Potential Analysis 3: 339-357, 1994.
© 1994 KluwerAcademic Publishers. Printed in the Netherlands'.
Properties at Infinity of Diffusion Semigroups and
Stochastic Flows via Weak Uniform Covers
XUE-MEI LI
Department of Mathematics, University of Warwick, Coventry CV4 7AL, U.K.
(Received: 5 August 1992; accepted: 19 September 1993)
Abstract. A unified treatment is given of some results of H. Donnelly, P. Li and L. Schwartz concerning
the behaviour of heat semigroups on open manifolds with given compactifieations, on one hand, and the
relationship with the behaviour at infinity of solutions of related stochastic differential equations on the
other. A principal tool is the use of certain covers of the manifold: which also gives a non-explosion
test. As a corollary we obtain known results about the behaviour of Brownian motions on a complete
Riemannian manifold with Ricci curvature decaying at most quadratically in the distance function.
Mathematics Subject Classifications(1991). 31B25, 60H30, 60J50, 60J60.
Key words.Brownian motion, diffusion, heat semigroup, Riemannian manifold, compactification, boundary,
uniform cover.
1. Introduction
Let M be a manifold, {f~, Y , ~ , P} a filtered p r o b a b i l i t y space. A diffusion process
on M is a stochastic process which is p a t h c o n t i n u o u s a n d s t r o n g l y M a r k o v . These
diffusions are often given by solutions to stochastic differential e q u a t i o n s of the
following type:
dx t = X(x,)o dB t + A(xt)dt
(1)
Here B t is a R " valued B r o w n i a n m o t i o n , X is C 2 from M x R " to the t a n g e n t b u n d l e
T M with X ( x ) : R " - - . T~M a linear m a p for each x in M, A is a C 1 vector field on
M. Let u be an M - v a l u e d r a n d o m variable i n d e p e n d e n t of ~'~o- D e n o t e by F~(u) the
s o l u t i o n s t a r t i n g from u, ~(u) the e x p l o s i o n time. T h e n {Ft(x)} is a diffusion for each
x ~ M . A s s o c i a t e d with F t there are also the p r o b a b i l i s t i c s e m i g r o u p Pt a n d the
c o r r e s p o n d i n g infinitesimal g e n e r a t o r ~ .
I. MAIN RESULTS
T h e m a i n aim [17] of this article is to give unified t r e a t m e n t to s o m e of the results
from H. D o n n e l l y , P. Li a n d L. Schwartz. It gives first a p r o b a b i l i s t i c i n t e r p r e t a t i o n
340
X U E - M E I LI
and extends part of the latter. We first introduce weak uniform covers in an analogous
way to uniform covers, which gives a nonexplosion test by using estimates on exit
times of the diffusion considered. As a corollary this gives the known result on
nonexptosion of a Brownian motion on a complete Riemannian manifold with Ricci
curvature decaying at most quadratically in the distance function [14].
One interesting example is that a solution to a stochastic differential equation on
R" whose coefficients have linear growth has no explosion and has the C o property.
Notice under this condition, its associated generator has quadratic growth. On the
other hand let M = R n, and let L be an elliptic differential operator:
L = ~ al;'
~2
b"~
i c?x~---j+ ~. ~c~xi
where aij and bi are C a. Let (sij) be the positive square root of the matrix (a~j). Let
X ~ = Zj s~j c3/c?xj, A = Zj bj ~?/~xj. Then the s.d.e, defined by:
(ItO)
dx t = ~ X i (x~)dB ti + A(x~)dt
has generator L. Furthermore if t(aij)l has quadratic growth and b~ has linear growth,
then both X and A in the s.d.e, above have linear growth. In this case any solution
u, to the following partial differential equation:
~u~ = Lu t
8t
satisfies: ut e Co(M), if u o ~ Co(M) (see next part).
II. P R E L I M I N A R I E S
A diffusion process is said to be a C o diffusion if its semigroup leaves invariant C o (M),
the space of continuous functions vanishing at infinity, in which case the semigroup
is said to have the Co property. A Riemannian manifold is said to be stochastically
complete if the Brownian motion on it is complete, it is also said to have the C o
property if the Brownian motion on it does. A Brownian motion is, by definition, a
where A denotes the
path continuous strong Markov process with generator z/k,
i
Laplacian Beltrami operator. Equivalently a manifold is stochastically complete if
the minimal heat semigroup satisfies: P,I = 1.
One example of a Riemannian manifold which is stochastically complete is a
complete manifold with finite volume. See Gaffney [11]. More generally a complete
Riemannian manifold with Ricci curvature bounded from below is stochastically
complete and has the Co property as proved by Yau [20]. See also Ichihara [14],
Dodziuk [6], L. Karp and P. Li [16], Bakry [2], Grigory/m [12], Hsu [13], Davies
PROPERTIES AT INFINITY
341
[5], and Takeda [19] for further discussions in terms of volume growth and bounds
on Ricci curvature. For discussions on the behaviour at infinity of diffusion processes,
and the C o property, we refer the reader to Azencott [1] and Elworthy [10].
Those papers above are on a Riemannian manifold except for the last reference,
for a manifold without a Riemannian structure, Elworthy [9] following It6 [15]
showed that the diffusion solution to (1) does not explode if there is a uniform cover
for the coefficients of the equation. See also Clarke [4]. In particular this shows that
the s.d.e. (1) does not explode on a compact manifold if the coefficients are reasonably
smooth. See [3], [8]. To apply this method to check whether a Riemannian manifold
is stochastically complete, we usually construct a stochastic differential equation
whose solution is a Brownian motion.
HI. HEAT EQUATIONS, SEMIGROUPS, AND FLOWS
Let M be a compactification of M, i.e., a compact Hausdorff space. We assume M
is first countable. Let h be a continuous function on M. Consider the following heat
equation with initial boundary conditions:
O--j= zSf,
xeM
(2)
f(x,O)=h(x),
x~I
(3)
f(x, t) = h(x),
x e ~?M
(4)
It is known that there is a unique minimal solution satisfying the first two equations
on a stochastically complete manifold, the solution is in fact given by the semigroup
associated with Brownian motion on the manifold applied to h. So the above equation
is not solvable in general. However, with a condition imposed on the boundary of
the eompactification, Donnelly and Li [7] showed that the heat semigroup satisfies
(4). Here is the condition and the theorem:
THE BALL C O N V E R G E N C E CRITERION. Let {x,} be a sequence in M
converging to a point 2 on the boundary, then the geodesic balls ~,(x,), centered at
x, of radius r, converge to 0~ as n goes to infinity for each fixed r.
An example of a manifold which satisfies the ball convergence criterion is R" with
sphere at infinity, but not the compactification of a cylinder with a circle at infinity
added at each end. The one point compactification also satisfies the ball convergence
criterion.
T H E O R E M 1.1 (H. Donnelly and P. Li). Let M be a complete Riemannian manifold
with Ricci curvature bounded from below. The over determined system (2)-(4) is solvable
342
XUE-MEI LI
for any given continuous function h on M, if and only if the ball convergence criterion
holds for ~ .
Notice if the Brownian motion starting from x converges to some point on the
boundary to which x converges, (2)-(4) is clearly solvable. See Section 5 for details.
We would also like to consider the opposite question: Do we get any information
on the diffusion processes if we know the behaviour at infinity of the associated
semigroups? This is true for many cases. In particular for the one point compactification,
Schwartz has the following theorem [18-1, which provides a partial converse to 1-10]:
T H E O R E M 1.2 (L. Schwartz). Let F t be the standard extension of F t to M = M w { ~ },
the one point compactification. Then the map (t, x ) ~ F~(x) is continuous from R + × ]9I
to L°(f~, P, M), the space of measurable maps with topology induced from convergence
in probability, if and only if the semigroup Pt has the Co property and the map t ~ P t f
is continuous from R + to C o (M).
2. Weak Uniform Cover and Nonexplosion
D E F I N I T I O N 2.t [9-1. A stochastic dynamical system (1) is said to admit a uniform
cover (radius r > 0, bound k), if there are charts {~b,, U,} of diffeomorphisms from
open sets U, of the manifold onto open sets ~b,(U,) of R", such that:
1. B3~ c qSi(Ui), each i. (B~ denotes the open ball about 0, radius e).
2. The open sets {q~Tl(Br)} cover the manifold.
3. If (qSi),(X) is given by:
(~i),(X)(v)(e) = (D~i)¢; ,(~,)X(dp~ lv)(e)
with (qS~),(A) simitary defined, then both (~b,),(X) and d(~bl) are bounded by k on
B2r, here d is the generator for the dynamical system.
~i(Ui)
Ui
PROPERTIES AT INFINITY
343
Let {ej} be an orthonormal basis for R.,, and define X j to be X(x)(ej). Then for
Stratonovich equation (1) on M,
d(e)~) = ~
k,l = 1
xkx~(~,) + d4),(A).
Let M = R". If we replace the Stratonovich differential in (1) by the It6 differential, then
1
m
~(~i)(X) -~- 2 k~l=l O2~i(x)(Xk(x)' xl(x)) + O~i(x)(a(x)),
which does not involve any of the derivatives of XJ's.
D E F I N I T I O N 2.2. A diffusion process {Ft, 3) is said to have a weak uniform cover
if there are pairs of connected open sets {U °, U,}, and a non-increasing sequence
{6,} with 6, > 0, such that:
1. U ° c U,, and the open sets {U °} cover the manifold. For x ~ U ° denote by C(x)
the first exit time of F,(x) from the open set U,. Assume z" < ~ unless C = co.
2. There exists {K,},~I, a family of increasing open subsets of M with w K , = M,
such t h a t each U, is contained in one of these sets and intersects at most one
boundary from {~K,,}m~=l.
3. Let x s U ° and U. ~ K m, then for t < 6.,:
P{co:C(x) < t} <~Ct 2
(5)
4. E,~ 1 6, = oe.
Km
Notice the introduction of {K,} is only for giving an order to the open sets {U.}.
This is quite natural when looking at concrete manifolds. In a sense the condition
says the geometry of the manifold under consideration changes slowly as far as the
diffusion process is concerned. In particular if 6, can be taken all equal, we take
K. = M; e.g., when the number of open sets in the cover is finite. On a Riemannian
manifold the open sets are often taken as geodesic balls.
344
XUE-MEI LI
L E M M A 2.1. Assume there is a uniform cover for the stochastic dynamic system (1),
then the solution has a weak uniform cover with 6, = 1, all n.
Proof This comes directly from Lemma 5, p. 127 in [9].
REMARKS.
1. Let T be a stopping time, the inequality (5) gives the following from the strong
Markov property of the process:
let V c U °, and V c K,,, then when t < 6,,:
(6)
P{z"(FT(X)) < tlFT(X)a V} <~ Ct 2,
since
l'
P {z"(Fr(x)) < t, Fr(x ) ~ V} = Jv P(z"(y) < t)Pr(x, dy) <<.Ct2p{Fr(x)
V},
here PT(X,dy) denotes the distribution of Fr(x ).
2. Denote by pV. the heat solution on U, with Dirichlet boundary condition, then
(6) is equivalent to the following: when x ~ U °,
1 - PV"(1)(x) ~< Ct 2
(7)
3. The methods in this article work in infinite dimensions to give analogous results.
EXIT TIMES. Given such a cover, let x e U °. We define stopping times {Tk(x)} as
follows: Let To = 0. Let Tl(x ) = inf{t > 0:Ft(x, to)¢U-£,} be the first exit time of f t ( x )
from the set U,. Then Frl(X, to) must be in one of the open sets {U°}. Let
nl
= {o):Frl(x)(to)e U °, Tl(X, to) < oo}
n,~
=
to:F~(x)e u °
U i ,oT
--
1 <
co
j=l
Then {gl~} are disjoint sets such that ~
= {T1 < oo}. In general we only need to
consider the nonempty sets of such. Define further the following: Let T2 = m, if
T1 = oo. Otherwise if to a gl~, let:
Tz(X , co) = TI(x, to) + zk(FTI(X, CO))
(8)
In a similar way, the whole sequence of stopping times {Tj(x)} and sets {II~}L1
are defined for j = 3, 4 . . . . . Clearly IY~ so defined is measurable with respect to the
sub-algebra ~-rj.
PROPERTIES AT INFINITY
345
L E M M A 2.2. Given a weak uniform cover as above. Let x ~ U ° and U, ~ Km. Let
t < 5m+ k. Then
P{co:Tk(x, co)
- Tk_l(X,(.O ) < t, Tk_ 1 < 00} ~
C t 2.
(9)
P r o o f Notice for such x, FTk(X)~Km+k_ P Therefore for t < 5m+ k we have
P{o~:r~(x, co) - r~_l(x, co) < t , ~ _ , < oo}
j=l
= ~ P{zJ(FT~_,(x))< t,~)~-1}
j=l
.<
ct2
ct2,
j=l
as in R e m a r k 1. Here ZA is the characteristic function for a measurable set A, and E
denotes taking expectation.
•
L E M M A 2,3. I f Z , t, = 0% t, > 0 non-increasing. Then there is a non-increasing
sequence {s,}, such that 0 < s, <~ t,:
(i) Y.s.= oo
(ii) Z s 2 < oo.
P r o o f Assume t. ~< 1, all n. G r o u p the sequence {t.} in the following way:
t 1; t 2 , • • • , t k 2 ; t k 2 + 1 , " • • , t k a ; t k a + 1 • • •
such that 1 ~< t 2 + " " q'- tka ~ 2, 1 <. tk,+l + "'" + tk,+, ~< 2, i /> 2. Let s 1 = tl,
s z = tz/2 . . . . . Sk2 = tkJ2, Sgz + 1 ---- tk2 + 1/3, "'" , Sk~ = tga/3, Sk3+ 1 = tk3 + 1/4 . . . . . Clearly
the s,'s so defined satisfy the requirements.
•
So without losing generality, we m a y assume from now on that the constants {6,}
for a weak uniform cover fulfill the two conditions in the above lemma. With these
established, we can now state the nonexplosion result. The p r o o f is analogous to the
a r g u m e n t of T h e o r e m 6 on p. 129 in [9].
T H E O R E M 2.4. I f the solution F,(x) o f the equation (1) has a weak uniform cover,
then it is complete (nonexplosion),
P r o o f Let x ~ K , , t > 0, 0 < e < 1. Pick a n u m b e r p (may be depending on e and
n), such that ~i=,+l';'"+v~6 i > t, which is possible from: Z ®i=l 6~ = Go. So
346
XUE-MEI LI
P{~(x) < t} ~< P{Tp(x) < t, Tp_ 1 < or}
P { ~kI (Tk(x) - Tk-I(X)) < t' TP-1 < ~ }
P
~< 2 P{T~(x)- Tk_l(x) <e6.+k, Tk_ 1 < o0}
k=l
n+p
~<
Z
k=n
<
2
k=l
Let e ~ 0, we get: P {~(x) < t} = 0.
REMARK. The argument in the above proof is valid if the definition of a weak
uniform cover is changed slightly, i.e. replacing the constant C by C. (with some slow
growth condition) but keep all 6. equal.
To understand more of the weak uniform cover, we look into an example:
EXAMPLE 1. Let {U,} be a family of relatively compact open (proper) subsets of
M such that U, c Un+ 1 and w,~ 1 U. = M. Assume there is a sequence of numbers
{6.} with Y., 5. = o% such that the following inequality holds when t < 5,_ i and
x e U . _ l : P { z V " ( x ) < t} <<.ct 2. Then the diffusion concerned does not explode by
taking {U.+ 1 - U._I, Un -- U._I} to be a weak uniform cover and K. = t7..
3. Boundary Behaviour of Diffusion Processes
To consider the boundary behaviour of diffusion processes, we introduce the following
concept:
D E F I N I T I O N 3.1. A weak uniform cover {U °, U,} is said to be regular (at infinity
for M), if the following holds: let {xj} be a sequence in M converging to £~0M, and
xj~U,,jet° suO~,j,=l, then the corresponding open sets {U,,j}j~=lc {U,}~ converges to
:~ as well. A regular uniform cover can be defined in a similar way.
p
Go
For a point x in M, there are a succession of related open sets {W~ )p=l, which are
defined as follows: Let W~ be the union of all open sets from {U,} such that U °
contains x, and W~ be the union of all open sets from { U. } such that U ° intersects
one of the open sets U.° defining W~. The {Wf } are defined similarly. These sets are
well defined and in fact form an increasing sequence.
PROPERTIES AT INFINITY
347
L E M M A 3.1. Assume F t has a regular weak uniform cover. Let {x,} be a sequence in
M which converges to a point ~ ~ ~M. Then I4(.Xnp converges to ~ as well for each fixed p.
Proof We only need to prove the following: let {Zk} be a sequence Zke W~Vk,then
Zk--O~, as n----~o9. First, let p = 2.
By definition, for each xk, Zk, there are open sets U,° and U°~ such that x k e U nk~
°
z k e U°~ and U,° ~ U°~ ¢ ~ . F u r t h e r m o r e U,~ ~ Y as k---. oe. Let {Yk } be a sequence
of points with Yk ~ U,° c~ U°~. But Yk ~ "2 since U,~ does. So Um~---""2 again from the
definition of a regular weak uniform cover. Therefore Zk converges to 2 as k---, 0%
which is what we want. The rest can be proved by induction.
•
T H E O R E M 3.2. I f the diffusion F t admits a regular weak uniform cover for 3~, with
5, = 6, all n, then the map Ft(--): M--~ M can be extended to the compactification ~i
continuously in probability with the restriction to the boundary to be the identity map,
uniformly in t in finite intervals. (We will say F t extends.)
Proof Take )~e OM and a sequence {x,} in M converging to 2. Let U be a
n e i g h b o u r h o o d of ~ in M. We want to prove for each t:
lim P{co:F~(x.,(o)q~ U, for some s < t} = 0.
n~oo
Since x, converges to 2, there is a n u m b e r N(p) for each p, such that if n > N = N(p),
W.Xpn c U. Let t > 0, choose p such that 2tip < 6. F o r a n u m b e r n > N(p) fixed, we have:
P{oo:F~(x,,e))¢U, for some s < t} c P{(o:F~(x,,e))¢ W~,, for some s < t}
<~ P{o):Tp(x,)(o9) < t, Tp_a(xn) < oo}
<~ ~ P Tk(X.) -- Tk_ l(x.) < p Tk- l(X") <
k=l
Ct 2
<<.-P
Here C is the constant in the definition of the weak uniform cover. Let p go to infinity
to complete the proof.
•
R E M A R K . If5, can be taken all equal, Theorem 2.4, Theorem 3.2 hold if(5) is relaxed to:
P{z'(X ) < t} <~f(t),
for some nonnegative function f satisfying limt~ o f(t)/t = O.
The weak uniform cover condition is not very easy to apply for a general diffusion
on a manifold. F o r a Brownian m o t i o n in a Riemannian manifold it is relatively
simple. We often start with a uniform cover, ref. L e m m a 2.1.
348
XUE-MEI LI
E X A M P L E 2. Let M = R" with the one point compactification. Consider the
following s.d.e.:
(It6)
dx, = X(x,)dB, + A(xt}dt
T h e n if b o t h X and A have linear growth, the solution has the C O property.
Proof There is a well k n o w n uniform cover for this system. See [4], or [9]. A
slight change gives us the following regular uniform cover:
T a k e a c o u n t a b l e set of points {P,},~o c M such that the o p e n sets
O°={z:lz-p,l<lp,I/3},
n=l,2,..,
and Uo° = { z : E z - p o l < 2 }
cover R". Let
U o = {z:lz - Pol < 6; and U, = {z:lz - P,I < Ip, I/2}, for n ¢ 0. Let ~b, be the chart
m a p on U,:
4.(z)
- z -
p.
Ip.l
This certainly defines a uniform cover (for details see E x a m p l e 3 below). F u r t h e r m o r e
if z,---+ oo and z , e U °, then any y e U, satisfies the following:
lyl >
> 5 t z . I - - - , oo,
since ]p,] i> 3]z,I/4. Thus we have a regular uniform cover which gives the required
C o property.
E X A M P L E 3. Let M = R', compactified with the sphere at infinity: M" = R" vo S ' - 1.
Consider the same s.d.e, as in the example above. Suppose both X and A have
sublinear growth of power ~ < 1:
IX(x)[ ~< K(lxf ~ + 1)
IA(x)J <~K({x[ ~ + 1)
for a constant K. T h e n there is no explosion. M o r e o v e r the solution F, extends.
Proof. T h e p r o o f is as in E x a m p l e 2, we only need to construct a regular uniform
cover for the s.d.e.:
T a k e points Po, Pl, P2,... in R" (with [Po] = 1, ]p,] > 1), such that the open sets.{ U ° }
defined by: U o = {z:lz - P0l < 2}, U ° = {z:lz - Pil < IP,[=/6} cover R'.
Let U o = {z:lz - Po[ < 6}, U~ = {z:iz - Pil < IPil'/2), and let ~bi be the chart m a p
from U~ to R":
6(z - Pi)
,~(z) - -
tp~l ~
349
PROPERTIES AT INFINITY
Then {qS~,Ui} is a uniform cover for the stochastic dynamical system. In fact, for
i ~ 0, and y e B 3 c R":
I(q~),(X)(y)t ~< K(I + [q~-a(y)l=)
g
lP~l~
~< i i~(1 + 21p~[~) < 18K
Similarly [(q~),(A)(y)f ~< 18K, and D2q~i = 0.
Next we show this cover is regular. Take a sequence x k converging to 2 in OR".
Assume x k e U °. Let Zk e Uk. We want to prove {zk} converge to ~. First the norm
of z k converges to infinity as k---~ 0% since [Pk] > 21Xk[/3 and [zkt > IP~[ - ½[PY.
z
Let 0 be the biggest angle between points in Uk, then
0
[z-p.1
tan~ ~< sup - ~v° Ip, I
~<
tp.I" Ip.U 1
~
~< - - 2
~0.
Thus { U,, 4), } is a uniform cover satisfying the convergence criterion for the sphere
compactification. The required result holds from the theorem.
This result is sharp in the sense there is a s.d.e, with coefficients having linear growth
but the solution to it does not extend to the sphere at infinity to be identity:
E X A M P L E 4. Let B be a one dimensional Brownian motion. Consider the following
s.d.e, on the complex plane ~:
dx z = ixtdB ~
The solution starting from x is in fact xe ~*+t/2, which does not continuously extend
to be the identity on the sphere at infinity.
4. Boundary Behaviour Continued
A diffusion process is a C O diffusion if its semigroup has the C O property. This is
equivalent to the following [1]: let K be a compact set, and TK(x ) the first entrance
time to K of the diffusion starting from x, then limx_,ooP{TK(x ) < t} = 0 for each
t > 0, and each compact set K.
350
XUE-MEI LI
The following theorem follows from Theorem 3.2 when 6, in the definition of the
weak uniform cover can be taken all equal:
T H E O R E M 4.1. Let ffl be the one point compactification. Then if the diffusion process
Ft(x ) admits a regular weak uniform cover, it is a C O diffusion.
Proof Let K be a compact set with K c Kfi here {Kj} is as in Definition 2.2. Let
e > 0, t > 0, then there is a number N = N(e, t) such that:
t
6 j + 2 -~ 6 j + 4 -F "'" -F" 6 j + 2 N - 2 > - .
Take xCKj+2N.Assume xeKr,, some m > j + 2N. Let TO be the first entrance
time of Ft(x) to Kj+zN_ 1, T1 be the first entrance time of F~(x) to Ki+zN_ 3 after To,
(if To < oo), and so on. But P { Ti < t, T/_ 1 < oo} -%<Ct 2 for t < 6j+zN-2~, i > 0, since
any open sets from the cover intersects at most one boundary of sets from {K, }. Thus
P{TK(x ) < t} <~P
Ti(x) < t, TN_ 2 <
N-1
E P{Ti(x) < e6~+2u-2i, Ti-,(x) < oo}
i=1
N-1
oo
~< Ce2 ~ 6j+~N-21
2
<<. C~2 ~ 6~.
1
1
The proof is complete by letting e---~O.
E X A M P L E 5. Let M be a complete Riemannian manifold, p a fixed point in M.
Denote by p(x) the distance between x and p, B,(x) the geodesic ball centered at x
of radius r, and Ricci(x) the Ricci curvature at x.
A S S U M P T I O N A.
v/~dr
= ~.
(10)
Here K is defined as follows:
K(r)= -Iinf
Ricci(x)A 0}.
I.Br(p)
Let Xt(x) be a Brownian motion on M with Xo(x ) = x. Consider the first exit time
of Xt(x ) from BI(X):
r = inf {t >~ O:p(x, Xt(x)) = 1}.
PROPERTIES AT INFINITY
351
Then we have the following estimate on T from [13]:
If L(x) >
,JK(p(x) + t), then
P {T(x) <.% C~(X)}<<.e-c~L~)
for all x e M . Here cl, c 2 are positive constants independent of L.
This can be rephrased into the form we are familiar with: when 0 ~< t <
Cl/~.g(p(x ) "}- 1),
P{T(x) <<,t} ~ e -( .... /t)
But limt~ o e - cc,c2/o/r2 = 0. So there is a 6 o > 0, such that: e - tc,c2/o ~< t 2, when t < 60. Thus:
E S T I M A T I O N O N EXIT TIMES: when t <
cl/~+
1)/x 60,
P{T(x) < t} ~< t 2.
Let 6, =
(11)
1/~fK(3n + 1) ^ 60, then we also have the following:
-
-
~1 6, >~ 1 x/K(3n + 1)
>/
f;'
dr=oo
(12)
~K(3r)
with this we may proceed to prove the following from Hsu:
C O R O L L A R Y [13]. A complete Riemannian manifold M with Ricci curvature
satisfying Assumption A is stochastically complete and has the C Oproperty.
Proof. There is a regular weak uniform cover as follows:
First take any pEM, and let K , = B3,(p ). Take points p~ such that U ° = Bl(p~)
covers the manifold. Let Ui = B2(pi). Then {U °, Ui} is a regular weak uniform cover
for M u Z X .
R E M A R K . Grigoryfin has the following volume growth test on nonexplosion. The
Brownian motion does not explode on a manifold if there is a function f on M satisfying:
foo
r
Ln (Vol (B R)) dr
=C~O.
Here Vol(BR) denotes the volume of a geodesic ball centered at a point p in M. This
result is stronger than the corollary obtained above by the following comparison
theorem on an n dimensional manifold: let o , _ 1 denote the volume of the n - 1
sphere of radius 1,
352
X U E - M E I LI
VoltB.).
Sinht / 7__lir)
dr
Notice K(R) is positive when R is sufficiently big provided the Ricci curvature is not
nonnegative everywhere. So Grigory/m's result is stronger than the one obtained above.
The definition of weak uniform cover is especially suitable for the one point
compactification. For general compactifications the following definition explores more
of the geometry of the manifolds and gives a better result:
D E F I N I T I O N 4.1. Let !~ be a compactification of M, 2e~?M. A diffusion process
F, is said to have a uniform cover at the point 2, if there is a sequence A, of open
neighbourhoods of 2 in M and positive numbers 6, and a constant c > 0, such that:
1. The sequence of A, is strictly decreasing, with c~A, = 2, and A, D ~A,+I.
2. The sequence of numbers 6, is non-increasing with E 6, = oc and E 6~ < oe.
3. When t < 6,, and x e A, - A, + 1,
p{~A--~(x) < t} < ct ~.
Here ra"(x) denotes the first exit time of F,(x) from the set A,.
P R O P O S I T I O N 4.2. I f there is a uniform cover for ,Y~ ~M, then Ft(x) converges to
,2 continuously in probability, uniformly in t in finite interval, as x----~2.
Proof The existence of {A,} will ensure F~A,,(x) c A,_ 1, which allows us to apply
a similar argument as in the case of the one point compactification. Here we denote
by zA the first exit time of the process Ft(x) from a set A.
Let U be a neighbourhood of 2. For this U, by compactness of M, there is a
2 1/2
number m such that A,. c U, since C~k=l°°A k = ~. Let 0 < e < 1, g = (e/e Z k 6 k) . We
may assume g < 1. Choose p = p(e) > 0 such that:
t
6,. + 6,.+ 1 + "'" + 6,.+p- 1 > =.
Let xEA"+p+2. Denote by To(x) the first exit time of F,(x) from Am+p+ l, TI(x )
PROPERTIES AT INFINITY
353
the first exit time of Fro(x) from A,,+p where defined. Similarly T~, i > 1 are defined.
Notice if Ti(x) < o% then FT,(:,)~A,,,+p_ i - A,,,+p+l_ i, for i = 0, 1,2 . . . . . Thus for
i > 0 there is the following inequality from the definition and the Markov property:
e{r (x) <
<
•
Therefore we have:
P{~U(x) < t} <~P{zA~(x) < t}
e{T +--- + 71 < t,
<
}
P
i=1
P
<~ cg2 y, Om+p-i
2
<~ 8.
i=l
This finishes the proof.
5. Property at Infinity of Semigroups
Recall a semigroup is said to have the C O property, if it sends Co(M ), the space of
continuous functions on M vanishing at infinity, to itself. Let M be a compactification
of M. Denote by /~ the point at infinity for the one point compactification.
Corresponding to the C o property of semigroups we consider the following C,
property for M:
D E F I N I T I O N 5.1. A semigroup Pt is said to have the C. property for .Q, if for each
continuous function f on M, the following holds: let {x,} be a sequence converging
to ~ in ~M, then
lira Ptf(x,,) = f(ff).
(13)
n--* ~9
To justify the definition, we notice if 2~ is the one point compactification, condition
C. will imply the C o property of the semigroup. On the other hand if Pt has the C O
property, it has the C , property for M w/X assuming nonexplosion. This is observed
by subtracting a constant function from a continuous function f on M w A: let
9(x) = f ( x ) - f ( A ) , then g ~ Co(M). So Ptg(x) = P t f ( x ) - f ( A ) . Thus
lira P J ( x n ) = lim Ptg(x,) + f ( A ) = f ( A ) ,
tl-+ c13
n~oO
if lim,~ ~ x, = / ~ ,
In fact the C, property holds for the one point compactification if and only if there
is no explosion and the Co property holds. These properties are often possessed by
354
XUE-MEI Lt
processes, e.g. a Brownian motion on a Riemannian manifold with Ricci curvature
which satisfies (10) has this property.
Before proving this claim, we observe first that:
L E M M A 5.1. If Pt has the C. property for any compactification M, it must have the
C. property for the one point compactification.
Proof Let f e C ( M ~ A). Define a map fi from M to M u A: N x ) = x on the
interior of M, and fl(x) = A, if x belongs to the boundary. Then fl is a continuous
map from b2 to M w A , since for any compact set K, the inverse set
/~ - K = fl- t(M w A - K) is open in _M.
Let g be the composition map o f f with fi: g = f ° fl: .~t--+ R. Thus g(x) I M = f(x) IM,
and g(x)iou = f ( A ) . So for a sequence {x,} converging to 2~8M, lim, P J ( x , ) =
lim, Ptg(x.) = g(2) = f(A).
•
We are ready to prove the following theorem.
T H E O R E M 5.2. If a semigroup Pt has ~he C, property, the associated diffusion process
F t is complete,
Proof We may assume the compactification under consideration is the one point
compactification from the lemma above. Take f - 1 ,
Ptf(x)= P{t < ¢(x)}. But
P{t < ¢(x)} ~ 1 as x--* A from the assumption. More precisely for any e > 0, there
is a compact set K, such that if x ~ K~, P{t < ~(x)} > 1 - e.
Let K be a compact set containing K,. Denote by ~ the first exit time of F,(x) from
K. So F~(x)¢K, on the set {r < oe}. Thus:
P{t <
~(x)} >i P{~ < 0%t < ~(F,(x))} + P{r = oo}
= E{Z~<~E{z,<{(v,(x))I W~}} + P{z = oo}.
Here Za denotes the characteristic function of set A. Applying the strong Markov
property of the diffusion we have:
E{Z,< ~E{zt<e(v~(x))l ~}} = E{Z~<~E{t < ~(y) l F~ -- y}}.
However
E{Z,<¢(r) I F~ = Y} > 1 -- z,
SO
=
1 -
eP{r
<
Therefore P{t < ¢(x)} = 1, since e is arbitrary.
oe}.
•
PROPERTIES AT INFINITY
355
In the following we examine the relation between the behaviour at oe of diffusion
processes and the diffusion semigroups.
T H E O R E M 5.3. The semigroup P~ has the C, property if and only if the diffusion
process F t is complete and can be extended to 37I continuously in probability with
Ft(x) 1aM = x.
Proof Assume F t is complete and extends. Take a point ff s c~M, and sequence {x, }
converging to 2. Thus
lim Ptf(x,) = lim Ef(Ft(x,) ) = Ef(x) = f(x)
n ~
cg)
n -+ oo
for any continuous function on M, by the dominated convergence theorem.
On the other hand Pt does not have the C , property if the assumption above is
not true. In fact, let x, be a sequence converging to 2, such that for some neighbourhood
U of 2, and a number 6 > 0:
limm P{F,(x,)¢ U} = (~.
n--+ oo
There is therefore a subsequence {x,, } such that:
lim P{F,(x,,)~ U} = 6.
Thus there exists N > 0, such that if i > N:
(f
P{Ft(xn,)e2ffl-
U} > ~.
But since 2~r is a compact Hausdorff space, there is a continuous function f from M
to [0,1] such that f t ~ - v = 1, and f(x)[ o = 0, for any open set G in U. Therefore
Ptf(x.) = Ef(Ft(x.) )
>1 ~{~:F~(x,)~pa-v} f(Ft(x"))P(d°°)
= P{f,(x.)e~
So lim Ptf(x,) ~ f(x) = O.
6
- U} > ~.
[]
C O R O L L A R Y 5.4. Assume the diffusion process F t admits a weak uniJbrm cover regular
for M w A, then its diffusion semigroup P, has the C. property Jbr M • A . The same
is true for a general eompactifieation if all 6, in the weak uniform cover can be taken equal.
356
XUE-MEI LI
EXAMPLE 6 [7]. Let M be a complete connected Riemannian manifold with Ricci
curvature bounded from below. Let ~r be a compactification such that the ball
convergence criterion holds (ref. Section 1). In particular the over determined equation
(2)-(4) is solvable for any continuous function f on M if the ball convergence criterion
holds.
Proof We keep the notation of Example 5 here. Let K = -{infxRicci(x ) ^ 0},
6 = c~/k, where c 1 is the constant in Example 5. Let p ~ M be a fixed point, and
K, = ~
be compact sets in m. Take points {Pi} in m such that {B~(pi)} cover
the manifold. Then {Bl(pi),B2(pi)} is a weak uniform cover from (11) and (12).
Moreover this is a regular cover if the ball convergence criterion holds for the
compactification.
Acknowledgements
This paper is developed from my Warwick Msc. dissertation. It is a pleasure to thank
my supervisor Professor D. Elworthy for pointing out the problem and for discussions
on the proofs and for critical reading of the paper. I am grateful to Professor Zhankan
Nie of Xi'an Jiaotong University and for the support of the Sino-British Friendship
Scheme. I also benefited from the EC programme SCI-0062-C(EDB). I also like to
thank Professor I. Chavel for helpful comments.
References
1. Azencott, R.: 'Behaviour of diffusion semigroups at infinity', Bull. Soc. Math, France 102 (1974), 193-240.
2. Bakry, D.: 'Un critrre de non-explosion pour certaines diffusions sur une vari&6 Riemannienne
complrte', C. R. Acad. Sci. Paris 303, Srrie I(1) (1986), 23-26.
3. Carverhill A. and Elworthy, D.: "Flows of stochastic dynamical systems: the functional analytic
approach', Z. Wahrscheinlichkeitstheorie Verw. Gebiete 65 (1983), 245-267.
4. Clark, J. M. C.: 'An introduction to stochastic differential equations on manifolds', in D. Q. Mayne
and R. W. Brockett, eds, Geometric Methods in Systems Theory (1973), pp. 131-149.
5. Davies, E. B.: 'Heat kernel bounds, conservation of probability and the Feller property', preprint (1989).
6. Dodziuk, J.: 'Maximal principle for parabolic inequalities and the heat flow on open manifolds', preprint.
7. Donnelly, H. and Li, P.: 'Heat equation and compactifications of complete Riemannian manifold',
Duke Mathematical Journal 51 (t984), 667-673.
8. Elworthy, K. D.: 'Geometric aspects of diffusions on manifolds', in P. L. Hennequin, ed., Ecole d'gtd
de probabilitds de Saint-Flour XV-XVII, 1985, 1987. Lecture Notes in Mathematics, volume 1362, pages
276-425 (1988), Springer-Verlag.
9. Etworthy, K. D.: Stochastic Differential Equations on Man~olds. Lecture Notes Series 70, Cambridge
University Press, Cambridge, 1982.
10. Elworthy, K. D.: 'Stochastic flows and the co property', Stochastics 0 (1982), I-6.
11. Gaffney, M. P.: 'The conservation property of the heat equation on Riemannian manifolds', Comm.
Pure. Appl. Math. 12 (1959).
12. Grigoryfin, A.: 'On stochastically complete manifolds', Soviet Math. Dokl. 34(2) (1987), 310-313.
13. Hsu, P.: 'Heat semigroup on a complete Riemannian manifold', Ann. of Prob. 17(3) (1989), 1248-1254.
PROPERTIES AT INFINITY
357
14. Ichihara, K.: °Curvature, geodesics and the Brownian motion on a Riemannian manifold II: Explosion
properties', Nagoya Math. J. 87 (1982), 115-125.
15. It6, K.:'On stochastic differential equations on a differentiable manifold', Nagoya Math. J. 1 (1950), 35-47.
t6. Karp, L. and Li, P,: 'The heat equation on complete Riemannian manifolds', preprint.
17. Li, X.-M.: 'Behaviour at infinity of stochastic solutions and diffusion semi-groups', Warwich Msc.
dissertation (1989).
18. Schwartz, L.: 'Le semi-groupe d'une diffusion en liaison avec les trajectories', in Lecture Notes in
Mathematics 1372, pp. 326-342 (1989), Springer-Verlag.
19. Takeda, M.: 'On the conservation of the Brownian motion on a Riemannian manifold', Bull. London
Math. Soc. 23 (1991), 86-88.
20. Yau, S. T.: 'On the heat kernel of a complete Riemannian manifold', J. de Math. Pures Appl. 57(2)
(1978), 191-201.