Algorithmic Game Theory, Summer 2019 Lecture 24 (8 pages)

Voting
Thomas Kesselheim Last Update: July 8, 2019

Today, we will turn to a mechanism-design problem, in which you once again don’t want to
have money involved: Voting. Say, you want to design presidential elections. How would you
design the rules? If the choice will be only among two candidates, this is easy.

Example 24.1. If there are only two candidates A and B, let each voter vote for one of the
candidates. The winner is whoever receives the majority of votes. (Ignore the issue of tie-breaking
here.)

This “voting rule” fulfills almost every design goal that we could ask for. For example, there
is no other “better” outcome. Only a minority of voters would prefer the other candidate. All
votes count the same, so the identities of voters do not matter. While this may sound boring,
just consider the US presidential elections. There, frequently candidates get elected that get less
votes than another candidates.
Another advantage is that participating as a voter is easy. (We always assume that you know
your preferences.) You simply vote for the candidate that you prefer. No voter can influence the
outcome in his favor by not voting for his favorite candidate.
If we have three or more candidates, things are not nearly as clear, clean, and simple. There
are different voting and ranking rules and some of them fulfill more desirable properties than
others. As we will see today and next time, there is a plethora of possible voting rules, some of
which are better than others, but none of them actually fulfills all the desired properties at once.
Indeed, we will also prove formally that this is just not possible.

1 Formal Problem Statement

Again we have n agents N . Furthermore, there are m candidates Γ. Each of the agents has a
complete preference list over the candidates without ties. That is, agent i has some preference
relation i , which is a total order over Γ. We write a i b i c if agent i strictly prefers
candidate a over candidate b and b over c. Note that this also implies a i c (transitivity).
There are two problems of interest:

• In the voting problem, given preferences 1 , . . . , n , we have to come up with one winner
from Γ.

• In the ranking problem, given preferences 1 , . . . , n , we have to come up with a society

ranking > of all candidates. That is, besides finding a winner, we also have to choose a
second, a third, and so on.

So, our goal is to design a voting rule or a ranking rule. In the case of two candidates, these
two notions coincide.
These problems are different from most of what we have seen so far. We do not have an
“objective function”. There are different ways to evaluate the outcome and naturally there are
trade-offs. Furthermore, the outcome affects the entire society. In all other cases, if we simply
took the union of two sets of agents, we could still have treated them as separate groups.

2 Examples of Voting and Ranking Rules

Let us see possible examples of voting and ranking rules with three or more candidates and let
us find out their weaknesses.
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 2 of 8)

2.1 Plurality Voting

In plurality voting, every voter can vote for one candidate. The candidate who got the most
votes wins the election. This can be translated into a voting rule within our framework by
soliciting the entire preferences but only counting how often one candidate is ranked in the
top-most position by an agent. We can easily also turn this voting rule into a ranking rule by
ordering the candidates by decreasing number of votes.
We all know examples in which plurality voting fails to give a “natural outcome”. Suppose
there are two candidates A and B from one party (so their policy is very similar) and another
candidate C from a very different party.
60 % of the voters prefer the party of A and B, 40 % prefer C’s party. In terms of preferences
this may look as follows:

35 % : A i B i C
25 % : B i A i C
40 % : C i A i B

In this case, C wins the election despite the fact that most voters would be in favor of the
party of A and B.
In the light of strategic behavior this also means that we cannot expect truthful reports.
The agents in the B i A i C group could decide to all report A i B i C instead. Then
candidate A wins, who they prefer to C.

2.2 Runoff Voting and Instant-Runoff Voting

A very common way to deal with the deficiencies of plurality voting is to run a second round in
which the voters can choose between the two candidates that got the highest number of votes
in the first round. This is called runoff voting. In our framework it can be captured by first
counting two first preferences just as in plurality voting. Among the top two candidates, call
them X and Y , one counts for how many i one has X i Y in comparison to Y i X. Whoever
gets the majority wins the vote.
If one has the full preference lists of all agents, one can also extend this procedure to multiple
stages. Again, we solicit the preferences of all agents and count how often a candidate is ranked
first. The candidate who receives the smallest number of first-rank votes is then eliminated
(that is, he gets removed from all preferences) and the procedure is re-iterated until only a single
candidate is left. For three candidates, this is equal to runoff voting, for more the result is
potentially different.
Obviously, the outcome on the above example is different now. First, candidate B gets
eliminated. Then, candidate C is eliminated because only 40 % prefer him to A.
However, sometimes it may still be interesting for agents to misreport. For example, consider
the following preferences.

30 % : A i B i C
45 % : B i C i A
25 % : C i A i B

In this case, candidate A wins. The B i C i A group could instead report C i A i B, then
candidate C would win, who they prefer to B.

2.3 Dictatorship
An obvious feasible choice of a ranking rule is to take agent i’s preferences (or the ones of any
other fixed agent). While this is clearly not what we have in mind when speaking of elections in
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 3 of 8)

a democracy, it does have a positive property: No agent can gain anything from misreporting
the preferences.

3 Who should win elections?

What should ideal voting or ranking rules fulfill? Let us introduce four examples.

• Anonymity: The identities of the agents do not matter. That is, the outcome on preference
profile (1 , . . . , n ) is the same as on (π(1) , . . . , π(n) ) for any permutation π : N → N .

• Monotonicity: If one agent moves a candidate up in the ranking, keeping everything else
fixed, the candidate does not move down in the combined society’s ranking.

• Condorcet winner criterion: If, in pairwise comparisons, one candidate is always preferred
to any other candidate in the majority of of all preference profiles, it is the winner of the
election. That is, if there is A such that for all B there are more i with A i B than
B i A, then A is the winner of the election.

• Condorcet loser criterion: If for some A for all B there are more i with B i A than
A i B, then A does not win the election.

3.1 Plurality Voting

Let us come back to our initial example of plurality voting. We choose the candidate who
is ranked highest in the plurality of agents’ preferences. This voting rule is anonymous and
monotone.
Recall the following example, in which C is the winner.

35 % : A i B i C
25 % : B i A i C
40 % : C i A i B

Observe that A is the Condorcet winner. In A vs. B, he gets 75 % of the votes; in A vs. C,
he gets 60 % of the votes. So, the Condorcet winner criterion is not fulfilled.
Furthermore, C is the Condorcet loser in this case. We observe that in C vs. A and also in
C vs. B he gets 40 % of the votes. That is, the Condorcet loser criterion is not fulfilled either.
Instant-runoff voting does not fulfill monotonicity or the Condorcet winner criterion (see page
221 in the Karlin/Peres book). It does fulfill the Condorcet loser criterion because a Condorcet
loser would never survive the final round.

3.2 Borda Count and Positional Voting

A very general voting/ranking approach works as follows. Depending on the position in the
preference list, each candidate gets a score from every agent. The social ranking is determined
by ordering the candidates by sum of scores. Every vector of a1 ≥ . . . ≥ am is such a positional
voting rule. Plurality voting is recovered by a1 = 1, a2 = . . . , am = 0. Again, any of these rules
is anonymous and monotone.
Another common rule is the Borda count, in which one sets a1 = m, a2 = m − 1, . . . , am = 1.
Let us consider our example from above:

35 % : A i B i C
25 % : B i A i C
40 % : C i A i B
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 4 of 8)

If we have n = 100 voters, A gets a total score of 35 × 3 + 25 × 2 + 40 × 2 = 235, B gets a score

of 35 × 2 + 25 × 3 + 40 × 1 = 185, C gets a score of 35 × 1 + 25 × 1 + 40 × 3 = 180.
Note that equivalently, we could perform pairwise votes between any two candidates and
sum up the number of votes each candidate gets in his m − 1 pairwise comparisons.
Other voting rules follow the same scheme. For example, the Eurovision Song Contest uses
a1 = 12, a2 = 10, a3 , = 8, a4 = 7, a5 = 6, a6 = 5, a7 = 4, a8 = 3, a9 = 2, a10 = 1, a11 = . . . =
am = 0, where agents correspond to countries participating in the vote.
While in our example above, indeed the Condorcet winner wins and the Condorcet loser
loses, this is generally not true for any of these rules. Indeed, no such rule fulfills the Condorcet
winner criterion. Consider the following example, in which B wins Borda count, despite the fact
that A is the Condorcet winner.

60 % : A i B i C
40 % : B i C i A

We could also ask at this point whether the Condorcet winner criterion is actually desirable.
In this particular example, A is polarizing: Some agents really dislike him, so B might be a
reasonable compromise. The reason is that the criterion only makes a binary comparison. It
does not matter how much an agent likes one candidate better than the other.

3.3 Copeland Rule

Yet another approach is to perform pairwise comparisons and to count how many a candidate
wins. The Copeland score of a candidate is defined to be the difference of how many pairwise
comparisons this candidate wins and how many he loses. So, it is an integer in the range of
−(m − 1) to m − 1. The candidates are then ranked by decreasing Copeland score.
This rule fulfills the Condorcet winner and loser criteria. A Condorcet winner always has
the highest possible score of m − 1, a Condorcet loser always has score −(m − 1), which is the
lowest possible score.

4 Strategic Vulnerability
As we will show, it is not a coincidence that agents want to lie about their preferences in the
above voting rules. Indeed, there are theorems by Arrow (1950) and Gibbard and Satterthwaite
(1973/1975), which show that every ranking or voting rule has these deficiencies unless it is a
dictatorship.
For the remainder of today, we will show Arrow’s theorem that deals with ranking rules.
That is, given the preferences (1 , . . . , n ), we have to come up with a combined single ranking
> expressing the preferences of society.

Definition 24.2. A ranking rule is strategically vulnerable if there are an agent i, a preference
profile (1 , . . . , n ), an alternative preference report 0i , and two candidates A and B such that
A i B and B > A but A >0 B, where > is the social ranking under (1 , . . . , n ) and >0 is the
social ranking under (1 , . . . , i−1 , 0i , i+1 , . . . , n ).

The first observation will be that we can rephrase the property of strategic vulnerability
using the following property.

Definition 24.3. A ranking rule is independent of irrelevant alternatives (IIA) if for any two
candidates A and B, the fact whether A is ranked before or after B in society’s ranking only
depends on how the agents rank A and B against each other (in pairwise comparisons). That is,
given two preference profiles (1 , . . . , n ) and (01 , . . . , 0n ) such that A i B whenever A 0i B,
then A > B if and only if A >0 B.
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 5 of 8)

B B B B B B B B B B

··· → ···

B B B B B B B B B B
B ranked last B ranked first

Figure 1: Candidate B is always ranked first or last in the society ranking under such preference
profiles and IIA. If more and more agents change their preference, there has to be a transition
from last to first. Therefore, a single agent can make the difference

This also seems be a natural property. Interestingly, any ranking rule that is not strategically
vulnerable has to fulfill it.

Lemma 24.4. If a ranking rule does not fulfill IIA, it is strategically vulnerable.

Proof. If the ranking rule does not fulfill IIA, there have to be (1 , . . . , n ) and (01 , . . . , 0n )
such that A i B whenever A 0i B but A > B and B >0 A.
We consider the sequence of n + 1 preference profiles (1 , . . . , t , 0t+1 , . . . , 0n ). In one of
the n steps in between, the relative ranks of A and B have to switch. Let this switch happen
from i to i + 1. That is, A is ranked before B on (1 , . . . , i−1 , i , 0i+1 , . . . , 0n ) but after on
(1 , . . . , i−1 , 0i , 0i+1 , . . . , 0n ).
If B i A, agent i can strategically misreport 0i instead of i . Otherwise, we have A 0i B
and he can misreport i instead of 0i .

5 Arrow’s Impossibility Theorem

Clearly, every dictatorship fulfills IIA because we simply use the preferences of one single agent.
Another way of fulfilling IIA is to return an arbitrary order, which does not depend on the
agents’ preferences at all. As a matter of fact, these are essentially the only ranking rules that
fulfill IIA. Every non-trivial ranking rule that you will come up with does not fulfill IIA. This is
the statement of Arrow’s theorem.

Definition 24.5. A ranking rule is unanimous if the following condition holds. If candidate A
is preferred to B by every agent i, then this candidate gets a better rank in the overall ranking.
That is, if A i B for all i, then A > B.

Now, we are ready to state the theorem.

Theorem 24.6 (Arrow’s theorem). A ranking rule for three or more candidates fulfills unanimity
and IIA only if it is a dictatorship.

The general idea of the proof is not too complicated. One considers preference profiles with
the property that the is one candidate B who is ranked highest or lowest by each agent (but not
consistently so). By IIA, candidate B will be either highest or lowest in the society ranking as
well. Because of unanimity, there is a preference profile such that a change in the preference of
one agent moves B from the bottom to the top.
This observation has an interesting consequence: The agent who decides whether B is ranked
first or last also determines the relative ranking of all other candidates. This is due to the fact
that we could move B between any two other candidates in his preference. Due to IIA, this
would not change their relative ranking.
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 6 of 8)

6 Outlook
As we have seen, there is a plethora of different voting rules. In each of them, agents sometimes
have an incentive to misreport their preferences. Crucially, the counterexamples are often cyclic.
That is, there are preferences of the forms A i B i C, B i C i A, and C i A i B. One
can show, for example, that there are non-trivial strategy-proof voting rules if preferences have
more of a structure. Most prominently, they could correspond to a left-right political spectrum
and therefore be not cyclic.
Regarding other properties of voting rules, there are some that are less disputable than
others. There is clearly no single best voting rule, simply because there are multiple ways to
define who should be the winner.

Further Reading
• Chapter 13 in the Karlin/Peres book

• K. Arrow, Social Choice and Individual Values, 1951

• A. Gibbard, Manipulation of voting schemes: A general result, Econometrica, 1973

• M. Satterthwaite, Strategy-proofness and Arrow’s conditions: Existence and correspondence

theorems for voting procedures and social welfare functions, Journal of Economic Theory,
1975

A Bonus: Proof of Arrow’s Theorem

Definition 24.7. A candidate B is polarizing with respect to a preference profile if each agent
i ranks B highest or lowest.

Lemma 24.8. Consider a ranking rule that fulfills unanimity and IIA. If there is a polarizing
candidate B in the strategy profile, then the ranking rule ranks B highest or lowest.

Proof. We prove the claim by contradiction. Suppose, it is false. That is, given a preference
profile (1 , . . . , n ) such that each agent i ranks B highest or lowest, the outcome of the ranking
rule is such that there are A and C for which A > B > C. We now derive a preference profile
(01 , . . . , 0n ) on which we will apply IIA.
For every i, we distinguish the relative rank of A, B, and C in i . There are the following
four cases: (i) B i A i C, (ii) B i C i A, (iii) A i C i B, (iv) C i A i B.
In case (i), define 0i such that B 0i C 0i A; in case (iii), define 0i such that C 0i A 0i B;
in the other cases, define 0i to be simply i .
What ranking >0 does the ranking rule decide on (01 , . . . , 0n )? By IIA, we have to have
A >0 B and B >0 C because (01 , . . . , 0n ) keeps the positions of B relative to both A and C.
Furthermore, by unanimity, C >0 A because every agent prefers C to A in the new preferences.
This is a contradiction, so either A or C cannot exist.

Definition 24.9. Given a candidate B, we call an agent i B-pivotal if there is a preference

profile (∗1 , . . . , ∗n ) and an alternative preference ∗∗
i such that B is polarizing and ranked lowest
under (∗1 , . . . , ∗n ) and polarizing and ranked highest under (∗1 , . . . , ∗i−1 , ∗∗ ∗ ∗
i , i+1 , . . . , n ).

Lemma 24.10. Consider a ranking rule that fulfills unanimity and IIA. For every candidate B,
there is a B-pivotal agent.
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 7 of 8)

(0) (n)
Proof. We consider a sequence of n + 1 different preference profiles (i )i∈N , . . . , (i )i∈N . In
(0) (n) (t−1)
(i )i∈N , all agents rank B lowest, in (i )i∈N , all agents rank B highest. From (i )i∈N
(t) (0) (n)
to (i )i∈N , we switch agent t’s preferences from t to t .
(t)
By Lemma 24.8, candidate B is ranked highest or lowest in the outcome on every (i )i∈N .
(0)
By unanimity, he is ranked lowest in the outcome on (i )i∈N and highest in the outcome
(n)
on (i )i∈N . At some point t, there has to be a transition such that B is ranked lowest in
(t−1) (t)
(i )i∈N and highest in (i )i∈N . These profiles fulfill the claim.

Lemma 24.11. Consider a ranking rule that fulfills unanimity and IIA. Let B be any candidate
and let i be the agent with the property guaranteed by Lemma 24.10. Then i is a dictator on
Γ \ {B}. That is, for A, C 6= B, we have A > C if and only if A i C.

Proof. It suffices to show that if A i C then also A > C. We will construct a preference profile
(01 , . . . , 0n ) and apply IIA. Note that by IIA everything that matters are the relative positions
of A, B, and C. In other words, we can assume without loss of generality that Γ = {A, B, C}.
We use the preferences (∗1 , . . . , ∗n ) and ∗∗ 0 0 0
i to construct (1 , . . . , n ). For every i 6= i,
we know that B is either ranked first or last in ∗i0 . Construct 0i0 by moving B to the same
position as in i0 and keeping the relative order of A and C fixed. Furthermore, choose 0i as
A 0i B 0i C.
First of all, by IIA considering candidates A and C, the relative order of A and C in the
outcome on (01 , . . . , 0n ) has to be the same as on (1 , . . . , n ). Next, by IIA but this time
regarding candidates A and B, A has to be ranked before B on (01 , . . . , 0n ) because the relative
order is just as in (∗1 , . . . , ∗n ). Finally, B has to be ranked before C on (01 , . . . , 0n ) because
their relative order is like in (∗1 , . . . , ∗i−1 , ∗∗ ∗ ∗
i , i+1 , . . . , n ).

Proof of Theorem 24.6. Consider three distinct candidates A, B, and C. By Lemma 24.11, we
know that there is a B-pivotal agent i, who is a dictator on Γ \ {B}. There is also a C-pivotal
agent i0 , who is a dictator on Γ \ {C}. We claim that i = i0 .
To see this, consider the preference profile (∗1 , . . . , ∗n ) and the preference ∗∗
i that make
i a B-pivotal agent. Suppose that i = 6 i0 . Then A would be ranked higher than B if and only
if A ∗i0 B. However, what makes i B-pivotal is the fact that only his preference changes and
makes B move from bottom to top, whereas the preferences of i0 do not change. This is a
contradiction.

B Bonus: Gibbard-Satterthwaite Theorem

There is a similar statement regarding voting rules. In this case, the “truthfulness” property is
called strategy-proofness. It is defined as follows.

Definition 24.12. A voting rule is strategy-proof if for all preference profiles 1 , . . . , n , all
agents i, and candidates A and B the following holds. If A i B and B wins under 1 , . . . , n ,
then A does not win under any false report 0i of agent i.

So, in words, this definition requires that no agent can enforce an outcome that he likes
better by lying. This, once again, is conceptually the same as truthfulness in mechanism design
with money.
The following theorem was discovered by Allan Gibbard (1973) and Mark Satterthwaite
(1975) independently.

Theorem 24.13. If a voting rule for three or more candidates is onto (that is, every candidate
can be elected) and strategy-proof, then it is a dictatorship. That is, there is some agent i such
that always agent i’s most preferred candidate wins.
Algorithmic Game Theory, Summer 2019 Lecture 24 (page 8 of 8)

The assumption that the voting rule is onto mirrors unanimity in Arrow’s theorem. If a
ranking rule fulfills unanimity, then, by definition, it can output every possible ranking, namely
if all agents agree on this one. For a voting rule to be onto, a similar property is sufficient: If all
agents agree that one candidate is the best, then this candidate gets elected.
We will not prove this theorem because the proof is not very enlightening on the technical
level. The general argument is to use Arrow’s theorem and to show that a voting rule fulfilling
the assumptions of the theorem would be a contradiction to Arrow’s theorem.
In more details, such a voting rule could be turned into a ranking rule as follows. The fact
whether A > B or not is determined by whether A would win a vote against B if they were both
moved to the top of the preference profiles. From the assumptions, we get that this ranking
rule fulfills unanimity and IIA. If the voting rule that we started from is not a dictatorship,
then the created ranking rule isn’t either: For every agent, there are preference profiles, such
that some part of the social ranking does not correspond to the individual preference. This is a
contradiction, therefore the voting rule cannot exist.