groups
Citation for published version (APA):
Overdijk, D. A. (1979). The recurrence behaviour of random walks on locally compact groups. (Memorandum COSOR; Vol. 7906). Technische Hogeschool Eindhoven.
Document status and date: Published: 01/01/1979
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
PROBABILITY THEORY, STATISTICS, OPERATIONS RESEARCH, AND SYSTEMS THEORY GROUP
Memorandum COSOR 79-06
The recurrence behaviour of random walks on locally compact groups
by
D.A. Overdijk
Eindhoven, May 1979 The Netherlands
D.A. Overdijk
Introduction and summary
+
Let (x,I) be a measurable space and let
M
be the class of nonnegative ex-tended real valued measurable functions on (X, I). A Markov process on (X,~) is a mapping P of M+ into itself such that( 1) 00
I
n=i a. pf n n P 00I
n=i a. f n n (a.n + ~ 0, f EM) , n ( 2) Pi ~ 1If we put P(x,A) PiA(x) for every x E X and every A E I, then P(·,·) is
the (sub) transition probability describing the process on the state space
(X,I). Conversely, every (sub) transition probability on (X,~) determines
a Markov process on (X,I).
Sometimes we want to consider Markov processes in a more global sense i.e. modulo a a-finite measure ~ on the state space (X,I). Let (X,I,~) be a
a-+
finite measure space and let
M
(~) be the space of equivalence classes of~-almost everywhere equal nonnegative extended real valued measurable func-tions on (X,I). A Markov process on (X,I,~) is a mapping P of M+(~) into itself such that (1) and (2) hold ~-a.e.
When P is a Markov process on a measurable space (X,I) and ~ is a a-finite measure on (X, I) , then in general P is not a Markov process on the measure space (X,I,~). This is the case if and only if P is nonsingular with respect to ~ i.e. for every A E I with meA) = 0 we have m({Pi
A >
oJ)
= O. Conversely, if P is a Markov process on a measure space (X,L,~), then in general there does not exist a (sub) transition probability on the measurable space (X,I) describing the process.In this note we shall discuss the recurrence behaviour of a random walk on a locally compact abelian group (G,I), where ~ denotes the a-algebra of the Borel sets of G.
Our main tool to investigate the recurrence behaviour of random walks will be the use of embedded processes. Let P be a Markov process on a measurable space (X,~), then for every A E L the embedded proces P
A is the Markov pro-cess on (X,I) given by
00
I
n=O
+
Here AI stands for x\A and I for multiplication by the indicator function
A
1 of the set A. The expression P 1 (x) is the probability that, starting
A A B
in x, at its first visit to the set A the process will enter the set B. Section one gives preliminaries concerning Markov processes. In section two we shall show that a random walk is either conservative or dissipative and
the random walk is conservative if and only if the random walk is recurrent. Section three gives criteria for a random walk to be Harris (cf. [5J, ch. 3 theorem 4.4) and to be ergodic. In section four some recurrence criteria for random walks onmn (cf. [2J) will be given.
1. Preliminaties concerning Markov processes
In this section we collect some results about Markov processes, which can be found in [7J. P will be a Markov process on an arbitrary measurable space
(X,I). A a-finite measure ~ on (X,I) is said to be invariant with respect to
P if
J
Pf(x)~(dx)
J
f (x)~
(dx) for all fEM.
+Proposition 1.1. If P is a Markov process on the measurable space (X,I) and the a-finite measure ~ on (X,I) is invariant with respect to P, then P is nonsingular with respect to ~ and therefore P can also be considered as a Markov process on the measure space (X,L,~).
Proof. Suppose A E I with ~(A)
=
O. Since ~ is invariant we haveHence PiA
o
~-a.e. and therefore ~({P1 >a})
A O.
o
Proposition 1.2. If P is a Markov process on the measurable space (X,I), then for all sets A c I and all f E M+ we have
y
n=lProof. Replac~ in [7J proposition 2.2 A by AI, B by A and f by IAf.
0
Let ~ be a a-finite measure on (X,I) such that the Markov process P on (X,I) is nonsingular with respect to ~, then P is also a Markov process on the mea-sure space (X,I,~). The decomposition theorem of E. Hopf states, t~at there
exists a mod ~ unique decomposition of the state space X in a conservative part C and a dissipative part D ([3J, ch. II). In the following theorem we collect some important properties of the conservative and dissipative parts.
Theorem 1.1. Let P be a Markov process on the a-finite measure space (X,L,~) and let C and D be tile conservative and dissipative parts of X with respect to P, then we have
i) For all A E L with Ace
!
pnlA
=
00~-a.e.
on A . n=Oii) There exists a measurable partition of D ql ,q2' ... such that PD. 1 ~ qi < 1 ~-a.e. on Di ~ D 1 u D2 u ••• and a sequence 00
I
n=O 1~ ~-a.e. on X for all i • 1 - q.
~
For a proof of theorem 1.1 see [7J, theorem 2.1 and 2.2.
Theorem 1.2. If P is a conservative Markov process on a measure space (X,L,~)
+
and for a function f E
M
we have Pf ~ f ~-a.e., then Pf=
f ~-a.e.For a proof of theorem 1.2 see [3J, ch. II theorem B.
Let P be a Markov process on a measure space (X,L,~). A set R E L is said to
be invariant, if P1
R 2 lR ~-a.e. Intuitively, this means that i t is
impossi-ble to leave the set R under the action of the process. There is no unanimity in the definition of invariant set. The definition in [3J differs from our definition, but for conservative Markov processes they are equivalent because of theorem 1. 2.
Proposition 1.3. If P is a conservative Markov process on a measure space
*
(X,L,~), then for every A E L we have PAl
=
lA
*
~-a.e., where A is the mod ~ smallest invariant set containing A.Proposition 1.4. Let P be a Markov process on the measure space (X,~,~) and
th th th {dV > O} b 1
v a finite invariant measure wi v«~, en e set d~ e ongs to
the conservative part.
For a proof of proposition 1.4 see [3J Ch. IV theorem E.
If P is a Markov process on a measure space (X,~,~), then i t is easily veri-fied that for every A E ~ the embedded process P is also a Markov process
A
on (X,~,~). Let C(P) be the conservative part of X with respect to P and C(P
A) the conservative part of X with respect to PA'
Proposition 1.5. Let P be a Markov process on the measure space (X,~,~), then for every A E ~ we have C(P
A) = A n C(p) .
For a proof of proposition 1.5 see [7J, proposition 2.4.
Proposition 1.6. Let P be ~ Markov process on the a-finite measure space
(X,I,~) such that ~ is invariant with respect to P. If A E I with 0 < ~(A) <00
and PAl = 1 ~-a.e. on A then A c C(P) •
Proof. The restriction of ~ to the set A will be denoted by ~A' We shall prove that ~A is invariant with respect to P
A" Using the invariance of ~ with respect to P, it is a straightforward verification by induction, that for all N E ~ and all bounded functions f E
M+
we haveHence
(1)
for all B E I . Define the measure v on (X,~) by
v(B) =
J
IAPA1B(x)~(dx)
Since v «
~
the Radon-Nikodym derivative~~
exists and from (1) we conclude dv < 1~-a"e.
d~ - A
Since PAl = 1 ~-a.e. on A we have
d\l
Hence - -
=
1 ~-a.e.d~ A
+
For every f E M we now have
J
PAf(x)~A(dx)
=
f
=
J
=
f
f (x)~A(dx) •Hence ~A is invariant withe respect to P
A• From proposition 1.4 we conclude A C C(P
A) and by propositior11.5 we have A C C(P). , 0
2. Recurrence and conservatively for random walks
Throughout the sections two and three G will be a locally compact abelian group. We shall suppose that G is metrizable and d is a metric on G compati-ble with the topology. The group operation in G will be written additively. The a-algebra of the Borel sets of G will be denoted by L and the Haar mea-sure on (G,L) by A.
A random walk on (G,L) is determined by a probability measure p on (G,L). Following Revuz (see [5J, p. 27) the probability measure pis called the law of the random walk. A random walk with law p is a Markov process on
(G,L) with transition probability P(x,A) =
J
1A(x+Y)P(dy) for all x i G and all A E L •
The corresponding Markov operator P on M+ is
Pf(x) =
f
f(x+y)p(dy) for all x E G and all f E M+ •In the sequel we shall frequently use the translation operator T (a E G)
a
M
+.on For every a E G define
T f (x) = f (x+a)
a
Proposition 2.1. Let P be and every A E L we have
i) PTa
=
T P,
a
ii) PATa T P
a A+a
. M+ for all x E G and f E •
a random walk with law p, then for every a € G
+
Proof. For every f E M we have
PT f.(x)a =
f
T f(x+y)p(dy)a =J
f(x+y+a)p(dy) = T Pf (x)Hence the operators P and T commute, which proves i). For every BEE we d. have Hence We conclude 1 B+a(x+a)f(x+a) T Ia B+af (x) • P T A a 00
L
n=Oo
Proposition 2.2. The Haar measure A is invariant with respect to every ran-dom walk.
+
Proof. Let P be a random walk with law p, then for every f E
M
we have using Fubini 's theoremf
Pf(x) A(dx)f
(J
f(x+y)p(dY))A(dx)=
J
(f
f(x+Y)A(dx))p(dy)f
f(x)A(dx)f
p(dy)= J
f(x)A(dx) •0
Proposition 2.3. If A,B E E with A(A) > 0 and A(B) > 0, then there exists a set C E L withA(C) > 0 such that A«A+p) n B) > 0 for all p E C.
Proof. Suppose A( (A + p) n B)
=
0 A-a. e. on G. Then we haveo
=J
.>"«A+p) n B)A(dp) =f
(f
1A(x-p)lB(x)A(dX))A(dp)= A(-A) A(B) > 0 •
This contradiction proves the proposition.
o
Let P be a random walk on (G,L) with law p. Because of the proposition 1.1 and 2.2 the random walk P is also a Markov p~oces on the measure space
(G,X:,A) •
Theorem~. Let P be a random walk on (G,L) with law p. The random walk as
Markov process on the measure space (G,L,A) is either conservative or dissi-pative.
Proof. Let C and D be the conservative and dissipative parts of C with res-pect to P and suppose A(C) > 0 and A(D) > O. It follows from theorem 1.1 that
there exist a Borel set A c D and a number s such that A(A) > 0 and
I:=o
pnl A~
s A-a.e.From proposition 2.3 we conclude, that there exists a point pEG such that A«A+p) n C) > O. Put B = (A+p) n C, then B c
c,
A(B) > 0 and by proposi-tion 2.1 00 00I
n=O 00 00 T \ pn1 ~ s A-a.e. -p L A n=O Since B c C we conclude from theorem 1.1 A(B)the theorem.
O. This contradiction proves
o
Remark~ When we say a random walk is conservative (dissipative), t~e random
walk is considered as a Markov process on the measure space (G,L,A).
Let P be a random walk on (G,L). A set A E L is said to be recurrent if for
every starting point x E A the random walk will return to the set A with probability one i.e. PAl (x) = 1 for all x E A. The random walk P is said to be recurrent if every open set is recurrent (cf. [1
J,
def. 3.31). Let flbe a a-finite measure on (G,L), then a set A E L is said to be fl-recurrent if PAl = 1 fl-a.e. on A. It follows from theorem 1.1 that the random walk is conservative if and only if every set A E L is A-recurrent.
One might think, that a recurrent Markov process is conservative. This, how-ever is not the case, there exists a recurrent Markov process on OR,L,A), which is dissipative (see [6J). For random walks however recurrency and con-servativity are equivalent.
Theorem 2.2. A random walk is conse~vativeif and only if the random walk is recurrent.
Proof. First suppose the random walk is conservative. Let A be an open set such that there exists a point pEA with PAl (pi
=
s < 1. For all E > 0 and all x E G put Ue(x) = {y E GI
d(x,y) < e}.Choose e > 0 such that U
2e(p) c A.
Because of proposition 2.1
we
have for all q E U (p) EP U ( ) 1(q) £ p = Pu (p) _q+pl (p) £ P T 1(p) U (p)-q+p q-p £ s < 1 •
By theorem 1.1 we have U (p) belongs to the dissipative part of G and hence
£
A (U (p» =
o.
£
This contradiction proves that the random walk is recurrent.
Conversely, suppose the random walk is recurrent. Let A be an open set with
a
< A(A) < 00. From proposition 1.6 we conclude that A belongs to the vative part of G. It follows from theorem 2.1, that the random walk is conser-vative.Corollary. A random walk is conservative if and only if there exists a A-recurrent set A E E with
a
< A(A) < 00.D
Proof. If the random walk is conservative then it follows from theorem 1.1 that every set A E E is recurrent. Conversely, suppose there exists a
A-recurrent set A E E with
a
< A(A) < 00, then it is an immediate consequence of proposition 1.6 and theorem 2.1 that the random walk is conservative.D
3. Harris recurrence and ergodicity
We start with some definitions and properties of probability measures on
(G, E) •
If ~ and v are two probabilities on (G,E), then the convolution ~
*
V is a probability on (G,E) defined by(~
*
v) (A) ==J
J
lA(x + y)
~
(dx)v(dy) for all A E E • *nThe n-fold convolution of ~ will be denoted by ~
Proposition 3.1. If ~ is a probability on (G,E) with ~ « A, then for all probabilities V we have ~
*
v « A.Proof. Suppose A E E with A(A)
=
O. Since ~ « A and A(A)~(A- y)
=
a
for all y E G. Hencea
we haveq + r with q « A and r singular with respect to A.
n n n n
Let ~ be a probability on (G,E), then for n ~ 1 there exists the
decomposi-. *n
t~on ~
Proposition 3.2. If ~ is a probability on (G,E), then r ~ r
*
r for alln+m n m
nand m.
Proof. ~
*
(n+m)=
~*n*
~*m=
(q +r )*
(q +r )n n n m
I t follows from proposition 3.1 that r ~ r
*
r .n+m n m IJ
Following Revuz ([5J, p. 91) we say that a probability ~ is spread out, if
*p ,
there exists an integer p such that ~ is nonsingular with respect to A
i.e. qp ~
o.
Proposi tion 3.3. If the probabili ty ~ on (G,E) is spread out, then r (G) .j..
a
n
if n -+ 00.
Proof. By proposition 3.2 we have
r 1 (G) ~ (r
*
r) (G)n+ n
Hence there exists a number g ~ 0 such that r (G) .j.. g if n -+ 00.
n
Since ~ is spread out there exists an integer p such that r (G) = s < 1. From
p proposition 3.1 we conclude for all k
Hence g
o.
k
s
o
In the sequel we shall use the following well known fact (see e.g. [4J, ch.
3, § 6.1).
Proposition~. Given are two functions f and g on G such that f E £1 (A)
and g E £oo(A). The function ¢(x) =
J
f(x + y)g(y)A(dy) is a continuousfunc-tion on G.
A random walk P on (G,E) with law p is called recurrent.inthe sense of
Harris (or shortly Harris) if for every starting point x E G and every A E E
i.e. for every A E E with A(A) > 0 we have PAl (x)
=
1 for all x E G (cf. [5J, ch. 3, de£. 2.8).The random walk is said to be sp~e~dou~ if the law p is spread out.
Theorem3~1. A random walk P with lawp on a connected group G is Harris if and only if the random walk is recurrent and spread-out.
Proof. First suppose the random walk is Harris.
It follows from theorem 1.1 that the random walk is conservative and by theorem 2.2 recurrent.
Suppose the random walk is not spread out. Then there exists a set A E
E
*n
with A(A)
=
0 and p (A)=
1 for all n. Then we haveP A,l(O) = 00 00
L
n=OJ
*n .lA' (y)p (dy) = 0 •
Since A(A') > 0 we conclude that the random walk is not Harris. This contra-diction proves that the random walk is spread but.
Conversely, suppose the random walk is recurrent and spread out. Take AE E wi th A(A) > O.
Since PPAl
=
PIAPAl + PIA,PAl
=
PAl - PIA(l - PAl) there exists a functionM+ n
g E such that P PAl
+
g if n + 00. Because of the a-additivity of P we havePg = P (lim pnpA1) n+oo g • Hence Pg
=
g. We now conclude g (x + y) r (dy) nf
g(x+y)r n (dy) g(x+y)qn(dy) +f
dq n g (x + y) dA (y) dy +=
png (x)=
f
g (x + y) p*n(dy)f
J
g(x) Since g ~ 1 we haveJ
dqo
~ g(x) - . g(x+Y)dAn(y)dY ~ rn(G) •J
dq
g(x) = lim g(x+Y)dAn(Y)dY, n-700
where the convergence is uniform. Using proposition 3.4 we get g is a conti-nuous function.
Since PP
A1 ~ PAl and the random walk is conservative i t follows from theorem 1.2 that g PAl A-a.e. By proposition 1.3 we get that g takes values one or zero A-a.e. Since g is continuous g takes values one or zero only. Since G is connected g is identically one or zero. By theorem 1. 1 we have g = 11.-a.e. on A and therefore g = 1 everywhere on G.
We now have 1 = g <- P 1A is Harris.
~ 1. Hence PAl = 1 everywhere on G and therefore P
o
Remark: Some condition like the connectedness of G in theorem 3.1 is neces-sary, which can be illustrated by the following example. The identity on~ is a recurrent random walk, which is spread out but not Harris.
A random walk P on (G,L) is said to be ergodic if for every A E L with
A(A) > 0 the probability that the random walk will visit the set A is posi-tive for A-almost all starting points i.e. PAl> 0 A-a.e. for all A E L with
A(A) > O. Usually ergodicity is defined by the fact, that every invariant set or its complement is empty mod the measure under consideration. It is easily verified that our definition is equivalent with the latter definition.
Theorem 3.2. A random walk P on (G,L) with law p is ergodic if and only if
the random walk will visit every non-empty open set A with positive proba-\'00 *n
bility when started in the point zero i.e. L
n=l p (A) > 0 for all nonempty open sets A.
*n
Prpof. First suppose there exists a nonempty open set A, such that p (A) = 0 for all n ~ 1. Since
00 00
L
n=lI
*n 1 A(y)p (dy) = 00L
n=l p*n(A)o ,
we have PAlCO) = O.Choose E > 0 such that n
I
I
A n (A +T)) has a nonempty T) <Ebe a nonempty open subset of n
I
I
A n (A +T)) • T) <e:For all x E U (0) we have by proposition 2.1
e:
P l(x) = P T l(x) = T P l(x) = P 1(0) ~ PAl (0)
B B -x -x B-x B-x
we conclude, that P is not ergodic, since A(B) > 0 and P
Bl = 0
Conversely, suppose for every nonempty open set A we have I:=l
Let A E E with A(A) > O. In order to show that P is ergodic i t
o
suffices to
show that for all FEE with 0 < A(F) < 00 there exists an integer n such
n *n
tha t P 1A (x) = p (A - x) > 0 on a subset of F wi th posi ti ve Haar measure
Le.
Using Fubini's theorem we get
f
IF(x)
(J
lA(x+y)p*n(dY»A(dx)J (J
IF(x)lA(X+ Y)A(dX))p*n(dY)
Since I
F E £1 (A) and lA E £oo(A) i t follows from proposition 3.4 that
ep(y) =
T
IF(x)lA(x+ Y)A(dx) is a continuous function. From
J
ep(y)A(dy) = A(F)A(A) > 0 we conclude that {z E GI
ep(z) > O} is anon-J
*nempty open set. Hence I
F(x) p (A - x) A(dx) > 0 for some integer nand
therefore P is ergodic. 0
If a random walk P is conservative and ergodic, then i t follows from
propo-sition 1.3 that for every A E E with A(A) > 0 we have PAl = 1 A-a.e. Hence
Harris recurrency of a random walk is stronger than conservativity and er-godocity. We end this section with an example of a random walk, which is conservative and ergodic but not Harris.
Consider the random walk P on CIR,E,A) with law p, determined by
1
p({n}) = 1 + n and p ({-1 }) = ~l~+~n~n~
Since n is irrational the set A = {nn - m n E:IN, m E:IN} is dense in]R and
i t follows from theorem 3.2 that P is ergodic. Since
J
xp(dx) = 0 i t followsfrom theorem 4.5 that P is conservative. On the other hand, A(A) = 0 and *n
p (A) = 1 for all n and therefore the random walk is not spread out and by
4. Recurrence criteria for random walks onmn
In this section we shall consider random walks on ORn,L,A) with n ~ 1, where L is the a-algebra of the Borel sets ofmn and A the Lebesgue measure on
ORn,L). The theorems and propositions of this section are slight modifica-tions of corresponding theorems in [8J. The techniques used in this section for random walks onmn can also be used for random walks on locally compact groups. We start with some notations. For all x and y inmn we write
(x,y) n
I
XiYi i=lI
xI
= I(x,x) d(x,y) = Ix - ylProposition 4.1. Let P be a dissipative random walk on ORn,L,A), then for
I
oo n nevery bounded set A E L we have 0 P 1 is bounded onm •
n= A
Proof. Put U = {x
I
Ixl < 1}. Because of the corollary of theorem 2.2 there exists a point p E U such that Pul (p) s < 1. Choose e: > 0 such that U2e:(p) c U. For every q E U (p) we have by proposition 2.1
e: Pu ( )T 1(q) p p-q e: Hence P 1(p) :5 U2e: s < 1 • s .1u (p) • e: We shall prove pn 1 Ue:(p) n n-l Suppose Pu (p)l :5 s e: n-l
:5 s for n ~ 1. For n = 1 the assertion is true.
for some n > 1. Then
pn+1 1 U (p) e: n n-l We conclude P ue:( )p 1 :5 s for n ~ 1. Hence 00
I
P~
(p)l :5 1~
s n=l E n onm 1 n :5 1 on m • By propos i-- s for all a E JRn.From proposition 1.2 we conclude
I
oo=l pn1 ()00 n n 1 Ue: Pn
tion 2. 1 we have \' P 1 < .. .. on "TO
Ln=l Ue:(p)+a - 1 - s ~
Since A can be covered by a finite number of translates of the hypersphere U (p) the proposition has been proved.
proposition.4.2. every open set A
n
Let P be a conservative random walk on OR ,I, X), then for \'00 n
we have Ln=O P l A = 00 on A. Proof. From theorem 2.2 we conclude PAl have pnl = 1 on A. Then
A
1 on A. Suppose for some n > 1 we
n PAl
=
1 on \'00 n Ln=l P l A We conclude 1.2 we haveA for n
~
1. Hence I:=lP~l
=
=
00 on A.00 on A. By proposition
o
We now introduce a class T of functions on JRn. A function f on JRn belongs to T if the following conditions are satisfied.
1) f is a nonnegative continuous even function on En such that f(O) > O. 2) There exists a nonnegative nonincreasing function ~ on the nonnegative
real numbers such that
i) f(x) $;~(lxP forallXEJRn ,
ii)
1
~(lxI>X(dx)
< 00 • JRn
Theorem 4.1. Let P be a random walk on OR ,E,X). The random walk is conser-vative if and only if there exists a function f E T such that
00
I
n=O
00 •
Proof. Suppose the random walk is conservative. Let A be an open set around zero such that f ~ ~f(O)lA. Then we have
00
~f(O)
I
pnl A(0) n=OFrom proposition 4.2 we conclude \'00 0 pnl (0) = 00 and therefore \'00 0 pnf (0) =00.
Ln
=
A Ln=
\'00 n
Conversely, suppose Ln=O P f(O) = 00 for some f E T. Let $ be a function as-sociated with f as described in the definition of T. Define h(x)
=
$(Ixl) for all x E JRn, then we have I~=o P~(O) = 00. Putn
K = {x E JR
I
0 $; xi < 1 for i=
1,2, ••• ,n} • There exists a countable partition of JR~K1,K2, •••} such that Ki is a trans-late of K for all i. Define the function k on JRn
k (x)
I
sup h (Y) 1 K. (x) i=1 YEK i ~ n for all x c EIt is easily verified that k (x)
s
1jJ(I
xI -
In)
forI
xI
~In
and thereforef
k(x)A(dx) < 00. Since h(x) s k(x) we haveL:=O
P~(O) = 00. Suppose the random walk is dissipative, then we conclude from proposition 4.1 and 2.1 that there exists a number w such that00
L
pn1K s w for all i ElN.
n=O ~
We now get
00 00 00 00 00
1: p~(x) 1: pn
L
I
hey)L
n
sup hey) 1K. (x) sup p l
K. (x)
n=O n=O i=1 YEK. ~ i=1 YEK
i n=O ~ ~ 00 s w 1: sup hey) i=1 YEK i w
f
k(x)A(dx) < 00 •Hence 1::=0 P~(O) < 00. This contradiction proves, that the random walk is
conservative.
o
Proposition4~3. Let P be a random walk on ORn,E,A) and define the function
n f on E by 2 n sin x. f (x) = 11 ~)n for all x E lRn 2
.
i=1 x. ~The random walk P is cons erva ti ve if and only if
Proof. It follows from theorem 4.1 that i t suffices to show f E T. Since . 2 sin2x
for all x E lR we have s~n x + 2 s 2 we have
x Hence . 2 s~nx s 2 x 2 2 1+x for all x EE . f(x)
=
n II i=1 n II i=1 (. 1 )n 2 1 +x. ~The function f satisfies the two conditions in the definition of the class
o
for r ~ 0 . ~(r)
T with the function ~ given by
2
2n
Let p be a probability on ([Rn,L;), then the characteristic fu.nction p of p is defined by
for all t E: IRn •
-For every function f E £1(A) the Fourier tran~form f of f is defined by
f (t)
f
f(x)e-i(x,t)A(dx) for all t E: IRn •Proposition 4.4. Let P be a random walk onIRn with law p and let f E: T such that
f
If(t) IA(dt) < 00 then00
. 1. lim
f
f
(t)Re 1 _r~
(-t) dt •(2'IT)n rtl
Proof. Suppose f E T and
f
If(t) IA(dt) < 00 We have/"....
Pf(t)
f
e-i(x,t)Pf(x)A(dx)J
e-i(X,t)(f f(x+y)p(dy))A(dx)f
ei(y,t)(f
f(x+y)e-i(x+y,t)A(dx))p(dy)=
J
ei(y,t)f(t)p(dy) Hence~
=
f(p)n for n~
O.Choose 0 < r < 1. Then we have
f
e-i(x,t) n=O!
r~nf(x)A(dx)00
I
n=O
00
I
nn(The series r P f(x) converges uniformly and is integrable.)
~
f Hence L O r P fn= = 1- rp~.Since f is a bounded continuous function, i t follows from the dominated con-vergence theorem, that pnf is continuous for all n. Hence
I:=o
r~nf
is a continuous function such that its Fourier transform is absolutely integra-ble. From a well known inversion theorem (see e.q. [4J, p. 2) we concludeI
r~nf(x)
1I
ei(x,t)f(t) .l~
dtn=O = (21T)n 1 - rp (t) for all x
E lRn •
Since f is an even function the function f is real valued and we get
00
I
n=O pnf(0) = :.-...;;.;1... lim (21T) n r t lf
-
1 f (t) Re 1 _ rp (t) dt •o
Proposition 4.5. Let p be a probability on ORn,~) with characteristic func-tion p. If for some t E:IRn we have p(t) = 1, then
p
is periodic with period t. Proof. Suppose p(t) = 1. Then we havep(t)
J
cos(x,t)p(dx) + if
sin(x,t)p(dx) = 1Hence cos(x,t) = 1 and sin(x,t) = 0 for p-almost all x ElRn and therefore
i(x,~ n
e = 1 for p-almost all x E lR • Hence
p
(x) •o
Proposi tion 4.6. Let p be a probability on ORn I~) with characteristic func-tion p. If there exists a number w > 0 such that
lim
I
Re 1 _ :, ( ) dt 1= 00rt1 rp t
I
tl~wthen for all a. > 0 we have
lim
f
Re 1 _1~
( ) dt 1= 00 •rt1 rp t
Itl~a.
Proof. Suppose there exists a number w > 0 such that
lim
I
rt1
Itl~w
and let a. > O.1
There exists a sequence r t 1 (n + 00) such that n lim
I
Re ..1~
.... dt=
t < 00 • 1 - r p (t) n~ Itl~w n 1 limI
rtl B~
a. and x~ u
kj
=
1 U (a.)}. Sincew J B is compact and< 1 on B we conclude from the dominated convergence theorem that
Ipi
compact and there exists k
that A cu. 1 U (a.).
J= w J
{x E: En
I I
xI
Put A
=
{x E: EnI I
xI
~ a. andp
(x)=
1}. Since p is continuous the set A is a finite number of points a1,a2, ••. ,ak in A such Put B
Using proposition 4.5 we conclude lim sup rtl
f
Rel~;P(t)
dt~
kt + s < 00I
tI
~a. and therefore lim rtlf
Re l-;P(t) dt-F
00 • Itl~a. IJTheorem 4.2. Let P be a random walk on (IRn,E,A) with law p. The random walk is conservative if and only if there exists a number a. > Osuch that
lim
f
Re rtl \tI
~a.
1 - rp(t) dt 00 • Proof. Let f be Put Q {x E: En ~ n n l Q(t) 2 Hi =l rier transformsthe function as defined in proposition 4.3.
I
Ixil ~ 1 for i=
1,2, ..• ,n} then the Fourier transform sin x./x .. It follows from the convolution theorem forFou-~ ~
that
~ 2
1*2n
=
4- n fQ
and from a well known inversion theorem ([4J, p. 2) that
1*2n(x)
=
4-n2.
f
f(t)ei(x,t)dt •Q (21T)n
2
Hence f(t)
=
4n (21T)n 1*2n(_t) and therefore f~
0 and has compact support.Q
Hence there exists a number w > 0 such that f(t)
=
0 for all ItI
> wand we conclude from proposition 4.400
f
I
pnf(0)=
1 lim f (t) Re 1 dt • rp(t) n=O (21T) n rtl 1 -ItlswFirst suppose there exists a number a. > 0 such that
lim
f
Re . 1 dtrp(t) 00 ,
rtl 1
-Itl~a
then i t follows from proposition 4.6 that for all S > 0 we have
lim r Re 1
-~P(t)
rtlIti~S
dt 00
Since f(O)
#
0 there exists a number v with 0 < v ~ wand f(t) ~ ~f(O) for all It I ~ v. We then have00
- 1
f
1~ ~f(O) . .. Re ...,;., . . dt=oo.
(2 )n 1 - rp (t)
1T
I
tI
svFrom proposition 4.3 we conclude the random walk is conservative. Converse-ly, suppose the random walk is conservative. Then by proposition 4.3 we have
I:=o
pnf(O) = 00. Put M = max f(t), then we getItl~w Hence 00 00 =
I
n=O pnf (0) ~ ~M_ lim inf (21T)n rtlf
1 Re ~l--~rp~(t-)- dt • Itl~w limJ
Re 1 dt 0 rp(t) 00 rtl 1 -Itl~w Corollary. Iff
Re 1 _ 1 dtP
(t) = 00 Itl~a.for some a. > 0, then the random walk with law p is conservative.
Proof. From Fatou's lemma we conclude
Hence
J
Re 1_l~(t)
dt~
I
tI
~a. lim inf rtlI
Re "7"l...._-'-'-~..,.;~r.(-t":'")
dt • Itl~a.lim rtl
I
R 1 e 1 - r~(t) Itl::;a dt (J()and therefore the random walk is conservative.
o
A probability p on ORn,~) is said to be ~educible if there exists a linear
n
subspace L ofm such that dim L < nand p(L) = 1. A probability which is
not reducible is called irreducible. A random walk with law p is called ~
reducible if its law p is irreducible.
Proposi tion4.7. Let p be a probability on ORn,~) and define for all a ;::: 0
n the function </> on m by a </> (x) a
I
(x,y) 2p (dy) IYI::;a for all x E mn •The probability p is irreducible if and only if there exists a number a > 0
such that </>a > 0 on {x I Ixl = 1}.
Proof. First suppose p is reducible, then there exists a linear subspace L
of mn with dim L < nand p(L) = 1. Let y E mn with Iy
I
1 and y orthogonalto L. Then </>a(y) = 0 for all a > O. Conversely, suppose p is irreducible.
For all n EIN put Wn = {x E mn
I
Ixl = 1 and </>n(x) = A}. If n:=l Wnf:.
~,then there exists a point y E mn with Iyl = 1 and
f
(z,y)2p (dy) = O. Itfol-lows p({x
I
(x,y) = O}) = 1 and hence p is reducible. This contradiction(W ) n
o
</> > 0 on {xI I
xI
= 1}. m W ncontinuous the set W is compact and hence
n
~. Since the sequence
(J()
proves nn=1 Wn =~. Since ~~n is
m
there exists a number m EIN such that nn=l
is nonincreasing we conclude W =~. Hence
m
Proposi tion 4.8. If P is an irreducible probability on ORn,E) with
characte-ristic function
p,
then there exist positive numbers Aand ~ with~
I
2Re(l - p(t» ;::: A tl for It
I ::;
~•
Proof. Since p is irreducible i t follows from proposition 4.7 that there
exist positive numbers a and cr such that
for Iy
I
= 1 •2
(x,y) p(dx) ;::: cr
J
Ixl::;a
Isin
~(x,t)
I
~
J
(x,t)I
1T i f Ixl
~ l~1
.
Choose t E lRn such that 0 < It I ~ 1T We now have
ct Re (1- p(t»
=
2f
sin2~(x,t)P(dx) ~
2r
sin2~(x/t)p(dx)
Ixl$l~r
Hence~
1
f
(x,t>2p (dX);l
f
(x,t)2p (dx) 1T 1 T .Ixl~~
Ixl~ct
21t l2f
(
t)2 (d )~
20 ltl 2 • =7
X'N
p x 1T Ixl$ct ~ 20I 12 Re(1 - P (t» ~ -"- t 1T for It Io
nTheorem 4.3. If n ~ 3 , then every irreducible random walk on
OR
,E,A) isdissipative.
Proof. Let P be a random walk with law p. We have for 0 < r < 1 1 ~e(1-
rp
(t) ) Re(1 ~ rp (t) ) Re = $ 1-
rp (t) 2 ~ 2 11 - rp(t) I (Re (1 - rp (t) ) .1 1 Re(1-
rp (t) ) ~ Re (1 - p(t) ).
f
dt~ltl2
• Itl~J.l~
Altl2 for all ItI
$ ].1. HenceJ
1 1 dt $ Re(1 - p(t» dt~'r
Itl$J.l 0<J
Re . 1. .. - 1 - rp(t) \tl$J.lFrom proposition 4.8 we conclude, that there exist two positive numbers A and ].1 such that Re(1 - p(t»
Since we conclude lim rtl 1 dt
t:
00 - rp(t)and therefore by theorem 4.2 the random walk is dissipative.
o
Because of theorem 4.3 the situation with respect to conservativity of ran-dom walks onlRn if n ~ 3 is quite simple. Things are less simple for
ran-2
Proposition 4.9. Let p be a probability on ORn,E) with
J
jxlp(dx) < 00, thenProof.
~~
Re (lit'
(t» = 0 •Re(l-p(t»
=
2f
sin2~(x,t)p(dx)
S;21tlf
Ixllsin2~(x,t)lp(dX)
• (x,t).
ISin2~(x
t)I .
S~nce '(x,t)' , ~s bounded and tends to zero if t ~ 0 we conclude from the dominated convergence theorem that
o
Proposition 4.10. Let p be a probability on ORn,E) such that
J
Ixlp(dx) < 00 and cr2J
Ixj2p(dx) < 00, theni) Re(l - p(t» S;
~cr21t12
,ii) if
f
xp(dx)=
0, then lIm p(t)I
S;~cr21t12
•Proof. Re(l-p(t» =2
J
sin2~(x,t)p(dX) s;~ltI2
J
jxI2p(dx) =~cr21t12
which proves i). sin(x,t)=
(x,t)-~(x,t)2sin(e(x,t»
where e depends on(x,t) and O,<e < 1. Hence
Im p(t) =
J
sin(x,t)p(dx)=
f
(x,t)p(dx) -~
J
(x,t)2sin (e(x,t»P(dx). FromJ
xp(dx) = 0 we concludeJ
(x,t)p(dx)=
(t,J xp(dx»=
0 for all tEIRn•Hence IIm
p
(t)I
S;~i
It 12• 0Proposi tion 4.,,1 L Let p be a probab ili ty on OR, E) such that
J
IxI
p (dx) < 00. If we put ~=
J
xp(dx) thenlim
~mp(t)
=~
•t~ t
Proof. rm p(t)
=
I
sin(x,t)p(dx)=
tf
xSi~txt
p(dx) From the dominated convergence theorem we concludelim
I~P(t)
=
~
.
t~ t
Proposition 4.12. Let p be a probability on CIR,E) such that
J
\xlp(dx) < 00 then1
I
Re(1 -1-
pet)~ t2 dt < 00 • Proof. Define 1ep(x)
I
1 - cos .xt dt for all x E lR •2 0 t Then 1 x2n+2t2n
f
00 <I>(x)L
(-1) n dt • (2n + 2) ~ 0 n=OThe uniform convergence of the series permits interchanging of summation and integration and we get
00 <I> (x) =
I
n=O 2n+2 (-1)n x (2n+2) ~(2n+1)and therefore </>" (x) = sin x
x
I
<1>' (x)I
~1
and <I>(x) ~I
x 1f
Re (1 -1J
OO sin t dt 1T• Since <1>' (0) = 0 and 0 t =
2
we have for all x E lR. Hence1 00
J
:i
f
(1 - cos xt) p (dx)-1 -00
00
f
<I> (x)p(dx)~If
Ixlp(dx) < 0 0 .D
2 Theorem 4.4. Let P be an irreducible random walk with law p onlR orlR such that
J
Ixlp(dx) < 00 and 02=
J
IxI 2p(dx) < 00. If~
=J
xp(dx) = 0, then the random walk is conservative.Proof. From proposition 4.8 we conclude that there exist positive numbers
Aand p such that
if
I
tl ~ p •Since ~
o
we conclude from proposition 4.10~41~14
(Re(1 - p(t») 2
~ ~
i:lt.L.
4 4~ 2 0 t
Hence if It
I
~ p we have Re(1 ~p
(t) ) ~ -,-2_A--:-~ 2 ~ 2 ~4ItI2' (Re(l-p(t») + (1m p(t» v 1 Re 1 - p(t) = Thereforef
Re~~1,...,.-
dt 1 -P
(t) Itl~p>.?2.
f
~='"
- 4 2 ' aItl~p
ItI
From the corollary of theorem 4.2 we conclude, that P is conservative. 0
Theorem 4.5. Let P be a random walk onJR with law p such that
J
Ixlp(dx) <"'. The random walk P is conservative if and only if ~=
f
xp(dx)=
o.
Proof. First suppose ~
=
O. Given any E > 0 i t follows from the propositions 4.9 and 4.11 that there exists a number a > 0 such that 11m p(t)I
~ Eltl and IRe(l - p(t»I
~ Eltl if It I ~ a. For 0 < r < 1 we have1 1 - r
Re ~
~--""'""""'-"""""_::""""--':=--~-_"""':"'--'-'-'--:-1 - rp(t) 2 2 ~ 2"
(Re(1 - rp (t) ) ) + r (1m p (t) )
It is easily verified, that for 0 < r < 1 and 0 < a < 1 we have
2 2 2 2
(1 - raj ~ 2(1 - r) + 2r (1 - a)
Therefore we can choose a small enough such that
(Re(1 - rp (t) ) )~ 2 ~ 2 (1 - r) 2 + 2r (Re (1 - P t»)2 ~(2 for It
I
~ a • We now have a af
Re 1 dtJ
(l ,.. r)dt 1 - rp(t) ~ ~ 2 ~ 2 (Re(1-
rp (t) ) ) + (1m p(t» -a -a a_._._.
a 1-r (l - r)J
dt 1J
dt ~ > -2 3r2E2t2 - 3 1+~2t2.
2(1-r) + -a-
_._.-
a Hence 1-r a limJ
Re 1 dt ~ L rtl 1 - rp(t) 3E -a and therefore a limJ
Re 1 dt rp(t) '" rtl 1 --aand by theorem 4.2 P is conservative. conversely, suppose ~ ~
a
1 Re 1 - r~(t) = Re(l :-P
(t) + (1 :- r) Rep
(t) .. ~ 2 2 ~ 2 (Re(l - rp (t) ) ) + r (1m p (t) ) :;; Re(1 .:-P(
t)) + --,-,-_l,--,--~r_~ r2(lm p(t»2 11-rp(t) 12 By proposition 4.11 we can choose a >a
such thatI
1mp
(t) I~
I~2t
1 for It1 :;; a . Hence -a af
Re 1 dt :;; 4 2. 1 - rp(t) ~ dt •By proposition 4.12 the first integral on the right hand side is finite. We are now going to estimate the second integral.
1 - r dt \1 - rp(t) 1 2 t :-r dt 2 ~ 2 r (1m p (t) )
J
1-r:;;lt!:;;aJ
dt2= 2+8(l :-r)2 2(..J..._1)
1-r a 1-r:;;lt l:;;a t r ~ dt ~ + 1 - rf
~dt+1 - rJ
It l:;;l-r \l-rp(t) 1 2 1-r:;;!tl:;;af
It!:;;l-r :;; 2 + 4(1 - r ) 2 2 r ~ aJ
1 -r -~~--::- dt \1 - rp(t) 1 2 -a Hence a limJ
Re 1 _~P(t)
dt~
00 rtl -aand therefore by theorem 4.2 the random walk is dissipative.
o
References
[lJ Breiman, L.: Probability. Reading, Massachusets: Addison-Wesley 1968. [2J Chung, K.L. and W.H.J. Fuchs: On the distribution of sums of random
va-riables. Mem. A.M.S. nr. 6, 1-12 (1951).
[3J Foguel, S.R.: The ergodic theory of Markov processes. New York: Van Nos-trand 1969.
[4J Reiter, H.: Classical Harmonic Analysis and locally compact groups, Oxford University Press, 1968.
[5J Revuz, D.: Markov Chains, North-Holland Publishing Company, 1975.
[6J Simons, F.H~ and D.A. averdijk: A recurrent dissipative Markov process. Internal publication
casaR
77-02. Eindhoven University of Techno-logy, 1977.[7J Simons, F.H. and D.A. averdijk: Recurrent and sweep-out sets for Markov processes. Mh. Math. 86, 305-326 (1979).