#### Uniquely Determined Uniform Probability on the Natural Numbers

Uniquely Determined Uniform Probability on the Natural Numbers
Timber Kerkvliet 0
Ronald Meester 0
0 Department of Mathematics, Faculty of Sciences, VU University , De Boelelaan 1081a, 1081 HV Amsterdam , The Netherlands
In this paper, we address the problem of constructing a uniform probability measure on N. Of course, this is not possible within the bounds of the Kolmogorov axioms, and we have to violate at least one axiom. We define a probability measure as a finitely additive measure assigning probability 1 to the whole space, on a domain which is closed under complements and finite disjoint unions. We introduce and motivate a notion of uniformity which we call weak thinnability, which is strictly stronger than extension of natural density. We construct a weakly thinnable probability measure, and we show that on its domain, which contains sets without natural density, probability is uniquely determined by weak thinnability. In this sense, we can assign uniform probabilities in a canonical way. We generalize this result to uniform probability measures on other metric spaces, including Rn .
Uniform probability; Foundations of probability; Kolmogorov axioms; Finite additivity
1 Introduction and Main Results
Within the bounds of the Kolmogorov axioms [5], a probability measure on N =
{1, 2, 3, . . .} cannot assign the same probability to every singleton, and therefore, a
uniform probability measure on N does not exist. Despite this, we have some
intuition about what a uniform probability measure on N should look like. According to
this intuition, for example, we would assign probability 1/2 to the subset of all odd
numbers. If we want to capture this intuition in a mathematical framework, we have
to violate at least one of the axioms of Kolmogorov.
One suggestion by De Finetti [3] is to relax countable additivity of the measure to
finite additivity. To see why this suggestion is reasonable, we must first understand
why it is possible, within the axioms of Kolmogorov, to set up uniform (Lebesgue)
measure on [0, 1]. The type of additivity we demand plays a crucial role here. In the
standard theory one always demands countable additivity. If every singleton has the
same probability, in an infinite space, every singleton must have probability zero. With
countable additivity this means that every countable set must have probability zero.
This is no problem if we are working on the uncountable [0, 1], since we still have
freedom to assign different probabilities to different uncountable subsets of [0, 1]. The
interval [0, 1/2], for example, has Lebesgue measure 1/2, while it is equipotent with
[0, 1], which has Lebesgue measure 1. This works because the cardinality of the set
over which we sum is smaller than the cardinality of the space itself.
On N the problem of countable additivity is immediately clear: since every subset
of N is countable, every subset should have probability zero, which is impossible
because the probability of N itself should be 1. In analogy with Lebesgue
measure, we want finite subsets to have probability zero, and we want to be able to
assign different probabilities to countable subsets. To do this, we should change
the type of additivity to finite additivity. In short: since the cardinality of the space
changes from uncountable to countable, the additivity should change from countable
to finite.
Schirokauer and Kadane [8] study three different collections of finitely additive
probability measures on N which may qualify as uniform: the set L of measures that
extend natural density, the set S of shift-invariant measures and the set R of measures
that measure residue classes uniform. They show that N S R where the inclusions
are strict. If a set A N is without natural density, i.e.,
does not converge as n , different measures in L assign different probabilities
to A. So even the smallest collection discussed by Schirokauer and Kadane does not
lead to a uniquely determined uniform probability for sets which do not have a natural
density. This observation brings us to the main goal of this paper.
Main goal find a natural notion of uniformity, stronger than extension of natural density
such that all probability measures that are uniform under this notion assign the same
probability to a large collection of sets. In particular, this collection of sets should be
larger than the collection of sets having a natural density.
In this paper, we introduce and study a notion of uniformity which is stronger than
the extension of natural density. A uniform probability measure on [0, 1] or on a finite
space is characterized by the property that if we condition on any suitable subset,
the resulting conditional probability measure is again uniform on that subset. It is
this property that we will generalize, and the generalized notion will be called weak
thinnability. (The actual definition of weak thinnability is given later, and will also
involve two technical conditions.)
We allow probability measures to be defined on collections of sets that are closed
under complements and finite disjoint unions. This is because we think there is no
principal reason to insist that all sets are measured, just like not all subsets of R are
Lebesgue measurable. We should, however, be cautious when allowing domains that
are not necessarily algebras, for the following reason. De Finetti [3] uses a Dutch
Book argument to conclude that, under the Bayesian interpretation of probability, a
probability measure has to be coherent. He shows that if the domain of the probability
measure is an algebra, the finite additivity of the probability measure implies
coherence. On domains only closed under complements and finite disjoint unions, however,
this implication no longer holds. Therefore, someone sharing de Fenittis view of
probability would like to add coherence as additional constraint. For completeness,
we study both the case with and the case without coherence as additional constraint
on the probability measure.
Definition 1.1 Let X be a space and write P(X ) for the power set of X . An f -system
on X is a nonempty collection F P(X ) such that
1. A, B F with A B = implies that A B F ,
2. A F implies that Ac F .
1. A, B F with A B = implies that ( A B) = ( A) + (B),
2. (X ) = 1.
A coherent probability measure is a probability measure : F [0, 1] such that for
all n N, 1, . . . , n R, A1, . . . , An F
i I Ai (x ) ( Ai ) 0.
A probability pair on X is a pair (F , ) such that F is an f -system on X and is a
probability measure on F .
Remark 1.2 Schurz and Leitgeb [9, p. 261] call an f -system a pre-Dynkin system,
since in case of closure under countable unions of mutually disjoint sets, such a
collection is called a Dynkin system.
Remark 1.3 Expression 1.2 has the following interpretation. If i 0, we buy a bet
on Ai that pays out i for ( Ai ). If i < 0, we sell a bet on Ai that pays out |i | for
|i |( Ai ). Then (1.2) expresses there is no guaranteed amount of net loss.
We aim at uniquely determining the probability of as many sets as possible. In
particular, we are interested in probability pairs with an f -system consisting only of
sets with a uniquely determined probability. So we are not only interested in probability
pairs satisfying our stronger notion of uniformity, but in the canonical ones, where
canonical is to be understood in the following way.
Definition 1.4 Let P be some collection of probability pairs. A pair (F , ) P is
canonical with respect to P if for every A F and every pair (F , ) P with
A F we have ( A) = ( A).
Before we give a more detailed outline of our paper, we need the following
definition. Set
[a2i1, a2i ) : 0 a1 a2 a3 .
i=1
Note that M is an algebra on [0, ). It turns out that by working on [0, ) instead of
N, where we restrict ourselves to sub- f -systems of M, we can formulate and prove
our claims much more elegantly. Here, we view the elements of P(N) embedded in
M by the injection
[n 1, n).
nA
We should emphasize, however, that conceptually there is no difference between
[0, ) and N and that the work we do in Sects. 2 and 3 can be done in the same
way for N. After working on M, we explicitly translate our result to N and other
metric spaces in Sect. 4.
For A M we define A : [0, ) [0, 1] by A(0):=0 and
for x > 0. Also set
which are the elements of M that have natural density and let : C [0, 1] be given
by
We write L for the collection of probability pairs (F , ) on [0, ) such that C F
M and ( A) = ( A) for A C. Our earlier observation about the indeterminacy of
probability under L gets the following formulation in terms of L: a pair (F , ) L
is canonical with respect to L if and only if F = C. We write W T for the collection
of probability pairs that are a weakly thinnable pair (WTP), that is, a probability pair
that satisfies the condition of weak thinnability. The collection W T is a proper subset
of L and contains pairs (F , ) canonical with respect to W T such that F \ C = . In
other words, with restricting L to W T we are able to assign a uniquely determined
probability to some sets without natural density. Finally, we write W T C W T L
for the elements (F , ) W T such that is coherent.
The structure of this paper is as follows. In Sect. 2, we discuss weak thinnability
and motivate why this is a natural notion of uniformity. In Sect. 3, we introduce the
probability pair (Auni, ) where
A M : L
dy L = 0
Remark 1.5 The expression in (1.9) is sometimes called the logarithmic density of A
[11, p. 272].
We end Sect. 3 with the following theorem, which is the main result of our paper.
Theorem 1.6 (Main theorem) The following holds:
1 The pair (Auni, ) is a WTP, is extendable to a WTP (F , ) with F = M and
2 The pair (Auni, ) is canonical with respect to both W T and W T C .
3 If a pair (F , ) is canonical with respect to W T or W T C , then F Auni.
In Sect. 4, we derive from (Auni, ) analogous probability pairs on certain metric
spaces including Euclidean space. The proofs of the results in Sects. 24 are given in
Sect. 5.
We write N0:={0, 1, 2, . . .}. For real-valued sequences x , y or real-valued functions
x , y on [0, ) we write x y or xi yi if limi(xi yi ) = 0. Since we work
only on [0, ) in Sects. 2 and 3, every time we speak of an f -system, probability pair
or probability measure it is understood that this is on [0, ).
2 Weak Thinnability
Let m be the Lebesgue measure on R. For Lebesgue measurable Y R with 0 <
m(Y ) < the uniform probability measure on Y is given by
for all Lebesgue measurable X Y . Let A B C be all Lebesgue measurable
with m(B) > 0 and m(C ) < . Observe that
M:= { A M : m( A) = } .
Consider for A M the map f A : A [0, ) given by f A(x ):=SA(x ). The map
f A gives a one-to-one correspondence between A and [0, ). If A M and B M,
we want to introduce notation for the set
{ f A1(b) : b B},
that gives the subset of A that corresponds to B under f A. Inspired by van Douwen
[12], we introduce the following operation.
Definition 2.1 For A, B M, define
A B:={x [0, ) : x A SA(x ) B}.
A B = { f A1(b) : b B}.
We can view this operation as thinning A by B because we create a subset of A, where
B is deciding which parts of A are removed. We also can view the operation A B
as thinning out B over A, since we spread out the set B over A. Taking for example
B =
Let (F , ) be a probability pair and let A F M. If B M, the set A B is
the subset of A corresponding to B. We can use this to transform into a measure on
A as follows. We set FA:={ A B : B F } and then define A : FA [0, 1] by
is a natural generalization of (2.2). Using (2.12) this translates into
3 The Pair (Auni, )
For A M set A : (0, )2 [0, 1] given by
1A(y)dy,
We now have the restriction that A F M. However, if A F \ M, then any
uniform probability measure should assign 0 to A and since A B A (2.14) still
holds. In Sect. 6.2, we show that the condition that (2.14) holds for all A, B F is
so strong that only probability pairs with relatively small f -systems satisfy it. Since
it is our goal to find a notion of uniformity that allows for a canonical pair with a
large f -system, we choose to use a weakened version of this property which asks that
(C A) = (C )( A) for every C C and A F .
Weak thinnability also involves two technical conditions. Let (F , ) be a probability
pair, let A, B F and suppose it is true for every x [0, ) that
SA(x ) SB (x ).
Since this inequality is true for every x , the set B is sparser than A. Therefore, it is
natural to ask that ( A) (B). We call this property preserving ordering by S.
Since we have C F , it seems natural to also ask C = , but it turns out to
be sufficient to ask the weaker property that ([c, )) = 1 for every c [0, ).
So, to reduce redundancy we require the latter and then prove that C = . Putting
everything together, we obtain the following definition.
Definition 2.2 A probability pair (F , ) with F M is a WTP if it satisfies the
following conditions:
P1 For every C C and A F we have C A F and (C A) = (C )( A),
P2 preserves ordering by S,
P3 ([c, )) = 1 for every c [0, ).
That every WTP extends natural density is implied by the following result.
lim inf A(x ) ( A) lim sup A(x ).
x x
Definition 3.1 We define
U ( A):= lim sup sup A(D, x )
D x(0,)
L( A):= liDminf xi(n0,f) A(D, x ).
uni:={ A M : L( A) = U ( A)}.
log( A):={log(a) : a A [1, )}.
Auni:={ A M : log( A) W
Notice that Definition 3.1 gives a definition of (Auni, ) that is slightly different
from (1.8) and (1.9). For a justification of Eqs. 1.8 and 1.9, see the proof of Lemma
5.2. Our first concern is that coincides with natural density.
A typical example of a set in Auni that is not in C, is
n=0
n=0
A =
log( A) =
[2n, 2n + 1),
It is easy to check that A C, but
so log( A) Wuni with (log( A)) = 1/2. Hence A Auni with ( A) = 1/2.
That (Auni, ) is a probability pair follows directly from the fact that (Wuni,
is a probability pair. The pair (Auni, ) is also a WTP.
Remark 3.4 We use free ultrafilters in the proof of Theorem 3.3 to show there exists an
extension to a W T P ith M as f -system. The existence of free ultrafilters is guaranteed
by the Boolean Prime Ideal Theorem, which cannot be proven in ZF set theory, but
is weaker than the axiom of choice [4]. The existence of a atomfree or nonprincipal
(i.e., every singleton has measure zero) finite additive measure defined on the power
set of N cannot be established in ZF alone [10]. Consequently, a version of the axiom
of choice is always necessary to construct a probability measure on M that assigns
measure zero to all bounded intervals.
Theorem 3.6 If (F , ) W T is canonical with respect to W T , then F Auni. If
(F , ) W T C is canonical with respect to W T C , then F Auni.
4 Generalization to Metric Spaces
In this section we derive probability pairs on a class of metric spaces that are analogous
to (Auni, ). Of course one could also try to construct such a probability measure by
working more directly on these metric spaces, instead of constructing a derivative of
(Auni, ). Since probability pairs on [0, ), motivated from the problem of a uniform
probability measure on N, is the priority of this paper, we do not make such an effort
here.
Let us first sketch the idea of the generalization. Let A M. Whether A is in Auni
depends completely on the asymptotic behavior of A (Lemma 5.2). If A Auni, then
also ( A) only depends on the asymptotic behavior of A (Lemma 5.2). Now suppose
that on a space X , we can somehow define a density functions B : [0, ) [0, 1]
for (some) subsets B X in a canonical way. Then, by replacing by , we get the
analogue of (Auni, ) in X . The goal of this section is to make this idea precise.
Let (X, d) be a metric space. For x X and r 0, write
B(x , r ):={y X : d(y, x ) < r }.
Write B(X ) for the Borel -algebra of X . We need a uniform measure on this space
to measure density of subsets in open balls. It is clear that the measure of an open
ball should at least be independent of where in the space we look, i.e., it should only
depend on the radius of the ball. This leads to the following definition.
Definition 4.1 We say that a Borel measure on X is uniform if for all r > 0 and
x , y X we have
0 < (B(x , r )) = (B(y, r )) < .
On Rn with Euclidean metric, the standard Borel measure as obtained by assigning
to a product of intervals the product of the lengths of those intervals, is a uniform
measure. In general, on normed locally compact vector spaces, the invariant measure
with respect to vector addition, as given by the Haar measure, is a uniform measure.
A result by Christensen [1] tells us that uniform measures that are Radon measures
are unique up to multiplicative constants on locally compact metric spaces. This,
however, does not cover all cases. The set of irrational numbers, for example, is
not locally compact, but the Lebesgue measure restricted to Borel sets of irrational
numbers is a uniform measure and unique up to a multiplicative constant. We give a
slightly more general version of the result of Christensen.
Proposition 4.2 If 1 and 2 are two uniform measures on X , then there exists some
c > 0 such that 1 = c2.
Proposition 4.2 gives us uniqueness, but not existence. To see that there are metric
spaces without a uniform measure, consider the following example. Let X be the set of
vertices in a connected graph that is not regular. Let d be the graph distance on X . If we
suppose that is a uniform measure on X , from (4.2) with r < 1 it follows that for some
C > 0 we have ({x }) = C for every x X . But then (B(x , 2)) = C (1 + deg(x ))
for every x V , which implies (4.2) cannot hold for r = 2 since the graph is not
regular. A characterization of metric spaces on which a uniform measure exist, does
not seem to be present in the literature.
We now assume X has a uniform measure and that (X ) = . In addition to
that, we write h(r ):=(B(x , r )) for r 0 and assume that
= 1,
r (u):= sup {r [0, ) : h(r ) u} ,
r +(u):=r + 1
for u [0, ). Note that h(r (u)) u and h(r +(u)) u. Write (X, L(X ), ) for
the (Lebesgue) completion of (X, B(X ), ). Fix some o X . For A L(X ) define
the map A : [0, ) [0, ) given by A(0):=0 and
for r > 0. The value A(u) is the density of A in the biggest open ball around o of
at most measure u. Notice that A is independent of the choice of as a result of
Proposition 4.2. The function A does depend on the choice of o, but in Proposition
4.3 we show that the asymptotic behavior of A does not depend on the choice of o.
We also show in Proposition 4.3 that the asymptotic behavior of A is not affected if
we replace r (u) by r +(u) in (4.5).
(B(x , r (u)) A)
h(r (u))
Remark 4.4 Proposition 4.3 is not necessarily true if we do not assume (4.3), as
illustrated by the following example. Suppose X is the set of vertices of a 3-regular tree
graph and d is the graph distance. Let be the counting measure, which is a uniform
measure on this metric space. Then clearly (4.3) is not satisfied. Now pick any x X
and let y be a neighbor of x . Let A P(X ) be the connected component containing
y in the graph where the edge between x and y is removed. Then
r
r
Then we set
Auni(X ):=
A L(X ) : lim sup sup A(D, x ) = liDminf xin>f1 A(D, x )
D x>1
Proposition 4.5 Suppose X = Rn and d is Euclidean distance. Let be the surface
measure on the unit sphere in Rn. Then for A L(Rn) we can replace A(D, x ) in
(4.9) and (4.10) by
where K A : [0, ) [0, 1] is given by
Sn1
5 Proofs
First we show that every f -system of a WTP is closed under translation and that every
probability measure of a WTP is invariant under translation.
A :={c + a : a A} F
Proof Let (F , ) be a WTP. Let A F and c [0, ). Set B:=[c, ). We have
B C F and by P3 we have (B) = 1. Therefore, A = B A F and
Proof of Propositon 2.3 Let (F , ) be a WTP and A F . Set u:= lim supx A(x ).
If u = 1 there is nothing to prove, so assume u < 1. Let > 0 be given. Let
u [0, 1] Q such that u > u and u u < . The idea is to construct a Y M
such that we can easily see that (Y ) = u and A(x ) Y (x ) for all x , so that with
P2 we get ( A) u .
First we observe that there is a K > 0 such that for all x K we have A(x ) u .
We can write u as u = qp for some p, q N0 with p q. Now we introduce the set
Y given by
Y :=[0, K )
i=0
Note that Y C F . Lemma 5.1 and the fact that is a probability measure, gives
us that (Y ) = u . Further, observe that for each x [0, ) we have A(x ) Y (x ),
so with P2 we get
By applying this to Ac we find
( A) (Y ) = u < u + .
( A) u = lim sup A(x ).
x
( A) = 1 ( Ac) 1 lim sup Ac (x ) = lixminf A(x ).
x
Before we prove Proposition 3.2 and Theorem 3.3, we present the following
alternative representation of (Auni, ). We define for A M the map A : (1, )2 [0, 1]
given by
P:= (s, f ) : s S, f (1, )
for the collection of all such sequences.
For every (s, f ) P we set
Lemma 5.2 (Alternate representation) We have
s, f :={ A M : lim A(sn, fn) exists}
n
s, f ( A):= nlim A(sn, fn).
(s, f )P
with for any (s, f ) P and A Auni
1
log(A)(log(D), log(x )) = log(D) log(x)
u=x
This implies that for (s, f ) P we have
A M : lim log(A)(log(sn), log( fn)) exists
n
with for A As, f
Since for any A M and (s, f ) P
s, f ( A) = nlim log(A)(log(sn), log( fn)).
L(log( A)) linminf log(A)(log(sn), log( fn))
lim sup log(A)(log(sn), log( fn)) U (log( A)),
n
for every n N. Then (s, f ) P with
1
log(A)(log(sn), log(x )) log(A)(log(sn), log( fn)) < n
In the same way choose (s , f ) P such that
(s, f )P
A
So assume A (s, f )P As, f . Suppose we have (s, f ) P such that s, f ( A) =
L(log( A)) and (s , f ) P such that s , f ( A) = U (log( A)). Then we can create a
new sequence given by
Choose f (1, )N such that
Proof of Proposition 3.2 Let A C and (s, f ) P. Since A(y) ( A), we have
A(sn, fn ) ( A), so s, f ( A) = ( A). The result now follows by Lemma 5.2.
Proof of Theorem 3.3 Notice that any intersection of f -systems closed under weak
thinning is again closed under weak thinning. Therefore, if we show that (As, f , s, f )
is a WTP for every (s, f ) P, it follows from Lemma 5.2 that (Auni, ) is a WTP.
Let (s, f ) P. It immediately follows that (As, f , s, f ) is a probability pair and
that P2 and P3 hold, so we have to verify P1. Note that for every A, B M and x > 0
we have
Let A C and B As, f . Then
1A(y)1B (SA(y))dy
1B (u)du
= SAx(x ) SA1(x ) 0 SA(x) 1B (u)du
= A(x )B (SA(x )) = A(x )B (xA(x )).
If ( A) = 0 it is clear that A B As, f with s, f ( A B) = 0 = ( A)s, f (B). If
( A) > 0, then we see that
dy =
du
du
= 2 log
Thus A B As, f and since ( A) = s, f ( A) (see the proof of Propositon 3.2), we
have
s, f ( A B) = ( A)s, f (B) = s, f ( A)s, f (B).
( A):=U - nlim A(sn, fn).
Since the U -limit is multiplicative it follows completely analogous that (M, ) is a
WTP. Hence every (As, f , s, f ) can be extended to a WTP with M as its f -system.
In particular, by Lemma 5.2, this means that (Auni, ) can be extended to a WTP with
M as its f -system.
From de Finetti [2] it follows that if can be extended to a finitely additive
probability measure on an algebra, then is coherent. Since we have showed that can
be extended to M, which is an algebra, it follows that (Auni, ) W T C . Notice that
we showed that s, f can be extended to M for every (s, f ) P, so we also have
(As, f , s, f ) W T C for every (s, f ) P.
P(x ):=
and Q(D, x ):=
1A(y)dy.
Also set for C > 1 and A M
Lemma 5.3 For every A M we have
= C j1(C 1) C j1
U (C, A):= lim sup sup
n kN n
1 k+n1
j=k
Proof Let A M and fix C > 1.
Step 1 We show that
for D, x (0, ).
U (log( A)) = lim sup sup
D x(0,)
= lim sup sup
D x(0,) D
= lim sup sup
D x(0,) D j=P(x)+1 C j1
C Q(D,x) 1A(u)
C j1
in terms of A(C, j ).
If we set for j N
We now observe that
Step 2 We give an upper and lower bound for
C j1
du = log
C j1
1A(y)dy = A(C, j )(C 1)C j1,
du
C j1
du = log
C j ( j )
= log(C ) log (1 + (C 1)(1 A(C, j ))) .
The fact that log(1 + y) y for every y 0, combined with (5.40), (5.41) and (5.42)
gives
Step 3 We combine Step 1 and Step 2 to finish the proof.
Observe that
1 k+n1
j=k
We use (5.43) and (5.44) to find an upper bound for the expression in (5.37), giving
lim sup sup
D x(0,)
C 1 lim sup sup 1
= log(C ) D x(0,) Q(D, x ) P(x ) j=P(x)+1
Analogously, we find that
Combining (5.45) and (5.46) we obtain
which implies
We also need the following lemma.
The idea is to introduce a set B M for which we have lim supx B (x) C S
and A(x) B (x) for all x. Set
j=1
B:=
[C j1, C j1 + SC j1(C 1)).
SC j1(C 1)
S(Cn+1 1)
CCCn (1 + S(C 1))
By Proposition 2.3 we then find
( A) lim sup A(x) C S.
x
We are ready to give the proof of Theorem 3.5.
We give the following example to give an idea of the proof that follows. Set
Z1:=
Z2:=
Z3:=
i=1
[2i, 2i + 1) = [2, 3) [4, 5) [6, 7) ,
Observe that for j 3
A :=Z1 A + Z2 A + Z3 A.
1
A (2, j ) = 2 (A(2, j 1) + A(2, j 2)) .
So we constructed a set A that on each interval [2 j1, 2 j ) with j 3 has an average
that equals the average of the averages of A on two consecutive intervals. By weak
thinnability we find that ( A ) = 21 ( A) + 41 ( A) + 41 ( A) = ( A). If A (2, j ) is
convergent or only oscillates a little, we can give a good upper bound of ( A) using
Lemma 5.4. Applying this strategy not only for C = 2 but for any C > 1 and averages
of not only two but arbitrarily many averages on consecutive intervals, is what happens
in the proof.
Step 1 We construct a A F .
Fix C > 1 and n N. We split up [C j1, C j ) into intervals of length 1 plus a
remainder interval for every j . Set for j N
and for j N and l {1, . . . , N j }
so that for every j N we have
Nu+ j (Nn+ j + 1)
[C j1, C j ) =
C j1 + N j , C j
p=0
l=1
which can be done since N j is asymptotically equivalent with C j1(C 1). For
p {0, .., n}, k {1, .., C p } and j N we set
For l T set
I p,k ( j ):=
i=0
p1
i=0
l 1
i=0
that evenly distributes mass l over the interval [0, T ). Note that (5.60) guarantees
that
for every j N, so
j=1
Z ( p, k):=
I p,k (u + j ) C n p+ j1(C 1), m(I p,k (u + j ))
From this it directly follows that
We now introduce
p=0 k=1
m([Cu+ j1, Cu+ j ) A) =
C p m([C j+np1, C j+np) A). (5.68)
p=0
p=0
C p CpA(C, j + n p)
Cnu
We apply Lemma 5.4 for A and find with (5.69) that
Combining (5.70) and (5.71) gives
C un
C
1 sup
n + 1 11/C kN n + 1 j=k
Step 3 We take limits in (5.72).
Unfix n and C . We first take the limit superior for n in (5.72), giving
( A) C lim sup sup
n kN n + 1 j=k
Then we take the limit superior for C 1 and find by Lemma 5.3 that
( A) lim sup U (C, A) = U (log( A)).
C1
The lower bound we can now easily obtain by applying our upper bound for the
complement of A. Doing this, we see that
U (log( Ac))
= 1 L(log( A)),
Proof of Theorem 3.6 We prove the contrapositive. Let (F , ) be a WTP with F \
Auni = . Let A F \ Auni. By Lemma 5.2, this means that there is a (s, f ) P
such that
I := linminf A(sn, fn ) = lim sup A(sn, fn ) =: S.
n
s , f ( A) = I and s , f ( A) = S.
In the proof of Theorem 3.3 we showed that (As , f , s , f ) and (As , f , s , f ) are
both in WTC. Thus (F , ) is not canonical with respect to W T and in case is
coherent, (F , ) is not canonical with respect to W T C .
Proof of Proposition 4.2 We give a proof along the lines of Mattila [6, p. 45], with
small adaptations for completeness and more generality.
Let (X, d) be a metric space and 1, 2 uniform measures on X . Write
h1(r ):=1(B(x, r )) and h2(r ):=2(B(x, r )) for r > 0, which are well defined since
1 and 2 are uniform. We show that 1 = c2 for some c > 0. It is sufficient to show
that 1 = c2 on all open sets.
First let A be an open set of (X, d) with 1( A) < and 2( A) < . Suppose that
r > 0 is such that h2 is continuous in r . Then
|2( A B(x, r )) 2( A B(y, r ))| 2(B(x, r ) B(y, r ))
2(B(x, r + d(x, y)) \ B(x, r ))
= h2(r + d(x, y)) h2(r ).
Hence x 2( A B(x, r )) is a continuous mapping from X to [0, ). Since h2 is
nondecreasing, it can have at most countable many discontinuities. So we can choose
r1, r2, r3, . . . such that limn rn = 0 and h2 is continuous in every rn.
For n N let fn : X [0, 1] be given by
Notice that by our previous observation fn is continuous on A, hence fn is measurable.
Because A is open, we have limn fn(x) = 1 for every x A. With Fatous Lemma
we find
1
linminf h2(rn) A 2( A B(x, rn))1(dx)
1
linminf h2(rn) X A 1B(x,rn)(y)2(dy)1(dx).
1
= linminf h2(rn) A 1(B(y, rn))2(dy)
= linminf hh21((rrnn)) 2( A).
2( A) linminf hh21((rrnn)) 1( A).
Note that lim infn hh21((rrnn)) > 0 since (5.80) would otherwise imply that all open
balls are null sets. So we may rewrite (5.80) as
Hence v1( A) = cv2( A) with
v1( A)
linminf hh21((rrnn)) 2( A).
h1(rn)
c := linminf h2(rn) > 0.
1( A) = nlim 1( An) = nlim c2( An) = c2( A).
= rlim
= 1.
Observe that for any r [0, ) we have
(B(x, r (u)) A)
h(r (u))
By (4.3), it follows that
(B(x , r (u)) A)
h(r (u))
Combining (5.85) and (5.87) gives the desired result.
Proof of Proposition 4.5 Suppose X = Rn with d Euclidean distance. Set
!
Let be the Borel measure on Rn . Note that h(r ) = n1nr n. If we set u = n nn1 y,
then
dy =
r n1 K A(r )dr dy
r n1 K A(r )dr du.
Now observe that by partial integration
r n1 K A(r )dr du = un
n
r n1 K A(r )dr + n
If we set for D, x (1, )
r n1 K A(r )dr
u=x
1
A(Dn, n1n x n ) = A(D, x ) + log(D) x
Since |A(D, x )| log1(D) , the desired result follows.
6 Discussion
6.1 Algebra Versus f -System
The natural analogue of an -algebra in finite additive probability theory is an algebra.
It has been remarked [9,13] that the restriction of M to C is problematic since C is not
an algebra. However, any collection extending C that is not M itself, is not an algebra
since a(C) = M. This can be seen as follows. Let A M and set
A+:={a + 1 : a A}
A:={a 1 : a A \ [0, 1)}
M2:=M1c,
X :=( A M1) ( Ac+ M2),
Y :=( A M2) ( Ac M1).
Then M1, M2, X, Y C with (M1) = (M2) = (X ) = (Y ) = 1/2 and A =
(M1 X ) (M2 Y ). Hence A a(C) and since A M was arbitrary, we have
a(C) = M.
This observation brings us to the conclusion that the requirement of an algebra,
despite the fact that an algebra is the natural analogue of an -algebra, is too
restrictive. Furthermore, finite additivity only dictates how a probability measure behaves
when taking disjoint unions, and thus only suggests closedness under disjoint unions.
Coherence is a concern since, as remarked before, it is not guaranteed on f -systems,
whereas it is guaranteed on algebras. Coherence, however, can also be achieved on
f -systems, as does, and therefore, coherence not being guaranteed is in itself not
an argument against f -systems. Therefore, we think the requirement of an f -system
rather than an algebra in Definition 1.1 is justified.
It should be noted that even if one prefers M as domain, by Theorem 3.3 (Auni, )
can be extended to a WTP with M as f -system. Such a pair is not canonical with
respect to W T or W T C (Theorem 3.6), but still has Auni included as an f -system
within the domain on which probability is uniquely determined.
6.2 Thinnability
Suppose that in Definition 2.2 we replace P1 by the property that for every A, B F
we have A B F and ( A B) = ( A)(B). Instead of weak thinnability, we call
this thinnability. Now consider the set
n=0
A =
We have A, Ac Auni with ( A) = ( Ac) = 1/2. But also, we have A Ac Auni
with
"2n log(2) + log(1 + 1/6), 2n log(2) + log(1 + 1/3))
So (Auni, ) is not a thinnable pair. Since every thinnable pair is also a WTP, by
Theorem 3.5 we see that a thinnable probability measure on Auni, does not exist.
Notice that we are not necessarily looking for the strongest notion of uniformity,
but for a notion that allows for a canonical probability pair with a big f -system.
This is the reason why we are interested in weak thinnability rather than thinnability.
There may, of course, be other notions of uniformity that lead to canonical pairs with
bigger f -systems than Auni. At this point, we cannot see any convincing motivation
for such notions.
In this paper, we only studied the notion of weak thinnability from the interest in
canonical probability pairs. There are, however, interesting open questions about the
property of weak thinnability itself that we did not address in this paper. Some examples
are:
6.4 Size of Auni
A typical example of a set in M that does not have natural density, but is assigned a
probability by , is
than M. If we could construct a uniform probability measure on M by the method of
Sect. 4, we could determine the probability of Auni if Auni Auni(M). To construct
such a probability measure, we need to equip M with a metric d such that (M, d) has
a uniform measure. It is, however, not at all clear how we should choose d. So at this
point, it is not clear if there is a useful way of measuring the collections C and Auni.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution,
and reproduction in any medium, provided you give appropriate credit to the original author(s) and the
source, provide a link to the Creative Commons license, and indicate if changes were made.
1. Christensen , J.P.R. : On some measures analogous to Haar measure . Math. Scand. 26 , 103 - 106 ( 1970 )
2. de Finetti, B. : Sul significato soggettivo della probabilita . Fundam. Math. 17 , 298 - 329 ( 1931 )
3. de Finetti, B. : Theory of Probability, vols. 1 & 2 ( A. Machi & A. Smith , Trans.). Wiley, New York ( 1974 )
4. Halpern , J.D. : The independence of the axiom of choice from the Boolean prime ideal theorem . Fundam. Math. 55 , 57 - 66 ( 1964 )
5. Kolmogorov , A.N. : Grundbegriffe der Wahrscheinlichkeitsrechnung . Springer, Berlin ( 1933 )
6. Mattila , P. : Geometry of Sets and Measures in Euclidean Spaces: Fractals and Rectifiability . Cambridge University Press, Cambridge ( 1995 )
7. Pier , J.-P.: Amenable Locally Compact Groups. Wiley, New York ( 1984 )
8. Schirokauer , O. , Kadane , J. : Uniform distributions on the natural numbers . J. Theor. Probab . 20 , 429 - 441 ( 2007 )
9. Schurz , G. , Leitgeb , H. : Finitistic and frequentistic approximation of probability measures with or without -additivity . Stud. Log. 89 ( 2 ), 257 - 283 ( 2008 )
10. Solovay , R.M.: A model of set theory in which every set of reals is Lebesgue measurable . Ann. Math. 92 , 1 - 56 ( 1970 )
11. Tenenbaum , G. : Introduction to Analytic and Probabilistic Number Theory . Cambridge University Press , Cambridge ( 1995 )
12. van Douwen , E.K. : Finitely additive measures on N. Topol. Appl . 47 ( 3 ), 223 - 268 ( 1992 )
13. Wenmackers , S. , Horsten , L.: Fair infinite lotteries . Synthese 190 ( 1 ), 37 - 61 ( 2013 )