The Burkill-Cesari integral on spaces of absolutely continuous

games
F. Centrone
Dipartimento di Studi per lEconomia e lImpresa
Universita del Piemonte Orientale
Via Perrone 18, 28100 Novara
francesca.centrone@eco.unipmn.it
Phone: +390321375320
Fax: +390321375305
A. Martellotti
Dipartimento di Matematica e Informatica
Universita di Perugia
Via Vanvitelli 1, 06123 Perugia
amart@dmi.unipg.it

Abstract
We prove that the Burkill-Cesari integral is a value on a subspace of AC and then discuss
its continuity with respect to both the BV and the Lipschitz norm. We provide an example of
value on a subspace of AC strictly containing pN A as well as an existence result of a Lipschitz
continuous value, different from the Aumann and Shapleys one, on a subspace of AC .

Key words: TU games, derivatives of set functions, Burkill-Cesari integral, value, semivalue,
Lipschitz games.
MSC2010: 28A15, 28A25, 91A12, 91A13,
JEL classification: C71

1 Introduction

Since the seminal Aumann and Shapleys book [1], it is widely recognized that the theory of value
of nonatomic games is strictly linked with different concepts of derivatives. A few papers, up to the
recent literature, have investigated these relations (see for example [6], [14], [16]). In [1] Aumann
and Shapley proved the existence and uniqueness of a value on the space pN A, namely the space
spanned by the powers of nonatomic measures (which, under suitable hypotheses, contains for
instance games of interest in Mathematical Economics such as transferable utilility economies
with finite types). Moreover, in Theorem H of [1], the authors provided an explicit formula for
the value of games in pN A in terms of a derivative of their ideal set function .

Corresponding author

1
To the best of our knowledge, the more general contribution so far on the link between derivatives
of set functions and value theory is Mertens [14]; his results led to the proof of the existence of
a value on spaces larger than pN A. A more recent contribution on the same subject is due to
Montrucchio and Semeraro [16]. The problem of the existence of a value on the whole space AC
of absolutely continuous games (which contains pN A) is instead still unsolved and challenging.
Therefore, proofs of the existence of a value on other subspaces of AC, beyond pN A, can represent
a step forward, and investigations of this kind appear to be in order.
In Epstein and Marinacci [6] the question of the relation between their refinement derivative and
the value was posed, and a possible direction sketched; in Montrucchio and Semeraro [16], the
authors applied their more general (i.e. without the nonatomicity restriction) notion of refinement
derivative to the study of the value on certain spaces of games by extending the potential approach
of Hart and Mas-Colell [8] to infinite games.
In a previous paper ([3]) we had pointed out that, in a nonatomic context, the refinement derivative
is connected with the classical Burkill-Cesari (BC) integral of set functions and, for BC integrable
functions, the BC integral coincides with the refinement derivative at the empty set. Though less
general, the BC integral is analytically more treatable.
Motivated by all these facts, in [3] we have started the study of the BC integral in the framework
of transferable utility (TU) games.
In this paper we extend our investigation to develop the connection with the theory of value or of
semivalue, also in the light of the problem exposed above. In Section 2 we introduce the general
class of BC integrable games and prove that, under natural assumptions, regular measure games
belong to this class. Moreover, the class of BC integrable games contains a dense subspace of the
largely used space pN A, where continuous values and semivalues are largely described in the
literature (see for instance [5]). In addition, on the subspace of BC integrable games in AC, the
BC integral turns out to be a semivalue. Then, as natural, one considers the subspace of feasible
BC integrable games, that is, the space where the BC integral becomes indeed a value. We provide
examples of feasible BC integrable games that do not belong to pN A. Actually by means of these
examples, we provide a large class of subspaces of AC where the BC integral is a value, and we
also exhibit an example of a subspace of AC strictly containing pN A on which a value can be
defined as a sort of direct sum of the usual Aumann-Shapley value and this new set function.
Unfortunately, the BC integral proves to be not continuous with respect to the BV norm on the
BC integrable games in AC. As continuity appears to be a crucial property for many questions
concerning the value on subspaces of BV , in Section 3 we specialize to the subspace AC AC of
the so called Lipschitz games, where a suitable finer norm (the kk -norm) is defined and used as an
alternative (see [9], [15]). We completely characterize the scalar measure games (where the measure
is nonnegative) that belong to AC and we show that the BC integral on an appropriate subspace
is a Milnor (therefore || || -continuous) semivalue. Then again we turn our attention to the

2
subspace of feasible BC integrable games in AC , and to its closure in the k k -norm, F EAS .
In the final part of the paper we consider and somehow characterize the space F EAS pN A
(namely the k k -closure of vector measure games generated by polynomials), and we show that
the BC integral is not the unique value on it, in that it does not coincide with the Aumann-Shapley
value.

2 A semivalue on a space of Burkill-Cesari integrable games

From now on we will denote by (, ) a standard Borel space (i.e. is a Borel set of a Polish
space, and the family of its Borel subsets). represents a set of players, and the -algebra
of admissible coalitions.
A set function : R such that () = 0 is called a transferable utility (TU) game.
For the sake of brevity we refer the reader to [1] and to [11] for the terminology concerning TU
games: in particular BV will denote the space of all bounded variation games, endowed with the
variation norm || ||BV . The subspace of nonatomic countably additive measures will be denoted
by N A and the cone of the nonnegative elements of N A by N A+ .
Throughout the paper we shall write  to mean that absolute continuity by chains hold:

Definition 2.1 ([1]). A chain C is a non-decreasing family of sets:

= S0 S1 Sn = .

A link of a chain is a set of two consecutive elements {Si1 , Si }. A subchain of a chain is any set
of links.

A chain will be identified with the subchain consisting of all the links. Given a game and a
subchain of a chain C, the variation of over is defined as
X
kk := |(Si ) (Si1 )|,

where the sum ranges over all indexes i such that {Si1 , Si } is a link in the subchain.

Definition 2.2 ([1]) If and w are two games defined on , is said to be absolutely continuous
with respect to w if for every > 0 there exists a > 0 such that for every chain C and every
subchain of C,
kwk = kk .

The space AC BV introduced by Aumann and Shapley ([1]) is the space of all games for
which there exists N A+ such that is absolutely continuous with respect to .

We also refer the reader to [6] and to our previous paper [3] for details about the Epstein-Marinacci
refinement derivative.

3
A partition D of a set E is a finite family of pairwise disjoint elements of , whose union is E.
By (E) we shall denote the set of all the partitions of E. A partition D (E) is a refinement
of another partition D (E) if each element of D is union of elements of D.
As in [4], given a monotone nonatomic game one defines the mesh of a partition D as

(D) = max{(I), I D}. (1)

and the Burkill-Cesari (BC) integral of a game w.r.t. as:

Z X
E 7 = lim (I). (2)
E (D) 0 ID
D (E)

We denote by BC the space of games such that there exists N A+ so that is BC integrable
with respect to the mesh .
The BC integral does not depend upon the integration mesh (see Proposition 5.2 in [3]); in other
words, for every N A+ such that is -BC integrable, the BC integral is the same. Moreover,
the BC integral of a game is a finitely additive measure and, as observed in [3], it coincides with
+
the Epstein-Marinacci outer derivative at the empty set (, ) (see [6]). Hence, from now, on
+
we shall use the notation (, ).
As we shall see, the space BC contains many games which are of interest in the literature: we begin
by recalling a sufficient condition for vector measure games to be in BC, which is an immediate
consequence of Theorem 6.1 in [3].

Proposition 2.1 Let P : Rn be a nonatomic vector measure, and let f : Rn R be

a function with f (0) = 0. If f is differentiable at 0, then the game = f P BC, and
+
(, F ) = f (0) P (F ), F .

Anyway, the class of BC integrable games is not limited to smooth measure games. Indeed note
that the converse implication of the previous proposition does not hold: consider as in [3] (Example
3.2) f : R R to be any discontinuous solution to the functional equation

f (x + y) = f (x) + f (y), x, y R.

Then = f P is additive, and therefore for each F , and each D (F ) one has
X X
(I) = f [P (I)] = f [P (F )] = (F )
ID ID

and hence is BC integrable with respect to P although f is not differentiable at 0.

As for the relation between the spaces AC and BC, it is well known (see Theorem C of [1]) that

the game = , (with the Lebesgue measure on [0,1]) belongs to AC. Anyway / BC. To see
it, one has to show that is not BC integrable with respect to any mesh determined by some

4
 N A+ . Indeed is not refinement differentiable at , and then it cannot be BC integrable with
respect to ; to get convinced that does not admit outer refinement derivative at , observe
that for every partition Do () every > 0 we can provide a refinement D0 = {I1 , . . . , Ik , Io }
such that (I1 ) = . . . = (Ik ) and (Io ) < (see [13], Lemma 3.5). Clearly we can choose k quite

larger than say ]Do . Also we can chose = () determined by the uniform continuity of x 7 x
on [0, 1]. Thus  
X p 
(I) k(1 ) <
 

 0

ID
which shows that the refinement limit does not exist.
1
In fact ( \ Io ) > 1 , and therefore (Ij ) > , j = 1, . . . , k
r k
1
whence (Ij ) , j = 1, . . . , k, while (Io ) < .
k
So
k
1 X X 1
k (I) = (Io ) + (Ij ) + k .
k ID0 j=1
k
On the contrary, any strongly nonatomic finitely additive measure that is not countably addi-
+
tive provides an example of game belonging to BC (for (, ) = ()), but not to AC (see the
different notions of absolute continuity in [3]).

Therefore, we can now consider the space V = BC AC.

The following result states that the same measure can be used for the absolute continuity and the
BC integrability of a game in V .

Proposition 2.2 The space V can be equivalently defined as the space of games such that there
exists N A+ such that  and is BC integrable with respect to .

Proof. The fact that each game for which there exists N A+ such that  and is BC
integrable with respect to lies in V is straightforward.
Conversely, let V ; then there are 1 , 2 N A+ such that  1 (since AC) and is
2 - BC integrable. Then consider = 1 + 2 ; evidently  and, in view of Proposition 5.2
in [3]), is also -BC integrable. 2

From [1] we recall the following

Definition 2.3 Let G denote the space of automorphisms of (, ), that is isomorphisms of the
space onto itself; then each G induces a linear mapping of BV onto itself, defined by

( )(E) = ((E)) (3)

for E . A subspace that is invariant under for every G is called symmetric.

Proposition 2.3 The space V is symmetric.

5
Proof. We need to prove that for every G and every game V the game defined in
(3) is in V , namely it is  with respect to some nonatomic measure, and it is BC integrable too.
Let be a measure in N A+ with respect to which we have  and is - BC integrable
(thanks to Proposition 2.2 we can always assume that the default measure is the same).
Fix ; note that preserves set operations, therefore easily = is in N A+ .
It is also immediate to check that that  , because transforms chains and subchains into
chains and subchains as well.
It remains to prove that is BC integrable with respect to the mesh . Indeed we shall prove
that

+ +
( , F ) = (, (F )) (4)

for every F .
To this aim, for any F and any > 0 fixed, one has to find (, F ) > 0 such that for every
partition D (F ) with (D) < there holds
 
X 
+
(I) (, (F )) < . (5)
 

 
ID

Since is BC integrable, to each > 0 there corresponds (, (F )) > 0 such that for each
partition D [(F )] with (D) < there follows
 
X 
+
(J) (, (F )) < .
 

 
JD

Clearly we can rewrite (5) as

 
X 
+
[(I)] (, (F )) < .
 

 
ID

We choose (, F ) = (, (F )); thus if D (F ) has (D) < , the corresponding partition

D0 = {(I), I D} [(F )] has (D0 ) < = since, for each I D clearly (I) = (I) =
[(I)] < = and hence
   
X  X 
+ +
[(I)] (, (F )) =  (J) (, (F )) < .
   

   0

ID JD

According to [5] we remind the following definition.

Definition 2.4 A linear mapping : Y N A on a symmetric subspace Y of BV is called a

semivalue provided it satisfies the properties

V.1 (symmetry): = ( ) for each G;

6
V.2 (positivity): is positive, that is for every monotone game the measure () is non negative;

V.3 (projection axiom): is the identity operator on N A Y.

When satisfies also

V.4 (efficiency): for each Y there holds ()() = ()

is called a value on Y (compare with [1]).

+
The following result immediately derives from (4) and the definition of .

+
Corollary 2.1 The mapping : V NA is a semivalue on V .

+
Consider now the space F EAS0 = { BC : (, I) = (I)} and define the space F EAS =
+
F EAS0 V = F EAS0 AC. Obviously is a value on F EAS.
The next example shows that this space contains several games which do not belong to pN A, and
+
through these one finds several new subspaces of AC on which defines a value.

Example 2.1 Let be a signed measure on a measurable space (, ) with range say [-1,1] and,
for instance () = ]0, 1[, and let f : [1, 1] R be defined as

2 if x [1, [

f (x) = x if x [, ]
2
if x ], 1].

Let = f and consider the space E = span{ : G} . Then AC \ pN A so that

+
E AC but E 6 pN A and is a value on E. In fact

+
Fact I F EAS0 ; indeed, since f 0 (0) exists, by Proposition 2.1, BC and (, ) =
+
whence immediately (, ) = 2 = ().

Fact II 6 pN A: indeed, according to Kohlberg ([10]) a measure game f , where is a signed

measure, is in pN A iff the function f is continuously differentiable in ] 1, 1[ .

Fact III AC; for the proof, see the Appendix.

+
Finally E is symmetric, and by linearity and relationship (4), is efficient on E; therefore it is
a value on E.

It is clear that one can use functions of different form to provide classes of measure games f with
+
signed that are in AC \ pN A and in F EAS0 , and hence similar subspaces of AC on which
is a value. These subspaces will not be contained in pN A because the generating game 6 pN A.
However we can go a little further; in fact the next example shows that, by means of a similar
construction, there are subspaces in AC strictly containing pN A on which a value can be defined.

7
Example 2.2 On = [1, 1] equipped with the usual Borel -algebra, consider the signed
R
measure (S) =  S (signx)dx
 .    
1 1 1 1
Denote by I1 = 1, , I2 = , , I3 = , 1 and take the function f : R defined
2 2 2 2
by
1

x if x I1
2






f (x) = 0 if x I2




x 1


if x I3
2
Define the scalar measure game = f and take then w = + where denotes the usual
Lebesgue measure.
As in the previous example, and hence w are in F EAS0 , w 6 pN A (for 6 pN A) and w AC
(again, for the proof, see the Appendix).
Furthermore (and hence w) is in pN A0 , since f is continuous ([17]).
Thus F EAS \ pN A.
Again take E = span{ w : G} and denote by pN AE the smallest linear subspace containing
Xm
pN A and E. Define : pN AE N A as follows:, for each = ci i w + p pN AE, set
i=1
m
X
+
() = ci (i w, ) + AS (p), where AS is the usual Aumann and Shapley value.
i=1
is a value on pN AE (see the Appendix for the proof).

+
Finally, the following example shows that is not continuous on V equipped with the variation
norm.

Example 2.3 Consider the sequence of scalar measure games n = fn P , where fn (x) =
+ 0
min {x, 1/n} and P N A1 . Then kn kBV 0 as n +. However (n , S) = fn (0)P (S) =
P (S) does not converge to 0.

3 The operator + on subspaces of Lipschitz games

In [15] the author considers the class AC of Lipschitz games, that is games in BV for which
there exists a measure N A+ such that both and + are monotone games. The reason
why these games are called Lipschitz is the fact that the condition can be equivalently labelled in
the following form: for every link S T in there holds

|(T ) (S)| (T ) (S). (6)

The connection to the Lipschitz condition is made even stronger by the following

Proposition 3.1 For a scalar measure game = f the following are equivalent:

8
1. AC ;

2. f is Lipschitz on the interval [0, ()] (with () as Lipschitz constant, for each N A+ for
which (6) above is satisfied);

3. (6) holds for = L, with L Lipschitz constant for f .

Proof. To prove that 1. implies 2. , assume that = f is a Lipschitz game, and let N A+
be a measure for which (6) holds. For simplicity we assume that is a probability measure. Let
t, t0 [0, 1] with say t < t0 .
Then, by Lyapounoff Theorem, there exist sets t t0 such that

(, )(t ) = t(, )(), (, )(t0 ) = t0 (, )().

Hence f (t) = f [(t )] = (t ) and f (t0 ) = (t0 ); then from (6)

|f (t) f (t0 )| = |(t ) (t0 )| (t \ t0 ) = (t0 t)().

The fact that a Lipschitz function f generates a Lipschitz game (where one can precisely choose
L = in (6)) is immediate, so 2. implies 3.
Also 3. implies 1. trivially. 2

It is immediate to note that AC AC. However the smaller space can be equipped with an
alternative norm defined in the following way; for every N A+ such that (6) holds, write
  . Then we set

kk = inf{(), N A+ ,   }. (7)

Then AC is a Banach space when equipped with the above norm.

Again from [15] we quote the following definition.

Definition 3.1 Let AC and define the following two subsets of N A:

D = { N A|  }, D = { N A|  }.

Then the following two measures exist: = g.l.b.D , = l.u.b.D , and immediately
(although the symbol should be distinguished from , as the first one refers to setwise ordering,
the second to the order induced by the cone of monotonic games, in the case of measures they
actually assume the same meaning).
Let N A Q AC be a linear subspace, and let : Q N A be a linear operator; we shall say
that is a Milnor operator (MO) provided for every Q we have

9
Consider now the vector subspace Q = BC AC of Lipschitz games that are BC integrable.
Q is strictly included in BC, for there are easy examples of games in BC \ AC .
For instance, consider the function f : [0, 1] R defined as

2
x

if 0 x
2

f (x) =
1 x2 if 2 x 1



2
and the scalar measure game = f where represents the usual Lebesgue measure. Then
+
BC with () = f 0 (0) = thanks to Proposition 2.1, but 6 AC since f is not
Lipschitz on [0, 1].

Also the inclusion Q AC is a strict one, for there are Lipschitz games that are not in BC. To
see this we need the following result, which is a partial converse of Proposition 2.1.

Proposition 3.2 Let the scalar measure game = f , N A+ , 6= 0 be in AC ; then the

following are equivalent

1. f admits right hand-side derivative at 0;

2. is -BC integrable and hence Q.

Proof. The implication 1.= 2. follows from Proposition 2.1.

We turn then to the implication 2. = 1.
As Q, there exists N A+ such that is -BC integrable. Since AC we already
f (x)
know that f is Lipschitz; hence the ratios are bounded. Assume by contradiction that f 0 (0)
x
does not exist. Then it can only happen that
f (x) f (x)
< `1 = lim inf < lim sup = `2 < +.
x0 x x0 x

Choose then two decreasing sequences {x0n }, {x00n } ]0, ()] with lim x0n = lim x00n = 0 and
n n

f (x0n ) f (x00n )
lim = ` 1 , lim = `2 .
n x0n n x00n

Fix F with (F ) > 0 and ]0, (F )]. Then there exists n N such that for each n > n
 f (x0n )  f (x00n )
   
 < .


 x0 ` 1
 < , 
 3(F )  x00 ` 2  3(F )
n n

e > n such that |f (x)| < whenever x x0ne ;
By means of the continuity of f at 0, choose next n
3
0 0 (F )

e can be chosen so that |`1 |xne < and such that xne
also n < where is the parameter
3 (F ) 3
of BC integrability.

10
Choose now the following D (F ): by means of Lyapounoff Theorem, divide F into finitely
  k
(F ) 0 [
many sets, say I1 , . . . , Ik , each with (, )(Ij ) = x0ne , xne , until F \ Ij x0ne and
(F )
j=1
k
[ (F )
then choose Ik+1 = F \ Ij ; thus easily (Ik+1 ) =
(Ik+1 ).
(F )
j=1

Then for D = {I1 , . . . , Ik , Ik+1 } one has (D) < .
3
We have then, similarly to the computation in Proposition 2.1

X k
X
|f [(I)] `1 (I)| = |f (x0ne ) `1 x0ne | + |f [(Ik+1 )] `1 (Ik+1 )|
ID n=1

k k 
 f (x0ne ) `1 x0ne  0

 x + + .
X X
|f (x0ne ) `1 x0ne | + |f [(Ik+1 )]| + |`1 |xe0k = 
0  ne 3 3
 xne
n=1 n=1

As for the first sum we have the following estimate

k  k
 f (x0ne ) `1 x0ne  0

X
x <
X (F \ Ik+1 )

0 n x0ne = < .
 xne  e
3(F ) 3 (F ) 3
n=1 n=1

In conclusion
X
|f [(I)] `1 (I)| < .
ID

Clearly we can repeat this construction with x00ne and find another partition D (F ) with

(D ) < as above; again
3
X
|f [(I)] `2 (I)| < .
ID

It is then clear that, since `1 6= `2 , the game is not -BC integrable. 2

Therefore for instance, taking f : [0, 1] R defined as


x sin log x if x 6= 0,
f (x) =
0 if x = 0

the game = f AC , since for x 6= 0 one has f 0 (x) = sin log x + cos log x L , but, as
f 0 (0) does not exist, according to the previous result, 6 BC.

We shall need in the sequel the following Lemma.

Lemma 3.1 The space AC is symmetric and the following equality holds for every G

k k = kk . (8)

Proof. Fix > 0 and choose N A+ such that   and () < kk + .

11
Let = N A+ . If A B then (A) (B) and therefore, by monotonicity, ( )[(A)]
( )[(B)], or else (A) (A) (B) (B), which is the same as to say that
is monotone, and hence  .
In a completely analogous way, as () = (), one reaches ()  (). In conclusion
  .
Moreover, () = [()] = [()] = () whence

k k () = () < kk + .

To prove the converse inequality, first of all, for N A consider the game 1
defined in the
following fashion: for every B set A = 1 (B) and set

1
(B) = (A)

so that [1 1
] = . It is a routine computation, based on the properties of , to show that
is a countably additive measure as well.
Again fix > 0 and choose N A such that   and () < k k + .
Take = 1
defined above.
Now  = (or else + monotone) implies in turn that + is monotone too,
and similarly  .
Hence
kk () = [1 ()] = () < k k +

which concludes the proof of relationship (8). 2

In Q we have the following result.

+
Proposition 3.3 The BC integral is a Milnor semivalue on Q.

Proof. Let be any game in Q, D ; then is a monotone game, and hence ()(E) 0
+ +
for each E , which in turn implies immediately that ( ) 0, namely () 0
+ +
setwise in . () being a measure, this equivalently says that () is monotone, that
+ +
is  (). In complete analogy if D then  ().
+ +
Hence for Q we have necessarily () which proves that is a MO.
+
Finally, we deduce from (4) the symmetry of the operator , and the proof is thus complete. 2
+
According to Theorem 1.8 in [15], can be extended to the whole space AC in such a way
+
that the extension, which we shall label as f
, remains a linear MO. Recall moreover that MO
on subspaces of AC are continuous with respect to the norm k k (Lemma 1.6 in [15]).
Let Q denote the k k -closure of Q. Then on Q we have

+
Theorem 3.1 f
is a Milnor semivalue on Q .

12
Proof. If Q , there exists a sequence in Q, say (k )k that k k -converges to .
kk + kk +
Because of (8), for each G, we have that k . But then ( k ) f ( ) too.
kk kk
Similarly + ( ) f
k
+
(), and then, again by (8), [ + ( )] [f
k
+
()].

+
In conclusion [f +
()] = ( ).
f 2

Since powers of probabilities belong to Q, there immediately follows that

+
is a k k -continuous semivalue on pN A .
Corollary 3.1 f

We point out that, as Q, Q , pN A are symmetric subspaces of AC , there follows from [15]
(Theorem 3.1) that + and f

+
are diagonal.

Moreover from [15] (Theorem 2.1), there exists a Borel measure on [0, 1] such that the following
+
representation of f
on pN A holds
Z 1
+
f
(, S) = (t1 , 1S )d, S ,
0

where is the ideal extension of the game defined in [1], Theorem G.

Define now the space F EAS1 = Q F EAS. Note that F EAS1 = Q F EAS0 . Also define
+
F EAS = { Q : f
(, ) = ()}.

Proposition 3.4 F EAS is the k k -closure of F EAS1 .

Proof. Denote by [QF EAS] the kk -closure of F EAS1 . Indeed it is immediate to prove that
[Q F EAS] F EAS : take [Q F EAS] , and a sequence (k )k F EAS1 converging in
the k k norm to . Then Q . Moreover, the convergence of (k )k to implies that k ()
k
+
converges to (). Furthermore, + ( , ) converges in variation (and hence setwise) to f (, ).

Therefore:
+ +
() = lim k () = lim (k , ) = f
(, ),
k k
so F EAS .
For the converse inclusion, take F EAS . Hence is the k k - norm limit of a sequence
(k )k of elements in Q. Construct now the sequence
+
k = k + [ (k , ) k ()]2 ,

for some N A1 . We claim that k F EAS1 and k k k -converges to .

+ +
Indeed, by the feasibility of and by the fact that k () () and (k , ) f
(, ), it
+ +
follows that k k k - converges to . Moreover, as (k , ) = (k , ), it immediately follows
that k F EAS. 2

+
Obviously f is a value of F EAS but, unfortunately, on the subspace pN A F EAS we lose
+
uniqueness, in that f does not agree with the Aumann-Shapley (AS) value.

13
 x2 x y2
Example 3.1 Let f (x, y) = 1 (x) + 2 (y) where 1 (x) = + and 1 (y) = + y . Let 1
2 2 2
and 2 be two linearly independent measures in N A1 , and = f where = (1 , 2 ) .
+
We claim that F EAS , but f () 6= (), where  denotes the AS value.
Z 1  Z 1 
0 0
1 2
The function f C (R ) and we know that () = fx (t, t)dt 1 + fy (t, t)dt 2 and
0 0
+ 0 0 + 1 2
() = fx (0)1 + fy (0)2 . Hence f
() = 2 + 2 , while () = 1 + 2 . It is also immediate
to check that F EAS .

Indeed one can provide infinitely many examples of subspaces of pN A where a value different
from the AS one is defined. By means of Theorem 2.1. in [15], each probability measure on
[0, 1] generates a Milnor semivalue on pN A which clearly becomes a value on the subspace
F EAS = { pN A : ()() = ()}. In our case for the Dirac measure 0 based at 0, we
have precisely F EAS0 = pN A F EAS . However we point out that our main interest in this
paper is not uniqueness of the value but the fact that the Burkill Cesari integral and hence the
Epstein-Marinacci derivative constitute a value.
It may anyway be of interest to characterize games in pN A F EAS . Remember that pN A is
the kk -closure of the linear span B of powers of nonatomic probability measures. With the same
technique used in Proposition 3.4 one proves that pN A F EAS is in fact the k k -closure
n
X
of B F EAS0 . It is now easy to characterize these games: in fact if = k kk B F EAS0 ,
k=1
n
X
+
with k N A+ then (, ) = 1 1 . So F EAS0 if and only if k [k ()k ] = 0.
k=2

4 Appendix

We shall first prove that the measure game in Example 2.1 is in AC.
Let I1 = [1, [, I2 = [, ], I3 =], 1] and let m = || (the total variation of ).
Fix a Hahn decomposition (P, N ) of .
We shall prove that  m.
Fix > 0 and choose = . Let C be a chain, and a subchain with kmk < . We claim
that kk < .
In fact, let (Si , Si+1 ) be a link in . Then, as m(Si+1 ) m(Si ) < , necessarily if (Si ), (Si+1 )
do not belong to the same Ij , then they are at most in two contiguous Ij s.
Indeed, suppose for instance that (Si ) I1 , (Si+1 ) I3 , namely (Si ) < , (Si+1 ) > .
Hence (Si P ) + (Si N ) = m(Si P ) m(Si N ) < and analogously m(Si+1 P )
m(Si+1 N ) > . Thus

[m(Si+1 P ) m(Si+1 N )] [m(Si P ) m(Si N )] > 2

14
that is
[m(Si+1 P ) m(Si P )] [m(Si+1 N ) m(Si N )] > 2

namely
m[(Si+1 \ Si ) P ] m[(Si+1 \ Si ) N ] > 2

whence, a fortiori, m(Si+1 \ Si ) > 2 which contradicts the initial assumption kmk < .
Then an easy computation shows that in both cases (either (Si ), (Si+1 ) belong to the same Ij
or they lye in two contiguous Ij s)

|(Si+1 ) (Si )| |m(Si+1 ) m(Si )|

and thus the required inequality, since < 1.

In a completely analogous way one shows that the game in Example 2.2 is in AC .
To prove that : pN AE N A in the same example is a value, the only property that needs to
be checked is positivity (the others are obvious).
Xm
Assume that = ci i w + p pN AE, p pN A is monotone.
i=1
m
X m
X
We shall also use the alternative form = ci i + u, where u = ci i + p.
i=1 i=1
Now by (4)
m
X m
X
+
() = ci (i w, ) + AS (p) = ci i + AS (p) = AS (u)
i=1 i=1

+ +
since (, ) = 0, while (, ) = . Hence we need to prove that AS (u) 0.
Note first that since pN A0 , the ideal set function exists,
Z and it actually coincides with

f ([1] Proposition 22.16 page 152), and that () = d. Hence the ideal function

exists too.
(t, S), (t, S), t [0, 1], S from [17].
Recall now the definition of +
In particular, since
(t1 + 1S ) (t1 )
+ (t, S) = lim
0
and
 Z Z   Z 
(t1 + 1S ) (t1 ) = f t 1 d + 1S d f t 1 d = f [t()+ (S)]f [t()] = f [ (S)]

(for () = 0), we find

f [ (S)]
+ (t, S) = lim = f 0 (0)(S) = 0
0
and similarly for the left-hand side limit. In other words (t, S) = 0 for each t, S.
We note now that, since is a measure game, for every G, ( )+ = +
= 0 (see [1]

Proposition 22.16, p. 152, and the proof of Theorem 3 in [17]).

15
 (t, S), (t, S) are non negative for each t, S. Hence
Now, since is monotone, both +

m
ci (i )+ (t, S) + u (t, S) = u (t, S)
X

0 + (t, S) =
i=1

where in the last summand we do not need to distinguish between u+ and u as u pN A and
hence u exists.
But then Z 1
0 u (t, S)dt = AS (u)(S), S
0
which is precisely what we wanted to prove.

References

[1] R.J. Aumann, L.S. Shapley, Values of non-atomic games, Princeton University Press, Prince-
ton, NJ, 1974.

[2] D. Candeloro, A. Martellotti, Geometric properties of the range of two-dimensional quasi-

measures with respect to the Radon-Nikodym Property, Advances in Mathematics 93 (1992)
9-24.

[3] F. Centrone, A. Martellotti, A mesh-based notion of differential for TU games, J. Math. Anal.
Appl. 38 (2012) 1323-1343.

[4] L. Cesari, Quasi additive set functions and the concept of integral over a variety, Trans. Amer.
Mat. Soc. 102 (1962) 94-113.

[5] P. Dubey, A. Neyman, J. Weber, Value Theory without efficiency, Math. Oper. Res. 6 (1981)
122-128.

[6] L. Epstein, M. Marinacci, The core of large differentiable TU games, J. Econom. Theory 100
(2001) 235-273.

[7] A. Ghizzetti, A. Ossicini, Polinomi ortogonali e problema dei momenti, Istituto per le Appli-
cazioni del Calcolo Mauro Picone, Roma, 1981.

[8] S. Hart, A. Mas-Colell, Potential, value and consistency, Econometrica 57 (1989) 589-614.

[9] S. Hart, D. Monderer, Potentials and weighted values of nonatomic games, Math. Oper. Res.
22 (1997) 619-630.

[10] E. Kohlberg, On non-atomic games: conditions for f pN A, Int. J. Game Theory 2

(1973), 87-98.

16
[11] M. Marinacci, L. Montrucchio, Introduction to the mathematics of ambiguity, in: Uncertainty
in Economic Theory, I. Gilboa (ed.), New York, Routledge, 2004, pp. 46-107.

[12] A. Martellotti, Countably additive restrictions of vector-valued quasi measures with respect
to range preservation, Boll. U.M.I. (7) 2 (1988) 445-458.

[13] A. Martellotti, Core equivalence theorem: countably many types of agents and commodities
in L1 (), Decisions in Econ. and Finance 30 (2007) 51-70.

[14] J.F. Mertens, Values and derivatives, Math. Oper. Res. 5 (1980) 523-552.

[15] D. Monderer, A Milnor condition for nonatomic Lipschitz games and its applications, Math.
Oper. Res. 15 (1990) 714-723.

[16] L. Montrucchio, P. Semeraro, Refinement derivatives and values of games, Math. Oper. Res.
33 (2008) 97-118.

[17] Y. Tauman, Value on a class of Non-Differentiable Market Games, Int. Journal of Game
Theory 10 Issue 3-4 (1981) 155-162.

17