133 64 11MB
English Pages 304 [296] Year 1974
Lecture Notes in Mathematics Edited by A. Dold and 8. Eckmann
426 Martin L. Silverstein
Symmetrie Markov Processes
Springer-Verlag Berlin · Heidelberg · NewYork 1974
Prof. Martin L. Silverstein University of Southern California Dept. of Mathematics University Park Los Angeles, CA 90007/USA
Library of Congress Cataloging in Publication Data
1939Silverstein, Martin L Symmetric Markov processes. (Lecture notes in mathematics ; 426)
Bibliography: p.
1. Markov processes. 2. Potential, Theory of. I. Title. II. Series. Lecture notes in mathematics (Berlin) ; 426. 510'.8s [519.2'33]74-22376 QA3.L28 no. 426 [QA274.7]
AMS Subject Classifications (1970): 60J25, 60J45, 60J50 ISBN 3-540-07012-5 Springer-Verlag Berlin · Heidelberg · New York ISBN 0-387-07012-5 Springer-Verlag New York · Heidelberg · Berlin This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically those of translation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machine or similar means, and storage in data banks. Under § 54 of the German Copyright Law where copies are made for other than private use, a fee is payable to the publisher, the amount of the fee to be determined by agreement with the publisher.
© by Springer-Verlag Berlin · Heidelberg 1974. Printed in Germany. Offsetdruck: Julius Beltz, Hemsbach/Bergstr.
DEDICA TED
TO
W. FELLER
Introduction This monograph is concerned with symmetric Markov processes and especially with Dirichlet spaces as a tool for analyzing them. The volume as a whole focuses on the problern of classifying the symmetric submarkovian semigroups which dominate a given one. The main results are contained in Chapter III and especially in Section 20. A modified reflected space is determined by a boundary together with an intensity for jumping to
t:,
t:,
rather than to the dead point.
Every dominating semigroup which is actually an extension is subordinate to at least one modified reflected space.
The extensions subordinate to a
given modified reflected space are classified by certain Dirichlet spaces which live on the appropriate
t:,.
When the intensity for jumping to
t:,
vanishes identically, the subordinate extensions all have the same local generator as the given one.
The most general dominating semigroup
which is not an extension is obtained by first suppressing jumps to the dead point and/or replacing them by jumps within the state space and then taking an extension. Some general theory is developed in Chapter I. A decomposition of the Dirichlet form into killing 11 , 11
11
jumping
11
and "diffusion" is
accomplished in Chapter II. Examples are discussed in Chapter N. Each chapter is prefaced by a short summary. The main prerequisite is familiarity with the theory of martingales as developed by P. A. Meyer and his school. Little is needed from the theory of Markov processes as such, except from the point of view of motivation.
VI
For a treatment of classification theory in the context of diffusions we refer to [20] and [30]. In fact it is M. Fukushima 's paper [20] that inspired our own research in this area and his influence is apparent throughout the volurne. For more inforrnation on "sample space constructions" for extensions of a given process we refer to Freedman's book [52] where current references to the literature can be found. The expert typing was dorre by Elsie E. Walker at the University of Southern California.
Notations
g:
Throughout the volume
is a separable locally compact
Hausdorff space and dx is a Radon measure on Q which charges
quasi
every nonempty open set. The indicator of a setwill be denoted both by lA and I(A). a
condi.tion
and
$ I(X t
E
The integral of a function
over the set determined by
such as "X E ru will be denoted both by $[X Er: s] t t r)
s.
The measure which is absolutely continuous with
respect to a given measure cp • f..! •
s
f..! and has density cp will often be represented
The subcollection of bounded functions in ,!; will be denoted by Jb.
All functions arereal valued.
2 2 In particular L (dx) or L (X,dx) is the
real Hilbert space of square integrable functions on the measure space
(g:, dx)
and
C com (Q),
C (:lQ 0
are the collections of real valued
respectively with compact support and
continuous functions on
~
"vanishing at infinity."
Questions of measurability are generally taken
for granted-thus functions are usually understood to be measurable with respect to the obvious sigma algebra.
Table of Contents I.
General Theory _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
3
1. Transience and Recurrence_ _ _ _
_
3
_ _ _ _ _ _ _ _ _ _ _
20
2. Regular Dirichlet Spaces _ _
_ _ _ _
_ _ _
3. Same Potential Theory _ _ _ _ _ _ _
_ _ _ _ _
_ _
_ _
5. An Approximate Markov Process __ _ _
_ _ _ _ _
7. Balayage _ _ _ _ _ _
_ _ _ _
_
_
_
_ _ _ _
_ _ _ _ _ _ _
8. Random Time Change _ _
_
_
_ _ __
_
_ _ _ _ _
_ _ __ _ _ _ _ _ _ _
Decomposition of the Dirichlet Form _ _ _ - - - 9. Potentials in the Wide Sense_ _ _ _ _ _ _
-
_
24
39
_
61
_ _
69
_ _
_
78
_
_ _ _ _
84
-
- - -
-
97
_ _ _
98
_ _ _ _ _
10.
TheLevyKernel _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
102
11.
The Diffusion F o r m _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
112
J _ _ _ _ _ _ _ _ _ _ _ _ __
126
12. Characterization of III.
_ _
4. Construction of Processes_ _ _ _ _ _ _ _ _
6. Additive Functi0nals _ _ _
II.
_
_ _ _ _ _ _
x and
Structure Theory_ _ _ _ _ _ 13.
Preliminary Formula _ _ _
_ _ _ _ _
_ _
_ __ _ _ _ _ _
_ _ _ _ _ _ __ _ _ _ _ _ ___
14. The Reflected Dirichlet Space _ _ _ _
_ _ _
_ _ _ __ _ _ _ _
130 133 143
15. First Structure Theorem _ _ _ _ _ _ _ _ _ _ _ _ _ _ ·- _ _
152
16. The Recurrent Case _ _ _ _ _ _ _
_ _ _ __
158
17. Scope of First Structure Theorem- __ - - - - - - - - - - -
165
18. The Enveloping Dirichlet Space _ _ _ _ _ _ _ _ _ _
173
_ _
_ _ _ _
_ _ _
_
X 19. Equivalent Regular Representations __ _ _ _ _ _ _ _ _ _ _ _
178
20. Second Structure T h e o r e m - - - - - - - - - - - - - - - - -
183
Third Structure Theorem_---- _ _ - - - - - - - - -
216
Examples _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
220
_ _
222
21.
N.
22.
Diffusions with Bounded Scale; No Killing _ _ _ _ _ _ _
23.
Diffusions with Bounded Scale; Nontrivial Killing __ - _ - -
225
24. Unbounded Scale _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __
237
Infinitely Divisible Processes _ _ _ _ _ _ _ _ _ - - - - - -
248
25.
_ _ _
254
27. General Markov Chains _ _ _ _ _ _ _ _ _ _ _ - - - - _
258
26. Stable Markov Chains _ _ _ _ _ _ _ _ _ _ _
_ _ _
Chapter I. General Theory This chapter unifies and extends some of the results in [ 44] and [46]. In Section l we establish the connection between the submarkov property for symmetric semigroups for Dirichlet spaces
(~, E).
P
t
and the contractivity property
This was first discovered by A. Beurling
and J. Deny [l] but apparently it was M. Fukushima who first appreciated its significance for Markov processes. Also
in Section l we introduce
what seems to be the appropriate notion of "irreducibility" and we distinguish the transient and recurrent cases. We define the extended Dirichlet space
F
= (e)
by completing
__F relative to the
E form alone
(that is, without adding a piece of the standard inner product) and we show that
(~(e)
,E) is an honest Hilbert space when
(~,E)
is transient.
In Section 2 we show how a given Dirichlet space can be transformed into a regular one by introducing an appropriate modification of the state space.
Our construction differs only slightly from Fukushima's in [21].
Some potential theory for regular Dirichlet spaces is developed in Section 3 and used in Section 4 to construct a "decent" Markov process. The main result was first established by Fukushima [22]. Our approach differs from his in that we avoid Ray resolvents and quasi-homeomorphisms. In Section 5 we adapt G. A. Hunt' s construction of "approximate Markov chains" to our situation. In Section 6 we introduce various additive functionals, some of which are used to develop a theory of balayage in Section 7. In Section 8 we study random time change. We show that the time
2
I.2 changed process is symmetric relative to the "time changing measure" and we identify the time changed Dirichlet space. One immediate application is that if
(~, E)
is in the extended space
is recurrent then the constant function l
J (e)
and the norm E(l,l) = 0.
In particular
&(e) is not a Hilbert space which complements the result in Section l for the transient case.
3
J.
Transience and Recurrence
LJ, Definition. A symmetric (submarkovian) reso1vent on L 2 (dx) is a family of bounded symmetric linear operators
symmetric
G
u
f
> 0 whenever
f > 0 and
(v-u) G G
symmetric
U V
2 G , u > 0 on L (dx) satisfying. u
uG f < 1 whenever f < l.
u -
.///
Of course the inequalities in 1.1.1 are in the almost everywhere sense. It follows from 1. 1.1 tagether with the Cauchy-Schwarz inequality that each uGu
2 is a contraction on L (dx).
(Irrdeed this is the case on Lp(dx), 1
Consider now a given symmetric resolvent thecommonrangeofthe G
u
2 isdensein L (dx).
! Gu , u
~ p~ + ... )
> 0 l and as sume that
Itfollowsfrom 1.1.2 that
there exists a unique non-positive definite self adjoint operator A
2 on L (dx)
suchthat (1.1)
G
u
(u- A)
-1
u
> 0.
For t > 0 let (1. 2)
with the right side defined by the usual operator calculus. that the Pt
It is easy to check
satisfy the following conditions. 2 is a symmetric contraction on L (dx).
1. 2.1,
Each
1.2.2.
PtPs
1.2.3.
Limt~
1.2.4.
Gu =
Pt+s 0
for
s, t > 0.
2 Pt= 1 in the strong operator topology on L (dx).
J0"" dt e -ut Pt
for u
> 0.
4
there
whenever f > 0
P tf _:: 0
f < 1 whenever
and
1.2
f < 1. I I I
there is no difficulty in interpreting the integral in 1. 2. 4
Because of 1. 2. 3
Ta establish l. 2. 5 it suffices to apply Laplace inversion
as a Riemann integral.
as in [16, XIII. 4] to 1. 1. 1. Let
/_A be the unique nonnegative definite self adjoint square root of -A
and put (1. 3)
=
f
Jdx /-A
E(f,g)
Also for u
>0
domain / -A f(x)
/-A g(x)
f, g in
f.
put
{1.4)
Eu(f,g)
We will refer to the pair
Jdx
E(f, f)
+ u
(f,E)
as the associated Dirichlet space.
hx).
The following
is quite useful for analyzing (f, E). 2 For general f in L (dx)
Lemma l.l.
J dx {f(x) -
(1/t) increases as
t
~
0 and u
increases as f
is
in F
u t
Ptf(x) } f(x)
oo •
Jdx
{f(x) - u GJ(x)} f(x)
Each of these expressions remains bounded if and only if
and in this case
(1. 5)
E(f, f)
= Lim
(1. 6)
E(f, f)
=
tJ
0
Lim t u
(1/t)
00
u
Jdx {f(x) -
Ptf(x)} f(x)
Jdx {f(x) - u G u f(x)}
f(x).
III
5
Proof.
cg:"':dx),
1.3
By the spectra1 theorem there exists a measure space 'U.
an isometry
2
""'L 2 (~'"':dx),
L (:g:,dx)
2 suchthat f in L 0.
V(b)
I
t2
and (1):)(1-e-ta)
I
(u)
{ a(u+a) - ua}
I
(u+a)
2
2 2 = a / (u+a) •
Then
and it suffices to show that
C1earl y
v(+ 0
is negative.
sufficentl y small
Finally
e-b-be-b-e-b=- be-b is negative and we are done. / / / It is elementary that
products
Eu.
We say that
g
f
is a Hilbert s pace relative to any of the inner
A deeper result can be established with the help of Lemma 1.1. f if there exist ~verywhere
is a normalized contraction of
defined versions of (l. 9)
cp
then
l 2 - b(1-b) + (1-b + zb ) - 1 = -
~~ (b)
ua/(u+a)
For the fir st expres sion it suffices to
cp (t) = (1ltl (1-e-ta).
For the second expression 1et
cp 1(t) = { tae -ta - 1 + e-ta}
the expressions
f
!g(x) I ~
With our hypotheses on
and
g such that
lf(x) I
lg(x) - g(y) I ~
~ there exists for u
>0
lf(x) - f(y) I • a symmetric measure
6
g: x g:
Gu (dx, dy) on
1.4
suchthat
Jdx f(x) Gug(x) for of
JJ Gu (dx, dy) f(x)g(y) f
f in
Then for
for
g a normalized contraction
and for
f
(1. ~0)
Jdx{ g(x) =
- uGug(x)} g(x)
Jdx l
(x) {1 - uGu l(x)}
+ u
Jdx g 2 (x) { 1-uGu 1(x)}
JJ Gu (dx, dy) { g 2 (x)
+~ u
- g(x)g(y)}
JJ Gu(dx, dy) { g(x)
- g(y)}
2
and it follows that
Jdx {g -
Gug(x)} g(x) ~
Thus by Lemma 1.1 g be1ongs to f
l. 3 .1.
f
1. 3. 2.
If f is in
f- Guf(x)} f(x).
and E(g, g) < E(f, f).
We summarize in
is a Hilbert space relative to any of the inner products
also g belongs to f
f
(1. 4).
and if g is a normalized contraction of f then
and E(g, g) ~ E (f, f).
In general any pair with E
Jdx{
(f, E)
a bilinear form on F
if it satisfies l. 3.1 and l. 3. 2.
with
f
III
2 a dense linear subset of L (dx) and
will be called a Dirichlet space
~
2 L (dx)
We have shown above that the Dirichlet space
2 associated with a symmetric submarkovian resolvent on L (dx) is a Dirichlet 2 space on L (dx).
We now prove conversely that every Dirichlet space (f,E)
2 on L (dx) is associated with a unique submarkovian resolvent. 2 Let(f,E) beaDirichletspaceon L (dx). unique bounded operator from
2 L (dx)
into
f
For u>O let Gu bethe
determined by
7
Jdx cp (x) g(x)
(l.ll)
for
({J
L (dx). Let
2 in L (dx)
.f.
g in
and
It is easy to verify the resolvent identity
and that for u > 0 the operator uGu is a symmetric contraction on
l.l. 2
2
l.S
In place of 1.1.1 we prove the following more genera1
result.
T be any mapping from the rea1s to the reals satisfying
(1.12)
I TcY -
=0
TO
~
I
0 Eu (f - u Gu cp , f -u Gucp)
Jdx f l(x) + u cp (x) Gucp (x) -2f(x) (x) 2 = E(f, f) + u Jdx I f(x) - cp (x) r + u Jdx ! u cp (x) Gucp (x) - cp (x) l.
= E(f, f) + u
Cf!
and therefore the functional
P (f)
= E(f, f) + u
:fJx
-cp (x)
u Gucp.
has the unique minimum f T uGucp = u Gucp.
! f(x'
l2
But in general
iP (Tf)
~
p (f)
and so
W e summarize these results in
Let
Theorem 1.2.
I Gu,
u >0
2 with common rangedensein L (dx). 2
Dirichlet space on L (dx).
l
2 be a submarkovian resolvent on L (dx)
Then the pair Cf,El defined by (1.3) is a
2 Conversely every Dirichlet space on L (dx)
i~
2 associated in this way with a unique submarkovian reso1vent on L (dx) with dense common range.
Remark l.
III
Our restriction to resolvents having dense range and
therefore to Dirichlet spaces for Theorem 1.2.
2 F which aredensein L (dx) is not necessary
2 In the general case we need only replace L (dx) by the
8
1.6 closure of the range of the resolvent operators. can be weakened considerably. Borel set in the sense of (39].
Remark 2.
Also our assumptions about
It suffices for example that
~
X
be an absolute
III
The connection between a submarkovian resolvent
Gu' u > 0
and a submarkovian semigroup Pt' t > 0 is a familiar tool in the theory of Markov pmces ses. To our knowledge it was first used systematically by W. Fell er in his fundamental papers on diffusions and solutions of the Kolmogorov equations. The connection between symmetric resolvents and Dirichlet spaces is less well The basic idea goes back to Beurling and Deny [1].
known.
However it seems
that M. Fukushima [20] first appreciated its significance in the context of Markov processes.
III
We continue to work now with a Dirichlet space Cf,E), as in Theorem 1.2 and its associated resolvent general f > 0
with
f
n
! Gu,u
>0
l
and semigroup
! Pt,t
>0
l·
For
define
any sequence of square integrable functions which increase to f
almost
everywhere. Also define
when the right side converges.
It then follows from symmetry and from l. 2. 5
that the extended operators are contractions on Lp(dx) ,
l ~ p ~
+ "'.
the operators uGu in a similar way and note that by l.l.l the uGu are contractions on Lp(dx) ,
< p < + "'.
Extend
9 1 in L (dx)
Lemma 1. 3, For f
Limt t 0
Jdx I f(x)- Ptf(x)!
0. / / /
We use Rota's well known device [41].
Proof.
w
!.7
Irom the half line [o, ", > into the augmented space
be the usual trajectory variables and let
0 be the set of maps
Let
:g: u 1a l ,
let xt, t > o
t? be the unique measure on
0 such
that jt?(dw) f (x (w)) ... fn(Xt (w)) 0 0 n
for 0 < t 1 < • · • < tn on
0
by Xs'
generated by s
> t.
and for x
0
f , .•. , fn _::: 0
0
and for t > 0 let
are uniformly integrable on
:g:.
f in
1.2.4 to Lp(dx},
1_::: p _:::
A of
~
Ptf(x), t
>0
2 1 L (dx}nL (dx).///
The same argument gives an Lp result for 1 _::: p < +
the weak* topology on L"' (dx).
=~ - A
and therefore the functions
But this proves the lemma because it follows
Also it follows from symmetry that
c
er- algebra
are uniformly integrable on
inmeasureas tso for
froml.2.3 that Ptf,..f
A
.70 be the
.7 t be the er -algebra generated
P tf(X ), t > 0 2 0
0 in the sense of Hunt [27, Section I. 6]
A subset
Let
For f in L\X)
and it follows that the functions
Remark.
:g:.
on
co.
Pt is continuous in t with respect to
Once this is established it is routine to extend
+ "' • ///
is proper invariant if neither A nor its complement
is dx null and if PtlA < lA for all t > 0.
It follows from
10
symmetry that
A is proper invariant if and only if its complement
A
c
1.8 is.
2 f belongs to L (dx) then
Moreover if
J
+ dx l
Thus if f belongs to
F
Ac
then so does
space by restricting everything to
f(x) {1 f(x)- Ptl c f(x)} Ac A
lA f
A.
and we can obtain a new Dirichlet
This suggests that an appropriate
generalization of irreducibility for Markov chains is
1.4.
Condition
ill'
Irreducibility.
There exist
no proper
From now on we assume that this condition is satisfied.
invariant sets./// Our feeling is that
this restriction will be harmles s in practice. Next we apply the techniques associated with the Hopf decomposition to distinguish the transient and recurrent cases. is the book of Foguel (17]. in the discrete time case.
Our source for these techniques
Webegin by adapting Garsia' s well known argument Let
f be in L 1(dx)
and for n > 0 define
z-n
(1.13)
S f(x) n
J0
dt Ptf(x) kZ-n
S f(x) n,m
max1 < k < m
J0
The integrals are well defined in the L
dt P l(x)
1
sense because of Lemma l. 3.
Clearly
11
1.9 k2 -n
I0
dt Ptf
= sn f + p 2 -n
(k-1)2 -n
J0
dt Ptf
and therefore
s n,m f 0]
J dx (Si.'n,m f) + -
>
and after passage to the limit
J ,,
(1.15)
[s···
f >
n
Since
in L
1
2n Snf
where
"* f
E
=
o]
U:=O
1
n (S "'
n, m
f) +
> 0
> 0.
as n t "'
[s:
2-
m t "'
dx Snf(x)
in L
jdx P
also
(Again we are using the concept of
f > 0 ].
uniform integrability as formu1ated in [2 7, I. 6 ]. ) Thus after multiplying (1.15) by 2n and passing to the limit n t "'
Lemma 1. 4.
we obtain
1 Let f be in L (dx)
and let k2-n
E =
fx
:- supn
2: O,
k >1
~
dt Ptf(x) > 0
J.
12
IE dx f(x)
Then
>
-
l.lO
o.lll
We deiine the Green•s operator G by
{1.16)
Gf
when it makes sense. Corollary 1. 5.
Let f, g > 0 almest everywhere be in L 1 {dx).
[Gf < + oo] = [Gg < + ce] Proof.
Let A
I
B
almost everywhere.
(Gf= "', Gg 0 almost
But if f :::_ 0 is nontrivial then by irreducibility G f > 0 almost 1
13
l.ll Gf 2:_ G G f 1
everywhere and since
1
and nontrivial in L (dx).
f > 0
of the choice of
it follows that the outcome is independent Thus the following makes
sense.
l. 5.
The Dirichlet space
Definition.
0~::,
E) is transient
if
Gf finite almost everywhere for all f in
1 L (dx) and recurrent if almost everywhere
Gf = + "' forall f>O
1 andnontrivialin L (dx).///
For h > 0
r
Jo stays bounded as t t Theorem 1 6
t+h
h ds p sl -
t
ds p
s
l
There follows
co •
If
J
Cf, E)
is recurrent, then Ptl
1 almost everywhere for
every t > 0. / / / Of course the converse to Theorem 1. 6 is false.
(Consider for example
standard Brownian motion in Rd for d > 3.) f belongs to the extended Dirichlet space
1. 6 Definiti.Qn.
exists a sequence fn is
almost
l. 6. 2.
rf
n
such that
is Cauchy relative to E.
fn ,..f
Condition 1. 6.1
1. 6 1'
f
E(fn' fn)
almost everywhere on
-z. I/I
can be replaced by the apparently weaker is bounded independent of
n. / / /
f (e)
if there
14
!.12 To prove this we adapt the proof of a well known theorem of Banach and Saks [40, p. 80].
fi n l
We assume that
satisfies
the Ces'aro means of a subsequence satisfy 1. 6,1. computation
we temporarilly introduce
by first identifying functions in completing.
cp weak1y in
:r
:r.
and we show that
To avoid excessive
the Hilbert space formed from
whose difference has E
We use the same symbol for functions in
equivalence classes in f n-+
f
:r
1.6.1'
J'
norm zero and then and their corresponding
After se1ecting a subsequence we can assume that
and after again se1ecting a subsequence we can assume
that
E(cp - fm, cp - fn) < (1/n) C1early l. 6.1
for
n > m.
will follow if we show that
(1.17) But the left side of (l. 17)
2 n (1 n ) L:k=l E(cp -fk, cp -fk)
I
and we are clone.
Remark,
There is an alternative noncomputational argument which uses
the fact that the weak closure of a convex subset of a Hilbert space is also the strong closure,
III
If
15
Lemma 1. 7. (i)
Let f be in ,I(e)
The limit
Limn t
approximating sequence
E(fn, fn)
co
! f n !.
I fn l
and let
1.13 be as in Definition 1. 6.
is independent of the choice of the
Therefore E
extends uniquely to
by continuity and
(1.18)
E(f, f)
(ii)
(1.19)
=
Lim t n
c:o
E(f
f ). n, n
The expressions
(1/t}
Jdx
{1-Ptl(x)} hx)
+ ~(1/tJJfPt(dx, dy){f(y)-
f(x)}
2
(1.19 1 }
arefinite for u, t
Proof.
> 0 and increase to E(f, f) as t
By the triangle inequality
will follow from (ii) increase to
t .
t
= Limn
J0
t
and u t"'
co E(fn'
if we show that the expressions
fn)
respectively./I I
exists.
Thus
(1.19) and (1.19')
We give the argument only for (1.19) and temporarilly
introduce the specia1 notation Et(f, f} for (1.19). Et(f
The estimate
-f , f -f ) < E(f -f , f -f ) mnmnmnmn
is valid by Lemma 1.1 and it follows with the he1p of Fatou' s 1emma that
By the triangle inequality Et(f, f) 1 jEI (f, f) -
1
E~
(fn' fn)
I~
is finite and 1
Lim supm t
"'E~fm -fn' fm -fn)
(i)
16
1.14 In particular Et(fn' fn)
Et(f, f) increase as
-1>
Et(f, f)
t 10
and so again by Lemma l.l the expressions
and are dominated by .e.
Finally convergence
to .e follows from the estimate
l/ +./ Et(fn ,fn)- Et(f,fl/ ,f n n )- Et(fn n l/ +I E(fn,f -< /.e-E(fn,f E(f -f,f -f)+/E(f,f)-Et(f,f)/.111 0 then actually Gcp _::: M
on the set where
Remark 2.
and E(l, 1)
Hilbert space relative to E.
almost everywhere./11
0. Thus
f(e)
cannot be made into a
This is also true for dx unbounded, but the
proof must wait until Section 8. Remark__l.
almost everywhere
E) is recurrent and if dx is bounded then by Theorem l. 6
If
the function 1 is in f
III
f relative to the E norm.
Therefore in the
transient case these operators extend by continuity to contractions on From now on we take these extensions for granted. strong Operator topology on f(e)
u
=
J0 dt e -ut p
Pt
It follows from the spectral theorem that the operators
and u Gu are contractive on
G
f
We summarize in
cp > 0
Moreover if
bounded on the set where rp > 0 then Gp f
Also it follows
Of course the limit function
Theorem1.9 . .!f (f,E) is transient then as inner product.
and to cp _:. 0 in
immediately
subsequence converges almost everywhere.
E
in f (e)
This extends immediately to f
as
t ~0
and
f
(e)'
Also Pt' u Gu-+ 1 in the u t "'•
the identity
00
topology on f(e)
t
is valid for the extension, and Pt -+ 0 in the strong operator
as t
tce.lll
19
1.17
Remark 4.
We have recently received a preprint from M. Fukushima
entitled "Almost polar sets and an ergodie theorem. " Among other things this paper deals with transience and recurrence in a nonsymmetric setting.
III
20 Dirichlet Dirichlet Dirichlet
.f.J.
Definition.
f n
2.1.1. 2 .l. 2. open s et
The Dirichlet space
ccom(~)
is uniformly densein ccomqp and That is
dx is everywhere dense. G.
2 (f, E) on L (dx) is regular if
III
Condition 2.1 .2
dense in .f.
1
for any nonempty
is harmless since we can always replace
by the
~
(.f, E) is regular and transient then also
Notice that if
support of dx.
JG dx > 0
E
.F n Ccomqp is Edensein (f, E)
We consider now a Dirichlet space is dense in
2
L (dx)
on
2 L (dx)
suchthat
and we construct a regular Dirichlet space
modifying the state space
f
by
Our construction differs only slightly from one
~-
first given by Fukushima [21]. This s ection is es sentially a r epetition of the appendix in [ 44]. Note first that if f in to
Jh
f
is bounded almost everywhere, then
and so the subcollection of f in
form an algebra.
f
f
2
belongs
which are integrable and bounded
Obviously this algebra is dense in
f
and since
f
itself
is separable (since it is the domain of a self adjoint operator on a separable Hilbert space) there exists a subset
2.2 .1.
~O
2.2.2.
~O
of J"
satisfying
is countaLle is an algebra over the rationals
2.2.3.
Every f in
2.2.4.
~O
is densein
The uniform closure
is integrable and bounded,
~O
~
f
(and therefore in
2 L (dx)
)./II
is a commutative Banach algebra and so the
well known techniques associated with the Gelfand transform of ~ are available.
21
2.2
Y be the collection of real valued functions
Let
2. 3.1.
y (fl
$: with full measure.
a unique mapping
f in
flo·
for f in
2.4.2.
J?.
satisfying
~
2.4.1.
for some f in
is then a separable locally
compact Hausdorff space which is compact if and only if l
let
not identically
on
00
y(af+bg)
2. 3. 3.
J?
f, g in
zero, which satisfy for
y
There exists
such that
{i x) (f) = f(x) for f in
J?o
and for x
in
:lfio.
Clearly
there exists a unique Borel measure dy
ly
dy cp(y)
=
on Y
is Borel measurable and so such that
22
for nonnegative cp that
dy
is
mapping of
Radon.
It follows from the integrability of cp
We use the same symbol
;!? onto c
in
B
0
i to denote the natural
Thus
(Y).
0 -
jf (y)
Clearly for
2.3
'f.
on
y (f)
f
E
:!?•
y
E
~·
f in ;!? and for any polynomial P iP(f)
and since any T
=
P(if)
satisfying (L 12)
is continuous and therefore can be
approximated uniformly by polynomials on compact sets, also iT(f) For
f, g _:_ 0
I dy
in
j E
~
= I dx
j f(ix) j f(ix)
= I dx
jx(f) j x(g)
= I dx
jx(fg)
= I dx
f(x)g(x)
j
be defined on
The desired
onto
is an isometry from
i :I?o by
jE(j f, i g)
product
T(if).
jf(y) jg(y >
and it follows that Let
=
= E(f, g).
Dirichlet space is the closure of i B
0
relative to the inner
23
2.4 tagether with the continuous extension of
To establish
to this closure.
E
regularity for this Dirichlet space it on1y remains to checkthat
dy
is dense.
For this it suffices to show that
n~=1 ! X
meas.
(2 .1)
:
e > 0 of
for any choice of
I fi
in
y
y
and of
is fa1se then there exist po1ynomia1s Pm Pm (fl' •.. , fn)
> 0
f , ••. , fn in 1
~.
If (2.1)
in n indeterminates such that
converges uniform1 y to
and it follows that gh be1ongs to in
l
(x) - '{ (fi) ) < e
~
whenever
h
does.
h
Since every
:§1 can be represented
and since
max ~=
by polynomia1s in
follows
that y (h)
1 I fi
- y (fi)
I
I f.1
- '{ (f.)
l
1
can be uniform1y approximated
i=l, ... , n
0 for all hin
J:?,
not containing the constant term, it
which possibility has been ruled out
by hypothesis.
Remark l. Remark 2.
The proof of (2.1) given on page 71 of[44] is incorrect.lll Clear1y the final state space
However by Theorem 2.1 in [22] for
is related to
y
be a
any
1'
J
depends on the choice of
:§1 0 .
resulting from a different choice
"capacity preserving quasi-homeomorph ism"
(see [22] for the precise definition) which is enough to guarantee that
1
are identical from the point of view of the processes constructed below.
return to this subject in Section 19.
III
and We
1
I
24 3. Theory Potential Theory Theory
Throughout this section 2 on L (dx). E 1. )
(.f, E) is a transient regular Dirichlet space
(The recurrent case can always be handled by replacing E
with
The point of view taken here goes back at least to H. Cartan [3,4] for the
classical Dirichlet spaces associated with the Laplacian and Brownian motion.
The general formulation in terms of regular Dirichlet spaces is
J. Deny.
due to A. Beurling and
(See [1].)
The results themselves were
first established by Fukushima [22] using an indirect a pproach. This section differs only slightly from sectionsection is in f(e)
and
f in f(e) isapotential if E(f, g)
g ~ 0
(i)
f is a potential.
(ii)
There exists a Radon measure
E(f, g) for
lf (e) n
g in
E(f+g, f+g)
(iv)
f,?:. u Guf f > Ptf
Proof.
g
=
f ( e )'
1-1 suchthat
J 1-1 (dx) g(x)
ccom (~).
(iii)
(v)
> 0 whenever
almost everywhere. / / /
Th .::..::;.:e:_.:cfo:::l::.:lc:::oc.:w:..:.i:.:cn:.....::a:::r...:e:....:::e.:.q.::u;:.iv.:..a:.:.l::ce::.:n:.:.t:.....:.fo::::..::..r__:__:i:::n
section
section
Section l in [ 44].
> E(f, f) whenever for all u
for all t
That (ii) implies
g
is in f ( e)
and
g > 0.
> 0.
> 0. / / / (i) is trivial.
To prove that (i)
implies (ii)
let f be a potential and consider the nonnegative linear functional I defined Oll
g in
~
n ccom
(~)
by I(g) = E(f, g).
If g n
decrease to 0
pointwise
then by Dini' s theorem they do so uniformly and after comparing to a fixed
25
3.2
n C com ()~)
11;
nonnegative g in
I(gn) ! 0.
we see that
Thus
which is _:: 1 on the support of
g1
follows by the Daniell approach to
(ii)
integration (as presented for examp1e in (34] ).
That
(i)
implies (iii)
follows from (3 .1)
E(f+ tg, f
for
t
=1
small.
+ tg)
= E(f, f)
+ 2t E(f, g) + t 2 E(g, g)
implies (i) follows from (3 .1) for t > 0
and that (iii)
sufficiently
Equivalence of (iv) and (v) is easily established with the help of
Laplace inversion.
Jdx
(3. 2)
That (i) impli es
j f(x) - u Gu f(x) ] cp (x)
cp > 0
which is valid for
(iv) follows from
as
in Theorem 1. 9.
That (v)
implies (i)
follows from (3. 3)
E(J
s
0
upon dividing by
Jds
du P f, g) u
s
j f(x)
P f(x) J g(x) 8
and passing to the limit
s l 0.
(Thc identity (3. 3)
follows in the same way as (1. 20).) / / /
Corollary 3, Z. If
(iii)
If f
E(/f/ ,
lrl
Every potential is nonnegative.
f, g are potentials, then so is min(f, g).
(ii)
Proof.
(i)
isapotential then so is min(f,c) for
(i)
follows since
l
E(f, f)
+ Ed
rl - f, Ir/
+
I
-fl
+ 2E(f,
E( f/ - f,
I f/
-f)
> E(/f/, if/ + E(/f/ -f,
/f/
-f)
.:: E(f, f)
c > 0. / / /
Ir! - f)
26
and therefore
E(
lfl - f,
= 0.
lfl - f)
Conclusions
directly from Lemma 3.1- (iv) or (v).
Lemma 3. 3.
(ii)
and
3.3
(iii) follow
III
Let g be a potential in F
and let f
satisfy 0 < f _:: g
and in addition
all t > 0
from or
u G
u
Then f is a potential in F
f
O.
and E(f, f) _:: ~·
I II
Again we consider only the first alternative.
Proof. follows from
Lemma 1.1 and the estimate
Jdx {f(x)
2
- Ptf(x)} f(x)
Jdx f(x)
0 suchthat
J~ (dx)f(x) for f in
has finite energy if there exists
~
l
2
c {E(f, f)}
2
The collection of all such measures is denoted by 'fl1.
27
3.4 cp 2:_ 0 such that cp • dx belongs to
The collection of Borel by '!!! o
'lll is denoted
II/ then there exists a unique potential, written
belongs to 'lJ?
Clearly if i.J. N i.J., suchthat
(3.4)
for
E(Ni.J., g)
g in
f
(e)
n
C com
qp.
= j 1-L (dx)g(x) We introduce the special notation
(3. 5)
Important
and call E(i.J.) the energy of i.J..
'lJ?
compactnes s properties of
are summarized in
Lemma 3.4. (i) indeed f (ii)
be a sequence in 'l/1.
_!! Ni.J.n converges weakly to f in f (e) then f is a potential and Ni.J. where
.!f
is the vague limit of the
i.J.
is bounded and if
E(i.J.n)
and Ni.J.n -+ Ni.J. (iii)
I I-ln I
Let
weakly in
J' (e)
1-l n • i.J.
I-ln -+ 1-l vaguely then
is in
'!!!
•
'lll is complete relative to the energy metric E.
/II
This lemma is an immediate consequence of regularity.
We omit the proof.
To make further progress we must validate (3.4) for general g in I'(e) which means in particular that we must represent g by a refinement which is specified up to
j.L
equivalence for every
i.J. in 'l/1
The main tool for this
is a capacity as sociated with E. J..,2..
Definition.
For G an open subset of Cap(G)
= inf
E(f, f)
~
let
28
as
runs over the functions in
f
g:
subset of
G
A
For
a general Borel
let inf Cap(G)
Cap{A) as
= + "' •
If no such f exist let Cap(G)
on G.
3. 5
suchthat f > 1 almost everywhere
E(e)
runs over the open supersets of
supersets
A
Borel set A
A. / / / is
if Cap(A)
= 0.
A
general
set is polar if it is a subset of a Borel set which is polar.///
subset
g:
Let G be an open subset of G
(i)
f
Th_......_e_r_e_e_x.;;;ic:..s.:..ts:;_;a.;_:u::::;ru.;::·.:;oq~u"'e-f:..:u:::n:.:cc:..:t==i.;:o.::n:.__,P_ __::i:::n
f (e)
is minimal a.mong f in (ii) (iii)
0 < p
G
~
l
>
satisfying f
G
p
+
co •
G G (e) .::.s.::.u:..:c;:;h:....::::th:.:ca::.t:__..;E:::..>.;(p"-~'-"p--')
almost everywhere on G. _G=l a 1mos t everywh ere on G • P
almost everywhere and
pG is a potential and indeed
Cap(G)
1 almost
f in f(e)
W is convex and closed and
Clearly
(i)
follows directly.
Concclusion (ii) follows upon noting that if f belongs to W then so does min (f, 1) and max(f, 0). and if g E(f,g)
~
~
prove
To
on G then E(f+tg, f+tg)
0
note first that if g is in
(iii) ~
E(f, f) for all t > 0
(e)
and therefore
It only remains to adapt the proof of Lemma 3.1 considering
0.
. /II
restrictions to
Proposition 3. 6.
For
v in 'lr 1
{3. 6)
f
v(G)
J
and for
G
an open subset of
~
1
2 ! Cap(G) l 2 . ///
If PG were in C com qp this wou1d follow from (3. 4).
In general this is
29
3.6
false and instead we must approximate with the help of ~emma
Let
3, 7,
v be in
'11/. 0
belongs to
(ii)
'11/
for t > 0,
Lim t J 0 (1/t) (1-Pt) Nv (x) dx"' v both vague1y and relative to the
energy metric E,
I II
For T > 0
Proof. T
Therefore
T+t
t
10
ds P
s
(1-Pt)Nv
(1-Pt) Nv is in
10 ds '11/
0
P Nvs
1T
ds P Nv s
and
t
J0
ds P Nv. s
and the lemma follows with the help of Lemma 3.4.
III
Now Proposition 3. 6 follows from v (G)
J
dx(l-Pt) Nv (x) < Lim inftl 0 (1/t) G
Lim t!O (1/t) E(pG, G(l-Pt)Nv) E(pG, Nv)
and the Cauchy-Schwa rz inequality. Corollary 3. 8.
III
If v belongs to
'11/ then v charges no polar set.
III
Next we establish some properties of Gap which permit the application of Choquet' s general theory.
30
3.7 Proposition 3 ._2_,
Cap(G ) 1
(i)
G
and p n t p (iii)
G
(i) is clear. g
For
open
G , G2 1
To prove(ii) observe first that if m
G
(3. 7)
If Cap(G) is finitethenalso
G
G
> E(g, g) + E(p m-g, p m_ g)
and we conclude that g
=p
G
m
G
p
and so the
If supn Cap(Gn)
everywhere sense.
But also
).
> E(p m.p
E(g, g)
G G E(p m, p m)
Cap(Gn) 2: Cap(A)
the sequence
and it suffices to take
Let Gn
vn be as in
For each n clearly
G. fl
any vague limit point of
(Actually the proof of Proposition 3. 9-(ii) shows that the entire vn converges vaguely.
)//I
32
3.9 Now we are ready to introduce refined Versions.
Now
A property is valid q_uasi-everywhere (abbreviated q. e.)
Definition.
Two functions are quasi-equivalent if they differ
if the exceptional set is polar.
only on apolar set, that is, if they are equal quasi-everywhere.lll Now Definition.
fn-+ f quasi-uniformly if there exists a decreasing
sequence of open sets
Um with Cap(Um) I 0
on ~- Um for each
m.lll
Now
A function f
on
~
suchthat fn -+f
is quasi-continuous on an open set G if there
exists a decreasing sequence of open sets
Um with Cap(Um) l 0
f is defined and continuous on G - U ll1 for each m.
Theore1Jh3.1J.
uniformly
(i) Each f in
f(e) has
suchthat
III
a representative uniquely
specified up to to quasi-equivalence such that
(3.9)
f is quasi-continuous on f(e)
(3.9)
Cap
(3.10) for
fx
(f(x))
> e
Proof.
and for
iJ
l