Int. Journal of Game Theory, Vol. 6, :ssue 1, page 23-31. Physica-Verlag, Vienna.

V a l u e s o f G a m e s in Partition Function Form I ) By R.B. Myerson, Evanston 2)

Abstract: Shapley's value axioms are adapted to games in partition function form, and the natural extension of the Shapley value for partition function games is derived.

Games in partition function form were introduced in Lucas and Thrall [ 1963] to generalize the characteristic function form. The Shapley value for games in characteristic function form was developed from three simple axioms in Shapley [ 1953]. In this paper we introduce the natural extension of the Shapley value to the partition function form. L e t N be a fixed nonempty finite set, and let the members of N be interpreted as players in some game situation. Given N, let CL be the set of all coalitions (nonempty subsets) of N,

CL-- {SlS_CN, S~=O}.

Let PT be the set of partitions of N, so (S 1. . . . .

S t} E PT iff: l U S i =N, V/S/~eO, V k S I ( ~ S k = 0 i f k 4:/.

i=t

(2)

And let ECL be the set of embedded coalitions, that is the set of coalitions together with specifications as to how the other player are aligned. Formally: ECL = ((S, Q) IS E Q E PT}. (3)

For any finite set L, let R L denote the set of real vectors indexed on the members ofL. We define a game in partition function form (or a partition function game for short) to be any vector in R ~ ' z . For any such w ~ R ECL and any embedded coalition 1) Research for this article was partly supported by an N.S.F. Graduate Fellowship. 2) Prof. R.B. Myerson, Northwestern University, Graduate School of Management, Evanston, Illinois 60201.

24

R.B. Myerson

(S, Q) E ECL, WS, Q (the (S, Q)-component of w) is interpreted as the wealth, measured
in units of transferable utility, which the coalition S would have to divide among its members if all the players were aligned into the coalitions of partition Q. We assume that, in any game situation represented by such a vector w E R ECL, the universal coalition N (embedded in {N)) will actually form, so that the players will have WN, {N)to divide among themselves. But we also anticipate that the actual allocation of this wealth will depend on all the other potential wealths ws,Q, as they influence the relative bargaining strengths of the players. So we are interested in value functions, which are functions from the space of partition function games R ECL into the space of payoff allocations R N. If 9 (-)is a value function and w E R ECL is a partition function game, then for any n E N we would like to interpret ~n (w) (the n-component of q~ ( w ) E R N) as the utility payoff which player n should expect from the game w. If the value function is to be realistic for these purposes, then it ought to satisfy certain properties, which we will develop. We will need some additional notation. For any Q E PT and Q E PT, define QAQEPTby: QAQ= {S A f f lS EQ, gEO.,S A f f 4=O).
^ A

(4)

Given any (S, Q) and (S, Q) in ECL, we write:

^ ^ A ^

(S, Q) >> (S, Q) iffS 3 S and QAQ =

^ A

(5)

(Read): "(S, Q) covers (S, Q)".) For any S _CN, let I S i be the number of players in S; for any Q E PT, let 1Q t be the number of coalitions in Q. For any w E R ECL and any Q E PT, we adopt the convention that Wr = 0, where 0 is the empty set. Now, suppose rI:N ~ N is any permutation of the set of players. Then without danger of confusion, we can also let [1 act as a permutation on CL and on ECL in the natural way: zr(S)= {rc(n) t n E S ) , V S ~ C L , and

/r (S 1, {S 1. . . . . sk))) = (Tr ($1), (Tr (S 1) . . . . . /r (sk))), V (S 1, (S 1. . . . . sk)) E ECL. Then, for any w E R ECL, define ~ro w CR ECL so that

(Tr o W)rr(S,Q) = WS, Q, V (S, Q) E ECL.

(6)

(equivalently, (n o W ) S , Q , = w n -1 (S,Q).) So/r o w is the partition function game which would result from w if we relabeled the players by permutation 7r. The first axiom asserts that the value should be invadant under such relabelling.

Values of Games in Partition Function Form

25

ValueAxiom 1: For any permutation rr : N ~ N ,

Vw E R ECL, Vn ~N, ~n (w) = ~,~(n) (n o w). Given w E R ECL and S E CL, we say that S is a carrier o f w iff w~,o =

= WSnS, O.A {S,N\S)' ~' (g" Q) E ECL. (Recall Wr = 0.) Note t h a t N is always a carrier. The second axiom suggests that all available wealth should be divided among the members of a carrier. ValueAxiom 2: V w E R ECL, VS E CL, ifS is a carrier
of w, then nES ~" dpn (W) = WN, (iV)" (Note that WN, (N} =

Ws, (S,N\S} whenever S is a carrier of w.)

Using the real vector space nature ofR ECL, we can talk about adding games. Our third axiom suggests that addition partition function games, which could be interpreted as merging games to be played simultaneously, should cause addition to their values,

ValueAxiom 3: V w 1 E R ECL, 'g' w 2 @R ECL,

(w 1 + w 2) = ~ (w 1) + ~ (w2).
These three axioms can be recognized as adaptations of the Shapley value axioms [see Shapley] to the partition function game. Our main result is that these three axioms determine a unique value function, which is the analogue of the Shapley value for partition function games.

Theorem 1 : There exists exactly one value function

qb :R ECL ~ R N which satisfies Value Axioms 1, 2, and 3. Henceforth, we will always use 9 to represent the value function of Theorem 1. The following theorem shows how to compute our value function.

Theorem 2: The value function ~ : R ECL --rR N defined

in Theorem 1 satisfies the following formula, V n EN, ~' w E R ECL :

~n(W)=

(S,Q)~ECL

(--1) IQI'I. (JQI-I)!

NI

s ~ 2 (IQ I - 1 ) ( I N I - I ~s n~

Ws, o '

26

R.B. Myerson

For example, i f N = (1, 2, 3}, then the formula for q~x (w) is: ~b, ( w ) = (1/3)" W(1,2,3},((1,2,3} } + (1/6)" W{1,2}, ({1,2},(3}}--(1/3) "W(3},{(1,2}(3}} + (1/6) 9 w(1,3}, {(1,3}, {2}}-- (1/3) 9 w{2}, {{1,3},{2}} + (2/3)'w(1},((11 ,(2,3}}-(1/3)'w{2,3},((1},{2,3) } + (1/6) "w(2),((1},(2} ' (3}}+ (1/6) 9 w(3},({1},{2},(3} }
(1/3)w(1},((1} ' (2},(3}}

(7)

In a corollary, we note that a stronger version of Value Axiom 2 is true. Given Q E PT and w E R ECL, we say that w is Q-decomposable iff:

V(S,Q)EECL, w~=

Z w~ SEQ S(3S,QAQ "

(Recall wr

= 0.)

Corollary 1 : If w E R ECL is Q-decomposable, then, for

anySEQ, I; 9 n ( w ) = w s , ( 2 . nES

A second corollary notes that our value is a consistent extension of the Shapley value.

Corollary 2: Suppose w E R ECL and v E R CL satisfy Ws, Q = Vs,

V (S, Q) E ECL. Then ~o (w) = ~0(v), where ~o :R cL ~ R N is the Shapley value operator.

Example 9 Suppose w (2,3},((1},(2,3)} = 1, w (1),({1},(2,3}} = - 1, and all other embedded coalitions have wealth zero, in a partition function game w on the set of playersN = (1, 2, 3}. The idea is that every coalition earns wealth zero if players 2 and 3 are separated, but if 2 and 3 are grouped together then they cooperate to earn one unit of wealth at a cost of one unit to the coalition containing player 1 (so w(1,2,3}, ((1,2,3 }} = 1 + ( - 1 ) = 0). Our value function prescribes O1 (w) = - 1, O2 (w) = 1/2 = O3 (w). This is a reasonable result, because this game is ({1), (2, 3}} -- decomposable and ~52 = 9 3 = 1/2 is the reasonable allocation for two players who earn zero apart but earn one unit wealth together.

Values of Games in Partition Function Form

27

The Shapley value for this game depends on how we try to reduce the partition function game to a characteristic function game. If we use the rule v S = Ws, (S,N\S) then the Shapley value is r (v) = ( - 2/3, 1/3, 1/3). If we use the rule Vs - W s , (s) u { {n) In~ S) then the Shapley value is r (v) = ( - 1/3, 1/6, 1/6). And if we use the rule v S = m~n Ws, Q, then the Shapley value is ~0 (v) = ( - 2/3, 1/3, 1/3). All of these values would be inconsistent with our insight that the game is decomposable, with no effective cooperation between player 1 and the other two players.

eroofs
Theorem 1 : For any (S, Q) E ECL and any real number t, define the partition function game x t,S,Q by
: t, if (S, Q) >> (S, Q)

(xt,S,O)~ ~ = S,Q
= 0, otherwise. (Recallequation (5).) Observe that ~r o x t'S'Q = x t'~r(S'Q), because (S, ~)) >> (S, Q)

iff~r (S, a) >> ~ (S, a).

Suppose (n, m} _CS E Q. Let n : N ~ N switch n and m, leaving all other players fixed. Then Axiom 1 requires that cbn (x t,S,a) = ~n(n)(rr o x t,S,Q) = dpm (x t,n(S,Q)) = cbm (xt, S~Q), Also, i f S E Q, then both S and S U S are carriers o f x t'S'Q, so Axiom 2 requires that:
A

= t, i f S = S;

dPn (x t,S,Q) --_ nES

^

= O, for any other S C Q.

Thus Value Axioms 1 and 2 require that

= t/ISI, i f n E S ; dPn (X t ' S ' Q ) = {

=0, ifn~S. Note that, in terms o f R ECL as a real vector space, x t'S,Q = t 9 x I'S'Q. So (t 9 x l , S , Q ) = cb (xt,S,O) = t 9 cb (xl,S,O). For simplicity, henceforth let xS, Q = x 1,S,Q. The next claim is that the collection (x S,Q I (S, Q) E ECL} forms a basis f o r R EEL considered as a real vector space. There are [ECL[ games in the collection, and dim (R ECL) = IECLI. So it suffices to show that the games in the collection are linearly independent.

28

R.B. Myerson

Suppose that there were some sequence of coefficients rS ,~, not all zero, such that E 9 x S,Q = 0 in R ECL. Let (S, Q) be any embedded coalition such that (S,Q)~ECL rS'o IQI -- max (IQI t rs,Q ~ 0} and r~ ~ 4= 0. Consider now that X r . ' (S,Q)~ECL S,Q 9 t~.S,Q'l^ ,, = 0. But (xS'Q)~,~ = 0 unless (S, ". Q) ^ >> (S, Q), and rs, o = 0 unless JS,Q
2~

IQI ~ IQI. There is only one embedded coalition (S, Q) which satisfies both (S, Q) >> (S, Q) and tQI ~> IQI, and that is ~,S, Q) = (S, Q). we can thus conclude that roS( ^ = 0, a contradiction of the way (S, Q) was chosen. Therefore the collection gQamesx S'Q is linearly independent, and forms a basis for R EC L. By induction in k, the number of summands, it easily follows from Axiom 3 that k k-1 k-1 k qb(i=lX wi) = $ ( ~ l Wi + wk) = cb ( Zi:l Wi) + dp (wk) = i= ~ l cb (wi)" Now for any w E R ECL, there is a unique collection of coefficients bs, Q such that w = ~ 9 x S,Q because the x S,Q games form a basis. But then we must (S,Q)~EcLbS,Q have 9 (w) = 9 ( (s,a)bs, ~ a 9xS,Q) = ~ d# (bs,Q 9 x S,Q) = ~ bs, Q . dp (xS,O). (s,a) (s,a) This formula, together with

(xS'Q) = / 1/iS[, i f n E S ,
~n

L0,

i f n ~ S,

determines the unique function r which can satisfy the three Axioms. To show that this value function actually satisfies the Axioms, henceforth let r be the function defined in the above paragraph. Because r is a linear map, it must satisfy Axiom 3. To prove Axiom 1, consider any permutation lr : N - + N and any n EN. Define the function f : R ECL -~R b y f ( w ) = Cn (w) - ~r(n) (Tr o w). T h e n f (.) is a linear function, because r (.) and Or o.) are linear. For any (S, Q) E ECL, cbn (x S,Q) = =r (x~r(S,Q)), because n E S iff n (n) E ~r (S). S o f ( x s,Q) = 0, because x~(S,Q) = 7r o x S'Q. But these x S'Q games spanR ECL, so f ( w ) = 0 for any w. To prove Axiom 2, consider any S E CL. Define the functiong : R ECL ~ R ECL by g-~,O (w) = W~ns,OA (S,N\S)" Notice that g is linear, and } =xS'Q, if(S, { S , N \ S ) ) > > ( ~ Q); t

__

g(xS,O) =

= 0, otherwise.
r e _

Thus games of the formx s'Q such that (S, {S, N \ S ) ) >>(S, Q) span the range ofg.

Values of Games in Partition Function F o r m

29

But for any game of this form, 2; ~ (xS'Q)=

nES n

2;_1
n~S

t lSt

= 1 =(xS'Q) N,{N}.

So by linearity o f ~ , if w is in the range ofg then n~SZCbn (w)=wN, (N}" But ifS is a carrier ofw then w = g ( w ) , so Z q~n (w) = WN, {N} follows.
nES

Theorem 2: For any finite set K, let PT (K) be the set of partitions of K. We will use
the following combinatorial fact: (--1) IPl-1 ( I P I - 1 ) ! = / 1 if IKI = 1

Z
PEPT(K)

to

if I K [ > I ,

(If I K I = 1, then the unique partition has I P [ = 1, so the equation is obvious.) (If i f I > 1, select any k E K, and observe that K \ (k} r 0. For each partition P E PT (K \^(k}), we can make I P I different partitions of K by adding k to any of the t/31 sets in P, and we can make one other distinct partition of K by putting k alone in (K}. Each P E PT (K) is made from exactly one P E PT (K \ (k}) by one of these procedures. Therefore E ( - 1) IPL-1 (l P I - 1)! =
eePT(K~ ^

=^P~PT(K2:\(k}) ((. 1)IPI-I(IP[ .

= ^ Y, P ~ P T ( K ' , (k})

1)r . 1 / 3 1 + ( -.- 1 ) I P J + l - l ( I / 3 1 + l

1)!)

(--

1) IPI'I

(I/31!-I/31 !)=03

The value function of Theorem 1 was characterized in the proof as being the unique linear map (P :R ECL --->R N such that __ [ 1/I~(I ifn ES,

cbn (x S,Q) =
0 ifn ~ ( ,

(*)

__ [1 if (S, Q) ~ (S, Q) where (x s'a) s,a =

0 otherwise.

So it suffices to show that the formula in Theorem 2 defines a function which satisfies these conditions as well. The function is obviously linear in w, so it remains to show that (*) is satisfied for every (S, Q) E ECL andevery n EN. Consider any n E N and (S, Q) E ECL. Let S be the coalition such that n E/~ E Q. Then the formula in Theorem 2 implies that:

30

R.B. Myerson

dO n (XS'Q) =

Y" - - (-- 1) IQII. (S,Q)>>(S,Q)

(I Q I -

1)!

iI
~
/

SeQ ~nqS

(IQ I- 1)(INI- iSi)

(-- l ) IPII

(IPI-

1)!

P~PT(Q)

I T~P(IPI--I)(INI--I
~s~r
sct T
1

S~T

u Sli) =
A

Tc_Q\(S,S}

INI-I SeT u S
1

( ~_ ( - 1 ) Le~-~( I P l - 1 ) ! ) = P~PT(Q\T)

IN[-I 0

s~Q\ (Y)

t~

S[

ifS=S

ifff~:S

{
n~S

1/ISl

ifn e S ,
ifn E S .

Corollary 1 : We are given a partition Q, and a Q-decomposable partition function garne w. For eachS E Q , define wS E R ECL by (wS)~',~) = W~nS,~.AO. Then w = ,,~ wS, SEQ
^

by Q-decomposability. For any S and S in Q, both S and S U S are carriers of w S, so

(wS)N, (N} = wS,Q' ifS = S;

^

don (wS) =
0, for any other S E Q . ~ don(w) = l~ do ( w S ) = w s , a. nES nES n Thus, f o r a n y S E Q ,

Corollary 2: Define the linear map 0 :R CL -+R ECL by OS, Q (v) = vS for any v E R CL and (S, Q) E ECL. Then our Value Axioms 1, 2, and 3 for do (.) directly imply the Shapley value axioms for do (0(-)). So by uniqueness of the Shapley value for these axioms [Shapley], ~ (v) = do (0 (v)).

Values of Games in Partition Function Form References

31

Lucas, W.F., and R.M. Thrall: n-Person Games in Partition Function Form. Naval Research Logistics Quarterly X, 1963, 281-298. Shapley, L,S. : A Value for n-Person Games, in Contributions to the Theory of Games I1, ed. by H.~r Kuhn andA.W. Tucker. Princeton 1953, 307-317.

Received September, 1976