82 E. B.

Dynkin
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

(/. i*) = S / (y) i* (»).

If (/, /x) < oo9 then we say that / i s [i-integrable, and also that M is/-finite.
A non-negative function/is called excesszve if P / < / * and harmonic if
/(x) < °° for all x and P/ = /. Similarly, a measure is called excessive if
JUP < n, and harmonic if /i(x) < °° for all x and JUP = /*.
One of the central problems before us is the description of all harmonic and
excessive functions and measures connected with the transition function
p(x, y). It is expedient here to consider only 7-integrable functions h and
/-finite measures v, where 7 and / are, respectively, a previously selected
reference measure and function on E. Of fundamental interest is the case when
p(x, y) is transient (see the definition in §3). In this case we are able to attach
to each point y G £ a 7-integrable excessive function ky and a /-finite excessive
measure Ky . By means of the kernels ky (x) and Ky (x) we construct two com-
pactifications E* and E of E, such that ky (x) is extended for each x G E con-
tinuously to E*, and KV(X) is extended continuously to E. From the sets
E*\E and E \E we single out Borel sets B and B, where ky is a harmonic
function and (ky, 7) = 1, for y G B, and Ky is a harmonic measure and
(/, Ky)= 1 f o r ^ G i .
It can be proved that every 7-integrable excessive function h is representable
uniquely in the form

and any /-finite excessive measure v is expressible uniquely in the form

Here \xh and /xy are finite measures which are determined uniquely by h and v,
Copyright @ 1982. Cambridge University Press.

respectively. We call them spectral measures.

The set B is called the exit space, and the set B the entrance space. The
origin of these terms becomes clear in the following section.
§2. Markov processes

Suppose that a particle moves in a space E, going through a sequence of

states a0, alf a2, . . . The path aoaYa2 • . . may be terminate or may continue
unboundedly. The set of all (terminating or non-terminating) paths is denoted
by £2. The set of all non-terminating paths is denoted by £2^ .
By / < g we mean that f(x) < g(x) for all x e E.

EBSCO : eBook Academic Collection (EBSCOhost) - printed on 11/14/2018 11:03 PM via UNIVERSIDAD
NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 552478 ; Dynkin, E. B..; Markov Processes and Related Problems of Analysis
Account: ns145102.main.ehost
 Boundary Theory of Markov Processes {The Discrete Case) 83
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

Among the subsets of SI the so-called simple sets play a special role. A simple
set [a0a1a2 . . .an] is composed of all paths beginning with the states
a0, a i, . . ., an and continuing in any manner after the moment n. We denote by
3* the a-algebra in SI generated by all simple sets.
We depend on the following theorem on measures in SI.
T H E O R E M A . Suppose that for any n and any ao,al9 . . .,an EE a non-
negative number p(a0, ax, . . .,an) is given, where

2 p (a0, a l t . . ., an) < p (aOf a lf . . ., a,t-i)- (3)

Then there exists a measure P, which is moreover unique, on the a-algebra &
such that
P \aQax . . . an] - p (a0, at a n ).

This theorem will be proved in the Appendix.1

From Theorem A it follows that for any measure v on E with v(x) < «> for
all x, there exists a measure Pv on 12, such that

P v [aoat . . . an] = v (a0) p (ao,at) . . . p (an-l9 an). (4)

An important role is played by the particular case when v is the unit measure
concentrated at the point x (when v{y) - b(x, y), where 8(x, y) = 1 if
x = y, 8(x, y) = 0 if x ^y). The corresponding measure in SI is denoted by Px ,
so that

Px laodt . . . an] - 6 (x, a0) p (aOf at) . . . p (a u _ lt a n ) t (4')

Px is the probability measure concentrated on the paths starting from x. We note

that for any v.
Pv= S v ( x ) P x

and that Pv (Sl)= v(E).

Copyright @ 1982. Cambridge University Press.

Each measure in the space of paths SI determines a random process.2 The

process determined by the measure Pv is called the Markov process with initial
distribution v and transition function p{x, y). The process corresponding to
the measure Px is called the Markov process with initial state x and transition
function p(x, y).
One of the basic results of boundary theory states that almost every non-

The necessity of the condition (3) is obvious, since [a0, ax,.. .,an] C [a0, ax, . . .,an_i ] for any an,
and since distinct [a0, ax,. . .,an] do not intersect.
2
If p(£l) = 1, then P(A) can be interpreted as the probability that the trajectories of motion belong to
A. In the general case P(A) may prove to be greater than 1 and even equal to «>.

EBSCO : eBook Academic Collection (EBSCOhost) - printed on 11/14/2018 11:03 PM via UNIVERSIDAD
NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 552478 ; Dynkin, E. B..; Markov Processes and Related Problems of Analysis
Account: ns145102.main.ehost
 84 E. B. Dynkin
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

terminating path tends to some point of the exit space B. The measure of the
set of paths for which this limit belongs to the Borel set T C B is

where ky is the harmonic function corresponding to the point y GB, and (JL1 is
the spectral measure of the excessive function 1.
To explain the role of the exit space, we have to introduce into the discussion
paths without beginning or end. These are functions at with range in E defined
for all integers t from - °° to + °°. The set of such paths is denoted by £2. We
use the term simple sets in 12 for the sets [am, am + x, . . .,an] nm consisting of
all paths passing at the moments m, m + 1, . . ., n through the points
a
m > am +1 > • • •' an • (Before the moment m and after the moment n they can
behave arbitrarily.) We denote by & the a-algebra in £2 generated by all simple
sets.
For the construction of measures in the space 12 we can use the following
modification of Theorem A.
THEOREM B. Suppose that for any integers mKn and any
a
m>am +1 > • • •><*/! E E * non-negative numberspnm(am, am + 1, . . .,an)are given,
where

S( . . , an-l)> (3')

) p ( + l a n)). (3")
am
Then there exists a unique measure P on the o-algebra $- such that

• • • 0/»]m — Pm (Orny «m+l» • • • i «n)-

The necessity of the conditions (3'), (3") is evident. For the proof of Theorem
B see the Appendix.
We suppose that the transition function p(x, y) satisfies (1) with the equality
Copyright @ 1982. Cambridge University Press.

sign and that v is a harmonic measure. Then the function

Pm ( f l m , < W i , . . . , an) ^ V (am) p (dm, am+i) . . . p (fln-i, dn)

satisfies the conditions (3')—(3"), and by Theorem B there exists a measure Pv

on & such that

Pv Kiflm+i • • • <*X = v (am) p (ami a m+1 ) . . . p (a n -,, an). (4")

A random process determined by Pv in 12 is called a stationary Markov process

with stationary distribution v and transition function p{x, y).

EBSCO : eBook Academic Collection (EBSCOhost) - printed on 11/14/2018 11:03 PM via UNIVERSIDAD
NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 552478 ; Dynkin, E. B..; Markov Processes and Related Problems of Analysis
Account: ns145102.main.ehost
 Boundary Theory of Markov Processes {The Discrete Case) 85
All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law.

In boundary theory it is proved that for such a process almost all paths
converge as t -> — °° to some point of the entrance space B. The measure of the
set of paths for which this limit belongs to a Borel set T C B is

f
where Ky is the harmonic measure corresponding to the point y E B and i? is
the spectral measure for v.
This result can be further generalized in several directions.
Random variables connected with Markov processes are ^"-measurable
functions defined on £2 or on a subset of this space (or £F -measurable functions
on £2 or a subset of £2). The integral of such a function £ over its domain
with respect to the measure ?v is denoted by Mv £, and with respect
toP^ X>yMx%.
Here are some examples.
f is the terminal moment of a path: if the last moment at which the path co
is defined is n, then f(co) = n; if the path does not terminate, then f(oo) = + °°.
xn is the position of a particle at the moment n. This function is defined on
the set ! or. f(co) >n\.\n the case of a stationary process

Pv \xn = y] = v (y)

for any n. For a process with the initial distribution v

Pv [Xn -"= y) = 2 Pv [*n-l "-= 2] P (2, y)- (5)

z
To prove this equation it suffices to note that
{co: xn =y} = U {co: xn_1 =z, xn =.y},to decompose the sets occurring here
into simple sets, and to use (4).
We put
p (ny x, y) = P x {xn = y).
From (5) it follows that
Copyright @ 1982. Cambridge University Press.

p(n, xy y) = S p ( « - l , x, z)p(z, y)y

z

and in view of the obvious relation

we have

Mxf(xn) ^ A / X V 6 (x,,, y) / (Xn) = 2 AT*« (^», if) / (if) -

(6)

EBSCO : eBook Academic Collection (EBSCOhost) - printed on 11/14/2018 11:03 PM via UNIVERSIDAD
NACIONAL ABIERTA Y A DISTANCIA - UNAD
AN: 552478 ; Dynkin, E. B..; Markov Processes and Related Problems of Analysis
Account: ns145102.main.ehost