Utilitarianism, Egalitarianism, and the Timing Effect in Social Choice Problems

Author(s): Roger B. Myerson

Source: Econometrica , Jul., 1981, Vol. 49, No. 4 (Jul., 1981), pp. 883-897
Published by: The Econometric Society

Stable URL: https://www.jstor.org/stable/1912508

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/1912508?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to
Econometrica

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 Econometrica, Vol. 49, No. 4 (July, 1981)

UTILITARIANISM, EGALITARIANISM, AND THE TIMING

EFFECT IN SOCIAL CHOICE PROBLEMS

BY ROGER B. MYERSON

Two theorems are derived about social choice functions, which are defined on compre-
hensive convex subsets of utility allocation space. Theorem 1 asserts that a linearity
condition, together with Pareto optimality, implies that a social choice function must be
utilitarian. Theorem 2 asserts that a concavity condition, together with Pareto optimality
and independence of irrelevant alternatives, implies that a social choice function must be
either utilitarian or egalitarian. These linearity and concavity conditions have natural
interpretations in terms of the timing of social welfare analysis (before or after the
resolution of uncertainties) and its impact on social choices.

1. INTRODUCTION

THE utilitarian principle in ethical theory asserts that the best social policy is the
one which gives the greatest total welfare to the individual members of society,
where "total welfare" is measured by summing utility numbers for all individuals.
This principle has been advocated by philosphers going back to Bentham, and
more recently by Harsanyi [5, 6] from the point of view of Bayesian decision
theory.
The egalitarian principle in ethical theory asserts that the best social policy is
the one which gives the greatest welfare subject to the constraint that all
individual members should enjoy equal benefits from society. As long as there is
any positive tradeoff between the utility payoffs to different individuals, this
egalitarian principle leads to the same social choices as the maximin principle,
which always maximizes the utility of the most unfortunate individuals in society.
Rawls [12, 13] has argued for the maximin principle in his theory of justice. The
egalitarian principle has been studied in a game-theoretic context by Kalai [7]
and Myerson [9].
As Shapley [18] has pointed out, these two ethical principles both use interper-
sonal comparisons of utility, but in very different ways. Translated into the
practical debates of daily life, the utilitarian principle asserts that "you should do
something for me if it will hurt you less than it will help me," whereas the
egalitarian principle asserts that "you should do something for me if you are
better off than I am (or if you have gained more from our cooperation than I
have)."
This paper will investigate some properties of social choice rules related to
these two principles, with the goal of helping to explain why these two principles
have been so important both in the development of ethical theories and in
practical social decision making. Other approaches to this same question have
been studied by Deschamps and Gevers [3, 4], Sen [17], D'Aspremont and
Gevers [2], Maskin [8], and Roberts [14]. Our approach will differ from these
others in that we will use the risk properties of von Neumann-Morgenstern utility
scales in an essential way. In this respect, we will follow Harsanyi [5, 6], who
883

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 884 ROGER B. MYERSON

derives utilitarianism from assumptions about social decision making in the

presence of risk.
We will analyze utilitarianism and egalitarianism with respect to a phenome-
non which we may call the timing effect in social welfare analysis. To understand
this timing effect, consider the following simple example. The -father of twin sons
wants to set them up in their careers. His financial resources allow him only two
alternatives: he may send both sons to school (for four years each) to become
teachers, or he may send one son to school (for eight years) to become a doctor
while the other son will have to become a clerk. Each son would prefer to be a
doctor over teacher and teacher over clerk; and each son would prefer a .50-.50
gamble between becoming a doctor or a clerk over becoming a teacher for sure.
The father is concerned only for the welfare of his two sons, and he wants to be
egalitarian. What should he do?
The problem is that the egalitarian principle leads to different conclusions
depending on when the welfare of the sons is evaluated. If we think about the
sons as they will be in their careers a few years from now, there appears to be
only one egalitarian option: both sons must be teachers. On the other hand, if we
think about the sons as they are today, then this plan of making both sons
teachers is not even Pareto optimal. Both sons would prefer today that the father
should plan to send one of them to medical school, letting a fair coin toss decide
who gets to become a doctor and who must be a clerk. Before the coin toss, this
"randomization" plan seems egalitarian and Pareto dominates the "both teach-
ers" plan; but after the coin toss the randomization plan will seem unegalitarian
(since one unfortunate son is a clerk) and the both teachers plan will be ex post
Pareto optimal.
The moral of this story is that simply specifying a social welfare function may
not be enough to fully determine a procedure for collective decision making. One
must also specify when the individuals' preferences or utility levels should be
evaluated; before or after the resolution of uncertainties. The timing of social
welfare analysis may make a difference.
The timing effect is often an issue in moral debates, as when people argue
about whether a social system should be judged with respect to its actual income
distribution or with respect to its distribution of economic opportunities. Those
who judge with respect to income are implicitly evaluating individuals' welfare at
a later point in time than do those who judge with respect to opportunities.
In this paper, we will show that utilitarian and egalitarian choice rules have
some special properties with respect to this timing-effect problem. In Section 2,
we will develop some basic definitions relating to social choice problems and the
choice functions which may be used to solve them. In Section 3, we will see that
the timing effect can be ignored if and only if the social choice function is linear,
and we will show that linearity and Pareto optimality together imply that a
choice function must be of a generalized utilitarian type. This result is closely
related to Harsanyi's Theorem V in [5], with our linearity condition playing the
same role as the sure-thing principle in Harsanyi's work.
In Section 4, we will show that, if the linearity condition from Section 3 is

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 885

replaced by a weaker concavity condition,

alternatives condition [10] is added, then our choice functions must be either
utilitarian or egalitarian. This concavity condition has a natural interpretation in
terms of the timing effect; it guarantees that the social choices generated by
current welfare analysis must ex ante Pareto dominate those choices generated
by future welfare analysis. (In our example, we saw that both sons would indeed
prefer ex ante the randomization plan, recommended by current egalitarian
analysis, over the both teachers plan, recommended by future egalitarian analy-
sis.) That is, concavity guarantees that all individuals should always want social
decisions to be planned using current expected utility as the data for welfare
analysis, so that the timing effect should never cause disputes over when to
evaluate social welfare. Perles and Maschler [11] have also used this concavity
property (which they call super-additivity) to characterize a new solution concept
for two-person bargaining games.

2. BASIC DEFINITIONS

In this paper, social choice problems are represented by the sets of feasible
utility allocations available to the society. We assume that there are n individuals,
numbered 1,2, . .. , n, in the group or society. We also assume that some
von Neumann-Morgenstern utility scale (assigning utility values to all imaginable
social situations) has been specified for each individual in the society. Thus, any
vector x = (xl, . . . , xn) in Rn can be interpreted as a utility allocation vector,
corresponding to some social situation or gamble in which each individual i gets
expected utility xi, as measured in his given utility scale.
We do not assume here that the given utility scales have any particular
significance for interpersonal comparisons. Of course, any increasing affine
transformations of the given utility scales would also be valid von Neumann-
Morgenstern utility scales for the individuals, with identical decision-theoretic
properties.
Given any vectors x and y in ER, we write x > y (or y < x) iff xi > yi for every
i = 1, 2, . .. , n. Similarly x > y (or y < x) means that xi > yi for every i.
The usual dot product is used for vectors in Rn, that is:

x * y = ExjyCi
i=l1

A set S C Rn is comprehensive iff: y < x and x E S together imply that y E S.

That is, a comprehensive set describes a choice situation in which free disposal of
any individual's utility is always possible.
A set S C Rn is convex iff: x E S andy E S and 0 < X < 1 together imply that
Xx + (1 - X)y E S, where Xx + (1 - X)y is the vector whose ith component is
Xxi + (1 - X)yi. Since we are measuring utilities in von Neumann-Morgenstern
scales, if the group can always plan to randomize between any two collective
choice options, then the set of feasible expected utility allocations will be convex.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 886 ROGER B. MYERSON

Given any finite collection of vectors { xi, xk} C R' (so that each xi
= (xi,x&, ... xnJ) is a vector in llRn), we define H(x . , xk) to be the smallest
convex and comprehensive set containing the set {xl, .. ., x k}. That is:

yE D Rn there exist numbers AX, Ak such that

H(x1, . . .,xk). ] | everyA >O, 0 R = l,and

kF

y < J AxJ

We call H(x', . . , x'k) the comprehensive-convex hull of {x , . . . k, x

In this paper, a choice problem is formally defined to be a nonempty, closed,
convex, and comprehensive subset of R', representing the set of feasible utility
allocations. That is, we shall always assume that randomized strategies and free
disposal of utility are possible in every choice problem.
Throughout, we let CP denote the set of choice problems to be studied. Also,
we define CP' to be the class of all choice problems which can be generated as
the comprehensive convex hulls of nonempty finite sets of allocations. That is:

CPO = (fH(x', . . ., Xk)I (XI, . . . , Xk'} is a finite subset of Rn}.

All the results which we derive will hold for the case of CP = CPO, but we ma
allow CP to represent more general classes of convex comprehensive sets as well.
For any two sets S C llR and T C R' and any number A, we define AS
+ (1 - A)T to be the set:

XS + ( -)T = {Ax + (1 -A)ylx e S andy e T}.

If 0 < X < 1, then we can interpret this set as follows. Suppose there is some
random variable which may take the value 0, with probability A, or may take the
value 1, with probability 1 - A. Suppose further that the individuals know that
they will learn the random variable's true value tomorrow; if the value is 0 then
the group will get a choice problem with feasible set S, and if the value is 1 then
the feasible set will be T. Now suppose the group decides to plan its choices
today, before learning the random variable. Then XS + (1 - X) T is the set of
expected utility allocations which can be generated by making such conditional
plans today.
Given any collection of choice problems CP, we define a choice function to be
a mapping F: CP -E Rn such that, for every S in CP:

F(S) = (F1(S), . . ., Fn(S)) E S.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 887

That is, a choice function should select a feasible utility allocation vector for
every choice problem. In the rest of this paper, we will study various properties
which we might want a choice function to satisfy, and we will characterize the
classes of choice functions which can satisfy these properties.

3. LINEAR AND UTILITARIAN CHOICE FUNCTIONS

In this section, we consider choice functions satisfying weak Pareto optimality

and linearity properties.
A function F: CP -* R' is weakly Pareto optimal (WPO) iff: (i) F(S) E S, and
(ii) x > F(S) implies x E S, for every S in CP and every x in llR'. That is, a
choice function is weakly Pareto optimal if it always chooses a feasible utility
allocation such that there is no other feasible allocation making everyone strictly
better off. If a choice function were not weakly Pareto optimal, then it could be
criticized as inefficient, in those cases where it chose a strictly dominated point.
A function F: CP - AR' is linear iff:

F(QS + (1 - X)T) = XF(S) + (1 - X)F(T)

for every pair of choice problems S and T in CP, and for every number X such
that 0 ? X < 1 and XS + (1 - X)T e CP.
The linearity property has a natural interpretation in terms of timing of social
choices. Suppose that a social decision will have to be made tomorrow, at which
time the set of feasible utility allocations either will be S, with probability X, or
will be T, with probability 1 - X. If F will be applied to the choice problem
tomorrow, then the chosen utility allocation will either be F(S), with probability
X, or F(T), with probability (1 - A). So the expected utility allocation (as
assessed today) is AF(S) + (1 - X)F(T), if the social choices are to be made
tomorrow. On the other hand, XS + (1 - A) T is the set of all expected utility
allocations which are now feasible by planning tomorrow's decisions today (using
contingency plans which may depend on the information to be learned tomor-
row). If F is applied to the choice problem on the planning level today, then
F(XS + (1 - X)T) should be the chosen utility allocation. Thus, a linear choice
function is one for which every individual can expect the same utility from a
social choice whether it is planned ahead today or made on a situational basis
tomorrow.
We say that a function F: CP -* R' is utilitarian iff there exists some vector
P = (pl,... p) in R;In such that: (i) EI pi = 1 and every pi > 0, and
(ii)p * F(S) = maximumx,Es p * x, for every S E CP. That is, a utilitarian
function is one which always maximizes some weighted average of the individu-
als' utilities.
One might argue that perhaps we should call these "weighted utilitarian"
choice functions, and that we should call a choice function "properly utilitarian"
only in the case of PI = P2 = ... = pn. However any choice function which is
utilitarian in our above sense could be represented as properly utilitarian if all

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 888 ROGER B. MYERSON

individuals with pi = 0 were droppe

als' utility scales were multiplied by
functions are precisely those which could be represented as properly utilitarian
for some selection of von Neumann-Morgenstern utility scales for the individu-
als, possibly after some restriction of membership in the society.
We say that CP is a convex collection of choice problems iff AS + (1 - A) T
C CP whenever S E CP, T E CP, and 0 < X < 1. With this regularity condition
on the set of choice problems considered, we can state our first main result.

THEOREM 1: Suppose CP is a convex collection of choice problems, and suppose

F:CCP -* Wn is a linear and weakly Pareto optimal choice function. Then F is
utilitarian.

PROOF: Let A = {p E R'l Z =Ipi, all pi > 0). For any S in CP, define the s
Q(S) = {q ED n for some x E S, q * x > q * F(S)}.

Notice that Q(S) must be an open subset of En.

If F is not utilitarian, then we must have A C Us cpQ(S), so the Q(S) sets
must form an open cover on A. Since A is compact, there must exist some finite
collection {S1,S2, . Sk, S } CP such that A c U>k=I Q(Si). Then consider
= (I/k)> Si. By WPO, F(S0) must be on the Pareto frontier of the closed
convex set So, so by the Supporting Hyperplane Theorem (see Rockafellar [15,
Section 11]), there must exist some vectorp El A such thatp * F(S0) > p * x for all
x in So. However, since the Q(Si) cover A, we may assume (renumbering if
necessary) that p E Q(S 1). So for some xl in S', p * x' > p F(S 1). Then by
linearity, we get

p F( SO) =P. ( k2F(Si) = k( P* F(Si)

k p

< (P
k + p2 - F(Si)
~~j=2

= P * x + E F(Si ))
But

k/ k

x EF(SI = J=S0
k ( j=2 ) k j=_1
so this inequality implies p E Q(S), a contradiction of the way p was con-
structed. To avoid this contradiction, we must conclude that F is utilitarian.
Q.E.D.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 889

4. CONCAVE AND EGALITARIAN CHOICE FUNCTIONS

Let E: CP -* Rn be the choice function which always selects the highest

feasible allocation giving all individuals equal utility. That is, E is weakly Pareto
optimal and satisfies: EI(S) = E2(S) = * = En(S) for every choice problem
S. Since all our choice problems are closed comprehensive sets, there will always
be a unique point on the weakly Pareto optimal frontier which satisfies this
equity condition, so E is well-defined.
The choice function E is consistent with the maximin principle recommended
by Rawls [13], in that:

minimum (Ei (S)) = maximum (minimum (xi)).

(If not, then there would exist some x in S such that xj > miniEi(S) = Ej(S) for
everyj, which would contradict the Pareto optimality of E(S).)
By generalizing the equity constraints which define E, we can introduce the
class of egalitarian choice functions. We say that F: CP -* lR is egalitarian iff F
is weakly Pareto optimal and there exist numbers u,1, U2. ..., un and cl > 0,
c2 > O0 . , cn > 0 such that, for every S in CP:

F1(S)-u1 F2(S)-U2 n(S)-_Un

Cl=,C2
. . .Cn

That is, when u

tion F always selects the best point in S of the form

F(S) = u + ac,

for some number a. Since we consider only closed and comprehensive sets, there
will always be a unique weakly Pareto optimal point in S on the line { u + ac j a
E 1R)}. So the egalitarian choice function is well-defined once the vectors u and
c > 0 are specified.
As before, one might argue that we should call these "rescaled egalitarian"
choice functions, reserving the term "properly egalitarian" for the choice func-
tion E above. However, any choice function which is egalitarian in our
above sense could also be represented as properly egalitarian if we used the
appropriate affine transformations of the individuals' utility scales (mapping
(xi - ui)/ci). That is, our egalitarian choice functions are precisely those w
could be represented as properly egalitarian under some selection of von Neu-
mann-Morgenstem utility scales for the individuals.
An egalitarian choice function like E is generally neither linear nor utili-
tarian. For example, (letting n = 2) suppose that S = H((4, 4), (0, 10)) and T =
H((4,4),(10,0)). Then E(S) = (4,4) and E(T) = (4,4). But L S + - T=
H((4, 4), (7,2), (2,7), (5,5)). So we get:

E( S + T) = (5,5) #- (4' 4) E(S) + } E(T).

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
890 ROGER B. MYERSON

Although egalitarian choice functions do not satisfy linearity, they do satisfy a

weaker property.
A choice function F: CP -* R' is concave iff:

F(AS + (1 - X)T) > XF(S) + (1 - X)F(T)

for every pair of choice problems S and T in CP, and for every number X such
that 0 < X < 1 and XS + (1 -)T E CP. (Notice that the formula above is a
vector inequality in Rl, meaning that the inequality holds in every component.)
In the last section we saw that a linear choice function is one for which every
individual can expect the same utility from a social choice whether it is planned
ahead "today" or made on a situational basis "tomorrow." In these terms, a
concave choice function is one for which every individual's expected utility from
planning ahead (F1(XS + (1 - X) T)) is always greater than or equal to his
expected utility from situational judgments (XFi(S) + (1 - X)Fi(T)). So, whe
concave choice function is used, the timing of social choices can make a
difference; but timing would never be a cause for dispute, because all individuals
would agree that earlier (planned-ahead) choices yield better expected outcomes.
For illustration, let us return to the numerical example given above, with
S = H((4,4), (0, 10)) and T = H((4,4), (10,0)). Suppose that a fair coin is about
to be tossed: if it comes up heads then the two individuals will be offered the
choice problem S, and if it comes up tails, then they will be offered the choice
problem T. Applying the choice function E after the coin toss will yield the
equitable (and ex post Pareto optimal) allocation (4,4), no matter how the coin
may fall. But if the individuals plan their group choice before the coin is tossed,
then the choice function E selects the equitable (and ex ante Pareto optimal)
allocation (5, 5), which is implemented by planning to take (0, 10) if heads and
(10,0) if tails. Thus, we have the concavity relation:

E( 2 S + 2 T) = (55) > (4 4) = 2 E(S) + 1 E(T)

So before the coin is tossed, there is no dispute about whether to plan the group
choices immediately or to wait until after the coin is tossed; both want to plan
the group choices immediately.
Of course any linear utilitarian choice function must also be concave, since
linearity implies concavity trivially. Furthermore, any egalitarian choice function
must be concave. To prove this fact, suppose that:

F(S) = u + ac, F(T) = u + 1c,

and
F(XS + (1 -X)T) = u + yc.

Since

aF(S) + (1 - X)F(T) = u + (om + (I-A))c En XS + (1 - X) T,

and since F(AS + (I1-AX)T) must be undominated in XS + (I1-AX)T, we must

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 891

have:

y> Xa + (1 - X),

and so:

F(XS + (1 - X)T) > XF(S) + (1 - X)F(T).

There are concave choice functions which are neither utilitarian nor egalitar-
ian. For example, let n = 2, and let F' and F" be linear utilitarian choice
functions such that F'(S) maximizes 3- X1 + 32 X2 over x E S, and
mizes 32 XI + 3 X2 over x E S. Then let

F"'(S) = I F'(S) + { F"(S)

(so that F"' is linear but not weakly Pareto optimal), and let F(S) = (F1(S),
F2(S)) be the unique point on the weakly Pareto optimal frontier of S such that:

F1(S) - F".(S) = F2(S) -F..(S).

It can be checked that the F: CP -* fR 2 thus defined is a concave and weakly
Pareto optimal choice function, but it is not utilitarian or egalitarian. However,
this choice function may also seem rather complicated, in that it depends on
F'(S) and F"(S), two alternatives which are not actually selected. That is, this
choice function violates a principle of independence of irrelevant alternatives.
Following Nash [10], we say that a choice function is independent of irrelevant
alternatives (IIA) iff, for any two choice problems S and T in CP: if S C T and
F(T) E S then F(S) = F(T). (Sen [16] calls this condition "Property a", to
distinguish it from the condition of the same name used by Arrow [1].) To
interpret this condition, suppose that society faces choice problem T and selects
F(T). Now suppose that it is discovered that some points in T really are not
feasible, and so the actual feasible set is S 5 T. Independence of irrelevant
alternatives requires that, if the old choice is still feasible (F(T) E S), then it
should still be chosen. Any choice function determined by maxmization of a
social welfare function, or by a social preference ordering over utility allocations,
will satisfy independence of irrelevant alternatives.
Recall that CPF is the set of all choice problems which can be generated as
comprehensive convex hulls of finite sets of points in R . If we impose the
regularity condition that CP D CPO, then we can state our second main res
characterizing utilitarianism and egalitarianism.

THEOREM 2: Suppose that CP D CPO, and suppose that F: CP -> fR is weakly

Pareto optimal, concave, and independent of irrelevant alternatives. Then F is either
utilitarian or egalitarian.

PROOF OF THEOREM 2: The proof will be broken up into a series of lemmas

and definitions.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
892 ROGER B. MYERSON

DEFINITION: For any allocations x and y in RO, let f(x) = F(H(x)) and let
f(x, y) = F(H(x, y)).

DEFINITION: Let M be the set of all allocations in R' which F could possibly
select; that is:

M = {xlx = F(S) for some S E CP}.

LEMMA 1: M= {xIx E R' and x = f(x)}.

PROOF: If x = f(x) = F(H(x)) then x E M, for S = H(x). On the other hand,

if x = F(S) for any S, then H(x) C S by WPO, and so x = F(H(x)) by IIA.

LEMMA 2: M is convex.

PROOF: Suppose x = f(x), y = f(y), 0 < X < 1, and z = Xx + (1 - X)y. Then

H(z) = XH(x) + (1 - X)H(y), so

f(z) = F(H(z)) > XF(H(x)) + (1 - X)F(H(y))

= Xx + (1 -X)y = Z.

So f(z) > z by concavity. But f(z) < z by WPO. So f(z) = z, and z E M.

DEFINITION: Let CP* be the set of all choice problems generated by utility
allocations in M. That is:

CP* = { H(x, y, . . ., z)Ix E M,y E M, ... , z E M)}.

LEMMA 3: If S E CP* and T E CP* and O < X < 1, then

F(XS + (1 - X)T) = XF(S) + (1 - X)F(T).

PROOF: Let x = F(S), y =F(T). Choose w E S n M and z E S n M so that

Xw + (1 -X)z > F(S + (1 -X)T).

(This can be done because, by convexity of M, S E CP* must be generated from
S n M by disposal of utility only. Furthermore, by feasibility, there must be
some w* E S and z* E T such that

Xw* + (1 -X)z* = F(XS + (1 -X)T).

Then choose w e S n M and z e T n M so that w> w* and z> z*.) Then by
IIA and convexity of M, we must have

Xw + (1 -X)z = F(XS + (1 -X)T).

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 893

By IIA, we have x = f(x, w), since H(x, w) 5 S. Also:

Xw + (1 -X)z = f(x + (I -X)z,Xw + ( -X)z)

by IIA from XS + (1 - X) T. But z E M implies that z = f(z). So by concavity

Xw + (1 -X)z > Xf(x, w) + (1 -X)f(z) = x + (1 -X)z.

Since X >0, we get w > x. So x E H(w) and x =f(w) by IIA. But w E M, so
x = w.
A similar argument shows that y = z. Thus

x + (1 -X)y = F(XS+ (1 -X)T).

LEMMA 4: There exists some vector p E R' such that pi > O for
and: for all S E CP*, and all x E S, p* x ? p* F(S).

PROOF: F is linear on CP* by Lemma 3. CP* is a convex collection of sets,

because M is convex. So by Theorem 1, F is utilitarian on CP*.

DEFINITION: Henceforth let p be the vector constructed in Lemma 4.

LEMMA 5: For any choice problem S E CP, for any x E S n M, p * F(S) > p*
x.

PROOF: Let y = F(S). Then y E M and x E M, so H(x, y) E CP*. By IIA,

F(H(x, y)) = y. So p * y > p * x by Lemma 4.

LEMMA 6: Suppose p * x > p * f(y) and x E M. Then x = f(x, y). (Notice that,
since p > 9 and y > f(y), the hypothesis of this lemma will hold if p * x > p * y a
x E M.)

PROOF: Let z = f(x, y). For some X such that 0 < X < 1, we have z < Xy
+ (1 - X)x. Our goal is to show that X = 0.
If X = 1 then we would have z < y, implying z = f(y) by IIA. But Lemma 5
requires p * z > p * x, which would contradict the hypothesis if z = f(y).
So we know X < 1. By IIA we know z = f(x, Xy + (1 - X)x). But by concavity:

f(x,Xy + (1-X)x) > Xf(x, y) + (1-X)f(x) = Xz + (1-X)x.

So z > Xz + (1 -X)x, and thus z > x, since 1- > 0. So Xy + (1 -X)x > z > x.
This implies that either X = 0 or y > x. But y > x is not possible, by Lemma 5
and the hypothesis of this lemma. So X = 0 and z < x. Therefore z = x.

LEMMA 7: Suppose x E M, y E M, X E , and z = x + (1 -X)y. Then p * f(z)

=p* z.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 894 ROGER B. MYERSON

PROOF: If 0 < X < 1, then this lemma follows trivially from Lemma 2.
Consider now the case X > 1. Suppose that, contrary to the lemma, for some
,E p * z -p * f(z) > e > O. Let w = f(z) +,El, where 1 = (1, . . . , l) E- Rln.
If f(z, w) were in H(z) then we would have f(z, w) = f(z) by IIA. But this
would violate WPO, since w > f(z). So z # f(z, w).
By construction, p z > p w = p * f(z) + E. (Notice p * 1 = 1.) Then by
Lemma 6,

f(x, w+ j Y)=x, since x= z + 1y.

(Remember X > 1 is assumed here.) But by concavity, we must have:

x = f(x w+ Y) > f(zw)+ w f(y)

= A (Z, W) + A; fyY)

Multiplying through by X, we get Xx + (1 - X)y > f(z, w), a contradiction of the

conclusion in the last paragraph. Thus our assumption of p * z > p * f(z) must be
impossible, and so p . z = p * f(z) if X > 1.
For the case of X < 0, the proof is similar to the X > 1 case, reversing the roles
of x andy, and of X and (1 - X).

LEMMA 8: Suppose that F is not utilitarian. Then there exist allocations u and v
in R such that u = f(v) and p * v > p * u. Also, there exists a vector c E Rn such
that c > O,p * c = 1, and u - c E M.

PROOF: If F is not utilitarian, then we can find some S E CP such that

supwvsp. w > p F(S). So we can choose some v E S such that p vp
F(S). Now f(v) < v so f(v) E S n M. By Lemma 5, p * F(S) > p f(v). So
p * v > p u, when we let u = f(v).
To construct c, first let d = u -f(u -1). We know u-d = f(u -1) E M.
Also, u - 1 > f(u - 1), so d > 1 > 9 andp * d > p . 1 = 1. Let c = (I/p * d)d. So
p * c = 1, c > 9, and

u-c =( 1)(u d) (P d 1

is in M (by convexity of M).

ASSUMPTION: In Lemmas 9 through 12 we will assume that F is not utilitarian.

We will prove that F must be egalitarian.

DEFINITION: Henceforth, under the assumption that F is not utilitarian, let

u, v, and c be as in Lemma 8.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 895

LEMMA 9: Suppose x EM andp p v >p. x >p u. Then x > u.

PROOF: Let a = (p . v-p * u + l)/(p * v-p * x), and let y = v + a(x - v).
Notice that a > 1 and p * y = p * u-1 < p * u. Since u E M, by Lemma 6,
f(u,y)= u.
Let w = (I/a)u + ((a -1)/la)v. Then w < v, since u < v and a > 1. By
Lemma 6, f(v, x) = x, so by IIA we get f(w, x)= x. But

H(w,x) = Iaa
H(u,y) + a H(v).
So by concavity of F:

x =f(w,x) > I f(u,y) + a 1 f(v)= u+ lu= u.

LEMMA 10: Suppose x EM and p. x =p u. Then x > u.

PROOF: Suppose not. Then, for some i, xi < ui. Let y = f(u + c). Since
and u-cEM, Lemma 7 impliespy=p(u+c)=pu+ 1. ButyEM, so
for any positive E(O < E < 1) we have (1 - E)x + Ey e M. If we choose E small
enough, we get

(1 - E)X + EY, < uj

so (1 - E)x + 9y # u. But

p > V p * ((1 - E)X +' y) =p * u +E >p * u

as long as E is small and positive. So (1 - E)x + sy will violate Lemma 9 if x # u.
So we must have x > u.

LEMMA 11: For any vector d in Rn, if u - d E M then, for every X in R,

u + Xd E M.

PROOF: Consider first the case of X > 0. Let y = f(u + Xd). Then p y =p
(u+Xd) by Lemma 7. So

1 + - i\X (u -d) E=M

(by convexity of M) and

p (1 I y + 1 (u-d)) =p u.
By Lemma 10, we conclude

l I ;\Y+ 1 AX (u-d) > u.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
896 ROGER B. MYERSON

So

y > (1 + X)u-X(u-d) = u + Xd
Buty =f(u +Xd) impliesy < u +Xd. So u +Xd =y E M.
For the case of X < 0, observe that the preceding case showed u - (- d) = u +
d E M. So, using our results for positive multipliers, we get u + Xd = u +
(-X)(-d) E M since-A > O.

LEMMA 12: Suppose x E M. Then x = u + /3c, where /B = p * x -p * u.

PROOF: By Lemma 11 and the definition of c, u - /c E M. So I x + 1 (u-/3c)

e M. But

p* (4x + I(u - /3c)) = p* x + I p u - I 3p * c

= ( x + p* u - /3) = p * u.
So I x + I (u- 8c) > u by Lemma 10. Solving for x, we get x > u + /c.
Now lety = u + (u - x). By Lemma 11 we knowy E M and also u + 83c E M.
So Iy + I (u + /3c) E M. Observe that

p (?Y + I(U + /3c)) = -p u -4 /3 + Ip * u + /3 = p * u.

So Iy + I(u + /3c) > u by Lemma 10. Thus y > u-/8c, and so x = 2u-y
< u + 83c. Together with x > u + 83c from the preceding paragraph, this proves
the lemma.

Since Lemma 12, stating that F is egalitarian, was proven under the assump-
tion that F is not utilitarian, we have proven the Theorem. Q.E.D.

Northwestern University

Manuscript received June, 1979; revision received April, 1980.

REFERENCES

[1] ARROW, K. J.: Social Choice and Individual Values, 2nd ed.. New York: John Wiley and Sons,
1963.
[2] D'ASPREMONT, C., AND L. GEVERS: "Equity and the Informational Basis of Collective Choice,"
Review of Economic Studies, 44(1977), 199-209.
[3] DESCHAMPS, R., AND L. GEVERS: "Separability, Risk-Bearing, and Social Welfare Judgments,"
European Economic Review, 10(1977), 77-94.
[4] "Leximin and Utilitarian Rules: A Joint Characterization," Journal of Economic Theory,
17(1978), 143-163.
[5] HARSANYI, J. C.: "Cardinal Welfare, Individualistic Ethics, and Interpersonal Comparisons of
Utility," Journal of Political Economy, 63(1955), 309-321.
[6] "Nonlinear Social Welfare Functions," Theory and Decision, 6(1975), 311-332.
[7] KALAI, E.: "Proportional Solutions to Bargaining Situations: Interpersonal Utility Compari-
sons," Econometrica, 45(1977), 1623-1630.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms
 SOCIAL CHOICE PROBLEMS 897

[8] MASKIN, E.: "A Theorem on Utilitarianism," Review of Economic Studies, 45(1978), 93-96.
[9] MYERSON, R. B.: "Two-Person Bargaining Problems and Comparable Utility," Econometrica,
45(1977), 1631-1637.
[101 NASH, J.: "The Bargaining Problem," Econometrica, 18(1950), 155-162.
[11] PERLES, M. A., AND M. MASCHLER: "The Super-Additive Solution for the Nash Bargaining
Game," mimeographed, Institute of Mathematics, Hebrew University of Jerusalem, 1980.
[12] RAWLS, J.: A Theory of Justice. Cambridge: Harvard University Press, 1971.
[13] : "Some Reasons for the Maximin Criterion," American Economic Review, 64(1974),
141-146.
[14] ROBERTS, K. W. S.: "Interpersonal Comparability and Social Choice Theory," Review of
Economic Studies, 47(1980), 421-439.
[15] ROCKAFELLAR, R. T.: Convex Analysis. Princeton: Princeton University Press, 1970.
[16] SEN, A. K.: Collective Choice and Social Welfare. San Francisco: Holden-Day, 1970.
[17] : "On Weights and Measures: Informational Constraints in Social Welfare Analysis,"
Econometrica, 45(1977), 1539-1572.
[18] SHAPLEY, L. S.: "Utility Comparison and the Theory of Games," in La Decision. Paris: Edition
du Centre de la Recherche Scientifique, France, 1969, pp. 251-263.

This content downloaded from

154.59.124.153 on Wed, 08 Dec 2021 22:23:48 UTC
All use subject to https://about.jstor.org/terms