Arrows Impossibility Theorem
Arrows Impossibility Theorem
• Each person i has preferences over the policy options %i , which are complete and transitive,
and for simplicity, we’ll assume they are strict as well (no indifferences)
9
2 What Are We Trying To Do?
• We’re looking for a way to aggregate preferences – that is, we want a way to turn each possible
set of individual preferences {1 , 2 , . . . , N } into a preference relation %∗ for “society”
• The mapping from individual preference relations into social preferences relations is alterna-
tively called a social welfare function, or a preference aggregation rule, or a Constitution
• I’ll use social welfare function – just remember that the SWF is not a set of preferences itself,
but a rule for generating a set of preferences for society for each set of individual preferences
– So if we let X be the set of all possible preference relations over the set of policies A, a
SWF is a mapping from X N to X
2.1 Social preferences need to be defined for any set of individual preferences.
• Our SWF has to specify some set of social preferences %∗ for any given set of individual
preferences {1 , 2 , . . . , N }
• This is also called “universal domain” – we’re not ruling out any possible preferences.
• we want the social preferences %∗ chosen by our SWF to be complete and transitive for each
set of individual preferences
Why? Suppose there are three people and three policies, with preferences
A 1 B 1 C
B 2 C 2 A
C 3 A 3 B
If we go by majority rule, two out of three prefer A to B, and two out of three prefer B to C,
and two out of three prefer C to A; so social preferences would have to be A ∗ B ∗ C but
C ∗ A, which isn’t transitive.
10
2.3 Social preferences should respect unanimity.
• If everyone in society agrees that policy a is strictly better than b, then the social preferences
defined by our SWF should also strictly prefer a to b.
• If not, it’s doing a pretty bad job of aggregating the individuals’ preferences.
• This rules out stupid rules like “society ranks all options equally” or “regardless of individual
preferences, society ranks policies alphabetically”
• Basically, this says that if we’re trying to figure out whether society prefers a to b, what
people think of c shouldn’t matter.
• Formally, suppose we start with some set of individual preferences {1 , . . . , N }, and the
SWF picks a social preference function under which a %∗ b
Now modify one guy’s preferences i such that his preference between a and b stays the same,
but his preferences for other things (including whether he prefers c to a or b) changes
Then under the new preferences, the SWF should still pick a social preference function that
prefers a to b.
Suppose there are three policies, and we ask everyone for a rank-order list; and then we give
a policy 3 points for each person who ranked it first, 2 points for each person who ranked it
second, and 1 point for every person who ranked it third, and we say society prefers policies
with more points.
A 1 B 1 C
B 2 C 2 A
A 1 C 1 B
B 2 A 2 C
it would prefer A to B. But in both cases, A 1 B and B 2 A – all we did was change parts
of preferences that “aren’t supposed to matter” for choosing between A and B. So a Borda
count rule violates IIA.
11
2.5 (Aside: note the ordinality of preferences)
• All we know is whether an individual prefers one policy to another – we have no language to
even talk about how much he prefers one to another
• We don’t have money, to allow people to barter (I prefer A to B, but I’ll prefer B if you give
me $11), or anything like that
• All we have is ordinal pairwise preferences, and we need to aggregate them to preferences for
society
• One more thing to define: a SWF is a dictatorship if the social preference always just reflects
the same one guy’s preferences, that is, if there’s some individual k such that regardless of
anyone else’s preferences, a ∗ b if and only if a k b.
3 The Result
Theorem (Arrow). Any SWF which respects transitivity, unanimity, and independence of irrele-
vant alternatives is a dictatorship.
• The proof has several steps. We assume we have a SWF which satisfies transitivity, unanimity,
and IIA, and then show that this means there must be some voter whose preferences always
match the social preferences, who is therefore a dictator.
12
4 The Proof
• Throughout the proof, we will maintain the assumptions that social preferences %∗ are always
transitive, and that the SWF satisfies unanimity and IIA.
Lemma 1 (Extremal Lemma). For any policy b, if every individual i ranks b either strictly best or
strictly worst, then %∗ must rank b either strictly best or strictly worst as well.
• Suppose the lemma were false. Then there would be some set of individual preferences and
two other policies a and c such that a %∗ b %∗ c, even though each individual has b as either
their favorite or least favorite policy.
For each individual i who has b at the top of their list: move c up to second on the list, so it
is now strictly above a if it wasn’t already.
And for each individual i who has b at the bottom of their list, move c to the top of the list,
so it’s strictly above a if it wasn’t already.
• This doesn’t change any individual’s preferences between b and a: if b was at the top of your
list, it’s still at the top; and if it was at the bottom, it’s still at the bottom. So by IIA, society
still prefers a to b
• And similarly, it doesn’t change any individual’s preferences between b and c, so by IIA, b is
still preferred to c.
• But now, everyone has c ranked above a, so unanimity would require c ∗ a, giving a
contradiction
• So that proves that b must be either strictly best or strictly worst according to %∗ – if it
wasn’t, we could generate this type of contradiction.
13
4.2 Part 2 – Find a pivotal guy and give him a name.
• We know (unanimity) that if everyone in society puts b last, then %∗ puts b last too; and if
everyone puts b first, %∗ puts b first
• And we just showed that if some people put b first and the rest put b last, then %∗ puts b
either first or last
• So start out with a set of individual preferences where everyone in society puts b last. By
unanimity, %∗ must put b last as well.
• Now change voter 1’s preferences by moving b from last to first. Since everyone in society
either likes b the most or the least, b must be either first or last in %∗ .
• Now change voter 2’s preferences by moving b from last to first. Again, since everyone in
society either likes b the most or the least, b must be either first or last in %∗ .
• Keep going like this. By the time we’ve switched everyone’s preferences to having b first, by
unanimity, society must have b first as well.
• Now find the voter where the first switch happened. That is, the first time that the social
preference %∗ switched from having b at the bottom to b at the top. Call him Bob.
That is, given the other preferences we started with, if everyone with a lower number than
Bob has b first on their list, and Bob and everyone after has b last on their list, then society
puts b last; but if Bob switches to having b first, then society puts b first.
• Call the first profile of preferences – where everyone up to Bob puts b first, and Bob and the
rest put it last – profile I; and call the second set of preferences – where everyone up to Bob,
and Bob himself, put b first, and the rest put it last – profile II.
• The rest of the proof involves showing that Bob is a dictator – that to satisfy IIA and
transitivity, %∗ has to always agree with Bob’s preferences, regardless of what everyone else
thinks
14
4.3 Part 3 – Proving Bob is a dictator
• That is, for any two policies a and c which aren’t b, we’ll show that if Bob prefers a to c, then
no matter what everyone else’s preferences are, %∗ must put a strictly ahead of c as well.
• Start with arbitrary preferences where a Bob c, and call these (true) preferences Profile IV.
Move a to the top of Bob’s preference list, and b to second on Bob’s preference list
• When we moved from IV to III, we didn’t change anyone’s ranking of a versus c – for Bob, a
was above c, and we moved it to the top; for everyone else, all we did was move b
So by IIA, the societal preference between a and c has to be the same at profile IV as at
profile III
• Now, recall that at profile I, society put b last, which means a ∗ b at profile I.
• At profile II, on the other hand, society put b first, which means b ∗ c at profile II
• So at profile IV, a ∗ c.
• So whenever a Bob c, a ∗ c.
15
Part 3b: Bob is also dictator when it comes to b
• All that’s left is to show is that Bob is also a dictator when one of the choices being considered
is b
• This is the easy part. We need to show that if b Bob a, b ∗ a; and if a Bob b, a ∗ b.
• So now pick a policy that’s different from those two – say, c – and repeat everything we
already did
• Starting with the extremal lemma, move preferences from c worst to c best, find the pivotal
guy, and prove that he has to be a dictator for any two policies that aren’t c
– So far, we don’t know whether this new dictator is Bob or someone else
– Just that someone is a dictator when c isn’t involved, meaning, when choosing between
a and b
– When we moved from profile I to II, Bob’s preferences for b were all that changed
And that shifted society from putting b last to putting b first
• So if the new dictator isn’t Bob, we have a contradiction; so the new dictator must also be
Bob
• Which means Bob is dictator over any pair of policies that excludes b, and over any pair of
policies that includes b
16
5 So What?
• So the result is, if you want to aggregate individual preferences and get a transitive social
preference that respects unanimity and IIA, all that’s left is a dictatorship.
– Basically, by assuming IIA, we’re ruling out inferring anything “cardinal” about prefer-
ences from where you rank other alternatives
– If there are 100 policy choices, and you have a and b ranked #45 and #46, chances are,
you’re pretty close to indifferent between them
OTOH, if you have a ranked #1 and b ranked #100, you probably like a a lot more
IIA says that for choosing between a and b, we have to treat those two cases the same
– all we’re allowed to consider is that you like a more than b
By assuming IIA, we’re really ruling out having any way to elicit cardinal preferences
Which is what makes this hard
– In some sense, markets behave well exactly because prices elicit cardinal information
– How much you’re willing to buy of something at a given price reveals exactly how much
you like it – not just whether you prefer it to something else
• In Arrow’s world, we can’t use prices – or anything else – to make this type of judgment
• One other thing to notice: we’ve completely ignored the problem of figuring out peoples’
preferences
– Depending on the social choice function, people might have an incentive to lie about
their preferences
– The rest of the semester will, in some sense, be focused on that part of the problem
References
• Original result: Kenneth Arrow (1951), Social Choice and Individual Values, New York: Wiley
• Our proof: John Geanakoplos (2005), “Three Brief Proofs of Arrow’s Impossibility Theorem,”
Economic Theory 26(1)
17