Integration by Parts and Quasi-Invariance For Heat Kernel Measures On Loop Groups

File: 580J 310301 . By:DS . Date:10:09:97 . Time:10:31 LOP8M. V8.0.
Page 01:01
Codes: 3806 Signs: 2161 . Length: 50 pic 3 pts, 212 mm
Journal of Functional AnalysisFU3103
journal of functional analysis 149, 470547 (1997)
Integration by Parts and Quasi-Invariance for Heat
Kernel Measures on Loop Groups
Bruce K. Driver*
Deparineni of Maihenaiics, 0112, Unitersiiy of California, San Diego,
1a Jolla, California 92093-0112
Received January 3, 1997; accepted January 29, 1997
Integration by parts formulas are established both for Wiener measure on the
path space of a loop group and for the heat kernel measures on the loop group. The
Wiener measure is defined to be the law of a certain loop group valued ``Brownian
motion'' and the heat kernel measures are time i, i>0, distributions of this
Brownian motion. A corollary of either of these integrations by parts formulas is
the closability of the pre-Dirichlet form considered by B. K. Driver and T. Lohrenz
[1996, J. 1unciional nal. 140, 381448]. We also show that the heat kernel
measures are quasi-invariant under right under right and left translations by finite
energy loops.

1997 Academic Press
Conienis
1. Iniroduciion. 1.1 Statement of results.
2. Noiaiion and prerequisiies.
3. Brownian noiion on loop groups. 3.1 L(q)-valued Bronian motion. 3.2. L(G)-
valued Brownian motion. 3.3. Generator of the process 2.
4. Iniegraiion by paris on ihe paih space of L(G). 4.1. Parallel translation.
4.2. Integration by parts. 4.3. Closability of the Dirichlet form.
5. The finiie dinensional appro\inaiions. 5.1. Finite dimensional integration by
parts formula. 5.2. Geometry of the finite dimensional approximations.
6. Iniegraiion by paris on ihe loop group. 6.1. Passing to the limit.
7. Quasi-intariance of ihe heai kernel neasure. 7.1. Finite dimensional
preliminearies. 7.2. Quasi-invariance for the heat kernel measure on L(G).
8. ppendi\: Fetiew of ihe Iio` iniegral in infiniie dinensions. 8.1. Continuous
Hilbert valued local martingales. 8.2. The Ito^ integral on our abstract Wiener
space. 8.3. Backwards Ito^ integrals.
article no. FU973103
470
0022-1236}97 25.00
Copyright 1997 by Academic Press
All rights of reproduction in any form reserved.
* This research was partially supported by NSF Grant DMS 92-23177 and DMS 96-12651.
E-mail: drivereuclid.ucsd.edu.
File: 580J 310302 . By:DS . Date:10:09:97 . Time:10:31 LOP8M. V8.0. Page 01:01
1. INTRODUCTION
Let G be a connected compact
1
Lie group equipped with an d
G
-
invariant inner product } , } , on the Lie algebra (q#T
e
G) of G. Let
L(G) denote the space of continuous loops in G based at the identity.
Following Malliavin [17], a L(G)-valued processes 2
i
]
i0
is con-
structed; see Theorem 3.8. In Theorem 3.10 below this processes is shown
to satisfy the martingale characterization of a Brownian motion on L(G).
Let v#Law(2
( } )
) and v
T
#Law(2
T
) so that v (Wiener measure) and v
T
(heat kernel measure) are probability measures on the path space of L(G)
and L(G), respectively. Two types of integration by parts formulas are
established.
The first integration by parts formula is for the measure v relative to a
certain class of vector fields on the path space. This version is an ``infinite''
dimensional version of the integration by parts theorem in Driver [4], see
Theorem 9.1 on p. 363.
The second is for the left-invariant first order differential operators on
L(G). This version is a infinite dimensional analogue of the fact that heat
kernel on a finite dimensional Lie group has a logarithmic derivative. Of
course, the finite dimensional version follows from the fact that the heat
kernel measure is absolutely continuous relative to the Riemannian volume
measure and the RadonNikodym density is smooth and never zero.
In Driver and Lohernz [6], a Logarithmic Sobolev inequality for cylin-
der functions was proved on a loop group with the underlying reference
measure being the heat kernel measure v
T
. The Logarithmic Sobolev
inequality as stated in [6] is really a collection of Logarithmic Sobolev
inequalities for certain finite dimensional approximations to the Loop
group with the constants being independent of the approximation. A
corollary of either of the integrations by parts formulas in this paper is that
the pre-Dirichlet form considered in [6] is closable. This elevates the
Logarithmic Sobolev inequality in [6] to a truly infinite dimensional
inequality.
We will also show that the heat kernel measure v
T
is quasi-invariant
under right and left translations by ``finite energy'' loops in L(G). This will
be done using an argument due to Cruzeiro [2] (see also Dennis Bell [1]
and Gunnar Peters [20, 21]) for proving quasi-invariance of flow from
integration integration by parts formulas.
471 LOOP GROUP QUASI-INVARIANCE
1
To avoid certain technical complications, G is assumed to be compact in the body of this
paper. However, it would be possible to extend the results in this paper to the case where G
is a Lie group of compact type, i.e., G=K_R
d
, where K is a compact Lie group and
d # 0, 1, 2, ...].
1.1. Siaieneni of Fesulis
Let G be a compact Lie group, q#T
e
G be the Lie algebra of G, and
} , } , be an d
G
invariant inner product on q. Let L=L(G) denote the
based loop group on G consisting of continuous paths g: [0, 1] G such
that g(0)=g(1)=e, where e # G is the identity element. Similarly, L(q)
will denote the continuous paths h: [0, 1] q such that h(0)=h(1)=0.
Given h # L(q), define (h, h)= if h is not absolutely continuous and
set (h, h)=
1
0
|h$(s)|
2
ds otherwise. Let
H
0
(q)#h: [0, 1] q | h(0)=h(1)=0 and (h, h)<]. (1.1)
Hence H
0
(q)/L(q) is Hilbert space with inner product (h, k)=
1
0
h$(s),
k$(s), ds. The Hilbert space H
0
(q) is to be thought of as the Lie algebra of
L(G). Using left translation, we may extend the inner product ( } , } ) to a
``Riemannian metric'' on the Cameron-Martin tangent space (TL) to L.
Explicitly,
TL#X: [0, 1] TG | 0X, # H
0
(q)], (1.2)
where (0X,)(s)#0X(s), and 0 is the Maurer Cartan form on G, i.e.,
0_,=1
g
&1
*
_ # q for all _ # T
g
G and g # G. Let : TGG denote the
projection of a tangent verctor in TG to its base point. Given g # L, the
tangent space to L at g is
T
g
L#X # TL: b X=g]/TL.
The length (X, X) of a tangent vector X # TL is now defined by
(X, X)#(0X,, 0X,)
H
0
(q)
.
In this way, L is to be thought of as an infinite dimensional ``Riemannian''
manifold.
The Levi-Civita covariant derivative (D) acting on H
0
(q), which should
be identified with left-invariant vector fields on L, is determined by
(D
k
h)(s)#
|
s
0
[k(o), h$(o)] do&s
|
1
0
[k(o), h$(o)] do, (1.3)
where h, k # H
0
(q). See Proposition 1.6 in Freed [10] and Definition 3.6
and Theorem 3.12 in Driver and Lohernz [6]. As for finite dimensional Lie
groups,
2
Eq. (1.3) uniquely determines the Levi-Civita covariant derivative
{ acting on paths in TL. Namely, if i X(i) is path in TL such that
472 BRUCE K. DRIVER
2
For the case of finite dimensional Lie groups see Section 6 in [5].
h(i)#0X(i), # H
0
(q) and g(i)#b X(i) # L are sufficiently smooth
paths, then Levi-Civita covariant derivative of X( } ) is
{X(i)}di=1
g(i)
*
h (i)+D
[ (i)
h(i)],
where [(i)#0g` (i), # H
0
(q). In particular, parallel translation (}}) along a
sufficiently smooth path i g(i) # L(G) is defined by }}
i
=1
g(i)
*
U(i),
where U solves the ordinary differential equation,
dU(i)
di
+D
[(i)
U(i)=0 with U(0)=I
H
0
(q)
. (1.4)
Let [(i)]
i0
be an L(q)-valued Brownian motion with covariance
determined by Hilbert norm ( } , } ). A more precise description of [ is that
[=[(i, s)]
i0, s # [0, 1]
is a jointly continuous two parameter q-valued
Gaussian process with mean zero and covariance given by
E[, [(i, s),B, [(t, o),]=, B,(i .t)(s .o&so),
where , B# q, i, t # [0, ), s, o # [0, 1], and s .o#min(s, o). (See
Section 3.1 for a more detailed discussion.) Following Malliavin [17],
we have the following theorem which is proved in Section 3 below, see
Theorem 3.8.
Theorem 1.1 (Brownian Motion on L). Giten g
0
# L(G), ihere is a
joinily coniinuous soluiion 2(i, s) io ihe siochasiic differeniial equaiion
2(oi, s)=1
2(i, s)
*
[(oi, s) wiih 2(0, s)=g
0
(s) \s # [0, 1], (1.5)
where for each fi\ed s # [0, 1], 2(oi, s) and [(oi, s) denoie ihe Siraionotich
differeniials of ihe processes i 2(i, s) and i [(i, s) respeciitely. (In ihe
sequel, for concreieness we will assune ihai 2 is ihe process defined in
Eq. (1.5) abote wiih g
0
(s)#e for all 0s1.)
Noiaiion 1.2. The iener space based on L=L(G) is the set of paths
(L)#o # C([0, ), L) : o(0)=e # L]. (1.6)
Similarly, let H(H
0
(q)) be the set of continuous functions h: [0, )
H
0
(q) such that h(0)=0 and there is a function h # 1
2
([0, ), di; H
0
(q))
such that h(i)=
i
0
h (t) dt for all i # [0, ). (The integral is taken to be the
Bochner integral. As in the scalar valued case, one may show that such a
function h is absolutely continuous, the derivative of h exists for almost
every i # [0, ), and dh(i)}di=h (i) a.e.) Then H(H
0
(q)) becomes a Hilbert
space with inner product defined by
(h, k)
H(H
0
(q))
#
|

0
(h (i), k (i))
H
0
(q)
di
for all h, k # H(H
0
(q)).
Definition 1.3 (Cylinder Functions). A function f : LR is said to be
a snooih cylinder funciion on L if f has the form
f ( g)=1(g(s
1
), ..., g(s
n
)) (1.7)
for some partition P=0<s
1
<s
2
< } } } <s
n
<1] of [0, 1] and some
1 # C
(G
n
). The collection of smooth cylinder functions on L will be
denoted by FC
(L). A function f : (L) R is said to be a snooih

cylinder funciion on (L) if f can be written in the form
f (o)=1(o(i
1
, s
1
), ..., o(i
n
, s
n
)) \o # (L), (1.8)
where 1 # C
(G
n
) and (i
i
, s
i
)]
n
i=1
/[0, )_(0, 1). The collection of
smooth cylinder functions on (L) will be denoted by FC
((L)).
To simplify notation in the sequel we will let
g
P
#( g(s
1
), ..., g(s
n
)) (1.9)
when g # L and P=0<s
1
<s
2
< } } } <s
n
<1]. With this notation
Eq. (1.7) may be written as f( g)=1( g
P
).
Theorem 3.10 below shows that the process 2
i
#2(i)#2(i, } ) # L(G)
deserves to be called Brownian motion on L(G) starting at g
0
. Let v
denote the law of 2(i, s)]
i0, s # [0, 1]
and v
T
denote the law of 2
T
#
2(T, } ). Also let }}
i
be ``stochastic parallel translation'' along the Brownian
motion 2(i). In analogy to the smooth case as above, }}
i
#1
2(i)
*
U(i),
where U(i) is process taking values in the unitary group of H
0
(q) which
``solves'' the Stratonovich stochastic differential equation,
oU(i)+D
[(oi)
U(i)=0, with U(0)=I
H
0
(q)
. (1.10)
See Theorem 4.1, Definition 4.2, Lemma 4.3 and the discussion at the
beginning of Section 4 for more details. The following integration by parts
theorem for (L) is completely analogous to the well known integration
by parts theorem (see for example Theorem 9.1 in [4]) for the Wiener
space (M) of compact Riemannian manifold M.
474 BRUCE K. DRIVER
Theorem 1.4 (Integration by Parts on (L)). 1or each h # H(H
0
(q))
lei X
h
denoie ihe tecior-field on (L(G)) defined by
X
h
i
(2)=}}
i
h(i)=1
2(i)
*
U(i) h(i).
Then for all snooih cylinder funciions f on (L)
E[(X
h
f )]=E[ f (2(T)) :
T
(h)], (1.11)
where
(X
h
f )(2)#
d
du }
0
f(2e
uX
h
(2)
), (1.12)
and :
T
(h) is a randon tariable described in Eq. (4.17) below.
This theorem is proved in Section 4 using the method which has been
described in Hsu [12] and Sections 2 and 3 of Driver [5] when the loop
group is replaced by a finite dimensional Riemannian manifold, see
Theorems 4.10 and 4.12 below. The next theorem describes an integration
by parts formula for the left invariant vector fields on L.
Theorem 1.5 (Integration by Parts on L). 1ei i>0, h # H
0
(q), f be a
cylinder funciion on L, and hf ( g)#(d}di)|
0
f ( ge
ih
). (So h is a firsi order lefi
intariani differeniial operaior on L.) Define H(t), t # [0, i], io be ihe solu-
iion io ihe Siraionotich siochasiic differeniial equaiion:
dH(t)+D
o[(t)
H(t)=0 wiih final daia H(i)=h. (1.13)
(The precise neaning of ihis equaiion is e\plained in Theoren 6.1 below.)
Then
E[(hf )(2
i
)]=
1
i
E
_
f (2
i
)
|
i
0
\{
I&
1
2
t Ric
=
H(t)
d[(t)
+&
, (1.14)
where

d[ in Eq. (1.14) denoies ihe backwards siochasiic differeniial and Ric
is ihe Ficci iensor on L. See Seciion 8.3 of ihe ppendi\ for a shori retiew
of ihe backwards Iio` iniegral and Definiiion 2.4 below for ihe neaning of ihe
Ficci iensor Ric.
Theorem 1.5 is a special case of Theorem 6.2 below. Theorems 1.5 and
6.2 turn out to be more delicate than Theorem 1.4. The proof is based on
Corollary 6.4 in Driver [5], which is a finite dimensional analogue to
Theorem 1.5. The basic idea of the proof is to apply Corollary 6.4 in [5]
to certain finite dimensional approximations to the loop group and then
to pass to the limit of finer and finer approximations. The necessary
geometry and estimates for the finite dimensional approximations, which
are needed to carry out this limiting procedure, are developed in Section 5.
See in particular Theorems 5.8 and 5.10.
An application of either of the above integration by parts formulas is the
closability (see Theorem 4.14 below) of the symmetric pre-Dirichlet form
on L
2
(L(G), v
T
) defined as: D(E
0
)=FC
(L) and for f # D(E

0
),
E
0
( f, f )#
|
L
&{9 f ( g)&
2
H
0
(q)
v
T
(dg).
Here {9 f ( g) denotes the gradient of f at g # L, i.e., {9 f ( g) is the unique
element in H
0
(q) such that
({9 f ( g), h)=(hf )( g) \h # H
0
(q). (1.15)
A second application of Theorem 1.5 is the quasi-invariance of the heat
kernel measure v
T
under left and right translations by ``finite energy'' loops
in L(G), see Corollary 7.7 and 7.10 in Section 7 below. The quasi-
invariance under right translations by finite energy loops will be proved
using the second integration by parts formula coupled with an argument
due to Cruzeiro [2] (see also Dennis Bell [1]) for proving quasi-
invariance of flows from integration by parts formulas. The quasi-
invariance under left translations by finite energy loops then follows
easily from the fact that v
T
is invariant under the transformation
g # L(G) g
&1
# L(G), see Proposition 7.9 below.
2. NOTATION AND PREREQUISITES
This section gathers some needed additional notation and results
from Driver and Lohrenz [6] and in Driver [5]. Let HS(H
0
(q))$
H
0
(q)*H
0
(q) be the Hilbert Schmidt operators on H
0
(q), S
0
/H
0
(q) be
an orthonormal basis for H
0
(q) and q
0
/q be an orthonormal basis of q.
For # q, let be the unique left invariant vector field on G such that
(e)=. The following theorem may be found in Lemma 3.9 and
Theorem 3.12 of [6].
Theorem 2.1. 1or k # H
0
(q), lei D
k
: H
0
(q) H
0
(q) denoie ihe operaior
defined in Eq. (1.3). e will also ihink of D as an operaior fron H
0
(q)
HS(H
0
(q)) tia (Dh) k#D
k
h. Then D is a bounded operaior such ihai
&D&
2
op
# sup
&h&
H
0
(q)
=1
_
k # S
0
&D
k
h&
2
<
and D
k
is skew adjoini operaior on H
0
(q) for all k # H
0
(q).
476 BRUCE K. DRIVER
If T is a Hilbert space, we will say that f : LT is a smooth cylinder
function if f has the form
f( g)= _
n
i=1
f
i
( g) \
i
\g # L,
where f
i
# FC
(L) and \
i
# T. The set of smooth cylinder functions on L
with values in T will be denoted by FC
(L, T). The left invariant vector

fields h for h # H
0
(q) extend naturally to operators on FC
(L, T), namely

hf( g)#
d
di }
0
f ( ge
ih
).
Definition 2.2 (Covariant Derivative). Let h # H
0
(q). Define {
h
via:
1. if f # FC
(L), set {
h
f #hf.
2. If f # FC
(L, H
0
(q)), set {
h
f #hf +D
h
f, where (D
h
f )( g)#
D
h
( f ( g)).
3. If f # FC
(L, H
0
(q)*), set {
h
f #h&D
ir
h
f, where (D
ir
h
f )( g)#
D
ir
h
( f ( g)) and D
ir
h
: H
0
(q)*H
0
(q)* is the transpose of the operator D
h
;
i.e., D
ir
h
l#l b D
h
for l # H
0
(q)*.
Definition 2.3 (Laplacian). For f # FC
(L) or f # FC
(L, H
0
(q))
or f # FC
(L, H
0
(q)*), the 1aplacian of f is defined by
Af # _
h # S
0
{
2
h
f # _
h # S
0
{
h
({
h
f ). (2.1)
The existence of the above sum is guaranteed by Proposition 4.19 of [6].
We now introduce the Ricci tensor on L, see Freed [10] and Driver and
Lohrenz [6] for more details and motivation. This tensor naturally
appears in all of the integration by parts formulas that we consider.
Definition 2.4 (Ricci Tensor). The Ricci tensor is the symmetric
quadratic form on H
0
(q) defined by
Rich, k,=&
|
1
0
|
1
0
G
0
(o, s) Kh$(s), k$(o), do ds \h, k # H
0
(q), (2.2)
where G
0
(o, s)#o.s&so and
KB, C,# _
# q
0
ad
B, ad
C,=&tr(ad
B
ad
C
),
for all B, C # q. That is K is the negative of the Killing form on q. We will
also view Ric as a bounded symmetric linear operator on H
0
(q), explicitly
(Ric h, k)=Rich, k,.
The following theorem summarizes the properties of the gradient
(defined in Eq. (1.15)), the Laplacian and the Ricci tensor that we will need
in the sequel.
Theorem 2.5. 1ei f # FC
(L), be giten as in Eq. (1.7), ihen

Af ( g)= _
# q
0
_
n
i, j=1
G
0
(s
i
, s
j
)(
(i )
( j )
1)( g(s
1
), ..., g(s
n
)), (2.3)
and
{9 f ( g)= _
# q
0
_
n
i=1
(
(i )
1)( g
P
) G
0
(s
i
, } ) , (2.4)
where for # q,
(i )
is ihe lefi intariani tecior-field on G
n
defined by
(
(i )
1)(g
1
, ..., g
n
)#
d
dc }
0
1( g
1
, ..., g
i&1
, g
i
e
c
, g
i+1
, g
n
). (2.5)
(s abote, q
0
/q is an orihonornal basis of q.)
The Bochner Wietzenbock fornula in ihis conie\i is
([A, {9 ] f )#A{9 f &{9 Af =Ric {9 f. (2.6)
If H
0
(q) is tiewed as ihe subspace of consiani funciions in FC
(L, H
0
(q)),
ihen
A
(1)
#A|
H
0
(q)
= _
k # S
0
D
2
k
.
This sun is sirongly contergeni and A
(1)
is a bounded self-adjoini operaior
on H
0
(q).
Froof. See [6] Proposition 4.19 for Equation (2.3), Theorem 4.26 for
Eq. (2.6), and Lemma 4.20 for the assertions concerning A
(1)
. Equa-
tion (2.4) is easily checked using the definition of {9 f in Eq. (1.15) and the
reproducing kernel property of G
0
, see Eq. (3.11) in [6] or the discussion
preceding Eq. (3.3) below. Q.E.D.
478 BRUCE K. DRIVER
3. BROWNIAN MOTION ON LOOP GROUPS
Let L(q)#\ # C([0, 1] q | $0)=\(1)=0] be the continuous based
loops in q. It is well known that (H
0
(q), L(q)) is an abstract Wiener space
as introduced by Gross in [11]. As usual in the abstract Wiener space
setting, we have L(q)*/H
0
(q)*$H
0
(q)/L(q). Let us recall the explicit
description of L(q)* in H
0
(q)*.
To this end we will say, for h # H
0
(q), that h$ is of bounded tariaiion if
there is a right continuous function (z) of bounded variation such that
h$(s)=z(s) a.e.. Let
H
B1
0
#h # H
0
| h$ is of bounded variation].
Now suppose that k # H
0
(q), then by an integration by parts (see for
example Theorem 3.30 of [9])
(h, k)=
|
1
0
h$(s), k$(s), ds=
|
1
0
z(s), dk(s),=&
|
1
0
k(s), dz(s),. (3.1)
In the future we will abuse notation and write
1
0
k(s), dh$(s), for
1
0
k(s),
dz(s),.
Lemma 3.1. 1or each h # H
B1
0
and \ # L(q) lei z
h
($#&
1
0
$s),
dh$(s),. Then ihe nap h # H
B1
0
z
h
# L(q)* is an isonorphisn. Moreoter
z
h
(k)=(h, k) for all k # H
0
(q).
Froof. The last assertion of the Lemma clearly follows from (3.1). Now
suppose that z
h
#0 then 0=z
h
(k)=(h, k) for all k # H
0
(q) which implies
that h=0 in H
0
(q). Therefore h z
h
is injective.
Since L(q)*/H
0
(q)*, for z # L(q)* there exists h # H
0
(q) such that
z(k)=(h, k) for all k # H
0
(q). Since H
0
(q) is dense in L(q), if we can show
that h # H
B1
0
, it will follow that z=z
h
. Hence the map h z
h
is surjective.
Noting that L(q) is a closed subspace of C([0, 1], q), the Hahn-Banach
theorem asserts that z has an extension (z` ) to a bounded linear functional
on C([0, 1], q). By the Riesz theorem (e.g., Theorem 7.17 of [9]) there is
a q-valued measure such that
z`($=
|
1
0
$s), (ds), \\ # C([0, 1], q).
Define z(s)#([0, s]) # q for s # [0, 1]. Then z is of bounded variation
and we have
z($=
|
1
0
$s), dz(s), \\ # L(q).
Restricting this last identity to k # H
0
(q) and then doing an integration by
parts shows that
|
1
0
h$(s), k$(s), ds=(h, k)=z(k)=
|
1
0
k(s), dz(s),
=&
|
1
0
z(s), k$(s), ds.
Since k$ # 1
2
([0, 1], q) | k # H
0
(q)] is the orthogonal compliment of
the constant functions in 1
2
([0, 1], q) the above equation implies that
h$(s)=z(s)&
1
0
z(s) ds a.e. This proves that h # H
B1
0
. Q.E.D.
Noiaiion 3.2. In the sequel, we will write (h, $ instead of z
h
(\) when
h # H
B1
0
and \ # L(q)*.
3.1. L(q)-1alued Brownian Moiion
Let G denote the smallest o-field on L such that all of the smooth
cylinder functions in FC
(L) are measurable. For the sequel, fix a filtered

probability space (W, F
i
]
i0
, F, F) and a L(q)-valued process [(i)]
i>0
on W with the following properties:
1. F
i
/F
i$
/F for all 0ii$.
2. F
i
is right continuous, i.e., F
i
=F
i+
#
c>0
F
i+c
,
3. F
0
contains all of the null sets of F,
4. for all c # W, the map i # [0, ) [(i)(c) # L(q) is continuous.
5. [(i) is F
i
}G-measurable for all i0, and
6. [(i)]
i0
is a mean-zero Gaussian process with convariance,
E[(h, [(i))(k, [(t))]=i .t(h, k), (3.2)
where h, k # H
B1
0
and i, t # [0, ).
We say that such a process [(i)]
i0
is a L(q)-valued Brownian noiion.
Fenark 3.3. The existence of an L(q)-valued Brownian motion [(i) is
well known. In fact, it is known more generally that Brownian motions
exist on arbitrary abstract Wiener spaces. One possible construction is to
first use Kolomogorov's existence theorem to construct a Brownian motion
[(i) satisfying all of the properties above except for the continuity. Then by
Fernique's theorem (e.g., Theorem 3.1 of Kuo [15]) and scaling it can be
seen that Kolomogorov's continuity criteria may be applied to yield a
version [(i) of [(i) which is z holder continuous for all z # (0, 1}2).
480 BRUCE K. DRIVER
Suppose that h(s)#G
0
(s, u) where # q and as above G
0
(s, u)=
s .u&su. Then for \ # L(q)
(h, \)=&
|
1
0
\(s), dh$(s),=&
|
1
0
\(s), , d
s
1
su
&u]
=&
|
1
0
\(s), ,(&o
u
(ds))=\(u), ,,
where o
u
is the Dirac measure concentrated at u. Let k(s)#G
0
(s, t) B.
Notice that (h, k)=, B, G
0
(u, t).
Write [(i, s) for the q-valued random variable determined by [(i, s)(c)
#[(i)(c)(s). Using the previous paragraph and (3.2), for all , B # q and
i, t # [0, ) with it,
E[, [(i, u),B, [(t, t),]
=E[, [(t, u),+, [(i, t)&[(t, u),]B, [(t, t),]
=E[, [(t, u),B, [(t, t),]
=, B, tG
0
(u, t)
=, B,(i .t) G
0
(u, t). (3.3)
For each h # H
0
(q) and i0, let
(h, [(i))#1
2
- lim
n
(h
n
, [(i)),
where h
n
]/H
B1
0
and h
n
h in H
0
(q) as n . Then it is easily checked
that i (h, [(i)) is a (not necessarily continuous) Brownian motion
with variance (h, h). Let [
h
(i) denote a continuous version of (h, [(i)).
Such a version exists by Kolomogorov's continuity criteria. Then [
h
is a
Brownian motion with variance (h, h) on the filtered probability space
(W, F
i
], F, F). The next Lemma records the mutual quadratic variation
O[
h
, [
k
o for h, k # H
0
(q).
Lemma 3.4. 1or each h, k # H
0
(q),
O[
h
, [
k
o
i
=(h, k) i a.s.
Froof. Decompose h as h=zk+j, where j = k and (h, k)=z(k, k).
Then [
h
is indistinguishable from z[
k
+[
j
. Since the pair [
j
, [
k
] is a
Gaussian process and
E( [
j
(i) [
k
(t))=i .t( j, k)=0,
it follows that [
j
and [
k
are independent Brownian motions. Hence
O[
h
, [
k
o
i
=zO[
k
, [
k
o
i
+O[
j
, [
k
o
i
=z(k, k) i+0=(h, k) i. Q.E.D.
Corollary 3.5. 1ei u, t # [0, 1] and , B # q, ihen
O, [( } , u),, B, [( } , t),o
i
=i, B, G
0
(u, t). (3.4)
Froof. Take h=G
0
( } , u) and k=G
0
( } , u) B in Lemma 3.4 and use
(h, k)=, B, G
0
(u, t). Q.E.D.
3.2. L(G)-1alued Brownian Moiion
Noiaiion 3.6. Given an L-valued process 2(i)]
i0
on W, let
2(i, s)(c)#2(i)(c)(s). In this way we will identify L-valued processes on
W with two parameter G-valued processes.
In preparation for proving the existence of a ``Brownian Motion'' on
L(G), we will introduce a metric on G.
Definition 3.7. The distance metric d: G_GG is defined by
d( g, h)=inf
|
1
0
|o$(s)| ds,
where the infimum is taken over all C
1
-paths o in G such that o(0)=g and
o(1)=h. Also set
| g| =* d( g, e) \g # G.
Notice that
d(\g, \h)=d( g, h)
for all g, h, \ # G. Indeed, if o is a curve joining g to h, then \o( } ) is a curve
joining \g to \h which has the same length as o. Set | g| #d( g, e)=d(e, g),
then because of the above displayed equation,
d( g, h)=| g
&1
h| =|h
&1
g|.
Setting h=e in this equation shows that | g| =| g
&1
| for all g # G.
482 BRUCE K. DRIVER
The next theorem is stated in Malliavin [17]. For the readers con-
venience we will supply a proof.
Theorem 3.8 (Malliavin). Suppose ihai G is a conpaci 1ie group and
} , } , is an d
G
intariani inner produci on q. There e\isis a coniinuous
adapied process 2(i)]
i0
on ihe filiered probabiliiy space (W, F
i
]
i0
,
F, F) such ihai for each s # [0, 1], 2( } , s) soltes ihe siochasiic differeniial
equaiion:
2(oi, s)=2(i, s) [(oi, s) wiih 2(0, s)=g
0
(s). (3.5)
More precisely, Eq. (3.5) is shorihand noiaiion for ihe siochasiic differeniial
equaiion
2(oi, s)= _
# q
0
(2(i, s)) [
(oi, s) wiih 2(0, s)=g

0
(s), (3.6)
where q
0
/q is an orihonornal basis of q, is ihe lefi intariani tecior field
on G such ihai (e)=, and [
(i, s)#, [(i, s),. Here [
(oi, s) denoies
ihe Siraionotich differeniial of ihe process i [
(i, s). In ihe sequel, we will

use ``o'' for Siraionotich differeniial and ``d'' for ihe differeniial of a seni-
nariingale.
Before starting the proof of this theorem, let us recall the following easy
lemma.
Lemma 3.9. 1ei M and N be iwo finiie dinensional nanifolds, X
i
]
n
i=1
and Y
i
]
n
i=1
be a colleciion of snooih tecior fields on M and N respeciitely,
and b(i)=(b
1
(i), b
2
(i), ..., b
n
(i)) ( for i0) be an R
n
-talued coniinuous seni-
nariingale. (s usual b is defined on a filiered probabiliiy space saiisfying
ihe usual hypoihesis.) Suppose ihai \ and y are seni-nariingales on M and
N which saiisfy ihe siochasiic differeniial equaiions,
o\= _
n
i=1
X
i
(\) ob
i
and oy= _
n
i=1
Y
i
( y) ob
i
,
respeciitely. Then (\, y) is a M_N-talued seni-nariingale saiisfying ihe
siochasiic differeniial equaiion
o(\, y)= _
n
i=1
(X
i
(\, y)+Y
i
(\, y)) ob
i
,
where X
i
and Y
i
are ihe snooih tecior fields on M_N defined by X
i
(\, y)#
(X
i
(\), 0
y
) and Y
i
(\, y)#(0
\
, Y
i
( y)). Here 0
\
and 0
y
denoie ihe :ero
iangeni teciors in T
\
M and T
y
M respeciitely.
Froof. Let : be the M_N-valued semi-martingale which solves the
stochastic differential equation,
o:= _
n
i=1
(X
i
(:)+Y
i
(:)) ob
i
with :(0)=(\(0), y(0)). (3.7)
To finish the proof it suffices to show that :=(\, y).
Define
1
: M_NM and
2
: M_NN to be projections onto the
first and second factor of M_N respectively and :
i
#
i
(:). If f # C
(M)
then f b
1
# C
(M_N). So by definition of : solving Eq. (3.7), we have

o( f b
1
(:))= _
n
i=1
[(X
i
(:)+Y
i
(:))( f b
1
)] ob
i
.
Now X
i
(:)( f b
1
)=(
*
X
i
(:)) f =X
i
(:
i
) f and similarly Y
i
(:)( f b
1
)=
(
*
Y
i
(:)) f =0. Therefore, the last displayed equation may be written as
o( f (:
1
))= _
n
i=1
[X
i
(:
1
) f ] ob
i
,
i.e., :
1
is a semi-martingale on M solving the stochastic differential
equation,
o:
1
= _
n
i=1
X
i
(:
1
) ob
i
with :
1
(0)=\(0).
Since this is the same equation solved by \ it follows by uniqueness of
solutions that :
1
=\. The same argument also shows that :
2
=y. That is
(\, y)=:. Q.E.D.
Froof of Theoren 3.8. For the purposes of the proof we will adopt the
following notation. If f # C
(G), let f $ # C
(G, q*) be defined by

f $( g),=df ( g),=
d
di }
0
f ( ge
i
)
and let
A
G
f # _
# q
0
2
f.
484 BRUCE K. DRIVER
We also start with the special case where g
0
(s)#e for all s # [0, 1]. For
each s # [0, 1], let 2
0
( } , s) be a solution to the stochastic differential equa-
tion in Eq. (3.5) (or equivalently Eq. (3.6)) with initial condition 2
0
(0, s)=
g
0
(s)#e # G. (For the existence of solutions to this equation see for
example [8, 13, 14].) In this way we construct a G-valued two parameter
process 2
0
(i, s). Our immediate goal is to show that there exists a con-
tinuous version (2(i, s)) of this process.
For the moment, fix t # [0, ) and let u(i)#2
0
(t, s)
&1
2
0
(i, s) for it.
Then u(t)=e # G and u solves the stochastic differential equation,
ou(i)=1
2
0
(t, s)
&1
*
_
# q
0
(2
0
(i, s)) [
(oi, s)
= _
# q
0
(2
0
(t, s)
&1
2
0
(i, s)) [
(oi, s)
= _
# q
0
(u(i)) [
(oi, s),
wherein we have used the left invariance of along with Eq. (3.6). So if
f # C
(G) such that f (e)=0, then using Corollary 3.5 we find for all it
that
f (u(i))=f (u(t))+
|
i
t
_
# q
0
(f )(u(r)) [
(or, s)
=
|
i
t
_
# q
0
(f )(u(r)) [
(dr, s)+
1
2
|
i
t
_
# q
0
(
2
f )(u(r)) G
0
(s, s) dr
=
|
i
t
_
# q
0
(f )(u(r)) [
(dr, s)+
G
0
(s, s)
2
|
i
t
(Af )(u(r)) dr.
For any p # [2, ) and it, it follows from Burkholder's inequality that
E | f (u(i))|
p
C
p
( f ) E
{
|
i
t
| f $(u(r))|
2
G
0
(s, s) dr
=
p}2
+C
p
( f ) E
}
|
i
t
A
G
f (u(r)) G
0
(s, s) dr
}
p
C
p
( f )(i&t)
p}2
+(i&t)
p
], (3.8)
where C
p
( f ) denotes a constant depending only on p and bounds on f $ and
A
G
f. Let f
i
]
n
i=1
/C
(G) be a suitable collection of functions such that

f
i
(e)=0 for all i and
| g| :=d( g, e) _
n
i=1
| f
i
( g)| \g # G,
cf., the Whitney imbedding theorem. This equation and Eq. (3.8) implies,
for all it and all s # [0, 1], that
E[d(2
0
(i, s), 2
0
(t, s))]
p
=E |2
0
(t, s)
&1
2
0
(i, s)|
p
=E |u(i)|
p
C
p
(i&t)
p}2
+(i&t)
p
],
where C
p
#n
( p&1)
_
n
i=1
C
p
( f
i
). So we have proven for all s # [0, 1] and
i, t # [0, ) that
Ed(2
0
(t, s), 2
0
(i, s))]
p
C
p
|i&t|
p}2
+|i&t|
p
]. (3.9)
Now fix s, o # [0, 1], and set u(i)#2
0
(i, s) 2
0
(i, o)
&1
. In the case that
G is a matrix group, we may compute du(i) to find
du(i)=2
0
(i, s)[(oi, s)&[(oi, o)] 2
0
(i, o)
&1
=u(i) B(oi),
where
B(i)=
|
i
0
d
2
0
(t, o)
[(ot, s)&[(ot, o)]. (3.10)
To prove the analogous formula in the general case we will use
Lemma 3.9. To this end let f # C
(G) and 1( g, k)#f ( gk

&1
) so that
f(u(i))=1(2
0
(i, s), 2
0
(i, o)). For # q, we have
d
dc }
0
1( ge
c
, k)=
d
dc }
0
f ( gk
&1
ke
c
k
&1
)=
d
dc }
0
f ( gk
&1
e
cd
k
)
=((d
k
) f )( gk
&1
)
and, by essentially the same computation,
d
dc }
0
1( g, ke
c
)=&((d
k
) f )( gk
&1
).
486 BRUCE K. DRIVER
therefore,
d[ f (u(i))]= _
# q
0
((d
2
0
(i, o)
) f )(u(i)) } ( [
(oi, s)&[
(oi, o))
= _
, C# q
0
d
2
0
(i, o)
, C,(Cf )(u(i)) } [(oi, s)&[(oi, o), ,
= _
C # q
0
(Cf )(u(i)) } d
2
0
(i, o)
( [(oi, s)&[(oi, o)), C,
= _
C # q
0
(Cf )(u(i)) oB
C
(i),
where B is the process defined in Eq. (3.10).
We now claim that B may be expressed as
B(i)=
|
i
0
d
2
0
(t, o)
[(dt, s)&[(dt, o)]. (3.11)
The main point here is that
d
i
(d
2
0
(i, o)
)= _
# q
0
(d
( } )
)(2
0
(i, o)) o[
(i, o)
= _
# q
0
d
2
0
(i, o)
ad
o[
(i, o).
Using this equation we find
B(i)=
|
i
0
d
2
0
(t, o)
[(dt, s)&[(dt, o)]
+
1
2
|
i
0
_
# q
0
d
2
0
(i, o)
ad
[(dt, s)&[(dt, o)] [
(dt, o)
from which Eq. (3.11) follows because
_
# q
0
ad
[(dt, s)&[(dt, o)] [
(dt, o)
= _
, C # q
0
ad
C } d
t
O[
C
( } , s)&[
C
( } , o)], [
( } , o)o
t
= _
# q
0
ad
} (G
0
(s, o)&G
0
(o, o)) dt=0.
Since d
g
: q q acts isometrically for all g # G, we conclude from
Eq. (3.11) that B(i) is again a Brownian motion with the same covariance
as the Brownian motion i [(i, s)&[(i, o). This covariance is
E( [(i, s)&[(i, o)) ( [(i, s)&[(i, o))]#i1(s, o) I,
where I#_
# q
0
# qq and 1(s, o)#G
0
(s, s)+G
0
(o, o)&2G
0
(s, o).
Notice for each fixed o # [0, 1] that s 1(s, o) is a piecewise C
1
-function
such that 1(o, o)=0 and
|1(s, o)}s| =|1&2s&2(1
so
&o)| 4.
Therefore
1(s, o)4 |s&o|, \s, o # [0, 1].
By computations similar to those which lead to Eq. (3.8), if f # C
(G)
and f (e)=0, then
E | f (u(i))|
p
C
p
E
{
|
i
0
| f $(u(t))|
2
1(s, o) dt
=
p}2
+C
p
E
}
|
i
0
A
G
f (u(t)) 1(s, o) dt
}
p
C
p
( f )(i1(s, o))
p}2
+(i1(s, o))
p
],
where C
p
( f ) denotes a constant depending only on p and the bounds on
f $ and A
G
f as above. As in the proof of Eq. (3.9), we may conclude
Ed(2
0
(i, s), 2
0
(i, o))]
p
=E |2
0
(i, s)
&1
2
0
(i, o)|
p
=E |u(i)|
p
C
p
(i1(s, o))
p}2
+(i1(s, o))
p
]
C
p
(i |s&o| )
p}2
+(i |s&o| )
p
], (3.12)
where C
p
is a constant only depending on p and the compact group G.
The triangle inequality and the estimates in (3.9) and (3.12) yields
Ed(2
0
(i, s), 2
0
(t, o))]
p
C
p
i |s&o|
p}2
+(i |s&o| )
p
+|i&t|
p}2
+|i&t|
p
],
where C
p
is a constant dependig only on p and the compact group G.
Consequently, for each T # (0, ), there is a constant C
p
(T) such that
Ed(2
0
(i, s), 2
0
(t, o))]
p
C
p
(T)|s&o|
p}2
+|i&t|
p}2
],
488 BRUCE K. DRIVER
for all s, o # [0, 1] and i, t # [0, T]. Hence, by Kolmogorov's continuity
criteria (see for example Theorem 1.4.4 of Kunita [14] and Theorem 53,
Chapter 4 of Protter [22]) there is a continuous version (2(i, s)) of
2
0
(i, s) such that for all [ # (0, 1}2) there exists a positive random variable
(K
[
) on W such that
d(2(i, s), 2(t, o))K
[
|i&t|
[
+|s&o|
[
] a.s. (3.13)
Furthermore, EK
p
[
< for all p # (1, ). Since, for each s # [0, 1], 2( } , s)
is a version of 2
0
( } , s), it follows that 2 satisfies all the hypothesis of the
theorem when g
0
(s)=e.
For the general case let 2 be as in the special case just proved and define
2(i, s)#g
0
(s) 2(i, s). Then 2(i)]
i0
is a continuous adapted L-valued
process satisfying the differential equation in Eq. (3.5). Q.E.D.
3.3. Generaior of ihe Frocess 2
The next theorem shows that the L(G)-valued process 2(i) constructed
above satisfies the standard martingale criteria of a Brownian motion. For
this reason 2(i) deserves to be called an L(G)-valued Brownian motion.
Theorem 3.10. 1ei P=0<s
1
<s
2
< } } } <s
n
<1] be a pariiiion on
[0, 1], G
P
be ihe sei of funciions fron P io G, 2(i) be ihe L-talued process
in Theoren 3.8 and f : [0, T]_LR be a funciion of ihe forn f (i, g)=
1(i, g
P
). ssune ihai 1: [0, T]_G
P
R is a coniinuous funciion
saiisfying:
1. 1|
(0, T)_G
P is snooih and
2. The deritaiites of 1|
(0, T)_G
P up io second order e\iend io
coniinuous funciions on [0, T]_G
P
.
Then
f (i, 2(i))=f (0, 2(0))+M
i
+
|
i
0
\\

t
+
1
2
A
+
f(t, } )
+
(2(t)) dt,
where M
i
is ihe nariingale:
M
i
= _
n
i=1
_
# q
0
|
i
0
(
(i )
1)(t, 2
P
(t))[(dt, s
i
), ,
=
|
i
0
(({9 f )(t, 2(t)), [(dt)).
Froof. Set
2
P
(i)#(2(i)
P
)=(2(i, s
1
), ..., 2(i, s
n
)).
Then, by Lemma 3.9, 2
P
(i) # G
n
satisfies the stochastic differential
equation
o2
P
(i)= _
n
i=1
_
# q
0
(i )
(2
P
(i))[(oi, s
i
), ,, (3.14)
where for # q,
(i )
is defined in Eq. (2.5).
Equations (3.4) and (2.3) allows us to compute, using Ito^ 's lemma in
finite dimensions, the differential of f (i, 2(i)) as
d[ f (i, 2(i))]
=f(i))}i+ _
n
i=1
_
# q
0
(
(i )
1)(i, 2
P
(i))[(oi, s
i
), ,
=f(i, 2(i))}i+ _
n
i=1
_
# q
0
(
(i )
1)(i, 2
P
(i))[(di, s
i
), ,
+
1
2
_
n
i=1
_
# q
0
_
n
j=1
_
B # q
0
(B
( j )
(i )
1)(i, 2
P
(i))[(di, s
i
), ,[(di, s
j
), B,
=dM
i
+f (i, 2(i))}i+
1
2
_
n
i=1
_
# q
0
_
n
j=1
(
( j )
(i )
1)(i, 2
P
(i)) G
0
(s
i
, s
j
) di
=dM
i
+f (i, 2(i))}i+
1
2
(Af )(i, 2(i)) di.
In the above computation we have used [(i, s), ,=[
G
0
(s, } )
(i), see the
proof of Corollary 3.5, and Eq. (2.4) to conclude that
_
n
i=1
_
# q
0
(
(i )
1)(i, 2
P
(i))[(di, s
i
), ,=({9 f (i, 2(i)), d[(i)). Q.E.D.
Noiaiion 3.11. For definiteness, in the remainder of this paper let
2(i)]
i0
denote the L(G) valued process constructed in Theorem 3.8
with g
0
(s)#e for all s # [0, 1].
490 BRUCE K. DRIVER
By the proof of Theorem 3.10, if 1 # C
(G
P
) then
d1(2
P
(i))= _
n
i=1
_
# q
0
(
(i )
1)(2
P
(i))[(di, s
i
), ,
+
1
2
_
# q
0
_
n
i, j=1
(
( j )
(i )
1)(2
P
(i)) G
0
(s
i
, s
j
) di.
Thus 2
P
(i) is a diffusion process on G
P
with generator
A
P
#
1
2
_
# q
0
_
n
i, j=1
G
0
(s
i
, s
j
)
( j )
(i )
. (3.15)
Lemma 3.12. 1ei P=0<s
1
<s
2
< } } } <s
n
<1] be a pariiiion of
[0, 1], ihen ihe nairi\ G
0
(s
i
, s
j
)]
n
i, j=1
is posiiite definiie. In pariicular A
P
is a second order ellipiic differeniial operaior on G
P
.
Froof. Let H
0
(R) be the set of absolutely continuous functions
l : [0, 1] R such that l(1)=l(0)=0 and (l, l )#
1
0
(l $(s))
2
ds<. Choose
an orthonormal basis h of H
0
(R). Then by Lemma 3.8 of [6], G
0
(s, o)=
_
l # h
l(s) l(o) with the sum being absolutely convergent therefore, if
_=(_
1
, _
2
, ..., _
n
) # R
n
,
_
n&1
i, j=1
G
0
(s
i
, s
j
) _
i
_
j
= _
n&1
i, j=1
_
l # h
l(s
i
) l(s
j
) _
i
_
j
= _
l # h
_
_
n&1
i=1
l(s
i
) _
i
&
2
0
with equality iff _
n&1
i=1
l(s
i
) _
i
=0 for all l # h. Since h is an orthonormal
basis for H
0
(R) and the map l # H
0
(R) l(s) # R is a continuous linear
functional, the condition _
n&1
i=1
l(s
i
) _
i
=0 for all l # h, is equivalent to
_
n&1
i=1
l(s
i
) _
i
=0 for all l # H
0
(R). Choosing l # H
0
(R) such that l is nonzero
on exactly one of the partition points in P allows us to conclude that
_
n&1
i, j=1
G
0
(s
i
, s
j
) _
i
_
j
=0 iff _=0. Q.E.D.
Before ending this section let us record a slight extension of
Theorem 3.10 which will be needed in the sequel. (The proof will be left to
the reader.) We will first need the following definition.
Definition 3.13. Given a Hilbert space T, let {
0
denote the ``flat''
covariant derivative on FC
(L, T) defined by {
0
h
f #hf for all f #
FC
(L, T) and h # H
0
(q). Let A
0
denote the ``flat'' Laplacian on
FC
(L, T) defined by
A
0
f # _
h # S
0
({
0
h
)
2
f = _
h # S
0
h
2
f.
Theorem 3.14. Suppose ihai T is a Hilberi space and f: [0, T]_
L(G) T is a funciion of ihe forn f =_
n
i=1
f
i
_
i
where f
i
: [0, T]_
L(G) T are funciions saiisfying ihe assunpiions in Theoren 3.10 and
_
i
# T for i=1, 2, ..., n. Then
d( f (i, 2(i)))=({
0
d[(i)
f )(2(i))+
\\

i
+
1
2
A
0
+
f
+
(i, 2(i)) di. (3.16)
4. INTEGRATION BY PARTS ON THE PATH SPACE OF L(G)
In this section, integration by parts formulas on the path space of the
loop group are derived. As a corollary we will show that the pre-Dirichlet
form introduced in Driver and Lohrenz [6] is closable. Before doing this
however, it is first necessary to discuss parallel translation (}}
i
) along the
Brownian motion 2(i). Pretending for the moment that L(G) is a finite
dimensional Lie group, }}
i
k
0
=1
2(i)
*
k(i), where k(i) is the solution to the
stochastic differential equation,
dk(i)+D
o[(i)
k(i)=0 with k(0)=k
0
, (4.1)
where o[ denotes the Stratonovich differential of [. For motivation, see
Theorem 6.3 in [5]. Formally, writting ``[=_
h # S
0
[
h
h,''
D
o[
k=D
d[
k+
1
2
D
d[
dk=D
d[
k&
1
2
D
d[
D
d[
k
=D
d[
k&
1
2
_
h # S
0
D
h
D
h
k di
=D
d[
k&
1
2
A
(1)
k di.
Hence we should interpret Eq. (4.1) as the Ito^ equation,
dk(i)=&D
d[(i)
k(i)+
1
2
A
(1)
k(i) di with k(0)=k
0
. (4.2)
See the Appendix (Section 8) for a review of the Ito^ integral in this context.
4.1. Farallel Translaiion
Theorem 4.1 (Parallel Translation). 1or each k
0
# H
0
(q), ihere e\isis a
unique soluiion io Eq. (4.2). Moreoter if h
0
# H
0
(q) and h is ihe soluiion io
Eq. (4.2) wiih k and k
0
replaced by h and h
0
respeciitely, ihen for all i0,
(h(i), k(i))=(h
0
, k
0
) alonosi surely.
492 BRUCE K. DRIVER
Froof. Eq. (4.2) may be solved using the usual Picard iterates scheme.
Recall that A
(1)
=A|
H
0
(q)
is a bounded operator and notice that
E
"
|
i
0
D
[(dt)
k(t)
"
2
=E
|
i
0
_
h # S
0
&D
h
k&
2
dt&D&
2
op
E
|
i
0
&k&
2
dt
when k(i)]
i0
is a continuous adapted H
0
(q)-valued process such that
E
i
0
&k&
2
dt<.
To simplify notation, if f is a possibly random function on [0, ) taking
values in a normed space T, let f *(i)#sup& f (t)&: 0ti]. If
k
n
(i)#k
0
&
|
i
0
D
[(dt)
k
n&1
(t)+
1
2
|
i
0
A
(1)
k
n&1
(t) dt, (4.3)
then
k
n+1
(i)&k
n
(i)=&
|
i
0
D
[(dt)
(k
n
(t)&k
n&1
(t))
+
1
2
|
i
0
A
(1)
(k
n
(t)&k
n&1
(t)) dt.
Hence using Burkholder's inequality
E(k
n+1
&k
n
)*
2
(i)2E
\
|
}
0
D
[(dt)
(k
n
(t)&k
n&1
(t))
+
*
2
(i)
+
i
2
&A
(1)
&
2
op
|
i
0
E(k
n
&k
n&1
)*
2
(t) dt
4 &D&
2
op
|
i
0
E &k
n
(t)&k
n&1
(t)&
2
dt
+
i
2
&A
(1)
&
2
op
|
i
0
E(k
n
&k
n&1
)*
2
(t) dt
4 &D&
2
op
|
i
0
E(k
n
&k
n&1
)*
2
(t) dt
+
i
2
&A
(1)
&
2
op
|
i
0
E(k
n
&k
n&1
)*
2
(t) dt
=
\
4 &D&
2
op
+
i
2
&A
(1)
&
2
op
+
|
i
0
E(k
n
&k
n&1
)*
2
(t) dt.
Fix T > 0 and let K
T
# (4 &D&
2
op
+ (T}2) &A
(1)
&
2
op
), and f
n
(i) =
E(k
n
&k
n&1
)*
2
(i). Then for 0iT,
f
n+1
(i)K
T
|
i
0
f
n
(t) dt
which implies after iteration that
f
n
(i)
(K
T
i)
n
n!
f
0
*(T).
Thus _
n=0
f
n
*(T)f
0
*(T) } exp(K
T
T), from which we learn that k
n
(i) is
1
2
-uniformly convergent for i in compact subsets of [0, ) to a continuous
process, say k(i). Passing to the limit in Eq. (4.3) shows that k solves,
k(i)#k
0
&
|
i
0
D
[(dt)
k(t)+
1
2
|
i
0
A
(1)
k(t) dt. (4.4)
Let k be as above and h(i) be a solution to Eq. (4.2) (or equivalently
(4.4)) with h(0)=h
0
and set 1
i
l#&D
l
h(i) and G
i
l#&D
l
k(i) for all
l # H
0
(q). Then
dh(i)=1
i
d[(i)+
1
2
A
(1)
h(i) di
and
dk(i)=G
i
d[(i)+
1
2
A
(1)
k(i) di.
Therefore by Ito^ 's Lemma, see Theorem 8.5 of the Appendix,
d(h(i), k(i))=(1
i
*k(i), d[(i))+
1
2
(A
(1)
h(i) di, k(i))
+(G
i
*h(i), d[(i))+
1
2
(h(i), A
(1)
k(i) di)
+(1
i
, G
i
)
HS
di,
where, for all u, t # HS(H
0
(q))$H
0
(q)*H
0
(q) (the Hilbert Schmidt
operators on H
0
(q)) and
(u, t)
HS
# _
h # S
0
(uh, th). (4.5)
494 BRUCE K. DRIVER
For all l # H
0
(q),
(1
i
*k(i), l )+(G
i
*h(i), l )=(k(i), 1
i
l )+(h(i), G
i
l )
=(k(i), &D
l
(i))+(h(i), &D
l
k(i))=0,
since D
l
is skew adjoint. Also
(1
i
, G
i
)
HS
= _
l # S
0
(1
i
l, G
i
l )= _
l # S
0
(D
l
h(i), D
l
k(i))
= _
l # S
0
(h(i), &D
2
l
k(i))=&(h(i), A
(1)
k(i)).
Combining the last four equations and using A
(1)
is self-adjoint shows that
d(h(i), k(i))=0. Q.E.D.
Definition 4.2. Let O(H
0
(q)) be the group of unitary operators on
H
0
(q) and U(i) be the O(H
0
(q))-valued process defined by
U(i) h
0
# _
k
0
# S
0
(k
0
, h
0
) k(i) for all h
0
# H
0
(q), (4.6)
where for each k
0
# S
0
, k(i) is the solution to (4.2) with k(0)=k
0
.
Lemma 4.3. Suppose ihai h
0
# H
0
(q) and h(i)#U(i) h
0
, ihen h is a solu-
iion io (4.2) wiih h(0)=h
0
. Moreoter, i U(i) is a.s. sirongly coniinuous.
Froof. Let S
n
] be an increasing sequence of finite subsets of S
0
such
that S
n
=S
0
. Set
H
n
(i)# _
k
0
# S
n
(k
0
, h
0
) k(i),
so that
&(U(i) h
0
&H
n
(i)&
2
=
"
_
k
0
# S
0
"S
n
(k
0
, h
0
) k(i)
"
2
= _
k
0
# S
0
"S
n
| (k
0
, h
0
)|
2
which tends to zero uniformly in i as n. This shows that U(i) h
0
is
continuous, i.e., U(i) is strongly continuous.
Let h(i) denote the solution to (4.2) with h(0)=h
0
. In order to prove
that U(i) h
0
solves Eq. (4.2) with initial condition h
0
, it suffices to show
that U(i) h
0
=h(i) or equivalently that
lim
n
E &H
n
(i)&h(i)&
2
=0.
Now it is clear that H
n
solves (4.2) with initial condition h
n
#
_
k
0
# S
n
(k
0
, h
0
) k
0
. Therefore we have
h(i)&H
n
(i)=(h
0
&h
n
)&
|
i
0
D
[(dt)
(h(t)&H
n
(t))
+
1
2
|
i
0
A
(1)
(h(t)&H
n
(t)) dt,
from which it follows that
E &H
n
(i)&h(i)&
2
3 &(h
0
&h
n
)&
2
+3E
|
i
0
_
l # S
0
&D
l
(h(t)&H
n
(t))&
2
dt
+
3
2
E
}
|
i
0
&A
(1)
(h(t)&H
n
(t))& dt
}
2
3 &(h
0
&h
n
)&
2
+3(&D&
2
op
+&A
(1)
&
2
i}2)
_
|
i
0
E &(h(t)&H
n
(t))&
2
dt.
It now follows by Gronwall's Lemma that E &H
n
(i)&h(i)&
2
0 as n .
Q.E.D.
Theorem 4.4. Suppose ihai h : [0, )_WH
0
(q) is a progressitely
neasurable process, i.e., h |
[0, i]_W
is B([0, i]) F
i
}B(H
0
(q))-neasurable
for all i # (0, ). (Here B([0, i]) and B(H
0
(q)) are ihe Borel o-algebras on
[0, i] and H
0
(q) respeciitely.) lso assune ihai
E
{
|
i
0
&h (t)&
2
dt
=
< \i>0 (4.7)
and sei h(i)=
i
0
h (t) dt. Hence h is alnosi surely absoluiely coniinuous wiih
deritaiite giten by h (i). Then
d(U(i) h(i))=&D
d[(i)
(U(i) h(i))+
1
2
A
(1)
(U(i) h(i)) di+U(i) h (i) di (4.8)
496 BRUCE K. DRIVER
or equitalenily, because h(0)=0,
U(i) h(i)=&
|
i
0
D
d[(t)
(U(t) h(t))+
|
i
0
1
2
A
(1)
(U(t) h(t)) dt
+
|
i
0
U(t) h (t) dt. (4.9)
Froof. Let us first assume there is a constant M< such that almost
surely, sup
i # [0, )
&h (i)&
H
0
(q)
M. Let =0=i
0
<i
1
<i
2
< } } } <i
n
=i]
/[0, i] be a partition of [0, i] and || #max|i
i+1
&i
i
| : i # 0, 1, 2, ...,
n&1]]. For t=i
i
# let t+#i
(i+1) .n
. Then
U(i) h(i)&h(0)= _
t #
U(t+) h(t+)&U(t) h(t)]
= _
t #
(U(t+)&U(t)) h(t)+ _
t #
U(t)(h(t+)&h(t))
+ _
t #
(U(t+)&U(t))(h(t+)&h(t))
=:
+B
+C
.
For _ # (t, t+], let h
(_)=h(t), and U
(_)=U(t), o
(_)#(U(t+)&
U(t)). With this notation we have
=&
|
i
0
D
d[(_)
(U(_) h
(_))+
|
i
0
1
2
A
(1)
(U(_) h
(_)) d_,
B
=
|
i
0
U
(_) h (_) d_,

and
C
=
|
i
0
o
(_) h (_) d_.

If
#&
|
i
0
D
d[(t)
(U(t) h(t))+
|
i
0
1
2
A
(1)
(U(t) h(t)) dt
and
B#
|
i
0
U(t) h (t) dt,
then
E &&
&
2
2E
|
i
0
_
l # S
0
&D
l
(U(_)(h(_)&h
(_))&
2
d_
+2E
\
|
i
0
1
2
&A
(1)
(U(_)(h(_)&h
(_))& d_
+
2
2 &D&
2
op
E
|
i
0
&h(_)&h
(_)&
2
d_+2 &A
(1)
&
2
op
_E
\
|
i
0
&h(_)&h
(_)& d_
+
2
which tends to zero as || 0 by the Dominated convergence theorem.
Similarly
&B&B
&
|
i
0
&(U(_)&U
(_)) h (_)& d_.

Therefore, the strong continuity of U and the dominated convergence
theorem implies that BB
as || 0 a.s. Finally the estimate

&C
&
|
i
0
&o
(_) h (_)& d_,

the strong continuity of U, and the dominated convergence theorem
implies that lim
|| 0
C
=0. This proves Eq. (4.9) in the case that h is

bounded.
For the general case, let h
n
(i)#1
[0, n]
(&h (t)&) h (t) and
h
n
(i)#
|
i
0
1
[0, n]
(&h (t)&) h (t) dt.
Since &h
n
(i)&n, we know that Eq. (4.9) holds with h replaced by h
n
, i.e.,
U(i) h
n
(i)=&
|
i
0
D
d[(t)
(U(t) h
n
(t))+
|
i
0
1
2
A
(1)
(U(t) h
n
(t)) dt
+
|
i
0
U(t) h
n
(t) dt. (4.10)
498 BRUCE K. DRIVER
By the Dominated convergence theorem,
|
i
0
&h (t)&h
n
(t)&
2
dt=
|
i
0
|1&1
[0, n]
(&h
n
(t)&)|
2
&h (t)&
2
dt 0
a.s. n . (4.11)
We also have the Sobolev estimate,
&h(i)&h
n
(i)&
2
\
|
i
0
&h (t)&h
n
(t)& dt
+
2
i
|
i
0
&h (t)&h
n
(t)&
2
dt. (4.12)
Using equations (4.11), (4.12), the facts that A
(1)
is bounded and U(i) is
unitary, and the 1
2
-isometry property of the Ito^ integral, it is easy to let
n in Eq. (4.10) to conclude that Eq. (4.9) holds for this general h as
well. Q.E.D.
4.2. Inegraiion by Faris
Lemma 4.5. 1ei G # FC
(L, H
0
(q)) and k # H
0
(q), ihen
1
2
(AG( g), k)=
1
2
(A
0
G( g), k)+(G( g), A
(1)
k)]&({
0
G( g), Dk)
HS
,
where ( } , } )
HS
is defined in Eq. (4.5), A in Definiiion 2.3, A
0
in Definiiion 3.13
and A
(1)
=A|
H
0
(q)
in Theoren 2.5.
Froof. Using ihe skew synneiry properiies of D
h
we find
(AG( g), k)= _
h # S
0
({
2
h
G( g), k)= _
h # S
0
((h+D
h
)
2
G( g), k)
=(A
0
G( g), k)+2 _
h # S
0
(D
h
hG( g), k)+ _
h # S
0
(D
2
h
G( g), k)
=(A
0
G( g), k)&2 _
h # S
0
({
0
h
G( g), D
h
k)+ _
h # S
0
(G( g), D
2
h
k)
=(A
0
G( g), k)&2({
0
G( g), Dk)
HS
+(G( g), A
(1)
k). Q.E.D.
Noiaiion 4.6. Suppose that f # FC
(L) of the form f ( g)=1( g

P
),
where P is a partition of [0, 1] and 1: G
P
R is a smooth function.
Define (e
iA}2
f ) : L(G) R by
(e
iA}2
f )( g)#(e
iA
P
}2
1)( g
P
), (4.13)
where A
P
is defined in Eq. (3.15) above.
Since A
P
is the generator of 2
P
( } ), we could also write Eq. (4.13) as
(e
iA}2
f )(g)#E[1( g
P
2
P
(i))]=E[ f ( g2(i))], (4.14)
where 2 solves Eq. (3.5) with 2(0, s)#e for all s # [0, 1]. It should also be
noted that
(e
iA}2
f )( g)
i
=
1
2
(A
P
e
iA
P
}2
1)( g
P
)=
1
2
(Ae
iA}2
f )( g),
where we have used Eqs. (2.3) and (3.15) to conclude that
(A
P
e
iA
P
}2
1)( g
P
)=(Ae
iA}2
f )( g). (4.15)
Proposition 4.7. 1ei h # H(H
0
(q)), f # FC
(L), and T>0. Sei 1

i
#
e
(T&i) A}2
f ihen
d({
9
1
i
(2(i)), U(i) h(i))=(({
d[(i)
{
9
1
i
)(2(i)), U(i) h(i))
+({
9
1
i
(2(i)), U(i) h (i)+
1
2
Ric U(i) h(i)) di.
Froof. By Ito^ 's Lemma, Theorem 3.14 above and the equalities
{
9
1
i
}i={
9
A1
i
}2 and 1
i
}i=A1
i
}2, we have
d({
9
1
i
(2(i)))=(({
0
d[(i)
{
9
1
i
)(2(i))+
1
2
(A
0
{
9
1
i
(2(i))&{
9
A1
i
(2(i))) di.
Using this equation, Theorem 4.4 and Ito^ 's Lemma (see Theorem 3.14
above and Theorem 8.5 in the Appendix),
d({
9
1
i
(2(i)), U(i) h(i))
=(({
0
d[(i)
{
9
1
i
)(2(i))+
1
2
(A
0
{
9
1
i
(2(i))&{
9
A1
i
(2(i))) di, U(i) h(i))
+({
9
1
i
(2(i)), &D
d[(i)
(U(i) h(i))+
1
2
A(U(i) h(i)) di+U(i) h (i) di)
+({
0
{
9
1
i
(2(i)), &D(U(i) h(i)))
HS
di.
The above expression may be simplified using {={
0
+D and Lemma 4.5
above to get
d({
9
1
i
(2(i)), U(i) h(i))=(({d[(i) {
9
1
i
)(2(i)), U(i) h(i))
+({
9
1
i
(2(i)), U(i) h (i)) di
+
1
2
((A {
9
1
i
&{
9
A1
i
)(2(i)), U(i) h(i)) di.
500 BRUCE K. DRIVER
This equation and the Bochner Wietzenbock Formula in Eq. (2.6) proves
the proposition. Q.E.D.
Noiaiion 4.8. For each unitary map U: H
0
(q) H
0
(q), let Ric
U
#
U
&1
Ric U, where Ric is defined in Eq. (2.2).
Corollary 4.9. Coniinuing ihe noiaiion fron Froposiiion 4.7,
E({
9
f (2(T)), U(T) h(T))=E
|
T
0
({
9
1
i
(2(i)), U(i) H (i)) di,
where
H(i)#h(i)+
1
2
|
i
0
Ric
U(t)
h(t) dt. (4.16)
Froof. By Proposition 4.7 and the assumption that h(0)=0,
({
9
f (2(T)), U(T) h(T))=
|
T
0
(({
d[(i)
{
9
1
i
)(2(i)), U(i) h(i))
+
|
T
0
({
9
1
i
(2(i)), U(i) H (i)) di.
The proof is completed by taking expectations of both sides of this
equation. Q.E.D.
We now may state the first version of the main theorem of this section.
Theorem 4.10 (Integration by Parts I). 1or each h # H(H
0
(q)) and
f # FC
(L),
E[({
9
f (2(T)), U(T) h(T))]=E[ f(2(T)) :
T
(h)],
where
:
T
(h)#
|
T
0
(U(i) (h (i)+
1
2
Ric
U(i)
h(i)), d[(i))
=
|
T
0
(U(i) h (i)+
1
2
Ric U(i) h(i), d[(i)). (4.17)
Froof. Let H be defined as in Eq. (4.16). Then using the 1
2
-isometry
property of the Ito^ integral and Ito^ 's lemma (Theorem 3.10) we find
E[ f (2(T)) :
T
(h)]
=E
_{
1
0
(2(0))+
|
T
0
({
9
1
i
(2(i)), d[(i))
=
|
T
0
(U(i) H (i), d[(i))
&
=E
_
|
T
0
({
9
1
i
(2(i)), U(i) H (i)) di
&
=E({
9
f (2(T)), U(T) h(T)),
wherein the last equality we have used Corollary 4.9. Q.E.D.
Definition 4.11. 1or each h # H(H
0
(q)) lei X
h
denoie ihe tecior-filed on
(L(G)) defined by
X
h
i
(2)=1
2(i)
*
U(i) h(i).
Theorem 4.12 (Integration by Parts II). 1or each h # H(H
0
(q)) and
f # FC
((L)),
E[(X
h
f )(2)]=E[ f (2(T)) :
T
(h)], (4.18)
where X
h
f is defined in Eq. (1.12) and :
T
(h) is defined in Eq. (4.17).
Froof. Write f (2)=1(2) where 2#(2(i
1
), ..., 2(i
k
)), 0<i
1
<i
2
< } } }
<i
k
, and 1: L
k
R is a smooth cylinder function. That is
1( g
1
, ..., g
k
)=1(( g
1
)
P
, ..., ( g
k
)
P
) \g
i
# L,
where P=0<s
1
<s
2
< } } } <s
n
<1] is partition of [0, 1] and 1 #
C
(G
kn
). We will prove the theorem by induction on k. The case k=1 is
the content of Theorem 4.10. Suppose k>1 and the theorem is true when
there are k&1 i
i
's. The induction step will be completed by showing that
Eq. (4.18) holds for f (2)=1(2) described above.
For h # H
0
(q), let h
(i )
1 denote the action of h on the i th variable of 1,
i.e.,
(h
(i )
1)(2)#
d
dc }
0
1((2(i
1
), ..., 2(i
i
) e
ch
, ..., 2(i
k
)) \h # H
0
(q).
502 BRUCE K. DRIVER
Also let
A
(i )
1# _
h # S
0
(h
(i )
)
2
1
and ({
9
(i )
1)(2) denote the gradient of 1 in the i th variable, i.e., ({
9
(i )
1)(2)
is the unique element of H
0
(q) such that
(({
9
(i )
1)(2), h)=(h
(i )
1)(2) \h # H
0
(q).
Then
E(X
h
f )(2)= _
k&1
i=1
E(({
9
(i )
1)(2), U(i
i
) h(i
i
))+E(({
9
(k)
1)(2), U(i
k
) h(i
k
))
=: S+T. (4.19)
Let o#(2(i
1
), ..., 2(i
k&1
)), u=i
k&1
, t=i
k
, o#i
k
&i
k&1
, and 1
i
(o, g)
#(e
(t&i) A
(k)
}2
1)(o, g). Then by Proposition 4.7 and Corollary 4.9,
T=E
_
(({
9
(k)
1
u
)(o, 2(u)), U(u) h(u))+
|
t
u
d(({
9
(k)
1
i
)(o, 2(i)), U(i) h(i))
&
=E
_
(({
9
(k)
1
u
)(o, 2(u)), U(u) h(u))+
|
t
u
(({
9
(k)
1
i
)(o, 2(i)), U(i) H (i)) di
&
=E
_
(({
9
(k)
1
u
)(o, 2(u)), U(u) h(u))
+
|
t
u
(({
9
(k)
1
i
)(o, 2(i)), d[(i)) }
|
t
u
(U(i) H (i), d[(i))
&
.
By Theorem 3.14,
|
t
u
(({
9
(k)
1
i
)(o, 2(i)), d[(i))=1
t
(o, 2(t))&1
u
(o, 2(u))
=1(2)&1
u
(o, 2(u)).
Since t
t
0
(U(i) H (i), d[(i)) is a Martingale,
E[1
u
(o, 2(u))
|
t
u
(U(i) H (i), d[(i))]=0.
Combining the three above displayed equations gives
T=E
{
(({
9
(k)
1
u
)(o, 2(u)), U(u) h(u))+1(2) }
|
t
u
(U(i) H (i), d[(i))
=
. (4.20)
Using the Markov property, S may be written as
S= _
k&1
i=1
E(({
9
(i )
1
u
)(o, 2(u)), U(i
i
) h(i
i
)). (4.21)
Set 1(2)=1
u
(o, 2(u)) so that 1 # FC
((L)), and notice that

(X
h
1)(2)= _
k&1
i=1
(({
9
(i )
1
u
)(o, 2(u)), U(i
i
) h(i
i
))
+(({
9
(k)
1
u
)(o, 2(u)), U(u) h(u)). (4.22)
Therefore, by Equations (4.204.22),
S+T=E
{
(X
h
1)(2)+1(2) }
|
t
u
(U(i) H (i), d[(i))
=
=E
{
1(2)
|
u
0
(U(i) H (i), d[(i))+1(2) }
|
t
u
(U(i) H (i), d[(i))
=
,
wherein the second equality we have used the induction hypothesis. Using
the Markov property once again,
E
{
1(2)
|
u
0
(U(i) H (i), d[(i))
=
=E
{
1
u
(o, 2(u))
|
u
0
(U(i) H (i), d[(i))
=
=E
{
1(o, 2(t))
|
u
0
(U(i) H (i), d[(i))
=
=E
{
1(2)
|
u
0
(U(i) H (i), d[(i))
=
.
The Theorem now follows from the last two equations and (4.19). Q.E.D.
4.3. Closabiliiy of ihe Dirichlei 1orn
Recall that v
T
is the Law of 2
T
=2(T), where 2 is the Brownian motion
on L constructed in Theorem 3.8 with 2(0, s)#e for all s # [0, 1].
504 BRUCE K. DRIVER
Definition 4.13. Let E
0
T
denote the symmetric quadratic form on
1
2
(L(G), v
T
) with domain D(E
0
T
)=FC
(L) and for u, t # FC
(L),
E
0
T
(u, t)#
|
L(G)
({
9
u( g), {
9
t(g))
H
0
(q)
v
T
(dg)=E[({
9
u(2
T
), {
9
t(2
T
))
H
0
(q)
].
Theorem 4.14. The quadraiic forn E
0
T
is closable.
Froof. To simplify notation, let E
0
T
( f )=E
0
T
( f, f ). Suppose that f
n
#
FC
(L) such that lim

n
f
n
=0 in 1
2
(v
T
) and
E
0
T
( f
n
&f
n
) 0 as n, n . (4.23)
We must show that lim
n
E
0
T
( f
n
)=0. Because of (4.23) the functions
G
n
#{
9
f
n
# FC
(L, H
0
(q)) form a Cauchy sequence in 1
2
(L(G), v
T
;
H
0
(q)). Hence there exists G # 1
2
(L(G), v
T
; H
0
(q)) such that 1
2
&
lim
n
G
n
=G. Since
E
0
T
( f
n
)=
|
L(G)
&G
n
&
2
dv
T
=E &G
n
(2
T
)&
2
,
it follows that lim
n
E
0
T
( f
n
)=E &G(2
T
)&
2
. So to finish the proof it
suffices to show that G(2
T
)=0 a.s.
To this end let h # H
0
(q), Q # FC
((L)), and set k(i)=(i}T) h. Then

k # H(H
0
(q)) and using the integration by parts Theorem 4.10 we find
E(G(2(T)), U(T) h) Q(2)]= lim
n
E({
9
f
n
(2(T)), U(T) k(T)) Q(2)]
= lim
n
E(X
k
f
n
(2
T
)) Q(2)]
= lim
n
E f
n
(2
T
)(&X
k
+:
T
(k)) Q(2)]=0.
Because FC
((L)) is dense in 1
2
((L), v) and because Q# FC
((L))
was arbitrary, the last displayed equation implies (G(2
T
), U(T) h)=0 a.s.
Hence
&G(2
T
)&
2
= _
h # S
0
(G(2
T
), U(T) h)
2
=0 a.s.,
i.e., G(2
T
)#0 a.s. Q.E.D.
Fenark 4.15. Theorem 4.14 may be stated equivalently as saying that
the gradient operator {
9
with domain FC
(L) has a densely defined 1

2
-
adjoint. However, the method of proof does not give any explicit informa-
tion as to what is in the domain in 1
2
(L, v
T
; H
0
(q)) of the adjoint
operator {
9
*. This deficiency will be remedied in Theorem 6.2 of Section 6
below where it is shown that FC
(L, H
0
(q))/D({
9
*).
5. THE FINITE DIMENSIONAL APPROXIMATIONS
5.1. 1iniie Dinensional Iniegraiion by Faris 1ornula
In this section let P=0<s
1
<s
2
< } } } <s
n
<1] be a partition of
[0, 1]. In order to prove Theorem 1.5 above, we will apply Corollary 6.5 of
Driver [5] to the Lie group G
P
and then pass to the limit of finer and finer
partitions P. In order to carry out this procedure it is necessary to
introduce the unique Riemannian metric, ( } , } )
P
, on G
P
for which A
P
in
Eq. (3.15) will become the Laplace Beltrami operator on (G
P
, ( } , } )
P
).
Let q
P
be the Lie algebra of G
P
which may naturally be identified with
the set of functions from P to q. In the sequel, h and k will typically denote
elements of q
P
or H
0
(q).
Proposition 5.1 (Metric on G
P
). 1ei P=0<s
1
<s
2
< } } } <s
n
<1]
be a pariiiion of [0, 1], and Q be ihe interse io ihe nairi\ G
0
(s
i
, s
j
)]
n
i, j=1
,
and
(h, k)
P
# _
n
i, j=1
Q
ij
h(s
i
), k(s
j
), for all h, k # q
P
.
e e\iend ( } , } )
P
io a lefi intariani Fienannian neiric on G
P
which will
siill be denoied by ( } , } )
P
. Then ihe ellipiic differeniial operaior A
P
defined
in (3.15) is ihe 1aplace Belirani operaior on G
P
wiih neiric ( } , } )
P
.
Froof. It is an exercise in linear algebra to check that A
P
may be
written as A
P
=_
h # 1
h
2
, where 1 is an orthonormal basis of (q
P
, ( } , } )
P
)
and h denotes the unique left invariant vector field on G
P
such that h(e)
=h. It is well known that _
h # 1
h
2
is the Laplace-Beltrami operator on G
P
because G
P
is compact and hence uni-modular, see for example
Remark 2.2 in Driver and Gross [7]. Q.E.D.
Noiaiion 5.2. Let [
P
(i) be the standard Brownian motion on
(q
P
, ( } , } )
P
) given by [
P
(i)=( [(i, s
1
), [(i, s
2
), ..., [(i, s
n
)), and {
P
and
Ric
P
be the Levi-Civita covariant derivative and the Ricci tensor on
(G
P
, ( } , } )
P
) respectively. For all h, k # q
P
, define D
P
h
k=({
P
h
k)(e) with
e # G
P
being the identity element. Also define U
P
to the solution to the
Stratonovich differential equation,
dU
P
(i)+D
P
o[
P
(i)
U
P
(i)=0 with U
P
(0)=I # End(q
P
).
506 BRUCE K. DRIVER
The operators D
P
h
and Ric
P
are computed explicitly in Proposition 5.7
and Eq. (5.9) below. We now may state a finite dimensional verion of
Theorem 1.5. Q.E.D.
Theorem 5.3. 1ei T>0 and l : [0, T] R be an absoluiely coniinuous
funciion such ihai l(0)=0, l(T)=1 and
T
0
l
2
(i) di<. Suppose ihai
f # FC
(L) is of ihe forn f ( g)=1( g

P
) where 1 # C
(G
P
) and P=
0<s
1
<s
2
< } } } <s
n
<1] is a pariiiion of [0, 1]. Then for all h # H
0
(q),
E[(hf )(2(T))]=E
_
f (2(T))(U
P
(T)
&1
h
P
,
|
T
0
U
&1
P
(i)(l(i)&
1
2
l(i) Ric
P
) d[
P
(i))
&
, (5.1)
where h
P
#h |
P
and d[
P
denoies ihe backwards Iio` differeniial.
Froof. By Proposition 5.1 and the dicussion leading up to Eq. (3.15),
2
P
and [
P
are Brownian motions on G
P
and q
P
respectively which are
related to each other by Eq. 3.14. Using this remark and the identities,
f(2(T))=1(2
P
(i)) and (hf )(2(T))=(h
P
1)(2
P
(i)), Eq. (5.1) follows as
an application of Corollary 6.5 in [5]. Q.E.D.
The proof of Theorem 1.5 will be given in Section 6 by passing to the
limit of finer and finer partitions P of [0, 1] in Eq. (5.1). In order to take
this limit it is necessary to understand the geometry on G
P
and its rela-
tionship to the geometry on L(G). This is the topic of the next subsection.
5.2. Geoneiry of ihe 1iniie Dinensional ppro\inaiions
To facilitate our computations, it will be convenient to identify q
P
with
the orthogonal compliment to the null space, nul(H
P
)/H
0
(q), where
H
P
: H
0
(q) q
P
is defined by H
P
(h)#h |
P
. The next lemma shows that
nul(H
P
)
=
=H
P
(q)#h # H
0
(q) :C
2
((0, 1)"P, q) : h"=0 on [0, 1]"P].
(Notice that h # H
P
(q) iff h # H
0
(q) and h is piecewise linear.) The following
notation will be used in the next lemma and the remainder of this section.
Noiaiion 5.4. Given a partition P=0<s
1
<s
2
< } } } <s
n
<1] of [0, 1]
and h # H
0
(q), for i=0, 1, 2, ..., n, let o
i
h#h(s
i+1
)&h(s
i
) and o
i
#s
i+1
&s
i
where s
0
#0 and s
n+1
#1.
Lemma 5.5. The orihogonal conplineni H
P
(q)
=
of H
P
(q) in H
0
(q) is
H
P
(q)
=
=nul(H
P
)=h # H
0
(q) : h |
P
#0].
Froof. Suppose that h # H
0
(q) and k # H
P
(q), then
(h, k)= _
n
i=0
|
s
i+1
s
i
h$(o), k$(o), do= _
n
i=0
o
i
h, o
i
k,}o
i
. (5.2)
Hence if h # nul(H
P
), i.e., h |
P
=0, then (h, k)=0 for all k # H
P
(q). Hence
nul(H
P
)/H
P
(q)
=
.
For the other inclusion, suppose that
i
# q is given such that
_
n
i=0

i
=0. Define
k(s)#
|
s
0
\
_
n
i=0
1
(s
i
, s
i+1
]
(o)
i
}o
i
+
do= _
n
i=0
(s
i+1
.s&s
i
.s)
i
}o
i
.
Since k$(s)=o
i
k}o
i
=
i
}o
i
for s # (s
i
, s
i+1
] and k(1)=_
n
i=0

i
=0, k is in
H
P
(q). If h # H
P
(q)
=
then, using the k just constructed in Eq. (5.2),
0= _
n
i=0
o
i
h,
i
,}o
i
(5.3)
for all
i
# q such that _
n
i=0

i
=0. Since
_
n
i=0
o
i
h=h(1)&h(0)=0,
we may put
i
#o
i
h in (5.3) to find
0= _
n
i=0
o
i
h, o
i
h,}o
i
,
i.e., o
i
h#0 for all i=0, 1, 2, ..., n&1. Because h(0)=0, this implies that
h|
P
#0. Thus we have shown that if H
P
(q)
=
/nul(H
P
). Q.E.D.
In general H
P
(q) is a subspace of H
0
(q) but not a Lie subalgebra with
the inherited pointwise commutator. In order to remedy this, let F
P
:
H
0
(q) H
P
(q) denote the orthogonal projection map and define [ } , } ]
P
on H
P
(q) by [h, k]
P
#F
P
[h, k]. One may check that (H
P
(q), [ } , } ]
P
) is
a Lie algebra. Indeed the only non-trivial property to verify is the Jacobi
identity. Since [h, [k, l ]
P
]
P
=F
P
[h, [k, l ]
P
] is uniquely determined by it
values on P, i.e., by the values [h(s), [k, l ]
P
(s)]=[h(s), [k(s), l(s)]] for
s # P, the Jacobi identity simply follows from the Jacobi identity for the Lie
bracket ([ } , } ]) on q.
Lemma 5.6. 1ei H
P
(q) and q
P
be ihe 1ie algebras described abote
equipped wiih ihe inner producis ( } , } )=( } , } )
H
0
(q)
and ( } , } )
P
respeciitely.
508 BRUCE K. DRIVER
Then linear nap H
P
: H
P
(q) q
P
(H
P
(h)#h
P
#h |
P
) is an isoneiric 1ie
algebra isonorphisn.
Froof. The only assertion which is not obvious to check the asser-
tion that H
P
is an isometry. For =(
1
,
2
, ...,
n
) # q
n
let h
#
_
n
i=1
G
0
(s
i
, } )
i
and notice that h
# H
P
(q). Using the reproducing kernel
property for G
0
, see Eq. 3.11 in [6],
(h
, h
B
)= _
n
i, j=1
(G
0
(s
i
, } ), G
0
(s
j
, } ))
H
0
(R)

i
, B
j
,= _
n
i, j=1
G
0
(s
i
, s
j
)
i
, B
j
,.
Because G
0
(s
i
, s
j
)]
n
i, j=1
is a positive definite matrix, the last equation with
B=, shows that # q
n
h
# H
P
(q) is injective hence surjective by the
rank nullity theorem. On the other hand,
(H
P
h
, H
P
h
B
)
P
= _
n
i, j, k, l=1
Q
k, l
h
(s
k
), h
B
(s
l
),
= _
n
i, j, k, l=1
Q
k, l
G
0
(s
i
, s
k
) G
0
(s
j
, s
l
)
i
, B
j
,
= _
n
i, j=1
G
0
(s
i
, s
j
)
i
, B
j
,.
Comparing the last two displayed equations proves the isometry assertion.
Q.E.D.
liernaie proof of ihe isoneiry properiy. In this proof we will use the
fact that second order elliptic differential operators on a manifold induce a
unique Riemannian metric on the manifold.
Let 1 # C
(G
P
), f #1b
P
# FC
(L), and S
P
be an orthonormal basis
for H
P
(q). Then, using Lemma 5.5 and the fact that the sum defining the
Laplace operator A in Eq. (2.1) is basis independent, we have
Af ( g)= _
h # S
P
(h
2
f )( g)= _
h # S
P
(h
2
P
1)( g
P
) \g # L.
On the other hand by Eq. (2.3), Af ( g)=(A
P
1)( g
P
). Hence we learn that
A
P
=_
h # S
P
h
2
P
which is the Laplace Beltrami operator on G
P
equipped
with the metric on q
P
for which the map H
P
is an isometry. But this inner
product must agree with ( } , } )
P
, since we have seen in the proof of
Proposition 5.1 that A
P
is also the Laplacian relative to the metric ( } , } )
P
on G
P
. Q.E.D.
In the sequel we will identify q
P
with H
P
(q).
Proposition 5.7. 1or h # H
P
(q), lei D
P
h
: H
P
(q) H
P
(q) denoie 1ie
algebra tersion of ihe 1eti-Citiia cotariani deritaiite ({
P
) on G
P
as defined
in Noiaiion 5.2 (Fecall ihai we are ideniifying q
P
wiih H
P
(q) as in
1enna 5.6.) Then D
P
h
=F
P
D
h
, where as abote F
P
is ihe orihogonal projec-
iion of H
0
(q) onio H
P
(q).
Froof. We need to check that D
P
h
is metric compatible and Torsion
free. Both of these properties follow directly from the corresponding
properties of D
h
described in Theorem 3.12 of Driver and Lohrenz [6].
Indeed if h, k # H
P
(q), then
(D
P
h
k, k)=(F
P
D
h
k, k)=(D
h
k, k)=0
and
D
P
h
k&D
P
k
h=F
P
(D
h
k&D
k
h)=F
P
([h, k])=[h, k]
P
. Q.E.D.
Theorem 5.8. 1ei S
P
be an orihonornal basis for H
P
(q) and
A
(1)
P
: H
P
(q) H
P
(q) be defined by
A
(1)
P
# _
k # S
P
D
P
k
D
P
k
.
Then
lim
|P| 0
&F
P
(A
(1)
&A
(1)
P
) F
P
&
op
=0,
where &} &
op
is ihe operaior norn on bounded linear operaiors on H
0
(q).
The following lemma is used in the proof of this theorem.
Lemma 5.9. 1ei S
P
be an orihonornal basis for H
P
(q) and , B # q.
Then
_
k # S
P
[k(s
i
), ], [k(s
j
), B],=G
0
(s
i
, s
j
) K, B, (5.4)
for all i, j # 1, 2, ..., n].
Froof. It is easily checked that the left member in Eq. (5.4) is independent
of the orthonormal basis S
P
of H
P
(q). So to simplify the computation, we
510 BRUCE K. DRIVER
may take S
P
#lC]
l # h
P
, C # q
0
where q
0
is an orthonormal basis for q and
h
P
is an orthonormal basis for
H
P
(R)#l # H
0
(R) :C
2
((0, 1)"P, R) : l "#0 on (0, 1)"P].
Because
_
C # q
0
[C, ], [C, B],= _
C# q
0
ad
C, ad
B
C,= _
C# q
0
&C, ad
ad
B
C,
=&tr(ad
ad
B
)=K, B,,
it follows that
_
k # S
P
[k(s
i
), ], [k(s
j
), B],= _
l # h
P
, C# q
0
l(s
i
) l(s
j
)[C, ], [C, B],
= _
l # h
P
l(s
i
) l(s
j
) K, B,.
To evaluate _
l # h
P
l(s
i
) l(s
j
), let h/H
0
(R) be an orthonormal basis of
H
0
(R) which contains h
P
. Notice if l # h"h
P
, then l # H
P
(R)
=
and hence
l |
P
#0. Therefore
_
l # h
P
l(s
i
) l(s
j
)= _
l # h
l(s
i
) l(s
j
)=G
0
(s
i
, s
j
),
where the last equality verified in Lemma 3.8 of [6]. Q.E.D.
Froof of Theoren 5.8. Let h, J # H
P
(q), then
(A
(1)
P
h, J)= _
k # S
P
(D
P
k
D
P
k
h, J)
=& _
k # S
P
(D
P
k
h, D
P
k
J)
=& _
k # S
P
_
n
i=0
o
i
(D
k
h), o
i
(D
k
J),}o
i
.
Now
o
i
(D
k
h)=
|
s
i+1
s
i
[k, dh]&o
i
|
1
0
[k, dh]
and
|
s
i+1
s
i
[k, dh]=
|
s
i+1
s
i
_
k(s
i
)+
(o&s
i
)
o
i
o
i
k, o
i
h}o
i
&
do
=
_
k(s
i
) o
i
+
o
2
i
2o
i
o
i
k, o
i
h}o
i
&
=
_
k(s
i
)+
1
2
o
i
k, o
i
h
&
=[(k(s
i
)+k(s
i+1
))}2, o
i
h].
Set k
a
i
#(k(s
i
)+k(s
i+1
))}2, then the above two displayed equations show
that
o
i
(D
k
h)=[k
a
i
, o
i
h]&o
i
_
n
j=0
[k
a
j
, o
j
h]. (5.5)
Thus, using Lemma 5.9,
(A
(1)
P
h, J)=& _
k # S
P
_
n
i=0
o
&1
i
[k
a
i
, o
i
h]&o
i
_
n
j=0
[k
a
j
, o
j
h],
[k
a
i
, o
i
J]&o
i
_
n
j=0
[k
a
j
, o
j
J]
=& _
k # S
P
_
n
i=0
o
&1
i
{
[k
a
i
, o
i
h], [k
a
i
, o
i
J],
+o
2
i
_
n
j=0
[k
a
j
, o
j
h], _
n
l=0
[k
a
l
, o
l
J]
=
& _
k # S
P
_
n
i=0
o
&1
i
{
o
i
_
n
j=0
[k
a
j
, o
j
h], [k
a
i
, o
i
J]
&o
i
[k
a
i
, o
i
h], _
n
j=0
[k
a
j
, o
j
J]
=
=& _
k # S
P
_
n
i=0
o
&1
i
[k
a
i
, o
i
h], [k
a
i
, o
i
J],
&
_
n
j=0
[k
a
j
, o
j
h], [k
a
i
, o
j
J]
=
= _
n
i=0
{
&G
ii
o
&1
i
Ko
i
h, o
i
J,+ _
n
j=0
G
ij
Ko
j
h, o
i
J,
=
=&S+T,
512 BRUCE K. DRIVER
where
G
ij
=
1
4
G
0
(s
i
, s
j
)+G
0
(s
i+1
, s
j
)+G
0
(s
i
, s
j+1
)+G
0
(s
i+1
, s
j+1
)].
We now work on the two terms S and T separately.
S= _
n
i=0
|
s
i+1
s
i
G
ii
Kh$(o), J$(o), do=
|
1
0
j
P
(o) Kh$(o), J$(o), do,
where
j
P
(s)# _
n
i=0
G
ii
1
(s
i
, s
i+1
]
(s).
Similarly,
T= _
n
i=0
_
n
j=0
G
ij
Ko
j
h, o
i
J,
= _
n
i=0
_
n
j=0
G
ij
|
s
i+1
s
i
do
|
s
j+1
s
j
ds Kh$(s), J$(s),
=
|
1
0
do
|
1
0
ds G
P
(o, s) Kh$(s), J$(o),,
where
G
P
(o, s)# _
n
i, j=0
G
ij
1
(s
i
, s
i+1
]
(o) } 1
(s
j
, s
j+1
]
(s). (5.6)
Assembling the above computations gives
(A
(1)
P
h, J)=&
|
1
0
j
P
(o) Kh$(o), J$(o), do
+
|
1
0
do
|
1
0
ds G
P
(o, s) Kh$(s), J$(o),.
From Eq. (4.42) in [6],
(A
(1)
h, J)=&
|
1
0
G
0
(o, o) Kh$(o), J$(o), do
+
|
1
0
do
|
1
0
ds G
0
(o, s) Kh$(s), J$(o),.
Combining the last two equations gives
((A
(1)
&A
(1)
P
) h, J)=
|
1
0
( j
P
(o)&G
0
(o, o)) Kh$(o), J$(o), do
+
|
1
0
do
|
1
0
ds(G
0
(o, s)&G
P
(o, s)) Kh$(s), J$(o),.
It is now a simple matter to use this equation to show that
&F
P
(A
(1)
&A
(1)
P
) F
P
&
B(H
P
(q))
Cc(P),
where
c(P)# max
s, i # [0, 1]
| j
P
(s)&G
0
(s, s)| +|G
0
(s, i)&G
P
(s, i)| ].
By the uniform continuity of G
0
, lim
|P| 0
c(P)=0. This proves the
theorem. Q.E.D.
We now work on the Ricci tensor.
Theorem 5.10. 1ei Ric
P
be ihe 1ie algebra tersion of ihe Ficci iensor
on G
P
. (e will inierchangeably tiew Ric
P
as a bi-linear forn or an
operaior on H
P
(q). Then
&F
P
(Ric&Ric
P
) F
P
&
op
0 as |P| 0.
Froof. Let h # H
P
(q) and S
P
/H
P
(q) be an orthonormal basis of
H
P
(q). Since the expressions of interest are independent of the choice of
orthonormal basis, we may assume with out loss of generality that S
P
is
a ``good basis,'' i.e., [h(s), h(o)]=0 for all s, o # [0, 1]. (For example take
the basis used in the proof of Lemma 5.9.) Then
Ric
P
h, h,= _
k # S
P
(F
P
h, k, k, h)
= _
k # S
P
(D
P
h
D
P
k
k&D
P
k
D
P
h
k&D
P
[h, k]
P
k, h)
= _
k # S
P
(&D
P
k
D
P
h
k&D
P
[h, k]
P
k, h)
= _
k # S
P
(D
P
h
k, D
P
k
h)&(D
P
[h, k]
P
k, h)]
=: S&T.
514 BRUCE K. DRIVER
Using Eq. (5.5), S may be written as
S= _
k # S
P
_
n
i=0
o
&1
i
o
i
D
P
h
k, o
i
D
P
k
h,
= _
k # S
P
_
n
i=0
o
&1
i
[h
a
i
, o
i
k]&o
i
_
n
j=0
[h
a
j
, o
j
k],
[k
a
i
, o
i
h]&o
i
_
n
j=0
[k
a
j
, o
j
h]
= _
k # S
P
_
n
i=0
{
o
&1
i
[h
a
i
, o
i
k], [k
a
i
, o
i
h],& _
n
j=0
[h
a
i
, o
i
k], [k
a
j
, o
j
h],
=
= _
k # S
P
_
n
i=0
{
o
&1
i
[h
a
i
, o
i
k], [k
a
i
, o
i
h],& _
n
j=0
[k
a
i
, o
i
h], [k
a
j
, o
j
h],
=
=& _
k # S
P
_
n
i=0
o
&1
i
[k
a
i
, [h
a
i
, o
i
k]], o
i
h,& _
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
=& _
k # S
P
_
n
i=0
o
&1
i
ad
k
i
a
ad
h
i
a
o
i
k, o
i
h,& _
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
,
wherein the fourth equality we did a summation by parts. Namely we have
used
o
i
[h, k]=[h
a
i
, o
i
k]+[o
i
h, k
a
i
]=[h
a
i
, o
i
k]&[k
a
i
, o
i
h]
and the fact that _
n
i=0
o
i
[h, k]=0.
Similarly using Eq. (5.5) and _
n
i=0
o
i
h=0, T may be expressed as
T= _
k # S
P
_
n
i=0
o
&1
i
{
[[h, k]
a
i
, o
i
k]&o
i
_
n
j=0
[[h, k]
a
j
, o
j
k], o
i
h
=
= _
k # S
P
_
n
i=0
o
&1
i
ad
[h, k]
i
a
o
i
k, o
i
h,.
So combining the expressions for S and T shows
Ric
P
h, h,=S&T
=& _
k # S
P
_
n
i=0
o
&1
i
ad
k
i
a
ad
h
i
a
o
i
k+ad
[h, k]
i
a
o
i
k, o
i
h,
& _
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
.
Now using the assumption that S
P
is a good basis,
ad
k
i
a
ad
h
i
a
o
i
k+ad
[h, k]
i
a
o
i
k=[ad
k
i
a
, ad
h
i
a
] o
i
k+ad
[h, k]
i
a
o
i
k
=ad
[k
i
a
, h
i
a
]
o
i
k+ad
[h, k]
i
a
o
i
k
=ad
[h, k]
i
a
&[h
i
a
, k
i
a
]]
o
i
k
=
1
4
ad
[o
i
h, o
i
k]
o
i
k.
Assembling the last two equations implies
Ric
P
h, h,=&
1
4
_
k # S
P
_
n
i=0
o
&1
i
ad
[o
i
h, o
i
k]
o
i
k, o
i
h,& _
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
=
1
4
_
k # S
P
_
n
i=0
o
&1
i
|[o
i
h, o
i
k]|
2
& _
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
. (5.7)
The above sums on k # S
P
may be computed using Lemma 5.9,
1
4
_
k # S
P
_
n
i=0
o
&1
i
|[o
i
h, o
i
k]|
2
= _
n
i=0
o
&1
i
Ko
i
h, o
i
h, z
i
=
|
1
0
z
P
(o) Kh$(o), h$(o), do,
where
z
i
#G
0
(s
i
, s
i
)+G
0
(s
i+1
, s
i+1
)&2G
0
(s
i
, s
i+1
)]}4
and
z
P
(s)# _
n
i=0
z
i
1
(s
i
, s
i+1
]
(s). (5.8)
Similarly,
_
k # S
P
}
_
n
i=0
[k
a
i
, o
i
h]
}
2
= _
n
i, j=0
Ko
i
h, o
j
h, G
ij
=
|
1
0
do
|
1
0
ds G
P
(o, s) Kh$(s), h$(o),,
where G
P
is defined in Eq. (5.6). Hence
Ric
P
h, h,=
|
1
0
z
P
(o) Kh$(o), h$(o), do
&
|
1
0
do
|
1
0
ds G
P
(o, s) Kh$(s), h$(o),. (5.9)
516 BRUCE K. DRIVER
The polarization of Eq. (5.9) and Eq. (2.2) shows
Ric
P
h, k,&Rich, k,
=
|
1
0
z
P
(o) Kh$(o), k$(o), do
&
|
1
0
do
|
1
0
dsG
P
(o, s)&G
0
(o, s)] Kh$(s), k$(o),
for all h, k # H
P
(q).
Let &} &
u
denote the supremum norm on functions, then it easily follows
from the last equation, for all h, k # H
P
(q), that
|((Ric&Ric
P
) h, k)|
C &z
P
&
u
&h& &k&+
|
1
0
do
|
1
0
ds |G
0
(o, s)&G
P
(o, s)| |Kh$(s), k$(o),|
C(&z
P
&
u
+&G
0
&G
P
&
u
) &h& &k&.
Hence
&F
P
(Ric&Ric
P
) F
P
&
op
C(&z
P
&
u
+&G
0
&G
P
&
u
). (5.10)
Looking at the definitions of z
P
and G
P
and using uniform continuity of
G
0
, it is easily seen that &z
P
&
u
+&G
0
&G
P
&
u
0 as |P| 0. This observa-
tion and Eq. (5.10) finishes the proof. Q.E.D.
6. INTEGRATION BY PARTS ON THE LOOP GROUP
In this section we will prove Theorem 1.5 by passing to the limit in
Theorem 5.3. Before doing this it is first convenient to rewrite Eq. (5.1) as
E[(hf )(2(T))]
=E
_
f (2(T))
|
T
0
(U
P
(i) U
P
(T)
&1
h
P
, (l(i)&
1
2
l(i) Ric
P
)

d[
P
(i))
&
.
Setting U
P
(i, T)#U
P
(i) U
P
(T)
&1
, H
P
(i)#U
P
(i, T) h
P
=U
P
(i) U
P
(T)
&1
h
P
,
we see that this last equation may be written as
E[(hf )(2(T))]=E
_
f (2(T))
|
T
0
(H
P
(i), (l(i)&
1
2
l(i) Ric
P
)

d[
P
(i))
&
,
(6.1)
where H
P
(i) is the solution to the Stratonovich stochastic integral equation
h&H
P
(i)+
|
T
i
D
P
o[
P
(t)
H
P
(t)=0. (6.2)
The reader should notice that the process H
P
is not adapted to the filtra-
tion F
i
]
i0
. Nevertheless the integral in (6.2) may be defined as the usual
1
2
limit of Riemann sum approximations of the form
|
T
i
D
P
o[
P
(t)
H
P
# lim
|| 0
1
2
_
t #
D
P
([(t+)&[(t))
(H
P
(t)+H
P
(t+)), (6.3)
where denotes a partition of [i, T], for t # , t+ denotes the successor
to t in , and || is the mesh size of the partition. To show the con-
vergence, notice from the usual adapted theory that
|
T
i
D
P
o[
P
(t)
U
P
(t)# lim
|| 0
1
2
_
t #
D
P
( [(t+)&[(t))
(U
P
(t)+U
P
(t+)), (6.4)
where the sum exists in 1
2
. Since Eq. (6.3) is obtained from Eq. (6.4) by
multiplication on the right by U
&1
P
(T) h
P
, Eq. (6.3) has the same con-
vergence properties as Eq. (6.4). Moreover, by the discussion in Section 4.1
in [5], the integrals in Eq. (6.3) and (6.4) may be expressed in terms of
Backwards Ito^ integrals,
|
T
i
D
P
o[
P
(t)
H
P
(t)=
|
T
i
D
P
d[
P
(t)
H
P
(t)&
1
2
|
T
i
D
P
d[
P
(t)
dH
P
(t)
=
|
T
i
D
P
d[
P
(t)
H
P
(t)+
1
2
_
h # S
P
|
T
i
D
P
h
D
P
h
H
P
(t) dt
=
|
T
i
D
P
d[
P
(t)
H
P
(t)+
1
2
|
T
i
A
(1)
P
H
P
(t) dt,
where
|
T
i
D
P
d[
P
(t)
H
P
(t)# lim
|| 0
_
t #
D
P
( [(t+)&[(t))
H
P
(t+),
518 BRUCE K. DRIVER
Therefore (6.2) is equivalent to the backwards stochastic integral equation,
h&H
P
(i)+
|
T
i
D
P
d[
P
(t)
H
P
(t)+
1
2
|
T
i
A
(1)
P
H
P
(t) dt=0. (6.5)
Following the notation and discussion in Section 8.3 of the Appendix
below, let [
T
(i)#[(T&i)&[(T) and H
T
P
(i)#H
P
(T&i) for i # [0, T]
and F
T
i
]
i # [0, T]
denote the filtration generated by [
T
(i)]
i # [0, T]
appropriately completed. Then [
T
(i)]
i # [0, T]
is again a standard
F
T
i
]
i # [0, T]
-Brownian motion and Eq. (6.5) may be expressed as
h&H
P
(i)&
|
T&i
0
D
P
d[
T
P
(t)
H
T
P
(t)+
1
2
|
T&i
0
A
(1)
P
H
T
P
(t) dt=0.
Replacing i by T&i in this last equation shows that Eq. (6.5) (for H
P
) is
equivalent to the following standard forward stochastic differential
equation (for H
T
P
):
H
T
P
(i)=h&
|
i
0
D
P
d[
T
P
(t)
H
T
P
(t)+
1
2
|
i
0
A
(1)
P
H
T
P
(t) dt. (6.6)
This last equation may be written in differential form as
dH
T
P
(i)+D
P
d[
T
P
(i)
H
T
P
(i)&
1
2
A
(1)
P
H
T
P
(t)=0 with H
T
P
(0)=h. (6.7)
In analogy to Eq. (6.5), for h # H
0
(q), we let H(i) denote the solution to
the backwards stochastic differential equation
h&H(i)+
|
T
i
D
d[
(t)
H(t)+
1
2
|
T
i
A
(1)
H(t) dt=0. (6.8)
Theorem 6.1 (Backwards Parallel Translation). Giten T>0 and
h # H
0
(q), ihere e\isis a unique H
0
(q)-talued coniinuous backwards seni-
nariingale H(i), relaiite io ihe filiraiion F
T
i
]
i # [0, T]
, solting Eq. (6.8).
Moreoter ihere e\isis a process i # [0, T] U(i, T) # O(H
0
(q)) such ihai
for all h # H
0
(q), H(i)#U(i, T) h is ihe unique soluiion io Eq. (6.8).
Froof. Using Definition 8.6 of the backwards stochastic integral in
Section 8.3 of the Appendix and the same argument used above in passing
from Eq. (6.5) to Eq. (6.7), we find that Eq. (6.8) is equivalent to
H
T
(i)=h&
|
i
0
D
d[
T
(t)
H
T
(t)+
1
2
|
i
0
A
(1)
H
T
(t) dt, (6.9)
where H
T
#H(T&i). With this observation, the theorem follows from
Theorem 4.1 and Lemma 4.3 above. Q.E.D.
Theorem 6.2 (Integration by Parts). 1ei T>0, l # H(R) such ihai
l(T)=1, h # H
0
(q), and H(i) ( for i # [0, T]) be ihe unique soluiion io
Eq. (6.8). Then for all f # FC
(L),
E[(hf )(2
T
)]=E
_
f (2
T
)
|
T
0
(l(i)&
1
2
l(i) Ric] H(i),

d[(i))
&
, (6.10)
where

d[ denoies ihe backwards siochasiic differeniial. In pariicular,
Eq. (1.14) of Theoren 1.5 follows fron Eq. (6.10) by choosing l(i)#i}T.
Fenark 6.3. The backwards stochastic integral appearing in Eq. (6.10)
is well defined and we have the estimate
E:
2
:=E
|
T
0
&l(i)&
1
2
l(i) Ric] H(i)&
2
di
&h&
2
}
|
T
0
[|l(i)+
1
2
|l(i)| &Ric&
op
]
2
di<.
More generally, using Burkholder's inequality, for all p # [2, ) there are
constants C
p
< such that
&:&
1
p
(F)
C
p
&h&
\
|
T
0
[|l(i)+
1
2
|l(i)| &Ric&
op
]
2
di
+
1}2
<.
6.1. Fassing io ihe 1inii
The rest of this section will now be devoted to the proof of Theorem 6.2
which will be carried out by letting |P| 0 in Eq. (6.1). The following
theorem is the key result needed to take this limit. Recall the notation
used above Eq. (6.1), namely U
P
(i, T)#U
P
(i) U
P
(T)
&1
and H
P
(i)#
U
P
(i, T) h
P
=U
P
(i) U
P
(T)
&1
h
P
.
Theorem 6.4. 1ei T # (0, ), h # H
P
0
(q), and P
n
be a sequence of parii-
iions of [0, 1] such ihai P
n+1
#P
n
for all n and lim
n
|P
n
| =0. Sei
:#
|
T
0
(l(i)&
1
2
l(i) Ric] U(i, T) h,
d[(i)), (6.11)
520 BRUCE K. DRIVER
and
:
n
#
|
T
0
(l(i)&
1
2
l(i) Ric
P
n
(i, T) h,

d[
P
n
(i)). (6.12)
Then :
n
conterges io : in 1
2
.
Froof. To simplify notation let F
n
#F
P
n
: H
0
(q) H
0
(q) denote ortho-
gonal projection onto H
P
n
(q), Ric
n
#Ric
P
n
, and U
n
#U
P
n
. We will first
show that F
n
I strongly as n . To prove this it suffices to show, since
ran(F
n+1
)#ran(F
n
) for all n, that D#
n
ran(F
n
) is dense in H
0
(q). To
see that D is dense, first notice that G
0
(s, } ) # D for all s # Q#
n
P
n
and
# q. Hence if h = D, then h(s), ,=(h, G
0
(s, } ) )=0 for all s # Q and
# q. Since h is continuous and Q/[0, 1] is dense, it follows that h#0.
Therefore D is dense in H
0
(q) and hence F
n
I strongly.
In the remainder of the proof, c
n
will be used to denote any generic
sequence of non-negative real numbers such that lim
n
c
n
=0. (The value
of c
n
may vary from line to line in the following proof, but in all cases
lim
n
c
n
=0.)
Using the isometry property of the Ito^ Integral,
E |:&:
n
|
2
=E
}
|
T
0
(l(i)&
1
2
l(i) Ric] U(i, T) h
&l(i)&
1
2
l(i) Ric
n
] U
n
(i, T) h,

d[(i))
}
2
=E
|
T
0
&l(i)&
1
2
l(i) Ric] U(i, T) h
&l(i)&
1
2
l(i) Ric
n
] U
n
(i, T) h&
2
di
2
|
T
0
E &F
n
l(i)&
1
2
l(i) Ric] U(i, T) h
&l(i)&
1
2
l(i) Ric
n
] U
n
(i, T) h&
2
di
+2
|
T
0
E &(I&F
n
)l(i)&
1
2
l(i) Ric] U(i, T) h&
2
di.
Because F
n
I
H
0
(q)
strongly as n , U(i, T) is unitary and Ric is a
bounded operator, we may apply the dominated convergence theorem to
find
lim
n
|
T
0
E &(I&F
n
)l(i)&
1
2
l(i) Ric] U(i, T) h&
2
di=0. (6.13)
The last two displayed equations imply that
E |:&:
n
|
2
4
|
T
0
l
2
(i) E &F
n
U(i, T)&U
n
(i, T)] h&
2
di
+
|
T
0
l
2
(i) E &F
n
Ric U(i, T)&Ric
n
U
n
(i, T)] h&
2
di+c
n
.
(6.14)
As in the proof of Eq. (6.13),
lim sup
n
|
T
0
l
2
(i) E &F
n
Ric(I&F
n
) U(i, T) h&
2
di
lim sup
n
|
T
0
l
2
(i) &Ric&
2
E &(I&F
n
) U(i, T) h&
2
di=0,
which along with Eq. (6.14) implies that
E |:&:
n
|
2
4
|
T
0
l
2
(i) E &F
n
U(i, T)&U
n
(i, T)] h&
2
di
+2
|
T
0
l
2
(i) E &F
n
Ric F
n
U(i, T)&Ric
n
U
n
(i, T)] h&
2
di+c
n
.
(6.15)
Recall that Theorem 5.10 asserts that
lim
n
&F
n
Ric F
n
&Ric
n
F
n
&
2
op
=0
522 BRUCE K. DRIVER
and in particular this implies that C#sup
n
&Ric
n
F
n
&
2
op
<. Therefore
|
T
0
l
2
(i) E &F
n
Ric F
n
U(i, T)&Ric
n
U
n
(i, T)] h&
2
di
2
|
T
0
l
2
(i) &F
n
Ric F
n
&Ric
n
F
n
&
2
op
E &U(i, T) h&
2
di
+2
|
T
0
l
2
(i) E &Ric
n
F
n
U(i, T)&U
n
(i, T)] h&
2
di
2 &F
n
Ric F
n
&Ric
n
F
n
&
2
op
&h&
2
}
|
T
0
l
2
(i) di
+2
|
T
0
l
2
(i) &Ric
n
F
n
&
2
op
E &F
n
U(i, T)&U
n
(i, T)] h&
2
di
2C
|
T
0
l
2
(i) E &F
n
U(i, T)&U
n
(i, T)] h&
2
di+c
n
.
Using this estimate in Eq. (6.15) gives
E |:&:
n
|
2
4
|
T
0
(l
2
(i)+Cl
2
(i))
_E &F
n
U(i, T)&U
n
(i, T)] h&
2
di+c
n
. (6.16)
Since &F
n
U(i, T)&U
n
(i, T)] h&2 &h&, the theorem follows from
Eq. (6.16) and the dominated convergence theorem provided that
lim
n
E&F
n
U(i, T)&U
n
(i, T)] h&
2
=0.
This is the content of the next lemma. Q.E.D.
Lemma 6.5. Keeping ihe noiaiion of ihe pretious iheoren,
lim
n
sup
i # [0, T]
E &F
n
U(i, T)&U
n
(i, T)] h&
2
=0.
Froof. Recall that [ is the L(q)-valued Brownian motion described in
Section 3.1. Let [
n
#F
n
[=[
P
n
, D
n
#D
P
n
, A
(1)
n
#A
(1)
P
n
, Q
n
#I&F
n
, H(i)#
U(i, T) h, H
n
(i)#U
n
(i, T) h, and z
n
(i)#H
n
(i)&F
n
H(i). By Eq. (6.8),
dH=&D
d[
H&
1
2
A
(1)
H di=0 with H(T)=h
and Eq. (6.5)
dH
n
=&D
d[
n
H
n
&
1
2
A
(1)
n
H
n
di=0 with H
n
(T)=h.
Therefore,
&dz
n
=&dH
n
+F
n
dH
=D
n
d[
n
H
n
&F
n
D
d[
H+
1
2
(A
(1)
n
H
n
&F
n
A
(1)
H) di
=F
n
D
d[
n
(H
n
&H)&F
n
D
Q
n
d[
H
+
1
2
[A
(1)
n
(z
n
+F
n
H)&F
n
A
(1)
(F
n
H+Q
n
H)] di
=F
n
D
d[
n
z
n
&F
n
D
d[
n
Q
n
H&F
n
D
Q
n
d[
H
+
1
2
A
(1)
n
z
n
di+
1
2
(A
(1)
n
&F
n
A
(1)
F
n
) F
n
H di&
1
2
F
n
A
(1)
Q
n
H di,
with z
n
(T)=0. More precisely we have
z
n
(i)=
n
(i)&B
n
(i)&C
n
(i)+
1
2
(D
n
(i)+E
n
(i)&1
n
(i)),
where
n
(i)#
|
T
i
F
n
D
d[
n
z
n
, B
n
(i)#
|
T
i
F
n
D
d[
n
Q
n
H,
C
n
(i)#
|
T
i
F
n
D
Q
n
d[
H, D
n
(i)#
|
T
i
A
(1)
n
z
n
(t) dt,
E
n
(i)#
|
T
i
(A
(1)
n
&F
n
A
(1)
F
n
) F
n
H(t) dt,
and
1
n
(i)#
|
T
i
F
n
A
(1)
Q
n
H(t) dt.
Let us now estimate the 1
2
-norms of the four terms not containing z
n
.
B
n
-term,
E &B
n
(i)&
2
=E
|
T
i
_
l # S
0
&F
n
D
F
n
l
Q
n
H(u)&
2
du
E
|
T
0
_
l # S
0
&D
l
Q
n
H(u)&
2
du
E
|
T
0
&D&
2
op
&Q
n
H(u)&
2
du,
524 BRUCE K. DRIVER
which tends to zero by the dominated convergence theorem and the fact
that Q
n
converges strongly to zero.
C
n
-term,
E &C
n
(i)&
2
=E
|
T
i
_
l # S
0
&F
n
D
Q
n
l
H(u)&
2
du
E
|
T
0
_
l # S
0
&D
Q
n
l
H(u)&
2
du
=E
|
T
0
&(Q
ir
n
I ) DH(u)&
2
du, (6.17)
where Q
ir
n
: H
0
(q)*H
0
(q)* is the transpose of Q
n
. Since Q
n
is an
orthogonal projection operator, it easily follows that Q
n
is unitarily equiv-
alent to Q
ir
n
under the natural unitary isomorphism between H
0
(q) and
its dual H
0
(q)*. In particular Q
ir
n
0 strongly as n and hence
(Q
ir
n
I ) 0 strongly as n , see the proposition on p. 299 of Reed and
Simon [23]. So again by the dominated convergence theorem, it follows
from Eq. (6.17) that lim
n
E &C
n
(i)&
2
=0.
E
n
-term,
E &E
n
(i)&
2
(T&i) &A
(1)
n
&F
n
A
(1)
F
n
&
2
op
E
|
T
i
&H(u)&
2
du
=T &A
(1)
n
&F
n
A
(1)
F
n
&
2
op
|
T
i
&h&
2
du
T
2
&A
(1)
n
&F
n
A
(1)
F
n
&
2
op
&h&
2
,
which tends to zero as n by Theorem 5.8.
1
n
-term,
E &1
n
(i)&
2
(T&i) &F
n
A
(1)
&
2
op
E
|
T
i
&Q
n
H(u)&
2
du
T &A
(1)
&
2
op
E
|
T
0
&Q
n
H(u)&
2
du,
which again tends to zero as n because of the dominated convergence
theorem and the fact that Q
n
is strongly convergent to the zero.
Combining the above four estimates with the expression for z yields
E &z
n
(i)&
2
2E &
n
(i)+D
n
(i)&
2
+c
n
4E &
n
(i)&
2
+4E &D
n
(i)&
2
+c
n
4E
"
|
T
i
F
n
D
d[
n
z
n
"
2
+E
"
|
T
i
A
(1)
n
z
n
(u) du
"
2
+c
n
=4 _
l # S
0
|
T
i
E &F
n
D
F
n
l
z
n
(u)&
2
du+E
"
|
T
i
A
(1)
n
z
n
(u) du
"
2
+c
n
4 &D&
2
op
|
T
i
E &z
n
(u)&
2
du+(T&i) E
|
T
i
&A
(1)
n
z
n
(u)&
2
du+c
n
(4 &D&
2
op
+T sup
n
&A
(1)
n
&
2
op
) E
|
T
i
&z
n
(u)&
2
du+c
n
,
where c
n
denotes a generic sequence of positive numbers with lim
n
c
n
=0. By Theorem 5.8, sup
n
&A
(1)
n
&
2
op
< and hence the proof of the Lemma
may be concluded with an application of Gronwall's inequality. Q.E.D.
Froof of Theoren 6.2. Let i>0 and f # FC
(L). Choose a partition

P of [0, 1] so that f =1b
P
for some C
1
-function 1 on G
P
. Let P
0
be a
partition which refines P (i.e., P/P
0
) and for the moment assume that
h # H
P
0
(q).
Let P
n
be a sequence of partitions such that P
0
/P
n
/P
n+1
for all
n=1, 2, ... . Let : and :
n
be the random variables as in Theorem 6.4, see
Eqs. (6.11) and (6.12). By Eq. (6.1), with P replaced by P
n
,
E[(hf )(2
T
)]=E[ f (2
T
) :
n
] (6.18)
holds for all n. By Theorem 6.4, we may let n in (6.18) to conclude
that
E[(hf )(2
T
)]=E[ f (2
T
) :]
=E
_
f (2
T
)
|
T
0
(l(i)&
1
2
l(i) Ric] U(i, T) h,

d[(i))
&
. (6.19)
By Remark 6.3 and the fact that f is bounded, the right hand side of (6.19)
is continuous in h # H
0
(q). Similarly, since the H
0
(q) norm is stronger than
the supremum norm and df is bounded, it follows that the left-hand-side of
Eq. (6.18) is also continuous in h # H
0
(q). The continuity of both sides of
Eq. (6.19), coupled with the fact that the span of the union of H
P
0
(q) over
all finite partitions P
0
of [0, 1] which refine P is dense in H
0
(q), implies
that (6.19) is valid for all h # H
0
(q). Q.E.D.
526 BRUCE K. DRIVER
Corollary 6.6. 1or each h # H
0
(q), ihe differeniial operaior h wiih
donain FC
(L(G)) is a densely defined closable operaior on 1

2
(L(G), v
T
).
Moreoter ihe 1
2
-adjoini of h* of h saiisfies
h*|
FC
(L(G))
=&h+z
h
, (6.20)
where z
h
: L(G) R is a Borel neasurable funciion such ihai
z
h
(2
T
)#
1
T
E
_
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+ }
o(2
T
)
&
a.s. (6.21)
Froof. Let u, t # FC
(L) and l(i)=i}T. Then apply Theorem 6.2 with

f # FC
(L) replaced by ut # FC
(L) to find
E[(hu)(2
T
) t(2
T
)+u(2
T
)(ht)(2
T
)]=E[u(2
T
) z
h
(2
T
)]
or equivalently
|
L(G)
(hu)( g) t( g) dv
T
( g)=
|
L(G)
u( g)[&(ht)( g)+z
h
( g) t( g)] dv
T
( g).
This proves Eq. (6.20) and the fact that h is closable, since the properties
of being closable and having a densely defined adjoint are equivalent.
Q.E.D.
7. QUASI-INVARIANCE OF THE HEAT KERNEL MEASURE
In this section we will show that the measure v
T
#Law(2
T
) is quasi-
invariant under both right and left translations by finite energy paths in
L(G), see Corollary 7.7 and 7.10 below. Our method will be modeled on
a technique in Cruzeiro [3] (see also Dennis Bell [1] and Gunnar Peters
[20, 21]) for proving quasi-invariance of flows of certain vector fields on
an abstract Wiener space. In order to carry out the proof it is necessary to
recall a few results from the finite dimensional case.
7.1. 1iniie Dinensional Frelininaries
Let M be a finite dimensional manifold and X be a smooth complete
vector field on M. We will denote the flow of X by e
iX
]
i # R
so that
e
iX
: MM is a diffeomorphism for all i # R such that e
0X
=id
M
and
de
iX
}di=Xb e
iX
. Suppose that o is Borel measure on M such that, in every
coordinate chart, o has a smooth positive density relative to Lebesgue
measure. Then the standard change of variable theorem guarantees that
e
*
iX
o#ob e
&iX
is absolutely continuous relative to o and that the Radon-
Nikodym derivative Z
i
(n)#(de
*
iX
o}do)(n) may be chosen to be a smooth
positive function of (i, n) # R_M.
Definition 7.1 (Divergence). The ditergence div
o
(X) of X relaiite io o
nay be defined as
div
o
(X)=&
d
di }
0
Z
i
. (7.1)
(The reason for ihe ninus sign is io adhere io ihe siandard sign conteniions
for ihe ditergence defined by oiher neans.)
The following proposition summarizes some well known properties of Z
i
and div
o
(X).
Proposition 7.2. 1ei X, Z
i
, and div
o
(X) be as abote.
1. Suppose ihai B/M is a Borel subsei of M such ihai B is conpaci,
ihen
d
di }
0
o(e
iX
(B))=
|
B
div
o
(X) do. (7.2)
So div
o
(X) neasures ihe raie of spreading of ihe flow e
iX
as seen by ihe
neasure o.
2. 1iewing X as a firsi order differeniial operaior, for f # C
1
(M), lei
X*f =&Xf &div
o
(X) f. (7.3)
Then for all f, g # C
1
(M) such ihai ihe produci fg has conpaci suppori,
|
M
(Xf ) g do=
|
M
f (X*g) do. (7.4)
3. The Fadon-Nikodyn deritaiite Z
i
nay be recotered fron ihe flow
e
iX
and div
o
(X) by ihe fornula
Z
i
(n)=e
&
i
0
div
o
(X) b e
&tX
(n) dt
(\n # M). (7.5)
Froof. The key point is that for f: MR, bounded and measurable
with compact support, we have by the definition of Z
i
that
|
M
f b e
iX
do=
|
M
f Z
i
do.
528 BRUCE K. DRIVER
Hence
d
di
|
M
f b e
iX
do=
|
M
f
Z
i
i
do. (7.6)
Taking f to be the characteristic function of B and i=0 in Eq. (7.6) implies
Eq. (7.2). If we replace f by fg, where f, g # C
1
(M) such that fg has compact
support, then differentiating under the integral sign in Eq. (7.6) implies
|
M
X( f b e
iX
) } g b e
iX
+f b e
iX
X( g b e
iX
)] do=
|
M
fg
Z
i
i
do. (7.7)
Taking i=0 implies Eq. (7.3). Now suppose that g#1 in Eq. (7.7), then
|
M
f
Z
i
i
do=
|
M
X( f b e
iX
) } 1 do=
|
M
( f b e
iX
) } X*1 do
=&
|
M
( f b e
iX
) } div
o
X do
=&
|
M
( f b e
iX
) } (div
o
X) b e
&iX
b e
iX
do
=&
|
M
f } (div
o
X) b e
&iX
Z
i
do.
Since f # C
1
c
(M) is arbitrary in this last equation, Z
i
must satisfy the
differential equation:
Z
i
i
=&Z
i
} (div
o
X) b e
&iX
with Z
0
#1.
The unique solution of this equation is given in Eq. (7.5). Q.E.D.
For the infinite dimensional application to the loop group, it will be
necessary to recall the following key estimate of Ana Bela Cruzeiro
(see Corollary 2.2 in [2]) for the 1
p
-norms of Z
i
in terms of div
o
X. For
the readers convenience I will also give the short proof.
Theorem 7.3 (Cruzeiro). 1ei p # (1, ) and M, X, o, Z
i
and z#
&div
o
X be as abote. ssune now ihai o is probabiliiy neasure and wriie
E
o
f for
M
f do. If for a giten T>0, I(T)#sup
|t| T
E
o
[Z
p
t
]<, ihen for
all i # [&T, T];
E
o
[Z
p
i
] sup
|s| |i|
E
o
exp
{
p
2
p&1
sz
=
E
o
exp
{
p
2
p&1
|i| |z|
=
. (7.8)
Froof. Let J
s
#[0, s] if s0 and J
s
#[s, 0] if s<0. Using Jensen's
inequality, Eq. (7.5), the definition of Z
s
and Holder's inequality we find for
|s| |i| T that
E
o
[Z
p
s
]=E
o
exp
\
p
|
s
0
z b e
&tX
dt
+
=E
o
exp
\
ps
|
J
s
z b e
&tX
dt
|s| +
E
o
|
J
s
dt
|s|
exp( psz b e
&tX
)=
|
J
s
dt
|s|
E
o
(e
psz
} Z
&t
)
|
J
s
dt
|s|
(E
o
e
pqsz
)
1}q
(EZ
p
&t
)
1}p
(E
o
e
pqsz
)
1}q
I(i)
1}p
,
where 1}q+1}p=1. Hence it follows that
I(i) sup
|s| |i|
(E
o
e
pqsz
)
1}q
I(i)
1}p
.
Solving this equation for I(i) shows that
I(i) sup
|s| |i|
(E
o
e
pqsz
)= sup
|s| |i|
E
o
exp
{
p
2
p&1
sz
=
E
o
exp
{
p
2
p&1
|i| |z|
=
.
Q.E.D.
7.2. Quasi-Intariance for ihe Heai Kernel Measure on L(G)
Let 2 be the L(G)-valued Brownian motion constructed in Theorem 3.8
with 2(0, s)#e for 0s1 and v
T
#Law(2
T
). For h # H
0
(q), let
z
h
: L(G) R be a Borel measurable function as in Corollary 6.6. By
taking u=f # FC
(L) and t=1 in Corollary 6.6 we find

E
T
(hf )=E[(hf )(2
T
)]=E[ f (2
T
) z
h
(2
T
)]=E
v
T
( fz
h
). (7.9)
Since the flow of the vector field h is e
ih
( g)=ge
ih
, Eq. (7.9) and the finite
dimensional discussion above motivates the following theorem.
Theorem 7.4. 1ei h # H
0
(q), ihen v
T
is quasi-intariani under ihe irans-
fornaiion g # L(G) ge
h
# L(G). Moreoter, lei D
0
denoie ihe sei of loops
g # L(G) such ihai
1
0
|z
h
( ge
&uh
)| du<, ihen v
T
(D
0
)=1, ihe funciion
Z
h
: L(G) R defined by
Z
h
( g)=1
D
0
( g) exp
\
|
1
0
z
h
( ge
&uh
) du
+
(7.10)
530 BRUCE K. DRIVER
is in 1
1
(dv
T
) and
|
L(G)
f ( ge
h
) dv
T
( g)=
|
L(G)
f ( g) Z
h
( g) dv
T
( g)
for all bounded neasurable funciions f on L(G). (This lasi equaiion nay
also be equitalenily e\pressed as E[ f (2
T
e
h
)]=E[ f (2
T
) Z
h
].)
Fenark 7.5. For each finite partition of P of [0, 1],
P
*
v
T
is the
smooth measure on G
P
given by (d
P
*
T
}dz
P
)(\)=p
P
T
(e, \), where \ # G
P
,
p
G
T
is the heat kernel on G
P
associated Riemannian inner product ( } , } )
P
,
i.e., p
P
T
is the integral kernel of the operator e
T A
P
}2
.
Froof of Theoren 7.4. Let P
0
/P
1
/P
2
} } } be a nested sequence of
partitions of [0, 1] such that lim
n
|P
n
| =0. Suppose n
0
# 0, 1, 2, ...]
and f: LR is function such that f =1b
P
n
0
for some bounded Borel
measurable function 1: G
P
n
0 R. Let G
n
be the smallest o-algebra on L(G)
such that the projection
P
n
: L(G) G
P
n
is measurable, where G
P
n
is
given the Borel o-algebra. Set
z`
h, n
#&div
P
n
V v
T
h |
G
P
n
and z
h, n
#z`
h, n
b
P
n
. Then by Proposition 7.2 and Corollary 6.6,
z
h, n
(2
T
)#
1
T
E
_
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+ }
G
n
&
v-a.e. (7.11)
Therefore, by Proposition 7.2 and Remark 7.5,
E[ f (2
T
e
h
)]=E[ f (2
T
) Z
h, n
], (7.12)
where
Z
h, n
#exp
{
|
1
0
z
h, n
(2
T
e
&sh
) ds
=
. (7.13)
The proof of Theorem 7.4 will continue after the following key lemma.
Lemma 7.6. 1ei Z
h, n
be defined as in Eq. (7.13). Then for all p # (1, ),
sup
n
E
v
T
Z
p
h, n
exp
1
2 \
p
4
&h&
2
( p&1)
2
T
2
|
T
0
"
I&
1
2
t Ric
"
2
op
dt
+
=: M( p, h)<, (7.14)
and Z
h, n
]
n=1
is Cauchy in 1
p
(v
T
).
Froof. Since G
P
n
is compact and g # G
P
n
exp
1
0
z`
h, n
( ge
&sh
) ds is
smooth, Cruzeiro's Theorem 7.3 may be applied to show
E
v
T
Z
p
h, n
sup
|s| 1
E
v
T
_
exp
{
p
2
p&1
sz
h, n
=&
= sup
|s| 1
E
v
T
\
exp
{
p
2
p&1
s
T
E
v
T
_
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+ }
G
n
&=+
sup
|s| 1
E
v
T
\
E
v
T
_
exp
{
p
2
p&1
s
T
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+= }
G
n
&+
= sup
|s| 1
E
v
T
_
exp
{
p
2
p&1
s
T
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+=&
, (7.15)
wherein the second inequality we use Jensen's inequality. For fixed
s # [&1, 1] and T as above, set
M
i
#
p
2
p&1
s
T
|
T
T&i
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+
for 0iT.
Then M
i
is a martingale such that the quadratic variation of M at T is
given by
OMo
T
=
p
4
s
2
( p&1)
2
T
2
|
T
0
"{
I&
1
2
t Ric
=
H(t)
"
2
dt.
and thus, because &H(t)&=&h&,
&OMo
T
&
1
(v)
p
4
&h&
2
s
2
( p&1)
2
T
2
|
T
0
"
I&
1
2
t Ric
"
2
op
dt<. (7.16)
Hence, Novikov's criterion (see Proposition 1.15, p. 308 in [24]) implies
that Ee
M
T
&(1}2)M,
T
=1 so that
Ee
M
T
=E(e
M
T
&(1}2)M,
T
} e
(1}2)M,
T
)e
(1}2) &M,
T
&
1
exp
1
2 \
s
2
p
4
&h&
2
( p&1)
2
T
2
|
T
0
"
I&
1
2
t Ric
"
2
op
dt
+
. (7.17)
(Alternatively, see Lemma 1.4 in Kusuoka and Stroock [16].) Combining
equations (7.15) and (7.17) proves the bound in Eq. (7.14).
532 BRUCE K. DRIVER
Let n>n be two positive integers. By the fundamental theorem of
calculus, for all \, y # R,
e
y
&e
\
=( y&\)
|
1
0
e
(uy+(1&u) \)
du=( y&\)
|
1
0
(e
y
)
u
(e
\
)
1&u
du.
Applying this equation with \=
1
0
z
h, n
(2
T
e
&sh
) ds and y=
1
0
z
h, n
(2
T
e
&sh
) ds gives
|Z
h, n
&Z
h, n
| =
}
|
1
0
z
h, n
(2
T
e
&sh
)&z
h, n
(2
T
e
&sh
) ds
}
}
|
1
0
Z
u
h, n
Z
(1&u)
h, n
du.
Using Holder's inequality we find
E |Z
h, n
&Z
h, n
|
|
1
0
&z
h, n
(2
T
e
&sh
)&z
h, n
(2
T
e
&sh
)&
1
3}2 ds
|
1
0
&Z
u
h, n
Z
(1&u)
h, n
&
1
3 du.
(7.18)
Now by the bound in Eq. (7.14)
&Z
u
h, n
Z
(1&u)
h, n
&
1
3&Z
u
h, n
&
1
6 &Z
(1&u)
h, n
&
1
6M(6, h)
1}3
,
which combined with Eq. (7.18) shows that
E |Z
h, n
&Z
h, n
| M(6, h)
1}3
|
1
0
&z
h, n
(2
T
e
&sh
)&z
h, n
(2
T
e
&sh
)&
1
3}2 ds.
(7.19)
Since n>n,
&z
h, n
(2
T
e
&sh
)&z
h, n
(2
T
e
&sh
)&
1
3}2
=[E( |z
h, n
(2
T
)&z
h, n
(2
T
)|
3}2
Z
&sh, n
)]
2}3
=&|z
h, n
(2
T
)&z
h, n
(2
T
)| (Z
&sh, n
)
2}3
&
1
3}2
&z
h, n
(2
T
)&z
h, n
(2
T
)&
1
2 } &(Z
&sh, n
)
2}3
&
1
6
(M(4, h))
1}6
} &z
h, n
(2
T
)&z
h, n
(2
T
)&
1
2 .
This equation and Eq. (7.19) shows that
E |Z
h, n
&Z
h, n
| M(6, h)
1}3
(M(4, h))
1}6
} &z
h, n
(2
T
)&z
h, n
(2
T
)&
1
2
(v)
.
(7.20)
Now
&z
h, n
(2
T
)&z
h, n
(2
T
)&
1
2
#
1
T "
(E
n
&E
n
)
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+"
1
2
, (7.21)
where E
n
#E[ } | G
n
] denotes conditional expectation relative to the
o-algebra G
n
. Since E
n
converges strongly to E[ } | o(2
T
)] in 1
2
it follows
from Eq. (7.21) that
lim
n, n
&z
h, n
(2
T
)&z
h, n
(2
T
)&
1
2=0.
In view of Eq. (7.20) this finishes the proof of the lemma. Q.E.D.
We now continue the proof of Theorem 7.4. Let Z
h
be the 1
p
(v
T
)-limit
of Z
h, n
. Of course Z
h
inherits the bounds in Eq. (7.14), namely that
E
v
T
Z
p
h
M( p, h)< for all p # (1, ).
By the previous lemma, we may let n tend to infinity in Eq. (7.12) to find
E[ f (2
T
e
h
)]=E[ f (2
T
) Z
h
(2
T
)]. (7.22)
Eq. (7.22) is valid for all f # FC
(L) which are based on P

n
for some
positive integer n. So by a monotone class argument or Dynkin's &z
theorem, one may easily show that this equation is in fact valid for all
bounded measurable functions on L(G).
Setting o#v
T
, and o
h
#F
e
h
V
v
T
, we have shown that o
h
<<o and that
do
h
}do=Z
h
. We now show that o<<o
h
. To this end let f : L(G) R be
a bounded measurable function, then
E
o
( f )=Ef (2
T
)=Ef (2
T
e
&h
e
h
)
=E[ f (2
T
e
h
) Z
&h
(2
T
)]=E[ f (2
T
e
h
) Z
&h
(2
T
e
h
e
&h
)]
=E
o
h
( f Z
&h
(( } ) e
&h
)).
Therefore o<<o
h
and do}do
h
=Z
&h
(( } ) e
&h
).
534 BRUCE K. DRIVER
So to finish the proof we need only show that Z
h
defined in Eq. (7.10)
is well defined and Z
h
(2
T
)=Z
h
v-a.e. First consider
|
L(G)
\
|
1
0
|z
h
( ge
&uh
)| du
+
2
dv
T
( g)=E
\
|
1
0
|z
h
(2
T
e
&uh
)| du
+
2
|
1
0
E |z
h
(2
T
e
&uh
)|
2
du
=
|
1
0
E( |z
h
(2
T
)|
2
Z
&uh
(2
T
)) du
(E |z
h
(2
T
)|
4
)
1}2
M(2, h)<.
(Note that E |z
h
(2
T
)|
4
< because of Remark 6.3 and the fact that
conditional expectations are contractions on 1
p
-spaces.) This shows that
1
0
|z
h
(ge
&uh
)| du< for v
T
-a.e. g and hence that Z
h
is well defined.
Set _#
1
0
z
h
(2
T
e
&uh
) du and _
n
#
1
0
z
h, n
(2
T
e
&uh
) du. Since Z
h
(2
T
)=e
_
and Z
h
(2
T
)=lim
n
e
_
n
, to show that Z
h
(2
T
)=Z
h
(2
T
) a.e. it suffices to
show that _
n
converges to _ in 1
1
. We start with the estimate
E |_&_
n
|
|
1
0
E |z
h, n
(2
T
e
&uh
)&z
h
(2
T
e
&uh
)| du
=
|
1
0
E[|z
h, n
(2
T
)&z
h
(2
T
)| Z
&uh
(2
T
)] du
&z
h, n
(2
T
)&z
h
(2
T
)&
1
2
|
1
0
&Z
&uh
(2
T
)&
1
2
(F)
du
&z
h, n
(2
T
)&z
h
(2
T
)&
1
2
(F)
-M(2, h).
Now by Eq. (7.11),
&z
h, n
(2
T
)&z
h
(2
T
)&
1
2
(F)
#
1
T "
(E
n
&E
)
|
T
0
\{
I&
1
2
t Ric
=
H(t),

d[(t)
+"
1
2
(F)
, (7.23)
where E
n
#E[ } | G
n
] and E
( } )#E[ } | o(2
T
)]. This finishes the proof
because E
n
converges strongly in 1
2
to E
as n . Q.E.D.
We now wish to extend Theorem 7.4 to include right translations by
k # L
1
0
(G), where L
1
0
(G) denotes the space of contractible loops in L(G)
which have finite energy. A loop k # L(G) is said to have finiie energy
provided k is absolutely continuous and
|
1
0
|0k$(s),|
2
ds=
|
1
0
|1
k(s)
&1
V
k$(s)|
2
ds<.
Corollary 7.7. 1or each k # L
1
0
(G), v
T
quasi-intariani under ihe righi
iranslaiion nap F
k
: L(G) L(G) giten by F
k
g=gk.
We will need the following simple lemma for the proof of this corollary.
Lemma 7.8. 1ei (L, G, ) be a probabiliiy space and T: LL be an
interiible neasurable nap wiih a neasurable interse. ssune ihai T
*
#
b T
&1
, T
*
&1
#b T, and are nuiually absoluiely coniinuous. lso lei
f: L[0, ) be a neasurable funciion such ihai f >0 -a.s. and
L
f du=1. If f denoies ihe probabiliiy neasure defined by ( f)()#
f d for all # G, ihen T

*
( f)#( f) b T
&1
, T
*
&1
( f)#( f) b T, and
are all nuiually absoluiely coniinuous as well.
Froof. Let Z=dT
*
}d and g: L[0, ) be a measurable function.
Then
|
L
g d[T
*
( f))]=
|
L
g b T } f d=
|
L
( g } f b T
&1
) b T d
=
|
1
g } ( f b T
&1
) Z d.
This shows that
d[T
*
( f))]
d
=Z } ( f b T
&1
). (7.24)
Since ( f
&1
(0]))=0 and and T
*
&1
are mutually absolutely con-
tinuous, we have 0=(b T)( f
&1
(0]))=(g # L: f b T
&1
( g)=0]). That
is f b T
&1
is positive -a.s. Thus it follows from Eq. (7.24) that T
*
( f) and
are mutually absolutely continuous. By symmetry, T
*
&1
( f) and are
mutually absolutely continuous as well. Q.E.D.
Froof of Corollary 7.7. By Theorem 7.4 and repeated use of Lemma
7.8, it suffices to prove: for any k # L
1
0
(G) there is an integer n and
h
i
# H
0
(q) such that
k(s)=e
h
1
(s)
e
h
2
(s)
} } } e
h
n+1
(s)
. (7.25)
536 BRUCE K. DRIVER
To prove (7.25), choose a ball B in q centered at 0 such that
1 :=e
_
: _ # B] is open and the map _ # Be
_
# 1 is a diffeomorphism
with inverse denoted by log. Let L(1)=g # L(G) : g([0, 1])/1], so
that L(1) is an open neighborhood of L
0
(G). It is easily shown that
#
n=1
L(1)
n
is both open and closed in L(G) (with the sup-norm
topology) and hence =L
0
(G)-the connected component of the identity
in L(G). (The space L
0
(G) may also be described as the space of contrac-
tible loops in L(G).) Therefore there is an integer n # Z
+
and k
i
# L(1)
such that k=k
1
k
2
} } } k
n
. Let u
i
(s)#log k
i
(s), then u
i
# L(q) and
e
&u
n
e
&u
n&1
} } } e
&u
1
k is the constant path sitting at e # G. Choose h
i
# H
0
(q)
sufficiently close to u
i
in the sup-norm topology on L(q) such that
e
&h
n
e
&h
n&1
} } } e
&h
1
k # L(1), Define h
n+1
#log(e
&h
n
e
&h
n&1
} } } e
&h
1
k) # H
0
(q).
Then
e
h
n+1
=e
&h
n
e
&h
n&1
} } } e
&h
1
k,
which is equivalent to (7.25). Q.E.D.
Proposition 7.9. The heai kernel neasure v
T
is intariani relaiite io ihe
interse nap g # L(G) g
&1
# L(G).
Froof. It suffices to show that each of the finite dimensional distribu-
tions,
P
*
v
T
#v
T
b
&1
P
(where P is a finite partition of [0, 1]) is invariant
under the inverse map g # G
P
g
&1
# G
P
. But this property is know to
hold, in general, for heat kernel measures on uni-modular Lie groups
equipped with a left invariant Riemannian metric, see for example
Remark 2.2 and Proposition 3.1 in Driver and Gross [7] Q.E.D.
Corllary 7.10. 1or each k # L
1
0
(G), v
T
is quasi-intariani under ihe lefi
iranslaiion nap 1
k
: L(G) L(G) giten by 1
k
g=kg.
Froof. This a direct consequence of Corollary 7.7 and Proposition 7.9
above. Indeed, let f : LR be a bounded and measurable function and for
each k # L
1
0
(G) set Z
k
#dv
T
b F
&1
k
}dv
T
. Then
Ef (k2
T
)=Ef (k2
&1
T
)=Ef ((2
T
k
&1
)
&1
)
=E( f (2
&1
T
) Z
k
&1(2
T
))=E( f (2
T
) Z
k
&1(2
&1
T
)).
This shows that v
T
b 1
&1
k
<<v
T
and dv
T
b 1
&1
k
}dv
T
( g)=Z
k
&1( g
&1
) for v
T
almost every g # L(G). Since Z
k
&1>0 v
T
-a.s. and g Z
k
&1( g
&1
) has the
same distribution as Z
k
&1 , it follows that dv
T
b 1
&1
k
}dv
T
>0 v
T
-a.s. Hence v
T
is absolutely continuous relative to v
T
b 1
&1
k
as well. Q.E.D.
8. APPENDIX: REVIEW OF THE ITO INTEGRAL IN
INFINITE DIMENSIONS
As in the body of the text, let (W, F, F
i
]
i0
, F) be a filtered probab-
ility space satisfying the usual hypothesis as described in the beginning of
Section 3.1. The purpose of this appendix is to set up notation and review
some very basic facts about Hilbert space valued martingales and the Ito^
integral 1 d[. For Hilbert space martingale theory the reader is referred
to Me tivier [18]. For the Ito^ integral on abstract Wiener space, see
Sections III.5 of Kuo [15], p. 188207, especially Theorem 5.1 of [15].
Also see Kusuoka and Stroock [16] p. 5 for a very short description of the
Ito^ integral in this context. For the notion and basic properties of condi-
tional expectations for Banach space valued Random variables, see
Section 8.3 in Chapter 2 in Me tivier [18].
8.1. Coniinuous Hilberi 1alued 1ocal Mariingales
Let K be a Hilbert space. We will use ( } , }) to denote the inner product
on both of the Hilbert space H
0
(q) and K.
Theorem 8.1 (Quadratic Variations). Suppose ihai M and N are iwo
coniinuous local nariingales wiih talues in a Hilberi space K. Then ihere is
a real talued process of bounded tariaiion OM, No such ihai for any
increasing sequence of pariiiions
n
]
n=1
of [0, ) such ihai |
n
| 0 as
n,
OM, No
i
= lim
n
_
i
i
#
n
(M
i .i
i+1
&M
i .i
i
, N
i .i
i+1
&N
i .i
i
), (8.1)
where ihe linii e\isis in probabiliiy unifornly for i in conpaci subseis of
[0, ). Moreoter ihe following properiies hold:
1. |OM, No| -OMo} ONo a.s., where OMo#OM, Mo.
2. |OMo&ONo| -OM&No} OM+No a.s.
3. EM
i
*
2
4E &M
0
&
2
+4EOMo
i
.
4. The following ihree condiiions are equitaleni:
(a) M is a square iniegrable nariingale,
(b) EM
i
*
2
< for all i0 and
(c) E &M
0
&
2
< and EOMo(i)< for all i0.
538 BRUCE K. DRIVER
5. If M
n
and M are coniinuous K-talued 1
2
-nariingales such ihai
E &(M
n
&M)
i
&
2
0 as n ihen OM
n
o(i) OMo(i) in 1
1
.
6. Suppose ihe M
n
]
n=1
is a sequence of K-talued coniinuous local
nariingales such ihai M
n
(0)=0 for all n and OM
n
o0 a.s. as n .
Then M
n
0 in probabiliiy unifornly on conpaci subseis of [0, ).
For a proof of this theorem see, for example, Theorems 20.5 and 20.6 in
Me tivier [18] and Me tivier and Pellaumail [19].
8.2. The Iio` Iniegral on Our bsiraci iener Space
For the rest of this Appendix we will adopt the notation in Section 3.1
of the body of the paper.
Theorem 8.2. Suppose ihai f
i
]
i0
is an (W, F
i
]
i0
, F, F)-adapied
and coniinuous process wiih talues in H
0
(q). Then ihere is a coniinuous
local nariingale N such ihai for any orihonornal basis h
n
]
n=1
of H
0
(q),
N=_
n=1
( f
i
, h
n
) d[
h
n
, where ihe sun is contergeni in probabiliiy
unifornly for i in conpaci subseis of [0, ). e will wriie N
i
as
i
0
f d[ or
N= f d[ for shori. The quadraiic tariaiion of f d[ is giten by
O f d[o
i
=
i
0
& f
t
&
2
dt or O f d[o= & f &
2
dt for shori.
Froof. Let k # Z
+
and N
(k)
#_
k
n=1
( f
i
, h
n
) d[
h
n
a local martingale.
Then for k$>k,
ON
(k$)
&N
(k)
o=
|
_
k$
n=k+1
|( f
i
, h
n
)|
2
dt 0 a.s. as k, k$ .
Using theorem 8.1, this shows that N
(k)
converges uniformly on compacts
in probability to a local martingale N and moreover
ONo=
|
_
n=1
|( f
t
, h
n
)|
2
dt=
|
& f
t
&
2
dt.
Now suppose that l
n
]
n=1
is another orthonormal basis for H
0
(q) and
that
Q
(k)
# _
k
n=1
|
( f
i
, l
n
) d[
l
n
.
Let F
k
and F
k
be orthogonal projections onto spanh
1
, h
2
, ..., h
k
] and
spanl
1
, l
2
, ..., l
k
] respectively. Then
OQ
(k)
&N
(k)
o=O _
k
n=1
{
|
( f
i
, h
n
) d[
h
n
&
|
( f
i
, l
n
) d[
l
n
=
o
= _
k
n, n=1
O
{
|
( f
i
, h
n
) d[
h
n
&
|
( f
i
, l
n
) d[
l
n
=
,
{
|
( f
i
, h
n
) d[
h
n
&
|
( f
i
, l
n
) d[
l
n
=
o
=_
k
n
|
|( f
i
, h
n
)|
2
+|( f
i
, l
n
)|
2
] di
&2 _
k
n, n=1
|
( f
i
, h
n
)( f
i
, l
n
)(h
n
, l
n
) di
=
|
&F
k
f
i
&
2
+&F
k
f
i
&
2
&2(F
k
f
i
, F
k
f
i
)] di 0
a.s. as k ,
where we have used the fact that F
k
and F
k
are strongly convergent to I
as k along with the dominated convergence theorem. This shows that
f d[ is basis independent. Q.E.D.
Theorem 8.3 (Associativity). Suppose f
i
]
i0
and g
i
]
i0
are (W,
F
i
]
i0
, F, F)-adapied and coniinuous process wiih talues in H
0
(q) and R
respeciitely. Sei M# ( f, d[), ihen
|
g dM=
|
( gf, d[). (8.2)
Froof. Let h
n
]
=1
be orthonormal basis for H
0
(q) and for N # Z
+
set
M
N
# _
N
n=1
|
( f, h
n
) d[
h
n
.
Then
O
|
gdM&
|
g dM
N
o=
|
g
2
dOM&M
N
o=
|
g
2
\
_
n=N+1
|( f, h
n
)|
2
+
di,
540 BRUCE K. DRIVER
and this last expression tends to zero almost surely as N. Therefore
g dM
N
g dM uniformly on compacts in probability. On the other
hand, by associativity of the finite dimensional Ito^ integral,
|
g dM
N
= _
N
n=1
|
gd
\
|
( f, h
n
) d[
h
n
+
= _
N
n=1
|
g( f, h
n
) d[
h
n
= _
N
n=1
|
( gf, h
n
) d[
h
n
|
( gf, d[) as N.
So we have shown that g dM
N
converges to both g dM and ( gf, d[).
Q.E.D.
Theorem 8.4. Suppose ihai K is anoiher separable Hilberi space,
1
i
]
i0
is an F
i
-adapied and coniinuous process wiih talues in
HS(H
0
(q), K)ihe Hilberi Schnidi operaiors fron H
0
(q) io K. Then ihere
is a coniinuous K-talues local nariingale N such ihai, for any orihonornal
basis k
n
]
n=1
of K, _
n=1
(1*k
n
, d[) k
n
conterges unifornly on conpacis
in probabiliiy io N. e will wriie N
i
as
i
0
1 d[ or N= 1 d[ for shori.
The quadraiic tariaiion of 1 d[ is giten by O 1 d[o
i
=
i
0
&1&
2
HS
dt or
O 1 d[o= &1&
2
HS
dt.
Froof. Let k # Z
+
and N
(k)
#_
k
n=1
(1*k
n
, d[) k
n
a K-valued local
martingale. Then for k$>k,
ON
(k$)
&N
(k)
o=O
|
_
k$
n=k+1
|
(1*k
n
, d[) k
n
o
= _
k$
n, n=k+1
O
|
(1*k
n
, d[),
|
(1*k
n
, d[)o(k
n
, k
n
)
= _
k$
n=k+1
|
&1*k
n
&
2
HS
dt.
Recall that
&1&
2
HS
= _
n=1
&1h
n
&
2
= _
n=1
_
n=1
|(1h
n
, k
n
)|
2
= _
n=1
_
n=1
| (h
n
, 1*k
n
)|
2
=&1*&
2
HS
.
Therefore, _
n=1
&1*k
n
&
2
dt< and hence,
ON
(k$)
&N
(k)
o= _
k$
n=k+1
|
&1*k
n
&
2
dt 0 as k, k$ .
Again by Theorem 8.1, this shows that N
(k)
converges uniformly on
compacts in probability to a K-valued local martingale N and also that
ONo=
|
&1
t
&
2
HS
dt.
Now suppose that l
n
]
n=1
is another orthonormal basis for H
0
(q) and
that
Q
(k)
# _
k
n=1
|
(1*l
n
, d[) l
n
.
Again let F
k
be orthogonal projection onto span k
1
, k
2
, ..., k
k
]. Then
OQ
(k)
&N
(k)
o=O _
k
n=1
{
|
(1*k
n
, d[) k
n
&
|
(1*l
n
, d[) l
n
=
o
=O _
k
n=1
|
(1*k
n
, d[) k
n
o+O _
k
n=1
|
(1*l
n
, d[) l
n
o
&2 _
k
n=1
_
k
n=1
O
|
(1*k
n
, d[) k
n
,
|
(1*l
n
, d[) l
n
o
= _
k
n=1
|
&1*k
n
&
2
dt+ _
k
n=1
|
&1*l
n
&
2
dt
&2
|
G
k
dt, (8.3)
where
G
k
# _
k
n=1
_
k
n=1
(1*k
n
, 1*l
n
)(k
n
, l
n
)= _
k
n=1
(1*F
k
l
n
, 1*l
n
).
Let F$
k
#I&F
k
and notice that
|G
k
|
n=1
&1*F
k
l
n
&
2
n=1
&1*l
n
&
2
=&1*F
k
&
HS
} &1*&
HS
&1&
2
HS
542 BRUCE K. DRIVER
and
}
G
k
& _
k
n=1
(1*l
n
, 1*l
n
)
}
_
k
n=1
| (1*F$
k
l
n
, 1*l
n
)|
_
k
n=1
&1*F$
k
l
n
&
2
}
_
k
n=1
&1*l
n
&
2
&1*F$
k
&
HS
} &1*&
HS
=
n=k+1
&1*h
n
&
2
} &1&
HS
.
Therefore we are justified in applying the dominated convergence theorem
in Eq. (8.3) to find that
lim
k
OQ
(k)
&N
(k)
o=
|
&1*&
2
HS
+&1*&
2
HS
&2 &1*&
2
HS
] di=0.
Hence Q
(k)
&N
(k)
0 in probability which proves that 1 d[ is basis
independent. Q.E.D.
Theorem 8.5 (Ito^ 's Lemma). Suppose ihai K is a separable Hilberi
space, 1
i
]
i0
is an (W, F
i
]
i0
, F, F)-adapied and coniinuous process
wiih talues in HS(H
0
(q), K). Then
"
|
i
0
1 d[
"
2
=2
|
i
0
\
1
t
*
|
t
0
1 d[, d[(t)
+
+
|
i
0
&1
t
&
2
HS
dt. (8.4)
Froof. Let k
n
]
n=1
be orthonormal basis for K and M
n
#(1*k
n
, d[).
Then
"
|
1 d[
"
2
= _
n=1
\
|
(1*k
n
, d[)
+
2
= _
n=1
M
2
n
= _
n=1
{
2
|
M
n
dM
n
+OM
n
o
=
=2 _
n=1
|
(M
n
1*k
n
, d[)+ _
n=1
|
&1*k
n
&
2
dt
=2 _
n=1
|
(M
n
1*k
n
, d[)+
|
&1&
2
HS
dt, (8.5)
wherein the third equality we used the Associativity Theorem 8.3.
Let F
N
be orthogonal projection onto the subspace spanned by
k
1
, k
2
, ..., k
N
]. Then for any h # H
0
(q)
\
h, 1*
|
F
N
1 d[
+
=
\
1h,
|
F
N
1 d[
+
=
\
1h, _
n=1
k
n
|
((F
N
1)* k
n
, d[))
= _
N
n=1
(k
n
, 1h)
|
(1*k
n
, d[)= _
N
n=1
M
n
(1*k
n
, h),
and hence
_
n=1
|
(M
n
1*k
n
, d[)= lim
N
_
N
n=1
|
(M
n
1*k
n
, d[)
= lim
N
|
\
1*
|
F
N
1 d[, d[
+
. (8.6)
So to finish the proof it suffices to consider
O
|
\
1*
|
1 d[, d[
+
&
|
\
1*
|
F
N
1 d[, d[
+
o
i
=
|
i
0
"
1
t
*
|
t
0
(I&F
N
) 1 d[
"
2
dt
|
i
0
&1
t
&
2
op
"
|
t
0
(I&F
N
) 1 d[
"
2
dt. (8.7)
Since, &(I&F
N
) 1h
n
&
2
&1h
n
&
2
and _
n=1
&1h
n
&
2
=&1&
2
HS
<, it follows
by the dominated convergence theorem that
O
|
(I&F
N
) 1 d[o=
|
&(I&F
N
) 1&
2
HS
dt 0 as N, (8.8)
and hence (I&F
N
) 1 d[ converges to zero uniformly on compacts in
probability. Therefore, the right side of Eq. (8.7) tends to zero as N.
This implies that (1* F
N
1 d[, d[) (1* 1 d[, d[) in probability as
N. Equation (8.4) now follows from this limit and equations (8.5)
and (8.6). Q.E.D.
8.3. Backwards Iio` Iniegrals
Let T>0 be fixed. For the moment suppose that 1 is a finite dimen-
sional vector space, X(i)]
i0
is a continuous 1-valued process and
544 BRUCE K. DRIVER
(i)]
i0
is continuous End(1)-valued process. Let =0=i
0
<i
1
<i
1
<
i
2
< } } } <i
n
=T] denote a partition of [0, T], || # max
i
|i
i+1
&i
i
|. For
t=i
i
# , let t+#i
(i+1)
be the successor to t in . (By convention
i
n+1
#T.) Then the forward siochasiic and respectively backwards
siochasiic iniegral of relative to X is
|
i
0
dX# lim
|| 0
_
t #
(t)(X(i .(t+))&X(t .i)) (8.9)
and
|
T
i

dX# lim
|| 0
_
t #
(t+)(X(i (t+))&X(t i)), (8.10)
provided that limits exists in probability uniformly for i in [0, T]. For
example, if and X are semi-martingales then the above limit exists and
|
T
i

dX=
|
T
i
dX+
|
T
i
d dX,
where
T
i

dX=
T
0

dX&
i
0

dX and
|
i
0
d dX# lim
|| 0
_
t #
((i .t+)&(t))(X(i .(t+))&X(t .i)) (8.11)
is the joint quadratic variation between and X. Set
T
(i)#(T&i),
X
T
(i)#X(T&i)&X(T) and for each partition of [0, T] as above let
T
denote the partition
T
=0=T&i
n
<T&i
n&1
< } } } <T&i
1
<T&i
0
=T].
Noting that |
T
| =|| and
X(T&(T&i) .(t+))&X(T&(T&i) .t)
=X(i (T&t+))&X(i (T&t)),
we have
|
T&i
0
T
dX
T
# lim
|| 0
_
t #
_
t #
(T&t)(X(T&(T&i) .(t+))
&X(T&(T&i) .t))
= lim
|| 0
_
t #
T
(t+)[X(i t)&X(i t+)]
=&
|
T
i

dX. (8.12)
We will now use this last relationship as a definition for our infinite
dimensional backwards Ito^ integrals. We now formulate the precise defini-
tion that is used in the body of this paper. As in the last subsection let (W,
F
i
]
i0
, F, F) be the filtered probability space and [(i)]
i0
be the
L(q)-valued Brownian motion as in the body of the text. Fix T>0 and
set [
T
(i)#[(T&i)&[(T) for i # [0, T], F
T
denote the completion of
the o-algebra generated by [
T
(i) : 0iT] and F
T
i
be the o-algebra
generated by [
T
(t) : 0ti]=[(t)&[(T) : T&itT] augmented
by the null sets of F
T
.
Definition 8.6. Suppose that H(i) is a continuous (for simplicity)
H
0
(q)-valued process such that H(i) is F
T
T&i
-measurable. (Notice that
F
T
T&i
is the o-algebra generated by [(t)&[(T) : itT] augmented by
the null sets of F
T
.) Then H
T
(i)#H(T&i) is a continuous process
adapted to the filtration F
T
i
]
0iT
and we define, for 0iT, the
backwards stochastic integral of H as
|
T
i
(H(t),

d[(t))#&
|
T&i
0
(H(T&t), d[
T
(t)). (8.13)
Fenark 8.7. Notice that the backward Ito^ -integral defined in Eq. (8.13)
inherits the basis the 1
2
-isometry property from the forward Ito^ integral,
namely
E
_
|
T
i
(H(i),

d[(i))
&
2
=E
|
T
i
&H(i)&
2
di (8.14)
provided the right side of Eq. (8.14) is finite.
ACKNOWLEDGMENTS
I thank Johan Van Biesen for illuminating discussions on topics related to this work and
Trevor Carson for his suggestions and corrections to the manuscript.
Noie added in proof. Theorem 1.5 and Theorem 1.4 have been extended to unbased loop
groups in Trevor Carson's 1997 University of California, San Diego, Ph.D. thesis.
REFERENCES
1. Denis Bell, A quasi-invariance theorem for measures on Banach spaces, Trans. ner.
Maih. Soc. 290, No. 2 (1985), 851855.
2. Ana-Bela Cruzeiro, E quations diffe rentielles ordinaires: Non explosion et mesures quasi-
invariantes, J. 1unciional nal. 54 (1983), 193205.
3. Ana-Bela Cruzeiro, E quations diffe rentielles sur l'espace de Wiener et formules de
Cameron-Martin non lineaires, J. 1unciional nal. 54 (1983), 206227.
546 BRUCE K. DRIVER
4. B. K. Driver, A Cameron-Martin type quasi-invariance theorem for Brownian motion on
a compact Riemannian manifold, J. 1unciional nal. 110 (1992), 272376.
5. B. K. Driver, Integration by parts for heat kernel measures revisted, UCSD preprint,
34 pp., December 1996 [To appear October 1997 in J. Maih. Fures ppl.]
6. B. K. Driver and T. Lohrenz, Logarithmic Sobolev inequalities for pinned loop groups,
J. 1unciional nal. 140 (1996), 381448.
7. B. K. Driver and L. Gross, Hilbert spaces of holomorphic functions on complex Lie
groups, in ``Proceedings of the 1994 Taniguchi Symposium: New Trends in Stochastic
Analysis'' (K. D. Elworthy, S. Kusuoka, and I. Shigekawa, Eds.), pp. 76106, World
Scientific, New Jersey, 1997.
8. M. Emery, ``Stochastic Calculus in Manifolds,'' Springer-Verlag, Berlin}Heidelberg}
New York, 1989.
9. Gerald B. Folland, ``Real Analysis, Modern Techniques and their Applications,'' Wiley,
New York, 1984.
10. D. S. Freed, The geometry of loop groups, J. Diff. Geon 28 (1988), 223276.
11. L. Gross, Abstract Wiener Spaces, Proc. 5th Berkeley Symposium Math., Siai. Frob. 2
(1965), 3142.
12. Elton P. Hsu, Lectures given at the IAS}Park City Mathematics Institute, Summmer
Session, held June 23July 13, 1996 at the Institute for Advanced Study.
13. N. Ikeda and S. Watanabe, ``Stochastic Differential Equations and Diffusion Processes,''
2nd ed., North-Holland, Amsterdam}Oxford}New York, 1989.
14. H. Kunita, ``Stochastic Flows and Stochastic Differential Equations,'' Cambridge Univ.
Press, Cambridge, 1990.
15. H.-H. Kuo, ``Gaussian Maesures in Banach Spaces,'' Lecture Notes in Mathematics, 463,
Springer-Verlag, Berlin, 1975.
16. S. Kusuoka and D. Stroock, Precise asymptotics of certain Wiener functionals, J. 1unc-
iional nal 99, No. 1 (1991), 174.
17. P. Malliavin, Hypoellipticity in infinite dimension, in ``Diffusion Processes and Related
Problems in Analysis. Vol. I. Diffusions in Analysis and Geometry. Papers from the Inter-
national Conference Held at Northwestern University, Evanston, Illinois, October 2327,
1989,'' (Mark A. Pinsky, Ed.), Progress in Probability, Vol. 22, Birkhau ser, Boston, 1990.
18. Michel Me tivier, ``Semimartingales, a Course on Stochastic Processes,'' Gruyter, Berlin}
New York, 1982.
19. M. Me tivier and J. Pellaumail, ``Stochastic Inte gration,'' Academic Press, New York,
1980.
20. Gunnar Peters, Anticipating flows on the Wiener space generated by vector fields of low
regularity, J. 1unci. nal. 142 (1996), 129192.
21. Gunnar Peters, Flows on the Wiener space generated by vector fields with low regularity,
C. F. cad. Sci. Faris Serie I 320 (1995), 10031008.
22. P. Protter, ``Stochastic Integration and Differential Equations; A New Approach,''
Springer, Berlin}Heidelberg}New York, 1990.
23. Michael Reed and Barry Simon, ``Methods of Modern Mathematical Physics: I Func-
tional Analysis,'' Academic Press, New York, 1980.
24. Revuz and Yor, ``Continuous Martingales and Brownian Motion,'' Springer-Verlag,
Berlin}Heidelberg}New York, 1991.

Integration by Parts and Quasi-Invariance For Heat Kernel Measures On Loop Groups

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Integration by Parts and Quasi-Invariance For Heat Kernel Measures On Loop Groups

Uploaded by

Copyright:

Available Formats

File: 580J 310301 . By:DS . Date:10:09:97 . Time:10:31 LOP8M. V8.0.

(L). A function f : (L) R is said to be a snooih

(L) and for f # D(E

(L, T). The left invariant vector

(L, T), namely

(L), be giten as in Eq. (1.7), ihen

(L) are measurable. For the sequel, fix a filtered

(oi, s) wiih 2(0, s)=g

(i, s)#, [(i, s),. Here [

(i, s). In ihe sequel, we will

(M_N). So by definition of : solving Eq. (3.7), we have

(G, q*) be defined by

(G) be a suitable collection of functions such that

(G) and 1( g, k)#f ( gk

[(dt, s)&[(dt, o)] [

[(dt, s)&[(dt, o)] [

(_) h (_) d_,

(_) h (_) d_.

(_)) h (_)& d_.

as || 0 a.s. Finally the estimate

(_) h (_)& d_,

=0. This proves Eq. (4.9) in the case that h is

(L) of the form f ( g)=1( g

(L), and T>0. Sei 1

((L)), and notice that

(L) and for u, t # FC

(L) such that lim

((L)), and set k(i)=(i}T) h. Then

(L) has a densely defined 1

(L) is of ihe forn f ( g)=1( g

(L). Choose a partition

(L(G)) is a densely defined closable operaior on 1

(L) and l(i)=i}T. Then apply Theorem 6.2 with

(L) and t=1 in Corollary 6.6 we find

(L) which are based on P

f d for all # G, ihen T

You might also like