Kingman1967 CompletelyRandomMeasures
Kingman1967 CompletelyRandomMeasures
Kingman1967 CompletelyRandomMeasures
> {A)
(4) 0 g \t{AL) g oo ,
and we have
(Here an expression of the form 'Z a.s/ stands for P{Z} = 1.)
If Au A,, . . . are countably many disjoint sets in @, with union
A, then
(7) = µeS;
62 J. F. C. KINGMAN
f= (x) , ,
where
(i) f and x ar e independent, completely random measures
satisfying ^,
(ii) f is given by
(14) /= ^<p(x)8m,
(15) }) = 0 a.s.,
X1(Anj) = a/n .
Thus
E\e u
^\ = e~aln ,
(A) = £ (Anj) ,
and therefore (A) must be infinitely divisible (or, in the terminology
of [9], decomposable).
Since (A) is nonnegative, it follows from the Levy Khinchin re
presentation theorem (in the special form for nonnegative variables
given for instance, in [6]) that there exists a finite measure (A, .)
on the class 33+ of Borel subsets of R+, such that, for t ^ 0,
where
Then, for any countable dissection of E into Borel sets Em, we have
(19)
m n n m
(20) f J( ^ = ( Jdµ%.
J B+ n n J R+
Now the function which sends A into the measure (A, •) de~
fined in the previous section is a function from % into ^ . We show
first that is additive, and then that it may be extended from §>
COMPLETELY RANDOM MEASURES 65
Since this holds for all t ^ 0, it follows that (A, .) = '(A, .), and
hence we have proved that, if {An} is a measurable dissection of A e §,
then
(21)
Here {Cn} is the dissection occurring in condition <g=% and the defini
tion has meaning since A Cn e §. Moreover (21) shows that
(A, .) = (A, .)
for Ae$, so that x extends from § to @. Again, it follows at
once from (19) that is additive in its first argument Finally, for
any Ae&,
oo
Remarks. (1) Since, for fixed , fc( , z) is bounded away from zero
and infinity, (A, R+) < oo if and only if t(<A) < ^ , i.e. if and only
if Ae$.
(2) A knowledge of determines E{e~t U)} for all , and hence
determines the finite dimensional distributions of . Conversely, for
a given , is given uniquely by (23).
(3) Since the idea of a function of two arguments which is
additive in each is perhaps unfamiliar, it may be helpful to remark
that there is a one to one correspondence between the class of func
tions satisfying conditions (i) and (ii) of the theorem and the
measures * on the product space S x R+, given by
t U)
(25) E {e } = e x p { 1 fc(t, z) *(A x dz)\ .
Then πv(m, z) and rv(.) are measures on S which are finite on 3, and
are thus finite (by ^ ) , and moreover πv(., z) is absolutely continuous
with respect to 7 V(.). For any z in the set Q+ of finite rationals in
R+ let i^(., z) be (a version of) the Radon Nikodym derivative of
7 V(., ^) with respect to v(.), so that
= lim I Fv(x, z + ^ ^^
so that, for Tr^ almost all x, Fv(x, z +) = F (x, z). A similar argument
shows that, for ^ almost all x,
lim Fv(x, n) = 1 .
n—>oo
(31) p9(x, (z , z2]) = Fv(x, z2) F9(x, zj, (z, <z2e Q+) ,
which satisfies
Hence, by (29),
pv(x,.)(v = l,2, . . . ; ? e S )
on R+ with pv(., E) measurable for each Ee$$+, such that, for all
t> 0,
where
REMARK. It may seem that we can say more about the measures
COMPLETELY RANDOM MEASURES 69
(42) a =
then
(43) = (P, + d + o
(44) 0/ = *(*)«..
Then f is a completely random measure satisfying (39). It is clearly
concentrated on the countable set s/ %
The second component is constructed trivially by putting
(45) d =
(46) <P =
eπ
is a completely random measure satisfying (41).
Thus the proof of the theorem is complete.
If we call two random measures equivalent if they have the same
finite dimensional distributions, then we have proved that any comple
tely random measure satisfying ^ is equivalent to one of the form
implied by equations (42) (46). In particular we have the following
result.
(52) = Q v
v =l
(53) π=X πv .
v l l
11. All the above analysis has been carried out using the condi
tion ^ and it is worthwhile noting the points at which this condition
enters the argument. If it is not assumed, the measures \t can still
be defined by (3), but they no longer need be finite. Thus, if §>
is the (/ ideal generated by % (the class of countable unions of sets in
3), §> may be a proper sub ideal of @.
The first application of <& occurs in § 4, where it is used to
establish the countability of the set s^f of fixed atoms. If ^ is not
assumed, j ^ may be uncountable, and indeed nonmeasurable. When
this is so, there is no obvious way of 'splitting off' the fixed atomic
component as is done in Theorem 1.
The analysis of Theorem 2 goes through, as long as we restrict
attention to sets in %. Thus is defined for Ae%, and then ex
tended as in § 6 to %. If we then set
shows that <g> does not imply <jg". It is also true that <g" does not
imply <g*; in fact it is possible to construct a completely random
measure on the Borel subsets of R with the following properties
(m denoting Lebesgue measure):
(i) is finite for all ,
(ii) if m{A) > 0, then (A) = co a.s.,
(ill) if m(A) = 0, then <P(A) = 0 a.s.
In this case 3 (and so also §) consists of the sets of Lebesgue me
asure zero and so ^ is not satisfied. The details of the construction
are given in Appendix 3; the result shows the depth of pathology
which can be reached if ^ is not assumed.
— / l " \>n
cn = 2~na~\ µ = £ onµn
Bn = U A k (n = 0 , 1 , 2 , . . . ) ,
74 J. F . C. KINGMAN
then
X(Ak) < 2~ k = 2~ n ,
k=n+l k=n+l
and therefore, if
we have ( oo) = 0.
Hence µn{BJ) = 0 for all n, and so µiB^) = 0, and it suffices to
prove that µ is finite on S — B^. Now
S JBo. = 0 ( 5.)
ft — n u U /* ft L. . . . i /» n J _ ^ /» //
r=n+l
On the right hand side the first n terms are finite on An — Bn, and
it suffices to show that the last is also finite. It is in fact totally
fiaite on An — Bn, since
(55)
COMPLETELY RANDOM MEASURES 75
are independent,
(ii) if µ(A) < 001 then N(A) has a Poisson distribution with mean
µ(A),
(iii) if µ(A) = oof then N(A) = 00 a.s.
Thus, in the terminology of this paper, is a Poisson process with
measure µ if and only if N is a completely random measure with (in
the notation of §3)
(56) Xt(A) =
It is far from obvious that, given a space (S, @, µ), there does
exist a Poisson process as described above. The object of this appendix,
which is the outcome of discussions with Professor D. G. Kendall, is
to show how, under suitable conditions, such a process can be con
structed.
We first remark that, for a Poisson process to exist, µ must be
nonatomic. For, if x e S has µ{x} = a > 0, we have
) = # ({x} ) £ 1 .
It is, of course, possible to modify our definition to avoid this difficulty,
but this seems hardly worthwhile, and we shall be content to assume
that µ is nonatomic.
The obvious way to proceed to the construction of is as follows.
As in §2, the assumption that N is a completely random measure
satisfying (56) leads to a consistent family of finite dimensional dis
tributions for the stochastic process
{N(A);Ae&}.
Hence, by the Daniell Kolmogorov theorem, there does exist a stochastic
process with these finite dimensional distributions. In particular, N
is almost surely a finitely additive, integer valued set function satisfy
ing (i) (iii). The theorem is not, however, sufficiently strong to ensure
that N is additive. Kendall (private communication) has shown how
this sort of difficulty may be avoided by powerful general methods,
but in the simple case of a Poisson process it is possible to proceed
76 J. F. C. KINGMAN
(58) (A) = X (C O + r ) .
r eq
Hence
Now let E be the set of for which (A) = < >, so that m(E) > 0.
From the definition of it is clear that
eE, reQ, + r e (0, l)=* + r eE .
Hence, if G(x) = m{E n (0, a?]}, (0 < x < 1), we have
G(x) = xG(l) =
Thus, for almost all x e (0, 1),
E (x) = G'(»)
and since ^ =0 or 1 and m(E) > 0 it follows that m(J?) = 1. Thus
P{ (A) = ^} = ME) = 1 ,
and the proof of (ii) is complete.
References
1. M. S. Bartlett, Spectral analysis of point processes, J. Roy. Statist. Soc. Ser. B.
2 5 (1963), 264 296.
2. V. E. Benes, General Stochastic Processes in the Theory of Queues, Addison Wesley,
1963.
3. J. L. Doob, Stochastic Processes, Wiley, 1953.
4. J. Feldman, Subinvariant measures for Markoff operators, Duke Math. J. 2 9 (1962),
71 98.
5. P. R. Halmos, Measure Theory, van Nostrand, 1950.
6. D. G. Kendall, Extreme point methods in stochastic analysis, Z. Wahrscheinlich
keitstheorie 1 (1963), 295 300.
7. J. F. C. Kingman, The stochastic theory of regenerative events, Z. Wahrscheinlich
keitstheorie 2 (1964), 180 224.
8. P. Levy, Complement a Vetude des processus de Markoff, Ann. Sci. £cole Norm.
Sup. Ill 69 (1951), 203 212.
9. M. Loeve, Probability Theory, van Nostrand, 1963.
10. C. Ryll Nardzewski, Remarks on processes of calls, Proc. 4th Berkeley Sympos.
Math. Statist, and Prob. 1961, II, 455 465.
11. P. Whittle, On the variation of yield variance with plot size, Biometrika 4 3
(1956), 337 343.
EDITORS
H. SAMELSON J. DUGUNDJI
Stanford University- University of Southern California
Stanford, California Los Angeles. California 90007
ASSOCIATE EDITORS
E. F. BECKENBACH B. H. NEUMANN F WOLF K. YOSIDA
SUPPORTING INSTITUTIONS
UNIVERSITY OF BRITISH COLUMBIA STANFORD UNIVERSITY
CALIFORNIA INSTITUTE OF TECHNOLOGY UNIVERSITY OF TOKYO
UNIVERSITY OF CALIFORNIA UNIVERSITY OF UTAH
MONTANA STATE UNIVERSITY WASHINGTON STATE UNIVERSITY
UNIVERSITY OF NEVADA UNIVERSITY OF WASHINGTON
NEW MEXICO STATE UNIVERSITY * * *
OREGON STATE UNIVERSITY AMERICAN MATHEMATICAL SOCIETY
UNIVERSITY OF OREGON CHEVRON RESEARCH CORPORATION
OSAKA UNIVERSITY TRW SYSTEMS
UNIVERSITY OF SOUTHERN CALIFORNIA NAVAL ORDNANCE TEST STATION