Applications
A. S.
Ust unel
2
Introduction
The aim of this book is to give a rigorous introduction for the graduate
students to Analysis on Wiener space, a subject which has grown up very
quickly these recent years under the new impulse of the Stochastic Calculus
of Variations of Paul Malliavin (cf. [55]). A portion of the material exposed
is our own research, in particular, with Moshe Zakai and Denis Feyel for the
rest we have used the works listed in the bibliography.
The origin of this book goes back to a series of seminars that I had given
in Bilkent University of Ankara in the summer of 1987 and also during the
spring and some portion of the summer of 1993 at the Mathematics Insti
tute of Oslo University and a graduate course dispensed at the University of
Paris VI. An initial and rather naive version of these notes has been pub
lished in Lecture Notes in Mathematics series of Springer at 1995. Since then
we have assisted to a very quick development and progress of the subject in
several directions. In particular, its use has been remarked by mathemati
cal economists. Consequently I have decided to write a more complete text
with additional contemporary applications to illustrate the strength and the
applicability of the subject. Several new results like the logarithmic Sobolev
inequalities, applications of the capacity theory to the local and global dier
entiability of Wiener functionals, probabilistic notions of the convexity and
logconcavity, the Monge and the MongeKantorovitch measure transporta
tion problems in the innite dimensional setting and the analysis on the path
space of a compact Lie group are added.
Although some concepts are given in the rst chapter, I assumed that the
students had already acquired the notions of stochastic calculus with semi
martingales, Brownian motion and some rudiments of the theory of Markov
processes.
The second chapter deals with the denition of the (socalled) Gross
Sobolev derivative and the OrnsteinUhlenbeck operator which are indispens
able tools of the analysis on Wiener space. In the third chapter we begin the
proof of the Meyer inequalities, for which the hypercontractivity property
of the OrnsteinUhlenbeck semigroup is needed. We expose this last topic
in the fourth chapter and give the classical proof of the logarithmic Sobolev
inequality of L. Gross for the Wiener measure. In chapter V, we complete
the proof of Meyer inequalities and study the distribution spaces which are
dened via the OrnsteinUhlenbeck operator. In particular we show that
the derivative and divergence operators extend continuously to distribution
spaces. In the appendix we indicate how one can transfer all these results
to arbitrary abstract Wiener spaces using the notion of time associated to a
3
continuous resolution of identity of the underlying CameronMartin space.
Chapter VI begins with an extension of Clarks formula to the distribu
tions dened in the preceding chapter. This formula is applied to prove the
classical 01law and as an application of the latter, we prove the positivity
improving property of the OrnsteinUhlenbeck semigroup. We then show
that the functional composition of a nondegenerate Wiener functional with
values in IR
n
, (in the sense of Malliavin) with a realvalued smooth function
on IR
n
can be extended when the latter is a tempered distribution if we look
at to the result as a distribution on the Wiener space. This result contains the
fact that the probability density of a nondegenerate functional is not only
C
. . . . . . . . . . . . . . . . . . . . . . 202
13.3 OrnsteinUhlenbeck operator and the Wiener chaos . . . . . . 203
13.4 Some useful formulea . . . . . . . . . . . . . . . . . . . . . . . 207
13.5 Right derivative . . . . . . . . . . . . . . . . . . . . . . . . . . 210
13.6 Quasiinvariance . . . . . . . . . . . . . . . . . . . . . . . . . . 214
13.7 Anticipative transformations . . . . . . . . . . . . . . . . . . . 220
13.7.1 Degree type results . . . . . . . . . . . . . . . . . . . . 225
Chapter 1
Introduction to Stochastic
Analysis
This chapter is devoted to the basic results about the Wiener measure, Brow
nian motion, construction of the Ito stochastic integral, CameronMartin and
Girsanov theorems, representation of the Wiener functionals with stochastic
integrals and the ItoWiener chaos decomposition which results from it. The
proofs are rather sketchy whenever they are given; for complete treatment
of these results we refer the reader to the excellent references given in the
bibliography.
1.1 The Brownian Motion and the Wiener
Measure
Let W = C
0
([0, 1]), dene W
t
as to be the coordinate functional, i.e., for
w W and t [0, 1], let W
t
(w) = w(t) . If we note by B
t
= W
s
; s t,
then, the following theorem is wellknown (cf. for instance [81]):
Theorem 1.1.1 There is one and only one measure on W which satises
the following properties:
i) w W : W
0
(w) = 0 = 1,
ii) For any f c
b
(IR), the stochastic process process
(t, w) f(W
t
(w))
1
2
_
t
0
f(W
s
(w))ds
is a (B
t
, )martingale, where denotes the Laplace operator. is
called the (standard) Wiener measure.
1
2 Brownian Motion
From Theorem 1.1.1, it follows that, for t > s,
E
_
e
i(W
t
W
s
)
[B
s
_
= exp
_
1
2
2
(t s)
_
,
hence (t, w) W
t
(w) is a continuous additive process (i.e.,a process with
independent increments) and (W
t
; t [0, 1]) is also a continuous martingale.
1.2 Stochastic Integration
The stochastic integration with respect to the Brownian motion is rst de
ned on the adapted step processes and then extended to their completion
by isometry. A mapping K : [0, 1] W IR is called a step process if it can
be represented in the following form:
K
t
(w) =
n
i=1
a
i
(w) 1
[t
i
,t
i+1
)
(t), a
i
(w) L
2
(B
t
i
) .
For such a step process, we dene its stochastic integral with respect to the
Brownian motion, which is denoted by
I(K) =
_
1
0
K
s
dW
s
(w)
as to be
n
i=1
a
i
(w)
_
W
t
i+1
(w) W
t
i
(w)
_
.
Using the the independence of the increments of (W
t
, t [0, 1]), it is easy to
see that
E
_
_
1
0
K
s
dW
s
2
_
= E
_
1
0
[K
s
[
2
ds ,
i.e., I is an isometry from the adapted step processes into L
2
(), hence it
has a unique extension as an isometry from
L
2
([0, 1] W, /, dt d)
I
L
2
()
where / denotes the sigma algebra on [0, 1] W generated by the adapted,
left (or right) continuous processes. The extension of I(K) is called the
stochastic integral of K and it is denoted as
_
1
0
K
s
dW
s
. If we dene
I
t
(K) =
_
t
0
K
s
dW
s
Ito Formula 3
as
_
1
0
1
[0,t]
(s)K
s
dW
s
,
it follows from the Doob inequality that the stochastic process t I
t
(K) is a
continuous, square integrable martingale. With some localization techniques
using stopping times, I can be extended to any adapted process K such that
_
1
0
K
2
s
(w)ds < a.s. In this case the process t I
t
(K) becomes a local
martingale, i.e., there exists a sequence of stopping times increasing to one,
say (T
n
, n IN) such that the process t I
tT
n
(K) is a (square integrable)
martingale. Vector (i.e. IR
n
) valued Brownian motion is dened as a pro
cess whose components are independent, realvalued Brownian motions. A
stochastic process (X
t
, t 0) with values in a nite dimensional Euclidean
space is called an Ito process if it can be represented as
X
t
= X
0
+
_
t
0
a
s
dW
s
+
_
t
0
b
s
ds ,
where (W
t
, t 0) is a vector valued Brownian motion and a and b are
respectively matrix and vector valued, adapted, measurable processes with
_
t
0
([a
s
[
2
+ [b
s
[)ds < almost surely for any t 0. In the sequel the no
tation
_
t
0
H
s
dX
s
will mean
_
t
0
H
s
a
s
dW
s
+
_
t
0
H
s
b
s
ds, we shall also denote by
([X, X]
t
, t 0) the DoobMeyer process dened as
[X, X]
t
=
_
t
0
trace (a
s
a
s
)ds .
This is the unique increasing process such that ([X
t
[
2
[X, X]
t
, t 0) is a
(continuous) local martingale. It can be calculated as the limit of the sums
lim
n
i=1
_
[X
t
i+1
[
2
[X
t
i
[
2
_
= lim
n
i=1
_
[X
t
i+1
[ [X
t
i
[
_
2
= lim
n
i=1
E
_
[X
t
i+1
[
2
[X
t
i
[
2
[T
t
i
_
,
where the limit is taken as the length of the partition t
1
, . . . , t
n+1
of [0, t],
dened by sup
i
[t
i+1
t
i
[, tends to zero.
1.3 Ito formula
The following result is one of the most important applications of the stochas
tic integration:
4 Brownian Motion
Theorem 1.3.1 Let f C
2
(IR) and let (X
t
, t [0, 1]) be an Ito process,
i.e.,
X
t
= X
0
+
_
t
0
K
r
dW
r
+
_
t
0
U
r
dr
where X
0
is B
0
measurable, K and U are adapted processes with
_
1
0
_
[K
r
[
2
+[U
r
[
_
dr < (1.3.1)
almost surely. Then
f(X
t
) = f(X
0
) +
_
t
0
f
t
(X
s
)K
s
dW
s
+
1
2
_
t
0
f
tt
(X
s
)K
2
s
ds
+
_
t
0
f
t
(X
r
)U
r
dr .
Remark 1.3.2 This formula is also valid in the several dimensional case. In
fact, if K is and U are adapted processes with values in IR
n
IR
m
and IR
m
respectively whose components are satisfying the condition (1.3.1), then we
have, for any f C
2
(IR
m
),
f(X
t
) = f(X
0
) +
_
t
0
i
f(X
r
)K
ij
(r)dW
j
r
+
_
t
0
i
f(X
r
)U
i
(r)dr
+
1
2
_
t
0
2
ij
f(X
r
)(K
r
K
r
)
ij
dr
almost surely.
To prove the Ito formula we shall proceed by
Lemma 1.3.3 Let X = (X
t
, t 0) and Y = (Y
t
, t 0) be two Ito real
valued processes, then
X
t
Y
t
= X
0
Y
0
+
_
t
0
X
s
dY
s
+
_
t
0
Y
s
dX
s
+ [X, Y ]
t
, (1.3.2)
almost surely, where [X, Y ] denotes the DoobMeyer process. In particular
X
2
t
= X
2
0
+ 2
_
t
0
X
s
dX
s
+ [X, X]
t
. (1.3.3)
Proof: Evidently it suces to prove the relation (1.3.3), since we can obtain
(1.3.2) via a polarization argument. Since X has almost surely continuous
trajectories, using a stopping time argument we can assume that X is almost
surely bounded. Assume now that t
1
, . . . , t
n
is a partition of [0, t] and
Ito Formula 5
denote by (M
t
, t 0) the local martingale part and by (A
t
, t 0) the nite
variaton part of X. We have
X
2
t
X
2
0
=
n
k=1
(X
t
k
X
t
k1
)
2
+ 2
n
k=1
X
t
k1
(X
t
k
X
t
k1
) (1.3.4)
=
n
k=1
(X
2
t
k
X
2
t
k1
) + 2
n
k=1
X
t
k1
(M
t
k
M
t
k1
) (1.3.5)
+2
n
k=1
X
t
k1
(A
t
k
A
t
k1
) . (1.3.6)
Now, when sup
k
[t
k
t
k1
[ 0, then the rst term at the right hand side
of (1.3.5) converges to [X, X]
t
and the sum of the second term with (1.3.6)
converges to 2
_
t
0
X
s
dX
s
in probability.
Proof of the Ito formula:
Using a stopping argument we can assume that X takes its values in a
bounded interval, say [K, K]. The interest of this argument resides in the
fact that we can approach a C
2
function, as well as its rst two derivatives
uniformly by the polynomials on any compact interval. On the other hand,
using Lemma 1.3.3, we see that the formula is valid for the polynomials. Let
us denote by (f)
t
the random variable
f(X
t
) f(X
0
)
_
t
0
f
t
(X
s
)dX
s
1
2
_
t
0
f
tt
(X
s
)d[X, X]
s
.
Assume moreover that (p
n
, n 1) is a sequence of polynomials such that
(p
n
, 1), (p
t
n
, 1) and (p
tt
n
, 1) converge uniformly on [K, K] to f, f
t
and
to f
tt
respectively. Choosing a subsequence, if necessary, we may assume that
sup
x[K,K]
([f(x) p
n
(x)[ +[f
t
(x) p
t
n
(x)[ +[f
tt
(x) p
tt
n
(x)[) 1/n.
Using the Doob and the Chebytchev inequalities, it is easy to see that (f)
t
(p
n
)
t
converges to zero in probability, since (p
n
)
t
= 0 almost surely, (f)
t
should be also zero almost surely and this completes the proof of the Ito
formula.
As an immediate corollary of the Ito formula we have
Corollary 1.3.4 For any h L
2
([0, 1]), the process dened by
c
t
(I(h)) = exp
__
t
0
h
s
dW
s
1
2
_
t
0
h
2
s
ds
_
is a martingale.
6 Brownian Motion
Proof: Let us denote c
t
(I(h)) by M
t
, then from the Ito formula we have
M
t
= 1 +
_
t
0
M
s
h
s
dW
s
,
hence (M
t
, t [0, 1]) is a local martingale, moreover, since I(h) is Gaus
sian, M
1
is in all the L
p
spaces, hence (M
t
, t [0, 1]) is a square integrable
martingale.
1.4 Alternative constructions of the Wiener
measure
A) Let us state rst the celebrated theorem of ItoNisio about the con
vergence of the random series of independent, Banach space valued random
variables (cf. [42]):
Theorem 1.4.1 (ItoNisio Theorem) Assume that (X
n
, n IN) is a se
quence of independent random variables with values in a separable Banach
space B whose continuous dual is denoted by B
. The sequence (S
n
, n IN)
dened as
S
n
=
n
i=1
X
i
,
converges almost surely in the norm topology of B if and only if there exists
a probability measure on B such that
lim
n
E
_
e
i<,S
n
>
_
=
_
B
e
i<,y>
(dy)
for any B
.
We can give another construction of the Brownian motion using Theorem
1.4.1 as follows: Let (
i
; i IN) be an independent sequence of N
1
(0, 1)
Gaussian random variables. Let (g
i
) be a complete, orthonormal basis of
L
2
([0, 1]). Then W
t
dened by
W
t
(w) =
i=1
i
(w)
_
t
0
g
i
(s)ds
converges almost surely uniformly with respect to t [0, 1] and (W
t
, t [0, 1])
is a Brownian motion. In fact to see this it suces to apply Theorem 1.4.1
to the sequence (X
n
, n IN) dened by
X
n
(w) =
n
(w)
_
0
g
n
(s)ds .
Ito Formula 7
Remark 1.4.2 In the sequel we shall denote by H the socalled Cameron
Martin space H([0, 1], IR
n
) (in case n = 1 we shall again write simply H or,
in case of necessity H([0, 1])) i. e., the isometric image of L
2
([0, 1], IR
n
) under
the mapping
g
_
0
g()d .
Hence for any complete, orthonormal basis (g
i
, i IN) of L
2
([0, 1], IR
n
),
(
_
0
g
i
(s)ds, i IN) is a complete orthonormal basis of H([0, 1], IR
n
). The
use of the generic notation H will be preferred as long as the results are
dimension independent.
B) Let (, T, P) be any abstract probability space and let H be any sep
arable Hilbert space. If L : H L
2
(, T, P) is a linear operator such that
for any h H, E[exp iL(h)] = exp
1
2
[h[
2
H
, then there exists a Banach space
with dense injection
H
W
dense, hence
W
H
is also dense and there exists a probability measure on W such that
_
W
expw
, w)d(w) = exp
1
2
[ j
(w
) [
2
H
and
L(j
(w
))(w) = w
, w)
almost surely. (W, H, ) is called an Abstract Wiener space and is the
Wiener measure (cf. [37]). In the case H is chosen to be
H([0, 1]) =
_
h : h(t) =
_
t
0
h(s)ds, [h[
H
= [
h[
L
2
([0,1])
_
then is the classical Wiener measure and W can be taken as C
0
([0, 1]).
Remark 1.4.3 In the case of the classical Wiener space, any element of
W
is a signed measure on [0, 1], and its image in H = H([0, 1]) can be
represented as j
()(t) =
_
t
0
([s, 1])ds. In fact, we have for any h H
(j
h(s)ds
8 Brownian Motion
=
_
1
0
(([0, 1]) ([0, s])
h(s)ds
=
_
1
0
([s, 1])
h(s)ds.
1.5 CameronMartin and Girsanov Theorems
In the sequel we shall often need approximation of the Wiener functional
with cylindrical smooth functions on the Wiener space. This kind of prop
erties hold in every Wiener space since this is due to the analyticity of the
characteristic function of the Wiener measure. However, they are very easy
to explain in the case of classical Wiener space, that is why we have chosen
to work in this frame. In particular the CameronMartin theorem which is
explained in this section is absolutely indispensable for the development of
the next chapters.
Lemma 1.5.1 The set of random variables
_
f(W
t
1
, . . . , W
t
n
); t
i
[0, 1], f o(IR
n
); n IN
_
is dense in L
2
(), where o(IR
n
) denotes the space of innitely dierentiable,
rapidly decreasing functions on IR
n
.
Proof: It follows from the martingale convergence theorem and the monotone
class theorem.
Lemma 1.5.2 The linear span of the set
=
_
exp
__
1
0
h
s
dW
s
1
2
_
1
0
h
2
s
ds
_
: h L
2
([0, 1])
_
is dense in L
2
().
Proof: It follows from Lemma 1.5.1, via the Fourier transform.
Remark: Although the set separates the points of L
2
(), it does not give
any indication about the positivity.
Lemma 1.5.3 The polynomials are dense in L
2
().
Proof: The proof follows by the analyticity of the characteristic function of
the Wiener measure, in fact, due to this property, the elements of the set in
Lemma 1.5.2 can be approached by the polynomials.
Ito Formula 9
Theorem 1.5.4 (CameronMartin Theorem) For any bounded Borel mea
surable function F on C
0
([0, 1]) and h L
2
([0, 1]), we have
E
_
F
_
w +
_
0
h
s
ds
_
exp
_
_
1
0
h
s
dW
s
1
2
_
1
0
h
2
s
ds
__
= E
[F] .
This assertion implies in particular that the process (t, w) W
t
(w) +
_
t
0
h
s
ds
is again a Brownian motion under the new probability measure
exp
_
_
1
0
h
s
dW
s
1
2
_
1
0
h
2
s
ds
_
d.
Proof: It is sucient to show that the new probability has the same char
acteristic function as : if x
, then x
, w)
W
=
_
1
0
W
s
(w)x
(ds)
= W
t
(w) x
([0, t])
1
0
_
1
0
x
([0, t])dW
t
(w)
= W
1
x
([0, 1])
_
1
0
x
([0, t]).dW
t
=
_
1
0
x
((t, 1])dW
t
.
Consequently
E
__
exp i
_
1
0
x
([t, 1])dW
t
__
w +
_
0
h
s
ds
_
c(I(h))
_
= E
_
exp
_
i
_
1
0
x
([t, 1])dW
t
+ i
_
1
0
x
([t, 1])h
t
dt
_
1
0
h
t
dW
t
1
2
_
1
0
h
2
t
dt
__
= E
_
exp
_
i
_
1
0
(ix
([t, 1]) h
t
)dW
t
_
exp
_
i
_
1
0
x
([t, 1])h
t
dt
1
2
_
1
0
h
2
t
dt
__
= exp
_
1
2
_
1
0
(ix
([t, 1]) h
t
)
2
dt + i
_
1
0
x
([t, 1])h
t
dt
1
2
_
1
0
h
2
t
dt
_
= exp
1
2
_
1
0
(x
([t, 1]))
2
dt
= exp
1
2
[j(x
)[
2
H
,
and this achieves the proof.
The following corollary is one of the most important results of the modern
probability theory:
10 Brownian Motion
Corollary 1.5.5 (Paul Levys Theorem) Suppose that (M
t
, t [0, 1]) is
a continuous martingale with M
0
= 0 and that (M
2
t
t, t [0, 1]) is again a
martingale. Then (M
t
, t [0, 1]) is a Brownian motion.
Proof: From the Ito formula
f(M
t
) = f(0) +
_
t
0
f
t
(M
s
) dM
s
+
1
2
_
t
0
f(M
s
) ds .
Hence the law of (M
t
: t [0, 1]) is .
As an application of Paul Levys theorem we can prove easily the following
result known as the Girsanov theorem which generalizes the CameronMartin
theorem. This theorem is basic in several applications like the ltering of the
random signals corrupted by a Brownian motion, or the problem of optimal
control of Ito processes.
Theorem 1.5.6 (Girsanov Theorem) Assume that u : [0, 1] W IR
n
is a measurable process adapted to the Brownian ltration satisfying
_
1
0
[u
s
[
2
ds <
almost surely. Let
t
= exp
_
_
t
0
(u
s
, dW
s
) 1/2
_
t
0
[u
s
[
2
ds
_
.
Assume that
E [
1
] = 1 . (1.5.7)
Then the process (t, w) W
t
(w) +
_
t
0
u
s
(w)ds is a Brownian motion under
the probability
1
d.
Remark 1.5.7 The condition (1.5.7) is satised in particular if we have
E
_
exp
1
2
_
1
0
[u
s
[
2
ds
_
< .
This is called the Novikov condition (cf. [67, 101]). There is another, slightly
more general sucient condition due to Kazamaki [45], which is
E
_
exp
1
2
_
1
0
u
s
dW
s
_
< .
Note that the dierence between the CameronMartin theorem and the Gir
sanov theorem is that in the former the mapping w w +
_
0
h(s)ds is an
invertible transformation of the Wiener space W and in the latter the corre
sponding map w w +
_
0
u
s
(w)ds is not necessarily invertible.
Ito Formula 11
1.6 The Ito Representation Theorem
The following result is known as the Ito representation formula:
Theorem 1.6.1 Any L
2
() can be represented as
= E[] +
_
1
0
K
s
dW
s
where K L
2
([0, 1] W) and it is adapted.
Proof: Since the Wick exponentials
c(I(h)) = exp
__
1
0
h
s
dW
s
1/2
_
1
0
h
2
s
ds
_
can be represented as claimed and since their nite linear combinations are
dense in L
2
(), the proof follows.
Remark 1.6.2 Let be an integrable real random variable on the Wiener
space. We say that it belongs to the class H
1
if the martingale M = (M
t
, t
[0, 1]) satises the property that
E[< M, M >
1/2
1
] < .
The Ito representation theorem extends via stopping techniques to the ran
dom variables of class H
1
.
1.7 ItoWiener chaos representation
For any h L
2
([0, 1]), dene K
t
=
_
t
0
h
s
dW
s
, t [0, 1]. Then, from the Ito
formula, we can write
K
p
1
= p
_
1
0
K
p1
s
h
s
dW
s
+
p(p 1)
2
_
1
0
K
p2
s
h
2
s
ds
= p
_
1
0
_
(p 1)
_
t
1
0
K
p2
t
2
h
t
2
dW
t
2
+
(p 1)(p 1)
2
_
t
1
0
K
p3
t
2
h
2
t
2
dt
2
_
dW
t
1
+
where p is a positive integer. Iterating this procedure we see that K
p
1
can be
written as the linear combination of the multiple integrals of deterministic
integrands of the type
J
p
=
_
0<t
p
<t
p1
<<t
1
<1
h
t
1
h
t
2
. . . h
t
p
dW
i
1
t
1
. . . dW
i
p
t
p
,
12 Brownian Motion
i
j
= 0 or 1 with dW
0
t
= dt and dW
1
t
= dW
t
. Hence we can express the
polynomials as multiple WienerIto integrals. Let us now combine this ob
servation with the Ito representation:
Assume that L
2
(), then from the Ito representation theorem :
= E[] +
_
1
0
K
s
dW
s
.
Iterating the same procedure for the integrand of the above stochastic inte
gral:
= E[] +
_
1
0
E[K
s
]dW
s
+
_
1
0
_
t
1
0
E[K
1,2
t
1
,t
2
]dW
t
2
dW
t
1
+
_
1
0
_
t
1
0
_
t
2
0
K
1,2,3
t
1
t
2
t
3
dW
t
3
dW
t
2
dW
t
1
.
After n iterations we end up with
=
n
p=0
J
p
(K
p
) +
n+1
and each element of the sum is orthogonal to the other one. Hence (
n
; n
IN) is bounded in the Hilbert space L
2
() and this means that it is weakly
relatively compact. Let (
n
k
) be a weakly convergent subsequence and
=
lim
k
n
k
. Then it is easy from the rst part that
is orthogonal to the
polynomials, therefore
p=0
J
p
(K
p
)
exists and it is equal to almost surely. Let
S
n
=
n
p=0
J
p
(K
p
) ,
then, from the weak convergence, we have
lim
n
E[[S
n
[
2
] = lim
n
E[S
n
] = E[[[
2
] ,
hence (S
n
, n 1) converges weakly to and its L
2
norm converges to the
L
2
norm of and this implies that the series
p=1
J
p
(K
p
)
Ito Formula 13
converges to in the strong topology of L
2
(). Let now
K
p
be an element
of
L
2
[0, 1]
p
(i.e. symmetric), dened as
K
p
= K
p
on C
p
= t
1
< < t
p
.
We dene I
p
(
K
p
) = p!J
p
(K
p
) in such a way that
E[[I
p
(
K
p
)[
2
] = (p!)
2
_
C
p
K
2
p
dt
1
. . . dt
p
= p!
_
[0,1]
p
[
K
p
[
2
dt
1
. . . dt
p
.
Let
p
=
K
p
p!
, then we have proven
Theorem 1.7.1 Any element of L
2
(), can be decomposed as an orthog
onal sum of multiple WienerIto integrals
= E[] +
p=1
I
p
(
p
)
where
p
is a symmetric element of L
2
[0, 1]
p
. Moreover, this decomposition
is unique.
Remark: In the following chapters we shall give an explicit representation
of the kernels
p
using the GrossSobolev derivative.
Notes and suggested reading
The basic references for the stochastic calculus are the books of DellacherieMeyer [21]
and of StroockVaradhan [81]. Especially in the former, the theory is established for
the general semimartingales with jumps. For the construction of the Wiener measure on
Banach spaces we refer the reader to [37] and especially to [49].
14 Brownian Motion
Chapter 2
Sobolev Derivative, Divergence
and OrnsteinUhlenbeck
Operators
2.1 Introduction
Let W = C
0
([0, 1], IR
d
) be the classical Wiener space equipped with the
Wiener measure. We want to construct on W a Sobolev type analysis in
such a way that we can apply it to the random variables that we encounter
in the applications. Mainly we want to construct a dierentiation operator
and to be able to apply it to practical examples. The Frechet derivative is
not satisfactory. In fact the most frequently encountered Wiener functionals,
as the multiple (or single) Wiener integrals or the solutions of stochastic
dierential equations with smooth coecients are not even continuous with
respect to the Frechet norm of the Wiener space. Therefore, what we need
is in fact to dene a derivative on the L
p
()spaces of random variables,
but in general, to be able to do this, we need the following property which is
essential: if F, G L
p
(), and if we want to dene their directional derivative,
in the direction, say w W, we write
d
dt
F(w +t w)[
t=0
and
d
dt
G(w +t w)[
t=0
.
If F = G a.s., it is natural to ask that their derivatives are also equal a.s.
For this, the only way is to choose w in some specic subspace of W, namely,
the CameronMartin space H:
H =
_
h : [0, 1] IR
d
/h(t) =
_
t
0
h(s)ds, [h[
2
H
=
_
1
0
[
h(s)[
2
ds
_
.
In fact, the theorem of CameronMartin says that for any F L
p
(), p > 1,
15
16 Derivative, Divergence
h H
E
_
F(w + h) exp
_
_
1
0
h(s) dW
s
1
2
[h[
2
H
__
= E
[F] ,
or equivalently
E
[F(w + h)] = E
_
F(w) exp
__
1
0
h
s
dW
s
1
2
[h[
2
H
__
.
That is to say, if F = G a.s., then F( + h) = G( + h) a.s. for all h H.
2.2 The Construction of and its properties
If F : W IR is a function of the following type (called cylindrical ):
F(w) = f(W
t
1
(w), . . . , W
t
n
(w)), f o(IR
n
),
we dene, for h H,
h
F(w) =
d
d
F(w + h)[
=0
.
Noting that W
t
(w + h) = W
t
(w) + h(t), we obtain
h
F(w) =
n
i=1
i
f(W
t
1
(w), . . . , W
t
n
(w))h(t
i
),
in particular
h
W
t
(w) = h(t) =
_
t
0
h(s)ds =
_
1
0
1
[0,t]
(s)
h(s)ds.
If we denote by U
t
the element of H dened as U
t
(s) =
_
s
0
1
[0,t]
(r)dr, we have
h
W
t
(w) = (U
t
, h)
H
. Looking at the linear map h
h
F(w) we see that it
denes a random element with values in H
_
1
0
h(s)dW
s
2
2
_
1
0
[
h
s
[
2
ds
__
=0
= E
_
F
n
(w)
_
h
(w) + (w)
_
1
0
h(s)dW
s
__
n
0
since (F
n
, n IN) converges to zero in L
p
().
Proposition 2.2.1 tells us that the operator can be extended to larger
classes of Wiener functionals than the cylindrical ones. In fact we dene rst
the extended L
p
domain of , denoted by Dom
p
() as
Denition 2.2.2 F Dom
p
() if and only if there exists a sequence (F
n
; n
IN) of cylindrical functions such that F
n
F in L
p
() and (F
n
) is Cauchy
in L
p
(, H). Then, for any F Dom
p
(), we dene
F = lim
n
F
n
.
The extended operator is called GrossSobolev derivative .
Remark 2.2.3 Proposition 2.2.1 implies that the denition of F is inde
pendent of the choice of the approximating sequence (F
n
).
Now we are ready to dene
Denition 2.2.4 We will denote by ID
p,1
the linear space Dom
p
() equipped
with the norm F
p,1
= F
p
+F
L
p
(,H)
.
Remark 2.2.5 1. If is a separable Hilbert space we can dene ID
p,1
()
exactly in the same way as before, the only dierence is that we take
o
h
(w) =
n
i=1
i
f(W
t
1
, . . . , W
t
n
)h(t
i
) ,
hence is again a random variable which is B
t
measurable. In fact this
property is satised by a larger class of Wiener functionals:
Proposition 2.3.1 Let ID
p,1
, p > 1 and suppose that is B
t
measurable
for a given t 0. Then is also B
t
measurable and furthermore, for any
h H, whose support is in [t, 1],
h
= (, h)
H
= 0 a.s.
Proof: Let (
n
) be a sequence of cylindrical random variable converging to
in ID
p,1
. If
n
is of the form f(W
t
1
, . . . , W
t
k
), it is easy to see that, even
if
n
is not B
t
measurable, E[
n
[B
t
] is another cylindrical random variable,
say
n
(W
t
1
t
, . . . , W
t
k
t
). In fact, suppose that t
k
> t and t
1
, . . . , t
k1
t.
We have
E[f(W
t
1
, . . . , W
t
k
)[B
t
] = E[f(W
t
1
. . . , W
t
k1
, W
t
k
W
t
+ W
t
)[B
t
]
=
_
IR
f(W
t
1
, . . . , W
t
k1
, W
t
+ x)p
t
k
t
(x)dx
= (W
t
1
, . . . , W
t
k1
, W
t
) ,
and o if f o(IR
k
), where p
t
denotes the heat kernel. Hence we
can choose a sequence (
n
) converging to in ID
p,1
such that
n
is B
t

measurable for each n IN. Hence is also B
t
measurable. If h H has
OrnsteinUhlenbeck Operators 19
its support in [t, 1], then, for each n, we have
h
n
= 0 a.s., because
n
has its support in [0, t] as one can see from the explicit calculation for
n
.
Taking an a.s. convergent subsequence, we see that
h
= 0 a.s. also. .
Let now K be an adapted simple process:
K
t
(w) =
n
i=1
a
i
(w)1
(t
i
,t
i+1
]
(t)
where a
i
ID
p,1
and B
t
i
measurable for any i. Then we have
_
1
0
K
s
dW
s
=
n
i=1
a
i
(W
t
i+1
W
t
i
)
and
h
_
1
0
K
s
dW
s
=
n
i=1
h
a
i
(W
t
i+1
W
t
i
)
+
n
i=1
a
i
(h(t
i+1
) h(t
i
))
=
_
1
0
h
K
s
dW
s
+
_
1
0
K
s
h(s)ds .
Hence
_
1
0
K
s
dW
s
2
H
2
_
_
1
0
K
s
dW
s
2
H
+
_
1
0
[K
s
[
2
ds
_
and
E
_
_
_
1
0
K
s
dW
s
2
H
_
p/2
_
2
p
E
__
_
1
0
K
s
dW
s
p
H
+
_
1
0
[K
s
[
2
ds
_
p/2
_
.
Using the BurkholderDavisGundy inequality for the Hilbert space valued
martingales, the above quantity is majorized by
2c
p
E
_
__
1
0
[K
s
[
2
H
ds
_
p/2
+
__
1
0
[K
s
[
2
ds
_
p/2
_
= c
p

K
p
L
p
(,HH)
+
K
L
p
(,H)
,
where
K. =
_
0
K
r
dr .
Thanks to this majoration, we have proved:
20 Derivative, Divergence
Proposition 2.3.2 Let
K ID
p,1
(H) such that K
t
=
d
K(t)
dt
be B
t
measurable
for almost all t. Then we have
_
1
0
K
s
dW
s
=
_
1
0
K
s
dW
s
+
K (2.3.1)
almost surely.
Remark 2.3.3 The relation 2.3.1 means that, for any h H, we have
h
_
1
0
K
s
dW
s
=
_
1
0
h
K
s
dW
s
+
_
1
0
K
s
h(s)ds .
Corollary 2.3.4 If = I
n
(f
n
), f
n
L
2
([0, 1]
n
), then we have, for h H,
h
I
n
(f
n
) = n
_
[0,1]
n
f(t
1
, . . . , t
n
)
h(t
n
)dW
t
1
, . . . , dW
t
n1
dt
n
.
Proof: Apply the above proposition ntimes to the case in which, rst f
n
is
C
([0, 1]
n
), then pass to the limit in L
2
().
The following result will be extended in the sequel to much larger classes of
random variables:
Corollary 2.3.5 Let : W IR be analytic in Hdirection. Then we have
= E[] +
n=1
I
n
_
E[
n
]
n!
_
,
where
I
n
(g), for a symmetric g H
n
, denotes the multiple Wiener integral
of
n
g
t
1
. . . t
n
(t
1
, . . . , t
n
) .
In other words the kernel
n
L
2
[0, 1]
n
of the Wiener chaos decomposition
of is equal to
n
t
1
. . . t
n
E[
n
]
n!
.
Proof: We have, on one hand, for any h H,
E[(w + h)] = E
_
exp
_
1
0
h
s
dW
s
1
2
_
1
0
h
2
s
ds
_
= E[c(
I(h))] .
OrnsteinUhlenbeck Operators 21
On the other hand, from Taylors formula:
E[(w + h)] = E[] +
1
E
_
(
n
(w), h
n
)
n!
_
= E[] +
1
1
n!
(E[
n
], h
n
)
H
n
= E[] +
1
1
n!
E[
I
n
(E[
n
])
I
n
(h
n
)]
n!
= E[] +
1
E
_
I
n
(E[
n
])
n!
I
n
(h
n
)
n!
_
hence, from the symmetry, we have
I
n
(
n
) =
1
n!
I
n
(E[
n
]) ,
where we have used the notation
I
1
(h) =
I(h) =
_
1
0
h
s
dW
s
and
I
n
(
n
) =
_
[0,1]
n
n
t
1
. . . t
n
(t
1
, . . . , t
n
)dW
t
1
. . . dW
t
n
.
2.4 The divergence operator
The divergence operator, which is the adjoint of the Sobolev derivative with
respect to the Wiener measure, is one of the most important tools of the
Stochastic Analysis. We begin with its formal denition:
Denition 2.4.1 Let : W H be a random variable. We say that
Dom
p
(), if for any ID
q,1
(q
1
+ p
1
= 1), we have
E[(, )
H
] c
p,q
().
q
,
and in this case we dene by
E[ ] = E[(, )
H
] ,
i.e., =
, where
h(s)dW
s
.
To see this, it is sucient to test this relation on the exponential mar
tingales: if k H, we have
E
_
h exp
__
1
0
k
s
dW
s
1
2
_
1
0
k
2
s
ds
__
= E[(h, c(I(k))
H
)]
= E[(h, k)
H
c(I(k))]
= (h, k)
H
.
On the other hand, supposing rst h W
,
E[I(h) c(I(k))] = E[I(h)(w + k)]
= E[I(h)] + (h, k)
H
= (h, k)
H
.
Hence in particular, if we denote by
1
[s,t]
the element of H such that
1
[s,t]
(r) =
_
r
0
1
[s,t]
(u)du, we have that
(
1
[s,t]
) = W
t
W
s
. (2.4.3)
OrnsteinUhlenbeck Operators 23
3.) Let now K be an adapted, simple process
K
t
(w) =
n
1
a
i
(w).1
[t
i
,t
i+1
[
(t) ,
where a
i
ID
p,1
and B
t
i
measurable for each i. Let
K be
_
0
K
s
ds. Then
from the identity (2.4.3), we have
K =
_
n
1
a
i
.
1
[t
i
,t
i+1
[
_
=
n
1
_
a
i
(
1
[t
i
,t
i+1
[
) (a
i
,
1
[t
i
,t
i+1
[
)
_
.
From the relation (2.4.3), we have (
1
[t
i
,t
i+1[
) = W
t
i+1
W
t
i
, further
more, from the Proposition 2.3.1, the support of a
i
is in [0, t
i
], con
sequently, we obtain
K =
n
i=1
a
i
(W
t
i+1
W
t
i
) =
_
1
0
K
s
dW
s
.
Hence we have the important result which says that
Theorem 2.4.2 Dom
p
() (p > 1) contains the set consisting of the primi
tives of adapted stochastic processes satisfying
E
__
_
1
0
K
2
s
ds
_
p/2
_
< .
Moreover one has
__
0
K
s
ds
_
=
_
1
0
K
s
dW
s
.
2.5 Local characters of and
Before proceeding further, we shall prove the locality of the GrossSobolev
derivative and the divergence operators in this section:
Lemma 2.5.1 Let ID
p,1
for some p > 1, then we have, for any constant
c IR,
= 0 on = c ,
almost surely.
Proof: Replacing by c, we may assume that c = 0. Let now f be
a positive, smooth function of compact support on IR such that f(0) = 1.
24 Derivative, Divergence
Let f
() u] = E[(F
(), u)
H
]
= E[f
()(, u)
H
]
E[1
=0
(, u)
H
]
as 0. On the other hand [F
()[ f
L
1
(IR,dt)
, hence it follows that
E[1
=0
(, u)
H
] = 0 ,
since such us are dense in L
q
(, H), the proof follows.
The divergence operator has an analogous property:
Lemma 2.5.2 Assume that u Dom
p
(), p > 1, and that the operator norm
of u, denoted by u
op
is in L
p
(). Then
u = 0 a.s. on w W : u(w) = 0 .
Proof: Let f
_
[u[
2
H
_
u
_
= E
_
f
t
_
[u[
2
H
_ _
u, [u[
2
H
_
H
_
+E
_
f
_
[u[
2
H
_
(u, )
H
_
. (2.5.4)
Note that
f
t
_
[u[
2
H
_ _
u, [u[
2
H
_
H
[u[
2
H
f
t
_
[u[
2
H
_
u
op
sup
IR
[xf
t
(x)[u
op
.
Hence from the dominated convergence theorem, the rst term at the right
of (2.5.4) tends to zero with . Evidently the second one also converges to
zero and this completes the proof.
Remark 2.5.3 Using the local character of the Sobolev derivative one can
dene the local Sobolev spaces as we shall see later.
OrnsteinUhlenbeck Operators 25
2.6 The OrnsteinUhlenbeck Operator
For a nice function f on W, t 0, we dene
P
t
f(x) =
_
W
f
_
e
t
x +
1 e
2t
y
_
(dy) , (2.6.5)
this expression for P
t
is called Mehlers formula. Since (dx)(dy) is invariant
under the rotations of W W, i.e., ( )(dx, dy) is invariant under the
transformation
T
t
(x, y) =
_
xe
t
+ y(1 e
2t
)
1/2
, x(1 e
2t
)
1/2
ye
t
_
,
we have obviously
P
t
f(x)
p
L
p
()
_ _
[(f 1)(T
t
(x, y))[
p
(dx)(dy)
=
_ _
[(f 1)(x, y)[
p
(dx)(dy)
=
_
[f(x)[
p
(dx) ,
for any p 1, P
t
f
L
p f
L
p ; hence also for p = by duality. A
straightforward calculation gives that, for any h H W
(= W
),
P
t
(c(I(h)) = c(e
t
I(h))
=
n=0
e
nt
I
n
(h
n
)
n!
.
Hence, by homogeneity, we have
P
t
(I
n
(h
n
)) = e
nt
I
n
(h
n
)
and by a density argument, we obtain
P
t
I
n
(f
n
) = e
nt
I
n
(f
n
) ,
for any f
n
L
2
([0, 1]
n
). Consequently P
s
P
t
= P
s+t
, i.e., (P
t
) is a measure
preserving Markov semigroup. Its innitesimal generator is denoted by /
and is / is called the OrnsteinUhlenbeck or the number operator. Evidently,
we have
/I
n
(f
n
) = nI
n
(f
n
) (2.6.6)
and this relation means that the Wiener chaos are its eigenspaces. From the
denition, it follows directly that (for a
i
being T
t
i
measurable)
P
t
_
a
i
(W
t
i
+1
W
t
i
)
_
= e
t
(P
t
a
i
)(W
t
i
+1
W
t
i
),
26 Derivative, Divergence
that is to say
P
t
_
1
0
H
s
dW
s
= e
t
_
1
0
P
t
H
s
dW
s
,
and by dierentiation
/
_
1
0
H
s
dW
s
=
_
1
0
(I +/)H
s
dW
s
. (2.6.7)
Also we have
P
t
= e
t
P
t
. (2.6.8)
The following lemma is a consequence of the relation (2.6.6):
Lemma 2.6.1 Assume that L
2
() with the Wiener chaos representation
=
n=0
I
n
(
n
)
satisfying
n=1
n(n!)
n

2
H
n < .
Then
= /,
where is the divergence operator
1
.
Proof: It is sucient to prove for = c(I(h)). In this case from the identity
(2.4.2)
( ) = (h c(I(h)))
=
_
I(h) [h[
2
H
_
c(I(h))
= /c(I(h)) .
Remark 2.6.2 Let us dene for the smooth functions , a seminorm
[[[[[[
p,k
= (I +/)
k/2

L
p
()
.
At rst glance, these seminorms (in fact norms), seem dierent from the one
dened by 
p,k
=
k
0

j

L
p
(,H
j
)
. We will show in the next chapters
that they are equivalent.
1
Sometimes, in the classical case, it is also called HitsudaRamerSkorohod integral.
OrnsteinUhlenbeck Operators 27
2.7 Exercises
These exercises are aimed to give some useful formulas about the iterated divergence
operator and related commutation properties.
1. Prove that
P
t
= e
t
P
t
(2.7.9)
and
P
t
u = e
t
P
t
u (2.7.10)
for any ID
p,1
and u ID
p,1
(H).
2. Assume that u : W H is a cylindrical random variable. Prove that
u =
i=1
(u, e
i
)
H
e
i
e
i
(u, e
i
)
H
,
for any complete, orthonormal basis (e
i
, i IN) of H. In particular, in the nite
dimensional case we can write
u(w) =< u(w), w > trace u(w) ,
although in innite dimensional case such an expression is meaningless in general.
In case the trace u exists, the remaining part is called the Stratonovitch integral.
3. Assume that u : W H is a cylindrical random variable. Prove that
E[(u)
2
] = E[[u[
2
H
] +E[ trace (uu)].
4. Let u be as above, prove the identity
2
u
2
= (u)
2
[u[
2
H
trace (uu) 2(
u
u) ,
where
2
u
2
is dened by the integration by parts formula as
E[
2
u
2
] = E[(
2
, u
2
)
2
] ,
for any test function and (, )
2
denotes the inner product of the space of Hilbert
Schmidt operators on H. Prove that more generally one has
=
2
( ) + trace ()
+(
) + (, )
H
,
where and are two Hvalued, cylindrical random variables.
5. With the same hypothesis as above, show that one has
p+1
u
p+1
= u
p
u
p
u
(
p
u
p
) (
p
u
p
u) .
6. For a u : W H as above, prove that
(u)
p
=
_
u(u)
p1
_
+(u)
p2
_
(p 1)[u[
2
H
+ (p 2) ((
u
u) + trace (uu))
i=0

i

L
p
(,H
i
)
,
and

p,k
= (I +/)
k/2

L
p
()
,
for any p > 1 and k IN. The key point is the continuity property of
the Riesz transform on L
p
([0, 2], dx), i.e., from a totally analytic origin,
although the original proof of P. A. Meyer was probabilistic (cf. [62]). Here
we develop the proof suggested by [28].
3.1 Some Preparations
Let f be a function on [0, 2], extended to the whole IR by periodicity. We
denote by
f(x) the function dened by
f(x) =
1
p.v.
_
0
f(x + t) f(x t)
2 tan t/2
dt , (3.1.1)
where p.v. denotes the the principal value of the integral in (3.1.1). The
famous theorem of M. Riesz, cf. [105], asserts that, for any f L
p
[0, 2],
29
30 Meyer Inequalities
f L
p
([0, 2]), for 1 < p < with

f
p
A
p
f
p
,
where A
p
is a constant depending only on p. Most of the classical functional
analysis of the 20th century has been devoted to extend this result to the
case where the function f was taking its values in more abstract spaces than
the real line. We will show that our problem also can be reduced to this one.
In fact, the main result that we are going to show will be that
(I +/)
1/2

p

p
by rewriting (I + /)
1/2
as an L
p
(, H)valued Riesz transform. For this
we need rst, the following elementary
Lemma 3.1.1 Let K be any function on [0, 2] such that
K()
1
2
cot
2
L
([0, ]) ,
then the operator f T
K
f dened by
T
K
f(x) =
1
p.v.
_
0
(f(x + t) f(x t))K(t)dt
is again a bounded operator on L
p
([0, 2]) with
T
K
f
p
B
p
f
p
for any p (1, )
where B
p
depends only on p.
Proof: In fact we have
T
K
f
f
(x)
1
_
0
[f(x + t) f(x t)[
K(t)
1
2
cot
t
2
dt
c f
L
p
_
_
_K
1
2
cot
2
_
_
_
L
.
Hence
T
K
f
p
_
c
_
_
_K
1
2
cot
2
_
_
_
L
+ A
p
_
f
p
.
Riesz Transform 31
Remark 3.1.2 If for some a ,= 0, aK()
1
2
cot
2
L
1
[a[
_
_
_
_aT
K
f
f
_
_
_
p
+
_
_
_
f
_
_
_
p
_
1
[a[
__
_
_aK
1
2
cot
2
_
_
_
L
f
p
+ A
p
f
p
_
c
p
f
p
with another constant c
p
.
Corollary 3.1.3 Let K be a function on [0, ] such that K = 0 on
_
2
,
_
and K
1
2
cot
2
L
__
0,
2
__
. Then T
K
dened by
T
K
f(x) =
_
/2
0
[f(x + t) f(x t)] K(t)dt
is continuous from L
p
([0, 2]) into itself for any p [1, ) .
Proof: We have
cK()1
[0,
2
]
1
2
cot
2
L
([0, ])
since on the interval
_
2
,
_
, sin
2
_
2
2
, 1
_
, then the result follows from the
Lemma 3.1.1.
3.2 (I + /)
1/2
as the Riesz Transform
Let us denote by R
(x, y) =
_
x cos + y sin , x sin + y cos
_
.
Note that R
= R
+
. We have also, putting e
t
= cos ,
P
t
f(x) =
_
W
f(e
t
x +
1 e
2t
y)(dy)
=
_
W
(f 1)(R
(x, y))(dy)
= P
log cos
f(x) .
32 Meyer Inequalities
Let us now calculate (I +/)
1/2
using this transformation:
(I +/)
1/2
(x) =
_
0
t
1/2
e
t
P
t
(x)dt
=
_
/2
0
(log cos )
1/2
cos
_
W
( 1)(R
(x, y))d
_
.
On the other hand, we have, for h H
h
P
t
(x)
=
d
d
P
t
(x + h)[
=0
=
d
d
_
_
e
t
(x + h) +
1 e
2t
y
_
(dy)[
=0
=
d
d
_
_
e
t
x +
1 e
2t
_
y +
e
t
1 e
2t
h
_
_
(dy)[
=0
=
d
d
_
_
e
t
x +
1 e
2t
y
_
c
_
e
t
1 e
2t
I(h)
_
(y)(dy)[
=0
=
e
t
1 e
2t
_
W
_
e
t
x +
1 e
2t
y
_
h(y) (dy) .
Therefore
h
(I +/)
1/2
(x)
=
_
0
t
1/2
e
t
h
P
t
(x)dt
=
_
0
t
1/2
e
2t
1 e
2t
_
W
h(y)
_
e
t
x +
1 e
2t
y
_
(dy)dt
=
_
/2
0
(log cos )
1/2
cos
2
sin
tan
_
h(y) ( 1) (R
(x, y))(dy)d
Since (dy) is invariant under the transformation y y, we have
_
h(y)( 1)(R
(x, y))(dy) =
_
h(y)( 1)(R
(x, y))(dy),
therefore:
h
(I +/)
1/2
(x)
Riesz Transform 33
=
_
/2
0
(log cos )
1/2
.
_
h(y)
( 1)(R
(x, y))
2
(dy)d
=
_
W
h(y)
_
/2
0
K() (( 1)(R
((0, /2]).
Proof: The only problem is when 0. To see this let us put e
t
= cos ,
then
cot
2
=
1 + e
t
1 e
t
2
t
and
K() =
e
t
t
hence
2K() cot
2
L
__
0,
2
__
.
Using Lemma 3.1.1, Remark 3.1.2 following it and Corollary 3.1.3, we see
that the map f p.v.
_
/2
0
(f(x + ) f(x ))K()d is a bounded map
from L
p
[0, ] into itself. Moreover
Lemma 3.2.2 Let F : W W IR be a measurable, bounded function.
Dene TF(x, y) as
TF(x, y) = p.v.
_
/2
0
[F R
(x, y) F R
(x, y)))K()d ,
34 Meyer Inequalities
this is the Riesz transform for xed (x, y) W W, hence we have
_
/2
0
[TF(R
(x, y))[
p
d c
p
_
0
[F(R
(x, y))[
p
d ,
taking the expectation with respect to , which is invariant under R
,
we have
E
_
0
[TF(R
(x, y))[
p
d = E
_
0
[TF(x, y)[
p
d
=
2
E[[TF[
p
]
c
p
E
_
0
[F(R
(x, y))[
p
d
= c
p
E[[F[
p
] .
We have
Theorem 3.2.3 (I + /)
1/2
: L
p
() L
p
(, H) is a linear continuous
operator for any p (1, ).
Proof: With the notations of Lemma 3.2.2, we have
h
(I +/)
1/2
=
_
W
h(y) T( 1)(x, y)(dy) .
From Holder inequality:
[
h
(I +/)
1/2
(x)[ h
q
__
W
[T( 1)(x, y)[
p
(dy)
_
1/p
c
p
[h[
H
__
W
[T( 1)(x, y)[
p
(dy)
_
1/p
,
where the last inequality follows from the fact that y h(y) is an N
1
(0, [h[
2
H
)
Gaussian random variable. Hence
[(I +/)
1/2
(x)[
H
__
W
[T( 1)(x, y)[
p
(dy)
_
1/p
consequently, from Lemma 3.2.2
(I +/)
1/2

p
p
_
WW
[T( 1)(x, y)[
p
(dx)(dy)
 1
p
L
p
()
= 
p
p
and this completes the proof.
Riesz Transform 35
Corollary 3.2.4 We have
(I +/)
1/2

p
c
p

p
,
for any L
p
(; H) and for any p (1, ).
Proof: It suces to take the adjoint of (I +/)
1/2
.
Corollary 3.2.5 The following identities are valid for any ID:
1. 
p
c
p
(I +/)
1/2

p
2. (I +/)
1/2

p
c
p
(
p
+
p
),
where c
p
and c
p
are two constants independent of .
Proof: The rst identity follows easily as

p
= (I +/)
1/2
(I +/)
1/2

p
c
p
(I +/)
1/2

p
.
To prove the second we have
(I +/)
1/2

p
= (I +/)
1/2
(I +/)
p
= (I +/)
1/2
(I + )
p
(I +/)
1/2

p
+(I +/)
1/2

p

p
+ c
p

p
,
where the last inequality follows from Corollary 3.2.4.
Notes and suggested reading
The inequalities studied in this chapter are due to P. A. Meyer in his seminal paper [62].
He discusses at the last part of it already about the space of test functions dened by the
OrnesteinUhlenbeck operator and proves that this space is an algebra. Then the classical
duality results give birth immediately to the space of the distributions on the Wiener
space, and this is done in [102]. Later the proof of P. A. Meyer has been simplied by
several people. Here we have followed an idea of D. Feyel, cf. [28].
36 Meyer Inequalities
Chapter 4
Hypercontractivity
Introduction
We know that the semigroup of OrnsteinUhlenbeck is a bounded opera
tor on L
p
(), for any p [1, ]. In fact for p (1, ), it is more than
bounded. It increases the degree of integrability, this property is called hy
percontractivity and it is used to show the continuity of linear operators
on L
p
()spaces dened via the Wiener chaos decomposition or the spectral
decomposition of the OrnsteinUhlenbeck operator. We shall use it in the
next chapter to complete the proof of the Meyer inequalities. Hypercontrac
tivity has been rst discovered by E. Nelson, here we follow the proof given
by [66]. We complete the chapter by an analytic proof of the logarithmic
Sobolev inequality of Leonard Gross (cf. [36], [22]) for which we shall give
another proof in the fth chapter.
4.1 Hypercontractivity
In the sequel we shall show that this result can be proved using the Ito
formula. Let (, /, P) be a probability space with (B
t
; t IR
+
) being a
ltration. We take two Brownian motions (X
t
; t 0) and (Y
t
; t 0) which
are not necessarily independent, i.e., X and Y are two continuous, real mar
tingales such that (X
2
t
t) and (Y
2
t
t) are again martingales (with respect
to (B
t
)) and that X
t
X
s
and Y
t
Y
s
are independent of B
s
, for t > s.
Moreover there exists (
t
; t IR
+
), progressively measurable with values in
[1, 1] such that
(X
t
Y
t
_
t
0
s
ds, t 0)
37
38 Hypercontractivity
is again a (B
t
)martingale. Let us denote by
t
= (X
s
; s t),
t
= (Y
s
; s t)
i.e., the corresponding ltrations of X and Y and by and by their
respective supremum.
Lemma 4.1.1 1. For any L
1
(, , P), t 0, we have
E[[B
t
] = E[[
t
] a.s.
2. For any L
1
(, , P), t 0, we have
E[[B
t
] = E[[
t
] a.s.
Proof: Since the two claims are similar, we shall prove only the rst one.
From Paul Levys theorem, we have also that (X
t
) is an (
t
)Brownian mo
tion. Hence
= E[] +
_
0
H
s
dX
s
where H is (
t
)adapted process. Hence
E[[B
t
] = E[] +
_
t
0
H
s
dX
s
= E[[
t
] .
Let T be the operator T : L
1
(, , P) L
1
(, , P) dened as the restric
tion of E[ [] to the space L
1
(, , P). We know that T : L
p
() L
p
()
is a contraction for any p 1. If we impose supplementary conditions to ,
then we have more:
Proposition 4.1.2 If [
t
(w)[ r (dt dP a.s.) for some r [0, 1], then
T : L
p
() L
q
() is a bounded operator, where
p 1 r
2
(q 1) .
Proof: p = 1 is already known. So suppose p, q ]1, [ . Since L
() is
dense in L
p
(), it is enough to prove that TF
q
F
p
for any F L
().
Moreover, since T is a positive operator, we have [T(F)[ T([F[), hence we
can work as well with F L
+
(). Due to the duality between L
p
spaces, it
suces to show that
E[T(F)G] F
p
G
q
,
_
1
q
t
+
1
q
= 1
_
,
39
for any F L
+
(), G L
+
(). Since bounded and positive random
variables are dense in all L
p
+
for any p > 1, we can suppose without loss of
generality that F, G [a, b] almost surely for some 0 < a < b < . Let
M
t
= E[F
p
[
t
]
N
t
= E[G
q
[
t
] .
Then, from the Ito representation theorem we have
M
t
= M
0
+
_
t
0
s
dX
s
N
t
= N
0
+
_
t
0
s
dY
s
where is adapted, is adapted, M
0
= E[F
p
], N
0
= E[G
q
]. From the
Ito formula, we have
M
t
N
t
= M
0
N
0
+
_
t
0
M
1
s
N
s
dM
s
+
_
t
0
M
s
N
1
s
dN
s
+
+
1
2
_
t
0
M
s
N
s
A
s
ds
where
A
t
= ( 1)
_
t
M
t
_
2
+ 2
t
M
t
t
N
t
t
+ ( 1)
_
t
N
t
_
2
and =
1
p
, =
1
q
t
= M
0
+
_
t
0
M
1
s
s
dX
s
+
( 1)
2
_
t
0
M
2
s
2
s
ds
N
t
=
and then as
M
t
N
t
M
0
N
0
=
_
t
0
M
s
dN
s
+
_
t
0
N
s
dM
s
+
_
t
0
M
1
s
N
1
s
s
s
ds
=
_
t
0
M
s
_
N
1
s
s
dY
s
+
( 1)
2
N
2
s
2
s
ds
_
+
_
t
0
N
s
_
M
1
s
s
dX
s
+
( 1)
2
M
2
s
2
s
ds
_
+
_
t
0
M
1
s
N
1
s
s
s
ds
40 Hypercontractivity
and nally to pick up together all the integrands integrated with respect to
the Lebesgue measure ds.
As everything is square integrable, it comes
E[M
] = E
_
E[F
p
[
E[G
q
_
= E[F G]
=
1
2
_
0
E[N
t
M
t
A
t
]dt + EM
0
N
0
= E[F
p
]
E[G
q
+
1
2
_
0
E[M
t
N
t
A
t
]dt .
Consequently
E[FG] X
p
Y 
q
=
1
2
_
0
E
_
M
t
N
t
A
t
_
dt .
Look now at A
t
as a polynomial of second degree with respect to
M
. Then
4
=
2
2
t
( 1)( 1) .
If 0 then the sign of A
t
is the same as the sign of ( 1) 0, i.e., if
2
t
( 1)( 1)
=
_
1
1
__
1
1
_
= (p 1)(q
t
1)
a.s., then we obtain
E[FG] = E[T(F)G] F
p
G
q
1
2
z,
] =
_
W
F
_
x +
_
1
2
z
_
(dz).
Proof: For any G L
), we have
E[F(w) G(x)] = E
_
F(w)G
_
w +
_
1
2
z
__
= E
_
F
_
w +
_
1
2
z
_
G(w)
_
=
_ _
F
_
w +
_
1
2
z
_
G( w) (d w)(d z)
= E
_
G(x)
_
F
_
x +
_
1
2
z
_
(d z)
_
41
where w, z represent the dummy variables of integration.
Corollary 4.1.4 Under the hypothesis of the above lemma, we have
_
_
_
_
_
W
F
_
x +
_
1
2
z
_
(d z)
_
_
_
_
L
q
()
F
L
p
()
for any
(p 1)
2
(q 1) .
4.2 Logarithmic Sobolev Inequality
Let (P
t
, t 0) be the OrnsteinUhlenbeck semigroup. The commutation
relation (cf. 2.7.9)
P
t
f = e
t
P
t
f
is directly related to the logarithmic Sobolev inequality of L. Gross:
E
_
f
2
log f
2
_
E[f
2
] log E[f
2
] 2E
_
[f[
2
H
_
.
In fact, suppose that f is strictly positive and lower and upper bounded. We
have
E[f log f] E[f] log E[f] =
_
0
E
_
d
dt
P
t
f log P
t
f
_
dt
=
_
0
E [/P
t
f log P
t
f] dt
=
_
0
E
_
[P
t
f[
2
H
P
t
f
_
dt
=
_
0
e
2t
E
_
[P
t
f[
2
H
P
t
f
_
dt . (4.2.1)
Now insert in 4.2.1 the following,
[P
t
(f)[
2
H
=
P
t
_
f
1/2
f
f
1/2
_
2
H
(P
t
f) P
t
_
[f[
2
H
f
_
42 Hypercontractivity
which is a consequence of the Holder inequality, to obtain
E[f log f] E[f] log E[f]
_
0
e
2t
E
_
P
t
_
[f[
2
H
f
__
dt
=
_
0
e
2t
4E[[
_
f[
2
H
]dt
= 2E[[
_
f[
2
H
] ,
replacing f by f
2
completes the proof of the inequality.
Remark 4.2.1 Here we have used the fact that if f > 0 almost surely, then
P
t
f > 0 also. In fact one can prove, using the Cameron Martin theorem,
that, if g > 0 > 0, then P
t
g > 0 almost surely. P
t
is called a positivity
improving semigroup (cf. Corollary 6.1.7).
Notes and suggested reading
The hypercontractivity property of the OrnsteinUhlenbeck semigroup is due to E. Nelson.
The proof given here follows the lines given by J. Neveu, cf. [66]. For the logarithmic
Sobolev inequality and its relations to hypercontractivity cf. [22].
AND DISTRIBUTIONS
Chapter 5
L
p
Multipliers Theorem, Meyer
Inequalities and Distributions
5.1 L
p
Multipliers Theorem
L
p
Multipliers Theorem gives us a tool to perform some sort of symbolic
calculus to study the continuity of the operators dened via the Wiener chaos
decomposition of the Wiener functionals. With the help of this calculus we
will complete the proof of the Meyers inequalities.
Almost all of these results have been discovered by P. A. Meyer (cf. [62])
and they are consequences of the Nelsons hypercontractivity theorem ([65]).
First let us give rst the following simple and important result:
Theorem 5.1.1 Let F L
p
(), p > 1, denote by I
n
(F
n
) the projection of F
on the nth Wiener chaos, n 1. Then the map F I
n
(F
n
) is continuous
on L
p
().
Proof: Suppose rst p > 2. Let t be such that p = e
2t
+ 1, then we have
P
t
F
p
F
2
.
Moreover
P
t
I
n
(F
n
)
p
I
n
(F
n
)
2
F
2
F
p
but P
t
I
n
(F
n
) = e
nt
I
n
(F
n
), hence
I
n
(F
n
)
p
e
nt
F
p
.
For 1 < p < 2 we use the duality: let F I
n
(F
n
) = J
n
(F). Then
I
n
(F)
p
= sup
G
q
1
[G, J
n
(F))[
43
44 Multipliers and Inequalities
= sup [J
n
(G), F)[
= sup [J
n
G, J
n
F)[
sup e
nt
G
q
F
p
= e
nt
F
p
.
Proposition 5.1.2 (Meyers Multipliers theorem) Let the function h
be dened as
h(x) =
k=0
a
k
x
k
be analytic around the origin with
k=1
[a
k
[
_
1
n
_
k
<
for n n
0
, for some n
0
IN. Let (x) = h(x
) and dene T
on L
p
() as
T
F =
n=0
(n)I
n
(F
n
) .
Then the operator to T
is bounded on L
p
() for any p > 1.
Proof: Suppose rst = 1. Let T
= T
1
+ T
2
where
T
1
F =
n
0
1
n=0
(n)I
n
(F
n
), T
2
F = (I T
1
)F .
From the hypercontractivity, F T
1
F is continuous on L
p
(). Let
n
0
F =
n=n
0
I
n
(F
n
).
Since
(I
n
0
)(F) =
n
0
1
n=0
I
n
(F
n
),
n
0
: L
p
L
p
is continuous, hence P
t
n
0
: L
p
L
p
is also continuous.
Applying RieszThorin interpolation theorem, which says that if P
t
n
0
is
L
q
L
q
and L
2
L
2
then it is L
p
L
p
for any p such that
1
p
is in the
interval
_
1
q
,
1
2
_
, we obtain
P
t
n
0

p,p
P
t
n
0

2,2
P
t
n
0

1
q,q
P
t
n
0

2,2

n
0

1
q,q
Distributions 45
where
1
p
=
2
+
1
q
, (0, 1) . Choose q large enough such that 1 (if
necessary). Hence we have
P
t
n
0

p.p
e
n
0
t
K, K = K(n
0
, ) .
A similar argument holds for p (1, 2) by duality.
We then have
T
2
(F) =
nn
0
(n)I
n
(F
n
)
=
nn
0
_
k
a
k
_
1
n
_
k
_
I
n
(F
n
)
=
k
a
k
nn
0
_
1
n
_
k
I
n
(F
n
)
=
k
a
k
nn
0
/
k
I
n
(F
n
)
=
k
a
k
/
k
n
0
F .
We also have
/
1
n
0
F
p
=
_
_
_
_
0
P
t
n
0
Fdt
_
_
_
p
K
_
0
e
n
0
t
F
p
dt K
F
p
n
0
/
2
n
0
F
p
=
_
_
_
_
0
_
0
P
t+s
n
0
Fdt ds
_
_
_
p
K
F
p
(n
0
)
2
,
/
k
n
0
F
p
KF
p
1
(n
0
)
k
.
Therefore
T
2
(F)
p
k
KF
p
1
n
k
0
KF
p
1
n
k
0
by the hypothesis (take n
0
+ 1 instead of n
0
if necessary).
For the case (0, 1) , let
()
t
(ds) be the measure on IR
+
, dened by
_
IR
+
e
s
()
t
(ds) = e
t
.
Dene
Q
t
F =
n
e
n
t
I
n
(F
n
) =
_
0
P
s
F
()
t
(ds) .
Then
Q
t
n
0
F
p
F
p
_
0
e
n
0
s
()
t
(ds)
= F
p
e
t(n
0
)
.
,
46 Multipliers and Inequalities
the rest of the proof goes as in the case = 1.
Examples:
1) Let
(n) =
_
1 +
1 + n
_
s
s (, )
= h
_
_
1
n
_
_
, h(x) =
_
1 + x
1 + x
2
_
s
.
Then T
: L
p
L
p
is bounded. Moreover
1
(n) =
1
(n)
= h
1
__
1
n
_
,
h
1
(x) =
1
h(x)
is also analytic near the origin, hence T
1 : L
p
L
p
is also a
bounded operator.
2) Let (n) =
1+n
2+n
then h(x) =
_
x+1
2x+1
satises also the above hypothesis.
3) As an application of (2), look at
(I +/)
1/2

p
= (2I +/)
1/2

p
(I +/)
1/2
(2I +/)
1/2

p
= (2I +/)
1/2
(I +/)
1/2

p
= T
(I +/)
1/2
(I +/)
1/2

p
c
p
(I +/)
p
.
Continuing this way we can show that

k

L
p
(,H
k
)
c
p,k

p,k
(= (I +/)
k/2

p
)
c
p,k
(
p
+
k

L
p
(,H
k
)
)
and this completes the proof of the Meyer inequalities for the scalarvalued
Wiener functionals. If is a separable Hilbert space, we denote with ID
p,k
()
the completion of the valued polynomials with respect to the norm

ID
p,k
()
= (I +/)
k/2

L
p
(,)
.
We dene as in the case = IR, the Sobolev derivative , the divergence
, etc. All we have said for the real case extend trivially to the vector case,
including the Meyer inequalities. In fact, in the proof of these inequalities
the main step is the Riesz inequality for the Hilbert transform. However this
Distributions 47
inequality is also true for any Hilbert space (in fact it holds also for a class of
Banach spaces which contains Hilbert spaces, called UMD spaces). The rest
is almost the transcription of the real case combined with the Khintchine
inequalities. We leave hence this passage to the reader.
Corollary 5.1.3 For every p > 1, k IR, has a continuous extension as
a map ID
p,k
ID
p,k1
(H).
Proof: We have

p,k
= (I +/)
k/2

p
= (2I +/)
k/2

p
c
p
(1 +/)
1/2
(2I +/)
k/2

p
(I +/)
(k+1)/2

p
= 
p,k+1
.
Corollary 5.1.4 =
: ID
p,k
(H) ID
p,k1
is continuous for all p > 1
and k IR.
Proof: The proof follows from the duality.
In particular we have :
Corollary 5.1.5 The Sobolev derivative and its adjoint extend to distribu
tion spaces as explained below:
Sobolev derivative operates as a continuous operator on
: ID =
p,k
ID
p,k
D(H) =
p,k
ID
p,k
(H)
and it extends continuously as a map
: ID
t
=
_
p,k
ID
p,k
ID
t
(H) =
_
p,k
ID
p,k
(H).
The elements of the space ID
t
are called MeyerWatanabe distributions.
48 Multipliers and Inequalities
Consequently its adjoint has similar properties:
:
p,k
ID
p,k
(H) = ID(H) ID
is continuous and this map has a continuous extension to
: ID
t
(H) ID
t
Proof: Everything follows from the dualities
(ID)
t
= ID
t
, (ID(H))
t
= ID
t
(H).
Denition 5.1.6 For n 1, we dene
n
as (
n
)
with respect to .
Here is the generalization of Corollary 2.3.5 promised in Chapter 2:
Proposition 5.1.7 For L
2
(), we have
= E[] +
n1
1
n!
n
(E[
n
]) . (5.1.1)
Proof: If f is a symmetric element of H
n
, we e shall denote by
I
n
(f) the
nth order multiple ItoWiener integral of the density of f with respect to the
Lebesgue measure on [0, 1]
n
(cf. also Corollary 2.3.5). With this notational
convention, suppose that h (w + h) is analytic for almost all w. Then
we have
(w + h) = (w) +
n1
(
n
(w), h
n
)
H
n
n!
.
Take the expectations:
E[(w + h)] = E[ c(h)]
= E[] +
n
(E[
n
], h
n
)
n!
= E[] +
n1
E
_
I
n
(E[
n
])
n!
c(h)
_
.
Since the nite linear combinations of the elements of the set c(h); h H
is dense in any L
p
(), we obtain the identity
(w) = E[] +
n1
I
n
(E[
n
])
n!
.
Distributions 49
Let ID, then we have (with E[] = 0),
, ) =
n1
E[
I
n
(
n
)
I
n
(
n
)]
=
n
E
_
I
n
(E[
n
])
n!
I
n
(
n
)
_
=
n
(E[
n
],
n
)
=
n
1
n!
(E[
n
], E[
n
])
=
n
1
n!
E[(E[
n
],
n
)]
=
n
1
n!
E[
n
(E[
n
]) ]
hence we obtain that
=
n
1
n!
n
E[
n
].
In particular it holds true that
n
E[
n
] =
I
n
(E[
n
]) .
Evidently this identity is valid not only for the expectation of the nth deriva
tive of a Wiener functional but for any symmetric element of H
n
.
Remark 5.1.8 Although in the litterature Proposition 5.1.7 is announced
for the elements of ID, the proof given here shows its validity for the elements
of L
2
(). In fact, although
n
is a distribution, its expectation is an ordi
nary symmetric tensor of order n, hence the corresponding multiple Wiener
integrals are welldened. With a small extra work, we can show that in fact
the formula (5.1.1) holds for any
kZ
ID
2,k
.
Let us give another result important for the applications:
Proposition 5.1.9 Let F be in some L
p
() with p > 1 and suppose that the
distributional derivative F of F, is in some L
r
(, H), (1 < r). Then F
belongs to ID
rp,1
.
Proof: Without loss of generality, we can assume that r p. Let (e
i
; i IN)
be a complete, orthonormal basis of the CameronMartin space H. Denote
by V
n
the sigmaeld generated by e
1
, . . . , e
n
, and by
n
the orthogonal
50 Multipliers and Inequalities
projection of H onto the subspace spanned by e
1
, . . . , e
n
, n IN. Let us
dene F
n
by
F
n
= E[P
1/n
F[V
n
],
where P
1/n
is the OrnsteinUhlenbeck semigroup at t = 1/n. Then F
n
belongs to ID
r,k
for any k IN and converges to F in L
r
(). Moreover, from
Doobs lemma, F
n
is of the form
F
n
(w) = (e
1
, . . . , e
n
),
with being a Borel function on IR
n
, which is in the intersection of the
Sobolev spaces
k
W
r,k
(IR
n
,
n
) dened with the OrnsteinUhlenbeck operator
L
n
= +x on IR
n
. Since L
n
is elliptic, the Weyl lemma implies that
can be chosen as a C
function. Consequently, F
n
is again V
n
measurable
and we nd , using the very denition of conditional expectation and the
Mehler formula, that
F
n
= E[e
1/n
n
P
1/n
F[V
n
].
Consequently, from the martingale convergence theorem and from the fact
that
n
I
H
in the weak operators topology, it follows that
F
n
F,
in L
r
(, H), consequently F belongs to ID
r,1
.
Appendix: Passing from the classical Wiener
space to the Abstract Wiener Space (or vice
versa):
Let (W, H, ) be an abstract Wiener space. Since, `a priori, there is no notion
of time, it seems that we can not dene the notion of anticipation, non
anticipation, etc. This diculty can be overcome in the following
way:
Let (p
= p
(H), where p
(H) in H.
Denition 5.1.10 We will denote by T
almost surely.
Distributions 51
Lemma 5.1.11 We have
= B(W)
up to negligeable sets.
Proof: We have already
_
T
such that h
n
h
in H. Hence h
n
h in L
p
(), for all p 1. Since each h
n
is
_
T

measurable, so does h. Since B(W) is generated by h; h H the proof
is completed.
Denition 5.1.12 A random variable : W H is called a simple, adapted
vector eld if it can be written as a nite sum:
=
i<
F
i
(p
i+1
h
i
p
i
h
i
)
where h
i
H, F
i
are T
i
measurable (and smooth for the time being) random
variables.
Proposition 5.1.13 For each adapted simple vector eld we have
i) =
i<
F
i
(p
i+1
h
i
p
i
h
i
)
ii) with Itos isometry:
E
_
[[
2
_
= E
_
[[
2
H
_
.
Proof: The rst part follows from the usual identity
[F
i
(p
i+1
p
i
)h
i
] = F
i
[(p
i+1
p
i
)h
i
]
_
F
i
, (p
i+1
p
i
)h
i
_
H
and from the fact that the second term is null since F
i
H
almost surely.
The verication of the second relation is left to the reader.
Remark 5.1.14 If we denote F
i
1
]
i
,
i+1
]
() h
i
by
(), we have the follow
ing relations:
=
_
()dp
with 
2
2
= E
_
d(
, p
) = 
2
L
2
(,H)
,
which are signicantly analogous to the relations that we have seen before.
52 Multipliers and Inequalities
The Ito representation theorem can be stated in this setting as follows:
suppose that (p
h, k)
H
is continuous for any h, k H. Then
Theorem 5.1.15 Let us denote with ID
a
2,0
(H) the completion of adapted sim
ple vector elds with respect to the L
2
(, H)norm. Then we have
L
2
() = IR + : ID
a
2,0
(H) ,
i.e., any L
2
() can be written as
= E[] +
for some ID
a
2,0
(H). Moreover such is unique up to L
2
(, H)equivalence
classes.
The following result explains the reason of the existence of the Brownian
motion (cf. also [90]):
Theorem 5.1.16 Suppose that there exists some
0
H such that the set
p
0
; has a dense span in H (i.e. the linear combinations from it
is a dense set). Then the realvalued (T
)martingale dened by
b
= p
0
is a Brownian motion with a deterministic time change and (T
; ) is
its canonical ltration completed with the negligeable sets.
Example: Let H = H
1
([0, 1]), dene A as the operator dened by Ah(t) =
_
t
0
s
h)(t) =
_
t
0
1
[0,]
(s)
h(s)ds
and
0
(t) =
_
t
0
1
[0,1]
(s)ds satises the hypothesis of the above theorem.
0
is
called the vacuum vector (in physics).
This is the main example, since all the (separable) Hilbert spaces are
isomorphic, we can carry this time structure to any abstract HilbertWiener
space as long as we do not need any particular structure of time.
Distributions 53
5.2 Exercises
1. Give a detailed proof of Corollary 5.1.4, in particular explain the why of the exis
tence of continuous extensions of and .
2. Prove the last claim of Remark 5.1.8.
Notes and suggested reading
To complete the series of the Meyer inequalities, we have been obliged to use the hypercon
tractivity property of the OrnsteinUhlenbeck semigroup as done in [62]. Once this is done
the extensions of and to the distributions are immediate via the duality techniques.
Proposition 5.1.7 is due to Stroock, [80] with a dierent proof. The results of the appendix
are essentially due to the author, cf. [90]. In [101] a stochastic calculus is constructed in
more detail.
54 Multipliers and Inequalities
Chapter 6
Some Applications
Introduction
In this chapter we give some applications of the extended versions of the
derivative and the divergence operators. First we give an extension of the
ItoClark formula to the space of the scalar distributions. We refer the reader
to [11] and [70] for the developments of this formula in the case of Sobolev
dierentiable Wiener functionals. Let us briey explain the problem: al
though, we know from the Ito representation theorem, that each square inte
grable Wiener functional can be represented as the stochastic integral of an
adapted process, without the use of the distributions, we can not calculate
this process, since any square integrable random variable is not necessarily
in ID
2,1
, hence it is not Sobolev dierentiable in the ordinary sense. As it
will be explained, this problem is completely solved using the dierentiation
in the sense of distributions. Afterwards we give a straightforward appli
cation of this result to prove a 0 1 law for the Wiener measure. At the
second section we construct the composition of the tempered distributions
with nondegenerate Wiener functionals as MeyerWatanabe distributions.
This construction carries also the information that the probability density of
a nondegenerate random variable is not only innitely dierentiable but also
it is rapidly decreasing. The same idea is then applied to prove the regularity
of the solutions of the Zakai equation for the ltering of nonlinear diusions.
55
56 Applications
6.1 Extension of the ItoClark formula
Let F be any integrable random variable. Then we the celebrated Ito Rep
resentation Theorem 1.6.1 tells us that F can be represented as
F = E[F] +
_
1
0
H
s
dW
s
,
where (H
s
; s [0, 1]) is an adapted process such that, it is unique and
_
1
0
H
2
s
ds < + a.s.
Moreover, if F L
p
(p > 1), then we also have
E
__
_
1
0
[H
s
[
2
ds
_
p/2
_
< .
One question is how to calculate the process H. In fact, below we will
extend the Ito representation and answer to the above question for any F
ID
t
(i.e., the MeyerWatanabe distributions). We begin with
Lemma 6.1.1 Let ID(H) be represented as (t) =
_
t
0
s
ds, then de
ned by
(t) =
_
t
0
E[
s
[T
s
]ds
belongs again to ID(H). In other words : ID(H) ID(H) is a linear
continuous operator.
Proof: Let (P
t
, t IR
+
) be the OrnsteinUhlenbeck semigroup. Then it is
easy to see that, for any [0, 1], if L
1
() is B
measurable, then so
is also P
t
for any t IR
+
. This implies in particular that / = /.
Therefore

p,k
= E
__
_
1
0
[(I +/)
k/2
E[
s
[T
s
][
2
ds
_
p/2
_
=
= E
__
_
1
0
[E[(I +/)
k/2
s
[T
s
][
2
ds
_
p/2
_
c
p
E
__
_
1
0
[(I +/
k/2
s
[
2
ds
_
p/2
_
(c
p
= p)
where the last inequality follows from the convexity inequalities of the dual
predictable projections (c.f. [21]).
ItoClark Formula 57
Lemma 6.1.2 : ID(H) ID(H) extends as a continuous mapping to
ID
t
(H) ID
t
(H).
Proof: Let ID(H), then we have, for k > 0,

p,k
= (I +/)
k/2

p
= (I +/)
k/2

p
c
p
(I +/)
k/2

p
c
p

p,k
,
then the proof follows since ID(H) is dense in ID
t
(H).
Before going further let us give a notation: if F is in some ID
p,1
then
its GrossSobolev derivative F is an Hvalued random variable. Hence
t F(t) is absolutely continuous with respect to the Lebesgue measure on
[0, 1]. We shall denote by D
s
F its RadonNikodym derivative with respect
to the Lebesgue measure. Note that D
s
F is ds dalmost everywhere well
dened.
Lemma 6.1.3 Let ID, then we have
= E[] +
_
1
0
E[D
s
[T
s
]dW
s
= E[] + .
Moreover ID(H).
Proof: Let u be an element of L
2
(, H) such that u(t) =
_
t
0
u
s
ds with
( u
t
; t [0, 1]) being an adapted and bounded process. Then we have, from
the Girsanov theorem,
E
_
(w + u(w)). exp
_
_
1
0
u
s
dW
s
2
2
_
1
0
u
s
ds
__
= E[].
Dierentiating both sides at = 0, we obtain:
E[((w), u)
_
1
0
u
s
dW
s
] = 0 ,
i.e.,
E[(, u)] = E[
_
1
0
u
s
dW
s
].
Furthermore
E
_
_
1
0
D
s
u
s
ds
_
= E
_
_
1
0
E[D
s
[T
s
] u
s
ds
_
= E[(, u)
H
]
= E
__
_
1
0
E[D
s
[T
s
]dW
s
__
_
1
0
u
s
dW
s
__
.
58 Applications
Since the set of the stochastic integrals
_
1
0
u
s
dW
s
of the processes u as above
is dense in L
2
0
() = F L
2
() : E[F] = 0, we see that
E[] =
_
1
0
E[D
s
[T
s
]dW
s
= .
The rest is obvious from the Lemma 6.1.1 .
Lemma 6.1.3 extends to ID
t
as:
Theorem 6.1.4 For any T ID
t
, we have
T = T, 1) + T .
Proof: Let (
n
) ID such that
n
T in D
t
. Then we have
T = lim
n
n
= lim
n
E[
n
] +
n
= lim
n
E[
n
] + lim
n
n
= lim
n
1,
n
) + lim
n
n
= 1, T) + T
since : ID
t
ID
t
(H), : ID
t
(H) ID
t
(H) and : ID
t
(H) ID
t
are all
linear, continuous mappings.
Here is a nontrivial application of the Theorem 6.1.4:
Theorem 6.1.5 (01 law)
Let A B(W) such that
h
1
A
= 0, h H, where the derivative is in the
sense of the distributions. Then (A) = 0 or 1.
Remark: In particular, the above hypothesis is satised when A + H A.
Proof: Let T
A
= 1
A
, then Theorem 6.1.4 implies that
T
A
= T
A
, 1) = (A) ,
hence (A)
2
= (A). Another proof can be given as follows: let T
t
be dened
as P
t
1
A
, where (P
t
, t 0) is the OrnsteinUhlenbeck semigroup. Then, from
the hypothesis, T
t
= e
t
P
t
1
A
= 0, consequently T
t
is almost surely a
constant for any t > 0, this implies that lim
t0
T
t
= 1
A
is also a constant.
Lifting of the Functionals 59
Remark 6.1.6 From DoobBurkholder inequalities, it follows via a duality
technique, that
E
_
__
1
0
[(I +/)
k/2
E[D
s
[T
s
][
2
ds
_
p/2
_
c
p
(I +/)
k/2

p
L
p
()
c
p,k

p
p,k
,
for any p > 1 and k IR. Consequently, for any ID
p,k
, ID
p,k
(H)
and the Ito integral is an isomorphisme from the adapted elements of ID
p,k
(H)
onto ID
o
p,k
= ID
p,k
: , 1) = 0 (cf. [85] for further details).
Corollary 6.1.7 (Positivity improving) Let F L
p
(), p > 1 be a non
negative Wiener functional such that F > 0 > 0, denote by (P
, IR
+
)
the OrnsteinUhlenbeck semigroup. Then, for any t > 0, the set A
t
= w :
P
t
F(w) > 0 has full measure, in fact we have
A
t
+ H A
t
.
Proof: From the Mehler and CameronMartin formulae, we have
P
t
F(w + h) =
_
W
F(e
t
(w + h) +
1 e
2t
y)(dy)
=
_
W
F(e
t
w +
1 e
2t
y)(
t
h(y))(dy)
where
t
=
e
t
1 e
2t
and
(h) = exp
_
h 1/2[h[
2
H
_
.
This proves the claim about the Hinvariance of A
t
and the proof follows
from Theorem 6.1.5.
6.2 Lifting of o
/
(IR
d
) with random variables
Let f : IR IR be a C
1
b
function, F ID. Then we know that
(f(F)) = f
t
(F)F .
60 Lifting of o
t
(IR
d
)
Now suppose that [F[
2
H
L
p
(), then
f
t
(F) =
((f(F)), F)
H
[F[
2
H
Even if f is not C
1
, the right hand side of this equality has a sense if we look
at (f(F)) as an element of ID
t
. In the following we will develop this idea:
Denition 6.2.1 Let F : W IR
d
be a random variable such that F
i
ID,
for all i = 1, . . . , d, and that
[det(F
i
, F
j
)]
1
p>1
L
p
().
Then we say that F is a nondegenerate random variable.
Lemma 6.2.2 Let us denote by
ij
= (F
i
, F
j
)
H
and by =
1
(as a
matrix). Then ID(IR
d
IR
d
), in particular det ID.
Proof: Formally, we have, using the relation = Id,
ij
=
k,l
ik
jl
kl
.
To justify this we dene rst
ij
=
ij
+
ij
, > 0. Then we can write
ij
= f
ij
(
), where f : IR
d
IR
d
IR
d
IR
d
is a smooth function of
polynomial growth. Hence
ij
ID. Then from the dominated convergence
theorem we have
ij
ij
in L
p
and
k
ij
ij
in L
p
(, H
k
) (the latter
follows again from
= Id).
Lemma 6.2.3 Let G ID. Then, for all f o(IR
d
), the following identities
are true:
1.
E[
i
f(F).G] = E[f(F) l
i
(G)] ,
where G l
i
(G) is linear and for any 1 < r < q < ,
sup
G
q,1
1
l
i
(G)
r
< +.
2. Similarly
E[
i
1
...i
k
f F.G] = E[f(F) l
i
1
...i
k
(G)]
and
sup
G
q,1
1
l
i
1
...i
k
(G)
r
< .
61
Proof: We have
(f F) =
d
i=1
i
f(F)F
i
hence
((f F), F
j
)
H
=
d
j=1
ij
i
f(F) .
Since is invertible, we obtain:
i
f(F) =
ij
((f F), F
j
)
H
.
Then
E[
i
f(F).G] =
j
E [
ij
((f F), F
j
)
H
G]
=
j
E [f F F
j
ij
G] ,
hence we see that l
i
(G) =
j
F
j
ij
G. Developing this expression gives
l
i
(G) =
j
[((
ij
G), F
j
)
H
ij
G/F
j
]
=
j
_
_
ij
(G, F
j
)
H
k,l
ik
jl
(
kl
, F
j
)
H
G
ij
G/F
j
_
_
.
Hence
[l
i
(G)[
j
_
kl
[
ik
jl
[ [
kl
[ [F
j
[ [G[
+[
ij
[ [F
j
[ [G[ + [
ij
[ [G[ [/F
j
[] .
Choose p such that
1
r
=
1
p
+
1
q
and apply Holders inequality:
l
i
(G)
r
d
j=1
_
k,l
G
q

ik
jl
[
kl
[
H
[F
j
[
H

p
+
+
ij
[F
j
[
p
[G[
q
+
ij
/F
j

p
G
q
_
G
q,1
_
d
j=1

ik
jl
[F
kl
[[F
j
[
p
+
+
ij
[F
j
[
p
+
ij
/F
j

p
_
.
62 Lifting of o
t
(IR
d
)
To prove the last part we iterate this procedure for i > 1.
Remember now that o(IR
d
) can be written as the intersection (i.e., pro
jective limit) of the Banach spaces S
2k
which are dened as below:
Let A = I +[x[
2
and dene f
2k
= A
k
f

2k
(G)
L
1
and
sup
G
q,2k
1
[E[f F.G][ A
k
f
sup
G
q,2k
1

2k
(G)
1
= Kf
2k
.
Consequently
f F
p,2k
Kf
2k
.
Corollary 6.2.5 The linear map f f F from o(IR
d
) into ID extends
continuously to a map from o
t
(IR
d
) into ID
t
whenever F ID(IR
d
) is non
degenerate.
As we have seen in Theorem 6.2.4 and Corollary 6.2.5, if F : W IR
d
is
a nondegenerate random variable, then the map f f F from o(IR
d
)
ID has a continuous extension to o
t
(IR
d
) ID
t
which we shall denote by
T T F.
63
For f o(IR
d
), let us look at the following Pettis integral:
_
IR
d
f(x)c
x
dx ,
where c
x
denotes the Dirac measure at x IR
d
. We have, for any g o(IR
d
),
_
_
f(x)c
x
dx, g
_
=
_
f(x)c
x
, g)dx
=
_
f(x)c
x
, g)dx
=
_
f(x)g(x)dx = f, g) .
Hence we have proven:
Lemma 6.2.6 The following representation holds in o(IR
d
):
f =
_
IR
d
f(x)c
x
dx .
From Lemma 6.2.6, we have
Lemma 6.2.7 We have
_
c
y
(F), )f(y)dy = E[f(F) ],
for any ID, where , ) denotes the bilinear form of duality between ID
t
and ID.
Proof: Let
be a mollier. Then c
y
c
y
in o
t
on the other hand
_
IR
d
(c
y
)(F)f(y)dy =
_
IR
d
(F + y) f(y)dy =
=
_
IR
d
(y)f(y + F)dy
0
f(F) .
On the other hand, for ID,
lim
0
_
IR
d
< (c
y
)
dx
(x) = p
F
(x),
moreover p
F
o(IR
d
) (i.e., the probability density of F is not only C
but
it is also a rapidly decreasing function).
Proof: We know that, for any ID, the map T T(F), ) is continuous
on o
t
(IR
d
) hence there exists some p
F,
o(IR
d
) such that
E[T(F) .] =
S
p
F,
, T)
S
.
Let p
F,1
= p
F
, then it follows from the Lemma 6.2.6 that
E[f(F)] =
_
c
y
(F), 1)f(y)dy
=
_
p
F
(y)f(y)dy .
Remark 6.2.9 From the disintegration of measures, we have
E[f(F) ] =
_
IR
d
p
F
(x)E[[F = x]f(x)dx
=
_
IR
d
f(x)c
x
(F), )dx
hence
E[[F = x] =
c
x
(F), )
p
F
(x)
dxalmost surely on the support of the law of F. In fact the right hand side
is an everywhere dened version of this conditional probability.
6.2.1 Extension of the Ito Formula
Let (x
t
) be the solution of the following stochastic dierential equation:
dx
t
(w) = b
i
(x
t
(w))dt +
i
(x
t
(w))dw
i
t
x
0
= x given,
where b : IR
d
IR
d
and
i
: IR
d
IR
d
are smooth vector elds with bounded
derivatives. Let us denote by
X
0
=
d
i=1
b
i
0
x
i
, X
j
=
j
i
x
j
Filtering 65
where
b
i
(x) = b
i
(x)
1
2
k,
(x)
k
(x).
If the Lie algebra of vector elds generated by X
0
, X
1
, . . . , X
d
has dimen
sion equal to d at any x IR
d
, then x
t
(w) is nondegenerate cf. [102]. In fact
it is also uniformly nondegenerate in the following sense:
E
_
t
s
[det(x
i
r
, x
j
r
)[
p
dr < ,
forall 0 < s < t and p > 1.
As a corollary of this result, combined with the lifting of o
t
to ID
t
, we can
show the following:
Theorem 6.2.10 For any T o
t
(IR
d
), one has the following:
T(x
t
) T(x
s
) =
_
t
s
AT(x
s
)ds +
_
t
s
ij
(x
s
)
j
T(x
s
)dW
i
s
,
where the Lebesgue integral is a Bochner integral, the stochastic integral is
as dened at the rst section of this chapter and we have used the following
notation:
A =
b
i
i
+
1
2
a
ij
(x)
2
x
i
x
j
, a(x) = (
)
ij
, = [
1
, . . . ,
d
] .
6.2.2 Applications to the ltering of the diusions
Suppose that we are given, for any t 0,
y
t
=
_
t
0
h(x
s
)ds + B
t
where h C
b
(IR
d
) IR
d
, B is another Brownian motion independent of w
above. The process (y
t
; t [0, 1]) is called an (noisy) observation of (x
t
, t
IR
+
). Let
t
= y
s
; s [0, t] be the observed data till t. The ltering
problem consists of calculating the random measure f E[f(x
t
)[
t
]. Let
P
0
be the probability dened by
dP
0
= Z
1
1
dP
where
Z
t
= exp
__
t
0
(h(x
s
), dy
s
)
1
2
_
t
0
[h(x
s
)[
2
ds
_
.
66 Applications
Then for any bounded,
t
measurable random variable Y
t
, we have:
E[f(x
t
).Y
t
] = E
_
Z
t
Z
t
f(x
t
).Y
t
_
= E
0
[Z
t
f(x
t
)Y
t
]
= E
0
_
E
0
[Z
t
f(x
t
)[
t
] Y
t
_
= E
_
1
E
0
[Z
t
[
t
]
E
0
[Z
t
f(x
t
)[
t
] Y
t
_
,
hence
E[f(x
t
)[
t
] =
E
0
[Z
t
f(x
t
)[
t
]
E
0
_
Z
t
[
t
_
.
If we want to study the smoothness of the measure f E[f(x
t
)[
t
], then
from the above formula, we see that it is sucient to study the smoothness of
f E
0
[Z
t
f(x
t
)[
t
]. The reason for the use of P
0
is that w and (y
t
; t [0, 1])
are two independent Brownian motions
1
under P
0
Remark 6.2.11 Let us note that the random distribution f
t
(f) dened
by
t
(f) = E
0
[Z
t
f(x
t
)[
t
]
satises the Zakai equation:
t
(f) =
0
(f) +
_
t
0
s
(Af)ds +
_
t
0
s
(h
i
f)dy
i
s
,
where A denotes the innitesimal generator of the diusion process (x
t
, t
IR
+
)
After this preliminaries, we can prove the following
Theorem 6.2.12 Suppose that the map f f(x
t
) from o(IR
d
) into ID has
a continuous extension as a map from o
t
(IR
d
) into ID
t
. Then the measure
f E[f(x
t
)[
t
] has a density in o(IR
d
).
Proof: As explained above, it is sucient to prove that the (random) mea
sure f E
0
[Z
t
f(x
t
)[
t
] has a density in o(IR
d
). Let /
y
be the Ornstein
Uhlenbeck operator on the space of the Brownian motion (y
t
; t [0, 1]). Then
we have
/
y
Z
t
= Z
t
_
_
t
0
h(x
s
)dy
s
+
1
2
_
t
0
[h(x
s
)[
2
ds
_
p
L
p
.
1
This claim follows directly from Paul Levys theorem of the characterization of the
Brownian motion.
Filtering 67
It is also easy to see that
/
k
w
Z
t
p
L
p
.
From these observations we draw the following conclusions:
Hence Z
t
(w, y) ID(w, y), where ID(w, y) denotes the space of test
functions dened on the product Wiener space with respect to the laws
of w and y.
The second point is that the operator E
0
[ [
t
] is a continuous mapping
from ID
p,k
(w, y) into ID
0
p,k
(y), for any p 1, k Z), since /
y
commutes
with E
0
[ [
t
] .
Hence the map
T E
0
[T(x
t
)Z
t
[
t
]
is continuous from o
t
(IR
d
) ID
t
(y). In particular, for xed T o
t
,
there exist p > 1 and k IN such that T(x
t
) ID
p,k
(w). Since
Z
t
ID(w, y),
Z
t
T(x
t
) ID
p,k
(w, y)
and
T(x
t
).(I +/
y
)
k/2
Z
t
ID
p,k
(w, y).
Consequently
E
0
[T(x
t
) (I +/
y
)
k/2
Z
t
[
t
] ID
p,k
(y).
Finally it follows from the latter that
(I +/)
k/2
E
0
[T(x
t
)(I +/
y
)
k/2
Z
t
[
t
] = E
0
[T(x
t
)Z
t
[
t
]
belongs to L
p
(y). Therefore we see that:
T E
0
[T(x
t
)Z
t
[
t
]
denes a linear, continuous (use the closed graph theorem for instance)
map from o
t
(IR
d
) into L
p
(y).
Since o
t
(IR
d
) is a nuclear space, the map
T
E
0
[T(x
t
)Z
t
[
t
]
is a nuclear operator. This implies that can be represented as
=
i=1
i
f
i
i
68 Applications of Ito Clark Formula
where (
i
) l
1
, (f
i
) o(IR
d
) and (
i
) L
p
(y) are bounded sequences.
Dene
k
t
(x, y) =
i=1
i
f
i
(x)
i
(y) o(IR
d
)
1
L
p
(y)
where
1
denotes the projective tensor product topology. It is easy now to
see that, for g o(IR
d
)
_
IR
d
g(x)k
t
(x, y)dx = E
0
[g(x
t
) Z
t
[
t
]
and this completes the proof.
6.3 Some applications of the Clark formula
6.3.1 Case of nondierentiable functionals
In this example we use the Clark representation theorem for the elements
of ID
t
and the composition of the tempered distributions with the non
degenerate Wiener functionals: Let w (w) be the sign of the ran
dom variable w W
1
(w) where W
1
denotes the value of the Wiener path
(W
t
, t [0, 1]) at time t = 1. We have, using Theorem 6.2.4
E[D
t
[T
t
] = 2 exp
_
W
2
t
2(1 t)
_
1
_
2(1 t)
,
dt dalmost surely. Hence
= 2
_
1
0
exp
_
W
2
t
2(1 t)
_
1
_
2(1 t)
dW
t
,
almost surely. Note that, although is not strongly Sobolev dierentiable,
the integrand of the stochastic integral is an ordinary square integrable pro
cess. This phenomena can be explained by the fact that the conditional
expectation tames the distribution, in such a way that the result becomes an
ordinary random variable.
Here is another application of the Clark formula:
Proposition 6.3.1 Assume that A is a measurable subset of W, then from
Theorem 6.1.4, there exists an e
A
L
2
(, H) which can be represented as
e
A
(t) =
_
t
0
e
A
()d, t [0, 1], such that e
A
is adapted and
1
A
= (A) + e
A
.
LogSobolev Inequality 69
If B is another measurable set, then A and B are independent if and only if
E [(e
A
, e
B
)
H
] = 0 .
Proof: It suces to observe that
(A B) = (A)(B) + E[(e
A
, e
B
)
H
] , (6.3.1)
hence A and B is independent if and only if the last term in (6.3.1) is null.
6.3.2 Logarithmic Sobolev Inequality
As another application of the Clark representation theorem, we shall give a
quick proof of the logarithmic Sobolev inequality of L. Gross
2
(cf. [36]).
Theorem 6.3.2 (logSobolev inequality) For any ID
2,1
, we have
E[
2
log
2
] E[
2
] log E[
2
] + 2E[[[
2
H
] .
Proof: Clearly it suces to prove the following inequality
E[f log f]
1
2
E
_
1
f
[f[
2
H
_
,
for any f ID
2,1
which is strictly positive, lower bounded with some > 0
and with E[f] = 1. Using the ItoClark representation theorem, we can write
f = exp
_
_
_
1
0
E[D
s
f[T
s
]
f
s
dW
s
1
2
_
1
0
_
E[D
s
f[T
s
]
f
s
_
2
ds
_
_
,
where f
s
= E[f[T
s
]. It follows from the Ito formula that
E[f log f] =
1
2
E
_
_
f
_
1
0
_
E[D
s
f[T
s
]
f
s
_
2
ds
_
_
.
Let be the probability dened by d = f d. Then we have
E[f log f] =
1
2
E
_
_
f
_
1
0
_
E[f D
s
log f[T
s
]
f
s
_
2
ds
_
_
2
The proof which is given here is similar to that of B. Maurey.
70 Applications
=
1
2
E
__
1
0
(E
[D
s
log f[T
s
])
2
ds
_
1
2
E
_
1
0
(D
s
log f)
2
ds
=
1
2
E[f [log f[
2
H
]
=
1
2
E
_
[f[
2
H
f
_
,
Remark 6.3.3 We have given the proof in the frame of the classical Wiener
space. However this result extends immediately to any abstract Wiener space
by the use of the techniques explained in the Appendix of the fourth chapter.
Remark 6.3.4 A straightforward implication of the Clark representation,
as we have seen in the sequel of the proof, is the Poincare inequality which
says that, for any F ID
2,1
, one has
E[[F E[F][
2
] E[[F[
2
H
] .
This inequality is the rst step towards the logarithmic Sobolev inequality.
Exercises
1. Assume that F : W X is a measurable Wiener function, where X is a separable
Hilbert space. Assume further that
F(w +h) F(w +k)
X
K[h k[
H
almost surely, for any h, k H. Prove that there exists F
= F almost surely
such that
F
(w +h) F
(w +k)
X
K[h k[
H
for any w W and h, k H.
2. Deduce from this result that if A is a measurable subset of W, such that A+h A
almost surely, then A has a modication, say A
(IR
d
)
is due to Kuo [50]. Watanabe has generalized it to more general Wiener functionals,
[102, 103]. Later it has been observed by the author that this implies automatically
the fact that the density of the law of a nondegenerate Wiener functional is a rapidly
decreasing C
function. This last result remains true for the conditional density of the
nonlinear ltering as it has been rst proven in [89].
72 Applications
Chapter 7
Positive distributions and
applications
7.1 Positive MeyerWatanabe distributions
If is a positive distribution on IR
d
, then a wellknown theorem says that
is a positive measure, nite on the compact sets. We will prove an analogous
result for the MeyerWatanabe distributions in this section, show that they
are absolutely continuous with respect to the capacities dened with respect
to the scale of the Sobolev spaces on the Wiener space and give an application
to the construction of the local time of the Wiener process. We end the
chapter by making some remarks about the Sobolev spaces constructed by
the second quantization of an elliptic operator on the CameronMartin space.
We will work on the classical Wiener space C
0
([0, 1]) = W. First we have
the following:
Proposition 7.1.1 Suppose (T
n
) ID
t
and each T
n
is also a probability
on W. If T
n
T in ID
t
, then T is also a probability and T
n
T in the
weakstar topology of measures on W.
For the proof of this proposition, we shall need the following:
Lemma 7.1.2 (GarsiaRademichRamsey lemma) Let p, be two con
tinuous, stritly increasing functions on IR
+
such that (0) = p(0) = 0 and
that lim
t
(t) = . Let T > 0 and f C([0, T], IR
d
). If
_
[0,T]
2
_
[f(t) f(s)[
p([t s[)
_
ds dt B,
73
74 Positive Distributions
then for any 0 s t T, we have
[f(t) f(s)[ 8
_
ts
0
1
_
4B
u
2
_
p(du) .
Proof: [of the Proposition] It is sucient to prove that the sequence
of probability measures (
n
, n 1) associated to (T
n
, n 1), is tight. In
fact, let S = ID C
b
(W), if the tightness holds, then we would have, for
= w lim
n
(taking a subsequence if necessary), where w lim denotes
the limit in the weakstar topology of measures,
() = T() on S .
Since the mapping w e
iw,w
)
(w
such that
[w(s) w(t)[ K
[t s[
15
4
,
for 0 s < t 1, hence A
= W
almost surely. Let C
(x) = (x ). We have
n
(A
c
)
_
W
(G(w))
n
(dw) .
We claim that _
W
(G)d
n
=
(G), T
n
).
To see this, for > 0, write
G
(w) =
_
[0,1]
2
[w(t) w(s)[
8
( +[t s[)
3
ds dt .
75
Then
(G
) S (but not
(G
)d
n
=
(G
), T
n
).
Moreover
(G
(G
), T
n
) =
(G), T
n
) .
From the dominated convergence theorem, we have also
lim
0
_
(G
)d
n
=
_
(G)d
n
.
Since T
n
T in ID
t
, there exist some k > 0 and p > 1 such that T
n
T in
ID
p,k
. Therefore
(G), T
n
) = (I +/)
k/2
(G), (I +/)
k/2
T
n
)
_
_
_(I +/)
k/2
(G)
_
_
_
q
sup
n
_
_
_(I +/)
k/2
T
n
_
_
_
p
.
From the Meyer inequalities, we see that
lim
_
_
_(I +/)
k/2
(G)
_
_
_
q
= 0 ,
in fact, it is sucient to see that
i
(
(G)) 0 in L
p
for all i [k] +1, but
this is obvious from the choice of
sup
n
n
(A
c
)
sup
n
_
_
_(I +/)
k/2
T
n
_
_
_
p
lim
_
_
_(I +/)
k/2
(G)
_
_
_
p
= 0 ,
which implies the tightness and the proof is completed.
Corollary 7.1.3 Let T ID
t
such that T, ) 0, for all positive ID.
Then T is a Radon measure on W.
Proof: Let (h
i
) H be a complete, orthonormal basis of H. Let V
n
=
h
1
, . . . , h
n
. Dene T
n
as T
n
= E[P
1/n
T[V
n
] where P
1/n
is the Ornstein
Uhlenbeck semigroup on W. Then T
n
0 and it is a random variable
in some L
p
(). Therefore it denes a measure on W (it is even absolutely
continuous with respect to ). Moreover T
n
T in ID
t
, hence the proof
follows from Proposition 7.1.1.
76 Capacities
7.2 Capacities and positive Wiener functio
nals
We begin with the following denitions:
Denition 7.2.1 Let p [1, ) and k > 0. If O W is an open set, we
dene the (p, k)capacity of O as
C
p,k
(O) = inf
p
p,k
: ID
p,k
, 1 a.e. on O .
If A W is any set, dene its (p, k)capacity as
C
p,k
(A) = infC
p,k
(O); O is open O A .
We say that some property takes place (p, k)quasi everywhere if the
set on which it does not hold has (p, k)capacity zero.
We say N is a slim set if C
p,k
(N) = 0, for all p > 1, k > 0.
A function is called (p, k)quasi continuous if for any > 0 , there
exists an open set O
such that C
p,k
(O
.
A function is called quasi continuous if it is (p, k)quasi continuous
for any p > 1, k IN.
The results contained in the next lemma are proved by Fukushima & Kaneko
(cf. [33]):
Lemma 7.2.2 1. If F ID
p,k
, then there exists a (p, k)quasi continuous
function
F such that F =
F a.e. and
F is (p, k)quasi everywhere
dened, i.e. if
G is another such function, then C
p,k
(
F ,=
G)) = 0.
2. If A W is arbitrary, then
C
p,k
(A) = inf
p,k
: ID
p,k
, 1 (p, r) q.e. on A
3. There exists a unique element U
A
ID
p,k
such that
U
A
1 (p, k)quasi
everywhere on A with C
p,k
(A) = U
A

p,k
, and
U
A
0 (p, k)quasi
everywhere. U
A
is called the (p, k)equilibrium potential of A.
Local Time 77
Theorem 7.2.3 Let T ID
t
be a positive distribution and suppose that T
ID
q,k
for some q > 1, k 0. Then, if we denote by
T
the measure associated
to T, we have
T
(A) T
q,k
(C
p,k
(A))
1/p
,
for any set A W, where
T
denotes the outer measure with respect to
T
.
In particular
T
does not charge the slim sets.
Proof: Let V be an open set in W and let U
V
be its equilibrium potential
of order (p, k). We have
P
1/n
T, U
V
) =
_
P
1/n
T U
V
d
_
V
P
1/n
T U
V
d
_
V
P
1/n
Td
=
P
1/n
T
(V ) .
Since V is open, we have, from the fact that
P
1/n
T
T
weakly,
liminf
n
P
1/n
T
(V )
T
(V ) .
On the other hand
lim
n
P
1/n
T, U
V
) = T, U
V
)
T
q,k
U
V

p,k
= T
q,k
C
p,k
(V )
1/p
.
7.3 Some Applications
Below we use the characterization of the positive distributions to give a
dierent interpretation of the local times. Afterwards the 01 law is revisited
via the capacities.
7.3.1 Applications to Ito formula and local times
Let f : IR
d
IR be a function from o
t
(IR
d
) and suppose that (X
t
, t 0)
is a hypoelliptic diusion on IR
d
which is constructed as the solution of the
78 Ito Formula
following stochastic dierential equation with smooth coecients:
dX
t
= (X
t
)dW
t
+ b(X
t
)dt (7.3.1)
X
0
= x IR
d
.
We denote by L the innitesimal generator of the diusion process (X
t
, t
0). For any t > 0, X
t
is a nondegenerate random variable in the sense of
Denition 6.2.1. Consequently we have the extension of the Ito formula
f(X
t
) f(X
u
) =
_
t
u
Lf(X
s
)ds +
_
t
u
ij
(X
s
)
i
f(X
s
)dW
j
s
,
for 0 < u t 1. Note that, since we did not make any dierentiability hy
pothesis about f, the above integrals are to be regarded as the elements of ID
t
.
Suppose that Lf is a bounded measure on IR
d
, from our result about the pos
itive distributions, we see that
_
t
u
Lf(X
s
)ds is a measure on W which does not
charge the slim sets. By dierence, so does the term
_
t
u
ij
(X
s
)
i
f(X
s
)dW
j
s
.
As a particular case, we can take d = 1, L =
1
2
(i.e. = 1), f(x) = [x[ and
this gives
[W
t
[ [W
u
[ =
1
2
_
t
u
[x[(W
s
)ds +
_
t
u
d
dx
[x[(W
s
)dW
s
.
As
d
dx
[x[ = sign(x), we have
_
t
u
d
dx
[x[(W
s
)dW
s
=
_
t
u
sign(W
s
)dW
s
= M
u
t
is a measure absolutely continuous with respect to . Since lim
u0
M
u
t
= N
t
exists in all L
p
, so does
lim
u0
_
t
u
[x[(W
s
)ds
in L
p
for any p 1. Consequently
_
t
0
[x[(W
s
)ds is absolutely continuous
with respect to , i.e., it is a random variable. It is easy to see that
[x[(W
s
) = 2c
0
(W
s
) ,
where c
0
denotes the Dirac measure at zero, hence we obtain
_
t
0
2c
0
(W
s
)ds =
_
t
0
[x[(W
s
)ds
= 2l
0
t
which is the local time of Tanaka. Note that, although c
0
(W
s
) is singular
with respect to , its Pettis integral is absolutely continuous with respect to
.
0 1Law 79
Remark 7.3.1 If F : W IR
d
is a nondegenerate random variable, then
for any S o
t
(IR
d
) with S 0 on o
+
(IR
d
), S(F) ID
t
is a positive distri
bution, hence it is a positive Radon measure on W. In particular c
x
(F) is a
positive Radon measure.
7.3.2 Applications to 0 1 law and to the gauge func
tionals of sets
In Theorem 6.1.5 we have seen that an Hnvariant subset of W has measure
which is equal either to zero or to one. In this section we shall rene this
result using the capacities. Let us rst begin by dening the gauge function
of a measurable subset of W: if A B(W), dene
q
A
(w) = inf[[h[
H
: h (A w) H] , (7.3.2)
where the inmum is dened as to be innity on the empty set. We have
Lemma 7.3.2 For any A B(W), the map q
A
is measurable with respect
to the completion of B(W). Moreover
[q
A
(w + h) q
A
(w)[ [h[
H
(7.3.3)
almost surely, for any h H and q
A
< = 0 or 1.
Proof: Without loss of generality, we may assume that A is a compact subset
of W with (A) > 0. Then the set K(w) = (A w) H ,= almost surely.
Therefore w K(w) is a multivalued map with values in the nonempty
subsets of H for almost all w W. Let us denote by G(K) its graph, i.e.,
G(K) = (h, w) : h K(w) .
Since (h, w) h+w is measurable from HW to W when the rst space is
equipped with the product sigma algebra, due to the continuity of the map
(h, w) w + h, it follows that G(K) is a measurable subset of H W.
From a theorem about the measurable multivalued maps, it follows that
w K(w) is measurable with respect to the  completed sigma eld B(W)
(cf. [16]). Hence there is a countable sequence of Hvalued measurable
selectors (u
i
, i IN) of K (i.e., u
i
: W H such that u
i
(w) K(w) almost
surely) such that (u
i
(w), i IN) is dense in K(w) almost surely. To see the
measurability, it suces to remark that
q
A
(w) = inf([u
i
(w)[
H
: i IN) .
80 Applications
The relation 7.3.3 is evident from the denition of q
A
. To complete the proof
it suces to remark that the set Z = w : q
A
(w) < is Hinvariant,
hence from Theorem 6.1.5, (Z) = 0 or 1. Since Z contains A and (A) > 0,
(Z) = 1.
The following result renes the 0 1law (cf. [29], [52]):
Theorem 7.3.3 Assume that A W is an Hinvariant set of zero Wiener
measure. Then
C
r,1
(A) = 0
for any r > 1.
Proof: Choose a compact K A
c
with (K) > 0. Denote by B
n
the ball
of radius n of H and dene K
n
= K +B
n
. It is easy to see that
n
K
n
is an
Hinvariant set. Moreoever
(
n
K
n
) A = ,
otherwise, due to the Hinvariance of of A, we would have A K ,= . We
also have (K
n
) 1. Let
p
n
(w) = q
K
n
(w) 1 .
From Proposition 5.1.9, we see that p
n
p
ID
p,1
. Moreover p
n
(w) = 1 on
K
c
n+1
(hence on A) by construction. Since p
n
= 0 on K
n
, from Lemma 2.5.1
p
n
= 0 almost surely on K
n
. Consequently
C
r,1
(A)
_
([p
n
[
r
+[p
n
[
r
H
)d
=
_
K
c
n
([p
n
[
r
+[p
n
[
r
H
)d
2(K
c
n
) 0
as n .
7.4 Local Sobolev spaces
In Chapter II we have observed the local character of the Sobolev derivative
and the divergence operator. This permits us to dene the local Sobolev
spaces as follows:
Local Sobolev Spaces 81
Denition 7.4.1 We say that a Wiener functional F with values in some
separable Hilbert space X belongs to ID
loc
p,1
(X), p > 1, if there exists a sequence
(
n
, n 1) of measurable subsets of W whose union is equal to W almost
surely and
F = F
n
a.s. on
n
,
where F
n
ID
p,1
(X) for any n 1. We call ((
n
, F
n
), n 1) a localizing
sequence for F.
Lemma 2.5.1 and Lemma 2.5.2 of Section 2.5 permit us to dene the local
Sobolev derivative and local divergence of the Wiener functionals. In fact, if
F ID
loc
p,1
(X), then we dene
loc
F as
loc
F = F
n
on
n
.
Similarly, if ID
loc
p,1
(X H), then we dene
loc
=
n
on
n
.
From the lemmas quoted above
loc
F and
loc
are independent of the choice
of their localizing sequences.
Remark: Note that we can dene also the spaces ID
loc
p,k
(X) similarly.
The most essential property of the Sobolev derivative and the divergence
operator is the fact that the latter is the adjoint of the former under the
Wiener measure. In other words they satisfy the integration by parts formula:
E[(, )
H
] = E[] .
In general this important formula is no longer valid when we replace and
with
loc
and
loc
respectively. The theorem given below gives the exact
condition when the local derivative or divergence of a vector eld is in fact
equal to the global one.
Theorem 7.4.2 Assume that ID
loc
p,1
(X), and let ((
n
,
n
), n IN) be
a localizing sequence of . A neccessary and sucient condition for
ID
p,1
(X) and for =
loc
almost surely, is
lim
n
C
p,1
(
c
n
) = 0 . (7.4.4)
Proof: The neccessity is trivial since, from Lemma 7.2.2. To prove the
suciency we can assume without loss of generality that is bounded. In
fact, if the theorem is proved for the bounded functions, then to prove the
general case, we can replace by
k
=
_
1 +
1
k

X
_
1
,
82 Applications
which converges in ID
p,1
(X) as k due to the closedness of the Sobolev
derivative. Hence we shall assume that is bounded. Let > 0 be arbitrary,
since C
p,1
(
c
n
) 0, by Lemma 7.2.2, there exists some F
n
ID
p,1
such that
F
n
1 on
c
n
quasieverywhere and F
n

p,1
C
p,1
(
c
n
) + 2
n
, for any
n IN. Evidently, the sequence (F
n
, n IN) converges to zero in ID
p,1
. Let
f : IR [0, 1] be a smooth function such that f(t) = 0 for [t[ 3/4 and
f(t) = 1 for [t[ 1/2. Dene A
n
= f F
n
, then A
n
= 0 on
c
n
quasi
everywhere and the sequence (A
n
, n IN) converges to the constant 1 in
ID
p,1
. As a consequence of this observation A
n
=
n
A
n
almost surely and
by the dominated convergence theorem, (
n
A
n
, n IN) converges to in
L
p
(, X). Moreover
(A
n
) = (
n
A
n
)
= A
n
loc
+ A
n
loc
in L
p
(, X H) since (A
n
, n IN) and are bounded. Consequently
(
n
A
n
)
loc
in L
p
(, XH), since is a closed operator on L
p
(, X)
the convergence takes place also in ID
p,1
(X) and the proof is completed.
We have also a similar result for the divergence operator:
Theorem 7.4.3 Let be in ID
loc
p,1
(H) with a localizing sequence ((
n
,
n
), n
IN) such that L
p
(, H) and
loc
L
p
(). Assume moreover
lim
n
C
q,1
(
c
n
) = 0 , (7.4.5)
where q = p/(p 1). Then Dom
p
() and
loc
= almost surely.
Proof: Due to the hypothesis (7.4.5), we can construct a sequence (A
n
, n
IN) as in the proof of Theorem 7.4.2, which is bounded in L
(), converging
to the constant function 1 in ID
q,1
such that A
n
= 0 on
c
n
. Let ID be
bounded, with a bounded Sobolev derivative. We have
E
_
A
n
(
loc
)
_
= E [A
n
n
]
= E [A
n
(
n
, )
H
] + E [(A
n
,
n
)
H
]
= E [A
n
(, )
H
] + E [(A
n
, )
H
]
E [(, )
H
] .
Moreover, from the dominated convergence theorem we have
lim
n
E[A
n
(
loc
)] = E[(
loc
)] ,
(A)Distributions 83
hence
E
_
(
loc
)
_
= E [(, )
H
] .
Since the set of functionals with the above prescribed properties is dense
in ID
q,1
, the proof is completed.
7.5 Distributions associated to (A)
It is sometimes useful to have a scale of distribution spaces which are dened
with a more elliptic operator than the OrnsteinUhlenbeck semigroup. In
this way objects which are more singular than Meyer distributions can be
interpreted as the elements of the dual space. This is important essentially
for the constructive Quantum eld theory, cf. [77]. We begin with an abtract
Wiener space (W, H, ). Let A be a selfadjoint operator on H, we suppose
that its spectrum lies in (1, ), hence A
1
is bounded and A
1
 < 1. Let
H
n
Dom(A
n
) ,
hence H
is dense in H and (A
h, h)
H
is increasing. Denote by H
the completion of H
= (A
h, h); IR.
Evidently H
t
= H
n=0
I
n
(
n
), dene the second quantization of A
(A) = E[] +
n=1
I
n
(A
n
n
) .
Denition 7.5.1 For p > 1, k Z, IR, we dene ID
p,k
as the completion
of polynomials based on H
. If
is a separable Hilbert space, ID
p,k
() is dened similarly except that is
taken as an valued polynomial.
Remark 7.5.2 If = exp(h
1
2
[h[
2
) then we have
(A) = exp
_
(Ah)
1
2
[Ah[
2
_
.
Remark 7.5.3 ID
p,k
is decreasing with respect to , p and k.
84 Applications
Theorem 7.5.4 Let (W
, H
. Let us denote by ID
()
p,k
the Sobolev
space on W
dened by

ID
()
p,k
= (I +/)
k/2

L
p
(
,W
)
Then ID
()
p,k
and ID
p,k
are isomorphic.
Remark: This isomorphism is not algebraic, i.e., it does not commute with
the pointwise multiplication.
Proof: We have
E[e
i(A
/2
h)
] = exp
1
2
[A
/2
h[
2
= exp
[h[
2
2
which is the characteristic function of
on W
.
Theorem 7.5.5 1. For p > 2, IR, k Z, there exists some >
2
such that

ID
p,k

ID
2,k
consequently
,k
ID
2,k
=
,p,k
ID
p,k
.
2. Moreover, for some > we have

ID
p,k

ID
2,0
,
hence we have also
ID
2,0
=
,p,k
ID
p,k
.
Proof: 1) We have

ID
p,k
= 
n
(1 + n)
k/2
I
n
((A
/2
)
n
n
)
L
p
=
_
_
_
(1 + n)
k/2
e
nt
e
nt
I
n
((A
/2
)
n
n
)
_
_
_
L
p
.
From the hypercontractivity of P
t
, we can choose t such that p = e
2t
+1 then
_
_
_
(1 + n)
k/2
e
nt
e
nt
I
n
(. . .)
_
_
_
p
_
_
_
(1 + n)
k/2
e
nt
I
n
(. . .)
_
_
_
2
.
(A)Distributions 85
Choose > 0 such that A
 e
t
, hence
_
_
_
(1 + n)
k/2
e
nt
I
n
(. . .)
_
_
_
2
_
_
_
(1 + n)
k/2
(A
)(A
)e
nt
I
n
((A
/2
)
n
n
)
_
_
_
2
(1 + n)
k/2
I
n
((A
+/2
)
n
n
))
2
= 
ID
2+
2,k
.
2) If we choose A
 < e
t
then the dierence suces to absorb the action
of the multiplicator (1 +n)
k/2
which is of polynomial growth and the former
gives an exponential decrease.
Corollary 7.5.6 We have similar relations for any separable Hilbert space
valued functionals.
Proof: This statement follows easily from the Khintchine inequality.
As another corollary we have
Corollary 7.5.7 Let us denote by (H
) the space
(H
). Then
1. : (H
) and : (H
) and :
t
(H
)
t
.
2. is an algebra.
3. For any T
t
, there exists some
t
(H
) such that
T = T, 1) + .
Proof: The rst claim follows from Theorems 7.5.4 and 7.5.5. To prove the
second one it is sucient to show that
2
if . This follows from the
multiplication formula of the multiple Wiener integrals. (cf. Lemma 8.1.1).
To prove the last one let us observe that if T
t
, then there exists some
> 0 such that T ID
2,0
, i.e., T under the isomorphism of Theorem 7.5.4
is in L
2
(
, W
1
, . . . , h
n
) with h
i
are the eigenvectors
of A.
86 Applications
7.6 Applications to positive distributions
Let T
t
be a positive distribution. Then, from the construction of the dis
tribution spaces, there exists some ID
p,k
such that T ID
p,k
and T, ) 0
for any ID
q,k
, 0. Hence i
. Therefore i
n=0
1
n!
E [[u[
n
H
]
1/2
< . (7.6.6)
Then the mapping dened by
E[(w + u(w))] =< L
u
, >
is a positive distribution and it can be expressed as
n=0
1
n!
n
u
n
.
Moreover this sum is weakly uniformly convergent in ID
2,0
, for any > 0
such that A
1

2
<
1
2
.
Proof: It follows trivially from the Taylor formula and from the denition
of
n
as the adjoint of
n
with respect to , that
< L
u
, >=
n=0
1
n!
E
_
n
u
n
_
for any cylindrical, analytic function . To complete the proof it suces
to show that < L
u
, > extends continuously to . If has the chaos
decomposition
=
k=0
I
k
(
k
) ,
with
k
H
k
, then
E
_

n

2
H
k
_
=
kn
(k!)
2
(k n)!

k

2
H
k
kn
c
k
(k!)
2
(k n)!

k

2
H
k
,
(A)Distributions 87
where c
n=0
1
n!
[ <
n
, u
n
> [
n=0
1
n!
E [[u[
n
H
]
1/2
_
_
kn
c
k
(k!)
2
(k n)!

k

2
H
k
_
_
1/2
n=0
1
n!
E [[u[
n
H
]
1/2
_
_
kn
(2c
)
k

k

2
H
k
_
_
1/2
.
Choose now such that c
k=0
k!
k

2
H
k
= 
2
ID
2,0
.
Hence the sum is absolutely convergent provided that u satises the condition
(7.6.6).
7.7 Exercises
1. Let K be a closed vector subspace of H and denote by P the orthogonal projection
associated to it. Denote by T
K
the sigma algebra generated by k, k K. Prove
that
(P)f = E[f[T
K
] ,
for any f L
2
().
2. Assume that M and N are two closed vector subspaces of the CameronMartin space
H, denote by P and Q respectively the corresponding orthogonal projections. For
any f, g L
2
() prove the following inequality:
[E [(f E[f])(g E[g])][ PQf
L
2
()
g
L
2
()
,
where PQ is the operator norm of PQ.
3. Prove that (e
t
I
H
) = P
t
, t 0, where P
t
denotes the OrnsteinUhlenbeck semi
group.
4. Let B is a bounded operator on H, dene d(B) as
d(B)f =
d
dt
(e
tB
)f[
t=0
.
Prove that
d(B)f = Bf
and that
d(B)(fg) = f d(B)g +g d(B)f ,
(i.e., d(B) is a derivation) for any f, g ID whenever B is skewsymmetric.
88 Applications
Notes and suggested reading
The fact that a positive Meyer distribution denes a Radon measure on the Wiener space
has been indicated for the rst time in [3]. The notion of the capacity in an abstract frame
has been studied by several people, cf. in particular [12], [56] and the references there.
Application to the local times is original, the capacity version of 0 1law is taken from
[52]. Proposition 7.6.1 is taken from [46], for the more general distribution spaces we refer
the reader to [46, 48, 63] and to the references there.
Chapter 8
Characterization of
independence of some Wiener
functionals
Introduction
In probability theory, probably the most important concept is the indepen
dence since it is the basic property which dierentiates the probability theory
from the abstract measure theory or from the functional analysis. Besides
it is almost always dicult to verify the independence of random variables.
In fact, even in the elementary probability, the tests required to verify the
independence of three or more random variables get very quickly quite cum
bersome. Hence it is very tempting to try to characterize the independence
of random variables via the local operators as or that we have studied
in the preceding chapters.
Let us begin with two random variables: let F, G ID
p,1
for some p > 1.
They are independent if and only if
E[e
iF
e
iG
] = E[e
iF
]E[e
iG
]
for any , IR, which is equivalent to
E[a(F)b(G)] = E[a(F)]E[b(G)]
for any a, b C
b
(IR).
Let us denote by a(F) = a(F) E[a(F)], then we have:
F and G are independent if and only if
E[ a(F) b(G)] = 0 , a, b C
b
(IR) .
89
90 Independence
Since e
ix
can be approximated pointwise with smooth functions, we can
suppose as well that a, b C
1
b
(IR) (or C
0
(IR)). Since / is invertible on the
centered random variables, we have
E[ a(F)b(G)] = E[//
1
a(F) b(G)]
= E[/
1
a(F) b(G)]
= E[(/
1
a(F), (b(G)))
H
]
= E[((I +/)
1
a(F), (b(G)))]
= E[((I +/)
1
(a
t
(F)F), b
t
(G)G)
H
]
= E[b
t
(G) ((I +/)
1
(a
t
(F)F), G)
H
]
= E[b
t
(G) E[((I +/)
1
(a
t
(F)F, G)
H
[(G)]] .
In particular choosing a = e
ix
, we nd that
Proposition 8.0.1 F and G (in ID
p,1
) are independent if and only if
E
_
((I +/)
1
(e
iF
F), G)
H
(G)
_
= 0 a.s.
8.1 The case of multiple Wiener integrals
Proposition 8.0.1 is not very useful, because of the nonlocalness property
of the operator /
1
. Let us however look at the case of multiple Wiener
integrals:
First recall the following multiplication formula of the multiple Wiener
integrals:
Lemma 8.1.1 Let f
L
2
([0, 1]
p
), g
L
2
([0, 1]
q
). Then we have
I
p
(f) I
q
(g) =
pq
m=0
p! q!
m!(p m)!(q m)!
I
p+q2m
(f
m
g) ,
where f
m
g denotes the contraction of order m of the tensor f g, i.e., the
partial scalar product of f and g in L
2
([0, 1]
m
).
By the help of this lemma we will prove:
Theorem 8.1.2 I
p
(f) and I
q
(g) are independent if and only if
f
1
g = 0 a.s. on [0, 1]
p+q2
.
91
Proof: () : By independence, we have
E[I
2
p
I
2
q
] = p!f
2
q!g[[
2
= p!q![[f g
2
.
On the other hand
I
p
(f)I
q
(g) =
pq
0
m!C
m
p
C
m
q
I
p+q2m
(f
m
g) ,
hence
E[(I
p
(f)I
q
(g))
2
]
=
pq
0
(m!C
m
p
C
m
q
)
2
(p + q 2m)!f
m
g
2
(p + q)!f
g[[
2
(dropping the terms with m 1) .
We have, by denition:
f
g
2
=
_
_
_
1
(p + q)!
S
p+q
f(t
(1)
, . . . , t
(p)
)g(t
(p+1)
, . . . , t
(p+q)
)
_
_
_
2
=
1
((p + q)!)
2
,S
p+q
,
,
where S
p+q
denotes the group of permutations of order p + q and
,
=
_
[0,1]
p+q
f(t
(1)
, . . . , t
(p)
)g(t
(p+1)
, . . . , t
(p+q)
)
f(t
(1)
, . . . , t
(p)
)g(t
(p+1)
, . . . , t
(p+q)
)dt
1
. . . dt
p+q
.
Without loss of generality, we may suppose that p q. Suppose now
that ((1), . . . , (p)) and ((1), . . . , (p)) has k 0 elements in common. If
we use the block notations, then
(t
(1)
, . . . , t
(p)
) = (A
k
,
A)
(t
(p+1)
, . . . , t
(p+q)
) = B
(t
(1)
, . . . , t
(p)
) = (A
k
,
C)
(t
(p+1)
, . . . , t
(p+q)
) = D
where A
k
is the subblock containing elements common to (t
(1)
, . . . , t
(p)
)
and (t
(1)
, . . . , t
(p)
). Then we have
,
=
_
[0,1]
p+q
f(A
k
,
A)g(B) f(A
k
,
C)g(D)dt
1
. . . dt
p+q
.
92 Independence
Note that A
k
A B = A
k
C D = t
1
, . . . , t
p+q
,
A
C = . Hence we
have
A B =
C D. Since
A
C = , we have
C B and
A D. From
the fact that (
A, B) and (
C, D) are the partitions of the same set, we have
D
A = B
,
=
=
_
[0,1]
p+q
f(A
k
,
A)g(
C, B
C) f(A
k
,
C)g(
A, D
A)dt
1
. . . dt
p+q
=
_
[0,1]
p+q
f(A
k
,
A)g(
C, B
C)f(A
k
,
C)g(
A, B
C)dA
k
d
Ad
Cd(B
C)
=
_
[0,1]
qp+2k
(f
pk
g)(A
k
, B
C)(f
pk
g)(A
k
, B
C) dA
k
d(B
C)
= f
pk
g
2
L
2
([0,1]
qp+2k
)
where we have used the relation D
A = B
,
(k = p) +
,
(k ,= p))
_
_
.
The number of
,
with (k = p) is exactly
_
p+q
p
_
(p!)
2
(q!)
2
, hence we have
p!q!f
2
g
2
p!q!f g
2
+
p1
k=0
c
k
f
pk
g
2
L
2
([0,1]
qp+2k
)
with c
k
> 0. For this relation to hold we should have
f
pk
g = 0 , k = 0, . . . , p 1
in particular for k = p 1, we have
f
1
g = 0 .
(): From the Proposition 8.0.1, we see that it is sucient to prove
((I +/)
1
e
iF
F, I
q
(g)) = 0 a.s.
with F = I
p
(f), under the hypothesis f
1
g = 0 a.s. Let us write
e
iI
p
(f)
=
k=0
I
k
(h
k
) ,
93
then
e
iI
p
(f)
I
p
(f) = p
k=0
I
k
(h
k
) I
p1
(f)
= p
k=0
k(p1)
r=0
p,k,r
I
p1+k2r
(h
k
r
f) .
Hence
(I +/)
1
e
iF
F = p
k
k(p1)
r=0
(1 + p + k 1 2r)
1
I
p1+k2r
(h
k
p
f) .
When we take the scalar product with I
q
(g), we will have terms of the type:
(I
p1+k2r
(h
k
r
f), I
q1
(g))
H
=
=
i=1
I
p1+k2r
(h
k
r
f(e
i
))I
q1
(g(e
i
)) .
If we use the multiplication formula to calculate each term, we nd the terms
as
i=1
_
(h
k
r
f(e
i
))(t
1
, . . . , t
p+k2r1
)g(e
i
)(t
1
, . . . , t
q1
)dt
1
dt
2
. . .
=
_ _
1
=0
(h
k
r
f())(t
1
, . . . , t
p+k2r1
)g(, t
1
, . . . , t
q1
)d dt
1
. . .
From the hypothesis we have
_
1
0
f(, t
1
. . .)g(, s
1
. . . , )d = 0 a.s. ,
hence the Fubini theorem completes the proof.
Remark: For a more elementary proof of the suciency of Theorem 8.1.2
cf. [43].
Remark 8.1.3 In the proof of the necessity we have used only the fact
that I
p
(f)
2
and I
q
(g)
2
are independent. Hence, as a byproduct we obtain
also the fact that I
p
and I
q
are independent if and only if their squares are
independent.
Corollary 8.1.4 Let f and g be symmetric L
2
kernels respectively on [0, 1]
p
and [0, 1]
q
. Let
S
f
= spanf
p1
h : h L
2
([0, 1])
p1
94 Independence
and
S
g
= spang
q1
k; k L
2
(]0, 1]
q1
) .
Then the following are equivalent:
i) I
p
(f) and I
q
(g) are independent,
ii) I
p
(f)
2
and I
q
(g)
2
are independent,
iii) S
f
and S
g
are orthogonal in H,
iv) the Gaussiangenerated elds I
1
(k); k S
f
and I
1
(l); l S
g
are independent.
Proof: As it is indicated in Remark 8.1.3, the independence of I
p
and I
q
is
equivalent to the independence of their squares.
(iiii): The hypothesis implies thatf
1
g = 0 a.s. If a S
f
, b S
g
then
they can be written as nite linear combinations of the vectors f
p1
h
and g
q1
k respectively. Hence, it suces to assume, by linearity, that
a = f
p1
h and b = g
q1
k. Then it follows from the Fubini theorem
(a, b) = (f
p1
h, g
q1
k) = (f
1
g, h k)
(L
2
)
p+q2
= 0 .
(iiii) If (f
1
g, h k) = 0 for all h L
2
([0, 1]
p1
), k L
2
([0, 1]
q1
), then
f
1
g = 0 a.s. since nite combinations of h k are dense in L
2
([0, 1]
p+q2
).
Finally, the equivalence of (iii) and (iv) is obvious.
Proposition 8.1.5 Suppose that I
p
(f) is independent of I
q
(g) and I
p
(f) is
independent of I
r
(h). Then I
p
(f) is independent of I
q
(g), I
r
(h).
Proof: We have f
1
g = f
1
h = 0 a.s. This implies the independence of
I
p
(f) and I
g
(g), I
r
(h) from the calculations similar to those of the proof of
suciency of the theorem.
In a similar way we have
Proposition 8.1.6 Let I
p
(f
); J and I
q
(g
); K be two arbi
trary families of multiple Wiener integrals. The two families are independent
if and only if I
p
(f
) is independent of I
q
(g
) for all (, ) J K.
Corollary 8.1.7 If I
p
(f) and I
q
(g) are independent, so are also I
p
(f)(w+
h)
and I
q
(g)(w +
k) for any
h,
k H.
95
Proof: Let us denote, respectively, by h and k the Lebesgue densities of
h
and
k. We have then
I
p
(f)(w +
h) =
p
i=0
_
p
i
_
(I
pi
(f), h
i
)
H
i .
Let us dene f[h
i
] L
2
[0, 1]
pi
by
I
pi
(f[h
i
]) = (I
pi
(f), h
i
).
If f
1
g = 0 then it is easy to see that
f[h
i
]
1
g[k
j
] = 0 ,
hence the corollary follows from Theorem 8.1.2.
From the corollary it follows
Corollary 8.1.8 I
p
(f) and I
q
(g) are independent if and only if the germ
elds
I
p
(f), I
p
(f), . . . ,
p1
I
p
(f)
and
I
q
(g), . . . ,
q1
I
q
(g)
are independent.
Corollary 8.1.9 Let X, Y L
2
(), Y =
0
I
n
(g
n
). If
X
1
g
n
= 0 a.s. n,
then X and Y are independent.
Proof: This follows from Proposition 8.0.1.
Corollary 8.1.10 In particular, if
h H, then
h
= 0 a.s. implies that
and I
1
(h) =
h are independent.
8.2 Exercises
1. Letf
L
2
([0, 1]
p
) and h L
2
([0, 1]). Prove the product formula
I
p
(f)I
1
(h) = I
p+1
(f h) +pI
p1
(f
1
h) . (8.2.1)
2. Prove by induction and with the help of (8.2.1), the general multiplication formula
I
p
(f) I
q
(g) =
pq
m=0
p! q!
m!(p m)!(q m)!
I
p+q2m
(f
m
g) ,
where f
L
2
([0, 1]
p
) and g
L
2
([0, 1]
q
).
96 Independence
Notes and suggested reading
All the results of this chapter are taken from [93, 94], cf. also [43] for some simplication of
the suciency of Theorem 8.1.2. Note that, in Theorem 8.1.2, we have used only the inde
pendence of I
p
(f)
2
and I
q
(g)
2
. Hence two multiple ItoWiener integrals are independent
if and only if their squares are independent.
Chapter 9
Moment inequalities for Wiener
functionals
Introduction
In several applications, as limit theorems, large deviations, degree theory of
Wiener maps, calculation of the RadonNikodym densities, etc., it is impor
tant to control the (exponential) moments of Wiener functionals by those
of their derivatives. In this chapter we will give two results on this subject.
The rst one concerns the tail probabilities of the Wiener functionals with
essentially bounded GrossSobolev derivatives. This result is a straightfor
ward generalization of the celebrated Ferniques lemma which says that the
square of the supremum of the Brownian path on any bounded interval has an
exponential moment provided that it is multiplied with a suciently small,
positive constant. The second inequality says that for a Wiener functional
F ID
p,1
, we have
E
w
E
z
[U(F(w) F(z))] E
w
E
z
_
U
_
2
I
1
(F(w)
_
(z)
_
, (9.0.1)
where w and z represent two independent Wiener paths, E
w
and E
z
are
the corresponding expectations, and I
1
(F(w))(z) is the rst order Wiener
integral with respect to z of F(w) and U is any lower bounded, convex
function on IR. Then combining these two inequalities we will obtain some
interesting majorations.
In the next section we show that the logSobolev inequality implies the ex
ponential integrability of the square of the Wiener functionals whose deriva
tives are essentially bounded. In this section we study with general measures
which satisfy a logarithmic Sobolev inequality.
97
98 Moment Inequalities
The next inequality is an interpolation inequality which says that the
Sobolev norm of rst order can be upper bounded by the product of the
second order and of the zeroth order Sobolev norms.
In the last part we study the exponential integrability of the Wiener
functionals in the divergence form, a problem which has gained considerable
importance due to the degree theorem on the Wiener space as it is explained
in more detail in the notes at the end of this chapter.
9.1 Exponential tightness
First we will show the following result which is a consequence of the Doob
inequality:
Theorem 9.1.1 Let ID
p,1
for some p > 1. Suppose that L
(, H).
Then we have
[[ > c 2 exp
_
_
_
(c E[])
2
2
2
L
(,H)
_
_
_
for any c 0.
Proof: Suppose that E[] = 0. Let (e
i
) H be a complete, orthonormal
basis of H. Dene V
n
= e
1
, . . . , e
n
and let
n
= E[P
1/n
[
n
], where
P
t
denotes the OrnsteinUhlenbeck semigroup on W. Then, from Doobs
Lemma,
n
= f
n
(e
1
, . . . , e
n
).
Note that, since f
n
p,k
W
p,k
(IR
n
,
n
), the Sobolev embedding theorem
implies that after a modication on a set of null Lebesgue measure, f
n
can
be chosen in C
(IR
n
). Let (B
t
; t [0, 1]) be an IR
n
valued Brownian motion.
Then
[
n
[ > c = P[f
n
(B
1
)[ > c
P sup
t[0,1]
[E[f
n
(B
1
)[B
t
][ > c
= P sup
t[0,1]
[Q
1t
f
n
(B
t
)[ > c ,
where P is the canonical Wiener measure on C([0, 1], IR
n
) and Q
t
is the heat
kernel associated to (B
t
), i.e.
Q
t
(x, A) = PB
t
+ x A .
Exponential Tightness 99
From the Ito formula, we have
Q
1t
f
n
(B
t
) = Q
1
f
n
(B
0
) +
_
t
0
(DQ
1s
f
n
(B
s
), dB
s
) .
By denition
Q
1
f
n
(B
0
) = Q
1
f
n
(0) =
_
f
n
(y) Q
1
(0, dy)
=
_
IR
n
f
n
(y)e
1
2
[y[
2 dy
(2)
n/2
= E
_
E[P
1/n
[V
n
]
_
= E
_
P
1/n
_
= E[]
= 0 .
Moreover we have DQ
t
f = Q
t
Df, hence
Q
1t
f
n
(B
t
) =
_
t
0
(Q
1s
Df
n
(B
s
), dB
s
) = M
n
t
.
The DoobMeyer process (M
n
, M
n
)
t
, t IR
+
) of the martingale M
n
can be
controlled as
M
n
, M
n
)
t
=
_
t
0
[DQ
1s
f
n
(B
s
)[
2
ds
_
t
0
Df
n

2
C
b
ds = tf
n

2
C
b
= tf
n

L
(
n
)
t
2
L
(,H)
.
Hence from the exponential Doob inequality, we obtain
P
_
sup
t[0,1]
[Q
1t
f
n
(B
t
)[ > c
_
2 exp
_
_
c
2
2
2
L
(,H)
_
_
.
Consequently
[
n
[ > c 2 exp
_
_
c
2
2
2
L
(,H)
_
_
.
Since
n
in probability the proof is completed.
100 Moment Inequalities
Corollary 9.1.2 Under the hypothesis of the theorem, for any
<
_
2
L
(,H)
_
1
,
we have
E
_
exp [[
2
_
< .
Proof: The rst part follows from the fact that, for F 0 a.s.,
E[F] =
_
0
PF > tdt .
Remark: In the next sections we will give more precise estimate for E[exp F
2
].
In the applications, we encounter random variables F satisfying
[F(w + h) F(w)[ c[h[
H
,
almost surely, for any h in the CameronMartin space H and a xed con
stant c > 0, without any hypothesis of integrability. For example, F(w) =
sup
t[0,1]
[w(t)[, dened on C
0
[0, 1] is such a functional. In fact the above
hypothesis contains the integrability and Sobolev dierentiability of F. We
begin rst by proving that under the integrability hypothesis, such a func
tional is in the domain of :
Lemma 9.1.3 Suppose that F : W IR is a measurable random variable
in
p>1
L
p
(), satisfying
[F(w + h) F(w)[ c[h[
H
, (9.1.2)
almost surely, for any h H, where c > 0 is a xed constant. Then F
belongs to ID
p,1
for any p > 1.
Remark: If in (9.1.2) the negligeable set on which the inequality is satised
is independent of h H, then the functional F is called HLipschitz.
Proof: Since, for some p
0
> 1, F L
p
0
, the distributional derivative of F,
F exists . We have
k
F ID
t
for any k H. Moreover, for ID, from
the integration by parts formula
E[
k
F ] = E[F
k
] + E[Fk ]
=
d
dt
[
t=0
E[F (w + tk)] + E[Fk ]
=
d
dt
[
t=0
E [F(w tk) (tk)] + E [Fk ]
= lim
t0
E
_
F(w tk) F(w)
t
_
,
Exponential Tightness 101
where (k) denotes the Wick exponential of the Gaussian random variable
k, i.e.,
(k) = exp
_
k
1
2
[k[
2
_
.
Consequently,
[E[
k
F ][ c[k[
H
E[[[]
c[k[
H

q
,
for any q > 1, i.e., F belongs to L
p
(, H) for any p > 1. Let now (e
i
; i
IN) be a complete, orthonormal basis of H, denote by V
n
the sigmaeld
generated by e
1
, . . . , e
n
, n IN and let
n
be the orthogonal projection
onto the the subspace of H spanned by e
1
, . . . , e
n
. Let us dene
F
n
= E[P
1/n
F[V
n
],
where P
1/n
is the OrnsteinUhlenbeck semigroup at the instant t = 1/n.
Then F
n
k
ID
p
0
,k
and it is immediate, from the martingale convergence
theorem and from the fact that
n
tends to the identity operator of H point
wise, that
F
n
= E[e
1/n
n
P
1/n
F[V
n
] F,
in L
p
(, H), for any p > 1, as n tends to innity. Since, by construction,
(F
n
; n IN) converges also to F in L
p
0
(), F belongs to ID
p
0
,1
. Hence we can
apply the Corollary 9.1.2.
Lemma 9.1.4 Suppose that F : W IR is a measurable random variable
satisfying
[F(w + h) F(w)[ c[h[
H
,
almost surely, for any h H, where c > 0 is a xed constant. Then F
belongs to ID
p,1
for any p > 1.
Proof: Let F
n
= [F[ n, n IN. A simple calculation shows that
[F
n
(w + h) F
n
(w)[ c[h[
H
,
hence F
n
ID
p,1
for any p > 1 and [F
n
[ c almost surely from Lemma
9.1.3. We have from the ItoClark formula (cf. Theorem 6.1.4),
F
n
= E[F
n
] +
_
1
0
E[D
s
F
n
[T
s
]dW
s
.
102 Moment Inequalities
From the denition of the stochastic integral, we have
E
_
__
1
0
E[D
s
F
n
[T
s
]dW
s
_
2
_
= E
__
1
0
[E[D
s
F
n
[T
s
][
2
ds
_
E
__
1
0
[D
s
F
n
[
2
ds
_
= E[[F
n
[
2
]
c
2
.
Since F
n
converges to [F[ in probability, and the stochastic integral is bounded
in L
2
(), by taking the dierence, we see that (E[F
n
], n IN) is a sequence
of (degenerate) random variables bounded in the space of random variables
under the topology of convergence in probability, denoted by L
0
(). There
fore sup
n
E[F
n
] > c 0 as c . Hence lim
n
E[F
n
] = E[[F[] is nite.
Now we apply the dominated convergence theorem to obtain that F L
2
().
Since the distributional derivative of F is a square integrable random vari
able, F ID
2,1
. We can now apply the Lemma 9.1.3 which implies that
F ID
p,1
for any p.
Remark: Although we have used the classical Wiener space structure in the
proof, the case of the Abstract Wiener space can be reduced to this case
using the method explained in the appendix of Chapter IV.
Corollary 9.1.5 (Ferniques Lemma) For any <
1
2
, we have
E[exp w
2
W
] < ,
where w is the norm of the Wiener path w W.
Proof: It suces to remark that
[w + h w[ [h[
H
for any h H and w W.
9.2 Coupling inequalities
We begin with the following elementary lemma (cf. [71]):
Coupling Inequalities 103
Lemma 9.2.1 Let X be a Gaussian random variable with values in IR
d
.
Then for any convex function U on IR and C
1
function V : IR
d
IR, we
have the following inequality:
E[U(V (X) V (Y ))] E
_
U
_
2
(V
t
(X), Y )
IR
d
__
,
where Y is an independent copy of X and E is the expectation with respect
to the product measure.
Proof: Let X
)d
=
_
[0,/2]
(V
t
(X
), X
t
)
IR
d d
=
2
_
[0,/2]
(V
t
(X
), X
t
)
IR
d d
where d
=
d
/2
. Since U is convex, we have
U(V (X) V (Y ))
_
/2
0
U
_
2
(V
t
(X
), X
t
)
_
d
.
Moreover X
and X
t
2
(V
t
(X), Y )
__
d
= E
_
U
_
2
(V
t
(X), Y )
__
.
Now we will extend this result to the Wiener space:
Theorem 9.2.2 Suppose that ID
p,1
, for some p > 1 and U is a lower
bounded, convex function (hence lower semicontinuous) on IR. We have
E[U((w) (z))] E
_
U
_
2
I
1
((w))(z)
__
where E is taken with respect to (dw)(dz) on WW and on the classical
Wiener space, we have
I
1
((w))(z) =
_
1
0
d
dt
(w, t)dz
t
.
104 Moment Inequalities
Proof: Suppose rst that
= f(h
1
(w), . . . , h
n
(w))
with f smooth on IR
n
, h
i
H, (h
i
, h
j
) =
ij
. We have
I
1
((w))(z) = I
1
_
n
i=1
i
f(h
1
(w), . . . , h
n
(w))h
i
_
=
n
i=1
i
f(h
1
(w), . . . , h
n
(w))I
1
(h
i
)(z)
= (f
t
(X), Y )
IR
n
where X = (h
1
(w), . . . , h
n
(w)) and Y = (h
1
(z), . . . , h
n
(z)). Hence the
inequality is trivially true in this case.
For general , let (h
i
) be a complete, orthonormal basis in H,
V
n
= h
1
, . . . , h
n
and let
n
= E[P
1/n
[V
n
] ,
where P
1/n
is the OrnsteinUhlenbeck semigroup on W. We have then
E[U(
n
(w)
n
(z))] E
_
U
_
2
I
1
(
n
(w))(z)
__
.
Let
n
be the orthogonal projection from H onto span h
1
, . . . , h
n
. We have
I
1
(
n
(w))(z) = I
1
(
w
E
w
[P
1/n
[V
n
])(z)
= I
1
(E
w
[e
1/n
P
1/n
n
[V
n
])(z)
= I
1
(
n
E
w
[e
1/n
P
1/n
[V
n
])(z)
= E
z
[I
z
1
(E
w
[e
1/n
P
w
1/n
[V
n
])[
V
n
]
where
V
n
is the copy of V
n
on the second Wiener space. Then
E
_
U
_
2
I
1
(
n
(w))(z)
__
E
_
U
_
2
I
1
(E
w
[e
1/n
P
1/n
[V
n
])(z)
__
= E
_
U
_
2
e
1/n
E
w
[I
1
(P
1/n
(w))(z)[V
n
]
__
E
_
U
_
2
e
1/n
I
1
(P
1/n
(w))(z)
__
= E
_
U
_
2
e
1/n
P
w
1/n
I
1
((w))(z)
__
Coupling Inequalities 105
E
_
U
_
2
e
1/n
I
1
((w))(z)
__
= E
_
U
_
2
P
(z)
1/n
I
1
((w))(z)
__
E
_
U
_
2
I
1
((w))(z)
__
.
Now Fatous lemma completes the proof.
Let us give some consequences of this result:
Theorem 9.2.3 The following Poincare inequalities are valid:
i) E[exp( E[])] E
_
exp
_
2
8
[[
2
H
_ _
,
ii) E[[ E[][]
2
E[[[
H
].
iii) E[[ E[][
2k
]
_
2
_
2k
(2k)!
2
k
k!
E[[[
2k
H
], k IN.
Remark 9.2.4 Let us note that the result of (ii) can not be obtained with
the classical methods, such as the ItoClark representation theorem, since the
optional projection is not a continuous map in L
1
setting. Moreover, using
the Holder inequality and the Stirling formula, we deduce the following set
of inequalities:
 E[]
p
p
2

L
p
(,H)
,
for any p 1 . To compare this result with those already known, let us recall
that using rst the ItoClark formula, then the BurkholderDavisGundy
inequality combined with the convexity inequalities for the dual projections
and some duality techniques, we obtain, only for p > 1 the inequality
 E[]
p
Kp
3/2

L
p
(,H)
,
where K is some positive constant.
Proof: Replacing the function U of Theorem 9.2 by the exponential function,
we have
E[exp( E[])] E
w
E
z
[exp((w) (z))]
E
w
_
E
z
_
[exp
2
I
1
((w))(z)
__
= E
_
exp
2
8
[[
2
H
_
.
(ii) and (iii) are similar provided that we take U(x) = [x[
k
, k IN.
106 Moment Inequalities
Theorem 9.2.5 Let ID
p,2
for some p > 1 and that [[
H
L
(, H).
Then there exists some > 0 such that
E[exp [[] < .
In particular, this hypothesis is satised if 
2

op
L
(), where  
op
denotes the operator norm.
Proof: From Theorem 9.2.3 (i), we know that
E[exp [ E[][] 2E
_
exp
2
2
8
[[
2
_
.
Hence it is sucient to prove that
E
_
exp
2
[[
2
_
<
for some > 0. However Theorem 9.1.1 applies since [[ L
(, H).
The last claim is obvious since [[[
H
[
H

2

op
almost surely.
Corollary 9.2.6 Let F ID
p,1
for some p > 1 such that [F[
H
L
().
We then have
E[exp F
2
] E
_
_
1
_
1
2
4
[F[
2
H
exp
_
E[F]
2
1
2
4
[F[
2
_
_
_
, (9.2.3)
for any > 0 such that [F[
H

2
L
()
2
4
< 1.
Proof: Ley Y be an auxiliary, realvalued Gaussian random variable, living
on a separate probability space (, , P) with variance one and zero expec
tation. We have, using Theorem 9.2.3 :
E[exp F
2
] = E E
P
[exp
2FY ]
E E
P
_
exp
_
2E[F]Y +[F[
2
Y
2
2
4
__
= E
_
_
1
_
1
2
2
[F[
2
H
exp
_
E[F]
2
1
2
2
[F[
2
_
_
_
,
where E
P
denotes the expectation with respect to the probability P.
Remark: In the next section we shall obtain a better estimate then the one
given by (9.2.3).
Exponential integrability 107
9.3 LogSobolev inequality and exponential
integrability
There is a close relationship between the probability measures satisfying the
logSobolev inequality and the exponential integrability of the random vari
ables having essentially bounded Sobolev derivatives. We shall explain this in
the frame of the Wiener space: let be a probability measure on (W, B(W))
such that the operator is a closable operator on L
2
(). Assume that we
have
E
[H
(f
2
)] KE
[[f[
2
H
]
for any cylindrical f : W IR, where H
(f
2
) = f
2
(log f
2
log E
[f
2
]).
Since is a closable operator, of course this inequality extends immediately
to the extended L
2
 domain of it.
Lemma 9.3.1 Assume now that f is in the extended L
2
domain of such
that [f[
H
is essentially bounded by one. Then
E
[e
tf
] exp
_
tE
[f] +
Kt
2
4
_
, (9.3.4)
for any t IR.
Proof: Let f
n
= min([f[, n), then it is easy to see that [f
n
[
H
[f[
H
almost surely. Let t IR and dene g
n
as to be e
t
2
f
n
. Denote by (t) the
function E[e
tf
n
]. Then it follows from the above inequality that
t
t
(t) (t) log (t)
Kt
2
4
(t) . (9.3.5)
If we write (t) =
1
t
log (t), then lim
t0
(t) = E[f
n
], and (9.3.5) implies
that
t
(t) K/4, hence we have
(t) E
[f
n
] +
Kt
4
,
therefore
(t) exp
_
tE
[f
n
] +
Kt
2
4
_
. (9.3.6)
It follows from the monotone convergence theorem that E[e
tf
] < , for any
t IR. Hence the function (t) = E[e
tf
] satises also the inequality (9.3.5)
which implies the inequality (9.3.4).
Using now the inequality (9.3.4) and an auxillary Gaussian random vari
able as in Corollary 9.2.6, we can show easily:
108 Moment Inequalities
Proposition 9.3.2 Assume that f L
p
() has essentially bounded Sobolev
derivative and that this bound is equal to one. Then we have, for any > 0,
E
[e
f
2
]
1
1 K
exp
_
2E
[f]
2
1 K
_
,
provided K < 1.
9.4 An interpolation inequality
Another useful inequality for the Wiener functionals
1
is the following inter
polation inequality which helps to control the L
p
 norm of F with the help
of the L
p
norms of F and
2
F.
Theorem 9.4.1 For any p > 1, there exists a constant C
p
, such that, for
any F ID
p,2
, one has
F
p
C
p
_
F
p
+F
1/2
p

2
F
1/2
p
_
.
Theorem 9.4.1 will be proven, thanks to the Meyer inequalities, if we can
prove the following
Theorem 9.4.2 For any p > 1, we have
(I +/)
1/2
F
p
4
(1/2)
F
1/2
p
(I +/)F
1/2
p
.
Proof: Denote by G the functional (I +/)F. Then we have F = (I +/)
1
G.
Therefore it suces to show that
(I +/)
1/2
G
p
4
(1/2)
G
1/2
p
(I +/)
1
G
1/2
p
.
We have
(I +/)
1/2
G =
2
(1/2)
_
0
t
1/2
e
t
P
t
Gdt,
where P
t
denotes the semigroup of OrnsteinUhlenbeck. For any a > 0, we
can write
(I +/)
1/2
G =
2
(1/2)
__
a
0
t
1/2
e
t
P
t
Gdt +
_
a
t
1/2
e
t
P
t
Gdt
_
.
1
This result has been proven as an answer to a question posed by D. W. Stroock, cf.
also [19].
Interpolation Inequality 109
Let us denote the two terms at the right hand side of the above equality,
respectively, by I
a
and II
a
. We have
(I +/)
1/2
G
p
2
(1/2)
[I
a

p
+II
a

p
].
The rst term at the right hand side can be upper bounded as
I
a

p
_
a
0
t
1/2
G
p
dt
= 2
aG
p
.
Let g = (I +/)
1
G. Then
_
a
t
1/2
e
t
P
t
Gdt =
_
a
t
1/2
e
t
P
t
(I +/)(I +/)
1
Gdt
=
_
a
t
1/2
e
t
P
t
(I +/)gdt
=
_
a
t
1/2
d
dt
(e
t
P
t
)dt
= a
1/2
e
a
P
a
g +
1
2
_
a
t
3/2
e
t
P
t
gdt,
where the third equality follows from the integration by parts formula. There
fore
II
a

p
a
1/2
e
a
P
a
g
p
+
1
2
_
a
t
3/2
e
t
P
t
g
p
dt
a
1/2
g
p
+
1
2
_
a
t
3/2
g
p
dt
= 2a
1/2
g
p
= 2a
1/2
(I +/)
1
G
p
.
Finally we have
(I +/)
1/2
G
p
2
(1/2)
_
a
1/2
G
p
+ a
1/2
(I +/)
1
G
p
_
.
This expression attains its minimum when we take
a =
(I +/)
1
G
p
G
p
.
Combining Theorem 9.4.1 with Meyer inequalities, we have
Corollary 9.4.3 Suppose that (F
n
, n IN) converges to zero in ID
p,k
, p >
1, k Z, and that it is bounded in ID
p,k+2
. Then the convergence takes place
also in ID
p,k+1
.
110 Moment Inequlaties
9.5 Exponential integrability of the divergence
We begin with two lemmas which are of some interest:
Lemma 9.5.1 Let L
p
(), p > 1, then, for any h H, t > 0, we have
h
P
t
(x) =
e
t
1 e
2t
_
W
(e
t
x +
1 e
2t
y)h(y)(dy)
almost surely, where
h
P
t
represents (P
t
, h)
H
.
Proof: From the Mehler formula (cf. 2.6.5), we have
h
P
t
(x) =
d
d
=0
_
W
_
e
t
(x + h) +
1 e
2t
y
_
(dy)
=
d
d
=0
_
W
_
e
t
x +
1 e
2t
_
y +
e
t
1 e
2t
h
__
(dy)
=
d
d
=0
_
W
_
e
t
x +
1 e
2t
y
_
_
e
t
1 e
2t
h
_
(y)(dy)
=
_
W
_
e
t
x +
1 e
2t
y
_
e
t
1 e
2t
h(y)(dy),
where (h) denotes exp(h 1/2[h[
2
H
).
Lemma 9.5.2 Let L
p
(, H), p > 1 and for (x, y) W W, t 0,
dene
R
t
(x, y) = e
t
x + (1 e
2t
)
1/2
y
and
S
t
(x, y) = (1 e
2t
)
1/2
x e
t
y .
Then S
t
(x, y) and R
t
(x, y) are independant, identically distributed Gaussian
random variables on (WW, (dx)(dy)). Moreover the following identity
holds true:
P
t
(x) =
e
t
1 e
2t
_
W
I
1
((R
t
(x, y)))(S
t
(x, y))(dy),
where
I
1
((R
t
(x, y)))(S
t
(x, y))
denotes the rst order Wiener integral of (R
t
(x, y)) with respect to the in
dependent path S
t
(x, y) under the product measure (dx) (dy).
Divergence 111
Proof: The rst part of the lemma is a wellknown property of the Gaussian
random variables and left to the reader. In the proof of the second part, for
the typographical facility, we shall denote in the sequel by e(t) the function
(exp t)/(1 exp 2t)
1/2
. Let now be an element of ID, we have, via
duality and using Lemma 9.5.1
< P
t
, > = < , P
t
>
=
i=1
<
i
,
h
i
P
t
>
=
i
e(t)E
__
W
i
(x)(R
t
(x, y)) h
i
(y)(dy)
_
where (h
i
; i IN) W
1 e
2t
y
y
1 e
2t
x e
t
y.
We then obtain
< P
t
(), >= e(t)
_
W
(x)I
1
((R
t
(x, y))) (S
t
(x, y)) (dx)(dy),
for any ID and the lemma follows from the density of ID in all L
p
spaces.
We are now ready to prove the following
Theorem 9.5.3 Let > 1/2 and suppose that ID
2,2
(H). Then we have
E [exp ] E
_
exp
_
[(2I +/)
[
2
H
__
,
for any satisfying
1
2
_
1
()
_
IR
+
t
1
e
2t
1 e
2t
dt
_
2
,
where / denotes the OrnsteinUhlenbeck or the number operator on W.
Proof: Let = (2I +/)
__
E
_
exp [[
2
H
_
,
112 Moment Inequlaties
where we have used the identity
(I +/)
=
_
(2I +/)
_
.
We have from the resolvent identity and from the Lemma 9.5.2,
(I +/)
=
1
()
_
IR
+
t
1
e
t
P
t
dt
=
_
IR
+
W
e
t
()
1 e
2t
t
1
e
t
I
1
((R
t
(x, y)))(S
t
(x, y))(dy)dt.
Let
0
=
1
()
_
IR
+
t
1
e
2t
1 e
2t
dt
and
(dt) = 1
IR
+
(t)
1
0
()
t
1
e
2t
1 e
2t
dt.
Then, from the Holder inequality
E
_
exp
_
(I +/)
__
= E
_
exp
_
0
_
IR
+
_
W
I
1
((R
t
(x, y)))(S
t
(x, y))(dy)(dt)
__
_
IR
+
_
W
_
W
exp
0
I
1
((R
t
(x, y)))(S
t
(x, y)) (dx)(dy)(dt)
= E
_
exp
_
2
0
2
[[
2
H
__
,
which completes the proof.
In the applications, we need also to control the moments like E[exp 
2
2
]
(cf. [101]), where is an Hvalued random variable and  . 
2
denotes the
HilbertSchmidt norm. The following result gives an answer to this question:
Proposition 9.5.4 Suppose that > 1/2 and that ID
2,2
(H). Then we
have
E[exp 
2
2
] E[exp c[(I +/)
[
2
H
],
for any
c c
0
=
_
1
()
_
IR
+
t
1
e
2t
(1 e
2t
)
1/2
dt
_
2
.
In particular, for = 1 we have c 1/4.
Divergence 113
Proof: Setting = (I +/)

2
2
_
E
_
exp c[[
2
H
_
.
Let (E
i
, i IN) be a complete, orthonormal basis of H H which is the
completion of the tensor product of H with itself under the HilbertSchmidt
topology. Then

2
2
=
i
K
i
(, E
i
)
2
,
where (., .)
2
is the scalar product in H H and K
i
= (, E
i
)
2
. Let (t) be
the function
1
()
t
1
e
2t
(1 e
2t
)
1/2
and let
0
=
_
0
(t)dt. From Lemmas 9.5.1 and 9.5.2, we have
(x)
2
2
= (I +/)
(x)
2
2
=
i
K
i
(x)
_
IR
+
(t)
_
W
(I
1
(E
i
)(y), (R
t
(x, y)))
H
(dy)dt
=
_
IR
+
W
(t) (I
1
((x))(y), (R
t
(x, y)))
H
(dy)dt
_
IR
+
(t)
__
W
[I
1
((x))(y)[
2
H
(dy)
_
1/2
__
W
[(R
t
(x, y))[
2
H
(dy)
_
1/2
dt
=
_
IR
+
(t)(x)
2
(P
t
([[
2
H
))
1/2
dt,
where I
1
((x))(y) denotes the rst order Wiener integral of (x) with
respect to the independent path (or variable) y. Consequently we have the
following inequality:

2
_
IR
+
(t)
_
P
t
([[
2
H
)
_
1/2
dt.
Therefore
E[exp 
2
2
] E
_
exp
_
_
IR
+
2
0
P
t
([[
2
H
)
(t)
0
dt
__
E
_
IR
+
(t)
0
exp
_
2
0
P
t
([[
2
H
)
_
dt
E
_
IR
+
(t)
0
exp
_
2
0
[[
2
H
_
dt
= E
_
exp
2
0
[[
2
H
_
.
114 Moment Inequlaties
As an example of application of these results let us give the following theorem
of the degree theory of the Wiener maps (cf. [101]):
Corollary 9.5.5 Suppose that ID
2,2
(H), > 1/2, satises
E
_
exp a
(2I +/)
2
H
_
< ,
for some a > 0. Then for any
_
a
4c
0
and h H, we have
E
_
e
i(h+(h,)
H
)
_
= exp
1
2
[h[
2
H
,
where is dened by
= det
2
(I
H
+ ) exp
_
h
2
2
[[
2
H
_
.
In particular, if we deal with the classical Wiener space, the path dened by
T
(w) = w + (w),
is a Brownian motion under the new probability measure E[[(T
)]d, where
(T
.
Proof: This result follows from the degree theorem for Wiener maps (cf.
[94]). In fact from the Theorem 3.2 of [94] (cf. also [101]), it follows that
E[] = 1. On the other hand, from the Theorem 3.1 of the same reference,
we have
E[F T
] = E[F]E[].
Hence the proof follows.
Notes and suggested reading
The results about the exponential tightness go back till to the celebrated Lemma of X.
Fernique about the exponential integrability of the square of seminorms (cf. [49]). It is
also proven by B. Maurey in the nite dimensional case for the Lipschitz continuous maps
with the same method that we have used here (cf. [71]). A similar result in the abstract
Wiener space case has been given by S. Kusuoka under the hypothesis of Hcontinuity,
i.e., h (w+h) is continuous for any w W. We have proven the actual result without
this latter hypothesis. However, it has been proven later that the essential boundedness of
Divergence 115
the Sobolev derivative implies the existence of a version which is Hcontinuous by Enchev
and Stroock (cf. [24]). Later it has been discovered that the exponential integrability is
implied by the logarithmic Sobolev inequality (cf. [2]). The derivation of the inequality
(9.3.6) is attributed to Herbst (cf. [54]).
In any case the exponential integrability of the square of the Wiener functionals has
found one of its most important applications in the analysis of nonlinear Gaussian func
tionals. In fact in the proof of the Ramer theorem and its extensions this property plays
an important role (cf. Chapter X, [97], [98] and [101]). Corollary 9.5.5 uses some results
about the degree theory of the Wiener maps which are explained below:
Theorem 9.5.6 Assume that and r be xed strictly positive numbers such that r >
(1 +)1. Let u ID
r,2
(H) and assume that
1.
u
L
1+
(),
2.
u
(I
H
+u)
1
h L
1+
(, H) for any h H,
where
u
= det
2
(I
H
+u) exp
_
u
1
2
[u[
2
H
_
.
Then, for any F C
b
(W), we have
E [F(w +u(w))
u
] = E[
u
]E[F] .
In particular, using a homotopy argument, one can show that, if
exp
_
u +
1 +
2
u
2
2
_
L
1+
() ,
for some > 0, > 0, then E[
u
] = 1. We refer the reader to [101] for further information
about this topic.
116 Moment Inequlaties
Chapter 10
Introduction to the Theorem of
Ramer
Introduction
The Girsanov theorem tells us that if u : W H is a Wiener functional
such that
du
dt
= u(t) is an adapted process such that
E
_
exp
_
_
1
0
u(s)dW
s
1
2
_
1
0
[ u(s)[
2
ds
__
= 1,
then under the new probability Ld, where
L = exp
_
_
1
0
u(s)dW
s
1
2
_
1
0
[ u(s)[
2
ds
_
,
w w+u(w) is a Brownian motion. The theorem of Ramer studies the same
problem without hypothesis of adaptedness of the process u. This problem
has been initiated by Cameron and Martin. Their work has been extended by
Gross and others. It was Ramer [74] who gave a main impulse to the problem
by realizing that the ordinary determinant can be replaced by the modied
CarlemanFredholm determinant via dening a Gaussian divergence instead
of the ordinary Lebesgue divergence. The problem has been further studied
by Kusuoka [51] and the nal solution in the case of (locally) dierentiable
shifts in the CameronMartin space direction has been given by
Ust unel and
Zakai [97]. In this chapter we will give a partial ( however indispensable for
the proof of the general ) result.
To understand the problem, let us consider rst the nite dimensional
case: let W = IR
n
and let
n
be the standard Gauss measure on IR
n
. If
u : IR
n
IR
n
is a dierentiable mapping such that I +u is a dieomorphism
117
118 Theorem of Ramer
of IR
n
, then the theorem of Jacobi tells us that, for any smooth function F
on IR
n
, we have
_
IR
n
F(x + u(x))[ det(I + u(x))[ exp
_
< u(x), x >
1
2
[u[
2
_
n
(dx)
=
_
IR
n
F(x)
n
(dx),
where u denotes the derivative of u. The natural idea now is to pass to the
innite dimension. For this, note that, if we dene det
2
(I + u) by
det
2
(I + u(x)) = det(I + u(x)) e
trace[ u(x)]
=
i
(1 +
i
) exp
i
,
where (
i
) are the eigenvalues of u(x) counted with respect to their multi
plicity, then the density of the left hand side can be written as
= [det
2
(I + u(x))[ exp
_
< u(x), x > +trace u(x)
1
2
[u[
2
_
and let us remark that
< u(x), x > trace u(x) = u(x),
where is the adjoint of the with respect to the Gaussian measure
n
.
Hence, we can express the density as
= [det
2
(I + u(x))[ exp
_
u(x)
[u(x)[
2
2
_
.
As remarked rst by Ramer, cf. [74], this expression has two advantages:
rst det
2
(I + u), called CarlemanFredholm determinant, can be dened
for the mappings u such that u(x) is with values in the space of Hilbert
Schmidt operators rather than nuclear operators (the latter is a smaller class
than the former), secondly, as we have already seen, u is welldened for a
large class of mappings meanwhile < u(x), x > is a highly singular object in
the Wiener space.
10.1 Ramers Theorem
After these preliminaries, we can announce, using our standard notations,
the main result of this chapter:
119
Theorem 10.1.1 Suppose that u : W H is a measurable map belonging
to ID
p,1
(H) for some p > 1. Assume that there are constants c and d with
c < 1 such that for almost all w W,
u c < 1
and
u
2
d < ,
where   denotes the operator norm and  
2
denotes the HilbertSchmidt
norm for the linear operators on H. Then:
Almost surely w T(w) = w + u(w) is bijective. The inverse of T,
denoted by S is of the form S(w) = w + v(w), where v belongs to
ID
p,1
(H) for any p > 1, moreover
v
c
1 c
and v
2
d
1 c
,
almost surely.
For all bounded and measurable F, we have
E[F(w)] = E[F(T(w)) [
u
(w)[]
and in particular
E[
u
[ = 1,
where
u
= [det
2
(I +u)[ exp u
1
2
[u[
2
H
,
and det
2
(I + u) denotes the CarlemanFredholm determinant of I +
u.
The measures , T
and S
(respectively S
d
= [
u
[ ,
dT
d
= [
v
[,
where
v
is dened similarly.
120 Theorem of Ramer
Remark 10.1.2 If u 1 instead of u c < 1, then taking u
=
(1 )u we see that the hypothesis of the theorem are satised for u
. Hence
using the Fatou lemma, we obtain
E[F T [
u
[] E[F]
for any positive F C
b
(W). Consequently, if
u
,= 0 almost surely, then
T
i=1
i
(h
1
, . . . , h
n
)h
i
,
with
i
C
(IR
n
) with bounded rst derivative, h
i
W
are orthonormal
1
in H. Suppose furthermore that  c < 1 and that 
2
d as above.
Then we have
Almost surely w U(w) = w + (w) is bijective.
The measures and U
(w)[]
for all bounded and measurable F and in particular
E[[
[] = 1,
where
= [det
2
(I +)[ exp
1
2
[[
2
H
.
1
In fact h
i
W
i=1
i
(h
1
, . . . , h
n
)h
i
,
such that 
c
1c
and 
2
d
1c
.
Proof: Note rst that due to the Corollary 9.1.2 of the Chapter VIII,
E[exp [[
2
] < for any <
1
2c
. We shall construct the inverse of U
by imitating the xed point techniques: let
0
(w) = 0
n+1
(w) = (w +
n
(w)).
We have
[
n+1
(w)
n
(w)[
H
c[
n
(w)
n1
(w)[
H
c
n
[(w)[
H
.
Therefore (w) = lim
n
n
(w) exists and it is bounded by
1
1c
[(w)[
H
. By
the triangle inequality
[
n+1
(w + h)
n+1
(w)[
H
[(w + h +
n
(w + h)) (w +
n
(w))[
H
c[h[
H
+ c[
n
(w + h)
n
(w)[
H
.
Hence passing to the limit, we nd
[(w + h) (w)[
H
c
1 c
[h[
H
.
We also have
U(w + (w)) = w + (w) + (w + (w))
= w + (w) (w)
= w,
hence U (I
W
+ ) = I
W
, i.e., U is an onto map. If U(w) = U(w
t
), then
[(w) (w
t
)[
H
= [(w
t
+ (w
t
) (w)) (w
t
)[
H
c[(w) (w
t
)[
H
,
which implies that U is also injective. To show the Girsanov identity, let
us complete the sequence (h
i
, i n) to a complete orthonormal basis whose
122 Theorem of Ramer
elements are chosen from W
i=1
h
i
(w)h
i
,
where the sum converges almost surely in the norm topology of W. Let F
be a nice function on W, denote by
n
the image of the Wiener measure
under the map w
in
h
i
(w)h
i
and by the image of under w
i>n
h
i
(w)h
i
. Evidently =
n
. Therefore
E[F U [
[] =
_
IR
n
E
_
_
F
_
_
w +
in
(x
i
+
i
(x
1
. . . , x
n
))h
i
_
_
[
[
_
_
IR
n
(dx)
= E[F],
where
IR
n
(dx) denotes the standard Gaussian measure on IR
n
and the equal
ity follows from the Fubini theorem. In fact by changing the order of inte
grals, we reduce the problem to a nite dimensional one and then the result
is immediate from the theorem of Jacobi as explained above. From the con
struction of V , it is trivial to see that
(w) =
in
i
(h
1
, . . . , h
n
)h
i
,
for some vector eld (
1
, . . . ,
n
) which is a C
mapping from IR
n
into itself
due to the nite dimensional inverse mapping theorem. Now it is routine to
verify that
= (I +)
V,
hence

2
I + V 
2
(1 +) V 
2
d
_
1 +
c
1 c
_
=
d
1 c
.
Lemma 10.1.4 With the notations and hypothesis of Lemma 10.1.3, we
have
V = [[
2
H
+ trace [( V ) ] ,
almost surely.
123
Proof: We have
=
i=1
(, e
i
)
H
e
i
e
i
(, e
i
)
H
,
where the sum converges in L
2
and the result is independent of the choice
of the orthonormal basis (e
i
; i IN). Therefore we can choose as basis
h
1
, . . . , h
n
that we have already used in Lemma 10.1.3, completed with the
elements of W
i=1
(, h
i
)
H
h
i
h
i
(, h
i
)
H
.
From the Lemma 10.1.3, we have V = and since, h
i
are originating from
W
V . Consequently,
applying all this, we obtain
V =
n
1
( V, h
i
)
H
(h
i
+ (h
i
, )
H
) (
h
i
(, h
i
)
H
) V
= ( V, )
H
+ ( V ) +
n
h
i
( V, h
i
)
H
h
i
(, h
i
)
H
V
= [[
2
H
+
n
1
( V [h
i
], [h
i
])
H
= [[
2
H
+ trace( V ),
where [h] denotes the HilbertSchmidt operator applied to the vector
h H.
Remark 10.1.5 Since and are symmetric, we have U = and
consequently
U = [[
2
H
+ trace [( U) ] .
Corollary 10.1.6 For any cylindrical function F on W, we have
E[F V ] = E [F [
[] .
E[F U] = E [F [
[] .
Proof: The rst part follows from the identity
E [F [
[] = E [F V U [
[]
= E[F V ].
124 Theorem of Ramer
To see the second part, we have
E[F U] = = E
_
F U
1
[
[ V
U [
[
_
= E
_
F
1
[
[ V
_
.
From Lemma 10.1.4, it follows that
1
[
[ V
=
1
[det
2
(I +) V [
exp
_
+ 1/2[[
2
H
_
V
=
1
[det
2
(I +) V [
exp
_
1/2[[
2
H
+ trace(( V ) )
_
= [
[,
since, for general HilbertSchmidt maps A and B, we have
det
2
(I +A) det
2
(I +B) = exp trace(AB) det
2
((I +A)(I +B)) (10.1.1)
and in our case we have
(I + V ) (I +) = I .
Remark: In fact the equality (10.1.1) follows from the multiplicative prop
erty of the ordinary determinants and from the formula (cf. [23], page 1106,
Lemma 22):
det
2
(I + A) =
i=1
(1 +
i
)e
i
,
where (
i
, i IN) are the eigenvalues of A counted with respect to their
multiplicity.
Proof of Theorem 10.1.1: Let (h
i
, i IN) W
be a complete orthonor
mal basis of H. For n IN, let V
n
be the sigma algebra on W generated
by h
1
, . . . , h
n
,
n
be the orthogonal projection of H onto the subspace
spanned by h
1
, . . . , h
n
. Dene
n
= E
_
n
P
1/n
u[V
n
_
,
125
where P
1/n
is the OrnsteinUhlenbeck semigroup on W with t = 1/n. Then
n
in ID
p,1
(H) for any p > 1 (cf., Lemma 9.1.4 of Chapter IX). Moreover
n
has the following form:
n
=
n
i=1
n
i
(h
1
, . . . , h
n
)h
i
,
where
n
i
are C
n
= E
_
n
e
1/n
P
1/n
u[V
n
_
,
hence

n
 e
1/n
E
_
P
1/n
u[V
n
_
,
and the same inequality holds also with the HilbertSchmidt norm. Conse
quently, we have

n
 c , 
n

2
d ,
almost surely. Hence, each
n
satises the hypothesis of Lemma 10.1.3. Let
us denote by
n
the shift corresponding to the inverse of U
n
= I +
n
and let
V
n
= I +
n
. Denote by
n
and L
n
the densities corresponding, respectively,
to
n
and
n
, i.e., with the old notations
n
=
n
and L
n
=
n
.
We will prove that the sequences of densities
n
: n IN and L
n
: n IN
are uniformly integrable. In fact we will do this only for the rst sequence
since the proof for the second is very similar to the proof of the rst case.
To prove the uniform integrability, from the lemma of de la VallePoussin, it
suces to show
sup
n
E [[
n
[[ log
n
[] < ,
which amounts to show, from the Corollary 10.1.6, that
sup
n
E [[ log
n
V
n
[] < .
Hence we have to control
E
_
[ log det
2
(I +
n
V
n
)[ +[
n
V
n
[ + 1/2[
n
V
n
[
2
_
.
126 Theorem of Ramer
From the Lemma 10.1.4, we have
n
V
n
=
n
[
n
[
2
H
+ trace(
n
V
n
)
n
,
hence
E[[
n
V
n
[] 
n

L
2
()
+ E[[
n
[
2
] + E[
n
V
n

2

n

2
]

n

L
2
(,H)
+
n

2
L
2
(,H)
+
n

L
2
(,HH)
+
d
2
1 c

n

L
2
(,H)
+
n

2
L
2
(,H)
+
d(1 + d)
1 c
,
where the second inequality follows from

L
2
()

L
2
(,HH)
+
L
2
(,H)
.
From the Corollary 9.1.2 of Chapter IX, we have
sup
n
E
_
exp [
n
[
2
H
_
< ,
for any <
(1c)
2
2d
2
, hence
sup
n
E[[
n
[
2
] < .
We have a wellknown inequality (cf. [101], Appendix), which says that
[det
2
(I + A)[ exp
1
2
A
2
2
for any HilbertSchmidt operator A on H. Applying this inequality to our
case, we obtain
sup
n
[log det
2
(I +
n
V
n
)[
d
2
2
and this proves the uniform integrability of (
n
, n IN). Therefore the
sequence (
n
, n IN) converges to
u
in L
1
() and we have
E[F T [
u
[] = E[F],
for any F C
b
(W), where T(w) = w + u(w).
To prove the existence of the inverse transformation we begin with
[
n
m
[
H
[
n
V
n
m
V
n
[
H
+[
m
V
n
m
V
m
[
H
[
n
V
n
m
V
n
[
H
+ c [
n
m
[
H
,
127
since c < 1, we obtain:
(1 c)[
n
m
[
H
[
n
V
n
m
V
n
[
H
.
Consequently, for any K > 0,
[
n
m
[
H
> K [
n
V
n
m
V
n
[
H
> (1 c)K
= E
_
[
n
[1
[
n
m
[>(1c)K
_
0,
as n and m go to innity, by the uniform integrability of (
n
; n IN) and by
the convergence in probability of (
n
; n IN). As the sequence (
n
; n IN)
is bounded in all L
p
spaces, this result implies the existence of an Hvalued
random variable, say v which is the limit of (
n
; n IN) in probability. By
uniform integrability, the convergence takes place in L
p
(, H) for any p > 1
and since the sequence (
n
; n IN) is bounded in L
n
[1
[v
n
[
H
>a/2
_
0,
by the uniform integrability of (
n
; n IN), therefore almost surely, we
have S T(w) = w.
128 Theorem of Ramer
10.2 Applications
In the sequel we shall give two applications. The rst one consists of a very
simple case of the Ramer formula which is used in Physics litterature (cf.
[20] for more details). The second one concerns the logarithmic Sobolev
inequality for the measures T
with respect
to .
Proof: Let e
i
, i 0 be the complete, orthonormal basis of H correspond
ing to the eigenfunctions of K and denote by
i
, i 0 its eigenvalues. We
can represent K as
K =
i=0
i
e
i
e
i
and
K
2
=
i=0
2
i
e
i
e
i
.
VanVleck Formula 129
Since K =
i
i
e
i
e
i
, we have
K
2
H
=
i=0
2
i
e
2
i
=
i=0
2
i
(e
2
i
1) +
i=0
2
i
=
i=0
2
i
(e
i
.e
i
) + trace K
2
=
(2)
K
2
+ trace K
2
.
Theorem 10.2.3 Let K L
2
(H) be a symmetric HilbertSchmidt operator
such that (I + K) is invertible and let h
1
, . . . , h
n
be n linearly independent
elements of H. Denote by
(2)
_
K +
1
2
K
2
__
h = x
_
= exp
_
1
2
trace K
2
_
det
2
(I + K)
1 q
K
(x)
q
0
(x)
,
where q
0
(x) and q
K
(x) denote respectively the densities of the laws of the
Gaussian vectors (h
1
, . . . , h
n
) and
_
(I + K)
1
h
1
, . . . , (I + K)
1
h
n
_
.
Proof: By the Ramer formula (cf. Theorem 10.1.1), for any nice function f
on IR
n
, we have
E
_
f(
h)[det
2
(I + K)[ exp
_
(2)
K
1
2
K
2
H
__
= E
_
f(
h) T
1
K
(w)
_
.
Hence
_
IR
n
E
_
exp
_
(2)
_
K +
1
2
K
2
__
[
h = x
_
f(x)q
0
(x) dx
= exp
_
1
2
trace K
2
_
det
2
(I + K)
1
_
IR
n
f(x)q
K
(x)dx.
130 Theorem of Ramer
Corollary 10.2.4 Suppose that A is a symmetric HilbertSchmidt operator
whose spectrum is included in (1/2, 1/2). Let h
1
, . . . , h
n
be n linearly inde
pendent elements of H and dene the symmetric, HilbertSchmidt operator
K as K = (I + 2A)
1/2
I. Then the following identity holds:
E
_
exp(
(2)
A) [
h = x
_
=
1
_
det
2
(I + 2A)
q
K
(x)
q
0
(x)
, (10.2.2)
for any x = (x
1
, . . . , x
n
) IR
n
.
Proof: Since the spectrum of A is included in (1/2, 1/2), the operator
I + 2A is symmetric and denite. It is easy to see that the operator K
is HilbertSchmidt. We have K + K
2
/2 = A, hence the result follows by
Theorem 10.2.3.
10.2.2 Logarithmic Sobolev inequality
Recall that the logarithmic Sobolev inequality for the Wiener measure says
E
_
f
2
log
f
2
E[f
2
]
_
2E[[f[
2
H
] , (10.2.3)
for any f ID
2,1
. We can extend this inequality easily to the measures
= T
, where T = I
W
+ u satises the hypothesis of Theorem 10.1.1
Theorem 10.2.5 Assume that is a measure given by = T
, where
T = I
W
+u satises the hypothesis of Theorem 10.1.1, in particular u c
almost surely for some c (0, 1). Then, we have
E
_
f
2
log
f
2
E[f
2
]
_
2
_
c
1 c
_
2
E
[[f[
2
H
] (10.2.4)
for any cylindrical Wiener functional f, where E
_
[f[
2
H
I
H
+u S
2
_
= 2E
_
[f[
2
H
(I
H
+v)
1

2
_
2
_
c
1 c
_
2
E
_
[f[
2
H
_
and this completes the proof.
We have also the following:
Theorem 10.2.6 The operator is closable in L
p
() for any p > 1.
Proof: Assume that (f
n
, n IN) is a sequence of cylindrical Wiener func
tionals, converging in L
p
() to zero, and assume also that (f
n
, n IN) is
Cauchy in L
p
(, H), denote its limit by . Then, by denition, (f
n
T, n IN)
converges to zero in L
p
(), hence ((f
n
T), n IN) converges to zero in
ID
p,1
(H). Moreover
(f
n
T) = (I
H
+u)
f
n
T ,
hence for any cylindrical ID(H), we have
lim
n
E[((f
n
T), )
H
] = lim
n
E[(f
n
T, (I
H
+u))
H
]
= E[( T, (I
H
+u))
H
]
= 0 .
Since T is invertible, the sigma algebra generated by T is equal to the Borel
sigma algebra of W upto the negligeable sets. Consequently, we have
(I
H
+u)
T = 0
almost surely. Since I
H
+ u is almost surely invertible, almost surely
we have T = 0 and this amounts up to saying = 0 almost surely.
Notes and suggested reading
The Ramer theorem has been proved, with some stronger hypothesis (Frechet regularity
of u) in [74], later some of its hypothesis have been relaxed in [51]. The version given
here has been proved in [97]. We refer the reader to [101] for its further extensions and
applications to the degree theory of Wiener maps (cf. [98] also). The VanVleck formula
is wellknown in Physics, however the general approach that we have used here as well as
the logarithmic Sobolev inequalities with these new measures are original.
132 Theorem of Ramer
Chapter 11
Convexity on Wiener space
Introduction
On an innite dimensional vector space W the notion of convex or concave
function is wellknown. Assume now that this space is equipped with a
probability measure. Suppose that there are two measurable functions on
this vector space, say F and G such that F = G almost surely. If F is
a convex function, then from the probabilistic point of view, we would like
to say that G is also convex. However this is false; since in general the
underlying probability measure is not (quasi) invariant under the translations
by the elements of the vector space. If W contains a dense subspace H
such that w w + h (h H) induces a measure which is equivalent to
the initial measure or absolutely continuous with respect to it, then we can
dene a notion of Hconvexity or Hconcavity in the direction of H for
the equivalence classes of real random variables. Hence these notions will be
particularly useful for the probabilistic calculations.
The notion of Hconvexity has been used in [101] to study the absolute
continuity of the image of the Wiener measure under the monotone shifts.
In this chapter we study further properties of such functions and some ad
ditional ones in the frame of an abstract Wiener space, namely Hconvex,
Hconcave, log Hconcave and log Hconvex Wiener functions, where H de
notes the associated CameronMartin space. In particular we extend some
nite dimensional results of [73] and [13] to this setting and prove that some
nite dimensional convexityconcavity inequalities have their counterparts in
innite dimensions.
133
134 Convexity
11.1 Preliminaries
In the sequel (W, H, ) denotes an abstract Wiener space, i.e., H is a separa
ble Hilbert space, called the CameronMartin space. It is identied with its
continuous dual. W is a Banach or a Frechet space into which H is injected
continuously and densely. is the standard cylindrical Gaussian measure on
H which is concentrated in W as a Radon probability measure.
In the sequel we shall use the notion of second quantization of bounded
operators on H; although this is a wellknown subject, we give a brief outline
below for the readers convenience (cf. [8], [30], [77]). Assume that A : H
H is a bounded, linear operator, then it has a unique, measurable (i.e.,
measurable with respect to the completion of B(W)) extension, denoted
by
A, as a linear map on W (cf. [8, 30]). Assume in particular that A 1
and dene S = (I
H
A
A)
1/2
, T = (I
H
AA
)
1/2
and U : H H H H
as U(h, k) = (Ah+Tk, Sh+A
w +
Sy)(dy) ,
which happens to be a Markovian contraction on L
p
() for any p 1. (A)
can be calculated explicitly for the Wick exponentials as
(A) exp
_
h 1/2[h[
2
H
_
= exp
_
Ah 1/2[Ah[
2
H
_
(h H) .
This identity implies that (AB) = (A)(B) and that for any sequence
(A
n
, n IN) of operators whose norms are bounded by one, (A
n
) converges
strongly to (A) if lim
n
A
n
= A in the strong operator topology. A particular
case of interest is when we take A = e
t
I
H
, then (e
t
I
H
) equals to the
OrnsteinUhlenbeck semigroup P
t
. Also if is the orthogonal projection of
H onto a closed vector subspace K, then () is the conditional expectation
with respect to the sigma eld generated by k, k K.
11.2 Hconvexity and its properties
Let us give the notion of Hconvexity on the Wiener space W:
Denition 11.2.1 Let F : W IR be a measurable function. It is
called Hconvex if for any h, k H, [0, 1]
F(w + h + (1 )k) F(w + h) + (1 )F(w + k) (11.2.1)
Hconvexity 135
almost surely.
Remarks:
This denition is more general than the one given in [99, 101] since F
may be innite on a set of positive measure.
Note that the negligeable set on which the relation (11.2.1) fails may
depend on the choice of h, k and of .
If G : W IR is a measurable convex function, then it is
necessarily Hconvex.
To conclude the Hconvexity, it suces to verify the relation (11.2.1)
for k = h and = 1/2.
The following properties of Hconvex Wiener functionals have been proved
in [99, 100, 101]:
Theorem 11.2.2 1. If (F
n
, n IN) is a sequence of Hconvex functionals
converging in probability, then the limit is also Hconvex.
2. If F L
p
() (p > 1) is Hconvex if and only if
2
F is positive and
symmetric HilbertSchmidt operator valued distribution on W.
3. If F L
1
() is Hconvex, then P
t
F is also Hconvex for any t 0,
where P
t
is the OrnsteinUhlenbeck semigroup on W.
The following result is immediate from Theorem 11.2.2 :
Corollary 11.2.3 F
p>1
L
p
() is Hconvex if and only if
E
_
2
F(w), h h
_
2
_
0
for any h H and ID
+
, where ( , )
2
denotes the scalar product for the
HilbertSchmidt operators on H .
We have also
Corollary 11.2.4 If F L
p
(), p > 1, is Hconvex and if E[
2
F] = 0,
then F is of the form
F = E[F] + (E[F]) .
136 Convexity
Proof: Let (P
t
, t 0) denote the OrnsteinUhlenbeck semigroup, then P
t
F is
again Hconvex and Sobolev dierentiable. Moreover
2
P
t
F = e
2t
P
t
2
F.
Hence E[
2
P
t
F] = 0, and the positivity of
2
P
t
F implies that
2
P
t
F = 0
almost surely, hence
2
F = 0. This implies that F is in the rst two Wiener
chaos.
Remark: It may be worthwhile to note that the random variable which
represents the share price of the Black and Scholes model in nancial math
ematics is Hconvex.
We shall need also the concept of cconvex functionals:
Denition 11.2.5 Let (e
i
, i IN) W
n
= w w
n
, then
a Wiener functional f : W IR is called cconvex if, for any such basis
(e
i
, i IN), for almost all w
n
, the partial map
w
n
f(w
n
+ w
n
)
has a modication which is convex on the space spane
1
, . . . , e
n
IR
n
.
Remark: It follows from Corollary 11.2.3 that, if f is Hconvex and in some
L
p
() (p > 1), then it is cconvex. We shall prove that this is also true
without any integrability hypothesis.
We begin with the following lemma whose proof is obvious:
Lemma 11.2.6 If f is cconvex then it is Hconvex.
In order to prove the validity of the converse of Lemma 11.2.6 we need some
technical results from the harmonic analysis on nite dimensional Euclidean
spaces that we shall state as separate lemmas:
Lemma 11.2.7 Let B B(IR
n
) be a set of positive Lebesgue measure. Then
B + B contains a nonempty open set.
Proof: Let (x) = 1
B
1
B
(x), where denotes the convolution of functions
with respect to the Lebesgue measure. Then is a nonnegative, continuous
function, hence the set O = x IR
n
: (x) > 0 is an open set. Since B
has positive measure, can not be identically zero, hence O is nonempty.
Besides, if x O, then the set of y IR
n
such that y B and x y B
has positive Lebesgue measure, otherwise (x) would have been null. Con
sequently O B + B.
The following lemma gives a more precise statement than Lemma 11.2.7:
Hconvexity 137
Lemma 11.2.8 Let B B(IR
n
) be a set of positive Lebesgue measure and
assume that A IR
n
IR
n
with B B = A almost surely with respect to the
Lebesgue measure of IR
n
IR
n
. Then the set x + y : (x, y) A contains
almost surely an open subset of IR
n
.
Proof: It follows from an obvious change of variables that
1
A
(y, x y) = 1
B
(y)1
B
(x y)
almost surely, hence
_
IR
n
1
A
(y, x y)dy = (x)
almost surely, where (x) = 1
B
1
B
(x). Consequently, for almost all x IR
n
such that (x) > 0, one has (y, x y) A, this means that
x IR
n
: (x) > 0 u + v : (u, v) A
almost surely.
The following lemma is particularly important for the sequel:
Lemma 11.2.9 Let f : IR
n
IR
+
be a Borel function which is nite
on a set of positive Lebesgue measure. Assume that, for any u IR
n
,
f(x)
1
2
[f(x + u) + f(x u)] (11.2.2)
dxalmost surely (the negligeable set on which the inequality (11.2.2) fails
may depend on u). Then there exists a nonempty, open convex subset U of
IR
n
such that f is locally essentially bounded on U. Moreover let D be the
set consisting of x IR
n
such that any neighbourhood of x D contains a
Borel set of positive Lebesgue measure on which f is nite, then D U, in
particular f = almost surely on the complement of U.
Proof: From the theorem of Fubini, the inequality (11.2.2) implies that
2f
_
x + y
2
_
f(x) + f(y) (11.2.3)
dxdyalmost surely. Let B B(IR
n
) be a set of positive Lebesgue measure
on which f is bounded by some constant M > 0. Then from Lemma 11.2.7,
B + B contains an open set O. Let A be the set consisting of the elements
of B B for which the inequality (11.2.3) holds. Then A = B B almost
surely, hence from Lemma 11.2.8, the set = x + y : (x, y) A contains
almost surely the open set O. Hence for almost all z
1
2
O, 2z belongs to
138 Convexity
the set , consequently z =
1
2
(x + y), with (x, y) A. This implies, from
(11.2.3), that f(z) M. Consequently f is essentially bounded on the open
set
1
2
O.
Let now U be set of points which have neighbourhoods on which f is
essentially bounded. Clearly U is open and nonempty by what we have
shown above. Let S and T be two balls of radius , on which f is bounded
by some M > 0. Assume that they are centered at the points a and b
respectively. Let u =
1
2
(b a), then for almost all x
1
2
(S + T), x + u T
and x u S, hence, from the inequality (11.2.2) f(x) M, which shows
that f is essentially bounded on the set
1
2
(S+T) and this proves the convexity
of U.
To prove the last claim, let x be any element of D and let V be any
neighbourhood of x; without loss of generality, we may assume that V is
convex. Then there exists a Borel set B V of positive measure on which
f is bounded, hence from the rst part of the proof, there exists an open
neighbourhood O B + B such that f is essentially bounded on
1
2
O
1
2
(V + V ) V , hence
1
2
O U. Consequently V U ,= , and this implies
that x is in the closure of U, i.e. D U. The fact that f = almost surely
on the complement of U is obvious from the denition of D.
Theorem 11.2.10 Let g : IR
n
IR be a measurable mapping such
that, for almost all u IR
n
,
g(u + x + y) g(u + x) + g(u + y) (11.2.4)
for any , [0, 1] with + = 1 and for any x, y IR
n
, where the
negligeable set on which the relation (11.2.4) fails may depend on the choice
of x, y and of . Then g has a modication g
t
which is a convex function.
Proof: Assume rst that g is positive, then with the notations of Lemma
11.2.9, dene g
t
= g on the open, convex set U and as g
t
= on U
c
.
From the relation (11.2.4), g
t
is a distribution on U whose second derivative
is positive, hence it is convex on U, hence it is convex on the whole space
IR
n
. Moreover we have g
t
,= g U and U has zero Lebesgue measure,
consequently g = g
t
almost surely. For general g, dene f
= e
g
( > 0),
then, from what is proven above, f
has a modication f
t
which is convex
(with the same xed open and convex set U), hence limsup
0
f
= g
t
is
also convex and g = g
t
almost surely.
Theorem 11.2.11 A Wiener functional F : W IR is Hconvex if
and only if it is cconvex.
log Hconcave 139
Proof: We have already proven the suciency. To prove the necessity, with
the notations of Denition 11.2.5, Hconvexity implies that h F(w
n
+
w
n
+ h) satises the hypothesis of Theorem 11.2.10 when h runs in any
ndimensional Euclidean subspace of H, hence the partial mapping w
n
F(w
n
+w
n
) has a modication which is convex on the vector space spanned
by e
1
, . . . , e
n
.
11.3 Log Hconcave and clog concave Wiener
functionals
Denition 11.3.1 Let F be a measurable mapping from W into IR
+
with
F > 0 > 0.
1. F is called log Hconcave, if for any h, k H, [0, 1], one has
F(w + h + (1 )k) F(w + h)
F(w + k)
1
(11.3.5)
almost surely, where the negligeable set on which the relation (11.3.5)
fails may depend on h, k and on .
2. We shall say that F is clog concave, if for any complete, orthonormal
basis (e
i
, i IN) W
n
+w
n
) is log
concave (cf. Denition 11.2.5 for the notation), up to a modication,
on spane
1
, . . . , e
n
IR
n
.
Let us remark immediately that if F = G almost surely then G is also
log Hconcave. Moreover, any limit in probability of log Hconcave random
variables is again log Hconcave. We shall prove below some less immediate
properties. Let us begin with the following observation which is a direct
consequence of Theorem 11.2.11:
Remark: F is log Hconcave if and only if log F is Hconvex (which may
be innity with a positive probability), hence if and only if F is clog concave.
Theorem 11.3.2 Suppose that (W
i
, H
i
,
i
), i = 1, 2, are two abstract Wiener
spaces. Consider (W
1
W
2
, H
1
H
1
,
1
2
) as an abstract Wiener space.
Assume that F : W
1
W
2
IR
+
is log H
1
H
2
concave. Then the map
w
2
_
W
1
F(w
1
, w
2
) d
1
(w
1
)
is log H
2
concave.
140 Convexity
Proof: If F is log H Hconcave, so is also F c (c IR
+
), hence we may
suppose without loss of generality that F is bounded. Let (e
i
, i IN) be a
complete, orthonormal basis in H
2
. It suces to prove that
E
1
[F](w
2
+ h + l) (E
1
[F](w
2
+ h))
(E
1
[F](w
2
+ l))
w +
Sy)
is log H Hconcave on W W. In fact, for any + = 1, h, k, u, v H,
one has
F(
A
w +
Sy + (A
h + Sk) + (A
u + Sv)) (11.3.6)
F(
A
w +
Sy + A
h + Sk)
F(
A
w +
Sy + A
u + Sv)
,
ddalmost surely. Let us recall that, since the image of under the
map (w, y)
A
w+
Sy is , the terms in the inequality (11.3.6) are dened
without ambiguity. Hence
(A)F(w) =
_
W
F(
A
w +
Sy)(dy)
is log Hconcave on W from Theorem 11.3.2.
Corollary 11.3.4 Let F : W IR
+
be a log Hconcave functional. Assume
that K is any closed vector subspace of H and denote by V (K) the sigma
algebra generated by k, k K. Then the conditional expectation of F
with respect to V (K), i.e., E[F[V (K)] is again log Hconcave.
Proof: The proof follows from Theorem 11.3.3 as soon as we remark that
(
K
)F = E[F[V (K)], where
K
denotes the orthogonal projection associ
ated to K.
Corollary 11.3.5 Let F be log Hconcave. If P
t
denotes the Ornstein
Uhlenbeck semigroup on W, then w P
t
F(w) is log Hconcave.
Proof: Since P
t
= (e
t
I
H
), the proof follows from Theorem 11.3.3.
Here is an important application of these results:
Theorem 11.3.6 Assume that F : W IR is an Hconvex Wiener
functional, then F has a modication F
t
which is a Borel measurable convex
function on W. Any log Hconcave functional G has a modication G
t
which
is Borel measurable and logconcave on W.
142 Convexity
Proof: Assume rst that F is positive, let G = exp F, then G is a positive,
bounded clog concave function. Dene G
n
as
G
n
= E[P
1/n
G[V
n
] ,
where V
n
is the sigma algebra generated by e
1
, . . . , e
n
, and (e
i
, i IN)
W
= e
F
, then from above, there exists a modica
tion of f
, say f
t
.
The rest is now obvious.
Under the light of Theorem 11.3.6, the following denition is natural:
Denition 11.3.7 A Wiener functional F : W IR will be called
almost surely convex if it has a modication F
t
which is convex and Borel
measurable on W. Similarly, a nonnegative functional G will be called almost
surely logconcave if it has a modication G
t
which is logconcave on W.
The following proposition summarizes the main results of this section:
Theorem 11.3.8 Assume that F : W IR is a Wiener functional
such that
F < > 0 .
Then the following are equivalent:
1. F is Hconvex,
log Hconcave 143
2. F is cconvex,
3. F is almost surely convex.
Similarly, for G : W IR
+
, with G > 0 > 0, the following properties are
equivalent:
1. G is log Hconcave,
2. G is log cconcave,
3. G is almost surely logconcave.
The notion of a convex set can be extended as
Denition 11.3.9 Any measurable subset A of W will be called Hconvex
if its indicator function 1
A
is log Hconcave.
Remark: Evidently any measurable convex subset of W is Hconvex. More
over, if A = A
t
almost surely and if A is Hconvex, then A
t
is also Hconvex.
Remark: If is an Hconvex Wiener functional, then the set
w W : (w) t
is Hconvex for any t IR.
We have the following result about the characterization of the Hconvex
sets:
Theorem 11.3.10 Assume that A is an Hconvex set, then there exists a
convex set A
t
, which is Borel measurable such that A = A
t
almost surely.
Proof: Since, by denition, 1
A
is a log Hconcave Wiener functional, from
Theorem 11.3.6, there exists a logconcave Wiener functional f
A
such that
f
A
= 1
A
almost surely. It suces to dene A
t
as the set
A
t
= w W : f
A
(w) 1 .
Example: Assume that A is an Hconvex subset of W of positive measure.
Dene p
A
as
p
A
(w) = inf ([h[
H
: h (A w) H) .
Then p
A
is Hconvex, hence almost surely convex (and HLipschitz c.f.
[101]). Moreover, the w : p
A
(w) is an Hconvex set for any IR
+
.
144 Convexity
11.4 Extensions and some applications
Denition 11.4.1 Let (e
i
, i IN) be any complete orthonormal basis of H.
We shall denote, as before, by w
n
=
n
i=1
e
i
(w) e
i
and w
n
= ww
n
. Assume
now that F : W IR is a measurable mapping with F < > 0.
1. We say that it is aconvex (a IR), if the partial map
w
n
a
2
[w
n
[
2
+ F(w
n
+ w
n
)
is almost surely convex for any n 1, where [w
n
[ is the Euclidean norm
of w
n
.
2. We call G alogconcave if
w
n
exp
_
a
2
[w
n
[
2
_
G(w
n
+ w
n
)
is almost surely logconcave for any n IN.
Remark: G is alogconcave if and only if log G is aconvex.
The following theorem gives a practical method to verify aconvexity or log
concavity:
Theorem 11.4.2 Let F : W IR be a measurable map such that
F < > 0. Dene the map F
a
on H W as
F
a
(h, w + h) =
a
2
[h[
2
H
+ F(w + h) .
Then F is aconvex if and only if, for any h, k H and , [0, 1] with
+ = 1, one has
F
a
(h + k, w + h + k) F
a
(h, w + h) + F
a
(k, w + k) (11.4.8)
almost surely, where the negligeable set on which the inequality (11.4.8)
fails may depend on the choice of h, k and of .
Similarly a measurable mapping G : W IR
+
is alogconcave if and
only if the map dened by
G
a
(h, w + h) = exp
_
a
2
[h[
2
H
_
G(w + h)
satises the inequality
G
a
(h + k, w + h + k) G
a
(h, w + h)
G
a
(k, w + k)
, (11.4.9)
almost surely, where the negligeable set on which the inequality (11.4.9)
fails may depend on the choice of h, k and of .
Extensions 145
Proof: Let us denote by h
n
its projection on the vector space spanned by
e
1
, . . . , e
n
, i.e. h
n
=
in
(h, e
i
)
H
e
i
. Then, from Theorem 11.3.8, F is
aconvex if and only if the map
h
n
a
2
_
[w
n
[
2
+ 2(w
n
, h
n
) +[h
n
[
2
_
+ F(w + h
n
)
satises a convexity inequality like (11.4.8). Besides the term [w
n
[
2
being
kept constant in this operation, it can be removed from the both sides of the
inequality. Similarly, since h
n
(w
n
, h
n
) is being ane, it also cancels from
the both sides of this inequality. Hence aconvexity is equivalent to
F
a
(h
n
+ k
n
, w + h
n
+ k
n
) F
a
(h
n
, w + h
n
) + F
a
(k
n
, w + k
n
)
where k
n
is dened as h
n
from a k H.
The second part of the theorem is obvious since G is alogconcave if and
only if log G is aconvex.
Corollary 11.4.3 1. Let
L
0
() be the space of the equivalence classes
of IR valued random variables regarded as a topological semi
group under addition and convergence in probability. Then F
L
0
()
is convex if and only if the mapping
h
2
[h[
2
H
+ F(w + h)
is a convex and continuous mapping from H into
L
0
().
2. F L
p
(), p > 1 is convex if and only if
E
__
(I
H
+
2
F)h, h
_
H
_
0
for any ID positive and h H, where
2
F is to be understood in
the sense of the distributions ID
t
.
Example: Note for instance that sin h with [h[
H
= 1, is a 1convex ran
dom variale and that exp(sin h) is 1logconcave.
The following result is a direct consequence of Prekopas theorem:
Proposition 11.4.4 Let G be an alog concave Wiener functional, a [0, 1],
and assume that V is any sigma algebra generated by the elements of the rst
Wiener chaos. Then E[G[V ] is again alogconcave.
146 Convexity
Proof: From Corollary 11.4.3, it suces to prove the case V is generated by
e
1
, . . . , e
k
, where (e
n
, n IN) is an orthonormal basis of H. Let
w
k
=
ik
e
i
(w)e
i
z
k
= w w
k
z
k,n
=
k+n
i=k+1
e
i
(w)e
i
and let z
k,n
= z
k
z
k,n
. Then we have
E[G[V ] =
_
G(z
k
+ w
k
)d(z
k
)
= lim
n
1
(2)
n/2
_
IR
n
G(z
k,n
+ z
k,n
+ w
k
)e
z
k,n

2
2
dz
k,n
.
Since
(z
k,n
, w
k
) exp
_
1
2
(a[w
k
[
2
+[z
n,k
[
2
)
_
G(z
k,n
+ z
k,n
+ w
k
)
is almost surely logconcave, the proof follows from Prekopas theorem (cf.
[73]).
The following theorem extends Theorem 11.3.3 :
Theorem 11.4.5 Let G be an alogconcave Wiener functional, where a
[0, 1). Then (A)G is alogconcave, where A L(H, H) (i.e. the space
of bounded linear operators on H) with A 1. In particular P
t
G is a
logconcave for any t 0, where (P
t
, t 0) denotes the OrnsteinUhlenbeck
semigroup on W.
Proof: Let (e
i
, i IN) be a complete, orthonormal basis of H, denote
by
n
the orthogonal projection from H onto the linear space spanned by
e
1
, . . . , e
n
and by V
n
the sigma algebra generated by e
1
, . . . , e
n
. From
Proposition 11.4.4 and from the fact that (
n
A
n
) (A) in the strong
operator topology as n tends to innity, it suces to prove the theorem when
W = IR
n
. We may then assume that G is bounded and of compact support.
Dene F as
G(x) = F(x)e
a
2
[x[
2
= F(x)
_
IR
n
e
a(x,)
d() .
Extensions 147
From the hypothesis, F is almost surely logconcave. Then, using the nota
tions explained in Section 2:
e
a
x
2
2
(A)G(x)
=
_ _
F(A
x + Sy) exp
_
a
[x[
2
2
+
a(A
x + Sy, )
_
d(y)d()
= (2)
n
_ _
F(A
x + Sy) exp
(x, y, )
2
dyd ,
where
(x, y, ) = a[x[
2
2
a(A
x + Sy, ) +[y[
2
+[[
2
= [
ax A[
2
+[
ay S[
2
+ (1 a)[y[
2
,
which is a convex function of (x, y, ). Hence the proof follows from Prekopas
theorem (cf. [73]).
The following proposition extends a wellknown nite dimensional in
equality (cf. [41]):
Proposition 11.4.6 Assume that f and g are Hconvex Wiener functionals
such that f L
p
() and g L
q
() with p > 1, p
1
= 1 q
1
. Then
E[f g] E[f]E[g] + (E[f], E[g] )
H
. (11.4.10)
Proof: Dene the smooth and convex functions f
n
and g
n
on W by
P
1/n
f = f
n
P
1/n
g = g
n
.
Using the fact that P
t
= e
t/
, where / is the number operator / =
and the commutation relation P
t
= e
t
P
t
, for any 0 t T, we have
E [P
Tt
f
n
g
n
] = E[P
T
f
n
g
n
] +
_
t
0
E [/P
Ts
f
n
g
n
] ds
= E[P
T
f
n
g
n
] +
_
t
0
e
(Ts)
E [(P
Ts
f
n
, g
n
)
H
] ds
= E[P
T
f
n
g
n
] +
_
t
0
e
(Ts)
E [(P
T
f
n
, g
n
)
H
] ds
+e
2T
_
t
0
_
s
0
e
s+
E
__
P
T
2
f
n
,
2
g
n
_
2
_
dds
E[P
T
f
n
g
n
]
+E [(P
T
f
n
, g
n
)
H
] e
T
(e
t
1) (11.4.11)
148 Convexity
where (, )
2
denotes the HilbertSchmidt scalar product and the inequality
(11.4.11) follows from the convexity of f
n
and g
n
. In fact their convexity
implies that P
t
2
f
n
and
2
g
n
are positive operators, hence their Hilbert
Schmidt tensor product is positive. Letting T = t in the above inequality we
have
E[f
n
g
n
] E [P
T
f
n
g
n
] + (1 e
T
)E [(P
T
f
n
, g
n
)
H
] . (11.4.12)
Letting T in (11.4.12), we obtain, by the ergodicity of (P
t
, t 0), the
claimed inequality for f
n
and g
n
. It suces then to take the limit of this
inequality as n tends to innity.
Proposition 11.4.7 Let G be a (positive) logconcave Wiener functional
with [0, 1]. Then the map h E[G(w+h)] is a logconcave mapping on
H. In particular, if G is symmetric, i.e., if G(w) = G(w), then
E[G(w + h)] E[G] .
Proof: Without loss of generality, we may suppose that G is bounded. Using
the usual notations, we have, for any h in any nite dimensional subspace L
of H,
E[G(w + h)] = lim
n
1
(2)
n/2
_
W
n
G(w
n
+ w
n
+ h) exp
_
[w
n
[
2
2
_
dw
n
,
from the hypothesis, the integrand is almost surely logconcave on W
n
L,
from Prekopas theorem, the integral is logconcave on L, hence the limit is
also logconcave. Since L is arbitrary, the rst part of the proof follows. To
prove the second part, let g(h) = E[G(w + h)], then, from the logconcavity
of g and symmetry of G, we have
E[G] = g(0)
= g (1/2(h) + 1/2(h))
g(h)
1/2
g(h)
1/2
= g(h)
= E[G(w + h)] .
Remark: In fact, with a little bit more attention, we can see that the map
h exp
1
2
(1 )[h[
2
H
E[G(w + h)] is logconcave on H.
We have the following immediate corollary:
logSobolev inequality 149
Corollary 11.4.8 Assume that A W is an Hconvex and symmetric set.
Then we have
(A + h) (A) ,
for any h H.
Proof: Since 1
A
is log Hconcave, the proof follows from Proposition 11.4.7.
Proposition 11.4.9 Let F L
p
() be a positive log Hconvex function.
Then for any u ID
q,2
(H), we have
E
F
_
(u E
F
[u])
2
_
E
F
_
[u[
2
H
+ 2(
u
u) + trace (u u)
_
,
where E
F
denotes the mathematical expectation with respect to the probability
dened as
F
E[F]
d.
Proof: Let F
be P
F, where (P
, IR
+
) denotes the OrnsteinUhlenbeck
semigroup. F
(, H H) where  
2
denotes the HilbertSchmidt norm. Then, for any r > 1, there exists some
t
r
> 0 such that, for any 0 t < t
r
, the image of the Wiener measure
under w w + tu(w) is equivalent to with the RadonNikodym density
L
t
L
r
(). Hence w F
is log Hconvex.
Consequently t E[F
F
[][
2
d
F
_
W
_
(I
H
+
2
F)
1
,
_
H
d
F
. (11.5.14)
In particular, if F is an Hconvex Wiener functional, then the condition
(11.5.13) is satised with = 1.
Proof: Assume rst that W = IR
n
and that F is a smooth function on
IR
n
satisfying the inequality (11.5.13) in this setting. Assume also for the
typographical facility that E[e
F
] = 1. For any smooth function function
on IR
n
, we have
_
IR
n
[ E
F
[][
2
d
F
=
1
(2)
n/2
_
IR
n
e
F(x)[x[
2
/2
[(x) E
F
[][
2
dx .
(11.5.15)
The function G(x) = F(x)+
1
2
[x[
2
is a strictly convex smooth function. Hence
BrascampLieb inequality (cf. [13]) implies that:
_
IR
n
[ E
F
[][
2
d
F
_
IR
n
_
(Hess G(x))
1
(x), (x)
_
IR
n
d
F
(x)
=
_
IR
n
_
(I
IR
n
+
2
F)
1
,
_
IR
n
d
F
.
To prove the general case we proceed by approximation as before: indeed
let (e
i
, i IN) be a complete, orthonormal basis of H, denote by V
n
the
sigma algebra generated by e
1
, . . . , e
n
. Dene F
n
as to be E[P
1/n
F[V
n
],
where P
1/n
is the OrnsteinUhlenbeck semigroup at t = 1/n. Then from the
martingale convergence theorem and the fact that V
n
is a smooth sigma al
gebra, the sequence (F
n
, n IN) converges to F in some ID
p,2
. Moreover F
n
satises the hypothesis (with a better constant in the inequality (11.5.13))
since
2
F
n
= e
2/n
E[Q
2
n
2
F[V
n
], where Q
n
denotes the orthogonal pro
jection onto the vector space spanned by e
1
, . . . , e
n
. Besides F
n
can be
represented as F
n
= (e
1
, . . . , e
n
), where is a smooth function on IR
n
satisfying
((I
IR
n
+
2
(x))y, y)
IR
n
[y[
2
IR
n ,
logSobolev inequality 151
for any x, y IR
n
. Let w
n
=
Q
n
(w) =
in
(e
i
)e
i
, W
n
=
P
n
(W) and
W
n
= (I
W
Q
n
)(W) as before. Let us denote by
n
the probability measure
corresponding to F
n
. Let us also denote by V
n
the sigma algebra generated
by e
k
, k > n. Using the nite dimensional result that we have derived,
the Fubini theorem and the inequality 2[ab[ a
2
+
1
b
2
, for any > 0, we
obtain
E
n
_
[ E
n
[][
2
_
=
_
W
n
W
n
e
F
n
(w
n
)
[(w
n
+ w
n
) E
n
[][
2
d
n
(w
n
)d
n
(w
n
)
(1 + )
_
W
e
F
n
[ E[e
F
n
[V
n
][
2
d
+
_
1 +
1
__
W
e
F
n
[E[e
F
n
[V
n
] E
n
[][
2
d
(1 + )E
n
__
(I
H
+
2
F
n
)
1
,
_
H
_
+
_
1 +
1
__
W
e
F
n
[E[e
F
n
[V
n
] E
n
[][
2
d, (11.5.16)
where F
t
n
denotes F
n
log E[e
F
n
]. Since V
n
and V
n
are independent sigma
algebras, we have
[E[e
F
n
[V
n
][ =
1
E[e
F
n
]
[E[e
F
n
[V
n
][
1
E[e
F
n
]
E[e
F
n
[V
n
]
= 
,
hence, using the triangle inequality and the dominated convergence theorem,
we realize that the last term in (11.5.16) converges to zero as n tends to
innity. Since the sequence of operator valued random variables ((I
H
+
2
F
n
)
1
, n IN) is essentially bounded in the strong operator norm, we can
pass to the limit on both sides and this gives the claimed inequality with a
factor 1 + , since > 0 is arbitrary, the proof is completed.
Remark: Let T : W W be a shift dened as T(w) = w + u(w), where
u : W H is a measurable map satisfying (u(w +h) u(w), h)
H
[h[
2
.
In [99] and in [101], Chapter 6, we have studied such transformations, called
monotone shifts. Here the hypothesis of Theorem 11.5.1 says that the shift
T = I
W
+F is monotone.
The Sobolev regularity hypothesis can be omitted if we are after a Poincare
inequality with another constant:
152 Poincare inequality
Theorem 11.5.2 Assume that F
p>1
L
p
() with E
_
e
F
_
is nite and
that, for some constant > 0,
E
__
(I
H
+
2
F)h, h
_
H
_
[h[
2
H
E[] ,
for any h H and positive test function ID, where
2
F denotes the
second order derivative in the sense of the distributions. Then we have
E
F
_
[ E
F
[][
2
_
1
F
[[[
2
H
] (11.5.17)
for any cylindrical Wiener functional . In particular, if F is Hconvex,
then we can take = 1.
Proof: Let F
t
be dened as P
t
F, where P
t
denotes the OrnsteinUhlenbeck
semigroup. Then F
t
satises the hypothesis of Theorem 11.5.1, hence we
have
E
F
t
_
[ E
F
t
[][
2
_
1
F
t
_
[[
2
H
_
for any t > 0. The claim follows when we take the limits of both sides as
t 0.
Example: Let F(w) = w +
1
2
sin(h) with [h[
H
1, where   denotes
the norm of the Banach space W. Then in general F is not in
p>1
ID
p,2
,
however the Poincare inequality (11.5.17) holds with = 1/2.
Theorem 11.5.3 Assume that F is a Wiener functional in
p>1
ID
p,2
with
E[exp F] < . Assume that there exists a constant > 0 such that
_
(I
H
+
2
F)h, h
_
H
[h[
2
H
(11.5.18)
almost surely, for any h H. Let us denote by
F
the probability measure
on (W, B(W)) dened by
d
F
= exp
_
F log E
_
e
F
__
d.
Then for any smooth cylindrical Wiener functional , we have
E
F
_
2
_
log
2
log 
2
L
2
(
F
)
__
2
F
_
[[
2
H
_
. (11.5.19)
In particular, if F is an Hconvex Wiener functional, then the condition
(11.5.18) is satised with = 1.
logSobolev inequality 153
Proof: We shall proceed as in the proof of Theorem 11.5.1. Assume then
that W = IR
n
and that F is a smooth function satisfying the inequality
(11.5.18) in this frame. In this case it is immediate to see that function
G(x) =
1
2
[x[
2
+ F(x) satises the BakryEmery condition (cf. [9], [23]),
which is known as a sucient condition for the inequality (11.5.19). For
the innite dimensional case we dene as in the proof of Theorem 11.5.1,
F
n
,
n
, V
n
, V
n
. Then, denoting by E
n
the expectation with respect to the
probability expF
t
n
d, where F
t
n
= F
n
log E[e
F
n
], we have
E
n
_
2
_
log
2
log 
2
L
2
(
F
)
__
= E
n
_
2
_
log
2
log E[e
F
2
[V
n
]
__
+E
n
_
2
_
log E[e
F
2
[V
n
] log E
n
[
2
]
__
E
n
_
[[
2
H
_
+E
n
_
2
_
log E[e
F
2
[V
n
] log E
n
[
2
]
__
, (11.5.20)
where we have used, as in the proof of Theorem 11.5.1, the nite dimensional
logSobolev inequality to obtain the inequality (11.5.20). Since in the above
inequalities everything is squared, we can assume that is positive, and
adding a constant > 0, we can also replace with
= + . Again by
the independance of V
n
and V
n
, we can pass to the limit with respect to n
in the inequality (11.5.20) for =
to obtain
E
F
_
_
log
2
log 

2
L
2
(
F
)
__
2
F
_
[
[
2
H
_
.
To complete the proof it suces to pass to the limit as 0.
The following theorem fully extends Theorem 11.5.3 and it is useful for
the applications:
Theorem 11.5.4 Assume that G is a (positive) logconcave Wiener func
tional for some [0, 1) with E[G] < . Let us denote by E
G
[ ] the
expectation with respect to the probability measure dened by
d
G
=
G
E[G]
d.
Then we have
E
G
_
2
_
log
2
log E
G
[
2
]
__
2
1
E
G
[[[
2
H
] , (11.5.21)
for any cylindrical Wiener functional .
154 Poincare inequality
Proof: Since G c, c > 0, is again logconcave, we may suppose with
out loss of generality that G is bounded. Let now (e
i
, i IN) be a com
plete, orthonormal basis for H, denote by V
n
the sigma algebra generated
by e
1
, . . . , e
n
. Dene G
n
as to be E[P
1/n
G[V
n
]. From Proposition 11.4.4
and Theorem 11.4.5, G
n
is again a logconcave, strictly positive Wiener
functional. It can be represented as
G
n
(w) = g
n
(e
1
, . . . , e
n
)
and due to the Sobolev embedding theorem, after a modication on a set
of zero Lebesgue measure, we can assume that g
n
is a smooth function on
IR
n
. Since it is strictly positive, it is of the form e
f
n
, where f
n
is a smooth,
convex function. It follows then from Theorem 11.5.3 that the inequality
(11.5.21) holds when we replace G by G
n
, then the proof follows by taking
the limits of both sides as n .
Example: Assume that A is a measurable subset of W and let H be a
measurable Wiener functional with values in IR . If G dened by
G = 1
A
H is logconcave with [0, 1), then the hypothesis of Theorem
11.5.4 are satised.
Denition 11.5.5 Let T ID
t
be a positive distribution. We say that it is
alogconcave if P
t
T is an alogconcave Wiener functional. If a = 0, then
we call T simply logconcave.
Remark: From Corollary 7.1.3, to any positive distribution on W, it corre
sponds a positive Radon measure
T
such that
< T, >=
_
W
(w)d
T
(w)
for any ID, where
represents a quasicontinuous version of .
Example: Let (w
t
, t [0, 1]) be the onedimensional Wiener process and
denote by p
(w
1
) is logconcave, where
0
denotes the Dirac measure at zero.
The following result is a Corollary of Theorem 11.5.4:
Theorem 11.5.6 Assume that T ID
t
is a positive, logconcave distribu
tion with [0, 1). Let be the probability Radon measure dened by
=
T
< T, 1 >
.
logSobolev 155
Then we have
E
2
_
log
2
log E
[
2
]
__
2
1
E
[[[
2
H
] , (11.5.22)
for any smooth cylindrical function : W IR.
Here is an application of this result:
Proposition 11.5.7 Let F be a Wiener functional in ID
r,2
for some r > 1.
Suppose that it is pnondegenerate in the sense that
_
F
[F[
2
_
L
p
() (11.5.23)
for any ID, for some p > 1. Assume furthermore that, for some x
0
IR,
(F x
0
)
2
F +F F 0 (11.5.24)
almost surely. Then we have
E
_
2
_
log
2
log E
_
2
[F = x
0
__
[F = x
0
_
2 E
_
[[
2
H
[F = x
0
_
for any smooth cylindrical .
Proof: Note that the nondegeneracy hypothesis (11.5.23) implies the exis
tence of a continuous density of the law of F with respect to the Lebesgue
measure (cf. [56] and the references there). Moreover it implies also the fact
that
lim
0
p
(F x
0
) =
x
0
(F) ,
in ID
t
, where
x
0
denotes the Dirac measure at x
0
and p
_
_
f
2
log
_
_
f
2
f
2
L
2
()
_
_
_
_
2E
_
[(I
H
+
2
/
1
F)
1
f[
2
H
_
(11.6.26)
and
E
[[f E
[f][
2
] E
_
[(I
H
+
2
/
1
F)
1
f[
2
H
_
(11.6.27)
for any smooth, cylindrical f.
Proof: Let F
n
= E[P
1/n
F[V
n
], where V
n
is the sigma algebra generated
by e
1
, . . . , e
n
and let (e
n
, n IN) be a complete, orthonormal basis of
H. Dene
n
by /
1
F
n
, then
n
is (1 a)strongly monotone (cf. [99]
or [101]) and smooth. Consequently, the shift T
n
: W W, dened by
T
n
(w) = w +
n
(w) is a bijection of W (cf. [101] Corollary 6.4.1), whose
inverse is of the form S
n
= I
W
+
n
, where
n
(w) = g
n
(e
1
, . . . , e
n
) such
that g
n
: IR
n
IR
n
is a smooth function. Moreover the images of under
T
n
and S
n
, denoted by T
n
and S
n
respectively, are equivalent to and we
have
dS
d
=
n
dT
d
= L
n
logSobolev 157
where
n
= det
2
(I
H
+
n
) exp
_
1
2
[
n
[
2
H
_
L
n
= det
2
(I
H
+
n
) exp
_
1
2
[
n
[
2
H
_
.
The hypothesis (11.6.25) implies the uniform integrability of the densities
(
n
, n 1) and (L
n
, n 1) (cf. [100, 101]). For any probability P on
(W, B(W)) and any positive, measurable function f, dene H
P
(f) as
H
P
(f) = f(log f log E
P
[f]). (11.6.28)
Using the logarithmic Sobolev inequality of L. Gross for (cf. [36]) and the
relation
(I
H
+
n
) T
n
= (I
H
+
n
)
1
,
we have
E[
n
H
n
d
(f
2
)] = E[H
(f
2
S
n
)]
2E[[(f S
n
)[
2
H
]
= 2E[[(I
H
+
n
)f S
n
[
2
H
]
= 2E[
n
[(I
H
+
n
)
1
f[
2
H
] . (11.6.29)
It follows by the aconvexity of F that
(I
H
+
n
)
1

1
1 a
almost surely for any n 1, where   denotes the operator norm. Since the
sequence (
n
, n IN) is uniformly integrable, the limit of (11.6.29) exists in
L
1
() and the proof of (11.6.26) follows. The proof of the inequality (11.6.27)
is now trivial.
Corollary 11.6.2 Assume that F satises the hypothesis of Theorem 11.6.1.
Let Z be the functional dened by
Z = det
2
(I
H
+
2
/
1
F) exp
1
2
[/
1
F[
2
H
and assume that Z, Z
1
L
(I
H
+
2
/
1
F)
1
f
2
H
_
(11.6.30)
158 Change of Variables
and
E
_
e
F
f E[e
F
f]
2
_
KE
_
e
F
(I
H
+
2
/
1
F)
1
f
2
H
_
(11.6.31)
for any smooth, cylindrical f, where K = Z
L
()
Z
1

L
()
.
Proof: Using the identity remarked by Holley and Stroock (cf. [40], p.1183)
E
P
_
H
P
(f
2
)
_
= inf
x>0
E
P
_
f
2
log
_
f
2
x
_
(f
2
x)
_
,
where P is an arbitrary probability measure, and H is dened by the relation
(11.6.28), we see that the inequality (11.6.30) follows from Theorem 11.6.1
and the inequality (11.6.31) is trivial.
Exercises
1. Assume that A
1
, . . . , A
n
are almost surely convex and symmetric sets. Prove the
following inequality:
_
n
i=1
(A
i
+h
i
)
_
_
n
i=1
A
i
_
, (11.6.32)
for any h
1
, . . . , h
n
H.
2. Assume that F is a positive, symmetric, almost surely logconcave Wiener func
tional such that F > 0 > 0. Denote by
F
the probability dened by
d
F
=
F
E[F]
d.
Prove the inequality (11.6.32) when is replaced by
F
.
3. Let A and B be two almost surely convex sets. For [0, 1], dene the map
(, w) f(, w) as
f(, w) = 1
C
(w) ,
where C
) (A)
(B)
1
.
4. Let F and G be two almost surely convex, symmetric Wiener functionals from ID
2,2
.
Prove that
E[(F, G)
H
] 0 .
5. Let W be the classical Wiener space C
0
([0, 1], IR) and let f and g be two Hconvex
functions in L
2
(). With the help of the Clarks formula, prove that
E[E[D
t
f[T
t
] E[D
t
g[T
t
]] E[D
t
f]E[D
t
g] ,
dtalmost surely.
logSobolev 159
Notes and references
The notion of convexity for the equivalence classes of Wiener random variables is a new
subject. It has been studied for the rst time in [31]. Even in the nite dimensional case
it is not evident to nd a result about the Hconvexity.
The logSobolev inequalities given here are wellknown in the nite dimensional case
except the content of the last section. The fact that logconcavity is preserved under
the action of certain semigroups and especially its implications concerning logconcave
distributions seem to be novel.
160 Change of Variables
Chapter 12
MongeKantorovitch Mass
Transportation
12.1 Introduction
In 1781, Gaspard Monge has published his celebrated memoire about the
most economical way of earthmoving [64]. The congurations of excavated
earth and remblai were modelized as two measures of equal mass, say
and , that Monge had supposed absolutely continuous with respect to the
volume measure. Later Amp`ere has studied an analogous question about the
electricity current in a media with varying conductivity. In modern language
of measure theory we can express the problem in the following terms: let
W be a Polish space on which are given two positive measures and , of
nite, equal mass. Let c(x, y) be a cost function on WW, which is, usually,
assumed positive. Does there exist a map T : W W such that T =
and T minimizes the integral
_
W
c(x, T(x))d(x)
between all such maps? The problem has been further studied by Appell
[6, 7] and by Kantorovitch [44]. Kantarovitch has succeeded to transform
this highly nonlinear problem of Monge into a linear problem by replacing
the search for T with the search of a measure on W W with marginals
and such that the integral
_
WW
c(x, y)d(x, y)
is the minimum of all the integrals
_
WW
c(x, y)d(x, y)
161
162 MongeKantorovitch
where runs in the set of measures on W W whose marginals are and
. Since then the problem adressed above is called the Monge problem and
the quest of the optimal measure is called the MongeKantorovitch problem.
In this chapter we study the MongeKantorovitch and the Monge problem
in the frame of an abstract Wiener space with a singular cost. In other words,
let W be a separable Frechet space with its Borel sigma algebra B(W) and
assume that there is a separable Hilbert space H which is injected densely
and continuously into W, hence in general the topology of H is stronger than
the topology induced by W. The cost function c : W W IR
+
is
dened as
c(x, y) = [x y[
2
H
,
we suppose that c(x, y) = if x y does not belong to H. Clearly, this
choice of the function c is not arbitrary, in fact it is closely related to Ito
Calculus, hence also to the problems originating from Physics, quantum
chemistry, large deviations, etc. Since for all the interesting measures on
W, the CameronMartin space is a negligeable set, the cost function will be
innity very frequently. Let (, ) denote the set of probability measures
on W W with given marginals and . It is a convex, compact set under
the weak topology (, C
b
(W W)). As explained above, the problem of
Monge consists of nding a measurable map T : W W, called the optimal
transport of to , i.e., T =
1
which minimizes the cost
U
_
W
[x U(x)[
2
H
d(x) ,
between all the maps U : W W such that U = . The Monge
Kantorovitch problem will consist of nding a measure on W W, which
minimizes the function J(), dened by
J() =
_
WW
[x y[
2
H
d(x, y) , (12.1.1)
where runs in (, ). Note that infJ() : (, ) is the square of
Wasserstein metric d
H
(, ) with respect to the CameronMartin space H.
Any solution of the MongeKantorovitch problem will give a solution
to the Monge problem provided that its support is included in the graph
of a map. Hence our work consists of realizing this program. Although in
the nite dimensional case this problem is wellstudied in the pathbreaking
papers of Brenier [14] and McCann [59, 60] the things do not come up easily
in our setting and the diculty is due to the fact that the cost function is
not continuous with respect to the Frechet topology of W, for instance the
1
We denote the pushforward of by T, i.e., the image of under T, by T.
Mass Transportation 163
weak convergence of the probability measures does not imply the convergence
of the integrals of the cost function. In other words the function [x y[
2
H
takes the value plus innity very often. On the other hand the results we
obtain seem to have important applications to several problems of stochastic
analysis that we shall explain while enumerating the contents of this chapter.
Section 12.3 is devoted to the derivation of some inequalities which con
trol the Wasserstein distance. In particular, with the help of the Girsanov
theorem, we give a very simple proof of an inequality, initially discovered
by Talagrand ([83]); this facility gives already an idea about the eciency
of the innite dimensional techniques for the MongeKantorovitch problem
2
.
We indicate some simple consequences of this inequality to control the mea
sures of subsets of the Wiener space with respect to second moments of
their gauge functionals dened with the CameronMartin distance. These
inequalities are quite useful in the theory of large deviations. Using a dier
ent representation of the target measure, namely by constructing a ow of
dieomorphisms of the Wiener space (cf. Chapter V of [101]) which maps
the Wiener measure to the target measure, we obtain also a new control of
the KantorovitchRubinstein metric of order one. The method we employ
for this inequality generalizes directly to a more general class of measures,
namely those for which one can dene a reasonable divergence operator.
In Section 12.4, we solve directly the original problem of Monge when the
rst measure is the Wiener measure and the second one is given with a den
sity, in such a way that the Wasserstein distance between these two measures
is nite. We prove the existence and the uniqueness of a transformation of
W of the form T = I
W
+, where is a 1convex function in the Gaussian
Sobolev space ID
2,1
such that the measure = (I
W
T) is the unique solu
tion of the problem of MongeKantorovitch. This result gives a new insight
to the question of representing an integrable, positive random variable whose
expectation is unity, as the RadonNikodym derivative of the image of the
Wiener measure under a map which is a perturbation of identity, a problem
which has been studied by X. Fernique and by one of us with M. Zakai (cf.,
[26, 27, 101]). In [101], Chapter II, it is shown that such random variables
are dense in L
1
1,+
() (the lower index 1 means that the expectations are equal
to one), here we prove that this set of random variables contains the random
variables who are at nite Wasserstein distance from the Wiener measure.
In fact even if this distance is innite, we show that there is a solution to
this problem if we enlarge W slightly by taking IN W.
Section 12.5 is devoted to the immediate implications of the existence
and the uniqueness of the solutions of MongeKantorovitch and Monge prob
2
In Section 12.7 we shall see another illustration of this phenomena.
164 MongeKantorovitch
lems constructed in Section 12.4. Indeed the uniqueness implies at once
that the absolutely continuous transformations of the Wiener space, at nite
(Wasserstein) distance, have a unique decomposition in the sense that they
can be written as the composition of a measure preserving map in the form
of the perturbation of identity with another one which is the perturbation
of identity with the Sobolev derivative of a 1convex function. This means
in particular that the class of 1convex functions is as basic as the class of
adapted processes in the setting of Wiener space.
In Section 12.6 we prove the existence and the uniqueness of solutions of
the MongeKantorovitch and Monge problems for the measures which are at
nite Wasserstein distance from each other. The fundamental hypothesis we
use is that the regular conditional probabilities which are obtained by the
disintegration of one of the measures along the orthogonals of a sequence
of regular, nite dimensional projections vanish on the sets of codimension
one. In particular, this hypothesis is satised if the measure under question
is absolutely continuous with respect to the Wiener measure. The method we
use in this section is totally dierent from the one of Section 12.4; it is based
on the notion of cyclic monotonicity of the supports of the regular conditional
probabilities obtained through some specic disintegrations of the optimal
measures. The importance of cyclic monotonicity has rst been remarked by
McCann and used abundently in [59] and in [34] for the nite dimensional
case. Here the things are much more complicated due to the singularity of the
cost function, in particular, contrary to the nite dimensional case, the cyclic
monotonicity is not compatible with the weak convergence of probability
measures. A curious reader may ask why we did not treat rst the general
case and then attack the subject of Section 12.4. The answer is twofold:
even if we had done so, we would have needed similar calculations as in
Section 12.4 in order to show the Sobolev regularity of the transport map,
hence concerning the volume, the order that we have chosen does not change
anything. Secondly, the construction used in Section 12.4 has an interest by
itself since it explains interesting relations between the transport map and
its inverse and the optimal measure in a more detectable situation, in this
sense this construction is rather complementary to the material of Section
12.6.
Section 12.7 studies the MongeAmp`ere equation for the measures which
are absolutely continuous with respect to the Wiener measure. First we
briey indicate the notion of second order Alexandro derivative and the
Alexandro version of the OrnsteinUhlenbeck operator applied to a 1convex
function in the nite dimensional case. With the help of these observations,
we write the corresponding Jacobian using the modied CarlemanFredholm
determinant which is natural in the innite dimensional case (cf., [101]).
Mass Transportation 165
Afterwards we attack the innite dimensional case by proving that the ab
solutely continuous part of the OrnsteinUhlenbeck operator applied to the
nite rank conditional expectations of the transport function is a submartin
gale which converges almost surely. Hence the only diculty lies in the calcu
lation of the limit of the CarlemanFredholm determinants. Here we have a
major diculty which originates from the pathology of the RadonNikodym
derivatives of the vector measures with respect to a scalar measure as ex
plained in [84]: in fact even if the second order Sobolev derivative of a Wiener
function is a vector measure with values in the space of HilbertSchmidt op
erators, its absolutely continuous part has no reason to be HilbertSchmidt.
Hence the CarlemanFredholm determinant may not exist, however due to
the 1convexity, the detereminants of the approximating sequence are all with
values in the interval [0, 1]. Consequently we can construct the subsolutions
with the help of the Fatou lemma.
Last but not the least, in section 12.7.1, we prove that all these diculties
can be overcome thanks to the natural renormalization of the Ito stochastic
calculus. In fact using the Ito representation theorem and the Wiener space
analysis extended to the distributions, we can give the explicit solution of
the MongeAmp`ere equation. This is a remarkable result in the sense that
such techniques do not exist in the nite dimensional case.
12.2 Preliminaries and notations
Let W be a separable Frechet space equipped with a Gaussian measure of
zero mean whose support is the whole space. The corresponding Cameron
Martin space is denoted by H. Recall that the injection H W is compact
and its adjoint is the natural injection W
L
2
(). The triple
(W, , H) is called an abstract Wiener space. Recall that W = H if and
only if W is nite dimensional. A subspace F of H is called regular if
the corresponding orthogonal projection has a continuous extension to W,
denoted again by the same letter. It is wellknown that there exists an
increasing sequence of regular subspaces (F
n
, n 1), called total, such that
n
F
n
is dense in H and in W. Let (
F
n
)
3
be the algebra generated by
F
n
, then for any f L
p
(), the martingale sequence (E[f[(
F
n
)], n 1)
converges to f (strongly if p < ) in L
p
(). Observe that the function
f
n
= E[f[(
F
n
)] can be identied with a function on the nite dimensional
abstract Wiener space (F
n
,
n
, F
n
), where
n
=
n
.
Let us recall some facts from the convex analysis. Let K be a Hilbert
space, a subset S of K K is called cyclically monotone if any nite subset
3
For the notational simplicity, in the sequel we shall denote it by
F
n
.
166 MongeKantorovitch
(x
1
, y
1
), . . . , (x
N
, y
N
) of S satises the following algebraic condition:
y
1
, x
2
x
1
) +y
2
, x
3
x
2
) + +y
N1
, x
N
x
N1
) +y
N
, x
1
x
N
) 0 ,
where , ) denotes the inner product of K. It turns out that S is cyclically
monotone if and only if
N
i=1
(y
i
, x
(i)
x
i
) 0 ,
for any permutation of 1, . . . , N and for any nite subset (x
i
, y
i
) :
i = 1, . . . , N of S. Note that S is cyclically monotone if and only if any
translate of it is cyclically monotone. By a theorem of Rockafellar, any
cyclically monotone set is contained in the graph of the subdierential of a
convex function in the sense of convex analysis ([75]) and even if the function
may not be unique its subdierential is unique.
Let now (W, , H) be an abstract Wiener space; a measurable function f :
W IR is called 1convex if the map
h f(x + h) +
1
2
[h[
2
H
= F(x, h)
is convex on the CameronMartin space H with values in L
0
(). Note that
this notion is compatible with the equivalence classes of random variables
thanks to the CameronMartin theorem. It is proven in Chapter 11 that this
denition is equivalent the following condition: Let (
n
, n 1) be a sequence
of regular, nite dimensional, orthogonal projections of H, increasing to the
identity map I
H
. Denote also by
n
its continuous extension to W and dene
n
= I
W
n
. For x W, let x
n
=
n
x and x
n
=
n
x. Then f is 1convex
if and only if
x
n
1
2
[x
n
[
2
H
+ f(x
n
+ x
n
)
is
n
almost surely convex.
12.3 Some Inequalities
Denition 12.3.1 Let and be two probabilities on (W, B(W)). We say
that a probability on (W W, B(W W)) is a solution of the Monge
Kantorovitch problem associated to the couple (, ) if the rst marginal of
is , the second one is and if
J() =
_
WW
[xy[
2
H
d(x, y) = inf
__
WW
[x y[
2
H
d(x, y) : (, )
_
,
Mass Transportation 167
where (, ) denotes the set of all the probability measures on WW whose
rst and second marginals are respectively and . We shall denote the
Wasserstein distance between and , which is the positive squareroot of
this inmum, with d
H
(, ).
Remark: Since the set of probability measures on WW is weakly compact
and since the integrand in the denition is lower semicontinuous and strictly
convex, the inmum in the denition is always attained even if the functional
J is identically innity.
The following result is an extension of an inequality due to Talagrand [83]
and it gives a sucient condition for the Wasserstein distance to be nite:
Theorem 12.3.2 Let L ILlog IL() be a positive random variable with
E[L] = 1 and let be the measure d = Ld. We then have
d
2
H
(, ) 2E[Llog L] . (12.3.2)
Proof: Without loss of generality, we may suppose that W is equipped with a
ltration of sigma algebras in such a way that it becomes a classical Wiener
space as W = C
0
(IR
+
, IR
d
). Assume rst that L is a strictly positive and
bounded random variable. We can represent it as
L = exp
_
_
0
( u
s
, dW
s
)
1
2
[u[
2
H
_
,
where u =
_
0
u
s
ds is an Hvalued, adapted random variable. Dene
n
as
n
(x) = inf
_
t IR
+
:
_
t
0
[ u
s
(x)[
2
ds > n
_
.
n
is a stopping time with respect to the canonical ltration (T
t
, t IR
+
) of
the Wiener process (W
t
, t IR
+
) and lim
n
n
= almost surely. Dene u
n
as
u
n
(t, x) =
_
t
0
1
[0,
n
(x)]
(s) u
s
(x)ds .
Let U
n
: W W be the map U
n
(x) = x+u
n
(x), then the Girsanov theorem
says that (t, x) U
n
(x)(t) = x(t) +
_
t
0
u
n
s
ds is a Wiener process under the
measure L
n
d, where L
n
= E[L[T
n
]. Therefore
E[L
n
log L
n
] = E
_
L
n
_
_
0
( u
n
s
, dW
s
)
1
2
[u
n
[
2
H
__
=
1
2
E[L
n
[u
n
[
2
H
]
=
1
2
E[L[u
n
[
2
H
] .
168 MongeKantorovitch
Dene now the measure
n
on W W as
_
WW
f(x, y)d
n
(x, y) =
_
W
f(U
n
(x), x)L
n
(x)d(x) .
Then the rst marginal of
n
is and the second one is L
n
.. Consequently
inf
__
WW
[x y[
2
H
d :
1
= ,
2
= L
n
.
_
_
W
[U
n
(x) x[
2
H
L
n
d
= 2E[L
n
log L
n
] .
Hence we obtain
d
2
H
(L
n
., ) = J(
n
) 2E[L
n
log L
n
] ,
where
n
is a solution of the MongeKantorovitch problem in (L
n
., ).
Let now be any cluster point of the sequence (
n
, n 1), since J()
is lower semicontinuous with respect to the weak topology of probability
measures, we have
J() liminf
n
J(
n
)
sup
n
2E[L
n
log L
n
]
2E[Llog L] ,
since (L., ), it follows that
d
2
H
(L., ) 2E[Llog L] .
For the general case we stop the martingale E[L[T
t
] appropriately to obtain
a bounded density L
n
, then replace it by P
1/n
L
n
to improve the positivity,
where (P
t
, t 0) denotes the OrnsteinUhlenbeck semigroup. Then, from
the Jensen inequality,
E[P
1/n
L
n
log P
1/n
L
n
] E[Llog L] ,
therefore, using the same reasoning as above
d
2
H
(L., ) liminf
n
d
2
H
(P
1/n
L
n
., )
2E[Llog L] ,
and this completes the proof.
Mass Transportation 169
Corollary 12.3.3 Assume that
i
(i = 1, 2) have RadonNikodym densities
L
i
(i = 1, 2) with respect to the Wiener measure which are in ILlog IL. Then
d
H
(
1
,
2
) < .
Proof: This is a simple consequence of the triangle inequality (cf. [10]):
d
H
(
1
,
2
) d
H
(
1
, ) + d
H
(
2
, ) .
Let us give a simple application of the above result in the lines of [58]:
Corollary 12.3.4 Assume that A B(W) is any set of positive Wiener
measure. Dene the Hgauge function of A as
q
A
(x) = inf([h[
H
: h (A x) H) .
Then we have
E[q
2
A
] 2 log
1
(A)
,
in other words
(A) exp
_
E[q
2
A
]
2
_
.
Similarly if A and B are Hseparated, i.e., if A
= x W : q
A
(x) , then
(A
c
)
1
(A)
e
2
/4
and consequently
(A) (B) exp
_
2
4
_
.
Remark: We already know that, from the 0 1law, q
A
is almost surely
nite, besides it satises [q
A
(x + h) q
A
(x)[ [h[
H
, hence E[exp q
2
A
] <
for any < 1/2 (cf. [101]). In fact all these assertions can also be proved
with the technique used below.
Proof: Let
A
be the measure dened by
d
A
=
1
(A)
1
A
d.
170 MongeKantorovitch
Let
A
be the solution of the MongeKantorovitch problem, it is easy to see
that the support of
A
is included in W A, hence
[x y[
H
inf[x z[
H
: z A = q
A
(x) ,
A
almost surely. This implies in particular that q
A
is almost surely nite.
It follows now from the inequality (12.3.2)
E[q
2
A
] 2 log (A) ,
hence the proof of the rst inequality follows. For the second let B = A
c
and
let
AB
be the solution of the MongeKantorovitch problem corresponding to
A
,
B
. Then we have from the Corollary 12.3.3,
d
2
H
(
A
,
B
) 4 log (A)(B) .
Besides the support of the measure
AB
is in AB, hence
AB
almost surely
[x y[
H
and the proof follows.
For the distance dened by
d
1
(, ) = inf
__
WW
[x y[
H
d :
1
= ,
2
=
_
we have the following control:
Theorem 12.3.5 Let L IL
1
+
() with E[L] = 1. Then we have
d
1
(L., ) E
_
(I +/)
1
L
H
_
. (12.3.3)
Proof: To prove the theorem we shall use a technique developed in [18]. Us
ing the conditioning with respect to the sigma algebra V
n
= e
1
, . . . , e
n
,
where (e
i
, i 1) is a complete, orthonormal basis of H, we reduce the prob
lem to the nite dimensional case. Moreover, we can assume that L is a
smooth, strictly positive function on IR
n
. Dene now = (I +/)
1
L and
t
(x) =
(x)
t + (1 t)L
,
for t [0, 1]. Let (
s,t
(x), s t [0, 1]) be the ow of dieomorphisms
dened by the following dierential equation:
s,t
(x) = x
_
t
s
(
s,
(x))d .
Mass Transportation 171
From the standart results (cf. [101], Chapter V), it follows that x
s,t
(x)
is Gaussian under the probability
s,t
., where
s,t
= exp
_
t
s
(
)(
s,
(x))d
is the RadonNikodym density of
1
s,t
with respect to . Dene
H
s
(t, x) =
s,t
(x) t + (1 t)L
s,t
(x) .
It is easy to see that
d
dt
H
s
(t, x) = 0
for t (s, 1). Hence the map t H
s
(t, x) is a constant, this implies that
s,1
(x) = s + (1 s)L(x) .
We have, as in the proof of Theorem 12.3.2,
d
1
(L., ) E[[
0,1
(x) x[
H
0,1
]
E
_
0,1
_
1
0
[
t
(
0,t
(x))[
H
dt
_
= E
__
1
0
t
(
0,t
1
0,1
)(x)
H
dt
_
= E
__
1
0
t
(
1
t,1
(x))
H
dt
_
= E
__
1
0
[
t
(x)[
H
t,1
dt
_
= E[[[
H
] ,
and the general case follows via the usual approximation procedure.
12.4 Construction of the transport map
In this section we give the construction of the transport map in the Gaussian
case. We begin with the following lemma:
Lemma 12.4.1 Let (W, , H) be an abstract Wiener space, assume that f :
W IR is a measurable function such that it is Gateaux dierentiable in the
direction of the CameronMartin space H, i.e., there exists some f : W
H such that
f(x + h) = f(x) +
_
1
0
(f(x + h), h)
H
d ,
172 MongeKantorovitch
almost surely, for any h H. If [f[
H
L
2
(), then f belongs to the
Sobolev space ID
2,1
.
Proof: Since [[f[[
H
[f[
H
, we can assume that f is positive. Moreover,
for any n IN, the function f
n
= min(f, n) has also a Gateaux derivative
such that [f
n
[
H
[f[
H
almost surely. It follows from the Poincare
inequality that the sequence (f
n
E[f
n
], n 1) is bounded in L
2
(), hence
it is also bounded in L
0
(). Since f is almost surely nite, the sequence
(f
n
, n 1) is bounded in L
0
(), consequently the deterministic sequence
(E[f
n
], n 1) is also bounded in L
0
(). This means that sup
n
E[f
n
] < ,
hence the monotone convergence theorem implies that E[f] < and the
proof is completed.
Theorem 12.4.2 Let be the measure d = Ld, where L is a positive
random variable, with E[L] = 1. Assume that d
H
(, ) < (for instance
L ILlog IL). Then there exists a 1convex function ID
2,1
, unique upto
a constant, such that the map T = I
W
+ is the unique solution of the
original problem of Monge. Moreover, its graph supports the unique solution
of the MongeKantorovitch problem . Consequently
(I
W
T) =
In particular T maps to and T is almost surely invertible, i.e., there
exists some T
1
such that T
1
= and that
1 =
_
x : T
1
T(x) = x
_
=
_
y W : T T
1
(y) = y
_
.
Proof: Let (
n
, n 1) be a sequence of regular, nite dimensional orthogonal
projections of H increasing to I
H
. Denote their continuous extensions to W
by the same letters. For x W, we dene
n
x =: x
n
= x
n
x. Let
n
be
the measure
n
. Since is absolutely continuous with respect to ,
n
is
absolutely continuous with respect to
n
:=
n
and
d
n
d
n
n
= E[L[V
n
] =: L
n
,
where V
n
is the sigma algebra (
n
) and the conditional expectation is taken
with respect to . On the space H
n
, the MongeKantorovitch problem, which
consists of nding the probability measure which realizes the following in
mum
d
2
H
(
n
,
n
) = inf J() : M
1
(H
n
H
n
) , p
1
=
n
, p
2
=
n
n
(x) +
n
(y) (x, y)
H
for all x, y H
n
and that
n
(x) +
n
(y) = (x, y)
H
n
almost everywhere. Hence the support of
n
is included in the graph of
the derivative
n
of
n
, hence
n
n
=
n
and the inverse of
n
is equal
to
n
. Let
n
(x) =
n
(x)
1
2
[x[
2
H
n
(y) =
n
(y)
1
2
[y[
2
H
.
Then
n
and
n
are 1convex functions and they satisfy the following rela
tions:
n
(x) +
n
(y) +
1
2
[x y[
2
H
0 , (12.4.4)
for all x, y H
n
and
n
(x) +
n
(y) +
1
2
[x y[
2
H
= 0 , (12.4.5)
n
almost everywhere. From what we have said above, it follows that
n

almost surely y = x +
n
(x), consequently
J(
n
) = E[[
n
[
2
H
] . (12.4.6)
Let q
n
: W W H
n
H
n
be dened as q
n
(x, y) = (
n
x,
n
y). If is any
solution of the MongeKantorovitch problem, then q
n
(
n
,
n
), hence
J(
n
) J(q
n
) J() = d
2
H
(, ) . (12.4.7)
174 MongeKantorovitch
Combining the relation (12.4.6) with the inequality (12.4.7), we obtain the
following bound
sup
n
J(
n
) = sup
n
d
2
H
(
n
,
n
)
= sup
n
E[[
n
[
2
H
]
d
2
H
(, ) = J() . (12.4.8)
For m n, q
m
n
(
m
,
m
), hence we should have
J(
m
) =
_
WW
[
m
x
m
y[
2
H
d
m
(x, y)
_
WW
[
m
x
m
y[
2
H
d
n
(x, y)
_
WW
[
n
x
n
y[
2
H
d
n
(x, y)
=
_
WW
[x y[
2
H
d
n
(x, y)
= J(
n
) ,
where the third equality follows from the fact that we have denoted the
n
on H
n
H
n
and its image in W W by the same letter. Let now be
a weak cluster point of the sequence of measures (
n
, n 1), where the
word weak
4
refers to the weak convergence of measures on W W. Since
(x, y) [x y[
H
is lower semicontinuous, we have
J() =
_
WW
[x y[
2
H
d(x, y)
liminf
n
_
WW
[x y[
2
H
d
n
(x, y)
= liminf
n
J(
n
)
sup
n
J(
n
)
J() = d
2
H
(, ) ,
from the relation (12.4.8). Consequently
J() = lim
n
J(
n
) . (12.4.9)
Again from (12.4.8), if we replace
n
with
n
E[
n
] and
n
with
n
+E[
n
]
we obtain a bounded sequence (
n
, n 1) in ID
2,1
, in particular it is bounded
4
To prevent the reader against the trivial errors let us emphasize that
n
is not the
projection of on W
n
W
n
.
Mass Transportation 175
in the space L
2
() if we inject it into latter by
n
(x)
n
(x)1(y). Consider
now the sequence of the positive, lower semicontinuous functions (F
n
, n 1)
dened on W W as
F
n
(x, y) =
n
(x) +
n
(y) +
1
2
[x y[
2
H
.
We have, from the relation (12.4.5)
_
WW
F
n
(x, y)d(x, y) =
_
W
n
d +
_
W
n
(y)d +
1
2
J()
=
1
2
(J() J(
n
)) 0 .
Consequently the sequence (F
n
, n 1) converges to zero in L
1
(), therefore it
is uniformly integrable. Since (
n
, n 1) is uniformly integrable as explained
above and since [x y[
2
has a nite expectation with respect to , it follows
that (
n
, n 1) is also uniformly integrable in L
1
() hence also in L
1
().
Let
t
be a weak cluster point of (
n
, n 1), then there exists a sequence
(
t
n
, n 1) whose elements are the convex combinations of some elements of
(
k
, k n) such that (
t
n
, n 1) converges in the norm topology of ID
2,1
and
almost everywhere. Therefore the sequence (
t
n
, n 1), constructed from
(
k
, k n), converges in L
1
() and almost surely. Dene and as
(x) = limsup
n
t
n
(x)
(y) = limsup
n
t
n
(y) ,
hence we have
G(x, y) = (x) + (y) +
1
2
[x y[
2
H
0
for all (x, y) WW, also the equality holds almost everywhere. Let now
h be any element of H, since x y is in H for almost all (x, y) W W,
we have
[x + h y[
2
H
= [x y[
2
H
+[h[
2
H
+ 2(h, x y)
H
almost surely. Consequently
(x + h) (x) (h, x y)
H
1
2
[h[
2
H
almost surely and this implies that
y = x +(x)
176 MongeKantorovitch
almost everywhere. Dene now the map T : W W as T(x) = x+(x),
then
_
WW
f(x, y)d(x, y) =
_
WW
f(x, T(x))d(x, y)
=
_
W
f(x, T(x))d(x) ,
for any f C
b
(W W), consequently (I
W
T) = , in particular T = .
Let us notice that any weak cluster point of (
n
, n 1), say
, satises
(x) = y x
almost surely, hence almost surely we have
= . This implies that
(
n
, n 1) has a unique cluster point , consequently the sequence (
n
, n
1) converges weakly in ID
2,1
to . Besides we have
lim
n
_
W
[
n
[
2
H
d = lim
n
J(
n
)
= J()
=
_
WW
[x y[
2
H
d(x, y)
=
_
W
[[
2
H
d,
hence (
n
, n 1) converges to in the norm topology of ID
2,1
. Let us
recapitulate what we have done till here: we have taken an arbitrary optimal
(, ) and an arbitrary cluster point of (
n
, n 1) and we have
proved that is carried by the graph of T = I
W
+ . This implies that
and are unique and that the sequence (
n
, n 1) has a unique cluster
point .
Certainly (
n
, 1) converges also in the norm topology of L
1
(). More
over, from the nite dimensional situation, we have
n
(x) +
n
(y) = 0
n
almost everywhere. Hence
E
[[
n
[
2
H
] = E[[
n
[
2
H
]
this implies the boundedness of (
n
, n 1) in L
2
(, H) (i.e., Hvalued
functions). To complete the proof we have to show that, for some measurable,
Hvalued map, say , it holds that x = y + (y) almost surely. For this
let F be a nite dimensional, regular subspace of H and denote by
F
the
projection operator onto F which is continuously extended to W, put
F
=
I
W
F
. We have W = F F
, with F
= ker
F
=
F
(W). Dene the
measures
F
=
F
() and
F
=
F
(). From the construction of , we know
Mass Transportation 177
that, for any v F
F
almost all v F
, one has
(A[
F
= v) > 0 .
It then follows from Lemma 3.4 of Chapter 11, and from the fact that the
regular conditional probability ( [
F
= v) is absolutely continuous with
respect to the Lebesgue measure of F, that u (u + v) is ( [
F
= v)
almost everywhere dierentiable on F for
F
almost all v F
. It then
follows that, almost surely, is dierentiable in the directions of F, i.e.,
there exists
F
F almost surely. Since we also have
(y + k) (y) (x y, k)
H
1
2
[k[
2
H
,
we obtain, almost surely
(
F
(y), k)
H
= (x y, k)
H
,
for any k F. Consequently
F
(y) =
F
(x y)
almost surely. Let now (F
n
, n 1) be a total, increasing sequence of
regular subspaces of H, we have a sequence (
n
, n 1) bounded in L
2
()
hence also bounded in L
2
(). Besides
n
(y) =
n
x
n
y almost surely.
Since (
n
(x y), n 1) converges in L
2
(, H), (
n
, n 1) converges in
the norm topology of L
2
(, H). Let us denote this limit by , then we have
x = y +(y) almost surely. Note that, since
n
=
n
, we can even write
in a weak sense that = . If we dene T
1
(y) = y + (y), we see that
1 = (x, y) W W : T T
1
(y) = y
= (x, y) W W : T
1
T(x) = x ,
and this completes the proof of the theorem.
Remark 12.4.3 Assume that the operator is closable with respect to ,
then we have = . In particular, if and are equivalent, then we have
T
1
= I
W
+ ,
where is is a 1convex function.
178 MongeKantorovitch
Remark 12.4.4 Assume that L IL
1
+
(), with E[L] = 1 and let (D
k
, k
IN) be a measurable partition of W such that on each D
k
, L is bounded.
Dene d = Ld and
k
= ([D
k
). It follows from Theorem 12.3.2, that
d
H
(,
k
) < . Let then T
k
be the map constructed in Theorem 12.4.2
satisfying T
k
=
k
. Dene n(dk) as the probability distribution on IN given
by n(k) = (D
k
), k IN. Then we have
_
W
f(y)d(y) =
_
WIN
f(T
k
(x))(dx)n(dk) .
A similar result is given in [27], the dierence with that of above lies in the
fact that we have a more precise information about the probability space on
which T is dened.
12.5 Polar factorization of the absolutely con
tinuous transformations of the Wiener
space
Assume that V = I
W
+v : W W be an absolutely continuous transforma
tion and let L IL
1
+
() be the RadonNikodym derivative of V with respect
to . Let T = I
W
+ be the transport map such that T = L.. Then it
is easy to see that the map s = T
1
V is a rotation, i.e., s = (cf. [101])
and it can be represented as s = I
W
+ . In particular we have
+ s = v . (12.5.10)
Since is a 1convex map, we have h
1
2
[h[
2
H
+ (x + h) is almost surely
convex (cf. Chapter11). Let s
t
= I
W
+
t
be another rotation with
t
: W
H. By the 1convexity of , we have
1
2
[
t
[
2
H
+ s
t
1
2
[[
2
H
+ s +
_
+ s,
t
_
H
,
almost surely. Taking the expectation of both sides, using the fact that s
and s
t
preserve the Wiener measure and the identity (12.5.10), we obtain
E
_
1
2
[[
2
H
(v, )
H
_
E
_
1
2
[
t
[
2
H
(v,
t
)
H
_
.
Hence we have proven the existence part of the following
Proposition 12.5.1 Let 1
2
denote the subset of L
2
(, H) whose elements
are dened by the property that x x +(x) is a rotation, i.e., it preserves
Mass Transportation 179
the Wiener measure. Then is the unique element of 1
2
which minimizes
the functional
M
v
() = E
_
1
2
[[
2
H
(v, )
H
_
.
Proof: To show the uniqueness, assume that 1
2
be another map mini
mizing J
v
. Let be the measure on W W, dened as
_
WW
f(x, y)d(x, y) =
_
W
f(x + (x), V (x))d.
Then the rst marginal of is and the second marginal is L.. Since
= (I
W
T) is the unique solution of the MongeKantorovitch problem,
we should have
_
[x y[
2
H
d(x, y) >
_
[x y[
2
H
d(x, y) = E[[[
2
H
] .
However we have
_
WW
[x y[
2
H
d(x, y) = E
_
[v [
2
H
_
= E
_
[v[
2
H
_
+ 2M
v
()
= E
_
[v[
2
H
_
+ 2M
v
()
= E
_
[v [
2
H
_
= E
_
[ s[
2
H
_
= E
_
[[
2
H
_
=
_
WW
[x y[
2
H
d(x, y)
= J()
and this gives a contradiction to the uniqueness of .
The following theorem, whose proof is rather easy, gives a better un
derstanding of the structure of absolutely continuous transformations of the
Wiener measure:
Theorem 12.5.2 Assume that U : W W be a measurable map and L
ILlog IL a positive random variable with E[L] = 1. Assume that the measure
= L is a Girsanov measure for U, i.e., that one has
E[f U L] = E[f] ,
for any f C
b
(W). Then there exists a unique map T = I
W
+ with
ID
2,1
is 1convex, and a measure preserving transformation R : W W
such that U T = R almost surely and U = R T
1
almost surely.
180 MongeKantorovitch
Proof: By Theorem 12.4.2 there is a unique map T = I
W
+ , with
ID
2,1
, 1convex such that T transports to . Since U = , we have
E[f U L] = E[f U T]
= E[f] .
Therefore x U T(x) preserves the measure . The rest is obvious since
T
1
exists almost surely.
Another version of Theorem 12.5.2 can be announced as follows:
Theorem 12.5.3 Assume that Z : W W is a measurable map such that
Z , with d
H
(Z, ) < . Then Z can be decomposed as
Z = T s ,
where T is the unique transport map of the MongeKantorovitch problem for
(, Z) and s is a rotation.
Proof: Let L be the RadonNikodym derivative of Z with respect to . We
have, from Theorem 12.4.2,
E[f] = E[f T
1
T]
= E[f T
1
L]
= E[f T
1
Z] ,
for any f C
b
(W). Hence T
1
Z = s is a rotation. Since T is uniquely
dened, s is also uniquely dened.
Although the following result is a translation of the results of this section,
it is interesting from the point of view of stochastic dierential equations:
Theorem 12.5.4 Let (W, , H) be the standard Wiener space on IR
d
, i.e.,
W = C(IR
+
, IR
d
). Assume that there exists a probability P which is the
weak solution of the stochastic dierential equation
dy
t
= dW
t
+ b(t, y)dt ,
such that d
H
(P, ) < . Then there exists a process (T
t
, t IR
+
) which is a
pathwise solution of some stochastic dierential equation whose law is equal
to P.
Mass Transportation 181
Proof: Let T be the transport map constructed in Theorem 12.4.2 corre
sponding to dP/d. Then it has an inverse T
1
such that T
1
T(x) =
x = 1. Let be the 1convex function such that T = I
W
+ and denote by
(D
s
, s IR
+
) the representation of in L
2
(IR
+
, ds). Dene T
t
(x) as the
trajectory T(x) evaluated at t IR
+
. Then it is easy to see that (T
t
, t IR
+
)
saties the stochastic dierential equation
T
t
(x) = W
t
(x) +
_
t
0
l(s, T(x))ds , t IR
+
,
where W
t
(x) = x(t) and l(s, x) = D
s
T
1
(x).
12.6 Construction and uniqueness of the trans
port map in the general case
In this section we call optimal every probability measure
5
on W W such
that J() < and that J() J() for every other probability having the
same marginals as those of . We recall that a nite dimensional subspace F
of W is called regular if the corresponding projection is continuous. Similarly
a nite dimensional projection of H is called regular if it has a continuous
extension to W.
We begin with the following lemma which answers all kind of questions
of measurability that we may encounter in the sequel:
Lemma 12.6.1 Consider two uncountable Polish spaces X and T. Let t
t
be a Borel family of probabilities on X and let T be a separable sub
algebra of the Borel algebra B of X. Then there exists a Borel kernel
N
t
f(x) =
_
X
f(y)N
t
(x, dy) ,
such that, for any bounded Borel function f on X, the following properties
hold true:
i) (t, x) N
t
f(x) is Borel measurable on T X.
ii) For any t T, N
t
f is an Tmeasurable version of the conditional
expectation E
t
[f[T].
5
In fact the results of this section are essentially true for the bounded, positive measures.
182 MongeKantorovitch
Proof: Assume rst that T is nite, hence it is generated by a nite partition
A
1
, . . . , A
k
. In this case it suces to take
N
t
f(x) =
k
i=1
1
t
(A
i
)
__
A
i
fd
t
_
1
A
i
(x)
_
with 0 =
0
0
_
.
For the general case, take an increasing sequence (T
n
, n 1) of nite sub
algebras whose union generates T. Without loss of generality we can assume
that (X, B) is the Cantor set (Kuratowski Theorem, cf., [21]). Then for every
clopen set (i.e., a set which is closed and open at the same time) G and any
t T, the sequence (N
n
t
1
G
, n 1) converges
t
almost everywhere. Dene
H
G
(t, x) = limsup
m,n
[N
n
t
1
G
(x) N
m
t
1
G
(x)[ .
H
G
is a Borel function on T X which vanishes
t
almost all x X, besides,
for any t T, x H
G
(t, x) is Tmeasurable. As there exist only countably
many clopen sets in X, the function
H(t, x) = sup
G
H
G
(t, x)
inherits all the measurability properties. Let be any probability on X, for
any clopen G, dene
N
t
1
G
(x) = lim
n
N
n
t
1
G
(x) if H(t, x) = 0 ,
= (G) if H(t, x) > 0 .
Hence, for any t T, we get an additive measure on the Boolean algebra of
clopen sets of X. Since such a measure is additive and extends uniquely
as a additive measure on B, the proof is completed.
Remark 12.6.2 1. This result holds in fact for the Lusin spaces since
they are Borel isomorphic to the Cantor set. Besides it extends easily
to countable spaces.
2. The particular case where T = /
1
(X), i.e., the space of probability
measures on X under the weak topology and t
t
being the identity
map, is particularly important for the sequel. In this case we obtain
a kernel N such that (x, ) N
f is an
Tmeasurable version of E
[f[T].
Mass Transportation 183
Lemma 12.6.3 Let and be two probability measures on W such that
d
H
(, ) <
and let (, ) be an optimal measure, i.e., J() = d
2
H
(, ), where J is
given by (12.1.1). Assume that F is a regular nite dimensional subspace of
W with the corresponding projection
F
from W to F and let
F
= I
W
F
. Dene p
F
as the projection from W W onto F with p
F
(x, y) =
F
x and
let p
F
(x, y) =
F
x. Consider the Borel disintegration
() =
_
F
W
( [x
(dz
)
=
_
F
( [x
(dx
)
along the projection of WW on F
, where
is the measure
F
, ( [x
)
denotes the regular conditional probability ( [p
F
= x
) and
is the mea
sure p
F
. Then,
and
almost surely ( [x
) is optimal on (x
+F)W.
Proof: Let p
1
, p
2
be the projections of W W dened as p
1
(x, y) =
F
(x)
and p
2
(x, y) =
F
(y). Note rst the following obvious identity:
p
1
( [x
) = ( [x
) ,
and
/
1
(F F) and C as
B = (x
, ) : (p
1
( [x
), p
2
( [x
))
C = (x
) ,
where /
1
(F F) denotes the set of probability measures on F F. Let
K be the projection of C on F
from K to /
1
(F F) such that,
almost surely, (x
,
x
) C. Dene
() =
_
K
x
()d
(x
) +
_
K
c
( [x
)d
(x
) .
Then (, ) and we have
J() =
_
K
J(
x
)d
(x
) +
_
K
c
J(( [x
))d
(x
)
<
_
K
J(( [x
))d
(x
) +
_
K
c
J(( [x
))d
(x
)
= J() ,
184 MongeKantorovitch
hence we obtain J() < J() which is a contradiction to the optimality of .
Lemma 12.6.4 Assume that the hypothesis of Lemma 12.6.3 holds and let F
be any regular nite dimensional subspace of W. Denote by
F
the projection
operator associated to it and let
F
= I
W
F
. If
F
almost surely, the
regular conditional probability ( [
F
= x
+F
whose Hausdor dimension are at most equal to dim(F)1, then there exists
a map T
F
: F F
F such that
__
(x, y) W W :
F
y = T
F
(
F
x,
F
x)
__
= 1 .
Proof: Let C
x
be the support of the regular conditional probability ( [x
)
in (x
)
is optimal in (
1
( [x
),
2
( [x
)) < for
almost
everywhere x
+ u, v) C
x
), i.e., ( [x
+ F of co
dimension one, the subdierential under question, denoted as U
F
(u, x
) is
( [x
)
__
(u, v) C
x
:
F
v = U
F
(u, x
)
__
= 1 ,
,u
= v W : (u, v) K
x
.
Then K
x
,u
consists of a single point for almost all u with respect to ( [x
).
Let
N =
_
(u, x
) F F
: Card(K
x
,u
) > 1
_
,
note that N is a Souslin set, hence it is universally measurable. Let be
the measure which is dened as the image of under the projection x
(
F
x,
F
x). We then have
(N) =
_
F
(dx
)
_
F
1
N
(u, x
)(du[x
)
= 0 .
Mass Transportation 185
Hence (u, x
) K
x
,u
= y is and almost surely welldened and it
suces to denote this map by T
F
to achive the proof.
Theorem 12.6.5 Suppose that and are two probability measures on W
such that
d
H
(, ) < .
Let (
n
, n 1) be a total increasing sequence of regular projections (of H,
converging to the identity map of H). Suppose that, for any n 1, the
regular conditional probabilities ( [
n
= x
) vanish
n
almost surely on
the subsets of (
n
)
1
(W) with Hausdor dimension n1. Then there exists
a unique solution of the MongeKantorovitch problem, denoted by (, )
and is supported by the graph of a Borel map T which is the solution of
the Monge problem. T : W W is of the form T = I
W
+ , where H
almost surely. Besides we have
d
2
H
(, ) =
_
WW
[T(x) x[
2
H
d(x, y)
=
_
W
[T(x) x[
2
H
d(x) ,
and for
n
almost almost all x
n
, the map u u + (u + x
n
) is cyclically
monotone on (
n
)
1
x
n
, in the sense that
N
i=1
_
u
i
+ (x
n
+ u
i
), u
i+1
u
i
_
H
0
n
almost surely, for any cyclic sequence u
1
, . . . , u
N
, u
N+1
= u
1
from
n
(W). Finally, if, for any n 1,
n
almost surely, ( [
n
= y
) also
vanishes on the n 1Hausdor dimensional subsets of (
n
)
1
(W), then T
is invertible, i.e, there exists S : W W of the form S = I
W
+ such that
H satises a similar cyclic monotononicity property as and that
1 = (x, y) W W : T S(y) = y
= (x, y) W W : S T(x) = x .
In particular we have
d
2
H
(, ) =
_
WW
[S(y) y[
2
H
d(x, y)
=
_
W
[S(y) y[
2
H
d(y) .
186 MongeKantorovitch
Remark 12.6.6 In particular, for all the measures which are absolutely
continuous with respect to the Wiener measure , the second hypothesis is
satised, i.e., the measure ( [
n
= x
n
) vanishes on the sets of Hausdor
dimension n 1.
Proof: Let (F
n
, n 1) be the increasing sequence of regular subspaces
associated to (
n
, n 1), whose union is dense in W. From Lemma 12.6.4,
for any F
n
, there exists a map T
n
, such that
n
y = T
n
(
n
x,
n
x) for almost
all (x, y), where
n
= I
W
n
. Write T
n
as I
n
+
n
, where I
n
denotes the
identity map on F
n
. Then we have the following representation:
n
y =
n
x +
n
(
n
x,
n
x) ,
almost surely. Since
n
y
n
x =
n
(y x)
=
n
(
n
x,
n
x)
and since y x H almost surely, (
n
y
n
x, n 1) converges almost
surely. Consequently (
n
, n 1) converges , hence almost surely to a
measurable . Consequently we obtain
((x, y) W W : y = x + (x)) = 1 .
Since J() < , takes its values almost surely in the CameronMartin space
H. The cyclic monotonicity of is obvious. To prove the uniqueness, assume
that we have two optimal solutions
1
and
2
with the same marginals and
J(
1
) = J(
2
). Since J() is linear, the measure dened as =
1
2
(
1
+
2
) is also optimal and it has also the same marginals and . Consequently,
it is also supported by the graph of a map T. Note that
1
and
2
are
absolutely continuous with respect to , let L
1
(x, y) be the RadonNikodym
density of
1
with respect to . For any f C
b
(W), we then have
_
W
fd =
_
WW
f(x)d
1
(x, y)
=
_
WW
f(x)L
1
(x, y)d(x, y)
=
_
W
f(x)L
1
(x, T(x))d(x) .
Therefore we should have almost surely, L
1
(x, T(x)) = 1, hence also L
1
= 1
almost everywhere and this implies that =
1
=
2
. The second part
about the invertibility of T is totally symmetric, hence its proof follows along
the same lines as the proof for T.
Mass Transportation 187
Corollary 12.6.7 Assume that is equivalent to the Wiener measure ,
then for any h
1
, . . . , h
N
H and for any permutation of 1, . . . , N, we
have, with the notations of Theorem 12.6.5,
N
i=1
_
h
i
+ (x + h
i
), h
(i)
h
i
_
H
0
almost surely.
Proof: Again with the notations of the theorem,
k
almost surely, the graph
of the map x
k
x
k
+
k
(x
k
, x
k
) is cyclically monotone on F
k
. Hence, for
the case h
i
F
n
for all i = 1, . . . , N and n k, we have
N
i=1
_
h
i
+ x
k
+
k
(x
k
+ h
i
, x
k
), h
(i)
h
i
_
H
0 .
Since
i
(x
k
, h
(i)
h
i
)
H
= 0, we also have
N
i=1
_
h
i
+
k
(x
k
+ h
i
, x
k
), h
(i)
h
i
_
H
0 .
We know that
k
(x
k
+h
i
, x
k
) converges to (x+h
i
) almost surely. Moreover
h (x + h) is continuous from H to L
0
() and the proof follows.
12.7 The MongeAmp`ere equation
Assume that W = IR
n
and take a density L ILlog IL. Let ID
2,1
be the
1convex function such that T = I + maps to L . Let S = I + be
its inverse with ID
2,1
. Let now
2
a
be the second Alexandrov derivative
of , i.e., the RadonNikodym derivative of the absolutely continuous part
of the vector measure
2
with respect to the Gaussian measure on IR
n
.
Since is 1convex, it follows that
2
I
IR
n
in the sense of the distribu
tions, consequently
2
a
I
IR
n
almost surely. Dene also the Alexandrov
version /
a
of / as the RadonNikodym derivative of the absolutely contin
uous part of the distribution /. Since we are in nite dimensional situation,
we have the explicit expression for /
a
as
/
a
(x) = ((x), x)
IR
n
trace
_
2
a
_
.
Let be the Gaussian Jacobian
= det
2
_
I
IR
n
+
2
a
_
exp
_
/
a
1
2
[[
2
IR
n
_
.
188 MongeKantorovitch
Remark 12.7.1 In this expression as well as in the sequel, the notation
det
2
(I
H
+ A) denotes the modied CarlemanFredholm determinant of the
operator I
H
+ A on a Hilbert space H. If A is an operator of nite rank,
then it is dened as
det
2
(I
H
+ A) =
n
i=1
(1 + l
i
)e
l
i
,
where (l
i
, i n) denotes the eigenvalues of A counted with respect to their
multiplicity. In fact this determinant has an analytic extension to the space
of HilbertSchmidt operators on a separable Hilbert space, cf. [23] and Ap
pendix A.2 of [101]. As explained in [101], the modied determinant exists
for the HilbertSchmidt operators while the ordinary determinant does not,
since the latter requires the existence of the trace of A. Hence the modied
CarlemanFredholm determinant is particularly useful when one studies the
absolute continuity properties of the image of a Gaussian measure under non
linear transformations in the setting of innite dimensional Banach spaces
(cf., [101] for further information).
It follows from the change of variables formula given in Corollary 4.3 of [60],
that, for any f C
b
(IR
n
),
E[f T ] = E
_
f 1
(M)
_
,
where M is the set of nondegeneracy of I
IR
n
+
2
a
,
(x) =
1
2
[x[
2
+ (x)
and denotes the subdierential of the convex function . Let us note that,
in case L > 0 almost surely, T has a global inverse S, i.e., ST = T S = I
IR
n
almost surely and ((M)) = (S
1
(M)). Assume now that > 0
almost surely, i.e., that (M) = 1. Then, for any f C
b
(IR
n
), we have
E[f T] = E
_
f T
T
1
T
_
= E
_
f
1
T
1
1
(M)
_
= E[f L] ,
where T
1
denotes the left inverse of T whose existence is guaranteed by
Theorem 12.4.2. Since T(x) (M) almost surely, it follows from the
above calculations
1
= L T ,
Mass Transportation 189
almost surely. Take now any t [0, 1), the map x
1
2
[x[
2
H
+ t(x) =
t
(x)
is strictly convex and a simple calculation implies that the mapping T
t
=
I +t is (1 t)monotone (cf. [101], Chapter 6), consequently it has a left
inverse denoted by S
t
. Let us denote by
t
the Legendre transformation of
t
:
t
(y) = sup
xIR
n
(x, y)
t
(x) .
A simple calculation shows that
t
(y) = sup
x
_
(1 t)
_
(x, y)
[x[
2
2
_
+ t
_
(x, y)
[x[
2
2
(x)
__
(1 t)
[y[
2
2
+ t
1
(y) .
Since
1
is the Legendre transformation of
1
(x) = [x[
2
/2 + (x) and since
L ILlog IL, it is nite on a convex set of full measure, hence it is nite
everywhere. Consequently
t
(y) < for any y IR
n
. Since a nite, convex
function is almost everywhere dierentiable,
t
exists almost everywhere
on and it is equal almost everywhere on T
t
(M
t
) to the left inverse T
1
t
, where
M
t
is the set of nondegeneracy of I
IR
n
+ t
2
a
. Note that (M
t
) = 1. The
strict convexity implies that T
1
t
is Lipschitz with a Lipschitz constant
1
1t
.
Let now
t
be the Gaussian Jacobian
t
= det
2
_
I
IR
n
+ t
2
a
_
exp
_
t/
a
t
2
2
[[
2
IR
n
_
.
Since the domain of is the whole space IR
n
,
t
> 0 almost surely, hence,
as we have explained above, it follows from the change of variables formula
of [60] that T
t
is absolutely continuous with respect to and that
1
t
= L
t
T
t
,
almost surely.
Let us come back to the innite dimensional case: we rst give an inequality
which may be useful.
Theorem 12.7.2 Assume that (W, , H) is an abstract Wiener space, as
sume that K, L IL
1
+
() with K > 0 almost surely and denote by T : W W
the transfer map T = I
W
+, which maps the measure Kd to the measure
Ld. Then the following inequality holds:
1
2
E[[[
2
H
] E[log K + log L T] . (12.7.11)
190 MongeKantorovitch
Proof: Let us dene k as k = K T
1
, then for any f C
b
(W), we have
_
W
f(y)L(y)d(y) =
_
W
f T(x)K(x)d(x)
=
_
W
f T(x)k T(x)d(x) ,
hence
T =
L
k
..
It then follows from the inequality 12.3.2 that
1
2
E
_
[[
2
H
_
E
_
L
k
log
L
k
_
= E
_
log
L T
k T
_
= E[log K + log L T] .
Suppose that ID
2,1
is a 1convex Wiener functional. Let V
n
be the sigma
algebra generated by e
1
, . . . , e
n
, where (e
n
, n 1) is an orthonormal
basis of the CameronMartin space H. Then
n
= E[[V
n
] is again 1convex
(cf. Chapter 11), hence /
n
is a measure as it can be easily veried. However
the sequence (/
n
, n 1) converges to / only in ID
t
. Consequently, there
is no reason for the limit / to be a measure. In case this happens, we
shall denote the RadonNikodym density with respect to , of the absolutely
continuous part of this measure by /
a
.
Lemma 12.7.3 Let ID
2,1
be 1convex and let V
n
be dened as above and
dene F
n
= E[[V
n
]. Then the sequence (/
a
F
n
, n 1) is a submartingale,
where /
a
F
n
denotes the absolutely continuous part of the measure /F
n
.
Proof: Note that, due to the 1convexity, we have /
a
F
n
/F
n
for any
n IN. Let X
n
= /
a
F
n
and f ID be a positive, V
n
measurable test
function. Since /E[[V
n
] = E[/[V
n
], we have
E[X
n+1
f] /F
n+1
, f)
= /F
n
, f) ,
where , ) denotes the duality bracket for the dual pair (ID
t
, ID). Conse
quently
E[f E[X
n+1
[V
n
]] /F
n
, f) ,
Mass Transportation 191
for any positive, V
n
measurable test function f, it follows that the absolutely
continuous part of /F
n
is also dominated by the same conditional expectation
and this proves the submartingale property.
Lemma 12.7.4 Assume that L ILlog IL is a positive random variable
whose expectation is one. Assume further that it is lower bounded by a con
stant a > 0. Let T = I
W
+ be the transport map such that T = L.
and let T
1
= I
W
+. Then / is a Radon measure on (W, B(W)). If L
is upper bounded by b > 0, then / is also a Radon measure on (W, B(W)).
Proof: Let L
n
= E[L[V
n
], then L
n
a almost surely. Let T
n
= I
W
+
n
be the transport map which satises T
n
= L
n
. and let T
1
n
= I
W
+
n
be its inverse. We have
L
n
= det
2
_
I
H
+
2
a
n
_
exp
_
/
a
1
2
[
n
[
2
H
_
.
By the hypothesis log L
n
log a. Since
n
is 1convex, it follows from
the nite dimensional results that det
2
(I
H
+
2
a
n
) [0, 1] almost surely.
Therefore we have
/
a
n
log a ,
besides /
n
/
a
n
as distributions, consequently
/
n
log a
as distributions, for any n 1. Since lim
n
/
n
= / in ID
t
, we obtain
/ log a, hence log a / 0 as a distribution, hence / is a
Radon measure on W. This proves the rst claim. Note that whenever L is
upperbounded, = 1/L T is lowerbounded, hence the proof of the second
claim is similar to that of the rst one.
Theorem 12.7.5 Assume that L is a strictly positive bounded random vari
able with E[L] = 1. Let ID
2,1
be the 1convex Wiener functional such
that
T = I
W
+
is the transport map realizing the measure L. and let S = I
W
+ be
its inverse. Dene F
n
= E[[V
n
], then the submartingale (/
a
F
n
, n 1)
converges almost surely to /
a
. Let () be the random variable dened as
() = lim inf
n
n
=
_
liminf
n
det
2
_
I
H
+
2
a
F
n
_
_
exp
_
/
a
1
2
[[
2
H
_
192 MongeKantorovitch
where
n
= det
2
_
I
H
+
2
a
F
n
_
exp
_
/
a
F
n
1
2
[F
n
[
2
H
_
.
Then it holds true that
E[f T ()] E[f] (12.7.12)
for any f C
+
b
(W), in particular ()
1
LT
almost surely. If E[()] = 1,
then the inequality in (12.7.12) becomes an equality and we also have
() =
1
L T
.
Proof: Let us remark that, due to the 1convexity, 0 det
2
(I
H
+
2
a
F
n
)
1, hence the liminf exists. Now, Lemma 12.7.4 implies that / is a Radon
measure. Let F
n
= E[[V
n
], then we know from Lemma 12.7.3 that (/
a
F
n
, n
1) is a submartingale. Let /
+
denote the positive part of the measure /.
Since /
+
/, we have also E[/
+
[V
n
] E[/[V
n
] = /F
n
. This implies
that E[/
+
[V
n
] /
+
a
F
n
. Hence we nd that
sup
n
E[/
+
a
F
n
] <
and this condition implies that the submartingale (/
a
F
n
, n 1) converges
almost surely. We shall now identify the limit of this submartingale. Let
/
s
G be the singular part of the measure /G for a Wiener function G such
that /G is a measure. We have
E[/[V
n
] = E[/
a
[V
n
] + E[/
s
[V
n
]
= /
a
F
n
+/
s
F
n
,
hence
/
a
F
n
= E[/
a
[V
n
] + E[/
s
[V
n
]
a
almost surely, where E[/
s
[V
n
]
a
denotes the absolutely continuous part of the
measure E[/
s
[V
n
]. Note that, from the Theorem of Jessen (cf., for exam
ple Theorem 1.2.1 of [101]), lim
n
E[/
+
s
[V
n
]
a
= 0 and lim
n
E[/
s
[V
n
]
a
= 0
almost surely, hence we have
lim
n
/
a
F
n
= /
a
,
almost surely. To complete the proof, an application of the Fatou lemma
implies that
E[f T ()] E[f]
= E
_
f T
1
L T
_
,
Mass Transportation 193
for any f C
+
b
(W). Since T is invertible, it follows that
()
1
L T
almost surely. Therefore, in case E[()] = 1, we have
() =
1
L T
,
and this completes the proof.
Corollary 12.7.6 Assume that K, L are two positive random variables with
values in a bounded interval [a, b] (0, ) such that E[K] = E[L] = 1. Let
T = I
W
+ , ID
2,1
, be the transport map pushing Kd to Ld, i.e,
T(Kd) = Ld. We then have
L T () K ,
almost surely. In particular, if E[()] = 1, then T is the solution of the
MongeAmp`ere equation.
Proof: Since a > 0,
dT
d
=
L
K T
b
a
.
Hence, Theorem 12.7.12 implies that
E[f T L T ()] E[f L]
= E[f T K] ,
consequently
L T () K ,
the rest of the claim is now obvious.
For later use we give also the folowing result:
Theorem 12.7.7 Assume that L is a positive random variable of class ILlog IL
such that E[L] = 1. Let ID
2,1
be the 1convex function corresponding to
the transport map T = I
W
+ . Dene T
t
= I
W
+ t, where t [0, 1].
Then, for any t [0, 1], T
t
is absolutely continuous with respect to the
Wiener measure .
194 MongeKantorovitch
Proof: Let
n
be dened as the transport map corresponding to L
n
=
E[P
1/n
L
n
[V
n
] and dene T
n
as I
W
+
n
. For t [0, 1), let T
n,t
= I
W
+t
n
.
It follows from the nite dimensional results which are summarized in the
beginning of this section, that T
n,t
is absolutely continuous with respect to
. Let L
n,t
be the corresponding RadonNikodym density and dene
n,t
as
n,t
= det
2
_
I
H
+ t
2
a
n
_
exp
_
t/
a
t
2
2
[
n
[
2
H
_
.
Besides, for any t [0, 1),
_
(I
H
+ t
2
a
n
)h, h
_
H
> 0 , (12.7.13)
almost surely for any 0 ,= h H. Since
n
is of nite rank, 12.7.13 implies
that
n,t
> 0 almost surely and we have shown at the beginning of this
section
n,t
=
1
L
n,t
T
n,t
almost surely. An easy calculation shows that t log det
2
(I + t
2
a
n
)
is a nonincreasing function. Since /
a
n
/
n
, we have E[/
a
n
] 0.
Consequently
E [L
t,n
log L
t,n
] = E [log L
n,t
T
n,t
]
= E [log
t,n
]
= E
_
log det
2
_
I
H
+ t
2
n
_
+ t/
a
n
+
t
2
2
[
n
[
2
H
_
E
_
log det
2
_
I
H
+
2
n
_
+/
a
n
+
1
2
[
n
[
2
H
_
= E [L
n
log L
n
]
E[Llog L] ,
by the Jensen inequality. Therefore
sup
n
E[L
n,t
log L
n,t
] <
and this implies that the sequence (L
n,t
, n 1) is uniformly integrable for
any t [0, 1]. Consequently it has a subsequence which converges weakly in
L
1
() to some L
t
. Since, from Theorem 12.4.2, lim
n
n
= in ID
2,1
, where
is the transport map associated to L, for any f C
b
(W), we have
E[f T
t
] = lim
k
E [f T
n
k
,t
]
= lim
k
E [f L
n
k
,t
]
= E[f L
t
] ,
Mass Transportation 195
hence the theorem is proved.
12.7.1 The solution of the MongeAmp`ere equation via
Itorenormalization
We can interpret the MongeAmp`ere equation as follows: given two proba
bility densities K and L, nd a map T : W W such that
L T J(T) = K
almost surely, where J(T) is a kind of Jacobian to be written in terms of T.
In Corollary 12.7.6, we have shown the existence of some () which gives
an inequality instead of the equality. Although in the nite dimensional case
there are some regularity results about the transport map (cf., [15]), in the
innite dimensional case such techniques do not work. All these diculties
can be circumvented using the miraculous renormalization of the Ito calculus.
In fact assume that K and L satisfy the hypothesis of the corollary. First let
us indicate that we can assume W = C
0
([0, 1], IR) (cf., [101], Chapter II, to
see how one can pass from an abstract Wiener space to the standard one) and
in this case the CameronMartin space H becomes H
1
([0, 1]), which is the
space of absolutely continuous functions on [0, 1], with a square integrable
Sobolev derivative. Let now
=
K
L T
,
where T is as constructed above. Then . is a Girsanov measure for the map
T. This means that the law of the stochastic process (t, x) T
t
(x) under
. is equal to the Wiener measure, where T
t
(x) is dened as the evaluation
of the trajectory T(x) at t [0, 1]. In other words the process (t, x) T
t
(x)
is a Brownian motion under the probability .. Let (T
T
t
, t [0, 1]) be its
ltration, the invertibility of T implies that
t[0,1]
T
T
t
= B(W) .
is upper and lower bounded almost surely, hence also .almost surely.
The Ito representation theorem implies that it can be represented as
= E[
2
] exp
_
_
1
0
s
dT
s
1
2
_
1
0
[
s
[
2
ds
_
,
196 MongeKantorovitch
where () =
_
0
s
ds is an Hvalued random variable. In fact can be
calculated explicitly using the ItoClark representation theorem, and it is
given as
t
=
E
[D
t
[T
T
t
]
E
[[T
T
t
]
(12.7.14)
dt dalmost surely, where E
_
1
0
s
dT
s
1
2
_
1
0
[
s
[
2
ds
_
= K ,
almost surely, where is given with (12.7.14).
Chapter 13
Stochastic Analysis on Lie
Groups
Introduction
This chapter is a partial survey of the construction of Sobolevtype analysis
on the path space of a Lie group. The word partial refers to the fact that
we give some new results about the quasiinvariance of anticipative transfor
mations and the corresponding measure theoretical degree theorems in the
last section. Almost all the theory has been initiated by S. Albeverio and
R. H.Krohn ([4]), L. Gross ([38, 39]) and M. P. Malliavin and P. Malliavin
([57]). Although the study of the similar subjects has already begun in the
case of manifolds (cf. [17]), we prefer to understand rst the case of the Lie
groups because of their relative simplicity and this will give a better idea of
what is going on in former situation; since the frame of the Lie groupvalued
Brownian motion represents the simplest nonlinear and nontrivial case in
which we can construct a Sobolev type functional analysis on the space of
the trajectories.
After some preliminaries in the second section we give the denitions of
the basic tools in the third section, namely the left and right derivatives
on the path space. The fourth section is devoted to the left divergence, in
the next one we study the OrnsteinUhlenbeck operator, Sobolev spaces and
some applications like the zeroone law. Sixth section is a compilation of
the formulas based essentially on the variation of the constants method of
the ordinary linear dierential equations which are to be used in the follow
ing sections. Section seven is devoted to the right derivative which is more
technical and interesting than the left one; since it contains a rotation of the
path in the sense of [96]. We also dene there the skewsymmetric rotational
197
198 Stochastic Analysis on Lie Groups
derivative and study some of its properties. Eighth section is devoted to the
quasiinvariance at the left and at the right with respect to the multiplicaton
of the path with deterministic paths of nite variation. Loop space case is
also considered there.
Section nine deals with the absolute continuity of the path and loop mea
sures under the transformation which consists of multiplying from the left
the generic trajectory with some random, absolutely continuous and antic
ipative path. We prove a generalization of the CampbellBakerHausdor
formula which is fundemental. To prove this we have been obliged to employ
all the recent sophisticated techniques derived in the at case. Afterwards,
the extension of the Ramer and the degree theorems are immediate.
In this chapter we have focuse our attention to the probabilistic and func
tional analytic problems. For the more general case of Riemannian manifolds
cf. [56] and the references therein.
13.1 Analytic tools on group valued paths
Let G be a nite dimensional, connected, locally compact Lie group and (
be its Lie algebra of left invariant vector elds which is isomorphic to the
tangent space at identity of G, denoted by T
e
(G) which is supposed to be
equipped with an inner product. C = C
G
denotes C
e
([0, 1], G) (i.e., p(0) = e
for p C
G
). C
denotes C
0
([0, 1], (). Let
H = H
=
_
h C
:
_
1
0
[
h(t)[
2
dt = [h[
2
<
_
.
Our basic Wiener space is (C
into C
G
and we denote by the image of under this
mapping. Similarly, if h H then we denote by e(h) the solution of the
following dierential equation:
e
t
(h) = e +
_
t
0
e
s
(h)
h
s
ds . (13.1.1)
Analytic Tools 199
Theorem 13.1.1 (CampbellBakerHausdor Formula) For any h
H the following identity is valid almost surely:
p(w + h) = e(
Adp(w)h)p(w) , (13.1.2)
where
Adp(w)h is the Hvalued random variable dened by
_
Adp(w)h
_
(t) =
_
t
0
Adp
s
(w)
h(s)ds .
Remark: In case we work with matrices,
Adp(w)h is dened as
_
t
0
p
s
(w)
h(s)p
1
s
(w)ds .
Remark: This theorem implies in particular that the C
G
valued random
variable w p(w) has a modication, denoted again by the same letter p,
such that h p(w + h) is a smooth function of h H for any w C
.
Calculation of (f(p
t
(w))): We have f(p
t
(w + h)) = f(e
t
(
Adph)p
t
)
where e
t
(h), h H is dened by the equation (13.1.1) . Let us write g =
p
t
(w) and F(x) = f(xg). Then
F(e
t
(
Adph)) = F(e) +
_
t
0
Adp
s
h
s
F(e
s
(
Adph))ds .
Hence
d
d
F(e
t
(
Adph))[
=0
=
_
t
0
Adp
s
h(s)F(e)ds .
Now if X is a left invariant vector eld on G, then we have XF(x) =
X(f(xg)) = X(f(gg
1
xg)) = (Adg
1
X)f(gx) by the left invariance of X.
In particular, for x = e, we have XF(e) = (Adg
1
X)f(g). Replacing g with
p
t
(w) above, we obtain
h
(f(p
t
)) = Adp
1
t
_
t
0
Adp
s
h
s
f(p
t
)ds (13.1.3)
=
_
Adp
1
t
_
t
0
Adp
s
h
s
ds
_
f(p
t
) . (13.1.4)
Notation: In the sequel, we shall denote the map h
_
0
Adp
s
h(s)ds by
p
h or by
Adph as before, depending on the notational convenience.
200 Stochastic Analysis on Lie Groups
Denition 13.1.2 If F : C
G
IR is a cylindrical function, h H, we
dene
L
h
F(p) =
d
d
F(e(h)p)[
=0
(13.1.5)
R
h
F(p) =
d
d
F(pe(h))[
=0
, (13.1.6)
where p is a generic point of C
G
. L is called the left derivative and R is
called the right derivative.
A similar calculation as above gives us
L
h
f(p
t
) = Adp
1
t
h
t
f(p
t
) (13.1.7)
R
h
f(p
t
) = h
t
f(p
t
) . (13.1.8)
If F(p) = f(p
t
1
, , p
t
n
), then
L
h
F(p) =
n
i=1
Adp
1
t
i
h
t
i
f(p
t
1
, , p
t
n
) (13.1.9)
R
h
F(p) =
n
i=1
h
t
i
f(p
t
1
, , p
t
n
) (13.1.10)
h
(F p(w)) =
n
i=1
Adp
1
t
i
(w)
p(w)
h
t
i
f(p
t
1
, , p
t
n
)(w) . (13.1.11)
Proposition 13.1.3 L
h
is a closable operator on L
p
() for any p > 1 and
h H. Moreover, we have
(L
h
F)(p(w)) =
1
p(w)
(h)
(F(p(w))) .
Proof: Suppose that (F
n
) is a sequence of cylindrical functions on C
G
con
verging to zero in L
p
() and that (L
h
F
n
) is Cauchy in L
p
(). Then, from the
formulas (7) and (9), we have
(L
h
F
n
)(p(w)) =
1
p(w)
(h)
(F
n
(p(w))) ,
since is a closed operator on L
p
(), we have necessarily lim
n
L
h
F
n
= 0
almost surely.
Remark 13.1.4 On the cylindrical functions we have the identity
R
h
F(p(w)) =
m(h)
(F(p(w)))
Analytic Tools 201
where m(h)
t
= Adp
t
(w)
_
t
0
Adp
1
s
e(h)
h
F(p e(h)) ,
where
e(h)
k H is dened as
e(h)
k(t) =
_
t
0
Ade
s
(h)
k(s)ds .
Remark 13.1.6 On the extended domain of L, we have the identity
L
h
F p(w) =
1
p(w)
(h)
(F p) (13.1.12)
= (
1
p(w)
(F p), h) (13.1.13)
= (
p(w)
(F p), h) (13.1.14)
if we assume that the scalar product of ( is invariant with respect to the inner
automorphisms, in which case G becomes of compact type, hence linear, i.e.,
a space of matrices and
p
becomes an isometry of H.
Proposition 13.1.7 If : C
G
H is a measurable random variable, then
we have
(L
F) p =
1
p(w)
(p)
(F p)
= (
p
(F p), p) .
202 Stochastic Analysis on Lie Groups
Proof: By denition, F Dom(L) i F p Dom() and in this case
h L
h
F induces an Hvalued random variable, denoted by LF. Then, for
any complete orthonormal basis (h
i
, i IN) of H
L
F p =
i
L
h
i
F p(, h
i
) p
=
1
h
i
(F p)(, h
i
)
H
p
=
1
h
i
(F p)(
1
p,
1
h
i
)
H
=
1
p
(p)
(F p)
=
_
p
(F p), p
_
H
13.2 The left divergence L
If : C
G
H is a cylindrical random variable and if F is a smooth function
on C
G
, we have
E
[L
F] = E
[(L
F) p]
= E
1
(p)
(F p)]
= E
[F p (
1
( p))] .
Since L is a closed operator, its adjoint with respect to is welldened and
we have
E
[L
F] = E
[F L
]
= E
[F p (L
) p] .
We have
Proposition 13.2.1 The following identity is true:
(L
) p = (
1
( p)) .
Proof: We have already tested this identity for cylindrical and F. To
complete the proof it is sucient to prove that the cylindrical F are dense
in L
p
(). Then the proof will follow from the closability of L. The density
follows from the fact that (p
t
; t [0, 1]) and the Wiener process generate the
same sigma algebra and from the monotone class theorem.
Wiener Chaos 203
Lemma 13.2.2 Let (H
t
, t [0, 1]) be the ltration (eventually completed)
of the process (p
t
, t [0, 1]) and (T
t
, t [0, 1]) be the ltration of the basic
Wiener process. We have
E
[[H
t
] p = E
[ p[T
t
]
almost surely.
Proof: Let f be a smooth function on IR
n
. Then
E
[ p f(p
t
1
(w), . . . , p
t
n
(w))] = E
[f(p
t
1
, . . . , p
t
n
)]
= E
[E
[[H
t
]f(p
t
1
, . . . , p
t
n
)]
= E
[E
[[H
t
] pf(p
t
1
(w), . . . , p
t
n
(w))] ,
since E
[[H
t
] p is T
t
measurable, the proof follows.
If F is a nice random variable on C
G
and denote by the optional pro
jection with respect to (T
t
). Using ItoClark representation theorem, we
have
F p = E
[F p] + [(F p)]
= E
[F] +
_
1
p
(F p)
_
= E
[F] +
_
1
p
p
(F p)
_
= E
[F] +
_
1
p
(LF p)
_
= E
[F] +
_
1
p
( LF) p)
_
= E
[F] + (L
( LF)) p
almost surely, where denotes the optional projection with respect to the
ltration (H
t
). Consequently, we have proved the following
Theorem 13.2.3 Suppose that F L
p
(), p > 1 such that F p D
p,1
.
Then we have
F = E
[F] + L
LF
almost surely.
13.3 OrnsteinUhlenbeck operator and the Wiener
chaos
Let F be a nice function on C
G
, then
(L
LF) p = L
(LF) p (13.3.15)
204 Stochastic Analysis on Lie Groups
=
_
1
p
(LF p)
_
(13.3.16)
=
_
1
((F p))
_
(13.3.17)
= (F p) (13.3.18)
= /(F p) , (13.3.19)
where / = is the OrnsteinUhlenbeck operator on W.
Denition 13.3.1 We denote by / the operator L
, i.e.,
P
t
f(w) =
_
C
G
f(e
t
w +
1 e
2t
y)(dy) .
Then it is easy to see that
d
dt
Q
t
F(p)[
t=0
= /F(p) .
Hence we can dene the spaces of distributions, verify Meyer inequalities,
etc. , as in the at case (cf. [72]): Let be an equivalence class of random
variables on (C
G
, ) with values in some separable Hilbert space X. For
q > 1, k IN, we will say that is in S
q,k
(X), if there exists a sequence of
cylindrical functions (
n
) which converges to in L
q
(, X) such that (
n
p)
is Cauchy in ID
q,k
(X). For X = IR, we write simply S
q,k
instead of S
q,k
(IR).
We denote by S(X) the projective limit of the spaces (S
q,k
; q > 1, k IN).
Using Meyer inequalities and the fact that w p(w) is smooth in the Sobolev
sense, we can show easily that, for q > 1, k Z
1. the left derivative L possesses a continuous extension from S
q,k
(X) into
S
q,k1
(X H), where
S
q,k
(X) =
_
>0
S
q,k
(X) .
2. L
q,k1
(X).
3. Consequently L maps S(X) continuously into S(X H) and L
maps
S(X H) continuously into S(X).
Wiener Chaos 205
4. By duality, L and L
[F p] +
i=1
1
n!
n
E
[
n
(F p)] .
On the other hand
(F p) =
1
(LF p) = (
1
LF) p
almost surely. Iterating this identity, we obtain
n
(F p) = ((
1
L)
n
F) p .
Therefore
E
[
n
(F p)] = E
[((
1
L)
n
F) p] (13.3.20)
= E
[(
1
L)
n
F] . (13.3.21)
206 Stochastic Analysis on Lie Groups
On the other hand, for K in H
n
(i.e., the symmetric tensor product), we
have
E
[
n
KH p] = E
[(K,
n
(H p))
n
]
= E
_
(K, ((
1
L)
n
H) p)
n
_
= E
_
(K, (
1
L)
n
H)
n
_
= E
[(L
)
n
K H]
= E
[((L
)
n
K) p H p] ,
for any cylindrical function H on C
G
, where (, )
n
denotes the scalar product
in H
n
. We have proved the identity
n
K = ((L
)
n
K) p ,
consequently the following Wiener decomposition holds:
Theorem 13.3.4 For any F L
2
(), one has
F = E
[F] +
n=1
1
n!
(L
)
n
_
E
[(
1
L)
n
F]
_
where the sum converges in L
2
.
The ItoClark representation theorem suggests us a second kind of Wiener
chaos decomposition. First we need the following:
Lemma 13.3.5 The set
=
_
exp
_
L
h
1
2
[h[
2
H
_
; h H
_
is dense in L
p
() for any p 1.
Proof: We have
L
h p = (
1
p(w)
h)
=
_
1
0
(Adp
1
s
h(s), dW
s
)
=
_
1
0
(
h(s), Adp
s
dW
s
) .
By Paul Levys theorem, t B
t
=
_
t
0
Adp
s
dW
s
denes a Brownian motion.
Hence, to prove the lemma, it suces to show that W and B generate the
Some Useful Formulea 207
same ltration. To see this, note that the process (p
t
) satises the following
stochastic dierential equation:
df(p
t
) = H
i
f(p
t
) dW
i
t
,
(f C
(G)), replacing dW
t
by Adp
t
dB
t
we obtain
df(p
t
) = Adp
1
t
H
i
f(p
t
) dB
i
t
.
Since everything is smooth, we see that p(w) is measurable with respect to
the ltration of B. But we know that the ltrations of p and W are equal
from the lemma 13.2.2.
Remark 13.3.6 Using the Brownian motion B
t
dened above we can also
represent the Wiener functionals, this gives another Wiener chaos decompo
sition.
13.4 Some useful formulea
Let us rst recall the variation of constant method for matrixvalued equa
tions:
Lemma 13.4.1 The solution of the equation
t
(h) =
t
+
_
t
0
s
(h)
h(s)ds
is given by
t
(h) =
0
+
_
_
t
0
d
ds
s
e
s
(h)
1
ds
_
e
t
(h) .
Corollary 13.4.2 We have
d
d
e
t
(h) =
__
t
0
Ade
s
(h)
h(s)ds
_
e
t
(h) (13.4.22)
= (
e(h)
h)(t)e
t
(h) . (13.4.23)
Corollary 13.4.3 We have
d
d
Ade
t
(h)
k
t
=
__
t
0
Ade
s
(h)
h
s
ds, Ade
t
(h)
k
t
_
.
208 Stochastic Analysis on Lie Groups
Corollary 13.4.4 We have
d
d
Ade
1
t
(h)
k
t
= Ade
1
t
(h)
__
t
0
Ade
s
(h)
h
s
ds,
k
t
_
.
Proof: Since AdeAde
1
= I, we have
0 =
d
d
Ade
t
(h)Ade
1
t
(h)
k
t
=
_
d
d
Ade
t
(h)
_
Ade
1
t
(h)
k
t
+ Ade
t
(h)
d
d
Ade
1
t
(h)
k
t
,
hence
d
d
Ade
1
t
(h)
k
t
= Ade
1
t
(h)
_
d
d
Ade
t
(h)
_
Ade
1
t
(h)
k
t
= Ade
1
t
(h)
__
t
0
Ade
s
(h)
h
s
ds, Ade
t
(h)Ade
1
t
(h)
k
t
_
= Ade
1
t
(h)
__
t
0
Ade
s
(h)
h
s
ds,
k
t
_
.
In further calculations we shall need to control the terms like
[Ade
1
t
(v)
h
t
Ade
1
t
()
h
t
[
.
For this, we have
Ade
1
t
(v)
h
t
Ade
1
t
()
h
t
=
_
1
0
d
d
Ade
1
t
(v + (1 ))
h
t
d
=
_
1
0
d
d
Ade
1
t
((v ) + )
h
t
d.
From the Corollary 6.3, we have
d
d
Ade
1
t
((v ) + )
h
t
=
Ade
1
t
((v ) + )
__
t
0
Ade
s
((v ) + )( v
s
s
)ds,
h
t
_
.
Therefore
[Ade
1
t
(v)
h
t
Ade
1
t
()
h
t
[
_
1
0
__
t
0
Ade
s
((v ) + )( v
s
s
)ds,
h
t
_
d.
Some Useful Formulea 209
Now we need to control the (norm of the Lie brackets: for this we introduce
some notations: let (e
i
) be a complete, orthonormal basis of (. Since [e
i
, e
j
]
( we should have
[e
i
, e
j
] =
n
k=1
k
ij
e
k
.
For h, k (,
[h, k] =
_
i
h
i
e
i
,
i
k
i
e
i
_
=
i,j
h
i
k
j
[e
i
, e
j
]
=
i,j,k
h
i
k
j
k
i,j
.
Consequently
[[h, k][
2
l
_
_
i,j
h
i
k
j
l
ij
_
_
2
l
_
_
i,j
h
2
i
k
2
j
_
_
_
_
i,j
(
l
ij
)
2
_
_
=
l
[h[
2
[k[
2
[
l
[
2
2
= [h[
2
[k[
2
l
[
l
[
2
2
,
where [ [
2
refers to the HilbertSchmidt norm on (. Although this is well
known, let us announce the above result as a lemma for later reference:
Lemma 13.4.5 For any h, k (, we have
[[h, k][
[h[
[k[
l
[
l
[
2
2
_
1/2
.
We have also the immediate consequence
Lemma 13.4.6 For any h, k H
Ade
1
t
(v)
h
t
Ade
1
t
()
h
t

2
[
h
t
[
_
t
0
[ v
s
s
[
ds,
where 
2
2
=
[
l
[
2
2
.
210 Stochastic Analysis on Lie Groups
Lemma 13.4.7 We have
d
d
(e(h)p) =
_
L(e(h)p),
Ade(h)h
_
H
.
Proof: We have
e
t
(ah)e
t
(bh) = e
t
(a
Ade
1
(bh)h + bh) ,
hence
e
t
(ah + bh) = e
t
(a
Ade
1
(bh)h)e
t
(bh)
= e
t
(b
Ade
1
(ah)h)e
t
(ah) ,
therefore
e
t
(( + )h) = e
t
(
Ade(h)h)e
t
(h) ,
which gives
d
d
(e(h)p) =
_
L(e(h)p),
Ade(h)h
_
H
.
13.5 Right derivative
Recall that we have dened
R
h
(p) =
d
d
(p e(h))[
=0
.
Since ( consists of left invariant vector elds, we have, using the global
notations :
R
h
f(p
t
) = (h
t
f)(p
t
) ,
where h
t
f is the function obtained by applying the vector eld h
t
to the
smooth function f. The following is straightforward:
Lemma 13.5.1 We have
p
t
(w)e
t
(h) = p
t
__
0
Ade
1
s
(h)dW
s
+ h
_
,
where e
t
(h) for h H is dened in (13.1.1).
Right Derivative 211
Lemma 13.5.2 We have
E
[R
h
F p] = E
[F p
_
1
0
h
s
dW
s
] ,
for any cylindrical function F.
Proof: From the Lemma 13.5.1, p
t
(w)e
t
(h) = p
t
(h +
_
0
Ade
1
s
(h)dW
s
).
Since
_
0
Ade
1
s
(h)dW
s
is a Brownian motion, it follows from the Girsanov
theorem that
E
_
F(p(w)e(h)) exp
_
_
1
0
(
h
s
, Ade
1
s
(h)dW
s
)
2
2
[h[
2
H
__
= E[F] ,
dierentiating at = 0 gives the result.
Denition 13.5.3 For h H and F smooth, dene
Q
h
F(w) by
Q
h
F(w) = F
__
0
Ade
1
s
(h)dW
s
_
,
note that since
_
0
Ade
1
s
(h)dW
s
is a Brownian motion, the composition
of it with F is welldened.
And
X
h
F(w) =
d
d
Q
h
F(w)
=0
.
Example 13.5.4 Let us see how the derivation operator X
h
operates on the
simple functional F = exp k, k H: we have
Q
h
F = exp
_
1
0
(
k
s
, Ade
1
s
(h)dW
s
)
= exp
_
1
0
(Ade
s
(h)
k
s
, dW
s
) ,
hence
X
h
e
k
= e
k
_
1
0
_
[h(s),
k
s
], dW
s
_
.
Proposition 13.5.5 We have the following identity:
(R
h
F) p =
h
(F p) + X
h
(F p) ,
for any F : C
G
IR smooth. In particular, R
h
and X
h
are closable operators.
212 Stochastic Analysis on Lie Groups
Remark 13.5.6 From the above denition, we see that
(R
2
h
)
1 =
2
h
2
_
1
0
_
[h
s
,
h
s
], dW
s
_
.
Hence R
n
does not give the pure chaos but mixes them with those of lower
order. Here enters the notion of universal envelopping algebra.
Notation : For h H, we will denote by
adh the linear operator on H
dened as
adh(k)(t) =
_
t
0
[h(s),
k(s)] ds .
Remark 13.5.7 Suppose that R
h
k = 0, i.e.,
(h, k) +
_
1
0
[h
s
,
k
s
] dW
s
= 0 .
Then (h, k) = 0 and [h(t),
k(t)] = 0 dtalmost surely. Hence this gives more
information than the independence of h and k.
Remark 13.5.8 Suppose that R
h
F = 0 a.s. for any h H. Then we have,
denoting F =
I
n
(f
n
), R
h
F = 0 implies
nf
n
(h) + d(
adh)f
n1
= 0 , k H .
Since f
1
= 0 (this follows from E[R
h
F] = E[
h
F] = 0), we nd that f
n
(h) =
0 for any h H, hence f
n
= 0, and F is a constant.
Remark 13.5.9 If X
h
F = 0 for any h H, we nd that
d(
adh)f
n
= 0
for any h H and for any n. Therefore f
n
s take their values in the tensor
spaces constructed from the center of (.
Recall that in the case of an abstract Wiener space, if A is a deterministic
operator on the CameronMartin space H, then the operator d(A) is dened
on the Fock as
d(A) =
d
dt
(e
tA
)[
t=0
for any cylindrical Wiener functional . We will need the following result
which is wellknown in the Quantum Field Theory folklore:
Right Derivative 213
Lemma 13.5.10 Suppose that A is a skewsymmetric operator on H (i.e.,
A + A
, H, ), we have
the following commutation relation:
[
h
, X
k
] F =
adk(h)
F ,
where h, k H.
We have also
Proposition 13.5.12 Let be a cylindrical function on (C
, H, ) and h
H. We have
E
[(X
h
)
2
] 
2
2
[h[
2
H
E
_
[[
2
H
+
2

2
2
_
,
where is the structure constant of ( and  
2
denotes the HilbertSchmidt
norm of H H.
214 Stochastic Analysis on Lie Groups
Proof: From Lemma 13.5.10, we have X
h
=
_
adh()
_
. Hence
E
_
(X
h
)
2
_
= E[[
adh[
2
H
] + E
_
trace
_
adh
_
2
_
.
From Lemma 13.4.5, we have
adh
2
H

2
2
[h[
2
H
[[
2
H
and
trace
_
adh
_
2

2
2
[h[
2
H

2

2
2
.
Suppose that u D(H) and dene X
u
F, where F is a cylindrical function
on C
, as
aduF. Then using similar calculations, we see that
Corollary 13.5.13 We have the following majoration:
E[[X
u
F[
2
] 
2
E
_
[u[
2
H
[F[
2
H
_
+ 2
2
E
_
[u[
2
H

2
F
2
2
+u
2
2
[F[
2
H
_
.
13.6 Quasiinvariance
Let
t
be a curve in G such that t
t
is absolutely continuous. We can
write it as
d
t
=
t
dt
=
t
1
t
t
dt
Hence
t
= e
t
(
_
1
s
s
ds) provided
_
1
0
[
1
t
t
[
2
dt < . Under these hypoth
esis, we have
t
p
t
(w) = p
t
_
w +
_
0
Adp
1
s
(w)(
1
s
s
)ds
_
.
For any cylindrical : G IR, we have
E
[( p) J
] = E
[]
where
J
p(w) = exp
_
_
1
0
(Adp
1
s
(
1
s
s
), dW
s
)
1
2
_
1
0
[
1
s
s
[
2
ds
_
.
Quasiinvariance 215
Similarly
p
t
(w)
t
= p
t
__
0
Ad
1
s
dW
s
+
_
0
1
s
s
ds
_
,
hence
E
[(p ) K
] = E
[]
where
K
p(w) = exp
_
_
1
0
(
1
s
s
, Ad
1
s
dW
s
)
1
2
_
1
0
[
1
s
s
[
2
ds
_
= exp
_
_
1
0
(
s
1
s
, dW
s
)
1
2
_
1
0
[
1
s
s
[
2
ds
_
. (13.6.24)
As an application of these results, let us choose = e
h
and denote by K
h
the RadonNikodym density dened by
E
[F(pe
h
)] = E[F K
h
] .
Since K
h
is analytic, from Remark 13.1.5, for smooth, cylindrical F,
we have
E[F(p e
h
)] =
n=0
n
n!
E[R
n
h
F(p)]
=
n=0
n
n!
E[F(p) R
n
h
1] ,
hence we have the identity
K
h
=
n=0
n
n!
R
n
h
1 .
Let us now choose F(p) of the form f(p
1
), where f is a smooth function on
G. Then
E[f(p
1
e
h(1)
)] =
n=0
n
n!
E[R
n
h
f(p
1
)]
=
n=0
n
n!
E[h(1)
n
f(p
1
)] .
Let q(x)dx be the law of p
1
where dx is the right invariant Haar measure on
G. Then
E[h(1)
n
f(p
1
)] =
_
G
h(1)
n
f(x) q(x)dx
= (1)
n
_
G
f(x)
h(1)
n
q(x)
q(x)
q(x)dx .
Hence we have proved
216 Stochastic Analysis on Lie Groups
Proposition 13.6.1 We have the following identity:
E
[K
h
[p
1
= x] =
n=0
(1)
n
n!
h(1)
n
q(x)
q(x)
,
for all x G. In particular, if h(1) = 0 then
E
[K
h
[p
1
= x] = 1
for all x G.
Proof: The only claim to be justied is all x instead of almost all x. This
follows from the fact that x E
[K
h
[p
1
= x] is continuous due to the non
degeneracy of the random variable p
1
in the sense of the Malliavin calculus.
Although the analogue of the following result is obvious in the at case, in
the case of the Lie groups, the proof requires more work:
Proposition 13.6.2 The span of K
h
; h H is dense in L
r
() for any
r > 1.
Proof: Let us denote by the span of the set of the densities. Suppose that
F L
s
with E
[
h
F + X
h
F]
= E
[
h
F] ,
since X
h
+ X
h
= 0. For n = 2
0 = E
[R
2
h
F]
= E
[
2
h
F + X
h
h
F +
h
X
h
F + X
2
h
F]
= E
[
2
h
F] + E
[
h
X
h
F] .
Also we have from the calculations of the rst order
E
[
h
X
h
F] = E
[X
h
Fh]
= E
[F(
adh(h))]
= E
[(F,
adh(h))
H
]
= 0 .
Quasiinvariance 217
By polarization, we deduce that, as a tensor in H
, E
[
2
F] = 0. Suppose
now that E
[
i
F] = 0 for i n. We have
E
[R
n+1
h
F] = E
[
n+1
h
F] + supplementary terms .
Between these supplementary terms, those who begin with X
h
or its powers
have automatically zero expectation. We can show via induction hypothesis
that the others are also null. For instance let us take the term E
[
h
X
n
h
F]:
E
[
h
X
n
h
F] = E
[X
n
h
F h]
= (1)
n
E
[F ((
adh)
n
h)]
= 0 ,
the other terms can be treated similarly.
We shall apply these results to the loop measure by choosing a special
form of . Let us rst explain the strategy: replace in the above expressions
the random variable (p(w)) by p(w)f(p
1
(w)). Then we have
_
(p(w)) f((1)p
1
)J
p(w)
_
= E
[ p(w)f(p
1
(w))]
and
_
(p(w)) f(p
1
(1))K
p(w)
_
= E
[ p(w)f(p
1
(w))] .
We shall proceed as follows: let f : G IR be a smooth cylindrical function.
Replace in the above expressions the map p by p f(p
1
(w)) where f is
a smooth function on G. Then we have on the one hand
E
[(p) f((1)p
1
) J
] = E
[(p) f(p
1
)] (13.6.25)
and on the other hand
E
[(p) f(p
1
(1)) K
] = E
[(p) f(p
1
)] . (13.6.26)
Choose such that (1) = e (i.e., the identity of G). Hence (13.6.25) becomes
E
[(p) f(p
1
) J
] = E[(p) f(p
1
)] ,
therefore
_
G
E
_
(p)J
(p)
p
1
= x
_
f(x)q
1
(x)dx =
_
G
E
_
(p)
p
1
= x
_
f(x)q
1
(x)dx ,
218 Stochastic Analysis on Lie Groups
where dx is the Haar measure on G and q
1
is the density of the law of p
1
with
respect to Haar measure which is smooth and strictly positive. Consequently
we obtain
E
_
(p)J
(p)
p
1
= x
_
= E
_
(p)
p
1
= x
_
.
Since both sides are continuous with respect to x, this equality holds every
where. We obtain a similar result also for the right perturbation using the
relation (13.6.26).
A natural candidate for for the loop measure based at e, i.e., for the
measure E
[ [p
1
= e] which we will denote by E
1
, would be
t
(h) = e
t
(h)e
1
1
(th) .
From the calculations of the sixth section, we have
t
(h) = e
t
(h)[
h
t
e
1
1
(th) e
1
1
(th)(
e(th)
h)(1)] .
Hence
Lemma 13.6.3 For
t
(h) = e
t
(h)e
1
1
(th), we have
1
t
(h)
t
(h) = Ade
1
(th)
h
t
(
e(th)
h)(1) .
In this case J
becomes
J
p = exp
_
1
0
_
Adp
1
s
[Ade
1
(sh)
h
s
(
e(sh)
h)(1)], dW
s
_
exp
1
2
_
1
0
[Ade
1
(sh)
h
s
(
e(sh)
h)(1)[
2
ds .
For K
we have
Ad
t
(
1
t
t
) =
t
1
t
= Ade
t
(h)
h
t
Ad(e
t
(h)e
1
1
(th))(
e(th)
h)(1) .
Since [ [ is Adinvariant, we have
K
p = exp
_
1
0
_
Ade
t
(h)
h
t
Ad(e
t
(h)e
1
1
(th))(
e(th)
h(1)), dW
t
_
exp
_
1
0
[
h
t
Ade
1
1
(th)(
e(th)
h(1))[
2
dt .
Remark 13.6.4 Note that as chosen above satises the following dier
ential equation:
t
=
t
(h)[Ade
1
(th)
h
t
e(th)
h(1)] .
Quasiinvariance 219
Let us calculate
d
d
(p(h))[
=0
and
d
d
((h)p)[
=0
for cylindrical . Denote by P
0
: H H
0
the orthogonal projection dened
by
P
0
h(t) = h(t) th(1) .
Then it is easy to see that
d
d
((h)p)[
=0
= L
P
0
h
(p)
and
d
d
(p(h))[
=0
= R
P
0
h
(p) .
Moreover, we have
d
d
J
(h)
(p(w))[
=0
=
_
1
0
_
Adp
1
s
(w)(
h
s
h(1)), dW
s
_
and
d
d
K
(h)
(p)[
=0
=
_
1
0
_
h
s
h(1), dW
s
_
Consequently we have proven
Theorem 13.6.5 For any cylindrical function on the loop space of G, we
have
E
1
[L
P
0
h
] = E
1
[L
P
0
h]
and
E
1
[R
P
0
h
] = E
1
[P
0
h]
for any h H. In particular, the operators L
P
0
h
and R
P
0
h
are closable on
L
p
(
1
) for any p > 1.
Before closing this section let us give a result of L. Gross (cf. [38]):
Lemma 13.6.6 For < 1 the measure ( [p(1) = e) is equivalent to on
(C
G
, H
[F[p(1) = e] = E
_
F
q
1
(p
, e)
q
1
(e, e)
_
,
where q
t
is the density of the law of p
t
with respect to the Haar measure.
220 Stochastic Analysis on Lie Groups
Proof: Without loss of generality we can suppose that F is a continuous and
bounded function on C
G
. Let g be a nice function on G, from the Markov
property, it follows that
E
[F g(p(1))] = E
_
F
_
G
q
1
(p
, y)g(y)dy
_
.
On the other hand, from the disintegration of measures, we have
E
[F g(p(1))] =
_
G
E
[F[p(1) = y]g(y)q
1
(e, y)dy .
Equating both sides gives
E
[F[p(1) = y] =
1
q
1
(e, y)
E
[F q
1
(p
, y)]
dyalmost surely. Since both sides are continuous in y the result follows if
we put y = e.
Remark 13.6.7 Note that we have the following identity:
E
[F(p)[p(1) = e] = E
[F p(w)[p
1
(w) = e]
for any cylindrical function F on C
G
.
13.7 Anticipative transformations
In this section we shall study the absolute continuity of the measures which
are dened as the image of under the mappings which are dened as the left
multiplication of the path p with the exponentials of anticipative (valued
processes. To be able to use the results of the at case we need to extend
the CampbellBakerHausdor formula to this case. We begin by recalling
the following
Denition 13.7.1 Let (W, H, ) be an abstract Wiener space. A random
variable F dened on this space is said to be of class R
p
,k
if F D
q,r
for
some q > 1, r 1 and
sup
[h[
H
[
k
F(w + h)[ L
p
() .
If p = 0, we write F R
0
,k
,
Anticipative Transformations 221
We write F R
p
,k
if the above condition holds for any > 0, and
F R
p
,
if F R
p
,k
for any k IN.
Finally, we say that F R() if F R
p
,
for any p > 1.
Remark 13.7.2 The importance of this class is easy to realize: suppose
that u is an Hvalued random variable, and let F R
0
,
. If (u
n
) is a
sequence of random variables of the form
i<
h
i
1
A
i
converging in prob
ability to u, with A
i
A
j
= for i ,= j, h
i
H, then we can dene
F(w + u
n
(w)) as
F(w + h
i
)1
A
i
(w) and evidently the sequence (F T
n
)
converges in probability, where T
n
= I
W
+ u
n
. Furthermore, the limit is in
dependent of the particular choice of the elements of the equivalence class of
u. Moreover, if we choose a sequence approximating F as F
n
= E[P
1/n
F[V
n
],
where (h
n
) is a complete basis of H, V
n
is the sigma algebra generated by
h
1
, , h
n
and P
1/n
is the OrnsteinUhlenbeck semigroup at t = 1/n, then
sup
n
sup
[h[
[
k
F
n
(w + h)[ < almost surely for any > 0, k IN, and
we can show, using an equicontinuity argument (cf. [101]) that the limit of
(F T
n
) is measurable with respect to the sigma algebra of T = I
W
+ u.
Lemma 13.7.3 For any t 0, the random variable w p
t
(w) belongs
to the class R(). Consequently, for any Hvalued random variable u, the
random variable w p
t
(w + u(w)) is welldened and it is independent of
the choice of the elements of the equivalence class of u.
Proof: In fact in [101], p.175, it has been proven that any diusion with
smooth coecients of compact support belongs to R
0
,
. In our particular
case it is easy to see that
sup
[h[
H

k
p
t
(w + h)
p
L
p
()
for any > 0 and k, n IN, where  is the Euclidean norm on M(IR
n
)H
k
and M(IR
n
) denotes the space of linear operators on IR
n
.
Lemma 13.7.4 Suppose that R
,
(H) D(H), and R
,
and
that u D(H) with [u[
H
a almost surely. Denote by T the mapping
I
W
+ u, then we have
() T = ( T) + ( T, u)
H
+ trace( T u) ,
almost surely.
222 Stochastic Analysis on Lie Groups
Proof: Let (e
i
) be a complete, orthonormal basis in H, denote by V
k
the
sigma algebra generated by e
1
, , e
k
, by
k
the orthogonal projection
of H onto the vector space generated by e
1
, , e
k
. Let u
n
be dened as
E[
n
P
1/n
u[V
n
], then [u
n
[
H
a almost surely again. From the nite dimen
sional Sobolev injection theorem one can show that the map T
n
is
continuous from D into itself and we have
( T
n
) = (I +u
n
)
T
n
(cf. [101]). For as above, it is not dicult to show the claimed identity,
beginning rst with a cylindrical then passing to the limit with the help of
the continuity of the map T
n
. To pass to the limit with respect to
n, note that we have
[ T
n
T[ sup
[h[
H
[(w + h)[
H
[u
n
(w) u(w)[
H
,
and, from the hypothesis, this sequence converges to zero in all the L
p
spaces.
For the other terms we proceed similarly.
Theorem 13.7.5 Let u be in D
q,1
(H) for some q > 1, then we have
p
t
T(w) = e
t
(
p
u)p
t
(w) ,
where e
t
(
p
u) is the solution of the ordinary dierential equation given by
e
t
= e
t
Adp
t
u
t
.
Proof: Suppose rst that u is also bounded. From Lemma 13.7.3, p
t
belongs
to R() hence the same thing is also true for the Stratonovitch integral
_
t
0
p
s
dW
s
. We can write the Stratonovitch integral as the sum of the Ito
integral of p
s
plus
1
2
_
t
0
Cp
s
ds, where C denotes the Casimir operator (cf. [25]).
Since sup
rt
[e
r
(
p
h)[ exp t[h[
H
, t p
t
T is almost surely continuous.
Moreover, it is not dicult to see that
_
t
0
Cp
s
ds is in R(). Hence we can
commute the Lebesgue integral with the composition with T. Consequently
we have, using Lemma 13.7.4,
__
t
0
p
s
dW
s
_
T =
_
t
0
p
s
T W
s
+
_
t
0
1
2
Cp
s
T ds
+
_
t
0
p
s
T u
s
ds
+
_
t
0
_
s
0
(D
r
p
s
) T D
s
u
r
drds
Anticipative Transformations 223
where W
s
denotes the Skorohod integral and D
s
is the notation for the
Lebesgue density of the Hvalued random variable . We can write this
expression simply as
__
t
0
p
s
dW
s
_
T =
_
t
0
p
s
Td
W
s
+
_
t
0
p
s
T u
s
ds ,
where d
W
s
represents the anticipative Stratonovitch integral, i.e., we add
the trace term to the divergence, whenever it is welldened. Therefore we
obtain the relation
p
t
T = e +
_
t
0
p
s
Td
W
s
+
_
t
0
p
s
T u
s
ds .
Let us now develop e
t
(
p
u)p
t
(w) using the Ito formula for anticipative pro
cesses (cf. [90]):
e
t
(
p
u)p
t
(w) = e +
_
t
0
e
s
(
p
u)p
s
(w)d
W
s
+
_
t
0
e
s
(
p
u)Adp
s
u
s
p
s
ds
= e +
_
t
0
e
s
(
p
u)p
s
(w)d
W
s
+
_
t
0
e
s
(
p
u)p
s
u
s
ds .
Hence, both p
t
T and e
t
(
p
u)p
t
satisfy the same anticipative stochastic
dierential equation with the obvious unique solution, therefore the proof is
completed for the case where u is bounded. To get rid of the boundedness
hypothesis, let (u
n
) be a sequence in D
q,1
(H) converging to u (with respect
to (q, 1)Sobolev norm) such that [u
n
[
H
2n + 1 and u
n
= u on the set
w : [u(w)[
H
n. Then from the bounded case, we have p
t
(w + u
n
(w)) =
e
t
(
p
u
n
)(w)p
t
(w) almost surely. Moreover both sides of this equality converge
in probability respectively to p
t
T and e
t
(
p
u)p
t
and the proof is completed.
The following results now follow immediately from the at case and The
orem 13.7.5: using the change of variable formula for the anticipative shifts
on the abstract Wiener spaces (cf. [94]), we can prove
Theorem 13.7.6 Suppose that u : C
G
H be a random variable such that
1. Lu
L
(,HH)
< ,
2.  Lu
op

L
()
c < 1, where c is a xed constant.
Then we have
E
[F(e(
p
u(p))p) [J
u
[] = E
[F]
for any F C
b
(C
G
), where
J
u
= det
2
(I
H
+
1
p
Lu) exp L
(
p
u)
1
2
[u[
2
.
224 Stochastic Analysis on Lie Groups
Proof: Let us denote by u
t
(w) the random variable u p which is dened on
W = C([0, 1], (). From CampbellBakerHausdor formula, we have
p(w + u
t
(w)) = e(
p(w)
u
t
(w))p(w)
(in fact here we are dealing with anticipative processes but the calculations
go as if the things were adapted thanks to the Stratonovitch integral which
denes the trajectory p). We know from [94] that
E
[F(p(w + u
t
(w)) [
u
[] = E
[F(p(w))]
where
u
= det
2
(I
H
+u
t
(w)) exp u
t
1
2
[u
t
[
2
.
To complete the proof it suces to remark that
u
t
(w) = (u p(w))
=
1
p
Lu p(w)
u
t
(w) = (u p)(w)
= L
(
p
u) p(w) .
We shall observe rst the based loop space case. We need the following
notations: if (t) is an absolutely continuous curve with values in G, we
denote by () the curve with values in ( dened by
()
t
=
_
t
0
1
s
s
ds ,
where we use, as before, the matrix notation.
Theorem 13.7.7 Suppose that : [0, 1] C
G
G be a random variable
which is absolutely continuous with respect to dt and that (0) = (1) = e,
where e denotes the unit element of G. Suppose moreover that
1. L
1
p
()
L
(,HH)
< ,
2.  L
1
p
()
op

L
()
c < 1,
3. J
S
r,1
for some r > 1, where S
r,1
is the Sobolev space on C
G
which
consists of the completion of the cylindrical functionals with respect to
the norm 
r,1
= 
L
r
()
+L
L
r
(,H)
.
Degree Type Results 225
Then we have
E
1
[F((p)p) [J
[] = E
1
[F]
for any F C
b
(C
G
), where
J
= det
2
(I
H
+
1
p
L
1
p
()) exp
_
L
()
1
2
[()[
2
_
.
Proof: It is sucient to take u =
1
p
() in the preceding theorem and then
apply the usual conditioning trick to obtain
E
[F((p)p)[J
[[p(1) = y] = E
[F[p(1) = y]
dyalmost surely. Note that by the hypothesis, there is some q > 1 such that
J
L
1+a
(),
2. J
_
I
H
+
1
p
L
1
p
()
_
h L
1+a
(), for any h H,
3. () S
r,2
(H), for some r >
1+a
a
, where S
r,2
is the Sobolev space of
Hvalued functionals as dened before.
Then we have
E
[F((p)p)J
] = E
[F]E
[J
] ,
for any F C
b
(C
G
).
The following is a consequence of Theorem 3.2 of [98]:
226
Proposition 13.7.9 Suppose that () S
q,1
(H) for some q > 1 and that
exp
_
L
1
p
() + 1/2L
1
p
()
2
2
_
L
1+b
() ,
for some b > 1. Then
E
[J
] = 1 .
Let us look at the loop space case:
Proposition 13.7.10 Let be as in Theorem 13.7.8, with (1) = e and
suppose moreover that J
S
c,1
, for some c > 1. Then
E
1
[F((p)p)J
] = E
1
[F]E
[J
] ,
for any smooth, cylindrical function F.
Proof: Let f be a nice function on G. From Theorem 9.4, we have
E
[F((p)p)f(p
1
)J
] = E
[F((p)p)f(
1
(p)p
1
)J
]
= E
[F(p)f(p
1
)]E
[J
]
hence
E
[F((p)p)J
[p
1
= y] = E
[F(p)[p
1
= y]E
[J
]
dy almost surely. Since both sides are continuous with respect to y, the
equality remains true for every y G.
Remark 13.7.11 Note that the degree of , namely E
[J
] remains the
same in both path and loop spaces.
Bibliography
[1] T. Abdellaoui and H. Heinich: Sur la distance de deux lois dans le
cas vectoriel. CRAS, Paris Serie I, Math., 319, 397400, 1994.
[2] S. Aida, T. Masuda and I. Shigekawa: Logarithmic Sobolev in
equalities and exponential integrability. Journal of Functional Anal
ysis,126, p. 83101, 1994.
[3] H. Airault and P. Malliavin: Integration geometrique sur lespace de
Wiener. Bull. Sci. Math. Vol.112, 1988.
[4] S. Albeverio and R. HeghKrohn: The energy representation of
Sobolev Lie groups. Composito Math.,36, p. 3752, 1978.
[5] R. D. Anderson and V.L. Klee, Jr.: Convex functions and upper
semicontinuous collections. Duke Math. Journal, 19, 349357, 1952.
[6] P. Appell: Memoire sur deblais et les remblais des syst`emes continus
ou discontinus. Memoires presentees par divers savants `a lAcademie
des Sciences de lInstitut de France. Paris, I. N. 29, 1208, 1887.
[7] P. Appell: Le probl`eme geometrique des deblais et des remblais.
Memorial des Sciences Mathematiques, fasc. XXVII, Paris, 1928.
[8] A. Badrikian: Derni`eres oeuvres. Annales Mathematiques Blaise Pas
cal, Numero Special, hors serie, 1996.
[9] D. Bakry and M. Emery: Diusions hypercontractives. Seminaire
de Probabilites XIX, p.179206. Lecture Notes in Math. Vol.1123.
Springer, 1985.
[10] P. J. Bickel and D. A. Freedman: Some asymptotic theory for the
bootstrap. The Annals of Statistics, Vol. 9, No. 6, 11961217, 1981.
227
228
[11] J.M. Bismut: Martingales, the Malliavin calculus and hypoelliptic
ity under general Hormander conditions. Zeit. Wahr. verw. Geb.56,
p.469505 (1981).
[12] N. Bouleau and F. Hirsch: Dirichlet Forms and Analysis on Wiener
Sace De Gruyter Studies in Math., Vol. 14, BerlinNew York, 1991.
[13] H. J. Brascamp and E. H. Lieb: On extensions of the Brunn
Minkowski and PrekopaLeindler theorems, including inequalities for
logconcave functions, and with an application to the diusion equa
tion, Journal of Functional Analysis, 22, p.366389, 1976.
[14] Y. Brenier: Polar factorization and monotone rearrangement of vec
tor valued functions. Comm. pure Appl. Math, 44, 375417, 1991.
[15] L. A. Caarelli: The regularity of mappings with a convex potential.
Jour. Amer. Math. Soc., 5, 99104, 1992.
[16] C. Castaing and M. Valadier: Convex Analysis and Measurable Mul
tifunctions. Lecture Notes in Math. 580, Springer 1977.
[17] A.B. Cruzeiro and P. Malliavin: Renormalized dierential geometry
on path space: structural equation, curvature. Jour. of Func. Anal.,
139, p. 119181, 1996.
[18] B. Dacarogna and J. Moser: On a partial dierential equation involv
ing the Jacobian determinant. Ann. Inst. Henri Poincare, Analyse
nonlineaire, 7, 126, 1990.
[19] L. Decreusefond, Y. Hu and A.S.
Ust unel: Une inegalite
dinterpolation sur lespace de Wiener. CRAS, Paris, Serie I, Vol.
317, p.10651067 (1993).
[20] L. Decreusefond and A.S.
Ust unel: On the conditional characteristic
functions of second order Wiener functionals. In Stochastic Anal
ysis and Related Fields, p.235245. Progress in Probability, vol.42.
Birkhauser, 1998.
[21] C. Dellacherie and P. A. Meyer: Probabilites et Potentiel II, Theorie
des Martingales. Hermann, 1980.
[22] J.D. Deuschel and D.W. Stroock: Large Deviations. Academic Press,
1989.
229
[23] N. Dunford and J. T. Schwarz: Linear Operators, vol. II. Interscience,
1957.
[24] O. Enchev and D. W. Stroock: Rademachers theorem for Wiener
functionals. The Annals of Probability, 21, p. 2534 (1993).
[25] H. D. Fegan: Introduction to Compact Lie Groups. Series in Pure
Mathematics, Vol. 13. World Scientic, 1991.
[26] X. Fernique: Extension du theor`eme de CameronMartin aux trans
lations aleatoires, Comptes Rendus Mathematiques, Vol. 335, Issue
1, 6568, 2002.
[27] X. Fernique: Comparaison aux mesures gaussiennes, espaces au
toreproduisants. Une application des proprietes isoperimetriques.
Preprint.
[28] D. Feyel: Transformations de HilbertRiesz. CRAS. Paris, t.310,
Serie I, p.653655 (1990).
[29] D. Feyel and A. de la Pradelle: Demonstration geometrique dune
loi de tout ou rien. C.R.A.S. Paris, t.316, p.229232 (1993).
[30] D. Feyel and A. de la Pradelle : Operateurs lineaires gaussiens.
Potential Analysis, 3, p.89105, 1994.
[31] D. Feyel and A. S.
Ust unel: Convexity and concavity on Wiener
space. Journal of Functional Analysis, 176, p. 400428, 2000.
[32] D. Feyel and A. S.
Ust unel: Transport of measures on Wiener space
and the Girsanov theorem. Comptes Rendus Mathematiques, Vol.
334, Issue 1, 10251028, 2002.
[33] M. Fukushima and H. Kaneko: On (r, p)capacities for general
Markovian semigroups in Innite Dimensional Analysis and Stochas
tic Processes, S. Albeverio (Ed.), Longman, Harlow, p. 4147 (1985).
[34] W. Gangbo and R. J. McCann: The geometry of optimal transporta
tion. Acta Mathematica, 177, 113161, 1996.
[35] B. Gaveau and P. Trauber: lIntegrale stochastique comme opera
teur de divergence dans lespace fonctionnel. J. Funct. Anal. 46,
p.230238 (1982).
230
[36] L. Gross: Logarithmic Sobolev inequalities. Amer. Jour. Math.97,
p.10611083 (1975).
[37] L. Gross: Integration and nonlinear transformation in Hilbert
space. Trans. Am. Math. Soc. 94, p.404440 (1960).
[38] L. Gross: Logarithmic Sobolev inequalities on loop groups. Jour. of
Func. Anal., 102, p. 268313, 1991.
[39] L. Gross: Uniqueness of ground states for Schrodinger operators over
loop groups. Jour. of Func. Anal.,112, p. 373441, 1993.
[40] R. Holley and D. Stroock:Logarithmic Sobolev inequalities and
stochastic Ising models. Journal of Statistical Physics, 46, p.1159
1194, 1987.
[41] Y. Hu: A unied approach to several inequalities for Gaussian mea
sures and diusion measures. Preprint.
[42] K. Ito and M. Nisio: On the convergence of sums of independent
Banach space valued random variables. J. Math., 5, Osaka, p. 3548
(1968).
[43] O. Kallenberg: On an independence criterion for multiple Wiener
integrals. The Annals of Probability, 19, p.483485, 1991.
[44] L. V. Kantorovitch: On the transfer of masses. Dokl. Acad. Nauk.
SSSR 37, 227229, 1942.
[45] N. Kazamaki: The equivalence of two conditions on weighted norm
inequalities for martingales. Proc. Intern. Symp. SDE Kyoto 1976
(ed. by K. Ito), p.141152. Kinokuniya, Tokyo, 1978.
[46] H. Korezlioglu and A. S.
Ust unel: A new class of distributions on
Wiener spaces. Procedings of Silivri Conf. on Stochastic Analysis
and Related Topics, p.106121. Lecture Notes in Math. Vol. 1444.
Springer, 1988.
[47] P. Kree: Continuite de la divergence dans les espaces de Sobolev
relatifs `a lespace de Wiener. CRAS, 296, p. 833834 (1983).
[48] I. Kubo and S. Takenaka: Calculus on Gaussian white noise. Proc.
Japan Acad. 56, 1980; 56, 1980; 57, 1981; 58, 1982.
231
[49] H.H. Kuo: Gaussian Measures on Banach Spaces. Lecture Notes in
Math. Vol. 463. Springer, 1975.
[50] H.H. Kuo: Donskers delta function as a generalized Brownian func
tional and its application. Lecture Notes in Control and Inf. Sci.
Vol.49, p.167. Springer, 1983.
[51] S. Kusuoka: The nonlinear transformation of Gaussian measures on
Banach space and its absolute continuity, I, J. Fac. Sci. Univ. Tokyo,
Sect. IA, Math., 29, p. 567598, 1982.
[52] S. Kusuoka: On the foundations of WienerRiemannian manifolds.
In Stoch. Anal., Path Integration and Dynamics, eds. K.D. Elworthy
and J.C. Zambrini, Pitman Res. Notes in Math. Longman Sci., 1989.
[53] S. Kusuoka: Analysis on Wiener spaces, I. Nonlinear maps. Journal
of Functional Analysis, 98, p. 122168 (1991).
[54] M. Ledoux: Concentration of measure and logarithmic Sobolev in
equalities. Seminaire de Probabilites XXXIII, p. 120216. Lecture
Notes in Math., Vol. 1709. Springer, 1999.
[55] P. Malliavin: Stochastic calculus of variations and hypoelliptic op
erators. In International Symp. SDE Kyoto, p.195253, Kinokuniya,
Tokyo, 1978.
[56] P. Malliavin: Stochastic Analysis. Springer Verlag, 1997.
[57] M. P. Malliavin and P. Malliavin: Integration on loop groups I.
Quasiinvariant measures. Jour. of Func. Anal., 93, p. 207237, 1990.
[58] K. Marton: Bounding
ddistance by informational divergence: a
method to prove measure concentration. Annals of Probability, 24,
no.2, 857866, 1996.
[59] R. J. McCann: Existence and uniqueness of monotone measure
preserving maps. Duke Math. Jour., 80, 309323, 1995.
[60] R. J. McCann: A convexity principle for interacting gases. Advances
in Mathematics, 128, 153179, 1997.
[61] H. P. McKean: Geometry of dierential space. The Annals of Prob
ability, Vol.1, No.2, p.197206, 1973.
232
[62] P. A. Meyer: Notes sur les processus dOrnsteinUhlenbeck.
Seminaire de Probabilites XVI, p. 95133. Lecture Motes in Math.
Vol. 920. Springer, 1982.
[63] P. A. Meyer and J. A. Yan: Distributions sur lespace de Wiener
(suite) dapr`es Kubo et Y. Yokoi. Seminaire de Probabilites XXIII,
p. 382392. Lecture Notes in Math. Vol. 1372. Springer 1989.
[64] G. Monge: Memoire sur la theorie des deblais et des remblais. His
toire de lAcademie Royale des Sciences, Paris, 1781.
[65] E. Nelson:The free Markov eld. Journal of Functional Analysis,
12, p. 17227, 1973.
[66] J. Neveu: Sur lesperance conditionnelle par rapport `a un mouvement
brownien. Ann. Inst. Henri Poincare, (B), Vol. XII, p. 105110 (1976).
[67] A. A. Novikov: On moment inequalities and identities for stochastic
integrals. Proc. Second JapanUSSR Symp. Prob. Theor., Lecture
Notes in Math., 330, p. 333339. Springer, 1973.
[68] D. Nualart: Noncausal stochastic integrals and calculus., Proceed
ings of Silivri Workshop on Stochastic Analysis, Lect. Notes in Math.
Vol.1316, Springer 1988.
[69] D. Nualart and A.S.
Ust unel: Mesures cylindriques et distributions
sur lespace de Wiener. In Proceedings of Trento meeting on SPDE,
G. Da Prato and L. Tubaro (Eds.). Lecture Notes in Math., Vol. 1390,
p.186191. Springer, 1989.
[70] D. Ocone: Malliavin calculus and stochastic integral representation
of functionals of diusion processes. Stochastics 12, p.161185 (1984).
[71] G. Pisier: Probabilistic Methods in the Geometry of Banach Spaces.
In Probability and Analysis, p.167241. Lecture Notes in Math. Vol.
1206. Springer, 1986.
[72] M. Pontier and A. S.
Ust unel: Analyse stochastique sur lespace de
LieWiener. CRAS, Paris, vol. 313, p. 313316, 1991.
[73] A. Prekopa: Logarithmic concave measures with application to
stochastic programming, Acta Sci. Math. (Szeged), 32, (1971), 301
315.
233
[74] R. Ramer: Nonlinear transformations of Gaussian measures, J.
Funct. Anal., Vol. 15, pp. 166187 (1974).
[75] T. Rockafellar: Convex Analysis. Princeton University Press, Prince
ton, 1972.
[76] I. Shigekawa: Derivatives of Wiener functionals and absolute con
tinuity of induced measures. J. Math. Kyoto Univ. 20, p.263289,
1980.
[77] B. Simon: The P()
2
Euclidean (Quantum) Field Theory. Princeton
Univ. Press. Princeton, 1974.
[78] B. Simon: Functional Integration and Quantum Physics, Academic
Press, 1979.
[79] A. V. Skorohod:On a generalization of stochastic integral. Theory
of Proba. Appl. 20, p. 219233, 1975.
[80] D. W. Stroock: Homogeneous chaos revisited. Seminaire de Proba
bilites XXI, p. 18. Lecture Notes in Math. Vol. 1247. Springer, 1987.
[81] D.W. Stroock and S.R.S. Varadhan: Multidimensional Diusion Pro
cesses. Grundlehren der math. Wiss. 233. Springer, 1979.
[82] V. N. Sudakov: Geometric problems in the theory of innite di
mensional probability distributions. Proc. Steklov Inst. Math., 141,
1178, 1979.
[83] M. Talagrand: Transportation cost for Gaussian and other product
measures. Geom. Funct. Anal., 6, 587600, 1996.
[84] E. Thomas: The LebesgueNikodym theorem for vector valued
Radon measures. Memoirs of A.M.S., 139, 1974.
[85] A. S.
Ust unel: Representation of distributions on Wiener space and
Stochastic Calculus of Variations. Jour. Funct. Analysis, 70, p. 126
139 (1987).
[86] A. S.
Ust unel: Integrabilite exponentielle de fonctionnelles de
Wiener. CRAS, Paris, Serie I, Vol. 315, p.279282 (1992).
[87] A. S.
Ust unel: Exponential tightness of Wiener functionals. In
Stochastic Analysis and Related Topics, Proceedings of Fourth Oslo
Silivri Workshop on Stochastic Analysis, T. Lindstrm, B. ksendal
234
and A. S.
Ust unel (Eds.), p. 265274 (1993). Gordon and Breach,
Stochastic Monographs, Vol. 8.
[88] A. S.
Ust unel: Some exponential moment inequalities for the Wiener
functionals. Journal of Functional Analysis, 136, p.154170, 1996.
[89] A. S.
Ust unel: Some comments on the ltering of diusions and the
Malliavin Calculus. Procedings of Silivri Conf. on Stochastic Analy
sis and Related Topics, p.247266. Lecture Notes in Math. Vol.1316.
Springer, 1988.
[90] A. S.
Ust unel: Construction du calcul stochastique sur un espace de
Wiener abstrait. CRAS, Serie I, Vol. 305, p. 279282 (1987).
[91] A. S.
Ustunel: Introduction to Analysis on Wiener Space. Lecture
Notes in Math. Vol.1610. Springer, 1995.
[92] A. S.
Ustunel:Damped logarithmic Sobolev inequalities on Wiener
space. Preprint, to appear in the 7th Silivri Proceedings.
[93] A. S.
Ust unel and M. Zakai: On independence and conditioning on
Wiener space. Ann. of Proba. 17, p.14411453 (1989).
[94] A. S.
Ust unel and M. Zakai: On the structure of independence.
Jour. Func. Analysis, 90, p.113137 (1990).
[95] A. S.
Ust unel and M. Zakai: Transformations of Wiener measure
under anticipative ows. Proba. Theory Relat. Fields, 93, p.91136
(1992).
[96] A. S.
Ust unel and M. Zakai: Applications of the degree theorem to
absolute continuity on Wiener space. Probab. Theory Relat. Fields,
95, p. 509520 (1993).
[97] A. S.
Ust unel and M. Zakai: Transformation of the Wiener measure
under noninvertible shifts. Probab. Theory Relat. Fields, 99, p. 485
500 (1994).
[98] A. S.
Ust unel and M. Zakai: Degree theorem on Wiener space.
Probab. Theory Relat. Fields, 108, p. 259279 (1997).
[99] A. S.
Ust unel and M. Zakai: Measures induced on Wiener space
by monotone shifts. Probab. Theory Relat. Fields, 105, p. 545563
(1996).
235
[100] A.S.
Ust unel and M. Zakai: On the uniform integrability of the
RadonNikodym densities for Wiener measure, J. Functional Anal.
159 (1998) 642663.
[101] A. S.
Ust unel and M. Zakai: Transformation of Measure on Wiener
Space. Springer Monographs in Math. Springer 1999.
[102] S. Watanabe: Stochastic Dierential Equations and Malliavin Calcu
lus. Tata Institute of Fundemental Research, Vol. 73. Springer, 1984.
[103] S. Watanabe: Donskers functions in the Malliavin calculus .
Stochatic Analysis, Liber Amicorum for Moshe Zakai, p.495502. E.
MayerWolf, E. Merzbach and A. Shwartz (Eds.). Academic Press,
1991.
[104] M. Zakai:On the optimal ltering of diusion processes. Z.
Wahrscheinlichkeitstheorie und verwandte Gebiete, 11, p. 230243,
1969.
[105] A. Zygmund:Trigonometric Series. Second Edition, Cambridge Uni
versity Press, 1959.
Index
cylindrical, 16
stochastic integral, 2
abstract Wiener space, 7
CampbellBakerHausdor Formula,
197
capacity, 73
CarlemanFredholm determinant, 116
closable, 16, 129
contraction, 88
coupling inequality, 100
divergence operator, 21
equilibrium potential, 74
exponential tightness, 96
ltering, 65
Girsanov Theorem, 10
GrossSobolev derivative, 17
HLipschitz, 98
HitsudaRamerSkorohod integral,
26
Hypercontractivity, 37
independence, 87
Ito formula, 3
ItoClark formula, 56
ItoWiener chaos representation, 11
lifting, 59
logarithmic Sobolev inequality, 18
local time, 76
localizing sequence, 78
logarithmic Sobolev inequality, 41,
69, 128
Mehlers formula, 25
Meyer inequalities, 29
MeyerWatanabe distributions, 47
multiplication formula , 88
multiplier, 43
number operator, 25
nondegenerate, 60
Novikov condition, 10
OrnsteinUhlenbeck operator, 25
Paul Levys Theorem, 10
positivity improving, 42, 59
quasicontinuous, 74
quasieverywhere, 73
Ramers Theorem, 115
resolution of identity, 50
Riesz transformation, 31
second quantization, 81
slim set, 73
stochastic integration, 2
Stratonovitch integral, 27
vacuum vector, 52
Wick exponentials, 11
236
237
Weyls lemma, 50
Wiener measure, 1
Zakai equation, 66
Notations
Adp(w)h, 197
B
t
, 1
det
2
(I +u), 116
C
p,k
, 73
ID, 47
ID
p,k
(X), 18
loc
, 78
u, 116
ID
t
, 47
ID
p,1
, 17
D
s
, 57
, 21
loc
, 78
Dom
p
, 17
Dom
p
(), 22
c
x
, 63
c(I(h)), 11
f
m
g, 88
(A), 80
H, 16
I(K), 2
I
p
, 12
/, 26
, 1
, 16
k
, 17
0
, 52
P
0
, 66
P
t
, 25
, 56
p.v., 29
p
, 50
, 51
o(IR
n
), 8
o, 16
o
t
(IR
d
), 59
W, 1
W
t
, 1
238