Sunteți pe pagina 1din 593

Jan A.

Van Casteren

Markov processes, Feller


semigroups and evolution
equations

– Monograph –

April 29, 2008

Springer
The author wants to dedicate this book to his mathematics
teacher Rudi Hirschfeld at the occasion of his 80th birthday.
Preface

Writing the present book has been a long time project which emerged more
than five years ago. One of the main sources of inspiration was a mini-course
which the author taught at Monopoli (University of Bari, Italy). This course
was based on the text in [241]. The main theorem of the present book (The-
orem 1.39), but phrased in the locally compact setting, was a substantial
part of that course. The title of the conference was International Summer
School on Operator Methods for Evolution Equations and Approximation
Problems, Monopoli (Bari), September 15–22, 2002. The mini-course was en-
titled “Markov processes and Feller semigroups”. Other papers which can be
considered as predecessors of the present book are [238, 240, 247, 248]. In
this book a Polish state space replaces the locally compact state space in the
more classical literature on the subject. A Polish space is separable and com-
plete metrizable. Important examples of such spaces are separable Banach and
Frechet spaces. The generators of the Markov processes or diffusions which
play a central role in the present book could be associated with stochastic
differential equations in a Banach space. In the formulation of our results we
avoid the use of the metric which turns the state space into a complete metriz-
able space; see e.g. the Propositions 3.23 and 8.8. As a rule of thumb we phrase
results in terms of (open) subsets rather than using a metric. As one of the
highlights of the book we mention Theorem 1.39 and everything surrounding
it. This theorem gives an important relationship between the following con-
cepts: probability transition functions with the (strong) Feller property, strong
Markov processes, martingale problems, generators of Markov processes, and
uniqueness of Markov extensions. In this approach the classical uniform topol-
ogy is replaced by the so-called strict topology. A sequence of bounded contin-
uous functions converges for the strict topology if it is uniformly bounded, and
if it converges uniformly on compact subsets. It can be described by means of
a certain family of semi-norms which turns the space of bounded continuous
functions into a sequentially complete locally convex separable vector space.
Its topological dual consists of genuine complex measures on the state space.
This is the main reason that the whole machinery works. The second chapter
VIII Preface

contains the proofs of the main theorem. The original proof for the locally
compact case, as exhibited in e.g. [34], cannot just be copied. Since we deal
with a relatively large state space every single step has to be reproved. Many
results are based on Proposition 2.2 which ensures that the orbits of our pro-
cess have the right compactness properties. If we talk about equi-continuity,
then we mean equi-continuity relative to the strict topology: see e.g. Theorem
1.7, Definition 1.16, Theorem 1.18, Corollary 1.19, Proposition 2.4, Corollary
2.4, Corollary 2.10, equation (3.114). In §3.4 a general criterion is given in or-
der that the sample paths of the Markov process are almost-surely continuous.
In addition this section contains a number of results pertaining to dissipativ-
ity properties of its generator: see e.g. Proposition 3.11. A discussion of the
maximum principle is found here: see e.g. Lemma 3.22 and Proposition 3.23.
In Section 3.3 we discuss Korovkin properties of generators. This notion is
closely related to the range property of a generator. In Section 3.5 we discuss
(measurability) properties of hitting times. In Chapters 4 and 5 we discuss
backward stochastic differential equations for diffusion processes. A highlight
in Chapter 4 is a new way to prove the existence of solutions. It is based on
a homotopy argument as explained in Theorem 1 (page 87) in Crouzeix et
al [63]: see Proposition 4.36, Corollary 4.37 and Remark 4.38. A martingale
which plays an important role in Chapter 5 is depicted in formula (5.2). A
basic result is Theorem 5.1. In Chapter 6 we discuss for a time-homogeneous
process a version of the Hamilton-Jacobi-Bellmann equation. Interesting the-
orems are the Noether theorems 6.13 and 6.17. In Chapters 7. 8, and 9 the
long time behavior of a recurrent time-homogeneous Markov process is inves-
tigated. Chapter 8 is analytic in nature; it is inspired by the Ph.-D. thesis
of Katilova [129]. Chapter 7 describes a coupling technique from Chen and
Wang [55]: see Theorem 8.3 and Corollary 8.4. The problem raised by Chen
and Wang (see §8.3) about the boundedness of the diffusion matrix can be
partially solved by using a Γ2 -condition instead of condition (8.5) in Theorem
8.3 without violating the conclusion in (8.6): see Theorem 8.71 and Example
8.77, Proposition 8.79 and the formulas (8.247) and (8.248). For more details
see Remark 8.41 and inequality (8.149) in Remark 8.53. Furthermore Chapter
8 contains a number of results related to the existence of an invariant σ-
additive measure for our recurrent Markov process. For example in Theorem
8.8. Conditions are given in order that there exist compact recurrent sub-
sets. This property has far-reaching consequences: see e.g. Proposition 8.16,
Theorem 8.18, and Proposition 8.24. Results about uniqueness of invariant
measures are obtained: see Corollary 8.35. The results about recurrent sub-
sets and invariant measures are due to Seidler [207]. Poincarë type inequalities
are proved: see the propositions 8.55 and 8.73, and Theorem 8.18. The results
on the Γ2 -condition are taken from Bakry [16, 17], and Ledoux [144]. In Chap-
ter 9 we collect some properties of relevant martingales. In addition, we prove
the existence and uniqueness of an irreducible invariant measure: see Theorem
9.12 and the results in §9.3. In Theorem 9.25 we follow Kaspi and Mandelbaum
[127] to give a precise relationship between Harris recurrence and recurrence
Preface IX

phrased in terms of hitting times. Theorem 9.36 is the most important one
for readers interested in an existence proof of a σ-additive invariant measure
which is unique up to a multiplicative constant. Assertion (e) of Proposition
9.40 together with Orey’s theorem for Markov chains (see Theorem 9.4) yields
the interesting consequence that, up to multiplicative constants, σ-finite in-
variant measures are unique. In §9.4 Orey’s theorem is proved for recurrent
Markov chains. In the proof we use a version of the bivariate linked forward
recurrence time chain as explained in Lemma 9.50. We also use Nummelin’s
splitting technique: see [162], §5.1 (and §17.3.1). The proof of Orey’s theo-
rem is based on Theorems 9.53 and 9.62. Results Chapter 9 go back to Meyn
and Tweedie [162] for time-homogeneous Markov chains and Seidler [207] for
time-homogeneous Markov processes.

Interdependence

From the above discussion it is clear how the chapters in this book are related.
Chapter 1 is a prerequisite for all the others except Chapter 7. Chapter 2
contains the proofs of the main results in Chapter 1; it can be skipped at a
first reading. Chapter 3 contains material very much related to the contents
of the first chapter. Chapter 5 is a direct continuation of 4, and is somewhat
difficult to read and comprehend without the knowledge of the contents of
Chapter 4. Chapter 6 is more or less independent of the other chapters in
Part 2. For a big part Chapter 7 is independent of the other chapters: most of
the results are phrased and proved for a finite-dimensional state space. The
chapters 8 and 9 are very much interrelated. Some results in Chapter 8 are
based on results in Chapter 9. In particular this is true in those results which
use the existence of an invariant measure. A complete proof of existence and
uniqueness is given in Chapter 9 Theorem 9.36. As a general prerequisite for
understanding and appreciating this book a thorough knowledge of probability
theory, in particular the concept of the Markov property, combined with a
comprehensive notion of functional analysis is very helpful. On the other hand
most topics are explained from scratch.

Acknowledgement

Part of this work was presented at a Colloquium at the University of Gent,


October 14, 2005, at the occasion of the 65th birthday of Richard Delanghe
and appeared in a very preliminary form in [244]. Some results were also pre-
sented at the University of Clausthal, at the occasion of Michael Demuth’s
60th birthday September 10–11, 2006, and at a Conference in Marrakesh, Mo-
rocco, “Marrakesh World Conference on Differential Equations and Applica-
tions”, June 15–20, 2006. Some of this work was also presented at a Conference
on “The Feynman Integral and Related Topics in Mathematics and Physics:
In Honor of the 65th Birthdays of Gerry Johnson and David Skoug”, Lincoln,
X Preface

Nebraska, May 12–14, 2006. Finally, another preliminary version was pre-
sented during a Conference on Evolution Equations, in memory of G. Lumer,
at the Universities of Mons and Valenciennes, August 28–September 1, 2006.
The author also has presented some of this material during a colloquium at
the University of Amsterdam (December 21, 2007), and at the AMS Special
Session on the Feynman Integral in Mathematics and Physics, II, on January
9, 2008, in the Convention Center in San Diego, CA.
The author is obliged to the University of Antwerp (UA) and FWO Flan-
ders (Grant number 1.5051.04N) for their financial and material support. He
was also very fortunate to have discussed part of this material with Karel
in’t Hout (University of Antwerp), who provided some references with a cru-
cial result about a surjectivity property of one-sided Lipschitz mappings: see
Theorem 1 in Croezeix et al [63]. Some aspects concerning this work, like
backward stochastic differential equations, were at issue during a conservation
with Étienne Pardoux (CMI, Université de Provence, Marseille); the author is
grateful for his comments and advice. The author is indebted to J.-C. Zambrini
(Lisboa) for interesting discussions on the subject and for some references. In
addition, the information and explanation given by Willem Stannat (Technical
University Darmstadt) while he visited Antwerp are gratefully acknowledged.
In particular this is true for topics related to asymptotic stability: see Chap-
ter 8. The author is very much obliged to Natalia Katilova who has given
the ideas of Chapter 7; she is to be considered as a co-author of this chapter.
Finally, this work was part of the ESF program “Global”.

Key words and phrases, subject classification

Some key words and phrases are: backward stochastic differential equation,
parabolic equations of second order, Markov processes, Markov chains, ergod-
icity conditions, Orey’s theorem, theorem of Chacon-Ornstein, invariant mea-
sure, Korovkin properties, maximum principle, Kolmogorov operator, squared
gradient operator, martingale theory.
AMS Subject classification [2000]: 60H99, 35K20, 46E10, 60G46, 60J25.

Antwerp, Jan A. Van Casteren


February 2008
Contents

Part I Strong Markov processes

1 Strong Markov processes on polish spaces . . . . . . . . . . . . . . . . . 3


1.1 Strict topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Theorem of Daniell-Stone . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.2 Measures on polish spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.3 Integral operators on the space of bounded continuous
functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Strong Markov processes and Feller evolutions . . . . . . . . . . . . . . 27
1.2.1 Generators of Markov processes and maximum principles 31
1.3 Strong Markov processes: main result . . . . . . . . . . . . . . . . . . . . . . 35
1.3.1 Some historical remarks and references . . . . . . . . . . . . . . . 40
1.4 Dini’s lemma, Scheffé’s theorem, and the monotone class
theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.4.1 Dini’s lemma and Scheffé’s theorem . . . . . . . . . . . . . . . . . . 42
1.4.2 Monotone class theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2 Strong Markov processes: proof of main result . . . . . . . . . . . . . 47


2.1 Proof of the main result: Theorem 1.39 . . . . . . . . . . . . . . . . . . . . . 47
2.1.1 Proof of item (a) of Theorem 1.39 . . . . . . . . . . . . . . . . . . . 47
2.1.2 Proof of item (b) of Theorem 1.39 . . . . . . . . . . . . . . . . . . . 69
2.1.3 Proof of item (c) of Theorem 1.39 . . . . . . . . . . . . . . . . . . . 72
2.1.4 Proof of item (d) of Theorem 1.39 . . . . . . . . . . . . . . . . . . . 76
2.1.5 Proof of item (e) of Theorem 1.39 . . . . . . . . . . . . . . . . . . . 94
2.1.6 Some historical remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

3 Space-time operators and miscellaneous topics . . . . . . . . . . . . . 99


3.1 Space-time operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.2 Dissipative operators and maximum principle . . . . . . . . . . . . . . . 111
3.3 Korovkin property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.4 Continuous sample paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
XII Contents

3.5 Measurability properties of hitting times . . . . . . . . . . . . . . . . . . . 150


3.5.1 Some side remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
3.5.2 Some related remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

Part II Backward Stochastic Differential Equations

4 Feynman-Kac formulas, backward stochastic differential


equations and Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.2 A probabilistic approach: weak solutions . . . . . . . . . . . . . . . . . . . . 191
4.3 Existence and Uniqueness of solutions to BSDE’s . . . . . . . . . . . . 194
4.4 Backward stochastic differential equations and Markov
processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

5 Viscosity solutions, backward stochastic differential


equations and Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
5.1 Comparison theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
5.2 Viscosity solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
5.3 Backward stochastic differential equations in finance . . . . . . . . . 248
5.4 Some related remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

6 The Hamilton-Jacobi-Bellman equation and the stochastic


Noether theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6.2 The Hamilton-Jacobi-Bellman equation and its solution . . . . . . 258
6.3 The Hamilton-Jacobi-Bellman equation and viscosity solutions 267
6.4 A stochastic Noether theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
6.4.1 Classical Noether theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 289
6.4.2 Some problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290

Part III Long time behavior

7 On non-stationary Markov processes and Dunford


projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.2 Kolmogorov operators and weak∗ -continuous semigroups . . . . . 296
7.3 Kolmogorov operators and analytic semigroups . . . . . . . . . . . . . . 301
7.3.1 Ornstein-Uhlenbeck process . . . . . . . . . . . . . . . . . . . . . . . . . 318
7.4 Ergodicity in the non-stationary case . . . . . . . . . . . . . . . . . . . . . . 357
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Contents XIII

8 Coupling methods and Sobolev type inequalities . . . . . . . . . . . 383


8.1 Coupling methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
8.2 Some related stability results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
8.3 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457

9 Miscellaneous topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459


9.1 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
9.2 Stopping times and time-homogeneous Markov processes . . . . . 463
9.3 Markov Chains: invariant measure . . . . . . . . . . . . . . . . . . . . . . . . . 465
9.3.1 Some definitions and results . . . . . . . . . . . . . . . . . . . . . . . . 465
9.3.2 Construction of an invariant measure . . . . . . . . . . . . . . . . 479
9.4 A proof of Orey’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
9.5 About invariant (or stationary) measures . . . . . . . . . . . . . . . . . . . 552
9.6 Weak and strong solutions to stochastic differential equations . 553

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Part I

Strong Markov processes


1
Strong Markov processes on polish spaces

1.1 Strict topology

Throughout this book E stands for a complete metrizable separable topo-


logical space, i.e. E is a polish space. The Borel field of E is denoted by
E. We write Cb (E) for the space of all complex valued bounded continu-
ous functions on E. The space Cb (E) is equipped with the supremum norm:
kf k∞ = supx∈E |f (x)|, f ∈ Cb (E). The space Cb (E) will be endowed with
a second topology which will be used to describe the continuity properties.
This second topology, which is called the strict topology, is denoted as Tβ -
topology. The strict topology is generated by the semi-norms of the form pu ,
where u varies over H(E), and where pu (f ) = supx∈E |u(x)f (x)| = kuf k∞ ,
f ∈ Cb (E). Here a function u belongs to H(E) if u is bounded and if for every
real number α > 0 the set {|u| ≥ α} = {x ∈ E : |u(x)| ≥ α} is contained in
a compact subset of E. It is noticed that Buck [44] was the first author who
introduced the notion of strict topology (in the locally compact setting). He
used the notation β instead of Tβ .
Remark 1.1. Let H + (E) be the collection of those functions u ∈ H(E) with
the following properties: u ≥ 0 and for every α > 0 the set {u ≥ α} is a
compact subset of E. Then every function u ∈ H + (E) is bounded, and the
strict topology is also generated by semi-norms of the form {pu : u ∈ H + (E)}.
Every u ∈ H + (E) attains its supremum at some point x ∈ E. Moreover
a sequence (fn )n∈N converges to a function f ∈ Cb (E) if and only if it is
uniformly bounded and if for every compact subset K of E the equality
limm→∞ supn≥m supx∈K |fn (x) − fm (x)| = 0 holds. Since Tβ -convergent se-
quences are Tβ -bounded, from Proposition 1.3 below it follows that a Tβ -
convergent sequence is uniformly bounded. The same conclusion is true for
Tβ -Cauchy sequences. Moreover, a Tβ -Cauchy sequence (fn )n∈N converges to a
bounded function f . Such a sequence converges uniformly on compact subsets
of the space E. Since the space E is polish, it follows that the limit function
f is continuous. Consequently, the space (Cb (E), Tβ ) is sequentially complete.
4 1 Strong Markov processes

Observe that continuity properties of functions f ∈ Cb (E) can be formulated


in terms of convergent sequences in E which are contained in compact subsets
of E. The topology of uniform convergence on Cb (E) is denoted by Tu .

1.1.1 Theorem of Daniell-Stone

In Proposition 1.3 below we need the following theorem. It says that an


abstract integral is a concrete integral. Theorem 1.2 will be applied with
S = E, H = Cb+ , the collection of non-negative functions in Cb (E), and
+
for I : CB → [0, ∞) we take the restriction to Cb+ of a non-negative linear
functional defined on Cb (E) which is continuous with respect to the strict
topology.
Theorem 1.2 (Theorem of Daniell-Stone). Let S be any set, and let H
be a non-empty collection of functions on S with the following properties:
(1) If f and g belong to H, then the functions f + g, f ∨ g and f ∧ g belong
to H as well;
(2) If f ∈ H and α is a non-negative real number, then αf , f ∧ α, and
+
(f − α) = (f − α) ∨ 0 belong to H;
(3) If f , g ∈ H are such that f ≤ g ≤ 1, then g − f belongs to H.
Let I : H → [0, ∞] be an abstract integral in the sense that I is a mapping
which possesses the following properties:
(4) If f and g belong to H, then I (f + g) = I(f ) + I(g);
(5) If f ∈ H and α ≥ 0, then I (αf ) = αI(f );
(6) If (fn )n∈N is a sequence in H which increases pointwise to f ∈ H, then
I (fn ) increases to I (f ).
Then there exists a non-negative σ-additive measureRµ on the σ-field generated
by H, which is denoted by σ(H), such that I(f ) = f dµ, for f ∈ H. If there
exists a countable family of functions
S∞ (fn )n∈N ⊂ H such that I (fn ) < ∞ for
all n ∈ N, and such that S = n=1 {fn > 0}, then the measure µ is unique.

Proof. Define the collection H ∗ of functions on S as follows. A function f :


S → [0, ∞] belongs to H ∗ provided there exists a sequence (fn )n∈N ⊂ H
which increases pointwise to f . Then the subset H ∗ has the properties (1)
and (2) with H ∗ instead of H. Define the mapping I ∗ : H ∗ → [0, ∞] by

I ∗ (f ) = lim I (fn ) , f ∈ H ∗ ,
n→∞

where (fn )n∈N ⊂ H is a sequence which pointwise increases to f . The defini-


tion does not depend on the choice of the increasing sequence (fn )n∈N ⊂ H.
In fact let (fn )n∈N and (gn )n∈N be sequences in H which both increase to
f ∈ H ∗ . Then by (6) we have

lim I (fn ) = sup I (fn ) = sup sup I (fn ∧ gm ) = sup sup I (fn ∧ gm )
n→∞ n∈N n∈N m∈N m∈N n∈N
1.1 Strict topology 5

= sup I (gm ) = lim I (gm ) . (1.1)


m∈N m→∞

From (1.1) it follows that I ∗ is well-defined. The functional I ∗ : H ∗ → [0, ∞]


has the properties (4), (5), and (6) (somewhat modified) with H ∗ instead of
H and I replaced by I ∗ . In fact the correct version of (6) for H ∗ reads as
follows:
(6∗ ) Let (fn )n∈N be a sequence H ∗ which increases pointwise to a function
f . Then f ∈ H ∗ , and I ∗ (fn ) increases to I ∗ (f ).
We also have the following assertion:
(3∗ ) Let f and g ∈ H ∗ be such that f ≤ g. Then I ∗ (f ) ≤ I ∗ (g).
We first prove (3∗ ) if f and g belong to H and f ≤ g. From (6), (3) and (4)
we get

I(g) = sup I (g ∧ m) = sup (I (g ∧ m − f ∧ m) + I (f ∧ m))


m∈N m∈N
≥ sup I (f ∧ m) = I (f ) . (1.2)
m∈N

Here we used the fact that by (3) the functions g ∧ m − f ∧ m, m ∈ N,


belong to H. Next let f and g be functions in H such that f ≤ g. Then there
exist increasing sequences (fn )n∈N and (gn )n∈N in H such that fn converges
pointwise to f ∈ H ∗ and gn to g ∈ H ∗ . Then

I ∗ (f ) = sup I (fn ) ≤ sup I (fn ∨ gn ) = I ∗ (g) . (1.3)


n∈N n∈N

Next we prove (6∗ ). Let (fn )n∈N be a pointwise increasing sequence in H ∗ ,


and put f = supn∈N fn . Choose for every n ∈ N an increasing sequence
(fn,m )m∈N ⊂ H such that supm∈N fn,m = fn . Define the functions gm , m ∈ N,
by
gm = f1,m ∨ f2,m ∨ · · · ∨ fm,m .
Then gm+1 ≥ gm and gm ∈ H for all m ∈ N. In addition, we have

sup gm = sup max fn,m = sup sup fn,m = sup fn = f. (1.4)


m∈N m∈N 1≤n≤m n∈N m≥n n∈N

Hence f ∈ H ∗ . For 1 ≤ n ≤ m the inequalities fn,m ≤ fn ≤ fm hold pointwise,


and hence gm ≤ fm . From (3∗ ) we infer

I ∗ (f ) = sup I (gm ) = sup I ∗ (gm ) ≤ sup I ∗ (fm ) ≤ I ∗ (f ) , (1.5)


m∈N m∈N m∈N

and thus supm∈N I ∗ (fm ) = I ∗ (f ).


Next we will get closer to measure theory. Therefore we define the col-
lection G of subsets of S by G = {G ⊂ S : 1G ∈ H ∗ }, and the mapping
µ : G → [0, ∞] by µ (G) = I ∗ (1G ), G ∈ G. The mapping µ possesses the
following properties:
6 1 Strong Markov processes

(10 ) If the subsets G1 and G2 belong to G, then the same is true for the
subsets G1 ∩ G2 and G1 ∪ G2 ;
(20 ) ∅ ∈ G;
(30 ) If the subsets G1 and G2 belong to G and if G1 ⊂ G2 , then µ (G1 ) ≤
µ (G2 );
(40 ) If the subsets G1 and G2 belong to G, then the following strong additivity
holds: µ (G1 ∩ G2 ) + µ (G1 ∪ G2 ) = µ (G1 ) + µ (G2 );
(50 ) µ (∅) = 0;
(60 ) If (Gn )n∈N is a sequence ¡in G such¢ that Gn+1 ⊃ Gn , n ∈ N, then
S S
n∈N Gn belongs to G and µ n∈N Gn = supn∈N µ (Gn ).

These properties are more or less direct consequences of the corresponding


properties of I ∗ : (1∗ )–(6∗ ).
Using the mapping µ we will define an exterior or outer measure µ∗
on the collection of all subsets of S. Let A be any subset of S. Then
we put µ∗ (A) = ∞ if for no G ∈ G we have A ⊂ G, and we write
µ∗ (A) = inf {µ (G) : G ∈ G, G ⊃ A}, if A ⊂ G0 for some G0 ∈ G. Then
µ∗ has the following properties:
(i) µ∗ (∅) = 0;
(ii) µ∗ (A) ≥ 0, for all subsets A of S;
(iii)µ∗ (A) ∗
S∞≤ µ (B),P whenever A and B are subsets of S for which A ⊂ B;

(iv)µ ( n=1 An ) ≤ n=1 µ∗ (An ) for any sequence (An )n∈N of subsets of S.

The assertions (i), (ii) and (iii) follow directly from the definition of µ∗ .
In order to prove (iv) we choose a sequence (An )n∈N , An ⊂ S, such that
µ∗ (An ) < ∞ for all n ∈ N. Fix ε > 0, and choose for every n ∈ N an subset
Gn of S which belongs to G and which has the S following properties:
S SmAn ⊂ Gn
and µ (GnS) ≤ µ∗ (An ) + ε2−n . By the equality n∈N Gn = m∈N n=1 Gn we
see that n∈N Gn belongs to G. From the properties of an exterior measure
we infer the following sequence of inequalities:
à ! à ! à ! Ãm !
[ [ [ [
∗ ∗
µ An ≤ µ Gn = µ Gn = sup µ Gn
m∈N n=1
n∈N n∈N n∈N
à m
! m
¡ ¢ X X
= sup I ∗ 1∪m
n=1 Gn
≤ sup I ∗ 1Gn = sup I ∗ (1Gn )
m∈N m∈N n=1 m∈N n=1
m
X ∞
X ∞
¡ ¢ X
= sup µ (Gn ) ≤ µ∗ (An ) + ε2−n = µ∗ (An ) + ε.
m∈N n=1 n=1 n=1
(1.6)
¡S ¢ P∞
Since ε > 0 was arbitrary we see that µ∗ n∈N An ≤

n=1 µ (An ). Hence
assertion (iv) follows.
Next we consider the σ-field D which is associated to the exterior measure
µ∗ , and which is defined by
1.1 Strict topology 7

D = {A ⊂ S : µ∗ (D) ≥ µ∗ (A ∩ D) + µ∗ (Ac ∩ D) for all D ⊂ S}


= {A ⊂ S : µ (D) ≥ µ∗ (A ∩ D) + µ∗ (Ac ∩ D)
for all D ∈ G with µ(D) < ∞} . (1.7)

Here we wrote Ac = S \ A for the complement of A in S. The reader is


invited to check the equality in (1.7). According to Caratheodory’s theorem
the exterior measure µ∗ restricted to the σ-field D is a σ-additive measure. We
will prove that D contains G. Therefore pick G ∈ G, and consider for D ∈ G
for which µ(D) < ∞ the equality

µ∗ (G ∩ D) + µ∗ (Gc ∩ D) = µ (G ∩ D) + inf {µ (U ) : U ∈ G, U ⊃ Gc ∩ D} .
(1.8)
Choose h ∈ H ∗ in such that h ≥ 1Gc ∩D . For 0 < α < 1 we have
1
1Gc ∩D ≤ 1{h>α} ≤ h.
α
Since 1{h>α} = supm∈N 1 ∧ (m(h − α)+ ) we see that the set {h > α} is a
member of G. It follows that T ∗ (h) ≥ αµ ({h > α}) ≥ αµ∗ (Gc ∩ D), and
hence

µ∗ (Gc ∩ D) ≤ inf {I ∗ (h) : h ≥ 1Gc ∩D , h ∈ H ∗ }


≤ inf {I ∗ (1U ) : U ⊃ Gc ∩ D, U ∈ G} = µ∗ (Gc ∩ D) . (1.9)

From (1.9) the equality

µ∗ (Gc ∩ D) = inf {I ∗ (h) : h ≥ 1Gc ∩D , h ∈ H ∗ }

follows. Next choose the increasing sequences (fn )n∈N and (gn )n∈N in such a
way that the sequence fn increases to 1D and gn increases to 1G . Define the
functions hn , n ∈ N, by

hn = 1D − fn ∧ gn = sup {(fm − fn ) + (fn − fn ∧ gn )} .


m≥n

Since the functions fm − fn , m ≥ n, and fn − fn ∧ gn belong to H we see that


hn belongs to H ∗ . Hence we get:

∞ > µ(D) = I ∗ (1D ) = I ∗ (hn ) + I ∗ (fn ∧ gn ) = I ∗ (hn ) + I (fn ∧ gn ) . (1.10)

In addition we have hn ≥ 1Gc ∩D . Consequently,

µ∗ (G ∩ D) + µ∗ (Gc ∩ D)
≤ µ (G ∩ D) + inf I ∗ (hn )
n∈N
= µ (G ∩ D) + µ (D) − sup I (fn ∧ gn )
n∈N
= µ (G ∩ D) + µ (D) − µ (G ∩ D) = µ (D) . (1.11)
8 1 Strong Markov processes

The equality in (1.11) proves that the σ-field D contains the collection G,
and hence that the mapping µ, which originally was defined on G in fact the
restriction is of a genuine measure defined on the σ-field generated by H,
which is again called µ, to G. R
We will show the equality I(f ) = f dµ for all f ∈ H. For f ∈ H we have
Z Z Z n2n
∞ ∞

¡ ¢ 1 X ∗¡ ¢
f dµ = µ {f > ξ} dξ = I 1{f >ξ} dξ = sup n I 1{f >j2−n }
0 0 n∈N 2 j=1
 
n2 n µ Z ¶
1 X ∞
= sup I ∗  n 1{f >j2−n }  = I ∗ x 7→ 1{f >ξ} (x)dξ
n∈N 2 j=1 0

= I ∗ (f ) = I (f ) . (1.12)
Finally we will prove the uniqueness of the measure Rµ. Let µ1 and
R µ2 be two
measures on σ(H) with the property that I(f ) = f dµ1 = f dµ2 for all
f ∈ H. Under the extra condition in Theorem 1.2 that there exist countable
manySfunctions (fn )n∈N such that I (fn ) < ∞ for all n ∈ N and such that

S = n=1 {fn > 0} we shall show that µ1 (B) = µ2 (B) for all B ∈ σ(H).
Therefore
© we Rfix a function
R f ∈ª H for which I(f ) < ∞. Then the collection
B ∈ σ(H) : B f dµ1 = B f dµ2 is a Dynkin system containing all sets of
the form {g > β} with g ∈³ H and β > 0. Fix ξ > 0, β > ´0 and g ∈ H. Then
+ +
the functions gm,n := min m (g − β) ∧ 1, n (f − ξ) ∧ 1 , m, n ∈ N, belong
to H. Then we have
Z
µ1 [{g > β} ∩ {f > ξ}] = lim lim gm,n dµ1 = lim lim I (gm,n )
m→∞ n→∞ m→∞ n→∞
Z
= lim lim gm,n dµ2 = µ2 [{g > β} ∩ {f > ξ}] . (1.13)
m→∞ n→∞

Upon integration the extreme terms


R in (1.13) with
R respect to the Lebesgue
measure dξ shows the equality {g>β} f dµ1 = {g>β} f dµ2 . It follows that
© R R ª
the collection B ∈ σ(H) : B f dµ1 = B f dµ2 contains all sets of the form
{g > β} where g ∈ H and β > 0. Such collection of sets is closed under finite
intersection. Hence, by a Dynkin argument, we infer the equality
½ Z Z ¾
B ∈ σ(H) : f dµ1 = f dµ2 = σ(H).
B B

The same argument applies with (nf ) ∧ 1 replacing f . By letting n tend to


∞ this shows the equality
σ(H) = {B ∈ σ(H) : µ1 [B ∩ {f > 0}] = µ2 [B ∩ {f > 0}]} . (1.14)
Since the set H is closed under taking finite maxima,
S∞ I (f ∨ g) ≤ I(f )+I(g) <
∞ whenever I(f ) and I(g) are finite, and S = n=1 {fn > 0} with I (fn ) < ∞,
n ∈ N, we see that
1.1 Strict topology 9
· ½ ¾¸
µ1 (B) = lim µ1 B ∩ max fj > 0
n→∞ 1≤j≤n
· ½ ¾¸
= lim µ2 B ∩ max fj > 0 = µ2 (B) (1.15)
n→∞ 1≤j≤n

for B ∈ σ(H).
This finishes the proof of Theorem 1.2.

1.1.2 Measures on polish spaces

Our first proposition says that the identity mapping f 7→ f sends Tβ -bounded
subsets of Cb (E) to k·k∞ -bounded subsets.
Proposition 1.3. Every Tβ -bounded subset of Cb (E) is k·k∞ -bounded. On
the other hand the identity is not a continuous operator from (Cb (E), Tβ ) to
(Cb (E), k·k∞ ), provided that E itself is not compact.

Proof. Let B ⊂ Cb (E) be Tβ -bounded. If B were not uniformly bounded, then


there exists sequences (fn )n∈N ⊂ B and (xn )n∈N ⊂ E such that |fn (xn )| ≥ n2 ,
X∞
1
n ∈ N. Put u(x) = 1x . Then the function u belongs to H(E), but
n=1
n n
sup pu (f ) ≥ sup pu (fn ) ≥ sup u (xn ) |f (xn )| ≥ sup n = ∞. The latter shows
f ∈B n∈N n∈N n∈N
that the set B is not Tβ -bounded. By contra-position it follows that Tβ -
bounded subsets are uniformly bounded.
Next suppose that E is not compact. Let u be any function in H(E). Then
limn→∞ u (xn ) = 0. If the imbedding (Cb (E), Tβ ) → (Cb (E), Tu ) were contin-
uous, then there would exist a function u ∈ H + (E) such that kf k∞ ≤ kuf k∞
for all f ∈ Cb (E). Let K be a compact subset of E such that 0 ≤ u(x) ≤ 21 for
x∈ / K. Since 1 ≤ kuk∞ = u (x0 ) for some x0 ∈ E, and since by assumption
E is not compact we see that K 6= E. Choose an open neighborhood O of
K, O 6= E, and a function f ∈ Cb (E) such that 1 − 1O ≤ f ≤ 1 − 1K .
In particular, it follows that f = 1 outside of O, and f = 0 on K. Then
1 1
1 = kf k∞ ≤ kuf k∞ ≤ supx∈K / |u(x)f (x)| ≤ 2 kf k∞ ≤ 2 . Clearly, this is a
contradiction.
This concludes the proof of Proposition 1.3.

The following proposition shows that the dual of the space (Cb (E), Tβ ) coin-
cides with the space of all complex Borel measures on E.
Proposition 1.4. 1. Let µ be a complex¯RBorel¯ measure on E. Then there
exists a function u ∈ H(E) such that ¯ f dµ¯ ≤ pu (f ) for all f ∈ Cb (E).
2. Let Λ : Cb (E) → C be a linear functional on Cb (E) which is continuous
with respect to the strict topology.R Then there exists a unique complex
measure µ on E such that Λ(f ) = f dµ, f ∈ Cb (E).
10 1 Strong Markov processes

Proof. 1 Since on a polish space every bounded Borel measure is inner regular,
there exists an increasing sequence of compact subsets (Kn )n∈N in E such that
|µ| (E \ Kn ) ≤ 2−2n−2 |µ| (E), n ∈ N. Fix f ∈ Cb (E). Then we have
¯Z ¯ X ¯ ¯
∞ ¯Z ¯ X ∞ Z
¯ ¯ ¯ ¯
¯ f dµ¯ ≤ ¯ f dµ¯ ≤ |f | d |µ|
¯ ¯ ¯ Kj+1 \Kj ¯
j=0 Kj+1 \Kj
j=0

X ° °
≤ °1K \K f ° |µ| (Kj+1 \ Kj )
j+1 j ∞
j=0
X∞
° °
≤ °1K \K f ° |µ| (E \ Kj )
j+1 j ∞
j=0

X ° °
≤ 2−2j−2 °1Kj+1 \Kj f °∞ |µ| (E)
j=0
X∞
≤ 2−2j−2 2j+1 kuf k∞ ≤ kuf k∞ (1.16)
j=0
P∞
where u(x) = j=1 2−j 1Kj (x) |µ| (E).
2 We decompose the functional Λ into a combination of four positive func-
+ − + −
tionals: Λ = (<Λ) − (<Λ) + i (=Λ) − i (=Λ) where the linear function-
+ −
als (<Λ) and (<Λ) are determined by their action on positive functions
f ∈ Cb (E):
+
(<Λ) (f ) = sup {< (Λ(g)) : 0 ≤ g ≤ f, g ∈ Cb (E)} , and

(<Λ) (f ) = sup {< (−Λ(g)) : 0 ≤ g ≤ f, g ∈ Cb (E)} .
+ −
Similar expressions cam be employed for the action of (=Λ) and (=Λ) on
functions f ∈ Cb+ . Since the complex linear functional Λ : Cb (E) → C is Tβ -
continuous there exists a function u ∈ H + (E)
¯ such that
¯ |Λ(f )| ≤ kuf k∞ for
¯ + ¯
all f ∈ Cb (E). Then it easily follows that ¯(<Λ) (f )¯ ≤ kuf k∞ for all real-
¯ ¯ √
¯ + ¯
valued functions in Cb (E), and ¯(<Λ) (f )¯ ≤ 2 kuf k∞ for all f ∈ Cb (E),

which in general take complex values. Similar inequalities hold for (<Λ) (f ),
+ −
(=Λ) (f ), and (=Λ) (f ). Let (fn )n∈N be a sequence of functions in Cb+ (E)
which pointwise increases to a function f ∈ Cb+ (E). Then limn→∞ Λ (fn ) =
Λ(f ). This can be seen as follows. Put gn = f − fn , and fix ε > 0. Then
the sequence (gn )n∈N decreases pointwise to 0. Moreover it is dominated by
f . Choose a strictly positive real number α in such a way that α kf k∞ ≤ ε.
Then it follows that
¡° ° ° ° ¢
|Λ (gn )| ≤ kugn k∞ = max °u1{u≥α} gn °∞ , °u1{u<α} gn °∞
¡ ° ° ¢
≤ max kuk °1{u≥α} gn ° , α kf k
∞ ∞
≤ε
∞ (1.17)
1.1 Strict topology 11
° °
where N chosen so large that kuk∞ °1{u≥α} gn °∞ ≤ ε for n ≥ N . By Dini’s
lemma such a choice of N is possible. An application of Theorem 1.2 then
yields the existence of measures µj , 1 ≤ j ≤ 4, defined on the Baire field of E
+ R − R + R
such that (<Λ) (f ) = f dµ1 , (<Λ) (f ) = f dµ2 , (=Λ) (f ) = f dµ3 , and
− R R R
(=Λ)
R (f ) =R f dµ4 forR f ∈ Cb (E). It follows that Λ(f ) = f dµ1 − f dµ2 +
i f dµ3 − i f dµ4 = f dµ for f ∈ Cb (E). Here µ = µ1 − µ2 + iµ3 − iµ4
and each measure µj , 1 ≤ j ≤ 4, is finite and positive. Since the space E is
polish it follows that Baire field coincides with the Borel field, and hence the
measure µ is a complex Borel measure.
This concludes the proof of Proposition 1.4.

The next corollary gives a sequential continuity characterization of linear



functionals which belong to the space (Cb (E), Tβ ) , the topological dual
of the space Cb (E) endowed with the strict topology. We say that a se-
quence (fn )n∈N ⊂ Cb (E) converges for the strict topology to f ∈ Cb (E) if
limn→∞ ku (f − fn )k∞ = 0 for all functions u ∈ H + (E). It follows that a
sequence (fn )n∈N ⊂ Cb (E) converges to a function f ∈ Cb (E) with respect
to the strict topology if and only if this sequence is uniformly bounded and
limn→∞ k1K (f − fn )k∞ = 0 for all compact subsets K of E.
Corollary 1.5. Let Λ : Cb (E) → C be a linear functional. Then the following
assertions are equivalent:

(1) The functional Λ belongs to (Cb (E), Tβ ) ;
(2) limn→∞ Λ (fn ) = 0 whenever (fn )n∈N is a sequence in Cb+ (E) which con-
verges to the zero-function for the strict topology;
(3) There exists a finite constant C ≥ 0 such that |Λ(f )| ≤ C kf k∞ for all
f ∈ Cb (E), and limn→∞ Λ (gn ) = 0 whenever (gn )n∈N is a sequence in
Cb+ (E) which is dominated by a sequence (fn )n∈N in Cb+ (E) which de-
creases pointwise to 0;
(4) There exists a finite constant C ≥ 0 such that |Λ(f )| ≤ C kf k∞ for all
f ∈ Cb (E), and limn→∞ Λ (fn ) = 0 whenever (fn )n∈N is a sequence in
Cb+ (E) which decreases pointwise to 0; R
(5) There exists a complex Borel measure µ on E such that Λ(f ) = f dµ for
all f ∈ Cb (E).
In (3) we say that a sequence (gn )n∈N in Cb+ (E) is dominated by a sequence
(fn )n∈N if gn ≤ fn for all n ∈ N.

Proof. (1) =⇒ (2). First suppose that Λ belongs to (Cb (E), Tβ ) . Then there
exists a function u ∈ H + (E) such that |Λ(f )| ≤ kuf k∞ for all f ∈ Cb (E).
Hence, if the sequence (fn )n∈N ⊂ Cb+ (E) converges to zero for the strict
topology, then limn→∞ kufn k∞ = 0, and so limn→∞ Λ (fn ) = 0. This proves
the implication (1) =⇒ (2).
(2) =⇒ (3). Let (fn )n∈N be a sequence in Cb (E) which converges
³ to´0 for
+
the uniform topology. From (2) it follows that the sequences (<fn ) ,
n∈N
12 1 Strong Markov processes
³ ´ ³ ´ ³ ´
− + −
(<fn ) , (=fn ) , and (=fn ) converge to 0 for the strict
n∈N n∈N n∈N
topology Tβ , and hence limn→∞ Λ (fn ) = 0. Consequently, the functional Λ :
Cb (E) → C is continuous if Cb (E) is equipped with the uniform topology,
and hence there exists a finite constant C ≥ 0 such that |Λ(f )| ≤ C kf k∞
for all f ∈ Cb (E). If (fn )n∈N is a sequence in Cb+ (E) which decreases to 0,
then by Dini’s lemma it converges uniformly on compact subsets of E to 0.
Moreover, it is uniformly bounded, and hence it converges to 0 for the strict
topology. If the sequence (gn )n∈N ⊂ Cb+ (E) is such that gn ≤ fn . Then the
sequence (gn )n∈N converges to 0 for the strict topology. Assertion (2) implies
that limn→∞ Λ (gn ) = 0.
(3) =⇒ (4). This implication is trivial.
(3) =⇒ (5). The boundedness of the functional Λ, i.e. the inequality
|Λ(f )| ≤ C kf k∞ , f ∈ Cb (E), enables us to write Λ in the form Λ =
+ − +
Λ1 −Λ2 +iΛ3 −iΛ4 in such a way that Λ1 = (<Λ) , Λ2 = (<Λ) , Λ3 = (=Λ) ,

and Λ3 = (=Λ) . From the definitions of these functionals (see the proof of
assertion 2 in Proposition 1.4) assertion (3) implies that limn→∞ Λj (fn ) = 0,
1 ≤ j ≤ 4, whenever the sequence (fn )n∈N ⊂ Cb+ (E) decreases to 0. From
Theorem 1.2 we infer that each functional
R Λj , 1 ≤ j ≤ 4, can be represented
by a Borel measure
R µj : Λ j (f ) = f dµ j , 1 ≤ j ≤ 4, f ∈ CB (E). It follows
that Λ(f ) = f dµ, f ∈ Cb (E), where µ = µ1 − µ2 + iµ3 − iµ4 .
(4) =⇒ (5). From the apparently weaker hypotheses in assertion (4) com-
pared to (3) we still have to prove that the functionals Λj , 1 ≤ j ≤ 4, as de-
scribed in the implication (3) =⇒ (5) have the property that limn→∞ Λj (fn ) =
0 whenever the sequence (fn )n∈N ⊂ Cb+ (E) decreases pointwise to 0. We
+
will give the details for the functional Λ1 = (<Λ) . This suffices because
+ + +
Λ2 = (< (−Λ)) , Λ3 = (< (−iΛ)) , and Λ4 = (< (iΛ)) . So let the sequence
+
(fn )n∈N ⊂ Cb (E) decreases pointwise to 0. Fix ε > 0, and choose 0 ≤ g1 ≤ f1 ,
g1 ∈ Cb (E), in such a way that

+ 1
Λ1 (f1 ) = (<Λ) (f1 ) ≤ < (Λ (g1 )) + ε. (1.18)
2
Then wePchoose sequence of functions (uk )k∈N ⊂ Cb+ (E) such that g1 =
aP
n ∞
supn∈N k=1 uk = k=1 uk (which is a pointwise increasing limit), and such
that uk ≤ fk − fk+1 , k ∈ N. In Lemma
Pn1.6 below we will show that such a
decomposition is possible. Then g1 − k=1 uk decreases pointwise to 0, and
hence by (4) we have
à n !
X 1
<Λ (g1 ) ≤ <Λ uk + ε, for n ≥ nε . (1.19)
2
k=1

From (1.18) and (1.19) we infer for n ≥ nε the inequality


1.1 Strict topology 13

+ 1
Λ1 (f1 ) = (<Λ) (f1 ) ≤ < (Λ (g1 )) + ε
à n ! 2
X Xn n
X +
≤ <Λ uk + ε = <Λ (uk ) + ε ≤ (<Λ) (fk − fk+1 ) + ε
k=1 k=1 k=1
n
X
= Λ1 (fk − fk+1 ) + ε = Λ1 (f1 ) − Λ1 (fn+1 ) + ε. (1.20)
k=1

From (1.20) we deduce Λ1 (fn ) ≤ ε for n ≥ nε + 1. Since ε > 0 was arbitrary,


this shows limn→∞ Λ1 (fn ) = 0. This is true for the other linear functionals
Λ2 , Λ3 and Λ4 as well. As in the proof of the implication (3) =⇒ (5) from
Theorem 1.2 it follows that eachRfunctional Λj , 1 ≤ j ≤ 4, can be represented
by a Borel
R measure µj : Λj (f ) = f dµj , 1 ≤ j ≤ 4, f ∈ Cb (E). It follows that
Λ(f ) = f dµ, f ∈ Cb (E), where µ = µ1 − µ2 + iµ3 − iµ4 .
(5) =⇒ (1). The proof of assertion 1 in Proposition 1.4 then shows that

the functional Λ belongs to (Cb (E), Tβ ) .

Lemma 1.6. Let the sequence (fn )n∈N ⊂ Cb+ (E) decrease pointwise to 0,
and 0 ≤ g ≤ f1 be a continuous function. Then there exists a sequence of
continuous functions (uPk )nk∈N such P
that 0 ≤ uk ≤ fk − fk+1 , k ∈ N, and

such that g = supn∈N k=1 uk = k=1 uk which is a pointwise monotone
increasing limit.
Pn
Proof. We write g = v1 = u1 + v2 = k=1 uk + vn+1 , and vn+1 = un+1 + vn+2
where u1 = g ∧ (f1 − f2 ), un+1 = vn+1 ∧ (fn+1 − fn+2 ), and vn+2 = vn+1 −
un+1 . Then 0 ≤ vn+1 ≤ vn ≤ fn . Since the sequence (fn )n∈N decreases
Pn to 0,
the sequence (vn )n∈N also decreases to 0, and thus g = supn∈N k=1 uk .
The latter shows Lemma 1.6.

In the sequel we write M(E) for the complex vector space of all complex Borel
measures on the polish space E. The space is supplied with the weak topology
σ (E, Cb (E)). We also write M+ (E) for the convex cone of all positive (= non-
negative) Borel measures in M(E). The notation M+ 1 (E) is employed for all
probability measures in M+ (E), and M+ ≤1 (E) stands for all sub-probability

measures in M+ (E). We identify the space M(E) and the space (Cb (E), Tβ ) .
Theorem 1.7. Let M be a subset of M(E) with the property that for every
sequence (Λn )n∈N in M there exists a subsequence (Λnk )k∈N such that
¡ ¢ ¡ ¢
lim sup < i` Λnk (f ) = sup < i` Λ(f ) , 0 ≤ ` ≤ 3,
k→∞ 0≤f ≤1 0≤f ≤1

for some Λ ∈ M(E). Then M is a relatively weakly compact subset of M(E)


if and only if it is equi-continuous viewed as a subset of the dual space of

(Cb (E), Tβ ) .
14 1 Strong Markov processes

Proof. First suppose that M(E) is relatively weakly compact. Since the weak
topology on M(E) restricted to compact subsets is metrizable and separa-
ble, the weak closure of M is bounded for the variation norm. Without loss
of generality we may and do assume that M itself is weakly compact. Fix
+
f ∈ Cb (E), f ≥ 0. Consider the mapping Λ 7→ (<Λ) (f ), Λ ∈ M(E). Here we

identify Λ = Λµ ∈ (Cb (E), Tβ ) and the corresponding
R complex Borel mea-
sure µ = µΛ given by the equality Λ(g) = gdµ, g ∈ Cb (E). The mapping
+
Λ 7→ (<Λ) (f ), Λ ∈ M(E), is weakly continuous. This can be seen as fol-
+
lows. Suppose Λn (g) → Λ(g) for all g ∈ Cb (E). Then (<Λn ) (f ) ≥ <Λn (g)
+
for all 0 ≤ g ≤ f , g ∈ Cb (E), and hence lim inf n→∞ (<Λn ) (f ) ≥
+
lim inf n→∞ <Λn (g) = (<Λ) (g). It follows that lim inf n→∞ (<Λn ) (f ) ≥
+ + +
sup0≤g≤f (<Λ) (g) = (<Λ) (f ). Since limn→∞ (<Λn ) (1) = (<Λ) (1) we
+ +
also have lim inf n→∞ (<Λn ) (1−f ) ≥ sup0≤g≤1−f (<Λ) (g) = (<Λ) (1−f ).
+ +
Hence we see lim supn→∞ (<Λn ) (f ) ≤ (<Λ) (f ).

In what follows we write K(E) for the collection of compact subsets of E.


Theorem 1.8. Let M be a subset of M(E). Then the following assertions are
equivalent:
(a) For every sequence (fn )n∈N ⊂ ZCb (E) which decreases pointwise to the zero
function the equality inf sup fn d |µ| = 0 holds;
n∈N µ∈M
(b) The equality inf sup |µ| (E \ K) = 0 holds, and sup |µ| (E) < ∞;
K⊂E, K∈K(E) µ∈M µ∈M
(c) There exists a function u¯R∈ H +¯ (E) such that for all f ∈ Cb (E) and for
all µ ∈ M the inequality ¯ f dµ¯ ≤ kuf k∞ holds.
Moreover, if M ⊂ M(E) satisfies one of the equivalent conditions (a), (b) or
(c), then M is relatively weakly compact.
Let Λ : Cb (E) → C be a linear functional such that inf n∈N |Λ| (fn ) = 0
for every sequence (fn )n∈N ⊂ Cb (E) which decreases pointwise to zero.
Here the linear functional |Λ| is defined in such a way that |Λ| (f ) =
sup {|Λ(v)| : |v| ≤ f, v ∈ Cb (E)} for all f ∈ Cb+ (E). Then by RCorollary 1.5
there exists a complex Borel measure µ such that Λ(f ) = f dµRfor all
f ∈ Cb (E). The positive Borel measure |µ| is such that |Λ| (f ) = f d |µ|
for all f ∈ Cb (E).
Proof. (a) =⇒ (b). By choosing the sequence fn = n−1 1 we see that
supµ∈M |µ| (E) < ∞. Next let ρ be a metric on E which it a polish space, let
(xn )n∈N be a dense sequence in E, and put Bk,n = {x ∈ E : ρ (x, xk ) ≤ 2−n }.
Choose continuous functions wk,n ∈ Cb (E) such that 1Bk,n
c ≤ wk,n ≤ 1Bk,n+1
c .
Put v`,n = min1≤k≤` wk,n . Then for every n ∈ N the sequence ` 7→ w`,n
decreases pointwise to Rzero. So for given ε > 0 and for given n ∈ N there
−n
exists
³ `n (ε) such
´ that w`n (ε),n d |µ| ≤ ε2 for all µ ∈ M . It follows that
` (ε)
c
|µ| ∩k=1
n
Bk,n ≤ ε2−n , and hence
1.1 Strict topology 15
³ ´
` (ε)
|µ| ∪∞ c
n=1 ∩k=1 Bk,n ≤ ε,
n
µ ∈ M.

` (ε)
Put K(ε) = ∩∞ n=1 ∪k=1 Bk,n . Then K(ε) is closed, and thus complete, and
n

completely ρ-bounded. Hence it is compact. Moreover, |µ| (E \ K(ε)) ≤ ε for


all µ ∈ M . Hence (b) follows from (a).
(b) =⇒ (c). This proof follows the lines of proof of assertion 1 of Proposition
1.4. Instead of considering just one measure we now have a family of measures
M.
(c) =⇒ (a). Essentially speaking this is a consequence of Dini’s lemma.
¯R Here
¯
we use the following fact. If for some µ ∈ M(E) ¯the inequality ¯ f dµ¯ ≤
R ¯
kuf k∞ holds for all f ∈ Cb (E), then we also have ¯ f d |µ|¯ ≤ kuf k∞ for all
f ∈ Cb (E). Fix α > 0. If (fn )n∈N is any sequence in Cb+ (E) which decreases
pointwise to zero, then for µ ∈ M we have the following estimate
Z
¡° ° ° ° ¢
fn d |µ| ≤ max °u1{u≥α} fn °∞ , °u1{u<α} fn °∞
à !
≤ max kuk∞ sup fn (x), α sup fn (x)
x∈{u≥α} x∈E
à !
≤ max kuk∞ sup fn (x), α sup f1 (x) . (1.21)
x∈{u≥α} x∈E

From the fact that the set {u ≥ α} is contained in a compactR subset of


E from (1.21) and Dini’s lemma we deduce that inf n∈N supµ∈M fn d |µ| ≤
α supx∈E f1 (x) for all α > 0. Consequently, (a) follows.
Finally we prove that if M satisfies (c), then M is relatively weakly com-
pact. First observe that µ ∈ M implies |µ| (E) ≤ kuk∞ . So the subset M is
uniformly bounded, and since E is a polish space, the same is true for the
ball {µ ∈ M(E) : |µ| (E) ≤ kuk∞ } endowed with the weak topology. There-
fore, if (µn )n∈N is a sequence
R in M it contains a subsequence (µnk )k∈N such
that Λ(f ) := limk→∞ f dµnk exists for all f ∈ Cb (E). Then it follows that
|Λ(f )| ≤ kuf k∞ for all f ∈ Cb (E). Consequently,
R the linear functional Λ can
be represented as a measure: Λ(f ) = f dµ, f ∈ Cb (E). It follows that the
weak closure of the set M is weakly compact.
Definition 1.9. A family of complex measures M ⊂ M(E) is called tight
if it satisfies one of the equivalent conditions in Theorem 1.8. Let M f be a
collection of linear functionals on Cb (E) which are continuous for the strict
topology. Then each Λ ∈ M f can be represented by a measure: Λ(f ) = inf f dµΛ ,
f ∈ Cb (E). Then the collection Mf of linear functionals is called tight, provided
n o
the same is true for the family M = µΛ : Λ ∈ M f .

Remark 1.10. In fact if M satisfies (a), then M satisfies Dini’s condition in


the sense that a sequence of functions µ 7→ |µ| (fn ) which decreasing pointwise
16 1 Strong Markov processes

to zero in fact converges uniformly on M . Assertion (b) says that the family
M is tight in the usual sense as it can be found in the standard literature.
Assertion (c) says that the family M is equi-continuous for the strict topology.

The following corollary says that if for M in Theorem 1.8 we choose a col-
lection of positive measures, then the family M is tight if and only if it is
relatively weakly compact. Compare these results with Stroock [224].
Corollary 1.11. Let M be a collection of positive Borel measures. Then the
following assertions are equivalent:
(a) The collection M is relatively weakly compact.
(b) The collection M is tight in the sense that supµ∈M µ(E) < ∞ and
inf K∈K(E) supµ∈M µ (E \ K) = 0. ¯R ¯
(c) There exists a function u ∈ H + (E) such that ¯ f dµ¯ ≤ kuf k∞ for all
µ ∈ M and for all f ∈ Cb (E).

Remark 1.12. Suppose that the collection M in Corollary 1.11 consists of prob-
ability measures and is closed with respect to the Lévy metric. If M satisfies
one of the equivalent conditions in 1.11, then it is a weakly compact subset of
P (E), the collection of Borel probability measures on E.

Proof. Corollary 1.11 follows more or less directly from Theorem 1.8. Let
M be as in Corollary 1.11, and (fn )n∈N be a sequence in Cb (E) which de-
creases
R to the zero
R function. Then observe that the sequence of functions
µ 7→ fn d |µ| = fn dµ, µ ∈ M , decreases pointwise to zero. Each of these
functions is weakly continuous. Hence, if M is relatively weakly compact, then
Dini’s lemma implies that this sequence converges uniformly on M to zero. It
follows that assertion (a) in Corollary 1.11 implies assertion (a) in Theorem
1.8. So we see that in Corollary 1.11 the following implications are valid: (a)
=⇒ (b) =⇒ (c). If M ⊂ M+ (E) satisfies (c), then Theorem 1.8 implies that
M is relatively weakly compact. This means that the assertions (a), (b) and
(c) in Corollary 1.11 are equivalent.

We will also need the following theorem.


Theorem 1.13. Let (µn )n∈N ⊂ M(E) be a tight R sequence (see Definition
1.9) with the property that Λ(f ) := limn→∞ f dµn exists for all f ∈
Cb (E). Let Φ ⊂ Cb (E) be a family of functions which is equi-continuous
and bounded.
¯Z Then ΛZ can be
¯ represented as a complex Borel measure µ, and
¯ ¯
lim sup ¯¯ ϕdµn − ϕdµ¯¯ = 0.
n→∞ ϕ∈Φ

Remark 1.14. According to the Theorem of Arzela-Ascoli an equi-continuous


and uniformly bounded family of functions restricted to a compact subset K
is relatively compact in Cb (K).
1.1 Strict topology 17

Proof. The fact that the linear functional Λ can be represented by a Borel
measure follows from Corollary 1.5 and Theorem 1.8. Assume to arrive at a
contradiction that
¯Z Z ¯
¯ ¯
lim sup sup ¯¯ ϕdµn − ϕdµ¯¯ > 0.
n→∞ ϕ∈Φ

Then there exist ε > 0, a subsequence (µnk )k∈N , and a sequence (ϕk )k∈N ⊂ Φ
such that ¯Z Z ¯
¯ ¯
¯ ϕk dµn − ϕk dµ¯ > ε, k ∈ N. (1.22)
¯ k ¯
Choose a compact subset of E in such a way that
ε
sup kϕk∞ × sup |µn | (E \ K) ≤ . (1.23)
ϕ∈Φ n∈N 16

By the Bolzano-Weierstrass theorem for bounded equi-continuous families of


functions, there exists a continuous function ϕK ∈ C(K) and a subsequence
of the sequence (ϕk )k∈N , which we call again (ϕk )k∈N , such that
¯ ¯
lim sup ¯ϕk (x) − ϕK (x)¯ = 0. (1.24)
k→∞ x∈K

By Tietze’s extension theorem there exists a continuous function ϕ ∈ Cb (E)


such that ϕ restricted to K coincides with ϕK and such that |ϕ| ≤ 2 sup kψk∞ .
ψ∈Φ
From (1.24) it follows there exists kε ∈ N such that for k ≥ kε the inequality
ε
sup |µn | (E) k1K (ϕk − ϕ)k∞ ≤ . (1.25)
n∈N 8

From (1.23)and (1.25) we obtain the following estimate:


¯Z Z ¯
¯ ¯
¯ ϕk dµn − ϕk dµ¯
¯ k ¯
¯Z Z ¯
¯ ¯
≤ ¯¯ (ϕk − ϕ) dµnk − (ϕk − ϕ) dµ¯¯
K K
¯Z Z ¯ ¯Z Z ¯
¯ ¯ ¯ ¯
¯ ¯ ¯
+¯ (ϕk − ϕ) dµnk − (ϕk − ϕ) dµ¯ + ¯ ϕdµnk − ϕdµ¯¯
¯ E\K E\K ¯
≤ k1K (ϕk − ϕ)k∞ (|µnk | (K) + |µ| (K))
¯Z Z ¯
¯ ¯
+ 4 sup kψk∞ (|µnk | (E \ K) + |µ| (E \ K)) + ¯ ϕdµnk − ϕdµ¯¯
¯
ψ∈Φ

≤ 2 k1K (ϕk − ϕ)k∞ sup |µnk | (K)


k∈N
¯Z Z ¯
¯ ¯
+ 8 sup kψk∞ sup |µnk | (E \ K) + ¯ ϕdµnk − ϕdµ¯¯
¯
ψ∈Φ k∈N
18 1 Strong Markov processes
¯Z Z ¯
3 ¯ ¯
≤ ε + ¯¯ ϕdµnk − ϕdµ¯¯ . (1.26)
4
¯R R ¯
Since limn→∞ ¯ ϕdµn − ϕdµ¯ = 0 the equality in (1.26) implies
¯Z Z ¯
¯ ¯
¯ ϕk dµn − ϕk dµ¯ < ε (1.27)
¯ k ¯

for k large enough. The conclusion in (1.27) contradicts our assumption in


(1.22).
This proves Theorem 1.13.

Occasionally we will need the following version of the Banach-Alaoglu


R the-
orem; see e.g. Theorem 7.18. We use the notation hf, µi = E f (x) dµ(x),
f ∈ Cb (E), µ ∈ M (E). For a proof of the following theorem we refer to e.g.
Rudin [205]. Notice that any Tβ -equi-continuous family of measures is con-
tained in Bu for some u ∈ H(E). Here Bu is the collection defined in (1.28)
below.
Theorem 1.15. (Banach-Alaoglu) Let u be a function in H(E), and define
the subset Bu of M (E) by

Bu = {µ ∈ M (E) : |hf, µi| ≤ kuf k∞ for all f ∈ Cb (E)} . (1.28)

Then Bu is σ (M (E), Cb (E))-compact.


Since the space (Cb (E), Tβ ) is separable, it follows that for every sequence
(µn )n∈N in Bu there exists a measure µ ∈ M (E) and a subsequence (µnk )k∈N
such that lim hf, µnk i = hf, µi for all f ∈ Cb (E).
k→∞
Instead of “σ (M (E), Cb (E))”-convergence we often write “weak∗ -conver-
gence”, which is a functional analytic term. In a probabilistic context people
usually write “weak convergence”.

1.1.3 Integral operators on the space of bounded continuous


functions

We insert a short digression to operator theory. Let E1 and E2 be two polish


spaces, and let T : Cb (E1 ) → Cb (E2 ) be a linear operator with the property
that its absolute value |T | : Cb (E1 ) → Cb (E2 ) determined by the equality

|T | (f ) = sup {|T g| : |g| ≤ f } , f ∈ Cb (E1 ) , f ≥ 0,

is well-defined and acts as a linear operator from Cb (E1 ) to Cb (E2 ).


Definition 1.16. A family of linear operators {Tα : α ∈ A}, where every
Tα ∈ L (Cb (E1 ) , Cb (E2 )) is called equi-continuous for the strict topology
if for every v ∈ H (E2 ) there exists u ∈ H (E1 ) such that the inequality
kvTα f k∞ ≤ kuf k∞ holds for all α ∈ A and for all f ∈ Cb (E1 ).
1.1 Strict topology 19

So the notion “equi-continuous for the strict topology” has a functional ana-
lytic flavor.
Definition 1.17. A family of linear operators {Tα : α ∈ A}, where every Tα
is a continuous linear operator from Cb (E1 ) to Cb (E2 ) is called tight if for
every compact subset K of E2 the family of functionals {Λα,x : α ∈ A, x ∈ K}
is tight in the sense of Definition 1.9. Here the functional Λα,x : Cb (E1 ) → C
is defined by Λα,x (f ) = Tα f (x), f ∈ Cb (E1 ). Its absolute value |Λα,x | has
then the property that |Λα,x | (f ) = |Tα | f (x), f ∈ Cb (E1 ).
The following theorem says that a tight family of operators {Tα : α ∈ A} is
equi-continuous for the strict topology and vice versa. Both spaces E1 and E2
are supposed to be polish.
Theorem 1.18. Let A be some index set, and let for every α ∈ A the mapping
Tα : Cb (E1 ) → Cb (E2 ) be a linear operator, which is continuous for the
uniform topology. Suppose that the family {Tα : α ∈ A} is tight. Then for every
v ∈ H (E2 ) there exists u ∈ H (E1 ) such that

kvTα f k∞ ≤ kuf k∞ , for every α ∈ A and for all f ∈ Cb (E1 ). (1.29)

Conversely, if the family {Tα : α ∈ A} is equi-continuous in the sense that for


every v ∈ H (E2 ) there exists u ∈ H (E1 ) such that (1.29) is satisfied. Then
the family {Tα : α ∈ A} is tight.
If the family {Tα : α ∈ A} satisfies (1.29), then the family {|Tα | : α ∈ A} sat-
isfies the same inequality with |Tα | instead of Tα . The argument to see this
goes in more or less the same way as we will prove the first part of Proposi-
tion 1.28 below. Fix f ∈ Cb (E1 ), α ∈ A, and x ∈ E1 , and let the functions
u ∈ H (E1 ) and v ∈ H (E2 ) be such that (1.29) is satisfied. Choose ϑ ∈ [−π, π]
in such a way that
¡ ¡ ¢¢ ³ ¡ ¢+ ´
|v(x) |Tα | (f )(x)| = |v(x)| |Tα | < eiϑ f (x) ≤ |v(x)| |Tα | < eiϑ f (x)

(definition of |Tα |)
n ¡ ¢+ o
= sup |v(x)Tα g(x)| : |g| ≤ < eiϑ f
n ¡ ¢+ o
≤ sup kugk∞ : |g| ≤ < eiϑ f ≤ kuf k∞ . (1.30)

From (1.30) we see that the inequality in (1.29) is also satisfied for the oper-
ators |Tα |, α ∈ A.
Corollary 1.19. Like in Theorem 1.18 let A be some index set, and let for
every α ∈ A the mapping Tα : Cb (E1 ) → Cb (E2 ) be a positivity preserving lin-
ear operator. Then the family {Tα : α ∈ A} is Tβ -equi-continuous if and only
if for every sequence (ψm )m∈N which decreases pointwise to 0, the sequence
{Tα (ψm f ) : m ∈ N} decreases pointwise to 0 uniformly in α ∈ A.
20 1 Strong Markov processes

Proof (Proof of Corollary 1.19.). Choose v ∈ H + (E). The proof follows by


considering the family of functionals Λα,x : Cb (E) → C, α ∈ A, x ∈ E, de-
fined by Λα,x f (x) = u(x)Tα f (x), f ∈ Cb (E). If the family {Tα : α ∈ A} is
Tβ -equi-continuous, then the family {Λα,x : α ∈ A, x ∈ E} is tight. For exam-
ple, it then easily follows that {Λu,α,x fm : α ∈ A, x ∈ E} converges uniformly
in α ∈ A, x ∈ E, to 0, provided that the sequence (fm )m∈N decreases point-
wise to 0. Conversely, suppose that for any given v ∈ H + (E), and for any
sequence of functions (fm )m∈N ⊂ Cb (E) which decreases pointwise to 0, the
sequence {Λv,α,x fm : α ∈ A, x ∈ E}m∈N converges uniformly to 0. Then the
family {Λα,x : α ∈ A, x ∈ E} is tight: see 1.19.

Proof (Proof of Theorem 1.18.). Like in Definition 1.17 the functionals Λα,x ,
α ∈ A, x ∈ E1 , are defined by Λα,x (f ) = [Tα f ] (x), f ∈ Cb (E1 ). First we
suppose that the family {Tα : α ∈ A} is tight. Let (fn )n∈N ⊂ Cb+ (E1 ) be
sequence of continuous functions which decreases pointwise to zero, and let v ∈
H (E2 ) be arbitrary. Since the family {Tα : α ∈ A} is tight, it follows that, for
every compact subset K the collection of functionals {Λα,x : α ∈ A, x ∈ K}
is tight. Then, since the sequence (fn )n∈N ⊂ Cb+ (E1 ) decreases pointwise to
zero, we have

lim sup |Λα,x | (fn ) = 0 for every compact subset K of E1 . (1.31)


n→∞ α∈A, x∈K

From (1.31) it follows that limn→∞ supα∈A, x∈K |v(x)| |Λα,x | (fn ) = 0. Hence
the family of functionals {|v(x)| Λα,x : α ∈ A, x ∈ E1 } is tight. By Theorem
1.8 (see Definition 1.9 as well) it follows that there exists a function u ∈ H (E1 )
such that
|v(x) [Tα f ] (x)| = |v(x)Λα,x (f )| ≤ kuf k∞ (1.32)
for all f ∈ Cb (E1 ), for all x ∈ E and for all α ∈ A. The inequality in (1.32)
implies the equi-continuity property (1.29).
Next let the family {Tα : α ∈ A} be equi-continuous in the sense that
it satisfies inequality (1.29). Then the same inequality holds for the family
{|Tα | : α ∈ A}; the argument was given just prior to the proof of Theorem
1.18. Let K be any compact subset of E1 and let (fn )n∈N ⊂ Cb+ (E1 ) be a
sequence which decreases to zero. Then there exists a function u ∈ H (E1 )
such that

sup [|Tα | fn ] (x) = k1K |Tα | fn k∞ ≤ kufn k∞ . (1.33)


α∈A, x∈K

From (1.33) it readily follows that limn→∞ supα∈A, x∈K [|Tα | fn ] (x) = 0. By
Definition 1.17 it follows that the family {Tα : α ∈ A} is tight.
This completes the proof of Theorem 1.18.

Theorem 1.20. Let E1 and E2 be two polish spaces, and let U : Cb (E1 , R) →
Cb (E2 , R) be a mapping with the following properties:
1.1 Strict topology 21

1. If f1 and f2 ∈ Cb (E1 ) are such that f1 ≤ f2 , then U (f1 ) ≤ U (f2 ). In


other words the mapping f 7→ U f , f ∈ Cb (E1 , R) is monotone.
2. If f1 and f2 belong to Cb (E1 , R), and if α ≥ 0, then U (f1 + f2 ) ≤ U (f1 )+
U (f2 ), and U (αf1 ) = αU (f1 ).
3. U is unit preserving: U (1E1 ) = 1E2 .
4. If (fn )n∈N ⊂ Cb (E1 , R) is a sequence which decreases pointwise to zero,
then so does the sequence (U (fn ))n∈N .
Then for every v ∈ H + (E2 ) there exists u ∈ H + (E1 ) such that
sup v(y)U (<f ) (y) ≤ sup u(x)<f (x), for all f ∈ Cb (E1 ) and hence
y∈E2 x∈E1

sup v(y)U |f | (y) ≤ sup u(x) |f (x)| , for all f ∈ Cb (E1 ). (1.34)
y∈E2 x∈E1

If the mapping U maps Cb (E1 ) to L∞ (E, R, E), then the conclusion about its
continuity as described in (1.34) is still true provided it possesses the properties
(1), (2), (3), and (4) is replaced by
40 If (fn )n∈N ⊂ Cb (E1 , R) is a sequence which decreases pointwise to zero,
then the sequence (U (fn ))n∈N decreases to zero uniformly on compact sub-
sets of E2 .
Proof. Put
½
MvU = ν ∈ M + (E1 ) : ν (E1 ) = sup v(y), < hg, νi ≤ sup v(y) (U <g) (y)
<
y∈E2 y∈E2
¾
for all g ∈ Cb (E1 ) and
½
|·|
MvU = ν ∈ M + (E1 ) : ν (E1 ) = sup v(y), |hg, νi| ≤ sup v(y) (U |g|) (y)
y∈E2 y∈E2
¾
for all g ∈ Cb (E1 ) . (1.35)

A combination of Theorem 1.8 and its Corollary 1.11 shows that the collections
< |·|
MvU and MvU are tight. Here we use hypothesis 4. We also observe that
< |·| |·|
MvU = MvU . This can be seen as follows. First suppose that ν ∈ MvU and
choose g ∈ Cb (E1 ). Then we have
h<g + k<gk∞ , νi ≤ sup v(y) (U |<g + kgk∞ |) (y)
y∈E2

≤ sup (v(y)U (<g) (y)) + sup v(y) kgk∞


y∈E2 y∈E2

= sup (v(y)U (<g) (y)) + ν (E1 ) kgk∞ . (1.36)


y∈E2

|·|
From (1.36) we deduce < hg, νi ≤ supy∈E1 (v(y)U (<g) (y)), and hence MvU ⊂
<
MvU . The reverse inclusion is show by the following arguments:
22 1 Strong Markov processes
­ ¡ iϑ ¢ ®
|hg, νi| = sup < e g ,ν
ϑ∈[−π,π]
¡¯ ¡ ¢¯¢
≤ sup sup v(y)U ¯< eiϑ g ¯ (y)
ϑ∈[−π,π] y∈E2

≤ sup sup v(y)U (|g|) (y) = sup v(y)U (|g|) (y). (1.37)
ϑ∈[−π,π] y∈E2 y∈E2

< |·|
From (1.37) the inclusion MvU ⊂ MvU follows. So from now on we will write
|·|
MvU = MvU = MvU . There exists a function Ru ∈ H + (E) such that for all
<

f ∈ Cb (E) and for all µ ∈ M the inequality < f dµ ≤ supx∈E < (u(x)f (x))
holds. The result in Theorem 1.20 follows from assertion in the following
equalities
sup v(y)U <f (y) = sup {< hf, νi : ν ∈ MvU } , and (1.38)
y∈E2

sup v(y)U |f | (y) = sup {|hf, νi| : ν ∈ MvU } . (1.39)


y∈E2

The equality in (1.38) follows from the Theorem of Hahn-Banach. In the


present situation it says that there exists a linear functional Λ : Cb (E1 , R) →
R such that Λ(f ) ≤ sup v(y)U f (y), for all f ∈ Cb (E1 , R), and
y∈E2

Λ (1E1 ) = sup v(y)U (1E1 ) (y) = sup v(y)1E2 (y) = sup v(y). (1.40)
y∈E2 y∈E2 y∈E2

Let f ∈ Cb (E1 , R), f ≤ 0. Then Λ(f ) ≤ sup v(y)U f (y) ≤ 0. Again using
y∈E2
Hypothesis 4 shows that Λ can be identified with a positive Borel measure on
E1 , which than belongs to MvU . Consequently, the left-hand side of (1.38) is
less than or equal to its right-hand side. Since the reverse inequality is trivial,
the equality in (1.38) follows. The equality in (1.39) easily follows from (1.38).
The assertion about a sub-additive mapping U which sends functions in
Cb (E1 ) to functions in L∞ (E, R, E) can easily be adopted from the first part
of the proof.
This concludes the proof of Theorem 1.20.
The results in Proposition 1.21 should be compared with Definition 3.6. We
describe two operators to which the results of Theorem 1.20 are applicable.
Let L be an operator with domain and range in Cb (E), with the property
that for all µ > 0 and f ∈ D(L) with µf − Lf ≥ 0 implies f ≥ 0. There
is a close connection between this positivity property (i.e. positive resolvent
property) and the maximum principle: see Definition 3.4 and inequality (3.46).
In addition, suppose that the constant functions belong to D(L), and that
L1 = 0. Fix λ > 0, and define the operators Uλj : Cb (E, R) → L∞ (E, R, E),
j = 1, 2, by the equalities (f ∈ Cb (E, R)):
Uλ1 f = sup inf {g ≥ f 1K : λg − Lg ≥ 0} , and (1.41)
K∈K(E) g∈D(L)

Uλ2 f = inf {g ≥ f : λg − Lg ≥ 0} . (1.42)


g∈D(L)
1.1 Strict topology 23

Here the symbol K(E) stands for the collection of all compact subsets of E.
Observe that, if g ∈ D(L) is such that λg − Lg ≥ 0, then g ≥ 0. This follows
from the maximum principle.
Proposition 1.21. Let the operator L be as above, and let the operators Uλ1
and Uλ2 be defined by (1.41) and (1.42) respectively. Then the following asser-
tions hold true:
(a) Suppose that the operator Uλ1 has the additional property that for every
sequence
¡ 1 ¢ (fn )n∈N ⊂ Cb (E) which decreases pointwise to zero the sequence
Uλ fn n∈N does so uniformly on compact subsets of E. Then for every
u ∈ H + (E) there exists a function v ∈ H + (E) such that

sup u(x)Uλ1 f (x) ≤ sup v(x)f (x), and


x∈E x∈E

sup u(x)Uλ1 |f | (x) ≤ sup v(x) |f (x)| for all f ∈ Cb (E, R). (1.43)
x∈E x∈E

(b) Suppose that the operator Uλ2 has the additional property that for every
sequence
¡ 2 ¢ (fn )n∈N ⊂ Cb (E) which decreases pointwise to zero the sequence
Uλ fn n∈N does so uniformly on compact subsets of E. Then for every
u ∈ H + (E) there exists a function v ∈ H + (E) such that the inequalities
in (1.43) are satisfied with Uλ2 instead of Uλ1 . Moreover, for f ∈ D (Ln ),
µ ≥ 0, and n ∈ N, the following inequalities hold:
n
µn f ≤ Uλ2 (((λ + µ) I − L) f ) , and (1.44)
n
µn kuf k∞ ≤ kv ((λ + µ) I − L) f k∞ . (1.45)

In (1.45) the functions u and v are the same as in (1.43) with Uλ2 replacing
Uλ1 .
The inequality in (1.45) could be used to say that the operator L is Tβ -
dissipative: see inequality (3.14) in Definition 3.5. Also notice that Uλ1 (f ) ≤
Uλ 2(f ), f ∈ Cb (E, R). It is not clear, under what conditions Uλ1 (f ) = Uλ2 (f ).
In Proposition 1.22 below we will return to this topic. The mapping Uλ1 is
heavily used in the proof of (iii) =⇒ (i) of Theorem 3.10. If the operator L in
Proposition 1.21 satisfies the conditions spelled out in assertion (a), then it is
called sequentially λ-dominant: see Definition 3.6.
Proof. The assertion in (a) and the first assertion in (b) is an immediate
consequence of Theorem 1.20. Let f ∈ D(L) be real-valued. The inequality
(1.45) can be obtained by observing that
24 1 Strong Markov processes

Uλ2 ((λ + µ) I − L) f
= inf {g ≥ ((λ + µ) I − L) f : λg − Lg ≥ 0}
g∈D(L)

= inf {g ≥ ((λ + µ) I − L) f :
g∈D(L)

(λ + µ) g − Lg ≥ µg ≥ ((λ + µ) I − L) (µf )}
= inf {g ≥ ((λ + µ) I − L) f : λg − Lg ≥ 0, g ≥ µf } ≥ µf. (1.46)
g∈D(L)

Repeating the arguments which led to (1.46) will show the inequality in (1.44).
From (1.46) and (1.43) with Uλ2 instead of Uλ1 we obtain

sup u(x) (µn f ) (x) ≤ sup Uλ2 ((λ + µ) f − Lf ) (x)


x∈E x∈E
n
≤ sup v(x) ((λ + µ)I − L) f (x), (1.47)
x∈E

for µ ≥ 0 and f ∈ D (Ln ). The inequality in (1.45) is an easy consequence of


(1.47). This concludes the proof of Proposition 1.21.

Proposition 1.22. Let the operator L with domain and range in Cb (E) have
the following properties:
1. For every λ > 0 the range of λI − L coincides with Cb (E), and the inverse
−1
R(λ) := (λI − L) exists as a positivity preserving bounded linear opera-
tor from Cb (E) to Cb (E). Moreover, 0 ≤ f ≤ 1 implies 0 ≤ λR(λ)f ≤ 1.
2. The equality lim λR(λ)f (x) = f (x) holds for every x ∈ E, and f ∈
λ→∞
Cb (E).
3. If (fn )n∈N ⊂ Cb (E) is any sequence which decreases pointwise to zero,
then for every λ > 0 the sequence (λR(λ)fn )n∈N decreases to zero as well.
Fix λ > 0, and define the mappings Uλ1 and Uλ2 as in (1.41) and (1.42)
respectively. Then the (in-)equalities
n o
k
sup (µR (λ + µ)) f ; µ > 0, k ∈ N ≤ Uλ1 (f ) ≤ Uλ2 (f ) (1.48)

hold for f ∈ Cb (E, R). Suppose that f ≥ 0. If the function in the left extremity
of (1.48) belongs to Cb (E), then the first two terms in (1.48) are equal. If it
belongs to D(L), then all three quantities in (1.48) are equal.

Proof. First we observe that for every (λ, x) ∈ (0, ∞) × E there exists a
Borel
R measure B 7→ r (λ, x, B) such that λr (λ, x, E) ≤ 1, and R(λ)f (x) =
E
f (y)r (λ, x, dy), f ∈ Cb (E). This result follows by considering the func-
tional Λλ,x : Cb (E) → C, defined by Λλ,x (f ) = R(λ)f (x). In fact

r (λ, x, B) = sup inf {R(λ)f (x) : f ≥ 1K } , B ∈ E.


K∈K(E), K⊂B

This result follows from Corollary 1.5. Often we write


1.1 Strict topology 25
Z
R(λ) (f 1B ) = f (y)r (λ, x, dy) , B ∈ E, f ∈ Cb (E).
B

Observe that the mapping B 7→ R(λ) (f 1B ) is a positive Borel measure on E.


Moreover, by Dini’s lemma we see that

lim sup sup λR(λ)fn (x) = 0, λ0 > 0. (1.49)


n→∞ λ≥λ0 x∈E

whenever the sequence (fn )n∈N ⊂ Cb (E) decreases pointwise to zero. From
Theorem 1.18 and its Corollary 1.19 it then follows that the family of operators
{λR(λ) : λ ≥ λ0 } is equi-continuous for the strict topology Tβ , i.e. for every
function u ∈ H + (E) there exists a function v ∈ H + (E) such that

λ kuR(λ)f k∞ ≤ kvf k∞ for all λ ≥ λ0 and all f ∈ Cb (E). (1.50)

Fix f ∈ Cb (E, R) and λ > 0. Next we will prove the


n o
k
Uλ1 (f ) ≥ sup (µ ((λ + µ)I − L)) : µ > 0, k ∈ N . (1.51)

A version of this proof will be more or less retaken in (3.137) in the proof of
the implication (iii) =⇒ (i) of Theorem 3.10 with D1 + L instead of L. First
we observe that for g ∈ D(L) we have

λg(x) − Lg(x) = lim µ (g(x) − µR (λ + µ) g(x)) , x ∈ E. (1.52)


µ→∞

If g ∈ D(L) is such that λg − Lg ≥ 0, then (λ + µ) g − Lg ≥ µg, and hence


g ≥ µR(λ + µ)g for all µ > 0. If g ≥ µR(λ + µ)g, then µ (g − µR(λ + µ)g) ≥ 0,
and by (1.52) we see λg − Lg ≥ 0. So that we have the following equality of
subsets

{g ∈ D(L) : λg − Lg ≥ 0} = {g ∈ D(L) : g ≥ µR (λ + µ) g for all µ > 0} .


(1.53)
From (1.53) we infer
( )
k
{g ∈ D(L) : λg − Lg ≥ 0} = g ∈ D(L) : g ≥ sup (µR (λ + µ)) g .
µ>0, k∈N
(1.54)
Let g ∈ D(L) be such that g ≥ f 1K and such that λg − Lg ≥ 0, then (1.54)
k k
implies g ≥ sup (µR (λ + µ)) (f 1K ). Since the operators (µR (λ + µ)) ,
µ>0, k∈N
µ > 0, k ∈ N, are integral operators, and bounded Borel measures are inner-
regular (with respect to compact subsets), we obtain
k
g≥ sup (µR (λ + µ)) f,
µ>0, k∈N

and hence
26 1 Strong Markov processes
k
sup inf {g ≥ f 1K : λg − Lg ≥ 0} ≥ sup (µ ((λ + µ) I − L)) f.
K∈K(E) g∈D(L) µ>0, k∈N
(1.55)
The inequality in (1.55) implies (1.51) and hence, since the inequality Uλ1 (f ) ≤
Uλ2 (f ) is obvious, the inequalities in (1.48) follow. Here we employ the fact
that λg − Lg ≥ 0 implies g ≥ 0. Fix a compact subset K of E, and f ≥ 0,
k
f ∈ Cb (E). If the function g = sup (µ (λ + µ) I − L) f belongs to Cb (E),
µ>0, k∈N
then g ≥ f 1K , and g ≥ µR (λ + µ) g for all µ > 0. Hence it follows that
k
sup (µ (λ + µ) I − L) f ≥ inf {g ≥ f 1K : g ≥ µR (λ + µ) g, g ∈ Cb (E)} .
µ>0, k∈N
(1.56)
Next we show that τβ - lim αR(α)f = f . From the assumptions 2 and 3, and
α→∞ ³ ´
−1
from (1.50) it follows that D(L) = R (βI − L) is Tβ -dense in Cb (E).
Therefore let g be any function in D(L), and let u ∈ H + (E). Consider, for
α > λ0 the equalities
f − αR(α)f = f − g − αR(α) (f − g) + g − αR(α)g
= f − g − αR(α) (f − g) − R(α) (Lg) , (1.57)
and the corresponding inequalities
ku (f − αR(α)f )k∞ ≤ ku (f − g)k∞ + kuαR(α) (f − g)k∞ + kuR(α) (Lg)k∞
kuk∞
≤ ku (f − g)k∞ + kv (f − g)k∞ + kLgk∞ . (1.58)
α
So that for given ε > 0 we first choose g ∈ D(L) in such a way that
2
ku (f − g)k∞ + kv (f − g)k∞ ≤ ε. (1.59)
3
kuk∞ 1
Then we choose αε ≥ λ0 so large that kLgk∞ ≤ ε. From the latter,
α 3
(1.58), and (1.59) we conclude:
ku (f − αR(α)f )k∞ ≤ ε, for α ≥ αε . (1.60)
From (1.60) we see that Tβ - lim αR(α)f = f . So that the inequality in (1.56)
α→∞
implies:
k
sup (µ (λ + µ) I − L) f ≥ inf {g ≥ f 1K : g ≥ µR (λ + µ) g, g ∈ D((L)} ,
µ>0, k∈N
(1.61)
k
and consequently Uλ1 (f ) ≤ f λ := sup (µ (λ + µ) I − L) f . It follows that
µ>0, k∈N
f λ = Uλ1 (f ) provided that f and f λ both belong to Cb (E). If f λ ∈ D(L),
then f λ = Uλ1 (f ) and f λ ≥ µR(λ + µ)f λ , and consequently λf λ − Lf λ . The
conclusion Uλ2 (f ) = f λ is then obvious.
This finishes the proof of Proposition 1.22.
1.2 Strong Markov processes and Feller evolutions 27

In the following proposition we see that a multiplicative Borel measure is a


point evaluation.

RProposition R 1.23.R Let µ be a non-zero Borel measure with the property that
f gdµ = f dµ R gdµ for all functions f and g ∈ Cb (E). Then there exists
x ∈ E such that f dµ = f (x) for f ∈ Cb (E).
R R
Proof. Since µ 6= 0 there exists f ∈ Cb (E) such that 0 6= f dµ = f 1dµ =
R R R ¡R ¢2 R
f dµ 1dµ, and hence 0 6= 1dµ = 1dµ . Consequently, 1dµ = 1. Let
+
f and g be functions in Cb (E). Then we have
Z ½¯Z ¯ ¾
¯ ¯
¯ ¯
f gd |µ| = sup ¯ hdµ¯ : |h| ≤ f g, h ∈ Cb (E)
½¯Z ¯ ¾
¯ ¯
¯ ¯
= sup ¯ h1 h2 dµ¯ : |h1 | ≤ f, |h2 | ≤ g, h1 , h2 ∈ Cb (E)
½¯Z ¯ ¾
¯ ¯
¯ ¯
= sup ¯ h1 dµ¯ : |h1 | ≤ f, h1 ∈ Cb (E)
½¯Z ¯ ¾
¯ ¯
¯ ¯
× sup ¯ h2 dµ¯ : |h2 | ≤ g, h2 ∈ Cb (E)
Z Z
= f d |µ| gd |µ| . (1.62)

From (1.62) it follows that the variation measure |µ| is multiplicative as well.
Since E is a polish space, the measure |µ| is inner regular. So there exists a
compact subset K of E such that |µ| (E \ K) ≤ 1/2, and hence |µ| (K) > 1/2.
Since |µ| is multiplicative it follows that |µ| (K) = 1 = |µ| (E). It follows that
the multiplicative measure |µ| is concentrated on the compact subset K, and
hence it can be considered as a multiplicative measure on C(K). But then
there exists a point x ∈ K such that |µ| = δx , the Dirac measure at x. So
there exists a constant cx such that µ = cx |µ| = cx δx . Since µ(E) = δx (E) = 1
it follows that cx = 1. This proves Proposition 1.23.

1.2 Strong Markov processes and Feller evolutions


In the sequel E denotes a separable complete metrizable topological Hausdorff
space. In other words E is a polish space. The space Cb (E) is the space of all
complex valued bounded continuous functions. The space Cb (E) is not only
equipped with the uniform norm: kf k∞ := supx∈E |f (x)|, f ∈ Cb (E), but
also with the strict topology Tβ . It is considered as a subspace of the bounded
Borel measurable functions L∞ (E), also endowed with the supremum norm.
Definition 1.24. A family {P (s, t) : 0 ≤ s ≤ t ≤ T } of operators defined on
L∞ (E) is called a Feller evolution or a Feller propagator on Cb (E) if it pos-
sesses the following properties:
28 1 Strong Markov processes

(i) It leaves Cb (E) invariant: P (s, t)Cb (E) ⊆ Cb (E) for 0 ≤ s ≤ t ≤ T ;


(ii) It is an evolution: P (τ, t) = P (τ, s) ◦ P (s, t) for all τ , s, t for which
0 ≤ τ ≤ s ≤ t and P (t, t) = I, t ∈ [0, T ];
(iii) It consists of contraction operators: kP (s, t)f k∞ ≤ kf k∞ for all t ≥ 0
and for all f ∈ Cb (E);
(iv) It is positivity preserving: f ≥ 0, f ∈ Cb (E), implies P (s, t)f ≥ 0;
(v) For every f ∈ Cb (E) the function (s, t, x) 7→ P (s, t)f (x) is continuous
on the diagonal of the set {(s, t, x) ∈ [0, T ] × [0, T ] × E : 0 ≤ s ≤ t ≤ T }
in the sense that for every element (t, x) ∈ (0, T ] × E the equality
lim P (s, t)f (y) = f (x) holds, and for every element (s, x) ∈ [0, T ) × E
s↑t,y→x
the equality lim P (s, t)f (y) = f (x) holds.
t↓s,y→x
(vi) For every t ∈ [0, T ] and f ∈ Cb (E) the function (s, x) 7→ P (s, t)f (x) is
Borel measurable and if (sn , xn )n∈N is any sequence in [0, t] × E such that
sn decreases to s ∈ [0, t], xn converges to x ∈ E, and lim P (sn , t) g (xn )
n→∞
exists in C for all g ∈ Cb (E), then lim P (sn , t) f (xn ) = P (s, t)f (x).
n→∞
(vii) For every (t, x) ∈ (0, T ] × E and f ∈ Cb (E) the following equality holds:
lims↑t, s≥τ P (τ, s) f (x) = P (τ, t) f (x), τ ∈ [0, t).

Remark 1.25. Since the space E is polish, the continuity as described in (v) can
also be described by sequences. So (v) is equivalent to the following condition:
for all element (t, x) ∈ (0, T ] × E and (s, x) ∈ [0, T ) × E the equalities

lim P (sn , t) f (yn ) = f (x) and lim P (s, tn ) f (yn ) = f (x) (1.63)
n→∞ n→∞

hold. Here (sn )n∈N ⊂ [0, t] is any sequence which increases to t, (tn )n∈N ⊂
[s, T ] is any sequence which decreases to s, and (yn )n∈N is any sequence in
E which converges to x ∈ E. If for f ∈ Cb (E) and t ∈ [0, T ] the function
(s, x) 7→ P (s, t)f (x), (s, x) ∈ [0, t] × E, is continuous, then (vi) and (vii) are
satisfied. If the function (s, t, x) 7→ P (s, t) f (x) is continuous on the space
{(s, t, x) ∈ [0, T ] × [0, T ] × E : s ≤ t}, then the propagator P (s, t) possesses
property (v) through (vii). In Proposition 1.26 we will single out a closely
related property. Its proof is part of the proof of part (b) in Theorem 1.39.

Proposition 1.26. Let the family {P (τ, t) : 0 ≤ τ ≤ t ≤ T } which possesses


properties (i) through (iv) of Definition 1.24. Suppose that for every f ∈ Cb (E)
the function (τ, t, x) 7→ P (τ, t) f (x) is continuous on the space

{(τ, t, x) ∈ [0, T ] × [0, T ] × E : τ ≤ t} . (1.64)

Then for every f ∈ Cb ([0, T ] × E) the function (τ, t, x) 7→ P (τ, t) f (t, ·) (x)
is continuous on the space in (1.64).

It is noticed that assertions (iii) and (iv) together are equivalent to


(iii0 ) If 0 ≤ f ≤ 1, f ∈ Cb (E), then 0 ≤ P (s, t)f ≤ 1, for 0 ≤ s ≤ t ≤ T .
1.2 Strong Markov processes and Feller evolutions 29

In the presence of (iii), (ii) and (i), property (v) is equivalent to:
(v0 ) lim ku (P (s, t)f − f )k∞ = 0 and lim ku (P (s, t)f − f )k∞ = 0 for all f ∈
t↓s s↑t
Cb (E) and u ∈ H(E). So that a Feller evolution is in fact Tβ -strongly
continuous in the sense that, for every f ∈ Cb (E) and u ∈ H(E),

lim ku (P (s, t) f − P (s0 , t0 ) f )k∞ = 0, 0 ≤ s0 ≤ t0 ≤ T. (1.65)


(s,t)→(s0 ,t0 )
s≤s0 ≤t0 ≤t

Remark 1.27. Property (vi) is satisfied if for every t ∈ (0, T ] the function
(s, x) 7→ P (s, x; t, E) = P (s, t) 1(x) is continuous on [0, t] × E, and if for
every sequence (sn , xn )n∈N ⊂ [0, t] × E for which sn decreases to s and xn
converges to x, the inequality lim supn→∞ P (sn , t) f (xn ) ≥ P (s, t) f (x) holds
for all f ∈ Cb+ (E). Since functions of the form x 7→ P (s, t)f (x), f ∈ Cb (E),
belong to Cb (E), it is also satisfied provided for every f ∈ Cb (E) we have

lim P (sn , t) f = P (s, t) f, uniformly on compact subsets of E.


n→∞

This follows from the inequality:

|P (sn , t) f (xn ) − P (s, t) f (x)|


≤ |P (sn , t) f (xn ) − P (s, t) (xn )| + |P (s, t) f (xn ) − P (s, t) f (x)|

where sn ↓ s, xn → as n → ∞, and f ∈ Cb (E).


Proposition 1.28. Let {P (s, t) : 0 ≤ s ≤ t ≤ T } be a family of operators
having property (i) and (ii) of Definition 1.24. Then property (iii0 ) is equiva-
lent to the properties (iii) and (iv) together.
Moreover, if such a family {P (s, t) : 0 ≤ s ≤ t ≤ T } possesses property (i),
(ii) and (iii), then it possesses property (v) if and only if it possesses (v0 ).
Proof. First suppose that the operator P (s, t) : L∞ (E) → L∞ (E) has the
properties (iii) and (iv), and let f ∈ Cb (E) be such that 0 ≤ f ≤ 1. Then by
(iii) and (iv) we have 0 ≤ P (s, t)f (x) ≤ supy∈E f (y) ≤ 1, and hence (iii0 ) is
satisfied. Conversely, let f ∈ Cb (E) and x ∈ E. Then by (iii0 ) the operator
P (s, t) satisfies

<P (s, t)f (x) = [P (s, t)<f ] (x) ≤ sup <f (y) ≤ k<f k∞ . (1.66)
y∈E

There exists ϑ ∈ [−π, π] such that by (1.66) we have


£ ¤
|P (s, t)f (x)| = < eiϑ P (s, t)f (x)
£ ¡ ¢¤ ° ¡ ¢°
= P (s, t)< eiϑ f (x) ≤ °< eiϑ f °∞ ≤ kf k∞ ,

from which (iii0 ) easily follows.


Next, suppose that the family {P (s, t) : 0 ≤ s ≤ t ≤ T } possesses property
(v0 ). Then, by taking s0 = t0 , it clearly has property (v). Fix (s0 , t0 ) ∈ [0, T ]×
30 1 Strong Markov processes

[0, T ] in such a way that s0 ≤ t0 . For the converse implication we employ


Theorem 1.18 with the families of operators

{P (sm , s0 ) : 0 ≤ sm ≤ sm+1 ≤ s0 } and {P (t0 , tm ) : t0 ≤ tm+1 ≤ tm ≤ T }


(1.67)
respectively. Let (fn )n∈N be a sequence functions in Cb+ (E) which decreases
pointwise to zero. Then by Dini’s lemma and assumption (v) we know that

lim sup sup P (sm , s0 ) fn (x) = lim sup sup P (t0 , tm ) fn (x) = 0 (1.68)
n→∞ m∈N x∈K n→∞ m∈N x∈K

for all compact subsets K of E. From (1.68) we see that the sequences of
operators in (1.67) are tight. By Theorem 1.18 it follows that they are equi-
continuous. If the pair (s, t) belongs to [0, s0 ] × [t0 , T ], then we write

P (s, t) f − P (s0 , t0 ) f = P (s, t0 ) (P (t0 , t) − I) f + (P (s, s0 ) − I) P (s0 , t0 ) f.


(1.69)
Let v be a function in H(E). Since the first sequence in (1.67) is equi-
continuous and by invoking (1.69) there exists a function v ∈ H(E) such
that the following inequality holds for all m ∈ N and all f ∈ Cb (E):

ku (P (sm , tm ) f − P (s0 , t0 ) f )k∞


≤ kv (P (t0 , tm ) − I) f k∞ + ku (P (sm , s0 ) − I) P (s0 , t0 ) f k . (1.70)

In order to prove the equality in (1.65) it suffices to show that the right-hand
side of (1.70) tends to zero if m → ∞. By the properties of the functions u
and v it suffices to prove that

lim k1K (P (sm , s0 ) f − f )k∞ = lim k1K (P (t0 , tm ) f − f )k∞ = 0 (1.71)


m→∞ m→∞

for every compact subset K of E and for every function f ∈ Cb (E). The
equalities in (1.71) follow from the sequential compactness of K and (v) which
imply that

lim P (sm , s0 ) f (xm ) = f (x0 ) = lim P (t0 , tm ) f (xm )


m→∞ m→∞

whenever sm increases to s0 , tm decreases to t0 and xm converges to x0 .


This completes the proof of Proposition 1.28.

Definition 1.29. Let for every (τ, x) ∈ [0, T ] × E, a probability measure Pτ,x
on FTτ be given. Suppose that for every bounded random variable Y : Ω → R
the equality ¡ ¯ ¢
Eτ,x Y ◦ ∨t ¯ Ftτ = Et,X(t) [Y ◦ ∨t ]
holds Pτ,x -almost surely for all (τ, x) ∈ [0, T ] × E and for all t ∈ [τ, T ]. Then
the process

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} (1.72)


1.2 Strong Markov processes and Feller evolutions 31

is called a Markov process. If the fixed time t ∈ [τ, T ] may be replaced with a
stopping time S attaining values in [τ, T ], then the process in (1.72) is called
a strong Markov process. By definition Pτ,4 (A) = 1A (ω4 ) = δω4 (A). Here
A belongs to F, and ω4 (s) = 4 for all s ∈ [0, T ]. If

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)}

is a Markov process, then we write

P (τ, x; t, B) = Pτ,x (X(t) ∈ B), B ∈ E, x ∈ E, τ ≤ t ≤ T, (1.73)

for the corresponding transition function. The operator family (of evolutions,
propagators)
{P (s, t) : 0 ≤ s ≤ t ≤ T }
is defined by
Z
[P (s, t)f ](x) = Es,x [f (X(t))] = f (y)P (s, x; t, dy) , f ∈ Cb (E), s ≤ t ≤ T.

Let S : Ω → [τ, T ] be an (Ftτ )t∈[τ,T ] -stopping time. Then the σ-field FSτ is
defined by
FSτ = ∩t∈[τ,T ] {A ∈ FTτ : A ∩ {S ≤ t} ∈ Ftτ } .
Of course, a stochastic variable S : Ω → [τ, T ] is called an (Ftτ )t∈[τ,T ] -stopping
time, provided that for every t ∈ [τ, T ] the event {S ≤ t} belongs to Ftτ .
This is perhaps the right place to explain the compositions F ◦ ∨t , F ◦ ∧t ,
and F ◦ ϑt , if F : Ω → C is FT0 -measurable, and if t ∈ [0,QT ]. Such functions
n
F are called stochastic variable. If F is of the form F = j=1 fj (tj , X (tj )),
where the functions fj , 1 ≤ j ≤ n, are bounded Borel functions, defined on
[0, T ] × E, then, by definition,
n
Y n
Y
F ◦ ∨t = fj (tj ∨ t, X (tj ∨ t)) , F ◦ ∧t = fj (tj ∧ t, X (tj ∧ t)) , and
j=1 j=1
Yn
F ◦ ϑt = fj ((tj + t) ∧ T, X ((tj + t) ∧ T )) . (1.74)
j=1
¡ ¢
If t is a Ft0 t∈[0,T ] -stopping time, then a similar definition is applied. By the
Monotone Class Theorem, the definitions in (1.74) extend to all FT0 measurable
variables F , i.e. to all stochastic variables. For a discussion on the Monotone
Class Theorem see Subsection 1.4.2

1.2.1 Generators of Markov processes and maximum principles

We begin with the definition of of the generator of a time-dependent Feller


evolution.
32 1 Strong Markov processes

Definition 1.30. A family of operators L(t), 0 ≤ t ≤ T , is said to be the


(infinitesimal) generator of a Feller evolution {P (s, t) : 0 ≤ s ≤ t ≤ T }, if
P (s, t)f − f
L(s)f = Tβ -lim , 0 ≤ s ≤ T . This means that a function f belongs
t↓s t−s
P (s, t)f − f
to D (L(s)) whenever L(s)f := lim exists in Cb (E), equipped with
t↓s t−s
the strict topology. It is the same as saying
½ that the function L(s)f
¾ belongs to
P (s, t)f − f
Cb (E), that the family of functions : t ∈ (s, T ) is uniformly
t−s
bounded and that convergence takes place uniformly on compact subsets of E.
Such a family of operators is considered as an operator L with domain in the
space Cb ([0, T ] × E). A function f ∈ Cb ([0, T ] × E) is said to belong to D(L)
if for every s ∈ [0, T ] the function x 7→ f (s, x) is a member of D(L(s)) and
if the function (s, x) 7→ L(s)f (s, ·) (x) belongs to Cb ([0, T ] × E). Instead of
L(s)f (s, ·) (x) we often write L(s)f (s, x). If a function f ∈ D(L) is such that
the function s 7→ f (s, x) is continuously differentiable, then we say that f

belongs to D(1) (L). The time derivative operator is often written as D1 .
∂s
(1)
Its domain is denoted by D (D1 ), and hence D (L) = D (D1 ) ∩ D(L).
Definition 1.31. The family of operators L(s), 0 ≤ s ≤ T , is said to generate
a time-inhomogeneous Markov process

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ τ ) , (∨t : τ ≤ t ≤ T ) , (E, E)} (1.75)

if for all functions u ∈ D(L), for all x ∈ E, and for all pairs (τ, s) with
0 ≤ τ ≤ s ≤ T the following equality holds:
· ¸
d ∂u
Eτ,x [u (s, X(s))] = Eτ,x (s, X(s)) + L(s)u (s, ·) (X(s)) . (1.76)
ds ∂s
Here it is assumed that the derivatives are interpreted as limits from the right
which converge uniformly on compact subsets of E, and that the differential
quotients are uniformly bounded.
So these derivatives are Tβ -derivatives.
Definition 1.32. By definition the Skorohod space D ([0, T ], E) consists of
all functions from [0, T ] to E which¡ posses left
¢ limits in E and are right-
continuous. The Skorohod space D [0, T ], E 4 consists of all functions from
[0, T ] to E 4 which posses left limits in E 4 and are right-continuous.
¡ ¢More
precisely, a path (or function) ω : [0, T ] → E 4 belongs to D [0, T ], E 4 if it
possesses the following properties:
(a) if ω(t) ∈ E, and s ∈ [0, t], then there exists ε > 0 such that X(ρ) ∈ E for
ρ ∈ [0, t + ε], and ω(s) = lim ω(ρ) and ω(s−) := lim ω(ρ) belong to E.
ρ↓s ρ↑s
(b) if ω(t) = 4 and s ∈ [t, T ], then ω(s) = 4. In other words 4 is an
absorbing state.
1.2 Strong Markov processes and Feller evolutions 33

Observe that the range of ω ∈ D ([0, T ], E) is contained in totally bounded


subset of
¡ E. Such ¢sets are relatively compact. Also observe that the range of
ω ∈ D [0, T ], E 4 restricted to an interval of the form [0, t] is ¡also totally¢
bounded provided that ω(t) ∈ E. It follows that paths ω ∈ D [0, T ], E 4
restricted to intervals of the form [0, t] have relatively compact range as long
as they have not reached the absorption state 4, i.e. as long as ω(t) ∈ E.
The assertions (a), (b), (c), and (d) of the following theorem are well-
known in case E is a locally compact second countable Hausdorff space. In
fact the sample space Ω should depend on τ . This is taken care of by assuming
that the measure Pτ,x is defined on the σ-field FTτ .
Let L be a linear operator with domain D(L) and range R(L) in Cb (E).
The following definition should be compared with Definition 3.4 below, and
with assertion (b) in Proposition 3.11.
Definition 1.33. Let E0 be subset of E. The operator L satisfies the maxi-
mum principle on E0 , provided

sup < (λf (x) − Lf (x)) ≥ λ sup <f (x), for all λ > 0, and for all f ∈ D(L).
x∈E0 x∈E0
(1.77)
If L satisfies (1.77) on E0 = E, then the operator L satisfies the maximum
principle of Definition 3.4.
The following definition is the same as the one in Definition 3.14 below.
Definition 1.34. Let E0 be a subset of E. Suppose that the operator L has the
property that for every λ > 0 and for every x0 ∈ E0 the number <h (x0 ) ≥ 0,
whenever h ∈ D(L) is such that < (λI − L) h ≥ 0 on E0 . Then the operator
L is said to satisfy the weak maximum principle on E0 .
The following proposition says that the concepts in the definitions 1.33 and
1.34 coincide, provided 1 ∈ D(L) and L1 = 0.
Proposition 1.35. If the operator L satisfies the maximum principle on E0 ,
then L satisfies the weak maximum principle on E0 . Suppose that the constant
functions belong to D(L), and that L1 = 0. If L satisfies the weak maximum
principle on E0 , then it satisfies the maximum principle on E0 .

Proof. First we observe that (1.77) is equivalent to

inf < (λf (x) − Lf (x)) ≤ λ inf <f (x), for all λ > 0, and for all f ∈ D(L).
x∈E0 x∈E0
(1.78)
Hence, if λf − Lf ≥ 0 on E0 , then (1.78) implies that <f (x0 ) ≥ 0 for all
x0 ∈ E0 .
Conversely, suppose that 1 ∈ D(L) and that L1 = 0. Let f ∈ D(L), put
m = inf {<f (y) : y ∈ E0 }, and assume that

inf < (λf − Lf ) (x) > λ inf <f (y) = λm. (1.79)
x∈E0 y∈E0
34 1 Strong Markov processes

Then there exists ε > 0 such that inf < (λf − Lf ) (x) ≥ λ (m + ε). Hence,
x∈E0
since L1 = 0, inf < (λI − L) (f − m − ε) (x) ≥ 0. Since the operator L satis-
x∈E
fies the weak maximum principle, we see < (f − m − ε) ≥ 0 on E0 . Since this
is equivalent to <f ≥ m + ε on E0 , which contradicts the definition of m.
Hence, our assumption in (1.79) is false, and consequently,

inf < (λf − Lf ) (x) ≤ λ inf <f (y). (1.80)


x∈E0 y∈E0

Since (1.80) is equivalent to (1.77) this concludes the proof of Proposition


1.35.

Definition 1.36. Let an operator L, with domain and range in Cb (E), sat-
isfy the maximum principle. Then L is said to possess the global Korovkin
property, if there exists λ0 > 0 such that fore every x0 ∈ E, the subspace
S (λ0 , x0 ), defined by

S (λ0 , x0 ) = {g ∈ Cb (E) : for every ε > 0 the inequality


sup {h1 (x0 ) : (λ0 I − L) h1 ≤ < g + ε, h1 ∈ D(L)}
≥ inf {h2 (x0 ) : (λ0 I − L) h2 ≥ < g − ε, h2 ∈ D(L)} is valid} ,
(1.81)

coincides with Cb (E).

Remark 1.37. Let D be a subspace of Cb (E) with the property that for every
x0 ∈ E the space S(x0 ), defined by

S (x0 ) = {g ∈ Cb (E) : for every ε > 0 the inequality


sup {h1 (x0 ) : h1 ≤ < g + ε, h1 ∈ D}
≥ inf {h2 (x0 ) : h2 ≥ < g − ε, h2 ∈ D} holds} , (1.82)

coincides with Cb (E). Then such a subspace D could be called a global Ko-
rovkin subspace of Cb (E). In fact the inequality in (1.82) is pretty much the
same as the one in (1.81) in case L = 0.

In what follows the symbol Kσ (E) denotes the collection of σ-compact subsets
of E. The set E0 in the following definition is in practical situation a member
of Kσ (E).
Definition 1.38. Let E0 be subset of E. Let an operator L, with domain and
range in Cb (E), satisfy the maximum principle on E0 . Then L is said to
possess the Korovkin property on E0 , if there exists λ0 > 0 such that for every
x0 ∈ K, the subspace Sloc (λ0 , x0 , E0 ), defined by
1.3 Strong Markov processes: main result 35
½
Sloc (λ0 , x0 , E0 ) = g ∈ Cb (E) : for every ε > 0 the inequality

sup {h1 (x0 ) : (λ0 I − L) h1 ≤ < g + ε, on E0 }


h1 ∈D(L)
¾
≥ inf {h2 (x0 ) : (λ0 I − L) h2 ≥ < g − ε, on E0 } ,
h2 ∈D(L)
(1.83)

coincides with Cb (E).

1.3 Strong Markov processes: main result


The following theorem contains the basic results about strong Markov pro-
cesses on polish spaces, their sample paths, and their generators. Item (a)
says that a Feller evolution (or propagator) can be considered as the one-
dimensional distributions, or marginals, of a strong Markov process. Item (b)
describes the reverse situation: with certain Markov processes we may asso-
ciate Feller propagators. In item (c) the intimate link between unique solutions
to the martingale problem and the strong Markov property is established. Item
(d) contains a converse result: Markov processes can be considered as solutions
to the martingale problem. Finally, in (e) operators which possess unique lin-
ear extensions which generate Feller evolutions are described, and for which
the martingale problem is uniquely solvable. For such operators the martingale
problem is said to be well-posed. A Hunt process is a strong Markov process
which is quasi-left continuous with respect to the minimum completed ad-
missible filtration {Ftτ }τ ≤t≤T . For assertion (a) in the locally compact setting
and a time-homogeneous Feller evolution (i.e. a Feller-Dynkin semigroup) the
reader may e.g. consult R.M. Blumenthal and R.K. Getoor [34].
Theorem 1.39.
Let {P (τ, t) : τ ≤ t ≤ T } be a Feller evolution in Cb (E). Then there exists
a strong Markov process (in fact a Hunt process)

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} , (1.84)

such that [P (τ, t)f ] (x) = Eτ,x [f (X(t))] , f ∈ Cb (E), t ≥ 0. Moreover this
Markov process is normal (i.e. Pτ,x [X(τ ) = x] = 1), is right continuous
(i.e. limt↓s X(t) = X(s), Pτ,x -almost surely for τ ≤ s ≤ T ), possesses left
limits in E on its life time (i.e. limt↑s X(t) exists in E, whenever ζ > s),
and¡ is quasi-left
¢ continuous (i.e. if (τn : n ∈ N) is an increasing sequence
τ
of Ft+ -stopping times, X(τn ) converges Pτ,x -almost surely to X (τ∞ ) on
the event {τ∞ < ζ}, where τ∞ = supn∈N τn ). Here ζ is the life time of the
process t 7→ X(t): ζ = inf {s > 0 : X(s) = 4}, when X(s) = 4 for some
s ∈ [0, T ], and elsewhere ζ = T . Put
τ
Ft+ = ∩s∈(t,T ] Fsτ = ∩s∈(t,T ] σ (X(ρ) : τ ≤ ρ ≤ s) . (1.85)
36 1 Strong Markov processes

Let F : Ω → C be a bounded FTs -measurable stochastic variable. Then


£ ¯ ¤ £ ¯ τ ¤
Es,X(s) [F ] = Eτ,x F ¯ Fsτ = Eτ,x F ¯ Fs+ (1.86)

Pτ,x -almost surely for all τ ≤ s and x ∈ E. Consequently, the process


defined in (1.84)
¡ τ ¢ is in fact a Markov process with respect to the right closed
filtrations: Ft+ t∈[τ,T ]
, τ ∈ [0, T ]. Moreover, the events {X(t) ∈ E} and
{X(t) ∈ E, ζ ≥ t} coincide Pτ,x -almost surely for τ ≤ t ≤ T and x ∈ E.
Even more is true, the process defined in (1.84) is strong Markov with
respect to the filtrations (Ftτ )t∈[τ,T ] , τ ∈ [0, T ], in the sense that
£ ¯ τ ¤
ES,X(S) [F ◦ ∨S ] = Eτ,x F ◦ ∨S ¯ FS+ (1.87)

for ¡all bounded


¢ FT0 -measurable stochastic variables F : Ω → C and for
τ τ
all Ft+ t∈[τ,T ] -stopping times S : Ω → [τ, T ]. The σ-field FS+ is de-
fined in Remark 1.40: see (1.91). As Ω the Skorohod
¡ space
¢ D ([0, T ], E), if
P (τ, x : t, E) = 1 for all 0 ≤ τ ≤ t ≤ T , or D [0, T ], E 4 , otherwise, may
be chosen.
(a)(b) Conversely, let

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} (1.88)

be a strong Markov process which is normal, right continuous, and pos-


sesses left limits in E on its life time. Put, for x ∈ E and 0 ≤ τ ≤ t ≤ T ,
and f ∈ L∞ ([0, T ] × E, E),
Z
[P (τ, t)f (t, ·)] (x) = Eτ,x [f (t, X(t))] = P (τ, x; t, dy) f (t, y) , (1.89)

where P (τ, x; t, B) = Pτ,x [X(t) ∈ B], B ∈ E. Suppose that the function


(s, t, x) 7→ P (s, t)f (x) is continuous on the set

{(s, t, x) ∈ [0, T ] × [0, T ] × E : s ≤ t}

for all functions f belonging to Cb (E), 0 ≤ s ≤ t ≤ T . Then the


family {P (s, t) : T ≥ t ≥ s ≥ 0} is a Feller evolution. Moreover, for f ∈
Cb ([0, T ] × E) the function (s, t, x) 7→ P (s, t)f (t, ·) (x) is continuous on
the same space. The operators ∨t : Ω → Ω, t ∈ [τ, T ], have the property
that for all (τ, x) ∈ [0, T ] × E the equality X(s) ◦ ∨t = X (s ∨ t) holds Pτ,x
for all t ∈ [τ, T ].
(c) Let the family L = {L(s) : 0 ≤ s ≤ T } be the generator of a Feller evo-
lution in Cb (E) and let the process in (1.84) be the be the corresponding
Markov process. For every f ∈ D(L(s)) and for every (τ, x) ∈ [0, T ] × E,
the process
Z tµ ¶

t 7→ f (t, X(t)) − f (τ, X(τ )) − + L(s) f (s, X(s)) ds (1.90)
τ ∂s
1.3 Strong Markov processes: main result 37

is a Pτ,x -martingale for the filtration (Ftτ )T ≥t≥τ , where each σ-field Ftτ ,
T ≥ t ≥ τ ≥ 0, is (some completion
T of ) σ (X(u) : τ ≤ u ≤ t). In fact the σ-
field Ftτ may be taken as Ftτ = s>t σ (X(ρ) : τ ≤ ρ ≤ s). ItR is also possible
to complete Ftτ with respect to Pτ,µ , given by Pτ,µ (A) = Pτ,x (A)dµ(x).
For Ftτ the following σ-field may be chosen:
\ \
Ftτ = {Pτ,µ -completion of σ (X(u) : τ ≤ u ≤ s)} .
µ∈P (E) T ≥s>t

(d) Conversely, let L = {L(s) : 0 ≤ s ≤ T } be a family of Tβ -densely defined


linear operators with domain D(L(s)) and range R(L(s)) in Cb (E), such
that D(1) (L) is Tβ -dense in Cb ([0, T ] × E). Let

((Ω, FTτ , Pτ,x ) : (τ, x) ∈ [0, T ] × E)

be a unique family of probability spaces, together with³state variables ´


(X(t) : t ∈ [0, T ]) defined on the filtered measure space Ω, (Ftτ )τ ≤t≤T
with values in the state space (E, E) with the following properties: for all
pairs 0 ≤ τ ≤ t ≤ T the state variable X(t) is Ftτ -E-measurable, for all
pairs (τ, x) ∈ [0, T ] × E, Pτ,x [X(τ ) = x] = 1, and for all f ∈ D(1) (L) the
process
Z tµ ¶

t 7→ f (t, X(t)) − f (τ, X(τ )) − + L(s) f (s, X(s)) ds
τ ∂s
is a Pτ,x -martingale with respect to the filtration (Ftτ )τ ≤t≤T . Then the fam-
ily of operators L = {L(s) : 0 ≤ s ≤ T } possesses a unique extension

L0 = {L0 (s) : 0 ≤ s ≤ T } ,

which generates a Feller evolution in Cb (E). It is required that the operator


D1 + L is sequentially λ-dominant in the sense of Definition 3.6; i.e. for
every sequence of functions (ψm )m∈N ⊂ C©b ([0, T ] × E) ª which decreases
pointwise to zero implies that the sequence ψnλ : n ∈ N , defined by

ψnλ = sup inf {g ≥ ψn 1K : g ∈ D (D1 + L) , (λI − D1 − L) g ≥ 0} ,


K∈K([0,T ]×E)

decreases uniformly on compact subsets of [0, T ] × E to zero.


¡ In addition,
¢
the sample space Ω is supposed to the Skorohod space D [0, T ] , E 4 ; in
particular X(t) ∈ E, τ ≤ s < t, implies X(s) ∈ E.
(e) (Unique Markov extensions) Suppose that the Tβ -densely defined linear
operator ½ ¾

D1 + L = + L(s) : 0 ≤ s ≤ T ,
∂s
with domain and range in Cb ([0, T ] × E), possesses the global Korovkin
property and satisfies the maximum principle, as exhibited in Definition
38 1 Strong Markov processes

1.33. Also suppose that L assigns real functions to real functions. Then
the family L = {L(s) : 0 ≤ s ≤ T } extends to a unique generator L0 =
{L0 (s) : 0 ≤ s ≤ T } of a Feller evolution, and the martingale problem is
well posed for the family of operators {L(s) : 0 ≤ s ≤ T }. Moreover, the
Markov process associated with {L0 (s) : 0 ≤ s ≤ T } solves the martingale
problem uniquely for the family L = {L(s) : 0 ≤ s ≤ T }.
Let E0 be a subset of E which is polish for the relative metric. The same
conclusion is true with E0 instead of E if the operator D1 + L possesses
the following properties:
1. If f ∈ D(1) (L) vanishes on E0 , then D1 f + Lf vanishes on E0 as well.
2. The operator D1 + L satisfies the maximum principle on E0 .
3. The operator D1 + L is positive Tβ -dissipative on E0 .
4. The operator D1 + L is sequentially λ-dominant on E0 for some λ > 0.
5. The operator D1 + L has the Korovkin property on E0 .
The notion of maximum principle on E0 is explained in the definitions
1.34 and 1.33: see Proposition 1.35 as well. The concept of Korovkin prop-
erty on a subset E0 can be found in Definition © 1.38. Let (D1 + L) ª ¹E0
be the operator defined by D ((D1 + L) ¹E0 ) = f ¹E0 : f ∈ D(1) (L) , and
(D1 + L) ¹E0 (f ¹E0 ) = D1 f + Lf ¹E0 , f ∈ D(L). Then the operator L ¹E0
possesses a unique linear extension to the generator L0 of a Feller semigroup
on Cb (E0 ).
For the notion of Tβ -dissipativity the reader is referred to inequality (3.14)
in Definition 3.5, and for the notion of sequentially λ-dominant operator see
Definition 3.6. In Proposition 1.21 the function ψnλ in assertion (d) is denoted
by Uλ1 (ψn ). The sequential λ-dominance will guarantee that the semigroup
which can be constructed starting from the other hypotheses in (d) and (e) is
a Feller semigroup indeed: see Theorem 3.10.
Remark 1.40. Notice that in (1.87) we cannot necessarily write
£ ¯ ¤
ES,X(S) [F ◦ ∨S ] = Eτ,x F ◦ ∨S ¯ FSτ ,

because events of the form {S ≤ t} may not be Ftτ -measurable, and hence the
σ-field FSτ is not well-defined. In (1.87) the σ-field FS+
τ
is defined by
τ
© ª
FS+ = ∩t≥0 A ∈ FTτ : A ∩ [0, t] ∈ Ft+τ
. (1.91)

Remark 1.41. Let d : E × E → [0, 1] be a metric on E which turns E into a


complete metrizable space, and let 4 be an isolated point of E 4 = E ∪ {4}.
The metric d4 : E 4 × E 4 → [0, 1] defined by
¯ ¯
d4 (x, y) = d (x, y) 1E (x)1E (y) + ¯1{4} (x) − 1{4} (y)¯

turns E 4 into
¡ a complete
¢ metrizable space. Moreover, if (E, d) is separable,
then so is E 4 , d
¡ 4¢ 4 . We also notice that the function x 7→ 1E (x), x ∈ E 4 ,
belongs to Cb E .
1.3 Strong Markov processes: main result 39

Remark 1.42. Let {P (τ, t) : 0 ≤ τ ≤ t ≤ T } be an evolution family on Cb (E).


Suppose that for any sequence of functions (fn )n∈N which decreases pointwise
to zero limn→∞ P (τ, t) fn (x) = 0, 0 ≤ τ ≤ t ≤ T . Then there exists a family
of Borel measures {B 7→ P (τ, x; t, B) : 0 ≤ τ ≤ t ≤ T } such that
Z
P (τ, t) f (x) = f (y)P (τ, x; t, dy) , f ∈ Cb (E). (1.92)

This is a consequence of Corollary 1.5. In addition the family

{B 7→ P (τ, x; t, B) : 0 ≤ τ ≤ t ≤ T }

satisfies the equation of Chapman-Kolmogorov:


Z
P (τ, x; s, dz) P (s, z; t, B) = P (τ, x; t, B) , 0 ≤ τ ≤ s ≤ t ≤ T, B ∈ E.
(1.93)
Next, for B ∈ E4 , and 0 ≤ τ ≤ t ≤ T we put

N (τ, x; t, B) = P (τ, x; t, B ∩ E) + (1 − P (τ, x; t, E)) 1B (4), x ∈ E, and


N (τ, 4; t, B) = 1B (4) . (1.94)

Then the family {B 7→ N (τ, x; t, ¡B) : 0 ≤ τ ¢≤ t ≤ T } satisfies the Chapman-


Kolmogorov equation on E 4 , N τ, x; t, E 4 = 1, and N (τ, 4; t, E) = 0. So
that if B 7→ P (τ, x; t, B) is a sub-probability on E, then B 7→ N (τ, x; t, B) is
a probability measure on E4 , the Borel field of E 4 .

Remark 1.43. Besides the family of (maximum) time operators {∨t : t ∈ [0, T ]}
we have the following more or less natural families: {∧t : t ∈ [0, T ]} (min-
imum
© T time operators),
ª and the time translation or time shift operators
ϑt : t ∈ [0, T ] . Instead of ϑTt we usually write ϑt . The operators ∧t :
Ω → Ω have the basic properties: ∧s ◦ ∧t = ∧s∧t , s, t ∈ [0, T ], and
X(s) ◦ ∧t = X (s ∧ t), s, t ∈ [0, T ]. The operators ϑt : Ω → Ω, t ∈ [0, T ],
have the following basic properties: ϑs ◦ ϑt = ϑs+t , s, t ∈ [0, T ], and
X(s) ◦ ϑt = X ((s + t) ∧ T ) = X (ϑs+t (0)).

It is clear that if a diffusion process (Xt , Ω, Ftτ , Pτ,x ) generated by the family
of operators Lτ exists, then for every pair (τ, x) ∈ [0, T ] × Rd , the measure
Pτ,x solves the martingale problem π(τ, x). Conversely, if the family Lτ is
given, we can try to solve the martingale problem for all (τ, x) ∈ [0, T ] × Rd ,
find the measures Pτ,x , and then try to prove that Xt is a Markov process
with respect to the family of measures Pτ,x . For instance, if we know that
for every pair (τ, x) ∈ [0, T ] × Rd the martingale problem π(τ, x) is uniquely
solvable, then the Markov property holds, provided that for there exists op-
erators ∨s : Ω → Ω, 0 ≤ s ≤ T such that Xt ◦ ∨s = Xt∨s , Pτ,x -almost surely
for τ ≤ t ≤ T , and τ ≤ s ≤ T . For the time-homogeneous case see, e.g.,
[84] or [109]. The martingale problem goes back to Stroock and Varadhan
40 1 Strong Markov processes

(see [225]). It found numerous applications in various fields of Mathematics.


We refer the reader to [147], [136], and [135] for more information about and
applications of the Martingale problem. In [80] the reader may find singular
diffusion equations which possess or which do not possess unique solutions.
Consequently, for (singular) diffusion equations without unique solutions the
martingale problem is not uniquely solvable.
Examples of (Feller) semigroups can be manufactured by taking a contin-
uous function ϕ : [0, ∞) × E → E with the property that
ϕ (s + t, x) = ϕ (t, ϕ (s, x)) ,
for all s, t ≥ 0 and x ∈ E. Then the mappings f 7→ P (t)f , with P (t)f (x) =
f (ϕ (t, x)) defines a semigroup. It is a Feller semigroup if limx→4 ϕ (t, x) = 4.
An explicit example of such a function, which does not provide a Feller
x
semigroup on C0 (R) is given by ϕ(t, x) = q (example due to
1 + 12 tx2
∂u
V. Kolokoltsov). Put u(t, x) = P (t)f (x) = f (ϕ(t, x)). Then (t, x) =
∂t
∂u
−x3 (t, x). In fact this (counter-)example shows that solutions to the mar-
∂x
tingale problem do not necessarily give rise to Feller-Dynkin semigroups.
These are semigroups which preserve not only the continuity, but also the
fact that functions which tend to zero at 4 are mapped to functions with the
same property. However, for Feller semigroups we only require that continu-
ous functions with values in [0, 1] are mapped to continuous functions with
the same properties. Therefore, it is not needed to include a hypothesis like
(1.95) in item (d) of Theorem 1.39. Here (1.95) reads as follows: for every
(τ, s, t, x) ∈ [0, T ]3 × E, τ < s < t, the equality
Pτ,x [X(t) ∈ E] = Pτ,x [X(t) ∈ E, X(s) ∈ E] (1.95)
holds.
In fact the result as stated is correct, but in case E happens to be locally
compact, then the resulting semigroup need not be a Feller-Dynkin semigroup.
This means that the corresponding family of operators assigns bounded con-
tinuous functions to functions in C0 (E), but they need not vanish at 4. This
means that the main result, Theorem 2.5, as stated in [48] is not correct.
That is solutions to the martingale problem can, after having visited 4, still
be alive. In the case of a non-compact space the metric without the Lévy part
is not adequate enough. That is why we have added the Lévy term. The prob-
lem is that the limits of the finite-dimensional distributions, given in (2.105)
below, on its own need not be a measure, and so there is no way of applying
Kolmogorov’s extension theorem.

1.3.1 Some historical remarks and references


Whereas in [48] we used only the first term in the distance dL of formula
(2.104) this is not adequate in the non-compact case. The reason for this
1.3 Strong Markov processes: main result 41

is that the second term in the right-hand side of the definition of the metric
dL (P2 , P1 ) in (2.104) ensures us that the limiting “functionals” are probability
measures indeed. Here we use a concept due to Lévy: the Lévy metric. In [74]
the authors Dorroh and Neuberger also use the strict topology to describe the
behavior of semigroups acting on the space of bounded continuous functions
on a Polish space. In fact the author of the present book was at least partially
motivated by their work to establish a general theory for Markov processes
on Polish spaces. Another motivation is provided by results on bi-topological
spaces as established by e.g. Kühnemund in [140]. Other authors have used
this concept as well, e.g. Es-Sarhir and Farkas in [82]. The notion of “strict
topology” plays a dominant role in Hirschfeld [103]. As already mentioned
Buck [44] was the first author who introduced the notion of strict topology
(in the locally compact setting). He denoted it by β in §3 of [44]. There are
several other authors who used it and proved convergence and approxima-
tion properties involving the strict topology: Buck [43], Prolla [190], Prolla
and Navarro [191], Katsaras [131], Ruess [206], Giles [91], Todd [235], Wells
[253]. This list is not exhaustive: the reader is also referred to Prolla [189],
and the literature cited there. In [250] Varadhan describes a metric on the
space D ([0, 1], R) which turns it into a complete metrizable separable space;
i.e. the Skorohod topology turns D ([0, 1], R) into a Polish space. On the other
hand it is by no means necessary that the ¡Skorohod¢ topology is the most
natural topology to be used on the space D [0, 1], Rd . For example in [113]
Jakubowski employs a quite different topology on this space. In [114] elab-
orates on Skorohod’s ideas about sequential convergence of distributions of
stochastic processes. After that the S-topology, as introduced by Jakubowski,
has been used by several others as well: see the references in [39] as well. Def-
inition 1.44 below also appears in [39]. Although the definition is confined to
R-valued paths, the S-topology also extends easily to the finite dimensional
Euclidean space Rd . By V+ ⊂ D ([0, T ], R) we denote the space of nonneg-
ative and nondecreasing functions V : [0, T ] → [0, ∞) and V = V+ − V+ .
We know that any element V ∈ V+ determines a unique positive measure
dV on [0, T ] and V can be equipped with the topology of weak convergence
RT RT
of measures; i.e. the equality limn→∞ 0 ϕ(s)dVn (s) = 0 ϕ(s)dV (s) for all
functions ϕ ∈ C ([0, T ], R) describes the weak convergence of the sequence
(Vn )n∈N ⊂ V to V ∈ V. Without loss of generality we may assume that the
functions V ∈ V are right-continuous and possess left limits in R.
Definition 1.44. Let (Y n )1≤n≤∞ ⊂ D ([0, T ], R). The sequence (Y n )n∈N is
said to converges to Y ∞ with respect to the S-topology, if for every ε > 0 there
exist elements (V n,ε )1≤n≤∞ ⊂ V such that kV n,ε − Y n k∞ ≤ ε, n = 1, . . . , ∞,
Z T Z T
and lim ϕ(s) dV n,ε (s) = ϕ(s) dV ∞,ε (s), for all ϕ ∈ C ([0, T ], R).
n→∞ 0 0
42 1 Strong Markov processes

1.4 Dini’s lemma, Scheffé’s theorem, and the monotone


class theorem
The contents of this section is taken from Appendix E in [70]. In this section
we formulate and discuss these three theorems.

1.4.1 Dini’s lemma and Scheffé’s theorem

The contents of this subsection is devoted to Dini’s lemma and Scheffé’s the-
orem. Another proof of Dini’s lemma can be found in Stroock [222], Lemma
7.1.23, p. 146.
Lemma 1.45. (Dini) Let (fn : n ∈ N) be a sequence of continuous functions
on the locally compact Hausdorff space E. Suppose that fn (x) ≥ fn+1 (x) ≥ 0
for all n ∈ N and for all x ∈ E. If limn→∞ fn (x) = 0 for all x ∈ E, then,
for all compact subsets K of E, limn→∞ supx∈K fn (x) = 0. If the function f1
belongs to C0 (E), then limn→∞ supx∈E fn (x) = 0.
Proof. We only prove the second assertion. Fix η > 0 and consider the subset
\
{x ∈ E : fn (x) ≥ η} .
n∈N

Since, by assumption, the function f1 belongs to C0 (E), and lim fn (x) = 0,


T n→∞

S the intersection n∈N {x ∈ E : fn (x) ≥ η} is void. As


x ∈ E, it follows that
a consequence E = n∈N {fn < η}. Let ε > 0 and put K = {f1 ≥ ε}. The
subset K is compact. By the preceding argument there exist nε ∈ N for which
K ⊆ {fnε < ε}. For n ≥ nε , we have 0 ≤ fn (x) ≤ ε for all x ∈ E.
In Definition 1.46 and in Theorem 1.50 of this subsection (E, E, m) may be
any measure space with m(B) ≥ 0 for B ∈ E.
Definition 1.46. A collection of functions {fj : j ∈ J} in L1 (E, E, m) is uni-
formly L1 -integrable if for every ε > 0 there exists g ∈ L1 (E, E, µ), g ≥ 0, for
which Z
sup |fj | dm ≤ ε.
j∈J {|fj |≥g}

Remark 1.47. If the collection {fj : j ∈ J} is uniformly L1 -integrable, and if


{gj : j ∈ J} is a collection for which |gj | ≤ |fj |, m-almost everywhere, for all
j ∈ J, then the collection {gj : j ∈ J} is uniformly L1 -integrable as well.
Remark 1.48. Cauchy sequences in L1 (E, E, m) are uniformly L1 -integrable.
Remark 1.49. Let f ≥ 0 beR a function in L1 (Rν , B, m), where m is the
Lebesgue measure. Suppose f (x)dm(x) = 1 and limn→∞ nν f (nx) = 0 for
all x 6= 0. Put fn (x) = nν f (nx), n ∈ N. Then the sequence is not uniformly
L1 -integrable. This will follow from Theorem 1.50 below.
1.4 Dini’s lemma, Scheffé’s theorem, and the monotone class theorem 43

A version of Scheffé’s theorem reads as follows. Our proof uses the arguments
in the proof Theorem 3.3.5 (Lieb’s version of Fatou’s lemma) in Stroock [222],
p. 54. Another proof can be found in Bauer [25], Theorem 2.12.4, p. 103.
Theorem 1.50. (Scheffé) Let (fn : n ∈ N) be a sequence in L1 (E, E, m). If
limn→∞ fn (x) = f (x), m-almost everywhere, then the sequence (fn : n ∈ N)
is uniformly L1 -integrable if and and only
Z Z
lim |fn (x)| dm(x) = |f (x)| dm(x).
n→∞

Proof. Consider the m-almost everywhere pointwise inequality

0 ≤ |fn − f | + |f | − |fn | ≤ 2 |f | . (1.96)

First suppose that the sequence {fn : n ∈ N} is uniformly L1 -integrable. Then,


by Fatou’s lemma,
Z Z Z
|f (x)| dm(x) = lim inf |fn (x)| dm(x) ≤ lim inf |fn (x)| dm(x)

R
(choose g ∈ L1 (E, m) such that {|fn |≥g}
|fn (x)| dm(x) ≤ 1)
Z Z
≤ lim inf |fn (x)| dm(x) + |fn (x)| dm(x)
{|fn |≥g} {|fn |≤g}
Z
≤ 1 + g(x)dm(x). (1.97)

From (1.97) we see that the function f belongs to L1 (E, m). From Lebesgue’s
dominated convergence theorem in conjunction with (1.97) we infer
Z
lim (|fn − f | + |f | − |fn |) dm = 0. (1.98)
n→∞

Since the sequence {fn : n ∈ N} is uniformly L1 -integrable,


R and since for m-
almost all x, limn→∞ fn (x) = f (x), we see that limn→∞ |fn − f | dm = 0.
So from (1.98) we get
Z Z
lim |fn | dm = |f | dm < ∞. (1.99)
n→∞

Conversely, suppose (1.99) holds. Then f belongs to L1 (E, m). Again we may
invoke Lebesgue’s dominated convergence theorem
R to conclude (1.98) from
(1.96). Again using (1.99) implies limn→∞ |fn − f | dm = 0. An appeal to
Remark 1.48 yields the desired result.
44 1 Strong Markov processes

1.4.2 Monotone class theorem

Our presentation of the monotone class theorems is taken from Blumenthal


and Getoor [34], pp. 5–7. For other versions of this theorem see e.g. Sharpe
[208], pp. 364–366. Theorems 1.52, 1.53, and Propositions 1.54, 1.55 we give
closely related versions of this theorem.
Definition 1.51. Let Ω be a set and let S be a collection of subsets of Ω.
Then S is a Dynkin system if it has the following properties:
(a) Ω ∈ S;
(b) if A and B belong to S and if A ⊇ B, then A \ B belongs to S;
(c) if
S∞ (An : n ∈ N) is an increasing sequence of elements of S, then the union
n=1 An belongs to S.

The following result on Dynkin systems is well-known.


Theorem 1.52. Let M be a collection of subsets of of Ω, which is stable
under finite intersections. The Dynkin system generated by M coincides with
the σ-field generated by M.

Theorem 1.53. Let Ω be a set and let M be a collection of subsets of of Ω,


which is stable (or closed) under finite intersections. Let H be a vector space
of real valued functions on Ω satisfying:
(i) The constant function 1 belongs to H and 1A belongs to H for all A ∈ M;
(ii)if (fn : n ∈ N) is an increasing sequence of non-negative functions in H
such that f = supn∈N fn is finite (bounded), then f belongs to H.
Then H contains all real valued functions (bounded) functions on Ω, that are
σ(M) measurable.

Proof. Put D = {A ⊆ Ω : 1A ∈ H}. Then by (i) Ω belongs to D and D ⊇ M.


If A and B are in D and if B ⊇ A, then B \ A belongs to D. If (An : n ∈ N)
is an increasing sequence in D, then 1∪An = supn 1An belongs to D by (ii).
Hence D is a Dynkin system, that contains M. Since M is closed under finite
intersection, it follows by Theorem 1.52 that D ⊇ σ(M). If f ≥ 0 is measurable
with respect to σ(M), then
Xn2n
f = sup 2−n 1{f ≥j2−n } . (1.100)
n j=1

Since 1{f ≥j2−n } , j, n ∈ N, belong to σ(M), we see that f belongs to H. Here


we employed the fact that σ(M) ⊆ D. If f is σ(M)-measurable, then we write
f as a difference of two non-negative σ(M)-measurable functions.

The previous theorems, i.e. Theorems 1.52 and 1.53, are used in the following
form. Let Ω be a set and let (Ei , Ei )i∈I be a family of measurable spaces,
indexed by an arbitrary set I. For each i ∈ I, let Si denote a collection of
1.4 Dini’s lemma, Scheffé’s theorem, and the monotone class theorem 45

subsets of of Ei , closed under finite intersection, which generates the σ-field


Ei , and let fi : Ω → Ei be a map from Ω to Ei . In our presentation of the
Markov property the space Ei are all the same, and the maps fi , i ∈ I, are
the state variables X(t), t ≥ 0. in this context the following two propositions
follow.
\
Proposition 1.54. Let M be the collection of all sets of the form fi−1 (Ai ),
i∈J
Ai ∈ Si , i ∈ J, J ⊆ I, J finite. Then M is a collection of subsets of Ω which
is stable under finite intersection and σ(M) = σ (fi : i ∈ I).

Proposition 1.55. Let H be a vector space of real-valued functions on Ω such


that:
(i) the constant function 1 belongs to H;
(ii)if (hn : n ∈ N) is an increasing sequence of non-negative functions in H
such that h = supn hn is finite (bounded),
Q then h belongs to H;
(iii)H contains all products of the form i∈J 1Ai ◦ fi , J ⊆ I, J finite, and
Ai ∈ Si , i ∈ J.
Under these assumptions H contains all real-valued functions (bounded) func-
tions in σ(fi : i ∈ I).

Definition 1.56. Theorems 1.52 and 1.53, and Propositions 1.54 and 1.55
are called the monotone class theorems.

Other theorems and results on integration theory, not explained in the book,
can be found in any textbook on the subject. In particular this is true for
Fatou’s lemma and Fubini’s theorem on the interchange of the order of inte-
gration. Proofs of these results can be found in Bauer [25] and Stroock [222].
The same references contain proofs of the Radon-Nikodym theorem. This the-
orem may be phrased as follows.
Theorem 1.57. (Radon-Nikodym) If a finite measure µ on some σ-finite
measure space (E, E, m) is absolutely continuous with respect
R to m, then there
exists a function f ∈ L1 (E, E, m) such that µ(A) = A f (x)dm(x) for all
subsets A ∈ E.
The measure µ is said to be absolutely continuous with respect to m if m(A) =
0 implies µ(A) = 0, and the measure m is said to be σ-finiteS if there exists
an increasing sequence (En : n ∈ N) in E such that E = n∈N En and for
which m (En ) < ∞, n ∈ N. A very important application is the existence of
conditional expectations. This can be seen as follows.
Corollary 1.58. Let (Ω, F, P) be a probability space and let F0 be a sub-field
of F, and let Y : Ω → [0, ∞] be a F-measurable function (random vari-
able) in L1 (Ω, F, P). Then there exists a function G ∈ L1 (E, F0 , m) such that
E [Y 1A ] = µ(A) = E [G1A ] for all A ∈ F0 .
46 1 Strong Markov processes
£ ¯ ¤
By convention the random variable G is written as G = E Y ¯ F0 . It is called
the conditional expectation on the σ-field F0 .
Proof. Put m(A) = E [Y 1A ], A ∈ F, and let µ be the restriction of m to F0 .
If for some A ∈ F0 , m(A) − 0, then µ(A) = 0. The Radon-Nikodym theorem
yields the existence of a function G ∈ L1 (E, F0 , m) such that E [Y 1A ] =
µ(A) = E [G1A ] for all A ∈ F0 .
2
Strong Markov processes: proof of main result

2.1 Proof of the main result: Theorem 1.39


In the present section we will prove item (a), (b), (c), (d) and (e) of Theorem
1.39. We will need a number of auxiliary results which can be found in the
current section or in the sections 3.1 and 3.2. We will always give the relevant
references. We need the following definition.
Definition
¡ 2.1.
¢ Let {X(t)}t∈[0,T ] , and {Y (t)}t∈[0,T ] be stochastic processes
on Ω, FT0 , P . The process {X(t)}t∈[0,T ] is a modification of {Y (t)}t∈[0,T ] if
P [X(t) = Y (t)] = 1 for all t ∈ [0, T ].

2.1.1 Proof of item (a) of Theorem 1.39

This subsection contains the proof of part (a). It employs the Kolmogorov’s
extension theorem and it uses the polish nature of the state space E in an
essential way.
Proof (Proof of item (a) of Theorem 1.39). We begin with the proof of the
existence of a Markov process (1.84), starting from a Feller evolution: see
Definition 1.24. First we assume P (τ, t) 1 = 1. Remark 1.42 will be used to
prove assertion (a) in case P (τ, t) 1 < 1. Temporarily we write Ω = E [0,T ] en-
dowed with the product topology, and product σ-algebra (or product σ-field),
which is the smallest σ-field on Ω which renders all coordinate mappings, or
state variables, measurable. The state variables X(t) : Ω → E are defined by
X(t, ω) = X(t)(ω) = ω(t), ω ∈ Ω, and the maximal mappings ∨s : Ω → Ω,
s ∈ [0, T ], are defined by ∨s (ω)(t) = ω (s ∨ t). Let the family of Borel measures
on
{B 7→ P (τ, x; t, B) : B ∈ E, (τ, x) ∈ [0, T ] × E, t ∈ [τ, T ]}
be determined by the equalities:
Z
P (τ, t) f (x) = f (y)P (τ, x; t, dy) , f ∈ Cb (E). (2.1)
48 2 Proof of main result

By Kolmogorov’s extension theorem there exists a family of probability spaces


(Ω, FTτ , Pτ,x ) , (τ, x) ∈ [0, T ] × E,
such that
Eτ,x [f (X (t1 ) , . . . , X (tn ))]
Z Z
= . . . f (y1 , . . . , yn ) P (τ, x; t1 , dy1 ) . . . P (tn−1 , yn−1 ; tn , dyn ) (2.2)
| {z }

¡ n¢
where τ ≤ t1 < · · · < tn ≤ T , and f ∈ L∞ E n , E⊗ . For f ∈ Cb ([0, T ] × E),
0 ≤ f , and α > 0 given we introduce the following processes:
Z ∞
t 7→ αR(α)f (t, X(t)) = α e−α(ρ−t) P (t, ρ ∧ T ) f (ρ ∧ T, ·) (X(t)) dρ
t
Z ∞
=α e−α(ρ−t) Et,X(t) [f (ρ ∧ T, X (ρ ∧ T ))] , t ∈ [0, T ], and (2.3)
t
s 7→ P (s, t) f (t, ·) (X(s)) = Es,X(s) [f (t, X(t))] , s ∈ [0, t], t ∈ [0, T ]. (2.4)
The processes in (2.3) and (2.4) could have been more or less unified by
considering the process:
Z ∞
(s, t) 7→α e−α(ρ−t) P (s, ρ ∧ T ) f (ρ ∧ T, ·) (X(s)) dρ
t
= αP (s, t) R(α)f (t, ·) (X(s)) , 0 ≤ s ≤ t ≤ T. (2.5)
Observe that limα→∞ αR(α)f (t, X(t)) = f (t, X(t)), t ∈ [0, T ]. Here we
use the continuity of the function ρ 7→ P (t, ρ) f (ρ, ·) (X(t)) at ρ = t. In
addition, for y ∈ E fixed, we have that the family of functionals f 7→
αR(α)f (t, ·) (X(t)), α ≥ 1, t ∈ [τ, T ], is Pτ,x -almost surely equi-continuous
for the strict topology.
Our first task will be to prove that for every (τ, x) ∈ [0, T ] × E the orbit
{(t, X(t)) : t ∈ [τ, T ]} is a Pτ,x -almost surely sequentially compact. Therefore
we choose an infinite sequence (ρn , X (ρn ))n∈N where ρn ∈ [τ, T ], n ∈ N.
This sequence contains an infinite subsequence (sn , X (sn ))n∈N such that
sn < sn+1 , n ∈ N, or an infinite subsequence (tn , X (tn ))n∈N such that
tn > tn+1 , n ∈ N. In the first case we put s = supn∈N sn , and in the sec-
ond case we write t = inf n∈N tn . In either case we shall prove that there exists
a subsequence which is a Cauchy sequence in [τ, T ] × E for a compatible uni-
formly bounded metric Pτ,x -almost surely. First we deal with the case that tn
decreases to t ≥ τ . Then we consider the stochastic process in (2.4) given by
ρ 7→ Et,X(t) [f (ρ, X(ρ))] where f is an arbitrary function in Cb ([0, T ] × E).
By hypothesis on the transition function P (τ, x; t, B) we have
Z
lim Et,X(t) [f (tn , X (tn ))] = lim P (t, X(t); tn , dy) f (tn , y) = f (t, X(t)) .
n→∞ n→∞
(2.6)
2.1 Proof of the main result: Theorem 1.39 49
h i
2
By applying the argument in (2.6) to the process ρ 7→ Et,X(t) |f (ρ, X(ρ))| ,
ρ ∈ [t, T ], the Markov property implies
h¯ ¯2 i
Eτ,x ¯Et,X(t) [f (ρ, X(ρ))] − f (ρ, X(ρ))¯
h i h¯ ¯2 i
= Eτ,x |f (ρ, X(ρ))| + Eτ,x ¯Et,X(t) [f (ρ, X(ρ))]¯
2

h i
− 2<Eτ,x f (ρ, X(ρ))Et,X(t) [f (ρ, X(ρ))]
£ ¯ ¤
(Markov property: Et,X(t) [f (ρ, X(ρ))] = Eτ,x f (ρ, X(ρ)) ¯ Ftτ Pτ,x -almost
surely)
h i h¯ ¯2 i
= Eτ,x |f (ρ, X(ρ))| + Eτ,x ¯Et,X(t) [f (ρ, X(ρ))]¯
2

h i
− 2<Eτ,x Et,X(t) [f (ρ, X(ρ))]Et,X(t) [f (ρ, X(ρ))]
h i h¯ ¯2 i
= Eτ,x |f (ρ, X(ρ))| − Eτ,x ¯Et,X(t) [f (ρ, X(ρ))]¯ .
2
(2.7)
h i
2
Applying the argument in (2.6) to the process ρ 7→ Et,X(t) |f (ρ, X(ρ))| ,
ρ ∈ [t, T ], and employing (2.7) we obtain:
h¯ ¯2 i
lim Eτ,x ¯Et,X(t) [f (tn , X (tn ))] − f (ρ, X (ρ))¯ = 0. (2.8)
n→∞

Again using (2.6) and invoking (2.8) we see that

lim f (tn , X (tn )) = f (t, X(t))


n→∞

in the space L2 (Ω, FTτ , Pτ,x ). Hence there exists a subsequence denote by
(f (tnk , X (tnk )))k∈N which converges Pτ,x -almost surely to f (t, X(t)). Let
d : E × E → [0, 1] be a metric on E which turns it into a polish space, and
let (xj )j∈N be a countable dense sequence in E. The previous arguments are
applied to the function f : [0, T ] × E → R defined by

X
f (ρ, x) = 2−j (d (xj , x) + |ρj − ρ|) , (2.9)
j=1

where the sequence (ρj )j∈N is a dense sequence in [0, T ]. From the previous
arguments we see that there exists a subsequence (tnk , X (tnk ))k∈N such that

lim f (tnk , X (tnk )) = f (t, X(t)) , Pτ,x -almost surely. (2.10)


k→∞

It follows that limk→∞ tnk = t. From (2.10) we also infer that

lim d (xj , X (tnk )) = d (xj , X(t)) , Pτ,x -almost surely for all j ∈ N.
k→∞
(2.11)
50 2 Proof of main result

Since the sequence (xj )j∈N is dense in E we see that

lim d (y, X (tnk )) = d (y, X(t)) , Pτ,x -almost surely for all y ∈ E. (2.12)
k→∞

The substitution y = X(t) in (2.12) shows that

lim (tnk , X (tnk )) = (t, X(t)) , Pτ,x -almost surely. (2.13)


k→∞

Again let f ∈ Cb ([0, T ] × E) be given. Next we consider the situation where


we have an infinite subsequence (sn , X (sn ))n∈N such that sn < sn+1 , n ∈ N.
Put s = supn∈N sn , and consider the process
Z
ρ 7→ Eρ,X(ρ) [f (s, X(s))] = P (ρ, X(ρ); s, dy) f (s, y) = P (ρ, s) f (s, ·) (X(s))
(2.14)
which is Pτ,x -martingale with respect to the filtration (Ftτ )τ ≤ρ≤s . Since the
process in (2.14) is a martingale we know that the limit

lim Esn ,X(sn ) [f (s, X(s))]


n→∞

exists. We also have


h i £ ¤
Eτ,x lim Esn ,X(sn ) [f (s, X(s))] = lim Eτ,x Esn ,X(sn ) [f (s, X(s))]
n→∞
£ £ ¯ n→∞
¤¤
= lim Eτ,x Eτ,X(τ ) f (s, X(s)) ¯ Fsτn = Eτ,x [f (s, X(s))] . (2.15)
n→∞

Like in (2.7) we write


h¯ ¯2 i
Eτ,x ¯Eρ,X(ρ) [f (s, X(s))] − f (ρ, X(ρ))¯
h¯ ¯2 i h i
= Eτ,x ¯Eρ,X(ρ) [f (s, X(s))]¯ + Eτ,x |f (ρ, X(ρ))|
2

h i
− 2<Eτ,x f (ρ, X(ρ))Eρ,X(ρ) [f (s, X(s))] . (2.16)

The expression in (2.16) converges to 0 as ρ ↑ s. Here we used the following


identity:

lim Eτ,x [g (ρ, X(ρ))] = lim P (τ, ρ) g (ρ, ·) (x)


ρ↑s ρ↑s

= P (τ, s) g (s, ·) (x) = Eτ,x [g (s, X(s))] , (2.17)


¡ ¢
Consequently, the Pτ,x -martingale Esn ,X(sn ) [f (s, X(s))] n∈N converges Pτ,x -
almost surely and in the space L2 (Ω, FTτ , Pτ,x ) to the stochastic variable
f (s, X(s)). In addition, the sequence (f (sn , X (sn )))n∈N converges in the
space L2 (Ω, FTτ , Pτ,x ) to the same stochastic variable f (s, X(s)). Then there
exists a subsequence (f (snk , X (snk )))n∈N which converges Pτ,x -almost surely
to f (s, X(s)). Again we employ the function in (2.9) to prove that
2.1 Proof of the main result: Theorem 1.39 51

lim X (snk ) = X(s), Pτ,x -almost surely. (2.18)


k→∞

The equalities (2.13) and (2.18) show that the orbit {(ρ, X(ρ)) : ρ ∈ [τ, T ]}
is Pτ,x almost surely a sequentially compact subset of E. Since the space
E is complete metrizable we infer that this orbit is Pτ,x -almost surely a
compact
n subset of E.oWe still have to show that there exists a modifica-
e
tion X(s) : s ∈ [0, T ] of the process {X(s) : s ∈ [0, T ]} which possesses left
limits, is right-continuous Pτ,x -almost surely, and is such that
h ³ ´i
e
P (τ, t) f (x) = Eτ,x [f (X(t))] = Eτ,x f X(t) , f ∈ Cb (E). (2.19)

. In order to achieve this we begin by using a modified version of the process


in (2.3):
Z ∞
t 7→ e−αt R(α)f (t, X(t)) = e−αρ P (t, ρ ∧ T ) f (ρ ∧ T, ·) (X(t)) dρ, (2.20)
t

for t ∈ [0, T ]. ¡The¢ process in (2.20) is a Pτ,x -supermartingale with respect to


the filtration Fρτ τ ≤ρ≤T . Since the process in (2.20) is a Pτ,x -supermartingale
on the interval [τ, T ] we deduce that for t varying over countable subsets its left
and right limits exist Pτ,x -almost surely. Then the process in (2.3) shares this
property as well. For a detailed argument which substantiates this claim see
the propositions 2.4 and 2.5 below. Since the orbit {(ρ, X(ρ)) : ρ ∈ [τ, T ]} is
Pτ,x -almost surely compact, and since the function f belongs to Cb ([0, T ] × E)
we infer that for sequences the process t 7→ f (t, X(t)) possesses Pτ,x -almost
surely left and right limits in E. Again an appeal to the function f in (2.9)
shows that the limits lims↑t, s∈D X(s) and limt↓s, t∈D X(t) exist Pτ,x -almost
surely for t ∈ (τ, T ] and s ∈ [τ, T ]. Here we wrote D = {k2−n : k ∈ N, n ∈ N}
for the collection of non-negative dyadic numbers. A redefinition (modifica-
e
tion) X(ρ) of the process X(ρ), ρ ∈ [0, T ], reads as follows:

e
X(ρ) = lim e ) = X(T ).
X(t), ρ ∈ [0, T ), X(T (2.21)
t↓ρ, t∈D∩(ρ,T ], t>ρ

So we obtain the following intermediate important result.


n o
Proposition 2.2. The process X(ρ) e : ρ ∈ [0, T ] is continuous from the
right and has left limits in E Pτ,x -almost surely. Moreover, its Pτ,x -distribution
coincides with that of the process
n {X(ρ)o: ρ ∈ [0, T ]}. n³ Fix (τ, x)´∈ [0, T ] × E o
e
and t ∈ [τ, T ]. On the event X(t) ∈ E the orbits e
s, X(s) : s ∈ [τ, t]
are Pτ,x -almost surely compact subsets of [τ, T ] × E.
³ ´
Fix 0 ≤ τ ≤ t ≤ T , and let S, S1 and S2 be F eτ -stopping times. In
t
t∈[τ,T ]
what follows we will make use of the following σ-fields:
52 2 Proof of main result
³ ´
e τ = σ X(ρ)
F e : τ ≤ρ≤t ;
t
³ ´
e τ = ∩0<ε≤T −t σ X(ρ)
F e : τ ≤ ρ ≤ t + ε = ∩0<ε≤T −t F eτ ; (2.22)
t+ t+ε
³³ ´ ´
e S,∨ = σ ρ ∨ S, X
F e (ρ ∨ S) : 0 ≤ ρ ≤ T ; (2.23)
T
n o
e S1 ,∨ = ∩s∈[0,T ] A ∈ F
F e S1 ,∨ : A ∩ {S2 ≤ s} ∈ F e0 ; (2.24)
S2 T s
n o
e
F S ,∨ e S1 ,∨ : A ∩ {S2 < s} ∈ F e0
S2 + = ∩s∈[0,T ] A ∈ FT
1
s
n o
= ∩0<ε≤T ∩s∈[0,T −ε] A ∈ F e S1 ,∨ : A ∩ {S2 ≤ s} ∈ F
e0
T s+ε

e S1 ,∨
= ∩ε>0 F (S2 +ε)∧T . (2.25)

The σ-field in (2.22) is called the right closure of F e τ , the σ-field in (2.23) is
t
called the σ-field after time S, the σ-field in (2.24) is called the σ-field between
time S1 and S2 , and finally the one in (2.25) is called the right closure of the
one in (2.24).
Proof (Continuation of the proof of assertion (a) of Theorem 1.39). Our most
important aim is to prove that the process
n³ ´ ³ ´ o
e τ , Pτ,x , X(t),
Ω, F e τ ≤ t ≤ T , (∨t : τ ≤ t ≤ T ) , (E, E) (2.26)
T

is a strong Markov process. We begin by proving the following equalities:


£ ¯ ¤
Es,X(s)
e [F ◦ ∨s ] = Eτ,x F ◦ ∨s ¯ Fsτ (2.27)
£ ¯ τ ¤ h ¯ τ i
= Eτ,x F ◦ ∨s ¯ Fs+ = Eτ,x F ◦ ∨s ¯ F e
s+ . (2.28)
³ ´
e
First we take F of the form F = f X(s) where f ∈ Cb (E). By an approx-
imation
³ ´argument it then follows that (2.27) and (2.28) also
h hold foriF =
e
f X(s) e
with f ∈ L∞ (E, E). So let f ∈ Cb (E). Since Ps,y X(s) =y = 1
³ ´ ³ ´
e
and f X(s) e
◦ ∨s = f X(s) we see
h ³ ´ i h ³ ´i ³ ´
Es,X(s)
e
e
f X(s) ◦ ∨s = Es,X(s)
e
e
f X(s) e
= f X(s) . (2.29)
³ ´
Since the stochastic variable f X(s)e is measurable with respect to the σ-
τ
field Fs+ by (2.29) we also have the Pτ,x -almost sure equalities:
h ³ ´ ¯ τ i h ³ ´¯ i ³ ´
e
Eτ,x f X(s) ◦ ∨s ¯ Fs+ = Eτ,x f X(s) e ¯ Fs+
τ e
= f X(s) . (2.30)

Next we calculate, while using the Markov property of the process t 7→ X(t)
and right-continuity of the function t 7→ P (s, t) f (y), s ∈ [τ, T ], y ∈ E,
2.1 Proof of the main result: Theorem 1.39 53
h ³ ´ ¯ τ i h ³ ´¯ i
e
Eτ,x f X(s) ◦ ∨s ¯ Fs+ e
= Eτ,x f X(s) ¯ Fsτ
£ ¯ ¤
= lim Eτ,x f (X(s + ε)) ¯ Fsτ = lim Es,X(s) [f (X(s + ε))]
ε↓0 ε↓0

= lim P (s, s + ε) f (·)) (X(s)) = f (X(s)) . (2.31)


ε↓0

In order to complete the


³ arguments
´ for the proof of (2.27) and (2.28) for
e
F of the form F = f X(s) , f ∈ Cb (E), we have to show the equality
³ ´
e
f X(s) = f (X(s)) Pτ,x -almost surely. This will be accomplished by the
following identities:
·¯ ³ ´ ¯2 ¸
¯ e ¯
Eτ,x ¯f X(s) − f (X(s))¯
h i h i h i
2 2
= lim Eτ,x |f (X(t))| − 2 lim <Eτ,x f (X(s))f (X(t)) + Eτ,x |f (X(s))|
t↓s t↓s

(Markov property for the process t 7→ X(t))


h i h i
2
= lim Eτ,x |f (X(t))| − 2 lim <Eτ,x f (X(s))Es,X(s) [f (X(t))]
t↓s t↓s
h i
2
+ Eτ,x |f (X(s))|

(relationship between Feller propagator and Markov property of X)


h i
2
= lim P (τ, t) |f (·)| (x) − 2 lim < P (τ, s) f (·)P (s, t) f (·) (x)
t↓s t↓s
2
+ P (τ, s) |f (·)| (x)
h i
2 2
= P (τ, s) |f (·)| (x) − 2 P (τ, s) f (·)f (·) (x) + P (τ, s) |f (·)| (x)
= 0. (2.32)
³ ´
From (2.32) we infer that f X(s) e = f (X(s)) Pτ,x -almost surely. From
(2.30), (2.31), and (2.32) we deduce
³ ´ equalities in (2.27) and (2.28) for
the
a variable F of the form F = f X(s) e , f ∈ Cb (E). An approximation argu-
ments then yields (2.27) and (2.28) for f ∈ L∞ (E, E).
In order to prove (2.27) in full generality it suffices by the Monotone Class
Theorem and an approximation argument to prove ³ the ´equalities in (2.28) for
Qn e (sj ) , where the functions
stochastic variables F of the form F = j=0 fj X
fj , 0 ≤ j ≤ n, belong³ to C´b (E) and where s = s0 < s1 < s2 < · · · < sn ≤ T .
e
Since the equality f X(s) = f (X(s)) holds Pτ,x -almost surely, it is easy to
see that by using the equalities (2.30), (2.31), ³and (2.32)
´ that it suffices to
Qn+1 e
take the variable F of the form F = j=1 fj X (sj ) where as above the
54 2 Proof of main result

functions fj , 1 ≤ j ≤ n + 1, belong to Cb (E) and where s < s1 < s2 < · · · <


sn < sn+1 ≤ T . For n = 0 we have Pτ,x -almost surely
h ³ ´¯ i £ ¯ ¤
Eτ,x f1 X e (s1 ) ¯ Fsτ = lim Eτ,x f1 (X (s1 + ε)) ¯ Fsτ
ε↓0

(Markov property of the process X)

= lim Es,X(s) [f1 (X (s1 + ε))] = lim P (s, s1 + ε) f (X(s))


ε↓0 ε↓0

= lim P (s, s1 ) P (s1 , s1 + ε) f (X(s)) = P (s, s1 ) f (X(s))


ε↓0
³ ´ h ³ ´i
e
= P (s, s1 ) f X(s) =E e f Xe (s1 ) . (2.33)
s,X(s)

³ ´
The equalities in (2.33) imply (2.27) with F = f1 X e (s1 ) where f1 ∈ Cb (E)
and s < s1 ≤ T . Then we apply induction ´ with respect to n to obtain (2.27)
Qn+1 ³ e
for F of the form F = j=1 fj X (sj ) where as above the functions fj ,
1 ≤ j ≤ n+1, belong to Cb (E) and where s < s1 < s2 < · · · < sn < sn+1 ≤ T .
In fact using the measurability of X e (sj ) with respect to the σ-field Fsτ + ,
n
1 ≤ j ≤ n, and the tower property of conditional expectation we get Pτ,x -
almost surely:
 
Y ³
n+1 ´¯
Eτ,x  fj Xe (sj ) ¯ Fsτ 
j=1
 
n
Y ³ ´ h ³ ´¯ i
= Eτ,x  fj e (sj ) Eτ,x fn+1 X
X e (sn+1 ) ¯ Fsτ Fsτ 
n
j=1

(Markov property for n = 1)


 
n
Y ³ ´ h ³ ´i
= Eτ,x  fj e (sj ) E e
X e Fsτ 
sn ,X(sn ) fn+1 X (sn+1 )
j=1

(induction hypothesis)
2.1 Proof of the main result: Theorem 1.39 55
 
n
Y ³ ´ h ³ ´i
= Es,X(s)
e
 fj Xe (sj ) E e fn+1 Xe (sn+1 ) 
sn ,X(sn )
j=1
 
n
Y ³ ´ h ³ ´¯ i
= Es,X(s)  fj e (sj ) E e
X e ¯ Fss 
e s,X(s) fn+1 X (sn+1 ) n
j=1
 
n
Y ³ ´ ³ ´
= Es,X(s)
e
 fj e (sj ) fn+1 X
X e (sn+1 ) 
j=1
 
n+1
Y ³ ´
= Es,X(s)
e
 e (sj )  .
fj X (2.34)
j=1

Qn+1 ³ e ´
So that (2.34) proves (2.27) for F = j=1 fj X (sj ) where the functions
fj , 1 ≤ n + 1, belong to Cb (E), and s < s1 < · · · < sn+1 . As remarked
above from (2.30), (2.31), and (2.32) the equality in ³ (2.27)´ then also follows
Qn e (sj ) with fj ∈ Cb (E)
for all stochastic variables of the form F = j=0 fj X
for 0 ≤ j ≤ n and 0 = s0 < s1 < · · · < sn ≤ T . By the Monotone Class
Theorem and approximation arguments it then follows that (2.27) is true for
all bounded FTτ stochastic variables F .
Next we proceed with a proof of the equalities in (2.28). Since F e τ ⊂ Fτ ,
s+ s+
τ
and the variable Es,X(s)
e [F ◦ ∨s ] is Fs+ -measurable, it suffices to prove the
first equality in (2.28), to wit
£ ¯ τ ¤
Eτ,x F ◦ ∨s ¯ Fs+ = Es,X(s)
e [F ◦ ∨s ] (2.35)

for any bounded FTτ -measurable stochastic variable F . We will not prove the
equality in (2.35) directly, but we will show the following ones instead:
£ ¯ τ ¤ £ ¯ s ¤ h ¯ s i
Eτ,x F ◦ ∨s ¯ Fs+ = Es,X(s) F ◦ ∨s ¯ Fs+ = Es,X(s) e
F ◦ ∨s ¯ F
e e s+ ,
(2.36)
under the condition that the function (s, x) 7→ P (s, t)f (x) is Borel measurable
on [τ, t] × E for f ∈ Cb (E), which is part of (vi) in Definition 1.24. In order to
prove the equalities in (2.36) it
³ suffices´ by the Monotone Class Theorem to take
Qn e (sj ) with s = s0 < s1 < · · · < sn ≤ T and
F of the form F = j=0 fj X
where de functions fj , 0 ≤ j ≤ n, are bounded Borel measurable functions.
By another approximation argument we may assume that the functions fj ,
0 ≤ j ≤ n, belong to Cb (E).
³ An ´ induction
³ argument
´ shows that it suffices to
e e
prove (2.36) for F = f0 X (s0 ) f1 X (s1 ) where s = s0 < s1 ≤ T , and
the functions f0 and f1 are members of Cb (E). The case f1 = 1 ³was taken
´
e
care of in the equalities (2.29) and (2.30). Since the variable f0 X(s) is
s
Fs+ -measurable the proof of the equalities in (2.36) reduces to the case where
56 2 Proof of main result
³ ´
e
F = f X(t) where τ < s < t ≤ T and f ∈ Cb (E). The following equalities
show the first equality in (2.36). With s < sn+1 < sn < t and limn→∞ sn = s
we have
h ³ ´¯ i h h ³ ´¯ ¯ i¯ i
e
Eτ,x f X(t) ¯ Fs+
τ
= Eτ,x Eτ,x f X(t) e ¯ Fsτ ¯ Fsn ¯ Fs+ τ
n
h h ³ ´i ¯ i
e
= Eτ,x Esn ,X(sn ) f X(t) ¯ Fs+
τ

h h ³ ´i ¯ i
= Eτ,x Esn ,X(s e
e n ) f X(t)
¯ Fs+
τ

h h ³ ´i ¯ i
= Eτ,x lim Esn ,X(se n ) f X(t)
e ¯ Fs+
τ
n→∞
h ³ ´i
= lim Esn ,X(s
e n) f e
X(t) (2.37)
n→∞
h h ³ ´i ¯ i
= Es,X(s)
e lim E e e
f X(t) ¯ Fs+
s
n→∞ sn ,X(sn )
h h ³ ´i ¯ i
= Es,X(s)
e Esn ,X(s e
e n ) f X(t)
¯ Fs+
s

h h ³ ´¯ i¯ i
= Es,X(s)
e Es,X(s)
e
e
f X(t) ¯ Fss ¯ Fs+ s
n
h ³ ´¯ i
= Es,X(s)
e
e
f X(t) ¯ Fs+
s
(2.38)
h ³ ´i
In these equalities we used the fact that the process ρ 7→ Eρ,X(ρ)e
e
f X(t) ,
s < ρ ≤ t is Ps,y -martingale for (s, y) ∈ [0, t)×E. The equality in (2.38) implies
the first equality in (2.36). The second one can be obtained by repeating the
four final steps in the proof of (2.38) with F e s instead of Fs . Here we use
s+ s+
that the stochastic variable in (2.37) is measurable with respect to the σ-field
e s , which is smaller than Fs .
F s+ s+
In order to deduce (2.35) from (2.36) we will need the full strength of
property (vi) in Definition 1.24. In fact using the representation in³ (2.37) ´
and using the continuity property in (vi) shows (2.35) for F = f X(t) e ,
f ∈ Cb (E). By the previous arguments the full assertion in (2.28) follows. In
fact Proposition 2.4 gives a detailed proof of the equalities in (2.72) below.
The equalities in (2.37) then follow from the Monotone Class Theorem.
Next we want to prove that the process t 7→ ³X(t) e
´possesses the strong
Markov property. This means that for any given Ft+ e τ
-stopping time
t∈[τ,T ]
S : Ω → [τ, T ] we have to prove an equality of the form (see (1.87))
h ¯ τ i
ES,X(S) [F ◦ ∨S ] = Eτ,x F ◦ ∨S ¯ Fe
e S+ , (2.39)

and this for all bounded FTτ -measurable stochastic variables F . By the Mono-
tone Class Theorem it follows that it suffices to³prove (2.39) for bounded
´
Qn e (sj ∨ S) where
stochastic variables F of the form F = j=0 fj sj ∨ S, X
the functions fj , 0 ≤ j ≤ n, are bounded Borel functions on [τ, T ] × E, and
2.1 Proof of the main result: Theorem 1.39 57

τ = s0 < s1 < · · · < sn ≤ T . By another approximation argument it suffices


to replace the bounded Borel functions fj , 0 ≤ j ≤ n, by bounded continuous
e τ -measurable.
functions on [τ, T ] × E. By definition the stopping time S is F S+
e τ
Let us show that X(S) is FS+ -measurable. Therefore we approximate the
stopping time S from above by stopping times Sn , n ∈ N, of the form
» ¼
T − τ 2n (S − τ )
Sn = τ + . (2.40)
2n T −τ

If t ∈ [τ, T ], then
¹ º
t−τ
2n
T[− τ ½ ¾
k−1 k
{Sn ≤ t} = (T − τ ) + τ < S ≤ (T − τ ) + τ , (2.41)
2n 2n
k=0
¡ τ ¢
and hence Sn is Ft+ t∈[τ,T ] -stopping time. Moreover, on the event
½ ¾
k−1 k
(T − τ ) + τ < S ≤ n (T − τ ) + τ
2n 2

k(T − τ )
the stopping time Sn takes the value Sn = tk,n , where tk,n = τ + .
2n
Consequently, we have the following equality of events:
½ ¾ ½ ¾
k(T − τ ) k−1 k
Sn = τ + = tk,n = (T − τ ) + τ < S ≤ n (T − τ ) + τ ,
2n 2n 2

2n (t − τ )
so that for k ≤ , which is equivalent to tk,n ≤ t, the event
½ T −τ
¾
k(T − τ ) e τ -measurable, and on this event the state vari-
Sn = τ + is F t+
2n
able Xe (Sn ) = X eτ
e (tk,n ) is F tk,n + -measurable. As a consequence we see that
on the event {Sn ≤ t} e τ -measurable. Then the
e (Sn ) is F
³ the state variable
´ X t+
space-time variable Sn , X e (Sn ) is measurable with respect to the σ-field
e τ . In addition, we have
FS+

T −τ
S ≤ Sn+1 ≤ Sn ≤ S + , (2.42)
2n
³ ´
and hence the space-time variable S, X e (S) is Fe τ -measurable as well. This
S+
³ ´
proves the equality in (2.39) in case F = f τ ∨ S, X e (τ ∨ S) where f ∈
Qn ³ ´
Cb ([τ, T ] × E). As a preparation for the case F = j=0 fj sj ∨ S, X e (sj ∨ S)
58 2 Proof of main result

where the functions fj , 0 ≤ j ≤ n, are bounded Borel functions on [τ, T ] × E,


and τ = s0 < s1 < · · · < sn ≤ T , we first consider the case (τ < t ≤ T )
³ ´ ³ ´
F = f t ∨ S, X e (t ∨ S) 1{S≤t} = f t, X(t)
e 1{S≤t} (2.43)

where f ∈ Cb ([τ, T ] × E). On the event {S ≤ t} we approximate the stopping


time S from above by stopping times Sn , n ∈ N, of the form
» ¼
t − τ 2n (S − τ )
Sn (t) = τ + n . (2.44)
2 t−τ

Then on the event {S ≤ t} we have the following inclusions of σ-fields:

e τ ∩ {S ≤ t} = F
F eτ eτ eτ
S+ S∧t+ ∩ {S ≤ t} ⊂ FSn+1 (t)+ ∩ {S ≤ t} ⊂ FSn (t)+ ∩ {S ≤ t}
(2.45)

and

∩∞ eτ eτ
n=1 FSn (t)+ ∩ {S ≤ t} = FS+ ∩ {S ≤ t} . (2.46)

Here we wrote F ∩ A0 = {A ∩ A0 : A ∈ F} when F is any σ-field on Ω and


A0 ⊂ Ω. Then we have
h ³ ´ ¯ τ i
Eτ,x f t ∨ S, Xe (t ∨ S) 1{S≤t} ¯ F e
S+
h ³ ´ ¯ τ i
= Eτ,x f t, X e (t) 1{S≤t} ¯ F e
S+
h h ³ ´ ¯ τ i¯ i
= Eτ,x Eτ,x f t, X e (t) 1{S≤t} ¯ F e ¯Feτ
Sn (t)+ S+
h h ³ ´i ¯ τ i
= Eτ,x ESn (t),X(S e
e n (t)) f t, X (t) 1{S≤t} ¯ Fe
S+
h h ³ ´i ¯ τ i
= lim Eτ,x ESn (t),X(S e
e n (t)) f t, X (t)
e
1{S≤t} ¯ F S+
n→∞
h h ³ ´i ¯ τ i
= Eτ,x lim ESn (t),X(S e
e n (t)) f t, X (t)
e
1{S≤t} ¯ F S+
n→∞

(employ (2.46) and the arguments leading to equality (2.36))


h ³ ´i
= lim ESn (t),X(S e
e n (t)) f t, X (t) 1{S≤t}
n→∞
h ³ ´¯ i
= ES,X(S)
e f t, X e S,∨ 1{S≤t}
e (t) ¯ F
S+

(appeal to (2.35) which relies on property (vi) of Definition 1.24)


h ³ ´i
= ES,X(S)
e
e (t) 1{S≤t} .
f t, X (2.47)
2.1 Proof of the main result: Theorem 1.39 59

e S,∨ -measurability of the stochastic state variable


From (2.47) and the F
³ ´ S+
e
S, X(S) we infer
h ³ ´¯ i
Eτ,x f t ∨ S, X e (t ∨ S) ¯ F eτ
S+
h ³ ´ ¯ τ i
= Eτ,x f t ∨ S, X e (t ∨ S) 1{S≤t} ¯ F e
S+
h ³ ´ ¯ τ i
+ Eτ,x f t ∨ S, X e (t ∨ S) 1{S>t} ¯ F e
S+
h ³ ´ ¯ τ i
= Eτ,x f t, X e (t) 1{S≤t} ¯ F e
S+
h ³ ´ ¯ τ i
+ Eτ,x f S, X e (S) 1{S>t} ¯ F e
S+
h ³ ´ i ³ ´
= ES,X(S)
e f t, X e (t) 1{S≤t} + f S, X e (S) 1{S>t}
h ³ ´i
= ES,X(S)
e f t ∨ S, X e (t ∨ S) . (2.48)
Qn+1 ³ ´
Next we consider the case F = j=0 fj sj ∨ S, X e (sj ∨ S) where the func-
tions fj , 0 ≤ j ≤ n + 1, are bounded Borel functions on [τ, T ] × E, and
τ = s0 < s1 < · · · < sn+1 ≤ T . From (2.48) and the F e S,∨ -measurability
³ ´ S+
of the stochastic state variable S, X(S) e we obtain (2.39) in case F =
³ ´ ³ ´
e ) f1 s1 , X
f0 τ, X(τ e (s1 ) , and thus
³ ´ ³ ´
F ◦ ∨S = f0 τ ∨ S, X(τe ∨ S) f1 s1 ∨ S, X e (s1 ∨ S) .

So that the cases n = 0 and n = 1 have been taken care of. The remaining
part of the proof uses induction. From (2.48) with the maximum operator
sn ∨ S replacing S together with the induction hypothesis we get
 
Y ³
n+1 ´¯
Eτ,x  e (sj ∨ S) ¯ FS+
fj sj ∨ S, X τ 

j=0

n
Y ³ ´
= Eτ,x  e (sj ∨ S)
fj sj ∨ S, X
j=0

h ³ ´¯ i¯
× Eτ,x fn+1 e (sn+1 ∨ S) ¯ Fsτ
sn+1 ∨ S, X ∨S+
¯ FS+
τ 
n


n
Y ³ ´
= Eτ,x  e (sj ∨ S)
fj sj ∨ S, X
j=0

h ³ ´i ¯
× Esn ∨S,X(s
e n ∨S) fn+1
e (sn+1 ∨ S) ¯ Fτ 
sn+1 ∨ S, X S+
60 2 Proof of main result

(induction hypothesis)

n
Y ³ ´
= ES,X(S)
e
 e (sj ∨ S)
fj sj ∨ S, X
j=0

h ³ ´i
× Esn ∨S,X(s
e n ∨S) fn+1
e (sn+1 ∨ S) 
sn+1 ∨ S, X

n
Y ³ ´
= ES,X(S)
e
 e (sj ∨ S)
fj sj ∨ S, X
j=0

h ³ ´¯ i
× ES,X(S)
e
e (sn+1 ∨ S) ¯ FS,∨
fn+1 sn+1 ∨ S, X 
sn ∨S+

  
n+1
Y ³ ´¯
= ES,X(S)
e
E e  fj e (sj ∨ S) ¯ FS,∨ 
sj ∨ S, X
S,X(S) sn ∨S+
j=0
 
n+1
Y ³ ´
= ES,X(S)
e
 fj e (sj ∨ S)  .
sj ∨ S, X (2.49)
j=0

The strong Markov property of the process X e follows from (2.49), an approx-
imation argument and the Monotone Class Theorem.
We still need to redefine our process and probability measures Pτ,x on
the Skorohod space D ([0, T ], E), (τ, x) ∈ [0, T ] × E in such a way that the
e is preserved. This can be done replacing (2.26)
distribution of the process X
with the collection
n³ ´ ³ ´ o
e F
Ω, eτ , P
eτ,x , X(t),
e τ ≤ t ≤ T , (∨t : τ ≤ t ≤ T ) , (E, E) (2.50)
T

where Ω eτ,x is determined by the equality E


e = D ([0, T ], E), and P e τ,x [F ] =
e
Eτ,x [F ◦ π]. Here F : Ω → C is a bounded variable which is measurable with
respect to the σ-field generated by the coordinate variables: X(t)e :ωe 7→ ωe (t),
t ∈ [τ, T ], ω e e e
e ∈ Ω. Notice that the restriction of X(t) to Ω is evaluation of
ω
e∈Ω e at t. The mapping π : Ω → Ω e is defined by π(ω) = X(t,
e ω), t ∈ [0, T ],
0 0
ω ∈ Ω . Here Ω has the property that for all (τ, x) ∈ [0, T ]×E its complement
in Ω is Pτ,x -negligible. We will describe the space Ω 0 . Let D be the collection
of positive dyadic numbers. For Ω 0 we may choose the space:

Ω 0 : = {ω ∈ Ω : t 7→ ω(t), t ∈ D ∩ [0, T ] has left and right limits in E}


∩ {ω ∈ Ω : the range {ω(t) : t ∈ D ∩ [0, T ]} is totally bounded in E} .
(2.51)
2.1 Proof of the main result: Theorem 1.39 61

Let (xj )j∈N be a sequence in E which is dense, and let d be a metric on


E ×E which turns E into a polish space. Put B (x, ε) = {y ∈ E : d(y, x) < ε}.
Define, for any finite subset of [0, T ] with an even number of members U =
{t1 , . . . , t2n } say, and ε > 0, the stochastic variable Hε (U ) by
n
X
Hε (U )(ω) = 1{d(X(t2j−1 ),X(t2j ))≥ε} (ω).
j=1

We also put

Hε (D ∩ [0, T ])
= sup {Hε (U ) : U ⊂ D ∩ [0, T ], U contains an even number of elements} .

Then the subset Ω 0 of Ω = E [0,T ] can be described as follows:


© ª
Ω 0 = ∩∞
n=1 ω ∈ Ω : H1/n (D ∩ [0, T ]) (ω) < ∞ (2.52)
\ n o
∩∞ ∞ n
m=1 ∪n=1 ω ∈ Ω : (X(s)(ω))s∈D∩[0,T ] ⊂ ∪j=1 B (xj , 1/m) .

The description in (2.52) shows that the subset Ω 0 is a measurable subset of


Ω. In addition we have Pτ,x (Ω 0 ) := Pτ,x (Ωτ0 ) = 1 for all (τ, x) ∈ [0, T ] × E.
Here
Ωτ0 = {ω ∈ Ω 0 : ω(ρ) = ω(τ ), ρ ∈ D ∩ [0, τ ]} , (2.53)
© 0
ª
which may be identified with ω ¹[τ,T ] : ω ∈ Ω which is a measurable subset
of Ωτ = E [τ,T ] . In order to complete the construction and the proof of assertion
(a) in Theorem 1.39 we need to prove the quasi-left continuity of the process
e So let (τn )
X. τ
n∈N be an increasing sequence of (Ft )t∈[τ,T ] -stopping times with
values in [τ, T ]. Put τ∞ = supn∈N τn . Let f and g be functions Cb+ (E), and
let h > 0. Then by the strong Markov property we have for m ≤ n
h ³ ´ ³ ´i
Eτ,x f X e (τm ) g X e ((τm + h) ∧ T )
h ³ ´ h ³ ´ii
= Eτ,x f X e (τm ) E e g Xe ((τm + h) ∧ T )
τm ,X(τm )
h ³ ´ h ³ ´i i
= Eτ,x f X e (τm ) E e g Xe ((τm + h) ∧ T ) , τm + h ≥ τ∞
τm ,X(τm )
h ³ ´ h ³ ´i i
+ Eτ,x f X e (τm ) E e g Xe ((τm + h) ∧ T ) , τm + h < τ∞
τm ,X(τm )

h ³ ´i
(the process ρ 7→ Eρ,X(ρ)
e
e
g X(s) is a right-continuous Pτ,x -martingale on
[τ, s])
h ³ ´ h ³ ´i i
= Eτ,x f Xe (τm ) E e g e ((τm + h) ∧ T ) , τm + h ≥ τ∞
X
τm ,X(τm )
h ³ ´ h ³ ´i i
+ Eτ,x f Xe (τm ) E e e
τm ,X(τm ) g X ((τm + h) ∧ T ) , τm + h < τ∞
62 2 Proof of main result
h ³ ´ ³ ´ i
= Eτ,x f Xe (τm ) P (τn , (τm + h) ∧ T ) g Xe (τn ) , τm + h ≥ τ∞
h ³ ´ h ³ ´i i
+ Eτ,x f Xe (τm ) E e g e ((τm + h) ∧ T ) , τm + h < τ∞ .
X
τm ,X(τm )
(2.54)

Put L = limn→∞ X e (τn ). Upon taking limits, as n → ∞, and employing the


fact that the propagator P (τ, t) is continuous from the left on the diagonal in
(2.54) we obtain:
h ³ ´ ³ ´i
Eτ,x f X e (τm ) g X e ((τm + h) ∧ T )
h ³ ´ ³ ´
= lim Eτ,x f X e (τm ) P (τn , τ∞ ) P (τ∞ , (τm + h) ∧ T ) g X e (τn ) ,
n→∞
i
τm + h ≥ τ∞
h ³ ´ h ³ ´i i
+ Eτ,x f X e (τm ) E e g e ((τm + h) ∧ T ) , τm + h < τ∞
X
τm ,X(τm )
h ³ ´ i
= Eτ,x f X e (τm ) P (τ∞ , (τm + h) ∧ T ) g (L) , τm + h ≥ τ∞
h ³ ´ h ³ ´i i
+ Eτ,x f X e (τm ) E e e
τm ,X(τm ) g X ((τm + h) ∧ T ) , τm + h < τ∞ .
(2.55)

Next we let m → ∞ in (2.55) to get


h ³ ´i
Eτ,x f (L) g X e ((τ∞ + h) ∧ T −) = Eτ,x [f (L) P (τ∞ , (τ∞ + h) ∧ T ) g (L)]
(2.56)
where we invoked property (vii) of Definition 1.24. Next we let h decrease to
zero in (2.56). This yields
h ³ ´i
Eτ,x f (L) g X e (τ∞ ) = Eτ,x [f (L) P (τ∞ , τ∞ ) g (L)] = Eτ,x [f (L) g (L)] .
(2.57)
Since f and g are arbitrary in Cb+ (E), the equality in (2.57) implies that
h ³ ´i
Eτ,x h L, Xe (τ∞ ) = Eτ,x [h (L, L)] (2.58)

for all bounded Borel measurable functions h ∈ L∞ (E × E, E ⊗ E). In particu-


lar we may take a bounded continuous metric h(x, y) = d(x, y), (x, y) ∈ E ×E.
From (2.58) it follows that
h ³ ´i
Eτ,x d L, Xe (τ∞ ) = Eτ,x [d (L, L)] = 0,

and hence
e (τn ) = X
L = lim X e (τ∞ ) , Pτ,x -almost surely. (2.59)
n→∞
2.1 Proof of the main result: Theorem 1.39 63

Essentially speaking this proves part (a) of Theorem 1.39 in case we are dealing
with conservative Feller propagators, i.e. Feller propagators with the property
that P (s, t) 1 = 1, 0 ≤ s ≤ t ≤ T . In order to be correct the process, or rather
the family of probability spaces in (2.26) has to be replaced with (2.50).
This completes the proof of Theorem 1.39 assertion (a) in case the Feller
propagator is phrased in terms of probabilities P (τ, x; t, E) = 1, 0 ≤ τ ≤ t ≤
T , x ∈ E. The case P (s, t) 1 ≤ 1 is treated in the continuation of the present
proof.

Proof (Continuation of the proof of (a) of Theorem 1.39 in case of sub-


probabilities.). We have to modify the proof in case a point of absorption
is required. Most of the proof for the case that P (τ, x; t, E) = 1 can be re-
peated with the probability transition function N (τ, x; t, B), B ∈ E4 . This
function was defined in (1.94) of Remark 1.42. However, we need to show
that the E 4 -valued process X e does not enter the absorption state 4 prior
to reentering the state space E. This requires an extra argument. We will
use a stopping time argument and Doob’s optional sampling time theorem to
achieve this: see Proposition 2.3 in which the transition function N (τ, x; t, B)
is also employed.
For further use we will also need a Skorohod space with a point of absorp-
¡ ¢[0,T ]
tion 4. The space Ω 4,0 consists of those ω ∈ E 4 whose restrictions to
D ∩ [0, T ] have left and right limits in E 4 , and which are such that for some
t = t(ω) ∈ [0, T ] the range {ω(s) : s ∈ D ∩ [0, t0 ]} is totally bounded in E for
all t0 ∈ (0, t), and such that ω(s) = 4 for s ∈ D ∩ [t(ω), T ]. Again using a
metric d4 on E 4 × E 4 which renders E 4 polish, it can be shown that Ω 4,0
¡ ¢[0,T ]
is a measurable subset of Ω = E 4 . In fact Ω 4,0 can be written as

Ω 4,0
[
= {ω ∈ Ω : s 7→ ω(s), s ∈ D ∩ [0, r] has left and right limits in E}
r∈D∩[0,T ]

\ \ [
({ω ∈ Ω : ω (D ∩ [0, r1 ]) is totally bounded in E}
m=1r1 <r2 , r2 −r1 <1/m
r1 , r2 ∈D∩[0,T ]

∩ {ω ∈ Ω : ω(s) = 4 for all s ∈ D ∩ [r2 , T ]}) . (2.60)


¡ ¢[0,T ]
From (2.60) it follows that Ω¡ 4,0 is¢ a measurable subset of Ω = E 4 .
Again it turns out that Pτ,x Ω 4,0 = 1. This fact follows from Proposition
2.3 and the fact that for all t ∈ D ∩ [0, T ]

Pτ,x [ω ∈ Ω : s 7→ ω(s), s ∈ D ∩ [0, t]


has left and right limits in E, and X(t) ∈ E]
= Pτ,x [ω ∈ Ω : ω(t) ∈ E] . (2.61)
64 2 Proof of main result

The equality in (2.61) follows in the same way as the corresponding result
in case P (τ, x; t, B), B ∈ E, but now with N (τ, x; t, B), B ∈ E4 . Again the
construction which led to the process in (2.50) can be performed to get a
strong Markov process of the form:
n³ ´ ³ ´ ¡ ¢o
e F
Ω, eτ , P
eτ,x , X(t),
e τ ≤ t ≤ T , (∨t : τ ≤ t ≤ T ) , E 4 , E4 , (2.62)
T

¡ ¢
e is the Skorohod space D [0, T ], E 4 .
where Ω
Since for functions f ∈ Cb (E) we have
Z Z
P (τ, t) f (x) = P (τ, x; t, dy) f (y) = N (τ, x; tdy) f (y) (2.63)

provided f (4) = 0, it follows that the process X e is quasi-left continuous on


its life time ζ. For the definition of N (τ, x; t, B) see Remark 1.42. In order to
be correct the process, or rather the family of probability spaces in (2.26) has
to be replaced with (2.62).
The arguments in Proposition 2.3 below then complete the proof of asser-
tion (a) of Theorem 1.39 in case the Feller propagator is phrased in terms of
sub-probabilities P (τ, x; t, E) ≤ 1, 0 ≤ τ ≤ t ≤ T , x ∈ E.
In the final part of assertion (a) of Theorem 1.39 we needed the following
proposition.
Proposition 2.3. Suppose the transition function P (τ, x; t, B), which satis-
fies the equation of Chapman-Kolmogorov, consists of sub-probability Borel
measures. Let N (τ, x; t, B), B ∈ E4 be the Feller transition function as con-
structed in Remark 1.42, which now consists of genuine Borel probability mea-
sures on the Borel field E4 of E 4 . As in (2.26) construct the corresponding
Markov process
n³ ´ ³ ´ ¡ ¢o
e τ , Pτ,x , X(t),
Ω, F e τ ≤ t ≤ T , (∨t : τ ≤ t ≤ T ) , E 4 , E4 . (2.64)
T

Fix (τ, x) ∈ [0, T ] × E and t ∈ [τ, T ]. Then the orbit


n³ ´ o
e
s, X(s) : τ ≤ s ≤ t, X(t) ∈ E

is Pτ,x -almost surely a relatively compact subset of E.


Proof. A proof can be based on a stopping time argument and Doob’s optional
sampling theorem. Let the life time ζ : Ω → [0, T ] be defined by
 n o
 inf s > 0 : X(s)
e e
= 4 , if X(s) = 4 for some s ≤ T ,
ζ=
T otherwise.
³ ´

Then ζ is an F -stopping time and we have:
t
t∈[τ,t]
2.1 Proof of the main result: Theorem 1.39 65
h i
e
Pτ,x X(t) ∈E
h h ii
= Eτ,x Pζ∧t,X(ζ∧t)
e
e
X(t) ∈E
h h i i
= Eτ,x Pζ∧t,X(ζ∧t)
e
e
X(t) ∈E ,ζ≤t
h h i i
+ Eτ,x Pζ∧t,X(ζ∧t)
e
e
X(t) ∈E ,ζ>t
h h i i h h i i
= Eτ,x Pζ,X(ζ)
e
e
X(t) ∈ E , ζ ≤ t + Eτ,x Pt,X(t)
e
e
X(t) ∈E ,ζ>t
h h i i h i
= Eτ,x Pζ,4 X(t) e e
∈ E , ζ ≤ t + Pτ,x X(t) ∈ E, ζ > t

(see Remark 1.42)


h i
e
= Eτ,x [N (ζ, 4; t, E) , ζ ≤ t] + Pτ,x X(t) ∈ E, ζ > t
h i
e
= Pτ,x X(t) ∈ E, ζ > t . (2.65)
n o
e
From (2.65) it follows that on the event X(t) ∈ E the orbits
n³ ´ o
e
s, X(s) : s ∈ [τ, t]

are Pτ,x -almost surely contained in compact subsets of [τ, t] × E.

In the proof of Proposition 2.5 we need the following result. Notice that in
this Proposition as well as in Proposition 2.5 the conservative property (2.66)
is employed. Proposition 2.3 contains a result which can be used in the non-
conservative situation. The possibility of non-conservativeness plays a role in
the proof of item (d) as well: see the inequalities in (2.119) and (2.120), and
their consequences.
Proposition 2.4. Let (τ, x) be an element in [0, T ] × E, and assume

Pτ,x [X(t) ∈ E] = P (τ, t) 1E (x) = P (τ, x; t, E) = 1 (2.66)

for all t ∈ [τ, T ]. Let (fm )m∈N be a sequence in Cb+ ([τ, T ] × E) which decreases
pointwise to zero. Denote by D the collection of positive dyadic numbers. Then
the following equality holds Pτ,x -almost surely:

inf sup sup Es,X(s) [fm (t, X(t))] = 0. (2.67)


m∈N t∈D∩[τ,T ] s∈D∩[0,t]

Consequently, the collection of linear functionals Λs,t : Cb ([τ, T ] × E) → C


defined by Λs,t (f ) = Es,X(s) [f (t, X(t))], f ∈ Cb ([τ, T ] × E), τ ≤ s ≤ t ≤ T ,
s, t ∈ D, is Pτ,x -almost surely equi-continuous for the strict topology Tβ .
Let (sn , tn ) be any sequence in [τ, T ] × [τ, T ] such that sn ≤ tn , n ∈ N.
Then the collection
66 2 Proof of main result

{Λs,t : τ ≤ s ≤ t ≤ T, s, t ∈ D or (s, t) = (sn , tn ) for some n ∈ N}

is Pτ,x -almost surely equi-continuous as well.


Proof. Let (fm )m∈N ⊂ Cb+ ([τ, T ] × E) be as in Proposition 2.4. For every
m ∈ N and t ∈ [τ, T ] we define the Pτ,x -martingale s 7→ Mt,m (s), s ∈ [τ, T ],
by Mt,m (s) = Es∧t,X(s∧t) [fm (t, X(t))]. Then the process

s 7→ sup Es∧t,X(s∧t) [fm (t, X(t))] = sup Es∧t,X(s∧t) [fm (t, X(t))]
t∈[τ,T ] t∈D∩[τ,T ]

is a Pτ,x -submartingale. Fix η > 0. By Doob’s submartingale inequality we


have
" #
ηPτ,x sup sup Es,X(s) [fm (t, X(t))] ≥ η
t∈D∩[τ,T ] s∈D∩[t,T ]
" #
= ηPτ,x sup sup Mm,t (s) ≥ η
t∈D∩[τ,T ] s∈D∩[τ,T ]
" #
= ηPτ,x sup sup Mm,t (s) ≥ η
s∈D∩[τ,T ] t∈D∩[τ,T ]
" # " #
≤ Eτ,x sup Mm,t (T ) = Eτ,x sup Et,X(t) [fm (t, X(t))]
t∈D∩[τ,T ] t∈D∩[τ,T ]
" #
= Eτ,x sup fm (t, X(t)) . (2.68)
t∈D∩[τ,T ]

Since the orbit {(t, X(t)) : t ∈ D ∩ [τ, T ]} is Pτ,x -almost surely contained in a
compact subset of E, Dini’s lemma implies that

sup fm (t, X(t)) decreases to 0 Pτ,x -almost surely,


t∈D∩[τ,T ]

which implies " #


lim Eτ,x sup fm (t, X(t)) = 0. (2.69)
m→∞ t∈D∩[τ,T ]

A combination of (2.68) and (2.69) yields (2.67). So the first part of Proposi-
tion 2.4 has been established.
The second assertion follows from (2.67) together with Theorem 1.8.
The third assertion follows from the fact that for f ∈ Cb+ ([τ, T ] × E) and
τ ≤ sn ≤ tn ≤ T the inequality

Esn ,X(sn ) [f (tn , X (tn ))] ≤ sup sup Es,X(s) [f (t, X (t))]
t∈D∩[τ,T ] s∈D∩[τ,t]

holds Pτ,x -almost surely.


This shows Proposition 2.4.
2.1 Proof of the main result: Theorem 1.39 67

The next proposition was used in the proof of item (a) of Theorem 1.39.
Proposition 2.5. Let (τ, x) ∈ [0, T ] × E, and assume the conservative prop-
erty (2.66). In addition, let f ∈ Cb ([0, T ] × E) and let ((sn , tn ))n∈N be se-
quence in [τ, T ] × [τ, T ] such that sn ≤ tn , n ∈ N and such that lim (sn , tn ) =
n→∞
(s, t). Then the limit
lim Esn ,X(sn ) [f (tn , X (tn ))] = lim [P (sn , tn ) f (tn , ·)] (X (sn ))
n→∞ n→∞
= [P (s, t) f (t, ·)] (X (s)) = Es,X(s) [f (t, X(t))] (2.70)
exists Pτ,x -almost surely. In particular if sn = tn for all n ∈ N, then s = t
and
lim Etn ,X(tn ) [f (tn , X (tn ))] = lim f (tn , X (tn ))
n→∞ n→∞
= f (t, X(t)) , Pτ,x -almost surely. (2.71)
In addition, by taking tn = t and letting the sequence (sn )n∈N decrease or
increase to s ∈ [τ, t] it follows that the process s 7→ Es,X(s) [f (t, X(t))] is Pτ,x -
almost surely a left and right continuous martingale. Moreover, the equalities
£ ¯ τ ¤ £ ¯ ¤
Eτ,x f (t, X(t)) ¯ Fs+ = Es,X(s) [f (t, X(t))] = Eτ,x f (t, X(t)) ¯ Fsτ (2.72)
hold Pτ,x -almost surely.
The equalities in (2.37) then follow from (2.72) together with the Monotone
Class Theorem.
Proof. In the proof of Proposition 2.5 we will employ the properties of the
process in (2.5) to its full extent. In addition we will use Proposition 2.4 which
implies that continuity properties of the process
Z ∞
(s, t) 7→α e−α(ρ−t) Es,X(s) [f (ρ ∧ T, X (ρ ∧ T ))] dρ
t
Z ∞
=α e−α(ρ−t) P (s, ρ ∧ T ) f (ρ ∧ T, ·) (X(s)) dρ
t
= αP (s, t) R(α)f (t, ·) (X(s))
Z ∞ h ³³ ρ´ ³³ ρ´ ´´i
= e−ρ Es,X(s) f t + ∧ T, X t + ∧T dρ, (2.73)
0 α α
0 ≤ s ≤ t ≤ T , Pτ,x -almost surely carry over to the process
(s, t) 7→P (s, t) f (t, ·) (X(s)) = Es,X(s) [f (t, X(t))]
Z ∞
= lim α e−α(ρ−t) P (s, ρ ∧ T ) f (ρ ∧ T, ·) (X(s)) dρ
α→∞ t
Z ∞ ³ ³ ρ´ ´ ³³ ρ´ ´
= lim e−ρ P s, t + ∧T f t+ ∧ T, · (X(s)) dρ
α→∞ 0 α α
Z ∞ h ³³ ´ ³³
ρ ρ´ ´´i
= lim e−ρ Es,X(s) f t + ∧ T, X t + ∧T dρ.
α→∞ 0 α α
(2.74)
68 2 Proof of main result

Let (sn , tn )n∈N be a sequence in [τ, T ] × [τ, T ] for which sn ≤ tn . Put


Z ∞
Λα,s,t f = α e−α(ρ−t) P (s, ρ ∧ T ) f (ρ ∧ T, ·) (X(s)) dρ.
t

The equality in (2.73) in conjunction with Proposition 2.4 shows that the
collection of functionals
{Λα,s,t : τ ≤ s ≤ t ≤ T, s, t ∈ D or (s, t) = (sn , tn ) for some n ∈ N, α ≥ 1}
is Pτ,x -almost surely Tβ -equi-continuous. Therefore the family of its limits
Λt,s = limα→∞ Λα,s,t inherits the continuity properties from the family
{Λα,s,t : τ ≤ s ≤ t ≤ T, s, t ∈ D or (s, t) = (sn , tn ) for some n ∈ N}
where α ∈ (0, ∞) is fixed.
We still have to prove that
lim Esn ,X(sn ) [f (tn , X (tn ))] = Es,X(s) [f (t, X(t))] (2.75)
n→∞

Pτ,x -almost surely, whenever f ∈ Cb ([τ, t] × E) and the sequence (sn , tn )n∈N
in [τ, T ] × [τ, T ] is such that limn→∞ (sn , tn ) = (s, t) and sn ≤ tn for all n ∈ N.
In view of the first equality in (2.20) and the previous arguments it suffices
to prove this equality for processes of the form
Z ∞
(s, t) 7→ α e−α(ρ−t) Es,X(s) [f (ρ ∧ T, X (ρ ∧ T ))] dρ
t

instead of
(s, t) 7→ Es,X(s) [f (t, X (t))]
It is easy to see that this convergence reduces to treating the case where, for
ρ ∈ (τ, T ] fixed and for sn → s, s ∈ [τ, ρ],
lim Esn ,X(sn ) [f (ρ, X(ρ))] = Es,X(s) [f (ρ, X(ρ))] (2.76)
n→∞

Here we will distinguish two cases: sn increases to s and sn decreases to s.


In both cases we will prove that the equality in (2.76). In case of an in-
creasing the result follows more or less directly from the martingale mar-
tingale property and from the left continuity on the diagonal. In case of
a decreasing sequence we employ the fact that a subspace of the form
{P (ρ, u) g : u ∈ (ρ, T ], g ∈ Cb (E)} is Tβ -dense in Cb (E). First we consider
the situation where sn increases to s ∈ [τ, ρ]. Then we have
£ ¯ ¤
Esn ,X(sn ) [f (ρ, X(ρ))] = Eτ,x f (ρ, X(ρ)) ¯ Fsτn
£ £ ¯ ¤¯ ¤
= Eτ,x Eτ,x f (ρ, X(ρ)) ¯ Fsτ ¯ Fsτn
£ ¯ ¤
= Eτ,x Es,X(s) [f (ρ, X(ρ))] ¯ Fsτn
£ ¤
= Esn ,X(sn ) Es,X(s) [f (ρ, X(ρ))]
= (P (sn , s) Es,· [f (ρ, X(ρ))]) (X (sn )) . (2.77)
2.1 Proof of the main result: Theorem 1.39 69

In (2.77) we let n → ∞ and use the left continuity of the propagator (see
property (v) in Definition 1.24) to conclude

lim Esn ,X(sn ) [f (ρ, X(ρ))] = Es,X(s) [f (ρ, X(ρ))] . (2.78)


n→∞

The equality in (2.78) shows the Pτ,x -almost sure left continuity of the process
s 7→ Es,X(s) [f (ρ, X(ρ))] on the interval [τ, ρ]. Next assume that the sequence
(sn )n∈N decreases to s ∈ [τ, ρ]. Then we get Pτ,x -almost surely

Es,X(s) [f (ρ, X(ρ))] = P (s, ρ) f (ρ, ·) (X(ρ))

(employ (vi) of Definition 1.24)

= lim P (sn , ρ) f (ρ, ·) (X (sn )) = lim Esn ,X(sn ) [f (ρ, X(ρ))]


n→∞ n→∞
h ¯ s i
= Es,X(s) lim Esn ,X(sn ) [f (ρ, X(ρ))] ¯ Fs+
n→∞
£ ¯ s ¤
= lim Es,X(s) Esn ,X(sn ) [f (ρ, X(ρ))] ¯ Fs+
n→∞
£ £ ¯ ¤¯ s ¤
= lim Es,X(s) Es,X(s) f (ρ, X(ρ)) ¯ Fssn ¯ Fs+
n→∞

(tower property of conditional expectation)


£ ¯ s ¤
= Es,X(s) f (ρ, X(ρ)) ¯ Fs+
£ £ ¯ s ¤¯ τ ¤
= Eτ,x Es,X(s) f (ρ, X(ρ)) ¯ Fs+ ¯ Fs+
h ¯ τ i
= Eτ,x lim Esn ,X(sn ) [f (ρ, X(ρ))] ¯ Fs+
n→∞
£ ¯ τ ¤
= lim Eτ,x Esn ,X(sn ) [f (ρ, X(ρ))] ¯ Fs+
n→∞

(Markov property)
£ £ ¯ ¤¯ τ ¤
= lim Eτ,x Eτ,x f (ρ, X(ρ)) ¯ Fsτn ¯ Fs+
n→∞

(tower property of conditional expectation)


£ ¯ τ ¤
= Eτ,x f (ρ, X(ρ)) ¯ Fs+ . (2.79)

The equality in (2.79) is the same as the first equality in (2.72). The second
equality is a consequence of the Markov property with respect to the filtration
(Ftτ )t∈[τ,T ] .
This completes the proof of Proposition 2.4.

2.1.2 Proof of item (b) of Theorem 1.39

Here we have to prove that Markov processes with certain continuity proper-
ties give rise to Feller evolutions.
70 2 Proof of main result

Proof (Proof of item (b) in Theorem 1.39.). Let the operators P (τ, t), τ ≤ t,
be as in (1.89). We have to prove that this collection is a Feller evolution. The
properties (i), (iii) and (iv) of Definition 1.24 are obvious. The propagator
property (ii) is a consequence of the Markov property of the process in (1.88).
To be precise, let f ∈ Cb (E) and 0 ≤ τ < s < t ≤ T . Then we have:
£ ¤
P (τ, s) P (s, t) f (x) = Es,x [P (s, t) f (X(s))] = Eτ,x Es,X(s) [f (X(t))]
£ £ ¯ ¤¤
= Eτ,x Eτ,x f (X(t)) ¯ Fsτ
= Eτ,x [f (X(t))] = P (τ, t) f (x). (2.80)

Let f be any function in Cb (E). The continuity of the function (τ, t, x) 7→


P (τ, t) f (x), 0 ≤ τ ≤ t ≤ T , x ∈ E, implies the properties (v) through (vii)
of Definition 1.24. Let f ∈ Cb ([0, T ] × E). In addition we have to prove that
the function (τ, t, x) 7→ P (τ, t) f (t, ·) (x) is continuous. The proof of this fact
requires the following steps:
1. The Feller evolution {P (τ, t) : 0 ≤ τ ≤ t ≤ T } is Tβ -equi-continuous.
2. Define the operators R(α) : Cb ([0, T ] × E) → Cb ([0, T ] × E), α > 0, as
in (3.6) below;
Z ∞
R(α)f (t, x) = e−α(ρ−t) P (t, ρ ∧ T ) f (ρ ∧ T, ·) (x)dρ, (2.81)
t

f ∈ Cb ([0, T ] × E). Then the functions (τ, t, x) 7→ P (τ, t) [R(α)f (t, ·)] (x),
0 ≤ τ ≤ t ≤ T , x ∈ E, α > 0, are continuous for all f ∈ Cb ([0, T ] × E).
3. The family {R(α) : α > 0} is a resolvent family, and hence the range of
R(α) does not depend on α > 0. The Tβ -closure of its range coincides
with Cb ([0, T ] × E).
From 3, 1 and 2 it then follows that functions of the form P (τ, t) f (t, ·) (x),
0 ≤ τ ≤ t ≤ T , f ∈ Cb ([0, T ] × E), are continuous. So we have to prove 1
through 3.
Let (ψm )m∈N be a sequence of functions in C+ (E) which decreases point-
wise to zero. Since, by assumption, the functions (τ, t, x) 7→ P (τ, t) ψm (x),
m ∈ N, are continuous, the sequence P (τ, t) ψm (x) decreases uniformly on
compact subsets to 0. By Theorem 1.18 it follows that the Feller evolution
{P (τ, t) : 0 ≤ τ ≤ t ≤ T } is Tβ -equi-continuous. This proves 1.
Let f ∈ Cb ([0, T ] × E), and fix α > 0. Then the function P (τ, t) R(α)f
can be written in the form
Z ∞
P (τ, t) [R(α)f (t, ·)] (x) = e−α(ρ−t) P (τ, ρ ∧ T ) f (ρ ∧ T, ·) (x)dρ,
t

which by inspection is continuous, because for fixed ρ ∈ [0, T ] the function


(τ, x) 7→ P (τ, ρ) f (ρ, ·) (x) is continuous. This proves assertion 2.
The family {R(α) : α > 0} is a resolvent family, i.e. it satisfies:

R(β) = R(α) + (α − β)R(α)R(β), α, β > 0. (2.82)


2.1 Proof of the main result: Theorem 1.39 71

Consequently, the range R(α)Cb ([0, T ] × E) does not depend on α > 0. Next
fix f ∈ Cb ([0, T ] × E). Then limα→∞ αR(α)f (t, x) = f (t, x) for all (t, x) ∈
[0, T ] × E. By dominated convergence it also follows that
Z
lim αR(α)f (t, x) dµ(t, x)
α→∞ [0,T ]×E
Z Z ³ ³

ρ´ ´ ³³ ρ´ ´
= lim e−ρ P t, t + ∧T f t+ ∧ T, · (x)dρ dµ(t, x)
α→∞ [0,T ]×E 0 α α
Z
= f (t, x) dµ(t, x), (2.83)
[0,T ]×E

where µ is a complex Borel measure on [0, T ] × E. From (2.83) and Corol-


lary 1.5 we see that the space R(α)Cb ([0, T ] × E) is Tβ -weakly dense in
Cb ([0, T ] × E). It follows that it is Tβ -dense. Let K be a compact subset
of E. Since the Feller evolution is Tβ -equi-continuous there exists a bounded
function u ∈ H + ([0, T ] × E) such that

sup |P (τ, t) f (x)| ≤ kuf k∞ , f ∈ Cb (E). (2.84)


(t,x)∈[0,T ]×E

Fix ε > 0. For α0 > 0 and f ∈ Cb ([0, T ] × E) fixed, there exists a function
g ∈ Cb ([0, T ] × E) such that

sup sup |u(s, y)f (s, y) − α0 R (α0 ) g (s, y)| ≤ ε. (2.85)


s∈[0,T ] y∈E

From (2.84) and (2.85) we infer:

sup sup |P (τ, t) [f (t, ·) − α0 R (α0 ) g (t, ·)] (x)|


0≤τ ≤t≤T x∈K

≤ sup sup |u(y) (f (s, y) − α0 R (α0 ) (s, y))| ≤ ε. (2.86)


0≤s≤T y∈E

As a consequence of (2.86) the function (τ, t, x) 7→ P (τ, t) f (t, ·) (x) inherits


its continuity properties from functions of the form

(τ, t, x) 7→ P (τ, t) R (α0 ) f (t, ·) (x), 0 ≤ τ ≤ t ≤ T, x ∈ E.

Since the latter functions are continuous, the same is true for the function
P (τ, t) f (t, ·) (x).
This concludes the proof of part (b) of Theorem 1.39.
As a corollary we mention the following: its proof follows from the argu-
ments leading to the observation that for all f ∈ Cb ([0, T ] × E) the function
(τ, t, x) 7→ P (τ, t) f (t, ·) (x) is continuous. It will be used in the proof of The-
orem 3.10 below.
72 2 Proof of main result

Corollary 2.6. Let the family {P (τ, t) : 0 ≤ τ ≤ t ≤ T } be a Feller evolu-


tion in Cb (E). Extend these operators to the space Cb ([0, T ] × E) by the for-
mula
n Pe (τ, t) f (τ, x) = P o
(τ, t) f (t, ·) (x), f ∈ Cb ([0, T ] × E). Then the family
e
P (τ, t) : 0 ≤ τ ≤ t ≤ T is again Tβ -equi-continuous. In addition define the
Tβ -continuous semigroup {S(t) : t ≥ 0} on Cb ([0, T ] × E) by

S(t)f (τ, x) = P (τ, (τ + t) ∧ T ) f ((τ + t) ∧ T, ·) (x),


f ∈ Cb ([0, T ] × E) .
(2.87)
Then the semigroup {S(t) : t ≥ 0} is Tβ -equi-continuous.

In the sequel we will not use the notation Pe (τ, t) for the extended Feller
evolution very much: we will simply ignore the difference between Pe (τ, t) and
P (τ, t). For more details on the semigroup defined in (2.87) see (3.5) below.
Proof (Proof of Corollary 2.6.). Let f ∈ Cb ([0, T ] × E). From the proof of
(b) of Theorem 1.39 (see the very end) we infer that the function (τ, t, x) 7→
Pe (tau, t) f (τ, x) is continuous. Let (ψm )m∈N be a sequence of functions in
Cb ([0, T ] × E) which decreases pointwise to 0. Let u ∈ H + ([0, T ] × [0, T × E).
Then the functions Pe (τ, t) (ψm f ) (x) alson decrease uniformlyo to 0. From
Corollary 1.19 it follows that the family Pe (τ, t) : τ ≤ t ≤ T is Tβ -equi-
continuous. From the representation (2.87) of the semigroup {S(t) : t ≥ 0}, it
is also clear that this semigroup is Tβ -equi-continuous.

2.1.3 Proof of item (c) of Theorem 1.39

In this part and in (d) of Theorem 1.39 we will see the intimate relationship
which exists between solutions to the martingale problem and the correspond-
ing (strong) Markov processes.
Proof (Proof of part (c) of Theorem 1.39.). In the proof of item (c) we will
use the fact that an operator L generates a Feller evolution if and only if it
generates the corresponding Markov process: see Proposition 3.1 below. So
we may assume that the corresponding Markov process is that of part (a)
of Theorem 1.39: see (1.84). Among other things this means that it is right
continuous, and has left limits in E in its life time. In addition, it is quasi-left
continuous on its life time. Let f ∈ Cb ([0, T ] × E) belong to the domain of
D1 +L. We will show that the process in (1.90) is a Pτ,x -martingale. Therefore,
fix s ∈ [τ, t], and put
Z sµ ¶

Mτ,f (s) = f (s, X(s)) − f (τ, X(τ )) − + L(ρ) f (ρ, ·) (X(ρ)) dρ.
τ ∂ρ

Then by the Markov property we have


£ ¯ ¤
Eτ,x Mτ,f (t) ¯ Fsτ − Mτ,f (s)
2.1 Proof of the main result: Theorem 1.39 73
£ ¯ ¤
= Eτ,x Ms,f (t) ¯ Fsτ = Es,X(s) [Ms,f (t)]
= Es,X(s) [f (t, X(t))] − Es,X(s) [f (s, X(s))]
Z t ·µ ¶ ¸

− Es,X(s) + L(ρ) f (ρ, ·) (X(ρ)) dρ
s ∂ρ

(the operator L generates the involved Markov process)


Z t
d
= Es,X(s) [f (t, X(t))] − Es,X(s) [f (s, X(s))] − Es,X(s) [f (ρ, X(ρ))] dρ
s dρ
¯ρ=t
= Es,X(s) [f (t, X(t))] − Es,X(s) [f (s, X(s))] − Es,X(s) [f (ρ, X(ρ))] ¯ρ=s = 0.
(2.88)

The equality in (2.88) proves the first part of assertion (b). Proposition 2.7
below proves more than what is claimed in (b) of Theorem 1.39. Therefore
the proof of item (b) in Theorem 1.39 is completed by 2.7.

Proposition 2.7. Let the Markov family of probability spaces be as in (a),


formula (1.84) of Theorem 1.39. Let ∨t , ∧t , ϑt : Ω → Ω, t ∈ [0, T ],
be time transformations with the following respective defining properties:
X(s) ◦ ∨t = X (s ∨ t), X(s) ◦ ∧t = X (s ∧ t), and X(s) ◦ ϑt = X ((s + t) ∧ T ),
for all s, t ∈ [0, T ]. Let the σ-fields Ftt21 , 0 ≤ t1 ≤ t2 ≤ T , be defined by
Ftt21 = σ (X(s) : t1 ≤ s ≤ t2 ). Fix t ∈ [0, T ]. Then the mapping ∨t is Ftt21∨t ∨t
-
t1 t1 ∧t t1 (t +t)∧T
Ft2 -measurable, the mapping ∧t is Ft2 ∧t -Ft2 -measurable, and ϑt is F(t2 +t)∧T t -
1

Ftt21 -measurable.
Fix τ ∈ [0, T ], and τ ≤ t1 ≤ t2 ≤ T . Let µ be a Borel probability measure
on E, and define the probability measure Pτ,µ on FTτ by the formula Pτ,µ (A) =
R ¡ ¢τ,µ
P (A)dµ(x), A ∈ FTτ . Let Ftt21
E τ,x
be the Pτ,µ -completion of the σ-field
Ftt21 . Then (Pτ,µ -a.s. means Pτ,µ -almost surely)
¡ t1 ¢τ,µ n τ,µ
o
F t2 = A ∈ (FTτ ) : 1A ◦ ∨t1 ◦ ∧t2 = 1A , Pτ,µ -a.s. , (2.89)

and
¡ ¢τ,µ \ n τ,µ
o
Ftt21+ = A ∈ (FTτ ) : 1A ◦ ∨t1 ◦ ∧t2 +ε = 1A , Pτ,µ -a.s. .
ε∈(0,T −t2 ]
(2.90)

In addition the following equalities are Pτ,µ -almost surely valid for all bounded
τ,µ
stochastic variables F which are (FTτ ) -measurable:
£ ¯ τ ¤ £ ¯ ¤
Eτ,µ F ¯ Ft+ = Eτ,µ F ¯ Ftτ , (2.91)
h ¯¡ ¢ τ,µ i £ ¯ τ ¤
Eτ,µ F ¯ Ft+ τ = Eτ,µ F ¯ Ft+ . (2.92)
74 2 Proof of main result
¡ τ ¢τ,µ
If the variable F is Ft+ -measurable, then the equalities
h ¯¡ ¢τ,µ i £ ¯ τ ¤ £ ¯ ¤
F = Eτ,µ F ¯ Ft+τ = Eτ,µ F ¯ Ft+ = Eτ,µ F ¯ Ftτ (2.93)

τ,µ
hold Pτ,µ -almost surely. If the bounded stochastic variable F is (FTt ) -
measurable, then Pτ,µ -almost surely
h ¯¡ ¢τ,µ i
Eτ,µ F ¯ Ft+τ = Et,X(t) [F ] . (2.94)

¡ t ¢τ,µ
Finally, if F is Ft+ -measurable, then

F = Et,X(t) [F ] , Pτ,µ -almost surely. (2.95)

In particular such variables are Pτ,x -almost surely functions of the space-time
variable (t, X(t)).

Proof. Let F be a bounded Fss21 -measurable variable. The measurability prop-


erties of the time operator ∨t follow from the fact that F ◦ ∨t is Fss21∨t∨t
-
measurable. Similar statements hold for the operators ∧t and ϑt .
The equality

Ftt21 = {A ∈ FTτ : 1A ◦ ∨t1 ◦ ∧t2 = 1A , Pτ,µ -a.s.} (2.96)

is clear, and so the left-hand is included in the right-hand side of (2.89). This
τ,µ
can be seen as follows. Let A ∈ Ftt21 . Then there exist subsets A1 and
A2 ∈ Ftt21 such that A1 ⊂ A ⊂ A2 and Pτ,µ [A2 \ A1 ] = 0. Then we have

1A1 − 1A2 = 1A1 − 1A2 ◦ ∨t1 ◦ ∧t2


≤ 1A − 1A ◦ ∨t1 ◦ ∧t2 ≤ 1A2 − 1A1 ◦ ∨t1 ◦ ∧t2 = 1A2 − 1A1 . (2.97)

From (2.97) we see that 1A −1A ◦∨t1 ◦∧t2 and hence the left-hand side of (2.89)
is
n includedτ,µ
in the right-hand side. Since by othe same argument the σ-field
τ
A ∈ (FT ) : 1A ◦ ∨t1 ◦ ∧t2 = 1A , Pτ,µ -a.s. is Pτ,µ -complete and since
¡ ¢
{A ∈ (FTτ ) : 1A ◦ ∨t1 ◦ ∧t2 = 1A , Pτ,µ -a.s.} ⊂ Ftt21 , (2.98)

we also obtain that the right-hand side of (2.89) is contained in the left-hand
side. The equality in (2.90) is an immediate consequence of (2.89), and the
definition of Ftt21+ .
By the Monotone Class Theorem and an approximation Qn argument the
proof of (2.91) can be reduced to the case where F = j=1 fj (X (tj )) where
τ ≤ t1 < · · · tk ≤ t < tk+1 < · · · < tn ≤ T , and fj ∈ Cb (E), 1 ≤ j ≤ n.
Then by properties of conditional expectation and the Markov property with
respect to the filtration (Ftτ )t∈[τ,T ] we have
2.1 Proof of the main result: Theorem 1.39 75
 
n
Y
£ ¯ τ ¤ ¯ τ
Eτ,µ F ¯ Ft+ = Eτ,µ  fj (X (tj )) ¯ Ft+ 
j=1
   
k
Y n
Y ¯ τ ¯ τ
= fj (X (tj )) Eτ,µ Eτ,µ  fj (X (tj )) ¯ Ftk+1  ¯ Ft+ 
j=1 j=k+1

(Markov property)

k
Y £ ¯ τ ¤
= fj (X (tj )) Eτ,µ g (X (tk+1 )) ¯ Ft+ , (2.99)
j=1
hQ i
n
where g(y) = fk+1 (y)Etk+1 ,y j=k+1 fj (X (tj )) . Again we may suppose
that the function g belongs to Cb (E). Then we get, for t < s < tk+1 ,
£ ¯ τ ¤ £ £ ¯ ¤¯ τ ¤
Eτ,µ g (X (tk+1 )) ¯ Ft+ = Eτ,µ Eτ,µ g (X (tk+1 )) ¯ Fsτ ¯ Ft+

(Markov property)
£ ¯ τ ¤
= Eτ,µ Es,X(s) [g (X (tk+1 ))] ¯ Ft+
£ ¯ τ ¤
= lim Eτ,µ Es,X(s) [g (X (tk+1 ))] ¯ Ft+
s↓t
£ ¯ τ ¤
= Eτ,µ Et,X(t) [g (X (tk+1 ))] ¯ Ft+
= Et,X(t) [g (X (tk+1 ))]

(again Markov property)

= Eτ,µ [g (X (tk+1 )) Ftτ ] . (2.100)

Inserting the result of (2.100) into (2.99) and reverting the arguments which
τ τ
led
Qn to (2.99) with Ft instead of Ft+ shows the equality in (2.91) for F =
j=1 fj (X (tj )) where the functions fj , 1 ≤ j ≤ n, belong to Cb (E). As men-
tioned earlier this suffices to obtain (2.91) for all bounded random variables
τ,µ
F which are (FTτ ) -measurable. Here we use the fact that for any σ-field
τ,µ τ,µ
F ⊂ (FTτ ) , and any bounded £(FTτ¯) ¤ -measurable stochastic variable F an
equality of the form F = Eτ,µ F ¯ F holds Pτ,µ -almost surely. This argu-
ment also shows that the equality in (2.92) is a consequence of (2.91). The
equalities in (2.93) follow from the definition of conditional expectation and
the equalities (2.91) and (2.92). The equality in (2.94) also follows from (2.91)
and (2.92) together with the Markov property. Finally, the equality in (2.100)
is a consequence of (2.99) and the definition of conditional expectation.
Altogether this proves Proposition 2.7.
76 2 Proof of main result

2.1.4 Proof of item (d) of Theorem 1.39

In this subsection we will establish the fact that unique solutions to the mar-
tingale problem yield strong Markov processes.
Proof (Proof of item (d) of Theorem 1.39.). The proof of this result is quite
technical. The first part follows from a well-known theorem of Kolmogorov
on projective systems of measures. In the second part we must show that
the indicated path space has full measure, so that no information is lost. The
techniques used are reminiscent the material found in for example Blumenthal
and Getoor [34], Theorem 9.4. p. 46. The result in (d) is a consequence of the
propositions 2.8, 2.9, and 2.11 below.
In (d) as anywhere else in the book L = {L(s) : 0 ≤ s ≤ T } is considered as a
linear operator with domain D(L) and range R(L) in the space Cb ([0, T ] × E).
Suppose that the domain D(L) of L is Tβ -dense in Cb ([0, T ] × E). The prob-
lem we want to address is the following. Give necessary and sufficient condi-
tions on the operator L in order that for every (τ, x) ∈ [0, T ] × E there exists
a unique probability measure Pτ,x on FTτ with the following properties:
(i) For every f ∈ D(L), which is C (1) -differentiable in the time variable the
process
Z t
f (t, X(t)) − f (τ, X(τ )) − (D1 f + Lf ) (s, X(s)) ds, t ∈ [τ, T ],
τ

is a Pτ,x -martingale;
(ii) Pτ,x [X(τ ) = x] = 1.
¡ ¢
Here we suppose Ω = D [0, ∞], E 4 is the Skohorod space associated with
E 4 , as described in Definition 1.32, and FTτ is the σ-field generated by the
state variables X(t), t ∈ [τ, T ]. The probability measures Pτ,x are defined on
the σ-field FTτ . The following procedure extends them to FT0 . If the event A
belongs to FT0 , then we put Pτ,x [A] = Eτ,x [1A ◦ ∨τ ]. The composition 1A ◦ ∨τ
is defined in (1.74). With this convention in mind the equality in (ii) may be
replaced by
(ii)0Pτ,x [X(s) = x] = 1 for all s ∈ [0, τ ].
Let P (Ω) be the set of all probability measures on FT0 and define the subset
P00 (Ω) of P (Ω) by
[ ½
P00 (Ω) = P ∈ P (Ω) : P [X(τ ) = x] = 1
(τ,x)∈[0,T ]×E 4

and for every f ∈ D(L) ∩ D (D1 ) the process


Z t
f (t, X(t)) − f (τ, X(τ )) − (D1 + L) f (s, X(s)) ds, t ∈ [τ, T ],
τ
2.1 Proof of the main result: Theorem 1.39 77
¾
is a P-martingale . (2.101)

Instead of D(L) ∩ D (D1 ) we often write D(1) (L): see the comments following
Definition 1.30. Let (vj : j ∈ N) be a sequence of continuous functions defined
on [0, T ] × E with the following properties:
(i) v0 = 1E , v1 = 1{4} ;
(ii) kvj k∞ ≤ 1, vj belongs to D(1) (L) = D(L) ∩ D (D1 ), and vj (s, 4) = 0 for
j ≥ 2; ¡ ¢
(iii)The linear span of vj , j ≥ 0, is dense in Cb [0, T ] × E 4 for the strict
topology Tβ .
In addition let (fk : k ∈ N) be a sequence in D(1) (L) such that the linear
span of {(fk , (D1 + L) fk ) : k ∈ N} is Tβ dense in the graph G (D1 + L) :=
{(f, (D1 + L) f ) : f ∈ D(L)} of the operator D1 +L. Moreover, let (sj : j ∈ N)
be an enumeration of the set Q∩[0, T ]. A subset P 0 (Ω), which is closely related
to P00 , may be described as follows (see (2.52) as well):
P 0 (Ω)
∞ \
\ ∞ \
∞ \ \
= {P ∈ P (Ω) :
n=1 k=1 m=0 (j1 ,...,jm+1 )∈Nm+1 0≤sj1 <...<sjm+1 ≤T
£ ¡ ¢ ¤
P [X (sjk ) ∈ E, 1 ≤ k ≤ m + 1] = P X sjm+1 ∈ E , and
Z m
¡ ¡ ¡ ¢¢ ¢Y
fk sjm+1 , X sjm+1 − fk (sjm , X (sjm )) vjk (sjk , X (sjk )) dP
k=1
Z ÃZ sjm+1
! m
Y
)
= (D1 + L) fk (s, X(s)) ds vjk (sjk , X (sjk )) dP . (2.102)
s jm k=1

Let P (Ω) be the collection of probability measures on FT0 . For a concise


formulation of the relevant distance between probability measures in P (Ω)
we introduce kind of Lévy numbers. Let P1 and P2 ∈ P (Ω). Then we write,
for Λ ⊂ [0, T ], Λ finite or countable,
© £ ¡ ¢¤
LΛ (P2 , P1 ) = lim inf η > 0 : P2 (X(s))s∈Λ ⊂ ∪`j=1 B xj , 2−m
`→∞
¡ ¢ £ ¤ ª
≥ 1 − η2−m P1 (X(s))s∈Λ ⊂ E , for all m ∈ N (2.103)
where B (x, ε) is a ball in E centered at x and with radius ε > 0. Notice that
in (2.103) lim`→∞ may be replaced with inf `∈N . In fact we shall prove that,
if for the operator L the martingale problem is solvable, that then the set
P 0 (Ω) is complete metrizable and separable for the metric d(P1 , P2 ) given by
dL (P1 , P2 )
¯ ¯
X X ¯Z ¯
¯ Y −j−` ¡ ¡ ¢¢ ¯
= 2−|Λ| ¯ 2 j
v s , X s d (P − P )¯
Λ⊂N,|Λ|<∞ ¯
(`j )j∈Λ ¯
j `j `j 2 1 ¯
j∈Λ ¯
78 2 Proof of main result

X ¡ ¢
+ 2−k LQ∩[0,sk ] (P2 , P1 ) + LQ∩[0,sk ] (P1 , P2 ) . (2.104)
k=1

If a sequence of probability measures (Pn )n∈N converges to P with respect


to the metric in (2.104), then the first term on the right-hand side says that
the finite-dimensional distributions of Pn converge to the finite-dimensional
distributions of P. The second term says, that the limit P is a measure indeed,
and that the paths of the process are P-almost surely totally bounded. The
following result should be compared to the comments in 6.7.4. of [225], pp.
167–168. It is noticed that in Proposition 2.8 the uniqueness of the martingale
problem is used to prove the separability.
Proposition 2.8. The set P 0 (Ω) supplied with the metric dL defined in
(2.104) is a separable complete metrizable Hausdorff space.
Proof. Let (Pn : n ∈ N) be a Cauchy sequence in (P 0 (Ω) , d). Then for ev-
ery m ∈ N, for every m-tuple (j1 , . . . , jm )R Q in Nm and for every m-tuple
m
(sj1 , . . . , sjm ) ∈ Qm ∩[0, T ] the limit lim`→∞ k=1 vjk (sjk , X (sjk )) dPn` ex-
ists. We shall prove that for every every m ∈ N, for every m-tuple (j1 , . . . , jm )
in Nm and for every m-tuple (tj1 , . . . , tjm ) ∈ [0, T ]m the limit
Z Y
m
lim ujk (tjk , X (tjk )) dPn (2.105)
n→∞ k=1

exists for all sequences (uj )j∈N in Cb ([0, T ] × E). Since, in addition,

lim lim LQ∩[0,sk ] (Pn , Pm ) = lim lim LQ∩[0,sk ] (Pn , Pm ) = 0, (2.106)


n→ m→∞ m→∞ n→∞

for all k ∈ N, it follows that the sequence (Pn )n∈N is tight in the sense that the
paths {X(s) : s ∈ Q ∩ [0, sk ]} are Pn -almost surely totally bounded uniformly
in Pn for all n simultaneously. The latter means that for every ε > 0 there
exists n(ε) ∈ N and integers (`m (ε))m∈N such that
h ¢i
`m (ε) ¡
Pn2 (X(s))s∈Q∩[0,sk ] ⊂ ∪j=1 B xj , 2−m
¡ ¢ h i
≥ 1 − ε2−m Pn1 (X(s))s∈Q∩[0,sk ] ⊂ E (2.107)

for all n2 , n1 ≥ n(ε), and for all m ∈ N. By enlarging `m (ε) we may and do
assume that
h ¢i
`m (ε) ¡
Pn (X(s))s∈Q∩[0,sk ] ⊂ ∪j=1 B xj , 2−m
¡ ¢ h i
≥ 1 − ε2−m Pn(ε) (X(s))s∈Q∩[0,sk ] ⊂ E , (2.108)

and
h ¢i
`m (ε) ¡
Pn (X(s))s∈Q∩[0,sk ] ⊂ ∪j=1 B xj , 2−m
2.1 Proof of the main result: Theorem 1.39 79
¡ ¢ h i
−m
≥ 1 − ε2 Pn (X(s))s∈Q∩[0,sk ] ⊂ E (2.109)

for all n ∈ N. It follows that


h ¡ ¢i
`m (ε)
Pn (X(s))s∈Q∩[0,sk ] ⊂ ∩∞ ∪
m=1 j=1 B xj , 2 −m

h i
≥ (1 − ε) Pn (X(s))s∈Q∩[0,sk ] ⊂ E , (2.110)

for all n ∈ N. But then there exists, by Kolmogorov’s extension theorem, a


probability measure P such that
Z Y Z Y
m m
lim ujk (tjk , X (tjk )) dPn = ujk (tjk , X (tjk )) dP, (2.111)
n→∞ k=1 k=1

for all m ∈ N, for all (j1 , . . . , jm ) ∈ Nm and for all (tj1 , . . . , tjm ) ∈ [0, T ]m .
From the description (2.101) of P 0 (Ω) it then readily follows that P is a
member of P 0 (Ω). So the existence of the limit in (2.105) remains to be
verified, together with the following facts: the limit P is a martingale solution,
and D([0, ∞], E 4 ) has full P-measure. Let t be in Q ∩ [0, T ]. Since, for every
j ∈ N, the process
Z s
vj (s, X(s)) − vj (0, X(0)) − (D1 + L) vj (σ, X(σ)) dσ, s ∈ [0, T ],
0

is a martingale for the measure Pn` , we infer


Z Z t Z Z
(D1 + L) vj (s, X(s)) dsdPn` = vj (t, X(t)) dPn` − vj (0, X(0)) dPn` .
0
Z Z t
and hence the limit lim (D1 + L) vj (s, X(s)) dsdPn` exists. Next let t0
`→∞ 0
be in [0, T ]. Again using the martingale property we see
Z
vj (t0 , X (t0 )) d (Pn` − Pnk )
Z µZ t ¶
= (D1 + L) vj (s, X(s)) ds d (Pn` − Pnk )
Z 0
+ vj (0, X(0)) d (Pn` − Pnk )
Z µZ t ¶
− (D1 + L) vj (s, X(s)) ds d (Pn` − Pnk ) , (2.112)
t0

where t is any number in Q ∩ [0, T ]. From (2.112) we infer


¯Z ¯
¯ ¯
¯ vj (t0 , X (t0 )) d (Pn − Pn )¯
¯ ` k ¯
80 2 Proof of main result
¯Z µZ t ¶ ¯
¯ ¯
≤ ¯¯ (D1 + L) vj (s, X(s)) ds d (Pn` − Pnk )¯¯
¯Z 0 ¯
¯ ¯
+ ¯¯ vj (0, X(0)) d (Pn` − Pnk )¯¯ + 2 |t − t0 | k(D1 + L) vj k∞ .(2.113)

If we let ` and k tend to infinity, we obtain


¯Z ¯
¯ ¯
lim sup ¯ vj (t0 , X (t0 )) d (Pn` − Pnk )¯¯ ≤ 2 |t − t0 | k(D1 + L) vj k∞ . (2.114)
¯
`,k→∞
R
Consequently for every s ∈ [0, T ] the limit lim`→∞ vj (s, X(s)) dPn` exists.
The inequality
¯Z Z ¯
¯ ¯
¯ vj (t, X(t)) dPn − vj (t0 , X (t0 )) dPn ¯
¯ ` `¯

¯Z Z t ¯
¯ ¯
= ¯¯ (D1 + L) vj (s, X(s)) ds dPn` ¯¯
t0
≤ |t − t0 | k(D1 + L) vj k∞
R
shows that the functions t 7→ lim`→∞ vj (t, X(t)) dPn` , j ∈ N, are contin-
uous. Since the linear span of (vj : j ≥ 2) is dense in Cb ([0, T ] × E) for the
strict topology, it follows that for every v ∈ Cb ([0, T ] × E) and for every
t ∈ [0, T ] the limit
Z
t 7→ lim v (t, X(t)) dPn` , t ∈ [0, T ], (2.115)
`→∞

exists and that this limit, as a function of t, is continuous. The following step
consists in proving that for every t0 ∈ [0, ∞) the equality
Z
lim lim sup |vj (t, X(t)) − vj (t0 , X (t0 ))| dPn` = 0 (2.116)
t→t0 `→∞

holds. For t > s the following (in-)equalities are valid:


µZ ¶2 Z
2
|vj (t, X(t)) − vj (s, X(s))| dPn` ≤ |vj (t, X(t)) − vj (s, X(s))| dPn`
Z Z
2 2
= |vj (t, X(t))| dPn` − |vj (s, X(s))| dPn`
Z
− 2< (vj (t, X(t)) − vj (s, X(s))) v j (s, X(s))dPn`
Z Z
2 2
= |vj (t, X(t))| dPn` − |vj (s, X(s))| dPn`
Z µZ t ¶
− 2< (D1 + L) vj (σ, X(σ)) dσ v j (s, X(s))dPn`
s
2.1 Proof of the main result: Theorem 1.39 81
Z Z
2 2
≤ |vj (t, X(t))| dPn` − |vj (s, X(s))| dPn`

+ 2(t − s) k(D1 + L) vj k∞ . (2.117)

Hence (2.115) together with (2.117) implies (2.116). By (2.116), we may ap-
ply Kolmogorov’s extension theorem to prove that there exists a probability
¡ ¢[0,T ]
measure P on Ω 0 := E 4 with the property that
Z Y
m Z Y
m
vjk (sjk , X (sjk )) dP = lim vjk (sjk , X (sjk )) dPn , (2.118)
n→∞
k=1 k=1

holds for all m ∈ N and for all (sj1 , . . . , sjm ) ∈ [0, T ]m . It then follows
¡ that the¢
equality in (2.118) is also valid for all m-tuples f1 , . . . , fm in Cb [0, T ] × E 4
instead of for vj1 , . . . , vjm . This
¡ is true because
¢ the linear span of the sequence
(vj )j∈N is Tβ -dense in Cb [0, T ] × E 4 . In addition we conclude that the
processes
Z t
f (t, X(t)) − f (0, X(0)) − (D1 + L) f (s, X(s)) ds,
0

t ∈ [0, T ], f ∈ D(1) (L), are P-martingales. We still have to show that


D([0, T ], E 4 ) has P-measure 1. From (2.116) it essentially follows that set
¡ ¢[0,T ]
of ω ∈ E 4 for which the left and right hand limits exist in E 4 has
“full” P-measure. First let f ≥ 0 be in Cb ([0, T ] × E). Then the process
·Z ∞ ¸
¯ 0
[Gλ f ] (t) := E e −λσ ¯
f (σ ∧ T, X(σ ∧ T )) dσ Ft
t
¡ ¢
is a P-supermartingale with respect to the filtration Ft0 t∈[0,T ] . It follows that
the limits limt↑t0 [Gλ f ] (t) and limt↓t0 [Gλ f ] (t) both exist P-almost surely for
all t0 ≥ 0 and for all f ∈ Cb ([0, T ] × E). In particular these limits exist P-
almost surely for all f ∈ D(1) (L). By the martingale property it follows that,
for f ∈ D(1) (L),
¯ ¯
¯f (t, X(t)) − λeλt [Gλ f ] (t)¯
¯ ·Z ∞ ¸¯
¯ ¯ ¯
= ¯¯λeλt E e−λσ (f (σ ∧ T, X(σ ∧ T )) − f (t, X(t))) dσ ¯ Ft0 ¯¯
¯ ·Zt ∞ µZ σ ¶ ¸¯
¯ λt ¯ ¯
= ¯¯λe E e−λσ (D1 + L) f (s, X(s)) ds dσ ¯ Ft0 ¯¯
Z ∞t t

≤ λeλt e−λσ (σ − t) k(D1 + L) f k∞ dσ = λ−1 k(D1 + L) f k∞ . (2.119)


t

Consequently, we may conclude that, for all s, t ≥ 0,

|f (t, X(t)) − f (s, X(s))|


82 2 Proof of main result
¯ ¯
≤ 2λ−1 k(D1 + L) f k∞ + ¯λeλt [Gλ f ] (t) − λeλs [Gλ f ] (s)¯ , (2.120)

Again using (2.108), (2.109) and (2.110) it follows that the the path

{X(s) : s ∈ Q ∩ [0, t], X(t) ∈ E}

is P-almost surely totally bounded. By separability and Tβτ -density of D(1) (L)
it follows that the limits limt↓s X(t) and lims↑t X(s) exist in E P-almost surely
for all s respectively t ∈ [0, T ], for which X(s) respectively X(t) belongs to
E. See the arguments which led to (2.12) and (2.13) in the proof of (a) of
Theorem 1.39. Put Z(s)(ω) = limt↓s,t∈Q∩[0,T ] X(t)(ω). Then, for P-almost all
ω the mapping s 7→ Z(s)(ω) is well-defined, possesses left limits in t ∈ [0, T ]
for those paths ω ∈ Ω for which ω(t) ∈ E and is right continuous. In addition
we have

E [f (s, Z(s))g(s)] = E [f (s, X(s+))g(s, X(s))]


= lim E [f (t, X(t))g(s, X(s))] = E [f (s, X(s))g(s, X(s))] ,
t↓s

for all f , g ∈ Cb ([0, T ] × E) and for all s ∈ [0, T ]: see (2.116). But then we
may conclude that X(s) = Z(s) P-almost surely for all s ∈ [0, T ]. Hence we
may replace X with Z and consequently (see the arguments in the proof of
(a) of Theorem 1.39, and see Theorem 9.4 in Blumenthal and Getoor [34], p.
49])
P [Ω] = 1, and so P ∈ P 0 (Ω) = P00 (Ω) (2.121)
¡ ¢ ¡ ¢
where Ω = D [0, T ] × E 4 . For the definition of D [0, T ] × E 4 see Defini-
tion 1.32, and for the definition of P 0 (Ω), and P00 (Ω) the reader is referred
to (2.102) and (2.101).
We also have to prove the separability. Denote by Convex the collection
of all mappings

α : Pf (N) × Pf (Q ∩ [0, T ]) → Q ∩ [0, 1],

which take only finitely many non-zero values, such that


X
α (Λ0 , Λ) = 1, Λ ∈ Pf (Q ∩ [0, T ]) ,
Λ0 ∈Pf (N)

and let {wΛ0 : Λ0 ∈ Pf (N)} be a countable family of functions from Q ∩ [0, T ]


to E 4 such that for every finite subset Λ = {sj1 , . . . , sjn } ∈ Pf (Q ∩ [0, T ])
the collection
{(wΛ0 (sj1 ) , . . . , wΛ0 (sjn )) : Λ0 ∈ Pf (N)}
¡ ¢(sj1 ,...,sjn ) ¡ 4 ¢Λ
is dense in E 4 = E . For example the value of the wΛ0 (sj` )
could be xk` , 1 ≤ ` ≤ n, where Λ0 = (k1 , . . . , kn ). Here (xk )k∈N is a dense
sequence in E4 . The countable collection of probability measures
2.1 Proof of the main result: Theorem 1.39 83

{Pα,w,Λ : α ∈ Convex, Λ ∈ Pf (N)}

determined by
£ ¡ ¢¤ X ¡ ¢
Eα,w,Λ F (s, X(s))s∈Λ = α (Λ0 , Λ) F (s, wΛ0 (s))s∈Λ
Λ0 ∈Pf (N)

is dense in P (Ω) endowed with the metric dL . Since P 0 (Ω) is a closed subspace
of P (Ω), it is separable as well.
Finally we observe that X(t) ∈ E, τ < s < t, implies ¡ X(s) ∈ E. ¢ This
follows from the assumption that the Skorohod space D [0, T ], E 4 is the
sample space on which we consider the martingale problem: see Definition
1.32. In particular it is assumed that X(s) = 4, τ < s ≤ t, implies X(t) = 4,
and L(ρ)f (ρ, ·) (X(ρ)) = 0 for s < ρ < t. Consequently, once we have X(s) =
4, and t ∈ (s, T ], then X(t) = 4, and by transposition X(t) ∈ E, s ∈ [τ, t)
implies X(s) ∈ E.

Proposition 2.9. Suppose that for every (τ, x) ∈ [0, T ] × E the martingale
problem is uniquely solvable. In addition, suppose that there exists λ > 0
such that the operator D1 + L is sequentially λ-dominant: see Definition
3.6. Define the map F : P 00 (Ω) → [0, T ] × E by F (P) = (τ, x), where
P ∈ P 00 (Ω) is such that P(X(s) = x) = 1, for s ∈ [0, τ ]. Then F is a
homeomorphism from the polish space P 00 (Ω) onto [0, T ] × E. In fact it fol-
lows that for every u ∈ Cb ([0, T ] × E) and for every s ∈ [τ, T ], the function
(τ, s, x) 7→ Eτ,x [u (s, X(s))], 0 ≤ τ ≤ s ≤ T , x ∈ E, is continuous.

Here P 00 (Ω) := {Pτ,x : (τ, x) ∈ [0, T ] × E}.


Proof. Since the martingale problem is uniquely solvable for every (τ, x) ∈
[0, T ] × E the map F is a one-to-one map from the polish space (P 00 (Ω) , dL )
onto [0, T ] × E (see Proposition 2.8 and (2.104)). Let for (τ, x) ∈ [0, T ] × E
the probability Pτ,x be the unique solution to the martingale problem:
(i) For every f ∈ D(1) (L) the process
Z t
f (t, X(t)) − f (τ, X(τ )) − (D1 + L) f (s, X(s))ds, t ∈ [τ, T ],
τ

is a Pτ,µ -martingale;
(ii) The Pτ,µ -distribution of X(τ ) is the measure µ. If µ = δx , then we write
Pτ,δx = Pτ,x , and Pτ,x [X(τ ) = x] = 1.
Then, by definition F (Pτ,x ) = (τ, x), (τ, x) ∈ [0, T ] × E. Moreover, since for
every (τ, x) ∈ [0, T ] × E the martingale problem is uniquely solvable we see
P 0 (Ω) = {Pτ,µ : (τ, µ) ∈ [0, T ] × P (E)}. Here P (E) is the collection of Borel
probability measures on E. This equality of probability spaces can be seen as
follows. If the measure Pτ,µ is a solution to the martingale problem, then it
84 2 Proof of main result

is automatically a member of P 0 (Ω). If P is a member of P 0 (Ω) which starts


at time τ , then by uniqueness of solutions we have:
£ ¯ ¤¯
P A ¯ σ (X(τ )) ¯X(τ )=x = Pτ,x [A] , A ∈ FTτ . (2.122)

In addition, P = Pτ,µ , where µ(B) = P [X(τ ) ∈ B], B ∈ E. Let ((t` , x` ))`∈N be


a sequence in [0, T ] × E with the property that lim`→∞ dL (Pt` ,x` , Pτ,x ) = 0
for some (τ, x) ∈ [0, T ] × E. Then for some stochastic variable ε the orbit
{(s, X(s)) : s ∈ (τ − ε, τ + ε)} is totally bounded Pt` ,x` -almost surely for all t`
and τ simultaneously. It follows that the sequence {x` = X (t` ) : ` ∈ N} ∪ {x}
is contained in a compact subset of E. Then lim`→∞ |vj (t` , x` ) − vj (τ, x)| =
0, for all j ∈ N, where, as above, the span of the sequence (vj )j≥2 is Tβ -
dense in C ([0, T ] × E). It follows that lim`→∞ (t` , x` ) = (t, x) in [0, T ] × E.
Consequently the mapping F is continuous. Since F is a continuous bijective
map from one polish space

P 00 (Ω) := {Pτ,x : (τ, x) ∈ [0, T ] × E} (2.123)

onto another such space [0, T ] × E, its inverse is continuous as well. Among
other things this impliesR that, for every s ∈ Q ∩ [0, ∞) and for every j ≥ 2,
the function (τ, x) 7→ vj (s, X(s)) dPτ,x belongs to Cb ([0, T ] × E). Since
the linear span of the sequence (vj : j ≥ 2) is Tβ -dense in Cb ([0, T ] × E)
it
R also follows that for every v ∈ Cb ([0, T ] × E), the function (τ, x) 7→
v (s, X(s)) dPτ,x belongs to Cb ([0, T ] × E). Next let s0 ∈ [0, T ] be arbitrary.
For every j ≥ 2 and every s ∈ Q ∩ [0, T ], s > s0 , we have by the martingale
property:

sup |Eτ,x (vj (s, X(s))) − Eτ,x (vj (s0 , X (s0 )))|
(τ,x)∈[0,s0 ]×E
¯Z s ¯
¯ ¯
= sup ¯ Eτ,x (Lvj (σ, X(σ))) dσ ¯¯
¯
(τ,x)∈[0,s0 ]×E s0

≤ (s − s0 ) k(D1 + L) vj k∞ . (2.124)

Consequently, for every s ∈ [0, T ], the function (τ, x) 7→ Eτ,x [vj (s, X(s))],
j ≥ 1, belongs to Cb ([0, T ] × E). It follows that, for every v ∈ Cb ([0, T ] × E)
and every s ∈ [0, T ], the function (τ, x) 7→ Eτ,x [v (s, X(s))] belongs to
Cb ([0, T ] × E). These arguments also show that the function (τ, s, x) 7→
Eτ,x [v (s, X(s))], 0 ≤ τ ≤ s ≤ T , x ∈ E, is continuous for every v ∈
Cb ([0, T ] × E). The continuity in the three variables (τ, s, x) requires the se-
quential λ-dominance of the operator D1 + L for some λ > 0. The arguments
run as follows. Using the Markov process

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} (2.125)

we define the semigroup {S(ρ) : ρ ≥ 0} as follows

S(ρ)f (τ, x) = P (τ, (ρ + s) ∧ T ) f ((ρ + s) ∧ T, ·) (x)


2.1 Proof of the main result: Theorem 1.39 85

= Eτ,x [f ((ρ + s) ∧ T, X ((ρ + s) ∧ T ))] . (2.126)

Here (τ, x) ∈ [0, T ] × E, ρ ≥ 0, and f ∈ Cb ([0, T ] × E). Let λ > 0 and


f ∈ Cb ([0, T ] × E). We want to establish a relationship between the semigroup
{S(ρ) : ρ ≥ 0} and the operator D1 + L. Therefore we first prove that the
process

t 7→e−λt f (t ∧ T, X (t ∧ T )) − e−λτ f (τ, X(τ ))


Z t
+ e−λρ (λI − D1 − L) f (ρ ∧ T, X (ρ ∧ T )) dρ, t ≥ τ, (2.127)
τ

is a Pτ,x -martingale with respect to the filtration (Ftτ )t∈[τ,T ] . Let τ ≤ s < t ≤
T , and y ∈ E. Then integration by parts shows:
Z t
−λt −λs
e f (t, X(t)) − e f (s, X(s)) + e−λρ (λI − D1 − L) f (ρ, X(ρ)) dρ
s
Z t
= e−λt f (t, X(t)) − e−λs f (s, X(s)) + λ e−λρ f (ρ, X(ρ)) dρ (2.128)
s
Z t Z t
− e−λt (D1 + L) f (ρ, X(ρ)) dρ − λ e−λρ (f (ρ, X(ρ)) − f (s, X(s))) dρ.
s s

Then by the martingale property the Ps,y -expectation of the expression in


(2.128) is zero. Employing the Markov property we obtain
·
Eτ,x e−λt f (t, X(t)) − e−λτ f (τ, X(τ ))
Z t ¸
¯ τ
+ e −λρ ¯
(λI − D1 − L) f (ρ, X(ρ)) dρ Fs
τ
µ
− e−λs f (s, X(s)) − e−λτ f (τ, X(τ ))
Z s ¶
−λρ
+ e (λI − D1 − L) f (ρ, X(ρ)) dρ
τ
·
= Eτ,x e−λt f (t, X(t)) − e−λs f (s, X(s))
Z t ¸
¯
+ e−λρ (λI − D1 − L) f (ρ, X(ρ)) dρ ¯ Fsτ
s

(Markov property)
·
= Es,X(s) e−λt f (t, X(t)) − e−λs f (s, X(s))
Z t ¸
−λρ
+ e (λI − D1 − L) f (ρ, X(ρ)) dρ = 0 (2.129)
s
86 2 Proof of main result

where in the final step in (2.129) we used the fact that the Ps,y -expectation,
y ∈ E, of the expression in (2.128) vanishes. Consequently, the process in
(2.127) is a Pτ,x -martingale. From the fact that the process in (2.127) is a
Pτ,x -martingale we infer by taking expectations that for t ≥ 0

e−λ(t+τ ) Eτ,x [f ((t + τ ) ∧ T, X ((t + τ ) ∧ T ))] − e−λτ Eτ,x [f (τ, X(τ ))]
Z t+τ
+ e−λρ Eτ,x [(λI − D1 − L) f (ρ ∧ T, X (ρ ∧ T ))] dρ = 0. (2.130)
τ

The equality in (2.130) is equivalent to

Eτ,x [f (τ, X(τ ))] − e−λt Eτ,x [f ((t + τ ) ∧ T, X ((t + τ ) ∧ T ))]


Z t+τ
= e−λ(ρ−τ ) Eτ,x [(λI − D1 − L) f (ρ ∧ T, X (ρ ∧ T ))] dρ = 0. (2.131)
τ

In terms of the semigroup {S(ρ) : ρ ≥ 0} the equality in (2.131) can be rewrit-


ten as follows:
Z t
f (τ, x) − e−λt S(t)f (τ, x) = e−λρ S(ρ) (λI − D1 − L) f (τ, x) dρ. (2.132)
0

By letting t → ∞ in (2.132) we see


Z ∞
f (τ, x) = e−λρ S(ρ) (λI − D1 − L) f (τ, x) dρ
0
= R(λ) (λI − D1 − L) f (τ, x) dρ (2.133)

where the definition


¡ ¢ of R(λ), λ > 0, is self-explanatory. Define the opera-
tor L(1) : D L(1) = R(λ)Cb ([0, T ] × E) → Cb ([0, T ] × E) by L(1) R(λ)f =
¡ ¢
λR(λ)f −f , f ∈ Cb ([0, T ] × E). Then by definition we see λI − L(1) R(λ)f =
¡ ¢
f , and thus R(λ) λI − L(1) R(λ)f = R(λ)f , f ∈ Cb ([0, T ] × E). Put g =
¡ ¢
λI − L(1) R(λ)f − f . Then by the resolvent identity we see that R(α)g = 0
for all α > 0, and hence S(ρ)g (τ, x) = Eτ,x [g ((ρ + τ ) ∧ T, X ((ρ + τ ) ∧ T ))] =
0 for all ρ > 0. By the right-continuity
¡ ¢ of the process ρ 7→ X(ρ), we see
that g = 0. Consequently, λI − L(1) R(λ)f − f = 0, f ∈ Cb ([0, T ] × E).
If f ∈ ¡D(1) (L),
¢ then¡ (2.133) ¢reads f = R(λ) (λI − D1 − L) f , and hence
f ∈ D L(1) , and λI − L(1) f = (λI − D1 − L) f , or what amounts to
¡ ¢
the same f ∈ D L(1) , and L(1) f = D1 f + Lf . In other words the op-
erator L(1) extends D1 + L. As in (1.41) define the sub-additive mapping
Uλ1 : Cb ([0, T ] × E, R) → L∞ ([0, T ] × E, R) by

Uλ1 f = sup inf {g ≥ f 1K : λg − D1 g − Lg ≥ 0} , and (2.134)


K∈K(E) g∈D (1) (L)

Since L(1) extends D1 + L, from (2.134) we get


2.1 Proof of the main result: Theorem 1.39 87
n o
Uλ1 f ≥ sup inf g ≥ f 1K : λg − L(1) g ≥ 0 , and (2.135)
K∈K(E) g∈D (L(1) )

Then, as explained in Proposition 1.22, formula (1.48), we have


n o
k
sup (µR (λ + µ)) f ; µ > 0, k ∈ N ≤ Uλ1 (f ), f ∈ Cb ([0, T ] × E, R) .
(2.136)
As is indicated in the proof of (iii) =⇒ (i) of Theorem 3.10 the following
equality also holds:
n o © ª
k
sup (µR (λ + µ)) f ; µ > 0, k ∈ N = sup e−λρ S(ρ) : ρ ≥ 0 , (2.137)

where f ∈ Cb ([0, T ] × E, R). For this observation the reader is referred to


the formulas (3.20), (3.21) and (3.22). Next let (fn )n∈N ⊂ Cb ([0, T ] × E) be a
sequence which decreases pointwise to zero. Using the sequential λ-dominance
of the operator D1 + L and using the equality in (2.136) and the inequality
in (2.137) we see that sup e−λρ S(ρ)fn (τ, x) decreases to zero uniformly on
ρ≥0
compact subsets of [0, T ] שE: see Definitionª3.6. From Proposition 1.21 it
follows that the semigroup e−λρ S(ρ) : ρ ≥ 0 is Tβ -equi-continuous. In ad-
dition, by the arguments above, every operator S(ρ), ρ ≥ 0, assigns to a
function f ∈ D(1) (L) = D (D1 ) ∩ D(L) a function S(ρ)f ∈ Cb ([0, T ] × E).
By the Tβ -continuity of S(ρ), and by the fact that D(1) (L) is Tβ -dense in
Cb ([0, T ] × E), the mapping S(ρ) extends to a Tβ -continuous linear contin-
uous operator from Cb ([0, T ] × E) to itself. This extension is again denoted
by S(ρ). In addition, for v ∈ D(1) (L), the function (τ, ρ, x) 7→ S(ρ)v(τ, x) is
continuous on [0, T ] × [0, ∞) × E; see (2.124). Fix f ∈ Cb ([0, T ] × E). Using
the sequential
© λ-dominance and ª its consequence of Tβ -equi-continuity of the
semigroup e−λρ S(ρ) : ρ ≥ 0 we see that the function (τ, s, x) 7→ S(ρ)f (τ, x)
is continuous on [0, T ]×[0, ∞)×E, and hence the same is true for the function
(τ, s, x) 7→ Eτ,x [f (s, X(s))]. Here we again used the Tβ -density of D(1) (L) in
Cb ([0, T ] × E).
This completes the proof of Proposition 2.9.
Notice that in the proof of the implication (iii) =⇒ (i) of Theorem 3.10 ar-
guments very similar to the ones in the final part of the proof of Proposition
2.9 will be employed.
Corollary 2.10. Suppose that the martingale problem is well posed for the
operator D1 + L, and that the operator D1 + L is sequentially λ-dominant for
some λ > 0. Let {(Ω, FTτ , Pτ,x ) : (τ, x) ∈ [0, T ] × E} be the solutions to the
martingale problem. Let the process in (2.125) be the corresponding Markov
process, and let the semigroup {S(ρ) : ρ ≥ 0}, as defined in (2.126), be the
corresponding Feller semigroup. Then this semigroup is Tβ -equi-continuous,
and its generator extends D1 + L.
88 2 Proof of main result

Proof.
© −λρ From Proposition
ª 1.21 it follows that for some λ > 0 the semigroup
e S(ρ) : ρ ≥ 0 is Tβ -equi-continuous: see the proof of Proposition 2.9.
Since S(ρ) = S(T ) for ρ ≥ T , we see that the semigroup {S(ρ) : ρ ≥ 0}
itself is Tβ -equi-continuous. Moreover, it is a Feller semigroup in the sense
that it consists of Tβ -continuous linear operators, and Tβ - lim S(t)f = S(s)f ,
t→s
f ∈ Cb ([0, T ] × E). From the proof of Proposition 2.9 it follows that the
generator of the semigroup {S(ρ) : ρ ≥ 0} extends D1 + L.
This proves Corollary 2.10.
The proof of the following proposition may be copied from Ikeda and Watan-
abe [109], Theorem 5.1. p. 205.
Proposition 2.11. Suppose that for every
¡ (τ, x) ∈ ¢[0, T ] × E the martingale
problem, posed on the Skorohod space D [0, T ], E 4 as follows,
(i) For every f ∈ D(1) (L) the process
Z t
f (t, X(t)) − f (τ, X(τ )) − (D1 + L) f (s, X(s))ds, t ∈ [τ, T ],
τ

is a P-martingale;
(ii)P(X(τ ) = x) = 1,
has a unique solution P = Pτ,x .
Then the process

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} , (2.138)

is
¡ τa ¢strong Markov process with respect to the right-continuous filtration
Ft+ t∈[τ,T ] .
τ
For the definition of FS+ the reader is referred to (1.91) in Remark 1.40.
Proof. Fix (τ, x) ∈ [0,
£ T ] × E¯ and ¤let S beτ a stopping time and choose a
realization A 7→ Eτ,x 1A ◦ ∨S ¯ FS+
τ
, A ∈ FT . Fix any ω ∈ Ω for which
£ ¯ τ ¤
A 7→ Qs,y [A] := Eτ,x 1A ◦ ∨S ¯ FS+ (ω),

is defined for all A ∈ FTτ . Here, by definition, (s, y) = (S(ω), ω(S(ω))). Notice
that this construction can be performed for Pτ,x -almost all ω. Let f be in
D(1) (L) = D (D1 ) ∩ D(L) and fix T ≥ t2 > t1 ≥ 0. Moreover, fix C ∈ Ftτ1 .
Then ∨−1 τ ∨S
S (C) is a member of Ft1 ∨S+ . Put Mf (t) = f (t, X(t)) − f (X(τ )) −
Rt
τ
(D1 + L) f (s, X(s))ds, t ∈ [τ, T ]. We have

Es,y [Mf (t2 )1C ] = Es,y [Mf (t1 )1C ] . (2.139)

We also have
Z µ Z t2 ¶
f (t2 , X(t2 )) − f (τ, X(τ )) − Lf (X(s))ds 1C dQs,y (2.140)
τ
2.1 Proof of the main result: Theorem 1.39 89

= Eτ,x f (t2 ∨ S, X (t2 ∨ S)) − f (S, X(S))

Z ¶ #
t2 ¯ τ
− (D1 + L) f (s ∨ S, X(s ∨ S)) ds (1C ◦ ∨S ) ¯ FS+ (ω)
τ

= Eτ,x f (t2 ∨ S, X(t2 ∨ S)) − f (S, X(S))

Z ! #
t2 ∨S ¯ τ
− (D1 + L) f (X(s)) ds (1C ◦ ∨S ) ¯ FS+ (ω)
S
" "Ã
= Eτ,x Eτ,x f (t2 ∨ S, X(t2 ∨ S)) − f (S, X(S))

Z ! # #
t2 ∨S ¯ τ ¯ τ
− (D1 + L) f (s, X(s)) ds ¯ Ft ∨S+ 1C ◦ ∨S ¯ FS (ω).(2.141)
1
S

By Doob’s optional sampling theorem, and right-continuity of paths, the pro-


cess
Z t∨S
f (t ∨ S, X(t ∨ S)) − f (S, X(S)) − (D1 + L) f (s, X(s)) ds
S

is a Pτ,x -martingale with respect to the filtration consisting of the σ-fields


τ
Ft∨S+ , t ∈ [τ, T ]. So from (2.140) we obtain:
Z µ Z t2 ¶
f (t2 , X(t2 )) − f (τ, X(τ )) − Lf (X(s))ds 1C dQs,y
τ

= Eτ,x f (t1 ∨ S, X(t1 ∨ S)) − f (S, X(S)) (2.142)

Z ! #
t1 ∨S ¯ τ
− ¯
(D1 + L) f (s, X(s)) ds (1C ◦ ∨S ) FS+ (ω)
S
Z µ Z t1 ¶
= f (t1 , X(t1 )) − f (τ, X(τ )) − (D1 + L) f (s, X(s))ds 1C dQs,y .
τ

It follows that, for f ∈ D(L), the process Mf (t) is a Ps,y - as well as a Qs,y -
martingale. Since Ps,y [X(s) = y] = 1 and since
£ ¯ τ ¤
Qs,y [X(s) = y] = Eτ,x 1{X(S)=y} ◦ ∨S ¯ FS+ (ω)
£ ¯ τ ¤
¯
= Eτ,x 1{X(S)=y} FS+ (ω) = 1{X(S)=y} (ω) = 1,(2.143)

we conclude that the probabilities Ps,y and Qs,y are the same. Equality (2.143)
follows, because, by definition, y = X(S)(ω) = ω(S(ω)). Since Ps,y = Qs,y , it
then follows that
90 2 Proof of main result
£ ¯ τ ¤
PS(ω),X(S)(ω) [A] = Eτ,x 1A ◦ ∨S ¯ FS+ (ω), A ∈ FTτ .

Or putting it differently:
£ ¯ τ ¤
PS,X(S) [1A ◦ ∨S ] = Eτ,x 1A ◦ ∨S ¯ FS+ , A ∈ FTτ . (2.144)

However this is exactly the strong Markov property.


This concludes the proof of Proposition 2.11.
The following proposition can be proved in the same manner as Theorem 5.1.
Corollary in Ikeda and Watanabe [[109], p. 206].
Proposition 2.12. If an operator family L = {L(s) : 0 ≤ s ≤ T } generates
a Feller evolution {P (s, t) : 0 ≤ s ≤ t ≤ T }, then the martingale problem is
uniquely solvable for L.

Proof. Let {P (τ, t) : 0 ≤ τ ≤ t ≤ T } be a Feller evolution generated by L and


let

{(Ω, FTτ , Pτ,x ) , (X(t), τ ≤ t ≤ T ) , (∨t : τ ≤ t ≤ T ) , (E, E)} , (2.145)

be the associated strong Markov process (see Theorem 1.39 (a)) If f belongs
to D(1) (L), then the process
Z t
Mf (t) := f (t, X(t)) − f (τ, X(τ )) − (D1 + L) f (s, X(s))ds, t ∈ [τ, T ],
τ

is a Pτ,x -martingale for all (τ, x) ∈ [0, T ] × E. This can be seen as follows. Fix
T ≥ t2 > t1 ≥ 0. Then
£ ¯ ¤
Eτ,x Mf (t2 ) ¯ Ftτ1 − Mf (t1 )
· Z t2 ¸
¯ τ
= Eτ,x f (t2 , X(t2 )) − ¯
(D1 + L) f (X(s))ds Ft1 − f (t1 , X(t1 ))
t1
(Markov property)
· Z t2 ¸
= Et1 ,X(t1 ) f (t2 , X(t2 )) − (D1 + L) f (s, X(s))ds − f (t1 , X(t1 ))
t1
Z t2
= Et1 ,X(t1 ) [f (t2 , X(t2 ))] − Et1 ,X(t1 ) [(D1 + L) f (s, X(s))] ds
t1
− f (t1 , X(t1 ))

(see Proposition 3.1 below)


Z t2
d
= Et1 ,X(t1 ) [f (t2 , X(t2 ))] − Et ,X(t1 ) [f (s, X(s))] ds − f (t1 , X(t1 ))
t1 ds 1
= 0. (2.146)
2.1 Proof of the main result: Theorem 1.39 91

Hence from (2.146) it follows that the process Mf (t), t ≥ 0, is a Pτ,x -


martingale. Next we shall prove the uniqueness of the solutions of the mar-
tingale problem associated to the operator L. Let P1τ,x and P2τ,x be solutions
“starting” in x ∈ E at time τ . We have to show that these ¡probabilities
¢
coincide. Let f belong to D(1) (L) and let S : Ω → [τ, T ] be an Ft+τ
t∈[τ,T ]
-
stopping time. Then, via partial integration, we infer
Z (

λ e−λt f ((t + S) ∧ T, X ((t + S) ∧ T ))
0
Z )
t+S
− (D1 + L) f (ρ ∧ T, X (ρ ∧ T )) dρ − f (S, X(S)) dt + f (S, X(S))
S
Z (

−λt
=λ e f ((t + S) ∧ T, X ((t + S) ∧ T ))
0
Z )
t+S
− (D1 + L) f (ρ ∧ T, X (ρ ∧ T )) dρ dt
S
Z ∞
=λ e−λt f ((t + S) ∧ S, X ((t + S) ∧ T )) dt
0
Z ∞ Z t
−λ e−λt (D1 + L) f ((t + S) ∧ T, X ((ρ + S) ∧ T )) dρ dt
0 0
Z ∞
=λ e−λt f ((t + S) ∧ T, X ((t + S) ∧ T )) dt
0
Z ∞ µZ ∞ ¶
−λt
−λ e dt (D1 + L) f ((ρ + S) ∧ T, X ((ρ + S) ∧ T )) dρ
0 ρ
Z ∞
= e−λt [(λI − D1 − L)f ] ((t + S) ∧ T, X ((t + S) ∧ T )) dt. (2.147)
0

From Doob’s optional sampling theorem together with (2.147) we obtain:


Z ∞
£ ¯ τ ¤
e−λt E1τ,x (λI − D1 − L) f ((t + S) ∧ T, X ((t + T ) ∧ T )) ¯ FS+ dt
0
Z "(

=λ e−λt E1τ,x f ((t + S) ∧ T, X ((t + S) ∧ T ))
0
Z ) #
t+S ¯
¯ τ
− (D1 + L) f (ρ ∧ T, X (ρ ∧ T )) dρ − f (S, X(S)) ¯ FS+ dt
S

+ f (S, X(S))
= f (S, X(S))
Z ∞ "(
−λt 2
=λ e Eτ,x f ((t + S) ∧ T, X ((t + S) ∧ T ))
0
92 2 Proof of main result
Z ) #
t+S ¯
¯ τ
− (D1 + L) f (ρ ∧ T, X (ρ ∧ T )) dρ − f (S, X(S)) ¯ FS+ dt
S

+ f (S, X(S))
Z ∞ h ¯ i
¯ τ
= e−λt E2τ,x (λI − D1 − L) f ((t + S) ∧ T, X ((t + S) ∧ T )) ¯ FS+ dt.
0
(2.148)

As in (3.5) below we write:

S(ρ)f (t, x) = P (t, (ρ + t) ∧ T ) f ((ρ + t) ∧ T, ·) (x), f ∈ Cb ([0, T ] × E) ,

ρ ≥ 0, (t, x) ∈ [0, T ] × E. Then the family {S(ρ) : ρ ≥ 0} is a Tβ -continuous


semigroup. Its resolvent is given by
Z ∞
[R(λ)f ] (τ, x) = e−λt [P (τ, (τ + t) ∧ T ) f ((τ + t) ∧ T, ·)] (x)dt
0
Z ∞
= e−λt S(t)f (τ, x)dt, (2.149)
0

for x ∈ E, λ > 0, and f ∈ Cb ([0, T ] × E). Let L(1) be its generator. Then, as
will be shown in Theorem 3.2 below, L(1) is the Tβ -closure of D1 + L, and
³ ´
λI − L(1) R(λ)f = f, f ∈ Cb ([0, T ] × E) ,
³ ´ ³ ´
R(λ) λI − L(1) f = f, f ∈ D L(1) . (2.150)

Since L(1) is the Tβ -closure of D1 + L, the equalities in (2.148) also hold for
L(1) instead of D1 + L. Among other things we see that
³ ´
R λI − L(1) = Cb ([0, T ] × E) , λ > 0.

From (2.148), with L(1) instead of D1 + L, (2.149), and (2.150) it then follows
that for g ∈ Cb ([0, T ] × E) we have
Z ∞
£ ¯ τ ¤
e−λt E1τ,x g ((t + S) ∧ T, X ((t + T ) ∧ T )) ¯ FS+ dt
0
Z ∞
= e−λt [S(t)g] (S, X(S)) dt
Z0 ∞
£ ¯ τ ¤
= e−λt E2τ,x g ((t + S) ∧ T, X ((t + T ) ∧ T )) ¯ FS+ dt. (2.151)
0

Since Laplace transforms are unique, g belongs to Cb ([0, T ] × E) and paths


are right continuous, we conclude
£ ¯ τ ¤
E1τ,x g ((t + S) ∧ T, X ((t + S) ∧ T )) ¯ FS+
2.1 Proof of the main result: Theorem 1.39 93

= [S(t)g] (S, X(S))


£ ¯ τ ¤
= E2τ,x g ((t + S) ∧ T, X ((t + S) ∧ T )) ¯ FS+ , (2.152)
¡ τ ¢
whenever g belongs to Cb ([0, T ] × E), t ∈ [0, ∞] and S is an Ft+ t∈[τ,T ]
-
stopping time. The first equality in (2.152) holds P1τ,x -almost surely and the
second P2τ,x -almost surely. In (2.152) we take for S a fixed time s ∈ [τ, T − t]
and we substitute ρ = t + s. Then we get
£ ¯ τ ¤ £ ¯ τ ¤
E1τ,x g ((ρ, X (ρ))) ¯ Fs+ = [S(ρ − s)g] (s, X(s)) = E2τ,x g (ρ, X (ρ)) ¯ Fs+ .
(2.153)
For s = τ the equalities in (2.153) imply
£ ¯ ¤ £ ¯ ¤
E1τ,x g ((ρ, X (ρ))) ¯ Fττ + = [S(ρ − τ )g] (τ, X(τ )) = E2τ,x g (ρ, X (ρ)) ¯ Fττ + ,
(2.154)
and by taking expectations in (2.154) we get

E1τ,x [g ((ρ, X (ρ)))] = [S(ρ − τ )g] (τ, x) = E2τ,x [g (ρ, X (ρ))] (2.155)

where we used the fact that X(τ ) = x P1τ,x - and P2τ,x -almost surely. It follows
that the one-dimensional distributions of P1τ,x and P2τ,x coincide. By induction
with respect to n and using (2.153) several times we obtain:
hYn i hYn i
E1τ,x fj (tj , X (tj )) = E2τ,x fj (tj , X (tj )) (2.156)
j=1 j=1

for n = 1, 2, . . . and for f1 , . . . , fn in Cb ([0, T ] × E). But then the probabilities


P1τ,x and P2τ,x are the same.
This proves Proposition 2.12.

Proposition 2.13. Let L be a densely defined operator for which the mar-
tingale problem is uniquely solvable. Then there exists a unique closed linear
extension L0 of L, which is the generator of a Feller semigroup.

Proof. Existence. Let {Pτ,x : (τ, x) ∈ [0, T ] × E} be the solution for L. Put

[S(t)f ] (τ, x) = Eτ,x [f ((τ + t) ∧ T, X ((τ + t) ∧ T ))] ,


Z ∞
[R(λ)f ] (τ, x) = e−λs [S(s)f ] (τ, x)ds,
0
L0 (R(λ)f ) := λR(λ)f − f, f ∈ Cb ([0, T ] × E) .

Here t ∈ [0, T ] and λ > 0 are fixed. Then, as follows from the proof of Theorem
3.2 the operator L0 extends D1 + L and generates a Tβ -continuous Feller
semigroup.
Uniqueness. Let L1 and L2 be closed linear extensions of L, which both gen-
erate Feller evolutions. Let
©¡ ¢ ª
Ω, FTτ , P1τ,x , (X(t) : t ∈ [0, T ]), (∨t : t ∈ [0, T ]), (E, E)
94 2 Proof of main result

respectively
©¡ ¢ ª
Ω, FTτ , P2τ,x , (X(t) : t ∈ [0, T ]), (∨t : t ∈ [0, T ]), (E, E)
be the corresponding Markov processes. For every f ∈ D(L), the process
Z t
f (t, X(t)) − f (τ, X(τ )) − (D1 + L) f (s, X(s))ds, t ≥ 0,
τ

is a martingale with respect to P1τ,x


as well as with respect to P2τ,x . Uniqueness
1 2
implies Pτ,x = Pτ,x and hence L1 = L2 .
Proof (Proof of item (d) of Theorem 1.39: conclusion). In this final part of the
proof we mainly collect the results, which we proved in part (a) of Theorem
1.39, and the propositions 2.8, 2.9, 2.10, 2.11, 2.12, and 2.13. The main work we
have to do is to organize these matters into a proof of item (d) of Theorem 1.39.
More details follow. As in (2.123) let P 00 (Ω) = {Pτ,x : (τ, x) ∈ [0, T ] × E}, be
the collection of unique solutions to the martingale problem. Then the process
n o
(Ω, FTτ , Pτ,x )(τ,x)∈[0,T ]×E , (X(t), t ∈ [0, T ]) , (∨t : t ∈ [0, T ]) , (E, E)

is strong Markov process, and the function P (τ, x; t, B) defined by


P (τ, x; t, B) = Pτ,x [X(t) ∈ B] , 0 ≤ τ ≤ t ≤ T, x ∈ E, B ∈ E,
is a Feller evolution. Here the ¡ state variables
¢ X(t) : Ω → E 4 are defined
4
by X(t) = ω(t), ω ∈ Ω = D [0, T ], E . The sample path space is supplied
with the standard filtration (Ftτ )τ ≤t≤T . The strong Markov property follows
from Proposition 2.11. The Feller property is a consequence of Proposition 2.9
(which in turn is based on Proposition 2.8 where completeness and separability
of the space P 00 (Ω) is heavily used). Its Tβ -continuity and Tβ -equi-continuity
is explained in Corollary 2.10 to Proposition 2.11. Define the Feller semigroup
{S(ρ) : ρ ≥ 0} on Cb ([0, T ] × E) as in (2.126), and let L(1) be its generator.
From Corollary 2.10 we see that L(1) extends the operator D1 + L. Since the
martingale problem is uniquely solvable for the operator L, it follows that the
martingale problem is uniquely solvable for the operator L(1) (but now as a
time-homogeneous martingale problem). Therefore, Proposition 2.12 implies
that the operator L(1) is the unique extension of D1 + L which generates
a Feller semigroup. It follows that L(1) − D1 is the unique Tβ -extension of
L which generates a Feller evolution. This Feller evolution is given by the
original solution to the martingale problem: this claim follows from item (a)
of Theorem 1.39.
Finally, this completes the proof of item (d) of Theorem 1.39.

2.1.5 Proof of item (e) of Theorem 1.39


In this subsection we will show that under certain conditions, like possess-
ing the Korovkin property, satisfying the maximum principle, and Tβ -equi-
continuity a Tβ -densely defined operator in Cb (E) has a unique extension
which generates a (strong) Markov process.
2.1 Proof of the main result: Theorem 1.39 95

Proof (Proof of item (e) of Theorem 1.39). Let E0 be a subset of [0, T ] × E


which is polish for the relative topology. First suppose that the operator D1 +
L possesses the Korovkin property on K. Also suppose that it satisfies the
maximum principle on E0 . By Proposition 3.17 and its Corollary 3.18 there
exists a family of linear operators {R(λ) : λ > 0} such that for all (τ0 , x0 ) ∈ E0
and g ∈ Cb (E0 ) the following equalities hold:

λR(λ)g (τ0 , x0 )
½ · µ ¶ ¸ ¾
1
= inf max h (τ0 , x0 ) + g − I − (D1 + L) h (τ, x)
h∈D (1) (L) (τ,x)∈E0 λ
½ µ ¶ ¾
1
= inf h (τ0 , x0 ) : I − (D1 + L) h ≥ g on E0
h∈D (1) (L) λ
½ µ ¶ ¾
1
= sup h (τ0 , x0 ) : I − (D1 + L) h ≤ g on E0
h∈D (1) (L) λ
½ · µ ¶ ¸ ¾
1
= sup min h (τ0 , x0 ) + g − I − (D1 + L) h (τ, x) .
h∈D (1) (L) (τ,x)∈E0 λ
(2.157)

As will be shown in Proposition 3.17 the family {R(λ) : λ > 0} has the resol-
vent property: R(λ) − R(µ) = (λ − µ) R(µ)R(λ), λ > 0, µ > 0. It also follows
that R(λ) (λI − D1 − L) f = f on E0 for f ∈ D(1) (L). This equality is an
easy consequence of the inequalities in (2.157): see Corollary 3.18. Fix λ > 0
and f ∈ Cb ([0, T ] × E). We will prove that f = Tβ - lim αR(α)f . If f is of
α→∞
the form f = R(λ)g, g ∈ Cb (E0 ), then by the resolvent property we have

α αR(α)g
αR(α)f − f = αR(α)R(λ)g − R(λ)g = R(λ)g − R(λ)g − .
α−λ α−λ
(2.158)
Since kαR(α)gk∞ ≤ kgk∞ , the equality in (2.158) yields

k·k∞ - lim αR(α)f − f = 0 for f of the form f = R(λ)g, g ∈ Cb (E0 ).


α→∞

Since g = R(λ) (λI − D1 − L) g on E0 , g ∈ D(1) (L), it follows that

lim kαR(α)g − gk∞ = 0 for g ∈ D(1) (L) = D (D1 ) ∩ D(L). (2.159)


α→∞

As will be proved in Corollary 3.20 there exists λ0 > 0 such that the family
{λR(λ) : λ ≥ λ0 } is Tβ -equi-continuous. Hence for u ∈ H + (E0 ) there exists
v ∈ H + (E0 ) that for α ≥ λ0 we have

kuαR(α)gk∞ ≤ kvgk∞ , g ∈ Cb (E0 ) . (2.160)

Fix ε > 0, and choose for f ∈ Cb (E0 ) and u ∈ H + (E0 ) given g ∈ D(1) (L) in
such a way that
96 2 Proof of main result

2
ku(f − g)k∞ + kv(f − g)k∞ ≤ ε. (2.161)
3
Since D(L) is Tβ -dense in Cb ([0, T ] × E) such a choice of g is possible by. The
inequality (2.161) and the identity

αR(α)f − f = αR(α)(f − g) − (f − g) + αR(α)g − g yield


ku (αR(α)f − f )k∞
≤ ku (αR(α)(f − g))k∞ + ku(f − g)k∞ + kuαR(α)g − gk∞
≤ kv(f − g)k∞ + ku(f − g)k∞ + kuαR(α)g − gk∞
2
≤ ε + kuαR(α)g − gk∞ . (2.162)
3
From (2.159) and (2.162) we infer Tβ - lim αR(α)f = f , f ∈ Cb (E0 ). Of
α→∞
course the same arguments if E0 = [0, T ] × E. In fact the detailed arguments
which prove the fact that the operator D1 + L, confined to E0 , extends to
the unique generator of a Feller semigroup are found in the proof of Theorem
3.21.
We still have to show that the martingale problem for the operator L
restricted to E0 is well posed. Saying that the martingale is well posed for
L ¹E0 is the same as saying that the martingale problem is well posed for the
operator (D1 + L) ¹E0 . More precisely, if
n o
(Ω, FTτ , Pτ,x )(τ,x)∈[0,T ]×E , (X(t), t ∈ [0, T ]) , (E0 , E0 ) (2.163)

is a solution to the martingale problem associated to L ¹E0 , then the time-


homogeneous family
½³ ´ ¾
¡ ¢
e e (0)
Ω, F, Pτ,x , (Y (t), t ≥ 0) , [0, T ] × E, B[0,T ] ⊗ E0 (2.164)
(τ,x)∈[0,T ]×E0

is a solution to the martingale problem associated with (D1 + L) ¹E0 . Here


Ωe = [0, T ] × Ω, Y (t)(τ, ω) = ((τ + t) ∧ T, X ((τ + t) ∧ T )), (τ, ω) ∈ [0, T ] × Ω,
(0)
and the measure Pτ,x is determined by the equality
   
Yn n
Y
E(0)
τ,x
 fj (Y (tj )) = Eτ,x  fj ((τ + tj ) ∧ T, X ((τ + tj ) ∧ T ))
j=1 j=1
(2.165)
where the functions fj , 1 ≤ j ≤ n, are bounded Borel measurable functions
(0)
on [0, T ] × E0 , and where 0 ≤ t1 < · · · tn . Conversely, if the measures Pτ,x in
(2.164) are known, then those in (2.163) are also determined by (2.165):
   
n
Y n
Y
Eτ,x  fj (tj , X (tj )) = E(0)
τ,x
 fj (Y (tj − τ )) (2.166)
j=1 j=1
2.1 Proof of the main result: Theorem 1.39 97

where the functions fj , 1 ≤ j ≤ n, are again bounded Borel measurable func-


tions on [0, T ] × E0 , and where τ ≤ t1 < · · · tn ≤ T . In fact in (2.166) the
functions fj , 1 ≤ j ≤ n, only need to be defined on E0 , and such a function fj
can be identified with a function on [0, T ] × E0 which does not depend on the
time variable: (s, y) 7→ fj (y), (s, y) ∈ [0, T ]×E0 . It follows that instead of con-
sidering the time-inhomogeneous martingale problem associated with L ¹E0
we may consider the time-homogeneous martingale problem associated with
(D1 + L) ¹E0 . However, the martingale problem for the time-homogeneous
case is taken care of in the final part of Theorem 3.21.
So combining the above observations with Theorem 3.21 completes the
proof of item (e) of Theorem 1.39.
Remark 2.14. Prove that the process in (2.3) is a supermartingale indeed.
+
Remark 2.15. Let (ψm )m∈N be a sequence in Cn³ b ([τ, T ] ´
× E) which decreases
o
pointwise to the zero function. Since the orbit e
t, X(t) : t ∈ [τ, T ] is Pτ,x -
almost surely compact we know that
³ ´
inf sup ψm t, X(t)e = 0, Pτ,x -almost surely.
m∈N t∈[τ,T ]

2.1.6 Some historical remarks


The Lévy numbers in (2.103) are closely related to the Lévy metric, which
in turn is related to approach structures. The definition of Lévy metric and
Lévy-Prohorov metric can be found in Encyclopaedia of Mathematics, edited
by Hazewinkel [102]. In the area of convergence of measures the Encyclopae-
dia contains contributions by V. M. Zolotarev. In fact special sections are
devoted to the Lévy metric, the Lévy-Prokhorov metric, and related topics
like convergence of probability measures on complete metrizable spaces. The
Lévy metric goes back to Lévy: see [146]. The Lévy-Prohorov metric general-
izes the Lévy metric, and has its origin in Prohorov [188]. For completeness
we insert the definition of Lévy-Prohorov metric.
Definition 2.16. Let (E, d) be a metric space with its Borel sigma field E.
Let P(E) denote the collection of all probability measures on the measurable
space (E, E). For a subset A ⊆ E, define the ε-neighborhood of A by
[
Aε := {x ∈ E : there exists y ∈ A such that d(x, y) < ε} = Bε (y)
y∈A

where Bε (y) is the open ball of radius ε centered at y. The Lévy-Prohorov


metric dLP : P(E)2 → [0, +∞) is defined by setting the distance between two
probability measures µ and ν as
dLP (µ, ν)
= inf {ε > 0 : µ(A) ≤ ν (Aε ) + ε and ν(A) ≤ µ (Aε ) + ε for all A ∈ E} .
(2.167)
98 2 Proof of main result

For probability measures we clearly have dLP (µ, ν) ≤ 1.


Some authors omit one of the two inequalities or choose only open or
closed subsets A; either inequality implies the other, but restricting to open
or closed sets changes the metric as defined in (2.167). The Lévy-Prohorov
metric is also called the Prohorov metric. The interested reader should com-
pare the definition of Lévy-Prohorov metric with that of approach structure
as exhibited in e.g. Lowen [152]. When discussing convergence of measures
and constructing appropriate metrics the reader is also referred to Billingsley
[30], Parthasarathy [186], Zolotarev [261], and others like Bickel, Klaassen,
Ritov and Wellner in [29], appendices A6–A9. A book which uses the notion
of Korovkin set to a great extent is [6]. For applications of Korovkin sets to
ergodic theory see e.g. Marsden and Riemenschneider [155], Nishiraho [166],
Labsker [142], Chapter 7 and 8 in Donner [72], and Krengel [138]. Another
book of interest is [28] edited by Bergelson, March and Rosenblatt. For the
convergence results we also refer to the original book by Korovkin [137]. The
reader also might want to consult (the references in) Bukhalov [45]. In the
terminology of test sets, or Korovkin sets, our space D(1) = D (D1 ) ∩ D (L)
−1
in Cb ([0, T ] × E) is a Korovkin set for the resolvent family (λI − D1 − L) ,
λ > 0. From the proof of item (e) of Theorem 1.39 it follows that we only need
the Korovkin property for some fixed λ0 > 0: see the definitions 3.13 and 1.36.
In the finite-dimensional setting these Korovkin sets may be relatively small:
see e.g. Özarslan and Duman [180]. Section 5.2 in the recent book on func-
tional analysis by Dzung Minh Ha [96] carries the title “Korovkin’s theorem
and the Weierstrass approximation theorem”.
3
Space-time operators and miscellaneous topics

3.1 Space-time operators

In this section we will discuss in more detail the generators of the the time-
space Markov process (see (1.75):

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ τ ) , (∨t : τ ≤ t ≤ T ) , (E, E)} (3.1)

In Definition 1.30 we have introduced the family of generators of the corre-


sponding Feller evolution {P (τ, t) : 0 ≤ τ ≤ t ≤ T } given by P (τ, t) f (x) =
Eτ,x [f (X(t))], f ∈ Cb (E). In fact for any fixed t ∈ [0, T ] we will consider
the Feller evolution as an operator from Cb ([0, T ] × E) to Cb ([0, t] × E).
This is done in the following manner. To a function f ∈ Cb ([0, T ] × E) our
Feller evolution assigns the function (τ, x) 7→ P (τ, t) f (t, ·) (x). We will also
consider the family of operators L := {L(t) : t ∈ [0, T )} as defined in Def-
inition 1.30, and which is considered as a linear operator which acts on a
subspace of Cb ([0, T ] × E). It is called the (infinitesimal) generator of the
P (s, t)f − f
Feller evolution {P (s, t) : 0 ≤ s ≤ t ≤ T }, if L(s)f = Tβ -lim ,
t↓s t−s
0 ≤ s ≤ T . This means that a function f belongs to D (L(s)) whenever
P (s, t)f − f
L(s)f := lim exists in Cb (E), equipped with the strict topology.
t↓s t−s
As explained earlier, such a family of operators is considered as an operator
L with domain in the space Cb ([0, T ] × E). A function f ∈ Cb ([0, T ] × E)
is said to belong to D(L) if for every s ∈ [0, T ] the function x 7→ f (s, x)
is a member of D(L(s)) and if the function (s, x) 7→ L(s)f (s, ·) (x) belongs
to Cb (E). Instead of L(s)f (s, ·) (x) we often write L(s)f (s, x). If a function
f ∈ D(L) is such that the function s 7→ f (s, x) is differentiable, then e say
that f belongs to D(1) (L). We will show that such a generator also gener-
ates the corresponding Markov process in the sense of Definition 1.31. For
convenience of the reader we repeat here the defining property. A family of
operators L := {L(s) : 0 ≤ s ≤ T }, is said to generate a time-inhomogeneous
100 3 Space-time operators

Markov process, as described in (1.75), if for all functions u ∈ D(L), for all
x ∈ E, and for all pairs (τ, s) with 0 ≤ τ ≤ s ≤ T the following equality holds:
· ¸
d ∂u
Eτ,x [u (s, X(s))] = Eτ,x (s, x) + L(s)u (s, ·) (X(s)) . (3.2)
ds ∂s
Our first result says that generators of Markov processes and the corre-
sponding Feller evolutions coincide.
Proposition 3.1. Let the Markov process in (3.1) and the Feller evolu-
tion {P (τ, t) : 0 ≤ τ ≤ t ≤ T } be related by P (τ, t) f (x) = Eτ,x [f (X(t))],
f ∈ Cb (E). Let L = {L(s) : 0 ≤ s ≤ T } be a family of linear operators with
domain and range in Cb (E). If L is a generator of the Feller evolution, then
it also generates the corresponding Markov process. Conversely, if L generates
a Markov process, then it also generates the corresponding Feller evolution.
Proof. First suppose that the Feller evolution {P (τ, t) : 0 ≤ τ ≤ t ≤ T } is gen-
erated by the family L. Let the function f belong to the domain of L and
suppose that D1 f is continuous on [0, T ] × E. Then we have
· ¸
∂f
Eτ,x (s, X(s)) + L(s)f (s, ·) (X(s))
∂s
∂f
= P (τ, s) (s, ·) (x) + P (τ, s) L(s)f (s, ·) (x)
∂s · ¸
∂f P (s, s + h) f (s, ·) − f (s, ·)
= P (τ, s) (s, ·) (x) + P (τ, s) lim (x)
∂s h↓0 h
· ¸
∂f P (s, s + h) f (s, ·) − f (s, ·)
= P (τ, s) (s, ·) (x) + lim P (τ, s) (x)
∂s h↓0 h
· ¸
∂f P (τ, s + h) f (s, ·) − P (τ, s) f (s, ·)
= P (τ, s) (s, ·) (x) + lim (x)
∂s h↓0 h
· ¸
∂f f (s + h, ·) − f (s, ·)
= P (τ, s) (s, ·) (x) − lim P (τ, s + h) (x)
∂s h↓0 h
· ¸
P (τ, s + h) f (s + h, ·) − P (τ, s) f (s, ·)
+ lim (x).
h↓0 h
∂f ∂f
= P (τ, s) (s, ·) (x) − P (τ, s) (s, ·) (x)
∂s ∂s
Eτ,x [f (s + h, X (s + h))] − Eτ,x [f (s, X (s))]
+ lim
h↓0 h
d
= Es,X(s) [f (s, X (s))] . (3.3)
ds
In (3.3) we used the fact that the function D1 f is continuous and its con-
f (s + h, y) − f (s, y)
sequence that lim converges uniformly for y in com-
h↓0 h
pact subsets of E. We also used the fact that the family of operators
{P (τ, t) : t ∈ [τ, T ]} is equi-continuous for the strict topology.
3.1 Space-time operators 101

In the second part we have to show that a generator L of a Feller process


(3.1) also generates the corresponding Feller evolution. Therefore we fix s ∈
[0, T ] and take f ∈ D(L(s)) ⊂ Cb (E). Using the fact that L generates the
Markov process in (3.1) we infer for h ∈ (0, T − s):

P (s, s + h)f (x) − f (x) d ¯


lim = P (s, s + h) ¯h=0
h↓0 h dh
d ¯
= Es,x [f (X (s + h))] ¯h=0 = Es,x [L(s)f (X(s))] = L(s)f (x). (3.4)
dh
To such a Feller evolution {P (τ, t) : 0 ≤ τ ≤ t ≤ T } we may also associate a
semigroup of operators S(ρ) acting on the space Cb ([0, T ] × E) and the cor-
responding resolvent family {R(α) : <α > 0}. The semigroup {S(ρ) : ρ ≥ 0}
is defined by the formula:

S(ρ)f (t, x) = P (t, (ρ + t) ∧ T ) f ((ρ + t) ∧ T, ·) (x)


= Et,x [f ((ρ + t) ∧ T, X ((ρ + t) ∧ T ))] , (3.5)

f ∈ Cb ([0, T ] × E), (t, x) ∈ [0, T ] × E. Notice that the operator S(ρ) does
not leave the space Cb (E) invariant: i.e. a function of the form (s, y) 7→ f (y),
f ∈ Cb (E), will be mapped to function S(ρ)f ∈ Cb ([0, T ] × E) which really
depends on the time variable. Then the resolvent operator R(α) which also
acts as an operator on the space of bounded continuous functions on space-
time space Cb ([0, T ] × E) is given by
Z ∞
R(α)f (t, x) = e−α(ρ−t) P (t, ρ ∧ T ) f (ρ ∧ T, ·) (x)dρ
Zt ∞
= e−αρ P (t, (ρ + t) ∧ T ) f ((ρ + t) ∧ T, ·) (x)dρ
0
Z ∞
= e−αρ S(ρ)f (t, x) dρ
0
·Z ∞ ¸
−αρ
= Et,x e f ((ρ + t) ∧ T, X ((ρ + t) ∧ T )) dρ , (3.6)
0

f ∈ Cb ([0, T ] × E), (t, x) ∈ [0, T ] × E. In order to prove that the family


{R(α) : <α > 0} is a resolvent family indeed it suffices to establish that the
family {S(ρ) : ρ ≥ 0} is a semigroup. Let f ∈ Cb ([0, T ] × E) and fix 0 ≤ ρ1 ,
ρ2 < ∞. Then this fact is a consequence of the following identities:

S (ρ1 ) S (ρ2 ) f (t, x)


= P (t, (ρ1 + t) ∧ T ) [y 7→ S (ρ2 ) f ((ρ1 + t) ∧ T, y)] (x)
= P (t, (ρ1 + t) ∧ T )
[y 7→ P ((ρ1 + t) ∧ T, (ρ2 + ρ1 + t) ∧ T ) f ((ρ2 + ρ1 + t) ∧ T, y)] (x)

(use evolution property)


102 3 Space-time operators

= P (t, (ρ2 + ρ1 + t) ∧ T ) f ((ρ2 + ρ1 + t) ∧ T, ·) (x)


= S (ρ2 + ρ1 ) f (t, x). (3.7)
(1)
Let D1 : Cb [0, T ] → Cb ([0, T ]) be the time derivative operator. Then the
space-time operator D1 + L defined by

(D1 + L) f (t, x) = D1 f (t, x) + L(t)f (t, ·) (x), f ∈ D (D1 + L) ,

turns out to be the generator of the semigroup {S(ρ) : ρ ≥ 0}. We also ob-
serve that once the semigroup {S(ρ) : ρ ≥ 0} is known, the Feller evolution
{P (τ, t) : 0 ≤ τ ≤ t ≤ T } can be recovered by the formula:

P (τ, t) f (x) = S (t − τ ) f (τ, x), f ∈ Cb (E), (3.8)

where at the right-hand side of (3.8) the function f is considered as the func-
tion in Cb ([0, T ] × E) given by (s, y) 7→ f (y).
Theorem 3.2. Let {P (τ, t) : 0 ≤ τ ≤ t ≤ T } be a Feller propagator. Define
the corresponding Tβ -continuous semigroup {S(ρ) : ρ ≥ 0} as in (3.5). Define
the resolvent
¡ family
¢ {R(α) : α > 0} as in (3.6). Let L(1)¡ be its generator.
¢
(1)
Then αI − L R(α)f = f , f ∈ Cb ([0, T ] × E), R(α) αI − L(1) f = f ,
¡ (1) ¢
f ∈ D L , and L(1) extends D1 + L. Conversely, if the operator L(1) is
defined by L(1) R(α)f = αR(α)f − f , f ∈ Cb ([0, T ] × E), then L(1) generates
the semigroup {S(ρ) : ρ ≥ 0}, and L(1) extends the operator D1 + L.
Proof. By definition we know that
1 ³ ´
L(1) f = Tβ - lim (S(t) − S(0)) f , f ∈ D L(1) . (3.9)
t↓0 t
¡ ¢
Here D L(1) is the subspace of those f ∈ Cb ([0, T ] × E) for which the limit
in (3.9) exists. Fix f ∈ Cb ([0, T ] × E), and α > 0. Then
Z
¡ ¢ ∞ −αρ
I − e−αt S(t) e S(ρ)f dρ
0
Z ∞ Z ∞
= e−αρ S(ρ)f dρ − e−αρ e−α(t+ρ) S (t) S (ρ) f dρ
0 0
Z ∞ Z ∞
= e−αρ S(ρ)f dρ − e−αρ S(ρ)f dρ
0 t
Z t
= e−αρ S(ρ)f dρ. (3.10)
0
¡ ¢ ¡ ¢
From (3.10) it follows that R(α)f ∈ D L(1) , and that αI − L(1) R(α)f =
¡ (1) ¢
f . Conversely, let f ∈ D L . Then we have
³ ´ 1¡ ¢
R(α) αf − L(1) f = R(α)Tβ - lim f − e−αt S(t)f dρ
t↓0 t
3.1 Space-time operators 103

1¡ ¢
= Tβ - lim R(α)f − R(α)e−αt S(t)f dρ
t↓0 t
Z
1 t −αρ
= Tβ - lim e S(ρ)f dρ = f. (3.11)
t↓0 t 0

The first part of Theorem 3.2 follows from (3.10) and (3.11). In order to show
that L(1) extends D1 + L we recall the definition of generator of a Feller
P (s, t)f − f
evolution as given in Definition 1.30: L(s)f = Tβ -lim . So that
¡ ¢ t↓s t−s
if f ∈ D(1) (L), then f ∈ D L(1) , and L(1) f = D1 f + Lf . Recall that
Lf (s, x) = L(s)f (s, ·) (x). Next, if the operator L0 is defined by L0 R(α)f =
αR(α)f − f , f ∈ Cb ([0, T ] × E). Then necessarily we have L0 = L(1) , and
hence L0 generates the semigroup {S(ρ) : ρ ≥ 0}.

Theorem 3.3. Let L be a linear operator with domain D(L) and range R(L)
in Cb (E). The following assertions are equivalent:
(i) The operator L is Tβ -closable and its Tβ -closure generates a Feller semi-
group.
(ii)The operator L verifies the maximum principle, its domain D(L) is Tβ -
dense in Cb (E), it is Tβ -dissipative and sequentially λ-dominant for some
λ > 0, and there exists λ0 > 0 such that the range R (λ0 I − L) is Tβ -dense
in Cb (E).

In the definitions 3.4 – 3.6 the notions of maximum principle, dissipativity,


and sequential λ-dominance are explained. In the proof we will employ the
results of Proposition 1.22.
Definition 3.4. An operator L with domain and range in Cb (E) is said to
satisfy the maximum principle, if for every f ∈ D (L) there exists a sequence
(xn )n∈N ⊂ E with the following properties:

lim <f (xn ) = sup <f (x), and lim <Lf (xn ) ≤ 0. (3.12)
n→∞ x∈E n→∞

In assertion (b) of Proposition 3.11 it will be shown that (3.12) is equivalent


to the inequality in (3.46).
Definition 3.5. An operator L with domain and range in Cb (E) is called
dissipative if

kλf − Lf k∞ ≥ λ kf k∞ , for all λ > 0, and for all f ∈ D(L). (3.13)

An operator L with domain and range in Cb (E) is called Tβ -dissipative if there


exists λ0 ≥ 0 such that for every function u ∈ H + (E) there exists a function
v ∈ H + (E) such that

kv (λf − Lf )k∞ ≥ λ kuf k∞ , for all λ ≥ λ0 , and all f ∈ D(L). (3.14)


104 3 Space-time operators

An operator L with domain and range in Cb (E) is called positive Tβ -dissipative


if there exists λ0 > 0 such that for every function u ∈ H + (E) there exists a
function v ∈ H + (E) for which

sup v(x)< (λf (x) − Lf (x)) ≥ λ sup u(x)<f (x), (3.15)


x∈E x∈E

for all λ ≥ λ0 , and for all f ∈ D(L).


The definition which follows is crucial in proving that an operator L (or its
Tβ -closure) generates a Tβ -continuous Feller semigroup. The symbol K(E)
stands for the collection of compact subsets of E. The mapping f 7→ Uλ1 (f ),
f ∈ Cb (E, R), was introduced in (1.41).
Definition 3.6. Let L be an operator with domain and range in Cb (E) and
fix λ > 0. Let f ∈ Cb (E, R), λ > 0, and put

Uλ1 (f ) = sup inf {g ≥ f 1K : (λI − L) g ≥ 0} . (3.16)


K∈K(E) g∈D(L)

The operator L is called sequentially λ-dominant if¡ for every sequence


¢ (fn )n∈N ,
λ 1
which decreases pointwise to zero, the sequence fn = Uλ (f ) n∈N defined as
in (3.16) possesses the following properties:
1. The function f¡nλ dominates
¢ fn : fn ≤ fnλ , and
λ
2. The sequence fn n∈N converges to zero uniformly on compact subsets of
E: lim sup fnλ (x) = 0 for all K ∈ K(E).
n→∞ x∈K

The functions fnλ automatically have the first property, provided that the con-
stant functions belong to D(L) and that L1 = 0. The real condition is given
by the second property. Some properties of the mapping Uλ1 : Cb (E, R) →
L∞ (E, E, R) were explained in Proposition 1.22.
If in Definition 3.6 Uλ1 is a mapping from Cb (E, R) to itself, then Dini’s lemma
implies that in (2) uniform convergence on compact subsets of E may be
replaced by pointwise convergence on E.
Remark 3.7. Suppose that the operator L in Definition 3.6 satisfies the max-
imum principle and that (µI − L) D(L) = Cb (E), µ > 0. Then the inverses
−1
R(µ) = (µI − L) , µ > 0, exist and represent positivity preserving opera-
tors. If a function g ∈ D(L) is such that (λI − L) g ≥ 0, then g ≥ 0 and
((λ + µ) I − L) g ≥ µg, µ ≥ 0. It follows that g ≥ µR (λ + µ) g, µ ≥ 0. In the
literature functions g ∈ Cb (E) with the latter property are called λ-super-
median. For more details see e.g. Sharpe [208]. If the operator L generates a
Feller semigroup {S(t) : t ≥ 0}, then a function g ∈ Cb (E) is called λ-super-
mean valued if for every t ≥ 0 the inequality e−λt S(t)g ≤ g holds pointwise. In
Lemma (9.12) in Sharpe [208] it is shown that, essentially speaking, these no-
tions are equivalent. In fact the proof is not very difficult. It uses the Hausdorff-
Bernstein-Widder theorem about the representation by Laplace transforms of
3.1 Space-time operators 105

positive Borel measures on [0, ∞) of completely positive functions. It is also


implicitly proved in the proof of Theorem 3.10 implication (iii) =⇒ (i): see
(in-)equalities (3.131), (3.132), (3.133), (3.134), and (3.140).

Proof (Proof of Theorem 3.3). (i) =⇒ (ii). Let L be the Tβ -closure


¡ of¢ L,
which is the Tβ -generator of the semigroup {S(t) : t ≥ 0}. Then R λI − L =
Cb (E), and the
R ∞ inverses of λI − L which we denote by R(λ) exist and satisfy:
R(λ)f (x) = 0 e−λt S(t)f (x). It follows that
Z ∞
λ< (R(λ)f (x)) = λ e−λt (S(t)<f ) (x)dt
0
Z ∞
≤λ e−λt dt sup <f (y) = sup <f (y), (3.17)
0 y∈E y∈E

and hence supx∈E λ< (R(λ)f (x)) ≤ supy∈E <f (y). The substitution f = λg −
Lg yields:
¡ ¢ ¡ ¢
λ sup <g(x) ≤ sup < λg(y) − Lg(y) , g ∈ D L . (3.18)
x∈E y∈E

In other words, the operator L satisfies the maximum principle, and so does
the operator L: see Proposition 3.11 assertion (b) below. Since the operator
L is Tβ -dissipative, the resolvent families {R(λ) : λ ≥ λ0 }, λ0 > 0, are Tβ -
equi-continuous.
R Hence every operator R(λ) can be written as an integral:
R(λ)f (x) = f (y)r (λ, x, dy), f ∈ Cb (E). For this the reader may consider the
arguments in (the proof
© of) Proposition
ª 1.22. Moreover, we have that for every
λ0 > 0, the family e−λ0 t S(t) : t ≥ 0 is Tβ -equi-continuous, and in addition,
lim S(t)f (x) = f (x), f ∈ Cb (E). It then follows that lim λR(λ)f (x) = f (x),
t↓0 λ→∞
f ∈ Cb (E). As in the proof of Proposition 1.22 we see that Tβ - lim λR(λ)f =
λ→∞
f , f ∈ Cb (E): see e.g. (1.58). Let f ≥ 0 belong to Cb (E), and consider the
function Uλ1 (f ) defined by

Uλ1 (f ) = sup inf {g ≥ f 1K : λg − Lg ≥ 0} . (3.19)


K∈K(E) g∈D(L)

In fact this definition is copied from (1.41). As was shown in Proposition 1.22,
we have the following equality:
n¡ ¢k o © ª
Uλ1 (f ) = sup (λ + µ) I − L f : µ > 0, k ∈ N = sup e−λt S(t)f : t ≥ 0 .
(3.20)
In fact in Proposition 1.22 the first equality in (3.20) was proved. The second
equality follows from the representations:
Z ∞
k µk
(µR (λ + µ)) f = tk−1 e−µt e−λt S(t)f dt and (3.21)
(k − 1)! 0
106 3 Space-time operators

X k
(µt) k
e−λt S(t)f = Tβ - lim e−µt (µR (λ + µ)) f. (3.22)
µ→∞ k!
k=0

A similar argument will be used in the proof of Theorem 3.10 (iii) =⇒ (i): see
(3.133) and (3.134). The representation in (3.20) implies that the operator L
is λ-dominant. Altogether this proves the implication (i) =⇒ (ii) of Theorem
3.3.
(ii) =⇒ (i). As in Proposition 3.11 assertion (a) below, the operator L is
Tβ -closable. Let L be its Tβ -closure. Then the operator L is Tβ -dissipative,
¡ ¢
λ-dominant, and satisfies the maximum principle. In addition R λI − L =
¡ ¢−1
Cb (E), λ > 0. Consequently, the inverses R(λ) = λI − L , λ > 0, exist.
The formulas in (3.21) and (3.22) can be used to represent the powers of the
resolvent operators, and to define the Tβ -continuous semigroup generated by
L. The λ-dominance is used in a crucial manner to prove that the semigroup
represented by (3.22) is a Tβ -equi-continuous semigroup which consists of
operators, which assign bounded continuous functions to such functions. For
details the reader is referred to the proof of Theorem 3.10 implication (iii)
=⇒ (i), where a very similar construction is carried for a time space operator
L(1) which is the Tβ -closure of D1 + L. In Theorem 3.10 the operator D1 is
taking derivatives with respect to time, and L generates a Feller evolution.

Proposition 3.8. Let L be a Tβ -closed linear operator with domain and


range in Cb (E). Suppose that the operator L satisfies the maximum prin-
ciple,
n and is such that R (λI o− L) = Cb (E), λ > 0. Then the resolvent family
−1
R(λ) = (λI − L) : λ > 0 consists of positivity preserving operators. In
addition, suppose that L possesses a Tβ -dense domain, and that the following
limits exist: for all (t, x) ∈ [0, ∞) × E and for all f ∈ Cb (E)

X (µt)k k
e−λt S(t)f (τ, x) = lim e−µt (µR (λ + µ)) f (τ, x), (3.23)
µ→∞ k!
k=0

and for all f ∈ D(L) and x ∈ E

lim µ (I − µR (λ + µ)) f (x) = λf (x) − Lf (x). (3.24)


µ→∞

Moreover, suppose that the operators R(λ), λ > 0, are Tβ -continuous. Fix
f ∈ Cb (E), f ≥ 0, and λ > 0. The following equalities and equality hold true:

sup inf {g ≥ f 1K : (λI − L) g ≥ 0} (3.25)


K∈K(E) g∈D(L)

= sup inf {g ≥ f 1K : g ≥ µR(λ + µ)g, for all µ > 0} (3.26)


K∈K(E) g∈Cb (E)
n o
k
≥ sup (µR (λ + µ)) f : µ ≥ 0, k ∈ N (3.27)
© ª
= sup e−λt S(t)f : t ≥ 0 . (3.28)
3.1 Space-time operators 107

If the
© function (t, x) ª7→ S(t)f (x) is continuous, then the function g =
sup e−λt S(t)f : t ≥ 0 is continuous, realizes the infimum in (3.26), and the
expressions (3.25) through (3.28) are all equal.

The proof of the following corollary is an immediate consequence of Proposi-


tion 3.8.
Corollary 3.9. Suppose that the operator L with domain and range in Cb (E)
be the Tβ -generator of a Feller semigroup {S(t) : t ≥ 0}. Let f ≥ 0 belong to
Cb (E). Then the quantities in (3.25) through (3.28) are equal.
Let g ∈ D(L). By assumption (3.24) we see that λg − Lg ≥ 0 if and only if
g ≥ µR (λ + µ) g for all µ > 0. Hence we have

inf {g ≥ f 1K : (λI − L) g ≥ 0}
g∈D(L)

= inf {g ≥ f 1K : g ≥ µR(λ + µ)g, for all µ > 0} . (3.29)


g∈D(L)

It is not so clear under what conditions we have equality of (3.29) and (3.26).
If f ∈ D(L) is such that λf − Lf ≥ 0, then the functions in (3.25) through
(3.28) are all equal to f .
Proof (Proof of Proposition 3.8). The representation in (3.23) shows that the
term in (3.28) is dominated by the one in (3.27). The equality
Z ∞
k µk
(µR (λ + µ)) f = tk−1 e−(λ+µ)t S(t)f dt, k ≥ 1, (3.30)
(k − 1)! 0

shows that the expression in (3.27) is less than or equal to the one in (3.28).
Altogether this proves the equality of (3.27) and (3.28). If the function g ∈
D(L) is such that g ≥ f 1K and (λI − L) g ≥ 0, then ((λ + µ) I − L) g ≥ µg,
and hence
k
g ≥ µR (λ + µ) g ≥ (µR (λ + µ)) g for all k ∈ N.

Consequently, the term in (3.25) dominates the second one. It also follows
that the expression in (3.26) is greater than or equal to
n o
k
sup sup (µR (λ + µ)) (f 1K ) : µ > 0, k ∈ N . (3.31)
K∈K(E)

k
Since the operators (µR (λ + µ)) , µ > 0 and k ∈ N, are Tβ -continuous the
expression in (3.31) is equal to the quantity in (3.27). Next we will show that
the expression in (3.26) is less than equal to (3.25). Therefore we chose an
arbitrary compact subset K of E. Let g ∈ Cb (E) be function with the following
properties: g ≥ f 1K , and g ≥ µR (λ + µ) g. Then for η > 0 arbitrary small
and α = αη > 0 sufficiently large we have αR(α) (g + η) ≥ g1K ≥ f 1K .
Moreover, the function gα,η := αR(α) (g + η) belongs to D(L) and satisfies
108 3 Space-time operators

gα,η ≥ µR (λ + µ) gα,η for all µ > 0 (3.32)

Here we employed the fact that D(L) is Tβ -dense in Cb (E). In fact we used the
fact that, uniformly on the compact subset K, g + η = limα→∞ αR(α) (g + η).
From (3.32) we obtain

(λI − L) gα,η = lim µ (I − µR (λ + µ)) gα,η ≥ 0, (3.33)


µ→∞

From (3.33) we obtain the inequality:

inf {g ≥ f 1K : g ≥ µR (λ + µ) g}
g∈Cb (E)

≥ inf {g ≥ f 1K : g ≥ µR (λ + µ) g} . (3.34)
g∈D(L)

The inequality in (3.34) shows that the expression in (3.26) is less than or equal
to the one in (3.25). Thus far we showed (3.25) = (3.26) ≥ (3.27) = 3.28). The
final assertion about the fact that the (continuous) function in (3.28) realizes
the equality in (3.27) being obvious, concludes the proof of Proposition 3.8.
In the following theorem (Theorem 3.10) we use the following subspaces of
the space Cb ([0, T ] × E):
½
(1)
CP,b = f ∈ Cb ([0, T ] × E) : all functions of the form (τ, x) 7→
Z τ +ρ ¾
P (τ, σ) f (σ, ·) (x)dσ, ρ > 0, belong to D (D1 ) ; (3.35)
τ
½
(1)
CP,b (λ) = f ∈ Cb ([0, T ] × E) : the function (τ, x) 7→
Z ∞ ¾
−λσ
e P (τ, σ) f (σ, ·) (x)dσ, belongs to D (D1 ) . (3.36)
τ

(1) (1)
Here λ > 0, and CP,b is a limiting case if λ = 0. The inclusion CP,b ⊂
(1)
∩λ0 >0 CP,b (λ0 ) follows from the representation of R (λ0 ) as a Laplace trans-
form:
Z ∞
R (λ0 ) f (τ, x) = e−λ0 ρ S(ρ)f (τ, x)dρ
0
Z ∞
= e−λ0 ρ P (τ, τ + ρ) f (τ + ρ, x)dρ
0
Z ∞ Z ρ
= λ0 e−λ0 ρ P (τ, τ + σ) f (τ + σ, x)dσ dρ
0 0
Z ∞ Z τ +ρ
= λ0 e−λ0 ρ P (τ, σ) f (σ, x)dσ dρ (3.37)
0 τ

From (3.37) we see that if for every ρ > 0 the function


3.1 Space-time operators 109
Z ρ Z τ +ρ
(τ, x) 7→ S(σ)f (τ, x)dσ = P (τ, σ ∧ T ) f (σ ∧ T, x)dσ
0 τ

belongs to D (D1 ), then so does theR function (τ, x) 7→ R (λ0 ) f (τ, x), pro-
ρ
vided that the function ρ 7→ e−λ0 ρ D1 0 S(σ)f dσ is Tβ -integrable in the space
(1) (1)
Cb ([0, T ] × E). The other inclusion, i.e. ∩λ0 >0 CP,b (λ0 ) ⊂ CP,b follows from
the following inversion formula:
Z τ +ρ Z ρ
P (τ, σ) f (σ, ·) (x)dσ = S(σ)f (τ, x) dσ
τ 0
Z ρ
2
= lim e−σλ eσλ R(λ) f (τ, x) dσ
λ→∞ 0
X∞ Z
1 ρ k k
= lim (σλ) e−σλ (λR(λ)) f (τ, x) dσ
λ→∞ k! 0
k=0
X∞ Z ρ
λk+1 k+1 −σλ k+1
= lim (σλ) e (R(λ)) f (τ, x) dσ
λ→∞ (k + 1)! 0
k=0
X∞ Z ρ Z ∞
λk+1 k+1 −σλ
= lim (σλ) e ρk1 e−λρ1 S (ρ1 ) f (τ, x) dρ1 dσ
λ→∞ (k + 1)!k! 0 0
k=0
X∞ Z k Z ∞
(−1)k λk+1 ρ k+1 −σλ ∂
= lim (σλ) e k
e−λρ1 S (ρ1 ) f (τ, x) dρ1 dσ
λ→∞ (k + 1)!k! 0 (∂λ) 0
k=0
X∞ k k+1 Z ρ k
(−1) λ k+1 −σλ ∂
= lim (σλ) e k
R(λ)f (τ, x)dσ (3.38)
λ→∞ (k + 1)!k! 0 (∂λ)
k=0

where the limits


R ρ have to taken in Tβ -sense. A similar limit representation is
valid for D1 0 S(ρ)f dρ(τ, x), provided that the family
½ ¾
λk+1
D1 R(λ)k f : λ > 0, k ∈ N
k!

is uniformly bounded. A simpler approach might be to use a complex inversion


formula:
Z ρ
(τ + ρ − σ) P (τ, (τ + σ) ∧ T ) f ((τ + σ) ∧ T, ·) (x)dσ
0
Z ρ Z ρ1 Z ω+i∞
1 1
= S(σ)f (τ, x)dσ dρ1 = eρλ R(λ)f (τ, x)dλ, (3.39)
0 0 2πi ω−i∞ λ2

and to assume that, for ω > 0, the family {λD1 R(λ)f : <λ ≥ ω} is uniformly
bounded. It is clear that the operator R(λ), <λ > 0, stands for
Z ∞
R(λ)f (τ, x) = e−λρ S(ρ)f (τ, x)dρ (3.40)
0
110 3 Space-time operators
Z ∞
= e−λρ P (τ, (τ + ρ) ∧ T ) f ((τ + ρ) ∧ T, ·) (x)dρ, f ∈ Cb ([0, T ] × E) .
0

It is also clear that the family of operators in (3.38) is a once integrated


integrated semigroup, and that the family in (3.39) is a twice integrated semi-
(1) (1)
group. In order to justify the inclusion ∩λ0 >0 CP,b (λ0 ) ⊂ CP,b in both ap-
proaches we need to know that the functions: λ 7→ R(λ)f , and λ 7→ D1 R(λ)f
are real analytic. For more details on inversion formulas for vector-valued
Laplace transforms and integrated semigroups see e.g. Bobrowski [37], Cho-
jnacki [57], Arendt [8], Arendt et al [9], and Miana [163]. For vector valued
Laplace transforms the reader is also referred to Bäumer and Neubrander [26].
Theorem 3.10. Let L be a linear operator with domain D(L) and range R(L)
in Cb ([0, T ] × E). Suppose that there exists λ > 0 such that the operator
D1 + L is sequentially λ-dominant in the sense of Definition 3.6. Under such
a hypothesis the following assertions are equivalent:
(i) The operator L is Tβ -closable, its Tβ -closure generates a Feller evolution,
the operator D1 + L is Tβ -densely defined, and there exists λ0 > 0 such
(1)
that the subspace CP,b (λ0 ) is Tβ -dense in Cb ([0, T ] × E).
(ii)The operator D1 + L is Tβ -closable and its Tβ -closure generates a Tβ -
continuous Feller semigroup in Cb ([0, T ] × E).
(iii)The operator D1 +L is Tβ -densely defined, is power Tβ -dissipative, satisfies
the maximum principle, and there exists λ0 > 0 such that the range of
λ0 I − D1 − L is Tβ -dense in Cb ([0, T ] × E).
Here we call the operator D1 + L power Tβ -dissipative if for some λ0 ≥ 0 and
for every k ∈ N there exists a Tβ -dense subspace Dk of Cb ([0, T ] × E) such
that for every u ∈ H + ([0, T ] × E) there exists v ∈ H + ([0, T ] × E) for which
the following inequality holds:
° °
° k °
λk kuf k∞ ≤ °v (λI − D1 − L) f ° for all f ∈ Dk and all λ ≥ λ0 . (3.41)

If the operator D1 + L is just Tβ -dissipative, then an inequality of the form


(3.41) holds, with a function v ∈ H + ([0, T ] × E) which depends on k. In (3.41)
the function v only depends on u (and the operator D1 + L), but it neither
depends on k, nor on f ∈ Dk or λ ≥ 1. Let the operator L(1) be an extension
of D1 + L which generates a Tβ -continuous semigroup {S0 (t) : t ≥ 0}, and
suppose that D1 + L satisfies (3.41). Then this semigroup is equi-continuous
in the sense that for every u ∈ H + ([0, T ] × E) there exists v ∈ H + ([0, T ] × E)
for which the following inequality holds:

kuS0 (t)f k∞ ≤ kvf k∞ for all f ∈ Cb ([0, ∞) × E) and all t ∈ [0, ∞). (3.42)

A closely related inequality is the following one


° °
° k °
°u (λR(λ)) f ° ≤ kvf k∞ , λ ≥ 1, f ∈ Cb ([0, T ] × E) . (3.43)

3.2 Dissipative operators and maximum principle 111

Notice that (3.43) is equivalent to (3.41) provided that the operator L(1) is
the Tβ -closure of D1 + L and the ranges of λI − L(1) , λ > 0, coincide with
Cb ([0, T ] × E). In fact the semigroup {S0 (t) : t ≥ 0} and the resolvent family
{R(λ) : λ > 0} are related as follows:
Z
k λk ∞ k−1 −λt
(λR (λ)) f = t e S0 (t)f dt, and (3.44)
k! 0
X∞
(λt)k k
S0 (t)f = Tβ - lim e−λt (λR(λ)) f. (3.45)
λ→∞ k!
k=0

The integral in (3.44) has to be interpreted in Tβ -sense. From (3.44) and


(3.45) the equivalence of (3.42) and (3.53) easily follows. We also observe that
(3.43) is equivalent to the following statement. For every sequence (fn )n∈N ⊂
Cb ([0, T ] × E) which decreases pointwise to zero it follows that
k
inf sup (λR(λ)) fn = 0.
n∈N λ≥1,k∈N

3.2 Dissipative operators and maximum principle


In the following proposition we collect some of the interrelationships which
exist between the concepts of closability, dissipativeness, and maximum prin-
ciple. A reformulation of assertion (f) in Proposition 3.11 can be found in
Lemma 7.3 in Chapter 7.
Proposition 3.11. (a1 ) Suppose that the operator L is dissipative and that
its range is contained in the closure of its domain. Then the operator L is
closable.
(a2 ) Suppose that the operator L is Tβ -dissipative and that its range is con-
tained in the Tβ -closure of its domain. Then the operator L is Tβ -closable.
(b) If the operator L satisfies the maximum principle, then
sup < (λf (x) − Lf (x)) ≥ λ sup <f (x), for all λ > 0, and for all f ∈ D(L).
x∈E x∈E
(3.46)
Conversely, if L satisfies (3.46), then the operator L satisfies the maxi-
mum principle. The inequality in (3.46) is equivalent to
inf < (λf (x) − Lf (x)) ≤ λ inf <f (x), for all λ > 0, and for all f ∈ D(L).
x∈E x∈E
(3.47)
(c) If the operator L satisfies the maximum principle, then L is dissipative.
(d) If the operator L satisfies the maximum principle, and if f ∈ D(L) is such
that λf − Lf ≥ 0 for some λ > 0, then f ≥ 0.
(e) If the operator L is dissipative, then
kλf − Lf k∞ ≥ <λ kf k∞ , for all λ with <λ > 0, and for all f ∈ D(L).
(3.48)
112 3 Space-time operators

(f ) The operator L is dissipative if and only if for every f ∈ D(L) there


exists a sequence (xn )n∈N ⊂ E such that lim |f (xn )| = kf k∞ , and
³ ´ n→∞
lim < f (xn )Lf (xn ) ≤ 0.
n→∞
(g) If the operator L is positive Tβ -dissipative, then it is Tβ -dissipative.
For the definition of an operator which is positive Tβ -dissipative or Tβ -
dissipative, the reader is referred to Definition 3.4. The same is true for the
other notions in Proposition 3.11.
Proof. (a1 ) Let (fn )n∈N ⊂ D(L) be any sequence with the following proper-
ties:
lim fn = 0, and g = lim Lfn
n→∞ n→∞

exists in Cb (E). Then we consider


° °
°(λfn + gm ) − λ−1 L (λfn + gm )° ≥ kλfn + gm k ,
∞ ∞

where (gm )m∈N ⊂ D(L) converges to g. First we let n tend to infinity, then
λ, and finally m. This limiting procedure results in

lim kgm − gk∞ ≥ lim kgm k∞ = kgk∞ .


m→∞ m→∞

Hence g = 0.
(a2 ) Let (fn )n∈N ⊂ D(L) be any sequence with the following properties:

Tβ - lim fn = 0, and g = Tβ - lim Lfn


n→∞ n→∞

exists in Cb (E). Let u ∈ H + (E) be given and let the function v be as in (3.14).
Then we consider
° ¡ ¢°
°v (λfn + gm ) − λ−1 L (λfn + gm ) ° ≥ ku (λfn + gm )k , (3.49)
∞ ∞

where (gm )m∈N ⊂ D(L) Tβ -converges to g. First we let n tend to infinity, then
λ, and finally m. The result will be

lim kvgm − vgk∞ ≥ lim kugm k∞ = kugk∞ ,


m→∞ m→∞

and hence g = 0.
(b) Let f ∈ D(L). Then choose a sequence (xn )n∈N ⊂ E as in (3.12). Then
we have

sup < (λf (x) − Lf (x)) ≥ lim < (λf (xn ) − Lf (xn )) ≥ λ sup <f (x)
x∈E n→∞ x∈E

which is the same as (3.46). Suppose that the operator L satisfies (3.46). Then
for every λ > 0 we choose xλ ∈ E such that
3.2 Dissipative operators and maximum principle 113

1
λ<f (xλ ) − <Lf (xλ ) ≥ λ sup <f (x) − . (3.50)
x∈E λ
From (3.50) we infer:
1
<Lf (xλ ) ≤ , and (3.51)
λ
1 1
sup <f (x) ≤ <f (xλ ) + − <Lf (xλ ) . (3.52)
x∈E λ2 λ
From (3.51) we see that lim supλ→∞ <Lf (xλ ) ≤ 0, and from (3.52) it follows
that lim supλ→∞ <f (xλ ) = supx∈E <f (x). From these observations it is easily
seen that (3.46) implies the maximum principle.
The substitution f → −f shows that (3.47) is a consequence of (3.46).
(c) Let f 6= 0 belong to D(L), choose α ∈ R and a sequence (xn )n∈N ⊂ E
in such a way that 0 <¢ kf k∞ = limn→∞ <eiα f (xn ) = supx∈E <eiα f (x), and
¡ iα
that limn→∞ <L e f (xn ) ≤ 0. Then
¡ ¢
kλf − Lf k∞ ≥ lim < eiα (λf − Lf ) (xn )
n→∞
¡ ¢ ¡ ¢
= lim λ< eiα f (xn ) − < eiα Lf (xn ) ≥ λ kf k∞ . (3.53)
n→∞

The inequality in (3.53) means that L is dissipative in the sense of Definition


3.5.
(d) Let f ∈ D(L) be such that for some λ > 0 λf (x) − Lf (x) ≥ 0 for all
x ∈ E. From (3.47) in (b) we see that
λ inf =f (x) = λ inf <(−if )(x) ≥ inf < (λ(−if )(x) − L(−if )(x))
x∈E x∈E x∈E
= inf = (λf (x) − L(f )(x)) = 0. (3.54)
x∈E

From (3.54) we get =f ≥ 0. If we apply the same argument to −f instead of


f we get =f ≤ 0. Hence =f ≡ 0, and so the function f is real-valued. But
then we have
0 ≤ inf (λf (x) − Lf (x)) ≤ inf f (x),
x∈E x∈E

and consequently f ≥ 0.
(e) From the proof it follows that L is dissipative if and only if for every

f ∈ D(L) there exists an element x∗ in Cb ([0, T ] × E) such that kx∗ k = 1,
∗ ∗
such that hf, x i = kf k∞ , and such that < hLf, x i ≤ 0. A proof of all this
runs as follows. Let L be dissipative. Fix f in D(L) and choose for each λ > 0

an element x∗λ in Cb ([0, T ] × E) in such a way that kx∗λ k ≤ 1 and
kλf − Lf k∞ = hλf − Af, x∗λ i . (3.55)
T
Choose an element x∗ in the intersection µ>0 weak∗ closure {x∗λ : λ > µ}.

Since the dual unit ball of Cb ([0, T ] × E) is weak∗ -compact such an element

x exists. From (3.55) it follows that
114 3 Space-time operators

< hLf, x∗λ i = λ< hf, x∗λ i − kλf − Lf k∞


≤ λ kf k∞ − kλf − Lf k∞ ≤ 0, λ > 0. (3.56)

Here we used the fact that A is supposed to be dissipative. From (3.55) we


also obtain the equality
° °
hf, x∗λ i = °f − λ−1 Lf °∞ + λ−1 hLf, x∗λ i , λ > 0. (3.57)

Since x∗ is a weak∗ limit point of {x∗λ : λ > µ} for each µ > 0 it follows from
(3.56) and (3.57) that

< hLf, x∗ i ≤ 0, and (3.58)


hf, x∗ i = kf k∞ , kx∗ k ≤ 1. (3.59)

Finally pick λ ∈ C with <λ > 0. From (3.58) and (3.59) we infer

kλf − Lf k∞ ≥ < hλf − Lf, x∗ i = < (λ hf, x∗ i) − < hLf, x∗ i


≥ < (λ kf k∞ ) − 0 = <λ kf k∞ . (3.60)

(f) If L is dissipative and if f ∈ D(L), then there exists a family (xλ )λ>0 ⊂ E
such that
kLf k∞
|λf (xλ ) − Lf (xλ )| ≥ λ kf k∞ − . (3.61)
λ
From (3.61) we infer

kLf k∞
λ |f (xλ )| + kLf k∞ ≥ λ kf k∞ − , (3.62)
λ

and
³ ´
2 2
λ2 |f (xλ )| − 2λ< f (xλ )Lf (xλ ) + |Lf (xλ )|
2
2 kLf k∞
≥ λ2 kf k2 − 2 kf k∞ kLf k∞ + . (3.63)
λ2
From (3.62) and (3.63) we easily infer

kLf k∞ kLf k∞
|f (xλ )| ≥ kf k∞ − − , (3.64)
λ λ2

and
³ ´
2 2
λ2 kf k∞ − 2λ< f (xλ )Lf (xλ ) + kLf k∞
2
2 kLf k∞
≥ λ2 kf k∞ − 2 kf k∞ kLf k∞ + . (3.65)
λ2
From (3.65) we get
3.2 Dissipative operators and maximum principle 115
³ ´ µ ¶
kf k∞ kLf k∞ 1 1 2
< f (xλ )Lf (xλ ) ≤ + 1 − 2 kLf k∞ . (3.66)
λ 2λ λ
³ ´
From (3.66) we obtain lim sup < f (xλ )Lf (xλ ) ≤ 0. From (3.64) we see
λ→∞
lim |f (xλ )| = kf k∞ . By passing to a countable sub-family we see that there
λ→∞
exists a sequence (xn )n∈N ⊂ E such that lim |f (xn )| = kf k∞ and such
³ ´ n→∞
that the limit lim < f (xn )Lf (xn ) exists and is ≤ 0. The proof of the
n→∞
converse statement is (much) easier. Let (xn )n∈N ⊂ ³E be a sequence
´ such
that lim |f (xn )| = kf k∞ and that the limit lim < f (xn )Lf (xn ) exists
n→∞ n→∞
and is ≤ 0. Fix f ∈ D(L). Then we have
³ ´
2 2 2
kλf − Lf k∞ ≥ λ2 |f (xn )| − 2λ< f (xn )Lf (xn ) + |Lf (xn )|
³ ´
2
≥ λ2 |f (xn )| − 2λ< f (xn )Lf (xn ) . (3.67)

From the properties of the sequence (xn )n∈N and (3.67) we obtain the inequal-
ity kλf − Lf k∞ ≥ λ kf k∞ , λ > 0, f ∈ D(L), which is the same as saying that
L is dissipative.
(g) Let the functions u and v ∈ H + (E) as in assertion (g), let f ∈ D(L), and
λ ≥ λ0 . Then we have
¡ ¡ ¢ ¡ ¢ ¢
kv (λf − Lf )k∞ = sup sup v(x)< λ eiϑ f (x) − L eiϑ f (x)
ϑ∈[−π,π] x∈E

(L is positive Tβ -dissipative)
¡ ¢
≥λ sup sup u(x)< eiϑ f (x) = λ kuf k∞ . (3.68)
ϑ∈[−π,π] x∈E

The inequality in (3.68) shows the dissipativity of the operator L.

Proof (Proof of Theorem 3.10). (i) =⇒ (ii). Let L be the Tβ -closure of L.


Then there exists a Feller evolution {P (s, t) : 0 ≤ s ≤ t ≤ T } such that

d ¡ ¢
P (τ, t) f (t, ·) (x) = P (τ, t) D1 + L(t) f (t, ·) (x), (3.69)
dt
¡ ¢
for all functions f ∈ D(1) L , 0 ≤ τ ≤ t ≤ T , x ∈ E. The functions f ∈
¡ ¢
D(1) L have the property that for every ρ ∈ [0, T ] the following Tβ -limits
exist:
P (ρ, ρ + h) f (ρ, ·) − f (ρ, ·)
(a) L(ρ)f (ρ, ·) (x) = Tβ - lim .
h↓0 h
∂ f (ρ + h, x) − f (ρ, x)
(b) f (ρ, x) = Tβ - lim .
∂ρ h→0 h
116 3 Space-time operators

As indicated the limits in (a) and (b) have to be interpreted in Tβ -sense.


Moreover, these functions as functions of the pair (ρ, x) are supposed to be
continuous. The equality in (3.69) was introduced in Definition 1.31. How-
ever, the reader is also referred to Proposition 3.1, and to equality (3.2). The
equality in (3.69) can also be written in integral form:
Z t µ ¶

P (τ, t) f (t, ·) (x) − f (τ, x) = P (τ, ρ) + L(ρ) f (ρ, ·) (x) (3.70)
τ ∂ρ
¡ ¢
for f ∈ D(1) L , 0 ≤ τ ≤ t ≤ T , x ∈ E. The Feller evolution is Tβ -equi-
continuous. This means that for every u ∈ H + (E), there exists v ∈ H + (E)
such that for all f ∈ Cb (E) the inequality

sup sup |u(τ, x)P (τ, t) f (·) (x)| ≤ sup |v(x)f (x)| (3.71)
τ ≤t≤T x∈E x∈E

holds
n for all f ∈ Cb (E). As
o was explained in Corollary 2.6, the Feller evolution
Pe (τ, t) : 0 ≤ τ ≤ t ≤ T , which is the same as {P (τ, t) : 0 ≤ τ ≤ t ≤ T }
considered as a family of operators on Cb ([0, T ] × E), is Tβ -equi-continuous
as well: see Corollary 1.19. As in (3.5) we define the semigroup Tβ -equi-
continuous semigroup {S(ρ) : ρ ≥ 0} by

S(ρ)f (t, x) = P (t, (ρ + t) ∧ T ) f ((ρ + t) ∧ T, ·) (x),


f ∈ Cb ([0, T ] × E) ,
(3.72)
where ρ ≥ 0, and (t, x) ∈ [0, T ] × E. Then the semigroup in (3.72) is Tτ -equi-
continuous. In fact we have

sup sup |u(τ, x)S(t)f (τ, x)| ≤ sup |v(x)f (τ, x)| (3.73)
τ ≤t≤T x∈E (τ,x)∈[0,T ]×E

where u ∈ H + ([0, T ] × E)
R ∞and v ∈ H + (E) are as in (3.71). Let L(1) be its
−λρ
generator, and R(λ)f = 0 e S(ρ)f dρ, f ∈ Cb ([0, T ] × E), its resolvent.
(1)
Then we will prove that L = D1 + L, and we will also show the following
well-known equalities (compare with (2.150)):
³ ´
λI − L(1) R(λ)f = f, f ∈ Cb ([0, T ] × E) ,
³ ´ ³ ´
R(λ) λI − L(1) f = f, f ∈ D L(1) . (3.74)

In order to understand the relationship between D1 + L and the Tβ -generator


of the semigroup {S(ρ) : ρ ≥ 0} we consider, for h > 0, λ > 0 the operators
(1) (1)
Lλ,h and ϑh Lλ,h , which are defined by

(1) 1¡ ¢
Lλ,h f (τ, x) = I − e−λh S(h) f (τ, x) (3.75)
h
1¡ ¢
= f (τ, x) − e−λh P (τ, (τ + h) ∧ T ) f ((τ + h) ∧ T, ·) (x)
h
3.2 Dissipative operators and maximum principle 117

1¡ ¢
= f (τ, x) − e−λh P (τ, (τ + h) ∧ T ) f (τ, ·) (x)
h
1 ¡ −λh ¢
− e P (τ, (τ + h) ∧ T ) (f ((τ + h) ∧ T, ·) − f (τ, ·)) (x)
h
³ ´1Z h
= λI − L(1) e−λρ S(ρ)f dρ (τ, x)
h 0

and

(1) 1¡ ¢
ϑh Lλ,h f (τ, x) = I − e−λh S(h) f ((τ − h) ∧ T ∨ 0, x) (3.76)
h
1
= (f ((τ − h) ∧ T ∨ 0, x) − f (τ, x))
h
1 ¡ −λh ¢
− e P ((τ − h) ∧ T ∨ 0, τ ) f (τ, ·) (x) − f (τ, x) .
h
The operator ϑh : Cb ([0, T ] × E) → Cb ([0, T ] × E) is defined by

ϑh f (τ, x) = f ((τ − h) ∧ T ∨ 0, x) , f ∈ Cb ([0, T ] × E) . (3.77)

Since
Z h
(1) (1) 1
Lλ,h R(λ)f = R(λ)Lλ,h f = e−λρ S(ρ)f dρ, f ∈ Cb ([0, T ] × E) ,
h 0
(3.78)
and L(1) is the Tβ -generator of the semigroup {S(ρ) : ρ ≥ 0}, the equalities
in (3.74) follow from (3.78). Since
(1) (1)
ϑh Lλ,h R(λ)f (τ, x) = Lλ,h R(λ)f ((τ − h) ∧ T ∨ 0, x) ,

it also follows that


³ ´
(1)
Tβ - lim ϑh Lλ,h R(λ)f (τ, x) = λI − L(1) R(λ)f (τ, x) . (3.79)
h↓0

A consequence of (3.79) and the second equality in (3.76) is that


³ ´ µ
(1) 1
λI − L f (τ, x) = lim (f ((τ − h) ∧ T ∨ 0, x) − f (τ, x)) (3.80)
h↓0 h

1 ¡ −λh ¢
− e P ((τ − h) ∧ T ∨ 0, τ ) f (τ, ·) (x) − f (τ, x)
h
µ
1¡ ¢
= lim f (τ, x) − e−λh P (τ, (τ + h) ∧ T ) f (τ, ·) (x)
h↓0 h

1
− ((f ((τ + h) ∧ T, ·) − f (τ, ·)) (x)) . (3.81)
h
¡ ¢
These limits exist in the strict sense; i.e. in the Tβ -topology. If f ∈ D L(1) ,
¡ ¢
and if f belongs to D (D1 ), then (3.80) and (3.81) imply that f ∈ D L , that
118 3 Space-time operators

1
L(f )(τ, x) = lim (P (τ, τ + h) f (τ, ·) (x) − f (τ, x))
h↓0 h
1
= lim (P (τ − h, τ ) f (τ, ·) (x) − f (τ, x)) , (3.82)
h↓0 h

and that
L(1) f = Lf + D1 f. (3.83)
Hence, in principle, the first term on the right-hand side in (3.80) con-
verges to¡ the negative
¢ of the time-derivative of the function f and the sec-
ond to λI − L f . The following arguments make this more precise. We
(1)
will need the fact that the subspace CP,b is Tβ -dense in Cb ([0, T ] × E). Let
f ∈ Cb ([0, T ] × E). In order to prove that, under certain conditions, ¡ ¢ the
operator L(1) is the closure of D1 + L, we consider for f ∈ D L(1) and
0 ≤ a ≤ b ≤ T the following equality:
Z b Z b
ϑρ S(ρ)L(1) f (τ, x) dρ = S(ρ)L(1) f ((τ − ρ) ∨ 0, x) dρ
a a
Z b

= ϑρ S (ρ) f (τ, x)dρ
∂τ a
+ P ((τ − b) ∨ 0, τ ) f (τ, x) − P ((τ − a) ∨ 0, τ ) f (τ, x). (3.84)

We first prove the equality on (3.84). Therefore we write


Z b

ϑρ S (ρ) f (τ, x)dρ + P ((τ − b) ∨ 0, τ ) f (τ, x) − P ((τ − a) ∨ 0, τ ) f (τ, x)
∂τ a
Z τ −a

= S (τ − ρ) f (ρ ∨ 0, x)dρ
∂τ τ −b
+ P ((τ − b) ∨ 0, τ ) f (τ, x) − P ((τ − a) ∨ 0, τ ) f (τ, x)
¡ ¢
(the function f belongs to D L(1) )
Z τ −a
= S (τ − ρ) L(1) f (ρ ∨ 0, x)dρ
τ −b
+ S (a) f ((τ − a) ∨ 0, x) − S (b) f ((τ − b) ∨ 0, x)
+ P ((τ − b) ∨ 0, τ ) f (τ, x) − P ((τ − a) ∨ 0, τ ) f (τ, x)
Z b Z b
= S (ρ) L(1) f ((τ − ρ) ∨ 0, x)dρ = ϑρ S (ρ) L(1) f (τ, x)dρ. (3.85)
a a

The equality in (3.85) shows (3.84). ¡In the¢ same manner the following equality
can be proved for λ > 0 and f ∈ D L(1) :
Z ∞ Z ∞
λ e−λρ ϑρ S(ρ)L(1) f dρ = λD1 e−λρ ϑρ S(ρ)f dρ
0 0
3.2 Dissipative operators and maximum principle 119
Z ∞
+ λ2 e−λρ ϑρ S(ρ)f dρ − λf. (3.86)
0
¡ ¢
As above let f ∈ D L(1) . From (3.86) we infer that
µ Z ∞ Z ∞ ¶
L(1) f = Tβ - lim λD1 e−λρ ϑρ S(ρ)f dρ + λ2 e−λρ ϑρ S(ρ)f dρ − λf .
λ→∞ 0 0
(3.87)
¡ ¢
If, in addition, f belongs to the domain of D1 , then it also belongs to D L ,
and
µ Z ∞ ¶
Lf = Tβ - lim λ2 e−λρ ϑρ S(ρ)f dρ − λf
λ→∞
µ Z0 ∞ ¶
= Tβ - lim λ2 e−λρ S(ρ)ϑρ f dρ − λf . (3.88)
λ→∞ 0

The second equality in (3.88) follows from (3.87). So far the result is not
conclusive. To finish the proof of the implication (i) =⇒ (ii) of Theorem 3.10
(1)
we will use the hypothesis that the space CP,b (λ0 ) is Tβ -dense for some λ0 > 0.
In addition, we will use the following identity for a function f in the domain
of the time derivative D1 :
Z ∞
λL(1) e−λρ S(ρ)ϑρ f dρ
0
Z ∞ ³ ´Z ∞
2 −λρ (1)
=λ e S(ρ)ϑρ f dρ − λf + λ λI − L e−λρ S(ρ) (I − ϑρ ) f dρ
0 0
Z ∞ Z ∞
2 −λρ
=λ e S(ρ)ϑρ f dρ − λf + λ e−λρ S(ρ)ϑρ D1 f dρ. (3.89)
0 0

However, this is not the best approach either. The following arguments will
(1)
show that the Tβ -density of CP,b (λ0 ) is dense in C0 ([0, T ] × E) entails that
D(1) (L) = D (L) ∩ D (D¡1 ) is ¢a core for the operator L(1) . From (3.83) it
follows that D(1) (L) ⊂ D L(1) . From (3.87), (3.88), and from (3.89) we also
¡ ¢ ¡ ¢
get D L(1) ∩ D (D1 ) = D L ∩ D (D1 ). Fix λ0 > 0 such that the space
(1) ¡ ¢ (1)
CP,b (λ0 ) is Tβ -dense in Cb ([0, T ] × E). Since R λ0 I − L(1) = CP,b (λ0 ), this
hypothesis has as a consequence that the range of the operator λ0 I − L − D1
(1)
is Tβ -dense in Cb ([0, T ] × E).
¡ The
¢ Tβ -dissipativity of the operator L (1) then
implies that the subspace D L ∩ D (D1 ) is a core for the operator L , and
consequently, the closure of the operator L + D1 coincides with L(1) . We will
show all this. Since the operator L(1) generates a Feller semigroup, the same
is true for the closure of L + D1 . The range of λ0 I − L − D1 coincides with
(1)
the subspace CP,b (λ0 ) defined in (3.36). It is easy to see that
½ Z ∞ ¾
(1) −λ0 ρ
CP,b (λ0 ) = f ∈ Cb ([0, T ] × E) : R (λ0 ) f = e S(ρ)f dρ ∈ D (D1 ) .
0
(3.90)
120 3 Space-time operators
(1) ¡ ¢
If f ∈ CP,b (λ0 ), then f = λ0 I − L(1) R (λ0 ) f where
³ ´ ¡ ¢
R (λ0 ) f ∈ D L(1) ∩ D (D1 ) = D L ∩ D (D1 ) , (3.91)

(1)
as was shown in (3.87) and (3.88). It follows that f ∈ CP,b (λ0 ) can be written
as ¡ ¢
f = λ0 I − L − D1 R (λ0 ) f. (3.92)
By (i) the range of λ0 I − L − D1 is Tβ -dense in Cb ([0, T ] × E). The second
equality in (3.277) follows from (3.91) and (3.92). Let f belong to the Tβ -
(1)
closure of λ0 I − L − D1 . Then there exists a net (gα )α∈A ⊂ CP,b (λ0 ) ⊂
¡ ¢
([0, T ] × E) such that f = limα λ0 I − L − D1 gα . From (3.14) we infer that
g = Tβ - limα gα . Since the operator Tβ -closed linear operator L(1) extends L +
D1 , it follows that L + D1 is Tβ -closable. Let L0 be its Tβ -closure. From (3.14)
it also follows that f = (λ0 I − L0 ) g. Since the range of λ0 I − L¡− D1¢ is Tβ -
dense, we see that R (λ0 I − L0 ) = Cb ([0, T ] × E). Next let g ∈ D L(1) . Then
¡ ¢
there exists g0 ∈ D (L0 ) such that λ0 I − L(1) g = (λ0 I − L0 ) g0 . Since L(1)
extends L0 , and since L(1) is dissipative (see (3.53), it follows that g = g0 ∈
D (L0 ). In other words, the operator L0 coincides with L(1) , and consequently,
the operator L + D1 is Tβ -closable, and its closure coincides with L(1) , the
Tβ -generator of the semigroup {S(ρ) : ρ ≥ 0}. This proves the implication (i)
=⇒ (ii) of Theorem 3.10.
(ii) =⇒ (iii). Let L(2) be the closure of the operator D1 + L. From (ii) we
(1)
know
¡ that ¢ L generates a Tβ -continuous semigroup {S2 (ρ) : ρ ≥ 0}. Since
D L(1) is Tβ -dense, it follows that D(1) (L) = D (D1 ) ∩ D(L) is Tβ -dense as
well. The generator of the Tβ -continuous semigroup {S(ρ) : ρ ≥ 0}, which we
denote by L(1) , extends D1 + L, and hence it also extends L(2) . Since L(2)
generates¡ a¢Feller semigroup, it ¡is dissipative,
¢ and so it satisfies (3.53). Let
g ∈ D L(1) , and choose g0 ∈ D L(2) such that
³ ´ ³ ´ ³ ´
λ0 I − L(1) g = λ0 I − L(1) g0 = λ0 I − L(2) g0 .
¡ ¢ ¡ ¢
The inequality in(3.53) implies that g = g0 ∈ D L(2) , and hence D L(2) =
¡ (1) ¢
D L . Moreover, L(1) extends L(2) . Therefore L(2) = L(1) . It also follows
that the semigroup {S2 (ρ) : ρ ≥ 0} is the same as {S(ρ) : ρ ≥ 0}. Moreover,
there exists λ0 > 0 such that the range of λ0 I − D1 − L is Tβ -dense in
Cb ([0, T ] × E). In fact this is true for all λ, <λ > 0. Finally, we will show
that the operator D1 + L is positive Tβ -dissipative. Let u ∈ H + ([0, T ] × E),
and consider the functionals f 7→ u(τ, x)λR(λ)f (τ, x), λ ≥ λ0 > 0, (τ, x) ∈
[0, T ] × E. Since L(1) generates a Tβ -continuous semigroup we know that
lim ku (f − λR(λ)f )k∞ = 0. (3.93)
λ→∞

If (fm )m∈N ⊂ Cb ([0, T ] × E) decreases pointwise to 0, then the sequence


(u(τ, x)λR(λ)fm (τ, x))m∈N also decreases to. By Dini’s Lemma and (3.93)
3.2 Dissipative operators and maximum principle 121

this convergence is uniform in λ ≥ λ0 and (τ, x) ∈ [0, T ] × E, because u ∈


H + ([0, T ] × E). From Theorem 1.8 it follows that there exists a function
v ∈ H + ([0, T ] × E) such that

kuλR(λ)f k∞ ≤ kvf k∞ , f ∈ Cb ([0, T ] × E) (3.94)

Since the operator L(1) sends real functions to real functions from (3.94), and
u ≥ 0, we derive for (σ, y) ∈ [0, T ] × E

< (u(σ, y)λR(λ)f (σ, y)) = u(σ, y)λR(λ) (<f ) (σ, y)


+
≤ u(σ, y)λR(λ) (<f ) (σ, y)
+
≤ sup v(τ, x) (<f ) (τ, x)
(τ,x)∈[0,T ]×E

≤ sup v(τ, x) (<f ) (τ, x). (3.95)


(τ,x)∈[0,T ]×E
¡ ¢
By the substitution f = λI − L(1) g in (3.95) we obtain:
³ ´
λ sup u(τ, x)<g(τ, x) ≤ sup v(τ, x)< λg(τ, x) − L(1) g(τ, x) .
(τ,x)∈[0,T ]×E (τ,x)∈[0,T ]×E
(3.96)
Since the operator L(1) extends D1 + L, the inequality in (3.96) displays the
fact that the operator D1 + L is positive Tβ -dissipative.
Altogether, this shows the implication (ii) =⇒ (iii) of Theorem 3.10.
(iii) =⇒ (i). Suppose that we already know that the Tβ -closure of D1 + L
generates a Tβ -continuous semigroup {S(ρ) : ρ ≥ 0}. Then we define the evo-
lution {P (τ, t) : 0 ≤ τ ≤ t ≤ T } by

P (τ, t) f (x) = S (t − τ ) [(s, y) 7→ f (y)] (τ, x) , f ∈ Cb (E). (3.97)

We have to prove that the family {P (τ, t) : 0 ≤ τ ≤ t ≤ T } is a Feller evolu-


tion indeed. First we show that it has the evolution property:

P (τ, t1 ) P (t1 , t) f (x) = S (t1 − τ ) [(s, y) 7→ P (t1 , t) f (y)] (τ, x)


= S (t1 − τ ) [(s, y) 7→ S (t − t1 ) f (s, y)] (τ, x)
= S (t1 − τ ) S (t − t1 ) [(s, y) 7→ f (s, y)] (τ, x)
= S (t − τ ) [(s, y) 7→ f (s, y)] (τ, x)
= P (τ, t) f (x). (3.98)

The equality in (3.98) exhibits the evolution property. The continuity of the
function (τ, t, x) 7→ P (τ, t) f (x) follows from the continuity of the function
(τ, t, x) 7→ S (t − τ ) [(s, y) 7→ f (y)] (τ, x): see (3.97).
Next we prove that the operator D1 + L is Tβ -closable, and that its closure
generates a Feller semigroup. Since the operator D1 + L is Tβ -densely defined
and Tβ -dissipative, it is Tβ -closable: see Proposition 3.11 assertion (a). Let
122 3 Space-time operators

L(1) be its Tβ -closure. Since there exists λ0 > 0 such that the range of λ0 I −
D1 − L is Tβ -dense in Cb ([0, T ] × E), and since D1 + L is T|beta -dissipative, it
¡ ¢ ¡ ¢−1
follows that R λ0 I − L(1) = Cb ([0, T ] × E). Put R (λ0 ) = λ0 I − L(1) ,
P∞ n n+1
and R (λ) = n=0 (λ0 − λ) (R (λ0 )) ¡, |λ − λ0 | < λ
¢ 0 . This series converges
in the uniform norm. It follows that R λI − L(1) = Cb ([0, T ] × E) for all
λ ¡∈ C for which
¢ |λ − λ0 | < λ0 . This procedure can be repeated to obtain:
R λI − L(1) = Cb ([0, T ] × E) for all λC with <λ > 0. Put
2
S0 (t)f = Tβ - lim e−λt etλ R(λ)
f, f ∈ Cb ([0, T ] × E) . (3.99)
λ→∞

Of course we have to prove that the limit in (3.99) exists. For brevity
¡ ¢ we write
A(λ) = λ2 R(λ) − λI = L(1) (λR(λ)), and notice that for f ∈ D L(1) we have
A(λ)f = λR(λ)L(1) f , and that
³ ´2 µ³ ´2 ¶
A(λ)f = λR(λ)L(1) f = R(λ) L(1) f + L(1) f, for f ∈ D L(1) .
(3.100)
Let 0 < λ < µ < ∞. From Duhamel’s formula we get

e−λt eλt(λR(λ)) f − e−µt eµt(µR(µ)) f


Z t
tA(λ) tA(µ)
=e f −e f= esA(λ) (A(λ) − A(µ)) e(t−s)A(µ) f ds. (3.101)
0
¡ ¢ ¡ ¢2
If f belongs to D L(1) , then A(λ)f − A(µ)f = (R(λ) − R(µ)) L(1) f , and
hence the equality in (3.101) can be rewritten as:

e−λt eλt(λR(λ)) f − e−µt eµt(µR(µ)) f


Z t ³ ´2
= esA(λ) (R(λ) − R(µ)) e(t−s)A(µ) L(1) f ds. (3.102)
0

From (3.102) we infer that for the uniform norm we have:


° °
° −λt λt(λR(λ)) °
°e e f − e−µt eµt(µR(µ)) f °

Z t° ³ ´2 °
° sA(λ) °
≤ ° (t−s)A(µ) (1)

°e (R(λ) − R(µ)) e L ° ds
0 ∞
Z t° ° ° °° ° ³ ´2 °
°
° sA(λ) ° ° °
≤ °e ° kR(λ) − R(µ)k °e(t−s)A(µ) ° ° ° L (1)

° ds
0 ∞
µ ¶ °³ ´2 °
1 1 ° ° L
°
≤t + (1)

° . (3.103)
λ µ ° ∞
³¡ ¢2 ´
From (3.103) we infer that for f ∈ D L(1) the limit

S0 (t)f (τ, x) = lim etA(λ) f (τ, x) (3.104)


λ→∞
3.2 Dissipative operators and maximum principle 123

exists uniformly in (τ, t, x) ∈ [0, T ] × [0, T ] × E. The ¡ next¢ step consists


¡ in
¢
showing that the limit in (3.104) exists for f ∈ D L(1) . Let f ∈ D L(1) ,
and λ > µ > 0. Then we have for λ0 > 0 sufficiently large:
° °
° tA(λ) °
°e f − etA(µ) f °

° °
° °
≤ °etA(λ) (f − λ0 R (λ0 ) f ) − etA(µ) (f − λ0 R (λ0 ) f )°

° °
° tA(λ) tA(µ) °
+ °e (λ0 R (λ0 ) f ) − e (λ0 R (λ0 ) f )°

³° ° ° °´ ° °
° tA(λ) ° ° tA(µ) ° ° (1) °
≤ °e ° + °e ° °R (λ0 ) L f °

° °
° tA(λ) tA(µ) °
+ °e (λ0 R (λ0 ) f ) − e (λ0 R (λ0 ) f )°

2 °° (1) °
° °
° tA(λ) tA(µ)
°
°
≤ °L f ° + °e (λ0 R (λ0 ) f ) − e (λ0 R (λ0 ) f )° . (3.105)
λ0 ∞ ∞

From (3.105)
¡ ¢ together with (3.104) it follows that (3.104) also holds for
f ∈ D L(1) . There remains to be shown that the limit in (3.104) also
exists in Tβ -sense, but now for f ∈ Cb ([0, T ] × E). Since the operator
L(1) is Tβ -dissipative, there exists, for u ∈ H + ([0, T ] × E), a function
v ∈ H + ([0, T ] × E) such that for all λ ≥ λ0 > 0 the inequality in (3.14)
in Definition 3.5 is satisfied, i.e.
° ³ ´° ³ ´
° °
°v λf − L(1) f ° ≥ λ kuf k∞ , for all λ ≥ λ0 , and for all f ∈ D L(1) .

(3.106)
From (3.106) we infer

λ kuR(λ)f k∞ ≤ kvf k∞ , f ∈ Cb ([0, T ] × E) . (3.107)

Let f ∈ Cb ([0, T ] × E). By Hausdorff-Bernstein-Widder inversion theorem


there exists a unique Borel-measurable function (τ, t, x) 7→ Se0 (t)f (τ, x) such
that
³ ´−1 Z ∞
R(λ)f (τ, x) = λI − L(1) f (τ, x) = e−λρ Se0 (ρ)f (τ, x)dρ, <λ > 0.
0
(3.108)
For the result in (3.108) see Widder [254] Theorem 16a, page 315. The re-
solvent property of the mapping λ 7→ R(λ), λ > 0, implies the semigroup
property of the mapping ρ 7→ S(ρ). To be precise we have:
Z ∞
¡ −λρ ¢
R(λ)f − R(µ)f = e − e−µρ Se0 (ρ)f dρ
0
Z ∞Z ρ
= (µ − λ) e−λ(ρ−s)−µs Se0 (ρ − s + s) f ds dρ
0 0
Z ∞Z ∞
= (µ − λ) e−λ(ρ−s)−µs Se0 (ρ − s + s) f dρ ds
0 s
124 3 Space-time operators
Z ∞ Z ∞
= (µ − λ) e−λρ−µs Se0 (ρ + s) f dρ ds. (3.109)
0 0

On the other hand we also have

R(λ)f − R(µ)f = (µ − λ) R(λ)R(µ)f


Z ∞Z ∞
= (µ − λ) e−λρ−µs Se0 (ρ) Se0 (s)f dρ ds. (3.110)
0 0

Comparing (3.109) and (3.110) shows the equality:

Se0 (ρ + s) f = Se0 (ρ)Se0 (s)f, ρ, s ≥ 0, f ∈ Cb ([0, T ] × E) .


n o
Hence the family Se0 (ρ) : ρ ≥ 0 is a semigroup. We have to show that the
function (τ, t, x) 7→ Se0 (t) f (τ, x) is a bounded continuous function. This will
be done in several steps. First we will prove the following representation for
Se0 (t)f , f ∈ Cb ([0, T ] × E),

X (λt)k
Se0 (t)f = lim e−λt
k
(λR(λ)) f = lim e−λt eλt(λR(λ)) f = S0 (t)f,
λ→∞ k! λ→∞
k=0
(3.111)
provided
¡ that
¢ the limit in (3.111) exists, and where S 0 (t) is as in (3.100). Let
f ∈ D L(1) . Then the function S0 (t)f is the uniform limit of functions of
the form (τ, t, x) 7→ etA(λ) f (τ, x), and such functions are continuous in the
variables (τ, t, x): see (3.105). Consequently,
¡ ¢the function S0 (t)f inherits this
continuity property. Again let f ∈ D L(1) . We will prove that R(µ)f =
R ∞ −µt
0
e S0 (t)f dt, µ > 0. Therefore we notice
Z ∞ Z ∞
e−µt S0 (t)f dt = e−µt lim etA(λ) f dt
0 0 λ→∞
Z ∞
−1
= lim e−µt etA(λ) f dt = lim (µI − A(λ)) f
λ→∞ 0 λ→∞
µ ¶µ ¶−1
λ 1 λµ
= lim I− L(1) I − L(1) f
λ→∞ λ + µ λ+µ λ+µ
³ ´−1
= µI − L(1) f. (3.112)

From (3.108) and (3.112) we infer the equality


³ ´
Se0 (t)f = S0 (t)f for f ∈ D L(1) . (3.113)

After that we will prove that the averages of the semigroup {S0 (ρ) : ρ ≥ 0}
is Tβ -continuous. As a consequence, for f ∈ Cb ([0, T ] × E) the function
Z
1 t −λρ
(τ, t, x) 7→ e S0 (ρ)f (τ, x) dρ,
t 0
3.2 Dissipative operators and maximum principle 125

is a bounded and continuous function, and the family of operators


½ Z t ¾
1 −λρ
e S0 (ρ)dρ : 0 ≤ t ≤ T is Tβ -equi-continuous. (3.114)
t 0

As above we write A(λ)f = λ2 R(λ)f − λf . Two very relevant equalities are:


³ ´−1 Z t ³ ´−1
R(λ)f = λI − L(1) f= e−λρ Se0 (ρ)f dρ + e−λt S0 (t) λI − L(1) f
0
Z t ³ ´−1
= e−λρ S0 (ρ)dρ f + e−λt S0 (t) λI − L(1) f, (3.115)
0

and
³ ´Z t
f = λI − L (1)
e−λρ Se0 (ρ)f dρ + e−λt Se0 (t)f. (3.116)
0

Here we wrote Z Z
t t
e−λρ S0 (ρ)dρ f = e−λρ Se0 (ρ)f dρ
0 0
Rt
to indicate that the operator f 7→ 0 e−λρ S0 (ρ)dρ f , f ∈ Cb ([0, T ] × E), is a
mapping from Cb ([0, T ] × E) to itself, whereas it is not is not so clear what
the target space is of the mappings Se0 (ρ), ρ > 0. In order to show that
the operators Se0 (t), t ≥ 0, are mappings from Cb ([0, T ] × E) into itself, we
need the sequential λ-dominance of the operator D1 + L for some λ > 0.
Moreover,
n it follows
o from this sequential λ-dominance that the semigroup
−λt e
e S0 (t) : t ≥ 0 is Tβ -equi-continuous. Once we know all this, then the
formula in (3.116) makes sense and is true.
For every measure ν on the Borel field of [0, T ] × E the mapping ρ 7→
R
e
S0 (ρ)f dν is a Borel measurable function on the the semi-axis [0, ∞). The for-
mula in (3.115) is correct, and poses no problem provided f ∈ Cb ([0, T ] × E).
In fact we have
Z ∞
1
e−µt e−λt S0 (t)R(λ)f dt = R(λ + µ)R(λ)f = (R(λ) − R(λ + µ)) f
0 µ
Z ∞ Z ∞Z ρ
1 − e−µρ −λρ e
= e S0 (ρ)f dρ = e−µt dt e−λρ Se0 (ρ)f dρ
µ
Z0 ∞ Z ∞ 0 0

−µt −λρ e
= e e S0 (ρ)f dρ dt, (3.117)
0 t

and hence
³ ´−1 Z ∞
e−λt S0 (t) λI − L(1) f = e−λt S0 (t)R(λ)f = e−λρ Se0 (ρ)f dρ. (3.118)
t

From (3.118) we infer


126 3 Space-time operators
Z t ³ ´−1
e−λρ Se0 (ρ)f dρ + e−λt S0 (t) λI − L(1) f
0
Z t Z ∞
= e−λρ Se0 (ρ)f dρ + e−λρ Se0 (ρ)f dρ
Z0 ∞ t
³ ´−1
−λρ e
= e S0 (ρ)f dρ = λI − L(1) f. (3.119)
0

The equality in (3.115) is the same as the one in (3.119). From the equality
Rt
in (3.115) it follows that the function (τ, t, x) 7→ 0 e−λρ Se0 (ρ)f (τ, x)dρ is
¡ ¢ ¡ ¢
continuous. Next let g ∈ D L(1) and put f = λI − L(1) g. From (3.115)
we get: Z t
g − e−λt S0 (t)g = e−λρ S0 (ρ)dρ f. (3.120)
0

From (3.120) we infer:


°Z t °
° ° ° °
°g − e−λt S0 (t)g ° = ° e −λρ
S0 (ρ)dρ f °
∞ ° °
0 ∞
Z t ° ° Z t ° °
−λρ ° e ° −λρ ° °
≤ e °S0 (ρ)f ° dρ ≤ e dρ °λf − L(1) f ° , (3.121)
0 ∞ 0 ∞
° °
and hence for λ = 0 we obtain kg − S0 (t)gk∞ ≤ t °L(1) f °∞ . This inequality
½ ¾
1
proves the uniform boundedness of the family (g − S0 (t)g) : t > 0 . Next
t
let us discuss its convergence. Therefore we again employ (3.115), and pro-
ceed as follows. Let (fn )n∈N be a sequence in Cb ([0, T ] × E) which decreases
¡ ¢ ¡ ¢
pointwise to the zero-function. Choose the sequence gnλ n∈N ⊂ D L(1) in
¡ ¢
such a way that λfn = λgnλ − L(1) gnλ . Then the sequence gnλ n∈N decreases to
zero as well. For t > 0 have
Z t
gnλ = e−λt S0 (t)gnλ + λ e−λρ S0 (ρ)dρ fn , (3.122)
0

or, what is equivalent,


Z t
eλt gnλ = S0 (t)gnλ +λ eλt−λρ S0 (ρ)dρ fn . (3.123)
0

Since the operator L(1) is Tβ -dissipative, it follows that sup gnλ decreases
λ, λ≥T −1
pointwise to zero. So that, with λ = t−1 , the equality in (3.122) implies
Z
1 t
sup S0 (ρ)dρ fn ↓ 0, as n → ∞. (3.124)
t, 0<t≤T t 0

Consequently, for any fixed λ ∈ R, the family of operators


3.2 Dissipative operators and maximum principle 127
½ Z t ¾
1
e−λρ S0 (ρ)dρ : T ≥ t > 0 (3.125)
t 0

is Tβ -equi-continuous: see Corollary 1.19. Let f ∈ Cb ([0, T ] × E). We will


show that Z
1 t −λρ
Tβ - lim e S0 (ρ)dρ f = f. (3.126)
t↓0 t 0

It suffices to prove (3.126) for λ > 0. First assume that f = R(λ) belongs to
the domain of L(1) . Then we have
Z Z Z ∞
1 t −λρ 1 t −λρ
e S0 (ρ)dρ f = e S0 (ρ) e−λσ S0 (σ) dσ dρg
t 0 t 0 0
Z Z
1 t ∞ −λ(σ+ρ)
= e S0 (σ + ρ) dσ g dρ
t 0 0
Z Z
1 t ∞ −λσ
= e S0 (σ) dσ g dρ. (3.127)
t 0 ρ
R∞
Since the function ρ 7→ ρ e−λσ S0 (σ) dσ g is continuous for the uniform norm
topology on Cb ([0, T ] × E), (3.127) implies
Z Z Z ∞ ³ ´
1 t ∞ −λσ
k·k∞ - lim e S0 (σ)dσ f = e−λσ S0 (σ)dσ g = f, f ∈ D L(1) .
t↓0 t 0 ρ 0
¡ ¢ (3.128)
Since D L(1) is Tβ -dense in Cb ([0, T ] × E), the equi-continuity of the family
in (3.125) implies that
Z
1 t −λρ
Tβ - lim e S0 (ρ)dρ f = f, f ∈ Cb ([0, T ] × E) . (3.129)
t↓0 t 0

From the equality in (3.120) together with (3.129) we see that

g − e−λt S0 (t)g ³ ´
Tβ - lim = f = λg − L(1) g, g ∈ D L(1) . (3.130)
t↓0 t
n o
So far we have proved that the semigroup Se0 (t) : t ≥ 0 maps the domain of
L(1) to bounded continuous functions, and that the family in (3.129) consists
of mappings which assign to bounded continuous again bounded bounded
continuous functions. What is not clear, is whether or not the operators Se0 (t),
t ≥ 0, leave the space Cb ([0, T ] × E) invariant. Fix λ > 0, and to every
f ∈ Cb ([0, T ] × E), f ≥ 0, we assign the function f λ defined by
n o
k
f λ = sup (µR (λ + µ)) f : µ > 0, k ∈ N , (3.131)

The reader is invited to compare the function f λ with (1.48) and other results
in Proposition 1.22. The arguments which follow are in line with the proof
128 3 Space-time operators

of Proposition 1.22. The function f λ is the smallest λ-super-median valued


function which exceeds f . A closely related notion is the notion of λ-super-
mean valued function. A function g : [0, T ] × E → [0, ∞) is called λ-super-
median valued if e−λt Se0 (t)g ≤ g for all t ≥ 0; it is called λ-super-mean valued
if µR (λ + µ) g ≤ g for all µ > 0. In Lemma (9.12) in Sharpe [208] it is shown
that, essentially speaking, these notions are equivalent. In fact the proof is
not very difficult. It uses the Hausdorff-Bernstein-Widder theorem about the
representation by Laplace transforms of positive Borel measures on [0, ∞) of
completely positive functions. The reader is also referred to Remark 3.7 and
Definition 3.6.
Let f ∈ Cb ([0, T ] × E) be positive. Here we use the representation

X (µt)k
e−λt Se0 (t)f = lim e−µt
k
(µR (λ + µ)) f ≤ f λ , (3.132)
µ→∞ k!
k=0

and hence
sup e−λt Se0 (t)f ≤ f λ . (3.133)
t>0

Since
Z ∞
µk
tk−1 e−µt e−λt Se0 (t)f dt
k
(µR (λ + µ)) f = (3.134)
(k − 1)! 0

we see by invoking (3.131) that the two expressions in (3.133) are the same.
In order to finish the proof of Theorem 3.10 we need the hypothesis that the
operator D1 + L is sequentially λ-dominant for some λ > 0. In fact, let the
sequence (fn )n∈N ⊂ Cb ([0, T ] × E) converge downward to zero, and select
functions gnλ ∈ D (D1 + L), n ∈ N, with the following properties:
1. fn ≤ gnλ ;
2. gnλ = sup inf {g ≥ fn 1K : (λI − D1 − L) g ≥ 0};
K∈K([0,T ]×E) g∈D(D1 +L)
3. lim gnλ (τ, x) =0 for all (τ, x) ∈ [0, T ] × E.
n→∞

In the terminology of (1.41) and Definition 3.6 the functions gnλ are denoted
by gnλ = Uλ1 (fn ), n ∈ N. Recall that K ([0, T ] × E) denotes the collection
of all compact subsets of [0, T ] × E. By hypothesis, the sequence as defined
in 2 satisfies 1 and 3. Let K be any compact subset of [0, T ] × E, and g ∈
D (D1 + L) be such that g ≥ fn 1K and (λI − D1 − L) g ≥ 0. Then we have
³ ´
(λ + µ) I − L(1) g = ((λ + µ) I − D1 − L) g ≥ µg. (3.135)

From (3.135) and g ≥ fn 1K we infer


k k
g ≥ µR (λ + µ) g ≥ (µR (λ + µ)) g ≥ (µR (λ + µ)) (fn 1K ) , (3.136)
and hence (3.136) together with (3.131) and (3.133) (which is in fact an equal-
ity) we see
3.2 Dissipative operators and maximum principle 129
n o
gnλ ≥ fnλ = sup sup e−λt Se0 (t) (fn 1K ) : t ≥ 0
K∈K([0,T ]×E)
n o
k
= sup sup (µR (λ + µ)) (fn 1K ) : µ > 0, k ∈ N
K∈K([0,T ]×E)
n o
k
= sup (µR (λ + µ)) fn : µ > 0, k ∈ N
n o
= sup e−λt Se0 (t)fn : t ≥ 0 . (3.137)

Since by hypothesis lim gnλ = 0 the inequality in (3.137) implies: lim fnλ = 0.
n→∞ n→∞
It follows that n o
lim sup e−λt Se0 (t)fn : t ≥ 0 = 0. (3.138)
n→∞

From Corollary 1.19 it follows that the family of operators


n o
k
(µR (λ + µ)) : µ ≥ 0, k ∈ N

is Tβ -equi-continuous. Hence for every function u ∈ H + ([0, T ] × E) there


exists a function v ∈ H + ([0, T ] × E) such that
° °
° k °
°u (µR (λ + µ)) f ° ≤ kvf k∞ , f ∈ Cb ([0, T ] × E) , µ ≥ 0, k ∈ N.

(3.139)
Since
n o n o
sup e−λt Se0 (t)f : t ≥ 0 = sup (µR (λ + µ)) f : µ ≥ 0, k ∈ N , f ≥ 0,
k

(3.140)
the inequality in (3.139) yields
° °
° −λt e °
°ue S0 (t)f ° ≤ kvf k∞ , f ∈ Cb ([0, T ] × E) , t ≥ 0. (3.141)

¡ ¢
Since D L(1) is Tβ -dense, and the operators Se0 (t), t ≥ 0, are mappings from
¡ (1) ¢
D L to Cb ([0, T ] × E) the Tβ -equi-continuity in (3.141) shows that the
operators Se0 (t), nt ≥ 0, are in fact mappings
o from Cb ([0, T ] × E) to itself, and
that the family e−λt Se0 (t) : t ≥ 0 is Tβ -equi-continuous.
However, all these observations conclude the proof of the implication (iii)
=⇒ (i) of Theorem 3.10.

Remark 3.12. The equality in (3.115) shows that the function g := R(λ)f ,
where f ≥ 0 and f ∈ Cb ([0, T ] × E) is λ-super-mean valued in the sense that
an inequality of the form e−λt S0 (t)g ≤ g holds. Such an inequality is equiv-
alent to µR (µ + λ) g ≤ g. For details on such functions and on λ-excessive
functions see Sharpe [208], page 17 and Lemma 9.12, page 45.
130 3 Space-time operators

3.3 Korovkin property

The following notions and results are being used to prove part (e) of Theorem
1.39. We recall the definition of Korovkin property.
Definition 3.13. Let K be a subset of E The operator L is said to possess
the Korovkin property if there exists a strictly positive real number λ0 > 0
such that for every x0 ∈ E0 the equality
½ · µ ¶ ¸ ¾
1
inf sup h(x0 ) + g − I − L h (x) (3.142)
h∈D(L) x∈E0 λ0
½ · µ ¶ ¸ ¾
1
= sup inf h(x0 ) + g − I − L h (x) (3.143)
h∈D(L) x∈E 0 λ 0

is valid for all g ∈ Cb (E).


Let g ∈ Cb (E) and λ > 0. The equalities
½ µ ¶ ¾
1
inf h(x0 ) : I − L h ≥ g on E0
h∈D(L) λ
µ · µ ¶ ¸ ¶
1
= inf sup h(x0 ) + g − I − L h (x)
h∈D(L) x∈E0 λ
½ · µ ¶ ¸ ¾
1
= inf sup min max h(x0 ) + g − I − L h (x) , (3.144)
Γ ⊂D(L) Φ⊂E0 h∈Γ x∈Φ λ
#Γ <∞ #Φ<∞

show that the Korovkin property could also have been defined in terms of
any of the quantities in (3.144). In fact, if L satisfies the (global) maximum
principle on E0 , i.e. if for every real-valued function f ∈ D(L) the inequality

λ sup f (x) ≤ sup (λf (x) − Lf (x)) (3.145)


x∈E0 x∈E0

holds for all λ > 0, then the Korovkin property (on E0 ) does not depend on
λ0 > 0. In other words, if it holds for one λ0 > 0, then it is true for all λ > 0.
This is part of the contents of the following proposition. In fact the maximum
principle as formulated in (3.145) is not adequate in the present context. The
correct version here is the following one, which is kind of a σ-local maximum
principle.
Definition 3.14. Let E0 be a subset of E. Suppose that the operator L has the
property that for every λ > 0 and for every x0 ∈ E0 it is true that h (x0 ) ≥ 0,
whenever h ∈ D(L) is such that (λI − L) h ≥ 0 on E0 . Then the operator L
is said to satisfy the weak maximum principle on E0 .
As we proved in Proposition 1.35 the notion weak maximum principle and
maximum principle coincide, provided 1 ∈ D(L) and l1 = 0.
3.3 Korovkin property 131

In order to be really useful, the Korovkin property on E0 should be accom-


panied by the maximum principle on E0 . To be useful the global Korovkin
property (see Definition 3.15) requires the global maximum principle (see
(3.145)). In addition we need the fact that the constant functions belong to
D(L) and that L1 = 0. If we only know the global maximum principle, in the
sense of (3.145), then the global Korovkin property is required:
Definition 3.15. The operator L is said to possess the global Korovkin prop-
erty if there exists a strictly positive real number λ0 > 0 such that for every
x0 ∈ E the equality
½ · µ ¶ ¸ ¾
1
inf sup h(x0 ) + g − I − L h (x) (3.146)
h∈D(L) x∈E λ0
½ · µ ¶ ¸ ¾
1
= sup inf h(x0 ) + g − I − L h (x) (3.147)
h∈D(L) x∈E λ 0

is valid for all g ∈ Cb (E).


First we treat the situation of a subset of E. The global version is obtained
from the one on E0 by replacing the subset E0 with the full state space E.
Again a resolvent family is obtained. In order to prove the equalities of (3.166)
through (3.175) the global maximum principle is used. In fact it is used to
show the equalities
½ · µ ¶ ¸ ¾
1
inf sup h(x0 ) + g − I − L h (x) (3.148)
h ∈ D(L) x ∈ E λ
½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ g on E (3.149)
h∈D(L) λ
½ µ ¶ ¾
1
= sup h (x0 ) : I − L h ≤ g on E (3.150)
h∈D(L) λ
½ · µ ¶ ¸ ¾
1
= sup inf h(x0 ) + g − I − L h (x) . (3.151)
h∈D(L) x∈E λ

In particular, if g = 0, and if L satisfies the global maximum principle, then


the expressions in (3.148) through (3.151) are all equal to 0. Put

λ0 R (λ0 ) g (x0 )
½ · µ ¶ ¸ ¾
1
= inf sup h(x0 ) + g − I − L h (x)
h∈D(L) x∈E0 λ0
½ · µ ¶ ¸ ¾
1
= sup inf h(x0 ) + g − I − L h (x) . (3.152)
h∈D(L) x∈E0 λ0

Then λ0 R (λ0 ) is a linear operator from Cb (E0 ) to Cb (E0 ). The following


proposition shows that there exists a family of operators {R(λ) : 0 < λ < 2λ0 }
132 3 Space-time operators

which has the resolvent property. The operator λR(λ) is obtained from (3.152)
by replacing λ0 with λ. It is clear that this procedure can be extended to the
whole positive real axis. In this way obtain a resolvent family {R(λ) : λ > 0}.
−1
The operator R(λ) can be written in the form R(λ) = (λI − L0 ) , where L0 is
a closed linear operator which extends L (in case E0 = E), and which satisfies
the maximum principle on E0 , and, under certain conditions, generates a Feller
semigroup and a Markov process. For convenience we insert the following
lemma. It is used for E0 = E and for E0 a subset of E which is polish with
respect to the relative metric. The condition in (3.155) is closely related to
the maximum principle.
Lemma 3.16. Suppose that the constant functions belong to D(L), and that
L1 = 0. Fix x0 ∈ E, λ > 0, and g ∈ Cb (E0 ). Let E0 be any subset of E. Then
the following equalities hold:
½ µ ¶ ¾
1
inf sup h (x0 ) + g(x) − I − L h(x)
h∈D(L) x∈E0 λ
½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ g on E0 , (3.153)
h∈D(L) λ

and
½ µ ¶ ¾
1
sup inf h (x0 ) + g(x) − I − L h(x)
h∈D(L) x∈E0 λ
½ µ ¶ ¾
1
= sup h (x0 ) : I − L h ≤ g on E0 . (3.154)
h∈D(L) λ

½ µ ¶ ¾
1
If inf h (x0 ) :I − L h ≥ 0 on E0 ≥ 0, then (3.155)
h∈D(L) λ
½ µ ¶ ¾
1
sup g(x) ≥ inf h (x0 ) : I − L h ≥ g on E0 ≥ inf g(x),
x∈E0 h∈D(L) λ x∈E0
(3.156)
and also
½ µ ¶ ¾
1
inf h (x0 ) : I − L h ≥ g on E0
h∈D(L) λ
½ µ ¶ ¾
1
≥ sup h (x0 ) : I − L h ≤ g on E0 . (3.157)
h∈D(L) λ

First notice that by taking h = 0 in the left-hand side of (3.153) we see that
the quantity in (3.153) is less than or equal to sup g(x), and that the quantity
x∈E0
in (3.154) is greater than or equal to inf g(x). However, it is not excluded
x∈E0
that (3.153) is equal to −∞, and that (3.154) is equal to ∞.
3.3 Korovkin property 133

Proof. Upon replacing g with −g we see that the equality in (3.154) is a


consequence of (3.153). We put
½ µ ¶ ¾
1
αE0 = inf h (x0 ) : I − L h ≥ g on E0 and
h∈D(L) λ
½ µ ¶ ¾
1
βE0 = inf sup h (x0 ) + g(x) − I − L h(x) . (3.158)
h∈D(L) x∈E0 λ

First assume that βE0 ∈ R. Let ε > 0. Choose hε ∈ D(L) in such a way that
for x ∈ E0 we have
µ ¶
1
hε (x0 ) + g(x) − I − L hε (x) ≤ βE0 + ε.
λ

Then
µ ¶
1
g(x) ≤ I − L hε (x) + βE0 + ε − hε (x0 )
λ
µ ¶
1
= I − L (hε − hε (x0 ) + βE0 + ε) (x). (3.159)
λ

The substitution e hε = hε − hε (x0 ) + βE0 + ε in (3.159) yields αE0 ≤ e


hε (x0 ) =
βE0 + ε. Since ε > 0 was arbitrary, we get αE0 ≤ βE0 . The same argument
with −n instead of βE0 + ε shows αE0 = −∞ if βE0 = −∞. Next we assume
that αE0 ∈ R. Again let ε > 0 be µ arbitrary.
¶ Choose a function hε ∈ D(L)
1
such that hε (x0 ) ≤ αE0 + ε, and I − L hε ≥ g on E0 . Then we have, for
λ
x ∈ E0 ,
µ ¶
1
hε (x0 ) + g(x) − I − L hε (x) ≤ hε (x0 ) ≤ βE0 + ε,
λ

and hence βE0 ≤ αE0 + ε. Since ε > 0 was arbitrary, we get βE0 ≤ αE0 . Again,
the argument can be adapted if αE0 = −∞: replace αE0 + ε by −n, and let n
tend to ∞. If condition (3.155) is satisfied, then with m = inf g(y) we have
y∈E0
½ µ ¶ ¾
1
αE0 ≥ inf h (x0 ) : I − L h ≥ inf g(y) on E0
h∈D(L) λ y∈E0
½ µ ¶ ¾
1
= inf h (x0 ) : I − L (h − m) ≥ 0 on E0 ≥ m. (3.160)
h∈D(L) λ

The inequality in (3.160) shows the lower estimate in (3.156). The upper
estimate is obtained by taking h = sup g(y). Next we prove the inequality
y∈E0
in (3.157). Therefore we observe that the functional Λ+
E0 : Cb (E, R) → R,
defined by
134 3 Space-time operators
½ µ ¶ ¾
1
Λ+
E0 (g) = inf h (x0 ) : I − L h ≥ g on E0 (3.161)
h∈D(L) λ
is sub-additive and positive homogeneous. The latter means that

Λ+ + + + +
E0 (g1 + g2 ) ≤ ΛE0 (g1 ) + ΛE0 (g2 ) , and ΛE0 (αg) = αΛE0 (g)

for g1 , g2 , g ∈ Cb (E, R), and α ≥ 0. Moreover,


½ µ ¶ ¾
1
−Λ+ E0 (−g) = sup h (x0 ) : I − L h ≤ g on E 0 . (3.162)
h∈D(L) λ

It follows that

Λ+ + +
E0 (g) + ΛE0 (−g) ≥ ΛE0 (0)
½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ 0 on E0 ≥ 0. (3.163)
h∈D(L) λ
The inequality in (3.157) is a consequence of (3.162) and (3.163).
This completes the proof of Lemma 3.16.
The definition of an operator L satisfying the maximum principle on a subset
E0 can be found in Definition 3.14
Proposition 3.17. Let 0 < λ < 2λ0 and g ∈ Cb (E) and E0 a subset of E.
Suppose the operator L satisfies the maximum principle on E0 . In addition,
let the domain of L contain the constant functions, and assume L1 = 0. Let
x0 ∈ E0 . Put

λR(λ)g(x0 )
= lim inf inf sup inf sup · · · inf sup
n→∞ h0 ∈D(L) x1 ∈E0 h1 ∈D(L) x2 ∈E0 hn ∈D(L) xn+1 ∈E0
n µ
X ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) (3.164)
j=0
λ0 λ0 λ0
= lim inf sup inf sup inf · · ·
n→∞ h ∈D(L) x1 ∈E0 h ∈D(L) x2 ∈E0
0 1

sup inf
hn ∈D(L) xn+1 ∈E0
n µ
X ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) . (3.165)
j=0
λ0 λ0 λ0

Then the following identities are true:

λR(λ)g(x0 )
 
Xn µ ¶j
λ λ j+1
= lim  1− (λ0 R (λ0 )) g (x0 ) (3.166)
n→∞ λ0 λ 0
j=0
3.3 Korovkin property 135
∞ µ ¶j
λ X λ j+1
= 1− (λ0 R (λ0 )) g (x0 ) (3.167)
λ0 j=0 λ0
= lim inf
n→∞ ∞ ³
P ´j−1
λ λ
hj ∈ D(L), (λI − L) h0 = λ0 1− λ0 (λI − L) hj
j≥0 j=1

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
max 1− hj (xj ) + g(xj+1 ) − I − L hj (xj+1 )
xj ∈E0
j=0
λ0 λ0 λ0
1≤j≤n+1
(3.168)
½ · µ ¶ ¸ ¾
1
= inf max h(x0 ) + g − I − L h (x) (3.169)
h ∈ D(L) x ∈ E0 λ
½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ g on E0 (3.170)
h∈D(L) λ
½ µ ¶ ¾
1
= sup h (x0 ) : I − L h ≤ g on E0 (3.171)
h∈D(L) λ
½ · µ ¶ ¸ ¾
1
= sup min h(x0 ) + g − I − L h (x) (3.172)
h∈D(L) x∈E0 λ
= lim sup
n→∞
P∞ ³ ´j−1
λ λ
hj ∈ D(L), (λI − L) h0 = λ0 1− λ0 (λI − L) hj
j≥0 j=1

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
min 1− hj (xj ) + g(xj+1 ) − I − L hj (xj+1 )
xj ∈E0
j=0
λ0 λ0 λ0
1≤j≤n+1
(3.173)
= lim inf max
n→∞ hj ∈D(L), 0≤j≤n xj ∈E0 , 1≤j≤n+1

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) . (3.174)
j=0
λ0 λ0 λ0
= lim sup min
n→∞ h ∈D(L), 0≤j≤n xj ∈E0 , 1≤j≤n+1
j

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) . (3.175)
j=0
λ0 λ0 λ0

Suppose that the operator possesses the global Korovkin property, and satisfies
the maximum principle, as described in (3.145). Put

λR(λ)g (x0 )
= lim inf inf sup inf sup · · · inf sup
n→∞ h0 ∈D(L) x1 ∈E h1 ∈D(L) x2 ∈E hn ∈D(L) xn+1 ∈E
136 3 Space-time operators
n µ
X ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) (3.176)
j=0
λ0 λ0 λ0
= lim inf sup inf sup inf · · · sup inf
n→∞ h ∈D(L) x1 ∈E h ∈D(L) x2 ∈E hn ∈D(L) xn+1 ∈E
0 1
n µ
X ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) . (3.177)
j=0
λ0 λ0 λ0

Then the quantities in (3.166) through (3.175) are all equal to λR(λ)g (x0 ),
provided that the set E0 is replaced by E.
In case we deal with the (σ-local) Korovkin property, the convergence of
∞ µ ¶j−1
λ X λ
(λI − L) h0 = 1− (λI − L) hj (3.178)
λ0 j=1 λ0

in (3.168) and (3.173) should be uniform onE0 . In case we deal with the
global Korovkin property, and the maximum principle in (3.145), then the
convergence in (3.178) should be uniform on E.
Corollary 3.18. Suppose that the operator L possesses the Korovkin property
on E0 . Then for all λ > 0 the quantities in (3.169), (3.170), (3.171), and
(3.172) are equal for all x0 ∈ E0 and all functions g ∈ Cb (E0 ). If L possesses
the global Korovkin property, then
½ · µ ¶ ¸ ¾
1
inf max h(x0 ) + g − I − L h (x) (3.179)
h ∈ D(L) x ∈ E λ
½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ g on E (3.180)
h∈D(L) λ
½ µ ¶ ¾
1
= sup h (x0 ) : I − L h ≤ g on E (3.181)
h∈D(L) λ
½ · µ ¶ ¸ ¾
1
= sup min h(x0 ) + g − I − L h (x) . (3.182)
h∈D(L) x∈E λ

Moreover, for λ > 0 and f ∈ D(L), the equality R(λ) (λI − L) f = f holds.

Proof. By repeating the result in Proposition 3.17 for all λ1 ∈ (0, 2λ0 ) instead
of λ0 we get these equalities for λ in the interval (0, 4λ0 ). This procedure can
be repeated once more. Induction then yields the desired result. That for
λ > 0 and f ∈ D(L), the equality R(λ) (λI − L) f = f holds can be seen by
the following arguments. By definition we have

λR(λ) (λI − L) f (x0 )


= inf {h (x0 ) : (λI − L) f ≥ h on E0 , h ∈ D(L)} ≤ f (x0 ) . (3.183)
3.3 Korovkin property 137

We also have

λR(λ) (λI − L) f (x0 )


= sup {h (x0 ) : (λI − L) f ≤ h on E0 , h ∈ D(L)} ≥ f (x0 ) . (3.184)

The stated equality is a consequence from (2.157) and (3.183).

Proof (Proof of Proposition 3.17). The equality of each term in (3.164) and
(3.165) follows from the Korovkin property on E0 as exhibited in the formulas
(3.142) and (3.143) of Definition 3.13, provided that the limit in (3.164) exists.
The existence of this limit, and its identification are given in (3.166) and
(3.167) respectively. For this to make sense we must be sure that the partial
sums of the first n + 1 terms of the quantities in (3.164) and (3.166) are equal.
In fact a rewriting of the quantity in (3.164) before taking the limit shows
that the quantity in (3.174) is also equal to (3.164); i.e.
 
X n 
inf sup inf sup · · · inf sup ···
h0 ∈D(L) x1 ∈E0 h1 ∈D(L) x2 ∈E0 hn ∈D(L) xn+1 ∈E0  
j=0
 
X n 
= inf max ··· .
hj ∈D(L), 0≤j≤n xj ∈E0 , 1≤j≤n+1  
j=0

In fact the same is true for the corresponding partial sums in (3.165) and
(3.175), but with inf instead of sup, and min instead of max. For 0 < λ < 2λ0 ,
we have |λ0 − λ| < λ0 . Since
¯ ¯
¯ λ0 − λ ¯
|λ0 − λ| kR (λ0 ) f k∞ ≤ ¯¯ ¯ kf k , f ∈ Cb (E, R) , (3.185)
λ0 ¯ ∞

the sum in (3.167) converges uniformly. The equality of the sum of the first
n + 1 terms in (3.164) and (3.166) can be proved as follows. For 1 ≤ k ≤ n we
may employ the following identities:

inf sup · · · inf sup


h0 ∈D(L) x1 ∈E0 hn ∈D(L) xn+1 ∈E0

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 )
j=0
λ0 λ0 λ0
= inf sup · · · inf sup
h0 ∈D(L) x1 ∈E0 hn−k ∈D(L) xn−k+1 ∈E0


n−k
λ
¶j ½
λ
µ
1
¶ ¾
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 )
j=0
λ0 λ0 λ0
n
X µ ¶j
λ λ j−(n−k)
+ 1− (λ0 R (λ0 )) g (xn−k+1 ) . (3.186)
λ0 λ0
j=n−k+1
138 3 Space-time operators

The equality in (3.186) can be proved by induction with respect to k, and


by repeatedly employing the definition of λ0 R (λ0 ) f , f ∈ Cb (E, R), together
with its linearity. Using (3.186) with k = n we get

inf sup · · · inf sup


h0 ∈D(L) x1 ∈E0 hn ∈D(L) xn+1 ∈E0

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 )
j=0
λ0 λ0 λ0
"½ µ ¶ ¾
λ 1
= inf sup h0 (x0 ) + g (x1 ) − I − L h0 (x1 )
h0 ∈D(L) x1 ∈E0 λ0 λ0

n µ ¶j
λ X λ j
+ 1− (λ0 R (λ0 )) g (x1 )
λ0 j=1 λ0
n µ ¶j
λ X λ j+1
= 1− (λ0 R (λ0 )) g (x0 ) . (3.187)
λ0 j=0 λ0

From the equality of (3.164) and (3.165), together with (3.187) we infer
n µ ¶j
λ X λ j+1
λR(λ)g (x0 ) = lim 1− (λ0 R (λ0 )) g (x0 ) . (3.188)
n→∞ λ0 λ 0
j=0

Notice that by (3.185) the series in (3.188) converges uniformly. Consequently,


the equalities of the quantities in (3.164), (3.165), (3.166), (3.167), (3.174),
and (3.175) follow, and all these expressions are equal to λR(λ)g (x0 ). Next
let (hj )j∈N ⊂ D(L) be any sequence with the following property:
∞ µ ¶j−1
λ X λ
(λI − L) h0 = 1− (λI − L) hj (3.189)
λ0 j=1 λ0

where the series in (3.189) converges uniformly. Then by the maximum princi-
∞ µ ¶j−1
λ X λ
ple the series 1− hj converges uniformly as well. So it makes
λ0 j=1 λ0
to write:
n+1 µ ¶j−1
λ X λ
h0 = 1− h0j where h0j = hj , 1 ≤ j ≤ n, and
λ0 j=1 λ0
X∞ µ ¶j−n−1
λ
h0n+1 = 1− hj . (3.190)
j=n+1
λ0

Again the series in (3.190) converges uniformly. From the representation of h0


in (3.190) we infer the equalities:
3.3 Korovkin property 139

inf max
hj ∈D(L), 0≤j≤n xj ∈E0 , 1≤j≤n+1
n µ
X ¶j ½ µ ¶ ¾
λ λ 1
1− hj (xj ) + g (xj+1 ) − I − L hj (xj+1 ) (3.191)
j=0
λ0 λ0 λ0
= inf
∞ ³
P ´j−1
λ λ
hj ∈ D(L), (λI − L) h0 = λ0 1− λ0 (λI − L) hj
j≥0 j=1

Xn µ ¶j ½ µ ¶ ¾
λ λ 1
max 1− hj (xj ) + g(xj+1 ) − I − L hj (xj+1 ) .
xj ∈E0
j=0
λ0 λ0 λ0
1≤j≤n+1
(3.192)

Hence, the equality of (3.174) and (3.168) follows. A similar argument shows
the equality of (3.175) and (3.173). Of course, here we used the equality of
(3.191) and (3.192) with inf instead of sup, and max replaced with min and
vice versa. So we have equality of the following expressions: (3.164), (3.165),
(3.166), (3.167), (3.168), (3.173), (3.174), (3.175). The proof of the fact that
these quantities are also equal to (3.169), (3.170), (3.171), and (3.172) is still
missing. Therefore we first show that the expression in (3.168) is greater than
or equal to (3.169). In a similar manner it is shown that the expression in
(3.173) is less than or equal to (3.172): in fact by applying the inequality
(3.168) ≥ (3.169) to −g instead of +g we obtain that (3.173) is less than or
equal to (3.172). From the (local) maximum principle it will follow that the
expression in (3.169) is greater than or equal to (3.172). As a consequence we
will obtain that, with the exception of (3.170) and (3.171), all quantities in
Proposition 3.17 are equal. Proving the equality of (3.169) and (3.170), and
of (3.171) and (3.172) is a separate issue. In fact the equality of (3.169) and
(3.170) follows from Lemma 3.16 equality (3.153), and the equality of (3.171)
and (3.172) follows from the same lemma equality (3.154).
(3.168) ≥ (3.169). Fix the subset E0 of E, and let (hj )j∈N ⊂ D(L) with the
n µ ¶j−1
λ X λ
following property: Lh0 = lim 1− hj . Here the convergence
n→∞ λ0 λ0
j=1
is uniform on E0 . In fact each hj may chosen equal to h0 . In (3.168) we choose
all xj = x ∈ E0 . Then we get
Xn µ ¶j ½ µ ¶ ¾
λ λ 1
1− h (xj ) + g (xj+1 ) − I − L hj (xj+1 )
j=0
λ0 λ0 λ0
Xn µ ¶ n µ ¶j
λ X
j
λ λ
= h0 (x0 ) + 1− hj (x) + 1− g(x) − h0 (x)
j=1
λ0 λ0 j=0 λ0
 
Xn µ ¶j Xn µ ¶j
λ 1 λ λ
− 1− hj (x) + L  1− hj  (x)
j=1
λ 0 λ λ 0 j=0
λ 0
140 3 Space-time operators
n µ ¶j µ ¶
λ X λ 1
= h0 (x0 ) + 1− g(x) − I − L h0 (x)
λ0 j=0 λ0 λ
 
µ ¶n+1 X ∞ µ ¶j−n−1
1 λ λ λ
− 1− L 1− hj  (x). (3.193)
λ λ0 λ0 j=n+1 λ0

The expression in (3.193) tends to


µ ¶
1
h0 (x0 ) + g(x) − I − L (x) uniformly on E0 , (3.194)
λ

and consequently, since h0 ∈ D(L) may be chosen arbitrarily, we see that


(3.168) ≥ (3.169).
(3.173) ≤ (3.172). The proof of this inequality follows the same lines as
the proof of (3.168) ≥ (3.169). In fact it follows from the latter inequality by
applying it to −g instead of g. The reader is invited to check the details.
(3.169) ≥ (3.172). Consider the mapping Λ+ : Cb (E, R) → [−∞ + ∞)
defined by
½ µ ¶ ¾
+ λ
Λ (g) = inf sup h (x0 ) + g(x) − I − L h(x) . (3.195)
h∈D(L) x∈E0 λ0

where g ∈ Cb (E, R). From the σ-local maximum principle (see Definition
3.14) and Lemma 3.16, inequality (3.156) it follows that Λ+ attains its values
in R. In addition, the functional Λ+ is sub-additive, and the expression in
(3.172) is equal to −Λ+ (−g). It follows that

Λ+ (g) + Λ+ (−g) ≥ Λ+ (0)


½ µ ¶ ¾
1
= inf h (x0 ) : I − L h ≥ 0, on E0 ≥ 0. (3.196)
h∈D(L) λ

In (3.196) we used the σ-local maximum principle: compare with the argu-
ments in (3.163) of the proof of inequality (3.157) in Lemma 3.16.

Proposition 3.19. Suppose that the operator L possesses the Korovkin prop-
erty on E0 . Then for all λ > 0 and f ∈ Cb (E) the quantities in (3.169),
(3.170), (3.171), and (3.172) are equal for all x0 ∈ E. These quantities are
also equal to
½ µ ¶ ¾
1
sup inf h (x0 ) : v I − L h ≥ vg (3.197)
v∈H + (E) h∈D(L) λ
½ µ ¶ ¾
1
= inf sup h (x0 ) : v I − L h ≤ vg . (3.198)
v∈H + (E) h∈D(L) λ
3.3 Korovkin property 141

Recall that H + (E) stands for all functions u ∈ H(E), u ≥ 0, with the property
that for every α > 0 the level set {u ≥ α} is a compact subset of E. Observe
that for every u ∈ H(E) there exists a function u0 ∈ H + (E) such that
|u(x)| ≤ u0 (x) for all x ∈ E.
Corollary 3.20. Suppose that the operator L possesses the Korovkin property
on E0 , and is positive Tβ -dissipative on E0 . Then the family {λR(λ) : λ ≥ λ0 },
as defined in Proposition 3.17, is Tβ -equi-continuous (on E0 ) for some λ0 > 0.
Proof. We use the representation in (3.171):
½ µ ¶ ¾
1
λR(λ)f (x0 ) = sup h (x0 ) : I − L h ≤ f on E0 . (3.199)
h∈D(L) λ

Let u ∈ H + (E) and x0 ∈ E. Since L is supposed to be positive Tβ -dissipative


on E0 , there exists λ0 > 0 and v ∈ H + (E0 ) such that
λu (x0 ) h (x0 ) ≤ sup v(x) (λh(x) − Lh(x)) (3.200)
x∈E0

for all h ∈ D(L) which are real-valued and for all λ ≥ λ0 . For the precise
definition of positive Tβ -dissipativity (on E) see (3.15) in Definition 3.5. From
(3.199) and (3.200) we infer:
u (x0 ) λR(λ)f (x0 )
= sup {u (x0 ) h (x0 ) : λh − Lh ≤ λf on E0 }
h∈D(L)

≤ sup {u (x0 ) h (x0 ) : λh − Lh ≤ λf on E0 }


h∈D(L)
½ ¾
1
≤ sup sup v(x) (λh (x) − Lh(x)) : λh − Lh ≤ λf on E0
h∈D(L) λ x∈E
≤ sup v(x)f (x). (3.201)
x∈E0

Since by construction <R(λ)f = R(λ)<f , (3.201) implies:


kuλR(λ)f k∞ ≤ kvf k∞ , f ∈ Cb (E0 ) , λ ≥ λ0 . (3.202)
The conclusion in Corollary 3.20 is a consequence of (3.202).
In the following theorem we wrap up more or less everything we proved so far
about an operator with the Korovkin property on a subset E0 of E. Theorem
3.21 and the related observations are used in the proof item (e) of Theorem
1.39.
Theorem 3.21. Let E0 be a polish subspace of the polish space E. Suppose
that every function f ∈ Cb (E0 ) can be extended to a bounded continuous func-
tion on E. Let L be a linear operator with domain and range in Cb (E) which
assigns the zero function to a constant function. Suppose that the operator L
possesses the following properties:
142 3 Space-time operators

1. Its domain D(L) is Tβ -dense in Cb (E).


2. The operator L assigns real-valued functions to real-valued functions:
< (Lf ) = L<f for all f ∈ D(L).
3. If f ∈ D(L) vanishes on E0 , then Lf vanishes on E0 as well.
4. The operator L satisfies the maximum principle on E0 .
5. The operator L is positive Tβ -dissipative on E0 .
6. The operator L is sequentially λ-dominant on E0 for some λ > 0.
7. The operator L has the Korovkin property on E0 .
Let L ¹E0 be the operator defined by D (L ¹E0 ) = {f ¹E0 : f ∈ D(L)}, and
L ¹E0 (f ¹E0 ) = Lf ¹E0 , f ∈ D(L). Then the operator L ¹E0 pos-
sesses a unique linear extension to the generator L0 of a Feller semigroup
{S0 (t) : t ≥ 0} on Cb (E0 ).
In addition, the time-homogeneous Markov process associated to the Feller
semigroup {S0 (t) : t ≥ 0} serves as the unique solution to the martingale prob-
lem associated with L.

Proof. Existence. First we prove that the restriction operator L ¹E0 is well-
defined and that it is Tβ -densely defined. The fact that it is well-defined
follows from 3. In order to prove that it is Tβ -densely defined, we use a Hahn-
Banach typeRargument. Let µ e be a bounded Borel measure on E0 such that
hf ¹E0 , µ
ei = E0 f de
µ = 0 for all f ∈ D(L). Define the measure µ on the Borel
field of E by µ(B) = µ e (B ∩ E0 ), B ∈ E. Then hf, µi = 0 for all f ∈ D(L).
Since D(L) is Tβ -dense in Cb (E), we infer hf, µi = 0 for all f ∈ Cb (E). Let
fe ∈ Cb (E). Then there exists f ∈ Cb (E) such that f = fe on E0 , and hence
D E
fe, µ
e = hf ¹E0 , µ
ei = hf, µi = 0. (3.203)

From (3.203) we see that a bounded Borel measure which annihilates D (L ¹E0 )
also vanishes on Cb (E0 ). By the theorem of Hahn-Banach in combination with
the fact that every element of the dual of (Cb (E0 ) , Tβ ) can be identified with
a bounded Borel measure on E0 , we see that the subspace D (L ¹E0 ) is Tβ -
dense in Cb (E0 ). Define the family of operators {λR(λ) : λ > 0} as in Propo-
sition 3.17, By the properties 4 and 7 such definitions make sense. Moreover,
the family {R(λ) : λ > 0} possesses the resolvent property: R(λ) − R(µ) =
(µ − λ) R(µ)R(λ), λ > 0, µ > 0. It also follows that R(λ) (λI − D1 − L) f = f
on E0 for f ∈ D(1) (L). This equality is an easy consequence of the inequalities
in (2.157): see Corollary 3.18. Fix λ > 0 and f ∈ Cb (E0 ). If f is of the form
f = R(λ)g, g ∈ Cb (E0 ), then by the resolvent property we have

α αR(α)g
αR(α)f − f = αR(α)R(λ)g − R(λ)g = R(λ)g − R(λ)g − .
α−λ α−λ
(3.204)
Since kαR(α)gk∞ ≤ kgk∞ , g ∈ Cb (E0 ), the equality in (3.204) yields

k·k∞ - lim αR(α)f − f = 0 for f of the form f = R(λ)g, g ∈ Cb (K).


α→∞
3.3 Korovkin property 143

Since g = R(λ) (λI − D1 − L) g on K, g ∈ D(1) (L), it follows that

lim kαR(α)g − gk∞ = 0 for g ∈ D(1) (L) = D (D1 ) ∩ D(L). (3.205)


α→∞

As was proved in Corollary 3.20 there exists λ0 > 0 such that the family
{λR(λ) : λ ≥ λ0 } is Tβ -equi-continuous. Hence for u ∈ H + (E0 ) there exists
v ∈ H + (K) that for α ≥ λ0 we have

kuαR(α)gk∞ ≤ kvgk∞ , g ∈ Cb (E0 ) . (3.206)

Fix ε > 0, and choose for f ∈ Cb (E0 ) and u ∈ H + (E0 ) given the function
g ∈ D (L ¹E0 ) in such a way that
2
ku(f − g)k∞ + kv(f − g)k∞ ≤ ε. (3.207)
3
Since D (L ¹E0 ) is Tβ -dense in Cb (E0 ) such a choice of g. The inequality
(3.207) and the identity

αR(α)f − f = αR(α)(f − g) − (f − g) + αR(α)g − g yield


ku (αR(α)f − f )k∞
≤ ku (αR(α)(f − g))k∞ + ku(f − g)k∞ + kuαR(α)g − gk∞
≤ kv(f − g)k∞ + ku(f − g)k∞ + kuαR(α)g − gk∞
2
≤ ε + kuαR(α)g − gk∞ . (3.208)
3
From (3.205) and (3.208) we infer

Tβ - lim αR(α)f = f, f ∈ Cb (E0 ) . (3.209)


α→∞

Define the operator L0 in Cb (E0 ) as follows. Its domain is given by D (L0 ) =


R(λ)Cb (E0 ), λ > 0. By the resolvent property the space R(λ)Cb (E0 ) does
not depend on λ > 0, and so D (L0 ) is well-defined. The operator L0 :
D (L0 ) → Cb (E0 ) is defined by L0 R(λ)f = λR(λ)f − f , f ∈ Cb (E0 ). Since
R(λ)f1 = R(λ)f2 , f1 , f2 ∈ Cb (E0 ), implies R(λ) (f2 − f1 ) = 0. By the resol-
vent property we see that αR(α) (f2 − f1 ) = 0 for all α > 0. From (3.209) we
infer f2 = f1 . In other words, the operator L0 is well-defined. Since the oper-
ators R(λ), λ > 0, are Tβ -continuous it follows that the graph of the operator
L0 is Tβ -closed. As in the proof of (iii) =⇒ (i) we have, like in (3.111),

X (λt)k
Se0 (t)f = lim e−λt
k
(λR(λ)) f = lim e−λt eλt(λR(λ)) f = S0 (t)f
λ→∞ k! λ→∞
k=0
(3.210)
e
where the operator S0 (t) is defined by using the Hausdorff-Bernstein-Widder
Laplace inversion theorem we have (compare with (3.108))
144 3 Space-time operators
Z ∞
e−λρ Se0 (ρ)f (x)dρ, <λ > 0, x ∈ E0 .
−1
R(λ)f (x) = (λI − L0 ) f (x) =
0
(3.211)
For a function f belonging to the space R(λ)Cb (E0 ) the equality S0 (t)f =e
S0 (t)f holds: see (3.113). Here S0 (t)f is defined as the uniform limit in (3.210).
Since the operator L ¹E0 is sequentially
n λ-dominant for some λo > 0 we in-
k
fer that the family of operators (µR (λ + µ)) : µ ≥ 0, k ∈ N is Tβ -equi-
continuous: see (1.43) in Proposition 1.21. Since for f ∈ R(α)Cb (E0 ) we have
© ª n o
k
sup e−λt S0 (t)f : t ≥ 0 = sup (µR (λ + µ)) f : µ ≥ 0, k ∈ N . (3.212)

From (3.212) combined with the Tβ -equi-continuity and the Tβ density of


D (L ¹E0 ) we see that, each operator S0 (t) has a Tβ -continuous extension
to all of Cb (E0 ). One way of achieving this is by fixing f ∈ Cb (E0 ), and
considering the family {αR(α)f : α ≥ λ}. Then Tβ - lim©α→∞ αR(α)f = f ª . Let
u ∈ H + (E0 ). By the Tβ -equi-continuity of the family e−λt S0 (t) : t ≥ 0 we
see that
° °
lim sup °ue−λt S0 (t) (βR(β)f − αR(α)f )°∞ = 0. (3.213)
α, β→∞ t≥0

Since the functions (t, x) 7→ S0 (t) (αR(α)f ) (x), α ≥ λ, are continuous


the same is true for the function (t, x) 7→ [S0 (t)f ] (x), where S0 (t)f =
Tβ - lim αR(α)f . Of course, for almost all t ≥ 0 we have S0 (t)f (x) = Se0 (t)t(x)
α→∞
for all x ∈ E0 . Since
1¡ ¢
Tβ - lim I − e−λt S0 (t) R(λ)f = R(λ)f, f ∈ Cb (E0 ) ,
t↓ t

we see that the operator L0 generates the semigroup {S0 (t) : t ≥ 0}. The
continuous extension of S0 (t), which was originally defined on R(λ)Cb (E0 ),
to Cb (E0 ) is again denoted by S0 (t). Let f ∈ D(L). Moreover, since

R(λ) (λf − Lf ) = f on E0 ,

we have D (L ¹E0 ) ⊂ D (L0 ), and

L0 f = L0 R(λ) (λI − L) f = λR(λ) (λI − L) f − (λI − L) f


= λf − λf + Lf = Lf (3.214)

on E0 . From (3.214) we see that the operator L0 extends the operator L ¹E0 .
Uniqueness of Feller semigroups. Let L1 and L2 be two extensions of the
operator L ¹E0 which generate Feller semigroups. Let {R1 (λ) : λ > 0} and
{R2 (λ) : λ > 0} be the corresponding resolvent families. Since L1 extends
L ¹E0 we obtain, for h ∈ D(L),
3.3 Korovkin property 145
µ ¶
1
λ0 R (λ0 ) I − L h = R (λ0 ) (λ0 I − L1 ) h = h. (3.215)
λ0

Then by the maximum principle and (3.215) we infer


µ µ ¶ ¶
1
sup inf h (x0 ) + g(x) − I − L h (x)
h∈D(L) x∈E0 λ0
µ µ µ ¶ ¶ ¶
1
≤ sup h (x0 ) + λ0 R (λ0 ) g − I − L h (x0 )
h∈D(L) λ0
µ µ ¶ ¶
1
= sup h (x0 ) + λ0 R1 (λ0 ) g (x0 ) − λ0 R1 (λ0 ) I − L h (x0 )
h∈D(L) λ0
= sup (h (x0 ) + λ0 R1 (λ0 ) g (x0 ) − h (x0 ))
h∈D(L)

= λ0 R1 (λ0 ) g (x0 )
µ µ ¶ ¶
1
≤ inf sup h (x0 ) + g(x) − I − L h(x) . (3.216)
h∈D(L) x∈E0 λ0

The same reasoning can be applied to the operator R2 (λ0 ) Since the ex-
tremities in (3.215) are equal we see that R1 (λ0 ) = R2 (λ0 ). Hence we get
−1 −1
(λ0 − L1 ) = (λ0 − L2 ) , and consequently L1 = L2 .

Of course the same arguments work if E0 = E.


Uniqueness of solutions to the martingale problem. Let L0 be the (unique)
extension of L, which generates a Feller semigroup {S0 (t) : t ≥ 0}, and let

{(Ω, F, Px ) , (X(t), t ≥ 0) , (ϑt , t ≥ 0) , (E, E)}

be the corresponding time-homogeneous Markov process with Ex [g(X(t)] =


S0 (t)g(x), g ∈ Cb (E), x ∈ E, t ≥ 0. Then the family {Px : x ∈ E} is a solu-
tion to the martingale problem associated to L. The proof of the uniqueness
part follows a pattern similar to the proof of the uniqueness part of linear
extensions of L which generate Feller semigroups. We will show that the
family of probability measures {Px : x ∈ E} is a solution to the martingale
problem associated to the operator L. Let f be a member of D(L) and put
Rt
Mf (t) = f (X(t)) − f (X(0)) − 0 Lf (X(s))ds. Then, for t2 > t1 we have
£ ¯ ¤ £ ¯ ¤
Ex Mf (t2 ) ¯ Ft1 − Mf (t1 ) = Ex Mf (t2 − t1 ) ◦ ϑt1 ¯ Ft1

(Markov property)

= EX(t1 ) [Mf (t2 − t1 )] . (3.217)

Since, in addition, by virtue of the fact that L0 , which is an extension of L,


generates the semigroup {S0 (t) : t ≥ 0}, we have
146 3 Space-time operators
Z t
Ez [Mf (t)] = S0 (t)f (z) − f (z) − S0 (u)Lf (z)du
0
Z t

= S0 (t)f (z) − f (z) − (S0 (u)f (z)) du
0 ∂u
= S0 (t)f (z) − f (z) − (S0 (t)f (z) − S0 (0)f (z)) = 0,
the assertion about the existence of solutions to the martingale problem fol-
lows from 3.217. Next we prove uniqueness of solutions to the martingale
problem. Its proof resembles the way we provedn the uniqueness
o n of extensions
o
(1) (2)
of L which generate Feller semigroups. Let Px : x ∈ E and Px : x ∈ E
be two solutions to the martingale problem for L. Let h ∈ D(L), and consider
Z ∞ · µ ¶ ¸
1 ¯
λ −λt (j)
e Ex h (X(s)) − I − L h (X(t + s)) Fs dt ¯
0 λ
Z ∞
£ ¯ ¤
= h (X(s)) − e−λt E(j)
x (λI − L) h (X(t + s)) ¯ Fs dt
0
Z ∞
£ ¯ ¤
= h (X(s)) − λ e−λt E(j)
x h (X(t + s)) ¯ Fs dt
0
Z ∞
£ ¯ ¤
+ e−λt E(j)
x Lh (X(t + s)) ¯ Fs dt
0

(integration by parts)
Z ∞
£ ¯ ¤
= h (X(s)) − λ e−λt E(j)
x h (X(t + s)) ¯ Fs dt
0
Z ∞ ·Z t ¸
¯
+λ e−λt E(j)
x Lh (X(ρ + s)) dρ ¯ F s dt
0 0
Z ∞
£ ¯ ¤
= h (X(s)) − λ e−λt E(j)
x h (X(t + s)) ¯ Fs dt
0
Z ∞ ·Z t+s ¸
¯
+λ e−λt E(j)
x Lh (X(ρ)) dρ ¯ F s dt
0 s

(martingale property)
∞Z
£ ¯ ¤
= h (X(s)) − λ e−λt E(j)
x h (X(t + s)) ¯ Fs dt
0
Z ∞
£ ¯ ¤
+λ e Ex h (X(t + s)) − h (X(s)) ¯ Fs dt = 0.
−λt (j)
(3.218)
0

Fix x0 ∈ E, g ∈ Cb (E), and s > 0. Then, from (3.218) it follows that, for
h ∈ D(L),
Z ∞
£ ¯ ¤
e−λt E(j) ¯
x0 g (X(t + s)) Fs dt
0
3.3 Korovkin property 147
Z ∞ · µ ¶ ¸
1 ¯
= e−λt E(j)
x0 h (X(s)) + g (X(t + s)) − I − L h (X(t + s)) ¯ Fs dt,
0 λ

and hence
Z ∞ £ ¯ ¤
Λ− (g, X(s), λ) ≤ λ exp(−λt)E(j) ¯ +
x0 g (X(t + s)) Fs dt ≤ Λ (g, X(s), λ) ,
0
(3.219)
for j = 1, 2, where
½ · µ ¶ ¸ ¾
1
Λ+ (g, x0 , λ) = inf sup min max h(x0 ) + g − I − L h (x)
Γ ⊂D(L) Φ⊂E0 h∈Γ x∈Φ∪{4} λ
#Γ <∞ #Φ<∞
· µ ¶ ¸
½ ¾
1
= inf sup h(x0 ) + g − I − L h (x) , and (3.220)
h∈D(L) x∈E0 λ
½ · µ ¶ ¸ ¾
− 1
Λ (g, x0 , λ) = sup inf max min h(x0 ) + g − I − L h (x)
Γ ⊂D(L) Φ⊂E0 h∈Γ x∈Φ∪{4} λ
#Γ <∞ #Φ<∞
½ · µ ¶ ¸ ¾
1
= sup inf h(x0 ) + g − I − L h (x) . (3.221)
h∈D(L) x∈E0 λ

We also have
Z ∞ · µ ¶ ¸
(j) 1
λ e−λt EX(s) h (X(0)) − I − L h (X(t)) dt
0 λ
Z ∞ Z ∞
(j) (j)
= h (X(s)) − λ e−λt EX(s) [h (X(t))] dt + e−λt EX(s) [Lh (X(t))] dt
0 0

(integration by parts)
Z ∞
(j)
= h (X(s)) − λ e−λt EX(s) [h (X(t))] dt
0
Z ∞ ·Z t ¸
(j)
+λ e−λt EX(s) Lh (X(ρ)) dρ dt
0 0

(martingale property)
Z ∞
(j)
= h (X(s)) − λ e−λt EX(s) [h (X(t))] dt
0
Z ∞
(j)
+λ e−λt EX(s) [h (X(t)) − h (X(0))] dt = 0 (3.222)
0

(j)
where in the first and final step we used X(0) = z Pz -almost surely. In the
same spirit as we obtained (3.219) from (3.222) we get
148 3 Space-time operators
Z ∞
(j)
Λ− (g, X(s), λ) ≤ λ e−λt EX(s) [g (X(t))] dt ≤ Λ+ (g, X(s), λ) , (3.223)
0

for j = 1, 2. Since, by Proposition 3.17 (formula (3.169) and (3.172)) the


identity Λ+ (g, x, λ) = Λ− (g, x, λ), is true for g ∈ Cb (E), x ∈ E, λ > 0, we
(1) (2)
obtain, by putting s = 0, Ex [g(X(t))] = Ex [g(X(t))], t ≥ 0, g ∈ Cb (E).
(1)
We also obtain, Px -almost surely,
£ ¯ ¤ (1)
E(1)
x g(X(t + s)) ¯ Fs = EX(s) [g(X(t))] ,

(2)
and, Px -almost surely,
£ ¯ ¤ (2)
E(2)
x g(X(t + s)) ¯ Fs = EX(s) [g(X(t))] , for t, s ≥ 0, and g ∈ Cb (E).

(1) (2)
It necessarily follows that Px = Px , x ∈ E. Consequently, the uniqueness
of the solutions to the martingale problem for the operator L follows.
This completes the proof Theorem 3.21.

3.4 Continuous sample paths

The following Lemma 3.22 and Proposition 3.23 give a general condition which
guarantee that the sample paths are Pτ,x -almost surely continuous on their
life time.
Lemma 3.22. Let P (τ, x; t, B), 0 ≤ τ ≤ t ≤ T , x ∈ E, B ∈ E, be a sub-
Markov transition function. Let (x, y) 7→ d(x, y) be a continuous metric on E×
E and put Bε (x) = {y ∈ E : d(y, x) ≤ ε}. Fix t ∈ (0, T ]. Then the following
assertions are equivalent:
(a) For every compact subset K of E and for every ε > 0 the following equality
holds:
P (s1 , x; s2 , E \ Bε (x))
lim sup = 0. (3.224)
s1 ,s2 →t, τ <s1 <s2 ≤t x∈K s2 − s1

(b) For every compact subset K of E and for every open subset G of E such
that G ⊃ K the following equality holds:

P (s1 , x; s2 , E \ G)
lim sup = 0. (3.225)
s1 ,s2 →t, τ ≤s1 <s2 ≤t x∈K s2 − s1

Proof. (a) =⇒ (b). Let G be an open subset of E and let K be a compact


subset of G. Then there exists ε > 0, n ∈ N, and xj ∈ K, such that

G ⊃ ∪nj=1 B2ε (xj ) ⊃ ∪nj=1 int (Bε (xj )) ⊃ K. (3.226)


3.4 Continuous sample paths 149

For any x ∈ K there exists j0 , 1 ≤ j0 ≤ n, such that d (x, xj0 ) < ε, and
hence for y ∈ int (Bε (x)) d (y, xj0 ) ≤ d (y, x) + d (x, xj0 ) < 2ε. It follows
that Bε (x) ⊂ G. Consequently, for x ∈ K and τ ≤ s1 < s2 < t we get
P (s1 , x; s2 , E \ G) ≤ P (s1 , x; s2 , Bε (x)). So (b) follows from (a).
(b) =⇒ (a). Fix ε > 0 and let K be any compact subset of E. Like in the proof
of the implication (a) =⇒ ¡ (b) we¢again choose elements xj ∈ K, 1 ≤ j ≤ n,
such that K ⊂ ∪nj=1 int Bε/4 (xj ) . Let x ∈ K ∩ Bε/4 (xj ) and y ∈ Bε/2 (xj ).
Then d(y, x) ≤ d (y, xj ) + d (xj , x) ≤ 12 ε + 14 ε = 43 ε < ε. Suppose that x ∈
K ∩ Bε/4 (xj ). For τ ≤ s1 < s2 < t it follows that
¡ ¡ ¢¢
P (s1 , x; s2 , E \ Bε (x)) ≤ P s1 , x; s2 , E \ int Bε/2 (xj ) ,

and hence

sup P (s1 , x; s2 , E \ Bε (x))


x∈K
¡ ¡ ¢¢
≤ max sup P s1 , x; s2 , E \ int Bε/2 (xj ) . (3.227)
1≤j≤n x∈K∩B
ε/4 (xj )

The inequality in (3.227) together with assumption in (b) easily implies (a).
This concludes the proof of Lemma 3.22.
Proposition 3.23. Let P (τ, x; t, B) be a sub-Markov transition function and
let the process X(t) be as in (a) of Theorem 1.39. Fix (τ, x) ∈ [0, T ] × E.
Suppose that for every t ∈ [τ, T ], and for every compact subset K and for
every open subset G for which G ⊃ K the equality
P (s1 , y; s2 , E \ G)
lim sup =0
s1 ,s2 ↑t, τ ≤s1 <s2 <t y∈K s2 − s1

holds. Then for every t ∈ (τ, T ] the equality

inf sup d (X(s), X(t)) 1[X(t)∈E] = 0, holds Pτ,x -almost surely.


ε>0 t−ε≤s≤t

Here d : E × E → [0, ∞) is a continuous metric on E × E.


Proof. Put tj,n = t − ε + j2−n ε, 0 ≤ j ≤ 2n . From Proposition 2.2 with X(s)
e
instead of X(s) it follows that it suffices to prove that for every η > 0 the
equality
· ¸
inf lim Pτ,x max n d (X (tj−1,n ) , X (tj,n )) 1{X(tj−1,n )∈K} 1{X(tj,n )∈K} > η
ε>0 n→∞ 1≤j≤2

=0 (3.228)

holds for all compact subsets K of E.


· ¸
Pτ,x max n d (X (tj−1,n ) , X (tj,n )) 1{X(tj−1,n )∈K} 1{X(tj,n )∈K} > η
1≤j≤2
150 3 Space-time operators
n
2
X £ ¤
≤ Pτ,x d (X (tj−1,n ) , X (tj,n )) 1{X(tj−1,n )∈K} 1{X(tj,n )∈K} > η
j=1

(Markov property)
n
2
X £ £ ¤
= Eτ,x Ptj−1,n ,X(tj−1 ,n) d (X (tj−1,n ) , X (tj,n )) 1{X(tj,n )∈K} > η
j=1
¤
×1{X(tj−1,n )∈K}
n
2
X £ ¤
≤ sup Ptj−1,n ,y d (y, X (tj,n )) 1{X(tj,n )∈K} > η
j=1 y∈K
n
2
X
= sup P (tj−1,n , y; tj,n , K \ Bη (y)) . (3.229)
j=1 y∈K

The result in Proposition 3.23 follows from (3.229) and Lemma 3.22.

3.5 Measurability properties of hitting times


In this section we study how fast Markov process reaches a Borel subset B
of the state space E. The material is taken from Chapter 2, Section 2.10 in
Gulisashvili et al [95]. Fix τ ∈ [0, T ]. Throughout this section we will assume
that the filtrations (Ftτ )t∈[τ,T ] are right-continuous and Pτ,µ -complete. Right-
τ
T
continuity means that Ft+ = Fsτ = Ftτ . By definition the σ-field Ftτ is
s∈(t,T ]
τ
Pτ,µ -complete if Pτ,µ -negligible events A belong to Ftτ . The σ-field Ft+ is the
τ τ
Pτ,µ -completion of a σ-field Ft+ if and only if for every A ∈ Ft+ there exist
τ
events A1 and A2 ∈ Ft+ such that A1 ⊂ A ⊂ A2 and Pτ,x (A2 \ A1 ) = 0. It
is also assumed that we are in the context of a backward Feller evolution (or
propagator) {P (s, t) : 0 ≤ s ≤ t ≤ T } in the sense of Definition 1.24 and the
corresponding strong Markov process with state space E:
n o
(Ω, FTτ , Pτ,x )(τ,x)∈[0,T ]×E , (X(t), t ∈ [0, T ]) , (E, E) . (3.230)

By P (E) will be denoted the collection of all Borel probability R measures on


the space E. For A ∈ FTτ and µ ∈ P (E), we put Pτ,µ (A) = Pτ,x (A)dµ(x).
For instance, if µ = δx is the Dirac measure concentrated at x ∈ E, then
Pτ,δx = Pτ,x . Let ζ be the first time the process X(t) arrives at the absorption
state 4:
(
inf {t > 0 : X(t) = 4} if X(t) = 4 for some t ∈ (0, T ],
ζ=
T if X(t) ∈ E for all t ∈ (0, T ).
3.5 Measurability properties of hitting times 151

Definition 3.24. Let (X(t), Pτ,x )¡ be a Markov


¢ process on Ω with state space
E and sample path space Ω = D [0, T ], E 4 , and let B be a Borel subset of
E 4 . Let τ ∈ [0, T ), and suppose that S : Ω → [τ, ζ] is a Ftτ -stopping time.
For the process X(t), the entry time of the set B after time S is defined by
 [

 inf {t : t ≥ S, X(t) ∈ B} on {S ≤ t, X(t) ∈ B} ,
S
DB = τ ≤t<T

 ζ elsewhere.
(3.231)
The pseudo-hitting time of the set B after time S is defined by
 [

 inf {t : t ≥ S, X(t) ∈ B} on {S ≤ t, X(t) ∈ B} ,
eS =
D τ <t<T
B

 ζ elsewhere.
(3.232)
The hitting time of the set B after time S is defined by
 [

 inf {t : t > S, X(t) ∈ B} on {S < t, X(t) ∈ B} ,
TBS = τ ≤t<T

 ζ elsewhere.
(3.233)
Observe that on the event {S = τ, X(τ ) ∈ B} we have DBS
= τ and DeS = T S.
B B
It is not hard to prove that
[ [
{S ≤ t, X(t) ∈ B} = {S ∨ t ≤ t, X (S ∨ t) ∈ B} and
t:τ ≤t<T t:τ ≤t<T
[ [
{S ≤ t, X(t) ∈ B} = {S ∨ t ≤ t, X (S ∨ t) ∈ B} .
t:τ <t<T t:τ <t<T

We also have
S
DB∪{4} S
= DB ∧ ζ, eS
D eS S S
B∪{4} = DB ∧ ζ, and TB∪{4} = TB ∧ ζ. (3.234)

In addition, we have DB S
≤De S ≤ T S . Next we will show that the following
B B
equalities hold:
n o n o
(ε+S)∧ζ (r+S)∧ζ
TBS = inf DB = inf DB . (3.235)
ε>0 r∈Q+
© ª
Indeed on TBS < ζ , the first equality in (3.235) can be obtained by using
the inclusion

{t ≥ (ε + S) ∧ ζ, X(t) ∈ B} ⊂ {t > S, X(t) ∈ B}

and the fact that for every t ∈ [τ, T ) and ω ∈ {S < t, X(t) ∈ B}, there exists
ε > 0 depending on ω such that ω ∈ {(ε + S) ∧ ζ ≤ t, X(t) ∈ B}. Since TBS ≤
152 3 Space-time operators
(ε+S)∧ζ © ª
DB , we see that on the event TBS = ζ the first equality in (3.235) also
holds. The second equality in (3.235) follows from the monotonicity of the
S
entry time DB with respect to S.
Our next goal is to prove that for the Markov process in (3.230) the entry
time DB S
, the pseudo-hitting time D e S , and the hitting time T S are stopping
B B
times. Throughout the present section, the symbols K(E) and O(E) stand for
the family of all compact subsets and the family of all open subsets of the
space E, respectively.
The celebrated Choquet capacitability theorem will be used in the proof
S eS
of the fact that DB , DB , and TBS are stopping times. We will restrict ourselves
to positive capacities and the pavement of the space E by compact subsets.
For more general cases, we refer the reader to [73, 160].
Definition 3.25. A function I from the class P(E) of all subsets of E into
the extended real half-line R̄+ is called a Choquet capacity if it possesses the
following properties:
(i) If A1 and A2 in P(E) are such that A1 ⊂ A2 , then I (A1 ) ≤ I (A2 ).
(ii)If An ∈ P(E), n ≥ 1, and A ∈ P(E) are such that An ↑ A, then I (An ) →
I(A) as n → ∞.
(iii)If Kn ∈ K(E), n ≥ 1, and K ∈ K(E) are such that Kn ↓ K, then
I (Kn ) → I(K) as n → ∞.
Definition 3.26. A function ϕ : K(E) → [0, ∞) is called strongly sub-
additive provided that the following conditions hold:
(i) If K1 ∈ K(E) and K2 ∈ K(E) are such that K1 ⊂ K2 , then ϕ (K1 ) ≤
ϕ (K2 ).
(ii)If K1 and K2 belong to K(E), then
ϕ (K1 ∪ K2 ) + ϕ (K1 ∩ K2 ) ≤ ϕ (K1 ) + ϕ (K2 ) . (3.236)
The following construction allows one to define a Choquet capacity starting
with a strongly sub-additive function. Let ϕ be a strongly sub-additive func-
tion satisfying the following additional continuity condition:
(iii)For all K ∈ K(E) and all ε > 0, there exists G ∈ O(E) such that K ⊂ G
and ϕ (K 0 ) ≤ ϕ (K) + ε for all compact subsets K 0 of G.
For any G ∈ O(E), put
I ∗ (G) = sup ϕ(K). (3.237)
K∈K(E);K⊂G

Next define a set function I : P(E) → R̄+ by


I(A) = inf I ∗ (G), A ∈ P(E). (3.238)
G∈O(E);A⊂G

It is known that the function I is a Choquet capacity. It is clear that for


any G ∈ O(E), I(G) = I ∗ (G). Moreover, it is not hard to see that for any
K ∈ K(E), ϕ(K) = I(K).
3.5 Measurability properties of hitting times 153

Definition 3.27. Let ϕ : K(E) → [0, ∞) be a strongly subadditive function


satisfying condition (iii), and let I be the Choquet capacity obtained from ϕ
(see formulas (3.237) and (3.238)). A subset B of E is said to be I-capacitable
if the following equality holds:

I(B) = sup {ϕ(K) : K ⊂ B, K ∈ K(E)} . (3.239)

Now we are ready to formulate the Choquet capacitability theorem (see, e.g.,
[73, 69, 160]). We will also need the following version of the Choquet capacity
theorem. For a discussion on capacitable subsets see e.g. Kiselman [133]; see
Choquet [58], [27] and [69] as well. For a general discussion on the foundations
of probability theory see e.g. [118].
Theorem 3.28. Let E be a polish space, and let ϕ : K(E) → [0, ∞) be a
strongly subadditive function satisfying condition (iii), and let I be the Choquet
capacity obtained from ϕ (see formulas (3.237) and (3.238)). Then every an-
alytic subset of E, and in particular, every Borel subset of E is I-capacitable.
The definition of analytic sets can be found in [73, 69]. We will only need the
Choquet capacitability theorem for Borel sets which form a sub-collection of
the analytic sets.
Lemma 3.29. Let τ ∈ [0, T ], and let {X(t) : t ∈ [τ, T ]} be an adapted, right-
continuous, and
µ quasi left-continuous ¶stochastic process on the filtered prob-
³ τ ´ τ
ability space X(t), Ft+ , Pτ,x . Suppose that S is an Ft+ -stopping
t∈[τ,T ]
time such that τ ≤ S ≤ ζ. Then, for any t ∈ [τ, T ] and µ ∈ P (E), the following
functions are strongly sub-additive on K(E) and satisfy condition (iii):
£ S ¤ h i
K 7→ Pτ,µ DK ≤ t , and K 7→ Pτ,µ D eKS
≤ t , K ∈ K(E). (3.240)

τ
We wrote Ft+ to indicate that this σ-field is right continuous and Pτ,x -
complete.
Proof. We have to check conditions (i) and (ii) in Definition 3.26 and also
condition (iii) for the set functions in (3.240). Let K1 ∈ K(E) and K2 ∈ K(E)
S S
be such that K1 ⊂ K2 . Then DK 1
≥ DK 2
, and hence
£ S ¤ £ S ¤
Pτ,µ DK 1
≤ t ≤ Pτ,µ DK 2
≤t .
£ S ¤
This proves condition (i) for the function K 7→ Pτ,µ DK ≤ t . The proof of
(i) for the second mapping in (3.240) is similar. £ S ¤
In order to prove condition (iii) for the mapping K 7→ Pτ,µ DK ≤ t , we
use assertion (a) in Lemma 3.34. More precisely, let K ∈ K(E) and Gn ∈
O(E), n ∈ N, be such as in Lemma 3.34. Then by part (a) of Lemma 3.34
below (note that part (a) of Lemma 3.34 also holds under the restrictions in
Lemma 3.29), we get
154 3 Space-time operators
£ S ¤ £ S ¤
Pτ,µ DK ≤t ≤ inf sup Pτ,µ DK 0 ≤ t
G∈O(E):G⊃K K 0 ∈K(E):K 0 ⊂G
£ S ¤
≤ inf sup Pτ,µ DK 0 ≤ t
n∈N K 0 ∈K(E):K 0 ⊂Gn
£ S ¤ £ S ¤
≤ inf Pτ,µ DG n
≤ t = Pτ,µ DK ≤t . (3.241)
n∈N

It follows from (3.241) that


£ S ¤ £ S ¤
Pτ,µ DK ≤t = inf sup Pτ,µ DK 0 ≤ t . (3.242)
G∈O(E):G⊃K K 0 ∈K(E), K 0 ⊂G

Now it is clear that £the equality


¤ in (3.242) implies property (iii) for the
S
mapping
h K i→
7 P τ,µ D K ≤ t . The proof of (iii) for the mapping K 7→
e S
Pτ,µ D ≤ t is similar. Here we use part (d) in Lemma 3.34 (note that
K
part (d) of Lemma 3.34 also holds under the restrictions
£ S in
¤ Lemma 3.29).
Next we will prove that the function K 7→ Pτ,µ DK ≤ t satisfies condition
(ii). In the proof the following simple equalities will be used: for all Borel
subsets B1 and B2 of E,
S S S
DB 1 ∪B2
= DB 1
∧ DB 2
, and (3.243)
S S S
DB 1 ∩B2
≥ DB 1
∨ DB 2
. (3.244)
By using (3.243) and (3.244) with K1 ∈ K(E) and K2 ∈ K(E) instead of B1
and B2 respectively, we get:
© S ª © S ª ¡© S ª © S ª¢ © S ª
DK1 ∪K2 ≤ t \ DK 2
≤ t = DK 1
≤ t ∪ DK 2
≤ t \ DK 2
≤t
© S ª © S ª © S ª ¡© S ª © S ª¢
= DK 1
≤ t \ DK 2
≤ t = DK 1
≤ t \ DK 1
≤ t ∩ DK 2
≤t
© S ª © S S
ª © S ª © S ª
= DK 1
≤ t \ DK 1
∨ DK 2
≤ t ⊂ DK 1
≤ t \ DK 1 ∩K2
≤ t . (3.245)
It follows from (3.245) that
£ S ¤ £ S ¤
Pτ,µ DK 1 ∪K2
≤ t + Pτ,µ DK 1 ∩K2
≤t
£ S ¤ £ S ¤
≤ Pτ,µ DK 1
≤ t + Pτ,µ DK 2
≤t . (3.246)
Now£ it is clear
¤ that (3.246) implies condition (ii) for the function K 7→
S
Pτ,µ DK ≤ t . The proof of condition (ii) for the second function in Lemma
3.29 is similar.
This completes the proof of Lemma 3.29.
S
The next theorem states that under certain restrictions, the entry time DB ,
the pseudo-hitting time De , and the hitting time T are stopping times.
S S
B B
τ
Recall that Ft+ denote the completion of the σ-field
\
τ
Ft+ = σ (X(ρ) : τ ≤ ρ ≤ s)
s∈(t,T ]

with respect to the family of measures {Ps,x : 0 ≤ s ≤ τ, x ∈ E}.


3.5 Measurability properties of hitting times 155

Theorem 3.30. Let τ ∈ [0, T ], and {X(t) : t ∈ [τ, T ]} be as in Lemma 3.29:


(i) The process X(t) is right-continuous and quasi left-continuous on [0, ζ).
τ
(ii)The σ-fields Ft+ are Pτ,x -complete and right-continuous for t ∈ [τ, T ] and
x ∈ E.
τ
Then for every τ ∈ [0, T ) and every Ft+ -stopping time S : Ω → [τ, ζ], the
S eS τ
stochastic variables DB , DB , and TBS are Ft+ -stopping times.

Proof. We will first prove Theorem 3.30 assuming that it holds for all open
and all compact subsets of E. The validity of Theorem 3.30 for such sets will
be established in lemmas 3.31 and 3.32 below.
Let B be a Borel subset of E, and suppose that we have already shown
(ε+S)∧ζ τ
that for any ε ≥ 0 the stochastic time DB is an Ft+ -stopping time. Since
(ε+S)∧ζ
TBS = inf DB
ε>0,ε∈Q+

τ
(see (3.235)), we also obtain that TBS is an Ft+ -stopping time. Therefore, in
τ
order to prove that TBS is an Ft+ -stopping time, it suffices to show that for
S τ
every Borel subset B of E, the stochastic time DB is an Ft+ -stopping time.
Since the process t 7→ X(t) is continuous from the right, it suffices to prove
the previous assertion with S replaced by (ε + S) ∧ ζ.
Fix t ∈ [τ, T ), µ ∈ P (E), and B ∈ E. By Lemma 3.29 and the Choquet
capacitability theorem, the set B is capacitable with respect to the£ capacity
¤
S
I associated with the strongly sub-additive function K 7→ Pτ,µ DK ≤t .
Therefore, there exists an increasing sequence Kn ∈ K(E), n ∈ N, and a
decreasing sequence Gn ∈ O(E), n ∈ N, such that

Kn ⊂ Kn+1 ⊂ B ⊂ Gn+1 ⊂ Gn , n ∈ N, and


£ S ¤ £ S ¤
sup Pτ,µ DK n
≤ t = inf Pτ,µ DG n
≤t . (3.247)
n∈N n∈N

The arguments in (3.247) should be compared with those in (3.262) below.


Put
[© ª \© ª
Λτ,µ,S
1 (t) = S
DK n
≤t and Λτ,µ,S
2 (t) = DGS
n
≤ t . (3.248)
n∈N n∈N

τ
Then Lemma 3.31 implies Λτ,µ,S
2 (t) ∈ Ft+ , and Lemma 3.32 gives Λτ,µ,S
1 (t) ∈
τ
Ft+ . Moreover, we have
© S ª
Λτ,µ,S
1 (t) ⊂ DB ≤ t ⊂ Λτ,µ,S
2 (t), (3.249)

and
h i £ S ¤
Pτ,µ Λτ,µ,S
2 (t) = inf Pτ,µ DG n
≤t
n∈N
156 3 Space-time operators
£ S ¤ h i
= sup Pτ,µ DK n
≤ t = Pτ,µ Λτ,µ,S
1 (t) . (3.250)
n∈N
h i
It follows from (3.249) and (3.250) that Pτ,µ Λτ,µ,S 2 (t) \ Λτ,µ,S
1 (t) = 0. By
© S ª
using (3.249) again, we see that the event DB ≤ t belongs to the σ-field
τ S τ
Ft+ . Therefore, the stochastic time DB is an Ft+ -stopping time. As we have
τ
already observed, it also follows that the stochastic time TBS is an Ft+ -stopping
time.
A similar argument with DB S
replaced by D e S shows that the stochastic
B
τ
times D e S , B ∈ E, are F -stopping times.
B t+
This completes the proof of Theorem 3.30.

Next we will prove two lemmas which have already been used in the proof of
Theorem 3.30.
τ
Lemma 3.31. Let S : Ω → [τ, ζ] be an Ft+ -stopping time, and let G ∈ O(E).
S eS τ
Then the stochastic times DG , DG , and TGS are Ft+ -stopping times.

Proof. It is not hard to see that


© S ª \ ½ 1
¾
S
DG ≤ t < ζ = DG < t + ∩ {t < ζ}
m
m∈N
\ [
= {S ≤ ρ, X(ρ) ∈ G} . (3.251)
m∈N τ ≤ρ<t+ m
1
,ρ∈Q+

We also have
© S ª © S ª © S ª
DG ≤ t = DG ≤ t < ζ ∪ {ζ ≤ t} = DG ≤ t < ζ ∪ {X(t) = 4} .
(3.252)
τ
The event on the right-hand side of (3.251) belongs to Ft+ , and hence from
S τ
(3.251) and (3.252) the stochastic time DG is an Ft+ -stopping time. The fact
τ
e S is an Ft+ -stopping time follows from
that D G

n o \ ½ 1
¾
e S
DG ≤ t < ζ = e S
DG < t + ∩ {t < ζ}
m
m∈N
\ [
= {S ≤ ρ, X(ρ) ∈ G}
m∈N ρ∈(τ,t+ m
1
)∩Q+

together with
n o n o
DeG
S
≤t = DeG
S
≤ t < ζ ∪ {X(t) = 4} . (3.253)

τ
The equality (3.235) with G instead of B implies that TGS is an Ft+ -stopping
time.
3.5 Measurability properties of hitting times 157
τ
Lemma 3.32. Let S : Ω → [τ, ζ] be an Ft+ -stopping time, and let K ∈
¡ ¢ τ
K E 4 . Then the S
stochastic times DK e S and T S are
,D Ft+ -stopping times.
K K

Proof. First let K be a compact subset of E, and let Gn , n ∈ N, be a sequence


of open subsets of E with the following properties: K ⊂ Gn+1 ⊂ Gn and
T S τ
n∈N Gn = K. Then every stochastic time DGn is an F t+ -stopping time (see
S
Lemma 3.31), and for every µ ∈ P (E) the sequence of stochastic times DG n
,
S
n ∈ N, increases Pτ,µ -almost surely to DK . This implies that the stochastic
S τ
time TK is an Ft+ -stopping time. The equality (3.235) with K instead of B
S τ
then shows that TK is an Ft+ -stopping time. Next we will show the Pτ,µ -
S S
almost sure convergence of the sequence DG n
, n ∈ N. Put DK = sup DG n
.
n∈N
S S S S
Since DG n
≤ DG n+1
≤ DK , it follows that DK ≤ DK . By Lemma 3.31, the
S τ
stochastic times DG n
, n ∈ N, are Ft+ -stopping times. It follows from the
quasi-continuity from the left of the process X(t), t ∈ [0, ζ), that
¡ S ¢
lim X DG n
= X (DK ) Pτ,µ -a.s.
n→∞

Therefore, \
X (DK ) ∈ Gn = K Pτ,µ -a.s.
n
S S S
Since DK ≥ S, we have DK ≤ DK Pτ,µ -almost surely, and hence DK = DK
Pτ,µ -almost surely. This establishes the Pτ,µ -almost sure convergence of the
S S
sequence DG n
, n ∈ N, to DK .
In order to finish the proof of Lemma 3.32, we will establish that for every
µ ∈ P (E), the sequence of stochastic times D e S increases Pτ,µ -almost surely
Gn
e S e e S
to DK . Put DK = sup DGn . Since
n∈N

eG
D S eG
≤D S eK
≤D S
,
n n+1

eK ≤ D
it follows that D e S . By using the fact that the process X(t), t ∈ [0, ζ),
K
is quasi-continuous from the left, we get
³ ´ ³ ´
lim X D eG
S
= X eK
D Pτ,µ -a.s.
n
n→∞

Therefore ³ ´ \
X DeK ∈ Gn = K Pτ,µ -a.s.
n

Since D e S ≥ S, we have D eS ≤ D e K Pτ,µ -almost surely, and hence D


eS = D eK
K K K
e S
Pτ,µ -almost surely. This equality shows that the stochastic time DK is an
τ
Ft+ -stopping time.
We still have to consider the case that 4 ∈ K. For this we use the equalities
DKS
= DKS
∧ ζ, and DeS eS
0 ∪{4} 0 K0 ∪{4} = DK0 ∧ ζ together with the fact that
158 3 Space-time operators

a compact subset K of E 4 is a compact subset of E or is of the form K =


K0 ∪ {4} where K0 ⊂ E is compact. Observe that on the event {ζ ≥ τ } ζ is
an Ftτ -stopping time.
This completes the proof of Lemma 3.32.

Let us return to the study of standard Markov processes. It was established in


Theorem 1.39 item (a) that if P is a transition sub-probability function such
that the backward free Feller propagator {P (s, t) : 0 ≤ s ≤ t ≤ T } associated
with P is a strongly continuous (backward) Feller propagator, then there
exists a standard Markov process as in 3.267 with (τ, x; t, B) 7→ P (τ, x; t, B)
as its transition function. Let τ ∈ [0, T ], and let (X(t), Ftτ , Pτ,x ) be a Markov
τ
process. Suppose that S is an Ft+ -stopping time such that τ ≤ S ≤ ζ. Fix a
S,∨
measure µ ∈ P (E), and denote by FT the completion of the σ-field FTS,∨ =
σ (S ∨ ρ, X (S ∨ ρ) : 0 ≤ ρ ≤ T ) with respect to the measure µ. The measure µ
is used throughout Lemma 3.34 below. The next theorem provides additional
examples of families of stopping times which can be used in the formulation
of the strong Markov property with respect to families of measures.
Theorem 3.33. Let (X(t), Ftτ , Pτ,x ) be a standard Markov process as in
(3.267), and let B ∈ E4 . Then the stopping times DB S e S , and T S are
, D B B
S,∨
measurable with respect to the σ-field FT .

Proof. Since the stopping time S attains its values in the interval [τ, ζ] we see
that {ζ ≤ ρ} = {ζ ≤ ρ ∨ S} = {X (ρ ∨ S) = 4} for all ρ ∈ [τ, T ]. This shows
S,∨ S S
that ζ is measurable with respect to FT . By (3.234) we see DB∪{4} = DB ∧ζ,
De S e S
= D ∧ ζ, and T S S
= T ∧ ζ, and hence we see that it suffices
B∪{4} B B∪{4} B
S
to prove that the stochastic times DB e S , and T S are FS,∨
, D T -measurable,
B B
whenever B is a Borel subset of E.

The proof of Theorem 3.33 is based on the following lemma. The same result
with the same proof is also true with E 4 instead of E.
Lemma 3.34. Let K ∈ K(E) and τ ∈ [0, T ). Suppose
T that Gn ∈ O(E),
n ∈ N, is a sequence such that K ⊂ Gn+1 ⊂ Gn and n∈N Gn = K. Then the
following assertions hold:
S
(a) For every µ ∈ P (E), the sequence of stopping times DG n
increases and
S
tends to DK Pτ,µ -almost surely.
© S ª
(b) For every t ∈ [τ, T ], the events DG ≤ t , n ∈ N, are FTS,∨ -measurable,
© S ª S,∨
n

and the event DK ≤ t is FT -measurable.


© ª
(c) For every t ∈ [τ, T ], the events TGSn ≤ t , n ∈ N, are FTS,∨ -measurable,
© S ª S,∨
and the event TK ≤ t is FT -measurable.
(d) For every µ ∈ P (E), the sequence of stopping times D e S increases and
Gn
e S
tends to DK Pτ,µ -almost surely.
3.5 Measurability properties of hitting times 159
n o
(e) For every t ∈ [τ, T ], the events D e S ≤ t , n ∈ N, are FS,∨ -measurable,
Gn T
n o S,∨
e S
and the event DK ≤ t is FT -measurable.

Proof. (a) Fix µ ∈ P (E), and let K ∈ K(E) and Gn ∈ O(E), n ∈ N, be as


S
in assertion (a) in the formulation of Lemma 3.34. Put DK = sup DG n
. Since
n∈N
S S S S
S ≤ DG n
≤ DG n+1
≤ DK , we always have S ≤ DK ≤ DK . Moreover, DK
is a stopping time. By using the quasi-continuity from the left of the process
t 7→ X(t) on [τ, ζ) with respect to the measure Pτ,µ , we see that
¡ S ¢
lim X DG n
= X (DK ) Pτ,µ -almost surely on {DK < ζ}.
n→∞

Therefore,
\
X (DK ) ∈ Gn = K Pτ,µ -almost surely on {DK < ζ}. (3.254)
n∈N

S S S
Now by the definition of DK we have DK ≥ S, and (3.254) implies DK ≤ DK
S
Pτ,µ -almost surely on {DK < ζ}, and hence DK = DK Pτ,µ -almost surely. In
S
the final we used the inequality DK ≤ DK which is always true.
(b) Fix t ∈ [τ, T ) and n ∈ N. By the right-continuity of paths on [0, ζ) we
have
© S ª \ ½ 1
¾
S
D Gn ≤ t < ζ = D Gn < t + ∩ {t < ζ}
m
m∈N
\ [
= {S ≤ ρ, X(ρ) ∈ Gn }
m∈N ρ∈[τ,t+ 1
m)
\ [
= {S ∨ ρ ≤ ρ, X (S ∨ ρ) ∈ Gn } . (3.255)
m∈N ρ∈[τ,t+ m
1
)∩Q+

It follows that © ª
S
DG n
≤ t < ζ ∈ FTS,∨ , 0 ≤ t ≤ T.
By using assertion (a), we see that the events
© S ª \© ª
S
DK ≤ t < ζ and DG n
≤t<ζ
n∈N

© S ª S,∨
coincide Pτ,µ -almost surely. It follows that DK ≤ t < ζ ∈ FT . It also
© S ª S,∨
follows that the event DK < ζ belongs to FT . In addition we notice the
equalities
© S ª © S ª © S ª
DK ≤ t = DK ≤ t < ζ ∪ DK ≤ t, ζ ≤ t

S
(DK ≤ ζ and S ≤ ζ)
160 3 Space-time operators
© S ª
= DK ≤ t < ζ ∪ {ζ ≤ S ∨ t}
© S ª
= DK ≤ t < ζ ∪ {X (S ∨ t) = 4} (3.256)
© S ª
From (3.256) we see that events of the form DK ≤ t , t ∈ [τ, T ], belong to
S,∨ S S,∨
FT . Consequently the stopping time DK is FT -measurable. This proves
assertion (b).
(c) Since the sets Gn are open and the process X(t) is right-continuous,
the hitting times TGSn and the entry times DG S
n
coincide. Hence, the first part
of assertion (c) follows from assertion (b). In order to prove the second part
of (c), we reason as follows. By assertion (b), for every r ∈ Q+ , the stopping
(r+S)∧ζ (r+S)∧ζ,∨
time DK is FT -measurable. Our next goal is to prove that for
every ε > 0,
(ε+S)∧ζ,∨
FT ⊂ FTS,∨ . (3.257)
Fix ε > 0, and ρ ∈ [τ, ζ], and put S1 = ((ε + S) ∧ ζ) ∨ ρ. Observe that for ρ,
t ∈ [0, T ], we have the following equality of events:

{S1 ≤ t} = {((ε + S) ∧ ζ) ∨ ρ ≤ t}
= {((ε + S) ∨ ρ) ∧ (ζ ∨ ρ) ≤ t}
= {S ∨ (ρ − ε) ≤ t − ε, ρ ≤ t} ∪ {ζ ≤ S ∨ t, ρ ≤ t}
= {S ∨ (ρ − ε) ≤ t − ε, ρ ≤ t} ∪ {X (S ∨ t) = 4, ρ ≤ t} . (3.258)

Therefore, the stopping time S1 = ((ε + S) ∧ ζ) ∨ ρ is FTS,∨ -measurable. Since


the process t 7→ X(t) is right-continuous, it follows from Proposition 3.39 that
X (S1 ) is FTS,∨ -measurable. This implies inclusion (3.257). Hence,
(ε+S)∧ζ,∨ S,∨
FT ⊂ FT , (3.259)
(ε+S)∧ζ S,∨
and we see that the for every ε > 0 the stopping time DK is FT -
(ε+S)∧ζ S
measurable. Since the family DK , ε > 0, decreases to TK , the hitting
S S,∨
time TK is FT -measurable as well.
(d) Fix µ ∈ P (E), and let K ∈ K(E) and Gn ∈ O(E), n ∈ N, be as in
e K = sup D
assertion (a). Put D e S . Since
Gn
n∈N

eG
D S eG
≤D S eK
≤D S
,
n n+1

we have D eK ≤ De S . It follows from the quasi-continuity from the left of the


K
process X(t) on [0, ζ) that
³ ´ ³ ´ n o
lim X D eS e eK < ζ .
Gn = X DK Pτ,µ -almost surely on D
n→∞

Therefore,
3.5 Measurability properties of hitting times 161
³ ´ \ n o
eK ∈
X D Gn = K Pτ,µ -almost surely on DeK < ζ .
n
n o
Now D e S ≥ S implies that D eS ≤ De K Pτ,µ -almost surely on D e K < ζ , and
K K
n o
e S e
hence DK = DK Pτ,µ -almost surely on D e K < ζ . As in (a) we get D
eS = DeK
K
Pτ,µ -almost surely.
(e) Fix t ∈ [τ, T ) and n ∈ N. By the right-continuity of paths,
n o \ ½ 1
¾
DeGS
≤ t < ζ = D S
Gn < t + ∩ {t < ζ}
n
m
m∈N
\ [
= {S ≤ ρ, X(ρ) ∈ Gn }
m∈N ρ∈(τ,t+ m
1
)
\ [
= {S ∨ ρ ≤ ρ, X (S ∨ ρ) ∈ Gn } . (3.260)
m∈N ρ∈(τ,t+ m
1
)∩Q+

n o
It follows that D e S ≤ t < ζ ∈ FS,∨ . By using assertion (d), we see that the
Gn
n o T nT o
events D e ≤ t < ζ and
S e S ≤ t < ζ coincide Pτ,µ -almost surely.
D
K Gn
n o n∈NS,∨
e S
Therefore, DK ≤ t < ζ ∈ FT . As in (3.256) we have
n o n o n o
DeK
S
≤t = DeK
S
≤t<ζ ∪ D eKS
≤ t, ζ ≤ t
n o
= DeK
S
≤ t < ζ ∪ {X (S ∨ t) = 4} . (3.261)

This proves assertion (e), and therefore the proof of Lemma 3.34 is complete.

Proof (Proof of Theorem 3.33: continuation). Let us return to the proof of


S
Theorem 3.33. We will first prove that for any Borel set B, the entry time DB
S,∨
is measurable with respect to the σ-field FT . Then the same assertion holds
S,∨
for the hitting time TBS . Indeed, if DB
S
is FT -measurable for all stopping times
(ε+S)∧ζ
S, then for every ε > 0, the stopping time DB is measurable with respect
(ε+S)∧ζ,∨ S,∨
to the σ-field FT . By using (3.259), we obtain the FT -measurability
(ε+S)∧ζ S,∨
of DB . Now (3.235) implies the FT -measurability of TBS .
Fix t ∈ [τ, T ), µ ∈ P (E), and B ∈ E. By Lemma£ S 3.29,
¤ the set B is
capacitable with respect to the capacity K 7→ Pτ,µ DK ≤ t . Notice that the
following argument was also employed in the proof of Theorem 3.30. Therefore,
there exists an increasing sequence Kn ∈ K(E), n ∈ N, and a decreasing
sequence Gn ∈ O(E), n ∈ N, such that

Kn ⊂ Kn+1 ⊂ B ⊂ Gn+1 ⊂ Gn , n ∈ N, and


£ S ¤ £ S ¤
sup Pτ,µ DK n
≤ t = inf Pτ,µ DG n
≤t . (3.262)
n∈N n∈N
162 3 Space-time operators

Next we put
[© ª \© ª
Λτ,µ,S
1 (t) = S
DK n
≤t and Λτ,µ,S
2 (t) = S
DG n
≤t . (3.263)
n∈N n∈N

The equalities in (3.248) which are the same as those in (3.263) show that the
S,∨
events Λτ,µ,S
1 (t) and Λτ,µ,S
2 (t) are FT -measurable. Moreover, we have
© S ª
Λτ,µ,S
1 (t) ⊂ DB ≤ t ⊂ Λτ,µ,S
2 (t), (3.264)
and
h i £ S ¤
Pτ,µ Λτ,µ,S
2 (t) = inf Pτ,µ DG n
≤t
n∈N
£ S ¤ h i
= sup Pτ,µ DK n
≤ t = Pτ,µ Λτ,µ,S
1 (t) . (3.265)
n∈N
h i
Now (3.264) and (3.265) give Pτ,µ Λτ,µ,S
2 (t) \ Λτ,µ,S
1 (t) = 0. By using (3.264),
© S ª S,∨
we see that the event DB ≤ t is measurable with respect to the σ-field FT .
S,∨ S
This establishes the FT -measurability of the entry time DB and the hitting
S e S is similar
time TB . The proof of Theorem 3.33 for the pseudo-hitting time D B
S
to that for the entry time DB .
The proof of Theorem 3.33 is thus completed.
Definition 3.35. Fix τ ∈ [0, T ], and let S1 : Ω → [τ, T ] be an (Ftτ )t∈[τ,T ] -
stopping time. A stopping time S2 : Ω → [τ, T ] is called terminal after S1 if
S1 ,∨
S2 ≥ S1 , and if S2 is FT -measurable.
The following corollary shows that entry and hitting times of Borel subsets
which are comparable are terminal after each other.
Corollary 3.36. Let (X(t), Ftτ , Pτ,x ) be a standard process, and let A and B
τ
be Borel subsets of E with B ⊂ A. Then the entry time DB is measurable with
D τ ,∨
respect to the σ-field FT A . Moreover, the hitting time TBτ is measurable with
T τ ,∨
respect to the σ-field FTA .
Proof. By Theorem 3.33, it suffices to show that the equalities
τ τ
D τ e TA = TBτ
DB A = DB and D B (3.266)
hold Pτ,µ -almost surely for all µ ∈ P (E). The first equality in (3.266) follows
from [ [
τ
{DA ≤ s, X(s) ∈ B} = {X(s) ∈ B} ,
τ ≤s<T τ ≤s<T

while the second equality in (3.266) can be obtained from


[ [
{TAτ ≤ s, X(s) ∈ B} = {X(s) ∈ B} .
τ <s<T τ <s<T

This proves Corollary 3.36.


3.5 Measurability properties of hitting times 163

τ
It follows from Corollary 3.36 that the families {DA : A ∈ E} and {TAτ : A ∈ E}
can be used in the definition of the strong Markov property in the case of
standard processes. The next theorem states that the strong Markov property
holds for entry times and hitting times of comparable Borel subsets.
Theorem 3.37. Let (X(t), Ftτ , Pτ,x ) be a standard process, and fix τ ∈ [0, T ].
Let A and B be Borel subsets of E such that B ⊂ A, and let f : [τ, T ]×E 4 → R
be a bounded Borel function. Then the following equalities hold Pτ,x -almost
surely:
h ¯ τ i
Eτ,x f (DBτ
, X (DB τ
)) ¯ FD τ
A
= EDτ ,X (Dτ ) [f (DB τ
, X (DBτ
))] and
A A
h ¯ i
Eτ,x f (TBτ , X (TBτ )) ¯ FTτ Aτ = ET τ ,X (T τ ) [f (TBτ , X (TBτ ))]
A A

© S ª
The
© S firstª one holds Pτ,x -almost surely on DA < ζ , and the second on
TA < ζ .

Proof. Theorem 3.37 follows from Corollary 3.36 and Remark 3.40.

Definition 3.38. The quadruple


½µ ³ ´ ¶ ¾
τ
Ω, Ft+ , Pτ,x , (X(t), t ∈ [0, T ]) , (∨t , t ∈ [0, T ]) , (E, E)
t∈[τ,T ]
(3.267)
is called a standard Markov process if it possesses the following properties:
³ τ ´
1. The process X(t) is adapted to the filtration Ft+ , right-continuous
t∈[τ,T ]
and possesses left limits in E on its life time.
τ
2. The σ-fields Ft+ , t ∈ [τ, T ], are right continuous and Pτ,x -complete.
3. The process (X(t) : t ∈ [0, T ]) is strong Markov with respect to the mea-
sures {Pτ,x : (τ, x) ∈ [0, T ] × E}
4. The process (X(t) : t ∈ [0, T ]) is quasi left-continuous on [0, ζ).
5. The equalities X(t) ◦ ∨s = X(t ∨ s) hold Pτ,x -almost surely for all (τ, x) ∈
[0, T ] × E and for s, t ∈ [τ, T ].
¡ ¢
If Ω = D [0, T ], E 4 and X(t)(ω) = ω(t), t ∈ [0, T ], ω ∈ Ω, then parts of
the items (1) and (2) are automatically satisfied. For brevity we often write
(X(t), Pτ,x ) instead of (3.267).
The following proposition gives an alternative way to describe stopping
times which are terminal after another stopping time: see Definition 3.35.
Proposition 3.39. Let S1 : Ω → [τ, T ] be an Ftτ -stopping time, and let the
stopping S2 : Ω → [τ, T ] be such that S2 ≥ S1 , and such that for every
t ∈ [τ, T ] the event {S2 > t} restricted to the event {S1 < t} = {S1 ∨ t < t}
only depends on FTt . Then S2 is FTS1 ,∨ -measurable. If the paths of the process X
are right-continuous, the state variable X (S2 ) is FTS1 ,∨ -measurable as well. It
follows that the space-time variable (S2 , X (S2 )) is FTS1 ,∨ -measurable. Similar
164 3 Space-time operators

results are true if the σ-fields FTt and FTS1 ,∨ by their Pτ,µ -completions for some
probability measure µ on E.

Proof. Suppose that for every t ∈ [τ, T ] the stochastic variable S2 is such that
on {S1 < t} = {S1 ∨ t < t} the event {S2 > t} only depends on FTt . Then
on {S1 < t} the event © {S2 > t} only depends ª © on the σ-field generated by ª
the state variables X(ρ) ¹{S1 ∨t<t} : ρ ≥ t = X (ρ ∨ S1 ) ¹{S1 ∨t<t} : ρ ≥ t .
Consequently, the event {S2 > t > S1 } is FTS1 ,∨ -measurable. Since S2 = S1 +
RT
τ
1{S2 >t>S1 } dt, we see that S2 is FTS1 ,∨ -measurable. This argument can be
adapted if we only know that for every t ∈ [τ, T ] on the event {S1 < t} the
event {S2 > t} only depends
© on the Pτ,µ -completion
ª of the σ-field generated
by the state variables X(ρ) ¹{S1 ∨t<t} : ρ ≥ t for some probability measure
µ on E.
If the process X(t) is right-continuous, and if S2 is a stopping time which
is terminal after the stopping time S1 : Ω → [0, T ], then the space-time
S1 ,∨
variable (S2 , X (S2 )) is FT -measurable. This result follows from the equality
in (2.44) with S2 instead of S:
» ¼
t − τ 2n (S2 − τ )
S2,n (t) = τ + n . (3.268)
2 t−τ

Then notice that the stopping times S2,n (t), n ∈ N, t ∈ (τ, T ], are FTS1 ,∨ -
measurable, provided that S2 has this property. Moreover, we have S2 ≤
S2,n+1 (t) ≤ S2,n (t) ≤ S2 + 2−n (t − τ ). It follows that the state variables
X (S2,n (t)), n ∈ N, t ∈ (τ, T ], are FTS1 ,∨ -measurable, and that the same is true
for X (S2 ) = lim X (S2,n (t)).
n→∞
This completes the proof of Proposition 3.39.

Remark 3.40. If in Theorem¡ 3.41 for


¢ the sample path space Ω we take the
Skorohod space Ω = D [0, T ], E 4 , X(t)(ω) = ω(t), ω ∈ Ω, t ∈ [0, T ], then
the process t 7→ X(t), t ∈ [0, T ], is right-continuous, has left limits in E on its
life time, and is quasi-left continuous on its life time as well.

Theorem 3.41. Let, as in Lemma 3.29,


n³ τ
´ o
Ω, FT , Pτ,x , (X(t), t ∈ [0, T ]) , (∨t , t ∈ [τ, T ]) , (E, E)

be a standard Markov process with right-continuous paths, which has left lim-
its on its life time, and is quasi-continuous from the left on its life time.
S,∨
For fixed (τ, x) ∈ [0, T ] × E, the σ-field FT is the completion of the σ-
field FTS,∨ = σ (S ∨ ρ, X (S ∨ ρ) : 0 ≤ ρ ≤ T ) with respect to the measure Pτ,x .
S1 ,∨
Then, if (S1 , S2 ) is a pair of stopping times such that S2 is FT -measurable
and τ ≤ S1 ≤ S2 ≤ T , then for all bounded Borel functions f on [τ, T ] × E 4 ,
the equality
3.5 Measurability properties of hitting times 165
h ¯ τ i
Eτ,x f (S2 , X (S2 )) ¯ FS1 = ES1 ,X(S1 ) [f (S2 , X (S2 ))] (3.269)

holds Pτ,x -almost surely on {S1 < ζ}.


First notice that the conditions on S1 and S2 are such that S2 is terminal
after S1 : see Definition 3.35. Also observe that the Markov process in (3.267) is
quasi-continuous from the left on its life time [0, ζ): see compare with Theorem
1.39 item (a). Let A and B be Borel subsets of E such that B ⊂ A. In (3.269)
τ τ
we may put S1 = DA together with S2 = DB , or S1 = TAτ and S2 = TBτ : see
Theorem 3.33 and Corollary 3.36.
Proof. This result is a consequence of the strong Markov property as exhibited
in Theorem 1.39 item (a).

3.5.1 Some side remarks

We notice that we have used the following version of the Choquet capacity
theorem.
Theorem 3.42. In a polish space every analytic set is capacitable.
For a discussion on capacitable subsets see e.g. Kiselman [133]; see Choquet
[58], [27] and [69] as well. For a general discussion on the foundations of
probability theory see e.g. [118].
Without the sequential λ-dominance of the operator D+ L the second
formula, i.e. the formula in (3.116), poses a difficulty as far as it is not
clear that the function e−λt Se0 (t)f belongs to Cb ([0, T ] × E) indeed. For
the moment suppose that the function f ∈ Cb ([0, T ] × E) is such that
Se0 (t)f ∈ Cb ([0, T ] × E). Then equality (3.115) yields:
³ ´−1 Z t ³ ´−1
λI − L (1)
f= e−λρ Se0 (ρ)f dρ + e−λt S0 (t) λI − L(1) f
0
Z t ³ ´−1
= e−λρ Se0 (ρ)f dρ + e−λt λI − L(1) Se0 (t)f. (3.270)
0
Rt ¡ ¢
Consequently, the function 0 e−λρ Se0 (ρ)f dρ belongs to D L(1) and the
equality in (3.116) follows from (3.270). Next, let (µm )m∈N be a sequence
in (0, ∞) which increases to ∞, and let (fn )n∈N be sequence in Cb ([0, T ] × E)
which decreases pointwise to the zero-function. From (3.113), (3.270) and
(3.116) we obtain the following equality:
³ ´Z t
(1)
µm R (µm ) fn = λI − L e−λρ S0 (ρ) (µm R (µm ) fn ) dρ
0
+ e−λt S0 (t) (µm R (µm ) fn ) , m, n ∈ N, and (3.271)
Z t
µm R (µm ) R(λ)fn = e−λρ S0 (ρ) (µm R (µm ) fn ) dρ
0
166 3 Space-time operators

+ e−λt S0 (t) (µm R (µm ) R(λ)fn ) , m, n ∈ N. (3.272)

Perhaps it is better to consider the following equalities rather than those in


(3.272). They also follow from the equalities in (3.113), (3.115) and (3.116):
R t2 R t −λρ R t2 −λt
t1 0
e S0 (ρ)dρ f dt e S0 (t)R(λ)f dt
R(λ)f = + t1 , (3.273)
t2 − t1 t2 − t1

and
R t2 R t R t2
³ ´ e−λρ S0 (ρ)dρ f dt e−λt S0 (t)f dt
(1) t1 0 t1
f = λI − L + . (3.274)
t2 − t1 t2 − t1
We have to investigate the equalities in (3.273) and (3.274) if for f we choose
a function fn from a sequence (fn )n∈N which decreases to zero. Then
R t2 −λt R t2 −λt
t1
e S0 (t)fn dt t1
e S0 (t)fn dt
inf sup = lim sup .
n∈N 0≤t1 <t2 ≤T t2 − t1 n→∞ 0≤t1 <t2 ≤T t2 − t1
(3.275)
From (3.274) we infer that
R t2 R t −λρ
t1 0
e S0 (ρ)dρ fn dt
lim sup = 0. (3.276)
n→∞ 0≤t1 <t2 ≤T t2 − t1

From (3.275) and our extra assumption we see that the limit in (3.276) van-
ishes, and hence that the semigroup {S0 (t) : t ≥ 0} consists of linear mappings
which leave the function space Cb ([0, T ] × E) invariant, and which is Tβ -equi-
continuous.
The operator L(1) is the Tβ -closure of the operator D1 +L, which is positive
Tβ -dissipative. Hence the operator L(1) inherits this property, and so it is
positive Tβ -dissipative as well.
The operator D1 + L is Tβ -densely defined, is Tβ -dissipative, satisfies the
maximum principle, and there exists λ0 > 0 such that the range of λ0 I −D1 −L
is Tβ -dense in Cb ([0, T ] × E).
Moreover, we have
(1) (1) ©¡ ¢ ¡ ¢ ª
CP,b = ∩λ0 >0 CP,b (λ0 ) = ∩λ0 >0 λ0 I − L − D1 g : g ∈ D L ∩ D (D1 ) .
(3.277)
The second equality in (3.277) follows
¡ ¢from (3.91)
¡ ¢and (3.92).
Consider for functions f ∈ D L(1) , g ∈ D L ∩ D (D1 ), and λ > 0 the
equalities:

L(1) f − Lg − D1 g
= L(1) f − λR(λ)L(1) f + λ2 R(λ)f − λf − Lg − D1 g
Z ∞
¡ ¡ ¢¢ ³ (1) ´
= e−ρ I − S λ−1 ρ L f dρ
0
3.5 Measurability properties of hitting times 167
Z ∞ ¡ ¡ ¢ ¢
+λ e−ρ S λ−1 ρ − I (f − g) dρ
0
Z ∞ (¡ ¡ ¢ ¢ )
−1
S λ ρ ϑλ −1 ρ − I g
+ ρe−ρ − Lg dρ
0 λ−1 ρ
Z ∞ á ¢ !
−ρ
¡ −1 ¢ I − ϑλ−1 ρ g
+ ρe S λ ρ − D1 g dρ
0 λ−1 ρ
Z ∞
© ¡ ¢ ª
+ ρe−ρ S λ−1 ρ D1 g − D1 g dρ. (3.278)
0

We recall that the time shift operators were defined in (3.77).

3.5.2 Some related remarks

In subsection 2.1.6 we already discussed to some length topics related to Ko-


rovkin families and convergence properties of measures. Here we will say some-
thing about the maximum principle, the martingale problem, and stopping
time arguments.
For more general versions than our Choquet capacitability theorem 3.28
the reader is referred to e.g. [73, 69, 160]). For a discussion on capacitable
subsets see e.g. Kiselman [133]; see Choquet [58], [27] and [69] as well. For a
general discussion on the foundations of probability theory see e.g. [118]. In
[95] the authors also made a thorough investigation of measurability proper-
ties of stopping times. However, in that case the underlying state space was
locally compact. In [48] the author makes an extensive study of the maximum
principle of an unbounded operator with domain and range in the space of
continuous functions which vanish at infinity where the state space is locally
compact. As indicated in Chapter 1 an operator L for which the martingale
problem is well-posed need posses a unique extension which is the generator
of a Dynkin-Feller semigroup. As indicated by Kolokoltsov in [136] there exist
relatively easy counter-examples: see comments after Theorem 1.39 in §1.3.
For the time-homogeneous case see, e.g., [84] or [109]. In fact [84] contains a
general result on operators with domain and range in C0 (E) and which have
unique linear extensions generating a Feller-Dynkin semigroup. The martin-
gale problem goes back to Stroock and Varadhan (see [225]). It found numer-
ous applications in various fields of mathematics. We refer the reader to [147],
[136], and [135] for more information about and applications of the martingale
problem. In [80] the reader may find singular diffusion equations which possess
or which do not possess unique solutions. Consequently, for (singular) diffusion
equations without unique solutions the martingale problem is not uniquely
solvable. Another valuable source of information is Jacob [110, 111, 112].
Other relevant references are papers by Hoh [104, 106, 105, 107]. Some of
Hoh’s work is also employed in Jacob’s books. In fact most of these references
discuss the relations between pseudo-differential operators (of order less than
168 3 Space-time operators

or equal to 2), the corresponding martingale problem, and being the generator
of a Feller-Dynkin semigroup.
Part II

Backward Stochastic Differential Equations


4
Feynman-Kac formulas, backward stochastic
differential equations and Markov processes

In this chapter we explain the notion of stochastic backward differential equa-


tions and its relationship with classical (backward) parabolic differential equa-
tions of second order. The chapter contains a mixture of stochastic processes
like Markov processes and martingale theory and semi-linear partial differen-
tial equations of parabolic type. Some emphasis is put on the fact that the
whole theory generalizes Feynman-Kac formulas. A new method of proof of
the existence of solutions is given. All the existence arguments are based on
rather precise quantitative estimates.

4.1 Introduction
This introduction serves as a motivation for the present chapter and also for
Chapter 5. Backward stochastic differential equations, in short BSDE’s, have
been well studied during the last ten years or so. They were introduced by
Pardoux and Peng [184], who proved existence and uniqueness of adapted
solutions, under suitable square-integrability assumptions on the coefficients
and on the terminal condition. They provide probabilistic formulas for solu-
tion of systems of semi-linear partial differential equations, both of parabolic
and elliptic type. The interest for this kind of stochastic equations has in-
creased steadily, this is due to the strong connections of these equations
with mathematical finance and the fact that they gave a generalization of
the well known Feynman-Kac formula to semi-linear partial differential equa-
tions. In the present chpter we will concentrate on the relationship between
time-dependent strong Markov processes and abstract backward stochastic
differential equations. The equations are phrased in terms of a martingale
problem, rather than a stochastic differential equation. They could be called
weak backward stochastic differential equations. Emphasis is put on existence
and uniqueness of solutions. The paper [246] deals with the same subject, but
it concentrates on comparison theorems and viscosity solutions. The proof of
the existence result is based on a theorem which is related to a homotopy
172 4 BSDE’s and Markov processes

argument as pointed out by the authors of [63]. It is more direct than the
usual approach, which uses, among other things, regularizing by convolution
products. It also gives rather precise quantitative estimates.
For examples of strong solutions which are driven by Brownian motion
the reader is referred to e.g. section 2 in Pardoux [181]. If the coefficients
x 7→ b(s, x) and x 7→ σ(s, x) of the underlying (forward) stochastic differential
equation are linear in x, then the corresponding forward-backward stochastic
differential equation is related to option pricing in financial mathematics. The
backward stochastic differential equation may serve as a model for a hedging
strategy. For more details on this interpretation see e.g. El Karoui and Quenez
[126], pp. 198–199. A rather recent book on financial mathematics in terms of
martingale theory is the one by Delbaen and Schachermeyer [68]. E. Pardoux
and S. Zhang [185] use BSDE’s to give a probabilistic formula for the solution
of a system of parabolic or elliptic semi-linear partial differential equation
with Neumann boundary condition. In [40] the authors also put BSDE’s at
work to prove a result on a Neumann type boundary problem.
In this chapter we want to consider the situation where the family of
operators L(s), 0 ≤ s ≤ T , generates a time-inhomogeneous Markov process

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)} (4.1)

in the sense that


d
Eτ,x [f (X(s))] = Eτ,x [L(s)f (X(s))] , f ∈ D (L(s)) , τ ≤ s ≤ T.
ds
We consider the operators L(s) as operators on (a subspace of) the space of
bounded continuous functions on E, i.e. on Cb (E) equipped with the supre-
mum norm: kf k∞ = supx∈E |f (x)|, f ∈ Cb (E), and the strict topology Tβ .
With the operators L(s) we associate the squared gradient operator Γ1 defined
by

Γ1 (f, g) (τ, x)
1
= Tβ - lim Eτ,x [(f (X(s)) − f (X(τ ))) (g (X(s)) − g (X(τ )))] , (4.2)
s↓τ s − τ

for f , g ∈ D (Γ1 ). Here D (Γ1 ) is the domain of the operator Γ1 . It consists of


those functions f ∈ Cb (E) = Cb (E, C) with the property that the strict limit
1 h ³ ´i
Tβ - lim Eτ,x (f (X(s)) − f (X(τ ))) f (X(s)) − f (X(τ )) (4.3)
s↓τ s−τ

exists. We will assume that D (Γ1 ) contains an algebra of functions in


Cb ([0, T ] × E) which is closed under complex conjugation, and which is Tβ -
dense. These squared gradient operators are also called energy operators: see
e.g. Barlow, Bass and Kumagai [23]. We assume that every operator L(s),
0 ≤ s ≤ T , generates a diffusion in the sense of the following definition.
4.1 Introduction 173

In the sequel it is assumed that the family of operators {L(s) : 0 ≤ s ≤ T }


possesses the property that the space of functions u : [0, T ] × E → R with
∂u
the property that the function (s, x) 7→ (s, x) + L(s)u (s, ·) (x) belongs to
∂s
Cb ([0, T ] × E) := Cb ([0, T ] × E; C) is Tβ -dense in the space Cb ([0, T ] × E).
This subspace of functions is denoted by D(L), and the operator L is defined
by Lu(s, x) = L(s)u (s, ·) (x), u ∈ D(L). It is also assumed that the family
A is a core for the operator L. We assume that the operator L, or that the
family of operators {L(s) : 0 ≤ s ≤ T }, generates a diffusion in the sense of
the following definition.
Definition 4.1. A family of operators {L(s) : 0 ≤ s ≤ T } is said to generate
a diffusion if for every C ∞ -function Φ : Rn → R, with Φ(0, . . . , 0) = 0, and
every pair (s, x) ∈ [0, T ] × E the following identity is valid

L(s) (Φ (f1 , . . . , fn ) (s, ·)) (x) (4.4)


Xn
∂Φ
= (f1 , . . . , fn ) L(s)fj (x)
j=1
∂x j
n
1 X ∂2Φ
+ (f1 , . . . , fn ) (x)Γ1 (fj , fk ) (s, x)
2 ∂xj ∂xk
j,k=1

for all functions f1 , . . . , fn in an algebra of functions A, contained in the


domain of the operator L, which forms a core for L.
Generators of diffusions for single operators are described in Bakry’s lecture
notes [16]. For more information on the squared gradient operator see e.g. [19]
and [17] as well. Put Φ(f, g) = f g. Then (4.4) implies L(s) (f g) (s, ·)(x) =
L(s)f (s, ·)(x)g(s, x) + f (s, x)L(s)g(s, ·)(x)Γ1 (f, g) (s, x), provided that the
three functions f , g and f g belong to A. Instead of using the full strength of
(4.4), i.e. with a general function Φ, we just need it for the product (f, g) 7→ f g:
see Proposition 4.24.
Remark 4.2. Let m be a reference measure on the Borel field E of E, and let
p ∈ [1, ∞]. If we consider the operators L(s), 0 ≤ s ≤ T , in Lp (E, E, m)-
space, then we also need some conditions on the algebra A of “core” type in
the space Lp (E, E, m). For details the reader is referred to Bakry [16].
By definition the gradient of a function u ∈ D (Γ1 ) in the direction of v ∈
D (Γ1 ) is the function (τ, x) 7→ Γ1 (u, v) (τ, x). For given (τ, x) ∈ [0, T ] × E
the functional v 7→ Γ1 (u, v) (τ, x) is linear: its action is denoted by ∇L u (τ, x).
Hence, for (τ, x) ∈ [0, T ] × E fixed, we can consider ∇L u (τ, x) as an element
in the dual of D (Γ1 ). The pair
¡ ¢
(τ, x) 7→ u (τ, x) , ∇Lu (τ, x)

may be called an element in the phase space ¡ of the family L(s), 0 ≤¢ s ≤ T ,


(see Jan Prüss [193]), and the process s 7→ u (s, X(s)) , ∇L
u (s, X(s)) will be
174 4 BSDE’s and Markov processes

called an element of the stochastic phase space. Next let f : [0, T ] × E × R ×



D (Γ1 ) → R be a “reasonable” function, and consider, for 0 ≤ s1 < s2 ≤ T
the expression:
Z s2
¡ ¢
u (s2 , X (s2 )) − u (s1 , X (s1 )) + f s, X(s), u (s, X(s)) , ∇Lu (s, X(s)) ds
s1
Z s2 µ ¶
∂u
− u (s2 , X (s2 )) + u (s1 , X (s1 )) + L(s)u (s, X(s)) + (s, X(s)) ds
s1 ∂s
(4.5)
Z s2
¡ ¢
= u (s2 , X (s2 )) − u (s1 , X (s1 )) + f s, X(s), u (s, X(s)) , ∇Lu (s, X(s)) ds
s1
− Mu (s2 ) + Mu (s1 ) , (4.6)

where

Mu (s2 ) − Mu (s1 )
Z s2 µ ¶
∂u
= u (s2 , X (s2 )) − u (s1 , X (s1 )) − L(s)u (s, X(s)) + (s, X(s)) ds
s1 ∂s
Z s2
= dMu (s). (4.7)
s1

Details on the properties of the function f will be given in the theorems 4.26,
4.30, 4.33, 4.34, and 4.42.
The following definition also occurs in Definition 1.29. In Definition 1.29
the reader will find more details about the definitions 4.3 and 4.4. It also
explains the relationship with transition probabilities and Feller propagators.
Definition 4.3. The process

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)} (4.8)

is called a time-inhomogeneous Markov process if


£ ¯ ¤
Eτ,x f (X(t)) ¯ Fsτ = Es,X(s) [f (X(t))] , Pτ,x -almost surely. (4.9)

Here f is a bounded Borel measurable function defined on the state space E


and τ ≤ s ≤ t ≤ T .
Suppose that the process X(t) in (4.8) has paths which are right-continuous
and have left limits in E. Then it can be shown that the Markov property
for fixed times carries over to stopping times in the sense that (4.9) may be
replaced with
£ ¯ ¤
Eτ,x Y ¯ FSτ = ES,X(S) [Y ] , Pτ,x -almost surely. (4.10)

Here S : E → [τ, T ] is an Ftτ -adapted stopping time and Y is a bounded


stochastic variable which is measurable with respect to the future (or terminal)
4.1 Introduction 175

σ-field after S, i.e. the one generated by {X (t ∨ S) : τ ≤ t ≤ T }. For this type


of result the reader is referred to Chapter 2 in Gulisashvili et al [95] and to
item (a) in Theorem 1.39. Markov processes for which (4.10) holds are called
strong Markov processes.
The following definition is, essentially speaking, the same as Definition
1.31. Its relationship with Feller propagators or evolutions (see Chapter 1,
Definition 1.30) is explained in Proposition 3.1 in Chapter 3. The derivatives
and the operators L(s), s ∈ [0, T ], have to be taken with respect to the strict
topology: see Section 1.1.
Definition 4.4. The family of operators L(s), 0 ≤ s ≤ T , is said to generate
a time-inhomogeneous Markov process

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)} (4.11)

if for all functions u ∈ D(L), for all x ∈ E, and for all pairs (τ, s) with
0 ≤ τ ≤ s ≤ T the following equality holds:
· ¸
d ∂u
Eτ,x [u (s, X(s))] = Eτ,x (s, X(s)) + L(s)u (s, ·) (X(s)) . (4.12)
ds ∂s

Next we show that under rather general conditions the process s 7→ Mu (s) −
Mu (t), t ≤ s ≤ T , as defined in (4.6) is a Pt,x -martingale. In the following
proposition we write Fst , s ∈ [t, T ], for the σ-field generated by X(ρ), ρ ∈ [t, s].
The proof of the following proposition could be based on item (c) in Theorem
1.39 in Chapter 1. For convenience we provide a direct proof based on the
Markov property.
Proposition 4.5. Fix t ∈ [τ, T ). Let the function u : [t, T ] × E → R be
∂u
such that (s, x) 7→ (s, x) + L(s)u (s, ·) (x) belongs to Cb ([t, T ] × E) :=
∂s
Cb ([t, T ] × E; C). Then the process s 7→ Mu (s) − Mu (t) is adapted to the
filtration of σ-fields (Fst )s∈[t,T ] .

Proof. Suppose that T ≥ s2 > s1 ≥ t. In order to check the martingale


property of the process Mu (s) − Mu (t), s ∈ [t, T ], it suffices to prove that
£ ¯ ¤
Et,x Mu (s2 ) − Mu (s1 ) ¯ Fst 1 = 0. (4.13)

In order to prove (4.13) we notice that by the time-inhomogeneous Markov


property:
£ ¯ ¤
Et,x Mu (s2 ) − Mu (s1 ) ¯ Fst 1 = Es1 ,X(s1 ) [Mu (s2 ) − Mu (s1 )]
·
= Es1 ,X(s1 ) u (s2 , X (s2 )) − u (s1 , X (s1 ))
Z s2 µ ¶ ¸
∂u
− L(s)u (s, X(s)) + (s, X(s)) ds
s1 ∂s
176 4 BSDE’s and Markov processes

= Es1 ,X(s1 ) [u (s2 , X (s2 )) − u (s1 , X (s1 ))]


Z s2 ·µ ¶¸
∂u
− Es1 ,X(s1 ) L(s)u (s, X(s)) + (s, X(s)) ds
s1 ∂s
Z s2
d
= Es1 ,X(s1 ) [u (s2 , X (s2 )) − u (s1 , X (s1 ))] − Es1 ,X(s1 ) [u (s, X(s))] ds
s1 ds
= Es1 ,X(s1 ) [u (s2 , X (s2 )) − u (s1 , X (s1 ))]
− Es1 ,X(s1 ) [u (s2 , X (s2 )) − u (s1 , X (s1 ))] = 0. (4.14)

The equality in (4.14) establishes the result in Proposition 4.5.

As explained in Definition 4.1 it is assumed that the subspace D(L) contains


an algebra of functions which forms a core for the operator L.
Proposition 4.6. Let the family of operators L(s), 0 ≤ s ≤ T , generate a
time-inhomogeneous Markov process

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)} (4.15)

in the sense of Definition 4.4: see equality (4.12). Then the process X(t) has
a modification which is right-continuous and has left limits on its life time.
For the definition of life time see e.g. item (a) in Theorem 1.39. The life time
ζ is defined by
(
inf {s > 0 : X(s) = 4} on the event {X(s) = 4 for some s ∈ (0, T )},
ζ=
ζ = T, if X(s) ∈ E for all s ∈ (0, T ).
(4.16)
In view of Proposition 4.6 we will assume that our Markov process has left
limits on its life time and is continuous from the right. The following proof is a
correct outline of a proof of Proposition 4.6. If E is just a polish space it needs
a considerable adaptation. Suppose that E is polish, and first assume that the
process t 7→ X(t) is conservative, i.e. assume that Pτ,x [X(t) ∈ E] = 1. Then,
by an important intermediate result (see Proposition 2.2 in Chapter 2 and
the arguments leading to it) we see that the orbits {X(ρ) : τ ≤ ρ ≤ T } are
Pτ,x -almost surely relatively compact in E. In case that the process t 7→ X(t)
is not conservative, i.e. if, for some fixed t ∈ [τ, T ], an inequality of the form
Pτ,x [X(t) ∈ E] < 1 holds, then a similar result is still valid. In fact on the
event {X(t) ∈ E} the orbit {X(ρ) : τ ≤ ρ ≤ t} is Pτ,x -almost surely relatively
compact: see Proposition 2.3 in Chapter 2. All details can be found in the proof
of item (a) of Theorem 1.39: see Subsection 2.1.1 in Chapter 2.
Proof. As indicated earlier the argument here works in case the space E is
locally compact. However, the result is true for a polish space E: see item (a)
in Theorem 1.39.
Let the function u : [0, T ] × E → R belong to the space D(L). Then the
process s 7→ Mu (s) − Mu (t), t ≤ s ≤ T , is a Pt,x -martingale. Let D[0, T ]
4.1 Introduction 177

be the set of numbers of the form k2−n T , k = 0, 1, 2, . . . , 2n . By a classical


martingale convergence theorem (see e.g. Chapter II in Revuz and Yor [199])
it follows that the following limit lim u (s, X(s)) exists Pτ,x -almost
s↑t, s∈D[0,T ]
surely for all 0 ≤ τ < t ≤ T and for all x ∈ E. In the same reference it is
also shown that the limit lim u (s, X(s)) exists Pτ,x -almost surely for
s↓t, s∈D[0,T ]
all 0 ≤ τ ≤ t < T and for all x ∈ E. Since the locally compact space [0, T ] × E
is second countable it follows that the exceptional sets may be chosen to be
independent of (τ, x) ∈ [0, T ] × E, of t ∈ [τ, T ], and of the function u ∈ D(L).
Since by hypothesis the subspace D(L) is Tβ -dense in Cb ([0, T ] × E) it follows
that the left-hand limit at t of the process s 7→ X(s), s ∈ D[0, T ] ∩ [τ, t], exists
Pτ,x -almost surely for all (t, x) ∈ (τ, T ] × E. It also follows that the right-hand
limit at t of the process s 7→ X(s), s ∈ D[0, T ] ∩ (t, T ], exists Pτ,x -almost
surely for all (t, x) ∈ [τ, T ) × E. Then we modify X(t) by replacing it with
X(t+) = lims↓t, s∈D[0,T ]∩(τ,T ] X(s), t ∈ [0, T ), and X(T +) = X(T ). It also
follows that the process t 7→ X(t+) has left limits in E.
The hypotheses in the following Proposition 4.7 are the same as those in
Proposition 4.6. The functions u and v belong to D(1) (L) = D (D1 ) ∩ D(L):
see Definition 1.30.
Proposition 4.7. Let the continuous function u : [0, T ] × E → R be such
that for every s ∈ [t, T ] the function x 7→ u(s, x) belongs to D (L(s)) and
suppose that the function (s, x) 7→ [L(s)u (s, ·)] (x) is bounded and continuous.
In addition suppose that the function s 7→ u(s, x) is continuously differentiable
for all x ∈ E. Then the process s 7→ Mu (s) − Mu (t) is a Fst -martingale with
respect to the probability Pt,x . If v is another such function, then the (right)
derivative of the quadratic co-variation process of the martingales Mu and Mv
is given by:
d
hMu , Mv i (t) = Γ1 (u, v) (t, X(t)) .
dt
In fact the following identity holds as well:
Mu (t)Mv (t) − Mu (0)Mv (0)
Z t Z t Z t
= Mu (s)dMv (s) + Mv (s)dMu (s) + Γ1 (u, v) (s, X(s)) ds. (4.17)
0 0 0

Here Fst ,
s ∈ [t, T ], is the σ-field generated by the state variables X(ρ), t ≤
ρ ≤ s. Instead of Fs0 we usually write Fs , s ∈ [0, T ]. The formula in (4.17) is
known as the integration by parts formula for stochastic integrals.
Proof. We outline a proof of the equality in (4.17). So let the functions u and
v be as in Proposition 4.7. Then we have
Mu (t)Mv (t) − Mu (0)Mv (0)
n
2X −1
¡ ¢¡ ¡ ¢ ¡ ¢¢
= Mu k2−n t Mv (k + 1)2−n t − Mv k2−n t
k=0
178 4 BSDE’s and Markov processes
n
2X
−1
¡ ¡ ¢ ¡ ¢¢ ¡ ¢
+ Mu (k + 1)2−n t − Mu k2−n t Mv k2−n t
k=0
n
2X −1
¡ ¡ ¢ ¡ ¢¢ ¡ ¡ ¢ ¡ ¢¢
+ Mu (k + 1)2−n t − Mu k2−n t Mv (k + 1)2−n t − Mv k2−n t .
k=0
(4.18)
Rt
The first term on the right-hand side of (4.18) converges to 0 Mu (s)dMv (s),
Rt
the second term converges to 0 Mv (s)dMu (s). Using the identity in (4.7) for
the function u and a similar identity for
R t v we see that the third term on the
right-hand side of (4.18) converges to 0 Γ1 (u, v) (s, X(s)) ds.
This completes the proof Proposition 4.7.

Remark 4.8. The quadratic variation process of the (local) martingale s 7→


Mu (s) is given by the process s 7→ Γ1 (u (s, ·) , u (s, ·)) (s, X(s)), and therefore
"¯Z ¯2 # ·Z s2 ¸
¯ s2 ¯
Es1 ,x ¯¯ ¯
dMu (s)¯ = Es1 ,x Γ1 (u (s, ·) , u (s, ·)) (X(s)) ds < ∞
s1 s1

under appropriate conditions on the function u. Very informally we may think


of the following representation for the martingale difference:
Z s2
Mu (s2 ) − Mu (s1 ) = ∇L
u (s, X(s)) dW (s). (4.19)
s1

Here we still have to give a meaning to the stochastic integral in the right-
hand side of (4.19). If E is an infinite-dimensional Banach space, then W (t)
should be some kind of a cylindrical Brownian motion. It is closely related to
a formula which occurs in Malliavin calculus: see Nualart [168] (Proposition
3.2.1) and [169].

Remark 4.9. It is perhaps worthwhile to observe that for Brownian motion


(W (s), Px ) the martingale difference Mu (s2 ) − Mu (s1 ), s1 ≤ s2 ≤ T , is given
by a stochastic integral:
Z s2
Mu (s2 ) − Mu (s1 ) = ∇u (τ, W (τ )) dW (τ ).
s1

Its increment of the quadratic variation process is given by


Z s2
2
hMu , Mu i (s2 ) − hMu , Mu i (s1 ) = |∇u (τ, W (τ ))| dτ.
s1

Next suppose that the function u solves the equation:


¡ ¢ ∂
f s, x, u (s, x) , ∇L
u (s, x) + L(s)u (s, x) + u (s, x) = 0. (4.20)
∂s
4.1 Introduction 179

If moreover, u (T, x) = ϕ (T, x), x ∈ E, is given, then we have


Z T ¡ ¢
u (t, X(t)) = ϕ (T, X(T )) + f s, X(s), u (s, X(s)) , ∇L
u (s, X(s)) ds
t
Z T
− dMu (s), (4.21)
t

with Mu (s) as in (4.7). From (4.21) we get

u (t, x) = Et,x [u (t, X(t))] (4.22)


Z T
£ ¡ ¢¤
= Et,x [ϕ (T, X(T ))] + Et,x f s, X(s), u (s, X(s)) , ∇L
u (s, X(s)) ds.
t

Theorem 4.10. Let u : [0, T ] × E → R be a continuous function with the


property that for every (t, x) ∈ [0, T ] × E the function s 7→ Et,x [u (s, X(s))] is
differentiable and that
· ¸
d ∂
Et,x [u (s, X(s))] = Et,x L(s)u (s, X(s)) + u (s, X(s)) , t < s < T.
ds ∂s

Then the following assertions are equivalent:


(a) The function u satisfies the following differential equation:

∂ ¡ ¢
L(t)u (t, x) + u (t, x) + f t, x, u (t, x) , ∇L
u (t, x) = 0. (4.23)
∂t
(b) The function u satisfies the following type of Feynman-Kac integral equa-
tion:
" Z T #
¡ L
¢
u (t, x) = Et,x u (T, X(T )) + f τ, X(τ ), u (τ, X(τ )) , ∇u (τ, X(τ )) dτ .
t
(4.24)
.
(c) For every t ∈ [0, T ] the process
Z s ¡ ¢
s 7→ u (s, X(s)) − u (t, X(t)) + f τ, X(τ ), u (τ, X(τ )) , ∇L
u (τ, X(τ )) dτ
t

is an Fst -martingale with respect to Pt,x on the interval [t, T ].


(d) For every s ∈ [0, T ] the process
Z T ¡ ¢
t 7→ u (T, X(T ))−u (t, X(t))+ f τ, X(τ ), u (τ, X(τ )) , ∇L
u (τ, X(τ )) dτ
t

is an FTt -backward martingale with respect to Ps,x on the interval [s, T ].


180 4 BSDE’s and Markov processes

Remark 4.11. Suppose that the function u is a solution to the following ter-
minal value problem:

 L(s)u (s, ·) (x) + ∂ u (s, x) + f ¡s, x, u (s, x) , ∇L (s, x)¢ = 0;
u
∂s (4.25)

u(T, x) = ϕ(T, x).
¡ ¢
Then the pair u (s, X(s)) , ∇L u (s, X(s)) can be considered as a weak solution
to a backward stochastic differential equation. More precisely, for every s ∈
[0, T ] the process
Z T ¡ ¢
t 7→u (T, X(T )) − u (t, X(t)) + f τ, X(τ ), u (τ, X(τ )) , ∇L
u (τ, X(τ )) dτ
t

is an FTt -backward martingale relative to Ps,x on the interval [s, T ]. The sym-
bol ∇L L
u v (s, x) stands for the functional v 7→ ∇u v (s, x) = Γ1 (u, v)(s, x), where
Γ1 is the squared gradient operator:

Γ1 (u, v)(s, x) (4.26)


1
= Tβ - lim Es,x [(u (s, X(t)) − u (s, X(s))) (v (s, X(t)) − v (s, X(s)))] .
t↓s t − s

Possible choices for the function f are


¡ ¢
f s, x, y, ∇L
u = −V (s, x)y and (4.27)
¡ ¢ ¯
1¯ L ¯ 2 1
f s, x, y, ∇L
u = ∇ (s, x)¯ − V (s, x) = Γ1 (u, u) (s, x) − V (s, x).
2 u 2
(4.28)

The choice in (4.27) turns equation (4.25) into the following heat equation:

 ∂ u (s, x) + L(s)u (s, ·) (x) − V (s, x)u(s, x) = 0;
∂s (4.29)

u (T, x) = ϕ(T, x).

The function v(s, x) defined by the Feynman-Kac formula


h RT i
v(s, x) = Es,x e− s V (ρ,X(ρ))dρ ϕ (T, X(T )) (4.30)

is a candidate solution to equation (4.29).


The choice in (4.28) turns equation (4.25) into the following Hamilton-Jacobi-
Bellman equation:

 ∂ u (s, x) + L(s)u (s, X(s)) − 1 Γ (u, u) (s, x) + V (s, x) = 0;
1
∂s 2 (4.31)

u (T, x) = − log ϕ(T, x),
4.1 Introduction 181

where − log ϕ(T, x) replaces ϕ(T, x). The function SL defined by the genuine
non-linear Feynman-Kac formula
h RT i
SL (s, x) = − log Es,x e− s V (ρ,X(ρ))dρ ϕ (T, X(T )) (4.32)

is a candidate solution to (4.31). Often these “candidate solutions” are viscos-


ity solutions. However, this was the main topic in [246] and is the main topic
in Chapter 5.

Remark 4.12. Let u(t, x) satisfy one of the equivalent conditions in Theorem
4.10. Put Y (τ ) = u (τ, X(τ )), and let M (s) be the martingale determined by
M (0) = Y (0) = u (0, X(0)) and by
Z s
¡ ¢
M (s) − M (t) = Y (s) + f τ, X(τ ), Y (τ ), ∇L
u (τ, X(τ )) dτ.
t

Then the expression ∇L u (τ, X(τ )) only depends on the martingale part M of
the process s 7→ Y (s). This entitles us to write ZM (τ ) instead of ∇L u (τ, X(τ )).
d
The interpretation of ZM (τ ) is then the linear functional N 7→ hM, N i (τ ),
¡ ¢ dτ
2 0
where¡ N 0is a Pτ,x
¢ -martingale in M Ω, FT , Pt,x2 .¡ Here0a process ¢ N belongs to
2
M Ω, FT , Pt,x whenever N is martingale in L Ω, FT , Pt,x . Notice ¡ that the¢
functional ZM (τ ) is known as soon as the martingale M ∈ M2 Ω, FT0 , Pt,x
is known. From our definitions it also follows that
Z T
M (T ) = Y (T ) + f (τ, X(τ ), Y (τ ), ZM (τ )) dτ,
0

where we used the fact that Y (0) = M (0).

Remark 4.13. Let the notation be as in Remark 4.12. Then the variables Y (t)
and ZM (t) only depend on the space-time variable (t, X(t)), and as a con-
sequence the martingale increments M (t2 ) − M (t1 ), 0 ≤ t1 < t2 ≤ T , only
depend on Ftt21 = σ (X(s) : t1 ≤ s ≤ t2 ). In Section 4.2 we give Lipschitz type
conditions on the function f in order that the BSDE
Z T
Y (t) = Y (T ) + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T ), τ ≤ t ≤ T,
t
(4.33)
possesses a unique pair of solutions

(Y, M ) ∈ L2 (Ω, FTτ , Pτ,x ) × M2 (Ω, FTτ , Pτ,x ) .

Here M2 (Ω, FTt , Pt,x ) stands for the space of all (Fst )s∈[t,T ] -martingales in
L2 (Ω, FTt , Pt,x ). Of course instead of writing “BSDE” it would be bet-
ter to write “BSIE” for Backward Stochastic Integral Equation. However,
since in the literature people write “BSDE” even if they mean integral
182 4 BSDE’s and Markov processes

equations we also stick to this terminology. Suppose that the σ (X(T ))-
measurable variable Y (T ) ∈ L2 (Ω, FTτ , Pτ,x ) is given. In fact we will prove
that
¡ the solution ¢(Y, M ¡) of the equation ¢ in (4.33) belongs to the space
S2 Ω, FTt , Pt,x ; Rk ×M2 Ω, FTt , Pt,x ; Rk . For more details see the definitions
4.18 and 4.28, and Theorem 4.42.
Remark 4.14. Let M and N be two martingales in M2 [0, T ]. Then, for 0 ≤
s < t ≤ T,
2
|hM, N i (t) − |hM, N i (s)||
≤ (hM, M i (t) − hM, M i (s)) (hN, N i (t) − hN, N i (s)) ,

and consequently
¯ ¯2
¯d ¯
¯ hM, N i (s)¯ ≤ d hM, M i (s) d hN, N i (s).
¯ ds ¯ ds ds
Hence, the inequality
Z T¯ ¯ Z Tµ ¶1/2 µ ¶1/2
¯d ¯ d d
¯ hM, N i (s)¯ ds ≤ hM, M i (s) hN, N i (s) ds
¯ ds ¯ ds ds
0 0
(4.34)
Z T¯ ¯
¯d ¯
follows. The inequality in (4.34) says that the quantity ¯ ¯
¯ ds hM, N i (s)¯ ds
0
is dominated by the Hellinger integral H (M, N ) defined by the right-hand
side of (4.34).
For a proof we refer the reader to [246]. We insert a proof here as well.
Proof (Proof of Theorem 4.10). For brevity, only in this proof, we write

F (τ, X(τ )) = f (τ, X (τ ) , u (τ, X (τ ))) .

(a) =⇒ (b). The equality in (b) is the same as the one in (4.22) which is a
consequence of (4.20).
(b) =⇒ (a). We calculate the expression
· Z s ¸
∂ ¡ L
¢
Et,x u (s, X(s)) + f τ, X (τ ) , u (τ, X (τ )) , ∇u (τ, X (τ )) dτ .
∂s t

First of all it is equal to


· ¸

Et,x u (s, X(s)) + L(s)u (s, X(s)) + F (s, X(s)) (4.35)
∂s
Next we also have by (4.24) in (b):
· Z s ¸
∂ ¡ L
¢
Et,x u (s, X(s)) + f τ, X (τ ) , u (τ, X (τ )) , ∇u (τ, X (τ )) dτ
∂s t
4.1 Introduction 183
" " Z # Z #
T s

= Et,x Es,X(s) u (T, X(T )) + F (τ, X(τ )) dτ + F (τ, X(τ )) dτ
∂s s t

(Markov property)
" " Z T # Z #
¯ t s
∂ ¯
= Et,x Et,x u (T, X(T )) + F (τ, X(τ )) dτ Fs + F (τ, X(τ )) dτ
∂s s t
" " Z T ##
∂ ¯
= Et,x Et,x u (T, X(T )) + F (τ, X(τ )) dτ ¯ Fst
∂s t
" Z T #

= Et,x u (T, X(T )) + F (τ, X(τ )) dτ = 0. (4.36)
∂s t

From (4.36) and (4.35) we get


· ¸
∂ ¡ L
¢
Et,x u (s, X(s)) + L(s)u (s, X(s))+f s, X (s) , u (s, X (s)) , ∇u (s, X (s))
∂s
= 0, s > t. (4.37)

Passing to the limit for s ↓ t in (4.37) we obtain:


· ¸
∂ ¡ ¢
Et,x u (t, X(t)) + L(t)u (t, X(t)) + f t, X (t) , u (t, X (t)) , ∇L
u (t, X (t))
∂t
=0 (4.38)

and, since X(t) = x Pt,x -almost surely, we obtain equality (4.23) in assertion
(a).
(a) =⇒ (c). If the function u satisfies the differential equation in (a), then
from the equality in (4.5) we see that
Z s
¡ ¢
0 = u (s, X (s)) − u (t, X (t)) + f τ, X(τ ), u (τ, X(τ )) , ∇Lu (τ, X(τ )) dτ
Zt s µ ¶
∂u
− u (s, X (s)) + u (t, X (t)) + L(τ )u (τ, X(τ )) + (τ, X(τ )) dτ
t ∂τ
(4.39)
Z s
¡ ¢
= u (s, X (s)) − u (t, X (t)) + f τ, X(τ ), u (τ, X(τ )) , ∇L u (τ, X(τ )) dτ
t
− Mu (s) + Mu (t) , (4.40)

where, as in (3.217),

Mu (s) − Mu (t)
Z s µ ¶
∂u
= u (s, X (s)) − u (t, X (t)) − L(τ )u (τ, X(τ )) + (τ, X(τ )) dτ
t ∂τ
184 4 BSDE’s and Markov processes
Z s
= dMu (τ ). (4.41)
t

Since the expression in (4.40) vanishes (by assumption (a)) we see that the
process in (c) is the same as the martingale s 7→ Mu (s) − Mu (t), s ≥ t. This
proves the implication (a) =⇒ (c).
The implication (c) =⇒ (b) is a direct consequence of assertion (c) and the
fact that X(t) = x Pt,x -almost surely.
The equivalence of the assertions (a) and (d) is proved in the same manner
as the equivalence of (a) and (c). Here we employ the fact that the process
t 7→ Mu (T ) − Mu (t) is an FTt -backward martingale on the interval [s, T ] with
respect to the probability Ps,x .
This completes the proof of Theorem 4.10
Remark 4.15. Instead of considering ∇Lu (s, x) we will also consider the bilinear
mapping Z(s) which associates with a pair of local semi-martingales (Y1 , Y2 )
a process which is to be considered as the right derivative of the co-variation
process: hY1 , Y2 i (s). We write
d
ZY1 (s) (Y2 ) = Z(s) (Y1 , Y2 ) = hY1 , Y2 i (s).
ds
The function f (i.e. the generator of the backward differential equation)
will
¡ then be of the ¢ form: f (s, X(s), Y (s), ZY (s)); the deterministic phase
u(s, x), ∇L u(s, x) is replaced with the stochastic phase (Y (s), ZY (s)). We
should find an appropriate stochastic phase s 7→ (Y (s), ZY (s)), which we
identify with the process s 7→ (Y (s), MY (s)) in the stochastic phase space
S2 × M2 , such that
Z T Z T
Y (t) = Y (T ) + f (s, X(s), Y (s), ZY (s)) ds − dMY (s), (4.42)
t t

where the quadratic variation of the martingale MY (s) is given by

d hMY , MY i (s) = ZY (s) (Y ) ds = Z(s) (Y, Y ) ds = d hY, Y i (s).

This stochastic phase space S2 × M2 plays a role in stochastic analysis very


similar to the role played by the first Sobolev space H 1,2 in the theory of
deterministic partial differential equations.
Remark 4.16. In case we deal with strong solutions driven by standard Brow-
nian
R s2 motion the martingale difference MY (s2 ) − MY (s1 ) can be written as
Z (s)dW (s), provided that the martingale MY (s) belongs to the space
s1 ¡ Y ¢
M2 Ω, G0T , P . Here G0T is the σ-field generated by W (s), 0 ≤ s ≤ T . If
Y (s) = u (s, X(s)), then this stochastic integral satisfies:
Z s2
ZY (s)dW (s) = u (s2 , X (s2 )) − u (s1 , X (s1 ))
s1
4.1 Introduction 185
Z s2 µ ¶

− L(s) + u (s, X (s)) ds. (4.43)
s1 ∂s

Such stochastic integrals are for example defined if the process X(t) is a
solution to a stochastic differential equation (in Itô sense):
Z s Z s
X(s) = X(t)+ b (τ, X(τ )) dτ + σ (τ, X(τ )) dW (τ ), t ≤ s ≤ T. (4.44)
t t

d
Here the matrix (σjk (τ, x))j,k=1 is chosen in such a way that

d
X
ajk (τ, x) = σj` (τ, x) σk` (τ, x) = (σ(τ, x)σ ∗ (τ, x))jk .
`=1

The process W (τ ) is Brownian motion or Wiener process. It is assumed that


operator L(τ ) has the form
d
1 X ∂2
L(τ )u(x) = b (τ, x) · ∇u(x) + ajk (τ, x) u(x). (4.45)
2 ∂xj xk
j,k=1

Then from Itô’s formula together with (4.43), (4.44) and (4.45) it follows that

the process ZY (s) has to be identified with σ (s, X(s)) ∇u (s, ·) (X(s)). For
more details see e.g. Pardoux and Peng [184] and Pardoux [181]. The equality
in (4.43) is a consequence of a martingale representation theorem: see e.g.
Proposition 3.2 in Revuz and Yor [199].
Remark 4.17. Backward doubly stochastic differential equations (BDSDEs)
could have been included in the present chapter: see Boufoussi, Mrhardy and
Van Casteren [41]. In our notation a BDSDE may be written in the form:
Z T µ ¶
d
Y (t) − Y (T ) = f s, X(s), Y (s), N 7→ hM, N i (s) ds
t ds
Z T µ ¶
d ←−
+ g s, X(s), Y (s), N 7→ hM, N i (s) d B (s)
t ds
+ M (t) − M (T ). (4.46)

Here the expression


Z T µ ¶
d ←−
g s, X(s), Y (s), N 7→ hM, N i (s) d B (s)
t ds

represents a backward Itô integral. The symbol hM, N i stands for the co-
variation process of the (local) martingales M and N ; it is assumed that this
process is absolutely continuous with respect to Lebesgue measure. Moreover,

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)}


186 4 BSDE’s and Markov processes

is a Markov process generated by a family of operators L(s), 0 ≤ s ≤ T , and


Ftτ = σ {X(s) : τ ≤ s ≤ t}. The process X(t) could be the (unique) weak or
strong solution to a (forward) stochastic differential equation (SDE):
Z t Z t
X(t) = x + b (s, X(s)) ds + σ (s, X(s)) dW (s). (4.47)
τ τ

Here the coefficients b and σ have certain continuity or measurability proper-


ties, and Pτ,x is the distribution of the process X(t) defined as being the
unique weak solution to the equation in (4.47). We want to find a pair
(Y, M ) ∈ S2 (Ω, Ftτ , Pτ,x ) × M2 (Ω, Ftτ , Pτ,x ) which satisfies (4.46).

We first give some definitions. Fix (τ, x) ∈ [0, T ] × E. In the definitions 4.18
and 4.19 the probability measure Pτ,x is defined on the σ-field FTτ . In Defi-
nition 4.28 we return to these notions. The following definition and implicit
results described therein shows that, under certain conditions, by enlarging
the sample space a family of processes may be reduced to just one process
without losing the S2 -property.
Definition 4.18. Fix ¡(τ, x) ∈ [0, T ] ×¢E. An Rk -valued process Y is said to
2 τ k τ
belong to the
· space S Ω,¸FT , Pτ,x ; R if Y (t) is Ft -measurable (τ ≤ t ≤ T )
2
and if Eτ,x sup |Y (t)| < ∞. It is assumed that Y (s) = Y (τ ), Pτ,x -almost
τ ≤t≤T
surely,
¡ for s ∈ [0, τ ]. The
¢ process Y (s), s ∈ [0, T ], is said to belong to the space
S2unif Ω, FTτ , Pτ,x ; Rk if
· ¸
2
sup Eτ,x sup |Y (t)| < ∞,
(τ,x)∈[0,T ]×E τ ≤t≤T

¡ ¢
and it belongs to S2loc,unif Ω, FTτ , Pτ,x ; Rk provided that
· ¸
2
sup Eτ,x sup |Y (t)| <∞
(τ,x)∈[0,T ]×K τ ≤t≤T

for all compact subsets K of E.


¡ ¢
If the σ-field Ftτ and Pτ,x are clear from the context we write S2 [0, T ], Rk
or sometimes just S2 .
Definition 4.19. Let the process M be such that the process t 7→ M (t)−M (τ ),
t ∈ [τ, T ], is a Pτ,x -martingale with the property that the stochastic variable
2 τ
M (T ) − M ¡ (τ ) τbelongs tok ¢L (Ω, FT , Pτ,x ). Then M is said to belong to the
2
space M Ω, FT , Pτ,x ; R . By the Burkholder-Davis-Gundy
· inequality
¸ (see
2
inequality (4.79) below) it follows that Eτ,x sup |M (t) − M (τ )| is finite
τ ≤t≤T
if and only if M (T ) − M (τ ) belongs to the space L2 (Ω, FTτ , Pτ,x ). Here an
Ftτ -adapted process M (·) − M (τ ) is called a Pτ,x -martingale provided that
4.1 Introduction 187
£ ¯ ¤
Eτ,x [|M (t) − M (τ )|] < ∞ and Eτ,x M (t) − M (τ ) ¯ Fsτ = M (s) − M (τ ),
Pτ,x -almost surely, for T ≥ t ≥ s ≥ τ . The martingale¡ difference s 7→¢
M (s) − M (0), s ∈ [0, T ], is said to belong to the space M2unif Ω, FTτ , Pτ,x ; Rk
if · ¸
2
sup Eτ,x sup |M (t) − M (τ )| < ∞,
(τ,x)∈[0,T ]×E τ ≤t≤T
¡ ¢
and it belongs to M2loc,unif Ω, FTτ , Pτ,x ; Rk provided that
· ¸
2
sup Eτ,x sup |M (t) − M (τ )| <∞
(τ,x)∈[0,T ]×K τ ≤t≤T

for all compact subsets K of E.


From the Burkholder-Davis-Gundy inequality (see inequality
¡ (4.79) below)
¢ it
follows that the process M (s)−M (0) belongs to M2unif Ω, FTτ , Pτ,x ; Rk if and
only if
h i
2
sup Eτ,x |M (T ) − M (τ )|
(τ,x)∈[0,T ]×E

= sup Eτ,x [hM, M i (T ) − hM, M i (τ )] < ∞.


(τ,x)∈[0,T ]×E

Here hM, M i stands for the quadratic variation process of the process t 7→
M (t) − M (0).
The notions in the definitions 4.18 and (4.19) will exclusively be used in
case the family of measures {Pτ,x : (τ, x) ∈ [0, T ] × E} constitute the distri-
butions of a Markov process which was defined in Definition 4.3.
Again let the Markov process, with right-continuous sample paths and
with left limits,

{(Ω, FTτ , Pτ,x ) , (X(t) : T ≥ t ≥ 0) , (E, E)} (4.48)

be generated by the family of operators {L(s) : 0 ≤ s ≤ t}: see definitions 4.3,


equality (4.9), and 4.4, equality (4.11).
Next we define the family of operators {Q (t1 , t2 ) : 0 ≤ t1 ≤ t2 ≤ T } by

Q (t1 , t2 ) f (x) = Et1 ,x [f (X (t2 ))] , f ∈ Cb (E) , 0 ≤ t1 ≤ t2 ≤ T. (4.49)

Fix ϕ ∈ D(L). Since the process t 7→ Mϕ (t) − Mϕ (s), t ∈ [s, T ], is a Ps,x -


martingale with respect to the filtration (Fts )t∈[s,T ] , and X(t) = x Pt,x almost
surely, the following equality follows:
Z t
Es,x [L(ρ)ϕ (ρ, ·) (X(ρ))] dρ + Et,x [ϕ (t, X(t))] − Es,x [ϕ (t, X(t))]
s
Z t · ¸
∂ϕ
= ϕ(t, x) − ϕ(s, x) − Es,x (ρ, X(ρ)) dρ. (4.50)
s ∂ρ
188 4 BSDE’s and Markov processes

The fact that a process of the form t 7→ Mϕ (t) − Mϕ (s), t ∈ [s, T ], is a Ps,x -
martingale follows from Proposition 4.5. In terms of the family of operators

{Q (t1 , t2 ) : 0 ≤ t1 ≤ t2 ≤ T }

the equality in (4.50) can be rewritten as


Z t
Q (s, ρ) L(ρ)ϕ (ρ, ·) (x) dρ + Q(t, t)ϕ (t, ·) (x) − Q(s, t)ϕ (t, ·) (x)
s
Z t
∂ϕ
= ϕ(t, x) − ϕ(s, x) − Q (s, ρ) (ρ, ·) (x)dρ. (4.51)
s ∂ρ

From (4.51) we infer that

Q(t, t)ϕ (t, ·) (x) − Q(s, t)ϕ (t, ·) (x)


L(s)ϕ(s, ·)(x) = − lim . (4.52)
t↓s t−s

Equality (4.51) also yields the following result. If ϕ ∈ D(L) is such that

∂ϕ
L(ρ)ϕ (ρ, ·) (y) = − (ρ, y),
∂ρ
then
ϕ (s, x) = Q (ρ, t) ϕ (t, ·) (x) = Es,x [ϕ (t, X(t))] . (4.53)
Since 0 ≤ s ≤ t ≤ T are arbitrary from (4.53) we see

Q (s, t0 ) ϕ (t0 , ·) (x) = Q (s, t) Q (t, t0 ) ϕ (t0 , ·) (x) 0 ≤ s ≤ t ≤ t0 ≤ T, x ∈ E.


(4.54)
If in (4.54) we (may) choose the function ϕ (t0 , y) arbitrary, then the family
Q(s, t), 0 ≤ s ≤ t ≤ T , automatically is a propagator in the space Cb (E) in
the sense that Q (s, t) Q (t, t0 ) = Q (s, t0 ), 0 ≤ s ≤ t ≤ t0 ≤ T . For details on
propagators or evolution families see [95].
Remark 4.20. In the sequel we want to discuss solutions to equations of the
form:
∂ ¡ ¢
u (t, x) + L(t)u (t, ·) (x) + f t, x, u (t, x) , ∇L
u (t, x) = 0. (4.55)
∂t
For a preliminary discussion on this topic see Theorem 4.10. Under certain
hypotheses on the function f we will give existence and uniqueness results.
Let m be (equivalent to) the Lebesgue measure in Rd . In a concrete
¡ situation
¢
where every operator L(t) is a genuine diffusion operator in L2 Rd , m we
consider the following Backward Stochastic Differential equation
Z T ¡ ¢
u (s, X(s)) = Y (T, X(T )) + f ρ, X(ρ), u (ρ, X(ρ)) , ∇L
u (ρ, X(ρ)) dρ
s
4.1 Introduction 189
Z T
− ∇L
u (ρ, X(ρ)) dW (ρ) . (4.56)
s

Here we suppose that the process t 7→ X(t) is a solution to a genuine stochastic


differential equation driven by Brownian motion and with one-dimensional
∂u
distribution u(t, x) satisfying L(t)u (t, ·) (x) = (t, x). In fact in that case
∂t
we will not consider the equation in (4.56), but we will try to find an ordered
pair (Y, Z) such that
Z T Z T
Y (s) = Y (T ) + f (ρ, X(ρ), Y (ρ) , Z (ρ)) dρ − hZ (ρ) , dW (ρ)i . (4.57)
s s

If the pair (Y, Z) satisfies (4.57), then u (s, x) = Es,x [Y (s)] satisfies (4.55).
Moreover Z(s) = ∇L L
u (s, X(s)) = ∇u (s, x), Ps,x -almost surely. For more de-
tails see section 2 in Pardoux [181].
Remark 4.21. Some remarks follow:
(a) In section 4.2 weak solutions to BSDEs are studied.
(b) In section 7 of [246] and in section 2 of Pardoux [181] strong solutions to
BSDEs are discussed: these results are due to Pardoux and collaborators.
(c) BSDEs go back to Bismut: see e.g. [32].
d d
1 X ∂2u X ∂u
(d) If L(s)u(s, x) = aj,k (s, x) (s, x) + bj (s, x) (s, x), then
2 ∂xj xk j=1
∂x j
j,k=1

d
X ∂u ∂v
Γ1 (u, v) (s, x) = aj,k (s, x) (s, x) (s, x).
∂xj ∂xk
j,k=1

As a corollary to theorems 4.10 and 4.34 we have the following result.


Corollary 4.22. Suppose that the function u solves the following

 ∂u (s, y) + L(s)u(s, ·) (y) + f ¡s, y, u(s, y), ∇L (s, y)¢ = 0;
u
∂s (4.58)
 2 τ
u (T, X(T )) = ξ ∈ L (Ω, FT , Pτ,x ) .
Let the pair (Y, M ) be a solution to
Z T
Y (t) = ξ + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T ), (4.59)
t

with M (τ ) = 0. Then
(Y (t), M (t)) = (u (t, X(t)) , Mu (t)) ,
where
Z t Z t
∂u
Mu (t) = u (t, X(t))−u (τ, X(τ ))− L(s)u (s, ·) (X(s)) ds− (s, X(s)) ds.
τ τ ∂s
190 4 BSDE’s and Markov processes

Notice that the processes s 7→ ∇Lu (s, X(s)) and s 7→ ZMu (s) may be identified
and that ZMu (s) only depends on (s, X(s)). The decomposition
Z tµ ¶
∂u
u (t, X(t)) − u (τ, X(τ )) = (s, X(s)) + L(s)u (s, ·) (X(s)) ds
τ ∂s
+ Mu (t) − Mu (τ ) (4.60)

splits the process t 7→ u (t, X(t)) − u (τ, X(τ )) into a part which is bounded
variation (i.e. the part which is absolutely continuous with respect to Lebesgue
measure on [τ, T ]) and a Pτ,x -martingale part Mu (t) − Mu (τ ) (which in fact
is a martingale difference part).
If L(s) = 12 ∆, then X(s) = W (s) (standard Wiener process or Brownian
motion) and (4.60) can be rewritten as
Z tµ ¶
∂u 1
u (t, W (t)) − u (τ, W (τ )) = (s, W (s)) + ∆u (s, ·) (W (s)) ds
τ ∂s 2
Z t
+ ∇u (s, ·) (W (s)) dW (s) (4.61)
τ
Rt
where τ
∇u (s, ·) (W (s)) dW (s) is to be interpreted as an Itô integral.
Remark 4.23. Suggestions for further research:
(a) Find “explicit solutions” to BSDEs with a linear drift part. This should
be a type of Cameron-Martin formula or Girsanov transformation.
(b) Treat weak (and strong) solutions BDSDEs in a manner similar to what
is presented here for BSDEs.
(c) Treat weak (strong) solutions to BSDEs generated by a function f which
is not necessarily of linear growth but for example of quadratic growth in
one or both of its entries Y (t) and ZM (t).
(d) Can anything be done if f depends not only on s, x, u(s, x), ∇u (s, x), but
also on L(s)u (s, ·) (x)?
In the following proposition it is assumed that the operator L generates a
strong Markov process in the sense of the definitions 1.30 and 1.31.
Proposition 4.24. Let the functions f , g ∈ D(L) be such that their product
f g also belongs to D(L). Then Γ1 (f, g) is well defined and for (s, x) ∈ [0, T ] ×
E the following equality holds:

L(s) (f g) (s, ·) (x) − f (s, x)L(s)g (s, ·) (x) − L(s)f (s, ·) (x)g(s, x)
= Γ1 (f, g) (s, x). (4.62)

Proof. Let the functions f and g be as in Proposition 4.24. For h > 0 we have:

(f (X(s + h)) − f (X(s))) (g (X(s + h)) − g (X(s)))


= f (X(s + h)) g (X(s + h)) − f (X(s)) g (X(s)) (4.63)
4.2 A probabilistic approach: weak solutions 191

− f (X(s)) (g (X(s + h)) − g (X(s))) − (f (X(s + h)) − f (X(s))) g (X(s)) .

Then we take expectations with respect to Es,x , divide by h > 0, and pass to
the Tβ -limit as h ↓ 0 to obtain equality (4.62) in Proposition 4.24.

4.2 A probabilistic approach: weak solutions


In this section and also in sections 4.3 we will study BSDE’s on a single prob-
ability space. In Section 4.4 and Chapter 5 we will consider Markov families
of probability spaces. In the present section we write P instead of P0,x , and
similarly for the expectations E and E0,x . Here we work on the interval [0, T ].
Since we are discussing the martingale problem and basically only the dis-
tributions of the process t 7→ X(t), t ∈ [0, T ], the solutions we obtain are of
weak type. In case we consider strong solutions we apply a martingale rep-
resentation theorem (in terms of Brownian Motion). In Section 4.4 we will
also use this result for probability measures of the form Pτ,x on the inter-
0
val [τ, T ]. In¡this section ¢we consider
¡ a pair of¢ Ft = Ft -adapted processes
2 k 2 k
(Y, M ) ∈ L Ω, FT , P; R × L Ω, FT , P : R such that Y (0) = M (0) and
such that
Z T
Y (t) = Y (T ) + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T ) (4.64)
t

where M is a P-martingale with respect to the filtration Ft = σ (X(s) : s ≤ t).


In [246] we will employ the results of the present section with P = Pτ,x , where
(τ, x) ∈ [0, T ] × E.
Proposition 4.25. Let the pair (Y, M ) be as in (4.64), and suppose that
Y (0) = M (0). Then
Z t
Y (t) = M (t) − f (s, X(s), Y (s), ZM (s)) ds, and (4.65)
0
" Z T #
¯
Y (t) = E Y (T ) + f (s, X(s), Y (s), ZM (s)) ds ¯ Ft ; (4.66)
t
" Z #
T ¯
M (t) = E Y (T ) + f (s, X(s), Y (s), ZM (s)) ds ¯ Ft . (4.67)
0

The equality in (4.65) shows that the process M is the martingale part of the
semi-martingale Y .
Proof. The equality in (4.66) follows from (4.64) and from the fact that M is
a martingale. Next we calculate
" Z T #
¯
E Y (T ) + f (s, X(s), Y (s), ZM (s)) ds ¯ Ft
0
192 4 BSDE’s and Markov processes
" Z #
T ¯
= E Y (T ) + f (s, X(s), Y (s), ZM (s)) ds ¯ Ft
t
Z t
+ f (s, X(s), Y (s), ZM (s)) ds
0
Z t
= Y (t) + f (s, X(s), Y (s), ZM (s)) ds
0

(employ (4.64))
Z T
= Y (T ) + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T )
t
Z t
+ f (s, X(s), Y (s), ZM (s)) ds
0
Z T
= Y (T ) + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T )
0
= M (T ) + M (t) − M (T ) = M (t). (4.68)

The equality in (4.68) shows (4.67). Since


Z T
M (T ) = Y (T ) + f (s, X(s), Y (s), ZM (s)) ds
0

the equality in (4.65) follows.


In the following theorem we write z = ZM (s) and y belongs to Rk .
Theorem 4.26. Suppose that there exist finite constants C1 and C2 such that
2
hy2 − y1 , f (s, x, y2 , z) − f (s, x, y1 , z)i ≤ C1 |y2 − y1 | ; (4.69)
2 d
|f (s, x, y, ZM2 (s)) − f (s, x, y, ZM1 (s))| ≤ C22 hM2 − M1 , M2 − M1 i (s).
ds
(4.70)

Then there exists a unique pair of adapted processes (Y, M ) such that Y (0) =
M (0) and such that the process M is the martingale part of the semi-
martingale Y :
Z T
Y (t) = M (t) − M (T ) + Y (T ) + f (s, X(s), Y (s), ZM (s)) ds
t
Z t
= M (t) − f (s, X(s), Y (s), ZM (s)) ds. (4.71)
0

The following proof contains just an outline of the proof of Theorem 4.26.
Complete and rigorous arguments are found in the proof of Theorem 4.33: see
Theorem 4.42 as well.
4.2 A probabilistic approach: weak solutions 193

Proof. The uniqueness follows from Corollary 4.32 of Theorem 4.30 below.
In the existence part of the proof of Theorem 4.26 we will approximate the
−1
function f by Lipschitz continuous functions fδ , 0 < δ < (2C1 ) , where each
−1
function fδ has Lipschitz constant δ , but at the same time inequality (4.70)
remains valid for fixed second variable (in an appropriate sense). It follows
that for the functions fδ (4.70) remains valid and that (4.69) is replaced with
1
|fδ (s, x, y2 , z) − fδ (s, x, y1 , z)| ≤ |y2 − y1 | . (4.72)
δ
In the uniqueness part of the proof it suffices to assume that (4.69) holds. In
Theorem 4.34 we will see that the monotonicity condition (4.69) also suffices
to prove the existence. For details the reader is referred to the propositions
4.35 and 4.36, Corollary 4.37, and to Proposition 4.39. In fact for M ∈ M2
fixed, and the function y 7→ f (s, x, y, ZM (s)) satisfying (4.69) the function
y 7→ y − δf (s, x, y, ZM (s)) is surjective as a mapping from Rk to Rk and
its inverse exists and is Lipschitz continuous with constant 2. The Lipschitz
continuity is proved in Proposition 4.36. The surjectivity of this mapping is a
consequence of Theorem 1 in [63]. As pointed out by Crouzeix et al the result
follows from a non-trivial homotopy argument. A relatively elementary proof
of Theorem 1 in [63] can be found for a continuously differentiable function
in Hairer and Wanner [98]: see Theorem 14.2 in Chapter IV. For a few more
details see Remark 4.38. Let fs,M be the mapping y 7→ f (s, y, ZM (s)), and
put ³ ´
−1
fδ (s, x, y, ZM (s)) = f s, x, (I − δfs,x,M ) , ZM (s) . (4.73)
−1
Then the functions fδ , 0 < δ < (2C1 ) , are Lipschitz continuous with con-
stant δ −1 . Proposition 4.39 treats the transition from solutions of BSDE’s
with generator fδ with fixed martingale M ∈ M2 to solutions of BSDE’s
driven by f with the same fixed martingale M . Proposition 4.35 contains the
passage from solutions (Y, N ) ∈ S2 × M2 to BBSDE’s with generators of the
form (s, y) 7→ f (s, y, ZM (s)) for any fixed martingale M ∈ M2 to solutions
for BSDE’s of the form (4.71) where the pair (Y, M ) belongs to S2 × M2 . By
hypothesis the process s 7→ f (s, x, Y (s), ZM (s)) satisfies (4.69) and (4.70).
Essentially speaking a combination of these observations show the result in
Theorem 4.26.

Remark 4.27. In the literature functions with the monotonicity property are
also called one-sided Lipschitz functions. In fact Theorem 4.26, with f (t, x, ·, ·)
Lipschitz continuous in both variables, will be superseded by Theorem 4.33
in the Lipschitz case and by Theorem 4.34 in case of monotonicity in the
second variable and Lipschitz continuity in the third variable. The proof of
Theorem 4.26 is part of the results in Section 4.3. Theorem 4.42 contains a
corresponding result for a Markov family of probability measures. Its proof is
omitted, it follows the same lines as the proof of Theorem 4.34.
194 4 BSDE’s and Markov processes

4.3 Existence and Uniqueness of solutions to BSDE’s


The equation in (4.55) can be phrased in a semi-linear setting as follows. Find
a function u (t, x) which satisfies the following partial differential equation:

 ∂u (s, x) + L(s)u (s, x) + f ¡s, x, u(s, x), ∇L (s, x)¢ = 0;
u
∂s (4.74)

u(T, x) = ϕ (T, x) , x ∈ E.

Here ∇L f2 (s, x) is the linear functional f1 7→ Γ1 (f1 , f2 ) (s, x) for smooth


enough functions f1 and f2 . For s ∈ [0, T ] fixed the symbol ∇L f2 stands for the
linear mapping f1 7→ Γ1 (f1 , f2 ) (s, ·). One way to treat this kind of equation is
considering the following backward problem. Find a pair of adapted processes
(Y, ZY ), satisfying
Z T
Y (t) − Y (T ) − f (s, X(s), Y (s), Z(s) (·, Y )) ds = M (t) − M (T ), (4.75)
t

where M (s), t0 < t ≤ s ≤ T , is a forward


¡ local
¢ Pt,x -martingale (for every
T > t > t0 ). The symbol ZY1 , Y1 ∈ S2 [0, T ], Rk , stands for the functional

d ¡ ¢
ZY1 (Y2 ) (s) = Z(s) (Y1 (·), Y2 (·)) = hY1 (·), Y2 (·)i (s), Y2 ∈ S2 [0, T ], Rk .
ds
(4.76)
If the pair (Y, ZY ) satisfies (4.75), then ZY = ZM .¡Instead of¢trying¡to find the¢
pair (Y, ZY ) we will try to find a pair (Y, M ) ∈ S2 [0, T ], Rk ×M2 [0, T ], Rk
such that
Z T
Y (t) = Y (T ) + f (s, X(s), Y (s), ZM (s)) ds + M (t) − M (T ).
t
¡ ¢ ¡ ¢
Next we define the spaces S2 [0, T ], Rk and M2 [0, T ], Rk : compare with
the definitions 4.18 and 4.19.
Definition 4.28. Let (Ω, F, P) be a probability space, and let Ft , t ∈ [0, T ],
be a filtration on F. Let t 7→ Y (t) be an stochastic process with values in Rk
which is adapted to the filtration Ft and which
¡ is P-almost
¢ surely continuous.
Then Y is said to belong to the space S2 [0, T ], Rk provided that
" #
2
E sup |Y (t)| < ∞.
t∈[0,T ]
¡ ¢
Definition 4.29.
¡ The space¢ of Rk -valued martingales in L2 Ω, F, P; Rk is
denoted by M2¡ [0, T ], Rk¢ . So that a continuous martingale t 7→ M (t) − M (0)
belongs to M2 [0, T ], Rk if
h i
2
E |M (T ) − M (0)| < ∞. (4.77)
4.3 Existence and Uniqueness of solutions to BSDE’s 195
2 2
Since the process t 7→ |M (t)| − |M (0)| − hM, M i (t) + hM, M i (0) is a mar-
tingale difference we see that
h i
2
E |M (T ) − M (0)| = E [hM, M i (T ) − hM, M i (0)] , (4.78)
¡ ¢
and hence
¡ a martingale
¢ difference t 7→ M (t)−M (0) in L2 Ω, F, P; Rk belongs
to M2 [0, T ], Rk if and only if E [hM, M i (T ) − hM, M i (0)] is finite. By the
Burkholder-Davis-Gundy inequality this is the case if and only if
· ¸
2
E sup |M (t) − M (0)| < ∞.
0<t<T

To be precise, let M (s), t ≤ s ≤ T , be a continuous local L2 -martingale


taking values in Rk . Put M ∗ (s) = supt≤τ ≤s |M (τ )|. Fix 0 < p < ∞. The
Burkholder-Davis-Gundy inequality says that there exist universal finite and
strictly positive constants cp and Cp such that
h i h i
2p p 2p
cp E (M ∗ (s)) ≤ E [hM (·), M (·)i (s)] ≤ Cp E (M ∗ (s)) , t ≤ s ≤ T.
√ (4.79)
If p = 1, then cp = 14 , and if p = 12 , then cp = 18 2. For more details and a
proof see e.g. Ikeda and Watanabe [109].
The following theorem will be employed to prove continuity of solutions
to BSDE’s. It also implies that BSDE’s as considered by us possess at most
unique solutions. The variables (Y, M ) and (Y 0 , M 0 ) attain their values in Rk
Pk
endowed with its Euclidean inner-product hy 0 , yi = j=1 yj0 yj , y 0 , y ∈ Rk .
Processes of the form s 7→ f (s, Y (s), ZM (s)) are progressively measurable
processes whenever the pair (Y, M ) belongs to the space mentioned in (4.80)
mentioned in next theorem.
Theorem 4.30. Let the pairs (Y, M ) and (Y 0 , M 0 ), which belong to the space
¡ ¢ ¡ ¢
L2 [0, T ] × Ω, FT0 , dt × P × M2 Ω, FT0 , P , (4.80)

be solutions to the following BSDE’s:


Z T
Y (t) = Y (T ) + f (s, Y (s), ZM (s)) ds + M (t) − M (T ), and (4.81)
t
Z T
Y 0 (t) = Y 0 (T ) + f 0 (s, Y 0 (s), ZM 0 (s)) ds + M 0 (t) − M 0 (T ) (4.82)
t

for 0 ≤ t ≤ T . In particular this means that the processes (Y, M ) and (Y 0 , M 0 )


are progressively measurable and are square integrable. Suppose that the coef-
ficient f 0 satisfies the following monotonicity and Lipschitz condition. There
exist some positive and finite constants C10 and C20 such that the following
inequalities hold for all 0 ≤ t ≤ T :
196 4 BSDE’s and Markov processes

hY 0 (t) − Y (t), f 0 (t, Y 0 (t), ZM 0 (t)) − f 0 (t, Y (t), ZM 0 (t))i


2 2
≤ (C10 ) |Y 0 (t) − Y (t)| , and (4.83)
2
|f 0 (t, Y (t), ZM 0 (t)) − f 0 (t, Y (t), ZM (t))|
2 d
≤ (C20 ) hM 0 − M, M 0 − M i (t). (4.84)
dt
Then the pair (Y 0 − Y, M 0 − M ) belongs to
¡ ¢ ¡ ¢
S2 Ω, FT0 , P; Rk × M2 Ω, FT0 , P; Rk ,

and there exists a constant C 0 which depends on C10 , C20 and T such that
· ¸
0 2 0 0
E sup |Y (t) − Y (t)| + hM − M, M − M i (T )
0<t<T
"
2
≤ C 0 E |Y 0 (T ) − Y (T )|

Z #
T
0 2
+ |f (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s))| ds . (4.85)
0

Remark 4.31. From the proof it follows that for C 0 we may choose C 0 =
2 2
260eγT , where γ = 1 + 2 (C10 ) + 2 (C20 ) .

By taking Y (T ) = Y 0 (T ) and f (s, Y (s), ZM (s)) = f 0 (s, Y (s), ZM (s)) it also


implies that BSDE’s as considered by us possess at most unique solutions. A
precise formulation reads as follows.
Corollary 4.32. Suppose that the coefficient f satisfies the monotonicity con-
dition (4.83) and the
¡ Lipschitz condition¢(4.84). ¡Then there¢ exists at most one
pair (Y, M ) ∈ L2 [0, T ] × Ω, FT0 , dt × P × M2 Ω, FT0 , P which satisfies the
backward stochastic differential equation in (4.81).

Proof (Proof of Theorem 4.30). Put Y = Y 0 − Y and M = M 0 − M . From


Itô’s formula it follows that
¯ ¯ ­ ® ­ ®
¯Y (t)¯2 + M , M (T ) − M , M (t)
Z T
¯ ¯2 ­ ®
= ¯Y (T )¯ + 2 Y (s), f 0 (s, Y 0 (s), ZM 0 (s)) − f 0 (s, Y (s), ZM 0 (s)) ds
t
Z T ­ ®
+2 Y (s), f 0 (s, Y (s), ZM 0 (s)) − f 0 (s, Y (s), ZM (s))
t
Z T ­ ®
+2 Y (s), f 0 (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s)) ds
t
Z T ­ ®
−2 Y (s), dM (s) . (4.86)
t
4.3 Existence and Uniqueness of solutions to BSDE’s 197

Inserting the inequalities (4.83) and (4.84) into (4.86) shows:


¯ ¯ ­ ® ­ ®
¯Y (t)¯2 + M , M (T ) − M , M (t)
Z T Z T µ ¶1/2
¯ ¯2 0 2
¯ ¯2 ¯ ¯ d ­ ®
¯ ¯
≤ Y (T ) + 2 (C1 ) ¯ ¯
Y (s) ds + 2C2 0 ¯ Y (s)¯ M , M (s) ds
t t ds
Z T
¯ ¯
+2 ¯Y (s)¯ |f 0 (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s))| ds
t
Z T ­ ®
−2 Y (s), dM (s) . (4.87)
t

b2
The elementary inequalities 2ab ≤ 2C20 a2 + and 2ab ≤ a2 + b2 , 0 ≤ a,
2C20
b ∈ R, apply to the effect that
¯ ¯ ¡­ ® ­ ® ¢
¯Y (t)¯2 + 1 M , M (T ) − M , M (t)
2
¯ ¯2 ³ ´Z T ¯ ¯
2
≤ ¯Y (T )¯ + 1 + 2 (C10 ) + 2 (C20 )
2 ¯Y (s)¯2 ds
t
Z T
2
+ |f 0 (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s))| ds
t
Z T Z t
­ ® ­ ®
−2 Y (s), dM (s) + 2 Y (s), dM (s) . (4.88)
0 0

For a concise formulation of the relevant inequalities we introduce the follow-


ing functions and the constant γ:
h¯ ¯2 i
AY (t) = E ¯Y (t)¯ ,
£­ ® ­ ® ¤
AM (t) = E M , M (T ) − M , M (t) ,
h i
2
C(t) = E |f 0 (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s))| ,
Z T Z T
B(t) = AY (T ) + C(s)ds = B(T ) + C(s)ds, and
t t
0 2 2
γ =1+ 2 (C1 ) + 2 (C20 ) . (4.89)

Using the quantities in (4.89) and remembering the fact that the final term
in (4.88) represents a martingale difference, the inequality in (4.88) implies:
Z T
1
AY (t) + AM (t) ≤ B(t) + γ AY (s)ds. (4.90)
2 t

Using (4.90) and employing induction with respect to n yields:


1
AY (t) + AM (t) (4.91)
2
198 4 BSDE’s and Markov processes
Z T n
X Z T n+2
γ k+1 (T − s)k γ (T − s)n+1
≤ B(t) + B(s)ds + AY (s)ds.
t k! t (n + 1)!
k=0

Passing to the limit for n → ∞ in (4.91) results in:


Z T
1
AY (t) + AM (t) ≤ B(t) + γ eγ(T −s) B(s)ds. (4.92)
2 t
RT
Since B(t) = AY (T ) + C(s)ds from (4.92) we infer:
t
à Z T !
1 γ(T −t)
AY (t) + AM (t) ≤ e AY (T ) + C(s)ds . (4.93)
2 t

By first taking the supremum over 0 < t < T and then taking expectations in
(4.88) gives:
· ¸ h¯ ´Z
¯ ¯2 ¯2 i ³ 2 2
T h¯ ¯2 i
E sup ¯Y (t)¯ ≤ E ¯Y (T )¯ + 1 + 2 (C10 ) + 2 (C20 ) E ¯Y (s)¯ ds
0<t<T 0
Z h T i
2
+ E |f 0 (s, Y (s), ZM (s)) − f (s, Y (s), ZM (s))| ds
0
· Z t ¸
­ ®
+ 2E sup Y (s), dM (s) . (4.94)
0<t<T 0
Rt­ ®
The quadratic variation of the martingale t 7→ 0 Y (s), dM (s) is given by
Rt¯ ¯2 ­ ®
the increasing process t 7→ 0 ¯Y (s)¯ d M , M (s). From the Burkholder-
Davis-Gundy inequality (4.79) we know that
Ã !1/2 
· Z t ¸ Z T
­ ® √ ¯ ¯2 ­ ®
E sup Y (s), dM (s) ≤ 4 2E  ¯Y (s)¯ d M , M (s) .
0<t<T 0 0

(4.95)
For more details on the Burkholder-Davis-Gundy inequality,
√ see e.g. Ikeda and
Watanabe [109]. Again we use an elementary inequality 4 2ab ≤ 14 a2 + 32b2
and plug it into (4.95) to obtain
 ÃZ !1/2 
· Z t ¸
­ ® √ ¯ ¯ T ­ ®
E sup Y (s), dM (s) ≤ 4 2E  sup ¯Y (t)¯ d M , M (s) 
0<t<T 0 0<t<T 0
· ¸
1 ¯ ¯2 £­ ® ¤
≤ E sup ¯Y (t)¯ + 32E M , M (T ) .
4 0<t<T
(4.96)

From (4.93) we also infer


4.3 Existence and Uniqueness of solutions to BSDE’s 199
Z Z Ã Z !
T T T
γ AY (s)ds ≤ γ eγ(T −s) AY (T ) + C(ρ)dρ ds
0 0 s
Z T
¡ ¢ ¡ ¢
= eγT − 1 AY (T ) + eγT − eγρ C(ρ)dρ. (4.97)
0

Insertingthe inequalities (4.96) and (4.97) into (4.94) yields:


· ¸ h¯ Z T · ¸
¯ ¯ ¯ i 1 ¯ ¯2
E sup ¯Y (t)¯2 ≤ eγT E ¯Y (T )¯2 + eγT C(s)ds + E sup ¯Y (t)¯
0<t<T 0 2 0<t<T
£­ ® ¤
+ 64E M , M (T ) . (4.98)
From (4.93) we also get
£­ ® ¤
E M , M (T ) = AM (0) (4.99)
à Z ! à Z !
T h¯ ¯2 i T
≤ 2eγT AY (T ) + C(s)ds = 2eγT ¯ ¯
E Y (T ) + C(s)ds .
0 0

A combination of (4.99) and (4.98) results in


· ¸ Ã Z !
¯ ¯2 h¯ ¯2 i T
¯
E sup Y (t) ¯ ≤ 258eγT
E ¯Y (T )¯ + C(s)ds . (4.100)
0<t<T 0

Adding the right- and left-hand sides of (4.98) and (4.99) proves Theorem 4.30
2 2
with the constant C 0 given by C 0 = 260eγT , where γ = 1 + 2 (C10 ) + 2 (C20 ) .
¡ ¢ ¡ ¢
In the definitions 4.28 and 4.29 the spaces S2 [0, T ], Rk and M2 [0, T ], Rk
are defined.
In Theorem 4.34 we will replace the Lipschitz condition (4.101) in Theorem
4.33 for the function Y (s) 7→ f (s, Y (s), ZM (s)) with the (weaker) monotonic-
ity condition (4.123). Here we write y for the variable Y (s) and z for ZM (s).
It is noticed ¡that¢ we consider a probability space (Ω, F, P) with a filtration
(Ft )t∈[0,T ] = Ft0 t∈[0,T ] where FT = F.
¡ ¢∗
Theorem 4.33. Let f : [0, T ] × Rk × M2 → Rk be a Lipschitz continuous
in the sense that there exists finite constants C¡1 and C2 ¢such that ¡ for any¢
two pairs of processes (Y, M ) and (U, N ) ∈ S2 [0, T ], Rk × M2 [0, T ], Rk
the following inequalities hold for all 0 ≤ s ≤ T :
|f (s, Y (s), ZM (s)) − f (s, U (s), ZM (s))| ≤ C1 |Y (s) − U (s)| , and (4.101)
µ ¶1/2
d
|f (s, Y (s), ZM (s)) − f (s, Y (s), ZN (s))| ≤ C2 hM − N, M − N i (s) .
ds
(4.102)
hR i
T 2
Suppose that E 0 |f (s, 0, 0)| ds < ∞. Then there exists a unique pair
¡ ¢ ¡ ¢
(Y, M ) ∈ S2 [0, T ], Rk × M2 [0, T ], Rk such that
200 4 BSDE’s and Markov processes
Z T
Y (t) = ξ + f (s, Y (s), ZM (s)) ds + M (t) − M (T ), (4.103)
t
¡ ¢
where Y (T ) = ξ ∈ L2 Ω, FT , Rk is given and Y (0) = M (0).
For brevity we write
¡ ¢ ¡ ¢
S2 × M2 = S2 [0, T ], Rk × M2 [0, T ], Rk
¡ ¢ ¡ ¢
= S2 Ω, FT0 , P; Rk × M2 Ω, FT0 , P; Rk .

In fact we employ this theorem with the function f replaced with fδ , 0 < δ <
−1
(2C1 ) , where fδ is defined by
³ ´
−1
fδ (s, y, ZM (s)) = f s, (I − δfs,M ) , ZM (s) . (4.104)

Here fs,M (y) = f (s, y, ZM (s)). If the function f is monotone (or one-sided
Lipschitz) in the second variable with constant C1 , and Lipschitz in the second
variable with constant C2 , then the function fδ is Lipschitz in y with Lipschitz
constant δ −1 .
Proof. The proof of the uniqueness part follows from Corollary 4.32.
In order to prove existence we proceed as follows. By induction we define
a sequence (Yn , Mn ) in the space S2 × M2 as follows.
" Z T #
¯
Yn+1 (t) = E ξ + ¯
f (s, Yn (s), ZMn (s)) ds Ft , and (4.105)
t
" Z #
T ¯
Mn+1 (t) = E ξ + f (s, Yn (s), ZMn (s)) ds ¯ Ft , (4.106)
0

Then, since the process s 7→ f (s, Yn (s), Mn (s)) is adapted we have:


Z T
ξ+ f (s, Yn (s), ZMn (s)) ds + Mn+1 (t) − Mn+1 (T )
t
Z " Z #
T T ¯
=ξ+ f (s, Yn (s), ZMn (s)) ds + E ξ + f (s, Yn (s), ZMn (s)) ds ¯ Ft
t 0
" Z #
T ¯
−E ξ+ f (s, Yn (s), ZMn (s)) ds ¯ FT
0
Z " Z #
T T ¯
=ξ+ f (s, Yn (s), ZMn (s)) ds + E ξ + f (s, Yn (s), ZMn (s)) ds ¯ Ft
t 0
Z T
−ξ− f (s, Yn (s), ZMn (s)) ds
0
4.3 Existence and Uniqueness of solutions to BSDE’s 201
" Z # Z
T ¯ t
=E ξ+ f (s, Yn (s), ZMn (s)) ds ¯ Ft − f (s, Yn (s), ZMn (s)) ds
0 0
" Z #
T ¯
=E ξ+ f (s, Yn (s), ZMn (s)) ds ¯ Ft = Yn+1 (t). (4.107)
t

Suppose that the pair (Yn , Mn ) belongs S2 × M2 . We first prove that the pair
(Yn+1 , Mn+1 ) is a member of S2 × M2 . Therefore we fix α = 1 + C12 + C22 ∈ R
where C1 and C2 are as in (4.101) and (4.102) respectively. From Itô’s formula
we get:
Z T Z T
2 2
e2αt |Yn+1 (t)| + 2α e2αs |Yn+1 (s)| ds + e2αs d hMn+1 , Mn+1 i (s)
t t
2
= e2αT |Yn+1 (T )|
Z T
+2 e2αs hYn+1 (s), f (s, Yn (s), ZMn (s)) − f (s, Yn (s), 0)i ds
t
Z T
+2 e2αs hYn+1 (s), f (s, Yn (s), 0) − f (s, 0, 0)i ds
t
Z T Z T
+2 e2αs hYn+1 (s), f (s, 0, 0)i ds − 2 e2αs hYn+1 (s), dMn+1 (s)i .
t t
(4.108)

We employ (4.101) and (4.102) to obtain from (4.108):


Z T Z T
2αt 2 2αs 2
e |Yn+1 (t)| + 2α e |Yn+1 (s)| ds + e2αs d hMn+1 , Mn+1 i (s)
t t
Z T µ ¶1/2
2 d
≤ e2αT |Yn+1 (T )| + 2C2 e2αs |Yn+1 (s)| hMn , Mn i (s) ds
t ds
Z T
+ 2C1 e2αs |Yn+1 (s)| |Yn (s)| ds
t
Z T Z T
+2 e2αs |Yn+1 (s)| |f (s, 0, 0)| ds − 2 e2αs hYn+1 (s), dMn+1 (s)i .
t t
(4.109)

b2
The elementary inequalities 2ab ≤ 2Cj a2 + , a, b ∈ R, j = 0, 1, 2, with
2Cj
C0 = 1, in combination with (4.109) yield
Z T Z T
2αt 2 2αs 2
e |Yn+1 (t)| + 2α e |Yn+1 (s)| ds + e2αs d hMn+1 , Mn+1 i (s)
t t
Z T Z T
2 2 1
≤ e2αT |Yn+1 (T )| + 2C22 e2αs |Yn+1 (s)| ds + e2αs d hMn , Mn i (s)
t 2 t
202 4 BSDE’s and Markov processes
Z T Z T
2 1 2
+ 2C12 e2αs |Yn+1 (s)| ds + e2αs |Yn (s)| ds
t 2 t
Z T Z T
2 2
+ e2αs |Yn+1 (s)| ds + e2αs |f (s, 0, 0)| ds
t t
Z T
−2 e2αs hYn+1 (s), dMn+1 (s)i , (4.110)
t

and hence by the choice of α from (4.110) we infer:


Z T Z T
2 2
e2αt |Yn+1 (t)| + e2αs |Yn+1 (s)| ds + e2αs d hMn+1 , Mn+1 i (s)
t t
Z T
+2 e2αs hYn+1 (s), dMn+1 (s)i
0
Z T Z T
2 1 1 2
≤ e2αT |Yn+1 (T )| + e2αs d hMn , Mn i (s) + e2αs |Yn (s)| ds
2 t 2 t
Z T Z t
2
+ e2αs |f (s, 0, 0)| ds + 2 e2αs hYn+1 (s), dMn+1 (s)i . (4.111)
t 0

The following steps can ¡ be justified


¢ by observing that the process Yn+1 be-
longs to the space L2 Ω, FT0 , P , and that sup0≤t≤T |Yn+1 (t)| < ∞ P-almost
surely. By stopping the process Yn+1 (t) at the stopping time τN being the first
time t ≤ T that |Yn+1 (t)| exceeds N . In inequality (4.111) we then replace
t by t ∧ τN and proceed as below with the stopped processes instead of the
processes itself. Then we use the monotone convergence theorem to obtain in-
equality
hR (4.114). By the same approximation
i argument we may assume that
T 2αs
E t e hYn+1 (s), dMn+1 (s)i = 0. Hence (4.111) implies that
" Z Z #
T T
2 2
E e2αt |Yn+1 (t)| + e2αs |Yn+1 (s)| ds + e2αs d hMn+1 , Mn+1 i (s)
t t
"Z #
h i 1 T
2αT 2 2αs
≤e E |Yn+1 (T )| + E e d hMn , Mn i (s)
2 t
"Z #
T
1 2
+ E e2αs |Yn (s)| ds
2 t
"Z #
T
2αs 2
+E e |f (s, 0, 0)| ds < ∞. (4.112)
t

Invoking the Burkholder-Davis-Gundy inequality and applying the equality


¿Z · Z · À
2αs 2αs
e hYn+1 (s), dMn+1 (s)i , e hYn+1 (s), dMn+1 (s)i (t)
0 0
Z t
2
= e4αs |Yn+1 (s)| d hMn+1 , Mn+1 i (s)
0
4.3 Existence and Uniqueness of solutions to BSDE’s 203

to (4.111) yields:
· ¸
2αt 2
E sup e |Yn+1 (t)|
0<t<T
"Z #
h i 1 T
2αT 2 2αs
≤e E |Yn+1 (T )| + E e d hMn , Mn i (s)
2 0
"Z #
T
1 2αs 2
+ E e |Yn (s)| ds
2 0
"Z # "Z #
T T
2αs 2 2αs
+E e |f (s, 0, 0)| ds − 2E e hYn+1 (s), dMn+1 (s)i
0 0
Ã !1/2 
√ Z T
 2 
+ 8 2E e4αs |Yn+1 (s)| d hMn+1 , Mn+1 i (s)
0

hR i
T
(without loss of generality assume that E 0
e2αs hYn+1 (s), dMn+1 (s)i = 0)
"Z #
h i 1 T
2αT 2 2αs
≤e E |Yn+1 (T )| + E e d hMn , Mn i (s)
2 0
"Z # "Z #
T T
1 2αs 2 2αs 2
+ E e |Yn (s)| ds + E e |f (s, 0, 0)| ds
2 0 0
 ÃZ !1/2 
√ T
+ 8 2E  sup eαt |Yn+1 (t)| e2αs d hMn+1 , Mn+1 i (s) 
0<t<T 0

√ a2
(8 2ab ≤ + 64b2 , a, b ∈ R)
2
"Z #
h i 1 T
2αT 2 2αs
≤e E |Yn+1 (T )| + E e d hMn , Mn i (s)
2 0
"Z # "Z #
T T
1 2αs 2 2αs 2
+ E e |Yn (s)| ds + E e |f (s, 0, 0)| ds
2 0 0
· ¸ "Z #
T
1 2αt 2 2αs
+ E sup e |Yn+1 (t)| + 64E e d hMn+1 , Mn+1 i (s)
2 0<t<T 0

(apply (4.112))
"Z #
h i 65 T
2αT 2 2αs
≤ 65e E |Yn+1 (T )| + E e d hMn , Mn i (s)
2 0
204 4 BSDE’s and Markov processes
"Z #
T
65 2αs 2
+ E e |Yn (s)| ds
2 0
"Z # · ¸
T
2 1 2
+ 65E e2αs |f (s, 0, 0)| ds + E sup e2αt |Yn+1 (t)| . (4.113)
0 2 0<t<T

From (4.113) it follows that


· ¸
2αt 2
E sup e |Yn+1 (t)|
0<t<T
"Z #
h i T
2 2
≤ 130e2αT E |Yn+1 (T )| + 130E e2αs |f (s, 0, 0)| ds (4.114)
0
"Z # "Z #
T T
2αs 2αs 2
+ 65E e d hMn , Mn i (s) + 65E e |Yn (s)| ds < ∞.
0 0

From (4.112) and (4.114) it follows that the pair (Yn+1 , Mn+1 ) belongs to
S2 × M2 .
Another application of Itô’s formula shows:
Z T
2αt 2 2
e |Yn+1 (t) − Yn (t)| + 2α e2αs |Yn+1 (s) − Yn (s)| ds
t
Z T
+ e2αs d hMn+1 − Mn , Mn+1 − Mn i (s)
t
2αT 2
=e |Yn+1 (T ) − Yn (T )|
Z T ­ ¡ ¢®
+2 e2αs 4Yn (s), f (s, Yn (s), ZMn (s)) − f s, Yn (s), ZMn−1 (s) ds
t
Z T ­ ¡ ¢ ¡ ¢®
+2 e2αs 4Yn (s), f s, Yn (s), ZMn−1 (s) − f s, Yn−1 (s), ZMn−1 (s) ds
t
Z T
−2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i , (4.115)
t

where for brevity we wrote 4Yn (s) = Yn+1 (s) − Yn (s). From (4.101), (4.102),
and (4.115) we infer
Z T
2αt 2 2
e |Yn+1 (t) − Yn (t)| + 2α e2αs |Yn+1 (s) − Yn (s)| ds
t
Z T
+ e2αs d hMn+1 − Mn , Mn+1 − Mn i (s)
t
2αT 2
≤e |Yn+1 (T ) − Yn (T )|
Z T µ ¶1/2
2αs d
+ 2C2 e |Yn+1 (s) − Yn (s)| hMn − Mn−1 , Mn − Mn−1 i (s) ds
t ds
4.3 Existence and Uniqueness of solutions to BSDE’s 205
Z T
+ 2C1 e2αs |Yn+1 (s) − Yn (s)| |Yn (s) − Yn−1 (s)| ds
t
Z T
−2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i
t
Z T
2 2
≤ e2αT |Yn+1 (T ) − Yn (T )| + 2C22 e2αs |Yn+1 (s) − Yn (s)| ds
t
Z T
1
+ e2αs d hMn − Mn−1 , Mn − Mn−1 i (s)
2 t
Z T Z T
2 1 2
+ 2C12 e2αs |Yn+1 (s) − Yn (s)| ds + e2αs |Yn (s) − Yn−1 (s)| ds
t 2 t
Z T
−2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i . (4.116)
t

Since α = 1 + C12 + C22 the inequality in (4.116) implies:


Z T
2 2
e2αt |Yn+1 (t) − Yn (t)| + 2 e2αs |Yn+1 (s) − Yn (s)| ds
t
Z T
+ e2αs d hMn+1 − Mn , Mn+1 − Mn i (s)
t
2
≤ e2αT |Yn+1 (T ) − Yn (T )|
Z
1 T 2αs
+ e d hMn − Mn−1 , Mn − Mn−1 i (s)
2 t
Z
1 T 2αs 2
+ e |Yn (s) − Yn−1 (s)| ds
2 t
Z T
−2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i
0
Z t
+2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i . (4.117)
0

Upon taking expectations in (4.117) we see


"Z #
h i T
2αt 2 2αs 2
e E |Yn+1 (t) − Yn (t)| + 2E e |Yn+1 (s) − Yn (s)| ds
t
"Z #
T
2αs
+E e d hMn+1 − Mn , Mn+1 − Mn i (s)
t
h i
2
≤ e2αT E |Yn+1 (T ) − Yn (T )|
"Z #
T
1 2αs
+ E e d hMn − Mn−1 , Mn − Mn−1 i (s)
2 t
206 4 BSDE’s and Markov processes
"Z #
T
1 2αs 2
+ E e |Yn (s) − Yn−1 (s)| ds . (4.118)
2 t

In particular it follows that


"Z #
T
2αs 2
2E e |Yn+1 (s) − Yn (s)| ds
t
"Z #
T
+E e2αs d hMn+1 − Mn , Mn+1 − Mn i (s)
t
"Z #
T
1 2αs 2
≤ E e |Yn (s) − Yn−1 (s)| ds
2 t
"Z #
T
1
+ E e2αs d hMn − Mn−1 , Mn − Mn−1 i (s) ,
2 t

provided that Yn+1 (T ) = Yn (T ). As a consequence we see that the sequence


(Yn , Mn ) converges with respect to the norm k·kα defined by
°µ ¶°2 "Z Z T #
° Y ° T
° ° 2αs 2 2αs
° M ° =E e |Y (s)| ds + e d hM, M i (s) .
α 0 0

Employing a similar reasoning as the one we used to obtain (4.113) and (4.114)
from (4.117) we also obtain:
2
sup e2αt |Yn+1 (t) − Yn (t)|
0≤t≤T
2
≤ e2αT |Yn+1 (T ) − Yn (T )|
Z
1 T 2αs
+ e d hMn − Mn−1 , Mn − Mn−1 i (s)
2 0
Z
1 T 2αs 2
+ e |Yn (s) − Yn−1 (s)| ds
2 0
Z T
−2 e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i
0
Z t
+ 2 sup e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i . (4.119)
0≤t≤T 0

By taking expectations in (4.119), and invoking the Burkholder-Davis-Gundy


inequality (4.79) for p = 12 we obtain:
· ¸
2
E sup e2αt |Yn+1 (t) − Yn (t)|
0≤t≤T
h i
2
≤ e2αT E |Yn+1 (T ) − Yn (T )|
4.3 Existence and Uniqueness of solutions to BSDE’s 207
"Z #
T
1
+ E e2αs d hMn − Mn−1 , Mn − Mn−1 i (s)
2 0
"Z #
T
1 2αs 2
+ E e |Yn (s) − Yn−1 (s)| ds
2 0
· Z t ¸
+ 2E sup e2αs hYn+1 (s) − Yn (s), dMn+1 (s) − dMn (s)i
0≤t≤T 0
h i
2αT 2
≤e E |Yn+1 (T ) − Yn (T )|
"Z #
T
1 2αs
+ E e d hMn − Mn−1 , Mn − Mn−1 i (s)
2 0
"Z #
T
1 2αs 2
+ E e |Yn (s) − Yn−1 (s)| ds
2 0
Ã !1/2 
√ Z T
2
+ 8 2E  e4αs |Yn+1 (s) − Yn (s)| d hMn+1 − Mn , Mn+1 − Mn i (s) 
0

(insert the definition of k·kα )

h i 1° µ ¶°
° Yn − Yn−1 °2
E |Yn+1 (T ) − Yn (T )| + ° °
2αT 2
≤e
2 ° Mn − Mn−1 °α


+ 8 2E  sup eαs |Yn+1 (s) − Yn (s)|
0≤s≤T

ÃZ !1/2 
T
× e2αs d hMn+1 − Mn , Mn+1 − Mn i (s) 
0

√ 1
(8 2ab ≤ a2 + 64b2 , a, b ∈ R)
2
h i 1° µ ¶°
° Yn − Yn−1 °2
≤e 2αT 2
E |Yn+1 (T ) − Yn (T )| + ° ° °
2 Mn − Mn−1 °α
· ¸
1 2αs 2
+ E sup e |Yn+1 (s) − Yn (s)|
2 0≤s≤T
"Z #
T
+ 64E e2αs d hMn+1 − Mn , Mn+1 − Mn i (s) . (4.120)
0

Employing inequality (4.118) (with t = 0) together with (4.120), and the


definition of the norm k·kα yields the inequality
208 4 BSDE’s and Markov processes
· ¸ "Z #
T
2αt 2 2αs 2
E sup e |Yn+1 (t) − Yn (t)| + 129E e |Yn+1 (s) − Yn (s)| ds
0≤t≤T 0
"Z #
T
1 2αs
+ E e d hMn+1 − Mn , Mn+1 − Mn i (s)
2 0

131 2αT h i 131 ° µ ¶°


° Yn − Yn−1 °2
≤ e
2
E |Yn+1 (T ) − Yn (T )| + ° °
2 4 ° Mn − Mn−1 °α
· ¸
1 2αs 2
+ E sup e |Yn+1 (s) − Yn (s)| . (4.121)
2 0≤s≤T

(In order to justify the transition from (4.119) to (4.121) like in passing from
inequality (4.111) to (4.114) a stopping time argument might be required.)
Consequently, from (4.121) we get
· ¸
2αt 2
E sup e |Yn+1 (t) − Yn (t)|
0≤t≤T
"Z #
T
2αs
+E e d hMn+1 − Mn , Mn+1 − Mn i (s)
0
h i °µ ¶°2
2αT 2 131 ° °
° Yn − Yn−1 ° .
≤ 131e E |Yn+1 (T ) − Yn (T )| + ° (4.122)
2 Mn − Mn−1 °α
£ ¯ ¤
Since by definition Yn (T ) = E ξ ¯ FTT for all n ∈ N, this sequence also con-
verges with respect to the norm k·kS2 ×M2 defined by
°µ ¶°2 · ¸
° Y °
° ° = E sup |Y (s)|
2
+ E [hM, M i (T ) − hM, M i (0)] ,
° M ° 2 2 0<s<T
S ×M

because
" Z #
T ¯ 0
Yn+1 (0) = Mn+1 (0) = E ξ + fn (s, Yn (s), ZMn (s)) ds ¯ F0 , n ∈ N.
0

This concludes the proof of Theorem 4.33.


In the following theorem we replace the Lipschitz condition (4.101) in Theorem
4.33 for the function Y (s) 7→ f (s, Y (s), ZM (s)) with the (weaker) monotonic-
ity condition (4.123). Here we write y for the variable Y (s) and z for ZM (s):
see Theorem 4.34 in Chapter 4.
¡ ¢∗
Theorem 4.34. Let f : [0, T ]×Rk × M2 → Rk be monotone in the variable
y and Lipschitz in z. More precisely, suppose that there exist finite constants
C1 ¡ and C2 such
¢ that
¡ for any¢ two pairs of processes (Y, M ) and (U, N ) ∈
S2 [0, T ], Rk × M2 [0, T ], Rk the following inequalities hold for all 0 ≤ s ≤
T:
4.3 Existence and Uniqueness of solutions to BSDE’s 209
2
hY (s) − U (s), f (s, Y (s), ZM (s)) − f (s, U (s), ZM (s))i ≤ C1 |Y (s) − U (s)| ,
(4.123)
µ ¶1/2
d
|f (s, Y (s), ZM (s)) − f (s, Y (s), ZN (s))| ≤ C2 hM − N, M − N i (s) ,
ds
(4.124)

and

|f (s, Y (s), 0)| ≤ f (s) + K |Y (s)| . (4.125)


hR ¯ ¯2 i
T
If E 0 ¯f (s)¯ ds < ∞, then there exists a unique pair
¡ ¢ ¡ ¢
(Y, M ) ∈ S2 [0, T ], Rk × M2 [0, T ], Rk

such that
Z T
Y (t) = ξ + f (s, Y (s), ZM (s)) ds + M (t) − M (T ), (4.126)
t
¡ ¢
where Y (T ) = ξ ∈ L2 Ω, FT , Rk is given and where Y (0) = M (0).
In order to prove Theorem 4.34 we need the following proposition, the proof
of which uses the monotonicity condition (4.123) in an explicit manner.
¡ ¢
Proposition 4.35. Suppose that for every ξ ∈ L2 Ω, FT0 , P and M ∈ M2
there exists a pair (Y, N ) ∈ S2 × M2 such that
Z T
Y (t) = ξ + f (s, Y (s), ZM (s)) ds + N (t) − N (T ). (4.127)
t
¡ ¢
Then for every ξ ∈ L Ω, FT0 , P there exists a unique pair (Y, M ) ∈ S2 × M2
2

which satisfies (4.126).


The following proposition can be viewed as a consequence of Theorem 12.4
in [98]. The result is due to Burrage and Butcher [46] and Crouzeix [64]. The
obtained constants are somewhat different from ours.
Proposition 4.36. Fix a martingale M ∈ M2 , and choose δ > 0 in such
a way that δC1 < 1. Here C1 is the constant which occurs in inequality
(4.123). Choose, for given y ∈ R k e k
³ , the stochastic
´ variable Y (t) ∈ R in
such a way that y = Ye (t) − δf t, Ye (t), ZM (t) . Then the mapping y 7→
³ ´
f t, Ye (t), ZM (t) is Lipschitz continuous with a Lipschitz constant which is
µ ¶
1 δC1
equal to max 1, . Moreover, the mapping y 7→ I − δf (t, y, ZM (t))
δ 1 − δC1
is surjective and has a Lipschitz continuous inverse with Lipschitz constant
1
.
1 − δC1
210 4 BSDE’s and Markov processes

k k
Proof (Proof of Proposition 4.36). Let the pair
³ (y1 , y2 ) ∈ ´ R × R and the
pair of Rk × Rk -valued stochastic variables Ye1 (t), Ye2 (t) be such that the
following equalities are satisfied:
³ ´ ³ ´
y1 = Ye1 (t) − δf t, Ye1 (t), ZM (t) and y2 = Ye2 (t) − δf t, Ye2 (t), ZM (t) .
(4.128)
We have to show that there exists a constant C(δ) such that
¯ ³ ´ ³ ´¯
¯ ¯
¯f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t) ¯ ≤ C(δ) |y2 − y1 | . (4.129)

In order to achieve this we will exploit the inequality:


D ³ ´ ³ ´E
Ye2 (t) − Ye1 (t), f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t)
¯ ¯2
¯ ¯
≤ C1 ¯Ye2 (t) − Ye1 (t)¯ . (4.130)

Inserting the equalities in (4.128) into (4.130) results in


D ³ ´ ³ ´E
y2 − y1 , f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t)
¯ ³ ´ ³ ´¯2
¯ ¯
+ δ ¯f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t) ¯
D ³ ´ ³ ´E
≤ C1 |y2 − y1 | + 2δC1 y2 − y1 , f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t)
2

¯ ³ ´ ³ ´¯2
¯ ¯
+ C1 δ 2 ¯f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t) ¯ . (4.131)

Notice that (4.131) is equivalent to:


¯ ³ ´ ³ ´¯2
¯ ¯
δ ¯f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t) ¯
2
≤ C1 |y2 − y1 |
µ ¶
1 D ³ ´ ³ ´E
+ 2 δC1 − y2 − y1 , f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t)
2
¯ ³ ´ ³ ´¯2
¯ ¯
+ C1 δ 2 ¯f t, Ye2 (t), ZM