DOI 10.1007/s0018200600292
ORI GI NAL ARTI CLE
Cores of nonatomic market games
M. Amarante · F. Maccheroni · M. Marinacci ·
L. Montrucchio
Received: 15 October 2005 / Revised: 15 May 2006 /
Published online: 10 August 2006
© SpringerVerlag 2006
Abstract We study the cores of nonatomic market games, a class of
transferable utility cooperative games introducedby AumannandShapley (Val
ues of nonatomic games, 1974), and, more in general, of those games that admit
a nacontinuous and concave extension to the set of ideal coalitions, studied by
Einy et al. (Int J Game Theory 28:1–14, 1999). We show that the core of such
games is norm compact and some related results. We also give a Multiple Priors
interpretation of some of our ﬁndings.
1 Introduction
In their studies of values of exchange economies with a continuum of players,
Aumann and Shapley (1974) introduced (nonatomic) market games, a class
of transferable utility (TU) cooperative games that includes those arising from
exchange economies. These games have been extensively studied in value the
ory (see, e.g., Neyman 2002, Sect. 12) and our purpose in this paper is to provide
a thorough study of their cores, a fundamental solution concept for TU games.
M. Amarante
Department of Economics, Columbia University,
New York, USA
F. Maccheroni (B)
Istituto di Metodi Quantitativi and IGIER, Università Bocconi, Milano, Italy
email: fabio.maccheroni@unibocconi.it
M. Marinacci · L. Montrucchio
Dipartimento di Statistica e Matematica Applicata and Collegio Carlo Alberto,
Università di Torino, Torino, Italy
400 M. Amarante et al.
Hart (1977b) showed that under weak conditions the cores of market games
arising from exchange economies are generically singleton and always ﬁnite
dimensional subsets of nonatomic measures. Our main result, Theorem 1,
complements his result by showing that the cores of all market games are norm
compact subsets of nonatomic measures. This shows that the cores of these
games have a very strong structure and they are never “large”, the ﬁnite dimen
sional form being the most important one they can take. The results of Hart
(1977b) can thus be viewed as “typical” for the cores of market games.
Theorem 1 actually holds for the larger class of games that admit a
nacontinuous and concave extension to the set of ideal coalitions, a class stud
ied by Einy et al. (1999) that contains large production games with a production
function which satisﬁes decreasing returns to scale.
We then provide, in Theorem2, a classiﬁcation of market games based on the
properties of their naextensions and on generalizations of the linear produc
tions games of Owen(1975) andBillera andRaanan(1981). Proposition3 shows
that the cores of these generalized linear production games admit a convenient
representation, and in this way we extend earlier results of Billera and Raanan
(1981), Einy et al. (1999), and Marinacci and Montrucchio (2003). Finally, in
Theorem 3 we provide a full characterization of exact market games.
Some of our results can be interpreted in the Multiple Priors model
axiomatized by Gilboa and Schmeidler (1989). In particular, we will show how
normcompact sets of priors consisting of nonatomic probability measures have
some useful properties that do not hold for general sets of priors. For example,
Multiple Priors preferences featuring such sets of priors are never Choquet
Expected Utility preferences, unless they are Subjective Expected Utility. In
other words, these Multiple Priors preferences have only a “trivial” overlap
with Schmeidler (1989)’s Choquet Expected Utility ones.
The paper is organized as follows. Section 2 introduces notation and prelim
inaries, Sect. 3 establishes the paper’s main results, whose proofs are relegated
to the Appendix, while Sect. 4 contains some examples. Finally, Sect. 5 provides
the Multiple Priors interpretation of our results, and the reader only interested
in this issue can move directly to this section, after having a look at Subsects. 2.1
and 2.2 for some notation and terminology.
2 Notation and preliminaries
2.1 Transferable utility games
Let be the set of players, the σalgebra of admissible coalitions. A (trans
ferable utility) game is a real valued function ν : →R such that ν (∅) = 0. A
game ν is:
1
1
In the sequel subsets of are understood to be in even where not stated explicitly and they
are referred to both as sets and as coalitions.
Cores of nonatomic market games 401
• positive if ν (A) ≥ 0 for all A;
• bounded if sup
A∈
[ν (A)[ < ∞;
• continuous if lim
n→∞
ν (A
n
) = lim
n→∞
_
ν () −ν
_
A
c
n
__
= 0 whenever
A
n
↓ ∅;
• additive (a charge) if ν (A∪ B) = ν (A) ÷ ν (B) for all pairwise disjoint A
and B;
• countably additive (a measure) if ν
__
∞
i=1
A
i
_
=
∞
i=1
ν (A
i
) for all count
able collections of pairwise disjoint sets {A
i
]
∞
i=1
;
• superadditive if ν (A∪ B) ≥ ν (A) ÷ν (B) for all pairwise disjoint A and B;
• convex if ν (A∪ B) ÷ν (A∩ B) ≥ ν (A) ÷ν (B) for all A and B.
The variation of a game ν : → R is the function [ν[ : → [0, ∞] deﬁned
by
[ν[ (A) = sup
n
i=1
¸
¸
ν (A
i
) −ν
_
A
i−1
_¸
¸
, ∀A ∈
where the supremum is taken over all ﬁnite chains ∅=A
0
⊆A
1
⊆ · · · ⊆A
n
=A.
A set function is of bounded variation if [ν[ () < ∞. The map ν .→ [ν[ ()
deﬁnes the variation norm on the vector space bv () all games of bounded var
iation (see Aumann and Shapley 1974, Sect. 4). The variation norm is complete.
Given a set function ν : → R, a coalition N is ν null, or simply null, if
ν (A∪ N) = ν (A) for all A in . An atom of ν is a nonnull coalition A such
that for every B ⊆ A, either B or A ¸ B is null. If ν has no atoms, ν is called
nonatomic (see Aumann and Shapley 1974, p. 14; Marinacci and Montrucchio
2004, p. 55).
The set ba () of all charges of bounded variation, its subset ca () of all mea
sures, and na () of all nonatomic measures are closed subspaces of bv (). As
well known, ba () is (isometrically isomorphic to) the norm dual of the space
B() of all bounded and measurable functions (endowed with the supnorm),
the duality being ¸g, µ) =
_
gdµ for all g in B() and µ in ba (). We will
sometimes write µ(g) instead of
_
gdµ. For all α ∈ R, the closed and convex
subset of ba () consisting of all measures taking value α at the grand coalition
is denoted by ba
α
(); ca
α
() and na
α
() are deﬁned analogously.
2.2 Exact games
The core of a game ν : →R is the set
core(ν) = {µ ∈ ba () : µ() = ν () and µ(A) ≥ ν(A) for all A ∈ ] .
The core is a weak* compact subset of ba (). Games having nonempty cores
are called balanced.
402 M. Amarante et al.
A balanced game ν is exact if
ν (A) = min
µ∈core(ν)
µ(A) , ∀A ∈ .
An exact game ν is naturally extended to B() by the function
ν
e
(g) = min
µ∈core(ν)
µ(g) , ∀g ∈ B() . (1)
Clearly, a game is exact if and only if it is the lower envelope
ν
K
(A) = inf
φ∈K
φ (A) , ∀A ∈ (2)
of a nonempty bounded subset K of ba
α
().
2
In this case, the weak* closed
and convex hull co
∗
(K) of K is contained in core (ν
K
). Notice that in general
the inclusion is strict, namely, co
∗
(K) ,= core (ν
K
) (see Examples 6, 7, and 8 of
Sect. 4).
The next proposition is basically proved inMarinacci and Montrucchio (2004,
p. 54–58).
Proposition 1 Let K be a nonempty bounded subset of ba
α
(). The game ν
K
is continuous if and only if K is a relatively weak compact subset of ca
α
(). In
this case, ν
K
is nonatomic if and only if K ⊆ na
α
().
2.3 Market and premarket games
An ideal coalition is an element of B() taking values in
_
0, 1
_
, and the set
of all ideal coalitions is denoted by B
1
(). The set B
1
() can be endowed
with the natopology due to Aumann and Shapley (1974), which is the coarsest
topology that makes continuous all the functionals g .→ µ(g) with µ ∈ na ().
3
By the Lyapunov Theorem, the characteristic functions are na dense in B
1
().
Therefore, any game ν, when viewed as a function 1
A
.→ ν (A) over the charac
teristic functions, has at most one nacontinuous extension to B
1
(). Following
Aumann and Shapley (1974), we denote this extension by ν
∗
. They show that
(Prop. 44.27) if this extension exists then core (ν) ⊆ na ().
We say that a game ν is a (nonatomic) market game if it is superadditive
and admits a positively homogeneous nacontinuous extension ν
∗
[see, e.g.,
Mertens (1980) for a similar deﬁnition]. This name is justiﬁed by the fact that,
under suitable conditions exchange economies with a continuum of agents can
2
When we consider subsets of a normed space, by bounded we mean norm bounded.
3
Of course, any subset of B() can be endowed with the natopology.
Cores of nonatomic market games 403
be modeled as market games (see Hart 1977b). For this reason, market games
play an important role in value theory.
4
Aumann and Shapley (1974, Proposition 27.1) show that the superadditivity
of a market game ν is inherited by its extension ν
∗
, so that ν
∗
is superlinear
when ν is a market game. In particular, this implies that a game ν is a market
game if and only if it has a superlinear nacontinuous extension ν
∗
.
We will also consider games having a concave nacontinuous extension ν
∗
,
and we will call them premarket games. This class, introduced by Einy et al.
(1999), contains all vector measure games of the form f ◦ µ, where µ is a ﬁnite
dimensional vector of nonatomic measures and f is a concave and continuous
function deﬁned on the range of µ.
5
Like market games, also vector measure
games play an important role in value theory (see, e.g., Hart and Neyman 1988;
Neyman 2001). Einy et al. (1999) show that a premarket game ν is balanced if
and only if ν
∗
is positively homogeneous along the diagonal
_
β1
: β ∈
_
0, 1
__
of B
1
(), and it is totally balanced if and only of it is a market game.
6
In the sequel we will make use of the following result, due to Marinacci and
Montrucchio (2005, Proposition 4), which shows that convex and premarket
games have only a trivial overlap.
7
Lemma 1 A bounded convex game is a premarket game if and only if it is a
nonatomic measure.
3 Main results
3.1 Cores
We are now ready to state our main result.
Theorem 1 The core of a premarket game is a norm compact set of nonatomic
measures.
In other words, cores of premarket and, a fortiori, of market games are norm
compact subsets of na (). Notice that, as observed, while market games are
necessarily balanced, premarket games are not. Example 1 in Sect. 4 shows
that the concavity of the nacontinuous extension is crucial for this result.
The following result is an important step in the proof of Theorem 1.
8
4
When (, ) is a standard Borel space, the class of games having nacontinuous extensions coin
cides with Aumann and Shapley’s class pNA
/
, while market games are the subclass they denote by
H
/
(see Aumann and Shapley 1974, p. 273; Hart 1977a).
5
The case of inﬁnite dimensional µ is considered by Milchtaich (1998).
6
A game is totally balanced if for every A ∈ the restriction of ν to ∩ A (i.e., the subgame
determined by A) has nonempty core.
7
See the discussion of the proof of Theorem 1 in the Appendix for a proof.
8
In the statement ∂ν
∗
(t1
) and Dν
∗
(t1
; ·) denote, respectively, the superdifferential and the
directional derivative of ν
∗
at t1
in the sense of Convex Analysis, while ν
e
is the function deﬁned
by (1).
404 M. Amarante et al.
Proposition 2 Let ν be a balanced premarket game. For all t ∈ (0, 1), core (ν) =
∂ν
∗
(t1
), ν
e
= Dν
∗
(t1
; ·), and ν
e
is continuous on the unit ball of B() in the
natopology.
Besides Theorem 1, Proposition 2 has few other interesting consequences.
Next we present two of them. The ﬁrst is a known result on the games in
pNA (see Aumann and Shapley 1974, Theorem J), reported here for sake of
completeness.
9
Corollary 1 If (, ) is a standard Borel space, then the core of any balanced
premarket game in pNAis a singleton and it coincides with the Aumann–Shapley
value.
The next sum rule is a second consequence of Proposition 2, and it shows
that cores of premarket games are stable under summation.
Corollary 2 If ν
1
and ν
2
are premarket games, then
core (ν
1
÷ν
2
) = core (ν
1
) ÷core (ν
2
) . (3)
Note that (3) implies that core (ν
1
÷ν
2
) = ∅ if core (ν
i
) = ∅ for some i. The
sum rule (3) does not hold in general, and bounded convex games are the only
other class of games for which it is known to hold (see Danilov and Koshevoy
2000; Marinacci and Montrucchio 2003, Corollary 9). However, by Lemma 1,
bounded convex games and premarket games are essentially disjoint classes of
games, and so Corollary 2 provides a new important class of games for which
the sum rule holds.
3.2 Linear production games
Using the extension properties of ν
∗
we can provide a classiﬁcation of market
games. The simplest class of market games is given by the linear production
games of Owen (1975) and Billera and Raanan (1981). They are games of the
form ϕ ◦ µ, where µ : →R
n
is a nonatomic vector measure and ϕ : R
n
→R
is given by ϕ (x) = min
t=1,...,T
a
t
· x, with each a
t
∈ R
n
.
Marinacci and Montrucchio (2003, 2005) have generalized this notion by
considering games of the form
ν (A) = min
µ∈
µ(A) , (4)
where is a ﬁnite dimensional compact subset of na (). These games are called
generalized linear production games, and the linear production games are the
special case corresponding to a ﬁnite set .
9
pNA is the closure in the variation norm of the linear space of games that is generated by all
powers of nonatomic probability measures.
Cores of nonatomic market games 405
Motivated by Theorem 1, here we further generalize these notions by
considering
ν (A) = min
µ∈
µ(A)
for some norm compact ⊆ na (). We call such a game an abstract linear
production game.
We can now state the announced classiﬁcation of market games based on the
extension properties of ν
∗
.
Theorem 2 The following properties are equivalent for a game ν:
(i) ν is an abstract (generalized, resp.) linear production game,
(ii) ν is a market game such that its naextension ν
∗
admits a further super
linear extension to the unit ball of B() (to the entire B(), resp.) that is
nalower semicontinuous at 0.
In other words, abstract linear production games are the market games whose
naextension can be extended from B
1
() to the unit ball of B(), while gen
eralized linear production games are the market games whose naextension can
be further extended to the entire space B().
In view of Theorem 2, generalized and abstract linear production games
are natural subclasses of market games. Interestingly, their cores admit a neat
representation.
Proposition 3 If ν (A) = min
µ∈
µ(A) is an abstract linear production game,
then
core (ν) = co{µ ∈ : µ() = ν ()] . (5)
Proposition 3 extends to abstract linear production games the results of
Billera and Raanan (1981) and Marinacci and Montrucchio (2003) on linear
production games and generalized linear production games, respectively.
Along with Lemma 1, Theorem 2 implies that abstract linear production
games are not convex, unless they are additive.
Corollary 3 An abstract linear production game is convex if and only if it is
additive.
Corollary 3 shows, inter alia, that neither the core nor the set of the extreme
points of the core of a nonadditive and bounded convex game ν can be a com
pact (e.g., ﬁnite) subset of na (). In fact, denote by either one of these two
sets; if is a compact subset of na (), then ν (·) = min
µ∈
µ(·) is an abstract
linear production game, something impossible by Corollary 3.
We conclude this subsection by considering exact market games, a special
class of abstract linear production games, and we fully characterize them and
their cores. Observe that the lower probabilities considered in Sect. 5 are games
of this type.
406 M. Amarante et al.
Theorem 3 The following conditions are equivalent for a game ν:
(i) ν is an exact market game;
(ii) ν is superadditive, and it admits an nacontinuous extension ν
∗
to B
1
()
such that
ν
∗
(αg ÷(1 −α) β1
) = αν
∗
(g) ÷(1 −α) βν
∗
(1
) (6)
for all α, β ∈
_
0, 1
_
and all g ∈ B
1
();
(iii) ν is exact and ν
e
is nacontinuous on B
1
();
(iv) ν is exact, continuous, nonatomic, and its core is norm compact;
(v) ν is the lower envelope of some norm relatively compact subset K of
na
α
();
(vi) ν is a uniform limit of exact linear production games.
In this case, for any K such that (v) holds and for any sequence ν
n
of exact
linear production games that uniformly converges to ν, we have:
co (K) = co
∗
(K) = core (ν) = lim
n
core (ν
n
) , (7)
where the limit is taken in the Hausdorff metric.
In particular, an exact game is a market game if and only if its core is a
norm compact set of nonatomic measures. Marinacci and Montrucchio (2005)
showed that, in this case, the core is also the unique von Neumann–Morgenstern
stable set of the game.
4 Examples
In this section we illustrate our results by means of some examples.
Example 1 Consider the bounded convex game ν (A) = λ
2
(A), where λ is the
Lebesgue measure on [0,1]. This game has the (convex) nacontinuous exten
sion ν
∗
(f ) =
__
fdλ
_
2
. Then its core is a weak compact subset of nonatomic
probability measures (see Lemma 2 in Appendix). Since ν is nonadditive, its
core cannot be norm compact (see also Example 5).
Example 2 Not all market games are abstract linear production games. Con
sider for example the vector measure game ν = ϕ (λ, µ), where ϕ (x, y) =
√
xy
for all (x, y) ∈ R
2
÷
and λ, µ are two non atomic probability measures on a stan
dard Borel space. The function ϕ is not Lipschitz and the game ν is a market
game in pNA. Its extension ν
∗
cannot be extended as a superlinear function to
the unit ball of B(). Thus ν is not an abstract linear production game. Notice
that, however, this game can be represented as the minimum over a family of
nonatomic measures. In fact,
_
λ (A) µ(A) = min
_
aλ (A) ÷bµ(A) : (a, b) ∈ R
2
÷
, ab ≥ 1
_
, ∀A ∈ .
Cores of nonatomic market games 407
The set
_
aλ ÷bµ : (a, b) ∈ R
2
÷
, ab ≥ 1
_
is not norm compact.
Example 3 Let ϕ (x, y) = −
_
x ÷y
÷
_
÷
for all (x, y) ∈ R
2
. The measure game
ν = ϕ (λ, µ), with λ, µ ∈ na (), admits a natural extension to B(). Hence, ν
is a linear production game. It is actually a glove market game in that it is easy
to see that it has the representation ν = min {0, −λ, −λ −µ].
Example 4 We provide a simple example of an abstract linear production game.
Let B be the Borel σalgebra of the unit interval I, and let λ be the Lebesgue
measure. Denote by π
n
the uniform partition of I with cardinality n, and by B
n
the ﬁnite σalgebra generated by π
n
. Fix an element f ∈ L
1
(λ) and deﬁne in
L
1
(λ) the sequence:
f
n
= E
_
f [ B
n
_
=
A∈π
n
_
A
f dλ
λ (A)
1
A
.
By the classic Martingale Convergence Theorem(see, e.g., Diestel andUhl 1977,
p. 67), f
n
→ f in the L
1
norm. Denoting N
∗
= N ∪ {∞], consider the collection
{µ
n
]
n∈N
∗ of nonatomic measures such that dµ
n
/dλ = f
n
and dµ
∞
/dλ = f . The
game ν (A) = min
n∈N
∗ µ
n
(A) is clearly an abstract linear production game. If, in
addition, f is not B
n
measurable for any n, this game is not a linear production
game. By Theorem3, ν is an exact market game and core (ν) = co{µ
n
: n ∈ N
∗
].
A slight modiﬁcation of this example delivers a nonexact abstract linear pro
duction game. It sufﬁces to deﬁne ν (A) = min
n∈N
∗ α
n
µ
n
(A), where α
n
is a
scalar sequence approaching 1.
Example 5 Let ν
(A) = min
µ∈
µ(A), where the set ⊆ na
α
() is weakly
compact but not norm compact. By Proposition 1, ν is an exact, continuous and
nonatomic game. However, in viewof Theorem3, ν is not a market game. Note
that ⊆ core (ν) and the inclusion may be strict.
By Theorem 3, any norm compact and convex subset K of na () such that
µ
1
() = µ
2
() for all µ
1
, µ
2
∈ K is the core of an exact market game. The next
examples show that this is not true for compact and convex sets not contained
in na (). More generally, these examples show that (7) may fail altogether if
the nonatomicity assumption is removed.
Example 6 Let = {ω
1
, ω
2
, ω
3
] and = 2
. Set
K =
_
µ ∈ ba () : µ is a probability measure and µ(ω
1
) ≤ µ(ω
2
)
_
.
The probability measure µsuchthat µ(ω
1
) = µ(ω
3
) = 1/2 belongs tocore (ν
K
),
but it does not belong to co(K)( = co(K) = co
∗
(K)).
Example 7 Let = {ω
1
, ω
2
, ω
3
] and = 2
. Set K = {µ, φ], where µ and
φ are probability measures such that µ(ω
1
) = µ(ω
2
) = 1/2 and φ (ω
3
) = 1,
respectively. We have:
core (ν
K
) =
_
(ξ
1
, ξ
2
, ξ
3
) ∈
_
0, 1/2
_
_
0, 1/2
_
_
0, 1
_
: ξ
1
÷ξ
2
÷ξ
3
= 1
_
,
408 M. Amarante et al.
where (ξ
1
, ξ
2
, ξ
3
) is the measure taking value ξ
i
on ω
i
. Hence, core (ν
K
) is larger
than co(K) (= co (K) = co
∗
(K) ). Since ν
K
is convex, this example, based on
Huber and Strassen (1973) and Wasserman and Kadane (1990), shows that (7)
fails even for sets of measures generating convex games.
Example 8 Let =
_
0, 1
_
and its Borel σalgebra. Set K = {δ
0
, λ], where δ
0
is the Dirac measure concentrated at 0 and λ is the Lebesgue measure on
_
0, 1
_
.
Then,
ν
K
(A) =
_
λ (A) if 0 ∈ A,
0 if 0 / ∈ A,
is a continuous and convex game. A probability measure µ belongs to core (ν
K
)
if and only if µ(A) ≥ λ (A) for all A containing 0, and
core (ν
K
) =
__
1 −
_
f dλ
_
δ
0
÷f dλ : 0 ≤ f ≤ 1 λa.e.
_
,
where f dλ is the measure taking value
_
A
f dλ on A. Clearly, core (ν
K
) is not
ﬁnite dimensional and it is therefore much larger than the ﬁnite dimensional
set co(K) (= co (K) = co
∗
(K)).
5 Multiple priors interpretation
Some of our results can be interpreted in the Multiple Priors (MP) decision
model axiomatized by Gilboa and Schmeidler (1989). Recall that in the MP
model beliefs are represented by a set C of priors, and the nonsingleton nature
of this set reﬂects the limited information on which decision makers base the
quantiﬁcation of their beliefs.
Formally, priors are probability charges and payoff prospects are elements of
B(). Gilboa and Schmeidler (1989) give necessary and sufﬁcient conditions
on a preference relation on B() that guarantee the existence of a weak*
compact set C of priors such that
f g ⇔ min
p∈C
p(f ) ≥ min
p∈C
p(g) . (8)
The lower probability
ν
C
(A) = min
p∈C
p(A) , ∀A ∈
can be viewed as the willingness to bet of the decision maker (see Ghirardato
and Marinacci 2002).
The next result extends Chateauneuf et al. (2005, Theorem 2), which gives
necessary and sufﬁcient conditions on the preference that guarantee that C is
Cores of nonatomic market games 409
a subset of na (). Here we show that a simple technical condition on ensures
that C is a norm compact (ﬁnite dimensional, resp.) subset of na ().
Proposition 4 Let be a MP preference with a weak* compact set of priors C.
The following conditions are equivalent:
(i) C is a norm compact (ﬁnite dimensional, resp.) subset of na ();
(ii) if f
d
, g
d
are naconvergent bounded nets (naconvergent nets, resp.) in
B() such that f
d
g
d
for all d, then lim
d
f
d
lim
d
g
d
.
The ﬁrst important property of this kind of MP preferences is that they are
uniquely determined by their willingness to bet:
Corollary 4 Let
1
and
2
be two MP preferences with norm compact sets of
priors C
1
and C
2
contained in na(). Then the following conditions are equiva
lent:
(i) f
1
k ⇒ f
2
k for all f ∈ B() and k ∈ R.
(ii) ν
C
1
≤ ν
C
2
.
In particular,
1
coincides with
2
if and only if ν
C
1
= ν
C
2
.
Notice that point (i) amounts to say that
1
is more ambiguity averse than
2
in the sense of Ghirardato and Marinacci (2002). Moreover, by Proposition
3 we have:
Corollary 5 If a norm compact set of priors C is contained in na (), then
core (ν
C
) = co (C) .
As a result, it is easy to derive the core of the lower probability ν
C
when C is
a norm compact set of priors consisting of nonatomic probability measures. In
particular, core (ν
C
) = C if C itself is convex, so that convex and norm compact
subsets of nonatomic probability measures are always the cores of the lower
probabilities they generate.
All these considerations are false for general sets of priors, as Examples 6, 7,
and 8 show.
By Corollary 2, lower probabilities generated by norm compact subsets of
na () form a convex set, and the map ν .→ core (ν) is afﬁne.
Corollary 6 Let ν
C
1
and ν
C
2
be lower probabilities generated by norm compact
sets of priors C
1
and C
2
contained na (). Then, for each t ∈
_
0, 1
_
, the set
function tν
C
1
÷(1 −t) ν
C
2
is the lower probability generated by tC
1
÷(1 −t) C
2
,
and
core
_
tν
C
1
÷(1 −t) ν
C
2
_
= t core
_
ν
C
1
_
÷(1 −t) core
_
ν
C
2
_
.
Corollary 3 takes the following form in the MP context.
Corollary 7 The lower probability ν
C
generated by a norm compact set C of
priors contained in na () is convex if and only if C is a singleton.
410 M. Amarante et al.
As observed after Corollary 3, this implies that if the core of a convex lower
probability consists of nonatomic measures (that is, if the lower probability is
continuous and nonatomic), then it cannot have ﬁnitely many extreme points
(unless the lower probability is a nonatomic probability measure itself).
Another important consequence of Corollaries 4 and 7 is that MP prefer
ences featuring sets of priors that are norm compact subsets of na () are never
Choquet Expected Utility preferences (see Schmeidler 1989), unless they are
Subjective Expected Utility preferences.
10
Acknowledgements We thank Achille Basile, Marco Li Calzi, two anonymous Referees, the
Associate Editor, and William Thomson (the Editor) for helpful discussions and suggestions.
The ﬁnancial support of the Ministero dell’Istruzione, dell’Università e della Ricerca is gratefully
acknowledged
A Proofs and related material
A.1 Preliminaries
If λ is a positive measure, we denote by I
∞
(λ) the positive unit ball of L
∞
(λ),
i.e., I
∞
(λ) = {g ∈ L
∞
(λ) : 0 g 1 λa.e.], and by ca (, λ) the set of all
λabsolutelycontinuous measures, i.e., ca (, λ) = {µ ∈ ca () : µ _ λ]. Clearly,
g ∈ I
∞
(λ) iff there is f ∈ B
1
() such that g (ω) = f (ω) for λalmost every ω
in . The set ca (, λ) is a closed subspace of ca() (of na () if λ is nonatomic)
and it is isometrically isomorphic to L
1
(λ). We will use repeatedly the fact that
L
∞
(λ) is (isometrically isomorphic to) the norm dual of L
1
(λ).
Given a subset F of L
∞
(λ), the relative σ
_
L
∞
(λ) , L
1
(λ)
_
topology (or
σ (L
∞
(λ) , ca (, λ))) has as neighborhood base at f ∈ F the sets V
f
(ε; µ
1
, . . . ,
µ
n
) of the form:
V
f
(ε; µ
1
, . . . , µ
n
) = {g ∈ F : [¸g, µ
i
) −¸f , µ
i
)[ < ε ∀i = 1, . . . , n]
where each µ
i
_ λ and ε > 0. That is, the relative σ
_
L
∞
(λ) , L
1
(λ)
_
topology
is the relative weak* topology of F.
The functional ν
e
: B() →Rgiven by ( 1) is deﬁned for any balanced game
ν (on the other hand, it is an extension of ν iff ν is exact).
Lemma 2 If a balanced game ν admits a nacontinuous extension ν
∗
to B
1
(),
then core (ν) ⊆ na () and ν
∗
(f ) ≤ ν
e
(f ) for all f ∈ B
1
().
Proof By Aumann and Shapley (1974, Proposition 44.27), if µ ∈ core (ν), then
µ ∈ na () and µ is nacontinuous. For any g ∈ B
1
() there exists a net
1
E
d
→
na
g. Since µ
_
1
E
d
_
≥ ν
∗
_
1
E
d
_
for all d, passing to the na limits we obtain
µ(g) ≥ ν
∗
(g). Therefore, min
µ∈core(ν)
µ(g) ≥ ν
∗
(g) for all g ∈ B
1
(). ¬.
10
In other words, support functionals generated by norm compact subsets of na () are comono
tonic additive if and only if they are additive (see Schmeidler (1986)).
Cores of nonatomic market games 411
A.2 Norm compactness and support functionals
Let X be a Banach space and X
∗
its norm dual. The bounded weak* topol
ogy is the ﬁnest topology on X
∗
which on every ball coincides with the weak*
topology.
11
If K is a bounded nonempty subset of X, the support functional
σ
K
: X
∗
→R of K is deﬁned by
σ
K
_
x
∗
_
= sup
x∈K
_
x, x
∗
_
, ∀x
∗
∈ X
∗
.
Hormander (1954, Theorem 6) shows that K is relatively (norm) compact and
ﬁnite dimensional iff σ
K
is continuous in the weak* topology. The next lemma
shows that the counterpart of relative norm compactness alone is bounded
weak* continuity.
Lemma 3 A nonempty subset of a Banach space is norm relatively compact
iff it is bounded and its support functional is continuous in the bounded weak*
topology.
Proof Let K be a nonempty bounded subset of the Banach space X, and
B
X
∗ = {x
∗
∈ X
∗
: x
∗
 ≤ 1] be the unit ball in X
∗
. Since σ
K
is positively homo
geneous, it is continuous in the bounded weak* topology iff its restriction to
B
X
∗ is continuous in the relative weak* topology.
Let K be relatively norm compact. Given ε > 0, choose x
1
, . . . , x
n
∈ K such
that for all x ∈ K, there exists i such that x −x
i
 ≤ ε. For all x
∗
∈ B
X
∗ , consider
the relative weak* neighborhood of x
∗
given by
U
x
∗
(x
1
, . . . , x
n
; ε) =
_
y
∗
∈ B
X
∗ :
¸
¸
_
x
j
, x
∗
−y
∗
_¸
¸
< ε ∀j = 1, . . . , n
_
.
If y
∗
∈ U
x
∗ , for all x ∈ K,
_
x, x
∗
−y
∗
_
−
_
x
i
, x
∗
−y
∗
_
=
_
x −x
i
, x
∗
−y
∗
_
≤ x −x
i

_
_
x
∗
−y
∗
_
_
≤ 2ε, i.e.,
_
x, x
∗
−y
∗
_
≤
_
x
i
, x
∗
−y
∗
_
÷2ε ≤ max
j=1,...,n
_
x
j
, x
∗
−y
∗
_
÷2ε ≤ 3ε.
Whence σ
K
(x
∗
−y
∗
) = sup
x∈K
¸x, x
∗
−y
∗
) ≤ 3ε; analogous considerations
yield σ
K
(y
∗
−x
∗
) ≤ 3ε (just switch x
∗
and y
∗
). By the subadditivity of the
support function, we can conclude that
¸
¸
σ
K
_
x
∗
_
−σ
K
_
y
∗
_¸
¸
≤ max
_
σ
K
_
x
∗
−y
∗
_
, σ
K
_
y
∗
−x
∗
__
≤ 3ε,
for all y
∗
∈ U
x
∗ , and so the restriction of σ
K
to B
X
∗ is continuous in the relative
weak* topology.
11
See, e.g., Dunford and Schwartz (1954, Ch. V.5) or Megginson (1998, Ch. 2.7) for details.
412 M. Amarante et al.
Conversely, assume K is bounded and σ
K
is continuous in the bounded weak*
topology. Then, the restriction of σ
K
to 2B
X
∗ is continuous in the relative weak*
topology. For all ε > 0 there exists a relative weak* neighborhood of 0 (in 2B
X
∗ )
U
0
(x
1
, . . . , x
n
; δ) =
_
z
∗
∈ 2B
X
∗ :
¸
¸
_
x
j
, z
∗
_¸
¸
< δ ∀j = 1, . . . , n
_
suchthat σ
K
(z
∗
) ≤ ε for all z
∗
∈ U
0
. For all x
∗
∈ B
X
∗ consider the relative weak*
neighborhood U
x
∗
(x
1
, . . . , x
n
; δ). For all y
∗
∈ U
x
∗ we have x
∗
−y
∗
, y
∗
−x
∗
∈ U
0
.
Therefore, σ
K
(x
∗
−y
∗
) ≤ ε and σ
K
(y
∗
−x
∗
) ≤ ε, so that
sup
y
∗
∈U
x
∗
sup
x∈K
¸
¸
_
x, x
∗
−y
∗
_¸
¸
≤ ε.
When the elements of X are regarded as weak* continuous functions on
the weak* compact set B
X
∗ , what we proved amounts to the fact that: for
all ε >0 and all x
∗
∈B
X
∗ , there exists a neighborhood U
x
∗ of x
∗
such that
sup
x∈K
sup
y
∗
∈U
x
∗
[¸x, x
∗
) −¸x, y
∗
)[ ≤ ε. That is, K is equicontinuous. Being
bounded (sup
x
∗
∈B
X
∗
[¸x, x
∗
)[ = x ≤ sup
y∈K
y for all x ∈ K) and equicon
tinuous, K is relatively compact by the Ascoli–Arzelà Theorem in the space
C(B
X
∗ ) of all weak* continuous functions on B
X
∗ endowed with the supnorm.
Notice that
d
X
(x, y) = x −y = max
x
∗
∈B
X
∗
¸
¸
_
x, x
∗
_
−
_
y, x
∗
_¸
¸
= d
C(B
X
∗ )
(¸x, ·) , ¸y, ·)) .
Since every sequence in K admits a Cauchy subsequence in C(B
X
∗ ), it admits
a Cauchy subsequence in X. We conclude that K is norm relatively compact in
X (since X is complete). ¬.
For future reference we state the following immediate reformulation of
Lemma 3.
Lemma 4 Anonempty subset K of a Banach space X is normrelatively compact
iff it is bounded and the functional x
∗
.→ inf
x∈K
¸x, x
∗
) is continuous on B
X
∗ in
the relative weak* topology.
We close with a result on superdifferentials.
Lemma 5 Let X be a Banach space and ϕ : X →R∪{−∞] a concave function.
If ¨ x ∈ dom(ϕ) is such that ϕ (α¨ x) = αϕ (¨ x) for all α ∈
_
0, 1
_
, then
∂ϕ (α¨ x) =
_
x
∗
∈ ∂ϕ (0) : x
∗
(¨ x) = ϕ (¨ x)
_
=
_
x
∗
∈ ∂ϕ (¨ x) : x
∗
(¨ x) = ϕ (¨ x)
_
for all α ∈ (0, 1).
Proof Let α ∈ (0, 1) and x
∗
∈ ∂ϕ (α¨ x). By deﬁnition,
ϕ (y) −x
∗
(y) ≤ ϕ (α¨ x) −x
∗
(α¨ x) = α
_
ϕ (¨ x) −x
∗
(¨ x)
_
(9)
Cores of nonatomic market games 413
for all y ∈ X. For all y = β¨ x with β ∈ (0, 1)¸{α], we get (β −α)
_
ϕ (¨ x) −x
∗
(¨ x)
_
≤
0, whence ϕ (¨ x) = x
∗
(¨ x). Hence, getting back to (9), we conclude that ϕ (y) −
x
∗
(y) ≤ 0 = ϕ (¨ x) −x
∗
(¨ x). This implies both x
∗
∈ ∂ϕ (0) and x
∗
∈ ∂ϕ (¨ x). That is
∂ϕ (α¨ x) ⊆
_
x
∗
∈ ∂ϕ (0) : x
∗
(¨ x) = ϕ (¨ x)
_
∩
_
x
∗
∈ ∂ϕ (¨ x) : x
∗
(¨ x) = ϕ (¨ x)
_
.
Conversely:
• if x
∗
∈ ∂ϕ (0) and x
∗
(¨ x) = ϕ (¨ x), then for all y ∈ X
ϕ (y) −x
∗
(y) ≤ 0 = ϕ (¨ x) −x
∗
(¨ x) = ϕ (α¨ x) −x
∗
(α¨ x);
• if x
∗
∈ ∂ϕ (¨ x) and x
∗
(¨ x) = ϕ (¨ x), then for all y ∈ X
ϕ (y) −x
∗
(y) ≤ ϕ (¨ x) −x
∗
(¨ x) = 0 = ϕ (α¨ x) −x
∗
(α¨ x);
in any case x
∗
∈ ∂ϕ (α¨ x) . ¬.
Lemma 6 Let X be a Banach space, K
∗
a weak* compact subset of X
∗
, and
ϕ (x) = min
x
∗
∈K
∗
_
x, x
∗
_
, ∀x ∈ X.
Then ∂ϕ (x) = co
w
∗
{x
∗
∈ K
∗
: ¸x, x
∗
) = ϕ (x)] for each x ∈ X.
Proof Clearly,
ϕ (x) = min
x
∗
∈co
w
∗
(K
∗
)
_
x, x
∗
_
, ∀x ∈ X,
and so ∂ϕ (0) = co
w
∗
(K
∗
) as co
w
∗
(K
∗
) is convex and weak* compact. More
over, ∂ϕ (x) = {x
∗
∈ ∂ϕ (0) : ¸x, x
∗
) = ϕ (x)] for all x ∈ X . Since K
∗
is weak*
compact ext (∂ϕ (0)) = ext
_
co
w
∗
(K
∗
)
_
⊆ K
∗
(see Megginson 1998, Corollary
2.10.18).
Arbitrarily choose x ∈ X. ∂ϕ (x) is a weak* compact and convex subset of
X
∗
. Next we show that ∂ϕ (x) is extremal in ∂ϕ (0). In fact, if x
∗
, y
∗
∈ ∂ϕ (0),
t ∈ (0, 1), and tx
∗
÷(1 −t) y
∗
∈ ∂ϕ (x), then
_
x, x
∗
_
≥ ϕ(x),
_
x, y
∗
_
≥ ϕ(x), and
t
_
x, x
∗
_
÷(1 −t)
_
x, y
∗
_
=
_
x, tx
∗
÷(1 −t) y
∗
_
= ϕ(x);
therefore ¸x, x
∗
) = ¸x, y
∗
) = ϕ (x) and x
∗
, y
∗
∈ ∂ϕ (x).
Since ∂ϕ (x) is extremal in ∂ϕ (0), if u
∗
is an extreme point of ∂ϕ (x), then u
∗
is an extreme point of ∂ϕ (0) and it belongs to K
∗
. That is
ext (∂ϕ (x)) = ext (∂ϕ (0)) ∩ ∂ϕ (x) ⊆ K
∗
∩ ∂ϕ (x) =
_
x
∗
∈ K
∗
:
_
x, x
∗
_
= ϕ (x)
_
.
414 M. Amarante et al.
By the Krein–Milman Theorem
∂ϕ (x) = co
w
∗
(ext (∂ϕ (x))) ⊆ co
w
∗
_
x
∗
∈ K
∗
:
_
x, x
∗
_
= ϕ (x)
_
⊆ ∂ϕ(x). ¬.
A.3 A separation result
In the sequel we will need the following separation lemma, which is of some
independent interest.
Lemma 7 Let K
1
and K
2
be nonempty, disjoint, norm compact, and convex
subsets of na
α
(). Then, there exists A in such that
min
φ
1
∈K
1
φ
1
(A) > max
φ
2
∈K
2
φ
2
(A) . (10)
Proof Since K
i
is weakly compact and it consists of nonatomic measures, there
exists a nonatomic probability measure λ
i
such that φ
i
_ λ
i
for all φ
i
in K
i
,
i = 1, 2 (see Dunford and Schwartz 1954, Theorem IV.9.2). Therefore, all mea
sures in K
1
∪ K
2
are absolutely continuous w.r.t. λ = 2
−1
(λ
1
÷λ
2
). Therefore,
K
1
and K
2
are norm compact subsets of ca (, λ) .
The Separating Hyperplane Theoremguarantees that there exist f ∈ L
∞
(λ)¸
{0] and β ∈ R such that
min
φ
1
∈K
1
φ
1
(f ) > β > max
φ
2
∈K
2
φ
2
(f ) . (11)
Since K
1
, K
2
⊆ na
α
(), w.l.o.g. we can choose f ∈ I
∞
(λ). In other words,
setting (g) = min
φ
1
∈K
1
φ
1
(g) − max
φ
2
∈K
2
φ
2
(g), there exists f ∈ I
∞
(λ) such
that (f ) > 0. By Lemma 3, is continuous w.r.t. the relative weak* topology
on I
∞
(λ). Hence, since the set {1
A
]
A∈
is dense in I
∞
(λ) by the Lyapunov
Theorem, we conclude that (10) holds for some A ∈ . ¬.
Corollary 8 Let K
1
and K
2
be nonempty subsets of na
α
(). If K
2
is norm
relatively compact, then
inf
φ
1
∈K
1
φ
1
(A) ≥ inf
φ
2
∈K
2
φ
2
(A) ∀A ∈ ⇒ K
1
⊆ co (K
2
) .
Proof Assume, by contradiction that inf
φ
1
∈K
1
φ
1
(A) ≥ inf
φ
2
∈K
2
φ
2
(A) for all
A ∈ andthere exists φ ∈ K
1
−co (K
2
). Since co (K
2
) = co
_
K
2
_
andK
2
is norm
compact, by the Mazur Compactness Theorem (see, Megginson 1998, Theorem
2.8.15) co(K
2
) is norm compact. Direct application Lemma 7 to co (K
2
) and
{φ], yields that there exists B ∈ such that
Cores of nonatomic market games 415
inf
φ
2
∈K
2
φ
2
(B) ≥ min
µ∈co(K
2
)
µ(B) > φ (B) ≥ inf
φ
1
∈K
1
φ
1
(B) ,
which is absurd. ¬.
A.4 Proofs of the results
Proof of Theorem 1 If core (ν) = ∅ the result is trivial. Assume that core (ν) ,=
∅, and denote by ν
∗
the concave and na continuous extension of ν to B
1
().
Lemma 2 guarantees that core (ν) ⊆ na (), it remains to show that core (ν) is
norm compact.
Claim 1 core (ν) = ∂ν
∗
(t1
) for all t ∈ (0, 1), where ∂ν
∗
is the superdifferential
of ν
∗
.
The proof of this claim follows the same lines of Einy et al. (1999, Theorem
A). First we prove that ν
∗
(α1
) = αν
∗
(1
) for all α ∈
_
0, 1
_
, namely, ν
∗
is
positively homogeneous along the diagonal. If µ ∈ core (ν), then µ(f ) ≥ ν
∗
(f )
for all f ∈ B
1
() (see Lemma 2). In particular, αν () = αµ() = µ(α1
) ≥
ν
∗
(α1
). Concavity of ν
∗
implies ν
∗
(α1
) ≥ αν
∗
(1
). The claim immediately
follows. Let t ∈ (0, 1). If µ ∈ core (ν), for all f ∈ B
1
() we have µ(f ) ≥ ν
∗
(f )
and µ(t1
) = tµ(1
) = tν
∗
(1
) = ν
∗
(t1
), whence
µ(f ) −µ(t1
) ≥ ν
∗
(f ) −ν
∗
(t1
) , ∀f ∈ B
1
() , (12)
that is µ ∈ ∂ν
∗
(t1
). Conversely, if µ ∈ ∂ν
∗
(t1
), µ satisﬁes (12), setting f = 0
we obtain tµ() ≤ tν (), setting f = 1
we obtain (1 −t) µ() ≥ (1 −t) ν (),
whence
µ() = ν () and µ(f ) ≥ ν
∗
(f ) , ∀f ∈ B
1
() ,
a fortiori, µ ∈ core (ν).
Given the concave function ν
∗
: B
1
() →R, let f
0
belong to the (supnorm)
interior of B
1
(), and choose a positive scalar η such that f −f
0
 ≤ η implies
f ∈ B
1
(). Denote by Dν
∗
(f
0
; h), with h ∈ B(), the directional derivative,
i.e.,
Dν
∗
(f
0
; h) = lim
t→0
÷
ν
∗
(f
0
÷th) −ν
∗
(f
0
)
t
= sup
t>0
ν
∗
(f
0
÷th) −ν
∗
(f
0
)
t
.
Since ν
∗
is supnorm continuous, we have (see Barbu and Precupanu 1986,
Proposition 2.3) that ∂ν
∗
(f
0
) is a nonempty, convex, and weak* compact sub
set of ba (); moreover,
Dν
∗
(f
0
; h) = min
φ∈∂ν
∗
(f
0
)
φ (h) , ∀h ∈ B() . (13)
416 M. Amarante et al.
In particular, if t ∈ (0, 1), then Dν
∗
(t1
; h) = min
φ∈∂ν
∗
(t1
)
φ (h) = min
φ∈core(ν)
φ (h) = ν
e
(h).
Claim 2 If ∂ν
∗
(f
0
) ⊆ na (), the function Dν
∗
(f
0
; ·) is continuous on the unit
ball B
B()
in the natopology.
Inviewof (13), Dν
∗
(f
0
; ·) is naupper semicontinuous onB
B()
, providedthat
∂ν
∗
(f
0
) ⊆ na (). Consequently, it sufﬁces to check that the function Dν
∗
(f
0
; ·)
is nalower semicontinuous on B
B()
. This is obtained with a technique inspired
by Hart (1977a).
Fix h
0
∈ B
B()
and ε > 0. We can ﬁnd a scalar positive scalar t ≤ η such that:
ν
∗
_
f
0
÷th
0
_
−ν
∗
(f
0
)
t
≥ Dν
∗
(f
0
; h
0
) −ε/2 (14)
(the scalar η has been previously deﬁned to guarantee that f ∈ B
1
() if
f −f
0
 ≤ η, hence f
0
÷ th ∈ B
1
() for all h ∈ B
B()
and ν
∗
_
f
0
÷th
_
is
well deﬁned).
As ν
∗
is nacontinuous on B
1
(), there exists a naneighborhood
U
f
0
÷th
0
(µ
1
, . . . , µ
n
; δ) =
_
f ∈ B
1
() : max
j=1,..,n
¸
¸
µ
j
_
f −
_
f
0
÷th
0
__¸
¸
< δ
_
(15)
of f
0
÷th
0
in B
1
() such that
¸
¸
ν
∗
(f ) −ν
∗
_
f
0
÷th
0
_¸
¸
< εt/2, ∀f ∈ U
f
0
÷th
0
(µ
1
, . . . , µ
n
; δ) .
Consider now the following naneighborhood of h
0
in B
B()
:
U
h
0
_
µ
1
, . . . , µ
n
; δ/t
_
=
_
h ∈ B
B()
: max
j=1,..,n
¸
¸
µ
j
(h −h
0
)
¸
¸
< δ/t
_
.
It is immediately checked that if h ∈ U
h
0
, then f
0
÷ th ∈ U
f
0
÷th
0
. In fact,
f
0
÷th ∈ B
1
() and
max
j=1,..,n
¸
¸
µ
j
_
f
0
÷th −
_
f
0
÷th
0
__¸
¸
= t max
j=1,..,n
¸
¸
µ
j
(h −h
0
)
¸
¸
< δ.
Hence, h ∈ U
h
0
implies
¸
¸
ν
∗
_
f
0
÷th
_
−ν
∗
_
f
0
÷th
0
_¸
¸
< εt/2 and
ν
∗
_
f
0
÷th
_
> ν
∗
_
f
0
÷th
0
_
−εt/2.
In view of (14), we have
Dν
∗
(f
0
; h) ≥
ν
∗
_
f
0
÷th
_
−ν
∗
(f
0
)
t
>
ν
∗
_
f
0
÷th
0
_
−εt/2 −ν
∗
(f
0
)
t
=
ν
∗
_
f
0
÷th
0
_
−ν
∗
(f
0
)
t
−ε/2 ≥ Dν
∗
(f
0
; h
0
) −ε
Cores of nonatomic market games 417
for all h ∈ U
h
0
. This means that Dν
∗
(f
0
; ·) is nalower semicontinuous at
h
0
. Since h
0
was chosen arbitrarily, Dν
∗
(f
0
; ·) is nalower semicontinuous on
B
B()
.
Claim 3 If is a weak* compact subset of ba () and : B() → R, deﬁned
by
(f ) = min
µ∈
µ(f ) , ∀f ∈ B() ,
is nacontinuous on the unit ball B
B()
of B(), then is a norm compact
subset of na ().
Clearly is nacontinuous on every ball tB
B()
with t > 0. Then, for all
µ ∈ , µ : B
1
() → R is continuous in the natopology, and Lemma 2 implies
µ ∈ core (µ) ⊆ na (). Then ⊆ na () is weak* compact in ba (), hence
it is weak compact (Marinacci and Montrucchio 2004, p. 53), a fortiori there
exists a nonatomic probability measure λ ∈ na () such that ⊆ ca (, λ) (see
Dunford and Schwartz 1954, Thm. IV.9.2). As a consequence, setting
ρ (f ) = min
µ∈
µ(f ) , ∀f ∈ L
∞
(λ)
deﬁnes a function ρ : L
∞
(λ) → R. Next we show that under the identiﬁcation
of L
∞
(λ) with the norm dual of ca (, λ), the functional ρ is continuous on
B
L
∞
(λ)
in the relative weak* topology σ (L
∞
(λ) , ca (, λ)), that is, (being
weak compact and – by Lemma 4 – norm relatively compact) is norm compact.
Notice that (f ) = ρ (f ) for all f ∈ B(). Let f
/
∈ B
L
∞
(λ)
and ε > 0. Choose
f ∈ B
B()
such that f
/
= f λa.e. and let
U
f
(µ
1
, . . . , µ
n
; δ) =
_
g ∈ B
B()
: [µ
i
(g) −µ
i
(f )[ < δ ∀i = 1, . . . , n
_
be a naneighborhood of f such that
_
U
f
_
⊆ ( (f ) −ε, (f ) ÷ε). W.l.o.g.,
suppose each µ
i
is a probability measure.
Let µ
i
= µ
a
i
÷ µ
s
i
be the Lebesgue decomposition of each µ
i
, with µ
a
i
_ λ
and µ
s
i
⊥ λ. Set
V
f
=
_
h ∈ B
B()
:
¸
¸
µ
a
i
(h) −µ
a
i
(f )
¸
¸
<
δ
2
and
¸
¸
µ
s
i
(h) −µ
s
i
(f )
¸
¸
<
δ
2
∀i = 1, . . . , n
_
.
Clearly, V
f
⊆ U
f
and so
_
V
f
_
⊆ ( (f ) −ε, (f ) ÷ε).
Consider the weak* neighborhood of f
/
W
f
/ =
_
g
/
∈ B
L
∞
(λ)
:
¸
¸
µ
a
i
_
g
/
_
−µ
a
i
_
f
/
_¸
¸
<
δ
2
∀i = 1, . . . , n
_
.
418 M. Amarante et al.
For each g
/
∈ W
f
/ , let g ∈ B
B()
be such that g
/
= g λa.e.. Choose E ∈ such
that λ (E) = 0 and µ
s
i
(E
c
) = 0 for each i = 1, . . . , n.
12
Set h = f 1
E
÷g1
E
c . Then
h = g = g
/
λa.e. and ρ
_
g
/
_
= ρ (h) = (h). Moreover, for all i = 1, . . . , n
¸
¸
µ
a
i
(h) −µ
a
i
(f )
¸
¸
=
¸
¸
µ
a
i
_
g
/
_
−µ
a
i
_
f
/
_¸
¸
<
δ
2
,
while
¸
¸
µ
s
i
(h) −µ
s
i
(f )
¸
¸
=
¸
¸
µ
s
i
(f 1
E
) ÷µ
s
i
(g1
E
c ) −µ
s
i
(f 1
E
) −µ
s
i
(f 1
E
c )
¸
¸
= 0,
so that h ∈ V
f
. In sum, for each g
/
∈ W
f
/ there exists h ∈ V
f
such that ρ
_
g
/
_
=
(h) and
ρ
_
g
/
_
= (h) ∈
_
V
f
_
⊆ ( (f ) −ε, (f ) ÷ε) =
_
ρ
_
f
/
_
−ε, ρ
_
f
/
_
÷ε
_
.
This proves that ρ is continuous on B
L
∞
(λ)
in the relative weak* topology. ¬.
The proof is concluded by observing that: ∂ν
∗
_
2
−1
1
_
= core (ν) ⊆ na ()
(Claim 1), hence ν
e
= Dν
∗
_
2
−1
1
; ·
_
is nacontinuous on B
B()
(Claim 2), but
ν
e
= min
φ∈core(ν)
φ (·), then core (ν) is norm compact (Claim 3). ¬.
Proposition 2 immediately descends from inspection of the above Claims
1 and 2. In particular, if a premarket game ν is exact, the uniqueness of the
nacontinuous extension to B
1
() implies that ν
∗
= ν
e[B
1
()
and ν
∗
is positively
homogeneous; therefore ν is a market game.
In turn this implies that a bounded convex (hence exact) premarket game
is a market game and (Marinacci and Montrucchio 2005, Prop. 4) is equivalent
to Lemma 1.
Proof of Corollary 1 By Proposition 2, core (ν) = ∂ν
∗
(t1
) for all t ∈ (0, 1).
By (Aumann and Shapley, 1974, Prop. 24.1), for almost every t ∈ (0, 1) the
naextension ν
∗
is Frechet differentiable at t1
, provided ν ∈ pNA. Hence,
core (ν) is a singleton and ∂ν
∗
(t1
) agrees with the derivative. Moreover, by
the diagonal formula (see Aumann and Shapley 1974, Thm. H) the game value
is
(ϕν) (S) =
1
_
0
∂ν
∗
(t1
) (S) dt = µ(S)
where µ ∈ core (ν). Accordingly, ϕν = µ ∈ core (ν). ¬.
Proof of Corollary 2 By Einy et al. (1999, Thm. A) (see also Claim 1 of the
proof of Theorem 1), if ν
i
is balanced then ν
∗
i
is positively homogeneous along
12
For i = 1, . . . , n there exists A
i
such that λ (A
i
) = µ
s
i
_
A
c
i
_
= 0 (µ
s
i
⊥ λ), set E =
_
n
i=1
A
i
.
Cores of nonatomic market games 419
the diagonal, and core (ν
i
) = ∂ν
∗
i
_
2
−1
1
_
= ∂ν
∗
i
(t1
) for all t ∈ (0, 1), by
Proposition 2. Analogously, if ν
∗
i
(t1
) = tν
∗
i
(1
) for some t ∈ (0, 1), then
∅ ,= ∂ν
∗
i
(t1
) = core (ν
i
). Suppose ﬁrst that, for instance, core (ν
1
) = ∅. Then,
ν
∗
1
_
2
−1
1
_
> 2
−1
ν
1
(). This implies
(ν
1
÷ν
2
)
∗
_
2
−1
1
_
=
_
ν
∗
1
÷ν
∗
2
_
_
2
−1
1
_
= ν
∗
1
_
2
−1
1
_
÷ν
∗
2
_
2
−1
1
_
> 2
−1
(ν
1
÷ν
2
) (),
and the game ν
1
÷ν
2
has empty core. Suppose nowthat core (ν
i
) ,= ∅for i = 1, 2.
It follows that:
core (ν
1
÷ν
2
) = ∂ (ν
1
÷ν
2
)
∗
_
2
−1
1
_
= ∂
_
ν
∗
1
÷ν
∗
2
_
_
2
−1
1
_
= ∂ν
∗
1
_
2
−1
1
_
÷∂ν
∗
2
_
2
−1
1
_
= core (ν
1
) ÷core(ν
2
),
where we are using the sum rule for superdifferentials (see, e.g., Barbu and
Precupanu 1986, Ch. 3, Thm. 2.6). ¬.
Proof of Theorem 2 (i) ⇒ (ii). Assume there exists a norm compact (and ﬁnite
dimensional, resp.) subset of na () such that ν (A) = min
µ∈
µ(A) for all
A ∈ . There exists a nonatomic probability measure λ ∈ na () such that
⊆ ca (, λ) (see Dunford and Schwartz 1954, Thm. IV.9.2). Lemma 4 guaran
tees that the functional
ˆ ν : L
∞
(λ) → R
f .→ min
µ∈
µ(f )
is continuous in the relative weak* topology of B
L
∞
(λ)
, a fortiori, the exten
sion of ν to B
B()
deﬁned by ν
∗
(g) = min
µ∈
µ(g) is superlinear and σ
_
B
B()
,
ca (, λ)
_
continuous, hence nacontinuous. If, moreover, is ﬁnite dimen
sional, Hormander (1954, Thm. 6) guarantees that ˆ ν is continuous on (the
entire) L
∞
(λ) in the weak* topology.
(ii) ⇒ (i). Assume that a market game ν admits an extension ν
∗
to B
B()
which is superlinear and nalower semicontinuous at 0. ν
∗
can be extended
(by positive homogeneity) to a superlinear function on B(), which we still
denote by ν
∗
. A standard argument guarantees that ν
∗
is nacontinuous on
every bounded subset of B(). A fortiori, ν
∗
is supnorm continuous on B()
and
ν
∗
(f ) = min
µ∈
µ(f ) , ∀f ∈ B() (16)
where is a weak* compact and convex subset of ba (). Claim 3 of the proof
of Theorem 1 shows that is a norm compact subset of na (). A similar argu
ment, building on Hormander (1954, Thm. 6) rather than Lemma 4, shows that
420 M. Amarante et al.
if ν
∗
is nalower semicontinuous at 0, as a function on B() rather than on
B
B()
, then is ﬁnite dimensional. ¬.
Proof of Proposition 3 If ν (A) = min
µ∈
µ(A), with a norm compact subset
of na (), then, as observed in the proof of Theorem 2,
ν
∗
(f ) = min
µ∈
µ(f ) , ∀f ∈ B
1
() .
Consider the extension of ν
∗
to B() deﬁned by
ν
∗∗
(f ) = min
µ∈
µ(f ) , ∀f ∈ B() .
Since is weak* compact in ba (), by Lemma 6,
∂ν
∗∗
(f
0
) = co
w
∗
_
µ ∈ : µ(f
0
) = ν
∗∗
(f
0
)
_
, ∀f
0
∈ B() .
Finally,
core (ν) = ∂ν
∗
_
2
−1
1
_
= ∂ν
∗∗
_
2
−1
1
_
= co
w
∗
_
µ ∈ : µ
_
2
−1
1
_
= ν
∗∗
_
2
−1
1
__
= co
w
∗
{µ ∈ : µ() = ν ()]
= co {µ ∈ : µ() = ν ()] ,
where the last equality holds since the Mazur Compactness Theoremguarantees
norm compactness of the set co{µ ∈ : µ() = ν ()] ¬.
Proof of Theorem 3 We ﬁrst show that (i) ⇒ (ii) ⇒ (iii) ⇒ (iv) ⇒ (v) ⇒ (i)
and the ﬁrst part of (7).
(i) ⇒ (ii). Being a market game, ν admits a concave and nacontinuous
extension ν
∗
to B
1
(), exactness guarantees core (ν) ,= ∅. Proposition 2 yields
that ν
e
is continuous on the unit ball B
B()
in the natopology. A fortiori ν
e
is
continuous on B
1
(), and exactness guarantees that ν
e[B
1
()
is an extension of
ν to B
1
(). It follows that ν
e[B
1
()
= ν
∗
and (6) holds.
(ii) ⇒(iii). As ν is superadditive, by Aumann and Shapley (1974, Prop. 27.1)
ν
∗
is superadditive. While setting β = 0 in (6) guarantees positive homogeneity.
A routine exercise shows that ν
∗
admits an extension ν
∗∗
to the entire space
B() which is superlinear and such that ν
∗∗
(g ÷β1
) = ν
∗∗
(g) ÷ βν () for
all g ∈ B() and all β ∈ R. Next we show that ν
∗∗
is supnorm continuous.
In fact, if g
n
→ g in the supnorm, there exist α ∈ (0, 1) and β ∈ R such that
αg
n
÷β1
, αg÷β1
∈ B
1
() andαg
n
÷β1
→ αg÷β1
inthe supnorm. There
fore µ(αg
n
÷β) → µ(αg ÷β) for all µ ∈ na (). Since ν
∗
is na continuous on
B
1
(), then ν
∗
(αg
n
÷β) → ν
∗
(αg ÷β) and
αν
∗∗
(g
n
) ÷β = ν
∗
(αg
n
÷β) → ν
∗
(αg ÷β) = αν
∗∗
(g) ÷β.
Cores of nonatomic market games 421
Therefore,
ν
∗∗
(g) = min
µ∈∂ν
∗∗
(0)
µ(g) , ∀g ∈ B() , (17)
where ∂ν
∗∗
(0) = {µ ∈ ba () : µ(g) ≥ ν
∗∗
(g) for all g ∈ B()]. For all µ ∈
∂ν
∗∗
(0) and all A ∈ we have µ(A) ≥ ν (A); moreover, −µ() = µ(−1
) ≥
ν
∗∗
(−1
) = −ν (), whence µ() = ν () and µ ∈ core (ν). Therefore,
∂ν
∗∗
(0) ⊆ core (ν) and for all g ∈ B
1
()
ν
∗
(g) = min
µ∈∂ν
∗∗
(0)
µ(g) ≥ min
µ∈core(ν)
µ(g) = ν
e
(g) ≥ ν
∗
(g) .
The last inequality follows from Lemma 2. This proves that ν
e
= ν
∗
is nacon
tinuous on B
1
() and that ν is exact.
(iii) ⇒ (iv). Exactness guarantees that ν
e[B
1
()
is a concave extension of
ν to B
1
(). Then nacontinuity of ν
e[B
1
()
= ν
∗
, together with Theorem 1,
guarantees that core (ν) is a norm compact subset of na (). Continuity and
nonatomicity of ν follow from Proposition 1.
(iv) ⇒ (v). Continuity and nonatomicity of ν, together with Proposition 1,
guarantee core (ν) ⊆ na
ν()
(). Just set K = core (ν).
(v) ⇒ (i) and the ﬁrst part of (7). Assume ν is the lower envelope of a
norm relatively compact subset K of na
α
(). Obviously ν is exact and co(K) ⊆
co
∗
(K) ⊆ core (ν). Proposition 1 guarantees that ν is continuous and non
atomic, a second application of Proposition 1 yields that core (ν) ⊆ na ().
Since
min
µ∈core(ν)
µ(A) = ν (A) = ν
K
(A) = inf
φ∈K
φ (A) , ∀A ∈ ,
Corollary 8 implies core (ν) ⊆ co (K). Therefore co(K) = co
∗
(K) = core (ν) ⊆
na () and the Mazur Compactness Theorem implies that of core (ν) is norm
compact. Exactness of ν and Theorem 2 imply that ν is a market game.
In sum, (i)–(v) are equivalent and the ﬁrst part of (7) holds.
Next we show the equivalence between (i) and (vi) and that the second part
of (7) holds.
(i) ⇒ (vi). Since core (ν) is a norm compact subset of na (), for all ε > 0,
there is a ﬁnite subset M of core (ν) such that for all µ ∈ core (ν) there exists
φ ∈ M with µ −φ < ε. Consider the exact linear production game ν
M
.
Clearly, ν
M
≥ ν. Let Abe any coalition and µ ∈ core (ν) such that µ(A) = ν (A).
If φ ∈ M satisﬁes µ −φ < ε, we have
0 ≤ ν
M
(A) −ν (A) ≤ φ (A) −µ(A) ≤ φ −µ < ε.
So that sup
A∈
[ν
M
(A) −ν (A)[ ≤ ε, which implies that ν belongs to the sup
norm closure of the set of exact linear production games.
(vi) ⇒ (i) and the second part of (7). Let ν
n
be a sequence of exact linear
production games uniformly converging to ν. Notice that ν
∗
n
is a nacontinuous
422 M. Amarante et al.
and bounded function on B
1
() for all n ∈ N. Next we show that ν
∗
n
is a Cauchy
sequence (in the space C
b
(B
1
()) of all bounded and nacontinuous functions
on B
1
() endowed with the supnorm). For all ε > 0 there exists p ∈ N such
that for all m, n ≥ p
sup
A∈
[ν
m
(A) −ν
n
(A)[ ≤ ε,
but, the function
¸
¸
ν
∗
m
−ν
∗
n
¸
¸
: B
1
() →Ris nacontinuous and the characteristic
functions are nadense in B
1
(), therefore
sup
f ∈B
1
()
¸
¸
ν
∗
m
(f ) −ν
∗
n
(f )
¸
¸
≤ ε.
Let ν
∗
be the limit of ν
∗
n
in C
b
(B
1
()) with the supnorm. Checking that ν is
superadditive, ν
∗
is an extension of ν to B
1
() (obviously nacontinuous), and
that ν
∗
satisﬁes condition (6) of (ii) is an easy exercise. Hence ν is an exact
market game.
Now, notice that by (iii) we have that ν
∗
= ν
e[B
1
()
and ν
∗
n
= (ν
n
)
e[B
1
()
. The
uniform convergence of ν
∗
n
to ν
∗
amounts to say that
lim
n
_
sup
f ∈B
1
()
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(f ) − min
µ∈core(ν)
µ(f )
¸
¸
¸
¸
_
= 0. (18)
Let λ
0
∈ na () be a probability measure such that core (ν) ⊆ ca (, λ
0
) and
λ
n
∈ na () be a probability measure such that core (ν
n
) ⊆ ca (, λ
n
) for all
n ≥ 1. The nonatomic probability measure
λ =
1
2
λ
0
÷
1
2
n≥1
1
2
n
λ
n
is such that core(ν), core (ν
n
) ⊆ ca (, λ) for all n ≥ 1. Therefore, (18) implies
0 ≤ lim inf
n
_
sup
f ∈I
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(f ) − min
µ∈core(ν)
µ(f )
¸
¸
¸
¸
_
= 0.
thus
0 ≤ lim inf
n
_
sup
g∈B
L
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(g) − min
µ∈core(ν)
µ(g)
¸
¸
¸
¸
_
≤ lim sup
n
_
sup
g∈B
L
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(g) − min
µ∈core(ν)
µ(g)
¸
¸
¸
¸
_
= lim sup
n
_
sup
f ∈I
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(2f −1
) − min
µ∈core(ν)
µ(2f −1
)
¸
¸
¸
¸
_
Cores of nonatomic market games 423
= lim sup
n
_
sup
f ∈I
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(2f ) − min
µ∈core(ν)
µ(2f ) ÷ν () −ν
n
()
¸
¸
¸
¸
_
≤ lim sup
n
_
sup
f ∈I
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(2f ) − min
µ∈core(ν)
µ(2f )
¸
¸
¸
¸
÷[ν () −ν
n
()[
_
= lim
n
_
2 sup
f ∈I
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(f ) − min
µ∈core(ν)
µ(f )
¸
¸
¸
¸
÷[ν () −ν
n
()[
_
= 0.
This concludes the proof since the Hausdorff distance between core (ν
n
) and
core (ν) in ca (, λ) (and hence in ba ()) is given by (see, e.g., Aliprantis and
Border 1999, Ch. 6.7):
d
H
(core (ν
n
) , core (ν)) = sup
g∈B
L
∞
(λ)
¸
¸
σ
core(ν
n
)
(g) −σ
core(ν)
(g)
¸
¸
= sup
g∈B
L
∞
(λ)
¸
¸
¸
¸
min
µ
n
∈core(ν
n
)
µ
n
(g) − min
µ∈core(ν)
µ(g)
¸
¸
¸
¸
.
¬.
Proof of Proposition 4 Set V
C
(f ) = min
p∈C
p(f ) for all f ∈ B().
(i) ⇒(ii). If C is a norm compact (resp. ﬁnite dimensional) subset of na (),
then (see the proof of Theorem 2) V
C
is nacontinuous on B
B()
(resp. B()),
and (ii) follows.
(ii) ⇒(i). For all r ∈ R, {V
C
≤ r]∩B
B()
(resp. {V
C
≤ r]) and{V
C
≥ r]∩B
B()
(resp. {V
C
≥ r]) are naclosed. Therefore, V
C
is a superlinear and nacontinuous
extension of ν
C
to B
B()
(resp. B()). The technique used at the end of the
proof of Theorem 2 delivers (i). ¬.
Proof of Corollary 4 Set V
C
i
(f ) = min
p∈C
i
p(f ) for all f ∈ B(), i = 1, 2.
(i) ⇒ (ii). For all f ∈ B(), f ∼
i
V
C
i
(f ) 1
for i = 1, 2, hence V
C
2
(f ) 1
∼
2
f
2
V
C
1
(f ) 1
and V
C
2
(f ) ≥ V
C
1
(f ), in particular ν
C
1
≤ ν
C
2
.
(ii) ⇒ (i). By Corollary 8, C
2
⊆ co(C
1
). Therefore, V
C
2
≥ V
co(C
1
)
= V
C
1
,
and (i) immediately follows. ¬.
Corollaries 5, 6, and 7 are just restatements of Proposition 3, Corollaries 2,
and 3.
References
Aliprantis CD, Border KC (1999) Inﬁnite dimensional analysis. Springer, Berlin Heidelberg New
York
Aumann RJ, Shapley LS (1974) Values of nonatomic games. Princeton University Press, Princeton
Barbu V, Precupanu T (1986) Convexity and optimization in Banach spaces. Reidel Publishing
Company, Dord recht
Billera LJ, RaananJ (1981) Cores of nonatomic linear productiongames. MathOper Res 6:420–423
Chateauneuf A, Maccheroni F, Marinacci M, Tallon JM (2005) Monotone continuous multiple
priors. Econ Theory 26:973–982
424 M. Amarante et al.
Danilov VI, Koshevoy GA (2000) Cores of cooperative games, superdifferentials of functions, and
Minkowski difference of sets. J Math Anal Appl 247:1–14
Diestel J, Uhl JJ (1977) Vector measures. American Mathematical Society, Providence
Dunford N, Schwartz JT (1954) Linear operators, part I: general theory. Wiley, London
Einy E, Moreno D, Shitovitz B (1999) The core of a class of nonatomic games which arise in
economic applications. Int J Game Theory 28:1–14
Ghirardato P, Marinacci M (2002) Ambiguity made precise: a comparative foundation. J Econ
Theory 102:251–289
Gilboa I, Schmeidler D (1989) Maxmin expected utility with a nonunique prior. J Math Econ
18:141–153
Hart S (1977a) Asymptotic values of games with a continuum of players. J Math Econ 4:57–80
Hart S (1977b) Values of nondifferentiable markets with a continuum of traders. J Math Econ
4:103–116
Hart S, Neyman A (1988) Values of nonatomic vector measure games. J Math Econ 17:31–40
Hormander L (1954) Sur la fonction d’appui des ensembles convexes dans une espace localement
convexe. Arkiv Matematik 3:181–186
Huber PJ, Strassen V (1973) Minimax tests and the Neyman–Pearson lemma for capacities. Ann
Stat 1:251–263
Marinacci M, Montrucchio L (2003) Subcalculus for set functions and cores of TU games. J Math
Econ 39:1–25
Marinacci M, Montrucchio L (2004) Introduction to the mathematics of ambiguity. In: Gilboa I
(ed) Uncertainty in economic theory, Routledge, New York, pp 46–107
Marinacci M, Montrucchio L (2005) Stable cores of large games. Int J Game Theory 33:189–213
Megginson R (1998) An introduction to Banach space theory. Springer, Berlin Heidelberg New
York
Mertens JF (1980) Values and derivatives. Math Oper Res 5:523–552
Milchtaich I (1998) Vector measure games based on measures with values in an inﬁnite dimensional
vector space. Games Econ Behav 24:25–46
Neyman A (2001) Values of nonatomic vector measure games. Isr J Math 124:1–27
Neyman A (2002) Values of games with inﬁnitely many players. In: Aumann RJ, Hart S (eds)
Handbook of game theory, vol III. Elsevier, Amsterdam
Owen G (1975) On the core of linear production games. Math Program 9:358–370
Schmeidler D (1986) Integral representation without additivity. Proc Am Math Soc 97:255–261
Schmeidler D (1989) Subjective probability and expected utility without additivity. Econometrica
57:571–587
Wasserman LA, Kadane JB (1990) Bayes’ theorem for Choquet capacities. Ann Stat 18:1328–1339