Untitled
Untitled
Untitled
Editors
M anaging Editors
B. Eckmann S. R. S. Varadhan
Kai Lai Chung
Lectures from
Markov Processes
to Brownian Motion
With 3 Figures
Preface VII
Chapter I
Markov Process
1.1. Markov Property
1.2. Transition Function 6
1.3. Optional Times 12
1.4. Martingale Theorems 24
1.5. Progressive Measurability and the Projection Theorem 37
Notes 44
Chapter 2
Basic Properties
2.1. Martingale Connection 45
2.2. FeUer Process 48
2.3. Strong Markov Property and Right Continuity of Fields 56
2.4. Moderate Markov Property and Quasi Left Continuity 66
Notes 73
Chapter 3
Hunt Process
3.1. Defining Properties 75
3.2. Analysis of Excessive Functions 80
3.3. Hitting Times 87
3.4. Balayage and Fundamental Structure 96
3.5. Fine Properties 106
3.6. Decreasing Limits 116
3.7. Recurrence and Transience 122
3.8. Hypothesis (B) 130
Notes 135
Chapter 4
Brownian Motion
4.1. Spatial Homogeneity 137
4.2. Preliminary Properties of Brownian Motion 144
VI Contents
Chapter 5
Potential Developments
5.1. Quitting Time and Equilibrium Measure 208
5.2. Some Prineiples of Potential Theory 218
Notes 232
Bibliography 233
Index 237
Preface
This book evolved from several stacks of lecture notes written over a decade
and given in classes at slightly varying levels. In transforming the over-
lapping material into a book, I aimed at presenting some of the best features
of the subject with a minimum of prerequisities and technicalities. (Needless
to say, one man's technicality is another's professionalism.) But a text frozen
in print does not allow for the latitude of the classroom; and the tendency
to expand becomes harder to curb without the constraints of time and
audience. The result is that this volume contains more topics and details
than I had intended, but I hope the forest is still visible with the trees.
The book begins at the beginning with the Markov property, followed
quickly by the introduction of option al times and martingales. These three
topics in the discrete parameter setting are fully discussed in my book A
Course In Probability Theory (second edition, Academic Press, 1974). The
latter will be referred to throughout this book as the Course, and may be
considered as a general background; its specific use is limited to the mate-
rial on discrete parameter martingale theory cited in §1.4. Apart from this
and some dispensable references to Markov chains as examples, the book
is self-contained. However, there are a very few results which are explained
and used, but not proved here, the first instance being the theorem on pro-
jection in §1.6. The fundamental regularity properties of a Markov process
having a Feller transition semigroup are established in Chapter 2, together
with certain measurability questions which must be faced. Chapter 3 con-
tains the basic theory as formulated by Hunt, including some special topics
in the last three sections. Elements of a potential theory accompany the
development, but a proper treatment would require the setting up of dual
structures. Instead, the relevant circle of ideas is given a new departure in
Chapter 5. Chapter 4 grew out of a short compendium as a particularly
telling example, and Chapter 5 is a splinter from unincorporated sections
of Chapter 4. The venerable theory of Brownian motion is so well embel-
lished and ramified that once begun it is hard to know where to stop. In
the end I have let my own propensity and capability make the choice. Thus
the last three sections of the book treat several recent developments which
have engaged me lately. They are included here with the hope of inducing
further work in such fascinating old-and-new themes as equilibrium,
energy, and reversibility.
Vlll Preface
Markov Process
T = [0, (0).
Let E be a locally compact separable metric space; and let C be the minimal
Borel field in E containing all the open sets. The reader is referred to any
standard text on real analysis for simple topological notions. Since the
Euclidean space Rd of any dimension d is a well known particular case of
an E, the reader may content hirnself with thinking of Rd while reading
about E, which is not a bad practice in the learning process.
For each tE T, let
and we say that X t is a randorn variable taking values in (E, 0"). For E = R 1,
0" = g&1, this reduces to the farniliar notion of areal randorn variable. Now
any farnily {X t , tE T} is called a stochastic process. In this generality the
notion is of course not very interesting. Special dasses of stochastic pro-
cesses are defined by irnposing certain conditions on the randorn variables
Xt, through their joint or conditional distributions. Such conditions have
been forrnulated by pure and applied rnathernaticians on a variety of grounds.
By far the most irnportant and developed is the dass of Markov processes
that we are going to study.
Borel field is also called <T-field or <T-algebra. As a general notation, for
any farnily of randorn variables {Za, IX E A}, we will denote the (J-field
generated by it by (J(Za, IX E A). Now we put specifically
(1)
Symmetrically, we have
From here on we shall often omit such qualifying phrases as "Vt E T". As
a general notation, we denote by b'1f the dass of bounded real-valued
'§-measurable functions; by Ce the dass of continuous functions on E with
compact supports.
Form (ii) ofthe Markov property is the most useful one and it is equivalent
to any of the following:
(iia) VYEbff;:
E{YIg;;} = E{YIXJ
(iic) Vu 2. t, f E Ce(E):
4 I. Markav Pracess
Lemma 1. For each open set G, there exists a sequence or./imctions [J;,l in Ce
such that
lim i In = IG·
n
(2)
Now apply Lemma 2 to the space E. Let [D be the dass of open sets, C the
dass of sets A satisfying
(3)
where t :s; U I < ... < U n , and!i E bg for 1 :s; j :s; n. For such a Y with n = 1,
(iia) is just (iib). To make induction from n - 1 to n, we write
for some 9 E bg. Substituting this into the above and using the induction
hypothesis with ln-I' 9 taking the place of ln-I, we see that the last term in
(4) is equal to
Then Q E C, [D ce and C has the properties (a) and (b) in Lemma 2. Hence
by Lemma 2, C ::::J O"([D) which is just ff;. Thus (iia) is true for any indicator
Y E .'F; [that is (ii)], and so also for any Y E bff; by approximations. The
equivalence of (iia), (iib), (iic), and (ii), is completely proved.
Finally, (iic) is equivalent to the following: for arbitrary integers n ~ 1
and 0 :s; t l < ... t n < t < u, and I E CAE) we have
(5)
6 1. Markov Process
This is the oldest form of the Markov property. It will not be needed below
and its proof is left as an exercise.
Definition. The collection {PsI,'), 0 ::s; s < t < cx::} is a M arkov transition
function on (E, 0") iff Vs < t < U we have
(a) Vx E E:A --> Ps,,(x,A) is a probability measure on If,
(b) VA E 0":x --> Ps)x,A) is 0"-measurable;
(c) Vx E E, VA E 0":
P,jx,A) = SE Ps"(x,dy)P,,u(y,A).
This function is called [temporally ] homogeneous iff there exists a collection
{P,( . , . ), 0 < t} such that Vs < t, x E E, A E 0" we have
In this case (a) and (b) hold with Ps" replaced by P" and (c) may be rewritten
as follows (Chapman-Kolmogorov equation):
(1)
(2)
1.2. Transition Function 7
Observe that the left side in (2) is defined as a function of (j) (not shown!)
only up to a set of P-measure zero, whereas the right side is a complete1y
determined function of (j) since X, is such a function. Such a relation should
be understood to mean that one version of the conditional expectation on
the left side is given by the right side.
Henceforth a homogeneous Markov process will simply be called a
Markov process.
The distribution J1. of X ° is called the initial distribution of the process. If
0::; t 1 < ... < t n and! E b~n, we have
f f f
= J1.(dx o) P,JXo,dx 1 )' .. P'n-'n_,(Xn-l,dxn)!(X1' ... ,xn) (3)
(4)
(5)
(6)
(7)
(8)
E{ Y ~
er I J"rJ
Ob\ -- EX,)\ Y} . (9)
The relations (8) and (9) follow from (7) by Lemma 2 of §1.1.
Does a shift exist as defined by (6)? If Q is the space of all functions on T
to E: Q = ET , as in the construction by Kolmogorov's theorem mentioned
above, then an obvious shift exists. In fact, in this ca se each co in Q is just the
sampie function X(·, co) with domain T, and we may set
which is another such function. Since X,(co) = X(S, co) the equation (6) is a
triviality. The same is true if Q is the space of all right continuous (or contin-
uous) functions, and such aspace will serve for our later developments. For
an arbitrary Q, a shift need not ex ist but it is always possible to construct a
shift by enlarging Q without affecting the probability structure. We will not
detail this but rather postulate the existence of a shift as part of our basic
machinery for a Markov process.
For an arbitrary probability measure f.l on IJ, we put
contrast to P which is given on :#' => :#,0. Later we shall extend pli to a larger
u-field by completion.
The transition function Pt (·, .) has been assumed to be a strict probabiJity
kernei, namely Pt(x, E) = 1 for every tE T and x E E. We will extend this by
allowing
Pt(X, E) ::; 1, I::ft E T, x E E. (11)
Such a transition function is called submarkovian, and the case where equality
holds in (11) [strictl y] M arkovian. A simple device converts the former to the
latter as follows. We introduce a new 0 f E and put
tff a = u{tff,{o}}.
The new point 0 may be considered as the "point at infinity" in the one-point
compactification of E. If E is itself compact, 0 is nevertheless adjoined as an
isolated point. We now define P; as follows for t > and A E 1&': °
P;(X, A) = Pt(x, A),
P;(x,o) = 1 - Pt(x, E), if x i= D; (12)
I::fw, I::fs ;:::: 0: {X,(w) = o} c {Xt(w) = 0 for all t ;:::: s}. (13)
where, as a standard convention, inf 0 = CfJ for the empty set 0. Thus
((co) = CfJ if and only if Xt(w) =f= (~ for all tE T, in other words Xt((I)) E E for
all t E T. The random variable ( is called the lifetime of the process X.
The observant reader may remark that so far we have not defined P o( ',').
There are interesting cases where P o(x, .) need not be the point mass I;A'),
then x is called a "branching point". There are also cases where P 0(',') should
be left undefined. However, we shall assurne until further notice that we are
in the "normal" case where
10 1. Markov Process
(15)
Before proceeding furt her let us give a few simple examples of (homo-
geneous) Markov processes.
The conditions (a) and (c) in the definition of transition function become in
this case:
(a) Vi E E: L Pij(t) = 1;
jE E
while (b) is trivially true. For the submarkovian case, the" =" in (a) is replaced
by ":::;". If we add the condition
For x E R 1, t 2: 0, we put
1.2. Transition Function 11
Starting from any point x, the process moves deterministically to the right
with uniform speed. This trivial example turns out to be the source of many
counterexamples to facile generalities. A slight modification yields an
example for which (15) is false. Let E = {o} u (- 00, -1] u [1, (0),
P 0(0, { - I}) = 1;
Pt(X,') = 8 x +k), ifx 2:: 1; t 2:: 0;
P t (x,')=8 x -k), ifx::S; -1;t2::0;
Pt(O,') = H81+k) + Li-k)}·
Note that although Po is not the identity, we have P OPt = PtP 0 = Pt for
t 2:: 0.
if m < n,
if m 2:: n.
Note that in this case there is spatial homogeneity, namely: the function of
the pair (n, m) exhibited above is a function of m - n only.
Pt(x, y) =
1 [(y - X)2]
n:::-. exp -
'\/ 2m 2t
A major device to "tarne" the continuum of time is the use of random times.
The idea is very useful even for discrete time problems and has its origin in
considering "the first time when a given event occurs". In continuous time
it becomes necessary to formalize the intuitive notions in terms of various
O'-fields.
We complete the increasing family {ff" tE Tl by setting
eJb-VeJb
,;,- f ' - ,;·7"1"
lET
Recall that the notation on the right side above means the minimal O'-field
including all ff" wh ich is not the same as UIET ff,. Although we shall apply
the considerations below to a Markov process, we need not specify the
family [ff, 1to begin with.
Definition. The function T: Q ..... [0, ex;] is called optional relative to {.il;} iff
The preceding relation then holds also for t =CX; by letting I i:y.J through a
sequence. Define
SElO.I)
It follows from the definition that the family {.~+} is right continuous. Note
the analogy with a real-valued increasing function t ..... f(t).
{T ~ t - ~} E ~t _ ljn) + C ~
and consequently
{T < t} = U Cf) {
T ~ t - -
1} E~.
n=l n
and consequently
{T ~ t} = nN {
T < t +- 1} E 1\
Y]
~ + ljn = ~ +. D
n=l n n=l
EXAMPLES.
2. Suppose,'1'o contains all P-null sets and Vt: P{T = t} = 0. Then we have
Vt:{T~t}-{T<t}E~
3. Consider a Poisson process with left continuous paths and T = first jump
time. Then T is optional relative to {,'1'~+} but not to {,'1'~}, because
{T = t} ~ ,'1'~[ = ,'1'~_].
Sv T, S + T.
14 I. Markov Proccss
Proof. The first two are easy; to treat the third we consider for each I:
where Qt = Q n (0, t). Clearly each member of the union is in .~; hence so
is the union. 0
Definition. If T is strictly optional, :!Fr is the dass of subsets of .'F such that
Let us verify, for example, that .C;;:T+ is a IT-field. We see from (3) that
Q E:FT+ because T is optional. It then follows that :FT+ is dosed under
complementation. It is also dosed under countable union; hence it is a
1.3. Optional Timcs 15
cr-field. For T = a constant t, :#'T reduces to !!i'; , and :#'T+ reduces to !!i';+.
Thus our definition and notation make sense and it is a leading idea beJow
to extend properties concerning constant times to their analogues for op-
tional tim es.
It is easy to see that if {!!i';} is right continuous, then any optional T is
strictly optional, and furthermore we have
(4)
Theorem 5.
(a) If T is optional, then TE :#'T+ ;
(b) If Sand T are both optional and S ::; T, then :#'s+ C :#'T+;
(c) If T n is optional, T n + I ::; T n for each n;:::: 1, and lim n T n = T, then
w
Proof. We leave (a) and (b) as exercises. To prove (c), we remark that T
is optional by Proposition 3. Hence we have by (b):
X;
:#'T+ C 1\
n=1
:#'T n +' (6)
where Qr = Q n (0, t). For each r, {r< T< t} E:Fr, {S < r} E .'Fr C .Y,.
Hence the union above belongs to :Fr and this shows : S < T} E ·?FT +.
Next we have
since for each r < t, {r < Tl = {T:S; rJ" E :Fr + C .Y,. Hence, {S < Tl E .'F~".
Combining the two results we obtain {S < Tl E .?Fs + 1\ .'FTt . Since
Definition. For any function T from Q to [0, x], ·?FT - is the IT-field generated
by .'Fo + and the dass of sets:
As defined above, .?FT _ need not be induded in ff: but this is the case if
TE .?F, as we shall assurne in what folIows. It is obvious that TE .YT -.
If T is optional, then .?Fo + c 'YT; by Theorem 5, (b). Next, for each
u ;::: 0, the set
1.3. Optional Times 17
is empty for u ~ t, and belongs to ff'u if t < u and A E !Fr. Hence each gen-
erating set of ff'T- in (7) belongs to ff'T+, by definition of the latter. In other
words, we have for an optional T:
(9)
(10)
Proof. The first assertion is easy. Under the hypo thesis of the second as-
sertion, we have for any A:
Finally,
An {T = O} = [A n {S = O}] n {T = O} E ff'T-'
( 12)
Ob
,-7"1'_ -VJb
- .7"1',,+' (13)
n=1
Also 9'"0 + C '~1'" _ for each n. Hence each generating set of 9'"1' _ belongs to
the right member of(12), and (12) is proved.1f {Tn } announces the predictable
T, then '~1'- ::J 9'"1',,+ by Proposition 7, and so we have "::J" in (13). Since
9'"1'" + ::J 9'"1'" _ for each n, the reverse inclusion follows from (12). 0
is in !Jß x .~, where (JjJ is the Euclidean Borel field on T. X is said to have
right limits iff for each w, the sampIe function t --> X(t, w) has right limits
everywhere in T; X is said to be right continuous iff for each co, the sampIe
function t --> X(t, w) is right continuous in T. Similarly, "left" versions of
these properties are defined in (0, Cf) ].
(14)
1.3. Optional Times 19
It is clear that
(k
+1 )
(t,w)---> Xn(t,w) = k~O l[k/2".k+l/2")(t)X ~,w
Cf)
for (t, w) E T x Q. Hence Xis Borel measurable; see the Lemma in §1.5 below
for a proof. If Xis left continuous we need only modify the definition in (14)
by omitting the "+ 1" in the right member. 0
(16)
(17)
where X o. = X O·
Note. If T(w) = 00, the left members of(16) and (17) are defined to be zero
by convention, even when X 70 + or XX) _ are not defined.
Proof. Define for n 2': 0:
(18)
This is just the usual dyadic approximation of T from above. For each n, T n
is strictly optional (proof ?) and T n > T, lim n t T n = T. For convenience
let us introduce the "dyadic set"
( 19)
20 I. Markov Process
by Theorem 5.
I[ T is predictable, let {T n } announce T. Then
on {o < T < oo} provided X has left limits in (O,.Xl), and triviallyon {T = O}
provided we define X(O - ) = X(O). Now for each n, we have just shown that
X(T n ) E '~Tn+; hence
by Theorem 8, ifwe note that T E 9'T-' Ifperchance X has left limits in (O,x]
so that X is defined, then we have X(T -) E .?i'T-'
Cj)_ D
We can extend the shift et to eT for any function T from Q to [0, 'l.J 1
Observe that the defining equations for the shift, (6) of§1.2, is an identity in s
and t for each w. It follows that on the set {T <x} we have
(20)
Xt ·" er = X T +t
(21)
1.3. Optional Times 21
So far so good, but the next step is to consider the inverse mapping ():r 1.
This would be awkward if ()T were to be defined on {T < oo} only. Here is
the device to extend it to Q. Put
Vw E Q: X x(w) = o.
Postulate the existence of a point Wo in Q for which
Vw E Q: OAw) = WiJ.
In equation (21) we can now omit the factor 1{T < 00 j' and the result is equiv-
alent to
where u(X t) is the u-field generated by Xt. Since ffT + t + increases with t it
follows by the usual extension that
(22)
(23)
(24)
22 1. Markov Process
and each member of the union above belongs to §',. Hence the set above
belongs to .~ and this proves the theorem. D
What does the random variable in (23) mean? Let A be a set in If and
define
Diw) = inf{t ~ üIX(t,w) E A}. (25)
This is called the "first entrance time into A", and is a prime example of an
optional time. Indeed, it was also historically the first such time to be con-
sidered. The reader may weil recall instances of its use in the classical theory
of probability. However for a continuous time Markov process, it is not a
trivial matter to verify that DA is optional for a general A. This is easy only
for an open set A, and we shall treat other sets in §3.3. Suppose this has been
shown; we can identify the random variable in (23) when S = DA as folIows:
T(w) + DA 8T (w)
0 = inf{t ~ T(w)IX(t,w) E A}. (26)
entrance time into A on or after the first entrance time into B." Of course, a
similar interpretation is meaningful for any optional time T. We shall see
many applications to Markov processes involving such random times.
The definitions and results ofthis section have been given without mention
of probability. It is usually obvious how to modify them when exceptional
sets are allowed. For instance, if each T n is defined only P-a.e., and lim n T"
exists only P-a.e., then there is a set Qo with P(Qo) = 1 on which all Tn's are
defined and lim n T n exists for all w in Qo. Furthermore if {T n ::;; t} differs
from a set in :lF, by a P-null sets, then Qo n {T" ::;; t} E.~ provided that
(g;, P) is complete and 9'"0 ( c .~) is augmented (by all P-null sets). Thus T"
is optional considered on the trace of (Q, 9'") on Qo. In this way we can apply
the preceding discussions to a reduced probability space, and we shall often
do so tacitly. On the other hand, there are deeper results concerning option-
ality in which probability considerations are expedient; see for example
Problem 12 below.
1.3. Optional Times 23
Exercises
1. If T is optional and S = ({J(T) > T where ({J is a Borel function, then S
is strictly optional, indeed predictable. In general a strictly positive pre-
dictable time is strictly optional.
2. :#'T- is also genera ted by :#'0+ and one ofthe two classes ofsets below:
(i) {t:::::; T} (\ A, where 0:::::; t < 00 and A E :?F;-;
(ii) {t< T} (\ A, where 0:::::; t < 00 and A E :?F;+.
3. If Sand Tare functions from Q to [0, 00] such that S:::::; T, then :#'s- C ,'#'T-
if and only if SE :#'T-' This is the case if S is optional. On the other hand
if S is arbitrary and T is optional, :#'s- C :#'T+ may be false.
4. Define for any function T from Q to [0,00] the a-field
7J
Prove that if T is optional, this coincides with the definition given in (3).
5. If T is optional then
%
,'#'T+ = V :#'(TAn)+'
n=l
6. Let S be optional relative to {:?F;} and define '§t = ffs+1' where '§ 0 = .~s+·
Prove that if T ~ S, T is optional relative to {:?F;} if and only if T - S is
optional relative to {'§J
7. If T is predictable and A E :#'T-' then the TA defined in (4) is predictable.
8. Let {T(k)} be a sequence of predictable times. If T(k) increases to T,
then T is predictable. If T(k) "settles down" to T, namely T(k) decreases
and for each w, there exists an integer N(w) far which TN(W)(w) = T(w),
then T is "almost surely predictable". [Hint: for the second assertion
let {T~k)} announce T(k) such that p(T~k) + 2- n < T) < 2- n - k for all
k and n. Let Sn = infk T~k); then {Sn} announces Talmost surely.]
9. If S is optional and T is arbitrary positive, then
:#'s+ (\ {S < T} E ''#'T-;
11. Let {9';} be an increasing family of (J-fields, and C§I be the (J-field generated
by 9'; and all P-null sets. Suppose S is optional relative to {C§J Prove
that there exists T which is optional relative to {.~} and P(S = T) = 1.
Furthermore if A E C§ s then there exists ME .?fCT+ such that P(A D M) = 0.
[Hint: approximate S as in (18); for the second assertion, consider
[SA = T <x} where SA is defined as in (4).J
12. If T is predictable, there exists an "almost surely" announcing sequence
{Tn } such that each T n is countably-valued with values in the dyadic
set. Namely, we have for each n, T n ::;: T n + 1 ::;: T, T n < Ton {T > O},
and P{lim n T n = T} = 1. Furthermore each T n is predictable. [Hint:
let T(k) k J + 1)/2
n = ([2 T n
k n ) >
' Show that there exists lf knJ1 such that P lf T(k
n
-
T i.o.} = 0. Put Sm = infn~m T~,k,,) and show that Sm is dyadic-valued. For
the last assertion use Problems 1 and 8.J
We write 5 H t to mean "5 > t, S ---+ t", and 5 iI t to mean "5 < I, S ---+ t".
Let S be a countable dense subset ofT; for convenience we suppose S :=J N.
(4)
See Theorem 9.4.2 of Course. Observe that this inequality really applies
to a supermartingale indexed by any finite linearly ordered set with m as
the last index. In other words, the bound given on the right side of (4) de-
pends on the last random variable of the supermartingale sequence, not
on the number of terms in it (as long as the number is finite). Hence if we
consider the supermartingale {X t } with the index t restricted to [0, m] n S',
where S' is any finite subset of S containing m, and denote the corresponding
upcrossing number by U([O,m] n S'; [a,b]), then exactly the same bound
as in (4) applies to it. Now let S' increase to S, then the upcrossing number
increases to U([O,m] n S; [a,b]). Hence by the monotone convergence
theorem, we have
f []
E1U(O,m nS; a,b)
[]}
s b-a+ a <00.
E(X;;;) (5)
This is true for each mE N and each pair of rational numbers (a, b) such
that a < b. Therefore (6) holds simultaneously far all mE N and all such
26 1. Markov Proccss
pairs in a set D o with P(D o) = 1. We claim that for each W E D o, the limit
in (2) must exist. lt is sufficient to consider the right limit. Suppose that for
some t, !imsES . .,w X(s, w) does not exist. Then we have firstly
But this means that the sam pie function X(·, w) crosses from (/ to h infinitely
many times on the set S in any right neighborhood of t. Since wEDa this
behavior is ruled out by definition of Da.
To prove the second assertion of the theorem, we recall the following
inequality for a discrete parameter supermartingale (Theorem 9.4.1 of
Course). For any ;. > 0:
p{'E sup
[O.m]nS
IX,I = X;} = 0, or p{'E sup
[U.m]nS
IX,I <x} = I.
o
We shall say "almost surely" as paraphrase for "for P-a.e. (1)".
function X(·, ) is right continuous but does not have a left limit at some t,
then it must oscillate between two numbers a and b infinitely many times in
a left neighborhood of t. Because the function is right continuous at the
points where it takes the values a and b, these oscillations also occur on the
countable dense set S. This is almost surely impossible by the theorem. 0
A+a
E{U(S; [a,b])} ~ b _ a' (8)
where A = 0 under (a), Adenotes the sup under (b). Now let U(w; [a,b])
denote the total number of upcrossings of Ca, b] by the unrestricted sampIe
function X(-,w). We will show that if X(-,w) is right continuous, then
For ifthe right member above be :::O:k, then there exist SI < t l < ... < Sk < t k
such that X(Sj' w) < a and X(t j, w) > b for 1 ~ j ::;; k. Hence by right con-
tinuity at these 2k points, there exist S'I < (1 < ... < s~ < t~, all members
of S such that 51 < S'I < t l < (1 < ... < Sk < s~ < t k < t~ such thaiX(sj, w) < a
and X(tj, w} > b for 1 ::;; j ::;; k. Thus the left member of (9) must also be :::0: k.
Of course it cannot exceed the right member. Hence (9) is proved. We have
therefore
(10)
X t 2 E{XIJ~}. (13)
Now the supermartingale [XJ with s in the index set {t, ... ,t", ... ,tl) is
uniformly integrable; see Theorem 9.4.7(c) of Course. Hence we obtain (11)
by letting n -->XJ in (13). Next, let Sn E S, Sn Ii t; then
Letting 11--> X we obtain (12) since :Fs .. i .~_; here we use Theorem 9.4.8,
(ISa) of Course. Finally, let U n > U > t n > C, U" E S, t n E S, Un 11 u, t" H t, and
A E :ll';+. Then we have
set. We shall not delve into this question but proceed to find a good version
for a supermartingale, under certain conditions.
From now on we shall suppose that (Q,:F, P) is a complete probabi!ity
space. A sub-<T-field of :F is called augmented iff it contains all P-null sets
in:F. We shall assurne for the increasing family {~} that:Fa is augmented,
then so is ~ for each t. Now if {r;} is aversion of {X t }, and X t E ~, then
also Y; E ~. Moreover if {XI>~} is a (super) martingale, then so is {y;, ~},
as can be verified trivially.
where the second equation follows from the uniform integrability of { Y;J
and the right continuity of Y. Hence (15) is true.
Conversely suppose (15) is true; let S be a countable dense subset of T
and define X t + as in (2). Since X t + E ~+ = ~, we have by (11)
P-a.e. (16)
(17)
where the first equation follows from uniform integrabi!ity and the second
from (15). Combining (16) and (17) we obtain
\:It: P(X t = X t +) = 1;
namely {X t+} is aversion of {X t}. Now for each w, the function t --> X(t +, w)
as defined in (2) is right continuous in T, by elementary analysis. Hence
{X t +} is a right continuous version of {Xt}, and {Xt+,~} is a supermar-
tingale by a previous remark. This is also given by Theorem 2, but the point
is that there we did not know whether {X t +} is aversion of {Xt}. D
JE(YI
l %) , d
.Tt :$1
tf (18)
Proof. Let us first observe that condition (19) is satisfied with Y == 0 if X t ;:::. 0,
Vt. Next, it is also satisfied when there is equality in (19), namely for the dass
of martingales exhibited in (18). The general case of the theorem amounts to
a combination ofthese two cases, as will be apparent below. We put
Then Sn and T n are optional relative to the family {$'(k/2 n), k E N}, indeed
strictly so, and Sn ::; T n. Note that X(Sn) = X(S) = X 00 on {S = oo}; X(T n) =
X(T) = X 'Xl on {T = oo}. We now invoke Theorem 9.3.5 of Course to obtain
(21)
Since 9's+ = A:'= 1 $'Sn+ by Theorem 5 of§1.3, this implies: for any A E $'s+
we have
(22)
Letting n --+ 00, then X(Sn) --+ X(S) and X(T n) --+ X(T) by right continuity.
Furthermore {X(Sn), $'(Sn)} and {X(Tn), $'(Tn)} are both supermartingales
on the index set N = {... ,n, ... , 1, O} which are uniformly integrable. To
see this, e.g., for the second sequence, note that (21) holds when Sn is replaced
=
by T n- 1 since T n- 1 Z T n; next if we put Sn 0 in (21) we deduce
L X(S)dP z L
X(T)dP. (23)
The truth of this relation for each A E $'s+ is equivalent to the assertion (b).
The final sentence of the theorem is proved by considering { - X t} as weB as
{XI}' Theorem 4 is complete1y proved.
32 1. M arkov Process
(24)
Now apply Theorem 4(b) with equality and with T for S, CD for T there. The
resuIt is, for any option al T:
since :#'T+ C :#'n. We may replace ''#'T+ by .'1'r above since {.~] is right
continuous by hypothesis.
Next, let T be predictable, and {T n } announce T. Recalling that [XI} has
left limits by Corollary 1 to Theorem 1, we have
(26)
(28)
It is easy to show that these last relations do not hold for an arbitrary optional
T (see Example 3 of ~1.3). In fact, Dellacherie [2] has given an example in
which X T- is not integrable for an optional T.
The next theorem, due to P. A. Meyer [1], may be used in the study of
excessive functions. Its proof shows the need for deeper notions of mea-
surability and serves as an excellent introduction to the more advanced
theory.
(29)
Proof. The fact that {XI' g;;} is a positive supermartingale is trivial; only the
right continuity of t -+ X, is in question. Let D be the dyadic set, and put for
each t ~ 0:
Next let T be any optional time and T k = ([2 k T] + 1)j2 k • Then we have by
Theorem 4,
(31 )
Since X(Td converges to Y(T) by the definition of the latter, it follows from
(31) and Fatou's lemma that
Y(T(w) + 6j , w) - X(T(w) + 6j , w) ~ D.
because X 2': x(n). Since both t -> Y(t, W) and t -> x(n)(t, w) are right con-
tinuous the inequality above implies that
as !5 j --+ 0. Note that the case where all ()j = 0 is included in this argument.
Letting n ->X) we obtain
Therefore (34) holds on the set {T < .XJ}, and consequently (32) is possible
only if P {T <X)} = 0. Since <; is arbitrary, this means we have alm ost surely:
Together with (30) we conclude that equality must hold in (35), namely that
P{Vt 2': 0: Y; = Xt] = 1. The theorem is proved since t -> Y(t,(!)) is right
continuous for (J) E Qo· 0
y(n)
t
= (x(n)
t
_ E lf X(l)
u
I !F})
t
1\ k
~
is right continuous alm ost surely in [0, u]. Now let first k iX), then u iY.::.
An important complement to Theorem 5 under astronger hypothesis will
next be given.
Theorem 6. U nder the hypotheses oJ Theorem 5, suppose Jurther that the index
set is [O,X)] and (29) holds for t E [0, oc]. Suppose also that Jor an.\' sequence
oJ optional times {Tn } increasing to Talmost surely, we have
8prlT'
n
< 00 1.J <
- E{X(T e
n
) - Xn
(Te).
"'
Ten < 00 1.J
Note that for any optional T, E {X(T)} ::;; E {X(O)} < 00 by Theorem 4. By
the same theorem,
Thus
The condition (36) is satisfied for instance when all sam pie functions of X
are left continuous, therefore continuous since they are assumed to be right
36 1. M arkov Process
When this is the case, the superpotential is said to be "of dass D"-a nomen-
dature whose origin is obscure. This dass plays a fundamental role in the
advanced theory, so let us prove one little result about it just to make the
acquaintance.
Proof. If T n --> 00, then X T n --> 0 almost surely; hence uniform integrability
of {X TJ implies (39). To prove the converse, let nE N and
r
JtXT:o>n}
X T dP ~ r
J{T:o>T n }
X T dP = r
Ja
X s dP ~ r
Ja
X T dP
n
1.5. Progressive Measurability and the Projection Theorem 37
°
where the middle equation is due to X 00 == and the last inequality due to
Doob's stopping theorem. Since the last term above converges to zero as
n -+ 00, {X T:T E 0} is uniformly integrable by definition. 0
°
For t :::::: let Tl = [0, t] and ggl be the Euclidean (classical) Borel field on
Tl' Let {'~l} be an increasing family of u-fields.
(s, w) -+ X(s, w}
restricted to Tl X Q is measurable ggl x .?r for each t :::::: 0. This implies that
Xl E.?r for each t, namely {Xl} is adapted to {.?r} (exercise).
T(w) T
Thus the introduction of such times is analogous to the study of plane geom-
etry by means of curves- a venerable tradition. This apparent analogy has
been made into a pervasive concept in the "general theory of stochastic
processes;" see Dellacherie [2] and Dellacherie-Meyer [1].
According to the new point of view, a function X(·, .) on T x Q which is
in:!J x .~ is precisely a Borel measurable process as defined in §1.3. In partic-
ular, for any subset H of T x Q, its indicator function 1H may be identified
38 1. Markov Process
Proof. We will prove the result for right continuity, the left case being similar.
Fix t and define for n 2:: 1 :
x(n)(s, (1)) k + 1 t,
X ( -----y;- ) . [k k+l) < k < 2" - l'
2"'t -2"- t ,0-
= (J) If S E -
- ,
{(s,w)lx(n)(S,W)EB} =
k k+
2" - 1 ( [
kVOyt,-----y;-t 1) x {wlX (k + 1 ) })-'2/11,(1) EB
Lemma. Let si be any u-field of subsets of T x Q. Suppose that X(II) E .01 fOI"
each n and X(II) ---> X everywhere. Then X E .01.
Proof. Let IC be the dass of sets B in g such that X - '(B) E .01. Let G be open
and G = U:'=l Fk where Fk = {xld(x,GC) 2:: l/k] an d is a metric. Then we
have by pointwise convergence of x(n) to X:
X-'(G) =
x
UU
k=l n=1 m=n
n (x(m))-'(F k )·
1.5. Progressive Measurability and the Projection Theorem 39
This belongs to si, and so the dass IC contains all open sets. Since it is a
O'-field by properties of X-I, it contains the minimal O'-field containing all
open sets, namely tff. Thus X E si /tff. 0
The next result is an extension of Theorem 10 of §1.3.
(I)
{w T( w) E A; w E A}
J = {T E A} n A E :?l's + c g;;.
(2)
\. H
w --------------------------~~_____
It is clear that
[Observe that ifwe replace [0, t) by [0, t] in the second term above we cannot
conclude that it is equal to the third term with " < " replaced by ":-s; ".] By
the theory of analytical sets (see Dellacherie-Meyer [1], Chapter 3) the
projection of each set in !!J t x :F; on Q is an "fft-analytic" set, hence .~
measurable when (Q,:F;, P) is complete. Thus for each t we have {D/J < t} E
:F;, namely D H is optional. 0
1.5. Progressive Measurability and the Projection Theorem 41
But X(T l ) = °
on (TI <oo} by the definition of TI and right eontinuity
of paths. Henee the right membcr of (5) is also equal to zero. Sinee X ~ °
this implies that
P{X(T I + t) = °
for all tE Q (\ [0,00); TI< oo} = P{T I < CD}.
We may omit Q in the above by right eontinuity, and the result is equivalent
to the assertion of the theorem when T is replaeed by Tl.
The situation is different for T 2 beeause we ean no longer say at onee
that X(T z) = 0, although this is part of the desired eonclusion. It is eon-
eeivable that X(t) --> as t ° n
T 2 but jumps to a value different from at
Tz. To see that this does not happen, we must make a more detailed analysis
°
of sam pie funetions of the kind frequently needed in the study of Markov
proeesses later. We introduee the approximating optional times as folIows:
(6)
Sinee the left member does not exeeed I In, letting n -->CXJ we obtain
E {X(S), S <:1)} = 0.
The rest is as before. D
1.5. Progressive Measurability and the Projection Theorem 43
We used the fact that {S < oo} c nn{Sn < oo}. The reverse indusion is
not true!
Exercises
1. Let {XI'~} be a martingale with right continuous paths. Show that
there exists an integrable Y such that X t = E( Y I~) for each t, if and only
if {X t} is uniformly integrable.
2. Show that if T is the firstjump in a Poisson process (see Example 3 of §1.3),
then E{XTI%T-} =1= X T-·
3. In the notation ofthe proof ofTheorem 6 of§l.4, show that Xl n ) converges
to X t uniformly in each finite t-interval, P-a.e., if and only if for each
°
c > we have P{lim n T~ = oo} = 1.
4. If(t,w)--+ X(t,w) is in.0B x %0 then for each A E tff', the function (t,x)--+
Pt(x, A) is in gg x tff'. Hence this is the case if {X t} is right [or left] con-
tinuous. [Hint: consider the dass of functions ((J on T x Q such that
(t,x)--+P{((J} belongs to fJ6xtff'. It contains functions of the form
IB(t)lA(w) where BE gg and A E %0.]
5. If {X t } is adapted to {~}, and progressively measurable relative to
{~+ ,} for each 8 > 0, then {X t } is progressively measurable relative to
f%}
l t·
similarly for [eS, T]], ((S, T)), ((S, TJ]. Show that all these four sets are
progressively measurable relative to {~}.
The (J-field generated by [eS, T)) when Sand T range over all optional
times such that S S T is called the optional field; the (J-field generated by
[eS, T)) when Sand T range over all predictable times such that S S T is
called the predictable field. Thus we have
These are the fundamental (J-fields for the general theory of stochastic
processes.
44 I. Markov Process
NOTES ON CHAPTER 1
§1.1. The basic notions of the Markov property, as weil as optionality, in the discrete
parameter case are treated in Chapters 8 and 9 ofthe Course. A number ofproofs carry
over to the continuous parameter case, without change.
§1.2. Among the examples of Markov processes given here, only the case of Brownian
motion will be developed in Chapter 4. But for dimension d = 1 the theory is somewhat
special and will not be treated on its own merits. The case ofMarkov chains is historically
the oldest, but its modern development is not covered by the general theory. It will only
be mentioned here occasionally for peripheral illustrations. The c1ass of spatially
homogeneous Markov processes, sometimes referred to as U:vy or additive processes,
will be briefly described in §4.1.
§1.3. Most of thc material on optionality may be found in Chung and Doob [1] in
a more general form. For deeper properties, which are sparingly used in this book, see
Meyer [1]. The latter is somewhat da ted but in certain respects more readable than the
comprehensive new edition which is Dellacherie and Meyer [1].
§1.4. It is no longer necessary to attribute the foundation of martingale theory to
Doob, but his book [1] is obsolete especially for the treatment ofthe continuous param-
eter case. The review here borrows m uch from M eyer [1] and is confined to later nceds,
except for Theorems 6 and 7 which are given for the sake of illustration and general
knowledge. Meyer's proof of Theorem 5 initiated the method of projection in order to
establish the optionality of the random time T defined in (33). This idea was later
developed into a powerful methodology based on the two CT-fields mentioned at the
end ofthe section. A very c1ear exposition ofthis "general theory" is given in Dellacherie
[2].
A curious incident happened in thc airplane from Zürich to Beijing in May of 1979.
At the prodding ofthe author, Doob produced a new proof ofTheorem 5 without using
projection. Unfortunately it is not quite simple enough to be inc1uded here, so the
interested reader must await its appearance in Doob's forthcoming book.
Chapter 2
Basic Properties
It follows from (1) that its right member is a decreasing function of t, hence
limit in (2) always exists.
Hence
(t, x) -+ PJ(x)
Definition. Let fE hJf, rx > 0, then the rx-potential of f is the function Uaf
given by
Uaf(x) = So' e-'tPJ(x)dt
( 5)
U' is also defined for fE Jf + but may take the value + oc. The family of
operators { U', rx > O} is also known as the "resolvent" of the semigroup (Pt),
or by abuse of language, of the process (X t ). We postpone a discussion of
the relevant facts. For the present the next two propositions are important.
Proof We have
2.1. Martingale Connection 47
(6)
Proof. Let
We have
Hence we have
r
JIIE(YI(1)12:/l1
IE(YI~§)ldP<
-
r
Jlllc(YIWll2:/I}
IYldP',
1
p{IE(YI~§)1 ~ 11) ::;; - E(IYI) -> O.
11
Hence {r;} is uniformly integrable by (8). Finally, the first term on the right
side of (7) is continuous in t, hence it is progressively measurable as a process.
On the other hand since U a! E Cf, it is dear that [U'!(X t )} as a process is
progressively measurable if {Xt} is, relative to {.~}. D
Then {At} is an increasing process, namely for each w, At(w) increases with
t; and Ar. = Y x by definition of the latter. The decomposition given in (6),
together with (8), may be rewritten as:
'-"U'[(X)
i: . t -- EJA
( y
I .Tt!
Vb 1. - A t· (9)
IIIII = SLIp
XEE(-
II(x)l·
2.2. Feiler Process 49
then limx~r1 fo(x) = O. Let 1C0 denote the subclass of IC vanishing at ri ("van-
ishing at infinity"); lC e denote the subclass of 1C0 having compact supports.
Recall that IC and 1C0 are both Banach spaces with the norm 11 11; and that 1C0
is the uniform closure (or completion) of lC e .
Let (Pt) be a submarkovian transition semigroup on (E,6"). It is extended
to be Markovian on (Ei;' 6" a), as shown in §1.2. The following definition applies
to this extension.
lt turns out that the condition in (1), which requires convergence in norm, is
equivalent to the apparently weaker condition below, which requires only
pointwise convergence:
(ii') '1f E IC, x E E,<
lim PJ(x) = fex). (2)
t~O
Proof. Consider (t, x,f) fixed and (s, y, g) variable in the inequality below:
by the Markov property. Since PJg E I[ and PJg ---> g, we have by bounded
convergence, as b 1 0:
(3)
Take h to be ametrie of the space Ea. Then the limit above is egual to zero,
and the result asserts that X tH converges to X t in probability. Next if
o < b < t, then we have for each x:
e{f(Xt-J)g(X t)} = e{f(Xt_ö)Ex,-o[g(XJ)]J
(4)
= e{f(Xt-tl)P"g(X t - o)} = P,-oUP"g)(x).
hence also
E(Y) = EI'(Y).
Given any probability measure f.1 on S, and the transition function {P t (',·),
tE T}, we can construct a Markov process with f.1 as its initial distribution,
as reviewed in §1.2.1t is assumed that such a process exists in the probability
space (Q, ff, P). Any statement concerning the process given in terms of
P and E is therefore ipso facto true for any initial f.1, without an explicit
claimer to this effect. One may regard this usage as an editorial license to
save print! For instance, with this understanding, formula (3) above contains
its duplicate when E there is replaced by EX, for each x. The resulting relation
may then be written as
which may be easier to recognize than (3) itself. This is what is done in (4),
since the corresponding argument for convergence as r5 ! 0, under E instead
of EX, may be less obvious.
On the other hand, suppose that we have proved a result under p x and
EX, for each x. Then integrating with respect to f.1 we obtain the result under
pI' and EIl, for any f.1. Now the identification of P(A) with PI'(A) above shows
that the result is true under P and E. This is exactly what we did at the end
of the proof of Proposition 2, but we have artfully concealed f.1 from the
view there.
To sum up, for any A E ffo, P(A) = 1 is just a cryptic way of writing
Vx E E: PX(A) = 1.
We shall say in this case that A is almost sure. The apparently stronger state-
ment also follows: for any probability measure f.1 on S, we have PI'(A) = 1.
After Theorem 5 of §2.3, we can extend this to any A E ff -.
52 2. Basic Properties
The next proposition concerns the IX-potential Var. Let us re cord the
following basic facts. If (Pt) is Fellerian, thcn
(5)
Vf E C: lim C(--+'Xl
IIIXV~f - fll = O. (6)
(7)
Prooj'. For any x -:f. y, there exists On such that x E On and Y ~ 0". Hence
«Jn(x) = 0 < «Jn(y), namely <(Jn separates x and y. We have by (7), for suffi-
ciently large IX E N:
IIXV'qJn(X) - qJn(X)1 < !«Jn(Y),
IIXV'<{Jn(Y) - qJn(y)1 < !«Jn(Y)·
Hence IXV'qJn{X) -:f. IXV'qJn(Y), namely V'qJn separates x and y. [Of course
we can use (6) to get IIIXV'qJn - qJnll < !qJn(Y)·] 0
2.2. Feiler Process 53
restriction to S.
Proof. Suppose for some t, his does not have a right limit at t. This means
that there exist t n E S, t n H t, t~ E S, t~ !! t such that
There exists gE [J) such that g(x) #- g(y). Since 9 is continuous, we have
This contradicts (9) and proves the proposition since the case of the left
limit is similar. 0
Proof. Let 9 be a member of the dass [J) in (8), thus 9 = Ukep where k E N,
ep E C. By Proposition 2 of §2.1,
restricted to S has right and left limits as asserted. Clearly the factor e- kt
may be omitted in the above. Let Q* = ()9E[D Qg, then P(Q*) = 1. If W E Q*,
then the preceding statement is true for each 9 E [J). Since [J) separates points
54 2. Basic Properties
Theorem 6. Suppose that each ff, is augmented. Then each of the processes
{X t } and {X t } is aversion of {X,}; hence it is a Feiler process with the same
transition semigroup (PJ as {X,}.
(11)
E} = P,(x, E) = 1 for each x and t, it does not follow that we have X(t, w) E E
for all tE T for any w. In other words, it may not be bounded in each finite
interval of t. The next proposition settles this question for a general sub-
markovian (PJ
Theorem 7. Let {X" ff,} be a Feiler process with right continuous paths having
lejt limits; and let
Then we have almost surely X(( + t) = a for all t ~ 0, on the set {( <J.J}.
2.2. Feller Process 55
Then <p vanishes only at C. Since <p E C, we have Pr<p ---> <P as t 1 0, and
consequently
Corollary. l{ (Pt) is strictly Markovian, then almost surely the sampie function
is bounded in each finite t-interval. N amely not only X(t, w) E E .f(Jr all t ~ 0,
°
but X(t-,w) E Efor all t > as weil.
Pg< t} = Pg < t; X r E E} = 0.
Exercises
1. Consider the Markov chain in Example 1 of §1.2, with E the set of positive
integers, no ci, and (f the discrete topology on E. Show that iflimlj 0 Pii(t) =
1 uniformly in all i, then (Pt) is a Feiler semigroup. Take aversion which is
right continuous and has left limits. Show that for each sam pie function
56 2. Basic Propcrties
X(·, w), there is a discrete set of tat which X(t -, w) =1= X(t, (1)), and X(', (1))
is constant between two consecutive values of I in this set.
2. Show that the semigroups of Examples 2, 3 and 4 in §1.2 are all Fellerian.
3. For a Markov chain suppose that there exists an i such that lim t10 [I -
Pii(t)J/t = + w. (Such astate is called instantaneous. It is not particularly
easy to construct such an example; see Chung [2J, p. 285.) Show that for
any 6 > 0 we have pi( X(t) = i for all tE [0, 6J] = O. Thus under pi, no
version of the process can be right continuous at t = O. Hence (Pt) cannot
be Fellerian. Prove the last assertion analytically.
4. Prove that in the definition of Feiler property, condition (2) implies
condition (I). [Hint: by the Riesz representation theorem the dual space
ofC is the space offinite measures on Er. Use this and the Hahn-Banach
thorem to show that the set of functions of the form (Uacp}, where rx > 0
and cp E C, is dense in C. Iff belongs to this set, (I) is true.J
(1)
Proof. Observe first that on {T = w}, both sides of (1) reduce to f((} Let
.Q?;
'/'T+ =
1\ %
'''"1',,'
n=l
Since both fand PJ are bounded continuous, and X(·) is right continuous,
we obtain by letting n ---> OC! in the first and last members of (2):
The truth ofthis for each A E :FT + is equivalent to the assertion in (1). 0
which reduces to :F; when T = t. Then we have for each integrable Y in :F~:
This is the extension of(iia) in §1.1. Alternatively, we have for each integrable
Y in ffo:
(3)
Definition. The Markov process {Xt,~} is said to have the strong Markov
property iff (3) is true for each optional time T.
Proof. Recall that T + t is optional, in fact strictly so for t > 0; and XT+' E
.Y'1+t+ by Theorem 10 of ~1.3. Now apply (1) with T replaced by T + t:
(4)
This proves the first assertion of the theorem. As for the second, it is a con-
sequence of the re-statement of Theorem 1 because [X T+" .9'1+,+] IS a
Feiler process with right continuous paths. 0
Theorem 3. Let S :::::: T and S E ''#'T +. Then we haue ji)r each f E IC:
(5)
The quantity in the right member of(5) is the value ofthe function (t, x)-->
e{f(X t )} = PJ(x) at (S(w) - T(w), XT(w)), with (!) omitted. It may be
written out more explicitly as
(6)
Proof of Theorem 3. Observe first that (5) reduces to fra) = f(?) on [S =XJ J
by the various conventions, so we may suppose S < x in wh at folIows. Let
U
[2"(S - T)]
=-----~-.
+1
" 2" '
then U" 1 (S - T). Since Sand T belong to ·'J7T ~, so does U" for each I!.
Consequently. for each d E D (the dyadic set) we have {U" = d} E .1'/+. Now
let A be any set in .9'/+, and put A d = A n {U" = d] E .1'1+' We have
for each f EI[:
L fex T+u.)dP= L
dED
L d
fex T+d)dP
= L L Pdf(XT)dP =
dED d
L uJ(X
p 1 )dP (7)
2.3. Strong Markov Property and Right Continuity of Fields 59
where the second equation is by (1). Now t --* PJ(x) is continuous for each
x andf E C by Theorem 1 of§2.2. Letting n --* 00 in the first and last members
of (7), we obtain by right continuity:
namely '§* is the minimal O'-field induding '§ and .H. '§* is called the aug-
mentation of,§ with respect to (Q,:?, P). It is characterized in the following
proposition, in which a "function" is from Q into [ - 00, + 00]. Recall that
for a function Yon E, Y E '§* is an abbreviation of Y E '§* /t! where t! is the
Borel field of E.
Lemma. A subset A of Q belongs to '§* if and only if there exists B E '§ such
that A /::" BE JV'. A function Y belongs to '§* if and only !f there exists a
function Z E '§ such that {Y =f. Z} E .H.
Proof. Let .S!I denote the class of sets A described in the lemma. Observe
that A L. B = C is equivalent to A = B /::" C for arbitrary sets A, Band C.
Since '§* contains all sets of the form B L. C where BE '§ and CE .;(1', it is
clear that '§* ::::J .91. To show that .91 ::::J '§* it is sufficient to verify that .91 is a
O'-field since it obviously includes both '§ and .H·. Since AC LI B C= A L1 B,
.r4' is closed under complementation. It is also closed under countable union
by the inclusion relation
and the remark that if the right member above is in .AI, then so is the !eft
member. Thus .si is a O'-field and we have proved the first senten ce of the
Lemma.
Next, !et Z E '§ and {Y =f. Z} E .H. Then for any subset S of [ - 00, + 00],
we have the trivial inclusion
{YE S} L. {Z E S} c {Y =f. Z}.
60 2. Basic Properties
Remark. If (Q, .?l", P) is not complete, the Lemma is true provided that we
assume Y E ff in the second assertion. We leave this to the interested reader.
(a) ZE~§;
(b) for each W E b~l}, we have
Note that (8) is then also true for W E hctj*. It follows that if Z is a mem ber of
the dass E( Y IctJ), and Z* is a member of the dass E( Y IctJ*), then {Z i= Z*} E
.N·. We state this as folIows:
We have seen that there is advantage in augmenting the a-fields ~~+ appear-
ing in (10). Since ~?+ ~ ~? ~ ~g, it is sufficient to augment ~g to achieve
this, for then fl' will be induded in it, and a fortiori in the others.
(11)
(12)
Lemma 2 of §l.l may be used to show that the dass of sets Y satisfying (11)
constitutes a a-field. We have just seen that this dass indudes ~? and ~;.
Since ~o = a(~?, ~;) for each t, the dass must also indude ~o. Hence (11)
is true for each Y E ~o. For any Y E ~?+ (c ~o), a representative of
E{YI~?+} is Y itself. Hence Y E~? by (11), and this means .~?+ c .~?
Therefore ~?+ = ~? as asserted. 0
and more gene rally for Y E b.~o, and each T optional relative to {~n:
(13)
cf. (3) above. The notation EIl indicates that the conditional expectation is
with respect to (Q, ~/l, P/l); and any two representatives ofthe left number in
62 2. Basic Properties
(13) differs by a pl'-null function. lt will be shown below (in the proof of
Theorem 5) that (13) is also true for Y E hY;l'. Now if Y E .'F 0 , then Y 0, E
y;o. But if Y E Y;I', we cannot conelude that Y c Ot E .'FI', even if Y = feX J
This difficulty will be resolved by the next step.
Namely, we introduce a smaller O'-field
'F-
't = /\ 'F I
'
't· (15)
I'
In both (14) and (15), J1 ranges over all finite measures on IJ. lt is easy to see
that the result is the same if J1 ranges over all O'-finite measures on {;'. Leaving
aside astring of questions regarding these fields (see Exercises), we state the
useful result as folIows.
Proof. We have al ready proved the right continuity of {.'Fn and will use
this to prove that of {y;n. Let S(J1, s) be a elass of sets in any space, for each
J1 and s in any index sets. Then the following is a trivial set-theoretic identity:
Applying this to S(J1, s) = .'F~, for all finite measures J1 on 0', and all s E (t, oc),
we obtain
/\ (,'Fn+ = (Y;~)+ (17)
I'
where
(}};,I't ) +
(J. = /\ (}};'I'
,,;r .... , (.'F~) + = / \ .'F;.
s>t s>t
These are to be distinguished from (g;;+)1' and (.~+r = /\/l(g;;+)I', but see
Exercise 6 below. Since (.'Fi) + = .'Fi, (17) reduces to .'F~ = (Y;~)+ as asserted.
o
After this travail, we can now extend (13) to Y in h.'#'-. But first we must
consider the function x --> e {Y}. We know this is a Borel function if Y E b.'F°;
what can we say if Y E bY;-? The reader will do weil to hark back to the
theory of Lebesgue measure in Euclidean spaces for a elue. There we may
begin with the O'-field fJj) of classical Borel sets, and then complete :ßj with
2.3. Strong Markov Property and Right Continuity of Fields 63
respect to the Lebesgue measure. But we can also complete /J)J with respect to
other measures defined on /!/J. In an analogous manner, beginning with :F 0 ,
we have completed it with respect to pI' and called the result .~'". Then we
have taken the intersection of all these completions and called it :F-. Exactly
the same procedure can be used to complete rffD (the Borel field of Ea) with
respect to any finite measure J.l defined on $. The result will be denoted by
«,'"; we then put
(18)
(19)
is in rS"-. For each J.l and each T optional relative to {:Fn, we haue
(20)
(21)
Proof. Observe first that for each x, since Y E :FEx, EX(Y) is defined. Next, it
folio ws easily from the definition of the completion :FI' that Y E h:F1' if and
only ifthere exist Y! and Yz in h:F ü such that P'"{ Y! #- Yz} = 0 and Y! :s:; Y:s:;
Yz. Hence ifwe put
i = 1,2;
qJ! and qJz are in rS", qJ! :s:; qJ :s:; qJz and J.l({qJl #- qJz}) = O. Thus qJ E rS"1' by
definition. This being true for each J.l, we have qJ E IS'-. Next, define a measure
v on rS"a as folIows, for each f E hrS":
(22)
64 2. Basic Properties
If Y E ff-, then there exists ZI and Z2 in ffo such that r{Zj =F Z2} = 0
and ZI ::::; Y::::; Z2' We have Zi 0 GT E .'F11 by (22) of §1.3, with {.~n for {.~}
(re ca II ffl1 = ff~,), and
(23)
(24)
Now put t/Ji(X) = e(zJ, i = 1,2. Since Zi E ffo, t/Ji(X 1') is a representative of
P{Zi • GTlffj} by (13). On the other hand, we have t/Jl ::::; <p::::; t/J2' hence
(26)
and
We end this section with a simple but important result which is often
called Blumenthal's zero-or-one law.
Praaf. Suppose first that A E .'Fg. Then A = X 0 I(A) for some A E {;. Since
PX{X o = x} = 1, we have
which can only take the value 0 or 1. If A E .~~, then for each x, there
exists A X such that r(A D A X ) = 0, so that PX(A) = PX(A X ) which is either
oor 1 as just proved. 0
If we think that ffg is a sm all (J-field, and .'F~ is only "trivially" larger.
the result appears to be innocuous. Actually it receives its strength from the
fact that ff; = (ff-)o+, as part of the Corollary to Theorem 4. Whether an
event will occur "instantly" may be portentous. For instance, let T be optional
relative to {ff;. tE T}. then {T = O} E .~; and consequently the event
{T = O} has probability 0 or 1 under each p x . When T is the hitting time
2.3. Strong Markov Property and Right Continuity of Fields 65
ofa set, this dichotomy leads to a basic notion in potential theory (see §3.4)~
the regularity of a point for a set or the thinness of a set at a point. For a
Brownian motion on the line, it yields an instant proof that almost every
sampie function starting at 0 must change sign infinitely many times in any
sm all time interval.
Exercises
1. Prove that if 1ft ~ 0: :Fr = .~+, then for any T which is optional relative
to {.~}, we have 1ft ~ 0: ''#'T+t = :FT + t +.
2: Let N* denote the dass of all sets in :F 0 such that PX(A) = 0 for every
x E E. Let /F* = lT(:F° v N*). Show that :F* c .'#'-. The lT-field :F* is not
to be confused with :F-; see Problem 3.
3. Let Q = R l and E = R l . Define Xlw) = w + t; then {Xt, t ~ O} is the
uniform motion (Example 2 of §1.2). Show that:F~ = ggl, the usual Borel
field on R l , and p x = Gx , the point mass at x. Show that
1\ (:F0)"x is the dass of all subsets of R 1;
xERl
where f.1 ranges over alI probability measures on gol. The dass N* defined
in Problem 2 is empty so that :F* = gol. (This example is due to Getoor.)
4. Show that
V :FI( = :FI'.
tE T
Is it true that
V Ob-=
.;r t Ob-?•
•.er
tET
5. F or each niet <§ n be a dass of positive (~O) functions dosed under the
operation oftaking the lim sup ofa sequence; let.Yf be a dass ofpositive
functions dosed under the operation of taking the lim inf of a sequence.
Denote by <§n + Yt' the dass offunctions ofthe form gn + h where gn E ~t}n
and hE Yt'. Suppose that <§n::;) <§n+l for every n and <§ = n~j=l <§n'
Prove that
namely that the two operations of "taking the intersection over (t,Xi)"
and "augmentation with respect to pll" performed on {JF?, sET} are
commutative.
7. 1f both Sand T are optional relative to {JF;-}, then so is T + SJ Or.
[Cf. Theorem 11 of §1.3.]
From now on we write .~ for .'F;- and JF for JF-, and consider the Feiler
process (XI,:lFr, PIl) in the probability space (Q, ff, PIl) for an arbitrary fixed
probability measure J1 on 15. We shall omit the superscript J1 in what follows
unless the occasion requires the explicit use. For each w in Q, the sampIe
function X(·, w) is right continuous in T = [0, Xi) and has left limits in
(0, Xi). The family {:lFr, t E T} is right continuous; hence each optional time
T is strictly optional and the pre- T field is .'Fr( = .'F] +). The process has the
strong Marov property. Recall that the latter is a consequence of the right
continuity of paths without the intervention of left limits. We now proceed
to investigate the left limits. We begin with a lemma separated out for its
general utility, and valid in any space (Q, .F, P).
Lemma 1. Let ~§ be a sub-(J~field oj .'F and X E qj; Y E .'F. Suppose that j(n'
each f
E Co we have
(2)
see Lemma 1 of §1.1. Take A = {X ~ U}, and integrate (2) over Il:
o
Corollary. Let X and Y be two randorn variables such that .f(n· any I E Ce.
g E Ce we have
E{f(X)g(X)} = E{f(Y)g(X)}. (4)
Then P{X = Y} = 1.
2.4. Moderate Markov Property and Quasi Left Continuity 67
Proof. As be fore (4) is true if g = lA, where Ais any open set. Then it is also
true if A is any Borel set, by Lemma 2 of §1.1. Thus we have
r
J{XEA}
f(X)dP = r
J{XEAj
f(Y)dP'
,
and consequently
The next lemma is also general, and will acquire further significance
shortly.
p{lim X s =
s~t
Xt} = 1.
Since X is stochastically continuous at t, there exist Sn ii t and tn H t such
that almost surely we have
lim X Sn = Xt = lim X tn '
n n
But the limits above are respectively X t - and X t + since the latter are assumed
to exist almost surely. Hence P{ X t - = X t = X t +} = 1 wh ich is the assertion.
o
Remark. It is sufficient to assume that X t - and X t + exist almost surely,
for each t.
The property we have just proved is sometimes referred to as folIows:
"the process has no fixed time of discontinuity." This implies stochastic
continuity but is not implied by it. For example, the Markov chain in
Example 1 of §1.2, satisfying condition (d) and with afinite state space E, has
this property. But it may not if Eis infinite (when there are "instantaneous
states"). On the other hand, the simple Poisson process (Example 3 of §1.2)
has this property even though it is a Markov chain on an infinite E. Of course,
the said property is weaker than the "almost sure continuity" of sampie
functions, which means that alm ost every sam pie function is a continuous
function. The latter is a rather special situation in the theory of Markov
68 2. Basic Properties
Theorem 3. For each predictable T, we have for each f E 1[, al1d u 2': 0:
(5)
where X o - = X o.
Proof. Remember that limt~w X t need not exist at all, so that X T - is unde-
fined on {T = oo}. Since T E .~T-' we may erase the two appearances of
11T < c j in (5) and state the resuIt "on the set {T < oo}". However, since T
lf
does not necessarily belong to :#'Tn, we are not allowed to argue blithely
below "as if T is finite".
Let {T n} announce T, then we have by Theorem 1 of §2.3:
(7)
· j·(X {/ ·(X)
) =. on 1.rT -- 0 1j,
11m T +u
u
r 1.
n~7 n f(XT+u-) on lT> OJ.
(8)
Letting M --+ 00 we obtain (5) with X T+u- replacing X T+u in the left member.
2.4. Moderate Markov Property and Quasi Left Continuity 69
Now define X'lj _ (as weil as X exJ to be aeven where limt~CX'. X t exists and is
not equal to a. Then the factor 1(T<x.) in the above may be cancelled. Since
X T- E ~T- by Theorem 10 of §1.3, it now follows from Lemma 1 that we
have almost surely:
XT=X T-· (9)
This is then true on {T < oo} without the arbitrary fixing ofthe value of X.x;_.
For each optional T and u > 0, T + u is predictable. Hence we have just
proved that X T+u = X T+u- almost surelyon {T < oo}. This remark enables
us to replace X T + u _ by X 1"+ u in the left mem ber of (8), th us com pleting the
proof of Theorem 3. D
(10)
The details of the proof are left as an exercise. The similarity between (10)
above and (13) of §2.3 prompts the following definition.
Thus, this is the ca se for a Feiler process whose sam pIe functions are right
continuous and have left limits. The adjective "moderate" is used here in
default of a better one; it does not connote "weaker than the strong".
There is an important supplement to Theorem 3 which is very useful in
applications because it does not require T to be predictable. Let {Tn } be a
sequence of optional times, increasing (loosely) to T. Then T is optional but
not necessarily predictable. Now for each w in Q, the convergence of Tn(w)
to T(w) may happen in two different ways as described below.
Ca se (i). "In: T n < T. In this case X T " ---> X T - if T< 00.
Theorem 4. Let TI! be optional and increase to T. Then we haue almost surely
Proof. Put
A = [\In: TI! < T};
and define
on {TI! < T}, T on A,
T' = {TI! T'- {
n 00 on{Tn=T}; 'AJ on Q - A.
Since {TI! < T} E .'FTn and A E:'FT by Proposition 6 of §1.3. T;, and T' are
optional by Proposition 4 of §1.3. We have
and consequently
In view of (12), (13) and (14), Equation (11) is true on A n {T <J'~ }. But (11)
is trivially true on Q - A. Hencc Theorem 4 is proved. D
We are now ready to prove the optionality of the hitting time for a closed
set as weIl as an open set. For any set A in {j'4> define
This is called the (first) hitting time of A. Compare with the first entrance
time DA defined in (25) of §1.3. The difference between them can be crucial.
2.4. Moderate Markov Property and Quasi Left Continuity 71
tion 3 of §1.3. It is easy to see that the assertion remains true when {ff~} is
replaced by {ffn, for each 11.
Now it is obvious that DA is the debut of the set
HA = {(t,w)IX(t,w) E A} (17)
Theorem 5. IJ the M arkov process {X t} is right continuous, then Jor each open
set A, DA and TA are both optional relative to {ff~}. IJ the process is right
continuous almost surely and is also quasi left continuous, then Jor each closed
set as weil as each open set A, DA and TA are both optional relative to {.?";-}.
where Qt = Q n [0, t), provided all sampie functions are right continuous.
To see this suppose DA(W) < t, then X(s, w) E A for so me s E [D A(W), t), and
by right continuity X(r, w) E A for some rational r in (D A(W), t). Thus the left
member of (18) is a subset of the right. The converse is trivial and so (18) is
true. Since the right member clearly belongs to ff~, we conclude that
{D A < t} E ff~.
Now suppose that only P-almost all sam pie functions are right continuous.
Then the argument above shows that the two sets in (18) differ by a P-null
set. Hence {DA< t} is in the augmentation of ff~ with respect to (Q, .?", P).
Translating this to pli, we have {DA< t} E ff': for each 11, hence {DA< t} E
ff;-. Since {ff;-} is right continuous we have {DA:::; t} E ff;-.
72 2. Basic Properties
(19)
We now make the important observation that for any Borel set B, we have
almost surely
(21 )
For if DB(w) < 00, then for each () > 0, there exists tE [DB(w), DB(w) + (5)
such that X(t, w) E B. Hence (21) follows by right continuity. Thus we have
X(D AJ EAn for all n and therefore by quasi left continuity and (19):
almost surelyon {O::;; S < oo}. The case S = 0 is of course trivial. This
implies S ;::: DA; together with (20) we conclude that DA = S a.s. on {S <Xc }.
But DA ;::: S = Cf) on {S =00 }. Hence we have proved a.s.
(22)
(23)
Exercises
1. Here is a shorter proof of the quasi left continuity of a Feller process
(Theorem 4). For IX> 0, fE bC+, {e-~tU"f(Xt)} is a right continuous
positive supermartingale because U"j is continuous. Hence if T n i T,
and Y = limn X(T n) (which exists on {T < oo}), we have a.s.
NOTES ON CHAPTER 2
§2.1. This chapter serves as an interregnum between the more concrete Feiler
processes and Hunt's axiomatic theory. It is advantageous to introduce some of the
basic tools at an early stage.
§2.2. Feiler process is named after William Feiler who wrote aseries of pioneering
papers in the 1950's. His approach is essentially analytic and now rarely cited. The
sampie function properties of his processes were proved by Kinney, Dynkin, Ray,
Knight, among others. Dynkin [lJ developed Feller's theory by probabilistic methods.
His book is rich in content but difficult to consult owing to excessive codification.
Hunt [lJ and Meyer [2J both discuss Feiler processes before generalizations.
§2.3. It may be difficult for the novice to appreciate the fact that twenty five years
ago a formal proof of the strong Markov property was a major event. Who is now
interested in an example in which it does not hold?
74 2. Basic Properties
Hunt Process
Theorem 1. Almost surely the sampie paths haue lelt limits in (0, IX)).
(I)
where d denotes a metric of the space Ea. Our first task is to show that T is
optional, indeed relative to {:F7}. For this purpose let {zd be a countable
76 3. Hunt Process
dense set in Ea, and B kn be the closed ball with center Zk and radius n ~ 1 : B kn =
{xld(x,Zk):-S; n~l}. Then {B kn } forms a countable base ofthe topology. Put
Then Tkn is optional relative to {,F?} by Theorem 6 of §2.4, since the set
{xld(x,B kn ) > c} is open. We claim that
It is clear that right member of (2) is a subset of the left member. To see the
converse, suppose T(w) < t. Then there is s(w) < t such that d(Xs(w),
X o(w)) > c; hence there are n and k (both depending on w) such that d(Xs(w),
Xo(w)) > B + 2n~1 and Xo(w) E Bk"" Thus d(Xs(w), Bkn ) > I; and Tkn(O)) < t;
namely 0) belongs to the right member of (2), establishing the identity. Since
the set in the right member of (2) belongs to .~?, T is optional as claimed.
Next, we define T o == 0, Tl == T and inductively for n 2: 1:
Let Q* = n~= 1 Q I/rn; then P(Q*) = 1. We assert that if 0) E Q*, then X(-, 0))
must have left limits in (0,00). For otherwise there exists tE (0, 00) and m
such that X(·, 0)) has oscillation > 2/m in (t - (), t) for every () > 0. Thus
t tJ [T~/m, T~~md for all n 2: 0, which is impossible by (3) with B = I/rn. D
other words, the process has no fixed time of discontinuity. For a Feiler
process, this was remarked in §2.4.
Much stronger conditions are needed to ensure that almost all paths are
continuous. One such condition is given below which is particularly adapted
to a Hunt process. Another is given in Exercise 1 below.
. 1
hm - sup [1 - Pt(x, B(x, E))] = 0 (4)
t-O t XEK
where B(x, E) = {y E E cI d(x, y) :-:;; c:}. Then almost all paths are continuous.
max
O~k~n-l
d (f (~),f (~))
n n
> E. (5)
Proof of the Lemma. If f is not continuous in [0, 1], then there exists t E
(0,1] and c: > 0 such that d(f(t - ), f(t)) > 2E. For each n ~ 1 define k by
kn- 1 < t:-:;; (k + 1)n- 1 . Then for n z no(E) we have d(f(kn- 1 ), f(t-)) < <:/2,
d(f( (k + l)n -1 ),f(t)) < E/2; hence d(f(kn - 1),f( (k + l)n - 1)) > [; as asserted.
Conversely, if f is continuous in [0,1], then f is uniformly continuous
there and so for each c: > 0 (5) is false for all sufficiently large n. This is a
stronger conclusion than necessary for the occasion.
To prove the theorem, we put for a fixed compact K:
Mn={wl sup
O~k~n-l
d(X(~'W),X(~'W))>E;
n n
Using the condition (4) with t = n -1, we see that the last quantity above
tends to zero as n ~ 00. Hence P(liminfn M~):::; lim n P(M~) = and P(M) =
0. Now replace the interval [0,1] by [//2,1/2 + 1] for integer 1 ;;:: 1, and K
°
by Km U {a} where Km is a compact subset of E and Km i E. Denote the
resulting M by M(l, m) and observe that we may replace K by K U {cJ
in (4) because ais an absorbing state. Thus we obtain
= 2 JEI"
J2ni [u J
2
exp - 2t du
{2 IX ~(exp[-1A~JLl.)dU:::; (2 ~exp[-~~J.
=
~ mJ, u 2t t ~ m B 2t
Exercises
1. A stochastic process {X(t), t ;;:: O} is said to be separable iff there exists
a countable dense set S in [0,00) such that for almost every w, the sam pIe
function X(·, w) has the following property. For each t ;;:: 0, there exists
Sn E S such that Sn ~ t and X(S.,w) ~ X(t,w). [The sequence {Sn} depends
3.1. Defining Properties 79
Take S to be the dyadics and prove that X(s, w) with SES is continuous
on S for almost every w. Finally, verify the condition above for the
Brownian motion in R 1. [Hint: for the last assertion estimate L~: c/
P{IX((k + l)r n) - X(k2- n)1 > n- 2 } and use the Borel-Cantelli lemma.
For two dyadics sand s', X(s) - X(s') is a finite sum of terms of the
form X((k + 1)2-n) - X(kr n). The stated criterion for continuity is due
to Kolmogorov.]
2. Give an example to show that the Lemma in this section becomes false
if the condition "f has left limits in (0, 1]" is dropped. This does not
seem easy, see M. Steele [1].
3. Let X be a homogeneous Markov process. A point x in E is called a
°
"holding point" ifffor some b > we have P X { X(t) = x for all t E [0, b]} >
°
0. Prove that in this case there exists A ~ such that PX{T{x}c > t} =
e- Jet for all t > 0. When A = 0, x is called an "absorbing point". Prove
that if X has the strong Markov property and continuous sam pIe func-
tions, then each holding point must be absorbing.
4. For a homogeneous Markov chain (Example 1 of §1.2), the state i is
holding (also called stahle) if and only if
. 1 - pu(t)
11m < 00.
tlO t
The limit above always exists but may be + 00, when finite it is equal
to the A in Problem 3. In general, a Markov chain does not have aversion
which is right continuous (even if all states are stable), hence it is not a
Hunt process. But we may suppose that it is separable as in Exercise 1
above.
5. For a Hunt process: for each x there exists a countable collection of
optional times {T n} such that for PX-a.e. w, the set of discontinuities
of X(-,w) is the union UnTn(w). [Hint: for each e> Odefine S<e) =
inf{t > 0ld(Xt-,X t) > e}. Show that each S(e) is optional. Let Sie) = S(f.),
s~eL = s~e) + S(e) e(s~f.)) for n ~ 1. The collection {S~l/m)}, m ~ 1, n ~ 1,
0
(1)
thus p? = P,. For each rt., (P~) is also a Borelian semigroup. It is necessarily
submarkovian for a > O. The corresponding potential kernel is where V"
vaf
. =
Jor' paj' dt.
l.
(2)
Here f E h6+ or f E {; +. If fE bff+ and a > 0, then the function V~f is finite
everywhere, whereas this need not be true if fE ff +. This is one reason why
we have to deal with va sometimes even when we are interested in VO = V.
But the analogy is so nearly complete that we can save a lot of writing by
treating the casea = 0 and daiming the result fora;:::: O. Only the finiteness
of the involved quantities should be watched, and the single forbidden
operation is "XJ -XJ".
The rt.-potential has been introduced in §2.l, as are rt.-superaveraging and
rt.-excessive functions. The dass of rt.-excessive functions will be denoted by
sa, and SO = S.
We begin with a fundamental lemma from elementary analysis, of which
the proof is left to the reader as an essential exercise.
(3)
Here we have a glimpse of the interplay between the semigroup and the
process. Incidentally, speaking logically, we should have said "an associated
Markov process" in the above.
82 3. Hunt Process
Proposition 4. If rx < ß, then S" c Sß. For each rx 2: 0, S" = nß>a Sß.
We have already observed that the limit above is monotone. Hence we may
define Pt+f(x) = lim s11t PJ.
lim PJ = O. (6)
Then we haue
I = lim
hlü
i u(L- Pd). h
(7)
3.2. Analysis of Excessive Functions 83
Praaf. The hypothesis Pt! < 00 allows us to subtract below. We have for
h > 0:
If we divide though by h above, then let t i 00, the last term converges to
zero by (6) and we obtain
(8)
The integrand on the left being positive because J ?: PhJ, this shows that
the limit above is the potential shown in the right member of (7). When
h 1 0, the right member increases to the limit J, because lims!ü i Ps! = f.
This establishes (7). 0
(9)
Praaf. This is proved by the following calculations, for J E blff + and rJ. "'" ß:
84 3. Hunt Process
Note that the steps above are so organized that it is immaterial where
ß - r:t. > 0 or < O. This remark estab!ishes the second as weil as the first
equation in (9). 0
we have
(10)
Proof. Let 0 < ß < r:t. and fE brff +. Then we have by (9):
(12)
Here subtraction is allowed because U i7f <00. If (10) is true, then gaß =
f - (r:t. - ß)U1 ;:::: O. Since UIJg,ß is ß-excessive, Uaf is ß-excessive for all
ß E (0, r:t.). Hence U'I E S by Proposition 4. Next, we see from (12) by simple
arithmetic that
we have f* ES by Proposition 2.
For a general fE rff + satisfying (0), let fn = f 1\ n. Then r:t.U"f~ ::;: j~ by
(10) and the inequa!ity r:t.U'n ::;: n. Hence we may apply what has just been
proved to each fn to obtain r::
= lim,lw i r:t.U~fn. Also the inequa!ity
r:t.U"j~ ;:::: ßUßln leads to r:t.U"l;:::: ßUßf, if r:t. ;:::: ß. Hence we have by Lemma 1:
Theorem 9. Ir I E S, then Ior each ß > 0, there exists gn E brff + such that
(15)
n
Prool. Suppose first that I E brff +. Then as in the preceding proof, we have
Ua(rx.f) = Uß(rxgaß)' Thus we obtain (15) from (11) with gn = ngnß' In general
we apply this to obtain for each k, I/\ k = lim n i Ußg~k). It follows then by
Lemma 1 that
etc. But the preceding proof is more general in the sense that it makes use
only of (10) and (11), and the resolvent equation, without going back to the
semlgroup.
It is an important observation that Theorem 9 is false with ß = 0 in (15);
see Exercise 3 below. An additional assumption, which turns out to be of
particular interest for Hunt processes, will now be introduced to permit
ß = 0 there. Since we are proceeding in an analytic setting, we must begin by
assuming that
"Ix E E: U(x,E) > O. (16)
This is trivially true for a process with right continuous paths; for a more
general situation see Exercise 2 below. Remember however that U(a, E) = O.
To avoid such exceptions we shall agree in what follows that a "point" means
a point in E, and a "set" means a subset ofE, without specific mention to the
86 3. Hunt Process
0< Uh< 00 on E. (\ 7)
Proposition 10: Ir (Pt) is transient, and fES, then ther!! !!xists 4n Eh(; + such
that
(19)
Proof. Put
where
gnk = k(J~ - P 1 / k .1;,):::;: kn.
Ug nk = k S.
Ilk
()
~
PsIndx:::;: n.
3.3. Hitting Times 87
lim
n
i J:. = lim i lim i
n k
Ug nk = lim
n
i Ug nn ·
lim
n
i J:. = lim
n
i lim i
,~o
Pt!n = lim
,~o
i lim i
n
Pt!n = lim Pt! = f·
,~o
A of E a, we put
This is the projection on Q of the set of (s, w) in [0, t] x Q such that X(s, w) E A.
The mapping A -> A' from the dass of all subsets of Er to the dass of all
subsets of Q has the following properties:
(i) Al C A 2 =A'1 C A~;
(ii) (Al U A 2 )' = A'l U A~;
(iii) =
An i A A~ i A'.
In view of (i) and (ii), (iii) is equivalent to
(iii') (Un AnY = Un A~ for an arbitrary sequence {An}·
So far these properties are trivial and remain valid when the fundamental
internal [0, t] is replaced by [0, O'J). The next proposition will depend
essentially on the compactness of [0, t J.
We shall write Al (A z to denote P(A 1 \A z ) = 0, and Al == A z to denote
P(A I L::,. A z ) = O.
K =nGn(=n G
n n
n ). (1)
Then we have
K' == n G~. (2)
K' == n
m
G~ ::::J n
m
K~m ::::J n
K~ ::::J K',
Corollary. For each compact K, and c; > 0, there exists open G such that
K c G and
P(K') ::::; P(G') ::::; P(K') + t:.
In particular
P(K') = inf P(G') (4)
G=oK
From here on in this section, the letters A, B, G. Kare reserved for ar bi-
trary, BoreI, open, compact sets respectively.
(6)
Proof. We have, as a superb example ofthe facility (felicity) of reasoning with
sampIe functions:
= G~ - (GI n G 2 )'.
90 3. Hunt Process
Up to here GI and G2 may be arbitrary sets. Now we use definition (a) above
to convert the preceding inequality into (6). D
The same argument shows C is also strongly additive over all compact
sets, because of (5). Later we shall see that C is strongly additive over all
capacitable sets by the same token.
(7)
(8)
Prooj'. For m = 1. (8) reduces to (7). Assume (8) is true as shown and observe
that
(9)
We now apply the strong subadditivity of Cover the two open sets U:= I Gn
and Gm + bits monotonicity, and (9) to obtain
Theorem 5. Wehave
(i) Al c Az=C(A I ) ~ C(A z );
(ii) An i A=ClAn) i C(A);
(iii) Kn 1K = C(K n ) 1 C(K).
3.3. Hitting Times 91
Proof. We have already mentioned (i) above. Next, using the notation of
Lemma 4, we have
rn
Since 8 is arbitrary, (ii) follows from this inequality and (i). Finally, it follows
from (5), Proposition 2, and the monotone property of P that
(12)
(13)
and consequently
(14)
It follows that
Theorem 7. For each B E IS", DB and TB are hoth optional relative to {.~).
Proof. Recall that t is fixed in the foregoing discussion. Let us now denote
the B' above by B'(t). For the sake of explicitness we will replace (Q, .'1', P)
above by (Q, .~, PIl) for each probability measure f1 on IS, as fully discussed
in §2.3. Put also Qt = (Q n [0, t)) u {tl. Since G is open, the right continuity
of S --+ X s implies that
G'(t) = {w[:ls E Qt:Xs(w) E G}.
Hence G'(t) E gF~ c .'F'(. Next, we have Pi'{B'(t) D G~(t)} = n,.
by (13),
hence B'(t) E .'F'( because .'F'( contains all PIl-null sets. Finally, a careful
°
scrutiny shows that for each B E ffa and t > 0:
{w[DB(w)<t}= U {W[:lsE[O,r]:Xs(w)EB}
rE Qn[O.t)
U B'(r) E gF'(.
rE Qn[O.t)
3.3. Hitting Times 93
Theorem 8(a). Far each J1 and BE rffo, there exist K n c B, K ni and Gn ::::J B, Gnl
such that
(15)
Praaf. The basic idea is to apply the Corollary to Theorem 6 to B'(r) for all
rE Q. Thus for each rE Q, we have
(16)
such that
( 17)
n
For n ~ j, we have
hence
94 3. H un t Process
Hence by (17),
lim pl1{ G~(r) - K~(r)} = 0
n
and consequently by (18), PI1{D' > r j > DU} = O. This being true for all r j ,
we conclude that PII{D' = D B = DU} = 1, which is the assertion in (15). 0
Theorem 8(b). For each /l and B Elfe, there exists K n c B,K ni such that
Pll-a.s. (19)
(20)
applied to a sequence Sk 11 O. Let /lk = /lP'k' Then for each k :::::: 1, we have
by part (a): there exist K kn C B, Kkni as l1i such that
which means
(22)
P"-a.s.
This proves (19) and it is important to see that the additional condition
°
/leB) = is needed for (20). Observe that TB = DB unless X 0 E B; hence under
the said condition we have pli { TB = D B } = 1. lt follows from this remark
that T G = D G for an open G, because if X 0 E G we have T G = 0 (which is
not the ca se for an arbitrary B). Thus we have by part (a):
o
3.3. Hitting Times 95
Remark. The most trivial counterexamplc to (20) when 1l(B) i= 0 is the case
ofuniform motion with B = {0},11 = Eo. Under pI', it is obvious that TB = Cf]
but T G = 0 for each open G ::::J B. Another example is the case of Brownian
motion in the plane, if B = {x} and 11 = Ex; see the Example in §3.6.
As a first application of Theorem 8, we consider the "left entrance time"
and "left hitting time" of a Borel set B, defined as folIows:
Da = DG ; Ta = T G;
Dii 2: DB; Tii 2: TB'
Proof. Since the relation (21) holds for Tii and Dii as weil, it is sufficient to
consider the D's.
If X t E G and G is open, then there is an open set GI such that GI c G
and X t E GI' The right continuity of paths then im pli es that for each w there
exists 60 (w) > 0 such that X tH E GI for 0< 6 :$; 60 and so X tH - E GI' This
observation shows Da :$; DG . Conversely if t > 0 and X t - E G, then there
exist t n ii t such that X tn E G; whereas if X o - E G then X o E G. This shows
DG :$; Da and consequently D G = Da.
For a general B, we apply the second part of (15) to obtain Gn ::::J B such
that DGn i DB, PI'-a.s. Since Dii 2: Dan = DGn , it follows that Dii 2: DB, PI'-a.s.
D
(24)
This implies: on the set {t < 0, the closure of the set of ualues USE[O.tl X(s, w)
is a compact subset of E.
We dose this section by making a "facile generalization" which turns
out to be useful, in view oflater developments in §3.4.
Definition. A set A c E a is called nearly Borel iff for each finite measure 11 on
there exist two Borel sets BI and B 2 , depending on 11, such that BI c A c
(j'",
B 2 and
(25)
96 3. Hunt Process
(26)
One can describe g' in a folksy way by saying that the "poor" Hunt process
{X t } cannot distinguish a set in,r;" from a set in ~;o. It should be obvious that
the hitting time of a nearly Borel set is optional, and the approximation
theorems above hold for it as weil. eWe do not need "nearly compact" or
"nearly open" sets!] Indeed, if A E g' then for each 11 there exists B E ~;, such
that DB = DA, PI'-a.s. Needless to say, this B depends on 11.
A function f on Ea to [ - 00, + 00] is nearly Borel when fE IJ '. This is
the case if and only if for each 11, there exist two Borel functions fl and .t~,
depending on 11, such that fl :S: f :S: .t~ and
(27)
(1)
each f, then we may omit "T < 00" in the expression above. We shall fre-
quently do so without repeating this remark.
The following composition property is fundamental. Recall that if Sand
T are optional relative to {g;;}, so is S + Tc Os: see Exercise 7 of §2.3.
Proposition 1. We have
(2)
3.4. Balayage and Fundamental Structure 97
= E'{e-a(S+ToOslj(X S+ToOs )} . o
Here and henceforth we will adopt the notation E'(' .. ) to indicate the
function x -+ E X (, • ').
(4)
The set of all points in Eo which are regular for A will be denoted by Ar; and
the union A u Ar is called the fine closure of A and denoted by A *. The
nomenc1ature will be justified in §3.5.
(5)
for each rx> 0; indeed r{TA = O} may be regarded as lim aico P~l(x).
Finally if x E Ar, then for any ! we have
Applying the strong Markov property at TA, we see that the right member
of (6) is equal to the PX-probability of the set
where TA c 8TA means the "time lapse between the first hitting time of A and
the first hitting time of A thereafter." If X T)W) r/: A and TA 0T)W) >
then the sampie function X(·, W) is not in A for TE [T A(W), TA(w) +
°
TA 0 0T)W)), a nonempty interval. This is impossible by the definition of TA-
Hence the right member of (6) must be equal to zero, proving the assertion
~~ilioo~. 0
The following corollary for rx = 0 is just (21) of §2.4, wh ich is true for TB
as well as D B .
(7)
(8)
Theorem 3.
(a) p~UJ':::;,U'i;
(b) p~UJ' = U'i, if 11 c A;
(c) If we have
Uaf:::;' uag (9)
Letting n -> 00, using (a) and (b), and (10) for both land y, we obtain
(11)
( 12)
This is verified exaetly like (16) of §2.4, and the intuitive meaning is equally
obvious. We have from (7):
It follows from this and (13) that P~ U"y E S. For I E S, we have by Theorem
9 of §3.2, I = lim k i U"Yk where rx > O. Henee P~I = lim k i P~ U'Yk and so
P~I ES" by Proposition 2 of §3.2. Finally P AI = lim a10 i P~I; henee
P AI E S by Proposition 4 of §3.2. Next if Ale A 2, then it follows at onee
from (7) that P~, U'g ::::; P~2U'g. Now the same approximations just shown
establish (11), whieh includes PAI : : ; I as a partieular ease.
Next, given x and A, let {K n } be as in the preeeding proof. As in (10), we
have
3.4. Balayage and Fundamental Structure 101
= lim
n
r lim r Ptuagk(X) =
k
lim r p~J(x).
n
The next result is crucial, though it will soon be absorbed into the bigger
Theorem 6 below.
Proof. Since x E Ar, f(x) = P Af(x) as noted above. It follows from Theorem
8(b) of §3.3 that for each c > 0, there exists K c A such that
(14)
We have arrived at the fOllOwing key result. In its proof we shall be more
circumspect than elsewhere about pertinent measurability questions.
Theorem 6. Let fES. Then almost surely t -..... f(X t ) is right continuous on
[0,00) and has left limits (possibly + 00) in (0, 00]. M oreover, f E <ff •.
102 3. Hunt Process
An argument similar to that for the Tin (2) of §3.1 shows that Sr. is optional
(relative to (~». It is clear that we have
n
k= I
{S\lk>O}c{wllimf(Xt(W»=f(Xo(W»}=A,
tlO
(15)
say. For each x put A = {y E Ealp(.f(x),f(y» > e}. Then A ES"; and S"
reduces to the hitting time TA under p x . It is a consequence of Theorem 5
that x f/= Ar; for obviously the value f(x) does not lie between the bounds of
fon A. Hence for each x, PX{Sr. > O} = 1 by Blumenthal's zero-one law.
Now {Sr. > O} E ,~, hence r{S, > O} E S- by Theorem 5 of §2.3, and this
implies PfJ{S, > O} = 1 for each probability measure /1 on Si!' lt now follows
from (15) that A E fffJ since (,97 fJ ,pfJ) is complete. This being true for each /1,
we have proved that A E ,~ and A is an almost sure set. For later need let
us remark that for each optional S, A " es E ff by Theorem 5 of §2.3.
Our next step is to define inductively a family of optional times T~ such
that the sampie function t -+ f(X t ) oscillates no more than 2e in each interval
[T~, Td 1)' This is analogous to our procedure in the proof of Theorem 1
of §3.1, but it will be necessary to use transfinite induction here. The reader
will find it instructive to compare the two proofs and scrutinize the difference.
Let us denote by (D the well-ordered set of ordinal numbers before the first
uncountable ordinal (see e.g., Kelley [1; p. 29]). Fix e, put T o == 0, Tl == Sr.,
and for each IX E (D, define T~ as folIows:
(i) if IX has the immediate predecessor IX - 1, put
T~ = T~-1 + Tl 0 eTO _ 1 ;
Then each Ta. is optional by induction. [Note that the sup in ca se (ii) may be
replaced by the sup of a countable sequence.] Now for each IX E (D, we have
for each /1:
3.4. Balayage and Fundamental Structure lO3
by the strong Markov property and the fact that r(A) = 1 for each x, proved
above. It follows that P/J-a.e. on the set {T~ < oo}, the limit relation in the
rJ.E il).
°
Now the set il) has uncountable cardinality. If b~ > for all rJ. E il), then the
uncountably many intervals (Cd b C~), rJ. E il), would be all nonempty and
disjoint, which is impossible because each must contain a rational number.
Therefore there exists rJ.* E il) for which b~. = 0; namely that
For each rJ. < rJ.*, if sand t both belong to [T~, Td 1)' then p(f(Xs),f(Xt ))::;; 28.
This means: there exists Q~ with P/J(Q,) = 1 such that if W E Q~, then the
sampIe function t~ I(X(t,w)) does not oscillate more than 28 to the right
Q:
01 any tE [0,00). It follows that the set of w satisfying the latter require-
ment belongs to f7 and is an almost sure set. Since nk'= 1 QT/k is contained
in the set of w for which t ~ I(Xt(w)) is right continuous in [0,00), we have
proved that the last-mentioned set belongs to .'#i and is an almost sure set.
This is the main assertion of the theorem.
In order to fully appreciate the preceding proof, it is necessary to reflect
that no conc1usion can be drawn as to the existence ofleft limits ofthe sam pIe
function in (0, 00). The latter result will now follows from an earlier theorem
on supermartingales. In fact, if we write In = I 1\ n then {fn(X t),~} is a
bounded supermartingale under each r, by Proposition 1 of §2.1, since J"
is superaveraging. Almost every sampIe functions t ~ J,,(X t ) is right con-
tinuous by what has just been proved, hence it has left limits in (0, 00 ] by
Corollaries 1 and 2 to Theorem 1 of §1.4. Since n is arbitrary, it follows
(why?) that t ~ I(X t ) has left limits (possibly + 00) as asserted.
It remains to prove that I E S implies I E co. In view of Theorem 9 of
§3.2, it is sufficient to prove this for U~g where rJ. > 0 and 9 E bC:. For each
j1, there exist gl and g2 in Cil such that gl ::;; 9 ::;; g2 and j1U~(92 - 91) = 0.
For each t ;?: 0, we have
It follows that
104 3. Hunt Process
Therefore (17) is also true simultaneously for all rational t > O. But
U~(g2 - gl) E S~ n t&', hence the first part of our proof extended to S' shows
that the function of t appearing in (17) is right continuous, hence must
vanish identically in t, PIl_ a.s. This means U~g E t&'. as asserted. Theorem 6
is completely proved. 0
If fES, when can f(X t ) = (fJ? The answer is given below. We introduce
the notation
Vt E (0, (fJ]: f(X t ) _ = !im f(X.). (18)
sHt
( 19)
It is obvious that on the set {D F > O},f(X t ) = (fJ for t < DF, andf(X t )- =
00 for t ::;; DF ; but f(X(D F)) may be finite or infinite.
There are numerous consequences ofTheorem 6. One ofthe most import-
ant is the following basic property of excessive functions which can now be
easily established.
Exercises
1. Let f E bg +, T n i T. Prove that for a H un t process
then we have
Remark, Under the hypothesis that (20) holds for any optional T n i T,
the function t --> Uaf(X(t)) is left continuous. This is a complement to
Theorem 6 due to P. A. Meyer: the proof requires projection onto the pre-
dictable field, see Chung [5].
2. Let A E g' and suppose T n i TA' Then we have for each :x> 0, a.s. on
the set nn{Tn < TA< oo}:
!im E X (1n){e- aTA } = 1.
[Hint: consider EX{e- aTA ; T n < TA IffTJ and observe that TA E Vn ·~T,,·]
3. Let X o be a holding but not absorbing point (see Problem 3 of§3.l). Define
T = inf{t > O:X(t) # xo}.
106 3. Hu n t Process
Indeed, it follows that for any optional S, pxo{ 0 < S < T} = O. [Hint: the
second assertion follows from the first, by contraposition and transfinite
induction, as suggested by J. B. Walsh.J
4. Let fE 0+ and f 2': PKf for every compact K. Let gE IJ such that
U(g+) /\ U{g-) < (XJ (namely U(g) is defined). Then
12': Ug on lH..::.
8. The optional random variable T is called terminal iff for every t 2': 0
we have
T = t + T Ot on {T > t}
and
T = lim(t + Tc 0t).
tt 0
Theorem 5 of §3.4 may be stated as folIows: if fES, then the bounds ofi on
the fine closure of Aare the same as those on A itself. Now if I is continuous
in any topology, then this is the case when the said closure is taken in that
topology. We proceed to define such a topology.
3.5. Fine Properties 107
Definition. An arbitrary set A is called finely open iff for each x E A, there
exists B E tff such that x E B c A, and
(1)
Proof. Assume the right continuity. For any real constant c, put
A = {f> q},
f(X(t,w)) < q,
Since X(tn, w) E A and A is finely open, the point X(t n, w) is certainly regular
for A, hence cpiX(tn, w)) = 1 for all n. Therefore we have cpiX(t, w)) = 1 by
(2). But B = {f< q} is also finely open and X(t, w) E B, hence by definition
the point X(t, w) is not regular for Be, a fortiori not regular for A since
Be:=J A. Thus Cpq(X(t, w)) < 1. This contradiction proves the claim. A similar
argument shows that limsHJ(X(s,w));:::: f(X(t,w)) for all t;:::: 0, PX-a.s.
Hence t --+ f(X(t, w)) is right continuous in [0, CX)), a.s. 0
(3)
Proof. (3) is easy. To prove (4) let f = g·{e- TA }. Then f(x) = 1 if and only
if x E Ar. If X E (Ar)' then fix) ;:::: inf f by theorem 5 of §3.4. Hence fix) = 1.
To prove (5), note first that Ar E g' so that TAr is optional. We have by
Theorem 2 of §3.4, almost surely:
where the last equation is by (4). Hence TA eTAr = 0 by the strong Markov
0
property applied at T Ar, and this says that TA cannot be strict1y greater
than TAr. 0
3.5. Fine Properties 109
We shall consider the amount of time that the sampie paths of a Hunt
process spend in a set. Define for A c E a :
(6)
(7)
(8)
Proposition 3. If Ua (., A) == 0 for some iJ. ~ 0, then Ua (., A) == 0 for all iJ. ~ O.
Proof. A set is dense in any topology iff its closure in that topology is the
whole space; equivalently iff each nonempty open set in that topology
contains at least one point of the set. Let 0 be a finely open set and XE O.
Then almost every path starting at x spends a nonempty initial interval of
time in a nearly Borel subset B of 0, hence we have EX{m(J B)} > O. Since
P{m(JA)} = 0 we have EX{m(JBnAc)} > O. Thus 0 n AC is not empty for
each finely open 0, and so AC is finely dense. 0
For they are both finely continuous and agree on a finely den se set. The
assertion is therefore reduced to a familiar topological one.
From certain points of view a set of zero potential is not so "smalI" and
scarcely "negligible". Of course, what is negligible depends on the context.
For example a set of Lebesgue measure zero can be ignored in integration
but not in differentiation or in questions of continuity. We are going to
define certain sm all sets which play great roles in potential theory.
The last condition is equivalent to: e{e-'TA} == P~l(x) == 0 for each CI. > O.
For comparison, A is thin if and only if
(10)
Note that (11) does not imply that SUPXE E CfJ A(X) < 1. An example is furnished
by any singleton {xo} in the uniform motion, since CfJ{xo}(xo) = 0 and
lim xil Xo CfJ {xo}(x) = 1.
Proof. Recall that ({JA is IX-excessive; hence by Theorem 5 of §3.4, the sup in
(11) is the same if it is taken over A * instead of A. But if x E Ar, then ({J A(X) = 1.
Hence under (11) Ar = 0 and A is thin.
Next, denote the sup in (11) by so that e e<
1. Define Tl == TA and for
n:;::: 1: T n+ 1 = T n + TA Tn0. e
Thus {Tn, n:;::: I} are the successive hitting
times of A. These are actually the successive entrance times into A, because
X(Tn ) E A* = A here. The adjective "successive" is justified as folIows. On
[T n < oo}, X(Tn ) is a point which is not regular for A; hence by the strong
Markov property we must have TA 0 e
Tn > 0, namely T n + 1 > Tn' Now we
have, taking IX = 1 in cP A:
oo}
e1
n
=
° as n ---+ 00. Let T n i TX). Then E{e- T C}
1. Therefore, we have almost surely
= 0;
JA = U {T,,}I{Tn <xj;
n~l
(12)
(13)
Then we have
(14)
Hence each A n A(n) is a very thin set. If Ais thin, then (14) exhibits A as a
countable union of very thin sets. Therefore each semipolar set by its de-
finition is also such a union. Finally, (14) also shows that A - (A n Ar) is
such a union, hence semipolar. 0
This follows at once from the first assertion of the theorem and (12). Note
however that the collection {T n } is not necessarily weII-ordered (by the
increasing order) as in the case of a very thin set A.
We summarize the hierarchy of small sets as folIows.
Proposition 7. A polar set is very thin; a very thin set is thin; a thin set is
semipolar; a semipolar set is oJ zero potential.
Proof. The first three implications are immediate from the definitions. If A is
semipolar, then JA is countable and so m(J A) = 0 almost surely by Theorem 6.
Hence U(·, A) == 0 by (8). This proves the last implication. 0
Let us ponder a little more on a thin set. Such a set is finely separated in
the sense that each ofits points has a fine neighborhood containing no other
point of the set. If the fine topology has a countable base (namely, satifies
the second axiom of countability), such a set is necessarily countable. In
general the fine topology does not even satisfy the first axiom of countability,
so a finely separated set may not be so sparse. Nevertheless Theorem 6
asserts that almost every path meets the set only countably often, and so
only on a countable subset (depending on the path). The question arises if
the converse is also true. Namely, if A is a nearly Borel set such that almost
every path meets it only countably often, or more generally only on a count-
able subset, must A be semipolar? Another related question is: if every
compact subset of A is semipolar, must A be semipolar? [Observe that if
every compact subset of A is polar, then A is polar by Theorem 8 of §3.3.]
These questions are deep but have been answered in the affirmative by
Dellacherie under an additional hypo thesis which is widely used, as folIows.
It follows easily that ifwe put ( = (oU" for any lJ. > 0, then Ais ofpotential
zero ifand only if ((A) = O. Hence under Hypothesis (L) such a measure exists.
It will be called a reJerence measure and denoted by ( below. For example,
for Brownian motion in any dimension, the corresponding Borel-Lebesguc
measure is a reference measure. It is trivial that there exists a probability
measure which is equivalent to (. We can then use P~ with ( as the initial
distribution.
3.5. Fine Properties 113
Theorem 8. Assume Hypothesis (L). Let A E g. and suppose that almost surely
the set JA in (6) is countable. Then A is semipolar.
Proposition 9. If f1 and f2 are two excessive functions such that f1 :::; f2 ~ -a.e.,
then f1 :::; f2 everywhere. In particular the result is true with equalities re-
J
placing the inequalities. If f is excessive and E c f d~ = 0, then f == O.
Proposition 10. Let A E 1%' •• Under (L) there exists a sequence of compaet sub-
sets K n of A such that K n i and for each x:
(15)
c = lim ~(~).
n
On the other hand, for each x it follows from Theorem 4 of §3.4 that there
exists a sequence of compact subsets Ln(x) such that
Therefore (16) also holds with the inequality reversed, hence it holds with
the inequality strengthened to an equality. Recalling the definition of f we
have proved that lim n P'i) = P~1. Let T Kn 1 S 2 TA- We have e{e- aS } =
e{e- aTA }, hence r{S = TA} = 1 which is (15). D
(17)
Proposition 12. Let A he a semipolar set. Then Jor each x, there exists F =
u~~ 1 K n where the Kn's are compact subsets of A, such that PX_a.s. we haue
TA - F = CD.
Prao/. Replace ~ by x in (17) and call the resulting measure V Since this is X
•
Exercises
1. Prove that g' is a O'-field and that rc g-.
2. Prove that if A is finely dense, then Ar = Ei" Furthermore for almost
every w, the set J A(W) is dense in T. [Hint: consider EX(e - TA:.]
3. Suppose that Hypothesis (L) holds. Then for ry, ?: 0, eachry,-excessive
function is Borelian. If A E 6, then Ar E 6. F or each t ?: 0, X -> P T ß ::;; t:
X
[
is Borelian. [Hint: the last assertion follows from the first through the
Stone-Weierstrass theorem; this is due to Getoor.]
4. Suppose that for some ry, > 0, all ry,-excessive functions are lower semi-
continuous. Then Hypothesis (L) holds. [Hint: consider AU a where J, =
Ln 2- ncx" and {x n} is a dense set in K]
5. Let AcE and A be a fine neighborhood of x. Then for ry, > 0 there exists
an ry,-excessive function cp and a compact set K such that {cp < I} c K c
A, and {cp < 1} is a fine neighborhood of x. [Hint: Let Y be open with
compact closure, B = AC u y c. By Theorem 8(a) of §3.3, there exists
open G::J B such that cp(x) = P{e- aTG } < 1. Take K = Ge.]
6. Fix ry, > O. If /!7 is a topology on Ea which renders all ry,-excessive functions
continuous, then /!7 is a finer topology than the fine topology, namely
each finely open set is open in .'!7. Thus, the fine topology is the least fine
topology rendering all ry,-excessive functions continuous. [Hint: consider
the cp in Problem 5.]
7. If .f is bounded and finely continuous, then for each ß ?: 0, lima~x
rt.Ua+ßf = f. As a consequence, if the function.f in Problem 4 of §3.2 is
finely continuous, then it is excessive.
8. Let A E g' and x E Ar and consider the following statement: "for almost
every w, X(t,w) = x implies that there exists t n 11 t such that X(tn,w) E
A." This statement is true for a fixed t (namely when t is a constant
independent of w), or when t = T(w) where T is optional; but it is false
when t is generic (namely, when t is regarded as the running variable in
t-> X(t,w)). [Hint: take x = 0, A = {O} in the Brownian motion in R I .
See §4.2 for a proof that 0 E [0]'.]
9. Let A be a finely open set. Then for almost every w, X(t, w) E A implies
that there exists 6(w) > 0 such that X(u,w) E A for all u E [t, t + 6(w)).
Here t is generic. The contrast with Problem 8 is interesting. [Hint:
suppose A E g' and let cp(x) = EX{e- TAc }. There exists 6'(w) ::;; 1 such
that cp(X(u,w))::;; A < 1 for u E [t, t + 6'(w)). The set B = AC n {cp::;; A}
is very thin. By Proposition 5 of §3.5, JB(w) n [t, t + 6'(w)) is a finite
set. Take 6(w) to be the minimum element of this set.]
10. If A is finely closed, then alm ost surely J A( w) is closed from the right,
namely, t n 11 t and t n E J A(W) implies t E J A(W), for a generic t.
116 3. Hunt Process
(4)
{J>J+c}= U {J/\m>(]/\m)+c}
m=l
00
Corollary. If the limit function f above vanishes except on a set of zero potential,
then it vanishes except on a polar set.
(5)
The corollary has a facile generalization which will be stated below to-
gether with an analogue. Observe that the condition (6) below holds for an
excessive f.
Proposition 2. Let f E 13": and suppose that we have for each compact K:
(6)
Proof. The first assertion was proved above. Let K c {f = oo}, then we have
as in (5):
P (x)
t
= _1 ex p
2nt
(_llx2tI12 ),
The next theorem deals with a more special situation than Theorem 1. I t
is an important part of the Riesz representation theorem for an excessive
function (see Blumenthai and Getoor [1], p. 84).
(8)
This relation says that the first hitting time on A after t is the first hitting
time of A after 0, provided that A has not yet been hit at time t. It follows
from (7) that
(9)
Thus the right member of(9) decreases as n increases. Hence for each k ~ no
we have for all n ~ k:
(11)
(12)
f = h + fo (15)
Proo/. For each x, the assumption (14) implies that there exists 10 = t o(x) such
that ptJ(x) < -co. Since PJ ::::; PtJfort ::::: t o, and PJPtof)(x) ::::; Ptof(x) <co
for s ::::: 0, the following limit relation holds by dominated convergence, for
each s ::::: 0:
Put j~ = f - h, so that .fo(x) =co if f(x) = co. Then we have for 1 ::::: (0:
(20)
A !ittle reflection shows that the set within the braces in (20) is exactly the
set of w for which J A(W) is an unbounded subset ofT, where JA is defined in
(6) of §3.5.
Definition. The set A is recurrent iff the probability in (19) is equal to one
for every x E E; and transient iff it is equal to zero for every x E E. Of course
this is not a dichotomy in general, but for an important dass of processes
it indeed is (see Exercise 6 of §4.1).
Clearly if A is recurrent then PA 1 == 1 (on E). The converse is true if and only
if(P t ) is strictIy Markovian. The phenomenon ofrecurrence may be described
in a more vital way, as folIows. For each A E rff· we introduce a new function:
°
where sup 0 = as by standard and meaningful convention. We call LA the
"quitting time of A", or "the last exit time from A". Its contrast to the hitting
time TA should be obvious. To see that LA E .'F-, note that for each w E Q:
It follows that
ro
{LA< oo}= U {TA 0 e n = cD}; (23)
n=1
n {TA
CI)
It is dear from the first equivalence in (22) that LA is not optional; indeed
{LA::s; t} E cr(X" SE (t, (0)), augmented. Such a random time has special
properties and is a kind of dual to a hitting time. We shall see much of it in
§5.l.
Exercises
I. Let I E t'+ and T = TA where A E fr Suppose that for each .\ and t > 0,
we have
I(x)?: EX {{(X t )1{t<n}·
Show that ifI(x) <x, then {{(X t ) 1{t < Ti' /#'10 PX} is a supermartingale.
2. In the notation of Case 1 of Theorem 3, show that for each compact
K cE, we have g = PK'g on {g <x}.
3. Under Hypothesis (L), let F be a dass of excessive functions. Then there
exists a decreasing sequence {In} in F such that if u = lim n 1 In we have
u?: inffEFI?: u. [Hint: choose in decreasing and Iim n dln/(l + f~)) =
inffEF~(f/(1 +f)).]
A Hunt process will be called recurrent iff any, hence every, one of the equi-
valent conditions in the following theorem is satisfied. We use the convention
below that an unspecified point, set, function is of, in, on E, not Ei). Thus
a function is said to be a constant ifit is a constant on E; its value at C, when
defined, is not at issue. Neverthless, the proof of Theorem 1 is complicated
by the fact that it is not obvious that assertion (1) holds prima Iacie under
each of the conditions (i) to (iv) there. The additional ca re needed to avoid
a
being trapped by is weil worth the pain from the methodological point of
view. Recall that , = T{(ll'
Theorem 1. The Iollowing Iour propositions are equivalent Ior a Hunt process,
provided that E contains more than one point.
(i) Each excessive Iunction is a constant.
(ii) F or each I E $ '+, either VI == 0 or VI == 00.
(iii) F or each B in $' which is not thin, we have P BI == 1.
(iv) For each B in $', either PBi == 0 or P Bl == 1. Namely, each nearly
Borel set is either polar or recurrent as defined at the end ol §3.6.
M oreover, any of these conditions implies the jcJ/lowing:
(1)
Proof (i) => (ii). Since V f is excessive, it must be a constant, say c, by (i).
Suppose c =1= 00. We have for t ~ 0:
(2)
and for n ~ 1:
Since X(ToJ E 0 1 C (0 2 )<, we have Tl < T 2 on {Tl< oo}. The same argu-
ment then shows that the sequence {Tn } is strictly increasing. Let lim n T n =
T; then on {T < oo}, X(T-) does not exist because X(Tn ) oscillates between
0 1 and O2 at a strictly positive distance apart. This is impossible for a
Hunt process. It follows that almost surely we have T = 00, and this implies
(1) because 0 is absorbing. Hence Ptc = C for all t.
We now have by (2) and dominated convergence:
Remark, It is instructive to see why (i) does not imply (ii) when E reduces
to one point.
(ii) => (iii). Fix BE C' and write qJ for P B 1. By Theorem 4 of §3.6, we have
qJ = h + qJo where h is invariant and qJo is purely excessive. Hence by Theorem
6 of§3.2, there existsfn E C~ such that qJo = lim n i Vfn. Since Vj,,:::; qJo:::; 1,
condition (ii) implies that Vfn == 0 and so qJo == O. Therefore qJ = PtqJ for
every t ~ o. If B is not thin, let Xo E Br so that qJ(x o) = 1. Put A = {x E EI
qJ(x) < I}. We have for each t ~ 0:
Since Pt(xo, Ei) = 1, the equation above entails that Pt(xo, A) = O. This
being true for each t, we obtain VI A (xo) = O. Hence condition (ii) implies
that VIA == O. But A is finely open since qJ is finely continuous. If XE A, then
VI A (x) > 0 by fine openness. Hence A must be empty and so qJ == 1.
124 3. Hunt Process
(iii) => (iv). Under (iii), P B l == 1 for any nonempty open set B. Hence (1) holds
as under (i). If B is not a polar set, there exists X o such that P Bl(xo) = i5 > O.
Put
(3)
(4)
As t ~ 00, the first member in (5) as weIl as the first term in the third member
converges to I{TH<"'}' Hence the se co nd term in the third member must
converge to zero. Now we have X(t) E E for t < 00 and therefore pX(ll[TB <
00] 2:: 6/2 by (4). It follows that limt~.x. 1{'<TB} = 0, PX-almost surely. This
isjust P B l(x) = 1, and x is arbitrary.
(iv) => (i). Let f be excessive. If fis not a constant, then there are real n umbers
a and b such that a < b, and the two sets A = {f < a}, B = U> b} are both
nonempty. Let x E A, then we have by Theorems 4 and 2 of §3.4:
since f 2:: b on B* by fine continuity. Thus P B l(x) < 1. But B being finely
open and nonempty is not polar, hence P B l == 1 by (iv). This contradiction
proves that f is a constant. 0
FinaIly, we have proved above that (i) implies (1), hence each of the other
three conditions also implies (1), which is equivalent to P,(x, E) = 1 for
every t 2:: 0 and x E E.
3.7. Recurrence and Transience 125
Remark. For the argument using (5) above, cf. Course, p. 344.
(7)
Since each i is regular for {i}, the only thin or polar set is the empty set.
Hence condition (iv) re duces to: p i {1{j) < oo} = 1 for every i andj. Indeed,
(8)
in the notation of (21) of §3.6. For a general Markov chain (7) and (8) are
equivalent provided that all states communicate, even when the process is
not a Hunt process because it does not satisfy the hypotheses (i), (ii) and (iii);
see Chung [2].
(10)
tinite limit time the path must hit the center by continuity. These statements
can be made quantitative by a more detailed study of the process.
Theorem 2. In general, ei/her (ii) or (iv) belmv implies (iii). Under [he condition
(11), (iii) implies (iv). Under the condition (12), (iv) implies (i).
Proof. (ii) = (iii) under (11). Let K n be compact and increase to E, and pul
(14)
3.7. Recurrence and Transience 127
CIJ 00 1
h= L I k2~n;n.
n=l k=l
(15)
Clearly h ::; 1. For each x, there exist n and k such that XE A nk by assump-
tion (ii). Hence h > 0 everywhere and so also Uh > O. It follows from (14)
that Uh ::; 1. Thus h has all the properties (and more!) asserted in (ii).
(iii) ~ (iv) under (11). Since Uh is finite, we have
(16)
Uh ) 1 1 (17)
PKI ::; PK ( C(K) = C(K) PKUh ::; C(K) Uh.
(18)
Define g as folio ws :
g = f
n= 1
gnn
n2 n ·
Then g::; 1 and 0< Ug::; 1 by (19). Now put h = uag; then h::; Ug::; 1.
It follows from the resolvent equation that
1 1
Uh = UUag ::; - Ug ::;-. (20)
rx rx
implies that
Hence Vh > 0 and h satisfies the conditions in (iii). Finally, under (12)
h is lower semi-continuous. Since h > 0, for each K we have infxEK h(x) =
C(K) > 0; hence by (20):
Vh 1
VI <-~<-
K - C(K) - a.C(K)'
namely (i) is true. Theorem 2 is completely proved. o
Let us recall that (iv) is equivalent to "every compact subset of E is
transient", i.e.,
VK: L K < CfJ almost surely. (21)
(22)
in particular for d = 3:
dy
V(x,dy) = -1--1'
-2nx-y (23)
I r dy nr 2
U(o, B) = 2n JB TYr = 2'
where B is the ball with center x and radius (j. The first term in the last
member of(24) is bounded uniformly in x by (2n)-11IfIIU(o,(j) = 4- 1 1IfIW.
The second term there is continuous at x by bounded convergence. It is
also bounded by (j -lllfllm(K). Thus Ufis bounded continuous. In particular
if f = 1K , this implies condition (i) of Theorem 2. The result also implies
condition (11) as remarked before. Hence all the conditions in Theorem 2
are satisfied. The case d > 3 is similar.
We can also compute for d = 3 and (J( ~ 0:
e-Izl~
For (J( > 0, an argument similar to that indicated above shows that VlIf
belongs to Co. This is stronger than the condition (12). For d ~ 3, (J( ~ 0,
ulI can be expressed in terms of Bessel functions. For odd values of d, these
reduce to elementary functions.
Let us take this occasion to introduce a new property for the semigroup
(Pt) which is often applicable. It will be said to have the "strong FeBer
property" iff for each t > 0 and f E btff +, we have Pd E C, namely continuous.
[The reader is warned that variants of this definition are used by other
authors.] This condition is satisfied by the Brownian motion in any dimen-
sion. To see this, we write
Exercises
1. For a Hunt process, not necessarily transient, suppose A Elf, BEg such
that Be A and U(x o, A) < 00. Then if inCll U(x, A) > 0, we have
pxo{L ll <oo} = 1, where L ll is the last exist time from B. [Hint: define
T n = n + T ll 8n, then U(xo,A) ~ PO(U(X(Tn),A); T n <00].]
2. Let X be a transient Hunt process. Let Albe compact, Anl and nnAn = B.
Put f = lim n PA)' Prove that fis superaveraging and its excessive regu-
larization J is equal to P B1. [Hint: f ~ J ~ PBl. If x ~ B, then
This section is devoted to a new assumption for a Hunt process which Hunt
listed as Hypothesis (B). What is Hypothesis (A)? This is more or less the
set of underlying assumptions for a Hunt process given in §3.l. Unlike
3.8. Hypothesis (B) 131
Hypothesis (B). Far same rx :2: 0, any A E ~., and open G such that Ac G,
we have
P~ = P'bP~. (1)
This means of course that the two measures are identical. Recalling
Proposition 1 of §3.4, for S = T G and T = TA, we see that (1) is true if
On the other hand, if rx > 0, then the following particular case of (1)
(3)
whereas the left member in (2) never exceeds the right member which is
inf{t> T G : X(t) E A}. We will postpone a discussion ofthe case rx = until °
the end of this section.
Next, the two members of (2) are equal on the set {T G < TA}. This is
an extension of the "terminal" property of TA expressed in (8) of §3.6. Since
G :::J A, we have always T G S; TA- Can T G = TA? [If A c G, this requires
the sampie path to jump from Ge into .4:.] There are two cases to consider.
Case (i). T G= TA < 00 and X(T Gl = X(T A) E Ar. In this case, the strong
Markov property at T G entails that TA ()ra = 0. Hence the two members
0
Now suppose that (2) holds and A is thin. Then neither case (i) nor case
(ii) above can occur. Hence for any open Gn :::J A we have alm ost surely
(5)
132 3. Hunt Process
Suppose first that x ~ A; then by Theorem 8(b) of §3.3 there exist open sets
Gn => A, Gn ! such that T Gn i TA, PX-a.s. Since left limits exist everywhere
for a Hunt process, we have limn~oo X(TGJ = X(T A-); on the other hand,
the limit is equal to X(T A) by quasi left continuity. Therefore, X(·) is con-
tinuous at TA, PX-a.s. For an arbitrary x we have
PX{TA < 00; X(T A-) = X(T A)} = lim PX{t < TA< 00; X(T A-) = X(T A)}
tlO
= lim PX{t < T A,. pX(t)[TA < 00·,
tlO
X(TA-)=X(T A)]}
= lim PX{t < TA; pX(tl[TA < oo]}
tlO
= lim PX{t < TA; P Al(X t)}
tlO
In the above, the first equation follows from Px { TA > O} = 1; the second by
Markov property; the third by what has just been proved because X(t) ~ A
for t < TA; and the fifth by the right continuity of t --> PA l(X t). The result is
as follows, almost surely:
(6)
Equation (6) expresses the fact that the path is almost surely continuous
at the first hitting time of A. This will now be strengthened to assert con-
tinuity whenever the path is in A. To do so suppose first that A is very thin.
Then by (12) of §3.5, the incidence set JA consists of a sequence {Tn , n ~ I}
of successive hitting times of A. By the strong Markov property and (6), we
have for n ~ 2, alm ost surely on {Tn_ 1 < oo}:
PX{Tn < 00; X(T n - ) = X(Tn )} = pX(Tn-d{TA < 00; X(T A-) = X(TAn
= pX(Tn-d{TA < oo} = PX{Tn < oo}.
Since (7) is true for each very thin set A, it is also true for each semi-polar set
by Theorem 6 of §3.5.
Thus we have proved that (2) implies the truth of (7) for each thin set
A c G. The converse will now be proved. First let A be thin and contained
in a compact K which in turn is contained in an open set G. We will show
that(4) is true. Otherwisethereis anx such thatPX{TG < oo;X(TG ) E A} > o.
Since A is at a strictly positive distance from Ge, this is ruied out by (7). Hence
3.8. Hypothesis (B) 133
Finally let A E C·, Ac G. By Theorem 8(b) of §3.5, for each x there ex ist
compact sets K n, K n c A, K ni such that T K n 1 TA, PX-almost surely. On
{TA < oo} we have T K n < 00 for all large n. Hence (8) with K = K n yields
(2)asn ~ 00.
We summarize the results in the theorem below.
Theorem 1. For a Hunt process the following four propositions are equivalent:
(i) Hypothesis (B), namely (1), is true for some IX> O.
(ii) Equation (3) is true for some IX > O.
(iii) For each semi-polar A and open G such that A c G, we have
P G(x, A) = 0 for each x.
(iv) For each semi-polar A, (7) is true.
Theorem 2. 1f the Hunt process is transient in the sense that it satisfies condition
(iii) of Theorem 2 of §3.7, then the following condition is also equivalent to the
conditions in Theorem 1.
(v) Hypothesis (B) is truefor IX = O.
Proof. Let h > 0 and Uh ::; 1. Then applying (1) with IX = 0 to the function
Uh, we obtain
positive. Hence the preceding equation forces this set to be alm ost surely
empty. Thus (2) is true, which implies direct1y (1) for every IX 2 O. D
Needless to say, we did not use the full strength of (v) above, since it is
applied only to a special kind of function. It is not apparent to wh at extent
the transience assumption can be relaxed.
In Theorem 9 of §3.3, we proved that for a Hunt process, TB ::; TB a.s. for
each B E C·. It is trivial that TB = TB if all paths are continuous. The next
134 3. Hunt Process
Theorem 3. Under Hypotheses (L) and (B), we have TB = TB a.s. for each
BE rf,u ••
Proof. The key idea is the following constrained hitting time, for each BEg':
A = B n{xlcp(x) s 1 - F.}.
Define a sequence ofoptional times as follows: R 1 = SA' and for n:?: 1:
Hence Rn i Cf) a.s. Put /(W) = (t > 0IX(t,W) E A}. By the definition of Rn'
for a.e. W such that Rn(w) < 00, the set I(w) n (Rn(w),Rn+1(w)) is contained
in the set of t where X(t-,w) =f. X(t,w), and is therefore a countable set.
Since Rn(w) i 00, it follows that /(w) itselfis countable. Hence by Theorem 8
of §3.5, A is a semi-polar set. This being true for each c;, we conclude that the
set B' = B n {x 1 cp(x) < I} is semi-polar.
Now we consider three cases on {TB <oo}, and omit the ubiquitous "a.s."
(i) X(T B-) = X(T B). Since B is compact, in this case X(T B-) E Band
so TB S TB'
(ii) X(T B-) =f. X(T B) and X(T B) E B - B'. In this case cp(X(TBll = 1
and so SB 8 TB = 0. Since TB S SB we have TB S TB'
0
(iii) X(T B- ) =f. X(T B) and X(T B) E B'. Since B' is semi-polar this is ruled
out by Theorem 1 under Hypothesis (B).
3.8. Hypothesis (B) 135
We have thus proved for each compact B that Ti :s:: TB on {TB < oo}; hence
Ti = TB by Theorem 9 of §3.3. For an arbitrary BEg·, we have by Theorem
8(b) of§3.3, for each x a sequence of compacts K n c B, K n i such that PX{ T K n 1
TB} = 1. Since T Kn = T K nas just proved and Ti :s:: T Kn for n ;:::: 1, it follows
that p x {Ti :s:: TB} = 1. As before this implies the assertion ofTheorem 3. D
Exercises
1. Give an example such that P AI = P GP AI but PA;:::: PGP A is false. [Hint:
starting at the center of a circle, after holding the path jumps to a fixed
point on the circle and then execute uniform motion around the circle.
This is due to Dellacherie. It is a recurrent Hunt process. Can you make
it transient?]
2. Under the conditions of Theorem 1, show that (7) remains true if the set
{XI E A} is replaced by {XI _ E A}. [Hint: use Theorem 9 of §3.3.]
3. If (7) is true for each compact semi-polar set A, then it is true for every
semi-polar set. [Hint: Proposition 11 of §3.5 yields a direct proof without
use of the equivalent conditions in Theorem 1.]
4. Let E = {O} u [1, (0); let 0 be a holding point from which the pathjumps
to 1, and then moves with uniform speed to the right. Show that this is a
Hunt process for which Hypothesis (B) does not hold.
5. Let A be a thin compact set, Gn open, and Gn U A. Show that under
Hypothesis (B), we have under px, x ~ A:
NOTES ON CHAPTER 3
§3.1. Hunt process is named after G. A. Hunt, who was the author's classmate in
Princeton. The basic assumptions stated here correspond roughly with Hypothesis
(A), see Hunt [2]. The Borelian character of the semigroup is not always assumed. A
more general process, called "standard process", is treated in Dynkin [1] and BlumenthaI
and Getoor [1]. Another technical variety called "Ray process" is treated in Getoor [1].
§3.2. The definition of an excessive function and its basic properties are due to Hunt
[2]. Another definition via the resolvents given in (14) is used in BlumenthaI and Getoor
[1]. Though the former definition is more restrictive than the latter, it yields more
transparent results. On the other hand, there are results valid only with the resolvents.
136 3. Hunt Process
§3.3. We follow Hunt's exposition in Hunt [3]. Although the theory of analytical sets
is essential here, we follow Hunt in citing Choquet's theorem without details. A proof
of the theorem may be found in Helms [1]. But for the probabilistic applications (pro-
jection and section) we need an algebraic version of Choquet's theory without endowing
the sampie space Q with a topology. This is given in Meyer [1] and Dellacherie and
Meyer [1].
§3.4 and 3.5. These seetions form the co re of Hunt's theory. Theorem 6 is due to
Doob [3], who gave the prooffor a "subparabolic" function. Its extension to an excessive
function is carried over by Hunt using his Theorem 5. The transfinite induction used
there may be concealed by some kind of maximal argument, see Hunt [3] or Blumenthai
and Getoor [1]. But the quiekest proof of Theorem 6 is by means of a projection onto
the optional field, see Meyer [2]. This is a good place to learn the new techniques
alluded to in the Notes on §1.4.
Dellacherie's deep result (Theorem 8) is one of the cornerstones of wh at may be
called "random set theory". Its proof is based on a random version of the Cantor-
Bendixson theorem in classical set theory, see Meyer [2]. Oddly enough, the theorem
is not included in Dellacherie [2], but must be read in Dellacherie [1], see also
Dellacherie [3].
§3.6. Theorem 1 originated with H. Cartan for the Newtonian potential (Brownian
motion in R 3 ); see e.g., Helms [1]. This and several other important results for Hunt
processes have companions wh ich are valid for left continuous moderate Markov
processes, see Chung and Glover [1]. Since a general Hunt process reversed in time has
these properties as mentioned in the Notes on §2.4, such developments may be interesting
in the future.
"Last exit time" is hereby rechristened "quitting time" to rhyme with "hitting time".
Although it is the obvious concept to be used in defining recurrence and transience, it
made a belated entry in the current vocabulary ofthe theory. No doubt this is partly due
to the former prejudice against a random time that is not optional, despite its demon-
strated power in older theories such as random walks and Markov chains. See Chung
[2] where a whole section is devoted to last exits. The name has now been generalized to
"co-optional" and "co-terminal"; see Meyer-Smythe-Walsh [1]. Intuitively the quitting
time of a set becomes its hitting time when the sense of time is reversed ("from infinity",
unfortunately). But this facile interpretation is only a heuristic guide and not easy to
make rigorous. See §5.1 for a vital application.
§3.7. For a somewhat more general discussion of the concepts of transience and
recurrence, see Getoor [2]. The Hunt theory is mainly concerned with the transient
case, having its origin in the Newtonian potential. Technically, transience can always
be engineered by considering (P~) with C( > 0, instead of (Pt). This amounts to killing
the original process at an exponential rate e-· t , independently ofits evolution. 11 can be
shown that the resulting killed process is also a Hunt process.
§3.8. Hunt introduced his Hypothesis (B) in order to characterize the balayage
(hitting) operator PA in a way recognizable in modern potential theory, see Hunt [1; §3.6].
He stated that he had not found "simple and general conditions" to ensure its truth.
Meyer [3] showed that it is implied by the duality assumptions and noted its importance
in the dual theory. It may be regarded as a subtle generalization ofthe continuity ofthe
paths. Indeed Azema [1] and Smythe and Walsh [1] showed that it is equivalent to the
quasi left continuity of a suitably reversed process. We need this hypo thesis in §5.1 to
extend a fundamental result in potential theory from the continuous case to the general
case. A new condition for its truth will also be stated there.
Chapter 4
Brownian Motion
(1)
(2)
(3)
where the integral is over E and * denotes the convolution. This is valid for
s ;::::: 0, t ;::::: 0, with n o the unit mass at 0. We add the condition:
and Ua(x, A) = Ua(A - x). Then U a is the ex-potential kernel and we have
for f E et+:
u1 = f f(x + y)Ua(dy). (6)
= f [f m(dX)f(x)]n,(dY ) = f m(dx)f(x).
(7)
mP,=m, Vt z O. (8)
Proof. To prove the first assertion, suppose A E tff and m(A) = O. Then
U"(x, A) = 0 for m-a.e. x by (9). Since U"l A is lower semi-continuous, this
implies U"(A) = U"(o, A) = O. Next, if U" « mIet u" be a Radon-Nikodym
derivative of U" with respect to m. We may suppose u" 2:': 0 and u" E tff.
Then (6) may be written as
Since u" E L l(m), a c1assical result in the Lebesgue theory asserts that
(11)
(see e.g. Titchmarsh [1], p. 377). Since fis bounded, the last term in (10)
shows that U"f is continuous by (11). D
It is c1ear that (XI) is a spatially homogeneous process. Let PI' 0", fi l , fl" be
the quantities associated with it. Thus
It is convenient to introduce
140 4. Brownian Motion
Thus we have
(13)
and
(In the last relation it is tempting to write V"(A, x) far V"'(x, A). This is indeed
a good practice in many lengthy formulas.) Now let J E b{,u +, then
for a measure fl. If follows from (15) and (16) that if fl is a (I-finite measure
as weil as m, we have by Fubini's theorem:
f Vj(y)fl(dy) = f m(dx)J(X)U'fl(X);
Under these conditions we say the processes are in duality. In this case the
relation (17) holds for any JE 0"+ and any (I-finite measure fl; clearly it
also holds with U'" and (j'" interchanged. Further conditions may be put
on the dual potential density function u(·,·). We content ourselves with
one important result in potential theory wh ich f10ws from duality.
Theorem 3. Assume duality. Let fl and v be two (I-finite measures such that Jor
same rx ~ 0 we haue
U"'fl = U"'v< 00. (19)
Then fl == v.
4.1. Spatial Homogeneity 141
Thus if (19) is true it is also true when IX is replaced by any greater value.
To illustrate the methodology in the simplest situation let us first suppose
that p. and v are finite measures. Take f E bC, then
boundedly because (X,) is a Hunt process. It folIo ws from this, the duality
relation (17), and the finiteness of p. that
This is also true when p. is replaced by v, hence by (19) for alIlarge values of
IX we obtain
°
Put h = vag; then clearly hE blff, h > and Shdp. < 00. Moreover, h being
an IX-potential for (X,), is IX-excessive for (P,) (calIed "IX-co-excessive"). Thus
it follows from (11) of §3.2 that lim a _ oo IXVah = h. But we need a bit more
than this. We know [rom the Corollary to Theorem 1 of §3.5 that h is finely
continuous with respect to (X,) (calIed "co-finely continuous"); hence so is
fh for any f E bC, because continuity implies fine continuity and continuity
in any topology is preserved by multiplication. It follows (see Exercise 7
of §3.5) that
142 4. Brownian Motion
Proof. We leave the converse part to the reader (cf. Theorem 9.2.2 of Course).
To prove the direct assertion we use Fourier transform (characteristic
functions). We write for x ERd, Y ERd:
if x = (Xl' . . . ,Xd ), Y = (Yl, ... ,Yd)' Let i = P and Uk ERd; and set
X(L d = o. We have
(21 )
n+l
fito(uo) n ntk-tk_l(Uk),
k=1
where fit is the characteristic function ofthe distribution {Lt of X(t). Assertion
(a) follows from this by a standard result on independence (see Theorem 6.6.1
of Course), and assertion (b) from taking expectations in (22). Note that 1ts
is the distribution of X(s) - X(O). 0
Exercises
1. Is m the unique a-finite measure satisfying (8), apart from a multiplicative
constant? [This is a hard question in general. For the Brownian motion
semigroup the answer is "yes"; for the Poisson semigroup the answer is
"no".]
2. For the uniform motion semigroup (Pt) (Example 2 of §1.2), show that
U~ « m and find u~(x, y). In this case Plx, . ) is singular with respect to m
for each t and x. For the Poisson semigroup (Pt) (Example 3 of §1.2),
find an invariant measure mo (on N). In this case Pt(n,·) « mo for each
t and n.
The next three exercises are valid for a process having stationary in-
dependent increments.
144 4. Brownian Motion
4. !>ut
Here
The proof requires Exercise 2 of §6.6 of Course, which states that mutually
orthogonal Gaussian random variables are actually independent.
A trivial extension of the definition consists of substituting a 2 t for t in (1),
where a 2 > O. We deern this an unnecessary nuisance. In this section we
review and adduce a few basic results for the Brownian motion. We begin
with the most important one.
(I) There is aversion of the Brownian motion with continuous sam pIe
paths.
For d = 1 the proofis given in the Example in §3.1. The general case is an
immediate consequence of the coordinate representation (even without
independence among the coordinates). This fundamental property is often
taken as part of the definition of the process. For simplicity we may suppose
all sam pIe functions to be continuous.
(11) The Brownian motion semigroup has both the FeIler property and
the strong FeIler property.
The latter is verified in §3.7, and the former in Exercise 2 of §2.2. As
defined, the strong FeIler property does not imply the FeIler property!
Recall also from Example 4 of §3.7 that Uf is bounded continuous if fE btff
and f has compact support, and Uaf for IX > 0 is bounded continuous for
fE btff. These properties can often be used as substitutes for the FeIler
properties.
(111) The Brownian motion in Rd is recurrent for d = 1 and d = 2; it is
transient for d ~ 3.
This is verified in Examples 2 and 4 of §3.7. Let us mention that for d = 1
and d = 2, there are discrete-time versions of recurrence which supplement
the conelusions ofTheorem 1 of §3.7. For each (j > 0, let {X(n(j), n ~ I} be a
"skeleton" ofthe Brownian motion {X(t), t ~ O}. Then {X(n(j) - X((n - 1)(j),
n ~ I} is a sequence of independent and identically distributed random
variables, with mean zero and finite second moment (the latter being needed
only in R 2 ). Hence the recurrence criteria for a random walk in R 1 or R 2
(§8.3 of Course) yield the following result. For any nonempty open set G:
This is just one illustration of the elose tie between the theories of random
walk and of Brownian motion. In fact, all the elassicallimit theorems have
146 4. Brownian Motion
their analogues for Brownian motion, which are frequently easier to prove
because there are ready sharp estimates. Since our emphasis here is on the
process we shall not delve into those questions.
Let us also recall that for any d ::?: 1, each set is either recurrent or transient
by Exercise 6 of §4.1.
(IV) A singleton is recurrent for Brownian motion in R 1, and polar in
Rd for d::?: 2.
The statement for R 1 is trivial by recurrence and continuity of paths.
The statement for Rd, d ::?: 2, is proved in the Example of §3.6. Another
proof will be given in §4.4 below.
As a consequence, we can show that the fine topology in R d , d ::?: 2, is
strict1y finer than the Euclidean topology. For example, if Q is the set of
points in R d with rational coordinates then R d - Q is a finely open set
because Q is a polar set. It is remarkable that all the paths "live" on this
set full of holes.
(V) Let
F = {x E Rd Ix I = O},
(3)
Then F c G; n G;.
To prove this we need consider only the ca se d = 1, but we will argue
in general. Let x E F. By the symmetry of the distribution of X l(t), we have
Since PX{X(t) E F} = °
by the continuity of the distribution of XI(t), we
deduce that r{X(t) E GI} =~. Hence for each u > 0:
(4)
4.2. Preliminary Properties of Brownian Motion 147
This is a consequence of the continuity of paths, but let us give the details
of the argument. By the definition of TB, on the set {TB< oo} for any e > 0
there exists tE [TB, TB + e) such that X(t) E B. Hence either X(T B) E B,
or by the right continuity of the path,
On the other hand, on the set {O< TB}, X(t) E Be for 0< t < TB, hence
by the left continuity of the path,
(5)
(VIII) Let .~ be either .~? or fi'; (see §2.3). Then for any x,
(6)
Thus (2) is true and the characterization mentioned there yields the
conclusion.
(X) Let B'be a Borel set such that m(B) < w. Then there exists e > 0
such that
sup e{exp(eTBc)} < W. (8)
XE E
This number may be made ::::;1 if t is large enough. Fix such a value of t
from here on. It follows from the Markov property that for x E 13 and
n;::: 1;
00
00
This simple method yields other useful results, see Exercise 11 below.
Let B(x, r) denote the open ball with center x and radius r, namely
Then under p x , T, is just TBc where B = B(x, r); it is also equal to TaB. The
next result is a useful consequence ofthe rotational symmetry ofthe Brownian
motion.
(XI) For each r > 0, the random variables T, and X(T,) are independent
under any P. Furthermore X(T,) is uniformly distributed on S(x, r) under
PX.
A rigorous proof of this result takes longer than might be expected, but
it will be given. To begin with, we identify each w with the sampIe function
X(-, w), the space Q being the dass of all continuous functions in E = R d .
Let qJ denote a rotation in E, and qJW the point in Q which is the function
X(·, qJw) obtained by rotating each t-coordinate of X(·, w), namely:
(14)
150 4. Brownian Motion
It foHows from (VI) that X(Tr) E S(X(O), r). Let X(O) = x, t 2 0 and A be
a Borel subset of S(x, r); put
(16)
Observe the double usage of <p in <p - 1 A and <p -1 A above. We now claim
that if x is the center of the rotation <p, then
(17)
a(A)
A E g, A c S(x, r) (19)
cr(r) ,
Since the total mass of the measure in (18) is equal to P{Tr ::::; t}, it follows
that the left member of(18) is equal to this number multiplied by the number
in (19). This establishes the independence of T r and X(Tr) since t and A
are arbitrary, as weH as the asserted distribution of X(TJ
Let us introduce also the foHowing notation for later use:
f rd
v(r) = m(B(x, r)) = J~ cr(s) ds = d cr(l). (21 )
4.2. Preliminary Properties of Brownian Motion 151
Proposition. Let {Xt} be a Markov process with transition function (Pt). Let
cp be a Borel mapping of E into E such that for all t ~ 0, X E E and A E tff, we
have
Then the process {cp(X t)} under px has the same finite-dimensional distribu-
tions as the process {X t} under p'P(x).
The left member of (22) is, by the Markov property and the induction
hypo thesis, equal to
W{f1(cpXtJPXC1[JicpX'2-tJ'" !z(cpX,,-t,_J]}
= EX{f1(cpXt,)P'PX"[J2(XtrtJ ... .h(Xt,-t,_ J]}
(23)
are both universally measurable, namely in tff-/fJ6 where fJ6 is the Borel
fie1d on R 1 •
152 4. Brownian Motion
Thus by definition f(X t ) E /\11 .,#,11 = g;-. Now let T r be as in (10), then
under px, X(T r) E S(x, r) by (VI). But the Lebesgue measurability of f does
not guarantee its measurability with respect to the area measure (J on S(x, r),
when f is restricted to S(x, r). In particular if 0 ~ f ~ 1 we can alter the
fl and f2 above to make fl = 0 and f2 = 1 on S(x, r), so that E''lf~(Xt) -
fl(X t )} = (J(S(x,r)). It should now be clear how the universal measurability
of f is needed to overcome the difficulty in dealing with f(X(T r )). Let us
observe also that for a general Borel set B the "surface" aB need not have
an area. Yet X(T,oB) induces a measure on aB under each pl1, and if f is
universally measurable P{f(X(TcB ))} may be defined.
Exercises
Unless otherwise stated, the process discussed in the problems is the
Brownian motion in Rd , and (Pt) is its semigroup.
1. If d ~ 2, each point x has uncountably many fine neighborhoods none
of which is contained in another. For d = 1, the fine topology coincides
with the Euclidean topology.
2. For d = 2, each line segment is a recurrent set; and each point on it
is regular for it. For d = 3, each line is a polar set, whereas each point
of a nonempty open set on a plane is regular for the set. [Hint: we can
change coordinates to make a given line a coordinate axis.]
3. Let D be an open set and x E D. Then P T DC = T"D) = 1. Give an
X
(
aD there exists a line segment yy' E D<, then PX{TDc = TU)c} = I for
every XE D. [Hint: use (V) after a change of coordinates.]
4. Let B be abalI. Compute e {T ßC} for all x E E.
5. If BE tff and t > 0, then PX{TB = t) = O. [Hint: show Jp X{ TB = s} dx = 0
for all but a countable set of s.]
4.2. Preliminary Properties of Brownian Motion 153
6. For any Hunt process with continuous paths, and any closed set C,
we have Tc E fJ'0. Consequently x --+ PX{Tc :::; t} and x --+ EX{f(X(Td)
are in C for each t ~ 0 and fE bC or C +. Extend this to any CE C
by Proposition 10 of §3.5. [Hint: let C be closed and Gn t! C where
Gn is open; then nn{3tE[a,b]:X(t)EGn} = {3tE[a,b]:X(t)EC}.]
7. For any Hunt process satisfying Hypothesis (L), BE C, fE flJ and fE C+
respectively, the functions x --+ EX{f(TBn and x --+ EX{f(X(TB))} are
both in C. [Hint: by Exercise 3 of §3.5, x --+ EX{e-~TB} is in C for each
0( > 0; use the Stone-Weierstrass theorem to approximate any function
This is called the "(first) exit time" from D. The usual convention that Tl) =
°
+ 00 when X(t) E D for all t > is meaningful; but note that X(oo) is generally
notdefined. By(X)of§4.2, wehavePX{T v < oo} = 1 forallx E Difm(D) <00,
in particular if D is bounded.
The cJass ofBorel measurable functions from R d to [ -00, +x] or [0, cx;]
will be denoted by ,g or ,g +; with the prefix "hO' when it is bounded. We say
f is 10ca1\y integrable in D and write fE Lioc(D) iff JK dm <x for each Ifl
compact K c D, where m is the Lebesgue measure. We say A is "strictly
contained" in Band write A € B iff A c D.
Letr E g +, and define a function h by
This has been denoted before by PTlJ or PvJ. Although h is defined for al\
x E E, and is cJearly equal to f(x) if x E (D')O = (15)', we are mainly concerned
with it in 15. It is important to note that h is universally measurable, hence
Lebesgue measurable by (XII) of§4.2. More specifically, it is Borel measurable
by Exercise 6 of §4.2 because D' is cJosed and the paths are continuous.
Finally ifr is defined (and Borel or universal\y measurable) only on aD and
we re pi ace T v by T"l) in (2), the resulting function agrees with h in D (Exercise
1).
Recall the notation (11), (12) and (20) [rom §4.2.
Ir(x) = -I-I.
rT(r) S(x.r)
h(Y)rT(dy). (3)
Furthermore, h is continuous in D.
Prooj". Write B for B(x, r), then almost surely TB < x; under P\ TB< T v so
that Tl) = TB + T v GTB • Hence it follows from the fundamental balayage
formula (Proposition 1 of §3.4) that
(4)
4.3. Harmonie Funetion 155
which is just
(5)
This in turn becomes (3) by the second assertion in (XI) of§4.2. Let us observe
that (4) and (5) are valid even when both members ofthe equations are equal
to + 00. This is seen by first replacing f by f 1\ n and then letting n ~ 00 there.
A similar remark applies below. Replacing r by s in (3), then multiplying by
O"(s) and integrating we obtain
h(x)
JoIr O"(s) ds = Ir JS(x,s)
Jo I h(y)O"(dy) ds = JB(x,r)
I h(y)m(dy). (6)
where in the last step a trivial use of polar coordinates is involved. Recalling
(21) of §4.2 we see that (6) may be written as
h(x) = _(1) I
v r JB(x,r)
h(y)m(dy), (7)
h(x) = _(1) I
vp JB(x,p)
h(y)m(dy) ~ _(1) I
V p JB(xQ,r)
h(y)m(dy)
(8)
Thus h < 00 in B(x o, r). This shows that the set F = D n {x Ih(x) < oo} is
open. But it also shows that F is c10sed in D. For if X n E Fand X n --+ X oo E D,
then X oo E B(x n , 15) E D for some 15 > 0 and some large n, so that X oo E F by
the argument above. Since D is connected and our hypothesis is that F is not
empty, we conc1ude that F = D. The relation (7) then holds for every x in D
with the left member finite, showing that h E L~c(D).
To prove that h is continuous in D, let r be so small that both B(x, r) and
B(x', r) are strictly contained in D. Applying (7) for x and x' we obtain
where C = B(x, r) 6. B(x', r). If x is fixed and x' --+ x then it is obvious that
m( C) --+ 0, and the integral in (8) converges to zero by the proven integrability
of h. Thus h is continuous and Theorem 1 is proved. D
h(x)v(r) = Ir I
Jo JS(x.sl
h(Y)(J(dy)ds.
By the proof ofTheorem 1, if Dis a domain and h E g +, h oJ= Cf.:; in D, then (b)
entails (a). Another easy eonsequenee is as folIows.
Ll = Ld (cl)2
l
j= 1 (X j
is the Laplacian in R d .
Llh = O. (10)
For each X in D, there exists <5 > 0 such that h is bounded in B(x, <5) and (3)
holds for 0 < r < <5. It follows that
h(x) I
= Jo CXl
[~( I )
a r JS(x,r)
h(y)a(dY)] <p(r)a(r) dr
I
= Jo Xl
fS(x,r)
h(y)<p(ilx - yll)a(dy)dr (12)
= SE h(y)<p(llx - yli)m(dy).
Since <p has support in B(o, <5), and all partial derivatives of<p are bounded
continuous in B(o, <5), we may differentiate the last-written integral with
respect to x under the integral. Since <p(llx - yll) is infinitely differentiable
in x for each y, we see that h is infinitely differentiable.
To prove (10) let us recall Gauss's "divergence formula" from calculus, in
a simple situation. Let B be a ball and h be twice continuously differentiable
in a neighborhood of B, then the formula may be written as follows:
where oh/on is the outward normal derivative. Now let us denote the right
member of (3) by A(x, r). Making the change of variables y = x + rz we have
A(x, r) I
= a (11) JS(o,l) h(x + rz)a(dz). (14)
, d ah oh
h (x + rz) = I Zj;l (x + rz) = -;- (x + rz); (15)
j=l uZj un
A'(x, r) I
= a (11) JS(o,l) ~h (x + rz)a(dz)
un
1 I ah
= a(r) JS(x,r) on (y)a(dy)
= ~( )I
a r JB(x,r)
,1h(y)m(dy). (16)
The differentiation of(13) under the integral is correct because h has bounded
continuous partial derivatives in a neighborhood of S(o, 1). The first member
of (16) vanishes for all sufficiently small values of r by (3), hence so does the
integral in the last member (rid of the factor l/a(r)). This being true for all
158 4. Brownian Motion
Proof. Even without the use ofDefinition 2, it is easy to see that we need only
prove that h x • is harmonie in eaeh bounded subdomain D J @ D. On D J ,
4.3. Harmonie Funetion 159
logllxll = _(1)
0" r
r
JS(x,r)
10gllzII0"(dz) =~) r logllx -
O"(r JS(o,r)
yIIO"(dy);
similarly for the seeond integral in (17). However, to evaluate the integrals
when Ilxll < r we need the following proposition.
Proqf. Writing p for Ilxll, we see that the Laplaee equation for h reduees to
dZ d- 1 d
-h+---h=O. (19)
df? p dp
160 4. Brownian Motion
_1
v(r)
r
JS(o.r)
logllx - Yllv(dy) = 10g(llxll v r), d = 2; (20)
cl ~ 3. (21)
Proo.f. Denote the left member of (20) by fix). It is harmonie in B(o, r) and
is a funetion of Ilxll alone upon a geometrie inspeetion. Henee by Proposition
5, fIx) = a logllxll + b. Far x = 0 we have f(o) = log r by inspeetion; henee
a = 0 and b = log r. It follows that fIx) = log r for Ilxll < r. Similarly, fIx) =
a' logllxll + b' far Ilxll > r. It is easily seen that limllxll~ [fix) - logllxll =
0; henee a' = 1 and b' = O. Thus fix) = logllxll for Ilxll > r. We have already
observed this above. It remains to show that f is finite eontinuous at Ilxll = r.
This is a good exereise in ealculus, not quite trivial and left to the reader.
The evaluation of (21) is similar. 0
The exaet ealculation of formulas like (20) and (21) farms a vital part of
classieal potential theory, and is the souree of many interesting results.
Exercises
1. Far any open set D show that PX[ X(TD) E cD} = 1 for a11 XE 15. However,
if XE ilD it is possible that PX{T D < T cv } = 1. An example is known as
Lehesyue' s thorn, see Port and Stone [1], p. 68.
2. Show that if h is harmonie then so are aU its partial derivatives of a11
orders. Verify the harmonieity of the following funetions in R 3 :
2x~ - xi - x~.
Ilxll'
3. Let h be harmonie in R d , d ~ 1 such that Ptlhl <x for all t > O. Then
h = Pth for a11 t ~ 0, where (Pt) is the semigroup of the Brownian motion
The proeess {h(X t), §;o t ~ O} is a eontinuous martingale. [Hint: use
polar coordinates and the sphere-averaging property; a by-produet is
the value of v(l).]
4. Let h E g, 0::;; h < CD and h = P 1h. Then h is a eonstant. [Hint: if h 1= x
then h is 10ea11y integrable; now eompare the values of P)J at two points
as n -+ 00, by elementary analysis. This solution is due to Hsu Pei.]
4.3. Harmonie Funetion 161
[Hint: the c1assieal proof uses inequalities whieh are derived from the
Poisson representation for a harmonie funetion in a ball (see §4.4). But a
proof follows from (8), followed by the usual "ehain argument."].
10. Let a > 0, b > O. Evaluate the following integrals by means of Pro-
position 6:
S 10g(a
OZ1t
Z + bZ - 2ab eos 8)d8;
12. Compute
Let D be an open set and J E bIS on iJD, and eonsider the funetion h defined
in 15 by (2) of §4.3. Changing the notation, we shall denote h by H DJ, thus
XE 15. (1)
This formula defines a measure H D(X, .) on the boundary iJD, ealled the
harmonie measure (of x with respeet to D). For eaeh Borel subset A of DD,
the funetion H(', A) is harmomie in D by the Corollary to Theorem 1 of
§4.3. We may regard the measure H(x,') as defined for all Borel sets, though
it is supported by iJD. The measure H(x,') is also defined for XE DC and is
harmonie in (15Y = (DC)o. For example if GD is the horizontal axis in R 2 ,
H Df is a harmonie funetion both in the upper and lower open half-planes,
but not neeessarily in the wh oie plane.
We now study the behavior of h(x) as x approaehes the boundary. Sinee
h is defined only in 15, we shall not repeat the restrietion that XE 15 below.
Reeall that a point z on iJD is regular or irregular Jor D aeeording as P r D =C Z
[
is lower semi-continuous in E.
n 0{VS G,
1 E t }X(S) E D}.
The set above is in ~7, and
(2)
The probability on the right side of (2) is of the form pt/ncp(x), where cp(x) E
bC. Hence by the strong Feller property «11) of §4.2), x ~ pt/ncp(x) is bounded
continuous. Therefore the left member of (2) is upper semi-continuous in x,
which is equivalent to the proposition. 0
(4)
Hence the corresponding limit exists and is equal to one. Note that here x
°
need not be restricted to [5. Let T r be as in (10) of §4.2. For r > and e > 0,
PX{T, > e} does not depend on x, and for any fixed r we have
because T, > ° almost surely. Now for any x, an elementary inequality gives
164 4. Brownian Motion
(7)
and so by (5):
lim r{'D < T r } = 1. (8)
On the set {'D< T r }, we have IIX('D) - X(O)II < rand so IIX('D) - zll <
IIX(O) - zll + r. Since f is continuous at z, given b > 0, there exists I] such
that if Y E oD and IIY - zll < 21], then If(Y) - f(z)1 < b. Now let XE l5 and
Ilx - zll < 1]; put r = 1]. Then under r, X('D) E oD (Exercise I of §4.3) and
IIX('D) - zll < 21] on {'D < T~}; hence If(X('D)) - f(z) I < b. Thus we have
P{lf(X('D)) - f(z)I;, D< oo} < bPX{'D < T~} + 211J11r{ T q ::; 'D < (0). (9)
When x -+ z, the last term above converges to zero by (8) with I] for r. Since
f> is arbitrary, we have proved that the left member of (9) converges to zero.
As a consequence,
lim h(x) = f(z) lim P X{, D < oo} = f(z)
by (4). o
Let us supplement this result at once by showing that the condition of
regularity of z cannot be omitted for the validity of (3) in Rd, d ~ 2. Note:
in R 1 every z E oD is regular for {z} c DC•
Proof. Indeed (10) is true for any fE bc.{oD) such that f(z) = 1 and f < 1 on
GD - {z}. An example of such a function is given by (1 -llx - .:11) v O.
Since {z} is polar by (IV) of §4.2, and z is irregular for D we have p z {X(, D) =
C,
where the equation follows from lim dO PZ{Tr < 'D} = 1, and the strong
Markov property applied at T r on {Tr < 'D}. The last member above is equal
to the left member of (10). 0
4.4. Dirichlet Problem 165
Praof. We may suppose that z is the origin and the flat cone lies in the
hyperplane ofthe first d - 1 coordinate variables. In the following argument
all paths issue from o. Let
is regular, see Figure (p. 165). Of course, a ball or a cube is regular; for a
simpler proof, see Exercise 1.
The most classical form of the Dirichlet boundary value problem may be
stated as follows. Given an open set D and a bounded continuous funetion
f on oD, to find a function whieh is harmonie in D, eontinuous in D, and
equal to f on oD. We shall refer to this problem as (D,j). The Corollary to
Theorem 1 of §4.3, and Theorem 2 above taken together establish that the
funetion HDf is a solution to the problem, provided that D is regular as
just defined. When B is a ball, for instanee, this is ealled the "interior Dirichlet
problem" if D = B, and the "exterior Dirichlet problem" if D = (ßt The
neeessity of regularity was aetually diseovered long after the problem was
posed. If there are irregular boundary points, the problem may have no
solution. The simplest example is as follows. Let D = B(o, 1) - {o} in R Z,
namely the punetured unit disk; and define f = 0 on oB(o, 1), f(o) = 1. The
function HDf reduees in this ease to the constant zero (why ?), whieh does
not satisfy the boundary eondition at o. But is there a true solution to the
Dirichlet problem? Call this h l ; then h l is harmonie in D, eontinuous in D
and equal to 1 and 0 respectively on oB(o, 1) and at o. Let cp be an arbitrary
rotation about the origin, and put hz(x) = hl(cpx) for XE D. Whether by
Definition 1 or 2 of §4.3, it is easy to see that h2 is harmonie in D and satisfies
the same boundary eondition as h l . Henee h l is identieal with h 2 (see Prop-
osition 5 below). In other words, we have shown that h l is rotationally
invariant. Therefore by Proposition 5 of §4.3, it must be of the form
Cl logllxll + C2 in D. The boundary eondition implies that Cl = 0, C2 = 1. Thus
h l is identically equal to one in D. Ii eannot eonverge to zero at the origin.
[The preeeding argument is due to Ruth Williams.] It is obvious that 0 is an
irregular boundary point of D, but one may quibble about its being isolated
so that the eontinuity of fon oD is a fietion. However, we shall soon diseuss
a more general kind of example by the methods of probability to show
where the trouble lies. Let us state first two basic results in the simplest case,
the first of which was already used in the above.
hex) = HDh(x), Vx E D.
4.4. Dirichlet Problem 167
Praaf. There exist regular open sets D n such that D n Ti D. In fact each D n
may be taken to be the union of cells of a grid on R d, seen to be regular by
the cone condition. We have h = HDnh in 15n by Proposition 5. Let first
XE D, then there exists m such that x E D n for n ~ m; hence
Letting n -+ 00, we have 'Dn i 'D r-a.s. by (VII) of §4.2. Hence the right
member above converges to H Dh(X) by bounded convergence. Next if XE oD
and x is regular for De, then HDh(x) = h(x) trivially. Finally if XE oD and x
is not regular for De, then PX{'D > O} = L We have for each t > 0:
because X(t) E D on {t < 'D}' so the second equation follows from what we
have already proved. As t t 0, the first term above converges to HDh(x), the
third term to EX{O < 'D; h(X(O))} = h(x) by bounded convergence. 0
h(x) = EX{h(X('DJ)}
(11)
= r{TilDo < TilBJ + EX{TilBn < TilDo ; h(X(TilBJ)}·
Since h is bounded in 15, it follows that the last term in (11) converges to
zero as n --+ 00. Consequently we obtain
This being true for x in D n for every n, we have proved that h == 1 in D. Thus
the Dirichlet problem (D,f) has no solution.
168 4. Brownian Motion
where y is any fixed point on aB. But aetually this is just the identity UI B =
PBUI B eontained in Theorem 3(b) of §3.4, followed by the observation that
U(y, B) is a eonstant for y E aB by rotational symmetry. Now the left member
of(14), apart from a eonstant, is equal to
r dy
JB Ilx _ Ylld 2
whieh eonverges to zero as Ilxll-- 00. Henee PBI(x) -- 0 as Ilxll-- 00, and so
Cl= O. Sinee PBI = 1 on aB, Cl = bd - 2 . Thus .
Ilxll ~ b. (15)
Consider the Diriehlet problem (D,f), where! is equal to one on S(o, a), zero
on S(o, b). The probabilistie solution given by PDc! signifies the probability
that the path hits S(o, a) before S(o, b). It is the unique solution by Proposition
5. Sinee it is a funetion of Ilxll alone it is of the form C-LIIXI12-d + Cl' The
eonstants Cl and Cl are easily determined by its valuefor Ilxll = a and Ilxll = b.
The result is reeorded below:
x ) _llxlll-d - bl - d
P {Ts(o,a) < TS(o,b)f -
a 2-d
- b2 - d ' a ~ Ilxll ~ b (17)
4.4. Dirichlet Problem 169
If h j + CXJ then TS(o.b) j 00 (why?), and we get (15) again with h replaeed
bya.
Exaetly the same method yields for d = 2;
Xf _ logllxll - log h
P l TS(o,a) < TS(o,b)} - log a - log h ' a ~ Ilxll ~ b. (18)
Ifwe let b j + 00, this time the limit ofthe above probability is 1, as it should
be on aeeount of reeurrenee. On the other hand, if we fix band let a 1 0 in
(18), the result is PX{ T{O} < TS(o,b)} = 0 (why?). Now let b j'X! to eoncludethat
P{o}l(x) = 0 for x#- o. Sinee Pt(x, {o}) = 0 for any x in E and t > 0, it follows
that P tP{o)l(x) = 0; henee P{o}l(x) = lim tlO P tP{o}l(x) = O. Thus {v} is apolar
set. This is the seeond proof ofa fundamental result reviewed in (IV) of§4.2.
. 1
j(x)=-
2ni
f f(z)
--dz
S(o,r) X - Z
(19)
d 2z. a 1 1 IIzl12 - II x l1 2
L d~2 -~ II
j= 1 - (;X j X - Z W- 2 Ilx - zlld-2 Ilx - zlld .
r r 2 -ll x l1 2 (J(dz)
JS(o,r) Ilx - zW .
Then this is also harmonie for x E B(o, r) (why?). Inspeetion shows that it is
a funetion of Ilxll alone, henee it is ofthe form c 1 11x11 2 - d + c 2 . Sinee its value
at x = 0 is equal to rZ-d(J(r) we must have Cl = 0 and C2 = r 2 - d (J(r) = m(l).
Reeall the value of (J(1) from (20) of §4.2 whieh depends on the dimension d
and will be denoted by (JAl) in this example. Now let f be any eontinuous
funetion on S(o, r) and put
. _ 1 r r2 - Il x 11 2 . _
(20)
I(x,f) - mAl) JS(o,rj Ilx _ zW j(z)(J(dL.).
170 4. Brownian Motion
This is the Poisson integral off for B(o, r). We have just shown that lex, 1) = 1
for XE B(o, r). We will now prove that l(x,f) = H B(X,f) for all fE q8B),
namely that the measure HB(x,') is given explieitly by lex, A) for all A E @,
Ac aB.
The argument below is general and will be so presented. Consider
1 r2 -llxl12
hex, z) = mAl) Ilx _ zlld ' Ilxll < r, Ilzll = r,
then h > 0 where it is defined, and SS(o,r) hex, z)(J(dz) = 1. Furthermore if
Ilyll = r, y =F z, then limx~z hex, y) = 0 boundedly for y outside any neighbor-
hood of z.1t follows easily that the probability measures lex, dz) = hex, z)(J(dz)
eonverge vaguely to the unit mass at z as x --+ z. Namely for any f E qaB),
l(x,f) eonverges to fez) as x --+ z. Sinee l(',f) is harmonie in B(o, r), l(',f) is
a solution to the Diriehlet problem (D,f). Henee it must eoineide with
HBf = HB(-,f) by Proposition 5.
Let us remark that H B(X, .) is the distribution of the "exit plaee" of the
Brownian motion path from the ball B(o, r), if it starts at x in the ball. This
is a simple-so unding problem in so-ealled "geometrie probability". An
expliet formula for the joint distribution of TB and X(TB) is given in Wendel
[1 ].
In view of the possible unsolvability of the original Diriehlet problem, we
will generalize it as follows. Given D and f as before, we say that h is a
solution to the generalized Dirichlet problem (D,f) iff h is harmonie in D and
eonverges to f at every point of aD whieh is regular for DC• Thus we have
proved above that this generalized problem always has a solution given by
HDf. If Dis bounded, this is the unique bounded solution. The proof will
be given in Proposition 11 of §4.5. We shall turn our attention to the general
question of non-uniqueness for an unbounded domain D.
Consider the funetion g defined by
Proof. There exists a sequenee of bounded regular open sets Dn such that
Dn @ Dn+ 1 for all n, and Un Dn = D. Such a sequenee exists by a previous
remark in the proof of Proposition 6. Suppose that h is bounded and har-
monie in D. Consider the Dirichlet problem (Dm h). It is plain that h is a
solution to the problem, henee we have by Proposition 5:
XE Dn• (23)
This implies the first part of the lemma below, in whieh we write T n for
to lighten the typography; also T o = O.
'D n ' Tfor 'D
nli~
... h(X(Tn))
~,
= {fe(X(T)) on {T < oo};
on {T = oo};
(24)
P{h(X(Tn))I.?(Tn- 1 )} = EX(Tn-tl{h(X(Tn))}
= HDnh(X(Tn- 1 )) = h(X(Tn_ 1 ))
where the last equation comes from (23). This proves the first assertion, and
eonsequently by martingale eonvergenee theorem the limit in (24) exists.
172 4. Brownian Motion
(25)
(26)
n~ CD
for every x ERd. Let x E D; then under p x and on the set {T = oo}, we have
T n > k and consequently h(X(Tn» ()k = h(X(Tn» for all sufficiently large
0
values of n. Hence in this case the limit in (26) coincides with that in (24).
Therefore the latter is equal to Z ()k which belongs to "§k. This being true
0
for each k 2 1, the limit in (24) is also equal to limk~oo Z ()k which belongs
0
to "§. Since "§ is trivial, the upper limit is a constant PX-a.s. for each XE Rd •
This means: for each x there is a number c(x) such that
°
°
By Exercise 4 below, g > in at most one component domain D o of D, and
g == in D - Do. Choose any Xo in Do and put for all x in D:
It is clear from the definition of Z that for any ball B = B(x, r) (§' D, q>(x) =
PBcq>(x) as in Theorem 1 of §4.3. Since q> is bounded, it is harmonie in D.
°
Since q> 2 and q>(xo) = 0, as shown above, it follows by Proposition 3 of
°
§4.3 (minimum principle) that q> == 0 in D o. As g == in D - D o, we have
q> == 0 in D. Thus we may replace c(x) by c(x o) in (27), proving the second
line in (24). For XE D, PX-a.s. on the set {T < oo}, we have T n i T, X(Tn ) E D
and X(Tn ) ~ X(T); since h(y) ~ fez) as y E D, y ~ Z E GD by hypothesis, we
4.4. Dirichlet Problem 173
have h(X(Tn )) - t j(X(T)). This proves the first line of (24). The lemma is
proved.
Now it follows by (23), (25) and bounded convergence that for each x E D:
Exercises
1. Without using Theorem 4, give a simple proof that a ball or a cube in Rd
is regular. Generalize to any solid with a "regular" surface so that there
is anormal pointing outward at each point. [Hint: (V) of §4.2 is sufficient
for most occasions.]
2. Let D be the domain obtained by deleting a radius from the ball B(o, 1)
in R 3 ; and let j = Ilxll on oD. The original Dirichlet problem (D,j) has
no solution.
3. In classical potential theory, a bounded domain D is said to be regular
iff the Dirichlet problem (D,j) is solvable for every continuous j on D.
Show that this definition is equivalent to the definition given here.
4. For any open set D, show that the function g defined in (21) is either
identically zero or strict1y positive in each connected component
(domain) of D. Moreover there is at most one component in which g > O.
Give examples to illustrate the various possibilities.
5. What is the analogue of (17) and (18) for R 1 ?
6. Derive the Poisson integral in R 2 from Cauchy's formula (19), or the
Taylor series of an analytic function (see Titchmarsh [1]).
7. Derive the Poisson integral corresponding to the exterior Dirichlet
problem for the ball B = B(o, r); namely find the explicit distribution of
X(ToB )1{ToB< 001 under px, where x E Be.
8. Let d 2 2; A be the hyperplane {x E Rdlxd = O}; and let D =
{xERdlxd>O}. Compute HD(x,'), namely the distribution of X(T A )
under px, for XE D. [Hint: we need the formula for Brownian motion in
R 1 :P{T{ol E dt} = Ixl/(2nt)3/2 exp[ -x 2/2t] dt; the rest is an easy com-
putation. Answer: HD(x,dy) = r(d/2)Xd1t-d/21Ix - yll-d l1 (dy) where 11 is
the area on A.]
174 4. Brownian Motion
9. Let D be open in Rd , d ;;::: 2. Then D is polar if and only if there does not
C
Since u is bounded below in B(x, r) we may suppose u ;;::: 0 there, and u(x) <x.
We then obtain (2) from (1) as we did in (6) of §4.3. It now follows that if u
is superharmonic in a domain D, then it is locally integrable there. To see this
we may again suppose u ;;::: 0 beeause u is bounded below on each eompact
4.5. Superharmonie Function and Supermartingale 175
subset of D. Let A be the set of points x in D such that u has a finite integral
over some ball with center at x. It is trivial that A is open. Observe that if
x E A, then there is X n arbitrarily near x for which u(x n) < 00, and con-
sequently U has a finite integral over any B(x m r) (§ D. Therefore this is also
true for any B(x, r) (§ D by an argument of inclusion. Now let X n E A, x n --+
X E D. Then for sufficiently large n and small r we have x E B(x n , r) (§ D. Thus
XE A and so A is closed in D. Since D is a domain and A is not empty we
have proved A = D as claimed.
We now follow an unpublished method of Doob's to analyse super-
harmonie functions by means of supermartingales. Suppose first that D is
bounded and u is superharmonic in an open set containing 15. Let p denote
the distance function in R d • For a fixed 8 > 0 and each x in D we put
thus S(x) is the sphere with center x and radius r(x) varying with x as pre-
scribed. Define a sequence of space-dependent hitting times {Tm n ~ O} as
folIows: T o = 0 and for n ~ 1:
(4)
provided that u(x) < 00. The integrability of u(X(Tn)) is proved by induction
»,
on nunder this proviso. Thus {u(X(Tn :F(Tn), n ~ 1} is a supermartingale
under PX • Since D is bounded all the terms of the supermartingale are
bounded below by a fixed constant. Hence it may be treated like a positive
supermartingale; in particular, it may be extended to the index set
"1 :$; n :$; 00" in a trivial way.
Let us first prove the crucial fact that for each x E D, we have
pX{lim T = 'D} = 1.
n~oo
n (5)
For it is clear that PX_a.s., Tni and T n < 'D < 00 because D is bounded. Let
T = lim n i Tn- Then X(T (x,) = limn~oo X(Tn) by continuity and so X(T,J E
00
15. Suppose if possible that X(T 00) E D. Then for all sufficiently large values
ofn, wehavep(X(Tn_1),oD) > 2- 1p(X(T),8D) > Oandp(X(Tn_d,X(Tn») <
»
4- 1 p(X(T),8D)/\8. But by (3), p(X(Tn-d,X(Tn = 2- 1 p(X(Tn_d,8D)/\r..
These inequalities are incompatible. Hence X(T 00) E 8D and consequently
T 00 = 'D·
176 4. Brownian Motion
(7)
is a supermartingale under Px provided u(x) < oc. Here X(T J = X(T D),
ff(T xJ = .~(TD) by (5). It follows that
by Fatou's lemma since a1l the terms are bounded be10w. Now !im n
X(T nAN ) = X(T N ) whether N is finite or infinite, and by the 10wer semi-
continuity of u in 15 (because u is superharmonic in an open set containing 15)
we conclude that
(9)
PX-a.s. on {t < T V }' On the set {t 2 T v }, N(ß) = + oc for a1l E > 0 and X(T N(I.») =
X(T D ) by (5). Hence (10) is also true. Using (10) in (9) and the lower semi-
continuity of u again, we obtain
(11 )
is a supermartingale under px Jor each XE D Jor which u(x) < 00. When u is
harmonie in an open set containing 15, then (13) is a martingale under Px Jor
each x E D.
By this definition Gis made an absorbing state, and T n the lifetime of X; see
(13) and (14) of §1.2. Define a transition semigroup (Qt, t ~ 0) as follows; for
any bounded Borel function <p on Da and x E Da:
(15)
In particular, Qo<P = <po We call X the "Brownian motion killed outside D"
or "kilIed at the boundary GD". It is easy to check that (Qt) is the transition
semigroup of X, but the next theorem says much more. Recall {~} is the
family of augmented a-fields associated with the unrestricted Brownian
motion {Xt, t ~ O}.
Proof. First of all, almost all sampie paths are continuous in [0, 00), in the
topology of D", To see this we need only check the continuity at T D when
Tn < 00. This is true because as ti Tn < 00, X(t) -+ G = X(T n). Since con-
tinuity clearly implies quasi left continuity, it remains only to check the strong
Markov property. Let J be a continuous function on Da, then it is bounded
because Da is compacL Let T be optional with respect to {~}, and A E ffT •
178 4. Brownian Motion
since Qt!(ö) = f(ö) for t ~ O. By (15), the first term on the right side above is
equal to
which is equal to
for every continuous f on Da. Therefore, X has the strong Markov property
(cf. Theorem 1 of §2.3), indeed with respect to a larger u-field than that
genera ted by X itself. 0
Proof· Let each D n be bounded open and D n ii D. Write 'n = 'D"' , = 'D
below. Theorem 1 above is applicab1e to each D.. Hence if x E D and u(x) <
00 we have
(17)
(18)
4.5. Superharmonie Function and Supermartingale 179
The two relations (18) and (19), together with the banality u(o) = Qtu(o),
show that u is excessive with respect to (Qt).
Conversely, if u is excessive with respect to (Qt), then u ;;::: 0; and the
inequality (1) is a very special case of u ;;::: P AU in Theorem 4 of §3.4. Observe
that QAc(X,') = P AC(X, .) if x E D and A c D, where QA is defined for X as
PA is for X. To show that u is superharmonic in D it remains to show that u
is lower semi-continuous at each x in D. Let B(x,2r) @ D. Then for all Y E
B(x, r) = B, we have pY {-r ::s; t} ::s; pY {Tr ::s; t} and the latter does not depend
on Y (Tr is defined in (10) of §4.2). It follows that
Suppose first that u is bounded, then Ptu is continuous by the strong FeIler
property; while the last term in (21) is less than s if 0< t< <5(s), for all Y E B.
We have then
u(y) ;;::: Qtu(y) ;;::: Ptu(y) - s;
lim u(y) ;;::: lim Ptu(y) - s = Ptu(x) - s ;;::: Qtu(x) - E.
y-x y-x
Proposition 4. Let z E oD; then z is regular Jor DC if and only if there exists a
barrier at z.
180 4. Brownian Motion
(24)
Ad
u(x, y) = Ilx _ Ylld 2'
Then u(x, y) denotes the potential density at y for the Brownian motion
starting from x. For a fixed y, u(·, y) is harmonie in Rd - {y} by Proposition
5 of §4.3. Since it equals + 00 at y it is trivial that the inequality (1) holds for
x = y; since it is continuous in the extended sense, u(·, y) is superharmonie
in R d • Now let D be an open set and for a fixed y E D put
where Pnc operates on the function x -+ u(x, y). It follows from the preceding
discussion that gn(·, y) is superharmonie in D, harmonie in D - { y}, and
vanishes in (15)" and at points of oD which are regular for DC ; if Dis regular
°
then it vanishes in D C• Since u(·, y) is superharmonie and ~ in Rd , it is
excessive with respect to (Pt) by an application of Theorem 3 with D = Rd •
Of course this fact can also be verified direct1y without the intervention
°
of superharmonicity; cf. the Example in §3.6. Now gn ~ by Theorem 4 of
§3.4. Hence it is excessive with respect to (Qt) by another application of
Theorem 3. For any Borel measurable function f such that Ulfl(x) < 00 we
have
The first term on the right side of (27) does not exceed
thus w( (0) = w. Then Jor any x E D, we have PX-a.s.: t ~ w(t) is right con-
tinuous in [0,(0) and has leJt limits in (0,00]; and {w(t), ~(t), 0 < t :S X)}
is a supermartingale under p x . In ease u(x) < 00, the parameter value t = 0
may be added to the supermartingale. IJ u is harmonie and bounded in D, then
{w(t), ~(t), 0 :S t:S oo} is a continuous martingale under p x Jor any x E D.
4.5. Superharmonie Function and Supermartingale 183
These relations imply the assertions of the theorem, the details of which
are left to the reader. D
Proof· Let 0::;; t l < ... < t n ::;; e, and jj E tff+ for 1 ::;;j::;; n. To show that
Em{Ö }-I
jj(X(t))} = Em{Ö }-I
jj(Xc(t))} (32)
1
r . Jdxnfn(x 0 p(tj+
n) .
}=n-I
1 - t j; x j+b x)jj(x) dX j
= r·Jfl(xl)dx 1 n
n-I
J=I
P(tj+l-tj;Xj,Xj+l)jj+I(Xj+l)dxj+l'
with respect to m. Hence for Ps(x,' )-a.e. y, the result holds under pY. For
°
s :::::: let
As = {w u(X(-, w)) is continuous in es, oo]}.
1
Theorem 9 is stated for u defined in the whole space. It will now be ex-
tended as folIows. Let D be an open set, and u be positive and superharmonie
in D. The function t -> u(X(t,w)) is defined in the set I(w) = {t > X(t,w) E 01
D}, which is a.s. an open set in (0,00). We shall say that u(X) is (right, left)
continuous wherever defined iff t -> u(X(t, w)) is so on I(w) for a.e. w. Let
rand r' be rational numbers, r < r'. It follows from Theorem 7 that for
each x
pX {X(t) E D for all tE er, r'], and t -> u(X(t)) is right continuous
in (r,r')} = PX{X(t) E D for all tE [r,r']}. (34)
Theorem 10. For the Brownian motion process, a semi-polar set is polar.
186 4. Brownian Motion
(35)
t > O. We have
Proof. We know from §4.4 that HDf is a solution. Let h be another solution
and put u = h - H Df. Then u is harmonie in D and converges to zero
on a* D, the set of "regular boundary points". By the Corollary to Theorem
10, DC\(D')' is a polar set. Observe that it contains aD - c7* D. For each
x in D, the following assertions are true PX_a.s. As t ii 'D' X(t) -+ X('D) E
iJ*D because aD - a*D being a subset of DC\(Dcy cannot be hit. Hence
u(X(t)) as t ii 'D converges to zero, and consequently the random variable
w defined in (28) is equal to zero. Applying the last assertion of Theorem 7
to u, we conclude that u(x) = P{w} = O. In other words, h == HDf in D.
D
4.5. Superharmonie Function and Supermartingale 187
on oB and ju there, then HBln S u on oB; for the neeessity part use
Proposition 5 of §4.4.J
6. Extend the neeessity part of Exercise 5 to any open set B with eompaet
Be D. [Hint: if B is not regular, Proposition 5 of §4.4 is no longer
suffieient, but Proposition 6 iso Alternatively, we may apply Exereise
4 to u - h.J
7. If u is superharmonie in R d and t > 0, then Ptu is eontinuous [Hint: split
the exponential and use p t / 2 u(0) < oo.J
8. Let D be bounded, u be harmonie in D and eontinuous in 15. Then the
proeess in (13) is a martingale. Without using the martingale verify
direetly that u(x) = E u(X(t J\, D))} for 0 S t S 00. Note that D need
X
{
not be regular.
9. Let D be an open set and define for t ~O:
11. Show that the first funetion in (17) of §4.3 is subharmonie in R 2 ; the
seeond superharmonie in R d, d ~ 3.
12. Let u be superharmonie in D. For (j > 0 put
uö(x) = fRd <Pö(llx - yll)u(y)dy
In the next two exercises, proofs of some of the main results above are
sketched which do not use the general Hunt theory as we did.
17. Let u be a positive superharmonic function in R d . We may suppose it
bounded by first truncating it. Prove that alm ost surely:
(a) t -> u(X(t)) is lower semi-continuous;
(b) if Q is the set of positive rationals, then
lim u(X(t)) = u(X(O));
Q3tlO
The Dirichlet problem deals with the solution of Laplace's equation Llq> = 0
in D with the boundary condition q> = f on aD. The inhomügeneous case
of this equation is called Poisson's equation: Llq> = 9 in D where 9 is a given
function. Formally this equation can be solved if there is an operator .1- 1
inverse to .1, so that .1(Ll- 1 g) = g. We have indeed known such an inverse
under a guise, and now proceed to unveil it.
Suppose that 9 is bounded continuüus and Uigl < 00. Then we have
· -1 (Pt - /)Ug
I1m
t!O t
=
I'1m --1
tJ,O t
It Psgds
0
= -g (1)
. 1
I1m
.si = - (Pt - /) (2)
tJ,O t
·
I1m (PI - l)j'(x)
~-- LI.
= - f' (x), (3)
110 t 2
Proof. Let x be fixed and f belong to q2) in the ball B(x, r). By Taylor's
theorem with an integral remainder term, we have for Ilyll < r:
d d
where J; and J;j denote the first and seeond partial derivatives of f. Note
that they are bounded in B(x, r). Now write
put y = zJt, and split the range ofthe integral into two parts: Ilyll = Ilzll.Jt <
rand Ilyll = Ilzll Jt 2: r respeetive1y. Observe that p(t; 0, z Jt) d(z Jt) =
p(l; 0, z) dz. The seeond part of the integral is bounded by
-1 ~ _p(l;o,z) { .jtLzJ(x)+t
_ d Ld ZiZj
SI t;)x+sz-y't)(l-s)ds
r- }
dz
t Ilzll <rl, I i= 1 i.j= 1 0 .
d
=
Jllzll<rl,1
L
r _ p(1;o,z) i.j=l ZiZj r
Jo
l
f)x + szJt)(l - s)dsdz, (6)
r
Jllzil <rlvl
_p(l :o,z)zjdz = O.
4.6. The Role of the Laplacian 191
where ({Jii t , z) = {f
0
/;j(X + szJt)(l - s)ds, if Jt[[z[[ < r;
(7)
0, if Jt[[z[[;:o: r.
Ug(x) = 2n
1 f [[xg(y)_ y[[ dy
is bounded continuous in x. In fact Ug E C(I). The same argument there
applies to any d ;:0: 3. Thus in this case Ug is smoother than g. In order that
Ug belong to C(2), we need a condition which is stronger than C(O) (continuity)
but weaker than C(1). There are several forms of such a condition but we
content ourselves with the following. The functions 9 is said to satisfy a
Hölder condition ifffor each compact K there exists rx > and M < 00 (rx and
M may depend on K) such that
°
[g(x) - g(y)[ $; M[[x - y[[~ (8)
more general assumptions, see Port and Stone [1], pp. 115 -118. An easy
ease of it is given in Exereise 1 below.
We are now ready to show that the two operators - L1/2 and U aet as
inverses to eaeh other on eertain classes of funetions.
Theorem 3 (for Rd , d ;:0: 3). Ij" 9 E IHl n then - (L1j2)( U g) = g. Ir 9 E e~2), thell
U( - (L1/2)g) = g.
. lt P, (P
Ptg - 9 = hm -hg-- -g) ds = lt Ps (.hm--Phg -
- g)
hjO 0 h 0 hjO
h - cis
(9)
If K is the eompaet support of g, then IPtgl :s; IlgllPtl K ..... 0 as t ..... .J"j beeause
K is transient in R d , ci ;:0: 3. Sinee lL1gl is bounded and has eompaet support,
the last member of(9) eonverges to U((L1/2)g) as t ..... JJ. We have thus proved
that - 9 = U( (L1j2)g). 0
Sinee e~2) c IHl n Theorem 3 implies that - L1j2 and U are inverses to eaeh
other on e~2).
Next, reealling (26) of §4.5 but writing 9 for the f there, we know that if
Uigl < JJ, then
(10)
where
(11)
and D is any bounded open set in Rd , d ?: 3. This is then true if 9 E IHle. More-
over ifwe apply the Laplaeian to all the terms in (10), we obtain by Theorem 3
To see the uniqueness, suppose <p is sueh a solution. Then <p - GDg is
harmonie in D, eontinuous in l5, and vanishes on aD. Henee it vanishes in
l5 by the maximum (minimum) prineiple for harmonie funetions.
We proeeed to obtain similar results for R 2 . Sinee the Brownian motion
in R 2 is reeurrent, the potential kernel U is of no use. We use a substitute
eulled from Port and Stone [1]. Let Ilxoll = 1 and put for all x E R 2 :
u*(x) = fo CD
pi(x) dt.
We have
_ -
- 1
2n
1""(e
0
-e s -lds
-llxWs-s)
1 In f.rlllxw -1 -r 1 1
= 2n Jo r S ds e dr = ~ log 1I~' (13)
Now put
pi(x, y) = pi(Y - x), u*(x, y) = u*(y - x),
Pi(x,dy) = pi(x,y)dy;
The kernel U* is the logarithmic potential. Note that not only u*(x) is positive
or negative aeeording as Ilxll :s; 1 or z 1; but the same is true for pi(x) for
eaeh t z O. If g is oounded with eompaet support, then U*g is bounded
eontinuous in R 2 . In this respeet the analogy of U* with U in R d for d z 3
is perfeet ; Theorem 2 also goes over as follows.
U *g - PIU *g = SI P,gds - hm
0 . JA + A I P,gds = S'oP,gds (15)
A-'> ·f
Proof. The first assertion follows from Theorems 1 and 2' and (15), in exactly
the same way as the corresponding ca se in Theorem 3. To prove the second
assertion we can use an analogue of (9), see Exercise 3 below. But we will
vary the technique by first proving the assertion when ;1y E IHle. U nder this
stronger assumption we may apply the first assertion to Llg to obtain
LI
- 2" U*(Llg) = Lly.
where Cf>b is defined in Exercise 6 of §4.3. Then gij E 1[:( x). Ir in addition g beloflgs
to I[:b2 ) in an open set D, then gb -> g and Llgij -> Llg both boundedly in D as
<5 t O.
g + ~ U*(Llg) = O. D
Next we consider Poisson's equation in R 2 . Since U* is not a true potential,
the analogue of (10) is in doubt. We make adetour as follows. Let y E ([~2),
then (9) is true. We now rewrite (9) in terms of the process
Ifwe put
(17)
then a general argument (Exercise 10) shows that {Mt, g;, t ~ O} is a mar-
tingale under p x for each x. Let D be a bounded open set, then {MtAtD ,
g;, t ~ O} is a martingale by Doob's stopping theorem. Since g and Ag are
bounded and EX{'D} < 00 for XE D, it follows by dominated convergence
that if XE D:
(18)
This is the exact analogue of (10). It follows that Theorem 4 holds intact in
R 2•
We have made the assumption of compact support in several places above
such as Theorems 2 and 2', mainly to lighten the exposition. This assumption
can be replaced by an integrability condition, but the following lemma
(valid in Rd for d ~ 1) shows that in certain situations there is no loss of
generality in assuming compact support.
Proof· Let D o be a bounded open set such that D o ::J jj and p(aD, aD o) =
c5 o. Let 0 < c5 < c5 0 and put
t/I(x) = f
JDo qJ;;(x - y)dy
where qJ;; is as in Lemma 6. Then t/I E 1[(<10). If p(x, Do) > c5, then t/I(x) = O. If
XE D, then t/I(x) = 1. It is clear that the function g = f· t/I has the required
properties. 0
We do not need Theorem 2 or 2' here, only the easier Exercise 1 to conclude
that Ug or U*g belongs to IC(<XJ); hence GDg E 1C(00) by (10) or (19). We have
by Theorem 4,
L1cp = g = 1 in D.
(20)
Since the first term on the right side of (20) is harmonie in D, we have
a neat verification.
Recall the notation T h = inf{t > O:IX(t) - X(0)12 h}. For each XE (a,b),
let [x - h, x + h] c (a,b). Then r{T h <, < oo} = 1. We have by the strong
Markov property:
Without loss of generality we may suppose that g :::::: 0. Then it follows from
(21) that
q;(x}:::::: Hq;(x + h) + q;(x - h)}.
Since q; is bounded this implies that q; is continuous, indeed concave (see e.g.
Courant [1], Vol. 2, p. 326). Next as h 1 0, it is easy to see by dominated
convergence that the right member of (22) converges to
It turns out that in high er dimensions there are also certain substitutes
for the Laplacian similar to the generalized second derivative above. Some
of the analytical difficulties disappear if these substitutes are used instead;
on the other hand under more stringent conditions they can be shown to be
the Laplacian after all, as in the case above.
EXAMPLE 2. Let 0< a < b < 00. Compute the expected occupation times of
B(a, a) before the first exit time from B(a, b).
We will treat the problem in R 2 which is the most interesting case since
the expected occupation time of B(a, a) is infinite owing to the recurrence
of the Brownian motion. In the notation of (19), the problem is to compute
GD! when D = B(a, b), ! = 1 in B(a, a) and ! = °
in B(a, b) - B(a, a). This
! is only "piecewise continuous" but we can apply (19) to two domains
separately. We seek q; such that
Clearly, q> depends on Ilxll only, nt:nce we can solve the equations above by
the polar form ofthe Laplacian given in (19) of§4.3. Let Ilxll = r; we obtain
r2
q>(r) = -2 + CI + C2 log r, 0< r < a;
a < r < b.
a2 - r2 b
q>(r) = 2 + a2 log~, o ~ r ~ a;
b
= a 2 10g-, a< r < b.
r
Exercises
1. In R d , d ~ 3, if 9 E lC~k), k ~ 1, then Vg E lC(k). In R 2 , if 9 E lC~k), k ~ 1,
then V*g E lC(k).
2. Prove Lemma 6. [Hint: use Green's formula for B(x, b o), b o > b to show
Llüx) = JLlf(x - y)q>ö(y)dy.]
3. Prove the analogue of (9) for R 2 :
Ptg-g= S~P:(~9)dS
and deduce the analogue of(10) from it.
4. ls the following dual of (15) true?
[Hint: gD(.X, y) = 2(x - a)(b - y)/(b - a) if a < x::; y::; h; 2(b - x)(y - a)1
(b - a) if a < y::; x < b.]
8. Solve the problem in Example 2 for R 3 .
9. Let 11 be a a-finite measure in R d , d ;;:: 2. Suppose U11 is harmonie in an
open set D. Prove that Il(D) = O. [Rint: let f E 1[(2) in B es D and f = 0
outside B where B is abalI; h(UIl)L1fdm = hd U(L1.f)dll; use Green's
formula.]
10. The martingale in (17) is a case of a useful general proposition. Let
{MI' t ;;:: O} be associated with the Markov process {XI':Fr, t;;:: O} as
folIows: (i) Mo = 0; (ii) Mt E:Fr; (iii) M s + t = M s + Mt' es where {OS'
s;;:: O} is the shift; (iv) for each x, EX{X t} = O. Then {Mw~t, t;;:: O} is a
martingale under each P X. Examples of Mt are g(X t) - g(X 0) and
Shcp(Xs)ds, where gEM, cpEbg; and their sumo Condition (iii) is the
additivity in additive functionals.
In this section we discuss the boundary value problem for the Schrödinger
equation. This includes Dirichlet's problem in §4.4 as a particular case. The
probabilistic method is based on the following functional of the Brownian
motion process. Let q E bIS, and put for t ;;:: 0:
(1)
(3)
(4)
For 0 < s ::; 2r, we have T s < r Dunder pxo, since B(x o,2r) (§ D. Hence
by the strong Markov property
The next crucial step is the stochastic independence of T s and X(TJ under
any px, proved in (XI) of §4.2. U sing this and (3) we obtain
(7)
For all XE B(x o, r), we have B(x, r) c B(x o, 2r) @ D. Hence we obtain
similarly:
u(x) = EX{eq(Ts)u(X(Ts))}
::; e{eQTsu(X(Ts))} ::; 2E X{u(X(Ts))}' (8)
4.7. The Feynman-Kac Functional and the Schrödinger Equation 201
This leads to the first inequality below, and then it follows from (7) that
u(x) s -v(r)
2 lB(x,r) u(y)dy s -v(r)2 lB(xQ,2r) u(y)dy
4v(2r) ( )
s~uxo. (9)
We have thus proved (4), in particular that u(x) < 00. As a consequence, the
set of x in D for which u(x) < 00 is open in D. To show that it is also closed
relative to D, let X n -+ xx. E D where u(x < 00 for all n. Then for a sufficiently
p)
large value of n, we have Ilx oo - xnll < tJ 1\ (p(X., oD)j2). Hence the inequality
(4) is applicable with x and Xo replaced by X oo and X n , yielding u(x oo ) < 00.
Since D is connected, the first assertion of the theorem is proved.
Now let Do be a subdomain with Do C D. Let 0 < r < tJ 1\ (p(D o, oD)j2).
Using the connectedness of D o and the compactness of Do, we can prove the
existence of an integer N with the following property. For any two points
x and x' in Do, there exist n points Xl' ... ,X n in Do with 2 S n S N + 1, such
that x = Xl' x' = Xn , Ilxj+ 1 - xjll < rand p(xj+ 1, oD) > 2r for 1 sj s n - 1.
A detailed proof ofthis assertion is not quite trivial, nor easily found in books,
hence is left as Exercise 3 with sketch of solution. Applying (4) successively
to x j and Xj +l ' 1 sj s n - 1 (sN), we obtain Theorem 1 with A = 2(d+2)N.
If K is any compact subset of D, then there exists D o as described above such
that K c Do. Therefore the result is a fortiori true also for K. D
Theorem 3. Let u be as in (2), but suppose in addition that fis bounded on oD.
If u =1= 00 in D, then u is bounded in D.
Put for XE E:
u 1 (X) = EX{elrD)f(X(T D)); TE < T D},
u 2(x) = EX{elrD)f(X(TD));T E = T D }.
202 4. Brownian Motion
By the strong Markov property, since In = IE + In' (JrE on the set (I E < I [)}:
(11 )
We shall denote the u(x) in (2) more specifically by u(D, qJ; x).
In analogy with the notation used in 94.6, let :?ß(D) denote the dass of
functions defined in D which are bounded and Lebesgue measurable; IHl(D)
the dass of functions defined in D which are bounded and satisfy (8) of 94.6
for each compact K cD. Thus hC(1)(D) c IHl(D) c hC(O)(D). Let us state the
following analytic lemma.
For d = 1 this is elementary. For d ~ 2, the results follow from (10) and
(19) of §4.6, via Theorems 2 and 2' there, with the observation that GDy =
G D(1Dg) by (11) of §4.6. Note however that we need the versions of these
theorems localized to D, which are implicit in their proofs referred to there.
For a curious deduction see Exercise 4 below.
(12)
4.7. The Feynman-Kac Functional and the Schrödinger Equation 203
( 13)
[Note that in the integral above the values of q and u on aD are irrelevant.]
This shows that the function l{t<tDJlq(X t)u(X t )1 of(t,w) is dominated in the
integration with respect to the produet measure m j x p x over [0, (0) x Q,
where m j denotes the Lebesgue measure on [0, (0). Therefore, we ean use
Fubini's theorem to transform the following integral as shown below:
The third member above is obtained from the seeond by first reversing the
order of the two integrations, then applying the Markov property at eaeh
tunder EX, noting that CD = t + CD 0 8t on {t < CD}' We ean now perform
the trivial integration with respect to t to obtain
(14)
(16)
As x -+ z, this converges to zero by (8) of §4.4, and (16) above. Next we have
for x E B(z, r):
Let us remark that contrary to the Laplace case, the uniqueness of solution
in the Schrödinger case is in general false. The simplest example is given in
R 1 by the equation u" + u = 0 in D = (0, n). The particular solution u(x) =
sin x vanishes on aD! In general, unicity depends on the size of the domain
D as weil as the function q. Such questions are related to the eigenvalue
4.7. The Feynman-Kac Functional and the Schrödinger Equation 205
problem associated with the Schrödinger operator. Here we see that the
quantity u(D, q, 1, x) serves as a gauge in the sense that its finiteness for some
x in Densures the unique solvability of all continuous boundary value
problems.
Exercises
1. Prove that the function u in (2) is Borel measurable.
2. If fE g + in (2), then either u := 0 in D or u > 0 in D. [Here D need not
be bounded.]
3. (a) Let D be a domain. Then there exist domains D n strictly contained in
D and increasing to D. [Hint: let Un be the union of all balls at dis-
tance > I/n from GD. Fix an X o in D and let D n be the connected
component of U n which contains xo. Show that Un D n is both open
and closed relative to D.]
(b) Let Do be a bounded domain strictly contained in D. Let 0 < r <
tp(D o, GD), and 15 0 c Uf= 1 B(Xi' r12) where all Xi E 15 0 . Define a
connection "~" on the set of centers S = {Xi' 1 ::;; i::;; N} as follows:
Xi ~ X j if Ilx i - Xjll < r. Use the connectedness of D oto show that for
any two elements X a and Xb of S, there exist distinct elements Xi'
1 ::;; j ::;; I, such that Xi, = X a , Xi, = Xb' and Xij ~ X ij +' for 1 ::;; j ::;; 1- 1.
In the language of graph theory, the set S with the connection ~ forms
a connected graph. [This formulation is due to M. Steele.]
(c) Show that the number N whose existence is asserted at the end of the
proof ofTheorem 1 may be taken to be the number N in (b) plus one.
In the following problems D is a bounded domain in Rd , d ;:::: 1; q E M'.]
4. (a) Let D 1 be a subdomain @D. If ~ c D - D 1 , then GDg is harmonie in
D1 •
(b) Let D 1 @ D2 @ D. If gE IHl(D) then there exists g1 such that g1 E
IHlc(R d ) and gl = 9 in D], 9 1 = () in Rd - D 2 . [Hint: multiply 9 by a
function in CXJ) as in Lemma 7 of §4.6.]
(c) Prove Proposition 4 by using Theorems 2 and 2' of §4.6. [This may
be putting the horse behind the cart as alluded to in the text, but it is
a good exercise!]
5. If u(D, q, 1; .) =1= 00 in D, then it is bounded away from zero in D. More-
over there exists a constant C > 0 such that
u(D,q, 1; x);:::: Cu(E,q, 1; x)
[Hint: for some t o > 0 we have two constants Cl > 0 and C z > 0 such
that Cl::; EX{elr D); 0< 'D::; I} ::; C 2 for all XE D; now estimate
EX{eq('D); n < 'D::; n + I} and add.]
7. Suppose D is regular and EX{ellqlltD} < 00 for some XE D. Then for any
fE qaD), u(D, q,f,.) is the unique solution of (12) with boundary value
f. [Hint: let ep be a solution which vanishes on aD. Show that ep = GD(qep).
Prove by induction on n that
NOTES ON CHAPTER 4
As an introduction to the c1assical viewpoint the old book by Kellogg [1] is still valuable,
particularly for its discussion of the physical background. A simpler version may be
found in Wermer [1]. Ahlfors [1] contains an elementary discussion of harmonie
functions and the Dirichlet problem in R 2 , and the connections with analytic functions.
Brelot [1] contains many modern developments as weil as an elegant (French style)
exposition of the Newtonian theory.
The proof of Theorem 8 by means of Lemma 9 may be new. The slow pace adopted
here serves as an example of the caution needed in certain arguments. This is probably
one of the reasons why even probabilists often bypass such proofs.
§4.5. Another method of treating superharmonie functions is through approximation
with smooth ones, based on results such as Theorem 12; see the books by Rao and
Port-Stone. This approach leads to their deeper analysis as Schwartz distributions. We
choose Doob's method to give further credance to the viability of paths. This method
is longer but ties several items together. The connections between (sub)harmonic
functions and (sub)martingales were first explored in Doob [2], mainly for the loga-
rithmic potential. In regard to Theorems 2 and 3, a detailed study of Brownian motion
killed outside a domain requires the use of Green's function, namely the density of the
kernel Q, defined in (15), due to Hunt [1]. Here we regard the case as a worthy illustration
ofthe general methodology (8 and all).
Doob proved Theorem 9 in [2] using H. Cartan's results on Newtionian capacity.
A non-probabilistic proof ofthe Corollary to Theorem 10 can be found in Wermer [1].
The general proposition that "semipolar implies polar" is Hunt's Hypothesis (H) and
is one of the deepest results in potential theory. Several equivalent propositions are
discussed in Blumenthai and Getoor [1]. A proof in a more general case than the
Brownian motion will be given in §5.2.
§4.6. The role of the infinitesimal generator is being played down here. For the
one-dimensional case it is quite useful, see e.g., Ho [1] for some applications. In higher
dimensions the domain of the operator is hard to describe and its full use is neither
necessary nor sufficient for most purposes. It may be said that the substitution of integral
operators (semigroup, re solvent, balayage) for differential ones constitutes an essential
advance of the modern theory of Markov processes. Gauss and Koebe made the first
fundamental step in identifying a harmonie function by its averaging property (Theorem
2 in §4.3). This is indeed a lucky event for probability theory.
§4.7. This section is added as an afterthought to show that "there is still sap from
the old tree". For a more complete discussion see Chung and Rao [3] where D is not
assumed to be bounded but m(D) < 00. The one-dimensional case is treated in Chung
and Varadhan [1]. The functional eq(t) was introduced by Feynman with a purely
imaginary q in his "path integrals"; by Kac [1] with a nonpositive q. Hs application
to the Schrödinger equation is discussed in Dynkin [1] with q ~ 0, Khas'minskii [1]
with q :?: O. The general case of a bounded q requires a new approach partly due to the
lack of a maximum principle.
Let us alert the reader to the necessity of a meticulous verification of domination,
such as given in (13), in the sort of ca1culations in Theorem 5. Serious mistakes have
resulted from negligence on this score. For instance, it is not sufficient in this case to
verify that u(x) < 00, as one might be misled to think after a preliminary (illicit) integra-
tion with respect t.
Comparison of the methods used here with the c1assical approach in elliptic partial
differential equations should prove instructive. For instance, it can be shown that the
finiteness of u(D, q, 1; .) in D is equivalent to the existence of a strictly positive solution
belonging to (:<2)(D) n (:<0)(1». This is also equivalent to the proposition that all eigen-
values A of the Schrödinger operator, written in the form (,1/2 + q)<p = A<p, are strictly
negative; see a forthcoming paper by Chung and Li [1]. Further results are on the way.
Chapter 5
Potential Developments
°
Of course u ;;:: and u E 6' x 0"; but no further condition will be imposed
untillater. Let A E 6 and
5.1. Quitting Time and Equilibrium Measure 209
,( )_{sUP{t>OIXt(W)EA}, if W E Ll A ,
IA W -
0, if W E Q - Ll A ;
TA is the hitting time, YA is the quitting time (or last exit time) of A (denoted
by LA in (21) of §3.6). Their dual relationship is obvious from the above.
Recall that A is transient if and only if
(3)
from which it follows that YA E :Fx' The next result is trivial but crucial:
From here on we fix the transient set A and omit it from the notation until
further notice. Let .f E blC (bounded continuous), I:: > 0, and consider the
following approximation of[(X(y- )):
(7)
Put
1
.1, (x) = -[;
'1',' 1 < "{ <
pxfO - I::}', (8)
(9)
210 5. Potential Developments
Using this we apply the simple Markov property at time t to see that the
integrand in (7) is equal to
Hence the quantity in (7) is equal to U(f 'Ij;,)(x), Let us verify that this is
uniformly bounded with respcct to G, Since f is bounded it is sufficient to
check the following:
= r
J{y>Oj
(~G Jrl r c)'
y
1 dt) PX(dw) ~P X (}, > O} ~ 1.
r
Jli>O}
1 rl'
~ Jü-c)' f(Xf)dtPX(dw), (ll)
Since fE bC, and X has a left limit at}' ( <:1:;), as I; 1 0 the limit ofthe quantity
in (11) is equal to
r
Jh'>O}'
f'(X(,'-))PX(dw)
I
= EXI"
tJ.
> 0; f(X("-
I
))1,
j
Mc(dy) = Ij;,(y)~(dy),
SL(x,dy)
( ) cp(y) =
.
hm
Scp(y)M,(dy) (13)
u x, y ,tO
because u(x, y)u(x, y) - 1 = 1 for y E E - {x}, and the point set {x} may be
ignored since M, is diffuse. Since for each x, L(x, .) is a finite measure and
u(x, y)-l is bounded on each compact, L(x, dy)u(x, y)-l is aRadon measure.
It is weIl known that two Radon measures are identical if they agree on all
cp in Ce (Exercise 1). Hence the relation (13) implies that there exists a single
Radon measure p on g such that
Let us pause to marvel at the fact that the integral above does not depend
on x. This suggests an ergodic phenomenon which we shall discuss in §5.2.
Since M, is aRadon measure for each G > 0, it follows also from (13) that
pis the unique vague limit of M, as G ! 0; but this observation is not needed
here. We are now going to turn (14) around:
When B = {x} in (14) the left member is equal to zero by condition (ii).
Hence p is diffuse. Next putting B = {y} with y i= x in (14) we obtain
L(x,B) = l B\{x)
u(x,y)
L(x,dy)
(
u x, Y
) +L(x,Bn{x})
= S B\{x)
u(x, y)p(dy) + L(x, B n {x})
It follows firstly that ~({x}) > 0 and secondly U(x, {x}) = w. Together with
the hypothesis that x is holding the latter condition implies that {x} is a
recurrent set under px. This is a familiar resuIt in the theory ofMarkov chains
(where the state space is a countable set). Moreover, another basic resuIt in
the latter theory asserts that it is almost impossible for thc path to go from
a recurrent set to a transient set. 80th resuIts can be adapted to the case in
question, the details of which are left in Exercise 8 below (strong Markov
property is needed). In condusion, we have proved that if x is holding then
PX{TA < w} = 0, which implies (18). We summarize the results abovc with
an important addition as folIows.
Theorem 1. Let X be a Hunt process with the potential kerne! in (1) satisfving
conditions (i) and (ii). Then for each transient set A, there exists aRadon
measure fJ.A such that for anJ' x E E a/1(/ BE ß:
(20)
If almost all paths of the process are continuous, then fJ.A Iws support in r:A.
In general if A is open then fJ.A has support in A.
5.1. Quitting Time and Equilibrium Measure 213
It is essential to see why for a compact A the argument above does not
show that ~A has support in A. For it is possible on Ll A that X(YA) E A while
X(y A -)i A; namely the path may jump from anywhere to cA and then
quit A forever.
Corollary. Wehave
(21)
where sup 0 = 0. Since for a Hunt process left limits exist in (0, 00), and
t -+ X I_ is left continuous, we have X(y A -) E A, regardless if the sup in (22)
is attained or not. This is the key to the next result.
Proof. Let us beware that "left" notions are not necessarily the same as
"right" ones! The definition oftransience is based on XI = X t +, not X t -, and
it is not obvious that the transient set A will remain "transient" for the left
limits of the process. The latter property means pX{yA' < oo} = 1 for every
x. That this is indeed true is seen as follows. We have by Theorem 9 of §3.3,
TA ;:::: TA a.s. for any Borel set A. On the other hand, the left analogue of (4)
is true: {YA' > t} = {TA' BI < oo}. It follows that
0
214 5. Potential Developmcnts
Since left notions are in general rather different from their right analogues,
Theorem 2 would require re-thinking of several basic notions such as "left-
regular" and "left-polar" in the developments to foIlow. Fortunately under
Hypotheses (B) and (L), we know by Theorem 3 of §3.8 that TA = T~ a.s.
It foIlows that YA = YA: a.s. as weIl (Exercise 6) and so under these hypotheses
Theorem 2 contains Theorem 1 as a particular case. We state this as a
coroIlary. We shaIl denote the support of a measure P by ~.
Since (24) is true for aIl fE M+, it foIlows that for each x there exists N x
with ~(N xl = 0 such that if y 1= N x:
Now the measure ~ charges every nonempty open set (why?). Hence for an
arbitrary y we have Yn 1= N x' Yn -> y, so that u(z, Yn) -> u(z, y) for every z by
condition (i). Therefore by Fatou:
This proves the first assertion of the proposition. The second follows from
Proposition 3 of §3.2 and the remark following it. 0
For each y, x --+ g(x, y) is excessive by Proposition 5 of §3.2. Observe that the
function g may not satisfy the conditions (i) and (ii), in particular g(x, x) may
not be infinite. The following results is proved in Chung and Rao [1] but
the proof is too difficult to be given here. A simpler proof would be very
interesting indeed.
Proof. We begin with the following basic relation. For each x E AC u Ar:
(27)
This has been proved before in this book (where?), though perhaps not
exactly in this form. The reader should ponder why the conditions on x and
on the transience of An are needed, as weil as the quasi left continuity of the
216 5. Potential Developments
n (Oe'
l fAn > O} = 1f yA > 0 1J . (28)
We will omit below the diche "alm ost surely" when it is obvious. Clearly
YA" 1 and YA" ~ }'k Let f3 = lim n ['An' Then on {YA > O}, we have X(y An - ) =
X(y"';: n - ) E An as shown above on account of Hypothesis (B). It follows by
right continuity that if {3 < YA" for all 11 then
This is trivial if Xis continuous at }'k The general argument below is due to
John B. Walsh and is somewhat delicate. If X has a jump at lA' this jump
time must be one of a countable collection of optional times {cx n }, by Exercise
5 of§3.1. This means that for alm ost every w, }'A(W) = cxn(w) where n depends
on w. We can apply the strong Markov property at !Y. n for all n to "cover
YA" whenever X is discontinuous there. (We cannot apply the property at
}'A!) Two cases will be considered for each CXn' written as CX below.
Case 1. X(!Y.) ~ A. Applying (27) with x = X(cx), we see that since TA (Ja =
CIJon {cx = YA}, there exists N(w) < CIJ such that TA" ., 0, =CIJ, hence !Y. = ;'A"
for n ~ N(w). Thus (30) is trivially true because i'A" = j'A for all sufficiently
large values of 11.
Case 2. X(!Y.) E A. Then on {cx = YA} we must have X(cx) E A\A r because the
path does not hit A at any time strictly after cx. Since ()( is a jump time and
A\A r is semipolar by Theorem 6 of §3.5, this possibility is ruled out under
Hypothesis (B), by Theorem 1 (iv) of §3.8. This ends the proof of (30).
Now let x E AC u Ar and{ E blC. Then we have by (28), {I'A" > O} 1 [}'A > O}
PX-a.s.; hence by (30) and bounded convergence:
(32)
5.1. Quitting Time and Equilibrium Measure 217
Exercises
1. Let (E,0") be as in §1.1. A measure 11 on 0" is called aRadon measure iff
Il(K) < 00 for each compact sub set K on E. Prove that if 11 and v are
two Radon measures such that Sf dll = Sf dv for all f E Ce> then 11 == v.
[Hint: use Lemma 1 of §1.1 to show that Il(D) = v(D) < 00 for each
relatively compact open D; then use Lemma 2 of §1.1 to show that
Il(B n D) = v(B n D) for all BE 0". Apply this to a sequence of Dk such
that D k i E. This exercise is given here because it does not seem easy to
locate it in textbooks.]
In Exercises 2 and 5, we assume the conditions of Theorem 1.
2. Let A be transient. If IlA(E) = 0, then Ais polar. Conversely if P A1(X) = 0
for some x, then IlA(E) = O.
3. Assume U(x, K) < 00 for each x and compact K. Let f be any excessive
function. If fex) > 0 for some x then fex) > 0 for all x. [Hint: use
Proposition 10 of §3.2.]
4. Assume each compact set is transient. Let f be an excessive function. If
fex) < 00 for some x then the set {x Ifex) = oo} is a polar set. [Hint:
use Theorem 7 of §3.4 and U > 0.]
5. Under the same conditions as Exercise 4 prove that each singleton is
polar. [Hint: let Dn be open relatively compact, Dn 11 {x o}; P v) = U Iln
with Illn c Vn- Show that (a subsequence of) {Iln} converges vaguely to
A6 xo ' and AU(X, x o) s limn U Iln(x) S 1 so that ), = O. For x I tt Vb limn
Ulln(X I ) = 0 because U(x I ,') is bounded continuous in VI' Now use
Exercise 2.]
6. Prove that for a Hunt process, YA ?: YA a.s. for any A E 0". lf TA = T~
a.s. then YA = YA a.s.
7. Prove that a polar set is left-polar; a thin set is left-thin.
8. The following is true for any Hunt process (even more gene rally). Let x
be a holding but not absorbing point. Define the sojourn time S = inf {t >
0IX t =1= x}, and the re-entry time R = inf{t > SIXt = x}. Show that
218 5. Potential Developmcnts
PX{S> t} = e- A1 for some }.: 0< A < Cf..;. Let PX{R < Cf.]} = p. Show
that U(x, {x}) = ),-1(1 - p)-I; hence U(x,{x}) = 00 if and only if {x}
is "recurrent under p Let A E $; prove that if P TA < oo} > 0 then
X
". X {
In this section we continue the study of a Hunt process with the potential
kernel given in (1) of §5.1, under further conditions. Recall that Hypothesis
(L) is in force with ~ as the reference measure. The principal new condition
is that of symmetry:
V(x, y): u(x, y) = u(y,x). (S)
There are various connections between the conditions. Recall that (TI) and
(U I) imply (T 2 ) by Theorem 2 of§3.7; (R) and (U tl imply (U 2) by Proposition
3 of §5.1 ; (R), (T 2) and either (U I) or (U 2) imply Hypothesis (B) by Theorem 4
of §5.1 ; (R) and (S) imply (U 1) trivially.
Readers who do not wish to keep track of the various conditions may
assume (R), (S) and (T 1)' and be assured that all the results below hold true
with the possible exception ofthose based on the energy principle. However,
it is one of the fascinating features of potential theory that the basic results
are interwoven in the manner to be illustrated below. This has led to the
development ofaxiometic potential theory by Brelot's schoo!.
The Equilibrium Principle of potential theory may be stated as folIows.
For each compact K, there exists a finite measure J1K with support in K such
5.2. Some Principles of Potential Theory 219
that
(E)
The principle holds under (R), (T 2) and either (U 1) or (U 2), because then
Hypothesis (B) holds and (E) holds by the Corollary to Theorem 2 of §5.1.
We are interested in the mutual relationship between several major
principles such as (E). Thus we may assurne the validity of (E) itself in some
of the results below. Let us first establish a basic relation known as Hunt's
"switching formula" in the duality theory. The proof is due to K. M. Rao.
From here on in this section the letters K, D, and Bare reserved respectively
for a compact, relatively compact open, and (nearly) Borel set.
(1)
hence for each x, y --> PBu(x, y) is excessive by simple analysis (cf., Exercise 8
of §4.l). On the other hand for each y, x --> PBu(x, y) is excessive by (U 2) and
a general property of excessive function (Theorem 4 of §3.4).
Let K c D, we have
Since U(X, y) 2:: PDu(x, y) by Theorem 4 of §3.4, and K is arbitrary for a fixed
D, it follows from (3) that
for each x and ~-a.e. y in D. Both members of (4) are excessive in y for each
fixed x, hence (4) holds for each x in E and y E D, by the Corollary to Pro-
position 4 of §3.5. [This property of U is of general importance.]
Nextwehave
220 5. Potential Developments
~f PJ)(y,dw)u(x, w)
= PDu(y,x).
(6)
Integrating this with respect to ~(dx) over an arbitrary compact set C, we have
y ~ K\K r • (8)
It is essential to notice that the last member above is finite by (Tl) because
it does not exceed U(y, Cl. Since C is arbitrary we deduce from (8) that
(9)
far each y rf= K\K r , and ~-a.e. x. Both members of (9) are excessive in each
variable when the other variable is fixed and ~(K\Kr) = 0 because K\,K' is
semipolar. Therefore (9) holds in fact for all x and y: since it is symmetrie in
x and y, we conclude that
(10)
Now given Band x, there exist compacts K n c B, such that T K" 1 TB,
PX-a.s. by Theorem 8(b) of §3.3. Since for each y, u(X" y) is right continuous
in tunder (U 2) by Theorem 6 of §3.4, it follows by Fatou and (10) that
This principle must be intuitively obvious to a physicist, since it says that the
potential induced by acharge is greatest where the charge lies. Yet its proof
seems to depend on some sort of duality assumption on the distribution of
the charge. This tends to show that physical processes carry an inherent
duality suggested by the reversal of time.
Theorem 2. The maximum principle (M) halds under (S), (Tl) and (U 2)'
because ltt c K and for each y E K, Pßu(y,x) = u(y,x) trivially since y E Br.
Ün the other hand, P ß(x, .) has support in B by Theorem 2 of §3.4, since B
is finely closed. Hence we have
Putting (13) and (14) together we conclude that Ufl ~ M since c is arbitrary.
o
The argument leading to (13) is significant and recorded below.
(15)
222 5. Potential Developments
(16)
Remark. For a thin set this says that if it cannot be hit immediately from
some point, then it cannot be hit at all from any point!
Proo.f. By Theorem 8(b) of §3.3 and Theorem 6 of §3.5, it is sufficient to prove
that each compact thin set is polar. Let K be such a set, and define for each
n;:C: 1:
T, = T,-t + T K BT , . , :
C
where ß ranges over all ordinals preceding IX. The following assertions
are almost surely true for eachlX. Since A is polar, we have X(T,) E C on
5.2. Some Principles of Potential Theory 223
{Ta< oo}. For a limit ordinal a this is a consequence of quasi left continuity.
In view of (17), we have for each a:
n {T
00
On the other hand, since K is thin, strong Markov property implies that
Ta < Ta + 1 on {T, < CXl}. It follows as in the proof of Theorem 6 of §3.4
that there exists a first countable ordinal a* (not depending on w) for which
Ta' = 00. This a* must be a limit ordinal by (18). Therefore, on the set of (j)
where K is ever hit, it is hit at a sequence of tim es which are all finite and
increase to infinity. This contradicts the transience of K. Hence K cannot
be hit at all, so it is polar. 0
In particular for any open D, the set ofpoints on cD which are not regular
for D is a polar set, because it is a subset of DC\(DT. This is the form of the
C
(19)
to be the capa city of A. Under (T 2), C(K) is defined for each compact K
and is a finite number since flK is aRadon measure. For any two O"-finite
measures )'1 and ,1.2' we put
(20)
(21 )
Vx E K: Uv(x):::; 1. (22)
224 5. Potential Developments
Then we have
v(K) ~ C(K). (23)
Proo{ Let K c D; under (T 2 ) both K and D are transient and so flD as weil
as flK exist. Under (S) we have as a case of (20):
(24)
Note that if the ineguality in (22) holds for all x E E, then the intervention
of (M) is not needed in the preceding proof. This is the case for Corollary 1
below.
Proof. Let An = {xl Vv(x) ~ n}. If \' charges a polar set it must charge a
polar subset of An for some n, and so also a compact polar subset of An,
because V is necessarily inner regular (Exercise 1). Let VI be the restriction
ofv to such a set K, and 1'2 = 1'1/11. Then VVl ~ Ion K, hence by the theorem
v2 (K) ~ C(K). But C(K) = 0 by Exercise 2 of §5.1. Thus 1'2 == 0 and \' does
not charge K after all. 0
provided that all four terms on the right are finite. This amounts to the
condition that (). + v, ), + v) < XJ. Let us remark that if )'1 - },2 = )'3 - )'4
and both (}'I - A2 , )'1 - )'2) and <i
3 - },4, },3 - A4 ) are defined then they
are egual. If the energies of AI - ;'2 and of ;'3 - },4 are defined then so is
the energy of (AI - A2 ) - (A 3 - J.4) = (/'1 + A4 ) - (Al + ;'3)' In particular
we can define the energy of a signed measure using its Hahn-Jordan decom-
position, but there is no need for that. We are ready to announce another
principle.
5.2. Some Principles of Potential Theory 225
Theorem 5. Suppose that the Hunt proeess has a transition density funetion
Pt with respeet to ~ whieh is symmetrie, namely for t > 0:
for all (x, y); where Pt ~ 0, Pt E ß x ß. Then the energy principle holds.
Proof. We have
and
I Jl.(dx}Ptf(x) = 0, if t rt N. (28)
Since X is right continuous, lim tLo Prf(x) = f(x) for each x. Taking a
sequence of t rt N, t 1 0 in (28) we obtain
Vf E bC: If dA - If dv = 0,
where both integrals are finite. Hence A == v (see Exercise 1 of §5.1). D
226 5. Potential Developments
From here on all the results so far proved will be used wherever appro-
priate. Let K be a fixed compact set and <P(K) denote the dass of finite
measures v such that ~ c K and <v, v) < CIJ. lt follows from Corollary 2
to Theorem 4 that such a v does not charge any polar set, since JUl'dl' < x.
For the equilibrium measure IlK' we have
Hence !lK E <P(K). But then IlK does not charge the set A below:
(30)
<
It follows from (31) and (33) that the energy of v - IlK is equal to v, v) -
C(K). We have therefore proved the following result under the energy
principle.
Theorem 6. Let C(K) > O. For any v E <P(K) with I'(K) = C(K), IVe haue
(35)
where the L~)'s are the iterates of L K = L~). The measure n is an invariant
(stationary) measure for L, namely nL K = n. Now under (S) there is an
obvious invariant measure obtained by norma!izing the equilibrium
measure:
1 1
1l'K = IlK(E) J1.K = C(K) J1.K· (37)
P A l(x)=-
2n
1 i IlA(dy)
'11 x - Y 11'
,A
(38)
This representation is very fruitful. First, it yields the following formula für
the capacity of A:
C(A) = 2n lim IlxIlPAl(x). (39)
X--I',y
Finally, for any two Borel sets A and B we have the inequality:
(41)
follows that
6(8B n A) 1
,uB(A) = 2nr 2 = -2 6(8B n A);
4nr r
in particular
C(B) = C(8B) = ,uB(8B) = 2nr. (44)
For another application of (39), let Albe a Borel set with its dosure
contained in the interior of A. Then it is obvious that P Al(x) = P A-AII(x) for
x E AC. Hence C(A) = C(A - AI) by (39). lt is interesting to see how this
result can be sharpened and extended to the dass of processes considered
1ll §5.1, with continuous paths. We have by (19)
for any x. Choose x E AC. Then the continuity of paths implies that under
p x we have rA = rA-Al almost surely, because the path from outside A can-
not hit A without hitting A - Ab and cannot quit A - Al (forever) without
quitting A. Therefore, LA(x,·) = LA-AJX,·) which says that the equilibrium
measures for A and A - Al are the same. The fact that ,uli = ,uDA may be
regarded as a limiting case.
In the language of electrostatics, if the interior of asolid conductor is
partially hollowed out, distribution of the induced charge on the outside
surface remains unchanged when it is grounded (in equilibrium). An analytic
proof of this physical observation may be based on the energy minimizing
characterization of the equilibrium measure (Theorem 6). To some of us
such a proof may seem more devious than the preceding one, but this is a
matter of subjective judgment and previous conditioning. Be it as it may,
here is an appropriate place to end these notes (for the moment), leaving the
reader with thoughts on the empricial origin of mathematical theory, the
grand old tradition of analysis, and the relatively new departure founded on
the theory of probability.
Exercises
1. Let (E, 0') be as specified in §3.1. Show that any 6-finite measure v on 0'
is inner regular; namely: v(B) = SUPKCB v(K). [Hint: for a finite measure
this follows from Exercise 12 of §2.1 of Course.]
230 5. Potential Developments
2. Show that under the hypothesis of Theorem 5 the duality relation (18)
of §4.1 holds with Ei" = U". Assurne also (P). Prove the uniqueness of the
equilibrium measure J1K in the following form: if Uv = UJ1K on K and
~ c Kr, then V = J1K' [Hint: use the Corollary to Theorem 2; and note
that U:K C Kr by (P).]
3. Under (S) and (E), prove that if J1 and V are a-finite measures such that
U J1 :::;; U v, then J1(E) :::;; v(E). [Hint: integrate with respect to J1K']
4. Let cP ± (K) denote the class of } = )"1 - )"2 where )'1 E CP( K), A2 E CP( K).
Under the energy principle prove that J1K is the unique member of
cp±(K) which minimizes G below:
G(A) = (),)") - 2A(K).
This is the quadratic actually used by Gauss.
5. Suppose that <A,A) ~ 0 for every } E cp±(K) as defined in Exercise 4.
This is called the Positivity Principle and forms part of the energy
principle. Show that under (S) we have the Cauchy-Schwarz inequality:
<A, V)2 :::;; O",A)<V, v) for any A and v in CP±(K).
6. Generalize (34) as folIows. Let K be compact: v a finite measure on K: U
defined on K x K is symmetric and ~ O. Define L(x, dy) = u(x, y)v(dy)
and assurne that for every x E K we have L(x, K) :::;; I. Prove that for
any measure } on K with A(K) :::;; 1 we have
Hence if the positivity principle in Exercise 5 holds, then <AL, )L) :::;;
(), A). Thus the energy of a subprobability measure decreases after each
transformation by L. This is due to J. R. Baxter. [Hint: put cp(y) =
JA(dx)u(x,y). Then S(Lcp)2dv:::;; SL(cp2)dv:::;; Scp 2 dv; hence <AL,AL) 0=
J(cpLcp)dv:::;; Scp 2 dv = <cp,cpL).]
The next few problems are taken from Frostman [1] and meant to give
some idea of the classical approach to the equilibrium problem. They are
analytical propositions without reference to a process, where (E,0") =
(R d, gßd), ~ = m (Lebesgue measure).
7. Let u ~ 0 and (x, y) -+ u(x, y) be lower semi-continuous. Prove that
there exists a probability measure v such that
<v, v) = inf<;"A)
Je
v == Je under the energy principle. [Hint: consider the energy of(1 - c)v + cA
for I; 1 O.J
9. Suppose (A,A) < 00. Then Uv = <v, v), ),-a.e. on K. [Hint: let <v, v) =
c> b > a > 0 and suppose that Uv < a on A with Je(A) > O. It follows
from lower semi-continuity that there exists a ball B with v(B) > 0 such
that Uv > bon B. Now transfer the mass v(B) and re-distribute it over A
proportionately with respect to Je; namely put for all S: v'(S) = v(S\B) +
<
V(B)A(S n A)A(A) - I. Compute v, Vi) to get a contradiction with Exer-
cise 8. Hence Uv : : :-: C, A-a.e. on K.J
Note. It follows from (31) and (37) that UV K = <v K , vK ) on K n Kr, hence
except for a polar set by (P). If }. has finite energy then }. does not charge
any polar set, hence the result in Exercise 9 is true by the methods used in
§§5.1-5.2. The notion of "capacity" was motivated by such considerations.
10. Let u(x, y) = Ilx - yll-' where rx > 0 in R 2 , and rx > 1 in R 3 . Prove that
if U J1 is continuous, then (M) holds. [Hint: the Laplacian of U J1 at a
point where maximum is attained must be <0 by calculus. This is
called the "elementary maximum principle" by Frostman and is an
illustration of the humble origin of a noble result.J
11. Let u be as in Exercise 10 but 1 < rx < 3 in R 3 . Let J1 have the compact
support K and suppose that U J1 considered as a function on K only is
continuous there. Then it is continuous everywhere. This is known as the
Continuity Principle and is an important tool in the classical theory.
[Hint: UJ1 is continuous in KC. It is sufficient to show that rpo(x) =
h(x,o) u(x, Y)J1(dy) 1 0 uniformly in x as b 1 o. For x E K this is true by
the hypo thesis and Dini's theorem. For x 1= K let y be a point in K
nearest x. Then IIY - zll :s 211x - zll for all z E K; if Ilx - yll :S b then
rpo(x) :S 2'rp20(y)·J
12. Establish the celebrated M. Riesz convolution formula below:
E"{ e- aTB } -I = EO{ exp[ Jbx l(t)J} = ;~) J'B exp[ J2ax IJO"(dy).
Both results above for R d , d ;::: 3' are in Getoor [3J].
15. For any bounded Borel set A, and any Borel set S, prove that
Apart from the use of ")' A the result (valid für R d, d ;::: 3) is implicit m
Spitzer [l].J
NOTES ON CHAPTER 5
The title ofthis ehapter is a double entendre. It is not the intention ofthis book to treat
potential theory exeept as a eoneomitant of the underlying processes. Nevcrthclcss we
shall proeeed far enough to show several faeets of this old theory in a new light.
§5.1 The eontent of this seetion is based on Chung [1]. This was an attempt to
utilize the inherent duality of a Markov proeess as evideneed by its hitting and quitting
times. Formula (20) or its better known eorollary (21) is a partieular ease of the repre-
sentation of an exeessive funetion by the potential of a measure plus a generalized
harmonie funetion. Such a representation is proved in Chung and Rao [I J for the dass
of potentials whieh satisfies the eonditions of Theorem 1. In classieal potential theory
this is known asF. Riesz's deeomposition of a subharmonie funetion; see Brelot [I].
We stop short of this fundamental result as it would take us too far into a well-
entrenehed field.
Let us point out that Robin's problem of determining the equilihrium charge
distribution is solved by the formula (14). One might employ a kind of Monte Carlo
method to simulate the last exit probability there to obtain empirieal results.
§5.2. The various prineiples of Newtonian potential theory are diseussed in Rao [I].
The extension of these to M. Riesz potentials was a major advanee made by Frostman
[I]. Their mutual relations form the base of an axiomatic potential theory founded by
Brelot. Wermer [IJ gives abrief aeeount of the physieal eoneepts of capacity and
energy leading to the equilibrium potential.
In prineiple, it should be possible to treat questions of duality by the method of
time reversal. Although a general theory of reversing has existed for some time, it seems
still too diffieult for applications. A prime example is the polarity prineiple (Theorem 3)
which is not prima facie such a problem. Yet all known probabilistic proofs use some
kind of reversing (cf. Theorem 10 of §4.5). It would be extremely interesting to obtain
this and perhaps also the maximum principle (Theorem 2) by a manifest reversing
argument. Such an approach is also suggested by the concept of reversibility of physical
processes from which potential theory sprang.
Bibliography
L. V. Ahlfors
[I] Complex Analysis, Second edition, McGraw Hill Co., New York, 1966.
J. Azema
[I] Theorie generale des processus et retournement du temps, Ann. Eeole Norm Sup. Sero 4,
6 (1973), 459-519.
R. M. Blumenthai and R. K. Getoor
[1] Markov Processes and Potential Theory, Academic Press, New York, 1968.
M. Brelot
[I] Elements de la Theorie Classique du Potentiei, Fourth Edition, Centre de Documentation
Universitaire, Paris, 1969.
K. L. Chung
[I] A Course in Prohahility Theory, Second Edition, Academic Press, New York, 1974.
[2] Markov Chains with Stationary Transition Prohabilities, Second Edition, Springer-
Verlag 1967.
[3] Boundary Theory for Markov Chains, Princeton University Press, Princeton NJ, 1970.
[4] A simple proof of Doob's theorem, Sem. de Prob. V, (1969/70), p. 76. Lecture Notes in
Mathematies No. 191, Springer-Verlag, Berlin Heidelberg New York.
[5] On the fundamental hypotheses of Hunt Processes, Instituto di alta matematica, Sympo-
sia Mathematica IX (1972), 43-52.
[6] Probabilistic approach to the equilibrium problem in potential theory, Ann. Inst. Fourier
23 (1973), 313-322.
[7] Remarks on equilibrium potential and energy, Ann. Inst. Fourier 25 (1975), 131-138.
[8] Excursions in Brownian motion, Arkivför Mat. 14 (1976), 155-177.
K. L. Chung and J. L. Doob
[1] Fields, optionality and measurability, Amer. J. Math. 87 (1965), 397 -424.
K. L. Chung and J. Glover
[1] Left continuous moderate Markov processes, Z. Wahrseheinlichkeitslheorie 49 (1979),
237-248.
K. L. Chung and P. Li
[I] Comparison of probability and eigenvalue methods for the Schrödinger equation,
Advances in Math. (to appear).
K. L. Chung and K. M. Rao
[I] A new setting for potential theory (Part I), Ann. Inst. Fourier 30 (1980),167- 198.
[2] Equilibrium and energy, Probab. Math. Stat. 1 (1981).
[3] Feynman-Kac functional and the Schrödinger equation, Seminar in Stochastic Processes
1 (1981), 1-29.
K. L. Chung and S. R. S. Varadhan
[I] Kac functional and Schrödinger equation, Studia Math. 68 (1979), 249-260.
K. L. Chung and J. B. Walsh
[I] To reverse a Markov process, Acta Math. 123 (1969),225-251.
R. Courant
[I] Differential and Integral Caleulus, Interscience 1964.
C. Dellacherie
[I] Ensembles aleatories I, 11, Sem. de Prob. III (1967/68), pp. 97-136, Leeture Notes in
Mathematics No. 88, Springer-Verlag.
[2] Capacites et Processus Stochastiques, Springer-Verlag Berlin Heidelberg New Y ork 1972.
234 Bibliography
[3J Mesurabilite des debuts et theoreme de section: le lot a la portee de toutes les bourscs.
Sem, de Prob. XV (1979/80), pp. 351--370 Leeture Notes in Mathematics No .• Springer-
Verlag, Berlin Heidelberg New York.
C. Dellacherie and P. A. Meyer
[I J Probahilill;,\' et Potentiei, Hermann, Paris 1975.
[2J Probahilith et Potentiei: Tlu;orie des Martingales, Hermann. Paris 1980.
J. L. Doob
[IJ Stoehastie Processes. Wiley and Sons. New York, 1953.
[2J Semimartingales and subharmonic functions, Trans. Amer. Math. SOl'. 77 (1954). 86 121.
[3J A probability approach to the heat equation, Trans. Amer. Math. Soc. 80 (1955).216- 280.
E. B. Dynkin
[lJ Markov Processes, Springer-Verlag, Berlin Heidelberg New York 1965.
E. B. Dynkin and A. A. Yushekevich
[I J Theorems and Problems in Markov Processes, (in Russian), Nauk, Moscow, 1967.
O. Frostman
[I J Potentiel d'equilibre et capacites des ensembles avec quelques applications a la theorie
des fonctions, Medd. Lunds Univ. Math. Sem. 3 (1935), 1--118.
R. K. Getoor
[I J Markov processes : Ray processes and right processes, Leeture Notes in Marhematics
No. 440, Springer-Verlag, Berlin Heidelberg New York 1975.
[2J Transience and recurrence 01' Markov Processes, Sem. de Prob. XIV (1978/79). pp, 397-
409, Leeture Notes in Mathematics No. 784, Springer-Verlag, Berlin Heidelberg New
York.
[3J The Brownian eseape process, Ar/l1. Probability 7 (1979),864--867.
I. I. Gihman and A, V. Skorohod
[I J The Theorv 0(' Stochastie Processes 1l (translated from the Russian), Springer-Verlag
Berlin Heidelberg New York 1975.
L. L. Helms
[I J Introduction to Potential Theory, Wiley and Sons, New York, 1969.
G. A. Hunt
[I J Some theorems concerning Brownian motion, Trans. Amer. Math. SOl'. 81 (1956).
294-319.
[2J Markoff processes and potentials, I, Illinois J. Math, I (1957),44 93.
[3J Martingales et Processus de Markov, Dunod, Paris, 1966.
K.Ito
[I J Leetures on Stochastic Proeesses, Tate Institute, Bombay 1961.
M. Kac
[I J On some connections between probability theory and differential and integral equations.
Proc. Second Berkeley Symposium on Math. Stal. and Prohahilitr, 189 215. University
of California Press, Berkeley 1951.
S. Kakutani
[I J Two-dimensional Brownian motion and harmonie functions, Proc. Imp. Acad. Tokl'o
22 (1944), 706 714.
J. L. Kelley
[I J General Topology, D. Van Nostrand Co., New York 1955.
O. D. Kellogg
[I J Foundations o('Porential Theory, Springer-Verlag Berlin, 1929.
R. Z. Khas'minskii
[I J On positive solutions ofthe equation du + Vu = O. Theory ofProhahility alld its Applica-
tions (translated from the Russian) 4 (1959),309-318.
P. Levy
[I J Theorie de ['Addition des Variahles Aleatoires, Second edition, Gauthicr-Villars. Paris,
1954.
P. A. Meyer
[1 J Probahility and Potentials, Blaisdell Publishing Co. 1966.
[2J Processus de Markov, Leeture Notes in Mathematics No. 26, Springer-Verlag, Berlin
Heidelberg New York 1967.
[3J Processus de Markov: 1a frontiere de Martin. Leeture Notes in Mathematics No. 77,
Springer- Verlag, Berlin Heidelberg New Y ork 1968.
Bibliography 235
thin 110
Radon measure 212 totally inaccessiblc 80
reeurrent process 122 transient process 86, 126
set 121 set 121
reference measure 112 transition function 6
relleetion principle 153 transition probability dcnsity 11
regular point 97
domain 165
regularization 82 uniform motion 10
remote field 144 uniqueness fnr potentials of
resolvent 46 measures 140
equation 83 universal Illeasurability 63
reversal 01' time, reverse process 183
(F.) Riesz decomposition 118
Riesz potentials (see under M. Riesz version 28
potentials) volumc of ball 150
right continuity
of augmcnted fields 61
of excessive funetion on paths 101 zero-or-one law 64, 144
of fields 12 zero potential 109
Grundlehren der mathematischen Wissenschaften
ASeries of Comprehensive Studies in Mathematics
A Selection
153. Federer: Geometrie Measure Theory
154. Singer: Bases in Banaeh Spaees I
155. Müller: Foundations of the Mathematieal Theory of Eleetromagnetie Waves
156. van der Waerden: Mathematieal Statisties
157. Prohorov/Rozanov: Probability Theory. Basie Coneepts. Limit Theorems. Random Processes
158. Constantineseu/Cornea: Potential Theory on Harmonie Spaees
159. Köthe: Topologieal Vector Spaces I
160. Agrest/Maksimov: Theory of Ineomplete Cylindrieal Funetions and their Applieations
161. Bhatia/Szegö: Stability of Dynamieal Systems
162. Nevanlinna: Analytie Funetions
163. Stoer/Witzgall: Convexity and Optimization in Finite Dimensions I
164. Sario/Nakai: Classifieation Theory of Riemann Surfaees
165. MitrinovielVasie: Analytie Inequalities
166. GrothendieekiDieudonnt\: Elements de Geometrie Aigebrique I
167. Chandrasekharan: Arithmetieal Funetions
168. Palamodov: Linear Differential Operators with Constant Coeffieients
169. Rademacher: Topies in Analytic Number Theory
170. Lions: Optimal Control of Systems Governed by Partial Differential Equations
171. Singer: Best Approximation in Normed Linear Spaees by Elements of Linear Subspaces
172. Bühlmann: Mathematieal Methods in Risk Theory
173. Maeda/Maeda: Theory of Symmetrie Lattices
174. Stiefel/Scheifele: Linear and Regular Celestial Meehanie. Perturbed Two-body
Motion - Numerieal Methods - Canonieal Theory
175. Larsen: An Introduetion to the Theory of Multipliers
176. GrauertiRemmert: Analytische Stellenalgebren
177. Flügge: Praetieal Quantum Meehanies I
178. Flügge: Praetieal Quantum Mechanies 11
179. Giraud: Cohomologie non abelienne
180. Landkof: Foundations of Modern Potential Theory
181. Lions/Magenes: Non-Homogeneous Boundary Value Problems and Applieations I
182. Lions/Magenes: Non-Homogeneous Boundary Value Problems and Applieations II
183. Lions/Magenes: Non-Homogeneous Boundary Value Problems and Applications III
184. Rosenblatt: Markov Proeesses. Structure and Asymptotic Behavior
185. Rubinowiez: Sommerfeldsehe Polynommethode
186. Handbook for Automatie Computation. Vol. 2. Wilkinson/Reinsch: Linear Algebra
187. SiegeliMoser: Leetures on Celestial Meehanies
188. Warner: Harmonie Analysis on Semi-Simple Lie Groups I
189. Warner: Harmonie Analysis on Semi-Simple Lie Groups II
190. Faith: Algebra: Rings. Modules. and Categories I
191. Faith: Algebra 11. Ring Theory
192. Malleev: Aigebraie Systems
193. P6lyalSzegö: Problems and Theorems in Analysis I
194. Igusa: Theta Funetions
195. Berberian: Baer*-Rings
196. AthreyalNey: Branehing Proeesses
197. Benz: Vorlesungen über Geometrie der Aigebren
198. Gaal: Linear Analysis and Representation Theory
199. Nitsehe: Vorlesungen über Minimalflächen
200. Dold: Leetures on Aigebraie Topology
201. Beck: Continuous Flows in the Plane
202. Schrnetterer: Introduction to Mathematieal Statisties
203. Schoeneberg: Elliptic Modular Funetions
204. Popov: Hyperstability of Control Systems
205. Nikollskii: Approximation of Funetions of Several Variables and Imbedding Theorems
206. Andre: Homologie des Aigebres Commutatives
207. Donoghue: Monotone Matrix Funetions and Analytie Continuation
208. Laeey: The Isometrie Theory of Classieal Banaeh Spaees
209. Ringel: Map Color Theorem
2 !O. Gihman/Skorohod: The Theory of Stoehastie Proeesses I
211. Comfort/Negrepontis: The Theory of Ultrafilters
212. Switzer: Aigebraie Topology- Homotopy and Homology
213. Shafarevieh: Basic Aigebraie Geometry
214. van der Waerden: Group Theory and Quantum Meehanies
215. Schaefer: Banaeh Lattices and Positive Operators
216. Polya/Szegö: Problems and Theorems in Analysis II
217. Stenstriim: Rings of Quotients
218. Gihman/Skorohod: The Theory of Stoehastie Proeesses II
219. Duvant/Lions: Inequalities in Meehanies and Physies
220. Kirillov: Elements of the Theory of Representations
221. Murnford: Aigebraic Geornetry I: Complex Projeetive Varieties
222. Lang: Introduetion to Modular Forms
223. Bergh/Liifström: Interpolation Spaces. An Introduetion
224. Gilharg/Trudinger: Elliptic Partial Differential Equations of Seeond Order
225. Schütte: Proof Theory
226. Karouhi: K-Thcory. An Introduetion
227. Grauert/Remrnert: Theorie der Steinsehen Räume
228. Segal/Kunze: Integrals and Operators
229. Hasse: Number Theory
230. Klingenherg: Lcetures on C10sed Geodesies
231. Lang: Elliptie Curves: Diophantine Analysis
232. Gihman/Skorohod: The Theory of Stoehastie Proecsses 111
233. Stroock/Varadhan: Multi-dimensional Diffusion Processes
234. Aigner: Combinatorial Theory
235. Dynkin/Yushkevich: Markov Control Proeesses and Their Applieations
236. Grauert/Remmert: Theory of Stein Spaees
237. Köthe: Topologieal Vector-Spaees 11
238. Graham/MeGehec: Essays in Comrnutative Harmonie Analysis
239. Elliott: Probahilistie Nurnber Theory I
240. Elliott: Probabilistie Number Theory 11
241. Rudin: Funetion Theory in the Unit Ball of C"
242. Blaekburn/Huppert: Finite Groups I
243. Blaekburn/Huppert: Finite Groups 11
244. Kubert/Lang: Modular Units
245. Cornfeld/Fomin/Sinai: Ergodie Theory
246. Naimark: Thenry of Group Representations
247. Suzuki: Group Theory I