Van Casteren
– Monograph –
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
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”.
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.
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Part I
I ∗ (f ) = lim I (fn ) , f ∈ H ∗ ,
n→∞
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
(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 ).
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:
Ã ! Ã ! Ã ! Ãm !
[ [ [ [
∗ ∗
µ An ≤ µ Gn = µ Gn = sup µ Gn
m∈N n=1
n∈N n∈N n∈N
Ã 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
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
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
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→∞
for B ∈ σ(H).
This finishes the proof of Theorem 1.2.
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.
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.
+ 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
Ã n !
X 1
<Λ (g1 ) ≤ <Λ uk + ε, for n ≥ nε . (1.19)
2
k=1
+ 1
Λ1 (f1 ) = (<Λ) (f1 ) ≤ < (Λ (g1 )) + ε
Ã 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
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
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 ).
` (ε)
Put K(ε) = ∩∞ n=1 ∪k=1 Bk,n . Then K(ε) is closed, and thus complete, and
n
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.
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
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
(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 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
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
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
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
|·|
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
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)
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)| 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
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
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
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
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
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.
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.
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).
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),
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
<P (s, t)f (x) = [P (s, t)<f ] (x) ≤ sup <f (y) ≤ k<f k∞ . (1.66)
y∈E
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
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
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
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
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
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
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
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 .
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,
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
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
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
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
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
L0 = {L0 (s) : 0 ≤ s ≤ T } ,
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)
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
{B 7→ P (τ, x; t, B) : 0 ≤ τ ≤ t ≤ T }
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
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
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
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→∞
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→∞
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
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
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
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
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→∞
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 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
lim d (y, X (tnk )) = d (y, X(t)) , Pτ,x -almost surely for all y ∈ E. (2.12)
k→∞
h i
− 2<Eτ,x f (ρ, X(ρ))Eρ,X(ρ) [f (s, X(s))] . (2.16)
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)
e
X(ρ) = lim e ) = X(T ).
X(t), ρ ∈ [0, T ), X(T (2.21)
t↓ρ, t∈D∩(ρ,T ], t>ρ
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
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
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
(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
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
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)
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
We also put
Hε (D ∩ [0, T ])
= sup {Hε (U ) : U ⊂ D ∩ [0, T ], U contains an even number of elements} .
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)
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.
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 ]
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)
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
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:
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 ]
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
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]
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
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→∞
In (2.77) we let n → ∞ and use the left continuity of the propagator (see
property (v) in Definition 1.24) to conclude
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
(Markov property)
£ £ ¯ ¤¯ τ ¤
= lim Eτ,x Eτ,x f (ρ, X(ρ)) ¯ Fsτn ¯ Fs+
n→∞
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.
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)
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
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
Fix ε > 0. For α0 > 0 and f ∈ Cb ([0, T ] × E) fixed, there exists a function
g ∈ Cb ([0, T ] × E) such that
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
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.
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
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.
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
In particular such variables are Pτ,x -almost surely functions of the space-time
variable (t, X(t)).
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
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 ))]
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
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
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
exists for all sequences (uj )j∈N in Cb ([0, T ] × E). Since, in addition,
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)
h i
≥ (1 − ε) Pn (X(s))s∈Q∩[0,sk ] ⊂ E , (2.110)
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
¯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 `→∞
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
Again using (2.108), (2.109) and (2.110) it follows that the the path
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
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
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.
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
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
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
(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
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
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
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
"Ã
= 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
"Ã
= 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
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)
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 ))
+ 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)
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
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
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
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
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
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
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
In this section we will discuss in more detail the generators of the the time-
space Markov process (see (1.75):
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
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
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:
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).
lim <f (xn ) = sup <f (x), and lim <Lf (xn ) ≤ 0. (3.12)
n→∞ x∈E n→∞
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
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
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.
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:
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.
inf {g ≥ f 1K : (λI − L) g ≥ 0}
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
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
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
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
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
kuS0 (t)f k∞ ≤ kvf k∞ for all f ∈ Cb ([0, ∞) × E) and all t ∈ [0, ∞). (3.42)
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
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
Hence g = 0.
(a2 ) Let (fn )n∈N ⊂ D(L) be any sequence with the following properties:
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
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→∞
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
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
Finally pick λ ∈ C with <λ > 0. From (3.58) and (3.59) we infer
(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
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
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
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)
(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
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) ,
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)
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→∞
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
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
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
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
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
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
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
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
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
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)
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→∞
(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
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
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
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
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
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
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
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)
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
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
We also have
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:
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
Xµ
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
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
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
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
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
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
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)
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
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
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
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 ,
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
= 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 .
(Markov property)
(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
(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.
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
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
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
=0 (3.228)
(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.
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
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
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
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
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.
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
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
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
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)
eG
D S eG
≤D S eK
≤D S
,
n n+1
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.
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
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.
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.
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)
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
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 ∞ Ã¡ ¢ !
−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
or equal to 2), the corresponding martingale problem, and being the generator
of a Feller-Dynkin semigroup.
Part II
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
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
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
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 ] .
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
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.
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].
∂ ¡ ¢
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
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:
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).
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)
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
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
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
(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
(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
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
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
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)
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,
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
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,
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 }
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
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
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:
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.
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)
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
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
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 ) .
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
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
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)
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
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)
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
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 (4.112) 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
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
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
(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
and
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
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)
¯ ³ ´ ³ ´¯2
¯ ¯
+ C1 δ 2 ¯f t, Ye2 (t), ZM (t) − f t, Ye1 (t), ZM (t) ¯ . (4.131)