Agreeing to Disagree

Author(s): Robert J. Aumann

Source: The Annals of Statistics, Vol. 4, No. 6 (Nov., 1976), pp. 1236-1239
Published by: Institute of Mathematical Statistics
Stable URL: https://www.jstor.org/stable/2958591
Accessed: 07-10-2018 05:19 UTC

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 [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and

extend access to The Annals of Statistics

This content downloaded from 119.148.3.107 on Sun, 07 Oct 2018 05:19:22 UTC
All use subject to https://about.jstor.org/terms
 The Annals of Statistics
1976, Vol. 4, No. 6, 1236-1239

AGREEING TO DISAGREE'

BY ROBERT J. AUMANN

Stanford University and the Hebrew University of Jerusalem

Two people, 1 and 2, are said to have common knowledge of an event
E if both know it, 1 knows that 2 knows it, 2 knows that 1 knows is, 1
knows that 2 knows that 1 knows it, and so on.

THEOREM. If two people have the same priors, and their posteriors for an
event A are common knowledge, then these posteriors are equal.

If two people have the same priors, and their posteriors for a given event A
are common knowledge, then these posteriors must be equal. This is so even
though they may base their posteriors on quite different information. In brief,
people with the same priors cannot agree to disagree.
We publish this observation with some diffidence, since once one has the ap-
propriate framework, it is mathematically trivial. Intuitively, though, it is not
quite obvious; and it is of some interest in areas in which people's beliefs about
each other's beliefs are of importance, such as game theory2 and the economics
of information.3 A "concrete" illustration that may clarify matters (and that
may be read at this point) is found at the end of the paper.
The key notion is that of "common knowledge." Call the two people 1 and
2. When we say that an event is "common knowledge," we mean more than
just that both 1 and 2 know it; we require also that 1 knows that 2 knows it
knows that 1 knows it, 1 knows that 2 knows that 1 knows it, and so on. For
example, if 1 and 2 are both present when the event happens and see each other
there, then the event becomes common knowledge. In our case, if 1 and 2 tell
each other their posteriors and trust each other, then the posteriors are common
knowledge. The result is not true if we merely assume that the persons know
each other's posteriors.

Formally, let (Q, M, p) be a probability space, 9q and 92 partitions

whose join4 2 v 2 consists of nonnull events.5 In the interpretation, (Q, M)
is the space of states of the world, p the common prior of 1 and 2, and !i the

Received November 1975.

1 This work was supported by National Science Foundation Grant SOC74-11446 at the Institute
for Mathematical Studies in the Social Sciences, Stanford University.
2 Cf. Harsanyi (1967-1968); also Aumann (1974), especially Section 9j (page 92), in which the
question answered here was originally raised.
I Cf. e.g., Radner (1968) and (1972); also the review by Grossman and Stiglitz (1976) and the
papers quoted there.
4Coarsest common refinement of x and a 2.
I Events whose (prior) probability does not vanish.
AMS 1970 subject classifications. Primary 62A15, 62C05; Secondary 90A05, 90D35.
Key words and phrases. Information, subjective probability, posterior, statistics, game theory,
revising probabilities, concensus, Harsanyi doctrine.

1236

This content downloaded from 119.148.3.107 on Sun, 07 Oct 2018 05:19:22 UTC
All use subject to https://about.jstor.org/terms
 AGREEING TO DISAGREE 1237

information partition of i; that is, if the true state of the world is w, then i is
informed of that element Pi(@) of .9 that contains w. Given Ct in Q, an event
E is called common knowledge at w if E includes that member of the meet6 91 A 9
that contains w. We will show below that this definition is equivalent to the
informal description given above.
Let A be an event, and let qi denote the posterior probability p(A I 9i) of
given i's information; i.e., if w e Q, theni qi((o) = p(A n Pi(o))1p(Pi(0))
PROPOSITION. Let w GE Q, and let q1 and q2 be numbers. If it is common knowle
at wo that q, = q, and q2 = q2, then q, = q2.
PROOF. Let P be the member of A 92 that contains w. Write P = uj
where the Pi are disjoint members of i1. Since q, = q, throughout P, we
p(A n Pi)/p(Pi) = q, for all j; hence p(A n Pi) = qlp(Pi), and so by summ
over j we get p(A n P) = q1p(P). Similarly p(A n P) = q2p(P), and so q, =
This completes the proof.

To see that the formal definition of "common knowledge" is equivalent to the

informal description, let 0) E Q, and call a member w' of Q reachable from wO if
there is a sequence Pl, P2, . . ., pk such that so e PI, (o e pk, and consecutive Pi
intersect and belong alternatively to .9 and .2. Suppose now that o is the true
state of the world, Pl = P1Qjo), and E is an event. To say that 1 "knows" E
means that E includes PI. To say that 1 knows that 2 knows E means that E
includes all PI in _2 that intersect Pl. To say that 1 knows that 2 knows that
1 knows E means that E includes all P3 in 9P that-intersect P2 in .92 that intersect
PI. And so on. Thus all sentences of the form "i knows that i' knows that i
knows ... E" (where i' = 3 -- i) are true if and only if E contains all co' reachable
from s). But the set of all Go' reachable from co is a member of .) A ,?2; so the
desired equivalence is established.
The result fails when people merely know each other's posteriors. Suppose
Q has 4 elements a, S3 r, o Of equal (prior) probability, 9 = {ca, r3l, 2 =
{aPr, a}, A = ad, and o = a. Then 1 knows that q2 is 3, and 2 knows that q
is 1; but 2 thinks that 1 may not know what q2 is (3 or 1).
2'3
Worthy Qf note is the implicit assumption that the information partitions p
and 2 are themselves common knowledge. Actually, this constitutes no loss
of generality. Included in the full description of a state cl of the world is the
manner in which information is imparted to the two persons. This implies that
the information sets P1(w) and P2(w) are indeed defined unambiguously as func-
tions of co, and that these functions are known to both players.
Consider next the assumption of equal priors for different people. John
Harsanyi (1968) has argued eloquently that differences in subjective probabilities
should be traced exclusively to differences in information-that there is no ra-
tional basis for people who have always been fed precisely the same information
to maintain different subjective probabilities. This, of course, is equivalent to

6Finest common coarsening of S90 and _32.

This content downloaded from 119.148.3.107 on Sun, 07 Oct 2018 05:19:22 UTC
All use subject to https://about.jstor.org/terms
 1238 ROBERT J. AUMANN

the assumption of equal priors. The result of this paper might be considered
evidence against this view, as there are in fact people who respect each other's
opinions and nevertheless disagree heartily about subjective probabilities. But
this evidence is not conclusive: even people who respect each other's acumen
may ascribe to each other errors in calculating posteriors. Of course we do not
mean simple arithmetical mistakes, but rather systematic biases such as those
discussed by Tversky and Kahnemann (1974). In private conversation, Tversky
has suggested that people may also be biased because of psychological factors,
that may make them disregard information that is unpleasant or does not con-
form to previously formed notions.
There is a considerable literature about reaching agreement on subjective
probabilities; a recent paper is DeGroot (1974), where a bibliography on the
subject may be found. A "practical" method is the Delphi technique (see, e.g.,
Dalkey (1972)). It seems to me that the Harsanyi doctrine is implicit in much
of this literature; reconciling subjective probabilities makes sense if it is a ques-
tion of implicitly exchanging information, but not if we are talking about "innate"
differences in priors. The result of this paper might be considered a theoretical
foundation for the reconciliation of subjective probabilities.
As an illustration, suppose 1 and 2 have a uniform prior on the parameter of
a coin, and let A be the event that the coin will come up H (heads) on the next
toss. Suppose that each person is permitted to make one previous toss, and that
these tosses come up H and T (tails) respectively. If each one's information
consists precisely of the outcome of his toss, then the posteriors for A will be 2
and I respectively. If each one then informs the other one of his posterior, then
they will both conclude that the previous tosses came up once H and once T, so
that both posteriors will be revised to 2
Suppose now that each person is permitted to make several previous tosses,
but that neither one knows how many tosses are allowed the other one. For
example, perhaps both make 4 tosses, which come up HHHT for 1, and HTTT
for 2. They then inform each other that their posteriors are 3 and I respectively.
Now these posteriors may result from a single observation, from 4 observations,
or from more. Since neither one knows on what observations the other's pos-
terior is based, he may be inclined to give more weight to his own observations.
Some revision of posteriors would certainly be called for even in such a case;
but it does not seem clear that it would necessarily lead to equal posteriors.
Presumably, such a revision would take into account each person's prior on
the number of tosses available to him and to the other person. By assumption
these two priors are the same, but each person gets additional private informa-
tion-namely, the actual number of tosses he is allotted. By use of the prior
and the information that the posteriors are, respectively, 2 and 3, new posteriors
may be calculated. If the players inform each other of these new posteriors, fur-
ther revision may be called for. Our result implies that the process of exchanging
information on the posteriors for A will continue until these posteriors are equal.

This content downloaded from 119.148.3.107 on Sun, 07 Oct 2018 05:19:22 UTC
All use subject to https://about.jstor.org/terms
 AGREEING TO DISAGREE 1239

REFERENCES

[1] AUMANN, R. J. (1974). Subjectivity and correlation in randomized strategies. J. Math.

Econom. 1 67-96.
[2] DALKEY, N. C. (1972). Studies in the Quality of Life. Lexington Books, Lexington, Mass.
[3] DEGROOT, M. H. (1974). Reaching a consensus. J. Amer. Statist. Assoc. 69 118-121.
[4] GROSSMAN, S. J. and J. E. Stiglitz (1976). Information and competitive price systems. Amer.
Econom. Rev. 46 246-253.
[5] HARSANYI, J. (1967-1968). Games of incomplete information played by Bayesian players,
Parts I-III. Management Sci. 14 159-182, 320-334, 486-502.
[6] RADNER, R. (1968). Competitive equilibrium under uncertainty. Econometrica 36 21-58.
[7] RADNER, R. (1972). Existence of equilibrium plans, prices, and price expectations in a se-
quence of markets. Econometrica 40 289-304.
[8] TVERSKY, A. and D. KAHNEMAN (1974). Judgment under uncertainty: Heuristics and biases
Science 185 1124-1131.
THE INSTITUTE FOR MATHEMATICAL STUDIES
IN THE SOCIAL SCIENCES
STANFORD UNIVERSITY
STANFORD, CALIFORNIA 94305

This content downloaded from 119.148.3.107 on Sun, 07 Oct 2018 05:19:22 UTC
All use subject to https://about.jstor.org/terms