2
1 1
.2/ .t/ D b.t/ D
1Ct 3.1 C t/2
1 1
.3/ .t/ D p b.t/ D
1Ct 2.1 C t/
9.6 (p. 562) Let b 2 Rm , an m d matrix, B a d-dimensional Brownian motion.

Let us consider the SDE
X
d
di .t/ D bi i .t/ dt C i .t/ ij dBj .t/; i D 1; : : : ; m
jD1
(9.50)
0 D x :
a) Find an explicit solution. Prove that if xi > 0 then P.i .t/ > 0 for every t > 0/ D
1 and compute EŒi .t/.
b) Prove that the processes t 7! i .t/j .t/ are in M 2 and compute EŒi .t/j .t// and
the covariances Cov.i .t/; j .t//.
9.7 (p. 564) Let B be a two-dimensional Brownian motion and let us consider the
two processes
d1 .t/ D r1 1 .t/ dt C 1 1 .t/ dB1 .t/

d2 .t/ D r2 2 .t/ dt C 2 2 .t/ dB2 .t/
where r1 ; r2 ; 1 ; 2 are real numbers.

p
a) Prove that both processes Xt D 1 .t/2 .t/ and Zt D 1 .t/2 .t/ are solutions of
SDEs to be determined.
b) Answer the same questions as in a) if we had
d1 .t/ D r1 1 .t/ dt C 1 1 .t/ dB1 .t/

p
d2 .t/ D r2 2 .t/ dt C 2 1 2 2 .t/ dB2 .t/ C 2 2 .t/ dB1 .t/ ;
where 1 1.
9.8 (p. 565) Let be the solution (geometric Brownian motion) of the SDE
dt D bt dt C t dBt

0 D 1 :
a) Determine a real number ˛ such that .t˛ /t is a martingale.

b) Let be the exit time of out of the interval 12 ; 2Œ. Compute P. D 2/.
9.9 (p. 566) Let B be a two-dimensional Brownian motion and, for 2 R, let us
consider the SDE
dt D t dB1 .t/

dt D t dt C t dB2 .t/
0 D x; 0 D y > 0 :
a) Find an explicit expression for .t /t and .t /t . Is .t /t a martingale?

b) Compute E.t / and Var.t /.
c) Prove that, if < 12 , the limit
lim t
t!0C
exists a.s. and in L2 .
9.10 (p. 567) Let us consider the one-dimensional SDE
dt D .a C bt / dt C . C t / dBt

(9.51)
0 D x :
a) Let v.t/ D EŒt . Show that v satisfies an ordinary differential equation and
compute EŒt .
b1) Prove that if b < 0, then limt!C1 EŒt D ab .
b2) Prove that if x D ab then the expectation is constant and EŒt x whatever
the value of b.
b3) Assume b > 0. Prove that
(
1 if x0 < ab
lim EŒt D
t!C1 C1 if x0 > ab
dt D .a C bt / dt C . C t / dBt

(9.52)
0 D x :
Recall (Example 9.2) that if a D D 0 then a solution is given by 0 .t/x, where

2
0 .t/ D e.b 2 /tC Bt
.
a1) Determine a process C such that t 7! 0 .t/Ct is a solution of (9.52) and produce
explicitly the solution of (9.52).
a2) Prove that if D 0 then is a Gaussian process and compute EŒt , Var.t /.
a3) Prove that, if D 0 and b < 0, converges in law as t ! C1 to a limit
distribution to be determined.
b) We want to investigate the equation
dYt D .b C
log Yt /Yt dt C Yt dBt
(9.53)
Y0 D y > 0 :
Show that t D log Yt is the solution of an SDE to be determined. Show

that (9.53) has a unique solution such that Yt > 0 for every t a.s. What is
the law of Yt ? Prove that, if
< 0, .Yt /t converges in law as t ! C1 and
determine the limit distribution.
9.12 (p. 570) Let .˝; F ; .Ft /t ; .Bt /t ; P/ be a one-dimensional Brownian motion
and let be the Ornstein–Uhlenbeck process that is the solution of the SDE
dt D t dt C dBt
0 D x ;
where 6D 0.
a) Show that
2
Zt D e2 t t2 C
2
2
b) Let Yt D t2 C 2 . What is the value of E.Yt /? And of limt!1 E.Yt /? (It will be
useful to distinguish the cases > 0 and < 0.)

q 1 q
dt D 1 C t2 C t dt C 1 C t2 dBt
2 (9.54)
0 D x 2 R :
a) Does this equation admit strong solutions?

b1) Let
q
Yt D log 1 C t2 C t :
What is the stochastic differential of Y?

b2) Deduce an explicit solution of (9.54).
p
Hint: z 7! log 1 C z2 C z is the inverse function of y 7! sinh y.
9.14 (p. 572) Let .t ; t / be the solution of the SDE
dt D t dt C dB1 .t/

p
dt D t dt C dB1 .t/ C 1 2 dB2 .t/
with the initial conditions .0 ; 0 / D .x; y/, where B D .B1 ; B2 / is a two-dimensio-
nal Brownian motion and 1 1.
a) Compute the laws of t and of t and explicitly describe their dependence on the
parameter .
b) Compute the joint law of .t ; t /. What is the value of Cov.t ; t /? For which
values of is this covariance maximum? For which values of does the law of
the pair .t ; t / have a density with respect to Lebesgue measure?
c) What is the differential generator of the diffusion .t ; t /?
9.15 (p. 573) Let B be a Brownian motion. Let t D .1 .t/; 2 .t// be the diffusion
process that is the solution of the SDE
d1 .t/ D 12 1 .t/ dt 2 .t/ dBt

d2 .t/ D 12 2 .t/ dt C 1 .t/ dBt
with the initial condition 1 .0/ D 0; 2 .0/ D 1.

a) What is the generator of .t /t ? Is it elliptic? Uniformly elliptic?
b) Let Yt D 1 .t/2 C 2 .t/2 . Show that .Yt /t satisfies an SDE and determine it. What
can be said of the law of Y1 ?
9.16 (p. 574) Let 2 R, > 0 and

Z t
2
Xt D x C e. 2 /uC Bu
du
0
2
Yt D e. 2 /tC Bt
:
a) Prove that the two-dimensional process Zt D .Xt ; Yt / is a diffusion and compute

its generator.
b) Prove that
Xt
t D
Yt
is a (one-dimensional) diffusion which is the solution of an SDE to be deter-

mined.
c) Let r 2 R, > 0. Find an explicit solution of the SDE
dZt D .1 C rZt / dt C Zt dBt

(9.55)
Z0 D z 2 R
and compare with the one obtained in Exercise 9.11. Determine the value of
EŒZt and its limit as t ! C1 according to the values of r; ; z.
9.17 (p. 576) (Do Exercise 9.11 first) Let us consider the SDE
dt D t .a bt / dt C t dBt

(9.56)
0 D x > 0 ;
where b > 0. Note that the conditions for the existence of solutions are not satisfied,
as the drift does not have a sublinear growth at infinity.
a) Compute, at least formally, the stochastic differential of Zt D 1t and determine
an SDE satisfied by Z.
b) Write down the solution of the SDE for Z (using Exercise 9.11) and derive a
solution of (9.56). Prove that t > 0 for every t 0.
9.18 (p. 577) Let be the solution, if it exists, of the equation

Z t
dt D s ds dt C dBt
0
0 D x ;
where ; x 2 R, > 0. This is not an SDE, as the drift depends not only on the
value of the position of at time t but also on its entire past behavior.
a) Prove that if
Z t
t D s ds
0
then the two-dimensional process Zt D .t ; t / is a diffusion. Write down its

generator and prove that is a Gaussian process.
b) Prove (or take as granted) that if

0a
MD
b0
then
p pa p !
qcosh ab sinh ab
e D M
b
p b
p
a
sinh ab cosh ab
which of course means, if ab < 0,

p p p !
qcos ab ab sin ab
e DM p p : (9.57)
ba sin ab cos ab
c) Compute the mean and variance of the distribution of t and determine their
behavior as t ! C1, according to the values of ; .
9.19 (p. 579) (Have a look at Example 8.9 first) Let B be a Brownian motion and
the Ornstein–Uhlenbeck process solution of the SDE
dt D bt dt C dBt

0 D x ;
where b; 2 R. Let t D t2 .

Prove that is the solution of an SDE to be determined.
9.20 (p. 580) Let W R ! R be a function satisfying hypothesis (A’). Moreover,
let us assume that .x/ c > 0 for every x 2 R and let be the solution of the SDE
dt D .t / dBt ; 0 D 0 :
Let I D Œa; b, where a; b > 0, and let be the exit time from I. Show that
< C1 a.s. and compute P. D b/.
9.21 (p. 581) Let a; b > 0 and let be the Ornstein–Uhlenbeck process that is the
solution of
dt D t dt C dBt
0 D x :
a) Prove that the exit time, , of from a; bŒ is finite a.s. for every starting
position x and give an expression of P. D b/ as a function of ; ; a; b.
b1) Prove that if a > b then
Z a
2
e z dz
lim Z0 b D C1 : (9.58)
!C1 2
e z dz
0
b2) Assume a > b. Prove that
lim Px . D b/ D 1
!C1
whatever the starting point x 2 a; bŒ.

b3) In b2) it appears that, even if x is close to a, the exit from a; bŒ takes place
mostly at b if is large. Would you be able to give an intuitive explanation of
this phenomenon?
9.22 (p. 582) Let " be the Ornstein–Uhlenbeck process solution of the SDE
dt" D t" dt C " dBt

0" D x ;
where 2 R, > 0.
a) Prove that, for every t 0
L
t" ! e t x :
"!0
b) Prove that the laws of the processes " (which, remember, are probabilities on the
space C D C .Œ0; T; R/) converge in distribution to the Dirac mass concentrated
on the path x0 .t/ D e t x that is the solution of the ordinary equation
Pt D t
(9.59)
0 D x :
In other words, the diffusion " can be seen as a small random perturbation of
the ODE (9.59). See Exercise 9.28, where the general situation is considered.
9.23 (p. 583) (Example of explosion). Let us consider the SDE, in dimension 1,
dt D t3 dt C t2 dBt

(9.60)
0 D x
1
and let x D infftI Bt D x
g for x 6D 0, 0 D C1. Prove that, for t 2 Œ0; x Œ, a
solution is given by
x
t D
1 xBt
9.24 (p. 584) Let L be the operator in dimension 2
1 @2 @
LD 2
Cx (9.61)
2 @x @y
and the diffusion having L as its generator.

a) Is L elliptic? Uniformly elliptic? Compute the law of t with the starting
condition 0 D x. Does it have a density with respect to the Lebesgue measure
of R2 ?
b) Answer the same questions as in a) for the operator
1 @2 @
L2 D Cy (9.62)
2 @x2 @y
9.25 (p. 585) (Continuation of Exercise 9.24) Let us consider the diffusion
associated to the generator
1X X
m m
@2 @
LD aij C bij xj , (9.63)
2 i;jD1 @xi @xj i;jD1 @xi
where a and b are constant m m matrices.

a) Compute the transition function and show that, if the matrix a is positive definite,
then the transition function has a density with respect to Lebesgue measure.
b) Show that for the operator L in (9.63) the transition function has a density if and
only if the kernel of a does not contain subspaces different from f0g that are
invariant for the matrix b .
9.26 (p. 586) Let be the solution of the SDE in dimension 1
dt D b.t / dt C .t / dBt

0 D x :
Assume that b is locally Lipschitz continuous, is C1 and such that .x/ ı > 0
for some ı and that b and have a sublinear growth at infinity.
a) Prove that there exists a strictly increasing function f W R ! R such that the
process t D f .t / satisfies the SDE
dt D e
b.t / dt C dBt (9.64)
for some new drift e

b.
b) Prove that, under the assumptions above, the filtration Ht D .u ; u t/
coincides with the natural filtration of the Brownian motion.
9.27 (p. 587) (Doss 1977; Sussmann 1978). In this exercise we shall see that,
in dimension 1, the solution of an SDE can be obtained by solving an ordinary
differential equation with random coefficients. It is a method that can allow us to
find solutions explicitly and to deduce useful properties of the solutions.
Let us assume m D 1. Let b and be Lipschitz continuous functions R ! R

and let us assume that is twice differentiable and strictly positive. Let us denote
by h.x; y/ the solution of
u0 .y/ D .u.y//
(9.65)
u.0/ D x :
Let t 7! Dt .!/ be the solution of the ODE with random coefficients
D0t D f .Dt ; Bt .!//; D0 D x ;
where
h i Z
1 z
f .x; z/ D . 0 /.h.x; z// C b.h.x; z// exp 0 .h.x; s// ds :
2 0
a) Prove that t D h.Dt ; Bt / is the solution of

(9.66)
0 D x :
b) Let e
b be a Lipschitz continuous function such that e
b.x/ b.x/ for every x 2 R
and let e
be the solution of
de
t D e
b.t / dt C .t / dBt
e (9.67)
0 D e
x:
with ex x. Note that the two SDEs (9.66) and (9.67) have the same diffusion
coefficient and are with respect to the same Brownian motion. Show that e
t t
for every t 0 a.s.
a) Thanks to differentiability results for the solutions of an ODE with respect to a parameter, h is
twice differentiable in every variable. Giving this fact as granted, prove that
h Z i
@h y
@2 h
.x; y/ D exp 0 .h.x; s// ds ; .x; y/ D 0 .h.x; y// .h.x; y// (9.68)
@x 0 @y2
and apply Ito’s formula to t 7! h.Dt ; Bt /.

b) Recall the following comparison theorem for ODEs: if g1 ; g2 are Lipschitz continuous functions
R RC ! R such that g2 .x; t/ g1 .x; t/ for every x; t and i ; i D 1; 2 are solutions of the
ODEs
i0 .t/ D gi .i .t/; t/

i .s/ D xi
with x2 x1 , then 2 .t/ 1 .t/ for every t 0.

9.28 (p. 588) Let " be the solution of the SDE, in dimension m,
dt" D b.t" / dt C ".t" / dBt

0" D x ;
where b and are Lipschitz continuous with Lipschitz constant L and is bounded.
Intuitively, if " is small, this equation can be seen as a small random perturbation of
the ordinary equation
t0 D b.t /
(9.69)
0 D x :
In this exercise we see that " converges in probability as " ! 0 to the path that
is the solution of (9.69) (and with an estimate of the speed of convergence).
a) If "t D t" t , show that " is a solution of
d"t D Œb."t C t / b.t / dt C "."t C t / dBt

"0 D 0 :
b) Prove that, for T > 0; ˛ > 0,

n ˇ Z s ˇ o n o
ˇ ˇ
sup ˇ" ."u C u / dBu ˇ < ˛ sup j"s j < ˛eLT
0sT 0 0sT
and therefore
h 1 ˛ 2 e2LT i
P sup js" s j > ˛ 2m exp 2 :
0sT " 2mTkk21
Deduce that " W ˝ ! C .Œ0; T; Rm / converges in probability (and therefore in

law) as " ! 0 to the r.v. concentrated on the path a.s.
c) Are you able to weaken the hypothesis and show that the same result holds if b
and satisfy only Assumption (A’)?
b) Use Gronwall’s inequality and the exponential inequality (8.42). c) Localization. . .
9.29 (p. 589) We have seen in Sect. 9.3 some Lp estimates for the solutions of a
SDE, under Assumption (A’). In this exercise we go deeper into the investigation of
the tail of the law of the solutions. Let B be a d-dimensional Brownian motion and,
for a given process Z, let us denote sup0st jZs j by Zt .
a) Let X be the m-dimensional process
Z t Z t
Xt D x C Fs ds C Gs dBs
0 0
where F and G are processes in M 1 .Œ0; T/ and M 2 .Œ0; T/, respectively, that
are m- and m d-dimensional, respectively; let us assume, moreover, that G is
bounded and that the inequality
jFt j M.1 C jXt j/
holds (which is satisfied, in particular, if F is bounded). Then for every T > 0,

K > 0, there exist c D cT > 0 and R0 > 0 such that, for R > R0 and for every x
such that jxj K,
2
P.XT > R/ ecR : (9.70)
b) Let us assume that b and satisfy Assumption (A’) and that, moreover, the md
matrix is bounded. Let be the solution of

u D x :
Then for every T > 0 there exists a constant c D cT > 0 such that for large R
2
P.T > R/ ecR
uniformly for x in a compact set.

c) Let us assume again that b and satisfy Assumption (A’) and drop the
assumption of boundedness for . Show that Yt D log.1 C jt j2 / is an Ito process
and compute its stochastic differential. Deduce that, for every T > 0, there exists
a constant c D cT > 0 such that for large R
1
P.T > R/
Rc log R
uniformly for x in a compact set.
d) Show that T 2 Lp for every p (this we already know thanks to Theorem 9.1).

Moreover, if is bounded, all exponential moments T are finite (i.e. e T is
integrable for every 2 R) and even
2
EŒe˛T < C1
for ˛ < c for some constant c > 0.

Rt
a) Use the exponential inequality, Proposition 8.7, in order to estimate P.sup0tT j 0 Gs dBs j >
/ and then Gronwall’s inequality. d) Recall the formula of Exercise 1.3.
Chapter 10
PDE Problems and Diffusions
In this chapter we see that the solutions of some PDE problems can be represented
as expectations of functionals of diffusion process. These formulas are very useful
from two points of view. First of all, for the investigation and a better understanding
of the properties of the solutions of these PDEs. Moreover, in some situations, they
allow to compute the solution of the PDE (through the explicit computation of the
expectation of the corresponding functional) or the expectation of the functional (by
solving the PDE explicitly). The exercises of this chapter and Exercise 12.8 provide
some instances of this way of reasoning.
10.1 Representation of the solutions of a PDE problem
Let D be a domain of Rm and u a solution of the Dirichlet problem

(
Lu D 0 on D
(10.1)
uj@D D ;
where L is a second-order differential operator and a function @D ! R. Without

bothering about regularity assumptions, let B be a Brownian motion and x the
diffusion associated to L and starting at x 2 D obtained as seen in the previous
chapter by solving an SDE with respect to the Brownian motion B. Then by Ito’s
formula applied to the function u (that we assume to exist and to be regular enough)
we have
Z t Z t
u.tx / D u.x/ C Lu.sx / ds C u0 .sx /.Xs / dBs ;
0 0

306 10 PDE Problems and Diffusions
u0 denoting the gradient of u (recall Remark 9.1). If denotes the exit time of x
from D, then
Z t^ Z t^
x
u.t^ / D u.x/ C Lu.sx / ds C u0 .sx /.sx / dBs
0 „ƒ‚… 0
D0
and if the process t 7! u0 .tx / belongs to M 2 , taking the expectation we find

x
EŒu.t^ / D u.x/ :
As x 2 @D and u D on @D, taking the limit as t ! 1 we find finally, if is

finite and u bounded,
u.x/ D EŒu.x / D EŒ.x / ; (10.2)
i.e. the value of u at x is equal to the mean of the boundary condition taken with
respect to the exit distribution from the domain D of the diffusion associated to L and
starting at x. Note that, denoting by .C ; M ; .Mt /t0 ; .Xt /t0 ; .Px /x / the canonical
diffusion associated to L, (10.2) can also be written as
u.x/ D Ex Œ.X / : (10.3)
The idea is therefore really simple. There are, however, a few things to check in
order to put together a rigorous proof. More precisely:
a) We need conditions guaranteeing that the exit time from D is a.s. finite. We shall
also need to know that it is integrable.
b) It is not allowed to apply Ito’s formula to a function u that is defined on D only
and not on the whole of Rd . Moreover, it is not clear that the solution u can be
extended to the whole of Rd in a C2 way. Actually, even the boundary condition
itself might not be extendable to a C2 function.
c) We must show that t 7! u0 .tx / belongs to M 2 .
In the next sections we deal rigorously with these questions and we shall find similar
representation formulas for the solutions of other PDE problems.
Note that, as stated above, the previous argument leads to the representation of
the solutions of problem (10.1), once we know already that this problem actually
has a solution. The PDE theory is actually well developed and provides many
existence results for these problems. Once the representation (10.2) is obtained we
can, however, work the argument backwards: in a situation where the existence of
a solution is not known we can consider the function defined in (10.2) and try to
show that it is a solution. In other words, the representation formulas can serve as
a starting point in order to prove the existence in situations in where this is not
guaranteed a priori. This is what we shall do in Sects. 10.4 and 10.6.
10.2 The Dirichlet problem 307
10.2 The Dirichlet problem
We need a preliminary result. Let be the solution of the SDE

Z t Z t
t D x C b.u ; u/ du C .u ; u/ dBu ;
s s
where b and satisfy Assumption (A’). Let D be a bounded open set containing x.
We see now that, under suitable hypotheses, the exit time from D, D infftI t
s; t .!/ … Dg, is finite and actually integrable. Let, as usual,
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ ,
2 i;jD1 @xi @xj iD1
@x i
where a D . Let us consider the following assumption.
Assumption H0 . There exists a C2;1 .Rm RC / function ˚, positive on DRC

and such that, on D RC ,
@˚
C Lt ˚ 1:
@t
Under Assumption H0 the exit time from D is finite and integrable.
Proposition 10.1 Under Assumption H0 and if D is bounded, E./ < C1

for every .x; s/ 2 D RC .
Proof By Ito’s formula, for every t s we have

@˚ X @˚
d˚.t ; t/ D C Lt ˚ .t ; t/ dt C .t ; t/ij .t ; t/ dBj .t/
@t i;j
@xi
hence for s < t
˚.t^ ; t ^ / ˚.x; s/
Z @˚ Z t^ X
t^
@˚
D C Lu ˚ .u ; u/ du C .u ; u/ij .u ; u/ dBj .u/ :
s @u s i;j
@xi
As and the first derivatives of ˚ are bounded on D Œs; t, the last term is a
martingale and has mean 0; hence for every t > s,
Z t^ @˚
EŒ˚.t^ ; t ^ / ˚.x; s/ D E C Lu ˚ .u ; u/ du E.t ^ / C s :
s @u
As ˚ is positive on D RC , we obtain ˚.x; s/ C s E.t ^ /. Now just take the

limit as t ! C1 and apply Fatou’s lemma. t
u
The following two assumptions, which are easier to check, both imply that
Assumption H0 is satisfied.
Assumption H1 . There exist two constants > 0, c > 0 and an index i,

1 i m, such that
aii .x; t/ ; bi .x; t/ c
for every x 2 D, t 0.
Assumption H2 . On D RC b is bounded and the matrix field a is uniformly

elliptic, i.e. such that ha.x; t/z; zi jzj2 for some > 0 and for every
.x; t/ 2 D RC , z 2 Rm .
Proposition 10.2 H2 ) H1 ) H0 .
Proof H2 ) H1 : we have aii .x; t/ D ha.x; t/ ei ; ei i , where ei denotes the

unitary vector pointing in the i-th direction.
H1 ) H0 : let ˚.x; t/ D ˇ.e˛R e˛xi /, where R is the radius of a sphere
containing D, and ˛; ˇ are positive numbers to be specified later (˚ therefore does
not depend on t); then ˚ > 0 on D and
@˚
C Lt ˚ D ˇe˛xi . 12 ˛ 2 aii .x; t/ C ˛bi .x; t// ˇe˛xi . 12 ˛ 2 ˛c/
@t
˛ˇe˛R . 12 ˛ c/
and we can make the last term 1 by first choosing ˛ so that 12 ˛ c > 0 and
then ˇ large. t
u
In particular, the exit time from a bounded open set D is integrable if the
generator of the diffusion is elliptic.
Let us assume that the open bounded set D is connected and with a C2
boundary; let
1X X
m m
@2 @
LD aij .x/ C bi .x/
2 i;jD1 @xi @xj iD1
@xi
be a differential operator on D such that

a) L is uniformly elliptic on D, i.e. ha.x/ z; zi jzj2 for some > 0 and for
every x 2 D, z 2 Rm ;
b) a and b are Lipschitz continuous on D.
The following existence and uniqueness theorem is classical and well-known (see
Friedman 1964, 1975).
Theorem 10.1 Let be a continuous function on @D and c and f Hölder

continuous functions D ! R with c 0. Then, if L satisfies conditions a) and
b) above, there exists a unique function u 2 C2 .D/ \ C.D/ such that
(
Lu cu D f on D
(10.4)
uj@D D :
We prove now that the solution u of (10.4) can be represented as a functional of

a suitable diffusion process, in particular giving a rigorous proof of (10.2).
By Proposition 9.2 there exists an m m matrix field , Lipschitz continuous
on D, such that a D . As and b are assumed to be Lipschitz continuous
on D, they can be extended to the whole of Rm with Assumption (A) satisfied.
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian motion and x the
solution of the SDE
Z t Z t
dt D x C
x
b.s / ds C
x
.sx / dBs :
0 0
Rt
Moreover, let Zt D e 0 c.s / ds , so that dZt D c.tx /Zt dt. Let, for " > 0, D" be a
x
regular open set such that D" D and dist.@D" ; @D/ "; let " be the exit time
from D" : obviously " and, moreover, < C1 by Propositions 10.1 and 10.2;
as the paths are continuous, " % a.s. as " & 0. Let u" be a bounded C2 .Rm /
function coinciding with u on D" . Then Ito’s formula gives
dZt u" .tx / D u" .tx / dZt C Zt du" .tx /

X m
@u" x
D c.tx /Zt u" .tx / dt C Zt Lu" .tx / dt C .s /ij .sx / dBj .s/ :
i;jD1
@xi
Therefore, P-a.s.,
Z t^"

Zt^" u" .t^
x
"
/ D u " .x/ C Lu" .sx / c.sx /u" .sx / Zs ds
Z t^0" X m (10.5)
@u" x
C .s /Zs ij .sx / dBj .s/ :
0 i;jD1
@x i
As the derivatives of u on D" are bounded, the stochastic integral above has 0 mean
and as u coincides with u" on D" we get, taking the expectation for x 2 D" ,
Z t^"
x
EŒu.t^"
/Zt^" D u.x/ C E f .sx /Zs ds :
0
Now, taking the limit as " & 0,

hZ t^ i
x
EŒu.t^ /Zt^ D u.x/ C E Zs f .sx / ds :
0
We now take the limit as t ! C1: we have EŒu.t^ x

/Zt^ ! EŒu.x /Z as u
is bounded (recall that 0 Zt 1). As for the right-hand side, since Zs 1 for
every s,
ˇZ t^ ˇ
ˇ ˇ
ˇ Zs f .sx / dsˇ k f k1 :
0
As is integrable by Propositions 10.1 and 10.2, we can apply Lebesgue’s theorem,

which gives
hZ i
EŒu.x /Z D u.x/ C E Zs f .sx / ds :
0
Finally, as x 2 @D a.s., u.x / D .x / and we obtain

hZ i
u.x/ D EŒ.x /Z E Zs f .sx / ds :
0
This formula can also be expressed in terms of the canonical realization:
Theorem 10.2 Under the assumptions of Theorem 10.1, the solution of the
PDE problem (10.4) is given by
hZ i
u.x/ D Ex Œ.X /Z Ex Zs f .Xs / ds ; (10.6)
0
where .C ; M ; .Mt /t0 ; .Xt /t0 ; .Px /x /R is the canonical diffusion associated to
t
the infinitesimal generator L, Zt D e 0 c.Xs / ds and D infftI Xt 62 Dg.
In the case c D 0, f D 0, (10.6) becomes simply
u.x/ D Ex Œ.X / ; (10.7)
i.e. the value at x of the solution u is the mean of the boundary value taken with
respect to the law of X , which is the exit distribution of the diffusion starting at x.
As remarked at the beginning of this chapter, formulas such as (10.6) or (10.7)
are interesting in two ways:
• for the computation of the means of functionals of diffusion processes, such as
those appearing on the right-hand sides in (10.6) and (10.7), tracing them back
to the computation of the solution of the problem (10.4),
• in order to obtain information about the solution u of (10.4). For instance, let
us remark that in the derivation of (10.6) we only used the existence of u.
Our computation therefore provides a proof of the uniqueness of the solution
in Theorem 10.1.
The Poisson kernel of the operator L on D is a family .˘.x; //x2D of measures
on @D such that, for every continuous function on @D, the solution of
(
Lu D 0 on D
(10.8)
uj@D D
is given by
Z
u.x/ D .y/ ˘.x; dy/ :
@D
(10.7) states that, under the hypotheses of Theorem 10.1, the Poisson kernel always
exists and ˘.x; / is the law of X with respect to Px (i.e. the exit distribution from D
of the diffusion X starting at x 2 D). This identification allows us to determine the
exit distribution in those cases in which the Poisson kernel is known (see the next
example and Exercise 10.4, for example).
Example 10.1 Let L D 12 4 and BR be the open sphere centered at 0 with

radius R in Rm , m 1; let .dy/ be the surface element of @BR and !m D
2 m=2
.m=2/ the measure of the surface of the sphere of radius 1 in R . Then it is
m
known that, for y 2 @BR ; x 2 BR , if
1 R2 jxj2 ,
NR .x; y/ D
R!m jx yjm
then the Poisson kernel of 12 4 on BR is given by
˘.x; dy/ D NR .x; y/ .dy/ : (10.9)
Therefore (10.9) gives the exit distribution from BR of a Brownian motion

starting at a point x 2 BR . If x Dthe origin we find the uniform distribution on
@BR , as we already know from Exercise 3.18.
Example 10.2 Let .Xt /t be a one-dimensional diffusion process associated to

some SDE with Lipschitz continuous coefficients b, . Let Œa; b be some
finite interval and x 2 a; bŒ and assume that > ˛ > 0 on Œa; b. We know
from Propositions 10.1 and 10.2 that the exit time, , from Œa; b is a.s. finite.
We want to compute the probability
P.X D a/ :
We can use the representation formula (10.7) with D D a; bŒ and given by
.a/ D 1, .b/ D 0 (the boundary @D here is reduced to fa; bg). By 10.7
therefore
u.x/ D Px .X D a/ D Ex Œ.X /
is the solution of the ordinary problem

8
< 2 00
bu0 C u D0 on a; bŒ ;
2
:
u.a/ D 1; u.b/ D 0 :
(continued )

1
Let us compute the solution: if v D u0 the ODE becomes bv C 2
2 v 0 D 0,
i.e.
v0 2b
D 2 ,
v
Rx 2b
which gives log v.x/ D a 2 .y/ dy C c1 i.e.
Rx2b.y/
dy
v.x/ D c1 e a 2 .y/
and therefore
Z x Rz 2b.y/
dy
u.x/ D c2 C c1 e a 2 .y/
dz :
a
The condition u.a/ D 1 gives at once c2 D 1, from which also c1 is easily

computed giving
Z x R z 2b.y/
dy
e a 2 .y/ dz
u.x/ D 1 Z b R z 2b.y/
a

a 2 dy
e .y/ dz
a
2
It is useful to note that to solve the equation bu0 C 2 u00 D 0 is equivalent
to finding a function u such that .u.Xt /t /t is a martingale. We know already
that problems of determining the exit distribution are often solved using an
appropriate martingale.
In the following examples, conversely, formulas (10.6) and (10.7) are used in order
to prove properties of the solutions of the Dirichlet problem (10.4).
Remark 10.1 (The maximum principle) Let D be a bounded domain and
1X X
m m
@2 @
LD aij .x/ C bi .x/
2 i;jD1 @xi @xj iD1
@x i
a uniformly elliptic differential operator on D; let us assume a and b are

Lipschitz continuous on D. The maximum principle states that for a function
u which is in C2 .D/ \ C.D/ and such that Lu D 0, we have
min u u.x/ max u :

@D @D
(continued )

This is now an immediate consequence of (10.7), as it is immediate that
min u Ex Œu.X / max u :

@D @D
If u satisfies the weaker condition Lu D f 0 (but still c D 0) then it enjoys

a weaker form of the maximum principle: starting from (10.6) we have
hZ i
u.x/ D Ex Œu.X / Ex f .Xs / ds Ex Œu.X / max u ;
0 @D
(c D 0 implies Zt 1). Recall that we have already come across the

maximum principle in Sect. 6.3 (Proposition 6.5).
Remark 10.2 Let D be a bounded domain and u W D ! R a harmonic

function, i.e. such that 12 4u D 0 on D. Then u enjoys the mean property:
if we denote by BR .x/ the ball of radius R and by the normalized Lebesgue
measure of the surface @BR .x/, then
Z
u.x/ D u.y/ d.y/ (10.10)
@BR .x/
for every x 2 D and R > 0 such that BR .x/ D.

Indeed, from (10.7), applied to D D BR .x/, we have u.x/ D Ex Œu.XR /,
where R denotes the exit time from BR .x/ and we know that the exit distri-
bution of a Brownian motion from a ball centered at its starting point is the
normalized Lebesgue measure of the surface of the ball (see Exercise 3.18).
Actually the mean property (10.10) characterizes the harmonic functions
(see also Example 6.3)
Formula (10.6) also gives a numerical approach to the computation of the solution
of a PDE problem as we shall see in Sect. 11.4.
10.3 Parabolic equations
A representation formula can also be obtained in quite a similar way for the Cauchy–
Dirichlet problem.
10.3 Parabolic equations 315
Let D Rm be a bounded connected open set having a C2 boundary and let

Q D D Œ0; TŒ and
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/ (10.11)
2 i;jD1 @xi @xj iD1
@xi
be a differential operator on Q such that

a) ha.x; t/ z; zi jzj2 , where > 0 for every .x; t/ 2 Q, z 2 Rm ;
b) a and b are Lipschitz continuous in Q.
Then (Friedman 1964, 1975) the following existence and uniqueness result holds.
Note that assumption b) above requires that the coefficients are also Lipschitz
continuous in the time variable.
Theorem 10.3 Let be a continuous function on D and g a continuous

function on @D Œ0; T such that g.x; T/ D .x/ if x 2 @D; let c and f be
Hölder continuous functions Q ! R. Then, under the previous hypotheses a)
and b), there exists a unique function u 2 C2;1 .Q/ \ C.Q/ such that (Fig. 10.1)
8 @u
ˆ
ˆ L u cu C Df on Q
< t @t
(10.12)
û.x; T/ D .x/ on D
:̂
u.x; t/ D g.x; t/ on @D Œ0; T :
By Proposition 9.2 there exists a Lipschitz continuous matrix field on Q

such that D a; moreover, the coefficients and b can be extended to
g(x, t)
⎧
⎪
⎪
•
⎪
⎪
⎪
⎪
⎪
⎪
⎨
D ← (x)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
•
0 T
g(x, t)
Fig. 10.1 In the problem (10.12) the boundary values are given on the boundary of D (i.e. on
@D Œ0; T) and at the final time T (i.e. on D fTg). Note that the hypotheses of Theorem 10.3
require that on @D fTg, where these boundary conditions “meet”, they must coincide
Rm Œ0; T preserving Lipschitz continuity and boundedness. Hence b and satisfy

Assumption (A) (see p. 260). Actually they are even Lipschitz continuous in the
time variable, which is not required in Assumption (A).
Theorem 10.4 Let .C ; M ; .Mts /ts ; .Xt /t ; .P x;s /x;s / be the canonical diffu-
sion associated to the SDE with coefficients b and . Then under the
hypotheses of Theorem 10.3 we have the representation formula
h R i
u.x; t/ D Ex;t g.X ; / e t c.Xs ;s/ ds 1f <Tg
h RT i
C Ex;t .XT / e t c.Xs ;s/ ds 1f Tg (10.13)
h Z ^T Rs i
Ex;t f .Xs ; s/ e t c.Xu ;u/ du ds ;
t
where is the exit time from D.
Proof The argument is similar to the one in

R s the proof of (10.6): we must apply Ito’s
formula to the process Ys D u" .Xs ; s/ e t c.Xr ;r/ dr , where u" approximates u in a
way similar to (10.6). t
u
In the particular case g D 0, c D 0, f D 0, (10.13) becomes
u.x; t/ D Ex;t Œ.XT /1f Tg :
Note that this relation follows from Theorem 10.4 if we assume that vanishes
on @D, otherwise the condition g.x; T/ D .x/ if x 2 @D, which is required in
Theorem 10.3, is not satisfied.
Remark 10.3 If (Fig. 10.2) b; ; f ; g; c do not depend on t then if v.x; t/ D

u.x; T t/, v is a solution of
8 @v
ˆ
ˆ Lv cv D f .x/ on Q
< @t
(10.14)
ˆv.x; 0/ D .x/ on D
:̂
v.x; t/ D g.x/ on @D Œ0; T :
For the solution v of (10.14) we have the representation (recall that now the
diffusion associated to L is time homogeneous, so that Ex;t Œ.XT /1f Tg can
be written as Ex Œ.XTt /1f Ttg )
v.x; t/ D u.x; T t/ D Ex Œ.Xt /1f tg :

10.4 The Feynman–Kac formula 317
g(x, t)
⎧
⎪
⎪
•
⎪
⎪
⎪
⎪
⎪
⎪
⎨
D ← (x)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
•
0 T
g(x, t)
Fig. 10.2 In the problem (10.14) the boundary values are given on @D Œ0; T and at time 0, i.e.
it is an “initial value” problem. Note that, with respect to (10.12), the term in @t@ has the sign
10.4 The Feynman–Kac formula
In this section we investigate representation formulas similar to the one of Theo-

rem 10.4, for the domain Rm Œ0; T. We shall see applications of them in the next
chapters. The main difficulties come from the fact that Rm is not bounded.
Let ; f be two real continuous functions defined respectively on Rm and Rm
Œ0; T and let us make one of the assumptions
j.x/j M.1 C jxj /; j f .x; t/j M.1 C jxj / .x; t/ 2 Rm Œ0; T

(10.15)
for some 0 (condition of polynomial growth), or
0; f 0: (10.16)
Let Lt be a differential operator on Rm Œ0; TŒ, as in (10.11), and let us consider its
associated canonical diffusion .C ; M ; .Mts /ts ; .Xt /t ; .P x;s /x;s /.
Theorem 10.5 Let c W Rm Œ0; T ! R be a continuous function bounded

below (i.e. c.x; t/ K > 1) and w a C2;1 .Rm Œ0; TŒ/ function,
continuous on Rm Œ0; T and a solution of the problem
8
<Lt w C @w cw D f on Rm Œ0; TŒ
@t (10.17)
:w.x; T/ D .x/ :
(continued)

Let us assume that the coefficients a and b are Lipschitz continuous on Œ0; T
Rm and that a is uniformly elliptic. Assume also that the functions ; f , in
addition to being continuous, satisfy one of the hypotheses (10.15) or (10.16).
Assume, finally, that w has a polynomial growth, i.e. there exists M1 ; > 0
such that
jw.x; t/j M1 .1 C jxj / (10.18)
for every t 2 Œ0; T, x 2 Rm . Then we have the representation formula

RT
hZ T Rs i
w.x; t/ D Ex;t .XT / e t c.Xs ;s/ ds Ex;t f .Xs ; s/ e t c.Xu ;u/ du ds :
t
(10.19)
Proof Let R be the exit time of X from the sphere of radius R. Then, by
Theorem 10.4 applied to D D BR , we have, for every .x; t/ 2 BR Œ0; TŒ,
h Z T^R Rs i
w.Xt ; t/ D Ex;t f .Xs ; s/ e t c.Xu;u/ du ds
t R R
CEx;t w.XR ; R / e t c.Xu ;u/ du 1fR <Tg (10.20)

RT
t c.Xu ;u/ du

CE x;t
.X T/ e 1fR Tg
D I1 C I2 C I3 :
Let us first assume that (10.15) holds. By Lemma 9.1 we know that
h i
Ex;t max jXu jq Cq .1 C jxjq / (10.21)
tuT
for every q 2, where the constant Cq does not depend on x. Then by the Markov
inequality we have

P x;t .R T/ D P x;t max jXu j R Cq Rq .1 C jxjq / ; (10.22)
tuT
hence P x;t .R T/ ! 0 as R ! C1, faster than every polynomial. The map

R ! R is increasing and R ^ T ! T as R ! C1 for every !. Therefore if
we let R ! C1 in the first term in (10.20) and apply Lebesgue’s theorem, thanks
to (10.21) and to the second relation in (10.15), we find that it converges to
hZ T Rs i
E x;t
f .Xs ; s/ e t c.Xu ;u/ du ds :
t
A similar argument works for I3 : as R ! C1 the r.v. inside the expectation

converges to
RT
.XT / e t c.Xu ;u/ du
;
being bounded by j.XT /j eKT which, thanks to (10.21) and to the first relation
in (10.15), is an integrable r.v. We still have to prove that I2 ! 0 as R ! C1.
In fact
ˇ R ˇ
ˇw.XR ; R / e t R c.Xu ;u/ du 1f Tg ˇ M1 eKT .1 C R /1f Tg
R R
and therefore
RT
Ex;t w.XR ; R / e t c.Xu ;u/ du 1fR >Tg M1 eKT .1 C R / P x;t .R T/
and the left-hand side tends to 0 as R ! C1 as in (10.22) (just choose q > ).
If, conversely, (10.16) holds, then we argue similarly, the only difference being
that now the convergence of I1 and I3 follows from Beppo Levi’s Theorem. t
u
Remark 10.4 The hypothesis of polynomial growth (10.18) can be weakened

if, in addition, is bounded. In this case the process .Xt /t enjoys a stronger
integrability property than (10.21). In Exercise 9.29 we prove that if
sup j.x/ .x/j k ; jb.x/j Mb .1 C jxj/ ;

x2Rm
where j j denotes the operator norm of the matrix, then

h i
Ex;t exp ˛ max jXu j2 K
tuT
for every ˛ < c D e2Mb T .2Tmk /1 . It is easy then to see that, if we
replace (10.18) by
2
jw.x; t/j M1 e˛jxj (10.23)
for some ˛ < c, then we can repeat the proof of Theorem 10.5 and the
representation formula (10.19) still holds. In this case, i.e. if is bounded,
also the hypotheses of polynomial growth for and f can be replaced by
2 2
j.x/j Me˛jxj ; j f .x; t/j Me˛jxj .x; t/ 2 Rm Œ0; T (10.24)
for some ˛ < c D e2MT .2Tmk /1 .
Theorem 10.5 gives an important representation formula for the solutions of the
parabolic problem (10.17). We refer to it as the Feynman–Kac formula.
It would, however, be very useful to have conditions for the existence of a
solution of (10.17) satisfying (10.18). This would allow us, for instance, to state
that the function defined as the right-hand side of (10.19) is a solution of (10.17).
Existence results are available in the literature: see Friedman (1975, p. 147), for
example, where, however, boundedness for the coefficients a and b is required, an
assumption that is often too constraining and which will not be satisfied in some
applications of Chap. 13.
Until now we have given representation formulas for the solutions of PDE
problems as functionals of diffusion processes. Let us conversely try to construct
a solution and therefore prove that a solution exists.
Let us assume c and f to be locally Hölder continuous and let, for .x; t/ 2 Rm
Œ0; T,
h Z i
RT
T Rs
u.x; t/ D Ex;t .XT / e t c.Xs ;s/ ds Ex;t f .Xs ; s/ e t c.Xu ;u/ du
ds ; (10.25)
„ ƒ‚ …
WDu .x;t/
„ t
ƒ‚ …
1
WDu2 .x;t/
where, as above, Lt is a differential operator on Rm Œ0; TŒ, as in (10.11), and

.C ; M ; .Mts /ts ; .Xt /t ; .P x;s /x;s / is the associated canonical diffusion.
We shall now prove that such a function u is continuous on Rm Œ0; T, that it is
2;1
C .Rm Œ0; TŒ/ and a solution of (10.17).
Let BR be the ball centered at the origin and of radius R large enough so that
x 2 BR . The idea is simply to show that, for every R > 0, u satisfies a relation of the
same kind as (10.13), with a function g to be specified on @BR Œ0; TŒ. Theorems 10.3
and 10.4 then allow us to state that u has the required regularity and that it is a
solution of (10.12) on BR Œ0; TŒ. In particular, by the arbitrariness of R, we shall
have
@u
Lt u C cu D f on Rm Œ0; TŒ : (10.26)
@t
The key tool will be the strong Markov property.
Lemma 10.1 Let us assume that

• the coefficients a and b are Lipschitz continuous on Œ0; T Rm and that a
is uniformly elliptic
• that the function c is continuous and bounded below, on Rm Œ0; T, by a
constant K.
Then if is continuous and has polynomial growth, the function u1 defined
in (10.25) is continuous in .x; t/. The same holds for u2 if f is continuous and
has polynomial growth.
If, moreover, the diffusion coefficient is bounded, it is sufficient to assume
that and f have an exponential growth, i.e. that there exist constants k1 ; k2
0 such that j.x/j k1 ek2 jxj , j f .x; t/j k1 ek2 jxj for every 0 t T.
Proof This is an application of the continuity property of a diffusion with respect to

the initial conditions (Sect. 9.8). The idea is simple but, to be rigorous, we need to
be careful. One can write
RT x;t
u1 .x; t/ D E .Tx;t / e t c.u ;u/ du ;
where x;t denotes, as in the previous chapter, the solution of
ds D b.s ; s/ ds C .s ; s/ dBs

t D x
with respect to a Brownian motion .˝; F ; .Ft /t ; .Bt /t ; P/.

By Theorem 9.9 we can assume that .x; t; s/ 7! sx;t .!/ is a continuous function
for every ! 2 ˝. If .xn ; tn / ! .x; t/ as n ! 1, let R be a number large enough so
that x; xn 2 BR .
Let R be a bounded continuous function coinciding with on BR and let us
denote by and n , respectively, the exit times of x;t and xn ;tn from BR . Then
ju1 .xn ; tn / u1 .x; t/j

hˇ RT xn ;tn RT ˇi
E ˇ.Txn ;tn / e tn c.u ;u/ du .Tx;t / e t c.u ;u/ du ˇ
x;t
hˇ RT xn ;tn RT ˇi
E ˇR .Txn ;tn / e tn c.u ;u/ du R .Tx;t / e t c.u ;u/ du ˇ
x;t
(10.27)
hˇ ˇ R T xn ;tn i
CE ˇ.Txn ;tn /ˇ e tn c.u ;u/ du 1fn ^ <Tg
hˇ ˇ R T x;t i
CE ˇ.Tx;t /ˇ e t c.u ;u/ du 1fn ^ <Tg :
Applying Lebesgue’s theorem, we first have hat

RT x ;tn RT x;t
R .Txn ;tn / e tn c.un ;u/ du
! R .Tx;t / e t c.u ;u/ du
n!1
and then (recall that R is bounded and c K) that

ˇ RT xn ;tn RT ˇ
lim E ˇR .Txn ;tn / e tn c.u ;u/ du R .Tx;t / e t c.u ;u/ du ˇ D 0 :
x;t
n!1
It remains to show that the last two terms in (10.27) can be made arbitrarily small
uniformly in n for large R. To this end let us observe that
1fn ^ <Tg 1fn <Tg C 1f <Tg D 1fsuptuT jux;t j>Rg C 1fsupt x ;tn

jun j>Rg :
n uT
Thanks to Hölder’s inequality (1.3) and using the upper bound j.x/j M.1 C jxj /
for some M > 0; > 0, we have for every z, s 0
RT z;s
E j.Tz;s /j e s c.u ;u/ du 1fsupsuT juz;s j>Rg

h i
MeKT E 1 C sup juz;s j 1fsupsuT juz;s j>Rg
suT
h i1=2 1=2
M2 E 1 C sup juz;s j2 P sup juz;s j > R
suT suT
and the last quantity tends to zero as R ! C1, uniformly for z; s in a compact set.
Actually thanks to Proposition 9.1
h i
E 1 C sup juz;s j2 < C1
suT
and thanks to Remark 9.3

P sup juz;s j > R ! 0:
suT R!C1
Hence the last two terms in (10.27) can be made uniformly arbitrarily small for
large R.
The proof of the continuity of u2 follows the same pattern. u
t
Lemma 10.2 Let be a t-stopping time such that T a.s. Then

R
u1 .x; t/ D Ex;t e t c.Xs;s/ ds u1 .X ; / ;

hZ Rs i h R i
u2 .x; t/ D E x;t
f .Xs ; s/ e t c.Xu ;u/ du ds C Ex;t e t c.Xu ;u/ du u2 .X ; /
t
We prove Lemma 10.2 later.

Let R D inffsI s t; Xs 62 BR g be the first exit time from BR after time t. This
is of course a stopping t-time (see the definition p. 160). Let D R ^ T and let us
deal first with u1 .
Lemma 10.2 applied to the t-stopping time gives
u.x; t/ D u1 .x; t/ u2 .x; t/

R i hZ Rs i
D Ex;t e t c.Xs ;s/ ds
u.X ; / Ex;t f .Xs ; s/ e t c.Xu ;u/ du ds
t
R R
RT
D Ex;t u.XR ; R / e t c.Xu ;u/ du 1fR <Tg C Ex;t .XT / e t c.Xs ;s/ ds 1fR Tg
h Z R ^T Rs i
Ex;t f .Xs ; s/ e t c.Xu ;u/ du ds :
t
Comparing with the representation formula of Theorem 10.4 with g D u this allows
us to state that if
• f and c are Hölder continuous and
• the function u is continuous,
then u coincides with the solution of
8
ˆ @w
ˆ
<Lt w C @t cw D f
ˆ on BR Œ0; TŒ
ˆw.x; T/ D .x/
ˆ on BR
:̂w.x; t/ D u.x; t/ on @BR Œ0; T
and, in particular, is of class C2;1 . By the arbitrariness of R, u is of class C2;1 and a

solution of (10.26).
We have proved the following
Theorem 10.6 Let us assume that

• the coefficients a and b of Lt are Lipschitz continuous on Rm Œ0; T and
moreover that for every R > 0 there exists a R > 0 such that ha.x; t/z; zi
R jzj2 for every .x; t/, jxj R; 0 t T, z 2 Rm .
• W Rm ! R is continuous and has polynomial growth.
• the real functions f and c, defined on Rm Œ0; T, are locally Hölder
continuous; c is bounded below, f has polynomial growth.
Then there exists a function u, continuous on Rm Œ0; T and C2;1 .Rm
Œ0; TŒ/, which is a solution of
8
<Lt u C @u cu D f on Rm Œ0; TŒ
@t (10.28)
:u.x; T/ D .x/ :
It is given by
h Z i
RT
T Rs
u.x; t/ D Ex;t .XT / e t c.Xs ;s/ ds Ex;t f .Xs ; s/ e t c.Xv ;v/ dv
ds ;
t
(continued)

where .C ; M ; .Mt /t ; .Xt /t ; .P x;s /x;s / is the canonical diffusion associated to
Lt . Moreover, u is the unique solution with polynomial growth of (10.28).
If, moreover, the diffusion coefficient is bounded, then if and f have an
exponential growth, u is the unique solution with exponential growth.
Proof of Lemma 10.2 The proof consists in a stronger version of the strong Markov
property. We shall first make the assumption that takes a discrete setRof values.

Taking the conditional expectation with respect to Mt , as the r.v. t c.Xs ; s/ ds
is Mt -measurable,
R RT
u1 .x; t/ D Ex;t e t c.Xs ;s/ ds Ex;t .XT / e c.Xs;s/ ds jMt : (10.29)
Let us denote by s1 ; : : : ; sm the possible values of . Then if C 2 Mt we have
h RT i hX
m RT i
E x;t
1C .XT / e c.Xs ;s/ ds
D Ex;t 1C\f Dsk g .XT / e c.Xs ;s/ ds
kD1
hX
m RT i
D Ex;t 1C\f Dsk g .XT / e sk c.Xs ;s/ ds
kD1
hX
m RT i
D Ex;t 1C\f Dsk g Ex;t .XT / e sk c.Xs ;s/ ds jMstk :
kD1
RT
Note now that the r.v. .XT / e sk c.Xs ;s/ ds is M1sk
-measurable. Hence by Proposi-
tion 6.1
RT RT
Ex;t .XT / e sk c.Xs;s/ ds jMstk D EXsk ;sk .XT / e sk c.Xs ;s/ ds P x;t -a.s.
Hence, as
X
m X
m RT
u1 .X ; / D 1f Dsk g u1 .Xsk ; sk / D 1f Dsk g EXsk ;sk .XT / e sk c.Xs ;s/ ds ;
kD1 kD1
we have
h RT i hX
m RT i
Ex;t 1C .XT / e c.Xs;s/ ds D Ex;t 1C\f Dsk g EXsk ;sk .XT / e sk c.Xs ;s/ ds
kD1
hX
m i
D Ex;t 1C\f Dsk g u1 .Xsk ; sk / D Ex;t 1C u1 .X ; /

kD1
and therefore
RT
Ex;t .XT / e c.Xs ;s/ ds jMt D u1 .X ; /
and going back to (10.29) we see that the first equality of Lemma 10.2 is proved
for a discrete stopping time . We obtain the result for a general stopping time
T with the usual argument: let .n /n be a sequence of discrete stopping times
decreasing to . We have proved that
RT
R n
Ex;t .XT / e t c.Xs ;s/ ds jMtn D e t c.Xs;s/ ds u1 .Xn ; n / : (10.30)
Conditioning both sides with respect to Mt , as Mt Mtn ,
RT
R n
Ex;t .XT / e t c.Xs ;s/ ds jMt D Ex;t e t c.Xs ;s/ ds u1 .Xn ; n /jMt :
Now clearly
R n R
e t c.Xs ;s/ ds
u1 .Xn ; n / ! e t c.Xs ;s/ ds
u1 .X ; /
n!1
and thanks to (10.30) the sequence on the left-hand side above is uniformly
integrable so that, P x;t -a.s.,
R n
R
Ex;t e t c.Xs ;s/ ds u1 .Xn ; n /jMt ! Ex;t e t c.Xs ;s/ ds u1 .X ; /jMt
n!1
R
D e t c.Xs ;s/ ds
u1 .X ; /
which proves the first equality of Lemma 10.2. The second one is proved along the
same lines. t
u
Example 10.3 Let us consider the problem

8
< 1 4u.x; t/ C @u .x; t/ jxj2 u.x; t/ D 0 on Rm Œ0; TŒ
2 @t
:
u.x; T/ D 1 :
By Theorem 10.6 a solution is therefore given by

RT 2
u.x; t/ D Ex;t e t jXs j ds

(continued )

(here 1, f 0), where X D .˝; F ; .Fts /ts ; .Xt /t ; .P x;t /x;t / is a
realization of the diffusion associated at generator 12 4, i.e.
R Tt 2
u.x; t/ D E e 0 jBs Cxj ds ; (10.31)
where B is a Brownian motion. This solution is unique among those with a

polynomial growth. In Exercise 12.8 the expectation on the right-hand side
of (10.31) will be computed giving the explicit solution
h p
p m=2 2 jxj2 p i
u.x; t/ D cosh 2 .T t/ exp tanh 2 .T t/ :
2
10.5 The density of the transition function, the backward

equation
In the examples of diffusion processes we have met so far (all of them being Rm -
valued for some m 1), the transition probability p had a density, i.e. there existed
a positive measurable function q.s; t; x; y/ such that
p.s; t; x; dy/ D q.s; t; x; y/ dy ;
dy denoting the Lebesgue measure of Rm . We now look for general conditions

ensuring the existence of a density.
Definition 10.1 A fundamental solution of the Cauchy problem on Rm Œ0; T

associated to Lt is a function .s; t; x; y/ defined for x; y 2 Rm , 0 s < t T,
such that if, for every continuous compactly supported function ,
Z
v.x; s/ D .s; t; x; y/.y/ dy ;
then v is C2;1 .Rm Œ0; tŒ/, is bounded and is a solution of

8
<Ls v C @v D 0 on Rm Œ0; tŒ
@s (10.32)
:lim
s!t v.x; s/ D .x/ :
Let us assume that the operator Lt satisfies the assumptions of Theorem 10.6. If the
transition function p.s; t; x; dy/ has a density q.s; t; x; y/, then this is a fundamental
10.5 The density of the transition function, the backwardequation 327
solution. Actually, by Theorem 10.6 there is a unique bounded solution of (10.32),

given by
Z Z
v.x; s/ D Ex;s Œ.Xt / D .y/ p.s; t; x; dy/ D .y/q.s; t; x; y/ dy :
Therefore D q is a fundamental solution. Conversely, if a fundamental solution

exists, by the same argument, necessarily
Z Z
.y/ p.s; t; x; dy/ D .y/ .s; t; x; y/ dy
for every compactly supported continuous function . This implies that the transi-
tion function has density .
We still have to investigate conditions under which the transition function has a
density, or equivalently, the differential generator has a fundamental solution.
Let us make the following assumptions

• ha.x; t/ z; zi jzj2 for some > 0 and for every z 2 Rm , .x; t/ 2 Rm
Œ0; T.
• a and b are continuous and bounded for .x; t/ 2 Rm Œ0; T. Moreover, a
is continuous in t uniformly in x.
• a and b are Lipschitz continuous in x uniformly for t 2 Œ0; T.
Theorem 10.7 Under the assumptions above there exists a unique funda-
mental solution of the Cauchy problem on Rm Œ0; T associated to Lt .
Moreover, it satisfies the inequalities
h jx yj2 i
j .s; t; x; y/j C.t s/m=2 exp c
ts
ˇ @ ˇ h jx yj2 i
(10.33)
ˇ ˇ .mC1/=2
ˇ .s; t; x; y/ˇ C.t s/ exp c ;
@xi ts
where C and c are positive constants. Furthermore the derivative of with

respect to s and its second derivatives with respect to x exist and are continuous
for t > s. Finally, as a function of .x; s/, s < t, is a solution of
@
Ls C D0: (10.34)
@s
Theorem 10.7 is a classical application of the parametrix method. See Levi (1907)
and also Friedman (1964, Chapter 1), Friedman (1975, Chapter 6) or Chapter 4
of Azencott et al. (1981). Note, however, that Assumption (H) is only a sufficient
condition. In Exercise 9.25 b) the transition density (and therefore the fundamental
solution) is computed explicitly for an operator L whose second-order matrix
coefficient a.x/ is not elliptic for any value of x.
As we have identified the transition density q.s; t; x; y/ and the fundamental
solution, q satisfies, as a function of s; x (the “backward variables”), the backward
equation
@q
Ls q C D0: (10.35)
@s
10.6 Construction of the solutions of the Dirichlet problem
In this section we construct a solution of the Dirichlet problem

8
ˆ
ˆ u 2 C2 .D/ \ C.D/
<
1
4u D 0 on D (10.36)
ˆ2
:̂
uj@D D
without using the existence Theorem 10.1. In particular, we shall point out the
hypotheses that must be must satisfied by the boundary @D in order to have a
solution; the C2 hypothesis for @D will be much weakened.
Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical realization of an m-dimen-
sional Brownian motion, D a connected bounded open set of Rm , the exit time
from D. Let
u.x/ D Ex Œ.X / : (10.37)
Let us recall that in Example 6.3 we have proved the following result.
Proposition 10.3 For every bounded Borel function on @D, u is harmonic

in D.
(10.37) therefore provides a candidate solution of the Dirichlet problem (10.36). We

see now which conditions on and @D are needed for u to be continuous on D and
verify the condition at the boundary. It is clear that, for z 2 @D, u.z/ D .z/ and in
order for u to be continuous at x it will be necessary to show that if x 2 D is a point
near z, then the law of X with respect to Px is concentrated near z. The investigation
10.6 Construction of the solutions of the Dirichlet problem 329
of this type of property is the object of the remainder of this section. It is suggested
at this time to go back and have a look at Example 6.4.
Proposition 10.4 Let g W Rm ! R be a bounded Borel function. Then the

function .x/ D Ex Œg.Xt / is C1 , for every fixed t > 0.
Proposition 10.4 is just a repetition of Remark 6.4.

Let 0 .!/ D infft > 0I Xt … Dg. 0 is not the exit time from D, because of the >
c
instead of . Of course, 0 D if X0 2 D or X0 2 D ; and 0 can be different if
the initial point of the process lies on the boundary @D, as then D 0 (as @D Dc ),
whereas we may have 0 > 0 (if the path enters D immediately).
It is immediate that 0 is a stopping time of the filtration .MtC /t . Indeed
f 0 D 0g D fthere exists a sequence .tn /n of times with tn > 0; tn & 0 and Xtn … Dg
and this is an event of M" for every " > 0. Therefore f 0 D 0g 2 M0C . Moreover,
for u > 0,
f 0 ug D f ug n f D 0g [ f 0 D 0g 2 Mu MuC :
Definition 10.2 x 2 @D is said to be a regular point if Px . 0 D 0/ D 1.
It is immediate that, as Brownian motion is a continuous Feller process, by

Blumenthal’s 0–1 law (Proposition 6.23), Px . 0 D 0/ can take the values 0 or 1
only.
Lemma 10.3 For every u > 0 the function x 7! Px . 0 > u/ is upper

semicontinuous.
Proof Let s .!/ D infft sI Xt … Dg. Then s & 0 as s & 0, therefore Px .s > u/
decreases to Px . 0 > u/ as s & 0. We now just need to prove that x 7! Px .s > u/ is
a continuous function, as the lower envelope of a family of continuous functions is
known to be upper semicontinuous. Using translation operators we can write s D
s C ı
s : in fact ı
s .!/ is the time between time s and the first exit from D of !
after time s. By the strong Markov property, Proposition 6.2,
Px .s > u/ D Px . ı
s > u s/ D Ex Œ1f >usg ı
s

D Ex ŒE.1f >usg ı
s jFs / D Ex PXs . > u s/ D Ex Œg.Xs / ;
where g.x/ D Px . > u sg. Therefore x 7! Px .s > ug is even C1 by

Proposition 10.4. t
u
Lemma 10.4 If z 2 @D is a regular point and u > 0, then
lim
x!z
Px . > u/ D 0 :
x2D
Proof If x 2 D, then Px . > u/ D Px . 0 > u/ and by Lemma 10.3
0 lim
x!z
Px . > u/ D lim
x!z
Px . 0 > u/ Pz . 0 > u/ D 0 :
x2D x2D
t
u
The following lemma states that if z 2 @D is a regular point, then, starting from a
point x 2 D near to z, the Brownian motion exits from D mainly at points near to z.
Lemma 10.5 If z 2 @D is a regular point, then for every " > 0 and for
every neighborhood V of z there exists a neighborhood W of z such that, if
x 2 D \ W, Px .X … V \ @D/ ".
Proof By the Doob and Chebyshev inequalities

˛ 4 h i 16m 16tm
Px sup jXs xj 2 Ex sup jXs xj2 2 Ex ŒjXt xj2 D 2
0st 2 ˛ 0st ˛ ˛
˛
Let ˛ > 0 be such that B˛ .z/ V; then if jx zj < 2
Px .jX zj ˛/ Px . > u/ C Px .jX xj ˛2 ; u/

˛ 16um
Px . > u/ C Px sup jXs xj Px . > u/ C
0su 2 ˛2
16um "
Now just choose u small so that ˛2
< 2 and apply Lemma 10.4. t
u
Proposition 10.5 If z 2 @D is a regular point and W @D ! R is continuous

at z, then the function u.x/ D Ex Œ.X / satisfies lim x!z u.x/ D .z/.
x2D
"
Proof Let > 0 be such that j.y/ .z/j < 2 for every y 2 @D with jy zj < ;
then
ju.x/ .z/j Ex Œj.X / .z/j

Ex j.X / .z/j1fjX zj<g C Ex j.X / .z/j1fjX zjg

"
C 2kk1 Px .jX zj / :
2
Now, by Lemma 10.5, there exists a neighborhood V of z such that if x 2 V then
Px .jX zj / ", which allows us to conclude the proof. t
u
The following statement is now a consequence of Propositions 10.3 and 10.5.
Theorem 10.8 Let D Rm be a bounded open set such that every point in
@D is regular. Then if is a continuous function on @D, the function u.x/ D
Ex Œ.X / is in C2 .D/ \ C.D/ and is a solution of
8
< 1 4u D 0 on D
2
:u D :
j@D
It is now natural to investigate conditions for the regularity of points of @D.
Proposition 10.6 (Cone property) Let z 2 @D and let us assume that there
exist an open cone C with vertex z and a neighborhood V of z such that C\V
Dc . Then z is regular.
Proof Let ˛ > 0 be such that the ball B˛ .z/ of radius ˛ and centered at z is contained
in V; then, as f 0 > tg & f 0 > 0g, we have
Pz . 0 > 0/ D lim Pz . 0 > t/

t!0C
but for every t > 0
Pz . 0 > t/ Pz .Xt … C \ B˛ .z// D 1 Pz .Xt 2 C \ B˛ .z//
so that
Pz . 0 > 0/ 1 lim Pz .Xt 2 C \ B˛ .z// : (10.38)

t!0C
But if C1 D C z, C1 is an open cone with vertex at the origin,

p which is invariant
under multiplication by a positive constant, hence C1 D C1 = t and
p
Pz .Xt 2 C \ B˛ .z// D P0 .Xt 2 C1 \ B˛ / D P0 . t X1 2 C1 \ B˛ /
Z
1 2
D P0 .X1 2 C1 \ B˛=pt / D ejyj =2 dy :
.2/ m=2
C1 \B˛=pt
Taking the limit as t ! 0C in (10.38) we find

Z
1 2 =2
Pz . 0 > 0/ 1 ejyj dy < 1 :
.2/m=2 C1
As Pz . 0 > 0/ can be equal to 0 or 1 only, we must have Pz . 0 > 0/ D 0 and z is a

regular point. t
u
By Proposition 10.6 all the points of the boundary of an open convex set are
therefore regular, as in this case we can choose a cone which is a whole half-plane.
One easily sees that all the points at which the boundary is C2 are regular. The
cone condition requires that the boundary does not have inward pointing thorns, see
Fig. 10.3.
The cone hypothesis of Proposition 10.6 is only a sufficient condition for the
regularity of the boundary, as we see in the following example.
............ ..
........ ...........
......... ... D c . .......
.......
...... D c ................
........ .. ... ........ ..... .....
....... ....... ..... .
....
..
......
...... .. .. ....... .....
.....
...... ... .. ...
....... ....
... ....
...... . ... ...
..... ... .. ...... ...
...
..... .
..... . ... ......... ...
... ... .
.
..... .. ....
....... ............. ... ..
... ..
..... .
...... ..... ... ...
.... ....... ......
D ... ....
... .. D .....
...
• ...
• .
z z
Fig. 10.3 The domain on the left-hand side enjoys the cone property at z, the one on the right-hand
side doesn’t
Example 10.4 Let D R2 be a domain formed by a disc minus a segment S

as in Fig. 10.5. Let us prove that all the points of @D are regular. This fact is
obvious for the points of the circle so we need to consider only the points of
the segment S. We can assume that S belongs to the axis y D 0 and that one
of its ends is the origin. Let us prove that the origin is a regular point.
Let V be a ball centered at 0 and with a small enough radius and let
VS D V \ S; V1 D V \ .fy D 0gnS/; V0 D VS [ V1
(see Fig. 10.4). Consider the three stopping times
00 D infft > 0I Xt 2 VS g; 1 D infft > 0I Xt 2 V1 g; 0 D infft > 0I Xt 2 V0 g:
The idea of the proof is simple: starting at 0, by the Iterated Logarithm

Law applied to the component X2 of the Brownian motion X, there exists
a sequence of times .tn /n decreasing to 0 such that X2 .tn / D 0 for every n.
Now we just need to show that with positive probability, possibly passing to
a subsequence, we have X1 .tn / 2 VS for every n. We have
1) f 00 D 0g f 0 D 0g.
2) f0 D 0g D f1 D 0g [ f 00 D 0g. In one direction the inclusion is
obvious. Conversely, if ! 2 f0 D 0g then there exists a sequence .tn /n ,
decreasing to 0 and such that Xtn 2 V0 . As V0 D VS [ V1 , there exists a
subsequence .tnk /k such that Xtnk 2 V1 for every k, or Xtnk 2 VS for every
k; therefore ! 2 f1 D 0g or ! 2 f 00 D 0g, from which we derive the
opposite inclusion.
3) P0 . 00 D 0/ D P0 .1 D 0/, which is obvious by symmetry.
4) P0 . 0 D 0/ 12 P0 .0 D 0/. From 2) and 3) we have
P0 .0 D 0/ P0 . 00 D 0/ C P0 .1 D 0/ D 2 P0 . 00 D 0/ 2 P0 . 0 D 0/:
5) P0 .0 D 0/ P0 .sup0t˛ jX1 .t/j < / > 0 where is the radius of V

and ˛ is any number > 0. Indeed f0 D 0g contains the event
n X2 .t/ X2 .t/ o
lim 1=2
D 1; lim 1=2
D 1; sup jX1 .t/j <
t!0C 2t log log 1t t!0C 2t log log 1
t
0t˛
˚
and this event is equal to sup0t˛ jX1 .t/j < by the Iterated Logarithm
Law.
From 4) and 5), we have P0 . 0 D 0/ > 0 and by Blumenthal’s 0–1 law 0
is a regular point (Fig. 10.5).
..........................
............. ........
........ ......
...... ......
......... .....
.
..... .....
...
.... ...
... ...
...
..
. ...
. .... ....
.... V ....
.. .
... ...
...
... ... ... S ...
...
... ... .. ..
... ... V1 VS...
. ..
... ....
... .... .... .... ..
... ..
... ...
... ...
... ...
..... .
......
..... .
...... .....
....... .....
......... .......
........................................
Fig. 10.4 The different objects appearing in Example 10.4
..................................
.......... .......
....... ......
...... .....
..
...... .....
....
..... ...
...
. ...
...
... ...
.... ...
... ...
....
... S ...
...
... ..
... ...
... 0 ...
... ..
.
... ...
...
... ...
... ...
.....
...
....
..... .
...... .....
....... .....
......... ......
.......................................
Fig. 10.5 The domain D is the interior of the disc minus the segment S
Exercises
10.1 (p. 591)

a) Let D Rm be a bounded open set with a C2 boundary. Let L be the differential
operator
1X X
m m
@2 @
LD aij .x/ C bi .x/
2 i;jD1 @xi @xj iD1
@x i
and assume that the coefficients a and b satisfy the conditions a) and b) stated
p. 309 before Theorem 10.1. Let u 2 C2 .D/ \ C.D/ be the solution of
(
Lu D 1 on D
(10.39)
uj@D D 0 :
Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical diffusion associated to L

and the exit time from D. Prove that
u.x/ D Ex ./ :
b) Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical realization of an m-di-

mensional Brownian motion. Let be the exit time from the sphere of radius 1.
Solve (10.39) explicitly and prove that
1
Ex ./ D .1 jxj2 /
m
Compare the result of b) with Exercise 5.10 c).
10.2 (p. 592) Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical diffusion
associated to the SDE in dimension 1
dXt D b.Xt / dt C dBt :
Let 0 < a < b and let us assume that b.x/ D ıx for a x b and that, on R, b
satisfies Assumption (A’). Let be the exit time of X from a; bŒ.
a) Show that < C1 Px -a.s. for every x.
b) Prove that, if ı 6D 12 and for a < x < b,
1 . ax /
Px .X D b/ D (10.40)
1 . ab /
with D 2ı 1.
c) Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical realization of an m-dimen-
sional Brownian motion with m 3 and let D D fa < jxj < bg. Let us denote
by the exit time of X from the annulus D. What is the value of Px .jB j D b/
for x 2 D? How does this probability behave as m ! 1?
Go back to Exercise 8.24 for the SDE satisfied by the process t D jXt xj.
10.3 (p. 593) Let a; b > 0, > 0, 2 R.

a) Let x be the solution of the SDE
dt D dt C dBt
0 D x
and the exit time of x from the interval a; bŒ.

a1) Prove that < C1 a.s. whatever the starting point x.
a1) What is the generator, L, of the diffusion ? Compute P.0 D b/ ( 0 is the
solution starting at x D 0).
b) Let x be the solution of the SDE

dt D dt C p dBt
1 C 2t 1 C 2t
0 D x :
Let now be the exit time of x from the interval a; bŒ.

b1) What is the generator of the diffusion ? Can you say that < C1 a.s.?
Whatever the starting point x?
b2) Compute P.0 D b/ (0 is the solution starting at x D 0).
Compare the result of a) with Exercise 5.20.
10.4 (p. 594) Let B be a two-dimensional Brownian motion, the circle of radius
1, x D . 21 ; 0/ and the exit time of B from the ball of radius 1. If 0 denotes
the set of the points of the boundary with a positive abscissa, compute Px .B 2 0 /
(Fig. 10.6).
See Example 10.1.
. . .......
......
. .....
. .....
. ...
...
. ...
. ...
. ...
...
.
. • ..
...
.
. 1 ....
. 2 ..
.
. ...
. ...
....
.
. ..
......
..
. . .......
. . ....................
Fig. 10.6 The starting point and the piece of boundary of Exercise 10.4
10.5 (p. 595) (The Laplace transform of the exit time, to be compared with
Exercise 5.32)
a) Let X be the canonical diffusion on Rm associated to the generator
1X X
m m
@2 @
LD aij .x/ C bi .x/
2 i;jD1 @xi @xj iD1
@xi
that we assume to be uniformly elliptic and satisfying Assumption (A’). If D

Rm is a bounded open set having a C2 boundary and
2 R, let u (if it exists)
be a solution of
(
Lu C
u D 0 on D Œ0; TŒ
(10.41)
u@D D 1 :
We know that, if
0, a unique solution exists by Theorem 10.1 and is given
by u.x/ D Ex Œe
. Let, for every " > 0, D" be an open set such that D" D
and dist .@D" ; @D/ " and let us denote by ; " the respective exit times of X
from D; D" .
a1) Show that if Mt D e
t u.Xt / then, for every " > 0, .Mt^" /t is a martingale.
a2) Prove that if u 0 then
u.x/ D Ex Œe
:
b) Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical realization of a real

Brownian motion and D D a; aŒ, a > 0.
b1) Prove that
p

cos. 2
x/ , 2
E Œe D
x
p
<
cos. 2
a/ 8a2
2
and Ex Œe
D C1 for
8a 2.
b2) Deduce that the r.v. is not bounded but that there exist numbers ˇ > 0 such
that Px . > R/ const eˇR and determine them.
b3) Compute Ex Œ.
b) Note that u.x/ D Ex ŒMt^" for every t and prove first, using Fatou’s lemma, that the r.v. e
is
Px -integrable.
10.6 (p. 596) Let us consider the problem

8 2
<1 @ u @u
C D0 on R Œ0; TŒ
2 @x2 @t (10.42)
:
u.x; T/ D .x/ :
a) Find a solution for .x/ D x2 . What can be said about uniqueness?

b) Show that if is a polynomial, then the unique solution of (10.42) having
polynomial growth is a polynomial in x; t.
10.7 (p. 597)

a) Find a solution of the problem
8 2 2
< 2 @ u @u @u
x 2 .x; t/ C bx .x; t/ C .x; t/ D 0 on R Œ0; TŒ
2 @x @x @t (10.43)
:
u.x; T/ D x :
b) What if the boundary condition was replaced by u.x; T/ D x2 ?
10.8 (p. 597) Let
2 X @2 X
m m
@
LD C xi
2 iD1 @x2i iD1
@x i
a) Find a solution of the problem

8
<Lu C @u D 0 on Rm Œ0; TŒ
@t (10.44)
:
u.x; T/ D cos.h
; xi/ ;
where
2 Rm .
b) What can be said of x 7! u.x; 0/ as T ! C1?
10.9 (p. 598) Compute the fundamental solution of the Cauchy problem (in
dimension 1) of the operator
1 2 @2 @
LD ax 2 C bx ,
2 @x @x
where a > 0; b 2 R.
10.10 (p. 599) Let x;s be the solution of the SDE in dimension 1
dt D b.t ; t/t dt C .t ; t/t dBt

s D x ;
where b and are continuous functions in x; s, Lipschitz continuous in x and

bounded. Let us assume .x; t/ > 0 for every x; t.
a) Assume x > 0. Show that the process x;s x;s
t D log.t / is defined for every s > 0
and is the solution of an SDE to be determined. Deduce that t > 0 for every t
a.s.
b) Let W R ! R be a continuous function having polynomial growth and f ; c
functions Rm Œ0; T ! R measurable in x; t and locally Hölder continuous in
x, such that f has a polynomial growth as a function of x and c is bounded below.
Let
h Z i
RT x;t
T Rs x;t
u.x; t/ D E .Tx;t / e t c.v ;v/ dv E f .sx;t ; s/ e t c.v ;v/ dv
ds
t
and let Lt be the differential operator
1 @2 d
Lt D .x; t/2 x2 2 C b.x; t/x
2 @x dx
Show that u is a solution of
8
<L u C @u cu D f on RC Œ0; TŒ
t
@t (10.45)
:
u.x; T/ D .x/ :
• Note that in b) Theorem 10.6 cannot be applied directly as Lt does not satisfy
all the required assumptions (the diffusion coefficient vanishes at 0).
b) Trace back to the diffusion introduced in a) and apply the Feynman–Kac formula.
10.11 (p. 600) Let X D .C ; M ; .Mts /t ; .Xt /t ; .P x;t /x;t / be the canonical realization
of an m-dimensional Brownian motion and W Rm ! R a bounded Borel function.
Let
u.x; t/ D Ex;t Œ.XT / :
a) Show that u is C1 .Rm Œ0; TŒ/.

b) Show that u.x; s/ D Ex;s Œu.Xt ; t/ for every t; s t < T.
c) Show that u is a solution of
1 @u
4u C D0 on Rm Œ0; TŒ (10.46)
2 @t
lim u.x; t/ D .x/ for every x of continuity for : (10.47)
t!T
d) For m D 1, find a solution of (10.46) such that

(
1 if x > 0
lim u.x; t/ D
t!T 0 if x < 0 :
What is the value of limt!T u.0; t/?
10.12 (p. 601) Let X D .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical diffusion
associated to the differential operator
1X X
m m
@2 @
LD aij .x/ C bi .x/ (10.48)
2 i;jD1 @xi @xj iD1
@x i
that we assume to satisfy the same hypotheses as in Theorem 10.1. Let D be an open
set of Rm , x 2 @D; we shall say that @D has a local barrier for L at x if there exists
a function w.y/ defined and twice differentiable in a neighborhood W of x and such
that Lw 1 on W \ D, w 0 on W \ D and w.x/ D 0.
Then
a) if @D has a local barrier for L at x, then x is regular for the diffusion X.
b) (The sphere condition) Let x 2 @D and assume that there exists a ball S Dc
such that S \ D D fxg. Then x is regular for X.
a) Apply Ito’s formula and compute the stochastic differential of t 7! w.Xt /. b) Construct a local
barrier at x of the form w.y/ D kŒjx zjp jy zjp , where z is the center of S.
Chapter 11

Simulation
Applications often require the computation of the expectation of a functional of a

diffusion process. But for a few situations there is no closed formula in order to
do this and one must recourse to approximations and numerical methods. We have
seen in the previous chapter that sometimes such an expectation can be obtained by
solving a PDE problem so that specific numerical methods for PDEs, such as finite
elements, can be employed. Simulation of diffusion processes is another option
which is explored in this chapter.
11.1 Numerical approximations of an SDE
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a d-dimensional Brownian motion and the

m-dimensional solution of the SDE

(11.1)
u D
where
a) b is an m-dimensional vector field and a m d matrix field satisfying
assumptions to be made precise later;
b) is an Fu -measurable square integrable r.v.
We have already discussed the question of the simulation of the paths of a process
in the chapter about Brownian motion in Sect. 3.7. Of course, the method indicated
there cannot be extended naturally to the case of a general diffusion unless the
transition function p of is known and easy to handle, as is the case, for instance,
for the Ornstein–Uhlenbeck process, as explained in Example 9.3.


342 11 Simulation
This is not, however, a common situation: the transition function is in most cases
unknown explicitly or difficult to sample. Hence other methods are to be considered,
possibly taking into account that we shall have only an approximate solution.
The simplest method in this direction is the so-called Euler scheme, which
borrows the idea of the scheme that, with the same name, is used in order to
solve numerically Ordinary Differential Equations. Sometimes it is called the Euler–
Maruyama scheme, G. Maruyama (1955), being the first to apply it to SDEs.
The idea is to discretize the time interval Œu; T into n small intervals of length
h D 1n .T u/. Let tk D uCkh, k D 0; 1; : : : ; n. Then we consider the approximation
Z tk Z tk
tk D tk1 C b.s ; s/ ds C .s ; s/ dBs
tk1 tk1 (11.2)
' tk1 C b.tk1 ; tk1 /h C .tk1 ; tk1 /.Btk Btk1 / :
Let .Zn /npbe a sequence of d-dimensional independent N.0; I/-distributed r.v.’s; then
the r.v.’s h Zk have the same joint distributions as the increments Btk Btk1 of the
Brownian motion. We can construct the subsequent positions of an approximating
.n/ .n/
process by choosing the initial value .u/ by sampling with the same law as
(and independently of the Zk ’s) and then through the iteration rule
.n/ .n/ .n/ .n/ p

tk D tk1 C b. tk1 ; tk1 /h C . tk1 ; tk1 / h Zk : (11.3)
Of course we need to prove that, as h ! 0, the process obtained in this way

converges, in some appropriate way, to the solution of the SDE (11.1).
There are two points of view concerning the convergence. The first one is the
so-called strong approximation. A result about strong approximation is a statement
.n/
that gives an estimate of how close the values of the approximants tk are to those
of the solution tk when Zk D p1h .Btk Btk1 / as in (11.2).
The second point of view, the weak approximation, concerns the estimation of
.n/
how close the law of is to the law of . Typically it consists in estimates of how
.n/
close the value of EŒ f . T / is to the “true” value EŒ f .T / for a function f that is
sufficiently regular.
.n/
Definition 11.1 An approximation scheme is said to be strongly conver-
gent of order ˇ if for every k, 1 k n,
.n/
EŒj tk tk j2 1=2 const hˇ : (11.4)
(continued)
11.2 Strong approximation 343
Definition 11.1 (continued)

.n/
Let W be a class of functions Rm ! R. An approximation scheme is said
to be weakly convergent of order ˇ for the class of functions W if for some
T >0
ˇ ˇ
ˇEŒ f . .n/ / EŒ f .T /ˇ const hˇ (11.5)
T
for every function f 2 W .
Of course, for every Lipschitz continuous function f with Lipschitz constant L we

have
ˇ ˇ
ˇEŒ f . .n/ / EŒ f .T /ˇ EŒj f . .n/ / f .T /j LEŒj .n/ T j (11.6)
T T T
so that strong approximation gives information of weak type, at least for some class
of functions f .
In the next sections we shall often drop the superscript .n/ . We shall use the
notations
• n: number of subintervals into which the interval Œu; T is partitioned;

• h D 1n .T u/: length of each of these subintervals;
• tk D u C hk, k D 0; : : : ; n: endpoints of these subintervals.
11.2 Strong approximation
We shall make the following assumptions.
Assumption (E) We say that the coefficients b and satisfy Assumption (E)
if there exist constants L > 0; M > 0 such that for every x; y 2 Rm , t 2 Œ0; T,
jb.x; t/ b.x; s/j L.1 C jxj/jt sj1=2 ; (11.7)

j.x; t/ .x; s/j L.1 C jxj/jt sj1=2 ; (11.8)
jb.x; t/j M.1 C jxj/; j.x; t/j M.1 C jxj/; (11.9)
jb.x; t/ b. y; t/j Ljx yj; j.x; t/ . y; t/j Ljx yj : (11.10)

344 11 Simulation
In practice this is almost the same as Assumption (A) p. 260, but for the fact that we
require more regularity in the time variable.
Let us consider an Euler scheme whose subsequent positions t0 ; : : : ; tn are
defined in (11.3), with the choice Zk D p1h .Btk Btk1 /. It will be useful to define
an approximating process b interpolating the values of the Euler scheme between
the times tk . More precisely, for tk1 t tk , let
b
t D tk1 C b. tk1 ; tk1 /.t tk1 / C . tk1 ; tk1 /.Bt Btk1 / : (11.11)
Let us define jn .s/ D tn for tn s < tnC1 ; then b

satisfies
Z t Z t
b
t D C b.b
jn .s/ ; jn .s// ds C .b
jn .s/ ; jn .s// dBs : (11.12)
0 0
In particular, b
is a Ito process.
Let us prove first that b has finite moments of order p for every p 1.
Lemma 11.1 If the coefficients b; satisfy (11.9) (sublinear growth) and the
initial value belongs to Lp , p 2, then
h i h i
E sup j tk jp E sup jb
t jp < C1 :
kD0;:::;n utT
Proof The proof follows almost exactly the steps of the proof of Theorem 9.1. For
R > 0, let b R .t/ D b
.t ^ R /, where R D infftI t T; jb
t j > Rg denotes the exit
b
time of from the open ball of radius R. Then a repetition of the steps of the proof
of Theorem 9.1 gives the inequality
h i Z t

E sup jb
R .s/jp c1 . p; T; M/ 1 C Ejjp C c2 . p; T; M/ E jb
R . jn .r//jp dr
ust u
Z h i
t
c1 . p; T; M/ 1 C Ejjp C c2 . p; T; M/ E sup jb
R .s/jp dr :
u usr

Let now v.t/ D E supust jb
R .s/jp : from the previous inequality we have
Z s
v.t/ c1 . p; T; M/.1 C Ejjp / C c2 . p; T; M/ v.r/ dr :
u
As v.s/ R we can apply Gronwall’s inequality, which gives

v.T/ c. p; T; M/ 1 C EŒjjp ;
i.e.
h i
E sup jb
s jp c. p; T; M/ 1 C EŒjjp :
usT^R
We remark that the right-hand side does not depend on R and we can conclude the
proof by letting R ! C1 with the same argument as in the proof of Theorem 9.1.
t
u
The following theorem gives the main strong estimate.
Theorem 11.1 Assume that b and satisfy Assumption (E). Let B be a

d-dimensional Brownian motion, 2 Lp , p 2, an m-dimensional r.v.
independent of Fu and the solution of

u D :
Let b
D b .n/ be the approximating process defined by the Euler
scheme (11.11) and with initial condition b
u D , u 0. Then for every
p 1 and T > u, we have
h i
E sup jb
t t jp const hp=2 ; (11.13)
utT
where the constant c depends only on T, p, L, M and the Lp norm of .
Proof The idea of the proof is to find upper bounds allowing us to apply Gronwall’s
inequality. We shall assume for simplicity that u D 0. Let jn .s/ D tn for tn s <
tnC1 , then, thanks to (11.12), we have
ˇZ t ˇp
ˇ ˇ
jb
t t jp 2p1 ˇ .b.b jn .s/ ; jn .s// b.s ; s// dsˇ
Z
0
ˇ t ˇp
ˇ ˇ
C2p1 ˇ ..b jn .s/ ; jn .s// .s ; s// dBs ˇ
0
and, by Hölder’s inequality and the Lp bound of Proposition 8.9,

h i hZ t i
E sup jb u u jp 2p1 T p E jb.b
jn .s/ ; jn .s// b.s ; s/jp ds
0ut 0
hZ t i (11.14)
j.b
p2
C2 c. p; m; d/T E
p1 2 jn .s/ ; jn .s// .s ; s/jp ds :
0

346 11 Simulation
We have, for tn s tnC1 ,
jn .s/ ; jn .s// b.s ; s/j D jb.b

jb.b tn ; tn / b.s ; s/j
jb.s ; s/ b.tn ; s/j C jb.tn ; s/ b.tn ; tn /j C jb.tn ; tn / b.b

tn ; tn /j
hence, for tn v < tnC1 , under Assumption (E),

Z v
jb.b
jn .s/ ; jn .s// b.s ; s/jp ds
tn
Z v Z v Z v
3p1 Lp js tn jp ds C .1 C jtn j/p js tn jp=2 ds C jtn b
tn jp ds :
tn tn tn
The same arguments give

Z v
j.b
jn .s/ ; jn .s// .s ; s/jp ds
tn
Z v Z v Z v
3 p1 p
L js tn j ds C .1 C jtn j/
p p
js tn j p=2
ds C jtn b
tn jp ds :
tn tn tn
Thanks to the Lp estimates for the solution of an SDE with coefficients with a
sublinear growth, (9.13) and (9.12), p. 261, we have for tn s < tnC1 and denoting
by c. p; T/, c.L; p; T/ suitable constants,
EŒjs tn jp c. p; T/.s tn /p=2 .1 C EŒjtn jp / c. p; T/.s tn /p=2 .1 C EŒjjp / :
Substituting into (11.14) we find

h i n Z T
b
E sup j u u j c.L; T; p/ .1 C EŒjj /
p p
js jn .s/jp=2 ds
0ut Z t 0
o (11.15)
C EŒjjn .s/ b
jn .s/ jp ; ds :
0
As
Z T n1 Z
X tkC1
n T
js jn .s/jp=2 ds D .s tk /p=2 D p h1Cp=2 D p hp=2
0 kD0 tk 2 C1 2 C1
and
Z t Z t h i
EŒjjn .s/ b
jn .s/ jp ds E sup ju b
u jp ds ;
0 0 0us
for the function v.t/ D EŒsup0ut j u u jp we have the inequality

Z t
v.t/ hp=2 c.L; T; p/.1 C EŒjjp / C c.L; T; p/ v.s/ ds :
0
As by Lemma 11.1 and Theorem 9.1 the function v is bounded, we can apply
Gronwall’s inequality, which gives
h i
E sup jb
u u jp c.L; T; p/eTc.L;T;p/ .1 C EŒjjp / hp=2
0uT
concluding the proof.

t
u
1
By Theorem 11.1 with p D 2 the Euler scheme is strongly convergent of order 2
.
Theorem 11.1 states that the L2 difference between the r.v.’s tk and tk tends to
0 as h ! 0. This is enough in many situations, typically if we need to estimate
a functional of the type EŒ f .T /. Sometimes we might be concerned with more
complicated functionals of the path of , which leads to the following extension.
Let us define a process, which we shall denote again by , by setting, for tk
t tkC1 , t as the value that is obtained by interpolating linearly between the values
tk and tkC1 . More precisely, if tk t tkC1 , we have
1 1
tD .tkC1 t/tk C .t tk /tkC1
h h
and we define
1 1
t D .tkC1 t/ tk C .t tk / tkC1
h h (11.16)
1
D tk C b. tk ; tk /.t tk / C . tk ; tk /.t tk /.BtkC1 Btk / :
h
The processes and b coincide at the discretization times tk but differ between
the times tk and tkC1 because the stochastic components are different. has the
advantage that it can be numerically simulated.
Corollary 11.1 Under the assumptions of Theorem 11.1, as n ! 1 the

.n/
Euler approximations (11.16) converge in law to the law of on
C .Œu; T; Rm / for every T > 0 and in C .RC ; Rm /.
Proof Again we shall assume u D 0. We shall prove that, for a fixed T, from every
.n0j /
sequence .nj /j converging to C1 we can extract a subsequence .n0j /j such that

348 11 Simulation
converges to uniformly on Œ0; T. As a.s. convergence implies convergence in law,

this will conclude the proof. If we define hj D n1j , then we have thanks to (11.13)
and the Markov inequality, for every " > 0,
.nj /
1 h .nj /
i hj cT
P sup j tk tk j " 2 E sup j tk tk j2 c 2 D 2
kD0;:::;nj " kD0;:::;nj " n j"
hence certainly there exists a subsequence .n0j /j such that
1
X
.n0 /
P sup j tk j tk j " < C1 :
jD1 kD0;:::;n0j
By the Borel–Cantelli lemma, for every " > 0 the event

n .n0 /
o
sup j tk j tk j " for infinitely many indices j
kD0;:::;n0j
has probability 0. Let ! 2 ˝. Then as the map t 7! t .!/ is continuous, for every
" > 0 there exists a ı" > 0 such that if jtsj ı" then jt .!/s .!/j ". Let us fix
.n0 /
" > 0 and let j0 > 0 be such that h0j ", h0j ı" and supkD0;:::;n0j j tk j .!/ tk .!/j <
" for every j j0 . Then we have, for tk t tkC1 ,
jt .!/ tkC1 .!/j " and jt .!/ tk .!/j " (11.17)
and also, for j > j0 ,
.n0 / .n0 /
jtk .!/ tk j .!/j " and jtkC1 .!/ tkC1
j
.!/j " : (11.18)
.n0 / .n0 / .n0 / 1

By (11.16) we have t j .!/ D ˛ tk j .!/ C .1 ˛/ tkC1
j
.!/ with ˛ D h0j
.tkC1
t/. Thanks to (11.17) and (11.18), we have by adding and subtracting the quantity
˛tk .!/ C .1 ˛/tkC1 .!/, if tk t tkC1 ,
.n0 /
jt .!/ t j .!/j
jt .!/ .˛tk .!/ C .1 ˛/tkC1 .!//j
.n0 / .n0 /
Cj.˛tk .!/ C .1 ˛/tkC1 .!// .˛ tk j .!/ C .1 ˛/ tkC1
j
.!/ j
„ ƒ‚ …
.n0j /
D t .!/
˛jt .!/ tk .!/j C .1 ˛/jt .!/ tkC1 .!/j

.n0 / .n0 /
C˛jtk .!/ tk j .!/j C .1 ˛/jtkC1 .!/ tkC1
j
.!/j
2" :
11.3 Weak approximation 349
.n0 / .n0j /
Hence we have proved that, for j j0 , sup0tT jt t j j 2" hence that . /j
converges a.s. to in C .Œ0; T; Rm /.
.n/
The convergence in law of to in C .RC ; Rm / is easily deduced recalling that
convergence in C .RC ; Rm / means uniform convergence on Œ0; T for every T. The
t
u
We remarked above that Theorem11.1 and (11.6) guarantee that the Euler scheme
is weakly convergent of order 12 for the class of Lipschitz functions. In the next
section we shall see that, for a class W of regular functions and under regularity
assumptions on the coefficients b and , the weak rate of convergence is of order 1.
11.3 Weak approximation
The following theorem gives an estimate concerning the weak convergence of the
Euler scheme in the case of a time homogeneous diffusion. We shall skip the proof,
putting the focus on the applications of the results.
Theorem 11.2 (The Talay–Tubaro expansion) Let us assume that

a) the coefficients b and are time homogeneous and differentiable infinitely
many times with bounded derivatives of all orders (b and may be
unbounded themselves);
b) the initial condition is x 2 Rm ;
c) f W Rm ! R is differentiable infinitely many times and all its derivatives
have polynomial growth, i.e. for every multi-index ˛ D .˛1 ; : : : ; ˛m / there
exist numbers C˛ ; p˛ > 0 such that
ˇ @j˛j f ˇ
ˇ ˇ
ˇ ˛ ˇ C˛ .1 C jxjp˛ / :
@x
Then there exists a constant c such that
EŒ f .T / EŒ f . T / D c h C O.h2 / : (11.19)
The Talay–Tubaro theorem, Talay and Tubaro (1990) or Graham and Talay (2013,
p. 180), actually gives more precision about the constant c. Note also that Theo-
rem 11.2 does not just give a bound of the error as in Theorem 11.1, but gives an
expansion of the error, which is a more precise statement. This fact will be of some
importance in Sect. 11.6.

350 11 Simulation
Example 11.1 The usual weak convergence estimate, such as the one of
Theorem 11.2, gives information concerning how much the expectation
EŒ f . T / differs from the true value EŒ f .T /, but what about the discrepancy
when considering the expectation of a more complicated function of the path
of the diffusion ? For instance, if W Rm ! R is a regular function how do
we estimate the difference between
hZ T i
E .s / ds (11.20)
0
and, for instance, its Euler approximation
h Xn1 i
E h . tk / ‹ (11.21)
kD0
A nice approach can be the following. Let us consider the .mC1/-dimensional

diffusion that is the solution of the SDE

(11.22)
dt D .t / dt :
RT
Hence T D 0 .s / ds and we have written our functional of interest as a
function of the terminal value of the diffusion .; /. The Euler scheme gives
for the component the approximation
X
n1
T D h . tk / : (11.23)
kD0
If satisfies the assumptions of Theorem 11.2 and is differentiable infinitely

many times with bounded derivatives, then (11.19) guarantees for the Euler
approximation (11.23) a rate of convergence of order h.
11.4 Simulation and PDEs
The representation results of Chap. 10 allow us to compute numerically the solution

of a PDE problem via the numerical computation of expectations of functionals of
diffusion processes. Let us consider, as in Sect. 10.4, the differential operator
1X X
m m
@2 @
Lt D aij .x; t/ C bi .x; t/
2 i;jD1 @xi @xj iD1
@x i
11.4 Simulation and PDEs 351
and the parabolic problem

8
<Lt u cu C @u D f on Rm Œ0; TŒ
@t
:u.x; T/ D .x/ x 2 Rm :
We know that, under the hypotheses of Theorem 10.6, a solution is

h Z i
RT
T Rs
u.x; t/ D Ex;t .XT / e t c.Xs;s/ ds Ex;t f .Xs ; s/ e t c.Xv ;v/ dv
ds ; (11.24)
t
where .C ; M ; .Mt /t ; .Xt /t ; .Px;s /x;s / is the canonical diffusion associated to Lt .

Hence, thanks to Corollary 11.1, as the functionals inside the expectations in (11.24)
x;t
are continuous functionals of the paths .Xt /t , we have that, denoting by Pn the law
on .C ; M / of the Euler approximation with time step h D Tn , if
x;t
RT
x;t
hZ T Rs i
un .x; t/ D En .XT / e t c.Xs ;s/ ds
En f .Xs ; s/ e t c.Xv ;v/ dv
ds
t
then
lim un .x; t/ D u.x; t/ :

n!1
In practice, an approximation of the value of the solution u at .x; t/ can be obtained

by simulating, with the Euler scheme, N paths 1 ; : : : ; N of the diffusion associated
to the generator L and with starting condition x; t and considering the approximation
1 X
N
un;N .x; t/ D ˚k ;
N kD1
where
RT
Z T Rs
˚ k D . k .T// e t c. k .s/;s/ ds
f . k .s/; s/ e t c. k .v/;v/ dv
ds :
t
In the simpler situation f 0, c 0 the approximation becomes
1X
N
un;N .x; t/ D . k .T// :
N kD1

352 11 Simulation
The solutions of the Dirichlet and Cauchy–Dirichlet problems of Sects. 10.2

and 10.3 can be numerically approximated in the same spirit. But then an additional
problem appears. Let us consider the Dirichlet problem
(
Lu c D f on D
uj@D D :
Theorem 10.2 states that, under suitable assumptions, the solution is given by
hZ i
u.x/ D E Œ.X /Z E
x x
Zs f .Xs / ds ; (11.25)
0
where .C ; M ; .Mt /t ; .Xt /t ; .Px /x / denotes

Rt
the canonical realization of the diffusion
associated to the generator L, Zt D e 0 c.Xs / ds and is the exit time from D.
We can think of simulating the diffusion with its Euler approximation in order to
approximate u.x/ with
x x
hZ i
u.x/ D E Œ.X /Z E Zs f .Xs / ds (11.26)
0
x
where by E we denote the expectation with respect to the law of the Euler
approximation with step h.
We know from Corollary 11.1 that under mild assumptions the laws of the Euler
x
approximations P converge weakly, as h ! 0, to Px . Unfortunately the map W
C ! RC is not continuous so we cannot immediately state that u.x/ converges to
u.x/ as h ! 0. In this section we prove that, in most cases, this is not a difficulty.
The idea is very simple. In the next statement we prove that the exit time from
an open set is a lower semicontinuous functional C ! RC , whereas the exit time
from a closed set is upper semicontinuous. In Theorem 11.3 it will be proved that,
under suitable assumptions, the exit time from an open set D is Px -a.s. equal to the
exit time from its closure for every x. Hence will turn out to be Px -a.s. continuous,
which is sufficient in order to guarantee the convergence of u.x/ to u.x/.
Proposition 11.1 Let D be an open (resp. closed) set. Then the exit time
from D as a functional C ! RC is lower (resp. upper) semicontinuous.
Proof Assume that D is an open set and let 2 C . For every " > 0 the set D
fx 2 DI x D t for some t ./ "g is a compact set contained in D. Hence
ı WD d.; @D/ > 0 (the distance between a compact set and a disjoint closed set is
strictly positive). Now U D fw 2 C I sup0t . /" jwt t j 2ı g is a neighborhood
of such that for every path w 2 U .w/ ./ ". Hence, by the arbitrariness
of ",
lim .w/ ./ :

w!
Assume now that D is closed and let be such that ./ < 1. Then there exist
arbitrarily small values of " > 0 such that . /C" 2 Dc . Let ı WD d. . /C" ; D/ > 0.
Again if U D fw 2 C I sup0t . /C" jwt t j 2ı g, for every w 2 U we have
.w/ ./ C ". t
u
We must now find conditions ensuring that, for a given open set D, the exit time
from D and from its closure coincide Px -a.s.
The intuitive explanation of this fact is that the paths of a diffusion are subject
to intense oscillations. Therefore as soon as a path has reached @D, hence has gone
out of D, it immediately also exits from D a.s. This is an argument similar to the
one developed in Example 6.4 (and in particular in Fig. 6.1). It will be necessary
to assume the boundary to have enough regularity and the generator to be elliptic,
which will ensure that oscillations take place in all directions.
The formal proof that we now develop, however, shall take a completely different
approach.
Let D Rm be a regular open set and let Dn be a larger regular open set with
D Dn and such that dist.@D; @Dn / ! 0 as n ! 1. Let us consider the functions
u, un that are the solutions respectively of the problems
(
Lu D 1 on D
(11.27)
uj@D D 0
and
(
Lun D 1 on Dn
(11.28)
unj@Dn D 0 :
As a consequence of Theorem 10.2 (see also Exercise 10.1) we have, denoting by

, and n the exit times from D, D, Dn respectively,
u.x/ D Ex Œ; un .x/ D Ex Œn
hence, for every x 2 D, as n
u.x/ Ex Œ un .x/ : (11.29)

354 11 Simulation
But, of course, un is also the solution of

(
Lun D 1 on D
(11.30)
un .x/ D un .x/ x 2 @D
and thanks to the representation formula (10.6), for x 2 D,
Ex Œn D un .x/ D Ex Œun .X / C Ex Œ : (11.31)
Later on in this section we prove that, for every " > 0, there exists an n0 such that
un .x/ " on @D for n n0 . This is rather intuitive, as un D 0 on @Dn and the
boundaries of D and Dn are close. Hence (11.31) will give
un .x/ Ex Œ C " ;
which together with (11.29) gives
Ex Œ Ex Œ Ex Œ C " :
Hence, by the arbitrariness of ", Ex Œ D Ex Œ and as this entails D a.s.

The result is the following. It requires the boundary of D to be C2;˛ , i.e. locally
the graph of a function that is twice differentiable with second derivatives that are
Hölder continuous of order ˛. It also requires that D is convex, an assumption that
can certainly be weakened.
Theorem 11.3 Let .C ; M ; .Mt /t ; .Xt /t ; .Px /x / be the canonical diffusion

associated to the differential operator
1 X X
m m
@2 @
LD aij .x/ C bi .x/ (11.32)
2 i;jD1 @xi @xj iD1
@x i
and let D be a bounded open set with a C2;˛ boundary with 0 < ˛ < 1. We
assume that there exists an open set e
D such that D e
D and
• the coefficients a and b are locally Lipschitz continuous on e D;
• L is uniformly elliptic on e
D, i.e. ha.x/ z; zi jzj2 for some > 0 and for
every x 2 eD, z 2 Rm .
Then the exit time from D is a.s. continuous with respect to Px for every x.
In order to complete the proof we must prove that for every " > 0 there exists an n0
such that un .x/ " on @D for n n0 . This will follow as soon as we prove that the
gradient of un is bounded uniformly in n, and this will be a consequence of classical
estimates for elliptic PDEs.
Let us introduce some notation. For a function f on a domain D and 0 < ˛ 1,
let us introduce the Hölder’s norms
j f .x/ f . y/j
j f j˛ D k f k1;D C sup
x;y2D;x6Dy jx yj˛
and, for a positive integer k and denoting by ˇ D .ˇ1 ; : : : ; ˇm / a multi-index,
Xk X ˇ jˇj ˇ
ˇ@ f ˇ
j f jk;˛ D ˇ ˇ ˇ :
@x ˛
jˇjDh
hD0
The following result (see Gilbarg and Trudinger 2001, Theorem 6.6, p. 98 for
example) is a particular case of the classical Schauder estimates.
Theorem 11.4 Let D be an open set with a C2;˛ boundary and let u be a
solution of (11.27), where L is the differential operator (11.32). Assume that
the coefficients satisfy the conditions
ha.x/; i jj2 (11.33)
for every x 2 D, 2 Rm and, for some ˛, 0 < ˛ < 1,
jaij j˛ ; jbi j˛ ; (11.34)
for every i; j D 1; : : : ; m, for some constants ; > 0. Then for the solution
u of (11.27) we have the bound
juj2;˛ C.kuk1 C 1/ (11.35)
for some constant C D C.m; ˛; ; ; D/.
End of the proof of Theorem 11.3 We must prove that, for a suitable family .Dn /n
as above, the gradients of the solutions un of (11.30) are bounded uniformly in n.
This will be a consequence of the Schauder inequalities of Theorem 11.4, as the
Hölder norm j j2;˛ majorizes the supremum norm of the gradient.

356 11 Simulation
The first task is to prove that the constant C appearing in (11.35) can be chosen
to hold for Dn for every n.
Let us assume that the convex set D contains the origin so that we can choose
Dn as the homothetic domain .1 C 1n /D. Note that, as D is assumed to be convex,
D Dn . Let us define rn D 1 C 1n and let n be such that Dn e D. Let
e
un .x/ D un .rn x/
1 X .n/ X
m m
e @2 e.n/ @ (11.36)
Ln D e
aij .x/ C bi .x/
2 i;jD1 @xi @xj iD1
@x i
.n/
with eaij .x/ D r12 aij .rn x/, e
b.n/ .x/ D 1
rn bi .rn x/. Then e
Lne
un .x/ D .Ln un /.rn x/, hence
n
e
un is the solution of
(
e
Lne
un D 1 on D
(11.37)
e
unj@D D 0 :
Let us consider on D the differential operators eLn , L and let us verify that there exist
constants , such that (11.33) and (11.34) are satisfied for all of them. Indeed if
ha.x/; i jj2 for every x 2 Dn , then, as rn 2,
2 2
ha.n/.x/; i jj jj
rn2 4
and if is such that jaij j˛ (the Hölder norm being taken on Dn ), then
.n/ rn˛
jaij j˛
rn2
and a similar majorization holds for the coefficients e

b.n/ . Hence by Theorem 11.3
there exists a constant C such that, for every n,
rn kun 0 k1 D ke
un 0 k1 je
uj2;˛ C.ke
un k1 C 1/ D C.kun k1 C 1/ : (11.38)
Finally, simply observe that, as Dn Dn and the generator L is elliptic on Dn , the

exit time from Dn is integrable thanks to Proposition 10.1 and we have the bound
kun k1 D sup Ex Œn sup Ex Œn < C1 (11.39)

x2Dn x2Dn
and, putting together (11.38) and (11.39), we obtain that the supremum norm of the
gradient of un is bounded uniformly in n, thus concluding the proof of Theorem 11.3.
t
u
Note that the assumption of convexity for the domain D is required only in order
to have D .1 C 1n /D and can be weakened (D starshaped is also a sufficient
assumption, for example).
Example 11.2 Consider the Poisson problem in dimension 2

(
1
2
4u D0 on D
(11.40)
uj@D .x1 ; x2 / D x2 _ 0 ;
where D is the ball centered at 0 of radius 1. The representation formula (10.6)

gives
u.x/ D Ex ŒX2 ./ _ 0 ;
where .C ; F ; .Ft /t ; .Xt /t ; .Px /x / denotes the realization of the diffusion

process associated to L D 12 4 so that, under Px , X D .X1 ; X2 / is a two-
dimensional Brownian motion starting at x and is the exit time from D. The
value u.x/ of the solution at x 2 D can therefore be computed by simulating
many paths of the Brownian motion starting at x and taking the mean of the
values of X2 ./_0, i.e. the positive part of the ordinate of the exit point. Doing
this for a grid of starting points in D, the result would appear as in Fig. 11.1.
This can be compared with the solution obtained with the aid of some
specific numerical method for PDEs (finite elements for instance), shown in
Fig. 11.2.
It is apparent that finite elements are much more accurate, as a comparison
with the exact solution given by the formulas of Example 10.1 would show.
Simulation therefore performs poorly in comparison with these numerical
methods and requires, for an accurate result, a very small value of the
discretization step h and a large number of simulated paths. However it has
two great advantages.
a) It is very simple to implement.
b) Very often it is the only method available. This is the case, for instance,
if the dimension is larger than, say, 4. In these situations other numerical
methods are unavailable or become really complicated to implement.

358 11 Simulation
Fig. 11.1 The solution of (11.40) computed numerically by simulation of a two-dimensional

Brownian motion. At each point the value was determined as the average on n D 1600 paths,
1
each obtained by a time discretization h D 50
Fig. 11.2 The solution of (11.40) computed numerically with the finite elements method
11.5 Other schemes 359
11.5 Other schemes
The Euler scheme is possibly the most natural approach to the approximation of the
solution of an SDE but not the only one, by far. In this section we provide another
approximation scheme. Many more have been developed so far, the interested reader
can refer to the suggested literature.
Going back to (11.2) i.e.
Z tk Z tk
tk D tk1 C b.s ; s/ ds C .s ; s/ dBs ; (11.41)
tk1 tk1
a natural idea is to find a better approximation of the two integrals. The Euler scheme
can be thought of as a zero-th order development: what if we introduce higher order
developments of these integrals?
In this direction we have the Milstein scheme. Let us apply Ito’s formula to the
process s 7! .s ; s/: assuming that is twice differentiable, we have for tk1
s < tk ,
ij .s ; s/ Dij .tk1 ; tk1 /

Z s
@ij X @ij 1 X @2 ij
m m
C C bl C arl .u ; u/ du
tk1 @u lD1
@xl 2 r;lD1 @xr @xl
Z d X
X m
s
@ij
C rl .u ; u/ dBl .u/ :
tk1 lD1 rD1
@xr
The integral in du is of order h and, after integration from tk1 to tk in (11.41), will
give a contribution of order o.h/, which is negligible with respect to the other terms.
Also we can approximate
Z d X
X m
s
@ij
rl .u ; u/ dBl .u/
tk1 lD1 rD1
@xr
X
m
@ij X
d
' .tk1 ; tk1 / rl .tk1 ; tk1 /.Bl .s/ Bl .tk1 // ;
rD1
@xr lD1
which gives finally the Milstein scheme, which is obtained by adding to the Euler
iteration rule (11.2) the term whose i-th component is
X
d X
m X
d Z
@ij tk
.tk1 ; tk1 / rl .tk1 ; tk1 / .Bl .s/ Bl .tk1 // dBj .s/ :
jD1 rD1
@xr lD1 tk1

360 11 Simulation
If d D 1, i.e. the Brownian motion is one-dimensional and is an m1 dimensional

matrix, by Ito’s formula we have
Z tk
1
.Bs Btk1 / dBs D .Btk Btk1 /2 h
tk1 2
and the iteration rule becomes, for i D 1; : : : ; m,
i .tk / Di .tk1 / C bi .tk1 ; tk1 /h C i .tk1 ; tk1 /.Btk Btk1 /

Xm Z tk
@i
C .tk1 ; tk1 /l .tk1 ; tk1 / .Bs Btk1 / dBs
lD1
@xl tk1
Di .tk1 / C bi .tk1 ; tk1 /h C i .tk1 ; tk1 /.Btk Btk1 /
1 X @i
m
C .tk1 ; tk1 /l .tk1 ; tk1 / .Btk Btk1 /2 h :
2 lD1 @xl
In practice, the subsequent positions of the approximating process will be obtained

by choosing u and iterating
p
i .tk / D i .tk1 / C bi . tk1 ; tk1 /h C h i . tk1 ; tk1 /Zk
h X @i
m
C . ; tk1 /l . tk1 ; tk1 /.Zk2 1/
2 lD1 @xl tk1
for a sequence .Zk /k of independent N.0; 1/-distributed r.v.’s.

In dimension larger than 1 this scheme requires us to be able to simulate the joint
distribution of the r.v.’s
Z tk
Bi .tj /; .Bl .t/ Bl .tk1 // dBj .t/; i; j; l D 1; : : : ; d; k D 1; : : : ; n ;
tk1
which is not easy. In addition this scheme requires the computation of the derivatives
of . The Milstein scheme in practice is confined to the simulation of diffusions in
dimension 1 or, at least, with a one-dimensional Brownian motion.
Also for the Milstein scheme there are results concerning strong and weak
convergence. Without entering into details, the Milstein scheme is of strong order
of convergence equal to 1 (i.e. better that the Euler scheme) and of weak order of
convergence also 1 (i.e. the same as the Euler scheme).
Let us mention, among others, the existence of higher-order schemes, usually
involving a higher order development of the coefficients.
11.6 Practical remarks 361
11.6 Practical remarks
In this section we discuss some issues a researcher may be confronted with when
setting up a simulation procedure.
In practice, in order to compute the expectation of a continuous functional of the
diffusion process it will be necessary to simulate many paths 1 ; : : : ; N using the
Euler or some other scheme with discretization step h and then to take their average.
If the functional is of the form ˚./ for a continuous map ˚ W C .Œ0; T; Rm / !
R, the expectation EŒ˚./ will be approximated by the average
1 X
N
˚ N WD ˚. k / :
N kD1
In order to evaluate the error of this approximation it is natural to consider the mean
square error

E .˚ N EŒ˚.//2 :
We have, adding and subtracting EŒ˚./,

E .˚ N EŒ˚.//2

D E .˚ N EŒ˚.//2 C 2E ˚ N EŒ˚./ EŒ˚./ EŒ˚./
C.EŒ˚./ EŒ˚.//2 :
The double product above vanishes as EŒ˚ N D EŒ˚./ so that the mean square
error is the sum of the 2 terms

I1 WD E .˚ N EŒ˚.//2 ; I2 WD .EŒ˚./ EŒ˚.//2 :
The first term, I1 , is equal to N1 Var.˚. 1 //, i.e. is the Monte Carlo error. The
second one is the square of the difference between the expectation of ˚./ and
the expectation of the approximation ˚./. If we consider the Euler scheme and the
functional ˚ is of the form ˚./ D f .T / and the assumptions of Theorem 11.2 are
satisfied, we then have I2 c h2 .
These remarks are useful when designing a simulation program, in order to
decide the best values for N (number of simulated paths) and h (amplitude of the
discretization step). Large numbers of N make the Monte Carlo error I1 smaller,
whereas small values of h make the bias error I2 smaller.

362 11 Simulation
In the numerical approximation problems of this chapter it is possible to take

advantage of the classical Romberg artifice, originally introduced in the context of
the numerical solution of Ordinary Differential Equations. Assume that we know
that an approximation scheme has an error of order h˛ , i.e. that
.n/
ah WD EŒ f . T / D EŒ f .T / C ch˛ C o.h˛ / : (11.42)
Then if we have an estimate ah for the value h and an estimate ah=2 for h=2, we have
2˛ ah=2 ah D .2˛ 1/EŒ f .T / C o.h˛ /
i.e., the linear combination
2˛ ah=2 ah
D EŒ f .T / C o.h˛ /
2˛ 1
of the two approximations ah and ah=2 gives an approximation of higher order.
Note that the knowledge of the constant c in (11.42) is not needed. In order to
apply this artifice the value of ˛ must be known, otherwise there is no improvement
(but the estimate should not become worse by much).
In the case of an Euler approximation, in the assumptions of Theorem 11.2 we
have ˛ D 1 and the Romberg estimator will give an approximation of order h2 .
Example 11.3 Imagine that you want to obtain an approximation of the

expectation at time T of a geometric Brownian motion using the Euler
scheme. Of course, the true value is known and for geometric Brownian
motion there are more reasonable ways to perform a simulation, but this
example allows us to make some explicit estimates.
By Theorem 11.2 the error is of order h. The Romberg artifice says that if,
instead of a simulation with a very small value of h, you make two simulations
with moderately small values h and h=2 you may obtain a better estimate at
a lower price. Table 11.1 shows some results. Here the drift and diffusion
coefficient are b D 1 and D 0:5 and we set T D 1. Hence the true value is
e D 2:7182.
Table 11.1 The quantity of interest appears to be better estimated making two sets of approxima-
tions with the values h D 0:02 and h D 0:01 than with a single simulation with h D 0:001, which
is probably more costly in time. Here 106 paths were simulated for every run
Value Error
h D 0:02 2.6920 0.0262
h D 0:01 2.7058 0.0124
2a0:01 a0:02 2.7196 0.0014
h D 0:001 2.7166 0.0016
Example 11.4 Let us go back to Example 3.3 where we were dealing with
the estimation by simulation of the probability

P sup Bt > 1 :
0t1
We remarked there that the discrepancy between the estimate provided by

the Euler scheme (which in this case coincides with a discretization of the
Brownian motion) and the true value was decreasing very slowly. Actually it
is believed that it is of the order h1=2 . Granting this fact, the Romberg artifice
1 1
gives, combining the estimates for h D 200 and h D 400 , the value
p
2 0:3031 0:2975
p D 0:3166
21
with a 0:15% of relative error with respect to the true value 0:3173. Note that
1
this value is much better than the result obtained by simulation with h D 1600 .
Exercises
11.1 (p. 602) Let us consider the geometric Brownian motion that is the solution
of the SDE
dt D bt dt C t dBt

(11.43)
0 D x :

364 11 Simulation
In this exercise we compute explicitly some weak type estimates for the Euler
scheme for this process. Let us denote by the Euler scheme with the time interval
Œ0; T divided into n subintervals. Of course the discretization step is h D Tn . Let
.Zk /k be a sequence of independent N.0; 1/-distributed r.v.’s.
a1) Prove that
Y
n
p
T x .1 C bh C h Zk / : (11.44)
iD1
a2) Compute the mean and variance of T and verify that they converge to the mean
and variance of T , respectively.
a3) Prove that
jEŒT EŒ T j D c1 h C o.h/ (11.45)

2
jEŒT2 EŒ T j D c2 h C o.h/ : (11.46)
b1) Find a formula similar to (11.44) for the Milstein approximation of T .

b2) Let us denote this approximation by e T . Prove that
jEŒT EŒe
T j D e
c1 h C o.h/ (11.47)
and compare the valuese

c1 and c1 of (11.45).
Chapter 12
Back to Stochastic Calculus
12.1 Girsanov’s theorem
Example 12.1 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. We

know that
1 2t
Mt D e Bt 2
is a positive martingale with expectation equal to 1. We can therefore define

on .˝; F / a new probability Q letting, for every A 2 F ,
Z
Q.A/ D EŒMT 1A D MT dP :
A
What has become of .Bt /t with respect to this new probability? Is it still a
Brownian motion?
We can first investigate the law of Bt with respect to Q. Its characteristic
function is
EQ Œei
Bt D EŒei
Bt MT :
Let us assume t T. Then, as Bt is Ft -measurable,

1 2
EŒei
Bt MT D E EŒei
Bt MT jFt D EŒei
Bt Mt D EŒe. Ci
/Bt 2 t
1 2 1 2 tCi
t 1 2 tCi
t
D EŒe. Ci
/Bt 2 . Ci
/ t e 2
D e 2
;
„ ƒ‚ …
D1
(continued )

366 12 Back to Stochastic Calculus

where the expectation of the quantity over the brace is equal to 1 because
we recognize an exponential complex martingale. The quantity on the right-
hand side is the characteristic function of an N. t; t/ distribution. Hence, with
respect to Q, Bt N. t; t/; it is not a centered r.v. and .Bt /t cannot be a
Brownian motion under Q. What is it then?
The previous computation suggests that Bt might be equal to a Brownian
motion plus a motion t 7! t. To check this we must prove that Wt D Bt t
is a Brownian motion with respect to Q. How can this be done? Among the
possible ways of checking whether a given process W is a Brownian motion
recall Theorem 5.17, reducing the task to the verification that
1 2 1 2t
Yt D ei
Wt C 2
t D ei
.Bt t/C 2
is a Q-martingale for every

2 R. This is not difficult: if s t T and
A 2 Fs , we have, again first conditioning with respect to Ft ,

EQ ŒYt 1A D EŒYt 1A MT D E Yt 1A EŒMT jFt D EŒYt 1A Mt

1 2 1 2 1 2
D EŒei
.Bt t/C 2
t e Bt 2 t 1A D EŒe„. Ci
/Bƒ‚
t 2 t. Ci
/
… 1A
WDZt
. Ci
/Bs 12 s. Ci
/2
D EŒZs 1A D EŒe 1A D EŒYs Ms 1A D EŒYs MT 1A D EQ ŒYs 1A ;
as we recognized that Z is a complex martingale with respect to P.

Hence, for 0 t T, B is a Brownian motion plus a drift t 7! t.
The example above introduces a subtle way of obtaining new processes from old:
just change the underlying probability. In this section we develop this idea in full
generality, but we shall see that the main ideas are more or less the same as in this
example.
From now on let B D .˝; F ; .Ft /t ; .Bt /t ; P/ denote an m-dimensional Brownian
2
motion. Let ˚ be a Cm -valued process in Mloc .Œ0; T/ and
hZ t
1
Z t i
Zt D exp ˚s dBs ˚s2 ds ; (12.1)
0 2 0
where, if z 2 Cm , we mean z2 D z21 C C z2m . By the same computation leading

to (8.40) we have
dZt D Zt ˚t dBt (12.2)

12.1 Girsanov’s theorem 367
and therefore Z is a local martingale. If ˚ is Rm -valued, Z is a positive supermartin-

gale and E.Zt / E.Z0 / D 1. If E.ZT / D 1 for T > 0, we can consider on .˝; F /
the new probability dQ D ZT dP. Girsanov’s Theorem below investigates the nature
of the process .˝; F ; .Ft /t ; .Bt /t ; Q/ in general.
First let us look for conditions guaranteeing that Z is a martingale. Let jzj2 D
Pm 2
RT 2

kD1 jzk j for z 2 C . If E 0 jZt ˚t j dt < C1, then Z˚ is an m-dimensional

m
process whose real and imaginary parts are in M 2 .Œ0; T/ and therefore, recall-
ing (12.2), Z is a complex martingale.
RT
Proposition 12.1 If 0 j˚s j2 ds K for some K 2 R then .Zt /0tT is a
complex martingale bounded in Lp for every p.
Proof Let Z D sup0sT jZs j and let us show first that E.Z p / < C1. We shall
use the elementary relation (see Exercise 1.3)
Z C1
E.Z p / D prp1 P.Z > r/ dr : (12.3)
0
Therefore we must just show that r 7! P.Z r/ goes to 0 fast enough. Observe
that, as the modulus of the exponential of a complex number is equal to the
exponential of its real part,
hZ t Z
1 t
Z
1 t i
2
jZt j D exp Re ˚s dBs jRe ˚s j ds C jIm ˚s j2 ds
0 2 0 2 0
Z
h t Z i
1 t
exp Re ˚s dBs C jIm ˚s j2 ds :
0 2 0
Therefore, if r > 0, by the exponential inequality (8.41),
hZ t Z
1 t i
P.Z r/ P sup exp Re ˚s dBs C jIm ˚s j2 ds r
0tT 0 2 0
Z t Z
1 T
D P sup Re ˚s dBs log r jIm ˚s j2 ds
0tT 0 2 0
Z t
K h .log r K /2 i
2
P sup Re ˚s dBs log r 2 exp :
0tT 0 2 2K
2
1
The right-hand side at infinity is of order ec.log r/ D rc log r and therefore r 7!
˛
P.Z r/ tends to 0 as r ! 1 faster than r for every ˛ > 0 and the integral
in (12.3) is convergent for every p > 1. We deduce that ˚Z 2 M 2 , as
Z T Z T
2 2 2 2
E j˚s j jZs j ds E Z j˚s j2 ds KE.Z / < C1
0 0
and, thanks to (12.2), Z is a martingale. It is also bounded in Lp for every p as we

have seen that Z 2 Lp for every p.
t
u
If ˚ is Rm -valued, then Zt 0 and if Z is a martingale for t 2 Œ0; T, then E.ZT / D
E.Z0 / D 1. In this case we can consider on .˝; FT / the probability Q having density
ZT with respect to P. From now on EQ will denote the expectation with respect to Q.
Lemma 12.1 Let .Yt /t be an adapted process. A sufficient condition for it to

be an .Ft /t -martingale with respect to Q is that .Zt Yt /t is an .Ft /t -martingale
with respect to P.
Proof Let s t T. From Exercise 4.4 we know that
E.ZT Yt jFs /
EQ .Yt jFs / D (12.4)
E.ZT jFs /
But if .Zt Yt /t is an .Ft /t -martingale with respect to P

E.ZT Yt jFs / D E E.ZT Yt jFt /jFs D E Yt E.ZT jFt /jFs D E.Zt Yt jFs / D Zs Ys
so that from (12.4)
Zs Ys
EQ .Yt jFs / D D Ys :
Zs
t
u
Theorem 12.1 (Girsanov) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-

2
dimensional Brownian and ˚ 2 Mloc .Œ0; T/. Let us assume that the process
Z defined in (12.1) is a martingale on Œ0; T and let Q be the probability on
.˝; F / having density ZT with respect to P. Then the process
Z t
e
Bt D Bt ˚s ds
0
is an .Ft /t -Brownian motion on Œ0; T with respect to Q.
Proof Thanks to Theorem 5.17 we just need to prove that, for every 2 Rm ,
Yt D eih ;e
1 2t
Bt iC 2 j j
is an .Ft /t -martingale with respect to Q. This follows from Lemma 12.1 if we verify
that Xt WD Zt Yt is a .Ft /t -martingale with respect to P. We have
hZ t
1
Z
1 2i t
Xt D Zt Yt D exp ˚s dBs j j t ˚s2 ds C ih ; e
Bt i C
0 0 2 2
hZ t Z
1 t 2
Z t
1 i
D exp .˚s C i / dBs ˚s ds i h ; ˚s i ds C j j2 t
0 2 0 0 2
hZ t 1
Z t i
D exp .˚s C i / dBs .˚s C i /2 ds
0 2 0
Pm 2
(recall that z2 is the function Cm ! C defined as z2 D kD1 zk ). If the r.v.
RT 2
0 j˚s j ds is bounded then X is a martingale by Proposition 12.1. In general, let
Z
˚ t
n D inf t TI j˚s j2 ds > n (12.5)
0
and ˚n .s/ D ˚s 1fs<n g . We have

hZ t
1
Z t i
exp ˚n .s/ dBs j˚n .s/j2 ds
0 2 0
hZ t^n
1
Z t^n i
D exp ˚.s/ dBs j˚.s/j2 ds D Zt^n :
0 2 0
If
h˝ Z t i
˛ 1
Yn .t/ D exp i ; Bt ˚n .s/ ds C j j2 t ;
0 2
RT
then, as 0 j˚n .s/j2 ds n, Xn .t/ D Zt^n Yn .t/ is a P-martingale by the first part of
the proof. In order to show that Xt D Zt Yt is a martingale, we need only to prove
that Xn .t/ ! Xt as n ! 1 a.s. and in L1 . This will allow us to pass to the limit as
n ! 1 in the martingale relation
EŒXn .t/jFs D Xn .s/ a:s:
First of all Zt^n !n!1 Zt a.s., as n !n!1 C1. If Hn D min.Zt^n ; Zt /, then

Hn !n!1 Zt and, as 0 Hn Zt , by Lebesgue’s theorem EŒHn ! EŒZt . Finally,
as EŒZt D EŒZt^n D 1 (both .Zt /t and .Zt^n /t are martingales equal to 1 for t D 0),
EŒjZt Zt^n j D EŒZt Hn C EŒZt^n Hn D 2EŒZt Hn ! 0

n!1
and therefore Zt^n !n!1 Zt in L1 . Moreover, note that Yn .t/ !n!1 Yt a.s. and
1 2
jYn .t/j e 2 j j t . Then

E jXn .t/ Xt j D E jZn .t/Yn .t/ Zt Yt j

E jZn .t/ Zt jjYn .t/j C E Zt jYn .t/ Yt j
and by Lebesgue’s theorem both terms on the right-hand side tend to 0 as n ! 1.

It follows that, for every t, Xn .t/ !n!1 Xt in L1 which concludes the proof.
t
u
Girsanov’s theorem has many applications, as the next examples will show. Note,
however, a first important consequence: if X is an Ito process with stochastic
differential
dXt D At dt C Gt dBt
then, with the notations of Theorem 12.1,
dXt D .At C Gt ˚t / dt C Gt de
Bt :
2 1
As both .Gt /t and .˚t /t belong to Mloc , the process t 7! Gt ˚t belongs to Mloc , so
that X is also an Ito process with respect to the new probability Q and with the
same stochastic component. Because of the importance of Girsanov’s theorem it
is useful to have weaker conditions than that of Proposition 12.1 ensuring that Z
is a martingale. The following statements provide some sufficient conditions. Let
us recall that anyway Z is a positive supermartingale: in order to prove that it is a
martingale, it is sufficient to show that E.ZT / D 1.
2
Rt
Theorem 12.2 Let ˚ 2 Mloc .Œ0; T/, Mt D 0 ˚s dBs , 0 t T, and
1
Zt D eMt 2 hMit ; tT:
Consider the following properties:

1 1
RT 2
a) EŒe 2 hMiT D EŒe 2 0 j˚s j ds < C1 (the Novikov’s criterion).
1
b) .Mt /0tT is a martingale that is bounded in L2 and EŒe 2 MT < C1 (the
Kazamaki’s criterion).
c) .Zt /0tT is a uniformly integrable martingale.
Then a))b))c).
Proof If a) is true then the positive r.v. hMiT has a finite Laplace transform at
D 12 .
It is therefore integrable:
hZ T i
EŒhMiT D E j˚s j2 ds < C1
0
hence ˚ 2 M 2 .Œ0; T/ and .Mt /0tT is a martingale which is bounded in L2 .

Moreover, note that
1 1=2 1
e 2 MT D ZT .e 2 hMiT /1=2
so that, thanks to the Cauchy–Schwarz inequality,

1 1 1
EŒe 2 MT EŒZT 1=2 EŒe 2 hMiT 1=2 EŒe 2 hMiT 1=2 < C1
(recall that EŒZT 1) so that b) is true.

Let us prove that b))c). By the stopping theorem, for every stopping time T
we have M D EŒMT jF and by Jensen’s inequality
1 1
e 2 M EŒe 2 MT jF :
1 1
As e 2 MT is assumed to be integrable, the r.v.’s EŒe 2 MT jF with ranging among
the stopping times that are smaller than T form a uniformly integrable family
1
(Proposition 5.4). Hence the family of r.v.’s e 2 M (which are positive) is also
uniformly integrable.
.a/ a
Let 0 < a < 1 and Yt D e 1Ca Mt . We have
1 2 hMi 1 a2 .a/ 1a2

eaMt 2 a t
D eMt 2 hMit Yt :
1Ca
If A 2 FT and T is a stopping time we have by Hölder’s inequality, as 2a > 1,
1 2 1 2 .a/ 2
EŒ1A eaM 2 a hMi EŒeM 2 hMi a EŒ1A Y 1a
.a/ 2 1Ca 2a 2 1 (12.6)
EŒ1A Y 1a EŒ1A .Y .a/ / 2a 1Ca .1a / D EŒ1A e 2 M a.1a/ :
1
We used here the relation EŒeM 2 hMi D EŒZ EŒZT 1, which follows from
the stopping theorem applied to the supermartingale Z.
1
As we know that the family .e 2 M / for ranging among the stopping times
that are smaller than T forms a uniformly integrable family, by (12.6) and using the
1 2
criterion of Proposition 5.2, the same is true for the r.v.’s eaM 2 a hMi .
1 2
Let us prove that .eaMt 2 a hMit /t is a uniformly integrable (true) martingale: as
we know that it is a local martingale, let .n /n be a sequence of reducing stopping
times. Then for every s t, A 2 Fs ,
1 2 hMi 1 2 hMi
EŒeaMt^n 2 a t^n
1A D EŒeaMs^n 2 a s^n
1A :
1 2 1 2
As eaMt^n 2 a hMit^n 1A !n!C1 eaMt 2 a hMit 1A and these r.v.’s are uniformly inte-
grable, we can take the limit in the expectations and obtain the martingale relation
1 2 hMi 1 2 hMi
EŒeaMt 2 a t
1A D EŒeaMs 2 a s
1A :
Again by Jensen’s inequality, repeating the argument above for D T, A D ˝, we

have
1 2 hMi 1 2 .a/ 2 1 2 1
1 D EŒeaMT 2 a T
EŒeMT 2 hMiT a EŒYT 1a EŒeMT 2 hMiT a EŒe 2 MT a.1a/ :
1
Letting a ! 1 this inequality becomes EŒeMT 2 hMiT 1, which allows us to
conclude the proof.
t
u
Corollary 12.1 With the notations of Theorem 12.2, if for some > 0
2
EŒej˚t j < C < C1 for every 0 t T, then .Zt /0tT is a martingale.
Proof By Jensen’s inequality (the exponential is a convex function)

h1 Z t2 i h 1
Z t2
1 i
exp j˚s j2 ds D exp .t2 t1 /j˚s j2 ds
2 t1 t2 t1 t1 2
Z
1 t2
1 2
e 2 .t2 t1 /j˚s j ds :
t2 t1 t1
Let t2 t1 < 2. With this constraint we have

h 1 Z t2 i 1
Z t2
2
2
E exp j˚s j ds EŒej˚s j ds < C1 :
2 t1 t2 t1 t1
Let 0 D t1 < t2 < < tnC1 D T be chosen so that tiC1 ti < 2 and let
˚i .s/ D ˚s 1Œti ;tiC1 Œ .s/. Each of the processes ˚i ; i D 1; : : : ; n, satisfies condition a)
of Theorem 12.2 as
h 1 RT 2
i h 1 Z tiC1 i
E e 2 0 j˚i .s/j ds D E exp j˚s j2 ds < C1
2 ti
and if
hZ t
1
Z t i
Zti D exp ˚i .s/ dBs j˚i .s/j2 ds
0 2 0
then Z i is a martingale. We have ZT D ZT1 : : : ZTn ; moreover, Ztii D 1, as ˚i D 0

on Œ0; ti Œ and therefore E.ZTi jFti / D Ztii D 1. Hence it is easy now to deduce that
EŒZT D 1 as
E.ZT / D E.ZT1 : : : ZTn / D EŒZT1 : : : ZTn1 E.ZTn jFtn / D E.ZT1 : : : ZTn1 / D D 1

t
u
Example 12.2 Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. Let us

prove that
Z t Z

2 t 2
Zt D exp
Bs dBs B ds
0 2 0 s
is a martingale for t 2 Œ0; T.

Novikov’s criterion requires us to prove that

2 R T 2
E e 2 0 Bs ds < C1 :
Here, however, it is much easier to apply the criterion of Corollary 12.1, which
requires us to check that for some > 0
2 2
E e
Bt < C1 (12.7)
for every t 2 Œ0; T. Recalling Remark 3.3 we have that if 2

2 T < 1 then
2 2
EŒe
Bt D .1 2
2 T/1=2 , so that (12.7) is satisfied as soon as
1
<
2
2 T
Therefore the criterion of Corollary 12.1 is satisfied and .Zt /0tT is a

martingale.
We can therefore consider on F the probability Q having density ZT with
respect to P. What can be said of B under this new probability Q?
Girsanov’s theorem states that, for 0 t T,
Z t
Wt D Bt
Bs dBs
0
(continued )

is a Q-Brownian motion. Therefore, under Q, .Bt /0tT has the differential
dBt D
Bt dt C dWt
and, with respect to Q, .Bt /0tT is an Ornstein–Uhlenbeck process.
Example 12.3 As an example of application of Girsanov’s theorem, let us

investigate the distribution of the passage time a at a > 0 of a Brownian
motion with a drift . Recall that the distribution of a for the Brownian
motion has been computed in Exercise 3.20. Let B be a Brownian motion
2
and Zt D eBt 2 t . Then Z is a martingale and, by Girsanov’s formula, if
dQ D Zt dP, then, with respect to Q, Ws D Bs s is, for s t, a Brownian
motion. Hence, with respect to Q, Bs D Ws C s.
If a D infftI Bt ag, let us compute Q.a t/, which is then the partition
function of the passage time of a Brownian motion with drift . We have

Q.a t/ D EQ Œ1fa tg D EP 1fa tg Zt :
Recall now that fa tg 2 Fa \ Ft D Fa ^t and that Z is a martingale.

Therefore, by the stopping theorem for martingales,

EP 1fa tg Zt D EP EŒ1fa tg Zt jFa ^t D EP 1fa tg Za ^t D EP 1fa tg Za ;
the last equality coming from the fact that, of course, on fa tg we have
a ^ t D a . But
2 2
Za D eBa 2 a
D ea 2 a
and, under P, B is a Brownian motion so that the law of a under P is given

by (3.23). In conclusion,
2
Q.aZ t/ D ea EP Œ1fa tg e 2 a
t 2
a a2 =2s 2 s
D ea 1=2 3=2
e e ds (12.8)
Z t0 .2/ s 1
a 2
D 1=2 s3=2
e 2s .sa/ ds :
0 .2/
(continued )
.9
.6
.3
0 1 2 3 4 56
3
Fig. 12.1 The graph of the densities of the passage time a for a D 1 and D 4
(solid) and
D 0 (dots). The first one decreases much faster as t ! C1

If > 0, then we know that a is Q-a.s. finite and taking the derivative we
find that the passage time a of the Brownian motion with a positive drift
has density (Fig. 12.1)
a 1 2
f .t/ D e 2t .ta/ : (12.9)
.2/1=2 t3=2
In particular,
Z C1 1
a 2
e 2s .sa/ ds D 1 : (12.10)
0 .2/1=2 s3=2
Note the rather strange fact that the integral in (12.10) does not depend on ,
as far as 0.
For > 0 the passage time has a finite expectation, as its density goes
to 0 at C1 exponentially fast. This is already known: in Exercise 5.33 we
obtained, using a Laplace transform argument, that EŒa D a .
If < 0, conversely, we know that a D C1 with probability 1 e2a
(Exercise 5.17).
Example 12.4 (Same as the previous one, but from a different point of view)
Imagine we want to compute the quantity
EŒ f .a /
(continued )

a being the passage time at a of a Brownian motion and f some bounded
measurable function. Imagine, moreover, that we do not know the distribution
of a (or that the computation of the integral with respect to its density is too
complicated) so that we are led to make this computation by simulation. This
means that we are going to simulate N paths of the Brownian motion and for
each one of these to note the passage times, T1 ; : : : ; TN say. Then by the law
of large numbers
1X
N
f .Ti /
n jD1
is, for N large, an approximation of EŒ f .a /.

However, in practice this procedure is unfeasible as a is finite but can take
very large values (recall that a does not have finite expectation). Therefore
some of the values Ti may turn out to be very large, so large that the
computer can stay on a single path for days before it reaches the level a (see
Exercise 3.20 b)).
The computation of the previous example suggests that we can do the
following: instead of a Brownian motion, let us simulate a Brownian motion

with a positive drift , collect the passage times, T1 ; : : : ; TN say, and consider
the quantity
ea X
N 2

f .Ti /e 2 Ti :
N jD1
2
Let us check that ea EŒ f .Ti /e 2 Ti D EŒ f .a / so that, again by the law of
large numbers, this is also an estimate of EŒ f .a /. Actually, denoting g ; g0

the densities of the r.v.’s Ti and Ti ,respectively (given by (12.9) and (3.23)
respectively), we have
Z C1
a

2
a
2
e E f .Ti /e 2 Ti De f .t/ e 2 t
g .t/ dt
0
Z C1 1
2 a 2
D ea f .t/ e 2 t
e 2t .ta/ dt
0 .2/1=2 t3=2
Z C1 a2
a
D f .t/ e 2t dt
0 .2/1=2 t3=2
Z C1
D f .t/g0 .t/ dt D E f .Ti / :

0
(continued )

Now, as the passage time of the Brownian motion with a positive drift
has finite expectation, the simulation of these paths is much quicker and the
problem mentioned above disappears.
In other words, Girsanov’s theorem can by applied to simulation in the
following way: instead of simulating a process X (B in the example above),
which can lead to problems, simulate instead another process Y (Bt C t
above) that is less problematic and then compensate with the density of the
law of X with respect to the law of Y.
Example 12.5 Let X be a Brownian motion and a real number. Let a > 0
and
D infftI Xt a C tg :
We want to compute P. T/, i.e. the probability for X to cross the linear
barrier t 7! a C t before time T. For D 0 we already know the answer
thanks to the reflection principle, Corollary 3.4. Thanks to Example 12.3 we
know, for < 0, the density of so we have
Z T 1
a 2
P. T/ D e 2t .ta/ dt ;
0 .2/1=2 t3=2
but let us avoid the computation of the primitive by taking another route and
applying Girsanov’s theorem directly. The idea is to write

P. T/ D P sup .Xt t/ a
0tT
and then to make a change of probability so that, with respect to the new
probability, t 7! Xt t is a Brownian motion for which known formulas are
available. Actually, if
1 2
ZT D e XT 2 T
and dQ D ZT dP, then with respect to Q the process
Wt D Xt t
(continued )

is a Brownian motion up to time T. Hence
h i
P sup .Xs s/ a D EQ ZT1 1fsup0sT .Xs s/ag
0sT
h 1 2
i
D EQ eXT C 2 T 1fsup0sT .Xs s/ag
1 2
h i
D e 2 T EQ e WT 1fsup0sT Ws ag :
The expectation on the right-hand side can be computed analytically as, from
Corollary 3.3, we know the joint density of WT and sup0sT Ws , i.e.

P sup .Xs s/ a
0sT
Z C1 Z
2 1 2
s
1 2
D p e 2 T ds .2s x/ e 2T .2sx/ e x dx :
2T 3 a 1
With the change of variable y D 2s x and using Fubini’s theorem

Z C1 Z C1
2 1 2 1 2
P sup .Xs s/ a D p e 2 T ds ye 2T y e.2sy/ dy
0sT 2T 3 a s
Z C1 Z y
2 1 2 1 2
Dp e 2 T ye 2T y ey dy e2s ds
2T 3 a a
Z C1
1 12 2 T 1 2
D p e ye 2T y ey .e2a e2y / dy :
2T 3 a
Now, integrating by parts,

1 ˇC1
1 2 1 2 ˇ
P sup .Xs s/ a D p e 2 T e 2T y ey .e2a e2y /ˇ
0sT 2T „ ƒ‚ a
…
D0
Z C1
1 1 2 1 2
Cp e 2 T
e 2T y .e.y2a/ C ey / dy
2T a
and with the change of variable z D p1 .y T/

T
Z C1
1 1 2 1 2
p e2a e 2 T e 2T y ey dy
2T a
(continued )
12.2 The Cameron–Martin formula 379

Z C1
1 2a 1 2
Dp e e 2T .yT/ dy
2T a
Z C1 a T
1 z2
D p e2a p e dz D e2a 1 ˚ p :
2 .aT/= T T
Similarly
Z C1 Z C1
1 12 2 T 1 2
2T y y 1 1 2
p e e dy p e e 2T .yCT/ dy
2T a 2T a
Z C1 a C T
1 z2 =2
p p e dy D 1 ˚ p
2 .aCT/= T T
and finally

P sup .Xs s/ a
0sT
a T a C T (12.11)
2a
De 1˚ p C 1˚ p :
T T
12.2 The Cameron–Martin formula
In this section we see how we can use Girsanov’s Theorem 12.1 in order to construct
weak solutions, in the sense of Definition 9.1, for SDEs that may not satisfy
Assumption (A’).
Let us assume that satisfies Assumption (A’) and, moreover, that it is a
symmetric d d matrix field and that, for every .x; t/, the smallest eigenvalue of
.x; t/ is bounded below by a constant > 0. The last hypothesis implies that the
matrix field .x; t/ 7! .x; t/1 is well defined and bounded. Let b W Rm RC ! Rm
be a bounded measurable vector field. We know, by Theorem 9.4, that there exists a
solution of the SDE
dt D .t ; t/ dBt

(12.12)
0 D x
which, moreover, is unique in law. Let now, for a fixed T > 0,

hZ T
1
Z T i
ZT D exp 1 .s ; s/b.s ; s/ dBs j 1 .s ; s/b.s ; s/j2 ds : (12.13)
0 2 0
As 1 and b are bounded, by Proposition 12.1 and Theorem 12.1 EŒZT D 1 and if
e
P is the probability on .˝; F / with density ZT with respect to P and
Z t
e
Bt D Bt 1 .s ; s/b.s ; s/ ds ;
0
then e
B D .˝; F ; .Ft /0tT ; .e
Bt /0tT ; e
P/ is a Brownian motion; hence (12.12) can
be written as
dt D b.t ; t/ dt C .t ; t/ de

Bt
(12.14)
0 D x :
Therefore, with respect to e

P, is a solution of (12.14).
Theorem 12.3 Under the previous assumptions equation (12.14) admits a

weak solution on Œ0; T for every T and there is uniqueness in law.
Proof The existence has already been proved. As for the uniqueness the idea of
the proof is that for Eq. (12.12) there is uniqueness in law (as Assumption (A’) is
satisfied) and that the law of the solution of (12.14) has a density with respect to the
law of the solution of (12.12).
B0t /0tT ; e
Let .˝ 0 ; F 0 ; .Ft0 /0tT ; .t0 /0tT ; .e P0 / be another solution of (12.14),
i.e.
dt0 D b.t0 ; t/ dt C .t0 ; t/ de

B0t :
We must prove that 0 has the same law as the solution constructed in (12.14). Let
h Z T
1
Z T i
e 0
Z T D exp 1
.s0 ; s/b.s0 ; s/ de
Bs j 1 .s0 ; s/b.s0 ; s/j2 ds
0 2 0
and let P0 be the probability on .˝ 0 ; F 0 / having density e Z 0T with respect to e

P0 . Then,
by Proposition 12.1 and Theorem 12.1, the process
Z t
0 e 0
Bt D Bt C 1 .s0 ; s/b.s0 ; s/ ds
0
is a P0 -Brownian motion. Expressing B0 in terms of e B0 in (12.14), it turns out that

.˝ ; F ; .Ft /0tT ; .t /0tT ; .Bt /0tT ; P / is a solution of (12.12). Moreover, e
0 0 0 0 0 0
P0
0
has density with respect to P given by
hZ T Z
1 T 1 0 i
e0 1
ZT0 D Z T D exp 1 0 0
.s ; s/b.s ; s/ dBs C 0
j .s ; s/b.s0 ; s/j2 ds
2Z 0
hZ T
0
1 T i
D exp 1 .s0 ; s/b.s0 ; s/ de
B0s j 1 .s0 ; s/b.s0 ; s/j2 ds :
0 2 0
12.3 The martingales of the Brownian filtration 381
As Eq. (12.12) satisfies Assumption (A’), there is uniqueness in law and the laws
.P/ and 0 .P0 / coincide. The law of the solution of (12.14) is .P/ (image of P
through ), but
.P/ D .ZT e
P/ :
Also the joint laws of .; ZT / and . 0 ; ZT0 / (Lemma 9.2) coincide and from this it
follows (Exercise 4.3) that the laws .ZT P/ D .e P/ and 0 .ZT0 P/ D 0 .e
P0 / coincide,
which is what had to be proved.
t
u
The arguments of this section stress that, under the hypotheses under consideration,
if is a solution of (12.14), then its law is absolutely continuous with respect to
the law of the solution of (12.12). More precisely, with the notations of the proof of
Theorem 12.3, if A is a Borel set of the paths space C ,
Q
P. 2 A/ D EP .ZT 1f2Ag / ; (12.15)
where ZT is defined in (12.13). (12.15) is the Cameron–Martin formula. In fact,

it holds under hypotheses weaker than those considered here. These are actually
rather restrictive, since often one is led to the consideration of diffusions having
an unbounded drift or a diffusion coefficient that it is not uniformly invertible. In
this case the vector field .x; t/ 7! .x; t/1 b.x; t/ is not bounded and therefore
Proposition 12.1 cannot be used to obtain that E.ZT / D 1; it will be necessary to
turn instead to Proposition 12.2. In Exercise 12.5 the absolute continuity of the laws
of the Ornstein–Uhlenbeck process with respect to Wiener measure is considered:
this is a simple example where the hypotheses of this section do not hold, the drift
being linear.
12.3 The martingales of the Brownian filtration
Let B be an m-dimensional Brownian motion with respect to a filtration .Ft /t . We

have seen in Sect. 8.4 that the stochastic integral of B with respect to an integrand
of M 2 is a square integrable .Ft /t -martingale. Is the converse also true? Is it true
that every square integrable .Ft /t -martingale is a stochastic integral with respect to
some integrand of M 2 ?
We see in this section that the answer is yes if .Ft /t D .G t /t , the natural
filtration augmented with the events of probability zero of F which was introduced
in Remark 3.1. This is a particularly important representation result with deep
applications, some of which will appear in Chap. 13. Now just observe that
this result implies that every .G t /t -martingale is continuous (has a continuous
modification, to be precise). We know from Proposition 4.3 that .G t /t is right-
continuous, so that .˝; F ; .G t /t ; .Bt /t ; P/ is a standard Brownian motion.
The key result is the following, which we have already mentioned in

Remark 7.2.
Theorem 12.4 Let B D .˝; F ; .G t /t ; .Bt /t ; P/ be an m-dimensional Brown-

ian motion with respect to its augmented natural filtration and let T > 0. Then
every r.v. Z 2 L2 D L2 .˝; G T ; P/ is of the form
Z T
Z D cC Hs dBs ; (12.16)
0
where c 2 R and H 2 M 2 .Œ0; T/ (in particular, H is .G t /t -adapted). Moreover,

the representation (12.16) is unique.
Proof The uniqueness is obvious, as c is determined by c D EŒZ whereas, if H1

and H2 were two processes in M 2 .Œ0; T/ satisfying (12.16), then from the relation
Z T
H1 .s/ H2 .s/ dBs D 0
0
we have immediately, by the isometry property of the stochastic integral, that

Z T ˇ ˇ
E ˇH1 .s/ H2 .s/ˇ2 ds D 0
0
and therefore H1 .s/ D H2 .s/ for almost every s 2 Œ0; T a.s.

As for the existence, let us denote by H the space of the r.v.’s of the form (12.16)
and let us prove first that H is a closed subset of L2 . If .Zn /n H is a Cauchy
sequence in L2 and
Z T
Zn D cn C Hn .s/ dBs D cn C In
0
then .cn /n is also a Cauchy sequence, as cn D EŒZn and
jcn cm j E.jZn Zm j/ E.jZn Zm j2 /1=2 :
Therefore both .cn /n and .In /n are Cauchy sequences (in R and in L2 respectively).
As the stochastic integral is an isometry between L2 and M 2 .Œ0; T/ (Theorem 7.1),
it follows that .Hn /n is a Cauchy sequence in M 2 .Œ0; T/. Therefore there exist c 2 R
and H 2 M 2 .Œ0; T/ such that
lim cn D c
n!1
lim Hn D H in M 2 .Œ0; T/
n!1
whence we get
Z T
lim Zn D c C Hs dBs in L2 :
n!1 0
Let us prove now that H is also dense in L2 . If f 2 L2 .Œ0; T; Rm / ( f is therefore a

deterministic function), for t T, and we set
Z t
1
Z t
ef .t/ WD exp f .s/ dBs j f .s/j2 ds
0 2 0
then we know that

Z T
ef .T/ D 1 C ef .s/f .s/ dBs : (12.17)
0
RT
Moreover, ef .T/ is square integrable, as 0 f .s/ dBs is a Gaussian r.v. More precisely
(see also Remark 7.3).
h Z T i Z T
EŒef .T/2 D E exp 2 f .s/ dBs exp jf .s/j2 ds
0 0
Z T Z T Z T
D exp 2 jf .s/j2 ds exp jf .s/j2 ds D exp jf .s/j2 ds :
0 0 0
Therefore
Z T Z T
EŒef .s/2 jf .s/j2 ds EŒef .T/2 jf .s/j2 ds
0 0
Z T Z T
exp jf .s/j2 ds jf .s/j2 ds ;
0 0
which proves that the process Hf .s/ D f .s/ef .s/ belongs to M 2 .Œ0; T/ hence, thanks
to (12.17), the r.v.’s of the form ef .T/ belong to H . Let us prove that they form a
total set, i.e. that their linear combinations are dense in H . To this end let Y 2
L2 .G T / be a r.v. that is orthogonal to each of the r.v.’s ef .T/ and let us prove that
Y D 0 a.s.
First, choosing f D 0 so that ef .T/ D 1 a.s., Y D Y C Y must have mean
zero and therefore EŒY C D EŒY . If both these mathematical expectations are
equal to 0, Y vanishes and there is nothing to prove. Otherwise we can multiply Y
by a constant so that E.Y C / D E.Y / D 1. The remainder of the proof consists
in checking that the two probabilities Y C dP and Y dP coincide on .˝; G T /, which
will imply Y C D Y and therefore Y D 0 a.s.
P
If f .s/ D njD1 j 1Œtj1 ;tj Œ .s/, where 0 D t0 < t1 < < tn D T and 1 ; : : : ; n 2
RT P
Rm , then 0 f .s/ dBs D njD1 h j ; Btj Btj1 i. We have
Xn
1X
n
ef .T/ D exp h j ; Btj Btj1 i j j j2 .tj tj1 /
jD1
2 jD1
and orthogonality of Y with respect to ef .T/ implies

h X
n i
0 D E Y exp h j ; Btj Btj1 i ; (12.18)
jD1
i.e.
h X
n i h X
n i
C
E Y exp h j ; Btj Btj1 i D E Y exp h j ; Btj Btj1 i :
jD1 jD1
Allowing the vectors 1 ; : : : ; n to take all possible values, this implies that the
laws of the random vector .Bt1 ; Bt2 Bt1 ; : : : ; Btn Btn1 / with respect to the two
probabilities Y C dP and Y dP have the same Laplace transforms and therefore
coincide (see Sect. 5.7 for more details). Recalling the definition of the law of a
r.v., this also implies that

E Y C ˚.Bt1 ; Bt2 Bt1 ; : : : ; Btn Btn1 / D E Y ˚.Bt1 ; Bt2 Bt1 ; : : : ; Btn Btn1 /
for every bounded measurable function ˚, from which we deduce that Y C dP and
Y dP coincide on the -algebra .Bt1 ; Bt2 Bt1 ; : : : ; Btn Btn1 /, which is equal to
.Bt1 ; : : : ; Btn /.
As the union of these -algebras for all possible choices of n, and of 0 D
t0 < t1 < < tn D T, forms a family that is stable with respect to finite
intersections and generates GT , the two probabilities Y C dP and Y dP coincide on
GT by Carathéodory’s criterion, Theorem 1.1. They also coincide on G T (just repeat
the argument of Remark 4.4): let C be the class of the events of the form A \ N
with A 2 GT and N 2 F is either a negligible set or N D ˝. Then C is stable with
respect to finite intersections, contains both GT and the negligible sets of F and
therefore generates G T . Moreover, the two probabilities Y C dP and Y dP coincide
on C and therefore also on G T , again by Carathéodory’s criterion.
t
u
An immediate consequence is the following
Theorem 12.5 Let .Mt /0tT be a square integrable martingale of the filtra-
tion .G t /t . Then there exist a unique process H 2 M 2 .Œ0; T/ and a constant
(continued)

c 2 R such that
Z t
Mt D c C Hs dBs a.s.
0
for t 2 Œ0; T. In particular, .Mt /0tT has a continuous modification.
Proof As MT 2 L2 , by Theorem 12.4 there exists a unique process H 2 M 2 .Œ0; T/

such that
Z T
MT D c C Hs dBs
0
and therefore
Z t
Mt D E.MT jG t / D c C Hs dBs a:s:
0
t
u
Note that in the statement of Theorem 12.5 we make no assumption of continuity.
Therefore every square integrable martingale of the filtration .G t /t always admits a
continuous version.
The representation Theorem 12.5 can be extended to local martingales.
Theorem 12.6 Let .Mt /0tT be a local martingale of the filtration .G t /t .

2
Then there exists a unique process H 2 Mloc .Œ0; T/ and a constant c 2 R
such that, for t 2 Œ0; T,
Z t
Mt D c C Hs dBs a.s.
0
Proof The idea of the proof is to approximate M with square integrable martingales,
but in order to do this properly we first need to prove that .Mt /0tT is itself
continuous, or, to be precise, that it has a continuous modification. Let us first
assume that .Mt /0tT is a martingale of the filtration .G t /t (not necessarily square
integrable). As MT is integrable and L2 .G T / is dense in L1 .G T /, let .Zn /n be a
sequence of r.v.’s in L2 .G T / such that
kZn MT k1 2n :
Let Mn .t/ D E.Zn jG t /; Mn is a square integrable martingale for every n and

has a continuous modification by Theorem 12.5. .Mt /0tT has a right continuous
modification thanks to Theorem 5.14. Then the supermartingale .jMt Mn .t/j/t is
right-continuous and we can apply the maximal inequality (5.16), which gives
1 1
P sup jMt Mn .t/j D P inf jMt Mn .t/j
0tT k 0tT k
kE.jMT Zn j/ k2n :
As we have on the right-hand side the general term of a convergent series, by the
Borel–Cantelli lemma,
1
sup jMt Mn .t/j <
0tT k
eventually a.s. In other words, .Mn .t//t converges a.s. uniformly to .Mt /t , which is
therefore a.s. continuous.
If M is a local martingale of the filtration .G t /t and .n /n is a sequence of
reducing stopping times, then the stopped process M n is a martingale. Therefore,
M is continuous for t < n for every n and, as limn!1 n D C1, M is continuous.
The fact that M is continuous allows us to apply the argument of Remark 7.6 and
we can assume that the sequence .n /n is such that the stopped processes M n are
bounded martingales, and therefore square integrable. By Theorem 12.5, for every
n there exist cn 2 R and a process H .n/ 2 M 2 .Œ0; T/ such that
Z t
Mtn D Mt^n D cn C Hs.n/ dBs a.s.
0
def
Obviously cn D cnC1 D M0 D c. Moreover, the two processes M n and M nC1
coincide for t n , therefore
Z n Z n
Hs.n/ dBs D Hs.nC1/ dBs a.s.
0 0
By the uniqueness of the representation of the r.v.’s of L2 .G T / of Theorem 12.4,

.n/ .nC1/
it follows that the two processes .Ht^n /t and .Ht^n /t are modifications of one
.n/
another. This allows us to define a process H by setting Ht D Ht on ft < n g, as
this definition does not depend on the choice of n. The process thus defined is such
that
Z n
E Hs2 ds < C1
0
12.4 Equivalent probability measures 387
RT
for every n; as n ! C1 as n ! 1, this implies 0 Hs2 ds < C1 a.s., and
2
therefore H 2 Mloc .Œ0; T/. We still have to prove that
Z t
Mt D c C Hs dBs a.s.;
0
but
Z t^n Z t^n
Mt^n D c C Hs.n/ dBs D c C Hs dBs a.s.
0 0
and we can take the limit as n ! 1.

t
u
12.4 Equivalent probability measures
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian motion. Girsanov’s

Theorem 12.1 states that if .Zt /t is a martingale of the form
hZ t
1
Z t i
Zt D exp ˚s dBs j˚s j2 ds ;
0 2 0
2
where .˚t /t is an Rm -valued process in Mloc .Œ0; T/, then dQ D ZT dP is a
probability on .˝; FT / that is equivalent to P.
Let us show now that if Ft D G t , then all the probabilities on .˝; FT / that are
equivalent to P are of this form.
Theorem 12.7 Let B D .˝; F ; .G t /t ; .Bt /t ; P/ be an m-dimensional Brow-

nian motion with respect to its natural augmented filtration .G t /t . If a
probability Q on .˝; G T / is equivalent to P then there exists a progressively
2
measurable m-dimensional process ˚ 2 Mloc .Œ0; T/ such that dQ D ZT dP,
where
hZ t Z
1 t i
Zt D exp ˚s dBs j˚s j2 ds ; (12.19)
0 2 0
Rt
is a martingale under P. Moreover, e
Bt D Bt 0 ˚s ds is a G t -Brownian motion
under Q.
Proof Let
dQjG t
Zt D
dPjG t
We know that .Zt /0tT is a martingale. Obviously we have, for every t T,
Q.Zt > 0/ D EP .Zt 1fZt >0g / D EP .Zt / D 1 :
Therefore Zt > 0 Q-a.s. and also P-a.s., as P and Q are assumed to be equivalent.
2
By Theorem 12.6 there exists a process .t /t 2 Mloc .Œ0; T/ such that
Z t
Zt D 1 C s dBs :
0
We want now to apply Ito’s formula in order to compute d log Zt . This is not possible
directly, as log is not a C2 function on R. Let f be a function such that f .x/ D log x
for x 1n and then extended to R so that it is of class C2 . Then, by Ito’s formula,
1
df .Zt / D f 0 .Zt / dZt C f 00 .Zt /jt j2 dt :
2
The derivatives of f coincide with those of log x for x 1n , i.e. f 0 .x/ D 1x , f 00 .x/ D
x12 , so that, if n D infftI Zt 1n g,
Z Z
t^n
0 1 t^n 00 2
log Zt^n D f .Zs /s dBs C f .Zs /j sj ds
0 2 0
Z t^n Z
s 1 t^n js j2
D dBs ds :
0 Zs 2 0 Zs2
By Exercise 5.8 b), Z being a martingale, we have P.Zt > 0 for every t T/ D 1
for every t T and therefore t 7! Zt .!/ never vanishes a.s. Therefore, Z being
continuous, for every ! 2 ˝ there exists an " > 0 such that Zt .!/ " for every
t T. Therefore ˚s D Zss 2 Mloc
2
.Œ0; T/ and
Z t^n
1
Z t^n
Zt^n D exp ˚s dBs ˚s2 ds :
0 2 0
Now just let n ! 1 and observe that, again as Zt > 0 for every t a.s., n ! C1.
t
u
Exercises
12.1 (p. 603) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a Brownian motion. Let us
consider the three processes
.Bt /t ; Yt D Bt C ct; Zt D Bt ;
where c; are real numbers. On the canonical space .C ; M ; .Mt /t ; .Xt /t / let us
consider the probabilities PB ; PY ; PZ , respectively the law of B (i.e. Wiener measure)
and the laws of the processes Y and Z. Then
a) PB and PY are equivalent (PB PY and PY PB ) on Mt for every t 0, but,
unless c D 0, not on M where they are actually orthogonal.
b) If jj 6D 1 then PB and PZ are orthogonal on Mt for every t > 0.
a) Use Girsanov’s theorem in order to find a probability Q of the form dQ D Z dP with respect
to which .Bt /t has the same law as .Yt /t . b) Look for an event having probability 1 for PB and
probability 0 for PZ .
12.2 (p. 604) Given two probabilities ; on a measurable space .E; E /, the
entropy H.I / of with respect to is the quantity
8Z
< d log d d if
H.I / D E d d
:
C1 otherwise :
Note that we also have, for ,

Z
d
H.I / D log d : (12.20)
E d
The chi-square discrepancy of with respect to is the quantity

8Z ˇ ˇ
< ˇ d 1ˇ2 d if
2
.I / D E d
:
C1 otherwise :
a) Show that H.I / 0 and H.I / D 0 if and only if D . Show, with an

example, that it is possible that H.I / 6D H.I /.
b) Let C D C .Œ0; T; Rm / and let .C ; MT ; .Ft /0tT ; .Xt /0tT ; P/ be the
canonical space endowed with the Wiener measure P. Let be a path almost
a.e. differentiable and with square integrable derivative and let P1 be the law of
Wt D Xt C t , i.e. of a Brownian motion with a deterministic drift .
b1) What is the value of the entropy H.P1 I P/? And of H.PI P1 /?
b2) Compute 2 .P1 I P/ and 2 .PI P1 /.
a) Use Jensen’s inequality, since x 7! x log x is a strictly convex function. b1) Girsanov’s theorem
gives the density of P1 with respect to P. b2) Look first for a more handy expression of the 2
discrepancy (develop the square inside the integral).
Xt D Bt
t;
> 0. As limt!C1 Xt D 1 (thanks, for instance, to the Iterated
Logarithm Law), we know that supt>0 Xt < C1. In this exercise we use Girsanov’s
theorem in order to compute the law of supt>0 Xt . We shall find, through a different
argument, the same result as in Exercise 5.17.
The idea is to compute the probability of the event fsupt>0 Xt > Rg with a change
of probability such that, with respect to the new probability, it has probability 1 and
then to “compensate” with the density of the new probability with respect to the
Wiener measure P.
2
a) Let Zt D e2
Bt 2
t . Show that .Zt /t is an .Ft /t -martingale and that if, for a fixed
T > 0,
dQ D ZT dP
Bt D Bt 2
t, then .˝; F ; .Ft /0tT ; .e
then Q is a probability and, if e Bt /0tT ; Q/
is a Brownian motion.
b) Show that .Zt1 /t is a Q-martingale. Note that Zt1 D e2
Xt .
1
c) Let R > 0 and R D infftI Xt D Rg. Show that P.R T/ D EQ .1fR <Tg ZT^ R
/
and
P.R T/ D e2
R Q.R T/ : (12.21)
d) Show that P.R < C1/ D e2

R , hence the r.v. supt>0 Xt has an exponential law
with parameter 2
.
c) fR Tg 2 FT^R and we can apply the stopping theorem to the Q-martingale .Zt1 /t .
12.4 (p. 607) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a standard Brownian motion. In
Exercise 8.14 it is proved that
Z Z
1 B2t t
Bs 1 t
B2s
Zt D p exp D exp dBs ds
1t 2.1 t/ 0 1s 2 0 .1 s/2
is a martingale for t 2 Œ0; T for every T < 1. Let Q be a new probability on .˝; F /
defined as dQ D ZT dP. Show that, with respect to Q, .Bt /t is a process that we have
already seen many times.
12.5 (p. 607) Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be an m-dimensional Brownian
motion. The aim of this exercise is to compute
h Z t i
J D E exp jXs j2 ds :
0
It may help to first have a look at Exercise 1.12 and, for d), at the properties of the
Laplace transform in Sect. 5.7.
a) Let
Z t Z

2 t
Zt D exp
Xs dXs jXs j2 ds :
0 2 0
Show that E.Zt / D 1 for every t 0.

b) Let Q be the probability on .˝; FT / defined as dQ D ZT dP. Prove that, with
respect to Q, on the time interval Œ0; T X is an Ornstein–Uhlenbeck process (see
Sect. 9.2). What is the mean and the covariance matrix of Xt , 0 t T, with
respect to Q?
2
c) Prove that, if D
2 and t T,

2
J D EQ e 2 .jXt j mt/
and determine the value of J for 0.

d) Compute J for every 2 R.
12.6 (p. 609) Let X D .˝; F ; .Ft /t ; .Xt /t ; P/ be a real Brownian motion.
a1) Let b W R ! R be a bounded continuously differentiable function and x a fixed
real number. Determine a newR probability Q on .˝; F / such that, with respect
t
to Q, the process Bt D Xt 0 b.Xs C x/ ds is a Brownian motion for t T.
Prove that, with respect to Q, the process Yt D x C Xt is the solution of an SDE
to be determined.
a2) Let U be a primitive of b. Prove that dQ D ZT dP with
Z
1 t 0
Zt D exp U.Xt C x/ U.x/ Œb .Xs C x/ C b2 .Xs C x/ ds : (12.22)
2 0
b1) Let b.z/ D k tanh.kz C c/ for some constant k. Prove that b0 .z/ C b2 .z/ k2 .
b2) Let Y be the solution of
dYt D k tanh.kYt C c/ dt C dBt

(12.23)
Y0 D x :
Compute the Laplace transform of Yt and show that the law of Yt is a mixture of
Gaussian laws, i.e. of the form ˛1 C .1 ˛/2 , where 0 < ˛ < 1 and 1 ; 2
are Gaussian laws to be determined.
b3) Compute EŒYt .
b2) A primitive of z 7! tanh z is z 7! log cosh z.
12.7 (p. 611) (Wiener measure gives positive mass to every open set of C0 ) Let as
usual C D C .Œ0; T; Rm / endowed with the topology of uniform convergence and
let PW be the Wiener measure on C . Let us denote by C0 the closed subspace of C
formed by the paths such that 0 D 0.
a) Show that PW .C0 / D 1.
b) Recall that, for a real Brownian motion, if denotes the exit time from Œr; r,
r > 0, then the event f > Tg has positive probability for every T > 0
(Exercise 10.5 c)). Deduce that PW .A/ > 0 for every open set A C containing
the path 0. Rt
c) Note that the paths of the form t D 0 ˚s ds, with ˚ 2 L2 .RC ; Rm /, are dense
in C0 (they form a subset of the paths that are twice differentiable that are dense
themselves). Deduce that PW .A/ > 0 for every open set A C0 .
b) A neighborhood of the path
0 is of the form V D fwI sup0tT jwt j < g for some
T; > 0. c) Use Girsanov’s formula to “translate” the open set A to be a neighborhood of the
origin.
12.8 (p. 612) In this exercise we use the Feynman–Kac formula in order to find
explicitly the solution of the problem
8
< 1 4u.x; t/ C @u .x; t/ jxj2 u.x; t/ D 0 if .x; t/ 2 Rm Œ0; TŒ
2 @t (12.24)
:
u.x; T/ D 1 ;
where 0. Let X D .˝; F ; .Ft /t ; .Bt /t ; P/ be a m-dimensional Brownian

motion. By Theorem 10.6, a solution of (12.24) is given by
h RT 2
i
u.x; t/ D Ex;t e t jXs j ds ;
where .˝; F ; .Ft /t ; .Xt /t ; .Px;t /x;t / is the canonical diffusion associated to the
operator L D 12 4. We know that, with respect to Px;t , the canonical process .Xs /st
has the same law as .Bst C x/t , where .Bt /t is a Brownian motion. Hence for x D 0,
t D 0 we shall recover the result of Exercise 12.5.
a) For x 2 Rm and
2 R, let
Z t Z

2 t
Zt D exp
.Bs C x/ dBs jBs C xj2 ds :
0 2 0
Show that E.ZT / D 1.

b) Let Q be the probability on .˝; FT / defined as dQ D ZT dP. Prove that, with
respect to Q, Ys D Bs Cx is an Ornstein–Uhlenbeck process (see Sect. 9.2). What
is the mean and the covariance matrix of BT C x with respect to Q?
c) Prove that, under P,

Z T
1
.Bs C x/ dBs D .jBT C xj2 jxj2 mT/ (12.25)
0 2
and deduce that, for every

0,
h
2 Z T i
2
h
2
i
E exp jBs C xj2 ds D e 2 .mTCjxj / EQ e 2 jBT Cxj :
2 0
d) Prove (or take as granted) that, if W is an m-dimensional N.b; 2 I/-distributed

r.v., then
2

E.e
jWj / D .1 2 2
/m=2 exp jbj 2
(12.26)
1 2 2
for every
< .2 2 /1 .
2 RT
e) What is the value of E exp

2 0 jBs C xj2 ds ?
f) Derive from b) and e) that a solution of (12.24) is
h p
p m=2 2 jxj2 p i
u.x; t/ D cosh 2 .T t/ exp tanh 2 .T t/ :
2
Is it unique?
a) Use the criterion of Proposition 12.2 b) and Exercise 1.12.
12.9 (p. 615) Let B be an m-dimensional Brownian motion, T > 0 and f W Rm ! R

a bounded measurable function. We know (Theorem 12.4) that there exists a process
X 2 M 2 such that
Z T
f .BT / D EŒ f .BT / C Xs dBs : (12.27)
0
In this exercise we determine X explicitly (Theorem 12.4 is not necessary).

a) Determine a function W Œ0; TŒRm ! R such that
EŒ f .BT /jFt D .Bt ; t/
and show that is actually differentiable infinitely many times on Rm Œ0; TŒ.
What is the value of .0; 0/?
b) Write the stochastic differential of Zt D .Bt ; t/. Because of what possible
reason can you state that
@ 1
C 4 D0
@t 2
without actually computing the derivatives? Prove that, for every t < T,
Z t
0
.Bt ; t/ D .0; 0/ C x .Bs ; s/ dBs : (12.28)
0
c1) Determine a process X 2 M 2 .Œ0; T/ such that (12.27) holds.

c2) Determine X explicitly if the dimension of B is m D 1 and f .x/ D 1fx>0g .
d) Prove that if, in addition, f is differentiable with bounded first derivatives,
then (12.27) holds with Xs D EŒ f 0 .BT /jFs .
Chapter 13
An Application: Finance
Stochastic processes are useful as models of random phenomena, among which a

particularly interesting instance is given by the evolution of the values of financial
(stocks, bonds, . . . ) and monetary assets listed on the Stock Exchange.
In this chapter we develop some models adapted for these situation and we
discuss their application to some typical problems.
13.1 Stocks and options
In a stock exchange, besides the more traditional stocks, bonds and commodities,
there are plenty of securities or derivative securities which are quoted and traded. A
derivative security (also called a contingent claim), as opposed to a primary (stock,
bond,. . . ) security, is a security whose value depends on the prices of other assets of
the market.
An option is an example of a derivative: a call option is a contract which

guarantees to its holder the right (but not the obligation) to buy a given asset
(stock, commodity, currency, . . . ), which we shall refer to as the underlying
asset, at a given time T (the maturity) and at a price K fixed in advance (the
strike price).
These types of contracts have an old history and they became increasingly practised
at the end of the 60s with the institution of the Chicago Board Options Exchange. A
call option is obviously intended to guarantee the holder of being able to acquire the
underlying asset at a price that is not larger than K (and therefore being safe from
market fluctuations).

396 13 An Application: Finance
If at maturity the underlying asset has a price greater than K, the holder of the
option will exercise his right and will obtain the asset at the price K. Otherwise he
will choose to drop the option and buy the asset on the market at a lower price.
To be precise, assume that the call option is written on a single asset whose price
S is modeled by some process .˝; F ; .Ft /t ; .St /t ; P/. Then the value Z that the
issuer of the option has to pay at maturity is equal to 0 if ST K (the option is not
exercised) and to ST K if ST > K. In short it is equal to .ST K/C . This quantity
.ST K/C is the payoff of the call option.
Many other kinds of options exist. For instance put options, which guarantee the
owner the right to sell at a maturity time T a certain asset at a price not lower than K.
In this case, the issuer of the option is bound to pay an amount K ST if the price ST
at time T is smaller than the strike K and 0 otherwise. The payoff of the put option
is therefore .K ST /C . Other examples are considered later (see Example 13.3 and
Exercise 13.4, for instance)
A problem of interest, since the appearance of these derivatives on the market, is
to evaluate the right price of an option. Actually the issuer of the option faces a risk:
in the case of a call option, if at time T the price of the stipulated asset turns out to
be greater than the strike price K, he would be compelled to hand to the owner the
amount ST K. How much should the issuer be paid in order to compensate the risk
that he is going to face?
A second important question, also connected with the determination of the price,
concerns the strategy of the issuer in order to protect himself from a loss (to
“hedge”).
Put and call options are examples of the so-called European options: each of these
is characterized by its maturity and its payoff. One can formalize this as follows
Definition 13.1 A European option Z with maturity T is a pair .Z; T/, where
T is the maturity date and Z, the payoff, is a non-negative FT -measurable r.v.
In the case of calls and puts the payoff is a function of the value of the underlying
asset at the maturity T. More generally, an option can be a functional of the whole
price process up to time T and there are examples of options of interest from
the financial point of view which are of this kind (see again Example 13.3 and
Exercise 13.4, for instance).
There are also other kinds of options, such as American options, which differ
from the European ones in the exercise date: they can be exercised at any instant
t T. Their treatment, however, requires tools that are beyond the scope of this
book.
In the next sections we develop the key arguments leading to the determination
of the fair price of an option. We shall also discuss which stochastic processes might
be reasonable models for the evolution of the price of the underlying asset. In the
last section, Sect. 13.6, we shall go deeper into the investigation of the most classical
13.2 Trading strategies and arbitrage 397
model in mathematical finance, the Black-Scholes model, and we shall be able to

derive explicit formulas for the price of the options and for some related quantities.
Constructing the models, we shall take into account the following facts.
• There are no transaction costs in buying or selling assets and no taxes.

• Short-selling is admitted, i.e. it is possible to sell assets that are not actually
in possession of the vendor: investors who do not own a stock can sell
shares of it and arrange with the owner at some future date to be payed an
amount equal to the price at that date.
• It is possible to buy or sell fractions of a security.
These facts constitute an ideal market that does not exist in real life. Hence
our models must be considered as a first approximation of the real markets. In
particular, models taking into account transaction costs (as in real markets) have
been developed, but they introduce additional complications and it is wiser to start
with our simple model in order to clarify the main ideas.
Some financial slang:

• holder of an option: the individual/company that has acquired the rights of
the option;
• issuer of an option: the individual/company that has issued the option and
has acquired the obligations connected to it;
• to be short (of something): this is said of someone who has short sold
something.
13.2 Trading strategies and arbitrage
Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be a d-dimensional Brownian motion.
Throughout this chapter we assume that Ft D G t , i.e. that the filtration .Ft /t
is the natural augmented filtration of B (see p. 32 and Sect. 4.5)
Actually it will be clear that this assumption is crucial in many instances.

We shall consider a market where m C 1 assets are present and we shall denote
their prices by S0 ; S1 ; : : : ; Sm . We shall, by now, assume that the prices S1 ; : : : ; Sm
are Ito processes with stochastic differentials
X
d
dSi .t/ D Ai .t/ dt C Gij .t/ dBj .t/ ; (13.1)
jD1
where Ai , Gij are continuous adapted processes. We shall assume, moreover, that the
solution S is such that Si .t/ 0 a.s for every t 2 Œ0; T. This is, of course, a condition
that must be satisfied by every good model . The process S0 will have differential
dS0 .t/ D rt S0 .t/ dt ; (13.2)
where .rt /t is a non-negative bounded progressively measurable process (the spot

interest rate). If we assume S0 .s/ D 1 then
Rt
S0 .t/ D e s ru du : (13.3)
Typically S0 is an investment of the type of treasury bonds or money deposited into

a bank account: it is the riskless asset (even if its evolution is however random, as
here we assume that the spot interest rate is a stochastic process).
The model (13.1), (13.2) is very general and we shall see later more explicit
choices for the processes Ai , Gij and r.
An investor may decide to invest an amount of money by acquiring shares of the
m C 1 assets of the market.
A trading strategy (or simply, a strategy) over the trading interval

Œ0; T is a progressively measurable .m C 1/-dimensional process Ht D
.H0 .t/; H1 .t/; : : : ; Hm .t// whose general component Hi .t/ stands for the
number of units of the i-th security held by an investor at time t. The portfolio
associated to the strategy H is the corresponding wealth process:
X
m
Vt .H/ D hHt ; St i D Hi .t/ Si .t/; t 2 Œ0; T : (13.4)
iD0
The initial value of the portfolio V0 .H/ represents the initial investment of the
strategy H.
At any moment the investor may decide to move part of his wealth from one asset
to another. A particularly important type of trading strategy, from our point of view,
is one in which he does not add or remove capital from the portfolio. The rigorous
definition is given below.
We shall assume that the trading strategy Ht D .H0 .t/; H1 .t/; : : : ; Hm .t// satisfies
the condition
Z T m Z
X T
jH0 .t/j dt C jHi .t/j2 dt < 1 a.s. (13.5)
0 iD1 0
The trading strategy H is said to be self-financing over the time interval Œ0; T
if it satisfies (13.5) and its associated portfolio Vt .H/ satisfies the relation
Xm
dVt .H/ D hHt ; dSt i D Hi .t/ dSi .t/; 0 t T : (13.6)
iD0
Therefore a strategy is self-financing if the variations of the associated portfolio,

in a small time period, depend only on increments of the asset prices S0 ; : : : ; Sm ,
i.e. changes in the portfolio are due to capital gains or losses and not to increase or
decrease of the invested funds.
Notice that the requirements (13.5) are technical and are necessary for the
differentials appearing in (13.6) to be well defined. In fact, as the processes t 7!
Ai .t/, t 7! Gij .t/ and t 7! rt are bounded on Œ0; T for every ! (Ai and Gij
1 1
have continuous paths), then H0 r 2 Mloc .Œ0; T/, Hi Ai 2 Mloc .Œ0; T/ for every
2
i D 1; : : : ; m, and also Hi Gij 2 Mloc .Œ0; T/ for every i; j (see also Remark 7.1 a)).
Of course (13.6) can be written as
X
m X
m X
d
dVt .H/ D H0 .t/ rt St0 dt C Hi .t/Ai .t/dt C Hi .t/ Gij .t/ dBj .t/ :
iD1 iD1 jD1
In particular, the portfolio V.H/ associated to a self-financing strategy H is a Ito

process. Let
Si .t/ Rt
e
Si .t/ D D e 0 rs ds Si .t/; i D 1; : : : ; m
S0 .t/
Vt .H/ Rt
e
V t .H/ D D e 0 rs ds Vt .H/ :
S0 .t/
P
Notice that, thanks to (13.4), e V t .H/ D H0 .t/ C m e
iD1 Hi .t/ Si .t/. We shall refer
e
to St D .S1 .t/; : : : ; Sm .t// as the discounted price process and to e
e e V.H/ as the
discounted portfolio. Intuitively, e Si .t/ is the amount of money that must be invested
at time 0 into the riskless asset in order to have the amount Si .t/ at time t. Note that
Rt
by Ito’s formula, as t 7! 0 rs ds has finite variation,
Rt Rt
de
Si .t/ D rt e 0 rs ds Si .t/ dt
Rt
C e 0 rs ds dSi .t/
(13.7)
D rteSi .t/ dt C e 0 rs ds dSi .t/ :
The following result expresses the property of being self-financing in terms of the
discounted portfolio.
Proposition 13.1 Let Ht D .H0 .t/; H1 .t/; : : : ; Hm .t// be a trading strategy

satisfying (13.5). Then H is self-financing if and only if
m Z
X t
e
V t .H/ D V0 .H/ C Hi .t/ de
Si .t/; t 2 Œ0; T : (13.8)
iD1 0
Proof Suppose that H is self-financing. Then, by Ito’s formula and (13.7),

Rt Rt Rt
de
V t .H/ D d e 0 rs ds Vt .H/ D rt e 0 rs ds Vt .H/ dt C e 0 rs ds dVt .H/
Rt X
m X
m
D e 0 rs ds
rt Hi .t/ Si .t/ dt C Hi .t/ dSi .t/
iD0 iD0
but, as H0 .t/dS0 .t/ D rt H0 .t/S0 .t/, the terms with index i D 0 cancel and we have
Rt X
m X
m
D e 0 rs ds rt Hi .t/ Si .t/ dt C Hi .t/ dSi .t/
iD1 iD1
X
m Rt Rt
D Hi .t/ rt e 0 rs ds Si .t/ dt C e 0 rs ds dSi .t/
„ ƒ‚ …
iD1
DdeSi .t/
and, as e
V 0 .H/ D V0 .H/, (13.8) holds. Conversely, if e
V.H/ satisfies (13.8), again by
Ito’s formula and (13.7)
Rt Rt Rt X
m
dVt .H/ D d e 0 rs dse
V t .H/ D rt e 0 rs dse
V t .H/ dt C e 0 rs ds Hi .t/ de
Si .t/
iD1
Rt X
m Rt
D rt Vt .H/ dt C e 0 rs ds Hi .t/ rte
Si .t/ dt C e 0 rs ds dSi .t/
iD1
X
m X
m
D rt Vt .H/ dt rt Hi .t/ Si .t/ dt C Hi .t/ dSi .t/
iD1 iD1
and again using the relation H0 .t/dS0 .t/ D rt H0 .t/S0 .t/,
X
m X
m X
m
D rt Vt .H/ dt rt Hi .t/ Si .t/ dt C Hi .t/ dSi .t/ D Hi .t/ dSi .t/ ;
iD0 iD0 iD0
„ ƒ‚ …
DVt .H/
i.e. H is self financing. t

u
Let us now introduce the following sub-class of self-financing strategies.
A self-financing strategy H is said to be admissible if Vt .H/ 0 for every t,

a.s.
To be precise, note that the processes Hi can take negative values corresponding to
short selling of the corresponding assets. In order for a strategy to be admissible it
is required, however, that the overall wealth Vt .H/ of the portfolio remains 0 for
every t (i.e. that the investor is solvable at all times).
A self-financing trading strategy H over Œ0; T is said to be an arbitrage

strategy if the associated portfolio Vt .H/ satisfies
V0 .H/ D 0
Vt .H/ 0 a.s. for every t T (13.9)
P.VT .H/ > 0/ > 0 :
Roughly speaking, an arbitrage portfolio requires no initial capital (this is the

assumption V0 .H/ D 0), its value is never negative (i.e. its admissible) and can
produce a gain with strictly positive probability.
As one can imagine, an arbitrage is an operation that is seldom possible. It can
happen for instance that, keeping in mind the parity between euro and yen, the dollar
is cheaper in Tokyo than in Milano. An operator might then buy dollars in Tokyo
and sell them in Milano; he could pay the purchase in Japan with the money from
the sale in Italy and make a gain with no risk and without the need of any capital.
Such a situation is therefore an arbitrage as it satisfies the three conditions (13.9).
In the real market these situations do appear, but seldom and in any case during a
very short time span: in the example above purchases in Tokyo and sales in Milano
would quickly provoke a raise of the exchange rate in Japan and a drop in Italy, thus
closing the possibility of arbitrage.
Therefore it is commonly assumed that in a reasonable model no arbitrage
strategy should be possible.
Definition 13.2 A market model is said to be arbitrage-free if every admis-

sible strategy H on Œ0; T with V0 .H/ D 0 is such that P.VT .H/ > 0/ D 0.
We shall see in the next section that the arbitrage-free property is equivalent to an
important mathematical property of the model.
13.3 Equivalent martingale measures
Definition 13.3 A probability P on .˝; FT / is called an equivalent martin-

gale measure if P is equivalent to P on FT and the discounted price processes
e
S1 .t/; : : : ; e
Sm .t/ are .Ft /t -martingales under P .
Equivalent martingale measures play a very important role in our analysis. Does an
equivalent martingale measure exist for the model (13.1)? Is it unique? We shall
investigate these questions later. In this section and in the next one we shall point
out the consequences of the existence and uniqueness of an equivalent martingale
measure.
Thanks to Theorem 12.7, if P is an equivalent martingale measure, there exists
2
a progressively measurable process ˚ 2 Mloc .Œ0; T/ such that
dPjFT RT
˚s dBs 12
RT
j˚s j2 ds
De 0 0 :
dPjFT
Therefore by Girsanov’s Theorem 12.1 the process

Z t
Bt D Bt ˚s ds
0
13.3 Equivalent martingale measures 403
is a P -Brownian motion and, recalling (13.1) and (13.7), under P the discounted
prices have a differential
de
Si .t/
Rt Rt X
d
D rte
Si .t/ C e 0 rs ds Ai .t/ C e 0 rs ds Gij .t/˚j .t/ dt
jD1 (13.10)
„ ƒ‚ …
Rt X
d
Ce 0 rs ds
Gij .t/ dBj .t/ :
jD1
Therefore the prices, which are supposed to be Ito processes under the “old”
probability P, are also Ito processes under P . Note also that properties of the trading
strategies as being self-financed, admissible or arbitrage are preserved under the new
probability.
The requirement that the components of e S are martingales dictates that the
quantity indicated by the brace in (13.10) must vanish. The following proposition is
almost obvious.
Proposition 13.2 If there exists an equivalent martingale measure P , the

discounted portfolio e V.H/ associated to any self-financing strategy H is a
local martingale under P . Moreover, if H is admissible on Œ0; T, then e
V.H/
is an .Ft /t -supermartingale under P on Œ0; T.
Proof Let H be a self-financing strategy. Then by Proposition 13.1 and (13.10),

under P
X
m Rt X
m X
d
de
V t .H/ D Hi .t/ de
Si .t/ D e 0 rs ds Hi .t/ Gij .t/ dBj .t/ ;
iD1 iD1 jD1
for t T. As t 7! Gij .t; !/ is continuous and therefore bounded for every i and j on
2
Œ0; T, and Hi 2 Mloc .Œ0; T/ for every i (recall that this condition is required in the
2
definition of a self-financing strategy), it follows that Hi Gij 2 Mloc .Œ0; T/ for every
i D 1; : : : ; m and therefore e
V.H/ is a local martingale on Œ0; T.
If H is admissible, then V.H/ is self-financing and such that Vt .H/ 0 a.s.
for every t under P. As P is equivalent to P, then also e V t .H/ 0 a.s. under P .
e
Hence under P , V.H/ is a positive local martingale and therefore a supermartingale
(Proposition 7.5). t
u
Note that, thanks to (13.10), under P the (undiscounted) price process S follows
the SDE
Rt Rt Rt
dSi .t/ D d e 0 rs dse
Si .t/ D rt e 0 rs dse
Si .t/ dt C e 0 rs ds de
Si .t/
Xd
(13.11)
D rt Si .t/ dt C Gij dBj .t/ :
jD1
In particular, the drift t 7! Ai .t/ is replaced by t 7! rt Si .t/ and the evolution of the
prices under P does not depend on the processes Ai .
The next statement explains the importance of equivalent martingale measures
and their relation with arbitrage.
Proposition 13.3 If there exists an equivalent martingale measure P , the

market model is arbitrage-free.
Proof We must prove that for every admissible strategy H over Œ0; T such that
V0 .H/ D 0 a.s. we must have VT .H/ D 0 a.s. By Proposition 13.2, eV t .H/ is a
non-negative supermartingale under P , hence
0 E Œe
V T .H/ E Œe
V 0 .H/ D 0:
Therefore, as VT .H/ 0, necessarily P .e

V T .H/ > 0/ D 0, and then, as P and P
are equivalent, also P.VT .H/ > 0/ D 0. t
u
Often we shall need the discounted portfolio to be a true martingale and not just a
local one. Thus, we introduce the following definition, which requires the existence
of an equivalent martingale measure P .
Let P be an equivalent martingale measure. For T > 0, MT .P / will denote

the class of the admissible strategies H on Œ0; T such that the associated
discounted portfolio e
V.H/ is an .Ft /t -P -martingale on Œ0; T.
13.4 Replicating strategies and market completeness
Definition 13.4 Assume that an equivalent martingale measure P exists. We

say that a European option .Z; T/ is attainable if there exists a strategy H 2
MT .P / such that VT .H/ D Z. Such a strategy H is said to replicate the option
.Z; T/ in MT .P /.
13.4 Replicating strategies and market completeness 405
Remark 13.1 If H 2 MT .P / then the corresponding discounted portfolio

e
V.H/ is a martingale under P , hence eV T .H/ is P -integrable and the same is
RT
true for VT .H/ D e 0 rs dse
V T .H/, as the spot rate r is assumed to be bounded.
Therefore the condition Z 2 L1 .˝; FT ; P / is necessary for the option
.Z; T/ to be attainable.
From now on we denote by E the expectation with respect to P .
Proposition 13.4 Assume that an equivalent martingale measure P exists

and let .Z; T/ be an attainable European option. Then for every replicating
strategy H 2 MT .P / the corresponding portfolio is given by
h RT ˇ i
Vt .H/ D E e t rs ds Z ˇ Ft : (13.12)
RT
Proof The r.v. e t rs ds Z is integrable under P , because Z is integrable under P
and r 0. Moreover, for every replicating strategy H 2 MT .P / for Z, e V.H/ is a
P -martingale, hence
Rt Rt
Rt RT
Vt .H/ D e 0 rs ds e
V t .H/ D e E e V T .H/ j Ft D e 0 rs ds E e 0 rs ds VT .H/ j Ft
0 rs ds
Rt RT
RT
D e 0 rs ds E e 0 rs ds Z j Ft D E e t rs ds Z j Ft :
t
u
RT ˇ

Let Vt D E Œe t
Z ˇ Ft . Proposition 13.4 suggests that (13.12) should be the
rs ds
right price of an option at time t. In fact, whenever an equivalent martingale measure

exists, the issuer of the option with an amount Vt can start at time t an admissible
strategy covering the payoff at maturity.
From another point of view, if the price at time t was not given by (13.12) there
would be a possibility of arbitrage. If the price was fixed at a level Ct > Vt , then
an investor could set up a portfolio by selling the option, thus acquiring the amount
Ct . Part of this amount would be used in order to set up the replicating portfolio.
He would invest the difference Ct Vt into the riskless asset. This operation does
not require any capital to be engaged. At maturity the replicating portfolio would
have the same value as the payoff ofR the option and the investor would remain with
T
a strictly positive gain of .Ct Vt / e t rs ds , which constitutes an arbitrage.
In quite a similar way an arbitrage portfolio can be produced if the price was
fixed at a level Ct < Vt .
RT ˇ
For t < T, the value Vt D E e t rs ds Z ˇ Ft is the no-arbitrage price of the

European option Z at time t.
This definition of price obviously depends on the martingale measure P . This might
not be unique but the next statement asserts that, if many equivalent martingale
measures exist, the respective prices and replicating portfolios coincide.
Proposition 13.5 Suppose there exist two equivalent martingale measures

P1 and P2 . Let .Z; T/ be a European option attainable both in MT .P1 / and
MT .P2 /. Then the values at any time t of the two replicating portfolios agree
and are equal to
h RT ˇ i h RT ˇ i
Vt D E1 e t rs ds Z ˇ Ft D E2 e t rs ds Z ˇ Ft ;
where Ei denotes the expectation under Pi , i D 1; 2. In particular, the no-
arbitrage prices under P1 and P2 agree.
Proof Let H1 and H2 be replicating strategies for .Z; T/ in MT .P1 / and MT .P2 /,
respectively. In particular, they are both admissible and Z is integrable both with
respect to P1 and P2 . Since P1 and P2 are both equivalent martingale measures, by
Proposition 13.4 e V.H1 / is a P2 -supermartingale and e
V.H2 / is a P1 -supermartingale.
RT
Moreover, VT .H1 / D Z D VT .H2 / and thus R e V T .H1 / D e 0 rs ds Z D

e
V T .H2 / and by Proposition 13.4, e V t .H1 / D E1 e 0 rs ds Z j Ft and e
T
V t .H2 / D
R
e
T
0 rs ds
E2 e Z j Ft . As V.H1 / is a P2 supermartingale we have P2 -a.s.
h RT ˇ i
V t .H2 / D E2 e 0 rs ds Z ˇ Ft D E2 Œe
e V T .H1 / j Ft e
V t .H1 / :
By interchanging the role of P1 and P2 , we obtain e

V t .H2 / eV t .H1 / P1 -a.s. as well
and hence, as P1 , P2 and P are equivalent (so that P1 -a.s. is the same as P2 -a.s.), the

two conditional expectations agree. t

u
In summary, the existence of an equivalent martingale measure P (i.e. a measure
equivalent to P under which the discounted price processes are martingales) has two
important consequences.
• Every discounted portfolio is a P -local martingale.
• The market is arbitrage free, which allows us to determine the non-arbitrage price
for attainable options and the existence of a replicating strategy.
13.5 The generalized Black–Scholes models 407
One might ask whether the converse of the last statement also holds: is it true
that absence of arbitrage implies existence of an equivalent martingale measure?
The answer is positive: this is the fundamental theorem of asset pricing. In the
literature there are several results in this direction, according to the model chosen
for the market: the interested reader can refer to Musiela and Rutkowski (2005) and
the references therein.
We have seen that if an equivalent martingale measure exists, the no arbitrage-
price is well defined for every attainable option. Therefore it would be nice if every
option (at least under suitable integrability assumptions) were attainable.
Assume that there exists an equivalent martingale measure P . The model is

said to be complete if every European option .Z; T/ with Z 2 L2 .˝; FT ; P /
is attainable in MT .P / for every equivalent martingale measure P . Other-
wise the market model is said to be incomplete.
We have the following fundamental result
Theorem 13.1 If the model is complete then the equivalent martingale

measure is unique.
Proof Suppose that there exist two equivalent martingale measures P1 and P2 . Let
RT
A 2 FT and consider the option .Z; T/ defined as Z D e 0 rs ds 1A . Notice that Z is
FT -measurable and, as r is assumed to be bounded, Z 2 Lp .˝; Pi / for every p and
i D 1; 2. Since the market is complete, Z is attainable both in MT .P1 / and MT .P2 /.
Hence, by Proposition 13.5,
h RT i h RT i
P1 .A/ D E1 e 0 rs ds Z D E2 e 0 rs ds Z D P2 .A/ :
As this holds for every A 2 FT , P1 P2 on FT , i.e. the equivalent martingale
measure is unique. t
u
13.5 The generalized Black–Scholes models
In the previous sections we investigated general properties of the very general

model (13.1), (13.2). In this section we introduce a more precise model, which is a
particular case of (13.1) and (13.2), and we investigate the existence of an equivalent
martingale measure and completeness.
Let us recall first some properties that a reasonable model should enjoy.
As remarked on p. 398, prices must remain positive at all times i.e., if we assume
a situation where only one asset is present on the market, it will be necessary for its
price St to be 0 a.s. for every t 0.
Moreover, the increments of the price must always be considered in a multi-
plicative sense: reading in a financial newspaper that between time s and time t an
p
increment of p% has taken place, this means that SSst D 1 C 100 . It is therefore
wise to model the logarithm of the price rather than the price itself. These and other
considerations lead to the suggestion, in the case m D 1 (i.e. of a single risky asset),
of an SDE of the form
dSt
D b.St ; t/ dt C .St ; t/ dBt
St (13.13)
Ss D x :
If b and are constants, this equation is of the same kind as (9.6) on p. 259 and we
know that its solution is a geometric Brownian motion, i.e.
2
St D xe.b 2 /.ts/C .Bt Bs / ; (13.14)
which, if the initial position x is positive, is a process taking only positive values.
More precisely, we shall consider a market where m C 1 assets are present with
prices denoted S0 ; S1 ; : : : ; Sm . We shall assume that S0 is as in (13.2), i.e.
Rt
S0 .t/ D e s ru du
for a non-negative bounded progressively measurable process r. As for the other

assets we shall assume that they follow the SDE
dSi .t/ X d
D bi .St ; t/ dt C ij .St ; t/ dBj .t/ (13.15)
Si .t/ jD1
where St D .S1 .t/; : : : ; Sm .t//. Recall that in this model there are m risky assets and
that their evolution is driven by a d-dimensional Brownian motion, possibly with
m 6D d.
We shall make the assumption
b and are bounded and locally Lipschitz continuous.
With this assumption the price process S is the solution of an SDE with coefficients
satisfying Assumption (A) on p. 260. In particular (Theorem 9.1), S 2 M 2 .
This is the generalized Black–Scholes (or Dupire) model. In the financial models
the diffusion coefficient is usually referred to as the volatility.
By Ito’s formula applied to the function f W z D .z1 ; : : : ; zm / 7!
.log z1 ; : : : ; log zm /, the process t D .log S1 .t/; : : : ; log Sm .t// solves the SDE
1 Xd
di .t/ D bi .ei .t/ ; t/ dt aii .ei .t/ ; t/ dt C ik .ei .t/ ; t/ dBk .t/ ; (13.16)
2 jD1
where a D . This not a rigorous application of Ito’s formula, as f is not even

defined on the whole of R, but note that the assumptions stated for b and guarantee
that the SDE (13.16) has a unique solution, , defined for every t > 0. It is then easy
by Ito’s formula, correctly applied to f 1 W .y1 ; : : : ; ym / 7! .ey1 ; : : : ; eym /, to check
that St D xet > 0 is a solution of (13.13).
Hence if Si .s/ D xi > 0 then Si .t/ > 0 for every t s a.s., which is a good thing,
as remarked above.
Let us investigate the properties of this model starting from the existence
and uniqueness of an equivalent martingale measure. The next theorem gives a
characterization of the equivalent martingale measures in the generalize Black–
Scholes model.
Theorem 13.2 The following are equivalent.

a) An equivalent martingale measure P exists.
2
b) There exists a process 2 Mloc such that
Rt Rt
0 u dBu 12 ju j2 du
t D e 0 (13.17)
is an .Ft /t -martingale and is a solution of the system of m equations (but

is d-dimensional)
.St ; t/.t/ D Rt b.St ; t/ ; (13.18)
where Rt denotes the m-dimensional process having all its components

equal to rt .
Proof Let us assume that b) holds and let P be the probability having density T
with respect to P. By Girsanov’s theorem, Theorem 12.1, the process
Z t
Bt D Bt u du (13.19)
0
is, for t T, a Brownian motion with respect to the new probability dP D T dP
and under P the discounted price process, thanks to (13.10), satisfies
deSi .t/ Xd Xd
D .bi .Si .t/; t/ rt / dt C ij .Si .t/; t/ dBj .t/ D ij .Si .t/; t/ dBj .t/ :
eSi .t/ jD1 jD1
Let us prove that e Si is a martingale with respect to P . Let us apply Ito’s formula
and compute the stochastic differential of t 7! log e Si .t/ or, to be precise, let f" be
a function coinciding with log on "; C1Œ and extended to R so that it is twice
differentiable. If " denotes the exit time of e
Si from the half-line "; C1Œ, then Ito’s
formula gives
Z log e
Si .t ^ " / Z
t^"
1 t^" 00 e
D log xi C f"0 .e
Si .u// deSi .u/ C f" .Si .u// dhe Si iu
s Z 2Zs
t^"
1 e 1 t^"
1
D log xi C d Si .u/ dheSi iu
s e
Si .u/ 2 s e
Si .u/2
Z t^" X Z
1 t^" X
d d
D log xi C ij .Si .u/; u/ dBj .u/ ij .Si .u/; u/2 du :
s jD1
2 0 jD1
(13.20)
As we know that Si .t/ > 0, hence also e

Si .t/ > 0, for every t > 0 a.s., we have
" ! C1 as " ! 0C, so that we can take the limit as " ! 0C in (13.20), which
gives for e
Si the expression
Z t X
d
1
Z tX
d
e
Si .t/ D xi exp ij .Si .u/; u/ dBj .u/ ij .Si .u/; u/2 du :
s jD1
2 s jD1
As is assumed to be bounded, e Si is a martingale, thanks to Proposition 12.1.

Conversely, if an equivalent martingale measure P exists, then by Theorem 12.7
2
there exists a process 2 Mloc such that the r.v. T defined by (13.17) is the density
of P with respect to P on FT . Then with respect to P the process B defined

in (13.19) is a Brownian motion up to time T so that, for 0 t T, e Si is a solution

of
deSi .t/ Xd Xd
D bi .Si .t/; t/ rt C ij .Si .t/; t//j .t/ dt C ij .Si .t/; t/ dBj .t/ :
eSi .t/ jD1 jD1
As e
Si is a martingale with respect to P , then necessarily the coefficient of dt in the
previous differential vanishes, i.e.
X
bi .Si .t/; t/ rt C ij .Si .t/; t//j .t/ D 0
jD1
for every i D 1; : : : ; m, i.e. (13.18). t

u
Recall that, as seen in the proof of Theorem 13.2, the processes e
Si satisfy under

P the relation
deSi .t/ Xd
D ij .St ; t/ dBj .t/ (13.21)
eSi .t/ jD1
whereas for the undiscounted price processes we have
dSi .t/ X d
D rt dt C ij .St ; t/ dBj .t/ : (13.22)
Si .t/ jD1
If the price processes e

Si are martingales under the probability P , the process
introduced in Theorem 13.2 is called the market price of risk.
Let us now investigate the existence of an equivalent martingale measure for the
generalized Black–Scholes model.
Proposition 13.6 Suppose d m (i.e. that the dimension of the driving

Brownian motion is larger than or equal to the number of risky assets), that
.x; t/ has maximum rank for every .x; t/ (i.e. .x; t/ has rank equal to m) and,
moreover, that the matrix field a D is uniformly elliptic with its smallest
eigenvalue bounded from below by > 0 (Definition 9.5 and (9.41)).
Then, there exists at least an equivalent martingale measure P and the
corresponding market price of risk is bounded a.s.
Moreover, if d D m the process is unique and is the unique solution
of (13.18). Conversely, if d > m then there is no uniqueness for the equivalent
martingale measure.
Proof We know, see Remark 9.7, that the assumption of ellipticity implies that
a.x; t/ is invertible for every x; t. Let us first consider the case d D m and therefore
that .x; t/ is invertible for every x; t. Equation (13.18) then has the unique solution
t D .St ; t/1 .Rt b.St ; t// :

Such a process is bounded. In fact

˝ ˛
jt j2 D .St ; t/1 Rt b.St ; t/ ; .St ; t/1 Rt b.St ; t/
˝ ˛ 1
D . .St ; t//1 Rt b.St ; t/ ; Rt b.St ; t/ jRt b.St ; t/j2 K :
„ ƒ‚ …
Da.x;t/1
(13.23)
The existence of an equivalent martingale measure is now a consequence of

Theorem 13.2. If d > m the argument is similar but we cannot argue in the same
way, as now .x; t/ is not invertible. Actually .x; t/ has, for every x; t, a .d m/-
dimensional kernel. However, for every x; t, let .x; t/ be an orthogonal d d matrix
such that its columns from the .m C 1/-th to the d-th form an orthogonal basis of
ker .x; t/. Then (13.18) can be written as
.S ; t/.S ; t/ 1 .St ; t/t D Rt b.St ; t/ : (13.24)

„ t ƒ‚ t …
WD .St ;t/
The columns from the .m C 1/-th to the d-th of the matrix .St ; t/ D .St ; t/.St ; t/
vanish, hence .x; t/ is of the form
.x; t/ D .e
.x; t/; 0m;dm / ;
where e
.x; t/ is an m m matrix and 0m;dm denotes an m .d m/ matrix of zeros.
As
a.x; t/ D .x; t/ .x; t/ D .x; t/ .x; t/ D e .x; t/ ;

.x; t/e
e
.x; t/ is invertible for every x; t. Let 1 be the m-dimensional process

.St ; t/1 Rt b.St ; t/
1 .t/ D e
and let 2 .t/ be any bounded progressively

measurable process taking values in
1 .t/
R . Then the process t D 2 .t/ satisfies the relation
dm
.St ; t/t D Rt b.St ; t/
and t D 1 .St ; t/t is a solution of (13.24). A repetition of the argument as
in (13.23), proves that the process 1 is bounded, hence also is bounded, thus
proving the existence of an equivalent martingale measure.
Finally, note that if d > m then there are many bounded solutions of (13.18) (for
any possible choice of a bounded progressively measurable process 2 ). Hence if
d > m the equivalent martingale measure is not unique. t
u
In particular, if d > m, the market cannot be complete, thanks to Proposition 13.1

(see, however, Remark 13.2 and Example 13.1 below). Let us investigate complete-
ness if d D m. The main point is contained in the next statement.
Lemma 13.1 Suppose d D m and that is uniformly elliptic. Let

P denote the unique equivalent martingale measure whose existence is
proved in Proposition 13.6. Let .Z; T/ be a European option such that Z 2
L1 .˝; FT ; P / and let
h RT ˇ i
e t D E e 0 rs ds Z ˇ Ft :
M
2
Then there exist m progressively measurable processes H1 ; : : : ; Hm 2 Mloc
such that
X
m
et D
dM Hi .t/ de
Si .t/ :
iD1
Proof The idea is to use the representation theorem of martingales, Theorem 12.6,
recalling that in this chapter we assume that .Ft /t is the natural augmented filtration
of B. In fact, as Me is a martingale of the Brownian filtration, we might expect to
have the representation:
h RT i Z t
e 0 rs ds
Mt D E e Z C e
Y s dBs ; t 2 Œ0; T ; (13.25)
0
where eY is a progressively measurable process in Mloc 2

. In order to conclude the
proof it would be sufficient to choose H as the solution of
X
m
Hi .t/e
Si .t/ik .St ; t/ D e
Y k .t/ ; k D 1; : : : ; m; (13.26)
iD1
i.e., being invertible,
Œ. /1 .St ; t/e

Y t i
Hi .t/ D i D 1; : : : ; m : (13.27)
eSi .t/
The integrability requirement for H is satisfied as . /1 is bounded (this has

already been proved in the proof of Proposition 13.6, as a consequence of the fact
that a D is uniformly elliptic) and e
St is a strictly positive continuous process.
2
Therefore Hi 2 Mloc for every i because e 2
Y k 2 Mloc for every k.
However, the representation theorem for Brownian martingales cannot be applied

in this setting. Indeed, as we are working under P , thus with the Brownian motion
B , the filtration to be taken into account in the representation theorem for Brownian
martingales is not .Ft /t but .Ft /t , i.e. the augmented filtration generated by B and
completed by the P -null sets (and in general, Ft Ft ). In other words, the above
RT
argument would be correct if one had to work with E e t rs ds Z j Ft , and not
RT
with E e t rs ds Z j Ft .
The argument that follows is necessary in order to take care of this difficulty. Let
denote the market price of risk (see Proposition 13.6) and let
Rt Rt

0 s dBs 12 js j2 ds
t D e 0
dP
jF
be the usual exponential martingale such that dPjFT
T
D T . Let us consider the
martingale
h RT ˇ i
M t D E e 0 rs ds Z T ˇ Ft :
Note that the R expectation is taken Runder the original measureR P and not P
T
T
T
and that E e 0 rs ds Z T D E e 0 rs ds Z , so that the r.v. e 0 rs ds Z T is P-
integrable. Therefore .M t /t2Œ0;T is an .Ft /t -martingale under P and we can apply
the representation theorem for the martingales of the Brownian filtration and obtain
2
that there exists a progressively measurable process Y 2 Mloc such that
h RT i Z t
M t D E e 0 rs ds ZT C Yt dBt :
0
Now, recalling the relation between conditional expectations under a change of

probability (Exercise 4.4),
h RT ˇ i h RT ˇ i
h RT ˇ i E e 0 rs ds Z T ˇ Ft E e 0 rs ds Z T ˇ Ft Mt
e t D E e
M 0 rs ds ˇ
Z Ft D D D
EŒT j Ft t t
e We have
Let us compute the stochastic differential of M.
dM t D Yt dBt D Yt .dBt C t dt/
and dt D t t dBt , hence
1 1 1 1 1 1
d D 2 dt C 3 dhit D t dBt C jt j2 dt D t dBt
t t t t t t
and with a final stroke of Ito’s formula
et D d Mt 1 Mt 1 1
dM D Yt .dBt C t dt/ t dBt Yt t dt D .Yt M t t / dBt :
t t t t t
This gives the representation formula (13.25) with e Y t D t1 .Yt M t t /, which
when inserted in (13.27) gives the process H we are looking for. It remains to prove
that e 2
Y 2 Mloc , which is easy and left to the reader. t
u
Theorem 13.3 If m D d and the matrix field a.x; t/ D .x; t/ is uniformly

elliptic then the generalized Black–Scholes model (13.15) is complete.
Proof Let Z 2 L1 .˝; FT ; P / be a positive r.v. We must prove that it is attainable.

Let
h RT ˇ i
Me t D E e 0 rs ds Z ˇ Ft ; t 2 Œ0; T :
By Lemma 13.1 there exist m progressively measurable processes H1 ; : : : ; Hm 2

2
Mloc such that
X
m
et D
dM Hi .t/de
Si .t/ :
iD1
Let
X
m
et
H0 .t/ D M Hi .t/e
Si .t/ ; t 2 Œ0; T ; (13.28)
iD1
and consider the trading strategy H D .H0 ; H1 ; : : : ; Hm / over Œ0; T. Notice that for
the corresponding portfolio we have
X
m
e
V t .H/ D H0 .t/ C Hi .t/e et;
Si .t/ D M t 2 Œ0; T ; (13.29)
iD1
hence VT .H/ D Z. Let us prove that H is admissible.

First observe that H satisfies (13.5). In fact, we already know that H1 ; : : : ; Hm 2
2
Mloc . Moreover, Me and eS are continuous; this proves that the Hi Si are also in Mloc 2
2 1
and from (13.28) H0 2 Mloc Mloc . By (13.29)
X
m
de et D
V t .H/ D dM Hi .t/de
Si .t/ ;
iD1
so that V.H/ is self-financing thanks to Proposition 13.1. Moreover, e et

V t .H/ D M
0 a.s. by construction, being the conditional expectation of a positive r.v. Therefore
H is admissible.
Moreover, H is a trading strategy in MT .P / as e V.H/ is a .Ft /t -martingale under
e e t and M
e t is an .Ft /t -P -martingale by construction.
P : V t .H/ D M RT RT
As, finally, VT .H/ D e 0 rs dse e T D Z, H replicates Z at time T
V T .H/ D e 0 rs ds M
and Z is attainable. t
u
In the previous statements we have assumed d m, i.e. that there are at least
as many Brownian motions as there are risky assets. If, conversely, d < m then the
solution of the equation
.St ; t/t D b.St ; t/ Rt
might not exist as .x; t/ is a matrix mapping Rd ! Rm . Hence an equivalent

martingale measure might not exist and arbitrage is possible. Exercise 13.5 gives an
example of this phenomenon.
Remark 13.2 Let us criticize the theory developed so far.

The definition of a European option given in Definition 13.1 is not really
reasonable. An option in the real market is something that depends on the
value of the prices. The call and put options are of this kind, as well as the
other examples that we shall see in the next sections.
Therefore a reasonable definition of a European option would be that it
is a pair .Z; T/ (payoff and maturity) such that Z is H T -measurable where
H t D .Ss ; s t; N / denotes the filtration generated by the price process S
and augmented. It is clear that H t Ft and in general Ht 6D Ft (recall that
we assume .Ft /t D .G t /t , see also Remark 9.4).
Similarly it would be useful if the trading strategy Ht at time t were an
H t -measurable r.v., i.e. a functional of the price process.
If m D d this turns out not to be a real problem: under the assumptions of
Theorem 13.3, i.e. the matrix field a.x; t/ D .x; t/ is uniformly elliptic, it
can be proved that the two filtrations .Ft /t and .H t /t coincide.
This equality of the two filtrations is not true if d > m. In this case,
however, one may ask whether a European option .Z; T/ such that Z is
H T -measurable is attainable. The next example provides an approach to the
question.
Example 13.1 Let us consider a generalized Black–Scholes market model

with one risky asset satisfying the SDE, with respect to the two-dimensional
(continued )
13.6 Pricing and hedging in the generalized Black–Scholes model 417

Brownian motion B,
dSt
D b.St / dt C 1 dB1 .t/ C 2 dB2 .t/ ; (13.30)
St
where 1 ; 2 > 0. In this case, the dimension of the Brownian motion is

d D 2, so that, by Proposition 13.6, there is no uniqueness of the equiv-
alent martingale measure and by Theorem 13.1 the model is not complete.
Hence there are options that are not attainable. But, denoting by .H t /t the
augmented filtration generated by S, we see now that every European option
.Z; T/ such that Z is H T -measurable is actually attainable.
Indeed, consider the real Brownian motion
1 B1 .t/ C 2 B2 .t/
Wt D q
12 C 22
Now (13.30) can be rewritten as

dSt
D b.St / dt C dWt ;
St
q
where D 12 C 22 and Theorem 13.3 guarantees that every option that
W W
is G T -measurable is attainable, where we denote by .G t /t the augmented
filtration of W (which is strictly smaller than FT : try to find an example of
W
an FT -measurable r.v. which is not GTW -measurable). As H T G T (again
thanks to Remark 9.4) our claim is verified.
W
Note also that in this case H T D G T (see Exercise 9.26).
13.6 Pricing and hedging in the generalized Black–Scholes

model
We have seen in Theorem 13.3 that under suitable conditions for the generalized
Black–Scholes model a unique equivalent martingale measure P exists and that
the model is complete. Hence every European option .Z; T/ such that Z is square
integrable with respect to P is attainable and by Proposition 13.4 its no-arbitrage
price is given by
h RT ˇ i
Vt D E e t rs ds Z ˇ Ft : (13.31)
The problem of the determination of the price is therefore solved.

But let us consider the problem also from the point of view of the issuer: which
strategy should be taken into account in order to deliver the contract? The value
Vt of (13.31) is also the value of a replicating portfolio but it is also important to
determine the corresponding strategy H. This would enable the issuer to construct
the replicating portfolio, which, at maturity, will have the same value as the payoff
of the option.
This is the hedging problem. In Theorem 13.1 we have proved the existence of
a replicating strategy H, but we made use of the martingale representation theorem,
which is not constructive. We make two additional assumptions to our model
• .x; t/ 7! .x; t/ is Lipschitz continuous.

• The randomness of the spot rate r is driven by the risky asset prices, i.e. is
of the form
rt D r.St ; t/ :
Note that this assumption contains the case where r is a deterministic

function. We shall require .x; t/ 7! r.x; t/ to be a bounded, non-negative
and Lipschitz continuous function.
Hence, recalling (13.22) and (13.21), the price process and the discounted price
process solve respectively the SDEs
dSi .t/ Xd
D r.St ; t/ dt C ik .Si .t/; t/ dBk .t/
Si .t/ kD1
(13.32)
Si .0/ D xi i D 1; : : : ; m
and
deSi .t/ Xd
D ik .Si .t/; t/ dBk .t//
eSi .t/ (13.33)
kD1
e
Si .0/ D xi i D 1; : : : ; m :
Note that with these assumptions on and r, the diffusion process S satisfies the
hypotheses of most of the representation theorems of Chap. 10 and in particular of
Theorem 10.6.
Let us consider an option Z of the form Z D h.ST /. Under P , S is an .m C 1/-
dimensional diffusion and thanks to the Markov property, Proposition 6.1,
h RT ˇ i
Vt D E e t r.Ss ;s/ ds h.ST / ˇ Ft D P.St ; t/ ;
where
h R T x;t i
P.x; t/ D E e t r.Ss ;s/ ds h.STx;t / ;
.Ssx;t /st denoting the solution of (13.33) starting at x at time t. The value of the
replicating discounted portfolio is then
Rt Rt
e
V t D e 0 r.Ss ;s/ ds
P.St ; t/ :
Now suppose that the function P is of class C2;1 (continuous, twice differentiable in
the variable x and once in t). Ito’s formula gives
Rt Rt
de
V t D r.St ; t/e 0 r.Ss ;s/ ds
P.St ; t/ dt C e 0 r.Ss ;s/ ds dP.St ; t/
Rt @P
D e 0 r.Ss ;s/ ds r.St ; t/P.St ; t/ C .St ; t/
@t
Xm
@P
C .St ; t/r.St ; t/Si .t/
iD1
@x i
(13.34)
1 X @2 P
m
C .St ; t/ aij .St ; t/ Si .t/ Sj .t/ dt
2 i;jD1 @xi @xj
Rt Xm
@P Xd
C e 0 r.Ss ;s/ ds
.St ; t/Si .t/ ik .St ; t/ dBk .t/ ;
iD1
@x i kD1
where a D as usual. Now, as e V is a P -martingale, necessarily the finite

variation part in the previous differential must vanish, i.e.
@P
.x; t/ C Lt P.x; t/ r.x; t/ P.x; t/ D 0; on Rm Œ0; TŒ ; (13.35)
@t
where Lt is the generator of S under the risk neutral measure i.e.
X
m
@ 1X
m
@2
Lt D r.x; t/ xi C aij .x; t/ xi xj
iD1
@xi 2 i;jD1 @xi @xj
Theorem 13.4 Let h W Rm ! RC be a continuous function such that h.x/

C.1 C jxj˛ / for some C; ˛ > 0 and let
h R T x;t i
P.x; t/ D E e t r.Ss ;s/ ds h.STx;t / (13.36)
(continued)

be the price function of the option h.ST / with maturity T in the generalized
Black–Scholes model. Then P is the solution of the PDE problem
8
< @P .x; t/ C L P.x; t/ r.x; t/ P.x; t/ D 0; on Rm Œ0; TŒ
t
@t (13.37)
:
P.x; T/ D h.x/ :
Proof In the computation above we obtained (13.37) under the assumption that P
is of class C2;1 , which is still to be proved. In order to achieve this point we use
Theorem 10.6, which states that the PDE problem (13.37) has a solution and that it
coincides with the left-hand side of (13.36).
Unfortunately Theorem 10.6 requires the diffusion coefficient to be elliptic,
whereas the matrix of the second-order derivatives of Lt vanishes at the origin
(and becomes singular on the axes), hence it cannot be applied immediately. The
somehow contorted but simple argument below is developed in order to circumvent
this difficulty.
The idea is simply to consider the process of the logarithm of the prices, whose
generator is elliptic, and to express the price of the option in terms of this logarithm.
For simplicity, for x 2 Rm , let us denote .ex1 ; : : : ; exm / by ex . By a repetition of
the argument leading to (13.16), if i .t/ D log Si .t/ then
1 Xd
.t/ .t/
di .t/ D r.e ; t/ dt a ii .e ; t/ dt C ik .e.t/ ; t/ dBk .t/ ;
2 jD1
where a D and with the starting condition i .s/ D log xi . Hence is a diffusion
with generator
Xm @ X m
1 @2
Lt D r.ex ; t/ aii .ex ; t/ C aij .ex ; t/ (13.38)
iD1
2 @xi i;jDm @xi @xj
Now for the option price we can write

h R T x;t x;t
i
P.ex ; t/ D E e t r.e ;s/ ds h.eT / :
s
As the generator Lt in (13.38) satisfies the assumptions of Theorem 10.6 (in partic-
ular, its diffusion coefficient is uniformly elliptic), the function P.x; t/ WD P.ex ; t/ is
a solution of
8
< @P
.x; t/ C Lt P.x; t/ r.ex ; t/ P.x; t/ D 0; on Rm Œ0; TŒ
@t
:
P.x; T/ D h.ex / ;
which easily implies that P satisfies (13.37). t

u
From (13.34) we now have
Xm
@P Xd Xm
@P
de
Vt D .St ; t/e
Si .t/ ik .St ; t/ dBk .t/ D .St ; t/ de
Si .t/
iD1
@x i kD1 iD1
@x i
so that, by Proposition 13.1, we find the replicating strategy, defined by

8 @
ˆ
ˆHi .t/ D P.St ; t/ i D 1; : : : ; m
< @xi
Xm (13.39)
ˆH .t/ D eVt Hi .t/ e
Si .t/ ;
:̂ 0
iD1
which is self-financing and, as e

V t 0 by construction, also admissible.
Equation (13.35) is sometimes called the fundamental PDE following from the
no-arbitrage approach. Theorem 13.4 gives a way of computing the price P of
the option with payoff h.ST /, just by solving the PDE problem (13.37), possibly
numerically.
Note the not-so-intuitive fact that the price does not depend on the drift b
appearing in the dynamic (13.15) of the prices in the generalized Black–Scholes
model.
This means in particular that it is irrelevant, in order to consider the price of an

option, whether the price of the underlying asset follows an increasing trend
(corresponding to bi .x; t/ > 0) or decreasing (bi .x; t/ < 0).
Rt
Recalling that e
P.St ; t/, (13.39) can also be written as
@P
Hi .t/ D .St ; t/; i D 1; : : : ; m
@xi
Rt Xm (13.40)
H0 .t/ D e 0 r.Ss ;s/ ds P.St ; t/ Hi .t/ Si .t/ :
iD1
The quantities Hi .t/, i D 1; : : : ; m, in (13.39) are also called the deltas of the option
and are usually denoted by :
@P
i .St ; t/ D .St ; t/; i D 1; : : : ; m :
@xi
The delta is related to the sensitivity of the price with respect to the values of the
underlying asset prices. In particular, (13.40) states that the replicating portfolio
must contain a large amount in the i-th underlying asset if the price of the option is
very sensitive to changes of the price of the i-th underlying.
The delta is a special case of a Greek.
The Greeks are quantities giving the sensitivity of the price with respect to the
parameters of the model. The name “Greeks” comes from the fact that they are
usually (but not all of them. . . ) denoted by Greek letters. They are taken into special
account by practitioners, because of the particular financial meanings of each of
them. The most used Greeks can be summarized as follows:
• delta: sensitivity of the price of the option w.r.t. the initial value of the price
@P
of the underlying: i D @x i
;
• gamma: sensitivity of the delta w.r.t. the initial values of the price of the
2P
underlying: ij D @x@i @x j
;
• theta: sensitivity of the price w.r.t. the initial time: D @P
@t I
@P
• Rho: sensitivity of the price w.r.t. the spot rate: RhoD @r I
@P
• Vega: sensitivity of the price w.r.t. the volatility: VegaD @ .
Obviously, in the Rho and Vega cases, the derivatives must be understood in a
suitably functional way whenever r and are not modeled as constants.
The last two Greeks, Rho and Vega, give the behavior of the price and then of
the portfolio with respect to purely financial quantities (i.e. the interest rate and the
volatility), whereas the other ones (delta, gamma and theta) give information about
the dependence of the portfolio with respect to parameters connected to the assets on
which the European option is written (the starting time and the prices of the assets).
13.7 The standard Black–Scholes model
In this section we derive explicit formulas for the price of a call and put option as
well as the associated Greeks in a classical one-dimensional model, the standard
Black–Scholes model. By this we mean the particular case where there is only one
risky asset and the volatility and the spot rate r are constant.
13.7 The standard Black–Scholes model 423
Under the risk-neutral measure P , the price of the risk asset evolves as
dSt
D r dt C dBt (13.41)
St
and the price at time t of the call option with maturity T is given by
h ˇ i
Pcall .St ; t/ D E er.Tt/ .ST K/C ˇ Ft ;
where K stands for the strike price. If we use the notation Sx;t to denote the solution
S of (13.41) starting at x at time t, then
1 2 /.st/C .B B /
Ssx;t D x e.r 2 s t ; st;
and we have
h i
Pcall .x; t/ D E er.Tt/ .STx;t K/C
h 1 2 C i (13.42)
D er.Tt/ E x e.r 2 /.Tt/C .BT Bt / K :
The expectation above can bepcomputed remarking that, with respect to P , BT Bt
has the same distribution as T t Z with Z N.0; 1/, so that
Z C1
1 .r 1 2 /.Tt/C pTt z C 2
Pcall .x; t/ D er.Tt/ p xe 2 K ez =2 dz :
2 1
1 2 /.Tt/C
p
The integrand vanishes if xe.r 2 Tt z
K, i.e. for z d0 .x; T t/, where
1 x 1
d0 .x; t/ D p log r 2 t
t K 2
so that
Pcall .x; t/
Z
r.Tt/ 1 C1 .r 1 2 /.Tt/C pTt z 2 (13.43)
De p xe 2 K ez =2 dz :
2 d0 .x;Tt/
This integral can be computed with a simple if not amusing computation, as already
developed in Exercise 1.13. If we denote by ˚ the partition function of a N.0; 1/-
distributed r.v. we have from (13.43)

Z C1 Z
x 1
p 2 Ker.Tt/ C1 2
Pcall .x; t/ D p e 2 .z Tt/ dz p ez =2 dz
2 d0 .x;Tt/ 2 d0 .x;Tt/
Z C1
x 2
Dp p ez =2 dz Ker.Tt/ ˚.d0 .x; T t//
2 d0 .x;Tt/ Tt
p
D x˚.d0 .x; T t/ C T t / Ker.Tt/ ˚.d0 .x; T t// :
It is customary to introduce the following two quantities
p 1 x 1
d1 .x; t/ D d0 .x; t/ C t D p log C r C 2 t
t K 2
1 x 1
d2 .x; t/ D d0 .x; t/ D p log C r 2 t
t K 2
so that finally the price of the call option is given by the classical Black-Scholes
formula (see Fig. 13.1)
Pcall .x; t/ D x ˚.d1 .x; T t// K er.Tt/ ˚.d2 .x; T t// : (13.44)
0.6
K=0.95
P(x, 0)
0.4
K=1.5 =
K 4
0.2
0
0 0.5 1 1.5 2
Fig. 13.1 Behavior of the price of a call option as a function of , on the basis of the Black-
Scholes formula for x D 1; r D :15; T D 1 and different values of the strike price K. As ! 0
the price tends to 0 if log Kx rT > 0, otherwise it tends to x KerT
Let us determine the hedging portfolio. Again by straightforward computations, one

obtains
@Pcall
call .x; t/ D .x; t/ D ˚.d1 .x; T t// : (13.45)
@x
Hence, recalling formulas (13.40), a hedging portfolio for the call option in this
model is given by
H1 .t/ D call .St ; t/ D ˚.d1 .St ; T t//

(13.46)
H0 .t/ D ert P.St ; t/ ˚.d1 .St ; T t//St :
Note that we can write

1 1 St 1 p
d1 .St ; T t/ D p log C r C 2 T t :
T t K 2
If the price St remains > K for t near T then d1 .St ; T t/ will approach C1 and,
thanks to (13.46), H1 .t/ will be close to 1. This is in accordance with intuition: if the
price of the underlying asset is larger than the strike it is reasonable to expect that
the call will be exercised and therefore it is wise to keep in the replicating portfolio
a unit of the underlying.
As for the put option, one could use similar arguments or also the call-put parity
property
Pcall .x; t/ Pput .x; t/ D er.Tt/ E Œ.STx;t K/ D x Ker.Tt/
as explained later in Remark 13.3. The associated price and delta are therefore given
by the formulas
Pput .x; t/ D K er.Tt/ ˚.d2 .x; T t// x ˚.d1 .x; T t//;

(13.47)
@Pput
put .x; t/ D .x; t/ D ˚.d1 .x; T t// 1 :
@x
The other Greeks can also be explicitly written as summarized below (denoting by
D ˚ 0 the standard N.0; 1/ density):
.d1 .x; T t//

call .x; t/ D p D put .x; t/
x T t
x .d1 .x; T t//
call .x; t/ D p r x er.Tt/ .d2 .x; T t//
2 T t
x .d1 .x; T t//
put .x; t/ D p C r x er.Tt/ ˚.d2 .x; T t// (13.48)
2 T t
Rhocall .x; t/ D x .T t/ er.Tt/ ˚.d2 .x; T t//
Rhoput .x; t/ D x .T t/ er.Tt/ ˚.d2 .x; T t//

p
Vegacall .x; t/ D x T t .d1 .x; T t// D Vegaput .x; t/ :
These expressions of the Greeks can be obtained by the Black–Scholes formula

with straightforward computations. Let us derive explicitly the formula for the Vega,
whose value has some important consequences. Let us consider the case of the call
option. We must take the derivative of the price with respect to . Instead of taking
the derivative in (13.44), it is simpler to differentiate (13.43) under the integral sign.
We then have, since the integrand in (13.43) vanishes at d0 .x; T t/,
@Pcall
.x; t/
@
er.Tt/ @d0 1 2 p 2 ˇˇ
D p .x; T t/ xe.r 2 /.Tt/C Tt z K ez =2 ˇ
2 @ „ ƒ‚
zDd0 .x;Tt/
…
D0
Z
e r.Tt/ C1
@ .r 1 2 /.Tt/CpTt z 2
C p xe 2 K ez =2 dz
2 d0 .x;Tt/ @
Z C1 p
x 1 2 p
2
Dp .T t/ C T t z e 2 .Tt/C Tt z ez =2 dz
2 d0 .x;Tt/
Z C1 p
x 1 p 2
Dp .T t/ C T t z e 2 .z Tt/ dz
2 d0 .x;Tt/
p
and with the change of variable y D z T t we finally obtain
Z C1 p
@Pcall x 2
.x; t/ D p p T t y ey =2 dy
@ 2 d0 .x;Tt/ Tt
p
x T t 1 .d0 .x;Tt/ pTt/2
D p e 2
2
p
which, recalling that d0 .x; T t/ T t D d1 .x; T t/, is the result that was
claimed in (13.48). In particular, the Vega in the standard Black–Scholes model is
strictly positive, i.e.
the price is a strictly increasing function of the volatility (Fig. 13.1).
Example 13.2 Let us assume 2 D 0:2, r D 0:1, T D 1, K D 1. What is

the price of a call option at time t D 0 if the price of the underlying asset is
x D 0:9? What is the composition, always at time 0, of the hedging portfolio?
This is a numerical computation, requiring appropriate software in order
to perform the evaluations requested by formulas (13.44) and (13.46). The
results are
Pcall .0:9; 0/ D 0:16

H1 .0/ D 0:58
H0 .0/ D 0:37 :
How do we expect these quantities to change if the price at t D 0 were x D

1:1?
In this case the option is more likely to be exercised, the initial price x
now being greater than the strike. Hence we expect the price to be higher and
the hedging portfolio to contain a larger amount of the underlying asset. The
numerical computation gives
Pcall .1:1; 0/ D 0:29

H1 .0/ D 0:75
H0 .0/ D 0:53
in accordance with what was expected.

And what if, still with x D 1:1, we assumed 2 D 0:5? We know that the
price will increase. In fact
Pcall .1:1; 0/ D 0:38

H1 .0/ D 0:74
H0 .0/ D 0:42 :
Example 13.3 (Barrier options) The payoff of a European option (Defini-

tion 13.1) can be any positive FT -measurable r.v. Puts and calls are particular
cases that are functions of the value of the price of the underlying asset at time
T. In practice more general situations are possible, in particular with payoffs
that are functions of the whole path of the price. These are generally referred
to as path dependent options.
A typical example is the payoff
Z D .ST K/C 1fsup0sT Ss Ug ; (13.49)
where U is some value larger than the strike K. The holder of this option
receives at time T the amount .ST K/C (as for a classical call option) but
under the constraint that the price has never crossed the level U (the barrier)
before time T. Many variations of this type are possible, combining the type
of option (put or call as in (13.49)) with the action at the crossing of the barrier
that may cancel the option as in (13.49) or activate it. In the financial jargon
the payoff (13.49) is an up and out call option.
In this example we determine the price at time t D 0 of the option
with payoff (13.49) under the standard Black–Scholes model. The general
formula (13.12) gives the value

p D E erT .ST K/C 1fsup0tT St Ug ; (13.50)
where, under P ,
1 2 /tC B
St D xe.r 2 t
and x D S0 is the price of the underlying at time 0. The computation of

the expectation in (13.50) requires the knowledge of the joint distribution of
ST and its running maximum sup0tT St or, equivalently, of the Brownian
motion with drift 1 .r 12 2 / and of its running maximum. The idea is to
first make a change of probability in order to “remove” the drift. The joint
distribution of a Brownian motion and of its running maximum is actually
known (Corollary 3.3). Let
1 1 1 1 2
ZT D exp r 2 BT 2 r 2 T ;
2 2 2
then with respect to Q, dQ D ZT dP , the process
1 2
Wt D Bt C r t
2
(continued )

is a Brownian motion. Hence
EŒerT .ST K/C 1fsup0tT St Ug

h i
D erT EQ ZT1 .xe WT K/C 1fsup0tT xe Wt Ug :
1 2
Now, as BT D WT
.r 2
/T,
h1
1 2 1 1 2 i
ZT1 D exp BT C 2 r 2 T
r
2 2 2
h1 1 2 1 1 2 2 i
D exp r WT 2 r T
2 2 2
so that the price p of the option is equal to
1 1 h 1 1 2 i
rT .r 2 2 /2 T Q
e 2 2 E e .r 2 /WT .xe WT K/C 1fsup0tT xe Wt Ug :
By Corollary 3.3 the joint law of .WT ; sup0tT Wt / has density

2 1=2 1 2
f .z; y/ D 3
.2y z/ e 2T .2yz/ ; y > z; y > 0 :
T
Hence
h 1 1 2 i
EQ e .r 2 /WT .xe WT K/C 1fsup0tT xe Wt Ug
Z 2 Z y 1
1 2
D dy e .r 2 /z .xe z K/C f .z; y/ dz
0 1
where 2 D 1 log. Ux /. Note that the integrand vanishes for z 1 WD

1
log. x /, so that we are left with the computation of
K
Z 2 Z y 1
1 2
dy e .r 2 /z .xe z K/f .z; y/ dz
0 1
2 1=2 Z 2 Z 2 1 1 2 1 2
D 3
dz e .r 2 /z .xe z K/ .2y z/ e 2T .2yz/ dy :
T 1 z_0
(continued )

Replacing the variable y with D 2y z the integral becomes
Z 2 Z 22 z
1 1 1 2 1 2
e .r 2 /z .xe z K/ dz e 2T d
2 1 2.z_0/z
Z 2 1 1 1
T 1 2 2 2
D e .r 2 /z .xe z K/.e 2T .2.z_0/z/ e 2T .22 z/ / dz :
2 1
Finally, the requested price is equal to

1 1
.r 2 2 /2 T 1
rT
2 2
e p
Z 2T
2 1 1 2 1 1 (13.51)
.r /z .2.z_0/z/2 2
e 2
.xe K/.e 2T
z
e 2T .22 z/ / dz :
1
From this it is possible to deduce a closed formula, having the flavor of the
Black–Scholes formula, only a bit more complicated.
Proposition 13.7 In the standard Black-Scholes model with spot inter-

est rate r and volatility , the price p at time 0 as a function of the initial
price x of an up-and-out call barrier option with strike K, maturity T and
barrier U is given by:
x .˚.a1 / ˚.a2 // KerT .˚.b1 / ˚.b2 //

2rTC 2 TC4 2 2
rT T2 . r 2 /2 T2 22 C 12 . p /
xe 2 T .˚.c1 / ˚.c2 // C (13.52)
2rT 2 TC4 2 2
rT T2 . r 2 /2 T2 22 C 12 . p /
CKe 2 T .˚.d1 / ˚.d2 // ;
where ˚ is again the cumulative distribution function of a N.0; 1/-

distributed r.v. and
1 1
1 D log. Kx /; 2 D log. Ux /;

2 p 2r C 2
1 p 2r C 2
a1 D p T ; a2 D p T ;
T 2 T 2
2 p 2r 2
1 p 2r 2
b1 D p T ; b2 D p T ;
T 2 T 2
2 2rT C 2 T C 42 1 2rT C 2 T C 42
c1 D p p ; c2 D p p ;
T 2 T T 2 T
2 2rT 2 T C 42 1 2rT 2 T C 42
d1 D p p ; d2 D p p
T 2 T T 2 T
(continued )

Proof The integral in (13.51) can be decomposed into the sum of four terms:
Z 2 Z 2
r 1 2 2 2 r 2 1 2
x e. C 2 /z 2T z dz x e T 2 e. C 2 C T 2 /z 2T z dz
1 1
Z 2 Z 2
1 2 2 2 2 1 2
e. 2 /z 2T z dz C Ke T 2 e. 2 C T 2 /z 2T z dz
r r
K
1 1
and we treat them separately. The first term on the right-hand side can be
rewritten as
Z 2 p 2 2
T 2rC 2 2 12 . pz T 2rC
2 /
xe 2 . 2 / e T dz;
1
p 2rC 2
so that with the change of variable u D pz T 2 we arrive at
T
Z p 2
p p2 T 2rC
T 2rC 2 2 T 2 1 u2
x 2T e 2 . 2 / p 2
p e 2 du

p1
T
T 2rC
2
2T
p T 2rC 2 2
D x 2T e 2 . 2 / .˚.a1 / ˚.a2 // :
With similar changes of variables, the second term gives
p 2 2C 1 .
2rTC 2 TC4 2 2
p /
x 2T e T 2 2 2 T .˚.b1 / ˚.b2 // ;
the third term

p T 2r 2 2
K 2T e 2 . 2 / .˚.c1 / ˚.c2 // ;
and the last one

p 2 2C 1 .
2rT 2 TC4 2 2
p /
K 2T e T 2 2 2 T .˚.d1 / ˚.d2 // :
1 2
Multiplying all these terms by erT 2 2 .r 2 /T
p1
2T
the proof is
completed. t
u
As an example let us compare the price of a barrier option with the values
x D 0:9, D 0:2, r D 0:1, T D 1, K D 1 and with the barrier U D 2.
Hence in this case the option is canceled if before the maturity T the price of
the underlying becomes larger that 2.
(continued )

We have obtained in Example 13.2 that without the barrier constraint the
price of the option is 0:15. Computing with patience the value in (13.7) we
find that the price of the corresponding barrier option becomes 0:09.
The barrier option is therefore considerably less expensive, but of course
it does not protect from spikes of the price of the underlying asset. Taking
U D 3 the price of the barrier option becomes 0:14, very close to the value
without barrier.
Remark 13.3 (The call-put parity) Let Ct , resp. Pt , denote the price of a call,
resp. a put, option on the same asset at time t. Assume that the spot rate r is
deterministic. Then the following call-put parity formula holds:
RT
Ct D Pt C St e t rs ds
K: (13.53)
In fact, using the relation zC .z/C D z,

h RT ˇ i h RT ˇ i
Ct Pt D E e t rs ds .ST K/C ˇ Ft E e t rs ds .K ST /C ˇ Ft
h RT ˇ i Rt RT
D E e t rs ds .ST K/ ˇ Ft D e 0 rs ds E Œe
ST jFt e t rs ds K
Rt RT RT
De e
S t e
0 rs ds t rs ds
K D S t e t rs ds
K:
The relation (13.53) can also be obtained from the requirement of absence of
arbitrage, without knowing the expression of the prices. Let us verify that a
different relation between these prices would give rise to an opportunity of
arbitrage. RT
Let us assume Ct > Pt C St Ke t rs ds . One can then establish a portfolio
buying a unit of the underlying asset and a put option and selling a call option.
The price of the operation is Ct Pt St and is covered through an investment
of opposite sign in the riskless asset. This operation therefore does not require
us to engage any capital. At maturity we dispose of a put option,
RT
a unit of the
underlying asset and an amount of cash .Ct Pt St / e t rs ds and we have
to fulfill a call.
There are two possibilities
• ST > K. In this case the call is exercised; we sell the underlying, which
allows us to honor the call and to collect an amount equal to K. The put is,
of course, valueless. The global balance of the operation is
RT
K .Ct Pt St / e t rs ds
>0:
(continued )

• ST K. In this case the call is not exercised. We use then the put in order
to sell the unit of underlying at
RT
the price K. The global balance of operation
is again K .Ct Pt St / e t rs ds > 0.
In a similar way it is possibleRT
to establish an arbitrage portfolio if
t rs ds
conversely Ct < Pt C St Ke .
Note that this arbitrage argument holds for very general models (the only
requirement is that the spot interest rate must be deterministic).
(13.53) is useful because it allows us to derive the price of a call option
from the price of the corresponding put and put options are somehow easier
to deal with, as their payoff is a bounded r.v.
Note that almost all quantities appearing in the Black–Scholes formulas (13.43)
and (13.47) are known in the market. The only unknown quantity is actually the
volatility .
In practice, is estimated empirically starting from the option prices already
known: let us assume that in the market an option with strike K and maturity T is
already traded with a price z. Let us denote by CK;T ./ the price of a call option
as a function of the volatility for the given strike and volatility K; T. As CK;T is
1
a strictly increasing function, we can determine the volatility as D CK;T .z/. In
this way options whose price is already known allow us to determine the missing
parameter and in this way also the price of options not yet on the market.
Another approach to the question is to estimate from the observed values of
the underlying: actually nowadays the price of a financial asset is known at a high
frequency. This means that, denoting the price process by S, the values St1 ; St2 ; : : :
at times t1 ; t2 ; : : : are known. The question, which is mathematically interesting in
itself, is whether it is possible starting from these data to estimate the value of ,
assuming that these values come from a path of a process following the Black–
Scholes model.
The fact that the option price is an invertible function of the volatility also allows
us to check the soundness of the Black–Scholes model. Assume that two options
on the same underlying asset are present in the market, with strike and maturity
K1 ; T1 and K2 ; T2 and prices z1 ; z2 , respectively. If the Black–Scholes model was a
good one, the value of the volatility computed by the inversion of the price function
should be the same for the two options, i.e. the two quantities
CK11 ;T1 .z1 / and CK12 ;T2 .z2 /
should coincide. In practice, it has been observed that this is not the case. The
standard Black–Scholes model, because of its assumption of constancy of the
volatility, thus appears to be too rigid as a model of the real world financial markets.
Nevertheless, it constitutes an important first attempt and also the starting point of,
very many, more complicated models that have been introduced in recent years.
Exercises
13.1 (p. 616) Assume that the prices St D .S0 .t/; S1 .t/; : : : ; Sm .t// follow the
generalized Black–Scholes model (13.15) and that the spot rate process .rt /t is inde-
pendent of .St /t (this assumption is satisfied in particular if .rt /t is deterministic).
Assume, moreover, that there exists an equivalent martingale measure P . Prove
that, for every i D 1; : : : ; m, and E ŒSi .0/ < C1,
Rt
E Œrs ds
E ŒSi .t/ D e 0 E ŒSi .0/ :
13.2 (p. 617) Let us consider the generalized Black–Scholes model (13.15) which,
we know, may not be complete in general, so that there are options that might
not be attainable. We want to prove, however, that, if the spot interest rate .rt /t is
deterministic and there exists an equivalent martingale measure P , then the option
Z T
ZD ˛Si .s/ C ˇ ds ;
0
˛; ˇ > 0, is certainly attainable, for every i D 1; : : : ; m.

a) Write down the Rdifferential of Si under P and compute E ŒSi .t/jFs .
T
b) Compute E Œe t ru du Z jFt and determine a replicating portfolio for the option
Z (check in particular that it is admissible).
c) Deduce that Z is an attainable option and compute its price at time 0 as a function
of the price x D Si .0/.
13.3 (p. 618) In a standard Black–Scholes model let us consider an investor

that constructs a self-financing portfolio V with the constraint that the value of
the component invested into the risky asset is constant and equal to some value
M > 0.
a) Assuming that the value V0 at time 0 is deterministic and positive, compute the
distribution of Vt and compute EŒVt , as a function of V0 and of the parameters
r; ; of the Black–Scholes model.
b) Is V admissible?
13.4 (p. 620) (Have a look at Example 12.5 first) Let us consider a standard Black–
Scholes model with parameters b; ; r. Consider an option, Z, that pays an amount
C if the price S crosses a fixed level K, K > 0, before some fixed time T and 0
otherwise.
a) Write down the payoff of this option.

b) Compute its no-arbitrage price at time 0.
13.5 (p. 621) Let us consider a market with a riskless asset with price S0 .t/ D ert
and two risky assets with prices
dSi .t/ D Si .t/.i dt C dBt /; Si .0/ D xi ;
where i 2 R, 1 6D 2 , r; > 0 and B is a real Brownian motion. Show that there

are arbitrage possibilities by proving that an equivalent martingale measure does not
exist and provide explicitly an arbitrage portfolio.
Solutions of the Exercises
1.1 If X and Y have the same law , then they also have the same p.f., as
FX .t/ D P.X t/ D . 1; t/ D P.Y t/ D FY .t/ :
Conversely, if X and Y have the same p.f. F, then, denoting by X and Y their
respective laws and if a < b,
X .a; b/ D P.a < X b/ D P.X b/ P.X a/ D F.b/ F.a/ :
Repeating the same argument for Y , X and Y coincide on the half-open intervals
a; b, a; b 2 R; a < b. The family C formed by these half-open intervals is stable
with respect to finite intersections (immediate) and generates the Borel -algebra.
Actually a -algebra containing C necessarily contains any open interval a; bŒ (that
is, the intersection of the half-open intervals a; b C 1n ) and therefore every open set.
By Carathéodory’s criterion, Theorem 1.1, X and Y coincide on B.R/.
1.2
a) If x > 0
Z x Z x
F.x/ D P.X x/ D f .t/ dt D e t dt D 1 e x ;
1 0
whereas the same formula gives F.x/ D 0 if x < 0 (f vanishes on the negative
real numbers). With some patience, integrating by parts, we find
Z C1 Z C1
1
EŒX D xf .x/ dx D x e x dx D
1 0
Z C1 Z C1
1 1
EŒX 2 D x2 f .x/ dx D x2 e x dx D C 2
1 0

P. Baldi, Stochastic Calculus, Universitext, DOI 10.1007/978-3-319-62226-2
438 Solutions of the Exercises
and Var.X/ D EŒX 2 EŒX2 D 1 .

b) We have
Z C1 Z 1
1
EŒU D xf .x/ dx D x dx D
1 0 2
Z C1 Z 1
2 2 1
EŒU D x f .x/ dx D x2 dx D
1 0 3
1 1 1
and Var.U/ D 3 4 D 12 .
c1) We have
P.Z z/ D P.˛U z/ D P.U ˛z / :
Now
8
ˆ
<0
ˆ 0
z
if ˛
P.U ˛/
z
D z
if 0 z
1
ˆ˛ ˛
:̂1 if z
1
˛
hence
8
ˆ
<0
ˆ if z 0
P.Z z/ D z
if 0 z ˛
ˆ˛
:̂1 if z ˛ :
Taking the derivative we find that the density of Z is

8
ˆ
<0
ˆ if z 0
fZ .z/ D 1
if 0 z ˛
ˆ˛
:̂0 if z ˛ ;
i.e. Z is uniform on the interval Œ0; ˛.

c2) If t > 0 we have
Z 1
FW .t/ D P. 1 log U t/ D P.U e t / D dx D 1 e t
e t
whereas FW .t/ D 0 for t < 0. Therefore W has the same p.f. as an exponential
law with parameter and, by Exercise 1.1, has this law.
Solutions of the Exercises 439
1.3
a) Let us denote by the law of X. By the integration rule with respect to an image
law, Proposition 1.1, and by Fubini’s theorem
Z C1 Z C1 Z x
EŒf .X/ D f .x/ d.x/ D d.x/ f .0/ C f 0 .t/ dt
0 0 0
Z C1 Z C1 Z C1
D f .0/ C f 0 .t/ dt d.x/ D f .0/ C f 0 .t/P.X t/ dt :
0 t 0
b) Imitating Fubini’s theorem,

1
X 1
X X
n X 1
1 X 1
X
EŒX D n P.X D n/ D P.X D n/ 1D P.X D n/ D P.X k/ :
nD1 nD1 kD1 kD1 nDk kD1
1.4
a) If is a vector in the kernel of C then, repeating the argument of (1.15),
0 D hC; i D EŒh; Xi2
and therefore h; Xi2 D 0 a.s., i.e. X is orthogonal to a.s. As C is symmetric, its
image coincides with the subspace of the vectors that are orthogonal to its kernel:
z 2 Im C if and only if hz; i D 0 for every vector such that C D 0. Hence,
if 1 ; : : : ; k , k m, is a basis of the kernel of C, then z 2 Im C if and only if
hz; i i D 0 for i D 1; : : : ; k. In conclusion
fX 2 Im Cg D fhX; 1 i D 0g \ \ fhX; k i D 0g
and therefore the event fX 2 Im Cg has probability 1, being the intersection of a

finite number of events of probability 1.
b) If C is not invertible, Im C is a proper subspace of Rm . The r.v. X E.X/ is
centered and has the same covariance matrix as X and, as we have seen, X
E.X/ 2 Im C with probability 1. If X had a density, f say, then we would have
Z
1 D P.X 2 Im C C E.X// D f .x/ dx :
Im CCE.X/
But the integral on the right-hand side is equal to 0 because the hyperplane ImCC
E.X/ has Lebesgue measure equal to 0 so that this is absurd.
1.5 It is well-known that a sequence of real numbers .an /n converges to a limit `

if and only if from every subsequence of .an /n we can extract a further subsequence
converging to `. Therefore kXn Xkp ! 0 as n ! 1 if and only if from every
subsequence of .kXn Xkp /n we can extract a further subsequence converging to 0.

For convergence in probability the argument is similar.
For a.s. convergence this argument cannot be repeated, because to say that a
subsequence converges a.s. means that there exists a negligible set N possibly
depending on the subsequence such that, on N c , Xnk .!/ !k!1 X.!/. The argument
above therefore allows us to say that the sequence .Xn /n converges to X outside the
union of all these negligible events. As the set of all possible subsequences of a
given sequence has a cardinal larger than countable, the union of all these negligible
events may have strictly positive probability.
Note also that, if this criterion was true for a.s. convergence, it would entail,
using the last statement of Proposition 1.5, that convergence in probability implies
a.s. convergence.
1.6 If X N.0; I/, then
Z Z
1 1 2 1 2 1 1 2 1 2
EŒeh
;Xi D eh
;xi e 2 jxj dx D e 2 j
j e 2 jx
j dx D e 2 j
j ;
.2/m=2 Rm .2/m=2 Rm
„ ƒ‚ …
D1
as we recognized the expression of the N.

; I/ density. If, more generally, X
N.b; /, then we know that X D 1=2 Z C b, where Z N.0; I/. Therefore
1=2 ZCbi 1=2
;Zi 1 1=2
j2
EŒeh
;Xi D EŒeh
; D eh
;bi EŒeh D eh
;bi e 2 j :
But j 1=2
j2 D h 1=2
; 1=2
i D h 1=2 1=2
;
i D h
;
i.
1.7 We have
b
XCY .
/ D EŒeih
;XCYi D EŒeih
;Xi eih
;Yi D EŒeih
;Xi EŒeih
;Yi D b
.
/b
.
/ ;
"
where the equality indicated by the arrow follows from Proposition 1.3, recalling
that X and Y are independent.
If X D .X1 ; : : : ; Xm / is a -distributed m-dimensional r.v., then the k-th marginal,
k , is nothing else than the law of Xk . Therefore, if we denote by e
the vector of
dimension m whose components are all equal to 0 but for the k-th one that is equal
to
,
Q
b .e
k .
/ D EŒei
Xk D EŒeih
;Xi D b
/ :
1.8 If
2 R then
Z Z Z
C1
C1 0
EŒe
X D e jxj e
x dx D e.
/x dx C e.
C /x dx :
2 1 2 0 1
It is apparent that in order for both integrals to converge the condition j

j < is
necessary. Once this condition is satisfied the integrals are readily computed and we
obtain the expression of the Laplace transform
1 1 2
EŒe
X D C D 2
2
C

2
As for the characteristic function, since x 7! sin

x is an odd function whereas
x 7! cos
x is even,
Z C1 Z C1 Z C1

EŒei
X D e jxj ei
x dx D e jxj cos
x dx D e jxj cos
x dx :
2 1 2 1 0
The last integral can be computed by parts with some patience and we obtain
2
EŒei
X D
2 C
2
We shall see in Sect. 5.7 that the characteristic function can be deduced from the
Laplace transform in a simple way thanks to the property of uniqueness of the
analytic continuation of holomorphic functions.
1.9 The vector .X; X C Y/ can be obtained from .X; Y/ through the linear map
associated to the matrix

10
AD :
11
As .X; Y/ is N.0; I/-distributed, by the stability property of the normal laws with
respect to linear-affine transformations as seen in Sect. 1.7, .X; X C Y/ has a normal
law with mean 0 and covariance matrix

10 11 11
AA D D :
11 01 12
p
Similarly the vector Z D .X; 2 X/ is obtained from X through the linear
transformation associated to the matrix

1
AD p
2
that allows us to conclude that Z is normal, centered, with covariance matrix AA .
Equivalently we might have directly computed the characteristic function of Z: if
D .
1 ;
2 /, then
p 1 p
EŒeih
;Zi D EŒei.
1 C 2
2 /X
D exp .
1 C 2
2 /2
2
1 2 p
D exp .
1 C 2
22 C 2 2
1
2 /
2
whence we get that it is a normal law with mean 0 and covariance matrix
p
p1 2
:
2 2
p
In particular, the two vectors .X; X C Y/ and .X; 2 X/ have the same marginals
(normal centered of variance 1 and 2 respectively) but different joint laws (the
covariance matrices are different).
1.10 First observe that the r.v. .X1 X2 ; X1 C X2 / is Gaussian, being obtained
from X D .X1 ; X2 /, which is Gaussian, through the linear transformation associated
to the matrix

1 1
AD :
1 1
We must at this point just check that the covariance matrix, , of .X1 X2 ; X1 C X2 /
is diagonal: this will imply that the two r.v.’s X1 X2 , X1 C X2 are uncorrelated and
this, for jointly Gaussian r.v.’s, implies independence. Recalling that .X1 ; X2 / has
covariance matrix equal to the identity, using (1.13) we find that
D AA D 2I:
p p
The same argument applies to the r.v.’s Y2 D 12 X1 23 X2 and Y2 D 12 X1 C 23 X2 . The
vector Y D .Y1 ; Y2 / is obtained from X through the linear transformation associated
to the matrix
p !
1 3

A D p23 2
1
: (S.1)
2 2
Therefore Y has covariance matrix AA D I and also in this case Y1 and Y2 are
independent. Furthermore, Y has the same law, N.0; I/, as X.
Taking a closer look at this computation, we have proved something more
general: if X N.0; I/, then, if A is an orthogonal matrix (i.e. such that A1 D A ),
AX also has law N.0; I/. The matrix A in (S.1) describes a rotation of the plane by
an angle equal to 3 .
1.11
a) Let us compute the law of eX with the method of the partition function. Let us
denote by ˚; and f; , respectively, the partition function and the density of an
N.; 2 / law; for y > 0 we have
P.eX y/ D P.X log y/ D ˚; .log y/ :
By taking the derivative we obtain the density of eX :
d 1 1 1
g; .y/ D ˚; .log y/ D f; .log y/ D p exp 2 .log y /2 :
dy y 2 y 2
b) If X N.; 2 /, then we can write X D C Z, where Z N.0; 1/. Therefore,

using Exercise 1.6, for every p 0,
1 2 2

EŒ.eX /p D EŒep ZCp D ep e 2 p :
2 =2
For p D 1 this gives the mean, which is therefore equal to e e ; as for the
variance
2 2 2 2 2
Var.eX / D EŒe2X EŒeX 2 D e2 e2 e e =2 D e2 e e 1 :
Note that the computations would have been more complicated if we tried to
compute the moments by integrating the density of the lognormal law, which
leads to the nasty looking integral
Z C1
yp g; .y/ dy :
0
1.12 Let us assume first that X N.0; 1/, then

Z C1 Z C1
tX 2 1 tx2 x2 =2 1 1 2 .12t/
EŒe D p e e dx D p e 2 x dx :
2 1 2 1
The integral diverges if t 12 . If, conversely, t < 1

2
we recognize in the integrand
Z C1
1 2
e 2 x .12t/
dx ;
1
up to a constant, the density of a normal

p law with mean 0 and variance .1 2t/1 .
2
Therefore the integral is equal to 2 .1 2t/1=2 , hence EŒetX D C1 if t 12
2
and EŒetX D .1 2t/1=2 if t < 12 . Recalling that if Z N.0; 1/ then X D Z
2 2 2
N.0; 2 /, we have EŒetX D EŒet Z . In conclusion, if X D N.0; 2 /,
(
1
tZ 2
C1 if t 2 2
EŒe D
p 1 if t < 1
:
12 2 t 2 2
1.13 Thanks to Proposition 1.1

Z C1
C 1 bC z C 2
EŒ.xe bC X
K/ D p e K ez =2 dz :
2 1
Note that the integrand vanishes if xebC z K < 0, i.e. if
1 K
z WD log b
x
Hence, with a few standard changes of variable,
Z C1
1 bC z 2
EŒ.xebC X K/C D p xe K ez =2 dz
2
Z C1 Z C1
x 2 1 2
Dp ebC zz =2 dz p K ez =2 dz
2 2
1 2 Z C1
xebC 2 1 2
D p e 2 .z / dz K˚./
2
1 2 Z C1
xebC 2 2
D p ez =2 dz K˚./
2
1 2
D xebC 2 ˚. C / K˚./ :
1.14
a) It is immediate to compute the characteristic functions of the r.v.’s Xn and their
limit as n ! 1: for every
2 Rm we have
1 1
Xn .
/ D eihbn ;
i e 2 hn
;
i ! eihb;
i e 2 h
;
i
n!1
and we recognize on the right-hand side the characteristic function of an

N.b; /-distributed r.v.
b1) X1 D ˛x C Z1 has a normal law (it is a linear-affine function of the normal r.v.
Z1 ) of mean ˛x and variance 2 .
X2 D ˛X1 C Z2 is also normal, being the sum of the two r.v.’s ˛X1 and Z2 ,
which are normal and independent. As
EŒX2 D EŒ˛X1 C EŒZ2 D ˛ 2 x

„ƒ‚…
D0
Var.X2 / D Var.˛X1 / C Var.Z2 / D ˛ 2 2 C 2
we derive that X2 N.˛ 2 x; .1 C ˛ 2 / 2 /. By recurrence
Xn N.˛ n x; 2 .1 C ˛ 2 C C ˛ 2.n1/ // :
Indeed let us assume that this relation is true for a value n and let us prove that
it holds also for nC1. As XnC1 D ˛Xn CZnC1 and the two r.v.’s Xn and ZnC1 are
independent and both normally distributed, XnC1 is also normally distributed.
We still have to check the values of the mean and the variance of XnC1 :
EŒXnC1 D EŒ˛Xn C EŒZnC1 D ˛ ˛ n x D ˛ nC1 x

Var.XnC1 / D Var.˛Xn / C Var.ZnC1 /
D ˛ 2 2 .1 C ˛ 2 C C ˛ 2.n1/ / C 2
D 2 .1 C ˛ 2 C C ˛ 2n / :
As j˛j < 1 we have ˛ n x!0 as n ! 1 and
2
2 .1 C ˛ 2 C C ˛ 2.n1/ / !
n!1 1 ˛2
2
Thanks to a) .Xn /n converges in law to an N.0; 1˛ 2 /-distributed r.v.
b2) The vector .Xn ; XnC1 / is Gaussian as a linear transformation of the vector
.Xn ; ZnC1 /, which is Gaussian itself, Xn and ZnC1 being independent and
Gaussian. In order to compute the limit in law we just need to compute the
limit of the covariance matrices n (we know already that the means converge
to 0). Now
Cov.Xn ; XnC1 / D Cov.Xn ; ˛Xn C ZnC1 / D ˛Cov.Xn ; Xn / C Cov.Xn ; ZnC1 /

D ˛ 2 .1 C ˛ 2 C C ˛ 2.n1/ / :
˛ 2
As n ! 1 this quantity converges to 1˛ 2 . We already know the value of the
limit of the variances (and therefore of the elements on the diagonal of n ), and
we obtain

2 1˛
lim n D :
n!1 1 ˛2 ˛1
The limit law is therefore Gaussian, centered and with this covariance matrix.
1.15
a) The result is obvious for p 2 (even without the assumption of Gaussianity),
thanks to the inequality between Lp norms (1.5). Let us assume therefore that
p 2 and let us consider first the case m D 1. If X is centered and EŒX 2 D 2 ,
we can write X D Z with Z N.0; 1/. Therefore
EŒjXjp D p EŒjZjp D cp EŒjXj2 p=2 :

„ ƒ‚ …
Dcp
For m 2, let X D .X1 ; : : : ; Xm /. Using the result obtained for m D 1 and as

p
2
1, thanks to both the inequalities of the hint,
p2 X
m
p2 X
m
EŒjXjp m 2 EŒjXi jp cp m 2 EŒjXi j2 p=2
iD1 iD1
p2
X
m p=2 p2
cp m 2 EŒjXi j2 D cp m 2 EŒjXj2 p=2 :
iD1
b) First of all L2 -convergence implies L1 -convergence and therefore the conver-

gence of the means: if mn D EŒXn , then limn!1 mn D 0. Let e X n D X n mn .
Then eX n is centered and also e
X n !n!1 0 in L2 . Thanks to a),

EŒjXn jp D EŒje
X n C mn jp 2p1 jmn jp C EŒje
X n jp

D 2p1 jmn jp C cp;m EŒje X n j2 p=2 ! 0:
n!1
1.16
a) For the partition function FY of Y we have, X and Z being independent,
FY .t/ D P.Y t/ D P.XZ t/ D P.XZ t; Z D 1/ C P.XZ t; Z D 1/

1 1
D P.X t; Z D 1/ C P.X t; Z D 1/ D P.X t/ C P.X t/ :
2 2
..............................
........
.....
. .......................................... .
.......
......
. ....................................................................
......................................................................................
...... ...... ...............................................................................................

.......................................................................................................
......
......
.............................................................................................................
. ...................................................................................................................
..
..
. ...... ..........................................................................................................................
.... ...... . ............................................................................................................................................................................
..
..
. ......
. . . . . . . . . . . . . . . . . . . . .
...............................................................................................
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . ................................... . . . . .
.....
.................................................................................................
....................................................................................................................................................
.
...
. ......
...........................................................................................................................................................
.
..... .......
...................................................................................................................................................................
..........................................................................................................................................................................
...
.. ........
..................................................................................................................................................................................
......
..........................................................................................................................................................................................................................................................
.
...
...
.
. . . . . . . . . . . . . . . . .
...........
. .
. . . ..... . . . . . . . . . . . . . . . ...........................................................................................
.
. ...........................................................................................................................................................................................................
.
.. ....................
............................
.. ................................................................................................................................................
. . ...................................................................................................................................................................................................................................
...........
. . . ........................................................................................................................................................................................................................................................
. . . .................................................................................................................................................................................................................................................................................................
0 t
..................................
........
. . ............................ .
............................................................. .
......
......
.............................................................................................................. .
...... ..............................................................
......
....................................................................................................................................... .
......
......
.................................................................................................................................................
...............................................................................................................................................................
.
..
. ...... ........................................................................................................................................................................
. .
.... ...... ......................................................................................................................................................................................
..
.. ...... ............................................................................................................................................................................................ .
.....
............................................................................................................................................................................................................ .
...
. ...... ............................................................................................................
...... ....... .........................................................................................................................................................................................................................

.......................................................................................................................................................................................................................................... .
.
...
... ........ ......................................................................................................................................................................................................................................................... .
...
...
...
...
....... ...........
......................
...........................................................................................................................................................................................................................................................................
.
.............................................................................................................................................................................................................................................................................................. .
....................... ........ ............................................................................................................................................................................................................................................................................................................................ . .

..
...................................................................................................................................................................................................................................................................................................................................................................................... . .
–t 0
Fig. S.1 The two shaded surfaces have the same area
But, with an obvious change of variable,

Z C1 Z t
1 2 1 2
P.X t/ D p ex =2 dx D p ex =2 dx D P.X t/
2 t 2 1
(see also Fig. S.1, for an intuitive explanation). Therefore, going back to the
computation of FY , P.Y t/ D P.X t/ for every t 2 R and X and Y have the
same law. It is also possible to compute the characteristic function of Y: always
using the independence of X and Z,
EŒei
Y D EŒei
XZ D EŒei
XZ 1fZD1g C EŒei
XZ 1fZD1g
1 1
D EŒei
X 1fZD1g C EŒei
X 1fZD1g D EŒei
X C EŒei
X
2 2
1
2 =2 1
2 =2 2
D e C e D e
=2 :
2 2
b) Let us compute the characteristic function of X C Y: we have EŒei
.XCY/ D
EŒei
X.1CZ/ and repeating the argument above for the characteristic function of
Y,
1 1 1 1 2
EŒei
.XCY/ D EŒ1fZD1g C EŒei2
X 1fZD1g D C EŒei2
X D C e2
:
2 2 2 2
It is easy to see that this cannot be the characteristic function of a normal
r.v.: for instance, note that X C Y has mean 0 and variance 2 D 2 (taking
the derivatives of the characteristic function at 0) and if it was Gaussian its
2
characteristic function would be
7! e
. The pair .X; Y/ cannot therefore
be jointly normal: if it was, then X C Y would also be Gaussian, being a linear
function of .X; Y/.
1.17
a) By the Borel–Cantelli lemma, Proposition 1.7, (1.29) holds if
1
X
P.Xn .˛ log n/1=2 / < C1 : (S.2)
nD1
Thanks to the inequality of the hint (the one on the right-hand side)
Z C1
1 2 =2
P.Xn .˛ log n/1=2 / D p ey dy
2 .˛ log n/1=2
1 1 1
p e 2 ˛ log n D p
2 ˛ log n 2 ˛ log n n˛=2
As ˛2 > 1, the series in (S.2) is summable.

b) Again in view of the Borel–Cantelli lemma we have to investigate the summa-
bility of P.Xn .2 log n/1=2 /. By the other inequality of the hint
1 1 log n 1
P.Xn .2 log n/1=2 / p .2 log n/1=2 C .2 log n/1=2 e p
2 2n log n
which is the term of a divergent series. Hence, by the Borel–Cantelli lemma
P.Xn .2 log n/1=2 infinitely many times/ D 1 :
c) The r.v.’s .log n/1=2 Xn have zero mean and variance, .log n/1 , tending to 0 as
n ! 1. By Chebyshev’s inequality, for every ˛ > 0,
ˇ X ˇ 1
P ˇp ˇ˛
n
2
! 0:
log n ˛ log n n!1
Therefore .log n/1=2 Xn !n!1 0 in probability. On the other hand, by b), we

have, with probability 1,
Xn p
p 2 for infinitely many indices n
log n
and therefore it is not possible for the sequence to converge to 0 a.s.
1.18 Let us first check the formula for f D 1A , with A 2 E . The left-hand side is
obviously equal to X .A/.
Note that 1A .X/ D 1fX2Ag: X.!/ 2 A if and only if ! 2 fX 2 Ag. Therefore the
right-hand side is equal to P.X 2 A/ and, by the definition of image probability, (1.7)
is true for every function that is the indicator of an event.
By linearity (1.7) is also true for every function which is a linear combination of
indicator functions, i.e. for every elementary function.
Let now f be a positive measurable function on E. By Proposition 1.11 there
exists an increasing sequence . fn /n of elementary functions converging to f .
Applying Beppo Levi’s theorem twice we find
Z Z Z Z
f dX D lim fn dX D lim fn .X/ dP D f .X/ dP (S.3)
E n!1 E n!1 ˝ ˝
so that (1.7) is satisfied for every positive function f and, by taking its decomposition
into positive and negative parts, for every measurable function f .
1.19 Let us consider the family, E 0 say, of the sets A 2 E such that X 1 .A/ 2 F .
Using the relations
[
1 [1
X 1 .A/c D X 1 .Ac /; X 1 An D X 1 .An /
nD1 nD1
it is easy to check that E 0 is a -algebra contained in E . It contains the class D

hence also E . Therefore X 1 .A/ 2 F for every A 2 E and X is measurable.
1.20
a) As Zn ! Z a.s. we also have Hn ! Z a.s. and, as 0 Hn Z, one can apply
Lebesgue’s theorem.
b) Note that if Z Zn then jZ Zn j D Z Zn D Z Hn whereas Zn Hn D 0 and
the relation is proved. Repeating the argument changing the roles of Z and Zn we
see that the relation also holds if Zn Z. We now have
EŒjZ Zn j D EŒZ Hn C EŒZn Hn D 2EŒZ Hn ! 0:

n!1
1.21
a) By the Central Limit Theorem the sequence
X1 C C Xn n
Yn D p
n
converges in law to an N.0; 1/-distributed r.v., where and 2 are respectively

the mean and the variance of X1 . Obviously D 12 , whereas
Z 1
1
EŒX12 D x2 dx D
0 3
and therefore 2 D 13 14 D 12
1
. W is nothing else than Y12 . We still have to see
whether n D 12 is a number large enough for Yn to be approximatively N.0; 1/
...
b) We have, integrating by parts,
Z Z C1
1 C1
1 3 x2 =2 ˇˇC1
2 =2 2

EŒX 4 D p x4 ex
dx D p x e ˇ C3 x2 ex =2 dx
2 1 2 1 1
Z C1
1 2
D3p x2 ex =2 dx D 3 :
2 1
„ ƒ‚ …
DVar.X/D1
The computation of the moment of order 4 of W is a bit more complicated. If

Zi D Xi 12 , then the r.v.’s Zi are independent and uniform on Œ 12 ; 12 and
EŒW 4 D EŒ.Z1 C C Z12 /4 :
As EŒZi D EŒZi3 D 0 (the Zi ’s are symmetric with respect to the origin), the
expectation of many terms appearing in the expansion of .Z1 C C Z12 /4
vanishes. For instance, as the r.v.’s Zi ; i D 1; : : : ; 12, are independent,
EŒZ13 Z2 D EŒZ13 EŒZ2 D 0 :
A moment of reflection shows that a non-zero contribution is given only by the

terms of the form EŒZi2 Zj2 D EŒZi2 EŒZj2 with i 6D j and those of the form E.Zi4 /.
The term Zi4 clearly has a coefficient D 1 in the expansion. In order to determine
the coefficient of Zi2 Zj2 ; i 6D j, we can observe that in the power series expansion
of the function
.x1 ; : : : ; x12 / D .x1 C C x12 /4
the monomial x2i x2j , for i 6D j, has the coefficient
1 @4 1
.0/ D 24 D 6 :
2Š2Š @x2i @x2j 4
Let us compute now

Z 1=2 Z 1=2
1 1
EŒZi2 D x2 dx D , EŒZi4 D x4 dx D
1=2 12 1=2 80
As, by symmetry, the terms of the form EŒZi2 Zj2 ; i 6D j, are all equal and there are
11 C 10 C C 1 D 12 12 11 of them, their contribution is
1 1 11
6 12 11 D
2 144 4
The contribution of the terms of the form EŒZi4 (there are 12 of them), is
conversely 12
80
. In conclusion
11 12
EŒW 4 D C D 2:9 :
4 80
In practice W turns out to have a law quite close to an N.0; 1/. It is possible to
compute its density and to draw its graph, which is almost indistinguishable from
the graph of the Gaussian density.
However it has some drawbacks: for instance W cannot take values outside
the interval Œ6; 6 whereas an N.0; 1/ can, even if with a very small probability.
In practice W can be used as a substitute for the Box–Müller algorithm of
Proposition 1.10 for tasks that require a moderate number of random numbers.
2.1
a) X and Y are equivalent because two r.v.’s that are a.s. equal have the same law: if
t1 ; : : : ; tn 2 T, then
[
n
f.Xt1 ; : : : ; Xtn / 6D .Yt1 ; : : : ; Ytn /g D fXti 6D Yti g :
iD1
These are negligible events, being finite unions of negligible events. Therefore,
for every A 2 E ˝n , the two events f.Xt1 ; : : : ; Xtn / 2 Ag and f.Yt1 ; : : : ; Ytn / 2 Ag
can differ at most by a negligible event and thus have the same probability.
b) As the paths of the two processes are a.s. continuous but for a negligible event,
if they coincide at the times of a dense subset D T, they necessarily coincide
on the whole of T. Let D D ft1 ; t2 ; : : : g be a sequence of times which is dense in
T (T \ Q, e.g.). Then
\ \
fXt D Yt for every tg D fXt D Yt g D fXti D Yti g :
t2T ti 2D
As P.Xti D Yti / D 1 for every i, it follows that

\
P fXti D Yti g D 1
ti 2D
so that also P.Xt D Yt for every t/ D 1 and the two processes are indistinguish-
able.
2.2
a) First it is clear that .Xt ; t T/ .Xt ; t 2 D/: every r.v. Xs , s 2 D, is obviously
.Xt ; t T/-measurable and .Xt ; t 2 D/ is by definition the smallest -algebra
that makes the r.v.’s Xs , s 2 D, measurable.
In order to show the converse inclusion, .Xt ; t T/ .Xt ; t 2 D/, we need
only show that every r.v. Xt ; t T, is measurable with respect to .Xt ; t 2 D/.
But if t T there exists a sequence .sn /n D such that sn ! t as n ! 1. As
the process is continuous, Xsn ! Xt and therefore Xt is the limit of .Xt ; t 2 D/-
measurable r.v.’s and is therefore .Xt ; t 2 D/-measurable itself.
b) The argument of a) can be repeated as is, but now we must choose the sequence
.sn /n decreasing to t and then use the fact that the process is right-continuous.
This implies that all the r.v.’s Xt with t < T are .Xt ; t 2 D/-measurable, but this
argument does not apply to t D T, as there are no times s 2 D larger than T.
Therefore if the process is only assumed to be right-continuous, it is necessary

to make the additional assumption T 2 D.
2.3 By hypothesis, for every u > 0, the map .Œ0; u ˝/; B.Œ0; u Fu / !
.E; E // defined as .t; !/ ! Xt .!/ is measurable. Now note that, as the composition
of measurable function is also measurable, .t; !/ ! .Xt .!// is measurable
.Œ0; u ˝/; B.Œ0; u Fu / ! .G; G //, i.e. t 7! .Xt / is progressively measurable.
2.4
a) To say that the sequence .Zn .!//n does not converge to Z1 .!/ is equivalent
to saying that there exists an m 1 such that for every n0 1 there exists an
n n0 such that jZn Z1 j m1 , which is exactly the event (2.2).
b1) Thanks to a) we know that
n o 1 \
[ 1 [
1
lim Xtn 6D ` D fjXtn `j m
g 2F (S.4)
n!1
mD1 n0 D1 nn0
is negligible. Let us prove that the corresponding event for e X is also negligible.
As the r.v.’s .Xtn ; : : : ; XtnCk / and .e
X tn ; : : : ; e
X tnCk / have the same distribution, the
events (belonging to different probability spaces)
[
k [
k
fjXtn `j 1
m
g and fje
Xtn `j 1
m
g
nDn0 nDn0
have the same probability. As these events are increasing in k, we have that
1
[ 1
[
fjXtn `j 1
m
g and fje
X tn `j 1
m
g (S.5)
nDn0 nDn0
also have the same probability. As the event in (S.4) is negligible and observing
that the events in (S.5) are decreasing in m; n0 , for m; n0 large
[
1
1
P fjXtn `j mg "
nDn0
therefore also
[
1
e
P fje
X tn `j 1
g ";
m
nDn0

! 2 e̋
b2) A repetition of the arguments of a) allows us to state that the event of the e
such that the limit (2.3) does not exist is
[ 1
1 \ [
fje
Xq `j 1
m
g
mD1 kD1 q2Q;jqtj 1
k
and, by a repetition of the argument of b1), this event has the same probability
as its analogue for X, i.e. 0.
2.5
a) We have E.Xs Xt / D E.Xs /E.Xt / D 0 for every s 6D t, the two r.v.’s being
independent and centered. Therefore the function .s; t/ 7! E.Xs Xt / vanishes but
on the diagonal s D t, which is a subset of Lebesgue measure 0 of Œa; b2 . Hence
the integral in (2.4) vanishes.
Rb
b) As .Xt /t is assumed to be measurable, the map ! 7! a Xs .!/ ds is a r.v. (recall
Example 2.2). By Fubini’s theorem
Z b Z b
E Xs ds D EŒXs ds D 0 : (S.6)
a a
Also the map .s; t; !/ 7! Xs .!/Xt .!/ is measurable and again by Fubini’s
theorem and a)
h Z b 2 i hZ b Z b i Z b Z b
E Xs ds DE Xs ds Xt dt D E.Xs Xt / ds dt D 0 :
a a a a a
(S.7)
Rb
The r.v. a Xs ds, which is centered by (S.6), has variance 0 by (S.7). Hence it is
equal to 0 a.s.
Rb
c) From b) we have that, for every a; b 2 Œ0; 1, a b, a Xs ds D 0 a.s.
Therefore, for almost every !, the function t 7! Xt .!/ is such that its integral
on a subinterval Œa; b vanishes for every a b and it is well-known that such a
function is necessarily 0 a.e.
Rb
Actually the previous argument is not completely correct as, if a Xs .!/ ds D
0 but for a negligible event, this event Na;b ˝ might depend on a and b,
whereas in the previous argument we needed a negligible event N such that
Rb
a Xs .!/ ds D 0 for every a; b. In order to deal with this question just set
[
ND Na;b :
a;b2Q
N is still negligible, being the union of a countable family of negligible events.

For every ! 62 N therefore
Z b
Xs .!/ ds D 0; for every a; b 2 Q
a
and a function whose integral on all the intervals with rational endpoints vanishes
is necessarily D 0 almost everywhere.
• To be rigorous, in order to apply Fubini’s theorem in (S.7) we should first
prove that s; t; ! 7! Xs .!/Xt .!/ is integrable. But this is a consequence of
Fubini’s theorem itself applied to the positivep measurable function s; t; ! 7!
jXs .!/Xt .!/j. Indeed, as EŒjXt j EŒXt2 D c,
hZ b Z b i Z b Z b
E ds jXs Xt j dt D E.jXs Xt j/ ds dt
a a a a
Z b Z b
D E.jXsj/E.jXt j/ ds dt c.b a/2
a a
hence s; t; ! 7! Xs .!/Xt .!/ is integrable.
2.6
a) We have
1
Z .A;t;" / D f!I jt Zt .!/j "g:
Therefore Z1 .A;t;" / 2 F , as this set is the inverse image through Zt of the
closed ball of Rm centered at t and with radius ".
b) As the paths are continuous,
1
˚
Z .U ;T;" /
D !I jt Zt .!/j " for every t 2 Œ0; T
˚
D !I jr Zr .!/j " for every r 2 Œ0; T \ Q
\ \
1
D f!I jr Zr .!/j "g D Z .A;r;" / :
r2Œ0;T\Q r2Œ0;T\Q
Therefore, thanks to a), Z1 .U ;T;" / 2 F , as a countable intersection of events

of F .
c) Keeping in mind Exercise 1.19 we must prove that the sets (2.5) generate
the Borel -algebra B.C /. For every 2 C the sets U ;T;" form a basis of
neighborhoods of and, as C is separable, every open set of C is the countable
union of sets of the form (2.5). Hence the -algebra generated by the sets (2.5)
contains the Borel -algebra.
3.1
a) With the usual trick of separating the actual position from the increment
EŒBs B2t D EŒBs .Bt Bs CBs /2 D EŒBs .Bt Bs /2 C2EŒB2s .Bt Bs /CEŒB3s D 0 :
More easily the clever reader might have argued that Bs B2t has the same law as
Bs B2t (.Bt /t is again a Brownian motion), from which EŒBs B2t D 0 (true even
if t < s).
b) Again
EŒB2s B2t D EŒB2s .Bt Bs C Bs /2

D EŒB2s .Bt Bs /2 C2 EŒB3s .Bt Bs / C EŒB4s D s.t s/ C 3s2 :
„ ƒ‚ … „ ƒ‚ … „ƒ‚…
Ds.ts/ D0 D3s2
p
c) We know that, if Z denotes an N.0; 1/-distributed r.v., then Bs s Z. Hence,
integrating by parts and recalling the expression for the Laplace transform of the
Gaussian r.v.’s (Exercise 1.6),
p Z C1 p
p p s 2
EŒBs eBs D EŒ s Ze s Z D p ze s z ez =2 dz
2 1
p p ˇ Z C1 p
s 2 ˇC1 s 2
D p e s z ez =2 ˇ Cp e s z ez =2 dz D ses=2 :
2 1 2 1
„ ƒ‚ …
D0
d) We have EŒBs eBt D EŒBs eBs eBt Bs and as Bs and Bt Bs are independent
EŒBs eBt D EŒBs eBs EŒeBt Bs D ses=2 e.ts/=2 D set=2 :
3.2
a) We have
p a
EŒ1fBt ag D P.Bt a/ D P. tB1 a/ D P B1 p
t
and therefore
1
lim EŒ1fBt ag D P.B1 0/ D
t!C1 2
b)
Z a x2 x2 ˇˇa
1 t
EŒBt 1fBt ag D p xe 2t dx D p e 2t ˇ
2t 1 2t 1
p 2
t a
D p e 2t ! 1 :
2 t!C1
p
3.3 Recalling that Bt t B1 , we are led to the computation of
Z C1
p 2 tZ 2 t3=2 2 1 2
lim t EŒtZ e D lim p x2 etx e 2 x dx
t!C1 t!C1 2 1
3=2 Z C1 Z C1
t 2 12 .2tC1/x2 t3=2 1 2
D lim p xe dx D lim p x2 e 2 2 x dx ;
t!C1 2 1 t!C1 2 1
1
where Z denotes an N.0; 1/-distributed r.v. and we have set 2 D 2tC1 . With this
position we are led back to the expression of the variance of a centered Gaussian
r.v.:
Z C1
1 1 2 t3=2
D lim t 3=2
p x2 e 2 2 x dx D lim 3 t3=2 D lim
t!C1 2 1 t!C1 t!C1 .2t C 1/3=2
„ ƒ‚ …
D 2
1
D 23=2 D p
8
3.4
a) As fBtm Btm1 2 m g is independent of Ftm1 whereas all the other events are
Ftm1 -measurable,
P.Btm Btm1 2 m ; : : : ; Bt1 Bs 2 1 ; A/

D P.Btm Btm1 2 m /P.Btm1 Btm2 2 m1 ; : : : ; Bt1 Bs 2 1 ; A/ :
Iterating this procedure m times we have
P.Btm Btm1 2 m ; : : : ; Bt1 Bs 2 1 ; A/

D P.Btm Btm1 2 m /P.Btm1 Btm2 2 m1 / : : : P.Bt1 Bs 2 1 /P.A/
D P.Btm Btm1 2 m ; : : : ; Bt1 Bs 2 1 /P.A/ :
b) The r.v.’s Btm Bs ; : : : ; Bt1 Bs are functions of Btm Btm1 , Btm1 Btm2 ,. . . ,
Bt1 Bs , so that they are .Btm Btm1 ; Btm1 Btm2 ; : : : ; Bt1 Bs /-measurable
and therefore
.Btm Bs ; : : : ; Bt1 Bs / .Btm Btm1 ; Btm1 Btm2 ; : : : ; Bt1 Bs / :
But Btm Btm1 ; Btm1 Btm2 ; : : : ; Bt1 Bs are also functions of Btm Bs ; : : : ; Bt1
Bs and by the same argument we have the opposite inclusion.
c) Thanks to a) and b) we have, for s t1 < < tm and 1 ; : : : ; m 2 B.Rd /,
P.Btm Bs 2 m ; : : : ; Bt1 Bs 2 1 ; A/ D P.Btm Bs 2 m ; : : : ; Bt1 Bs 2 1 /P.A/
if s t1 < < tm . But the events
fBtm Bs 2 m ; : : : ; Bt1 Bs 2 1 g
form a class that is stable with respect to finite intersections and generates .Bt
Bs ; t s// and we can conclude the argument using Remark 1.1.
3.5
a) It is immediate that C is stable with respect to finite intersections. Also C
contains Fs (just choose G D ˝) and G (choose A D ˝). Therefore the -
algebra generated by C also contains F e s D Fs _ G (which, by definition, is
the smallest -algebra containing Fs and G ). The converse inclusion .C /
Fs _ G is obvious.
b) We must prove that, for s t, Bt Bs is independent of F e s . By Remark 1.1
and a) it is enough to prove that, for every Borel set 2 B.R/ and for every
A 2 Fs ; G 2 G ,
P.fBt Bs 2 g \ A \ G/ D P.Bt Bs 2 /P.A \ G/ :
Now fBt Bs 2 g \ A 2 Ft is independent of G (G and Ft are independent).

Therefore
P.fBt Bs 2 g \ A \ G/ D P.fBt Bs 2 g \ A/P.G/

D P.fBt Bs 2 g/P.A/P.G/ D P.Bt Bs 2 /P.A \ G/
where we used the fact that fBt Bs 2 g and A are independent.
3.6
a) The joint law of .Bs ; Bt / is a centered Gaussian distribution with covariance
matrix

ss
CD :
st
We have

1 1 t s
C D
s.t s/ s s
so that, defining z D .x; y/, the joint density of .Bs ; Bt / is
1 1 1 1 1
.tx2 Csy2 2sxy/
fs;t .z/ D p e 2 hC z;z;i D p e 2s.ts/
2 s.t s/ 2 s.t s/
1 1
..ts/x2 C.sx2 Csy2 2sxy//
D p e 2s.ts/
2 s.t s/
1 1
2s.ts/ ..ts/x2 Cs.xy/2 /
D p e
2 s.t s/
from which (3.19) follows.

b) Of course
Z 0 Z C1
P.Bs < 0; B2s > 0/ D dx fs;2s .x; y/ dy
1 0
Z 0 Z C1
1 1 2 1 1 2
D p e 2s x dx p e 2s .yx/ dy
1 2s 0 2s
and with the change of variable z D x y in the inner integral,

Z 0 Z
1 1 2 x
1 1 2
P.Bs < 0; B2s > 0/ D p e 2s x dx p e 2s z dz :
1 2s 1 2s
Let
Z
1 x 1 2
˚s .x/ D p e 2s z dz
2s 1
be the partition function of the N.0; s/ distribution, then the previous relation can
be written as
Z 0 ˇ0
1 ˇ 1
P.Bs < 0; B2s > 0/ D ˚s0 .x/˚s .x/ dx D ˚s .x/2 ˇ D
1 2 1 8
1
since ˚s .0/ D 2 for every s > 0.
3.7
a1) X is a Gaussian process, as the random vector .Xt1 ; : : : ; Xtm / is a linear function
of the vector .Be2t1 ; : : : ; Be2tm / which is Gaussian itself. If s t
Cov.Xt ; Xs / D EŒXt Xs D et es EŒBe2t Be2s D e.tCs/ e2s D e.ts/ D ejtsj :
a2) We have, for 1 i; j m, Cov.Xti Ch Xtj Ch / D ejti Ch.tj Ch/j D

ejti tj j D Cov.Xti Xtj /. Therefore the two centered Gaussian random vectors
.Xt1 Ch ; : : : ; Xtm Ch / and .Xt1 ; : : : ; Xtm / have the same covariance matrix, hence
the same law.
a3) K, being the covariance kernel of a stochastic process, is necessarily positive
definite as explained in Remark 2.5.
b1) A repetition of the argument of a1) gives that W is a centered Gaussian process.
Moreover, if v u,
r
p p p 1 p v
EŒWu Wv D EŒ u X 1 log u v X 1 log v D uv e 2 .log ulog v/ D uv Dv
2 2 u
so that W is a natural Brownian motion thanks to Proposition 3.1.
b2) Two possibilities. First one may try to apply Theorem 2.1 (Kolmogorov’s
continuity theorem): the r.v. Xt Xs is centered Gaussian with covariance
Var.Xt Xs / D Var.Xt / C Var.Xs / 2Cov.Xt ; Xs / D 2.1 e.ts/ / :

p
Therefore Xt Xs 2.1 e.ts/ / Z, where Z N.0; 1/. Hence
ˇ=2
EŒjXt Xs jˇ D 2.1 e.ts/ / EŒjZjˇ :
Now, for every ˛ > 0, j1 e˛ j ˛, as the function x 7! ex has a derivative
that is 1 in absolute value for x 0. We have then
EŒjXt Xs jˇ cˇ jt sjˇ=2 D cˇ jt sj1C.ˇ=21/ :
Hence, by choosing ˇ large enough, X has a modification that is Hölder

continuous with exponent for every < 12 , very much similarly as for the
Brownian motion.
More quickly just note that W, being a Brownian motion, has a continuous
e then e
modification. If we denote it by W, X t D et We e2t is a continuous
modification of X.
3.8
a) It is immediate that X is a Gaussian process (for every t1 < t2 ; : : : tm the r.v.
.Xt1 ; : : : ; Xtm / is a linear function of .Bt1 ; : : : ; Btm / which is Gaussian) and that
Cov.X1 .t/; X2 .s// D .t ^ s/ :

Moreover, X2 is obviously a Brownian motion and as
Cov.X1 .t/; X1 .s// D 2 .t ^ s/ C .1 2 /.t ^ s/ D t ^ s
the same is true for X1 .

b) As X is a Gaussian process, then B is also Gaussian (same argument as in a)).
Let jj < 1. B2 being already a real Brownian motion we must prove that B1 is a
Brownian motion and that B1 and B2 are independent. Now, for s t,
Cov.B1 .t/; B1 .s//

1
D Cov.X1 .t/; X1 .s// Cov.X1 .t/; X2 .s//
1 2 1 2
2
2
Cov.X2 .t/; X1 .s// C Cov.X2 .t/; X2 .s//
1 1 2
s
D 2
1 2 2 C 2 D s
1
and
1
Cov.B1 .t/; B2 .s// D p Cov.X1 .t/; X2 .s// p Cov.X2 .t/; X2 .s// D 0 :
1 2 1 2
• Note that in b) the condition 6D 1; 1 is needed.
3.9 If X is a Brownian motion then Xt Xs D A.Bt Bs / must be N.0; .t s/I/-

distributed. It is Gaussian, being a linear function of a jointly Gaussian r.v., and its
covariance matrix is C D .t s/AA (see (1.13)). We therefore have the condition
AA D I, or A D A1 . A must therefore be orthogonal. On the other hand, if A is
orthogonal, X is a Brownian motion with respect to the filtration .Ft /t , as Xt Xs D
A.Bt Bs / is independent of Fs and a) of Definition 3.2 is immediate.
3.10 We have
Z C1
.SA / D 1fBt 2Ag dt :
0
Therefore, with two strokes of Fubini’s theorem,

hZ C1 i Z C1
EŒ .SA / D E 1fBt 2Ag dt D P.Bt 2 A/ dt
0 0
Z C1 Z jxj2
Z Z C1 jxj2
1 1
D dt e 2t dx D dx e 2t
dt :
0 .2t/m=2 A A 0 .2t/m=2
1
With the changes of variable s D t and subsequently u D 12 jxj2 s,
Z C1 jxj2
Z C1
1 2t 1 1 2
s2C 2 e 2 jxj s ds
m
e dt D
0 .2t/ m=2 .2/ m=2
0
Z C1 Z C1
1 2u 2C m2 2 u 1 2m m
D 2 2
e du D jxj u2C 2 eu du :
.2/ m=2
0 jxj jxj 2 m=2
0
The last integral diverges (at 0C) if 2 C m

2 1, i.e. if m 2. On the other hand,
if m 3, it is equal to . m2 1/.
3.11
a) If 2 G, as the paths of X are continuous, the integral in (3.20) is the limit, for
every !, of its Riemann sums, i.e.
Z X
X D Xs d.s/ D lim Xi=n .Œ ni ; .iC1/
n
Œ/ :
n!1
i0
„ ƒ‚ …
DIn . /
Now if 1 ; : : : ; m 2 G, then the vector In D .In .1 /; : : : ; In .m // is Gaussian,

being a linear function of the r.v.’s Xi=n ; i D 0; 1; : : : , which are jointly
Gaussian.
On the other hand limn!1 In D .X1 ; : : : ; Xm / a.s.; as a.s. convergence
implies convergence in law, by Proposition 1.9 the r.v. .X1 ; : : : ; Xm / is
Gaussian and therefore .X / 2G is a Gaussian family.
b1) In order to show that Y is a Gaussian process, just observe that we can write
Z t Z
Yt D Xs .ds/ D Xs .ds/ ;
0
where d D 1Œ0;t d. One can therefore apply what we have already seen in a).
b2) Yt is centered for every t as, by Fubini’s theorem,
Z t Z t
E.Yt / D E Xu d.u/ D E.Xu / d.u/ D 0 :
0 0
If Ks;t D E.Yt Ys / denotes its covariance function, then

Z t Z s Z t Z s
Ks;t D E Xu d.u/ Xv d.v/ D d.u/ E.Xu Xv / d.v/
0 Z 0t Z s 0 0
D d.u/ u ^ v d.v/ :
0 0
(S.8)
Therefore
Z t Z t
t2 D Kt;t D d.u/ u ^ v d.v/
0 0
Z t Z u Z t Z t
D d.u/ v d.v/ C d.u/ u d.v/ D I1 C I2 :
0 0 0 u
By Fubini’s theorem
Z t Z v
I2 D d.v/ u d.u/ D I1 :
0 0
Moreover,
Z u Z u Z v Z u Z u Z u
v d.v/ D d.v/ dr D dr d.v/ D .r; u/ dr
0 0 0 0 r 0
and therefore, using (3.21),

Z Z Z Z Z
t u t t
1 t
I1 D d.u/ .r; u/ dr D dr .r; u/ d.u/ D .r; t/2 dr ;
0 0 0 r 2 0
which concludes the computation of the variance of Yt . In order to compute the

covariance Cov.Ys ; Yt / D Ks;t , we have, starting from (S.8) and assuming s t,
Z s Z s Z t Z s
Ks;t D d.u/ u ^ v d.v/ C d.u/ u ^ v d.v/
0 0 s 0
Z s Z t Z s
2
D .r; s/ dr C d.u/ v d.v/
0 s 0
Z s Z s Z s
D .r; s/2 dr C .s; t/ .r; s/ dr D .r; t/.r; s/ dr :
0 0 0
Only for completeness let us justify (3.21). .r; t/2 is nothing else than the
measure, with respect to ˝ , of the square r; tr; t, whereas the integral
on the right-hand side is the measure of the shaded triangle in Fig. S.2. The
rigorous proof can be done easily with Fubini’s theorem.
t .....................................................................................................................................................................................................................
......................................................................................................................................................................................................................................................................................
.........................................................................................................................................
......................................................................................................................................
......................................................................................................................................................................................................................................................................
.
...............................................................................................................................
..........................................................................................................................................................................................................................................................
..................................................................................................................................................................................................................................................
.
..........................................................................................................................................................................................................................................
..................................................................................................................................................................................................................................
.
.........................................................................................................................................................................................................................
.
..................................................................................................................................................................................................................
..........................................................................................................................................................................................................
.
..................................................................................................................................................................................................
..............................................................................................
.............................................................................................
..........................................................................................
.
..............................................................................................................................................................................
.....................................................................................................................................................................
...............................................................................................................................................................
.
............................................................................
..........................................................................
..............................................................................................................................................
.
.....................................................................................................................................
..............................................................................................................................
............................................................
.
..................................................................................................................
..........................................................................................................
..................................................................................................
.
..........................................................................................
..................................................................................
.
.........................................................................
...................................
.................................
.........................................................
.
...........................
.............................................
.....................................
.
..................
..............
......................
.
........
.
r
......
. . . . . . . . . . . . . . .
......
r t
Fig. S.2 The integration domain in (3.21)
3.12
a) The integral that appears in the definition of Zt is convergent for almost
every ! as, by the Iterated Logarithm Law, with probability 1, jBt j
..2 C "/t log log 1t /1=2 for t in a neighborhood of 0 and the function t 7!
t1=2 .log log 1t /1=2 is integrable at 0C. Let us prove that Z is a Gaussian process,
i.e. that, for every choice of t1 ; : : : ; tm , the r.v. e
Z D .Zt1 ; : : : ; Ztm / is Gaussian. We
cannot apply Exercise 3.11 b) immediately because 1u du is not a Borel measure
(it gives infinite mass to every interval containing 0). But, by Exercise 3.11 a),
.n/ .n/ .n/
the r.v. e
Z .n/ D .Zt1 ; : : : ; Ztm / is indeed Gaussian, where Zti D 0 if ti < 1n and
Z ti
.n/ Bu
Zti D Bti du i D 1; : : : ; m
1=n u
otherwise. Then just observe that limn!1 e

Z .n/ D e
Z a.s. and that a.s. convergence
implies convergence in law. We deduce that e Z is Gaussian, by Proposition 1.9.
Clearly Z0 D 0 and Zt is centered. We only need to compute the covariance
function of Z and verify that it coincides with that of the Brownian motion.
Repeatedly using Fubini’s theorem we have for s t,
h Z Z
Bu i s
Bv t
E.Zs Zt / D E Bs du dv Bt
0 0 u v
Z s Z t Z s Z t
E.Bt Bv / E.Bs Bu / E.Bv Bu /
D E.Bs Bt / dv du C dv du
0 v 0 u 0 0 uv
Z s Z s Z t Z s Z t
s vû
Ds dv du du C dv du
0 0 u 0 0 uv
s
„ ƒ‚ …
DI
D s s.log t log s/ C I :
Let us compute the double integral:

Z Z v Z Z
s
1 s t
1
ID dv du C dv du
0 0 v 0 v u
Z s
DsC .log t log v/ dv D s C s log t s log s C s D 2s C s.log t log s/
0
so that, finally, E.Zs Zt / D s D s ^ t, which, thanks to Proposition 3.1, completes

the proof that .Zt /t is a natural Brownian motion.
b) .Zt /t is clearly adapted to .Ft /t (see Exercise 2.2 a)). In order to show that it is
not a Brownian motion with respect to .Ft /t , there are many possible approaches.
For instance, if it was Brownian, then Zt Zs would be independent of Bs , that is
an Fs -measurable r.v. Instead we have
Z Z
t
E.Bs Bu / t
s t
EŒ.Zt Zs /Bs D EŒ.Bt Bs /Bs du D du D s log 6D 0 :
s u s u s
c) Arguing as in a) Zs and Bt are jointly Gaussian. Hence, in order to prove that Bt

is independent of Gt , by Remark 1.1 it is enough to prove that Bt is independent
of Zs for every s t, i.e. that they are uncorrelated. Actually
Z Z
s
E.Bu Bt / s
E.Zs Bt / D E.Bs Bt / du D s du D 0 :
0 u 0
3.13 We have
p
fBt a t for every t Tg
n Bt a o
D 1=2 q for every t T
2t log log 1t 2 log log 1t
but by the Iterated Logarithm Law, with probability 1
Bt
lim q D1
2t log log 1t
t!0C
hence, with probability 1 there exists a sequence of times .tn /n such that
Btn
q ! 1
2tn log log t1n
n!1
whereas
a
q ! 0
2 log log t1n
n!1
and therefore the event in question has probability 0.

3.14
p
a) By the Iterated Logarithm Law, see also (3.8), jBt j .1 C "/ 2t log log t for t
large. Therefore if b < 0
lim Xt D 0
t!C1
whatever the value of . The same arguments give limt!C1 Xt D C1 if b >

0 (again for every ). If b D 0 and > 0 we know, by the behavior of the
Brownian motion as t ! C1 as described in Remark 3.4, that limt!C1 Xt D
C1, limt!C1 Xt D 0.
b) We have
2
EŒXt D ebt EŒe Bt D e.bC 2 /t
2
so that the limit (3.22) is finite if and only if b 2 and equal to C1
2
otherwise. Observe the apparent contradiction: in the range b 2 2 ; 0Œ we
have limt!C1 Xt D 0 a.s., but limt!C1 EŒXt D C1.
3.15
p
a) By the Iterated Logarithm Law jBu j .1 C "/ 2u log log u for t large. Hence,
if b > 0, ebuC Bu !u!C1 C1 (this is also Exercise 3.14 a)) and in this case
the integrand itself diverges, hence also the integral. If b < 0, conversely, we
have, for t large,
p
ebuC Bu exp bu C .1 C "/ 2u log log u ebu=2
and the integral converges to a finite r.v.

b1) The integral can vanish only if the integrand, which is 0, vanishes a.s.
However, we know by the Iterated Logarithm Law (see Remark 3.4) that the
Brownian path takes a.s. strictly positive values in every neighborhood of 0. As
the paths are continuous they are therefore strictly positive on a set of times of
strictly positive Lebesgue measure a.s.
b2) By a change of variable and using the scaling properties of the Brownian
motion, as v 7! p1 t Btv is also a Brownian motion,
Z t Z 1 Z 1 Z 1
L
1fBu >0g du D t 1fBtv >0g dv D t 1f p1 Btv >0g dv t 1fBv >0g dv :
0 0 0 t 0
Now
Z 1
lim t 1fBu >0g du D C1
t!C1 0
R1
as we have seen in b1) that the r.v. 0 1fBu >0g du is strictly positive a.s. Hence,
Rt R1
as the two r.v.’s 0 1fBu >0g du and t 0 1fBu >0g du have the same distribution for
every t, we have
Z t
lim 1fBu >0g du D C1 in probability :
t!C1 0
In order to prove the a.s. convergence, it suffices to observe that the limit
Z t
lim 1fBu>0g du
t!C1 0
exists a.s. as the integral is an increasing function of t. Hence the proof

is complete, because the a.s. limit and the limit in probability necessarily
coincide.
b3) It suffices to observe that
e Bt 1fBt 0g
and then to apply b2).

c) By Fubini’s theorem and recalling the expression of the Laplace transform of
the Gaussian distributions,
hZ C1 i Z C1 Z C1
2
E ebuC Bu du D EŒebuC Bu du D e.bC 2 /u
du :
0 0 0
2
The expectation is therefore finite if and only if b < 2 . The integral is then
easily computed giving, in conclusion,
8
hZ C1 i < 1
2 if b < 2
2
E e buC Bu
du D bC 2
0 :C1 otherwise :
3.16 The sequence .n /n is increasing and therefore converges to some r.v. .

Obviously, as n , also . Let us prove that actually D .
Let us consider first the event f < C1g. On it of course n < C1 and Xn 2
@Dn . X being continuous, Xn ! X and d.X ; @D/ D limn!1 d.Xn ; @D/ D 0, so
that X 2 @D and . Conversely, on f D C1g there is nothing to prove.
The argument above also implies that Xn ! X as n ! 1.
Note that this argument only works for an open set D because otherwise the
condition X 2 @D does not imply . Actually the statement is not true if D is
closed (try to find a counterexample. . . ).
3.17 If Xs D Bs=2 , then X is also a Brownian motion by Proposition 3.2 so that,
if
X D inffuI Xu 62 Dg ;
we have X . As Bs D Xs2
D inffsI Bs 62 Dg D inffsI Bs 62 Dg D inffsI Xs2 62 Dg

1 1
D 2
inffuI Xu 62 Dg 2
:
3.18
a) By the Iterated Logarithm
p Law, with probability 1 there exist values of t such that
X1 .t/ .1 "/ 2t log log t (X1 is the first component of the Brownian motion
X). There exists therefore, with probability 1, a time t such that jXt j > 1 and so
< C1 a.s.
Let us denote by the law of X and let O be an orthogonal matrix; if Yt D
OXt , then (Exercise 3.9) Y is also a Brownian motion. Moreover, as jYt j D jXt j
for every t, is also the exit time of Y from S. Therefore the law of Y D OX
coincides with the law of X , i.e. with . Therefore the image law of through
O (that defines a transformation of the surface of the sphere, @S, into itself) is
still equal to .
This allows us to conclude the proof since, as indicated in the hint, the only
probability on @S with this property is the normalized .m 1/-dimensional
Lebesgue measure. Figures S.3 and S.4 show the positions of some simulated
exit points.
b) Let and A be Borel sets respectively of @S and of RC ; we must show that
P.X 2 ; 2 A/ D P.X 2 /P. 2 A/. Repeating the arguments developed in
a), we have, for every orthogonal matrix O,
P.X 2 ; 2 A/ D P.X 2 O; 2 A/ :
Therefore, for every fixed A 2 B.RC /, the measure A on @S defined as

A . / D P.X 2 ; 2 A/ is rotationally invariant. It is therefore of the form
A D c and obviously the constant c is determined by c .@S/ D P. 2 A/.
Therefore
P. 2 A/
P.X 2 ; 2 A/ D c . / D . / D P.X 2 /P. 2 A/ :
.@S/
...........•..•......•.•..•.•........•...•....
...•..•..•..•..• ••• ...•..•...
...•.....•
..•.•.• ...•...
.......•...• .•..•...
••..•...
..
•
....•.• •..•.•.•..
....•.
• •....
.•.•. •.•.•.
.. .•..
. . •.•..
..•.• •.•..
•
. .. •....
.•.•.•. ••...
...
.•.. ••....
.•...
•
•..
••...
• ••...
..
••.•.. •
.••..
.
•..•.. •
...•
•... .•..
•..•.. .•..•
... .•.•.
•..•.
•..... ...•..
•..•.. ..•.
•....... ....•
..•.•..
.•..•... ...•..•
.•
•..•..•....
••..•.•.•..•.... ...•..
• ....•..•.•..•
•..•.•.....•..•.•..••...••.............•...•.•.•....•.•...•
Fig. S.3 The exit positions of 200 simulated paths of a two-dimensional Brownian motion from
the unit ball. The exit distribution appears to be uniform on the boundary
.....
.............•.....••• ......•....•...•...•...•.......
...•....•..• •••..•..•...
..•...•..• ......
. ..
•.....•.• ••.....
. •...•..
..... ••..•...
•
...•..•
. •....
.... •....
.
•.•. •.•.•.
•
.
.. •.•.•..
•
.... •.•...
... ••..
••...
.•.. ••...
....
.•.. • •..
••.
... •...
... •
.•.•.
•.... •
.•.
.
•
•..
... .•.•.
..•
...
... ..•
.•
•.•... ....•.•
•
..... ..•.••
..•... ..•.•.•
•.....•... ....•.•
..•...
.......
••......... ...
•.. ..•.•...•.
•..•......••.......... .......•......•.••...•.•.•..• . •
••••......•.•. •
Fig. S.4 The exit positions of 200 simulated paths from the unit ball for a two-dimensional
Brownian motion starting at . 12 ; 0/ (denoted by a black small circle). Of course the exit distribution
does no longer appears to be uniform and seems to be more concentrated on the part of the
boundary that is closer to the starting position; wait until Chap. 10 in order to determine this
distribution
3.19
a) Immediate as
ke f k1 D sup ef .s/ D esup0s1 f .s/ :

0s1
b1) The clever reader has certainly sensed the imminent application of the scaling
properties of Brownian
p motion. Replacing in the left-hand side the Brownian
motion B with s 7! tBs=t we have, with the substitution u D s=t,
Z t Z t p Z 1 p
L
e ds
Bs
e tBs=t
ds D t e tBu
du :
0 0 0
b2) The previous relation gives

Z t Z 1 p
L
log eBs ds log t e tBs
ds
0 0
and by the property of the Lp norms mentioned in the hint we have, in

distribution,
Z h Z 1 p 1=pt i
1 t
1
lim p log eBs ds D lim p log t C lim log e tBs
ds
t!C1 t 0 t!C1 t t!C1 0
D log ke k1 D sup B.s/ :

B
0s1
p
b3) Taking the log and dividing by t
Z t 1 Z t
lim P eBs ds1 D lim P p log eBs ds 0 DP sup Bs 0 D 0 :
t!C1 0 t!C1 t 0 s1
Similarly, by the Reflection Principle,

Z t p 1 Z t
:77337 t
lim P e ds e
Bs
D lim P p log eBs ds :77337
t!C1 0 t!C1 t 0
1
D P sup Bs :77337 D 1 2P.B1 > :77337/ D
0s1 2
3.20
a) By the reflection principle the partition function Fa of a is, for t > 0,

Fa .t/ D P.a t/ D P sup Bs > a D 2P.Bt > a/ D 2P.B1 > at1=2 /
0st
Z C1
2 2 =2
D p ex dx :
2 at1=2
Fa is differentiable, so that a has density, for t > 0,

r
2 a2 =2t d a a 2
fa .t/ D Fa0 .t/ D e p Dp ea =2t :
dt t 2 t 3=2
We have
Z Z
p C1 p C1
a 2
E. a / D tfa .t/ dt D p ea =2t dt D C1 ;
0 0 2 t
as the integrand behaves as 1t towards infinity. A fortiori E.a / D C1. Again by

the reflection principle, as t ! C1,
P.a > t/ D 1 2P.Bt > a/ D 1 2P.B1 > at1=2 /

D 1 P.B1 > at1=2 / P.B1 < at1=2 /
Z at1=2
1 2 1 2a
Dp ex =2 dx p 2at1=2 D p
2 at1=2 2 2t
b) The events fTj tg, j D 1; : : : ; N are independent, each of them having
probability
1
P.Tj t/ D 2P.Bt 1/ D 2P Z p ;
t
where Z N.0; 1/. Hence
P.Tj > t for at least one index j; j D 1; : : : ; N/

\N 1 N
D 1P fTj tg D 1 P.Tj t/N D 1 2P Z p :
jD1
t

For N D 10000 and t D 108 we have 2P Z 104 D 0:9999202 and
10000
1 2P Z 104 D 1 :55 D 45%
whereas similar computations for t D 1010 give

10000
1 2P Z 105 D :076 D 7:6% :
Therefore the program has the drawback that it can remain stuck on a single path
for a very very long time. We shall see some remedies to this problem later on
(see Example 12.4).
c) By Theorem 3.3, e Bt D Ba Ct Ba is a Brownian motion independent of Fa . If
we denote by e a the passage time at a of e
B, the two r.v.’s a and e
a have the same
law and are independent. Moreover, it is clear that 2a D a C e a . By recurrence
therefore the sum of n independent r.v.’s X1 ; : : : ; Xn each having a law equal to
that of a has the same law as na . Therefore the density of n12 .X1 C C Xn /
is
na a2 n2 a 2
n2 fna .n2 t/ D n2 p exp 2 D p ea =2t D fa .t/ :
2 .n2 t/3=2 2tn 2 t3=2
3.21
a) We have
p 1
sup Bs D sup But D t sup p But :
0st u1 u1 t
As . p1 t But /u is itself a Brownian motion, this means that

p
sup Bs t sup Bu
0st u1
(recall that means equality in law). Of course this implies (3.24).

b) From Exercise 1.3 and thanks to the reflection principle
h i Z C1 Z C1
E sup Bs D P sup Bs > x dx D 2 P.Bt > x/ dx
0st 0 0st 0
Z C1 Z C1 2 Z C1 2 Z
2 2 z
e 2t dz D p e 2t dz
z z
Dp dx dx
2t 0 x 2t 0 0
Z C1 ˇ r
2 z2 2 z2 ˇC1 2t
Dp
ze 2t dz D p
te 2t ˇ D
2t 0 2t 0
3.22
1
a1) The vector jzj z has modulus equal to 1, therefore X is a Brownian motion thanks
to Remark 3.1. We have
D infftI hz; Bt i kg D infftI jzjXt kg D infftI Xt jzj g

k
:
a2) As coincides with the passage time of X at a D jzj k

it has infinite expectation,
as seen in Exercise 3.20, and its density is given by (3.23).
b1) X is a centered Gaussian process starting at 0 whatever the value of ˛. In order
to check whether it is a Brownian motion we have to determine for which values
of ˛ > 0 condition 3) of Proposition 3.1 is satisfied. We have, for s t,
EŒXt Xs D EŒ˛.B1 .t/ C B2 .t//˛.B1 .s/ C B2 .s//

2

D ˛ EŒB1 .t/B1 .s/ C EŒB2 .t/B2 .s/ C EŒB1 .t/B2 .s/ C EŒB1 .s/B2 .t/

D ˛ 2 s C s C s C s D 2˛ 2 .1 C /s
hence, if > 1, we have the condition

1
˛Dp (S.9)
2.1 C /
If D 1 such an ˛ does not exist. Actually, if D 1 we have B1 .t/CB2 .t/ D

0 a.s.
b2) A repetition of the argument of a2): coincides with the passage time of the
Brownian motion X with the choice of ˛ as in (S.9) at a D .2 C 2/1=2 so that
has the density (3.23) with a D .2 C 2/1=2 and has infinite expectation. By
the reflection principle (Proposition 3.4)
P. 1/ D 2P.X1 .2 C 2/1=2 / ;
which is maximum for D 1.
3.23
a) We know that t 7! hw; Bt i is a Brownian motion if w is a vector having modulus
1
equal to 1 (Example 3.1). Hence v D jzj satisfies the requested condition.
b1) Of course
D infftI B1 .t/ C B2 .t/ D 1g D infftI hz; Bt i D 1g :
b2) We have jzj2 D 2, so that, thanks to a), Wt WD p12 hz; Bt i is a Brownian motion.
Hence
p
D infftI B1 .t/ C B2 .t/ D 1g D infftI 2 Wt D 1g D infftI Wt D p12 g
so that has the same distribution as the passage time of W at a D p1 and

2
therefore has density
a 2 =2t 1
f .t/ D ea D e1=4t
.2/1=2 t3=2 2 1=2 t3=2
and, by the reflection principle and using tables of the partition function ˚ of
an N.0; 1/-distributed r.v.,
1
P. 1/ D 2P W1 p D 0:48 :
2
c1) We have 1 D infftI 2B1 .t/ C B2 .t/ D 1g. We can repeat p the arguments of b) for
2B
p1 .t/ C B 2 .t/ D hz; B.t/i with z D .2; 1/; here jzj D 5 and 2B1 .t/ C B2 .t/ D
5 W1 .t/, where W1 is a Brownian motion. Therefore
p
1 D infftI 2B1 .t/ C B2 .t/ D 1g D infftI 5 W1 .t/ D 1g D infftI W1 .t/ D p1 g
5
so that 1 has density, for t 0,

1
f1 .t/ D e1=10t :
.10/1=2 t3=2
The same argument, for the new Brownian motion W2 .t/ D p1 B1 .t/
5
p2 B2 .t/, gives
5
p
5
2 D infftI 12 B1 .t/ B2 .t/ D 1g D infftI 2
W2 .t/ D 1g D infftI W2 .t/ D p2 g
5
with density
2
f2 .t/ D e1=5t :
.10/1=2 t3=2
c2) In order to prove that 1 and 2 are independent, it suffices to show that the two
Brownian motions W1 ; W2 are independent. We have
2 1 1 2
W1 .t/ D p B1 .t/ C p B2 .t/; W2 .s/ D p B1 .s/ p B2 .s/ :
5 5 5 5
Therefore, assuming s t,
2 2
Cov.W1 .t/; W2 .s// D EŒW1 .t/W2 .s/ D p EŒB1 .t/B1 .s/ p EŒB2 .t/B2 .s/ D 0:
5 5
Hence, as the r.v.’s W1 .t/; W2 .s/ are jointly Gaussian, they are independent for
every t; s. Alternatively, just observe that the two-dimensional process .W1 ; W2 /
is obtained from .B1 ; B2 / through the linear transformation associated to the
matrix
2 1
!
p p
5 5
p1 p25
5
which is orthogonal. Hence by Exercise 3.9 .W1 ; W2 / is a two-dimensional

Brownian motion which implies the independence of W1 and W2 .
c3) The event f 1g is the union of the events f1 1g and f2 1g, which are
independent. Therefore
P. 1/ D P.1 1/ C P.2 1/ P.1 1/P.2 1/ :
Now, by the reflection principle and again denoting by ˚ the partition function
of the N.0; 1/ distribution,

P.1 1/ D 2P W1 p1 D 2 1 ˚ p1 D 2.1 0:67// D 0:65 ;
5 5
P.2 1/ D 2P W2 p2 D 2 1 ˚ p25 D 2.1 0:81// D 0:37 ;
5
which gives
P. 1/ D 0:65 C 0:37 0:65 0:37 D 0:78 :
c4) The important thing is to observe that the event f1 2 g coincides with the
event “the pair .1 ; 2 / takes its values above the diagonal”, i.e., as 1 and 2 are
independent,
Z C1 Z t
P.1 2 / D f2 .t/ dt f1 .s/ ds : (S.10)
0 0
Let us compute this integral. To simplify the notations let us set a1 D p1 ,

5
a2 D p2 . By the reflection principle
5
Z Z C1
t
2 x2
f1 .s/ ds D P.1 t/ D 2P.W1 .t/ a1 / D p e 2t dx
0 2t a1
and substituting into (S.10)

Z C1 Z C1
a2 2 2 2
P.1 2 / D dx ea2 =2t p ex =2t dt
a1 .2/1=2 t3=2
0 2t
Z Z C1 Z ˇtDC1
a2 C1 1 1 .a22 Cx2 / a2 C1 2 2t1 .a22 Cx2 / ˇ
D dx e 2t dtD dx e ˇ
a1 0 t2 a1 a22 C x2 tD0
Z Z C1
a2 C1 2 2 1 2 x ˇˇC1
D dx D dx D arctan ˇ
a1 a22 C x2 a2 a1 1 C . a2 /2
x
a2 a1
2 a1 2 1
D arctan D 1 arctan ' 0:7 :
2 a2 2
• Note that the independence of 1 and 2 is a consequence of the orthogonality of

the two straight lines 2x C 1 D 1 and 12 x y D 1.
4.1 Statement a) looks intuitive: adding the information D, which is independent

of X and of G , should not provide any additional information useful for the
prediction of X. But the formulation of the exercise has certainly led the reader
to surmise that things are not really this way. Let us therefore start proving b); we
shall then look for a counterexample showing that the answer to a) is negative.
b) The events G \ D, G 2 G ; D 2 D form a class which is stable with respect to
finite intersections, generating G _ D and containing ˝. Let us prove that
EŒE.X jG /1G\D D E.X1G\D /
for every G 2 G , D 2 D. As the r.v. EŒX jG is obviously G _ D-measurable, by

Remark 4.2 this will prove that E.X jG / D E.X jG _ D/. As D is independent of
.X/ _ G and therefore also of G ,
EŒE.X jG /1G\D D EŒE.X1G jG /1D D EŒE.X1G jG /EŒ1D

D EŒX1G EŒ1D DE.X1G 1D / D E.X1G\D / ;
"
where " denotes the place where independence of D and .X/ _ G is used.
a) The counterexample is based on the fact that it is possible to construct three
r.v.’s X; Y; Z such that the pairs .X; Y/, .Y; Z/ and .Z; X/ are each formed by
independent r.v.’s but such that X; Y; Z are not independent globally.
An example is given by ˝ D f1; 2; 3; 4g, with the uniform probability P.k/ D
1
4
, k D 1; : : : ; 4, and the -algebra F of all subsets of ˝. Let X D 1f1;2g ,
Y D 1f2;4g and Z D 1f3;4g . Then we have
1
P.X D 1; Y D 1/ D P.f1; 2g \ f2; 4g/ D P.2/ D D P.X D 1/P.Y D 1/ :
4
Similarly it can be shown that P.X D i; Y D j/ D P.X D i/P.Y D j/ for

every possible value of i; j 2 f0; 1g, which implies that X and Y are independent.
Similarly we prove that .X; Z/ and .Y; Z/ are also pairs of independent r.v.’s. Let
G D .Y/ and D D .Z/. Then the -algebra G _ D contains the events
f1g D fY D 0; Z D 0g; f2g D fY D 1; Z D 0g ;

f3g D fY D 0; Z D 1g; f4g D fY D 1; Z D 1g
and therefore G _D D F . Hence EŒX jG _D D X, whereas EŒX jG D E.X/ D

1
2
a.s. as X and G are independent.
4.2 Let us first assume that Y is integrable. Then if Y were G -measurable we

would have
Y D EŒY jG a.s.
But on the other hand, as Y is independent of G ,
EŒY jG D EŒY a.s.

and these two relations can both be true only if Y D EŒY a.s. If Y is not integrable
let us approximate it with integrable r.v.’s. If Yn D Y _ .n/ ^ n then Yn is still
independent of G and G -measurable. Moreover, as jYn j n, Yn is integrable.
Therefore by the first part of the proof Yn is necessarily a.s. constant and, taking
the limit as n ! 1, the same must hold for Y.
4.3 Recall (p. 88) that f .x/ D EŒZ jX D x means that f .X/ D EŒZ j.X/ a.s., i.e.
that EŒ f .X/ .X/ D EŒZ .X/ for every bounded measurable function W E ! R.
Let A 2 E . Then we have fX 2 Ag 2 .X/ and
Z
Q .A/ D Q.X 2 A/ D EŒZ1fX2Ag D EŒ f .X/1fX2Ag D f .y/ dP .y/ ;
A
dQ
which proves simultaneously that Q P and that dP D f.
4.4
a) If G D fE.Z jG / D 0g, then G 2 G and therefore, as E.Z jG /1G D 0,
E.Z1G / D EŒE.Z1G jG / D EŒE.Z jG /1G D 0 : (S.11)
As Z 0, this implies Z D 0 a.s. on G, i.e. fZ D 0g fE.Z jG / D 0g a.s.

Taking the complements, fZ > 0g fE.Z jG / > 0g and 1fZ>0g 1fE.Z jG />0g .
Hence if Y 0 we have a.s.
EŒZY jG D EŒZ1fZ>0g Y jG EŒZ1fE.Z jG />0g Y jG D 1fE.Z jG />0g EŒZY jG :
As the opposite inequality is obvious, (4.28) is proved for Y 0. We obtain the

case of a general r.v. Y splitting into positive and negative parts.
b) We have, with a repetition of the argument of (S.11),
Q.EŒZ jG D 0/ D EŒZ1fEŒZ jG D0g D 0 ;
hence EŒZ jG > 0 Q-a.s. Moreover, as the right-hand side of (4.29) is clearly G -
measurable, in order to verify (4.29) we must prove that, for every G -measurable
bounded r.v. W,
h E.YZ jG / i
EQ W D EQ ŒYW :
E.Z jG /
But
h E.YZ jG / i h E.YZ jG / i
EQ W D E ZW
E.Z jG / E.Z jG /
and, as inside the expectation on the right-hand side Z is the only r.v. that is not
G -measurable,
h E.YZ jG / i
D E E.Z jG /W D EŒE.YZ jG /W D EŒYZW D EQ ŒYW
E.Z jG /
and (4.29) is satisfied.
• In the solution of Exercise 4.4 we left in the background a delicate point which
deserves some attention. Always remember that a conditional expectation (with
respect to a probability P) is not a r.v., but a family of r.v.’s, only differing
from each other by P-negligible events. Therefore the quantity EŒZ jG must be
considered with care when arguing with respect to a probability Q different from
P, as it might happen that a P-negligible event is not Q-negligible. In the case of
this exercise there are no difficulties as P Q, so that negligible events for P are
also negligible for Q.
4.5 Let us prove that every D-measurable real r.v. W is independent of X. The
characteristic function of the pair Z D .X; W/, computed at
D . ; t/, 2 Rm ; t 2
R, is equal to
EŒeih
;Zi D EŒeih ;Xi eitW D EŒeitW E.eih ;Xi jD/ D EŒeitW EŒeih ;Xi
so that X and W are independent by criterion 7 of Sect. 1.6. This entails the
independence of X and D.
4.6
a) Thanks to Example 4.5 and particularly (4.12), the requested characteristic
function is
Z C1 t
2 2
EŒeih
;B i D e 2 j
j e t dt D 1
D
0 C 2 j
j 2 2 C j
j2
b1) The characteristic function of X is

Z C1

X .
/ D ei
x ejxj dx :
2 1
jxj
Now x 7! sin.
x/ e is an odd function so that the imaginary part in the
integral above vanishes. Conversely, x 7! cos.
x/ ejxj is an even function so
that
Z C1 Z C1
x
X .
/ D cos.
x/ e dx D < ei
x ex dx
0 0
2
D< D
i
2 C
2
b2) Comparing the characteristic function of B computed in a) with that of a

p in b2) we see that if m D 1 then B has a Laplace
Laplace distribution computed
distribution with parameter 2 .
4.7 There are two possible methods: the best is the second one below. . .
a) First method. Let us check directly that X has the same finite-dimensional
distributions as a Brownian motion. Let t1 < t2 < < tm ,
D .
1 ; : : : ;
m / 2
Rm .
Thanks to the freezing lemma,
EŒei
1 Xt1 CCi
m Xtm D EŒei
1 .BCt1 B /CCi
m .BCtm B /

D E EŒei
1 .BCt1 B /CCi
m .BCtm B / j./ D EŒ˚./ ;
where
˚.s/ D EŒei
1 .BsCt1 Bs /CCi
m .BsCtm Bs / :
As we know that the increments .BsCt Bs /t form a Brownian motion (Proposi-

tion 3.2), ˚ does not depend on s and we have that
EŒei
1 Xt1 CCi
m Xtm D EŒei
1 Bt1 CCi
m Btm :
The process X has the same finite-dimensional distributions as B so that it is itself

a natural Brownian motion.
b) Second method. Let F e s D .Fs ; .// be the filtration that is obtained by
adding to .Ft /t the -algebra generated by . Thanks to Exercise 3.5, B is also
a Brownian motion with respect to .F e t /t and now is a stopping time for this
larger filtration as f tg 2 ./ and therefore f tg 2 F e t for every t. Then
the stopping theorem, Theorem 3.3, allows us to conclude that X is an .F e t /t -
Brownian motion.
4.8
a) Recall that the -algebras G1 D .B1 .u/; u 0/ and G2 D .B2 .u/; u 0/ are
independent (Remark 3.2 b)) and note that is G2 -measurable.
b) Recalling Remark 4.5 and particularly (4.11), the density of B1 ./ is given by
Z C1
1 2
g.x/ D p ex =2t d.t/ ;
0 2t
where denotes the law of . From Exercise 3.20 we know that has density
a 2
f .t/ D p ea =2t
2 t 3=2
for t > 0 hence the density of B1 ./ is

Z C1 Z C1
1 a a2 =2t x2 =2t a 1 .a2 Cx2 /=2t
g.x/ D p p e e dt D e dt :
0 2t 2 t3=2 2 0 t2
1
With the change of variable s D t
we obtain
Z C1
a 1 2 Cx2 /s a
g.x/ D e 2 .a ds D ,
2 0 .a2 C x2 /
which is a Cauchy law.
4.9 The idea is always to split Bu into the sum of Bs and of the increment Bu Bs .
As B2u D .Bu Bs C Bs /2 D .Bu Bs /2 C B2s C 2Bs .Bu Bs /, we have
Z t Z t Z t
B2u du D .t s/Bs C .Bu Bs /2 du C 2Bs .Bu Bs / du :
s s s
Bs is Fs -measurable whereas (recall Remark 4.5)

Z t ˇ
Z t ˇ
Z t
1
E .Bu Bs / du ˇ Fs D 2
E .Bu Bs /2 ˇ Fs du D .u s/ du D .t s/2 ;
s s s 2
the r.v.’s Bu Bs being independent of Fs . By the same argument

Z t Z t
ˇ ˇ
E 2Bs .Bu Bs / du ˇ Fs D 2Bs E Bu Bs ˇ Fs du D 0
s s
so that finally
Z t ˇ 1
E B2u du ˇ Fs D .t s/Bs C .t s/2 :
s 2
As this quantity is already .Bs /-measurable,

Z t h Z t ˇ ˇ i 1
E B2u dujBs DE E B2u du ˇ Fs ˇ Bs D .t s/Bs C .t s/2 :
s s 2
The meaning of the equality between these two conditional expectations will
become clearer in the light of the Markov property in Chap. 6.
4.10 Let us denote by Y , Z , respectively, the laws of Y and Z and by y the law
of .y/ C Z. We must prove that for every pair of bounded measurable functions
f W E ! R and g W G ! R we have
Z Z
EŒ f .X/g.Y/ D g.y/ dY .y/ f .x/ dy .x/ :
G E
But we have
Z Z
f .x/ dy .x/ D f ..y/ C z/ dZ .z/
E E
and
Z Z
EŒ f .X/g.Y/ D EŒ f ..Y/ C Z/g.Y/ D g.y/ dY .y/ f ..y/ C z/ dZ .z/
G E
Z Z
D g.y/ dY .y/ f .x/ dy .x/ :
G E
4.11
a) We must find a function of the observation, Y, that is a good approximation
of X. We know (see Remark 4.3) that the r.v. .Y/ minimizing the squared
L2 distance EŒ..Y/ X/2 is the conditional expectation .Y/ D E.X jY/.
Therefore, if we measure the quality of the approximation of X by .Y/ in the
L2 norm, the best approximation of X with a function of Y is E.X jY/. Let us go
back to formulas (4.23) and (4.24) concerning the mean and the variance of the
conditional laws of a Gaussian r.v.’s: here
mX D mY D 0; Cov.X; Y/ D Cov.X C W; X/ D 1; Var.Y/ D 1 C 2 ;
so that
Cov.X; Y/ y
E.X jY D y/ D mX C .y mY / D
Var.Y/ 1 C 2
The required best approximation is
Y

1 C 2
Note that the variance of the conditional distribution of X given Y D y is
Cov.X; Y/2 1 2
Var.X/ D1 D (S.12)
Var.Y/ 1 C 2 1 C 2
b) The computation follows the same line of reasoning as in a) but now Y is two-
dimensional and we shall use the more complicated relations (4.21) and (4.22).
If Y D .Y1 ; Y2 /, y D .y1 ; y2 /, we have mX D 0, mY D 0, CX D 1. As Var.Y1 / D

Var.Y2 / D 1 C 2 and Cov.Y1 ; Y2 / D Cov.X C W1 ; X C W2 / D 1,

1 C 2 1
CY D ; CX;Y D 1 1 :
1 1 C 2
A quick computation gives

1 1 C 2 1
CY1 D :
.1 C 2 /2 1 1 1 C 2
Therefore the conditional law of X given Y D y has mean

1 .1 C 2 /y1 y2 y1 C y2
CX;Y CY1 y D 11 2 D
2
2 C 4 .1 C /y2 y1 2 C 2
The best approximation of X given Y D y now is
Y1 C Y2

2 C 2
The variance of the conditional law is now

1 1 C 2 1 1
1 CX;Y CY1 CX;Y
D1 2
11
2 C 4 1 1 C 2 1
2
1 2 2
D1 2 1 1 D 1 D
2 C 4 2 2 C 2 2 C 2
,
which is smaller than the value of the conditional variance given a single
observation, as computed in (S.12).
4.12
a) We apply formulas (4.21) and (4.22) to X D .Bt1 ; : : : ; Btm /, Y D B1 . We have
CY D 1 whereas, as Cov.Bti ; B1 / D ti ^ 1 D ti ,
0 1
t1
B :: C
CX;Y D@:A :
tm
Therefore as
0 1
t1
B :: C
CX;Y CY1 CX;Y

D @ : A t1 : : : tm
tm
the matrix CX;Y CY1 CX;Y

has ti tj as its .i; j/-th entry. As CX is the matrix .ti ^ tj /i;j ,
the conditional law is Gaussian with covariance matrix CX CX;Y CY1 CX;Y D
.ti ^ tj ti tj /i;j , i.e.
0 1
t1 .1 t1 / t1 .1 t2 / : : : t1 .1 tm /
B t1 .1 t2 / t2 .1 t2 / : : : t2 .1 tm / C
B C
B :: C
@ : A
t1 .1 tm / t2 .1 tm / : : : tm .1 tm /
and mean (here EŒX D 0; EŒY D 0)

01
t1 y
B C
CX;Y CY1 y D @ ::: A :
tm y
b) Now Y D .B1 ; Bv / and

0 1
t1 t1
11 1 v 1 B :: :: C
CY D ; CY1 D ; CX;Y D@: :A
1v v 1 1 1
tm tm
so that
0 1 0 1
t1 t1 t1 0
1 B :: :: C v 1 D B :: :: C
CX;Y CY1 DD @: : A 1 1 @: :A
v1
tm tm tm 0
and
0 1
t1 0
B :: :: C t1 : : : tm
CX;Y CY1 CX;Y

D @ : :A ;
t1 : : : tm
tm 0
which is still the matrix with ti tj entries. Hence the covariance matrix of the
conditional law is the same as in a). The same holds for the mean which is
equal to
0 1
t1 y
1 y B :: C
CX;Y CY D@ : A :
v
tm y
In conclusion, to be repositioned the conditional law neither depends on x nor on

v and coincides with the conditional law given B1 D y. From the point of view
of intuition this is not surprising as the addition of the information Bv D x means
to add the information Bv B1 D x y. As the increments of the Brownian
motion are independent of the past it would have been reasonable to expect that
the addition of this new information does not change the conditional law. This
fact may be obtained directly using Exercise 4.1 b).
4.13 Thanks to Exercise 3.11 the joint law of the two r.v.’s is Gaussian. In order
to identify it, we just need to compute its mean and covariance matrix. The two r.v.’s
are obviously centered. Let us compute the variance of the second one:
h Z 1 2 i hZ 1 Z 1 i hZ 1 Z 1 i
E Bs ds DE Bs ds Bt dt D E Bs Bt ds dt
0 0 0 0 0
Z 1 Z 1 Z 1 Z t Z 1 Z 1
D s ^ t ds dt D dt s ds C dt t ds D I1 C I2 :
0 0 0 0 0 t
We have easily
Z 1 2
t 1
I1 D dt D
0 2 6
and similarly I2 D 16 ; therefore the variance is 13 . The variance of B1 is equal to 1,

of course, whereas for the covariance we find
h Z 1 i Z 1 Z 1
1
E B1 Bs ds D E.B1 Bs / ds D s ds D
0 0 0 2
R1
The covariance matrix of B1 and 0 Bs ds is therefore
1

1 2
1 1 :
2 3
R1
The best L2 estimate of the value of B1 given 0 Bt dt D x is the conditional mean,
which is easily computed using formula (4.23) and is equal to 32 x.
4.14
a) We have Cov.; Ys / D Cov.; s/ C Cov.; Bs / D s2 ( and Bs are
independent). By the same argument
Cov.Ys ; Yt / D Cov.s; t/ C Cov.Bs ; Bt / D st2 C .s ^ t/ 2 :
b) As is independent of B, .; .Bt /t0 / is a Gaussian family. Therefore for

every t1 ; : : : ; tm the vector .Yt1 ; : : : ; Ytm / is Gaussian, being a linear function of
.; Bt1 ; : : : ; Btm /.
c) must satisfy the relation Cov. Yt ; Ys / D 0 for every s t. We have
Cov. Yt ; Ys / D Cov.; Ys / Cov. Yt ; Ys / D s2 .st2 C s 2 /

D s.2 .t2 C 2 //
and therefore
2
D
2 C t2
With this choice of the r.v. Z D Yt is not correlated with each of the r.v.’s
Ys ; s t. As these generate the -algebra Gt and .; Yt ; t 0/ is a Gaussian
family, by Remark 1.2 Z is independent of Gt .
d) As Yt is Gt -measurable whereas Z D Yt is independent of Gt and EŒYt D t,
EŒjGt D EŒ Yt C Z jGt D Yt C EŒZ D Yt C EŒ Yt

2 Yt t2 2 Yt C 2
D Yt C .1 t/ D 2 C 1 D
C t2 2 C t2 2 C t2
We have
2 Yt C 2 2 Yt
lim EŒjGt D lim 2 2
D lim 2
t!C1 t!C1 C t t!C1 C t2
2
D lim .t C Bt / D a.s.
t!C1 2 C t2
as limt!C1 1t Bt D 0 a.s. by the Iterated Logarithm Law.
4.15
a) If t1 ; : : : ; tn 2 RC , then .Xt1 ; : : : ; Xtn / is Gaussian, being a linear function of
.Bt1 ; : : : ; Btn ; B1 /. Moreover, if t 1,
EŒXt B1 D EŒ.Bt tB1 /B1 D t ^ 1 t D 0 :

The two r.v.’s Xt and B1 , being jointly Gaussian and uncorrelated, are independent
for every t 1. Xt is centered and, if s t,
EŒXt Xs D EŒ.Bt tB1 /.Bs sB1 / D s st st C st D s.1 t/ : (S.13)

1t
b) As Xs and Xt 1s Xs are jointly Gaussian, we need only to show that they are
not correlated. We have, recalling that s t,
h 1t i 1t
E Xt Xs Xs D s.1 t/ s.1 s/ D 0 :
1s 1s
c) We have
h 1t 1 t ˇˇ i 1t
EŒXt jXs D E Xt Xs C Xs Xs D Xs (S.14)
1s 1s 1s
1t 1t
as, thanks to b), Xt 1s Xs is independent of Xs and centered whereas 1s Xs
is already .Xs /-measurable. The relation EŒXt jXs D EŒXt jGs follows if we
1t
show that Xt 1s Xs is independent of Gs and then repeat the same argument
1t
as in (S.14). In order to obtain this it is sufficient to show that Xt 1s Xs is
independent of Xu for every u s. This is true as, if u s,
h 1t i 1t
E Xt Xs Xu D u.1 t/ u.1 s/ D 0 :
1s 1s
d) We have, for s t 1,
EŒXt jFs D EŒBt tB1 jFs D Bs tBs D .1 t/Bs :
This result is different from

1t
EŒXt jGs D Xs
1s
1t
obtained in c), as it is easy to see that Var. 1s Xs / D .1t/2 1s
s
whereas Var..1
2
t/Bs / D .1 t/ s. Therefore the two -algebras Fs and Gs are different.
e) Let 0 t1 < < tm 1. The conditional law of the vector .Bt1 ; : : : ; Btm /
given B1 D 0 can be computed as explained in Sect. 4.4. It has actually already
been computed in Exercise 4.12 b). It is Gaussian, centered, and with covariance
matrix whose .i; j/-th entry is ti ^ tj ti tj , the same as the covariance matrix of
the Brownian bridge, as it can be computed from (S.13).
5.1 For every t > s, we have E.Xt jFs / Xs a.s. Therefore the r.v. U D Xs
E.Xt jFs / is positive a.s.; but it has zero mean, as
EŒU D EŒXs EŒE.Xt jFs / D E.Xs / E.Xt / D 0 :

We deduce that U D 0 a.s. (a positive r.v. having mean equal to 0 is D 0 a.s.).

Therefore E.Xt jFs / D Xs a.s. for every t > s.
5.2
a) We know that P.Xa;b D b/ D a
aCb , hence
lim P.Xa;b D b/ D 1 ;

a!C1
which was to be expected: if the left endpoint of the interval a; bŒ is far from
the origin, it is more likely for the exit to take place at b.
b) The event fXa;b D bg is contained in fb < C1g, hence, for every a > 0,
P.b < C1/ P.Xa;b D b/
so that, thanks to a), P.b < C1/ D 1.
5.3
a) Thanks to the law of large numbers we have a.s.
1 1
n Xn D n .Y1 C C Yn / ! EŒY1 D p q < 0 :
n!1
Therefore a.s. there exists an n0 > 0 such that 1n Xn 12 .p q/ < 0 for n n0 ,

which implies Xn n2 .p q/; hence Xn ! 1 a.s.
b1) Let Fn D .X1 ; : : : ; Xn / D .Y1 ; : : : ; Yn /. We have, YnC1 being independent
of Fn ,
h q XnC1ˇ i h Xn CYnC1ˇ i q Xn h q YnC1ˇ i
EŒZnC1 j Fn D E ˇFn D E q ˇ Fn D E ˇ Fn
p p p p
q Xn h q YnC1 i q Xn q p
D E D P.YnC1 D 1/ C P.YnC1 D 1/
p p p p q
q X n q p
q Xn q X n
D pC q D .p C q/ D D Zn :
p p q p p
Note that this argument proves that the product of independent r.v.’s having
mean equal to 1 always gives rise to a martingale with respect to their natural
filtration. Here we are dealing with an instance of this case.
b2) As, thanks to a), limn!1 Xn D 1 a.s. and qp > 1, we have limn!1 Zn D 0.
c) As n ^ is a bounded stopping time, by the stopping theorem EŒZn^ D
EŒZ1 D 1. By a) < C1, therefore limn!1 Zn^ D Z a.s. As qp > 1 and
a Xn^ b, we have . qp /a Zn^ . qp /b and we can apply Lebesgue’s
theorem and obtain that EŒZ D limn!1 EŒZn^ D 1.
d) As X can only take the values a; b,

h q X q b q a
1 D EŒZ D E D P.X D b/ C P.X D a/
p p p
q b q a
D P.X D b/ C .1 P.X D b// :
p p
Actually, as < C1, P.X D a/ D 1 P.X D b/. Hence
q a q b q a
1 D P.X D b/ ;
p p p
i.e.
1 . qp /a
P.X D b/ D
. qp /b . qp /a
5.4
a) Zk is Fk1 -measurable whereas Xk is independent of Fk , therefore Xk and
Zk are independent. Thus Zk2 Xk2 is integrable, being the product of integrable
independent r.v.’s and Yn is square integrable for every n (beware of a possible
confusion: “.Yn /n square integrable” means Yn 2 L2 for every n, “.Yn /n bounded
in L2 ” means supn>0 E.Yn2 / < C1). Moreover,
EŒYnC1 jFn D EŒYn C ZnC1 XnC1 jFn D Yn C ZnC1 EŒXnC1 jFn

D Yn C ZnC1 EŒXnC1 D Yn :
„ ƒ‚ …
D0
b) As Zk and Xk are independent, EŒZk Xk D EŒZk EŒXk D 0 and therefore EŒYk D

0. Moreover, we have
h X
n 2 i hX
n i X
n
EŒYk2 D E Zk Xk DE Zk Xk Zh Xh D EŒZk Xk Zh Xh
kD1 k;hD1 k;hD1
but in the previous sum all the terms with h 6D k vanish: let us assume k > h,
then Zk Xh Zh is Fk1 -measurable whereas Xk is independent of Fk1 . Therefore
EŒZk Xk Zh Xh D EŒXk EŒZk Zh Xh ;

„ƒ‚…
D0
hence, using again the independence of Zk and Xk ,
X
n X
n X
n
EŒYk2 D EŒZk2 Xk2 D EŒZk2 EŒXk2 D 2
EŒZk2 : (S.15)
kD1 kD1 kD1
The compensator .An /n of the submartingale .Mn2 /n is given by the relations A0 D

2
0, AnC1 D An C EŒMnC1 Mn2 jFn . Now
2
EŒMnC1 Mn2 jFn D EŒ.Mn C XnC1 /2 Mn2 jFn
D EŒMn2 C 2Mn XnC1 C XnC1
2
Mn2 jFn D EŒ2Mn XnC1 C XnC1
2
jFn
2
D 2Mn EŒXnC1 jFn C EŒXnC1 jFn :
As XnC1 is independent of Fn ,
EŒXnC1 jFn D EŒXnC1 D 0 a.s.

2 (S.16)
2
EŒXnC1 jFn D EŒXnC1 D 2 a.s.
and therefore AnC1 D An C 2 , which with the condition A0 D 0 gives An D n 2 .

In order to compute the compensator of .Yn2 /n just repeat the same argument.
If we denote it by .Bn /n , then
2
EŒYnC1 Yn2 jFn D EŒ.Yn C ZnC1 XnC1 /2 Yn2 jFn
2 2
D EŒ2Yn ZnC1 XnC1 C ZnC1 XnC1 jFn
2 2 2 2
D 2Yn ZnC1 EŒXnC1 jFn C EŒZnC1 XnC1 jFn D ZnC1 EŒXnC1 jFn D 2 ZnC1
2
:
2
From the relations B0 D 0 and BnC1 D Bn C EŒYnC1 Yn2 jFn we find
X
n
Bn D 2 Zk2 :
kD1
c) By (S.15)
Xn
1
EŒYn2 D 2 2

kD1
k
As the series on the right-hand side is convergent, the martingale .Yn /n is

bounded in L2 . It is therefore convergent a.s. and in L2 and therefore uniformly
integrable.
5.5
a) We have
EŒYk D 1 2k C 0 .1 21k / C 1 2k D 0 ;
Therefore .Xn /n , being the sum of independent centered r.v.’s, is a martingale

(Example 5.1 a)).
b) The r.v.’s Yk are square integrable and therefore Xn 2 L2 , as the sum of square
integrable r.v.’s. The associated increasing process of the martingale .Xn /n , i.e.
the compensator of the submartingale .Xn2 /n , is defined by A0 D 0 and for n 1
An D An1 C EŒ.Xn Xn1 /2 jFn1 D An1 C EŒYn2 jFn1 D An1 C EŒYn2 :
As EŒYn2 D 2 2n ,
1 1
An D 1 C C C n1 D 2.1 2n / :
2 2
(Note that the associated increasing process .An /n turns out to be deterministic,
as is always the case for a martingale with independent increments).
c) Thanks to b) the associated increasing process .An /n is bounded. As
An D EŒXn2
the martingale .Xn /n is bounded in L2 . Hence it converges a.s. and in L2 . It also

0
converges in L1 , as Lp convergence implies convergence in Lp for every p0 p.
5.6
a) We have
8
ˆ
ˆ; if u < s
<
f ug D A if s u < t
ˆ
:̂˝ if u t
therefore, in any case, f ug 2 Fu for every u.

b) If X is a martingale then the relation E.X / D E.X0 / for every bounded stopping
time is a consequence of the stopping theorem (Theorem 5.13) applied to the
stopping times 2 D and 1 D 0.
Conversely, assume that E.X / D E.X0 / for every bounded stopping time
and let us prove the martingale property , i.e. that, if t > s, for every A 2 Fs
E.Xt 1A / D E.Xs 1A / : (S.17)
The idea is to find two bounded stopping times 1 ; 2 such that from the relation
EŒX1 D EŒX2 (S.17) follows. Let us choose, for a fixed A 2 Fs , as in a) and
2 t. Now X D Xs 1A C Xt 1Ac and the relation EŒX D EŒXt can be written as
EŒXs 1A C EŒXt 1Ac D EŒX D EŒXt D EŒXt 1A C EŒXt 1Ac
and by subtraction we obtain (S.17).

5.7 Note first that Mt is integrable for every t > 0, thanks to the integrability
of eBt . Then, as indicated at the end of Sect. 5.1, it suffices to prove that g.x/ D
.ex K/C is a convex function. But this is immediate as g is the composition of
the functions x 7! ex K, which is convex, and of y 7! yC , which is convex and
increasing.
5.8
a) If s t, as fMs D 0g 2 Fs ,
E.1fMs D0g Mt / D E.1fMs D0g Ms / D 0 :
As Mt 0, we must have Mt D 0 a.s. on fMs D 0g (this is also a consequence

of Exercise 4.4 a)).
b1) is the infimum of the set of times at which M vanishes; therefore there exists
a sequence .tn /n with tn & such that Mtn D 0. Therefore, as M is right-
continuous, M D 0.
b2) Let T > 0 and D infftI Mt D 0g. Let us prove that P. T/ D 0. By the
stopping theorem and as M D 0 a.s.
E.MT / D E.MT^ / D E.MT 1f >Tg C M 1f Tg / D E.MT 1f >Tg / :
As MT > 0 a.s. this equality is possible only if P. > T/ D 1.
5.9
a) This is an extension to an m-dimensional Brownian motion of what we have
already seen in Example 5.2. Let s < t. As Bs is Fs -measurable whereas Bt Bs
is independent of Fs ,
1 2
E.Xt jFs / D E.eh ;Bs iCh ;Bt Bs i 2 j j t jFs /
1 2 1 2
D eh ;Bs i 2 j j t E.eh ;Bt Bs i jFs / D eh ;Bs i 2 j j t E.eh ;Bt Bs i /
1 2 1 2 .ts/ 1 2
D eh ;Bs i 2 j j t e 2 j j D eh ;Bs i 2 j j s D Xs :
b) Two ways of reasoning are possible.

First let us assume that m D 1. For large t, Bt .!/ ..2 C "/t log log t/1=2 by
the Iterated Logarithm Law. Therefore, a.s.,
1
Bt j j2 t
2
p
1=2 Bt 1 j j2 t
D ..2C"/t log log t/ ! 1
..2 C "/t log log t/1=2 2 ..2 C "/ log log t/1=2 t!C1
„ ƒ‚ …
!C1
and therefore
1 2t
Xt D e Bt 2 j j ! 0 a.s.
t!C1
1
If m > 1, then we know that Wt D j j h ; Bt i is a Brownian motion. We can
write
1 2t
Xt D ej jWt 2 j j
and then repeat the argument above with the Iterated Logarithm Law applied to
W.
The second approach is the following: let 0 < ˛ < 1. Then
1 2 1 2 1 2 j j2 t 1 2 t˛.1˛/
EŒXt˛ D e 2 ˛j j t EŒeh˛ ;Bt i D e 2 ˛j j t e 2 ˛ D e 2 j j :
Hence limt!C1 EŒXt˛ D 0. Let us denote by X1 the a.s. limit of X (whose

existence is guaranteed, X being a continuous positive martingale). Then by
Fatou’s lemma
˛
EŒX1 lim EŒXt˛ D 0 :
t!C1
˛
The positive r.v. X1 , having expectation equal to 0, is therefore D 0 a.s.
This second approach uses the nice properties of martingales (the previous
one with the Iterated Logarithm Law does not) and can be reproduced in other
similar situations.
c) If the martingale .Xt /t was uniformly integrable, it would also converge to 0 in
L1 and we would have E.Xt / ! 0 as t ! C1. But this it is not the case, as
E.Xt / D E.X0 / D 1 for every t 0.
5.10
a) The first point has already been proved in Remark 5.4 on p. 132. However, let
us produce a direct proof. If s t, as Bs is Fs -measurable whereas Bt Bs is
independent of Fs ,
E.Yt jFs / D EŒ.Bs C .Bt Bs //2 jFs t

D EŒB2s C 2Bs .Bt Bs / C .Bt Bs /2 jFs t
D B2s C 2Bs E.Bt Bs jFs / C EŒ.Bt Bs /2 t D B2s s D Ys :
„ ƒ‚ … „ ƒ‚ …
D0 Dts
If Y was uniformly integrable then it would converge a.s. and in L1 . This is not
possible, as we know by the Iterated Logarithm Law that there exists a sequence
of times .tn /n such that tn ! C1 and Btn D 0. Therefore limt!C1 Ytn D 1
a.s.
b) The requested equality would be immediate if was bounded, which we do not
know (actually it is not). But, for every t > 0, t ^ is a bounded stopping time
and by the stopping theorem 0 D EŒYt^ D EŒB2t^ .t ^ /. Therefore
E.B2t^ / D E.t ^ /
and we can take the limit as t ! C1
lim E.B2t^ / D E.B2 /

t!C1
by Lebesgue’s theorem (B2t^ max.a2 ; b2 /) and
lim E.t ^ / D E./
t!C1
by Beppo Levi’s theorem (t 7! t ^ is obviously increasing). In Example 5.3

we have seen that
b a
P.B D a/ D , P.B D b/ D
aCb aCb
Therefore we find
a2 b C b2 a
E./ D E.B2 / D D ab :
aCb
c1) Immediate as
Yt D B1 .t/2 t C C Bm .t/2 t
so that, thanks to a), .Yt /t turns out to be the sum of m .Ft /t -martingales.
c2) Recall that we know already that < C1 a.s. (Exercise 3.18). By the stopping
theorem, for every t > 0, EŒjB ^t j2 ^ t D 0, i.e.
EŒjB ^t j2 D mEŒ ^ t :
By a repetition of the argument of b), i.e. using Lebesgue’s theorem for the
left-hand side and Beppo Levi’s for the right-hand side, we find
EŒjB j2 D mEŒ
and, observing that jB j2 D 1 a.s., we can conclude that
1
EŒ D
m
5.11
a) As EŒMt Ms jFs D Ms EŒMt jFs D Ms2 we have
EŒ.Mt Ms /2 D EŒEŒ.Mt Ms /2 jFs

D EŒEŒMt2 2Mt Ms C Ms2 jFs D EŒMt2 Ms2 :
b) Let us assume M0 D 0 for simplicity. This is possible because the two

martingales .Mt /t and .Mt M0 /t have the same associated increasing process.
Note that the suggested associated increasing process vanishes at 0. We have then
EŒMt2 jFs D EŒ.Mt Ms CMs /2 jFs D EŒ.Mt Ms /2 C2.Mt Ms /Ms CMs2 jFs :
But EŒ.Mt Ms /2 jFs D EŒ.Mt Ms /2 D EŒMt2 Ms2 as M has independent

increments whereas EŒ.Mt Ms /Ms jFs D Ms EŒMt Ms jFs D 0. Therefore
EŒMt2 jFs D Ms2 C EŒMt2 Ms2 ;
from which it follows that Zt D Mt2 E.Mt2 / is a martingale, i.e. that hMit D
EŒMt2 .
c) M being a Gaussian process we know (Remark 1.2) that Mt Ms is independent
of Gs D .Mu ; u s/ if and only if, for every u s,
EŒ.Mt Ms /Mu D 0 :
But this relation is immediate as, M being a martingale,

EŒ.Mt Ms /Mu D E EŒ.Mt Ms /Mu jGs D E EŒ.Mt Ms /jGs Mu :

„ ƒ‚ …
D0
d) As hMi is deterministic and because of the independence of the increments,

1 2 ˇ
1 2 ˇ
E e
Mt 2
hMit ˇ Fs D e
Ms 2
hMit E e
.Mt Ms / ˇ Fs
1 2
D e
Ms 2
hMit E e
.Mt Ms / :

1 2
As Mt Ms is Gaussian, we have E e
.Mt Ms / D e 2
Var.Mt Ms / (recall
Exercise 1.6). As, thanks to a),
Var.Mt Ms / D EŒ.Mt Ms /2 D EŒMt2 Ms2 D hMit hMis ;
we have
1 2 ˇ
1 2 1 2
E e
Mt 2
hMit ˇ Fs D e
Ms 2
hMit e 2
.hMit hMis / D Zs :
5.12
a) We must prove that, for every n and for every bounded Borel function W
Rn ! R,
EŒYnC1 .Y1 ; : : : ; Yn / D EŒYn .Y1 ; : : : ; Yn / : (S.18)
But
EŒYnC1 .Y1 ; : : : ; Yn / D EŒXnC1 .X1 ; : : : ; Xn / ;

EŒYn .Y1 ; : : : ; Yn / D EŒXn .X1 ; : : : ; Xn /
as .Xn /n and .Yn /n are equivalent and, .Xn /n being a martingale,
EŒXnC1 .X1 ; : : : ; Xn / D EŒXn .X1 ; : : : ; Xn / ;
from which (S.18) follows.

b) The proof follows the same idea as in a). Let us denote by .Gt /t the natural
filtration of the process .Yt /t . By Remark 4.2, in order to show that EŒYt jGs D Ys ,
s t, we must just prove that
EŒYt 1C D EŒYs 1C
for every C in a class C of events that is stable with respect to finite intersections,
containing ˝ and generating G . If we choose as C the family of events of the
form fYs1 2 A1 ; : : : ; Ysn 2 An g, for n D 1; 2; : : : , s1 ; : : : ; sn s and A1 ; : : : ; An 2
B.R/, then we are led to show that
EŒYt 1fYs1 2A1 g : : : 1fYsn 2An g D EŒYs 1fYs1 2A1 g : : : 1fYsn 2An g :
But this follows from the fact that
EŒYt 1fYs1 2A1 g : : : 1fYsn 2An g D EŒXt 1fXs1 2A1 g : : : 1fXsn 2An g
EŒYs 1fYs1 2A1 g : : : 1fYsn 2An g D EŒXs 1fXs1 2A1 g : : : 1fXsn 2An g
as .Xt /t and .Yt /t are equivalent and
EŒXt 1fXs1 2A1 g : : : 1fXsn 2An g D EŒXs 1fXs1 2A1 g : : : 1fXsn 2An g
as .Xt /t is, by assumption, a martingale.

5.13
a) If t > s and recalling Remark 4.5, then
h Z t Z t
ˇ i

EŒYt jFs D E tBt ˇ

Bu du Fs D E t.Bs C Bt Bs /jFs E Bu jFs du
0 0
Z s Z t
D tBs C EŒt.Bt Bs jFs Bu du Bs du
„ ƒ‚ … 0 s
D0
Z s
D tBs Bu du .t s/Bs D Ys :
0
b) We have, for u s < t,

h Z t i
EŒ.Yt Ys /Bu D EŒ.tBt sBs /Bu E Bu Bv dv
s
Z t Z t
D .t s/u EŒBu Bv dv D .t s/u u dv D 0 :
s s
5.14
a) Let us prove first that X is F -measurable. Let A 2 B.R/, then we have
[
k [
k
fX 2 A; kg D fX 2 A; D mg D fXm 2 A; D mg : (S.19)
mD0 mD0
As f D mg D f mg n f m 1g 2 Fm Fk , from (S.19) we have

fX 2 A; kg 2 Fk for every k, i.e. fX 2 Ag 2 F .
We are left with the proof that EŒX1A D EŒX 1A for every A 2 F . If A 2
F , then A \ f D ng 2 Fn and
1
X 1
X 1
X
E.X1A / D E.X1A\f Dng/ D EŒ1A\f DngE.X jFn / D E.Xn 1A\f Dng/
nD0 nD0 nD0
1
X
D E.X 1A\f Dng/ D E.X 1A / :
nD0
b) If .n /n is a sequence of stopping times decreasing to and such that n takes

a discrete set of values (see Lemma 3.3), then repeating the proof of a) we
have E.X jFn / D Xn . As .Xt /t is right-continuous, Xn ! X as n ! 1
a.s. Moreover, the family .Xn /n is uniformly integrable by Proposition 5.4 and
therefore Xn ! X also in L1 . As E.X jF / D EŒE.X jFn /jF we deduce that
E.X jF / D E.Xn jF / ! E.X jF / D X :

n!1
The last equality in the relation above follows from the fact that X is F -measu-
rable, as a consequence of Propositions 2.1 and 3.6.
5.15 We have, recalling again Remark 4.5,

Z t
EŒMt jFs D e t E.Bt jFs / e u EŒBu jFs du
0
Z s Z t
t u
D e Bs e Bu du e u Bs du
0 s
Z s
D e t Bs e u Bu du .e t e s /Bs D Ms :
0
which proves the martingale property. Let us write down the increments of M, trying
to express them in terms of the increments of the Brownian motion. We have
Z t
t s
Mt Ms D e Bt e Bs e u Bu du
s
Z t Z t
t t s u
D e .Bt Bs / C .e e /Bs e .Bu Bs / du e u Bs du
s s
Z t
D e t .Bt Bs / e u .Bu Bs / du :
s
As Mt Ms is a function of the increments of B after time s, it follows that it is

independent of Fs (recall Remark 3.2).
5.16
a) Recalling the expression of the Laplace transform of Gaussian laws (Exer-
cise 1.6),
2
EŒe Bt t D e. 2 /t
2 2
so that the required limit is equal to C1, 1 or 0, according as 2 > , 2 D or
2
2
< .
b) Let t > s. With the typical method of factoring out the increment we have
2
E.e Bt t jFs / D e Bs t E.e .Bt Bs / jFs / D e Bs C 2 .ts/t
2
so that, if 2 D , .Xt /t is a martingale. Conversely, it will be a supermartingale
if and only if
2
.t s/ t s;
2
2
i.e. if 2 . The same argument also allows us to prove the result in the
submartingale case.
c) We have
Xt˛ D e˛ Bt ˛t :
It is obvious that we can choose ˛ > 0 small enough so that 12 ˛ 2 2 < ˛, so that,
for these values of ˛, X ˛ turns out to be a supermartingale, thanks to b). Being
positive, it converges a.s. Let us denote by Z its limit: as by Fatou’s lemma
1 2 2
lim E.Xt˛ / D lim E.e˛ Bt ˛t / D lim e. 2 ˛ ˛/t
D0;
t!1 t!1 t!1
we have
E.Z/ lim E.Xt˛ / D 0 ;

t!C1
from which, as Z 0, we deduce Z D 0 a.s. Therefore limt!1 Xt˛ D 0 and also

limt!1 Xt D 0 a.s.
d) The integral
Z C1
A1 D e Bs s ds
0
is well defined (possibly D C1), the integrand being positive. By Fubini’s

theorem
h Z C1 i Z C1
i
Z C1 2

Bs s Bs s
E.A1 / D E e ds D Ee ds D e. 2 /s ds :
0 0 0
2 2 2
Therefore, if 2 < , E.A1 / D . 2 /1 . If 2 then EŒA1 D C1.
If Wt D 1 B 2 t , we know, thanks to the scaling property, that .Wt /t is also a
Brownian motion. Therefore the r.v.’s
Z C1
e Bs s ds
0
and
Z C1 Z C1
e Ws s ds D eB 2 s s ds
0 0
have the same law. Now just make the change of variable t D 2 s.
• Note the apparent contradiction: we have limt!1 Xt D 0 for every value of
2 R; > 0, whereas, for 12 2 > , limt!1 E.Xt / D C1.
5.17
a) Let us consider the two possibilities: if x D C1, then
lim Mt^x D lim Mt D 0

t!C1 t!C1
whereas if x < C1, then Mt^x ! Mx D x as t ! C1 (the martingale is

continuous). Putting together the two possible cases we obtain (5.34).
b) As the stopping time t ^ x is bounded, by the stopping theorem, Theorem 5.13,
EŒMt^x D EŒM0 D 1. As the stopped martingale .Mt^x /t is bounded (actually
0 Mt^x x), by Lebesgue’s theorem,
1 D lim E.Mt^x / D EŒx1fx <C1g D xP.x < C1/ ;

t!C1
i.e. P.x < C1/ D 1x .

c) Just observe that, for x > 1, fM xg D fx < C1g, so that
1
P.M x/ D P.x < C1/ D (S.20)
x
d) By Exercise 5.9 we know that .Mt /t is a continuous martingale such that M0 D 1

and vanishing as t ! C1. Thanks to (S.20)

P.X x/ D P.2
X 2
x/ D P sup.2
Bt 2
2 t/ 2
x
t0
2

D P sup e2
Bt 2
t e2
x D P.M e2
x / D 1 P.M > e2
x / D 1 e2
x :
t0
We recognize the partition function of an exponential law with parameter 2

,
therefore X is exponential with parameter 2
.
• This is an example of a ruin problem. The process X is a ruin process modeling,
for instance, the net loss of an insurance company, coming from the difference
between the claims and the premia. In the average Xt tends to 1, as the premia
are computed in order to make a profit, but the outflow produced by the claims is
a random quantity which can (hopefully with small probability) become large.
If R denotes the reserves of the company at time 0, the event fX > Rg
corresponds to the fact that the company is unable to face its obligations, hence
of being bankrupt.
Investigating this kind of situation is therefore of interest and Exercise 5.17 is
a first example. The process Xt D Bt
t is not really a good ruin process and
more realistic ones have been developed (see Asmussen 2000 if you want to go
further). It is nevertheless a good first look.
5.18
1
a) If s D 1tt
, then t D sC1s
and 1 t D sC1 . (5.35) therefore holds if and only
if the process Bs D .1 C s/X sC1 s is a Brownian motion. As it is obviously a
centered Gaussian process that vanishes for s D 0, we just have to prove that
1
E.Bs Bt / D s ^ t. Note that s 7! sC1s
D 1 sC1 is increasing. Therefore if s t,
then also sC1 tC1 and, recalling the form of the covariance function of the
s t
Brownian bridge,
s t
E.Bs Bt / D .1 C s/.1 C t/E X sC1s X t D .1 C s/.1 C t/ 1 Ds
tC1 sC1 tC1
and therefore B is a Brownian motion.
b) We have, with the change of variable s D t
1t
,
1
P sup Xt > a D P sup .1 t/B 1t
t > a D P sup Bs > a
0t1 0t<1 s>0 s C 1
1
D P sup .Bs .s C 1/a/ > 0 D P sup Bs sa > a :
s>0 s C 1 s>0
Thanks to Exercise 5.17 the r.v. sups>0 Bs sa has an exponential law with
parameter 2a, therefore
2
P sup Xt > a D e2a
0t1
2
and the partition function of sup0t1 Xt is F.x/ D 1 e2x for x 0. Taking
2
the derivative, the corresponding density is f .x/ D 4xe2x for x 0.
5.19
1
a) As computed in Example 5.3 (here a D x, b D 1) P.B D x/ D 1Cx .
b) The important observation is that if Z x, i.e.
min Bt x ;
t1
then B has gone below level x before passing at 1, so that B D x. Therefore,
by a),
1 ,
P.Z x/ D P.B D x/ D
1Cx
1
i.e. the partition function of Z is P.Z x/ D 1 1Cx
. Taking the derivative, the
density of Z is
1
fZ .x/ D , x>0:
.1 C x/2
The expectation of Z would be

Z C1
x
dx
0 .1 C x/2
1
but the integral does not converge, as the integrand behaves like x as x ! C1.
Therefore Z does not have finite expectation.
5.20
2
a) We have Mt D e2Bt 2 t so that this is the martingale of Example 5.2 for

D 2.
b1) By the Iterated Logarithm Law. . .
b2) By the stopping theorem
E.M ^t / D E.M0 / D 1 : (S.21)
As a X ^t b, t 7! M ^t is bounded and we can apply Lebesgue’s theorem

taking the limit as t ! C1 in (S.21), which gives
1 D E.M / D E.e2X / D e2a P.X D a/ C e2b P.X D b/

D e2a .1 P.X D b// C e2b P.X D b/
hence
1 e2a
P.X D b/ D
e2b e2a
and the limit as ! C1 of this probability is equal to 1.
5.21
a) We have
EŒB3t 3tBt jFs D EŒ.Bs C .Bt Bs //3 3t.Bs C .Bt Bs //jFs

D EŒB3s C 3B2s .Bt Bs / C 3Bs .Bt Bs /2 C .Bt Bs /3 jFs 3tBs
D B3s C 3.t s/Bs 3tBs D Xs :
b) By the stopping theorem applied to the bounded stopping time ^ t,
0 D E.X ^t / D E.B3^t / 3E.. ^ t/B ^t / ;
i.e.
E.B3^t / D 3EŒ. ^ t/B ^t :
Now E.B3^t / !t!C1 E.B3 /, as jB3^t j max.a; b/3 and we can apply
Lebesgue’s theorem. The same argument allows us to take the limit on the right-
hand side, since j. ^ t/B ^t j max.a; b/ and we know (Exercise 5.31) that
is integrable. We can therefore take the limit and obtain
1 1 b3 a a3 b 1 ab 1
E.B / D E.B3 / D D .b2 a2 / D ab.ba/ :
3 3 aCb aCb 3 aCb 3
We have actually seen in Exercise 5.31 that P.B D a/ D b

aCb and P.B D
b/ D aCb
a
. Therefore, recalling that E.B / D 0, we find
1
Cov.; B / D E.B / D ab.b a/ :
3
Note that the covariance is equal to zero if a D b, in agreement with the fact that,
if a D b, then and B are independent (Exercise 3.18).
5.22 We must prove that, if s t,
EŒMt Nt 1A D EŒMs Ns 1A (S.22)
for every A 2 Hs or at least for every A in a subclass Cs Hs , which generates Hs ,

contains ˝ and is stable with respect to finite intersections (this is Remark 4.2). One
can consider the class of the events of the form A1 \A2 with A1 2 Ms , A2 2 Ns . This
class is stable with respect to finite intersections and contains both Ms (choosing
A2 D ˝) and Ns (with A1 D ˝). As Mt 1A1 and Nt 1A2 are independent (the first is
Mt -measurable whereas the second one is Nt -measurable) we have
EŒMt Nt 1A1 \A2 D EŒMt 1A1 Nt 1A2 D EŒMt 1A1 EŒNt 1A2 D EŒMs 1A1 EŒNs 1A2
D EŒMs 1A1 Ns 1A2 D EŒMs Ns 1A1 \A2 ;
hence (S.22) is satisfied for the events of the class Cs .

5.23
a) If .Mt2 /t is a martingale we have, for every s t,

0 D E Mt2 Ms2 jFs D E .Mt Ms C Ms /2 Ms2 jFs

E .Mt Ms /2 C 2Ms .Mt Ms /jFs D E .Mt Ms /2 jFs :
Now, taking the expectation, EŒ.Mt Ms /2 D 0, i.e. Mt D Ms a.s. Let us

determine a negligible event N such that, for every ! 62 N, Bt D Bs for every
t; s. For q; r 2 Q let Nq;r be the negligible event such that for ! 62 Nq;rSMq D Mr .
It is easy to verify that, as M is supposed to be continuous, N D q;r2Q Nq;r
satisfies this requirement.
0 0
b1) The function x 7! jxjp being convex, .jMt jp /t is a submartingale and, for s t,
0
0
E jMt jp jFs jMs jp : (S.23)
0
If p0 < p then x 7! jxjp=p is also convex so that

0 0
0
0
jMs jp D E jMt jp jFs D E .jMt jp /p=p jFs E jMt jp jFs /p=p ;
0
0
which gives E jMt jp jFs jMs jp and this inequality together with (S.23)
allows us to conclude the proof.
b2) For p D 2 this is already proved. If for p > 2 .jMt jp /t is a martingale, then it is
a also martingale for p D 2 by b1) and is therefore constant thanks to a).
5.24
a) With the usual method of factoring out the increment we have for s t,

E Bi .t/Bj .t/jFs D E Bi .s/ C .Bi .t/ Bi .s/ Bj .s/ C .Bj .t/ Bj .s/ jFs

D E Bi .s/Bj .s/ C Bi .s/.Bj .t/ Bj .s// C Bj .s/.Bi .t/ Bi .s//

C.Bi .t/ Bi .s//.Bj .t/ Bj .s//jFs :
Now just observe that, the increments being independent of Fs ,

E Bi .s/.Bj .t/ Bj .s//jFs D Bi .s/E Bj .t/ Bj .s/jFs D 0

E Bj .s/.Bi .t/ Bi .s//jFs D Bj .s/E Bi .t/ Bi .s/jFs D 0

E .Bi .t/ Bi .s//.Bj .t/ Bj .s//jFs D E .Bi .t/ Bi .s//.Bj .t/ Bj .s// D 0
and therefore EŒBi .t/Bj .t/jFs D Bi .s/Bj .s/.

b) This is proved using the same idea as in a), only more complicated to express:
one must factor out Bi1 .t/ : : : Bid .t/ as
.Bi1 .s/ C .Bi1 .t/ Bi1 .s// : : : .Bid .s/ C .Bid .t/ Bid .s// D Bi1 .s/ : : : Bid .s/ C : : :
where the rightmost : : : denotes a r.v. which is the product of some Bik .s/ (which
are already Fs -measurable) and of some (at least one) terms of the kind Bik .t/
Bik .s/. These are centered r.v.’s which are, moreover, independent with respect to
Fs . The conditional expectation of their product with respect to Fs is therefore
equal to 0, so that
EŒBi1 .t/ : : : Bid .t/jFs D Bi1 .s/ : : : Bid .s/ ;
hence t 7! Bi1 .t/ : : : Bid .t/ is a martingale.

c) det Bt is a sum of terms of the form B1;j1 B2;j2 : : : Bm;jm , with j1 ; : : : ; jm being
distinct indices. Thanks to b) every such term is a martingale. det Bt is therefore
a sum of martingales, hence a martingale itself.
Moreover, Yt D det.Bt Bt / D det Bt det Bt D .det Bt /2 D Xt2 , so that Y is
the square of a martingale and therefore a submartingale.
Throughout this part of the solution we neglected to prove that both Xt and Yt
are integrable, but this immediate.
• Note that each of the processes t 7! Bi .t/ is a martingale with respect to the
filtration t 7! .Bi .u/; u t/. As these are independent (Remark 3.2 b)), from
Exercise 5.22 it follows immediately that t 7! Bi .t/Bj .t/ is a martingale with
respect to the filtration Gi;j .t/ D .Bi .u/; u t/ _ .Bj .u/; u t/.
5.25
a) If t T, thanks to the freezing Lemma 4.1 we have, as BT Bt is independent
of Ft ,
EŒ1fBT >0g jFt D EŒ1fBt CBT Bt >0g jFt D .Bt ; t/ ;
where, denoting by ˚ the partition function of an N.0; 1/ law,

x
.x; t/ D EŒ1fxCBT Bt >0g D P.BT Bt > x/ D ˚ p :
T t
p
We use here the fact that BT Bt T t Z with Z N.0; 1/. Hence
B
t
EŒ1fBT >0g jFt D ˚ p :
T t
b) From a) Mt D EŒ1fBT >0g jFt hence M is a uniformly integrable martingale. As
8
ˆ
x <C1 if x > 0
ˆ
lim p D 0 if x D 0
t!T T t ˆ
:̂1 if x < 0
we have
8
ˆ
x <1
ˆ if x > 0
lim ˚ p D 1
if x D 0
t!T T t ˆ2
:̂0 if x < 0 :
Hence (the event fBT D 0g has probability 0)

B
t
lim ˚ p D 1fBT >0g a.s.
t!T T t
5.26
a) The integral converges absolutely as u !
7 p1 is integrable at the origin and the
u
path t 7! Bt .!/ is bounded (there is no need here of the Iterated Logarithm
Law. . . ).
b) In order to prove that it is a Gaussian process, we use the usual technique of

approximating the integral with Riemann sums, which are obviously Gaussian,
using the property of stability of the Gaussian r.v.’s in the limit (Proposition 1.9).
In this case it is not obvious that the Riemann sums converge to the integral, as
the integrand diverges at 0. This argument, however, allows us to see that, for
every " > 0, the process e X t D 0 for t < " and
Z t
e Bu
Xt D p du; t"
" u
is Gaussian. One then takes the limit as " & 0 and uses again the properties of
stability of Gaussianity with respect to limits in law. As the r.v.’s Xt are centered,
for the covariance we have, for s t,
Z Z
Z t Z s
EŒBu Bv
t s
u^v
Cov.Xs ; Xt / D EŒXs Xt D p p dv D
du du p p dv
0 0 u v 0 0 u v
Z Z Z t Z s
s s
u^v u^v
D du p p dv C du p p dv :
0 0 u v s 0 u v
The last two integrals are computed separately: keeping carefully in mind which
among u and v is smaller, we have for the second one
Z Z Z Z Z Z sp
t s
u^v t
v s t
1
du p p dv D du p p dv D p du v dv
s 0 u v s 0 u v s u 0
4 p p
D s3=2 . t s/ :
3
Whereas for the first one
Z s Z s Z s Z u Z s Z s
u^v 1 p p 1
du p p dv D p du v dv C u du p dv
0 0 u v 0 u 0 0 u v
Z Z s
2 s p p p 1 2 4 2 2 2
D u du C 2 u. s u/ du D s C s s2 D s :
3 0 0 3 3 3
In conclusion
2 2 4 3=2 p p
Cov.Xs ; Xt / D s C s . t s/ :
3 3
c) The most simple and elegant argument consists in observing that .Xt /t is a square
integrable continuous process vanishing at the origin. If it were a martingale the
paths would be either identically zero or with infinite variation a.s. Conversely
the paths are C1 . Therefore .Xt /t cannot be a martingale.
Alternatively one might also compute the conditional expectation directly and
check the martingale property. We have, for s t.
Z t
Bu ˇ
Z s
Bu Z t
Bu ˇ
E.Xt jFs / D E p du ˇ Fs D p du CE p du ˇ Fs :
0 u u u
„0 ƒ‚ … s
DXs
(S.24)
For the last conditional expectation we can write, as described in Remark 4.5,
Z t
Bu ˇ
Z t
1
Z t
1 p p
E p du ˇ Fs D p EŒBu jFs du D Bs p du D 2. t s/Bs ;
s u s u s u
p p
so that E.Xt jFs / D Xs C 2. t s/Bs and X is not a martingale.
5.27 Let us denote by EP , EQ the expectations with respect to P and Q,

respectively.
a) Recall that, by definition, for A 2 Fs , Q.A/ D EP .Zs 1A /. Let s t. We must
prove that, for every A 2 Fs , EP .Zt 1A / D EP .Zs 1A /. But as A 2 Fs Ft , both
these quantities are equal to Q.A/.
b) We have Q.Zt D 0/ D EP .Zt 1fZt D0g / D 0 and therefore Zt > 0 Q-a.s. Moreover,
as fZt > 0g fZs > 0g a.s. if s t (Exercise 5.8), then for every A 2 Fs ,
EQ .1A Zt1 / D EQ .1A\fZt >0g Zt1 / D P.A \ fZt > 0g/ P.A \ fZs > 0g/
D EQ .1A Zs1 /
and therefore .Zt1 /t is a Q-supermartingale.

c) Let us assume P Q: this means that P.A/ D 0 whenever Q.A/ D 0. Therefore
also P.Zt D 0/ D 0 and
EQ .Zt1 / D EP .Zt Zt1 / D 1 :
The Q-supermartingale .Zt /t therefore has constant expectation and is a Q-mar-

tingale by the criterion of Exercise 5.1. Alternatively, observe that
dPjFt
D Zt1
dQjFt
and .Zt1 /t is a martingale by a).
5.28
a) Note that the -algebra GnC1 is generated by the same r.v.’s Xsk=2n that generate
Gn and some more in addition, therefore GnC1 Gn .
W
Let, moreover, G 0 D n1 Gn . As Gn G for every n, clearly G 0 G .
Moreover, the r.v.’s Xsk=2n , k D 1; : : : ; 2n , n D 1; 2; : : : , are all G 0 -measurable.
Let now u s. As the times of the form sk=2n , k D 1; : : : ; 2n , n D 1; 2; : : : ,
are dense in Œ0; s, there exists a sequence .sn /n of times of this form such that
sn ! u as n ! 1. As the process .Xt /t is assumed to be continuous, Xsn ! Xu
and Xu turns out to be G 0 -measurable for every u s. Therefore G 0 G , hence
G0 D G.
b1) The sequence .Zn /n is a .Gn /n -martingale: as GnC1 Gn ,
Zn D EŒXt jGn D EŒE.Xt jGnC1 /jGn D EŒZnC1 jGn :
As EŒZn2 EŒXt2 by Jensen’s inequality (Proposition 4.2 d) with ˚.x/ D

x2 ), the martingale .Zn /n is bounded in L2 and converges a.s. and in L2 by
Theorem 5.6.W Its limit, by Proposition 5.5, is equal to EŒXt jG thanks to the
relation G D n1 Gn that was proved in a).
b2) If Xt is only assumed to be integrable, then .Zn /n is again a uniformly integrable
martingale, thanks to Proposition 5.4. Hence it converges a.s. and in L1 to
EŒXt jG . The difference is that in the situation b1) we can claim that EŒXt jGn
converges to a r.v. that is the best approximation in L2 of Xt by a G -measurable
r.v. (Remark 4.3), whereas now this optimality claim cannot be made.
5.29
a) This is the stopping theorem applied to the bounded stopping times ^ t and to
the martingales .Bt /t and .B2t t/t .
b) We have, as t ! C1, ^ t % and B ^t ! B . By Beppo Levi’s theorem
EŒ ^ t % EŒ < C1 ( is integrable by assumption), so that
EŒB2^t EŒ < C1
for every t 0. Therefore the martingale .B ^t /t is bounded in L2 and we can

apply Doob’s inequality, which gives
h i
E sup B2^t 4 sup EŒB2^t 4EŒ < C1 ;
t0 t0
i.e. (5.36).
c) Thanks to (5.36) the r.v. B D supt0 B ^t is square integrable and we have
jB ^t j B and B2^t B 2 . These relations allow us to apply Lebesgue’s
theorem and obtain
EŒB D lim EŒB ^t D 0

t!C1
EŒB2 D lim EŒB2^t D lim EŒ ^ t D EŒ :

t!C1 t!C1
5.30
a) By the stopping theorem applied to the bounded stopping time t ^ a we have
EŒMtâ D EŒM0 D 1. If, moreover, 0 the martingale .Ma ^t /t is bounded,
as Ba ^t a, so that Ma ^t e a . We can therefore apply Lebesgue’s theorem
and, recalling that a < C1 a.s.,
EŒMa D lim E.Ma ^t / D 1 :

t!C1
If < 0 then Ma ^t is not bounded and Lebesgue’s theorem can no longer be

applied.
b) The previous equality can be rewritten as
1 2 1 2
1 D EŒMa D EŒe Ba 2 a
D EŒe a 2 a
;
i.e.
1 2
EŒe 2 a D e a : (S.25)
p
Now if
D 12 2 , i.e. D 2
(recall that (S.25) was proved for 0 so
that we have to discard the negative root), (S.25) can be rewritten, for
0, as
p
EŒe
a D ea 2
For
> 0 the Laplace transform is necessarily equal to C1 as a consequence
of the fact that E.a / D C1 (see Exercise 3.20), thanks to the inequality a
1
a

e .
c) If X1 ; : : : ; Xn are i.i.d. r.v.’s, having the same law as a , then the Laplace transform
of n2 .X1 C C Xn / is, for
0,
q n p
a 2
exp a 2
n2 D e :
The laws of the r.v.’s n2 .X1 C C Xn / and X1 have the same Laplace transform
and therefore they coincide, as seen in Sect. 5.7; hence the law of a is stable with
exponent 12 .
2
5.31 We know from Example 5.2 that, for 2 R, Mt D e Bt 2 t is a martingale.
The stopping theorem gives
2
1 D EŒMt^ D EŒe Bt^ 2 .t^ /
:
As jBt^ j a we can apply Lebesgue’s theorem and take the limit as t ! C1. We
obtain
2
1 D EŒe B 2
:
As B and are independent, as a particular case, for m D 1, of Exercise 3.18 b),

we have
2
1 D EŒe B EŒe 2 : (S.26)
B takes the values a and a with probability 12 , hence
1 a
EŒe B D .e C e a / D cosh. a/ ;
2
so that by (S.26)
2 1
EŒe 2
D
cosh. a/
2
p
and we just have to put
D 2 , i.e. D 2
.
5.32
1 2
a) In a way similar to Exercise 5.9 we can show that Mt D ei Bt C 2 t is a (complex)
.Ft /t -martingale:
1 2 1 2
E.Mt jFs / D e 2 t E.ei Bs ei .Bt Bs / jFs / D e 2 t ei Bs E.ei .Bt Bs / /
1 2 1 2
D e 2 t ei Bs e 2 .ts/
D Ms :
This implies that the real part of M is itself a martingale and note now that
1 2
<Mt D cos. Bt / e 2 t . The martingale relation EŒXt jFs D Xs can also be
checked directly using the addition formulas for the cosine function, giving rise
to a more involved computation.
b) By the stopping theorem applied to the bounded stopping time ^ t,
1 2
1 D E.X0 / D EŒcos. B ^t / e 2 . ^t/
: (S.27)
But jB ^t j < a hence, with the conditions on , j B ^t j < 2 and recalling the
behavior of the cosine function, cos. B ^t / cos. a/ > 0. We deduce that
1 2
EŒe 2 . ^t/
cos. a/1
1 2
and, letting t ! 1, by Beppo Levi’s Theorem the r.v. e 2 is integrable. Thanks
to the upper bound
1 2 1 2
0 < cos. B ^t / e 2 . ^t/
e2
we can apply Lebesgue’s theorem as t ! C1 in (S.27). As < C1 a.s. and

jB j D a, we obtain
1 2 1
E.e 2 / D
cos. a/
from which, replacing

D 12 2 , (5.38) follows.
c) The expectation of can be obtained as the derivative of its Laplace transform at
the origin; therefore
1 ˇ
d ˇ
E./ D p ˇ D a2 :
d
cos.a 2
/
D0
This result has already been obtained in Exercise 5.31 b). Finally, for p 0
and " > 0, we have xp c."; p/ e"x for x 0 (just compute the maximum of
x 7! xp e"x , which is c."; p/ D pp "p ep ). Therefore p c."; p/ e" . Now just
2
choose an " with 0 < " < 8a 2.
d) First of all note that

D 0 is not an eigenvalue, as the solutions of 12 u00 D 0
are linear-affine functions, which cannot vanish at two distinct points unless they
vanish identically. For
6D 0 the general integral of the equation
1 00
u
u D 0
2
p p
2
is u.x/ D c1 ex C c2 ex 2
: The boundary conditions impose on the constants

the conditions
p p
2
c1 ea C c2 ea 2
D0
p p
a 2
a 2
c1 e C c2 e D0:
This system, in the unknowns c1 ; c2 , has solutions different from c1 D c2 D 0 if

and only if the determinant of the matrix
p p !
ea p2
eap 2
ea 2
ea 2
p p p
vanishes, i.e. if and only if e2a 2
e2a 2
D 0 which gives 2a 2
D ik for
k 2 Z. Therefore the eigenvalues are the numbers
k2 2
k D 1; 2; : : :
8a2
2
They are all negative and the largest one is of course 8a2.
• This exercise completes Exercise 5.30 where the Laplace transform of was
computed for negative reals. A more elegant way of obtaining (5.38) is to observe
that the relation
1
EŒe
D p ,
>0;
cosh.a 2
/
which is obtained in Exercise 5.30, defines the Laplace transform of as the

holomorphic function
1
z 7! p (S.28)
cosh.a 2z/
p
on =z < 0. But z 7! cosh.a 2z/ is a holomorphic p function on the whole
complex plane which can be written as z 7! cos.a 2z/ for <z > 0. Hence the
is holomorphic as far as <z is smaller than the first positive
function in (S.28) p
2
p
zero of
7! cos.a 2
/ i.e. 8a 2 . Note that z 7! cosh.a 2z/ is holomorphic on
C, even in the presence of the square root, because the power series development
of cosh only contains even powers, so that the square root “disappears”.
5.33
a) By the Iterated Logarithm Law . . .
b) We have
2 2
Mt D e .Bt Ct/. 2 C /t
D e Bt 2 t ;
which is a martingale (an old acquaintance. . . ). As t ^ is a bounded stopping

time, by the stopping theorem EŒMt^ D EŒM0 D 1.
2
c) As 2 C 0,
2
Xt^ C .t ^ / Xt^ a :
2
Hence 0 Mt^ e a . Moreover, Xt^ !t!C1 X D a, hence
2
Mt^ ! e a. 2 C /
t!C1
and by Lebesgue’s theorem
2
1 D lim EŒMt^ D EŒe a. 2 C /
;
t!C1
i.e. (5.40). p
2
d) Let > 0 such that 2
C D
, i.e. D 2 C 2
. With this choice of
(5.40) becomes
p
2 C2
/
EŒe
D ea. : (S.29)
The expectation of is obtained by taking the derivative at 0 of the Laplace

transform (S.29):
d a.p2 C2
/ ˇˇ a
E./ D e ˇ D
d

D0
• Remark: the passage time of a Brownian motion with a positive drift through a
positive level has a finite expectation, a very different behavior compared to the
zero drift situation (Exercise 3.20).
6.1
a) As in Sect. 4.4 and Exercise 4.10 b), we have, for s t,
1 1
Ct;s D Kt;s Ks;s ; Yt;s D Xt Kt;s Ks;s Xs :
1
Yt;s is a centered Gaussian r.v. with covariance matrix Kt;t Kt;s Ks;s Ks;t .
Moreover, by the freezing Lemma 4.1, for every bounded measurable function f
EŒ f .Xt /jXs D EŒ f .Ct;s Xs C Yt;s /jXs D ˚f .Xs / ;
where ˚f .x/ D EŒ f .Ct;s xCYt;s /. Therefore the conditional law of Xt given Xs D
1
x is the law of the r.v. Ct;s x C Yt;s , i.e. is Gaussian with mean Ct;s x D Kt;s Ks;s x
1
and covariance matrix Kt;t Kt;s Ks;s Ks;t .
b) The Markov property with respect to the natural filtration requires that, for every
bounded measurable function f W Rm ! R,
Z
EŒ f .Xt /jGs D f .y/p.s; t; Xs ; dy/ :
Let us first determine what the transition function p should be: it is the law of Xt
given Xs D x which, as seen in a), is the law of Ct;s x C Yt;s . In a) we have also
proved that
Z
EŒ f .Xt /jXs D f .y/p.s; t; Xs ; dy/ :
We must therefore prove that
EŒ f .Xt /jGs D EŒ f .Xt /jXs : (S.30)
Let us prove that (6.32) implies the independence of Yt;s and the -algebra Gs D
.Xu ; u s/. This will imply, again by the freezing Lemma 4.1,
EŒ f .Xt /jGs D EŒ f .Ct;s Xs C Yt;s /jGs D ˚f .Xs / :
The covariances between Yt;s and Xu are given by the matrix (recall that all these
r.v.’s are centered)
E.Ys;t Xu / D E.Xt Xu / E.Kt;s Ks;s

1
Xs Xu / D Kt;u Ks;t Ks;s
1
Ks;u ;
which vanishes if and only if (6.32) holds. Hence if (6.32) holds then the Markov
property is satisfied with respect to the natural filtration.
Conversely, let us assume that .Xt /t is a Markov process with respect to its
natural filtration. For s t we have, by the Markov property,
1
E.Xt jGs / D E.Xt jXs / D Kt;s Ks;s Xs
hence
Kt;u D E.Xt Xu / D EŒE.Xt jGs /Xu D E.Kt;s Ks;s

1
Xs Xu / D Kt;s Ks;s
1
Ks;u ;
i.e. (6.32).
6.2
a) The simplest approach is to observe that the paths of the process .Xt /t have finite
variation (they are even differentiable), whereas if it was a square integrable
continuous martingale its paths would have infinite variation (Theorem 5.15).
Otherwise one can compute the conditional expectation
E.Xt jFs /
and check that it does not coincide with Xs . This can be done directly:
Z t ˇ
Z t
E.Xt jFs / D Xs C E Bu du ˇ Fs D Xs C EŒBu jFs du D Xs C .t s/Bs :
s s
b) If s t,
Z t Z s Z t Z s
Kt;s D Cov.Xt ; Xs / D E Bu du Bv dv D du EŒBu Bv dv
0 0 0 0
Z t Z s Z t Z s Z s Z s
D du u ^ v dv D du v dv C du u ^ v dv
0 0 s 0 0 0
Z Z Z Z
s2 s u s s
D .t s/ C du v dv C du u dv
2 0 0 0 u
s2 s3
D .t s/ C
2 3
In order for .Xt /t to be Markovian, using the criterion of Exercise 6.1 b), the
following relation
1
Kt;u D Kt;s Ks;s Ks;u (S.31)
must hold for every u s t. But

s2 s3 3 u2 u3
1
Kt;s Ks;s Ks;u D .t s/ C .s u/ C :
2 3 s3 2 3
If we choose u D 1, s D 2, t D 3, this quantity is equal to 74 56 D 35

24 , whereas
2 3
Kt;u D u2 .t u/ C u3 D 12 C 23 D 56 . .Xt /t is not a Markov process.
c) If s t,
Z t Z t Z s Z t
s2
Cov.Xt ; Bs / D E Bu Bs du D u^s du D u duC s du D Cs.ts/ :
0 0 0 s 2
Conversely, if s t,
Z t Z t Z t
t2
Cov.Xt ; Bs / D E Bu Bs du D u ^ s du D u du D :
0 0 0 2
d) If Yt D .Bt ; Xt /, the covariance function of this process is, for s t,

!
s2
Cov.Bt ; Bs / Cov.Bt ; Xs / s
Kt;s D D s2 s3
2
s2
:
Cov.Xt ; Bs / Cov.Xt ; Xs / 2
C s.t s/ 3
C 2
.t s/
Therefore the right-hand side of (S.31) becomes here

! !1 !
s2 s2 u2
s s u
s2 s3
2
s2
2
s s3
2 u2 u3
2
u2
:
2 C s.t s/ 3 C 2 .t s/ 2 3 2 C u.s u/ 3 C 2 .s u/
Since
!1 !
s2 12 s3 s2
3 2
1 s
Ks;s D 2
s s3
2 D 4 2
2 3
s s2 s
we have
! !
12 s2 s3 2
1 s s2 1 0
Kt;s Ks;s D 4 s2 s3
2
s2
3
2 D
s 2
C s.t s/ 3
C 2
.t s/ s2 s ts 1
so that
!
u2
1 1 0 u
Kt;s Ks;s Ks;u D u2 u3 u2
2
ts 1 2
C u.s u/ 3
C 2
.s u/
!
u2
u
D u2 u3 u2
2 D Kt;u :
2
C u.t u/ 3
C 2
.t u/
Condition (S.31) is therefore satisfied and .Bt ; Xt /t is a Markov process.

• We have found an example of a process, .Xt /t , which is not Markovian, but which
is a function of a Markov process.
6.3
a1) Let us observe first that X is adapted to the filtration .Fg.t/ /t . Moreover,
X is a Gaussian process: its finite-dimensional distributions turn out to be
linear transformations of finite-dimensional distributions of a Brownian motion.
Hence we expect its transition function to be Gaussian.
Using the Markov property enjoyed by the Brownian motion, for a mea-
surable bounded function f and denoting by p the transition function of the
Brownian motion, we have with the change of variable z D h.t/y
Z
EŒ f .Xt /jFg.s/ D EŒ f .h.t/Bg.t/ /jFg.s/ D f .h.t/y/ p.g.s/; g.t/; Bg.s/; dy/
Z
1 C1
.y Bg.s/ /2
Dp f .h.t/y/ exp dy
2.g.t/ g.s// 1 2.g.t/ g.s//
Z C1 . z Xg.s/ /2
1 h.t/ h.s/
D p f .z/ exp dz
h.t/ 2.g.t/ g.s// 1 2.g.t/ g.s//
Z .z h.t/ 2
h.s/ Xg.s/ /
C1
1
D p f .z/ exp dz
h.t/ 2.g.t/ g.s// 1 2h.t/2 .g.t/ g.s//
from which we deduce that X is a Markov process with respect to the filtration
.Fg.t/ /t and associated to the transition function
h.t/
q.s; t; x; dy/ N x; h.t/2 .g.t/ g.s// : (S.32)
h.s/
a2) From (S.32) we have in general Xt N.0; h.t/2 g.t//. Hence under the
condition of a2) Xt N.0; t/, as for the Brownian motion. However, for the
transition function of X we have
ptg.s/ t
q.s; t; x; dy/ N p x; .g.t/ g.s//
sg.t/ g.t/
whereas recall that the transition function of a Brownian motion is
p.s; t; x; sy/ N.x; t s/

p
hence X cannot be a Brownian motion, unless g.t/ D t (and in this case
Xt D Bt ).
b) The clever reader has certainly sensed the apparition of the Iterated Logarithm
Law. Actually, by a simple change of variable
Xt h.t/Bg.t/
lim p D lim p
t!C1 2
2g.t/h .t/ log log g.t/ t!C1 2g.t/h2 .t/ log log g.t/
Bg.t/
D lim p D1:
t!C1 2g.t/ log log g.t/
6.4
a) X is clearly a Gaussian process and we know, by Exercise 6.1, that it is
Markovian if it satisfies the relation
1
Kt;u D Kt;s Ks;s Ks;u ; for u s t : (S.33)
Now, if s t,
Kt;s D Cov.Xs ; Xt / D e .tCs/ EŒBe2 t Be2 s D e .tCs/ e2 s D e .ts/ :
In particular, Ks;s D Var.Xs / 1. Therefore

1
Kt;s Ks;s Ks;u D e .ts/ e .su/ D e .tu/ D Kt;u
so that (S.33) is satisfied and X is Markovian. In order to determine its transition

function, just recall that p.s; t; x; dy/ is the conditional distribution of Xt given
Xs D x. X being a Gaussian process, we know that this is Gaussian with mean
Kt;s
E.Xt / C .x E.Xs // D e .ts/ x
Ks;s
and variance
2
Kt;s
Kt;t D 1 e2 .ts/ :
Ks;s
Therefore p.u; x; dy/ N.e u x; 1 e2 u /. As both mean and variance are
functions of t s only, X is time homogeneous.
b1) Xt is Gaussian, centered and has variance D 1. In particular, the law of Xt does
not depend on t.
b2) This is immediate: the two random vectors are both Gaussian and centered. As
we have seen in a1) that the covariance function of the process satisfies the
relation Kt1 ;t2 D Kt1 Ch;t2 Ch , they also have the same covariance matrix.
c) Under Px the law of Zt is p.t; x; dy/ N.e t x; 1 e2 t /. As t ! C1 the
mean of this distribution converges to 0, whereas its variance converges to 1.
Thanks to Exercise 1.14 a) this implies (6.33).
6.5
a) From Exercise 4.15 we know that .Xt /t1 is a centered Gaussian process and that,
for s t,
Kt;s D E.Bt Bs / C stE.B21 / sE.Bt B1 / tE.Bs B1 / D s st D s.1 t/ :
Xt therefore has variance t2 D t.1 t/. Going back to Exercise 6.1 a), the
conditional law of Xt given Xs D x is Gaussian with mean
Kt;s 1t
xD x (S.34)
Ks;s 1s
and variance
2
Kt;s s.1 t/2 1t
Kt;t D t.1 t/ D .t s/ : (S.35)
Ks;s 1s 1s
b) With the notations of Exercise 6.1 if u s t, Kt;u D u.1 t/, whereas
1 1
Kt;s Ks;s Ks;u D s.1 t/ u.1 s/ D u.1 t/ D Kt;u :
s.1 s/
Therefore condition (6.32) is satisfied and .Xt /t is a Markov process. The

transition function coincides with the conditional distribution of Xt given Xs D x,
1t 1t
i.e. p.s; t; x; / N. 1s x; 1s .t s//.
6.6
a) As the joint distributions of X are also joint distributions of B, X is also a Gaussian
process. As for the covariance function, for s t 1,
Kt;s D E.Xt Xs / D E.B1t B1s / D min.1 t; 1 s/ D 1 t :
1
b) As, for u s t, Kt;s Ks;s Ks;u D .1 t/.1 s/1 .1 s/ D 1 t D Kt;u , the
Markovianity condition (6.32) is satisfied, hence .Xt /t is Markovian with respect
to its natural filtration. Its transition function p.s; t; x; / is Gaussian with mean
and variance given respectively by (4.21) and (4.22), i.e. with mean
1 1t
Kt;s Ks;s xD x
1s
and variance
1 .1 t/2 1t
Kt;t Kt;s Ks;s Ks;t D .1 t/ D .t s/ :
1s 1s
• The transition function above is the same as that of the Brownian bridge (see
Exercise 6.5). The initial distribution is different, as here it is the law of B1 , i.e.
N.0; 1/.
6.7
a) In order to prove the existence of a continuous version we use Kolmogorov’s
Theorem 2.1. Let us assume that the process starts at time u with initial
distribution . Recalling (6.6) which gives the joint law of .Xs ; Xt /, we have
for u s t,
Z Z Z
ˇ
EŒjXt Xs j D .dz/ p.u; s; z; dx/ jx yjˇ p.s; t; x; dy/
Rm Rm Rm
Z Z
c jt sjmC" .dz/ p.u; s; z; dx/ D cjt sjmC" :
Rm Rm
Then by Theorem 2.1, X has a continuous version (actually with Hölder

continuous paths of exponent for every < ˇ" ). In order to prove that the
generator is local, let us verify condition (6.26): as jy xjˇ > Rˇ for y 62 BR .x/,
Z
1 1
p.s; s C h; x; BR .x/c / D p.s; s C h; x; dy/
h h BR .x/c
Z
Rˇ
jy xjˇ p.s; s C h; x; dy/ cRˇ jhjmC"1 ;
h Rm
which tends to 0 as h ! 0.
6.8
a) If f D 1A we have, recalling that 1Ax .y/ D 1A .x C y/,
Z Z
1A .y/ p.t; x; dy/ D p.t; x; A/ D p.t; 0; A x/ D 1Ax .y/ p.t; 0; dy/
Z
1A .x C y/ p.t; 0; dy/
so that (6.34) is satisfied if f is an indicator function. It is then also true

for linear combinations of indicator functions and by the usual approximation
methods (Proposition 1.11, for example) for every bounded Borel function
f.
b) Recall that p.h; x; / is the law of Xh under Px , whereas p.h; 0; / is the law of Xh
under P0 . Therefore, if we denote by tx f the “translated” function y 7! f .x C y/,
thanks to a) we have
Z Z
E Œ f .Xh / D f .y/ p.h; x; dy/ D f .x C y/ p.h; 0; dy/ D E0 Œtx f .Xh /
x
for every bounded Borel function f . If f 2 CK2 then
1 1 x
Lf .x/ D lim Th f .x/ f .x/ D lim E Œ f .Xh / f .x/
h!0Ch h!0C h
1 0
D lim E Œtx f .Xh / tx f .0/ D L.tx f /.0/ :
h!0C h
As we have
@f @.tx f / @2 f @2 .tx f /
.x/ D .0/; .x/ D .0/ ;
@yi @yi @y2i @y2i
equating the two expressions
X
m
@2 f X
n
@f
Lf .x/ D aij .x/ 2
.x/ C bi .x/ .x/
i;jD1
@xi iD1
@xi
X
m
@2 f X
n
@f
L.tx f /.0/ D aij .0/ 2
.x/ C bi .0/ .x/ ;
i;jD1
@xi iD1
@x i
which must be equal for every choice of f 2 CK2 , we have aij .x/ D aij .0/; bi .x/ D
bi .0/, for every i; j m.
6.9
a) ph obviously satisfies condition i) on p. 151. Moreover, ph .t; x; / is a measure on
E. As
Z
e˛t e˛t
p .t; x; E/ D
h
h.y/ p.t; x; dy/ D Tt h.x/ D 1 ;
h.x/ E h.x/
ph is also a probability. We still have to check the Chapman–Kolmogorov

equation. Thanks to the Chapman–Kolmogorov equation for p,
Z Z Z
e˛t e˛s
p .s; y; A/ p .t; x; dy/ D
h h
h.y/p.t; x; dy/ h.z/ p.s; y; dz/
E h.x/ E h.y/ A
Z
e˛.tCs/
D h.z/ p.t C s; x; dz/ D ph .s C t; x; A/ :
h.x/ A
b) We have
Z
e˛t e˛t
Tth g.x/ D h.y/g.y/p.h; x; dy/ D Tt .hg/.x/ :
h.x/ h.x/
Therefore, if gh D f 2 D.L/,
1 1 1
Lh g.x/ D lim ŒTth g.x/ g.x/ D lim Œe˛t Tt f .x/ f .x/
t!0C t h.x/ t!0C t
1 1 1
D lim .Tt f .x/ f .x// C lim .e˛t 1/Tt f .x/
h.x/ t!0C t t!0C t
1 1
D .Lf .x/ ˛f .x// D L.gh/.x/ ˛g.x/ :
h.x/ h.x/
(S.36)
c) If h 2 C2 and g 2 CK2 , then also gh 2 CK2 and
X
m
@g @h
L.gh/ D h Lg C g Lh C aij
i;jD1
@xi @xj
From (6.35) we have Tt h D e˛t h, hence h 2 D.L/ and Lh D ˛h so that, thanks

to the relation between L and Lh given in (S.36),
1X X
m m
@2 e @
Lh D aij C bi ,
2 i;jD1 @xi @xj iD1
@xi
where
1 X
m
e @h
bi .x/ D b.x/ C aij .x/
h.x/ jD1 @xj
d) Let us check that h.x/ D ehv;xi satisfies (6.35). We have

Z Z
1 1 2
Tt h.x/ D h.y/ p.t; x; dy/ D ehv;yi e 2t jxyj dy :
Rm .2t/m=2 Rm
We recognize in the rightmost integral the Laplace transform of an N.x; tI/ law
computed at v. Hence (see Exercise 1.6)
1 2
Tt h.x/ D ehv;xi e 2 tjvj :
(6.35) is therefore verified with ˛ D 12 jvj2 . Thanks to b), therefore,
1 X @2 X @
m m
Lh D 2
C vi
2 iD1 @xi iD1
@xi
6.10
a) If f is a bounded Borel function and is stationary, we have
Z Z Z Z
f .x/ .dx/ D Tt f .x/ .dx/ D .dx/ f .y/ p.t; x; dy/
E E E E
R
and now just observe that E .dx/p.t; x; / is the law of Xt , when the initial
distribution is .
b) This is a consequence of Fubini’s theorem: if f is a bounded Borel function with

compact support, then
Z Z Z
1 1 2
Tt f .x/ dx D dx e 2t jxyj f .y/ dy
Rm .2t/m=2 Rm Rm
Z Z Z
1 1 2
D f .y/ dy e 2t jxyj dx D f .y/ dy :
Rm Rm .2t/m=2 Rm
„ ƒ‚ …
D1
c) Let us assume that (6.38) holds for every x; if an invariant probability existed,
then we would have, for every t > 0 and every bounded Borel set A,
Z
.A/ D p.t; x; A/ d.x/ :
E
As 0 p.t; x; A/ 1 and limt!C1 p.t; x; A/ D 0, we can apply Lebesgue’s

theorem and obtain
Z
.A/ D lim p.t; x; A/ d.x/ D 0
t!C1 E
so that .A/ D 0 for every bounded Borel set A in contradiction with the
hypothesis that is a probability.
If p is the transition function of a Brownian motion then, for every Borel set
A Rm having finite Lebesgue measure,
Z
1 1 2 1
p.t; x; A/ D e 2t jxyj dy mis.A/ ! 0:
.2t/m=2 A .2t/m=2 t!C1
Therefore a Brownian motion cannot have an invariant distribution.

d) Let f W E ! R be a bounded continuous function. Then (6.39) implies
Z Z
lim Ts f .x/ D lim f .y/ p.s; x; dy/ D f .y/ d.y/ : (S.37)
s!C1 s!C1 E E
By the Feller property Tt f is also bounded and continuous and, for every x 2 E,
Z
lim Ts f .x/ D lim TsCt f .x/ D lim Ts .Tt f /.x/ D Tt f .y/ .dy/ :
s!C1 s!C1 s!C1 E
(S.38)
(S.37) and (S.38) together imply that the stationarity condition (6.37) is satisfied
for every bounded continuous function. It is also satisfied for every bounded
measurable function f thanks to the usual measure theoretic arguments as in
Theorem 1.5, thus completing the proof that is stationary.
6.11
a) Let f W G ! R be a bounded measurable function. Then we have for s t,
thanks to the Markov property for the process X,
Z
EŒ f .Yt /jFs D EŒ f ı ˚.Xt /jFs D f ı ˚.z/ p.s; t; Xs ; dz/
E
Z Z
D f ı ˚.z/ p.s; t; ˚ 1 .Ys /; dz/ D f .y/ q.s; t; Ys ; dy/ ;
E G
where we denote by q.s; t; y; / the image law of p.s; t; ˚ 1 .y/; / through the
transformation ˚. This proves simultaneously that q is a transition function
(thanks to Remark 6.2) and that Y satisfies the Markov property.
b1) This is simply the integration rule with respect to an image probability
(Proposition 1.1).
b2) If f W G ! R is a bounded measurable function and s t we have, thanks to
the Markov property of X,
Z Z
EŒ f .Ys /j Fs D EŒ f ı˚.Xs /j Fs D f ı˚.z/ p.s; t; Xs ; dz/D f .y/q.s; t; Ys ; dy/;
E E
which proves simultaneously the Markov property for Y and the fact that q is a
transition function (Remark 6.2 again).
c) We must show that the invariance property (6.40) is satisfied when X is a
Brownian motion and ˚.x/ D jx zj. This is rather intuitive by the property of
rotational invariance of the transition function of the Brownian motion. Let
us give a rigorous form to this intuition. Using (6.42), which is immediate
as p.s; x; dy/ is Gaussian with mean x and variance s, we must show that, if
jx zj D jy zj then
p p
P. sZ 2 A x/ D P. sZ 2 A y/
...........................................
..........
. . . . . .................... . . .
. ............................................................................... .
....... ....... . ............................................................................................................ .

. .................................................................................................................................
...... ......
.................................................................................................................................................... .
............................... .
..... ..................
............................................................................................................................................................................................
........ ..........
................................................. . . . . . . ........................................
......................................................... .
.... .....................................................
. ................................... .
..................................
... ......... ...... .

.................................
.............................................................
..... ........
...........................................
..
.. .
..........................................
...........................
............................
..........................
... ........
...........................
......................................
..... .........................
.... .....
......................................
..
. . .. .
...........................................
........................
.................................
... ....
. . . . .
. . . . . . . . . . . .................................
................................. .. . . . . . . . . .
.............................................
... ... .............................................
.. .... .
......................
.
. . . . . . . . . . .
... ..
..................... ....................
....................
. ............................
.
.......................................
... ...
....................
... ... . .............................
•y
. . ............... ..........................
...........................
. ... ... .............................
... .... .
..................................... ............................
......... ..........................
............................
. ... ... ............................
.... ...
.....................................
.........................
...................................
. ... ...
........
....................................
.........
•x
....................................
... ...
.........
........................... .
................................... . ... ... ...........................
........
....................................
... .............................
. . ... ... .........................
.....................
.................................
... ...
.........................
........
. . .. .. •z ....................................
.................
...................
....................................
... ...
.....................................
... ...
....................... .........................
...................................
... ...
.... . . . . . . .
....................................
.................................. . .. ...
. ........
..........................
... ...
..........................
. . .
. ... .....................................
.
.
................... ..........................
... ...
........
....................................
. .
. .
.............................
...................................
. ... ..
...................
.................. . . ..............
... ...
..................................... ............................
. ... ...
............................
..........................
... ... ............................ . ..............................
... ..
.....................................
... ....
............................
.
............................. . .......
................................
. . . ... .....
............................. . ............................
... ..... .......................
.................................
. . . . ..
.. ..............................................
.... ..... ...
...................... ...........................................
........................
.
..... ........
. . . . . . . . . . . . . . . . . . . ..
.
............................................. .................................
..... ..
..... ............. ......
......................... ....................................
...................................... ............................
.
....... ..........
........................... .....................................
......................................... .......................................................
...... ...........
................................
............................... . . . . . . . . . . . . . . . . .
............................................... . .. . . . . . . . . . . . . . . .
. .................................................
...... . .....
....................................
.................................... . .
.......................................... . . . . . .......................................................................... .
........ . ...
. ...................................................................................................................................................................
....
........................................................................................................ .
...........
................................................................................................ .
........................................
................................................................................................ .
. .................................................................................................................. .
. ................................................................................... .
. . ..................................... . . .
. . . . . .. . .
Fig. S.5 Thanks to the property of rotational invariance, the probability of making a transition into
the shaded area is the same starting from x or y, or from whatever other point on the same sphere
centered at z
for every A of the form A D ˚ 1 .e A/, e

A 2 B.RC /. Such a subset of Rm is
some kind of annulus around z (see Fig. S.5). In any case the set A z is clearly
rotationally invariant. Let O be an orthogonal matrix such that O.x z/ D
y z (the two vectors x z and y z have the same modulus). As the N.0; I/-
distributed r.v.’s are rotationally invariant,
p p
P. sZ 2 A x/ D P. sZ 2 A z .x z//
p p p
D P. sZ 2 O.Az.xz/// D P. sZ 2 Az.yz// D P. sZ 2 Ay/ :
7.1
a) We have EŒZ D 0, as Z is a stochastic integral with a bounded (hence belonging
to M 2 ) integrand. We also have, by Fubini’s theorem,
h Z 1 2 i hZ 1 i Z 1
EŒZ 2 D E 1fBt D0g dBt DE 1fBt D0g dt D P.Bt D 0/ dt D 0 :
0 0 0
Z, having both mean and variance equal to 0, is equal to 0 a.s.

b) EŒZ D 0, as the integrand is in M 2 . As for the variance we have
h Z 1 2 i hZ 1 i Z 1
1
EŒZ 2 D E 1fBt 0g dBt DE 1fBt 0g dt D P.Bt 0/ dt D :
0 0 0 2
Rt
7.2 Let s t. We have Bs D 0 1Œ0;sŒ .v/ dBv therefore, by Remark 7.1,
Z t Z t Z t Z t
E Bs Bu dBu D E 1Œ0;sŒ .v/ dBv Bu dBu D EŒ1Œ0;s .u/Bu du D 0 :
0 0 0 0
If t s, the same argument leads to the same result.

7.3 We have
Z t
1 2t
E.Yt2 / D e2s ds D .e 1/
0 2
and
hZ t i Z t Z t
1 2s 1 t
E Ys2 ds D EŒYs2 ds D .e 1/ ds D .e2t 1/ < C1 :
0 0 0 2 4 2
Hence Y 2 M 2 so that E.Zt / D 0 and

hZ t i 1 t
EŒZt2 D E Ys2 ds D .e2t 1/
0 4 2
Finally, using the argument of Example 7.1 c), if s t,

hZ t Z s i h Z s 2 i 1 s
EŒZt Zs D E Yu dBu Yv dBv D E Yu dBu D .e2s 1/
0 0 0 4 2
• In this and the following exercises we skip, when checking that a process is in M 2 ,
the verification that it is progressively measurable. This fact is actually almost
always obvious, thanks to Proposition 2.1 (an adapted right-continuous process
is progressively measurable) or to the criterion of Exercise 2.3.
7.4
a) We have, by Theorems 7.1 and 7.6,
h Z t 2 i h Z t 2 i hZ t i
E B2s Bu dBu DE Bs Bu dBu DE B2s B2u du :
s s s
Recalling that s u so that Bu Bs is independent of Bs ,
E.B2s B2u / D EŒB2s .Bu Bs CBs /2 D EŒB2s .Bu Bs /2 C2 EŒB3s .Bu Bs / CE.B4s / :
„ ƒ‚ … „ ƒ‚ …
Ds.us/ D0
p
We can write Bs D sZ where Z N.0; 1/ and therefore E.B4s / D s2 E.Z 4 / D
3s2 . In conclusion, E.B2s B2u / D s.u s/ C 3s2 and
h Z t 2 i Z t
1
E B2s Bu dBu D s.u s/ C 3s2 du D s.t s/2 C 3s2 .t s/ :
s s 2
Rt
b) One can write Z D 0 e X u dBu , where e
X u D Xu if u s and e
X u D 0 if 0 u < s.
Rt
e 2
It is immediate that X 2 M .Œ0; t/. As Bv D 0 1Œ0;v .u/ dBu and the product
e
X u 1Œ0;v .u/ vanishes for every 0 u t,
Z t Z t Z t
E.ZBv / D E e
X u dBu 1Œ0;v .u/ dBu D E e
X u 1Œ0;v .u/ du D 0 D E.Z/E.Bv / :
0 0 0
If Z was independent of Fs it would be EŒ.Z/W D EŒ.Z/EŒW for every

positive Borel function and every positive Fs -measurable r.v. W. Let us prove
that this is not true, in general,
R t with a counterexample. Using the computation of
a), let .Z/ D Z 2 with Z D s Bu dBu and W D B2s . Then
Z t
1 2
E.Z 2 / D EŒB2u du D .t s2 / ;
s 2
hence EŒZ 2 EŒB2s D s

2
.t2 s2 /, which is different from the value of E.B2s Z 2 /
computed in a).
2
7.5 We must check that s 7! eBs is a process in M 2 . For every t 0 and recalling
Remark 3.3,
hZ t i Z t Z t
1
2B2s 2B2s
E e ds D EŒe ds D p ds < C1 :
0 0 0 1 C 4s
In order to investigate whether X is bounded in L2 , just observe that

hZ t i Z t
1
2B2s
EŒXt2 DE e ds D p ds ! C1
0 0 1 C 4s t!C1
so that the answer is no.

7.6
a) Note that .˝; F ; .Ft /t ; .B2 .t//t ; P/ is a real Brownian motion and that B1 is
progressively measurable with respect to .Ft /t and also B1 2 MB2 2 .Œ0; t/ so that
we can define
Z t
2
Xt D eB1 .u/ dB2 .u/ :
0
Note also that .˝; F ; .Ft /t ; .B1 .t//t ; P/ is a real Brownian motion and that X is
progressively measurable with respect to .Ft /t . Moreover, for every t 0,
hZ t i Z t
1 1 p
2B1 .u/2
EŒXt2 DE e du D p du D 1 C 4t1 : (S.39)
0 0 1 C 4u 2
As we can write
Z t
Xs
Zt D dB1 .s/ (S.40)
0 1 C 4s
and
h i p
Xs2 1 1 C 4s 1
E 2
D ,
.1 C 4s/ 2 .1 C 4s/2
the integrand s 7! 1C4s

Xs
is in MB2 1 .Œ0; t/ for every t > 0 and Z is a martingale.
Using the notation (S.40) and thanks to Theorem 7.3 and Remark 7.1
Z Z
t
Xs2 t
Xs
hZit D ds; hZ; B1 it D ds :
0 .1 C 4s/2 0 1 C 4s
b) It is sufficient (and also necessary. . . ) to prove that Z is bounded in L2 . Now,

recalling (S.39),
Z Z t
t
EŒXs2 1 1 1
EŒZt2 D 2
ds D ds :
0 .1 C 4s/ 2 0 .1 C 4s/3=2 .1 C 4s/2
The rightmost integral being convergent as t ! C1, the L2 norm of Z is

bounded, so that Z converges as t ! C1 a.s. and in L2 . As L2 convergence
implies the convergence of the expectations we have EŒZ1 D 0. Also L2
convergence implies the convergence of the second-order moment, so that
Z C1
2 1 11
EŒZ1 D lim EŒZt2 D ds
t!C1 0 .1 C 4s/3=2 .1 C 4s/2
2
1 1 ˇC1 1 1 ˇC1 1
ˇ ˇ
D ˇ C ˇ D
4 .1 C 4s/1=2 0 8 1 C 4s 0 8
7.7
a) If f 2 L2 .Œs; t/
P is a piecewise constant function the statement is immediate.
Indeed, if f D niD1 i 1Œti1 ;ti Œ with s D t1 < < tn D t, then
Z t X
n
f .u/ dBu D i .Bti Bti1 /
s iD1
and all the increments Bti Bti1 are independent of Fs . In general, if .fn /n is a
sequence of piecewise constant functions (that are dense in L2 .Œs; t/) converging
to f in L2 .Œs; t/, then, by the isometry property of the stochastic integral,
Z t Z t
def L2 def
Bfn D fn .u/ dBu ! f .u/ dBu D Bf
s n!1 s
and, possibly taking a subsequence, we can assume that the convergence also
takes place a.s. We must prove that, for every bounded Fs -measurable r.v. W and
for every bounded Borel function ,
EŒW .Bf / D EŒWEŒ .Bf / : (S.41)
But, if is bounded continuous, then .Bfn / ! .Bf / a.s. as n ! 1 and we

can take the limit of both sides of the relation
EŒW .Bfn / D EŒWEŒ .Bfn / :
(S.41) is therefore proved for bounded continuous. We attain the general case
with Theorem 1.5.
Alternatively, in a simpler but essentially similar way, we could have used the
criterion of Exercise 4.5: for every 2 R we have
E.ei Bf jFs / D lim E.ei Bfn jFs / D lim E.ei Bfn / D E.ei Bf / ;
n!1 n!1
where the limits are justified by Lebesgue’s theorem.

• Note that the previous proof would be much simpler if we only had to prove
that Bf is independent of Gs D .Bv ; v s/. In this case, by the criterion of
Remark 1.2, it is sufficient to check that Bf and Bv are not correlated for every
v s, which is immediate: if we denote by e f the extension of f to Œ0; t obtained
by setting f .v/ D 0 on Œ0; sŒ, then
Z t Z t Z t
E.Bv Bf / D E 1Œ0;v dBu e
f .u/ dBu D 1Œ0;v .u/e
f .u/ du D 0 :
0 0 0
b) The r.v. B˚
t is F˚ 1 .t/ -measurable. This suggests to try to see whether B is a
˚
Brownian motion with respect to .F e t D F˚ 1 .t/ . If s t, then

e t /t , with F
Z ˚ 1 .t/ p
B˚
t B˚
s D ˚ 0 .u/ dBu
˚ 1 .s/
is independent of F e s thanks to the result of a). We still have to prove c) of

Definition 3.1. But B˚ ˚
t Bs is Gaussian by Proposition 7.1, is centered and has
variance
Z ˚ 1 .t/
EŒ.B˚ ˚ 2
t Bs / D ˚ 0 .u/ du D t s :
˚ 1 .s/
Rs
7.8 Let us assume first that f D 1Œ0;s ; therefore 0 f .u/ dBu D Bs . Note that the
r.v.’s Bs B2u and Bs B2u have the same law (B is also a Brownian motion) so that
EŒBs B2u D 0 and we have
Z t Z t
2
E Bs Bu du D E.Bs B2u / du D 0 :
0 0
Rt
In conclusion R 0 B2u du is orthogonal to all the r.v.’s Bs ; s > 0. It is therefore
orthogonal to 0 f .u/ dBu , when f is a piecewise constant function of L2 .Œ0; s/, as
s
in this case the stochastic integral is a linear combination of the r.v.’s Bu ; 0 u s.

The conclusion now follows by density.
7.9
a) We have
X
n1 X
n1 X
n1
BtiC1 .BtiC1 Bti / D Bti .BtiC1 Bti / C .BtiC1 Bti /2 :
iD0 iD0 iD0
The first term on the right-hand side converges, in probability, to the integral
Z t
Bs dBs
0
as the amplitude of the partition tends to 0 thanks to Proposition 7.4. Conversely,

by Proposition 3.4,
X
n1
.BtiC1 Bti /2 ! t
jj!0
iD0
in L2 , hence in probability. Therefore, in probability,
X
n1 Z t
BtiC1 .BtiC1 Bti / ! tC Bs dBs :
jj!0 0
iD0
b) We have again
X
n1 X
n1 X
n1
XtiC1 .BtiC1 Bti / D Xti .BtiC1 Bti / C .XtiC1 Xti /.BtiC1 Bti / :
iD0 iD0 iD0
Now, if X has paths with finite variation, we have a.s.
ˇXn1 ˇ X
n1
ˇ ˇ
ˇ .XtiC1 Xti /.BtiC1 Bti /ˇ sup jBtiC1 Bti j jXtiC1 Xti j
iD0 iD0;:::;n1 iD0
C sup jBtiC1 Bti j ! 0

iD0;:::;n1 jj!0
thanks to the continuity of B, where C D C.!/ is the variation of t 7! Xt in

the interval Œ0; t. Therefore the sums (7.38) and (7.39) have the same limit in
probability.
7.10
Rt
a) We know already that Xt D 0 f .s/ dBs is a martingale, as f 2 M 2 .Œ0; T/ for
every T. As, for every t 0,
Z t
EŒXt2 D f .s/2 ds kf k22 ;
0
.Xt /t is a martingale bounded in L2 and converges a.s. and in L2 .

b1) ThereR are various ways of showing that .Yt /t is a martingale. For instance, as
t
t 7! 0 g.s/ dBs is a Gaussian martingale with independent increments and its
Rt
associated increasing process is At D 0 g.s/2 ds, we have already seen this in
Exercise 5.11 c2). A direct computation is also simple:
Rt R
1 t 2 ˇ
R
1 t 2 Rt ˇ
EŒYt jFs D E e 0 g.u/ dBu 2 0 g .u/ du ˇ Fs D e 2 0 g .u/ du E e 0 g.u/ dBu ˇ Fs

Rs R
1 t 2 Rt ˇ
D e 0 g.u/ dBu 2 0 g .u/ du E e s g.u/ dBu ˇ Fs

Rs R
1 t 2 Rt
D e 0 g.u/ dBu 2 0 g .u/ du E e s g.u/ dBu

Rs Rt Rt
g.u/ dBu 12 g2 .u/ duC 12 g2 .u/ du
De 0 0 s D Ys :
Rt
As the r.v. 0 g.s/ dBs is Gaussian, we have
Rt
Rt 2 Rt Rt 2
Rt 2
E.Yt2 / D E e2 0 g.s/ dBs e 0 g.s/ ds D e2 0 g.s/ dBs e 0 g.s/ ds D e 0 g.s/ ds :
(S.42)
b2) If g 2 L2 .RC /, then by (S.42) .Yt /t is a martingale bounded in L2 and therefore

uniformly integrable.
b3)
Rt
˛ Rt 2 1
Rt 2
E.Yt˛ / D E e˛ 0 g.s/ dBs e 2 0 g.s/ ds D e 2 ˛.˛1/ 0 g.s/ ds ! 0:
t!C1
Therefore, defining Y1 D limt!C1 Yt (the a.s. limit exists because Y is a

continuous positive martingale), by Fatou’s lemma
˛
E.Y1 / lim E.Yt˛ / D 0 :
t!C1
˛ ˛
As Y1 0, this implies Y1 D 0 and therefore Y1 D 0. If .Yt /t
was a uniformly integrable martingale, then it would also converge in L1
and this would imply E.Y1 / D 1. Therefore .Yt /t is not uniformly inte-
grable
7.11
a1) Gaussianity is a consequence of Proposition 7.1. Moreover, for s t,
1 hZ
Z t
1 s i
E.Ys Yt / D .1 t/.1 s/E dBu dBv
0 1u 0 1v
Z
s
1 1
D .1 t/.1 s/ du D .1 t/.1 s/ 1 D s.1 t/ ;
0 .1 u/2 1s
i.e. the same covariance function as a Brownian bridge.

a2) From a1) for s D t, E.Yt2 / D t.1 t/ ! 0 for t ! 1 hence Yt ! 0 in L2 as
t ! 1.
b) .Ws /s is obviously a Gaussian continuous process. Moreover, observe that t 7!
A.t/ is increasing, therefore, for s t,
hZ A.s/
dBu
Z A.t/
dBv i
Z A.s/
1
E.Ws Wt / D E D du
0 1u 0 1v 0 .1 u/2
1
D 1D 1Cs1 DsD s^t:
1 A.s/
Therefore .Ws /s is a (natural) Brownian motion. The inverse of s 7! A.s/ is

t 7! .t/ D 1t
t
, i.e. A..t// D t. We have therefore
Z t
dBu
Yt D .1 t/ D .1 t/W .t/
0 1u
and
p W .t/
lim Yt D lim .1t/W .t/ D lim .1t/ 2.t/ log log .t/ p
t!1 t!1 t!1 2.t/ log log .t/
In order to conclude the proof, we simply recall that .t/ ! C1 as t ! 1

and therefore, by the Iterated Logarithm Law,
W .t/ W .t/
lim p D 1; lim p D 1 a.s.
t!1 2.t/ log log .t/ t!1 2.t/ log log .t/
whereas
p p
lim .1 t/ 2.t/ log log .t/ D lim 2t.1 t/ log log .t/ D 0
t!1 t!1
hence limt!1 Yt D 0 a.s.

7.12
a) Yt and Zt are both Gaussian r.v.’s, as they are stochastic integrals with a
deterministic square integrable integrand (Proposition 7.1). Both have mean
equal to 0. As for the variance,
h Z t 2 i Z t Z t
1
Var.Yt / D E e .ts/ dBs D e2 .ts/ ds D e2 u du D .1e2 t /
0 0 0 2
and in the same way
h Z t 2 i Z t
1
Var.Zt / D E e s dBs D e2 s ds D .1 e2 t / :
0 0 2
1
Both Yt and Zt have a Gaussian law with mean 0 and variance 2 .1 e2 t /.
b) .Zt /t is a martingale (Proposition 7.3). For .Yt /t , if s t, we have instead
hZ t ˇ i
Z s
EŒYt jFs D e t E e u dBu ˇ Fs D e t e u dBu D e .ts/ Ys :
0 0
Therefore .Yt /t is not a martingale.

c) We have
1
sup EŒZt2 D lim EŒZt2 D < C1 :
t0 t0 2
.Zt /t is therefore a martingale bounded in L2 and converges a.s. and in L2 .

d1) Obviously .Yt /t converges in law as t ! C1, as Yt and Zt have the same law
for every t and .Zt /t converges a.s. and therefore in law.
d2) We have
EŒ.YtCh Yt /2 D EŒYtCh
2
C EŒYt2 2EŒYt YtCh :
The two first terms on the right-hand side have already been computed in a).
As for the last, conversely,
Z t Z tCh Z t
.tCh/ t s s .2tCh/
EŒYt YtCh D e e E e dBs e dBs D e e2 s ds
0 0 0
1 h
D .e e .2tCh/ / :
2
Therefore, putting the pieces together,
1
lim EŒ.YtCh Yt /2 D lim 1e2 t C1e2 .tCh/ 2e h C2e .2tCh/
t!C1 t!C1 2
1
D .1 e h / :

.Yt /t is not therefore Cauchy in L2 as t ! C1, and cannot converge in L2 . We

shall see in Sect. 9.2 that .Yt /t is an Ornstein–Uhlenbeck process.
7.13
a) e
B is a Gaussian process by Proposition 7.1. In order to take advantage of
Proposition 3.1 let us show that E.e
Bse
Bt / D s ^ t. Let us assume s t. As
hZ t 12u 10u2
Z s
12u 10u2 i
EŒe
Bte
Bs D E 3 C 2 dBu 3 C 2 dBu
0 t t 0 s s
Z s 2 2
12u 10u 12u 10u
D 3 C 2 3 C 2 du
0 t t s s
Z s
36u 30u2 36u 144u2 120u3 30u2 120u3 100u4
D 9 C 2 C 2 C 2 2 C 2 2 du
0 s s t st st t t s s t
we have
2 2 2 3 3 3
s s s s s s
EŒe
Bte
Bs D 9s 18s C 10s 18 C 48 30 C 10 2 30 2 C 20 2 D s :
t t t t t t
b) As Y and eBt are jointly Gaussian, we must just check that they are uncorrelated.
Let t 1: we have
hZ 1 Z t
12u 10u2 i
EŒYe
Bt D E u dBu 3 C 2 dBu
0 0 t t
Z
h t Z t
12u 10u2 i
DE u dBu 3 C 2 dBu
0 0 t t
Z t
12u 10u2 3 10 2
D u 3 C 2 du D t2 4t2 C t D0:
0 t t 2 4
The proof for t > 1 is straightforward and is left to the reader.

e1 D .e
c) Let G Bt ; t 0/ and G1 D .Bt ; t 0/. Obviously G e1 G1 as
every r.v. e
Bt is G1 -measurable. On the other hand the r.v. Y defined in b) is G1 -
measurable but not G e1 -measurable. If it were we would have
e1 D Y
EŒY j G
whereas, conversely,
e1 D E.Y/ D 0
EŒY j G
as Y is independent of e e1 (Remark 1.2).

Bt for every t and therefore also of G
7.14
a) We have
Z T Z n
1fs<n g Xs2 ds D Xs2 ds n ;
0 0
hence the process .Xs 1fs<n g /s belongs M 2 .Œ0; T/ and

hZ n i hZ T i h Z T 2 i
E Xs2 ds D E 1fs<n g Xs2 ds DE 1fs<n g Xs dBs
0 0 0
h Z n 2 i h i
2
DE Xs dBs D E.Mt^n
/E sup Mt2 :
0 0tT
b) Assume (7.40): as n % T a.s., by Beppo Levi’s theorem,

h i hZ n i hZ T i
E sup Mt2 lim E Xs2 ds D E Xs2 ds ;
0tT n!1 0 0
hence X 2 M 2 .Œ0; T/. Conversely, if X 2 M 2 .Œ0; T/, .Mt /0tT is a square

integrable martingale bounded in L2 and (7.40) follows from Doob’s inequality
(the second one of the relations (7.23)).
7.15 Let .n /n be a sequence reducing the local martingale M. Then, for s t,
EŒMt^n jFs D Ms^n : (S.43)
As n % C1 as n ! 1, we have Mt^n ! Mt and Ms^n ! Ms as n ! 1.

Moreover, as the sequence .Mt^n /n is assumed to be uniformly integrable, the
convergence for n ! 1 also takes place in L1 and we can pass to the limit in (S.43)
and obtain that Mt is integrable for every t 0 and also the martingale relation
EŒMt jFs D Ms :
8.1
a) The computation of the stochastic differential of .Mt /t is of course an application
of Ito’s formula, but this can be done in many ways. The first that comes to mind
is to write Mt D f .Bt ; t/ with

1
f .x; t/ D .x C t/e.xC 2 t/ :
But one can also write Mt D Xt Yt , where
Xt D Bt C t
1
Yt D e.Bt C 2 t/ :
By Ito’s formula, considering that Yt D f .Bt C 12 t/ with f .x/ D ex , we have
1
1 1 1
dYt D e.Bt C 2 t/ dBt C dt C e.Bt C 2 t/ dt D Yt dBt ;
2 2
whereas obviously dXt D dBt C dt. Then, by Ito’s formula for the product, as
dhX; Yit D Yt dt,
dMt D d.Xt Yt / D Xt dYt C Yt dXt C dhX; Yit

D .Bt C t/Yt dBt C Yt dBt C Yt dt Yt dt
1
D .Bt C t C 1/e.Bt C 2 t/ dBt :
b) Thanks to a) .Mt /t is a local martingale. It is actually a martingale, as the process

1
t 7! .Bt C t 1/e.Bt C 2 t/ is in M 2 .
In order to show that it is a martingale we could also have argued without
computing the stochastic differential with the usual method of factoring out the
Brownian increments: if s t
1 ˇ
EŒMt jFs D E .Bs C .Bt Bs / C t/e.Bs C.Bt Bs /C 2 t/ ˇ Fs

1 1 ˇ
D E .Bs C t/e.Bs C 2 t/ e.Bt Bs / C .Bt Bs /e.Bs C 2 t/ e.Bt Bs / ˇ Fs

1
1
D .Bs C t/e.Bs C 2 t/ E e.Bt Bs / C e.Bs C 2 t/ E .Bt Bs /e.Bt Bs /

1 1 1 1
e.Bs C 2 t/ .Bs C t/e 2 .ts/ .t s/e 2 .ts/ D .Bs C s/e.Bs C 2 s/ D Ms :
This old method appears to be a bit more complicated.
8.2 It is convenient to write Xt D ab C .b a/Bt B2t C t, so that, recalling that

dB2t D 2Bt dBt C dt,
dXt D .b a/ dBt 2Bt dBt dt C dt D .b a 2Bt / dBt ;

i.e.
Z t
Xt D ab C .b a 2Bt / dBt :
0
The integrand obviously being in M 2 , X is a square integrable martingale.

8.3
a) Again Ito’s formula can be applied in many ways. For instance, if u.x; t/ D
et=2 sin x,
@u @u 1 @2 u
dXt D du.Bt ; t/ D .Bt ; t/ dt C .Bt ; t/ dBt C .Bt ; t/ dt
@t @x 2 @x2
and as
@u 1 @u @2 u
.x; t/ D et=2 sin x; .x; t/ D et=2 cos x; .x; t/ D et=2 sin x
@t 2 @x @x2
we find
1 1 t=2
dXt D et=2 sin Bt e sin Bt dt C et=2 cos Bt dBt D et=2 cos Bt dBt
2 2
and therefore X is a local martingale and even a martingale, being bounded on
bounded intervals. For Y the same computation with u.x; t/ D et=2 cos x gives
dYt D et=2 sin Bt dBt
so that Y is also a martingale. As the expectation of a martingale is constant we

have E.Xt / D 0, E.Yt / D 1 for every t 0.
More quickly the imaginative reader might have observed that
1
Yt C iXt D eiBt C 2 t
which is known to be an exponential complex martingale, hence both its real and
imaginary parts are martingales.
b) The computation of the stochastic differentials of a) implies that both X and Y
are Ito processes. It is immediate that
Z t
hX; Yit D es cos Bs sin Bs ds :
0
8.4
a) Just take the stochastic differential of the right-hand side: by Ito’s formula
1 2 1 2 1 2 1 1 2 1 2
de Bt 2 t D e Bt 2 t .dBt dt/ C e Bt 2 t 2 dt D e Bt 2 t dBt :
2 2
Hence the left and right-hand sides in (8.52) have the same differential. As they
both vanish at t D 0 they coincide.
2
b) Recall that if X 2 Mloc and
Rt Rt
Xu dBu 12 Xu2 du
Yt D e 0 0
then Y is a local martingale and satisfies the relation
dYt D Xt Yt dBt
and therefore
Z t
Xs Ys dBs D Yt 1 :
0
In our case Xt D Bt , hence

Z t Rs Rs Rt Rt
Bu dBu 12 B2u du Bu dBu 12 B2u du
e 0 0 Bs dBs D e 0 0 1:
0
8.5
a) If f .x; t/ D x3 3tx, by Ito’s formula,
@f @f 1 @2 f
df .Bt ; t/ D .Bt ; t/ dBt C .Bt ; t/ dt C .Bt ; t/ dt
@x @t 2 @x2
@f 1 @2 f @f
D .Bt ; t/ dBt C C .Bt ; t/ dt :
@x 2 @x2 @t
1 @2 f @f
It is immediate that 2 @x2 .x; t/ C @t .x; t/ D 3x 3x D 0; therefore, as f .0; 0/ D 0,
Z Z
t
@f t
f .Bt ; t/ D .Bs ; s/ dBs D .3B2s 3s/ dBs :
0 @x 0
The integrand belongs to M 2 .Œ0; C1Œ/ (it is a polynomial in t and Bt ) and

therefore Xt D f .Bt ; t/ is a (square integrable) martingale.
b) The computation above shows that Xt D Pn .Bt ; t/ is a martingale if and only if
1 @2 Pn @Pn
C D0:
2 @x2 @t
1 @2 P n
If Pn is of the form (8.53), the coefficient of xn2m2 tm in the polynomial 2 @x2 C
@Pn
@t is equal to
1
.n 2m/.n 2m 1/cn;m C .m C 1/cn;mC1 :
2
Requiring these quantities to vanish and setting cn;0 D 1, we can compute all
the coefficients sequentially one after the other, thus determining Pn , up to a
multiplicative constant. We find the polynomials
P1 .x; t/ D x
P2 .x; t/ D x2 t
P3 .x; t/ D x3 3tx
P4 .x; t/ D x4 6tx2 C 3t2 :
The first two give rise to already known martingales, whereas the third one is the
polynomial of a).
c) The stopping theorem, applied to the martingale .P3 .Bt ; t//t and to the bounded
stopping time ^ n, gives
E.B3^n / D 3 EŒ. ^ n/B ^n :
We can apply Lebesgue’s theorem and take the limit as n ! 1 (the r.v.’s B ^n
lie between a and b, whereas we already know that is integrable) so that
E.B3 / D 3 EŒB . Therefore, going back to Exercise 5.31 where the law of B
was computed
1 1 a3 b C ab3
E.B / D E.B3 / D D ab.b a/ :
3 3 aCb
If B and were independent we would have E.B / D E.B /E./ D 0.

Therefore B and are not independent if a 6D b. If conversely a D b, and
B are uncorrelated and we know already that in this case they are independent
(Exercise 3.18 b)).
8.6
a) Ito’s formula gives
dMt D e t Bt dt C e t dBt e t Bt dt D e t dBt

so that M is a square integrable martingale (also Gaussian. . . ) with associated

increasing process
Z t
1 2 t
hMit D e2 s ds D .e 1/ :
0 2
Note that, as M is Gaussian, this associated increasing process is deterministic

as was to be expected from Exercise 5.11.
b) If < 0, as
1 2 t
EŒMt2 D hMit D .e 1/ ;
2
M is bounded in L2 and therefore uniformly integrable and convergent a.s.

and in L2 . As M is Gaussian and Gaussianity is preserved with respect to
convergence in law (Proposition 1.9), the limit is also Gaussian with mean 0
and variance
1 2 t 1
lim .e 1/ D
t!C1 2 2
c1) Z is clearly an exponential martingale.

1
c2) If < 0, limt!C1 Mt exists in L2 and is a Gaussian N.0; 2 /-distributed r.v.
1 1
Hence Z1 is the exponential of an N. 4 ; 2 / distributed r.v., i.e. a lognormal
r.v. (Exercise 1.11).
c3) We have
p 2 t 1/ p2 p 2 t
EŒZt D EŒe pMt 4 .e 4 .e 1/
p
De :
If > 0, for p < 1 we have p2 p < 0, hence

p
lim EŒZt D 0 :
t!C1
Therefore, by Fatou’s lemma,

p
p
EŒZ1 lim EŒZt D 0 :
t!C1
p
The r.v. Z1 , being positive and with an expectation equal to 0, is equal to 0 a.s.
8.7 By Ito’s formula (see Example 8.6 for the complete computation)
dYt D t dBt
and therefore
1 1
d Yt t3 D t dBt t2 dt :
6 2
Hence, again by Ito’s formula,
1 3
1 1 1 3
dZt D eYt 6 t t dBt t2 dt C eYt 6 t dhYit
2 2
1 1
D Zt t dBt t2 dt C t2 dt D tZt dBt :
2 2
Therefore Z is a local martingale (and a positive supermartingale). In order to
prove that it is a martingale there are two possibilities. First we can prove that
Z 2 M 2 .Œ0; T/ for every T 0: as Y is a Gaussian process,
1 3 1 3 1 3 2 3 1 3
EŒZt2 D EŒe2Yt 3 t D e 3 t e2Var.Yt / D e 3 t e 3 t D e 3 t :
Therefore Z 2 M 2 and Z, being the stochastic integral of a process in M 2 , is a

square integrable martingale. A second method is to prove that EŒZt D 1 for every
t and to recall that a supermartingale with a constant expectation is a martingale
(Exercise 5.1). This verification is left to the reader.
8.8 By Corollary 8.1, there exists a Brownian motion W such that, if
Z t
ds
At D ,
0 2 C sin s
then Xt D WAt . In order to investigate the behavior of At as t ! C1 we must only

pay attention to the fact that the primitive indicated in the hint is not defined for
every value of t. Note that, as the integrand is a periodic function,
Z Z 2
2
ds
ds 2 t 1 ˇˇ 2
D D p arctan p tan C p ˇ D p
0 2 C sin s 2 C sin s 3 3 2 3 3
Hence, denoting by Œ the integer part function,

Z Z t
t
ds 2 h t i ds
At D Dp C
0 2 C sin s 3 2 2Œ 2
t
2 C sin s
so that limt!C1 1t At D 31=2 . We have therefore
Xt WAt .2At log log At /1=2

lim D lim
t!C1 .2t log log t/1=2 t!C1 .2At log log At /1=2 .2t log log t/1=2
As
.2At log log At /1=2

lim D 31=4 ;
by the Iterated Logarithm Law we find
Xt
lim D 31=4 ' 0:76 :
8.9 We have d.e t Bt / D e t Bt dt C e t dBt from which we derive

Z Z
t
s 1 1 t
Xt D e Bs ds D e t Bt C e s dBs :
0 0
The stochastic integral on the right-hand side is a martingale bounded in L2 and

therefore it converges a.s. and in L2 . As limt!C1 e t Bt D 0 a.s. (recall the Iterated
Logarithm Law) and in L2 , the limit in (8.54) also exists a.s. and in L2 , the limit
being the r.v.
Z C1
1
e s dBs ;
0
which is a centered Gaussian r.v. with variance

Z C1
1 1
e2 s ds D
2 0 2 3
8.10
a) We know (Proposition 7.1) that Xt" has a centered Gaussian distribution with
variance
Z t
s
2 sin2 ds :
0 "
Integrating by parts we find

Z t
s t t
2 sin2 ds D " sin cos C t : (S.44)
0 " " "
Hence, as " ! 0, Xt" converges in law to an N.0; t/-distributed r.v.

b) We know, by Theorem 8.2 and Corollary 8.1, that
Xt" D WA" "t
where W " is a Brownian motion and

Z t
s
A"t D2 sin2 ds :
0 "
The Brownian motion W " depends on " but of course it has the same law as B
for every ". Hence X " has the same distribution as .BA"t /t . As by (S.44) A"t ! t
uniformly in t, thanks to the continuity of the paths we have
ˇ ˇ
sup ˇBA"t Bt ˇ ! 0 :
0tT "!0
As a.s. convergence implies convergence in law, .BA"t /t , hence also .Xt" /t ,

converges in law as " ! 0 to the law of B, i.e. to the Wiener measure.
8.11 By Corollary 8.1 there exists a Brownian motion W such that, if

Z t
ds
At D D log.1 C t/ ;
0 1Cs
then Xt D WAt . Therefore, by the reflection principle applied to the Brownian motion
W,

P sup Xt 1 D P sup WAt 1 D P sup Ws 1
0t3 0t3 0slog 4
p Z C1
2 2 =2
D 2P.Wlog 4 1/ D 2P. log 4 W1 1/ D p ex dx ' 0:396 :
2 .log 4/1=2
8.12
a) We know that X is a time changed Brownian motion, more precisely (Corol-
lary 8.1)
Xt D WAt
where W is a Brownian motion and

Z t
At D .˛ C 1/ u˛ du D t˛C1 :
0
Therefore

P sup Xs 1 D P sup WAs 1 D P sup Wt 1 D 2P.WA2 1/
0s2 0s2 0tA2
p 1
D 2P. A2 W1 1/ D 2P.W1 2 2 .˛C1/ /
Z 1
2 2
D p et =2 dt :
21 .˛C1/
2 2
b) The previous computation with t instead of 2 gives

Z 1
2 2
P. t/ D P sup Xs 1 D p es =2 ds
21 .˛C1/
0st 2 t
and taking the derivative we find the density of :
2 1 1 ˛C1 .˛ C 1/ 1=.2t˛C1 /
f .t/ D p .˛ C 1/t 2 .˛C1/1 e1=.2t / D p e :
2 2 2t˛C3
In order for to have finite mathematical expectation the function t 7! tf .t/

must be integrable. At zero the integrand tends to 0 fast enough because of the
1
exponential whereas at infinity tf .t/ t 2 .˛C1/ . It is therefore necessary that
12 .˛ C 1/ < 1, i.e. ˛ > 1.
8.13 By Ito’s formula d.t Bt / D t0 Bt dt C t dBt . Therefore, for T > 0,

Z T Z T
t0 Bt dt C t dBt D T BT 0 B0 a.s.
0 0 „ƒ‚…
D0
If T is large enough, so that the support of is contained in Œ0; T, we have T D 0

and therefore
Z T Z T
t0 Bt dt D t dBt a:s:
0 0
Moreover, T can be replaced by C1, as both integrands vanish on T; C1Œ.

8.14
a) Let us apply Ito’s formula. If
1 x
.x; t/ D p exp
1t 2.1 t/
then Zt D .B2t ; t/ and we know that dB2t D 2Bt dBt C dt. Let us compute the
derivatives of :
@ 1 x x x
.x; t/ D exp exp
@t 2.1 t/3=2 2.1 t/ 2.1 t/5=2 2.1 t/
1 x
D .x; t/
2.1 t/ 2.1 t/2
and similarly
@ 1
.x; t/ D .x; t/
@x 2.1 t/
@2 1
2
.x; t/ D .x; t/ :
@x 4.1 t/2
Hence
@ @ 1 @2
dZt D .B2t ; t/ dt C .B2t ; t/ .2Bt dBt C dt/ C 2
.B2t ; t/ 4B2t dt
@t
h 1 @x 2 @x
B2t 1
D dt 2Bt dBt C dt
2.1 t/ 2.1 t/2 2.1 t/ i
1
C 4B2 dt .B2t ; t/
8.1 t/2 t
Bt 1
D .B2t ; t/ dBt D Bt Zt dBt :
1t 1t
(S.45)
This proves that Z is a local martingale. Being positive it is also a supermartin-

gale. In order to prove that it is actually a martingale it does not seem immediate
1
to verify that the stochastic integrand t 7! 1t Bt Zt is in M 2 , as it depends on Z
itself. It is however immediate that Z is bounded ( .1 t/1=2 ) in every interval
Œ0; T with T < 1 and we know that a bounded local martingale is a martingale
(Remark 7.7). It is also possible to check that, for every t < 1, EŒZt D EŒZ0 D 1,
and we know that a supermartingale having constant expectation is a martingale
(Exercise 5.1). This will be a byproduct of the computation of b).
b) We have
1 h pB2t i
p
EŒZt D E exp :
.1 t/p=2 2.1 t/
Now, denoting by W an N.0; 1/-distributed r.v. (recall Remark 3.3)
h p
pB2t i h ptW 2 i 1 1t
E exp D E exp D q Dp
2.1 t/ 2.1 t/ 1C pt 1 C .p 1/t
1t
so that
p 1
EŒZt D p
.1 t/.p1/=2 1 C .p 1/t
Therefore, for p D 1, EŒZt D 1, thus a second proof that Z is a martingale.

Moreover, Z belongs to Lp but, if p > 1, it is not bounded in Lp : the Lp norm
diverges as t ! 1.
c) Z being a continuous positive martingale, the limit limt!1 Zt WD Z1 exists a.s.
and is finite. From the computation of b) we have, for p < 1,
p
lim EŒZt D 0
t!1
and, as Z is positive, by Fatou’s lemma

p p
EŒZ1 lim EŒZt D 0 :
t!1
p
The r.v. Z1 being positive and having expectation equal to 0 is necessarily equal
to 0 a.s., i.e.
lim Zt D 0 a.s.
t!1
d) From (S.45) we recognize the differential of an exponential martingale. There-

fore a comparison between (S.45) and (8.11) suggests that
Z t
Bs 1
Z t
B2s
Zt D exp dBs 2
ds ;
0 1s 2 0 .1 s/
i.e.
Bs
Xs D
1s
This remark would be enough if we knew that (8.11) has a unique solution.
Without this we can nevertheless verify directly that, for t < 1,
Z Z
1 B2t t
Bs 1 t
B2s
p exp D exp dBs ds :
1t 2.1 t/ 0 1s 2 0 .1 s/2
(S.46)
Ito’s formula for the product gives
B2t 2Bt dBt C dt B2t

d D C dt
2.1 t/ 2.1 t/ 2.1 t/2
hence integrating
Z Z
B2t t
Bs t
B2s 1
D dBs C ds log.1 t/
2.1 t/ 0 1s 0 2.1 s/2 2
so that
Z Z
Bt p t
Bs 1 t
B2s
exp D 1 t exp dBs ds
2.1 t/ 0 1s 2 0 .1 s/2
and (S.46) follows.
8.15
a) It is immediate that W1 is a centered Gaussian process. Its variance is equal to
Z t Z t
2
.sin s C cos s/ ds D .1 C 2 sin s cos s/ ds D t C .1 cos 2t/ 6D t ;
0 0
hence W1 is not a Brownian motion.

b) We can write
Z t
W2 .t/ D u.s/ dBs
0
with u.s/ D .sin s; cos s/. As u.s/ is a vector having modulus equal to 1 for every
s, W2 is a Brownian motion thanks to Corollary 8.2.
c) As in b) we can write
Z t
W3 .t/ D u.s/ dBs
0
where now u.s/ D .sin B2 .s/; cos B2 .s//. Again, as u.s/ has modulus equal to
1 for every s, W3 is a Brownian motion. Note that, without taking advantage of
Corollary 8.2, it is not immediate to prove that W3 is a Gaussian process.
8.16 Let us denote by B the two-dimensional Brownian motion .B1 ; B2 /.

a) We can write
Z t
Xt D Zs dBs
0
with Zs D .sin B3 .s/; cos B3 .s//. As jZs j D 1 for every s, X is a Brownian

motion by Corollary 8.2. The same argument applies for Y.
b1) We have, for s t,
EŒXs Yt
hZ s Z t
DE sin.B3 .u// dB1.u/ cos.B3 .v// dB1 .v/
0 0
Z s Z t i
C cos.B3 .u// dB2.u/ sin.B3 .v// dB2 .v/
0 0
hZ s Z s i
DE sin.B3 .u// cos.B3 .u// du C cos.B3 .u// sin.B3 .u// du
0 0
Z s
D2 EŒsin.B3 .u// cos.B3 .u// du D 0 ;
0
where the last equality comes from the fact that sin.B3 .u// cos.B3 .u// has
the same distribution as sin.B3 .u// cos.B3 .u// D sin.B3 .u// cos.B3 .u//,
and has therefore mathematical expectation equal to 0. Hence Xs and Yt are
uncorrelated.
b2) If .Xt ; Yt /t were a two-dimensional Brownian motion the product t 7! Xt Yt
would be a martingale (see Exercises 5.22 or 5.24). This is not true because
Z t
hX; Yit D 2 sin.B3 .u// cos.B3 .u// du 6 0 :
0
c) With the new definition we can write

Z t
Xt
D Os dBs
Yt 0
with

sin.B3 .s// cos.B3 .s//
Os D :
cos.B3 .s// sin.B3 .s//
It is immediate that s 7! Os is an orthogonal matrix-valued process. Hence the

required statement follows from Proposition 8.8.
• Note the apparent contradiction: the processes X and Y in b) are each a Brownian
motion and are uncorrelated. But they do not form a two-dimensional Brownian
motion. Actually they are not jointly Gaussian.
8.17
a) Orthogonality with respect to Bv for v s imposes the condition
h Z s i Z v
0DE Bt ˚.u/ dBu ˛B1 Bv D v ˚.u/ du ˛v ;
0 0
i.e.
Z v
v.1 ˛/ D ˚.u/ du; for every v s (S.47)
0
and therefore ˚ 1 ˛ on Œ0; s. Orthogonality with respect to B1 conversely

requires
h Z s i Z s
0DE Bt ˚.u/ dBu ˛B1 B1 D t ˚.u/ du ˛ ;
0 0
i.e., taking into account that ˚ 1 ˛,
0 D t .1 ˛/s ˛ D t s ˛.1 s/ ;
i.e.
ts 1t
˛D ; ˚.u/
1s 1s
1t
b) Let X D 1sts
B1 C 1s Bs . In a) we have proved that the r.v. Bt X, which is
centered, is independent of G es -measurable,
es . Moreover, as X is G
es D EŒ.Bt X/ C X j G
EŒBt j G es
(S.48)
ts 1t
D X C EŒBt X D X D B1 C Bs :
1s 1s
Hence B is adapted to the filtration .G et /t -martingale, therefore

et /t but is not a .G
it cannot be a Brownian motion with respect to this filtration.
c1) As the r.v.’s eBt e
Bs ; B1 ; Bv ; v s form a Gaussian family, it suffices
(Remark 1.2) to show that e Bt e Bs is orthogonal to Bv , 0 v s, and to
B1 . We have
Z Z
t
EŒ.B1 Bu /B1 t
EŒ.e
Bt e
Bs /B1 D EŒ.Bt Bs /B1 du D t s du D 0
s 1u s
and, for v s,
Z t
EŒ.B1 Bu /Bv
Bt e
EŒ.e Bs /Bv D EŒ.Bt Bs /Bv du D 0 :
s 1u
c2) Let us prove that EŒeBte

Bs D s, for 0 s t. This is elementary, albeit
Bte
laborious. If s < t 1, note that EŒe Bs D EŒ.e
Bt e
Bs /e
Bs C EŒe
B2s D EŒe
B2s ,
thanks to c1). Hence we are reduced to the computation of
EŒe
B2s
s h Z Z s Z s h
2 B1 Bu i B1 Bv B1 Bu i
D E.Bs / 2 E Bs du C dv E du
0 1u 0 0 1v 1u
Z s Z s Z s
su 1uvCu^v
Ds2 du C dv du
0 1u 0 0 .1 v/.1 u/
D s 2I1 C I2 :
With patience one can compute I2 and find that it is equal to 2I1 , which gives the
result. The simplest way to check that I2 D 2I1 is to observe that the integrand
in I2 is a function of .u; v/ that is symmetric in u; v. Hence
Z Z Z s Z s
s
1uvCu^v
s
1uvCu^v
I2 D dv dv D 2 dv du
0 0 .1 v/.1 u/ 0 v .1 v/.1 u/
Z s Z s Z s
1u sv
D2 dv du D 2 dv D 2I1 :
0 v .1 v/.1 u/ 0 1 v
Hence EŒe BteBs D s ^ t. If s t, the r.v. e

Bt e
Bs is centered Gaussian as .e
B t /t
is clearly a centered Gaussian process, which together with c1), completes the
proof that e et /t -Brownian motion.
B is a .G
c3) We have, of course,
dBt D At dt C de
Bt
with
B1 Bt
At D
1t
et /t -adapted, B is an Ito process with respect to the new
Hence, since .At /t is .G
Brownian motion e B.
8.18
a) If there existed a second process .Ys0 /s satisfying (8.56), we would have
h Z T Z T 2 i h Z T 2 i
0DE Ys dBs Ys0 dBs DE .Ys Ys0 / dBs
0 0 0
hZ T i
DE .Ys Ys0 /2 ds
0
and therefore Ys D Ys0 a.e. with probability 1 and the two processes would be
indistinguishable.
b1) As usual, for s T,
Xs D EŒ.Bs C .BT Bs //3 jFs

D B3s C 3B2s EŒBT Bs C 3Bs EŒ.BT Bs /2 C EŒ.BT Bs /3 D B3s C 3Bs .T s/ :
By Ito’s formula,
dXs D 3B2s dBs C 3Bs ds C 3.T s/ dBs 3Bs ds D 3.B2s C .T s// dBs :
Note that the part in ds vanishes, which is not surprising as .Xs /s is clearly a
martingale.
b2) Obviously
Z T
B3T D XT D 3.B2 C .T s// dBs :
0 „ s ƒ‚ …
WDYs
2
c) We can repeat the arguments of b1) and b2): as s 7! e Bs 2 s
is a martingale,
2 2 2 2 2
T Bs 2 s
Xs D EŒe BT jFs D e 2 T
EŒe BT 2 T
jFs D e 2 e D e B s C 2 .Ts/
and
2 2
dXs D e 2 T
e Bs 2 s
dBs
and therefore
Z T Z T
2 2 2
e B T D X T D X 0 C e B s C 2 .Ts/
dBs D e 2 T
C e
„
Bs C 2
ƒ‚
.Ts/
… dBs :
0 0
WDYs
8.19
a1) We have
hZ T ˇ i
Z t hZ T ˇ i
Mt D EŒZ jFt D E Bs ds ˇ Ft D Bs ds C E Bs ds ˇ Ft
0 0 t
Z t
D Bs ds C .T t/Bt :
0
a2) Ito’s formula gives
dMt D Bt dt C .T t/ dBt Bt dt D .T t/ dBt
and therefore, as M0 D 0,
Z T Z T
Bs ds D MT D .T t/ dBt :
0 0 „ ƒ‚ …
WDXt
b) Let again
hZ T ˇ i
Z t hZ T ˇ i
Mt D E B2s ds ˇ Ft D B2s ds C E B2s ds ˇ Ft :
0 0 t
Now, recalling that t 7! B2t t is a martingale,

Z Z
T T
1 2
EŒB2s jFt ds D .EŒB2s sjFt C s/ ds D .T t/.B2t t/ C .T t2 /
t t 2
and putting things together

Z t
1 2
Mt D B2s ds C .T t/.B2t t/ C .T t2 / :
0 2
As dB2t D 2Bt dBt C dt,
dMt D B2t dt C .T t/.2Bt dBt C dt dt/ .B2t t/ dt t dt D 2.T t/Bt dBt
T2
and since M0 D 2
we have
Z Z
T
T2 T
B2s ds D MT D C2 .T t/Bt dBt :
0 2 0
8.20 Ito’s formula applied to the unknown process Z gives

2
dZt D Zt .
dMt dAt / C Zt .Bi .t/2 C Bj .t/2 / dt
2
1
D
Zt .Bi .t/ dBj .t/ C Bj .t/ dBi .t// C Zt
2 .Bi .t/2 C Bj .t/2 / dt dAt :
2
1 1
Z is therefore a local martingale if and only if dAt D 2
2 dhMit D 2
2 .Bi .t/2 C
Bj .t/2 / dt.
• Of course this leaves open the question of whether such a Z is a true martingale.
The interested reader will be able to answer this easily later using Proposi-
tion 12.2.
8.21
a) We have
Z C1 Z C1
12 ˛Z 2 ˛x2 x2 =2 1 1 2
EŒZ e D p xe e dx D p x2 e 2 .12˛/x dx
2 1 2 1
1=2 Z C1
.1 2˛/ 1 2
D .1 2˛/1=2 p x2 e 2 .12˛/x dx D .1 2˛/3=2
2 1
„ ƒ‚ …
as the quantity over the brace is equal to the variance of an N.0; .1 2˛/1 /-
distributed r.v. and is therefore equal to .1 2˛/1 .
b) We have limt!1 Ht .!/ D 0 for ! 62 fB1 D 0g, which is a set of probability
0. Hence t 7! Ht .!/ is continuous for t 2 Œ0; 1 if ! 62 fB1 D 0g, and H 2
2
Mloc .Œ0; 1/. In order to check whether H 2 M 2 .Œ0; 1/ we have, denoting by Z an
N.0; 1/-distributed r.v.,
hZ 1 i Z 11 h B2s i
E Hs2 ds D 3
E B 2
exp ds
0 .1 s/ 1s
s
0
Z h Z 1
1
1 2 sZ i s 2s 3=2
D E sZ exp ds D 1 C ds
0 .1 s/3 1s 0 .1 s/
3 1s
Z 1
s
D 3=2
D C1 ;
0 .1 s/ .1 C s/3=2
so that H 62 M 2 .Œ0; 1/.

c) Let
1 x2
f .x; t/ D p exp
1t 2.1 t/
and let us compute the stochastic differential, for 0 t < 1, of Xt D f .Bt ; t/. We
have
@f x x2
.x; t/ D exp ;
@x .1 t/3=2 2.1 t/
so that
@f
.Bt ; t/ D Ht
@x
and, with some patience,
@f 1 x2 x2
.x; t/ D 3=2
5=2
exp
@t 2.1 t/ 2.1 t/ 2.1 t/
@2 f 1 x2 x2
2
.x; t/ D 3=2
C 5=2
exp ;
@x .1 t/ .1 t/ 2.1 t/
so that
@f 1 @2 f
.x; t/ C .x; t/ D 0 :
@t 2 @x2
By Ito’s formula,
@f 1 @2 f @f @f
dXt D .Bt ; t/ C 2
.B t ; t/ dt C .Bt ; t/ dBt D .Bt ; t/ dBt D Ht dBt ;
@t 2 @x @x @x
from which, as X0 D 1, (8.57) follows. It is immediate that if ! 62 fB1 D 0g,

then limt!1 Xt D 0.
d) The integral
Z 1
Hs dBs
0
2
is well defined as H 2 Mloc .Œ0; 1/. Moreover, by continuity and thanks to (8.57),
Z 1 Z t
Hs dBs D lim Hs dBs D 1 lim Xt D 1 :
0 t!1 0 t!1
8.22
a) This was proved in Exercise 5.10 c).
.n/
b) In Example 8.9 it is proved that, if Xn .t/ D jBt j2 , then there exists a real
Brownian motion W such that
p
dXn .t/ D n dt C 2 Xn .t/ dWt
from which (8.58) follows.

c) We have, thanks to (8.58),
Z p
2 t^n
Rn .t ^ n / t ^ n D p Rn .s/ dWs :
n 0
p
As Rn .s/ 1 for s n , Zt D Rn .t ^ n / t ^ n is a square integrable
martingale. By Doob’s inequality,
h i
E sup jRn .t/ tj2 D E sup Zt2 4 sup E.Zt2 /
0tn t0 t0
16 hZ n i 16 16
D E Rn .s/ ds E.n / D
n 0 n n
2 16
In particular, E.jRn .n /n /j / n
and, recalling that Rn .n / D 1, by Markov’s
inequality,
16
P.j1 n /j "/ ,
n"2
which proves that n !n!1 1 in probability.
d) n;d .n / is the law of Xn D .B1 .n /; : : : ; Bd .n //, as indicated in the hint. As
n !n!1 1 in probability, from every subsequence of .n /n we can extract
a further subsequence converging to 1 a.s. Hence from every subsequence of
.Xn /n we can extract a subsequence converging to X D .B1 .1/; : : : ; Bd .1// a.s.
Therefore Xn ! N.0; I/ in law as n ! 1.
8.23
a) Let f .z/ D log jz xj so that Xt D f .Bt /. We cannot apply Ito’s formula
to f , which is not even defined at x, but we can apply it to a C2 function that
coincides with f outside the ball of radius 1n centered at x. Let us compute the
derivatives of f . As remarked in Sect. 8.5, we have for g.z/ D jz xj, z 6D x,
v
uX
@ u
m
@g t .zj xj /2 D zi xi
.z/ D
@zi @zi jD1 jz xj
and
@f zi xi
.z/ D
@zi jz xj2
@2 f 1 .zi xi /2
2
.z/ D C 2
@zi jz xj2 jz xj4
from which 4f .z/ D 0 for z 6D x. Hence

Z Z
t^n
1 t^n
Xt^n D log jxj C 4f .Bs / ds C f 0 .Bs / dBs
Z t^n 0 2 Z 0
B1 .s/ x1 t^n
B2 .s/ x2
D log jxj dB1 .s/ dB2 .s/ :
0 jB.s/ xj2 0 jB.s/ xj2
(S.49)
1
As jBs xj n for s n , we have
jBi .s/ xi j 1
2
n
jB.s/ xj jB.s/ xj
so that the integrands in (S.49) are processes in M 2 and .Xt^n /n is a square

integrable martingale.
b) Of course, n;M < C1 by the Iterated Logarithm Law. Moreover, by the
stopping theorem we have for every t > 0
log jxj D EŒXt^n;M
and, as log M Xt^n;M log 1n , we can apply Lebesgue’s theorem, taking
the limit as t ! C1. Therefore

log jxj D EŒXn;M D log 1n P.jBn;M xj D 1n /log M 1P.jBn;M xj D 1n /
from which we obtain
log M log jxj

P.jBn;M xj D 1n / D (S.50)
log M log 1n
c) We have fn < C1g fn n;M g D fjBn;M xj D 1n g. Therefore by (S.50)
log M log jxj

P.n < C1/
log M log 1n
This inequality holds for every M and taking the limit as M ! C1 we obtain
P.n < C1/ D 1. Hence B visits a.s. every neighborhood of x. The point x
being arbitrary, B visits every open set a.s. and is therefore recurrent.
d1) The probability P.k < k / is obtained from (S.50) by replacing M with k and
n by kk . Hence
log k log jxj log k log jxj

P.k < k / D 1
D
log k log kk .k C 1/ log k
and as k ! 1 this probability converges to 0.

d2) Note that k < : before reaching x, B must enter the ball of radius k1k centered
at x. Hence we have f < k g fk < k g and P. < k / P.k < k /. As
k ! C1 for k ! 1,
P. < C1/ D lim P. < k / lim P.k < k / D 0 :

k!1 k!1
8.24
a) We would like to apply Ito’s formula to z 7! jz C xj, which is not possible
immediately, as this is not a C2 function. In order to circumvent this difficulty
let gn W Rm ! R be such that gn .z/ D jz C xj for jz C xj 1n and extended for
jz C xj < 1n in such a way as to be C2 .Rm /. For jz C xj 1n therefore
@gn zi C xi @g2n ıij .zi C xi /.zj C xj /

.z/ D ; .z/ D
@zi jz C xj @zi @zj jz C xj jz C xj3
hence
X
m
@g2 m1
4gn .x/ D 2
n
.x/ D
iD1
@zi jz C xj
If n D infftI jBt C xj < 1n g, Ito’s formula applied to the function gn gives

Z Z t^n
t^n
1
jBt^n C xj D jxj C 4gn .Bs / ds C g0n .Bs / dBs
0 2 0
Z Z t^n
t^n
m1 Bs C x
D jxj C ds C dBs :
0 2 jBs C xj 0 jB s C xj
By Corollary 8.2 the process

Z t
Bs C x
Wt D dBs
0 jBs C xj
is a Brownian motion. Note, as seen in Sect. 8.5 (or in Exercise 8.23 for
dimension d D 2), that jBt C xj > 0 for every t > 0 a.s. and n ! C1 as
n ! 1. Therefore, taking the limit as n ! 1 and setting D jxj,
Z t
m1
Xt D C ds C Wt ;
0 2Xs
which gives the required stochastic differential.

b) If f 2 CK2 .0; C1Œ/, then by Ito’s formula
1
df .Xt / D f 0 .Xt / dXt C f 00 .Xt / dt D Lf .Xt / dt C f 0 .Xt / dWt ;
2
where
1 d2 m1 d
LD 2
C
2 dy 2y dy
As f has compact support 0; C1Œ,Rits derivative is bounded and therefore the
expectation of the stochastic integral 0 f 0 .Xs / dWs vanishes. In conclusion
t
Z
1 1 t
EŒ f .Xt / f ./ D EŒLf .Xs / ds ! Lf ./ :
t t 0 t!C0
9.1
a) Starting from the explicit solution, formula (9.3), the law of t is Gaussian with
mean e t x and variance
Z t
2 2
t D e2 .ts/ ds D .1 e2 t /
0 2
As > 0,
2
lim e t x D 0; lim t D
t!C1 t!C1 2
This implies (Exercise 1.14) that, for every x, the law of t converges weakly, as
2
t ! C1, to a Gaussian law with mean 0 and variance 2 .
b) This follows from Exercise 6.10 d) but let us verify this point directly. Let be a
1
Gaussian r.v. with distribution , i.e. N.0; 2 /, and independent of B. Then
a repetition of the arguments of Example 9.1 gives that a solution of (9.45) with
the initial condition 0 D is
Z t
t D e t C e t e s dBs :
„ƒ‚… 0
DY1 „ ƒ‚ …
DY2
Y1 and Y2 are independent and Gaussian, centered and with variances
2 2 t 2
e and .1 e2 t / ;
2 2
respectively. Therefore t is Gaussian, centered and with variance
2 2 t 2 2 ,
e C .1 e2 t / D
2 2 2
i.e. t .
9.2
a) Let us follow the idea of the variation of constants of Example 9.1. The solution
of the ordinary differential equation
x0t D b.t/xt
x0 D x
Rt
is xt D e .t/ x, where .t/ D 0 b.s/ ds. Let us look for a solution of (9.46) of the
form xt D e .t/ C.t/. One sees easily that C must be the solution of
e .t/ dC.t/ D .t/ dBt ;
i.e.
Z t
C.t/ D e .s/ .s/ dBs :
0
The solution of (9.46) is therefore

Z t
t D e .t/ x C e .t/ e .s/ .s/ dBs : (S.51)
0
As the stochastic integral of a deterministic function is, as a function of

the integration endpoint, a Gaussian process (Proposition 7.1), is Gaussian.
Obviously E.t / D e .t/ x. Let us compute its covariance function. Let s t and
let
Z t
Yt D e .t/ e .s/ .s/ dBs :
0
Then, by Proposition 8.5,
Z t Cov.t ; s / D E.YZt Ys /

h s i
D E e .t/ e .u/ .u/ dBu e .s/ e .v/ .v/ dBv
h Z 0t Z s 0 i (S.52)
.t/ .u/
De E e .u/ dBu e .v/ .v/ dBv e .s/
0 Z s 0
.t/ .u/
De e .u/ .u/ e .u/ du e .s/ :

0
In particular, if m D 1 the covariance is

Z s
e .t/ e .s/ e2 .u/ 2 .u/ du :
0
b) We have, for this equation,

Z t
1
.t/ D ds D log.1 t/ :
0 1s
Therefore e .t/ D 1 t and the solution of (9.47) is

Z t
dBs
t D .1 t/x C .1 t/
0 1s
Hence EŒt D .1 t/x. is a Gaussian process with covariance function

Z 1
s
1
K.t; s/ D .1 t/.1 s/ du D .1 t/.1 s/ 1 D s.1 t/ ;
0 .1 u/2 1s
for s t. If x D 0, then E.t / D 0 for every t and the process has the same mean
and covariance functions as a Brownian bridge.
9.3
a1) Recall, from Example 9.1, that the solution is
Z t
t D e t x C e t e s dBs :
0
By Corollary 8.1 there exists a Brownian motion W such that, if

Z t
2 2 t
At D e2 u 2 du D .e 1/ ;
0 2
then
Z t
e s dBs D WAt :
0
We have
t e t x C e t WAt e t WAt
lim p D lim p D lim p
t!C1 log t t!C1 log t t!C1 log t
Now, trying to go back to the Iterated Logarithm Law,
e t WAt e t .2At log log At /1=2 WAt

lim p D lim p
t!C1 log t t!C1 log t .2At log log At /1=2
and
WAt
lim D1
t!C1 .2At log log At /1=2
whereas
e t .2At log log At /1=2 log log A.t/ 1=2

lim p D lim .2e2 t A.t//1=2 lim
t!C1 log t t!C1 t!C1 log t
and we can conclude the proof as
2
lim 2e2 t At D
t!C1
and log log At log t as t ! C1. The the case of lim is treated in the same
way.
a2) Thanks to a1), for a sequence tn % C1 we have
p
tn p log tn :
2
Therefore
lim t D C1 :
t!C1
Similarly we obtain limt!C1 t D 1. Therefore the Ornstein–Uhlenbeck

process, in this case, is recurrent in the sense explained p. 243.
b) We can write
Z t
t D e t x C e u dBu :
0
The stochastic integral above is a martingale bounded in L2 , as its variance is

Rt 2
equal to 0 e2 u 2 du < 2 , as < 0. Therefore there exists a r.v. X such that
Z t
xC e u dBu ! X a.s.
0 t!C1
2
The limit X, moreover, is Gaussian with mean x and variance 2 . Therefore
limt!C1 t D C1 on the set A D fX > 0g and limt!C1 t D 1 on
2
Ac D fX < 0g. As X is N.x; 2 /-distributed, we have X D x C p Z, where
2
Z N.0; 1/. Hence
p
x 2
P.A/ D P.X > 0/ D P x C p Z>0 DP Z>
2
p p
x 2 x 2
DP Z< D˚ :

9.4
a) This is a particular case of the general situation of Exercise 9.2 a). The solution
of the “homogeneous equation”
1 t
dt D dt
2 1t
0 D x
is immediately found to be equal to

p
t D 1tx ;
p
and the “particular solution” is given by t 7! 1 t Ct , where C must satisfy
p p
1 t dCt D 1 t dBt ;
i.e. Ct D Bt and the solution of (9.48) is found to be

p
t D 1 t.x C Bt / :
is obviously a Gaussian process (this is a particular case of Exercise 9.2 a)).

b) We have
Var.t / D .1 t/t ;
which is the same as the variance of a Brownian bridge. As for the covariance
function we have, for s t,
p p
p p
Cov.t ; s / D E 1 t Bt 1 s Bs D s 1 s 1 t ;
which is different from the covariance function of a Brownian bridge. Note that,
if the starting point x is the origin, then, for every t; 0 t 1, the distribution
of t coincides with the distribution of a Brownian bridge at time t, but is not a
Brownian bridge.
9.5
a) Similarly to the idea of Example 9.2, if we could apply Ito’s formula to the
function log we would obtain the stochastic differential
1 1 2 .t/
d log t D dt 2 2 .t/ t2 dt D b.t/ dt C .t/ dBt (S.53)
t 2t 2
and we would have

Z Z
t 2 .s/ t
log t D b.s/ ds C .s/ dBs ;
0 2 0
which gives for (9.49) the solution

Rt 2 .s/ Rt
t D xe 0 b.s/ 2 ds C 0 .s/ dBs : (S.54)
It is now immediate to check by Ito’s formula correctly applied to the exponential

function that the process of (S.54) is a solution of (9.49).
b) In case (1) is square integrable, hence
Z t
.s/ dBs ! Z;
0 t!C1
R C1
where Z is a centered Gaussian r.v. with variance 0 2 .t/ dt D 1. The
convergence takes place in L2 and also a.s. (it is a martingale bounded in L2 ).
2
On the other hand b.s/ 2.s/ D 2.1Cs/
1
and
Z
C1
2 .s/
b.s/ ds D C1
0 2
and therefore t ! C1 a.s.

In case (2)
Z C1
2 .s/ ds D 1
0
Z Z C1
C1
2 .s/ 1 1
b.s/ ds D 2
ds D
0 2 0 6.1 C s/ 6
Rt
The martingale t 7! 0 .s/ dBs converges a.s. and in L2 , being bounded in L2 ,
therefore, as t ! C1, t ! xeZ where Z N. 16 ; 1/, the convergence taking
place also a.s.
2 .s/
Finally, in situation (3) we have b.s/ 2 D 0 but
Z C1 Z C1
2 1
.s/ ds D ds D C1 :
0 0 1Cs
Therefore, if
Z Z
t t
1
At D 2 .s/ ds D ds D log.1 C t/ ;
0 0 1Cs
then
Z t
.s/ dBs D WAt ;
0
where W is a new Brownian motion. As limt!C1 At D C1, by the Iterated

Logarithm Law,
Z t Z t
lim .s/ dBs D C1; lim .s/ dBs D 1
t!C1 0 t!C1 0
which implies
lim t D C1; lim t D 0 :

t!C1 t!C1
9.6
a) Using the same idea as in Example 9.2, applying Ito’s formula formally we have
1 1 2
˝Xd Xd
˛
d log i .t/ D di .t/ 2
i .t/ d ij B j .t/; ik Bk .t/ t
i .t/ 2i .t/ jD1 kD1
X
d
1X 2
d aii X d
D bi dt C ij dBj .t/ ij dt D bi dt C ij dBj .t/ ;
jD1
2 jD1 2 jD1
where we note a D . This would yield the solution
aii Xd
i .t/ D xi exp bi tC ij Bj .t/ : (S.55)
2 jD1
Actually, Ito’s formula cannot be applied in the way we did, log not being defined
on the whole of R. We can, however, apply Ito’s formula to the exponential
function and check that a process whose components are given by (S.55)
actually is a solution of (9.50). From (S.55) we have also that if xi > 0 then
i .t/ > 0 for every t a.s.
Recalling the expression of the Laplace transform of Gaussian r.v.’s we have
aii Pd
aii 1
Pd 2
EŒi .t/ D xi e.bi 2 /t E e jD1 ij Bj .t/ D xi e.bi 2 /tC 2 jD1 ij D xi ebi t :
b) We have
aii Xd
ajj Xd
2 2
i .t/ j .t/ D exp 2 bi tC2 ih Bh .t/ C 2 bj tC2 ik Bk .t/ ;
2 hD1
2 kD1
which is an integrable r.v. and, moreover, is such that t 7! EŒi .t/2 j .t/2 is
continuous (again recall the expression of the Laplace transform of Gaussian
r.v.’s), hence t 7! i .t/j .t/ is in M 2 .
Moreover, by Ito’s formula,
di .t/j .t/ D i .t/ dj .t/ C j .t/ di .t/ C dhi ; j it
X
d
D bj i .t/j .t/ dt C i .t/j .t/ jh dBh .t/ C bi j .t/i .t/ dt
hD1
X
d
Cj .t/i .t/ ik dBk .t/ C i .t/j .t/aij dt :
kD1
Writing the previous formula in integrated form and taking the expectation, we
see that the stochastic integrals have expectation equal to 0, as the integrands are
in M 2 . We find therefore
Z t
EŒi .t/j .t/ D xi xj C EŒi .s/j .s/.bi C bj C aij / ds :
0
If we set v.t/ D EŒi .t/j .t/, then v satisfies the ordinary equation
v 0 .t/ D .bi C bj C aij / v.t/ ;
hence
EŒi .t/j .t/ D v.t/ D xi xj e.bi Cbj Caij /t :
Therefore

Cov.i .t/; j .t// D EŒi .t/j .t/ EŒi .t/EŒj .t/ D xi xj e.bi Cbj Caij /t e.bi Cbj /t

D xi xj e.bi Cbj /t eaij t 1 :
• This is one of the possible extensions of geometric Brownian motion to a

multidimensional setting.
9.7
a) The clever reader has certainly observed that the processes 1 ; 2 are both
geometric Brownian motions for which an explicit solution is known, which
allows us to come correctly to the right answer. Let us, however, work otherwise.
Observing that the process h1 ; 2 i vanishes,
d1 .t/2 .t/ D 1 .t/ d2 .t/ C 2 .t/ d1 .t/

D 1 .t/ r2 2 .t/ dt C 2 2 .t/ dB2 .t/ C 2 .t/ r1 1 .t/ dt C 1 1 .t/ dB1 .t/ ;
hence

dXt D .r1 C r2 /Xt dt C Xt 1 dB1 .t/ C 2 dB2 .t/ :
If
1
Wt D q 1 B1 .t/ C 2 B2 .t/
12 C 22
then W is a Brownian motion and the above relation for dXt becomes
q
dXt D .r1 C r2 /Xt dt C 12 C 22 Xt dWt :
X is therefore also a geometric Brownian motion and
1 2 2
p
12 C22 Wt
Xt D x0 e.r1 Cr2 2 .1 C1 //tC : (S.56)
p
In order to investigate the case 1 .t/2 .t/ we can takepeither the square root
in (S.56) or compute the stochastic differential of Zt D Xt . The latter strategy
gives
p 1 1
dZt D d Xt D p dXt 3=2
.12 C 22 /Xt2 dt
2 Xt 8Xt
q p
1 1
D p .r1 C r2 /Xt dt C 12 C 22 Xt dWt .12 C 22 / Xt dt
2 Xt 8
1 q
1 1
D .r1 C r2 / .12 C 22 / Zt dt C 12 C 22 Zt dWt ;
2 8 2
which is again a geometric Brownian motion withpsome new parameters r and .

This argument is not completely correct as x 7! x is not a C2 function, but the
clever reader has certainly learned how to go round this kind of difficulty.
b) We can repeat the previous arguments: now dh1 ; 2 it D 1 2 1 .t/2 .t/ dt so
that
d1 .t/2 .t/ D 1 .t/ d2 .t/ C 2 .t/ d1 .t/ C dh1 ; 2 it
p
D 1 .t/ r2 2 .t/ dt C 2 1 2 2 .t/ dB2 .t/ C 2 2 .t/ dB1 .t/

C2 .t/ r1 1 .t/ dt C 1 1 .t/ dB1 .t/ C 1 2 1 .t/2 .t/ dt
so that, if Xt D 1 .t/2 .t/,

p
dXt D .r1 C r2 C 1 2 /Xt dt C Xt .1 C 2 / dB1 .t/ C 2 1 2 dB2 .t/ :
Now we have a new Brownian motion by letting
1 p
Wt D q .1 C 2 / B1 .t/ C 2 1 2 B2 .t/
12 C 22 C 21 2
and we find that X is again a geometric Brownian motion with stochastic

differential
q
dXt D .r1 C r2 C 1 2 /Xt dt C 12 C 22 C 21 2 Xt dWt :
The same arguments as above give

1 1 2
dZt D .1 C 22 C 21 2 / Zt dt
.r1 C r2 C 1 2 /
2 8
q
1
C 12 C 22 C 21 2 Zt dWt
2
1 q
1 2 2 1
D .r1 C r2 / .1 C 2 21 2 / Zt dt C 12 C 22 C 21 2 Zt dWt ;
2 8 2
which is again a geometric Brownian motion.
9.8
a) Two possibilities: by Ito’s formula, for every real number ˛
1
dt˛ D ˛t˛1 dt C ˛.˛ 1/t˛2 2 t2 dt
2
2
D ˛b C ˛.˛ 1/ t˛ dt C t˛ dBt :
2
If the coefficient of dt in the stochastic differential above vanishes then ˛ will be

a local martingale and actually a martingale (why?). The condition is therefore
˛ D 0 (obviously) or
2
bC .˛ 1/ D 0 ;
2
i.e.
2b
˛ D1 (S.57)
2
Second possibility: we know that has the explicit form
2
t D e.b 2 /tC Bt
and therefore
2
t˛ D e˛.b 2 /tC˛ Bt ;
which turns out to be an exponential martingale if
2 2
˛.b 2
/ D ˛ 2
2
and we obtain again (S.57).
Note, however, that the use of Ito’s formula above requires an explanation, as
the function x 7! x˛ is not defined on the whole of R.
b) By the stopping theorem we have, for every t 0 and the value of ˛ determined
in a),
EŒ˛^t D 1 :
As t 7! ˛^t remains bounded, Lebesgue’s theorem gives

1 D EŒ˛ D 2˛ P. D 2/ C 2˛ 1 P. D 2/ ;
from which we find

1 2˛
P. D 2/ D
2˛ 2˛
9.9
a) It is immediate that .t /t is a geometric Brownian motion and
1
t D ye. 2 /tCB2 .t/
and therefore
Z t
1
t D x C y e. 2 /tCB2 .t/ dB1 .t/ : (S.58)
0
This is a martingale, as the integrand in (S.58) is in M2 .Œ0; T/ for every T > 0
since
Z T Z T
.21/tC2B2 .t/
E e dt D e.21/tC2t dt < C1 :
0 0
Alternatively, in a more immediate way, the pair .; / is the solution of an

SDE with coefficients with sublinear growth, and we know that the solutions
are automatically in M2 .Œ0; T/ for every T > 0 (Theorem 9.1).
b) Obviously E.t / D x, as .t /t is a martingale and 0 D x. Moreover, if 6D 12 ,
h Z t i Z t
Var.t / D EŒ.t x/2 D E y2 e.21/tC2B2 .t/ dt D y2 EŒe.21/tC2B2 .t/ dt
0 0
Z t Z t 2
y
D y2 e.21/tC2t dt D y2 e.2C1/t dt D .e.2C1/t 1/
0 0 2 C 1
whereas, if D 12 , Var.t / D y2 t.
c) From b) if < 12 then E.t2 / D E.t /2 C Var.t / D x2 C Var.t / is bounded in
t. Therefore .t /t is a martingale bounded in L2 and therefore convergent a.s. and
in L2 .
9.10
a) We note that (9.51) is an SDE with sublinear coefficients. Therefore its solution
belongs to M 2 and taking the expectation we find
Z t
EŒt D x C .a C bEŒs / ds :
0
Thus the function v.t/ D EŒt is differentiable and satisfies
v 0 .t/ D a C bv.t/ : (S.59)
The general integral of the homogeneous equation is v.t/ D ebt v0 . Let us look
for a particular solution of the form t 7! ebt c.t/. The equation becomes
bebt c.t/ C ebt c0 .t/ D a C bebt c.t/ ;
i.e. c must satisfy
ebt c0 .t/ D a ;
hence
Z t
a
c.t/ D a ebs ds D .1 ebt /
0 b
so that ebt c.t/ D a

b .ebt 1/ and, as v.0/ D EŒ0 D x, we find
a bt
EŒt D v.t/ D ebt x C .e 1/ : (S.60)
b
b1) Immediate from (S.60).
b2) (S.60) can be written as
a bt a
EŒt D x C e ,
b b
therefore if x C a
b D 0, the expectation is constant and
a
EŒt Dx:
b
b3) Again by (S.60) limt!C1 EŒt D ˙1 according to the sign of x C ab .
9.11
a1) The process C must satisfy
d.0 .t/Ct / D .a C b0 .t/Ct / dt C . C 0 .t/Ct / dBt

D Ct 0 .t/ b dt C dBt C a dt C dBt :
Let us assume that dCt D t dtCt dBt for some processes ; to be determined.
We have d0 .t/ D b0 .t/ dt C 0 .t/ dt, hence, by Ito’s formula,
d.0 .t/Ct / D Ct d0 .t/ C 0 .t/ dCt C dhC; 0 it

D Ct 0 .t/ b dt C dBt C 0 .t/ dCt C t 0 .t/ dt :
Therefore C must satisfy
0 .t/ dCt C t 0 .t/ dt D 0 .t/.t dt C t dBt / C t 0 .t/ dt D a dt C dBt :
From this relation we obtain
t D 0 .t/1
t D .a /0 .t/1
and, recalling the expression of 0 ,

Z t 2
Z t
2
.b 2 /s Bs
Ct D .a / e ds C e.b 2 /s Bs dBs ;
0 0
which gives for (9.52) the general solution
t D 0 .t/x C 0 .t/Ct
Z t
2 2
D e.b 2 /tC Bt x C .a / e.b 2 /s Bs ds (S.61)
0Z
t
2

C e.b 2 /s Bs dBs :
0
a2) If D 0 (S.61) becomes

Z t Z t
bs
t D e bt
xCa e ds C ebs dBs :
0 0
The stochastic integral in this expression has a deterministic integrand and

is therefore Gaussian. The other terms appearing are also deterministic and
therefore is Gaussian itself. We have
a
EŒt D ebt x C .ebt 1/
Z t b
2 2bt (S.62)
Var.t / D 2 e2b.ts/ ds D .e 1/ :
0 2b
a3) being Gaussian it suffices to investigate the limits as t ! C1 of its mean
and its variance. (S.62) for b < 0 give
a 2
lim EŒt D ; lim Var.t / D
t!C1 b t!C1 2b
2
Therefore, t converges in law as t ! C1 to an N. b a
; 2b / distribution.
Observe that the mean of this distribution is the point at which the drift vanishes.
b) If Ito’s formula could be applied to the function log (but it cannot, as it is not
even defined on the whole real line) we would obtain
1 1 2
d.log Yt / D dYt 2 dhYit D .b C
log Yt / dt C dBt dt : (S.63)
Yt 2Yt 2
Therefore t D log Yt would be a solution of
2

dt D b 2
C
t dt C dBt
(S.64)
0 D x D log y ;
which is a particular case of (9.52). Once these heuristics have been performed,
it is immediate to check that if is a solution of (S.64) then Yt D et is a
solution of (S.63). We have therefore proved the existence of a solution of the
SDE (9.53), but this equation does not satisfy the existence and uniqueness
results of Chap. 9 (the drift does not have a sublinear growth) so that uniqueness
is still to be proved.
Let Y be a solution of (9.53) and " its exit time from the interval "; 1" Œ or,
which is the same, the exit time of log Yt from log 1" ; log 1" Œ. Let f" be a C2
function on R that coincides with log on "; 1" Œ. Then we can apply Ito’s formula
to f" .Yt /, which gives that
Z " ^t 2

log Y" ^t D log y C b 2 C
log.Y" ^s / ds C B" ^t :
0
Therefore by localization (Theorem 9.3) log Yt coincides with the solution

of (S.64) until the exit from the interval log 1" ; log 1" Œ. But Eq. (S.64) is a
nice equation with Lipschitz continuous coefficients and hence has a unique
solution. As we know that " ! C1 as " ! 0 (Remark 9.3), the uniqueness
of the solution of (S.64) implies the uniqueness of the solution of (9.53).
As Yt D et , necessarily Yt > 0 for every t a.s. Thanks to a), Yt has a
lognormal law (see Exercise 1.11 for its definition) with parameters t and t2
that can be obtained from (S.62) i.e.
b 12 2
t
t D e
t x C .e 1/

2 2
t
t2 D .e 1/ :
2
If
< 0,
1 2 2
lim t D b ; lim t2 D
t!C1
2 t!C1 2
and .Yt /t converges in law as t ! C1 to a lognormal distribution with these

parameters.
9.12
2
a) Let us apply Ito’s formula to the function u.x; t/ D e2 t .x2 C 2
/, so that dZt D
du.t ; t/. Recall the formula (see Remark 9.1)
@u @u
du.t ; t/ D C Lu .t ; t/ dt C .t ; t/ dBt ;
@t @x
where
2 @2 u @u
Lu.x; t/ D 2
.x; t/ C x .x; t/
2 @x @x
is the generator of the Ornstein–Uhlenbeck process. As
@u @u @2 u
D 2 u.x; t/; D 2xe2 t ; D 2e2 t
@t @x @x2
we have
Lu D 2 u.x; t/
and the term in dt of the stochastic differential vanishes. Therefore
@u
dZt D du.t ; t/ D .t ; t/ dBt D e2 t 2t dBt
@x
and .Zt /t is a martingale, being the stochastic integral of a process of M 2 .

b) We have
2 ,
EŒZt D Z0 D x2 C
2
hence
2
EŒYt D e2 t EŒZt D e2 t x2 C ;
2
which entails that limt!C1 EŒYt is equal to C1 or to 0 according as > 0 or

< 0.
9.13
a) The coefficients are Lipschitz continuous, therefore we have strong existence
and uniqueness. p p
b1) We have Yt D f .t / with f .z/ D log 1 C z2 C z . As 1 C z2 C z > 0 for
every z 2 R, f W R ! R is differentiable infinitely many times and
1 z 1
f 0 .z/ D p p C1 D p
2
1Cz Cz 1Cz2 1 C z2
z
f 00 .z/ D
.1 C z2 /3=2
Ito’s formula gives
1 t
dYt D p dt dhit
1 C t2 2.1 C t2 /3=2
1 q 1 t
D p 1 C t2 C t dt C dBt p dt
1 C t2 2 2 1 C t2
D dt C dBt :
p
Therefore, as Y0 D log 1 C x2 C x ,
p
Yt D log 1 C x2 C x C t C B t :
b2) We have t D f 1 .Yt /, assuming that f is invertible. Actually if y D f .z/, then

p
ey D 1 C z2 C z ;
p
i.e. ey z D 1 C z2 and taking the square e2y 2zey C z2 D 1 C z2 , i.e.
e2y 1 D 2zey , and finally
ey ey
zD D sinh y
2
so that
p
t D sinh log 1 C x2 C x C t C Bt :
9.14
a) If t D .t ; t /, we can write
dt D t dt C ˙ dBt
with the initial condition 0 D z D .x; y/, where

p0
˙D :
1 2
We know (Example 9.1) that the solution of this SDE is

Z t
t t
t D e zCe e s ˙ dBs
0
whence we obtain that t is Gaussian with mean e t z and covariance matrix
1
.1 e2 t / ˙˙ : (S.65)
2
Now
2 2
p0 p
˙˙ D D ;
1 2 0 1 2 2 2
hence each of the r.v.’s t ; t is Gaussian with variance
2
.1 e2 t / (S.66)
2
and mean e t x and e t y respectively. In particular, their (marginal) laws

coincide and do not depend on .
b) Their joint law has already been determined: it is Gaussian with covariance
matrix given by (S.65), i.e.

2 1
.1 e2 t / :
2 1
Their covariance is therefore equal to
2
.1 e2 t / ;
2
which is maximum for D 1.
A necessary and sufficient condition for a Gaussian law to have a density
with respect to Lebesgue measure is that the covariance matrix is invertible. The
determinant of ˙˙ is equal to 2 .1 2 /: the covariance matrix is invertible if
and only if 6D ˙1.
c)
2 @2 @2 @2 @ @
LD C C 2 x y
2 @x2 @y2 @x@y @x @y
9.15
a) Here the drift is b.x/ D . 12 x1 ; 12 x2 / whereas the diffusion coefficient is

x2
.x/ D
x1
whence the generator of the diffusion is the second-order operator

2 2
1X @2 X @
LD aij .x/ C bi .x/ ,
2 i;jD1 @xi @xj iD1
@x i
where
2
x2 x2 x2 x1
a.x/ D .x/.x/ D x2 x1 D ;
x1 x2 x1 x21
i.e.
1 2 @2 1 @2 @2 1 @ 1 @
LD x2 2 C x21 2 x1 x2 x1 x2
2 @x1 2 @x2 @x1 @x2 2 @x1 2 @x2
It is immediate that det a.x/ D 0 for every x, hence L is not elliptic. The lack
of ellipticity was actually obvious from the beginning as the matrix above has
rank 1 so that necessarily a D has rank 1 at most. This argument allows us
to say that, in all generality, the generator cannot be elliptic if the dimension of
the driving Brownian motion (1 in this case) is strictly smaller than the dimension
of the diffusion.
b) Let us apply Ito’s formula: if f .x/ D x21 C x22 , then dYt D df .t / and
@f @f 1 @2 f @2 f
df .t / D .t / d1 .t/ C .t / d2 .t/ C 2
.t /dh1 it C 2 .t /dh2 it
@x1 @x2 2 @x1 @x2
as the mixed derivatives of f vanish. Replacing and keeping in mind that dh1 it D
2 .t/2 dt, dh2 it D 1 .t/2 dt, we obtain
df .t /
1

D 21 .t/ .t/dt
2 1
2 .t/ dBt C 22 .t/ 12 2 .t/dt C 1 .t/ dBt

C 2 .t/2 C 1 .t/2 dt
D0
so that the process Yt D 1 .t/2 C 2 .t/2 is constant and Y1 D 1 a.s. The process
t D .1 .t/; 2 .t// takes its values in the circle of radius 1.
9.16
a) We note that .Yt /t is a geometric Brownian motion and that the pair .Xt ; Yt / solves
the SDE
dXt D Yt dt
dYt D Yt dt C Yt dBt
with the initial conditions X0 D x, Y0 D 1. Its generator is
1 2 2 @2 @ @
LD y 2
Cy C y
2 @y @x @y
b) Ito’s formula (assuming for an instant that it is legitimate to use it) gives
Xt Xt 1 Xt Xt
dt D 2
dYt C 3 dhYit C dXt D 2 Yt dt C dBt C 3 2 Yt2 dt Cdt :
Yt Yt Yt Yt Yt
Actually dhX; Yit D 0. This gives

dt D 1 C . 2 /t dt t dBt : (S.67)
However, we are not authorized to apply Ito’s formula, as the function W

.x; y/ 7! xy is not defined on the whole of R2 . We can, however, reproduce here
an idea developed earlier: let, for every " > 0, " be a function coinciding with
on fjyj > "g and then extended so that it is C2 . Let " be the first time that
Yt < ". Then the computation above applied to the function " gives
Z t^" Z t^"

t^" D x C 1 C . 2 /u du u dBu :
0 0
As P.Yt > 0 for every t 0/ D 1, we have " ! C1 as " ! 0 and, taking the
limit as " ! 0 in the relation above, we find
Z Z
t t
t D x C 1 C . 2 /u du u dBu :
0 0
c) Equation (9.55) is a particular case of (S.67) with D 2 r. With this choice

2 2 2
of we have 2 D 2
r and Yu D e. 2 r/uC Bu
so that a solution is, taking
x D z,
Z t
Xt 2 2
Zt D D e. 2 r/t Bt z C e. 2 r/uC Bu
du :
Yt 0
Recalling the expression of the Laplace transform of a Gaussian r.v. we have, for
r 6D 0,
Z t
2 2
EŒZt D z EŒe. 2 r/t Bt
C EŒe. 2 r/.tu/ .Bt Bu /
du
0
Z t
2 2 2 2
r/tC 2 t r/.tu/C 2 .tu/
D z e. 2 C e. 2 du
0
Z t
1 rt
D ze C rt
er.tu/ du D z ert C .e 1/ :
0 r
If r > 0 we have EŒZt 1r if z D 1r and

(
C1 if z > 1r
lim EŒZt D
t!C1 1 if z < 1r :
If r < 0 then limt!C1 EŒZt D 1r . whatever the starting point z.

If r D 0 the computation of the integrals above is different and we obtain
EŒZt D t C z ;
which converges to C1 as t ! C1.
9.17
a) Ito’s formula would give
1 1 1 1 1
dZt D 2
dt C 3 dhit D 2 t .a bt / dt 2 t dBt C 3 2 t2 dt ;
t t t t t
i.e.

dZt D b .a 2 /Zt dt Zt dBt (S.68)
1
at least as soon as remains far from 0. Of course Z0 D z with z D x > 0.
b) Exercise 9.11 gives for (S.68) the solution
2 Z t
2
.a 2 /t Bt
Zt D e zCb e.a 2 /sC Bs ds :
0
As z > 0, clearly Zt > 0 for every t a.s. We can therefore apply Ito’s formula
and compute the stochastic differential of t 7! Z1t . As the function z ! 1z is not
everywhere defined, this will require the trick already exploited in other exercises
(Exercise 9.16 e.g.): let " be the first time Zt < " and let " be a C2 function
coinciding with z 7! 1z for z ". Then Ito’s formula gives
Z
t^"
0 1
" .Zt^" / D " .z/ C " .Zs /.b .a 2 /Zs / C 00 2 2
" .Zs / Zs ds
0 2
Z t^"
0
C " .Zs /Zs dBs :
0
1
As " .z/ D z
for z ", we have for " < z
Z Z t^"
1 1 t^"
1 2 1 2 2 1
D C 2
.b .a /Z s / C 3
Z ds C Zs dBs
Zt^" z 0 Zs Zs s
0 Zs2
Z Z t^"
1 t^"
1 1 1
D C ab ds C dBs :
z 0 Zs Zs 0 Zs
1
As Zt > 0 for every t, we have " ! C1 as " ! 0 and the process t D Zt
satisfies (9.56). Hence (9.56) has the explicit solution
1 Z 1
1 2
t
2
t D D e.a 2 /tC Bt Cb e.a 2 /sC Bs
ds
Zt x 0
Of course, t > 0 for every t with probability 1.
9.18
a) Of course the pair .Xt ; Yt /t solves the SDE
dt D t dt C dBt
dt D t dt
with the initial conditions 0 D x, 0 D 0. This is an SDE with Lipschitz

continuous coefficients with a sublinear growth and therefore has a unique
solution. The generator is
2 @2 @ @
LD 2
C y Cx
2 @x @x @y
The process Zt D .t ; t / is the solution of the SDE
dZt D MZt dt C ˙ dBt ;
where M; ˙ are the matrices

0 0
MD ; ˙D :
10 00
We recognize the Ornstein–Uhlenbeck process of Example 9.1: a linear drift

and a constant diffusion matrix. Going back to (9.5), the solution is a Gaussian
process and we have the explicit formula
Z t
Zt D eMt z C eMt eMs ˙ dBs ;
0
where z D . 0x /.
b) We have

0a 0a ab 0
M2 D D
b0 b0 0 ab
and by recurrence

2n .ab/n 0 2nC1 0 a.ab/n
M D M D :
0 .ab/n b.ab/n 0
Therefore the elements in the diagonal of eM are

X1 X1 p
.ab/n . ab /2n p
D D cosh. ab/
nD0
.2n/Š nD0
.2n/Š
whereas the .1; 2/-entry is

1
X
r 1 p r
.ab/n a X . ab /2nC1 a p
a D D sinh. ab/ :
nD0
.2n C 1/Š b nD0 .2n C 1/Š b
A similar computation allows us to obtain the .2; 1/-entry and we find

p pa p !
q cosh ab b sinh ab
e D
M
b
p p
a
sinh ab cosh ab
and (9.57) follows from the relations cosh ix D cos x, sinh ix D i sin x.
c) The mean of Zt is eMt z with z D . 0x /. We have

0 t
Mt D
t 0
and, thanks to b2),

p p p !
cosh. t/ sinh. t/
e Mt
D p p
p1 sinh. t/ cosh. t/

p
so that EŒt D cosh. t/x. If > 0 this tends to ˙1 according to the sign
p
of x. If < 0 then EŒt D cos. t/x. This quantity oscillates between the
values x and x with fast oscillations if j j is large. The mean remains identically
equal to 0 if x D 0.
Let us now look at the variance of the distribution of t : we must compute
Z t
.tu/
eM.tu/ ˙˙ eM du :
0
As
2
0
˙˙ D
0 0
let us first compute the integrand:

p p p !

cosh. .t u// sinh. .t u// 2 0
e M.tu/
˙˙ D p p
p1 sinh. .t u// cosh. .t u// 0 0

p !
2 cosh. .t u// 0
D 2 p
p sinh
.t u/ 0
so that
M .tu/
eM.tu/ ˙ ˙ e
p ! p p !
1
2 cosh. .t u// 0 cosh. .t u// p sinh. .t u//

D 2 p p p p
p sinh. .t u// 0 sinh. .t u// cosh. .t u//

p 2 p p !
2 cosh2 . .tu// p

cosh. .tu// sinh. .tu//
D 2 p p 2 p :
p

cosh. .tu// sinh. .tu//
sinh2 . .tu//
The variance of t is therefore equal to

Z Z
t
p t
p
2 cosh2 . .t u// du D 2 cosh2 . u/ du
0 0
and it diverges as t ! C1. If < 0 the integral grows linearly fast (the
p
integrand becomes cos2 . u/ and is bounded), whereas if > 0 it grows
exponentially fast.
9.19 By Ito’s formula
dt D 2t dt C 2 dt D . 2 C 2bt2 / dt C 2t dBt :
If we define
Z t
s
Wt D p 1fs 6D0g C 1fs D0g dBs (S.69)
0 s
then, by Corollary 8.2, W is a real Brownian motion (the integrand is a process

having modulus equal to 1) and
p p t
t dWt D t p 1ft 6D0g C 1ft D0g dBt D t dBt :
t
Hence solves the SDE

p
dt D . 2 C 2bt / dt C 2 t dWt (S.70)
of course with the initial condition 0 D x2 .

• Note that the coefficients of (S.70) do not satisfy Condition (A’) (the diffusion
coefficient is not Lipschitz continuous at 0). In particular, in this exercise we
prove that (S.70) has a solution but are unable to discuss uniqueness.
9.20 Given the process

Z t
t D .s / dBs
0
we
R t know that there exists a Brownian motion W such that t D WAt , where At D
2
0 .t / dt. By the Iterated Logarithm Law therefore there exist arbitrarily large
values of t such that
p
t D WAt .1 "/ 2At log log At / :
As At c2 t, t takes arbitrarily large values a.s. and therefore exits with probability
1 from any bounded interval.
As we are under hypotheses of sublinear growth, the process and also ..t //t
belong to M 2 and therefore is a square integrable martingale. By the stopping
theorem therefore
E. ^t / D 0
for every t > 0. Taking the limit with Lebesgue’s theorem (t 7! ^t is bounded as
^t 2 Œa; b) we find
0 D E. / D aP. D a/ C bP. D b/
and, replacing P. D a/ D 1 P. D b/,

a
P. D b/ D
aCb
Note that this result holds for any , provided it satisfies the conditions indicated
in the statement. This fact has an intuitive explanation: the solution being a time
changed Brownian motion, its exit position from a; bŒ coincides with the exit
position of the Brownian motion.
9.21
a) Let us first look for a function u such that t 7! u.t / is a martingale. We know
(see Remark 9.1) that u must solve Lu D 0, L denoting the generator of the
Ornstein–Uhlenbeck process, i.e.
2 00
u .x/ xu0 .x/ D 0 :
2
2 2
2
If v D u0 , then v must solve v 0 .x/ D 2
x v.x/, i.e. v.x/ D c1 e 2 x , so that the
general integral of the ODE above is
Z x 2 2
u.x/ D c1 e 2 y dy Cc2 (S.71)
z0
„ ƒ‚ …
WDF.x/
for any z0 2 R. As u is a regular function, with all its derivatives bounded in

every bounded interval, t 7! u.t^ / is a martingale. This already proves that
< C1 a.s. Actually this is a bounded martingale, therefore a.s. convergent,
and convergence cannot take place on D C1 (see Sect. 10.2 for another
proof of this fact).
We have therefore, for every starting position x, u.x/ D Ex Œu.t^ / and
taking the limit as t ! C1 with Lebesgue’s theorem, we find
c1 F.x/ C c2 D u.x/ D Ex Œu. / D u.b/Px . D b/ C u.a/.1 Px . D b//

D c1 F.b/ F.a/ Px . D b/ C c1 F.a/ C c2 :
Let us choose z0 D a in (S.71), so that F.a/ D 0. Then the previous relation

becomes
Z x
2 2
e 2 z dz
F.x/
Px . D b/ D D Zab (S.72)
F.b/ 2 2
2 z
e dz
a
p
b1) With the change of variable y D z we have
p
Z a Z a
z2 2
e dz ey dy
lim Z0 b D lim Z0 p :
!C1 !C1 b
z2 y2
e dz e dy
0 0
Taking the derivative with respect to and applying L’Hospital rule this limit
is equal to
2
a e a
lim 2
,
!C1 b e b
which is equal to C1 if a > b.

b2) Using the symmetry of the integrand we can write (S.72) as
Z a Z x
2 2 2 2
e 2 z dz C e 2 z dz
Px . D b/ D Z0 a Z
0
b
2 2 2 2
z
e 2 dz C e 2 z dz
0 0
R a 2 2
and now one only has to divide numerator and denominator by 0 e 2 z dz: as
both b and jxj are smaller then a, the limit as ! C1 turns out to be equal to
1 thanks to b1).
b3) For large the process is affected by a large force (the drift x) taking it
towards 0. Therefore its behavior is as follows: the process is attracted towards
0 and stays around there until some unusual increment of the Brownian motion
takes it out of a; bŒ. The exit takes place mostly at b because this is the
closest to 0 of the two endpoints.
9.22
a) We know (see Example 9.1) that the law of t" is Gaussian with mean e t x and
variance
"2 2
.1 e2 t / :
2
By Chebyshev’s inequality
"2 2
P.jt" e t xj ı/ .1 e2 t / ! 0
2 ı 2 "!0
so that t" converges to e t x in probability and therefore in distribution.

b) It is sufficient to prove that

P sup jt" e t xj ı ! 0 :
0tT "!0
Actually this entails that the probability for " to be outside of a fixed neighbor-
hood of the path x0 .t/ D e t goes to 0 as " ! 0. Recall the explicit expression
for " :
Z t
t" D e t x C "e t e s dBs :
0
Therefore we are led to the following limit to be proved

ˇ Z t ˇ
ˇ t ˇ
P " sup ˇe e s dBs ˇ ı ! 0
0tT 0 "!0
which is immediate since the process inside the absolute value is continuous and
therefore the r.v.
ˇ Z t ˇ
ˇ ˇ
sup ˇe t e s dBs ˇ
0tT 0
is finite. Note that, using the exponential inequality of martingales we might

even give a speed of convergence. This will be made explicit in the general case
of Exercise 9.28.
9.23 The function f .y/ D 1xy x

is such that f .0/ D x and f 0 .y/ D f .y/2 , f 00 .y/ D
2f .y/3 . Therefore if it was legitimate to apply Ito’s formula we would have
1
df .Bt / D f 0 .Bt / dBt C f 00 .Bt / dt D f .Bt /2 dBt C f .Bt /3 dt
2
so that t D f .Bt / would be a solution of (9.60). This is not correct formally because
f is not C2 .R/ (it is not even everywhere defined). In order to fix this point, let us
assume x > 0 and let f" be a function that coincides with f on 1; 1x " and then
extended so that it is C2 .R/. If we denote by " the passage time of B at 1x ", then
Ito’s formula gives, for t D f" .Bt /,
Z t^" Z t^"
t^" D x C s3 ds C s2 dBs :
0 0
Therefore t D f .Bt / is the solution of (9.60) for every t " . Letting " & 0,
" % x and therefore .t /t is the solution of (9.60) on Œ0; x Œ.
• Note that limt!x t D C1 and that, by localization, any other solution must
agree with on Œ0; " Œ for every ", hence it is not possible to have a solution
defined in any interval Œ0; T. This is an example of what can happen when the
sublinear growth property of the coefficients is not satisfied.
9.24
a) The matrix of the second-order coefficients of L is

10
aD
00
whose square root, , is of course equal to a itself. a is not invertible, hence L is

not elliptic and a fortiori not uniformly elliptic. The corresponding SDE is then
dt D bt dt C dBt ;
where

10 00
D bD
00 10
and we know, by Example 9.1, that this SDE has the explicit solution
Z t
t D ebt x C ebt ebs dBs
0
and has a Gaussian law with mean ebt x and covariance matrix
Z t

t D ebs eb s ds :
0
We see that b2 D b3 D D 0. Hence

1
X
bk uk 10
e bu
D D I C bu D
kŠ u1
kD0
so that

u 10 10 1u 1 u
ebu eb D D
u1 00 01 u u2
and
!
t2
t
t D 2
t2 t3
:
2 3
t3
t being invertible for t > 0 (its determinant is equal to 12
) the law of t has a
density with respect to Lebesgue measure.
b) The SDE now is of the same kind but with

00 1 0
bD ; ebu D :
01 0 eu
Therefore

u 1 0 10 1 0 10
ebu eb D D
0 eu 00 0 eu 00
and

t 0
D ;
00
which is not invertible, so that there is no density with respect to Lebesgue

measure.
9.25
a) Let B be an m-dimensional Brownian motion and a square root of a. The SDE
associated to L is
dt D bt dt C dBt

0 D x :
We have seen in Example 9.1 that its solution is

Z t
t D e x C e
bt bt
ebs dBs :
0
The law of t is therefore Gaussian with mean ebt x and, recalling (S.52),
covariance matrix
Z t

t D ebu aeb u du : (S.73)
0
As the transition function p.t; x; / it is nothing else than the law of t with the
initial condition 0 D x, the transition function has density if and only if
is invertible (see Sect. 1.7 and Exercise 1.4). If a is positive definite then there
exists a number > 0 such that, for every y 2 Rm , hay; yi jyj2 . Therefore, if
y 6D 0,
Z t Z t Z t
b u b u b u
ht y; yi D he ae
bu
y; yi du D hae y; e yi du jeb u yj2 du > 0 :
0 0 0

Actually if y 6D 0, we have jeb u yj > 0, as the exponential of a matrix is
invertible.
b) Let us first assume that there exists a non-trivial subspace contained in the kernel
of a and invariant with respect to the action of b . Then there exists a non-zero
vector y 2 ker a such that b i y 2 ker a for every i D 1; 2; : : : It follows that also

eb u y 2 ker a. Therefore for such a vector we would have
Z t

ht y; yi D haeb u y; eb u yi du D 0
0
and t cannot be invertible.

Conversely, if t were not invertible, there would exist a vector y 2 Rm , y 6D 0,

such that ht y; yi D 0. As haeb u y; eb u yi 0, necessarily haeb u y; eb u yi D 0
for every u t and therefore

aeb u y D 0
for every u t. For u D 0 this relation implies y 2 ker a. Taking the derivative
with respect to u at u D 0 we find that necessarily ab y D 0. In a similar way,
taking the derivative n times and setting u D 0 we find ab n y D 0 for every n.
The subspace generated by the vectors y; b y; b 2 y; : : : is non-trivial, invariant
under the action of b and contained in ker a.
9.26 Let us compute the differential of with Ito’s formula, assuming that the
function f that we are looking for is regular enough. We have
1
dt D f 0 .t / dt C f 00 .t / 2 .t / dt
2 (S.74)
1
f 0 .t /b.t / C f 00 .t / 2 .t / dt C f 0 .t /.t / dBt :
2
Therefore, in order for (9.64) to be satisfied, we must have f 0 .t /.t / D 1. Let
Z z
1
f .z/ D dy :
0 .y/
The assumptions on entail that f 2 C2 so that Ito’s formula can be applied to

0 .x/
1
t D f .t /. Moreover, f is strictly increasing. As f 0 .x/ D .x/ and f 00 .x/ D .x/ 2,
Eq. (S.74) now becomes

1 1 0 1
1
dt D b.f .t // .f .t // dt C dBt ;
.f 1 .t // 2
„ ƒ‚ …
WDeb.t /
with the initial condition 0 D f .x/.

b) We have
Z t
Bt D t f .x/ e
b.t / dt :
0
The right-hand side is clearly measurable with respect to the -algebra .u ; u
t/, therefore Gt .u ; u t/, where Gt D .Bu ; u t/ as usual. As f is strictly
increasing, hence invertible, .u ; u t/ D .u ; u t/ D Ht . Hence Gt Ht
and as the converse inclusion is obvious (see Remark 9.4) the two filtrations
coincide.
9.27
a) Let us first admit the relations (9.68) and let us apply Ito’s formula to the process
t D h.Dt ; Bt /. As .Dt /t has finite variation,
@h @h 1 @2 h
dh.Dt ; Bt / D .Dt ; Bt /D0t dt C .Dt ; Bt / dBt C .Dt ; Bt / dt :
@x @y 2 @y2
But
@h
.Dt ; Bt / D .h.Dt ; Bt //
@y
@2 h
.Dt ; Bt / D 0 .h.Dt ; Bt //.h.Dt ; Bt //
@y2
@h h Z Bt i
.Dt ; Bt / D exp 0 .h.Dt ; s// ds :
@x 0
Moreover,
1 h Z Bt i
D0t 0
D .h.Dt ; Bt //.h.Dt ; Bt //Cb.h.Dt ; Bt // exp 0 .h.Dt ; s// ds :
2 0
Putting things together we find
dh.Dt ; Bt / D b.h.Dt ; Bt // dt C .h.Dt ; Bt // dBt ;
hence Zt D h.Dt ; Bt / is a solution of (9.66), which concludes the proof of a).

Let us check the two relations of the hint. We can write
@h
.x; y/ D .h.x; y//
@y
h.x; 0/ D x :
As we admit that h is twice differentiable in y,
@2 h @h
.x; y/ D 0 .h.x; y// .x; y/ D 0 .h.x; y//.h.x; y// :
@y2 @y
Similarly, taking the derivative with respect to x,
@2 h @h
.y; x/ D 0 .h.x; y// .x; y/
@y@x @x
@h
.x; 0/ D 1 :
@x
@h
Hence, if g.y/ D @x .x; y/, then g is the solution of the linear problem
g0 .y/ D 0 .h.x; y//g.y/

g.0/ D 1
whose solution is
hZ y i
g.y/ D exp 0 .h.x; s// ds :
0
b) Let us denote by e f the analogue of the function f with eb replacing b. Then clearly
e
f .x; z/ f .x; z/ for every x; z. If e
D is the solution of
e
D0t D e
f .e
Dt ; B t /
e
D0 D ex;
then e
Dt Dt for every t 0. As h is increasing in both arguments we have
e
t D h.e
Dt ; Bt / h.Dt ; Bt / D t :
9.28
a) We have
d"t D dt" t0 dt D .b.t" / b.t // dt C ".t" / dBt

D .b."t C t / b.t // dt C "."t C t / dBt :
ˇ Rs ˇ
b) If sup0sT ˇ" 0 ."u C u / dBu ˇ < ˛, we have, for t T,
Z t Z t
j"t j jb."s C s / b.s /j ds C ˛ L j"s j ds C ˛
0 0
and, by Gronwall’s inequality, sup0tT j"t j ˛eLT . Taking the complement we

have

P sup js" s j ˛ D P sup j"s j ˛
0sT 0sT
ˇZ s ˇ ˛ h 1 ˛ 2 e2LT i
ˇ ˇ
P sup ˇ ."u C u / dBu ˇ eLT 2m exp 2
0sT 0 " " 2mTkk21
thanks to the exponential inequality (8.42). As "s D s" s and by the

arbitrariness of ˛, this relation implies that " , as " ! 0, converges in probability
to a constant r.v. taking the value with probability 1.
c) Let us assume that Assumption (A’) holds. For fixed t, let R be the radius of some
ball centered at the origin of Rm and containing t for every t T. Let e b; e
be
a vector and a matrix field, respectively, bounded and Lipschitz continuous and
coinciding with b; , respectively, for jxj R C 1.
Let us denote by e " the solution of an SDE with coefficients e b and "e
and
with the same initial condition e "0 D x. We know then (Theorem 9.3) that the
two processes " and e " coincide up to the exit out of the ball of radius R C 1. It
follows that, for ˛ 1, the two events
n o n o
sup jt" t j ˛ and sup je
"t t j ˛
0tT 0tT
are a.s. equal. Therefore
LT i
h 1 ˛ 2 e2e
P sup jt" t j ˛ D P sup je "t t j ˛ 2m exp 2 :
0tT 0tT " 2mTke k21
9.29
a) Let k D sup0tT kGt Gt k, where k k denotes the norm as an operator, so that
hGt Gt
;
i k for every vector
of modulus 1. By Proposition 8.7
ˇZ t ˇ
ˇ ˇ 2
P sup ˇ Gs dBs ˇ 2mec0 ;
0tT 0
˚ ˇRt ˇ
where c0 D .2Tmk /1 . If we define A D sup0tT ˇ 0 Gs dBs ˇ < , then on
A we have, for t T,
Z t
jXt j jxj C M .1 C jXs j/ ds C ;
0
i.e.
Z t
Xt .jxj C MT C / C M Xs ds ;
0
and by Gronwall’s Lemma XT .jxj C MT C / eMT . Therefore, if jxj K,
ˇZ t ˇ
ˇ ˇ 2
P XT > .K C MT C / e MT
P sup ˇ Gs dBs ˇ 2mec0 :
0tT 0
Solving R D .K C MT C / eMT we have D ReMT .jxj C MT/ and therefore

2
P.XT > R/ 2m exp c0 ReMT .K C MT/ ;
from which we obtain that, for every constant c D cT strictly smaller than
e2MT ,
c0 e2MT D (S.75)
2Tmk
the inequality (9.70) holds for R large enough.
b) It is an obvious consequence of a), with Ft D b.t ; t/; Gt D .t ; t/.
c) If u.x/ D log.1 C jxj2 /, let us compute with patience the derivatives:
2xi
uxi .x/ D
1 C jxj2 (S.76)
2ıij 4xi xj
uxi xj .x/ D
1 C jxj2 .1 C jxj2 /2
In particular, as jxj ! C1 the first-order derivatives go to 0 at least as x 7! jxj1

and the second-order derivatives at least as x 7! jxj2 . By Ito’s formula the
process Yt D log.1 C jt j2 / has stochastic differential
X
m X
m
dYt D uxi .t /bi .t ; t/ C uxi xj .t /aij .t ; t/ dt
iD1 i;j
X
m X
d
C uxi .t /ij .t ; t/ dBj .t/
iD1 jD1
where a D . By (S.76), as b and are assumed to have sublinear growth, it

is clear that all the terms uxi bi , uxi xj aij , uxi ij are bounded. We can therefore apply
a), which guarantees that there exists a constant c > 0 such that, for large ,
2
P sup log.1 C jt j2 / log ec.log / ;
0tT
p 2 p
i.e. P T 1 ec.log / . Letting R D 1, i.e. D R2 C 1, the
inequality becomes, for large R,
2 C1//2 2 1
P.T R/ ec.log.R ec.log R/ D (S.77)
Rc log R
d) By Exercise 1.3, if p 1,
Z C1
EŒ.T /p D p tp1 P.T t/ dt :
0
Under Assumption (A’), by (S.77), the function t 7! P.T t/ tends to zero

faster than every polynomial as t ! C1 and therefore the integral converges
for every p 0. If, moreover, the diffusion coefficient is bounded, then
Z C1
2 2
EŒe˛.T / D 1 C 2˛te˛t P.T t/ dt
0
and again, by (9.70) and (S.75), for ˛ < e2MT .2Tmk /1 the integral is
convergent. Here k is any constant such that j.x/
j2 k for every x 2 Rm

and j
j D 1. A fortiori EŒe˛T < C1 for every ˛ 2 R.
10.1
a) Just apply the representation formula (10.6), recalling that here 0, Zt 1
and f 1.
b) If m D 1 then (10.39) becomes
(
1 00
2u D 1 on 1; 1Œ
u.1/ D u.1/ D 0
and has the solution u.x/ D 1 x2 . If m 2, as indicated in the hint, if we look

for a solution of the form u.x/ D g.jxj/, we have
1 1 m1 0
4u.x/ D g00 .jxj/ C g .jxj/ : (S.78)
2 2 2jxj
We therefore have to solve

1 00 m1 0
g .y/ C g .y/ D 1 (S.79)
2 2y
with the boundary condition g.1/ D 0 (plus another condition that we shall see
later). Letting v D g0 , we are led to the equation
1 0 m1
v .y/ C v.y/ D 1 : (S.80)
2 2y
Separating the variables, the homogeneous equation is
v 0 .y/ m1
D ,
v.y/ y
i.e. v.y/ D c1 ymC1 . With the method of the variation of the constants, let us
look for a solution of (S.80) of the form c.y/ymC1 . We have immediately that c
must satisfy
c0 .y/ymC1 D 2
and therefore c.y/ D m2 ym . In conclusion, the integral of (S.80) is v.y/ D

c1 ymC1 m2 y. Let us compute the primitive of v. If m 3 we find that the
general integral of (S.79) is
1 2
g.y/ D c1 ymC2 y C c2 :
m
The condition g.1/ D 0 gives c1 C c2 D m1 . However, we also need g to be

defined at 0, i.e. c1 D 0. In conclusion, the solution of (10.39) is
1 1
u.x/ D jxj2
m m
and, in particular, u.0/ D m1 . If m D 2 the computation of the primitive of v is

different and the general integral of (S.79) is
1 2
g.y/ D c1 log y y C c2
m
but the remainder of the argument is the same.
10.2
a) Follows from Propositions 10.2 and 10.1 (Assumption H2 is satisfied).
b) It is immediate that the constants and the function v.x/ D x12ı are solutions of
the ordinary equation
1 00 ı
v .x/ C v 0 .x/ D 0
2 x
for every x > 0. Therefore, if we denote by u.x/ the term on the right-hand side
in (10.40), as it is immediate that u.a/ D 0 and u.b/ D 1, u is the solution of the
problem
8
< 1 u00 .x/ C ı u0 .x/ D 0 x 2a; bŒ
2 x
:
u.a/ D 0; u.b/ D 1 :
Therefore, by (10.7), u.x/ D Px .X D b/.

c) Thanks to Exercise 8.24, we know that the process t D jBt C xj is a solution of
the SDE
m1
dt D dt C dWt
2t
0 D jxj
where W is a real Brownian motion. Therefore, denoting by Qjxj the law of ,
Px .jB j D b/ D Qjxj .X D b/ :
Comparing with the result of a), we have ı D m1

2 hence D 2ı 1 D m 2
which gives
1 . ax /m2
Px .jB j D b/ D
1 . ab /m2
As m ! 1 this probability converges to 1 for every starting point x, a < jxj < b.
10.3
a1) The solution is clearly x D Bt t C x and we can apply the usual Iterated
Logarithm argument.
2
a2) Lu D 2 u00 u0 . We know that the function u.x/ D P.x D b/ is the solution
of
8
< 2 00
Lu.x/ D u u0 D 0
2 (S.81)
:
u.a/ D 0; u.b/ D 1 :
Setting v D u0 we have

v0 D v;
2
i.e.
2
v.x/ D ce 2 x ;
and therefore
Z x 2
u.x/ D c1 e 2 z dz :
x0
We must enforce the constraints u.a/ D 0; u.b/ D 1, i.e. x0 D a and
1 2 1
c1 D R 2
D
b
e 2
z
dz 2 e 22 b e 22 a
a
Therefore
Z 2
0 2 2 2 1 e 2 a
P.0 D b/ D u.0/ D c1 e 2
z
dz D c1 1 e 2 a D 2 2

a 2 e 2 b e 2 a
(S.82)
b1) The generator of is
2 1
L2 u.x/ D 2
u00 .x/ 2
u0 .x/ D Lu.x/ :
2.1 C x / 1Cx 1 C x2
As the coefficient of the second-order derivative is const > 0 on the interval

considered and the coefficient of the first-order derivative is bounded, we know,
by Assumption H2 p. 308, that < C1 a.s. (and is even integrable) whatever
the starting point x.
b2) We know that the function u.x/ D P.x D b/ is the solution of
8
ˆ
<L u.x/ D 2
2 2
u00 .x/ u0 .x/ D 0
2.1 C x / 1 C x2 (S.83)
:̂u.a/ D 0; u.b/ D 1
but, factoring out the denominator 1 C x2 we see that the solution is the same
as that of (S.83), so that the exit probability is as in (S.82).
10.4 The exit law from the unit ball for a Brownian motion with starting point at
x has density (see Example 10.1)
1 1 jxj2
N.x; y/ D
2 jx yj2
with respect to the normalized one-dimensional measure of the circle. In this case
jxj D 12 so that 1 jxj2 D 34 and, in angular coordinates, we are led to the
computation of the integral
Z =2
3 1 1
d
;
4 2 =2 jx y.
/j2
where y.
/ D .cos
; sin
/. Therefore jx y.
/j2 D . 12 cos
/2 C sin2
D
5
4
cos
: The integral is computed with the change of variable
1 t2 2
cos
D , t D tan
2 , d
D dt
1 C t2 1 C t2
and therefore
Z 1 Z 1 ˇ1
3 1 2 3 1 1 ˇ
D 1t2
dt D dt D arctan.3t/ˇ
8 5
1 4 1 C t2 1 1 C 9t 2 1
1Ct2
2
D arctan 3 ' 0:795 :

10.5
a1) For " > 0 fixed, let us still denote by u a C2 .Rm / function that coincides with u
on D" . Then, by Ito’s formula,
1
t X
m
@2 u
dMt D
e
t u.Xt / dt C e
t u0 .Xt / dXt C e aij .Xt / .Xt / dt
2 i;jD1
@xi @xj

D e
t
u.Xt / C Lu.Xt / dt C e
t u0 .Xt / .Xt / dBt :
As M0 D u.x/ Px -a.s. and

u.Xt / C Lu.Xt / D 0 for t " ,
Z t^"
Mt^" D u.x/ C e
s u0 .Xs / .Xs / dBs ; Px -a.s.
0
This is a square integrable martingale, as u0 and are bounded on D" .

a2) By the relation
Ex Œe
.t^" / u.Xt^" / D Ex .Mt^" / D Ex .M0 / D u.x/ ;
taking the limit first as t ! C1 and then as " & 0 and using Fatou’s lemma
twice (recall that we assume u 0 and that u 1 on @D), we have
Ex Œe
D Ex Œe
u.X / u.x/ :
Therefore the r.v. e

is Px -integrable for every x 2 D. The r.v.’s e
.t^" / u.Xt^" /
are therefore positive and bounded above by e
maxx2D u.x/. Again taking the
limit as t ! C1 and then as " & 0 and using now Lebesgue’s theorem we
find Ex Œe
D u.x/.
b1) We must solve (10.41) for L D 12 4. The general integral of 12 u00 C
u D 0 is
p p
u.x/ D c1 cos. 2
x/ C c2 sin. 2
x/ :
p
With the conditions u.a/ D u.a/ D 1 we find c1 D cos. 2
a/1 , c2 D 0.
Therefore the solution of (10.41) is
p
cos. 2
x/
u.x/ D p
cos. 2
a/
p
As cos y > 0 for jyj < 2 , a positive solution exists if and only if 2
a < 2
2
i.e.
< 8a2 . therefore has a finite Laplace transform for these values of
.
2
Since Ex .e
/ % C1 as
% 8a 2 , this is the convergence abscissa.
b2) We have seen, in particular, that does not have a finite Laplace transform for
every
2 R and therefore cannot be bounded Px -a.s. By Markov’s inequality,
for every ˇ > 0,
Px . > R/ Px .eˇ > eˇR / eˇR Ex .eˇ / :
This relation gives an estimate of the tail of the distribution of for every
2
ˇ < 8a 2.
b3) The mean of is obtained as the derivative at 0 of the Laplace transform. As
p
d cos. 2
x/
p
d
cos. 2
a/
1 p p x p p a
D p cos. 2
a/ sin. 2
x/ p C sin. 2
a/ cos. 2
x/ p
cos2 . 2
/a 2
2
we have, for
D 0, Ex ./ D a2 x2 (the mean of the exit time has already
been computed in Exercises 5.10 and 10.1).
10.6
a) By Theorem 10.6 a solution is given by u.x; t/ D Ex;t ŒXT2 ; where Px;t is the law
of the diffusion associated to the differential generator Lu D 12 u00 with the initial
conditions x; t. With respect to Px;t , XT2 has the same law as .BTt C x/2 , where
.Bt /t is a Brownian motion. Therefore
u.x; t/ D E.B2Tt C 2xBTt C x2 / D T t C x2 :

The solution is unique in the class of solutions with polynomial growth by

Theorem 10.5.
b) If .x/ D xm then the unique solution with polynomial growth is
!
Xm
m k
u.x; t/ D EŒ.BTt C x/ D m
Tt / :
x E.Bmk
iD1
k
mk
As E.BTt / D 0 if mk is odd, whereas E.BTt
mk
/ D .T t/` E.B2`
1 / if mk D 2`,
u is a polynomial in the variables x; t. Now just observe that the solution u is
linear in .
10.7
a) Equation (10.43) can be written in the form
@u
Lu C D0;
@t
where L is the generator of a geometric Brownian motion. Hence a candidate
solution is
u.x; t/ D EŒTx;t ;
1 2 /.st/C .B
where sx;t D xe.b 2 s Bt / . Hence
u.x; t/ D xeb.Tt/ : (S.84)
We cannot apply Theorem 10.6 because the generator is not elliptic, but it is easy
to check directly that u given by (S.84) is a solution of (10.43).
b) Following the same idea we surmise that a solution might be
1 2 /.Tt/C2 .B 2 /.Tt/
u.x; t/ D EŒ.Tx;t /2 D EŒx2 e2.b 2 T Bt /
D x2 e.2bC :
Again it is easy to check that such a u is a solution of the given PDE problem.
10.8
a) Thanks to Theorem 10.6 a solution is given by
u.x; t/ D Ex;t Œcos.h

; XT i/ ;
where .C ; M ; .Mt /t ; .Xt /t ; .Px;t /x;t / denotes a realization of the diffusion process
associated to the generator L.
It is immediate that such a diffusion is an Ornstein–Uhlenbeck process and
that, with respect to Px;t , XT has a Gaussian distribution with mean e .Tt/ x and
2
covariance matrix 2 .1 e2 .Tt/ / I. Hence u.x; t/ is equal to the expectation
of cos.h
; Zi/ where Z is a Gaussian r.v. with these parameters. This is equal to
the real part of the characteristic function
.Tt/ xi
j
j2 2
EŒeih
;Zi D eih
;e exp .1 e2 .Tt/ / ;
4
i.e.
j
j2 2
u.x; t/ D cos.h
; e .Tt/ xi/ exp .1 e2 .Tt/ / :
4
b) As T ! C1 the value at time 0 of u.x; t/ converges, for every x; t, to the constant

j2 2
exp. j
4 /.
10.9 The SDE associated to L is
dt D bt dt C t dBt ; (S.85)

p
where D a. As Assumption (A) is satisfied, we know that the fundamental
solution is given by the density of the transition function, if it exists. In order to
show the existence of the transition function we cannot apply Theorem 10.7, whose
hypotheses are not satisfied since the coefficients are not bounded and the diffusion
coefficient is not elliptic. However, we know that the solution of (S.85) with the
initial condition 0 D x is
2
tx D xe.b 2
/tC Bt
(see Example 9.2). This is a time homogeneous diffusion and its transition function
p.t; x; / is the law of tx .
2
2
The r.v. e.b 2 /tC Bt is lognormal with parameters .b 2 /t and 2 t (see
Exercise 1.11) and therefore has density
1 1 2
2
g.y/ D p exp 2 log y .b 2 /t :
2t y 2 t
Its transition function p.t; x; / therefore has density
1 1 2
2
q.t; x; y/ D p exp log yx .b 2 /t
2t xy 2 2 t
and the fundamental solution is .s; t; x; y/ D q.t s; x; y/.

10.10
a) Let us apply Ito’s formula to the function log, with the usual care, as it is not
defined on the whole of R. Let us denote by " the exit time of x;s from the half
y;s
line "; C1Œ; then, writing t D t , y D log x and t D tx;s for simplicity, we
have
Z t^" Z t^"
1
t^" D y C b.u ; u/ .u ; u/2 du C .u ; u/ dBu
Zs t^" 2 sZ
1 t^"
DyC b.e u ; u/ .euu ; u/2 du C .euu ; u/ dBu :
s 2 s
(S.86)
If e
b.y; u/ D b.ey ; u/ 12 .ey ; u/2 , e
.y; u/ D .ey ; u/, the process coincides
therefore up to time " with the solution Y of the SDE
dYt D e
b.Yt ; t/ dt C e
.Yt ; t/ dBt
(S.87)
Ys D y :
As the coefficients e b; e
are bounded and locally Lipschitz continuous, by
Theorem 9.2, the SDE (S.87) has a unique solution. Moreover, as " coincides
with the exit time of Y from the half line log "; C1Œ, by Remark 9.3 " ! C1
as " ! 0 and, taking the limitx;s
as " ! 0 in (S.86), we find that is a solution
of (S.87) and that tx;s D et > 0 a.s. for every t 0.
b) The generator e Lt of the diffusion is
e 1 @2 d
Lt D e .y; t/2 2 C e
b.y; t/
2 @y dy
Moreover, if e
.y/ D .ey /, e
f .y; s/ D f .ey ; s/ and e
c.y; s/ D c.ey ; s/, we can write
h Z i
RT y;t
T Rs y;t
u.ey ; t/ D E e.T / e t Qc.v ;v/ dv E e Qc.v ;v/ dv
y;t
s ; s/ e
f .y;t t ds :
t
As and f have polynomial growth in x, e and e

f have exponential growth
in y. The operator e Lt satisfies Assumption (A’) and has a bounded diffu-
sion coefficient. By Theorem 10.6, the function eu.y; t/ D u.ey ; t/ is a solu-
tion of
8 @eu
<eLte
uC eceu De
f on R Œ0; TŒ
@t
:eu.y; T/ D e .y/ :
We easily see that

@ @
e
Lt C e u.y; t/ D Lt C c u.ey ; t/
c e
@t @t
and therefore u is a solution of (10.45).
10.11
a) We have
Z 1
1 jxyj2
u.x; t/ D .y/ e 2.Tt/ dy
.2.T t//m=2
thus u is C1 .Rm Œ0; TŒ/ as we can take the derivative under the integral sign
(this is a repetition of Remark 6.4 or of Proposition 10.4).
b) This is a consequence of the Markov property (6.13):
u.x; s/ D Ex;s ŒEx;s ..XT /jFts / D Ex;s ŒEXt ;t ..XT // D Ex;s Œu.Xt ; t/ :
c) Thanks to a), the function x 7! u.x; T 0 / is continuous for every T 0 < T.

Moreover, u is clearly bounded, as is bounded itself. By point b) and the
Feynman–Kac formula (Theorem 10.6) u is therefore a solution of (10.46) on
Rm Œ0; T 0 Œ and therefore, by the arbitrariness of T 0 , on Rm Œ0; TŒ.
In order to prove (10.47), let x be a continuity point of and let " > 0 be fixed.
Let ı > 0 be such that j.x/.y/j " if jxyj ı. Let us first prove that there
exists a number Nt > 0 such that Px;t .jXT xj > ı/ D Px;0 .jXTt xj > ı/ < "
for every T Nt < t < T. Actually with respect to p Px;0 , .Xt x/t is a Brownian
motion and therefore XTt x has the same law as T t Z where Z N.0; I/,
so that
p
Px;0 .jXTt xj > ı/ D P.j T t Zj > ı/ ! 0:
t!T
If T Nt < t < T we then have
ju.x; t/ .x/j Ex;t Œj.XT / .x/j

D Ex;t j.XT / .x/j1fjXT xj>ıg C Ex;t j.XT / .x/j1fjXT xjıg
kk1 Px;t .jXT xj > ı/ C " < ".kk1 C 1/
and we can conclude the proof thanks to the arbitrariness of ".

d) By a) and c) for D 1Œ0;C1Œ ,
Z C1
1 2 =2
u.x; t/ D Ex;t Œ1Œ0;C1Œ .XT / D p ey dy
.2/1=2 x= Tt
is a solution of (10.46),(10.47); x D 0 is the only point of discontinuity of D

1Œ0;C1Œ and it is immediate that u.0; t/ D 12 for every t T.
10.12
a) Let 0 D infftI t > 0; Xt 62 Dg. We must prove that, if @D has a local barrier for
L at x, then Px . 0 D 0/ D 1. Let us still denote by w a bounded C2 .Rm / function
coinciding with w in a neighborhood of x, again denoted by W. Let D W ^ 0 ,
where W is the exit time from W. As w.x/ D 0, Ito’s formula gives, for t > 0,
Z t^ Z t^
w.Xt^ / D Lw.Xs / ds C w0 .Xs / dBs Px -a.s.
0 0
The stochastic integral has mean equal to zero, since the gradient w0 is bounded
in W. Therefore
Z t^
E Œw.Xt^ / D E
x x
Lw.Xs / ds Ex .t ^ / :
0
As w 0 on W \ D, the left-hand side is 0 and necessarily Ex .t ^ / D 0 for

every t > 0, i.e. D 0 Px -a.s. As W > 0 Px -a.s. by the continuity of the paths,
necessarily 0 D 0 Px -a.s. and therefore x is regular for the diffusion X.
b) Let z be the center of the sphere and let us look for a barrier of the form w.y/ D
k.jx zjp jy zjp / where k; p are numbers > 0 to be made precise later. First
it is clear that w.x/ D 0 and w > 0 on D, as S Dc , and therefore, for y 2 D,
jy zj is larger than the radius of the sphere, which is equal to jx zj. Let us
compute the derivatives of w.
@w
.y/ D kpjy zjp2 .yi zi /
@yi
@2 w
.y/ D kp.p C 2/jy zjp4 .yi zi /.yj zj / C kpjy zjp2 ıij :
@yi @yi
Therefore
X
m
Lw.y/ D kp.p C 2/jy zjp4 aij .y/.yi zi /.yj zj /
i;jD1
X
m X
m
C kpjy zjp2 aii .y/ C kpjy zjp2 bi .y/.yi zi / :
iD1 iD1
Let now be a positive number such that ha.y/; i jj2 for every y 2 D and
2 Rm and M a number majorizing the norm of b.y/ and the trace of a.y/ for
every y 2 W. Then
Lw.y/ kp.p C 2/ jy zjp2 C kpMjy zjp2 C kpMjy zjp1

D kpjy zjp2 . .p C 2/ M Mjy zj/ :
If W is a bounded neighborhood of x, then the quantity jyzj remains bounded for

y 2 W. One can therefore choose p large enough so that .pC2/MMjyzj >
1 on W. With this choice we have, for y 2 W \ D,
Lw.y/ kpjy zjp2 kpjx zjp2
and, choosing k large enough, we have Lw.y/ 1, as requested.
11.1
a1) We have
p p
tkC1 D tk C b tk h C tk h Zk D tk .1 C bh C h Zk /
from which (11.44) is immediate.

a2) Recall that for the geometric Brownian motion of (11.43) EŒT D ebT and
2
EŒT2 D e.2bC /T : We have
Y
n
p bT n
EŒ T D x EŒ1 C bh C h Zk D x.1 C bh/n D x 1 C ! xebT
iD1
n n!1
and
2 Y
n
p
EŒ T D x EŒ.1 C bh C h Zk /2
iD1
Y
n
p p
Dx E 1 C b2 h2 C 2 hZk2 C 2bh C 2 h Zk C 2bh h Zk

iD1
Y
n
T T 2 n 2 /T
Dx 1C.2bC 2 /hCb2 h2 D x 1C.2bC 2 / Cb 2 ! e.2bC :
iD1
N N n!1
a3) We have
x.1 C bh/n D xen log.1Cbh/ :

Recall the developments
1
log.1 C z/ D z z2 C O.z3 /
2
ey D ey0 C ey0 .y y0 / C O.jy y0 j2 /
that give
1 2 ThCO.h2 / 1 bT 2
xen log.1Cbh/ D xebT 2 b D xebT xe b Th C O.h2 / :
2
Hence
ˇ ˇ
ˇEŒT EŒ ˇ D c1 h C O.h2 / ;
T
1
with c1 D 2 xebT b2 T. The computation for (11.46) is quite similar.
b1) We have
p 1
tkC1 D e
e tk C be
tk h C eh ZkC1 C 2e
tk 2
tk .hZkC1 h/
2
p 1
De
tk 1 C bh C h Zk C 2 h.ZkC1
2
1/ ;
2
hence
Y
n1 1 p
e
T D x 1 C h b C 2 .ZkC1
2
1/ C h ZkC1 :
kD0
2
b2) We have
n1 h
Y i
1 p
EŒe
T D x E 1 C h b C 2 .ZkC1
2
1/ C h ZkC1 D x.1 C bh/n :
kD0
2
The mean of the Milstein approximation in this case is exactly the same as that
of the Euler approximation. Hence (11.47) follows by the same computation
leading to (11.45).
12.1
a) By Girsanov’s theorem, if
1 2
e
Z t D ecBt 2 c t ;
then Ws D Bs cs is a Brownian motion for s t with respect to the probability

dQ D e Z t dP. Therefore, with respect to Q, writing Bs D Ws C cs, .Bs /0st has
the same law as .Ys /0st . We deduce that PY is the law of .Bt /t with respect to
Q. If 0 t1 < < tm t and f1 ; : : : ; fm are bounded Borel functions, we have
Y
EP Œ f1 .Xt1 / : : : fm .Xtm / D EQ Œ f1 .Bt1 / : : : fm .Btm /
1 2 1 2
D EŒ f1 .Bt1 / : : : fm .Btm / ecBt 2 c t D EP Œ f1 .Xt1 / : : : fm .Xtm / ecXt 2 c t :
B
1 2
Therefore the two probabilities PY and ecXt 2 c t dPB have the same finite-dimen-
sional distributions on Mt and hence coincide on Mt . This proves that PY has a
1 2
density with respect to PB so that PY PB . Let Zt D ecXt 2 c t . As Zt > 0, we
also have PB PY on Mt .
If c > 0, for instance, we have limt!C1 Yt D C1 hence with respect to
PY limt!C1 Xt D C1 a.s. Therefore the event flimt!C1 Xt D C1g has
probability 1 with respect to PY but probability 0 with respect to PB , as under
PB .Xt /t is a Brownian motion and therefore limt!C1 Xt D 1 a.s. by the
Iterated Logarithm Law.
b) As in the argument developed in the second part of a) we must find an event in
Mt which has probability 1 for PB and 0 for PZ . For instance, by the Iterated
Logarithm Law,
Xt
PB lim 1=2 D1 D1
t!0C 2t log log 1t
whereas, as .Xt /t under PZ has the same law as .Bt /t , we have, considering
separately the cases > 0 and < 0,
Xt
lim D jj PZ a:s:
t
hence, if jj 6D 1, the event

n Xt o
lim D 1
t
has probability 1 for PB and probability 0 for PZ .

Another event having probability 1 for PB and probability 0 for PZ can be
easily produced using P. Lévy’s modulus of continuity (Theorem 3.1).
12.2
a) Of course it is sufficient to consider the case . As f .x/ D x log x, with
the understanding f .0/ D 0, is convex and lower semi-continuous, by Jensen’s
inequality,
Z Z Z d
d d d
H.I / D log d D f d f d
E d d E d E d
Z
Df d D f .1/ D 0 :
E
As, moreover, f is strictly convex, the inequality is strict, unless the function
d is -a.s. constant. As the integral of d with respect to must be equal to
d d
1, this happens if and only if d D 1 -a.s. and therefore if and only if D .

d
A simple approach to finding an example showing that the entropy is not

symmetric in its two arguments is to consider a measure such that there
exists a set A with 0 < .A/ < 1 and then setting d D .A/1 1A d. The
measure is not absolutely continuous with respect to , as .Ac / D 0 whereas
.Ac / > 0; therefore H.I / D C1. Conversely, and one computes
immediately H.I / D log .A/ < C1.
b1) The process .Xt /0tT , under the Wiener measure P, is a Brownian motion. If
Z T
1
Z T
Z D exp s0 dXs js0 j2 ds
0 2 0
RT
then, as the r.v. 0 s0 dXs is Gaussian, we know that EŒZ D 1 and by Girsanov’s
Theorem 12.1 under the probability dQ D Z dP, the process Wt D Xt t is
a Brownian motion for 0 t T. Hence Xt D Wt C t and, with respect to
Q, X is, up to time T, a Brownian motion with the deterministic drift . Hence
Q D P1 and P1 is absolutely continuous on MT with respect to the Wiener
measure P and
Z Z
dP1 T
1 T
D exp s0 dXs js0 j2 ds :
dP 0 2 0
Using the expression (12.20) the entropy H.P1 I P/ is the mean with respect to P1
of the logarithm of this density. But, as with respect to P1 for t T Wt D Xt t
is a Brownian motion, we have,
dP1 Z T Z
1 T 02
H.P1 I P/ D EP1 log D EP1 s0 dXs js j ds
dP 0 2 0
Z T Z T Z
1 T 02
Z
1 T 02 1
EP1 s0 dWs C js0 j2 ds js j ds D js j ds D k 0 k22 :
0 0 2 0 2 0 2
Very similar to this is the computation of the entropy H.PI P1 /. We have, clearly,
Z Z
dP T
1 T
D Z 1 D exp s0 dXs C js0 j2 ds
dP1 0 2 0
and, as with respect to P .Xt /0tT is a Brownian motion,
Z Z Z
T
1 T
1 T 02 1
H.PI P1 / D E P
s0 dXs C js0 j2 ds D js j ds D k 0 k22 :
0 2 0 2 0 2
Here the two entropies H.PI P1 / and H.P1 I P/ coincide.

b2) Developing the square we have in general, if ,
Z h 2 i Z 2 Z
d d d d
2 .I / D 2 C 1 d D d 1 D d 1 :
E d d E d E d
In our case, i.e. D P and D P1 ,

Z h Z T Z
dP 1 T 0 2 i
dP D E exp
P
s0 dXs C js j ds
dP1 0 2 0
h Z T i 1 Z T
0
D E exp
P
s dXs exp js0 j2 ds
0 2 0
Z T 0 2
D exp js0 j2 ds D ek k2 :
0
Therefore
Z
dP 0 2
2 .P1 I P/ D dP 1 D ek k2 1 :
dP1
2 .PI P1 / can be computed similarly, giving the same result.
12.3
a) .Zt /t is a martingale and an old acquaintance (see Example 5.2). e
B is a Brownian
motion with respect to Q by Girsanov’s Theorem 12.1 (here ˚s 2
).
b) We have
Zt1 D e2
Bt C2
t D e2
e
2
Bt 2
t 2
:
Again .Zt1 /t is a Q-martingale as e B is a Q-Brownian motion. Alternatively,

simply observe that on Ft we have dP D Zt1 dQ, and therefore necessarily
2
.Zt1 /t is a Q-martingale (Exercise 5.27). Note also that Zt1 D e2
Bt C2
t D
e2
Xt .
c) As fR Tg 2 FT ,
P.R T/ D EQ .1fR Tg ZT1 / ;

but also fR Tg 2 FR and therefore fR Tg 2 FT^R . Conditioning with

respect to FT^R , by the stopping theorem of martingales (Theorem 5.13),

EQ .1fR Tg ZT1 / D EQ EQ .1fR Tg ZT1 jFT^R / D EQ 1fR Tg EQ .ZT1 jFT^R /
1
D EQ .1fR Tg ZT^R
/:
1 1
As ZT^ R
D e2
XT^R and XT^R D R on fR Tg, we have ZT^ 1
R fR Tg
D
2
R
e 1fR Tg hence (12.21).
d) As, with respect to Q, Xt D e Bt C
t, we have limt!C1 Xt D C1 Q-a.s. and
Q.R < C1/ D 1. Taking the limit as T ! C1 in (12.21) we have

P sup Xt R D P.R < C1/ D e2
R
t>0
so that the r.v. supt>0 Xt has an exponential law with parameter 2

.
• The observant reader has certainly noticed in the last limit a certain carelessness
(the probability Q itself actually depends on T). However, it is not difficult to
complete the argument with care.
12.4 By Girsanov’s theorem, the process

Z t
Bs
Wt D Bt C ds;
0 1s
t T, is a Brownian motion with respect to Q. Hence, with respect to Q, B is a

solution of the SDE
Bt
dBt D dt C dWt :
1t
As clearly B0 D 0 Q-a.s., we know (Exercise 9.2 b)) that the solution of this SDE is
a Brownian bridge.
12.5
a) In order to show that E.Zt / D 1, by Corollary 12.1 b) it is enough to prove that
2 2
E.e
jXs j / < C1 for some > 0 for every s t. If Xs D .X1 .s/; : : : ; Xm .s//,
then, as the components of X are independent,
2 jX j2 2X 2 2X 2 2X 2 2W2
E.e
s
/ D E.e
1 .s/
/ : : : E.e
m .s/
/ D E.e
1 .s/
/m D E.es
/m ;
where W N.0; 1/. This quantity, thanks to Exercise 1.12, is finite for every s
t if < .2
2 t/1 . Actually this is a repetition of the argument of Example 12.2.
b) By Girsanov’s Theorem 12.1, with respect to Q the process
Z s
Ws D Xs
Xu du
0
is an (m-dimensional) Brownian motion for s t. Therefore, for s t, X is the

solution of
dXs D
Xs ds C dWs
and is therefore, under Q, an Ornstein–Uhlenbeck process. As X0 D W0 D 0, as

seen in Sect. 9.2 we have
Z t
Xt D e
.ts/ dWs
0
and X is a Gaussian process; EQ .Xt / D 0 and, by Proposition 8.5, the covariance

matrix of Xt with respect to Q is equal to the identity matrix multiplied by
Z t
1 2
t
e2
.ts/ ds D .e 1/ : (S.88)
0 2
c) By Ito’s formula, with respect to P we have djXt j2 D 2Xt dXt C m dt, i.e.
Z t
1
Xs dXs D .jXt j2 mt/ :
0 2
2
Therefore, for D
2 ,

2
2
EQ Œe 2 .jXt j mt/ D EŒZt e 2 .jXt j mt/
h Z t
2 t
Z

i
D E exp
Xs dXs jXs j2 ds .jXt j2 mt/
0 2 0 2
h
2 Z t i
D E exp jXs j2 ds D J :
2 0
As the components of X are independent with respect to Q, thanks to b) we have

2 mt/ 1
2
J D EQ Œe 2 .jXt j D e 2 m
t EQ Œe 2 X1 .t/ m :
With respect to Q, X1 .t/ is a centered Gaussian r.v. with a variance given

by (S.88); therefore, recalling Exercise 1.12,

2 1=2 1 1=2
EQ Œe 2 X1 .t/ D 1 C
VarQ .X1 .t// D 1 C .e2
t 1/
2
1 1=2
D .e2
t C 1/ :
2
Putting all pieces together we obtain
1
1 m=2 e2
t C 1 m=2
J D e 2 m
t .e2
t C 1/ D
2 2e
t
p
D cosh.
t/m=2 D cosh. 2 t/m=2 :
Rt
d) Let us denote by the Laplace transform of the r.v. 0 jXs j2 ds so that J D . /.
p
We have seen that, for 0, . / D cosh. 2 t/m=2 . Recalling that the
Laplace transform
p is an analytic function
p (see Sect. 5.7), for 0 we have
. / D cosh.i 2 t/m=2 D cos. 2 t/m=2 , up to the convergence abscissa.
Keeping in mind that the first positive zero of the cosine function is 2 , the
2
convergence abscissa is D 8t 2 and in conclusion
8 p
ˆ m=2
<cosh.p 2 t/
ˆ if 0
J D cos . 2 t/m=2 2
if 0 < <
ˆ 8t2
:̂C1 if 2
:
8t2
12.6
Rt
a1) By Girsanov’s theorem, Wt D Xt 0 b.Xs C x/ ds is a Brownian motion for
t T with respect to the probability Q defined by dQ D ZT dP, where
Z t
1
Z t
Zt D exp b.Xs C x/ dXs b2 .Xs C x/ ds : (S.89)
0 2 0
Such a process .Zt /t is a martingale thanks to Proposition 12.1, b being

bounded.
With respect to Q, .Xt /0tT satisfies
Z t
Xt D b.Xs C x/ ds C Wt ;
0
hence Yt D Xt C x is a solution of
Z T Z T
Yt D x C b.Xs C x/ ds C Wt D x C b.Ys / ds C Wt :
0 0
a2) If U is a primitive of b, then by Ito’s formula
1
dU.Xt C x/ D b.Xt C x/ dXt C b0 .Xt C x/ dt
2
so that
Z Z
t
1 t
b.Xs C x/ dXs D U.Xt C x/ U.x/ b0 .Xs C x/ ds ;
0 2 0
whence substituting into (S.89) we find that Zt is given by (12.22).

b1) We have
sinh2 .z/ 1
tanh2 .z/ D 2
D1
cosh .z/ cosh2 .z/
1
tanh0 .z/ D ,
cosh2 .z/
hence
k2 1
b0 .z/ D 2
, b2 .z/ D k2 1 ;
cosh .kz C c/ cosh2 .kz C c/
so that b0 .z/ C b2 .z/ k2 .

b2) Let t T. The law of Yt coincides with the law of Xt C x with respect to the
probability Q defined as dQ D ZT dP, where ZT is as in (12.22) with b.z/ D
k tanh.kz C c/. Hence the Laplace transform of Yt is given by
EŒe
Yt D EQ Œe
.Xt Cx/ D EŒZT e
.Xt Cx/ D EŒEŒZT e
.Xt Cx/ jFt D EŒZt e
.Xt Cx/ :
A primitive of z 7! k tanh.kz C c/ is U.z/ D log cosh.kz C c/. Therefore

Z
1 t
U.Xt C x/ U.x/ Œb0 .Xs C x/ C b2 .Xs C x/ ds
2 0
1 2
D log cosh.kXt C kx C c/ log cosh.kx C c/ k t
2
and
cosh.kXt C kx C c/ 1 k2 t
Zt D e 2 :
cosh.kx C c/
Hence, as X is a Brownian motion,

1 2

.Xt Cx/ e 2 k t
EŒZt e D E cosh.kXt C kx C c/ e
.Xt Cx/
cosh.kx C c/
1 2
e 2 k t e
x
D E e.
Ck/Xt CkxCc C e.
k/Xt kxc
2 cosh.kx C c/
1 2
e 2 k t e
x 1 .
Ck/2 t kxCc 1 2
D e2 e C e 2 .
k/ t ekxc
2 cosh.kx C c/
1 1
2 t
.ktCx/ kxCc 1 2
e„2 ƒ‚e … e C„ e 2
t eƒ‚

.ktCx/ kxc
… e ;
2 cosh.kx C c/
b
1 .
/ b2 .
/
where we recognize that b 1 and b

2 are respectively the Laplace transform of
an N.kt C x; t/ and of an N.kt C x; t/ law. The law of Yt is therefore a mixture
of these two laws with weights (independent of t)
ekxCc ekxc
˛D , 1˛ D
ekxCc C ekxc ekxCc C ekxc
b3) Of course
Z Z
ekxCc .kt C x/ C ekxc .kt C x; /
EŒYt D ˛ x d1 .x/ C .1 ˛/ x d2 .x/ D
ekxCc C ekxc
D x C kt tanh.kx C c/ :
12.7
a) Let B D .˝; F ; .Ft /t ; .Bt /t ; P/ be an m-dimensional Brownian motion. Recall
that, by definition of Wiener measure, for every Borel set A C we have P.B 2
A/ D PW .A/. Hence PW .C0 / D P.B0 D 0/ D 1.
b) If A is an open set containing the path 0, then it contains a neighborhood of 0 of
the form U D fwI sup0tT jw.t/j < g; let us show that PW .U/ > 0. In fact
Y m
PW .U/ D P sup jBt j T/ > 0
0tT m
we have PW .U/ > 0. Rt

c) Let us denote by H1 C0 the set of the paths of the form t D 0 ˚s ds, ˚ 2 L2 .
As H1 is dense in C0 , if A C0 is an open set (with respect to the topology
induced by C ), then it contains a path 2 H1 . Hence e
A D A is an open set
containing the origin. Let V D fwI sup0tT jw.t/j < g be a neighborhood of
the origin of C0 contained in e
A. We have
P.B 2 A/ D P.B 2 e
A/ P.B 2 V/ :
Let
Z T
1
Z T
Z D exp s0 dBs js0 j2 ds ;
0 2 0
and let Q be the probability defined by dQ D Z dP. Then, by Girsanov’s theorem,

for t T the process Wt D Bt t is a Brownian motion with respect to Q and
we have
P.B 2 V/ D EQ ŒZ 1 1fB 2Vg D EQ ŒZ 1 1fW2Vg :
Now observe that Z 1 > 0 Q-a.s. and that, thanks to b), Q.W 2 V/ D P.B 2
V/ > 0 (W is a Brownian motion with respect to Q). Hence P.B 2 A/
EQ ŒZ 1 1fB2Vg > 0.
12.8
a) Let us prove that .Zt /t is a martingale for t T. We can take advantage
of Corollary 12.1, which requires us to prove that for some value of > 0
2 2
E.e
jBt Cxj / < C1 for every t T. But this is immediate as
2 jB 2 2 jxj2 2 jB j2 2 jxj2 2B 2
E.e
t Cxj
/ e2
E.e2
t
/ D e2
E.e2
1 .t/
/m
and the right-hand side is finite for every t T as soon as < .4

2 T/1
(Exercise 1.12).
b) By Girsanov’s theorem, with respect to Q the process
Z t
Wt D Bt
.Bs C x/ ds
0
is a Brownian motion for t T; therefore, if Yt D Bt C x,

Z t
Yt D Bt C x D x C
Ys ds C Ws
0
and .Yt /t is, with respect to Q, an Ornstein–Uhlenbeck process having x as its

initial value. Going back to the computations of Sect. 9.2, Yt is a Gaussian r.v.
with mean b D e
t x and covariance matrix D 2
1 .e2
t 1/I.
c) By Ito’s formula, with respect to P,
Z T
1 1
.Bs C x/ dBs D .jBT j2 mT/ C hx; BT i D .jBT C xj2 jxj2 mT/ ;
0 2 2
hence (12.25). Therefore

2
h
2
i
2
h
2
i
e 2 .mTCjxj / EQ e 2 jBT Cxj D e 2 .mTCjxj / E ZT e 2 jBT Cxj
h Z T Z

2 T
i
D E exp
.Bs C x/ dBs jBs C xj2 ds .jBT C xj2 jxj2 mT/
0 2 0 2
h Z

2 T i
D E exp jBs C xj2 ds :
2 0
d) We can write W D b C Z with Z N.0; I/. Therefore, thanks to the

independence of the components of Z,
h Xm i

jWj2
jbC Zj2
EŒe D EŒe D E exp
.bi C Zi /2
iD1
h
Xm i Y
m
2 2 2
D E exp
.b2i C 2Zi bi C 2 Zi2 / D e
bi E e2
bi Zi C
Zi :
iD1 iD1
Now
Z C1
2 2
1 2 2 2
E e2
bi Zi C
Zi D p e2
bi zC
z ez =2 dz
2 1
Z C1
1 1 2 2
D p e 2 ..12
/z 4
bi z/ dz :
2 1
This type of integral can be computed by writing the exponent in the form of the
square of a binomial times the exponential of a term not depending on z, i.e.
Z C1 h 1 2
2
1 4
bi z i
D p exp z2
2 1 2 1 2
2
Z C1 h 1 2
2
1 2
bi 2 i 2
2 2 b2
D p exp z dz exp i
:
2 1 2 1 2
2 1 2
2
The integral diverges if 2 2

1, otherwise, recognizing in the integrand (but
for the normalizing constant) the density of a Gaussian r.v. with variance .1
2
2 /1 we find
2 1 Y 2 2
2 2 b2
m
EŒe
jWj D 2
e
bi exp i
2
.1 2
/ m=2 1 2

iD1
(S.90)
1
D 2
exp jbj2 :
.1 2
/ m=2 1 2
2
e) Thanks to b) BT C x has, with respect to Q, a normal law with mean b D e

T x
and covariance matrix D 2 I, with 2 D 2
1 .e2
T 1/. By c) and d), replacing

with
2 in (S.90),
h Z i h
i

2 T

2 2
E exp jBs C xj2 ds D e 2 .mTCjxj / EQ e 2 jBT Cxj
2 0

2/ 1

D e 2 .mTCjxj 2
exp 2
jxj2 e2
T :
.1 C
/ m=2 2.1 C
/
As 1 C 2
D 12 .e2
T C 1/ we have

e 2 mT m=2

T m=2 1 2
T
D .e / .e C 1/
.1 C 2
/m=2 2
1 m=2
D .e
T C e
T / D cosh.
T/m=2
2
whereas

2

jxj2 2e2
T
e 2 jxj exp jxj 2 2
T
e D exp 1
2.1 C 2
/ 2 e2
T C 1

jxj2
D exp tanh.
T/ ;
2
i.e.
h Z i h
jxj2 i

2 T
E exp jBs C xj2 ds D cosh.
T/m=2 exp tanh.
T/ :
2 0 2
(S.91)
p
f) We have from c) above and (S.91) with
D 2 and T t instead of T
h R Tt 2
i
u.x; t/ D E e 0 jBs Cxj ds
p m=2 h p2 jxj2 p i
D cosh 2 .T t/ exp tanh 2 .T t/ :
2
This solution, which is bounded, is unique among the functions having polyno-
mial growth.
12.9
a) This is a typical application of Lemma 4.1 (the freezing lemma) as developed
in Example 4.4, where it is explained that
Z jzBt j2
1
EŒ f .BT /jFt D .Bt ; t/ WD f .z/ e 2.Tt/ dz :
.2.T t//m=2 Rm
Differentiability properties of (for t < T) follow from the theorem of

differentiation under the integral sign. Moreover, .0; 0/ D EŒ f .BT /.
b) Ito’s formula gives
@ 1
0
dZt D d .Bt ; t/ D .Bt ; t/ C 4 .Bt ; t/ dt C x .Bt ; t/ dBt
@t 2
the term in dt, however, must vanish as .Bt ; t/ D EŒ f .BT /jFt is a continuous
square integrable martingale!
c1) The candidate integrand is of course Xs D x0 .Bs ; s/. We must prove that such
a process belongs to M 2 .Œ0; T/, which is not granted in advance as is not
necessarily differentiable in the x variable for t D T (unless f were itself
differentiable, as we shall see in d)). We have, however, for every t < T, by
Jensen’s inequality,
h Z 2 i

t
EŒ f .BT /2 E EŒ f .BT /j Ft 2 D EŒ .Bt ; t/2 D E .0; 0/C 0
x .Bs ; s/ dBs :
0
Observing that .0; 0/ D EŒ f .BT / we have for t < T

h Z t 2 i
0
E .0; 0/ C x .Bs ; s/ dBs
0
hZ t i hZ t i
D EŒ f .BT /2 C E 0
x .Bs ; s/
2
ds C 2EŒ f .BT / E 0
x .Bs ; s/ dBs
0
„ 0 ƒ‚ …
D0
so that, for every t T,

Z t
0 2
EŒj x .Bs ; s/j ds EŒ f .BT /2 EŒ f .BT /2
0
and, taking the limit as t ! T, we obtain

Z T
0 2
EŒj x .Bs ; s/j ds EŒ f .BT /2 EŒ f .BT /2 < C1 :
0
Hence s 7! x0 .Bs ; s/ belongs to M 2 .Œ0; T/. Now (12.27) follows from (12.28),
as f .BT / D limt!T .Bt ; t/ a.s. p
c2) If f .x/ D 1fx>0g , then, using once more the fact that BT Bt T t Z with
Z N.0; 1/,
x
.x; t/ D EŒ f .xCBT Bt / D EŒ1fxCBT Bt >0g DP.BT Bt > x/ D ˚ p ;
T t
where we denote by ˚ the partition function of an N.0; 1/ law. Therefore

2
@ 1
Bs
Xs D .Bs ; s/ D p e 2.Ts/ :
@x 2.T s/
d) We have
Z jzxj2
1
.x; t/ D f .z/ e 2.Tt/ dz
.2.T t//m=2 Rm
Z jzj2
1
D f .x C z/ e 2.Tt/ dz :
.2.T t//m=2 Rm
If f is differentiable with bounded derivatives then the theorem of differentia-

tion under the integral sign gives
Z jzj2
0 1 2.Tt/
x .x; t/ D fx0 .x C z/ e dz
.2.T t//m=2 Rm
Z jzxj2
1 2.Tt/
D f 0 .z/ e dz
.2.T t//m=2 Rm
hence, again by the freezing lemma,

0
x .Bt ; t/ D EŒ fx0 .BT /jFt :
13.1 Recall (this is (13.22)) that under P the prices follow the SDE
X
d
dSi .t/ D rt Si .t/ dt C ij .St ; t/Si .t/ dBj .t/ ;
jD1
where .Bt /t is a P -Brownian motion. As in this model the processes Si , i D

1; : : : ; m, are in MB2 , the stochastic component of the differential is a martingale
and
Z t Z t

E ŒSi .t/ D E Œrs Si .s/ ds C EŒSi .0/ D E Œrs E ŒSi .s/ ds C E ŒSi .0/ :
0 0
Therefore the function v.t/ D E ŒSi .t/ satisfies the differential equation
v 0 .t/ D E Œrt v.t/

Rt
E Œrs ds
with the initial condition v.0/ D E ŒSi .0/, so that v.t/ D e 0 E ŒSi .0/.
13.2
a) If s t we have
Rt Rt Rt
E ŒSi .t/jFs D e 0 ru du E Œe
Si .t/jFs D e 0 ru du e
Si .s/ D e s ru du Si .s/ ;
whereas, of course, E ŒSi .t/jFs D Si .t/ if s t.

b) We have
Z T
RT RT
Vt WD E Œe Z jFt D e t ru du
t ru du
˛E ŒSi .s/jFt C ˇ ds
RT Z t 0Z
T
D e t ru du ˇT C ˛Si .s/ ds C ˛E ŒSi .s/jFt ds (S.92)
RT Z 0t t Z T R
s
t ru du
De ˇT C ˛Si .s/ ds C ˛Si .t/ e t ru du ds :
0 t
Rt
As S0 .t/ D e 0 ru du , we can write
RT Z t Z T R
T
V t D e 0 ru du
˛Si .s/ ds C ˇT S0 .t/ C ˛ e s ru du ds Si .t/ :
0
„ ƒ‚ … „ t ƒ‚ …
WDH0 .t/ WDHi .t/
Let us prove that H is admissible. The condition Vt 0 is verified by construc-

tion. In order to check that it is self-financing we shall verify condition (13.8) of
Proposition 13.1. From (S.92) we have
Rt RT Z t Z T Rs
e
V t D e 0 ru du V t D e 0 ru du
ˇT C ˛Si .s/ ds C ˛Si .t/ e t ru du
ds :
0 t
(S.93)
Hence, taking the differential,
RT n Z T Rs
de
V t D e 0 ru du
˛Si .t/ dt C ˛Si .t/ 1 rt e t ru du
ds dt
t
Z T Rs o
C˛ e t ru du ds dSi .t/
t
Z
RT T Rs
D e 0 ru du
˛ e t ru du
ds rt Si .t/ dt C dSi .t/ :
t
Recall now, this is (13.7), that

Rt
de
Si .t/ D e 0 ru du .rt Si .t/ dt C dSi .t/
so that
RT Z T Rs Z T RT
de
V t D e t ru du
˛ e t ru du
ds de
Si .t/ D ˛ e s ru du
ds de
Si .t/
t t
D Hi .t/ de
Si .t/ :
c) The arbitrage price of the option at time 0 is obtained by taking t D 0 in (S.93):

denoting S1 .0/ by x,
RT
Z T RT
V0 D e
V 0 D e 0 ru du
ˇT C ˛x e s ru du
ds :
0
• This is an example of an option that does not depend on the value of the
underlying at time T only, as is the case for calls and puts. Note also that the
price of the option at time t, Vt , is a functional of the price of the underlying Si
up to time t (and not just of Si .t/, as is the case with calls and puts).
13.3
a) The portfolio that we need to investigate enjoys two properties, first that it is
self-financing, i.e. is such that
dVt D H0 .t/ dS0 .t/ C H1 .t/ dS1 .t/ (S.94)
and, second, that
M
H1 .t/ D
S1 .t/
Now Vt D H0 .t/S0 .t/ C H1 .t/S1 .t/ D H0 .t/S0 .t/ C M. Hence
dVt D H0 .t/ dS0 .t/ C S0 .t/ dH0 .t/ ; (S.95)
so (S.94) and (S.95) together give
H0 .t/ dS0 .t/ C H1 .t/ dS1 .t/ D H0 .t/ dS0 .t/ C S0 .t/ dH0 .t/ ;
i.e. S0 .t/ dH0 .t/ D H1 .t/ dS1 .t/, from which we obtain
H1 .t/ M Mert
dH0 .t/ D dS1 .t/ D dS1 .t/ D dS1 .t/ :
S0 .t/ S0 .t/S1 .t/ S1 .t/
As in the Black–Scholes model
dS1 .t/ D S1 .t/.b dt C dBt /
we obtain
dH0 .t/ D Mert .b dt C dBt /
hence
Z t
Mb
H0 .t/ D H0 .0/ C .1 ert / C M ers dBs ;
r 0
which gives for the value of the portfolio
Vt D H0 .t/S0 .t/ C H1 .t/S1 .t/ D H0 .t/S0 .t/ C M

Z t
Mb
D ert H0 .0/ C .1 ert / C M ers dBs C M :
r 0
Of course,
Mb Mb
EŒVt D ert H0 .0/ C .1 ert / C M D ert V0 M C .1 ert / C M :
r r
b) In order for V to be admissible it is necessary that Vt > 0 a.s. for every t.
This is not true in this case as Vt has a Gaussian distribution with a strictly
positive variance and such a r.v. is strictly negative with a strictly positive
probability.
13.4 Let us denote by .St /t the price process.

a) Obviously
Z D C 1fsup0uT Su >Kg :
b) By definition (see Proposition 13.4) the arbitrage price of Z at time 0 is

V0 D erT E ŒZ D erT C P sup Su > K :
0uT
Under the equivalent martingale measure P the price process S with the starting
condition S0 D x is the geometric Brownian motion
2
St D xe.r 2 /tC Bt
:
Therefore

V0 D erT C P sup Su > K
0uT
2
D erT C P sup xe.r 2 /uC Bu
>K
0uT
2
D erT C P sup r u C Bu > log K log x
0uT 2
r 1
D erT C P sup u C Bu > log Kx :
0uT 2
This quantity is of course equal to erT C if x > K, as in this case the prices
are already larger than the level K at time 0. If x < K, hence log Kx > 0, the
probability above (probability of crossing a positive level before a time T for a
Brownian motion with drift) has been computed in Example 12.5. Going back
to (12.11) (we must replace with . r 2 / and a with 1 log Kx ) we have,
denoting by ˚ the partition function of an N.0; 1/-distributed r.v.
n 2 r K
h 1 1 K r i
V0 D erT C e . 2 / log x 1 ˚ p log C . /T
T x 2 2
1 1 K r o
C1 ˚ p log . /T :
T x 2 2
• This is another example where the payoff is not a function of the underlying asset
at the final time T (as in Exercise 13.2). One may wonder if in this case the price
of the option can also be obtained as the solution of a PDE problem. Actually
the answer is yes, but unfortunately this is not a consequence of Theorem 10.4
because not all of its assumptions are satisfied (the diffusion coefficient can
vanish and the boundary data are not continuous at the point .K; T/).
13.5 Let us first prove that there exist no equivalent martingale measure Q such
that the discounted price processes e
S1 ; e
S2 are both martingales. Actually, Eq. (13.18)
here becomes the system
(
t D r 1
t D r 2
and, if 1 6D 2 , a solution cannot exist. Note that here we have d D 1 (dimension

of the Brownian motion) smaller than m D 2 (number of underlying assets). In this
case the existence of an equivalent martingale measure is not ensured, as remarked
on p. 416.
Also note that, if 1 > 2 , then the strategy H0 0, H1 1, H2 xx12
(recall that x1 D S1 .0/, x2 D S2 .0/) produces an arbitrage portfolio. In words, it is
a portfolio whose composition is constant in time and is obtained by buying at time
0 a unit of asset S1 and short selling xx12 units of asset S2 . This operation requires no
capital. The value at time t of such a portfolio is
x1 2 x1 2
Vt .H/ D S1 .t/ S2 .t/ D x1 e.1 2 /tC Bt x2 e.2 2 /tC Bt
x2 x2
2
D x1 .e1 t e2 t /e 2 tC Bt
;
which proves simultaneously that it is admissible (Vt .H/ 0 for every t) and with
arbitrage, as VT .H/ > 0 a.s.
References
Asmussen, S. (2000). Ruin probabilities, volume 2 of Advanced Series on Statistical Science &
Applied Probability. World Scientific Publishing Co., Inc., River Edge, NJ.
Azencott, R. and al. (1981). Géodésiques et Diffusions en temps petit. Astérisque 84–85.
Doss, H. (1977). Liens entre équations différentielles stochastiques et ordinaires. Ann. Inst. H.
Poincaré, B, 13:99–126.
Fishman, G. S. (1996). Monte Carlo. Springer Series in Operations Research. Springer-Verlag,
New York. Concepts, algorithms, and applications.
Friedman, A. (1964). Partial Differential Equations of Parabolic Type. Prentice Hall.
Friedman, A. (1975). Stochastic Differential Equations and Applications, I and II. Academic
Press, New York.
Gallardo, L. (1981). Au sujet du contenu probabiliste d’un lemme d’Henri Poincaré. Ann. Sci. Univ.
Clermont-Ferrand II Math., (19):185–190. Saint-Flour Probability Summer Schools (Saint-
Flour, 1979/1980).
Gilbarg, D. and Trudinger, N. S. (2001). Elliptic partial differential equations of second order.
Classics in Mathematics. Springer-Verlag, Berlin. Reprint of the 1998 edition.
Glasserman, P. (2004). Monte Carlo methods in financial engineering, volume 53 of Applications
of Mathematics (New York). Springer-Verlag, New York. Stochastic Modelling and Applied
Probability.
Graham, C. and Talay, D. (2013). Stochastic simulation and Monte Carlo methods, volume 68 of
Stochastic Modelling and Applied Probability. Springer, Heidelberg. Mathematical foundations
of stochastic simulation.
Han, Q. and Lin, F. (1997). Elliptic partial differential equations, volume 1 of Courant Lecture
Notes in Mathematics. New York University, Courant Institute of Mathematical Sciences, New
York; American Mathematical Society, Providence, RI.
Ikeda, N. and Watanabe, S. (1981). Stochastic Differential Equations and Diffusion Processes.
North Holland.
Karatzas, I. and Shreve, S. (1991). Brownian Motion and Stochastic Calculus, 2nd edition.
Springer, Berlin, Heidelberg, New York.
Kloeden, P. E. and Platen, E. (1992). Numerical solution of stochastic differential equations,
volume 23 of Applications of Mathematics (New York). Springer-Verlag, Berlin.
Levi, E. E. (1907). Sulle equazioni lineari totalmente ellittiche alle derivate parziali. Rend. Circolo.
Mat. Palermo, 24:275–317.
Maruyama, G. (1955). Continuous Markov processes and stochastic equations. Rend. Circ. Mat.
Palermo (2), 4:48–90.
Musiela, M. and Rutkowski, M. (2005). Martingale methods in financial modelling, volume 36 of
Stochastic Modelling and Applied Probability. Springer-Verlag, Berlin, second edition.

624 References
Neveu, J. (1964). Bases Mathématiques du Calcul des Probabilités. Masson et Cie.

Priouret, P. (1974). Diffusions et équations différentielles stochastiques. In Ecole d’été de
Probabilités de St. Flour III-1973, Lect. Notes Math. 390. Springer.
Revuz, D. and Yor, M. (1999). Continuous martingales and Brownian motion, volume 293 of
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences]. Springer-Verlag, Berlin, third edition.
Rogers, L. C. G. and Williams, D. (2000). Diffusions, Markov processes, and martingales. Vol. 2.
Cambridge Mathematical Library. Cambridge University Press, Cambridge. Itô calculus,
Reprint of the second (1994) edition.
Stroock, D. and Varadhan, S. (1979). Multidimensional Diffusion Processes. Grundlehren der
Mathematisches Wissenschaften 233. Springer, Berlin, Heidelberg, New York.
Sussmann, H. (1978). On the gap between deterministic and stochastic differential equations. Ann.
Probab., 6:19–41.
Talay, D. and Tubaro, L. (1990). Expansion of the global error for numerical schemes solving
stochastic differential equations. Stochastic Anal. Appl., 8(4):483–509 (1990).
Index
-correlated Brownian motion, 76 chi-square discrepancy, 389

-algebra generated by a r.v., 5 compensator, 111
complete, market, 407
conditional
absolutely continuous, probability, 4 expectation, 85
admissible strategy, 401 law, 85, 98
associated increasing process, 111 convergence
continuous time, 131, 207 almost sure (a.s.), in Lp , in law, 11
weak of measures, 11
convolution, 10
backward equation, 328 covariance, 4
barrier options, 428 function, 40, 175
Bessel processes, 179, 254 matrix, 10
Bessel square processes, 238
Black–Scholes
formula, 424 delta, 422
generalized model, 409 derivative (in finance), 395
standard model, 422 diffusion coefficient, 256
Blumenthal, 0-1 law, 168 diffusion processes, 174
Borel–Cantelli, lemma, 13 discounted
Box–Müller algorithm, 22 portfolio, 399
Brownian bridge, 107, 145, 176, 291 price process, 399
Brownian motion, 45 Doob’s
-correlated, 76 decomposition, 111
canonical, 53 maximal inequality, 119
geometric, 259 measurability criterion, 5
multidimensional, 48 drift, 256
standard, 102 Dupire model, 409
Dynkin, criterion, 179
call-put, parity, 432

Cameron–Martin, formula, 381 entrance time, 65
Carathéodory, criterion, 6 entropy, 389
Chapman–Kolmogorov, equation, 151 equivalent martingale measure, 402
characteristic function, 13 equivalent, probabilities, 5
Chebyshev’s inequality, 3 Euler, simulation scheme, 342

626 Index
European options, 396 law

exit time, 65 of a process, 52
of a r.v., 4
laws
Feller, property, 164 Cauchy, 135, 479
Feynman–Kac, formula, 319 exponential, 25, 134
filtration, 31 lognormal, 27
natural augmented, 32 normal or Gaussian, 15
right-continuous, 34 stable, 82
finite variation, functions, 55 local barrier, 339
finite-dimensional distributions, 38
fundamental solution, 326
marginal, law, 9
market price of risk, 411
Gaussian family, 18 Markov
geometric Brownian motion, 259 canonical process, 161
Greeks, 422 process, 152
delta, 422, 425 property, 152
gamma, 422, 426 strong property, 160
Rho, 422, 426 Markov’s inequality, 3
theta, 422, 426 martingale, 109
Vega, 422, 426 local, 205
Gronwall’s inequality, 261 with independent increments, 143
mathematical expectation, 1
maturity, 395
Hölder’s inequality, 2 maximal inequalities, 113
maximum principle, 171, 313
mean property of harmonic functions, 314
increasing predictable process, 111 Milstein, simulation scheme, 359
independence Minkowski’s inequality, 3
of -algebras, 6, 7 modulus of continuity, 53
of random variables, 5 moment
infinitesimal generator, 170 absolute of order ˛, 3
local, 170 of order ˛, 3
invariant
measure, probability, 178
Iterated Logarithm Law, 57 natural filtration, 31, 46
Ito augmented, 51, 102
formula, 219, 237 negligible, event, 5
process, 215 no-arbitrage price, 406
Jensen’s inequality, 2 occupation time, 78

options
American, 396
kernel, Poisson, 311 European, 396
Kolmogorov’s barrier, 428
continuity theorem, 35 call, put, 396, 424, 432
existence theorem , 40 Ornstein–Uhlenbeck, process, 258, 281, 291,
292, 295, 298–300, 374, 391, 532
Laplace transform, 26, 133

convergence abscissas, strip, 134 partition function, 25
domain, 133 passage time, 68, 145, 148, 374, 375
Poincaré’s lemma, 253 stochastic process, 31

Poisson kernel, 311 adapted, 31
portfolios, 398 continuous, right-continuous, 33
positive definite, kernel, 41 elementary, 184
probabilities equivalent, indistinguishable, 32
absolutely continuous, 4 measurable, 33
equivalent, orthogonal, 5 modification, version, 32
processes progressively measurable, 33
-correlated Brownian motion, 76, 83 standard, 34
Bessel, 179, 254 stopping time, 63
Bessel square, 238 entrance, exit, 65
Brownian bridge, 107, 145, 176, 291 strike price, 395
geometric Brownian motion, 259 strong Markov, property, 160
Ito, 215 submartingale, supermartingale, 109
Ornstein–Uhlenbeck, 258, 281, 291, 292, superior limit of events, 13
295, 298–300, 374, 391, 532
pseudo-inverse, of an increasing function, 228
Talay–Tubaro, expansion, 349
trading strategy, 398
Radon–Nikodym, theorem, 4 admissible, 401
random variables arbitrage, 401
independent, 5 self-financing, 399
uncorrelated, 4 transition function, 151
realization, of a Markov process, 157 time homogeneous, 159
reflection principle, 70 translation, operators, 161
riskless asset, 398
uniform integrability, 121

scaling, 48 uniqueness
solutions in law, 256
strong, 256 pathwise, 256
weak, 256, 379 upcrossings, 114
sphere condition, the, 339
spot interest rate, 398
stationary variance, 3
measure, probability, 178 volatility, 409
stochastic differential, 215
stochastic integral, 187
of elementary processes, 184 Wiener measure, 53

No

Uploaded by

Copyright:

Available Formats

No

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

No

Uploaded by

Copyright:

Available Formats

Universitext

Universitext is a series of textbooks that presents material from a wide variety of

More information about this series at http://www.springer.com/series/223

ISSN 0172-5939 ISSN 2191-6675 (electronic)

Library of Congress Control Number: 2017948389

Mathematics Subject Classification (2010): 60H10, 60H05, 60H30, 60G42, 60G44

© Springer International Publishing AG 2017

Printed on acid-free paper

This Springer imprint is published by Springer Nature

and other branches of mathematics (algebraic structures, differential geometry, . . . )

Roma, Italy Paolo Baldi

4.3 Conditional laws. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98

9.8 Lp bounds and dependence on the initial data. . .. . . . . . . . . . . . . . . . . . . . 282

Solutions of the Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 437

Real, complex numbers, Rm

x_y D max.x; y/ the largest of the real numbers x and y

Mb .E/ Real bounded measurable functions on the topological space E

Cb .Rm / The Banach space of real bounded continuous functions on Rm

1.1 Probability spaces, random variables

A measurable space is a pair .E; E / where

EŒX D .EŒX1 ; : : : ; EŒXm / :

We say that two r.v.’s X; Y are equivalent if P.X D Y/ D 1.

© Springer International Publishing AG 2017 1

EŒX D EŒX C EŒX ;

Jensen’s inequality. Let X be an m-dimensional integrable r.v. and ˚ W Rm !

EŒ˚.X/  ˚.EŒX/ (1.1)

Hölder’s inequality. Let Z, W be real positive r.v.’s and ˛, ˇ positive real

EŒZ ˛ W ˇ  EŒZ˛ EŒWˇ : (1.2)

which also goes under the name of Hölder’s inequality.

Minkowski’s inequality. Let X, Y be real r.v.’s and p  1, then

kX C Ykp  kXkp C kYkp : (1.4)

Minkowski’s inequality, in particular, implies that Lp , p  1, is a vector space. If

kXkpp D EŒjXjp D EŒ˚.jXjq /  ˚.EŒjXjq / D EŒjXjq p=q

and therefore, taking the p-th root,

kXkp  kXkq : (1.5)

1.2 Variance, covariance, law of a r.v.

Var.X/ D EŒ.X EŒX/2 D EŒX 2 EŒX2 :

Markov’s inequality. For every ı > 0; ˇ > 0 we have

Chebyshev’s inequality. Let X 2 L2 . Then for every ˛ > 0

This is a very important inequality. It is, of course, a particular case of Markov’s

Cov.X; Y/ D EŒXY EŒXEŒY :

If Cov.X; Y/ D 0, X and Y are said to be uncorrelated.

X .A/ D P.X 1 .A//

Proposition 1.1 Let X W .˝; F ; P/ ! .E; E / be a r.v., X the law of X. Then

In particular, if X is a real r.v. then X has finite mathematical expectation if and

Also we have, for every ˛,

(the converse is obvious). P and Q are said to be equivalent if both P Q and

1.3 Independence, product measure

If E , E 0 are -algebras of events of E let us denote by E _ E 0 the smallest -

Lemma 1.1 (Doob’s measurability criterion) Let X W .˝; F ; P/ !

The r.v.’s X1 ; : : : ; Xm , taking values respectively in .E1 ; E1 /; : : : ; .Em ; Em /, are

P.X1 2 A01 ; : : : ; Xm 2 A0m / D P.X1 2 A01 / : : : P.Xm 2 A0m / :

The events A1 ; : : : ; Am 2 F are said to be independent if and only if

P.Ai1 \    \ Ai` / D P.Ai1 / : : : P.Ai` /

for every choice of 1  `  m and of 1  i1 < i2 <    < i`  m. A

P.A1 \    \ Am / D P.A1 / : : : P.Am / :

Proposition 1.2 In the previous notations the r.v.’s X1 ; : : : ; Xm are indepen-

The proof of this proposition makes use of the following theorem.

Theorem 1.1 (Carathéodory’s criterion) Let .E; E / be a measurable

EŒ˚.X/ ˚.EŒX/ (1.1)

EŒZ ˛ W ˇ EŒZ˛ EŒWˇ : (1.2)

Minkowski’s inequality. Let X, Y be real r.v.’s and p 1, then

kX C Ykp kXkp C kYkp : (1.4)

Minkowski’s inequality, in particular, implies that Lp , p 1, is a vector space. If

kXkpp D EŒjXjp D EŒ˚.jXjq / ˚.EŒjXjq / D EŒjXjq p=q

kXkp kXkq : (1.5)

X .A/ D P.X 1 .A//

Proposition 1.1 Let X W .˝; F ; P/ ! .E; E / be a r.v., X the law of X. Then

P.Ai1 \ \ Ai` / D P.Ai1 / : : : P.Ai` /

for every choice of 1 ` m and of 1 i1 < i2 < < i` m. A

P.A1 \ \ Am / D P.A1 / : : : P.Am / :

for some n 1; X1 ; : : : ; Xn 2 I ; A01 2 E1 ; : : : ; A0n 2 En and similarly define

for some k 1; Y1 ; : : : ; Yk 2 J ; B01 2 G1 ; : : : ; B0k 2 Gk . One verifies that C1

A1 D fX1 2 A01 ; : : : ; Xn 2 A0n g D f.X1 ; : : : ; Xn / 2 A01 A0n g

Theorem 1.2 (Fubini) Let 1 ; 2 be measures respectively on the measur-

Let be a probability on .Rm ; B.Rm // and let us denote by i W Rm ! R its

If X D .X1 ; : : : ; Xm / is an m-dimensional r.v. with law it is clear that its i-th

Proposition 1.4 If and are probabilities on Rm having density f and g

If X is an m-dimensional r.v. and denotes its law, let us define

For m D 1 this relation becomes b 00 .0/ D EŒX 2 . If X is centered, then

9. Let b 2 Rk and A a k m matrix. Then Y D AX C b is a Rk -valued r.v.; if

A probability on R is said to be N.a; 2 / (normal, or Gaussian, with mean a

and therefore k N.ak ; kk /.

By 7. of Sect. 1.6 we see then that D 1 ˝ ˝ m . Therefore

EŒjXYj EŒX 2 1=2 EŒY 2 1=2 < C1 :

Conversely, if X D .X1 ; : : : ; Xm / is Gaussian and 2 Rm , then 1 X1 C C

which is not possible, as necessarily limM!C1 P.Y M/ D 0. As .mn /n is