Gibbard-Satterthwaite Theorem

MATH3002 Mathematics Seminars Presentation Report

Tang Chiu Chak

(U. id.: 3035074660)

14th Feb 2014

Contents
1 Introduction 2
1.1 Motivation —- An Example Case . . . . . . . . . . . . . . . . 2
1.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Proof of Gibbard-Satterhwaite theorem 4

2.1 Idea Of The Proof . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Propositions Related To Non-Manipulability . . . . . . . . . . 5
2.3 A Vital Observation . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Revision On The Supplied Book Extract 15

3.1 Revision On The Crucial Observation . . . . . . . . . . . . . . 15
3.2 Revision On The Proof In Lemma 2 . . . . . . . . . . . . . . . 15

1
1 Introduction
1.1 Motivation —- An Example Case
Imagine that you and your secondary schoolmates are going to have vote on
where the next reuion take place. The voting profile happens to be as below :

Sch. mate 1 - 18 Sch. mate 19-28 Sch. mate 29-37 You

McDonald Cafe Restaurant X Restaurant X

Cafe Restaurant X Cafe Cafe
Restaurant X McDonald McDonald McDonald
Profile 1.1a
Suppose that the final decision will be made based on the Hare system,
that is, by iteratively deleting options (alternatives) that have the least vote
and reconsidering the profile again until a set of winner(s) that everyone
agrees on is produced. But then according to the profiile above, Cafe and
Restaurant X (each having only 10 votes) will be kicked in the first round as
they have the least votes (McDonald has 19 votes). Consequently, one can
verify that McDonald, the alternatives you least desire, will then be chosen
in this case.

Thus, to avoid this from happen, you secretly change your vote: you ex-
change the position of Cafe and Restaurant X in your choices, resulting in
the follow new profile:

Sch. mate 1 - 18 Sch. mate 19-28 Sch. mate 29-37 You

McDonald Cafe Restaurant X Cafe

Cafe Restaurant X Cafe Restaurant X
Restaurant X McDonald McDonald McDonald
Profile 1.1a
Consequently, based on the Hare system, one can verify that Restaurant
X (having 9 votes) will be deleted in the first round, McDonald (having 19
votes in the second round) being kicked away next, and the final decision
becomes Cafe (having 20 votes in second round)

2
Hurray! No more McDonald’s! But how can this happen? How can a
single voter can have such a power that he can somehow control the result
of a seemingly “democratic” procedure that based on the will of 40 people
by strategically giving an insincere vote?

We would like to investigate whether there can be an “ideal” social choice pro-
cedure, such that it can “reasonably” reflect the society’s will while it cannot
be controlled by anyone. That is why we discuss the Gibbard-Satterthwaite
Theorem this time.

1.2 Definition
To begin with, let us first state the theorem:

Gibbard-Satterhwaite Theorem– Suppose that there are at least

three elements in the set A of alternatives and the set I of voter individuals
is finite. The only social choice procedure for A and I which never yield ties
and satisfies both the Pareto condition and non-manipulability is a dictator-
ship.

What does this mean? let us explain the terms one by one:

† Pareto condition (unanimity)– For any alternatives x,y belongs

to A, if every individual prefers x to y, then y is not a social choice.
(That is, the social choice function has no ”aim”: it will simply reflect what
everybody agrees on, if any)

† Manipulability – Assign the voters into some arbitrary but fixed

order and denote the vote (preference list) of the k-th voter by Vk . A
social choice function is defined to be manipulable if there exists a profile
P = (V1 , V2 , ...Vi , ..., Vn ) producing a social choice x (A profile which we as-
sume to be representing the true preference lists of the n voters) and another
profile P 0 obtained from P by only switching Vi to Vi0 ( an insincere vote from
the i-th voter) so that P 0 produce another social choice y which the i-th voter
prefer more to x.
A social choice procedure is called non-manipulable if it is not manipulable,
i.e. there does not exist any such pair of profiles for the procedure.

3
† Dictatorship – A social choice procedure is defined to be a dicta-
torship if there exist a special voter, called a dictator, such that the choice
on top of his list will automatically be the social choice no matter how others
vote.

In other words, Gibbard-Satterhwaite Theorem tells us that there is no so-

cial choice procedure that produces only one winner each time, reasonable
in some sense (satisfying Pareto condition), can not be manipulated and is
under no dictatorship. This alarms us that with some condition bound, a
seemingly ”democratic” social choice procedure is susceptible to some ”tac-
tical voting”– someone with full understanding of how others vote may be
able to ”control” the outcome by giving an insincere vote.

So ,let try proofing the theorem.

2 Proof of Gibbard-Satterhwaite theorem

2.1 Idea Of The Proof
Before we start explaining how we are going to proof the theorem, let me
first introduce the idea of Dictating Set.

Definition. A subset X of the set I of all voters is defined to be a

dictating set if ”X can use a to block b” for any distinct alternatives a,b in
the alternatives set A. By ”X can use a to block b”, we mean that when-
ever everyone in X prefers a to b (that is everyone in X ranks a higher than
b in their vote) then the social choice will not be b.

The reasons why we would like to consider such set are that firstly, such
kind of sets must exist– when X = I, then X is bound to be a dictating set
since if everyone in X, that is, everyone in I rank a higher than b, then by
Pareto condition we have b must not be the social choice, that is, by Pareto
condition we have X = I can use a to block b. Secondly, if we are able to
show that the social choice procedure have a dictating set containing only
one element, say X = {p}, then by definition that particular element p is
a dictator. (Since if p put a alternative say l on the top of this list, then
by definition p can use l to block any other alternative and hence the social

4
choice will not be any other alternative but l)

Our strategy for proofing the theorem will be that starting with the largest
dictating set (which must exist as argued above), we will divide the set down
into two disjoint pieces (so that the size of the resulting sets will be smaller
as X is finite) and we would like prove that under the conditions stated in
the theorem, one of the pieces must also be a dictating set. If we can do so,
then by iterating the process we will end up with a dictating containing only
one element inside, which indicate the existence of a dictator.

2.2 Propositions Related To Non-Manipulability

In fact, instead of using non-manipulability, which is seemingly complicated,
to complete our proof, we would like to replace it by another property that all
non-manipulable social choice procedures satisfy– down-monotonicity. Let us
first state what down-monotonicity is:

Definition. For any alternative x, y belongs to A, where x is the so-

cial choice and y is not, A social choice procedure is said to satisfy down-
monotonicity if anyone changes his vote by moving down y one spot (or in
other words, exchanging the position between y and the alternatives ranked
right below y in his vote), then x is still the social choice.

Proposition 1. Non-manipulability implies down-monotonicity.

Proof.
(we would like proof the proposition by contrapositive, that is, we assume
that a social choice function does not satisfy down-monotonicity and we
would like to deduce that such social choice procedure is manipulable.)

Assume the produce does not satisfy down-monotonicity, then there must
exist two profiles P and P 0 and alternatives a, b, c, d (some of them, but not
all, are allowed to be the same element) such that

1. (Profile P ) Alternative a is the social choice, and some fixed individual i

ranks c (c 6= a as we would like to consider the case down-monotonicity
is not satisfy) immediately above d in his preference list (see profile

5
 2.2a) ;

2. (profile P 0 ) P 0 only differ from P in that the voter i as mentioned

above has interchanged the position of c and d and we have alternative
b (b 6= a, again, we assume this as we would like to consider the case
down-monotonicity is not satisfy) is the social choice in P 0 (see profile
2.2b)

.
... Voter i (Vi ) ... Output ... Voter i (V 0 i ) ... Output
.. .. .. .. .. ..
. . . a . . . b
.. .. .. ..
. c . . d .
.. .. .. ..
. d . . c .
.. .. .. .. .. ..
. . . . . .
Profile 2.2a Profile 2.2b
Notice that we can divide this situation into three cases:

1. voter i ranks b above a in profile P .

2. voter i ranks a above b in profile P 0 .

3. the negation of the above two, i.e. voter i ranks a above b in profile P
and ranks b above a in profile P 0 .

.
Case 1a: voter i ranks b above a in profile P .
Then suppose the vote of voter i in P is his sincere preference while the
one in P 0 is insincere. As voter i prefer b to a, by submitting his insincere
preference V 0 i , the social choice will become b instead of a. Hence the social
choice procedure is manipulable in this case.

Case 1b: voter i ranks a above b in profile P 0 .

(note that case 1a and case 1b are not disjointed)
Then suppose the vote of voter i in P 0 is his sincere preference while the one
in P is insincere. As voter i prefer a to b, by submitting his insincere preference
Vi , the social choice will become a instead of b. Again the social choice procedure
is consequently manipulable in this case.

Case 2: Otherwise

6
 Then we have voter i ranks a above b in profile P and ranks b above a in profile
P 0 . Such change in preferences list indicates that c = a and d = b since c and d
are the only alternatives that voter i has changed the relative position. But then
this contradict our construction that c 6= a. So Case 2 is impossible.

Combining the argument above, we have if a social choice procedure does not
satisfy down-monotonicity then it is manipulable. By contrapositive we have if a
social choice procedure is non-manipulable, then it satisfies down-monotonicity.

Proposition 2. Down-monotonicity implies monotonicity.

Proof.
Suppose alternative x is the social choice. If anyone changes his vote by mov-
ing up x one spot, that is, exchanging the position between x and the alternative,
say y, ranked intermediately above x in his preference list, which is the same as
saying that moving down y one spot, then by down-monotonicity, x is still the
social choice. This complete the proof.

Note that even though we were just moving the alternatives one slot, our proof
can still apply to the situation moving the alternatives several slots as these cases
can be expressed as moving the alternatives one slot several times .

2.3 A Vital Observation

In this section, we would like to make a paramount observation which will greatly
simplify our proof.

As later we are going to proof the theorem by finding out a smaller and smaller
dictating set, we need to find out a way to proof some set X dictating, or at least,
a way to show that some sets X can ”use a to block b”. How to do so? One
most straight forward idea will be simply list out all the cases, and see whether
the criterion is satisfied, that is, whether when everyone in X rank a over b we will
have b not be a social social choice.

Here we can see that we can simply ignore any cases such that someone in X
does not rank a over b as these cases do not bother our proof. In fact, as the social
choice procedures we are going to consider are subjected to down-monotonicity,
we can further simplify our examination to only considering one single case from
”worst case scenario”, which we would like to state this a vital observation:

7
A Vital Observation
For any subset X of the set I of individual and for any alternatives a, b in A, to
show that X can use a to block b, we just need to consider one single case with
the following conditions holds:

1. Everyone in X ranks a the top and b the second.

2. Everyone not in X ranks b the top.

3. The resulting social choice in this case is not b.

(Notice that the first condition implies everyone in X ranks a over b and the second
condition implies everyone not in X rank b over a.)

It is called ”the worst case scenario” in a sense that if we can provide a case
satisfying all the conditions listed above, then we must have b is not the social
choice in all other cases where everyone in X ranks a over b. Let’s proof this
claim.

Proof: Suppose that we have the set of alternative A = {a, b, x1 , x2 , . . . , xn }.

Assume that we can provide a case , denoted as C0 , such that it satisfies all the
three condition listed above. Then for any other cases such that every one in X
rank a over b, denoted as C 0 , we do the following step by step:

1. Rearranging the position of x1 , x2 , . . . , xn in the votes from everyone in X

and the the position of a, x1 , x2 , . . . , xn in the votes from everyone not in X
,without changing the position of a, b in votes’ in X and the position of b in
votes not in X, so that the relative positions of x1 , x2 , . . . , xn in votes from
everyone in X and the relative positions of a, x1 , x2 , . . . , xn in votes from
everyone not in X in the new profile obtained will be the same as the their
relative positions in C 0 . Denote the obtained new profile as C1 .

2. From C1 , move b down so the the relative positions of b, x1 , x2 , . . . , xn in

votes from everyone in X and the relative positions of all alternatives in
votes from everyone not in X in the new profile obtained will be the same
as the their relative positions in C 0 . Denote the obtained new profile as C2 .

3. From C2 , move a down so the the relative positions of all alternatives in the
new profile obtained will be the same as the their relative positions in C 0 .
Denote this final product as C3 , where we should have C3 = C 0 .

8
Firstly, we claim that after the first step, we still have the social choice in C1 not be
b. The reason is that we can decompose the position rearrangement as successive
steps of moving some alternatives xi up one slot, i.e. exchanging the position of xi
with the alternative immediately above it. Now if b becomes the social choice in
the process at some point, then we reverse the process from that point, i.e. move
some alternatives xi (which then must not be the social choice as the procedure
produce unique winner) down back one slot and we will consequently still have
b be the social choice by down-monotonicity. Continuously reversing the process
one should then be able to conclude that b is the social choice at the beginning
case, which contradict with our assumption that b is not the social choice in C0 .
Hence, we have b is still not a social choice in C1 .

Secondly, we claim that b is not the social choice after step 2. This follow im-
mediately from our first claim since if we have b not be a social choice in C1 , then
as C2 is obtained from C1 by moving b down, by down-monotonicity we have b is
not the social choice in C2 .

Finally, we claim that b is still not the social choice after step 3. The reason
is that again, we can express ”moving a down” as successive steps of moving the
alternatives immediately below a up one slot. Then if b become the social choice
halfway through the process, we will still have b be the social choice by down-
monotonicity when we reverse the process, which is a contradiction. So b is still
not a social choice. (Note that b will not be moved in this step as everyone in X
ranks a over b in case C 0

Now we have provided an algorithm to “transform” C0 to any other case C 0 which

everyone in X rank a over b and showed that under this algorithm, b will still not
be a social choice. Hence, we have shown that to verify a set X can use a to block
b, we only need to provide a single case such that in that case, everyone in X ranks
a the top and b the second, everyone not in X ranks b the top and the resulting
social choice in this case is not b.

(It is advised that one should try to follow the algorithm once if possible since
it is easier to understand then just reading the algorithm.)

2.4 The Proof

Here we will divide out proof into six parts : three lemmas about the “blocking”
property, two lemmas about breaking a dictating set into smaller dictating set and
the finally a proof for the theorem by applying the lemmas.

9
Lemma 1. Let a, b, c be three distinct alternatives in A. If X can use a
to block b, then X can use a to block c and X can use c to block b.
(Since the role of c is arbitrary. The lemma here in fact means that If X can use
a to block b, then X can use a to block anything and X can use anything to block b.)

Proof. We divide the proof into two parts:

Part 1: X can use a to block c — We consider the following case (Profile 2.4a)
:
Everyone in X Everyone not in X Output
a b a
b c
c a
.. ..
. .
Profile 2.4a

First notice that everyone ranks a, b, c higher than everything else and everyone
ranks b over c, hence by Pareto condition we must have the social choice be either
a or b. Second, since everyone in X ranks a over b and the assumption that X
can use a to block b, we must have the social choice be a but not b. Now, move
b down in everyone’s preference list (Profile 2.4b), we then will have, by down-
monotonicity, the social choice remain to be a .

Everyone in X Everyone not in X Output

a c a
c a
b b
.. ..
. .
Profile 2.4b

Now then we have provided a case where everyone in X ranks a the top and c the
second, everyone not in X ranks c the top and the resulting social choice is not c.
Based on our ”vital observation”, X can use a to block c.

Part 2: X can use c to block b —

We consider the following case (Profile 2.4c) :

10
 Everyone in X Everyone not in X Output
c b c
a c
b a
.. ..
. .
Profile 2.4c

First notice that everyone ranks a, b, c higher than everything else and everyone
ranks c over a, hence by Pareto condition we must have the social choice be either
c or b. Second, since everyone in X ranks a over b and the assumption that X
can use a to block b, we must have the social choice be c but not b. Now, move
a down in everyone’s preference list (Profile 2.4d), we then will have, by down-
monotonicity, the social choice remain to be c .

Everyone in X Everyone not in X Output

c b c
b c
.. ..
. .
Profile 2.4d

Now then we have provided a case where everyone in X ranks c the top and b the
second, everyone not in X ranks b the top and the resulting social choice is not b.
Based on our ”vital observation”, X can use c to block b.

Combining the argument in the two part, we have if X can use a to block b,
then X can use a to block anything (c) and use anything (c) to block b.

Lemma 2.– If X can use a to block b then X can use b to block a.

Proof.
Let a, b, c be three distinct elements in A, notice that this is possible since A has
at least three elements. As X can use a to block b, then by lemma 1, X can use
a to block c and by lemma 1 again X can use b to block c. Once more by lemma
1, X can use b to block a. This complete the proof.

Lemma 3 – If X can use a to block b , then X is a dictating set ,i.e.

X can use x to block y for any distinct element x, y in A.

Proof.

11
since we have either a = y or a 6= y

• Case 1: a = y— then by lemma 2 X can use b to block y = a and by

lemma 1 X can use x to block y.

• Case 2: a 6= y — then by lemma 1 X can use a to block y and by lemma

1 again, X can use x to block y.

In any case, we have X can use x to block y. This complete the proof.

Lemma 4. Let a, b, c be three distinct alternatives in A ( which is possible

as A has at least three elements) . Suppose X can use a to block b. If X is
partitioned into two sets Y and Z (empty set is allowed), then one of them can
use a to block c or use c to block b.
(That is, one of the pieces will inherit the abilities of using a to block others or
the abilities of keeping b be blocked)
Proof.
First , we consider the profile where everyone in X ranks a the top and b the
second while everyone not in X, ranks b the top (see profile 2.4e). Then by our
assumption we have X can use a to block b, i.e. b is not the social choice. Also,
together with unanimity one must be a be the social choice.

Everyone in X Everyone not in X Output

a b a
b a
.. ..
. .
Profile 2.4e

Now, partition X into 2 sets Y and Z. Consider the case obtained from profile2.4e
by moving up c such that it is as the same as shown in profile 2.4f

Everyone in X Everyone not in X Output

Everyone in Y Everyone in Z
a c b a/c
b a c
c b a
.. .. ..
. . .
Profile 2.4f

12
Notice that by down-monotonicity, after we moving up the position of c, the social
choice should either remain to be a or change to be c.
(For otherwise, if any other alternative d 6= a, c becomes the social choice, then
when we reverse the process, d should remain to be the social choice by down-
monotonicity, which is a contradiction).

Case 1 : a remains to be the social choice – As b is not the social choice, we

move b down and by down-monotonicity a will remain to be a social choice (profile
2.4g).And as the procedure always produce unique winner, we have c is not the
social choice.
Everyone in X Everyone not in X Output
Everyone in Y Everyone in Z
a c c a
c a a
b b b
.. .. ..
. . .
Profile 2.4g

Hence according to our ”vital observation”, we have provided a case (profile 2.4g)
such that Y can use a to block c.

Case 2 : c becomes the social choice – Then again as the procedure al-
ways produce unique winner, we have b and a are not the social choice. And as a
is not the social choice, we move a down and by down-monotonicity c will remain
to be a social choice (profile 2.4h).

Everyone in X Everyone not in X Output

Everyone in Y Everyone in Z
b c b c
c b c
a a a
.. .. ..
. . .
Profile 2.4h

Then according to our ”vital observation”,we have provided a case such that Z
can use c to block b.

Consequently, in any case, we have proof that if a set, X, which can use a to
block b is partitioned into two sets Y and Z (empty set is allowed), then one of

13
them can use a to block c or use b.

Lemma 5. Suppose is a dictating set and that X is split into two sets, Y
and Z , then one of them is a dictating set.

Proof. Let a, b, c be three distinct element in A, which is possible since there is

at least three element in A. Now as X is a dictating set, it can use a to block b.
Then by lemma 4, either Y or Z can either use a to block c or use c to block a.
Assume, W LOG, Y can use a to block c , then by lemma 3, Y is a dictating set.
This complete the proof.

Gibbard-Satterhwaite Theorem– Suppose that there are

at least three elements in the set A of alternatives and the set I of all voter indi-
viduals is finite, then the only social choice procedure for A and I which produce
always produce an unique winner and satisfies both the Pareto condition as well
as non-manipulability is a dictatorship.

Proof. By position 1 and 2 we have the social choice procedure must satisfy
down-monotonicity and monotonicity (and hence all the lemmas above can apply).
Suppose the set I contains n elements, and assign all voters an arbitrary but fixed
index from 1 to n. Notice that the set I itself is a dictating set. Partition I into two
disjoint sets, Y1 and Z1 where Z1 = {x1 } and Y1 = I \ Z1 . Then by lemma 5 either
Y1 or Z1 is the dictating set. If Z1 is the dictating set, then we are done— x1 is the
dictator. If Y1 is the dictating set, then we continue partitioning. Iterate this pro-
cess, i.e. partitioning Yk into Zk+1 = {xk+1 } and Yk+1 = I \ {x1 , x2 , . . . , xk , xk+1 }
until we get Zj is the dictating set for some 1 ≤ j ≤ n, which indicate that the
existence of a dictator ( i.e. the xj voter is a dictator.) We know that our iter-
ating process must shop for some point some as each step we produce a smaller
and smaller finite-size dictating set and their size are bounded below by 0. This
complete the proof.

14
3 Revision On The Supplied Book Extract
3.1 Revision On The Crucial Observation
The original three conditions stated in the crucial observation in the book extract
are listed as follows:

1. Everyone in X ranks a over b .

2. Everyone not in X ranks b over a.

3. The resulting social choice in this case is not b

We claim that these conditions are not sufficient enough to implies ”X can use a
to block b”. Here is to provide a counter example.

Counterexample— Consider the following profile (profile3.1a) where there are

only 2 voters , Vx and Vy where X = {Vx } and there are only three alternatives
a, b, c. Assume that Vy is a dictator.

Vx Vy Output
a c c
b b
c a
Profile 3.1a

One can verify that all three conditions listed is satisfy in profile 3.1a while X
cannot be a dictating set since Vy ∈
/ X is a dictator.

3.2 Revision On The Proof In Lemma 2

There is seemingly a logical error in the writer’s proof for his second lemma–

Suppose X can use a to block b and c is an alternative distinct from

a and b, then X can use a to block c and X can use c to block b

The error lies in that he named some set to be Y and deduce that the set Y
can use a to block c because of the naming –

...that Y is all of X and Z is the empty set. Lemma 1 then implies

15
that either Y can use a to block c ...

For your reference, his lemma 1 states that –

Suppose X can use a to block b and c is an alternative distinct from

a and b. Suppose that X is split into two sets Y and Z (either of which
may be the empty set) so that each element of X is in exactly one of
the two sets. Then either Y can use a to block c or Z can use c to block
b

But one should notice that since naming a set is an arbitrary issue: one can
call the set originally named Y to be Z and call that set originally named Z to be
Y . So in fact, what his lemma 1 suggest should be either one of the resulting
sets can use a to block c or Z can use c to block b. That is, if we partition a set
X, which can use a to block b, into two sets Y and Z, one should have one of the
following:

1. Y can use a to block c.

2. Y can use c to block b.

3. Z can use a to block c.

4. Z can use c to block b.

And we do not know which case will happen solely based on his lemma 1. Hence
what the writer has proven in his lemma 2 is actually: If X can a to block b, then
either X can use a to block c or X can use c to block b. instead of ‘‘X can use
a to block c and X can use c to block b" as desired.

Despite this, I am sorry that I cannot provide a counterexample for “disproving”

since what he want to prove turns out to be true (as we have proven in section 2.4
Lemma 1)

16