Purely finitely additive measures are non-constructible
objects∗
Luc Lauwers
Center for Economic Studies, K.U.Leuven
Naamsestraat 69, B-3000 Leuven, BELGIUM
Luc.Lauwers@econ.kuleuven.be
April 13, 2010
Abstract. The existence of a purely finitely additive measure cannot be proved
in Zermelo-Frankel set theory if the use of the Axiom of Choice is disallowed.
Key words. Finitely additive probabilities; Charges; Axiom of choice; Constructivism.
1
Introduction
A finitely additive measure µ on N assigns to each subset of N a nonnegative real number
and assigns to the union of disjoint sets the sum of their numbers. The measure µ is said to
be countably additive if the measure of a countable union of pairwise disjoint sets is equal
to the sum of the measures of those sets. The finitely additive measure ν is dominated
by µ (and we write ν ≤ µ) is for each subset S of N, we have ν(S) ≤ µ(S). The finitely
additive measure µ is said to be purely finitely additive if the inequalities 0 ≤ ν ≤ µ with
ν countably additive imply that ν = 0. From Yosida and Hewitt (1952) and Rao (1958) we
know that each finitely additive measure uniquely decomposes as the sum of a countably
additive and a purely additive measure. Typically, a purely finitely additive measure is
obtained by means of Hahn-Banach’s theorem or by means of a free ultrafilter.1 Let us
describe the second route.
∗
This note extends the result obtained in “The uniform distributions puzzle” (2007, 2009) to arbitrary
purely finitely additive measures. I thank professors J.B. Kadane, Philippe Mongin, Stephen Portnoy,
K.P.S. Bhaskara Rao, and T. Seidenfeld for their comments on these earlier versions.
1
A free ultrafilter on the set N of natural numbers is a collection U of subsets of N such that (i) N ∈ U
and ∅ ∈
/ U, (ii) if A ⊆ B and A ∈ U, then B ∈ U, (iii) if A, B ∈ U, then A ∩ B ∈ U, and (iv) for each
A ⊆ N, either A or N − A belongs to U. The existence of a free ultrafilter follows from Zorn’s Lemma.
1
A free ultrafilter F on N defines a limit on X. Consider a sequence x in X and all of its
limit points. Each limit point is the limit of a subsequence. There is only one limit point
with a converging subsequence xi1 , xi2 , . . . , xit , . . . for which the set {i1 , i2 , . . . , it , . . .} of
indices belongs to F. Define limF (x) = limt→∞ xit . Due to the intersection property of F,
we have limF (x + y) = limF (x) + limF (y) for each x and y in X. The ultrafilter-based-limit
limF defines a finitely additive measure:
µF (S) = lim st
F
with st =
#(S ∩ {1, 2, . . . , t})
,
t
and S a subset of N. If the sequence s1 , s2 , . . . , st , . . . has only one accumulation point,
then µF (S) coincides with ‘the’ limit of this sequence and is known as the natural density
of S. For example, the set of even numbers has a natural density equal to .5; the set of
all multiples of 20 has a natural density equal to .05. Not every subset of N has a natural
density. For example, the set
S1 = {1, 10, 11, . . . , 19, 100, 101, . . . , 199, 1000, 1001, . . .}
of all natural numbers having their first digit equal to 1 has no natural density. The
measure µF (S1 ) depends upon the particular (non-constructible!) ultrafilter F and can
take any value between 1/9 and 5/9.2
Both routes to obtain purely finitely additive measures (Hahn-Banach’s theorem and a
free ultrafilter) rely upon AC and involve non-constructive methods. Obviously, one cannot
conclude from this that purely finitely additive measures are non-constructible objects. The
knowledge that non-constructive methods can be used to obtain a purely finitely additive
measure, does not answer the question whether a purely finitely additive measure can be
obtained without recurse to non-constructive methods.
This note shows that the existence of a purely finitely additive measure on N entails
the existence of a non-Ramsey set. From Mathias (1977) we know that a non-Ramsey set
is a non-constructible object.
The next section touches the notion of constructivism and recalls the Axiom of Choice
and the Axiom of Dependent choice. Section 3 states and proves the main result. The
result extends to purely finitely measures on R.
2
Constructivism
The Axiom of Choice (AC) postulates for each nonempty family D of nonempty sets the
existence of a function f such that f (S) ∈ S for each set S in the family D. The function
f is referred to as a choice function. AC does not provide an explicit way to construct
such a choice function and provoked considerable criticism in the aftermath of Zermelo’s
2
In this example, lim inf(st ) is the limit of the sequence 1/9, 11/99, 111/999, . . . and is equal to 1/9;
lim sup(st ) is the limit of the sequence 1, 11/19, 111/199, . . . and is equal to 5/9.
2
formulation in 1904.3 Among the applications of AC, we mention Zorn’s Lemma, the
theorem of Hahn-Banach, and the existence of free ultrafilters. AC also implies a number
of paradoxes such as the decomposition of a sphere into a sphere of smaller size, and
the existence of a non-measurable set of real numbers. The nonconstructive character of
AC is further revealed by Dianonescu (1975) who showed that AC implies the law of the
excluded middle.4 Constructive mathematics rejects the law of the excluded middle and
hence rejects AC. On the other hand, the Axiom of Dependent Choice (DC) is generally
accepted by constructivists (Beeson, 1988, p. 42). Let S be a nonempty set and let R be
a binary relation in S such that for each a in S there is a b in S with (a, b) ∈ R. Then,
DC postulates the existence of a sequence (a1 , a2 , . . . , an , . . .) of elements in S such that
(ak , ak+1 ) ∈ R for each k = 1, 2, . . ..
The nonconstructive object used in this note is known as a non-Ramsey set. Let I be
an infinite set and let n be a positive integer. Let [I]n collect all the subsets of I with
exactly n elements. Ramsey (1928) shows that for each subset S of [I]n , there exists an
infinite set J ⊂ I such that either [J]n ⊂ S or [J]n ∩ S = ∅. When n is replaced by
countable infinity, then Ramsey’s theorem fails. There exists a subset S of [I]∞ such that
for each infinite subset J of I the class [J]∞ intersects S and its complement [I]∞ − S as
well. Such a set S is said to be non-Ramsey. Mathias (1977) showed that the existence of
non-Ramsey sets does not follow from ZF (without AC).5
3
Finitely additive measures
Let N be the set of natural numbers. Let F be the field of all subsets of N. A finitely
additive probability is a map µ : F → R+ that satisfies and the condition
µ A1 ∪ A2 ∪ . . . ∪ An = µ(A1 ) + µ(A2 ) + · · · + µ(An ),
where the sets A1 , A2 , . . . , An are disjoint, for all finite n. We now state the main result of
this note.
Theorem. The existence of a finitely additive measure µ on N that attaches zero probability
to each natural number entails the existence of a non-Ramsey set.
Proof. Rescale the measure µ such that µ(N) = 1. We use some additional notation. For
two natural numbers i > j, let [ i, j[ denote the set {i, i+1, . . . , j −1}. Furthermore, to each
3
AC is (i) consistent and (ii) independent: (i) AC can be added to the Zermelo-Fraenkel axioms of set
theory (ZF) without yielding a contradiction, and (ii) AC is not a theorem of ZF (Fraenkel et al, 1973).
4
The law of the excluded middle states the truth of ‘P or not-P ’ for each proposition P and can be
used to claim the
existence of certain objects without any hint to its construction. For example, the real
√
√
√
number c = 2 2 either
√ is rational (in which case one sets a = b = 2 ) or is not rational (inb which case
one sets a = c and b = 2). Conclude the existence of irrational numbers a and b for which a is rational.
5
More precisely, Solovay (1970) proposed a model in which ZF and DC are true and in which AC fails.
Mathias showed that in this Solovay-model a non-Ramsey set does not exist. Hence, the existence of a
non-Ramsey set is independent of ZF + DC.
3
infinite set A ⊆ N we connect a set, denoted by A0 , as follows. Let A = {n0 , n1 , . . . , nk , . . .}
with nk < nk+1 for each k, then A0 = [ n0 , n1 [ ∪ [ n2 , n3 [ ∪ . . . ∪ [ n2k , n2k+1 [ ∪ . . . .
Now, let µ satisfy the requirements listed in the theorem. We show that
S = {A ⊆ N | µ(A0 ) > 0.5}
is a non-Ramsey set. It is sufficient to show that each infinite set A = {n0 , n1 , . . . , nk , . . .}
includes an infinite subset B such either A or B belongs to S (the ‘either-or’ being exclusive). We distinguish three cases.
Case 1. A ∈
/ S, in particular µ(A0 ) < 0.5. Let B = A − {n0 }. Then, [ 0, n0 [ ∪A0 ∪ B0 = N.
Since µ([ 0, n0 [) = 0, µ(A0 ) < 0.5, and µ(N) = 1; we obtain that µ(B0 ) > 0.5. Therefore,
B ⊆ A and B ∈ S.
Case 2. A ∈
/ S, in particular µ(A0 ) = 0.5. Let B = {n0 , n3 , n4 , n7 , . . . , n4k , n4k+3 , . . .} and
′
let B = {n0 , n1 , n2 , n5 , n6 , n9 , . . . , n4k+2 , n4k+5 , . . .}. Then, A0 = B0 ∩ B0′ . Hence, we have
µ(B0 ) ≥ 0.5 and µ(B0′ ) ≥ 0.5. Furthermore,
[ 0, n0 [ ∪ (B0 ∆ B0′ ) ∪ A0 = N.
Conclude that the symmetric difference B0 ∆ B0′ has a measure equal to 0.5. Hence, at
least one of the sets B0 or B0′ has a measure strictly larger than 0.5. Select the subset B
or B ′ of A for which the corresponding set B0 or B0′ has the highest measure. The selected
subset of A belongs to S.
Case 3. A ∈ S. Similar to Case 1, we put B = A − {n0 }. Conclude that B ⊆ A and that
B∈
/ S.
✷
Finally, we indicate how this result extends to purely finitely additive measures on R.
Consider such a measure µ. Since, µ is not countably additive, there exists a countable
sequence of pairwise disjoint sets A1 , A2 , . . ., An , . . . (subsets of R) such that
µ(A1 ) + µ(A2 ) + · · · + µ(An ) + · · · < µ(A1 ∪ A2 ∪ · · · ∪ An ∪ · · ·).
Define a measure µ′ on N by µ′ (C) = µ(∪j∈C Aj ). This measure has a purely finitely
additive component.
References
Beeson MJ (1988) Foundations of constructive mathematics. Ergebnisse der Mathematik
und ihrer Grenzgebiete 3.6. Berlin: Springer-Verlag.
Bhaskara Rao KPS, Bhaskara Rao M (1983) Theory of charges, a study of finitely additive
measures. London: Academic Press.
Chichilnisky G (1998) The economics of global environmental risks. International Yearbook of Environmental and Resource Economics, Vol. II eds. T Tietenberg and H
Folmer, Edward Elgar, 235-273.
4
Chichilnisky G (2000) An axiomatic approach to choice under uncertainty with catastrophic risks. Resource and Energy Economics 22(3) 221 -231.
Chichilnisky G (2009) Avoiding extinction: equal treatment of the present and the future.
Economics: The Open-Access, Open-Assessment E-Journal 3, 2009-32.
http://www.economics-ejournal.org/economics/journalarticles/2009-32.
Chichilnisky G (2009) Subjective Probability with Black Swans. Journal of Probability
and Statistics, Special Issue on Actuarial and Financial Risks (in press).
Chichilnisky G (2009) The Foundations of Statistics with Black Swans. Mathematical
Social Sciences, doi:10.1016/j.mathsocsci.2009.09.007.
Chichilnisky G (2009) The Topology of Fear. Journal of Mathematical Economics 45
Issue 11-12.
Chichilnisky G (2009) The limits of econometrics: nonparametric estimation in Hilbert
spaces. Econometric Theory 25(04), 1070-1086.
Chichilnisky G (2009) Avoiding Extinction: Equal Treatment of the Present and the
Future. Economics: The Open-Access, Open-Assessment E-Journal, 2009, Vol. 3,
2009-32.
Deal RB, Waterhouse WC, Alder RL, Konheim AG, and Blau JG (1963) Probability measure for sets of positive integers. The American Mathematical Monthly (Advanced
Problems and Solutions) 70(2), 218-219.
Diaconis P (1977) The distribution of leading digits and uniform distribution mod 1. The
Annals of Probability 5(1), 72-81.
Dubins LE, Savage LJ (1976) Inequalities for stochastic processes, how to gamble if you
must. New York: Dover Publications.
Fishburn P (1986) The axioms of subjective probability. Statistical Science 1(3), 335-358.
Fraenkel AA, Bar-Hillel Y, and Levy A (1973) Foundations of set theory. Amsterdam:
North-Holland.
Kadane JB, O’Hagan A (1995) Using finitely additive probability: uniform distributions
on the natural numbers. Journal of the American Statistical Association 90, 626-631.
Kadane JB, Schervish M, and Seidenfeld T (1986) Statistical implications of finitely additive probability. p59-76 in Bayesian Inference and Decision Techniques: Essays in
Honor of Bruno de Finetti, edited by PK Goel and A Zellner, Amsterdam: North
Holland.
Kadane JB, Schervish M, and Seidenfeld T (1999) Rethinking the foundations of statistics.
Cambridge UP, 1999.
5
Mathias ARD (1977) Happy families. Annals of Pure and Applied Logic 12, 59-111.
Ramsey FP (1928) On a problem of formal logic. Proceedings of the London Mathematical
Society, Series 2. 30.4, 338-384.
Rao RR (1958) A note on finitely additive measures. The Indian Journal of Statistics 19,
27-28.
Savage LJ (1954) The foundations of statistics. New York: Dover Publications, 2nd edn,
1972.
Seidenfeld T, Schervish M (1983) A conflict between finite additivity and avoiding Dutch
book. Philosophy of Science 50, 1983, 398-412.
Schirokauer O, Kadane JB (2007) Uniform Distributions on the Natural Numbers Journal
of Theoretical Probability (2007) 20, 429-441.
Sierpiński W (1938) Fonctions additives non complètement additives et fonctions non
mesurables. Fundamenta Mathematicae 30, 96-99.
Solovay R (1970) A model of set theory where every set of reals is Lebesgue measurable.
Annals of Mathematics 92, 1-56.
Stinchcombe MB (1997) Countably Additive Subjective Probabilities. Review of Economic Studies 64(1), 125-146.
Wakker P (1993) Savage’s axioms usually imply violation of strict stochastic dominance.
Review of Economic Studies 60, 487-493.
Yosida K, Hewitt E (1952) Finitely additive measures. Transactions of the American
Mathematical Society 72, 46-66.
6
Copyright © 2010 @ t he aut hor( s) . Discussion papers are in dr aft form . This discussion paper
is dist ribut ed for purposes of com m ent and discussion only. I t m ay not be reproduced wit hout
perm ission of t he copyright holder. Copies of working papers are available from t he aut hor.