Probability and Its Applications

Published in association with the Applied Probability Trust

Editors: S. Asmussen, J. Gani, P. Jagers, T.G. Kurtz
Probability and Its Applications
Azencott et al.: Series of Irregular Observations. Forecasting and Model
Building. 1986
Bass: Diffusions and Elliptic Operators. 1997
Bass: Probabilistic Techniques in Analysis. 1995
Berglund/Gentz: Noise-Induced Phenomena in Slow-Fast Dynamical Systems:
A Sample-Paths Approach. 2006
Biagini/Hu/Øksendal/Zhang: Stochastic Calculus for Fractional Brownian Motion
and Applications. 2008
Chen: Eigenvalues, Inequalities and Ergodic Theory. 2005
Chen/Goldstein/Shao: Normal Approximation by Stein’s Method. 2011
Costa/Fragoso/Marques: Discrete-Time Markov Jump Linear Systems. 2005
Daley/Vere-Jones: An Introduction to the Theory of Point Processes I: Elementary
Theory and Methods, 2nd ed. 2003, corr. 2nd printing 2005
Daley/Vere-Jones: An Introduction to the Theory of Point Processes II: General
Theory and Structure, 2nd ed. 2008
de la Peña/Gine: Decoupling: From Dependence to Independence, Randomly
Stopped Processes, U-Statistics and Processes, Martingales and Beyond. 1999
de la Peña/Lai/Shao: Self-Normalized Processes. 2009
Del Moral: Feynman-Kac Formulae. Genealogical and Interacting Particle
Systems with Applications. 2004
Durrett: Probability Models for DNA Sequence Evolution. 2002, 2nd ed. 2008
Ethier: The Doctrine of Chances. 2010
Feng: The Poisson–Dirichlet Distribution and Related Topics. 2010
Galambos/Simonelli: Bonferroni-Type Inequalities with Equations. 1996
Gani (ed.): The Craft of Probabilistic Modelling. A Collection of Personal
Accounts. 1986
Gawarecki/Mandrekar: Stochastic Differential Equations in Infinite Dimensions.
2011
Gut: Stopped RandomWalks. Limit Theorems and Applications. 1987
Guyon: Random Fields on a Network. Modeling, Statistics and Applications. 1995
Kallenberg: Foundations of Modern Probability. 1997, 2nd ed. 2002
Kallenberg: Probabilistic Symmetries and Invariance Principles. 2005
Högnäs/Mukherjea: Probability Measures on Semigroups, 2nd ed. 2011
Last/Brandt: Marked Point Processes on the Real Line. 1995
Li: Measure-Valued Branching Markov Processes. 2011
Molchanov: Theory of Random Sets. 2005
Nualart: The Malliavin Calculus and Related Topics, 1995, 2nd ed. 2006
Rachev/Rueschendorf: Mass Transportation Problems. Volume I: Theory and
Volume II: Applications. 1998
Resnick: Extreme Values, Regular Variation and Point Processes. 1987
Schmidli: Stochastic Control in Insurance. 2008
Schneider/Weil: Stochastic and Integral Geometry. 2008
Serfozo: Basics of Applied Stochastic Processes. 2009
Shedler: Regeneration and Networks of Queues. 1986
Silvestrov: Limit Theorems for Randomly Stopped Stochastic Processes. 2004
Thorisson: Coupling, Stationarity and Regeneration. 2000
Leszek Gawarecki  Vidyadhar Mandrekar

Stochastic Differential
Equations in Infinite
Dimensions

with Applications to Stochastic

Partial Differential Equations
Leszek Gawarecki Vidyadhar Mandrekar
Department of Mathematics Department of Statistics and Probability
Kettering University Michigan State University
1700 University Ave A436 Wells Hall
Flint, MI 48504 East Lansing, MI 48823
USA USA
lgawarec@[Link] mandrekar@[Link]

Series Editors:
Søren Asmussen Peter Jagers
Department of Mathematical Sciences Mathematical Statistics
Aarhus University Chalmers University of Technology
Ny Munkegade and University of Gothenburg
8000 Aarhus C 412 96 Göteborg
Denmark Sweden
asmus@[Link] jagers@[Link]

Joe Gani Thomas G. Kurtz

Centre for Mathematics and its Applications Department of Mathematics
Mathematical Sciences Institute University of Wisconsin - Madison
Australian National University 480 Lincoln Drive
Canberra, ACT 0200 Madison, WI 53706-1388
Australia USA
gani@[Link] kurtz@[Link]

ISSN 1431-7028
ISBN 978-3-642-16193-3 e-ISBN 978-3-642-16194-0
DOI 10.1007/978-3-642-16194-0
Springer Heidelberg Dordrecht London New York

Mathematics Subject Classification (2010): 35-XX, 60-XX

© Springer-Verlag Berlin Heidelberg 2011

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, 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.

Cover design: VTEX, Vilnius

Printed on acid-free paper

Springer is part of Springer Science+Business Media ([Link])

We dedicate this book to our wives
Edyta and Veena
for their encouragement and patience during
the preparation of this book
Preface

Stochastic differential equations are playing an increasingly important role in

applications to finance, numerical analysis, physics, and biology. In the finite-
dimensional case, there are two definitive books: one by Gikhman and Sko-
rokhod [25], which studies the existence and uniqueness problem along with proba-
bilistic properties of solutions, and another by Khasminskii [39], which studies the
asymptotic behavior of the solutions using the Lyapunov function method. Our ob-
ject in this book is to study these topics in the infinite-dimensional case. The two
main problems one faces are the invalidity of the Peano theorem in the infinite-
dimensional case and the appearance of unbounded operators if one wants to apply
finite-dimensional techniques to stochastic partial differential equations (SPDEs).
Motivated by these difficulties, we discuss the theory in the deterministic case
from two points of view. The first method (see Pazy [63] and Butzer and Berens [6])
involves semigroups generated by unbounded operators and results in constructing
mild solutions. The other sets up the equation in a Gelfand triplet V → H → V ∗
of Hilbert spaces with the space V as the domain of the unbounded operator and V ∗
its continuous dual. In this case variational solutions are produced. This approach is
studied by Agmon [1] and Lions [48], who assume that either the injection V → H
is compact and the unbounded operator is coercive or that the unbounded operator
is coercive and monotone.
The systematic study of the first approach to SPDEs was first undertaken by
Ichikawa [32, 33] and is explained in the timely monographs of Da Prato and
Zabczyk [11, 12]. The approach of J.P. Lions was first used by Viot [75] (see also
Metivier [56] and Metivier and Viot [58]). Working under the assumption that the
embedding V → H is compact and that the coefficients of the equation are coer-
cive, the existence of a weak solution was proven. These results were generalized by
Pardoux [62], who assumed coercivity and monotonicity, and used the crucial deter-
ministic result of Lions [48] to produce the strong solution. Later, Krylov, and Ro-
zovskii [42] established the above-mentioned result of Lions in the stochastic case
and also produced strong solutions. The initial presentation was given by Rozovskii
in [66]. However, rather rigorous and complete exposition in a slightly general form
is provided by Prévôt and Röckner [64].

vii
viii Preface

In addition to presenting these results on SPDEs, we discuss the work of Leha and
Ritter [46, 47] on SDEs in R∞ with applications to interacting particle systems and
the related work of Albeverio et al. [2, 3], and also of Gawarecki and Mandrekar [20,
21] on the equations in the field of Glauber dynamics for quantum lattice systems.
In both cases the authors study infinite systems of SDEs.
We do not present here the approach used in Kalliapur and Xiong [37], as it re-
quires introducing additional terminology for nuclear spaces. For this type of prob-
lem (referred to as “type 2” equations by K. Itô in [35]), we refer the reader to
[22, 23], and [24], as well as to [37].
A third approach, which involves solutions being Hida distribution is presented
by Holden et al. in the monograph [31].
The book is divided into two parts. We begin Part I with a discussion of the
semigroup and variational methods for solving PDEs. We simultaneously develop
stochastic calculus with respect to a Q-Wiener process and a cylindrical Wiener pro-
cess, relying on the classic approach presented in [49]. These foundations allow us to
develop the theory of semilinear partial differential equations. We address the case
of Lipschitz coefficients first and produce unique mild solutions as in [11]; how-
ever, we then extend our research to the case where the equation coefficients depend
on the entire “past” of the solution, invoking the techniques of Gikhman and Sko-
rokhod [25]. We also prove Markov and Feller properties for mild solutions, their
dependence on the initial condition, and the Kolmogorov backward equation for the
related transition semigroup. Here we have adapted the work of B. Øksendal [61],
S. Cerrai [8], and Da Prato and Zabczyk [11].
To go beyond the Lipschitz case, we have adapted the method of approximating
continuous functions by Lipschitz functions f : [0, T ] × Rn → Rn from Gikhman
and Skorokhod [25] to the case of continuous functions f : [0, T ] × H → H [22].
This technique enabled us to study the existence of weak solutions for SDEs with
continuous coefficients, with the solution identified in a larger Hilbert space, where
the original Hilbert space is compactly embedded. This arrangement is used, as we
have already mentioned, due to the invalidity of the Peano theorem. In addition, we
study martingale solutions to semilinear SDEs in the case of a compact semigroup
and for coefficients depending on the entire past of the solution.
The variational method is addressed in Chap. 4, where we study both the weak
and strong solutions. The problem of the existence of weak variational solutions is
not well addressed in the existent literature, and our original results are obtained
with the help of the ideas presented in Kallianpur et al. [36]. We have followed the
approach of Prévôt and Röckner in our presentation of the problem of the existence
and uniqueness of strong solutions.
We conclude Part I with an interesting problem of an infinite system of SDEs that
does not arise from a stochastic partial differential equation and serves as a model of
an interacting particle system and in Glauber dynamics for quantum lattice systems.
In Part II of the book, we present the asymptotic behaviors of solutions to infinite-
dimensional stochastic differential equations. The study of this topic was undertaken
for specific cases by Ichikawa [32, 33] and Da Prato and Zabczyk [12] in the case
of mild solutions. A general Lyapunov function approach for strong solutions in a
Preface ix

Gelfand triplet setting for exponential stability was originated in the work of Khas-
minskii and Mandrekar [40] (see also [55]). A generalization of this approach for
mild and strong solutions involving exponential boundedness was put forward by
R. Liu and Mandrekar [52, 53]. This work allows readers to study the existence of
invariant measure [52] and weak recurrence of the solutions to compact sets [51].
Some of these results were presented by K. Liu in a slightly more general form
in [50].
Although we have studied the existence and uniqueness of non-Markovian solu-
tions, we do not investigate the ergodic properties of these processes, as the tech-
niques in this field are still in development [28].

East Lansing Leszek Gawarecki

October, 2010 Vidyadhar Mandrekar
Acknowledgements

During the time we were working on this book, we were helped by discussions
with various scholars. We thank Professors A.V. Skorokhod and R. Khasminskii for
providing insights in the problems studied in Parts I and II, respectively. We are
indebted to Professor B. Øksendal for giving us ideas on the organization of the
book. One of us had the privilege of visiting Professors G. Da Prato and Jürgen
Pothoff. The discussions with them helped in understanding several problems. The
latter provided an opportunity to present a preliminary version of the first part to
his students. Clearly, all the participants’ comments improved the presentation. The
content of Chap. 5 bears the influence of Professor Sergio Albeverio, from whom
we got an insight into applications to physics. Professor Kallianpur provided pre-
liminary knowledge of the field and encouragement by extending an invitation to a
conference on SPDEs in Charlotte. We will be amiss if we do not thank the partic-
ipants of the seminar on the subject given at Michigan State University. Our spe-
cial thanks go to Professors Peter Bates, Sheldon Newhouse, and Shlomo Levental,
whose comments and questions led to cleaning up some confusion in the book.
Finally, we thank two referees for their insightful comments which led to a sig-
nificant improvement of our presentation. We are grateful to Dr. M. Reizakis of
Springer Verlag for her timely administrative support and encouragement.

xi
Contents

Part I Stochastic Differential Equations in Infinite Dimensions

1 Partial Differential Equations as Equations in Infinite Dimensions . 3

1.1 The Heat Equation as an Abstract Cauchy Problem . . . . . . . . . 3
1.2 Elements of Semigroup Theory . . . . . . . . . . . . . . . . . . . 5
1.3 Commonly Used Function Spaces . . . . . . . . . . . . . . . . . . 10
1.4 The Abstract Cauchy Problem . . . . . . . . . . . . . . . . . . . . 11
1.5 The Variational Method . . . . . . . . . . . . . . . . . . . . . . . 15

2 Stochastic Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1 Hilbert-Space-Valued Process, Martingales, and Cylindrical
Wiener Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.1 Cylindrical and Hilbert-Space-Valued Gaussian Random
Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.2 Cylindrical and Q-Wiener Processes . . . . . . . . . . . . 19
2.1.3 Martingales in a Hilbert Space . . . . . . . . . . . . . . . 21
2.2 Stochastic Integral with Respect to a Wiener Process . . . . . . . . 23
2.2.1 Elementary Processes . . . . . . . . . . . . . . . . . . . . 25
2.2.2 Stochastic Itô Integral for Elementary Processes . . . . . . 26
2.2.3 Stochastic Itô Integral with Respect to a Q-Wiener Process 34
2.2.4 Stochastic Itô Integral with Respect to Cylindrical Wiener
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.5 The Martingale Representation Theorem . . . . . . . . . . 49
2.2.6 Stochastic Fubini Theorem . . . . . . . . . . . . . . . . . 57
2.3 The Itô Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.3.1 The Case of a Q-Wiener Process . . . . . . . . . . . . . . 61
2.3.2 The Case of a Cylindrical Wiener Process . . . . . . . . . 69

3 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . . 73

3.1 Stochastic Differential Equations and Their Solutions . . . . . . . 73
3.2 Solutions Under Lipschitz Conditions . . . . . . . . . . . . . . . . 84
3.3 A Special Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

xiii
xiv Contents

3.4 Markov Property and Uniqueness . . . . . . . . . . . . . . . . . . 107

3.5 Dependence of the Solution on the Initial Value . . . . . . . . . . 111
3.6 Kolmogorov’s Backward Equation . . . . . . . . . . . . . . . . . 117
3.7 Lipschitz-Type Approximation of Continuous Coefficients . . . . . 125
3.8 Existence of Weak Solutions Under Continuity Assumption . . . . 128
3.9 Compact Semigroups and Existence of Martingale Solutions . . . . 135
3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener Process . . 141
Appendix: Compactness and Tightness of Measures
in C([0, T ], M ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

4 Solutions by Variational Method . . . . . . . . . . . . . . . . . . . . 151

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.2 Existence of Weak Solutions Under Compact Embedding . . . . . 152
4.3 Strong Variational Solutions . . . . . . . . . . . . . . . . . . . . . 174
4.4 Markov and Strong Markov Properties . . . . . . . . . . . . . . . 181

5 Stochastic Differential Equations with Discontinuous Drift . . . . . . 185

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w ) . . . . . . 185
5.3 Locally Interacting Particle Systems, Solutions
in C([0, T ], RZ ) . . . . . . . . . . . . . . . . . . . .
d
. . . . . . 194

Part II Stability, Boundedness, and Invariant Measures

6 Stability Theory for Strong and Mild Solutions . . . . . . . . . . . . 203

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
6.2 Exponential Stability for Stochastic Differential Equations . . . . . 211
6.3 Stability in the Variational Method . . . . . . . . . . . . . . . . . 225
Appendix: Stochastic Analogue of the Datko Theorem . . . . . . . 230

7 Ultimate Boundedness and Invariant Measure . . . . . . . . . . . . . 233

7.1 Exponential Ultimate Boundedness in the m.s.s. . . . . . . . . . . 233
7.2 Exponential Ultimate Boundedness in Variational Method . . . . . 237
7.3 Abstract Cauchy Problem, Stability and Exponential Ultimate
Boundedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
7.4 Ultimate Boundedness and Invariant Measure . . . . . . . . . . . 255
7.4.1 Variational Equations . . . . . . . . . . . . . . . . . . . . 261
7.4.2 Semilinear Equations Driven by a Q-Wiener Process . . . . 266
7.4.3 Semilinear Equations Driven by a Cylindrical Wiener
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
7.5 Ultimate Boundedness and Weak Recurrence of the Solutions . . . 274

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Acronyms

(H,  · H ) Real separable Hilbert space

(X,  · X ) Banach space
(X ∗ ,  · X∗ ) Dual of a Banach space
T∗ Adjoint of a linear operator T
B(H ) Collection of Borel-measurable real-valued functions on H
Bb (H ) Collection of bounded Borel-measurable real-valued functions
on H
m
C (H ) Collection of real-valued m-times continuously Fréchet
differentiable functions on H
C0m (H ) Subspace of C m (H ) consisting of functions with compact
support
m
Cb (H ) Subspace of C m (H ) consisting of functions whose derivatives of
order k = 0, . . . , m are bounded
C(H ) = C (H ) Collection of real-valued continuous functions on H
0

C([0, T ], H ) Banach space of H -valued continuous functions on [0, T ]

C m ([0, T ], H ) H -valued m-times continuously differentiable functions on [0, T ]
C ∞ ([0, T ], H ) H -valued infinitely many times continuously differentiable
functions on [0, T ]
Lp ([0, T ], H ) Banach space of H -valued Borel measurable functions on [0, T ],
Bochner integrable in the pth power
∞
L ([0, T ], H ) Banach space of H -valued essentially bounded functions
on [0, T ]
MT2 (H ) Hilbert space of H -valued continuous square-integrable
martingales
W m,p (O) Sobolev space of functions f ∈ Lp (O), with D α f ∈ Lp (O),
|α| ≤ m
m,p
W0 (O) Completion of C0m (O) in W m,p (O)
B(H ) Borel σ -field on (H,  · H )
E (L (K, H )) L (K, H )-valued elementary processes

xv
xvi Acronyms

L (X, Y ) Bounded linear operators from X to Y

L1 (H ) Trace-class operators on H
L2 (H1 , H2 ) Hilbert–Schmidt operators from H1 to H2
Aλ Yosida approximation of an operator A
D(A) Domain of an operator A
R(A) Range of an operator A
R(λ, A) Resolvent of an operator A
ρ(A) Resolvent set of an operator A
tr(A) Trace of an operator A
τ (A) Trace norm of an operator A
S  Φ(t) Stochastic convolution t
Qt Covariance of the stochastic convolution 0 S(t − s) dWs
Wt Q-Wiener process
W̃t Cylindrical Wiener process
t
Φ(s) dWs Stochastic integral with respect to a Q-Wiener process
0t
0 Φ(s) d W̃s Stochastic integral with respect to a cylindrical Wiener process
M t Increasing process of a martingale Mt
M t Quadratic variation process of a martingale Mt
 Part I
Stochastic Differential Equations in
Infinite Dimensions
Chapter 1
Partial Differential Equations as Equations
in Infinite Dimensions

The purpose of this chapter is to explain how infinite-dimensional equations arise

from finite-dimensional partial differential equations and to consider their classical
and mild solutions. The mild solutions are studied using the semigroup methods as
in [63], and strong solutions are studied using variational methods as in [48] (see
also [71]).

1.1 The Heat Equation as an Abstract Cauchy Problem

Let us consider a PDE and explain how it is related to a semigroup of linear opera-
tors.

Example 1.1 Consider the one-dimensional heat equation


ut (t, x) = uxx (t, x), t > 0,
(1.1)
u(0, x) = ϕ(x), x ∈ R,

with the initial temperature distribution ϕ being modeled by a bounded, uniformly

continuous function on R.
A function u(t, x) is said to be a solution of (1.1) if it satisfies (1.1) for t > 0, u,
ut , ux , uxx are bounded, uniformly continuous functions in R for every t > 0 and
limt→0+ u(t, x) = ϕ(x) uniformly in x.

The uniqueness of the solution and the major premise of the Hughen’s principle
lead to the following implications. If the temperature distribution u(t, x) at t > 0
is uniquely determined by the initial condition ϕ(x), then u(t, x) can also be ob-
tained by first calculating u(s, x) for some intermediate time s < t and then by
using u(s, x) as the initial condition. Thus, there exist transformations G(t) on ϕ
defined by (G(t)ϕ)(x) = uϕ (t, x) satisfying the semigroup property
 
G(t)ϕ = G(t − s) G(s)ϕ

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 3

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_1, © Springer-Verlag Berlin Heidelberg 2011
4 1 Partial Differential Equations as Equations in Infinite Dimensions

and the strong continuity property

 
lim = G(t)ϕ − ϕ  = 0.
t→0+

Above, ϕ = supx∈R |ϕ(x)|.

The transformations are linear due to the linear dependence of the problem on the
initial condition. The relationship between the semigroup G(t) and the differential
operator ∂ 2 /∂x 2 in (1.1) can be explained by calculating
 ϕ     
 u (·, t + h) − uϕ (·, t)   G(h) − I 
lim    
= lim G(t) ϕ
h→0 + h  h→0+ h 
     
 G(h) − I   G(h) − I ϕ 
= lim   
G(t)ϕ  = lim   u (·, t)
h→0 + h h→0 + h ,

where I denotes the identity. Let us define the linear operator

G(t)ϕ − ϕ
Aϕ = lim (in norm)
t→0+ t
and denote by D(A), called the domain of A, the collection of all those functions ϕ
for which the limit exists. We arrive at an abstract formulation of the heat equation
in the form of an abstract Cauchy problem in the Banach space (X,  · ) of bounded
uniformly continuous functions on R,
⎧
⎨ du(t) = Au(t), t > 0,
dt (1.2)
⎩u(0) = ϕ ∈ X,

where the differentiation is understood in the Banach space X.

For a solution u(t), it is required that for t > 0, u(t) ∈ D(A), u is continuously
differentiable and limt→0+ u(t) − ϕ = 0. In the case of (1.1), there is an explicit
form of the semigroup G(t), given by the Gaussian semigroup
⎧
⎨ 1 
  exp −|x − y|2
/4t ϕ(y) dy, t > 0,
G(t)ϕ (x) = (4πt)1/2 R (1.3)
⎩
ϕ(x), t = 0.

The solution to (1.1) is known to be (see [7] Chap. 3 for detailed presentation)
 
u(t, x) = G(t)ϕ (x).

The operator A = d2 /dx 2 , and

 
df d2 f
D(A) = f ∈ X : f, are continuously differentiable, and ∈ X .
dx dx 2

Exercise 1.1 Show that the operators G(t) defined in (1.3) have the semigroup
property.
1.2 Elements of Semigroup Theory 5

We now review the fundamentals of the theory of semigroups of linear operators

with the goal of studying the existence of classical and mild solutions to an abstract
Cauchy problem.

1.2 Elements of Semigroup Theory

In this section we review the fundamentals of semigroup theory and refer the reader
to [6, 63], or [78] for proofs.
Let (X,  · X ) and (Y,  · Y ) be Banach spaces. Denote by L (X, Y ) the family
of bounded linear operators from X to Y . L (X, Y ) becomes a Banach space when
equipped with the norm

T L (X,Y ) = sup T xY , T ∈ L (X, Y ).

x∈X, xX =1

For brevity, L (X) will denote the Banach space of bounded linear operators on X.
The identity operator on X is denoted by I .
Let X ∗ denote the dual space of all bounded linear functionals x ∗ on X. X ∗ is
again a Banach space under the supremum norm
 
x ∗ X∗ = sup x, x ∗ ,
x∈X, xX =1

where ·, · denotes the duality on X × X ∗ .

For T ∈ L (X, Y ), the adjoint operator T ∗ ∈ L (Y ∗ , X ∗ ) is defined by
   
x, T ∗ y ∗ = T x, y ∗ , x ∈ X, y ∗ ∈ Y ∗ .

Let H be a real Hilbert space. A linear operator T ∈ L (H ) is called symmetric if

for all h, g ∈ H ,
T h, g H = h, T g H.

A symmetric operator T is called nonnegative definite if for every h ∈ H ,

T h, h H ≥ 0.

Definition 1.1 A family S(t) ∈ L (X), t ≥ 0, of bounded linear operators on a Ba-

nach space X is called a strongly continuous semigroup (or a C0 -semigroup, for
short) if
(S1) S(0) = I ,
(S2) (Semigroup property) S(t + s) = S(t)S(s) for every t, s ≥ 0,
(S3) (Strong continuity property) limt→0+ S(t)x = x for every x ∈ X.

Let S(t) be a C0 -semigroup on a Banach space X. Then, there exist constants

α ≥ 0 and M ≥ 1 such that
 
S(t) ≤ Meαt , t ≥ 0. (1.4)
L (X)
6 1 Partial Differential Equations as Equations in Infinite Dimensions

If M = 1, then S(t) is called a pseudo-contraction semigroup. If α = 0, then S(t) is

called uniformly bounded, and if α = 0 and M = 1 (i.e., S(t)L (X) ≤ 1), then S(t)
is called a semigroup of contractions. If for every x ∈ X, the mapping t → S(t)x is
differentiable for t > 0, then S(t) is called a differentiable semigroup. A semigroup
of linear operators {S(t), t ≥ 0} is called compact if the operators S(t), t > 0, are
compact.
For any C0 -semigroup S(t) and arbitrary x ∈ X, the mapping

R+ t → S(t)x ∈ X

is continuous.

Definition 1.2 Let S(t) be a C0 -semigroup on a Banach space X. The linear oper-
ator A with domain
 
S(t)x − x
D(A) = x ∈ X : lim exists (1.5)
t→0+ t

defined by
S(t)x − x
Ax = lim (1.6)
t→0+ t
is called the infinitesimal generator of the semigroup S(t).

A semigroup S(t) is called uniformly continuous if

 
lim S(t) − I L (X) = 0.
t→0+

Theorem 1.1 A linear operator A is the infinitesimal generator of a uniformly con-

tinuous semigroup S(t) on a Banach space X if and only if A ∈ L (X). We have
∞
 (tA)n
S(t) = etA = ,
n!
n=0

the series converging in norm for every t ≥ 0.

We will however be mostly interested in the case where A ∈

/ L (X), as in (1.2).
The following theorem provides useful facts about semigroups.

Theorem 1.2 Let A be an infinitesimal generator of a C0 -semigroup S(t) on a

Banach space X. Then
(1) For x ∈ X,
1 t+h
lim S(t)x ds = S(t)x. (1.7)
h→0 h t
1.2 Elements of Semigroup Theory 7

(2) For x ∈ D(A), S(t)x ∈ D(A) and

d
S(t)x = AS(t)x = S(t)Ax. (1.8)
dt
t
(3) For x ∈ X, 0 S(s)x ds ∈ D(A), and
 t 
A S(s)x ds = S(t)x − x. (1.9)
0

(4) If S(t) is differentiable then for n = 1, 2, . . ., S(t) : X → D(An ) and

S (n) (t) = An S(t) ∈ L (X).

(5) If S(t) is compact then S(t) is continuous in the operator topology for t > 0,
i.e.,
 
lim S(s) − S(t)L (H ) = 0. (1.10)
s→t, s,t>0

(6) For x ∈ D(A),

t t
S(t)x − S(s)x = S(u)Ax du = AS(u)x du. (1.11)
s s

(7) D(A) is dense in X, and A is a closed linear operator.

(8) The intersection ∞ n=1 D(A ) is dense in X.
n

(9) Let X be a reflexive Banach space. Then the adjoint semigroup S(t)∗ of S(t) is
a C0 -semigroup whose infinitesimal generator is A∗ , the adjoint of A.

If X = H , a real separable Hilbert space, then for h ∈ H , define the graph norm
 1/2
hD(A) = h2H + Ah2H . (1.12)

Then (D(A),  · D(A) ) is a real separable Hilbert space.

Exercise 1.2 Let A be a closed linear operator on a real separable Hilbert space.
Prove that (D(A),  · D(A) ) is a real separable Hilbert space.

Let B(H ) denote the Borel σ -field on H . Then D(A) ∈ B(H ), and
   
A : D(A), B(H )|D(A) → H, B(H ) .

Consequently, the restricted Borel σ -field B(H )|D(A) coincides with the Borel
σ -field on the Hilbert space (D(A),  · D(A) ), and measurability of D(A)-valued
functions can be understood with respect to either Borel σ -field.
8 1 Partial Differential Equations as Equations in Infinite Dimensions
t
Theorem 1.3 Let f : [0, T ] → D(A) be measurable, and let 0 f (s)D(A) < ∞.
Then
t t t
f (s) ds ∈ D(A) and Af (s) ds = A f (s) ds. (1.13)
0 0 0

Exercise 1.3 Prove Theorem 1.3.

Conditions under which an operator A can be an infinitesimal generator of a

C0 -semigroup involve the resolvent of A.

Definition 1.3 The resolvent set ρ(A) of a closed linear operator A on a Banach
space X is the set of all complex numbers λ for which λI − A has a bounded inverse,
i.e., the operator (λI − A)−1 ∈ L (X). The family of bounded linear operators

R(λ, A) = (λI − A)−1 , λ ∈ ρ(A), (1.14)

is called the resolvent of A.

We note that R(λ, A) is a one-to-one transformation of X onto D(A), i.e.,

(λI − A)R(λ, A)x = x, x ∈ X,

(1.15)
R(λ, A)(λI − A)x = x, x ∈ D(A).

In particular,
AR(λ, A)x = R(λ, A)Ax, x ∈ D(A). (1.16)
In addition, we have the following commutativity property:

R(λ1 , A)R(λ2 , A) = R(λ2 , A)R(λ1 , A), λ1 , λ2 ∈ ρ(A). (1.17)

The following statement is true in greater generality; however, we will use it only in
the real domain.

Proposition 1.1 Let S(t) be a C0 -semigroup with infinitesimal generator A on a

Banach space X. If α0 = limt→∞ t −1 ln S(t)L (X) , then any real number λ > α0
belongs to the resolvent set ρ(A), and
∞
R(λ, A)x = e−λt S(t)x dt, x ∈ X. (1.18)
0

Furthermore, for each x ∈ X,

lim λR(λ, A)x − xX = 0. (1.19)

λ→∞

Theorem 1.4 (Hille–Yosida) Let A : D(A) ⊂ X → X be a linear operator on a Ba-

nach space X. Necessary and sufficient conditions for A to generate a C0 -semigroup
S(t) are
1.2 Elements of Semigroup Theory 9

(1) A is closed and D(A) = X.

(2) There exist real numbers M and α such that for every λ > α, λ ∈ ρ(A) (the
resolvent set) and
  
 R(λ, A) r  ≤ M(λ − α)−r , r = 1, 2, . . . . (1.20)
L (X)

In this case, S(t)L (X) ≤ Meαt , t ≥ 0.

We will now introduce an important approximation of an operator A and of the

C0 -semigroup it generates.
For λ ∈ ρ(A), consider the family of operators

Rλ = λR(λ, A). (1.21)

Since the range R(R(λ, A)) ⊂ D(A), we can define the Yosida approximation of A
by
Aλ x = ARλ x, x ∈ X. (1.22)
Note that by (1.16)
Aλ x = Rλ Ax, x ∈ D(A).
Since λ(λI − A)R(λ, A) = λI , we have λ2 R(λ, A) − λI = λAR(λ, A), so that

Aλ x = λ2 R(λ, A) − λI,

proving that Aλ ∈ L (X). Denote by Sλ (t) the (uniformly continuous) semigroup

generated by Aλ ,
Sλ (t)x = etAλ x, x ∈ X. (1.23)
Using the commutativity of the resolvent (1.17), we have

Aλ1 Aλ2 = Aλ2 Aλ1 (1.24)

and, by the definition of Sλ (t) (1.23),

Aλ Sλ (t) = Sλ (t)Aλ . (1.25)

Proposition 1.2 (Yosida Approximation) Let A be an infinitesimal generator of a

C0 -semigroup S(t) on a Banach space X. Then

lim Rλ x = x, x ∈ X, (1.26)
λ→∞

lim Aλ x = Ax, x ∈ D(A), (1.27)

λ→∞
and
lim Sλ (t)x = S(t)x, x ∈ X. (1.28)
λ→∞
10 1 Partial Differential Equations as Equations in Infinite Dimensions

The convergence in (1.28) is uniform on compact subsets of R+ . The following esti-

mate holds:
  
Sλ (t) ≤ M exp tλα/(λ − α) (1.29)
L (X)
with the constants M, α determined by the Hille–Yosida theorem.

1.3 Commonly Used Function Spaces

We define commonly used function spaces. Let (H,  · H ) be a real separable

Hilbert space, and B(H ) be its Borel σ -field. The collection of measurable (respec-
tively, bounded measurable) real-valued functions on H will be denoted by B(H )
(respectively, Bb (H )).
For a positive constant T , let C([0, T ], H ) be the Banach space of H -valued
continuous functions on [0, T ], with the norm f C([0,T ],H ) = sup0≤t≤T f (t)H .
For a positive integer m, C m ([0, T ], H ) and C ∞ ([0, T ], H ) denote the spaces of H -
valued, respectively m-times and infinitely many times continuously differentiable
functions on [0, T ].
By C m (H ) we denote the collection of real-valued m-times continuously Fréchet
differentiable functions on H , and let C0m (H ) and Cbm (H ) be its subspaces consist-
ing respectively of those functions which have compact support and of those whose
derivatives of order k = 0, . . . , m are bounded. It is typical to suppress the super-
script m if m = 0 and write, for example, C(H ) for C 0 (H ). If the Hilbert space H
is replaced with an interval [0, T [ (resp. ]0, T ]), we consider the right derivatives at
t = 0 (resp. left derivatives at t = T ).
The Banach space Lp ([0, T ], H ) is the space of H -valued Borel-measurable
functions on [0, T ] that are Bochner integrable in the pth power, with the norm
 T 1/p

f Lp ([0,T ],H ) = f (t)p dt .
H
0

The space of square-integrable functions L2 ([0, T ], H ) equipped with the scalar

product
T 
f, g L2 (H ) = f (t), g(t) H dt
0
is a Hilbert space. We refer the reader to [14] for details of Bochner integration in a
Hilbert space.
The Banach space of H -valued essentially bounded functions on [0, T ] is de-
noted by L∞ ([0, T ], H ) and consists of functions f for which the norm
  
f L∞ ([0,T ],H ) = ess sup f (t), t ∈ [0, T ]

is finite.
1.4 The Abstract Cauchy Problem 11

We now define Sobolev spaces. For nonnegative integers α1 , . . . , αd , consider the

multiindex α = (α1 , . . . , αd ) of order |α| = di=1 αi and for x = (x1 , x2 , . . . , xd ) ∈
Rd , denote
∂ α1 ∂ α2 ∂ αd
Dα = α1 α2 . . . .
∂x1 ∂x2 ∂xdαd
Let O be an open subset of Rd . For a function f ∈ C m (O) with positive integer m,
define the norm
  1/p

f m,p = D α f (x)p dx < ∞,
O |α|≤m

where the differentiation is in the sense of distributions. The Sobolev space

W m,p (O) with 1 ≤ p < ∞ is the completion in the norm  · m,p of the sub-
space of C m (O) consisting of all functions f such that f m,p < ∞. Similarly,
m,p
the space W0 (O) is the completion in the norm  · m,p of C0m (O). The Sobolev
spaces W m,p (O) defined above consist of functions f ∈ Lp (O) whose distribu-
tional derivatives D α f of order up to m are also in Lp (O).
m,p
The spaces (W m,p (O),  · m,p ) and (W0 (O),  · m,p ) are Banach spaces, and
W m,2 (O) and W0m,2 (O) are Hilbert spaces with the scalar product

f, g W m,2 (O) = D α f (x)D α g(x) dx.
O |α|≤m

Note that the Sobolev space W m,p (O) is a subset of Lp (O). If O = Rd , then it is
known ([30], Chap. 10, Proposition 1.5) that
   
W01,2 Rd = W 1,2 Rd . (1.30)

1.4 The Abstract Cauchy Problem

Let A be a linear operator on a real separable Hilbert space H , and let us consider
the abstract Cauchy problem given by
⎧
⎨ du(t)
= Au(t), 0 < t < T ,
dt (1.31)
⎩u(0) = x, x ∈ H.

Definition 1.4 A function u : [0, T [→ H is a (classical) solution of the prob-

lem (1.31) on [0, T [ if u is continuous on [0, T [, continuously differentiable and
u(t) ∈ D(A) for t ∈]0, T [, and (1.31) is satisfied on [0, T [.

Exercise 1.4 Argue why if x ∈ D(A), then (1.31) cannot have a solution.
12 1 Partial Differential Equations as Equations in Infinite Dimensions

If A is an infinitesimal generator of a C0 -semigroup {St , t ≥ 0}, then for any

x ∈ D(A), the function ux (t) = S(t)x, t ≥ 0, is a solution of (1.31) ([63], Chap. 4,
Theorem 1.3). On the other hand, if x ∈ D(A), then the continuity at zero may not
be a problem (see Exercise 1.4), but ux (t) does not have to be differentiable, unless
the C0 -semigroup has additional properties, for example, when it is a differentiable
semigroup (refer to (4) in Theorem 1.2).
In this case, ux (t) = S(t)x is not a solution in the usual sense, but it can be
viewed as a “generalized solution,” which will be called a “mild solution.” In fact,
the concept of mild solution can be introduced to study the following nonhomoge-
neous initial-value problem:
⎧
⎨ du(t)
= Au(t) + f (t), 0 < t < T ,
dt (1.32)
⎩u(0) = x, x ∈ H,

where f : [0, T [→ H .
We assume that A is an infinitesimal generator of a C0 -semigroup so that the
homogeneous equation (1.31) has a unique solution for all x ∈ D(A). The definition
of a classical solution, Definition 1.4, extends to the case of the nonhomogeneous
initial-value problem by requiring that in this case, the solution satisfies (1.32).
We now define the concept of a mild solution.

Definition 1.5 Let A be an infinitesimal generator of a C0 -semigroup S(t) on H ,

x ∈ H , and f ∈ L1 ([0, T ], H ) be the space of Bochner-integrable functions on
[0, T ] with values in H . The function u ∈ C([0, T ], H ) given by
t
u(t) = S(t)x + S(t − s)f (s) ds, 0 ≤ t ≤ T,
0

is the mild solution of the initial-value problem (1.32) on [0, T ].

Exercise 1.5 Prove that the function u(t) in Definition 1.5 is continuous.

Note that for x ∈ H and f ≡ 0, the mild solution is S(t)x, which is not in general
a classical solution.
When x ∈ D(A), the continuity of f is insufficient to assure the existence of a
classical solution. To see this, following [63], consider f (t) = S(t)x for x ∈ H such
that S(t)x ∈ D(A). Then (1.32) may not have a classical solution even if u(0) = 0 ∈
D(A), as the mild solution
t
u(t) = S(t − s)S(s)x ds = tS(t)x
0

is not, in general, differentiable.

One has the following theorem ([63], Chap. 4, Theorem 2.4).
1.4 The Abstract Cauchy Problem 13

Theorem 1.5 Let A be an infinitesimal generator of a C0 -semigroup {S(t), t > 0},

let f ∈ L1 ([0, T ], H ) be continuous on ]0, T ], and let
t
v(t) = S(t − s)f (s) ds, 0 ≤ t ≤ T.
0

The mild solution u to the initial–value problem (1.32) is a (classical) solution on

[0, T [ for every x ∈ D(A) if
(1) v(t) is continuously differentiable on ]0, T [.
(2) v(t) ∈ D(A) for 0 < t < T , and Av(t) is continuous on ]0, T [.
If (1.32) has a solution u on [0, T [ for some x ∈ D(A), then v satisfies (1) and (2).

Exercise 1.6 Show that if f is continuously differentiable on [0, T ], then

t t
v(t) = S(t − s)f (s) ds = S(s)f (t − s) ds
0 0

is continuously differentiable for t > 0, and its derivative is given by

t t
v  (t) = S(t)f (0) + S(t)f  (t − s) ds = S(t)f (0) + S(t − s)f  (s) ds.
0 0

Conclude that in this case the initial-value problem (1.32) has a solution for every
x ∈ D(A).

We conclude with examples of the heat equation in Rd and in a bounded domain

O ⊂ Rd .

Example 1.2 Consider the heat equation in Rd


ut (t, x) = u(t, x), 0 < t < T ,
(1.33)
u(0, x) = ϕ(x),

x ∈ Rd . The Gaussian family of operators

⎧
⎨ 1 
  exp −x − y2Rd /4t ϕ(y) dy, t > 0,
Gd (t)ϕ (x) = (4πt)d/2 R (1.34)
⎩
ϕ(x), t = 0,

defines a C0 -semigroup of contractions on H = L2 (Rd ) with the infinitesimal gen-

erator
d
∂2
ϕ= ϕ,
i=1
∂xi2
14 1 Partial Differential Equations as Equations in Infinite Dimensions

whose domain is D( ) = W 2,2 (Rd ). Note the difference between a bounded region
and Rd . Here the domain of the infinitesimal generator has this simple form due
to (1.30). Consider the related abstract Cauchy problem
⎧
⎨ du
= u, 0 < t < T ,
dt (1.35)
⎩u(0) = ϕ ∈ L2 (Rd ).

It is known that (Gd (t)ϕ)(x) is a classical solution of problem (1.35) for any ϕ ∈
H = L2 (Rd ) in the sense of Definition 1.4, since the semigroup Gd (t) on H is
differentiable ([63], Chap. 7, Theorem 2.7 and Remark 2.9).

Example 1.3 Let O ⊂ Rd be a bounded domain with smooth boundary ∂O (i.e.,

for each x ∈ ∂O and some ball B centered at x, ∂O ∩ B has the form xi =
f (x1 , . . . , xi−1 , xi+1 , . . . , xd ) for some 1 ≤ i ≤ d with f being k ≥ 1 times dif-
ferentiable). Then the heat equation has an abstract Cauchy form
⎧
⎨ du
= u, t > 0,
dt (1.36)
⎩u(0) = ϕ ∈ W 2,2 (O) ∩ W 1,2 (O).
0

The Laplace operator is an infinitesimal generator of a C0 -semigroup of con-

tractions on H = L2 (O), and the initial-value problem (1.36) has a unique solution
u(t, x) ∈ C([0, ∞[, W 2,2 (O) ∩ W01,2 (O)). This is a consequence of a more general
result for strongly elliptic operators ([63], Chap. 7, Theorem 2.5 and Corollary 2.6).

To study nonlinear equations, one also needs to look at the Peano theorem in an
infinite-dimensional Hilbert space, that is, to study the existence of a solution of the
equation
du  
(t) = G u(t) , u(0) = x ∈ H,
dt
where G is a continuous function on H . We note that due to the failure of the
Arzela–Ascoli theorem in C([0, T ], H ), the proof in the finite-dimensional case
fails (see the proofs in [15] and [29]). In fact the Peano theorem in a Hilbert space
is not true [26]. However, if we look at semilinear equations
⎧
⎨ du(t)
= Au(t) + G(u(t)), t > 0,
dt
⎩u(0) = x, x ∈ H,

we can salvage the theorem if {S(t), t > 0} is a semigroup of compact operators. We

present this theorem in the general case of SPDEs in Chap. 2.
1.5 The Variational Method 15

1.5 The Variational Method

We shall now consider problem (1.32) with f (t) ≡ 0 and with further condition
of coercivity on A. In this case, we follow the variational method due to Lions
[48]. However, we use the approach of Tanabe [71]. We present the basic technique
of Lions, without presenting the proofs. However, we present one crucial lemma
of Lions [48] on which Pardoux’s [62] approach is based. This lemma has been
appropriately generalized for the stochastic case by Krylov and Rozovskii [42]. We
present the method where the work of Krylov and Rozovskii is essential.
Let us start with the variational set up. We have three Hilbert spaces

V → H → V ∗

with V a dense subspace of H , V ∗ the continuous dual of V , and the embeddings

→ being continuous. Also, with ·, · denoting the duality on V × V ∗ , we assume
that v, v  = v, v  H if v ∈ V and v  ∈ H .
Let us consider now an analogue of problem (1.32). Let A be a linear operator,
A : V → V ∗ such that

AvV ∗ ≤ MvV
and
2v, Av ≤ λv2H − αv2V , v ∈ V,

for some real number λ and M, α > 0. The following theorem is due to Lions [48].

Theorem 1.6 (Lions) Let x ∈ H and f ∈ L2 ([0, T ]), V ∗ ). Then there exists a
unique function u ∈ L2 ([0, T ], V ) with du(t)/dt ∈ L2 ([0, T ], V ∗ ) and satisfying
⎧
⎨ du(t)
= Au(t) + f (t), t > 0,
dt (1.37)
⎩u(0) = x.

For the proof, see [48], p. 150.

The crucial lemma needed for proving that the solution u ∈ C([0, T ], H ) and to
work out an analogue of Itô’s formula for the function u(t)2H is the following,

Lemma 1.1 (Lions) If u ∈ L2 ([0, T ], V ) and du(t)/dt ∈ L2 ([0, T ], V ∗ ), then u ∈

C([0, T ], H ). Furthermore,
 
d 
u(t)2 = 2 u(t), du(t) .
H
(1.38)
dt dt

Proof Let 0 < a < T ; we extend u to (−a, T + a) by putting u(t) = u(−t) for a <
t < 0 and u(t) = u(2T − t) for T < t < T + a. Observe that u ∈ L2 ((−a, T + a), V )
and u ∈ L2 ((−a, T + a), V ∗ ). Define

w(t) = θ (t)u(t),
16 1 Partial Differential Equations as Equations in Infinite Dimensions

where θ (t) is continuously differentiable function, θ ≡ 1 in [0, T ] and θ ≡ 0 in the

neighborhood of −a and T + a. By using a mollifier jn , we define
T +a
wn (t) = jn (t − s)w(s) ds.
−a

Then, as n → ∞, we have that wn → w in L2 ((−a, T + a), V ) and dwn /dt →

dw/dt in L2 ((−a, T + a), V ∗ ). Note that dwn /dt ∈ V , so that we have
  d
t 
wn (t) − wm (t)2 = wn (s) − wm (s)2 ds
H H
−a ds
t  dwn dwm

= 2 wn (s) − wm (s), (s) − (s) ds
−a ds ds

T +a  dw

2 T +a  
≤  n (s) − dwm (s) ds + wn (s) − wm (s)2 ds.
 ds ds  V
−a V∗ −a

Thus, in addition to the L2 convergence in V , {wn } is a Cauchy sequence in

C([−a, T + a], H ). Hence, modifying the values on a null set, we get that w(t) ∈ H
and wn → w uniformly on [−a, T + a]. This gives u(t) ∈ C([0, T ], H ), by the
choice of θ . Let ϕ ∈ C0∞ ([0, T ]) and consider the integral
T  dϕ
wn (t), v(t) (t) dt
0 dt
T d 
 T dwn

= v(t), ϕ(t)wn (t) dt − v(t), ϕ(t) (t) dt
0 dt 0 dt
T   T  
dv dwn
=− ϕ(t)wn (t), (t) dt − v(t), ϕ(t) (t) dt.
0 dt 0 dt H

Taking the limit, we obtain

T  dϕ T  dv
 
du

u(t), v(t) H (t) dt = − u(t), (t) + v(t), (t) ϕ(t) dt.
0 dt 0 dt dt

For v(t) = u(t), this shows that the statement (1.38) is valid in terms of distributions.
Since the RHS in (1.38) is an integrable function of t, we conclude that u(t)2H is
absolutely continuous. 
Chapter 2
Stochastic Calculus

2.1 Hilbert-Space-Valued Process, Martingales, and Cylindrical

Wiener Process

2.1.1 Cylindrical and Hilbert-Space-Valued Gaussian Random

Variables

We first introduce cylindrical Gaussian random variables and Hilbert-space-valued

Gaussian random variables and then define cylindrical Wiener process and Hilbert-
space-valued Wiener process in a natural way. Let (, F , P ) be a probability space,
and K be a real separable Hilbert space with the norm and scalar product denoted
by  · K and ·, · K . We will always assume that (, F , P ) is complete, i.e., that
F contains all subsets A of  with P -outer measure zero,

P ∗ (A) = inf P (F ) : A ⊂ F ∈ F = 0.

Definition 2.1 We say that X̃ is a cylindrical standard Gaussian random variable

on K if X̃ : K → L2 (, F , P ) satisfies the following conditions:
(1) The mapping X̃ is linear.
(2) For an arbitrary k ∈ K, X̃(k) is a Gaussian random variable with mean zero and
variance k2K .
(3) If k, k  ∈ K are orthogonal, i.e., k, k  K = 0, then the random variables X̃(k)
and X̃(k  ) are independent.

Note that if {fj }∞ ∞

j =1 is an orthonormal basis (ONB) in K, then {X̃(fj )}j =1 is a
sequence of independent Gaussian random variables with mean zero and variance
one. By linearity of the mapping X̃ : K → L2 (, F , P ), we can represent X̃ as
∞

X̃(k) = k, fj K X̃(fj ),
j =1

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 17

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_2, © Springer-Verlag Berlin Heidelberg 2011
18 2 Stochastic Calculus

with the series convergent P -a.s. by Kolmogorov’s three-series theorem ([5], Theo-
rem 22.3).
However, it is not true that there exists a K-valued random variable X such that
 
X̃(k)(ω) = X(ω), k K .

This can be easily seen since we can express

∞

   2
X(ω)2 = X(ω), fj K ,
K
j =1

with the series being P -a.s. divergent by the strong law of large numbers.
In order to produce a K-valued Gaussian random variable, we proceed as follows.
Denote by L1 (K) the space of trace-class operators on K,
  
L1 (K) = L ∈ L (K) : τ (L) := tr (LL∗ )1/2 < ∞ , (2.1)

where the trace of the operator [L] = (LL∗ )1/2 is defined by

∞
    
tr [L] = [L]fj , fj K
j =1

for an ONB {fj }∞j =1 ⊂ K. It is well known [68] that tr([L]) is independent of the
choice of the ONB and that L1 (K) equipped with the trace norm τ is a Banach
space. Let Q : K → K be a symmetric nonnegative definite trace-class operator.
Assume that X : K → L2 (, F , P ) satisfies the following conditions:
(1) The mapping X is linear.
(2) For an arbitrary k ∈ K, X(k) is a Gaussian random variable with mean zero.
(3) For arbitrary k, k  ∈ K, E(X(k)X(k  )) = Qk, k  K .
Let {fj }∞
j =1 be an ONB in K diagonalizing Q, and let the eigenvalues corresponding
to the eigenvectors fj be denoted by λj , so that Qfj = λj fj . We define
∞

X(ω) = X(fj )(ω)fj .
j =1

Since ∞ j =1 λj < ∞, the series converges in L ((, F , P ), H ) and hence P -a.s.
2

In this case, P -a.s.,

 
X(ω), k K = X(k)(ω),
so that X : Ω → K is F /B(K)-measurable, where B(K) denotes the Borel σ -
field on K.

Definition 2.2 We call X : Ω → K defined above a K-valued Gaussian random

variable with covariance Q.
2.1 Hilbert-Space-Valued Process, Martingales, and Cylindrical Wiener Process 19

Definition 2.3 Let K be a separable Hilbert space. The measure P ◦ X −1 induced

by a K-valued Gaussian random variable X with covariance Q on the measurable
Hilbert space (K, B(K)) is called a Gaussian measure with covariance Q on K.

Exercise 2.1 Let K be a separable Hilbert space, Q : K → K be a symmetric non-

negative definite trace-class operator, and
X be a cylindrical Gaussian random vari-
able on K. Show that Y = X ◦ Q1/2 = ∞ j =1 X(Q1/2 (f ))f is a K-valued Gaus-
j j
sian random variable with covariance Q.

2.1.2 Cylindrical and Q-Wiener Processes

Let (, F , {Ft }t≥0 , P ) be a filtered probability space, and, as above, K be a real
separable Hilbert space. We will always assume that the filtration Ft satisfies the
usual conditions
(1) F0 contains
 all A ∈ F such that P (A) = 0,
(2) Ft = s>t Fs .

Definition 2.4 A K-valued stochastic process {Xt }t≥0 defined on a probability

space (, F , P ) is called Gaussian if for any positive integer n and t1 , . . . , tn ≥ 0,
(Xt1 , . . . , Xtn ) is a K n -valued Gaussian random variable.

A standard cylindrical Wiener process can now be introduced using the concept
of a cylindrical random variable.

Definition 2.5 We call a family {W̃t }t≥0 defined on a filtered probability space
(, F , {Ft }t≥0 , P ) a cylindrical Wiener process in a Hilbert space K if:
(1) For an arbitrary t ≥ 0, the mapping W̃t : K → L2 (, F , P ) is linear;
(2) For an arbitrary k ∈ K, W̃t (k) is an Ft -Brownian motion;
(3) For arbitrary k, k  ∈ K and t ≥ 0, E(W̃t (k)W̃t (k  )) = tk, k  K .
√
For every t > 0, W̃t / t is a standard cylindrical Gaussian random variable, so
that for any k ∈ K, W̃t (k) can be represented as a P -a.s. convergent series
∞

W̃t (k) = k, fj K W̃t (fj ), (2.2)
j =1

where {fj }∞
j =1 is an ONB in K.

Exercise 2.2 Show that E(W̃t (k)W̃s (k  )) = (t ∧ s)k, k  K and conclude that
W̃t (fj ), j = 1, 2, . . . , are independent Brownian motions.
20 2 Stochastic Calculus

For the same reason why a cylindrical Gaussian random variable cannot be real-
ized as a K-valued random variable, there is no K-valued process Wt such that
 
W̃t (k)(ω) = Wt (ω), k K .

However, if Q is a nonnegative definite symmetric trace-class operator on K, then

a K-valued Q-Wiener process can be defined.

Definition 2.6 Let Q be a nonnegative definite symmetric trace-class operator on a

separable Hilbert space K, {fj }∞
j =1 be an ONB in K diagonalizing Q, and let the
corresponding eigenvalues be {λj }∞j =1 . Let {wj (t)}t≥0 , j = 1, 2, . . . , be a sequence
of independent Brownian motions defined on (, F , {Ft }t , P ). The process
∞
 1/2
Wt = λj wj (t)fj (2.3)
j =1

is called a Q-Wiener process in K.

We can assume that the Brownian motions wj (t) are continuous. Then, the se-
ries (2.3) converges in L2 (Ω, C([0, T ], K)) for every interval [0, T ], see Exer-
cise 2.3. Therefore, the K-valued Q-Wiener process can be assumed to be con-
tinuous. We denote
∞
 1/2
Wt (k) = λj wj (t)fj , k K
j =1

for any k ∈ K, with the series converging in L2 (Ω, C([0, T ], R)) on every interval
[0, T ].

Exercise 2.3 Use Doob’s inequality, Theorem 2.2, for the submartingale
 n 
 
 1/2 
 λj wj (t)fj 
 
j =m K

to prove that the partial sums of the series (2.3) defining the Q-Wiener process are
a Cauchy sequence in L2 (Ω, C([0, T ], K)).

Remark 2.1 A stronger convergence result can be obtained for the series (2.3). Since
  n    n 2
  1  
 1/2    1/2 
P sup  λj wj (t)fj  > ε ≤ 2 E  λj wj (T )fj 
0≤t≤T   ε  
j =m K j =m K

T 
n
= λj → 0
ε2
j =m
2.1 Hilbert-Space-Valued Process, Martingales, and Cylindrical Wiener Process 21

with m ≤ n, m, n → ∞, the series (2.3) converges uniformly on [0, T ] in proba-

bility P , and hence, by the Lévy–Itô–Nisio theorem ([45], Theorem 2.4), it also
converges P -a.s. uniformly on [0, T ].

Basic properties of a Q-Wiener process are summarized in the next theorem.

Theorem 2.1 A K-valued Q-Wiener process {Wt }t≥0 has the following properties:
(1) W0 = 0.
(2) Wt has continuous trajectories in K.
(3) Wt has independent increments.
(4) Wt is a Gaussian process with the covariance operator Q, i.e., for any k, k  ∈ K
and s, t ≥ 0,
 
E Wt (k)Ws (k  ) = (t ∧ s)Qk, k  K .
(5) For an arbitrary k ∈ K, the law L ((Wt − Ws )(k)) ∼ N (0, (t − s)Qk, k K ).
∞
Exercise 2.4 Consider a cylindrical Wiener process W̃t (k) = j =1 k, fj K W̃t (fj )

and a Q-Wiener process Wt = ∞
1/2
j =1 λj wj (t)fj , as defined in (2.2) and (2.3),
respectively. Show that

(a) Wt1 = W̃t ◦ Q1/2 = ∞
1/2
j =1 λj W̃t (fj )fj defines a Q-Wiener process;

(b) W̃t1 (k) = ∞ j =1 k, fj K wj (t) defines a cylindrical Wiener process.

2.1.3 Martingales in a Hilbert Space

Definition 2.7 Let H be a separable Hilbert space considered as a measurable space

with its Borel σ -field B(H ). We fix T > 0 and let (, F , {Ft }t≤T , P ) be a fil-
tered probability space and {Mt }t≤T be an H -valued process adapted to the filtra-
tion {Ft }t≤T . Assume that Mt is integrable, EMt H < ∞. Then Mt is called a
martingale if for any 0 ≤ s ≤ t,

E(Mt |Fs ) = Ms , P -a.s.

We note that because H is separable, the measurability of Mt with respect to

the σ -fields Ft and B(H ) is equivalent to the measurability of Mt , h H with re-
spect to Ft and B(R) for all h ∈ H , which implies the measurability of Mt H .
p
If EMt H < ∞, we will also write Mt ∈ Lp (, Ft , P ). The condition for Mt ∈
L1 (, Ft , P ) to be a martingale is equivalent to
 
E Mt , h H |Fs = Ms , h H , ∀h ∈ H, P -a.s.

if 0 ≤ s ≤ t.
If Mt ∈ Lp (, F , P ) is an H -valued martingale than for p ≥ 1, the process
p
Mt H is a real-valued submartingale. We have therefore the following theorem.
22 2 Stochastic Calculus

Theorem 2.2 (Doob’s Maximal Inequalities) If Mt ∈ Lp (, F , P ) is an H -valued

martingale, then
p
(1) P (sup0≤t≤T Mt H > λ) ≤ λ1p EMT H , p ≥ 1, λ > 0;
p  p p p
(2) E(sup0≤t≤T Mt H ) ≤ p−1 EMT H , p > 1.

We will now introduce the Hilbert space of square-integrable H -valued martin-

gales. Note that by Doob’s inequality, Theorem 2.2, we have
 
E sup Mt 2H ≤ 4EMT 2H .
t≤T

Definition 2.8 A martingale {Mt }0≤t≤T is called square integrable if

EMT 2H < ∞.

The class of continuous square-integrable martingales will be denoted by MT2 (H ).

Since Mt ∈ MT2 (H ) is determined by the relation Mt = E(MT |Ft ), the space

MT2 (H ) is a Hilbert space with scalar product
 
M, N M 2 (H ) = E MT , NT H .
T

In the case of real-valued martingales Mt , Nt ∈ MT2 (R), there exist unique

quadratic variation and cross quadratic variation processes, denoted by M t and
M, N t , respectively, such that Mt2 − M t and Mt Nt − M, N t are continuous
martingales. For Hilbert-space-valued martingales, we have the following definition.

Definition 2.9 Let Mt ∈ MT2 (H ). We denote by M t the unique adapted contin-

uous increasing process starting from 0 such that Mt 2H − M t is a continuous
martingale. We define a quadratic variation process M t of Mt as an adapted
continuous process starting from 0, with values in the space of nonnegative definite
trace-class operators on H , such that for all h, g ∈ H ,
 
Mt , h H Mt , g H − M t (h), g H

is a martingale.

Lemma 2.1 The quadratic variation process of a martingale Mt ∈ MT2 (H ) exists

and is unique. Moreover,
M t = trM t. (2.4)

Proof The lemma follows by applying the classical one-dimensional results. We

can assume without loss of generality that M0 = 0. Denote

Mti = Mt , ei H,
2.2 Stochastic Integral with Respect to a Wiener Process 23

where {ei }∞
i=1 is an ONB in H . Note that the quadratic variation process has to
satisfy
 
M t (ei ), ej H = Mi , Mj t .
Consequently, we define the quadratic variation process by
∞

 
M t (h), g H = Mi , Mj t ei , h H ej , g H . (2.5)
i,j =1

The sum in (2.5) converges P -a.s. and defines a nonnegative definite trace-class
operator on H , since
∞
 ∞

E trM t =E Mi t = E Mt , ei 2
H = EMt 2H < ∞.
i=1 i=1

Now equality (2.4) follows from (2.5). 

Exercise 2.5 Show that an H -valued Q-Wiener process {Wt }t≤T is a continuous
square-integrable martingale with W t = t (trQ) and W t = tQ.

Exercise 2.6 Let 0 < t1 , . . . , tn < t ≤ T be a partition of the interval [0, t], t ≤ T ,
and max{tj +1 − tj , 1 ≤ j ≤ n − 1} → 0. Denote ΔWj = Wtj +1 − Wtj . Show that


n
ΔWj 2K → t (tr Q), P -a.s.
j =1

2.2 Stochastic Integral with Respect to a Wiener Process

We will introduce the concept of Itô’s stochastic integral with respect to a Q-Wiener
process and with respect to a cylindrical Wiener process simultaneously.
Let K and H be separable Hilbert spaces, and Q be either a symmetric non-
negative definite trace-class operator on K or Q = IK , the identity operator on K.
In case Q is trace-class, we will always assume that its all eigenvalues λj > 0,
j = 1, 2, . . .; otherwise we can start with the Hilbert space ker(Q)⊥ instead of K.
The associated eigenvectors forming an ONB in K will be denoted by fk .
Then the space KQ = Q1/2 K equipped with the scalar product
∞
 1
u, v KQ = u, fj K v, fj K
λj
j =1

is a separable Hilbert space with an ONB {λj fj }∞

1/2
j =1 .
24 2 Stochastic Calculus

If H1 , H2 are two real separable Hilbert spaces with {ei }∞

i=1 an ONB in H1 , then
the space of Hilbert–Schmidt operators from H1 to H2 is defined as
 ∞


L2 (H1 , H2 ) = L ∈ L (H1 , H2 ) : Lei 2H2 < ∞ . (2.6)
i=1

It is well known (see [68]) that L2 (H1 , H2 ) equipped with the norm
 ∞
1/2

LL2 (H1 ,H 2) = Lei 2H2
i=1

is a Hilbert space. Since the Hilbert spaces H1 and H2 are separable, the space
L2 (H1 , H2 ) is also separable, as Hilbert–Schmidt operators are limits of sequences
of finite-dimensional linear operators.
Consider L2 (KQ , H ), the space of Hilbert–Schmidt operators from KQ to H .
If {ej }∞
j =1 is an ONB in H , then the Hilbert–Schmidt norm of an operator L ∈
L2 (KQ , H ) is given by
∞
 ∞

  1/2  2  2
L2L2 (KQ ,H ) = L λj f j , ei H = LQ1/2 fj , ei H
j,i=1 j,i=1
 2   ∗ 
= LQ1/2  L2 (K,H )
= tr LQ1/2 LQ1/2 . (2.7)

The scalar product between two operators L, M ∈ L2 (KQ , H ) is defined by

  ∗ 
L, M L2 (KQ ,H ) = tr LQ1/2 MQ1/2 . (2.8)

Since the Hilbert spaces KQ and H are separable, the space L2 (KQ , H ) is also
separable.
Let L ∈ L (K, H ). If k ∈ KQ , then
∞
  1/2  1/2
k= k, λj fj λ fj ,
KQ j
j =1

and L, considered as an operator from KQ to H , defined as

∞
  1/2  1/2
Lk = k, λj fj λ Lfj ,
KQ j
j =1

has a finite Hilbert–Schmidt norm, since

∞
 ∞

  1/2 2
L2L2 (KQ ,H ) = L λ fj  = λj Lfj 2H ≤ L2L (K,H ) tr(Q).
j H
j =1 j =1
2.2 Stochastic Integral with Respect to a Wiener Process 25

Thus, L (K, H ) ⊂ L2 (KQ , H ). If L, M ∈ L (K, H ), formulas (2.7) and (2.8) re-

duce to
 
L2L2 (KQ ,H ) = tr LQL∗ (2.9)

and
 
L, M L2 (KQ ,H ) = tr LQM ∗ , (2.10)

allowing for separation of Q1/2 and L∗ . This is usually exploited in calculations

where L ∈ L2 (KQ , H ) is approximated with a sequence Ln ∈ L (K, H ).
The space L2 (KQ , H ) consists of linear operators L : K → H , not necessar-
ily bounded, with domain D(L) ⊃ Q1/2 K, and such that tr((LQ1/2 )(LQ1/2 )∗ ) is
finite. If Q = IK , then KQ = K. We note that the space L2 (KQ , H ) contains gen-
uinely unbounded linear operators from K to H .

Exercise 2.7 Give an example of an unbounded linear operator from K to H , which

is an element of L2 (KQ , H ).

2.2.1 Elementary Processes

Let E (L (K, H )) denote the class of L (K, H )-valued elementary processes

adapted to the filtration {Ft }t≤T that are of the form


n−1
Φ(t, ω) = φ(ω)1{0} (t) + φj (ω)1(tj ,tj +1 ] (t), (2.11)
j =0

where 0 = t0 ≤ t1 ≤ · · · ≤ tn = T , and φ, φj , j = 0, 1, . . . , n − 1, are respec-

tively F0 -measurable and Ftj -measurable L2 (KQ , H )-valued random variables
such that φ(ω), φj (ω) ∈ L (K, H ), j = 0, 1, . . . , n − 1 (recall that L (K, H ) ⊂
L2 (KQ , H )).
Note that if Q = IK , then the random variables φj are, in fact, L2 (K, H )-valued.
We shall say that an elementary process Φ ∈ E (L (K, H )) is bounded if it is
bounded in L2 (KQ , H ).
We have defined elementary processes to be left-continuous as opposed to being
right-continuous. There is no difference if the Itô stochastic integral is constructed
with respect to a Wiener process. Our choice, however, is consistent with the con-
struction of a stochastic integral with respect to square-integrable martingales.
26 2 Stochastic Calculus

2.2.2 Stochastic Itô Integral for Elementary Processes

For an elementary process Φ ∈ E (L (K, H )), we define the Itô stochastic integral
with respect to a Q-Wiener process Wt by
t 
n−1
Φ(s) dWs = φj (Wtj +1 ∧t − Wtj ∧t )
0 j =0

for t ∈ [0, T ]. The term φW0 is neglected since P (W0 = 0) = 1. This stochastic
integral is an H -valued stochastic process.
We define the Itô cylindrical stochastic integral of an elementary process Φ ∈
E (L (K, H )) with respect to a cylindrical Wiener process W̃ by
  
n−1
t     
Φ(s) d W̃s (h) = W̃tj +1 ∧t φj∗ (h) − W̃tj ∧t φj∗ (h) (2.12)
0 j =0

for t ∈ [0, T ] and h ∈ H . The following proposition states Itô’s isometry, which is
essential in furthering the construction of the stochastic integral.

Proposition 2.1 For a bounded elementary process Φ ∈ E (L (K, H )),

 t 2
  t 
E  Φ(s) dWs 

 = E Φ(s)2
L (K ,H )
ds < ∞ (2.13)
2 Q
0 H 0

for t ∈ [0, T ].

Proof The proof resembles calculations in the real case. Without loss of generality
we can assume that t = T , then
 T 2 n−1 2
   
   
E Φ(s) dWs  = E  φj (Wtj +1 − Wtj )
0 H  
j =1 H
n−1 
 2 
n−1
 
=E φj (Wt 
j +1 − Wtj ) H + φj (Wtj +1 − Wtj ), φi (Wti+1 − Wti ) H .
j =1 i =j =1

Consider first the single term Eφj (Wtj +1 −Wtj )2H . We use the fact that the random
variable φj and consequently for a vector em ∈ H , the random variable φj∗ em is Ftj -
measurable, while the increment Wtj +1 − Wtj is independent of this σ -field. With
{fl }∞ ∞
l=1 and {em }m=1 , orthonormal bases in Hilbert spaces K and H , we have

∞

 2  2
E φj (Wtj +1 − Wtj )H = E φj (Wtj +1 − Wtj ), em H
m=1
∞
   2  
= E E φj (Wtj +1 − Wtj ), em H Ftj
m=1
2.2 Stochastic Integral with Respect to a Wiener Process 27

∞
   2  
= E E Wtj +1 − Wtj , φj∗ em K Ftj
m=1

∞
  ∞
2  
    
= E E Wtj +1 − Wtj , fl ∗
K φj em , fl K
F t
 j
m=1 l=1
∞
  ∞
 
   2 
= E E Wtj +1 − Wtj , fl 2
φj∗ em , fl K F t
K  j
m=1 l=1
∞
  ∞
   
+ E E Wtj +1 − Wtj , fl K φj∗ em , fl K
m=1 l =l  =1
 
  
× Wtj +1 − Wtj , fl K φj∗ em , fl  K Ftj

∞
 ∞
  2
= (tj +1 − tj ) λl φj∗ em , fl K
m=1 l=1
∞
   1/2  2
= (tj +1 − tj ) φj λl fl , em H = (tj +1 − tj )φj 2L2 (KQ ,H ) .
m,l=1

Similarly, for the single term Eφj (Wtj +1 − Wtj ), φi (Wti+1 − Wti ) H , we obtain that
 
E φj (Wtj +1 − Wtj ), φi (Wti+1 − Wti ) H
∞
 ∞
   
=E E Wtj +1 − Wtj , fl K φj∗ em , fl K
m=1 l,l  =1

    
× Wti+1 − Wti , fl  K
φi∗ em , fl  K Ftj =0

if i < j . Now we can easily reach the conclusion. 

We have the following counterpart of (2.13) for the Itô cylindrical stochastic
integral of a bounded elementary process Φ ∈ E (L (K, H )):
  2
t t  2
E Φ(s) d W̃s (h) = E Φ ∗ (s)(h)K ds < ∞. (2.14)
0 0

Exercise 2.8 Prove (2.14).

The idea now is to extend the definition of the Itô stochastic integral and cylindri-
cal stochastic integral to a larger class of stochastic processes utilizing the fact that
28 2 Stochastic Calculus
T T
the mappings Φ → 0 Φ(s) dWs and Φ ∗ (·)(h) → ( 0 Φ(s) d W̃s )(h) are isome-
tries due to. (2.13) and (2.14).
When the stochastic integral with respect to martingales is constructed, then the
natural choice of the class of integrands is the set of predictable processes (see, for
example, [57]), and sometimes this restriction is applied when the martingale is a
Wiener process. We will however carry out a construction which will allow the class
of integrands to be simply adapted and not necessarily predictable processes.
Let 2 (KQ , H ) be a class of L2 (KQ , H )-valued processes measurable as
mappings from ([0, T ] × , B([0, T ]) ⊗ F ) to (L2 (KQ , H ), B(L2 (KQ , H ))),
adapted to the filtration {Ft }t≤T (thus F can be replaced with FT ), and satisfying
the condition
T 
E Φ(t)2 dt < ∞. (2.15)
L2 (KQ ,H )
0

Obviously, elementary processes satisfying the above condition (2.15) are elements
of 2 .
We note that 2 (KQ , H ) equipped with the norm

 T 1/2

Φ2 (KQ ,H ) = E Φ(t)2 dt (2.16)
L2 (KQ ,H )
0

is a Hilbert space. The following proposition shows that the class of bounded ele-
mentary processes is dense in 2 (KQ , H ). It is valid for Q, a trace-class operator,
and for Q = IK .

Proposition 2.2 If Φ ∈ 2 (KQ , H ), then there exists a sequence of bounded ele-

mentary processes Φn ∈ E (L (K, H )) approximating Φ in 2 (KQ , H ), i.e.,

T 
Φn − Φ22 (KQ ,H ) = E Φn (t) − Φ(t)2 dt → 0
L2 (KQ ,H )
0

as n → ∞.

Proof We follow the idea in [61].

(a) We can assume that Φ(t, ω)L2 (KQ ,H ) < M for all t, ω. Otherwise, we de-
fine
⎧  
⎨n Φ(t, ω)
if Φ(t, ω)L (K ,H ) > n,
Φn (t, ω) = Φ(t, ω)L2 (KQ ,H ) 2 Q
⎩
Φ(t, ω) otherwise.

Then Φn − Φ2 (KQ ,H ) → 0 by the Lebesgue dominated convergence theorem

(DCT).
2.2 Stochastic Integral with Respect to a Wiener Process 29

(b) We can assume that Φ(t, ω) ∈ L (K, H ) since every operator L ∈ L2 (KQ , H )
can be approximated in L2 (KQ , H ) by the operators Ln ∈ L (K, H ) defined by


n
 1/2  1/2 
Ln k = L λj f j λj f j , k K .
Q
j =1

Indeed, for k ∈ K, we have

 n
∞ 
2
  1/2  1/2 
Ln k2H = L λj f j λj f j , k K , ei
Q
i=1 j =1 H
∞
 
n
  1/2  2 1
≤ L λj f j , ei H fj , k 2
K
λj
i,j =1 j =1

≤ Cn L2L2 (KQ ,H ) k2K ,

so that Ln ∈ L (K, H ), and

∞

    
Φ(t) − Φn (t)2 = Φ(t) λ1/2 fj 2 → 0.
L2 (KQ ,H ) j H
j =n+1

Then
T 
E Φ(t) − Φn (t)2 dt → 0
L2 (KQ ,H )
0
as n → ∞ by the Lebesgue DCT, so that Φn → Φ in 2 (KQ , H ).
(c) Now assume, in addition to (a) and (b), that for each ω, the function Φ(·, ω) :
[0, T ] → L2 (KQ , H ) is continuous. For a partition 0 = t0 ≤ t1 ≤ · · · ≤ tn = T ,
define the elementary process


n−1
Φn (t, ω) = Φ(0, ω)1{0} (t) + Φ(tj , ω)1(tj ,tj +1 ] (t).
j =1

With the size of the partition max{|tj +1 − tj | : j = 0, . . . , n} → 0 as n → ∞, we

have that Φn (t, ω) → Φ(t, ω) and
T 
Φ(t, ω) − Φn (t, ω)2 dt → 0
L2 (KQ ,H )
0

due to the continuity of Φ(·, ω). Consequently, due to the Lebesgue DCT,
T 
E Φ(t, ω) − Φn (t, ω)2 dt → 0.
L2 (KQ ,H )
0
30 2 Stochastic Calculus

(d) If Φ(t, ω) ∈ L (K, H ) is bounded for every (t, ω) but not necessarily contin-
uous, then we first extend Φ to the entire R by assigning Φ(t, ω) = 0 for t < 0 and
t > T . Next, we define bounded continuous approximations of Φ by
t
Φn (t, ω) = ψn (s − t)Φ(s, ω) ds, 0 ≤ t ≤ T.
0

Here ψn (t) = nψ(nt), and ψ(t) is any nonnegative

 bounded continuous function on
R with support in [−1, 0] and such that R ψ(t) dt = 1. We will use the technique
of an approximate identity (refer to [73], Chap. 9).
The functions Φn (t, ω) are clearly bounded, and we now justify their continuity.
With t + h ≤ T , we have
 
Φn (t + h, ω) − Φn (t, ω)
L2 (KQ ,H )
 t+h 
   t 
= ψ n s − (t + h) Φ(s, ω) ds − ψ n (s − t)Φ(s, ω) ds 

0 0 L2 (KQ ,H )
 
 t    
≤
 ψn s − (t + h) − ψn (s − t) Φ(s, ω) ds 

0 L2 (KQ ,H )
 
 T   
+
 ψn s − (t + h) Φ(s, ω)1[t,t+h] (s) ds 
 .
0 L2 (KQ ,H )

The second integral converges to zero with h → 0 by the Lebesgue DCT. The first
integral is dominated by
    
ψn s − (t + h) − ψn (s − t)Φ(s, ω) 1[0,t] (s) ds
L 2 (KQ ,H )
R
  
= ψn (u + h) − ψn (u)Φ(u + t, ω) 1 (u) du
L2 (KQ ,H ) [−t,0]
R
 1/2
 
= ψn (u + h) − ψn (u)2 du
R
 1/2
 
× Φ(u + t, ω))2 1 (u) du ,
L2 (KQ ,H ) [−t,0]
R

so that it converges to zero due to the continuity of the shift operator in L2 (R) (see
Theorem 8.19 in [73]). Left continuity in T follows by a similar argument.
Since the process Φ(t, ω) is adapted to the filtration Ft , we deduce from the
definition of Φn (t, ω) that it is also Ft -adapted.
We will now show that
T 
Φn (t, ω) − Φ(t, ω)2 dt → 0
L2 (KQ ,H )
0
2.2 Stochastic Integral with Respect to a Wiener Process 31

as n → ∞. Consider
 
   t   
Φn (t, ω) − Φ(t, ω) ≤ ψn (s − t) Φ(s, ω) − Φ(t, ω) ds 
L2 (KQ ,H )  
0 L2 (KQ ,H )
 
 t  
+
 ψn (s − t) − 1 Φ(t, ω) ds 
 .
0 L2 (KQ ,H )
t
For a fixed ω, denote wn (t) =  0 (ψn (s − t) − 1)Φ(t, ω) dsL2 (KQ ,H ) . Then wn (t)
converges to zero for every t as n → ∞ and is bounded by CΦ(t, ω)L2 (KQ ,H )
for some constant C. From now on, the constant C can change its value from line to
line. The first integral is dominated by
 
ψn (s − t)Φ(s, ω) − Φ(t, ω)L 1[0,t] (s) ds
2 (KQ ,H )
R
 
= ψn (u)Φ(u + t, ω) − Φ(t, ω)L 1[−t,0] (u) du
2 (KQ ,H )
R
 
≤ Φ(u + t, ω) − Φ(t, ω) 1/2 1/2
ψ (u)ψn (u)1[−t,0] (u) du
L2 (KQ ,H ) n
R
 1/2
 
≤ Φ(u + t, ω) − Φ(t, ω)2 ψ (u)1[−t,0] (u) du
L2 (KQ ,H ) n
R

by the Schwarz–Bunyakovsky inequality and the property that ψn (t) integrates to

one.
We arrive at
T 
Φn (t, ω) − Φ(t, ω)2 dt
L2 (KQ ,H )
0
T  
≤2 Φ(u + t, ω) − Φ(t, ω)2 ψ (u)1[−t,0] (u) du dt
L2 (KQ ,H ) n
0 R
T
+2 wn2 (t, ω) dt.
0

The second integral converges to zero as n → ∞ by the Lebesgue DCT and is

T
dominated by C 0 Φ(t, ω)2L2 (KQ ,H ) dt. Now we change the order of integration
in the first integral:
T  
Φ(u + t, ω) − Φ(t, ω)2 ψ (u)1[−t,0] (u) du dt
L2 (KQ ,H ) n
0 R
T  2
= ψn (u) 1[−t,0] (u)Φ(u + t, ω) − Φ(t, ω)L dt du
2 (KQ ,H )
R 0
    2
T v 
 v 
= ψ(v) 1[−t,0]  Φ + t, ω − Φ(t, ω)
 dt dv.
R 0 n n L2 (KQ ,H )
32 2 Stochastic Calculus

We note that, again by the continuity of the shift operator in L2 ([0, T ], L2 (KQ , H )),
    2
T v 
 v 
1[−t,0] Φ + t, ω − Φ(t, ω) dt
0 n  n 
L2 (KQ ,H )

T
   2

≤ Φ v + t, ω − Φ(t, ω) dt → 0
 n 
0 L2 (KQ ,H )

as n → ∞, so that the function

    2
T v 
 v 
ψ(v) 1[−t,0]  Φ + t, ω − Φ(t, ω)
 dt
0 n n L2 (KQ ,H )

converges to zero and is bounded by CΦ(·, ω)2L2 ([0,T ],L . This proves that
2 (KQ ,H ))

T 
rn (ω) = Φn (t, ω) − Φ(t, ω)2 dt → 0
L2 (KQ ,H )
0

as n → ∞ and rn (ω) ≤ CΦ(t, ω)2L2 ([0,T ],L . Therefore,

2 )(KQ ,H )

T 
E Φn (t, ω) − Φ(t, ω)2 dt = Ern (ω) → 0
L2 (KQ ,H )
0

as n → ∞ by the Lebesgue DCT. 

We will need the following lemma when using the Yosida approximation of an
unbounded operator.

Lemma 2.2 (a) Let T , Tn ∈ L (H ) be such that for every h ∈ H , Tn h → T h, and

let L ∈ L2 (KQ , H ). Then

Tn L − T LL2 (KQ ,H ) → 0. (2.17)

(b) Let A be the generator of a C0 -semigroup S(t) on a real separable Hilbert

space H , and An = ARn be the Yosida approximations of A as defined in (1.22).
T 2p
Then, for Φ(t) ∈ 2 (KQ , H ) such that E 0 Φ(t)L2 (KQ ,H ) dt < ∞, p ≥ 1,

t   
lim sup E  e(t−s)An − S(t − s) Φ(s)2p ds → 0. (2.18)
n→∞ t≤T L (K 2 Q ,H )
0

(c) Under the assumptions in part (b),

t   2p
lim E sup  esAn − S(s) Φ(s)L (K ds → 0.
n→∞ 2 Q ,H )
0 0≤s≤T
2.2 Stochastic Integral with Respect to a Wiener Process 33

Proof (a) Note that by the Banach–Steinhaus theorem,

!
max T L (H ) , sup Tn L (H ) < C
n

for some constant C. Hence, for an ONB {fj }∞

j =1 ⊂ K,we have

∞
  
Tn L − T L2L2 (KQ ,H ) = (Tn − T )LQ1/2 fj 2
H
j =1

∞
  2
≤ Tn − T 2L (H ) LQ1/2 fj H
j =1

∞
  2
≤ 4C 2 LQ1/2 fj H = 4C 2 L2L2 (KQ ,H ) .
j =1

The terms (Tn − T )LQ1/2 fj 2H converge to zero as n → ∞ and are bounded
by 4C 2 LQ1/2 fj 2H , so that (a) follows by the Lebesgue DCT with respect to the
counting measure δj .
(b) We will use two facts about the semigroup Sn (t) = etAn . By Proposition 1.2,
we have
 
supSn (t)L (H ) ≤ Me2αt , n > 2α,
n

and limn→∞ Sn (t)x = S(t)x, x ∈ H , uniformly on finite intervals.

Now, for n > 2α,
t   
sup E  Sn (t − s) − S(t − s) Φ(s)2p ds
L (K 2 Q ,H )
0≤t≤T 0
 ∞
p
t    
= sup E  Sn (t − s) − S(t − s) Φ(s)Q1/2 fj 2 ds
H
0≤t≤T 0 j =1
∞ p
T   
= sup E  Sn (t − s) − S(t − s) Φ(s)Q1/2 fj 2 1[0,t] (s) ds
H
0≤t≤T 0 j =1
" ∞ 
#
T       p
sup  Sn (t − s) − S(t − s) Φ(s)Q1/2 fj H 1[0,t] (s)
2
≤E ds
0 j =1 0≤t≤T
" ∞ 
#p
T     
≤E sup sup  1/2 2
Sn (t − s) − S(t − s) Φ(s)Q fj H 1[0,t] (s) ds
0 j =1 n>2α 0≤t≤T
34 2 Stochastic Calculus
" ∞
#p
T   2 4αT   
≤E 4M e  1/2 2
Φ(s)Q fj H ds
0 j =1

 p T 
= 4M 2 e4αT E Φ(s)2p ds < ∞.
L2 (KQ ,H )
0

The term
  2 
sup  Sn (t − s) − S(t − s) Φ(s)Q1/2 fj H 1[0,t] (s) → 0
0≤t≤T

as n → ∞ and is bounded by 4M 2 e4αT Φ(s)Q1/2 fj 2H ; hence, (b) follows by the

Lebesgue DCT relative to the counting measure δj and then relative to dP ⊗ dt.
(c) The proof of part (c) follows the arguments in the proof of part (b). 

2.2.3 Stochastic Itô Integral with Respect to a Q-Wiener Process

We are ready to extend the definition of the Itô stochastic integral with respect to a
Q-Wiener process to adapted stochastic processes Φ(s) satisfying the condition
T 
E Φ(s)2 ds < ∞,
L2 (KQ ,H )
0

which will be further relaxed to the condition

 T 
 
P Φ(s)2 ds < ∞ = 1.
L (K ,H ) 2 Q
0

Definition 2.10 The stochastic integral of a process Φ ∈ 2 (KQ , H ) with respect

to a K-valued Q-Wiener process Wt is the unique isometric linear extension of the
mapping
T
Φ(·) → Φ(s) dWs
0

from the class of bounded elementary processes to L2 (, H ), to a mapping

from 2 (KQ , H ) to L2 (, H ), such that the image of Φ(t) = φ1{0} (t) +
n−1 n−1
j =0 φj 1(tj ,tj +1 ] (t) is j =0 φj (Wtj +1 − Wtj ). We define the stochastic integral
t
process 0 Φ(s) dWs , 0 ≤ t ≤ T , for Φ ∈ 2 (KQ , H ) by
t T
Φ(s) dWs = Φ(s)1[0,t] (s) dWs .
0 0
·
Theorem 2.3 The stochastic integral Φ → 0 Φ(s) dWs with respect to a K-valued
Q-Wiener process Wt is an isometry between 2 (KQ , H ) and the space of contin-
2.2 Stochastic Integral with Respect to a Wiener Process 35

uous square-integrable martingales MT2 (H ),

 2
 t  t 
E
 Φ(s) dWs 
 =E
Φ(s)2
L2 (KQ ,H )
ds < ∞ (2.19)
0 H 0

for t ∈ [0, T ]. t
The quadratic variation process of the stochastic integral process 0 Φ(s) dWs
t
and the increasing process related to  0 Φ(s) dWs 2H are given by
 ·  t  ∗
Φ(s) dWs = Φ(s)Q1/2 Φ(s)Q1/2 ds
0 t 0
and
 ·  t   ∗ 
Φ(s) dWs = tr Φ(s)Q1/2 Φ(s)Q1/2 ds
0 t 0
t 
= Φ(s)2 ds.
L2 (KQ ,H )
0

Proof We note that the stochastic integral process for a bounded elementary pro-
cess in E (L (K, H )) is a continuous square-integrable martingale. Let the se-
quence of bounded elementary processes {Φn }∞ n=1 ⊂ E (L (K, H )) approximate
Φ ∈ 2 (KQ , H ). We can assume that Φ1 = 0 and
1
Φn+1 − Φn 2 (KQ ,H ) < . (2.20)
2n
Then by Doob’s inequality, Theorem 2.2, we have

∞   t  
 t  1

P sup Φn+1 (s) dWs − 
Φn (s) dWs  > 2
t≤T 0 0 H n
n=1
∞  
 T  2
≤ n4 E 
 Φ n+1 (s) − Φ n (s) dW s


n=1 0 H

∞ T ∞ 4

 n
= n4 E Φn+1 (s) − Φn (s)2 ds ≤ .
L2 (KQ ,H ) 2n
n=1 0 n=1

By the Borel–Cantelli lemma,

 t 
 t  1
sup Φn+1 (s) dWs − Φn (s) dWs 
 ≤ n2 , n > N (ω),
t≤T 0 0 H

for some N(ω), P -a.s. Consequently, the series

∞ 
 t t 
Φn+1 (s) dWs − Φn (s) dWs
n=1 0 0
36 2 Stochastic Calculus
t
converges to 0 Φ(s) dWs in L2 (, H ) for every t ≤ T and converges P -a.s. as a
t
series of H -valued continuous
· functions to a continuous version of 0 Φ(s) dWs .
Thus, the mapping Φ → 0 Φ(s) dWs is an isometry from the subset of bounded
elements in E (L (K, H )) into the space of continuous square-integrable martin-
gales MT2 (H ), and it extends to 2 (KQ , H ) with images in MT2 (H ) by the com-
pleteness argument. We only need to verify the formula for the quadratic variation
process. Note (see Exercise 2.9) that using representation (2.3), we have for h ∈ H ,

  ∞
 t
t
1/2 
Φ(s) dWs , h = λj Φ(s)fj , h H dwj (t),
0 H j =1 0

with the series convergent in L2 (, R). If h, g ∈ H , then

∞

  ∗     
Φ(s)Q1/2 Φ(s)Q1/2 h, g H = h, Φ(s)Q1/2 fj H g, Φ(s)Q1/2 fj H
j =1
∞
    
= λj h, Φ(s)fj H g, Φ(s)fj H .
j =1

Now, for 0 ≤ u ≤ t,
 t   t 
E Φ(s) dWs , h Φ(s) dWs , g
0 H 0 H
    
t  ∗ 
− Φ(s)Q1/2 Φ(s)Q1/2 ds (h), g F u

0 H
∞ ∞
 t
1/2    t
1/2  
=E λi Φ(s)fi , h H dwi (s) λj Φ(s)fj , g H
dwj (s)
i=1 0 j =1 0
∞
  
t    

− λj h, Φ(s)fj H
g, Φ(s)fj H ds Fu
j =1 0
∞ 
 t  t 
=E λj Φ(s)fj , h H dwj (s) Φ(s)fj , g H
dw j (s)
j =1 0 0

∞
  
t    

− λj h, Φ(s)fj H
g, Φ(s)fj H ds Fu
j =1 0
 ∞ 
 1/2 1/2
t 
+E λi λj Φ(s)fi , h H dwi (s)
i =j =1 0
2.2 Stochastic Integral with Respect to a Wiener Process 37

 
t 
× Φ(s)fj , g H dwj (s) Fu
0
∞ 
 u 
 u 
= λj Φ(s)fj , h H dwj (s) Φ(s)fj , g H dwj (s)
j =1 0 0

∞
 u    
− λj h, Φ(s)fj H g, Φ(s)fj H ds
j =1 0
∞ 
 u 
1/2 1/2  u 
+ λi λj Φ(s)fi , h H dwi (s) Φ(s)fj , g H dwj (s)
i =j =1 0 0
 u   u 
= Φ(s) dWs , h Φ(s) dWs , h
0 H 0 H
 u  
 ∗
− Φ(s)Q1/2 Φ(s)Q1/2 ds (h), g .
0 H

The formula for the increasing process follows from Lemma 2.1. 

Exercise 2.9 Prove that for Φ ∈ 2 (KQ , H ) and h ∈ H ,

  ∞
 t
t
1/2 
Φ(s) dWs , h = λj Φ(s)fj , h H dwj (t)
0 H j =1 0

with the series convergent in L2 (, R).

The following corollary follows from the proof of Theorem 2.3.

Corollary 2.1 For the sequence of bounded elementary processes Φn ∈ E (L (K,

H )) approximating Φ in 2 (KQ , H ) and satisfying condition (2.20), the corre-
sponding stochastic integrals converge uniformly with probability one,
  t  
 t 

P sup Φn (s) dWs − 
Φ(s) dWs  → 0 = 1.
t≤T 0 0 H

Exercise
t 2.10 Prove the following two properties of the stochastic integral process
0 Φ(s) dWs for Φ ∈ 2 (KQ , H ):
   
 t  1 T  2
P sup 


Φ(s, ω) dWs  > λ ≤ 2 E Φ(s, ω) ds, (2.21)
λ 2 (KQ ,H )
0≤t≤T 0 H 0

 2
 t  T  2
E sup 
 Φ(s, ω) dWs 
 ≤4 E Φ(s) ds. (2.22)
2 (KQ ,H )
0≤t≤T 0 H 0
38 2 Stochastic Calculus

Conclude that if Φ is approximated by a sequence {Φn }∞

n=1 in 2 (KQ , H ), then for
t t
every t ≤ T , 0 Φn (s) dWs → 0 Φ(s) dWs in L2 (, H ).

Remark 2.2 For Φ ∈ 2 (KQ , H ) such that Φ(s) ∈ L (K, H ), the quadratic varia-
t
tion process of the stochastic integral process 0 Φ(s) dWs and the increasing pro-
t
cess related to  0 Φ(s) dWs 2H simplify to
 ·  t
Φ(s) dWs = Φ(s)QΦ(s)∗ ds
0 t 0
and
 ·  t  
Φ(s) dWs = tr Φ(s)QΦ(s)∗ ds.
0 t 0

The final step in constructing the Itô stochastic integral is to extend it to the
class of integrands satisfying a less restrictive assumption on their second moments.
This extension is necessary if one wants to study Itô’s formula even for functions as
simple as x → x 2 . We use the approach presented in [49] for real-valued processes.
In this chapter, we will only need the concept of a real-valued progressively mea-
surable process, but in Chap. 4 we will have to refer to H -valued progressively
measurable processes. Therefore we include a more general definition here.

Definition 2.11 An H -valued stochastic process X(t), t ≥ 0, defined on a filtered

probability space (, F , {Ft }t≥0 , P ) is called progressively measurable if for ev-
ery t ≥ 0, the mapping
    
X(·, ·) : [0, t], B [0, t] ×(, Ft ) → H, B(H )

is measurable with respect to the indicated σ -fields.

It is well known (e.g., see Proposition 1.13 in [38]) that an adapted right-
continuous (or left-continuous) process is progressively measurable.

Exercise 2.11 Show that adapted right-continuous (or left-continuous) processes

are progressively measurable.

Let P(KQ , H ) denote the class of L2 (KQ , H )-valued stochastic processes

adapted to the filtration {Ft }t≤T , measurable as mappings from ([0, T ] × ,
B([0, T ]) ⊗ FT ) to (L2 (KQ , H ), B(L2 (KQ , H ))), and satisfying the condition
 T 
 2
P  
Φ(t) L (K ,H ) dt < ∞ = 1. (2.23)
2 Q
0

Obviously, 2 (KQ , H ) ⊂ P(KQ , H ). We will show that processes from P(KQ ,

H ) can be approximated in a suitable way by processes from 2 (KQ , H ) and, in
fact, by bounded elementary processes from E (L (K, H )). This procedure will al-
low us to derive basic properties of the extended stochastic integral.
2.2 Stochastic Integral with Respect to a Wiener Process 39

Lemma 2.3 Let Φ ∈ P(KQ , H ). Then there exists a sequence of bounded pro-
cesses Φn ∈ E (L (K, H )) ⊂ 2 (KQ , H ) such that
T 
Φ(t, ω) − Φn (t, ω)2 dt → 0 as n → ∞ (2.24)
L2 (KQ ,H )
0

in probability and P -a.s.

Proof For Φ ∈ P(KQ , H ), define

⎧  t 
⎪ 
⎪
⎪inf t ≤ T : Φ(s, ω)2 ds ≥ n ,
⎨ L2 (KQ ,H )
0
τn (ω) = (2.25)
⎪
⎪ T 
⎪
⎩T if Φ(s, ω)2 ds < n. L2 (KQ ,H )
0
t
The real-valued process 0 Φ(s, ω)2L2 (KQ ,H ) ds is adapted to the filtration Ft
and continuous, and hence it is progressively measurable. Therefore, τn is an
Ft -stopping time, and we can define the Ft -adapted process

Φn (t, ω) = Φ(t, ω)1{t≤τn (ω)} . (2.26)

By the definition of τn (ω) we have

T 
E Φn (t, ω)2 dt ≤ n,
L2 (KQ ,H )
0

so that Φn ∈ 2 (KQ , H ). Moreover, in view of (2.26),

 T 
 
P Φ(t, ω) − Φn (t, ω)2 dt > 0
L (K 2 Q ,H )
0
 T 

≤P Φ(t, ω)2 dt > n ,
L2 (KQ ,H )
0

so that Φn → Φ in probability in the sense of the convergence in (2.24).

By Proposition 2.2, for every n, there exists a sequence of bounded elementary
processes {Φn,k }∞
k=1 ⊂ E (L (K, H )) such that

T 
E Φn (t, ω) − Φn,k (t, ω)2 dt → 0 as n → ∞.
L2 (KQ ,H )
0

Then
 T 

P Φ(t, ω) − Φn,k (t, ω)2 dt > ε
L2 (KQ ,H )
0
 T 

≤P 2 Φ(t, ω) − Φn (t, ω)2 dt > 0
L2 (KQ ,H )
0
40 2 Stochastic Calculus
 T 

+P 2 Φn (t, ω) − Φn,k (t, ω)2 dt > ε
L2 (KQ ,H )
0
 T 

≤P Φ(t, ω) − Φn (t, ω)2 dt > 0
L2 (KQ ,H )
0
 T 
 ε
+P Φn (t, ω) − Φn,k (t, ω)2 dt >
L2 (KQ ,H ) 2
0
 T 

≤P Φ(t, ω)2 dt > n
L2 (KQ ,H )
0

2 T 
+ E Φn (t, ω) − Φn,k (t, ω)2 dt,
ε L2 (KQ ,H )
0

which proves the convergence in probability in (2.24) and P -a.s. convergence for a
subsequence. 

We can define a class of H -valued elementary processes E (H ) adapted to the

filtration {Ft }t≤T as all processes of the form


n−1
Ψ (t, ω) = ψ(ω)1{0} (t) + ψj (ω)1(tj ,tj +1 ] (t), (2.27)
j =0

where 0 = t0 ≤ t1 ≤ · · · ≤ tn = T , ψ is F0 -measurable, and ψj (j = 0, 1, . . . , n−1),

are Ftj -measurable H -valued random variables. Applying the same proof as in
Lemma 2.3, we can prove the following statement.

Lemma 2.4 Let Ψ (t), t ≤ T , be an H -valued, Ft -adapted stochastic process sat-

isfying the condition
 T 
 
P Ψ (t) dt < ∞ = 1.
H
0
Then there exists a sequence of bounded elementary processes Ψn ∈ E (H ) such that
T 
Ψ (t, ω) − Ψn (t, ω) dt → 0 as n → ∞ (2.28)
H
0

in probability and almost surely.

We will need the following useful estimate.

Lemma 2.5 Let Φ ∈ 2 (KQ , H ). Then for arbitrary δ > 0 and n > 0,
    
 t  n T 
P sup
 Φ(s) dWs 
 > δ ≤ + P Φ(s)2
2 (KQ ,H )
ds > n . (2.29)
t≤T 0 H δ2 0
2.2 Stochastic Integral with Respect to a Wiener Process 41

Proof Let τn be the stopping time defined in (2.25). Then

  t  
 

P sup Φ(s) dWs  > δ 
t≤T 0 H
   
 t  T 
= P sup
 Φ(s) dWs 
 > δ and
Φ(s)2
2 (KQ ,H )
ds > n
t≤T 0 H 0
  
 t  T 
+ P sup
 Φ(s) dWs 
 > δ and
Φ(s)2
2 (KQ ,H )
ds ≤ n .
t≤T 0 H 0
T
The first probability on the right is bounded by P ( 0 Φ(s)22 (KQ ,H ) ds > n),
while the second probability does not exceed
  t  
  1 T 
 
P sup 1[0,τn ] (s)Φ(s) dWs  > δ ≤ 2 E 1[0,τ ] (s)Φ(s)2 ds
n 2 (KQ ,H )
t≤T 0 H δ 0
n
≤
δ2
by Doob’s maximal inequality. 

We are ready to conclude the construction of the stochastic integral now.

Lemma 2.6 Let Φn be a sequence in 2 (KQ , H ) approximating a process Φ ∈

P(KQ , H ) in the sense of (2.24), i.e.,
 T 
 2
P  
Φn (t, ω) − Φ(t, ω) L (K ,H ) dt > 0 → 0.
2 Q
0

Then, there exists an H -valued FT -measurable random variable, denoted by

T
0 Φ(t) dWt , such that
T T
Φn (t) dWt → Φ(t) dWt
0 0
T
in probability. The random variable 0 Φ(t) dWt does not depend (up to stochastic
equivalence) on the choice of the approximating sequence.

Proof For every ε > 0, we have

 T 
 
lim P Φn (t, ω) − Φm (t, ω)2 > 0 = 0.
m,n→∞ L 2 (KQ ,H )
0

If δ > 0, then by (2.29)

 T  
 T 
lim sup P  Φn (t) dWt − Φm (t) dWt 
 > δ
m,n→∞ 0 0 H
42 2 Stochastic Calculus
 T 
ε  ε
≤ + lim P Φn (t, ω) − Φm (t, ω)2 dt > ε = 2,
δ 2 m,n→∞ 2 (KQ ,H ) δ
0

and since ε was arbitrary, the convergence

  
 T T 
lim P 
 Φn (t) dWt − Φm (t) dWt 
 > δ =0
m,n→∞ 0 0 H
T T
follows. The limit in probability 0 Φ(t) dWt = limn→∞ 0 Φn (t) dWt does not
depend on the choice of the approximating sequence. If Ψn is another approximating
sequence, then the two sequences can be merged into one. The resulting limit would
then have to coincide with the limits of its all subsequences. 
T
Definition 2.12 The H -valued random variable 0 Φ(t) dWt defined in Lemma 2.6
is called the stochastic integral of a process in P(KQ , H ) with respect to a
Q-Wiener
t process. For 0 ≤ t ≤ T , we define an H -valued stochastic integral pro-
cess 0 Φ(s) dWs by
t T
Φ(s) dWs = Φ(s)1[0,t] (s) dWs .
0 0

Exercise 2.12 Prove (2.29) for Φ ∈ P(KQ , H ).

Remark 2.3 We note that for Φ ∈ P(KQ , H ), the stochastic integral process
t
0 Φ(s) dWs also has a continuous T
version.
%
Indeed, let n = {ω : n − 1 ≤ 0 Φ(s)22 (KQ ,H ) < n}; then P ( − ∞
n=1 n )
= 0, and if Φn are defined as in (2.26), then on n ,

Φn (t) = Φn+1 (t) = · · · = Φ(t), t ≤ T .

t t
Therefore, on n , 0 Φ(s) dWs = 0 Φn dWs has a continuous version, and hence
it has a continuous version on .

The stochastic integral process for Φ ∈ P(KQ , H ) may not be a martingale, but
it is a local martingale. We will now discuss this property.

Definition 2.13 A stochastic process {Mt }t≤T , adapted to a filtration Ft , with val-
ues in a separable Hilbert space H is called a local martingale if there exists a
sequence of increasing stopping times τn , with P (limn→∞ τn = T ) = 1, such that
for every n, Mt∧τn is a uniformly integrable martingale.
t
Exercise 2.13 Show an example of Φ ∈ P(KQ , H ) such that 0 Φ(s) dWs is not a
martingale.

We begin with the following lemma.

2.2 Stochastic Integral with Respect to a Wiener Process 43

Lemma 2.7 Let Φ ∈ P(KQ , H ), and τ be a stopping time relative to {Ft }0≤t≤T
such that P (τ ≤ T ) = 1. Define
τ u 
Φ(t) dWt = Φ(t) dWt on the set ω : τ (ω) = u .
0 0

Then
τ T
Φ(t) dWt = Φ(t)1{t≤τ } dWt . (2.30)
0 0

Proof For an arbitrary process Φ ∈ P(KQ , H ), let Φn be a sequence of bounded

elementary processes approximating Φ as in Lemma 2.3. Since
T  T 
Φn (t)1{t≤τ } − Φ(t)1{t≤τ } 2 dt ≤ Φn (t) − Φ(t)2 dt,
2 (KQ ,H ) 2 (KQ ,H )
0 0

we conclude that
T T
Φn (t)1{t≤τ } dWt → Φ(t)1{t≤τ } dWt
0 0

in probability.
For bounded elementary processes Φ ∈ E (L (K, H )), equality (2.30) can be
τ T
verified by inspection, so that 0 Φn (t) dWt = 0 Φn (s)1{s≤τ } dWs .
τ u
On the set {ω : τ (ω) = u}, we have 0 Φn (t) dWt = 0 Φn (t) dWt . Also, for
every u ≤ T ,
u u
Φn (t) dWt → Φ(t) dWt
0 0
in probability. Thus, for every u ≤ T ,
T u
1{τ =u} Φn (t)1{t≤τ } dWt → 1{τ =u} Φ(t) dWt
0 0

in probability. This implies that for every u ≤ T ,

u T
1{τ =u} Φ(t) dWt = 1{τ =u} Φ(t)1{t≤τ } dWt P -a.s.
0 0
u
Since the stochastic integral process 0 Φ(t) dWt , u ≤ T is P -a.s. continuous, we
get that the above equality holds P -a.s. for all u ≤ T , and (2.30) follows. 
t
Now observe that for Φ∈P(KQ , H ), the stochastic integral process 0 Φ(s) dWs
is a local martingale, with the localizing sequence of stopping
 t∧τtimes τn defined
in (2.25). We have P (limn→∞ τn = T ) = 1, and, by (2.30), 0 n Φ(s) dWs is a
44 2 Stochastic Calculus

martingale with
 2 
 t∧τn  T 
E 
 Φ(s) dWs 
 =E Φ(s)2 1
L2 (KQ ,H ) {s≤τn }
ds ≤ n.
0 H 0
 t∧τn
This proves that for every n, the process 0 Φ(s) dWs is a square-integrable and
hence a uniformly integrable martingale.

2.2.4 Stochastic Itô Integral with Respect to Cylindrical Wiener

Process

We now proceed with the definition of the stochastic integral with respect to a cylin-
drical Wiener process. We will restrict ourselves to the case where the integrand
Φ(s) is a process taking values in L2 (K, H ), following the work in [18]. A more
general approach can be found in [57] and [19].
We recall that if Φ(s) is an elementary process, Φ(s) ∈ E (L (K, H )), then
Φ(s) ∈ L2 (K, H ), since Q = IK . Assume that Φ(s) is bounded in the norm of
L2 (K, H ). Using (2.14), with {ei }∞
i=1 an ONB in H , we calculate,

∞ 
  2 
∞
t t  2
E Φ(s) d W̃s (ei ) = E Φ ∗ (s)ei K ds
i=1 0 i=1 0
∞
t   t 
=E Φ ∗ (s)ei 2 ds = E Φ ∗ (s)2 ds
K L2 (H,K)
0 i=1 0

t 
=E Φ(s)2 ds.
L2 (K,H )
0
t
Then we define the stochastic integral 0 Φ(s) d W̃s of a bounded elementary process
Φ(s) as follows:

t ∞ 
 t  
Φ(s) d W̃s = Φ(s) d W̃s (ei ) ei . (2.31)
0 i=1 0

t
By the above calculations, 0 Φ(s) d W̃s ∈ L2 (, H ) and is adapted to the filtra-
tion Ft . The equality
 
 T 
 Φ(s) d W̃s  = Φ2 (K,H ) (2.32)
 
0 L2 (,H )

establishes the isometry property of the stochastic integral transformation.

2.2 Stochastic Integral with Respect to a Wiener Process 45

Definition 2.14 The stochastic integral of a process Φ ∈ 2 (K, H ) with respect

to a standard cylindrical Wiener process W̃t in a Hilbert space K is the unique
isometric linear extension of the mapping
T
Φ(·) → Φ(s) d W̃s
0

from the class of bounded elementary processes to L2 (, H )  to a mapping from

2 (K, H ) to L2 (, H ), such that the image of Φ(t)=φ1{0} (t)+ n−1
j =0 φj 1(tj ,tj +1 ] (t)
is
∞ 
 n
    
W̃tj +1 ∧t φj∗ (ei ) − W̃tj ∧t φj∗ (ei ) ei .
i=1 j =0
t
We define the stochastic integral process 0 Φ(s) d W̃s , 0 ≤ t ≤ T , for Φ ∈
2 (K, H ) by
t T
Φ(s) d W̃s = Φ(s)1[0,t] (s) d W̃s .
0 0

The next theorem can be proved in a similar manner as Theorem 2.3.

·
Theorem 2.4 The stochastic integral Φ → 0 Φ(s) d W̃s with respect to a cylin-
drical Wiener process W̃t in K is an isometry between 2 (K, H ) and the space of
continuous square-integrable martingales MT2 (H ). The quadratic variation process
t
of the stochastic integral process 0 Φ(s) d W̃s is given by
 ·  t
Φ(s) d W̃s = Φ(s)Φ ∗ (s) ds
0 t 0
and
  t
· t 
Φ(s) d W̃s = tr Φ(s)Φ ∗ (s) ds = Φ(s)2 ds.
L2 (K,H )
0 t 0 0

Similarly as in the case of the stochastic integral with respect to a Q-Wiener pro-
cess, we now conclude the construction of the integral with respect to a cylindrical
Wiener process.
t
Remark 2.4 Since for Φ ∈ 2 (K, H ), the process 0 Φ(s) d W̃s ∈ MT2 (H ), the con-
clusion of Lemma 2.5 holds in the cylindrical case.
Define P(K, H ) = P(KQ , H ) with Q = IK . We can construct the cylindri-
T
cal stochastic integral 0 Φ(t) d W̃t for processes Φ ∈ P(K, H ) by repeating the
arguments in Lemma 2.6.

Exercise 2.14 Verify the statements in Remark 2.4.

46 2 Stochastic Calculus

We arrive at the following definition.

Definition 2.15 Let Φn be a sequence in 2 (K, H ) approximating a process Φ ∈

P(K, H ) in the sense of (2.24), i.e.,
 T 

P Φn (t, ω) − Φ(t, ω)2 dt > 0 → 0.
L2 (K,H )
0
T T
Denote the limit in probability of the sequence 0 Φn (t) d W̃t by 0 Φ(t) d W̃t and
call the limit the stochastic integral of Φ with respect to the cylindrical Wiener
process W̃t . The integral is an H -valued FT -measurable random variable, and it
does not depend (up to stochastic equivalence) on the choice of the approximating
sequence. t
For 0 ≤ t ≤ T , we define an H -valued stochastic integral process 0 Φ(s) d W̃s
by
t T
Φ(s) d W̃s = Φ(s)1[0,t] (s) d W̃s .
0 0

Exercise 2.15 (a) Prove (2.30) in the cylindrical case, i.e., show that for a process
Φ ∈ P(K, H ) and a stopping time τ relative to {Ft }0≤t≤T such that P (τ ≤ T ) = 1,
τ T
Φ(t) d W̃t = Φ(t)1{t≤τ } d W̃t , (2.33)
0 0

where
τ u 
Φ(t) d W̃t = Φ(t) d W̃t on the set ω : τ (ω) = u .
0 0
t
(b) Show that the stochastic integral 0 Φ(s) d W̃s 0 ≤ t ≤ T , is a local martingale
and that it has a continuous version.

The following representation of the stochastic integral with respect to a Q-Wiener

process and with respect to a cylindrical Wiener process if Q = IK can be also used
as a definition.

Lemma 2.8 Let Wt be a Q-Wiener process in a separable Hilbert space K, Φ ∈

2 (KQ , H ), and {fj }∞
j =1 be an ONB in K consisting of eigenvectors of Q. Then

∞
 t
1/2   1/2 
t
Φ(s) dWs = Φ(s)λj fj d Ws , λj fj K
Q
0 j =1 0
∞
 t 
= Φ(s)fj dWs , fj K. (2.34)
j =1 0
2.2 Stochastic Integral with Respect to a Wiener Process 47

For a cylindrical Wiener process W̃t , if Φ ∈ 2 (K, H ) and {fj }∞

j =1 is an ONB in K,
then
∞
 t
t 
Φ(s) d W̃s = Φ(s)fj d W̃s (fj ). (2.35)
0 j =1 0

Proof We will prove (2.35), the cylindrical case only, since the proof for a
Q-Wiener process is nearly identical.
We first note that

∞ 2 ∞ 2
 t   ∞ 
 t 
 
E Φ(s)fj d W̃s (fj ) = E Φ(s)fj d W̃s (fj ), ei
 
j =1 0 H i=1 j =1 0 H
∞ 
 ∞ t 2
= E Φ(s)fj , ei H
i=1 j =1 0

t 
=E Φ(s)2 < ∞.
L2 (K,H )
0

∞  t
Thus, j =1 0 (Φ(s)fj ) d W̃s (fj ) ∈ H , P -a.s. For a bounded elementary process

Φ(s) = 1{0} φ + n−1
k=1 φk 1(tk ,tk+1 ] (s) ∈ E (L (K, H )) and any h ∈ H , we have that
a.s.

  
n−1
t     
Φ(s) d W̃s , h = W̃tk+1 φk∗ (h) − W̃tk φk∗ (h)
0 H k=1

 ∞
n−1 
  
= W̃tk+1 (fj ) − W̃tk (fj ) fj , φk∗ (h) K
k=1 j =1

∞ 
 n−1
   
= φk (fj ) W̃tk+1 (fj ) − W̃tk (fj ) , h H
j =1 k=1
 ∞ n−1 
  
= φk (fj ) W̃tk+1 (fj ) − W̃tk (fj ) , h
j =1 k=1 H
∞ 
 t 
= Φ(s)fj d W̃s (fj ), h ,
j =1 0 H

so that (2.35) holds in this case.

48 2 Stochastic Calculus

⊥
Let Pm+1 denote the orthogonal projection onto span{fm+1 , fm+2 , . . .}. Now
for a general Φ ∈ 2 (K, H ), we have for an approximating sequence Φn (s) ∈
E (L (K, H )), using (2.35) for elementary processes,

 2
 t 
m t  
 
E Φ(s) d W̃s − Φ(s)fj d W̃s (fj )
 0 
j =1 0 H
 
 t t t

= lim E  Φ(s) d W̃s − Φn (s) d W̃s + Φn (s) d W̃s
n→∞  0 0 0
2

m t  

− Φ(s)fj d W̃s (fj )

j =1 0 H
 ∞
  t  
m t 

= lim E  Φn (s)fj d W̃s (fj ) + Φn (s)fj d W̃s (fj )
n→∞ 
j =m+1 0 j =1 0
2

m t  

− Φ(s)fj d W̃s (fj )

j =1 0 H
 ∞ 2
  t  
 
= lim E  Φn (s)fj d W̃s (fj )
n→∞  
j =m+1 0 H
∞ 2
 t    
 ⊥ 
= lim E  Φn (s)Pm+1 fj d W̃s (fj )
n→∞  
j =1 0 H
 2
 t  
= lim E  ⊥
Φn (s)Pm+1 d W̃s 
n→∞  0

H
 2
 t  
= E

⊥
Φ(s)Pm+1 d W̃s 

0 H
t 
=E Φ(s)P ⊥ 2
m+1 L2 (K,H ) ds
0
∞

t  
=E Φ(s)fj 2 → 0, as m → ∞,
H
0 j =m+1

⊥ → ΦP ⊥ in  (K, H ) as n → ∞. This
where we have used the fact that Φn Pm+1 m+1 2
concludes the proof. 
2.2 Stochastic Integral with Respect to a Wiener Process 49

2.2.5 The Martingale Representation Theorem

We recall (Definition 2.6) that an H -valued Q-Wiener process is defined by

∞
 1/2
Wt = λj wj (t)fj ,
j =1

where {fj }∞ ∞
j =1 is an ONB in a separable Hilbert space K, {wj (t)}j =1 are inde-
∞
pendent Brownian motions, and {λj }j =1 are summable and assumed to be strictly
positive without loss of generality. Let us denote
∞ 
j  & j
Ft = σ wj (s) : s ≤ t , Gt = σ Ft ,
j =1

and

FtW = σ Ws (k) : k ∈ K, and s ≤ t .

Then clearly FTW = GT and

∞
'
   j 
L2 , FTW , P = L2 (, GT , P ) = L2 , FT , P .
j =1

Exercise 2.16 Prove that if Fi , i = 1, 2, . . . , are independent σ -fields, then

 ∞   ∞
& '
2
L , σ Fi , P = L2 (, Fi , P ).
i=1 i=1

In view of the fact that the linear span

T
h(t)dwj (t)− 12
T
h2 (t) dt
 !
span e 0 0 : h ∈ L2 [0, T ], R

j
is dense in L2 (, FT , P ) ([61], Lemma 4.3.2), we deduce that the linear span
T 
1 T 2   !
span e 0 h(t)dwj (t)− 2 0 h (t) dt : h ∈ L2 [0, T ], R , j = 1, 2, . . .

is dense in L2 (, FTW , P ).

Now following the proof of Theorem 4.3.4 in [61], we conclude that every real-
valued FtW -martingale mt in L2 (, FTW , P ) has a unique representation
∞
 t
1/2
mt (ω) = Em0 + λj φj (s, ω) dwj (s), (2.36)
j =1 0

∞ T
where j =1 λj E 0 φj2 (s, ω) ds < ∞.
50 2 Stochastic Calculus

Let H be a separable Hilbert space, and Mt be an FtW -martingale, such that

EMt 2H < ∞. Choose an ONB {ej }∞ j =1 ⊂ H . Then there exist unique processes
i
φj (t, ω) as in the representation (2.36) such that

∞
 t
1/2
Mt , ei H = EM0 , ei H + λj φji (s, ω) dwj (s).
j =1 0

∞
i=1 Mt , ei H < ∞, we have
Since E 2

∞

Mt = Mt , ei H ei .
i=1

Therefore,
∞
 ∞ 
 ∞ t
1/2
Mt = E M0 , ei H ei + λj φji (s, ω)ei dwj (s). (2.37)
i=1 i=1 j =1 0

Under the assumptions on Mt , we obtain that EM0 H < ∞, so that the first term
is equal to EM0 . Using the assumptions on Mt and the representations (2.36) and
(2.37) above, we obtain

∞
∞ 2 ∞ 
∞
  t
1/2
 t  2
E λj φji (s, ω) dwj (s) = λj E φji (s, ω) ds.
i=1 j =1 0 j =1 i=1 0

This justifies interchanging the summations in (2.37) to write Mt as

∞
 ∞
 t
1/2
Mt = EM0 + λj φji (s, ω)ei dwj (s). (2.38)
j =1 i=1 0

Define for k ∈ KQ , h ∈ H ,

∞ 
 ∞
 
Φ(s, ω)k, h H = λj h, ei H k, fj KQ φj (s, ω).
i

j =1 i=1

Then Φ(s, ω) ∈ 2 (KQ , H ), and, by the definition of the stochastic integral, the
second term in (2.38) is equal to
t
Φ(s, ω) dW (s).
0

We have the following theorem.

2.2 Stochastic Integral with Respect to a Wiener Process 51

Theorem 2.5 (Martingale Representation Theorem I) Let H and K be separable

Hilbert spaces, Wt be a K-valued Q-Wiener process, and Mt an H -valued con-
tinuous FtW -martingale such that EMt 2H < ∞ for all t ≥ 0. Then there exists a
unique process Φ(t) ∈ 2 (KQ , H ) such that
t
Mt = EM0 + Φ(s) dWs .
0

Remark 2.5 If EM0 = 0, then, by Theorem 2.3, the quadratic variation process
corresponding to Mt and the increasing process related to Mt 2H are given by
(see [57])
t  ∗
M t = Φ(s)Q1/2 Φ(s)Q1/2 ds,
0
t
(2.39)
 t   ∗ 
M t = Φ(s)2 ds = tr Φ(s)Q1/2 Φ(s)Q1/2 ds.
L2 (KQ ,H )
0 0

We shall now prove the converse. We need the following two results.

Theorem 2.6 (Lévy) Let Mt , 0 ≤ t ≤ T , be a K-valued continuous square-

integrable martingale with respect to a filtration {Ft }t≤T , Assume that its quadratic
variation process is of the form M t = tQ, t ∈ [0, T ]. Then Mt is a Q-Wiener pro-
cess with respect to the filtration {Ft }t≤T .

Proof Consider Mtn = ( 1/2 Mt , f1 K , . . . , 1/2 Mt , fn K ),

1 1
where {fj }∞
j =1 is an
λ1 λn
ONB in K such that Qfj = λj fj . Then, by the classical Lévy theorem, Mtn is
an n-dimensional Brownian motion with respect to the filtration {Ft }t≤T . This im-

plies that Mt = ∞
1/2
j =1 λj wj (t)fj , where wj (t) = 1/2 Mt , fj , j = 1, 2, . . . , are
1
λj
independent Brownian motions with respect to {Ft }t≤T . 

Using a theorem on measurable selectors of Kuratowski and Ryll–Nardzewski

[43], we obtain the following lemma in [11]. Below, for a separable Hilbert space
H , L1 (H ) denotes the (separable) space of trace-class operators on H .

Lemma 2.9 Let H be a separable Hilbert space, and Φ be a measurable function

from (, F ) to (L1 (H ), B(L1 (H ))) such that Φ(ω) is a nonnegative definite op-
erator for every ω. Then there exists a nonincreasing sequence of real-valued non-
negative measurable functions λn (ω) and H -valued measurable functions gn (ω)
such that for all h ∈ H ,
∞
  
Φ(ω)h = λn (ω) gn (ω), h H gn (ω).
n=1
52 2 Stochastic Calculus

Moreover, the sequences λn and gn can be chosen to satisfy


 
gn (ω) = 1 if λn (ω) > 0,
H 0 if λn (ω) = 0,

and gn (ω), gm (ω) H = δm,n , ∀ω ∈  .

t
We know that Mt = 0 Φ(s)dWs describes all continuous FtW -martingales such
t
that EMt 2H is finite, M t = 0 (Φ(s)Q1/2 )(Φ(s)Q1/2 )∗ ds, and EM0 = 0. The
t
related increasing process is given by M t = 0 tr((Φ(s)Q1/2 )(Φ(s)Q1/2 )∗ ) ds.
Now let (, F , {Ft }t≤T , P ) be a filtered probability space, and Mt be an H -valued
square-integrable martingale relative to Ft . Assume that
t
M t = QM (s) ds,
0

where the process QM (s, ω) is adapted to Ft with values in nonnegative definite

symmetric trace-class operators on H .
Then we can define a stochastic integral
t
Nt = Ψ (s) dMs (2.40)
0

with respect to Mt exactly as we did for the case of a Wiener process. The integrands
are Ft -adapted processes Ψ (t, ω) with values in linear, but not necessarily bounded,
operators from H to a separable Hilbert space G satisfying the condition
T   
tr Ψ (s)QM (s) Ψ (s)QM (s)∗ ds < ∞.
1/2 1/2
E (2.41)
0

The stochastic integral process Nt ∈ MT2 (G), and its quadratic variation is given by
t  ∗
1/2 1/2
N t = Ψ (s)QM (s) Ψ (s)QM (s) ds. (2.42)
0

In particular, we may have

  ∗
QM (s, ω) = Φ(s, ω)Q1/2 Φ(s, ω)Q1/2

with Φ(s, ω) ∈ 2 (KQ , H ) and Mt adapted to FtW for a Q-Wiener process Wt ,

and EM0 = 0. In this case
t
Mt = Φ(s) dWs
0
and (2.43)
t t
Nt = Ψ (s) dMs = Ψ (s)Φ(s) dWs .
0 0
2.2 Stochastic Integral with Respect to a Wiener Process 53

By (2.39), the quadratic variation process of Nt has the form

t  ∗
N t = Ψ (s)Φ(s)Q1/2 Ψ (s)Φ(s)Q1/2 ds. (2.44)
0

Exercise 2.17 Reconcile formulas (2.42) and (2.44). Hint: use Lemma 2.10 to show
that if L ∈ L (H ), then (LL∗ )1/2 = LJ , where J is a partial isometry on (ker L)⊥ .

Exercise 2.18 Provide details for construction of the stochastic integral (2.40) with
respect to square-integrable martingales for the class of stochastic processes
t satis-
fying condition (2.41). Prove property (2.42). Show (2.43) for Mt = 0 Φ(s) dWs .

We shall use this integral for representing a square-integrable

t martingale with
its quadratic variation process given by M t = 0 (Φ(s)Q1/2 )(Φ(s)Q1/2 )∗ ds, in
terms of some Wiener process. This provides the converse of the Martingale Repre-
sentation Theorem I (Theorem 2.5). The formulation and proof of the theorem we
present are taken directly from [11].

Theorem 2.7 (Martingale Representation Theorem II) Let Mt , 0 ≤ t ≤ T , be an

H -valued continuous martingale with respect to a filtration {Ft }∞
t=0 . Assume that its
t
quadratic variation process is given by M t = 0 (Φ(s, ω)Q )(Φ(s, ω)Q1/2 )∗ ds,
1/2

where Φ(s, ω) is an adapted 2 (KQ , H )-valued process. Then there exists a

K-valued Q-Wiener process on an extended probability space ( × , ˜ F × F˜ ,
P × P̃ ) adapted to filtration {Ft × F˜t } such that
t
Mt (ω) = Φ(s, ω) dWs (ω, ω̃).
0

In addition, the Wiener process can be constructed so that its increments Wt − Ws

are independent of Fs for t ≥ s.

Proof To simplify the notation, denote Ψ (s, ω) = Φ(s, ω)Q1/2 . We shall prove that
if Mt is an H -valued continuous Ft -martingale with the quadratic variation process
t
M t = Ψ (s, ω)Ψ ∗ (s, ω) ds
0
T t
such that E 0 tr(Ψ (s)Ψ ∗ (s)) ds < ∞, then Mt = 0 Φ(s, ω) dWs , where Wt is a
Q-Wiener process.
Since M t = 0 Ψ (s, ω)Ψ ∗ (s, ω) ds, the space Im(Ψ (s, ω)Ψ ∗ (s, ω)) will play
t

a key role in reconstructing Mt .

By Lemma 2.9 we get
∞

   
Ψ (s, ω)Ψ ∗ (s, ω) h = λn (s, ω) gn (s, ω), h H gn (s, ω),
n=1
54 2 Stochastic Calculus

where λn and gn are Ft -adapted. Let {fn }∞

n=1 be an ONB in K and define
∞

V (s, ω)k = k, fn K gn (s, ω).
n=1

Then
∞
  
V ∗ (s, ω)h = gn (s, ω), h H fn ,
n=1

and the Ft -adapted process

Π(s, ω) = V (s)V ∗ (s)

is an orthogonal projection on Im(Ψ (s, ω)Ψ ∗ (s, ω)). Thus we can write
t 
Mt = Π(s) + Π ⊥ (s) dMs
0
t t
= Π(s) dMs + Π ⊥ (s) dMs = Mt + Mt .
0 0
t
But M0 = 0 and M  t = 0 Π ⊥ (s)Ψ (s)Ψ ∗ (s)Π ⊥ (s) ds = 0, so that Mt = 0. In
conclusion,
t t
Mt = Mt + Mt = V (s)V ∗ (s) dMs = V (s) dNs
0 0

with
t
Nt = V ∗ (s, ω) dMs ,
0
a continuous K-valued square integrable martingale whose quadratic variation is
given by
t
N t = V ∗ (s, ω)Ψ (s, ω)Ψ ∗ (s, ω)V (s, ω) ds.
0
We define now
∞

 
Λ(s, ω)k = V ∗ (s, ω)Ψ (s, ω)Ψ ∗ (s, ω)V (s, ω) k = λn (s, ω)k, fn K fn ,
n=1

so that we can represent Nt through its series expansion in K,

∞

Nt = ηn (t)fn ,
n=1
t
where ηn (t) are real-valued martingales with increasing processes 0 λn (s, ω) ds.
2.2 Stochastic Integral with Respect to a Wiener Process 55

We consider now a filtered probability space

 
 ×  , F × F  , {Ft × Ft }t≤T , P × P  ,

where ( , F  , {Ft }t≤T , P  ) is another filtered probability space. Define


−1/2
λn (s, ω) if λn (s, ω) > 0,
δn (s, ω) =
0 if λn (s, ω) = 0,
and

0 if λn (s, ω) > 0,
γn (s, ω) =
1 if λn (s, ω) = 0.

Let βn (t) be independent Brownian motions on ( , F  , {Ft }t≤T , P  ).

We extend processes defined on either  or  alone to the product space  ×  ,
e.g., by M(t, ω, ω ) = M(t, ω), and define
t t
ŵn (t) = δn (s) dηn (s) + γn (s) dβn (s).
0 0

Then it is easy to verify that ŵn , ŵm t = tδn,m , using the mutual independence of
ηn and βm and the fact that
 
ηn , ηm t = N t fn , fm K = 0.

Thus, by Lévy’s theorem, ŵn (t) are independent Brownian motions, and the expres-
sion
∞
  
(t (k) =
W ŵn (t)fn , k K
n=1
defines a K-valued cylindrical Wiener process.
Since
t t
1/2 1/2
λn (s) d ŵn (s) = λn (s)δn (s) dηn (s) = ηn (t),
0 0
we get, using, for example, (2.35), that
t
Nt = (s .
Λ1/2 (s) d W
0

Thus we arrive at
t t
Mt = V (s) dNs = (s .
V (s, ω)Λ1/2 (s, ω) d W
0 0

All that is needed now is a modification of the integrand Ψ̂ (s) = V (s)Λ1/2 (s) to the
desired form Ψ (s), and the cylindrical Wiener process W (t needs to be replaced with
a Q-Wiener process.
56 2 Stochastic Calculus

It follows directly from the definitions of V (s) and Λ(s) that

Ψ (s)Ψ ∗ (s) = Ψ̂ (s)Ψ̂ ∗ (s).

Now we need the following general fact from the operator theory (refer to [11],
Appendix B).

Lemma 2.10 Let H be a Hilbert space, A, B ∈ L (H ), and

AA∗ = BB ∗ .

Denote by  the orthogonal projection onto (ker B)⊥ . Define

J = A−1 B : H → (ker A)⊥ ,

where A−1 : R(A) → D(A) is the pseudo-inverse operator, i.e., A−1 h is defined as
the element g of the minimal norm such that Ag = h. Then

B = AJ,

where J is a partial isometry on (ker B)⊥ , and J J ∗ is an orthogonal projection

onto (ker A)⊥ .

It follows from the formula defining the operator J that there exists an
Ft -adapted process J (t) : (ker Ψ̂ (t))⊥ → (ker Ψ (t))⊥ such that

Ψ̂ (t) = Ψ (t)J (t)

and such that J (t)J ∗ (t) is an orthogonal projection onto (ker Ψ (t))⊥ . We need an-
other filtered probability space ( , F  , {Ft }t≤T , P  ) and a cylindrical Wiener
process W ( and extend all processes trivially to the product filtered probability
t
space
  
 ×  ×  , F × F  × F  , Ft × Ft × Ft , P × P  × P  ,

e.g., by Mt (ω, ω , ω ) = Mt (ω, ω ) = Mt (ω). If we define

t t
Wt = (s +
Q1/2 J (s) d W (
Q1/2 K(s) d W s
0 0

with K(s) = (J (s)J ∗ (s))⊥ , the projection onto ker Ψ (s), then, using Lévy’s theo-
rem (Theorem 2.6), Wt is a Q-Wiener process, since by Theorem 2.4
t 
W t = Q1/2 J (s)J ∗ (s)Q1/2 + Q1/2 K(s)Q1/2 ds = tQ.
0
2.2 Stochastic Integral with Respect to a Wiener Process 57

Also
t t t
Φ(s) dWs = (s +
Ψ (s)J (s) d W (
Ψ (s)K(s) d W s
0 0 0
t
= Ψ (s = Mt .
((s) d W
0


We can now prove the following martingale representation theorem in the cylin-
drical case.

Corollary 2.2 Let Mt , 0 ≤ t ≤ T , be an H -valued continuous martingale with re-

spect to a filtration {Ft }∞
t=0 . Assume that its quadratic variation process is given
t ∗
by M t = 0 Ψ (s, ω)Ψ (s, ω) ds, where Ψ (s, ω) is an adapted 2 (K, H )-valued
process. Then there exists a cylindrical Wiener process W̃t in K on an extended
probability space ( × , ˜ F × F˜ , P × P̃ ) adapted to filtration {Ft × F˜t } such
that
t
Mt (ω) = Ψ (s, ω) d W̃s (ω, ω̃).
0
In addition, the cylindrical Wiener process can be constructed so that for any k ∈ K,
the increments W̃t (k) − W̃s (k) were independent of Fs , if t ≥ s.

Proof Let Q be a symmetric nonnegative definite trace–class operator on K with

strictly positive eigenvalues. Define Φ(s) = Ψ (s)Q−1/2 ∈ (KQ , H ). With the no-
tation from the proof of Theorem 2.7, we have
t t t
Mt = Φ(s) dWs = Ψ (s)Q−1/2 dWs = )s ,
Ψ (s) d W
0 0 0

)t (k) = Ws , Q−1/2 k

where the relation W K, k ∈ K, defines the desired cylindrical
Wiener process. 

2.2.6 Stochastic Fubini Theorem

The stochastic version of the Fubini theorem helps calculate deterministic integrals
of an integrand that is a stochastic integral process. In literature, this theorem is
presented for predictable processes, but there is no need for this restriction if the
stochastic integral is relative to a Wiener process.

Theorem 2.8 (Stochastic Fubini Theorem) Let (G, G , μ) be a finite measurable

space, and Φ : ([0, T ] ×  × G, B([0, T ]) ⊗ FT ⊗ G ) → (H, B(H )) be a mea-
surable map such that for every x ∈ G, the process Φ(·, ·, x) is {Ft }t≤T -adapted.
58 2 Stochastic Calculus

Let Wt be a Q-Wiener process on a filtered probability space (, F , {Ft }t≤T , P ).

If
 
9Φ9 = Φ(·, ·, x) (K ,H ) μ(dx) < ∞, (2.45)
2 Q
G
then
T
(1) 0 Φ(t, ·, x) dWt has a measurable version as a map from ( × G, FT ⊗ G )
 (H, B(H ));
to
(2) G Φ(·, ·, x) μ(dx) is {Ft }t≤T -adapted;
(3) The following equality holds P -a.s.:
 T  T 
Φ(t, ·, x) dWt μ(dx) = Φ(t, ·, x) μ(dx) dWt . (2.46)
G 0 0 G

Proof Note that condition (2.45) implies that

T 
E Φ(t, ω, x) dt μ(dx) < ∞,
L2 (KQ ,H )
G 0

so that Φ ∈ L1 ([0, T ] ×  × G). Also, Φ(·, ·, x) ∈ 2 (KQ , H ) for almost all x ∈ G.

We will carry the proof out in two steps.
(A) We can assume without loss of generality that for all x ∈ G,
 
Φ(·, ·, x) < M.
L (K ,H ) 2 Q

In order to prove (1)–(3) for an unbounded Φ ∈ L1 ([0, T ]× ×G) with Φ(·, ·, x) ∈
2 (KQ , H ) μ-a.e., we only need to know that (1)–(3) hold for a  · L2 (KQ ,H ) -norm
bounded sequence Φn such that 9Φn −Φ9 → 0. We define an appropriate sequence
by
⎧
⎨n Φ(t, ω, x)
if Φ(t, ω, x)L2 (KQ ,H ) > n,
Φn (t, ω, x) = Φ(t, ω, x)L2 (KQ ,H )
⎩
Φ(t, ω, x) otherwise.

By the Lebesgue DCT relative to the  · L2 (KQ ,H ) -norm, we have that μ-a.e.
 
Φn (·, ·, x) − Φ(·, ·, x) → 0.
2 (KQ ,H )

By the isometric property (2.19) of the stochastic integral,

T T
lim Φn (t, ·, x) dWt = Φ(t, ·, x) dWt (2.47)
n→∞ 0 0

in L2 (), and, by selecting a subsequence if necessary, we can assume that the

convergence in (2.47) is P -a.s.
2.2 Stochastic Integral with Respect to a Wiener Process 59

We prove (1) by defining an FT ⊗ G -measurable version of the stochastic inte-

gral
⎧ T
⎨
T
lim Φn (t, ·, x) dWt if the limit exists,
Φ(t, ·, x) dWt = n→∞ 0
0 ⎩
0 otherwise.

Again, by the Lebesgue DCT, we have that Φn → Φ in L1 ([0, T ] ×  × G),

so that we can assume, by selecting a subsequence if necessary, that Φn → Φ for
almost all (t, ω, x), and hence (2) follows for Φ.
To prove (3), we consider
     
 T T 
E
 Φn (t, ·, x) dWt μ(dx) − Φ(t, ·, x) dWt μ(dx)

G 0 G 0 H

≤ 9Φn − Φ9 → 0 (2.48)

and
 T  T  
 
E
 Φn (t, ·, x) μ(dx) dWt − Φ(t, ·, x) μ(dx) dWt 

0 G 0 G H
  T  2 1/2
   
≤ E
 Φn (t, ·, x) − Φ(t, ·, x) μ(dx) dWt 

0 G H
 
T
2 1/2
  
≤ E  Φn (t, ·, x) − Φ(t, ·, x) μ(dx)
 
0 G L2 (KQ ,H )
 
   
=
 Φn (·, ·, x) − Φ(·, ·, x) μ(dx)

G 2 (KQ ,H )
 
≤ Φn (·, ·, x) − Φ(·, ·, x) μ(dx)
2 (KQ ,H )
G

= 9Φn − Φ9 → 0. (2.49)

Now, (3) follows for Φ from (2.48) and (2.49), since it is valid for Φn .
(B) If Φ is bounded in the  · L2 (KQ ,H ) -norm, then it can be approximated in
9 · 9 by bounded elementary processes


n−1
Φ(t, ω, x) = Φ(0, ω, x)1{0} (t) + Φj (tj , ω, x)1(tj ,tj +1 ] (t), (2.50)
j =1

where 0 ≤ t1 ≤ · · · ≤ tn = T , and the size of partition of [0, T ] converges to 0.

60 2 Stochastic Calculus

This can be seen by replacing P with P ⊗ μ in Proposition 2.2. Since μ is finite

and Φ is bounded, we have
  T 1/2 2

E Φn (t, ω, x) − Φ(t, ω, x)2 dt μ(dx)
L2 (KQ ,H )
G 0
 T 

≤ μ(G) E Φn (t, ω, x) − Φ(t, ω, x)2 dt μ(dx)
L2 (KQ ,H )
G 0
T   
≤ Φn (t, ω, x) − Φ(t, ω, x)2 dt P (dω) ⊗ μ(dx)
L2 (KQ ,H )
G× 0

with Φ(t, ω, x), square integrable with respect to dt ⊗ dP ⊗ dμ, so that Proposi-
tion 2.2 gives the desired approximation.
Clearly, Φn (·, ·, x) is {Ft }t≤T -adapted for any x ∈ G, and the stochastic integral
T
0 Φn (t, ·, x) dWt is FT ⊗ G /B(H )-measurable.
Since for every t ∈ T and A ∈ L2 (KQ , H ),
 
Φn (t, ·, ·), A L
2 (KQ ,H )

is FT ⊗ G /B(R)-measurable and P ⊗ μ-integrable, then by the classical Fubini

theorem, the function

 
Φn (t, ·, x), A L μ(dx)
2 (KQ ,H )
G

is Ft -measurable, and so is the function

Φn (t, ·, x) μ(dx)
G

by the separability of L2 (KQ , H ).

Obviously, (3) holds for Φn .
Now (B) follows by repeating the arguments in (A), since Φn satisfies conditions
(1)–(3) and 9Φn − Φ9 → 0, so that (2.48) and (2.49) are also valid here. 

Let us now discuss the cylindrical case. In the statement of Theorem  t 2.8 we can
consider W̃t , a cylindrical Wiener process, and the stochastic integral 0 Φ(s) d W̃s .
Definitions 2.10 and 2.14 differ only by the choice of Q being either a trace-class op-
erator or Q = IK , but in both cases the integrands are in 2 (KQ , H ). Both stochas-
tic integrals are isometries by either (2.19) or (2.32). We have therefore the following
conclusion.

Corollary 2.3 Under the assumptions of Theorem 2.8, with condition (2.45) re-
placed with
2.3 The Itô Formula 61

 
9Φ91 = Φ(·, ·, x) μ(dx) < ∞, (2.51)
2 (K,H )
G
conclusions (1)–(3) of the stochastic Fubini theorem hold for the stochastic integral
T
0 Φ(t, ·, ·) d W̃t with respect to a standard cylindrical Wiener process {W̃t }t≥0 .

2.3 The Itô Formula

We will present a theorem that gives conditions under which a stochastic process
F (t, X(t)) has a stochastic differential, provided that X(t) has a stochastic differ-
ential. First we explain some generally used notation.

2.3.1 The case of a Q-Wiener process

If φ ∈ L2 (KQ , H ) and ψ ∈ H , then φ ∗ ψ ∈ L2 (KQ , R), since

∞
 ∞
 ∗ 2  ∗  1/2 2    2
φ ψ  = φ ψ λ fj = ψ, φ λ1/2 fj H
L2 (KQ ,R)
j =1 j =1

≤ ψ2H φ2L2 (KQ ,H ) .

Hence, if Φ(s) ∈ P(KQ , H ) and Ψ (s) ∈ H are Ft -adapted processes, then the
process Φ ∗ (s)Ψ (s) defined by
 ∗   
Φ (s)Ψ (s) (k) = Ψ (s), Φ(s)(k) H

has values in L2 (KQ , R). If, in addition, P -a.s., Ψ (s) is bounded as a function of
s, then
 T 
 ∗ 2
P  
Φ (s)Ψ (s) L (K ,R) ds < ∞ = 1,
2 Q
0

so that Φ ∗ (s)Ψ (s) ∈ P(KQ , R), and we can define

T  T
Ψ (s), Φ(s) dWs H
= Φ ∗ (s)Ψ (s) dWs .
0 0

Theorem 2.9 (Itô Formula) Let Q be a symmetric nonnegative trace-class operator

on a separable Hilbert space K, and let {Wt }0≤t≤T be a Q-Wiener process on a
filtered probability space (, F , {Ft }0≤t≤T , P ). Assume that a stochastic process
X(t), 0 ≤ t ≤ T , is given by
t t
X(t) = X(0) + Ψ (s) ds + Φ(s) dWs , (2.52)
0 0
62 2 Stochastic Calculus

where X(0) is an F0 -measurable H -valued random variable, Ψ (s) is an H -valued

Fs -measurable P -a.s. Bochner-integrable process on [0, T ],
T 
Ψ (s) ds < ∞ P -a.s.,
H
0

and Φ ∈ P(KQ , H ).
Assume that a function F : [0, T ] × H → R is such that F is continuous and
its Fréchet partial derivatives Ft , Fx , Fxx are continuous and bounded on bounded
subsets of [0, T ] × H . Then the following Itô’s formula holds:

    t   
F t, X(t) = F 0, X(0) + Fx s, X(s) , Φ(s)dWs H
0
t      
+ Ft s, X(s) + Fx s, X(s) , Ψ (s) H
0
1 *    ∗ +
+ tr Fxx s, X(s) Φ(s)Q1/2 Φ(s)Q1/2 ds (2.53)
2
P -a.s. for all t ∈ [0, T ].

Proof We will first show that the general statement can be reduced to the case of
constant processes Ψ (s) = Ψ and Φ(s) = Φ, s ∈ [0, T ]. For a constant C > 0,
define the stopping time
 
  t 
τC = inf t ∈ [0, T ] : max X(t)H , Ψ (s)H ds,
0
t  

Φ(s)2 ds ≥ C
L2 (KQ ,H )
0

with the convention that the infimum of an empty set equals T .

Then, with the notation XC (t) = X(t ∧ τC ), ΨC (t) = Ψ (t)1[0,τC ] (t), and
ΦC (t) = Φ(t)1[0,τC ] (t), we have
t t
XC (t) = XC (0) + ΨC (s) ds + ΦC (s) dWs , t ∈ [0, T ].
0 0

By (2.30), it is sufficient to prove Itô’s formula for the processes stopped at τC .

Since
 T 
P ΨC (s)H ds < ∞ = 1
0

and
T 
E ΦC (s)2 ds < ∞,
L2 (KQ ,H )
0
2.3 The Itô Formula 63

by Lemma 2.4 and Corollary 2.1 it follows that ΨC and ΦC can be approximated re-
spectively by sequences of bounded elementary processes ΨC,n and ΦC,n for which
P -a.s. uniformly in t ≤ T

t 
ΨC,n (s) − ΨC (s) ds → 0
H
0

and
 
 t t 
 ΦC,n (s) dWs − ΦC (s) dWs 
  → 0.
0 0 H

Define
t t
XC,n (t) = X(0) + ΨC,n (s) ds + ΦC,n (s) dWs .
0 0

Then
 
supXC,n (t) − XC (t)H → 0
t≤T

with probability one. Assume that we have shown Itô’s formula for the process
XC,n (t), that is,

    t   
F t, XC,n (t) = F 0, X(0) + Fx s, XC,n (s) , ΦC,n (s)dWs H
0
t      
+ Ft s, XC,n (s) + Fx s, XC,n (s) , ΨC,n (s) H
0

1 *     +
1/2 ∗
+ tr Fxx s, XC,n (s) ΦC,n (s)Q1/2
ΦC,n (s)Q ds
2
(2.54)

P -a.s. for all t ∈ [0, T ]. Using the continuity of F and the continuity and local
boundedness of its partial derivatives, we will now conclude that

    t   
F t, XC (t) = F 0, X(0) + Fx s, XC (s) , ΦC (s)dWs H
0
t      
+ Ft s, XC (s) + Fx s, XC (s) , ΨC (s) H
0

1 *     +
1/2 ∗
+ tr Fxx s, XC (s) ΦC (s)Q1/2
ΦC (s)Q ds. (2.55)
2

Clearly, the LHS of (2.54) converges to the LHS of (2.55) a.s.

64 2 Stochastic Calculus

For the stochastic integrals in (2.54) and (2.55), we have

 
 t    t    2
E  Fx s, XC,n (s) , ΦC,n (s) dWs H − Fx s, XC (s) , ΦC (s) dWs H 
0 0
t      2
≤2 E ΦC∗ (s) Fx s, XC,n (s) − Fx s, XC (s) L ds
2 (KQ ,R)
0
t    2
+2 E  ΦC∗ (s) − ΦC,n
∗
Fx s, XC,n (s) L ds
2 (KQ ,R)
0
t  2      
≤2 E ΦC (s)L Fx s, XC,n (s) − Fx s, XC (s) 2 ds
2 (KQ ,H ) H
0
t      
+2 E ΦC∗ (s) − ΦC,n
∗ 2
L
Fx s, XC,n (s) 2 ds.
2 (KQ ,H ) H
0

The first integral converges to zero, since the first factor is an integrable process, and
the second factor converges to zero almost surely, so that the Lebesgue DCT applies.
The second integral is bounded by M ΦC∗ (s) − ΦC,n ∗ 2
2 (KQ ,H ) for some constant
M, so that it converges to zero, since ΦC,n (s) → ΦC in the space 2 (KQ , H ).
In conclusion, the stochastic integrals in (2.54) converge to the stochastic integral
in (2.55) in mean square, so that they converge in probability.
We now turn to the nonstochastic integrals.
The first component, involving Ft , of the nonstochastic integral in (2.54) con-
verges P -a.s. to the corresponding component in (2.55) by the continuity and local
boundedness of Ft , so that the Lebesgue DCT can be applied.
Note that, P -a.s., ΨC,nk → ΨC in L1 ([0, t], H ), and Fx is locally bounded, so
that the functions s → Fx (s, XC,nk (s)) and s → Fx (s, XC (s)) are in L∞ ([0, t], H ).
The convergence with probability one of the second component follows from the
duality argument.
To discuss the last nonstochastic integral, we use the fact that
 
ΦC,n (s) − ΦC (s) →0
 (K ,H ) 2 Q

and select a subsequence nk for which

 
ΦC,n (s) − ΦC (s) → 0,
k L 2 (KQ ,H )

and therefore, for the eigenvectors fj of Q,

 
ΦC,n (s)fj − ΦC (s)fj  → 0 (2.56)
k H

a.e. on [0, T ] × . By Exercise 2.19,

   ∗

tr Fxx s, XC,nk (s) ΦC,nk (s)QΦC,n k
(s)
 ∗   
= tr ΦC,nk
(s)Fxx s, XC,nk (s) ΦC,nk (s)Q
2.3 The Itô Formula 65

∞
    
= λj Fxx s, XC,nk (s) ΦC,nk (s)fj , ΦC,nk (s)fj H .
j =1

Since XC,nk (s) is bounded, the continuity of Fxx and (2.56) imply that
   
Fxx s, XC,nk (s) ΦC,nk (s)fj , ΦC,nk (s)fj H
   
→ Fxx s, XC (s) ΦC (s)fj , ΦC (s)fj H .

By the Lebesgue DCT (with respect to the counting measure), we get that, a.e. on
[0, T ] × ,
   ∗
    
tr Fxx s, XC,nk (s) ΦC,nk (s)QΦC,nk
(s) → tr Fxx s, XC (s) ΦC (s)QΦC∗ (s)

and the LHS is bounded by the functions

  
ηn (s) = Fxx s, XC,nk (s) L (H ) ΦC,nk 22 (KQ ,H )

that converge P -a.s. to

  
η(s) = Fxx s, XC (s) L (H ) ΦC 22 (KQ ,H ) .
t
However, by the boundedness of the second derivative of F , 0 ηn (s) ds →
t
0 η(s) ds, so that we can apply the general Lebesgue DCT, Theorem 3.4, to con-
clude the convergence with probability one of the last nonstochastic integral.1
In conclusion, possibly for a subsequence, left- and right-hand sides of (2.54)
converge in probability to the left- and right-hand sides of (2.55), so that (2.55)
holds P -a.s.
By the additivity of the integrals we can further reduce the proof to the case
where
X(t) = X(0) + Ψ t + ΦWt ,
where Ψ and Φ are F0 -measurable random variables independent of t.
Now, define
 
u(t, Wt ) = F t, X(0) + Ψ t + ΦWt ;
then the function u is of the same smoothness order as F . We will now prove Itô’s
formula for the function u.

1 This elementary proof can be replaced by the following argument. The space of trace-class op-
erators L1 (H ) can be identified with the dual space to the space of compact linear operators on
H , the duality between the two spaces is the trace operator ([68], Chap. IV, Sect. 1, Theorem 1).
Hence, as a dual separable space, it has the Radon–Nikodym property ([14], Chap. III, Sect. 3, The-
orem 1). Thus, L1 ([0, T ], L1 (H ))∗ = L∞ ([0, T ], L1 (H )∗ ) ([14], Chap. IV, Sect. 1, Theorem 1).
But L∞ ([0, T ], L1 (H )∗ ) = L∞ ([0, T ], L (H )) ([68], Chap. IV, Sect. 1, Theorem 2). Thus the
convergence of the last nonstochastic integral follows from the duality argument.
66 2 Stochastic Calculus

Let 0 = t1 < t2 < · · · < tn = t ≤ T be a partition of an interval [0, t] and de-

note Δtj = tj +1 − tj and ΔWj = Wtj +1 − Wtj . Using Taylor’s formula, there exist
random variables t¯j ∈ [tj , tj +1 ] and θj ∈ [0, 1] such that

u(t, Wt ) − u(0, 0)

n−1
* + n−1
* +
= u(tj +1 , Wtj +1 ) − u(tj , Wtj +1 ) + u(tj , Wtj +1 ) − u(tj , Wtj )
j =1 j =1


n−1
= ut (t¯j , Wtj +1 )Δtj
j =1
n−1,
 -
  1 
+ ux (tj , Wtj ), ΔWj K
+ uxx (tj , W̄j )(ΔWj ), ΔWj K
2
j =1


n−1 
n−1
 
= ut (tj , Wtj +1 )Δtj + ux (tj , Wtj ), ΔWj K
j =1 j =1

1 
n−1

+ uxx (tj , Wtj )(ΔWj ), ΔWj K
2
j =1


n−1
* +
+ ut (t¯j , Wtj +1 ) − ut (tj , Wtj +1 ) Δtj
j =1

1 *
n−1
+ 
+ uxx (tj , W̄j )(ΔWj ) − uxx (tj , Wtj )(ΔWj ) , ΔWj K , (2.57)
2
j =1

where W̄j = Wtj + θj (Wtj +1 − Wtj ).

Assuming that max{tj +1 − tj , 1 ≤ j ≤ n − 1} → 0 and using the uniform conti-
nuity of the mapping [0, T ] × [0, T ] (s, r) → ut (s, Wr ) ∈ R (Exercise 2.20) and
the continuity of the map [0, T ] t → ux (t, Wt ) ∈ K ∗ , we get

n−1 
n−1
 
ut (tj , Wtj +1 )Δtj + ux (tj , Wtj ), ΔWj K
j =1 j =1
t t 
→ ut (s, Ws ) ds + ux (s, Ws ), dWs K
P -a.s.
0 0
by Lemma 2.6. Clearly,
n−1 
* + 
 
 ut (t¯j , Wtj +1 ) − ut (tj , Wtj +1 ) Δtj 
 
j =1
 
≤ T suput (t¯j , Wtj +1 ) − ut (tj , Wtj +1 ) → 0.
j ≤n
2.3 The Itô Formula 67

In view of the continuity of the mapping [0, T ] × K (t, x) → uxx (t, x) ∈

L (K, K), utilizing the result in Exercise 2.5, we have
n−1 
* +  

 uxx (tj , W̄j ) − uxx (tj , Wtj ) (ΔWj ), ΔWj K 
 
j =1

  
n−1
 
≤ sup uxx (tj , W̄j )(ΔWj ) − uxx (tj , Wtj )(ΔWj )L (K,K) ΔWj 2 → 0
K
j ≤n−1 j =1

with probability one as n → ∞ (using arguments as in Exercise 2.20).

It remains to show that


n−1
  t * +
uxx (tj , Wtj )(ΔWj ), ΔWj K
→ tr uxx (s, Ws )Q ds (2.58)
j =1 0

in probability P .
Let 1Nj = 1{max{Wti K ≤N,i≤j }} . Then 1j is Ftj -measurable, and, using the rep-
N

resentation (2.3), we get

   

E 1Nj uxx (tj , Wtj )(ΔWj ), ΔWj K Ftj
 ∞
 1/2  
= E 1N u
j xx j (t , W tj ) λk wk (tj +1 ) − wk (tj ) fk ,
k=1
∞
  
 1/2   
λl wl (tj +1 ) − wl (tj ) fl F t
 j
l=1 K
∞
     2  
= F t
j uxx (tj , Wtj )fk , fk K wk (tj +1 ) − wk (tj )
E λk 1N j
k=1
 
= tr 1N
j uxx (tj , Wtj )Q Δtj .

In view of the above and the fact that uxx is bounded on bounded subsets of [0, T ] ×
H , we obtain
n−1 2
   N  
E 1N
j uxx (tj , Wtj )(ΔWj ), ΔWj K − tr 1j uxx (tj , Wtj )Q Δtj
j =1


n−1
  N 2
= E 1j uxx (tj , Wtj )(ΔWj ), ΔWj K
j =1
  2 
− E tr 1N
j uxx (tj , Wtj )Q (Δtj )2
68 2 Stochastic Calculus

  
n−1
 
≤ sup uxx (s, h)2 EΔWj 4K − (tr Q)2 (Δtj )2
L (H )
s≤t,hH ≤N j =1

  
n−1
=2 sup uxx (s, h)2 Q2L2 (K) (Δtj )2 → 0.
L (H )
s≤t,hH ≤N j =1

Also, as N → ∞,
n−1
  
P 1 − 1N
j uxx (tj , Wtj )(ΔWj ), ΔWj K
j =1

 
− tr 1N
j uxx (tj , Wtj )QΔtj =0
 
≤ P sup Ws  > N → 0.
s≤t

This proves (2.58). Taking the limit in (2.57), we obtain Itô’s formula for the func-
tion u(t, Wt ),
t 1  

u(t, Wt ) = u(0, 0) + ut (s, Ws ) + tr uxx (s, Ws )Q ds
0 2
t 
+ ux (s, Ws ), dWs K
. (2.59)
0

To obtain Itô’s formula for F (t, X(t)), we calculate the derivatives

     
ut (t, k) = Ft t, X(0) + Ψ t + Φk + Fx t, X(0) + Ψ t + Φk , Ψ K ,
 
ux (t, k) = Φ ∗ Fx t, X(0) + Ψ t + Φk ,
 
uxx (t, k) = Φ ∗ Fxx t, X(0) + Ψ t + Φk Φ.

Noting that (see Exercise2.19)

*    ∗ + *   +
tr Fxx s, X(s) ΦQ1/2 ΦQ1/2 = tr Φ ∗ Fxx s, X(s) ΦQ ,

we arrive at the desired result (2.53). 

Exercise 2.19 Show that, for a symmetric operator T ∈ L (H ) and Φ ∈ L (K, H ),

   
tr T ΦQΦ ∗ = tr Φ ∗ T ΦQ .

Exercise 2.20 Let f : ([0, T ] × K) → R and g : [0, T ] → K be continuous. Show

that the mapping [0, T ] × [0, T ] (s, r) → f (s, g(r)) ∈ R is uniformly continuous.
2.3 The Itô Formula 69

2.3.2 The Case of a Cylindrical Wiener Process

As in the case of a Q-Wiener process, for Φ(s) ∈ P(K, H ) and a P -a.s. bounded
H -valued Ft -adapted process Ψ (s), Φ ∗ (s)Ψ (s) ∈ P(K, R). In addition, since
∞
 ∞
 ∗  2   2  2  2
Φ (s)Ψ (s) (fj ) = Ψ (s), Φ(s)(fj ) H ≤ Ψ (s)H Φ(s)L ,
2 (KQ ,H )
j =1 j =1

the process Φ ∗ (s)Ψ (s) can be considered as being K- or K ∗ -valued, and we can
define
T  T  T
Ψ (s), Φ(s) d W̃s H
= Φ ∗ (s)Ψ (s), d W̃s K
= Φ ∗ (s)Ψ (s) d W̃s .
0 0 0

Theorem 2.10 (Itô Formula) Let H and K be real separable Hilbert spaces, and
{W̃t }0≤t≤T be a K-valued cylindrical Wiener process on a filtered probability space
(, F , {Ft }0≤t≤T , P ). Assume that a stochastic process X(t), 0 ≤ t ≤ T , is given
by
t t
X(t) = X(0) + Ψ (s) ds + Φ(s) d W̃s , (2.60)
0 0

where X(0) is an F0 -measurable H -valued random variable, Ψ (s) is an H -valued

Fs -measurable P -a.s. Bochner-integrable process on [0, T ],
T 
Ψ (s) ds < ∞, P -a.s.,
H
0

and Φ ∈ P(K, H ).
Assume that a function F : [0, T ] × H → R is such that F is continuous and
its Fréchet partial derivatives Ft , Fx , Fxx are continuous and bounded on bounded
subsets of [0, T ] × H . Then the following Itô’s formula holds:

    t   
F t, X(t) = F 0, X(0) + Fx s, X(s) , Φ(s)d W̃s H
0
t      
+ Ft s, X(s) + Fx s, X(s) , Ψ (s) H
0

1 *    ∗ +
+ tr Fxx s, X(s) Φ(s) Φ(s) ds (2.61)
2

P -a.s. for all t ∈ [0, T ].

Proof The proof is nearly identical to the proof of the Itô formula for a Q-Wiener
process, and we refer to the notation in the proof of Theorem 2.9. The reduction to
70 2 Stochastic Calculus

the processes XC (t) = X(t ∧ τC ), ΨC (t) = Ψ (t)1[0,τC ] (t), ΦC (t) = Φ(t)1[0,τC ] (t),
with
t t
XC (t) = XC (0) + ΨC (s) ds + ΦC (s) d W̃s , t ∈ [0, T ],
0 0

is possible due to (2.33).

A further reduction to bounded elementary processes ΨC,n and ΦC,n for which,
P -a.s. uniformly in t ≤ T ,

t 
ΨC,n (s) − ΨC (s) ds → 0
H
0
and
 
 t t 
 ΦC,n (s) d W̃s − ΦC (s) d W̃s 
  →0
0 0 H

is achieved using Lemma 2.4 and Corollary 2.1 with Q = IK , so that we can define

t t
XC,n (t) = X(0) + ΨC,n (s) ds + ΦC,n (s) d W̃s
0 0

and claim that

 
supXC,n (t) − XC (t)H → 0
t≤T

with probability one. Then, using the isometry property (2.32) and the arguments
in the proof of Theorem 2.9 that justify the term-by-term convergence of (2.54)
to (2.55), we can reduce the general problem to the case

X(t) = X(0) + Ψ t + Φ W̃t , (2.62)

where, recalling (2.31) and (2.12),

∞
 ∞

    ∗ 
Φ W̃t = (Φ W̃t )ei ei = W̃t Φ ei ei ∈ H (2.63)
i=1 i=1

for Φ ∈ L2 (K, H ).
From here we proceed as follows. Define
 
u(t, ξt ) = F t, X(0) + Ψ t + ξt

with ξt = Φ W̃t ∈ MT2 (H ). Similarly as in the proof of Theorem 2.9, with 0 = t1 <
t2 < · · · < tn = t ≤ T , Δtj = tj +1 − tj , and Δξj = ξtj +1 − ξtj , using Taylor’s for-
2.3 The Itô Formula 71

mula, we obtain

u(t, ξt ) − u(0, 0)

n−1 
n−1
 
= ut (tj , ξtj +1 )Δtj + ux (tj , ξtj ), Δξj H
j =1 j =1

1 
n−1

+ uxx (tj , ξtj )(Δξj ), Δξj H
2
j =1


n−1
* +
+ ut (t˜j , ξtj +1 ) − ut (tj , ξtj +1 ) Δtj
j =1

1 *
n−1
+ 
+ uxx (tj , ξ̃j )(Δξj ) − uxx (tj , ξtj )(Δξj ) , Δξj H ,
2
j =1

= S1 + S 2 + S3 + S 4 + S 5 ,

where Δξj = Φ(W̃tj +1 − W̃tj ), ξ̃j = Φ W̃tj + θj Φ(W̃tj +1 − W̃tj ), and t˜j ∈ [tj , tj +1 ],
θj ∈ [0, 1] are random variables.
Using the smoothness of the function u, we conclude that S4 and S5 converge to
zero with probability one as n → ∞ and that
t t 
S1 + S2 → ut (s, ξs ) ds + Φ ∗ ux (s, ξs ), d W̃s K
0 0
t t 
= ut (s, ξs ) ds + ux (s, ξs ), Φ d W̃s H
.
0 0

To show that


n−1
  t * +
uxx (tj , ξtj )(Δξj ), Δξj H
→ tr uxx (s, W̃s )ΦΦ ∗ ds
j =1 0

in probability P , we let 1N j = 1{max{ξti H ≤N,i≤j }} . Then 1j is Ftj -measurable, and,

N

using the representation (2.63), we get

   

E 1N j uxx (tj , ξtj )(Δξj ), Δξj H Ftj
 ∞
     
= E 1N j u xx (t j , ξ t j ) W̃tj +1 Φ ∗ ei − W̃tj Φ ∗ ei ei ,
i=1
∞
  
      
W̃tj +1 Φ ∗ el − W̃tj Φ ∗ el el F t
 j
l=1 H
72 2 Stochastic Calculus

∞
     ∗   ∗ 2  
= F t
E 1Nj uxx (tj , ξtj )ei , ei H W̃tj +1 Φ ei − W̃tj Φ ei j
i=1
 ∗

= tr 1N
j uxx (tj , ξtj )ΦΦ Δtj .

Now, to complete the proof, we can now follow the arguments in the proof of The-
orem 2.9 with ΦΦ ∗ replacing Q. 
Chapter 3
Stochastic Differential Equations

3.1 Stochastic Differential Equations and Their Solutions

Let K and H be real separable Hilbert spaces, and Wt be a K-valued Q-Wiener
process on a complete filtered probability space (, F , {Ft }t≤T , P ) with the filtra-
tion Ft satisfying the usual conditions. We consider semilinear SDEs (SSDEs for
short) on [0, T ] in H . The general form of such SSDE is

dX(t) = (AX(t) + F (t, X)) dt + B(t, X) dWt ,
(3.1)
X(0) = ξ0 .

Here, A : D(A) ⊂ H → H is the generator of a C0 -semigroup of operators

{S(t), t ≥ 0} on H . Recall from Chap. 1 that for a C0 -semigroup S(t), we have
S(t)L (H ) ≤ M exp{αt} and if M = 1, then S(t) is called a pseudo-contraction
semigroup.
The coefficients F and B are, in general, nonlinear mappings,
 
F :  × [0, T ] × C [0, T ], H → H,
 
B :  × [0, T ] × C [0, T ], H → L2 (KQ , H ).

The initial condition ξ0 is an F0 -measurable H -valued random variable.

We will study the existence and uniqueness problem under various regularity
assumptions on the coefficients of (3.1) that include:
(A1) F and B are jointly measurable, and for every 0 ≤ t ≤ T , they are measurable
with respect to the product σ -field Ft ⊗ Ct on  × C([0, T ], H ), where Ct is
a σ -field generated by cylinders with bases over [0, t].
(A2) F and B are jointly continuous.
(A3) There exists a constant such that for all x ∈ C([0, T ], H ),
       
F (ω, t, x) + B(ω, t, x) ≤ 1 + sup x(s)H
H L (K ,H )
2 Q
0≤s≤T

for ω ∈  and 0 ≤ t ≤ T .

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 73

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_3, © Springer-Verlag Berlin Heidelberg 2011
74 3 Stochastic Differential Equations

For every t ∈ [0, T ], we define the following operator θt on C([0, T ], H ):


x(s), 0 ≤ s ≤ t,
θt x(s) =
x(t), t < s ≤ T .

Assumption (A1) implies that

F (ω, t, x) = F (ω, t, x1 ) and B(ω, t, x) = B(ω, t, x1 )

if x = x1 on [0, t]. Because θt x is a Borel function of t with values in C([0, T ], H ),

F (ω, t, θt x) and B(ω, t, θt x) also are Borel functions in t. With this notation,
(3.1) can be rewritten as

dX(t) = (AX(t) + F (t, θt X)) dt + B(t, θt X) dWt ,
X(0) = ξ0 .

We will say that F and B satisfy the Lipschitz condition if

(A4) For all x, y ∈ C([0, T ], H ), ω ∈ , 0 ≤ t ≤ T , there exists K > 0 such that
   
F (ω, t, x) − F (ω, t, y) + B(ω, t, x) − B(ω, t, y)
H L2 (KQ ,H )
 
≤ K sup x(s) − y(s)H .
0≤s≤T

To simplify the notation, we will not indicate the dependence of F and B on ω

whenever this does not lead to confusion.
There exist different notions of a solution to the semilinear SDE (3.1), and we
now define strong, weak, mild, and martingale solutions.1

Definition 3.1 A stochastic process X(t) defined on a filtered probability space

(, F , {Ft }t≤T , P ) and adapted to the filtration {Ft }t≤T
(a) is a strong solution of (3.1) if
(1) X(·) ∈ C([0, T ], H );
(2) X(t, ω) ∈ D(A) dt ⊗ dP -almost everywhere;
(3) the following conditions hold:
 T 

P AX(t) dt < ∞ = 1,
H
0
 
T     
P F (t, X) + B(t, X)2 dt < ∞ = 1;
H L 2 (KQ ,H )
0

1 A weak (mild) solution is called mild (respectively mild integral) solution in [9], where also a

concept of a weakened solution is studied.

3.1 Stochastic Differential Equations and Their Solutions 75

(4) for every t ≤ T , P -a.s.,

t  t
X(t) = ξ0 + AX(s) + F (s, X) ds + B(s, X) dWs ; (3.2)
0 0

(b) is a weak solution of (3.1) (in the sense of duality) if

(1) the following conditions hold:
 T 

P X(t) dt < ∞ = 1, (3.3)
H
0
 
T     
P F (t, X) + B(t, X)2 dt < ∞ = 1; (3.4)
H L 2 (KQ ,H )
0

(2) for every h ∈ D(A∗ ) and t ≤ T , P -a.s.,

  t     
X(t), h H = ξ0 , h H + X(s), A∗ h H + F (s, X), h H ds
0
t 
+ h, B(s, X) dWs H
; (3.5)
0

(c) is a mild solution of (3.1) if

(1) conditions (3.3) and (3.4) hold;
(2) for all t ≤ T , P -a.s.,
t t
X(t) = S(t)ξ0 + S(t − s)F (s, X) ds + S(t − s)B(s, X) dWs . (3.6)
0 0

We say that a process X is a martingale solution of the equation


dX(t) = (AX(t) + F (t, X)) dt + B(t, X) dWt ,
(3.7)
X(0) = x ∈ H (deterministic),

if there exists a filtered probability space (, F , {Ft }t∈[0,T ] , P ) and, on this proba-
bility space, a Q-Wiener process Wt , relative to the filtration {Ft }t≤T , such that Xt
is a mild solution of (3.7).
Unlike the strong solution, where the filtered probability space and the Wiener
process are given, a martingale solution is a system ((, F , {Ft }t≤T , P ), W, X)
where the filtered probability space and the Wiener process are part of the solution.
If A = 0, S(t) = IH (identity on H ), we obtain the SDE

dX(t) = F (t, X) dt + B(t, X) dWt ,
(3.8)
X(0) = x ∈ H (deterministic),

and a martingale solution of (3.8) is called a weak solution (in the stochastic sense,
see [77]).
76 3 Stochastic Differential Equations

Remark 3.1 In the presence or absence of the operator A, there should be no con-
fusion between a weak solution of (3.1) in the sense of duality and a weak solution
of (3.8) in the stochastic context.
Obviously, a strong solution is a weak solution (either meaning) and a mild solu-
tion is a martingale solution.

We will first study solutions to an SDE corresponding to the deterministic ab-

stract inhomogeneous Cauchy problem (1.32),
t t
X(t) = AX(s) ds + Φ(s) dWs . (3.9)
0 0
t
The role of the deterministic convolution 0 S(t − s)f (s) ds will now be played by
the stochastic process
t
S  Φ(t) = S(t − s)Φ(s) dWs , Φ ∈ P(KQ , H ), (3.10)
0

which will be called stochastic convolution. Let  · D(A) be the graph norm on
D(A),
 1/2
hD(A) = h2H + Ah2H .
The space (D(A),  · D(A) ) is a separable Hilbert space (Exercise 1.2). If f :
T
[0, T ] → D(A) is a measurable function and 0 f (s)D(A) < ∞, then for any
t ∈ [0, T ],
t t t
f (s) ds ∈ D(A) and Af (s) ds = A f (s) ds.
0 0 0

We have the following stochastic analogue of this fact.

Proposition 3.1 Assume that A is the infinitesimal generator of a C0 -semigroup

on H and that Wt is a K-valued Q-Wiener process. If Φ(t) ∈ D(A) P -a.s. for all
t ∈ [0, T ] and
 T 

P Φ(t)2 dt < ∞ = 1,
L2 (KQ ,H )
0
 T 

P AΦ(t)2 dt < ∞ = 1,
L2 (KQ ,H )
0
T
then P ( 0 Φ(t) dWt ∈ D(A)) = 1 and

T T
A Φ(t) dWt = AΦ(t) dWt P -a.s. (3.11)
0 0
3.1 Stochastic Differential Equations and Their Solutions 77

Proof Equality (3.11) is true for bounded elementary processes in E (L (K, D(A)).
Let Φn ∈ E (L (K, D(A)) be bounded elementary processes approximating Φ as in
Lemma 2.3,
T 
Φ(t, ω) − Φn (t, ω)2  →0 as n → ∞
L2 KQ ,D(A)
0

P -a.s. and hence,

t t
Φn (s) dWs → Φ(s) dWs ,
0 0
t t t
A Φn (s) dWs = AΦn (s) dWs → AΦ(s) dWs
0 0 0

in probability as n → ∞. Now (3.11) follows since the infinitesimal generator A is

a closed operator. 

Theorem 3.1 Assume that A is an infinitesimal generator of a C0 -semigroup of

operators S(t) on H and that Wt is a K-valued Q-Wiener process.
(a) If Φ ∈ P(KQ , H ) and, for h ∈ D(A∗ ) and every 0 ≤ t ≤ T ,
 
  t
∗
 t
X(t), h H = X(s), A h H ds + Φ(s) dWs , h P -a.s., (3.12)
0 0 H

then X(t) = S  Φ(t).

(b) If Φ ∈ Λ2 (KQ , H ), then for every 0 ≤ t ≤ T , S  Φ(t) satisfies (3.12).
(c) If Φ ∈ Λ2 (KQ , H ), Φ(KQ ) ⊂ D(A), and AΦ ∈ Λ2 (KQ , H ), then S  Φ(t) is
a strong solution of (3.9).

Proof (a) The proof in [11] relies on the fact which we make a subject of Exer-
cise 3.2. Another method is presented in [9]. We choose to use an Itô formula type
of proof which is consistent with the deterministic approach (see [63]).
Assume that (3.12) holds and let
 
u(s, x) = x, S ∗ (t − s)h H ,

where h ∈ D(A∗ ) is arbitrary but fixed, x ∈ H , and 0 ≤ s ≤ t ≤ T . The problem is

to determine the differential of u(s, X(s)).
Since the adjoint semigroup S ∗ (t) is a C0 -semigroup whose infinitesimal gener-
ator is A∗ (see Theorem 1.2), we have
 
us (s, x) = x, −A∗ S ∗ (t − s)h H ,
 
ux (s, x) = ·, S ∗ (t − s)h H ,
uxx (s, x) = 0.
78 3 Stochastic Differential Equations

Let 0 = s1 ≤ s2 ≤ · · · ≤ sn = s be a partition of an interval [0, s] and denote Δsj =

sj +1 − sj and ΔXj = X(sj +1 ) − X(sj ). According to (2.57), there exist random
variables s̃j ∈ [sj , sj +1 ] such that
   
u s, X(s) − u 0, X(0)

n−1
  
n−1
   
= us sj , X(sj +1 ) Δsj + ux sj , X(sj ) , ΔXj H
j =1 j =1


n−1
*   +
+ us s̃j , X(sj +1 ) − us (sj , Wsj +1 ) Δsj . (3.13)
j =1

Due to the continuity of us (s, X(s)),


n−1
  s   s 
us sj , X(sj +1 ) Δsj → us r, X(r) dr = X(r), −A∗ S ∗ (t − r)h H dr.
j =1 0 0

We consider the second sum,


n−1
    
n−1
 ∗ 
ux sj , X(sj ) , ΔXj H = S (t − sj )h, X(sj +1 ) − X(sj ) H
j =1 j =1

n−1
 sj +1   sj +1 
∗ ∗ ∗
= X(r) dr, A S (t − sj )h + Φ(r) dWr , S (t − sj )h dr
j =1 0 H 0 H
 sj   sj  
− X(r) dr, S ∗ (t − sj )h − Φ(r) dWr , S ∗ (t − sj )h
0 H 0 H
n−1
 sj +1   sj +1  
= X(r) dr, A∗ S ∗ (t − sj )h + Φ(r) dWr , S ∗ (t − sj )h .
j =1 sj H sj H

Due to the continuity of A∗ S ∗ (t − s)h = S ∗ (t − s)A∗ h, the fist sum converges to

s 
X(r), A∗ S ∗ (t − r)h H dr.
0
s  sj +1
Denote Ms = 0 Φ(r) dWr ∈ MT2 (H ). Then, sj Φ(r) dWr = Msj +1 − Msj .
By (2.43), the second sum converges in L2 (, R) to
 s   s 
S(t − r) dMr , h = S(t − r)Φ(r) dWr , h .
0 H 0 H

The last sum in (3.13) converges to zero, since it is bounded by

    
t sup  X(sj +1 ), A∗ S ∗ (t − s̃j )h H − X(sj +1 ), A∗ S ∗ (t − sj )h H  → 0
0≤j ≤n−1
3.1 Stochastic Differential Equations and Their Solutions 79

due to the uniform continuity of S(s)h on finite intervals and commutativity of A∗

and S ∗ on the domain of A∗ . We have proved that
     
u s, X(s) − u 0, X(0) = X(s), S ∗ (t − s)h H
 s 
= S(t − r)Φ(r) dWr , h . (3.14)
0 H

For s = t, we have
 
  t
X(t), h H = S(t − r)Φ(r) dWr , h .
0 H

Since D(A∗ ) is dense in H , (a) follows.

(b) For h ∈ D(A∗ ) and k ∈ K, consider the process defined by
  ∗ 
Ψ (s, ω, t)(k) = 1{(0,t]} (s) S(t − s)Φ(s) A∗ h (k)
 
= 1{(0,t]} (s)S(t − s)Φ(s)(k), A∗ h H .

Then Ψ : [0, T ] ×  × [0, T ] → L2 (KQ , R).

For every 0 ≤ t ≤ T , Ψ (·, ·, t) is {Fs }0≤s≤T -adapted, and
T 
9Ψ 9 = Ψ (·, ·, t) dt
Λ2 (KQ ,R)
0
T T   
1/2
= E 1{(0,t]} (s) S(t − s)Φ(s) ∗ A∗ h2 ds dt
Λ2 (KQ ,R)
0 0
 1/2
 ∗  T 
 
≤ T A h H Me E
αt Φ(s)2 ds
Λ2 (KQ ,H )
0
= CΦΛ2 (KQ ,H ) < ∞,

so that the assumptions of the stochastic Fubini theorem, Theorem 2.8, are satisfied.
We obtain
 t  s
∗
t 
S  Φ(s) ds, A h = S(s − u)Φ(u) dWu , A∗ h H ds
0 H 0 0
t t 
= Ψ (u, ω, s) dWu ds
0 0
t t 
= Ψ (u, ω, s) ds dWu
0 0
t t  

∗
= 1{(0,s]}(u) S(s − u)Φ(u)(·), A h H ds dWu
0 0
t  t  
= A S(s − u)Φ(u)(·) ds , h dWu
0 u H
80 3 Stochastic Differential Equations

t   
= S(t − u)Φ(u) − Φ(u) (·), h H dWu
0
 t 

= S(t − u)Φ(u) − Φ(u) dWu , h ,
0 H
t
where we have used the fact that for x ∈ H , the integral 0 S(r)x dr ∈ D(A), and
 t 
A S(r)x dr = S(t)x − x
0

(Theorem 1.2). Thus we conclude that

 t   
∗
  t
S  Φ(s) ds, A h = S  Φ(t), h H − Φ(s) dWs , h ,
0 H 0 H

proving (b).
(c) Recall from (1.22), Chap. 1, the Yosida approximation An = ARn of A, and
let Sn (s) = esAn be the corresponding semigroups. Then part (b) implies that
t t
Sn  Φ(t) = An Sn  Φ(s) ds + Φ(s) dWs . (3.15)
0 0

Part (b) of Lemma 2.2, Chap. 2, implies that

 2
sup E Sn  Φ(t) − S  Φ(t)H → 0. (3.16)
0≤t≤T

Recall the commutativity property (1.16) from Chap. 1 that for x ∈ D(A), ARn x =
Rn Ax. In addition, ASn (t)x = Sn (t)Ax for x ∈ D(A), see Exercise 3.1. Using
Proposition 3.1, we obtain
t
An Sn  Φ(t) = ARn Sn (t − s)Φ(s) dWs
0
t
= Rn Sn (t − s)AΦ(s) dWs
0
= Rn Sn  AΦ(t).

Hence,
 
 t  2
sup E 
 An Sn  Φ(s) − AS  Φ(s) ds 

0≤t≤T 0 H
t 
≤ T 2 sup E An Sn  Φ(s) − AS  Φ(s)2 ds
H
0≤t≤T 0
T 
≤ T 2E Rn Sn  AΦ(s) − S  AΦ(s)2 ds
H
0
3.1 Stochastic Differential Equations and Their Solutions 81

T  
≤ T 2E Rn Sn  AΦ(s) − S  AΦ(s) 2 ds
H
0
T 
+ T 2E (Rn − I )S  AΦ(s)2 ds
H
0
 T 
≤C E Sn  AΦ(s) − S  AΦ(s)2 ds
H
0
T 

+E (Rn − I )S  AΦ(s)2 ds .
H
0

The first summand converges to zero by (c) of Lemma 2.2, Chap. 2.

Since Rn x → x for x ∈ H , we have
 
(Rn − I )S  AΦ(s) → 0
H

and
   
(Rn − I )S  AΦ(s)2 ≤ C1 S  AΦ(s)2
H H
with
T  T s 
E S  AΦ(s)2 ds = E S(s − u)AΦ(u)2 du ds
H L2 (KQ ,H )
0 0 0

≤ C2 AΦ2Λ2 (KQ ,H ) < ∞,

and the second summand converges to zero by the Lebesgue DCT.

Summarizing,
 
 t  2
sup E 
 An Sn  Φ(s) − AS  Φ(s) ds 
 → 0. (3.17)
0≤t≤T 0 H

Combining (3.16) and (3.17), we obtain that both terms in (3.15) converge uniformly
in mean square to the desired limits, so that (3.9) is satisfied by S  Φ(t). This
concludes the proof. 

Exercise 3.1 Show that ASn (t)x = Sn (t)Ax, for x ∈ D(A).

After the preliminary discussion concerning stochastic convolution, we turn to a

general problem of the relationship among different types of solutions to the semi-
linear SDE (3.1).

Theorem 3.2 A weak solution to (3.1) is a mild solution. Conversely, if X is a mild

solution of (3.1) and
T 
E B(t, X)2 dt < ∞,
L2 (KQ ,H )
0
82 3 Stochastic Differential Equations

then X(t) is a weak solution of (3.1). If, in addition X(t) ∈ D(A), dP ⊗ dt almost
everywhere, then X(t) is a strong solution of (3.1).

Proof The techniques for proving parts (a) and (b) of Theorem 3.1 are applicable to
a more general case. Consider the process X(t) satisfying the equation
  t     
X(t), h H = ξ0 , h H + X(s), A∗ h H + f (s), h H ds
0
t 
+ h, Φ(s) dWs H
(3.18)
0

with an adapted process f (·) ∈ L1 (, H ), Φ ∈ P(KQ , H ), and h ∈ D(A∗ ).

As in (a) of Theorem 3.1, we let
 
u(s, x) = x, S ∗ (t − s)h H ,

where h ∈ D(A∗ ) is arbitrary but fixed, x ∈ H , and 0 ≤ s ≤ t ≤ T . Then, for-

mula (3.14) takes the form
       
u s, X(s) − u 0, X(0) = X(s), S ∗ (t − s)h H − X(0), S ∗ (t)h H
 s  n−1
 sj +1 
= S(t − r)Φ(r) dWr , h + lim S ∗ (t − sj )h, f (r) dr
0 n→∞ sj
H j =1 H
 s   s 
= S(t − r)Φ(r) dWr , h + S(t − r)f (r) dr, h .
0 H 0 H

For s = t, we have
 
  t
X(t), h H = S(t)ξ0 , h H + S(t − r)Φ(r) dWr , h
0 H
 s 
+ S(t − r)f (r) dr, h .
0 H

Now it follows that X(t) is a mild solution if we substitute f (t) = F (t, X) and
Φ(t) = B(t, X) and use the fact that D(A∗ ) is dense in H .
To prove the converse statement, consider the process
t
X(t) = S(t)ξ0 + S(t − s)f (s) ds + S  Φ(t),
0

where f (t) is as in the first part, and Φ ∈ Λ2 (KQ , H ). We need to show that
 
X(t), h H = ξ0 , h H
t s 
+ S(s)ξ0 + S(s − u)f (u) du + S  Φ(s), A∗ h ds
0 0 H
3.1 Stochastic Differential Equations and Their Solutions 83

t  
 t
+ f (s), h H ds + Φ(s) dWs , h .
0 0 H

Using the result in (b) of Theorem 3.1, we have that

   
  t t
S  Φ(t), h H = S  Φ(s) ds, A∗ h + Φ(s) dWs , h .
0 H 0 H
t
Since (see Theorem 1.2) for any ξ ∈ H , 0 S(s)ξ ds ∈ D(A) and

t
A S(s)ξ ds = S(t)ξ − ξ,
0

we get
  t 
S(t)ξ0 , h H = ξ0 , h H + S(s)ξ0 , A∗ h H ds.
0

Finally, using (deterministic) Fubini’s theorem,

 t s   t t 
∗ ∗
S(s − u)f (u) du ds, A h = S(s − u)f (u) ds du, A h
0 0 H 0 u H
 t t 
= A S(s − u)f (u) ds du, h
0 u H
 t t−u 
= A S(v)f (u) dv du, h
0 0 H
 t 

= S(t − u)f (u) − f (u) du, h ,
0 H

completing the calculations.

The last statement of the theorem is now obvious. 

The following existence and uniqueness result for linear SDEs is a direct appli-
cation of Theorem 3.2.

Corollary 3.1 Let {Wt , 0 ≤ t ≤ T } be a Q-Wiener process defined on a filtered

probability space (, F , {Ft }t≤T , P ), and A be the infinitesimal generator of a
C0 -semigroup {S(t), t ≥ 0}. Assume that B ∈ L (K, H ), f (·) ∈ L1 (, H ) is an
{Ft }t≤T -adapted process, and ξ0 is an H -valued F0 -measurable random variable.
Then the linear equation

dX(t) = (AX(t) + f (t)) dt + B dWt ,
(3.19)
X(0) = ξ0 ,
84 3 Stochastic Differential Equations

has a unique weak solution given by

t t
X(t) = S(t)ξ0 + S(t − s)f (s) ds + S(t − s)B dWs , 0 ≤ t ≤ T.
0 0

Exercise 3.2 Prove that if X(t) is a weak solution of the equation


dX(t) = AX(t) dt + B dWt ,
X(0) = ξ0 ,

with B ∈ L (K, H ), then for an arbitrary function ζ (·) ∈ C 1 ([0, T ], (D(A∗ ),

 · D(A∗ ) )) and t ∈ [0, T ],

  t  t 
X(t), ζ (t) H = X(s), ζ  (s) + A∗ ζ (s) H ds + ζ (s), B dWs H
. (3.20)
0 0

Hint: Prove the result for a linearly dense subset of C 1 ([0, T ], (D(A∗ ),  · D(A∗ ) ))
consisting of functions ζ (s) = ζ0 ϕ(s), where ϕ(s) ∈ C 1 ([0, T ], R).

Exercise 3.3 Apply (3.20) to a function ζ (s) = S ∗ (t − s)ζ0 with ζ0 ∈ D(A∗ ) to

show that if X(t) is a weak solution of the linear SDE (3.19) with ξ0 = 0 and f (t) ≡
0, then, P -a.s., X(t) = S  B(t). Extend this result to a general case of ξ0 and f (t).

3.2 Solutions Under Lipschitz Conditions

We first prove the uniqueness and existence of a mild solution to (3.1) in the case
of Lipschitz-type coefficients. This result is known (see Ichikawa [32]) if the coeffi-
cients F (t, ·) and B(t, ·) depend on x ∈ C([0, T ], H ) through x(t) only. We follow
a technique extracted from the work of Gikhman and Skorokhod, [25] and extend it
from R n to H -valued processes.
Note that conditions (A3) and (A4) imply, respectively, that
  b
 b    
 F (t, x) dt  1 + sup (θt x)(s)H dt
  ≤
a H a s≤T

and
 
 b   b   
 F (t, x) − F (t, y) dt  sup  θt (x − y) (s) dt.
  ≤K
a H a s≤T

We will now state inequalities useful for proving the existence, uniqueness, and
properties of solutions to the SDE (3.1). We begin with well-known inequalities
(refer to (7.8), (7.9) in [11] and (24) in [34]).
3.2 Solutions Under Lipschitz Conditions 85

Lemma 3.1 Let Φ ∈ Λ2 (KQ , H ) and p ≥ 1. Then

  2p   2p 
 t   T 
E sup 


Φ(s) dWs  ≤ c1,p E 
 Φ(s) dWs 

0≤t≤T 0 H 0 H
 T p

≤ c2,p E Φ(s)2 ds
L2 (KQ ,H )
0
 T 

≤ c2,p T p−1 E Φ(s)2p ds (3.21)
L2 (KQ ,H )
0

with the constants

 2p
2p
c1,p = ,
2p − 1
 p 2
c2,p = p(2p − 1) (c1,p )2p .

Proof The first inequality follows from the fact that the stochastic integral is an
Lp ()-martingale and from Doob’s maximal inequality, Theorem 2.2. The third is
just Hölder’s inequality. We now prove the second. For p = 1, it is the isometry
property of the stochastic integral.
2p
Assume now that p > 1. Let F (·) =  · H : H → R. Then F is continuous, and
its partial derivatives
  2(p−1)
Fx (x) (h) = 2pxH x, h H , h ∈ H,
  2(p−2)
Fxx (x) (h, g) = 4p(p − 1)xH x, h H x, g H
2(p−1)
+ 2pxH h, g H , h, g ∈ H,

are continuous and bounded on bounded subsets of H , with

 
Fxx (x) ≤ 2p(2p − 1)xH
2(p−1)
. (3.22)
L (H ×H,R)
t
Let M(t) = 0 Φ(s) dWs . Applying Itô’s formula (2.53) to F (M(t)) and taking ex-
pectations, we obtain, using (3.22), Hölder’s inequality, and Doob’s maximal in-
equality,
 
 2p s1 *   +
  =E  ∗ 
E M(s)  tr Fxx M(u) Φ(u)QΦ (u)  du
0 2
 s 
   
≤ p(2p − 1)E M(u)2(p−1) Φ(u)2 du
H L (K ,H ) 2 Q
0
 
 2(p−1) s 
≤ p(2p − 1)E sup M(u)H Φ(u)2
L2 (KQ ,H )
du
0≤u≤s 0
86 3 Stochastic Differential Equations

.  ,  p - 1
 2p / p−1 s  p
≤ p(2p − 1) E sup M(u)H Φ(u)2
p
E L2 (KQ ,H )
du
0≤u≤s 0
, 2p - p−1 ,  p - 1
2p  2p p s  p
≤ p(2p − 1)  
E M(s) H E Φ(u)2 du .
2p − 1 L2 (KQ ,H )
0
p−1
2p
Dividing both sides by (EM(s)H ) p , we obtain
 p
 2p c2,p s 
E M(s)H ≤ E Φ(u)2
L2 (KQ ,H )
du ,
c1,p 0

and (3.21) follows. 

Corollary 3.2 Let {S(t), 0 ≤ t ≤ T } be a C0 -semigroup and p ≥ 1. For Φ ∈

Λ2 (KQ , H ) and t ∈ [0, T ],
 2p  p
 t  t 
E


S(t − s)Φ(s) dWs  ≤ Cp,α,M,T E
1 Φ(s)2
L2 (KQ ,H )
ds
0 H 0
t 
≤ Cp,α,M,T
2
E Φ(s)2p ds. (3.23)
L2 (KQ ,H )
0

1
The constants Cp,α,M,T 2
and Cp,α,M,T depend only on the indicated parameters.

Proof We define G(s) = S(t − s)Φ(s). Then, for u ∈ [0, t], we have by Lemma 3.1
 u 2p  u 2p
   

E 
S(t − s)Φ(s) dWs  = E   G(s) dWs  
0 H 0 H
 u p
c2,p 
≤ E G(s)2 ds
c1,p L 2 (KQ ,H )
0
 u p
c2,p 
= E S(t − s)Φ(s)2 ds
c1,p L2 (KQ ,H )
0
 u p
c2,p 2p 2p αT 
≤ M e E Φ(s)2 ds .
c1,p L2 (KQ ,H )
0

In particular, for u = t, we get the first inequality in (3.23), the second is the Hölder
inequality. 

We will need inequalities of Burkholder type for the process of stochastic convo-
lution. We begin with a supporting lemma [10].

Lemma 3.2 Let 0 < α ≤ 1 and p > 1 be numbers such that α > 1/p. Then, for an
arbitrary f ∈ Lp ([0, T ], H ), the function
t
Gα f (t) = (t − s)α−1 S(t − s)f (s) ds, 0 ≤ t ≤ T, (3.24)
0
3.2 Solutions Under Lipschitz Conditions 87

is continuous, and there exists a constant C > 0 such that

 
sup Gα f (t)H ≤ Cf Lp ([0,T ],H ) .
0≤t≤T

Exercise 3.4 Prove Lemma 3.2.

Hint: use the Hölder inequality to show the bound for Gα f (t). To show the con-
tinuity, start with a smooth function f ∈ C ∞ ([0, T ], H ) vanishing near t = 0 and
show that
d t d
Gα f (t) = s α−1 S(s) f (t − s) ds
dt 0 dt
is bounded on [0, T ]. For general f (t), use the fact that C ∞ ([0, T ]) → Lp ([0, T ],
H ) densely.

The following Burkholder-type inequalities concern two cases. The first allows a
general C0 -semigroup but is restricted only to the powers strictly greater than two.
Its proof relies on a factorization technique developed in [10], and it is a conse-
quence of (3.21) and (3.23). The second inequality allows the power of two but
is restricted to pseudo-contraction semigroups only. Curiously, the general case of
power two is still an open problem.

Lemma 3.3 Let Wt be a K-valued Wiener process with covariance Q, and Φ ∈

Λ2 (KQ , H ).
(a) Let S(t) be a general C0 -semigroup and p > 1. If
 T 
 
E Φ(t)2p dt < ∞,
L (K ,H ) 2 Q
0

 t there exists a continuous modification of the stochastic convolution S  Φ(t) =

then
0 S(t − s)Φ(s) dWs .
For this continuous version, there exists a constant Cp,α,M,T , depending only on
the indicated parameters, such that for any stopping time τ ,
 2p T ∧τ  
E sup S  Φ(t)H ≤ Cp,α,M,T E Φ(t)2p
L2 (KQ ,H )
dt. (3.25)
0≤t≤T ∧τ 0

Let An = ARn be the Yosida approximations, and Sn (t) = eAn t . Then a continuous
version of S  Φ(t) can be approximated by the (continuous) processes Sn  Φ(t) in
the following sense:
 2p
lim E sup S  Φ(t) − Sn  Φ(t)H = 0. (3.26)
n→∞ 0≤t≤T

(b) Let S(t) be a C0 -pseudo-contraction semigroup and p ≥ 1. If

 T p
 
E Φ(t)2 dt < ∞,
L (K ,H ) 2 Q
0
88 3 Stochastic Differential Equations

then there exists a continuous modification of the stochastic convolution S  Φ(t).

For this continuous version, there exists a constant Cp,α,T , depending only on the
indicated parameters, such that for any stopping time τ ,
 T ∧τ p
 2p  2
 
E sup S  Φ(t) H ≤ Cp,α,T E  
Φ(t) L (K ,H ) dt . (3.27)
2 Q
0≤t≤T ∧τ 0

Let Φn (t) = Rn Φ(t). Then a continuous version of S  Φ(t) can be approximated

by the (continuous) processes S  Φn (t) in the following sense:
 2p
lim E sup S  Φ(t) − S  Φn (t)H = 0. (3.28)
n→∞ 0≤t≤T

Proof (a) We follow the proof of Proposition 7.3 in [11], which uses the factor-
ization method introduced in [10]. Let us begin with the following identity (see
Exercise 3.6):
t π
(t − s)α−1 (s − σ )−α ds = , 0 < α < 1, σ < t. (3.29)
σ sin πα
Using this identity and the stochastic Fubini theorem 2.8, we obtain
t
S(t − s)Φ(s) dWs
0
sin πα t t 
= (t − s)α−1 (s − σ )−α ds S(t − σ )Φ(σ ) dWσ
π 0 σ
t  s 
sin πα
= (t − s)α−1 S(t − s) (s − σ )−α S(s − σ )Φ(σ ) dWσ ds
π 0 0

sin πα t
= (t − s)α−1 S(t − s)Y (s) ds P -a.s.
π 0

with
s
Y (s) = (s − σ )−α S(s − σ )Φ(σ ) dWσ , 0 ≤ s ≤ T.
0
Hence, we have the modification
sin πα t
S  Φ(t) = (t − s)α−1 S(t − s)Y (s) ds, (3.30)
π 0

for which we need to prove the assertions in (a).

1
Let 2p < α < 12 . Applying Hölder’s inequality to the integral in (3.30), we obtain,
1
for some constant Cp,α,M,T ,
 2p
 t 
sup 
 S(t − s)Φ(s) dWs 

0≤t≤T 0 H
3.2 Solutions Under Lipschitz Conditions 89
 2p  2p−1 T
M sin πα T 2p 
≤ e 2pαT
(t − s)
(α−1)( 2p−1 )
ds Y (s)2p ds
π H
0 0
T 
≤ Cp,α,M,T
1 Y (s)2p ds,
H
0
2
since α > 1/(2p). By Corollary 3.2, there exists a constant Cp,α,M,T > 0 such that
T  2p
E Y (s)H ds
0
T s  2
p
−2α 
≤ Cp,α,M,T
2 E (s − σ ) Φ(σ )L dσ ds
2 (KQ ,H )
0 0
 p  T 
T 
≤ Cp,α,M,T
2 σ −2α
dσ E Φ(σ )2p dσ
L2 (KQ ,H )
0 0
 T 

≤ Cp,α,M,T
3 E Φ(σ )2p dσ (3.31)
L2 (KQ ,H )
0

3
with some constant Cp,α,M,T > 0, by the theorem about convolution in Lp (Rd ), see
Exercise 3.7. Now (3.25), in case τ =T , follows with Cp,α,M,T = Cp,α,M,T
1 3
Cp,α,M,T .
We will consider (3.25) with a stopping time τ . Let τn ↑ τ P -a.s. be an increasing
sequence of stopping times approximating τ , each τn taking kn values 0 ≤ t1 ≤ · · · ≤
tkn ≤ T . Then
 s 2p
 
E sup   S(s − r)Φ(r) dWr  
0≤s≤τn ∧t 0 H
  2p 

kn
 s 
= E 1{τn =ti } sup 
 S(s − r)Φ(r) dWr 

i=1 0≤s≤ti ∧t 0 H
 2p

kn
 s 
= E sup 
 1{τn =ti } S(s − r)Φ(r) dWr 

i=1 0≤s≤ti ∧t 0 H


kn ti ∧t  2p
≤ Cp,α,M,T E 1{τ Φ(r)L (K dr
n =ti } 2 Q ,H )
i=1 0
τn ∧t  
= Cp,α,M,T E Φ(r)2p dr,
L2 (KQ ,H )
0

and (3.25) is obtained by the monotone convergence theorem.

Note that by (3.31), the process Y (t) has almost surely 2p-integrable paths, so
that by Lemma 3.2,
sin πα
S  Φ(t) =
Gα Y (t) (3.32)
π
has a continuous version. Equation (3.32) is referred to as the factorization formula.
90 3 Stochastic Differential Equations

Finally, we justify (3.26). As in (3.30),

π t
Sn  Φ(t) = (t − s)α−1 Sn (t − s)Yn (s) ds
sin πα 0

with
s
Yn (s) = (s − σ )−α Sn (s − σ )Φ(σ ) dWσ , 0 ≤ s ≤ T.
0
Hence,
π t  
S  Φ(t) − Sn  Φ(t) = (t − s)α−1 S(t − s) − Sn (t − s) Y (s) ds
sin πα 0
π t  
+ (t − s)α−1 Sn (t − s) Y (s) − Yn (s) ds
sin πα 0
= In (t) + Jn (t).

Let us analyze the terms In (t) and Jn (t) separately. By the Hölder inequality,
 2p T   
sup In (t)H ≤ C  S(t − s) − Sn (t − s) Y (s)2p ds,
H
0≤t≤T 0

with the expression on the right-hand side converging to zero and being bounded by
a P -integrable function, so that
 2p
lim E sup In (t)H = 0
n→∞ 0≤t≤T

by the Lebesgue DCT.

The expression for Jn is an integral of the type in (3.30), so that, by applying the
Hölder inequality, we obtain
 2p
sup Jn (t)H
0≤t≤T
 
 sin πα t   2p
= sup 
 π (t − s)α−1 Sn (t − s) Y (s) − Yn (s) ds 

0≤t≤T 0 H
T 
≤ Cp,α,M,T
3 Y (t) − Yn (t)2p dt.
H
0

Similarly to (3.31), using the convolution inequality in Exercise 3.7 (with r = 1 and
s = p), we have
T  2p
E Y (t) − Yn (t)H dt
0

T
2p
t   
=E  (t − s)−α S(t − s) − Sn (t − s) Φ(s) dWs 
  dt
0 0 H
3.2 Solutions Under Lipschitz Conditions 91

T t   2
p
≤ Cp4 E (t − s)−2α  S(t − s) − Sn (t − s) Φ(s)L ds dt
2 (KQ ,H )
0 0
T t   2
p
≤ Cp4 E (t − s)−2α sup  S(u) − Sn (u) Φ(s)L ds dt
2 (KQ ,H )
0 0 0≤u≤T
 p  
T T   2p
≤ Cp4 E t −2α
dt sup  S(u) − Sn (u) Φ(t)L (K dt
2 Q ,H )
0 0 0≤u≤T
T   2p
≤ Cp,α
5
E sup  S(u) − Sn (u) Φ(t)L (K dt → 0
2 Q ,H )
0 0≤u≤T

by Lemma 2.2, part (c).

(b) We follow the idea in [72] for pseudo-contraction semigroups. Let An = ARn
be the Yosida approximations of A; then Φn = Rn Φ ∈ Λ2 (KQ , H ), Φn (KQ ) ⊂
D(A) (see Chap. 1, (1.15)), and AΦn = An Φ ∈ Λ2 (KQ , H ). By Theorem 3.1(c),

Xn = S  Φn (s)

is a strong solution of the equation

t t
X(t) = AX(s) ds + Φn (s) dWs . (3.33)
0 0

2p
Applying Itô’s formula to F (x) = xH , we get, similarly as in Lemma 3.1,
  s  2(p−1) 
Xn (s)2p ≤ 2p Xn (u)H Xn (u), Φn (u) dWu H
H
0
s  2(p−1) 
+ 2p Xn (u)H X(u), AX(u) H du
0
1 s  2(p−1)  
+ 2p(2p − 1)Xn (u)H Φn (u)2
L
du.
2 2 (KQ ,H )
0

Since S(t) is a pseudo-contraction semigroup,

Ax, x H ≤ αx2H , x ∈ D(A)

(see Exercise 3.5). Thus,

 
 2p  s    
sup Xn (s)H ≤ 2p sup  Xn (u)2(p−1) Xn (u), Φn (u) dWu 
H H
0≤s≤t 0≤s≤t 0
t  2p
+ 2pα sup Xn (u)H du
0 0≤u≤s

 2(p−1) t 
+ p(2p − 1) sup Xn (s)H Φn (s)2
L2 (KQ ,H )
ds.
0≤s≤t 0
92 3 Stochastic Differential Equations

To simplify the notation, denote

  t 
Xn∗ (t) = sup Xn (s)H , and φn (t) = Φn (s)2
L2 (KQ ,H )
ds.
0≤s≤t 0

Let τk = inft≤T {Xn∗ (t) > k}, k = 1, 2, . . . , with the infimum over an empty set being
equal T . By the Burkholder inequality for real-valued martingales, we have
 
 s∧τk    
E  Xn (u)2(p−1) Xn (u), Φn (u) dWu 
H H
0
 s∧τk  1/2
  
≤E Xn (u)2(2p−1) Φn (u)2 du
H L2 (KQ ,H )
0
 2p−1  1/2 
≤ E Xn∗ (s ∧ τk ) φn (s ∧ τk ) .

Now, by Hölder’s inequality,

 2p−1  1/2 
E Xn∗ (s ∧ τk ) φn (s ∧ τk )
  2p 1−1/2p   p 1/2p
≤ E Xn∗ (s ∧ τk ) E φn (s ∧ τk )

and
 2(p−1) 
E Xn∗ (s ∧ τk ) φn (s ∧ τk )
  2p 1−1/p   p 1/p
≤ E Xn∗ (s ∧ τk ) E φn (s ∧ τk ) .

We arrive at the following estimate

 2p   2p 1−1/2p   p 1/2p
E Xn∗ (t ∧ τk ) ≤ 2p E Xn∗ (t ∧ τk ) E φn (t ∧ τk )
  2p 1−1/p   p 1/p
+ p(2p − 1) E Xn∗ (t ∧ τk ) E φn (t ∧ τk )
t 2p
+ 2pαE Xn∗ (s ∧ τk ) ds,
0

since
t∧τk  2p t 2p
Xn∗ (s) ds ≤ Xn∗ (s ∧ τk ) ds.
0 0

This is an expression of the form

t
g(t) ≤ u(t) + pα g(s) ds
0
3.2 Solutions Under Lipschitz Conditions 93

with constants p, α > 0 and an integrable function u : [0, T ] → R. Now we use

Gronwall’s lemma and the fact that sup0≤s≤t u(s) = u(t):

t
g(t) ≤ u(t) + pα u(s)epα(t−s) ds
0
t
≤ u(t) + sup u(s) pα epα(t−s) ds = u(t)epαt .
0≤s≤t 0

Multiplying the obtained inequality by (E(Xn∗ (t ∧ τk ))2p )1/(2p)−1 we can see that
  ∗ 1/2p
E Xn (t ∧ τk )2p
   2p −1/2p  1/p
≤ e2pαt p(2p − 1) E Xn∗ (t ∧ τk ) (E φn (t ∧ τk )p
  1/2p 
+ 2p E φn (t ∧ τk )p .

Let z = (E(Xn∗ (t ∧ τk ))2p )1/2p , we have

 1/p  1/2p
z2 ≤ e2pαt p(2p − 1)(E φn (t ∧ τk )p + e2pαt 2p(E φn (t ∧ τk )p z,

giving

1  1/2  1/2p
z≤ 2pe2pαt + 4p 2 e4pαt + 4p(2p − 1)e2pαt Eφn (t ∧ τk )p
2
 1/2p
≤ Cp,T e2pαt Eφn (t ∧ τk )p .

Thus, we proved that

 p
 2p t 
E sup Xn (s ∧ τk )H ≤ Cp,T e2pαt E Φn (s)2
L2 (KQ ,H )
0≤s≤t 0

(dropping the stopping time in the RHS does not decrease its value).
Since sup0≤s≤t∧τk Xn (s)H ↑ sup0≤s≤t Xn (s)H , P -a.s., as k → ∞, by the
continuity of Xn (t) as a solution of (3.33), we get by the monotone convergence
that
 t 
 2p  2 p
 
E sup Xn (s) H ≤ Cp,T e 2pαt
E  
Φn (s) L .
2
0≤s≤t 0

In conclusion, note that Φn → Φ in Λ2 (KQ , H ) by the Lebesgue DCT, so that

Xn (t) = S  Φn (t) → S  Φ(t) = X(t), in L2 (, H ), for any 0 ≤ t ≤ T . In addition,
note that
 t p
 
E Φn (s) − Φ(s)2 →0
L (K ,H ) 2 Q
0
94 3 Stochastic Differential Equations

by the Lebesgue DCT. By applying inequality (3.27) to Xn (t) − Xm (t), we deduce

that Xn is a Cauchy sequence in the norm
 1/2p
 · L2p (,C([0,T ],H )) = E sup  · 2p ,
0≤t≤T

so that
 2p
E sup Xn (t) − X̃(t) → 0, (3.34)
0≤t≤T

where X̃(t) is a continuous modification of X(t), proving simultaneously (3.27) for

S  Φ(s) and (3.28).
The argument that (3.27) holds with a stopping time τ is the same as that
for (3.25). 

Remark 3.2 Under the assumptions of Lemma 3.3, part (a), the continuous modifi-
cation of the stochastic convolution S  Φ(s) defined by (3.30) can be approximated,
as in (3.28), by the processes Xn (t) = S  Φn (t) defined in the proof of part (b). This
is because for X̃ defined in the proof of part (b),
 
P X̃(t) = S  Φ(t), 0 ≤ t ≤ T = 1

for a continuous version of S  Φ(t).

Thus, for a general C0 -semigroup S(t), p > 1, and Φ ∈ Λ2 (KQ , H ), we have
two approximations

Sn  Φ(t) → S  Φ(t),
S  Φn (t) → S  Φ(t),

both converging in L2p (, C([0, T ], H )).

Exercise 3.5 (Lumer–Phillips) Prove that if A generates a pseudo-contraction semi-

group S(t) on H , then

Ax, x H ≤ αx2H , x ∈ D(A).

Hint: this is Theorem 4.3 in [63] if α = 0. See also [78].

Exercise 3.6 Prove (3.29).

Exercise 3.7 Prove that if f, g ∈ L1 (Rd ), then f ∗ g exists almost everywhere in

Rd , f ∗ g ∈ L1 (Rd ), and

f ∗ gL1 (Rd ) ≤ f L1 (Rd ) gL1 (Rd ) .

3.2 Solutions Under Lipschitz Conditions 95

More general, prove that if f ∈ Lr (Rd ) and g ∈ Ls (Rd ) with r, s ≥ 1, 1

r + 1
s ≥ 1,
p = r + s − 1, then f ∗ g(s) exists almost everywhere in R , and
1 1 1 d

f ∗ gLp (Rd ) ≤ f Lr (Rd ) gLs (Rd ) .

For an adapted process ξ(·) ∈ C([0, T ], H ), denote

t t
I (t, ξ ) = S(t − s)F (s, ξ ) ds + S(t − s)B(s, ξ ) dWs . (3.35)
0 0

Lemma 3.4 If F (t, x) and B(t, x) satisfy conditions (A1) and (A3), S(t) is either
a pseudo-contraction semigroup and p ≥ 1 or a general C0 -semigroup and p > 1,
then, for a stopping time τ ,
 
 2p t  2p
E sup I (s, ξ )H ≤ C t + E sup ξ(u)H ds (3.36)
0≤s≤t∧τ 0 0≤u≤s∧τ

with the constant C depending only on p, M, α, T and the constant .

Proof We note that

 2p
sup I (s, ξ )H
0≤s≤t∧τ
 2p  2p 
 s   s 
≤2  S(s − u)F (u, ξ ) du  S(s − u)B(u, ξ ) dWu 
 +
2p−1
sup   .
0≤s≤t∧τ 0 H 0 H

We can find a bound for the expectation of the first term,

 2p
 s 
E sup 
 S(s − u)F (u, ξ ) du

0≤s≤t∧τ 0 H
 s 2p
  
≤ E sup CM,α,t  
1 + sup ξ(r) H du
0≤s≤t∧τ 0 0≤r≤u
 s∧τ  2p 
≤ 22p−1 ( CM,α,t )2p t 2p + t E sup sup ξ(r)H du
0≤s≤t 0 0≤r≤u
 
t  2p
≤ Cp,M,α,T , t+ E sup ξu H ds ,
0 0≤u≤s∧τ

and, using (3.25) or (3.27), a bound for the expectation of the second term,
 2p
 s  t∧τ  
E sup 
 S(s − u)B(u, ξ ) dWu 
 ≤ Cp,M,α,t E
B(s, ξ )2p
L2 (KQ ,H )
ds.
0≤s≤t∧τ 0 H 0
96 3 Stochastic Differential Equations

The latter is dominated by

t∧τ    2p
Cp,M,α,t 2p
E 1 + sup ξ(u)H ds
0 0≤u≤s
 
t∧τ  2p
≤ 22p−1 Cp,M,α,T 2p
t +E sup ξ(u)H ds
0 0≤u≤s
 
t  2p

= Cp,M,α,T t+  
E sup ξ(u) H ds .
,
0 0≤u≤s∧τ

We complete the proof by combining the inequalities for both terms. 

Lemma 3.5 Let conditions (A1) and (A4) be satisfied, and S(t) be either a pseudo-
contraction semigroup and p ≥ 1 or a general C0 -semigroup and p > 1. Then,
 2p t  2p
E sup I (s, ξ1 ) − I (s, ξ2 )H ≤ Cp,M,α,T ,K E sup ξ1 (u) − ξ2 (u)H ds
0≤s≤t 0 0≤u≤s

with the constant Cp,M,α,T ,K depending only on the indicated parameters.

Proof We begin with the following estimate:

 2p
E sup I (s, ξ1 ) − I (s, ξ2 )H
0≤s≤t
 
 s   2p
≤ 22p−1 E sup  S(s − u) F (u, ξ1 ) − F (u, ξ2 ) du
 
0≤s≤t 0 H
 2p 
 s   
+
 S(s − u) B(u, ξ1 ) − B(u, ξ2 ) dWu 
 .
0 H

Considering the two terms separately, we obtain

 
 s   2p
E sup 
 S(s − u) F (u, ξ1 ) − F (u, ξ2 ) du

0≤s≤t 0 H
s  2p
≤ M 2p e2pαt K 2p 2p−1
t E sup sup ξ1 (r) − ξ2 (r)H du
0≤s≤t 0 0≤r≤u
t  2p
= Cp,M,α,T ,K E sup ξ1 (u) − ξ2 (u)H ds,
0 0≤u≤s

by Hölder’s inequality, and similarly, using (3.25) or (3.27),

 2p
 s   
E sup 
 S(s − u) B(u, ξ1 ) − B(u, ξ2 ) dWu 

0≤s≤t 0 H
3.2 Solutions Under Lipschitz Conditions 97

t 
≤ Cp,M,α,T E B(s, ξ1 ) − B(s, ξ2 )2p ds
L2 (KQ ,H )
0
t  2p
≤ Cp,M,α,T ,K E sup ξ1 (u) − ξ2 (u)H ds.
0 0≤u≤s

This completes the proof. 

Let H2p denote the space of C([0, T ], H )-valued random variables ξ such that
the process ξ(t) is jointly measurable, adapted to the filtration {Ft }t∈[0,T ] , with
2p
Esup0≤s≤T ξ(s)H < ∞. Then H2p is a Banach space with the norm
  2p  2p1
ξ H2p = E sup ξ(s)H .
0≤s≤T

Theorem 3.3 Let the coefficients F and B satisfy conditions (A1), (A3), and (A4).
Assume that S(t) is either a pseudo-contraction semigroup and p ≥ 1 or a general
C0 -semigroup and p > 1. Then the semilinear equation (3.1) has a unique continu-
2p
ous mild solution. If, in addition, Eξ0 H < ∞, then the solution is in H2p .
2p
If A = 0, then (3.8) has unique strong solution. If, in addition, Eξ0 H < ∞,
then the solution is in H2p , p ≥ 1.

2p
Proof We first assume that Eξ0 H < ∞. Let I (t, X) be defined as in (3.35), and
consider I (X)(t) = I (t, X). Then, by Lemma 3.4, I : H2p → H2p . The solution
can be approximated by the following sequence:

X0 (t) = S(t)ξ0 ,
(3.37)
Xn+1 (t) = S(t)ξ0 + I (t, Xn ), n = 0, 1, . . . .
2p
Indeed, let vn (t) = E sup0≤s≤t Xn+1 (s) − Xn (s)H . Then v0 (t) =
2p
E sup0≤s≤t X1 (s) − X0 (s)H ≤ v0 (T ) ≡ V0 , and, using Lemma 3.5, we obtain
 2p  2p
v1 (t) = E sup X2 (s) − X1 (s)H = E sup I (s, X1 ) − I (s, X0 )H
0≤s≤t 0≤s≤t
t  2p
≤C E sup X1 (u) − X0 (u)H ds ≤ CV0 t
0 0≤u≤s

and, in general,
t V0 (Ct)n
vn (t) ≤ C vn−1 (s) ds ≤ .
0 n!
Next, similarly to the proof of Gikhman and Skorokhod in [25], we show that
 
sup Xn (t) − X(t)H → 0 a.s.
0≤t≤T
98 3 Stochastic Differential Equations

for some X ∈ H2p . If we let εn = (V0 (CT )n /n!)1/(1+2p) , then, using Chebychev’s
inequality, we arrive at
      2p 
P sup Xn+1 (t) − Xn (t)H > εn = P sup Xn+1 (t) − Xn (t)H > εn
2p
0≤t≤T 0≤t≤T
 0 
V0 (CT )n [V0 (CT )n ]2p/(1+2p)
≤ = εn .
n! n!
∞
Because n=1 εn < ∞, by the Borel–Cantelli lemma, sup0≤t≤T Xn+1 (t) −
Xn (t)H < εn P -a.s. Thus, the series
∞
  
sup Xn+1 (t) − Xn (t)H
n=10≤t≤T

converges P -a.s., showing that Xn converges to some X a.s. in C([0, T ], H ).

Moreover,
 2p  2p
E sup X(t) − Xn (t)H = E lim sup Xn+m (t) − Xn (t)H
0≤t≤T m→∞ 0≤t≤T
n+m−1 2p
   
 
= E lim sup  Xk+1 (t) − Xk (t) 
m→∞ 0≤t≤T  
k=n H
n+m−1 2p
  
≤ E lim sup Xk+1 (t) − Xk (t)H
m→∞
k=n 0≤t≤T
n+m−1 2p
   1
= lim E sup Xk+1 (t) − Xk (t)H k
m→∞ 0≤t≤T k
k=n
∞
∞ p/q
  2p 2p 
≤ E sup Xk+1 (t) − Xk (t)H k k −2q
k=n 0≤t≤T k=n

with 1/2p + 1/2q = 1. Note that  q > 1/2; 2p hence,the second series converges.
The first series is bounded by: ∞ v
k=n k (T )k ≤ ∞ k=n V0 (CT ) k /k! → 0 as
k 2p

n → ∞.
To justify that X(t) is a mild solution to (3.1), we note that, a.s., F (s, Xn ) →
F (s, X) uniformly in s. Therefore,
t t
S(t − s)F (s, Xn ) ds → S(t − s)F (s, X) ds a.s.
0 0
2p
Using the fact, proved above, that E supt X(t) − Xn (t)H → 0, we obtain
 2p
 t   
E
 S(t − s) B(s, X) − B(s, Xn ) dWs 

0 H
3.2 Solutions Under Lipschitz Conditions 99

t  
≤ Cp,M,α,T E  B(s, X) − B(s, Xn ) 2p ds
L (K Q ,H )
0
 2p
≤ Cp,M,α,T ,K E sup X(t) − Xn (t)H → 0.
0≤t≤T

2p
Now, if Eξ0 H ≤ ∞, take the F0 -measurable random variable χk = 1{ξ0 <k} and
let
ξk = ξ0 χk .
Let X k (t) be a mild solution of (3.1) with the initial condition ξk . We will first show
that
X k χk = X k+1 χk .
Let Xnk and Xnk+1 be the approximations of mild solutions X k and X k+1 defined
by (3.37). Since

X0k (t) = S(t)ξ0 χk = S(t)ξ0 χk+1 χk = X0k+1 (t)χk ,

we deduce that

X0k χk = X0k+1 χk ,
   
F t, X0k χk = F t, X0k+1 χk ,
   
B t, X0k χk = B t, X0k+1 χk ,

so that
t   t  
X1k (t)χk = S(t)ξ0 χk + χk S(t − s)F s, X0k ds + χk S(t − s)B s, X0k dWs
0 0
t  
= S(t)ξ0 χk + χk S(t − s)F s, X0k+1 ds
0
t  
+ χk S(t − s)B s, X0k+1 dWs
0
   
= S(t)ξ0 χk+1 χk + I t, X0k+1 χk
= X1k+1 (t)χk .

This, by induction, leads to Xnk χk = Xnk+1 χk . Since

Xnk → X k and Xnk+1 → X k+1

in H2p , we also have by the generalized Lebesgue DCT, Theorem 3.4, that

Xnk χk → X k χk and Xnk+1 χk → X k+1 χk

100 3 Stochastic Differential Equations

in H2p , so that P -a.s., for all t ∈ [0, T ], X k (t)χk = X k+1 (t)χk . The limit

lim X k (t) = X(t)

k→∞

exists P -a.s. and is as an element of C([0, T ], H ). X(t) satisfies (3.1), since

X(t)χk = X k (t) P -a.s. (3.38)

and X k (t) satisfies (3.1), so that

t t
X(t)χk = S(t)ξ0 χk + χk S(t − s)F (s, X) ds + χk S(t − s)B(s, X) dWs
0 0
%
and P ( k {χk = 1}) = 1.
The obtained solution is unique. If X(t), Y (t) are two solutions to (3.1), then
consider the processes X k (t) = X(t)χk and Y k (t) = Y (t)χk , k ≥ 1. We define
2p
V (t) = E sups≤t X k (s) − Y k (s)H . By Lemma 3.5,
t  2p (Ct)n
V (t) ≤ C V (s) ds ≤ · · · ≤ E sup X k (s) − Y k (s)H →0
0 0≤s≤T n!

as n → ∞, giving V (t) = 0. Consequently,

X(t)χk = X k (t) = Y k (t) = Y (t)χk P -a.s. 

We have used the following facts in the proof of Theorem 3.3.

Exercise 3.8 Let A ∈ F0 , Φ ∈ P(KQ , H ). Prove that

t t
1A Φ(s) dWs = 1A Φ(s) dWs . (3.39)
0 0

Exercise 3.9 Prove the following theorem.

Theorem 3.4 (Generalized Lebesgue DCT) Let (E, μ) be a measurable space, and
gn be a sequence of nonnegative real-valued integrable functions such that gn (x) →
g(x) for μ-a.e. x ∈ E and

gn (x)μ(dx) → g(x)μ(dx).
E E

Let fn be another sequence of functions such that |fn | ≤ gn and fn (x) → f (x) for
μ-a.e. x ∈ E, then fn and f are integrable functions, and

fn (x)μ(dx) → f (x)μ(dx).
E E
3.2 Solutions Under Lipschitz Conditions 101

The Itô formula (2.53) is applicable to calculate differentials of functions of pro-

cesses of the form (2.52) (in particular, to strong solutions), but not of the form in
which a mild solution is provided. Hence, the following proposition will be essen-
tial, since the Itô formula can be applied to strong solutions (if they exist) of the
following approximating problems, where the unbounded operator A is replaced by
its Yosida approximations An = ARn with Rn = nR(n, A) and R(n, A) being the
resolvent operator of A:

dX(t) = (An X(t) + F (t, X)) dt + B(t, X) dWt ,
(3.40)
X(0) = ξ0 .

Proposition 3.2 Let the coefficients F and B satisfy conditions (A1), (A3), and
2p
(A4) and Eξ0 H < ∞. Let X(t) be a mild solution to the semilinear equa-
tion (3.1), and Xn (t) be strong solutions of the approximating problems (3.40). If
S(t) is a pseudo-contraction semigroup and p ≥ 1 or a general C0 -semigroup and
p > 1, then the mild solution X(t) of (3.1) is approximated in H2p by the sequence
of strong solutions Xn (t) to (3.40), that is,
 2p
lim E sup Xn (t) − X(t)H = 0
n→∞ 0≤t≤T

Proof First note that under the assumption on the coefficients of (3.40), by The-
orem 3.3, strong solutions Xn exist, and they are unique and coincide with mild
solutions. Moreover,

 2p   2p
E sup Xn (s) − X(s) H ≤ C E sup  Sn (s) − S(s) ξ0 H
 
0≤s≤t 0≤s≤t
t   
+E  Sn (t − s) − S(t − s) F (s, Xn )2p ds
H
0
t   
+E  Sn (t − s) − S(t − s) B(s, Xn )2p ds
L
0

 2p
+ E sup I (s, Xn ) − I (s, X)H .
0≤s≤t

Since for x ∈ H and n → ∞, (Sn (t) − S(t))(x) → 0 uniformly in t ∈ [0, T ], the

first three summands converge to zero by the Lebesgue DCT. Thus first three terms
are bounded by ε(n) → 0 as n → ∞. t
Lemma 3.5 implies that the last summand is bounded by C 0 E sups≤t Xn (s) −
2p
X(s)H ds. By the Gronwall lemma, we deduce that
 2p
E sup Xn (s) − X(s)H ds ≤ ε(n)eC t → 0
0≤t≤T

as n → ∞. 
102 3 Stochastic Differential Equations

The following estimate on moments for mild solutions of SSDEs is available.

t t
Lemma 3.6 Let X(t) = ξ0 + 0 S(t − s)F (s, X) ds + 0 S(t − s)B(s, X) dWs with
the coefficients F and B satisfying conditions (A1) and (A3) and an F0 -measurable
H -valued random variable ξ0 . Let S(t) be either a pseudo-contraction semigroup
and p ≥ 1 or a general C0 -semigroup and p > 1. Then
 2p  2p 
E sup X(s)H < Cp,M,α, ,T 1 + Eξ0 H eCp,M,α, ,T t
0≤s≤t

with Cp,M,α, ,T depending only on the indicated constants.

Proof Let τn = inf{t : X(t)H > n}. Lemma 3.4 implies that
 2p  2p
E sup X(s ∧ τn )H = E sup S(t ∧ τn )ξ0 + I (s ∧ τn , X)H
0≤s≤t 0≤s≤t
  2p  2p 
≤ 22p−1 E sup S(s)ξ0 H + E sup I (s, X)H
0≤s≤t∧τn 0≤s≤t∧τn
  2p 
sup I (s, X)H ds
2p
≤ 22p−1 Eξ0 H Meαt + E
0≤s≤t∧τn
 
t  2p
≤C
2p
Eξ0 H +t +  
E sup X(u ∧ τn ) H ds .
0 0≤u≤s

By the Gronwall lemma, we conclude that

 2p   t 
E sup X(s ∧ τn )H ≤ C Eξ0 H + t + C
2p 2p
Eξ0 H + t eC(t−s) ds
0≤s≤t 0
 2p  Cp,M,α, ,T t
≤ Cp,M,α, ,T 1 + Eξ0 H e .
Also, because τn → ∞ a.s., we have sups≤t X(s ∧ τn )H → sups≤t X(s)H a.s.
Therefore,
 2p  2p
E supX(s) = E lim supX(s ∧ τn )
H H
s≤t n→∞ s≤t
 2p
≤ lim inf E supX(s ∧ τn )H
n→∞ s≤t
 2p 
≤ Cp,M,α, ,T 1 + Eξ0 H eCp,M,α, ,T t . 

Example 3.1 Consider a K-valued Q-Wiener process Wt with sample paths in

C([0, t], K). For B ∈ L (K, H ), the process WtB = BWt is an H -valued Wiener
process with covariance BQB ∗ , and it can be realized in C([0, T ], H ). Consider
now the equation

dX(t) = (AX(t) + F (t, X)) dt + dWtB ,
(3.41)
X(0) = x ∈ H (deterministic),
3.3 A Special Case 103

with F satisfying conditions (A1), (A2), and (A4) and A generating a C0 -semigroup
of operators on H . Then, Theorem 3.3 guarantees the existence of a unique mild
solution to (3.41) in C([0, T ], H ), which is given by
t t
X(t) = S(t)x + S(t − s)F (s, X) ds + S(t − s)B dWt .
0 0

In case F ≡ 0, this process is called an H -valued Ornstein–Uhlenbeck process.

In Sect. 3.3 we will consider a special case where the coefficients F and B of an
SSDE (3.1) depend on X(t) rather than on the entire past of the solution. It is known
that even in the deterministic case where A ≡ 0, B ≡ 0, and F (t, X) = f (t, X(t))
with a continuous function f : R × H → H , a solution to a Peano differential equa-
tion

X  (t) = f (t, X(t)),
(3.42)
X(0) = x ∈ H,
may not exist (see [26] for a counterexample), unless H is finite-dimensional. Thus
either one needs an additional assumption on A, or one has to seek a solution in a
larger space. These topics will be discussed in Sects. 3.8 and 3.9.

3.3 A Special Case

We proved the existence and uniqueness theorem for mild solutions to (3.1), Theo-
rem 3.3, for general C0 -semigroups if p > 1, and for pseudo-contraction semigroups
if p = 1. This defect is due to the fact that it is still an open problem if the maximum
inequality (3.25) is valid for p = 1 and a general strongly continuous semigroup
S(t). In this section we include the case p = 1 and a general C0 -semigroup. As-
sume that the coefficients of (3.1) depend on the value of the solution at time t alone
rather than on the entire past, so that F (ω, t, x) = F (ω, t, x(t)) and B(ω, t, x) =
B(ω, t, x(t)) for x ∈ C([0, T ], H ). We note that if F̃ (ω, t, h) = F (ω, t, x(t)) and
B̃(ω, t, h) = B(ω, t, x(t)) for x(t) ≡ h, a constant function, then conditions (A1),
(A3), and (A4) on F and B in Sect. 3.1 imply the following conditions on F̃ and B̃:
(A1’) F̃ and B̃ are jointly measurable on  × [0, T ] × H , and for every 0 ≤ t ≤ T ,
and x ∈ H , they are measurable with respect to the σ -field Ft on .
(A3’) There exists a constant such that for all x ∈ H ,
     
F̃ (ω, t, x) + B̃(ω, t, x) ≤ 1 + xH .
H L (K ,H )
2 Q

(A4’) For x1 , x2 ∈ H ,
   
F̃ (ω, t, x1 ) − F̃ (ω, t, x2 ) + B̃(ω, t, x1 ) − B̃(ω, t, x2 )
H L 2 (KQ ,H )

≤ K x1 − x2 H .
104 3 Stochastic Differential Equations

On the other hand, if we define F (ω, t, x) = F̃ (ω, t, x(t)) and B(ω, t, x) =

B̃(ω, t, x(t)), then conditions (A1’), (A3’), and (A4’) imply conditions (A1), (A3),
and (A4).
We will remain consistent with our general approach and use the notation and
conditions developed for F (t, x) = F (t, x(t)) and B(t, x) = B(t, x(t)).
Let H˜2p denote the class of H -valued stochastic processes X that are measur-
able as mappings from ([0, T ] × , B([0, T ]) ⊗ F ) to (H, B(H )), adapted to the
filtration {Ft }t≤T , and satisfying sup0≤s≤T Eξ(s)H < ∞. Then H˜2p is a Banach
2p

space with the norm

  1
2p 2p
XH˜2p = sup EX(t)H .
0≤t≤T

Let I (t, ξ ) be as in (3.35). By repeating the proofs of Lemmas 3.4 and 3.5, with all
suprema dropped and with (3.23) replacing (3.25) and (3.27), we obtain the follow-
ing inequalities for ξ1 , ξ2 ∈ H˜2p , p ≥ 1, and a general C0 -semigroup:
 
 2p t  2p
 
E I (t, ξ ) H ≤ Cp,M,α,T , t +  
E ξ(s) H ds , p ≥ 1, (3.43)
0
 2p
E I (t, ξ1 ) − I (t, ξ2 )H
t  2p
≤ Cp,M,α,T ,K E ξ1 (s) − ξ2 (s)H ds, p ≥ 1. (3.44)
0

Inequality (3.44) implies

 2p  2p
E I (t, ξ1 ) − I (t, ξ2 )H ≤ Cp,M,α,T ,K t sup E ξ1 (s) − ξ2 (s)H
0≤s≤t
2p
≤ Cp,M,α,T ,K tξ1 − ξ2  , p ≥ 1. (3.45)
H˜2p

Hence, we have the following corollary to the two lemmas.

Corollary 3.3 Let ξ, ξ1 , ξ2 ∈ H˜2p , p ≥ 1. If F (t, x) = F (t, x(t)) and B(t, x) =

B(t, x(t)) satisfy conditions (A1) and (A3), then there exists a constant Cp,M,α,T , ,
such that
   
I (·, ξ )2p ≤ Cp,M,α,T , 1 + ξ 2p . (3.46)
˜
H2p ˜ H2p

If F (t, x) = F (t, x(t)) and B(t, x) = B(t, x(t)) satisfy conditions (A1) and (A4),
then there exists a constant Cp,M,α,T ,K such that
 
I (·, ξ1 ) − I (·, ξ2 )2p ≤ Cp,M,α,T ,K ξ1 − ξ2 2p . (3.47)
˜H2p ˜ H2p

The constants depend only on the indicated parameters.

Now we can prove the existence and uniqueness result.

3.3 A Special Case 105

Theorem 3.5 Let the coefficients F (t, x) = F (t, x(t)) and B(t, x) = B(t, x(t)) sat-
isfy conditions (A1), (A3), and (A4). Assume that S(t) is a general C0 -semigroup.
Then the semilinear equation (3.1) has a unique continuous mild solution. If, in
addition, Eξ0 H < ∞, p ≥ 1, then the solution is in H˜2p .
2p

In this case, either for p > 1 and a general C0 -semigroup or for p = 1 and a
pseudo-contraction semigroup, the solution is in H2p .

Proof We follow the proof of the existence and uniqueness for deterministic
Volterra equations, which uses the Banach contraction principle. The idea is to
change a given norm on a Banach space to an equivalent norm, so that the inte-
gral transformation related to the Volterra equation becomes contractive.
We first assume that
2p
Eξ0 H < ∞.
Let B be the Banach space of processes X ∈ H˜2p , equipped with the norm
  2p  2p1
XB = sup e−Lt E X(t)H ,
0≤t≤T

where L = Cp,M,α,T ,K , the constant in Corollary 3.3. The norms ·H˜2p and ·B
are equivalent since

e−LT /2p  · H˜2p ≤  · B ≤  · H˜2p .

With I (t, ξ ) as in (3.35), define

I˜(X)(t) = S(t)ξ0 + I (t, X). (3.48)

Note that I˜ : B → B by (3.43). We will find a fixed point of the transformation I˜.
Let X, Y ∈ B. We use (3.44) and calculate
   
I˜(X) − I˜(Y )2p = sup e−Lt E I (t, X) − I (t, Y )2p
B H
0≤t≤T
t  2p
≤ sup e−Lt L E X(s) − Y (s)H ds
0≤t≤T 0
t  2p
= sup e−Lt L eLs e−Ls E X(s) − Y (s)H ds
0≤t≤T 0
t
≤ LX − Y B sup e−Lt
2p
eLs ds
0≤t≤T 0

eLt −1
= LX − Y B sup e−Lt
2p
0≤t≤T L
 
≤ 1 − e−LT X − Y B ,
2p
106 3 Stochastic Differential Equations

proving that I˜ : B → B is a contraction, and we can use Banach’s contraction prin-

ciple to find its unique fixed point.
We need to show that the fixed point is a process having a continuous modifica-
tion. The only problem is when p = 1, since for p > 1, the continuity is ensured by
Theorem 3.3 (note that we allow S(t) to be a general C0 -semigroup). Similarly as
in the proof of Theorem 3.3, let

χk = 1{ξ0 <k} and ξk = ξ0 χk ,

2p
so that Eξk H < ∞ for any p > 1. Let X k (·) ∈ H2p be mild solutions of (3.1)
with the initial condition ξk . The processes X k (t) and the limit

lim X k (t) = X̃(t)

k→∞

are continuous (see (3.38)), and X̃(t) is a mild solution of (3.1) with the initial
condition ξ0 . By the uniqueness, X̃(t) is the sought continuous modification of X(t).
2p
The proof is complete for ξ0 satisfying Eξ0 H < ∞ and p ≥ 1.
2p
If Eξ H ≤ ∞, then we apply the corresponding part of the proof of Theo-
rem 3.3.
The uniqueness is justified as in the proof of Theorem 3.3.
The final assertion of the theorem is a direct consequence of Theorem 3.3. 

Denote FtX = σ {X(s), s ≤ t} and

 
FsW,ξ0 = σ FsW ∪ σ (ξ0 ) .

2p
Remark 3.3 In case Eξ0 H < ∞, the unique mild solution X constructed in The-
orem 3.5 can be approximated in H˜2p by the sequence

X0 (t) = S(t)ξ0 ,
Xn+1 (t) = I˜(Xn )(t)
t   t  
= S(t)ξ0 + S(t − s)F s, Xn (s) ds + S(t − s)B s, Xn (s) dWs .
0 0

W,ξ
Then Xn (t) and its limit X(t) are measurable with respect to Ft 0 .
2p
If Eξ0 H ≤ ∞, the mild solution X is obtained as a P -a.e. limit of mild solu-
W,ξ
tions adapted to Ft 0 , so that it is also adapted to that filtration.

We conclude with the following corollary to Proposition 3.2. It follows by drop-

ping suprema and using (3.44) in the proof of Proposition 3.2.

Corollary 3.4 Let the coefficients F (t, x) = F (t, x(t)) and B(t, x) = B(t, x(t))
2p
satisfy conditions (A1), (A3), and (A4) and Eξ0 H < ∞. Let X(t) be a mild
3.4 Markov Property and Uniqueness 107

solution to the semilinear equation (3.1), and Xn (t) be strong solutions of the ap-
proximating problems (3.40). If S(t) is a C0 -semigroup and p ≥ 1, then the mild
solution X(t) of (3.1) is approximated in H˜2p by the sequence of strong solutions
Xn (t) to (3.40),
 2p
lim sup E Xn (t) − X(t)H = 0.
n→∞ 0≤t≤T

3.4 Markov Property and Uniqueness

We examine the Markov property of mild solutions obtained under the Lipschitz
condition on the coefficients of the SSDE (3.1) in the setting of Sect. 3.3, that is, we
assume that the coefficients of (3.1) depend on t and the value of the solution at time
t alone, so that F (t, x) = F (t, x(t)) and B(t, x) = B(t, x(t)) for x ∈ C([0, T ], H ).
In addition, let F and B not depend on ω, so that we can think of F and B as
F : [0, T ] × H → H and B : [0, T ] × H → L2 (KQ , H ).

Definition 3.2 An H -valued stochastic process {X(t), t ≥ 0} defined on a probabil-

ity space (, F , P ) is called a Markov process if it satisfies the following Markov
property:
       
E ϕ X(t + h) FtX = E ϕ X(t + h) X(t) (3.49)
for all t, h ≥ 0 and any real-valued bounded measurable function ϕ.

Exercise 3.10 Prove that if (3.49) holds true for any real-valued bounded measur-
able function ϕ, then it is also valid for any ϕ, such that ϕ(X(t + h)) ∈ L1 (, R).

We now want to consider mild solutions to (3.1) on the interval [s, T ]. To that
end, let {Wt }t≥0 , be a Q-Wiener process with respect to the filtration {Ft }t≤T and
consider W̄t = Wt+s − Ws , the increments of Wt . The process W̄t is a Q-Wiener
process with respect to F¯t = Ft+s , t ≥ 0. Its increments on [0, T − s] are identical
to the increments of Wt on [s, T ].
Consider (3.1) with W̄t replacing Wt and F¯0 = Fs replacing F0 . Under the
assumptions of Theorem 3.5, there exists a mild solution X(t) of (3.1), and it is
unique, so that for any 0 ≤ s ≤ T and an Fs -measurable random variable ξ , there
exists a unique process X(·, s; ξ ) such that
t  
X(t, s; ξ ) = S(t − s)ξ + S(t − r)F r, X(r, s; ξ ) dr
s
t  
+ S(t − r)B r, X(r, s; ξ ) dWr . (3.50)
s

Let ϕ be a real bounded measurable function on H . For x ∈ H , define

  
(Ps,t ϕ)(x) = E ϕ X(t, s; x) . (3.51)
108 3 Stochastic Differential Equations

This definition can be extended to functions ϕ such that ϕ(X(t, s; x)) ∈ L1 (, R)
for arbitrary s ≤ t. Note that for any random variable η,
  
(Ps,t ϕ)(η) = E ϕ X(t, s; x) x=η .

Due to the uniqueness of the solution, we have the following theorem.

Theorem 3.6 Let the coefficients F and B satisfy conditions (A1), (A3), and (A4).
Assume that S(t) is a general C0 -semigroup. Then, for u ≤ s ≤ t ≤ T , the solutions
X(t, u; ξ ) of (3.50) are Markov processes, i.e., they satisfy the following Markov
property:
     
E ϕ X(t, u; ξ ) FsW,ξ = (Ps,t ϕ) X(s, u; ξ ) (3.52)
for any real-valued function ϕ, such that ϕ(X(t, s; ξ )) ∈ L1 (, R) for arbitrary
s ≤ t.

Proof Using the uniqueness, we have, for 0 ≤ u ≤ s ≤ t ≤ T ,

 
X(t, u; ξ ) = X t, s; X(s, u; ξ ) P -a.e.,

so that we need to prove

      
E ϕ X t, s; X(s, u; ξ ) FsW,ξ = (Ps,t ϕ) X(s, u; ξ ) .

We will prove that

   
E ϕ X(t, s; η) FsW,ξ = Ps,t (ϕ)(η) (3.53)

for all σ (X(s, u; ξ ))-measurable random variables η. By the monotone class theo-
rem (functional form) it suffices to prove (3.53) for ϕ bounded continuous on H .
Note that if η = x ∈ H , then clearly the solution X(t, s; x) obtained in Theo-
rem 3.3 is measurable with respect σ {Wt − Ws , t ≥ s} and hence independent of
W,ξ W,ξ
Fs , by the fact that the increments Wt − Ws , t ≥ s, are independent of Fs .
This implies
   
E ϕ X(t, s; x) FsW,ξ = (Ps,t ϕ)(x). (3.54)
Consider a simple function


n
 
η= xj 1Aj X(s, u; ξ ) ,
j =1

where {Aj , j = 1, 2, . . . , n} is a measurable partition of H , and x1 , x2 , . . . , xn ∈ H .

Then, P -a.e.,

n
 
X(t, s; η) = X(t, s; xj )1Aj X(s, u; ξ )
j =1
3.4 Markov Property and Uniqueness 109

and

    n
     
E ϕ X(t, s; η) FsW,ξ = E ϕ X(t, s; xj ) 1Aj X(s, u; ξ ) FsW,ξ .
j =1

W,ξ
Now ϕ(X(t, s; xj )), j = 1, 2, . . . , n are independent of Fs , and 1Aj (X(s, u; ξ ))
W,ξ
is Fs -measurable, giving that, P -a.e.,

    n
 
E ϕ X(t, s; η) FsW,ξ = (Ps,t ϕ)(xj )1Aj X(s, u; ξ )
j =1

= (Ps,t ϕ)(η). (3.55)

If Eη2H < ∞, then there exists a sequence of simple functions ηn of the above
form such that Eηn 2 < ∞ and Eηn − η2 → 0. Lemma 3.7, in Sect. 3.5, yields
  2
E X(t, s; ηn ) − E X(t, s; η) H → 0.

By selecting a subsequence if necessary, we can assume that X(t, s; ηn ) →

X(t, s; η) P -a.e. Since ϕ is continuous and bounded, (3.55) implies that
     
E ϕ X(t, s; η) FsW,ξ = lim E ϕ(X(t, s; ηn )|FsW,ξ
n
= lim(Ps,t ϕ)(ηn ) = (Ps,t ϕ)(η). (3.56)
n

For a general η, we consider the solutions X(t, s, ηn ) with initial conditions ηn =

ηχn , where χn = 1{η<n} as we did in the final step of the proof of Theorem 3.3.
Then, X(t, s, ηn ) → X(t, s, η) P -a.e., and we can repeat the argument in (3.56) to
conclude (3.53). 

Corollary 3.5 Under the assumptions of Theorem 3.5, for u ≤ s ≤ t ≤ T , the solu-
tions X(t, u; ξ ) of (3.50) satisfy the following Markov property:
     
E ϕ X(t, u; ξ ) FsX = (Ps,t ϕ) X(s, u; ξ ) (3.57)

with FsX = σ {X(r, u; ξ ), u ≤ r ≤ s} for every real-valued function ϕ such that

ϕ(X(t, s; ξ )) ∈ L1 (, R) for arbitrary s ≤ t.

W,ξ
Proof Since the RHS of (3.52) is FsX -measurable and FsX ⊂ Fs , which is a
consequence of Remark 3.3, it is enough to take conditional expectation with respect
to FsX in (3.52). 

Exercise 3.11 Show that if X(t) = X(t, 0; ξ0 ) is a mild solution to (3.1) as in The-
orem 3.5, then the Markov property (3.52) implies (3.49).
110 3 Stochastic Differential Equations

We now consider the case where F and B are independent of t, and assume that
x ∈ H . Then we get
t+s  
X(t + s, t; x) = S(s)x + S(t + s − u)F X(u, t; x) du
t
t+s  
+ S(t + s − u)B X(u, t, x) dWu
t
s  
= S(s)x + S(s − u)F X(t + u, t; x) du
0
s  
+ S(s − u)B X(t + u, t; x) d W̄u ,
0

where W̄u = Wt+u − Wt .

Since W̄u is a Wiener process with the same distribution as Wu , the processes
X(t + s, t; x) and X(s, 0; x) are mild solutions to the same SDE (3.1) but with
different Q-Wiener processes. By the uniqueness of the solution we have that
d 
X(t + s, t; x), s ≥ 0 = X(s, 0; x), s ≥ 0 ,

i.e., the solution is a homogeneous Markov process. In particular, we get

Ps,t (ϕ) = P0,t−s (ϕ), 0 ≤ s ≤ t,

for all bounded measurable functions ϕ on H . Let us denote by

Pt = P0,t .

Note that for ϕ ∈ Cb (H ), the space of bounded continuous functions on H , Pt (ϕ) ∈

Cb (H ), due to the continuity of the solution with respect to the initial condition,
which we will prove in Lemma 3.7. This property is referred to as the Feller property
of Pt . In addition, by the Markov property,

(Pt ◦ Ps )(ϕ) = Pt+s (ϕ), (3.58)

so that Pt is a Feller semigroup. A Markov process Xt whose corresponding semi-

group has Feller property is called a Feller process.

Exercise 3.12 Show (3.58).

Denote X x (t) = X(t, 0; x). In the case of time-independent coefficients F and B

and with x ∈ H , the Markov property (3.57) takes the form
   x  
E ϕ X x (t + s) FtX = (Ps ϕ) X x (t) (3.59)
x
with FtX = σ (X x (u), u ≤ t).
In Sect. 3.5 we examine the dependence of the solution on the initial condition
in detail.
3.5 Dependence of the Solution on the Initial Value 111

3.5 Dependence of the Solution on the Initial Value

As in Sect. 3.4, we consider the semilinear SDE (3.1) with the coefficients F (t, x) =
F (t, x(t)) and B(t, x) = B(t, x(t)) for x ∈ C([0, T ], H ) such that F and B do not
depend on ω.
Before we study the dependence on the initial condition, we need the following
lemma.

Lemma 3.7 Let {S(t), t ≥ 0} be a C0 -semigroup and for ξ ∈ L2 (, H ) and X ∈

C([0, T ], H ), extend the operator I˜ defined in (3.48) to
t   t  
I˜(ξ, X)(t) = S(t)ξ + S(t − s)F s, X(s) ds + S(t − s)B s, X(s) dWs .
0 0
(3.60)
Let F and B satisfy conditions (A1), (A3), and (A4). Then, for 0 ≤ t ≤ T ,
(1) EI˜(ξ, X)(t) − I˜(η, X)(t)2H ≤ C1,T Eξ − η2H .
T
(2) EI˜(ξ, X)(t) − I˜(ξ, Y )(t)2H ≤ C2,T 0 EX(t) − Y (t)2H dt.

Proof Condition (2) follows from (3.44) in Lemma 3.5. To prove (1), we let X ξ (t)
and X η (t) be mild solutions of (3.1) with initial conditions ξ and η, respectively.
Then,
 2
E X ξ (t) − X η (t)H
  t    
≤ 3CT Eξ − ηH + E 2 F s, X ξ (s) − F s, X η (s) 2 ds
H
0
t 
   
+ B s, X ξ (s) − B s, X η (s) 2 ds
L 2 (KQ ,H )
0
 
t  2
≤ 3CT Eξ − η2H + K 2
E X ξ (s) − X η (s)H ds .
0

We now obtain the result using Gronwall’s lemma. 

We now prove the continuity of the solution with respect to the initial value.

Theorem 3.7 Let Xn be mild solutions to the sequence of stochastic semilinear

equations (3.1) with coefficients Fn , Bn and initial conditions ξn , so that the follow-
ing equations hold:
t   t  
Xn (t) = S(t)ξn + S(t − r)Fn r, Xn (r) dr + S(t − r)Bn r, Xn (r) dWr .
0 0

Assume that Fn (t, x) and Bn (t, x) satisfy conditions (A1), (A3), and (A4) of
Sect. 3.1, and in addition, let the following conditions hold:
112 3 Stochastic Differential Equations

(IV1) supn Eξn 2 < ∞.

(IV2) With n → ∞, Fn (t, x) − F0 (t, x)2H + Bn (t, x) − B0 (t, x)L2 (KQ ,H ) → 0,
and Eξn − ξ0 2H → 0.
Then Xn (t) → X0 (t) in L2 (, H ) uniformly in t.

Proof For any t ≤ T ,

 2
E Xn (t) − X0 (t)H

 2
≤ 3 S(t)(ξn − ξ0 )H
 t 2
    
+ 2E S(t − s) Fn s, Xn (s) − Fn s, X0 (s)  ds
H
0
 2
 t      
+ 2E 
 S(t − s) Bn s, Xn (s) − Bn s, X0 (s) dWs 

0 H
 
 t      2
+ 2E 
 S(t − s) Fn s, X0 (s) − F0 s, X0 (s) ds 

0 H
 t 2 
      
+ 2E 
 S(t − s) Bn s, X 0 (s) − B0 s, X 0 (s) dW 
s
0 H
 t 

≤ 3e2αt ξn − ξ0 2H + 2K 2 Xn (s) − X0 (s)2 ds + 2α (n) (t) + 2α (n) (t)
H 1 2
0

by the Lipschitz condition (A4). Now,

t     
(n)
α1 (t) = E S(t − s) Fn s, X0 (s) − F0 s, X0 (s) 2 ds
H
0
T    
≤ e2αT E Fn s, X0 (s) − F0 s, X0 (s) 2 ds.
H
0

As Fn (s, X0 (s)) → F0 (s, X0 (s)), by condition (A3) and Lemma 3.6, we have, for
all s,
  2   2   
E Fn s, X0 (s) H ≤ 2 2 E 1 + X0 (s)H ≤ C 1 + Eξ0 2H

with the constant C independent of n. Using the uniform integrability, we obtain

α1(n) → 0 uniformly in t.

Similarly,
 2
 t      
α2 (t) =  S(t − s) Bn s, X0 (s) − B0 x, X0 (s) dWs 
(n)
E 
0 H
3.5 Dependence of the Solution on the Initial Value 113

T     
≤E S(t − s) Bn s, X0 (s) − B0 s, X0 (s) 2 ds → 0
L2 (KQ ,H )
0

uniformly in t ≤ T .
We obtain the result using Gronwall’s lemma. 

We discuss differentiability of the solution with respect to the initial value in the
case where the coefficients F : [0, T ] × H → H and B : [0, T ] × H → L2 (KQ , H )
of (3.1) are Fréchet differentiable in the second (Hilbert space) variable.2

Theorem 3.8 Assume that F : [0, T ] × H → H and B : [0, T ] × H → L2 (KQ , H )

satisfy conditions (A1), (A3), and (A4).
(a) If Fréchet derivatives DF (t, ·) and DB(t, ·) are continuous in H and
bounded,
   
DF (t, x)y  + DB(t, x)y  ≤ M1 yH (3.61)
H L (K ,H ) 2 Q

for x, y ∈ H , 0 ≤ t ≤ T , with the constant M1 ≥ 0, then I˜ : H × H˜2 → H˜2 is

continuously Fréchet differentiable, and its partial derivatives are
 ˜ 
∂ I (x, ξ )
y (t) = S(t)y,
∂x
 ˜ 
∂ I (x, ξ ) t  
η (t) = S(t − s)DF s, ξ(s) η(s) ds (3.62)
∂ξ 0
t  
+ S(t − s)DB s, ξ(s) η(s) dWs
0

P -a.s., with ξ, η ∈ H˜2 , x, y ∈ H , 0 ≤ t ≤ T .

(b) If in addition, second-order Fréchet derivatives D 2 F (t, ·) and D 2 B(t, ·) are
continuous in H and bounded,
 2   
D F (t, x)(y, z) + D 2 B(t, x)(y, z) ≤ M2 yH zH (3.63)
H L (K ,H ) 2 Q

for x, y, z ∈ H , 0 ≤ t ≤ T , with the constant M2 ≥ 0, then I˜ : H × H˜2 → H˜2 is

twice continuously Fréchet differentiable, and its second partial derivative is
 
∂ 2 I˜(x, ξ )
(x, ξ )(η, ζ ) (t)
∂ξ 2

2A reader interested in the theory of reaction–diffusion systems in a bounded domain O of Rd ,

perturbed by a Gaussian random field, and the related stochastic evolution equations in an infinite-
dimensional Hilbert space H , is referred to the work of Cerrai [8]. The author considers the non-
linear case with the coefficient F defined on H = L2 (O, Rd ), which is (necessarily) not Fréchet
differentiable but can be assumed Gateaux differentiable.
114 3 Stochastic Differential Equations

t   
= S(t − s)D 2 F s, ξ(s) η(s), ζ (s) ds
0
t   
+ S(t − s)D 2 B s, ξ(s) η(s), ζ (s) dWs (3.64)
0

P -a.s., with ξ, η, ζ ∈ H˜2 , x ∈ H , 0 ≤ t ≤ T .

Proof Consider

I˜(x + h, ξ )(t) − I˜(x, ξ )(t) − S(t)h S(t)(x + h) − S(t)x − S(t)h

= = 0,
hH hH

proving the first equality in (3.62). To prove the second equality, let

rF (t, x, h) = F (t, x + h) − F (t, x) − DF (t, x)h,

rB (t, x, h) = B(t, x + h) − B(t, x) − DB(t, x)h.

By Exercise 3.13,
   
rF (t, x, h) ≤ 2M1 hH and rB (t, x, h) ≤ 2M1 hH .
H L2 (KQ ,H )

∂ I˜(x,ξ )
Now with ∂ξ as defined in (3.62), we have

 ˜ 
˜ ˜ ∂ I (x, ξ )
rI˜ (x, ξ, η)(t) = I (x, ξ + η)(t) − I (x, ξ )(t) − η (t)
∂ξ
t   t  
= S(t − s)rF s, ξ(s), η(s) ds + S(t − s)rB s, ξ(s), η(s) dWs
0 0
= I1 + I2 .

We need to show that, as ηH˜2 → 0,

rI˜ (x, ξ, η)H˜2

→ 0.
ηH˜2

Consider
  t 2 1/2
sup0≤t≤T E  0 S(t − s)rF (s, ξ(s), η(s)) ds H
ηH˜2
 1/2
T rF (s, ξ(s), η(s))2H η(s)2H
≤C E 1{η(s)H =0} ds .
0 η(s)2H η2
H˜2
3.5 Dependence of the Solution on the Initial Value 115

Since F is Fréchet differentiable, the factor

rF (s, ξ(s), η(s))2H

→ 0,
η(s)2H

and, as noted earlier, it is bounded by 4M12 . In addition, the factor

η(s)2H
1{η(s)H =0} ≤ 1.
η2
H˜2

Consequently, by the Lebesgue DCT, I1 H˜2 /ηH˜2 → 0 as ηH˜2 → 0.

Because

 t 1/2
 
I2 H˜2 = sup E S(t − s)rB s, ξ(s), η(s) 2 ds ,
L 2 (KQ ,H )
0≤t≤T 0

we obtain that I2 H˜2 /ηH˜2 → 0 as ηH˜2 → 0, similarly as for I1 .

This concludes the proof of part (a), and the proof of part (b) can be carried out
using similar arguments. 

Exercise 3.13 Let H1 , H2 be two Hilbert spaces. For a Fréchet differentiable func-
tion F : H1 → H2 , define rF (x, h) = F (x + h) − F (x) − DF (x)h. Show that

   
rF (x, h) ≤ 2 sup DF (x)L (H hH1 .
H2 1 ,H2 )
x∈H1

We will use the following lemma on contractions depending on a parameter.

Lemma 3.8 Let X, U be Banach spaces, and f : X × U → U be a contraction with

respect to the second variable, i.e., for some 0 ≤ α < 1,

 
f (x, u) − f (x, v) ≤ αu − vU , x ∈ X, u, v ∈ V (3.65)
U

and let for every x ∈ X, ϕ(x) denote the unique fixed point of the contraction
f (x, ·) : U → U . Then the unique transformation ϕ : X → U defined by

 
f x, ϕ(x) = ϕ(x) for every x ∈ X (3.66)
116 3 Stochastic Differential Equations

is of class C k (X) whenever f ∈ C k (X × U ), k = 0, 1, . . . . The derivatives of ϕ can

be calculated using the chain rule; in particular,
, -−1  
∂f (x, ϕ(x)) ∂f (x, ϕ(x))
Dϕ(x)y = −I y ,
∂u ∂x
, -−1  2
∂f (x, ϕ(x)) ∂ f (x, ϕ(x))
D ϕ(x)(y, z) =
2 −I (y, z)
∂u ∂x 2
(3.67)
∂ 2 f (x, ϕ(x))   ∂ 2 f (x, ϕ(x))  
+ Dϕ(x)y, z + y, Dϕ(x)z
∂x∂u ∂u∂x

∂ 2 f (x, ϕ(x))  
+ 2
Dϕ(x)y, Dϕ(x)z .
∂ u

Let {fn }∞
n=1 be a sequence of mappings in C (X × U ) satisfying condition (3.65),
l

denote by ϕn : X → U the unique transformations satisfying condition (3.66), and

assume that for all x, x1 , . . . , xk ∈ X, u, u1 , . . . , uj ∈ U , 0 ≤ k + j ≤ l,

∂ k+j fn (x, u) ∂ k+j f (x, u)

lim k j
(u1 , . . . , uj , x1 , . . . , xk ) = (u1 , . . . , uj , x1 , . . . , xk ).
n→∞ ∂x ∂u ∂x k ∂uj
(3.68)
Then
lim D l ϕn (x)(x1 , . . . , xl ) = D l ϕ(x)(x1 , . . . , xl ). (3.69)
n→∞

Proof Let F (x, u) = u − f (x, u). Then F (x, u) = 0 generates the implicit func-
tion ϕ(x) defined in (3.66). In addition, Fu (x, u) = I − fu (x, u) is invertible, since
fu (x, u)L (U ) ≤ α < 1. The differentiability and the form of the derivatives of
ϕ(x) follow from the implicit function theorem (see VI.2 in [44]). The last state-
ment follows from the convergence in (3.68) and the form of the derivatives of ϕ(x)
given in (3.67). 

We are now ready to prove a result on differentiability of the solution with respect
to the initial condition.

Theorem 3.9 Assume that F : [0, T ] × H → H and B : [0, T ] × H → L2 (KQ , H )

satisfy conditions (A1), (A3), and (A4). Let Fréchet derivatives DF (t, ·), DB(t, ·),
D 2 F (t, ·), and D 2 B(t, ·) be continuous in H and satisfy conditions (3.61)
and (3.63). Then the solution X x of (3.1) with initial condition x ∈ H , viewed
as a mapping X · : H → H˜2 , is twice continuously differentiable in x and for any
y, z ∈ H , the first and second derivative processes DX x (·)y and D 2 X x (·)(y, z) are
mild solutions of the equations

dZ(t) = (AZ(t) + DF (t, X x (t))Z(t)) dt + DB(t, X x (t))Z(t) dWt ,
(3.70)
Z(0) = y,
3.6 Kolmogorov’s Backward Equation 117

and
⎧
⎪
⎨ dZ(t) = (AZ(t) + DF (t, X (t))Z(t) + D F (t, X (t))(DX (t)y, DX (t)z)) dt
x 2 x x x

+ (DB(X x (t))Z(t) + D 2 B(t, X x (t))(DX x (t)y, DX x (t)z)) dWt ,

⎪
⎩
Z(0) = 0.
(3.71)
If Xn is the solution to (3.40) with deterministic initial condition x ∈ H , then for
y, z ∈ H , we have the following approximations for the first and second derivative
processes:
  
lim  DXnx (·) − DX x (·) y H˜ = 0,
n→∞ 2
 2 x   (3.72)
lim  D Xn (·) − D X (·) (y, z)H˜ = 0.
2 x
n→∞ 2

Proof Consider the operator I˜ : B → B with I˜ defined in (3.60) and the Banach
space B defined in the proof of Theorem 3.5 in the case p = 1. Since B is just H˜2
renormed with the norm  · B that is equivalent to  · H˜2 , we can as well prove
the theorem in B, and the result will remain valid in H˜2 .
Since I˜ is a contraction on B, as shown in the proof of Theorem 3.5, the solution
X of (3.1) is the unique fixed point in B of the transformation I˜, so that
x

 
X x = I˜ X x . (3.73)

By Theorem 3.8, I˜ ∈ C 2 (H ×B) satisfies the conditions of Lemma 3.8, so that X x ∈

C 2 (H ) and formulas (3.70) and (3.71) follow from the chain rule and from (3.62)
and (3.64). The last part of the assertion follows from the approximation (3.69). 

3.6 Kolmogorov’s Backward Equation

We put an additional restriction on the coefficients in this section and assume that
F and B depend only on x ∈ H . We will now discuss analytical properties of the
transition semigroup Pt . Recall that for a bounded measurable function ϕ on H ,
  
Pt ϕ(x) = P0,t ϕ(x) = E ϕ X x (t) ,

where X x (t) = X(t, 0; x) is a solution of (3.50) with deterministic initial condi-

tion ξ0 = x ∈ H and s = 0, or simply, a solution of (3.1) with ξ0 = x. The smooth
dependence of the solution with respect to the initial condition results in smooth
dependence of the function

u(t, x) = Pt ϕ(x)
118 3 Stochastic Differential Equations

on t and x, and formulas (3.62) and (3.64) allow to establish a specific form of a
parabolic-type PDE for u(t, x), which is called Kolmogorov’s backward equation,
⎧  
⎪
⎪ ∂u(t, x) ∂u(t, x)
⎪
⎪ = Ax + F (x),
⎪
⎪ ∂t ∂x
⎪
⎪  2 H

⎨ 1 ∂ u(t, x)   
1/2 ∗
+ tr B(x)Q 1/2
B(x)Q , (3.74)
⎪
⎪ 2 ∂x 2
⎪
⎪
⎪
⎪ 0 < t < T , x ∈ D(A),
⎪
⎪
⎩u(0, x) = ϕ(x).

We follow the presentation in [11] and begin with the case where A is a bounded
linear operator on H .

Theorem 3.10 (Kolmogorov’s Backward Equation I) Assume that A ∈ L (H ) and

F , B do not depend on t, F : H → H , and B : H → L2 (KQ , H ). Let Fréchet
derivatives DF (x), DB(x), D 2 F (x), and D 2 B(x) be continuous and satisfy condi-
tions (3.61) and (3.63) (with t omitted). If conditions (A1), (A3), and (A4) hold, then
for ϕ ∈ Cb2 (H ), there exists a unique solution u to Kolmogorov’s backward equa-
tion, satisfying (3.74) on [0, T [3 such that u(t, ·) ∈ Cb2 (H ) and u(·, x) ∈ Cb1 ([0, T [).
The solution is given by
  
u(t, x) = Pt ϕ(x) = E ϕ X x (t) , 0 ≤ t ≤ T , x ∈ H, (3.75)

where X x (t) is the solution to (3.1) with deterministic initial condition ξ0 = x ∈ H .

Proof We first show that u(t, x) defined by (3.75) satisfies (3.74). Since the operator
A is bounded, the proof follows from the Itô formula applied to the function ϕ(x)
and the strong solution X x (t) of the SSDE (3.1),
 
 x  dϕ(X x (s))  x 
dϕ X (t) = , AX (t) + F X (t)
x
dx H
 2 
1 d ϕ(X (t))   x  1/2   x  1/2 ∗
x
+ tr B X (t) Q B X (t) Q dt
2 dx 2
 
dϕ(X x (s))  x 
+ , B X (t) dWt .
dx H

Let u(t, x) = Eϕ(X x (t)). Then by the Lebesgue DCT,

∂ + u(0, x) u(t, x) − ϕ(x)
= lim
∂t t→0 + t
  
1 t  x  dϕ X x (s)
= E lim AX (s) + F X (s) ,
x
ds
t→0+ t 0 dx H

3 Such solution is called a strict solution.

3.6 Kolmogorov’s Backward Equation 119
 2  x  
1 1 t d ϕ X (s)   x  1/2   x  1/2 ∗
+ E lim tr B X (s) Q B X (s) Q ds
2 t→0+ t 0 dx 2
   2 
dϕ(x) 1 d ϕ(x)   
1/2 ∗
= Ax + F (x), + tr B(x)Q1/2
B(x)Q . (3.76)
dx H 2 dx 2

Now Theorem 3.9 and the fact that ϕ ∈ Cb2 (H ) imply that u(t, x) is twice Fréchet
differentiable in x for 0 ≤ t ≤ T , and for y, z ∈ H , we have
   
∂u(t, x) ∂ϕ(X x (t))
,y = E , DX x (t)y ,
∂x H ∂x H
 2   2 
∂ u(t, x) ∂ ϕ(X x (t))
y, z = E x x
DX (t)y, DX (t)z (3.77)
∂x 2 H ∂x 2 H
 
∂ϕ(X x (t)) 2 x
+E , D X (t)(y, z) .
∂x H

But DX x (0) = I and D 2 X x (0) = 0, so that

   
∂u(0, x) ∂ϕ(x)
,y = ,y ,
∂x H ∂x H
 2   2 
∂ u(0, x) ∂ ϕ(x)
y, z = y, z ,
∂x 2 H ∂x 2 H

giving for x ∈ H ,
 
∂ + u(0, x) ∂u(0, x)
= Ax + F (x),
∂t ∂x H
 2 
1 ∂ u(0, x)   
1/2 ∗
+ tr B(x)Q1/2
B(x)Q .
2 ∂x 2
By (3.58), u(t + s, x) = u(t, u(s, x)). Hence,
∂ + u(s, x) ∂ + u(0, u(s, x)) Pt (Ps ϕ)(x) − (Ps ϕ)(x)
= = lim .
∂t ∂t t→0 + t

Note that (Ps ϕ)(x) = u(s, x) ∈ Cb2 (H ), so that we can repeat the calculations
in (3.76) with (Ps ϕ)(x) replacing ϕ(x) to arrive at
 
∂ + u(s, x) ∂u(s, x)
= Ax + F (x),
∂t ∂x H
 2 
1 ∂ u(s, x)   
1/2 ∗
+ tr B(x)Q1/2
B(x)Q
2 ∂x 2

for x ∈ H , 0 ≤ s < T . Note that the functions ∂u(s, x)/∂x and ∂ 2 u(s, x)/∂x 2 are
continuous in t, because they depend on derivatives of ϕ and the derivative processes
120 3 Stochastic Differential Equations

DX x (t)y and D 2 X x (t)(y, z) that are mild solutions of (3.70) and (3.71). Hence,
∂ + u(s, x)/∂t is continuous on [0, T [, which implies ([63], Chap. 2, Corollary 1.2)
that u(·, x) is continuously differentiable on [0, T [. Thus, u(t, x) satisfies (3.74) on
[0, T [.
To prove the uniqueness, assume that ũ(t, ·) ∈ Cb2 (H ), ũ(·, x) ∈ Cb1 ([0, T [), and
ũ(t, x) satisfies (3.74) on [0, T [. For a fixed 0 < t < T , we use Itô’s formula and,
for 0 < s < t, consider the stochastic differential
 
d ũ t − s, X x (s)
  
∂ ũ(t − s, X x (s)) ∂ ũ(t − s, X x (s))  x 
= − + , AX (s) + F X (s)
x
∂t ∂x H
 2 
1 ∂ ũ(t − s, X (s))   x  1/2   x  1/2 ∗
x
+ tr B X (s) Q B X (s) Q ds
2 ∂x 2
 
∂ ũ(t − s, X x (s))  x 
+ , B X (s) dWs .
∂x H

Since ũ(t, x) satisfies (3.74), we get

    t  ∂ ũ(t − s, X x (s))  x 

ũ 0, X(t) = ũ t, X(0) + , B X (s) dWs .
0 ∂x H

Therefore, applying expectation to both sides and using the initial condition
ũ(0, x) = ϕ(x) yields
  
ũ(t, x) = Eϕ X(t) .

Using Theorem 3.10 and the Yosida approximation, Da Prato and Zabczyk stated a
more general result when the operator A is unbounded.

Theorem 3.11 (Kolmogorov’s Backward Equation II) Assume that F and B do not
depend on t, F : H → H , and B : H → L2 (KQ , H ). Let the Fréchet derivatives
DF (x), DB(x), D 2 F (x), and D 2 B(x) be continuous and satisfy conditions (3.61)
and (3.63) (with t omitted). If conditions (A1), (A3), and (A4) hold, then for ϕ ∈
Cb2 (H ), there exists a unique solution u of Kolmogorov’s backward equation (3.74)
satisfying (3.74) on [0, T [ and such that
(1) u(t, x) is jointly continuous and bounded on [0, T ] × H ,
(2) u(t, ·) ∈ Cb2 (H ), 0 ≤ t < T ,
(3) u(·, x) ∈ Cb1 ([0, T [) for any x ∈ D(A).
Moreover, u is given by formula (3.75), where X x (t) is the solution to (3.1) with
deterministic initial condition ξ0 = x ∈ H .

Proof To prove that u(x, t) = Eϕ(X x (t)) is a solution, we approximate it with the
sequence un (t, x) = Eϕ(Xnx (t)), where Xn (t) are strong solutions to (3.40) with
3.6 Kolmogorov’s Backward Equation 121

the linear terms An being the Yosida approximations of A. By Corollary 3.4 with
p = 1, we know that the mild solution X(t) of (3.1) is approximated in H˜2p by the
sequence Xnx (t), i.e.,
 2
lim sup E Xnx (t) − X x (t)H = 0.
n→∞ 0≤t≤T

This implies, choosing subsequence if necessary, that Xnx (t) → X x (t) a.s., so that,
by the boundedness of ϕ,
   
un (t, x) = Eϕ Xnx (t) → Eϕ X x (t) = u(t, x). (3.78)

By Theorem 3.10 we have

⎧  
⎪
⎪ ∂un (t, x) ∂un (t, x)
⎪
⎪ = An x + F (x),
⎪
⎪ ∂t ∂x
⎪
⎪  2
H

⎨
1 ∂ un (t, x)   
1/2 ∗
+ tr B(x)Q 1/2
B(x)Q ,
⎪
⎪ 2 ∂x 2
⎪
⎪
⎪
⎪ 0 < t < T , x ∈ H,
⎪
⎪
⎩u (0, x) = ϕ(x).
n

By (3.72) we have, choosing a subsequence if necessary, that for y, z ∈ H ,

 
lim DXnx (t)y − DX x (t)y H = 0,
n→∞
 
lim D 2 Xnx (t)(y, z) − D 2 X x (t)(y, z)H = 0,
n→∞

uniformly in [0, T ]. Consequently, using the boundedness of dϕ(x)/dx,

   
∂un (t, x) dϕ(Xnx (t, x)
,y = E x
, DXn (t)y
∂x H dx H
 x   
dϕ(X (t, x) ∂u(t, x)
→E , DX (t)y =
x
,y ,
dx H ∂x H

with the last equality following by direct differentiation of Eϕ(X x (t)) under the
expectation. Hence,
   
∂un (t, x) ∂u(t, x)
, An x + F (x) → , Ax + F (x) . (3.79)
∂x H ∂x H

Next consider
 2 
∂ un (t, x)  1/2
 
1/2 ∗
tr B(x)Q B(x)Q
∂x 2
∞
  2 
d ϕ(Xnx (t))   
1/2 ∗
= E DX x
n (t) B(x)Q 1/2
B(x)Q e k , DX x
n (t)e k
dx 2 H
k=1
122 3 Stochastic Differential Equations

∞
  
dϕ(Xnx (t)) 2 x   
1/2 ∗
+ E , D Xn (t) B(x)Q1/2
B(x)Q ek , ek . (3.80)
dx H
k=1

Assume that ek (x) are the eigenvectors of (B(x)Q1/2 )(B(x)Q1/2 )∗ . Let us discuss
the convergence of the first term. We have

∞
  2 
 d ϕ(Xnx (t))   
1/2 ∗

E x
DXn (t) B(x)Q1/2
B(x)Q x
ek , DXn (t)ek
dx 2 H
k=1
  
d2 ϕ(X x (t))   ∗ 
− DX x (t) B(x)Q1/2 B(x)Q1/2 ek , DX x (t)ek 
dx 2 
H
 2
  d ϕ(X x (t)) d2 ϕ(X x (t)) 
∞
 
≤ E  n
− DXnx (t) B(x)Q1/2
dx 2 dx 2
k=1
 
 ∗ 
B(x)Q1/2 ek , DXnx (t)ek 
H
∞  2 
  d ϕ(X x (t))   
 1/2 ∗
+ E x
DXn (t) B(x)Q1/2
B(x)Q x
ek , DXn (t)ek
dx 2 H
k=1
  
d2 ϕ(X x (t))   ∗ 
− DX x (t) B(x)Q1/2 B(x)Q1/2 ek , DX x (t)ek 
dx 2 
H
= S1 + S 2 .

Now note that

 
sup sup DXnx y H˜ < ∞, p ≥ 1. (3.81)
2p
n yH ≤1

Indeed, DXnx (t)y are mild solutions to (3.70) whose coefficients satisfy the assump-
tions of Theorem 3.5, since we have assumed that the Fréchet derivatives of F
and B satisfy conditions (3.61). By Theorem 3.5 each solution can be obtained in
H˜2p , p ≥ 1 (the initial condition is deterministic), using the iterative procedure that
employs the Banach contraction principle and starting from the unit ball centered
at 0. In addition, the sequence of contraction constants can be bounded by a con-
stant strictly less than one, since eAn t L (H ) ≤ Menαt/(n−α) by estimate (1.29) in
Chap. 2.
Consider the series S1 . For each k, the sequence
 2 
 d ϕ(Xnx (t)) d2 ϕ(X x (t))  

E − DXnx (t) B(x)Q1/2
dx 2 dx 2
 
 ∗ 
B(x)Q1/2 ek , DXnx (t)ek 
H
3.6 Kolmogorov’s Backward Equation 123
 2 
 d ϕ(Xnx (t)) d2 ϕ(X x (t))   
≤ E
 2
− 2
 λk DX x (t)ek 2 → 0,
 n H
dx dx H

where λk (x) is the eigenvalue corresponding to the eigenvector ek (x). The sequence
converges to zero since, by (3.81) with p = 1, we take a scalar product in L2 ()
of two sequences, one converging to zero and one bounded. As functions of k, the
expectations are bounded by Cλk . We conclude that S1 → 0 as n → ∞.
Consider S2 :
∞
  2 
 d ϕ(X x (t))   
1/2 ∗
E  DX x
n (t) B(x)Q1/2
B(x)Q e k , DX x
n (t)e k
dx 2 H
k=1
  
d2 ϕ(X x (t))   
1/2 ∗

− DX x
(t) B(x)Q1/2
B(x)Q e , DX x
(t)e 
dx 2 k k 
H
∞  2
  d ϕ(X x (t))   ∗
≤ E  2
DXnx (t) B(x)Q1/2 B(x)Q1/2 ek
dx
k=1
 
  ∗ 
− DX (t) B(x)Q1/2 B(x)Q1/2 ek , DX x (t)ek
x 

H
∞  2
  d ϕ(X x (t))  
+ E  2
DXnx (t) B(x)Q1/2
dx
k=1
 
 ∗ 
B(x)Q1/2 ek , DX x (t)ek − DXnx (t)ek ,

H

so that the convergence of S2 to zero follows by similar arguments as above.

Finally, the second term in (3.80) is bounded by
  ∞
 dϕ(x)      ∗  

sup   E  D 2 Xnx (t) − D 2 X x (t) B(x)Q1/2 B(x)Q1/2 ek , ek H H
x∈H dx  H k=1
 
 dϕ(Xnx (t)) dϕ(X x (t)) 
+ E
 − 

dx dx H
∞
  2 x     
× D X (t) B(x)Q1/2 B(x)Q1/2 ∗ ek , ek  ,
H H
k=1

which converges to zero by similar arguments as that used for the first term, but
now we need to employ the fact that an analogue to the bound (3.81) holds for
D 2 Xnx (ek , ek ),
 2 x 
sup sup D X (y, z) < ∞, p ≥ 1. (3.82)
n H˜2p
n yH ,zH ≤1
124 3 Stochastic Differential Equations

Thus we have established that

 2 
∂ un (t, x)  1/2
 
1/2 ∗
tr B(x)Q B(x)Q
∂x 2
 2 
∂ u(t, x)   
1/2 ∗
→ tr B(x)Q 1/2
B(x)Q . (3.83)
∂x 2
Putting together (3.79) and (3.83), we have (at least for a subsequence)
 
∂un (t, x) ∂u(t, x)
lim = , Ax + F (x)
n→∞ ∂t ∂x H
 2 
∂ u(t, x)   
1/2 ∗
+ tr B(x)Q 1/2
B(x)Q . (3.84)
∂x 2
Using bounds (3.81) and (3.82) and the assumptions on ϕ, we conclude that the
left-hand side in (3.84) is bounded as a function of t for any fixed x ∈ H , so that, by
integrating both sides of (3.84) on [0, t], we get that
t  ∂u(s, x) 
lim un (t, x) = , Ax + F (x)
n→∞ 0 ∂x H
 2 
∂ u(s, x)   
1/2 ∗
+ tr B(x)Q 1/2
B(x)Q dt.
∂x 2
This, together with (3.78), proves that
   2 
∂u(t, x) ∂u(t, x) ∂ u(t, x)   
1/2 ∗
= , Ax + F (x) + tr B(x)Q1/2
B(x)Q .
∂t ∂x H ∂x 2
To prove the uniqueness, assume that ũ(t, x) satisfies (3.74) on [0, T [ and fulfills
conditions (1)–(3) of the theorem.
Let Xnx (t) be a mild solution of equation
⎧
⎨dX(t) = AX(t) + Rn F (X(t))dt + Rn B(X(t)) dWt
⎩
X(0) = x ∈ D(A)
that is,
t   t  
Xnx (t) = S(t)x + S(t − s)Rn F Xnx (s) ds + S(t − s)Rn B Xnx (s) dWs ,
0 0

where Rn is defined in (1.21).

Since x ∈ D(A) and Rn : H → D(A), the solution Xnx (t) ∈ D(A). Hence, by
Theorems 3.5 and 3.2 Xnx (t) is a strong solution. Then, using Itô’s formula, we can
consider the differential
  ∂ ũ(t − s, Xnx (s))
d ũ t − s, Xnx = − ds
∂t
3.7 Lipschitz-Type Approximation of Continuous Coefficients 125
 
∂ ũ(t − s, Xnx (s))  
+ , AXnx (s) + Rn F Xnx (s) ds
∂x H
 2 
1 ∂ ũ(t − s, Xnx (s))   x  1/2   x  1/2 ∗
+ tr R n B X n (s) Q R n B X n (s) Q ds
2 ∂x 2
 
∂ ũ(t − s, Xnx (s))  x 
+ , Rn B Xn (s) dWs
∂x H
 
∂ ũ(t − s, Xn (s))
x  x   x 
= , Rn F Xn (s) − F Xn (s) ds
∂x H
 2 
1 ∂ ũ(t − s, Xnx (s))   x  1/2   x  1/2 ∗
+ tr Rn B Xn (s) Q Rn B Xn (s) Q ds
2 ∂x 2
 2 
1 ∂ ũ(t − s, Xnx (s))   x  1/2   x  1/2 ∗
− tr B Xn (s) Q B Xn (s) Q ds
2 ∂x 2
 
∂ ũ(t − s, Xnx (s))  x 
+ , Rn B Xn (s) dWs .
∂x H

The second equality holds since ũ(t, x) is a solution of (3.74) and since the terms
containing A cancel.
Now, we integrate over the interval [0, t] and take expectation. Then we pass to
the limit as n → ∞ (note that the operators Rn are uniformly bounded (see (1.20))
and follow the argument provided in (6.20)). Finally, use the initial condition to
obtain that
 
ũ(t, x) = Eϕ X x (t) .
This concludes the proof. 

3.7 Lipschitz-Type Approximation of Continuous Coefficients

We now construct sequences of Lipschitz-type coefficients Fn and Bn , taking values
in H , which approximate continuous coefficients F and B uniformly on compact
subsets of C([0, T ], H ). The values of Fn and Bn are not restricted to any finite-
dimensional subspaces of H .

Lemma 3.9 Let F : [0, T ] × H → H , B : [0, T ] × H → L2 (KQ , H ) satisfy con-

ditions (A1)–(A3). There exist sequences Fn : [0, T ] × H → H and Bn : [0, T ] ×
H → L2 (KQ , H ) of functions satisfying conditions (A1)–(A4), with a universal
constant in the condition (A3), such that
   
sup F (t, x) − Fn (t, x) + sup B(t, x) − Bn (t, x)
H
→0
L2 (KQ ,H )
0≤t≤T 0≤t≤T

uniformly on any compact set in C([0, T ], H ).

126 3 Stochastic Differential Equations

Proof The sequences Fn and Bn can be constructed as follows. Let {en }∞ n=1 be an
ONB in H . Denote
      
fn (t) = x(t), e1 H , x(t), e2 H , . . . , x(t), en H ∈ R n ,
#n (t) = fn (kT /n) at t = kT /n and linear otherwise,
kT
γn (t, x0 , . . . , xn ) = xk at t = and linear otherwise, with xk ∈ R n ,
n
k = 0, 1, . . . , n.

Let g : R n → R benonnegative, vanishing for |x| > 1, possessing bounded deriva-

tive, and such that R n g(x) dx = 1. Let εn → 0. We define

  
Fn (t, x) = ··· F t, γn (·, x0 , . . . , xn ), e
  n      
εn  2 1
n
n ∧ t − xk dxk
fn kT
× exp − xk g . (3.85)
n εn εn
k=0 k=0

Above, (γn (·, x0 , . . . , xn ), e) = γn1 e1 + · ·

· + γnn en , where γn1 , . . . , γnn are the coor-
dinates of the vector γn in R , and xk = ni=1 (xki )2 , dxk = dxk1 . . . dxkn .
n 2

The coefficients Bn (t, x) are defined analogously using the ONB in KQ . We note
that conditions (A1)–(A4) are satisfied. To see that, note that the functions Fn and
Bn depend on a finite collection of variables fn (kT /n), and hence the arguments of
Gikhman and Skorokhod in [25] can be applied. We only need to verify the uniform
convergence on compact sets of C([0, T ], H ). We have
 
sup Fn (t, x) − F (t, x)H
0≤t≤T
     
≤ ··· sup F t, γn ·, fn (0) + x 0 , . . . , fn (T ∧ t) + x n , e
0≤t≤T
 
     1 n
x k dx k
− F t, γn ·, fn (0), . . . , fn (T ∧ t) , e H g
εn εn
k=0
     
+ sup F (t, x) − F t, γn ·, fn (0), . . . , fn (T ∧ t) , e H
0≤t≤T
  
    ε n
sup F t, γn (·, x0 , . . . , xn ), e H 1 − exp −
n
+ ··· 2
xk
0≤t≤T n
k=0
   kT   
1
n
fn n ∧ t − xk dxk
× g .
εn εn
k=0

We will now verify convergence for each of the three components of the sum
above.
3.7 Lipschitz-Type Approximation of Continuous Coefficients 127

Consider the first summand. If K ⊂ C([0, T ], H ) is a compact set, then the col-
lection of functions {γn (·, fn (0), . . . , fn (T )), e), n ≥ 1} ⊂ C([0, T ], H ), fn (t) =
(x(t), e1 H , . . . , x(t), en H ) with x ∈ K, is a subset of some compact set K1 . This
follows from a characterization of compacts in C([0, T ], H ) (see Lemma 3.14 in
the Appendix) and from Mazur’s theorem, which states that a closed convex hull of
a compact set in a Banach space is a compact set. Moreover,
 
sup sup γn (t, x0 + z0 , . . . , xn + zn ) − γn (t, x0 , . . . , xn )H < εn
x0 ,...,xn 0≤t≤T

if |zk | ≤ εn , k = 0, . . . , n, zk ∈ R n .
The set K1 + B(0, εn ) is not compact (B(0, εn ) denotes a ball of radius εn
centered at 0), but sup0≤t≤T F (t, u) − F (t, v)H can still be made arbitrarily
small if v is sufficiently close to u ∈ K1 . Indeed, given ε > 0, for every u ∈ K1
and t ∈ [0, T ], there exists δut such that if sup0≤t≤T v(t) − u(t)H < δut , then
F (t, v) − F (t, u)H < ε/2. Because t is in a compact set and F (t, u) is continu-
ous in both variables, δu = inft δut > 0. Therefore, for u ∈ K1 , sup0≤t≤T F (t, v) −
F (t, u)H < ε/2 whenever sup0≤t≤T v(t) − u(t)H < δu .
We take a finite covering B(uk , δuk /2) of K1 and let δ = min{δuk /2}. If
u ∈ K1 and sup0≤t≤T v(t) − u(t)H < δ, then for some k, sup0≤t≤T u(t) −
uk (t)H < δuk /2, and sup0≤t≤T v(t) − uk (t)H < δ + δuk /2 ≤ δuk . Therefore,
sup0≤t≤T F (t, v) − F (t, u)H ≤ ε.
Thus taking n sufficiently large and noticing that g(x k /εn ) vanishes if |x k | ≥ εn ,
we get
     
sup F t, γn ·, fn (0) + x 0 , . . . , fn (T ∧ t) + x n , e
0≤t≤T
    
− F t, γn ·, fn (0), . . . , fn (T ∧ t) , e H < ε

for any ε, independently of fn associated with x ∈ K. This gives the uniform con-
vergence to zero on K of the first summand.
Now we consider the second summand. Let Pn be the orthogonal projec-
tion onto the linear subspace spanned by {e1 , . . . , en }, and let Pn⊥ denote the
orthogonal projection onto the orthogonal complement of this space. We note
that sup0≤t≤T Pn⊥ x(t)H → 0 as n → ∞; otherwise, there would be a sequence
tn → t0 with Pn⊥ x(tn )H > c > 0, and Pn⊥ x(tn )H ≤ x(tn ) − x(t0 )H +
Pn⊥ x(t0 )H → 0.
Let N = N(x) be chosen such that if m ≥ N , then sup0≤t≤T Pm⊥ x(t)H < ε/3.
Thus also sup0≤t≤T PN⊥ (#m (x)(t), e)H < ε/3.
There exists M = M(x) ≥ N (x) such that for m ≥ M,
    
sup x(t) − #m (x)(t), e  ≤ sup P ⊥ x(t)
H N H
0≤t≤T 0≤t≤T
     
+ sup PN x(t) − PN #m (x)(t), e H + sup PN⊥ #m (x)(t), e H < ε,
0≤t≤T 0≤t≤T

because the middle term is a finite-dimensional approximation.

128 3 Stochastic Differential Equations

Actually, supx$K sup0≤t≤T x(t) − (#m (x)(t), e)H → 0 if K is compact. If

{x1 , . . . , xn }, is an ε/3 net in K, then
    
sup x(t) − #m (x)(t), e H ≤ sup x(t) − xk (t)H
0≤t≤T 0≤t≤T
      
+ sup xk (t) − #m xk (t), e H + sup  #m xk (t), e − #m x(t), e H < ε
0≤t≤T 0≤t≤T

if m ≥ max{M(xk )}.
The continuity of F and the fact that {(#m (x), e), x ∈ K} ∪ K is a subset of a
compact set, guarantees the uniform convergence to zero of the second summand.
The third summand converges uniformly to zero on compact sets since it is bounded
by
      2
1 + εn + sup x(t)H εn sup x(t)H + εn . 
0≤t≤T 0≤t≤T

Exercise 3.14 Prove that Fn and Bn defined in (3.85) satisfy conditions (A1)–(A4).

We will now consider methods for proving the existence of solutions to SDEs
and SSDEs that are based on some compactness assumptions. In the case of SDEs,
we will require that the Hilbert space H be embedded compactly into some larger
Hilbert space. In the case of SSDEs, the compactness of the semigroup S(t) will be
imposed to guarantee the tightness of the laws of the approximate solutions.
These cases will be studied separately.

3.8 Existence of Weak Solutions Under Continuity Assumption

To obtain weak convergence results, we need an estimate on moments for incre-
ments of solutions to SDEs.
t t
Lemma 3.10 Let ξ(t) = x + 0 F (s, ξ ) ds + 0 B(s, ξ ) dWs with x ∈ H and the
coefficients F and B satisfying conditions (A1) and (A3). Then
 4
E ξ(t + h) − ξ(t)H ≤ C(T , )h2

Proof Using Ito’s formula, we obtain

  t+h    
ξ(t + h) − ξ(t)4 = 4 ξ(u) − ξ(t)2 ξ(u) − ξ(t), F (u, θu ξ ) du
H H H
t
t+h      
+2 ξ(u) − ξ(t)2 tr B(u, θu ξ )Q1/2 B(u, θu ξ )Q1/2 ∗ du
H
t
t+h   ∗ * +⊗2
+2 B(u, θu ξ )Q1/2 B(u, θu ξ )Q1/2 ξ(u) − ξ(t) du
t
3.8 Existence of Weak Solutions Under Continuity Assumption 129

By Lemma 3.6, this yields the following estimate for the fourth moment:

 4 t+h  3    
E ξ(t + h) − ξ(t)H ≤ C1 E ξ(u) − ξ(t)H 1 + supξ(v)H du
t v≤u
t+h  2   2 
+ C2 E ξ(u) − ξ(t)H 1 + supξ(v)H du
t v≤u
, -3/4
t+h  4
≤ C3  
E ξ(u) − ξ(t) H du h1/4
t
, -1/2 
t+h  4
+ E ξ(u) − ξ(t)H du h1/2 ≤ Ch.
t

Substituting repeatedly, starting with C(u − t) for Eξn (u) − ξn (t)4Rn , into the
above inequality, we arrive at the desired result. 

The next lemma is proved by Gikhman and Skorokhod in [25], Vol. I, Chap. III,
Sect. 4.

Lemma 3.11 The condition supn E(ξn (t + h) − ξn (t)4H ) ≤ Ch2 implies that for
any ε > 0,
   
lim sup P sup ξn (t) − ξn (s)H > ε = 0.
δ→0 n |t−s|<δ

Corollary 3.6 Let Fn and Bn satisfy conditions (A1) and (A3) with a common con-
stant in the growth condition (A3) (in particular Fn and Bn can be the approximat-
ing sequences from Lemma 3.9). Let Xn be a sequence of solutions to the following
SDEs:

dXn (t) = Fn (t, Xn ) dt + Bn (t, Xn ) dWt ,
Xn (0) = x ∈ H.
Then
(1) the sequence Xn is stochastically bounded, i.e., for every ε > 0, there exists Mε
satisfying
   
sup P sup Xn (t)H > Mε ≤ ε, (3.86)
n 0≤t≤T

(2) for any ε > 0,

   
lim sup P sup Xn (t) − Xn (s)H > ε = 0. (3.87)
δ→0 n |t−s|<δ

It is known that even a weak solution Xt (·) ∈ C([0, T ], H ) to the SDE (3.8) may
not exist. Therefore our next step is to find a weak solution on a larger Hilbert space.
130 3 Stochastic Differential Equations

Let H−1 be a real separable Hilbert space such that the embedding J : H → H−1
is a compact operator with representation
∞

(J) Jx = λn x, en H hn , λn > 0, n = 1, 2, . . . .
n=1

In general, J has always the above representation; we are only assuming that λn = 0.
Here, {en }∞ ∞
n=1 ⊂ H and {hn }n=1 ⊂ H−1 are orthonormal bases. We will identify H
with J (H ) and P ◦ X with P ◦ X −1 ◦ J −1 . Thus, if x ∈ H , we will also write
−1

x ∈ H−1 . In particular, en = λn hn ∈ H−1 and x, hn H−1 = λn x, en H . Note that

xH−1 ≤ J xH and C([0, T ], H ) ⊂ C([0, T ], H−1 ).

Corollary 3.7 Let Fn , Bn , Xn be as in Corollary 3.6, and J : H → H−1 be a

compact embedding into a real separable Hilbert space H−1 . Then the sequence of
measures {P ◦ Xn−1 }∞
n=1 on C([0, T ], H−1 ) is relatively weakly compact.

Proof For any ε > 0, let Mε be the constant in condition (3.86) of Corollary 3.6.
The ball B(0, Mε ) ⊂ H of radius Mε , centered at 0, is relatively compact in H−1 .
Denote by B its closure in H−1 ; then
     
P J Xn (t) ∈ / B = P Xn (t)H > Mε < ε.

Condition (3.87) of Corollary 3.6 is also satisfied in H−1 . The relative compactness
now follows from the tightness criterion given in Theorem 3.17 in the Appendix,
and from Prokhorov’s theorem. 

In order to construct a weak solution to the SDE (3.8) on C([0, T ], H−1 ), we impose
some regularity assumptions on the coefficients F and B with respect to the Hilbert
space H−1 .
Assume that F : [0, T ]×C([0, T ], H−1 ) → H−1 and B : [0, T ]×C([0, T ], H−1 )
→ L2 (KQ , H−1 ) satisfy the following conditions:
(B1) F and B are jointly measurable, and for every 0 ≤ t ≤ T , they are measurable
with respect to the σ -field C˜t on C([0, T ], H−1 ) generated by cylinders with
bases over [0, t].
(B2) F and B are jointly continuous.
(B3) There exists a constant −1 such that ∀x ∈ C([0, T ], H−1 ),
     
F (t, x) + B(t, x)L (K ,H ) ≤ −1 1 + sup x(t)H−1 ,
H −1 2 Q −1
0≤t≤T

for ω ∈  and 0 ≤ t ≤ T .
Equation (3.8) is now considered in H−1 , and in the circumstances described above,
we can prove the existence of a weak solution to the SDE (3.8) on C([0, T ], H−1 ).
However we do not need all conditions (A1)–(A3) and (B1)–(B3) to hold simulta-
neously. We state the existence result as follows.
3.8 Existence of Weak Solutions Under Continuity Assumption 131

Theorem 3.12 Let H−1 be a real separable Hilbert space. Let the coefficients F
and B of the SDE (3.8) satisfy conditions (B1)–(B3). Assume that there exists a
Hilbert space H such that the embedding J : H → H−1 is a compact operator with
representation (J) and that F and B restricted to H satisfy
 
F : [0, T ] × C [0, T ], H → H,
 
B : [0, T ] × C [0, T ], H → L2 (KQ , H ),

and the linear growth condition (A3). Then the SDE (3.8) has a weak solution X(·) ∈
C([0, T ], H−1 ).

Proof Since the coefficients F and B satisfy assumptions (B1)–(B3), we can con-
struct approximating sequences Fn : [0, T ] × C([0, T ], H−1 ) → H−1 and Bn :
[0, T ]×C([0, T ], H−1 ) → L2 (KQ , H−1 ) as in Lemma 3.9. The sequences Fn → F
and Bn → B uniformly on compact subsets of C([0, T ], H−1 ).
Now we consider restrictions of the functions Fn , Bn to [0, T ] × C([0, T ], H ),
and we claim that they satisfy conditions (A1), (A3), and (A4). Let us consider
the sequence Fn only; similar arguments work for the sequence Bn . We adopt the
notation developed in Lemma 3.9.
If x ∈ C([0, T ], H ), then
  
Fn (t, x) = ··· F t, γn (·, x1 , . . . , xn ), h
  n      
εn  2 1
n
f˜n kT
n ∧ t − xk dxk
× exp − xk g ∈ H,
n εn εn
k=0 k=0

where
    
f˜n (t) = x(t), h1 H , . . . , x(t), hn H
−1 −1
     
= λ1 x(t), h1 H , . . . , λn x(t), hn H := λfn (t),
  
  e1 en
γn (·, x0 , . . . , xn ), h = γn (·, x0 , . . . , xn ), ,...,
λ1 λn
   
x0 xn  
= γn ·, , . . . , , e ∈ C [0, T ], R n ,
λ λ

xk  xk1 xkn 
and λ = λ1 , . . . , λn ∈ R .
n Let εn /λk < 1, k = 1, . . . , n. Then
    
x0 xn
Fn (t, x) = · · · F t, γn ·, , . . . , ,e
λ λ
  n     kT  xk   
εn  2 1 λ fn n ∧ t −
n
λ dxk
× exp − xk g
n εn εn
k=0 k=0
132 3 Stochastic Differential Equations

  
= ··· F t, γn (·, y0 , . . . , yn ), e
  n       
εn  1
n
n ∧ t − yk
λ fn kT dyk
× exp − (yk λ)2
g 2n .
n εn εn / i=1 λi
k=0 k=0

First, we observe that Fn are measurable with respect to the product σ -field on
[0, T ] × C([0, T ], H ) because Fn satisfy condition (B1), depend only on finitely

many variables, and J −1 y = ∞ k=1 1/λk y, hk H−1 ek on J (H ) is measurable from
J (H ) to H as a limit of measurable functions. The same argument justifies that Fn
are adapted to the family {Ct }t≤T .
The linear growth condition (A3) is satisfied with a universal constant. Indeed,

     
Fn (t, x) ≤ ··· F t, γn (·, y0 , . . . , yn ), e 
H H

n     kT
1
  
λ f n n ∧ t − yk dyk
× g 2n
εn εn / i=1 λi
k=0
    1
n
≤ 1 + sup x(t)H + max (εn /λk ) ··· g(zk ) dzk
0≤t≤T 1≤k≤n
k=0
   
≤ 
1 + sup x(t)H ,
0≤t≤T

because of the choice of εn .

Further, the function Fn depends only on finitely many variables f ( kT n ∧ t) =
((x( kT
n ∧ t), e 1 )H , . . . , (x( kT
n ∧ t), e n )H ) ∈ R n , k = 1, . . . , n.

By differentiating under the integral sign, we get that, with x ∈ C([0, T ], H ),

∂Fn (t, x)  
 kT  ≤ ··· 1 + max |yk |
∂fl n ∧ t 0≤k≤n
   
εn    λl
n
dyl
× exp − (yk λ)2 supg  (z) 2n
n z εn εn / i=1 λi
k=0

1n     kT   
λ fn n ∧ t − yk dyk
× g 2
εn εn / ni=1 λi
k=0,k =l
 2n n     n
  εn 
n 1
i=1 λi λl
≤C ··· 1 + max |yk | exp − (yk λ)2
dyk
εn εn 0≤k≤n n
k=0 k=0
< ∞.
3.8 Existence of Weak Solutions Under Continuity Assumption 133

Thus, Fn (t, x) has bounded partial derivatives, and hence it is a Lipschitz function
with respect to the variable x. Let Xn be a sequence of strong solutions to equations
t t
Xn (t) = x + Fn (s, Xn ) ds + Bn (s, Xn ) dWs
0 0

considered on C([0, T ], H ). By Corollary 3.12, the sequence of measures μn =

P ◦ Xn−1 = P ◦ Xn−1 ◦ J −1 on C([0, T ], H−1 ) (with H identified with J (H )) is
relatively weakly compact, and therefore, by choosing a subsequence if necessary,
we can assume that μn converges weakly to a measure μ on C([0, T ], H−1 ).
We now follow the ideas of Gikhman and Skorokhod in [25]. Let gt be a bounded
continuous function on C([0, T ], H−1 ) measurable with respect to the cylindrical
σ -field generated by cylinders with bases over [0, t]. Then for any u ∈ H−1 ,
, -
  t+h  
x(t + h) − x(t), u H − Fn (s, x), u H ds gt (x)μn (dx)
−1 −1
t
  t+h   
= E Xn (t + h) − Xn (t) − Fn (s, Xn ) ds gt (Xn ) , u = 0. (3.88)
t H−1

Let νn (dx) = (1 + sup0≤s≤T x(s)H−1 )μn (dx). As in the finite-dimensional case,

the measures νn are uniformly bounded and are weakly convergent.
Indeed, using the tightness of the sequence μn , for any ε > 0, we find a com-
pact set Kε ⊂ C([0, T ], H−1 ) with lim supn μn (Kεc ) < ε, and therefore if BN ⊂
C([0, T ], H−1 ) is a ball of radius N centered at 0 and if Kε ⊂ BN , we obtain
 1/2
  1/2  c   2
sup x(s)H dμn ≤ μn BN sup x(s)H dμn
c 0≤s≤T −1 c 0≤s≤T −1
BN BN
1/2   √
≤ Cμn Kεc = C ε,

and for a real-valued, bounded, continuous function f ,

  
        
 f (x) 1 + sup x(s)  
dμn − f (x) 1 + sup x(s) H dμ
 H −1 −1
0≤s≤T 0≤s≤T
  
  
≤  f (x) 1 + sup x(s)H ∧ N dμn
−1
0≤s≤T
 
  
− f (x) 1 + sup x(s)H ∧ N dμ
−1
0≤s≤T
 √  √ 
+ f ∞ C ε + ε = O ε .

Thus, there exists a compact set K̃ε ⊂ C([0, T ], H−1 ) such that lim supn→∞ νn (K̃εc )
≤ ε. Thus, because of the uniform convergence sup0≤s≤T F (s, x) − Fn (s, x)H−1
134 3 Stochastic Differential Equations

→ 0 on K̃ε and the growth condition,

t+h   
lim sup  F (s, x) − Fn (s, x), u g (x)  ds μn (dx)
H−1 t
n→∞ t
t+h F (s, x) − Fn (s, x)H−1 uH−1  
≤ C lim sup gt (x) ds νn (dx) = O(ε).
n→∞ t (1 + sup0≤s≤T x(s)H−1 )

It follows that
t+h   
lim  F (s, x) − Fn (s, x), u g (x)  ds μn (dx) = 0,
n→∞ H−1 t
t

and the weak convergence of the measures μn , together with the uniform integra-
bility, implies that, as n → ∞,
   
x(t + h) − x(t), u H gt (x) dμn → x(t + h) − x(t), u H gt (x) dμ (3.89)
−1 −1

and
t+h   t+h  
F (s, x), u H gt (x) ds dμn → F (s, x), u H gt (x) ds dμ. (3.90)
−1 −1
t t

Now, taking the limit in (3.88), we obtain

, -
  t+h  
x(t + h) − x(t), u H − F (s, x), u H ds gt (x)μ(dx) = 0,
−1 −1
t
t
proving that the process y(t) = x(t) − 0 F (s, x) ds is a martingale on C([0, T ],
H−1 ) endowed with the canonical filtration. Using the equality
 t+h 2
x(t + h) − x(t) − Fn (s, x) ds, u gt (x)μn (dx)
t H−1
t+h   ∗ * ⊗2 +
= Bn (s, x)Q1/2 Bn (s, x)Q1/2 u gt (x)μn (dx)
t

one can prove in a similar way that the process

 2 t   ∗ * ⊗2 +
y(t), u H − B(s, x)Q1/2 B(s, x)Q1/2 u ds
−1
0

is a martingale on C([0, T ], H−1 ) and obtain the increasing process

  t   ∗ * ⊗2 +
(y(·), u)H−1 t = B(s, x)Q1/2 B(s, x)Q1/2 u ds.
0
3.9 Compact Semigroups and Existence of Martingale Solutions 135

In view of Theorem 2.7, the measure μ is a law of a C([0, T ], H−1 )-valued process
X(t), which is a weak solution to the SDE (3.8) (we can let Φ(s) = B(s, X) in
Theorem 2.7). 

The following result was communicated to us by A.V. Skorokhod [69].

Theorem 3.13 Consider the equation

   
dXt = a X(t) dt + B X(t) dWt , (3.91)

where a : H → H , B : H → L (K, H ), and Wt is a K-valued Wiener process with

covariance Q. Assume the following conditions:
(1) a and B are jointly continuous and locally bounded,
(2) 2a(x), x H + tr(B(x)QB ∗ (x)) ≤ (1 + x2H ),
(3) there exists a positive symmetric compact linear operator S on H such that
     
2 S −1 a(Sx), x H + tr S −1 B(Sx)QB ∗ (Sx)S −1 ≤ 1 1 + x2H .

Then there exists a weak solution to (3.91) with X(0) = Sx0 , x0 ∈ H .

Exercise 3.15 Let us substitute conditions (1)–(3) with

(1’) a and B are jointly continuous,
(2’) a(x)H + B(x)L (K, H ) ≤ (1 + xH ),
(3’) there exists a positive symmetric compact linear operator S on H such that
 −1    
S a(Sx) + S −1 B S(x)  (K, H ) ≤ 1 (1 + xH ).
H L

Prove the assertion of Theorem 3.13.

Hint: define a norm x0 = S −1 xH on H0 = S(H ). Prove that the assump-
tions of Theorem 3.12 are satisfied with J : H0 → H .

3.9 Compact Semigroups and Existence of Martingale Solutions

In this section we present an existence theorem for martingale solutions in case
where the operator A generates a compact semigroup. This extends Theorem 8.1 in
[11], since we include coefficients F and B of (3.1) that may depend on the entire
past of a solution.
The following technical lemma will allow us to prove the tightness of a family
of probability measures. We will use it in the proof of Theorem 3.14 and again in
Sect. 7.4.

Lemma 3.12 Let p > 1 and 1/p < α ≤ 1. Consider the operator Gα defined
in (3.24),
t  
Gα f (t) = (t − s)α−1 S(t − s)f (s) ds, f ∈ Lp [0, T ], H .
0
136 3 Stochastic Differential Equations

Assume that {S(t), t ≥ 0} is a compact C0 -semigroup on H . Then Gα is compact

from Lp ([0, T ], H ) into C([0, T ], H ).

Proof It is enough to show that

   
Gα f ∈ Lp [0, T ], H : f Lp ≤ 1

is relatively compact in C([0, T ], H ). We will show that conditions (1) and (2) of
Lemma 3.14, in the Appendix, hold, i.e., that for any fixed 0 ≤ t ≤ T , the set

Gα f (t) : f Lp ≤ 1 (3.92)

is relatively compact in H and that

 
lim sup Gα f (t) − Gα f (s) = 0. (3.93)
δ→0+ H
|t−s|<δ,t,s∈[0,T ]

For ε ∈ (0, t], f ∈ Lp ([0, T ], H ), define

 ε  t−ε
Gα f (t) = (t − s)α−1 S(t − s)f (s) ds.
0

Then
t−ε
Gεα f = S(ε) (t − s)α−1 S(t − ε − s)f (s) ds.
0
Since S(ε) is compact, then so is Gεα . Let q = p/(p − 1). Now, using Hölder’s
inequality, we have
   
(Gα f )(t) − Gε f (t)
α H
 t 
 
= (t − s) S(t − s)f (s) ds 
α−1

t−ε H
 1  1
t  q q t   p
≤ (t − s)(α−1)q S(t − s) ds f (s)p p ds
L
t−ε t−ε
 1
ε (α−1)q+1 q
≤M f Lp
(α − 1)q + 1

with M = sups∈[0,T ] S(s)L (H ) . Since (α − 1)q + 1 = α − 1/p > 0, Gεα → Gα in

the operator norm in L(Lp ([0, T ], H ), H ); thus the limit Gα is a compact operator
from Lp ([0, T ], H ) to H , and the relative compactness of the set (3.92) follows.
Consider now the difference
 
Gα f (t) − Gα f (s)
H
s   
≤ (t − u)α−1 S(t − u) − (s − u)α−1 S(s − u) f (u) du
L (H ) H
0
3.9 Compact Semigroups and Existence of Martingale Solutions 137

t 
+ (t − u)α−1 S(t − u)f (u) du
H
s
 T 1
 q
≤ (v + u)α−1 S(v + u) − v α−1 S(v)q dv f Lp
L (H )
0
 t−s 1
q
+M v (α−1)q
dv f Lp
0
= I1 + I2 ,

where u = t − s > 0. Since, as u → 0, the expression

 
(v + u)α−1 S(v + u) − v α−1 S(v)q →0
L (H )

by the compactness of the semigroup, see (1.10), and it is bounded by (2M)q v (α−1)q ,
we conclude by the Lebesgue DCT that I1 → 0 as u = t − s → 0. Also, the second
term
(t − s)α−1/p
I2 ≤ M f Lp → 0
((α − 1)q + 1)1/q
as t − s → 0. This concludes the proof. 

Theorem 3.14 Assume that A is the infinitesimal generator of a compact

C0 -semigroup S(t) on a real separable Hilbert space H . Let the coefficients of
the SSDE (3.1) satisfy conditions (A1)–(A3). Then (3.1) has a martingale solution.

Proof As in the proof of Theorem 3.12, we begin with a sequence of mild solutions
Xn to equations
t  t
Xn (t) = x + AXn (s) + Fn (s, Xn ) ds + Bn (s, Xn ) dWs .
0 0

The coefficients Fn and Bn are the Lipschitz approximations of F and B as in

Lemma 3.9. Lemma 3.6 guarantees that
 2p
sup E sup Xn (t)H < C, p > 1. (3.94)
n 0≤t≤T

For p > 1 fixed, define

t
Yn (t) = (t − s)−α S(t − s)Bn (s, Xn ) dWs
0

with 1/(2p) < α < 1/2, then condition (A3) and inequality (3.94) imply that
T 
E Yn (s)2p ds < C  . (3.95)
H
0
138 3 Stochastic Differential Equations

Using the factorization technique, as in Lemma 3.3, we can express Xn (t) as fol-
lows:
t t
Xn (t) = S(t)x + S(t − s)Fn (s, Xn ) ds + S(t − s)Bn (s, Xn ) dWs
0 0
sin πα
= S(t)x + G1 Fn (·, Xn )(t) + Gα Yn (t),
π
where Gα : L2p ([0, T ], H ) → C([0, T ], H ), defined in (3.24), is a compact opera-
tor for 1/2p < α ≤ 1 by Lemma 3.12.
Inequalities (3.94) and (3.95) and the growth condition (A3) imply that for any
ε > 0, there exists ν > 0 such that for all n ≥ 1,
 T  1    1 
 2p π T  2p
P Yn (s)2p ds ≤ ν ∩ Fn (s, Xn )2p ds ≤ν
H sin πα H
0 0
≥ 1 − ε.

By the compactness of Gα and S(t) and the continuity of the mapping t → S(t)x,
we conclude that the set

K = S(·)x + Gα f (·) + G1 g(·) : f L2p ≤ ν, gL2p ≤ ν

is compact in C([0, T ], H ) and obtain the tightness of the measures μn = L (Xn )

on C([0, T ], H ).
We will now prove that
 t  t t
Xn (t) = x + A Xn (s) ds + F (s, Xn ) ds + B(s, Xn ) dWs . (3.96)
0 0 0

Let Am = mRm = mA(m − A)−1 ,

m > α, be the Yosida approximations of A (with
α determined by the Hille-Yosida Theorem 1.4). Consider the equations
t t t
Xn,m = x + Am Xn,m (s) ds + Fn (s, Xn,m ) ds + Bn (s, Xn,m ) dWs ,
0 0 0

whose strong and mild solutions coincide. Moreover, by Proposition 3.2, as m →

∞,
 2p
E sup Xn,m (t) − Xn (t)H ds → 0,
0≤t≤T

which implies, by selecting a subsequence if necessary, that a.s., as m → ∞,

 
Xn,m → Xn in C [0, T ], H .

Using the Lipschitz-type condition (A4), we obtain that

t t
Fn (s, Xn,m ) ds → Fn (s, Xn ) ds,
0 0
3.9 Compact Semigroups and Existence of Martingale Solutions 139

t t
Bn (s, Xn,m ) dWs → Bn (s, Xn ) dWs
0 0
t
a.s. in C([0, T ], H ), which implies that Am 0 Xn,m (s) ds is an a.s. convergent se-
quence in C([0, T ], H ). Because
t t
Xn,m (s) ds → Xn (s) ds
0 0

a.s. in C([0, T ], H ), Am = A(m(m − A)−1 ) = ARm , with the operators Rm uni-

formly bounded, and Rm (u) → u for u ∈ H , we obtain that, a.s.,
 t  t
Rm Xn,m (s) ds → Xn (s) ds for each t.
0 0
t t
The operator A is closed, and A(Rm ( 0 Xn,m (s) ds)) = Am ( 0 Xn,m (s) ds) con-
t t t
verges; therefore, 0 Xn (s) ds ∈ D(A) and Am ( 0 Xn,m (s) ds) → A 0 Xn (s) ds,
and, as we showed above, the convergence actually holds a.s. in C([0, T ], H ). This
proves that the representation (3.96) is valid.
Using the tightness of the measures μn = L (Xn ), we can assume, passing
to a subsequence if necessary, that Xn → X weakly for some process X(·) ∈
C([0, T ], H ). Using the Skorokhod theorem and changing the underlying proba-
bility space, we can assume that Xn → X a.s. as C([0, T ], H )-valued random vari-
ables.
The process
t t
Mn (t) = Xn (t) − x − A Xn (s) ds − Fn (s, Xn ) ds
0 0
t
= Bn (s, Xn ) dWs (3.97)
0

is a martingale with respect to the family of σ -fields Fn (t) = σ (Xn (s), s ≤ t). Be-
cause the operator A is unbounded, we cannot justify direct passage to the limit with
n → ∞ in (3.97), as we did in Theorem 3.12, but we follow an idea outlined in [11].
Consider the processes

Nn,λ (t) = (A − λI )−1 Mn (t)

t
= (A − λI )−1 Xn (t) − (A − λI )−1 x − (A − λI )−1 A Xn (s) ds
0
t
− (A − λI )−1 Fn (s, Xn ) ds
0

with λ > α (thus λ is in the resolvent set of A). The martingale Mn is square inte-
grable, and so is the martingale Nn,λ . The quadratic variation process of Nn,λ has
140 3 Stochastic Differential Equations

the form
t* +* +∗
Nn,λ t = (A − λI )−1 Bn (s, Xn )Q1/2 (A − λI )−1 Bn (s, Xn )Q1/2 ds.
0

We note that the operator (A − λI )−1 A is bounded on D(A) and can be extended
to a bounded operator on H . We denote this extension by Aλ .
Observe that we are now in a position to repeat the proof of Theorem 3.12
in the current situation (the assumption concerning a compact embedding J in
Theorem 3.12 is unnecessary if the sequence of measures is weakly convergent,
which now is the case, and we carry out the proof for C([0, T ], H )-valued pro-
cesses). The coefficients Fnλ (s, x) = Aλ x(s) + (A − λI )−1 Fn (s, x) and Bnλ (s, x) =
(A − λI )−1 Bn (s, x) satisfy assumptions (A1)–(A4), and the coefficients F λ (s, x) =
Aλ x(s) + (A − λI )−1 F (s, x) and B λ (s, x) = (A − λI )−1 B(s, x) satisfy assump-
tions (A1)–(A3) with Fnλ and Bnλ converging to F λ and B λ , respectively, uniformly
on compact subsets of C([0, T ], H ).
Moreover, by inequality (3.94), we have the uniform integrability,

 2 M2  2
E sup Nn,λ (t)H ≤ E sup Mn (t)H < ∞.
0≤t≤T (λ − α) 0≤t≤T
2

Thus we conclude, as in the proof of Theorem 3.12, that the process

Y λ (t) = (A − λI )−1 X(t) − (A − λI )−1 x

t t
− Aλ X(s) ds − (A − λI )−1 F (s, X) ds
0 0

is a square-integrable martingale on C([0, T ], H ) with respect to the family of

σ -fields Ft = σ (X(s), s ≤ t), and its quadratic variation process has the form
t* +* +∗
Y λ t = (A − λI )−1 B(s, X)Q1/2 (A − λI )−1 B(s, X)Q1/2 ds.
0

The representation theorem, Theorem 2.7, implies the existence of a Q-Wiener pro-
˜ F × F˜ , {Ft × F˜t }, P × P̃ ) such
cess Wt on a filtered probability space ( × ,
that
t t
(A − λI )−1 X(t) − (A − λI )−1 x − Aλ X(s) ds − (A − λI )−1 F (s, X) ds
0 0
t
= (A − λI )−1 B(s, X) dWs .
0

Consequently,
t t t
X(t) = x + (A − λI )Aλ X(s) ds + F (s, X) ds + B(s, X) dWs ,
0 0 0
3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener Process 141

and because for u ∈ D(A∗ ), ((A − λI )Aλ )∗ (u) = A∗ (u), we obtain

  t  t 
X(t), u H = x, u H + X(s), A∗ u H ds + F (s, X), u H ds
0 0
t 
+ u, B(s, X) dWs H
.
0

It follows by Theorem 3.2 that the process X(t) is a mild solution to (3.1). 

Exercise 3.16 Show that the operator (A − λI )−1 A is bounded on D(A).

3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener

Process

We now present an existence and uniqueness result from [12], which later will be
useful in discussing an innovative method for studying invariant measures in the
case of a compact semigroup (see Sect. 7.4.3).
Let K and H be real separable Hilbert spaces, and W )t be a cylindrical Wiener
process in K defined on a complete filtered probability space (, F , {Ft }t≤T , P )
with the filtration {Ft }t≤T satisfying the usual conditions. We consider the follow-
ing SSDE on [0, T ] in H , with an F0 -measurable initial condition ξ :

dX(t) = (AX(t) + F (X(t))) dt + B(X(t)) d W )t ,
(3.98)
X(0) = ξ.

Let the coefficients of (3.98) satisfy the following assumptions,

(DZ1) A is the infinitesimal generator of a strongly continuous semigroup {S(t), t ≥
0} on H .
(DZ2) F : H → H is a mapping such that for some c0 > 0,
   
F (x) ≤ c0 1 + xH , x ∈ H,
H
 
F (x) − F (y) ≤ c0 x − y, x, y ∈ H.
H

(DZ3) B : H → L (K, H ) is such that for any k ∈ K, the mapping x → B(x)k

is continuous from H to H and for any t > 0 and x ∈ H , S(t)B(x) ∈
L2 (K, H ), and there exists a locally square-integrable mapping K :
[0, ∞) → [0, ∞) such that
   
S(t)B(x) ≤ K (t) 1 + xH for t > 0, x ∈ H,
L2 (K,H )
 
S(t)B(x) − S(t)B(y) ≤ K (t)x − yH , for t > 0, x, y ∈ H.
L (K,H ) 2
142 3 Stochastic Differential Equations

(DZ4) There exists α ∈ (0, 1/2) such that

1
t −2α K 2 (t) dt < ∞.
0

We are interested here only in mild solutions and, taking into account the assumption
(DZ3), we have the following definition.

Definition 3.3 A stochastic process X(t) defined on a filtered probability space

(, F , {Ft }t≤T , P ), adapted to the filtration {Ft }t≤T , is a mild solution of (3.98)
if
 T 
 
P  
X(t) H dt < ∞ = 1, (3.99)
0
 
T      
P F X(t)  + S(t − s)B X(t) 2 dt < ∞ = 1, (3.100)
H L 2 (K,H )
0

and for all t ≤ T , P -a.s.,

t   t  
X(t) = S(t)ξ + S(t − s)F X(s) ds + )s .
S(t − s)B X(s) d W (3.101)
0 0

Recall from Sect. 3.3 that H˜2p denotes a Banach space of H -valued stochas-
tic processes X, measurable as mappings from ([0, T ] × , B([0, T ]) ⊗ F ) to
2p
(H, B(H )), adapted to the filtration {Ft }t≤T , and satisfying sup0≤s≤T Eξ(s)H <
∞ with the norm
 1
2p 2p
XH˜2p = sup EX(t)H .
0≤t≤T

We will need the following lemmas.

Lemma 3.13 Let Φ ∈ Λ2 (K, H ), p ≥ 1, then

 2p
 t 
sup E 
 Φ(s) d Ws 
)

0≤t≤T 0 H
 p
 p T  2p 1/p
≤ p(2p − 1) E Φ(s)L (K,H ) ds . (3.102)
0

Proof For p = 1, the result is just the isometric property (2.32) of the stochastic
t
integral. For p > 1, let M̃(t) = 0 Φ(s) d W)s , and we apply the Itô formula (2.61) to
2p
M̃(t)H and, as in the proof of Lemma 3.1, obtain
 
 2p s   
E M̃(s) ≤ p(2p − 1)E M̃(u)2(p−1) Φ(u)2
H L2 (K,H )
du .
0
3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener Process 143

Using the Hölder inequality, we have

 2p s  2p (p−1)/p   2p 1/p
E M̃(s) ≤ p(2p − 1) E M̃(u)H E Φ(u)L (K,H ) du
2
0
s  2p (p−1)/p   2p 1/p
≤ p(2p − 1) sup E M̃(v)H E Φ(u)L (K,H ) du.
2
0 0≤v≤u

Consequently,
 2p
sup E M̃(t)
0≤t≤T
T  2p (p−1)/p   2p 1/p
≤ p(2p − 1) sup E M̃(s)H E Φ(t)L (K,H ) dt
2
0 0≤s≤t
  2p (p−1)/p T  2p 1/p
≤ p(2p − 1) sup E M̃(s)H E Φ(t)L (K,H ) dt,
2
0≤t≤T 0

and (3.102) follows. 

Theorem 3.15 (a) Under conditions (DZ1)–(DZ3), for an arbitrary F0 -measurable

2p
initial condition ξ such that Eξ H < ∞, p ≥ 1, there exist a unique mild solution
to (3.98) and a constant C, independent of the initial condition ξ , such that
2p  2p 
X ≤ C 1 + Eξ H . (3.103)
H˜2p

(b) If, in addition, condition (DZ4) holds, then the solution X(t) is continuous
P -a.s.

Proof For X ∈ H˜2p , let

t   t  
I˜(X)(t) = S(t)ξ + S(t − s)F X(s) ds + )s .
S(t − s)B X(s) d W
0 0

First let us show that

   
I˜(X)2p ≤ C 1 + X2p , (3.104)
˜ ˜
H2p H2p

where the constant C may depend on ξ . We have


   
I˜(X)2p ≤ sup 32p−1 S(t)2p Eξ 2p
˜
H2p L (H ) H
t≤t≤T
 t 2p
 
+E S(t − s)F X(s)  ds
H
0
144 3 Stochastic Differential Equations
 t 2p 
   
+ E  S(t − s)B X(s) d Ws 
 )

0 H
   2p 
≤ C Eξ H + 1 + sup E X(t)H
2p
0≤t≤T
, -p 
t   2p 1/p
sup E S(t − s)B X(s) L (K,H )
2
0≤t≤T 0

2p 2p
≤ C Eξ H + 1 + X ˜
H2p
 p  
T  2p 
+2 2p−1
K (t − s) ds
2
1 + sup E X(t)H
0 0≤t≤T
 2p 
≤ Cξ 1 + X ˜ ,
H2p

where we have used Lemma 3.13 to find a bound on the norm of the stochastic
integral. Next, we compute
 t
     
 S(t − s) F X(s) − F Y (s) ds

0
2p
t      
+ S(t − s) B X(s) − B Y (s) d Ws 
)

0 H
 t 

2p−1 
     2p
≤2 
 S(t − s) F X(s) − F Y (s) ds 
0 H
 t 2p 
      
+ )
 S(t − s) B X(s) − B Y (s) d Ws 
0 H
 t  2p
≤2 2p−1
C1 E X(s) − Y (s)H ds
0
 p 
t   2p 1/p
+ C2 K 2 (t − s) E X(s) − Y (s)H ds .
0
T
Let L1 = 22p−1 max{C1 , C2 } and L2 = 2L1 (1 + p 0 K 2 (t) dt). Let, as in the
proof of Theorem 3.5, B denote the Banach space obtained from H̃2p by modi-
fying its norm to an equivalent norm
  2p  2p1
XB = sup e−L2 t E X(t)H .
0≤t≤T

Then,
 
I˜(X) − I˜(Y )2p
B
 t  2p
≤ sup e−L2 t L1 E X(s) − Y (s)H ds
0≤t≤T 0
3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener Process 145
 p 
t   2p 1/p
+ K 2 (t − s) E X(s) − Y (s)H ds
0
 t  2p
= sup e −L2 t
L1 E X(s) − Y (s)H eL2 s e−L2 s ds
0≤t≤T 0
 p 
t   2p 1/p
+ K (t − s) eL2 s e−L2 s E X(s) − Y (s)H
2
ds
0
 t  2p 
≤ sup e−L2 t L1 sup e−L2 s E X(s) − Y (s)H eL2 s ds
0≤t≤T 0 0≤s≤T
  p 
t   2p 1/p
+ K 2 (t − s)eL2 s/p sup E −L2 s E X(s) − Y (s)H ds
0 0≤s≤T
 t  t p 
≤ L1 X − Y B sup e−L2 t
2p
eL2 s ds + K 2 (t − s)eL2 s/p ds
0≤t≤T 0 0
  t p  t p 
2p −L2 t eL2 t−1
≤ L1 X − Y B sup e + K (s) ds 2
e L2 s/p
ds
0≤t≤T L2 0 0
  p  p 
1 − e−L2 T t 1 − e−L2 T 2p
≤ L1 + K (s) ds
2
p p
X − Y B
L2 0 L2
2p
≤ CB X − Y B
with the constant CB < 1.
Hence, I˜ is a contraction on B, and it has a unique fixed point, which is the
solution to (3.98).
To prove (3.103), note that
 2p
   t t 
I˜(0)2p = sup S(t)ξ + S(t − s)F (0) ds + S(t − s)B(0) d Ws 
)
˜
H2p  
0≤t≤T 0 0 H
 2p 
≤ C0 1 + ξ H
for a suitable constant C0 .
Since the fixed point can be obtained as a limit, X = limn→∞ I˜n (0) in B, using
the equivalence of the norms in H˜2p and B, we have
 ∞

  n+1   
XH˜2p ≤ C1 I˜ (0) − I˜n (0) + I˜(0)
B H˜2p
n=1
∞
 
n ˜

≤ C2 CB I (0)H˜
2p
n=0
 2p 
≤ C 1 + ξ H

for suitable constants C, C1 , C2 .

146 3 Stochastic Differential Equations

To prove part (b), the continuity of X(t), it is enough to show that the stochastic
convolution with respect to a cylindrical Wiener process,
t  
S ˜ B(X)(t) = )s ,
S(t − s)B X(s) d W
0

1 1
has a continuous version. Similarly as in the proof of Lemma 3.3, let 2p <α< 2
and define
s  
Y (s) = )σ .
(s − σ )−α S(s − σ )B X(σ ) d W
0

Using the stochastic Fubini Theorem 2.3 for a cylindrical Wiener process, we have
t   t
)s = sin πα
S(t − s)B X(s) d W (t − s)α−1 S(t − s)Y (s) ds.
0 π 0

Using the Hölder inequality, we have

 2p
 t  
sup 
 S(t − s)B X(s) d Ws 
)

0≤t≤T 0 H
 2p
 sin πα t 
= sup  π (t − s) α−1
S(t − s)Y (s) ds 

0≤t≤T 0 H
T 
≤C Y (s)2p ds.
H
0

However,
 2p
T  2p T  s   
E Y (s)H ds = E
 (s − σ ) −α
S(s − σ )B X(σ ) d Wσ 
)
 ds
0 0 0 H
 p
T s   2
≤C E (s − σ )−2α S(s − σ )B X(σ ) L dσ ds
2 (K,H )
0 0
T s   2 
p
≤C (s − σ ) −2α
K (s − σ ) 1 + X(σ )H dσ
2
ds
0 0

 T p
2p 
s
≤ C 1 + XH2p (s − σ )−2α K 2 (s − σ ) dσ ds < ∞.
0 0

Since the process Y (t) has almost surely 2p-integrable paths, Lemma 3.2 implies
that
sin πα
S ˜ B(X)(t) = Gα Y (t)
π
has a continuous version. 
3.10 Mild Solutions to SSDEs Driven by Cylindrical Wiener Process 147

We conclude with two important properties of the solutions to (3.98).

Proposition 3.3 The solution X ξ (t) of (3.98) as a function of the initial condition
ξ is continuous as mapping from L2p (, H ), p ≥ 1, into itself, and there exists a
constant C such that for ξ , η ∈ L2p (, H ),

 2p
E X ξ (t) − X η (t)H ≤ CEξ − ηH .
2p
(3.105)

Proof Using assumptions (DZ2) and (DZ3), Lemma 3.13, and Exercise 3.7, we
calculate
  2p
 2p t 
E X ξ (t) − X η (t)H ≤ C Eξ − ηH + E X ξ (s) − X η (s) ds
2p
H
0
 p 
t    2p 1/p
+ E K (t − s)X ξ (s) − X η (s)H ds
0
  t 2p

2p
≤ C Eξ − ηH + E X ξ (s) − X η (s) ds
H
0
 p 
t   2p 1/p
+ K 2 (t − s) E X ξ (s) − X η (s)H ds
0
  t 2p

2p
≤ C Eξ − ηH + E X ξ (s) − X η (s) ds
H
0
 p 
t t  2p 1/p
+ K 2 (s) ds E X ξ (s) − X η (s)H ds
0 0
 
t  2p
E X ξ (s) − X η (s)H ds .
2p
≤ C Eξ − ηH +
0

Now an appeal to the Gronwall lemma concludes the proof. 

Proposition 3.4 The solution of (3.98) is a homogeneous Markov and Feller pro-
cess.

We omit the proof since it follows nearly word by word the proof of Theorem 3.6
and the discussion in the remainder of Sect. 3.5 with the σ -field F W,ξ being re-
placed by
∞
 
&  
F W̃ ,ξ
=σ σ W)s (fj ), s ≤ t ∪ σ (ξ ) .
j =1
148 3 Stochastic Differential Equations

Appendix: Compactness and Tightness of Measures

in C([0, T ], M )

In this book we are dealing with the space X = C([0, T ], M ), where (M , ρM ) is

a complete separable metric space. For this case, we obtain analytic conditions for
the family of measures {μn } ⊂ P(X ) to be tight.
The space X with the usual supremum metric ρ∞ (x, y) = sup0≤t≤T ρM (x(t) −
y(t)), x, y ∈ M , is itself a complete separable metric space. We denote by B(X )
the σ -field of subsets of X generated by the open subsets of X . The elements of
B(X ) will be referred to as Borel subsets of X . The object is to study proba-
bility measures on the measurable space (X , B(X )). We denote the class of all
probability measures on X by P(X ). We recall that a sequence of probability
measures {μn }∞ n=1 ⊂ P(X ) converges weakly to μ ∈ P(X ) if for every bounded
continuous function f : X → R,

f (x) μn (dx) → f (x) μ(dx).

X X

The above convergence defines a metric ρ on P(X ) (Prokhorov’s metric), such

that (P(X ), ρ) is a complete separable metric space (see [4], Theorem 5, p. 238).
The concept of tightness is essential in verifying that a family of measures {μn } ⊂
P(X ) possesses weak limit points.

Definition 3.4 A family of measures {μn }∞

n=1 ⊂ P(X ) is tight if for each ε > 0,
there exists a compact set Kε ⊂ X such that
 
μn Kεc < ε

for all n.

The following theorem is due to Prokhorov [65].

Theorem 3.16 A family of measures {μn }∞

n=1 ⊂ (P(X ), ρ) is relatively compact
if and only if it is tight.

The modulus of continuity of a function x ∈ C([0, T ], M ) is defined by

 
w(x, δ) = sup ρM x(t), x(s) . (3.106)
|t−s|<δ

Lemma 3.14 (Compactness) A set A ⊂ C([0, T ], M ) has a compact closure if and

only if the two following conditions are fulfilled:
(1) There exists a dense subset T  ⊂ [0, T ] such that for every t ∈ T  , the closure of
the set {x(t) : x ∈ A} is a compact subset of M ,
(2) limδ→0+ supx∈A w(x, δ) = 0.
Appendix: Compactness and Tightness of Measures in C([0, T ], M ) 149

Proof The compactness of A implies condition (1) easily. Condition (2) follows
from the fact that the function w(x, δ) is a continuous function of x, monotonically
decreasing to 0 as δ → 0. By Dini’s theorem, w(x, δ) converges to 0 uniformly on
a compact set A.
If conditions (1) and (2) hold, then let xn (t) be a sequence of elements in A.
We form a sequence t1 , t2 , . . . from all elements of T  and select a convergent
subsequence xnk1 (t1 ). From this subsequence we select a convergent subsequence
xnk2 (t2 ), etc. Using the diagonal method, we construct a subsequence xnk such that
xnk (t) converges for every t ∈ T  . Denote the limit by y(t) and let yk (t) = xnk (t).
Then yk (t) → y(t) pointwise on T  . For any ε > 0, let δ be chosen, so that
supx∈A w(x, δ) < ε/3. Let t1 , . . . , tN be such that the length of each of the inter-
vals [0, t1 ], [t1 , t2 ], . . . , [tN , T ] is less than δ. Then
   
sup ρM yk (t), yl (t) ≤ sup ρM yk (ti ), yl (ti )
0≤t≤T 1≤i≤N
    
+ sup ρM yk (ti ), yk (t) + ρM yl (ti ), yl (t) < ε
|t−ti |≤δ

for k, l sufficiently large.

Hence, yk is a Cauchy sequence in the ρ∞ metric. Therefore, ρ∞ (yk , y) → 0. 

Theorem 3.17 (Tightness) A family Pn of probability measures on C([0, T ], M ) is

tight if and only if the following two conditions are fulfilled,
(1) There exists a dense subset T  ⊂ [0, T ] such that for every t ∈ T  , the family of
measures Pn ◦ x(t)−1 on M is tight,
(2) limδ→0 lim supn Pn (w(x, δ) > ε) = 0 for all ε > 0.

Proof Conditions (1) and (2) follow easily from the tightness assumption and
Lemma 3.14. Conversely, let {tk }k∈Z+ be a countably dense set in [0, T ]. For each
k, we can find a compact set Ck ⊂ M and δk > 0 such that supn Pn (x(tk ) ∈ / Ck ) ≤
ε/2k+2 (by Prokhorov’s theorem) and, for some n0 , supn≥n0 Pn (w(x, δk ) > 1k ) ≤
ε/2k+2 . Let
 c
&  
Kε = / Ck ∪ w(x, δk ) > 1/k
x(tk ) ∈
k>n0
3  
= x(tk ) ∈ Ck ∩ w(x, δk ) ≤ 1/k .
k>n0

The set Kε satisfies the assumptions of Lemma 3.14; therefore it has a compact
closure. Moreover,
∞
  
sup Pn Kεc ≤ 2 · ε/2k+2 = ε ⇒ inf Pn (Kε ) ≥ 1 − ε.
n≥n0
k=0
n≥n0 
Chapter 4
Solutions by Variational Method

4.1 Introduction
The purpose of this chapter is to study both weak and strong solutions of nonlinear
stochastic partial differential equations, or SPDEs. The first work in this direction
was done by Viot [75]. Since then, Pardoux [62] and Krylov and Rozovskii [42]
studied strong solutions of nonlinear SPDEs. We will utilize the recent publication
by Prévôt and Röckner [64] to study strong solutions.
In all these publications, the SPDEs are recast as evolution equations in a Gelfand
triplet,
V → H → V ∗ ,
where H is a real separable Hilbert space identified with its dual H ∗ . The space
V is a Banach space embedded continuously and densely in H . Then for its dual
space V ∗ , the embedding H → V ∗ is continuous and dense, and V ∗ is necessarily
separable. The norms are denoted by  · V , and similarly for the spaces H and V ∗ .
The duality on V × V ∗ is denoted by ·, · , and it agrees with the scalar product in
H , i.e., v, h = v, h H if h ∈ H .
By using the method of compact embedding of Chap. 3, the ideas from [36], and
the stochastic analogue of Lion’s theorem from [42], we show the existence of a
weak solution X in the space C([0, T ], H ) ∩ L∞ ([0, T ], H ) ∩ L2 ([0, T ] × , V )
such that
  2 
E sup X(t)H < ∞, (4.1)
t∈[0,T ]

under the assumption that the injection V → H is compact without using the as-
sumption of monotonicity. In the presence of monotone coefficients, as in [36], we
obtain a unique strong solution using pathwise uniqueness.
The approach in [64] is to consider monotone coefficients. Under weakened as-
sumptions on growth and without assuming compact embedding, using again the
stochastic analogue of Lion’s theorem, a unique strong solution is produced in
C([0, T ], H ) ∩ L2 ([0, T ] × , V ), which again satisfies (4.1). We will present this
method in Sect. 4.3.

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 151

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_4, © Springer-Verlag Berlin Heidelberg 2011
152 4 Solutions by Variational Method

Now, let K be another real separable Hilbert space, Q ∈ L1 (K) be a symmetric

nonnegative definite operator, and {Wt , t ≥ 0} be a K-valued Q-Wiener process
defined on a filtered probability space (, F , {Ft }t≥0 , P ).
Consider the variational SDE
   
dX(t) = A t, X(t) dt + B t, X(t) dWt (4.2)
with coefficients
A : [0, T ] × V → V ∗ and B : [0, T ] × V → L2 (KQ , H )
and H -valued F0 -measurable initial condition ξ0 ∈ L2 (, H ). Recall that
  ∗
B(t, v)Q1/2 B(t, v)Q1/2 ∈ L1 (H ).
Let us now define different notions of a solution to (4.2). Such solutions are often
called variational solutions.

Definition 4.1 An H -valued stochastic process X(t) defined on a given filtered

probability space (, F , {Ft }t≤T , P ) is a strong solution of (4.2) if
T 
(1) E X(t)2 dt < ∞,
V
0 T 
 
(2) P A t, X(t)  ∗ dt < ∞ = 1,
V
0
t  
(3) B s, X(s) dWs is a square-integrable H -valued martingale,
0
t   t  
(4) X(t) = ξ0 + A s, X(s) ds + B s, X(s) dWs P -a.s.
0 0
The integrants A(t, X(t)) and B(t, X(t)) are evaluated at a V -valued Ft -measurable
version of X(t) in L2 ([0, T ] × , V ).
A weak solution of (4.2) is a system ((, F , {Ft }t≤T , P ), W, X), where Wt is
a K-valued Q-Wiener process with respect to the filtration {Ft }t≤T , Xt is an H -
valued process adapted to Ft and satisfies conditions (1)–(3) above,
t   t  
X(t) = X(0) + A s, X(s) ds + B s, X(s) dWs , P -a.s.,
0 0

and P ◦ (X(0))−1 = L (ξ0 ).

Unlike in the case of a strong solution, the filtered probability space and the
Wiener process are part of a weak solution and are not given in advance.

4.2 Existence of Weak Solutions Under Compact Embedding

In this section we study the existence of weak solutions using mainly the techniques
in [36] and [22]. We assume that in the Gelfand triplet V → H → V ∗ , V is a real
separable Hilbert space.
4.2 Existence of Weak Solutions Under Compact Embedding 153

Let the coefficients A and B satisfy the following joint continuity and growth
conditions:
(JC) (Joint Continuity) The mappings

(t, v) → A(t, v) ∈ V ∗ and (t, v) → B(t, v)QB ∗ (t, v) ∈ L1 (H ) (4.3)

are continuous.
For some constant θ ≥ 0,
(G-A)
   
A(t, v)2 ∗ ≤ θ 1 + v2 , v ∈ V. (4.4)
V H
(G-B)
   
B(t, v)2 ≤ θ 1 + v2H , v ∈ V. (4.5)
L2 (KQ ,H )

In addition, we will impose the following coercivity condition on A and B:

(C) There exist constants α > 0, γ , λ ∈ R such that for v ∈ V ,
   2
2 A(t, v), v + B(t, v)L (K ,H ) ≤ λv2H − αv2V + γ . (4.6)
2 Q

Finally, we will require that the initial condition of (4.2) satisfies

(IC)
  2 
E ξ0 2H ln 3 + ξ0 2H < c0 (4.7)
for some constant c0 . It will become clear that this property will be used to
ensure the uniform integrability of the squared norm of the approximate solu-
tions.
We will first consider a finite-dimensional SDE related to the infinite-dimensional
equation (4.2). Let {ϕj }∞ j =1 ⊂ V be a complete orthonormal system in H , and
∞
{fk }k=1 be a complete orthonormal system in K. Define the map Jn : Rn → V
by

n
Jn (x) = xj ϕj = u ∈ V
j =1

and the coefficients (a n (t, x))j : [0, T ] × Rn → Rn , (bn (t, x))i,j : [0, T ] × Rn →
Rn × Rn , and (σ n (t, x))i,j : [0, T ] × Rn → Rn × Rn and the initial condition ξ0n by
 n   
a (t, x) j = ϕj , A(t, Jn x) , 1 ≤ j ≤ n,
 n   
b (t, x) i,j = Q1/2 B ∗ (t, Jn x)ϕi , fj K , 1 ≤ i, j ≤ n,
 n    T  (4.8)
σ (t, x) i,j = bn (t, x) bn (t, x) i,j ,
 n
ξ0 j = ξ0 , ϕj H .
154 4 Solutions by Variational Method

Note that

 n  
n
 n   
σ (t, x) i,j = b (t, x) i,k bn (t, x) j,k
k=1


n
   
= Q1/2 B ∗ (t, Jn x)ϕi , fk K
Q1/2 B ∗ (t, Jn x)ϕj , fk K
.
k=1

Lemma 4.1 The growth conditions (4.4) and (4.5) assumed for the coefficients A
and B imply the following growth conditions on a n and bn :
 n   
a (t, x)2 n ≤ θn 1 + x2 n , (4.9)
R R
    T   
tr σ n (t, x) = tr bn (t, x) bn (t, x) ≤ θ 1 + x2Rn . (4.10)

The coercivity condition (4.6) implies that

    T 
2 a n (t, x), x Rn + tr bn (t, x) bn (t, x)
    ∗ 
≤ 2 Jn x, A(t, Jn x) + tr B(t, Jn x)Q1/2 B(t, Jn x)Q1/2
≤ λJn x2H − αJn x2V + γ . (4.11)

In particular, for a large enough value of θ , the coercivity condition (4.6) implies
that
    T   
2 a n (t, x), x Rn + tr bn (t, x) bn (t, x) ≤ θ 1 + x2Rn . (4.12)
The constant θn depends on n, but θ does not.
The distribution μn0 of ξ0n on Rn satisfies
 2    2 2 
E ξ0n Rn ln 3 + ξ0n Rn < c0 . (4.13)

Exercise 4.1 Prove Lemma 4.1. In addition, show that for k ≥ n and x ∈ Rk , the
following estimate holds true:


n
 k  2  
a (t, x) j ≤ θn 1 + x2Rk . (4.14)
j =1

We will need the following result, Theorem V.3.10 in [17], on the existence of a
weak solution. Consider the following finite-dimensional SDE,
   
dX(t) = a t, X(t) dt + b t, X(t) dBtn , (4.15)

with an Rn -valued F0 -measurable initial condition ξ0n . Here Btn is a standard Brow-
nian motion in Rn .
4.2 Existence of Weak Solutions Under Compact Embedding 155

Theorem 4.1 There exists a weak solution to (4.15) if a : [0, ∞] × Rn → Rn and b :

[0, ∞] × Rn → Rn ⊗ Rn are continuous and satisfy the following growth conditions:
   
b(t, x)2 n ≤ K 1 + x2 n ,
L (R ) R
    (4.16)
x, a(t, x) Rn ≤ K 1 + x2Rn

for t ≥ 0, x ∈ Rn , and some constant K.

We will use the ideas developed in [70], Sect. 1.4, for proving the compactness of
probability measures on C([0, T ], Rn ). The method was adapted to the specific case
involving linear growth and coercivity conditions in [36]. Our first step in proving
the existence result in the variational problem will be establishing the existence and
properties of finite-dimensional Galerkin approximations in the following lemma.

Lemma 4.2 Assume that the coefficients A and B of (4.2) satisfy the assumptions
of joint continuity (4.3), growth (4.4), (4.5), and coercivity (4.6) and that the initial
condition ξ0 satisfies (4.7). Let a n , bn , and ξ0n be defined as in (4.8), and Btn be an
n-dimensional standard Brownian motion. Then the finite-dimensional equation
   
dX(t) = a n t, X(t) dt + bn t, X(t) dBtn (4.17)

with the initial condition ξ0n has a weak solution X n (t) in C([0, T ], Rn ). The laws
μn = P ◦ (X n )−1 have the property that for any R > 0,
    !
sup μn x ∈ C [0, T ], Rn : sup x(t)Rn > R
n 0≤t≤T
4   2
≤ 2c0 eC(θ)T 1 + R 2 ln 3 + R 2 (4.18)

and that
  2    2 
sup 1 + x(t)Rn ln ln 3 + x(t)Rn μn (dx) < C (4.19)
C([0,T ],Rn ) 0≤t≤T

for some constant C.

Proof Since the coefficients a n and bn satisfy conditions (4.9) and (4.10), we can
use Theorem 4.1 to construct a weak solution X n (t) to (4.17) for every n. Let
   2
f (x) = 1 + x2Rn ln 3 + x2Rn , x ∈ Rn .

Define for g ∈ C02 (Rn ) the differential operator

 n  
n
∂g   1   ∂ 2g
n n
 
Lt g (x) = (x) a n (t, x) i + (x) σ n (t, x) i,j .
∂xi 2 ∂xi ∂xj
i=1 i=1 j =1
156 4 Solutions by Variational Method

We leave as an exercise, see Exercise 4.2, to prove that

    
fx (x)2 n ≤ Cf (x) ln 3 + x2 n 2 (4.20)
R R

and that the coercivity condition (4.6) implies

Lnt f (x) ≤ Cf (x). (4.21)

Using Itô’s formula for the function f (x), we have

    t  
f X n (t) = f X n (0) + Lns f X n (s) ds + Mt , (4.22)
0

where Mt is a local martingale. Define a stopping time

  
τR = inf t : X n (t)Rn > R or T . (4.23)

Then Mt∧τR is a square-integrable martingale with the increasing process

t∧τR     
M t∧τR = bn s, X n (s) fx X n (s) 2 n ds
R
0
t∧τR   2       2
≤ Cθ 1 + X n (s)Rn f X n (s) ln 3 + X n (s)Rn ds
0
t∧τR  
= Cθ f 2 X n (s) ds
0
   t∧τR  
≤ Cθ sup f X n (s) f X n (s) ds,
0≤s≤t∧τR 0

where we have applied (4.20).

Using Burkholder’s inequality, Theorem 3.28 in [38], we calculate
   1/2
E sup |Mt∧τR | ≤ 4E M t∧τR
0≤s≤t
  1/2 
 n 1/2 t∧τR  
≤ 4(Cθ ) 1/2
E sup f X (s ∧ τR ) f X n (s) ds
0≤s≤t 0
  1/2 
 1/2 t  
≤E sup f X n (s ∧ τR ) 16Cθ sup f X (r ∧ τR ) ds
n
0≤s≤t 0 0≤r≤s
  
1  n  t  
≤ E sup f X (s ∧ τR ) + 16Cθ sup f X n (r ∧ τR ) ds .
2 0≤s≤t 0 0≤r≤s

Then, by (4.21) and (4.22),

    s  
f X n (s ∧ τR ) ≤ f X n (0) + C f X n (r ∧ τR ) dr + Ms∧τR ,
0
4.2 Existence of Weak Solutions Under Compact Embedding 157

and hence,

  t  
E sup f X n (s ∧ τR ) ≤ 2c0 + (2C + 16Cθ )E sup f X n (r ∧ τR ) ds.
0≤s≤t 0 0≤r≤s

By applying Gronwall’s lemma, we obtain the bound

 
E sup f X n (s ∧ τR ) ≤ 2c0 e(2C+16Cθ)T , 0 ≤ t ≤ T.
0≤s≤t

Then
       
P sup X n (s)Rn > R ≤ E sup f X n (s ∧ τR ) /f (R)
0≤s≤t 0≤s≤t
4   2
≤ 2c0 eC(θ)T 1 + R 2 ln 3 + R 2 ,

proving (4.18). To prove (4.19), denote g(r) = (1 + r 2 ) ln ln(3 + r 2 ), r ≥ 0. Since g

is increasing, we have

  2    2 
sup 1 + x(t)Rn ln ln 3 + x(t)Rn μn (dx)
C([0,T ],Rn ) 0≤t≤T
  
= sup g x(s)Rn μn (dx)
C([0,T ],Rn ) 0≤t≤T
∞     
= μn sup g x(s)Rn > p dp
0 0≤t≤T
∞    
= ln ln 3 + μn sup x(s)Rn > g −1 (p) dp
ln ln 3 0≤t≤T
∞    
≤ ln ln 3 + μn sup x(s)Rn > r g  (r) dr
0 0≤t≤T
∞ g  (r)
≤ ln ln 3 + 2c0 eC(θ)T dr < ∞,
0 (1 + r 2 )(ln(3 + r 2 ))2

with the very last inequality being left to prove for the reader in Exercise 4.3. 

Exercise 4.2 Prove (4.20) and (4.21).

Exercise 4.3 With the notation of Theorem 4.2, prove that

∞ g  (r)
dr < ∞.
0 (1 + r 2 )(ln(3 + r 2 ))2

We will need the following lemma from [36] (see also Sect. 1.4 in [70]).
158 4 Solutions by Variational Method

Lemma 4.3 Consider the filtered probability space (C([0, T ], Rn ), B, {Ct }0≤t≤T ,
P ), where B is the Borel σ -field, and Ct is the σ -field generated by the cylin-
ders with bases over [0, t]. Let the coordinate process mt be a square-integrable
Ct -martingale with quadratic variation m t satisfying
 
m t − m s = tr m t − m s ≤ β(t − s)

for some constant β and all 0 ≤ s < t ≤ T . Then for all ε, η > 0, there exists δ > 0,
depending possibly on β, ε, η, and T , but not on n, such that
 
P sup mt − ms Rn > ε < η.
|t−s|<δ

Proof Define the sequence of stopping times


τ0 = 0, τj = inf τj −1 < s ≤ T : ms − mτj −1 Rn > ε/4 or τj = T , j ≥ 1.

Let N = inf{j : τj = T } and α = inf{τj − τj −1 : 0 ≤ j ≤ N }. It is left as an exercise

to show that
!
sup mt − ms Rn > ε ⊂ {α < δ}. (4.24)
|t−s|<δ

Therefore, for any positive integer k,

 
P sup mt − ms Rn > ε ≤ P (α ≤ δ)
|t−s|<δ
≤ P (τj − τj −1 < δ for some j ≤ N, j ≤ k)
+ P (τj − τj −1 < δ for some j ≤ N, j > k)
≤ P (τj − τj −1 < δ for some j ≤ k) + P (N > k). (4.25)

Let from now on ε, η > 0 be fixed but arbitrary.

First, for any stopping time τ , consider the sub-σ -field of B

Cτ = {A ∈ B : A ∩ {τ ≤ t} ∈ Ct , for all 0 ≤ t ≤ T } (4.26)

and a regular conditional probability distribution of P given Cτ , denoted by

P τ (A, ω). Since C([0, T ], Rn ) is a Polish space, a regular conditional probabil-
ity distribution exists (see [25], Vol. I, Chap. I, Theorem 3 and [70] Sect. 1.3 for a
specific construction). The important property of the measure P τ is that it preserves
the martingale property. Specifically, if mt is a Ct -martingale with respect to the
measure P , then it also is a Ct -martingale for t ≥ τ (ω) with respect to each condi-
tional distribution P τ (·, ω), except possibly for ω outside of a set E of P -measure
zero.
It follows that for t ≥ 0, the process mt − mt∧τ is a Ct -martingale with respect
to P τ (·, ω), except possibly for ω ∈ E. This is left to the reader to verify as an
exercise (see Exercise 4.5), and more general results can be found in Sects. 1.2
and 1.3 of [70].
4.2 Existence of Weak Solutions Under Compact Embedding 159

We are now going to find bounds for the probabilities in the last line of (4.25).
We have, P -a.s.,

P (τj − τj −1 < t|Fτj −1 ) = P τj (τj − τj −1 < t)

 
= P τj sup ms+τj −1 − mτj −1 Rn > ε/4
0≤s≤t
 
= P τj sup ms − ms∧τj −1 Rn > ε/4
0≤s≤t+τj −1

64  
≤ E mt+τj −1 − mτj −1 2Rn |Fτj
ε2
64  
≤ 2 E m t+τj −1 − m τj −1 |Fτj
ε
64βt
≤ 2 ,
ε
where we have used properties of the regular conditional probability distribution in
the last two lines. Next, for t > 0, P -a.s.,
 
E e−(τj −τj −1 ) |Fτj −1
≤ P (τj − τj −1 < t|Fτj −1 ) + e−t P (τj − τj −1 ≥ t|Fτj −1 )
 
≤ e−t + 1 − e−t P (τj − τj −1 < t|Fτj −1 )
 
≤ e−t + 1 − e−t 64βt/ε 2 = λ < 1

for t small enough. Hence,

     
E e−τj Fτj −1 = e−τj −1 E e−(τj −τj −1 ) Fτj −1

≤ λe−τj −1 ≤ · · · ≤ λj ,

so that
 
P (N > k) = P (τk < T ) ≤ P e−τk > e−T ≤ eT λk < η/2
for k large enough, depending on T , λ, and η. Finally,


k
P (τj − τj −1 < δ for some j ≤ k) ≤ P (τj − τj −1 < δ for some j ≤ k)
j =1
 
≤ k 64βδ/ε 2 < η/2

for δ small enough, depending on k, β, ε, and η. Combining the last two inequalities
proves the lemma. 

Exercise 4.4 Prove (4.24).

160 4 Solutions by Variational Method

Exercise 4.5 Let (, F , {Ft }0≤t≤T , P ) be a filtered probability space, with  a
Polish space and F its Borel σ -field. Assume that Mt is an Rn -valued continuous
martingale and τ is a stopping time.
Show that Mt − Mt∧τ is an Ft -martingale with respect to the conditional prob-
ability P τ (·, ω), except possibly for ω in a set E of P -measure zero.
Hint: prove that
 
     
mt ω dP τ ω , ω dP ω
B∩{τ ≤s} A
 
     
= ms ω dP τ ω , ω dP ω
B∩{τ ≤s} A

for 0 ≤ s ≤ t ≤ T and A ∈ Fs , B ∈ Fτ , and use the fact that Fs is countably

generated. Conclude that for s ≥ τ (ω ), outside possibly of a set Es,t of P -measure
zero,
τ  
E P (·,ω) mt | Fs = ms P τ (ω)-a.s.
Choose a dense countable subset D of [0, T ] and show that the family {mt Rn , t ∈
D} is uniformly integrable.

Now we will use the compact embedding argument, previously discussed

in Sect. 3.8.

Theorem 4.2 Let the coefficients A and B of (4.2) satisfy conditions (4.3), (4.4),
(4.5), and (4.6). Consider the family of measures μn∗ on C([0, T ], V ∗ ) with support
in C([0, T ], H ), defined by
 
  
n
 
μn∗ (Y ) = μn x ∈ C [0, T ], R :
n
xi (t)ϕi ∈ Y , Y ⊂ C [0, T ], V ∗ ,
i=1

where μn are the measures constructed in Lemma 4.2. Assume that the embed-
ding H → V ∗ is compact. Then the family of measures {μn∗ }∞
n=1 is tight on
C([0, T ], V ∗ ).

Proof We will use Theorem 3.17. Denote by BC([0,T ],H ) (R) ⊂ C([0, T ], H ) the
closed ball of radius R centered at the origin. By the definition of measures μn∗ and
Lemma 4.2, for any η > 0, we can choose R > 0 such that
  n  
 c     
 
n
μ∗ BC([0,T ],H ) (R) = μ x ∈ C [0, T ], R : sup 
n n
xi (t)ϕi  > R
0≤t≤T  i=1 
H
    !
= μn x ∈ C [0, T ], Rn : sup x(t)Rn > R < η.
0≤t≤T
4.2 Existence of Weak Solutions Under Compact Embedding 161

Denote the closed ball of radius R centered at zero in H by BH (R). Then its closure
V∗
in V ∗ , denoted by BH (R) , is a compact subset of V ∗ , and we have

 V ∗
μn∗ ◦ x(t)−1 BH (R) ≥ 1 − η, 0 ≤ t ≤ T,

fulfilling the first condition for tightness in Theorem 3.17.

V∗
Again, using the compactness of BH (R) in V ∗ , for any ε > 0, we can find an
index n0 ≥ 1 such that
 ∞ 
  
 
 xj ϕj  < ε/4 if xH ≤ R. (4.27)
 
j =n0 +1 V∗

Since the embedding H → V ∗ is continuous and linear, we have xV ∗ ≤ CxH

for x ∈ H and some constant C, independent of η and R, so that
n  n 
0  0 
   
 (xj − yj )ϕj  ≤ C  (xj − yj )ϕj  .
   
j =1 V∗ j =1 H

Recall the modulus of continuity (3.106) and indicate the space, e.g., V ∗ , in the
subscript in wV ∗ (x, δ) if the V ∗ norm is to be used. Then, with BC([0,T ],Rn ) (R)
denoting the closed ball with radius R centered at the origin in C([0, T ], Rn ) and
n > n0 ,
  
μn∗ x ∈ C [0, T ], V ∗ : x ∈ BC([0,T ],H ) (R) , wV ∗ (x, δ) > ε
  n  
 
≤ μ x ∈ BC([0,T ],Rn ) (R) : wV ∗
n
x(·) j ϕj , δ > ε
j =1
 n 
0
     
 
≤μ n
x ∈ BC([0,T ],Rn ) (R) : sup  x(t) j − x(s) j ϕj 
0≤s,t≤T  j =1

V∗
|s−t|<δ
 n  
       
 
+ sup  x(t) j − x(s) j ϕj  >ε
0≤s,t≤T  j =n0 +1

V∗
|s−t|<δ
 n  
0
     
 
≤ μn x ∈ BC([0,T ],Rn ) (R) : C sup  x(t) j − x(s) j ϕj  + ε/4 > ε
0≤s,t≤T  j =1

H
|s−t|<δ
  1/2 

n0
    2
=μ n
x ∈ BC([0,T ],Rn ) (R) : sup x(t) j − x(s) j > 3ε/(4C) .
0≤s,t≤T j =1
|s−t|<δ
162 4 Solutions by Variational Method

For the stopping time τR = inf{0 ≤ t ≤ T : x(t) ∈

/ BC([0,T ],Rn ) (R)} or T , the
Rn -valued martingale
t∧τR  
t (x) = x(t ∧ τR ) −
mR a n s, x(s) ds
0

has the quadratic variation process given by

 R  t∧τR    T
m (x) t = bn s, x(s) (bn s, x(s) ds
0

with the function tr(bn (s, x(s))(bn (s, x(s))T ) bounded on bounded subsets of Rn
uniformly relative to the variable s, due to condition (4.10). Hence, for t ≥ s,
 R   R      
m (x) t − m (x) s = tr mR (x) t − mR (x) s ≤ β(R)(t − s)

with the constant β(R) not depending on n. Now, by Lemma 4.3, we have
     
μn x ∈ C [0, T ], Rn : wRn mR (x), δ > ε/(2C) < η (4.28)

for sufficiently small δ independent of n.

Let n ≥ n0 and sup0≤t≤T x(t)Rn ≤ R. Using (4.14), we have
 1/2

n0
 n   2  
a t, x(t) j ≤ θn0 1 + R 2 ;
j =1

hence, for a sufficiently small constant δ, we can write

n  2 1/2
0 t  
n
a t, x(t) j ≤ ε/(4C) whenever |t − s| < δ.
j =1 s

Also, whenever sup0≤t≤T x(t)Rn ≤ R,

t  
mnt = x(t) − a n s, x(s) ds = mR
t .
0

We can continue our calculations as follows:

 n 1/2 
 0
    2
μn x ∈ BC([0,T ],Rn ) (R) : sup x(t) j − x(s) j > 3ε/(4C)
0≤s,t≤T j =1
|s−t|<δ

 
≤ μ x ∈ BC([0,T ],Rn ) (R) : wRn mn , δ
n

 ·  
 
+ wRn0 a n s, x(s) ds, δ > 3ε/(4C)
0
4.2 Existence of Weak Solutions Under Compact Embedding 163
  
≤ μn x ∈ BC([0,T ],Rn ) (R) : wRn mn , δ > ε/(2C)
  
≤ μn wRn mR , δ > ε/(2C) ≤ η.

Summarizing, for any ε, η > 0 and sufficiently small δ > 0, there exists n0 such that
for n > n0 ,
  
μn∗ x ∈ C [0, T ], V ∗ : wV ∗ (x, δ) > ε
 c  
≤ μn∗ BC([0,T ],H ) (R) + μn∗ x ∈ BC([0,T ],H ) (R) : wV ∗ (x, δ) > ε
≤ 2η,

concluding the proof. 

We will now summarize the desired properties of the measures μn and μn∗ .

Corollary 4.1 Let X n (t) be solutions to (4.17), μn be their laws in C([0, T ], Rn ),

and μn∗ be the measures induced in C([0, T ], V ∗ ) as in Theorem 4.2. Then for some
constant C independent of n,
 2   2 
sup x(t)Rn ln ln 3 + x(t)Rn μn (dx) < C, (4.29)
C([0,T ],Rn ) 0≤t≤T

implying the uniform integrability of X n 2Rn . The  · H norm of x(t) satisfies the
following properties:
 2   2 
sup x(t)H ln ln 3 + x(t)H μn∗ (dx)
C([0,T ],V ∗ ) 0≤t≤T
  2   2 
=E sup Jn X n (t)H ln ln 3 + Jn X n (t)H < C. (4.30)
0≤t≤T

For a cluster point μ∗ of the tight sequence μn∗ ,

 2
sup x(t)H μ∗ (dx) < C. (4.31)
C([0,T ],V ∗ ) 0≤t≤T

There exists a constant C such that for any R > 0,

    !
μ∗ x ∈ C [0, T ], V ∗ : sup x(t)H > R < C/R 2 , (4.32)
0≤t≤T

and also,
  !
μ∗ x ∈ C [0, T ], V ∗ : sup x(t)H < ∞ = 1. (4.33)
0≤t≤T

Finally, the  · V norm of x(t) satisfies

T 
x(t)2 dt μn (dx) < C (4.34)
V ∗
C([0,T ],V ∗ ) 0
164 4 Solutions by Variational Method

and
T 
x(t)2 dt μ∗ (dx) < ∞. (4.35)
V
C([0,T ],V ∗ ) 0

Proof Property (4.29) is just (4.19) and inequality (4.30) is just a restatement
of (4.29)
To prove (4.31), assume, using the Skorokhod theorem, that Jn Xn → X a.s. in
C([0, T ], V ∗ ). We introduce the function αH : V ∗ → R by

αH (u) = sup v, u , v ∈ V , vH ≤ 1 .

Clearly αH (u) = uH if u ∈ H , and it is a lower semicontinuous function as a

supremum of continuous functions. For u ∈ V ∗ \ H , αH (u) = +∞ (Exercise 4.6).
Thus, we can extend the norm  · H to a lower semicontinuous function on V ∗ .
By the Fatou lemma and (4.29),
 2
sup x(t)H μ∗ (dx)
C([0,T ],V ∗ ) 0≤t≤T
  2 
=E sup X(t)H
0≤t≤T
  2 
≤ E lim inf sup Jn X n (t)H
n→∞ 0≤t≤T
  2 
≤ lim inf E sup Jn X n (t)H
n→∞ 0≤t≤T
 2
= lim inf sup x(t)H μn∗ (dx) < C.
n→∞ C([0,T ],V ∗ ) 0≤t≤T

Property (4.32) follows from the Markov inequality, and (4.33) is a consequence
of (4.31). To prove (4.34), we apply the Itô formula and (4.11) to obtain that
 2  2 t   
E Jn X n (t)H = E Jn ξ0n H + 2E a n s, X n (s) , X n (s) Rn ds
0
t     T 
+E tr bn s, X n (s) bn s, X n ds
0
t  2
≤ EJn ξ0 2H + λ E Jn X n (s)H ds
0
t  2
−α E Jn X n (s)V ds + γ .
0

Using the bound in (4.30), we conclude that

T  2
sup E Jn X n (t)V dt < ∞.
n 0
4.2 Existence of Weak Solutions Under Compact Embedding 165

Finally, we can extend the norm  · V to a lower semicontinuous function on V ∗

by introducing the lower semicontinuous function

αV (u) = sup v, u , v ∈ V , vV ≤ 1 ,

since αV (u) = uV if u ∈ V and, for u ∈ V ∗ \ V , αV (u) = +∞. Now (4.35) fol-
lows by the Fatou lemma. 

Exercise 4.6 Justify the statements about αH made in the proof of Corollary 4.1.

As we stated in the introduction, in order to identify the solution, weak or strong,

as a continuous H -valued process, we will need the following deep result, which is
a stochastic extension of a lemma of Lions. This result is included in [42], Theo-
rem I.3.1, and a detailed proof is given in [64], Theorem 4.2.5.

Theorem 4.3 Let X(0) ∈ L2 (, F0 , P , H ), and Y ∈ L2 ([0, T ] × , V ∗ ) and

Z ∈ L2 ([0, T ] × , L2 (KQ , H )) be both progressively measurable. Define the con-
tinuous V ∗ -valued process
t t
X(t) = X(0) + Y (s) ds + Z(s) dWs , t ∈ [0, T ].
0 0

If for its dt ⊗ P -equivalence class X̂, we have X̂ ∈ L2 ([0, T ] × , V ), then X is an

H -valued continuous Ft -adapted process,
  2 
E sup X(t)H < ∞ (4.36)
t∈[0,T ]

and the following Itô formula holds for the square of its H -norm P -a.s.:

    t    2 
X(t)2 = X(0)2 + 2 X̄(s), Y (s) + Z(s)L ds
H H 2 (KQ ,H )
0
t 
+2 X(s), Z(s) dWs H
, t ∈ [0, T ] (4.37)
0

for any V -valued progressively measurable version X̄ of X̂.

Remark 4.1 Note that the process

X̄(t) = 1{αV (X̂(t))<∞} X̂(t)

serves as a V -valued progressively dt ⊗ dP -measurable version of X̂.

We are now ready to formulate the existence theorem.

166 4 Solutions by Variational Method

Theorem 4.4 Let V → H → V ∗ be a Gelfand triplet of real separable Hilbert

spaces with compact inclusions. Let the coefficients A and B of (4.2) satisfy condi-
tions (4.3), (4.4), (4.5), and (4.6). Let the initial condition ξ0 be an H -valued random
variable satisfying (4.7). Then (4.2) has a weak solution X(t) in C([0, T ], H ) such
that
  2 
E sup X(t)H < ∞ (4.38)
0≤t≤T

and
T 
E X(t)2 dt < ∞. (4.39)
V
0

Proof Let X n (t) be solutions to (4.17), μn be their laws in C([0, T ], Rn ), and μn∗
be the measures induced in C([0, T ], V ∗ ) as in Theorem 4.2, with a cluster point
μ∗ . We need to show that μ∗ is the law of a weak solution to (4.2). Again, using the
Skorokhod theorem, assume that Jn X n (t) and X(t) are processes with laws μn∗ and
μ∗ , respectively, with Jn X n → X P -a.s. By (4.30) and (4.33)–(4.35), Jn X n and X
are P -a.s. in C([0, T ], V ∗ ) ∩ L∞ ([0, T ], H ) ∩ L2 ([0, T ], V ). Denote by {ϕj }∞
j =1 ⊂
V a complete orthogonal system in V , which is an ONB in H . Note that such a
system always exists, see Exercise 4.7. Then, the vectors ψj = ϕj /ϕj V form an
ONB in V . For x ∈ C([0, T ], V ∗ ) ∩ L∞ ([0, T ], H ) ∩ L2 ([0, T ], V ), consider
t  
Mt (x) = x(t) − x(0) − A s, x(s) ds.
0

Using (4.4) and (4.30), we have, for any v ∈ V and some constant C,

  2   2 
v, A s, x(s) ln ln 3 + x(s)H μn∗ (dx)

      
≤ A s, x(s) 2 ∗ v2 ln ln 3 + x(s)2 μn (dx)
V V H ∗

  2    2 
≤ θ 1 + x(s)H ln ln 3 + x(s)H v2V μn∗ (dx)

≤ Cθ v2V (4.40)

and, in a similar fashion, adding (4.31) to the argument,

  2
v, A s, x(s) μ∗ (dx) ≤ Cθ v2V . (4.41)

Properties (4.38) and (4.39) are just restatements of (4.31) and (4.35). By involv-
ing (4.41) we conclude that the continuous process v, Mt (·) is μ∗ -square inte-
grable. We will now show that for any v ∈ V , s ≤ t, and any bounded function gs on
C([0, T ], V ∗ ) which is measurable with respect to the cylindrical σ -field generated
4.2 Existence of Weak Solutions Under Compact Embedding 167

by the cylinders with bases over [0, s],

  
v, Mt (x) − Ms (x) gs (x) μ∗ (dx) = 0, (4.42)

i.e., that v, Mt (·) ∈ MT2 (R) (continuous square-integrable real-valued martin-
gales). First, assume that gs is continuous and extend the result to the general case
by the monotone class theorem
 (functional form).
Let for v ∈ V , v m = mj =1 v, ψj V ψj . Then, as m → ∞,

  
gs (x) v − v m , Mt (x)  μ∗ (dx) → 0

by uniform integrability, since |gs (x)|Mt (x)V ∗ v − v m V → 0. Hence,

    
gs (x) v, Mt (x) − v, Ms (x)
   
− gs (x) v m , Mt (x) − v m , Ms (x)  μ∗ (dx) → 0. (4.43)

By the choice of the vectors ϕj and ψj , we have, for x n (t) = (x1 (t), . . . , xn (t)) ∈
Rn ,
 n  n∧m
 m   
v , Jn x n (t) = xj (t)ϕj = xj (t)v, ϕj H .
j =1 j =1

For n ≥ m, the process

       t  
v m , Mt Jn x n (·) = v m , Jn x n (t) H − v m , x(0) H − v m , A s, Jn x n (s) ds
0

m  
   n    t  
= v m , ϕj H
x (t) j − x(0) j − a n s, x n (s) j ds
j =1 0

is a martingale relative to the measure μn . Hence, the above and the uniform inte-
grability of Mt (Jn X n ), v (that follows from (4.30) and (4.40)) imply that
  
gs (x) v m , Mt (x) − Ms (x) μ∗ (dx)
  
= E gs (X) v m , Mt (X) − Ms (X)
      
= lim E gs X n v m , Mt Jn X n − Ms Jn X n
n→∞
        
= lim gs Jn x n v m , Mt Jn x n − Ms Jn x n μn dx n = 0.
n→∞

The above conclusion, together with (4.43), ensures (4.42). Next, we find the in-
creasing process for the martingale v, Mt (x) . We begin with some estimates. For
168 4 Solutions by Variational Method

x, v ∈ V , we have
       ∗   2
v, B s, x(s) Q1/2 B s, x(s) Q1/2 v ≤ v2H B(s, x)L .
2 (KQ ,H )

Hence,
       ∗ 
v, B s, x(s) Q1/2 B s, x(s) Q1/2 v μn∗ (dx)
 
≤ v2H θ 1 + x2H μn∗ (dx)

≤ θ (1 + C)v2H (4.44)
by (4.30), and by (4.31)
       ∗ 
v, B s, x(s) Q1/2 B s, x(s) Q1/2 v μ∗ (dx) ≤ θ (1 + C)v2H .

As a consequence, we obtain that

 t 
       ∗ 
 v m , B u, x(u) Q1/2 B u, x(u) Q1/2 v m

s

    1/2    1/2 ∗  
− v, B u, x(u) Q B u, x(u) Q v gs (x) du μ∗ (dx)
    

≤ 2sup gs (x) T θ (1 + C)v m − v H vH < ε/2 (4.45)
x

for m sufficiently large. Next, observe that

 2  2 
v m , Mt (x) − v, Mt (x) gs (x) μ∗ (dx)
    m 2 1/2   2 1/2

≤ sup gs (x)  v − v, Mt (x) μ∗ (dx) v m + v, Mt (x) μ∗ (dx)
x
< ε/2, (4.46)

since by (4.31) and (4.41) the integrals above are bounded by Dv m − v2V and
Dv m + v2V , respectively, for some constant D.
By the uniform integrability of v, Mt (Jn X n ) 2 (ensured by (4.30) and (4.40)),
we have
 m 2  2 
v , Mt (x) − v m , Ms (x) gs (x) μ∗ (dx)
 2  2  
=E v m , Mt (X) − v m , Ms (X) gs (X)
  2   2   
= lim E v m , Mt Jn X n − v m , Ms Jn X n gs J n X n
n→∞
4.2 Existence of Weak Solutions Under Compact Embedding 169
 m   2 
 t    
= lim E X n
(t) − ξ0n − a u, X n (u) du v, ϕj
n
H gs J n X n
n→∞ 0
j =1 j
 m   2 
 s    
− lim E X n (s) − ξ0n − a u, X (u) du v, ϕj
n n
H gs J n X n
n→∞ 0
j =1 j
  m  
t     T   
= lim E bn u, X n (u) bn u, X n (u) jj
v, ϕj 2
H gs J n X
n
du
n→∞ s j =1
  m n  
t    ∗ 2  
= lim E B u, Jn X n (u))Q1/2 ϕj , fk K v, ϕj 2
H gs Jn X n
du .
n→∞ s j =1 k=1

Here, we have used the fact that the martingale

t   t  
X n (t) − ξ0n − a n s, X n (s) ds = bn s, X n (s) dBsn
0 0

t
has an increasing process given by 0 tr(b(s, X n (s))(b(s, X n (s))T ) ds.
By using the positive and negative parts of gs (x) separately, we can assume,
without any loss of generality, that gs (x) ≥ 0 in the following argument. Consider
the last expectation above. It is dominated by
  m ∞  
t     ∗ 2  
E B u, Jn X n (u) Q1/2 ϕj , fk K v, ϕj 2
H gs Jn X n
du
s j =1 k=1
  m  
t       
=E  B u, Jn X n (u) Q1/2 ∗ ϕj 2 v, ϕj 2 n
K H gs J n X du
s j =1
  m
t        ∗ 
=E B u, Jn X n (u) Q1/2 B u, Jn X n (u) Q1/2 ϕj , ϕj H
s j =1
 
 
v, ϕj 2H gs Jn X n
du

 t  
  1/2    1/2 ∗ m m   
=E n
B u, Jn X (u) Q n
B u, Jn X (u) Q n
v , v H gs Jn X du
s
 t  
  1/2    1/2 ∗ m m  
→E B u, X(u) Q B u, X(u) Q v , v H gs (X) du
s
t       ∗ 
= B u, x(u) Q1/2 B u, x(u) Q1/2 v m , v m H gs (x) du μ∗ (dx),
s
170 4 Solutions by Variational Method

using the weak convergence and uniform integrability of the integrand ensured
by (4.5) and (4.31). Hence,
  m n  
t     1/2 ∗ 2  
lim E n
B u, Jn X (u) Q ϕj , fk K v, ϕj H gs Jn X
2 n
du
n→∞ 0 j =1 k=1
t       ∗  
≤ B u, x(u) Q1/2 B u, x(u) Q1/2 v m , v m H gs (x) du μ∗ (dx).
s

To show the opposite inequality, note that if n ≥ r, then


m 
n
   ∗ 2  
lim inf B u, Jn X n (u) Q1/2 ϕj , fk K v, ϕj 2
H gs Jn X n
n→∞
j =1 k=1


m 
r
   ∗ 2  
≥ lim inf B u, Jn X n (u) Q1/2 ϕj , fk K v, ϕj 2
H gs Jn X n
n→∞
j =1 k=1


m 
r
   ∗ 2
= B u, Jn Xn (u) Q1/2 ϕj , fk K v, ϕj 2
H gs (X)
j =1 k=1
      ∗ 
→ B u, X(u) Q1/2 B u, X(u) Q1/2 v m , v m H gs (X),

and an application of the Fatou lemma gives the equality

 2  2  
v m , Mt (x) − v m , Ms (x) gs (x) μ∗ (dx)
t       ∗  
= B u, x(u) Q1/2 B u, x(u) Q1/2 v m , v m H gs (x) du μ∗ (dx).
s
(4.47)
Summarizing, calculations in (4.45), (4.46), and (4.47) prove that for v ∈ V , the
process v, Mt (x) is a square-integrable continuous martingale with the increasing
process given by
t      ∗ 
B u, x(u) Q1/2 B u, x(u) Q1/2 v, v du.
0

Let {ψj∗ }∞
j =1 be the dual orthonormal basis in V ∗ defined by the duality
 
u, ψj∗ V∗
= ψj , u , u ∈ V ∗.

Since by (4.4)
     
Mt (x)2 ∗ ≤ C 1 + sup x(t)2 ,
V H
0≤t≤T

the martingale Mt (x) ∈ MT2 (V ∗ ), i.e., it is a continuous μ∗ -square-integrable

V ∗ -valued martingale.
4.2 Existence of Weak Solutions Under Compact Embedding 171

Denote Mt (x) = Mt (x), ψj∗ V ∗ . Using (2.5) and the property of the dual basis,
j

its increasing process is given by

∞

    j    
M(x) t (u), v V ∗ = M (x), M k (x) t ψj∗ , u V ∗ ψk∗ , v V ∗
j,k=1
∞
  
= M j (x), M k (x) t ψj , u ψk , v u, v ∈ V ∗ .
j,k=1

Since
j   
Mt (x)Mtk (x) = ψj , Mt (x) ψk , Mt (x) ,
we can write
 j     
M (x), M k (x) t = ψj , M(x) , ψk , M(x) t
t      ∗ 
= B s, x(s) Q1/2 B s, x(s) Q1/2 ψj , ψk ds.
0

Define, for any 0 ≤ t ≤ T , a map Φ(s) : K → V ∗ by

∞
    ∗ 
Φ(s)(k) = B s, X(s) Q1/2 ψj , fm K ψj∗ , k ∈ K.
j =1

Then
∞
    ∗
Φ ∗ (s)(u) = ψj , u B s, X(s) Q1/2 ψj , u ∈ V ∗,
j =1

and we have, for u, v ∈ V ∗ ,

t 
Φ(s)Φ ∗ (s)u, v V∗
ds
0
∞

t    ∗    ∗ 
= B s, X(s) Q1/2 ψj , B s, X(s) Q1/2 ψk K ψj , u ψk , v ds
0 j,k=1

∞

t       ∗ 
= B s, X(s) Q1/2 B s, X(s) Q1/2 ψj , ψk ψj , u ψk , v ds
0 j,k=1
  
= M(X) t (u), v V ∗ ,

giving that
  t
M(X) t = Φ(s)Φ ∗ (s) ds.
0
172 4 Solutions by Variational Method

Note that Φ(s) ∈ L2 (K, V ∗ ), since

∞
 ∞

   ∗   2
Φ(s)fm 2 ∗ = ψj , B s, X(s) Q1/2 fm V ∗
V
m=1 m,j =1
∞
    2
= ψj , B s, X(s) Q1/2 fm
m,j =1
∞
    2
≤ ϕj , B s, X(s) Q1/2 fm H
m,j =1
∞
    
= B s, X(s) Q1/2 fm 2
H
m=1
  2
= B s, X(s) L < ∞,
2 (KQ ,H )

where we have used the assumption on the duality on Gelfand triplet and the fact that
ψj = ϕj /ϕj V with the denominator greater than or equal to one. Consequently,
the growth condition (4.5), together with (4.38), implies that

T 
E Φ(t)2 .
L2 (K,V ∗ )
0

Using the cylindrical version of the martingale representation theorem, Corol-

lary 2.2, we can write
t
Mt (X) = )s .
Φ(s) d W
0

Define
∞
  
Wt = )t Q1/2 fm fm .
W
m=1

By Exercise 2.4, Wt is a K-valued Q-Wiener process. Using Lemma 2.8, we calcu-

late
t  
X(t) − ξ0 − A s, X(s) ds = Mt (X)
0
t
= )s
Φ(s) d W
0
∞
 t 
= )s (fm )
Φ(s)fm d W
m=1 0
4.2 Existence of Weak Solutions Under Compact Embedding 173

∞
 ∞
t    
= )s (fm )
ψj , B s, X(s) Q1/2 fm ψj∗ d W
m=1 0 j =1
∞
 ∞
t      
= )s Q1/2 fm
ψj∗ , B s, X(s) fm V ∗ ψj∗ d W
m=1 0 j =1
∞
t     
= )s Q1/2 fm
B s, X(s) fm d W
0 m=1

t  
= B s, X(s) dWs .
0

We are now in a position to apply Theorem 4.3 to X(t), Y (t) = A(t, X(t)), and
Z(t) = B(t, X(t)) to obtain that X ∈ C([0, T ], H ), completing the proof. 

Exercise 4.7 Show that under the assumption of compact embedding in the Gelfand
triplet, there exists a vector system {ϕj }∞ j =1 ⊂ V which is a complete orthogonal
system in V and an ONB in H .
Hint: show that the canonical isomorphism I : V ∗ → V takes a unit ball in V ∗ to
a subset of the unit ball in V , which is relatively compact in H . For the eigenvectors
hn of I , we have

hn , hm H = hn , hm = hn , I hm V = λm hn , hm V.

We now address the problem of the existence and uniqueness of a strong solution
using a version of the Yamada and Watanabe result in infinite dimensions. Recall
the notion of pathwise uniqueness.

Definition 4.2 If for any two H -valued weak solutions (X1 , W ) and (X2 , W )
of (4.2) defined on the same filtered probability space (, F , {Ft }0≤t≤T , P ) with
the same Q-Wiener process W and such that X1 (0) = X2 (0) P -a.s., we have that
 
P X1 (t) = X2 (t), 0 ≤ t ≤ T = 1,

then we say that (4.2) has the pathwise uniqueness property.

We introduce here the weak monotonicity condition

(WM) There exists θ ∈ R such that for all u, v ∈ V , t ∈ [0, T ],
   2
2 u − v, A(t, u) − A(t, v) + B(t, u) − B(t, v)L ≤ θ u − v2H .
2 (KQ ,H )
(4.48)
The weak monotonicity is crucial in proving the uniqueness of weak and strong
solutions. In addition, it allows one to construct strong solutions in the absence of
the compact embedding V → H .
174 4 Solutions by Variational Method

Theorem 4.5 Let the conditions of Theorem 4.4 hold and assume the weak mono-
tonicity condition (4.48). Then the solution to (4.2) is pathwise unique.

Proof Let X1 , X2 be two weak solutions as in Definition 4.2, Y (t) = X1 (t) − X2 (t),
and denote a V -valued progressively measurable version of the latter by Ȳ . Apply-
ing the Itô formula and the monotonicity condition (4.48) yields

 2 t  2
e−θt Y (t)H = −θ e−θs Y (s)H ds
0
t      
+ e−θs 2 Ȳ (s), A s, X1 (s) − A s, X2 (s)
0
    2 
+ B s, X1 (s) − B s, X2 (s)  L2 (KQ ,H )
ds
t       
+2 e−θs Ys , B s, X1 (s) − B s, X2 (s) dWs H
0
≤ Mt ,

where Mt is a real-valued continuous local martingale represented by the stochastic

integral above. The inequality above also shows that Mt ≥ 0. Hence, by the Doob
maximal inequality, Mt = 0. 

As a consequence of an infinite-dimensional version of the result of Yamada and

Watanabe [67], we have the following corollary.

Corollary 4.2 Under the conditions of Theorem 4.5, (4.2) has a unique strong so-
lution.

4.3 Strong Variational Solutions

We will now study the existence and uniqueness problem for strong solutions.
A monotonicity condition will be imposed on the coefficients of the SDE (4.2),
and the compactness of embeddings V → H → V ∗ will be dropped. We empha-
size that the monotonicity condition will allow us to construct approximate strong
solutions using projections of a single Q-Wiener process, as opposed to construct-
ing finite-dimensional weak solutions in possibly different probability spaces. In the
presence of monotonicity, we can weaken other assumptions on the coefficients of
the variational SDE. A reader interested in exploring this topic in more depth is
referred to a detailed presentation in [64], where the authors reduce the conditions
even slightly further.
We assume that in the Gelfand triplet V → H → V ∗ , V is a real separable
Hilbert space.
4.3 Strong Variational Solutions 175

The joint continuity assumption is replaced by the spatial continuity:

(SC) For any t ∈ [0, T ], the mappings

V v → A(t, v) ∈ V ∗ and
  ∗ (4.49)
V v → B(t, v)Q1/2 B(t, v)Q1/2 ∈ L1 (H )

are continuous.
We now assume that the coefficient A satisfies the following growth condition:
(G-A’)
   
A(t, v)2 ∗ ≤ θ 1 + v2 , v ∈ V. (4.50)
V V

The coercivity condition (4.6) remains in force, and, in addition, we assume the
weak monotonicity condition (4.48).

Exercise 4.8 The coefficient B satisfies the following growth condition:

   
B(t, v)2 ≤ λv2H + θ  1 + v2V , v ∈ V , λ ∈ R, θ  ≥ 0.
L 2 (KQ ,H )

We will rely on the following finite-dimensional result for an SDE (4.15) with
the initial condition ξ0 . Its more refined version is stated as Theorem 3.1.1 in [64].

Theorem 4.6 Assume that a : [0, ∞] × Rn → Rn , b : [0, ∞] × Rn → Rn ⊗ Rn , and

a(t, x), b(t, x) are continuous in x ∈ Rn for each fixed value of t ≥ 0. Let
T    2 
sup a(t, x) + b(t, x) dt < ∞,
Rn
0 xRn ≤R

where b(t, x)2 = tr(b(t, x)bT (t, x)). Assume that for all t ≥ 0 and R > 0, on the
set {xRn ≤ R, yRn ≤ R}, we have
   2  
2 x, a(t, x) Rn + b(t, x) ≤ θ 1 + x2Rn

and
   2
2 x − y, a(t, x) − a(t, y) Rn + b(t, x) − b(t, y) ≤ θ x − y2Rn .

Let Eξ0 Rn < ∞.

Then there exists a unique strong solution X(t) to (4.15) such that for some
constant C,
 2   
E X(t)Rn ≤ C 1 + Eξ0 2Rn .

Here is the variational existence and uniqueness theorem for strong solutions.
176 4 Solutions by Variational Method

Theorem 4.7 Let V → H → V ∗ be a Gelfand triplet of real separable Hilbert

spaces, and let the coefficients A and B of (4.2) satisfy conditions (4.49), (4.50),
(4.6), and (4.48). Let the initial condition ξ0 be an H -valued F0 -measurable ran-
dom variable satisfying Eξ0 2H < c0 for some constant c0 . Then (4.2) has a unique
strong solution X(t) in C([0, T ], H ) such that
  2 
E sup X(t)H < ∞ (4.51)
0≤t≤T

and
T 
E X(t)2 dt < ∞. (4.52)
V
0

Proof Let {ϕi }∞

i=1 be an orthonormal basis in H obtained by the Gramm–Schmidt
orthonormalization process from a dense linearly independent subset of V . Define
Pn : V ∗ → Hn ⊂ V by


n
Pn u = ϕi , u ϕi , u ∈ V ∗.
i=1

By the assumption of the Gelfand triplet, Pn is the orthogonal projection of H

onto Hn . Also, let

n
1/2
Wtn = λi wi (t)fi
i=1
as in (2.3). Consider the following SDE on Hn :
   
dX n (t) = Pn A t, X n (t) dt + Pn B t, X n (t) dWtn (4.53)

with the initial condition X n (0) = Pn ξ0 and identify it with the SDE (4.15) in
Rn . It is a simple exercise to show that the conditions of Theorem 4.6 hold (Ex-
ercise 4.9); hence, we have a unique strong finite-dimensional solution X n (t) ∈ Hn .
We will now show its boundedness in the proper L2 spaces. Identifying X n (t) with
an Rn -valued process and applying the finite-dimensional Itô formula yield
 n 2   t  
X (t) = X n (0)2 + 2 X n (s), A s, X n (s)
H H
0
   2 
+ Pn B s, X n (s) P̃n  L2 (KQ ,H )
ds + M n (t),

where
t    
M n (t) = 2 X n (s), Pn B s, X n (s) dWsn H
0

is a local martingale, and P̃n denotes the orthogonal projection on span{f1 , . . . , fk }

⊂ K.
4.3 Strong Variational Solutions 177

Let σl , l = 1, 2, . . . , be stopping times localizing M n , and ηl = inf{t : X n (t)H

≥ l}. Then τl = σl ∧ ηl localizes M n , τl → ∞, and we can apply expectations as
follows:
 2   2  t    
E X n (t ∧ τl )H = E X n (0)H + E 1[0,τl ] (s)2 X n (s), Pn A s, X n (s)
0
   2 
+ Pn B s, X n (s) P̃n  L2 (KQ ,H )
ds.

Integration by parts and coercivity (4.6) yield

  2   2 
E e−λt X n (t ∧ τl ) − E X n (0) H H
t   2 
= E −λe−λs X n (s ∧ τl )H ds
0
t    
+ e−λs E 1[0,τl ] (s)2 X n (s), Pn A s, X n (s)
0
   2 
+ Pn B s, X n (s) P̃n  L2 (KQ ,H )
ds
t   2 
≤ E −λe−λs X n (s ∧ τl )H ds
0
t   2 
+ E λe−λs X n (s)H ds
0
t   2 
− E 1[0,τl ] (s)αe−λs X n (s)V ds + γ T . (4.54)
0

Rearranging and applying the Fatou lemma as l → ∞, we obtain

  2   2  t   2 
E e−λt X n (t)H − E X n (0)H + E αe−λs X n (s)V ds ≤ C.
0

Hence, Xn is bounded in L2 ([0, T ]

× , V ) and in L2 ([0, T ] × , H ). Us-
ing (4.6), we also have the boundedness of A(·, X n (·)) in L2 ([0, T ] × , V ∗ ) =
(L2 ([0, T ] × , V ))∗ . In addition, by Exercise 4.8, Pn B(·, X n (·)) is bounded in
L2 ([0, T ] × , L2 (KQ , H )). Therefore, by using the Alaoglu theorem and passing
to a subsequence if necessary, we can assume that there exist X, Y , and Z such that
   
X n → X weakly in L2 [0, T ] × , V , and weakly in L2 [0, T ] × , H ,
   
Pn A ·, X n (·) → Y weakly in L2 [0, T ] × , V ∗ , (4.55)
   
Pn B ·, X (·) → Z weakly in L [0, T ] × , L2 (KQ , H ) .
n 2

Also Pn B(·, X n (·))P̃n → Z weakly in L2 ([0, T ] × , L2 (KQ , H )) and

t   t  
Pn B s, X n (s) P̃n dWs = Pn B s, X n (s) dWsn ,
0 0
178 4 Solutions by Variational Method

we can claim that

t   t
Pn B s, X n (s) dWsn → Z(s) dWs
0 0

weakly in L2 ([0, T ] × , H ), the reason being that the stochastic integral is a con-
tinuous transformation from 2 (KQ , H ) to L2 ([0, T ] × , H ), and so it is also
continuous with% respect to weak topologies in those spaces (see Exercise 4.10).
For any v ∈ n≥1 Hn and g ∈ L2 ([0, T ] × , R), using the assumption on the
duality, we obtain
 T 
 
E g(t)v, X(t) dt
0
 T 

= lim E g(t)v, X (t) dt n
n→∞ 0
 T   t  
= lim E g(t)v, X n (0) + g(t)v, Pn A s, X n (s) ds
n→∞ 0 0
  
t  
+ Pn B s, X n (s) dWsn , g(t)v dt
0 H
 T 
  T T  
= lim E v, X (0) H n
g(t) dt + n
g(t)v dt, Pn A s, X (s) ds
n→∞ 0 0 s
T t  
 
+ Pn B s, X n (s) dWsn , g(t)v dt
0 0 H
 T t t  
=E g(t)v, X(0) + Y (s) ds + Z(s) dWs dt .
0 0 0

Therefore,
t t
X(t) = X(0) + Y (s) ds + Z(s) dWs , dt ⊗ dP -a.e.,
0 0

and applying Theorem 4.3, we conclude that X is a continuous H -valued process

with
  2 
E sup X(t)H < ∞.
0≤t≤T

We now verify that Y (t) = A(t, X(t)) and Z(t) = B(t, X(t)), dt ⊗ dP -a.e. For a
nonnegative function ψ ∈ L∞ ([0, T ], R), we have
T 
E ψ(t)X(t), X n (t) H dt
0
T 5   5   
≤E ψ(t)X(t)H ψ(t)X n (t)H dt
0
4.3 Strong Variational Solutions 179
 1/2  1/2
T  2 T  2
≤ E ψ(t)X(t)H dt E ψ(t)X n (t)H dt .
0 0

Hence, by the weak convergence of X n to X,

T  2 T 
E ψ(t)X(t)H dt = lim E ψ(t)X(t), X n (t) H dt
0 n→∞ 0
 1/2  1/2
T  2 T  2
≤ E ψ(t)X(t)H dt lim inf E ψ(t)X n (t)H dt < ∞,
0 n→∞ 0

giving
T  2 T  2
E ψ(t)X(t)H dt ≤ lim inf E ψ(t)X n (t)H dt. (4.56)
0 n→∞ 0

Let φ ∈ L2 ([0, T ] × , V ). Revisiting the calculations in (4.54), with the constant c

in the weak monotonicity condition (4.48), replacing λ, and taking the limit as
l → ∞ yield
  2   2 
E e−ct X n (t)H − E X n (0)H
t   2    2 
= e−cs E −cX n (s)H + Pn B s, X n (s) P̃n L ds
2 (KQ ,H )
0
t    
+ e−cs E 2 X n (s), Pn A s, X n (s) ds
0
t   2   2 
≤ e−cs E −cX n (s)H + B s, X n (s) L ds
2 (KQ ,H )
0
t    
+ e−cs E 2 X n (s), Pn A s, X n (s) ds
0
t   2     2
= e−cs E −cX n (s) − φ(s)H + B s, X n (s) − B s, φ(s) L
2 (KQ ,H )
0
    
+ 2 X n (s) − φ(s), Pn A s, X n (s) − A s, φ(s) ds
t   2  
+ e−cs E cφ(s)H − 2c X n (s), φ(s) H
0
       2
+ 2 B s, X n (s) , B s, φ(s) L (K ,H ) − B s, φ(s) L (K ,H )
2 Q 2 Q
 n       
+ 2 X (s), A s, φ(s) + 2 φ(s), Pn A s, X n (s) − A s, φ(s) ds.

Since by (4.48) the first of the last two integrals is negative, by letting n → ∞, using
the weak convergence (4.55) in L2 , and applying (4.56), we conclude that for any
180 4 Solutions by Variational Method

function ψ as above,

T   2  2 
ψ(t)E e−ct X(t)H − X(0)H dt
0
T t   2  
≤ ψ(t) e−cs E cφ(s)H − 2c X(s), φ(s) H
0 0
     2
+ 2 Z(s), B s, φ(s) L − B s, φ(s) L
2 (KQ ,H ) 2 (KQ ,H )

     
+ 2 X̄(s), A s, φ(s) + 2 φ(s), Y (s) − A s, φ(s) ds dt. (4.57)

Recall the Itô formula (4.37). With stopping times τl localizing the local martingale
represented by the stochastic integral, we have
 2   2 
E X(t ∧ τl )H − E X(0)H
t      2 
= E 1[0,τl ] (s) 2 X̄(s), Y (s) + Z(s)L ds.
2 (KQ ,H )
0

Since, by (4.36) and by the square integrability of Y and Z, we can pass to the limit
using the Lebesgue DCT, the above equality yields
 2   2 
E X(t)H − E X(0)H
t     2 
= E 2 X̄(s), Y (s) + Z(s)L ds. (4.58)
2 (KQ ,H )
0

Applying integration by parts, we get

  2   2 
E e−ct X(t)H − E X(0)H
t     2  2 
= e−cs E 2 X̄(s), Y (s) + Z(s)L − ce−cs X(s)H ds. (4.59)
2 (KQ ,H )
0

We now substitute the expression for the left-hand side of (4.59) into the left-hand
side of (4.57) and arrive at

T t    
E ψ(t) e−cs 2 X̄(s) − φ(s), Y (s) − A s, φ(s) (4.60)
0 0

   2   
+ B s, φ(s) − Z(s)L − c X(s) − φ(s)2 ds dt ≤ 0. (4.61)
2 (KQ ,H ) H

Substituting φ = X̄ gives that Z = B(·, X̄). Now let φ = X̄ − ε φ̃v with ε > 0, φ̃ ∈
L∞ ([0, T ] × , R), and v ∈ V . Let us divide (4.60) by ε and pass to the limit as
4.4 Markov and Strong Markov Properties 181

ε → 0 using the Lebesgue DCT. Utilizing (4.49) and (4.48), we obtain that
T  t   

E ψ(t) e−cs φ̃(s) v, Y (s) − A s, X̄(s) ds dt ≤ 0.
0 0

This proves that Y = A(·, X̄) due to the choice of ψ and φ̃.
The argument used in the proof of Theorem 4.5 can now be applied to show the
uniqueness of the solution. 

Exercise 4.9 Show that the coefficients of (4.53) identified with the coefficients
of (4.15) satisfy the conditions of Theorem 4.6.

Exercise 4.10 Show that if T : X → Y is a continuous linear operator between

Banach spaces X and Y , then T is also continuous with respect to weak topologies
on X and Y .

Exercise 4.11 Let X and Y be two solutions to (4.2). Using (4.58), show that under
the conditions of Theorem 4.5 or Theorem 4.7,
 2  2
E X(t) − Y (t)H ≤ ect E X(0) − Y (0)H , 0 ≤ t ≤ T .

Note that this implies the uniqueness of the solution, providing an alternative argu-
ment to the one used in text.

4.4 Markov and Strong Markov Properties

Similarly as in Sect. 3.4, we now consider strong solutions to (4.2) on the interval
[s, T ]. The process W̄t = Wt+s − Ws is a Q-Wiener process with respect to F¯t =
Ft+s , t ≥ 0, and its increments on [0, T − s] are identical with the increments of
Wt on [s, T ]. Consider (4.2) with W̄t replacing Wt and F¯0 = Fs replacing F0 .
By Theorem 4.7, there exists a unique strong solution X(t) of (4.2), so that for any
0 ≤ s ≤ T and an Fs -measurable random variable ξ , there exists a unique process
X(·, s, ξ ) such that
t   t  
X(t, s, ξ ) = ξ + A r, X(r, s, ξ ) dr + B r, X(r, s, ξ ) dWr . (4.62)
s s

As before, for a real bounded measurable function ϕ on H and x ∈ H ,

  
(Ps,t ϕ)(x) = E ϕ X(t, s; x) , (4.63)

and this definition can be extended to functions ϕ such that ϕ(X(t, s; x)) ∈
L1 (, R) for arbitrary s ≤ t. As usual, for a random variable η,
  
(Ps,t ϕ)(η) = E ϕ X(t, s; x) x=η .
182 4 Solutions by Variational Method

The Markov property (3.52) (and consequently (3.57)) of the solution now follows
almost word by word by the arguments used in the proof of Theorem 3.6 and by
Exercise 4.11. A proof given in [64] employs similar ideas.

Theorem 4.8 The unique strong solution to (4.2) obtained in Theorem 4.7 is a
Markov process.

Remark 4.2 In the case where the coefficients A and B are independent of t and
with x ∈ H ,
t+s   t+s  
X(t + s, t; x) = x + A X(u, t; x) du + B X(u, t, x) dWu
t t
s   s  
=x+ A X(t + u, t; x) du + B X(t + u, t, x) d W̄u ,
0 0

where W̄u = Wt+u − Wt . Repeating the arguments in Sect. 3.4, we argue that
d 
X(t + s, t; x), s ≥ 0 = X(s, 0; x), s ≥ 0 ,

i.e., the solution is a homogeneous Markov process with

Ps,t (ϕ) = P0,t−s (ϕ), 0 ≤ s ≤ t,

for all bounded measurable functions ϕ on H .

As before, we denote
Pt = P0,t .
Due to the continuity of the solution with respect to the initial condition, Pt is a
Feller semigroup, and X(t) is a Feller process.

In Sect. 7.5 we will need the strong Markov property for strong variational solu-
tions. We prove this in the next theorem. Consider the following variational SDE:
   
dX(t) = A X(t) dt + B X(t) dWt (4.64)

with the coefficients

A:V →V∗ and B : V → L2 (KQ , H )

and an H -valued F0 -measurable initial condition ξ ∈ L2 (, H ).

Definition 4.3 Let τ be a stopping time with respect to a filtration {Ft }t≥0 (an
Ft -stopping time for short). We define

Fτ = σ A ∈ F : A ∩ {τ ≤ t} ∈ Ft , t ≥ 0 .
4.4 Markov and Strong Markov Properties 183

Exercise 4.12 Show that FτW = σ {Ws∧τ , s ≥ 0} and that FτX = σ {Xs∧τ , s ≥ 0}
for a strong solution to (4.2).

Exercise 4.13 Let X(t) be a progressively measurable process with respect

to filtration {Ft }t≥0 , and τ be an Ft -stopping time. Show that Xτ 1{τ <∞} is
Fτ -measurable.

Definition 4.4 A solution X(t) of (4.64) in C([0, T ], H ) is called a strong Markov

process if it satisfies the following strong Markov property:
     
E ϕ X(τ + s; ξ ) FτW,ξ = (Ps ϕ) X(τ ; ξ ) P -a.s. on {τ < ∞}, (4.65)

for any real-valued function ϕ such that ϕ(X(t; ξ )) ∈ L1 (, R) and an

W,ξ
Ft -stopping time τ .

Theorem 4.9 Under the assumptions of Theorem 4.7, the unique strong solution
X(t) of (4.64) in C([0, T ], H ) is a strong Markov process.

Proof By the monotone class theorem (functional form) we only need to show that
W,ξ
for any bounded continuous function ϕ : H → R and A ∈ Fτ ,
       
E ϕ X(τ + s; ξ ) 1A∩{τ <∞} = E (Ps ϕ) X(τ ; ξ ) 1A∩{τ <∞} . (4.66)

If τ takes finitely many values, then A ∈ Fmax{τ (ω)} , and (4.66) is a consequence of
Theorem 4.8.
W,ξ
Let τn be a sequence of Ft -stopping times, each taking finitely many values,
W,ξ W,ξ
and τn ↓ τ on τ < ∞ (see Exercise 4.14). Since τn ≥ τ , we have Fτn ⊃ Fτ
W,ξ
and A ⊂ Fτn for all n. Consequently, (4.66) holds for τn ,
       
E ϕ X(τn + s; ξ ) 1A∩{τ <∞} = E (Ps ϕ) X(τn ; ξ ) 1A∩{τ <∞} .

By the continuity of ϕ and X(t),

       
E ϕ X(τn + s; ξ ) 1A∩{τ <∞} → E ϕ X(τ + s; ξ ) 1A∩{τ <∞} .

By the Feller property of the semigroup Ps ,

       
E (Ps ϕ) X(τn ; ξ ) 1A∩{τ <∞} → E (Ps ϕ) X(τ ; ξ ) 1A∩{τ <∞} .

This completes the proof. 

Exercise 4.14 For an Ft -stopping time τ , construct a sequence of Ft -stopping

times such that τn ↓ τ on {τ < ∞}.

Corollary 4.3 Under the assumptions of Theorem 4.9, the unique strong solution
X(t) of (4.64) in C([0, T ], H ) has the following strong Markov property:
     
E ϕ X(τ + s; ξ ) FτX = (Ps ϕ) X(τ ; ξ ) P -a.s. on {τ < ∞} (4.67)
184 4 Solutions by Variational Method

for any real-valued function ϕ such that ϕ(X(t; ξ )) ∈ L1 (, R) with FsX =
σ {X(r; ξ ), r ≤ s} and an FtX -stopping time τ .

Exercise 4.15 Show that if X(t) = X(t; ξ ) is as a solution to (4.64) as in Theo-

rem 4.9, then the strong Markov property (4.67) implies that
       
E ϕ X(τ + s) FτX = E ϕ X(τ + s) Xτ

for any a real-valued function ϕ such that ϕ(X(t; ξ )) ∈ L1 (, R) and any
FtX -stopping time τ .

We now refer the reader to Sect. 4.1 in [64], where several examples and further
references are provided.
Chapter 5
Stochastic Differential Equations
with Discontinuous Drift

5.1 Introduction
In this chapter, we consider genuine infinite-dimensional stochastic differential
equations not connected to SPDEs.
This problem has been discussed in Albeverio’s work on solutions to infinite-
dimensional stochastic differential equations with values in C([0, T ], RZ ), which
d

was motivated by applications to quantum lattice models in statistical mechanics

[2, 3]. Leha and Ritter [47] have studied the existence problem in C([0, T ], H w ),
where H is a real separable Hilbert space endowed with its weak topology and
applied their results to modeling unbounded spin systems. Our purpose here is to
extend the results obtained in [2, 3], and [47] using techniques developed in Chaps. 3
and 4.
We begin with the problem of the existence of weak solutions for SDEs with dis-
continuous drift in a Hilbert space H discussed in [47]. The discontinuity is modeled
by a countable family of real-valued functions. The solution has finite-dimensional
Galerkin approximation and is realized in C([0, T ], H w ). The result in [47] is gen-
eralized, and we also show that, under the assumptions in [47], both the Galerkin
approximation and the infinite-dimensional approximation of [47] produce solutions
with identical laws.
Next, we study the solutions in C([0, T ], RZ ) using ideas in [3] and the tech-
d

nique of compact embedding in [22].

5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w )

In Chap. 3, it was shown that with the usual continuity and growth assumptions on
the coefficients of a stochastic differential equation in an infinite-dimensional real
separable Hilbert space H , the solution can be obtained in a larger Hilbert space
H−1 such that the embedding H → H−1 is compact. The space H−1 was arbitrary,
however needed due to a (deterministic) example of Godunov [26]. This forced us
to extend continuously the coefficients of the SDE to H−1 .

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 185

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_5, © Springer-Verlag Berlin Heidelberg 2011
186 5 Stochastic Differential Equations with Discontinuous Drift

We exploit this idea to obtain solutions in C([0, T ], H w ) (H w denotes H with

its weak topology) using the technique of compact embedding in Chap. 3. As a con-
sequence of this approach, we are able to give an extension of an interesting result
on equations with discontinuous drift due to Leha and Ritter [47], by eliminating
local Lipschitz assumptions. When the drift is Lipschitz continuous, we relate the
solution constructed here to that obtained by Leha and Ritter. The advantage of the
construction presented here is that the weak solution has finite-dimensional Galerkin
approximation as opposed to the infinite-dimensional approximation given in [47].
Let us now consider H as a space isomorphic to
 ∞

  
l2 = x ∈ R ∞ :
2
xi < ∞
i=1

with the canonical isomorphism

   
h → x 1 , x 2 , . . . = h, e1 H , h, e2 H , . . . ,

where {ek }∞
k=1 is an ONB in H . Then the natural choice of the larger space is
(R∞ , ρR∞ ) with its metric defined for coordinate-wise convergence
∞
 1 |x k − y k |
ρR∞ (x, y) = ,
2k 1 + |x k − y k |
k=1

as the embedding J : l2 → R∞ is continuous and compact, see Exercise 5.1.

Exercise 5.1 Prove that the embedding J : l2 → R∞ is continuous and compact.

Show that for some constant C and x, y ∈ l2 ,

ρR∞ (J x, Jy) ≤ Cx − yH . (5.1)

Let (Ω, F , {Ft }t≤T , P ) be a filtered probability space, and Wt be an H -valued

Wiener process with covariance Q, a nonnegative trace-class operator on H , and
with eigenvalues λi > 0 and the associated eigenvectors ei , i =
1, 2, . . . . Denote by H τ the space H endowed with the topology induced by R∞
(under the identification with l2 ).

Theorem 5.1 Let F : H → H , and assume that F k (x) := F (x), ek H : H τ → R

k : R → R, k = 1, 2, . . . , are continuous. Assume that γ > 0 are constants
and q k
with ∞ k=1 γk < ∞. Let the following conditions hold:
(GF)
   
F (x)2 ≤ 1 + x2 , x ∈ H; (5.2)
H H

(Gq1)
 
uq k (u) ≤ γk + u2 , u ∈ R; (5.3)
5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w ) 187

(Gq2) Let q̄n (u1 , . . . , un ) = (q 1 (u1 ), . . . , q n (un )) ∈ Rn . There exists a positive in-
teger m, independent on n, such that
   
q̄n (un )2 n ≤ C 1 + un 2mn . (5.4)
R R

Then, there exists a weak solution X in C([0, T ], H w ) to the equation

t  
X(t) = x + F X(s) + q X(s) ds + Wt
0

in the following sense. There exists an H -valued Q-Wiener process Wt and a process
X(·) ∈ C([0, T ], H w ) such that for every k = 1, 2, . . . ,
t    
X k (t) = x k + F k X(s) + q k X k (s) ds + Wtk . (5.5)
0

Here, y k = y, ek H is the kth coordinate of y ∈ H .

Proof Consider the following sequence of equations:

t    
Xn (t) = Pn x + Pn F Pn Xn (s) + qn Pn Xn (s) ds + Wtn . (5.6)
0

Here,
n Pnk iskthe projection of H onto span{e1 , . . . , en }, qn : Pn H → Pn H , qn (y) =
k=1 q (y )e k , y ∈ Pn H , and W t
n is an H -valued Wiener process with covariance

Qn = Pn Q Pn .
We can consider (5.6) in Rn by identifying Pn H with Rn and treating Wtn as an
R -valued Wiener process. Denote
n

     
n
Gn (x) = F 1 (xn ) + q 1 x 1 , . . . , F n (xn ) + q n x n ∈ Rn , x ∈ R , xn =
n
x k ek .
k=1

Note that conditions (5.2) and (5.3) imply that

   
x, Gn (x) Rn ≤ C 1 + x2Rn , (5.7)

so that, by Theorem 4.1, there exists a weak solution ξn (·) ∈ C([0, T ], Rn ).

We now establish estimates for the moments of ξn (t). Denote
  
τR = inf ξn (t)Rn > R
0≤t≤T

or τR = T if the infimum is taken over an empty set. Using the Itô formula for the
function x2Rn on Rn , we get

  t∧τR   
ξn (t ∧ τR )2 n = Pn x2 + 2 ξn (s), Gn ξn (s) Rn ds
R H
0
188 5 Stochastic Differential Equations with Discontinuous Drift

t∧τR  
+ (t ∧ τR ) tr(Qn ) + 2 ξn (s), dWsn Rn
.
0

For the stochastic integral term, we use that 2a ≤ 1 + a 2 ; hence, by the Doob in-
equality, Theorem 2.2, and (5.7), we have

 2 t 
E sup ξn (s ∧ τR )Rn ≤ C1 + C2 ξn (s ∧ τR )2 n ds.
R
0≤s≤t 0

Now the Gronwall lemma implies that

 2
E sup ξn (t ∧ τR )Rn < C3 .
0≤t≤T

Taking R → ∞ and using the monotone convergence theorem imply

 2
sup E sup ξn (t)Rn < ∞. (5.8)
n 0≤t≤T

Let l be a positive integer. Using the Itô formula for the function x2l
Rn on R , we
n

get

  t∧τR     
ξn (t ∧ τR )2ln = Pn x2l + 2l ξn (s)2(l−1) ξn (s), Gn ξn (s) ds
R H nR Rn
0
t∧τR    1/2 
+ 2l(l − 1) ξn (s)2(l−2) Qn ξn (s)2 n ds
nR R
0
t∧τR  
+l ξn (s)2(l−1) tr(Qn ) ds
n R
0
t∧τR    
+ 2l ξn (s)2(l−1) ξn (s), dWsn Rn .
Rn
0

Taking the expectation to both sides and using (5.7) yield

 2l t  2(l−1)
E ξn (t ∧ τR )Rn ≤ C1 + C2 E ξn (t ∧ τR )Rn ds
0
t  2l
+ C3 E ξn (t ∧ τR )Rn ds
0
t  2l
≤ (C1 + C2 T ) + (C2 + C3 ) E ξn (t ∧ τR )Rn ds,
0

where we have used the fact that a 2(l−1) ≤ 1 + a 2l . By Gronwall’s lemma, for some
constant C,
 2l
E ξn (t ∧ τR )Rn ≤ C,
5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w ) 189

which, as R → ∞, leads to
 2l
E sup ξn (t)Rn ≤ C. (5.9)
0≤t≤T

Using (5.8) and (5.9) and essentially repeating the argument in Lemma 3.10, we now
obtain an estimate for the fourth moment of the increment of the process ξn (t ∧ τR ).
Applying the Itô formula for the function x4Rn on Rn , we get
 
ξn (t + h) − ξn (t)4 n
R
t+h     
=4 ξn (u) − ξn (t)2 n ξn (u) − ξn (t), Gn ξn (u) du
R Rn
t
t+h  
+2 ξn (u) − ξn (t)2 n tr(Qn ) du
R
t
t+h   
+2 Qn ξn (u) − ξn (t), ξn (u) − ξn (t) Rn du
t
t+h    
+4 ξn (u) − ξn (t)2 n ξn (u) − ξn (t), dW n n .
R u R
t

Taking the expectation of both sides and using assumptions (5.2) and (5.4), which
imply the polynomial growth of Gn , we calculate
 4
E ξn (t + h) − ξn (t)Rn
t+h     
≤ C1 E ξn (u) − ξn (t)3 n 1 + ξn (u) du
R
t
t+h  
+ C2 E ξn (u) − ξn (t)2 n du
R
t
, t+h  -3/4

≤ C3 E ξn (u) − ξn (t)4 n du h1/4
R
t
, t+h  -1/2 

+ E ξn (u) − ξn (t)4 n du h1/2 ≤ Ch
R
0

for a suitable constant C. Substituting repeatedly, starting with C(u − t) for

Eξn (u) − ξn (t)4Rn , leads to the following estimate for the fourth moment of the
increment of ξn :
 4
E ξn (t + h) − ξn (t) n ≤ Ch2 R
for some constant C independent of n.
By (5.1), the compactness of the embedding J and Lemma 3.14, the measures
μn = P ◦ Xn−1 are tight on C([0, T ], R∞ ). Let μ be a limit point of the sequence
190 5 Stochastic Differential Equations with Discontinuous Drift

μn . Since the function C([0, T ], R∞ ) x → sup0≤t≤T ∞ k=1 (x (t)) ∈ R is lower
k 2

semicontinuous, we conclude that

 ∞
  ∞

  k 2   k 2
Eμ sup x (t) ≤ lim inf Eμn sup x (t)
0≤t≤T k=1 n 0≤t≤T k=1
  2 
= lim inf EP sup Xn (t)H < ∞.
n 0≤t≤T

Here, Xn (t) = nk=1 ξnk (t)ek .
Using the Skorokhod theorem, by changing the underlying probability space, we
can assume that
 
Xn → X, P -a.s. in C [0, T ], R∞
(not in C([0, T ], H w ), which is not metrizable) with Xn (t), X(t) ∈ H , t ∈ [0, T ],
P -a.s., since
 ∞

  k 2
P sup X (t) < ∞ = 1.
0≤t≤T k=1

The process X(t) is in H ; hence the measure μ is concentrated on C([0, T ], H w ),

as the topologies on H τ and on H w coincide on norm-bounded sets of H , and a.e.
path is such a set. t
Now, the random variables (Xnk (t))2 and 0 [Fnk (Xn (s)) + q k (Xnk (s))]2 ds are
P -uniformly integrable in view of (5.2), (5.4), and (5.9). In addition, for n ≥ k,
Fnk (Xn (s)) = F k (Xn (s)), and hence,
t    
EP F k Xn (s) − F k X(s) 2 ds → 0
n
0

due to the continuity of F k on H τ . Consequently, the sequence of Brownian motions

t*    +
Ynk (t) = Xnk (t) − x k − Fnk Xn (s) + q k Xnk (s) ds
0

converges in L2 (Ω, P ) to the Brownian motion

t*    +
Y k (t) = X k (t) − x k − F k X(s) + q k X k (s) ds.
0

Define an H -valued Brownian motion

∞

Wt = Y k (t)ek ,
k=1

then X(t), Wt satisfy (5.5). 

5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w ) 191

Let us relate these results to the work of Leha and Ritter [47]. We begin with the
general uniqueness and existence theorem in [46].

Theorem 5.2 Let H be real separable Hilbert space, and Wt be a Q-Wiener pro-
cess. Assume that A : H → L (H ) and B ∈ L (H ) satisfy the following growth and
local Lipschitz conditions. There exist constants C and Cn , n = 1, 2, . . . , such that
(1) tr(B(x)QB ∗ (x)) ≤ C(1 + x2H ), x ∈ H ;
(2) x, A(x) H ≤ C(1 + x2H ), x ∈ H ;
(3) A(x) − A(y)L (H ) + B(x) − B(y)H ≤ Cn x − yH for xH ≤ n and
yH ≤ n.
Then there exists a unique strong solution to the equation
t   t  
X(t) = x + A X(s) ds + B X(s) dWs , t > 0.
0 0

Exercise 5.2 Prove Theorem 5.2.

Hint: assume global Lipschitz condition first to obtain the usual growth condi-
tions on A and B and produce solution X(t) as in Theorem 3.3. Define

A(x), xH ≤ n,
An (x) =
A(nx/xH ), xH > n,

and define Bn in a similar way. Show that An and Bn are globally Lipschitz and
show that for the corresponding solutions Xn (t), there exists a process X(t) such
that X(t ∧ τn ) = Xn (t ∧ τn ) with τn denoting the first exit time of Xn from the
ball of radius n centered at the origin. Finally, show that P (τn < t) → 0 by using
the estimate for the second moment of Xn obtained from the application of the Itô
formula to the function x2H .

Following [47], for a finite subset V of the set of positive integers, define q V :
H → H by

 V  q k (y k ), k ∈ V ,
q (y), ek H =
0, k∈/ V.
From Theorem 5.2 we know that for any fixed V , under conditions (5.2), (5.3) and
the local Lipschitz condition on the coefficients F and q V , there exists a unique
strong solution to the equation
t   
ξ V (t) = x + F ξ V (s) + q V ξ V (s) ds + Wt
0

with the solution ξ V (·) ∈ C([0, T ], H ).

Leha and Ritter proved further in Theorem 2.4 in [47] that, under condi-
tions (5.2), (5.3) and the local Lipschitz condition (3) of Theorem 5.2 on F and
192 5 Stochastic Differential Equations with Discontinuous Drift

q V (the Lipschitz constant possibly dependent on V ), there exists a weak solution ξ

in C([0, T ], H w ) constructed as a weak limit of ξ Vn satisfying, for each k,
t  k   
ξ k (t) = x k + F ξ(s) + q k ξ k (s) ds + W k (t).
0

We now show that, under the global Lipschitz condition independent on V , the laws
of the above solution ξ(t) and the solution X(t) obtained in Theorem 5.1 coincide.

Theorem 5.3 Assume that the functions F + q V : H → H satisfy global Lipschitz

conditions for every V , a finite subset of positive integers, with the Lipschitz con-
stants independent on V . Assume that conditions (5.2) and (5.3) hold. Then the
weak solution ξ of Theorem 2.4 in [47] and the weak solution X constructed in
Theorem 5.1, using the compact embedding argument, have the same law.

Proof We first note that by Theorem 5.2 under the Lipschitz condition, both ap-
proximating sequences ξnV of Theorem 2.4 in [47] and Xn of Theorem 5.1 can be
constructed as strong solutions on the same probability space. Let Fn = Pn ◦ F ◦ Pn .
Then
t    
Xn (t) = xn + Fn Xn (s) + qn Xn (s) ds + Wtn ,
0
t    
ξ Vn (t) = x + F ξ Vn (s) + q Vn ξ Vn (s) ds + Wt .
0

The laws L (ξ V ) are tight (see the proof of Theorem 2.4 in [47]). Therefore, for
a sequence Vn = {1, 2, . . . , n}, there is a subsequence Vnk such that L (ξ Vnk ) →
L (ξ ). Therefore, for simplicity, we assume as in Theorem 2.4 in [47] that Vn =
{1, . . . , n} and that ξ Vn → ξ weakly. Denote Yn (t) = Pn ξ Vn (t). Then
t    
Yn (t) = xn + Pn F ξ Vn (s) + qn Yn (s) ds + W n (t).
0

We obtain
 2 t     
E Xn (t) − Yn (t)H ≤ CE Pn F Xn (s) − Pn F ξ Vn (s) 2
H
0
    2 
+ qn Xn (s) − qn Yn (s) H ds
t     
≤ C1 E Xn (s) − ξ Vn (s)2 + Xn (s) − Yn (s)2 ds
H H
0
t     
≤ C1 E Yn (s) − ξ Vn (s)2 + 2Xn (s) − Yn (s)2 ds,
H H
0

where we have used the Lipschitz condition.

5.2 Unbounded Spin Systems, Solutions in C([0, T ], H w ) 193

Note that (5.9) also holds for ξ Vn (the same proof works, with H replacing Rn ).
Hence, by changing the underlying probability space to ensure the a.s. convergence
by means of Skorokhod’s theorem, we have that
t 
2 t 
E  
Yn (s) − ξ (s) H ds ≤ 3 E
Vn Yn (s) − Pn ξ(s)2 ds
H
0 0
t 
 t 
+E Pn ξ(s) − ξ(s)2 ds + E ξ(s) − ξ Vn (s)2 ds → 0
H H
0 0

as n → ∞, due to the uniform integrability, implying that with some εn → 0,

 2 t  2
E Xn (t) − Yn (t)H ≤ 2C1 εn + C1 E Xn (s) − Yn (s)H ds.
0

Using Gronwall’s lemma, we obtain that

 2
E Xn (t) − ξ Vn (t)H ≤ εn eC1 t

with the expression on the right-hand side converging to zero as n → ∞.

We conclude that Xn (t) − ξ Vn (t) → 0 in L2 (Ω). Thus, ξ Vn → ξ implies that
Xn (t) → ξ(t) weakly. Therefore, Law(X) = Law(ξ ) on C([0, T ], R∞ ), and conse-
quently the laws coincide on C([0, T ], H w ). 

Example 5.1 (Unbounded Spin Systems) In statistical mechanics an unbounded

spin system can be described by a family of interaction potentials and on-site en-
ergy functions. For a family of subsets V ⊂ Z+ with |V | < ∞, consider interaction
potentials ϕV : RV → R. A typical example is pair potentials ϕ{k,l} : R2 → R,

ϕ{k,l} (u, v) = −Jk,l uv, k = l.

On-site energy is modeled by functions ϕk : R → R.

The energy function H defined on the space of configurations with finite support
in R∞ into R∞ is now defined at every site “k” as
  
H k (x) = φk x k − Jk,l x k x l .
l =k

The components of the drift coefficient of an SDE takes the following form:
∂  
F k (x) = k
Jk,l x k x l = Jk,l x l ,
∂x
l =k l =k

q k (u) = −φk (u),

so that
∂
F k (x) + q k (x) = − H k (x).
∂xk
194 5 Stochastic Differential Equations with Discontinuous Drift

The growth condition (5.2) is satisfied if, for example,


1 if |k − l| = 1,
Jk,l =
0 otherwise,

that is, there is only the closest neighbor interaction.

Conditions (5.3) and (5.4) hold if, for example,

ϕk (u) = P (u) = an u2n + an−1 u2(n−1) + · · · + a1 u2 + a0 (5.10)

with an > 0.
The results show that there exists a weak solution only under the assumption that
q k (u) : R → R satisfy growth conditions and are continuous functions. It should be
noted that even if Jk,l = 0 for all k, l, the growth condition (5.3) is necessary for the
existence of a solution without explosion, and continuity is needed in the proof of
the Peano theorem.
In Euclidian quantum field theory continuous spin models serve as lattice ap-
proximations (see [60] for details) with R∞ replaced by RZ . In [60],
d

 
1 if dj =1 |kj − lj | = 1,
Jk,l =
0 otherwise,
 
ϕk (u) = d + m2 /2 u2 + P (u),

where k, j ∈ Zd .
We will study such models in the next section.

5.3 Locally Interacting Particle Systems, Solutions

in C([0, T ], RZ )
d

We use recent ideas from Albeverio et al. [3] to study the dynamics of an infinite
particle system corresponding to a Gibbs measure on the lattice. Our technique is to
study weak solutions of an infinite system of SDEs using the work in [22, 36], and
the methods in Chap. 4 related to the case of a SDE in the dual to a nuclear space.
This allows us to extend the existence result in [3] by removing the dissipativity
condition.
The work [3] provides results for the existence and uniqueness of solutions to
a system of SDEs describing a lattice spin-system model with spins taking values
in “loop spaces.” The space of configurations Ωβ = C(Sβ )Z , where Sβ is a circle
d

with circumference β > 0.

We consider a lattice system of locally interacting diffusions, which is a special
case of the system studied in [3], when the continuous parameter in Sβ is absent.
The resulting infinite-dimensional process is of extensive interest (see [3], Sect. 2,
for references).
5.3 Locally Interacting Particle Systems, Solutions in C([0, T ], RZ )
d
195

The fundamental problem is to study the dynamics corresponding to a given

Gibbs measure on RZ . We consider here a lattice system of locally interacting
d

diffusions and study weak solutions in C([0, T ], RZ ). The space Zd is equipped

d

with the Euclidian norm and RZ has the metric for coordinate-wise convergence,
d

similar as in Sect. 5.2.

We begin with the system of SDEs describing the lattice system of interest,
t    
X k (t) = F k X k (s) + q k X k (s) ds + Wtk . (5.11)
0

Here, for k ∈ Zd and x ∈ RZ , x k denotes the kth coordinate of x. The coefficient

d

F : RZ → RZ is defined by
d d

  
 1
F (x) = F (x) k∈Zd = −
k
a(k − j )x j
(5.12)
2 k∈Zd
j ∈BZd (k,ρ)

with BZd (k, ρ) denoting a sphere of radius ρ centered in k ∈ Zd .

Let l2 (Zd ) denote the Hilbert space of square-integrable sequences indexed by
elements of Zd . The “dynamical matrix” A = (ak,j )k,j ∈Zd ∈ L (l2 (Zd )) is lattice
translation invariant, and the interactions are local, i.e., ak,j = a(k −j ), and ak,j = 0
for |k − j |Zd > ρ.
The family of drifts q(x) = {qk (x k )}k∈Zd : RZ → RZ is in general a singular
d d

mapping on the scales of the Hilbert space l2 (Zd ). The functions qk : R → R are
the derivatives of potentials Vk (u). In [3], Vk (u) = λP (u) with P (u) as in (5.10),
which is the case in an important class of the so-called P (ϕ) models.
We note that
  
 k  1  j 2 1/2
F (x) ≤ A x , (5.13)
2
j ∈BZd (k,ρ)

where
  1/2
A = a (k − j )
2
.
j ∈BZd (k,ρ)

The assumptions we impose on functions qk : R → R are the same as in Sect. 5.2.

Denote
  
  2n  k 2
l2n = l2n Zd = x ∈ RZ :
d
1 + |k|Zd x <∞ ,
k∈Zd
∞
3 ∞
&
Φ= l2n → · · · → l21 → l2 = l20 → l2−1 → · · · → l2−n =: Φ  .
n=1 n=1

The embeddings between the Hilbert spaces l2n → l2m with m + d2 < n are compact
(in fact, Hilbert–Schmidt) operators. The space Φ endowed with the projective limit
196 5 Stochastic Differential Equations with Discontinuous Drift

topology is a nuclear space of fast decreasing sequences, and the space Φ  , endowed
with the inductive limit topology, is dual to Φ.
Let Q be a continuous quadratic form on Φ and denote its extension (which
always exists) to a nuclear form on some l2−m , m > 0, by the same symbol.

Theorem 5.4 Let F : RZ → RZ be as in (5.12), defined by a lattice translation-

d d

invariant “dynamical matrix” A, and the drifts q k : R → R, k ∈ Zd , be continuous

and satisfy conditions (5.3) and (5.4). Then there exists a weak solution X(·) ∈
−p
C([0, T ], l2 ), for some p > 0, to the equation
t   
X(t) = x + F X(s) + q X(s) ds + Wt , x ∈ l2 , (5.14)
0
−p
in the following sense. There exist an l2 -valued Q-Wiener process Wt and a pro-
−p
cess X(·) ∈ C([0, T ], l2 ) such that for every k ∈ Zd ,
t    
X k (t) = x k + F k X(s) + q k X k (s) ds + Wtk . (5.15)
0

Proof Let us show that for m > 0, F : l2−m → l2−m is Lipschitz continuous. If x, y ∈
l2−m , m > 0, then
   −2m  k 2
F (x) − F (y)2−m = 1 + |k|Zd F (x) − F k (y)
l 2
k∈Zd
   
−2m 1  j 2
≤ 1 + |k|Zd A2 x − yj
4
d
k∈Z j ∈BZd (k,ρ)

≤ C1 x − y2l −m .
2

Note that F and q k satisfy all conditions of Theorem 5.1. Following its proof,
we first construct solutions ξn of (5.6). Let |Bn | denote the cardinality of the ball
BZd (0, n) of radius n centered at 0 in Zd . Observe that
 
ξn ∈ C [0, T ], R|Bn | .
 j
Next, we obtain approximations Xn (t) = j ∈B d ξn hj , where we denote by
Z (0,n)
{hk }k∈Zd the canonical basis in l2 . We have
 
Xn ∈ C [0, T ], l2 .

Since for m > d/2, the embedding l2 → l2−m is compact and

 · l −m ≤ C · l2 ,
2

Lemma 3.14 guarantees that the measures μn = L (Xn ) are tight on C([0, T ], l2−m ).
5.3 Locally Interacting Particle Systems, Solutions in C([0, T ], RZ )
d
197

Let μ be a limit point of the sequence μn . Using the Skorokhod theorem, we can
assume that
 
Xn → X, P -a.s. in C [0, T ], l2−m .
t
The random variables (Xnk (t))2 and 0 [Fnk (Xn (s))+q k (Xnk (s))]2 ds are P -uniformly
integrable. By the same arguments as in the proof of Theorem 5.1, we conclude that
the sequence of Brownian motions
t*    +
Ynk (t) = Xnk (t) − x k − Fnk Xn (s) + q k Xnk (s) ds
0

converges in L2 (Ω, P ) to the Brownian motion

t*    +
Y k (t) = X k (t) − x k − F k X(s) + q k X k (s) ds,
0

and we define the l2−m -valued Brownian motion


Wt = Y k (t)h−m
k ,
k∈Zd

where {h−m
k }k∈Zd is the basis in l2
−m
obtained by applying the Gramm–Schmidt
orthonormalization of the vectors hk . Then X(t), Wt satisfy (5.15). 

When the assumptions on the drifts q k are more restrictive, (5.14) can be considered
in a Hilbert space. We will require that the functions q k (u) have linear growth.

Theorem 5.5 Let F : RZ → RZ be as in (5.12), defined by a lattice translation-

d d

invariant “dynamical matrix” A, and the drifts q k : R → R, k ∈ Zd , be continuous

and satisfy the linear growth condition
 k   
q (u) ≤ C 1 + |u| , u ∈ R.

Assume that B : Φ  → Φ  , B(x) ∈ L (l2−m ) if x ∈ l2−m for m > 0, and it is continu-

ous in x. In addition,
  ∗   
tr B(x)Q B(x) ≤ θ 1 + x2l −m .
2

−p
Then there exists a weak solution X(·) ∈ C([0, T ], l2 ), for some p > 0, to the
equation
t    t  
X(t) = x + F X(s) + q X(s) ds + B X(s) dWs , x ∈ l2 , (5.16)
0 0

−p
where Wt is an l2 -valued Q-Wiener process.
198 5 Stochastic Differential Equations with Discontinuous Drift

Proof If x ∈ l2−m , m > 0, then

   −2m  k  2
F (x) + q(x)2−m = 1 + |k|Zd F (x) + qk x k
l 2
k∈Zd
 −2m  k 2   k 2 
≤2 1 + |k|Zd F (x) + qk x
k∈Zd
   
−2m 1  2   2
≤2 1 + |k|Zd A2 xj + C 2 1 + x k 
4
k∈Zd j ∈BZd (k,ρ)
 
≤ C1 1 + x2l −m ,
2

showing that G(x) = F (x) + q(x) : l2−m → l2−m , m > d/2 (otherwise, we face a
divergent series, see Exercise 5.3), and providing the estimate
   
G(x)2−m ≤ C1 1 + x2−m .
l 2 l 2

Moreover, we know from the proof of Theorem 5.4 that F : l2−m → l2−m , m > 0, is
Lipschitz continuous.
We will now use the approach presented in Sect. 3.8. With {ek−r }k∈Zd , r > 0,
denoting the ONB in l2−r consisting of the eigenvectors of Q, let Pn : l2−r → l2−r be
defined by
  
Pn x = x, ek−r l −r ek−r .
2
k∈BZd (0,n)

Note that for 0 < m < r and Pn : l2−m → l2−m ,

  
Pn x = x, ek−m l −m ek−m , x ∈ l2−m .
2
k∈BZd (0,n)

Let us consider a weak solution of the equation

   
dXn (t) = Pn G Pn Xn (t) dt + dPn B Pn Xn (t) Wtn (5.17)

with a Pn QPn -Wiener process Wt in l2−m .

Equation (5.17) can be considered in an |Bn |-dimensional subspace of l2−m (|Bn |
denotes the cardinality of BZd (0, n)). In addition, we have the following estimates:
   
Pn G(Pn x)2−m ≤ C1 1 + Pn x2−m ,
l 2 l 2
  ∗   
tr Pn B(Pn x)(Pn QPn ) Pn B(Pn x) ≤ θ 1 + Pn x2l −m .
2

Since Pn Hl −m is finite-dimensional, the embedding J : Pn l2−m → Pn l2−m is com-

2
pact, and we conclude by Theorem 3.12 (or by Theorem 4.1) that (5.17) has a weak
5.3 Locally Interacting Particle Systems, Solutions in C([0, T ], RZ )
d
199

solution Xn (t) in the space C([0, T ], Pn l2−m ) satisfying the inequalities

 2j
E sup Xn (t)l −m < C  , j ≥ 1,
0≤t≤T 2
(5.18)
 4
E Xn (t + h) − Xn (t)l −m ≤ C  h2 ,
2

with the constant C

independent of n. Consider now the sequence of measures on
C([0, T ], l2−m ) induced by {Xn (t)}t∈[0,T ] and denote them by μn . Again, using the
−p
fact that for p > m + d2 , the embedding l2−m → l2 is Hilbert–Schmidt, we obtain
−p
the relative weak compactness of the sequence μn on C([0, T ], l2 ) for some p > 0.
−p −p
Because Xn (t) ∈ C([0, T ], l2 ), there exists X(t) ∈ C([0, T ], l2 ) such that
−p
Xn ⇒ X in C([0, T ], l2 ). By Skorokhod’s theorem, we can assume that Xn → X
−p −p
a.s. For x ∈ l2 , using the continuity of F on l2 and uniform integrability, we
obtain
  t      t   
Xn (t), x −p
l2
− F Xn (s) , x l −p ds → X(t), x l −p − F X(s) , x l −p ds
0 2 2 0 2

−p
a.s. and in L1 (Ω, P ). Next, let {xl }∞
j
l=1 be a sequence converging to x in l2 , xl = 0,
j∈/ BZd (0, l). We consider
                 
E  q Xn (t) , x l −p − q X(t) , x l −p  ≤ E  q Xn (t) , x l −p − q Xn (t) , xl l −p 
2 2 2 2
                 
+  q Xn (t) , xl l −p − q X(t) , xl l −p  +  q X(t) , xl l −p − q X(t) , x l −p 
2 2 2 2
      j   j  j 
 
≤ E q Xn (t) l −p x − xl l −p + E  q Xn (t) − q X (t) xl
j j   
2 2
j ∈BZd (0,l)
  
+ E q X(t) l −p xl − xl −p
2 2
     
≤ C E sup 1 + Xn (t) −p + X(t) −p x − xl  −p
l2 l2 l2
n
  j j    
+E q Xn (t) − q j X j (t) x j .
l
j ∈BZd (0,l)

Using the estimate in (5.18), we can choose l, independent of n, such that the first
summand is arbitrarily small. By choosing n large enough and using the continuity
of qk on R, we can make the second summand arbitrarily small. Using the uniform
integrability for the term involving q, we conclude that
  t   
Mn (t) = Xn (t), x l −p − G Xn (s) , x l −p ds
2 0 2

  t   
→ X(t), x l −p − G X(s) , x l −p ds = M(t) (5.19)
2 0 2

in L1 (Ω) and a.s. for some subsequence.

200 5 Stochastic Differential Equations with Discontinuous Drift

Since
        2 
tr Pn B Xn (s) Pn QPn B ∗ Xn (s) Pn ≤ θ 1 + Xn (s)l −p ,
2

the LHS is uniformly integrable with respect to the measure dP × ds, and it con-
verges dP × ds-a.e. to tr(B(X(s))QB ∗ (X(s))) (see Exercise 5.4), implying that
t      
E Pn QPn B ∗ Xn (s) Pn x, B ∗ Xn (s) Pn x ds
0
t      
→E Q B ∗ X(s) x, B ∗ X(s) x ds,
0

i.e., Mn t → M t in L1 (Ω). In conclusion, the process on the right-hand side

of (5.19) is a real continuous square-integrable martingale whose increasing process
is given by
t      
Q B ∗ X(s) x, B ∗ X(s) x ds,
0
An appeal to Lemma 2.1 and to the martingale representation theorem (Theo-
rem 2.7) completes the proof. 

Exercise 5.3 Show that for p > d/2, k∈Zd (1 + |k|)2p < ∞, and the series di-
verges otherwise.

Exercise 5.4 Show the convergence

          
tr Pn B Xn (s) Pn QPn B ∗ Xn (s) Pn → tr B X(s) QB ∗ X(s)

claimed in the proof of Theorem 5.5.

 Part II
Stability, Boundedness, and Invariant
Measures
Chapter 6
Stability Theory for Strong and Mild Solutions

6.1 Introduction
Let (X,  · X ) be a Banach space, and let us consider the Cauchy problem
⎧
⎨ du(t)
= Au(t), 0 < t < T ,
dt (6.1)
⎩u(0) = x ∈ X.

We know that if A generates a C0 -semigroup {S(t), t ≥ 0}, then the mild solution
of the Cauchy problem (6.1) is given by

ux (t) = S(t)x.

If X is finite-dimensional, with a scalar product ·, · X, Lyapunov proved the equiv-

alence of the following three conditions:
(1) ux (t)X ≤ c0 xX e−rt , r, c0 > 0.
(2) max{Re(λ) : det(λI − A) = 0} < 0.
(3) There exists a positive definite matrix R satisfying
(i) c1 x2X ≤ Rx, x X ≤ c2 x2X , x ∈ X, c1 , c2 > 0,
(ii) A∗ R + RA = −I .
If condition (1) is satisfied, then the mild solution {ux (t), t ≥ 0} of the Cauchy
problem (6.1) is said to be exponentially stable.
To prove that (1) implies (3), the matrix R is constructed using the equation
∞ 
Rx, x X = ux (t)2 dt.
X
0

When X is infinite-dimensional, then the interesting examples of PDEs result in an

unbounded operator A. In this case, if we replace condition (2) by
(2’) max{Re(λ) : λ ∈ σ (A)} < 0,
with σ (A) denoting the spectrum of A, the equivalence of (1) and (2’) fails
((2’)  (1)) due to the failure of the spectral mapping theorem (refer to [63], p. 117),

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 203

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_6, © Springer-Verlag Berlin Heidelberg 2011
204 6 Stability Theory for Strong and Mild Solutions

unless we make more restrictive assumptions on A (e.g., A is analytic). A sufficient

condition for exponential stability is given in [63], p. 116:
∞ 
S(t)x p dt < ∞, for p > 1.
X
0

In our PDE examples, we need p = 2 and X = H , a real separable Hilbert space.

In this case, condition (1) alone implies that R, given by
∞ 
Rx, y H = ux (t), uy (t) H dt,
0

exists as a bilinear form, and in fact, the equivalence of conditions (1) and (3) above
can be proved (see [13]). We now consider the Cauchy problem in a real Hilbert
space,
⎧
⎨ du(t)
= Au(t), 0 < t < T ,
dt (6.2)
⎩u(0) = x ∈ H.

Theorem 6.1 Let (H, ·, · H) be a real Hilbert space. The following conditions are
equivalent:
(1) The solution of the Cauchy problem (6.2) {ux (t), t ≥ 0} is exponentially stable.
(2) There exists a nonnegative symmetric operator R such that for x ∈ D(A),

A∗ Rx + RAx = −x.

Proof Define Rx, y H as above. Using condition (1), we have

∞ 
Rx, x H = S(t)x 2 dt < ∞. (6.3)
H
0

Clearly, R is nonnegative definite and symmetric. Now, for x, y ∈ H ,

d     
RS(t)x, S(t)y = RAS(t)x, S(t)y H + RS(t)x, AS(t)y H .
dt
But
  ∞ 
RS(t)x, S(t)y H
= S(u)x, S(u)y H
du
t
by the semigroup property. Hence, we obtain
   
RAS(t)x, S(t)y H + RS(t)x, AS(t)y H
d ∞ 
= S(u)x, S(u)y H du
dt t
 
= − S(t)x, S(t)y H ,
6.1 Introduction 205

since S(t) is strongly continuous. Thus, if x ∈ D(A), then

  
RA + A∗ R x, y H = −x, y H ,

giving (2).
From the above calculations condition (2) implies that
d   2
RS(t)x, S(t)x H = −S(t)x H .
dt
Hence,
t   
S(u)x 2 du = Rx, x H − RS(t)x, S(t)x H
H
0
≤ Rx, x H.
∞
Thus, 0 S(t)x2H dt < ∞.
We know that S(t)x → 0 as t → ∞ for each x (see Exercise 6.1). Hence, by the
uniform boundedness principle, for some constant M, we have S(t)L (H ) ≤ M
for all t ≥ 0.
Consider the map T : H → L2 (R+ , H ), T x = S(t)x. Then T is a closed linear
operator on H . Using the closed graph theorem, we have
∞ 
S(t)x 2 dt ≤ c2 x2 .
H H
0

Let 0 < ρ < M −1 and define

  
tx (ρ) = sup t : S(s)x H > ρxH , for all 0 ≤ s ≤ t .

Since S(t)xH → 0 as t → ∞, we have that tx (ρ) < ∞ for each x ∈ H , tx (ρ) is

clearly positive, and
tx (ρ)  
tx (ρ)ρ 2 x2H ≤ S(t)x 2 dt ≤ c2 x2 ,
H H
0

giving tx (ρ) ≤ (c/ρ)2 = t0 .

For t > t0 , using the definition of tx (ρ), we have
        
S(t)x  ≤ S t − tx (ρ)  S tx (ρ) x 
H L (H ) H
≤ MρxH .

Let β = Mρ < 1 and t1 > t0 be fixed. For 0 < s < t1 , let t = nt1 + s. Then
     
S(t) ≤ S(nt1 )L (H ) S(s)L (H )
L (H )
 n
≤ M S(t1 )L (H ) ≤ Mβ n ≤ M  e−μt ,

where M  = M/β and μ = −(1/t1 ) log β > 0. 

206 6 Stability Theory for Strong and Mild Solutions

In particular, we have proved the following corollary.

Corollary 6.1 If S(·)x ∈ L2 (R+ , H ) for all x in a real separable Hilbert space H ,
then
 
S(t) ≤ c e−rt , for some r > 0.
L (H )
∞
Exercise 6.1 (a) Find a continuous function f (t) such that 0 (f (t))2 dt < ∞ but
limt→∞ f (t) = 0. 
∞
(b) Show that if 0 S(t)x2H dt < ∞ for every x ∈ H , then limt→∞ S(t)xH
= 0 for every x ∈ H .
Hint: recall that S(t)L (H ) ≤ Meαt . Assume that S(tj )xH > δ for some se-
quence tj → ∞. Then, S(t)xH ≥ δ(Me)−1 on [tj − α −1 , tj ].

We note that Rx, x H does not play the role of the Lyapunov function, since
in the infinite-dimensional case, Rx, x H ≥ c1 x2 with c1 > 0 does not hold (see
Example 6.1). We shall show that if A generates a pseudo-contraction semigroup,
then we can produce a Lyapunov function related to R. The function Λ in Theo-
rems 6.2 and 6.3 is called the Lyapunov function. Let us recall that {S(t), t ≥ 0} is
a pseudo-contraction semigroup if there exists ω ∈ R such that
 
S(t) ≤ eωt .
L (H )

Theorem 6.2 (a) Let {ux (t) t ≥ 0} be a mild solution to the Cauchy problem (6.2).
Suppose that there exists a real-valued function Λ on H satisfying the following
conditions:
(1) c1 x2H ≤ Λ(x) ≤ c2 x2H for x ∈ H ,
(2) Λ (x), Ax H ≤ −c3 Λ(x) for x ∈ D(A),
where c1 , c2 , c3 are positive constants. Then the solution ux (t) is exponentially
stable.
(b) If the solution {ux (t) t ≥ 0} to the Cauchy problem (6.2) is exponentially
stable and A generates a pseudo-contraction semigroup, then there exists a real-
valued function Λ on H satisfying conditions (1) and (2) in part (a).

Proof (a) Consider ec3 t Λ(ux (t)). We have

d  c3 t  x       
e Λ u (t) = c3 ec3 t Λ ux (t) + ec3 t Λ ux (t) , Aux (t) H .
dt
Hence,
  t       
ec3 t Λ ux (t) − Λ(x) = ec3 s c3 Λ ux (s) + Λ ux (s) , Aux (s) H ds.
0

It follows, by condition (2), that

 
ec3 t Λ ux (t) ≤ Λ(x).
6.1 Introduction 207

Using (1), we have

 2
c1 ux (t)H ≤ e−c3 t Λ(x) ≤ c2 e−c3 t x2H ,

proving (a).
(b) Conversely, we first observe that for Ψ (x) = Rx, x H with R defined
in (6.3), we have Ψ  (x) = 2Rx by the symmetry of R. Since R = R ∗ , we can write
   
Ψ  (x), Ax H
= Rx, Ax H + Rx, Ax H = A∗ Rx, x H + x, RAx H
 
= A∗ Rx + RAx, x H = −x2H .

Consider now, for some α > 0 (to be determined later),

Λ(x) = Rx, x H + αx2H .

Clearly Λ(x) satisfies condition (1) in (a). Since S(t) is a pseudo-contraction semi-
group, there exists a constant λ (assumed positive WLOG) such that (see Exer-
cise 3.5)
x, Ax H ≤ λx2H , x ∈ D(A). (6.4)
We calculate
   
Λ (x), Ax H
= Ψ  (x), Ax H + 2αx, Ax H = x2H (2αλ − 1).

Choosing α small enough, so that 2αλ < 1, and using condition (1), we obtain (2)
in (a). 

Let us now consider the case of a coercive operator A (see condition (6.5)), with
a view towards applications to PDEs. For this, we recall some concepts from Part I.
We have a Gelfand triplet of real separable Hilbert spaces

V → H → V ∗ ,

where the embeddings are continuous. The space V ∗ is the continuous dual of V ,
with the duality on V × V ∗ denoted by ·, · and satisfying

v, h = v, h H

if h ∈ H .
Assume that V is dense in H . We shall now construct a Lyapunov function for
determining the exponential stability of the solution of the Cauchy problem (6.2),
where A : V → V ∗ is a linear bounded operator satisfying the coercivity condition

2v, Av ≤ λv2H − αv2V , v ∈ V , λ ∈ R, α > 0. (6.5)

208 6 Stability Theory for Strong and Mild Solutions

We note that the following energy equality in [72] holds for solutions ux (t) ∈
L2 ([0, T ], V ) ∩ C([0, T ], H ):
 x 2 t 
u (t) − x2 = 2 ux (s), Aux (s) ds. (6.6)
H H
0

We now state our theorem.

Theorem 6.3 (a) The solution of the Cauchy problem (6.2) with a coercive coeffi-
cient A is exponentially stable if there exists a real-valued function Λ that is Fréchet
differentiable on H , with Λ and Λ continuous, locally bounded on H , and satisfy-
ing the following conditions:
(1) c1 x2H ≤ Λ(x) ≤ c2 x2H .
(2) For x ∈ V , Λ (x) ∈ V , and the function
 
V x → Λ (x), v ∗ ∈ R

is continuous for any v ∗ ∈ V ∗ .

(3) For x ∈ V , Λ (x), Ax ≤ −c3 Λ(x), where c1 , c2 , c3 are positive constants.
In particular, if
 
2 Λ (x), Ax H = −x2V ,
then condition (3) is satisfied.
(b) Conversely, if the solution to the Cauchy problem (6.2) is exponentially stable,
then the real-valued function
∞ 
Λ(x) = ux (t)2 dt (6.7)
V
0

satisfies conditions (1)–(3) in part (a).

Proof Note that for t, t  ≥ 0,

     t d  x 
Λ ux (t) − Λ ux t  = Λ u (s) ds.
t ds

But, using (2) and (3), we have

d  x    x    
Λ u (s) = Λ u (s) , Aux (s) ≤ −c3 Λ ux (s) .
ds
Denoting Φ(t) = Λ(ux (t)), we can then write

Φ  (t) ≤ −c3 Φ(t)

or, equivalently, d(Φ(t)ec3 t )/dt ≤ 0, giving Φ(t)ec3 t ≤ Φ(0).

6.1 Introduction 209

Using condition (1), we have

 2  
c1 ux (t)H ≤ Λ ux (t) ≤ Λ(x)e−c3 t ≤ c2 x2H e−c3 t .

To prove (b), we observe that, by the energy equality,

 x 2 t 
u (t) = x2 + 2 Aux (s), ux (s) ds
H H
0
t  t 
≤ x2H + |λ| ux (s)2 ds − α ux (s)2 ds.
H V
0 0

Hence,
 x 2 t  t 
u (t) + α ux (s)2 ds ≤ x2 + |λ| ux (s)2 ds.
H V H H
0 0

Letting t → ∞ and using the fact that

 x 2
u (t) ≤ cx2 e−γ t (γ > 0),
H H

we obtain
∞  

ux (s)2 ds ≤ 1 1 + |λ|c x2 .
V H
0 α 2γ
Define
∞ 
Λ(x) = ux (s)2 ds,
V
0

then Λ(x) ≤ c2 x2H . Let x, y ∈ H and consider

∞ 
T (x, y) = ux (t), uy (t) V dt.
0

Using the fact that ux (s) ∈ L2 ([0, ∞), V ) and the Schwarz inequality, we can see
that T (x, y) is a continuous bilinear form on V , which is continuous on H . Hence,
T (x, y) = C̃x, y H . Since Λ (x) = 2C̃x (by identifying H with H ∗ ), we can see
that Λ and Λ are locally bounded and continuous on H . By the continuity of the
embedding V → H , we have that for v, v  ∈ V , T (v, v  ) = Cv, v  V for some
bounded linear operator C on V , and property (2) in (a) follows. Now,
 x 2 t 
u (t) − x2 = 2 Aux (s), ux (s) ds.
H H
0

But |ux (s), Aux (s) | ≤ c2 ux (s)2V , giving

 x 2 t 
u (t) − x2 ≥ −2c ux (s)2 ds.
H H 2 V
0
210 6 Stability Theory for Strong and Mild Solutions

Let t → ∞; then ux (t)2H → 0, so that

−x2H ≥ −2c2 Λ(x),

implying Λ(x) ≥ c1 x2H for c1 = 1/(2c2 ).
It remains to prove that Λ(x) satisfies condition (3) in (a).
Note that
  ∞ x 
Λ ux (t) = uu (t) (s)2 ds.
V
0
By the uniqueness of solution,
x (t)
uu (s) = ux (t + s).
Hence,
  ∞  ∞ 
Λ ux (t) = ux (t + s)2 ds = ux (s)2 ds.
V V
0 t
Observe that
d  x    x 
Λ u (s) = Λ u (s), Aux (s) .
ds

Since the map Λ : V → H is continuous, we can write
  t    t 
Λ ux (t) − Λ(x) = Λ ux (s) , Aux (s) ds = − ux (s)2 ds.
V
0 0
By the continuity of the embedding V → H , we have xH ≤ c0 xV , x ∈ V ,
c0 > 0, and hence,
t    1 t 
Λ ux (s) , Aux (s) ds ≤ − 2 ux (s)2 ds.
H
0 c0 0

Now divide both sides by t and let t → 0. Since Λ is continuous and ux (·) ∈
C([0, T ], H ), we get
   1
Λ (x), Ax ≤ − 2 x2H .
c0 

The following example shows that in the infinite-dimensional case, if we define

∞ 
Λ(x) = ux (t)2 dt,
H
0

then Λ(x) does not satisfy the lower bound in condition (2) of (a) of Theorem 6.3.

Example 6.1 Consider a solution of the following equation

⎧ 2  
⎨d u(t, x) = a 2 ∂ u dt + b ∂u + cu dt,
t
∂x 2 ∂t (6.8)
⎩
u(0, x) = ϕ(x) ∈ L2 (R) ∩ L1 (R).
6.2 Exponential Stability for Stochastic Differential Equations 211

Here, H = L2 (R), and V is the Sobolev space W 1,2 (R). We denote by ϕ̂(λ) the
Fourier transform of ϕ(x) and use the similar notation û(t, λ) for the Fourier trans-
form of u(t, x). Then (6.8) can be written as follows:
⎧
⎨ dû(t, λ)
= −a 2 λ2 û(t, λ) + (ibλ + c)û(t, λ),
dt (6.9)
⎩û(0, λ) = ϕ̂(λ).

The solution is
 
ûϕ (t, λ) = ϕ̂(λ) exp −a 2 λ2 + ibλ + c t .
By Plancherel’s theorem, uϕ (t, ·)H = ûϕ (t, ·)H , so that
 ϕ  ∞   
u (t, ·)2 = ϕ̂(λ)2 exp −2a 2 λ2 + 2c t dλ
H
−∞
≤ ϕ2H exp{γ t} (γ = 2c).

For c < 0, we obtain an exponentially stable solution.

Take A = −2a 2 , B = 2c. Then
∞ ∞    ∞ |ϕ̂(λ)|2
Λ(ϕ) = ϕ̂(λ)2 exp − Aλ2 + B t dλ dt = dλ
0 −∞ −∞ Aλ2 + B

does not satisfy Λ(ϕ) ≥ c1 ϕ2H (see condition (1) in part (a) of Theorem 6.3).
In the next section, we consider the stability problem for infinite-dimensional
stochastic differential equations using the Lyapunov function approach. We shall
show that the fact that a Lyapunov function for the linear case is bounded below can
be used to study the stability for nonlinear stochastic PDEs.

6.2 Exponential Stability for Stochastic Differential Equations

We recall some facts from Part I. Consider the following stochastic differential equa-
tion in H :

dX(t) = (AX(t) + F (X(t))) dt + B(X(t)) dWt ,
(6.10)
X(0) = x ∈ H,

where
(1) A is the generator of a C0 -semigroup {S(t), t ≥ 0} on H .
(2) Wt is a K-valued Ft -Wiener process with covariance Q.
(3) F : H → H and B : H → L (K, H ) are Bochner-measurable functions satis-
fying
     
F (x)2 + tr B(x)QB ∗ (x) ≤ 1 + x2 ,
H H
      
F (x) − F (y)2 + tr B(x) − B(y) Q B(x) − B(y) ∗ ≤ K x − y2 .
H H
212 6 Stability Theory for Strong and Mild Solutions

Then (6.10) has a unique Ft -adapted mild solution (Chap. 3, Theorem 3.5), which
is a Markov process (Chap. 3, Theorem 3.6) and depends continuously on the initial
condition (Chap. 3, Theorem 3.7). That is, the integral equation
t   t  
X(t) = S(t)x + S(t − s)F X(s) ds + S(t − s)B X(s) dWs (6.11)
0 0
%
has a solution in C([0, T ], L2p ((Ω, F , P ), H )), p ≥ 1. Here F = σ ( t≥0 Ft ).
In addition, the solution of (6.11) can be approximated by solutions Xn obtained
by using Yosida approximations of the operator A in the following manner.
Recall from (1.22), Chap. 1, that for n ∈ ρ(A), the resolvent set of A, R(n, A) de-
notes the resolvent of A at n, and if Rn = nR(n, A), then An = ARn are the Yosida
approximations of A. The approximating semigroup is Sn (t) = etAn . Consider the
strong solution Xnx of

dX(t) = (An X(t) + F (X(t))) dt + B(X(t)) dWt ,
(6.12)
X(0) = x ∈ H.

Then Xnx ∈ C([0, T ], L2p ((Ω, F , P ), H )), p ≥ 1, by Theorem 3.5 in Chap. 3. By

Proposition 3.2 in Chap. 3, for p ≥ 1,
 2p 
lim sup E Xnx (t) − X x (t)H = 0, (6.13)
n→∞ 0≤t≤T

where X x (t) is the solution of (6.11).

We also recall the Itô formula, Theorem 2.9 in Chap. 2, for strong solutions
2
of (6.10). Let Cb,loc ([0, T ] × H ) denote the space of twice differentiable functions
Ψ : [0, T ] × H → R with locally bounded and continuous partial derivatives Ψt ,
Ψx , and Ψxx . Let X x (t) be a strong solution of (6.10), and Ψ ∈ Cb,loc
2 ([0, T ] × H ).
Then, with x ∈ D(A),

  t    
Ψ t, X x (t) − Ψ (0, x) = Ψt s, X x (s) + L Ψ s, X x (s) ds
0
t     
+ Ψx s, X x (s) , B X x (s) dWs H , (6.14)
0

where
  1  
L Ψ (t, x) = Ψx (t, x), Ax + F (x) H + tr Ψxx (t, x)B(x)QB ∗ (x) . (6.15)
2

Clearly (6.14) is valid for strong solutions of (6.12), with x ∈ H and A replaced by
An in (6.15).
We are ready to discuss the stability of mild solutions of (6.10).
6.2 Exponential Stability for Stochastic Differential Equations 213

Definition 6.1 Let {X x (t), t ≥ 0} be a mild solution of (6.10). We say that X x (t) is
exponentially stable in the mean square sense (m.s.s.) if for all t ≥ 0 and x ∈ H ,
 2
E X x (t)H ≤ ce−βt x2H , c, β > 0. (6.16)

2 (H ), with p ≥ 1, the subspace of C 2 (H ) con-

It is convenient to denote by C2p
sisting of functions f : H → R whose first two derivatives satisfy the following
growth condition:
    
f (x) ≤ Cx2p and f  (x) ≤ CxH
2p
H H L (H )

for some constant C ≥ 0.

Theorem 6.4 The mild solution of (6.10) is exponentially stable in the m.s.s. if there
exists a function Λ : H → R satisfying the following conditions:
(1) Λ ∈ C2p
2 (H ).

(2) There exist constants c1 , c2 > 0 such that

c1 x2H ≤ Λ(x) ≤ c2 x2H for all x ∈ H.

(3) There exists a constant c3 > 0 such that

L Λ(x) ≤ −c3 Λ(x) for all x ∈ D(A)

with L Λ(x) defined in (6.15).

Proof Assume first that the initial condition x ∈ D(A). Let Xnx (t) be the mild solu-
tion of Theorem 3.5 in Chap. 3 to the approximating equation

dX(t) = AX(t) + Rn F (X(t)) dt + Rn B(X(t)) dWt ,
(6.17)
X(0) = x ∈ D(A),

that is,
t   t  
Xnx (t) = S(t)x + S(t − s)Rn F Xnx (s) ds + S(t − s)Rn B Xnx (s) dWs
0 0

with Rn defined in (1.21). We note that (6.17) is an alternative to (6.12) in approx-

imating the mild solution of (6.10) with strong solutions. This technique preserves
the operator A and we have used it in the proof of Theorem 3.11.
Since x ∈ D(A) and Rn : H → D(A), the solution Xnx (t) ∈ D(A). Moreover,
since the initial condition is deterministic, Theorem 3.5 in Chap. 3 guarantees that
Xnx ∈ H˜2 . Then, the linear growth of B and the boundedness of Rn and S(t) im-
plies that the conditions of Theorem 3.2 are met, so that Xnx (t) is a strong solution
of (6.17). We apply Itô’s formula (6.14) to ec3 t Λ(Xnx (t)) and take the expectations,
to obtain
    t     
ec3 t EΛ Xnx (t) − Λ Xnx (0) = E ec3 s c3 Λ Xnx (s) + Ln Λ Xnx (s) ds,
0
214 6 Stability Theory for Strong and Mild Solutions

where
  1     ∗ 
Ln Λ(x) = Λ (x), Ax + Rn F (x) H + tr Λ (x) Rn B(x) Q Rn B(x) . (6.18)
2
By condition (3),
c3 Λ(x) + Ln Λ(x) ≤ −L Λ(x) + Ln Λ(x).
The RHS of the above equals to
  
Λ (x), (Rn − I )F (x) H
1 *   ∗  ∗ +
+ tr Λ (x) Rn B(x) Q Rn B(x) − B(x)Q B(x) .
2
Hence,
 
ec3 t EΛ Xnx (t) − Λ(x)

t   x   
≤E e c3 s
Λ Xn (s) , (Rn − I )F Xnx (s) H
0
1  *     ∗
+ tr Λ Xnx (s) Rn B Xnx (s) Q Rn B Xnx (s)
2

    ∗ +
− B Xnx (s) Q B Xnx (s) ds. (6.19)

In order to pass to the limit, we need to show that

 2
sup E Xnx (t) − X x (t)H → 0. (6.20)
0≤t≤T

Consider
 2
E Xnx (t) − X x (t)H
 t
   x   x 
≤ E  S(t − s) Rn F Xn (s) − F X (s) ds
0
2
t   x   x  
+ S(t − s) Rn B Xn (s) − B X (s) dWs 

0 H
  
 t      2
≤ C E
 S(t − s)Rn F Xnx (s) − F X x (s) ds 

0 H
t     
+E S(t − s)Rn B X x (s) − B X x (s) 2 ds
n L 2 (KQ ,H )
0
 
 t   2
+ E
 S(t − s)(Rn − I )F X (s) ds 
x

0 H
t 
 
+E S(t − s)(Rn − I )B X x (s) 2 ds .
L 2 (KQ ,H )
0
6.2 Exponential Stability for Stochastic Differential Equations 215
t
The first two summands are bounded by CK E 0 Xnx (s) − X x (s)2H for
n > n0 (n0 sufficiently large), where C depends on sup0≤t≤T S(t)L (H ) and
supn>n0 Rn L (H ) , and K is the Lipschitz constant.
By the properties of Rn , the integrand in the third summand converges to zero,
and, by (2.17) in Lemma 2.2, Chap. 2, the integrand in the fourth summand con-
verges to zero. Both integrands are bounded by C X x (s)2H for some constant C
depending on the norms of S(t) and Rn , similar as above, and the constant in the
linear growth condition. By the Lebesgue DCT, the third and fourth summands can
be bounded uniformly in t by εn (T ) → 0.
An appeal to Gronwall’s lemma completes the argument.
The convergence in (6.20) allows us to choose a subsequence Xnxk such that

Xnxk (t) → X x (t), 0 ≤ t ≤ T, P -a.s.

We will denote such a subsequence again by Xnx .

Now we use assumption (1), the continuity and local boundedness of Λ ,
the continuity of F , the uniform boundedness of Rn L (H ) , and the convergence
(Rn − I )x → 0 to conclude that
t     
E ec3 s Λ Xnx (s) , (Rn − I )F Xnx (s) H ds → 0
0

by the Lebesgue DCT. Now, using Exercise 2.19, we have

      ∗ 
tr Λ Xnx (s) Rn B Xnx (s) Q Rn B Xnx (s)
  ∗     
= tr Rn B Xnx (s) Λ Xnx (s) Rn B Xnx (s) Q
∞
         
= λj Λ Xnx (s) Rn B Xnx (s) fj , Rn B Xnx (s) fj H
j =1

with
        
Λ Xnx (s) Rn B Xnx (s) fj , Rn B Xnx (s) fj H
       
→ Λ X x (s) B X x (s) fj , B X x (s) fj H .

Hence,
      ∗ 
tr Λ Xnx (s) Rn B Xnx (s) Q Rn B Xnx (s)
      ∗ 
→ tr Λ X x (s) B X x (s) Q B X x (s) .

Obviously,
      ∗ 
tr Λ Xnx (s) B Xnx (s) Q B Xnx (s)
      ∗ 
→ tr Λ X x (s) B X x (s) Q B X x (s) .
216 6 Stability Theory for Strong and Mild Solutions

Now we use assumption (1), the continuity and local boundedness of Λ and Λ ,
the growth condition on F and B, and the fact that
 2
sup E Xnx (s)H < ∞,
0≤t≤T

and apply Lebesgue’s DCT to conclude that the right-hand side in (6.19) converges
to zero.
By the continuity of Λ and (6.20), we obtain
 
ec3 t EΛ X x (t) ≤ Λ(x),
and finally, by condition (2),
 2 c2
E X x (t)H ≤ e−c3 t x2H , x ∈ D(A). (6.21)
c1
We recall that the mild solution X x (t) depends continuously on the initial condition
x ∈ H in the following way (Lemma 3.7):
 2
sup E X x (t) − X y (t)H ≤ cT x − y2H , T > 0.
t≤T

Then for t ≤ T ,
 2  2  2
E X x (t)H ≤ E X y (t)H + E X x (t) − X y (t)H
c2 −c3 t
≤ e y2H + cT x − y2H
c1
c2 c2
≤ e−c3 t 2x − y2H + e−c3 t 2x2H + cT x − y2H
c1 c1
for all y ∈ D(A), forcing inequality (6.21) to hold for all x ∈ H , since D(A) is
dense in H . 

The function Λ defined in Theorem 6.4, satisfying conditions (1)–(3), is called a

Lyapunov function.
We now consider the linear case of (6.10) with F ≡ 0 and B(x) = B0 x, where
B0 ∈ L (H, L (K, H )), and B0 x ≤ dxH ,

dX(t) = AX(t) dt + B0 X(t) dWt ,
(6.22)
X(0) = x ∈ H.

Mild solutions are solutions of the corresponding integral equation

t
X(t) = S(t)x + S(t − s)B0 X(s) dWs . (6.23)
0

The concept of exponential stability in the m.s.s. for mild solutions of (6.22) obvi-
ously transfers to this case. We show that the existence of a Lyapunov function is a
necessary condition for stability of mild solutions of (6.22). The following notation
6.2 Exponential Stability for Stochastic Differential Equations 217

will be used:
  1  
L0 Ψ (x) = Ψ  (x), Ax H + tr Ψ  (x)(B0 x)Q(B0 x)∗ . (6.24)
2

Theorem 6.5 Assume that A generates a pseudo-contraction semigroup of opera-

tors {S(t), t ≥ 0} on H and that the mild solution of (6.22) is exponentially stable
in the m.s.s. Then there exists a function Λ0 (x) satisfying conditions (1) and (2) of
Theorem 6.4 and the condition
(3’) L0 Λ0 (x) ≤ −c3 Λ0 (x), x ∈ D(A), for some c3 > 0.

Proof Let
∞  2
Λ0 (x) = E X x (t)H dt + αx2H ,
0
where the value of the constant α > 0 will be determined later. Note that X x (t)
depends on x linearly. The exponential stability in the m.s.s. implies that
∞  2
E X x (t)H dt < ∞.
0

Hence, by the Schwarz inequality,

∞  
T (x, y) = E X x (t), X y (t) H dt
0

defines a continuous bilinear form on H × H , and there exists a symmetric bounded

linear operator T̃ : H → H such that
∞  2
T̃ x, x H = E X x (t)H dt.
0

Let
Ψ (x) = T̃ x, x H.

Using the same arguments, we define bounded linear operators on H by

  t  2
T̃ (t)x, x H
= E X x (s)H ds.
0

Consider solutions {Xnx (t), t ≥ 0} to the following equation:


dX(t) = An X(t) dt + B0 X(t) dWt ,
X(0) = x ∈ H,

obtained using the Yosida approximations of A. Just as above, we have continu-

ous bilinear forms Tn , symmetric linear operators T̃n (t), and real-valued continuous
218 6 Stability Theory for Strong and Mild Solutions

functions Ψn (t), defined for Xn ,

t  y 
Tn (t)(x, y) = E Xnx (u), Xn (u) H du,
0
  t  2
T̃n (t)x, x H = E Xnx (u)H du,
0
 
Ψn (t)(x) = T̃n (t)x, x H .

We have
 
  t  y 2 
Ψn (t) Xnx (s) = E Xn (u)H du  .
0 y=Xnx (s)

Let ϕ : H → R, ϕ(h) = h2H , and

   
P̃t ϕ (x) = Eϕ X x (t)

be the transition semigroup. Using the uniqueness of the solution, the Markov prop-
erty (3.59) yields
  t  
EΨn (t) Xnx (s) = E (P̃u ϕ) Xnx (s) du
0
t    Xx 
=E E ϕ Xnx (u + s) Fs n du
0
t  2
= E Xnx (u + s)H du
0
= Ψn (t + s)(x) − Ψn (s)(x). (6.25)

With t and n fixed, we use the Itô formula for the function Ψn (t)(x), then take the
expectation of both sides to arrive at
   s   
E Ψn (t) Xnx (s) = Ψn (t)(x) + E Ln Ψn (t) Xnx (u) du, (6.26)
0

where
   
Ln Ψn (t)(x) = 2 T̃n (t)x, An x H + tr T̃n (t)(B0 x)Q(B0 x)∗ .
Putting (6.25) and (6.26) together, we have
s   
Ψn (t + s)(x) − Ψn (s)(x) = E Ln Ψn (t) Xnx (u) du + Ψn (t)(x).
0

Rearranging the above, dividing by s, and taking the limit as s → 0 give

Ψn (t + s)(x) − Ψn (t)(x) 1 s    Ψn (s)(x)
= E Ln Ψn (t) Xnx (u) du + . (6.27)
s s 0 s
6.2 Exponential Stability for Stochastic Differential Equations 219

We fix n and t, and intend to take the limit in (6.27) as s → 0.

The processes Xnx (u) are continuous in the mean-square, since
 2
E Xnx (u) − Xnx (v)H
  v  2
≤ 2 An 2L (H ) + B0 2L (H,L (K,H )) tr(Q) E Xnx (r)H dr.
u

Hence,
Ψn (s)(x) 1 s  2
lim = lim E Xnx (u)H du = x2H . (6.28)
s→0 s s→0 s 0
Now consider
 
ELn Ψn (t) Xnx (u)
         ∗ 
= E 2 T̃n (t)Xnx (u), An Xnx (u) H + E tr T̃n (t) B0 Xnx (u) Q B0 Xnx (u) .

Since
lim An Xnx (u) = An x, lim T̃n (t)Xnx (u) = T̃n (t)x,
u→0 u→0
and
     2
 T̃n (t)X x (u), An X x (u)  ≤ T̃n (t) An L (H ) Xnx (u)H ∈ L1 (Ω),
n n L (H )

the Lebesgue DCT gives

     
lim E 2 T̃n (t)Xnx (u), An Xnx (u) H = 2 T̃n (t)x, An x H .
u→0

For the term involving the trace, we simplify the notation and denote
j
Φn (u) = B0 Xnx (u) and xn (u) = Φn (u)fj ,

where {fj }∞j =1 is an ONB in K that diagonalizes the covariance operator Q. Using
Exercise 2.19, we have
  ∗   ∗ 
tr T̃n (t)Φn (u)Q Φn (u) = tr Φn (u) T̃n (t)Φn (u)Q
∞
  
= λj T̃n (t)Φn (u)fj , Φn (u)fj H
j =1
∞
  j j 
= λj T̃n (t)xn (u), xn (u) H
j =1
∞
 t  j 2
= λj E Xnxn (u) (s)H ds. (6.29)
j =1 0
220 6 Stability Theory for Strong and Mild Solutions

Denote x j = (B0 x)fj . Since B0 is continuous, as u → 0, B0 Xnx (u) → B0 x in

j
L (K, H ), so that xn (u) → x j in H . By the continuity of the solution Xn with
respect to the initial condition (Chap. 3, Lemma 3.7),
 j 2
sup E Xnxn (u) (s) − Xnx (s)H → 0 as u → 0,
j

0≤s≤T

so that, by Lebesgue’s DCT and by reversing the calculations in (6.29),

∞
 t  j 2
λj E Xnxn (u) (s)H ds
j =1 0

∞
 t  j 2  
→ λj E Xnx (s)H ds = tr T̃n (t)(B0 x)Q(B0 x)∗ .
j =1 0

Summarizing, we proved that

dΨn (t)(x)
= Ln Ψn (t)(x) + x2H .
dt

In the next step, we fix t and allow n → ∞. By the mean-square continuity of Xnx (t)
and the definition of T̃n (t)x, x H and T̃ (t)x, x H , we can calculate the derivatives
below, and the convergence follows from condition (6.13):

dΨn (t)(x)  2  2 dΨ (t)(x)

= E Xnx (t)H → E X x (t)H = .
dt dt

Now, we need to show that as n → ∞, for x ∈ D(A),

   
Ln T̃n (t)x, x H → L0 T̃ (t)x, x H . (6.30)

Consider
    
 T̃n (t)x, An x − T̃ (t)x, Ax 
H H
       
≤ T̃n (t)x H (An − A)x H +  T̃n (t) − T̃ (t) x, Ax H  → 0.

Since (6.13) implies that

T 
lim E X x (u) − X x (u)2 du = 0, (6.31)
n→∞ n H
0

we thus have the weak convergence of Tn (t)x to T (t)x, and, further, by the Banach–
Steinhaus theorem, we deduce that supn Tn (t)L (H ) < ∞. Using calculations sim-
6.2 Exponential Stability for Stochastic Differential Equations 221

ilar as in (6.29), we have

∞

   
tr T̃n (t)(B0 x)Q(B0 x)∗ = λj T̃n (t)x j , x j H
j =1
∞
  
→ λj T̃ (t)x j , x j H
j =1
 
= tr T̃ (t)B0 xQ(B0 x)∗ ,

by Lebesgue’s DCT, proving the convergence in (6.30). Summarizing, we have

dT̃ (t)x, x H  
= L0 T̃ (t)x, x H + x2H .
dt
We will now let t → ∞. Then, by the exponential stability condition,

dT̃ (t)x, x  2
= E X x (t)H → 0.
H
dt

Since T̃ (t)x, x H → T̃ x, x H , using the weak convergence of T̃ (t)x to T̃ x and
the Lebesgue DCT, exactly as above, we obtain that
     
L0 T̃ (t)x, x H = 2 T̃ (t)x, Ax H + tr T̃ (t)B0 xQ(B0 x)∗
 
→ 2T̃ x, Ax H + tr T̃ B0 xQ(B0 x)∗ = L0 T̃ x, x H .

In conclusion,
L0 Ψ (x) = −x2H , x ∈ D(A).
Now, Λ0 satisfies conditions (1) and (2). To prove condition (3’), let us note that, as
in Sect. 6.1, since S(t) ≤ eωt , inequality (6.4) is valid for some constant λ > 0.
Hence,
   
L0 x2H = 2x, Ax H + tr (B0 x)Q(B0 x)∗ ≤ 2λ + d 2 tr Q x2H (6.32)

gives
 
L0 Λ0 (x) ≤ −x2H + α 2λ + d 2 tr(Q) x2H ≤ −c3 Λ0 (x),
c3 > 0, by choosing α small enough. 

Remark 6.1 For the nonlinear equation (6.10), we need to assume F (0) = 0 and
B(0) = 0 to assure that zero is a solution. In this case, if the solution {X x (t), t ≥ 0}
is exponentially stable in the m.s.s., we can still construct
∞  2
Λ(x) = E X x (t)H dt + αx2H .
0
222 6 Stability Theory for Strong and Mild Solutions

We however do not know if it satisfies condition (1) of Theorem 6.4. If we assume

that it does, then one can show, as in Theorem 6.5, that it satisfies condition (2).
Then we can prove that Λ(x) also satisfies condition (3).
First, observe that for Ψ (x) = Rx, x H , as before,

L Ψ (x) = −x2H

and

L Λ(x) = L Ψ (x) + αL x2H

    
= −x2H + α 2 x, Ax + F (x) H + tr B(x)QB ∗ (x) ,

noting the form of the infinitesimal generator L of the Markov process X x (t). We
obtain
    
L Λ(x) ≤ −x2H + 2αλx2H + α 2 x, F (x) H + tr B(x)QB ∗ (x) .

Now using the fact that F (0) = 0, B(0) = 0, and the Lipschitz property of F and B,
we obtain
 
L Λ(x) ≤ −x2H + α 2λ + 2K + K 2 tr(Q) x2H .

Hence, for α small enough, condition (3) follows from condition (2).

As shown in Part I, the differentiability with respect to the initial value requires
stringent assumptions on the coefficients F and B. In order to make the result more
applicable, we provide another technique that uses first-order approximation. We
use trace norm of a difference of nonnegative definite operators in the approximation
condition. Recall that for any trace-class operator T , we defined the trace norm
in (2.1) by
 1/2 
τ (T ) = tr T T ∗ .

Note (see [68]) that for a trace-class operator T and a bounded operator S,
(a) |tr(T )| ≤ τ (T ),
(b) τ (ST ) ≤ Sτ (T ) and τ (T S) ≤ Sτ (T ).

Theorem 6.6 Assume that A generates a pseudo-contraction semigroup of oper-

ators {S(t), t ≥ 0} on H . Suppose that the solution {X0x (t), t ≥ 0} of the linear
equation (6.22) is exponentially stable in the m.s.s. Then the solution {X x (t), t ≥ 0}
of (6.10) is exponentially stable in the m.s.s. if

    β
2xH F (x)H + τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗ ≤ x2H . (6.33)
2c
6.2 Exponential Stability for Stochastic Differential Equations 223

Proof Let Λ0 (x) = T̃ x, x H + αx2H , as in the proof of Theorem 6.5. Note that
∞
T̃ x, x H = E 0 X0x (t)2H dt, so that
∞ c
T̃ x, x H ≤ ce−βt x2H dt = x2H .
0 β

Hence, T̃ L (H ) ≤ c/β. Clearly Λ0 satisfies conditions (1) and (2) of Theorem 6.4.
It remains to prove that
L Λ0 (x) ≤ −c3 Λ0 (x).
Consider

L Λ0 (x) − L0 Λ0 (x)
  1   
= Λ0 (x), F (x) H + tr Λ0 (x) B(x)QB ∗ (x) − (B0 x)Q(B0 x)∗
2
    
≤ 2 (T̃ + α)x, F (x) H + τ (T̃ + α) B(x)QB ∗ (x) − (B0 x)Q(B0 x)∗
     
≤ T̃ L (H ) + α 2xH F (x)H + τ B(x)QB ∗ (x) − (B0 x)Q(B0 x)∗
 
1 β
≤ +α x2H .
2 2c
It follows that
 
1 β
L Λ0 (x) ≤ L0 Λ0 (x) + +α x2H
2 2c
 
  1 αβ
≤ −x2H + α 2λ + d 2 tr(Q) x2H + + x2H .
2 2c
For α small enough, we obtain condition (3) in Theorem 6.4 using condition (2). 

We now consider stability in probability of the zero solution of (6.10).

Definition 6.2 Let {X x (t)}t≥0 be the mild solution of (6.10) with F (0) = 0 and
B(0) = 0 (assuring that zero is a solution). The zero solution of (6.10) is called
stable in probability if for any ε > 0,
   
lim P supX x (t)H > ε = 0. (6.34)
xH →0 t≥0

Once a Lyapunov function satisfying conditions (1) and (2) of Theorem 6.4
is constructed, the following theorem provides a technique for proving condi-
tion (6.34).

Theorem 6.7 Let X x (t) be the solution of (6.10). Assume that there exists a func-
tion Ψ ∈ C2p
2 (H ) having the following properties:
224 6 Stability Theory for Strong and Mild Solutions

(1) Ψ (x) → 0 as xH → 0.

(2) infxH >ε Ψ (x) = λε > 0.
(3) L Ψ (x) ≤ 0, when x ∈ D(A) and xH < δ for some δ > 0.
Then, {X x (t), t ≥ 0} satisfies condition (6.34).

Proof The proof is similar to the proof of Theorem 6.4. We assume first that the
initial condition x ∈ D(A) and consider strong solutions Xnx (t) of the approximating
equations (6.17), n = 1, 2, . . . ,
t   t  
Xnx (t) = S(t)x + S(t − s)Rn F Xnx (s) ds + S(t − s)Rn B Xnx (s) dWs .
0 0

Denote Bε = {x ∈ H : x < ε} and let

     
τε = inf t : X x (t)H > ε and τεn = inf t : Xnx (t)H > ε .

Applying Itô’s formula to Ψ (Xnx (t)) and taking the expectations yield
   t∧τε  
Eψ Xnx t ∧ τεn − ψ(x) = E Ln Ψ Xnx (s) ds,
0

where
  1     ∗ 
Ln Ψ (x) = Ψ  (x), Ax + Rn F (x) H + tr Ψ  (x) Rn B(x) Q Rn B(x) .
2
Let ε < δ. Then for x ∈ Bε , using condition (3), we get

Ln Ψ (x) ≤ −L Ψ (x) + Ln Ψ (x)

 
= Ψ  (x), (Rn − I )F (x) H
1 *   ∗  ∗ +
+ tr Ψ  (x) Rn B(x) Q Rn B(x) − B(x)Q B(x) .
2
Hence,
 
EΨ Xnx (t ∧ τεn ) − Ψ (x)
t∧τεn   x   
≤E Ψ Xn (s) , (Rn − I )F Xnx (s) H
0
1  *     ∗
+ tr Ψ  Xnx (s) Rn B Xnx (s) Q Rn B Xnx (s)
2
    ∗ +
− B Xnx (s) Q B Xnx (s) ds. (6.35)

Using (6.20) and passing to the limit, as in the proof of Theorem 6.4, show that
the RHS in (6.35) converges to zero. Using condition (2), we conclude that for
x ∈ D(A) ∩ Bε and any n,
     
Ψ (x) ≥ E Ψ Xnx t ∧ τεn ≥ λε P τεn < t . (6.36)
6.3 Stability in the Variational Method 225

By the a.s. (and hence weak) convergence of τεn to τε , we have that

 
Ψ (x) ≥ λε lim inf P τεn < t ≥ λε P (τε < t).
n→∞

To remove the restriction that x ∈ D(A), recall that

 2
sup E X x (t) − X y (t)H → 0 as y − xH → 0.
0≤t≤T

We can select a sequence yn → x, yn ∈ D(A), such that X yn (t) → X x (t) a.s. for
all t. Now using the assumptions on Ψ and the Lebesgue DCT, we obtain (6.36) for
all x ∈ H . Inequality (6.36), together with conditions (2) and (1), implies that for
x ∈ Bε ,
    Ψ (x)
P supXtx H > ε ≤ → 0, xH → 0,
t≥0 λε
giving (6.34). 

The following results are now obvious from Theorems 6.5 and 6.6.

Theorem 6.8 Assume that A generates a pseudo-contraction semigroup of opera-

tors {S(t), t ≥ 0} on H . If the solution X0x (t) of the linear equation (6.22) is expo-
nentially stable in the m.s.s., then the zero solution of (6.22) is stable in probability.

Theorem 6.9 Assume that A generates a pseudo-contraction semigroup of opera-

tors {S(t), t ≥ 0} on H . If the solution X0x (t) of the linear equation (6.22) is ex-
ponentially stable in the m.s.s. and condition (6.33) holds for a sufficiently small
neighborhood of x = 0, then the zero solution of (6.10) is stable in probability.

We note that the exponential stability gives degenerate invariant measures. To

obtain nondegenerate invariant measures, we use a more general concept introduced
in Chap. 7.

6.3 Stability in the Variational Method

We consider a Gelfand triplet of real separable Hilbert spaces

V → H → V ∗ .

The space V ∗ is the continuous dual of V , V is dense in H , and all embeddings are
continuous. With ·, · denoting the duality between V and V ∗ , we assume that for
h ∈ H,
v, h = v, h H.
226 6 Stability Theory for Strong and Mild Solutions

Let K be a real separable Hilbert space, and Q a nonnegative definite trace-class

operator on K. We consider {Wt , t ≥ 0}, a K-valued Q-Wiener process defined on
a filtered probability space (Ω, F , {F }t≥0 , P ).
Let M 2 ([0, T ], V ) denote the space of all V -valued measurable processes satis-
fying
(1) u(t, ·) is Ft -measurable,
T
(2) E 0 u(t, ω)2V dt < ∞.
Throughout this section we consider the following equation:

dX(t) = A(X(t)) dt + B(X(t)) dWt ,
(6.37)
X(0) = x ∈ H,

where A and B are in general nonlinear mappings, A : V → V ∗ , B : V →

L (K, H ), and
   
A(v) ∗ ≤ a1 vV and B(v) ≤ b1 vV , v ∈ V , (6.38)
V L (K,H )

for some positive constants a1 , b1 .

We recall the coercivity and weak monotonicity conditions from Chap. 4, which
we impose on the coefficients of (6.37)
(C) (Coercivity) There exist α > 0, γ , λ ∈ R such that for all v ∈ V ,
   
2 v, A(v) + tr B(v)QB ∗ (v) ≤ λv2H − αv2V + γ . (6.39)

(WM) (Weak Monotonicity) There exists λ ∈ R such that for all u, v ∈ V ,

     ∗ 
2 u − v, A(u) − A(v) + tr B(u) − B(v) Q B(u) − B(v)
≤ λu − v2H . (6.40)

Since conditions (6.38), (6.39), and (6.40) are stronger than the assumptions in The-
orem 4.7 (also in Theorem 4.4) of Chap. 4, we conclude that there exists a unique
strong solution {X x (t), t ≥ 0} of (6.37) such that
    
X x (·) ∈ L2 Ω, C [0, T ], H ∩ M 2 [0, T ], V .

Furthermore, the solution X x (t) is Markovian, and the corresponding semigroup

has the Feller property.
The major tool we will use will be the Itô formula due to Pardoux [62]. It was
introduced in Part I, Sect. 4.2, Theorem 4.3, for the function Ψ (u) = u2H .

Theorem 6.10 (Itô Formula) Suppose that Ψ : H → R satisfies the following con-
ditions:

(1) Ψ is twice Fréchet differentiable, and Ψ , Ψ  , Ψ are locally bounded.
(2) Ψ and Ψ  are continuous on H .

(3) For all trace-class operators T on H , tr(T Ψ (·)) : H → R is continuous.
6.3 Stability in the Variational Method 227

(4) If v ∈ V , then Ψ  (v) ∈ V , and for any v  ∈ V ∗ , the function Ψ  (·), v  : V → R

is continuous.
(5) Ψ  (v)V ≤ c0 (1 + vV ) for some constant c0 > 0 and any v ∈ V .
Let X x (t) be a solution of (6.37) in L2 (Ω, C([0, T ], H )) ∩ M 2 ([0, T ], V ). Then
  t   t     
Ψ X x (t) = Ψ (x) + L Ψ X x (s) ds + Ψ  X x (s) , B X x (s) dWs H ,
0 0

where
  1  
L Ψ (u) = Ψ  (u), A(u) + tr Ψ  (u)B(u)QB ∗ (u) .
2
We extend the notion of exponential stability in the m.s.s. to the variational case.

Definition 6.3 We say that the strong solution of the variational equation (6.37) in
the space L2 (Ω, C([0, T ], H )) ∩ M 2 ([0, T ], V ) is exponentially stable in the m.s.s.
if it satisfies condition (6.16) in Definition 6.1.

The following is the analogue of Theorem 6.4 in the variational context. The
proof for a strong solution is a simplified version of the proof of Theorem 6.4 and is
left to the reader as an exercise.

Theorem 6.11 The strong solution of the variational equation (6.37) in the space
L2 (Ω, C([0, T ], H )) ∩ M 2 ([0, T ], V ) is exponentially stable in the m.s.s. if there
exists a function Ψ satisfying conditions (1)–(5) of Theorem 6.10, and the following
two conditions hold:
(1) c1 x2H ≤ Ψ (x) ≤ c2 x2H , c1 , c2 > 0, x ∈ H .
(2) L Ψ (v) ≤ −c3 Ψ (v), c3 > 0, v ∈ V , with L defined in Theorem 6.10.

Exercise 6.2 Prove Theorem 6.11.

We now consider the linear problem analogous to (6.37). Let A0 ∈ L (V , V ∗ )

and B0 ∈ L (V , L (K, H )). In order to construct a Lyapunov function directly from
the solution, we assume a more restrictive coercivity condition
(C’) (Coercivity) There exist α > 0, λ ∈ R such that for all v ∈ V ,
 
2v, A0 v + tr (B0 v)Q(B0 v)∗ ≤ λv2H − αv2V . (6.41)

We denote by L0 the operator L with A and B replaced by A0 and B0 . Consider

the following linear problem:

dX(t) = A0 X(s) ds + B0 (X(s)) dWs ,
(6.42)
X(0) = x ∈ H.

Theorem 6.12 Under the coercivity condition (6.41), the solution of the linear
equation (6.42) is exponentially stable in the m.s.s. if and only if there exists a
228 6 Stability Theory for Strong and Mild Solutions

function Ψ satisfying conditions (1)–(5) of Theorem 6.10 and conditions (1) and (2)
of Theorem 6.11.

Remark 6.2 A function Ψ satisfying conditions in Theorem 6.12 is called a Lya-

punov function.

Proof It remains to prove the necessity. By the Itô formula applied to x2H , taking
expectations, and using condition (6.41), we have
 2 t 
E X x (t)H = x2H + 2E A0 X x (s), X x (s) ds
0
t   ∗ 
+E tr B0 X x (s)Q B0 X x (s) ds
0
t  2 t  2
≤ x2H + |λ| E X x (s)H ds − α E X x (s)V ds
0 0
  t t  2
≤ x2H 1 + |λ|c e−βs ds − α E X x (s)V ds
0 0

by exponential stability in the m.s.s. Let t → ∞. Then

∞  2 1 
E X x (s)V ds ≤ 1 + |λ|c/β x2H .
0 α
Define
∞  
T (x, y) = E X x (s), X y (s) V ds.
0
Then, by the preceding inequality and the Schwarz inequality, it is easy to see
that T is a continuous bilinear form on H × H . Since the embedding V → H
is continuous, T is also a continuous bilinear form on V × V . This fact can be
used to show that conditions (1)–(5) of Theorem 6.10 are satisfied by the function
Ψ (x) = T (x, x). Clearly, Ψ (x) ≤ c2 x2H . To prove the lower bound on Ψ (x), we
observe that
 
L0 v2H = 2v, A0 v + tr (B0 v)Q(B0 v)∗ ,
so that, for some constants m, c0 ,
 
L0 v2  ≤ c0 v2 + m tr(Q)v2 ≤ c v2 .
H V V 0 V

Again, by Itô’s formula, after taking the expectations, we obtain that

 2 t  2
E X x (t)H − x2H = EL0 X x (s)H ds
0
t  2
≥ −c0 E X x (s)V ds.
0
6.3 Stability in the Variational Method 229

As t → ∞, using exponential stability in the m.s.s., we can see that

Ψ (x) ≥ c1 x2H ,

where c1 = 1/c0 .
To prove the last condition, observe that, similarly as in (6.25), the uniqueness of
the solution and the Markov property (3.59) yield
  ∞  2
EΨ X x (t) = E X x (s + t)V ds
0
∞  2
= E X x (s)V ds
t
∞  2 t  2
≤ E X x (s)V ds − k E X x (s)H ds,
0 0

since kx2H ≤ x2V for some constant k. Hence, by taking the derivatives of both
sides at t = 0, we get
k
L0 Ψ (x) ≤ −kx2H ≤ − Ψ (x). 
c2

Remark 6.3 Note that in case where t → EX x (t)2V is continuous at zero, in the
last step of the proof of Theorem 6.12, we obtain that L0 Ψ (v) = −v2V for v ∈ V .

Let us now state analogues of Theorem 6.6 for the solutions in variational case.

Theorem 6.13 Let {X0 (t)}t≥0 be the solution of the linear equation (6.42) with the
coefficients satisfying condition (6.41). Assume that the function t → EX0 (t)2V is
continuous and that the solution X0 (t) is exponentially stable in the m.s.s. If for a
sufficiently small constant c,
   
2vV A(v) − A0 v V ∗ + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ cv2V , (6.43)

then the strong solution of (6.37) is exponentially stable in the m.s.s.

For the zero solution of (6.37) to be stable in probability, it is enough to as-
sume (6.43) for v ∈ (V ,  · V ) in a sufficiently small neighborhood of zero.

Theorem 6.14 Let {X0 (t)}t≥0 be the solution of the linear equation (6.42) with the
coefficients satisfying condition (6.41). Assume that the solution X0 (t) is exponen-
tially stable in the m.s.s. Let for v ∈ V , A(v) − A0 v ∈ H . If for a sufficiently small
constant c,
   
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ cv2H , (6.44)

then the strong solution of (6.37) is exponentially stable in the m.s.s.

For the zero solution of (6.37) to be stable in probability, it is enough to as-
sume (6.44) for v ∈ (H,  · H ) in a sufficiently small neighborhood of zero.
230 6 Stability Theory for Strong and Mild Solutions

Exercise 6.3 Verify that Theorem 6.7 holds for 6.37 (replacing (6.10)), under ad-
ditional assumptions (1)–(5) of Theorem 6.10.

Exercise 6.4 Prove Theorems 6.13 and 6.14.

Remark 6.4 Using an analogue of Theorem 6.7 with the function Ψ satisfying
conditions (1)–(5) of Theorem 6.10, we can also prove conclusions in Theo-
rems 6.8 and 6.9 for (6.37) and its linear counterpart (6.42) under conditions (6.43)
and (6.44).

Appendix: Stochastic Analogue of the Datko Theorem

Theorem 6.15 Let A generate a pseudo-contraction C0 semigroup {S(t), t ≥ 0}

on a real separable Hilbert space H , and B : H → L (K, H ). A mild solution
{X x (t), t ≥ 0} of the stochastic differential equation (6.22) is exponentially stable
in the m.s.s. if and only if there exists a nonnegative definite operator R ∈ L (H )
such that

L0 Rx, y H = −x, y H for all x, y ∈ H,

where L0 is defined in (6.24).

Proof The necessity part was already proved in Sect. 6.2, Theorem 6.5, with
∞  
Rx, y H = E X x (t), X y (t) H dt,
0

which, under stability assumption, is well defined by the Schwarz inequality. To

prove the sufficiency, assume that R as postulated exists; then
  
2Rx, Ay H = − I + (R) x, y H , (6.45)

where (R) = tr(R(B0 x)Q(B0 x)∗ ) I . The operator I + (R) is invertible, so that
we get
  −1 
2 R I + (R) x, y H = x, y H .

By Corollary 6.1,
 
S(t) ≤ Me−λt , λ > 0.
L (H )

We consider the solutions {Xnx (t), t ≥ 0} obtained by using the Yosida approxima-
tions An = ARn of A. Let us apply Itô’s formula to RXnx (t), Xnx (t) H and take the
expectations of both sides to arrive at
Appendix: Stochastic Analogue of the Datko Theorem 231

  t 
E RXnx (t), Xnx (t) H = Rx, x H + 2E RXnx (s), An Xnx (s) H ds
0
t 
+E (R)Xnx (s), Xnx (s) H ds.
0

From (6.45) with y = Rn Xnx it follows that

     
2 RXnx (s), ARn Xnx (s) H = − (R)Xnx (s), Rn Xnx (s) H − Xnx (s), Rn Xnx (s) H .

Hence,
  t 
E RXnx (t), Xnx (t) H = Rx, x H −E Xnx (s), Rn Xnx (s) H ds
0
t 
+E (R)Xnx (s), Xnx (s) − Rn Xnx (s) H ds.
0

We let n → ∞ and use the fact that supn supt≤T EXnx (t)2H < ∞ to obtain

  t  2
E RX x (t), X x (t) H = Rx, x H − E X x (s)H ds.
0

Let Ξ (t) = ERX x (t), X x (t) H. Then

 2
Ξ (t) ≤ RL (H ) E X x (t)H

and
 2 −1
Ξ  (t) = −E X x (t)H ≤ Ξ (t),
RL (H )
so that
−1
RL (H ) t
Ξ (t) ≤ Rx, x He ,
since Ξ (0) = Rx, x H. Hence,
 2
 x 2  2  t 

E X (t) H ≤ 2 S(t)x H + 2E 
  
 S(t − s)BX (s) dWs 

x
0 H
t  2
≤ 2M 2 e−2λt x2H + 2 tr(Q)M 2 B2L (H ) e−2λ(t−s) E X x (s)H ds.
0

We complete the proof by using

 2
Rx, x H ≤ RL (H ) x2H and E X x (s)H = −Ξ  (s). 

As shown in Sect. 6.1, Example 6.1, x → Rx, x , however, is not a Lyapunov

function, so that we cannot study stability of nonlinear equations using Theo-
rem 6.15.
Chapter 7
Ultimate Boundedness and Invariant Measure

We introduce in this chapter the concept of ultimate boundedness in the mean square
sense (m.s.s.) and relate it to the problem of the existence and uniqueness of invari-
ant measure. We consider semilinear stochastic differential equations in a Hilbert
space and their mild solutions under the usual linear growth and Lipschitz condi-
tions on the coefficients. We also study stochastic differential equations in the varia-
tional case, assuming that the coefficients satisfy the coercivity condition, and study
their strong solutions which are exponentially ultimately bounded in the m.s.s.

7.1 Exponential Ultimate Boundedness in the m.s.s.

Definition 7.1 We say that the mild solution of (6.10) is exponentially ultimately
bounded in the mean square sense (m.s.s.) if there exist positive constants c, β, M
such that
 2
E X x (t)H ≤ ce−βt x2H + M for all x ∈ H. (7.1)

Here is an analogue of Theorem 6.4.

Theorem 7.1 The mild solution {X x (t), t ≥ 0} of (6.10) is exponentially ultimately

bounded in the m.s.s. if there exists a function Ψ ∈ C2p
2 (H ) satisfying the following

conditions:
(1) c1 x2H − k1 ≤ Ψ (x) ≤ c2 x2H − k2 ,
(2) L Ψ (x) ≤ −c3 Ψ (x) + k3 ,
for x ∈ H , where c1 , c2 , c3 are positive constants, and k1 , k2 , k3 ∈ R.

Proof Similarly as in the proof of Theorem 6.4, using Itô’s formula for the solutions
of the approximating equations (6.17) and utilizing condition (2), we arrive at
    t   
EΨ X x (t) − EΨ X x (0) ≤ −c3 EΨ X x (s) + k3 ds.
0

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 233

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0_7, © Springer-Verlag Berlin Heidelberg 2011
234 7 Ultimate Boundedness and Invariant Measure

Hence, Φ(t) = EΨ (X x (t)) satisfies

Φ  (t) ≤ −c3 Φ(t) + k3 .

By Gronwall lemma,
 
k3 k3 −c3 t
Φ(t) ≤ + Φ(0) − e .
c3 c3

Using condition (1), we have, for all x ∈ H ,

 
 x 2  x  k3 k3 −c3 t
 
c1 E X (t) H − k1 ≤ EΨ X (t) ≤ + c2 xH − k2 −
2
e ,
c3 c3

and (7.1) follows. 

Theorem 7.2 Assume that A generates a pseudo-contraction semigroup of oper-

ators {S(t), t ≥ 0} on H . If the mild solution {X0x (t), t ≥ 0} of the linear equa-
tion (6.22) is exponentially ultimately bounded in the m.s.s., then there exists a
function Ψ0 ∈ C2p2 (H ) satisfying conditions (1) and (2) of Theorem 7.1, with the

operator L0 replacing L in condition (2).

Proof Since the mild solution X0x (t) is exponentially ultimately bounded in the
m.s.s., we have
 2
E X0x (t)H ≤ ce−βt x2H + M for all x ∈ H.

Let
T  2
Ψ0 (x) = E X0x (s)H ds + αx2H ,
0
where T and α are constants to be determined later.
First, let us show that Ψ0 ∈ C2p
2 (H ). It suffices to show that

T  2
ϕ0 (x) = E X0x (s)H ds ∈ C2p
2
(H ).
0

Now,
c  c
ϕ0 (x) ≤ 1 − e−βT x2H + MT ≤ x2H + MT .
β β
If x2H = 1, then ϕ0 (x) ≤ c/β + MT .
Since X0x (t) is linear in x, we have that, for any positive constant k, X0kx (t) =
kX0x (t). Hence, ϕ0 (kx) = k 2 ϕ0 (x), and for any x ∈ H ,
   
x c
ϕ0 (x) = x2H ϕ ≤ + MT x2H .
xH β
7.1 Exponential Ultimate Boundedness in the m.s.s. 235

Let c = c/β + MT . Then φ0 (x) ≤ c x2H for all x ∈ H . For x, y ∈ H , define

T  y 
τ (x, y) = E X0x (s), X0 (s) H ds.
0

Then τ (x, y) is a nonnegative definite bounded bilinear form on H × H since

ϕ0 (x) ≤ c x2H . Hence, τ (x, y) = Cx, y H , where C is a nonnegative definite
bounded linear operator on H with CL (H ) ≤ c . Therefore, ϕ0 = Cx, x H ∈
2 (H ), and Ψ ∈ C 2 (H ). Clearly Ψ satisfies condition (1) of Theorem 7.1. To
C2p 0 2p 0
prove (2), observe that by the continuity of the function t → EX0x (t)2H and be-
cause
  T  2 T +r  2
Eϕ0 X0x (r) = E X0x (r + s)H ds = E X0x (s)H ds,
0 r

we have

d  
L0 ϕ0 (x) = Eϕ0 X0x (r) 
dr r=0
Eϕ0 (X0x (r)) − Eϕ0 (x)
= lim
r→0 r
 
1 r  2 1 r+T  2
= lim − E X0x (s)H ds + E X0x (s)H ds
r→0 r 0 r T
 
= −x2H + E X0x (T )H
2

≤ −x2H + ce−βT x2H + M

 
≤ −1 + ce−βT x2H + M.

Therefore, since by (6.32), L0 x2H ≤ (2λ + d 2 tr(Q))x2H , we have

L0 Ψ0 (x) = L0 ϕ0 (x) + L0 x2H

   
≤ −1 + ce−βT x2H + α 2λ + d 2 tr(Q) x2H + M. (7.2)

If T > ln(c/β), then one can choose α small enough such that Ψ0 (x) satisfies con-
dition (2) with L replaced by L0 . 

The following theorem is a counterpart of Remark 6.1 in the framework of expo-

nential ultimate boundedness.

Theorem 7.3 If the mild solution of (6.10) is exponentially ultimately bounded in

the m.s.s. and, for some T > 0,
T  2
ϕ(x) = E X x (t)H dt ∈ C2p
2
(H ),
0
236 7 Ultimate Boundedness and Invariant Measure

then there exists a (Lyapunov) function Ψ ∈ C2p

2 (H ) satisfying conditions (1)

and (2) of Theorem 7.1.

Theorem 7.4 Suppose that the mild solution X0x (t) of the linear equation (6.22)
satisfies condition (7.1). Then the mild solution X x (t) of (6.10) is exponentially
ultimately bounded in the m.s.s. if
   
2xH F (x)H + τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗ < ω̃x2H + M1 , (7.3)

where ω̃ < maxs>ln(c/β) (1 − ce−βs )/(c/β + Ms).

Proof Let Ψ0 (x) be the Lyapunov function as defined in Theorem 7.2, with T >
ln(c/β), such that the maximum in the definition of ω̃ is achieved. It remains to
show that
L Ψ0 (x) ≤ −c3 Ψ0 (x) + k3 .
Since Ψ0 (x) = Cx, x H + αx2H for some C ∈ L (H ) with CL (H ) ≤ c/β +
MT and α sufficiently small, we have

L Ψ0 (x) − L0 Ψ0 (x)
     
≤ CL (H ) + α 2xH F (x)H + τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗
 
≤ (c/β + MT + α) ω̃x2H + M1 .

Using (7.2), we have

   
L Ψ0 (x) ≤ −1 + ce−βT x2H + α 2λ + d 2 tr(Q) x2H + M
 
+ (c/β + MT + α) ω̃x2H + M1
 
≤ −1 + ce−βT + ω̃(c/β + MT ) x2H
 
+ α 2λ + d 2 tr(Q) + ω̃ x2H + M + (c/β + MT + α).

Using the bound for ω̃, we have −1 + ce−βT + ω̃(c/β + MT ) < 0, so that we can
choose α small enough to obtain condition (2) of Theorem 7.1. 

Corollary 7.1 Suppose that the mild solution X0x (t) of the linear equation (6.22) is
exponentially ultimately bounded in the m.s.s. If, as xH → ∞,
       
F (x) = o xH and τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗ = o xH ,
H

then the mild solution X x (t) of (6.10) is exponentially ultimately bounded in the
m.s.s.

Proof We fix ω̃ < maxs>ln(c/β) (1 − ce−βt /(c/β + Ms), and using the assumptions,
we choose a constant K such that for xH ≥ K, condition (7.3) holds. But for
7.2 Exponential Ultimate Boundedness in Variational Method 237

xH ≤ K, by appealing to the growth conditions on F and B,

   
2xH F (x)H + τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗
 2  
≤ x2H + F (x)H + τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗
   
≤ x2H + 1 + x2H + B0 x2L (H ) tr(Q)
 
≤ x2H + 1 + x2H + d 2 x2H tr(Q)
 
≤ K 2 + 1 + K 2 + M .

Hence, condition (7.3) holds with the constant M1 = K 2 + (1 + K 2 ) + M  , and the

result follows from Theorem 7.4. 

Example 7.1 (Dissipative Systems) Consider SSDE (6.10) and, in addition to as-
sumptions (1)–(3) in Sect. 6.2, impose the following dissipativity condition:
(D) (Dissipativity) There exists a constant ω > 0 such that for all x, y ∈ H and
n = 1, 2, . . . ,
     
2 An (x − y), x − y H + 2 F (x) − F (y), x − y H + B(x) − B(y)L (K ,H )
2 Q

≤ −ωx − y2H , (7.4)

where An x = ARn x, x ∈ H , are the Yosida approximations of A defined

in (1.22).
Then the mild solution to (6.10) is ultimately exponentially bounded in the m.s.s.
(Exercise 7.1).

Exercise 7.1 (a) Show that condition (D) implies that for any ε > 0, there exists a
constant Cε > 0 such that for any x ∈ H and n = 1, 2, . . . ,
   
2An x, x H + 2 F (x), x H + B(x)L (K ,H ) ≤ −(ω − ε)x2H + Cε
2 Q

with An , the Yosida approximations of A. Use this fact to prove that the strong solu-
tions Xnx (t) of the approximating SDEs (6.12) are ultimately exponentially bounded
in the m.s.s. Conclude that the mild solution X x (t) of (6.10) is ultimately exponen-
tially bounded in the m.s.s.
(b) Prove that if zero is a solution of (6.10), then the mild solution X x (t) of (6.10)
is exponentially stable in the m.s.s.

7.2 Exponential Ultimate Boundedness in Variational Method

We study in this section strong solutions to (6.37) whose coefficients satisfy linear
growth, coercivity, and monotonicity assumptions (6.38)–(6.40).
238 7 Ultimate Boundedness and Invariant Measure

Definition 7.2 We extend Definition 7.1 of exponential ultimate boundedness in the

m.s.s. to the strong solution {X x (t), t ≥ 0} of (6.37) and say that X x (t) is exponen-
tially ultimately bounded in the m.s.s. if it satisfies condition (7.1).

Let us begin by noting that the proof of Theorem 7.1 can be carried out in this
case if we assume that the function Ψ satisfies conditions (1)–(5) of Theorem 6.10
and that the operator L is defined by
   
L Ψ (u) = Ψ  (u), A(u) + tr Ψ  (u)B(u)QB ∗ (u) . (7.5)

Hence, we have the following theorem.

Theorem 7.5 The strong solution {X x (t), t ≥ 0} of (6.37) is exponentially ulti-

mately bounded in the m.s.s. if there exists a function Ψ : H → R satisfying condi-
tions (1)–(5) of Theorem 6.10 and, in addition, such that

(1) c1 x2H − k1 ≤ Ψ (x) ≤ c2 x2H + k2 for some positive constants c1 , c2 , k1 , k2

and for all x ∈ H ,
(2) L Ψ (x) ≤ −c3 Ψ (x) + k3 for some positive constants c3 , k3 and for all x ∈ V .

In the linear case, we have both, sufficiency and necessity, and the Lyapunov
function has an explicit form under the general coercivity condition (C).

Theorem 7.6 A solution {X0x (t), t ≥ 0} of the linear equation (6.42) whose coeffi-
cients satisfy coercivity condition (6.39) is exponentially ultimately bounded in the
m.s.s. if and only if there exists a function Ψ0 : H → R satisfying conditions (1)–(5)
of Theorem 6.10 and, in addition, such that

(1) c1 x2H − k1 ≤ Ψ0 (x) ≤ c2 x2H + k2 for some positive constants c1 , c2 , k1 , k2

and for all x ∈ H ,
(2) L0 Ψ0 (x) ≤ −c3 Ψ0 (x) + k3 for some positive constants c3 , k3 and for all x ∈ V .

This function can be written in the explicit form

T t  2
Ψ0 (x) = E X0x (s)V ds dt (7.6)
0 0

with T > α0 (c|λ|/(αβ) + 1/α), where α0 is such that v2H ≤ α0 v2V , v ∈ V .

Proof Assume that the solution {X0x (t), t ≥ 0} of the linear equation (6.42) is ex-
ponentially ultimately bounded in the m.s.s., so that
 2
E X0x (t)H ≤ ce−βt x2H + M for all x ∈ H.
7.2 Exponential Ultimate Boundedness in Variational Method 239

Applying Itô’s formula to the function x2H , taking the expectations, and using the
coercivity condition (6.39), we obtain
 2 t  2
E X0x (t)H − x2H = EL0 X0x (s)H ds
0
t  2 t  2
≤λ E X0x (s)H ds − α E X0x (s)V ds + γ t. (7.7)
0 0

Hence,
 
t  x 2 1 t  x 2
 
E X0 (s) V ≤ λ  
E X0 (s) H ds + xH + γ t .
2
0 α 0

Applying condition (7.1), we have

t   
2 1 c|λ|    
E X0x (s)V ≤ 1 − e−βt x2H + x2H + |λ|M + γ t
0 α β
 
c|λ| 1 |λ|M + γ
≤ + x2H + t.
αβ α α
Therefore, with Ψ0 defined in (7.6),
 
T t  2 1 c|λ| |λ|M + γ 2
Ψ0 (x) = E X0x (s)V ds dt ≤ + T x2H + T . (7.8)
0 0 α αβ 2α
Now
 
L0 v2  ≤ 2a1 v2 + b2 tr(Q)v2 ≤ c v2
H V 1 V V
for some positive constant c . Therefore, we conclude that

L0 v2H ≥ −c v2V .

From (7.7) we get

 2 t  2
E X0x (t)H − x2H ≥ −c E X0x (s)V ds.
0

Using (7.1), we have

t  2  
c E X0x (t)V ds ≥ 1 − e−βt x2H − M.
0

Hence,
 
1 T  −βt
 1 c MT
Ψ0 (x) ≥  x2H 1−e dt − MT ≥  T − x2H −  .
c 0 c β c
Choose T > c/β to obtain condition (1).
240 7 Ultimate Boundedness and Invariant Measure

To prove that condition (2) holds, consider

  T t  Xx (r) 2
EΨ0 X0x (r) = E X0 0 (s)V ds dt.
0 0

By the Markov property of the solution and the uniqueness of the solution,

  T t  2 T t+r  2
EΨ0 X0x (r) = E X0x (s + r)V ds dt = E X0x (s)V ds dt.
0 0 0 r

We now need the following technical lemma that will be proved later.

Lemma 7.1 If f ∈ L1 ([0, T ]), T > 0, is a nonnegative real-valued function, then

  t+Δt
T t+Δt f (s) ds T f (s) ds T
lim t
dt = lim t dt = f (t) dt.
Δt→0 0 Δt 0 Δt→0 Δt 0

Assuming momentarily that Ψ0 satisfies conditions (1)–(5) of Theorem 6.10, we

have

d  
L0 Ψ0 (x) = EΨ0 X0x (r) 
dr r=0

T t+r EX x (s)2 ds T r  2
= lim t 0 V
dt − lim E X0x (s)V ds
r→0 0 r r→0 r 0
T  2 T 1 r  2
≤ E X0x (t)V dt − lim E X0x (s)H ds
0 r→0 α0 r 0

for α0 such that v2H ≤ α0 v2V . This gives

 
c|λ| 1 T |λ|M + γ
L0 Ψ0 (x) ≤ + − x2H + T. (7.9)
αβ α α0 α

With T > α0 ( c|λ|

α β + α ), condition (2) holds.
1

It remains to prove that Ψ0 satisfies conditions (1)–(5) of Theorem 6.10. We use

linearity of (6.42) to obtain, for any positive constant k,

X0kx (t) = kX0x (t).

Then Ψ0 (kx) = k 2 Ψ0 (x), and by (7.8), for xH = 1,

 
1 c|λ| |λ|M + γ 2
Ψ0 (x) ≤ + T+ T .
α αβ 2α
Hence, for x ∈ H ,
  ,  -
x 1 c|λ| |λ|M + γ 2
Ψ0 (x) ≤ x2H Ψ0 ≤ + T+ T x2H , (7.10)
xH α αβ 2α
7.2 Exponential Ultimate Boundedness in Variational Method 241

which implies that Ψ0 (x) ≤ c x2H for all x ∈ H . For x, y ∈ H , denote

T t   1 1
E X0x (s), X0 (s) H ds dt ≤ Ψ02 (x)Ψ02 (y) ≤ c xH yH .
y
τ (x, y) =
0 0

Then τ is a continuous bilinear form on H × H , and there exists C ∈ L (H ), with

CL (H ) ≤ c , such that
τ (x, y) = Cx, y H. (7.11)
Using the continuity of the embedding V → H , we conclude that τ (x, y) is a con-
tinuous bilinear form on V × V , and hence,

τ (x, y) = C̃x, y V for x, y ∈ V , (7.12)

with C̃ ∈ L (V ). Now it is easy to verify that Ψ0 satisfies conditions (1)–(5) of

Theorem 6.10. 

Proof of Lemma 7.1 We are going to use the Fubini theorem to change the order of
integrals,
 t+Δt
T f (s) ds 1 T  t+Δt 
t
dt = f (s) ds dt
0 Δt Δt 0 0
, Δt  s  T T 
1
= f (s) dt ds + f (s) dt ds
Δt 0 0 Δt s−Δt
T +Δt  T  -
+ f (s) dt ds
T s−Δt
, Δt T T +Δt -
1
= sf (s) ds + f (s)Δt ds + f (s)(T + Δt − s) ds
Δt 0 Δt T
, Δt T T +Δt -
1
≤ Δt f (s) ds + Δt f (s) ds + Δt f (s) ds
Δt 0 Δt T
Δt T T +Δt
= f (s) ds + f (s) ds + f (s) ds.
0 Δt T

The first and third terms converge to zero as Δt → 0, so that

 t+Δt
T f (s) ds T
lim t
dt ≤ f (t) dt.
Δt→0 0 Δt 0

The opposite inequality follows directly from Fatou’s lemma. 

By repeating the proof of Theorem 7.6, we obtain a partial converse of Theo-

rem 7.5.
242 7 Ultimate Boundedness and Invariant Measure

Theorem 7.7 Let the strong solution {X x (t), t ≥ 0} of (6.37) be exponentially ul-
timately bounded in the m.s.s. Let
T t  2
Ψ (x) = E X x (s)V ds dt (7.13)
0 0

with T > α0 (c|λ|/(αβ) + 1/α), where α0 is such that v2H ≤ α0 v2V , v ∈ V . Sup-
pose that Ψ (x) satisfies conditions (1)–(5) of Theorem 6.10. Then Ψ (x) satisfies
conditions (1) and (2) of Theorem 7.5.

To study exponential ultimate boundedness, i.e., condition (7.1), for the strong
solution of (6.37), we use linear approximation and the function Ψ0 of the corre-
sponding linear equation (6.42) as the Lyapunov function. We will prove the fol-
lowing result.

Theorem 7.8 Suppose that the coefficients of the linear equation (6.42) satisfy
the coercivity condition (6.39) and its solution {X0x (t), t ≥ 0} is exponentially ul-
timately bounded in the m.s.s. Let {X x (t), t ≥ 0} be the solution of the nonlinear
equation (6.37). Furthermore, we suppose that

A(v) − A0 v ∈ H for all v ∈ V

and that, for v ∈ V ,

   
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ ω̃v2H + k,

where ω̃ and k are constants, and

c
ω̃ < * 1 c|λ|  1 c|λ|  |λ|M  1 c|λ|  +
c 2
.
α0 β α + αβ + α + αβ + c
β + 2α α + αβ + β

Then X x (t) is exponentially ultimately bounded in the m.s.s.

Proof Let
T0 t  2
Ψ0 (x) = E X0x (t)V ds dt
0 0

with T0 = α0 (c|λ|/(αβ) + 1/α) + c/β. Then Ψ0 (s) satisfies conditions (1)–(5) of

Theorem 6.10, and for all x ∈ H ,

c1 x2H − k1 ≤ Ψ0 (x) ≤ c2 x2H + k2 .

It remains to prove that, for all x ∈ V ,

L Ψ0 (x) ≤ −c3 Ψ0 (s) + k3 .

7.2 Exponential Ultimate Boundedness in Variational Method 243

Then we can conclude the result by Theorem 7.5. Now, for x ∈ V ,

L Ψ0 (x) − L0 Ψ0 (x)
  1   
= Ψ0 (x), A(x) − A0 x + tr Ψ0 (x) B(x)QB ∗ (x) − B0 xQ(B0 x)∗
2
   1   
= Ψ0 (x), A(x) − A0 x H + tr Ψ0 (x) B(x)QB ∗ (x) − B0 xQ(B0 x)∗ .
2
But Ψ0 (x) = 2Cx and Ψ0 (x) = 2C for x ∈ V , where C is defined in (7.11). By
inequality (7.10),
 
1 c|λ| |λ|M + γ 2
CL (H ) ≤ + T0 + T0 .
α αβ 2α
Hence,
    
L Ψ0 (x)−L0 Ψ0 (x) ≤ 2 Cx, A(x)−A0 x H +τ C B(x)QB ∗ (x)−B0 xQ(B0 x)∗ ,

and we have
*
L Ψ0 (x) ≤ L0 Ψ0 (x) + CL (H ) 2xH A(x) − A0 xH
 +
+ τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗ .

When T = T0 , from (7.9) we have

c |λ|M + γ
L0 Ψ0 (x) ≤ − x2H + T0 ,
α0 β α
giving
c |λ|M + γ  
L Ψ0 (x) ≤ − x2H + T0 + CL (H ) ω̃x2H + k
α0 β α
 
c |λ|M + γ
≤ − + ω̃CL (H ) x2H + T0 + kCL (H ) .
α0 β α

Now, −c/(α0 β) + ω̃CL (H ) < 0 if ω̃ satisfies our original assumption, and we

arrive at L Ψ0 (x) ≤ −c3 Ψ0 (x) + k3 with c3 > 0. 

Remark 7.1 Note that the function Ψ0 (x) in Theorem 7.8 is the Lyapunov function
for the nonlinear equation.

Corollary 7.2 Suppose that the coefficients of the linear equation (6.42) satisfy
the coercivity condition (6.39), and its solution {X0x (t), t ≥ 0} is exponentially ul-
timately bounded in the m.s.s. Let {X x (t), t ≥ 0} be a solution of the nonlinear
equation (6.37). Furthermore, suppose that

A(v) − A0 v ∈ H for all v ∈ V

244 7 Ultimate Boundedness and Invariant Measure

and that, for v ∈ V ,

     
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ k 1 + v2H (7.14)

for some k > 0. If for v ∈ V , as vH → ∞,

   
A(v) − A0 v  = o vH
H
and     (7.15)
τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ = o v2H ,

then X x (t) is exponentially ultimately bounded in the m.s.s.

Proof Under assumption (7.15), for a constant ω̃ satisfying the condition of Theo-
rem 7.8, there exists an R > 0 such that, for all v ∈ V with vH > R,
   
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ ω̃v2H .

For v ∈ V and vH < R, by (7.14),

   
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗
   
≤ k 1 + v2H ≤ k 1 + R 2 .

Hence, we have
   
2vH A(v) − A0 v H + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗

≤ ω̃v2H + (k + 1)R 2 .

An appeal to Theorem 7.8 completes the proof. 

Theorem 7.9 Suppose that the coefficients of the linear equation (6.42) satisfy
the coercivity condition (6.39) and its solution {X0x (t), t ≥ 0} is exponentially ul-
timately stable in the m.s.s. with the function t → EX0x (t)2V being continuous for
all x ∈ V . Let {X x (t), t ≥ 0} be a solution of the nonlinear equation (6.37). If for
v∈V,
   
2vV A(v) − A0 v V ∗ + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ ω̃0 v2V + k0

for some constants ω̃0 , k0 such that

c
ω̃ < * 1 c|λ|  1 c|λ|  |λ|M  1 |λ|  +
c 2
,
(α0 + 1)β α + αβ + α + αβ + c
β + 2α α + αβ + β

then X x (t) is exponentially ultimately bounded in the m.s.s.

Proof Let, as before,

T0 t  2
Ψ0 (x) = E X0x (t)V ds dt
0 0
7.2 Exponential Ultimate Boundedness in Variational Method 245

with T0 = α0 (c|λ|/(αβ) + 1/α) + βc . For x ∈ V ,

L Ψ0 (x) − L0 Ψ0 (x)
  1   
= Ψ0 (x), A(x) − A0 x + tr Ψ0 (x) B(x)QB ∗ (x) − B0 xQ(B0 x)∗
2

with Ψ0 (x) = 2C̃x and Ψ0 (x) = 2C, where the operators C and C̃ are defined
in (7.11) and (7.12). By inequality (7.10) and the continuity of the embedding
V → H ,
 
1 c|λ| |λ|M + γ 2
CL (H ) ≤ + T0 + T0 ,
α αβ 2α
C̃L (V ) ≤ α0 CL (V ) .

Hence,
  
L Ψ0 (x) − L0 Ψ0 (x) ≤ 2C̃x, Ax − A0 x H + tr C B(x)QB ∗ (x) − B0 xQ(B0 x)∗ ,

and we have

L Ψ0 (x) ≤ L0 Ψ0 (x) + 2C̃L (V ) xV Ax − A0 xV ∗

 
+ tr CB(x)QB ∗ (x) − B0 xQ(B0 x)∗
 
≤ L0 Ψ0 (x) + CL (H ) + C̃L (V ) 2xV Ax − A0 xV ∗
 
+ τ B(x)QB ∗ (x) − B0 xQ(B0 x)∗ .

Since s → EX0x (s)2V is a continuous function, we obtain from earlier relations for
L0 Ψ0 (x) that
c |λ|M + γ
L0 Ψ0 (x) ≤ − x2V + T0 .
β α
Hence,

c |λ|M   
L Ψ0 (x) ≤ − x2V + T0 + CL (H ) + C̃L (V ) ω̃0 x2V + k0
β α
 
c  
≤ − + ω̃0 CL (H ) + C̃L (V ) x2V
β
 
|λ|M + γ
+ k0 CL (H ) + C̃L (V ) + T0 .
α

Since, with the condition on ω̃0 , −c/β + ω̃0 (CL (H ) + C̃L (V ) ) < 0, we see
that conditions analogous to those of Theorem 7.1 are satisfied by Ψ0 , giving the
result. 
246 7 Ultimate Boundedness and Invariant Measure

Corollary 7.3 Suppose that the coefficients of the linear equation (6.42) satisfy the
coercivity condition (6.39) and its solution {X0x (t), t ≥ 0} is exponentially ultimately
bounded in the m.s.s. with the function t → EX0x (t)2V being continuous for all
x ∈ V . Let {X x (t), t ≥ 0} be a solution of the nonlinear equation (6.37). If for
v ∈ V , as vV → ∞,
   
A(v) − A0 v  ∗ = o vV
V
and (7.16)
   
τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ = o v2V ,

then X x (t) is exponentially ultimately bounded in the m.s.s.

Proof We shall use Theorem 7.9. Under assumption (7.16), for a constant ω̃0 satis-
fying the condition of Theorem 7.9, there exists an R > 0 such that, for all v ∈ V
with vV > R,
   
2vV A(v) − A0 v V ∗ + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗ ≤ ω̃0 v2V .

Using that A(v)V ∗ , A0 (v)V ∗ ≤ a1 vV and B(v)L (K,H ) , B0 vL (K,H ) ≤
b1 vV , we have, for v ∈ V such that vH < R,
   
2vV A(v) − A0 v V ∗ + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗
 2 
≤ 4a1 v2V + B(v)L (K,H ) + B0 v2L (K,H ) tr(Q)
 
≤ 4a1 + 2b12 tr(Q) v2V
 
≤ 4a1 + 2b12 tr(Q) R 2 .

Hence, for v ∈ V ,
   
2vV A(v) − A0 v V ∗ + τ B(v)QB ∗ (v) − B0 vQ(B0 v)∗
 
≤ ω̃0 v2V + 4a1 + 2b12 tr(Q) R 2 .

An appeal to Theorem 7.9 completes the proof. 

7.3 Abstract Cauchy Problem, Stability and Exponential

Ultimate Boundedness
We present an analogue of a result of Zakai and Miyahara for the infinite-
dimensional case.

Definition 7.3 A linear operator A : V → V ∗ is called coercive if it satisfies the

following coercivity condition: for some α > 0, γ , λ ∈ R, and all v ∈ V ,

2v, Av ≤ λv2H − αv2V + γ . (7.17)

7.3 Abstract Cauchy Problem, Stability and Exponential Ultimate Boundedness 247

Proposition 7.1 Consider a stochastic evolution equation,


dX(t) = A0 X(t) dt + F (X(t)) dt + B(X(t)) dWt ,
(7.18)
X(0) = x ∈ H,

with the coefficients A0 and F satisfying the following conditions:

(1) A0 : V → V ∗ is coercive.
(2) F : V → H , B : V → L (K, H ), and there exists a constant K > 0 such that
for all v ∈ V ,
     
F (v)2 + B(v)2 ≤ K 1 + v2H .
H L (K,H )

(3) There exists a constant L > 0 such that for all v, v  ∈ V ,

     
F (v) − F (v  )2 + tr B(v) − B(v  ) Q B ∗ (v) − B ∗ (v  ) ≤ Lv − v  2 .
H H

(4) For v ∈ V , as vH → ∞,

       
F (v) = o vH , B(v) = o vH .
H L (K,H )

If the classical solution {ux (t), t ≥ 0} of the abstract Cauchy problem

⎧
⎨ du(t) = A u(t),
0
dt (7.19)
⎩u(0) = x ∈ H,

is exponentially stable (or even exponentially ultimately bounded), then the solution
of (7.18) is exponentially ultimately bounded in the m.s.s.

Proof Let A(v) = A0 v + F (v) for v ∈ V . Since F (v) ∈ H ,

   
2 v, A(v) + tr B(v)QB ∗ (v)
   
= 2v, A0 v + 2 v, F (v) + tr B(v)QB ∗ (v)
   2
≤ λv2H − αv2V + 2vH F (v)H + B(v)L (K,H ) tr(Q)

≤ λ v2H − αv2H + γ

for some constants λ and γ . Hence, the evolution equation (7.18) satisfies the coer-
civity condition (6.39). Under assumption (2)
       
F (v)2 + tr B(v)QB ∗ (v) ≤ F (v)2 + tr(Q)B(v)2
H H L (K,H )
   
≤ 1 + tr(Q) K 1 + vH ,
2
248 7 Ultimate Boundedness and Invariant Measure

so that condition (7.14) holds, and since

       
F (v) = o vH and τ B(v)QB ∗ (v) = o v2H as vH → ∞,
H

Corollary 7.2 gives the result. 

Example 7.2 (Stochastic Heat Equation) Let S 1 be the unit circle realized as the
interval [−π, π] with identified points −π and π . Denote by W 1,2 (S 1 ) the Sobolev
space on S 1 and by W (t, ξ ) the Brownian sheet on [0, ∞)×S 1 , see Exercise 7.2. Let
κ > 0 be a constant, and f and b be real-valued functions. Consider the following
SPDE:
⎧
⎨ ∂X(t) ∂ 2 X(t) ∂ 2W
(ξ ) = (ξ ) − κf (X(t)(ξ )) + b(X(t)(ξ )) ,
∂t ∂ξ 2 ∂t∂ξ (7.20)
⎩
X(0)(·) = x(·) ∈ L2 (S 1 ).

Let H = L2 (S 1 ) and V = W 1,2 (S 1 ). Consider

 2 
d
A0 (x) = − κ x
dξ 2

and mappings F, B defined for ξ ∈ S 1 and x, y ∈ V by

       
F (x)(ξ ) = f x(ξ ) , B(x)y (ξ ) = b x(·) , y(·) L2 (S 1 ) .

Let
 1/2
xH = 2
x (ξ ) dξ for x ∈ H,
S1
     1/2
dx(ξ ) 2
xV = x (ξ ) +
2
dξ for x ∈ V .
S1 dξ
Then we obtain the equation
   
dX(t) = A0 X(t) dt + F X(t) dt + B X(t) d W̃t ,

where W̃t is a cylindrical Wiener process defined in Exercise 7.2. We have

 
2 x, A0 (x) = −2x2V + (−2κ + 2)x2H
≤ −2x2H + (−2κ + 2)x2H = −2κx2H .

By Theorem 6.3(a), with Λ(x) = x2H , the solution of (7.19) is exponentially sta-
ble. If we assume that f and b are Lipschitz continuous and bounded, then con-
ditions (1)–(3) of Proposition 7.1 are satisfied. Using representation (2.35) of the
stochastic integral with respect to a cylindrical Wiener process, we can conclude
that the solution of the stochastic heat equation (7.20) is exponentially ultimately
bounded in the m.s.s.
7.3 Abstract Cauchy Problem, Stability and Exponential Ultimate Boundedness 249

Exercise 7.2 Let S 1 be the unit circle realized as the interval [−π, π] with identified
points −π and π . Denote by {fj (ξ )} an ONB in L2 (S 1 ) and consider
∞
 ζ
W (t, ζ ) = wj (t) fj (ξ ) dξ, t ≥ 0, −π ≤ ζ ≤ π, (7.21)
j =1 −π

where wj are independent Brownian motions defined on {, F , {Ft }t≥0 }, P }.

Show that the series (7.21) converges P -a.s. and that
 
Cov W (t1 , ζ1 )W (t2 , ζ2 ) = (t1 ∧ t2 )(ζ1 ∧ ζ2 ).

Conclude that the Gaussian random field W (·, ·) has a continuous version. This
continuous version is called the Brownian sheet on S 1 .
Now, let Φ(t) be an adapted process with values in L2 (S 1 ) (identified with
L (L2 (S 1 ), R)) and satisfying
∞ 
E Φ(t)2 2 dt < ∞.
L (S 1 )
0

Consider a standard cylindrical Brownian motion W̃t in L2 (S 1 ) defined by

∞

W̃t (k) = wj (t)k, fj L2 (S 1 ) .
j =1

Show that the cylindrical stochastic integral process

t
Φ(s) d W̃s (7.22)
0

is well defined in L2 (, R).

On the other hand, for an elementary processes of the form

Φ(t, ξ ) = 1[0,t] (s)1[−π,ζ ] (ξ ), (7.23)

define
∞
Φ ·W = Φ(s, ξ )W (ds, dξ ). (7.24)
0 S1
Clearly Φ · W = W (t, ζ ). Extend the integral Φ · W to general processes. Since
∞
Φ ·W = Φ(s) d W̃s
0

for elementary processes (7.23), conclude that the integrals are equal for general
processes as well.
250 7 Ultimate Boundedness and Invariant Measure

Example 7.3 Consider the following SPDE driven by a real-valued Brownian mo-
tion:
⎧  
⎪
⎪ ∂ 2 u(t, x) ∂u(t, x)
⎪
⎪ dt u(t, x) = α 2 +β + γ u(t, x) + g(x) dt
⎪
⎪ ∂x 2 ∂x
⎪
⎨  
∂u(t, x) (7.25)
⎪
⎪ + σ 1 + σ 2 u(t, x) dWt ,
⎪
⎪ ∂x
⎪
⎪
⎩u(0, x) = ϕ(x) ∈ L2 (−∞, ∞) ∩ L1 (−∞, +∞),
⎪

where we use the symbol dt to signify that the differential is with respect to t. Let
H = L2 ((−∞, ∞)) and V = W01,2 ((−∞, ∞)) with the usual norms
 +∞ 1/2
vH = v 2 dx , v ∈ H,
−∞
 +∞   2  1/2
dv
vV = v2 + dx , v ∈ V.
−∞ dx

Define the operators A : V → V ∗ and B : V → L (H ) by

d2 v dv
A(v) = α 2 2
+β + γ v + g, v ∈ V,
dx dx
dv
B(v) = σ1 + σ2 v, v ∈ V.
dx
Suppose that g ∈ L2 ((−∞, ∞)) ∩ L1 ((−∞, ∞)). Then, using integration by parts,
we obtain for v ∈ V ,
   
2 v, A(v) + tr Bv(Bv)∗
   2
2
2d v 2 dv
 dv 
= 2 v, α +β 
+ γ v + g + σ 1 + σ2 v 
dx 2 ∂x dx 
H
   
= −2α 2 + σ12 v2V + 2γ + σ22 + 2α 2 − σ12 v2H + 2v, g H
    1
≤ −2α 2 + σ12 v2V + 2γ + σ22 + 2α 2 − σ12 + ε v2H + g2H
ε
for any ε > 0. Similarly, for u, v ∈ V ,
    ∗ 
2 u − v, A(u) − A(v) + tr B(u − v) B(u − v)
   
≤ −2α 2 + σ12 u − v2V + 2γ + σ22 + 2α 2 − σ12 u − v2H .

If −2α 2 + σ12 < 0, then the coercivity and weak monotonicity conditions, (6.39)
and (6.40), hold, and we know from Theorem 4.7 that there exists a unique strong
7.3 Abstract Cauchy Problem, Stability and Exponential Ultimate Boundedness 251

solution uϕ (t) to (7.25) in L2 (Ω, C([0, T ], H ))∩M 2 ([0, T ], V ). Taking the Fourier
transform yields
 
dt ûϕ (t, λ) = −α 2 λ2 ûϕ (t, λ) + iλβ ûϕ (t, λ) + γ ûϕ (t, λ) + ĝ(λ) dt
 
+ iσ1 λûϕ (t, λ) + σ2 ûϕ (t, λ) dWt
  
= −α 2 λ2 + iλβ + γ ûϕ (t, λ) + ĝ(λ) dt

+ (iσ1 λ + σ2 )ûϕ (t, λ) dWt .

For fixed λ,

a = −α 2 λ2 + iλβ + γ ,
b = ĝ(λ),
c = iσ1 λ + σ2 .

By simple calculation (see Exercise 7.3),

 
 ϕ 2  2 bb + bϕ̂(λ)(a + a + cc) (a+a+cc)
   
E û (t, λ) = E ϕ̂(λ) + 2Re e
(a + a + cc)(a + cc)
   
b(a ϕ̂(λ) + b) at βb
− 2Re e + 2Re . (7.26)
a(a + cc) a(a + a + cc)

By Plancherel’s theorem,

 2 +∞  2
E uϕ (t)H = E ûϕ (t, λ) dλ
−∞

and
 2
 2  2 d ϕ 
E uϕ (t)V = E uϕ (t)H + E 
 dx u (t, x)

H
+∞    2
= 1 + λ2 E ûϕ (t, λ) dλ.
−∞

For a suitable T > 0,

T t  2
Ψ (ϕ) = E uϕ (s)V ds dt
0 0
+∞   T t  2
= 1 + λ2 E û(s, λ) ds dt dλ.
−∞ 0 0
252 7 Ultimate Boundedness and Invariant Measure

Thus it is difficult to compute a Lyapunov function explicitly. In view of Remark 7.1,

it is enough to compute a Lyapunov function of the linear SPDE
 
∂ 2 u(t, x) ∂u(t, x)
dt u(t, x) = α 2 + β + γ u(t, x) du
∂x 2 ∂x
 
∂u(t, x)
+ σ1 + σ2 u(t, x) dWt .
∂x

Define the operators A0 : V → V ∗ and B0 : V → L (H ) by

d2 v dv
A0 (v) = α 2 2
+β + γ v, v ∈ V,
dx dx
B0 (v) = B(v), v ∈ V

(since B is already linear). Taking the Fourier transform and solving explicitly, we
obtain that the solution is the geometric Brownian motion

ϕ 1 2
û0 (t, λ) = ϕ̂(λ)eat− 2 c t+cWt ,
 ϕ 2  2
E û0 (t, λ) = ϕ̂(λ) e(a+a+cc)t .
ϕ
The function t → Eu0 (t)2V is continuous for all ϕ ∈ V ,
   
A(v) − A0 (v) ∗ = gV ∗ = o vV as vV → ∞,
V

and
 
τ B(v)QB ∗ (v) − (B0 v)Q(B0 v)∗ ) = 0.
Thus, if {u0 (t), t ≥ 0} is exponentially ultimately bounded in the m.s.s., then the
Lyapunov function Ψ0 (ϕ) of the linear system is the Lyapunov function of the non-
linear system, and
 
+∞   T t  2
Ψ0 (ϕ) = 1 + λ2 E û0 (s, λ) ds dt dλ
−∞ 0 0
+∞  

2 exp{(−2α 2 + σ12 )λ2 + 2γ + σ22 )T }

= 1 + λ2 ϕ̂(λ)
−∞ ((−2α 2 + σ12 )λ2 + 2γ + σ22 )2

T 1
− − dλ.
(−2α 2 + σ12 )λ2 + 2γ + σ22 (−2α 2 + σ12 )λ2 + 2γ + σ22

Using Theorem 7.8, we can conclude that the solution of the nonlinear system is
exponentially ultimately bounded in the m.s.s.

Exercise 7.3 Complete the computations in (7.26).

7.3 Abstract Cauchy Problem, Stability and Exponential Ultimate Boundedness 253

Example 7.4 Consider an equation of the form


dXt = AX(t) dt + F (X(t)) dt + B(X(t)) dWt ,
X(0) = x ∈ H,

where F and B satisfy the conditions of Proposition 7.1. This example is moti-
vated by the work of Funaki. If −A is coercive, a typical case being A = , we
conclude that the solution of the deterministic linear equation is exponentially sta-
ble since the Laplacian has negative eigenvalues. Thus, the solution of the deter-
ministic equation is exponentially bounded, and hence, by Proposition 7.1, the so-
lution of the nonlinear equation above is exponentially ultimately bounded in the
m.s.s.

Example 7.5 Let O ⊆ Rn be a bounded open domain with smooth boundary.

Assume that H = L2 (O) and V = W01,2 (O), the Sobolev space. Suppose that
{Wq (t, x); t ≥ 0, x ∈ 0} is an H -valued Wiener process with associated covari-
ance operator Q, given by a continuous symmetric nonnegative definite kernel
q(x, y) ∈ L2 (O × O), q(x, x) ∈ L2 (O),

(Qf )(x) = q(x, y)f (y) dy.

O

By Mercer’s theorem [41], there exists an orthonormal basis {ej }∞

j =1 ⊂ L (O) con-
2

sisting of eigenfunctions of Q such that

∞

q(x, y) = λj ej (x)ej (y)
j =1

 
with tr(Q) = O q(x, x) dx = ∞ j =1 λj < ∞.
Let −A be a linear strongly elliptic differential operator of second order on O,
and B(u) : L2 (O) → L2 (O) with B(u)f (·) = u(·)f (·). By Garding’s inequality,
−A is coercive (see [63], Theorem 7.2.2). Then the infinite-dimensional problem is
as follows:

dt u(t, x) = Au(t, x) dt + u(t, x) dt Wq (t, x),

and we choose Λ(v) = v2H for v ∈ W01,2 (O). We shall check conditions under
which Λ is a Lyapunov function. With L defined in (6.15), using the spectral rep-
resentation of q(x, y), we have
   
L v2H = 2v, Av + tr B(v)QB ∗ (v)

= 2v, Av + q(x, x)v 2 (x) dx.

O
254 7 Ultimate Boundedness and Invariant Measure

Let
 
L (v2H )
λ0 = sup ,v ∈ W01,2 (O), v2H =0
v2H
 
2v, Av + Qv, v H
= sup , v ∈ W01,2 (O), v2H = 0 .
v2H

If λ0 < 0, then, by Theorem 6.4, the solution is exponentially stable in the m.s.s.
Consider the nonlinear equation in O,

dt u(t, x) = Ã(x, u(t, x)) dt + B̃(u(t, x))dt Wq (t, x),
(7.27)
u(0, x) = ϕ(x), u(t, x)|∂O = 0.

Assume that

Ã(x, v) = Av + α1 (x, v), B̃(x, v) = B(v) + α2 (x, v),

where αi (x, v) satisfy the Lipschitz-type condition

 
sup αi (x, v1 ) − αi (x, v2 ) < cv1 − v2 H ,
x∈O

so that the nonlinear equation (7.27) has a unique strong solution. Under the as-
sumption
αi (x, 0) = 0,
zero is a solution of (7.27), and if
   
sup αi (x, v) = o vH , vH → 0,
x∈O

then, by Theorem 6.14, the strong solution of the nonlinear equation (7.27) is expo-
nentially stable in the m.s.s.
On the other hand, let us consider the operator A as above and F and B satisfying
the conditions of Proposition 7.1. Then, under the condition
 
2v, Av 1,2
sup , u ∈ W 0 (O), v 2
H = 0 < 0,
v2H

the solution of the abstract Cauchy problem (7.19), with A0 replaced by A, is expo-
nentially stable, and we conclude that the solution of the equation

dX(t) = AX(t) dt + F (X(t)) dt + B(X(t)) dWt ,
X(0) = x ∈ H,

is ultimately exponentially bounded in the m.s.s.

7.4 Ultimate Boundedness and Invariant Measure 255

Consider now the SSDE (3.1) and assume that A is the infinitesimal generator of
a pseudo-contraction C0 -semigroup {S(t), t ≥ 0} on H (see Chap. 3) with the co-
efficients F : H → H and B : H → L (K, H ), independent of t and ω. We assume
that F and B are in general nonlinear mappings satisfying the linear growth con-
dition (A3) and the Lipschitz condition (A4) (see Sect. 3.3). In addition, the initial
condition is assumed deterministic, so that (3.1) takes the form

dX(t) = (AX(t) + F (X(t))) dt + B(X(t)) dWt ,
(7.28)
X(0) = x ∈ H.

By Theorem 3.5, there exists a unique continuous mild solution.

Using Corollary 7.1, we now have the following analogue of Proposition 7.1.

Proposition 7.2 Suppose that the classical solution {ux (t), t ≥ 0} of the abstract
Cauchy problem (7.19) is exponentially stable (or even exponentially ultimately
bounded) and, as hH → ∞,
   
F (h) = o hH ,
H
   
B(h) = o hH ,
L (K,H )

then the mild solution of (7.28) is exponentially ultimately bounded in the m.s.s.

7.4 Ultimate Boundedness and Invariant Measure

We are interested in the behavior of the law of a solution to an SDE as t →

∞. Let us begin with a filtered probability space (Ω, F , {Ft }t≥0 , P ) and an
H -valued time-homogeneous Markov process X ξ0 (t), X ξ0 (0) = ξ0 , where ξ0 is
F0 -measurable random variable with distribution μξ0 . Assume that its associated
semigroup Pt is Feller. We can define for A ∈ B(H ), the Markov transition proba-
bilities
P (t, x, A) = Pt 1A (x), x ∈ H.
Since a regular conditional distribution of X ξ0 (t) exists (Theorem 3, Vol. I, Sect. 1.3
in [25]), we have that
   
P (t, x, A) = P X ξ0 (t) ∈ A|ξ0 = x = P X ξ0 ∈ A|ξ0 = x μξ0 (dx), x ∈ H.
H

Then, for a bounded measurable function f on H (f ∈ Bb (H )),

(Pt f )(x) = f (y)P (t, x, dy). (7.29)

H
256 7 Ultimate Boundedness and Invariant Measure

The Markov property (3.52) takes the form

   ξ0   ξ  
E f X ξ0 (t + s) F Xt = (Ps f ) Xt 0 = f (y)P s, X ξ0 (t), dy ,
H

so that the transition probability P (t, x, A) is a transition function for a time-

homogeneous Markov process X ξ0 (t).
We observe that the following Chapman–Kolmogorov equation holds for Markov
transition probabilities

P (t + s, x, A) = P (t, y, A)P (s, x, dy), (7.30)

H

which follows from the semigroup property of Pt , (3.58) applied to ϕ(x) = 1A (x)
and from the fact that P (t, x, dy) is the conditional law of X ξ0 (t).

Exercise 7.4 Show (7.30).

Let us now define an invariant probability measure and state a general theorem
on its existence.

Definition 7.4 We say that a probability measure μ on H is invariant for a time-

homogeneous Markov process X x (t) with the related Feller semigroup {Pt , t ≥ 0}
defined by (7.29) if for all A ∈ B(H ),

μ(A) = P (t, x, A)μ(dx),

H

or equivalently, since H is a Polish space, if for all f ∈ Cb (H ),

(Pt f ) dμ = f (y) dμ.

H H

Let μ be a probability measure on H and define

1 tn
μn (A) = P (t, x, A) dt μ(dx) (7.31)
tn 0 H

for a sequence {tn }∞

n=1 ⊂ R+ , tn → ∞. In particular, for a real-valued bounded
Borel-measurable function f (x) on H , we have

1 tn
f (x) μn (dx) = f (y)P (t, x, dy) μ(dx) dt. (7.32)
H tn 0 H H

Theorem 7.10 If ν is weak limit of a subsequence of {μn }, then ν is an invariant

measure.
7.4 Ultimate Boundedness and Invariant Measure 257

Proof We can assume without loss of generality that μn ⇒ ν. Observe that, by the
Fubini theorem and the Chapman–Kolmogorov equation,

(Pt f )(x)ν(dx) = lim (Pt f )(x) μn (dx)

H n→∞ H

1 tn
= lim (Pt f )(y)P (s, x, dy) μ(dx) ds
n→∞ tn 0 H H
1 tn
= lim (Pt+s f )(x) μ(dx) ds
n→∞ tn 0 H
,  tn
1
= lim (Ps f )(x) μ(dx) ds
n→∞ tn 0 H
tn +t t -
+ (Ps f )(x) μ(dx) ds − (Ps f )(x) μ(dx) ds .
tn H 0 H

Since Ps f (x0 )H ≤ f (x0 H , the last two integrals are bounded by a constant,
and hence, using (7.32),
1 tn
(Pt f )(x) ν(dx) = lim (Ps f )(x0 ) μ(dx) ds
H n→∞ tn 0 H
1 tn
= lim f (y)P (s, x, dy) μ(dx) ds
n→∞ tn 0 H H

= lim f (x) μn (dx) = f (x) ν(dx). 

n→∞ H H

Corollary 7.4 If the sequence {μn } is relatively compact, then an invariant measure
exists.

Exercise 7.5 Show that if, as t → ∞, the laws of X x (t) converge weakly to a
probability measure μ, then μ is an invariant measure for the corresponding semi-
group Pt .

We shall now consider applications of the general results on invariant measures

to SPDEs. In case where {X ξ0 (t), t ≥ 0} is a solution of an SDE with a random
initial condition ξ0 , taking in (7.31) μ = μξ0 , the distribution of ξ0 , gives
 
P X ξ0 (t) ∈ A = P (t, x, A) μξ0 (dx). (7.33)
H

Thus, properties of the solution can be used to obtain tightness of the measures μn .

Exercise 7.6 Prove (7.33).

Before we apply the result on ultimate boundedness to obtain the existence of an

invariant measure, let us consider some examples.
258 7 Ultimate Boundedness and Invariant Measure

Example 7.6 (Navier–Stokes Equation [76]) Let D ⊆ R2 be a bounded domain with

smooth boundary ∂D . Consider the equation
⎧
∂vi (t, x)  ∂vi (t, x) 
2 2
⎪
⎪ 1 ∂P (t, x) ∂ 2 vi (t, x)
⎪
⎪ + v = − + ν + σi Ẇti (x),
⎪
⎪ ∂t
j
∂xj ρ ∂xi 2
⎪
⎨ j =1 j =1
∂x j
 2
∂v (7.34)
⎪
⎪
j
= 0, x ∈ D, ν > 0,
⎪
⎪
⎪
⎪ j =1
∂xj
⎪
⎩
i = 1, 2.

Let C0∞ = {v ∈ C0∞ (D) × C0∞ (D); ∇v = 0}, with ∇ denoting the gradient. Let
H = C0∞ in L2 (D) × L2 (D), and V = {v : W01,2 (D) × W01,2 (D), ∇v = 0}. Then
V ⊆ H ⊆ V ∗ is a Gelfand triplet, and the embedding V → H is compact.
It is known [76] that

L2 (D) × L2 (D) = H ⊕ H ⊥ ,

where H ⊥ is characterized by H ⊥ = {v : v = ∇(p) for some p ∈ W 1,2 (D)}.

Denote by  the orthogonal projection of L2 (D) × L2 (D) onto H ⊥ , and for
v ∈ C0∞ , define
* +
A(v) = v v −  (v · ∇)v .
Then A can be extended as a continuous operator form V to V ∗ .
Equation (7.34) can be recast as an evolution equation in the form

dX(t) = A(X(t)) dt + σ dWt ,
X(0) = ξ,

where Wt is an H -valued Q-Wiener process, and ξ ∈ V a.e. is an F0 -measurable

H -valued random variable. It is known (see [76]) that the above equation has a
unique strong solution {uξ (t), t ≥ 0} in C([0, T ], H ) ∩ L2 ([0, T ], V ), which is a
homogeneous Markov and Feller process, satisfying for T < ∞,
2 


 2 T  ∂uξ (t) 2 T
E uξ (T )H + ν  
 ∂x  dt ≤ Eξ H + 2 tr(Q).
2
0 i H
i=1

Using the fact that uξ (t)V is equivalent to ( 2i=1  ∂u(ξ ) 2 1/2
∂xi H ) , we have

1 T  2  c
sup E uξ (t)V dt ≤ tr(Q)
T T 0 2ν

with some constant c. By the Chebychev inequality,

1 T   
lim sup P uξ (t)V > R dt = 0.
R→∞ T T 0
7.4 Ultimate Boundedness and Invariant Measure 259

Hence, for ε > 0, there exists an Rε such that

1 T   
sup P uξ (t)V > Rε dt < ε.
T T 0

Thus, as tn → ∞,

1 tn  
sup )Rε dt μξ (dx) < ε,
P t, x, B
n tn H 0

where B)Rε is the image of the set {v ∈ V ; vV > Rε } under the compact embed-
ding V → H , and μξ is the distribution of ξ on H . Since B )Rε is a complement of a
compact set, we can use Prokhorov’s theorem and Corollary 7.4 to conclude that an
invariant measure exists. Note that its support is in V , by the weak convergence.

Example 7.7 (Linear equations with additive noise [79]) Consider the mild solution
of the equation

dX(t) = AX(t) dt + dWt ,
X(0) = x ∈ H,

where A is an infinitesimal generator of a strongly continuous semigroup {S(t),

t ≥ 0} on H . Denote
t
Qt = S(r)QS ∗ (r) dr,
0

and assume that tr(Qt ) < ∞. We know from Theorems 3.1 and 3.2 that
t
X(t) = S(t)x + S(t − s) dWs (7.35)
0
t
is the mild solution of the above equation. The stochastic convolution 0 S(t −
s) dWs is an H -valued Gaussian process with covariance
t
Qt = S(u)QS ∗ (u) du
0

for any t. The Gaussian process X(t) is also Markov and Feller, and it is called an
Ornstein–Uhlenbeck process. The probability measure μ on H is invariant if for
f ∈ Cb (H ) and any t ≥ 0,

  
f (x) μ(dx) = E f X x (t) μ(dx)
H H
 t 
= Ef S(t)x + S(t − s) dWs μ(dx).
H 0
260 7 Ultimate Boundedness and Invariant Measure

For f (x) = eiλ,x H , λ ∈ H , we obtain

  1
μ̂(λ) = μ̂ S ∗ (t)λ e− 2 Qt λ,λ H ,
where μ̂ denotes the characteristic function of μ. It follows that
 
μ̂(λ) ≤ e− 12 Qt ,λ,λ H ,

or
 
  1
Qt λ, λ H ≤ −2 lnμ̂(λ) = 2 ln .
|μ̂(λ)|
Since μ̂(λ) is the characteristic function of a measure μ on H , then by Sazonov’s
theorem [74], for ε > 0, there exists a trace-class operator S0 on H such that
|μ̂(λ)| ≥ 1/2 whenever S0 λ, λ H ≤ 1. Thus, we conclude that
Qt λ, λ H ≤ 2 ln 2
if S0 λ, λ H ≤ 1. This yields
0 ≤ Qt ≤ (2 ln 2)S0 .
Hence, supt tr(Qt ) < ∞.
On the other hand, if supt tr(Qt ) < ∞, let us denote by P the limit in trace norm
of Qt and observe that
∞ ∞
S(t)P S ∗ (t) = S(t + r)QS ∗ (t + r) dr = S(u)QS(u) du = P − Qt .
0 t

Thus,
1  1  1
S(t)P S ∗ (t)λ, λ H = P λ, λ H − Qt λ, λ H,
2 2 2
implying
1 1 ∗ (t)λ,S ∗ (t)λ 1
e− 2 P λ,λ H = e− 2 P S H e− 2 Qt λ,λ H .
1
In conclusion, μ with the characteristic functional e− 2 P λ,λ H is an invariant mea-
sure. We observe that the invariant measure exists for the Markov process X(t)
defined in (7.35) if and only if supt tr(Qt ) < ∞. Also, if S(t) is an exponentially
stable semigroup (i.e., S(t)L (H ) ≤ Me−μt for some positive constants M and μ)
or if St x → 0 for all x ∈ H as t → ∞, then the Gaussian measure with covariance
P is the invariant (Maxwell) probability measure.

Let {X(t), t ≥ 0} be exponentially ultimately bounded in the m.s.s., then, clearly,

 2
lim sup E X(t)H ≤ M < ∞ for all x ∈ H . (7.36)
t→∞

Definition 7.5 A stochastic process X(t) satisfying condition (7.36) is called ulti-
mately bounded in the m.s.s.
7.4 Ultimate Boundedness and Invariant Measure 261

7.4.1 Variational Equations

We focus our attention now on the variational equation with a deterministic initial
condition,

dX(t) = A(X(t)) dt + B(X(t)) dWt ,
(7.37)
X(0) = x ∈ H,
which is driven by a Q-Wiener process Wt . The coefficients A : V → V ∗ and B :
V → L (K, H ) are independent of t and ω, and they satisfy the linear growth,
coercivity (C), and weak monotonicity (WM) conditions (6.38), (6.39), (6.40). By
Theorem 4.8 and Remark 4.2 the solution is a homogeneous Markov process, and
the associated semigroup is Feller.
We note that in Theorem 7.5, we give conditions for exponential ultimate bound-
edness in the m.s.s. in terms of the Lyapunov function. Assume that Ψ : H → R
satisfies the conditions of Theorem 6.10 (Itô’s formula) and define
   
L ψ(u) = ψ  (u), A(u) + (1/2) tr ψ  (u)B(u)QB ∗ (u) . (7.38)
Let {X x (t), t ≥ 0} be the solution of (7.37). We apply Itô’s formula to Ψ (X x (t)),
take the expectation, and use condition (2) of Theorem 7.5 to obtain
    t  
EΨ X x (t) − EΨ X x (t  ) = E L Ψ X x (s) ds
t
t   
≤ −c3 EΨ X t (s) + k3 ds.
t

Let Φ(t) = EΨ (X x (t)), then Φ(t) is continuous, so that

Φ  (t) ≤ −c3 Φ(t) + k3 .
Hence,
 
 x  k3 k3 −c3 t
EΨ Xt ≤ + Ψ (x) − e .
c3 c3
Assuming that Ψ (x) ≥ c1 x2H − k1 , we obtain
 
 x 2 k3 k3 −c3 t
 
c1 E X (t) H − k1 ≤ + c2 xH −
2
e .
c3 c3
Thus we have proved the following:

Proposition 7.3 Let Ψ : H → R satisfy conditions (1)–(5) of Theorem 6.10 and

assume that condition (2) of Theorem 7.5 holds and that c1 x2H − k1 ≤ Ψ (x) for
x ∈ H and some constants c1 > 0 and k1 ∈ R. Then
 
 2 1 k3
lim sup E X x (t)H ≤ k1 + .
t→∞ c1 c3
In particular, {X x (t), t ≥ 0} is ultimately bounded.
262 7 Ultimate Boundedness and Invariant Measure

Let us now state the theorem connecting the ultimate boundedness with the exis-
tence of invariant measure.

Theorem 7.11 Let {X x (t), t ≥ 0} be a solution of (7.37). Assume that the embed-
ding V → H is compact. If X x (t) is ultimately bounded in the m.s.s., then there
exists an invariant measure μ for {X x (t), t ≥ 0}.

Proof Applying Itô’s formula to the function x2H and using the coercivity condi-
tion, we have
 2 t  2
E X x (t)H − x2H = EL X x (t)H ds
0
t  2 t  2
≤λ E X x (s)H ds − α E X x (s)V + γ t
0 0
with L defined in (7.38). Hence,
 
t  x 2 1 t  x 2
 
E X (s) V ds ≤ λ  
E X (s) H ds + xH + γ t .
2
0 α 0

Therefore,
1 T    1 EX x (t)2V T
P X x (t)V > R dt ≤ dt
T 0 T 0 R2
 T  
1 1 
≤ |λ| E X x (t)2 dt + x2 + γ T .
H H
αR 2 T 0

Now, by (7.36), EX x (t)2H ≤ M for t ≥ T0 and some T0 ≥ 0. But

 2
sup E X x (t) ≤ M  H
t≤T0

by Theorem 4.7, so that

1 T   
lim sup P X x (t)V > R dt
R→∞ T T 0
 T0 
|λ| 1  x 2 T  2
≤ lim sup E X (t) dt + E X x (t)H dt
R→∞ T αR 2 T 0 T0
 
|λ| T0  T − T0
≤ lim sup M + M
R→∞ T αR 2 T T
|λ|   
≤ lim 2
M +M , 0 ≤ T0 ≤ T .
R→∞ αR

Hence, given ε > 0, there exists an Rε such that

1 T   
sup P X x (t)V > Rε dt < ε.
T T 0
7.4 Ultimate Boundedness and Invariant Measure 263

By the assumption that the embedding V → H is compact, the set {v ∈ V :

vV ≤ Rε } is compact in H , and the result is proven. 

Remark 7.2 Note that a weaker condition on the second moment of X x (t), i.e.,

1 T  2
sup E X x (t)H dt < M for some T0 ≥ 0,
T >T0 T 0

is sufficient to carry out the proof of Theorem 7.11.

In Examples 7.2–7.6, we consider equations whose coefficients satisfy the condi-

tions imposed on the coefficients of (7.37) and the embedding V → H is compact,
so that an invariant measure exists if the solution is ultimately bounded in the m.s.s.

Theorem 7.12 Suppose that V → H is compact and the solution of {X x (t), t ≥ 0}

of (7.37) is ultimately bounded in the m.s.s. Then any invariant measure μ satisfies

x2V μ(dx) < ∞.

V

Proof Let f (x) = x2V and fn (x) = 1[0,n] (f (x)). Now fn (x) ∈ L1 (V , μ). We use
the ergodic theorem for a Markov process with an invariant measure (see [78],
p. 388). This gives

1 T
lim (Pt fn )(x) dt = fn∗ (x) μ-a.e.
T →∞ T 0

and Eμ fn∗ = Eμ fn , where Eμ fn = V fn (x) μ(dx).
By the assumption of ultimate boundedness, we have, as in the proof of Theo-
rem 7.11,
1 T  2 C|λ|
lim sup E X x (t)V dt ≤ , C < ∞.
T →∞ T 0 α
Hence,

1 T
fn∗ (x) = lim (Pt fn )(x) dt
T →∞ T 0

1 T 
≤ lim sup Pt f (x) dt
T →∞ T 0

1 T  2 C|λ|
= lim sup E X x (t)V dt ≤ .
T →∞ T 0 α

But fn (x) ↑ f (x), so that

C|λ|
Eμ f = lim Eμ fn = lim Eμ fn∗ ≤ . 
n→∞ n→∞ α
264 7 Ultimate Boundedness and Invariant Measure

Remark 7.3 (a) For parabolic Itô equations, one can easily derive the result using
Ψ (x) = x2H and Theorem 7.11.
(b) Note that if μn ⇒ μ and the support of μn is in V with the embedding
V → H being compact, then by the weak convergence the support of μ is in V
by the same argument as in Example 7.6.

Let us now consider the problem of uniqueness of the invariant measure.

Theorem 7.13 Suppose that for ε, δ, and R > 0, there exists a constant T0 (ε, δ, R)
> 0 such that for T ≥ T0 ,

1 T   
P X x (t) − X y (t)V ≥ δ dt < ε
T 0

for all x, y ∈ VR = {v ∈ V : vV ≤ R} with the embedding V → H being com-

pact. If there exists an invariant measure μ for a solution of (7.37), {X x0 (t), t ≥ 0},
X(0) = x0 , with support in V , then it is unique.

Proof Suppose that μ, ν are invariant measures with support in V . We need to show
that

f (x)μ(dx) = f (x)ν(dx)
H H
for f uniformly continuous bounded on H , since such functions form a determining
class.
For G ∈ B(H ), define

1 T  
μxT (G) = P X x (t) ∈ G dt, x ∈ H, T > 0.
T 0

Then, using invariance of μ and ν, we have

 
 
 f (x) μ(dx) − f (x) ν(dx)

H H
 
 * y +
=  f (x) μT (dx)μ(dy) − μzT (dx)ν(dz) 
H H
 
 
≤  f (x) μT (dx) −
y
f (x)μzT (dx) μ(dy)ν(dz).

H ×H H H

Let
 
 
F (y, z) =  f (x)μzT (dx).
y
f (x)μT (dx) −
H H
Then, using the fact that μ, ν have the supports in V , we have
 
   
 f (x) μ(dx) − f (x) ν(dx)≤ F (y, z) μ(dy)ν(dz).
 
H H V ×V
7.4 Ultimate Boundedness and Invariant Measure 265

Let VRc = V \ VR and choose R > 0 such that

   
μ VRc + ν VRc < ε.

Then,
 
     
 f (x) μ(dx) − f (x) ν(dx) ≤ F (y, z) μ(dy)ν(dz) + 4ε + 2ε 2 M,

H H VR ×VR

where M = supx∈H |f (x)|. But for δ > 0,

 
F (y, z)μ(dy)ν(dz)
VR ×VR
 
1 T     
≤ E f X y (t) − f X z (t)  μ(dy)ν(dz)
VR ×VR T 0

1 T     
≤ 2M sup P X y (t) − X z (t)V > δ + sup f (y) − f (z)
y,z∈VR T 0 y,z∈VR
y−z<δ

≤ 2Mε + ε

for T ≥ T0 , since f is uniformly continuous.  

Using the last inequality and the bound for | H f (x) μ(dx) − H f (x) ν(dx)|,
we obtain the result. 

Let us now give a condition on the coefficients of the SDE (7.37) which guaran-
tees the uniqueness of the invariant measure. We have proved in Theorem 7.11 (see
Remark 7.3), that the condition
 
1 T  
2

sup E X (t) H dt ≤ M for some T0 ≥ 0
x
T >T0 T 0

implies that there exists an invariant measure to the strong solution {X x (t), t ≥ 0},
whose support is in V .

Theorem 7.14 Suppose that V → H is compact, the coefficients of (7.37) satisfy

the coercivity condition (6.39), and that for u, v ∈ V ,
   2
2 u − v, A(u) − A(v) + B(u) − B(v)L (K ,H ) ≤ −cu − v2V ,
2 Q

where the norm  · L2 (KQ ,H ) is the Hilbert–Schmidt norm defined in (2.7). Assume
that the solution {X x (t), t ≥ 0} of (7.37) is ultimately bounded in the m.s.s. Then
there exists a unique invariant measure.

Proof By Itô’s formula, we have, for t > 0,

 2 t   t 
E X x (t)H = x2H + 2E X x (s), A X x (s) ds + E B X x (s) 2
L
ds.
2 (KQ ,H )
0 0
266 7 Ultimate Boundedness and Invariant Measure

Using the coercivity condition (C), (6.39), we have

 2 t    t 
E X x (t)H + αE X x (s)2 ds ≤ x2 + γ t + λE
V H
X x (s)2 ds.
H
0 0

It follows, similarly as in the proof of Theorem 7.11, that

1 T  2 |γ | + x2H /T0 |λ| T  2

sup E X x (s)V ds ≤ + sup E X x (s)H ds.
T >T0 T 0 α α T >T0 0

By the Chebychev inequality, we know that

 
1 T    1 1 T  2
P X x (s)V > R ≤ 2 E X x (s)V ds .
T 0 R T 0

Hence, using the arguments in Example 7.6, an invariant measure exists and is sup-
ported on V . To prove the uniqueness, let X x1 (t), X x2 (t) be two solutions with
initial values x1 , x2 . We apply Itô’s formula to X(t) = X x1 (t) − X x2 (t) and obtain

 2 t    
E X(t)H ≤ x1 − x2 2H + 2E X(s) − A X x1 (s) − A X x2 (s) ds
0
t    
+E B X x1 (s) − B X x2 (s) 2 ds.
L 2 (KQ ,H )
0

Using the assumption, we have

 2 t  2
E X(t)H ≤ x1 − x2 2H − c E X(s)V ds,
0

which implies that

t  2 1
E X(s)V ≤ x1 − x2 2H .
0 c
It now suffices to refer to the Chebychev inequality and Theorem 7.13 to complete
the proof. 

7.4.2 Semilinear Equations Driven by a Q-Wiener Process

Let us consider now the existence of an invariant measure for a mild solution of a
semilinear SDE with deterministic initial condition

dX(t) = (AX(t) + F (X(t))) dt + B(X(t)) dWt ,
(7.39)
X(0) = x ∈ H,
7.4 Ultimate Boundedness and Invariant Measure 267

where A is the infinitesimal generator of a pseudo-contraction C0 -semigroup S(t)

on H , and the coefficients F : H → H and B : H → L (K, H ), independent of
t and ω, are in general nonlinear mappings satisfying the linear growth condition
(A3) and the Lipschitz condition (A4) in Sect. 3.3. We know from Theorem 3.6
that the solution is a homogeneous Markov process and from Theorem 3.7 that it is
continuous with respect to the initial condition, so that the associated semigroup is
Feller.
We studied a special case in Example 7.7. Here we look at the existence under
the assumption of exponential boundedness in the m.s.s. We will use the Lyapunov
function approach developed earlier in Theorem 7.8 and Corollary 7.3. We first give
the following proposition.

Proposition 7.4 Suppose that the mild solution {X x (t)} of (7.39) is ultimately
bounded in the m.s.s. Then any invariant measure ν of the Markov process
{X x (t), t ≥ 0} satisfies

y2H ν(dy) ≤ M,
H
where M is as in (7.36).

The proof is similar to the proof of Theorem 7.12 and is left to the reader as an
exercise.

Exercise 7.7 Prove Proposition 7.4.

Theorem 7.15 Suppose that the solution {X x (t), t ≥ 0} of (7.39) is ultimately

bounded in the m.s.s. If for all R > 0, δ > 0, and ε > 0, there exists T0 =
T0 (R, δ, ε) > 0 such that for all t ≥ T0 ,
  
P X x (t) − X y (t)H > δ < ε for x, y ∈ BH (R) (7.40)

with BH (R) = {x ∈ H, x ≤ R}, then there exists at most one invariant measure
for the Markov process X x (t).

Proof Let μi , i = 1, 2, be two invariant measures. Then, by Proposition 7.4, for each
ε > 0, there exists R > 0 such that μi (H \ BH (R)) < ε. Let f be a bounded weakly
continuous function on H . We claim that there exists a constant T = T (ε, R, f ) > 0
such that
 
Pt f (x) − Pt f (y) ≤ ε for x, y ∈ BH (R) if t ≥ T .

Let C be a weakly compact set in H . The weak topology on C is given by the metric
∞
 1  
,
d(x, y) = ek , x − y H x, y ∈ C, (7.41)
2k
k=1

where {ek }∞
k=1 in an orthonormal basis in H .
268 7 Ultimate Boundedness and Invariant Measure

By the ultimate boundedness, there exists T1 = T1 (ε, R) > 0 such that for
T ≥ T1 ,
 
P X x (t) ∈ BH (R) > 1 − ε/2 for x ∈ BH (R).
Now f is uniformly continuous w.r.t. the metric (7.41) on BH (R). Hence, there
exists δ  > 0 such that x, y ∈ HR with d(x, y) < δ  imply that |f (x) − f (y)| ≤ δ,
and there exists J > 0 such that
∞
 1  
 ≤ δ  /2 for x, y ∈ BH (R).
ek , x − y H
2k
k=J +1

Since P (|ek , X x (t) − X y (t) | > δ) ≤ P (X x (t) − X y (t)H > δ), by the given as-
sumption we can choose T2 ≥ T1 such that for t ≥ T2 ,
 J 
   2 
P ek , X (t) − ek , X (t) > δ /2 ≥ 1 − ε/3
x y
(7.42)
k=1

for x, y ∈ BH (R). Hence, for t ≥ T2 ,

     
P f X x (t) − f X y (t)  ≤ δ
  
≥ P X x (t), X y (t) ∈ BH (R), d X x (t), X y (t) ≤ δ 
 
 J
1   
≥ P X (t), X (t) ∈ BH (R),
x y
ek , X (t) − X (t) H  ≤ δ /2
x y 
2k
k=1
   
≥ P X x (t), X y (t) ∈ BH (R),  ek , X x (t) − X y (t) H  ≤ δ  /2, k = 1, . . . , J
≥ 1 − ε/3 − ε/3 − ε/3 = 1 − ε,

since the last probability above is no smaller than that in (7.42).

Now, with M0 = sup |f (x)|, given ε > 0, choose T so that for t ≥ T ,
      ε
P f X x (t) − f X y (t)  ≤ ε/2 ≥ 1 − .
4M0
Then
     ε ε
E f X x (t) − f X y (t)  ≤ + 2M0 = ε.
2 4M0
Note that for invariant measures μ1 , μ2 ,

f (x)μi (dx) = (Pt f )(x)μi (dx), i = 1, 2.

H H

For t ≥ T , we have
 
 
 f (x)μ1 (dx) − f (y)μ2 (dy)

H H
7.4 Ultimate Boundedness and Invariant Measure 269
 
 * + 
=  f (x) − f (y) μ1 (dx)μ2 (dy)
H H
 
 * + 
=  (Pt f )(x) − (Pt f )(y) μ1 (dx)μ2 (y)
H H
  

=  + +
BH (R) H \BH (R) BH (R) H \BH (R)

* + 
× (Pt f )(x) − (Pt f )(y) μ1 (dx)μ2 (dy)

≤ ε + 2(2M0 )ε + 2M0 ε 2 .

Since ε > 0 is arbitrary, we conclude that

f (x) μ1 (dx) = f (x) μ2 (dx). 

H H

In case we look at the solution to (7.39), whose coefficients satisfy the linear
growth and Lipschitz conditions (A3) and (A4) of Sect. 3.1 in Chap. 3, we conclude
that under assumption (7.40) and conditions for exponential ultimate boundedness,
there exists at most one invariant measure.
Note that in the problem of existence of the invariant measure, the relative weak
compactness of the sequence μn in Theorem 7.10 is crucial. In the variational case,
we achieved this condition, under ultimate boundedness in the m.s.s., assuming that
the embedding V → H is compact. For mild solutions, Ichikawa [33] and Da Prato
and Zabczyk [11], give sufficient conditions. Da Prato and Zabczyk use a factoriza-
tion technique introduced in [10]. We start with the result in [32].

Theorem 7.16 Assume that A is a self-adjoint linear operator with eigenvectors

{ek }∞
k=1 forming an orthonormal basis in H and that the corresponding eigenvalues
−λk ↓ −∞ as k → ∞. Let the mild solution of (7.39) satisfy

1 T  2  
E X x (s)H ds ≤ M 1 + x2H . (7.43)
T 0

Then there exists an invariant measure for the Markov semigroup generated by the
solution of (7.39).

Proof The proof depends on the following lemma.

Lemma 7.2 Under the conditions of Theorem 7.16, the set of measures

1 t
μt (·) = P (s, x, ·) for t ≥ 0
t 0

with P (s, x, A) = P (X x (s) ∈ A) is relatively weakly compact.

270 7 Ultimate Boundedness and Invariant Measure

Proof Let yk (t) = X x (t), ek H . Then, by a well-known result about the weak com-
pactness ([25], Vol. I, Chap. VI, Sect. 2, Theorem 2), we need to show that the
expression
∞ 
1 T  2
Eyk (t) dt
T 0
k=1

is uniformly convergent in T .
Let S(t) be the C0 -semigroup generated by A. Since S(t)ek = e−λk t ek for each
k, yk (t) satisfies
t   
yk (t) = e−λk t xk0 + e−λk (t−s) ek , F X x (s) H ds
0
t    
+ e−λk (t−s) ek , B X x (s) dW (s) H ,
0
 2
 0 2  t   x  
Eyk2 (t) ≤ 3e −2λk t
xk + 3E  e −λk (t−s)
ek , F X (s) H ds 
0
 
 t    2
+ 3E  e−λk (t−s) ek , B X x (s) dWs  .
0

For N large enough, so that λN > 0, and any m > 0, using Exercise 7.8 and assump-
tion (7.43), we have

+m
N  2
1 T  t   x  
E  e −λk (t−s)
ek , F X (s) H ds  dt
T 0 0
k=N

1 T t    2
≤ e2(−λk +ε)(t−s)  ek , F X x (s) H  ds dt
2εT 0 0

1 T T    2
= e2(−λk +ε)(t−s) dt  ek , F X x (s) H  ds
T 0 r
T
0 EF (X x (s))2H ds c1 (1 + x2H )
≤ ≤
4ε(λN − ε)T ε(λN − ε)

for some constants ε > 0 and c1 > 0.

Utilizing the Hölder inequality, we also have that

+m
N  
1 T  t   x  2
E  e −λk (t−s)
ek , B X (s) dW (s)  dt
T 0 0
N
T
tr(Q) 0 EB(X x (t))2L (K,H ) dt c2 tr(Q)(1 + x2H )
≤ ≤
2λN T λN
7.4 Ultimate Boundedness and Invariant Measure 271

for some constant c2 > 0. Thus,

+m
N
1 T 3x2H
Eyk2 (t) dt ≤
T 0 2λN
N
, -
  1 tr(Q)
+ 3(c1 + c2 ) 1 + xH
2
+ .
δ(λN − δ) λN

Thus the condition in [25] holds. 

The proof of Theorem 7.16 is an immediate consequence of the lemma. 

Exercise 7.8 Let p > 1, and let g be a nonnegative locally p-integrable function on
[0, ∞). Then for all ε > 0 and real d,
 t p  p/q t
1
e d(t−r)
g(r) dr ≤ ep(d+ε)(t−r) g p (r) dr,
0 qε 0

where 1/p + 1/q = 1.

7.4.3 Semilinear Equations Driven by a Cylindrical Wiener

Process

We finally present a result in [12], which uses an innovative technique to prove the
tightness of the laws L (X x (t)). We start with the problem

dX(t) = (AX(t) + F (X(t))) dt + B(X(t)) d W̃t ,
(7.44)
X(0) = x ∈ H,

where W̃t is a cylindrical Wiener process in a separable Hilbert space K. Assume

that the coefficients and the solution satisfy the following hypotheses.

Hypothesis (DZ) Let conditions (DZ1)–(DZ4) of Sect. 3.10 hold, and, in addition,
assume that:
(DZ5) {S(t), t > 0} is a compact semigroup.
(DZ6) For all x ∈ H and ε > 0, there exists R > 0 such that for every T ≥ 1,

1 T   
P X x (t)H > R dt < ε,
T 0

where {X x (t), t ≥ 0} is a mild solution of (7.44).

272 7 Ultimate Boundedness and Invariant Measure

Remark 7.4 (a) Condition (DZ6) holds if {X x (t), t ≥ 0} is ultimately bounded in

the m.s.s.
(b) In the special case where Wt is a Q-Wiener process, we can replace B with
B̃ = BQ1/2 .

Theorem 7.17 Under Hypothesis (DZ), there exists an invariant measure for the
mild solution of (7.44).

Proof We recall the factorization formula used in Lemma 3.3. Let x ∈ H , and
t  
Y x (t) = (t − s)−α S(t − s)B X x (s) dWs .
0

Then
  sin πα
X x (1) = S(1)x + G1 F X x (·) (1) + Gα Y x (·)(1) P -a.s.
π
By Lemma 3.12, the compactness of the semigroup {S(t), t ≥ 0} implies that the
operators Gα defined by
t  
Gα f (t) = (t − s)α−1 S(t − s)f (s) ds, f ∈ Lp [0, T ], H ,
0

are compact from Lp ([0, T ], H ) into C([0, T ], H ) for p ≥ 2 and 1/p < α ≤ 1.
Consider γ : H × Lp ([0, 1], H ) × Lp ([0, 1], H ) → H ,

γ (y, f, g) = S(1)y + G1 f (1) + Gα g(1).

Then γ is a compact operator, and hence, for r > 0, the set


K(r) = x ∈ H : x = S(1)y + G1 f (1) + Gα g(1),

rπ
yH ≤ r, f Lp ≤ r, gLp ≤
sin πα
is relatively compact in H .
We now need the following lemma.

Lemma 7.3 Assume that p > 2, α ∈ (1/p, 1/2), and that Hypothesis (DZ) holds.
Then there exists a constant c > 0 such that for r > 0 and all x ∈ H with xH ≤ r,
   p
P X x (1) ∈ K(r) ≥ 1 − cr −p 1 + xH .

Proof By Lemma 3.13, using Hypothesis (DZ3), we calculate

1 
E Y x (s)p ds
H
0
7.4 Ultimate Boundedness and Invariant Measure 273

1
 p
s   
=E  (s − u)−α S(s − u)B X x (u) dWu 
  ds
0 0 H
1 s  
2
p/2
−2α 
≤ kE (s − u) S(s − u)B X (u) L x
(K,H )
du ds
2
0 0
1 s   2 
p/2
≤ k2p/2 E (s − u)−2α K 2 (s − u) 1 + X x (u)H du ds.
0 0

By (3.103) and Exercise 3.7,

1  p/2
p 1 1  2 p/2
E  x 
Y (s) H ds ≤ k2 p/2
t −2α
K (t) dt
2
E 1 + X x (u)H du
0 0 0
 p
≤ k1 1 + xH for some k1 > 0.

Also, using Hypothesis (DZ2), we get

1    
E F X x (u) p du ≤ k2 1 + xp , x ∈ H.
H H
0

By the Chebychev inequality,

 
  πr
P Y x (·)Lp ≤
sin απ
sinp απ  p  p
≥ 1 − r −p p
E Y x (·)Lp ≥ 1 − r −p π −p k1 1 + xH
π
  x  
 
P F X (·) Lp ≤ r
  p   p
≥ 1 − r −p E F X x (·) Lp ≥ 1 − r −p k2 1 + xH ,

giving
  
   x  πr   x  
P X (1) ∈ K(r) ≥ P
x Y (·) p ≤  
∩ F X (·) Lp ≤ r
L sin απ
  p
≥ 1 − r −p π −p k1 + k2 1 + xH . 

We continue the proof of Theorem 7.17.

For any t > 1 and r > r1 > 0, by the Markov property (recall Proposition 3.4)
and Lemma 7.3, we have
   
P X x (t) ∈ K(r) = P t, x, K(r)
 
= P 1, y, K(r) P (t − 1, x, dy)
H
 
≥ P 1, y, K(r) P (t − 1, x, dy)
yH ≤r1
274 7 Ultimate Boundedness and Invariant Measure

   p 
≥ 1 − c r −p 1 + r1 P (t − 1, x, dy)
yH ≤r1
   p    
= 1 − c r −p 1 + r1 P X x (t − 1)H ≤ r1 ,

giving

1 T    p 1
T   
P X x (t) ∈ K(r) dt ≥ 1 − cr −p 1 + r1 P X x (t)H ≤ r1 dt.
T 0 T 0

If we choose r1 according to condition (DZ6) and take r > r1 sufficiently large,

T
we obtain that T1 0 P (t, x, ·) dt is relatively compact, ensuring the existence of an
invariant measure. 

7.5 Ultimate Boundedness and Weak Recurrence of the

Solutions
In Sect. 4.3 we proved the existence and uniqueness for strong variational solutions,
and in Sect. 4.4 we showed that they are strong Markov and Feller processes. We
will now study weak (positive) recurrence of the strong solution of (7.45), which is
(exponentially) ultimately bounded in the m.s.s.
The weak recurrence property to a bounded set was considered in [59] for the so-
lutions of SDEs in the finite dimensions and in [33] for solutions of stochastic evo-
lution equations in a Hilbert space. This section is based on the work of R. Liu [51].
Let us consider a strong solution of the variational equation

dX(t) = A(X(t)) dt + B(X(t)) dWt ,
(7.45)
X(0) = x ∈ H.

We start with the definition of weak recurrence.

Definition 7.6 A stochastic process X(t) defined on H is weakly recurrent to a

compact set if there exists a compact set C ⊂ H such that
 
P x X(t) ∈ C for some t ≥ 0 = 1 for all x ∈ H,

where P x is the conditional probability under the condition X(0) = x. The set C is
called a recurrent region. From now on recurrent means recurrent to a compact set.

Theorem 7.18 Suppose that V → H is compact and the coefficients of (7.45) sat-
isfy the coercivity and the weak monotonicity conditions (6.39) and (6.40). If its
solution {X x (t), t ≥ 0} is ultimately bounded in the m.s.s., then it is weakly recur-
rent.

Proof We prove the theorem using a series of lemmas.

7.5 Ultimate Boundedness and Weak Recurrence of the Solutions 275

Lemma 7.4 Let {X(t), t ≥ 0} be a strong Markov process in H . If there exists

a positive Borel-measurable function ρ : H → R+ , a compact set C ⊂ H , and a
constant δ > 0 such that
   
P x X ρ(x) ∈ C ≥ δ for all x ∈ H,

then X x (t) is weakly recurrent with the recurrence region C.

Proof For a fixed x ∈ H , let τ1 = ρ(x), 1 = {ω : X(τ1 ) ∈ τ2 = τ1 + ρ(X(τ1 )),

/ C},
/ C}, τ3 = τ2 + ρ(X(τ2 )), etc. Define ∞ = ∞
2 = {ω : X(τ2 ) ∈ i=1 i . Since

ω : X(t, ω) ∈ / C for any t ≥ 0 ⊂ ∞ ,

it suffices to show that P x (∞ ) = 0. Note that

P x (1 ) < 1 − δ < 1.

Since ρ : H → R+ is Borel measurable and τi is a stopping time for each i, we can

use the strong Markov property to get
  
P x (1 ∩ 2 ) = E x E x 11 (ω)12 (ω)|Fτ1
  
= E x 11 (ω)E x 12 (ω)|Fτ1
  
= E x 11 (ω)E x 12 (ω)|X(τ1 )
    
= E x 11 (ω)P X(τ1 ) ω : X ρ(τ1 ) ∈/C .

But, by the assumption,

     
P X(τ1 ) ω : X ρ X τ1 (ω) ∈ / C < 1 − δ,

so that
P x (1 ∩ 2 ) < (1 − δ)2 .
By repeating the above argument, we obtain
 n 
3
Px i < (1 − δ)n ,
i=1

which converges to zero, and this completes the proof. 

Lemma 7.5 Let {X(t), t ≥ 0} be a continuous strong Markov process. If there exists
a positive Borel-measurable function γ defined on H , a closed set C, and a constant
δ > 0 such that
γ (x)+1  
P x X(t) ∈ C dt ≥ δ for all x ∈ H, (7.46)
γ (x)
276 7 Ultimate Boundedness and Invariant Measure

then, there exists a Borel-measurable function ρ : H → R+ such that γ (x) ≤

ρ(x) ≤ γ (x) + 1 and
   
P x X ρ(x) ∈ C ≥ δ for all x ∈ H. (7.47)

Proof By the assumption (7.46), there exists tx ∈ [γ (x), γ (x) + 1) such that
 
P x X(tx ) ∈ C ≥ δ.

Define
*    
ρ(x) = inf t ∈ γ (x), γ (x) + 1 : P x ω : X(t, ω) ∈ C ≥ δ .

Since the mapping t → X(t) is continuous and the characteristic function of a closed
set is upper semicontinuous, we have that the function
 
t → P x X(t) ∈ C

is upper semicontinuous for each x. Hence,

   
P x X ρ(x) ∈ C ≥ δ.

We need to show that the function x → ρ(x) is Borel measurable. Let us define
Bt (H ) = B(H ), for t > 0. Since {X(t), 0 ≤ t ≤ T } is a Feller process, the map
Θ : (t, x) → P x (ω : X(t) ∈ C) from ([0, T ] × H, B([0, T ] × H )) to (R1 , B(R1 ))
is measurable (see [54], [27]). Hence, Θ is a progressively measurable process with
respect to {Bt (H )}. By Corollary 1.6.12 in [16], x → ρ(x) is Borel measurable. 

Let us now introduce some notation. Let Br = {v ∈ V : vV ≤ r} be a sphere

in V with the radius r, centered at 0, and let B r be its closure in (H,  · H ). For
A ⊂ H , denote its interior in (H,  · H ) by A0 . If Brc = H \ Br , then (B r )c = (Brc )0 .

Lemma 7.6 Suppose that the coefficients of (7.45) satisfy the coercivity condi-
tion (6.39) and, in addition, that its solution {X x (t), t ≥ 0} exists and is ultimately
bounded in the m.s.s. Then there exists a positive Borel-measurable function ρ on
H such that
    1  
P x ω : X ρ(x), ω ∈ B r ≥ 1 − 2 |λ|M1 + M1 + |γ | , x ∈ H, (7.48)
αr
and
    0  1  
P x ω : X ρ(x), ω ∈ Brc ≤ 2 |λ|M1 + M1 + |γ | , x ∈ H, (7.49)
αr
where α, λ, γ are as in the coercivity condition, and M1 = M + 1 with M as in the
ultimate boundedness condition (7.36).
7.5 Ultimate Boundedness and Weak Recurrence of the Solutions 277

Proof Since lim supt→∞ E x X(t)2H ≤ M < M1 for all x ∈ H , there exist positive
numbers {Tx , x ∈ H } such that
 2
E x X(t)H ≤ M for t ≥ Tx .

Hence, we can define

 2 
γ (x) = inf t : E x X(s)H ≤ M1 for all s ≥ t .

Since t → E x X(t)2H is continuous, E x X(γ (x))2H ≤ M1 . The set

  2 
x : γ (x) ≤ t = x : E x X(s)H ≤ M1 for all s ≥ t
3  2 
= x : E x X(s)H ≤ M1
s≥t
s∈Q

is in B(H ), since the function x → E x X(s)2 is Borel measurable. Using Itô’s

formula (4.37) for x2H , then taking the expectations on both sides, and applying
the coercivity condition (6.39), we arrive at
  2   2
E x X γ (x) + 1 H − E x X γ (x) H
γ (x)+1          ∗ 
= Ex 2 X(s), A X(s) + tr B X(s) Q B X(s) ds
γ (x)
γ (x+1)  2 γ (x)+1  2
≤λ E x X(s)H ds − α E x X(s)V ds + γ .
γ (x) γ (x)

It follows that
γ (x)+1  2 1 
E X(s)V ds ≤ |λ|M1 + M1 + |γ | .
γ (x) α

Using Chebychev’s inequality, we get

γ (x+1)     1  
P x ω : X(t, ω)V > r dt ≤ 2 |λ|M1 + M1 + |γ | .
γ (x) αr

Hence,
γ (x)+1   0  1  
P x ω : X(t, ω) ∈ Brc ≤ 2 |λ|M1 + M1 + |γ | ,
γ (x) αr

and consequently,
γ (x)+1   1  
P x ω : X(t, ω) ∈ B r dt ≥ 1 − 2 |λ|M1 + M1 + |γ | .
γ (x) αr
278 7 Ultimate Boundedness and Invariant Measure

Using Lemma 7.5, we can claim the existence of a positive Borel-measurable

function ρ(x) defined on H such that γ (x) ≤ ρ(x) ≤ γ (x) + 1, and (7.48) and
then (7.49) follow for r > 0 and for all x ∈ H . 

We now conclude the proof of Theorem 7.18. Using (7.48), we can choose r
large enough such that
    1
P x ω : X ρ(x), ω ∈ B r ≥ for x ∈ H.
2

Since the mapping V → H is compact, the set B r is compact in H , giving that X(t)
is weakly recurrent to B r by Lemma 7.4. 

Definition 7.7 An H -valued stochastic process {X(t), t ≥ 0} is called weakly pos-

itive recurrent to a compact set if there exists a compact set C ⊂ H such that X(t)
is weakly recurrent to C and the first hitting time to C,

τ = inf t ≥ 0 : X(t) ∈ C ,

has finite expectation for any x = X(0) ∈ H .

Theorem 7.19 Suppose that V → H is compact and the coefficients of (7.45) sat-
isfy the coercivity condition (6.39) and the monotonicity condition (6.40). If its so-
lution {X x (t), t ≥ 0} is exponentially ultimately bounded in the m.s.s., then it is
weakly positively recurrent.

Proof We know that

 2
E x X(t)H ≤ ce−βt x2H + M for all x ∈ H.

Let M1 = M + 1, and w(r) = 1

β ln(1 + cr 2 ), r ∈ R. Then we have
 2  
E x X(t)H ≤ M1 for x ∈ H and t ≥ w xH ,

and
∞
 w((l + 1)N )
<∞ for any N ≥ 0. (7.50)
l2
l=1
√ √
Let K = (1 + Δ) |λ|M1 + M1 + |γ |/ α, and let us define the sets

E0 = B K ,
 c 0
El = B (l+1)K − B lK = B (l+1)K ∩ BlK for l ≥ 1,

where Br is a sphere in V with the radius r, centered at 0. We denote w  (l) =

w(lKα0 ) + 1 with α0 such that xH ≤ α0 xV for all x ∈ V . As in the proof of
7.5 Ultimate Boundedness and Weak Recurrence of the Solutions 279

Lemma 7.6, there exists a Borel-measurable function ρ(x) defined on H satisfying

w(xH ) ≤ ρ(x) ≤ w(xH ) + 1, and
    c 0  1  
P x ω : X ρ(x), ω ∈ BlK ≤ |λ|M1 + M1 + |γ |
α(lK)2
1
≤ for all x ∈ H. (7.51)
l 2 (1 + Δ)2
Let

τ1 = ρ(x), x1 (ω) = X(τ1 , ω), 1 = ω : x1 (ω) ∈
/ E0 ,
  
τ2 = τ1 + ρ x1 (ω) , x2 (ω) = X(τ2 , ω), 2 = ω : x2 (ω) ∈
/ E0 , ...,
∞
and so on. Let ∞ = i=1 i . As in the proof of Lemma 7.4,
∞ 
3
P x
i = 0.
i=1

Hence,  differs from

∞
& ∞
& 
ci = ω : xi (ω) ∈ E0
i=1 i=1

by at most a set of P x -measure zero. Let

&
i−1
 
Ai = ci − ω : xj (ω) ∈ E0 = ω : x1 (ω) ∈
/ E0 , . . . , xi−1 ∈
/ E 0 , xi ∈ E 0 .
j =1
%∞
Then  differs from i=1 Ai by at most a set of P x -measure zero. For i ≥ 2, let us
further partition
& j ,...,ji−1
Ai = Ai 1 ,
j1 ,j2 ,...,jn−1

where
j ,...,ji−1 
Ai 1 = ω : x1 (ω) ∈ Ej1 , . . . , xi−1 (ω) ∈ Eji−1 , xi (ω) ∈ E0 .

Let τ (ω) be first hitting time to E0 . Then for ω ∈ A1 = c1 ,

 
τ (ω) ≤ ρ(x) ≤ w xH + 1,

j ,...,ji−1
and for ω ∈ Ai 1 ,
 
τ (ω) ≤ τi (ω) ≤ τi−1 (ω) + ρ xi−1 (ω) .
280 7 Ultimate Boundedness and Invariant Measure

j ,...,ji−1
Moreover, for ω ∈ Ai 1 ,

xi−1 (ω) ∈ Eji−1 ⊂ B (ji−1 +1)K .

Hence,
   
xi−1 (ω) ≤ α0 xi−1 (ω) ≤ α0 (ji−1 + 1)K,
H V
giving
      
ρ xi−1 (ω) ≤ w xi−1 (ω)H + 1 ≤ w α0 (ji−1 + 1)K + 1 = w  (ji−1 + 1)

and
τ (ω) ≤ τi−1 + w  (ji−1 + 1).
Using induction,
 
τ (ω) ≤ w xH + 1 + w  (j1 + 1) + · · · + w  (ji−1 + 1).

By the strong Markov property,

 j ,...,ji−1   
P x Ai 1 = P x ω : x1 (ω) ∈ Ej1 , . . . , xi−1 (ω) ∈ Eji−1 , xi (ω) ∈ E0
 
≤ P x ω : x1 (ω) ∈ Ej1 , . . . , xi−1 (ω) ∈ Eji−1
  
= P x ω : x1 (ω) ∈ Ej1 , . . . , xi−2 (ω) ∈ Eji−2 ∩ xi−1 (ω) ∈ Eji−1
     !
≤ E x 1{ω:x1 (ω)∈Ej1 ,...,xi−2 ∈Eji−2 } P xi−2 (ω) ω̃ : X ρ xi−2 (ω) , ω̃ ∈ Eji−1 .

Since Eji−1 = B (ji−1 +1)K ∩ (Bjci−1 K )0 , we get by (7.51)

     
P xi−2 (ω) ω̃ : X ρ xi−2 (ω) , ω̃ ∈ Eji−1
      0 
≤ P xi−2 ω̃ : X ρ xi−2 (ω) , ω̃ ∈ Bjci−1 K
1
≤ 2 (1 + Δ)2
.
ji−1

Hence,
 j ,...,ji−1  1  
P x Ai 1 ≤ 2 (1 + Δ)2
P x ω : x1 (ω) ∈ Ej1 , . . . , xi−2 (ω) ∈ Eji−2 .
ji−1

By induction,
 j ,...,ji−1  1 1
P x Ai 1 ≤ ,
(1 + Δ)2(i−1) j12 · · · ji−1
2

which implies that P x (Ai ) < 1, for Δ large enough.

7.5 Ultimate Boundedness and Weak Recurrence of the Solutions 281

Now
  j ,...,ji−1 
E x (τ ) ≤ P x Ai 1
i,j1 ,...,ji−1 ≥1
*   +
× w  xH + 1 + w  (j1 + 1) + · · · + w  (ji−1 + 1)
∞ 
   1
≤ w xH + 1 +
(1 + Δ)2(i−1)
i=2
  
w  (xH ) + 1 + w  (j1 + 1) + · · · + w  (ji−1 + 1)
j1 ,...,ji−1 ≥1
j12 · · · ji−1
2

∞ 
   1    
= w xH + 1 + w xH + 1
(1 + Δ)2(i−1)
i=2
    
1 w  (j1 + 1)
+ (i − 1)
j 2 · · · ji−1
j1 ,...,ji−1 ≥1 1
2 j 2 · · · ji−1
j1 ,...,ji−1 ≥1 1
2

 ∞  i−1 
     A
= w xH + 1 1 +
(1 + Δ)2
i=2
∞ 
 i−2
B A
+ (i − 1),
(1 + Δ)2 (1 + Δ)2
i=2

 ∞ 1 
where A = ∞ l=1 l 2 , and B =
1
l=1 l 2 w (l + 1), with both series converging due
to (7.50).
Consequently, E x (τ ) is finite for Δ large enough. The set E0 is compact since
the embedding V → H is compact. 

We have given precise conditions using a Lyapunov function for exponential ulti-
mate boundedness in the m.s.s. We can thus obtain sufficient conditions for weakly
(positive) recurrence of the solutions in terms of a Lyapunov function.
We close with important examples of stochastic reaction–diffusion equations.
Let O ⊂ Rn be a bounded domain with smooth boundary ∂O, and p be a positive
integer. Let V = W 1,2 (O) and H = W 0,2 (O) = L2 (O). We know that V → H is a
compact embedding. Let

 ∂ α1 ∂ αn
A0 (x) = aα (x) · · · ,
∂x1α1 ∂xnαn
|α|≤2p

where α = (α1 , . . . , αn ) is a multiindex. If A0 is strongly elliptic, then by Garding

inequality ([63], Theorem 7.2.2) A0 is coercive.
282 7 Ultimate Boundedness and Invariant Measure

Example 7.8 (Reaction–diffusion equation) Consider a parabolic Itô equation


dX(t, x) = A0 X(t, x) dt + f (X(t, x)) dt + B(t, x)) dWt ,
(7.52)
X(0, x) = ϕ(x) ∈ H, X|∂O = 0,

where A0 , f , and B satisfy the following conditions:

(1) A0 : V → V ∗ is a strongly elliptic operator.
(2) f : H → H and B : H → L (K, H ) satisfy
     
f (h)2 + B(h)2 ≤ K 1 + h2H , h ∈ H.
H L (K,H )

(3) f (h1 ) − f (h2 )2H + tr((B(h1 ) − B(h2 ))Q(B(h1 ) − B(h2 ))∗ ) ≤ λh1 − h2 ,
h1 , h2 ∈ H .
If the solution to the equation

du(t, x) = A0 u(t, x) dt

is exponentially ultimately bounded and, as hH → ∞,

   
f (h) = o hH ,
H
   
B(h) = o hH ,
L (K,H )

then the strong variational solution of (7.52) is exponentially ultimately bounded in

the m.s.s. by Proposition 7.1, and consequently it is weakly positive recurrent.

Example 7.9 (Reaction–diffusion equation) Consider the following one-dimensional

parabolic Itô equation
⎧  2   
⎨dX(t, x) = α 2 ∂ X + β ∂X + γ X + g(x) dt + σ ∂X + σ X dW ,
1 2 t
∂x 2 ∂x ∂x (7.53)
⎩
u(0, x) = ϕ(x) ∈ L (O) ∩ L (O), X|∂O = 0,
2 1

where O = (0, 1), and Wt is a standard Brownian motion.

Similarly as in Example 7.3, if −2α 2 + σ12 < 0, then the coercivity and weak
monotonicity conditions (6.39) and (6.40) hold, and Theorem 4.7 implies the exis-
tence of a unique strong solution.
With Λ(v) = v2H and L defined by (6.15), we get
 
  2
2  dv 
  1
L Λ(v) ≤ −2α + σ1   + 2γ + σ22 + $ v2H + g2H .
2
dx H $

Since  dx H ≥ v2H (see Exercise 7.9), we have

dv 2

  1
L Λ(v) ≤ −2α 2 + σ12 + 2γ + σ 22 + $ v2H + g2H .
$
7.5 Ultimate Boundedness and Weak Recurrence of the Solutions 283

Hence, if −2α 2 + σ12 + 2γ + σ22 < 0, then the strong variational solution of (7.53) is
exponentially ultimately bounded by Theorem 7.5, and hence it is weakly positive
recurrent.

Exercise 7.9 Let f ∈ W 0,2 ((a, b)). Prove the Poincaré inequality
b b  df (x) 2
f 2 (x) dx ≤ (b − a)2 dx.
a a dx
References

1. S. Agmon. Lectures on Elliptic Boundary Value Problems, Mathematical Studies No. 2, Van
Nostrand, Princeton (1965).
2. S. Albeverio and R. Hoegh-Krohn. Homogeneous random fields and statistical mechanics,
J. Funct. Anal. 19, 242–272 (1975).
3. S. Albeverio, Yu. G. Kondratiev, M. Röckner, and T. V. Tsikalenko. Glauber dynamics for
quantum lattice systems, Rev. Math. Phys. 13 No. 1, 51–124 (2001).
4. P. Billingsley. Convergence of Probability Measures, Wiley, New York (1968).
5. P. Billingsley. Probability and Measure, Wiley, New York (1979).
6. P.L. Butzer and H. Berens. Semi-Groups of Operators and Approximation, Springer, New
York (1967).
7. J. R. Cannon. The One-Dimensional Heat Equation, Encyclopedia of Mathematics and Its
Applications 23, Addison–Wesley, Reading (1984).
8. S. Cerrai. Second Order PDE’s in Finite and Infinite Dimension, Lecture Notes in Mathematics
1762, Springer, Berlin (2001).
9. A. Chojnowska-Michalik. Stochastic Differential Equations in Hilbert Space, Banach Center
Publications 5, PWN–Polish Scientific Publishers, Warsaw (1979).
10. G. Da Prato, S. Kwapien, and J. Zabczyk. Regularity of solutions of linear stochastic equations
in Hilbert spaces, Stochastics 23, 1–23 (1987).
11. G. Da Prato and J. Zabczyk. Stochastic Equations in Infinite Dimensions, Encyclopedia of
Mathematics and its Applications 44, Cambridge University Press, Cambridge (1992).
12. G. Da Prato and J. Zabczyk. Ergodicity for Infinite Dimensional Systems, London Mathemat-
ical Society Lecture Note Series 229, Cambridge University Press, Cambridge (1996).
13. R. Datko. Extending a theorem of A. M. Liapunov to Hilbert space, J. Math. Anal. Appl. 32,
610–616 (1970).
14. J. Diestel and J.J. Uhl. Vector Measures, Mathematical Surveys 15, AMS, Providence (1977).
15. J. Dieudonné. Treatise on Analysis, Academic Press, New York (1969).
16. R. J. Elliott. Stochastic Calculus and Applications, Springer, New York (1982).
17. S. N. Ethier and T. G. Kurtz. Markov Processes: Characterization and Convergence. Wiley,
New York (1986).
18. B. Gaveau. Intégrale stochastique radonifiante, C.R. Acad. Sci. Paris Ser. A 276, 617–620
(1973).
19. L. Gawarecki. Extension of a stochastic integral with respect to cylindrical martingales, Stat.
Probab. Lett. 34, 103–111 (1997).
20. L. Gawarecki and V. Mandrekar. Stochastic differential equations with discontinuous drift in
Hilbert space with applications to interacting particle systems, J. Math. Sci. 105, No. 6, 2550–
2554 (2001). Proceedings of the Seminar on Stability Problems for Stochastic Models, Part I
(Naleczow, 1999).

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 285

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0, © Springer-Verlag Berlin Heidelberg 2011
286 References

21. L. Gawarecki and V. Mandrekar. Weak solutions to stochastic differential equations with dis-
continuous drift in Hilbert space. In: Stochastic Processes, Physics and Geometry; New In-
terplays, II (Leipzig, 1999), CMS Conf. Proc. 29, Amer. Math. Soc., Providence, 199–205
(2000).
22. L. Gawarecki, V. Mandrekar, and P. Richard. Existence of weak solutions for stochastic differ-
ential equations and martingale solutions for stochastic semilinear equations, Random Oper.
Stoch. Equ. 7, No. 3, 215–240 (1999).
23. L. Gawarecki, V. Mandrekar, and B. Rajeev. Linear stochastic differential equations in the
dual to a multi-Hilbertian space, Theory Stoch. Process. 14, No. 2, 28–34 (2008).
24. L. Gawarecki, V. Mandrekar, and B. Rajeev. The monotonicity inequality for linear stochastic
partial differential equations, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 12, No. 4,
1–17 (2009).
25. I.I. Gikhman and A.V. Skorokhod. The Theory of Stochastic Processes, Springer, Berlin
(1974).
26. A.N. Godunov. On Peano’s theorem in Banach spaces, Funct. Anal. Appl. 9, 53–55 (1975).
27. K. Gowrisankaran. Measurability of functions in product spaces, Proc. Am. Math. Soc. 31,
No. 2, 485–488 (1972).
28. M. Hairer. Ergodic properties of a class of non-Markovian processes. In: Trends in Stochastic
Analysis, London Mathematical Society Lecture Note Series 353. Ed. J. Blath et al. 65–98
(2009).
29. E. Hille. Lectures on Ordinary Differential Equations, Addison–Wesley, Reading (1969)
30. F. Hirsch and G. Lacombe. Elements of Functional Analysis, Graduate Texts in Mathematics
192, Springer, New York (1999).
31. H. Holden, B. Øksendal, J. Uboe, and T. Zhang. Stochastic Partial Differential Equations:
A Modeling, White Noise Functional Approach, Birkhauser, Boston (1996).
32. A. Ichikawa. Stability of semilinear stochastic evolution equations, J. Math. Anal. Appl. 90,
12–44 (1982).
33. A. Ichikawa, Semilinear stochastic evolution equations: boundedness, stability and invariant
measures, Stochastics 12, 1–39 (1984).
34. A. Ichikawa. Some inequalities for martingales and stochastic convolutions, Stoch. Anal. Appl.
4, 329–339 (1986).
35. K. Itô. Foundations of Stochastic Differential Equations in Infinite Dimensional Spaces,
CBMS–NSF 47 (1984).
36. G. Kallianpur, I. Mitoma, and R. L. Wolpert. Diffusion equations in dual of nuclear spaces,
Stoch. Stoch. Rep. 29, 285–329 (1990).
37. G. Kallianpur and J. Xiong. Stochastic differential equations in infinite dimensions: a brief
survey and some new directions of research. In: Multivariate Analysis: Future Directions,
North-Holland Ser. Statist. Probab. 5, North-Holland, Amsterdam, 267–277 (1993).
38. I. Karatzas and S. E. Shreve. Brownian Motion and Stochastic Calculus, Graduate Texts in
Mathematics, Springer, New York (1991).
39. R. Khasminskii. Stochastic Stability of Differential Equations, Sijthoff and Noordhoff, Alphen
aan den Rijn (1980).
40. R. Khasminskii and V. Mandrekar. On stability of solutions of stochastic evolution equations.
In: The Dynkin Festschrift, Progr. Probab. Ed. M. Freidlin, Birkhäuser, Boston, 185–197
(1994).
41. H. König. Eigenvalue Distribution of Compact Operators, Birkhäuser, Boston (1986).
42. N. V. Krylov and B. L. Rozovskii. Stochastic evolution equations, J. Sov. Math. 16, 1233–1277
(1981).
43. K. Kuratowski and C. Ryll-Nardzewki, A general theorem on selectors, Bull. Acad. Pol. Sci.
13, 349–403 (1965).
44. S. Lang. Analysis II, Addison–Wesley, Reading (1969).
45. M. Ledoux and M. Talagrand. Probability in Banach Spaces, Springer, Berlin (1991).
46. G. Leha and G. Ritter. On diffusion processes and their semigroups in Hilbert spaces with an
application to interacting stochastic systems, Ann. Probab. 12, No. 4, 1077–1112 (1984).
References 287

47. G. Leha and G. Ritter. On solutions to stochastic differential equations with discontinuous
drift in Hilbert space, Math. Ann. 270, 109–123 (1985).
48. J. L. Lions. Équations Différentielles Opérationelles et Problèmes aux Limites, Springer,
Berlin (1961).
49. R. S. Liptzer and A. N. Shiryaev. Statistics of Stochastic Processes, Nauka, Moscow (1974).
50. K. Liu. Stability of Infinite Dimensional Stochastic Differential Equations with Applica-
tions, Chapman & Hall/CRC Monographs and Surveys in Pure and Applied Mathematics
135 (2006).
51. R. Liu. Ultimate boundedness and weak recurrence of stochastic evolution equations, Stoch.
Anal. Appl. 17, 815–833 (1999).
52. R. Liu and V. Mandrekar. Ultimate boundedness and invariant measures of stochastic evolu-
tion equations, Stoch. Stoch. Rep. 56, No. 1–2, 75–101 (1996).
53. R. Liu and V. Mandrekar. Stochastic semilinear evolution equations: Lyapunov function, sta-
bility, and ultimate boundedness, J. Math. Anal. Appl. 212, No. 2, 537–553 (1997).
54. G.W. Mackey. A theorem of Stone and von Neuman, Duke Math. J. 16, No. 2, 313–326
(1949).
55. V. Mandrekar. On Lyapunov stability theorems for stochastic (deterministic) evolution equa-
tions. In: Stochastic Analysis and Applications in Physics, NATO Adv. Sci. Inst. Ser. C Math.
Phys. Sci. 449, Kluwer, Dordrecht, 219–237 (1994).
56. M. Metivier. Stochastic Partial Differential Equations in Infinite Dimensional Spaces, Scuola
Normale Superiore, Quaderni, Pisa (1988).
57. M. Metivier and J. Pellaumail. Stochastic Integration, Academic Press, New York (1980).
58. M. Metivier and M. Viot. On weak solutions of stochastic partial differential equations. In:
Stochastic Analysis, LNM 1322. Ed. M. Metivier, S. Watanabe, Springer, Berlin, 139–150
(1988).
59. Y. Miyahara. Ultimate boundedness of the system governed by stochastic differential equa-
tions, Nagoya Math. J. 47, 111–144 (1972).
60. E. Nelson. Probability Theory and Euclidian Field Theory, Lecture Notes in Phys. 25, Springer,
Berlin, 94–124 (1973).
61. B. Øksendal. Stochastic Differential Equations, Springer, New York (1998).
62. E. Pardoux. Stochastic partial differential equations and filtering of diffusion processes,
Stochastics 3, 127–167 (1979).
63. A. Pazy. Semigroups of Linear Operators and Applications to Partial Differential Equations,
Applied Mathematical Sciences, 44, Springer, New York (1983).
64. C. Prévôt and M. Röckner. A Concise Course on Stochastic Partial Differential Equations,
LNM 1905, Springer, Berlin (2007).
65. Yu. V. Prokhorov. Convergence of random processes and limit theorems in probability theory,
Theory Probab. Appl. 1, 157–214 (1956).
66. B. L. Rozovskii. Stochastic Evolution Systems: Linear Theory and Applications to Non-Linear
Filtering, Kluwer, Boston (1983).
67. M. Röckner, B. Schmuland, and X. Zhang. Yamada–Watanabe theorem for stochastic evolu-
tion equations in infinite dimensions, Condens. Matter Phys. 11, No. 2(54), 247–259 (2008).
68. R. S. Schatten. Norm Ideals of Continuous Operators, Springer, New York (1970).
69. A.V. Skorokhod. Personal communication.
70. D.W. Stroock and S.R.S. Varadhan. Multidimensional Diffusion Processes, Springer, New
York (1979).
71. H. Tanabe. Equations of Evolution, Pitman, London (1979).
72. L. Tubaro. An estimate of Burkholder type for stochastic processes defined by the stochastic
integral, Stoch. Anal. Appl. 62, 187–192 (1984).
73. R. Wheeden, A. Zygmund. Measure and Integral, Marcel Dekker, New York (1977).
74. N. N. Vakhania, V. I. Tarieladze, and S. A. Chobanyan. Probability Distributions on Banach
Spaces, Mathematics and Its Applications (Soviet Series) 14, Reidel, Dordrecht (1987).
75. M. Viot. Solutions faibles d’équations aux dérivées partielles non linéaires, Thése, Université
Pierre et Marie Curie, Paris (1976).
288 References

76. M. Vishik and A. Fursikov. Mathematical Problems of Statistical Hydrodynamics, Kluwer,

London (1988).
77. M. Yor. Existence et unicité de diffusions à valeurs dans un espace de Hilbert, Ann. Inst. H.
Poincaré B 10, 55–88 (1974).
78. K. Yosida. Functional Analysis, Springer, New York (1980).
79. J. Zabczyk. Linear stochastic systems in Hilbert spaces; spectral properties and limit be-
haviour, Banach Cent. Publ. 41, 591–609 (1985).
Index

A H
Abstract Cauchy problem, 4, 11 Heat equation, 3
in Rd , 13
B one–dimensional, 3
Burkholder inequality, 87 stochastic, 248
Hille–Yosida theorem, 8
C
Chapman–Kolmogorov equation, 256 I
Covariance Increasing process, 22
of a Gaussian random variable, 18 of Q–Wiener process, 23
of Gaussian stochastic convolution, 259 Infinitesimal generator, 6
Cylindrical Wiener process, 19 Invariant probability measure, 256
the linear case, 260
D Itô’s formula
Datko’s theorem the case of cylindrical Wiener process, 69
stochastic analogue of, 230 the case of Q–Wiener process, 61
Dissipative system, 237 the variational case, 226
Dissipativity condition, 237
Doob’s Maximal Inequalities, 22
K
Kolmogorov’s backward equation, 118
E
Eigenvalues, 18
Eigenvectors, 18 L
Energy equality, 208 Lebesgue’s DCT
generalized, 100
F Lévy’s theorem, 51
Factorization formula, 89 Linear equation with additive noise, 259
Factorization technique, 87 Lions’s lemma, 15
Feller property, 110 Lions’s theorem, 15
Lyapunov function, 206, 216
G
Gaussian measure with covariance Q, 19 M
Gaussian random variable, 17 Markov
cylindrical standard, 17 process, 107
with covariance Q, 18 property, 107
Gaussian semigroup, 4 transition function, 256
Gibbs measure, 194 transition probabilities, 255

L. Gawarecki, V. Mandrekar, Stochastic Differential Equations in Infinite Dimensions, 289

Probability and Its Applications,
DOI 10.1007/978-3-642-16194-0, © Springer-Verlag Berlin Heidelberg 2011
290 Index

Martingale, 21 S
square integrable, 22 Semigroup, 5
Martingale Representation Theorem I, 51 adjoint, 7
Martingale Representation Theorem II, 53 C0 , 5
Maxwell probability measure, 260 compact, 6
Monotonicity, 173 differentiable, 6
weak, 173 exponentially stable, 260
Feller, 110
N of contractions, 6
Navier-Stokes equation, 258 pseudo–contraction, 6
Norm semigroup property, 5
strong continuity property, 5
graph, 7
strongly continuous, 5
Hilbert–Schmidt, 24
uniformly bounded, 6
trace, 18
uniformly continuous, 6
Sequence
O stochastically bounded, 129
Operator Sobolev space, 11
adjoint, 5 Solution, 74
coercive, 246 classical, 11
Hilbert–Schmidt, 24 martingale, 75
non-negative definite, 5 mild, 12, 75
symmetric, 5 mild integral, 74
trace–class, 18 strong, 74
weak, 75
P weakened, 74
Poincaré inequality, 283 Spin system, 193
Probability space, 17 lattice, 194
complete, 17 Stability, 203
filtered, 19 deterministic condition for, 204
Process, 19 exponential in the m.s.s., 213
elementary, 25 of the mild solution, 213
bounded, 25 of the variational solution, 227
Feller, 110 exponential stability of the solution of the
Gaussian, 19 Cauchy problem, 203
Ornstein–Uhlenbeck, 259 in probability of the zero solution, 223
progressively measurable, 38 of a solution to a linear equation with
weakly positive recurrent, 278 additive noise, 259
Prokhorov’s theorem, 148 Stochastic convolution, 76
Gaussian, 259
Stochastic differential equation
Q of evolution type, 247
Q–Wiener process, 20 semilinear, 73
continuity, 20 variational, 152
properties, 21 Stochastic Fubini Theorem, 57
Quadratic variation process, 22 Stochastic integral
existence and uniqueness, 22 with respect to a Q–Wiener process
of Q–Wiener process, 23 of a process in P (KQ , H ), 42
of a process in 2 (KQ , H ), 34
R with respect to a standard cylindrical
Reaction–Diffusion Equation, 282 Wiener process, 45
Recurrent region, 274 Stochastic process, 260
Resolvent, 8 ultimately bounded in the m.s.s., 260
set, 8 weakly recurrent, 274
Index 291

Strong Markov process, 183 pathwise, 173

Strong Markov property, 183 Usual conditions, 19

T V
Theorem Variational method, 15
about convolution, 94 Variational solution, 152
Tight family of measures, 148 strong, 152
Trace, 18 weak, 152

U Y
Uniqueness, 173 Yosida approximation, 9