You are on page 1of 7

American Economic Association

Bayesian Decision Theory and Utilitarian Ethics

Author(s): John C. Harsanyi
Source: The American Economic Review, Vol. 68, No. 2, Papers and Proceedings of the
Ninetieth Annual Meeting of the American Economic Association (May, 1978), pp. 223-228
Published by: American Economic Association
Stable URL: .
Accessed: 14/08/2013 02:36

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .

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

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The
American Economic Review.

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions
Bayesian Decision Theory and Utilitarian Ethics

One of the great intellectual achievements similar economic and noneconomic vari-
of the twentieth century is the Bayesian ables affecting other individuals in the
theory of rational behavior under risk and society. Thus, mathematically, any situation
uncertainty. Many economists, however, are can be regarded as a point in a finite-dimen-
still unaware of how strong the case really is sional (say, r-dimensional) Euclidean space
for Bayesian theory, and many more fail to E'r
appreciate the far-reaching implications the In the case of risk and of uncertainty, an
Bayesian concept of rationality has for individual's choices can be modeled as
ethics and welfare economics. The purpose choices among different lotteries whose
of this paper is to argue that the Bayesian "prizes" are situations, that is, points in E'.
rationality postulates are absolutely ines- A lottery can be described as
capable criteria of rationality for policy de- L = (AI I|el . ** Ak I ek, . AK
(1) I eK)
cisions; and to point out that these Bayesian
rationality postulates, together with a indicating that this lottery L will yield
hardly controversial Pareto optimality re- prizes A1... AK, depending on which one
quirement, entail utilitarian ethics as a mat- of K mutually exclusive and exhaustive
ter of mathematical necessity. events el, ... eK occurs. These events el,
* . eK will be called conditioning events.
1. The BayesianRationalityPostulates Mathematically, any event ek (k = 1 ... ,
K) can be regarded as a measurable subset
In discussing the criteria for rational be- of the space S2of all possible "states of the
havior, I will distinguish behavior under world." A lottery L will be called a risky or
certainty, under risk, and under uncer- an uncertain lottery, depending on whether
tainty. Certainty obtains when we can pre- the decision maker does or does not know
dict the actual outcome of any action we the objective probability Pk = Prob (ek)
can take. Risk obtains when we know at associated with every event ek (k = 1.
least the objective probabilities associated K) used in this lottery L.
with alternative possible outcomes. Finally, In analyzing the behavior of any indi-
uncertainty obtains when even these ob- vidual i (i = 1 ... n), strict preference by
jective probabilities are partly or wholly him will be denoted by >i and nonstrict
unknown to us (or are possibly even unde- preference including indifference, or equiv-
fined). alence, by >?i
In the case of certainty, I will assume that Rational behavior by a given individual i
each individual chooses among various under certainty can be characterized by two
alternative situations, where each situation rationality postulates:
is characterized by finitely many economic 1. Complete preordering. The nonstrict
and noneconomic variables, such as his preferences of this individual i establish a
holdings of different commodities, includ- complete preordering over the space E' of
ing money, his health, his social position, all possible situations (or over some suitable
his social relationships, etc., as well as by closed subset of Er).
2. Continuity. Suppose that the sequence
*Professor of business administration and of eco- Al, A2 ... . of situations converges to a
nomics, University of California-Berkeley. I wish to particular situation AO, and that another
thank the National Science Foundation for supporting sequence B1, B2 .. - of situations con-
this research through grant SOC77-06394 to the Center
for Research in Management Science, University of verges to Bo, with Ak >i Bk for all k. Then,
California-Berkeley. AO >i BO.


This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions

For convenience I will call these two pos- 4. Sure-thing principle. Suppose that
tulates the basic utility axioms. Using these A* >i Ak fork = 1, ...,K. Then
two axioms, we can characterize rational
behavior as follows. (4) (A* Iel,..,AK eK)

>i (AI I el, -,_AK I eK)

THEOREM 1: Utility Maximization. If an In other words, other things being equal, a
individual'spreferences satisfy the two basic rational individual will not prefer a lottery
utility axioms, then his behavior will be equiv- yielding less desirable prizes over a lottery
alent to maximizing a well-defined (ordinal) yielding more desirable prizes. Note that
utility function.' (For proof, see Gerard the sure-thing principle is essentially identi-
Debreu, pp. 55-59.) cal with the game-theoretical principle that
a rational individual will avoid using any
To characterize rational behavior under (weakly or strongly) dominated strategy.
risk and under uncertainty, we need two Obviously, both postulates 3 and 4 are
additional rationality postulates: extremely compelling rationality require-
3. Probabilistic equivalence. Let ments. But they are subject to two qualifica-
L tions.
(2) = (AI eI ... SAK I eK)
(i) Both postulates presuppose that the
and utility U(A4k) of any prize Ak iS independent
of its conditioning event ek. This require-
L* = (A e
... AKe) ment can always be satisfied by suitable
and suppose that the decision maker knows definition of the prizes. For example, the
the objective probabilities associated with utility of an umbrella depends on whether
events e, . eK as well as with events it is raining or not (and whether there is a
e* ... IeK and knows that these probabili- heavy rain or a light rain). Therefore, it
ties satisfy would be inappropriate to make an um-
brella a prize of a lottery when the nature of
(3) Prob(ek) = Prob(e,*) the weather is the conditioning event;
for k = 1 . . . K rather, we must redefine the prize as staying
dry, or as getting slightly wet, or as getting
Then he will be indifferent between lotteries very wet since, as a rule, the utilities asso-
L and L*. ciated with these prizes can be assessed
In other words, a rational individual will without knowing the weather, etc.
be indifferent between two risky lotteries if (ii) More importantly, both postulates
these yield him the same prizes with the (and especially postulate 3) presuppose that
same probabilities even if the two lotteries the decision maker has no specific utility or
use quite different physical processes to gen- disutility for gambling as such, that is, for
erate these possibilities. Note that this pos- the nervous tension and the other psycho-
tulate implies von Neumann and Morgen- logical experiences directly connected with
stern's postulate on compound lotteries: gambling. In other words, the -two postu-
that a rational individual will be in- lates assume that the decision maker will
different between a two-stage lottery and a take a purely result-oriented attitude toward
one-stage lottery if both offer the same lotteries, and will derive all his utility and
prizes with the same probabilities. disutility from the prizes he may or may not
win through these lotteries, rather than
IFor some purposes it may be desirable to define from the act of gambling itself.
rational behavior without requiring that it should sat- Clearly, this assumption is seldom, if
isfy the continuity postulate (postulate 2). It can be
shown that if a given individual's preferences satisfy at
ever, satisfied in the case of gambling done
least the complete preordering postulate (postulate 1), primarily for entertainment. For example,
then his behavior will be equivalent to lexicographi- people who gamble in a casino will usually
cally maximizing a certain utility vector. do this because they are attracted by the

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions

nervous tension associated with gambling; ever, that for deriving the theorem, all we
and since the latter may strongly depend on need, apart from the two basic probability
the details of the physical process used to axioms, are the probabilistic equivalence
produce the relevant probabilities, they may postulate and the sure-thing principle, both
be far from indifferent to the nature of this of which represent absolutely compelling
physical process. (For example, they may rationality requirements for serious policy
not at all be indifferent between participat- decisions.
ing in a one-stage lottery and participating
in a probabilistically equivalent two-stage THEOREM 2: Expected-Utility. If an indi-
lottery.) vidual'spreferences satisfy postulates 1, 2, 3,
On the other hand, it is natural to expect and 4, then he will have a (cardinal) utility
that, in making important policy decisions, function Ui such that assigns, to any lottery L
responsible decision makers will take a re- ofform (1), a utility equal to its expected
sult-oriented attitude toward risk taking. utility, that is, equal to the quantity
This is probably a reasonably realistic de-
scriptive prediction; and it is certainly an K
obvious normative rationality requirement (5) Ui(L) = EPk Ui(Ak
as well as a moral requirement: responsible k= I
business executives using their shareholders'
money, and responsible political leaders where Pk (k - 1,...,K) is the probability
acting on behalf of their constituents, are associated with the conditioning event ek.
expected to do their utmost to achieve the More specifically, if L is a risky lottery, then
best possible results, rather than to gratify Pk must be interpreted as the objective prob-
their own personal desire for nervous ten- ability Pk = Prob(ek) of this event ek,
sion (or for avoiding nervous tension). Even whereas if L is an uncertain lottery, then Pk
clearer is the obligation of taking a purely must be interpreted as the subjective prob-
result-oriented attitude in making impor- ability Pk = Prob,*(ek) that the decision
tant moral decisions. maker chooses to assign to this event ek.
Thus, we can conclude that while postu-
lates 3 and 4 have little application to gam- Property (5) is called the expected-utility
bling done for entertainment, they are very property, and any utility function Ui pos-
basic rationality requirements for all serious sessing this property is called a von Neu-
policy decisions as well as for personal mann-Morgenstern utility function. For
moral decisions. short reference, I will call postulates 1, 2, 3,
and 4 the Bayesian rationality postulates.4
11. Expected-UtilityMaximization

The main conclusion of Bayesian theory will consistently act on the opinion that some events
is that a rational decision maker under risk are more likely to occur than not to occur, while other
events are more likely not to occur than to occur). In
and under uncertainty will act in such a way my own view, consistency in the use of qualitative or
as to maximize his expected utility; or, quantitative subjective probabilities should not be
equivalently, that he will assess the utility of assumed as an axiom, but rather should be inferred
any lottery to him as being equal to its ex- from some more basic-- and, one may hope, more
pected utility (expected-utility theorem). compelling---axioms. This is the approach taken by
F. J. Anscombe and Robert J. Aumann.
Different authors have used different 30wing to space limitations, the proof has been
axioms to derive this theorem, and some of omitted. But see my working paper.
these axioms had somewhat questionable 4As M. Hausner has shown, we can obtain a weaker
intuitive plausibility.2 It can be shown, how- form of the expected-utility theorem without using
postulate 2: it an individual's preferences satisfy at
2For example, Leonard Savagc's Postulate 4 directly least postulates 1, 3, and 4, then his behavior will be
assumcs that the decision makcr will act on the basis equivalent to lexicographically maximizing the ex-
of consistent qualitative subjective probabilities (i.e., he pected value of a certain utility vector.

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions

Ill. An Axiomatic Foundation a positive linear combination of all indi-

for Utilitarian Morality viduals' utility functions. I will assume that
society consists of n individuals. Consider
Now I propose to show that the Bayesian the social welfare function of individual j.
rationality postulates, together with a very The following three axioms will be used:
natural Pareto optimality requirement, log- Axiom a: Individual rationality. The
ically entail a utilitarian ethic. personal preferences of all n individuals sat-
According to Theorems I and 2, the be- isfy the four Bayesian rationality postulates.
havior of a rational individual i (i = 1 . . . , Axiom b. Rationality of moral prefer-
n) would be equivalent to that resulting ences. The moral preferences of individual j
from the maximization of the expected satisfy the four Bayesian rationality postu-
value of some cardinal utility function Ui lates.
expressing his personal preferences. In the Axiom c. Pareto optimality. Suppose
case of most individuals, these personal that at least one of the n individuals person-
preferences will not be completely selfish; ally prefers social situation A over social
but usually they will give greater weight to situation B, and that none of the other indi-
his own personal interests and to the in- viduals personally prefers B over A. Then,
terests of his family, friends, and other asso- individualj will morally prefer A over B.
ciates, than to the interests of complete Axiom a is an obvious rationality re-
strangers. quirement. So is Axiom b: it expresses the
Yet, there are situations where an indi- principle that an individual making a moral
vidual's behavior will not be guided by his value judgment must follow, if possible,
more or less self-centered personal prefer- even higher standards of rationality than
ences, but rather will be guided by much an individual merely pursuing his personal
more impartial and impersonal criteria. We interests. Thus, if rationality requires that
expect that judges and other public officials each individual should follow the Bayesian
will be guided in their official capacities by rationality postulates in his personal life as
some notions of public interest and of im- postulate I asserts, then he must even more
partial justice; and, more importantly, every persistently follow these rationality postu-
individual will have to be guided by certain lates when he is making moral value judg-
impartial and impersonal criteria when he is ments.5 While Axioms a and b are rational-
trying to make a moral value judgment. In- ity requirements, Axiom c is a moral
deed, by definition, any evaluative judg- principle-but it is surely a rather noncon-
ment based on biased, partial, and personal troversial moral principle.
criteria will not be a moral value judgment In view of Theorem 2, Axiom a implies
at all, but rather will be a mere judgment of that the personal preferences of each indi-
personal preference. vidual i can be represented by a von Neu-
I will describe the criteria guiding an indi- mann-Morgenstern (vN-M) utility function
vidual when he is honestly trying to make Ui, whereas Axiom b implies that the moral
an impartial and impersonal moral value preferences of individual j can be repre-
judgment as this individual's moral prefer- sented by a social welfare function Wjwhich
ences. likewise has the nature of a vN-M utility
In view of Theorems I and 2, if the moral function. Finally, the three axioms together
preferences of an individual i satisfy certain imply the following theorem.
consistency requirements, then his moral
value judgments will be such as if he tried to THEOREM 3: Linearity of the social wel-
maximize a special utility function ex-
pressing these moral preferences. This util-
ity function will be called his social welfare 5Axiom b, and, in particular, the assumption that
people's moral preferences should satisfy the sure-
function Wi. thing principle, was criticized by Peter Diamond. As I
I now propose to show that a rational have tried to show in my 1975 paper, his criticism is
individual's social welfare function must be invalid.

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions

farefunction. The social welfare function W accept somewhat stronger philosophical

of individualj must be a real-valued function commitments,then we can obtain a some-
over all social situations A, and must have the what stronger form of Theorem 3 (to the
mathematicalform effect that the social welfare function must
n be the arithmetic mean of all individual
(6) Wj(A) = ai Ui(A) utilities). What is more important, we can
achieve deeper philosophical insights into
with caistrictly positive f(r i = 1, . . n the nature of moral value judgments.
I have argued that moral value judgments
For the proof, see the author (1955, pp. must be based on impartial and impersonal
313-14).6 criteria. Now I propose to give a more spe-
Note that the proof of Theorem 3 does cific formal definition for this requirement
not assume the possibility of interpersonal of impartiality and of impersonality.
utility comparisons. The theorem will re- Suppose individual j expresses a value
main valid even if such comparisons are not judgment about the relative merits of one
admitted. Of course, if such comparisons possible social situation A as against an-
are ruled out, then the coefficients ai will other possible social situation B. How do
have to be based completely on individual we know whether he expresses a genuine
j's personal-and more or less arbitrary- moral value judgment, based on impartial
value judgments. and impersonal considerations? He would
On the other hand, if interpersonal utility certainly satisfy our impartiality and im-
comparisons (or at least interpersonal com- personality requirements if he did not know
parisons of utility differences) are admitted, how his choice between A and B would af-
then our three axioms can be supplemented fect him personally and, in particular, if he
by a fourth axiom: did not know what his own social position
Axiom d. Equal treatment of all indi- would be in situations A and B. More spe-
viduals. Individual j's social welfare func- cifically, let us assume he would think that
tion Wjwill assign equal weights to the util- in either situation he would have the same
ity functions Ul, . . . , Unof the n individuals probability I/n to occupy any one of the n
when these utility functions are expressed in possible social positions and,indeed,to be
equal utility units. put in the place of any one of the n indi-
Using this axiom, we can infer that in (6) viduals in the society. Then, he would
we must have clearly satisfy the impartiality and imper-
sonality requirements to the fullest possible
(7) a, an degree. I will call this assumption the equi-
probability model of moral value judg-
IV. The Equiprobability Model for ments.
Moral Value Judgments Obviously, this equiprobability model
cannot be taken literally. When individual j
The axiomatic analysis of Section III has makes a value judgment as to the relative
the advantage that it uses only extremely merits of situations A and B,he will often
weak philosophical assumptions: it derives have quite a good idea of the actual social
utilitarian ethics from two rationality re- position he would have in each situation;
quirements and one very natural moral re- and he will certainly know his own personal
quirement. However, if we are willing to identity. Nevertheless, his judgment as to
the relative merits of situations A and B will
61n view of Hausner's results, a weaker form of qualify as a genuine moral value judgment
Theorem 3 will remain true even if, in Axioms a and b, as long as he at least makes a serious at-
we redefine the Bayesian rationality postulates so as to tempt to disregard these morally irrelevant
omit postulate 2 (the continuity postulate). In this case
both the quantities Ui and the quantity W1will have to pieces of information in making this judg-
be reinterpreted as lexicographically ordered utility ment.
vectors. If we apply Theorem 2 to this equiprob-

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions

ability model, then we obtain the following tween Bayesian theory and utilitarian moral-
theorem. ity becomes even more obvious (Theorem 4).

THEOREM 4: The social welfare function

as an arithmetic mean of all individual utili- REFERENCES
ties. Suppose that individual j follows the
Bayesian rationality postulates. Then, under F. J. Anscombe and R. J. Aumann, "A Defini-
the equiprobability model, he will make his tion of Subjective Probability," Annals
moral value judgments in such a way as to Math. Statist., Mar. 1963, 34, 199-205.
maximize the social weljare fitic tion: Gerard Debreu, Theory of Value, New York
P. Diamond, "Cardinal Welfare, Individual-
(8) Wj(A) = - E Ui(A) istic Ethics, and Interpersonal Compari-
son of Utility: Comment," J. Polit. Econ.,
The theorem follows from the fact that Oct. 1967, 75, 765--66.
under the equiprobability model, j's ex- J. C. Harsanyi, "Cardinal Welfare, Indi-
pected utility will be given by the right-hand vidualistic Ethics, and Interpersonal
member of (8). Comparisons of Utility," J. Polit. Econ.,
Unlike Theorem 3, Theorem 4, as well as Aug. 1955, 63, 309-21.
the equiprobability model itself, does pre- , Nonlinear Social Welfare Func-
suppose the possibility of interpersonal tions," Theory Decn., 1975, 6, 311-32.
comparisons of utility differences (utility in- ,(1977a) "Morality and the Theory
crements). of Rational Behavior," Soc. Res., Winter
To sum up, we have found that the 1977, 44, 623- 56.
Bayesian rationality postulates, together , (1977b) "Bayesian Decision The-
with a Pareto optimality requirement, logi- ory," work. paper no. CP-404, Center
cally entail utilitarian ethics (Theorem 3)- Res. Manage. Sci., Univ. California-
even if interpersonal utility comparisons are Berkeley 1977.
not admitted. But, in actual fact, as I have M. Hausner, "Multidimensional Utilities,"
tried to show elsewhere (see the author, in Robert M. Thrall et al., eds., Decision
1977a), there are no valid arguments against Processes, New York 1954, 167-80.
such comparisons. Yet, once such compari- LeonardJ. Savage, The Foundations of Statis-
sons are admitted, the logical connection be- tics, Ncw York 1954.

This content downloaded from on Wed, 14 Aug 2013 02:36:49 AM

All use subject to JSTOR Terms and Conditions