lectures on random measuresboos/library/mimeo.archive/isms_1974_96… · abstract topological space...
TRANSCRIPT
LECTURES ON RANDOM MEASURES
by
Olav Kallenberg
Univepsity of ~tebopg, SWeden
and
Univepsity of Nopth Capolina at Chapel HillDepartment of Statistias
Institute of Statistics Mimeo Series No. 963
November, 1974
LECTURES ON RANDOM MEASURESby
Olav Kallenberg
PREFACE
Random measure theory is a new and rapidly growing branch of
probabi li ty of increasing interest both in theory and applications.
Loosely speaking, it is concerned with random quantities which can
only take non-negative values, such as e.g. the number of random variables
in a given sequence possessing a certain property, the time spent by a
random process in a certain region etc. In the special case when these
quantities are integer valued, our random measures reduce to point processes,
which constitute an important subclass of random measures. However, I am
convinced that general random measures are potentially just as important
from the point of view of applications. (See e.g. Kallenberg (1973c)
for some results supporting this opinion.)
The justification of random measure theory as a separate topic lies
in the overwhelming amount of important results which are specific to
random measures in the sense that they do not carryover to general random
processes. As a contrast, comparatively few results are specific in this
sense to the subclass of point processes. Indeed, it will he clear from
the development on the subsequent pages that a substantial part of point
process theory, as presented e.g. in the monograph of Kerstan, Matthes
and Mecke (1974), carries over to the class of general random measures.
..
The latter may therefore be considered as a natural scope of a theory.
These lectures contain research supported in part hy the Office of Naval eResearch under Contract NOOO1tl-67-I\-O~21-0002.
-2-
The origin of these notes is a series of lectures delivered at the
Statistics Department at Chapel Hill during the spring semester of 1974,
where I intended to collect the (in my opinion) central facts about random
measures from the scattered literature in the field. During the course of
my lectures, however, I often elaborated new ideas and approaches, and as
a result I found many improvements of known theorems and proofs, and also
a large number of new results. The present notes should therefore be
regarded both as a survey of known results and as an original contribution
to the research in the area.
Here follows a summary of the contents and an indication about
the nature of the most important novelties. Section 1 is introductory.
After having set up the general framework, we demonstrate how the basic
point processes may be defined by a simple mixing procedure. Here and
in subsequent sections, the Laplace transform plays the role of a powerful
general tool. In Section 2, the main purpose is to extend the classical
simplicity criteria for point processes into a general theory of atomic
properties for arbitrary random measures. Section 3 is devoted to the
question of uniqueness of random measure distributions. In particular,
it is shown that the distribution of a simple point process or diffuse
random measure ~ is uniquely determined by the set function ~t(B) =
= Ee-t~B for bounded Borel sets B, where t > 0 is arbitrary but fixed.
Note that, as t + 00, ~t(B) + P{~B = O} and our result turns formally
into a well-known one for point processes. In Section 4, we derive the
-3-
corresponding convergence criteria. A novelty here is a relationship
between the notions of convergence based on the weak and vague topologies
respectively. The idea of tightness leads in Section 5 to a simple
approach to the existence theorems of Harris and others.
Sections 6 and 7 are devoted to infinite dlivisibility and the limit
theorems for null-arrays. From a simple representation of an infinitely
divisible random measure l:;, we are able in Section 6 to draw interesting
conclusions about the relationship between the properties of l:;, itself
and its canonical spectral measure; in Section 7 we show among other things
how the above-mentioned uniqueness theorem in Section 3 leads in particular
cases to improved convergence criteria for null-arrays. In Section 8 on
compounding and thinning of point processes, the known results for the
case when the compounding variables f3 tend to zero in distributionn
are accompanied by an analogeous theory for the complementary case when
no subsequence tends to zero.
Symmetrically distributed random measures are examined in Section
9, and in particular we dwell on the case when the random measures are
at the same time infinitely divisible. The peak of our treatment is
the characterization of the class of canonical measures which can arise
in the corresponding spectral representations. After a general discussion
of Palm distributions in Section 10, containing improvements of known
uniqueness and continuity results, we finally enter in Section 11 into
the fascinating but difficult subject of characterizations by factorization
and invariance properties of the Palm distributions. As for the factorization
e
-4-
properties, we are able to give a very simple proof of the Kerstan
Kummer-Matthes theorem, while a modification of the original proof
yields a radically strengthened version of this result.
I hope that the technicalities of the above presentat,ion, mainly
intended for specialists, will not discourage newcomers to the field.
Great efforts have been made to provide a smooth approach to the subject,
and wherever I have relied on "known" facts, detailed references are given.
Still some requisites may prove helpful, including some basic knowledge of
measures on locally compact spaces (to be found in Bauer (1972)), and of
convergence of probability measures (as presented in Billingsley (1968),
Chapter 1).
Finally, I would like to express my thanks to my audience in Chapel
Hill, and in particular to Professor M. R. Leadbetter, for their interest
in this project and for several corrections and improvements of my original
draft. I am also grateful to the Institute of Statistics there for allowing
me to include these lecture notes in their Mimeo Series, and to Mrs. Melinda
Beck for her careful typing.
G~teborg in August 1974
O.K.
-5-
CONTENTS
1. Basic Concepts 6
2. Atomic Propert ies 18
3. Uniqueness ....................•..................... :;0
4. Convergence ..........•.......................•...... :;9
5. Existence 57
6. Null-Arrays and Infinite Divisibility ..........•.... 66
7. Further Results for Null-Arrays 87
8. Compound and Thinned Point Processes 99
9. Symmetrically Distrihuted Random Measures .....•.... 113
10. Palm Distributions 1:16
11. Characterizations Involving Palm Distributions 151
..
-6-
1. BASIC CONCEPTS
We could confine ourselves to random measures defined on the real
line R or on some Euklidean space kEN = {1,2, ... J , but it
turns out that we can choose a much more general framework without any
essential changes in notation or proofs. Hence we consider instead an
abstract topological space S which is assumed to be locally compact,
second countable and Hausdorff, (see e.g. Simmons (1963)). By choosing
a suitable metric p , we can convert S into a separable and complete
metric space. This possibility is often useful, but it is important to
realize that our definitions or results never depend on the actual choice
of p •
Let B be the Borel a-algebra in S, i.e. the a-algebra generated
by the class of open sets, and let B be the ring of all bounded B-sets.
(By definition, a set B is bounded whenever its closure B is compact.
In general, this is stronger than metric boundedness.) Write F = F(S)
for the class of !-measurable functions f: S + R = [0 00) and let+ "
F = F (S) denote its subclass of continuous functions with compact support.c c
To simplify statements, we introduce three types of subclasses of B.
By a DC-Bemiring (D for dissecting, C for covering) we mean a semi-ring
I c B (see e.g. Kolmogorov and Fomin (1970) for the definition) with the
property that, given any B E Band £ > 0 , there exists some finite
covering of B by I-sets of maximal diameter < £ in some fixed metric.
A DC-ring is a ring with the same property. Note that such classes always
exist and may be chosen countable. (We may e.g. take the ring generated
-7-
by some countable base in B.) On R, typical DC-semirings and -rings
consist of finite intervals and interval unions respectively, hence our
notation. Finally, we say that a class C c B is covering if any set
B E B can be covered by finitely many sets in C .
Note that the notions of DC-semiring and --ring are purely topological
and do not depend on the choice of metric p. In fact, suppose that p
and pI are two metrizations of S, and let B E: Band E: > 0 be
arbitrary.I
Choose a bounded open set G ~ B . (Note that any countable
base in B covers B, and that a finite subcDvering exists since B
is compact.) Since G is compact, the identity mapping of G onto itself
is uniformly continuous with respect to p' and p, and in particular we
may choose some E ' > 0 such that p(s,t) < E whenever s,t E: G with
pI (s, t) < E I . Thus any covering of B with p I -diameters <E" =
= E: I 1\ p (B, (;c) (Gc = S\G denoting the complement of (;) is included
in G and hence has p-diameters < E. It remains to verify that E" > 0
which follows easily from the facts that B is compact while G is open.
We now introduce the class M = M(S) of Borel measures ]J on (S, B)
which are locally finite, i.e. take finite values on B, and further the
subclass N = N(S) of Z -valued measures (Z = {O, 1, ... })+ +
Let M
and N be the smallest a-algebras in M and N respectively making
the mapping ]J +]JB of M or N into R measurable for each B E B+
Note that, by monotone convergence, the integral
I lIere and below, B denotes the closure of B.
-8-
is then automatically measurable for any f E F. More generally, assuming
f E F to be bounded with compact support and defining the measure f~ by
(f~)B = JBf(S)~(dS) , B E B , it is seen that the mapping ~ ~ f~ is
measurable. In particular, the restriction B~ = lB~ of ~ to B is
measurable for every B E B (As usual, lB denotes the indicator of
the set B, being equal to 1 on B and 0 elsewhere.)
LEMMA 1.1. The a-aZgebras M in M and N in N are both generated
by the mappings ~ ~ ~f , f E F ,and aZso by the mappings ~ ~ ~I ,c
I E r ~ for any fixed DC-semiring reB.
Proof. The same argument applies to both M and N , so let us e.g.
consider M Let M' be the a-algebra in M generated by the mappings
f E Fc By the definition of M, we have to prove that the
mapping ~ ~ ~B is M'-measurable for every B E B. If B is compact,
we may choose open bounded sets Gl ,G2, ... such that G + Bn By Urysohn's
lemma (cf. Simmons (1963)) there exist functions f l ,f2, ... E Fc such that
n EN, so for any ~ EM,
~B = ~lB ~ ~fn $ ~lG = ~Gn + ~B .n
Hence ~f + ~B , and the M'-measurability of ~B follows from that ofn
~f for n E Nn
We next consider an arbitrary fixed compact set C and define
v = {B E B: ~ ~ ~(B n C) is M'-measurable} .
-9-
Then V is clearly closed under monotone limits and proper differences
and it contains S. On the other hand, V contains the class C of
compact sets, which is closed under finite intersections. Hence by Dynkin's
2monotone class theorem (cf. Bauer (1974)), V contains the a-algebra a(C)
generated by C. But a(C) = B , and so Jl -+ Jl(B n C) is M'-measurable
for every B E B. Since C was an arbitrary compact set, this proves
that M' = M
The assertion involving I is proved by a similar argument, but
instead of f n we consider a finite covering of B by I-sets In1 , ... ,Ink
contained in Gn
disjoint, and then
Since I is a semiring, we may take the I .nJ
to be
JlB ~ Jl u Inj = I JlI . ~ JlG + JlB ,j j nJ n
proving that I· Jl 1 . -)- JlB This completes the proof.J nJ
LEMMA 1.2. N EM.
Proof. Let I c B be a countable DC-semiring, and let Jl be a fixed
measure in the set
M = {Jl E M: Jll E Z , I E I} .+
For any fixed finite union U of I-sets, we further define
V {B E B: Jl(B n U) E Z } .-I-
Then D lS closed under monotone limits and proper differences, and it
2ThoUI~h commonly ascribed to Oynkln, this theorem
Si erpinski, Fund. Math . ..!2 (l92R).
is actually due to
-10-
contains S since U may be written as a disjoint union of I-sets.
On the other hand, it is easily seen that V contains I, and I being
closed under finite intersections, it follows by Oykin's theorem that
V ~. a(I) ~ B and we obtain ~(B n U) E Z+ for every B E B. But for
given B E B , we may always choose U such that U 0 B , so we get
~B E Z for all B E B , proving that ~ EN. Hence MeN , and the+
converse relation being obvious, we have in fact N = M Since clearly
M EM, this completes the proof.
Let us now consider any fixed probability space (n,A,p). By a
random measure or point process ~ on S we mean any measurable mapping
of (n,A,p) into (M,M) or (N,N) respectively. By Lemma 1.2, a point
process may alternatively be defined as an N-valued random measure. Conversely,
any a.s. (almost surely w.r.t. P) N-valued random measure coincides with a
point process except on a P-null set (i.e. on an n-set of P-measure 0) .
For these reasons, we shall make no difference in the sequel between point
processes and a.s. N-valued random measures. Similarly, we shall allow a
random measure to take values outside M on a P-null set.
The class of random measures is clearly closed under addition and
multiplication by non-negative random variables. For a proof, it suffices
by the definition of M to observe that, if ~ and n are random measures
while a and S are R+-valued random variables, the quantity a~B + SnB
is a random variable for every B E B. Slightly less obvious is the
following
-11-
LEMMA 1.3. A series L- t;. of random measures on S ~s itself a randomJ J
measure iff L.~.B < 00 a.s. for eachJ J
B E B .
Proof. First note that, by monotone convergence, ~ = L·~·J Jis (everywhere
on Si) a-additive on B and is therefore measure valued. Next suppose
that C c B is a countable base in S
outside some fixed P-null set A E A
Then ~C < 00 for all C E C
and since C is covering, it follows
that ~B < 00 B E B , except on A Hence ~ E M a.s. It remains to
prove that ~ is measurable, and by the definition of M , this follows
from the elementary fact that
B E B .
~B = L.~.B is a random variable for everyJ J
The distribution of a random measure or point process F, is the
probability measure p~-l on (M,M) or (N,N) defined by
M E M or N.
We further define the intensity E~ of ~ by
(EOB = E(~B) B E B ,
where E denotes the "expectation" or integral w.r.t. P. By Fubini's
theorem, E~ is always a measure, though it need not be locally finite.
We finally define the L-transform
L~(f) = Ee-F,f
of (L for Laplace) by
f E F .
Note in particular that, for any kEN and Bl , .. , ,'\ E B, the function
-12-
LS[.r t.1B.] = E eXP[-.I t.~B.}J=1 J J J=1 J J
is the elementary L-transform of the random vector (sB1 , ... ,sBk)
One of the most important ways to define new random measures (or
rather their distributions) is by mixing. For a definition, let (8,T,Q)
be an arbitrary probability space and let {sS' S E 8} be a family of
random measures which mayor may not be defined on the same probability
space. If P{sS E M} is a T-measurab1e function of 8 for every M E M
we may then consider S as a random element in 8 with distribution Q
and form the mixture of with respect to Q , thus obtaining the
joint distribution R of the mixed random measure ~ and the mixing
element S. Formally, R is defined by the relation
is then a version of the conditional distribution of
(1 )
Note that
R(M x A) = JAP{~S E M} Q(dS)
Ps- 1S
M E M A E T .
given 8, and that
E'f(s,S) = E'E[f(~S'S) IS]
for any measurable function
for a complete discussion.)
f· M x 8 -+ R. + (See Lo~ve (1963) page 359
The measurability requirement above is not easy to verify. Here
is a cqnvenient criterion:
LEMMA 1.4. Let (8,T,Q) be an arbitrary probability space and let
{se ' 8 E 8} be a family of random measures on S. Then there exists
3 Here E' denotes expectation w.r.t. R .
-13-
a unique probability measure R on M x T satisfying (1) if Lt.: (f)e
is T-measurable in e for every fixed f E Fc
Proof. The necessity of our condition follows by monotone convergence
from simple functions. Conversely, suppose that Lt.: (f) is measurablee
in e for every f E Fc
Then
L~ (L t.f.)e j J J
= E exp(-L t.t.: f.)j J (I J
is measurable in e for any fixed keN
Now it follows by the entended Stone-Weierstrass
theorem (see Simmons (1963)) that any continuous function g(x l , ,xk)
on which vanishes at infinity can be uniformly approximated by linear
combinations of the functions eXP(-L/jXj ) with t l , ... ,tk E R+
Therefore Eg(!;efl' ... '!;efk) is measurable in e for any such g ,
and hence, by an easy application of Urysohn's lemma, it follows that
the probability
is measurable in e for any t l , ... ,tk E R+
Let us now define V as the class of all sets M c M such that
P{[,e c M} is measurable. Then V is closed under proper differences
and monotone limits, and it contains M. On the other hand, it was shown
above that V contains the class C of all sets
-14-
and since C is closed under finite intersections, it follows by Dynkin's
theorem that
f E Fc
But by Lemma 1.1, the right-hand side of this relation equals M, and
so P{~e E M} is indeed measurable in e for all M EM, as required.
We shall now introduce some important point processes and at the
same time illustrate the use of L-transforms and mixing.
For any s E 5 , let Os E N be defined by 0sB = lB(s) , B E B ,
and note that the mapping s -+ °Sis measurable B-+ M Hence is
a point process on 5 whenever , is a random element in (5,B) Writing
-1w = PT for the distribution of , , it is easily seen that 0, has
intensity -1E<\ = pT = W and L-transform
-f(T) -fE exp(-o,f) = Ee = we f E F .
Next suppose that n E Z+ and that T l' ... ,Tn
are independent random
elements in 5 with common distribution w E M Then any point processn
c;, which is distributed as L °T is called a sample process with intensityj=l j
nw. By independence, ~ has L-transform
-~fEe =n
-f nIT E exp(-o, f) = (we )
j =1 jf E F .
Here we may mix with respect to n = v , regarded as a random variable,
to obtain a mixed sample process with intensity Ev· wand L-transform
r - ~f E' ( -f) V ,I, ( -f)L:e =. we = If' we f E F ,
-15-
where
S E [0,1] ,
is the (probability) generating function of v
In the particular case when v is Poissonian with mean a, we get
1jJ (s) -a(l-s)= e S E [0,1] ,
and the L-transform of ~ becomes
(2)-~f f fEe = exp(- a(l - we- )) = exp(- 1..(1 - e- )) f E F ,
where we have put A = aw = Et;. This measure A is totally bounded, hut
we may also define point processes ~ satisfying (2) when A E M is arbitrary
unbounded. For this purpose, let Sl,52 , ... E B be a disjoint partition of 4ItS. Then the restrictions S.A of A to S. are totally bounded, and so
J J
there exist independent point processes ~1'~2' on S with L-transforms
(3) ( -f )E exp(-E;.f) = exp -(S.A) (1 - e )J J
f E F .
Since their sum (= LjSj satisfies
E~B = L E~.B = L (S.A)B. J . JJ J
= (L S.A)Bj J
AB < 00 B E B ,
it follows by Lemma 1. 3 that t, is a point process. Furthermore, it 15
seen from (3) and the independence of t,l'sZ' ... that, for any f ( F ,
(' -r,f:e· =-Sof
E If e Jj
-Cf= JT Ec J
J
cxp ( -L(S . A) (l. JJ
-16-
which shows that I; satisfies (2) . (The formal calculations here are easily
justified, using the fact that I;.f and (S.A) (1 -f non-negative.)- e ) areJ J
Any point process I; satisfying (2) will be called a Poisson process with
intensity A. If I; is a process of this type, it is easily seen from
(2) that, for any kEN and disjoint Bl , ... ,Bk E B , the random variables
I;B l , ... ,I;Bk are independent and Poissonian with means ABl , ... ,ABk .
Let us now consider a Poisson process I;a with intensity aA , where
a E R+ while ~ EM. By Lemma 1.4, we may then consider a as a random
variable and mix with respect to its distribution to obtain a mixed Poisson
process I; with L-transform
where
-E;f ( -f )Ee = E exp -aA(l - e ) f E F ,
Ht)-at= Ee t 2: 0 ,
is the L-transform of a. More generally, it is possible by Lemma 1.4
to consider A = n in (2) as a random measure, thus obtaining a Cox process
directed by n, possessing the L-transform
(4) f E F .
kLet us finally consider some fixed measure l.l E N , say l.l = I <s
j=l t j
where k E Z = Z u {oo} and t. E S , j ~ k Further suppose that+ + J
are independent and identically distributed R+-valued random
variables, say with LS
= ~ . Then it follows from Lemma 1.3 that
-17-
is a random measure on S , and we get for f E F
IT E exp(-S.f(t.)) = IT ~ 0 f(t.)j J J j J
~ = 1:.S.<\J J j
Ee -H = E exp(- 1: S.f(t.)) =. J JJ
= exp(I log ~ 0 f(t.)) =j J
exp(]..I log ~ 0 f)
By Lemma 1.4, it is permissible to mn with respect to ]..I
as a point process, thus obtaining
n when regarded
f E F .
Any random measure with this mixed distribution will be called a (j-compound
of n. In the particular case when S attains the values a and 1
only with probabilities 1 - P and p respectively, ~ will be called
a p-thinning of n. In this case, (5) takes the form
(6) f E F .
Scrutinizing the arguments used in the above constructions, it is
seen that they do not depend on the nature of S. Thus we may e.g. proceed
as above to construct Poisson processes with arbitrary a-finite intensities
on any measurable space (s,B) .
NOTES. !Iere and below, our main source is Kallenberg (1973a). However,
Lemma].4 is new and is essentially a strengthening of proposition 1.6.2
in Kerstan, Matthes and Mecke (1974). The present method on constructing
-18-
Poisson and Cox processes is simpler and more general than methods in
the current literature.
PROBLEMS.
1.1. By a o-ring ReB we mean a ring which is closed under non-decreasing
bounded limits. Show that, if 1 is a DC-semiring, then B is the
smallest a-ring generated by 1 In symbols: B = a(l) .
1.2. Show that Lemma 1.1 remains true for any semiring 1 c B with
a(I) = B.
1. 3. Let 1 c B be an arbitrary DC-semiring. Show that Lemma 1.4 remains
true with F replaced by the class of all functions of the form
L~=lc
LI I for arbitrary k E N , t l , ,tk E R andJ . +
JII' ,Ik E 1
1.4. Let ~ be a Cox process on S directed by n Show that ~
has independent increments, in the sense that ~Bl' ... ,~Bk are
independent for any kEN and disjoint Bl , ... ,Bk E B , iff
this is true for n Prove the corresponding fact for compound
point processes with 8 ~ O. (Hint: use analytic continuation.)
2. ATOMIC PROPERTIES
Let Md denote the class of diffuse (or non-atomic) measures in M.
Every measure ~ E M may clearly be written in the form
(1)k
~ = ~d + L b.o tj =1 J j
-19-
for some ~d E Md , k E Z+ = Z+ U {~} , bl , b2 , ... E R~ = (O,~) and
distinct t l , t 2 , ... E S , and this decomposition is clearly unique apart
from the order of terms. Moreover, 11 E N iff " - 0"d - and
When considering random measures, we often need a measu~able decomposition.
LEMMA 2.1. There exists a decomposition (1) of' every 11 E M such that
the quantities 11d E Md , k E Z+
~e measurable functions of 11
The measurable functions occurring in Lemn~ 2.1 are by no means unique,
not even apart from the order of terms. However, when considering random
measures ~ on S with ~S < ~ a.s. (which holds automatically when
S is compact), we may require the atom sizes 81 , 82, ... to be taken
in order 81 ~ 82 ~ ... and the corresponding atom positions Tl , T2' ...
to be chosen in random order within sets of e~lal 8j . (We shall always
assume the basic probability space (rl,A,P) to be rich enough to support
any such randomization.) Though the decomposition
(2)v
~ = ~d + Lj =1
8·0J T .J
is still not unique unless all the atom sizes are different with probability
1 , we obtain in this way a unique joint distribution of the random elements
occurring in (2).
Par the proof of Lemma 2.1, we shall need an auxiliary result, the
formulation and proof of which requires some further terminology and notation. _
-20-
denotes the diameterI. I
{B .} c B is a null-ar~y ofnJ
n E N form a finite disjoint
(Note that the last condition is
B E B , we shall say that
B
Given any set
par' ti tions of if the B. for fixednJ
partition of B and if max. 1 B 01 -+ 0 , whereJ nJ
in any fixed metrization p of S
independent of the choice of P.) For any p E M and £ > 0 , we define
*pEN and pi E M by£ £
* Lp B ::: I (p) B E B ,£ sEB {p{shd
p'B ::: pB - L lJ{s}l{lJ{s}~d (ll) B E B •£ sEB
LEMMA 2.2. Let II EM, £ > 0 and B E B , and let {B o} c B be anJ
null-array of partitions of B. Then
lim L I (p)J
o {lJB 0 ~e:}n-+<x> nJ*::: p Be:
PROOF. For sufficiently large n EN, all ll-atoms in B of size ~ £
will lie in different partitioning sets Bnj , and we get
L I (ll)o {pB .~e:}J nJ
*~ p Be:
Thus it remains to prove that
(2)
If (2) were false, there would exist some subsequence N' c N and some
numbers jn' n E N' , such that
-21-
(3) n E N' .
Choosing arbitrary n E N' , there exists by compactness
some further subsequence Nil c N' and some s E B such that sn -+ s ,
n E Nil , and sin¢e lB. I -+ 0 , it follows from (3) that ll'G ~ E fornJn £
every open set G E B containing s But then ll'{S} ~ e: , which con-E
tradicts the definition of II ' Hence (2) must be true, and the lemma£
is proved.
PROOF OF LEMMA 2.1. Define the set function ~ = ~(ll) by
(4) *= II B£
B E B E > 0 .
By the Caratheodory extension theorem, (4) defines for each II E M a
unique measure ~ on S x R: ' and it is easily seen that ~ is purely
atomic with all its atoms of unit size. Furthermore, it follows from
for some measurable
Lemmas 1.1 and 2.2 that ~
that the lemma is true for
is a measurable function ofk
~ , i.e. that ~ = I 0j =1 cr j
II Suppose
functions k E Z and cr l , °2' E S x R' Then (1) holds with+ +
cr. - (t . , b. ) , and since t. and b. are measurable functions of theJ J J J J
pair (t. , b. ) , the measurability in (1) will follow by Lemma 1.3. WeJ J
may therefore assume from now on that lld = 0 and bl :: b2 - - 1
Let s1' s2' ... be an arbitrary dense sequence in S and note
that p (s, sn) = p (t, sn) , n E N , implies s = t (To see this, let
N' c N be such that s -+ s n E N' , and note that thenn
-22-
p(S,t) 5 peS, s ) + pes, t) = 2 pes, s ) + 0n n n n € N' .)
We may therefore introduce a linear ordering of S by writing s < t
whenever, for some kEN,
pes, s.) = pet, s.)J J
j = 1, ... ,k - 1
For any fixed disjoint partition Bl , B2, '" E B of S, we now order
the atom positions t.J
of ~ , first according to their occurrance in
Bl , B2, ... , and then, within each Bn , according to their linear order.
We have to show that t l , t 2 , ... are measurable when ordered in
this way, and by induction it suffices to consider t l ' i.e. to prove
that the set {t l € B} is measurable for every fixed ~ E B. Now this
set is clearly expressible by means of countable set operations in terms
{p (c n r}) = k} cof the sets B. n {s: p(s, s. ) < with C = B or B
1 J
and with arbitrary i,j E N , kE Z+ and rational r > 0 Since
the latter sets are measurable, this completes the proof.
We shall now establish a simple relationship between the one-dimensional
distributions of a random measure ~ and the intensities of the corresponding
*point processes ~£'
o < a < b 5 00 •
* * *£ > 0 , or more generally of ~[a,b) = ~a - ~b '
THEOREM 2.3. Let ~ be a random measure on S, let B E B and let
{B .} c B be a null-array of partitions of B. Further suppose thatnJ
o < a < b s; 00 Then
(5 )
-23-
*E~[a,b)B = lim L P{~B . e [a,b)}n-+-<o j nJ
holds provided Et;'Ba< co , while in general
(6) 1iminf? P{t;Bnj e [a,b)} .n-+-<o J
PROOF. By taking differences in Lemma 2.2, we get
(7) lim? l{t;B .e[a b)}n-+-<o J nJ '
and hence, by Fatou's lemma,
1iminf E ? l{~B .era b)}n-+-<o J nJ '
= 1iminf L P{~B . E [a,b)} ,n~ j nJ
*proving (6). Since (6) implies (5) in the case E~[a,b) = co , it remains
*to prove (5) in the case when E~Ia,b)B and E~'B are both finite.a But
then (5) follows from (7) by dominated convergence, since we have
and hence
I} u {~'B .a nJ
~ a} ,
* \ -1~ L ~I b)B . + L a ~'Bh··j a, nJ j a J
~* B a- 1nB- ~[a,b) + sa
We next give a simple criterion ensuring t; to have no atoms of
size ~ £. Given any covering class C c B and any a > 0 we shall
-24-
say that ~ is a-regu~ar w.r.t. C, if there exists for every fixed
C E: C some array {C .} cC of finite coverings of CnJ (one for each
n E: N ), such that
(8) lim LP{~C . ~ a} = 0 .. nJn+oo J
THEOREM 2.4. Let ~ be a random measure on S, ~et C c B be a covering
cZass and ~et a > 0 *Then ~ = 0 a.s. if ~ is a-reguZar w.r.t. C.a
The converse is aZso true provided C is a DC-semiring and E~ E: M.
PROOF. Suppose that ~ is a-regular w.r.t. C, let C E: C be arbitrary
and let {C .} c C be an array of coverings of C such that (8) holds.nJ
Then
p{max ~{s} ~ a} S P{max ~C . ~ a} = P U {~C . ~ a} S I P{~C . ~ a} ,s C . nJ . nJ . nJ
J J J
*and it follows from (8) that ~ C = 0 a.s. Since C is covering, thisa*implies ~a = 0 a.s.
*Conversely, let E~ E: M and ~a = 0 a.s. , and suppose that C is
a DC-semiring. For every C E: C we may then choose a nUll-array {C .} c CnJ
of partitions of C , and we get by Theorem 2.3
*lim I P{~C . ~ a} + E~ C = 0 ,. nJ an+oo J
proving a-regularity of ~ w.r.t. C.
-25-
It is often useful to replace the global regularity condition (8)
by a local one. For a first step, note that if C is a DC-semiring and
if A E M is arbitrary, then (8) is implied by
(9) lim SUP{A~ P{~C ~ a}: C E C£+0
We shall show that the uniform convergence in (9) may essentially be replaced
by pointwise convergence.
THEOREM 2.5. Let ~ be a ~ndom measupe on S , let A E M and let
*a > O. Then ~a = 0 a.s. if
(10) lim sUP{Ai P{~I ~ a}: I E I£+0
sET S E S ,
fop some DC-semiping I c B 3 and also if
(11) lim sup{~~ P{~C ~ a}: C E C£+0
SEC S E S ,
whepe C is the class of closed or the class of open B-sets.
PROOF. Suppose that *P{~ ~ a} > a , and leta I c B be a DC-semiring.
Since I is covering, we may choose some *I E I such that Pis I > O} > aa
We may further choose some null-array {I.} c I of partitions of I , andnJ
we may assume {I.} to be nested in the sense that it proceeds by successivenJ
refinements. Then
(12) *Pis I > O}a* *= P U {~ II' > O} ~ I P{~ II' > a} ,
j a J j a J
and it follows that
(13)
-26-
1 *max xr-- pis II' > a} ~j lj a J
provided % is interpreted as a In fact, this follows trivially
from (12) if AI = a , while if AI > 0 , insertion of the converse of
(13) into (12) yields the contradiction
*P{s I > a} .a
By (13) there exists an index jl such that I = II' satisfies1 J l
1 * 1 *All P{sall > a} ~ rrP{sa l > O}
and proceeding inductively, we may construct a sequence 11 J 12 J •• ,
in I with lIn' ~ 0 and such that
1-- P{OAI n
nn EN.
Now {I} has a non-empty intersection {s}, and the last relation showsn
that (10) is violated at s.
Let us now introduce the class
I = {B E B: AaB = a} ,
where aB = B n BC
denotes the boundary of B If we can show that I
is a DC-ring, it follows that we may apply the first assertion to I.
Now the ring property of I follows from the elementary relations
a(B u C) c aB u ac a(B n C) c aB u ac
To see that I has the DC-property, it suffices to show that, for any
fixed t E S , the ball
-27-
S(t,r) = {s E S: p(s,t) < r}
belongs to I for arbitrarily small r > 0 For a proof, note first
that S(t,r) E B for small r, say for r < £ , since S is locally
compact. Next verify that
as(t,r) c is E s: p(s,t) = r} ,
and conclude that AaS(t,r) = 0 and hence S(t,r) E I for all but countably
many r < £ •
With the above choice of I we have AI = AI I E I , and hence
1IT p{O ~ a} I E I ,
which proves that (11) implies (10) when C is the class of closed sets.
In the case of open sets, let £ > 0 be arbitrary and approximate each
I E I by an open bounded set G :J I such that IGI < 2111 v £ and
rAJ AI > 0 ,AG ~
P{t;I ~ a} AI = 0
Then
1IGP{~G ~ { 2l~I P{~I ~ a}
a} ~
AI > 0 ,
AI = 0 ,
and again (11) implies (10).
...28-
From Theorem 2.4 it follows in particular that, given any covering
class C, a random measure ~ is a.s. diffuse provided ~ is a-regular
w.r.t. C for every a > 0 In this case we shall say, simply, that ~
is regular w.r.t. C
2.5.
Other diffuseness criteria are implicit in Theorem
Another case of particular interest is that of point processes. If
is a point process, we write *= ~1and say that is simple if
*~ ~ a.s. Since for point processes ~i = 0 a.s., (5) holds generally
with a = 1 Furthermore, we obtain convenient simplicity criteria by
taking a = 2 in Theorems 2.4 and 2.5.
We conclude this section with a simple but useful lemma.
LEMMA 2.6; Let ~ be a point process on S and let I c B be a DC-
semiring. Then ~ is simple iff
(14) *P{~I > I} ~ P{~ I > I} I ~ I .
PROOF. The necessity of (14) is obvious. Conversely, suppose that (14)
holds, and note that then
* *(15) 0 ~ P{~I > ~ I = I} = P{~I > I} - P{~ I > I} ~ 0 I E I .
Letting I E I be fixed and choosing a null-array {I.} c InJ
of partitions
of I , we obtain by (15) and Lemma 2.2
-29-
* * *P{sI > s I} = P U {sI . > ~ I .} = P U {sI . > s I . > I}j nJ nJ j nJ nJ
covering, it followsproving that
s P U {s*I . > I} = p{I 1 * >j nJ j {s I .>0
nJ*(s ~)I = 0 a.s. Since I is
*that ~ - s = 0 a.s., as asserted.
NOTES. In the point process case, Theorem 2.3 with a = 1 and b = 00
is known as Korolyuk' s theorem, while the converse part of Theorem 2.4
with a = 2 is known as Oobrushin's lemma, see Leadbetter (1972) and Belyaev
(1969). The present generalizations to random measures are new, although a
first ~tep towards Theorem 2.4 was given in Kallenberg (1975) while the
ideas behind Theorem 2.5 are implicit in Kallenberg (1973d). See also
Gikhman and Skorokhod (1969) for a version for random processes in 0[0,1] ~of the direct part of Theorem 2.4. We further point out that our conditions
(10) and (11) generalize the notion of analytic: orderliness in Daley (1974),
while (9) generalizes his notion of uniform analytic orderliness. Finally,
we mention that Lemma 2. I was given without proof in Kallenberg (l973b)
while Lemma 2.6 is taken from Kallenberg (1973a).
PROBLEMS.
*2.1. The measurability of ).l -+).l , a > 0 , is an immediate consequencea
of Lemma 2.1. Give a direct argument, proving the measurability
of ).l -+ ).l~ a > 0 and conclude that ).l -+).ld is measurable.
(All these facts are of course contained in Lemma 2.1.)
2.2. Prove that the events {s = ~d} and (in the point process case)*{t, = ~} are measurable. (Hint: considE~r the differences.)
-30-
2.3. Prove that the notion of null-array of partitions and the conditions
(10) and (11) are independent of the choice of metric p •
2.4. Let s be a Cox process directed by n. Show that s is simple
iff n is diffuse. Prove also that, if s is a p-thinning of
n (p>O) , then sand n are simultaneously simple. (Hint: consider
first the case of non-random n .)
2.5. Show that Lemma 2.6 follows easily from Theorem 2.3 in the particular
case when E~2 € M .
2.6. Let s be a random measure on S, let 1 c B be a DC-semiring*and let a > O. Prove that ~ = 0 a.s. iff P{sI ~ a} ~ P{s'I ~ a} ,a a
I € 1 .
2.7. Show by an example that the decomposition in Lemma 2.1 is not unique
in general, not even apart from the order of terms. (Hint: take
e.g. S = [0,1] , v = 2 Bl = B2 = 1 , and choose (T l , T2)
in two different ways.)
2.8. Show that the converse of Theorem 2.4 may be false when F~ ~ M .
(Hint: consider a suitable mixture w.r.t. n E N of the non-random
simple point processes on R with atoms at j2-n j EN.)
3. UNIQUENESS
The uniqueness results of this section will playa basic role in
the subsequent discussion of convergence in distribution.
with a theorem which applies to arbitrary random measures.
and V = (Bl , ... ,Bk) € Bk , kEN , we abbreviate (llBl ,
Let us start
For any p E M
Write d
-31-
for equality in distribution, i.e. d~, Tl iff
1 -1P~- = Pn .
THEOREM 3.1. Let E; be a random measure on S. Then PE;-l is uniquely
determined by PCE;f)-l, f E Fc ' and even by L~(f) , f E Fc It is
further determined by p(~V)-l , V E rk , kEN, for any fixed DC-semiring
reB.
PROOF. Let and be two random measures on S satisfying dE; n E;V = nV ,
V E rk k E N , for some fixed DC-semiring J c B . Letting V denote
the class of all sets M E M such that PE;-lM = Pn -1M , it is seen that
V contains M and is closed under proper differences and monotone limits.
Furthermore, V contains by assumption the class C of all sets of the
form
k f> N t l , ••. , t k E R+
Since this class is closed under finite intersE!ctions, it follows by Dynkin' s
theorem that V:J o(C) . But o(C) = M by Lemma 1.1, and we reach the, d
desired conclusion E; = n .
A similar argument shows that p~-l is de!termined by the class of
-1all distributions P(t;f l' ... , E;fk) kEN, f l' ... ,fk E Fc
But by the uniqueness theorem for mUlti-dimensional L-transforms,
-1,~fk) is in turn determined by
L f f (tl' ... , t k )f;, l'H.,f;, k
where f = \.t.f.LJ J J
Since f E F , this completes the proof of the firstc
assertion.
-32-
Note in particular that the L-transforms which were calculated in
Section 1 can be used to define the corresponding basic point processes.
COROLLARY 3.2. Let n be a random measure on S and let t;, be a Cox
directed by Then -1 is determined by pt;,-l This statementprocess n • Pn
is also true when n is a point process on S and t;, is a B-compound of
n "' provided PB- 1 is known and such that B ~ o .
PROOF. Suppose that t;, is a Cox process directed by n Solving for
f in 1 - e-f obtain from (1.4) for any fixed f F, we E c
(1)
where Ilfll = sup f(s) Now Lnf is analytic in the right half-plane,
and so it is completely determined by its values on any finite interval.
Hence by (1), is determined by pt;,-l , and since f E F wasc
arbitrary, the assertion follows from Theorem 3.1. A similar argument
applies to the case of compound point processes, except that (1) is then
replaced by
o ~ t < -I If11-1 log P{B = O} ,
where -1cP is the unique inverse of cP on the interval (p{ 8 = a} , 1] .
Stronger results than in Theorem 3.1 are attainable if we confine
ourselves to simple point processes or diffuse random measures.
-33-
THEOREM 3.3. Let ~ and n be two point proeesses on S" and suppose
that ~ is simp le. Further suppose that U c B is a DC-ring and I c B
a DC-semiring. Then d *~ = n iff
(2) P{~U = O} = P{nU = O} U E U ,
and we have even
(3)
d~ = n provided in addition
P{~I > I} ~ P{nI > I} I E I .
PROOF. The conditions (2) and (3) are clearly necessary. Conversely,
suppose that (2) holds. As before, the class V of all sets MEN
-1 -1satisfying p~ M = Pn M is closed under proper differences and monotone
limits, and it contains N By assumption, it: also contains the class C
of all sets of the form {~E N: ~U = O}, U E U. Now it is easily seen
that
{~U = O} n {~V = O} = {~CU u V) = O} u,V E U ,
and since U is closed under finite unions, it follows that C is closed
under finite intersections. Hence, by Dynkin's theorem, V:::> aCC). But
*from Lemma 2.2 with e: = I it is seen that the: mapping ~ -+ ~ is measurable
w. r. t. a (C) * d *so we get ~ = t;, = n , proving the first assertion. If in
addition (3) holds, then
*Pin I > I} = P{~I > I} ~ P{nI > I} I E I ,
and it follows by Lemma 2.6 that * dn = n = t;, • This completes the proof.
-34-
Note that a variety of related results may be obtained by using the
simplicity criteria of Section 2 rather than condition (3). A corresponding
remark applies to the next theorem.
THEOREM 3.4. Let ~ and n be ~o point ppocesses (pandom measupes)
on S, and suppose that ~ is simple (diffuse). Fupthep suppose that
U c B is a DC-ping and C c B a coveping class, and that sand t
ape fixed peal numbeps with 0 < s < t. Then ~ g n iff n is simple
(a.s. diffuse) and
(4) U E U ,
and also iff (4) holds and in addition
(5) E -s~C < E -snCe _ e C E C .
In the case E~ EM, we may peplace (5) by
(6 ) E~C 2': EnC C E C .
It is interesting to observe that (2) is a limiting case of (4) obtained
by letting t ~ 00 •
PROOF. Let us first consider the point process case, put p = 1-t
- e~
and let ~ and n be p-thinnings of ~ and n respectively. By (1.6)
we get
B E B
-35-
and letting x -+- (X) , we obtain by monotone convergence
(8) ~ ( ) -·t~BP{~B = O} = E exp ~B log(l - p) = Ee B E B .
By (4), (8) and the corresponding relation for n , we obtain
P{~U = O} = Ee-t~U = Ee-tnU = P{nU = O} U E U ,
and it follows from Theorem 3.3 that~* d ~*
t;, = n Now and are
and n (cf. Problemsimultaneously simple and correspondingly for
~ d ~*2.4), so we get ~ = n in general, and if n
n
is simple even~ d ~
t;, = n .
Thus the first assertion follows by Corollary ~;. 2.
To prove the second assertion, we may assume that
(5) holds. Choosing
~ d ~*
t;, = n and that
in (7) and in the corresponding relation for n , which is possible since
s < t , we obtain from (5)
~* ~ ~
Ee-xn C = Ee-x~C ~ Ee-xnC
or equivalently
C E C ,
C E C .
But here the random variable within brackets is non-negative, so it must~ ~*
in fact be a.s. 0, and it follows that nC = n C a.s., C E C since
-36-
~
x > O. Writing this result in the form (n~*
n )C = a a.s., C € C ,
and using the fact that C is covering, we obtain n = n a. s., so again~ d ~
~ = n , as required. The last assertion follows by a similar argument,~
based on the relations E~ = pE~ and En = pEn, which may e.g. be obtained
from (7) by differentiation.
The random measure case may be proved from Theorem 3.3 by a similar~
argument where we consider Cox processes ~ and n directed by t~
and tn respectively. Alternatively, we may choose any fixed x > t~
let ~ and n be Cox processes directed by x~ and xn respectively,
and apply the point process case of the present theorem.
As an application, we prove the following partial strengthening of
Corollary 3.2.
COROLLARY 3.5. Let ~ be a S-compound of n and suppose that
p = P{S > O} > a . Further suppose that Cn is known to be simple for
C E B with nC ~ a . Then -1 and PS-l are uniquely determinedsome Pn
by p~-l and p
PROOF. By (1.5),
(9) Ee-t~B = E exp(nB log ~(t)) B E B
and letting t 7 ~ , we obtain by monotone convergence
P{~B = a} = E exp(nB 10g(1 - p)) B E B .
-37-
Considering sets B contained in C , it follows by Theorem 3.4 that
P(Cn)-l is uniquely determined by PI;-l. Let us now write (9) for
B = C in the form
(10) LI;C = LnC 0 (-log <p) •
Since nC ~ 0 by assumption, the function L (~n... is strictly increasing,
so it must have a unique inverse
and we obtain from (10)
on its ,range (P{nC = Ol, 1] ,
proving that <p is determined by PI;-l. We may now apply Corollary
3.2 to complete the proof.
NOTES. In Theorem 3.1, the first assertion is due to Prokhorov (1961),
(see also von Waldenfels (1968) and Harris (1971)), while the second assertion
is essentially a standard result for random prbcesses. Theorem 3.3 was first
proved in the particular case of Poisson processes by R~nyi (1967), and then
in the general case, independently by Mt5nch (1971) (in a weaker formulation)
and Kallenberg (1973a). As for Theorem 3.4, the first assertion was proved
for diffuse random measures (again in a weaker formulation) by Mt5nch (1970),
and independently by Ka1lenberg (1973e) and Grande1l (1973). The point
process case of that theorem is new as are the last two assertions. Finally,
Corollary 3.2 is due in the thinning case to MElcke (1968) and in the case of
Cox processes to Kummer and Matthes (1970), while the assertions for compound
processes in Corollaries 3.2 and 3.5 are new.
-38-
PROBLEMS.
3.1. Show that, for any random measure ~,the distribution of ~ is
determined by L~(f) for all linear combinations f of indicators
IT' lEI, for any fixed DC-semiring 1.
3.2. Theorem 3.1 clearly extends to arbitrary semirings generating B.
Prove that the first assertion in Theorem 3.3 extends to arbitrary
rings U generating B. (Hint: Show that the class of sets U
satisfying (1) is closed under monotone bounded limits, and hence,
by the classical monotone class theorem in Lo~ve (1963), must contain
B = a(U) . Then apply Theorem 3.3 with U = B.) Even the second
part of Theorem 3.3 admits a similar strengthening, cf. Kerstan,
Matthes and Mecke (1974), page 43.
3.3. Prove an alternative to the second part of Theorem 3.4 by applying
the same method involving Cox processes to the second part of Theorem
3.3.
3.4. Prove that the distribution of an arbitrary point process ~ is
not even determined by P(E,;B)-l , B E B (Hint: consider two
and2 d andrandom vectors a 13 on {O,1,2} such that a ~ 13 ,
still a l ~ 13 a 2 ~ 13 and a l + a 2~ 13 + 13 2
.)1 2 - 1
3.5. Show that the ring property of U in Theorem 3.3 is essential.
(llint: we may e.g. introduce a suitable dependence in a stationary
renewal process.) Cf. Kerstan, Matthes and Mecke (1974), page
213, for a stronger result and further references.
3.6. Prove that, for an arbitrary random measure ~,the intensity
E~ determines p~-l iff E~ = a .
3.7. Show that
-39-
as t-l-O, B E B . This is
interesting since, for simple point processes and diffuse random
measures ~, p~-l is determined by the left-hand sides for arbitrarily
small t > 0 , and still it is not determined by the limit.
3.8. Assume that S is countable. Show that there exists some fixed
function f E F such that p~-l is determined by P(~f)-l for
every point process ~ on
numbers f(s) , s E S , be
above and below by positive
S with ~S <: 00 a.s. (Hint: let the
rationally independent and bounded
constants.)
3.9. Show that, if the class C in Theorem 3.4 is a DC-semiring, then
~ may be allowed to have atoms of fixed size and location. (Cf.
the proof of Theorem 7.8 below.)
4. CONVERGENCE
In view of Theorem 3.1, it appears natural to say that the random
measures
whenever
/;n' n EN, converge in distribution to ~ (wri tten
(1) ~ V ~ ~Vn
kEN ,
for some fixed DC-semiring I, and this is also the mode of convergence
used in much classical work (with I the class of real intervals). However,
in the sense of (1), (cf. the trouble
(1) requires too much in general since
does not necessarily imply o ~ 0T T
n
dT -l- Tn
for random elements in s
with pointwise convergence of distribution functions). Moreover, the above
-40-
mode of convergence depends strongly on the choice of I As will be
seen below, we can avoid these disadvantages by assuming that I satisfies
sal = 0 a.s., lEI. We prove that OC-semirings I with this property
do exist. Let us write B~ = {B E B: ~aB = 0 a.s.}
LEMMA 4.1. Bs is a DC-ring for every random measure ~ on S.
PROOF. Proceeding as in the proof of Theorem 2.5, we need only show that
(2) ~{s E S: p(s,t) = r} = 0 a.s.
for any fixed t E S and arbitrarily small r > O. For this purpose,
let s > 0 be such that
Set, s+) = {s E S: p(s,t) ::; dEB,
and define the random process X in O[O,s] by
Then
X(r) = ~S(t, r+) r E [O,s] .
~{s E S: p(s,t) = r} = X(r) - X(r-) r E (O,s] ,
and since we can have P{X(r) ~ X(r-)} > 0 for at most countably many
r E (D,s] (cf. Billingsley (1968), page 124), it follows that (2) holds
almost everywhere on [O,s] This completes the proof.
-41-
To be able to apply the powerful theory of weak convergence of pro-
bability measures, we shall now introduce a metrizable topology in the
measure space such that (1) subject to the above restriction on I is
equivalent to convergence in distribution in the sense of this theory.
For any
to )J
11, ]11' 11 2 , •••
vand write ~ ~ ~n
in M or
whenever
N , we shall say that l1n
11 f ~)Jf for all f E Fn c
tends vaguely
The so
induced vague topology in M or N is known to be metrizable and separable
(see Bauer (1972); it is indeed even Polish, see Lemma 10.1 below). It is
often useful to relate the notion of vague convergence to the better known
notion of weak convergence. For any 11, 11 1, )J2' ... E M with l1S < 00
we say that l1 n tends weakly to lJ and write whenever 11 f ~ l1fn
for every bounded continuous function f E F The reader should verify
that wIln -+ 11 iff v
11 ~)Jn and )J S ~ )JS < 00
n
The next lemma shows that the random elements in the topological
Borel spaces M and N coincide with our random measures and point processes
respectively.
LEMMA 4.2. M and N coincide with the a-algebras generated by the vague
topologies in M and N respectively.
PROOF. The vague topology in M (or N) is by definition the product
topology induced by the mappings )J ~ pf f E F , so the class Cc
kEN, is a topological base. Since M (N) is
separable, every vaguely open set is contained in a(C) , and so a(C)
-42-
is the a-algebra generated by the topology. On the other hand we have
aCe) = M (or N) be Lemma 1.1. This completes the proof.
Unless otherwise stated, convergence in distribution of random measures
d (d d d . ~) . 1 h f han point processes enote ~,or sometlmes ~ wll ence ort mean
weak convergence of the corresponding distributions with respect to the
vague topology, i.e. ~n i ~ iff Ef(~n) ~ Ef(~) for every bounded and
vaguely continuous functional f: M~ R+ or N ~ R+ respectively. Con
tinui ty in distribution holds for a large class of linear functionals.
Let Of denote the set of discontinuity points of the function f.
Let us further denote the mappings ~ ~ ~f and ~ ~ ~B by wf and
wB respectively.
~, ~l' ~2' ... be random measures on S such that
~ f i ~f for every bounded function f: S ~ R withn +
LEMMA 4.3. Let
d~ ~ ~. Thenn
compact support satisfying ~Df = 0 a.s. Furthermore~
kevery V E B~, kEN.
~ V i ~V forn
PROOF. Write C E B for the support of f and choose some open set
G E B with G ~ C By Urysohn's lemma, there exists some function
g E Fc such that Assuming v~ ~ ~ , we get for any boundedn
continuous function h E F
since gh E F , which proves thatcIf we further assume that
-43-
)JDf = a , it follows by Theorem 5.2 in Billingsley (1968), (which clearly
remains true for non-normalized measures), that
This proves that
(3)
The first assertion now follows by Theorem 5.1 in Billingsley (1968).
Specializing to indicators f = lB' B E B , it is seen from (3) that
B E B .
Hence the second assertion follows by applying Theorem 5.1 in Billingsley
k(1968) to the mapping )J + )JV, V E B~, kEN.
For point processes, convergence in distribution means the same thing
in M and N. In fact,
LEMMA 4.4. N is a vaguely closed subset of M. The class of point
processes on S 1.-S closed under convergence in distribution w.r.t. the
vague topology on M.
PROOF. Suppose that E N with v M By Lemma 4.3)Jl' )J2' )In + )J E
we have )J B + )JB for any B E B with )JaB = 0 , so )JB E Z for anyn +
such B Applying Dynkin's theorem as in the proof of Lemma 1. 2, it
follows that )JB E Z for all B E B , and so )J E N This proves that+
-44-
N is a vaguely closed subset of M. Now suppose that ~l' ~2' '"
are point processes on S converging in distribution to some random measure
~. By Theorem 2.1 in Billingsley (1968), we obtain
P{~ E N} ~ limsup P{~n E N} = 1 ,n-+oo
proving that ~ E N a.s., i.e. that ~ is a point process.
We shall now prove the basic fact that the classical notion of con-
vergence introduced above is essentially equivalent to convergence in
distribution with respect to the vague topologies.
THEOREM 4.5. Let ~, ~l' ~2' ... be random measures on S and Zet
I c B~ be a DC-semiring. Then the foZlowing statements are equivalent.
(i)
(ii)
d~ -+ ~ ,n
~ f i ~fn
f E Fc
(iii) L~ (f) -+ L~ (f)n
f E Fc
(iv) ~ V i ~Vn
kEN .
Two lemmas are needed for the proof.
LEMMA 4.6. Le t rl' r 2 , ... be random measures on S. Then {r}~ ~ . ~n
ia tight iff {t.: B}n
is tight for each B E B •
-45-
PROOF. We shall make use of the fact (cf. Bauer (1972)) that a set K
in M or N is relatively compact iff
(4) sup pB < 00
PEK
B E B .
Let us first suppose that {~n} is tight. By definition, there exists
for each £ > 0 some compact set K c M such that
(5) P{~ E K} ~ 1 - £n
n EN.
For fixed B E B we get by (4) m = sup pB < ':0 , and hence by (5)pEK
P{~ B > m} ~ P{~ 4 K} ~ £n n
n EN,
proving tightness of {~nB} .
Conversely, suppose that {~nBj is tight for each B E B. Let
Gl , G2
, ... be open B-sets with Gk + S , and choose for fixed £ > 0
the constants ml , m2 , ... > 0 such that
k , :n EN.
Define
K = n {p EM: p Gk ~ ~} ,k
and note that K is relatively compact, since there exists for any B E B
some k <:: N with Gk::J B , and therefore
sup llB ~ sup pGk ~ mk < 00 •
llEK llEK
-46~
Furthermore,
p{~ 4 K} = P u {~ Gk > ~ } S L P{~ Gk > mk}n k n K k n
proving tightness of {~n}
s; £ I 2-k= 2£
k
LEMMA 4.7. Let ~,n and ~l' ~2' be random measures on S, let
U c B be a DC-ring and let t > 0 be fixed.
and that ei the 1"
Further suppose that d~ 4- nn
(6)
or'
liminf P{~ U > £} S P{~U ~ £}n
U E U £ E (0,1) ,
(7) limsup E exp(-t~ U) ~ E exp(-t~U)J}+<lo n
U E U .
Then nF = 0 a.s. for every olosed Bet FEB with ~F = 0 a.s.
PROOF. Suppose that (6) holds and let FEB be closed with ~F = 0
a. s. By Lemma 4.1 there exists for every £ > 0 some closed set C E Bn
such that C ~ F and P{~C ~ £/2} s; £/2 We may next choose some U E U
such that U ~ C and P{~(U\C) ~ £/2} s; £/2. Then
P{~U ~ £} s P{~C ~ £/2} + P{~(U\C) ~ £/2} s £ .
Now ~ C ~ nC by Lemma 4.3, so we obtainn
P{nF > d s; P{nC > d s; liminf P{~ C > dnn4-<lO
s; liminf P{~ U > d S P{~U ~ d s: £ ,nn4-<lO
-47-
and since € was arbitrary, it follows that I1F = 0 a.s. as asserted.
A similar proof applies to the case when (7) holds.
PROOF OF THEOREM 4.5. The fact that (i) implies (ii) and (iv) was proved
in Lemma 4.3. Since (iii) follows trivially from (ii), it remains to
prove that (iii) and (iv) both imply (i).
Let us first suppose that (iii) holds. Then
L~ (tf) + L~(tf)n
f E Fc
so (ii) holds by the continuity theorem for L-transforms. But from (ii)
and Lemma 4.6 it is seen that { ~} is tight, and so it follows by Prohorov's"'n
theorem (Theorem 6.1 in Billingsley (1968)) that { ~} is relatively"'n
compact. This means that every sequence N' c N contains a subsequence
(ii)
By Theorem 3.1, this implies
E; .i E; (n E N) by Theoremn
was arbitrary, we getN'E; g n , and since
Nil such that ~ .in
we obtain E; f .i nfn
some n (n E: Nil) . From the implication (i) ==>
(n E Nil) , f € F , and since also ~ f ~ i;f ,c ndf E F ,we must have E;f = nf, f E Fc c
2.3 in Billingsley (1968).
Now this step is permitted by Lemma 4.3
To prove that (iv) implies (i), the same argument
for the point where we conclude from
k(n E Nil) , V E I , kEN.
d~ + n"'n
(n E Nil)
goes through, except
dthat E; V + nVn
provided I c B , and we shall show that this relation is indeed true.n
For this purpose, let U be the ring of finite unions of I-sets, and
note that these unions can always be taken disjoint. By (iv) and the
fact that addition is a continuous operation from to R , we ohtain+
-48-
d~ U ~ ~U, U c U , and son
liminf P{~ U > £} $ limsup P{~ U ~ £} $ P{~U ~ £}n nn-+oo n-+oo
U E U £ > a .
Hence it follows by Lemma 4.7 that nF = a a.s. for every closed set
FEB with E;F = 0 a.s., and in particular n31 = 0 a.s., I E I ,
as desired.
We shall now improve Theorem 4.5 in the particular cases of conver-
gence towards simple point processes and diffuse random measures. Given
any covering class C c B , and any a > 0 , we shall say that the sequence
{~n} of random measures on S is a-regular w.r.t. C if there exists for
every C E C some array {Cmk } c C of finite coverings of C (one for
each mEN) such that
(8) lim limsup L P{t; C k ~ a} = 0 .k
n mJII+<X> n-+oo
Note that, in the particular case when ~l = t;2 = ... = t; , this notion
reduces to a-regularity of E;, as defined in Section 2. As before, we
shall further say that {E;n} is regular, if it is a-regular for every
a > 0
THEOREM 4.8. Let t;, E;l' E;2' ... be point processes on S and suppose
that ~ is simple. Further suppose that U c BE; ~s a DC-ring and I c BE;
a DC-semiring. Then ~ ~ ~ iffn
(i)
(ii)
-49-
lim P{~ U = O} = P{~U = O}nn-+oo
limsup P{~ I > I} ~ P{~I > I}nn-+oo
U E U ,
lEI ,
(iii)
Moreover..
{~B} is tight for every B E B .n
t;, i ~ follows from (i) and (ii) if t;, '/-s 2-regular w.r. t.n
1 and from (i) alone if {~n} is 2-regular w.r.t. 1 or if
(9) 1imsup EE;, I ~ EO < 00nn-+oo
lEI .
PROOF. The necessity of (i) - (Hi) follows by Lenunas 4.3 and 4.6. Con-
versely, assuming (i) - (iii) • it follows by Lemma 4.6 that {~n} is
tight and hence relatively compact, so any sequence N' c N must contain
a sub-sequence N" such that ~ i some n (n E Nil) Here n is an
point process by Lemma 4.4. By (i) and Lemma 4. 7 we get naB = 0 a. s. ,
B E U u 1 , and so by Lemma 4.3
Hence
dt;, B -+ nBn
(n E Nil) BEUuI.
(10) r{nU O} = limnEN"
Pit;, U = O} =n
P{~U = O} U E U ,
p{ nI > I} = limnEN"
p{~ I > I} ~ limsup Pit;, I > I} ~n nn-+oo
P{t;,I > l} I E: 1 ,
and it follows by Theorem 3.3 that d~ = n Since N' was arbitrary,
this implies
-50-
Now suppose that ~ is 2-regular w.r.t. I and that (ii) holds.
Let I E I be fixed, and let {I.} c InJ
be an array of finite coverings
of I such that (2.8) holds with a = 2 and C . :: I .nJ nJ
If rm
is
the number of sets Imk for fixed mEN, we get by (ii)
limsup p{~ I > r }n mn+oo
~ limsup P U {~ I k > I}n+oo k n m
~ limsup L P{~ I k > I}n+oo k n m
Hence by (2.8)
lim limsup P{~ I > r } ~ lim L P{~I k > I} = 0 ,m-+oo n-+«> n m m-+oo k m
proving tightness of {~nI} . Therefore, (iii) follows from (ii) in this
case.
Next suppose that { ~} is 2-regular w.r.t."'n I and that (i) holds.
Then tightness of {~n} follows as above, so any sequence N' c N must
contain a sub-sequence Nil such that some n (n E Nil) As before
we may conclude from Lemma 4.7 that naB = 0 a.s., B E U u I so in
particular (10) remains true and ~ ~ n* follows by Theorem 3.3. Further-
more, it follows from (8) with a = 2 and Cmk :: Imk E I that
lim I P{nI k > I} = lim L lim P{, I k > I} ~ lim limsup L P{~ I k > I} = 0 ,m+"" k m m+= k nE Nil n m m-+oo n-+«> k n m
so n is 2-regular w.r.t. I, and hence simple by Theorem 2.4. Therefore
d~ = n , and again ~ ~, follows since N'
nwas arbitrary.
-51-
Finally suppose that (i) and (9) hold. By ~ebysev's inequality,
(9) implies tightness of {~ I}n
for every I E I , so is tight
by Lemma 4.6. Ifd
~ + n as n + 00 through some N" eN, then it followsn
as above that d * d~ = n and ~ I + nIn
(n EN") , I E I , so by (9) and
Fatou's lemma (in the formulation for convergence in distribution, see
Billingsley (1968), Theorem 5.3),
E~I*= En I ~ En! ~ liminf E~ I ~n
nEN"limsup E~ I ~ E~I
nn+OOI E I .
*lienee EnI = En I for all I E I , and n must be simple. As before,
we may conclude that The proof is now complete.
From Theorem 3.4 we obtain the following analogeous result which
may be proved in the same way.
THEOREM 4.9. Let ~, ~l' ~2' ... be point processes (random measures)
on S and suppose that ~ is simple (diffuse). Fu~ther suppose that
sand t are fixed real numbers with 0 < s < t and that U c B~ ~s
a DC-ring and C c B~ a covering class. Then _ d ~iff1- +
~n
-t~ U= Ee-t~U(i) lim Ee n U E U ,
n+oo-s~ C
> E -s~C(ii) liminf Ee n C E C- e ,n+oo
(iii) {~nC} is tight for every C E C .
Moreover, ~n 1. ~ follows from (i) alone if {~n} is 2-regular (Y'cgular')
w.r. t. C 01' if (9) holds with C in place of I
-52-
In many applications we need to consider convergence in distribution
with respect to the weak rather than the vague topology on M. When
wishing to emphasize the underlying topology, we write wd--+-
vdor --+- instead
of d-+ No new theory is needed for the case of the weak topology, since
this case may easily be reduced to that of the vague topology:
THEOREM 4.10. Let ~, ~1' ~2' ... be a.s. bounded random measures.
Then the following statements are equivalent.
(i) ~ ~ ~ ,n
(ii) ~ ~ and d~ ~ S -+ ~S •n n
(iii) ~ ~ ~ and inf 1imsup P{~ BC> d = a £ > an BEB nn-ro>
Note that (i) is trivially equivalent to
The point is that the latter condition may be replaced by the seemingly
much weaker conditions (ii) or (iii).
PROOF. Since ~ ~ ~ iff ~ ~ ~ and ~ S -+ ~S , the weak topology isn
metrizab1e, so the theory of convergence in distribution in metric spaces
applies to it. Write M(w) and M(v) for the space of bounded measures
in M endowed with the weak and vague topology respectively. Let us
first suppose that (i) holds. Since the mapping ~ -+ (~, ~S) from M(w)
into
(1968).
M(v) x R+
is continuous, (ii) follows by Theorem 5.1 in Billingsley
-53-
Next suppose that (ii) holds, and let £,0 > a be arbitrary. Since
{i; S} is tight, we may choose some a > a such thatn
(11) P{i; S > a} < 0n
n EN.
By dominated convergence, there further exists some B E B with
(12)_I;BC -a
E(l - e ) < £oe ,
and by Lemma 4.1 we may assume that l;aB = a a.s. Using (11), we obtain
a-I; S -1 -I; S-1 c n I; BC> I; S $ a] eaEI; BCe n
+ 0<' E[£ I; B e £ , + 8 $ £n n n n
I; BC -I; S -I; B -I; S~
-1 aE( n 1) e n+ 8 -1 aE( n en) 8E e e = £ e . e + ,
so by (ii), (12) and Lemma 4.3,
-I; B -I; SC -1 annlimsup P{I; B > £} $ £ e lim E(e - e ) + 8nn-+<x> n-+<x>
= £-leaE(e-I;B _ e-I;S) + 0 = £-leaE[e-I;B(l _ e-i;BC)] + 8
C$ £-l eaE(l _ e-I;B ) + 0 $ 8 + 8 = 28 .
Since 0 was arbitrary, this proves (iii).
Let us finally suppose that (iii) holds. Choose for each kEN
some Bk E BI; such that Bk t Sand
limsup P{l;n~ > k- 1} < k- 1
n-+<x>n EN.
-54-
Then
(13) lim limsup P{~nB~ > 8} = 0k-+-co n+co
£ > 0 .
We now define the metric p
by
in the topological product space M(v) x R+
lll' ll2 ( M
where is any metrization of M(v) Then
I~ Bk - ~ Sl = ~ BC
n n n k n,k EN,
so (13) may be written
Furthermore, we have in M(v) x R+
£ > 0 .
(15) (~, ~Bk) -+- (~, ~S) a.s. as k -+- co ,
and finally, since the mappings II -+- (ll, ll~) are continuous
M(v) -+- M(v) x R at all points II with lldBk = 0+
(16) (~n 'd
(E;, ~Bk) kEN .~nBk) -+- as n -+- co
Using Theorem 4.2 in Billingsley (1968), we may conclude from (14), (15)
and (16) that
(17)
Now the mapping
-55-
is continuous M(v) x R ~ M(w) , and therefore+
(i) follows from (17) by Theorem 5.1 in Billingsley (1968). This completes
the proof.
NOTES. The equivalence of (i) - (iii) in Theorem 4.5 was proved by Prohorov
(1961) for compact spaces and in general by von Waldenfels (1968), (see also
Harris (1971)). The remaining equivalence was given in Kallenberg (l973a).
This is also the source for Theorem 4.8, except for the assertion involving
(9) which was added by Kurtz (1973), while Theorem 4.9 improves and extends
a result by Kallenberg (1973e) and Grandel1 (1973). Theorem 4.10 is new,
although it was used implicitly in Ka1lenberg (1973c) and (1974b). We
finally refer to Jagers (1974) for Lemmas 4.2 and 4.4.
PROBLEMS.
4.1. Let P, PI' P2 , ... E M and let I E Bp be a DC-semiring. Prove
that p ~ p iff p I ~ pI, I E I .n n
4.2. Prove that p ~ p iff limsup p F :0; pF , FEB closed, andn nliminf 1J G ;:: :J.IG , G E B open. Note that one of these conditionsnis not enough.
4.3. Let F €: B be closed and G E B open, and let t E R Show+
that the p-sets {pF < t} and {pG > t} are open in M and N
(Hint: use Problem 4.2.)
4.4. Prove that, for bounded EM, w iff vp, PI' P2' ... Pn ~ P Pn ~ )J
and p S ~ pSn
on S and letwd
Ck~n -+ Ck~ ,for any bounded
W f = 0 a.s.
4.5.
-56-
Let ~'~l' ~2' ... be random measures
with Cko
t S. Prove that ~ ~ ~ iffvd n d
generally, ~ -+ ~ implies f~ ~ f~n n
f E F with bounded support satisfying
Cl , C2 , ... E B~
kEN. More
function
4.6. Prove that, if " '1' '2' are random elements in (S, B)
thend iff o ~ 0, + ,
n T ln
4.7. Prove Theorem 4.9.
4.8. Extend Theorem 4.8 to the case when ~1' ~2' ... are general random
measures.
d~n + ~ ,
a.s. if
4.9. Let ~'~1' ~2' ... be random measures on S such thatft .,.
and let 6: ~ 0 be fixed. Show that ~ = ~ - 0"'6:+ - "'(e:,oo) -
{~n} is a-regular w.r.t. B~ for every a > e: , and that the converse
is also true whenever E~ E M (Cf. Kallenberg (1975).)
4.10. Show that, if the class C in Theorem 4.9 is a DC-semiring, then
~ may be allowed to have atoms of any fixed size and position.
(Cf. Problem 3.9.)
4.11. Show that the equivalence of (i) and (iii) in Theorem 4.10 may essentially
be restated as a tightness criterion for random measures subject to the
weak topology in M. Give also a direct proof of this criterion.
4.12. Let ~1' ~2' .. . be Cox processes directed by n1 , n2 , ... respectively.d dShow that ~ + some ~ iff nn .+ some n , and that in this case ~ is
na Cox process directed by n . Prove the corresponding result for p-
thinnings with fixed p > 0
•
-57-
4.13. Let n, nl , n2, ... be random measures on S
suppose that the functions f, f l , f 2, '" € F
wi th uniformly bounded support and satisfy
dwith ~ + nand
are uniformly bounded
(Cf. the proof of TheoremShow that in this case
n{s € S: fn(sn) + f(s) for some sl' 52' ... + s} = 0 a.s.
d~fn + nf .
8.1 below.)
5. EXISTENCE
In the limit theorems of the preceding section, the limiting random
measure ~ was always assumed to exist. We shall now show that, under
mild addi tional conditions, the existence of E; will hold automatically
hy compactness. This provides a new approach to the general existence
problems in random measure theory.
We begin with the counterpart to Theorem 4.5.
LEMMA 5.1. Let E;l' E;2' ••• be random measures on S and let U c B
be a DC-ring. Suppose that either
(i) f € Fc or
(ii)d
~nV + some SV'
c,v are such that
un: th uk.y au .
k € N , where the random vectors
whenevey;' U, Ul ' u2 ' ..• € U
Then l; §. somen
l; satisfying
-58-
f E Fc 01'
kV E U ,
kEN , respectiveLy.
PROOF. In both cases, {l;n} is tight by Lemma 4.6, so by Prohorov's
theorem, /; §. some l; as n -+ 00 through some sequence Nt c N Butn
then /; f §. /;f (n E Nt) , f E F , by Theorem 4.5, so assuming (i).n c
we obtain I;f ~ I;f , and the assertion follows by Theorem 4.5.
In case of (ii), we may proceed as in the proof of Lemma 4.7 to conclude
that 1;3U = a a.s., U E U But then I; V ~ I;V (n E Nt) , V E Uk ,n
kEN, by Theorem 4.5, and the assertion follows as before.
Arguing in the same way, we can easily derive analogeous results
for convergence towards simple point processes and diffuse random measures.
For example,
LEMMA 5.2. Let 1;1' I;Z' be point processes on S, and Zet U c B
be a DC-ring and I c B a DC-semiring. Suppose that {I;n} is Z-reguLar
1.1'. r. t. I and that
(1) P{I; U = O} -+ some ~(U)nU E U .
Further suppose that ~(Un) -+ 1 whenever VI' Uz, ... E U and B E U u I
with U -t aBn
Then there exists some simpZe point process l; on S
(2)
dI; -+ I; andn
P{l;V = a} = ~(U) U E U .
-59-
We shall now use Lemma 5.1 to derive two important special cases
of a general existence theorem corresponding to the second part of Theorem
3.1. For notational convenience, we consider random vectors in 8S
random measures on the space {I, ... ,k} , kEN
THEOREM 5.3. Let U c B be a DC-T'ing and foT' any V E Uk, kEN,
let ~V be a T'andom vectoT' in R~ Then theT'e exists some T'andom meaSUT'e
~ on S satisfying U c B~ andkV E U , kEN, iff
(i) the distT'ibutions p~-l aT'e consistent,V
a . s . wheneve:'('(ii)
(iii)
~V{l} + ~v{2} = ~v{3}
with disjoint VI and V2 ,
~V ~ 0 wheneveT' V, VI' V2, ... E U withk
V -I- av .n
PROOF. The necessity of (i) - (iii) is obvious. Conversely, suppose
that (i) - (iii) hold. For fixed V E U , let
of nested partitions of V , choose arbitrary
{U .} c UnJ
Snj E Vnj
be a null-array
for all nand
j , and define the random measures ~l' ~2' ... on S by
~n = ~ ~v{j}os . ' where V = (UnI , Vn2 ' ... ) , n EN.J nJ
Writing U for the ring generated by V I' U 2' ... , we obtain by (i)n n n.
and (ii)
kEN ,
-60-
and so
d~ V + ~n V kEN ,
where U' = U U Now U' is a DC-ring in the subspace V , so itnn
follows by (iii) and Lemma 5.1 that ~ ~ some ~' satisfyingn
(3) kEN .
To see that (3) extends to U n V , let mEN and VI' ... ,Um E U n V
be arbitrary, and repeat the above argument with a refined null-array
such that U' is replaced by some U":::> U, U {Ul'
the existence of some random measure ~" such that
,V } . This yieldsm
(4) ~"V 9: t;<, - V kEN .
Comparing (3) and (4), we obtain
t; 'V d t;"V kEN ,
and so t;' ~ E;" by Theorem 3.1. Inserting this into (4) yields the desired
extension of (3).
To complete the proof, let VI' U2 , ... E U with U~ + S , and let
t;i, t;2 .. , be random measures t;' obtained as above with VI' U2, ...
as lJ. Then it follows by another application of Lemma 5.1 that t;k ~ some t,
with the asserted properties.
Rk . Then there exists some random measure+
V E Bk
, kEN, iff
THEOREM 5.4. For anyk
V E B ,
-61-
kEN , let ~ be a random veator in
d·on S satisfying ~V = ~V '
(i) the distributions PE;~l are aonsist;ent,
(i i) a. 8. whenever'
with disjoint B1 and B2 ,
(iii)d
E;B + ° whenevern
B1
, B2
, ... E B with B '" ~ .n
PROOF. The necessity is again obvious. Conversely, suppose that (i)
- (iii) hold. Define U = {B E B: E;aB = 0 a.s.} , and verify that U
is a ring. To see that U has the DC-property, let t E S be fixed
and put S(t,r) = {s E S: p(s,t) < r}
and suppose that 8 > ° is such that
Choose £ > 0 with S(t,£) E B ,
(5) P{E;aS(t,r) > 8} > 6
for infinitely many r E (0,£) , say for r l , r 2 , ... By (i) and Kolmogorov's
consistency theorem, there exist some random variables E;O' ~l' E;2' .••
such that
and we get by Fatou's lemma
-62-
00
contradicting the fact that, by (ii), L ~n ~ ~O < 00 a.s. Hence (5)1
can hold with fixed 0 > a for at most finitely many r E (O,E) , and
so the set {r Eo: (O,E): ~S(t,r) = a a.s.} must be dense in (O,E)
This proves that U contains a base and hence has the DC-property.
Applying Theorem 5.3, it is seen that there exists some random measure
~ on S satisfying
(6) kEN .
To extend (6) to B, let k E Z+ and V E Uk be fixed and let
B, Bl , B2, '" E B with
we get
B t B .n
By Theorem 4.4 in Billingsley (1968)
and so by continuity
(7)
d~ \ + ~ ~,V,B,B B V,B,pn
d~V B + ~V B .
, n '
A similar argument shows that (7) is also true when
other hand
B oj. B .n
On the
~(V, B ) + ~(V,B)na. s. , B + B ,
n
so by the monotone class theorem in Lo~ve (1963) we have dt;(V,B) = ~V B ',
B E B Proceeding inductively, we obtain the desired extension of (6).
-63-
Turning to the existence problem for point processes, let us first
consider a necessary condition. Define for any set function ~ on B
llA~(B) = ~(A u B) - ~(B) A,B E B •
Similarly, let llA ... llA ~(B) be the set function obtained by forming1 n
successive differences with respect to B. Say that ~ is corrpletely
monotone if
n EN.
LEMMA 5.5. Let ~ be a random measure on S and let ~ E (0,00] be
fixed. Then the set function -t~BEe , B E B , is completely monotone.
PROOF. Let us first suppose that t = 00
we get
For the first order difference
o s P{~A > 0, ~B = O} = P{~B = O}
= P{~B = O}
P{~A = 0, ~B = O} =
P{~(A u B) = O} = -l:IAP{~B = O} ,
and continuing inductively, we obtain
o s P{ ~A1 > 0, .,. ,~n > 0, ~ Bn= O} = (-1) l:I ... llA P{~B = O}
Al nn EN.
In the case t < 00 ,let n be a Cox process directed by t~ and note
that
p{nB = o}= Ee-t~B B c B .
-64-
Now suppose conversely that ~ is a completely monotone set function
defined on the DC-ring U and such that ~(Un) + 1 whenever U, Ul , U2' ... E U
with U + au. Let V E U be arbitrary and letn
be a null-array of nested disjoint partitions of
Unl ' ... ,Unkn
V. For fixed n EN, choose
process on S with all its mass confined to snl'
r{1; {s .} = 1, j E J; ~ {s .} = 0, j ¢ J} =n nJ n nJ
non-random points
(8)
s . E U . ,nJ nJ
j = 1, ... ,k .nLet ~n be a simple point
, snk and such thatn
= P{E; U . > 0, j E J; 1; U U . = O} =n nJ n j¢J nJ
= {j~J [-OUnj]}~[j~J Unj]for each subset J c {I, ... ,kn } . Note that ~ exists, since the<"n
right-hand sides of (8) are non-negative, and summation over J yields
1. (The latter fact is easily established by induction in the number
of partitioning sets.) By (8),
U E Un
n EN,
where
with
Un is the ring generated by Unl ' ... ,Unk ' so (1) is satisfiedn
U replaced by U, = U U If {1;} is known to be tight, it followsn nn
as in the proof of Theorem 5.3 that (2) is satisfied for some point process
*1; on S. Since (2) remains true for 1; , we reach the following result.
THEOREM 5.6. Let ~ be a set function on some DC-ring U c B. Then
there exists some simple point process E; on S satisfying (2) and such
that
-65-
U c B iffS
(i) ~ is completely monotone on U 3
(ii) ~(Un) ~ 1 whenever U, Ul , U2' ... E U with Un ~ au ,
(iii) for every V E U there exists some null-array {Unj } C U
of nested disjoint partitions of V such that the sequence
of point processes sn defined by (8) is tight.
NOTES. Theorem 5.4 is due with an entirely different proof to Harris
(1968) and (1971). For point processes, a much more general result (con
taining in particular the point process case of Theorem 5.3) is given by
Kerstan. Matthes and Mecke (1974), page 17. See also Prohorov (1961)
and Mecke (1972) for further general existence criteria. Finally, Theorem
5.6 is contained in the general results by Kurtz (1973) and Karbe (1973),
(cf. Kerstan, Matthes and Mecke (1974), page 35).
Our approach via Lemmas 5.1 and 5.2 seems to be new, although the
idea to prove existence by means of tightness arguments is well-established
in other fields, (see e.g. Billingsley (1968)).
PROBLEMS.
5.1. Prove Lemma 5.2 and the corresponding result for diffuse random
measures.
5.2. Show that, in the particular case of non-random t;v' Theorem 5.3
reduces to a version of Caratheodory's extension theorem.
~66-
5.3. Prove directly from (i) and (ii) of Theorem 5.4 that (ii) holds
with any finite number of terms. By (i) and Kolmogorovts consistency
theorem, PI;;-l can be consistently defined even for V E Ba> ShowV
by means of (iii) that (ii) then extends to infinite sums.
5.4. Suppose that S is discrete, i.e. that all its points are isolated,
and let ~ be a set function on U = B Show that (2) holds for
some point process I;; on S iff <jl is completely monotone and
<PC"') = 1 .
5.5. Show that Theorem 5.6 is false in general without condition (ii).
(Hint: let S = R and place a fictitious atom at the point 0+.
Let U be the ring generated by all intervals (a,b].)
5.6. Show that Theorem 5.6 is also false in general without condition
(iii) . (Hint: let I;; be a stationary "point process" on the unit
circle, whose atom positions have exactly one limit point.)
5.7. Show by a direct argument that the set function
is completely monotone for any random measure I;;
-tl;;BEe ,B E B ,
and fixed t < a>
6. NULL-ARRAYS AND INFINITE DIVISIBILITY
In this section we consider the general central limit problem for
random measures, i.e. the problem of convergence in distribution of sums
L I;; . , where the I;;nJ· , j,n £ N • form a nuZZ-appay of pandom measupes. nJJin the sense that the I;;nj are independent for fixed n and such that
(1) lim sup P{I;; .B > E} = 0n~ j nJ
E > 0 B E B .
-67-
Note that, in the particular case of point processes, (1) reduces to
lim sup P{~ .B > O} = 0n-+<>o j nJ
B E B .
Just as for distributions on R, the class of limiting distributions
of the sums \ ~ coincides with the class of infinitely divisible dis-Lj nj
tributions. A random measure ~ is said to be infinitely divisible4 if,
for every fixed n EN, we have for some independent
and identically distributed random measures ~1' In the point
process case, we require ~1' ... '~n to be point processes. Note that,
since a point process may be regarded as an N-va1ued random measure, we
have two notions of infinite divisibility for point processes, (which
are not equivalent since the measure n~ may belong to N even for ~
in M\N). If nothing else is stated, we mean infinite divisibility as
a point process.
In our derivation of canonical representations and convergence criteria,
we shall proceed in steps, beginning with the simple particular case of
R -valued random variables. In this case, (1) takes the form+
£ > 0 .lim sup p{~ . > £} = 0n-+<>o j nJ
We shall use K to denote the function 1 - e-x Furthermore, the pro-
jections and
= x·· t
are defined by
4 For a different notion of infinite divisibility, see O. Kallenberg,Infinitely divisible processes with interchangeable increments andrandom measures under convolution, Tech. Report, Dept. of Mathematics,Gtlteborg, 1974.
-68-
LEMMA 6.1. An R+-valued random variable ~ is infinitely divisible iff
(2)-'IT
- log Ee-~t = ta + A(l _ e t) t ~ 0 ,
for some a ~ 0 and some A E M(R')+
with AK < ~ , and in that case
a and A are unique. Moreover, any a and A with the stated properties
may occur in (2). If
then L.~ . &some ~J nJ
{~ .}nJ
iff
is a null-array of R+-valued random variables,
(3) \' -1 wK L'P~ . -+ some
J nJ
and then ~ , a and A satisfy (2). In the case of Z -valued random+
variables, (2) holds with a = 0 and bounded A E M(N) , while (3) is
equivalent L -1 w Nto . p~ . -+ A onJ nJ
Note that (3) is equivalent to the conditions
(i) L'P~-~ ~ A on R' with A{~} = 0 ,J nJ +
(ii) lim limsi L E ~ . = ae: nJ
,e:+O n~ j
where R' = (0 ,~J , limsi stands for both limsup and liminf , and+
E denotes the truncated expectation at the level e: > 0E:
PROOF. Writing ~ . = L~ ,we claim thatnJ ",.nJ
d\' ~ -+ somef "'njJ
iff
(4 ) ~n =? (1 - ~nj) -+ some continuous ~ ,J
and that in this case ~ = L~ = e-~ . In fact, \' ~ ~ ~ impliesLj nj
-69-
II .4> • ....." , whenceJ nJ
- I log ~ ..... - log 4> = ~ ,j nJ
and here ~ is continuous since 4> is an L-transform. Since (1) is
4> , and it follows that
(4) follows from this by con-4> .) .... a ,nJ
log 4> .nJ
Conversely, by a similar argument,
LI; . & some I;. nJJ
R and note+
~n = ~ p~-l. n E·N on1\ L "', ,
j nJ
(4) implies IT 4> ..... some continuousj nJ
We next introduce the measures
clearly equivalent to max(l -j
sidering a Taylor expansion of
that for t ~ 0 ,
<Xl
-tl; . -tl; .?J(1 _ e-tX)pl;-:(dx)~ (t) = L (1 - Ee nJ ) = LE (1 '" e nJ) = =n
j j nJJ 0
<Xl <Xl
f (1 - e-tX)~ (dx) J-txl-e
(K~n) (dx)= = .n l-e -x0 0
Hence (3) implies (4) with 1/1 given by<Xl <Xl
f -txf ( -tX)~(t)
l-e(d) (dx) o ,= ta + = ta + '1 - e . ~ (dx) t ~
-xo l-e 0
~ being continuous by monotone convergence. Conversely, suppose that
(4) holds for some continuous ~, and note that
~ (t + 1) - ~ (t)n n= f (e - tx - e- (t+1) X) ~n (dx)
o= J
o
-txe (KA) (dx)
If ~(l) ~ ~(O) , we may apply the continuity theorem for L-transforms
to the probability measures K~ /~ K , and conclude that KAn convergesn n
weakly on R+ But this is also true when ~(l) = ~(O) , since then
-70-
The uniqueness of a and A in (2) follows from the formula00
1/J(t + 1) - 1/J(t) = fo
-txe (a8
0+ KA)(dx)
and the uniqueness theorem for L-transforms. To see that every choice
of a and is possible, use (3) with -1An = [n , oo)A , n EN, to
construct a ~ with a = 0 and arbitrary A , and then replace ~ by
~ + a. In particular, the infinite divisibility of ~ in (2) follows
by dividing this relation by an arbitrary n E N Conversely, any infinitely
divisible L-transform ¢ = L~ can be difinition be written in the form
¢ = ¢n for every n E Nn so the above argument applies with ¢nj = ¢n
j = 1, ... ,n , showing that ~ has indeed the representation (2). The
last assertion follows easily from (3).
We state explicitly the corresponding tightness criterion.
LEMMA 6.2. Let {~nj} be a nuZZ-array of R+-vaZued random variabZes.
Then {I.~.} is tight iffJ nJ
(i) limsup L Et~ . < 00
n-+OO j nJt > 0 ,
(ii) lim limsup L P{~ . > t} = 0 .. nJ
t -+00 n-+OO J
If the ~nj are Z+-vaZued, then (i) may be repZaced by the condition
(i)' limsup I P{~ . > O} < 00 •
11+00 j nJ
PROOF. The sequence {d }n
-71-
of the last proof is vaguely relatively compact
on R+ iff (KAn) rO,t] is bounded for each fixed t > 0 , i.e. iff (i)
holds. The role of (ii) is to prevent mass from escaping to +~, thus
ensuring every vaguely convergent sub-sequence to be weakly convergent.
For point processes, (i) and (ii) imply (i)', which in turn implies (i),
so in this case, (i)' may replace (i) .
We now turn to the m-dimensiona1 version of Lemma 6.1, mEN
Let ~ be a one-point compactification of Rm and for+ '
t Rmx, E+
write
xt = x • t for the inner product of x and t
(2) holds for Bome
LEMMA 6.3. mAn R+-valued random vector
ma E R+ and some
~ is infinitely divisible iff
m -1A E M(R+\{Ol) with (A~t)K < ~
t E R: ' and in that case a and A are unique. Moreover, any a and
A with the stated properties may occur in (2). If {~njl is a null
array of R:-valued random vectors, then Lj~nj ~ some ~ iff there exist
some a and A as above such that
(i) L·P~-~XA on ~\{O} ,J nJ
(ii) lim limsi LE ~ .{kl =ark} k = l, ... ,m£+0 Jl+<X>
. £ nJJ
and then and satisfy (2) • In the case ofm randomE;, , a A Z -valued+
vectors, (2) holds with a = 0 and bounded A E M(Zm\{Ol) , and we may+
replace (i) and (ii) by the condition L'P~-~ ~ A on Zm\{O} .J nJ +
-72-
PROOF. Let {~ .} be a null-array satisfying (i) and (ii) for somenJ
a and A with the stated properties. Then it is easily seen that
so by Lemma 6.1,
continuity theorem for m-dimensiona1 L-transformsand it follows hy the
dthat I·~ . ~ some ~
J nJ
By Lemma 6.2, there
-TI
E exp(- I t~ .) ~ exp(-ta - A(l - e t)). nJJ
satisfying (2).
dConversely, suppose that I.~ . ~ some ~.
J nJ
exist for every sequence N' c N some subsequence Nil C N' and some
a and A the stated properties such that (i) and (ii) hold as n ~ ~
through Nil. By the direct assertion of the present lemma, it follows
that and satisfy (2). Now at is unique for every
by Lemma 6.1, and therefore a must be unique. As for A, write
1 := (l, ... ,1) E Rm and note that by (2)+
Applying the uniqueness theorem for m-dimensional L-transforms, it follows
that the bounded measure-TI
since 1 - e 1:.. > 0 on-
pendent of the choice of
-'IT 1(1 - e -)A is unique, and hence so is A
R:\{O} . This proves that a and A are inde-
N' , and hence that (i) and (ii) remain true as
n + 00 through N, (cf. Theorem 2.3 in Billingsley (1968)). The remaining
assertions now follow as in the proof of Lemma 6.1.
-73-
Before proceeding to arbitrary random measures, we prove a lemma.
LEMMA 6.4. Let ~ be a random measure on S and let I c B be a DC-
semiring. Then is infinitely divisible iff ~V is for evepy
kEN . In the point process case it suffices that be infinitely
divisible for every II' ,Ik E I, kEN.
PROOF. The necessity of our conditions is obvious. Conversely, suppose
that ~V is infinitely divisible in for every kV E I , kEN .
Since infinite divisibility of random vectors is preserved under convergence
in distribution, (as may be seen from the continuity theorem for L-transforms),
we may use the monotone class theorem in Lo~ve (1963) to extend this property
to arbitraryk
V E B , kEN It follows that there exist for every
n E N some random vectors ~V n satisfying,
(5)-n
(L~V) = L~V,n
k,n EN.
From (5) it is easily seen that the family {~V n} satisfies the conditions,in Theorem 5.4, and hence there exists some random measure ~n on S with
~ V ~ k k E N Writing (5) in the form~V n ' V E B ,n ,
n V E Bk k,n EN,L~V :: (L~ V)n
and using Theorem 3.1, it is seen that ~ is infinitely divisible.
Turning to the point process case, suppose that I·o.J J
is infinitely
divisible (as a random variable in Z+) for every II'
kEN. Since repetitions of I-sets may occur, this proves that, for
-74-
any fixed V = (11' ... ,Ik) ( Ik , kEN , tW = L·t.f;!. is infinitelyJ J J
divisible in Z+ for every t = (t l , ... ,tk) € Nk Turning to rationals
and using the above closure property, it follows that tt;V is indeed
infinitely divisible in R+ for every t = (t l , ... ,tk) € R~. Hence,
by Lemma 6.1, there exist some constants
such that
A ;:: 0p
- log Ee-utt;V = L (1 _ e-utp)Ap p
Moreover, making
In fact, there can be no constant component a here since the
infinitely divisible in Z+ and hence P{t;I. = O} > 0 .J
t;I.J
are
a Taylor expansion in (2), it is easily seen that the support of A is
contained in the support of pt;-l. (This is a simple special case of
Theorem 6.9 below.) Now suppose that n is an infinitely divisible random
vector in Z: with spectral measure A = \ A 0 , the existence of whichL.p P p
is ensured by Lemma 6.3 since LpAp< 00 Then Ltn coincides with
Ltt;V above, and so tn ~ tt;V . If the t. are chosen rationally independent,J
it follows that d t;V which proves that t;v is infinitely divisible.n :: ,
Since V was arbitrary, we may now proceed as in the first part of the
proof to conclude that ~ itself is infinitely divisible.
We are now ready to extend Lemmas 6.1 and 6.3 to random measures
defined on general spaces S.
THEOREM 6.5. fi pandom measure ~ on S is infinitely divisible iff
(6)-H
Log Ee '-1r
= af + A(l _ c f) f E F ,
-75-
-'
for some a E M and some measure _A on M\ {o} satisfying
B E B , and in this case a and A are unique. Moreover~ any a and A
with the stated properties may occur in (6). If
ofY'andom measures on S ~then I.F;. . ~ some t;,J nJ
a and A as above such that
{E;,nj} bS a null-array
iff there exist some
(7) \ -1 w -1KL'P(t;, .f) ~ afo O + K(A~f )J nJ
f E Fc
and then t;, ~ a and A satisfy (6). ~f I c B is a DC-semiring satiSfying
aaI = A{~aI > O} = 0 1 E I , then (7) is equivalent to the conditions
(8) I -1 v -1 ~\{o} y E rk kEN ,. P(t;, .Y) ~ A~V onJ nJ +
(9) lim limsi L E t;, .1 = al 1 E IE:~ n-+<>o j E: nJ
In the point process case~ we have a = 0 in (6) while A is confined
(9) can be replaced by (8) alone~ provided the vague convergence on
is strengthened to weak convergence on z:\{O}
to N\{O} and satiSfies -1(A~B )N < 00 B E B. In this case~ (8) and
It\{O}+
PROOF. Suppose that t;, is an infinitely divisible random measure. Then
E,Y is infini tely divisible for every Y = (Bl , ... , Bk) E Bk
, kEN
so by Lemma 6.4 there exist some unique ~ E R: and Ay E M(R:\{O})
such that
(10)-~
1 E -t(t;,V) t ~ (1 t)- og e = ay + A y - e
-76-
Choosing all components t{j} but one equal to 0, we obtain
ay = (aB ' ,aB ) , while choosing exactly one component equal to1 k
a , it is seen that {Ay } is in the obvious sense a consistent family
of measures. Let us now choose Y = (A,B, A u B) in (10), where A
and B are arbitrary disjoint sets in B. Defining by
we get for any t
A'C = AA B{(x,y): (x,y, x + y) E C} ,,
= (u v w) E R3
, , +
-7TA--Cl t) -__ log Ee-t(~V) -_uaA + vaB + waAuB + .-V - e
= _ log Ee-Cu+w)~A-(v+w)~B
= (u + w)aA
+ (v + w)aB
+ J (1 - e-Cu+w)x-(V+W)YPA,BCdXdy)
-7T
= uaA + vaB + wCaA
+ aB) + A'(l - e t)
By the uniqueness assertion in Lemma 6.4, it follows that
(11)
and A = A' and in particularV '
(12)
= A'{(X,y,z) E R3 : x + y ~ z} =+
= AA B{(x,y) E R:: x + y ~ x + y} = 0 .,
Next suppose that Bl , B2, ... E B with B (~ ~ , and conclude from (10)n
that -t a while
(13)
-77-
E: > 0 .
Combining the former relation with (11) , it is seen that the set function
B -+ Cl B , B E B is in fact a measure. As for the family fAv} let
B f B be fixed and define the measures A' E M(Rk) , V E Bk , k E N , hyV +
Since carries over to
{AV} , we may apply Theorem 5.4 to the latter family, and conclude from
(12) and (13) that there exists some unique measure A' onB
{IJB > O} E M
satisfying Defining the mea~ure on the same set by
A (d) (1 - e- IJB) -lA' (d1l)B IJ = B ~
we ohtain for any measurable set
f (1{IJVEC} J
-x -1= (1 - e ) An v(dX x C) =
R' ,+
-x -1 -y(1 - e ) (1 - e )AB B V(dxdy x C) =, ,
-- A (R' 2x C) = A (R' x C)B,B,V + B,V + '
and it follows that for any B' E B with B' ~ B
This proves that the measures AB are all restrictions of A = sup 1\B .BEB
-78-
It is now easy to verify that this A is the unique measure satisfying
(6). Using Theorem 5.4, it is further seen that any a and A with
the stated properties define the distribution of some random measure
~ satisfying (6), and it follows in particular that any such ~ is infinitely
divisible.
measure satisfying (6), it
and hence by Theorem 4.5,
L.~ . ~ some ~. ThenJ nJ
Now suppose that
is infinitely divisible by Lemma 6.4, and
follows by Lemma 6.1 that
d2.~ . ~~. Conversely, suppose that
J nJ
{~nj} is a null-array of random measures satisfying
A with the stated·properties. If ~ is a random
L'~ .f ~ ~f, f E FJ nJ c
and(7) for some
hence it satisfies (6) for some a and A. But then (7) follows by
Lemma 6.1. The same argument may be used to show that (8) and (9) are
necessary and sufficient, and therefore are equivalent to (7). Finally,
the assertions for point processes may easily be deduced from (8).
We shall now give two useful and illuminating interpretations of
the canonical representation (6). Let us write lCa,A) for any infinitely
divisible random measure ~ satisfying (6). In the point process case we
write 1(0,1..) = 1(1..) .
LEMMA 6.6. Let a and A be such as in Theopem 6.5 and let n be a
Poi8son process on M\{O} with intensity A. Then
(14) l(a,A) ~ a + f lJn(dlJ) .
PROOF.
-79-
~
Denote the 1eft- and right-hand sides of (14) by ~ and ~
Since for fixed f € F the mapping 1Tf : ].l -+ ].lf is measurable, it follows~
as in Section 1 that the function ~f = O'.f + n1Tf is a possibly infinite-
valued random variable, and we obtain by (1.2) and (6)
-~f -af -mTf -af ( -1rf ) -~f(15) Ee = e Ee = e exp -A(l - e ) = Ee .
In particular~ d~B = ~B < 00 a. s. , B E B , and it follows as in Lemma
~
1.3 that ~ is a random measure. Finally, (14) follows from (15) by
Theorem 3.1.
LEMMA 6.7. Let 0'. and A be such as in Theorem 6.5, and let Ml , M2, ...
be a disjoint partition of M\{O} such that AM < 00 , n EN.n For each
n ( N , we further assume that vn is a Poissonian random variable with
mean AM and thatn
common distribut'ion
dependent, we have
M VAM •n n
are random measures on Mh with the
Letting all these random elements be in-
(16) I(a,A) d a +
PROOF. The right-hand side of (16) has the L-transform, evaluated at
an arbitrary f E F ,
-80-
00_afe II
n=l
-AM 00
e n ~
k=O
00
e-a.f IIn=l
00
-af= e II
n=lexp{ - (M A) (1
n
-af= e exp{ -A (1-'IT
exp{-a.f - A(l _ e f)}
and by Theorem 6.5, this agrees with the L-transform of 1(a,A) . Thus
the assertion follows by Lemma 1.3 and Theorem 3.1.
The last two lemmas may be used to establish various relationships
between an infinitely divisible random measure ~ and the corresponding
canonical measures a and A. We shall give two results of this type.
~ g *THEOREM 6.8. Let 1(et,A) on S and let a > 0 Then ~a = 0
* * "A{ll{S} > o}a.s. iff (i) a = o ., (ii) A{l-! ;t O} = 0 ., and (iii) = o .,a a
s E S . In pa:I'ticular., ~ E Md a.s. iff a E Md and A~ = 0 •
*PROOF. Suppose that ~a = 0 a.s. By Lemma 6.7 we obtain (i), and
furthermore, *( ~.) = 0 a.s. whenever AM > O. But thensnJ a n
M A{l-! ;t O} = 0, n EN, and summing over n yields (ii). To proven a
(iii), suppose conversely that ~A{l-!{S} > E} > 0 for some n EN,
s E Sand E > O. Then we get the contradiction
P{~: ;t O} ~ p{~{S} ~ a} ~ p{Vn > [a/E] , ~nj{s} > E , j S[alE] +l} =
• P!Vn > [alE)}! P{Enj!S} > Ellla/EJ.1 > 0 •
-81-
so (iii) must indeed be true.
Conversely, suppose that (i) - (iii) hold. From (ii) it follows
that a.s. whenever AM > 0 , so it suffices to prove that,n
wi th probability one, no two terms in (16) can have atom positions in
common. But this follows by Fubini's theorem from the fact, implied by
(iii), that ~ .C = 0 a.s. for any countable set C c S .nJ
To state the next result, let SeA) denote the support of a measure
A on M w.r.t. the vague topology in M. Moreover, given any MeM,
define GM as the additive semigroup in M generated by M u {a} , and
write a = min M whenever a E M and a ~ ~, ~ EM.
THEOREM 6.9. Let dt; = l(a,A) on S. Then
(17)
(18) GS(A)-1= S (P t; ) - Ct •
PROOF. To avoid trivialities, we may assume that A ~ O. Let us first
consider the case of bounded A By Lemma 6.7, we may then assume that
t;. are independentJ
A/AM. From (19) it
d v(19) ~ a + L t;. ,
j=l J
where v is Poissonian with mean AM while the
of v and mutually independent with distribution
-AMis seen that ~ ~ a a.s. and P{t; = a} = e > 0 , proving (17). Since
addition and subtraction of Ct are continuous operations on M and
~82-
M + a respectively, we further obtain - a
and so it suffices to prove (18) in the case a = 0
Suppose that ~O E GS(A)\{O} , say ~O = ~l + ••. + ~k ,where kEN
and ~l' ... '~k E SeA) By the continuity of additien in M, there
Since Sep~-l) is closed, this proves that
exists for every E > 0 some <5 > 0 such that
P{p(E,;,\10
) < d ~ P{v = k , P(~j , ~. ) < 0 , j = 1, ... ,k} =J
k= p{v = k} II P{p (~ . , ~. ) < o} =
j=l J J
-AM (AM)k k= e k! II A{p ell, \1') < o} / AM > 0 ,
j=l J
proving that ~O E S(p~-l)
GS(A) c S(p~-l) .
Conversely, suppose that 110 E S(p~~l) . For n EN, let
obtained as ~ in (19) but with v replaced by v An. Then
a.s., so we get
(20) o < P{p(~,~O) < £} ~ liminf P{p(nn' llo) < £}n~
£ > 0 ,
and hence there exist some n EN, with If
we can show that lln E GS (A) , n € N it will follow that \1 0 E GS(A) ,
proving the relation S(p~-l) c GS(A) Keeping n E N fixed, we thus
assume that-1
Then there exists form E S(Pn ) every pEN somen
k E {O,l, ... ,n} satisfying
(21)
-83-
-1 .o < P{p(~, m) < p , \I = k} t:: P{P(~l + ... +-1
~k' ID) < P , \I = k}
-1= P{P(~l + ••• + ~k' m) < p }P{\I = k} ,
and since the set of possible k-va1ues is non-increasing in p, we can
choose k independent of p. Moreover, k = 0 implies m = 0 E GS(A) ,
so we can further assume that k ~ o. Let us now define the sets
Since
k -1Ap = {( 11 1 , ... , llk) EM: p (11 1 + ••• + llk' m) < p }
is separable in the product topology, each Ap
by countab1y many sets
pEN .
can be covered
j = I, ... ,k}
with (m1, ... ,mk) E Ap , 80 by (21) there must exist some vector
,mk ) E A satisfyingp p
(22) -1o < P{p(~, m. ) < pJP
kj = I, ... ,k} = II
j=l
-1P{p(~., m. ) < P }.
J JP
From vwe get m1p + ... + mkp ~ m as p ~ 00 , so
1imsup m. f ~ lim (m1 + ••• + mkp)f =~oo JP ~ P
mf < 00 f E Fc j = 1, ... ,k ,
showing that the sequences {mlp}' ,{mkp } are relatively compact. We
may therefore choose some sequence N' c N and some mI , ... ,mk E M
such that v (p EN') j 1, ,k It follows in particularm. ~ m. , =JP J
that v (p E N' )m1p + ... + mkp ~ m1 + .. . + mk , so m = ID1 + ... + mk
Let us now fix j E U, ... ,k} and E: > 0 , and choose pEN' so large
-84-
that p (m. , JD.) < £/2JP J
and -1P < £/2 . By (22), we then obtain
-1P{p(l;., m.) < d ~ P{p(~., m.) < E:/2} ~ P{p(~., m.) < p } > 0J J J JP J JP
so we get This shows that m E: GS(A) and hence
completes the proof in the case of bounded A .
In the case of unbounded A we still have ~ ~ a a.s. by Lemma
6.7, and if (18) is true, we get in addition -1a E: S(P~ ) , proving (17).
It thus remains to prove (18), and again we may assume that a = O. For
n E: N .~kj ,
n
Lk=l
~n =
as in Lemma 6.7, define
vk
Lj =1
=
~nj
An
and
Thenv
~n + ~ a.s., and moreover,
(23) n E: N ,
since the A are bounded and form a non-decreasing sequence. Supposen
that ~O E S(p~-l) Proceeding as in (20) , it is seen that there exist
S(p~-l) vsome ~n E: n E: N , with ~n + ~ , and since ~l' ~2' ... E: GS (A)n
by (23), it follows that ~O E: GS(A) .
Conversely suppose that ~O E: GS (A) , say ~ = ~l + ... + ~k forasome k E: Z and ~l' ... '~k E: SeA) Here we may assume without loss
+
that ~. ~ a for each j E: U, . .. ,k} , say that ~.f. ;t a where f. E: FJ J J J c
As M1 in Lemma 6.7 we can choose
k 1Ml = U {~ E M: ~f. >2~·f.}j =1 J J J
-85-
since by Theorem 6.5
k1
k-1 1
AMI ~ L A{~f . > '2 ~.f.} c: I AlT f. (2~·f., 00) < 00j=l J J J j=l J
J J
The set Ml being open and containing ~l' ,]1k ' we get
and hence by (23)
(24 ) n € N .
Let £ > 0 be arbitrary, and choose 0 € (0, £/2) such that
(25) p(]1, ]10) v p(]1I, 0) < 0 =? p(]1 + ]1', ]1) < £/2 ,
which is possible since the mapping (]1, ]1') + p(j1 + ]1',]1) is continuous.
Since ~ - ~ X0 a.s., we may further choose n € N such thatn
(26) P{p(~ - ~ , 0) < c} > 0 .n
By (24) - (26) and the independence of t" and':on ~ - ~n ' we obtain
= P{p(~ - ~n' 0) < o}P{p(~n' ]10) < &} > 0 ,
and since £ was arbitrary, it follows that
This completes the proof.
-1]10 € g(P~ ) , as desired.
-86-
NOTES. The canonical representations occurring in Lemmas 6.1 and 6.3
are classical, (see e.g. Feller (1968) and (1971)), while the one in Theorem
6.5 is due to Matthes (1963) and Lee (1967), (see also Mecke (1972)). The
corresponding convergence criteria are due to Kallenberg (1973a), (though
the proof of Lemma 6.1 follows that of Feller (1971) for general random
variables). As for Lemma 6.4, the first part is given in Kerstan, Matthes,
and Mecke (1974), page 49, while the second part is new. Lee (1967) gave a
proof of Lemma 6.7 along with a heuristic argument for Lemma 6.6; the
presentproof was supplied for point processes by Jagers (1973). Finally,
Theorem 6.8 extends a result for point processes by Kerstan, Matthes and
Mecke (1974), page 95, (see also Kallenberg (1973a)), while Theorem 6.9
is new.
The reader should consult Kerstan, Matthes and Mecke (1974) for
a different approach to the theory of infinitely divisible point processes,
where (16) plays the role of a basis for the whole theory while (14) and (16) fit
into their general theory of "showerfields".
PROBLEMS.
6.1. Prove the Z+-version of Lemma 6.1 directly by means of generating
functions.
6.2. State and prove tightness criteria corresponding to the convergence
assertions in Lemma 6.3 and Theorem 6.5.
6.3. dA ) n E N in Lemmas 6.1 and 6.3 and in TheoremLet F;, = l(a ,n n' n
6.5. Give criteria for convergence in distribution of {F;,n} in
terms of {a } and {A }n n
6.4. Assume that S is countable. Show that there exists some fixed
f E F such that, for any point process F;, on S with F;,S < 00
a.s., F;, is infinitely divisible iff F;,f is infinitely divisible
-87 -
vanishing constant component. (Hint: use the result in Problem
3.8.)
6.5. Let E, be a point process on S. Show that E, is infinitely
divisible iff E,f is infinitely divisible with P{E,f = O} > 0
for every f E Fc
6.6. For E, = I(O,A) on S with AM < 00 , state and prove a relationship
between the set of atom positions of pE,-l and that of A. Show
also that pE,-l can have no atoms if AM = 00 (Hint: Let f E F
be strictly positive and such that E,f < 00 a.s. Then-1ATI f R~ = A{~f > O} = AM , so the last assertion follows from the
corresponding fact for random variables due to Blum and Rosenblatt,
see Lukacs (1970), page 124.)
7. FURTHER RESULTS FOR NULL-ARRAYS.
In this section we shall see how the results of Section 6 can be
improved in certain special cases. Let us first consider the case of
random measures with independent increments, (sometimes called completely
random measures). By this we mean random measures E, such that
[,B l , ... ,E,Bk are independent for arbitrary disjoint Bl , ... ,Bk E B ,
kEN . Let us say that a measure ~ E M is degenerate if ~ = ho s
for some b;. 0 and s c S .
LEMMA 7.1. Let E, ~ l(a,A) on S. Then E, has independent increments
iff A is confined to the class of degenerate measures.
-88-
PROOF. Suppose that A has the asserted property, and consider arbitrary
kEN and disjoint Bl ,
for any t l , ... ,tk E R+
,Bk E B. Writing M' = M\{O} , we obtain
- log E exp(- L t.~B.)j J J
t.nB.J J
(1 - exp(- L t.llB.))A(dll) =j J J
= L t.aB. + L f ( 1 - exp (- t j lJ Bj )) A(dII ) =j J J j
lJB. >0J
= I {tjaB j + J (1 - exp(- t. lJB.))A(dJ1 )} =j J J
M'
= - log IT E exp(- t.~B.) ,j J J
proving that ;Bl , ... ,;Bk are independent. Conversely, suppose that
; has independent increments, and let A and B E B be disjoint. Since
a is additive, we obtain as above
(1) o .
But the function f(x) :: 1 - -xe is strictly concave on R , so+
f(x + y) < f(x) + fey) for every positive x and y, and hence (1)
is only possible provided A{J1A > 0 , )lB > O} = O. Making a partition
of S into countably many sets of diameter < € , it follows that, for
)l E M' a.e. A, at most one of these sets can have )l-measure > 0
Since € can be chosen arbitrarily small, this proves that lJ is degenerate
a.c. A.
-89-
THEOREM 7.2. A random measure ~ on S is infiniteZy divisibZe with
independent increments iff
(2) -H flog Ee = af +
R txs+
f E: F ,
for some a E: M and some y E: MeR t x S)+
with y(o x B)K < ~ , B E: B ,
and in this case a and yare unique. Moreover, any a and y with
the stated properties may occur in (2).
random measures on S and if I c B~
iff
If {~ .} is a nuZZ-arPay ofnJ
dis a DC-semiring, then I·~ . + ~] nJ
(i) \' -1 wK L . P (~ . I) + a 10
0+ Ky (. X I) ,
] nJI E: I , and
(ii) I·p{~ .1 1 A ~ .1 2 > £} + 0 for every £ > 0 and disjoint] nJ nJ
11' 12 E: 1 .
In the point process case we have a = 0 whiZe y 1.-S confined to N x S
Tn this case it suffices to take £ = 0 in (ii) and to repZace (i) by
(i) , I -1 w y (. I) I.P(~ .1) + x on N , I E:] nJ
PROOF. Suppose that ~ is infinitely divisible with independent increments.
Let M E M be the class of degenerate measures, and for ~ EM, let t be~
the unique atom position of ~. From Lemma 2.1 it is seen that M E: M and
that is a measurable mapping of M
Then (2) is
~ -> (~S, t )~
be the measure on R' x S+
induced by
into R t X S .+
A via this mapping.
Let y
obtained by transformation of the integral in (6.6), and since
(3)
-90 ..
y(dx x B) = A{~B € dx} x > 0 B € B ,
we have -1y(o x B)K = (A~B )K < m B € B . To see that y is unique,
note that the inverse mapping (xJt) + xO t transforms y into some A
and (2) into (6.6). From the above mappings it is further seen that any
a and y with'the stated properties are possible in (2).
As for the convergence assertion, the necessity of (i) and (ii)
follows easily from (3) and Theorem 6.5. Conversely, suppose that (i)
and (ii) hold. Then (6.9) follows from (i), and it reamins to prove (6.8).
It is then enough to take the components II' ... ,Ikof V in (6.8)
disjoint and to verify that
k k(4) L p n H: .r. ~ t.} + A n {~I. ~ t.}
j i=ln] 1 1 i=l 1 1
for any t = (t l , with t. = 0 or A{~I. = t.} = 0111
i = 1, ... ,k . Since events corresponding to t. = 01
can be omitted
from (4), we can assume that all the t.1
are > 0 But then (4) follows
from (i) for k = 1 and from (ii) and Lemma 7. 1 for k·~ 2. The last ,
assertion follows from the corresponding fact in Theorem 6.5.
The most general random measure with independent increments is obtained
from the one in Theorem 7.2 by addition of at most countably many fixed atoms:
COROLLARY 7.3. A random measure n on S has independent increments
iff it can be written in the foY'/11
(5)k
n = t;. + L S.Otj=l ] j
wheT'e ~ satisfies (2) faT' some
-91-
a and y with a{s} = y(R' x {s}) = 0 ,+
S E S , some k E Z and distinat t. :s; S .. j :s; k .. and some R -valued+ J +
T'andom va:fliables S. with P{S. > a} > 0 , j :s; k J whiah aT'e mutuallyJ J
independent and independent of ~ In this aase .. the above deaomposition
is unique apaT't fT'om the oT'deT' of terms.
PROOF. The sufficiency and uniqueness assertions are obvious. Next suppose
that 11 has independent increments. Like any random measure" 11 can have
at most countably many fixed atoms, i.e. I1{S} = a a.s. for all but countably
many s E S , (cf. the argument in Billingsley (1968), page 124). Let t.J
and S· ,J
j :~ k , be the corresponding atom positions and sizes, and let
~ be defined by (5). Since 11 has independent increments, it is easily
seen that the Sj are mutually independent and independent of ~,and
further that ~ has independent increments. Moreover, ~ has no fixed
atoms, so if {B .} c B is a null-array of partitions of some fixed setnJ
B E B , it follows by a simple compactness argument that {~B .}nJ
is a
null-array of random variables. Hence, by Lemma 6.1, the common row-sum
~B must be infinitely divisible. A similar argument shows more generally
that (E:Bl , ,~Bk) is infinitely divisible for any kEN and disjoint
Bl , ... ,Bk E B , and so we may conclude from Lemma 6.4 that ~ is itself
infinitely divisible. In view of Theorem 7.2, this completes the proof.
By comparison of (2) and (1.2), it is seen that a point process
~ on S is Poissonian iff it is infinitely divisible with independent
increments and with the measure y in (2) confined to the set {I} x S
-92-
In this case, ~ has intensity
E~B = yeO} x B) B E B •
Specializing Theorem 7.2 to this case, we obtain the important
COROLLARY 7.4. Let ~ be a Poisson process on S with intensity A EM,
and let {~nj } be a nuU-arTay of point prooesses on S Further suppose
that I c BA is a DC-semiring . Then I.~ . ~ ~ iffJ nJ
(i)
(ii)
2'P{~ .1 > O} + AIJ nJ
I.p{~ .B > l} + 0J nJ
IEI,and
B E B •
We are now going to consider the convergence of null-arrays towards
two special types of infinitely divisible random measures which play the
same role here as do the simple point processes and diffuse random measures
in Sections 3 - 5. But first we define regularity for canonical measures
A and for null-arrays {~nj} as follows. Given any covering class
C c B and any a > 0 , we shall say that A is a-regular w.r.t. C
if there exists for every C E C some array {Cmk } c C of finite coverings
of C such that
(6) lim 2 A{lJCmk ?: a} = 0 .]}t+OO k
The notion of a-regularity of {~nj } is defined analogeously but with
(6 ) replaced by
-93-
lim limsup L L P{~ .C k ~ a} = 0~ . k nJ mUl'- n-+oo J
"
As before, regularity means a-regularity for every a > 0 .
be a null-array of point processes on
is
S and
u c B~Further suppose that
Then \ ~ ~ E, if.'~Lj nj Ja DC-semiring.I c B~
Let {~ .}nJ
*with A{~ ~ ~ } = 0 .
THEOREM 7.5.
let [, ~ 1 (A)
a DC-ring and
(i) lim L P{E, .U > O} = A{~U > O}. nJn-+<oo J
U E U ,
(ii) limsup L P{E, .1 > I} ~ A{~1 > I} 1 < I ,n-+oo j nJ
(iii)
Moreover,
lim limsup L P{~ .B > k} = 0 B E B .k-+oo n-+oo j nJ
dLj~nj + ~ follows from (i) and (ii) if A &S 2-regular w.r.t.
I and from (i) alone if {E,nj} is 2-regular w.r.t. I or if
(7) limsup L EE, .1 ~ A~I < 00
n-+oo j nJI E I .
The proof of this theorem is similar to that of Theorem 4.8, except
that we now rely on the following two lemmas, playing the roles of Theorem
3.3 and Lemma 4.7 respectively.
LEMMA 7.6. Let A be the canonical measure of an infinitely divisible
point process, and suppose that U c B is a DC-ring and I c B a DC-
* *cerm:-ring. 7'hen A{~ ~ ~ } = 0 iff A{~1 > l} :0; A{ll 1 > l}, 1 E I ,
and in that case A &S uniquely determined by A{~U > O} U E U •
-94-
PROOF. For fixed e E U , let Ae be the measure on N' = N\{O} induced
by A via the mapping ~ + e~ , i.e.
MEN n N' .
Then
so either this quantity is zero, or Ae can be normalized to a probability
measure on N' , i.e. to the distribution of some simple point process.
Hence, by Theorem 3.3, I.e is completely determined by the quantities
U € U ,
Le. by all U E U But AC
determines the measure
for any f E F with support in C so by Theorem 6.5, the family {AC}
determines A. We may now complete the proof as in case of Theorem 3.3.
LEMMA 7.7. Let {E,;nj} be a nuZZ-array of random measures on S , and
thatd Further suppose that d
I (a,A) andsuppose I.E,; . + som~ n· E,; =J nJ
that U c B '[.,s a DC-ring satisfying
(8)-tE,; .U -tnuliminf I (1 - Ee nJ ) ~ taU + 1.(1 - e )
n-+oo jU EO U ,
for some t > 0 Then nC = 0 a.s. for every compact C E B with
~C = 0 a.s. In the point process case, (8) can be repZaced by
1iminf I P{~ .U > O} ~ A{~U > O}n-+oo j nJ
U t: U .
THEOREM 7.8.
on S and let
-95-
This may be proved in the same way as Lemma 4.7.
We now turn to the analogue of Theorem 4.9.
Let nnj} be a nuU-arTau of point processes (random measures)
~ ~ I (a,A) with a = 0 and A{}l;t /} = 0 (or' with AMc = 0d
respectively). Further suppose that sand t are fixed real numbers with
semi-ring) .
a covering class (DC-o < s < t
(i)
(ii)
and that U c BI; is a DC-ring and I c BI;
Then \ I; ~ I; if.fL.j nj
-tl; .U -twlim L (1 - Ee nJ ) = taU + A(1 _ e U)n-+oo j
-sl; .1 -STI
limsup r (1 - Ee n J ) ~ sal + A(l _ e I)n-+oo j
U E U ,
I € I ,
I E I .(iii)
Moreover.,
lim limsup I P{~ .1 > t} = 0t+oo n+oo j nJ
r·t; . ~!; follows from (i) alone if {!;nJ'} ~s 2-regular (regular)J nJ
w.r.t.
(9)
I or if
limsup I EI; .1 $ aC + ATII
< 00
n-+oo j nJI E I .
PROOF. We confine our attention to the random measure case, since the
point process case is similar but simpler. Suppose thatd
\.1; . + !;L. J nJ
and let B E U or I Then L.I; .B &!;B by Lemma 4.3, so by Lemma 6.1J nJ
This implies (i) and (ii), while (iii) follows by Lemmas 4.6 and 6.2.
-96-
Conversely, suppose that (i) - (iii) hold. Since -tx(1 - e )/x
is a decreasing function of x > 0 we have
and so we get for any r > 0 and U E U
o :-:; x :0; r ,
By Lemmas 4.6 and 6.2, it therefore follows from (i) and (iii) that {LjSnj}
is tight. But in that case every sequence N' c N has a subsequence Nil
by Lemma 7.7, it follows from (i), (ii) and
dsuch that Los, + some n
J nJ
6.5, and since U u I c Bn
(n E Nil) Now dn = some by Theorem
the necessity part of the present theorem that
-t7T -t7TtaU + ~ (1 e U) = telU + 1.(1 e U) U E U ,
-S7T -S7Tsal + ~(l - e l) :-:; sell + 1.(1 - e l) I E: I ,
which by Theorem 6.5 is equivalent to
(10) -tnU -tsU U E U ,Ee = Ee
(11) -snl -ssl I E: IEe ~ Ee
Furthermore, it follows from Theorem 6.8 that s is a.s. diffuse, possihly
apart from atoms of fixed size and location which arize from the atoms of
(1. Let us now consider any fixed XES Choosing Bl , B2, ... in U
or 1 respectively such that Bn ~ {x} , it follows by monotone convergence
-97-
from (10) and (11) that
(12) Ee-tnJx} Ee -tHx} -ta{x}= = e
(13) Ee-sn{x} :<: Ee-s~{x} -sa{x}= e
If n{x} were non-degenerate, it would follow from Jensen's inequality
for strictly convex functions that
(e -sa{x}) tIs = -ta{x}e
which is impossible. Hence n{x} must be degenerate, and by (12) we
obtain n{x} = a{x} a.s. Writing a' for the purely atomic part of
a , it follows that n:<: a' a.s., and since (10) and (11) may be written
in the form
(14 )-t (n-a')UEe = Ee-t(~-al)U U E U ,
Ee-s(n-a')I :<: Ee-s(~-al)I I E I ,
it from Theorem 3.4 that ~ ... a' d - a' i.e. that dis seen n , t;. = n
This shows that Ij~nj ~ ~ (n E Nil) , and since N' was arbitrary, even
that Lj t;.nj ~ t;. (n E N) , proving the first assertion.
The assertion involving regularity of {~nj} follows by the usual
arguments, so it remains to prove the assertion involving (9). In this
case we get by Fatou's lemma
(15) EnI :S liminf L E~ . ImiN" j nJ
:S limsup L E~ .1 ~ aI + AnI = E~In-+oo j nJ
I E I ,
-98-
replacing (11), so (13) is now replaced by
(16)
If n{x} were non-degenerate, we would obtain from (12), (16) and Jensen's
inequality
-ta{x}~ e ,
which is impossible, so again we get n{x} = a{x} a.s., x € S , proving
that n ~ a' a.s. Thus it follows by (14), (15) and .Theorem 1.4 that
~ - a' g n ~ a' , and the proof may be completed as before.
NOTES. The canonical representation in Theorem 7.2 and Corollary 7.3,
which is a well-known classical result for random measures on R, is
due in the present generality to Kingman (1967). The present approach
via Lemma 7.1 a~ well as the convergence criterion in Theorem 7.2 were
given in Kallenberg (l973a). Corollary 7.4 is a classical reSUlt, proved
'" "in particular cases by Palm (1943), Hincin (1960) and Ososkov (1956), and
in general (on R) by Grige1ionis (1963), (see also Goldman (1967) and
Jagers (1972)). Finally, Theorem 7.5 is taken from Ka1lenberg (1973a)
while Theorem 7.8 is new.
PROBLEMS.
7.1. Show in analogy with Lemma 6.6 that the random measure ~ in Theorem
7.2 is distributed like a + Jxotn(dxdt) , where n is a Poisson
process on R' x S with intensity y+
7.2.
7.3.
7.4.
-99-
Let {~nj} be a null-array of point processes on S. Show that
I.I;. is tight with only Poissonian limit distributions iffJ nJ
\.P{I; .B > O} is bounded and \.P{c .B > I} + 0 for every B E B .L J nJ L J snJ
State and prove a similar tightness criterion corresponding to the
convergence assertion in Theorem 7.2.
Prove Corollary 7.4 directly from Theorem 4.8 and Lemma 6.1 in the
particular case when A E Md
7.5. Prove Theorem 7.8 in the point process case. Show that I may
be assumed to be a covering class even in the random measure case,
provided a is known to be diffuse.
7.6. Let ~ ~ lCa,A) on S , let C c B be a covering class and let
a > 0 Show that A{ll ;t O} = 0 if A is a-regular w.r.t. Caand that the converse is also true provided C is a DC-semiring
and E~ E M
7.7. Let {~nj } be a null-array of random measures on S such that
I.~ . §. some *I (a, A) Show that A{ll ;t O} = 0 if {~nj } isJ nJ E:
a-regular w.r.t. B~ for every a > E: , and that the converse is
also true provided E~ E M.
8. COMPOUND AND THINNED POINT PROCESSES.
The notions of compounding and thinning were introduced in Section
1, and in Corollaries 3.2 and 3.5 we proved two uniqueness results for
compound point processes. The aim of the present section is to study
the convergence in distribution of such random measures, and we shall
-100-
first consider a case when the limits are themselves compound point processes.
THEOREM 8. 1• For n E Z+ ' let i:n be a 8 -compound of nn . Thennd
andd imply that
dConversely" suppose thatSn -+ 130 nn -+ nO ~n -+ ~o
d and that {nn} is tight. Then ~ is a So-compound ofE, -+ Bome E,n
no for some 80 and nO ' and if {Cnn } is 2-reguZar w.r.t. B~ for
some C E B with l;C ~ a"
there exist some constants PO' PI' ... E (0,1]
with inf 0 such thatd
8 ' and d , where ~p = Pn > 13' -+ n' -+ nO " 1-S a Pn-n 0 n
thinning of nn' n E Z+ ' while
(1) n E Z+
PROOF. Let us first suppose that
be a 13n-compound of nn' n E Z+
is then equivalent to
sn ~ 80 and ~ ~ nO ' and let l;n
dBy (1.5) and Theorem 4.5, l;n -+ £'0
(2) E exp(nnlog ~n 0 f) -+ E exp(nOlog ~O 0 f) f E Fc
where n E Z+ ' and by continuity and boundedness, (2) will
follow if we can prove that
(3) f E Fc
By Theorem 5.5 in Billingsley (1968), it 1S enough to prove (3) for non-
random nn' i.e. to show thatv
lln -+ 110 implies
(4) f E Fc
Considering a fixed
-101-
f E F ,we may assume thatcw
]In -+ ]JO ;t 0 (cf. Problem
4.5), and after a suitable normalization that the ]In are probability
measures. But then (4) may be written in the form
for some random elements in S with Since
f is bounded, the sequence {log </> 0 f}n is uniformly bounded, and hence
(5) is a consequence of the relation
(6)
Applying Theorem 5.S in Billingsley (1968) once more, it is seen that
the quantities Tn = sn may be considered as non-random with sn -+ So '
and since in that case xn = f(sn) -+ f(s) = x , it remains to prove that
whenever But this follows from the fact
that the convergence </>n -+ </>0 is uniform on bounded intervals.
Let us now suppose conversely that s ~ s and thatn
is tight.
Since 0 E M is a O-compound of 0 E N , we may assume that s ~ 0 Let
the sequence N' c N be such thatd (n EN') If
d0nn -+ some n n ,
we obtain for any fixed B E B
so by (1.5) and bounded convergence
d~ n B -+ 0 ,
n
-s BEe n = E exp(n B log </> (1)) -+ 1 ,
n n
-102:-
and hence F,B ~ 0, B E B , yielding the contradiction ~ = 0 a.s.
Thus n ~ 0 .
We shall use this fact to show that {Sn} is tight or, what amounts
to the same, that
(7) lim limsup (1 - ~n(t)} = 0 .t-+O n-+oo
Suppose on the contrary that the limit in (7) is greater than some E > a .
Then there exist arbitrarily small numbers t > a such that, for some
sequence N" eN' ,
(8) 1 - ~ (t) > En n E N" .
Choosing B E B~ n Bn
such that nB ~ 0 and noting that
(9) 1 - x ~ - log x ~ 2(1 - x)
it follows by (1. 5) and (8) that
-tl; B -En Bn E exp(nnB log ~n(t)) E exp[-nnB(l - ~n(t)}] Ee nEe = ~ ~
and I; B ~ I;B andd
obtain -t!;B -EnBsince ~B-+ n.B , we Ee ~ Ee Lettingn
A h th d·· 1 Be-En.B < 1 . h (7)·t -+ , we reac e contra lctl0n ~ , provIng t at IS
indeed true. Thus the sequences {Sn} and {nn} are both tight, and
in particular we may consider convergent subsequences and use the direct
part of the theorem to conclude that I; has the asserted form.
Let us now suppose that {C~} is 2-regular w.r.t. BI; for some
C E B~ with I;C ~ 0 • and assume that S ~ S' and n i n' as n -+ 00
'-> n n
-103-
through N' c N Since E; is then as' -compound of n ' , and moreover
S' ~ 0 it follows from (1. 5) that B = BE; and so Cn' must ben 'simple (cf. Problem 4.9) . Put P = P{S' > O} , let nO be a p-tFtinning
of n' and let So be an R+~valued random variable satisfying
(10) -1P{SO > x} = p P{S' > x}
Then ¢ = LS
and cP ' = LS' satisfy the relation cPt :: P<PO + (1 - p) ,00
so letting E;O be a So -compound of nO and using (1. 5) and (1.6), we
obtain for any f E F
LF,; (f)o
= L (-log cPO 0 f) = Ln , (-log[lnO
pel - <Po 0 f)J) =
= L ,T-log[l - (1 - <P' 0 f)]) = L ,(-log <p' 0 f) = Lt"(f) ,n n ~
and hence F,; g F,; by Theorem 3.1.o Now P{So > O} = 1 , and moreover
is simple by Problem 2.4, so it follows from Corollary 3.5 that
and are uni~uely determined by pF,;-l
Choose any fixed y > 0 satisfying N80 > y} > 0 and NsO = y} = 0 ,
and define p = P{s > y} , n E Z+ Then PO > 0 , and moreovern n
liminf p > () since if Pn -+ 0 as n -+ ClO through some sequence N' c N ,n n
we could choose some further subsequence Nil c N' such thatd
8n
-+ some S' ,
and then (10) would hold with p = P{8' > O} > 0 , and we would get the
contradiction
P = P{S > y} -+ P{S' > y} = P{B' > O}P{SO > y} > 0n n
(n E Nil) .
-104-
To obtain p = inf p > 0 , we may replace the finitely many zeros inn
the sequence {Pn} by arbitrary positive numbers. We can now let ~
be a Pn-thinning of nn' n E Z+ ' and define random variables 8 'n
satisfying (1).
Given any sequence N' eN, we may choose a subsequence Nil c N'
In this case the constant
such that d8 + somen 13' ,
p
dn + some n' andn
in (10) equals
p + p' = P{ 13' > y} (n E Nil) .n-1
p'PO ,and in particular
Using the direct part of the theorem, it is seen-1= p'PO
dNil n' +, 'Inthat, on no ' a p'-thinning of n' . On the other hand, it
was shown above that n is a (p' po1)-thinning of n' , and so nO is
dlike n' a po-thinning of n , and we get n' + n' (n E Nil) . Furthermore,o n 0
it is seen from (1) that 13' ~ 8" (n E Nil) , wheren 0
P{13' > O}
-1 -1 -1P{I3H > x} = pp' P{I3' > x} = pp' p'Po P{8
0> x} = -1
PPO P{13 0 > x} =
= N130 > x}
so 13" ~ 13' and it follows thato 0 13' ~ 8'n 0
(n E Nil) . In view of Theorem
2.3 in Billingsley (1968), this completes the proof.
We are now going
compounding variables
to consider the entirely different case when the
dI3n satisfy 8n + O. The class of possible limits
may then be described as follows.
Consider arbitrary a E R and A E M(R' ) with AK < 00 , and define+ +
-IT
(11) 1/J (t) = at + A(1 _ e t) t E R+
-105-
By Theorem 7.2, there exists for every Y E M some random_measure ~
with homogeneous (w.r.t. y) independent increments satisfring, for
any f E F ,
log Ee -~f = oJ.lf + J (l - e -Jef (t)) (A x )l) (dJedt) =R'xS
+
= t {af(t) + JR
, (1 - e-xf(t))A(dx)}y(dt) = )l(t/J 0 f) .
+
By Lemma 1.4, we may consider )l = n as a random measure and mix w.r.t.
its distribution, thus obtaining
(12) L (t/J 0 f)n
f E F .
Note in particular that, if a = 0 and A = 01 ' then ~ is a Cox process
directed by n. We shall need the following uniqueness result which is
similar to Corollary 3.5.
LEMMA 8.2. Let ~ be a random measure on S satisfying (11) and (12)
for some n ~ a and A ~ and suppose that a + AK = 1 whiZe Cn is
diffuse for some C E B wi th nC ~ 0 •
uniqueZy determined by p~-l.
PROOF. By (12),
Then -1Pn ~ a and A are
B E B ,
and replacing B by B n C , it follows by Theorem 3.4 that P(Cn)-l
is uniquely determined by p~-l. But then it follows as in the proof
-106-
of Corollary 3.5 that even ~ is uniquely determined, and by Lemma 6.1,
~ determines a. and A . The uniqueness of -1Pn now follows as in
case of Corollary 3.5.
To avoid trivial complications, let us assume that any compounding
dvariable B is such that f3 ~ 0 , and further that the a. and A in
(11) are normalized in such a way that a. + AK = 1 .
THEOREM 8.3. For n E N ~ let ~n
Suppose that Bn
i 0
be a Bn-compound
Then ~ 5! 0n
of ~ and put
"'f'+' d 0v J c n +n n
Purthermore~ the conditions
(i)
(ii)
dc n + n ,n n
imply thatd
~ + somen ~ satisfying (11) and (12). ConverseZy~
implies that ~ satisfies (11) and (12) for some n ~ a. and A ~ and
if moreover
then -1Pn ~
{c Cn }n n
a. and
is regular w.r.t. B~ for some C E B~
A are unique and satisfy (i) and (ii).
with ~C ~ 0 ,
PROOF. Writing ~ = Ln Bn
n EN, we get as before
-t~ B(13) Ee n = E exp(nnB log ~n(t)) B E B n EN.
Now B ~ 0n
large n E N
implies that ~ + 1 , so by (9) we get for sufficientlyn
-107-
and it follows from (13) that
2cn
-~ BE exp(-2cnnnB) ~ Ee n ~ E exp(-cnnnB) B E B ,
showing that
assertion.
r B ~ 0~n
iff dc n B + 0 •n n B E B • This proves the first
Next suppose that (i) and (ii) are satisfied. Proceeding as in the
proof of Theorem 8.1, we obtain ~ ~ ~ with p~-l determined by (11)n
and (12), provided we can show that
whenever x + x .n
But this follows from the
fact, essentially being proved in Lemma 6.1, that -1-cn log ~n tends to
W uniformly on bounded intervals.
Suppose conversely that ~n ~ some ~, and let B E B~ be arbitrary.
Then !; B ~ ~B by Lemma 4.3, so by (13)n
and since this limit tends to one as t + 0 , there exists for each E > 0
some t E (0,1) satisfying
(14) E exp(n B log ~ (t)) ~ 1 - E/2n nn EN.
Now it follows from (9) and the elementary inequality
that
1 -
-108-
-tx xe ~ tel - e- ) t E [0,1]
-tS- log ~ (t) ~ 1 - ~ (t) = E(l _ e n} ~n n
'" '"and so by (14) and Cebysev's inequality
-6tE (1 _ e n) = tc
n
P{ c n B >-1 2} P{tc n B > log 2} P{-n B log ~ (t) > log 2}t log = ~n n n n n n
= PO - exp(n B log ~ (t)) > !.} ~n n 2
~ 2EO - exp(n B log ~ (t))} ~ £ ,n n
proving tightness of {c 11 B}n n Since B E B~ was arbitrary, it follows
by Lemma 4.6 that {c n }n n
is tight.
Let us no\'l assume in addition that ~ 2 o. Proceeding as in the
proof of Theorem 8.1, we can then prove that
(15) lim limsup c-1(1 - ~ (t)) = a .t~O n+oo n n
V \IBy Cebysev's inequality, we further obtain for any r,t > a
-S tc-lE(l-e n) c-l(l_~ (t)}
> r} ~ _n = _n n _l_e-rt l_e-rt
and letting in order n +00 r ~ 00 and t ~ a , we get by (15)
lim limsup c-lp{S > r} = a ,n n
r~oo n~oo
which shows that the sequence {c-lK(pS- l )} of probability measures onn n
R is tight. In particular there exist some sequence N' ~ N and some+
-109-
n, a and A such that (i) and (ii) hold on Nl , and we may conclude
from the sufficiency part of the theorem that ~ has the asserted form.
(In the case ~ ~ 0 we may take n = 0 and choose arbitrary a and
A .)
Let us finally assume that is regular
w.r.t. B~ for some C E B~ with ~C 2 O. Since the sequences {cnnn}
and {c-lK(PS- l )} are both tight, any sequence N' c N must contain an n
subsequence N" c N' such that (i) and (ii) hold on N" for some n,
a and A But then ~ satisfies (11) and (12), and in particular
B~ = Bn
' so it follows by Problem 4.9 that Cn is a.s. diffuse. We
may thus conclude from Lemma 8.2 that -1Pn » a and A are uniquely
determined by -1p~ ,and hence, by Theorem 2.3 in Billingsley (1968),
that (i) and (ii) remain true for the original sequence. This completes
the proof.
The above convergence criterion takes a particularly simple form
in the case of thinnings:
COROLLARY 8.4. For n EN, let Pn E (0,1] and let ~ be a p -thinningn n
of some point process Suppose that P ~ 0 •n
Thend
~ ~ somen
iff P n i some n, and in this case ~ is a Cox process directed by n.n n
PROOF. The preceding proof applies with cn - Pn ' a = 0 and A = 61
(and it may even be simplified in the present case, since eii) is auto-
matically satisfied). No regularity assumption is required on account
of Corollary 3.2.
-110-
The last result leads to an interesting characterization of the class
of Cox processes:
COROLLARY 8.5. Let ~ be a point process on S. Then ~ is a Cox
process d-&l'ected by some random measure n iff ~ is for every p E (0, IJ
a p-thinning of some point process ~. In this case np
is a Cox process
directed by -1P n.
PROOF. Let p E (0,1] , let ~ and np be Cox processes directed by
and -1 respectively, and let ~p be a p-thinning of By11 P 11 np
(1. 4) and (1. 6) we then obtain for any f E F
= L (-log[l11p
p-thinning of some
so
L~ (f)P
pel - e-f)]) = L -1 (1 - [1 - pel - e-f )]) =
p n
= L -1 (p(l - e-f)) = L (1 - e-
f) = L~(f) ,
P n n ~
Conversely, suppose that ~ is for every p E (0,1]
-1Applying Corollary 8.4 with p =nandn
a
~n = ~ , n EN, it is seen that ~ must be a Cox process.
We conclude this section with a useful criterion for infinite divisibility.
LEMMA 8.6. Let ~ be a random measure on S and, for any c > 0 , let
11C
be a Cox process directed by c~. Then ~ is infinitely divisible
iff 11 c is infinitely divisible for every c > 0 .
PROOF. The necessity part is obvious. Suppose conversely that nc is
infinitely divisible for every c > O. Put p~-l = D and introduce
-111-
in the spaces of distributions on Nand M t~e operators n Vp P
and J corresponding to division by p , p-thinning and formation of
Cox processes respectively, p E: (O,lJ . From (1.4) and (1.6) it is seen
that V and J commute with convolution. Using this fact and the lastp
assertion in Corollary 8.S, we obtain
(V (In D)l/n)n = V In D = JDP P P P
P E: (0,1] n EN.
Thus (JD)l/n is a p-thinning for every p E: (0,1] and nE: N , so by
Corollary 8.5 there exists for every n E: N some distribution Dn on
M such that UD) lIn = JDn
But then
n E: N ,
and it follows by Corollary 3.2 that
o is infinitely divisible.
D = (D )nn n E: N , proving that
NOTES. Theorem 8.1 is new while Theorem 8.3 and Corollary 8.4 were given
in Kallenberg (1975). The latter result extends some classical work of
Renyi (1956), Nawrotzki (1962), Be1yaev (1963) and GOldman (1967), who
all consider p-thinnings of some fixed point process and change the "time"
scale by the factor p-1 to be able to obtain convergence. Finally,
Corollary 8.5 is due with a direct proof to Mecke (1968 ) and (1972). (ef.
Kerstan, Matthes and Mecke (1974), page 318), while Lemma 8.6 is due to
Kummer and Matthes (1970).
-112-
PROBLEMS.
8.1. State and prove the thinning case version of Theorem 8.1. Note
that no regularity condition is required in this case.
8.2. Let f, be a 8..f,Qmpound of s.ome simple point process*
p '" P{B > O} Show that E;O+ is then a p-thinning
this fact to give a new proof of Corollary 3.5.
n and put
of n. Use
8.3. For n E N , let E; be a B -compound of nn , and suppose that
B ~ 2 0n n d d
some 8 Show that E; + some E;, iff nn + some n ,n nand that E;, is then a B-compound of n State and prove the cor-
responding version of Theorem 8.3.
8.4. Let E;, be a B-compound of n and suppose that 8 is infinitely
divisible. Show that E;, has then infinitely many essentially dif
ferent representations as a 8'-compound of n' with 8' > 0 a.s.
(Hint: Replace n by nn, n E N Proceed as in the proof of
Theorem 8.1 to reduce to the case 8 > 0 a.s. Finally prove
that the representations thus obtained are really different.)
8.5. Let E;, be a 8-compound of n and suppose that n is infinitely
divisible. Show that E; is then also infinitely divisible. Prove
the corresponding result for the random measures defined by (11)
and (12).
8.6. Let E;, be a Cox process directed by n or a p-thinning of nShow that it is possible that E;, is infinitely divisible while
n is not. (Hint: It is enough to consider Z+-va1ued random variables.
To construct our n , we may start with the familiar Raikov-L~vy
example reproduced in Lukacs (1970), page 251, and add a suitable
infinitely divisible variable to make the thinning or Cox variable
E;, infinitely divisible.)
-113-
9. SYMMETRICALLY DISTRIBUTED RANDOM MEASURES.
Let w E Md be fixed. We shall say that a random measure s is
·e.-
symmetricaZZy distributed with respect to w if the distribution of
kEN , only depends
If wS E R' and if T is a random element in+
S with distribution w/wS, then this property is obvious for the point
process o , and since the property is persistent under addition of inT
dependent random measures and under mixing, it is shared by all mixed
Poisson and sample processes which are defined w.r.t. w
any such process must satisfy
In particular,
(1) P{sB = O} HwB) B E B ,
for some non-increasing function <p. If s is a mixed Poisson process
w.r.t. w with mixing random variable a , then
(2)-awB
P{sB = O} = Ee = ¢(wB) B E B ,
so in this case ¢ = L On the other hand, if wS E R' and if sa +
is a mixed sample process w.r.t. w/wS with mixing variable 'V , then
(3) P{~B O} = F. (I -wB)'V _
1/J (1 - ~~) = <PewB) B E BwS - wS ,
where 1/J is the generating function of 'V , so in this case
(4 ) <Pet) = 1/J(1 - w~) , t E: [0, wS)
1/J(t) = <t>(wS(l - t)) t E (0, I] .
-114-
rf ~ is a mixed Poisson or sample process, then p~-l is easily seen
to bc dctermined by w and ~. We shall use the symbol M(w,~) to
denote such a process ~. (Note, however, that the pair (w,~) is not
uniquely determined by p~-l since it can be replaced by (aw, ~('/a))
for any a > 0 .)
It turns out that every symmetrically distributed simple point process
is a mixed Poisson or sample process. We shall even be able to prove the
following stronger result.
THEOREM 9.1. Let ~ be a simpZe point process on S and Zet w E Md .
Further suppose that U c B is a DC-ring and that t E (O,~] is fixed.
dThen ~ = some M(w,~) iff there ex~st8 some non-increasing function
~t satisfying
(5 )
In that case
(5' )
-t~UEe = ~t CwU) U E U .
X E [0, wS] .
PROOF. Let t E R' and B E B+
and suppose that ~ is a mixed Poisson
process with mixing variable a. Then
On the other hand, we get in the mixed sample case, on account of (4)
-115-
Thus (5) holds in both cases with qit defined by (5'). For t", 00
this statement was proved in (2) and (3).
Suppose conversely that (5) holds, and let us first assume that
.-
t = 00
w ~ 0
Since w = 0 clearly implies ~ = 0 a.s., we may assume that
Our first aim is to extend (5) to B, i.e. to prove (1). For
this purpose, let sl' 52' ... E S be a bounded sequence of distinct
points belonging to the support of w, and consider for fixed s = sn
some sequence of sets Uk E U with kEN ,satisfying Uk ~ Is} .
Then 0 < wUk ~ w{s} = 0 while ~Uk ~ ~{s} a.s., so we get
Hence by Fatou's lemma
so qi(O+) = 1. To extend this continuity at 0 to (uniform) continuity
on [0, wS) (or even on [0, wSJ if w has bounded support), let £ > 0
be fixed and choose a > 0 such that 1 - qi(30) < £. For any s,t E [0, wS)
with s < t < s + cS , we can choose sets lJ,V E U with U c V and such that
(6 ) s - cS < wU S S < t < wV < t + 0 .
In fact, let C E U be such that wC ~ t and let {C .} cUbe a null-arraynJ
of partitions of C
for construction of
-116-
Then max wC . + 0 since w E Md ' so it sufficesj nJ
U and V to consider an"n E N' with max wC . < Qj nJ
From (5), (6) and the monotonicity of ~ we get
~(s) - ~(t) ~ ~(wU) - ~(wV) = P{~U = O} - P{~V = O} = P{~U = 0 , ~V > O} ~
~ P{~ (V\U) > O} = 1 - ~ (w(V\U)) =
= 1 - ~(wV - wU) ~ 1 - ~(3Q) < E •
We may now use the monotone class theorem to extend (5) to B.
For arbitTary n E Nand h < wS/n , it is possible to choose disjoint
sets Bl , ... ,Bn E B with wBI = ••• = wBn = ~ It. fact, let C E B
with wC > nh and let {C .} be a null-array of nested partitions ofmJ
fixed k > choose-1
C For 0 we can m so large that max wCmj < kj
and then there exist disjoint unions Bkl , ... ,Bkn of the partitioning
C such that h-1
~ h j 1,sets CmI' m2' ... - k < wBkj , = ... ,n
The sets Bkj may clearly be choosen non-decreasjng in' k for each
j , and in that case the sets
desired property.
B. = U BkJ
.J k
j = 1, ... ,n , have the
With this choice of Bl , ... ,Bn ' it is seen from (1) and Lemma
5.7 that
?: 0 ,
whichSmcans that the function ~ is completely monotone (cf. Feller (1971),
together with the corresponding inequality for arbitrary
pp. 223,439).
5
If wS = ~ , it follows that is the L-transform of some
x
-117-
R -valued random variable, and we get+ . by (2) and Theorem
-e
of some Z -valued random variable, and again+
3.3. On the other hand, if wS < ro then the function ~(t) = ~(wS(l - t))
t E (0,1) , being absolutely monotone, is the probability generating function
d~ = M(w,~) , now by (3) and
Theorem 3.3. This completes the proof in the case t = ro
In the case of finite t, let ~ be a p-thinning of ~,where
p = 1 --t
e As in the proof of Theorem 3.4, we obtain
U E U ,
and it follows as above that dn = M(w, ~t) . But then n is symmetrically
distributed w.r.t. w, and from (1.6) it is easily seen by analytic con-
tinuation that this is also true for ~. In particular, (1) must hold
for some ~,and it follows as above that
now complete.
The proof is
COROLLARY 9.2. Let ~ be a point process on S and let w EM. Then
d M( ) -:f.'+' B d M( .) f B B~ = some w,~ v J ~ = some wB' ~B or every E . We may
then take wB = Bw , in which case ~B becomes the restriction of ~
to [0, wB] . Furthermore, p~-l is then uniquely determined by P(t,;B)-l
for any fixed B E B with wB > °Note that the only if part of the first assertion contains a well-
known classical property of the Poisson process.
PROOF. Let us first suppose that w E Md' Then the first assertion
is an immediate consequence of Theorem 9.2, while the second one follows
-118-
by comparison of (Z) and (3). As for the last assertion, note that ~B
has probability generating function ~(wB(l - t)}, t E [O,lJ Thus
P(~B)-l determines ~ on [0, wB] and hence, by the uniqueness of analytic
continuations, on [0, wS)
The proof for general w E M requires randomization.
variables which are uniformly distributed on
and choose independently of {T.}J
a sequence {n.}J
[0,1]
Write ~ = L is. T.J J
of independent random
It is then easily
seen that ~ = L is is a point process on S x [0,1] and that. (T.,n.)J J J
Add~ = M(w x A,~) iff ~ = M(w,~) , where A denotes Lebesgue measure on
[0,1],..
Since w x A is diffuse, the assertion is true for ~, and
so it holds for ~ as well.
We next indicate by an example how Theorem 9.1 can be used to improve
Theorems 3.3 and 3.4 in the particular case of Poisson and sample processes.
COROLLARY 9.3. Let w E Md and~ for wS = 00 or < 00 ~ let ~ be a
Poisson or sample process respectively with intensity w
that n is a simple point process on S satisfying
Further suppose
P{nU = O} = ~(wU) U E U ,
for some DC-ring U c B and some non-increasing function ~ ThBn
dE;, = n iff theY'e exist two sets Bl , Bz E B with 0 < wBl < wBZ satisfying
(7) P{~B. = O} = P{nB. = O}J J
j = 1,2 .
PROOF. By Theorem 9.1 we have
-119-.
d .n = M(w,~) In the case wS = 00 we
get by (2) and (7)
(8) Ee -as -s= e Ee-at -t= e
where s = wBl and t = wB2 while a is a random variable with L-transform
~. If a were non-degenerate, it would follow by Jensen's inequality for
strictly convex functions that
-atEe . -t= e
which is impossible and proves that a is indeed degenerate. From (8)
we get a = 1 a. s., and so ~ (x) ==-xe which means that n is a Poisson
process with intensity w. A similar proof applies to the case wS < 00 •
We are now going to characterize the class of arbitrary symmetrically
distributed random measures. Consider an infinitely divisible random
measure ~ with homogeneous w.r.t. w independent increments such that
(9) -H f- log Ee = awf +
SxR'+
(1 - e-xf(s))(w x A) (dsdx) f E F ,
where a E R and A E M(R') with AK < 00. By Fubini's theorem, the+ +
integral Is (1 - e-xf(s)}w(ds) is a measurable function of x, and
therefore the right-hand side of (9) is jointly measurable in a and
A Hence, by Lemma 1.4, we may perform mixing with respect to a and
A considered as random elements. This gives the most general distribution
in case wS = 00. The case wS < 00 is entirely different:
-120-
THEOREM 9.4. Let W E Md\{O} and let ~ be a random measure on S .
Then ~ is symmetrically distributed w.r.t. w iff
(i) in the case wS = 00 ~ there exist some R+-valued random variable
a and some random measure A on R' with AK < 00 a.s. such+
that~ given a and A ~ the conditional L-transfoY'm of ~
is given by (9);
(ii) ~n the case wS < 00 , there exist some R+-valued random variable
is conditionally distributed
a and some point process S =
a.s. such that~ given a and
L'Oo] ..,.]
S ~ ~
on R I with+
L.S. < 00
] ]
as aw' + L.S.o ~ where WI = w/wS while T I , T 2 . are] ] T.
Jindependent random oZemp.nts in S with distribution WI
The random elements (a,A) and (a,S) respectively are a.s. unique
measurable functions of ~ .
We shall call (a,A) or (a,S) respectively the canonical random
elements of ~.
PROOF. The sufficiency of (i) and (ii) is obvious. Conversely, suppose
that ~ is symmetrically distributed w.r.t. w, and let us first assume
that wS < 00 , say that wS = 1 The symmetry assumption then extends
by monotone convergence to B and in particular it it seen that I;;S < 00
a.s. Hence by Lemma 2.1, f, has the representation
(10)
-121-
for some random elements sd E Md , 81 ~ 82 ~ ... ~ 0 and '1' '2' '" E S , <
the latter being a.s. distinct.
are taken in random order.assume that the corresponding ,.J
Whenever two or more B.J
coincide, we may
We shall first consider the case when S is the unit interval [0,1)
while w is Lebesgue measure. Using Theorem 3.1, it is easily seen that
Ps-l is symmetric in the sense that sT- l ~ s for every w-preserving
bi-measurable 1 - 1 mapping T of [0,1) onto itself. Since a = sdS
and 81, 82 , ... are invariant under such transformations, the conditional
distributions of S ,given (a, 81, 82, ... ) , are again symmetric, so we
may consider a, 61
, 82, ... as constants. Since sd is by Lemma 2.1 a
-1 -1 dmeasurable function of S , we have for any T as above sdT = (sT )d = sd '
so even Psdlis sYmmetric. Moreover, the process sd[O,t) - at, t E [0,1] 4It
is clearly a.S. uniformly continuous and of bounded variation so, writing
Inj = [(j - l)/n, j/n) and 'nnj = Sd1nj - o./n, j = 1, ... ,n, n EN,
we get
I Tln2
J' ~ 0: ITl . I) max In. I ~ °. . nJ . TIJJ J J
a. s. , n~oo.
From the symmetry and the obvious relations
we hence obtain by bounded convergence
and
2= E I Tl nj ~ °,
j
and so
K 2
~x E{jIl nnj} = m~X{kEn~l + k~ - llEnn1 "n2} ,~ nETl~l + nen - 1) !ETlnl Tln2 1 ~ 0 .
-122-
But this implies ~d[O,t) = at a.s. for every rational t E {O,l] ,
and therefore t;d = etA a.s.
Next suppose that 61> Q , and let v = v (~) be defined for fixed
n E N by '1 E I For k,m E Z (mod n) , let Tkm
be the transformationnv
of [0,1) which maps Ink onto In,k+m and In,k+m onto Ink ' and put
t;nm t;T-lBy symmetry we get for any A E Mm
n n
P{~T~-1P{t; E A} = L P{t; E A, \/ = k + m} = L E A, \/(t;Tkm
) = k + m} =k=l k=l
nP{t;T~
nP{I;T- l= L E A, \/ CO = k} = L E A, v(1;) = k} =
k=l k=l\/m
= P{t;T-1E A} = P{I; E A} .
\1m nm
(Note that this relation is even true when 81 = 82 ' although v is
not uniquely determined by t; in this case.) This proves that
for any nand m.
dt;nm = I;
For fixed s E (0,1] we now define t;s as the random measure obtained
from t; by replacing '1 by '1 + s (mod 1) Let m,n ~ 00 in such a way
that min ~ s , and suppose that '1 + s ~ 0 (mod 1)
for any interval I with s ¢ I , and hence also when
Then !; I ~ t; Inm s
s E 10 since
By Fubini's theorem, this implies for any
k = sup{j: 8. > O} ,J
Thereforesnm[O,l) = t;[0,1) =
d d .S =!; ~ t; , l.e.nm s
measurable function
I;s[O,l)
t;s ~ t;
kf: R+ ~ R+ '
wt;nm ~ t;s a. s. , and we get
-123-
Ef(,1' '2' ... ) = Ef('1 + s, '2' 0") = f: Ef('1 + s, '2' .. o)ds =
= Ef: f (,1 + s, '2' .. 0 ) ds = Ef: f (s, '2' .. 0 ) ds =
where , '1
is a random variable in rO,l) which is independent of
with distribution w. Hence 'I has the same properties, and continuing
inductively, it is seen that T l , T2
, are independent with distribution
u). This completes the proof of (ii) when S = [0,1)
Turning to the case of general Sand w with wS = 1 , note first
that the sYmmetry property of ~ extends by monotone convergence from B
to B. For convenience, we may compactify S or rather assume from the
beginning that S is compact. Let S = {S .} c B be a null-array of nestednJ
{I .} formsnJ
S there isS E
I =
For each[0,1]
in such a way that
S .nJ
II ·1 = wS .nJ nJ
S , and note that ~x wSnj+O by compactness of S and diffusenessJ
we can make correspond a subinterval I .nJ
satisfying
To each set
[0,1]
w •
a null-array of nested partitions of
of
of
parti tions of
is measurable,f
Each
s , and hence weS. containingnJ ns
by res) =f: S + [0,1]
a unique decreasing sequence of sets
n 1-.n nJns
t c [0,1] , may be written as a countable union of S-sets, so
can define a function
and it follows in particular that every measure ~ E M induces a measure -1~f
on [0,1]
M , so E,
Moreover, the mapping
induces a random measure
is measurable by definition of
[0,1] .
-124-
Writing D for the set of I-interval endpoints, we get by construction
of f
and hence for any t E D\{l}
-1t ::; wf [0 ,1] tED ,
-1 -1t / wf [O,t] = inf{wf [O,s),! ~ D, s > t} ~ inf{s E D: s > t} 3 t ,
which yields -1 Since is dense in [0,1] this togetherwf [0, t] = t D ,-1 implies that -1
with the fact that wf rO, 1] '" wS = 1 wf equals Lebesgue
measure. Moreover, ~f-1 is symmetrically distributed with respect to
-1wf since ~ is symmetrically distributed with respect to w.
Using the completeness of S, it is easily seen that f has a unique
inverse -1f on f(S)\D , and since and therefore
a.s., there is a.s. a unique correspondence between the atoms of ~ and
-1~f , and (10) is equivalent to
of the theorem is known to be true for
with a.s. diffuse ~ f- I and distinctd
f(T l ),f(T2), ... Now the statement
-1 -1-1~f , and so ~df, = awf a.s.
with a = ~dS , while f(T 1),f(T2), .. , are mutually independent and
-1independent of (a,6 1,6
2, ... ) with common distribution wf . For
every S . E SnJ
we obtain a.s.
~dS .nJ-1
= t;d(S . \f D)nJ-1= t;df (1 . \D)nJ
-1= awf (1. \D)nJ-1
= awf 1 . =nJ
= a I1 . I = awS .nJ nJ
-125-
and S being a OC-semi-ring, it follows by Oynkin's theorem that ~d = aw
a.s. Similarly, we have a.s. for any n, j and k
{'k E S .} ~ {'k € S .\f-lO} = {f('k) € I .\O} = {f('k) E I .} ,nJ nJ nJ nJ
so we get
k kP({(a, Sl' SZ' ... ) E A} (') h. E B.}) = P{(a, Sl' SZ' ... ) E A} IT wB.
j=l J J j=l J
for any measurable set A and for arbitrary kEN and. Bl ,
and by Oynkin's theorem, this extends to general Bl , The
proof of (ii) is now complete. The measurability of (a,S) is obvious.
We now consider the case when wS = <Xl Choose Cl , CZ' ... E B
with CO t S and wC l > a Since Cn~ is synnnetrically distributedn
with respect to C w for each n E N , we know thatn
Cn~ = a (C w) + \ S 0n n ~ nj T .J nJ
for some an' Snl' SnZ' ... and some 'nl' 'nZ' ... , the latter being
independent of (a , S l' S Z' ... )n n n and mutually independent with common
distribution C w/wCn n
Now C ~ = C (C~) for m ~ n , so writingm m n
a = 0. 1 ' Sn I 0 Pmn = wC /wC m $ n E N it is seen by identificationj Snj mn
that 0. 1 = aZ = ... = a a.s. while S is a p -thinning of Sn form mn
m :::' n E N Since Pmn + a as n + <Xl for each fixed m E N , it follows
by Corollary 8.5 that Sm is distributed as a Cox process on R' directed+
and that d Putting A A/weI it is seen thathy some A , A = P A = ,m m mnn
we may take A = (wC ) A for each n Let f E: F be arbitrary, andn n c
-126-
choose n E N so large that the support of f
Then
is contained in Cn
Ee-(~-aw)f = \Eexp{- L S .f(T .)} =j nJ nJ
EE[exp{- L S .f(T .)} I {S .}] =j nJ nJ nJ
= E ITJ. E[exp{-SnJ·f(TnJ·)} I SnJ'] = E IT f exp(-s .f(s))w(ds)/wC •
j S nJ n
and if we define
J-xf (s)hex) = - log e w(ds)/wC
S n
this expression may be written
E IT exp{-h(S .)}= Eexp{- Lh(S .)}. nJ . nJJ J
-S h= Ee n
-S h= EE[e n I 1<] =
= Eexp{-(wC ) f [1 - f e-xf(s)w(ds)/wC ]I«dx)} =n R' S n
+
= Eexp{ - f (1R'xS
+'
-xf(s))e (I< x w)(dxds)} .
Putting f(s) = tIc (s) •1
monotone convergence that
a.s. This proves that ~
t E R • it follows by Fubini's theorem and+
lim L~C (t) = 1 can only be true if I<K < 00
t+O 1is distributed as the random measure described
in (i). But for the in (i) we easily obtain S /wC X I< a.s. by then n
strong law of large numbers, so I< is a measurable function of ~ For
the ~ under consideration we may therefore redefine I< according to
-127-
this mapping, and doing so, (i) will automatically be satisfied. This
completes the proof of the theorem.
Of particular interest is the class of random measures which are
at the same time sYmmetrically distributed and infinitely divisible.
Here is a simple characterization of this class:
THEOREM 9.5. Let the random measure ~ be symmetrically distributed
with respect to some W E Md with canonical random elements (a,A)
or (a, B) Then ~ is infinitely divisible iff this is true for (a,A)
or (a,B) respectively.
PROOF. First suppose that wS < 00 , and let (a,S) be infinitely divisible.
For each n E N we may choose independent random elements
such that
By randomization we may then construct independent sYmmetrically distributed
random measures ~ ~ with canonical random elements'" l' ... , "'n
distributed with canonical random element
(a , B ) •n n
From Theorem 9.4 it is seen that + t;,n
is symmetrically
d... + B ) =
nd= (a, B) , and s 0 ~1 +
d+ t;, = F;.
nSince n was arbitrary, this proves
that F;. is infinitely divisible.
Conversely, suppose that F;. is infinitely divisible and choose for
fixed n E N some independent random measures d dF;.l = ... = F;.n such that
F;.l + ••• + t;,n
d= F;. . By considering the corresponding L-transforms, it
is easily seen that ~l' ... ,F;.n are symmetrically distributed, say with
-128-
canonical random elements (aI' 81), ... ,(an' Sn) . Again we may conclude
dfrom Theorem 9.4 that ~l + . "+~n = ~ has the canonical random element
so we must have+ S )n
+ S )nd= (a, S) Furthermore, the random elements
(aI' 61), ... ,(an' Sn) , being measurable functions of the independent
random measures sl' ... ,sn respectively, are themselves independent.
Since n was arbitrary, this proves that (a,S) is infinitely divisible.
In the case wS = 00 , let us first show that, if sl and s2 are
independent and symmetrically distributed with canonical random elements
(aI' AI) and (a 2 ' A2) respectively, then sl + s2 has canonical random
elements (a l + a2, Al + A2) To see this, note that (aI' AI) and
(a 2 ' A2
) are measurable functions of 1;1 and ~2 respectively, and
hence that 1;1 and 1;2 are conditionally independent with canonical
elements (aI' AI) and (a2 ' A2) , given (aI' a 2 , AI' A2
) Using (9),
it follows that sl + s2 has conditionally the canonical elements
(a l + a 2 , Al + A2) ,given (aI' a 2 , AI' A2) , and the assertion follows
by taking conditional expectations with respect to (a l + a2 , Al + A2)
Once this fact is established, we may proceed exactly as in the case
wS < 00 The proof is thus complete.
According to the last theorem, a symmetrically distributed random
measure t; is infinitely divisible iff this is true for the corresponding
pair of canonical random elements. In view of Theorem 6.5., the distribution
of ~ may thus be described in two different ways by means of spectral
representations of the form (6.6), and we may ask for the relationship
-129-
between the corresponding canonical measures. More precisely, it follows
from Theorem 6.5 that, if ~, (a,A) or (a,S) is infinitely divisible,
then
(11 ) f E F($) ,
or
t E R+
f E F(R') ,+
(13)
respectively, where in (11) rEM while A is a suitable measure on
AO
E M(R:) while y is a measure onM\{O} , in (12) a o E R+
M(R:)\{O} , and in (13) is a measure on N (R') \ {O}+
and we may ask for the relationship between these quantities. For a simple
description of r and A in terms of (ao ' AO' y) or (aD, y) , let
P and Q be the distributions of symmetrically distributed randomX,lJ X,lJ
measures on S with non-random canonical elements (x,lJ) in the cases
wS = 00 and wS = 1 respectively. In the former case we further denote
by Rx the measure on M\{O} induced by w via the mapping s -+ xc ,s
S E S , where x > 0 is arbitrary but fixed.
THEOREM 9.6. Let ~ be symmetricaZly distributed with respect to W E MduJiLh canonicaZ random elements (a,A) or (a,S) . Further suppose that
-130-
~ &s infiniteZy divisibZe, and Zet the quantities (r,A) and
(ao
' AO' y) or (aO' y) respectiveZy be defined by (11) - (13). Then
we have in the case wS = 00
(14) r = A = f P y(dxd~) + f R AO(dx) ,x,~ x
whiZe in the case wS = 1
(IS) r = A = J Q y(dxd~).x, ~
By comparison with Theorem 9.4, it is seen that the integrals
J Px,~dY and J Qx,~dY are formed in exactly the same way as the proba
bility distributions of general symmetrically distributed random measures,
except that the "mixing" measures y need no longer be normalized and
can in fact be infinite. In addition we have in (14) a term J RxdAO
which can not occur in the corresponding representation of distributions,
since the measure R is infinite and can not be normalized.x
PROOF. Let us first consider the case wS = 00
for any f E F
By Theorem 9.4 we get
log E[e-~f I a,A] = awf + J (1 - e-xf(s))(w x A) (dsdx) .
Wri ting
we thus obtain
g(x) = J (1 - e-xf(s))w(ds) x c R+
-131-
-~f -awf-Aglog Ee - - log Ee
and hence by (12)
Now
.-
-xc fe 5 )w(ds) =
giving rise to the second A-term in (14). Further observe that, by (9),
and so
= f dy(x,~) f (1 - e-mf)p (dm) =x,~
= f (1 - e-mf
) Jp (dm)dy(x,~),x,~
giving rise to the first A-term in (14). Finally, the term aOwf is
of the form rf with r = This completes the proof of (14).
Next suppose that wS = 1
f E F
We then get by Theorem 9.4 for any
-132-
Ee-sf = E exp(-awf - r 8.f(T.)) =j J J
EE (exp (-awf - LB. f (T .) I a, 8] =j J J
f -8.f(.. )= Ee -aw II E(e J JIB.] =
j J
-awf -8.£(T.)= Ee exp L log E[e J J I 8.] =
j J
where
-awf (\' )= Ee exp - ~ g(B j ) =
J
-awf-8gEe ,
g(x) = - log Je-xf(s)w(ds)
Using (13), we thus obtain
, X E R+
But from the above calculation it is seen that
II E N(R~) ,
and hence
= J dy(X,ll) J (1 - e-mf)Q (dm) =X,ll
= J (1 - e-mf
) J Q (dm)dy(x,ll).X,ll
Hence (15) holds, and the proof is complete.
-133-
We are now able to characterize the class of measures r and A
which can occur in the representation (11) of infinitely divisible sym-
metrically distributed random measure. Write for brevity K' (x) :: x .
THEOREM 9.7. Let W E Md and rEM, and let A be a measure on M\{O}
Then there exists some infinitely divisible random measure ~ which is
symmetrically distributed with respect to wand satisfies (11) iff
(i) in the case wS = 00 , (14) holds for some aO ~ a , some
1.. 0 E M(R') and some measure Y on (R+ x M(R~)J\{O} satisfying+
(16) A K < 00 Y{).1 K = oo} = a I (1 _ e-X-).IK)y(dxd).l) < 00
0
."
(ii) &n the case wS = 1 (15) holds for some a O ~ 0 and some
measure y on (R+ x (R~))\{O} satisfying
(17) Y{).IK' = oo} = 0
The quantities (aO' 1.. 0 ' y) or (aO' y) are then unique.
PROOF. Suppose that ~ is an infinitely divisible symmetrically distributed
random measure satisfying (11). Then (11) must be the canonical representation
occurring in Theorem 6.5, and therefore r and A are necessarily unique.
But then (14) or (15) holds by Theorem 9.6 for some (aO' 1.. 0 ' y) or
(uO' y) satisfying (12) or (13) respectively. In the case wS = 00 we
get in particular
r
(18) 1 E -t(a+AK) = a t +- og e 0 t ~ 0 ,
and since AK < 00 a.s. by Theorem 9.4, the last two terms on the right-
hand side of (18) must be finite for all t Furthermore, it follows
from (18) by monotone convergence as t + 0+ that 0 = Y{~K = oo} , which
completes the proof of (16). In the case wS = 1 , note that SKI < 00
a.s., and use a similar argument to prove (17). The uniqueness of
(ao' AO' y) or (aO
' y) follows by the uniqueness assertions in Theorems
6.5 and 9.4.
Conversely, let (aO
' AO
' y) be such as described in (i), and define
r and A by (14). Note that A exists according to the second relation
in (16). By Theorem 6.5 and the last relation in (16), there exist some
random variable a ~ 0 and some random measure on R'+ such that
(a,A) is infinitely divisible and satisfies (12). Using monotone con-
vergence as above, it follows from (12) that AK < 00 a.s., and hence
by Theorem 9.4, (a,A) can serve as the canonical random elements of
some sYmmetrically distributed random measure ~ which by Theorem 9.5
has to be infinitely divisible. Hence by Theorem 9.6, r and A are
related to ~ by (11), and therefore have the asserted property. A similar
proof works in the case wS = 1 .
NOTES. Theorem 9.1 is known for t = 00 in which case it was proved,
independently, by Davidson (1974) (in a special case) and Kallenberg (1973a).
The latter paper is also the source for Corollary 9.2, which extends a result
by Nawrotzki (1962)? Furthermore, Corollary 9.3 is a typical resul t from
Kallenberg (1974a), while Theorem 9.4 was proved by BUhlmann (1960) and
Cf. Kerstan, Matthes and Mecke (1974), p. 80.
-135-
by Kallenberg (1973b) and (1974b), (see also Kallenberg (1973c)). Finally,
Theorem 9.5 extends two results for point processes given by Kerstan,
Matthes and Mecke (1974), pages 66-67, while Theorems 9.6 and 9.7 are7
new.
PROBLEMS.
9.1. Let S = {I, ... ,n} , let w be counting measure on S and let
~ be counting measure on a subset of k elements in S chosen
at rando~. Show that ~ is (in the obvious sense) symmetrically
distributed w.r.t. wand still not a mixed sample process (unless
k = 0 or n). Thus the assumption w E Md is essential in Theorem
9.1.
9.2. Let ~ be a sample process on [0,1] with intensity n times
Lebesgue measure. Show that there does not exist any symmetrically
distributed (w.r.t. Lebesgue measure) point process n on some
interval [O,c], c > 1 , such that the restriction of n to
[0,1] is distributed like ~. (Hint: n[O,l] cannot be non-random
unless n = 0 a.s.)
9.3. Extend Corollary 9.2 to general symmetrically distributed random
measures. (Cf. Westcott (1973).)
Prove the
9.4. Let w E Md and suppose that
S ~ O. Show that ~ and n
distributed w.r.t. w
are
is a S-compound of n , where
simultaneously symmetrically
corresponding result for the
random measures defined by (8.11) and (8.12). (Hint: use analytic
continuation. )
9.5. Let
W •
W E Md be fixed and let ~ be symmetrically distributed w.r.t.
Show that p~-l is uniquely determined by p(C~)-l for any
7 Extensions of Theorems 9.5 - 9.6 are proved in Kallenberg: "Infinitelydivisible processes with interchangeable increments and random measuresunder convolution", Tech. flep0Y't, Dept. of Mathematics, Gljteborg Universlty.
-136-
fixed C € B with wC > O. (Hint: use the connection with thinnings
exploited in the proof of Theorem 9.4. Cf. Ka11enberg (1973c).)
9.6. Specialize Theorems 9.5 - 9.7 to the case of point processes.
10. PALM DISTRIBUTIONS.
Let t; be a random measure on S , and define the corresponding
C b II (p~-l)l S M bamp e measure ~ on x y
(1) B € B M € M.
If Et; is a-finite, which holds e.g. when Et; € M , we can define the
corresponding Radon-Nikodym derivatives P M bys
a.e.P Ms
= (pt;-l)l(ds XM) =
(pt;-l) 1 (dsxM)~[t;(ds);t;€M]
Et;(ds) S € S M € M.
More generally, suppose that f € F is such that Eft; is a-finite.
We can then define
P Ms= E[ft;(ds) ; t;€M]
Eft; (ds) S € S a.e. Eft;
without ambiguity, since the two expressions for P M clearly coincides
a.e. Et; on the support of f . We shall call P M the Palm probabilitys
of M w.r.t. s Note that, by definition, P M is 75-measurab1e in ss
for fixed M E M and that P M is almost a probability measure in Ms
for fixed s E S , in the sense that every probability defining property
holds a.s. Et; Just as in the case of conditional probabilities, we
-137-
may choose a version which is actually a probability measure in M (to
be called the Palm distribution) for every fixed s E S. This is possible
by the usual arguments for construction of conditional distributions (cf.
Bauer (1972)), on account of the following fact.
LEMMA 10.1. M and N are Polish spaces in their vague topologies.
PROOF. By Lemma 4.4, it is enough to consider M. Let G be a countable
base in S , and assume without loss that G is a semi-ring.
Approximate each indicator lG' G E G , from below by a sequence
of functions in Fc
be an enumeration of all these
functions for all G E G , and note that, by Dynkin's theorem, each ~ E M
{~n} , so any sequence
is uniquely determined by
In fact, the latter condition implies
must contain a subsequence
Furthermore, we have
N' c N
j EN.~f.J
j EN.for alliff ~ f. -+ ~f.n J J
compactness of
Nil such thatv
~' (n E Nil) But then ~ f. -+ ~ If. (n E Nil)~n -+ some ,n J J
j E N , so we get ~f. - ~ 'f. and hence ~ = ~ , , proving that indeedJ J
vIt is now easily that the function M2
-+ R defined by~n -+ ~ seen p:+
00
p(~, ~') = Lj =1
~, ~' EM,
is a complete metric on M generating the vague topology. To see that
M is separable, let A and B be countable dense subsets of Sand
R: respectively. Then the class of all finite sums of measures bo ,a
a E A, b E B , is clearly countable and dense in M.
-138-
Henceforth, we shall always assume the set functions ps to be proba-
bilities on M (and even on N if ~ is a point process), and for con-
venience of writing, we introduce for every s E S a random measure
~s ' defined on the same probability space as ~ and with distribution
ps
Just as in the case of conditional distributions, the expectations
corresponding to Palm distributions can be computed in two different ways.
More precisely, we have the following result, known as Campbell's theorem.
LEMMA 10.2. Let ~ be a random measure on S with E~ a-finite, and
let f: S x M-+ R be measurable. Then+
Ef(s, ~s) = f f(s, ~ (w))P(dw) = J f(s,~)P (d~) =Q s M s
= Ef(s,~)S(ds)
EHds) S E S a.e. E~.
PROOF. We have to prove that Ef(s, ~s) is B-measurable and satisfies
(2) J Ef(s, ~ )E~(ds) =B s
E J f(s,~)S(ds)B
B E B •
Now this is true by definition of ~s for indicators f of product sets
A x B, A E B, M EM, and it easily extends by Dynkin's theorem to
indicators of arbitrary sets in B x M. But then it holds for simple
functions, and finally, by monotone convergence, for arbitrary f .
-139-
In the particular case of simple point processes, PF;-l should bes
interpreted as the conditional distribution of F;, given that F;{s} = 1
(Note, however, that a definition along these lines only makes sense when
p{~{s} > o} > 0.) We could then expect F;s to have an atom at s, and
this can in fact be achieved by changing the definition on an EF;-null
set:
LEMMA 10.3. Let F; be a point process on S with EF; a-finite. Then
there exist versions of F;s such that F;s - Os is a point process for
each S E S • For any measurabZe function f: S x N + R we then have+
(3) S E S a.e. E~,
where f(s,~) may be arbitrariZy defined for ~ tN.
PROOF. Clearly
(4) {(s,~) E S x M: ~ - ° E N} = (S x N) n {(s,~) E S x N: ~{s} ~ I} .s
To see that the second set on the right of (4) belongs to B x N , let
B c B be arbitrary and let
of B Then
{B .} c B be a null-array of partitionsnJ
{ (5,11) E B x N: jJ{ s} ~ l} = n u (B . x {jJ EN: jJB . ~ I}) to B x N ,n j nJ nJ
and the measurability in (4) follows from the fact that S is a-compact.
By (2) and (4) we now obtain for any B E B
-140-
f P{~ - 0 E N}E~(ds) = E IBl{~_o EN}(s,~)~(ds) =
B s s s
= E IB1{ i;; {s hll (s ,0 i;; Cds) = E!;B ,
so
f (1 - P{~ - 0 E N})E~(ds) = 0 ,B s s
and hence
(5) p{!; - 0 E N} = 1 a.e. E~.s s
Redefining i;;s == 0 on the exceptional s-set in (5), we can make (5)s
hold for all s E S For each s E S we then redefine ~s = <5 ons
the exceptional w-set where i;;s - 0 4 N to make ~s - <5 E N hold fors s
all sand w. Since ~ B - 8 B remains A-measurable for each s E SS S
and B E B , it is now a point process. But then (3) is an immediate
consequence of Lemma 10.2.
Besides the random measures i;;s we shall also consider, for arbitrary
f E F with E~f E R' the mixture+ '
~f with distribution defined by
M EM.
A
When ~ is a point process, we also consider the point processes f,f
with distributions defined by
P{~f E M} = E~f Is P{!;s 8 E M}f(s)Eqds)5
= E~fEIs l{i;;_o EM}(s,i;;)f(s)~(ds) MEN .5
-141-
It is easily verified that Campbell's theorem remains valid for ~f andA
~f ' in the sense that
(6) Eh(sHfE~f
Here h is any measurable function from M or N respectively to R+
From the definitions of Palm distributions of a given random measure
~ it is obvious that they determine, together with the intensity E~,
the Campbell measure (Ps-l)l and hence also the distribution p~-l .
Indeed, a much stronger result is true, indicating a close relationship
between the Palm distributions for different s E S:
LEMMA 10.4. Let ~ be a random measure on S and let f E F with
E~f E R: . Then Esf and pE,;-l determine the restriction of p~-lf
to the set {~: ~f > O} They further determine p~-l for all s E Ss
a.e. E(fE,;)
Note that f may always be chosen such that f(s) > 0 for all
S E S Then ~f = 0 implies ~ = a , so in this case Esf and
alone determine the whole distribution of ~ .
PROOF. The first assertion follows from (6) by choosing for fixed M E M
~f > 0 ,
~f = 0 ,
-142-
since for this h, (6) takes the form
(8) p{i; E M, i;f > O} = Ei;f Eh(l;f) .
Furthermore, the probabilities in (8) determine the measure
E[ (fS) B; I; E M] E[Cfi;)13; i; E M, H > 0] 13 E B
and so the second assertion follows from the fact that
P{I; E M}s
E[i;(ds) ;I;EM]EI;(ds)
E((fS) Cds) ;i;EM]E (fl;) (ds) S E S a.e. E(fS) .
For f E F , the above uniqueness assertion can essentially bec
strengthened to continuity:
THEOREM 10.5. Let 1;, 1;1' 1;2' be random measures on S and let
condition dthen implies I;nf + I;f ~
Furthermore ~ I; !! I;n
EI; f + El;f ~ then
di;nf + i;fwhi le conversely
dand I;nf + i;f imply that
Under the assumptionEi;f E R~be such thatf ( Fc
implies
EI; f + El;f .n
PROOF. First assume that i; ~ I; and Ei; f + Ei;f , and let h: M+ Rn n +
be hounded and continuous. Then the mapping p -* h(p)pf is continuous,
so we get
(9) h(1; )F, f ~ hCF,)r,fn n
by Theorem 5.1 in Billingsley (1968). Furthermore, it follows from the
relations F, f ~ i;f and Ei; f + Er,f that {E; f} is uniformly integrablen n n
-143-
(cf. Theorem 5.4 in Billingsley (1968)), and hence so is
By (9) we thus obtain
{h(~ H f} .n n
(10)
so by (6)
(11)
Eh(~ )~ f + Eh(~)~f ,n n
is bounded and measurable with
If h:M+R+
Since h was arbitrary, this
Suppose conversely that
dproves that ~nf + ~f
d~nf + ~f and E~nf + E~f
~f ~ Of a.s., it follows by Theorem 5.2
in Billingsley (1968) that (11) holds, and by (6) this implies (10).
For any fixed bounded and continuous function g: S + R we may choose+
__ {]Jo (fg) /]Jfh (]J)
]Jf > a
In fact, h is bounded and measurable, and by (6) the set 0h = {]Jf = O}
satisfies
For this particular h , (10) takes the form
E(f~ )g = E~ (fg) + E~(fg) = E(f~)g ,n n
and since g was arbitrary, it follows by Theorem 4.10 that
as asserted.
wuf E: --)- fE: ,
n
Finally suppose that
-144-
~ i ~ andn
d~nf + ~f ' and choose
h(p) ::: (1 + ).If) -1 . Then h().I) and h().I)).If are both bounded and con-
tinuou5, so (10) and (11) are satisfied. Since moreover Eh(~f) > 0 ,
it follows from (6) that E~ f -+ E~f .n
In case of point processes, we may also prove a continuity theoremA
for ~f' f E Fc
THEOREM 10.6. Let be point processes on S ~ and suppose
that E~ E M\{O}. Then any two of the following statements imply the
third one.
(i)
(li)
(i ii)
Er X. Er<'n <"
A d A
~nf -+ ~f for all f E F with E~f > 0 .c
PROOF. Let f E Fc and let h: N+ R+ be bounded and continuous. Then
the function
(12) g().I) = J h().I - 0 )f(s)).I(ds)S s
).I eN,
is vaguely continuous. In fact, suppose that ).I{s}? 1 and ).I {s } > 1 ,n n
n EN, and that )l '! 11n ,.. in Nand s -+ 5n
in S Then
).I - 0n s
n
holds hy definition of vague convergence, and hence hy continuity
-145-
h (]J - 0 ) f (s ) -+ h (]J - 0 ) f (s ) .n s n s
n
Since hand f has bounded support, it follows by Theorem 5.5 in Billingsley
(1968) that g(]Jn) -+ g(]J). (Note that, in Billingsley's notation, we need
only assume that h (x ) -+ hex) for x -+ x such that x E E and x E En n n n n
= 1 .)
Now suppose that (i) and (ii) hold, and let h and g be as above.
Then Eg(t;) 1 Eg(t.:) follows by uniform integrability as in the preceedingn
proof, and it follows by (i) and (7) that I:h(t;nf) -+ Eh(t;f) provided
Et.;f> o. Since h was arbitrary, this implies (iii).
Next suppose that (ii) and (iii) hold, and define g by (12) with
]J EN.
Then h is bounded and continuous, so by (iii)
(13)A
Eh(snf) -+ Eh(€f) > 0 .
Moreover, g is continuous, and it is even bounded since
l+Jlf-f(s)+llfllf(s)Jl(ds)
IIence by (ii)
(14)
f(s)}l(ds)l+Jlf
Eg (s ) -~ Eg (S) .n
l]Jff <; 1 .+]J
Using (7), (13) and (14), we obtain Et:nf -+ Et;f. If U:f 0, we m~IY
-146-
choose f l , f 2 E Fc with f = f 1 - f 2 and E~fl = E~f2 > 0 , and conclude
that
E~f .
Thus E~ f + E~f is valid for all f E F ,and (i) follows.n c
Finally suppose that (i) and (iii) hold. By (7) we get for any
f E F and any bounded continuous h: N + Rc +
For any fixed
-tffe and put
f E Fc
h(ll) ==
and t E R+ ' we may replace the
-tllfe , thus obtaining
f in (15) by
E I exp{- t~ f + tf(s)}f(s)e-tf(s)~ (ds)S n n
+ E Js
exp{- t~f + tf(s)}f(s)e-tf(s)~(ds) ,
or equivalently
-t~ fE~ fe n + E~fe-t~f .
n
By Fubini's theorem and dominated convergence, we get
-~ f II -t~ fHe n = 1 _ E ~ fe n dt = 1 -
a n
= 1 -
Jl -t~ f II
E( fe n dt + 1 -o n a
I' I;, UC -tcfJt " Eo -(I' ,
EE,fc -tf;fdt
and since f was arbitrary, it follows by Theorcm 4.5 thatdr >- r,
'n
-147-
This completes the proof.
Palm distributions are often easily calculated by means of L-transforms.
In fact, we get by dominated convergence for any f E F with El;f E R'+
(16)
and Ll; is obtained from this derivative by normalization. As a: simples
example, suppose that l; is a Poisson process with intensity A E M\{O} .
By dominated convergence we get
so putting t = 0 ,
and hence by normalization
Lr (f) = Ll;(f)e-f(S) = Ll;(f)L8 (f) = Ll;+8 (f)J s 5 s
S E Sa. e. El;
By Theorem 3.1 thus obtain E;sd
E; 8 equivalently t: s 8d
(we = + or -s s
El; which yields E;fd
E; for f F with Et;f R' Wea.c. , every E E" +
shall usc this fact to prove the following complement to Corollary 7.4.
A ( r~\ {O} , Ze t; I ( BA
of point processes on s
be a DC-serni-ring and Zrt
Put s = L l; .• Thenn . nJ
J
{F, .}nJ
t; ~ t;n
l'e a null-or-rlay
and
-148-
for aU f E F with El;f E R' iff+
ei) L P{I; or > O} -+ AI r E I ,j nJ
(i i) L E[E; 0 B; F,njB > 1] -+ 0 B ( Bj nJ
PROOF. Suppose that (i) and (ii) hold. Then dE, -~ E,n
follows by Corollury
7.4 since
(17) L P{E, 08 > 1} ~ } ~ E[I; .8; I; .8 > 1] -+ 0j nJ J nJ nJ
B E B .
Furthermore, we get by ei), eii) and (17) for any I E I
L EI; .r := L P{I; or> O} - L P{I; or> l} + L E[I; 01; E, or> 1] -+ AI ,j nJ j nJ j nJ j nJ nJ
proving that
:= L EE, 0 X Aj nJ
EE; •
A d A dlienee it fa llows by Theorem 10.6 that E;nf -+ t.f E; for any f ( F with
EE;f E R'+
d A d d A
Suppose conversely that I;n -+ I; and that I;nf -+ S sf for any
f (F with EE;f E R'+
By Corollary 7.4 and Theorem 10.6, we get for
~ P{t; .8 > OJ -+ AB<, nJJ
L P{E; .B > I} -+ 0o nJJ
L U:nj
8,. AB .j
Ilere the fi rst relation contains (i) and moreover
L Efi; .B; ~ .B > lJ =. nJ nJ]
proving (ii).
-149-
L E~ .B - L P{~ .B > O} + I P{~ .B > I} ~ 0 ,j nJ j nJ j nJ
0'
NOTES, Palm probabilities were introduced for stationary point processes
on R hy Palm himself (1943), whose studies were continued by severalv
authors, including llinc'in (1960) and Slivnyak (1962). The present approach
is essentially due to nyll-Nardzewski (1961), (see also .Jagcrs (197:\),
Papangelou (1974), and Kerstan, Matthes and ~1ccke (1974), the latter ;Iuthors
working directly with the Campbell measures). Camphell 's theorem, in the
formulation of Lemma 10.2, was first stated by Kummer and ~Iatthes (1970),
while Lemma 10.3 is implicit in Kallenberg (1973a). The uniqueness Lemma
10,.4 is new. As for the corresponding continuity problem, a discussion
for stationary point processes is essentially given by Kerstan and Matthes
(1965), and for general point processes by Kallenberg (1973a), from which
Theorem 10.6 and Corollary 10.7 are taken. Theorem 10.5 extenns a result
in the same paper.
PROBLEMS.
10. 1. Prove the version of Theorem 5.5 in ~illjngsley (1068) which 1S
used in the proof of Theorem 10.6.
10.2. Let ~ be a random measure on S with Ei; a-finite, and let
M E M Show that P{i; E M} = 1 iff P{i; E M} = 1 , S E Ss
a.e. Ei;
10.3. Let E; be a random measure on S , let B c B with U,B < m
and let
-1 J= x B
x > 0 be fixed. Show that P{i;B x}
p{i; B = x}Ei;(ds) , (cf . .Jagers (1973)).s
-150-
10.4. Let F;, and n be independent random measures on S with T:r,
and En in M Calculate P (F;,-1 in of pF;,-l and+ n) s terms
-1 sPn s
T , S E S ,sSay that a random measure
and define the translation operatorsS = RLet
on M by (T ~)B = ~(B - s) , B c Bs
F, on R is stationa~y if P(TsF;,)-l 1S independent of s E R .
Show that in this case EF;, equals some constant A times Lebesgue-1measure, and that if A E R' , then peT F;,) is independent of s
+ sThe converse is also true. (Cf. Mecke (1967) and Jagersa.e. EF;,
(1973). )
10.5.
10.6. Let ~ be a random measure on R and let f E F (R) with
Show that it is possible to define a random measure ~f on R such
that
f P{T F;, E M}f(s)EF;,(ds)/EF;,fJ -s 5
= E f l{T F;, EM}fF;,(ds)/EF;,f-5 s
~1 EM.
State and prove the corresponding version of Campbell's theorem.
10.7. State and prove a version of Theorem 10.6 for the random measures
F;,f defined in Problem 10.6. Specialize to the case of stationary
random measures. (Cf. Kerstan and Matthes (1965) as well as Kallenberg
(l973a) . )
10.8. Let F;, be a random measure on S and let f (F be such that
EE,f c R'+
Define the di stri hut i on of a random element (n f , r: f )
on R x M by+
for any measurable set A
equals
-151-
in R x M , and show that this probability+
..
State and prove the corresponding version of Campbell's theorem.
11. CHARACTERIZATIONS INVOLVING PALM DISTRIBUTIONS.
Tn the preceding section we saw that any Poisson process [, on
S satisfies the relation ~ - 0 ~ ~, S E S a.e. E~. Actually,s s
this relation characterizes the class of Poisson processes among arbitrary
random measures:
Proposition 11.1. Let ~ be a random measure on S with E[, EM.
Then ~ &s a Poisson process iffd
~ = E. + 8s ' s S E S a.e. Et; .
PROOF.
n ( N
Suppose that d~ ~ + 0 a.e.s s Then we get for any B E Band
n} = !-E{~B; [,Bn n} = !- J E[[,(ds); t;B = n] = !- f P{[, B
n B nBsn}E[,(ds)
=!-J P{[,B = n - l}E~(ds)n 13
and hence by induction
P{~B = n} = P{~B
!- P{~B n - l}E~B ,n
([[,13/O} n!
In the same way we get more generally for any kEN, nl' .. , ,nk E Z+
-152-
and disjoint Bl , ... ,Bk E Bn.
k k k (EI;B. ) J
P n {I;B. = n. } = PH u B. = a} IT J
j =1 J J j =1 J j=l (n.) !J
/\ similar argument also shows that P{~B 1 z } = a ,+
B E B . Hence
must be a Poisson process, and the proof is complete.
The aim of the present section is to prove extensions in various
directions of the ahove proposition. We shall first consider the case
of infinitely divisihle random measures i; with Ei; EM. If a and
A are the corresponding canonical measures defined in Theorem 6.4, we
can define the corresponding Campbell measure A on S x M by the relation
A(B x M) = aBeaM + JM~BA (d~) B E B M E M
Formally, we may write A = ae + Al , where Al =J
pA (dlJ) , d. (10.1) .a
Since J lJBA (d~) = EI;B < 00 we can proceed as in the definition of Palm
distrihutions and define the distributions A s E S a.e. EI; bys
A Ms
= A(dsxM) =A(dsxM)
a(dS)OoM+JM~(dS)A(d~)
a(ds)+fM~(dS)A(dlJ)
Campbell's theorem remains valid for the
negative function f on S
AS
i.e. we have for any non-
A fs
a(ds)f(s,O)+ f(s,lJ)lJ(ds)A(dp)
a(ds)+ lJ(dS)A(d~)
a.e. U:.
-153-
Let us now use the method of the preceding section to calculate ps
For any f E: F t E R+ and B E B we obtain
a~ exp{- af - taB - f (1 - e-~f-t~B)A(d~)}
= L~(f + tlB){aB + J ~Be-~f-t~BA(d~)}
by dominated convergence. Hence
and so, by Campbell's theorem,
L~ (f) = L~ (f)s
Since the last expression is the L-transform of it follows
by Corollary 3.2 that p[-l = p~-l*i\ a. e. , and in particular's sp~-l I p~-l ( p~-l 1S a convolution factor of pt;-l ) . This factorizations s
property characterizes the class of infinitely divisible random measures:
THEOREM 11 .2. Let ~ be a nxndom measure on S z.,)i th H: (M Then
&s infinitely divisible iff s ( S a.e.
a.c.rr- l'5
c(we
1\ much stronger result, to he proved first, holds in the point process
case:
-154-
THEOREM 11 .3. Let E;. be a point procesB on S l,lith EE;. E M 'l'hen
E;. is infinitely divisible iff P{E;.f = O} > 0 and P(E;.f)-l
for every f E F with E;.f ~ 0c
PROOF.
P(E;.f) -1
If E;. is infinitely divisible and f E F with E;.f ~ 0 , thenc
I P(E;.ff)-l follows by the above calculations. Furthermore, Lemma
6.1 yields P{E;.B = O} > 0 B E B , which implies P{E;.f = O} > 0 ,
Conversely, let f E F be fixed withc E;.f ~ 0 and suppose that
f E Fc
Writing ~ = LE;.f ' it follows easily that
~'(t) = ~(t) Ee- ntEF;f t 2> 0 ,
for some random variable n E Z+ ' so if we divide by ~(t) and integrate
from 0 to t , we obt'ain
where
-1T
at + A(l - e t)
a = Pin = O}EF;f; A(dx)-1 -1
x Pn (dx)EF;f x > 0 .
By Lemma 6.1, this proves that E;.f is infinitely divisible. The sufficiency
assertion now follows by Problem 6.5.
Proof of Theorem 11.2. Suppose that IH-- l I -1s Pi';,s
S f: S a .c. ,: F, •
To show that ~ is infinitely divisible, it suffices hy Lemma 6.4 to
pr'ove that U:;BI
, ... ,E,Bk
) IS infinitely divisihle for every k ( N
and every disjoint '\, ... ,1\ ( 13 . Since clearly- 1
Jl r. for
-155-
any B E B with ~B ~ 0 , it is enough to prove the theorem for finite
S , say for S = {l, ... ,k}. For fixed c > 0 and s E S with ~{s} ~ 0 ,
cCE;s{l}, ... '~s{k}) respectively.
and hence also Pn-1 I PCn' + 0 )-1s
Pn,-l
~ = Lc~ , we further obtain
c C~{l},
Writing
, ~{k}) and
From Cl.4) it is seen that Pn- l Ibe Cox processes directed byn'andnlet
from C1. 4)-t -t
~ ~ (l-e 1, ,l-e k)
~ ~ CO, ,0)e-t
s
t 1 , , t k E R+
and so n' + 0S
proving that Since S E S was
arbitrary with ~{s} ~ 0 , we may conclude that PCnf)-l I PCnff)-l for
every f E F , and it follows by Theorem 11.3 thatc n is infinitely
divisible. The factor c > 0 being arbitrary, this implies by Lemma
8.6 that t; itself is infinitely divisible, and the proof is complete.
We are now going to show that it is enough in Theorem 11.2 to consider
one particular mixture of Palm distributions, provided we add an assumption
about the constant part of ~. Given any f E F with t;f < 00 a. s., we
define the non-random set function f~ on B by ft;B = essinf (f~)B =
= inf{x > 0: P{(f~)B < x} > o}. Since ft; is automatically finitely
superadditive and continuous at ~, only finite subadditivity is needed
to ensure n to be a measure. This holds in particular if essinf E;f = 0 .
-156-
THEOREM 11.4. Let ~ be a random measure on S and let f E F be strictly
positive with 0 < E~f < 00
finitely subadditive and
is infinitely divisible iff f~ is
PROOF. Suppose that ~ ~ I(a,A) . Then f~ = fa EM, (this is a simple
special case of Theorem 6.9), and it follows in particular that f~ 1S
subadditive.
Suppose conversely that f~ is subadditive (and hence a measure)
and that p~-l I p~il. Since f is strictly positive by assumption,
it is easily seen that ~ and f~ are simultaneously infinitely divisible,
and so it is enough to assume that f = I = IS' By Lemma 6.4, it suffices
to prove that (~Bl"" ,~Bk) is infinitely divisible for every kEN
,Bk E B. From the assumed factorization property it is
seen that
E[~S exp(-L.t.~B.)]E~~ J J = E exp(- r.t.~B.) E exp(~ L.t.n·)• JJ J JJJ
~
for some R+-valued random variables nl , ... ,nk ' and writing ~
we get
~ - I
(1)E[~S exp(-L.t.~B.)J E~S E exp(-L.t.n.)-~S
J J J Y. ~ J J J -----~......<---"--- = E exp ( - .t. ~B. )
~ 'J J JE~S E~S
Thus the numerator on the right is positive, and letting t l , ... , t k -~ 00
we obtain E~S P{n l = ••• =n k = O} 2 IS. But then the second factor on the
right of (1) must be an L-transform, and (1) shows that the assumed factori-~
zation property of ~ remains true for ~. We may therefore assume from
now on that ~ = 0
-157-
Let us now suppose that 11 = (11 0 , 11 1 , . o. ,11k) is;) Cox procc s s on
{O,l, ... ,k} directed by (I;;S, I;;B I , ,I;;Bk) 0 Proceeding as in the
proof of Theorem 11.2, it is seen that Pll-1 I Pll,-l where Pn,-l denotes
the Palm distribution of n w.r.t. the point 0, and if ~ is the gen-
erating function of n , we obtain
."
(2)t ~ I (t, t 1 ' ... , t k)
Ei;;St, t l , ... ,t
kE [0,1] ,
for some Z+-valued random variables S, sl' ... ,sk ' where the prime stands
for differentiation with respect to t. Now
11HO, t l , '0. ,tk) = E[t l
l
so it follows from (2) that t is a factor of the expectation on the
right, i.e. that P{s = a} = a Rewriting (2) in the form
1/J'(t,t l ,···,tk) 1 sl sk
H t , t 1 ' . . . , t k ) = EI;;S Et s - tl t k
and integrating w.r.t. t from a to 1, we thus obtain
In particular
a - log 1/J(I, ... ,1)-1
- log HO,I, ... ,1) - E~S Es
and hence by subtraction
- log 1/J(1, t l , ,tk) =
log HO,I, ,1) - log 1/J(0, t l , ... ,tk) + I:F;S Es -1
(1 - t:l
Since
-158-
-1Es < 00 , it is seen from Lemma 6.3 that the function e on
[O,l]k defined by
tl
, ... ,tk
E [0,1] ,
1S the generating function of some infinitely divisible random vector
on Zk+
Letting
and writing
for any p E
be a Cox process directed by
~p for its generating function, we obtain more generally
(0,1]
(3) - log ~p(l, t l , ,tk) =
= log ~p(O,l, ,1) - log ~p(O, t l , ... ,tk) - log ~(tl' ... ,tk)
where Gp is an infinitely divisible generating function. But n is
by Corollary 8.5 a p-thinning of n , and so by (1.6) and (3)p
where q = 1 - p. Here the last term corresponds by Lemma 8.6 to an
infinitely divisible distribution, so if we can show that
(4 ) log t/Jp(O,I, ... ,I) - log ~p(O, q + pt l ....• q + ptk) + 0 as p + 0 ,
-159-
it will follow that ~(l, t l , ... ,tk) is the limit of a sequence of
infinitely divisible generating functions, and hence itself infinitely
divisible.
.'
To prove (4), put ¢ = L , and note that by (1.4)~S, ~Bl ' ... , ~Bk
E exp{-p-l(~S(l - t) + L.~B.(l - t.))} =J J J
Hence
(-1 -1 -1)
= cpp (1 - t), P (1 - t 1)' ... ,p (1 - t k )
while
~ (0,1, ... ,1)P
-1 -1= ep(p , 0, ... ,0) = E exp (-p ~S) •
-1~ ~P (0 , q • ... , q) = cp (p ,1, '" ,1 )
= E exp(-p-l~S - L.~B.) ~ E exp(_(p-l + k)~S) .J J
Now it may be seen as in the proof of Theorem 11.3 that d~S = some l(a,>') ,
and since ~ = 0 we have a = O. Thus by Lemma 6.1, the difference in (4)
is at most equal to
log E exp(_p-l~S) - log E exp(_(8P-l + k)~S) =
which tends to zero as p + 0 by dominated convergence. This proves
that the Cox process (n l , ... ,nk) directed by (F,31 , ..• ,£:1\) IS
infinitely divisihle, and exactly the same argument shows more generally
-160-
that Cox processes directed by c(~B1'
for any c > o. By Lemma 8.6,
. .. , ~Bk)
(~B1' ...
are infinitely divisible
,~Bk) is then itself
infinitely divisible, and the proof is complete.
We shall now prove a related but quite different criterion for infinite
divisibility which applies to the class of purely atomic random measures
with diffuse intensity. Given any measure ~ E M and any E > 0 , we
define the measure ~ E M byE
~ B =E
B E B ,
and note that ~E is discrete in the sense that it is purely atomic with
finitely many atoms in each compact set. Note also that the mapping
~ ~ ~E is measurable by Lemma 2.1. Given any random measure ~ on
S , and any B E Band E > 0 with P{~ B = o} > 0 , we further defineE
the random measure ~B,E on S as one with distribution
P{~B E M} = P{~ E M I ~ B = o},E E
M EM.
THEOREM 11.5. Let ~ be a purely atomic random measure on S with
Md Then is infinitely divisible iff-1 I p~-l for everyE~ E ~ PB,E
E > 0 and B E B 7,n th P{~ B = o} > 0 . In the case 7'Jhen ~ &s a. r,.E
discrete~ this statement remains true w1:th E = 0
PROOF. The necessity is an immediate consequence of Lemma 6.6 and the
fact that Poisson processes have independent increments. To prove the
-161-
is independent of nB . For
for every B E: B satisfying P{~B = O} > 0
that ~d
~B where ~ ~ ~B 0assume + nB B ,s ( B a.e. E~ and any f E: F we obtainc
-1 I -1converse, let us first assume that ~ is a.s. discrete and that P~B,O P~
For any such B we may
L[ (f)"'5
(5)
and this holds simultaneously for all B belonging to some countable
base G and all $ E: B except for points s E: S in some fixed E~-null
set C. For every fixed s 4 C , consider a sequence Bl ,B2 , .. , E: G
such that Bn + {s} Since ~ is a.s. discrete and has no fixed atoms,
have P{~B O} 1 and ~Bd
~ Hence by (5) and Lemma 5.1,we -+ , so -+n
d p~-ln
nB s -+ some n , and we get p~-l We may thus conclude fromsn
Theorem 11. 2 that ~ is infinitely divisible.
We now consider the general case al1d hence assume that p~;~ I pt;-l
whenever p{~ B = O} > 0E
Writing d~ = t;BE + n BE with independent terms,
we get in place of (5)
and so
< 2
-162-
IL~ (f) - Lt,;(f)Ln (f) Is BES
Et,;BE(ds)E~(ds) + ILt,;BE (f) - L~(f) I
-t,;fE[~(ds);t,;cB=O] E[e ;t,;cB=O]
2 ~ ~ E -~f= =E-t,;(7'd-:--s-=-)=p"""{t,;=--=B-=-=-O'"} + ---'P"""{'-t,;---=-B=-'O"'""'}-- - .e
E E
E(t,;-t,;E) (ds) 1
2 ---=E=-:"t,;-=(-=-ds"""""')-- P{t,; B= O} + 2E
l-P{t,; B=O}E
P{t,; B=O}E
As before, this holds for all s outside some E~-null set C, simultaneously
for all E in some sequence E tending to 0 and all B with P{t,; B=O} > 0E
belonging to some countable base G. Now ~ - ~E X0 a.s. as E ~ 0 , so
E(~ - t,;E) X0 by dominated convergence. Choosing g (s) :: E(t,; - ~ ) (ds)/E~(ds)E E
non-decreasing in E E E , it follows from this by dominated convergence that
g (s) ~ 0 as E + 0 for s E S a.e. Et,;, say for s outside some null£:
set C' Furthermore, every ~E is clearly a.s. discrete and has no
fixed atoms, so for fixed E we have P{t,; B = O} ~ 0 asE
B -} {s}
Given arbi trary n €. Nand s 4 C u C' , we may therefore choose f. ( E
so small that g (s) < (8n)-1 and then BEG so small thatE
-1p{[ B > O} < (8n) . Writing nn for the corresponding random measure-E
nBES ' we then obtain
-1< n n E N f <: r
c
proving that By Lemma 5.1 it follows that some
-163-
. f' L L L d' p~-l I p~-l . A b f h" l'satls ylng t;, n = t;, , an agaln s Ss s e ore, t lS lmp less
infinite divisibility of t;" and the proof is complete.
We are now going to consider extensions of Proposition 11.1 in another
direction.
THEOREM 11.6. Let ~ be a point process on S lJith Et;, E: M Then
P(~s6 )-1 independent of S Et;, iff t;, d
M(Et;" tjl) fOY'- 1,fj S E: a.e.s
tjl Tn this tjl'(O) -1 and t;,s 6d
M(H, -tjl')some . case = -' - a.e.s
It is interesting to see how the specialization of Proposition 11.1
to point processes follows from this result simply by solving the differential
equation tjl = -tjl' subject to the initial condition tjl(O) = 1
The proof of Theorem 11.6 is based on the following lemma, giving
an interpretation of the notion of Palm distributions in terms of ordinary
conditional distributions.
LEr~MA 11. 7. Let ~ be a p01:nt process on S with EE; E: M -' and let
n I n , C E. B and M E: N For ~C > 0 u)C fuy,theY' asswne that T
1AJ one of ihe atoms of E, &n C -' chosen at random. Then
p{~ E. M I ~C = n, T = s}
a. e. u)i th Y'eBpect to the measure
P{t;, E: MIt;, cs sn} s ( C ,
E[E:(ds); ~C nJ = P{ E; Cs
n}EE;(ds) s c S .
PROOF. The n atoms in C being symmetric, we get for see
-164-
P{E; E M, E;C n, T E ds} = E[o (ds); t;, E M, E;C = n]T
-1t;,C = n]= n E[t;,(ds); t;, E M,
= n-lp{t;, E M, t;, C = n}Et;,(ds)s s
and in particular
P{t;,C = n, T E ds}
By the chain rule for Radon-Nikodym derivatives, we thus obtain a.e.
p{E; E M I E;C = n, T = s} = P{t;,EM, E;C=n, TEds}P{ E;C=n, TEds}
P{ t;, EM, t;, C=n}s sP{ t;, C=n}
s
as desired.
= P{E; E MIt;, C = n} ,s s
Proof of Theorem 11.6. Suppose that dt;, = M(Et;" ¢) . As shown in Section
9, E;B has then the generating function ¢(EE;B(l - t)) , t E [0,1] ,
and we get in particular
P{E;B = l} = d~ ¢(EE;B(l - t)) I t = ° = - EE;B ¢' (EE;B) B E B .
Using this fact and Corollary 9.2, we obtian for any B,C E B with B c C
f P{ (E;. - 0 ) B = o}EE; (ds )r: s s
= f P{E; Br: S
l}EE;(ds) = qr,C; r,B 1]
= P{E;B = l}E[~C I eB
- cf>'(EE;B)Ef,C .
1] _. 1;[,13 ,j,' ((;[,13) H.C'I' 1:F;B
On the other hand we get hy Corollary 9.2 for any disjoint B,C E B
-165-
I P{(I: -8)B O}EI:(ds) = Ic P{I: B O}Eqds) E[I:C; I:B 0]C S 5 5
00
I nP{I:C n, E;B = O}n=l
00
= I np{qB u C) = n} P{ E;C = n I E;(B u C) n}n=l
00
{ HC r= I nP{t;(B u C) n} EE; (BUC)n=l
00
EE;C d \' nEqBuC) dt L. t P{E;(B u C) = n} I
n=O t EE;Cjl:[, (sue)
EE;C d ( ) I= EqBuC) dt <PEE; (B u C) (1 - t)t = EE;CjEE;(BuC)
EEC ( )EE;(BuC) EE;(B u C) <P'EE;(B u C) - EE;C
- - <P' (EEB) EE;C
These calculations show that for any B E B
P{(~ - 8 )B = O} = - <P' (EEB)5 5
5 E S a.e. EE;.
If EE; c Md ' then
by Theorem 3.3 that
to evaluate
[, and hence also E; - 8 is simple, an<1it followss s
E - 8 ~ M(EE; -<P') Tn the general case we have5 5 ,.
n l' ... , (E; - 8 ) Bks 5keN
But these probabilities arc determined hy the expectations
-166-
E[s(ds); sBl = n l , ... ,sBk = nk] , so they are uniquely determined by
the finite-dimensional distributions of s which in turn depend on Es
and ~ only. Thus it makes no difference whether Es E Md or not, and
the above conclusion for diffuse Es must be generally valid.
Conversely, suppose that c - 0"'s s
d some n a.e. Es . Let C E B
and n c N be fixed with n} > 0 , let ,B E Bm be an
arbitrary disjoint partition of C and let k l , ... , k E Zm +
T as in Lemma 11.7, we get for s E Bl a.e. E[E;(ds); sC := n]
P{(s - 0,) Bl:= kl , ... , (s - o ) B := k I sC := n, , :: s}, m m
Defining
, sB := k I sC := n, , = s}m m
= P{ ssBl:= k + 1, ssB2
:= k2 , ... , ssBm:= k I s C := n}
1 m s
= P{(s os) Bl:= k l , ... , (ss - o )B := k I (ss - o )C := n - l}
s s m m s
= P{nB l:= kl , , nB = k I nC := n - l}
m m ,
and by the symmetry in Bl , ... ,Bm ' this remains true for SEC a.e.
E[E;(ds); E;C := n] . By Dynkin's theorem, we can extend this result to
P{E; - 0 E M I sC = n, ,, s} := P{n E M I nC n - l}
proving that , and s - 0 are conditionally independent, given that,
in C , we may easily conclude that
is a random cnumeration of the atom positionsr,C = n . If Tl
," n
,.J
is conditionally independent of
h.,i~j}]
for every j E {l, ... ,n} But then must
Ilc conditionally independent, and since they arc also equally distributcd,
-167-
Ci; is conditionally a sample process. Furthermore, we have for any
B E B with B c C
oW
-lE[eB I eC = ] = E[i;B; i;C=n]n ,s s ,n n P{i;C =n JPh 1 ( B I i;C n} E[o B I i;C = n]
'1
IBP{i;sc=nJEi;(dS) = ]BP{(i;s-os)c=n-1H:i;(dS)
n P{i;C=n} n P{i;C=n}
=P{nC=n-l}Ei;B
n P{i;C=n}
= Ei;Bfcp{(i;s-os)c=n-l}Ei;(dS)
n Ei;C P{i;C=n}
= Ei;B E[SC; i;C=n]n Ei;C P{i;C=n}
Ei;Bn Ei;C E[i;C; i;C=n] =
Ei;BEi;C '
showing that the common conditional distribution of the ,.J
is independent
of n. It follows that Ci; is unconditionally a mixed sample process,
and an appl ication of Coroll ary 9.2 completes the proof.
We shall finally consider an extension of Theorem 11.6 to gener;l1
random measures. I{ecall that a measure is deycner'ate if its mass
IS confined to one single point. If n is a random variahle while 1';
is a random measure on S , then (n,s) is said to be symmetrically
distributed w.r.t. some w EM, if the distrihution of (n, sBl'
for arbitrary kEN and disjoint Bl
, ... ,Bk E B only depends on
For convenience, the random measures descri hed in parts
(i) and (jj) of Theorem 9.4 arc said to he of Type 1\ and B respectively.
We shall a 11 ow the measure w there to he fi n i te in both cases. Note
that the two classes of random measures will then no longer he mutually
exclusive. The canonical random elements (lX,/-) and (a,()) :lrc tlefined
as in Theorem 9.4, and in case of Type 1\ random measures, we further
-168-
define the canonical random measure A on R+ by A(dx) = aoO(dx) + x\(dx) ,
x z 0 .
THEOREM 11.8. Let ~ be a random measure on S ~ith E~ E M Then
d~ - ~ {s}o ) = some (n,s) independently of s a.e. EF, iffs s s
(except possibly for a.s. degenerate ~) and ~ is symmetrically
distributed ~.r.t. E~ In this case, (n,s) is also symmetrically
distributed ~.r.t. E~, and furthermore, n is independent of s iff
either
(i)
(ii)
~ 'l-S of Type A ~ith A = pM a.s. for some random vaY'1:ab le
p ~ 0 and some non-random M E ~~ (R ) , or+
F, is of Type B ~ith a = 0 and S a mixed sample process
on R'+
Note by comparison with (i) that the point process S in (ii) is
also allowed to be a mixed Poisson process. In this case we can even
take a non-zero, proportional to the mixing variable.
To save space, we omit the rather complicated proof of this theorem
and refer the reader to Kallenherg (1974b).
NOTES. Proposition 11.1 was proved for point processes by Kerstan and
Matthes (1964) and Mecke (1967), but the prescnt proof is adapted from
.lagers (1973). Theorem 11.2 is clue to Kerstan and Matthes (1~)64), Mecke
(1~)67), Kummer ancl Matthes (1970) and others (cf. Kerstan, ~1atthes and
Mecke (1974), page 116), while the extensions given in Theorems 11.3 and
-169-
11.4 are new. Theorem 11.5 extends a result for point processes, proved
by Matthes (1969) with entirely different methods, (cf. Kerstan, Matthes
and Mecke (1974), page 97). Finally, Theorem 11.6 is due to Slivnyak
(1962), Papangelou (1974) and Kallenberg (1973a), while Theorem 11.8 was
given in Kallenberg (1974 b).
PROBLEMS.
11.1. Let r; be a point process on S with Et; E: M and let I c 13
be a DC-semiring. Show that t; is infinitely divisible iff
kfor every function f = I II ' k E: N ,
j =1 j
II' ... ,Ik E: I , such thatdt;f ~ 0 •
11.2. Suppose that S is countable. Show that there exists some fixed
such that, given any point processfunction
is infinitely divisible iffwith
and
f E: F
t;S < 00 a.s., t;
P(t;f) -1 I P(t;ff)-l (Hint:
E, on S
P{t; = O} > 0
use the result in Problem 6.4.)
11.3. Let t; be a point process on S and let B E: 13 be such that
Pit; ~ 0, E,B = a} = O. Show that E, is infinitely divisible iff
Pit; O} > a and pt;-l I pr;;l f = lB (Hint: argue with r;
as with the Cox processes in the proof of Theorem 11.4, and note
that the counterpart of (4) holds trivially in the present case.)
11.4. Show that the condition Er;; E: Md in Theorem 11.5 can he replaced
hy the assumption that EF, be a-finite and diffuse.
-170-
a.s., and let
~ ~ 0 has the con-
11.5. Let ~ be a random measure on
l be a random element in S
ditional distribution ~/~S
measurable f: M x S -)- R ,+
S with ~S < 00
which for given
Show that, for u > 0 S E S and
E[f(~,T) I ~S = U, T = s] = E[f(~s' s) I ~sS = u]
holds a.e. E[~(ds); ~S E du] , (cf. Kallenberg (1974b)).
REFERENCES
(1) Bauer, H.
New York.
(1972). Probability Theory and Elements of Measure Theory.
Holt, Rinehart and Winston.
(2) Belyaev, Yu. K. (1963). Limit theorems for dissipative flows.
Theor. Probab. Appl. ~, 165-173.
(3) Belyaev, Yu. K. (1969). Elements of the general theory of random
streams. (in Russian). Appendix 2 to the Russian edition of Cramer
and Leadbetter, Stationary and Related Stochastic Processes. Moskow.
MIR.
(4) Billingsley, P. (1968). Convergence of Probability Measures. New
York. Wiley.
(5) Bllhlmann, H. (1960). Austauschbare stochastische Variabeln lmd thre Grenz
wertslitze. Univ. California Publ. Statist. 3, 1-36.
(6) Daley, D. J. (1974). Various concepts of orderliness for point
processes. Tn Stochastic Geometry. pp. 148-161. New York. Wiley.
(7) Davidson, R. (1974). Stochastic processes of flats and exchangeabi 1ity.
In Stochastic Geometry. pp. 13-45. New York. Wiley.
(8) Feller, W. (1968). An Introduction to Probability Theory and its
Applications. Vol. I, 3rd ed. New York. Wiley.
-171-
(9) Feller, W. (1971). An Introduction to Probability Theory and its
Applications. Vol. II, 2nd ed. New York. Wiley.
(10) Gihman, I. I. and Skorohod, A. V. (1969). Introduction to the Theory
of Random Processes. Philadelphia, Saunders.
(11) Goldman, J. R. (1967). Stochastic point processes: limit theorems.
Ann. Math. Statist. 38, 771-779.
(12) Grandell, J. (1973). A note on characterization and convergence
of non-atomic random measures. Abstracts of communications. T.l,
175-176. International conference on probability theory and mathematical
statistics, Vilnius 1973.
(13) (;rigelionis, B. (1963). On the convergence of sums of random step
processes to a Poisson process. Theor. Probab. IIppZ. ~,172-l82.
(14 )
(15)
Harris, T. E. (1968). Counting measures, monotone random set functions.
Z. WahrscheinZichkeitstheorie verw. Gebiete. ~, 102-119.
Harris, T. E. (1971). Random measures and motions of point processes.
Z. WahrscheinZichkeitstheorie verw. Gebiete. ~, 85-115.
(16) lIin~in, A. (1960). Mathematical Methods in the Theory of Queueing.
London. Griffin.
(17) .lagers, P. (1972). On the weak convergence of superpositions of
point processes. Z. WahrscneinZichkeitstheorie verw. Gebiete. 22.
1-7.
(18) .lagers, P. (1973).
VPPI,). Gebiete. 26.
On Palm probabilities.
17-32.
7,. WahY'schrdn l ichkei tv theop/r
(19) Jagers, P. (1974). Aspects of random measures and point processes.
IIdvances in p~bability and related topics. 3. 179-239.
(20) Kallenberg, O. (l973a). Characterization and convergence of random
measures and point processes. 7,. Wahr'cchein11~ehkwttnth(?(Jr'1'cnerol,).
Gebiete. 27. 9-21.
-172-
(21) Kallenberg, O. (1973b). A canonical representation of symmetrically
distributed random measures. Mathematics and Statistics, essays in
honour of Harald Bergstr~m. 41-48. GBteborg, Teknologtryck.
(22) Kallenberg, O. (1973c). Canonical representati'Ons and convergence
cri teria for processes with interchangeable increments. /'. WahI'[whm:n
lichkeit8theo~ie veI~. Gebiete. 27. 23-36.
(23) Kallenberg, O. (l973d). Conditions for continuity of random processes
without discontinuities of the second kind. Ann. PI'obab. 1. 5l9-52().
(24) Kallenberg, O. (1973e). Characterization of continuous random processes
and signed measures. Studia Sci. Math. Hunga~. 8. 473-477.
(25) Kallenberg, O. (1974a). Extremali ty of Poisson and sample processes.
J. Stoch. Froc. Appl. 2. 73-83.
(26) Kallenberg, O. (1974b). On symmetrically distributed random measures.
T~ans. Ame~. Math. Soc. (to appear).
(27) Kallenberg, O. (1975). Limits of compound and thinned point processes.
J. Appl. P~obab. 12. (to appear).
(28) Karbe, W. (1973) . Konstruktion einfacher zuflill iger Punkt fo1 gen.
Diplomarbeit, Friedrich-Schil1er-Universittlt .lena, Sektion Mathematik.
(29) Kerstan, J. and Matthes, K. (1964). Verallgemeinerung cines Satzes
von Sliwnjak. Rev. Roumaine Math. Pu~cs Appl. 9. 811-829.
(30) Kerstan, J. and Matthes, K. (1965) . A gcneralization of the Palm-
lIin~in theorem (in Russian). Ukmin. Mat. I.. 17. 29-36
(31 ) Kerstan, J. , Matthes, K. and Mecke, J. (1974). lJnbegrenzt Tcilbare
Punktprozessc. Berlin, Akademi-Verlag.
(32) Kingman, J. F. C. (1967). Complctely random mC:Jsures. l'acific
J. Math. 21. 59-78.
(33) Kolmogorov, A. N. and Fomin, S. V. (1970). Introductory Re:Jl i\naly:.;i~.
Englewood Cliffs, Prentice-Hall.
-173-
(34) Kummer, G. and Matthes, K. (1970). Vera11gemeinerung eines Satzes
von Sliwnjak II-III. Rev. Roumaine Math. Pures Appl. ~, 845-870,
1631-1642.
,"
(35) Kurtz, T. (1973). Point Processes and completely monotone set functions.
Tech. report, Math. dept., Madison, Wisconsin.
(36) Leadbetter, M. R. (1972). On basic results of point process theory.
Proc. 6th Berkeley symp. math. statist. probab. I, 449-462. Berke ley,
Univ. of California Press.
(37) Lee, P. M. (1967). Infinitely divisible stochastic processes.
Z. WahY'scheinlichkeitstheoY'ic VCr'l,). Geli1:ete. 7. 147-160.
(1970). Characteristic Functions. 2nd ed. London.
(1963) . Probability Theory. 3rd ed. Princeton, Van
(1963). Unbeschrankt teilbare Verteilungsgesetze stationorere
Punktfolgen. Wise. z. Hochsch. ElektY'o. Ilmenau. 9.
(1969). Eine characterisierung der kontinuerlichen
unbegrenzt tei1baren Verteilungsgesetze zufalliger Punktfolgen.
[{eVU8 Roumaine Math. PUY'es Appl. 14. 1121-1127.
(38) Lo~ve, M.
Nostrand.
(39) Lukacs, E.
Griffin.
(40) Matthes, K.
zuHitl iger
235-238.
(41 ) Matthes, K.
(42) Mecke,.J. (1967). Stationare zufall ige Mosse auf lokalkompakten
Abe1schen Gruppen. 7.. WahNlcheinUchke1:tsthcoY'ic ver'l,). Gehiete.
9.36-58.
(43) Mecke, J. (1968). rine characteristischc Eigenschaft JCT doppclt
stochast ischen Poi ssonschen prozcsse. /'. h'ohr',:dz('{nliehkc1' L,:f)W()f'ir'
()(~ r'lJ). (;elJ1: (::I;e . 11 . 7'1 - R] .
(tl4) Mecke, .J. (1972). Zuf:illjge Maf)e ,lUi' lokalkompaktcn IlausJorffschen
W'iumen. Be1:tr'?t(!e ;;W' Analysis. 3. 7-:)0.
(45) Monch, G. (1970). Verallgemeinerung eines Satzes von Renyi. Disserta
tion, Technischc Ilochschule Karl-Marx-Stadt, Section Mathemat iL.
-174-
(46) Manch, G. (1971). Vera11gemeinerung eines Satzes von A. R~nyi.
Studia Sci. Math. Hungar. 6. 81-90.
(47) Nawrotzki, K.
Punktfo1gen.
(1962). Ein Grenzwertsatz fUr homogene zuf~llige
Math. Nachr. 24. LOl-217.
l48) Ososkov, G. A. (1956). A limit theorem for flows of homogeneous
events. Theor. Probab. Appl. 1. 248-255.
(49) Palm, C. (1943). Intensit~tsschwankungenin Fernsprechverkehr.
Ericsson Technics. 44. 1-189.
(50) Papangelou, F. (1974). On the Palm probabilities of processes of
points and processes of lines. In Stochastic Geometry. 114-147.
New York. Wiley.
(51 ) Prohorov, Yu. V.
Math. Dokl. 2.
(1961).
539-541.
Random measures on a compactum. Soviet
(52) Renyi, A. (1956). A characterization of Poisson processes. (In
Hungarian wi th Russian and English summaries.) Magyar Tud. Akad.
Nat. Kutato Int. Kozl. l, 519-527.
(53) Renyi, A. (1967). Remarks on the Poisson process. Studia Sci.
Math. Hungar. 2. 119-123.
(54) Ryll-Nardzewski, C. (1961). Remarks on processes of calls. Proc.
4th Berkeley Symp. Math. Statist. Probab. 2. 455-465. Berkeley,
lJniv. California Press.
(55) SimlIlOns, G. F. (1963). Introduction to Topology and Modern Analysis.
New York. McGraw-Hill.
(56) Slivnyak, I. M. (1962). Some properties of stationary flows of homogeneous
random events. Theor. Probab. fippl. 7. 336-341. 9. 168.
(57) Waldenfels, W. von (1968). Characteristische FlInktionale zuf!illiger
Ma13e. Z. Wahr'DcheinlichkeitDI,hcoY'1:e un'),). (,'ehieLf'. 10. 27~)-2R:i.
(58) Westcott, M. (1973). Some remarks on a property of the Poi sson
proccss. Sankhya~ Scr. 1\, 35, 29-34.
-175-
AUTHOR INDEX..
Bauer 4 9 41 45 137
Be1yaev 29 111
Bi 11 ings ley 4 40 43 44 47 51 52 54 55 62 65 72 91100 101 104 109 142 143 145 149
BUh1mann 134
Daley 29
Davidson 134
Feller 86 116
Fomin 6
Gihman 29
Goldman 98 III
Grandell 37 55
Grige lioni s 98
Harris 37 55 65
Hin~in 98 149
Jagers 55 86 98 149 150 168
Kallenberg 1 17 29 37 55 56 67 86 98 III 134 135136 149 150 168 169 170
Karbe 65
Kerstan 1 17 38 65 86 III 134 135 149 150 168 169
Kingman 98
Ko1mogorov 6
Kummer 37 111 149 168
Kurtz 55 65
Leadbetter 29
Lee 86
Lo~vc 12 38 62 73
Lukacs 87 112
M..'ltthes 1 17 37 38 65 86 111 134 135 149 150 16R 169
Mccke 1 17 37 38 65 86 III 134 ns 149 ISO 168 16~)
Monch 37
Nawrotzki III 134
Ososkov 98
-176-
Author Index (cont.)
Palm 98 149
Papange10u 149 169
Prohorov 37 SS 65
Renyi 37 III
Ryll-Nardzewski 149
Sierpinski 9
Simmons 63 85 138
Skorohod 29
Slivnyak 149 169
Wa1denfels 37 S5
Westcott 13S
-177
DEFINITION INDEX
bounded set
Campbell measure
Campbell's theorem
canonical rantlom clements
completely monotone
compound
covering
Cox process
DC-ring, DC-semiring
degenerate
diffuse
discrete
tlistribution
Dynkin's theorem
intlepcntlent increments
infinitely divisible
intensity
L-· transform
mixed Poisson process
mixed sample process
mixing
nested partitions
null-array of partitions
null-array of random measures
Palm distribution
point process
Poisson process
random measure
rcI!, lda r
sample process
simple point process
symllletrically distributed
6
136 152
138
120
63
17
7
16
6
87
18
160
11
9
87
67
11
11
16
14
12
2S
20
66
137
10
16, 17
10
24, 28, 48, 92, 93
14
28
113, 167
-179-
SYMBOL INDEX ,
p 6 aB 26
S 6 1;* 28
B 6 ].1 V 30
B 6d
31=F=F(S) 6 II . II 32
F =F (S) 6 d39-+
C C
M=M(S) 7 BI; 40
N=N(S) 7 ~, ~ 41
Z 7 ~ 42+wd].1f 7 ~ 52
M 7 t!.A 63
N 7 1Tt
,1Tf
,1TV
67
BC7 K 67
--B 7 lims i 68
f].l 8 R' 68+
Bp 8 E 68£
113
8 I(a,A),I(A) 78
0(' ) 9, 18 M(w,<P) 114
(rt,A,P) 10 P Q R 129-1 X,].1' X,].1' x
PI; 11 P 136s
EE.: 11 !;s 138~
I.e 11 !;f,l;f 140
0 14 f\ 152s st~J 18 I 153
Po 18 ft; 155
\{' 19 lJE: 160+
Z 19 E, IbO+ 13,£
I•I 20
11* 20,1J' 20
f
r k '"J h) 22" "
-180-
INDEX OF SOME SCATTERED TOPICS
atomic properties
canonical representations
compound point processes
convergence in distribution
Cox processes
independent increments
infinite divisibility
mixed Poisson and sample proc.
null-arrays of random measures
Palm distributions
Poisson processes
sample processes
simplicity and diffuseness
symmetric distributions
tightness criteria
thinned point processes
uniqueness
18ff, 56, 80, 99
68ff, 89, 120, 133ff
17, 32, 36, 99ff, 112, 135
39ff, 57f, 67ff, 87ff, 99ff, 142ff
16, 30, 32, 38, 56, 105, 109ff, 125, 157f
87ff, 105, 119
67ff, 87ff, 112, 127ff, 152ff
14, 16, 113ff, 163, 168
66ff, 89ff, 147
136ff, 151ff
16f, 78, 91f, 118,147,151
14, 118, 167
28, 32ff, 48ff, 58, 64, 93ff, 114ff
113ff, 167f
44, 70
17, 30, 34, 56, 100, 109ff, 112, 125, 136, 158
30ff, 93, 105, 141