This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
0
D,,,(E)
and
(8.36)
D__(E) =
U
1
0
D,._t(E).
Then D (E) is a complete countably normed space (Frichet space) and is its dual. The spaces DD,t(E), D (E), are denoted
by DD,t, D_, D_..... if E = R. Since D,, 3(E) c D,, 0(E) = L,(E) elements in
D,,t(E) U 1
o ,./2t)
ifs > 0,
are Wiener functionals in the usual sense,
but elements in D,,t(E) for s < 0 are usually not so. We have a clear analogy with the Schwartz distribution theory and it may be natural to call such elements as generalized Wiener functionals. The coupling D-oa(E)(0, F)D.(E),
45 E D__(E), F E D.(E).
MALLIAVIN'S STOCHASTIC CALCULUS
365
is also denoted by E(<'1, F>E) and it is called a generalized expectation.
In particular, 1 (= the functional identically equal to 1) E D. and, for 0 a D_,,, the coupling D__(cI, 1)D is called the generalized expectation of 0 and is denoted by the usual notation E(,P) because it is compatible when
45 ei
Ornstein-Uhlenbeck operator L: P(E) - P(E) can be extended uniquely to L: D_..(E) and it is continuous as a operator D,,,+2 (E) -- D,, (E) for every p e (1, oo) and s E R. The following important result is due to Meyer [218]. Unfortunatly, we can not have enough space here to give its complete proof. For the proof, see [225] and [232]. Theorem 8.4. (Meyer [218]). For p c- (1, oo) and k E Z+, there exist positive constants c,,k and C,,k such that k
(8.37)
c,,klIDkFlln c IIFIID.k <_ C,.kEIID'Fllp r-o
for every F E P(E). By using this we obtain the following theorem. Theorem 8.5. (Meyer [218], Sugita [225], P. Kree-M. Kree [208]). The operator D : P(E) -- P(H (D E) (or S(E) -- S(H (& E)) can be D_m(H (& E) so that its extended uniquely to an operator D: restriction D: D,,,+,(E) -- D,,,(H(D E) is continuous for every p e
(1,00) and sER. By the duality, we immediately have the following
Corollary. The operator D*: P(H Qx E) --- P(E) (or S(H (D E) -S(E)) can be extended uniquely to an operator D*: D_ ..(H (DE) D_.(E) so that its restriction D*: D,.,+,(H (& E) -- D,,,(E) is continuous for every p e (1, co) and s E R.
It is now clear that L = - D*D in this extended sense. These results imply, in particular, that operators D, D* and L, first defined on the space of polynomial functionals or smooth functionals, are closable and therefore can be extended to Sobolev space D,,3(E). It may be in-
DIFFUSION PROCESSES ON MANIFOLDS
366
teresting to note that the following operator on P; P 2) F- trace D2F E=- P
is not closable. Indeed, let r = 1 and Fk E P, k = 1,
be defined
by 2k
Fk(w) _
It is easy to show that Fk -- 0 in L, for every 1 < p < oo but trace DZFk = 2 for every k = 1, To prove Theorem 8.5, we prepare the following lemmas and propositions.
Lemma 8.1. (Commutation relations involving D). For a real sequence 0 = (0.,).'=O,
DTI = TT+D
P(E)
on
where
0+(n) = ¢(n + 1),
n = 0, 1,...,
Proof. For simplicity, we only show the lemma in By Proposition 8.1, it suffices to prove DTOH,(w) = TT+DHQ(w)
for every
case of E = R.
a E A,
where Ha(w) = jfl HH,([h,](w))
and [h,1,'!, is an ONB of H. Since TTH,(w) = 0(I a I)H4(w)
aEA,
we have (8.38)
DTTHQ(w) = cS(I a I)DHH(w).
Furthermore, noting that H;(x) = H,-,(x), we easily DH,(w) = aj>0 Z Haj-j([hi](w)) IIi*jH.,([hj](w))hi
see
MALLIAVIN'S STOCHASTIC CALCULUS
367
and hence
DHR(W) E C. This implies TO+DH,(w) _
n-1
O(n +
o(I a I)DH,(w).
Combining this with (8.38) yields the conclusion.
Proposition 8.6. (Hypercontractivity of the Ornstein-Uhlenbeck semigroup, Nelson [220]). Letp E (1, oo), t > 0 and q(t) = (p - 1)e2` + 1. Then for every G E
IIT,Gllgco c IIG1Ip
L.
Proof. The following proof is due to Neveu [221]. Let { B$'>, t > 0 }, i = 1, 2, be two n-dimensional Brownian motions with B0('> = 0 defined on a probability space {92, . ; Q }. We assume that { B;'', t > 0), i = 1,2 are independent. For given 2 E (0, 1), we set
q=q(1)=(p-l)1-2+1 and define q' = q'(2,) by
1 f-q,=1. It is easy to see that { (p - 1) (q' - 1) }1J2 = A. We define ]3> = Br(')
and
i ;2) = ABP) +
12B,(2) ,
and ._f"('), i = 1, 2, the proper reference families of o.[tt1>, 0 <
0
i.e..i7 _
s < t], i = 1, 2. Let f and g be C`-functions defined on R"
such that
a
forsome0
Define martingale { M,, 0 S t S 11 and { N1, 0 < t < 1 } by M, = E[f(Jiu)a' 1.
-t11)], N, = E[g(4112))D 1 _11',(2)],
0 < 1 < 1.
Then these can be written in the following forms (Theorem 11-6.6):
DIFFUSION PROCESSES ON MANIFOLDS
368
N, = No+fov/,dks2>.
M,=Mo+foO3d4,111,
By Ito's formula, we have EQ[M, /q' Nl.1/P] - EQ[Mo /ra'N (N/P]
2
EQ(J
IM -2 NP -2 L P
P) M,' V,1 - 2 pq, M`N`O,V, 11
1
Q
2
E
1_
l2
1
l
2 N`
LJ
-
1_2 (-, P1 +-L (1
1 N,0`/ Idt]
( by -,1(p
-1)(q'-1)=2 )
where EQ denotes the expectation with respect to Q. This implies that EQ[f( i1))g( i2))] = EQ[M'I N1i I lP
_
< E12[M101q, N'01']
Hence we obtain r
e
1
2
d1/q' {
J
Rnf(4)q,(N/1
27l
1
K
)"e-
Rng(S)P('V
2
n_ 1812
2e
_ If12
1
g(24 + /1 J Rn J Rn r
2 d dtt
n2 IfI2
27t
X2
Taking 2 = e ', t > 0, we have
{ f Rng(2 C
If
i - eryl)( /n_ "2
K
Rn g(S)P('V
27G
e
)ne_
jj fr 1/P
d4 2
1
812drl
z
IFI2
1
/1
d1/P
n_ I
2e
IZ
2d
j""
where 4(t)' = q(e °)'. By combining this with Remark 8.2 and Proposition 8.2, we easily obtain
f TG(w)F(w)P(dw)
s IIGIIPIIFIIQ(=),
for every G E LP and F E L4(,)I, which completes the proof.
MALLIAVIN'S STOCHASTIC CALCULUS
Proposition 8.7. J,,: L2 - C. is a bounded operator on L,,
369
PE
(1, cc).
Proof. Let p > 2 and t be the positive constant such that
e2.=p-1. Then by Proposition 8.6 we have
F E L, .
IITTFIIp <- IIFII2
In particular IIJVF112 s IIFII2 < IIFIlp
Since e-11114FII,,
this implies that (8.39)
for F E L,
c e''lIF11
For 1 < p < 2, consider the adjoint J,* of J. Then (8.39), applied to p'
such that 1/p + l/p' = 1, yields that IIJ.*FI1, < e-"'IIF11,
for F E P
But, for F E P, J,*F = 4F. By the denseness of P, the result follows. Lemma 8.2. (L, -multiplier theorem).* Let 0 positive sequence such that
0(n) = E a, (n--L ) k, for some no E N,
EIaxI(
for
0(n));.,, be a real
n > no
1 > a > 0 and o)k
Then for every I < p < oo there exists a positive constant C, such that First obtained by Meyer [218]. Proof given here is due to Shigekawa (Private communication).
DIFFUSION PROCESSES ON MANIFOLDS
370
F E P.
for every
IIT#FII, S C,IIFII,
Proof. First we consider the case a = 1. We set ng --1
To = n-0 L c(n)J + E 0(n)Jn
: = TO(1I + TO2)
X no
By Proposition 8.7, T0(1) is L; bounded, i.e., there exists a positive constant CD, (8.40)
IITT"FII, S C,IIFII,
for
F E L,.
We now show that there exists C > 0 such that (8.41)
IIT0(I - Jo - J, -...- Jn°-1)FII, < Ce-n°`IIFII for all t > 0
and
F E L,.
First we consider the case p = 2. For F (=- L2, we have
II Tt(I - J° - J, -- Jn°-1)FII22 (8.42)
= IIE ek,JkFl12 = Ek°n0 IIe kSJkFII2 k®n0 e20° k-n° E IIJkFII22 S e "ofIIFIIi
Next we consider the case p > 2. Take to such that p = e20 + 1. By Proposition 8.6, we obtain
II T°+# - J° - J, -...- Jn0-1)FII2 Jno-1)FII2 < II Tt(I - Jo - J,
< IIE e (8.43)
n-n0
e2n°n_'0 E IIJJFIIi e-2n0`IlFl12
< e2n°tIIFII2
=
e2n°t°e2n°a+t°,IIFII2
Furthermore for t :!!g to we obtain, by (8.13), 11 T, (I - J° -- J1 - ... - Jn°-1)FIIP
MALLIAVIN'S STOCHASTIC CALCULUS (8.44)
< 11(I - Jo - J,
< III - Jo - J, where III -
Jo - J, -
371
- J.0-1)FII, Jn0-111DIIFII,
- J.0_111, is the operator norm of (I - Jo
- J, - ... - JJ0_1) in L.
Combining (8.42), (8.43) and (8.44) we obtain (8.41) with p > 2. For 1 < p < 2, the (8.41) follows by duality. We now set
R"0 = fo
T,(I - J. - J, - ... -
Then, from (8.41) we obtain IIRnoFII, < Cno IIFII,.
Since
R"0F =
J of 0
T:+S(I - Jo - J, -
-
dtds
we obtain IIRn20FIIo < Cno IIFIID
and repeating this we have (8.45)
IIRn0FII, s C LIIFIIP,
k = 1, 2, ... .
Note that RnoF = nkJnF,
k = 1,
Hence we obtain
T1Z)F =n-n0k-0 E E akRnoJTF =k-0 E By (8.45),
if F E Cn, n
no.
DIFFUSION PROCESSES ON MANIFOLDS
372
IITT2'FIID < C(EIakl o)k)IIFIID
Combining this with (8.40), we can conclude that there exists C, > 0, such that
F E P.
IIT#FII < CPIIFIIP
For the general case i.e., 0 < a < 1, define Qia' = E e natJJF =
Ja
(T:F) a)(ds)
where p;°) is the probability measure defined by
f
asp160(ds) = 0
e
exp (-
A61t)
for every A > 0.
As in the case a = 1, write To = TOW + TO(2)
By the same reason, TT1' is L, bounded. From (8.41), we have II
Qia)(I -
< c$
J0 - J1
Jn0-1)FIID
IIFIIe-1'0°p,('1'(ds) = C exp (- not)IIFII,. o
Define R.0
f0
Qt-)(I
-J0-J1
and proceed as in the case a = 1. Then we obtain that TT2) is LD bounded. This concludes the proof.
Proof of Theorem 8.5. In case of E = R, we show the theorem. By Lemma 8.1, we obtain
I (I -
L)112DF
I H = I DR(I -- L)sl2F I H
where R is the operator given by n R=E(n+11
J,,.
FEP
MALLIAVIN'S STOCHASTIC CALCULUS
373
Note that R = To with 10,
1(n)
s/2
s/2 n (n+1) -(1+ n) ' 1
n=0 n> 1
and
h(x)=(1+x 1
is analytic near x = 0. By Theorem 8.4 and Lemma 8.2, there exist posi-
tive constants C, and C; such that II(I - L)1'2DFIIP = II DR(I - L)"ZFll, s
,
< C,II(I - L)TR(I - L)'TFII, = C,II R(I -
CII(I -
L)<s+ll'2Flip
Therefore IIDFII,,s < C;IIFII,,s+,
FEP
from which the result follows by a limiting argument.
Proposition 8.8. Let E, and E, be real separable Hilbert spaces. For any p, q c= (1, oo) and k = 0, 1, 2, such that
there exists a constant C,, q, k> 0 such that (8.46)
IIF0 GIIr,k C Cp.q.klIFIIp.kIIGIIq,k
for every F E P(E) and G E P(EZ). Proof. Indeed,
DIFFUSION PROCESSES ON MANIFOLDS
374
D(FQG)=DF®G+FQDG and hence ID(F(& G)IHeEIeEZ < IDFIS®E,IGIE,+ IFIE,IDGIR®Ey
from which (8.46) in the case of k = 1 is easily obtained by noting the equivalence of norms IIFII, + IIDFII, and IIFII,,1 (which is a consequence of Meyer's theorem). We can obtain (8.46) by the same argument successively.
From (8.46) and the duality (Proposition 8.3, (iii)), it is easy to obtain the following.
Corollary. For every p,q e (1, oo) such that 1
1 + q =
r
< 1
and k = 0, 1, , there exists a constant C,q,k > 0 such that (8.47)
IIF® GIIr.-k < CD q,kjIFII,,kIIGIiq.-k
for every F E P(E,) and G E P(E2). By (8.46) and (8.47), we see that F Q G E D,, k(E, (g EZ) is welldefined for every F E DD, k(E,) and G E Dq, k(EZ) and, F Q G E D,, _k(E,
(D E2) is well-defined for every F E D,, k(EI) and G E Dq, -k(EZ) provided
+ q = r < 1. In particular, D. is an algebra, i.e., if F,G
D_ then FG E D.. Furthermore, D (E) and D__(E) are D,.-modules, i.e.,
if F E D_ and G E D_(E) (D_..(E)), then FG E D_(E) (resp.
D--(E)). Note that if F E Dm and G E D__, the generalized expectation F>D,,. E(FG) of FG E D__ coincides with We note that chain rules (8.24)-(8.30) can be easily extended to Wiener functionals belonging to Sobolev spaces in an obvious way: In particular these formulas hold if P and P(E) are replaced by D.. and D_(E) respectively. Also the integration by parts formula (8.23) can be extended to the case F E=- D,,, and G E Dq,,(H) such that
I + q =1.
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
375
Of course, (8.23) holds in the case FED,,,+, and G E DQ, _,(H),
P + 9 =1,
sER,
if the both sides of (8.23) are understood as the natural coupling between D,, 3(H) and DQ, _,(H) (= (D,,,(H))') and D,, and Dq respectively.
9. Pull-back of Schwartz distributions under Wiener mappings and the regularity of induced measures (probability laws) We now apply the results of the previous section to the study of re-
gularity and asymptotics of probability density of finite dimensional Wiener functionals. For this, we regard a d-dimensional Wiener functional as a d-dimensional Wiener mapping F: W- Rd and discuss the pull-back of Schwartz tempered distributions on Rd under this mapping. This pull-back can be defined as an element in D_-, i.e., as a generalized Wiener functionals, if the Wiener mapping satisfies certain conditions on regularity and non-degeneracy usually referred to as Malliavin's conditions. If the Dirac delta function 8X at x E Rd is pulled back by a Wiener
mapping F: W- Rd, then the generalized expectation E[Sx(F)] of this pull-back 8z(F) coincides with the density p,(x) of the probability law of functional F with respect to the Lebesgue measure on Rd. We can discuss the continuous or differentiable dependence on parameters of the pull-back thereby deduce the regularity of the probability density. So let (W, P) be the r-dimensional Wiener space and let F(w) _ (F'(w), FZ(w), , Fd(w)) be a mapping W- Rd. We assume that it is smooth in the following sense of Malliavin: (A.1)
FED_(Rd),i.e.F'ED.,
i=
Then we can define (9.1)
or'i(w) =
i,.l = 1, 2, ... , d
and we know that a'J E D-. The matrix a(w) = (o'"(w)) is non-negative definite and hence det v(w) > 0 for almost all w(P). 6(w) is called the Malliavin convariance of F. Definition 9.1. We say that F is non-degenerate in the sense of Malliavin if it satisfies
DIFFUSION PROCESSES ON MANIFOLDS
376
(det v(w))-1 E f1 L,.
(A.2)
1
Here it is understood that 1/0 = oo. Assume that F: W-- Rd satisfies Malliavin's conditions (A.1) and (A.2). For every e > 0, define v8 = (oa) by aa1=au+6(/11
we have where t5'1 is Kronecker's delta. Then, setting (yt1) = y,8, E D_, i, j = 1, 2, ..., d because there exist tempered C"-functions f1(x), i, j = 1, 2, , d on R1 such that i, j = 1, 2, ... , d.
Y91= jt,(a8),
It is easy to prove by the chain rule that d
Dy,, = --
(9.2)
k,1-1
Y'rkYi1Do
Furthermore it is easy to deduce from (A.2) that in L,(H®k)
DkYi1- Dky,1
for every k=0,1,-.., 1
H®k = H Qx H Q k
i, j = 1, 2,
,
d
where
(Q`7)-1 This shows that
i,j= 1,2,...,d.
Yr1 e D_,
Also by (9.2), we have d
Dy1
(9.3)
= - k,E1-1
YrkYllDdkl
Let 0: Rd - R be a tempered C --function. Then, by (8.24), d
D(¢ -F)
(a,0 o F)DFr. t=1
Hence d
PULL-BACK OF SCHWARTZ DISTRIBUTIONS d
=
a,0°F. Qu [n1
and therefore, d
(9.4)
X. y,J
3r 5 °F
Jm1
Let g E D... Then, we obtain by (8.23) that fw ai¢ ° F(w) g(w) P(dw) d
_ 57, f
J-1
w
_ J-1 E
w
0° F(w) D*(yjJ(w)g(w)DFJ(w)]P(dw)
= f w 0 ° F(w)
01(w; g)P(dw)
where d
01(w; g) _ EJ-1D*[yu(w)g(w)DFJ(w)l d
Z (g(w)
(9.5)
J-1
+ 7,j(w)
by (8.26) and (8.30). Hence
f
w
(atlat2
... ask 0) °F(w) . g(w)P(dw)
= $o F(w) where 0'3, ,2, Oft, 12,
.
0t1, t2, ... r Ik (w; g)P(dw)
,k(w; g) E D. is determined successively by iv(w; g) = O1(w; 'Pt1r 12' ... '
-
; g))
We can conclude from (9.5) that 0,1, 1 .....:k(w; g) has a form
377
DIFFUSION PROCESSES ON MANIFOLDS
378
011. 12..... 'k(w; g) (9.7)
_ W0(w)g(w) + < .'1(w), Dg(w)>H + .. . + < Ylk(w), Dkg(w)> Hek
where
v=0,1,2,.
Tp EDW(H9%
which are obtained as polynomials in y,j, F= and their derivatives. a Similarly we have for d = E (a/ax`)2, i-3
J w111 + 1x12 - 2) ko)o F(w) . g(w)P(dw)
= f 0 - F(w)
?lzk(w; g)P(dw)
where 772k(w; g) E D. has a similar expression as (9.7). Indeed, 772k(w; g)
is obtained in the same way as 0,i, =z. . ,k (w; g) with some more polynomials of F multiplied in each step of the integration by parts. In particular, for every q > 1, (9.9)
sup{ II 2k(
; g)IIj; g E= Dq.2k, IlgIIq.2k s 1 } < 00
II, denotes the L,-norm. We introduce a system of Sobolev spaces on Rd as follows: Let .So(Rd) be the real Schwartz space of rapidly decreasing C"-functions on Rd. For 0 E .S'(Rd) and k = 0, + 1, + 2, , introduce the norm where II
(9.10)
110112k = II(1 + 1x12 -A/2)'011.
where II
II- is the sup-norm on Rd. Let S°2k, k = 0, ±1,±2,
be the completion of 9o(R2d) by the norm II
II2k. Then we have
c .9'2 c Y,, = : { f(x); continuous on Rd and lim I f(x) I 1X1--
= 0}cs°_2c n _9"2k = .7(Rd)
k>0
and
U -9"-2k = Y'(Rd).
k>O
Theorem 9.1. Let F: W--- Rd be a d-dimensional Wiener mapping satisfying Malliavin's conditions of (A.1) and (A.2). Then, for every p > 1 and k = 0, 1, 2, , there exists a positive constant C =
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
379
C(p, k; F) such that the following estimate holds: For every 0 E Y(Rd), (9.11)
IIsoFIIp.-2k -<- C11011-2k-
(Note that ¢ o F E D_).
Proof. Define q by 1/p + 1/q = 1. Then by the duality I I0e FI Ip, -zk = sup{
f w O o F(w) IIgllq.2k C
g(w)P(dw); g E Dq, zk, 1}
and, by (9.8) and (9.9),
i f 0 o F(w)
=I
f
g(w)P(dw) I
((I + 1x12 - d/2)k(l + I x 12 - d/2)-ko } o F(w)
x g(w)P(dw)
=I
f
{ (1 + 1x12 - A/2)-'o } o F(w)
< 11(1 + I x 12 - e/2)-k¢II W IIi72k(
<
172k(w; g)P(dw) I
, 8)111
II0II-2kll 2k( ; g)111-
Now (9.11) is established by setting C = C(p, k; F) to be the supremum in (9.9)
Corollary. The mapping Y(Rd) E3 - 0 o F E D. can be extended uniquely to a continuous linear mapping 5'-2k E) T- ToF E Dp, _2k
for every I < p < oo and k = 0, 1, 2, - - . In particular, for every T E 5"(R,), ToF E_ D__ is well-defined and -
(9.12)
ToFEU
(1
k-1 1
Dp, -2k-
This ToF, denoted also by T(F), is called the pull-back of Schwartz distribution T E S'(Rd) on Rd by the Wiener mapping F: W_ Rd, or the composite of distribution T on Rd and d-dimensional Wiener functional F.
DIFFUSION PROCESSES ON MANIFOLDS
380
The fact (9.12) is important because ToF can be coupled with test functionals in a class much larger than D_: By introducing the notations (9.13)
D_ = n
(9.14)
D__ = U
D,,
U
k-I I <,
fl
k-1 1<,
and
DD. -k,
if G E D_, then G. ToF E D_" is well-defined and hence the generalized expectation
D_
p_a, <1, is well-defined.
Example 9.1. Let r = I and so W = W,'. For w E W, set t
G(w) = exp{ a f w(s)2ds }, 0
a E R.
Then G(w) E D. if and only if a < rc2/8 and if f(x) is a C"-function on R with compact support, G(w)f(w(l)) E D" if a < n2/2. The proof of these facts can be easily provided by the formulas (6.9) and (6.10) of
Section VI-6. Note that G(w) E D_ if and only if a S 0. The above result can be applied to show the existence of smooth density of the probability law F,k(P) of F, i.e. the image measure of P under F. For this we first note the following fact.
Lemma 9.1. Let Sy E .9'(Rd) be the Dirac S-function at y e Rd and m be a positive integer. Then
Sy E Y-2. and D,Jy E -9"-'.-2k where a = (a,,, a2i -
+
,
if m> d and I a I S 2k,
ad), a, E Z+, is a multi-index, I a I = a1 + a2
+ ad and D. is the differential operator: Da =
as)a, 1
ax2
1a2
...
axd )ad
PULL-SACK OF SCHWARTZ DISTRIBUTIONS
381
Furthermore if m > d/2, then the mapping: Rd E) y -- Sy E .- 2m-2k
is C2k, i.e. 2k-times continuously differentiable.
Proof. First we note that the operator A = : (1 + I x 12 - d/2)-' is defined by a kernel A(x, y) :
Af(x) =
J Rd
A(x, Y)f(Y)dY
and A(x, y) is given by
A(x, y) = f e 'p,(t, x, y)dt where p1(t, x, y) is defined in (6.12) of Section VI-6 (in the case d=2, but similarly in general). From this, we see that A(x, y) > 0, A(x, y) is C°° in (x, y) E (Rd x Rd)\{ (x, y); x = Y1, and
const. x I x- y I- cd-2)
A(x, y)
as Ix - y j
d 0.
Also it is immediately seen that x, y e Rd
A(x, y) < A(x, y),
where A(x, y) is the kernel of the operator A = :(1 - A/2)-'. Hence ((1 + I x l 2- d/2) mo)(x)
- fRd
Rd
f...
fRd A(x,
f f ... f Rd
Rd
Rd
Yz) ... A(Ym-a, Y)dyidy2 ... dym-,
A(x) Y1)A(Y1, Y2) ... "(Ym-t, Y)dY1dY2 ... dym-I
I )d
2: f Rd 1+ I
12 m 2
)
and the right-hand side is in Soo as a function of x if m > d . From this
we can easily deduce that
DIFFUSION PROCESSES ON MANIFOLDS
382
(1 + I x z- d/2)-mb,, E .9%
for each y
and the mapping Rd EE) y --- (1 + I x 12 - d/2)-mby c- .9'0 is continuous
provided m > d/2. This proves the first assertion and the second assertion can be proved similarly. The last assertion is obvious if we notice that Day = (_ 1)"'(Dy)aby
where (Dy)a is the differentiation with respect to the parameter y.
Theorem 9.2. Let F: W- Rd satisfy the conditions (A.1) and (A.2) and let Fk(P) be the probability law of F. Then F*(P) has the smooth density pp(y) with respect to the Lebesgue measure dy on Rd and pp(y) can be given by (9.15)
PF(Y) = E[by(F)]
as generalized expectation of by(F). More generally, for every multiindex a, (9.16)
(Dapp)(Y) = E[((Dy)aby)(F)] = E[(-1)1a'(Daby)(F)].
Proof. Let k be a non-negative integer. By the lemma above and the continuity of the mapping: S 2m T -- T-F E Dp, _2m, we can deduce at once that the mapping: Rd E) y --- by(F) E D, -2m0-2k
is C2k where mo = [d/2] + 1 and hence, for every g E Dq,2mo+zk, I/P +
1/q = 1, the mapping: Rd ii) y -- E[gby(F)] is C21. In particular, the y - E[gby(F)] E R is C- if g E D,,. Thus the mapping: Rd m y -- E[by(F)] E R is C-. (9.15) can be verified by noting that for mapping: Rd
every c E So(Rd)
fRd
c(Y)E [by(F)]dy = E [J
xd
O(Y)by(F)dy]
= E[(f ad 0(Y)by( )dy) - F] = E[0 - F]. Now (9.16) is obvious. In the above proof, it was shown that if g E U Dp, 2m0+2k where mo p>1
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
383
[d/2] + 1 and k is a non-negative integer, then the mapping: Rd D y- E[g6y(F)] is C2k. It is easy to see that (9.17)
E[gby(F)] = pF(y)E[g I F = y]
for almost all y such that pF(y) > 0. Thus we have obtained a regularity result for conditional expectation:
Corollary. If g E U DD.2m0+Zk where mo = [d/2] + 1 and k is a y>1
non-negative integer, the conditional expectation E[g I F = y] has a C21 -version on the set {y; pF(y) > 0 ). In particular, it has a C°°-version
on the same set provided g E E. Example 9.2. Let r = d and F: W(= Wod) - Rd be defined, for fixed t > 0 and x E Rd, by F(w) = x + w(t). Then F E D_(Rd) (actually E P(Rd)) and Qtj(w) = out, i, j = 1, 2, , d. Hence yj = Stj/t where 81j is Kronecker's delta and hence (A. 1) and (A.2) are clearly satisfied. Let k E C2-(Rd) which is tempered in the sense D`xk, I a I :!S; 2m, are all polynomial growth order and
HEk(x)/Ix12=a
If a is sufficiently small (actually it is sufficient to assume a < 1/2t), we
can easily see that g(w) = exp{ f u k(x + w(s))ds) E pU1 D,.,,.. Hence
p(t, x, y): = pF(y) = E[(exp{ f' k(x + w(s))ds })O7(x + w(t))]
is C2k in y if k = m - mo > 0. By the Feynman-Kac formula (Theorem 3.2), it is easy to identify p(t, x, y) with the fundamental solution of heat equation au
I
at = 2 4u + ku.
The generalized Wiener functional O,(x + w(t)) is known as Donsker's 0-function (cf. [210]).
DIFFUSION PROCESSES ON MANIFOLDS
384
Now we consider a family { Fa(w) }ael of d-dimensional Wiener
functionals depending on a parameter a, the parameter set being assumed to be an m-dimensional interval I e Rm, for simplicity. We assume that Fa E for all a E I and also that, denoting by o(a) = (&'(a)), oif(a) =
for 1 < p < oo.
Then it is clear that the constant Ca = C(p, k; Fa) in Theorem 9.1 can be chosen such that sup C. < CO. aal
Hence, for T E 2k, if we choose 0. E .99(R4) such that 0. --- Tin S°_2k as n -- oo, then for every I < p < oo, 0.- F,, - T°Fu in D,, _2k as n -- oo uniformly in a E I. From this we can easily conclude the following: If the mapping: I a -- Fa E D. is continuous (2m-times differentiable) then, for T E S? 2k, the mapping: I Z) a - T°Fa E D,, -2k is continuous for every 1 < p < oo (resp. the mapping: I E) a --T°Fa (z D, _2k-2m is 2m-times differentiable for every 1 < p < cc). We can deduce from these facts the continuity or differentiability in a of the generalized expectation if the mapping: I iD a-. ga E DQ,2k is continuous or differentiable where 1lp + 1/q = 1. Note that, in the differentiable case, we have a
d aFQ 8T
as T ° Fa) _ E aa (3x"
°
Fa,
F. = (F1, F«, ... , Fad)
which is easily verified by a limiting argument as 0. - T, On e .9'(Rd). Finally, as an example of dependence on parameters, we discuss on
the asymptotic expansion of Wiener functionals. Consider a family { F(e, w) } of E-valued Wiener functional depending on a parameter e E
(0, 1] where E is a separable Hilbert space. We assume
F(e, w) E D (E)
for every e.
We define F(e, w) = O(em) in D.(E) as e 10 if F(e, w) = O(sm) in D,, k(E) for' every 1 < p < oo and k E N as a 0, i.e. lim e-mI!F(e, w)ll,.k < oo. #to
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
E
Let fo, f,, (9.17)
385
We define
fo + ef, + s2f2 + ..
ass J, 0 in if, for every n = 0, 1, 2, - - - ,
F(s, w)
F(e, w) - (1 0 + ef, + ... + e"fn) = O(e^+1) in D_(E) as a J, 0. Also, if { 0(e, w) }
is a family of elements in D_ depending one
(0, 1] and 1Vo, yr1,
-
- e D_..(E), we define in D__(E)
(9.18)
if, for every n = 0, 1, 2, that ds(c, w), s e (0, 1]
- -
, we can find
as s 10
kENand 1
and yro, yrli - - - , V/. are all in D,,-,(E)
and
0(e, w) - (+ilo + e'1 + ... + snpn) = O(en+1) in D, -k(E) as e 10. We define D_(E) and b_(E) for general E in the same way as (9.13) and (9.14), i.e.
D.(E) = k-1 fl 1
D,. k(E)
and
b_(E) =u fl k-1 1
Dp.-k(E)
If { G(e, w)) is a family of elements in D&(E) depending on e E (0.1] and go, g1, - - - E D_(E), we define (9.19)
G(e, w) --- go + eg, + e2g2 + -
if, for every n = 0, 1, that
in b_(E) as s y 0
and k e N, we can find 1 < p < oo such
G(e, w), e E (0, 1] and go, g1, -
-
, gn are all in D, k(E)
and
G(e, w) - (go + sg1 + ... + e"gn) = 0(s- 1) in Dp. k(E) ass 10.
DIFFUSION PROCESSES ON MANIFOLDS
386
If ( !P(e, w)) is a family of elements in D_..(E) depending on e E (0,1] and if yro, yrli E D_..(E), we define (9.20)
9'(e, w) -- yro + eyrl +
as e.L 0
, we can find k c- N such that
if, for every n = 0, 1, 2, Vl(e, w), a E (0,1]
in D_m(E)
and
, V. are all in
yro, yr1,
(1 D,,_k(E)
1
and, for every 1 < p < oo, TI(s, w) - (yro + eyrl + - - - + e"+v) = O(e°+1)
in D,, _k(E) as a J, 0.
Noting the continuity of multiplications in Sobolev spaces as given
by inequalities (8.46) and (8.47), it is easy to prove the following:
Proposition 9.3. (i) If F(e, w) E D_(E) and G(e, w) E D. (D_), e e (0,1] such that (9.21)
F(e, w) -fo + Ef1 +
in D_(E) as e J 0
with f, E D_(E), i = 0, 1, - - and (9.22)
G(e, w) - go + eg1 +
with g, E D. (resp.
i = 0, 1,
in D (resp. ,
as e.0
then H(e, w) = G(e, w)F(e, w)
satisfies (9.23)
H(e, w) - ho + eh1 + -
and ho, h1, tiplication : (9.24)
in
(resp.
as a 10
E D.(E) (resp. D-(E)) are obtained by the formal mul-
ho = gofo, h1 = gof1 + g1fo, h2 = go.fz + g1fi + gzfo, .. .
(ii) If G(e, w) E
and O(e, w) E
e E (0, 1] such
that (9.25)
G(s, w) --- go + sg1 +
with g1 a D. (resp. D_), i = 0, 1,
in D. (resp. D_) as e.0 and
387
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
(9.26)
00 + eO, + -
fi(s, w)
with 0, E D__(E), i = 0, 1,
in D_..(E)
as e
0
, then W(e, w) = G(e, w) fi(e, w) satis-
-
fies
(9.27)
V`(s, w) - V'o + eVr, + ...
in D_..(E) (resp. D__(E))
as 810 and Vro, Vr,,
.
are obtained by the formal
- E D__(E) (resp.
multiplication : (9.28)
Wo = 9000, w, = g001 + g160, Wz = 9002 + 8101 + g7-00,
.
--
(iii) If G(s, w) E D_ and fi(e, w) E D_(E), c E (0, 1] such that (9.29)
with g1ED_, i=0, 1, (9.30)
in D. ass 0
G(e, w) - go + eg, + and
in D_ . (E) as s
fi(a, w) --- Oo + e6, +
0
with 0, E i = 0, 1, , then (9.27) holds in D_..(E) with given by (9.28). Wr E D_..(E), i = 0, 1, -
Theorem 9.4. Let { F(e, w), a E (0,1] ) be a family of elements in such that it has the asymptotic expansion: (9.31)
F(e,w)-..fo+ef,+ i
(9.32)
= 0, 1,
and satisfies
limll(det
oo
for all
1
00
810
where a(s) = (o-"(s)) is the Malliavin covariance of F(e, w): 6'J(e) =
45(e, w)
00 +eO,+
and 01 E D_,,, i = 0, 1,
- -
in D__
asa10
are determined by the formal Taylor
DIFFUSION PROCESSES ON MANIFOLDS
388 expansion:
P(E, W) = T(fo + [Efi + E2f2 + ... l)
= Ea t1(D.T) o fo[efi + -2f,. + ...
(9.34)
]
= o +Eq1 + where (i) the summation is taken over all multi-indices and (ii) for every , ad) E Rd, we multi-index a = (a,, a2, , ad) and a = (a,, a2, set as usual a! = a1!a2! ... ad! and a' = a, Ia22 ... aadd.
In particular, denoting at = a/axt, i= 1, 2,
, d,
d
Oo = T(fo),
01 = i-1 Ef(a1T)(fo)
02 = tEff(a,T)(fo) + 3, (9.35)
03 = tn()(f0) + 2! Eff3(ataaT)(fo)
+ 3!t,J,k-1 E fff'fk(ata akn(fo) and etc. Here
fk = (fk,fk,
fkd) (E D..(Rd)),
k = 0, 1, 2, -
-
Remark. fo E D (Rd) satisfies the condition (A.2) by virtue of (9.32).
Proof. We set F(0, w) = fo, ai'(0) =
YtJ(0) + eYj' + El(2) + .. .
k=1,2,
with ¢1,
First, we show that, for every k e N, we can find s E R and o, ,vk-1 E n D,,, such that 1
T o F(e, w) E
fl
1
for all c E [0, 1]
PULL-BACK OF SCHWARTZ DISTRIBUTIONS
389
and (9.36)
ToF(e, w) = 00 + go, + ... + ek-10k-1 + O(ek) in Do., as e,I,0
for all 1 < p < oo. To prove this, we take m E N so large that 0 _ (1 + Ix 12 - d/2)-m T is a bounded function on Rd, k-times continuously
differentiable with bounded derivatives up to k-th order. Then, for every J E D", we have, by the same integration by parts as (9.8), w). J] = E[oo F(e, w)-1,(J)]
E[T o F(e,
where 18(J) E D. has a similar expression as (9.7): 2m
1e(J) = E
where P1(e, w) E D" (H®r), i = 0, 1, , 2m, which are polynomials in F(e, w), its derivatives and y(e). By the assumption on 0, we have OoF(e' w)
IalE-l -a!(Da0) ° A [F(e, w) - fola + Vk(e, w) zgk
where for some constant M > 0, I Vk(e, w) I< M Ial E -kI [F(e, w) - fo]a 1
Hence, for every p' E (1, oo), we can find C, > 0 such that 11 Vk(e, W)11,, < C,ek
for all e e [0, 1].
Let q' E (1, oo) such that l/p' + l/q' = 1 and choose q such that q > q'. Then a positive constant C2 exists such that 111e(J)11q1 < C211J11q,2m
for all e E [0, 1]
and J E D".
Hence, for all e E [0, 1] and J E D", I E[Vk(e, w)1e(J)]I C 11Vk(e, w)IID,111e(J)IIq,
C,C211Jll,,,2me'.
Also lalE-1-
(D60) °fo
[F(e, w) -folale(J)
(9.37)
<- (DaO) ° fo E I«E [Rk-i a!
[F(e, w) - fo]aP1(e, w), D`J>a®,
DIFFUSION PROCESSES ON MANIFOLDS
390
and since [F(e, w) - fo]aP1(e, w) has the asymptotic expansion elafea.r.o(w) + siai+lea,1.1(w) + .. .
[F(e, w) _ f0]aP,(e, w)
in D-(H®`)
as a 10,
we have
(9.37) = Zo(w) + aZ1(w) + ... + ek-1Zk-1(w) + Uk(e, w) where
Z,(w) = 1=0 E iai+v-l
< 11(Da0) °.fo - ea.t.v. D'J>R®i. a.
We can easily find C3 > 0 such that I E[Uk(e, w)] l S C3ekIIJIIq.2m
for all s E [0, 1] and J E D. Set
-f. 01 =/=0 E lal+v1 E_ (D*)a (1(Da0) a.
ea r v),
I = 0, 1, ... , k - 1.
It is easy to see that
01 E
II
1
Dp.-2m
and this, combined with the above, yields k-1
I E[T°F(e, w)
J] - E etE[O,J] I < (C1C2 + C3)ekIIJIIq.2m 1-0
for all a E [0,1] and J E D=. Hence we have k-1
II ToF(c, w) -
-"01lip. -2m
C (Cl C2 + C3)ek
where I/p + 1/q = 1. It is clear in the above argument thatp E (1, oo) can be chosen arbitrarily and thus (9.36) is obtained. It remains only to show that 1 can be determined through the expression (9.34). Since
APPLICATIONS TO HEAT KERNELS
07 = E imiE='
(D*)i(- (Da((1
391
+ Ix1z - d/2)--T}) o f. ea,j."
we see that the mapping ,7'(Rd) E) T- 01 a D_ is continuous in the sense that for every n c- N, there exist s > 0, p E (1, oo) and C such that
<
CIIT1I-2n.
On the other hand, it is clear that 01 is uniquely determined from T and must be calculated from the expression (9.34) in the case T E S°(Rd). Hence, it must also be given by (9.34) for general T E 9"'(Rd). This completes the proof.
10. The case of stochastic differential equations: Applications to beat kernels Let X = (X(t, x, w)) be the solution of a stochastic differential equation on Rd given in the same way as in Section 2:
dX(t) = a.(X(t))dw"(t) + b(X(t))dt
(2.1)
IX(O)
=x
which is realized on the r-dimensional Wiener space (W, P). We assume,
as before, that all coefficients a0(x) = (c,(x)) and b(x) = (b'(x)) are C" with bounded derivatives of all orders >_ 1. We know from the results of Section 2 that, for almost all w EE W (P), (i) (t, x) -> X(t, x, w) is continuous, (ii) x -> X(t, x, w) is a diffeomorphism of Rd. If Y(t) = (Yf(t)) is defined by YXI)
ax711(t, x, w),
then Y(t) satisfies (2.16)
dY(t) = a'(X(t))Y(t)dw°(t) + b'(X(t)) Y(t)dt
IY(0)=I
where
Q.(x)i = axe' a,' (x)
and
b'(x)i
=
xb'(x)
DIFFUSION PROCESSES ON MANIFOLDS
392
and I = (o) is the d x d identity matrix. The equations (2.1) and (2.16), put together, define a stochastic differential equation for the process r(t) = (X(t), Y(t)) and r(t) takes values in Rd x GL(d, R) which can be identified with the bundle of linear frames GL(R") over Rd. If we introduce a system of vector fields Va, a = 0, 1, 2, , r by Va(x) = a,(x)
a
a = It 2, ... , r
t,
and
V0(x) = [b'(x) - 2 aE aa(x)5o (x)] ax`
then (2.1) and (2.16) can be rewritten as dX(t) = E Va(X(t)) C dwa(t) + V0(X(t))dt a-1
(10.1)
X(O) = x and
dY(t) = E aVa(X(t))Y(t) o dwa(t) + a Vo(X(t)) Y(t)dt a-1
(10.2) I
Y(0) = I
where (aVa(x))i = axi Valx)
Proposition 10.1. Let t > 0 and x E Rd be fixed. Then (i) X(t, x, w) E D (Rd), i.e., XI(t, x, w) E Dm, i = 1, 2, (ii) the Malliavin covariance a(t) = (o`"(t)), a`f(t) _
,
d,
a'J(t) _
[Y(t)Y(s)_1Va(X(s))]1[Y(t)Y(s)-1Va(X(s))]'ds. a-1 J
0
Proof. For X(t) = X(t, x, w), we define X("'(t) = X1n'(t, x, w) as
If A = (A',) is a d x d matrix and x = (x') E R°, we denote (Ax)' = A "X"
i=1,2, ...,d.
APPLICATIONS TO HEAT KERNELS
393
in Section 2: X (n) (t) = x +
(10.4)
(O.(s))dW"(s) +
Qa(X (1> 0
' b(X (n) 0
(On(s)))ds
where
if s e [k/2n, (k + 1)/2"), k = 0, 1,
0"(s) = k/2"
.
We know that X (n) (t) E S and I I X (") (t) - X(t)I!, --- 0 as n -* oo for every 1 < p < oo. It is easy to see that D,,X (") (t) _ (DhX (") "t(t)), , defined by DhX(n)'i(t)
=
heH
satisfies DhX(n)(t)
+ Jf
=f
(
a .(X (n)(On(s)))DhX (n)(O \ (s))dw'(s)
r
0
b'(X (n) (0.(s)))DhX (n) (0.(s)) ds + J 0 Oa(X (n) (0.(s))) ha(s)ds
and hence, if we denote :
DhX (") (t) = f i t ,)(v)hfi(v)dv,
h = (hfl(t)) E H,
0
thenr (v), t
v, ft = 1, 2, f` J wn(v)nt
(10.5)
+ft
, r, satisfies
o (X (n> (cn(s))) ` (s),a(v)dwa(s) bt (X (n)
stnW M
n(s),Q(v)ds +
(n) (Y (v)),
where yin(v) = k/2"
if v E ((k - 1)/2", k/2"], k = 0, 1, .. .
By applying a slight modification of Lemma 2.1 to (10.4) and (10.5), we
have, for every I < p < oo, (10.6) and
E[ sup I X(")(s) - X(s) l'] - 0 se [0,:]
DIFFUSION PROCESSES ON MANIFOLDS
394
sup E[ sup
(10.7)
Je [y, (]
us,St
I
S (v) - :. (v) I'] -- 0
as n - co where ,, f(v), t > v > 0, satisfies fit. ,6(v) = f t o'a(X
,(v)dwa(s)
(10.8)
+f Note that (10.9)
b'(X(s))cs. (v)ds + a (X(v)). v
(v) is uniquely determined by (10.8) and is given by c,,,(v) = Y(t)Y(v)-'6.,(X(v)),
fl = 1, 2,
, r.
Indeed, it is immediately seen that ,, (v) given by (10.9) is a solution of (10.8) and the uniqueness is obvious. This proves
X(t) E (1 D, ,(Rd)
for each t > 0
1
and
(10.10)
1 jv)1 fi(v)dv,
hEH
where g,,,(v) From this, (10.3) can be concluded. Repeating this argument as in the proof of Proposition 2.1, we see easily X(t) E D (Rd). Next we shall investigate the assumption (A.2), namely the condition (10.11)
(det tr(t))-' E
(1 I
L,
where o(t) = (v"f(t)) is the Malliavin covariance of X(t, x, w). It is easy
to see that a(t) can be written as (10.12)
E f"0 Q`'(t) = a-1
where fea(r), r e GL(Rd), is defined by (2.34). Hence (10.13)
6(t)= Y(t)6(t)Y(t)*
where *
Y(t)* is the transpose of Y(t).
APPLICATIONS TO HEAT KERNELS
395
3'1(t) = , f r fva(r(s))fva(r(s))ds. a=1 0
(10.14)
We have seen in Section 2 that
for all 1 < p < co
E[II Y(t)I Ip + II Y(t)-' 11p] < oo
and hence (10.11) is equivalent to (det d(t))-' E fl Lp.
(10.15)
1
Note that f,,(r(t)), V E X(Rd), is a d-dimensional Ito process such that fv(r(t )) - fv(r(0)) = E f r fc0va, v](r(s)) - dwa(s) a.1
(10.16)
+ f' fc vo, I(r(s))ds
as we know from (2.36) of Section 2. Now we obtain a general result on Ito processes by which we can discuss the condition (10.15). First we state several simple lemmas. In are positive constants the following, c0, c1i c2, and ao, a,, a2, independent of n = 1, 2, and w E W. Lemma 10.1. Let q: W--- R be a real Wiener functional. Suppose
that c i = 0, 1, 2, 3 exist such that P[ I t7 l < n-°o] S c1 exp[-c2n`3], n = 1, 2,
(10.17)
Then
.
for all p > 1.
E[ I >7 -p] < oo
The proof is obvious and is ommitted. Let 5Fd be the totality of dx d symmetric, non-negative definite matrices. For g = (g11) E $#d and 1= (1') E Sd-1 = { x E Rd; I x j _ 1 }, we set
= MIgII
d
d
{i,J-1 Z(g`')2}I"2
and
g(1) _ Eg`11`11. i,/-1
We write g1 > g2, g1, g2 E Yd if g, - g2 is non-negative definite. Lemma 10.2. Let >7 and 4 be .Sod-valued Wiener functionals such that >7 ? rl a.s. Suppose that c, , i = 0, 1, 2, 3, 4 exist such that Ir/II < co a.a.w. and
DIFFUSION PROCESSES ON MANIFOLDS
396
sup P[il(1) < n-01] < c2 exp[- c3nc4], for n = 1, 2,
(10.18)
lcSd-1
Then
for all
E[(det r/)-p] < oo
Proof. Let W
1 < p < oo.
w; q > it and 11111 < c, 1. Then P(W) = 1. Set
W.(1) = { w e W; #(I) > n-I }
for I E Sd-' and n = 1, 2,
.
Then (10.18) implies that P[ W (1)c] < c2 exp[- c3n`4]. If w E W,
Ij(1) -i1(1')I c 2coI1-1'I
for all
Clearly there exist It E Sd-1, i = 1, 2,
1, 1' E Sd-1
, m such that
an
U {I E Sd-1; I I - lk I < 11(4con°1) } =
Sd-1
k-1
m
where m < can cd-11 °1 for some cs. TL l lien, w e (1
implies that
k-1
inf #(I) > 1/(2n`1) 111-1
and hence det
1/(2n°1)d/2. Since d
det r1 > det r/
if w E W, we have
an
P[det ri < l/(2n°1)T] 5 P[{ (1 Wf(Ik)}°] k-1
< csn
By Lemma 10.1, we have E[(det rl)-D] < oo
for all 1
which completes the proof. In the following, we consider the case when = (11J) E .7d is given by f1 Er
`' =
0 a-1
fa(s)ff(s)ds
and s - fa(s) (=- Rd, a = 1, 2, . , r, are {Rt }-adapted continuous processes where { .fit } is the proper reference family on W for the rdimensional Wiener process w(t) = (w"(t)) canonically realized on
APPLICATIONS TO HEAT KERNELS
397
(W, P). By Lemma 10.2, we can easily conclude the following: Lemma 10.3. Suppose that there exist constants c1, i = 0, 1, 2, 3, 4
and {0, }-stopping times 0 < a1 < a2 < I such that I fa(s) I < co for s E (a1, a21, a = 1, 2, , r a.s. and
sup P[z
(10.19)
f12
a-1 al
IeSd-1
[1
fa(S)]2dS < n-cl] s c2 exp[- c3n`4],
for every n = 1, 2,
,
where d
i -rats) = E 1fa(S). f=1
Then
for all 1 < p < co.
E[(det ri)-P] < oo
Proof is immediate if we notice 11(1)
= a=1 E
f1
0
[1
fa(s)]2ds
and choose
`' = E fa2fa(s)I' (s)ds, a=1 al
i,j=1,2,...,d,
in Lemma 10.2.
As usual, we call a continuous (-q,)-adapted process 4(t): [0, oo) . . , r, if
R an Ito process with the characteristic ta(t), a = 0, 1,
m) = S(0) +a=1 E f'0 Sa(S)th x(s) + f' y0(s)ds 0
where ha(s), a = 0, 1, such that
, r, are
(
.}-adapted measurable processes
r
f o I Us) I Zds +
Lo(s) I ds < oo
for every t > 0, a.s.
fo
We are now in position to state a key result for Ito processes which plays a fundamental role in obtaining sufficient conditions for (10.15). This result is first obtained in a weaker form by Malliavin [107] and then improved by ideas of Kusuoka and Stroock (cf. [211] and [212]).
DIFFUSION PROCESSES ON MANIFOLDS
398
Lemma 10.4. (Key lemma). Let i(t) be an Ito process with char, r. Assume furthermore, that ko(t) is acteristics fa(t), a = 0, 1, also an Ito process with characteristics to, a(t), a = 0, 1, - , r. Suppose that there exist c1, i = 0, 1, 2, 3 and {.q, }-stopping times 0 < a1 < az
< 1 such that (i) for almost all w (P) r I
r
(S) I +E I a(S) 12 + 14o(S) I + E 1 4o. Q(S) I -< Co
for all s e [al, az] and (ii) for every n = 1, 2, a1 <
]
< cl exp [- c2n°3].
n
Then for every given c4, there exist c1, i = 5, 6, 7, 8 (which depend only on c0, c1, c2, c3 and c4) such that r
o2A(ol+I/n)
(10.20)
P[f o1
o2A(o1+1/n)
14(s) I gds S l/n°5, E fof
14a(s) I gds
1/na4] S c6 exp [- c,n°$]
for every n = 1, 2, Here we present a proof of this lemma along the lines of [203] and [232]. It should be remarked that a much simplified different proof was given recently by Norris [222]. First we give a series of probabilistic lemmas.
Lemma 10.5. Let K > 0 and X(t) be a one-dimensional continuous semimartingale
X(t) = X(0) + m(t) + A(t) such that <m>(t) = f o a(s)ds, A(t) = f /3(s)ds and I a(s) I < K, I a(s) o
< K. Then
P(aa < A) < 4 exp [-
a2/(8K2)]
for all
a> 0
and
A E (0, a/2K]
APPLICATIONS TO HEAT KERNELS
399
where o, = inf { t; I X(t) - X(0) I > a). Proof.* By Theorem 11-7.2', there exists a one-dimensional Brow-
nian motion b(t) (b(0) = 0) such that X(t) - X(O) = b(<m>(t)) + A(t). Since { I X(t) - X(0) I > a } c { I b(<m>(t)) I > a12) U { I A(t) I > a/2 }, we have Qa > (a/2K) A (aa12/K) where 0,12 = inf { t ; I b(t) I > a12). Hence, if 0 < d < a12K
P(oa<1)
<
e- - 2 / 1 , -dx < J a/2
<
(K). exp[- a2/(8K).)})
exp[- a2/(MA)]
This completes the proof. Before stating the next lemma, we shall prepare some notation. For a finite interval I and a square-integrable function a(s) defined on I, we set
o,(a) = I j f r [a(s)2 - [22]ds I
where
a=
r f a(s)ds. I
It is easy to see that o1(a1 + a2) < ol(al) + o,(az)
Lemma 10.6. Let b = (b(t)) be a one-dimensional Brownian motion.
Then for every a > 0 and E > 0, (10.21)
P(oro a,(b) < 8)
exp - 2,82 a ].
Proof. Without loss of generality, we may assume that b(0) = 0. We use the following well known expansion formula of b(t) in Fourier series: * The proof was essentially given in Theoren IV-2.1. We shall repeat it however because of its importance.
DIFFUSION PROCESSES ON MANIFOLDS
400
= to +
[b(t)
At
(cos 27rkt - 1) 2rrk
+ ilk
sin 2irkt 1
2nk
J
where { Sk, r/k } are independent random variables with common distri-
bution N(0, 1) ([75], [136]).* Then
b(t) -j b(s)ds = (t - 1/2) 0 + .// T t [4k cos 2i kt + Ilk sin22iv t 27rk ] and since { cos22rkt } and { (t - 1/2), sin 27rkt } are mutually orthogonal in X2([0, 1]), we have 62
= Qio, ii(b)
k-i kl(2nk)2.
Hence for z > 0, E(e 12o2) < E(exp{ - 2z2 E 2 k/(2nk)2 }) k-1
J1 J E(exp{
-
Z142
k/2 r2k2 }) = 11 (1
+z2/iv2k2)-112
= v z/sinh z< y/2e-S 12. Consequently
P(cr < e) < exp[2z282] E(exp[- 2z2Q2]) < /-f exp[2z282 - z/4]
and by setting z = 1/1682 we have (10.22)
P(or < s) < /T
exp[_
L
21 2 a2
Since { )_b(at)}. { b(t) }, we have
13(b)
°3(b). Then(10.21)
follows from (10.22)
Lemma 10.7. Let b(t) be a one-dimensional Brownian motion on [0, A] where A is a positive constant. Then, for every 0 < y < 1/2, there exist positive constants a1i a2, a3 such that (10.23)
P[ sup L,SE(O. A] 1$S
I b(t)
- b()
I t - sly
I
> n] < alexp[- a2n°3]
* N(0, 1) stands for the normal distribution with mean 0 and variance 1.
APPLICATIONS TO HEAT KERNELS
for every n = 1, 2,
401
-
- -
.
Proof. First we note (10.24)
P[ sup
t,se[O,A] t*s
I b(t) - b(s) I < oo] = 1 It - SIY
for every 0 < y < 1/2.
Indeed it is proved in the proof of Corollary to Theorem 1-4.3 that
P[ sup IX(t)-X(s)I <00]=1 r,se[O,A]
t*s
It-SI
for any continuous stochastic process X(t), t e [0, A], satisfying
E[IX(t)-X(s)Ia)<MIt-sI'+B,
t,sE[0,A],
where a, /3, M are positive constants and a is any constant such that 0 < a < fl/a. Since
E[Ib(t)-b(s)I2m]=(2m- 1)(2m-3) ... 3 x 1 x It-slm,
m=1,2,-..7 (10.24) is now easily concluded. For one-dimensional Wiener space (Wo , P) with time interval restricted to [0, A], set
Wor={wE W0;IIwIIY
05sSA
I w(s)I +05s
Then Wo r is a Banach space which is however not separable and, if 0 < y < 1/2, (10.24) implies that P(Wo r) = 1. If WY is the closure of the Cameron-Martin Hilbert space H with respect to 11 I I ,-norm, then WY is a separable Banach space with respect to II IlY norm and it is
easy to see that Woi r, C WY
if 0 < y < y' < 1/2.
In particular, P(WY) = 1 for every 0 < y < 1/2. Then applying Fernique's theorem to the separable Banach space B = W. for 0 < y < 1/2,
DIFFUSION PROCESSES ON MANIFOLDS
402
we have
for some e > 0.
K.: = E[exp{sllwllf }] < -_ Hence
- b(s) I > n] P[ sup Ib(t) It - SI' tSA 0!9$
S P[Ilwllr > n] < Keexp[- en2],
n = 1, 2, -
-
For completeness, we give a proof of Fernique's theorem following Kuo [210).
Fernique's theorem: If B is a separable real Banach space with norm
II and u is a mean zero Gaussian measure (i.e. { 1(x), 1 E B' } c L2(B,,u) and it is a mean zero Gaussian system with respect to u). II
Then s > 0 exists such that Ja
exp[ellxll2]k(dx) < 00.
Proof. Consider the product probability space (B x
x+Y
x-y
and
/ 2
(D ,u). Then
for (x,y)EBxB
are B-valued random variables over this probability space such that they are mutually independent and both of them are p-distributed.
Hence, for t < s < 0, p(IIxII <_ s),u(IIxII > t)
O,u(Ilx-YII <s)pO,u(IIx+YII >t)
u
2
>t) =a® (Ilx - YII <sand IIx+YII ,./2 /
S
> %7
P(IIxII
= u(IIxII >
t
2s
)Z
and Ilyll > t
s1
APPLICATIONS TO HEAT KERNELS
For given e > 0, define t,,, n = 0, 1, ,/-f tn. Then ( / .L )n+1 - 1
to = 11/ 2 - 1
-
403
by to = s and ti}1 = s +
s = ((- 7)"' - 1) (-,I2 + 1)s.
Set
an = p(Ilxll > tn)lp(I xll S S). Then an+1 = p(Ilxll > s + %/ 2 tn)/u(Ilxll < s)
= iz(Ilxll > S + ./ 2 tn)AI xI i < s)/(u(llxll < S))2
and applying the above obtained inequality to the numerator, we obtain an+1 C u(llxll > tn)2/(.u(Ilxll C s))2 = an.
Hence
an S am = exp[2n log ao], that is n+1
4a(IIxii > (2
2
- l)(21 + 1)s) < iz(llxll < s)exp[2° log ao], I
n=1,2, .
Choose s so large that p(Ilxll < s) > 0 and ao =,u(llxll > s)/,u(llxii < s) < 1. Then the estimate obtained above can be written in the form p(llxll > u) < b exp[- au2]
for all u > c
with some positive constants a, b, c and the Fernique theorem follows
at once from this. Let ra(t), t E [0, 1], be an Ito process with bounded characteristic 'ia(s). By Theorem 11-7.2', we can find a Brownian motion b(t) such that rt(t) = ;,7(0) + b(a(t)) + f o ?7o(s)ds
404
DIFFUSION PROCESSES ON MANIFOLDS
where
a(t) _ a-1
f
r 0
17a(s)2ds.
Hence Lemma 10.7 implies immediately the following: Lemma 10.8. Let 17(t), t E [0, 1], be an Ito process with bounded characteristics.
Then c,, i = 0, 1, 2 exist such that (10.25)
I17(t) - rll 3)
P[ sup
os:
I
t-sI
> n]
co exp[- C1n`2]
forn=1,2, Finally we need the following real variable lemma due to Kusuoka and Stroock.
Lemma 10.9. Let f(t) be a continuous function on [a, b] and set g(t) = g(a) + f ' f(s)ds,
t c= [a, b].
O
If a
sup Sb
I f(t) -f(s)I It-SI1/3
n
and
f.' I f(s) gds > s2 for some s > 0 such that s3 < 22K3(b - a)512, then
(10.26)
(b - a)a2Ia.b](g)
2-133-IK-9611(b - a)-11/2.
Proof is easily provided if we note that, first to E [a, b] exists such
that
If(t0)I
>s(b-a)-1/2
and then an interval I, I e [a, b], exists with length s3K-32-3(b - a)`3/2 on which f is of constant sign and I f I > s2-1(b - a)-1/2. We now note
that there exists t1 E I such that
APPLICATIONS TO HEAT KERNELS
g(t)) _ (gl,), (Sl r) = 11-1
405
g(t)dt. J
Then
(b - a)a2ta.b)(g) ? f (g(s) - g)2ds
f, { f
r'
f (s)ds } 2dt
z 82 4-1(b - a)-'f (t - t,)2dt > (48)-'s2(b - a)-)III 3 which completes the proof. Proof of the key lemma. In proving (10.20), we may clearly assume are positive constants independent of n and w. We know(Theorem 11-7.2') that a one-
n ? 2 and co > 1. In the following a, and d1, i = 1, 2,
dimensional Brownian motion B(t) with B(O) = 0 exists such that
F,(t) = 4(o) + B(A(t)) + g(t), a, < t < 1 where
= A(t)
5f a-1
d
I ba(s) I gds
and
g(t) = f'al
4o(s)ds.
Set zn(o,+t/n)
W,
a=0
I 4a(t)12dt ;-> 1/nc4]
f',,'
and
w2 = [ f
I fi(t) 12dt < 1/n°5].
Here c, is given and c5 will be determinded later. Set o2A(a,+1/n)
Wi,
[f
o,
I o(t)12dt > 1/(2n`4), A(a2 A (a, + 1/n)) < 1/n-1]
and
W1.z = [A(az A (o + 1/n)) > 1/nal].
DIFFUSION PROCESSES ON MANIFOLDS
406
Then W1 c W1.1 U W1, 2
(10.27)
if a,
c, + 1
since
l/nal < 1/n`4+1 < 1/(2n°4)
for n ? 2.
Set
W3 = [as - a, ? l 1n], W4 = [
SUP
°15
10(t) _ 40(3)1 < n] I t- S I 1/3
and
W5 = [ sup I B(t )1 < 1 /n°2]. 0SK1/na1
By Lemma 10.9, we can find a3 such that W3 n W4 n [f I 0(t) 12dt >_ 1/(2W4)] In
(10.28)
c W3 n [a n(g) > 1/n°3]
where In = [a,, a, + 1/n]. Hence W3 n [0,2 (g) < 1/n°3, f I do(t) I Zdt >_ 1/(2n`4)] !n
(10.29)
c w3 n w4. We have w1,1 n w5 n W3 c [sup I B(A(t)) I < 1/n°2] n w3 tEln
C [an(B(A( ))) < 1/n2a2] n w3 c [a;,(B(A( ))) < 1/(4n°3)] n W3
if a2 > a3/2 + 1. Also
APPLICATIONS TO HEAT KERNELS
407
W2 n w3 c [fin 14(t) I gds/ I In I < l/n°s-'] n W3 C [ai,(k) < 1/n`5-1] n W3 C [o n W3
if c5 > a3+3.
Since
a,n(g) :!5: a,,,(B(A(
))) +
if a2 > a3/2+1 and c5 ? a3 + 3, w), n w2 n w3 n w3 (10.30)
C [Q n(g) <
1/na3] n W,,, n W3
CW; by (10.29). Hence, if a, > c, + 1, a2 > a3/2 + 1 and c3 > a3 + 3,
W,.,nW2CWWUW UWs. By Lemma 10.5, (10.31)
P[WS] < d,exp[- denal-2a2]
if a, > 2a2.
Hence we can conclude, by the assumption (ii) in Lemma 10.4, Lemma
10.8 and (10.31), that if c5 > a3 + 3 and a, > (a3 + 2) V (c4 + 1), (10.32)
P[W11 , n W2] < d3exp[- d,n,s].
Next we estimate P(W12 2 n W,). The idea is to use the fact a,(B(A( ))) is comparatively larger than a,(g) if the interval I is sufficiently small. If w E W3i I. C [6,, v2]. We divide I. into nm equal intervals
In.k = [Q) + k/n'+m, a, + (k + 1)/n' +m], k = 0, 1, ... , n^'- 1
where we choose m e N such that (10.33)
m > a,.
Then if w e W3 $In k I k(t) 12dt = $ Ink 14(al) + B(A(t)) + g(t) 12dt
> f I k(o) + J A
B(s) + g(A-)(s)) 12dA-)(s)
DIFFUSION PROCESSES ON MANIFOLDS
408
>
1
14(c'1) + B(s) + g(A-'(s)) I Zds.
A (Jn.k)
CO
Here we note that E I ha(s) I
co
a-1
by the assumption (i) in Lemma 10.4 and we set A(II, k) = { A(t); t E In. k }. Set
Jk = [A(6, + k/n'+m), A(o1 + k/n'+m) + 1/na,+m],
Then, setting g(s) = g(A-'(s)), W3 fl
{
I A(I,,, k) I ? 1/na'+m }
= W3 f l { A(IJ. k)
c W3 n {
J
Ink
Jk }
I fi(t) I Zdt >_
c W3 fl if
if co
I t(t) I Zdt
I
I fi(t) I Zdt >
I
Jn.k
I R(ay) + B(s) + g(s) I Zds }
g(' )) }
Co
Jn,k
c W3 n { f
Jk
I
co
[7'1(B) - oJk(g)]Z }
Since
g(s) = J r 40(u)du
and
140(u) I
<_ co on [a,, Qz],
we have
I g(t) - g(s) I < co I A-' (t) - A-' (s) I .
Therefore, setting t0 = A(o, + k/n'+m),
(7 40)
I Jk
S coZ
I
IJkI
f
f
J
I g(t) - g(t0) 12dt
k
Jk
(A-1(t) - A-1(t0))2dt
APPLICATIONS TO HEAT KERNELS
co[A-'(A(a, + (k +
409
1)/nl+m))- A-'(A(6, + k/nl+m))]2
C2/n2+2m 0
W3 n { I A(In, k) I > 1Ina1+m } and hence Jk C A(In, k). Therefore
if w
W3 n [O',/k(B) > 2Cp/nl+m, I A(I4, k) I > 1 /nal+m }
C W3 n [f
14(S) 12dS s
2
J
I CO
nC+m
I
n2+3m+a1
'n. k
Set
W6 =nm-, n0
{
6,k (B) > 2cp/n]+m }.
Since nm-1
A(C2 A (Q1 + 1/n)) = A(vl + 1/n) _k-0 Z IA(I,,.k)I > 1/na' on W3 n W1.2,
there exists k such that
I A(I,,, k) I
>_ 1/nal+m Hence
w1, 2 n w3 n W6 n,a-1
CU k-0
{
I A(In, k) I
1/na'+m' alk(B) > 2Cp/nl+m } n W3
nm-1
Cp/Y12+3m+a' } n W3
U .k I b(S) 12dS C k0 ln ('o2h >i +l ln)
C{
J
I 4(S) I
2
dS > 1/n 2+3m+4 1 }
a1
since we assumed co >_ I. Therefore if cs ? 2 + 3m + a,,
w1.2 n w2 c w3 u W6 and by Lemma 10.6 (note that A(u, + k/21+m) is a stopping time for B(t)), nn+-1
P[W6] S E P[v,k(B) < 2Cp/n'+ml k-0
nm-1
S E d6exp[- d, I Jk I k-0
d8exp[- d9nm+2-al]
1
(1/n1+m)-2]
DIFFUSION PROCESSES ON MANIFOLDS
410
d8exp[- d,n2]
because of (10.33). Noting the assumption (ii) in Lemma 10.4, we have thus obtained (10.34)
P[W12 2 n W2)
- d10exp[-
Therefore, if c5 is chosen sufficiently large, (10.32) and (10.34) imply the
estimate (10.20) because W, n w2 c (W,,, n W2) U (W,, 2 n w2). This completes the proof of Lemma 10.4.
Remark 10.1. In the key lemma, the variable n E N can be replaced obviously by a continuous variable T E [1, oo).
Remark 10.2. By examining the above proof carefully, we can deduce the following: Let 0 E (0, 1] and we assume in Lemma 10.4 that 4(t), ?;a(t), u, and a2 may depend on the parameter 0 but they satisfy the same assumptions (i) and (ii) where the constants CO, c1, c2, c3 are in-
dependent of 0. Then, for any given c, there exist c,, i = 5, 6, 7, 8, 9, 10 (which depend only on c0, c1, c2i C3, c4 and hence are independent of 0,
in particular) such that P[ d'2A(71+1I
J
a,
r
('o2h(a,i 1/n)
/, E J
(s) I2ds S o n°5
S c6nc7exp[- c80`9n`10]
a-O o,
2
I Za(s) I ds
for all n = 1 , 2,
01n1.]
and 0
(0, 1].
Now we return to the stochastic differential equation (10.1). We inX(M),
troduce the following notation: for V
(Va,V)=[Va,V]
a = 1, 2,
,r
(10.35)
(Vo, V) = [V0, V] + 2
[Va, [Va, V]].
Then, by (10.16), we have 1 - fv(r(t)) - I -.f,,(r(0)) r
(10.36)
= a=1 E J r0 1
- f[ va, 17(r(s)) o dwa(s) +
= E J o I- A',', v)(r(s))dwa(s) + f
f 0
r o
I - .f v0. 3(r(s))ds
1-.fcvo, )(r(s))ds
APPLICATIONS TO HEAT KERNELS
411
for every I E Sd-1 and V E 3; (Rd). For each n = 0, 1, set E. e X (Rd) successively by
,
define a
E,o={V1, V2, ...,V,} (10.37)
n={(Va, V);
,r} for n
1, 2,
and set
±n = Eo U E1 U ... U F,.,
(10.38)
for n = 0, 1,
.
Now we introduce the following hypoellipticity condition of Hormander
type for the vector fields Va, a = 0, 1,
, r.
, r satisfy the assumpAd E £M Ad(x) are linearly independent.
Definition 10.1. We say that Va, a = 0, 1,
-
tion (H) at x e Rd if there exists M E N and A1, A2, such that A1(x), A2(x), -
- -
,
Cleary (H) is satisfied at x if and only if there exists M E N such
that
inf E [1 - A(x)]2 > 0.
(10.39)
leSd-1 AeZM
If (10.39) is satisfied, then we can find S > 0 and bounded neighborhood U(x) in Rd and U(I) in GL(d, R) of x and I = (8;) respectively such that
inf E [1
(10.40)
f4(x)]2 >- 8
if r E U(x) X U(I).
leSd-1 Ae±M
Set a = inf{s; r(s) U(x)x U(I)} where r(s) is the solution of (10.1) and (10.2) put together. Set a, __ 0 and a2 = a A 1. Then a, and a2 satisfy the condition (ii) of Lemma 10.4 by Lemma 10.5. Cleary inf
1aSd-1
r02+ (a 11-1 /n)
J a1
{ Z [I - f4(r(s))12 Ids > 8/n AaZM
on the set W. = [a2 - 91 >- 1 /n] and
alexp[- a2n], n = 1, 2,
- -
. Let N
Then for
DIFFUSION PROCESSES ON MANIFOLDS
412
every I
Sd-1, we can find ao, a1, 1
,
ak, 0 S k < M such that
l n (0'2+1/-)
(10.41)
J al
[! f Vao.
ai ... , ak(r(s))]zds ? 8/(nN)
on W.
where
al. .... ak = (Vak, (Vak-1, (... (Val, V0)) ... ).
Va0.
Noting (10.36) and applying the key lemma successively, we can easily
deduce the following: For each j = 0, 1, , k, there exist positive constant cl, j, i = 1, 2, 3, 4, independent of n = 1, 2, and I E Sd-1
such that o2n(a1+1/n)
P[fJ o
[! ' f'ao, al, ..., aj (r(s))]2ds < n-0,.1] 1
<- cz.1exp[- C3,jna4.>]
for all n = 1, 2,
.
In particular, we can conclude that positive constants c,, i = 1, 2, 3, 4 exist independent of n such that
(10.42)
sup P[r az [I. f,(r(S))]2ds < n--l] a-1 J al
Ie Sd-1
S czexp[- cina4]
for every n = 1, 2,
By Lemma 10.3 and (10.14), we obtain 11(det 6(l))-11I, < oo
(10.43)
for all 1
Now introduce a parameter 1 > e > 0 and consider the following stochastic differential equations r
dX(t) = EE Va(X(t)) o dwa(t) + e2VO(X(t))dt a-i X(O) = x
(10.1)6
and
dY(t) = 81,EaVa(X(t))Y(t) e dwa(t) + ezaV0(X(t))dt (10.2)6
a-1
{(
Y(0) = I.
Denote the solution by r6(t) = (X8(t), Y6(t)). It is immediately seen that
APPLICATIONS TO HEAT KERNELS
413
( r(e2t), t ? 0 ) is equivalent in law to (re(t), t ? 0 }. In particular, r(e2) is equivalent in law to re(1) and 3e(1) defined similarly for re(t) is equivalent in law to 6(82). Suppose that ( Va, a = 0, 1, , r) satisfy the condition (H) at x. We can give the same arguments as above for the vector field Vq, e E_ (0, 1], where Va = SVa, a = 1, 2, - , r and Vg = &V0. -
In this case, however, b > 0 above is replaced by 8ek for some k e N and 6 > 0 which is independent of e e (0, 1] and we apply the key lemma in a generalized form as stated in Remark 10.2. Finally we obtain, instead of (10.42), su P [E a=i
tnd-1
Z [1 f vs (re(s))]2ds < ekn-ai]
for n = 1, 2,
and e E (0,1] where c1, i = 1, 2, 3, 4, 5, 6 are independent of n and e. Here all and a2 are also defined in the same way for
the process r8(t). From this we can deduce, as above, the following estimate:
P[det 38(1) S
T]
< a,Tazexp[- a3ea4Ta5]
for all T E [1, oo) and e E (0, 1] where a1, i = 1, 2, 3, 4, 5 are positive constants independent of e and T. We can now conclude that (10.44)
II(det de(1))-hII = II(det 3(e2))-'Il, <
K1(p)e-K2
for all pE(1,oo)and eE=(0,1] where K1(p) > 0 may depend on p but not on a and K2 > 0 does not depend on p and e. Thus we obtain the following result due to Kusuoka and Stroock. Theorem 10.2. Let V0, V,, , V. satisfy the assumption (H) at x Er Rd. Then, for every t > 0, X(t, x, w) E= D_(Rd) and it satisfies (A. 2): More precisely, there exists a positive constant K2 and for every
I < p < oo, there exists a positive constant K1(p) such that (10.45)
II(det a(t))-'IIp <
KI(p)t-K2
for all t E_ (0, 1] and 1 < p < oo. Remark 10.3. If (H) is satisfied everywhere in a domain D c Rd, the
DIFFUSION PROCESSES ON MANIFOLDS
414
estimate (10.44) holds uniformly in x E D for every ID clear from the above proof.
D. This is
Thus if (H) is satisfied at x E Rd, then for any Schwartz distribution
T E 7 '(Rd), T(X(t, x, w)) E D_ is defined for every t > 0. In particular, 8y(X(t, x, w)) is defined for every y E Rd and p(t, x, y) = E[8y(X(t, x, w))]
which is C" in y, coincides with the fundamental solution of 8u = Au
at where
r
A= -E Va,+V0 2 a_1
u(t, x)
= fRd p(t, x, y)f(y)dy
solves the initial value problem 8u = Au,
at
ur,_0 =f
uniquely among tempered (i.e. of polynomial growth) solutions for a tempered function f on Rd. More generally, for g(w) E b.,, E[g(w)by(X(t, x, w))]
is well-defined. In particular, if C(x) is a tempered C°-function such that
C(x) < K for some constant K > 0, then g(w) = exp[ f ' C(X(s, x, w))ds] E D_ and
E[g(w)oy(X(t, x, w))] = p°(t, x, y) * The vector field V. is regarded as a differential operator.
APPLICATIONS TO HEAT KERNELS
415
is the fundamental solution of au = (A + C)u.
at
Now assume that V., a = 0, 1, -
-
, r, are bounded on Rd and
satisfy (H) at x E Rd. Let O(f;) be a C--function on Rd such that O(t;) = 1
if 141 < 1/3 and 0(4) = 0 if I I > 2/3. Let y E Rd and define
u(z)_
10(
(_z-y 1X-Y1 ) 1
Then
if
x*y
if
x = y.
Way = ay and hence
p(t, x, y) = E[ay(X(t, x, w))] = E[V(X(t, x, w))by(X(t, x, w))].
By the integration by parts as discussed in Section 9, this is easily seen
to be a finite sum E[a,(X(t, x, w))F,(w)]
where a,(x) is of the form of a finite sum bi(x)DO#V(x)
with bounded continuous functions b,(x) and F,(w) E D. is a polynomial in the components of X(t, x, w), their derivatives and [det a(t)1-1. By Lemma 10.5, we can deduce that P[ I X(t, x, w) - x I
>-
I x - y I /3] < axeXp[- a2 I x - Y 12/t ]
where a, and a2 are positive constants independent of x, y E Rd and t E [0, T]. Combining this with (10.44), we obtain the following estimate:
Theorem 10:3. Suppose that Va, a = 0, 1,
, r, are bounded
and satisfy (H) everywhere in a domain D of Rd. Then, v > 0 exists and
for every compact set K C D and T > 0, positive constans c, and c2 exist such that (10.46)
p(t, x, y) S t° exp[- c2I x - y 12/t] for all
t E (0, T], x E K and y c- Rd.
DIFFUSION PROCESSES ON MANIFOLDS
416
Next, we shall study the short time asymptotics of the fundamental
solution p(t, x, y). For this we introduce, as before, a parameter e E (0, 1] and consider the equations (10.1)` and (10.2)e. We denote the solutions by Xe(t, x, w) and Ye(t, x, w), sometimes more simply by XX(t)
and P(t), respectively. Then we have (10.46)
x, y e Rd.
p(e2, x, y) = E[by(Xe(1, x, w))],
We introduce the following notation: For a = (a1, a2, {0, 1,2, ,r}m,m= 1, 2, ... , we set
am) E
hail=m+#{v; Also, let Sa(t, w) = f0 o dwa1(t1) (`' ° dwa2(t2)
frm-1 0
0
o
dwrm(tm)
be a multiple stochastic integral in the Stratonovich sense for a = (a,, a2,
,
am) E {0, 1, 2,
, r j- where we set w°(t) = t.
Theorem 10.4. Let x E Rd be fixed. Then X8(1, x, w) has the asymptotic expansion (10.47)
(10.48)
d as a J. 0 in D-(R)
X8(1, x, w) tifo + efi + e2f2 +
and f E D-(Rd), n = 0, 1,
D. (Rd)
, are given by
f0=x
and (10.49)
f = E Vam . Vam-1 ... Va2(Va)(x)Sa(1, w), a;Iag-
n=1,2, .
Here the vector field Va, regarded as a differential operator, is denoted by Ya, i.e.
Va(f)(x) = VV(x)ax, 2f (x),
.f (Z C-(Rd -* R)
and
VV(y) = (V«(y)) E C-(Rd -> Rd),
a = 0, 1, ... , r.
APPLICATIONS TO HEAT KERNELS
417
In particular, fl(w) _ E Va(x)wa(1).
(10.50)
ac,
The expression in (10.47) is uniform in x e K for any bounded set K in
Rd.
Proof. The proof is easily provided by successive applications of the Ito formula:
X8(1, x, w) - x
=8
a-i fo
Va(X8(S)) o dw"
0
aE Va(x)wa(1) + E
+ s'
o
J0
[Va(Xo(s)) - Va(x)] ° dWa
Vo(X8(s))ds
Ef + 82[i &1-
+f
+ elVo(X8(s))ds fl
a
Va2(Vai)(XE(u)) o dWa2(u)} o dwai(S)
J 0{ J 0
Vo(X 8(s))ds + E
f o{ f o
o r
Vo(Va)(X 8(u))du } o dwa(s)]
r
= efi + e2[Vo(x) + a, =i a2-1 E + 8 [ aE1 a2E f 't
f'
Va2(Va,)(x)Sca,.a2)(1)]
E(u))
Va2(Va,)(x)) ° dwa2(u) }
o dwa'(s) + f [Vo(X8(s)) - Vo(x)]ds o
+ ei f j f
V0(V.) (X8(u))du } o dwa(s)] 0
and so on. Continuing this process and applying Theorem 111-3.1, it is easy to see that X8(1, x, w) =fo + ef1 + ... + ,nfn + O(En+i)
in L,(R") for every p E (1, oo). Since DkX8(t, x, w) [h,, h2i
, hk]
are determined successively by stochastic differential equations as we saw in the proof of Proposition 10.1, it is easy to conclude that DkX8(1, x, w) = f0) + fl(k) + ... En ,(k) + 0(e+1)
DIFFUSION PROCESSES ON MANIFOLDS
418
in L,(H®k (3 Rd) as a J, 0. This completes the proof. X8(1, x, w) is not uniformly non-degenerate in the sense of (9.32) because f0 = x which is completely degenerate. However, if we consider
F(E,w)=
X8(l,x,w)-x,
(0, 1].
e
then we have the following: Theorem 10.5. The family F(e, w), e E (0, 1] defined above, satisfies (9.32) if and only if (Atf(x)) defined by A"(x) = E V.(x)V., (x) amt
is non-degenerate, i.e. det(A"(x)) > 0. Proof. Since in
F(e, w) - f1 + e. f2 + -2f3 + ...
where ft are given in Theorem 10.4 and the Malliavin covariance off, coincides with A(x) = (Atf(x)), A(x) must be non-degenerate in order that F(e, w) satisfies (9.32). Conversely, suppose that det(Atf(x)) > 0. Denoting by a(a) = (a"(e)) the Malliavin covariance of F(e, w), we have by Proposition 10.1 that a(e) = f 10 Set
r = inf{s; (Y8(s))-1A(XS(s))((YS(s))-1)* < A(x)/2}.
Applying Lemma 10.5, it is easy to see that P[a < 1/n] < clexp[- c2n`3],
n = 1, 2,
,
where c, i = 1, 2, 3, are positive constants independent of e E (0, 1] and n. Note also sup jldet(Y8(1))-111, < oo
86(U,1)
for all
p E (1, oo)
APPLICATIONS TO HEAT KERNELS
419
because (Y8(t))-l is the solution of the equation (2.27) in which a,', and b' are replaced by eca and e2b' respectively. Then det a(e)
(det Ye(1))2det f
IAT
(Ye(s))-1A(Xe(s))((Ye(s))-')*ds 0
2 (det Ye(1))Zdet(A(x))(1 A i)d
and we can now easily conclude that sup Ildet(a(E))-'llp < co
for all p E (1, oo),
se (o.1)
which complete the proof.
Suppose that det(A(x)) > 0. Then by Theorem 9.4, T(F(e, x, w)), T E .5(Rd), has the asymptotic expansion in f)--. Since 6.(Xe(l, x, w)) = 8X(x + eF(e, w)) = e'8o(F(E, w)), we have the following:
Theorem 10.6. Suppose that det(A(x)) > 0. Then 8,(X8(1, x, w)) has the following asymptotic expansion in b-. as a 10: (10.51)
8,(X8(1, x, w)) ^' a-d(0o + eS1s1 + 82452 + ... )
and 1 k E D__ can be obtained explicitly by Theorem 9.4: (10.52)
0o = ao(fi)
and
(10.53)
tkk(w) _
I
a
I
1
n2 ... fn'
fork= 1,2, where f. = (f9 is given by (10.49), f1 by (10.50) in particular. In (10.53), , the summation extends over all a = (a,, a2, , a,) E { 1, 2, such that n1 + n2 d}' and n = (n1i nz, , n,), n1 ? 2, 1 = 1 , 2,
+ ... +n,-1=k. Also D,,= aal aa2 ... as , = 1 if a = (a,, az,
.
ak =
xk
and a
, a).
From this we see that E[Ok(w)] = 0 if k is odd since, then, Pk(-w)
DIFFUSION PROCESSES ON MANIFOLDS
420
= - 0k(w) and the mapping: W D w -a - w E W preserves the measure P. Corollary. Suppose that det A(x) > 0 where A(x) = (Atf(x)). Then At, x, x), t > 0, x E Rd, has the asymptotic expansion (10.55)
p(t, x, x) ^' t-d12(Co(x) + c1(x)t +
and ct(x) is given by (10.56)
i = 0, 1,
ct(x) = E[O21],
where 0t is given by (10.52) and (10.53). In particular CO = { (27C)d det A(x) 1-1/2.
Remark 10.4. We can discuss by a similar method the asymptotic expansion of p(t, x, x) in the degenerate case det A(x) = 0 under some hypoellipticity condition and also the expansion of p(t, x, y), x * y, as t ,j 0, cf. [192], [212], [214], [215] and [228].
We now consider an application of our results to heat kernels on manifolds. Let M be a compact Riemannian manifold and b e X(M). Let e(t, x, y) be the fundamental solution of the heat equation at = 2 Amu + b(u)
uit-o =f where Am is the Laplce-Beltrami operator for the Riemannian metric
g on M i.e. u(t, x) = fm e(t, x, y)f(y)m(dy)
(m(dy): the Riemannian volume)
solves the above initial value problem. Let r(t, r, w) = (X(t, r, w), e(t, r, w)) be the stochastic moving frame realized on the d-dimensional Wiener space (W, P) by the solution of stochastic differential equation (4.43) with the initial value ro = r E O(M). Let S, be the Dirac 6-func-
tion at y E M, i.e. the Schwartz distribution on M such that 6Y(O) = 0(y)
for every 0 E F(M).
APPLICATIONS TO HEAT KERNELS
421
Then we can define S,,(X(t, r, w)) E D_ and e(t, x, y) can be expressed as
e(t, x, y) = E[6,.(X(t, r, w))]
where ir(r) = x. Actually, we can extend our results obtained so far to the case of manifolds: We can develop a general theory of Wiener mappings with values in manifolds and pull-back of Schwartz distributions on M by these mappings. We do not go into details on these topics, however, only referring interested readers to [230] and [190]. In order to apply our results above to the asymptotic expansion of e(t, x, x), we take two bounded coordinate neighborhoods U, and U2 of x such that U, c U2, view U2 as a part in Rd and extend the Riemannian metric (g,j(y)) and the vector field b = (b'(y)) outside UZ to a metric g' = (g,j(x)) and a vector field b' = (b"(x)) on Rd respectively, so that
g,j(y) = 8,f and b"(y) = 0 near oo. Let e'(t, x, y) be the fundamental solution of the heat equation
at = 2
d'u + Y(u)
on Rd
where d' is the Laplace-Beltrami operator for g'. Then we have (10.57)
sup I e(t, z, y) - e'(t, z, y) S c, exp[- c2/t]
y,zeU,
as t.0
for some positive constants c, and c2. For the proof, let v r=- C"(Rd) with
support contained in U, and set u(t, z) = fa, [e(t, z, y) - e'(t, z, y)]v(y)%/det g(y)dy,
z E U2.
Then u is a solution of
at = 2 Amu + b(u)
on U2
and
lim u(t, z) =0 do
on U2.
By Ito's formula applied to u(t-s, Xs) where X, = X(s, r, w) with 7L(r) _
z, we deduce that
DIFFUSION PROCESSES ON MANIFOLDS
422
u(t,z)=E[u(t-a,Xo);Q<_ t],
z E U2
where
a=inf{t;XXE8U2}. Hence
u(t, z) I <
(10.58)
sup
z E U2
I u(s, 1;) I ,
sE [o, r7, f e aU2
Noting that e(t, z, y) and e'(t, z, y) have estimates similar to (10.46), in particular, sup
Ceaa2,yG U1
I e(s, , y) - e'(s, , y) I < cl exp[- c2/s],
s E [0, 1]
we obtain from (10.58) that u(t, z) I < c1(exp[- c2t])IIvIIL,, z e U1, t E (0,1] and (10.57) follows from this. Hence, in order to obtain the asymptotic expansion of e(t, x, x) in powers of small t, we can instead take e'(t, x, x). This e' can be given by a generalized Wiener functional expectation as e'(e2, x, x) = E[bx(Xe(l, x, w)]
where X`(t, x, w) is the solution of the following stochastic differential equation on Rd: 1
dX`(t) = kE Qk(X(t))dwk(t) - e2[ Z L J,'gik(X(t))rk(X(t)) + b'(X(t))]dt X(0) = X.
Here, we denote the components of the above extended g' and b' by g,j and b`; rk is the Christoffel symbol for g,1. Also, (g`1) is the inverse of (g,a) and
is the square-root of (g`i), so that g`i _
vkvk. Then,
in the same way as above, we can obtain e(t,
x, x) = t-dJ2(co(x) + tc,(x) +
)
as
t.0
and each c,(x) can be expressed explicitly by a generalized Wiener func-
APPLICATIONS TO HEAT KERNELS
423
tional expectation. Usually, c,(x) is a very complicated polynomial in components of curvature tensor, their covariant derivatives and the vector fields b. By choosing our coordinates to be normal for the metric g, we can compute these expectations to obtain c0(x) = (27r)-l" and
cl(x) = (27r)-d/2( ZR(x)
-
-
div(b)(x) 2
IIbii2(x)) 2
where R(x) = R''(x) is the scalar curvature. However the details are left to the reader. For related topics, cf. [217] and [231].
Finally we apply our method to prove the Gauss-Bonnet-Chern theorem. Let M be a compact oriented manifold of even dimension d
= 21. As Section 4, let A,(M), p = 0, 1,
,
d be the spaces of
differential p-forms and
A(M) _
O+ A,(M). P-0
Consider the heat equation on A(M)
(5.4)
uit-o = a where
= - (d + 8)(d + S) = - (d8 + 8d) is the Laplacian of de
Rham-Kodaira. In Section 5, we solved this initial value problem by a probabilistic way in the form (5.32) where U(t, r) is the scalarization of u(t, x). For given r = (x, e) E O(M), there is a canonical isomorphism r`: i1Rd - ATx (M)
for each
p = 0, 1,
and hence
r: ARd = E ARd -),AT*(M) = P-0
E
AT,*(M),
P-0
defined by
r(Sal A da2 A ... A
aap) =P l Alan A ... fay
,
d,
DIFFUSION PROCESSES ON MANIFOLDS
424
where (fl, f"
, f d } is the ONB in T,*(M) which is dual to the
in TT(M) and ba is defined by ba = (0, 0, 1:'0' , 0), (cf. Section 5). Using this isomorphism, the scalarization F. E C°(O(M) -+ AR d) of a E A(M) can be written as ONB e = {e,, e2,
, ed }
a
Fa(r) = F -' (a(x)),
r = (x, e) E O(M).
Then by (5.32), the solution u(t, x) of (5.4) is expressed as (10.59)
u(t, r) = FE[M(t, r, w) F(t, r, w)-'a(X(t, r, w))]
where r(t, r, w) _ (X(t, r, w), e(t, r, w)) is the stochastic moving frame, i.e. the solution of (5.5). Note that the right-hand side of (10.59) is independent of a particular choice of a frame r = (x, e) E O(M) over x: This just corresponds to the fact that U(t, r) = F-'u(t, x) is 0(d)-equivariant, (cf. Section 5). Hence
u(t, x) = fM e(t, x, y)a(y)m(dy) and the fundamental solution e(t, x, y) E Hom(AT*(M), AT,*(M))
can be expressed by a generalized Wiener functional expectation as (10.60)
e(t, x, y) = E[FM(t, r, w)F(t, r, w)-'b,,(X(t, r, w))].
We now introduce the decomposition A(M) = A+(M) OO A_(M) where
A+(M) =p:even E Q Ap(M) and A_(M) _
QQ A,(M). p:odd
Define a linear operator (- 1)F on A(M) by
(- 1)Fw = 1-0)w
if w E A+(M) if co E A_(M).
Then the operator Q = (d +b) satisfies
Q(- 1)F = - (- 1)FQ,
APPLICATIONS TO HEAT KERNELS
i.e.,
425
{ (- 1)", Q) = (- 1)FQ + Q(- 1)F = 0. In particular, Q sends
A+(M) into A., (M). For .l >- 0, let H,i = {co E A(M); w + )co = 0)
and Hx = { w E A±(M); w + tw = 0 }. Then H,. = Hx Q Hi- and 0< dim Hx < co (cf. [145]). We have (10.61)
dim Ht = dim Hi"
if A > 0,
because the mapping Q: Hx a Hx is an isomorphism. Indeed, if w E HZ, then
Qw = - Q'w =
-.LQw
showing that Qw E Hi-. If co e HZ and Qw = 0, then co = (Q2w)/) = 0 showing that the mapping Q : Hx - Hr is one to one. Finally if w E Hj , then co = Q[(Qco)/.1] and (Qw)/.1 E Hx showing that the mapping Q: Ht -* Hi- is onto. In the same way, we define the decomposition ATx (M) = A+TX (M) Q+ A-T,*(M),
X E M,
and
AR" = A, R" (D A-Rd.
Then (- I)F E End(AT,*(M)) and (- 1)F E End(ARd) are defined similary as above. For A E End(AT,*(M)) or A E End(ARd), define the supertrace Str(A) of A by Str(A) = Tr((- 1)FA). If A E End(AT, (M)) leaves A,T,*(M) invariant, then clearly Str(A) = Tr(A 14+(T"(M)) - Tr(A I .t_(T. (M) )-
Now we claim that (10.62)
fm Str[e(t, x, x)]m(dx) = X(M),
t>0
X(M) being the Euler characteristic of M, (cf. [145]). Indeed, by the eigenfunction expansion of e(t, x, y), e(t, x, y) = E e-Ant n-0
E
c(x) 0 9S(y)
c5:ON8 in HA n
DIFFUSION PROCESSES ON MANIFOLDS
426
for some 0 S 20 < .1, < .. and hence Str[e(t, x, x)]
(10.63)
= -0 E e xnt where II
E #:ONB 3n HZn
II0(x)II2 --0 E
e_xn,
E
110(x)1 2
s:ONB in Hxn
is the Riemannian norm of AT*(M). Hence
II
f Str[e(t, x, x)]m(dx) = E e-xn(dim Hx; - dim Hxn)
-0
M
and, noting (10.61), this is equal to a
dim Ho - dim Ho = E (- 1)pdim HH(M) P-0
co E Ap(M); fjco = 0) is the space of harmonic p-
where HH(M) forms. Since
dim HH(M) = p-th Betti number of M by the de Rham theorem, (cf. [145]) (10.62) is proved. Hence if we can show that (10.64)
Str[e(t, x, x)] = C(x) + o(1)
as t 10 uniformly on M, then X (M) =
f
M
C(x)m(dx).
Now we show that (10.64) actually holds and we identity the limit C(x)
with an explicit polynomial in terms of curvature tensor as given by (10.89) below. This result was first obtained by Patodi [223]: In the same way as above, e(t, x, x) End(AT,*M) has the asymptotic expansion e(t, x, x)
t-d(co(x) + ci(x)t + ... + ce(x)t` + cr+1(x)tr+i + ... )
as t10, where I = d/2 and c,(x) E End(AT,*M), i = 0, 1, , c,(x) being a very complicated polynomial in components of curvature tensor and its covariant derivatives. However, Patodi showed that a remarkable cancellation takes place for the supertrace, namely,
APPLICATIONS TO HEAT KERNELS
427
i=0,1, ,1- 1
Str[c,(x)] = 0,
and furthermore, Str [c,(x)] is a polynomial of curvature tensor only, the terms involving covariant derivatives being completely cancelled. After Patodi, many simpler proofs were invented (cf. Gilkey [202], Getzler [201] etc.). A probabilistic method was first given by Bismut [193]. Our method, like Bismut, consists in obtaining cancellation at the level of functionals before taking expectation which simplifies much the proof of cancellation. In the study of short time asymptotics of e(t, x, x), we may assume as before, by choosing a coordinate neighborhood and extending the components of the metric to whole Euclidean space, that the metric tensor gjj(y) are defined globally on R' and furthermore, we may choose a normal coordinate around x so that x is the origin of Rd and near the origin (10.65)
g1(y) au +
Rtmk1(0)Ymyk + O(1 y 13)
and (10.66)
T7k(Y)
Rrikm(0)Ym -
3
Rik1m(0)Ym + O(I Y I Z)
Let (g'1(y)) and (vj(y)) be the inverse of (g,1(y)) and the square root of (g'1(x)) respectively. Consider the following stochastic differential equation on Rd x GL(d, R) with a parameter a (0, 1): dX'(t) = .6k'k(X(t))dW''(t) - 2 g1k(X(t))rJ (X(t))dt (10.67)
def(t) = - F.VX(t))ei(t) o dXm(t)
i,j=1,2,...,d. (X(0), e(0)) = (0, I)
Denote this solution by re(t) = (X8(t), ee(t)) and their components by X*)' and e8(t)j' respectively. We know that re(t) E O(M) _ {(y, e) (=R" X GL(d, R); g1(y)e' e'p = aQ } for all t 0 a.s. and that { r8(t) } is equivalent in law to { r(ezt, r, w) } where r(t, r, w) is the solution of (5.5). Let H8(t)E=- End(ARd) be defined by (10.68)
178(t): 6' A 3" A ... A 8r° -+ e8(t)" A e(t)'= A ... A ee(t)',
428
DIFFUSION PROCESSES ON MANIFOLDS
where e8(t)' _ (eg(t)', e`(t)i, , ee(t)`a) E Rd, i = 1, 2, fine Mg(t) E End(ARd) by the unique solution of (10.69)
dM dt t)
-
,
d. De-
= .ZM(t)DZ[ 2 J,,km(re(t))]
M(0) = I, (cf.(5.30)) of Section 5 for notations). Then, by (10.60), we can conclude
that e(e2, x, x) = e(e2, 0, 0) (10.70)
= E[M8(l)n8(l)So(X8(l))],
e E (0, 1].
By (10.65), (10.66) and (10.67), we can easily deduce as in Theorem 10.4
that (10.71)
X8(1) = ew(1) + O(e2)
in D-(Rd)
as e j 0
and, by Theorem 10.5, we see that X8(1)/e is uniformly non-degenerate in the sense of (9.32). Hence by Theorem 9.4, we have (10.72)
8o(X8(1)) = e-dSo(w(1)) -I- O(c d+1)
in
D--
as
a J, 0.
Similarly by (10.65), (10.66) and (10.67), we can deduce e8(1)5 = S] + (10.73)
3
[R[m,k(0) + Rr mk(0)] fo wk(s) o dwm(s)
+ O(s3)
in b.
as a J, 0, i, j = 1, 2,
, d.
From (10.67) and (10.68), we see that 178(t) is the unique solution of (10.74)
178(t) = I + f 178(s) o d08(s) o
where ®8(t) is an End(ARd)-valued semimartingale defined by d
(10.75)
08(t) = D1(08(t)) _ E 6,8,(t)a*aj
(cf. (5.23) for notations) where 08(t) is Rd x Rd-valued semimartingale given by
APPLICATIONS TO HEAT KERNELS (10.76)
O(t) = - f
429
Tm,(X8(s)) o dXr(s)m. O
By (10.66) we can deduce (10.77)
in D.
O1,(t) = e2Cj(t) + O(e3)
ass 10,
where
cu(t) =
[Rimjk(0) + R1fmk(0)3f wk(S) 0
3
dwm(s),
i,j=1,2,...,d.
By (10.74), we have
178(1) = I + f
(10.78)
378(s) o de2(s) o
= I +98(1) +
f
f" 0
n8(t2) c d08(2) c dO8(t,)
0
= i +98(1) + f 08(t,) c d98(t,) 0 o
+f
0
f o' f0
178(13) ° d98(t3) o d98(tz) o d98(t,)
and so on. Hence, setting (10.79)
A. = f f ... 0
0
f`m o d98(tm) - d08(tm_,) a ... -deco), 0
it is easy to conclude 178(1) = I + A, +
(10.80)
+ A, + 0(s21+2)
in
D_(End(ARd))
as a 10.
Furthermore, by (10.77), we can deduce for each m = 1, 2,
A. (10.81)
-', dD,[C(tm)l = Elm f f 0' ... f'--'-
that
0 dD,[C(tm-,)l
o
o dD,[C(tl)] + 0(e2m+,)
in
D_(End(ARd))
as a J0.
On the other hand, denoting by P(t) the Rd Qx Rd 3Rd ®x Rd-valued process J,Jkm(r8(t))/2, we see from (10.69) that
DIFFUSION PROCESSES ON MANIFOLDS
430
M8(1) = I + E2f l Me(t)D2[J8(t)]dt 0
= I `f'. elf l Dz[J8(t)]dt 0
+ e4f 'f"0 Me(t2)D2[J8(t2)]D2[`18(tl)]dt2dtl 7
O
and so on. Hence, setting (10.82)
B. =
e2n f
oJ
of ...
f
Dz[Je(tt)]
o
dtndtn_l ... dtl,
it is easy to conclude that (10.83)
M8(1) = I + Bl +
+ Bt + O(e21+2)
D.(End(ARd))
as a 10.
Furthermore, it is easy to deduce of each n = 1, 2,
, that
in
(10.84)
B. = n (D2[J])n + O(e2n+l) i
in
D..(End(ARd))
as a ,1. 0
where J = J8(0) = (Rljkm(0)/2)c- Rd 0 Rd (2) Rd ® Rd.
We now need the following algebraic lemma:
Lemma 10.10. Assume that d = 21 is even. (i) Let al, a2, , am E Rd 0 Rd and bl, b2i
, bn E Rd Q Rd Qx Rd ® Rd. Let A C- End(AR4) be a product of D1[al], Dl[a2], , Dj[an] and D2[bl], . 2[b2], , D2[bn] in some order. Then
(10.85)
Str(A) = 0
(ii) Let b = (10.86)
if m + 2n < d = 21. E Rd 0 Rd 0 Rd 0 Rd and suppose it satisfies
bukm = - bnkm = - brjmk = bkmij bifkm + btkm/ + bimik = 0.
Then
and
APPLICATIONS TO HEAT KERNELS
431
Str((D2([b])1)
(10.87)
=
E sgn(v)sgn(u)bv (1) v (2) p (J) p (2)
1
per'(2l)
21
by (21-1)P(21)p(21-1)p(21)
x by (3)y(4) 4(3)p(4)
where S(21) is the permutation group of order 21. Proof will be given at the end of this section. Remark 10.5. The curvature tensor R1jkm satisfies the relation (10.86).
By (i) of Lemma 10.10, we have
if 2n + m < 21 = d.
If 2n + m = 21 and m > 0 or if 2n + m > 21, then by (10.81) and (10.84), B.A. = O(e2m+2n) = O(e21+1)
as a 10.
in D_(End(ARd))
Hence we obtain by (10.80) and (10.83) that Str[M8(l)178(1)] = e2Str[B1] + 0(e"')
in D.
as a 10.
Therefore, it follows from (10.84), (10.87) and Remark 10.5 that 8
Str [M8(1)178(1)] _
Str[(D2[R1fkm(0)/2])1] +0(S21+1)
e2t
(10.88)
= 2211) v, E (zn
sgn(v)sgn(u)Rv (1) v (2) p (1) p (2) (0)
x Rv(3)v(4)µ(3)µ(4)(0)
+ O(E21+1)
. Rv(21-1)v(21)1U(21-1)p(2l)(0)
in D.
as a 10.
Combining this with (10.70), (10.71) and (10.72) and noting E[ao(w(l))] =
(23)-d/2 =
(27c)-1
we finally obtained the following: Str[e(e2, 0, 0)] = E[Str[M8(l).I78(l)]bo(X8(1))] (10.89)
1
= 231 111
v ,,
zq
sgn.(v)sgn(u)Rv (1) v (2) p (1) /I (2) (0)
432
DIFFUSION PROCESSES ON MANIFOLDS X Rv(3)v(4)IL (3)k(4)(0) ... Rv(21-I)v(21)k(21-1)A(21)(0)
+0(6) as e 10. It is clear that 0(e) in (10.89) can be estimated uniformly with respect to
x E M which we take to be the origin of the normal coordinate. Generally, denoting by J,fkm(r) the scalarization of the curvature tensor Rljkm(x) as before, define a function C(r) on O(M) by C(r) = 231ir111 k (2n sgn(v)sgn(p)Jv (I) v (2) P (I) ,u (2) (r) (10.90)
X Jy(3)v(4)p(3),(4)(r) ... Jv(2l-1)v(21)µ(21-1)µ(21)(r), r E O(M).
It is easy to see that C(ra) = C(r)
for every a e 0(d)
and hence C(r) depends only on 7t(r) = x E O(M). Thus we may write C(r) by C(n(r)). This function C(x) on M is known as the Chern polynomial. We have now established (10.64) and therefore X (Al) = J
M
C(x)m(dx).
This formula is known as the Gauss-Bonnet-Chern theorem. In the case
d = 2, C(x) = K(x)/2n where K(x) __
RI232(x)
det(glj(x))
is the Gauss total curvature and hence
1 f K(x)m(dx) = X (M) = 2(1 - g) 27c m
where g is the genus of the surface M. This is the classical Gauss-Bonnet theorem.
Proof of Lemma 10.10. A direct proof may be found in Patodi [223]. Following [194], we now give a proof based on a unifying idea of supersymmetry. Let r1Rd = E Q ARd be the exterior algebra over R4 and a*, al E End(AR4) P-0
APPLICATIONS TO HEAT KERNELS
433
be defined by (5.19) and its dual. Let ACRd be the complexification of ARd and let y, E End(A,Rd), i = 1, 2, , 2d, be defined by Yzt-i = at + ar
i= 1,2,
d.
Yzt = 1=11 (a* - at) Then from (5.20) we have { y,, yy } = 2S,f I and y* = yt. For every subset K = {u1 ,,t2i < luk of { 1, 2, , 2d} let ; Ik}, A, <,Uz < )k(k-1)12Yt<1Yuz ... yIIk
Ys=
and
Then y = I and y,*K = yr. Also, it is easy to see that Tr(y4) =
0
if A * 0
2d
ifA=0.
Indeed if A = 0, let u, = min{ i I i e A } and let A = A\{,u, }. Then,
if #A = k = even, Tr(Ys) _
)k-'Tr(Y,1,
_-
YA)
1)k-'Tr(YA
= - -1)k'Tr(Yk,
Yr,,) VA)
Tr(ye)
and hence, Tr(y4) = 0. If #A = odd, choose u E4 A and write Tr(ya) = Tr(Yr4 ' Yt4Y.4) = - Tr(Y,.
= - Tr(Y,. ' Y.t
Y" - Y"')
Y;') = - Tr(YA)
Hence Tr(y,,) = 0. It is easy to see that the system { y}, where K is a subset of { 1, 2, 2d }, forms a basis of End(A,Rd) : Independence of this system is easily seen from
Tr(y1y ,)
-
10 2d
if K* K' if K= K'.
Also, dim End(ACRd) _ (2d)2 = 2 zd = the number of the system {y1}
DIFFUSION PROCESSES ON MANIFOLDS
434
and the assertion follows. Thus every A E EndA(Rd) is expressed uniquely as (10.91)
A = E CS(A)ya
C1(A) E C
K
and (10.92)
Tr(A) = 2dCO(A).
(10.92) is known as the Berezin formula. Next we claim that (10.93)
(- 1)F = (- l)dyn.2.
,2dt
Indeed, if we denote the right-hand side by a, then { y,,, a) = 0 and hence
, d. From this, it is easy to conclude a = [at*, a) = 0, i = 1, 2, (- 1)F if we can show that a w = co for w e Ac(Rd) = C. But 1(akak - akak)
72k-I72k = %
and hence if co E C.
72k-I72k(t)
Thus acv = (- 1)d(%/
1)2d(2d-1) /2(
1)dco = (- 1)d(%/
)2dZco
= co.
Hence, combining this with (10.92), we have (10.94)
Str(A) = Tr((- l)FA) = (- 1)d2dC,1. 2, ...,2d)(A).
Now the proof of (i) is easy: If we express A in the form (10.91), then, in each term, y's appear at most 2m + 4n times. Hence if m + 2n < d, C41, 2...., 2,n (A) = 0.
For the proof of (ii), we first note that d
1 2N = 3L 2 brjkmY2t72JY2k-1Y2m-i (10.95)
+ a polynomial in y's of degree 2
APPLICATIONS TO HEAT KERNELS
435
if (btJkm) satisfies (10.86). Indeed, omitting the summation sign for repeated indices, we have 132[b] = biJkma$aJak am
= 24 Y,
1)`btlkmYt5JYkYm
where Yt = yet or Y21-, and the exponent a depends only on the way of this choice. Noting the well-known property of (btUkm) satisfying (10.86) that the alternation over any three indices vanishes, it is easy to deduce that b,jkma,*ajak am _ 4 { btJkmY2tY2JY2k-IY2m-1 + b1kmY21-1YZJ-1Y2kY2m bt)kmY21YZJ-1Y2k-1Y2. + bifkmY21-1Y2JY2kY2m-1
- btJkmYztY2J-1Y2kY2m-1 - bjjkmY21-1Y2JY2k-1Y2.)
+ a polynomial in y's of degree 2. It is easy to deduce from (10.86) that the first two terms are equal and the remaining four terms cancel. Hence the proof of (10.95) is finished. From this and (10.94), we can easily conclude (10.87).
Remark 10.5. If we take another supertrace Str 12) (A) = Tr[y (2, 4,
, u, A }
and do the same calculation as above, we can obtain the Hirzebruch signature theorem. Main difference in this case is that A. in (10.83) does not disappear in the process of cancellation, that is, an effect of the paral-
lel translation operator 17e(l) remains in the final form of signature theorem. For details, cf. [204]. Indeed, we can give a similar probabilistic proof of Atiyah-Singer index theorem for every classical complex, cf. [193] and [224].
CHAPTER VI
Theorems on Comparison and Approximation and their Applications
1. A comparison theorem for one-dimensional Ito processes Suppose now that we are given the following: (i) a strictly increasing function defined on (0, co) such that p(O) = 0 and (1.1) J
a+P(S)-Z a' = no;
(ii) a real continuous function 6(t, x) defined on (0, oo) x R' such that (1.2)
1 cr(t,x) - c(t,y) I < P(I x - y I),
x, y e k, t > 0;
(iii) two real continuous functions b1(t, x) and b2(t, x) defined on [0, oo) x R' such that (1.3)
b1(t, x) < b2(t, x)
Let (Q,.Y,P)
t > 0, x E R'.
be a probability space with a reference family (. )tz0.
Theorem 1.1. * Suppose that we are given the following stochastic processes: (1) two real (Y )-adapted continuous processes x1(t,c)) and x2(t,c));
(2) a one-dimensional (,)-Brownian motion
B(t,cv)
such that
B(0) = 0, a.s.;
* This theorem covers the previous results obtained by e.g. Skorohod [150], Yamada [182] and Malliavin [108]. 437
COMPARISON AND APPROXIMATION THEOREMS
438
(3) two real (-O';)-adapted well-measurable processes /f,(t,w) and ,62(40We assume that they satisfy the following conditions with probability one: (1.4)
x,(t)- xx(0) =
(1.5)
x,(0) < x2(0),
(1.6)
fi,(t) < b,(t,x,(t))
for every t
(1.7)
fl2(t) > b2(t,x2(t ))
for every
Ju
Q(s,x,(s))dB(s) + J'
b`(s)ds,
1= 1, 2,
0,
t > 0.
Then, with probability one, we have
xl(t) < x2(t)
(1.8)
for every t > 0.
Furthermore, if the pathwise uniqueness of solutions holds for at least one of the following stochastic differential equations
dX(t) = Q(t,X(t))dB(t) + b,(t,X(t))dt,
(1.9)
i = 1, 2,
then the same conclusion (1.8) holds under the weakened condition (1.3)'
b,(t,x) < b2(t,x)
for t > 0, x E R'.
Proof. * By a usual localization argument we may assume that o(t,x) and b,(t,x) are bounded. Step 1. We assume that b,(t,x) is Lipschitz continuous, i.e., there exists a constant K > 0 such that
I bi(t,x) - bi(t,y) I< K I x - y 1,
x, y E R'.
Choose a sequence V. (u), n = 1, 2, ... of continuous functions as in the proof of Theorem IV-3.2 and set
7n(x) =
0, fx
x<0, fY
f. dy J o
* The following proof is due to T. Shiga [146].
x > 0.
COMPARISON THEOREM FOR ITO PROCESSES
439
It is easy to see that 0. E C2(R'), 0.(x) = 0 for x < 0, 0 < 0'(X) < 1 and 0,(x) t x+ as n - co. An application of Ito's formula yields 0n(x1(t)-x2(t)) = 11(n) + 12(n) + 13(n),
where r1(n) = f 0n(x1(s)-x2(s)) {o(S,x1(S)) - Q(S,x2(s))}dB(s), 0
12(n)
=f
0
0n(xl(S)-x2(S)) {MS) - kS)}ds
and
= 2 f", 0, (x1(S)-x2(s)) {a(S,x1(s)) - 0,(s,x2(S))} 2ds.
13(n)
It is clear that E(11(n)) = 0 and E(I3(n)) < 2 E(f o cn'(x1(s)-x2(s))p(I x1(S)-x2(S)1)2ds) 5 n . Also, 12(H)
f 0'(x1(s)-xz(s)) {b1(s,x1(s)) -b2(s,x2(s))} ds o
= f' 0n(x1(s)-x2(s)) {b1(s,x1(s)) - b1(s,x2(s))} ds + f' O (xl(s)-x2(s)) {bl(s,x2(s)) - b2(s,x2(s))} ds :!9f
{b1(S,x1(S))-b1(s,x2(s))} ds o
01[.1cs»Y2c,»
Ix1(s) - x2(s)Ids
< K f o (x1(s) - x2(s))+ds.
Hence, by letting n -- oo, we have E[(x1(t) - x2(t ))+] < KE[ f 0 (x1(s) - x2(S))+ds]
= K f r0 E((x1(s) - x2(s))+)ds.
440
COMPARISON AND APPROXIMATION THEOREMS
We can deduce from this that
E[(x,(t) - x2(t))+] = 0
for all t > 0,
that is P(x,(t) < x2(t)) = 1
for all t > 0,
and so by the continuity of paths we conclude that (1.8) holds.
If instead we assume that b2(t,x) is Lipschitz continuous, then a similar argument leads to the same conclusion (1.8).
Step 2. In the general case we choose b(t,x) such that
b,(t,x) < b(t,x) < b2(t,x) and b(t,x) is Lipschitz continuous. Let X(t) be the unique solution of the following stochastic differential equation dX(t)
o(t,X(t))dB(t) + b(t,X(t))dt
X(0) = x2(0)
Then the result from step 1 gives that X(t) S x2(t) and x,(t) S X(t) for all t > 0 a.s. Consequently we can conclude that (1.8) holds. Step 3. We now turn to the proof of the second assertion. We assume that the pathwise uniqueness of solutions holds for one of the stochastic differential equations (1.9), say for i = 1. Let X(t) be the solution of the equation (1.10)
dX(t) = c(t,X(t))dB(t) + b,(t,X(t))dt
X(0) = x0)
For s > 0, let X(tg)(t) be the solutions of dX(t) = a(t,X(t))dB(t) + (b,(t,X(t)) ± s)dt
X(0) = x0) respectively. By what we have already proven,
V-4)(t) < X(t) < X(+B'(t)
for every t > 0,
APPLICATION TO OPTIMAL CONTROL PROBLEM
441
and if 0 < e2 < e1i then X(-ej)(t) < X(-e2)(t) and X`+e2)(t) < X`+aj1(t)
for every
t > 0.
Hence by the continuity of bl(t,x) and the pathwise uniqueness of solutions for (1.10), we have
lim X (-i)(t) =e1Olim X(t) = X(t) CIO
for every
t > 0.
Again applying what we have already proven to x1(t) and X'+E1(t), we have (1.11)
x,(t) S X (-')(t)
since fl1(t) < b1(t,x1(t)) ting s 10 in (1.11), (1.12)
x,(t) < X(t),
for every
a.s. and bi(t,x) < b1(t,x) + e. Hence, by letfor every t > 0.
2(t) > b2(t,x2(t)) a.s. and X'-e1(t) < x2(t); letting e j 0 gives Since
X(t) < x2(t)
t>0
for every
b2(t,x) > b1(t,x) - s, we have
t > 0.
Combining this with (1.12) we obtain the inequality
x1(t) < X(t) < x2(t),
for every t > 0,
which completes the proof of the second assertion.
2. An application to an optimal control problem As an example of an application of the comparison theorem in the previous section we shall consider the following stochastic optimization problem.
Let k(r) be a non-decreasing and non-negative function defined on [0, co). Let (Be, u) be a system of stochastic processes defined on a probability space (Q,- 9-,P) with a reference family (,') such that (i) B1 (Bo = 0) is a d-dimensional (.57;)-Brownian motion and (ii) u, is a d-dimensional (9)-well measurable process such that d u, I< 1 for all t> 0 a.s.
442
COMPARISON AND APPROXIMATION THEOREMS
Such a system (B,, u,) is called an admissible system or an admissible control. Let XE Rd be given and fixed. For a given admissible system, the response XX is defined by (2.1)
Xf = x + B, + E u,ds. o
The optimization problem then is to minimize the expectation E(k(I X,1)) among all possible admissible systems. The solution is given as follows. Let U(y) be defined by (2.2)
= o,Y11Y1,
y E 0 d\ E {01
U(Y)
a
Consider the following stochastic differential equation (2.3)
dX, = dB, + U(X,)dt Xo = X.
By the corollary to Theorem IV-4.2 we know that a solution (X,,, B,) exists uniquely. Set (2.4)
of = U(X,).
Then the admissible system (BOgo) gives an optimal control; that is, for any admissible system (B,,u,), we have (2.5)
E(k(I X, I )) <_ E(k(1 1 i I ))
We shall now prove this fact as a simple consequence of Theorem I.I. It was originally obtained by Benes [3] by different techniques.
Theorem 2.1. Let (B,,u,) be any given admissible system and, for a fixed x E Rd, let the response {Xf} be defined by (2.1). Then on an appropriate probability space we can construct Rd-valued processes 1±3 and 1±01 such that (i)
{XIM}
{'j},
(ii) {X°} and
(iii) 1 xf 1 < 1 'f 1
for every t > 0, a.s.
APPLICATION TO OPTIMAL CONTROL PROBLEM
443
Corollary. Let Wd = C((0, co) - Rd) and F(w) be a non-negative Borel function on Wd with the following property: if w,,w2 E Wd and
(2
F(wl) <
w2(t) I
I w1(t)
for every t Z 0, then
F(w2).
Then for any admissible system (Be, u,) we have
E(F(X`)).
(2.7)
That is, the solution IX,*) of (2.3) is optimal in the sense of minimizing the expectation of F(X').
Note that the particular case of F(w) = k(I w(l) 1) clearly satisfies (2.6).
In order to prove the theorem we first state the following lemma. Lemma 2.1. Let (X1, Bt) be a pair of d-dimensional continuous (-g;)adapted processes defined on a probability space (0,.Y1P) with a reference family (.J9'-,' such that {B,} is a d-dimensional Brownian motion with
B0 = 0. Let (Y,,B') be a similar pair defined on another space (S2',,`r'P') with (. '). Then we can construct a probability space (52,F P) with a reference family (,,) and a triple of d-dimensional (JP,')adapted processes such that (i) (X, B:) e (ii) (Y:, B'D
and
(iii) (B,) is a d-dimensional (.,9)-Brownian motion. The proof follows in exactly the same way as in Theorem IV-1.1.
Proof of the theorem. Let be a given admissible system and X, = x+B,+ f o u,ds be the response. Let (XO,B,) be a solution of (2.3). Choose an 0(d)-valued Borel function (p,j(x)) such that (2.8)
xJ/IxJ
x= (x',x2, ... ,x') k- 0
45,J
x=0.
p,j(x) =
Set Bt = f' p(X,)dB, and ff = f' p(XS)dB:. Then we have :
(2.9)
:
Xt" = x + f 0 P-'(Xs)dB, + f 0 u,ds
COMPARISON AND APPROXIMATION THEOREMS
444
and t
(2.10)
X: = x+ f
p-'(X s)dB s+ f' U(X o,)ds.
0
0
Now we apply Lemma 2.1 to (X",Br) and (X0,.§,0). We then have a triple
(t- Xo, Bt) of (.)-adapted processes on a probability space with a reference family (..) such that {Bt} is a d-dimensional (. )-Brownian motion and (Xu,. ) i' (X',B') Clearly there exists an (..)-well measurable d-dimensional process {12r} such that 19, I < 1 for
every t > 0 and =x+ft0p-'(.ftdBs-{-J0t tds. Applying Ito's formula to x1(t) = I to 12 and x2(t) = I X! 2, we have
dx2(t) = 2±j p-'(X jdBt + 21'. a dt + ddt = 211 7 I dB,' + [2 . U', + d]dt =2,/x,-(t)
dB t + [2t" ut + d] dt
and
dxl(t) = 2I .p-'(X°)dBt + 2A3o U(Xa,)dt + d dt
= 2I k I d d f + [-2I t I + d]dt dd.' + [-2 xl(t) + d]dt,
= 2/)
where st = (Bt, Bj, ... , Ba). (Note that Set a(t,x) = 2 x,p11(x) = ate 1 x I).
E -2 ,lx- O+d, fl1(t) = -2 xl(t)+d and
'(x)]1= x V 0,
xj(p-'(x));t =
b1(t,x) = b2(t,x) =
/32(t) = 2 ' tt+d. Then clearly fl1(t) = b1(t, x1(t)) and fi2(t) > -2 1 Xi I +d = b2(t, x2(t)). As we shall see in the next lemma, the pathwise uniqueness of solutions holds for the stochastic differential equation (2.11)
dX(t) = a(t, X(t))dB(t) + b1(t, X(t))dt = 2(X(t) V 0)112dB(t) + [-2(X(t) V 0)''2 + d]dt.
Therefore we can apply the second assertion of Theorem 2.1 and obtain the result that x1(t) < x2(t) for all t > 0, a.s., that is I Ito I < I X'j I for
all t > 0, a.s.
APPLICATION TO OPTIMAL CONTROL PROBLEM
445
Lemma 2.2. The pathwise uniqueness of solutions holds for the equation (2.11).
Proof. First we remark that the maximal and minimal solutions of (2.11) exist; that is, there exist strong solutions X,=F,(X1(0),B) and X2=F2(X2(0), B) of (2.11) such that if (X(t), B(t)) is any solution of (2.11)
and X1(0)=X(0)=X2(0), then X1(t) < X(t) < X2(t) for every t > 0, a.s. such that Indeed, choose smooth functions b( ,"(x) and
< ba'(x) < bi2)(x)
and
lim b;,"(x) = lim b,1,2'(x) = b(x).
Here b(x) = b,(t. x) = -2(x V 0)' 12+d. Then the solutions XX(t) of dX(t) = a(t,X(t))dB(t) + b,',"(X(t))dt { X(O) = x
satisfy that X (t) < n = 1, 2, ... , by Theorem 1.1 and we set F1(x, B) = lim X.(.). F2(x, B) is similarly defined. Clearly they possess the above property. Set
s(x) = f
1
exp I - f2z' dil dy, ,
x > 0.
Then s(0+) > - oo if d = 1 and s(0+)
oo
if d > 2. First we
consider the case d = 1. By Ito's formula s(X(t)) - s(X(0)) =
ft
s'(X(s))a(s,X(s))dB(s)
for any solution X(t) of (2.11) with X(0) > 0. Hence we have E(s(X(t))) = E(s(X(0))).
From this we can conclude that X1(t) = X2(t) a.s. for all t > 0 if X1(0) = X2(0) a.s. This implies the pathwise uniqueness of solutions for (2.11). If d > 2, we see by Theorem 3.1 given below that inf X(t) > 0 a.s. for every 05cST
COMPARISON AND APPROXIMATION THEOREMS
446
T > 0 if X(0) > 0 a.s. The uniqueness of solutions for (2.11) can be easily deduced from this; we leave the details to the reader.
The optimization problem discussed above is a special example of stochastic control problems. For the recent development of the stochastic control theory, we refer the reader to Krylov [91].*1 3. Some results on one-dimensional diffusion processes *2
Let I = (1, r) be an open interval in R1 (-oo < 1 < r < oo). Let 6(x) and b(x) be sufficiently smooth real valued functions*3 on I such that 62(x) > 0 for all x e I. For x E I, the stochastic differential equation
dX(t) = a(X(t))dB(t) + b(X(t))dt { X(0) = x
(3.1)
has a unique solution X'(t) up to the explosion time e= lim Tn, where nt""
Tn = inf {t; Xx(t)
[an, bn]} (an and b (n = 1, 2, ...) are chosen such that
I < an < bn < r and an j 1 and bn t r). By the same proof as in Lemma IV-2.1, we can show that limXx(t) exists and is equal to I or r a.s. on tie
the set {e < oo}. We define Xx(t) to be this limit for t > e on the set [I, r] {e < oo}. Let JP, be the set of all continuous paths w: [0, oo) such that w(0) E I and w(t) = w(e(w)) for all t > e(w): = inf {t; w(t) = I or w(t) = r} . e(w) is called the explosion time of the path w. Then XX = {X"(t)} defines a W,-valued random variable with e[Xi] = e. Let PX be the probability law on 41, of P. As we saw in Chapter IV, {Px} defines a diffusion process on I. It is called the minimal L-diffusion, where the differential operator L is defined by 2
(3.2)
L =
+ b(x).
62(x)
2
dxd 2
Let c e I be fixed and set (3.3)
s(x) = f c exp {- fy
y 62() dz} dy.
In particular, this book contains important results by Krylov on the estimates for the
distributions of stochastic integrals, an extension of Ito's formula and stochastic differential equations with measurable coefficients. *Z Cf. [731 for more complete information. #3 In the following discussion it is sufficient to assume that v and b are of class C'.
RESULTS ON ONE-DIMENSIONAL DIFFUSION PROCESSES
447
Then s(x) is a strictly increasing smooth function on I satisfying
Ls(x) = 0
(3.4)
on
1.
oo and s(r-) = co, then
Theorem 3.1. (1) If s(1+)
Px(e = oo) = Px(lim X(t) = r) = Px(lim X(t) = 1) = I
(3.5)
ft-
7T__
for every x.* In particular, the process is recurrent, i.e. Px(a,, < oo) = 1 for every x, y E I where ay = inf {t; X(t) = y} . (2) If s(1+) > -oo and s(r-) = oo, then lim X(t) exists a.s. (P.) and tte
Px(1im X(t) = 1) = Px(sup X(t) < r) = 1
(3.6)
tte
t<e
for every x. A similar assertion holds if the roles of 1 and r are interchanged.
(3) If s(1+) > - oo and s(r-) < oo, then lim X(t) exists a.s. and tle
(3.7)
Px(1im X(t) = 1) = 1 - Px(lim X(t) = r) = tte
tte
s(r-) -s(x) s(r-)
Thus the process is not recurrent in cases (2) and (3).
Proof: Let 1 < a < x < b < r and z = inf{t; Xt Er; [a, b]}. By Ito's formula and (3.4), s(Xx(tA r))
tAT
- s(x) = f0o
st (XX(s))a(Xx(s))dB(s)
and hence EE[s(X(t AT))] = s(x).
By letting t t oo, we have
EE[s(X(T))] = s(a)Px(X(r) = a) + s(b)Px(X(T) = b) = s(x). Combining this with
1 = Px(X(T) = a) + P,(X(T) = b) * X(t) = X(t, w) = w(t), w e a',. Also e = e(w).
COMPARISON AND APPROXIMATION THEOREMS
448
we have
PS(X(z) = a) =
(3.8)
s(b) - s(x) s(b) - s(a)'
PP(X(T) = b) = s(x) - s(a)
s(b) - s(a)
Suppose s(r-) = -s(1+) = oo. Then lim PX(X(z) = b) = 1. all
Hence
PP(sup X(t) > b) > lim P,,(X(a) = b) = I all
r<e
for every b < r and consequently P.,(sup X(t) = r) = 1. r<e
Similarly
P=(inf X(t) = 1) = 1. Now it is easy to conclude the assertion of (1). Next, t<e
assume that s(l+) > -oo and s(r-) = oo. We can deduce as above that (3.9)
PP(inf X(t) = 1) = 1. Ke
Now Y,-' = s(X(t A Sr)) - s(1+) is a non-negative martingale and, letting
a j 1 and b f r, we can easily see that Y, = s(X(tAe))-s(1+) is a nonnegative supermartingale. Consequently lim Y, = lim s(X,) - s(1+) exists :T"
tle
as a finite limit a.s. (Theorem 1-6.4). Therefore lim X, exists a.s. and in rte
view of (3.9) we must have (3.6). The proof of (3) is similar and thus omitted.
Let 1 < c < r be fixed and set (3.10)
lc(x) = 2
exp c
f 2b(z) dz IT C exp [fcn 2 b(z) dz1 d J dc, Q(z) 16(11) c0 2(Zj tfc xE=_ I.
Lemma 3.1. Let u(x) be the unique solution of
Lu(x) = u(x) (3.11)
u(c) = 1 u'(c) = 0.
RESULTS ON ONE-DIMENSIONAL DIFFUSION PROCESSES
449
Then (3.12)
x E 1.
1 + K(x) < u(x) < exp {K(x)},
Proof. u(x) is the solution of (3.11) if and only if it satisfies u(x) = 1+ f X ds(y) f Y u(z)dm(z), C
0
where
ds(y) = exp [_ f y 2"zj dzl dy
.z
I
l
and
&n(z) = 2 exp [f c Q() d i]
az) dz.
Hence u(x) = E uk(x) k.-0
u (x) > 0
with u0(x) - I and un(x) = f x ds(y) Jy and u,(x) = K(x). Consequently
u(x) > I + u,(x) = I + K(x). On the other hand, if we suppose that K(z)n/n!, then
f : ds(y) f u,,(z)&n(z) Y
n l s: ds(y) f K(z)°dm(z)
n f ' ds(y)K(y)" f: dm(z) = f : K(y)nd[K(Y)) t
t
K(x)n+1 (n + 1)!,
Consequently
450
COMPARISON AND APPROXIMATION THEOREMS x()n u(x) _ n=0
E=0
n!
= exp{x(x)}.
Remark 3.1. It is obvious that (i) x(r--) < oo s(r-) < o0 and
(ii) x(1+) < 00 = s(I+) > - 00. Theorem 3.2. (1) If x(r-) = x(1+) = oo, then (3.13) (2)
(3.14)
Px(e = oo) = 1
for all x E I.
If rc(r-) < oo or x(l+) < oo, then Px(e < oo) > 0
for all x E I.
Px(e
for all xEI
(3)
if and only if one of the following cases occurs (i) x(r-) < oo and x(1+) < 00. (ii) x(r-) < oo and s(1+) = -oo. (iii) x(1+) < oo and s(r-) = oo.
Proof. Let u(x) be defined by (3.11). Let 1 < a < x < b < r and z = inf {t; X(t)
(a, b)}. By Ito's formula
de tu(X(t)) = e-tu'(X(t))a(X(t))dB(t) + e-t(-u(X(t)) + (Lu)(X(t)))dt = e-tu'(X(t))c (X(t))dB(t) and hence a t^tu(X(t A z)) is a martingale. Letting a I I and b t r we immediately see that exp (-t A e)u(X(t A e)) is a non-negative supermartingale. If x(r-) = x(1-}-) = oo, then by Lemma 3.1, lim u(x) _ xII
lim u(x) = oo. Consequently Px(e < oo) > 0 is impossible because t xtr
exp(-(t A e))u(X(t A e)) is bounded a.s. as a non-negative supermartingale. This proves (1).
Next suppose that x(r-) < oo. Without loss of generality, we may assume that c < x. Let a' = inf {t; X(t) = c). By Lemma 3.1, u(r) < oo and hence (3.15)
exp (-tAr')u(X(tAr')) is a bounded Px martingale.
RESULTS ON ONE-DIMENSIONAL DIFFUSION PROCESSES
451
Consequently (3.16)
u(x) = EE[exp (- t A r')u(X(t A T'))].
Letting t t oo in (3.16), we have
u(x) = EE[exp (-e)u(r-); lim X(t) = r] rieA,'
4- EE[exp (-T')u(c); lim X(t) = c]. t t eAr'
If EE[exp(-e); lim X(t) = r] = 0 then we have t i ehr
lim X(t) = c] < u(c).
u(x) = u(c)EE[exp
r t ens'
This is clearly a contradiction. Therefore E.,[exp(-e); lim X(t) = r] > 0 t i cATI
which implies that Px(e < co) > 0. Finally we shall prove (3). If K(r-) < oo and if none of (i), (ii) or (iii) hold then we must have s(1+) > - oo and x(1+) = co. Then Px(lim X(t) _ tie
1) > 0. Since exp(-t A e)u(X(t A e)) is a non-negative supermartingale and lim u(x) = oo, we must have e = oo a.s. on the set {lim X(t) = 11. xlr
tie
Consequently Px(e = oo) > 0. Therefore if Px(e < oo) = 1, then at least one of the conditions (i), (ii) and (iii) must hold. It is now sufficient to show that each of conditions (i), (ii) and (iii) implies that Px(e < oo) = 1 for all x. First we shall assume (i). Define G(x, y), x, y E I, by
G(x, y) =
(s(x) - s(1-))(s(r-) - s(y)) s(r-) - s(1+) (s(y) - s(1-))(s(r-) - s(x))
s(r-) - s(1+)
x
y<x.
If f(x) is a bounded continuous function on I, then the condition (i) implies
that u(x) = f G(x, y)f(y)m(dy) is a bounded function. In particular
ul(x) = f G(x, y)m(dy)
COMPARISON AND APPROXIMATION THEOREMS
452
is a bounded function. It is easy to prove that u E C2(1), u(1+) = u(r-)
0 and Lu = -f. In particular, Lu, _ -1. Hence by Ito's formula, u,(X (t A e)) - u, (x) + t A e = f
ui(X(s))dB(s). o
Consequently
-E,(u,(X(tAe))) + u,(x) = E,(t/\e).
Letting t t oo yields u,(x) = E,(e) < oo. This proves that P,(e < 00) _
1. Next we shall assume
(ii).
For each n = 1, 2, ... , set 8n =
inf {t; X(t) = 1-x- 1/n} and a, = inf {t;X(t) = r). Then lim U. A o, = e. n,m
By the result of case (i), P,(5, A a, < 00) = 1 for all x. By Theorem 3.1,
1im P,(6 > a,) = 1 since s(r-) < eo and s(1+) = -co. Clearly {a, < n,m
oo} = U {a, < 8n} and hence P,(6, < oo) = 1. Consequently n
P,(e
4. Comparison theorem for one-dimensional projection of diffusion processes
Let a _ (aL(x)) be a sufficiently smooth function: Rd E) x i --- a(x) ERd(x Rd and b = (bt(x)) be a sufficiently smooth function: Rd E) x _ b(x)ERd. We consider a diffusion process X = (X(t)) on Rd defined by the solutions of the following stochastic differential equation (4.1)
dXf =
d
k-,
uu(X,)dBk + b'(X,)dt,
i= 1, 2, ... , d.
The diffusion X is defined up to the explosion time e (cf. Chapter IV, Section 2). As explained in Section 6 of Chapter IV, this is the diffusion process generated by the differential operator a2
(4.2)
L=
2 i.E,
al(x)
aaxe +
b`(x) ax`
COMPARISON THEOREM FOR DIFFUSION PROCESSES
453
where atf
d
(x) = E Qk(x)Qklx). k-l
Let p(x) be a smooth real function on Rd and let I = { = p(x); x e R,11. Then I is an interval in R'. Let S be the set (possibly empty) of all x ERd such that p(x) is an end point of I. Let I° be the maximal open interval contained in I and I be the minimal closed interval in [-cc, cc] which contains I. We assume that 12
PP(x) I = (t,±i a"(x) a (x) ax (x))
>0
for all x E Rd\S.
Set
(4.3)
a(x) = rE) a`1(x) a (x) ax (x),
(4.4)
b(x) = (Lp)(x)/a(x),
x E Rd\S
and
a+(c) = sup a(x), xe D(G;D)
(4.5)
b+(0 = sup b(x), xeD(I;D)
a-() = xED inf((;y)a(x) b-(c) = inf b(x), xED(4OD)
where D(g;p) = {x; p(x) = c} for E I°. We assume that at(e) and bt() are locally Lipschitz continuous functions on I° and at(e) > 0. On the interval I° we consider the following four minimal diffusion
processes tt =
which are generated by the operators L±± respec-
tively; 1 d2
L++ = a+() (2 dr2 + b+() d )' 2
L+- =
(2 de2 +
b-(0
(4.6)
L-+ = a-() (2 de2 + b+(0 2
L-- =
(2 de2 +
b-(0
d
d
454
COMPARISON AND APPROXIMATION THEOREMS
Each of the diffusions is given as in Section 3; thus, the sample paths **(t) are defined for all t > 0 and are !-valued continuous paths with
1\I° as traps. Let X, be a sample path of the above diffusion (4.1) starting at x0 e
Rd\S and set C = inf {t; t < e, Xr (z S). We assume that if C < oo then lim p(XS) exists in I. We set p(XX) = Sit
lim p(XS) for t > C. Thus p(X,) is defined for all t >_ 0 as an 1-valued Six
continuous path.
Theorem 4.1 ([54]). Let x° E Rd\S be fixed and S0 = p(x0) E 10. Let X = (Xr) be the above diffusion starting at x0. Then we can construct 1valued continuous stochastic processes rjr, rtf +, i7t+-, rtf +, ilt - on the same
probability space such that the following properties (i), (ii) and (iii) hold. (i) (,7r) has the same law as (p(XX)).
(ii) (r7t*) has the same law as** _ (**(t)) starting at 0 for each of four combinations of ± ±. (iii) If we set l r = max i1,
ii, -tir = min 0;!9j-'V
q±± =maxr;*
r7* =minr7;*, -t SsSr
ossst
and
o5s$r
then with probability one we have
It-<1,
(4.7)
for all t>0
and for all
(4.8)
t>0.
Proof. For simplicity we assume that C = oo a.s. and that are all conservative diffusion processes on I°; the general case can be proved with a slight modification. Set
+(t)
=
Jo
[a(XS)/a+(p(X:)))ds
and
0-(t) = $o [a(X)Ia-(p(X))]ds.
COMPARISON THEOREM FOR DIFFUSION PROCESSES
455
Then clearly q+(t) < t < O-(t)
for all
t > 0.
Let +(t) and e (t) be the inverse functions of t - 0+(t) and t 1---- 0-(t) respectively and set Xr+) = X()V+(t))
and
X;-) = X(Vl (t)).
As in Chapter IV, Section 4, we see that Xt+) = (X=+)',X(+'z , ... , X1+)°') and XI-) = (Xt-)1,X;-)z, ... X,(-)d) satisfy the following stochastic differential equations with appropriate d-dimensional Brownian motions B,+) _ (Bt+)',B,+)z, B,+)d) with Bo+) = 0 and B,-) _ (B,(-) 1,B,-)Z, ... ,BB-)d) with Bo) = 0 respectively: a
dX,+)r =
(a+(P(X,1+)))la(X,(+)))'1z E ak(X,+))dB,+)k k-1
(4.9)
+ Xo+) = xo, d
dXj-)'
= (a-(P(X,(-)))la(XI-))))rz E ak(X,-))dB(-)k k-1
(4.10)
+
i = 1,2, ... , d
Xo-) = xo.
By Ito's formula, dP(Xr+)) =
(a+(P(X,(+)))la(X,(+)))1iz
(4.11)
[a+(P(X,'+)))la(X,+))] (LP)(X,(+))dt
P(X o+)) = o and
ai(X,-))axt(x,(-))dBt-)'
dp(X,(-)) = (4.12)
+ [a-(P(X,(-)))la(X,(-))](LP)(X,(-))dt
P(Xi-)) = o Hence, if we set
456
COMPARISON AND APPROXIMATION THEOREMS
B° = t.Z l J 0
(l/a(X=t)))',2Qf(X(±))
LP
(X=t))aBst);
a then (B, ) and (BT) are 1-dimensional Brownian motions and we have (4.13)
(a+(p(X,(+')))';zdB
dp(XX+') =
+ a+(p(X (+)))b(X'+')dt
p(Xoc+> _
and
(4.14)
dp(X,-)) = (a-(p(X,(-')))'/2dB, + a-(p(Xc'))b(X( ))dt p(X o )) = o-
Let ,, = p(Xf+') and 7, = p(X,(-)). Then by (4.13) and (4.14) we have (4.15)
d?7t =zz(a+(t1,+))'nd.§ + a+(rl+b(X()lt
no =bo and
(4.16)
d,7 = (a-(17-))'12d.§;7 + a-(i7-)b(X=`))dt no = 0.
Consider the following stochastic differential equation (4.17)
drl, + =
(a+(j7t'+))'12dB+ + a+(r!s +)b+(ij, +)dt,
hoso
In Theorem 1.1, take
X'lt) = ili, a(t, ) =
X2(t) = 17r++2 (a+())''a,
fi'(t) = b(X,+')a+(ij ), R:(t) = b+(tl,++)a+(tl++)
and
b'(t, 4) = b2(t, c) = b+( a+(D are locally Lipschitz continuous, Since we assumed that a+() and the pathwise uniqueness of solutions holds for the equation (4.17) and hence, by Theorem 1.1, we have
COMPARISON THEOREM FOR DIFFUSION PROCESSES
(4.18)
for all t > 0,
r/+ < r/++
a.s.
Similarly, if r/+- is the solution of the stochastic differential equation d,lr+(4.19)
= (a+(,l+ -))1izd.§ + a+(?I, -)b-(rj; -)dt
lho = ,o,
then we have
(4.20)
r/+- < r/+
for all t >_ 0,
a.s.
Also, if 1f + is the solution of the stochastic differential equation (4.21)
,l-+ = (a-( +)) izdB + a-(r/,-+)b+(r7T +)dt 11/0 -+ = o,
then we have
(4.22)
r/< < r/ +
for all
t > 0,
a.s.
and if ?7,-- is the solution of
- _ (a (, -))11zdB + a-(ii -)b-(q -)dt (4.23)
then
(4.24)
?17 -
for all
t > 0,
a.s.
Finally, set tj, = p(XX). Then, since
max rI; = max rl o5s5r
OSsS
(c)
min r/+ = min 7,
OssSr
OSs ,+(t)
and t < yr+(t), we have (4.25) max rI+ > max r/, and min 7s < min rl,. 05.t OSsSt 0:915r 05s5t Similarly, using the fact that t > V -Q), we have (4.26)
max r/s < max ri, and min r/, > min r/,.
oar
0_<<5:5r
ossst
of ., t
457
458
COMPARISON AND APPROXIMATION THEOREMS
Combining (4.18), (4.24), (4.25) and (4.26), we have
max,s <maxna <maxris=rig 0<s5r 0<s5t
Osssr
and
rtj=maxrj,<maxris <maxna+=;t++, O5z:
osssr
ossst
Similarly we can obtain ni+<17,
This proves the theorem. Remark 4.1. If a(x) in Theorem 4.1 depends only onp(x), i.e., if there exists a function defined on I such that a(x) = a"(p(x)), then a-(0 = a() and therefore we may assume that ?1,+,+ = n7 + and n; - = n_ In this case we have
nj - < nr < j+,+
for all
t > 0,
a.s.
As an application of Theorem 4.1 we shall now investigate the possibility of explosions for non-singular diffusions. Let X=(X(t)) be the ddimensional diffusion process determined by the solution of (4.1). We assume d > 2 and det (a'J(x)) > 0 for all x E Rd. Our problem then is to find a criterion which determines whether or not explosion happens in finite time, i.e., whether e is finite or infinite. It is easy to see that X(t) leaves any bounded set containing X(0) in a finite time a.s., and hence we may assume for this problem that a'J(x) = 8tf and b'(x) = 0 for I x < 1
and i, j = 1, 2, ...,
d.
a
Let p(x) = j x j 2/2 = E (x')2/2, xE Ra. Then t-1
I= [0, oo) and S = {0} . In this case, a
a(x) = E a'°(x)x'xJ '.1-I
and
b(x) =a(x) 1{
t-t
a"(x)/2
Following (4.5), we set
+ E b'(x)x'} t-1
for x e Rd\ 101.
COMPARISON THEOREM FOR DIFFUSION PROCESSES
a+(r) = max a(x),
a-(r) = min a(x)
b+(r) = max b(x),
b-(r) = min b(x)
459
IXI -Zr
IXI-
and
XI-
I X1-,/V
for r E(0, oo). It is easy to see that these functions are locally Lipsc-
hitz continuous and at(r) > 0 for r E(0, oo). Let r+ = (r,) and r- _ (rr) be the minimal diffusion processes generated by the operators + dl r(1 d2 `(r) l 2 dr2 + b (r) dr)
and
1 d2 dl a-(r) (2 dr2 + - b-(r) drl respectively. Set
c+(r) = exp f 2b+(u)du,
c-(r) = exp f r 2b-(u)du
and
s+(r)
f i c+(u)
du,
s -(r)
=
f t c (u)
du
for r e (0, oo). By the above assumption,
s+(r) = s-(r) =
-(2 - log
- 1) (l/r"_' -1)
I r
d>2
d=2
for r e(0,1). Hence s+(0 -) = s-(0 -) = - oo. By Theorem 3.2, if e+ and e- are the explosion times for (r+(t)) and (r-(t)) respectively, then
P; {e+ = oo} = 1 or P, {e+ < oo} = 1 accordingly as
for r e (0, oo)
COMPARISON AND APPROXIMATION THEOREMS
460
JI
[l /c+(r)] fr [c+(u)/a+(u)]dudr = 00
or < oo;
also
P, {e- = oo} = 1 or P,{e- < 00} = 1
for r E(0, oo)
accordingly as
S1 [1/c-(r)] f 1 [c-(u)/a-(u)] dudr = o0
or
Here P, and P,- are the probability laws of r+ = (r+(t)) with r+(0)=r and r- = (r-(t)) with r-(0)=r respectively. By Theorem 4.1, we may assume that max r-(s) < max I X(s)12/2 < max r+(s), a.s. OSsSt
OSssr
O5s5i
and hence
e+ < e < e-,
a.s.
Consequently we have the following result. Theorem 4.2. (Hashiminsky [46]). Let P. be the probability law of the
solution X = (X(t)) with X(0) = x of (4.1). (i)
If
f 1/c+(r) f' [c+(u)/a+(u)]dudr = oo,
then PP(e = oo) = 1 for all (ii)
x E=- R4.
If
f 1/c-(r) f' [c-(u)/a-(u)]dudr < then Px(e < oo) = 1 for all x ERd. 5. Applications to diffusions on Riemannian manifolds First we shall introduce some necessary notions in differential geometry.
APPLICATIONS TO DIFFUSIONS ON MANIFOLDS
461
Let M be a d-dimensional complete Riemannian manifold and V be the Riemannian connection (cf. Chapter V, Section 4). To every pair X, Y E 3r(M) we associate a mapping Rx,.: 3r(M) - 3r(M), called the curvature transform, by (5.1)
RXY = Fcx.r) - (V VY - V4Px).
In local coordinates (x', x2, (5.2)
... , xd),
RxrZ = X'YJZkRalajak
where X = X'a1, Y = Y/a1 and Z = Z1a, (5.3)
where R1
(al
=
x1) ; furthermore,
and R,Jk, = gfhRhJk,
was defined in Chapter V, Section 5 by
R'Jk1
Al = ak{11J} - al{k1J} + ({1a J} {k'a} - {kal} {lla})-
A plane section at x e M is a 2-dimensional subspace of TI(M). Let be a plane section at x and let X,YeTT(M) be an orthonormal of is defined by basis for . The sectional curvature
K(c) =
It can be shown that depends on alone and is independent of the particular choice of X and Y. In local coordinates, K(b) = X'YJ YkXIR1Jk1
if X = X1a1i Y = Yla1 and {X, Y} are orthonormal. For X E=- X(M) such that 11XII = <X, X> =1, the Ricci curvature p(X) of the direction X is defined by d
p(X)=ZK({{X, Y1}}), 1-2
where {X,Y2i ... , Yd} is an orthonormal basis at TI(M) for each x and { {X, Y1} } is the plane section generated by X and Y, p(X) is independent of the choice of {YZ,Y3i ... ,Yd} . In local coordinates, *
is the inner product in each tangent space.
462
COMPARISON AND APPROXIMATION THEOREMS
p(X) = X'XJR1 j and
Rt j = Rkrjk
where X = X'at with IIXII = 1.
Definition 5.1. M is called a space of constant curvature if K(c) is independent of both the choice of the plane section at each x E M and of the point x E =-M.*'
Definition 5.2. Let y: I -- M be a smooth curve. y is called a geodesic if the tangent vectors y* = dt are parallel along y; in local coor-
dinates, y(t) = (y'(t),y2(t), ... ,yd(t)) is a geodesic if and only if (5.4)
dt2- (t) + {`kJ} dti dtj =
0,
t e I.
The following are a consequence of the existence and uniqueness theorems for the differential equation (5.4) :*2 for every x EMandX E TT(M) there exists a unique open interval J(X) in R' which contains 0 such that
(i) there exists a geodesic y: J(X) - M such that y(O) = x and y*(0) = X;
(ii) if a < 0 < b and c: (a,b) ----M is geodesic such that c(0) = x and c*(0) = X, then (a,b) C J(X) and c(t) = y(t) for t E (a,b). Definition 5.3. The above geodesic y is denoted by yx. We set
T,(M) = (X (E TX(M); 1 E J(X)) and define the exponential mapping exp: T=(M) --- M by (5.5)
exp X = yx(1).
It is also denoted by exp,X. Clearly
exp tX = y1(t).
Set
ro = rro(M) = max {r; there exists a ball in T.. (M) of radius r about 0 on which expxa is a diffeomorphism}. *t By Schur's theorem, the first property implies the second property if d Z 3 ([7]). *2 e.g. [47].
APPLICATIONS TO DIFFUSIONS ON MANIFOLDS
463
If M is a simply connected space of constant curvature c then r, is given as (5.6)
c>0
ro(c) )
c S 0.
100,
We also state the following fact known as Meyer's theorem. If the lower bound of the Ricci curvature is not less than (d-1)a2 > 0 * then M is compact and the diameter of M is not greater than 7r/a. Furthermore, the fundamental group of M is finite ([7] and [122]). Let = (01,02, ... , 9d-1) be the spherical polar coordinate on the sphere in Txo(M). (r = d(xo, x),0',92, ... ,9d-1) induces a local coordinate in a neighbourhood U of x0, called the geodesic polar coordinate, by the exponential map (r, ) 1-- expxo
Under this local coordinate, the Riemannian metric has the form 1
(gi,) =
0
ll
0(g,)I'
where (gt,(r, 01,02, ... , Bd-1)) is a Riemannian metric on the geodesic sphere S(xo;r)= {x; d(xo,x) = r, x E M} . The Laplace-Beltrami operator has the following expression L2 1
are
(a)a ,,/det
air
d-1 a + ,,/jet G i.,-1 (e(J/at---) a9! ae where G = (g11) ([47]).
Let h be the Jacobian determinant of expxo. Then we have (5.8)
ar log h
d
r 1- ar log
et G.
Also, if M is a space of constant curvature c, then * We assume a > 0.
464
COMPARISON AND APPROXIMATION THEOREMS
(5.9)
a a log A, ar log /det G = ar
(
where
1
- sin '/ a r)
0 < r < ro(c),
d-'
c>0
,
c=0
A=A(r,c)= rd-', 1
(
- sinh /=c r)
d'1
c < 0.
The following lemma plays an important role in our further discussions.
Lemma 5.1. Let the Ricci curvature p(X) satisfy that
(d - 1)b < p(X),
X E TT(M),
and the sectional curvature section . Then we have
ar
a for every plane
satisfy that
log h < ar log A(r,b) - d r
(i)
JlXii = 1,
log h > ar log A(r,a) -
1
,
d-1
r c- (0, r E (0, ro(a))
and
(ii)
h(r) < A(r, b)/rd-', h(r) > A(r, a)/rd-',
r E (0, rro(M)) r E (0, ro(a))
For a proof, see [7] (pp. 253 -r 256) and [42].
Generally, a coordinate mapping 0: U -- Rd, Uc M is called a normal coordinate mapping at x = ¢-'(0) if the inverse images of rays through OeRd are geodesics. A normal coordinate neighborhood N, the domain of a normal coordinate mapping 0, has the property that every y E N can be joined to 0-'(0) by a unique geodesic in N. As a corollary to Lemma 5.1 we have the following.
Corollary. If the Ricci curvature p satisfies p(X) > (d - 1)c, X E TT(M), JJXJJ = 1, x E M, then the volume of a normal coordinate ball
APPLICATIONS TO DIFFUSIONS ON MANIFOLDS
465
B(xo;r) is smaller than or equal to the volume of a normal coordinate ball of the same size in the simple space form of constant curvature c. By the simple space form we mean the sphere if c > 0, the Euclidean space if c = 0 and the hyperbolic space if c < 0, (cf. [7], p 256).
Now we shall apply the comparison theorem to the Brownian motions on M. Let x0 E M be fixed and let B(xo,r)= {x E M; d(xo,x) < r). For a fixed r* < r,,o(M), let X = (X,) be the minimal Brownian motion on B(xo,r*). As we saw in Chapter V, Section 4 and Section 8 the transition probability density p(t,x,y) of X exists with respect to the Riemannian volume m(dx): PX(X, E dy) = p(t, x, y)m(dy);
moreover p(t, x, y) is C" in (0, oo) x B(xo,r*) X B(xa,r*) and p(t, x, y) = p(t, y, x). Let r* < ro(c) and let ° = be the minimal L(c)-diffusion on (0, r*), where the operator L(C) is given by
Lcc> = I (d' 2 `dr
log A(r, c)) W'-) = + (d 1dr dr
1
1
d
2 A(r,c) dr (
A(r, c)
r
Let us assume that d > 2. Then
s () r
Y A (z, c) A(ro,c) = Jr,o' exp (-Jrro A(z, c) dz dy - Jf.o A(y,c) dy
*I
satisfies s(0 +) = -co. Hence the results of Section 3 imply that 0 cannot be reached in a finite time for °. Also, the transition probability density p (Q' (t, , r7)
exists with respect to the measure A(ri, c)dii, is C" in (0, oo) x
(0, r*)x(0, r*) and satisfies p(°'(t, g, r7) = p(°'(t, r7, ). Noting now (5.7), (5.8) (5.9) and Lemma 5.1, the comparison theorem immediately implies the following results. Theorem 5.1. Let the Ricci curvature p(X) satisfy that (d - 1)b < p(X)
for all X ETT(M), JJXIj = I and x EM. (Let the sectional curvature K(g) satisfy that
a for all plane sections
at every point x.) Let us
fix r* such that 0 < r* < rro(M) (respectively 0 < r* < ro(a)).*2 Then we can construct, on a suitable probability space, the above diffusions (X,) and (Ib)) ( respectively (X,) and such that o°) = d(xo,Xo)
*T 0
COMPARISON AND APPROXIMATION THEOREMS
466
(respectively o > = d(xo,Xo)) and the following is satisfied : if r, = d(xo,X,), then
fit') > r,
(5.10)
(respectively
t°' < r)
for t > 0, a.s. Corollary 1. (A.Debiard-B.Gaveau-E.Mazet [14]). (i)
limpca>(t,r,r1)
r110
(respectively lim p<°'(t, r, r) > p(t, y, x(,)). r,1o
(ii) If C is the explosion time (i.e. the hitting time to r* for (]") (resand the hitting time to 8B(x(,,r*) for (X,)), then E;')(C)
pectively
EY(O, y = (r, 0), 0 < r < r* (respectively E(o)(O > E,,())) where E() and E. stand for the expectations with respect to the probability measures with o°) = r and (X,) with Xo = x respectively. of
Remark 5.1. 0 is an entrance boundary for
and
in the sense
of Ito-McKean [73]. It is well known then that limp`°'(t, r, r1) and r,lo
lim p`"(t, r, rl) exist. rIl0
The proof is immediate from the above theorem and the corollary to Lemma 5.1.
Corollary 2. Let M be a connected, complete d-dimensional Riemannian manifold (d > 2) with non-positive sectional curvature for all plane sections. Furthermore, we assume that the Ricci curvature p satisfies the condition
0 > p(X) > (d - 1)c > -co, X E TX(M), IIXII = 1, x E M, then (X,) is conservative i.e.,
P(t,x,M)= 1
for all xE M.
Proof. Without loss of generality, we may assume that M is simply connected because the Brownian motion on M is obtained as the projection of the Brownian motion on the universal covering space M of M. Then r* = oo by the Cartan-Hadamard theorem ([7]) and the assertion follows immediately from the theorem.
467
STOCHASTIC LINE INTEGRALS
In this result, the condition on sectional curvature can be dropped. For details, cf. [186] and [206]. Similarly, we see that if M is a complete simply connected Riemannian manifold and satisfies the condition
0 > a > K(J > -oo
for all plane sections
then (X,) is non-recurrent. Further applications of comparison theorems, especially, applications to the smallest eigenvalue of the Laplacian, can be found in Azencott [191] Malliavin [108], A. Debiard-B. Gaveau-E. Mazet [14] and M. Pinsky [140].
6. Stochastic line integrals along the paths of diffusion processes Let M be a manifold and y be a smooth bounded curve in M. Then the line integral f r a is defined for any differential 1-form a. Now let
X = (X(t)) be an M-valued quasimartingale, i.e., t - f(X(t)) is a quasimartingale in the sense of Chapter III, Section 1, for every f E F0(M). We can then also define the "stochastic" line integral fxto tj a along the curve {X(s), s E [0,t]} for any differential 1-form a by making use of stochastic calculus; moreover the resulting process t f X10.03 a is a
quasimartingale. One standard way of defining this is as follows. We choose a locally finite covering (W.1 of M consisting of coordinate neighborhoods and choose coverings {UA} and of M such that
Define sequences {Q,F "} and {Tk )} of stopping times by
Q0, Tk> = inf {t ; t > ak711, XX E U.)
UP) = inf It ; t > Tk ), X,
Let a = a,(x)dxt and X(t) _ (X'(t)), rk) S t < o ), under the local coordinate in W,,. We define fxto,q a by setting f J x[o.r]
a
=
f
k
J k>nr
(X(s))odXi(s)
where { v,,} is a partition of unity subordinate to {
It is easy to see that
COMPARISON AND APPROXIMATION THEOREMS
468
f1ra.0 a is well defined and is independent of the particular choice of coordinate neighbourhoods (cf. [50] and [51]). In the following, we shall restrict ourselves to the case when X(t) is a non-singular diffusion on M. As we shall see, fx
dr(t) = Lk(r(t))odwk(t) + Eo(r(t))dt
(6.1)
{r(0)=r.*
Here {L1, L2, ... , L,,} is the system of canonical vector fields corresponding to the Riemannian connection and f, is the horizontal lift of a vector field b on M. The projection X(t) = n[r(t)] of r(t) onto M is the diffusion process on M generated by the operator
A+b. For simplicity,
we shall assume that X(t) is conservative. Let a EA1(M) be a differential
1-form and (a1(r), a2(r), ... , ad(r)) be its scalarization. Recall that ai(r) = aj(x)e;, i = 1, 2, ... , d, if a(x) = a f(x)dx> and r = (x', eD in local coordinates. Let a(b) E F(M) be defined by a(b)(x) = a,(x)b'(x) if a = a,(x)dx' and b = b'(x) ax' in local coordinates. Definition 6.1. We set (6.2)
J xro,:j a = ,1 o
ak(r(s))odwk(s) +
Jo
a(b)(X(s))&
fx[o,,I a is called the stochastic line integral of a E A1(M) along the curve {X(s), s E [0, t]}. This definition coincides with the above mentioned one. Indeed, if U is a coordinate neighbourhood and a < r are any stopping times such that X(t) E U for all t E [Q, r], then
ft o
ak(r(s))odwk(s) + f' a(b)(X(s))ds v
= f' a,(X(s))[ek(s)odwk(s) + bJ(X(s))ds] * As in Chapter V, Section 4, (w' (t)) is the canonical realization of the d-dimensional Wiener process.
STOCHASTIC LINE INTEGRALS
= fo a,(X(s))edXX(s)
469
for any t
It is obvious that t -- ixto.r7 a is a quasimartingale. Its canonical decomposition is given in the following theorem. Theorem 6.1. (6.3)
LOX a =
f
o
ak(r(s))dwk(s) + f o
(a(b)
-
2 aa) (X(s))ds
where b: d,(M) -- F(M) is defined as in Chapter V, Section 4. Proof. We have
f
J X[o, r)
a = f' ak(r(s))°dwk(s) + f' a(b)(X(s))ds 0
0
= fo ak(r(s))dwk(s) +
+f
2
f
o
d[ak(r(s))]
desk(s)
a(b)(X(s))ds. o
Since r(t) is the solution of (6.1),
d[ak(r(s))] = (Ljak)(r(s))°dw''(s) + (Loak)(r(s))ds
and hence d [ak(r(s))] . desk(s) _ (LJak)(r(s))dw'(s) dwk(s)
= Ek-l(Lkak)(r(s))ds. By Proposition V-4.1, we have
Lkak = (Pa)kk where {(pa)ts} is the scalarization of the (0,2)-tensor Pa. Consequently Lkak = (Pa),jerkee,
and by (4.22) and (5.14) of Chapter V,
COMPARISON AND APPROXIMATION THEOREMS
470
k-I
Lkak = E k-i(Va)!Jekek = gtJ(Pa)fJ = -8a.
Example 6.1. Let M = R2 and X = (X,(t), X2(t)) be the two dimensional Brownian motion. If
a=
(xidx2 _ x2dx'),
x = (x', x2),
2 then
2{ j X.(s)dX2(s) - f
LO ,r a
o
X2(s)d1l(s)}.
o
This stochastic integral in the case X(0) = 0 was introduced by P. Levy ([100] and [1011) as the stochastic area enclosed by a Brownian curve and its chord. We set
S(t) = f x10,=] a= 1 { f o X,(s)dX2(s)
(6.4)
-f
X2(s)dX,(s)} o
and
r(t) = ..IX,(t)2 + X2(t)2. Then by Ito's formula, r(2)2
(6.5)
=f
o
r(s)dB,(s) + t
where
B,(t)
'X2 (s) dX,(s) + f r(3) dX2(s). r = f -(s)) o
o
Clearly the system of martingales (S(t),B,(t)) satisfies the relations , = t, , = 0 and <S,S>, =
where ct is the inverse function of
t I--
4
f
r(s)Zds. 0
I f' r(s)2ds. Let B2(t) =S(Or)
4
o
STOCHASTIC LINE INTEGRALS
471
By Theorem 11-7.3, the Brownian motions B1(t) and B2(t) are mutually independent. We suppose from now on that X(O) = x is a fixed point in R2. Then we know that r(t) is the pathwise unique solution of (6.5) with r(O) = I x I (Chapter IV, Example 8.3). In particular, this implies that a[r(s), s < t] c a[B1(s), s S t], and consequently the processes {B2(t)} and {r(t)} are mutually independent. Thus we have obtained the following result. The stochastic area S(t) has the expression (6.6)
S(t) = B2(4 J'
r(s)2
)
where B2(t) is a one-dimensional Brownian motion independent of the pro-
cess {r(t)}. Suppose that X(0)=x 0. Then X(t)* 0 for all t > 0 and hence we can introduce the polar coordinate representation: X1(t) = r(t)cos0(t) and X2(t) = r(t)sinO(t). An application of Ito's formula yields
dS(t) = 2 {r(t)cos0(t)sin0(t)dr(t) + r(t)2(cos0(t))2d0(t) - r(t)cos0(t)sin0(t)dr(t) + r(t)2(sin0(t))2d0(t)}
= 2 r(t)2d0(t), and hence S(t) = 2 f o r(s)2d0(s).
Then
0(t) : = 0(t) - 0(0) =
f, r(s)2 dS(r),
and consequently d
= r( t)2
d
d<0, 0>, = r4 d<S, S>, = r(t)2 dt.
Again by Theorem 11-7.3, there exists a Brownian motion
{B3(t)}
(B3(0) = 0) independent of {B1(t)} (and hence independent of the process
{r(t)}) such that
COMPARISON AND APPROXIMATION THEOREMS
472
r
B(t) = B3
J r (1s) 2 is). o
The formula
(6.7)
[Xi(t) = r(t) cos [9(0) + B3 (f o r(s)2 ds)] X2(t) = r(t) sin [0(0) + B3 (f
o r(s)2
ds),
is known as the skew product representation of two dimensional Brownian motion. As an application of (6.6) we can obtain the following formula in the
case X(0) = 0: (6.8)
E(et1stt)) = (cosh Zt)-
for
E R.
Indeed, noting the independence of {B2(t)} and {r(t)}, we have E(eicsct)) = E(exp [icr B2
(4 f
r(s)2ds)]) o
= E (exp [- g2 fo r(s)2ds])
1)2 = = IE(exp[ - gf o,i(S)2dr])
where ti(t) is a one-dimensional Brownian motion with r/(0) = 0. Since I
c
ri(ct)}
(q(t)) for every c > 0,
E(er{sa>) = JE (exp [- 82 t-
f o'i(s)2ds])}
2.
Therefore it is sufficient to prove the formula (Cameron-Martin [10] and Kac [80]) (6.9)
E(exp [- A f o 17(s)2ds]) = (cosh
/2--A)-112
(I > 0).
Let K(t, s) = t As and consider the eigenvalue problem in X2([0,1]):
STOCHASTIC LINE INTEGRALS 2
J
0
K(t,s)g$(s)ds
473
_ (t)
The eigenvalues and eigenfunctions are given by
(n+2)7r
2n
2,
n=0, 1,
= 2 sin (n + 1 7rx,
...
Now it is easy to see that
q(t) = n
where the n are independent random variables with common distribution N(0,1). Then 2
f Jor72(s)ds=EAR, and hence
E(exp[-A
Jo
rt2(s)ds]) = n E (exp [0
A.
ll
z J)
82
= R (1 + (2n + 1)2n2)
= n-0 II
1+2 A.
-1/2
_ (cosh
2A
We can also obtain a formula which is more detailed than (6.8) (Gaveau [33]). Let X(t) be the two dimensional Brownian motion with X(0) = 0. Then for every x EE R2, t > 0 and E R, (6.10)
scr>X(t) = x) = E(e'I
r t exp I (1 L 2sinh Zt
- 2t coth Lt) x2t
2
2J
(6.8) is an easy consequence of (6.10) and hence the following proof also provides us with another proof of (6.8) (or (6.9)). * * Still another proof of (6.9) may be found in [114].
474
COMPARISON AND APPROXIMATION THEOREMS
Proof. By the rotation invariance of the Brownian motion it is easy to see that E(e"sa) I X(t) = x) = E(etls
I r(t) = r)
where r = I x 1. By (6.6), we have E(etcsct) I X(t)
= x) = E (exp [- 4
f
r(s)2ds I
I r(t) = r).
0
But we know (Chapter V, Section 3) that
u(t, x) = E((exp [-a f' I X + X(s) 12ds])f(x+X(t))), t > 0, x E R2, 0
where a > 0 is a given constant, is the solution of the initial value problem
at=zGu-aIxL2u
(6.11)
luIt_0 =.f Let {HH(x); n = 0, 1, ... } be the Hermite polynomials and set n m(x) = (exp C- I x 12])Hn(xl)Hm(x2) for x = (x),x2) and n, m = 0, 1, ...
Since n m
satisfies
(A - I x 12)5b,,,m + 2(n + m + 1)0n,m = 0,
we can solve (6.11) by the method of eigenfunction expansion:
u(t, x) = f RZPa(t, x, y)f(y)dy where
Pa(t, x, y) _ E
n,m-0
en.m(x)en,m( )
and
en m(x) =
A formula on Hermite polynomials yields
STOCHASTIC LINE INTEGRALS
475
H,((2a)i'4xr)Hn((2a)'14Y)
Pa(t,x,Y)=II E e `' r-, -0 (6.12) 2
exp
)` e eTa(z?+r2)12
("
[_, 2a coth
_
2"n!(2a)-,l4
/7G
2at(x? - 2xysech tat +y J]
27z sinh /Tat (2a)-'12 x = (x,, x2),
Y = (Y,, Y2)
Therefore,
E(e"() I X(t) = x) =
0, x) (27tt exp [- 121 2) _3 t
tt exp (1 - 2 cosh2)
12t
2
2 sinh Zt
Finally we consider the three dimensional process
X(t) = (X1(t), X2(t), X3(0)+S(t)),
where (X1(t), X2(t)) is a two-dimensional Brownian motion and S(t) is defined by (6.4). It is a diffusion process on R3 with generator
A=
(L21 + Li), 2
where
It is a degenerate diffusion, but since L dim a {L1, L2} x
L
a
=3 for every x. Thus the smooth transition probability
density p(t, x, y) exists by Theorem V-8.1. It is easy to obtain from (6.10) that m
z
P(t, 0, x) _ u;;)
J
sinh /2 eXp
x + X2 2t
{t x3
1/2
tanh
/2}
dd,
x = (x,, x2, x3)
476
COMPARISON AND APPROXIMATION THEOREMS
(cf. Gaveau [33]). Also, by Example 8.1 given below, we see that the topological support 5S'(P.) of the law P,, of {X(t)} with X(0) = x is the whole space W, = {w; C([0, oo) -- R3), w(O) = x). For a further detailed study of this diffusion, we refer the reader to Gaveau [33].
Remarks 6.1. The formulas (6.8) and (6.10) can be obtained more directly by the following Fourier series expansion, (cf. [33], [216]): Let X(t) = (X'(t), X2(t)) be a two-dimensional Brownian motion such that
X(0) = 0 and set, for i = I and 2,
v/2 f oin 2nkt dX() and
k = -./ 2
f1o
os 2irkt W(t),
k= 1, 2,
Note that [ X'(1), X2(1), k'), 412), Yfj1), rlk2), k = 1, 2, ... )is a system of independent Gaussian random variables with mean 0 and variance 1. Let S(1) be the stochastic area: .
S(l) = I f
0{XI(s)dX2(s) - X2(s)dX'(s)}.
Then it admits the following expansion in the sense of .2 (P) and almost surely :
M S(1)
27Ik
2 X2(1))Sk1 }.
`(i 1) - ., 2 X'(1)JSk)
This can be deduced, at least heuristically, from the Fourier series expansion of X'(t): X1(t) = tX°(1) + / T E (k° f 'sin 2xcks ds + rik° f 'cos 2xcks ds) k-t
0
i = 1,2. It can be proved rigorously as follows: For F(s, t)E-9 ([0, 1]2), the multiple stochastic integral I (F)
= J of
(s, t
o
)dX' (s)dX 2(t )
is defined to be the iterated stochastic integral
STOCHASTIC LINE INTEGRALS
477
f o{ f F(s, t)dX'(s)}dX2(s). 0
This is well-defined because, setting .
=v[X'(u),XZ(v); 0
for 0
45(t). = f (s, t)dX'(s) is (.9)-adapted measurable process and X2(t) is an (.9;)-Brownian motion. It is easy to see that I(F) is also given by f o{ f F(s, t)dX2(t) }dX'(s) 0
and that, if F, G E 9Z([0, 1 ]2), then E[{ I(F) - I(G) }2] =
f 1J [F(s, 0
t) - G(s, t)]2 dsdt.
0
In particular, if F,,, FE=-2'([0, T]2) and F. -* F in .2([O, 1]z), then 2'2(P). Also it is clear that, if F(s, t) = f(s)g(t), f,gE
I(F)
Y 2([O, 1]), then
f g(s)dX2(s).
I(F) = f 'f(s)dX'(s)
0
Now define F(s, t)E P2([0, 1]Z) by
0<s
1/2, F(s,t)-{-1/2,0
Then I(F) = S(l). Also F(s, t) can be expanded into Fourier series in .9' ([0, 1]Z):
F(s, t) =
°° ( (cos 27cks - .. I -T) sin 27rkt
2nk
_
sin 2ncks(cos 27rkt - ../_T) 2nk
Now, (*) can be easily calculated in 9'Z(P)-sense and, since this is a sum
of independent random variables, a.s. covergence also holds. Then, if I a ( < 2n,
478
COMPARISON AND APPROXIMATION THEOREMS
I: = E[eas(l) IX'(l) = x1, X2(l) = x2}
-
= E[exp[Ek-1°k2n{ (77k1)
=kfI E[exp[ 2nk (ylkl) L
2x2) T ]]]
2 x')S 2) - (77k(') 2
a (n1k) x')kk2)]]E[eXp[- 2nk
-
for (x1, x2) E R2. We see easily that, if a e R, E[exp[ 2nk (rly) -
= E exp [-
=
(2n)-1J m
(1 -
2 nk J
2 a)tL2)]]
(nk2) - - 2 a)tk' )
mexp[ 2nk 2
(x -
)exP[(1 -1/2
2 a)y - (x2 + y2)/2]dxdy
- (2nk)) (_2_nk) a ]. o
2
o
1
2 2
Hence
_
-( I - k?1(1
7
2nk)2)_.
= 2 sin a/2 exp[(1
1
exp[k-1(1 -(
- 2cot 2) I x
Q
2
2nk))
2nk)21
1
12
x
12]
by using well-known formulas sin x x
x2 ) = II( 1 - n22 n ,
1
cot x = x + 2x
1
x2-n2n2
By an analytic continuation and the scaling property of the Wiener process, we obtain (6.10). The proof of (6.8) is similar. For another interesting proof of (6.10), Cf. Yor [236]. Example 6.2. Let M be a Riemannian manifold. The horizontal Brownian motion r = {r(t)} is a diffusion process on O(M) determined by the equation (6.1) with La = 0: (6.13)
dr(t) = Lk(r(t))°dwk(t) I r(0) = r.
STOCHASTIC LINE INTEGRALS
479
Let b be a vector field on M. The differential 1-form co(b) is defined as
usual by co(b) = b,(x)dx' where bo(x) = g11(x)b'(x) and b = b'(x) ax' It is easy to see that the scalarizations of b and w(b) coincide:
cv(b):(r) = b,(x)e{ = b'(x)fi = b'(r),
i = 1, 2, ... , d.
Suppose that b'(r) = a(b)t(r) is bounded on O(M) for i = 1, 2,
... , d.
Then (6.14)
M(t) =exp { 5 f b'(r(s))dw'(s) - 2
Jo
[b'(r(s))]2ds}
o
is an exponential martingale. By the transformation of the drift determined by M(t) (cf. Chapter IV, Section 4.1), we obtain a d-dimensional Wiener
process w(t) = (W(t)) where V(t) = W(t) - f' b'(r(s))ds. The equation (6.1) now becomes
dr(t) = Lk(r(t))°dwk(t) + Lo(r(t))dt
(6.15)
r(0) = r
I
where Lo is a vector field on O(M) given by Lo(r) = bk(r)Lk(r). It is immediately seen that Lo coincides with the horizontal lift of b. Thus the solution of (6.1) is obtained from the solution of (6.13) by the transformation of drift determined by M(t). In particular, the diffusion generated by the
operator
d+b is obtained from the Brownian motion by the same
2 transformation. Note that M(t) can also be expressed in the form
M(t) = exp L1 xro,tl
.(b) +
a[W(b)](X(s))ds
Jo
2
2 J o Jkw(b)ll2(X(s))ds]
=
exp [J
X1010
w(b)
2Jo
div(b)(X(s))ds 2 J o I!b11Z(X(s))dr,.
This follows immediately from (6.3) and
COMPARISON AND APPROXIMATION THEOREMS
480
E Y(r)2 =
d
e{e;bj(x)b,(x) = gfl(x)bb(x)b,(x)
= g,,(x)b'(x)bt(x) = Ilbll2(x)
7. Approximation theorems for stochastic integrals and stochastic differential equations As we have seen in this book, stochastic integrals and solutions of stochastic differential equations are most important and typical examples of Wiener functionals. If, in these definitions of integrals or equations, a Brownian path is replaced by a smooth path, then they are defined by using ordinary calculus and we are thus provided with functionals defined on smooth paths. A question naturally arises: if we substitute in these functionals defined on smooth paths an approximation of Wiener process (i.e., a process consisting of smooth sample paths which converge to Brownian paths uniformly a.s.), do they converge to original Wiener functionals? Since these functionals are usually not continuous in the uniform topology of the path space, the answer is negative in general and the limiting Wiener
functionals, if they exist, need to be modified according to ways of approximations. For such familiar approximations as piecewise linear approximations or regularizations by convolution (mollifiers), however, the answer is affirmative if we adopt the symmetric multiplication in the definition of stochastic integrals or stochastic differential equations. Let (WW,PW) be the r-dimensional Wiener space * with the usual reference family {.fi}. PW is simply denoted by P below. The shift operator Wo is defined by (0,w)(s) = w(t+s)-w(t). We shall 0, (t > 0): Wo consider the following class of approximations for a Wiener process. Definition 7.1. By an approximation of a Wiener process, we mean a family {Bo(t,w)=(B(t,w),B2(t,w), ... ,Bj(t,w))}6>o of r-dimensional continuous processes defined over the Wiener space ( W0', P) such that (i) for every w, t -. B5(t, w) is piecewise continuously differentiable, (ii) Ba(0,w) is .9a-measurable,
(iii) Ba(t+k(5,w) = B5(t, 6kaw)+w(k6) for every k = 1, 2, ... , t > 0 and w, (iv)
E[B6'(0,w)] = 0
for
i=1,2, ... , r,
* Wo = 1w a C([0, oo) --- R'); w(0) = Of and Pw is the Wiener measure on W.
APPROXIMATION THEOREMS
481
(V)
E[ I Ba(O,w)16] < c63
for i = 1, 2, ... , r, *
E[(fa
I .$Xs,w) I d$)6] < C63
for
i = 1, 2, ... r,
o
where
=
Ea(s,w)
d
Ba(s,w)
If {Bo(t,w)} a,o satisfies the above conditions, then for every T > 0
we have that (7.1)
E{ max I w(t) - Ba(t,w)12} - 0 05r5T
as o 10.
Thus {Ba(t,w)} a>o actually approximates the Wiener process {w(t)} . (7.1) is an obvious consequence of Theorem 7.1 below. A simple application of Holder's inequality yields that
E[(f
a 0
I a (s,w) I ds)D'(
f6
I A6 (s,w) I ds)D2 ... (fa I Ea (s,w) I ds)pm] 0
0
if p,> 1 and p,+Pz+
<
By (iii), this estimate also holds if f Holder's inequality, it follows that o
E[(f "a I E`a (s,w) I ds)D'(f 0
(7.2)
"2a
(*+1)6. Again by is replaced by fk4
I Ea (s,w) I ds)P2 ... ( f "mal Ea (sw) I ds)Dm] 0
g0
<
+p.::96.
n"m(/'f(D1+D2+...+Dm>
if 1 < P,, p, +p2+
+Pm < 6 and n, E Z+.
Let us introduce the following notations:
(7.3)
fr
s,j(t, S): =
E 11
t
[Ba(s,w)l&a(s,w) - &(s,w).a(s,w)]ds} o
2
* c, c', c,, c2, ... are positive constants independent of 6.
COMPARISON AND APPROXIMATION THEOREMS
482
and
c,,(t, S): = l-E { f E (s,w)[Be(t,w) - BB(s,w)]ds}
(7.4)
u
for i, j = 1, 2, ... , r and t > 0. Clearly (s1,(t,S)) is a skew-symmetric r x r-matrix for each t and S. We shall make the following assumption.
Assumption 7.1. There exists a skew-symmetric rXr-matrix (st,) such that
s,,(S, S) - s,
(7.5)
510.*
as
Set
(7.6)
ct, = si, + 2 Sty,
i,j=1,2,...,r.
Lemma 7.1. Let k(b): (0, 1 ] -. Z+ such that k(S) t oo as S 10. Then (7.7)
lam ctj(k(6)6, S) = c,,.
Proof. We set S* = k(6)6. Then for every n = 1, 2, ... , 2ncsrl(na, S)
= k-
a+l) S
= k 0 E[
{Bgs, w).
(s, w) - B'(s, w)-66'(s, w)} ds]
E[f o {(BB(s, ekdw) + w°(kiS)),&6(s, ekow) - (B/s, ekaw) + w'(k6))E (s, ekaw)} ds]
=
0) +
n-
E[w`(kO)(BB(S, ekaw) - BB(0, 0 w))
r
- w'(k0)(Bb(a, ekdw) - Ba(0, Okaw))]
= W s,(a, S) +
`
n-1
k
w) - Ba(0, w)]
- E[wJ(kS)]E[Ba(S, w) - Ba(0, w)]} * We may consider the case that s,1(Sn, s,, for some sequence 6. 10. All results below also hold in this case if we replace {SI by the sequence {an} .
APPROXIMATION THEOREMS
483
= 2n5s,,(8, 6).
Hence a*st,(8, a)
= a*s,,(ar*, a) 2
E[
d r ds (B(s, w)B'(s, w))ds] - E[ J o* B (s, w)E(s, w)ds] J o*
= 2 E[B6((5*, w)B6'((5*,w)) - 2 E[Ba(0, w)&(0, w)]
+ E[
fob*
E (s, w) {BB(S*, w) - BB(s, w)} ds]
- E[BB(S*, w) {Ba(8*, w) - B6(0, w)} ] = 2 E[(Ba(0, Oo*w) + w°((5*))(B,1(0, 05*w) + wJ(8*))] 2 E[Ba(0, w)Bb(0, w)] + 6*ci,((5*, a)
-E[(B'a(0, 05*w) + wf(6*))(Ba(0, Oj*w) + w`(S*) - B6(0, w))]
= 2 E[Ba(0, w)Ba(0, w)] + 2 E[wi(8*)w'(6*)]
- 2 E[Ba(0, w)Ba(0, w)] + a*c,,(a*, S) - E[BE(0, w)BB(0, w)] - E[wt(8*)wf(S*)] + E[wJ(a*)B3'(0, w)]
= a*c,,(a*, d) -
E[wl(6*)wJ(S*)] - E[Ba(0, w)Bb(0, w)]
+ E[w1(&*)BX0,2 w)].
The assertion is now clear since E[w°(8*)wf(6*)] = 6,,a*, I E[Ba(0, w)Ba(0, w)] I < (E[Ba(0, w)'])' "(E[BgO, WAY c25 = o(o*) and I E[w'(8*)B6'(0,<_w)] I
(E[w'(S*)2])l'2(E[BXO, w)2])1 /2
s C,(a* /251/2= OW).
l
/Z
COMPARISON AND APPROXIMATION THEOREMS
484
Before stating our main results, we shall first give several examples. Let 0 be the space of continuously differentiable functions 0 on [0, 1] such
that ¢(0) = 0, 0(1) = 1. Let 0 = dt 0 and dkw' = w'(k8+b) - w'(kS). Example 7.1. Choose 0' E sls, i = 1, 2, ... , r, and set BB(t, w) = w'(k8) + 0'((t - k8)/8)dkw',
(7.8)
if kS
0
Also, it is clear that sj(5, 6) = 0 and hence Assumption 7.1 is satisfied with
sty = 0. (Thus c11 = 2 S,J in this example.) If, in particular, ¢'(t) = t for i = 1, 2, ... , r, then {Bo(t,w)} a>o is the familiar piecewise linear approximation. Example 7.2. (McShane [115]). Let r = 2 and choose 16'E rh, i = 1, 2. Set
(7.9)
w'(k8) + o'((t - kS)/8)dkw',
Bs(t, w)
w`(ka) + 03-`((t - k8)/8)Akw',
dkw'dkw2 Z 0, dkw'dkw2 < 0,
if k6
2f
o
_
{.B (s,
w)E (s, w) - Ba(s, w)aa(s, w)} ds
J dow d0w2 1
2
(1-2 f
APPROXIMATION THEOREMS
485
and hence Assumption 7.1 is satisfied with
(1 - 2 f 01(S)o2(S)dS).
S12 =
Example 7.3 (mollifiers). Let p be a non-negative C°°-function whose support is contained in [0, 1] and f' p(s)ds = 1. Set P,($) =
for 8> 0
P (S)
S
and
(7.10)
Bgt, w) = f W w'(s)pb(s - t)ds = 0
f w'(s + t)pa(s)ds 0 o
for
i = 1, 2, ... , r.
It is easy to verify that {Ba(t, w)} 610 is an approximation of a Wiener process. Indeed, (i) . (iv) in Definition 7.1 are obvious and (v) and (vi) are verified as follows. We have I Ba(0, w)
12m
= ( fo
wt(aS)P(s)dS)2mam
and hence
E[ I BB(O,w) 12m] = (SmE[( f ' w'(s)P(s)ds)2m].
Also,
f w'(s+BJp'()dd
961(s, w)
and hence
E[( f 0
ds)2m]
I A;(s,w) I
= E[(f
I .a( s, w) Ids) 2162m 0
= E[(f0 I f
w'(8(s + ))P'()dd 1 ,*)2m] 0
486
COMPARISON AND APPROXIMATION THEOREMS
= E[(f 11 0 f 01 W"'(" +
S))
= E[(f 1 j f 1 wt(s +
p'( dd
N/
0
0
pi(b)dS I ds)Zm]am
ds)sm]8'".
Clearly s j(6,E)= 0 and consequently Assumption 7.1 is satisfied with
stj=0. First, we shall consider the approximation of stochastic integrals. Let a be a differential 1-form on P given by
a = E at(x)dxt. t-1
We shall assume that all partial derivatives of at(x), i = 1, 2, ... , r, are bounded. In particular, we have
I a,(x)1 < K(1 + xI),
i =1, 2, ... , r,
for some positive constant K. Set (7.11)
A(t, a; w)
= Lo,t] aE f
at(w(s))adwt(s) 0
and
(7.12)
A(t, a; B5) = f
a = t-1 Ef ablo,t]
t
at(Bo(s))ha(s)ds. 0
Theorem 7.1. Let {B5(t, w)} ago be an approximation of a Wiener process satisfying Assumption 7.1. Then
lim E[ sup I A(t, a; B5) - A(t, a; w) (7.13)
J10
OStST
-ftE
0 ij-1 stj ataj(w(s))ds
1 Z] = 0
for every T > 0. Here 8a j = a a j. Proof. * It is sufficient to show that * We are indebted to S. Nakao for the main idea of the proof.
APPROXIMATION THEOREMS
487
lim E[ sup I f' u(Ba(s, w))h6(s, w)ds osrsr
j10
(7.14)
0
- f u(w(s))
° dw'(s)
0
-f
0
(Ei-1s1 u,(w(s)))ds I Z] = 0
for every T > 0 and j = 1, 2, ... , r. Here u: Rr --- R is a C2-function such that all partial derivatives are bounded and u1(x) =
8
u(x).
First we shall introduce the following notation. Choose n(8): (0,1] --Z+ such that
n(6)4S 10 and n(8) j oo as 6 j 0.
(7.15)
We shall then write a = n(8)&,
(7.16)
[s]+(a) _ (k + 1)3
(7.17)
if kb < s < (k + 1)3
[s]-(a) = kb
and (7.18)
m(t) = m(t)(6) = [t]-(a)/3
Integration by parts yields ck+1)a
f
u(Ba(s, w))dBB(s, w) ka
= -f
ck+1)a
u(B6(s, w))
ka
TS
[BB((k + 1)b, w) - Ba (s, w)]ds
= u(Ba(k3, w))[Bb(k3 + 3, w) - Ba(kb, w)] (7.19)
+if -1
(k+1)a kJ
u,(B5(s,
w) - BB(s, w)]ds
_ (u(Ba(k3, w)) - u(w(k3)))[Ba((k + 1)a, w) - Ba(k3, w)]
+ u(w(k3))[B4((k + 1)a, w) - w'((k + 1)b)] -u(w(k5))[Ba(k3, w) - wl(k3)] + u(w(kS))[w1((k + 1)b) - wf(kS)] (k+1)a
+ i-1 E f ka
u1(Ba(s, w))h (s, w)[B'((k + 1)S, w) - BB(s, w)]ds.
COMPARISON AND APPROXIMATION THEOREMS
488
Also,
u(BB(s, w))dBBs, w)
r
5
= -u(Ba(t, w))[BB([t]+((5), w)
- B (t, w)]
+ u(Ba([t]-(a), w))[BJ([t]+(a), w) - Bi[t]-(b), w)] ur(Ba(s, w))fa(s, w)[Ba([t]+(b), w) - BB(s, w)]ds
+riJ
(7.20)
= -[u(B4(t, w)) - u(B5([t]-(a), w))][Ba([t]+(a), w) - Bit, w)] [u(Ba([t]-(a), w)) - u(w([t]-(a)))][8 ([t]+(a), w) - B'$t, w)]
-
- u(w([t]-(a)))[B'([t]+(a), w) - Bit, w)] + u(Ba([t]-(a), w))[B'a([t]+(a), w) - Ba([t]-(b), w)] t
+E
w
ut(Ba(s, w))B6(s, w)[B[t ]+(a), w) -
w)]ds.
By (7.19) and (7.20) we have
f u(Ba(s, w))dB s, w) 0
-f
u(w(s)). dw'(s)
0
- f t Es,,u,(w(s))ds 0 r-1
fo u(Ba(s, w))dBgs, w) - f o u(w(s))dw'(s) - f
t
01-1
c, ju,(w(s))ds
= - [u(Ba(t, w)) - u(Bo([t]-(a), w))][BB[t]+(a), w) - B'(t, w)] - [u(Ba([t]-(a), w)) - u(w((t]-(a)))][B'([t]+(a), w) - BJ(t, w)] - u(w([t]-(a)))[B'a([t]+(a), w) - Ba(t, w)] + [u(Bo([t]-(b), w)) - u(w([t ]-(a)))][B'([t]+(a), w) - B ([t]-(a), w)] + u(w([t]-(a)))[B6([t]+(a), w) - BB([t]-(a), w)] (7.21)
+'f' ,_
Cr7-rte
ur(Bo(s, w))E (s, w)[B [t]+(a), w)
-
B'a(s, w)]ds
- u(w([t ]-(a)))[w'(t) - w'([t ]-(a))] [u(w([t]-(a)))(w>(t)
- w'((t]-(a))) -
f` u]
u(w(s))dw'(s)]
N cr1ut(w(s))ds [t]-m r r
+
m1 (t -[u(B5(ka, k-0
w)) - u(w(ka))][B1((k + 1)a, w) - B'(ka, w)]
APPROXIMATION THEOREMS
489
M(r) -I
+ E u(w(kS))[BB((k + 1)S, w) - w!((k + 1)S)]
-
k-0
m(t
1
u(w(ka))[Ba(ka, w) - w'(kb)]
k-o
1u(w(k6))(w'((k + 1)g) - w'(kb)) - f o]
+[
r m(t -1
+E t-1
0
k-0
u(w(s))dw'(s)1
(k+1)a
fka
[ut(B,(s, w)) - ut(w(ka))].a(s, w)[Ba((k + 1)S, w)
- B(s, w)]ds
+
t -1 r m(E
ut(w(kb)) f
(k+1)a
[$a (s, w)(BB ((k + 1)S, w)
- B(s, w)) - cj(S, 8)]ds r m(t -1
+ t-1 E k-
r m(t -1
-I-N E
k-0
ut(w(k3))[cj(3, S) - ctj]a
f
(k+1)a
[u(w(kb)) - ut(w(s))]ds c,
ka
r
= I1(t) + I2(t) + 13(t) + 14(t) + I5(t) + E Its(t) + I7(t) t-1 + 18(t) + 19(t) + 110(t) + 111(t) + I12(t) + I13(t) r r ' r r + E Ii4(t) + E Iis(t) + E Ii6(t) + E I17(t) t-1
1-1
t-1
1-1
It is clear that (7.22)
E[ sup I I9(t)12] -- 0
as 610
and, by a martingale inequality,
(7.23)
E[ sTj 113(t) + I8(t) 12] S c4E[f {u(w([t]-(a))) - u(w(t ))} 2dt] o
-0
as
By Lemma 7.1, it is clear that (7.24)
E[ sup 1116(t )1 2] o5t5r
We have E[ sup I I,7(t)12] 0:9 5r
0
as
&10.
510.
COMPARISON AND APPROXIMATION THEOREMS
490
< c, Et-1E[(f T 0I wt([s]-(J)) - wt(s) I ds)9 (7.25)
< c,T E f a E[ I w`((sl-(b)) - %I(s) 12lds 1-1
< c,T2rS - 0
b 10.
as
Also, by (7.2),
E[ sup
I11(t)12]
05t5 r
< C6 E E[ sup (Bgt, w) - B`a([t]-(b), w))2(B'd([t]+(3), w) - B'B(t, w))ZJ i-1
05t5 T
< c6 E E[ sup i-1
r (7.26)
C6
ftj+(
I
hAs,
w) I
05i5T J (]W (f(k+ua E[ max
t-1 r
(
05k5m(T)
(k+1)d
m((Tr)
E[(f
< c6(m(T) + 1)
+1)a
.a(s, I
I $'a(s, w) I ds)21
[t7- (d) ds)2 (f"41
I Ea(s, w) I i
ka
C6 E A E[(f kd
flt)Tcd)
ds)2(
w) I
ds)2(f
(k+1)3 kd
Z
s w) I ds) l a(,
I
Ef(s, I
w) I
ds)2J
I ,§b(s, w) I ds)2(f I E (s, w) Ids)-] o
< c7(m(T) + 1)n((5)4b2 < csn(b)36 _ 0
as
b j 0.
As for I2(t), E[ sup I
I2(t)121
05f- 5T
< c, E E[ sup (BB([tl-(b), w) - wt([tl-(3)))2 1-1
05t5 T
x (BB([t]+(b), w) - B(t, w))21 < c9 E E[ sup (BB([tl -(S), w) - wi([tl -(V)))2( i-1
0
5t5 r
f
(k+1)a
= C9i-1 E E[0$k<m max(T)(Bj(k(5, w) - wt(k5))2( (7.27)
< c9E
m(T)
i-1 k-0
E[(Ba(k3, w) - wt(kb'))2(f
ka
(k+1)a I kd
ca)
I 1%(S' w)
Ct]-(a)
s, w) I ds)2J $Ja(f
his,
w) I
ds)2l
d
< c9(m(T) + 1) E E[(BB(0, w))2( f 0 B$(s, w) I ds)2] i-1
I Ba(s, W) I d3)4J} 1/2
< c9(m(T) + 1) E {E[(B6(0, w))4JE[(f 1-1
0
I ds)2J
APPROXIMATION THEOREMS
491
1)(a2n(a)4a2)1/2
S c1o(m(T) +
< c118n(8) - 0
as
&10.
In the case of I,(t), we have E[ sup I I3(t)12] 0-5f: 5T
'+ (a)
S c17,i-1E E[OstsD sup (1 + I wr([t]-(a))12)(f[r]-(a) I.$j(s, w) I ds)9 ((k+1)a
< c17, E {E[ sup (1 + I wt(t)12)2]E[ max t-1
(7.28)
OStST
< c13(E[ mE k-0
OSkSn(T)
kJ
I
ha(s, w) I ds)4]} 112
(Irk+ua
I E (s, w) I ds)4])liz
ka
< c14((m(T) + 1)n(8)462)112 as
< c15(n(8)'d)"2 --- 0
&10.
Similarly, as in (7.27) and (7.28), we can prove that (7.29)
E[ sup 1 I4(t)1 2] -- 0
as 610
E[ sup I IS(t)12] J --- 0
as
OStS T
and
(7.30)
OSrST
610.
For I6(t), E[ sup I I5'(t )12] OSrST
< c16E[ sup (
OStS T J [r]
(T)
(a)
I ha(s, w) I ds
I h j(S, fI"") w) I ds} 2] 4-M
(k+1)a
(7.31) < C17 Eo E[ { f ka
(k+1)3
I B6(s, w) Ids f
I Ej(s, W) I ds} 2] ka
< c1,(m(T) + 1)E[ { f o I Es(s, w) I ds f 01 .j(s, w) I ds} 2] < c18 n(8)8 n(t5)462
= c18n(8)38 -- 0
As for I,(t), E[ sup II,(t)12] OSt T
as 6 j 0.
COMPARISON AND APPROXIMATION THEOREMS
492
c19E[ sup 0+ I w([t]-(b)) 12) sup I w'(t) - w'([t]-(b)) 12] 0515 T
O5t5 T
sup I w'(t + kb) - wJ(kS) 14]}' 12
< c20 {E[ max
OSksm(T) 0
(7.32)
mE
< c20 {
k-0
r 112 E[ sup I wJ(t + kJ) - wJ(kZ) 14]} 05t5a
c21
1
n(8)8
(n(8)8)1
= c21(n(6)6)112 --- 0
as 8 10.
We estimate 110(t) by E[ sup I I10(t) 12] 0:9t-19T
m(t)-1
< E[ sup
[u(Ba(kS, w)) - u(w(kS))]2
OSt$T k-0
m( t
-1
k-0
[Ba((k + 1)9, w)
- B'(kS, w)]2]
< E[
m(T)-i
m(T)- 1
[u(B5(kS, w)) - u(w(kS))]z E [BJB((k+ 1)S, w) k-0
k-0
(7.33)
S {E[m(T)
m(T
-
Ba(k3, w)]2]
1
[u(B5(kS, w)) - u(w(kS))]4) m(T)-1
X E[m(T) E [BB((k + 1)S, w) - Ba(kS,w)]4} 112 {m(E
< c22m(T) i-1
k-0
1 E[ I B6'(kS, w) - w'(ko) 14] m(7)-1
x E E[ I BB((k + 1)g, w) - BB(kS, w) 14} 1 iz < c23m(T)28 {2E[ I BJa(0, w) 14] + E[ I w(8) 14]} 111,
c24m(T)28S < c25n(8)-' --- 0
as 810.
In the case of 111(t), we first note that nn(w)
=
u(w(k3))(BB((k + 1)8, w) - wJ((k + 1)8)) n
=k
u(w(k3))Ba(0, B(k+llaw)
APPROXIMATION THEOREMS
493
Hence, by a martingale
is an {.} -martingale where inequality, E[ sup 1111(t)12] 0<s5 T
= E[ max
1 f]n(w)12] OSnSm(T)-1
5 C26E[J?lm(T)-l(W)12] (7.34)
i l E[u(w(k3))2(B'a(0, 0(,+l)aw))2]
= C26
= C26 m(E l E[u(w(kg))2]E[(B'a(0, w))2] k-0
S c27m(T)&
as 610.
= c28n(5-1- 0 We can show in the same way that E[ sup 1112(t)12] -- 0
(7.35)
as 8 10.
O5r T
As for 114(t), E[ sup I I'14(t) 1 21 051ST
m(1 -1
(k+1)a
f k-0 kJ
< E[ sup { 0$t5T
[ I u1(Ba(s, w)) - u1(w(kb)) I I Ea(s w) I X
W)
(`(k+1)1
, m(T)_1
(k+l)a
k-0 fk8
I d ]ds} 2]
6\
u B S w)) X r(k+ua
h1 S w)
u W kb
w)Idflds}2]
Jka
(7.36)
m(r)_1 ('(ktl)a [ I B6(0,0kaw)
C29 E E[ {
k-0 . ka
1-1
I
X II h AS' w) ,
< c29m(T)
+f S
m(T)-1
E[ k-0 E{f 1-1
+f'
f(k+l)a I g6( ,w) I dc]ds} 2]
(k+1)1 ka
[(I Ba(0, Okaw) ka+1)3
ka 0
ka E,(, , ekaw)dti
I N17, 0,,,w) I d1l) I -aa(s, w) I c29mm(T)2
r-1
J
w) I dal ds} 2]
E[ { I Bfa(0, w) I f a I Ea(s, w) I ds f a0 I Ea(s, w) I ds 0
COMPARISON AND APPROXIMATION THEOREMS
494
+ J'
I Ba(r1, w) I d j Jo I Ba(s, w) I ds f o
w) I d } zl
I
a
a
< c3om(fl2 E {E[ I Bl(0, w)12(f 0
1-1
+ E[(f
o
I BX77,
I Ea(s, W) I ds)2( f
W) I d3)2l
I
0
W) I dn)2(f 0 B'{s, w) 12ds)2( f o I Ba(s, w) 12ds)2]} I
S C30m(i )z !E {E[ I Ba(O, w) 1
671 ! 3E[( f
0
B6ls,
W)
I ds)3 (foIBSS'w)Id)3]z/3
x + E[(oI Ba'(,1, W) I dn)2(f 0 Ea(s, W) I ds)2(f o C31 n(6)262 (S(n(S)663)213
0
< C32n(8)48
as
I Ba(s,
W) I ds)2]}
+ n(6) 6&3)
S}0.
Finally we shall prove that (7.37)
E[ sup 11113(t) 12] -- 0 OSrgz
as S 10.
For this, we first note that (7.38)
3c,,(3, S) = 6cu(3, S) + (a - b)c,,(J - S, S)
where c*,(3, b) = E[
f6 E(s, w)[Ba3,w) - BB(s,w)]ds]/S. o
Hence
Iis(t) =
\
mc')-1
ck+1){B(s, w)(B'6((k + 1)a, w) - Ba(s, w)) u.(w(ka)) fka+a
-c,(S-5,5))ds
+ E m
(7.39)
k-0
4+6
1
ui(w(kS)) f kJ
{B`a(s, w)(BB((k + 1)6, w) - Ba(s, w)) - c *j(5, S)) ds
= ' J1(t) + J2(t).
APPROXIMATION THEOREMS
495
It is easy to see that '7.(W) = kFJO ut(w(kS))
J ka+6
a Ins, w)(B ((k + 1)b, w)
- Ba(s, w)) - c,(3 - 6, 8)} ds 3-6
_ k-0 u,(w(k3)) f0
-
{Ea(s, Bka+6w)(B6(3 - 8, eka+6w)
Ba(s, eka+Sw)) - c, (3 - (S, b)} ds
is an {..} -martingale where Y"' =
Consequently
I J)(t) 12] E[OStSsup
= E[0
-
c33E[I lm(T)-1IZJ
m(
c33
k-0 1 E[u,(w(k3))2(
-6 {h61s, eka+6w)(B(3
- (S, 04+0)
fa0
- B(s, Bka+6w)) - c11(8 - 6, 8)} ds)2] a-6
< c34m(T)E[(f
{E (s, w)(B (3 - 6, w) - Ba(s, w))
0
- c, (8 - 8, 8)} ds)2]
(7.40) 30-6
{Ba(s, w)(B'6(3 - 8, w) - Ba(s, w))} ds)2]
< c34m(T)E[(J a
c34m(T)E[(f0 I B]{s, w) I ds)2(f 0 I fts, w) I ds)2] < C35
-i- n((5)462
= c33n(8)38 --- 0
as
6 j 0.
Also E[ sup 14(t) 12] 0515 T
7.41)
m(t) - 1
< c36E[ sup ( O5t5T
k-0
k3+6
I f ka {Ba(s, w)(B'6((k + 1)8, w) - B(s, w))} ds 1)7, + m(T)2S2c(b, 8)2].
By (7.38) and Lemma 7.1, we have
COMPARISON AND APPROXIMATION THEOREMS
496
as 810
Scu(8,'5) --0 and consequently (7.42)
m(T)x8xc (b, 8)2
- c (a cu(b, 8i1
x
-- o
as
810.
On the other hand, m(:)-1
ka+6
E[05tST supk-0(E f ka S E[ sup m(t) < m(T)
m
Ea(s, w)(Ba(ka + b, w) - Ba(s, w))ds I)']
m(')-1
ka+a
E (f kJ
E 1 E[(J
-96'(s, w)(Ba(k8 + 8, w) - B](s, w))ds)x]
ka+a
h s, w)(Ba(k3 + 8, w) - BB(s, w))ds)x]
(7.43)
= m(T)2E[(f
0
B`a(s, w)(Ba(8, w) - Ba(s, w))ds)2]
< c38m(T)x {E[( f
a.b 0
(s, w)(Ba(8, w) - Ba(s,w))ds)x]
+ E[(Be(o, w) - B61(0, < c38m(T)x {E[( f o I
+ E[( f I Ee(s,
Ba`(s,
w))2(BB(3,
w) - B6J(8, w))2]}
W) I ds)2(f I Ba(s, W) I ds)x] o
w) I ds)2 (Ba(0, 0aw)S-
Ba(0, Oaw)
+ (wi(S) - w'(8))} 2]} < e39
1
n(6)xd2
(82 + 8xn(8)) -- 0
as 81 0.
This completes the proof.
Next we obtain an approximation theorem for solutions of stochastic differential equations. Such theorems have been discussed by several authors: McShane [115], Wong-Zakai [181], Stroock-Varadhan [158], Kunita [93], Nakao-Yamato [126], Malliavin [106] and Ikeda-NakaoYamato [52]. Our result covers most of these results in a unified way. Main idea of the proof is due to S. Nakao. Let Ao
b`(x)
8x`
and A. =
,
o (x)
X, ,
n = 1, 2, ... , r
APPROXIMATION THEOREMS
497
be vector fields on Rd such that 6n E CI(Rd) and b' Es CI(Rd) {Bo(t,w)} a>o be an approximation of Wiener process defined on the Wiener space (Wo, P). We assume that {B5(t,w)} a>o satisfies the Assumption 7.1. Consider the following equations *1 r
(7.44)
dXX(t,w) = Z A,(Xa(t, w))dB6n(t,w) + A0(X8(t,w))dt n-1
Xa(0,w)=x and
dX(t,w) _
A,,(X(t, w))odwn(t) + A0(X(t, w))dt
(7.45)
+ E snm[An,AmJ(X(t, w))dt lsn5m5r
X(0 , w) = X.
These are equivalent to the respective integral equations (7.44)'
I (t, w) - x' =
n
f
, (Xa(s,
o
w))B$(s, w)ds +
Jo
bt(Xo(S,
w))ds
and
X'(t, w) - x' (7.45)'
o (X(s, w))dwn(s) + n-1
0
J
r b'(X(s, w))ds 0
+ n.m-l ac_, E Cnm $ Qn (Va:Qm)(X(s, w))ds i = 1, 2, .. .
,
d.
0
For any given x E Rd, the solutions of (7.44) and (7.45) are unique; we shall denote them by Xa = (Xo(t,x,w)) and X = (X(t,x,w)). In the following discussion, the common initial value x is arbitrary but fixed and so we shall often suppress it.
Theorem 7.2. For every T > 0 we have (7.46)
lira E[ sup X(t, w) - X5(t, w) 1 Zj = 0. 05t r 810
Proof. *2 First we note by (7.44)' that *1 We use the same notations as in Chapter V. *ZIn the following, K, K KZ, ... are positive constants. S has the same meaning as in the proof of Theorem 7.1.
COMPARISON AND APPROXIMATION THEOREMS
498
(7.47)
w) -
1
w) I < K( nF f s
.
(u, w)1 du + (t - s))
for every 0 < s < t. We have for each i = 1, 2, ... , d that (7.48)
X(t, w) - X`(t, w) : = H1(t) + H2(t) + H3 + H4(t)
where
H1(t) =
n-
-
(7.49)
a (Xo(s, w))Bgs, w)ds
fr Cr]-(a)
n-1
ft
c,`,(X(s, w))dwn(s)
CAS-(a)
t
- En.m(-1 E Cnm f Ct]-(a) (6naaO'm)(X(s, w))ds, a-1 -m H2(t)
r
=
{Ja]
- fu]-'
(7.50)
°n(Xa(s, w))Ena(s, w)d
a.'(X(s, w))dw'(s)
a
E c,, -E m-1 a-1
f -E
H3 =
om06aan(X(s, a a
a
an(Xs(s, w))Bna(s, w)ds -
r
n-1
(7.51)
w))d },
16-(a)
n-1
0
f 6n(X(s, w))40(s) 0
d
E Cnm J a0 QnaaCm(X(S, w))ds
n.m-1 a-1
and
(7.52)
H4(t) = f
'00
b'(XX(s, w))ds - f 'o b1(X(s, w))ds.
It is clear that (7.53)
E[sp H4(t)12] < K1 f o E[ I Xo(s, w) - X(s, w)12lds
for every 0
H1(t) = H11(t) - En-1H"12(t) - H13(t)
APPROXIMATION THEOREMS
499
Then it is obvious that (7.54)
E[ sup I H13(t)1 2] <- K2b2. 0
By a similar estimate as for (7.28), we have (7.55)
BE uprI H11(t )12] <
K3(n(6)s6)112.
Also
H-2(t) = o (X([t]-(S), w))(w"(t) - w"([tl-(S))) t
+f
(cn(X(s, w)) - Q"(X([t ] (S), w)))dw"(s)
= Hi21(t) + H122(0By a similar estimate as for (7.32), we have (7.56)
BE sup I Hi21(t)1 2] < K4(n(3)5)1 i2 G-e
As for Hi22(t), E[ sup I H4122(t)1 2]
< E[ sup
sup { f
OSksm(T) OStsa
< kE)
E'[E[ sup
ka+'
(on(X(s,
w)) - Q (X([s]-(b), w)))dw"(s)} 2]
ka
{ f 0 (u.(X(s, y, w)) - 6n(y))dw"(s)l 2]
X> E'[E[ f0 (o"(X(s, y, w)) - Q"(y))2ds] I Y-X(ka.wl)l (7.57) < K, ffl(T) kN
< K6
/
E'[E[ f
O
X'(S, y, w) - Y' 1 gds] 1 y-X(ka.w')]
< K,(m(T) + 1) f o [s + s2]ds < K88.
Here we used the estimate
I X'(s, y, w) - y' 12 < K9[(' f" "-1
*
0
E' stands for the integration with respect to W.
y, w))dw"(b))2
1Y-X(ka.w')l
500
COMPARISON AND APPROXIMATION THEOREMS
+ (f
y, w))dd) 2],
o
and hence
E[ I XJ(s, y, w) - Y' 2] < K1o(s + s2).
Putting these estimate together, we conclude that as 310.
E[ sup I H1(t) 12] = o(1)
(7.58)
ggr
It is obvious that I H3 I S sup I H1(t) I and hence 0:r5 T
E[IH3I2] = o(l)
(7.59)
6 10.
as
Finally we shall estimate H2(t). We have w)ds
fka
_ on(X6(k3, w))[Ba((k+1)b, w) - Bg(k3, w)]
+ Ed
rck+lfa
$=1 J ka
0vn(X5(s, w))[E ofl(Xo(s, w))EJa(s, w) + bfl(Xo(s, w))] J®1
X [Ba((k+1)3, w) - Ba(s, w)]ds d
= J1(k) + E Jjx(k) fe1 Also
J1(k) = Qn(X6(ka - 6, w))(w"((k+1)3) - w"(k3)) + [o';(X6(k3,w)) - on(Xa(k3 - 6, w))](Ba((k+1)S, w) - B,-5(k3, w))
+ 6n(Xe(k3 - 6, w))(Ba((k+1)b, w) - w"((k+1)3)) + Q'(X5(kJ - b, w))(w"(kb) - Ba(k8, w)) = J11(k) + J12(k) + J13(k) + J14(k)-
Hence, writing H2(t) = E H-2(t), n-1
Hi(t) = I1(t) + IZ(t) + 13(t) + 14(t) + I5(t), {51} is the coefficient of the vector field A0 + E ;_[AA, AA,]. ALAI
APPROXIMATION THEOREMS
501
where
f
[cr(a)
m&)-1
I1(t) =
(7.60)
op(X(s, w))dw"(s)
a
m(r)-1
(7.61)
II(t) = E
(7.62)
15(t) = E { k-1 E Jj(k) --
M(r)-1
d fl-1
f [fl-WEr
w))ds}.
1-1
a
First we write 11(t) as
= I1(t)
[an(X5([sl'(a) - (5, w)) - an(X(s, w))]dw"(s)
and then use a martingale inequality to obtain E[ sup I I1(t) 12] < K1 1E[ I I1(t1) 12]
S K12 (7.63)
f
uiJ-a)
E[ I X5([s]-(3) - 8, w) - X(s, w) 12]ds
a
K13 { f 'o E[ I X6(s, w) - X(s, w)
+f S K1a{fo
w-(a) E[ I
12]d'
yX3([s]-(a) - 8, w) - X3(s, w) 12lds}
a
E[IXo(s,w)-X(s,w)I2]ds
[s]t(J)
+ f T E[1-IE (f is]-W-6
Al6(t, W) I d 2 + (b + 3)2]ds}
a
K1s { f E[ I Xj(s, w) - X(s, w) 12]ds + n((5)Za}
for every
0 S t1 S T.
In the case of I2(t), E[ sup I I2(t) 12] 0sC r
C
m(77-1
1
E[ E (an(Xa(kJ, w)) - Qn1(Xa(k8 Y - 6, w)))2 k-1
502
COMPARISON AND APPROXIMATION THEOREMS
x (7.64)
m(E
1)3, w)
k-1 m(T
< K16 {m(T)
-
Ba(k3, w))2]
1
- E[ I Xa(k3, w) - X6(k3 - 8, w) 14]
k-1
m(79-1
x m(T) E E[ I Ba((k+ 1)3, w) - Ba(k3, w) 14]] 1 J2 k-1
S K17(m(T)2U2m(T)23211)2 1
C K1S(n(a))-1 _.. 0
as
610.
Similarly as for (7.34), we can show that (7.65)
as 610
E[ sup 113(t) 12] < K19(n(!S))-' --- 0 05tST
and (7.66)
E[ stpTI I4(t) 12] <- K20(n(6))-'
as a j 0.
0
Finally we shall consider I5(t). Write I5(t) as
15(t) =g-1 E "'(0, where m(0-1
m(t)-1
E f k-1
S(t) - LIk-1f 12(k) r-1 m(t)-1
E3I + m(t)-1 r
(k+l)a
(k+1)a r
i
ka
E J-1
c1"(QSaQQn)(X(s, w))ds
[(a9a8Qn1(Xa(s, w))
w))]Ea(s, w)[B"l((k+1)b, w) - Bl(s, w)]ds
f
(k+1)a
(s, w))ds
ka
m"-' r(k+1)d
+EE
(asaso")(Xa(ka, w))[B(s, w)(Ba((k+1)8, w)
J ka
- Ba(s, w)) - c1"(b, S)]ds r m(:)-1
(k+l)a
+ E kr) Jfka r
[(apon)(xa(kb, w)) - (69afO"`)(X(s, w))]ds c1"
m(r) -I
+EJ-1E "(Qjfasa")(Xo(ka, w))(cj.(a, 6) - c1") k-1
APPROXIMATION THEOREMS r
r
503 r
r
= E 151(t) + 152(t) + F. I'53(t) + E I54(t) + J-1
155(t)
1-1
J-1
It is obvious that (7.67)
E[ sup 11-1.55(t)12]
As for 54(t), E[ sup I I54(t)1 2l 0SrSr,
< E[( o I (6iae6 (X'a([sl -(b), w))
S TE[ (7.68)
Ja
-
w)) I ds)Zki
w)) -
I
(6jfla,66n)(X(S w)) 12d$ jc12
<_ K22 f E[ I X(s, w) -- Xo([s]-(b), w) 12lds 0
< K23[ f"0E[ I X(s, w) - Xa(s, w)121ds + f 0 E[ I X5(s, w) - Xo([s]-(3), w) 12]ds]
< K24[
O
E[ I X(s, w) - XX(s, w) 12lds + n(l )26]
for 0
m(r)-I
(k+1)d
E[OSrST sup (k-]E f kd (7.69)
< K25E[ {
(k}1)d LL
w)) I ds f
I
m(T)-1 _ (k+i)a
f
k-1
ka
m(T)-1
K26b2m(T)
ES
I Ea(s, w) I ds} 2] (k+1)3
E[(f ka
I B$(s, w) I ds)2]
< K26V2m(T)2n(V)2S
< K2,n(8)2S -- 0 We estimate I51(t) by
as 610.
kd
I E (s, w) I ds)2]
504
COMPARISON AND APPROXIMATION THEOREMS
E[ sup J r5,(t) 12] 0
d
m(T)-1
l-i
k-1
< K,,, E E[ {
(k+1)a
f
I Xa(s, w) - Xa(k6, w) I I Bas, )+') I ds
ka
r(k+i)a
x m(T)-i
J
(k+l)S
(k+1)d
d
Kz9E[ k-1 { E (E f ka 1-1
I B`a(s, w) I ds + 6)f
f
x
(7.70)
< K30
1-1
m(T)
m(T
k-1
1
{E[(f
B"a(s, W) I ds} 23
(k+i)a
(k+1)a
I Ba(s, w)
(f
I Bs(s, w) I ds
ka
xa+1)a I
)2(fka
ka
x
+ 32E[(f
I Ba(s, w) I cuss} 2]
kJ
I Ba(s, W) I dy)2
(k+ua I B$(3, W) I ds)2]
kJ
+1)a
ka+ua
I B'a (s,
w)
Ids)'( f ka
IBa(s, w) I ds)2]}
< K,im(T)z(n(b)6E3 + n(8)684)
< K,,n(b)46 -. 0
as
61 0.
By repeating the same proof as for (7.37) we see that z
(7.71)
+ n(5)-1) --- 0
E[ upTI I53(t)121 S K33 (n(6)36 +
as 610.
Now combining (7.67), (7.68), (7.69), (7.70) and (7.71), we obtain the estimate (7.72)
01
E[ sup i I J5(t) 12] < K34 f E[ I X(s, w) - X5(s, w) 12]ds + 0(1)
as
610,
where o(1) is uniform in ti E [0, T]. By (7.63), (7.64), (7.65), (7.66) and (7.72), we have (7.73)
E[ sup I H2(t)12] <_ K35 f o E[ I X(s, W) - Xa(s, w) 12]ds + 0(1),
where again o(1) is uniform in t1E[0, T]. Finally, by (7.53), (7.58), (7.59) and (7.73), we have
APPROXIMATION THEOREMS
505
E[ sup I XX(t, w) - X(t, w) I'] 0st
(7.74)
1
S K36 f E[ I XS(s, w) - X(s, w) I Z]ds + o(1).
As before o(1) is uniform in t, E [0, T]. (7.46) then follows from (7.74) by a standard argument. This completes the proof.
Remark 7.1. If the vector fields A,, A2, ... , A, are commutative, i.e., [A., Am] = 0 for n, m = 1, 2, ... , r, then the limiting process of (X5(t, w)) does not depend on the choice of the approximation {BB(t, w)} as Theorem 7.2 shows. But this is seen more directly in this case if we notice that X(., w) is a continuous functional on Wo by the result of Doss (Example 2.2 of Chapter III, Section 2). If A,, AZ, ... , A, are not commutative, however, X(., w) is not continuous on Wo and the limiting processes of w) are controlled by {sj} . For piecewise linear approximations or approximations by mollifiers, we have sty = 0 and so the limiting process
is the solution of the stochastic differential equation corresponding to vector fields A,, A2i ... , A, and to A0 as was defined in Chapter V, Section 1. Remark 7.2. In the proof of Theorem 7.2, we have actually shown that (7.75)
lim sup E[ sup I Xa(t, x, w) - X(t, x, w)1 2] = 0. &10 xeRd
05t5T
In the following, we shall restrict ourselves to a class of piecewise linear approximations. It is an open question whether our results below can also be obtained for more general class of approximations. We shall assume that the coefficients an and b' of vector fields A0, A,, ... , A, are C°° and bounded together with derivatives of all orders. In the following discussion, the terms involving the coefficients b` do not cause any trouble and so we shall assume, just for simplicity of notations, that A0 = 0. Thus we consider the stochastic differential equation (7.76)
dX(t) = a(X(t))odw(t). *
Also, setting for each n = 1, 2,
... ,
wn(t)-nr(lnt)w(n)+(t- n)w(7n 1)1 * We shall use the matrix notations below; in particular a = (aa.) and w = (wt).
COMPARISON AND APPROXIMATION THEOREMS
506
jn < t < j+1 n
if
we consider the following ordinary differential equation
d "(t) =
(7.77)
d ° (t).
The solutions of (7.76) and (7.77) with the initial value xERd are denoted by X(t,x,w) and X"(t,x,w), respectively.
Lemma 7.2. Let T and N be arbitrary given positive constants. Then we have
sup sup E[ sup II D"X"(s, x, w)II'] < oo
(7.78)
IxISN
,.
05sST
for every p > 2 and multi-index a.
Proof. First we consider the case a = (0, 0, ... , 0), i.e., DOX. = X. We denote
dkw=w(kn1)-w(nk )' Then, if
k
X"(t, x, w) - X. (_t_ , x, w)
k/n
a(X"(s, x, w))dkw ds n
=Q(X"(n , x, w))dkwn(t
- n)
+ Jki" [ais, x, w)) - a (X°( and hence
X"(t, x, w) -- x
=%' x' w))dkw k-0 a(`Y"(n '
n , x, w))]ds dkw n
APPROXIMATION THEOREMS
+
(k+1)/n
(nr]-1
k-0 Ski,, k1.
507
[a(X(s,x'w))-o(X,1(n'x,W))]dsdkwn
+ a(X,11[nt)' x, wd(nt]wt - [n] )n r
J'nt]/n[a(Xr(S. x, w)) - o (X.1 ni' x, + =Il(t)+I2(t)+I3(t)+I4(t)
w))] ds AE t]w n
By Theorem III-3.1, E[ sup III1(t)II'] 05t5T
KIE1(Ent]
<
w))II2)''2]n
I
Ila(Xn(n f x,
<
D/z *1
K2nD/ln-y/2
=K2
E[ sup I I2(t) I') 05t5T
<
T°_'ln°-'E[05t5supT 1nt1 -1 k.E0
f(k/n k+l)/n
IIJ
-or < T °-lnP-1E
[t
[ci(Xn(s, x> w)) (X,(n,x,w))]dsdkwIl']n'*2
1 IIJ /`(k+l)/n
k-0
k/n
xf w))
L
-Q(Xn(n'x,w))]dsdkwll°]n' (nT]-1 f(k+1)/n
< T°-'n°-'n'n-(°-')k-0 E
J
k/n
- a (X,1(n < K3n' k-31J
k/n
E II Q(X,(s, x, w))
' x, w))] dkwlV]ds
1)/n E[II Xn(sf xf w) - X. (n , x, w)II° IldkwIl] dr
*1 K,, K2, ... are positive constants independent of n and x. k k *2 Generally, if A,,> A 2, ... , Ak are matrices of the same type IIE A,II 5 (E IIA,IIY
'
k
5 ka-1 E IIArII'. ,-1
508
COMPARISON AND APPROXIMATION THEOREMS [nn-1 (k+1)/n
Ef
K3n2p
s
E[II f k/n o(Xn(u, x, w))dkwdulI'II dkwII ']ds
k/n
[n77-1 f(k+1)/n
< K3n2p E
J k/n
[nTf)t-1
< K4n2p
ft
E[(J k/n
Lf
II d kwII du)1II dkwIl p]ds p
(k+1)/n I
k/n
S
n) ds EQldkwll2')
< KSn2pn n- (p+1) n-p
=K5
<E[ sup
Ilo'(Xn([n],
x, w))d[nt)wllp]
< K6E[ sup IId[nt)wlip] o
< K6E[ max Ildkwllp] 0
[nfl
< K6 E E[II dkwllp] k.-0
< K,([nT] + 1)n-p/2
< E[ sup
t
0
I I (6(Xn(S, x, W))
- o(X ([n t], x, < E[ sup
W)))d[nt)wlIpdS]npn-(p-1)
([nt)+1)ln
0<,
llo'(Xn(S, x, w))
6 (XX([nti, x, W))IIPII4[nt)wllpds] n
<
nfl
K9n[E
k-0 E
Kton1-p
(k+1)/n
[f k1n
II Xn(S, x, w) - xn(n , x, w) II°IldkwIIpdS]
APPROXIMATION THEOREMS
509
K11(1 + IxV').
E[ sup II Xx(t, x, w)II'] 05t5T
Thus (7.78) is proved for a = (0, 0, ... , 0). Next, we consider the case of the first order derivatives. Setting Y.01 x, w) = (8-J X.*' x, w))
and Da = (axJ 011.) '
we have
Yn(t, x, w) = I +
Do (X (s, x, w))YY(s, x, w)wn(s)ds
Hence
Yn(t,x,w)-I [n
I
k-00
Do(X"(nw)) (k+t)/n
Cnt]-1
+ A J k/n
n
w)dkw
IDc(X,(s, x, w))
-Dv(Xn(n,x,w))]dsYn(n,x,wAkwn (k+1)/n
Cat]-1
+
k-0
+ DQ(Xn ( ( J
Dcr(X,(s,
$ k/n [nt]
[cJis+
w))(Y,(s, x, w)
x,
, x, w)) Yn(
[nt]
, x, w)A[ntlw n(t
w))dsdkwn
)
, x, w))
[at]/n
- DQ(Xn ([nt]n +
- Yn(n ' x, - [nt]
x, w))] ds y n
([nt] x, w)Akw n n
ft
Cnt]/n
Dc(X,(s, x, w))(Y,(s, x, w) - Y"( nt]' n x,
w))ds Akw n
= J1(t) + J2(t) + J3(t) + J4(t) + Js(t) + J6(t ) *
We use the following notations: for A =
B = (bj) and C
ABC = (dp
where dJ = E ak ab;ca. It is easy to see II ABC II 5 IIA ll IIBU II C p where 11411 = (E I ai,a 12)1/2 k,a
1,,,a
COMPARISON AND APPROXIMATION THEOREMS
510
Let t1 E [0, T]. By Theorem 111-3. 1,
I -l
E[ sup (IJI(t)II°] ostst,
r
K12E[(EtkE
(
(
n'
I ljDa lX"1
(7.80)
<
r((tntl]-I
< K l,
Y.(n
w)Ilz)°/z] n-D/z
x,
7p" n-1/2 x, w)11)
K13E[(
K,an' E
1
x, w)111211
E(IIYY(kn,x,w)Il°)
E( suP II Yn(s, x, w)il°)dt. o
J0
5ssr
As for J2(t), E[ sup 0 tst,
IIJ2(t)II °]
E suP II O
,
[ntl-1 fk/n (k+1)/n
Stst' k-0 J
[Dox(s, x, w))
- Do(Xn(n ' x, w))] dS Yn(n [nt]-I
< TD-1n2y-1E sup f 11f 0Sts[, k-0
, x, w)dkwl [D]n'
(k+l)/n
IDQ(X°(s, x, w))
k/n
DQ(X4(k , x, w))]1 ds Yn(n x, w)dkwll°] I [nt1]-I
S T D-1n2D-I E
-
E[IIJ
Do'(Xn(nk
(k+I)/n
[DQ(Xn(s, x, w))
k/n
1
, x,w))]dS
< T°-ln2D-In (D-I) rntl]7]-1 E r(ktl)/n E[II k-0 J k/n
- Do (Xn(n , x, w))}
(7.81)
Dal]-1 /`(k+l)/n
k/n
Yn(kn'x,w)dkwll°]
IDa(Xn(s, x, w))
Yn(n ' x, w)d kwll D] k
E[IIX1(s,x,w)-X4(n ,x,w)IID
x l[Yn(n , x, w)II DIIdkwIID]ds (ktI)/n
[nt1]-1
- Kisn2D
o
f
kin
s
E[IIJ k/n a(Xn(u, x, w)) dkw dull'
APPROXIMATION THEOREMS
Yn(L, x, w) II°Ildkwll°]ds
x 11
(,t1)-l (k+1)/n < K16n2D
f
3q
k ° ds E [II Y ( k
(s - -) n
r(k+i)/n
[nt17-1
= Kl6n2D
k/n
(s
J k/n
511
w) II° Ildkwll2° n , x,
_ n) k yds E[IIYn(nk , x, w)ll° E[Ildkwll2p)
[nt1J-1
<
K,7n2yn-(y+l)n y
E E k-0
IIYn(n , x, w)II°
[nt1)-1
= K17n-' E E[IlY.(n, x, w)II° k-0
S K17 f
E[OSsSt sup II YY(s, x, w)lI°)dt.
0
For J,(t) E[ sup IIJ3(t)II°I 05t5t,
r
= E I sup I I L0St ,tj
[nr7-i
f k-0
(k+1)/n k/n
Da(XX(s, x, w)) Yn(s, x, w)
- Yn(n, X, w)1dsdkw rally]
(7.82)
< TI-'n°
[n'13-I 57,
k-0
f (k+l)/n E II Da('n(s, x, w#1111 US, x, w) k/n
- Yn(n , x, w)II°Ildkwll°]ds [rat
K18 ny
If
I (`(k+l)/n J k/n
k-0
(, E US, X, W) - Y. (n - , x, w)lI°II dkwI y]l A [ll
kn <s
Da(Xn(u, x, w)) Y,,(u, x, w)dkw dull2n2
k/n
< n9lldkwli2 f'
II Da(Xn(u, x, w))ll2du
k/n
S- kIf )f
k/n
f
=
II YY(u, x, w)112du k/n
ll Yn(u, x, w)ll2du
COMPARISON AND APPROXIMATION THEOREMS
512
<
X,
)V)112
+nlldkwll2f,
k/,,
From this inequality, we obtain the following estimate:
IIUs,x,w)- Y.
(k
,x,w)I12
x, w)IIZ exp{K2onlldkw112(s
< K2olldkwll2ll
-
and consequently, Y.
(k
,x,w)II°
(7.83) < K21JI4kwl1°II Y» (n , x, w)II° exp {K2211dkw112}.
Substituting (7.83) into (7.82) yields E[ sup I1J3(t)II1] 02 tsr,
S K23n°n-'
tnri7EII
(7.84)
x,
w) I1°
x E[Ildkw112° exp {KZZI1dkw112}] W13-1
S K24 n E E II
x, w) II°
k-0
< K24f 0
0stst
w)II°]dt
since E[Ildkwll2' exp {Kzzlldkw112} ]
-fR,(2acn)-.izIxl2pexp{K22IxI2n
<
K25n-'
J'4(t) is estimated as follows: E[ sup I14(t)JID] 0stst,
Ix12}dx
for sufficiently large n.
k )j
513
APPROXIMATION THEOREMS
= E[
x, w)) Yn([ntl x, w)a(t
S K26Er sup
[nt])dc.t3wllpi
x, w)II'IldCnt7wIl']
t:5t11IY.([n]
(7.85)
-
,x,w)II'Ildkwll']
k-0 Cnt13
EE k-o
SK27n-pie
J"
< K2$(n1-p/2
0
llYn(n,x,w)II'
E[ sup II Yn(S, x, w)IIldt o5.,:!gt
+n-p/2EElIYn([n1l, x,w)II°]) In the case of J5(t), E[ sup IIJS(t)II'] 05tst1
< K29nE
sup E f[nt]/n II XX(S, x, w) - X.([nt] n
05t5t,
(7.86)
x W II °dS
x HY.([nl, x, w)II'Ildxwll] (nt1
E(IIYn()1
,x,w II'
< K30(n'-p f" E[ sup II Yn(s, x, w)II']dt 0
05.,5:
x, w)IIp,)
+ by the same estimate as for J&). As for J,(t), E[ sup IIJ6(t)II'] Ostst,
Cnt1J
(k+1)/n
f".
E[IIYn(S,x,w)-Yn(n,x,w)II'Ildkwllp ds
(nt17
< K32n' Et E[II Yn(n x, w)II' < K32(n'-p f" E[ sup1 II Ya(S, x, w)II']dt 0
05,5
+ n-PEI IIYn([n1], x, w)II'])
COMPARISON AND APPROXIMATION THEOREMS
514
by the same estimate as for J3(t). Combining (7.80), (7.81), (7.84), (7.85), (7.86) and (7.87) together, it is easy to conclude that E[stp,11YY(t, x, w)IIP]
K33(1 + J o E sup IIYY(s, x, w)II'}dt).
By a standard truncation argument, we can easily deduce from this that (7.88)
K33e'33 T.
E[ sup II YY(t, x, w)II'] 0StST
a such that I a I = 1.*1 Now we proceed to the case of I a I = 2. First we note that if k/n < t S (k + 1)/n, we have Thus (7.78) is proved for every
(7.89)
x, w) - X.(-L, x, w) II'] < K34n-D12
E[II XX(t,
and, also by (7.83) and (7.88), E[II Yn(t, x,
(7.90)
w) - Yn(n , x, w)II'J < K35n12,
Set
Y51y?2(t, x, w) =
a2axj2
axI
X '.(t, x, w).
Then Yi,12(t, x, w)
= k-1 E a-1 EJ
t
oa(XX(s, x, w))kYY; i''z(S, x, w)wn(s)ds
t
+ k.1- a-1 J 0 o (Xn(s, x, w))$,Yn(s, x, w)%Yn(s, x,
w)12wa(s)ds.*2
We denote the second term on the right by an(t, x, w). Then, denoting as d
r
Cnt7-1
(m+I)/n
/ x, w) = E a,(t, k.l I E [ E m/n
t
+ Cnt]/n
*11t was actually proved that sup can be strengthened to sup. : IxlaN
012
.2 aa(x) j'=
a .'(x) zL
and
dQ (x)kt =
axkax<
aQ(x)-
APPROXIMATION THEOREMS d
515
r
E (HI(t; k,1, a) + Hz(t; k,1, a)), k.1-1 a-1 HI(t, k, 1, a)
+
Cnt]-I
f M-0
Iaa\Xn(n ,
(m+l)/, m/n
X ,W))kl Yn(
n n1 ! ln(n'X°W)) [o.;(x(s,x,W))tk! -Qa V(_ ki
X Yn( 7n ,X,
n
(m+i)/n
Cntt)-1
+
m-0
f
n' X, W)h Yn( n' X, W)J2dmwa
t
n1 W)i,k Yn(-,X,W dmWan n
)J2
m
r
oa(Xn(S, x, W))k'l Yn(S, X, w) , L
m/n
k
YA(n , x, w)1 ] ds 1
I
X Y. n , x, w) dmwan J2
+
Cnt]-1
m-0
f
(m+1)/n
o (Xn(S, X, W)),"'Y,(S2 X1 WA [Y-(S' X1 W)12
m/n
- Yn(n , x, w
)
l Jds dmwan
= KI(t) + K2(t) + K3(t) + K4(t). Then, by Theorem III-3.1 and (7.88), 2l
w)I14)°/J
E[ sup KI(t) I'] < K36E[(nL (nE 1 OS T m-0
II Yn(m , x,
n
Also, by (7.89) with p replaced by 2p, E[ sup K2(t) I'] 0!5t: 5T CnT]-1
KS8n'-I
F
- Xn(n , x,
nPn-(y-1)
1
E.*
J m/n
w)I!
20]112ds
II Xn(s, x, w)
E [II n(n , x, w) 1141]"' E[I dmwa 1.1]1/4
K39 < 00 . Using (7.90), similar estimates hold for K3(t) and K4(t). Thus we obtain
E[ sup HI(t; k, 1, a) I'] < K40 < oo . 05t5T
More easily we can obtain
E[ sup H2(t; k, 1, a) I o] < K41 < oo . OScS T
COMPARISON AND APPROXIMATION THEOREMS
516
Therefore sup E[ sup IIa.(t, x, w)II'] <_ K42 < 00 . ,.x
05t5T
Using this, sup E[ sup I Y5,),(t, x, w) 1,P] < 00 n,x
05t5T
can be proved as in the case of YY(t, x, w). By continuing this process step by step we can complete the proof of Lemma 7.2. Now (X(t, x, w), Y(t, x, w), ...) is the solution of stochastic differential equation dX(t) = cr(X(t))odw(t)
dY(t) = Dai(X(t))Y(t)odw(t) X(O) = x
Y(O) = I
The coefficients of this equation are not bounded and we cannot apply Theorem 7.2 directly. But Lemma 7.2 enables us to apply a standard truncation argument as in the proof of Lemma 2.1 of Chapter V and obtain the following: for every T > 0 and N > 0, (7.91)
lim supE[sup I DaXX(t, x, w) - DaX(t, x, w) I D] = 0
n-m IxISN
O5r5T
for every p Z 2 and multi-index a. Then, as in the proof of Proposition V-2.2, we can obtain the following result. Theorem 7.3. For every T > 0 and N > 0, (7.92)
lim E[ sup sup I DaX (t, x, w) - DaX(t, x, w) I p] = 0 r»m
05t5T IxJSN
for every p > 2 and multi-index a. We have assumed that coefficients are bounded together with all of their derivatives. Now we assume only that coefficients are C°° and that the
stochastic differential equation and approximating ordinary differential
SUPPORT OF DIFFUSION PROCESSES
517
equations possess global solutions for each fixed initial value. By a standard truncation argument, we can easily deduce from (7.92) the following: Corollary. There exists a subsequence {nk} such that, with probability
one, x, w) --- DaX(t, x, w)
as k - oo compact uniformly in (t, x) for every multi-index a. *1
8. The support of diffusion processes k = 1, 2, ... , r, be bounded smooth functions on Rd with bounded derivatives*Z and consider the Let o k(x) and b`(x), i = 1, 2, ... , d,
stochastic differential equation
dX`(t) _ E ug(X(t))odBk(t) + b`(X(t))dt (8.1)
i=1,2,...,d.
X(0)=x,
Let P. be the probability law of the solution X= {X(t)}. Then as we saw in Chapter IV, the system {Px} of probabilities on Wd constitutes the unique diffusion measure generated by the operator A, where
(8.3)
all(x) = E ak(x)ok(x) k-1
and
(8.4)
P(x) = b'(x) + 2 E
k-1
(axj ak(x))ok(x)
Now the path space Wd is a Frechet space with the metric defined in Chapter IV, Section 1, or equivalently, the topology of Wd is defined by *1 The subsequence {nk] can be chosen from any given subsequence of 1, 2, n,
..
: To be precise, the following discussions will be valid if the second order derivatives of a are uniformly continuous.
,
... ,
C;(Rd), bEC;(Rd) and
COMPARISON AND APPROXIMATION THEOREMS
518
the system of seminorms {II.11 T, T > 0} where (8.5)
IlwII
T = max
I w(t) I
OSrs T
for w E Wd.
The purpose of this section is to describe the topological support S"(Px) of the measure Px, i.e., the smallest closed subset of Wd that carries probability 1. In order to do this we need to introduce the following subclasses of the space Wo = {w e W'; w(0) = 0) ;
S'C.S',C Wo where
.', = 10 E Wo ; t i- c(t) is piecewise smooth) and
So = { C Wo; t i-- ¢(t) is smooth). For 0 E.S7, and x ERd, we obtain a d-dimensional curve
=(x, ¢)
4)) by solving of the ordinary differential equation
d ,' = E a (,)Ok(t)dt +
(8
o = x,
i=1,2,...,d.
We define the subclasses Sox and So; of Wd by (8.7)
and
So; = {c(x, c);
(8.8)
e .So}.
It is easy to see that the closures in Wd of .Sox and .9' coincide; i.e., Sox = .So Now we can state the main theorem due to Stroock and Varadhan [158].
Theorem 8.1. .7(Px) = Spx dt
for every x E Rd.
SUPPORT OF DIFFUSION PROCESSES
519
Proof. First we shall prove the inclusion .9'(P=) e ,7 x. For each n = 1, 2, ... and t > 0, we set t = [2°t]/2° and 1,, = ([2"t] + 1)/2". Let {B(t)} be a given r-dimensional Brownian motion with B(O) = 0 and define {B (t)} by
(t" -
t)
(tn - ta)
B(t,) +
(t -
B(t,,)
U. - tn)
E.5" . By Theorem 7.2, we can conThen B. EE S', and hence (x, clude that (x, X in Wd in probability. Hence Pz * P,, as n - 00 where Px is the probability law of (x, B.). Thus
Px(5';) > lim
=1
and consequently .So-x-= So; D V (Px). The converse inclusion .'(P.) ox will follow immediately from the following theorem. We consider the equation (8.1) on the Wiener space (Wo, PR') with respect to the canonical realization of the Wiener process; its solution is denoted by Xx=(X(t, w)).
Theorem 8.2. For every 0 e:,91, T > 0 and e > 0,
PW(IIXI-
(8.9)
(x,0)111'<EIIIw-0IIT<6) - 1
as (510.
Remark 8.1. It is well known that PR'(Ilw - oII T < 6) > 0 for every
0 ESo, T > 0 and b > 0, i.e., So(Pw) = Wo. Indeed, by the result of Chapter IV, Section 4,
Pw(IIw-oily<6)=Epw(M(w):llwlIT<6)>0 where
M(w) = exp[- E
f
o
k(s)dwk(s) - 2
Jo
I (s) 12 &J.
To prove Theorem 8.2 we first introduce several lemmas.
Lemma 8.1. There exist positive constants c, and c2 such that (8.10)
P '(IIwIIT < E)
c,
exp(- Z)
as e 10.
520
COMPARISON AND APPROXIMATION THEOREMS
Proof. Let P. be the r-dimensional Wiener measure starting at xE R'; so, in particular, Po = Ps'. Let D= {x E Rr; I x I < 1} and set o(w) = inf {t; w(t)
for w E F V.
D}
Then u(t, x)=Exf f(w(t))1,,(.)>j, x s D, t > 0, is the solution of the initial value problem in D,
at = 2 du ulaD=0
uit-o =f Consequently
u(t, x) = E e-xn' 75n(x) f 0n(Y)f(y)dy, n-1
D
where 0 < 11 < 2 < 2, < .. are eigenvalues and
{O (x)}
are cor-
responding eigenfunctions of the eigenvalue problem 2 do -{- 1o = 0
in D,
018D=0-
In particular, Pw(IIwfl, <
e) = PF'(max I w(t) I < e) = Pw( max I ew(t/e2) I < e)
PW(omrax sI
OstsT
0! tST
w(t) I < 1) = PW(a(w) > T/eZ)
e 'finT/62 95n(O) f D
3 (y)dy,
and consequently
P1(IIwIIT<e)_e12 01(0) f DOY(y)dy. Therefore (8.10) is proved with
SUPPORT OF DIFFUSION PROCESSES
C, _ 0,(0) f O,(x)dx
and
521
c2 = .1,T.
Lemma 8.2. Set (8.11)
e(t) =
2
f [w'(s)dwf(s) - w'(s)dw'(s)] o
for Q=1,2,. .
..,r.
Then we have (8.12)
lira sup P`r[ Ilillfll T > MS I IIwll T < 01= 0. Mt - o
1
Proof. Let i k j be fixed and set a(t)
4
f
[(w`)2(s) + - (wJ)2(s)]ds. o
We have remarked in Section 6 that B(t): = ij'J(a-'(t)) is a Brownian motion independent of {(w')2(t)+(wi)2(t)} (and hence, independent of the radial process { I w(t) I } and, in particular, of I I wI i T). Then P9 II11'II T > M8 I IIwI1 T < 0] = Pw[ IIB(a(t))II T > M8 I
< PR'[ max I B(s) I > M8] = = P'[ max I B(s) I > M], 05s5 T/4
0!-..j- g62 T/4
and this proves (8.12).
Corollary. Let (8.13)
'J (t) = f 'o w'(s)°dw'(s),
i, J = 1, 2, ... , r.
Then for all i, j = 1, 2, ... , r, (8.14)
lim
M6I IIwjjT<(5)=0-
Aft m o
In particular, for every e > 0, (8.15)
I IIwIIT<6)-0
as
6 1 0.
IiwIIT < 0]
COMPARISON AND APPROXIMATION THEOREMS
522
Proof. Since
Z wi(t)w'(t) -- VV) = 1i'(t), (8.14) follows at once from (8.12). The following is a key lemma for the proof of Theorem 8.2.
Lemma 8.3. Letf(x): Rd -- R be bounded and uniformly continuous. Then for all a > 0 and i, j = 1, 2, ... , r, (8.16)
PY ( 11 f' f ( X ( s, 0
w))diii(s)I I T> e I I I wI I T< fi) ---- 0
as
8 10.
Proof. First we shall assume that f e Cb(Rd). Then by Ito's formula, f ' f(X(s))ddif(s)
=.f(X(t))i'(t) - f ' fi(X(S))6ic(X(S))bi'(S)dWk(S)
-f
0
f f (X(s))QJ(X(s))w'(s)ds *1
= h(t) + 12(t) -l- 13(t) + I4(t) . Clearly it is sufficient to show that
PW(Illi(t)IIT> S I IIWIIr<6)-0 as 6 10 for every a > 0 and i = 1, 2, 3, 4. This follows from (8.15) for i = I and i = 3 and it is obvious for i = 4. So we only consider 12(t). For simplicity we set ak(x) = f (x)ak(x). Then, by Ito's formula, 12(t) = fak(X(s))Sh'(s)dwk(s) 0 zz zz ( = ak(X(t))S''(t)W"(t) - f ak.i(X(S))o (X(S))bi'(S)wk(S)dwm(S) *2 0
*1 fi = 'f
. We also adopt the usual convention for summation.
*2 ak,,(x) = axi ak(x)
SUPPORT zOF DIFFUSION PROCESSES
523
- f 0 (Aak)(X(s))S`I(s)wk(s)ds
f ak(X(s))wk(s)ddl'(s) 0
- f al(X(s))wt(s)ds - f 0
I'(s)ak.l(X(s))Um(X(S))J'-ds
S 0
f wk(s)ak,,(X(s))ai(X(s))w`(s)ds 0
= J7(t) + J2(t) + J3(t) + J4(t) + J5(t) + J6(t) + J7(t)-
Again it is sufficient to show that
PW(IIJ(t)IIT>e I IIWIIT<6)--0 as 6 l 0 for all e > 0 and i = 1, 2, ... , 7. This is no problem for i = 1, 3, 5, 6, 7. By Theorem II-7.2', there exists a one-dimensional Brownian motion B(t) such that
J2(t) = B(a(t)), where lak, l,at"](X(s))wk(s)w"(s)ds.
a(t) = <J2, J2)t = f t
Now PW( IIJZ(t)IIT> e
I
IIwIIT < a)
< PW(ll l'ilT> M6 I
IIwIIT < J)
+ PW(IIJ2(t)IIT> e, IIVIIT :! MoI IIwIIT < (5).
By (8.14), we can choose M > 0 for any given r/ > 0 such that the first term on the right is less than 17 for all 1 Z S > 0. Clearly I T < M6 and Ilwll T < 6 imply that a(t) < a(T) < c304,* and hence PW(I1J2(0IIT> e, T MS, IIwIIT < 6) < P''( max I B(t) I > s) = PW( max I B(t) I > e/ 0st c,64
ogtsJ
* in the following, c3, c4, ... are positive constants independent of S.
b2)
COMPARISON AND APPROXIMATION THEOREMS
524
=2
f31ne a-2
z
z
exp [-
I
] dx < ca exp [- c,,
2
as
.
By Lemma 8.1,
P(IIWIIT<6)>_ c6exp[-c,a21.
Hence Y
P '(IIJ2(t)II TT > s, (ISf'II T :! Moo I IIwII T < 6) c10321
< c8 exp I - c9c2
--- 0
as
as
610.
Consequently,
1iiPR'(IIJ2(t)IIT>E1 IIwtIr<0) and since ?I is arbitrary, we obtain the desired conclusion. Next we consider J4(t).
J4(t) = -
f' ak(X(s))wk(s)w`(s)dwJ(s) 0
iJ f a,, (X(s)) Wk (s) dw o
= K1(t) + K2(t). Then it is obvious that for any a > 0
PW(IIK2(t)IIT>e I IIwIIT<6)-.0
as 6 10.
K1(t) is a martingale with
[ak(X(S))Wk(S)W1(s)]2dS.
Therefore, if IIwIIT < 0 then
IIWIIT <
(5)
<- c12 exp L- C1382
8a
C1452]J
--0
0. Thus (8.16) is proved for f E Cb(Rd). Now suppose that f is f uniuniformly continuous and choose f E=-CZ(Rd) such that f as (5
525
SUPPORT OF DIFFUSION PROCESSES
formly. Set
Y.(t) = f 'f(X(S))dc'J(S) - J 0fn(X(s)) d'J(s) 0 I('
J O (({
-fn)(X(S))w'($)dwf(s) + 1 1 (f -fn)(X(s)) O
.
Then, for any given e > 0, we have by the same argument as above that PW(IIYyIIT> e I IIwIIT < S)
< PR'(II f o (f -fn)(X(S))w'(S)dwf(S)II T > 2 I
< C15 eX p _
C16E2
lI
IIf
< Cls exp
CI6 E
11W117. < a) z
I-T,
-ffIIZa2J
and 0 < 6 < 1. Hence
for all large n
P"(IffXSd(S)IIT> s I
IIwIIT < b)
< P"'( II f ofn(X(S))d J(S)lI T >
+ PW(IIYnIIT>
2
I
2
1
IIIIT < a)
IIwIIT < a)
For given ij > 0, we can choose n such that the second term is less than
rl for all 0 < 6 < 1. Then letting 8 j 0 the first term tends to 0. Consequently ME 610
P'(II f
' 0
T>eI
IIwIIT <
_ 11, &)<
and since ?7 is arbitrary, the proof of (8.16) is now complete.
Proof of Theorem 8.2. First we shall prove (8.9) when 0(t) - 0. In
this case, , =
d, = X.
Now
0)) is the solution of
526
COMPARISON AND APPROXIMATION THEOREMS
xv) - t(t) = E f k-1
t 0
ak(X(s))°dwk(s) +
f' [bt(X(s)) - b`( t)lds. 0
We shall prove that t
(8.17)
PW( II f ak(X(s))°dWk(s)II T > E I II WII T < a) ~ 0 0
as 8 j 0 for every s > 0. By Ito's formula f' ae(X(s))°dwk(s) 0
= ak(X(t))Wk(t) -
= aX(t))wk(t) -
ft 0
Wc(s)°d(ak(X(S)))
t
f ak.,(X(s))o,,(X(s))wk(s)°dwm(s) f ak,(X(s))b`(X(s))wk(s)ds t
= 11(t) + 12(t) + I3(t)-
It is sufficient to show that for every e > 0 and i = 1, 2, 3,
P'(III1(t)II >E I I1wIIr
as 6 j0.
Only the case of I = 2 needs to be examined. Then I2(t)
_ - f a k.t(X(s))am(X(s))-d km(s) - f' ak.1(X(s))am(X(s))dSkm(S) 0
zz
0
2
0 ax°
[ak.la»,l (X(s))aa(X(s))wk(s)S*mds
J1(t) + J2(t),
and we can conclude that for every r > 0 and i = 1, 2
P`'(IIJJ(t)Ilr>a I IIWIIT
as 61 0 .
SUPPORT OF DIFFUSION PROCESSES
527
Indeed, it is obvious for i=2 and the case i=1 follows from Lemma 8.3. On the set {Wi 1If ' 6k(X(s))odWk(S)IIT < E},
we have
I IIWIIT<6)--0 as
S 10 for every e > 0. Now we consider the case of general 0 E=-,91. Set
M(w) = exp { k f ok(s)dwk(s) - 2 f
I
I (s)zds}
o
and define the measure p on W,' (T): = the restriction of Won the interval [0, T) by
dPW
M(w).
Then by Theorem IV-4.1, w(t) = w(t)-O(t) is an r-dimensional Brownian motion for P, and X(t) satisfies
X(t) = x + f
a(X(s))odw(s) + f o b(s, X(s))ds o
where b(s, x) = b(x)+o (x)¢(s). * Since c _
fl
0)) satisfies
dot = b(t, ,)dt 1o=x,
we can conclude from the above that for every e > 0, * Thus we need to consider the case that the drift b dependent on t and b (-= C'([0, oo) x Rd). All the above results remain valid in such a case with an obvious modification of the proof.
528
COMPARISON AND APPROXIMATION THEOREMS
(8.18)
P(IIX(t)-Srllr>EIIIwIIT<(5)---0
as
6 10.
Noting that
M(w) = exp {kE k(i )wk(T) - k -1 J 0
wk(S)Ok(S)ds
f0 2
I
I (s)Zds}
0, then M(w)
is continuous in w (i.e., if IIw - wll (8.18) implies that IIIw-6IIT<
610
5)
PW(IIX(t)-S,IIT <s,
lim 65 X
11w-oiIT
Ew(M:IIw-oilT<6) PW(llw-0IIT<(5)
---1.
Example 8.1. Let Lo, L1, ... , L, be vector fields on Rd whose coefficients (in the Euclidean coordinates) are bounded and smooth with bounded derivatives. Let Xx = (X(t, w)) be the solution of L1(X(t))odw'(t) + Lo(X(t))dt
dX(t) X(0) = X.
Let Px be the probability law of Xx. Theorem 8.1 implies that
s°(P.)_ {cx,0);0ESo . (x, 0) is the solution of the dynamical system L0(,).
dtt = ,E 0 = X.
It is known (cf. [95]) that
that Z(?71)
J
dt ?7o = x,
+ Lo(t7r)
c E So contains all curves 1, such
SUPPORT OF DIFFUSION PROCESSES
529
where Z is an element of the Lie algebra 2(L,, L2, ... , L,) generated by
L), L2, ... , L,. In particular, if £(L,, L2, ... , L,) has rank d at every 0 E So = WX: = {w E Wd; w(0) = x} and consepoint, quently .So (Px) = Wx
Example 8.2. (maximum principle). Let A be the differential operator defined by (8.2). A function defined in a domain D e Rd is said to be Asubharmonic in D if it is upper semicontinuous and Au > 0 (in a certain weak sense to be specified later). A classical maximum principle for the Laplacian asserts that any subharmonic function in D which attains its maximum in D must be a constant. If the operator A is degenerate, however, such a maximum principle no longer holds in general. For example,
ifd=2,
2
A = x; + axe ,
x = (xi, x2)
and D is any domain intersecting the x,-axis, the function u(x) defined by
if if
x2 > 0 x2 < 0
is a nonconstant A-subharmonic function which attains its maximum in D.
We are interested in the following problem. For a given domain D and a point x E D, we want to determine a (relatively) closed subset D(x) of D having the following properties: (8.19) for any A-subharmonic function u(y) in D such that u(x) = max yED(x)
u(y), it holds that u(y) = u(x) for every y e D(x), (8.20) D(x) is maximal with respect to the property (8.19): i.e., if z E D\D(x), then there exists an A-subharmonic function u(y) in D such that u(x) = max u(y) and u(z) < u(x). ye D(x)
It is clear that D(x) is uniquely determined if it exists. The support theorem 8.1 enables us to describe the set D(x). Before proceeding, however, we shall first make precise the notion of A-subharmonicity. For simplicity we shall restrict ourselves to locally bounded functions. Let P. be the probability law of the solution of (8.1). Let D be a domain in Rd and be a compact exhaustion of D: D. are bounded subdomains of
COMPARISON AND APPROXIMATION THEOREMS
530
D such that D. c
and U D. =D. Let o =
inf {t; w(t)
,wEWd. Definition 8.1. A function u(x) defined in D is said to be A-subharmonic in D if (i) it is locally bounded and upper semicontinuous and (ii) for each n = 1, 2, ... and x E D, t u(x) is a u(w(t A Pz submartingale. If u is in C2(D), then u(w(t A a ))-u(x) = a martingale + ^ (Au)(w(s))ds. From this it is easy to see that u is A-subharmonic if
f0
and only if Au > 0 in D. Now we shall describe the set D(x). For simplicity, we shall assume that D has the property that t sup Ey[ f I,r(w(s))ds] < 00 yeD 0
for every compact set K C D, where z = -r(w) = inf {t; w(t)D}. Choose a compact exhaustion of D. Then the above assumption clearly implies that sup o for every n. Set Ye D
(8.21)
D(x) = {y; 3p1 E .7, 3t0 > 0 such that y = t(x, ¢)(t0)
and (x, O)(t); t E [0, to]) c D fl D. Theorem 8.3.* D(x) possesses the properties (8.19) and (8.20). Proof. First we shall prove that the set D(x) defined by (8.21) possesses
the property (8.19). Let y E D(x). It follows from Theorem 8.1 that for every neighbourhood U of y there exists an n such that Px(cr < cr) > 0,
where o = o (w) = inf {t; w(t) E U1. Let u(x) be an A-subharmonic function in D such that u(x) = max u(z). Then zCD(x)
u(x) < EE(u(w(o- A Q.)));
moreover, Theorem 8.1 implies that Px(w(a , A a.) E D(x)) = 1. Conu(x)) = 1. In particular, u(w(6 )) = u(x) on sequently, P(u(w(cr the set {au < c,,}. From this it is easy to find a sequence z ED such
that z - y and
u(x). Then
*_This was first proven by Stroock-Varadhan [158], but they only discussed the case of operators of parabolic type.
SUPPORT OF DIFFUSION PROCESSES
U(X) > u(y)
531
hm u(Z,) = u(x),
that is, u(y) = u(x). Next we shall prove that D(x) defined by (8.21) possesses the property (8.20). For this we need the following lemma.
Lemma 8.4. If f is a nonpositive continuous function in D with compact support, then the function u defined by u(y) = E,, [ f f(w(s))ds]
is a bounded A-subharmonic function in D. Assuming this lemma for a moment, we shall now complete the proof of Theorem 8.3. Let z E D\D(x). Then by Theorem 8.1 we can easily find a bounded neighbourhood U of z such that P.,(a,, <,r) = 0. We choose a
nonpositive continuous function fin D such that,f(z) = -1 and f(y) = 0 if y
U. Set
u(y) = EY[ f rf(w(s))ds]
Then u is a bounded nonpositive A-subharmonic function in D such that u(x) =0 and u(z) < 0. This proves that D(x) possesses the property (8.20). Proof of Lemma 8.4. u is bounded because of the above assumption on D. First we shall prove that it is upper semicontinuous. We saw in the proof of Proposition V-2.1 that if X(t, x) is the solution of (8.1), then for every t > 0, Rd D X , X(t, X) E Wd is continuous a.s. Then we can easily see that
is lower semicontinuous a.s. and hence x'-' I+r
is upper semicontinuous a.s. The upper semicontinuity of
x
u()c) = Ex( f o.f'(w(s))ds] = f E[f(X(t, o
follows at once from Fatou's lemma.
dt
COMPARISON AND APPROXIMATION THEOREMS
532
Next we shall prove that t -- u(w(t A v )) is a Pz submartingale for u(x) for every every n. It suffices to prove that EE[u(w(c A
stopping time v. But this is a consequence of the following formula (Dynkin's formula) which is immediately obtained from the strong Markov property :
Ex[ulw(cAo ))] - u(x) _ -Ex[
I0
Jlw(s))ds1
9. Asymptotic evaluation of the diffusion measure for tubes around a smooth curve
Consider a non-singular diffusion process X on a manifold M. We are sometimes interested in the following questions: given two smooth curves starting at the same point, which one is more probable for the diffusion process, or, among all possible smooth curves connecting two given points, which one is most probable for the diffusion process? One way to answer these questions is to evaluate the measure of tubes around a smooth curve ([154]). As we know from Chapter V, Section 4, we may assume that M is a Riemannian manifold and the diffusion is generated by the operator A where
A= 2 d -{- b.
Here d is the Laplace-Beltrami operator and b is a vector field. The Riemannian distance p(x, y) is defined (if x and y is sufficiently near) as the minimum length of a geodesic curve from x toy. For a given smooth curve
0: [0, T] -- M such that 0(0) = x, we want to evaluate p (0) = Px {w; p(X,(w), 0(t)) < a for all
t
[0, T]} .
If for any two such curves ¢ and yr, llm elo
exists and can be expressed as exp [ T L(O(s), 0(s))ds - f L(yr(s), i(s))ds] J
0
o
by some function L(z, x) on the tangent bundle TM, the above questions
ASYMPTOTIC EVALUATION OF DIFFUSION MEASURE
533
can be answered in terms of the function L Such a function is called the Onsager-Machlup function by physicists (cf. [20], [41], [56] and [134]. For simplicity, we discuss here the case of M = Rd and obtain the function L: we refer the reader to (199] and [227] for the results in general manifolds. So let M = Rd and let {Px} be the diffusion measure (on Wd) generated
by the operator
We shall assume that bi(x) E Cb(Rd), i=1,2,... , d. Let Pw be the d-dimensional Wiener measure (on Wd) starting at 0. By Lemma 8.1, we know that ,FT]
(9.1)
where A, is the first eigenvalue of the eigenvalue problem in D= {x E Rd; I x I < 11 given by
1 d0+20=0 (9.2)
OIaD=0, and c = 0,(0) f, q3,(x)dx, O,(x) being the normalized eigenfunction for A,. In the following, T > 0 and x E Rd are arbitrary but fixed.
Theorem 9.1. Let 0: [0, T] -- Rd be a smooth curve*2 such that 0(0) = x. Then we have (writing b(x) = (b'(x), bz(x), ... , bd(x)) Px(w IIw - 0IIT < a) (9.3)
exp
[- 2 f
I b(a(s)) - c(s) I gds o
-2f
(div b)(O(s))ds] P'(II wll r < a)
T
c exp
(- 2 f o I b(O(s)) - (s) I z ds o
T - 2 f (div b)(O(s))ds) exp [- E1 T o
" 11 wll7 = max I w(t) 1, w e W. Osi T
It is sufficient to assume that 0 e C2([O, Tl - Rd).
as
a 10.
COMPARISON AND APPROXIMATION THEOREMS
534
(div b) (x) = E
Here
ax`
b'(x).
Corollary. The Onsager-Machlup function L(x, x) is given, up to an additive constant, by
L(z, x)
(9.4)
2 12 - b(x)
Z I
-
(div b)(x).
Proof. By the transformation of drift discussed in Chapter IV, Section 4-1, we have T
PP(B) = Ew(exp [f0 b(x + w(s))dw(s)
(9.5)
2
f
o I
b(x + w(s)) I Zds]; B)
for BER(Wd). Similarly, PW(B)
O(s)
Ew[exp [-f o (s)dw(s)
2
(9.6)
f
I
I
Zds]
o
:(w;[w+¢-x]E B)].*i By combining (9.5) and (9.6), we have
P.(IIw-01IT<E)
= Ea'[exp [- J 0 (s)dw(s)
(9.7)
+f
o
2
f
o
I (s) I Zds
b(w(s) + O(s))d[w(s) + c(s)] - i f o I b(w(s) + O(s)) I Zds]
IIwIIT<E].
If w
satisfies IIwIIT < e, then
I f O(s) dw(s) I = I O(T) w(T) o
-f
w(s) (s)ds
A,c, *2
o
` For w e Wd and x e R , w + x E Wd is defined by (w + x](t) = w(t) + x. i2 All A21 ... and K Ks, ... are positive constants independent of e.
ASYMPTOTIC EVALUATION OF DIFFUSION MEASURE
535
b(o(s))c(s)ds I< AZE I f T b(w(s) + O(s))q$(s)ds - f T 0 0
and
I f I b(w(s) + ¢(s)) I zds - f I b(o(s)) I zds I< A,E. 0
0
We immediately have from this that
exp [- 2 f0 I c(s) - b(a(s)) I zds - (A, +A2+A3) < PP(II w -1II'T < s) {E1'(exp[ f o b(w(s) + 0(s))dw(s)]: IIII T < E]} -1
(9.8)
< exp [- 2
Jo
It(s) - b(o(s)) 12ds + (Al+Az+A3)E] .
Consequently it is sufficient to show that
Ew[exp[f o b(w(s) + ¢(s))dw(s) + 2
(9.9)
f
(div b) (O(s))ds] 111w11 r < E] o
as Ej0. Now
f
b(w(s) + O(s))dw(s) + 2 0
=f
o
+
f
(div b)(O(s))ds 0
b(¢(s))dw(s) +f b`J(¢(s))w1(s)dw'(s) .
f
o
(div b)(i(s))ds + Y; f
0t(s,
w)dw°(s)
o
where
0,(s,w) = b`(w(s)+O(s)) - b`(g(s)) - Z
(9.10)
Since
* b'j=a b' ax
COMPARISON AND APPROXIMATION THEOREMS
536
f
if
b(¢(s))dw(s) I < Ac
IIWIIT < e,
o
(9.9) is equivalent to d
E w exp {
1 i f ro b`,(¢(s))w'(s)dw`(s) p`(s, w)dw`(s)} I IIwIIr
(9.11)
+ 2 f o (div b)(O(s))ds + 571 fo
< EJ
as e10. Note that
t, J-1
f
T
b1,(c(s))w'(s)dw'(s) + 2 fo (div b)(O(s))ds 0
4J-1
f
0
b`j(O(s))
[wJ(s)dwi(s)
+ 2 a,, ds]
r
_ r, -1 f 0 b`,(O(s))dc "(s) where
V" (t) = fa w'(s)edw`(s) Generally, if random variables Y1, Y2 satisfy lim Ew[ e°'"r
I
11wIIr < e] < 1,
810
i= 1,2
for every real constant c, then
limE"[exp[Y1+ YZ] I IIWIIr<E] 810
< lim 810
{Ew( exp[2Y1] I !IwIIr
< c) Ew( exp[2Y2] I IIwIIT < e)} lie
< 1, and similary
lim E'(exp(-Y,-Y2) I IIwIIr < E) < 1. 810
ASYMPTOTIC EVALUATION OF DIFFUSION MEASURE
537
But
E`"(exp[Y, + Yz] I Ilwllr < E)ER'(exp[-Y,-Yz] I Ilwllr < e)
JEW(exp[Y 2 Yz]eXP[-Y2 Yz] Illwllr<s)}2=1, and hence lim EW(exp[Y, + Yz] I Ilwll r < e)
-ia
{alienE}(exp(-Y, - Y2) IIIwIIr<e)}-'> 1. In conclusion, we have lira EW( exp[Y, + Yz] I Ilwllr < e) = 1. alo
This result can be extended easily to an arbitrary number of random variables: if Y1, Yz,
... , Y.
satisfy
lien EW(exp[cY1] I Ilwllr < e) < 1
for every real constant c and i = 1, 2, ... , n, then lien EW( exp[Y,+Yz+
I
alo
Ilwllr < e) = 1.
Noting this fact, (9.11) will follow if we can show that for every real
constant c and i, j = 1, 2, ... , d, (9.12)
lien E W(exp [c f o b;,(O(s))dcfl(s)]
I
I I wl l r< e) S 1
and
(9.13)
11lm E"'[exp[c I o cP (s, w)dw'(s)] I Ilwllr < e] < 1.
First we shall prove (9.13). Since I co,(s, w) I < Asez
on the set
{w; Ilwllr < e} ,
COMPARISON AND APPROXIMATION THEOREMS
538
we have
PIVI
ft 0
cO,(s, w)dw'(s)lIr > 6 111w;!r < E)
< A6exp [ L
-
A162
4 A8EZ1
< Klexp
J
E
rE 1-K2
Sa]
if 0 < E < K38
by a standard estimate as in the proof of Theorem 8.2. From this it is easy to conclude that lim E w( exp [ I c f 0#1 w)dw`(s) I]
(9.14)
810
I
I I W1 I T< e) = 1.
o
Indeed, for each 80 > 0 and 0 < E < K3S0,
Ew(exp[Ic f o 0t(s, w)dw'(s)I] I
< eao + Kl f ee exp [- K4 2] dd + Kiebo exp [- K;ao] ao
(9.14) follows by first letting a } 0 and then letting So 10. Finally we shall prove (9.12). By writing cb'f = c(x) it is sufficient to show that limo E'(exp[ f c(O(s))ddf`(s)] o 81
(9.15)
I
lIwlIr < E) < 1
for every c(x) E Cb(Rd). Now
f
r
r 0
c(O(s)) ddf'(s) = c(O(T ))cf r(T) -
f0 f'(s)d [c(qS(s))]
Also cf'(t) = 2 w'(t)w'(t) - raf(t) where i7"(t) is defined by (8.11). As we saw in the proof of Lemma 8.2, i7'1(t) = B(
f
[(w`)2(s) + (w')2(s)]ds) o
where B is a Brownian motion independent of the radial process (and hence independent of IIwjjr). Then
w(t) I }
ASYMPTOTIC EVALUATION OF DIFFUSION MEASURE
E"'(exp[
<
Jo
eKgea
c(c(S))d `'(S)]
I
E '(exp [K5 max
539
IIWIIT < e) I B(t) I ])
0
=
eKgez
JoV
as e j 0.
exp
This completes the proof. In Chapter V, Section 4, we saw that the diffusion {P,} is symmetrizad
ble if and only if the differential 1-form co defined by w= E bt(x)dxt is
given as co = dF for some F e C°°(Rd --- R), or equivalently the line integral f r o) vanishes along every closed smooth curve. Using Theorem 9.1, this condition can be restated as follows. Theorem 9.2. The diffusion {Px} is symmetrizable if and only if (9.16)
lim PX(I I W -
011T < E) = 1 eloPx(IIW-0-IIT<E)-
for every x and every smooth curve 0: [0, T] -. Rd such that 0(0) _ 4(T) = x. Here the curve 0_ is defined by (9.17)
¢_(t)=O(T-t),
0
Proof. By Theorem 9.1, the limit in (9.16) is equal to exp(2
J0
exp(2
J
o(0).
Thus (9.16) holds if and only if f d co = 0. A similar probabilistic characterization of the symmetry was obtained by Kolmogorov [87] in the case of a Markov chain.
Bibliography
1. S. Agmon, Lectures on elliptic boundary value problems, Van Nostrand, Princeton, New Jersey, 1965. 2. H. Airault, Perturbations singulieres et solutions stochastiques de problemes de D. Neumann-Spencer, Jour. Math. Pures Appl., 55(1976), 233-268. 3. V. E. Benes, Composition and invariance methods for solving some stochastic control problems, Adv. Appl. Prob., 7 (1975), 299-329.
4. S. Bernstein, Principes de la theorie des equations differentielles stochastiques, Trudy Fiz-Mat. Stekrov Inst. Acad. Nauk, 5 (1934), 95-124. 5. S. Bernstein, Equations differentielles stochastiques, Act. Sci. et Ind. 738, Conf. intern. Sci. Math. Univ. Geneve, 5-31, Herman, Paris, 1938. 6. P. Billingsley, Convergence of probability measures, John Wiley and Sons, New York, 1968. 7. R. Bishop and R. J. Crittenden, Geometry of manifolds, Academic Press, New York, 1964.
8. R. M. Blumenthal and R. K. Getoor, Markov processes and potential theory, Academic Press, New York, 1968. 9. R. H. Cameron and W. T. Martin, Transformation of Wiener integrals under translations, Ann. Math., 45 (1944), 386-396. 10. R. H. Cameron and W. T. Martin, Evaluations of various Wiener integrals by use of certain Sturm-Liouville differential equations, Bull. Amer. Math. Soc., 51 (1945), 73-90. 11. K. L. Chung and It. Durret, Downcrossing and local time, Z. Wahr. verw. Geb., 35 (1976), 147-149.
12. B. S. Cirel'son, An example of a stochastic differential equation having no strong solution, Theor. Prob. Appl., 20 (1975), 416-418. 13. P. E. Conner, The Neumann's problem for differential forms on Riemannian manifolds, Mem. Amer. Math. Soc., 20 (1956). 14. A. Debiard, B. Gaveau and E. Mazet, Theoremes de comparaison en geometrie riemannienne, Publ. RIMS, Kyoto Univ., 12 (1976), 391-425. 15. C. Dellacherie, Capacites et processus stochastiques, Springer-Verlag, Berlin, 1972. 16. C. Dellacherie, Integrals stochastiques par rapport aux processus de Wiener et de Poisson, Seminaire de Prob. (Univ. de Strasbourg) VIII, Lecture Notes in Math., 381, 25-26, Springer-Verlag, Berlin, 1974. 17. C. Doleans-Dade and P. A. Meyer, Integrales stochastiques par rapport aux martingales locales, Seminaire de Prob. (Univ. de Strasbourg) IV, Lecture Notes in Math., 124,77-107, Springer-Verlag, Berlin, 1970. 18. J. L. Doob, Stochastic processes, John Wiley and Sons, New York, 1953. 19. H. Doss, Liens entre equations differentielles stochastiques et ordinaires, Ann. Inst. H. Poincarg, 13 (1977), 99-125. 20. D. Durr and A. Bach, The Onsager-Machlup function as Lagrangian for the most probable path of a diffusion process, Comm. Math. Phys., 60 (1978), 153-170. 21. E. B. Dynkin, Theory of Markov processes, Pergamon Press, Oxford, 1960. 22. E. B. Dynkin, Markov processes, 1, II, Springer-Verlag, Berlin, 1965. 23. J. Eells and K. D. Elworthy, Stochastic dynamical systems, Control theory and 541
BIBLIOGRAPHY
542
topics in functional analysis, III, Intern. atomic energy agency, Vienna, 1976, 179185.
24. A. Einstein, Investigations on the theory of the Brownian movement, Methuen, London, 1926. 25. K. D. Elworthy, Stochastic dynamical systems and their flows, Stochastic Analysis (ed. by A. Friednman and M. Pinsky), 79-95, Academic Press, New York, 1978. 26. W. Feller, Zur Theorie der Stochastischen Prozesse (Existenz- und Eindeutigkeitssatze), Math. Ann., 113 (1936), 113-160. 27. D. L. Fisk, Quasi-martingales and stochastic integrals, Tech. Rep. 1, Dept. Math. Michigan State Univ., 1963. 28. A. Friedman, Stochastic differential equations and applications Vol. 1 and Vol. 2, Academic Press, New York, 1975. 29. M. Fujisaki, G. Kallianpur and H. Kunita, Stochastic differential equations for the non linear filtering problem, Osaka J. Math., 9 (1972), 19-40. 30. M. Fukushima, Dirichlet forms and Markov processes, Kodansha, Tokyo, 1980. 31. T. Funat,d, Construction of a solution of random transport equation with boundary condition, J. Math. Soc. Japan, 31 (1979), 719-744. 32. A. M. Garsia, Martingale inequalities, Seminar Notes on Recent Progress, W. A. Benjamin, Reading, Massachusetts, 1973. 33. B. Gaveau, Principe de moindre action, propagation de la chaleur et estimees sous elliptiques sur certain groupes nilpotents, Acta Math., 139 (1977), 95-153. 34. R. K. Getoor and M. J. Sharpe, Conformal martingales, Invent. Math., 16 (1972), 271-308.
35. I. 1. Gihman, A method of constructing random processes, Dokl. Akad. Nauk SSSR, 58 (1947), 961-964, (in Russian). 36. I. I. Gihman, Certain differential equations with random functions, Ukr. Mat. Zh., 2 (1950), No. 3, 45-69, (in Russian). 37. I. I. Gihman, On the theory of differential equations of random processes, Ukr. Mat. Zh., 2 (1950), No. 4, 37-63, (in Russian). 38. 1. I. Gihman and A. V. Skorohod, Stochastic differential equations, Springer-Verlag, Berlin, 1972.
39. I. I. Gihman and A. V. Skorohod, The theory of stochastic processes III, SpringerVerlag, Berlin, 1979. 40. I. V. Girsanov, On transforming a certain class of stochastic processes by absolutely continous substitution of measures, Theor. Prob. Appl., 5 (1960), 285-301. 41. R. Graham, Path integral formulation of general diffusion processes, Z. Physik B, 26 (1977), 281-290. 42. R. E. Greene and H. Wu, Function theory on manifolds which possess a pole, Lecture Notes in Math., 699, Springer-Verlag Berlin, 1979. 43. B. Grigelionis, On representation of integer-valued random measures by means of stochastic integrals with respect to the Poisson measure, Lit. Mat. Sb., 11 (1971), 93-108, (in Russian).
44. B. Grigelionis, On the martingale characterization of stochastic processes with independent increments, Lit. Mat. Sb., 17 (1977), 52-60. 45. P. R. Halmos, Measure theory, Van Nostrand, New York, 1950. 46. R. Z. Hashiminsky, Ergodic properties of recurrent diffusion processes and stabilization of the solution of the Cauchy problem for parabolic equations, Theor. Prob. Appl., 5 (1960), 179-196. 47. S. Helgason, Differential geometry and symmetric spaces, Academic Press, New York, 1962. 48. L. Hormander, Hypoelliptic second order differential equations, Acta Math., 119 (1967), 147-171. 49. N. Ikeda, On the construction of two-dimensional diffusion processes satisfying Wentzell's boundary conditions and its application to boundary value problems, Mem. Coll. Sci. Univ. Kyoto Math., 33 (1961), 367-427.
50. N. Ikeda and S. Manabe, Stochastic integral of differential forms and its appli-
BIBLIOGRAPHY
543
cations, Stochastic Analysis (ed. by A. Friedman and M. Pinsky), 175-185, Academic Press, New York, 1978. 51. N. Ikeda and S. Manabe, Integral of differential forms along the path of diffusion processes, Pub!. RIMS, Kyoto Univ., 15 (1979), 827-852. 52. N. Ikeda, S. Nakao and Y. Yamato, A. class of approximations of Brownian motion, Publ. RIMS, Kyoto Univ., 13 (1977), 285-300. 53. N. Ikeda and S. Watanabe, The local structure of a class of diffusions and related problems, Proc. Second Japan-USSR Symp. Prob. Theor., Lecture Notes in Math., 330, 124-169, Springer-Verlag, Berlin, 1973. 54. N. Ikeda and S. Watanabe, A comparison theorem for solutions of stochastic differential equations and its applications, Osaka J. Math., 14 (1977), 619-633. 55. N. Ikeda and S. Watanabe, Heat equation and diffusion on Riemannian manifold
with boundary, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito), 75-94, Kinokuniya, Tokyo, 1978.
56. H. Ito, Probabilistic construction of Lagrangean of diffusion processes and its application, Prog. Theoretical Phys., 59 (1978), 725-741.
57. K. R6, Differential equations determining Markov processes, Zenkoku Shijo Silgaku Danwakai, 244 (1942), No. 1077, 1352-1400, (in Japanese). 58. K. Ito, On stochastic processes (1) (Infinitely divisible laws of probability), Japanese J. Math., 18 (1942), 261-301. 59. K. Ito, Stochastic integral, Proc. Imp. Acad. Tokyo, 20 (1944), 519-524.
60. K. It6, On a stochastic integral equation, Proc. Imp. Acad. Tokyo, 22 (1946), 32-35. 61. K. Ito, Stochastic differential equations in a differentiable manifold, Nagoya Math. J., 1 (1950), 35-47. 62. K. It6, On stochstic differential equations, Mem. Amer. Math. Soc., 4 (1951). 63. K. It6, On a formula concerning stochastic differentials, Nagoya Math. J., 3 (1951), 55-65. 64. K. Ito, Multiple Wiener integral, J. Math. Soc. Japan., 3 (1951),157-169. 65. K. Ito, Stochastic differential equations in a differentiable manifold (2), Mem. Coll. Sci. Univ. Kyoto Math., 28 (1953), 81-85. 66. K. Ito, Theory of probability, Iwanami, Tokyo, 1953 (in Japanese). 67. K. Ito, Lectures on stochastic processes., Tata Institute of Fundamental Research, Bombay, 1960. 68. K. Ito, The Brownian motion and tensor fields on Riemannian manifold, Proc. Intern. Congr. Math., Stockholm, 536-539, 1963. 69. K. 10, Stochastic processes, Lecture Notes Series, 16, Aarhus Univ. 1969. 70. K. Ito, Poisson point processes attached to Markov processes, Proc. Sixth Berkeley Symp. Math. Statist. Prob. III, 225-239, Univ. California Press, Berkeley, 1972.
71. K. Ito, Stochastic differentials, Appl. Math. Opt., 1 (1975), 374-381. 72. K. Ito, Stochastic parallel displacement, Probabilistic methods in differential equations, Lecture Notes in Math., 451, 1-7, Springer-Verlag, Berlin, 1975. 73. K. Ito and H.P. McKean, Jr., Diffusion processes and their sample paths, SpringerVerlag, Berlin, 1965. 74. K. Ito and M. Nisio, On stationary solutions of a stochastic differential equation, J. Math. Kyoto Univ., 4 (1964), 1-75. 75. K. It6 and M. Nisio, On the convergence of sums of independent Banach space valued random variables, Osaka J. Math., 5 (1968), 35-48. 76. K. It6 and S. Watanabe, Introduction to stochastic differential equations, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. It6), i-xxx, Kinokuniya, Tokyo, 1978. 77. J. Jacod, Calcul stochastique et problemes de martingales, Lecture Notes in Math., 714, Springer-Verlag, Berlin, 1979. 78. T. Jeulin, Grossissement d'une filtration et applications, Seminaire de Prob. XIII (Univ. de Strasbourg), Lecture Notes in Math., 721, 574-609, Springer-Verlag, Berlin, 1979.
544
BIBLIOGRAPHY
79. M. Kac, On distributions of certain Wiener functionals, Trans. Amer. Math. Soc., 65 (1949), 1-13. 80. M. Kac, On some connections between probability theory and differential and integral equations, Proc. Second Berkeley Symp. Math. Statist. Prob., 189-215, Univ. California Press, Berkeley, 1951. 81. N.E. Karoui and J.P. Lepeltier, Representation des processus ponctuels multivari6s a l'aide d'un processus de Poisson, Z. Wahr. verw. Geb., 39 (1977),111-133. 82. Y. Kasahara and S. Kotani, On limit processes for a class of additive functionals of recurrent diffusion processes, Z. Wahr. verw. Geb., 49 (1979), 133-159. 83. N. Kazamaki, The equivalence of two conditions on weighted norm inequalities for martingales, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito), 141-152, Kinokuniya, Tokyo, 1978. 84. F.B. Knight, A reduction of continuous square-integrable martingales to Brownian motion, Lecture Notes in Math., 190, Springer-Verlag, Berlin, 1971. 85. A.N. Kolmogoroff, Ober die analytischen Methoden in der Wahrscheinlichkeitsrechnung, Math. Ann., 104 (1931), 415-458. 86. A.N. Kolmogoroff, Zufallige Bewegungen, Ann. Math. II., 35 (1934), 116-117. 87. A.N. Kolmogoroff, Zur Theorie der Markoffschen Ketten, Math. Ann., 112 (1936), 155-160.
88. A.N. Kolmogoroff, Zur Umkehrbarkeit der statistischen Naturgesetze, Math. Ann., 113 (1937),766-772.
89. M.G. Krein and M.A. Rutman, Linear operators leaving invariant a cone in a Banach space, Amer. Math. Trans. Series 1, 10 (1950), 199-325. 90. N.V. Krylov, On Ito's stochastic integral equations, Theor. Prob. Appl.,14 (1969), 330-336. 91. N.V. Krylov, Control of diffusion type processes, Nauka, Moscow, 1977, (in Russian). 92. N.V. Krylov and M.V. Safonov, Estimate for the hitting probability of diffusion processes to a set of positive measure, Dokl. Akad. Nauk SSSR., 245 (1979), 18-20, (in Russian). 93. H. Kunita, Diffusion processes and control systems., Course at University of Paris VI, 1974. 94. H. Kunita, Estimation of stochastic processes, Sangyo Tosho, Tokyo, 1976 (in Japanese). 95. H. Kunita, Supports of diffusion processes and controllability problems, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito), 163-185, Kinokuniya, Tokyo, 1978. 96. H. Kunita, On the representation of solutions of stochastic differential equations, Seminaire de Prob. XIV, Lecture Notes in Math., 784, 282-304, Springer-Verlag, Berlin, 1980. 97. H. Kunita and S. Watanabe, On square integrable martingales, Nagoya Math. J., 30 (1967), 209-245. 98. K. Kuratowski, Topology I., Academic Press, New York, 1966. 99. P. Levy, Theorie de 1'addition des variables aleatoires, Gauthier-Villars, Paris, 1937. 100. P. Levy, Le mouvement brownien plan, Amer. J. Math., 62 (1940), 487-550. 101. P. Levy, Processus stochastiques et mouvement brownien, Gauthier-Villars, Paris, 1948.
102. R.S. Liptzer and A.N. Shiryaev, Statistics of stochastic processes, Nauka, Moscow, 1974, (in Russian). 103. B. Maisonneuve, Exit systems, Ann. Prob., 3 (1975), 399-411. 104. P. Malliavin, Formule de la moyenne, calcul de perturbations et theoremes d'annulation pour les formes harmoniques, J. Funct. Analy., 17 (1974), 274-291. 105. P. Malliavin, Champs de Jacobi stochastiques, C.R.A. Sc. Paris, 285 (1977), 789792.
106. P. Malliavin, Stochastic calculus of variation and hypoelliptic operators, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito),195-263, Kinokuniya, Tokyo, 1978.
13S/1
BIBLIOGRAPHY
545
107. P. Malliavin, Ck-hypoellipticity with degeneracy, Stochastic Analysis (ed. A. Friedman and M. Pinsky), 199-214, 327-340, Academic Press, New York, 1978. 108. P. Malliavin, Geometrie differentielle stochastique, Les Presses de I'Universite de
Montreal, Montreal, 1978. 109. S. Manabe, On the intersection number of the path of a diffusion and chains, Proc. Japan Acad., 55 (1979), 23-26.
110, G. Maruyama, On the transition probability functions of the Markov process, Nat. Sci. Rep. Ochanomizu Univ., 5 (1954), 10-20. 111. G. Maruyama, Continuous Markov processes and stochastic equations, Rend. Circ. Mate. Palermo., 4 (1955), 48-90. 112. Y. Matsushima, Differentiable manifolds, Marcel Dekker, New York, 1972. 113. H.P. McKean Jr., Stochastic integrals, Academic Press, New York, 1969. 114. H.P. McKean, Brownian local times, Adv. Math., 16 (1975), 91-111. 115. E.J. McShane, Stochastic differential equations and models of random processes, Proc. Sixth Berkeley Symp. Math. Statist. Prob. 111, 263-294, Univ. California Press, Berkeley, 1972. 116. E.J. McShane, Stochastic calculus and stochastic models., Academic Press, New York, 1974. 117. P.A. Meyer, Probability and potentials, Blaisdel, Waltham, Massachusetts, 1966. 118. P.A. Meyer, stochastiques I-IV, Seminaire de Prob. (Univ. de Strasbourg) I. Lecture Notes in Math., 39, 72-162, Springer-Verlag, Berlin, 1967.
119. P.A. Meyer, Demonstration simplifiee d'un theoreme de Knight, Seminaire de Prob. (Univ. de Strasbourg) V, Lecture Notes in Math., 191, 191-195, SpringerVerlag, Berlin, 1971.
120. P.A. Meyer, Martingales and stochastic integrals I, Lecture Notes in Math., 284, Springer-Verlag, Berlin, 1972. 121. P.A. Meyer, Un tours sur les integrales stochastiques, Seminaire de Prob. (Univ. de Strasbourg) X, Lecture Notes in Math., 511, 245-400, Springer-Verlag, Berlin, 1976.
122. J. Milnor, Morse theory, Princeton Univ. Press, Princeton, New Jersey, 1963. 123. M. Motoo, Diffusion process corresponding to 1/2 E 82/8x12 + E b`(x) a/8x0, Ann. Inst. Statist. Math., 12 (1960-61), 37-61. 124. M. Motoo, Periodic boundary problems of the two dimensional Brownian motion on upperhalf plane, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito), 265281, Kinokuniya, Tokyo, 1978. 125. S. Nakao, On the pathwise uniqueness of solutions of ona-dimensional stochastic differential equations, Osaka J. Math., 9 (1972), 513-518. 126. S. Nakao and Y. Yamato, Approximation theorem on stochastic differential equations, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. Ito), 283-296, Kinokuniya, Tokyo, 1978. 127. E. Nelson, The adjoint Markoff process, Duke Math. J., 25 (1958), 671-690. 128. E. Nelson, Tensor analysis, Mathematical Notes, Princeton Univ. Press, Princeton, New Jersey, 1967. 129. J. Neveu, Bases mathematiques du calcul des probabilites, Masson et Cie., Paris, 1964.
130. M. Nisio, On the existence of solutions of stochastic differential equations, Osaka J. Math., 10 (1973), 185-208. 131. K. Nomizu, Lie groups and differential geometry, Publ. Math. Soc. Japan, Tokyo, 1956.
132. A.A. Novikov, On moment inequalities and identities for stochastic integrals, Proc. Second Japan-USSR Symp. Prob. Theor., Lecture Notes in Math., 330, 333-339, Springer-Verlag, Berlin, 1973. 133. S. Ogawa, On a Riemann definition of the stochastic integral, (I), (II), Proc. Japan Acad., 46 (1970), 153-157, 158-161. 134. L. Onsager and S. Machlup, Fluctuations and irreversible processes, I, II, Phys. Rev., 91 (1953), 1505-1512, 1512-1515.
BIBLIOGRAPHY
546
135. S. Orey, Diffusion on the line and additive functionals of Brownian motion, Proc. conf. on stochastic differential equations and applications (ed. by J.D. Mason), 211230, Academic Press, New York, 1977.
136. R.E.A.C. Paley and N. Wiener, Fourier transforms in the complex domain, Amer. Math. Soc, Coll. Publ., 19, 1934. 137. G.C. Papanicolaou, D.W. Stroock and S.R.S. Varadhan, Martingale approach to some limit theorems, 1976 Duke Turbulence Conf., Duke Univ. Math. Series III, 1977.
138. K.P. Parthasarathy, Probability measures on metric spaces, Academic Press, New York, 1967. 139. M. Pinsky, Multiplicative operator functionals and their asymptotic properties, Advances in Probability (ed. by P. Ney and S. Port), 3, 1-100, Marcel Dekker, New York, 1974. 140. M. Pinsky, Stochastic Riemannian geometry, Probabilistic analysis and related topics (ed. by A.T. Bharucha-Reid), Academic Press, New York, 1978. 141. J.W. Pitman, One-dimensional Brownian motion and the three-dimensional Bessel process, Adv. Appl. Prob., 7 (1975), 511-526. 142. Yu.V. Prohorov, Convergence of random processes and limit theorems in probability theory, Theor. Prob. Appl., 1 (1956), 157-214. 143. K.M. Rao, On decomposition theorems of Meyer, Math. Scand., 24 (1969),66-78. 144. D.B. Ray and T.M. Singer, R-torsion and the Laplacian on Riemannian manifolds, Adv. Math., 7 (1971), 145-210. 145. G.de Rham, Varietes differentiables, Hermann, Paris, 1960. 146. T. Shiga, Diffusion processes in population genetics, J. Math. Kyoto Univ., Forthcoming. 147. T. Shiga and S. Watanabe, Bessel diffusions as a one-parameter family of diffusion processes, Z. Wahr. verw. Geb., 27 (1973), 37-46. 148. I. Shigekawa, Derivatives of Wiener functionals and absolute continuity of induced measures, J. Math. Kyoto Univ., 20 (1980), 263-289. 149. A.V. Skorohod, Stochastic equations for diffusion processes in a bounded region, Theor. Prob. Appl., 6 (1961), 264-274. 150. A.V. Skorohod, Studies in the theory of random processes, Addison-Wesley, Reading, Massachusetts, 1965. 151. S.L. Sobolev, Applications of functional analysis in mathematical physics, Amer. Math. Soc., Providence, 1963. 152. E.M. Stein, Singular integrals and differentiability properties of functions, Princeton Univ. Press, Princeton, New Jersey, 1970. 153. R.L. Stratonovich, Conditional Markov processes and their application to the theory of optimal control, Amer. Elsevier, New York, 1968. 154. R.L. Stratonovich, On the probability functional of diffusion processes, Selected Trans. in Math. Statist. Prob., 10 (1971), 273-286. 155. C. Stricker, Quasimartingales, martingales locales, semimartingales et filtration naturelle, Z. Wahr. verw. Geb., 39 (1977), 55-63. 156. D.W. Stroock, On the growth of stochastic integrals, Z. Wahr. verve. Geb., 18 (1971), 340-344. 157. D.W. Stroock and S.R.S. Varadhan, Diffusion processes with continuous coefficients, I, II, Comm. Pure Appl. Math., 22 (1969), 345-400, 479-530. 158. D.W. Stroock and S.R.S. Varadhan, On the support of diffusion processes with
applications to the strong maximum principle, Proc. Sixth Berkeley Symp. Math. Statist. Prob. Ill., 333-359, Univ. California Press, Berkeley, 1972. 159. D.W. Stroock and S.R.S. Varadhan, Diffusion processes, Proc. Sixth Berkeley Symp. Math. Statist. Prob. 111., 361-368, Univ. California Press, Berkeley, 1972. 160. D.W. Stroock and S.R.S. Varadhan, Multidimensional diffusion processes, SpringerVerlag, Berlin, 1979. 161. H. Tanaka, On the uniqueness of Markov process associated with the Boltzmann
BIBLIOGRAPHY
547
equation of Maxwellian molecules, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. It6), 409-425, Kinokuniya, Tokyo, 1978. 162. H. Tanaka and M. Hasegawa, Stochastic differential equations, Seminar on Prob., 19, 1964, (in Japanese). 163. H.F. Trotter, A property of Brownian motion paths, Illi. J. Math., 2 (1958),425433.
164. Y. Umemura (Y. Yamazaki), On the infinite dimensional Laplacian operator, J. Math. Kyoto Univ., 4 (1965), 477-492. 165. S.R.S. Varadhan, Stochastic processes, Lecture Notes, Courant Institute of Math. Sci. New York Univ. (1967/68), 1968. 166. S. Watanabe, On stochastic differential equations for multi-dimensional diffusion processes with boundary conditions, 1, 11, J. Math. Kyoto Univ., 11(1971), 169-180, 545-551. 167. S. Watanabe, Solution of stochastic differential equations by random time change, Appl. Math. Opt., 2 (1975), 90-96. 168. S. Watanabe, On time inversion of one-dimensional diffusion processes, Z. Wahr. verw. Geb., 31 (1975), 115-124.
169. S. Watanabe, Stochastic differential equations, Sangyo Tosho, Tokyo, 1975 (in Japanese). 170. S. Watanabe, Construction of diffusion processes by means Poisson point process of Brownian excursions, Proc. Third Japan-USSR Symp. Prob. Theor., Lecture Notes in Math., 550, 650-654, Springer-Verlag, Berlin, 1976. 171. S. Watanabe, Poisson point process of Brownian excursions and its applications to diffusion processes, Proc. Symp. Pure Math. Amer. Math. Soc., 31, 153-164, 1977.
172. S. Watanabe, Excursion point process of diffusion and stochastic integral, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. It0), 437-461, Kinokuniya, Tokyo, 1978. 173. S. Watanabe, Point processes and martingales, Stochastic Analysis, (ed. by A. Friedman and M. Pinsky), 315-326, Academic Press, New York, 1978. 174. S. Watanabe, Construction of diffusion processes with Wentzell's boundary conditions by means of Poisson point processes of Brownian excursions, Probability theory, Banach Center Publications, Vol. 5, 255-271, Polish Scientific Publishers, Warsaw, 1979. 175. A.D. Wentzell, On boundary conditions for multidimensional diffusion processes, Theor. Prob. Appl., 4 (1959), 164-177. 176. H. Whitney, Geometric integration theory, Princeton Univ. Press, Princeton, New Jersey, 1957.
177. N. Wiener, Differential space, J. Math. Phys., 2 (1923), 131-174. 178. N. Wiener, Nonlinear problems in random theory, M.T.T. Press, Cambridge and John Wiley and Sons, New York, 1958. 179. D. Williams, Path decomposition and continuity of local time for one-dimensional diffusions, 1, Proc. London Math. Soc. (3), 28 (1974), 738-768. 180. D. Williams, Levy's downcrossing theorem, Z. Wahr. verw. Geb., 40 (1977), 157158.
181. E. Wong and M. Zakai, On the relation between ordinary and stochastic differential equations, Intern. J. Engng. Sci., 3 (1965), 213-229.
182. T. Yamada, On a comparison theorem for solutions of stochastic differential equations and its applications, J. Math. Kyoto Univ., 13 (1973), 497-512. 183. T. Yamada and S. Watanabe, On the uniqueness of solutions of stochastic differential equations, J. Math. Kyoto Univ., 11 (1971),155-167. 184. Y. Yamato, Stochastic differential equations and nilpotent Lie algebras, Z. Wahr. verw. Geb., 47 (1979), 213-229. 185. K. Yano and S. Bochner, Curvature and Betti numbers, Ann. Math. Studies, 32, Princeton Univ. Press, Princeton, New Jersey, 1953. 186. S.T. Yau, On the heat kernel of a complete Riemannian manifold, J. Math. Pures Appl., 57 (1978), 191-201.
548
BIBLIOGRAPHY
187. M.P. Yershov, On stochastic equations, Proc. Second Japan-USSR Symp. Prob. Theor., Lecture Notes in Math., 330, 527-530, Springer-Verlag, Berlin, 1973. 188. M.P. Yershov, Localization of conditions on the coefficients of diffusion type equations in existence theorems, Proc. Intern. Symp. SDE Kyoto 1976 (ed. by K. It6), 493-507, Kinokuniya, Tokyo, 1978. 189. A.K. Zbonkin and N.V. Krylov, On. strong solutions of stochastic differential equations, Proc. School-Seminar on Theor. Random Processes (Druskininkai), Vilnius, Ac. Sci. Lit. SSR, part II, 9-88, 1975, (in Russian).
References Added in the Second Edition 190. H. Akiyama, Geometric aspects of Malliavin's calculus on vector bundles, J. Math. Kyoto Univ., 26, (1986), 673-696. 191. R. Azencott, Behavior of diffusion semi-groups at infinity, Bull. Soc. Math. France, 102 (1974), 193-240. 192. M. Bismut, Large deviations and the Malliavin calculus, Birkhause, Boston, 1984. 193. J-M. Bismut, The Atyah-Singer theorems: a probabilistic approach, 1, the index theorem, II, the Lefschetz fixed point formulas, J. Func. Anal., 57 (1984), 56-99 and 329-348.
194. H. J. Cycon, R. G. Froece, W. Kirsh and B. Simon, Schrodinger operators with application to quantum mechanics and global geometry, Texts and Monographs in Physics, Springer-Verlag, Berlin, 1987. 195. A. Debiard and B. Gaveau, Frontiere de Shilov des domains faiblement pseudoconvexes, Bull. Sci. Math., 100 (1976), 17-31. 196. R. Durrett, Brownian motion and matringales in analysis, Wadsworth, Inc., 1984. 197. K.D. Elworthy, Stochastic differential equations on manifolds, London Math, Soc. Lect. Note Ser., 70, Cambridge University Press, 1982. 198. M. Fukushima and M. Okada, On Dirichlet forms for plurisubharmonic functions, Acta Math., 159 (1987), 171-213. 199. T. Fujita and S. Kotani, The Onsager-Machlup function for diffusion processes, J. Math. Kyoto Univ., 22 (1982), 115-130.
200. B. Gaveau, Mdthodes de contrble optimal en analyse complexe. Resolution d'equation de Monge-Ampere complexe, J. Funct. Analy., 24 (1977), 391-411. 201. E. Getzler, A short proof of the Atiyah-Singer index theorems, Topology, 25 (1986), 111-117.
202. P.B. Gilkey, Invariance theory, the heat equation, and the Atiyah-Singer index theorem, Math. Lecture Set., 11, Publish or Perish, Inc., 1984. 203. N.'Ikeda and S. Watanabe, An introduction to Malliavin's calculus, Stochastic Analysis, Proc. Taniguchi Intern. Symp, on Stochastic Analysis, Katata and Kyoto 1982, (ed. by K. It6), 1-52, Kinokuniya/North-Holland, Tokyo, 1984. 204. N. Ikeda and S. Watanabe, Malliavin calculus of Wiener functionals and its applications, From local times to global geometry, control and physics, (ed. by K. D. Elworthy), Pitman Research, Notes in Math. Ser., 150, 132-178, Longman Scienti-
fic and Technical, Horlow, 1987. 205. H. Kaneko and S. Taniguchi, A stochastic approach to the Silov boundary, J. Funct. Anal., 74 (1987), 415-429. 206. S.W. Kendall, Stochastic differential geometry, An introduction, Acta App1. Math., 9 (1987), 29-60.
207. S. Kobayashi and K. Nomizu, Foundation of differential geometry, Vol. I, II, Interscience, New York, 1963 and 1969. 208. M. Kree and P. Kree, Continute de la divergence dans les espaces de Sobolev relatifs a 1'espace de Wiener, C.R. Acad. Sci. Paris, 296 (1983),833-836. 209. H. Kunita, Stochastic differential equations and stochastic flows of diffeomor-
BIBLIOGRAPHY
549
phisms, Ecole d'Ete de Probabilites de Saint-Flour XII-1982, Lecture Notes in Math., 1097, 143-303, Springer-Verlag, Berlin, 1984. 210. H-H. Kuo, Gaussian measures in Banach spaces, Lecture Notes in Math., 463, Springer-Verlag, Berlin, 1975.
211. S. Kusuoka and D.W. Stroock, Applications of the Malliavin calculus, Part I, Stochastic Analysis, Proc. Taniguchi Intern. Symp. on Stochastic Analysis, Katata and Kyoto, 1982, (ed. by K. It8), 271-306, Kinokuniya/North-Holland, Tokyo, 1984.
212. S. Kusuoka and D.W. Stroock, Applications of the Malliavin calculus, Part II, J. Fac. Sci. Univ. Tokyo Sect. IA, Math., 32 (1985), 1-76. 213. S. Kusuoka and D.W. Stroock, Applications of the Malliavin calculus, Part III, J. Fac. Sci., Univ. Tokyo Sect., IA, Math., 34 (1987), 391-442. 214. R. L6andre, Majoration en temps petit de la densit6 d'une diffusion d6g6ncr6e, Probab. Th. Rel. Fields, 74 (1987), 289-294. 215. R. L6andre, D6velopernent asymptotique de la densit6 d'une diffusion d6g6n6r6e,
to appear in J. Funct. Anal. 216. P. L6vy, Random functions: General theory with special reference to Laplacian random functions, Univ. of California Publ. in Statistics, 1 (1953), 331-388. 217. H.P. McKean and T.M. Singer, Curvature and eigenvalues of the Laplacian, J. Geom., 1 (1967), 43-69. 218. P.A. Meyer, Quelques resultats analytiques sur le semigroupe d'Ornstein-Uhlenbeck en dimension infinie, Theory and applications of random fields, Proc. IFIPWG 7/1 Working Conference, (ed. by G. Kallianpur), Lecture Notes in Contr. Inform., 49, 201-214, Springer-Verlag, Berlin, 1983. 219. M. Motoo and S. Watanabe, On a class of additive functionals of Markov processes, J. Math. Kyoto Univ., 4 (1964-65), 429-469. 220. E. Nelson, The free Markov field, J. Funct. Anal., 12 (1973), 217-227. 221. J. Neveu, Sur l'esperance conditionnelle par rapport a un movement brownien, Ann. Inst. Henri Poincare, B XII, (1976), 105-110. 222. J. Norris, Simplified Maliavin calculus, Seminaire de Prob. XX 1984/85 (ed. par J. Azema et M. Yor), Lect Notes in Math., 1204, 101-130, Springer-Verlag, Berlin, 1986.
223. V.K. Patodi, Curvature and eigenforms of the Laplace operator, J. Dif. Geom., 5 (1971), 233-249. 224. I. Shigekawa and N. Ueki, A stochastic approach for the Riemann-Roch theorem, Osaka J. Math., 25 (1988), 759-784.
225. H. Sugita, Sobolev spaces of Wiener functionals and Malliavin's calculus, J. Math. Kyoto Univ., 25 (1985), 31-48. 226. H. Sugita, On a characterization of the Sobolev spaces over an abstract Wiener space, J. Math. Kyoto, Univ., 25 (1985), 717-725. 227. Y. Takahashi and S. Watanabe, The probability functionals (Onsager-Machlup functions of diffusion processes, Stochastic integrals (ed. by D. Williams) Lecture Notes in Math., 851, 433-463, Springer-Verlag, Berlin, 1981. 228. S. Takanobu, Diagonal short time asymptotics of heat kernels for certain degenerate second order differential operators of Hdrmander type, Publ. RIMS, Kyoto Univ., 24 (1988), 169-203.
229. S. Takanobu and S. Watanabe, On the existence and uniqueness of diffusion processes with Wentzell's boundary conditions, J. Math. Kyoto Univ., 28 (1988), 71-80. 230. S. Taniguchi, Malliavin's stochastic calculus of variations for manifold-valued Wiener functionals and its applications, Z. Wahrsch, verw. Gebiete, 65 (1983), 269-290.
231. H. Uemura, On a short time expansion of the fundamental solution of heat equations by the method of Wiener functionals, J. Math. Kyoto Univ., 27 (1987), 417431.
550
BIBLIOGRAPHY
232. S. Watanabe, Lectures on stochastic differential equations and Maiiavin calculus, Tata Institute of Fundamental Research/Springer, 1984. 233. S. Watanabe, Analysis of Wiener functions (Malliavin calculus) and its applications to heat kernels, Ann. Probab., 15 (1987), 1-39.
234. S. Watanabe, Construction of semimartingales from pieces by the method of excursion point processes, Ann. Inst. Henri Poincare, Sup. au n° 2, 23 (1987), 297320.
235. D. Williams, Markov properties of Brownian local time, Bull. AMS, 75 (1969), 1035-1036.
236. M. Yor, Remarques sur une formule de Paul Levy, Seminaire de Prob. XIV, 1978/ 79, (ed.,par et M. Yer), Lect. Notes in Math., 784, 343-346, SpringerVerlag, Berlin, 1980.
Index
absolute boundary condition, 310 absorbing barrier Brownian
horizontal-, 298, 312 Lobachevskii-, 347 occupation time of-, 146 pinned-, 243 reflecting (barrier)-, 119, 209, 313
motion, 211 adapted, 21
(_97)--,21 admissible control, 442 admissible system, 442 affine connection, 278 approximation of a Wiener process,
-with drift, 208 -with random absorbtion, 208 bundle of linear frames, 267
-of orthonormal frames, 282 -of unitary frames, 345
480
approximation theorem, 480 arcsine law, 134 Ascoli-Arzala theorem, 17 asymptotic expansion of Wiener functional, 384
Cameron-Martin subspace, 351 canonical isomorphism, 423 canonical decomposition, 98 characteristic measure, 43 Chern polynomial, 432 Christoffel symbol, 282 comparison theorem, 437
Berezin formula, 434 Bessel diffusion, 237, 347 Borel
-cylinder set, 16
-for one-dimensional projection
-field, I
of diffusion processes, 452 compensated sum, 63 compensator, 60 complete, I complete probability space, I complex Brownian motion, 155 complex manifold, 341 component of connection, 278 conditional expectation, 12 conformal mortingale, 155 connection
-isomorphic, 13 -measurable, I
-set, I
boundary condition, 217 bounded variation part, 98
boundary operator of the Wentzell type, 217 Brownian bridge, 243 Brownian excursion, 240 Brownian functional, 349 Brownian motion, 40
affine-, 278
absorbing barrier-, 211 canonical realization of-, 40 complex-, 155 d-dimensional-, 40, 207 elastic barrier-, 211
-compatible with the Riemannian metric, 281
component of-, 278 -of Levi-Civita, 282 Riemannian-, 282 symmetric (torsion free)-, 281
(3')--, 41
551
INDEX
552
conservative, 203 convergence
-almost everywhere (almost surely), 8
-in law, 8
exponential quasimartingale, 149 extension, 89
standard-, 89 exterior derivative, 277 exterior product, 277
-in probability, 8 coordinate neighborhood, 247 holomorphic-, 3 41 countably determined, 14 counting measure, 43 covariant derivative, 279 covariant differentiation, 278 curvature
Gauss total-, 432 scalar-, 423 space of constant-, 462 tensor, 297
-transform, 461 differential one (p) form, 277 diffusion, 204 A --, 205, 213
(A, L)--, 218 Bessel-, 237 Gaussian-, 232
Fernigue's theorem, 402
Feynman-Kac -formula, 274, 383 -type weigh, 306 finite dimensional distribution, 17 first hitting time, 23, 316 Fisk integral, 97, 101 flow of diffeomorphisms, 254 frame, 267
unitary -, 345 Fubini-type theorem (for stochastic integrals), 116 fundamental solution, 420, 435 fundamental tensor field, 251
Gaussian diffusion, 232 Gauss-Bonnet-Chern theorem, 432 generalized expectation, 365 generalized Wiener functional, 364
Kahler-, 344 linear-, 232
geodesic, 462
-measure, 204 -measure generated by A, 205, 213 -measure generated by (A, L), 218 -process, 205
Girsanov transfomation, 192
boundary distinguished. 349 Doob-Kolmogorov's inequality, 28 Doob-Meyer's decomposition theorem, 34 drift
-term, 160 transformation of-, 192, 225 -vector field, 284 d -system, 22
elastic barrier Brownian motion, 211 equi-integrable, 30 equivariant, 280 Euler characteristic, 425 excursion, 123, 325
-formula, 328 -interval, 123, 325 -of Brownian motion, 123 existence theorem, 167 expectation, 11
conditional-, 12 generalized, 356 explosion time, 172, 173, 248 exponential mapping, 462 exponential martingale, 149
-polar coordinate, 463 harmonic form, 292 heat equation, 269
-for tensor field, 297 heat kernel, 391 Hilbert-Schmidt norm, 360 holomorphic coordinate neighborhood, 347 Hermite polynomial, 150, 354 Hermitian metric, 343 horizontal, 279
-Brownian motion, 298 canonical-vector fields, 280 -Laplacian, 298 -lift, 279, 280 -subspace, 279 hypercontractivity (of OrnsteinUhlenbeck semigroup), 367 hypoellipticity problem, 416 image measure, 2, independent
mutually-, l l
-of a a-ifield, II induced measure, 2 the regularity of, 375 inequality Doob-Kolmogorov's-, 28
INDEX
Jensen's-, 12 moment-for martingales, 110 initial distribution, 40, 205 inner regular, 3 integrable ( p-th, square), I1 integrable increasing process, 35
natural-, 35 invariant measure, 290 lt$'s formula, 66
-in the complex form, 157 Ito process, 64 Ito's stochastic integral, 45
Jensen's inequai lity, 12
553
exponential-, 149
local-, 52 localconformal-, 155 locally square-integrable-, 52
-part, 98 -problem, 74 square-integrable-, 47 -term, 160
-transform, 25 -with reversed time, 31 maximum principle, 529 measurable
-mapping, I -process, 21
K5hler -diffusion, 341, 344 -manifold, 344 -metric, 344 Knight's theorem, 86,92, 156 Langevin's equation, 233 Laplace-Beltrami operator, 285 Laplacian
de Rham-Kodaira's-, 292 horizontal-, 298 last exit formula, 328 last exit time, 316 Levy
-Khinchin formula, 65 -measure, 65 -process, 65 life time, 205 linear diffusion, 232 Lobachevskii Brownian motion, 347 local conformal martingal, 155 local coordinate, 247
-conplex-, 342 local martingale, 52 local time, 113, 220
-of Brownian motion, 113
-space, I universally-, 1 Meyer's theorem (on the equivalence of Soboldv norms), 365 minimal - A-diffusion, 290 -Brownian motion, 211 modification, 266
right continuous-, 33 mollifier, 485
moment inqualities for martingales,
110
motion with random acceleration, 202
multiplicative operator functional (MOF), 318, 321 natural, 35 non-degenerate, 288 non-sticky, 221 normal component (of d -form), 309
normal coordinate mapping, 464 normal derivative, 309 normally reflecting diffusion process, 313, 325, 341
number operator, 356
Lp-multipliplier theorem, 369 Malliavin convariance, 375
Malliavin's stochastic calculus of variation, 349 manifold, 247
complex-, 341
Kahler-, 344
Riemannian-, 281
-with boundary, 308 Markovian system, 204
one-dimensional diffusion process, 446 Onsager-Machlup function, 533 optimal control problem, 441 optional sampling theorem, 26, 34 optional stopping, 25 Ornstein-Uhlenbeck
-'s Brownian motion, 233 -operator, 356 -process, 351 -semigroup, 356
strongly-, 204 Markov time, 22 martingale, 25
conformal-, 155
'tr-system, 22
parallel displacement, 279
stochastic-, 297
554
INDEX
pathwise uniqueness, 162 piecewise linear appoximation, 484 Pitman's theorm, 140 plane section, 461 polynominal functional, 351 point function, 43 point process, 43
characteristic measur of-, 43
(3 )-adapted-, 59
-of Brownian excursions, 125 -of the class (QL), 59 Poisson-, 43, 140
a-finite-, 59 stationary-, 43 Poisson point process, 43
Ricci curvature, 461 Ricci tensor, 302 Riemannian
-connection, 282 -manifold, 281 -metric, 281 scalarization, 280 sectional curvature, 461 semi-martingale, 64
continuous martingale part of-, 64 representaion theorem of-, 84 a-field, I right-continuous-, 20
topological-, I
(,9;)-, 60
Poisson random measure, 42 intensity mesure of-, 43
mean measure of-, 43 predictable, 21,25 probability, I
-space, I -distribution, 2 -law, 2 process
adapted-, 21 bounded-, 21 continuous-, 16 measurable-, 21 -of time change, 102, 197 predictable-, 21 right-continuous-, 21 well-measurable-, 21 proper reference family, 27
pull-back of Schwartz distributions (under a Wiener mapping); 379
skew product, 472 Skorohod equation, 121 Sobolov norm of Wiener functionals, 362
Sobolov space of Wiener functionals, 364,365 smooth functional, 352 sojourn time density, 113 solution, 161
-admitting of explosions, 172 strong-, 163 uniqueness of-, 162, 221 standard measurable space, 13 state space, 203 sticky, 221
non--, 221 stochastic area, 470 stochastic differential, 97, 98 stochastic differential equation, 159 -, Markovian type, 161
-of the jump type, 244 quadratic variational process, 53 quasimartingale, 98 canonical decomposition of-, 98
bounded variation part of-, 98 exponential-, 149 martingale part of-, 98
time homogeneous-, 161 -, time independent Markovian type, 161
-with respect to Poisson point process, 244
-with respect to quasimartingale, 103
random variable, 2
complex-, 2 d-dimensional-, 2
real-, 2
reference family, 20
proper-, 27 reflecting (barrier), Brownian motion, 119, 209, 314
regular
-condiuitnal probability, 13, 15 -submartingale, 38 relatively compact, 7
stochastic integral
ItO's-, 45 -with respect to the Brownian motion, 49
-with respect to a martingale, 55, 57
-with respect to point processes, 59
stochastic line integral, 468 stochastic moving frame, 298, 424 stochastic parallel displacement, 297 stopping time, 22, 204
INDEX Stratonovich integral, 97, 101
strong Markov property -of Brownian motion, 78 -of Poisson point process, 78 strongly Markovian system, 204 strong renewal property, 78 subharmonic, 529, 530 submartingale, 25
regular-, 38 -of class (DL), 35
555
transformation of drift, 192, 225 transition probability, 204 transition semigroup, 290 uniqueness
-in the sense of probability law, 162
-of the regular conditional probability, 13
-of the solutions, 162, 221
successive approximation, 180
universally measurable, I
sum, 77 supermartingale, 25 supertrace, 425 suppprt, 517
variance, I I vector field, 248
symmetric multiplication, 97, 100 symmetrizable, 290
locally-, 291 tangent component (of p-form), 309 tangent space, 247 tangent vector, 247 tensor, 276. 297 tensor field, 277
fundamental-, 281 scalarization of-, 280 terminal point, 203 tight, 7 time change, 102, 197, 225
process of-, 102, 197 torsion free, 281 torsion tensor, 281 total, 206
transformation of Brownian motion, 224
basic-, 280 C° --, 248 system of canonical horizontal-, 280
weakly convergent, 4 WeitzenbSck's formula, 301 well-measurable, 21 Whiteney's imbedding theorem, 253 Wiener
-functional, 349 genesalized-functional, 350, 364
-'s homogeneous chaos 354 mapping, 375
-martingale, 100 -measure, 40 -process, 40 Williams's decomposition theorem, 135, 146
Williams's description of Brownian excursion law, 144