To SOLVE pure nash equilibrium:

ss
Lecture:

Social Choice Theory

●​ What are the properties of different voting systems?
●​ Which voting system best represents the group’s preferences>
○​ Can we even make a well-defined group ranking?
●​ What does it mean for a voting system to have different choices

E.g
●​ Three friends: Brad, Brenton, and Peter
●​ Where to go to lunch? Chipotle, Panera, or SatCo
●​ How does the group rank these three options

2. How does an individual rank options?

Model of choice :

Ingredient of the problem:

1.​ A set of actions that are available to the individual (or decision maker)
2.​ A specification of the individual’s preferences

A set is a collection of objects. The objects that make up a set are called the elements or
members of the set.

●​ We write a set as {Apple, Banana, Orange}

○​ The order does not matter
○​ Often use capital letters to denote the set: A = {Apple, Banana, Orange}
●​ We denote a member of set A as a belongs to A

Preferences
As individual’s preferences determines the individual’s ranking of the alternatives

That is, given two alternatives, which one do I prefer?

Preference relation:
A preference relation on a set A specifies, for two alternatives, x, y belong A, does the
individual:
●​ 1. Prefer x over y?
●​ 2. Prefer y over x?
●​ 3. Or is indifferent between x and y

Denote a preference relation as

Requirement to Rational Preferences

A preference relation is rational if it satisfies:

1.​ Completeness: for all alternatives, either x is preferred to y, y is preferred to x, or x

and y are indifferent.
2.​ Transitivity: for all x, y, z if x preferred to y, and y preferred to z, then x is preferred to
z

Key assumptions:
●​ Individuals are ration
●​ Act to maximizes preferences

Why do we make these assumptions?

●​ TO make predictions about how actors will behave
●​ Infer preferences from observed behaviours

Rationality= acting to maximise own consistent preferences

●​ Rationally =/ acting in accordance to normative or moral principle

Aug 28th 2024

Majority Rule and Other Voting Methods

Utility Functions:
Supposed that a voter prefers the D over R. This preference is represented by both u(DEM)
= 10, u(REP) =4 and u(DEM)=2, u(REP)=0.5

Representing a given preference relation is an ordinal property

●​ Only the ordering of alternatives matters
●​ Nothing about the intensity of preferences

Example: Utility functions over {Apple, Banana, Carrot}

Utility Functions II:

If u represents an individual’s preferences and v is another utility function for which v(a) >
v(b) when’re u(a) > u(b) - - > v also represents the DMs preferences

The numerical values of utility are cardinal properties

●​ If the exact numbers are important - different utility functions do canta in
different implications
●​ We will return to this with expected utility
Majority Rule
●​ The group prefers option x over option y if at least half of the member of the
group prefer x over y

Condorcet’s paradox:
Under majority rule, rational individual preferences do not need to lead to rational social
preferences. With > 2 options we can have intransitive social preferences.

Condorcet Winner
We say an alternative is a Condorcet winner if it wins in a pair wise majority rule vote
against every other alternatives
●​ If a condorcet winner exists, its naturally the one to choose

Goal: voting rule should deliver consistent group preference, we want a rule that leads to a
rational social preference relation- - Complete and transitive
Other Voting Rules

For > 2 options majority rule can create cycles, what other rules might work better?

Plurality Rule: Everyone votes for their most preferred option, the winner is the alternative
that gets the most votes (even if it does not get a majority)

E.g 7 friends with the following preferences over lunch spots {P,C,S}:
●​ 3 voters: S > C >P
●​ 2 voters: C > D > P
●​ 2 voters: P > C > S

Result: 3 votes for S, 2 votes for C, and 2 votes for P - S wins

●​ Social preference relation: S>C ~ P
●​ Result is nice and transitive

Ranked Choice Voting

Process:
1.​ Voters rank every candidate
2.​ If someone gets more than 50 percent of 1st place voters they win
3.​ If not, drop the candidate with the fewest first place votes
4.​ Ballots that picked the loser go the second choice
 5.​ Continué until someone wins a majority

Recap
Created cycles in the group preference (not transitive)
●​ Majority rule

Open to manipulation by introducing “irrelevant” alternatives

●​ Plurality rule
●​ Ranked choice voting

Criteria
We want to find a voting rule that satisfies
1.​ Rationality: The voting rule should yield a complete and transitive social preference.
2.​ Independence of irrelevant alternatives: For two preference profiles (1, 2…n) and
(1’,2’….n’) if all individuals have the same preference between a and b under both
profiles then the voting rule should yield the same social preference over a and b
under both profiles.
3.​ Universal Domain: The voting rule specifies a group preference for any given set of
individual preferences. We do not want to rule out any preferences that an individual
hold.
4.​ Unanimity: If every individual prefer alternative a over alternative b then the voting
rule should yield a group preference for a over b

A bit of formalization
1.​ We can think of a voting rule as a function

2.​ It takes as input the preferences of each individual and it outputs a group preference

3.​ Our goal is to find a voting rule (function) that has all four properties

Theorem (Arrow’s Theorem)

Assume there are 3 or more alternatives. A voting rule
satisfies universal domain, rationality, unanimity, and IIA if
an only if it is a dictatorship.
●​ No matter what tule you pick to make a decision in a group it must violate one of
these properties (or be a dictatorship)
●​ This only says that every (non-dictator) rule can fail these properties not that they
always do

Pairwise Majority rule & Arrow

We know majority rule fails rationality, what about unanimity? And Independence of
irrelevant alternatives?
 ●​ Satisfies unanimity, and IIA

Other Considerations/Example Preferences

Say we have our 7 friends with their preferences over lunch spots:
●​ 3 voters: s>c>p
●​ 2 voters: c>p>s
●​ 2 voters: p>c>s

Plurality Rule:
Under plurality rule we have S: 3 votes, C: 2 votes, P: 2 votes- S wins

Borda Count:
We have n alternatives: A = (a1, a2,……an)
●​ Each voter list the alternatives in the order of their preference
●​ An alternative gets 0 points for each time it is listed last, 1 point each time it is listed
second to last, and n-1 points each time it is ranked first
●​ The social preferences ranking over alternatives is determined in order of the points.

Strategy-Proofness

Definition: We say a voting rule is strategy proof if there does not exist a
preference profile (1>2>3…>n), and individual i, and a preference relation
>~ such that i prefer the outcome under one profile over the outcome
under another profile.

There are not cases where an individual is better off under the rule reporting a preference
different than their true preference

Theorem (Gibband-Satterhwaite)
Assume there are three or more alternatives and each individual is
allowed to have any ranking of there’s alternatives. The only voting rule
that satisfies strategy-proofness and unanimity is dictatorship.

(Recall unanimity: If every individual prefers alternative x over alternative y then the voting
rule should yield a group preference for x over y)

Arrow and Gibbard-Satterthwaite

Arrow’s theorem covers a very broad number of voting rules

●​ Very few restrictions

However, no accounting for intensity of preferecnes

Gibbard-Satterthwaite also applies to rules that use intensity

Indeed, rules that involve expressing intensity of preferences tend to be especially

susceptible to manipulation

Approval Voting

Voters list all alternatives that they “approve” - rather than just listing one alternative - the
alternative with the highest approval rating wins.

E.g:
●​ North Dakota for city local elections
●​ ELecting the pope(1294-1621)
●​ Secretary-General of the United Nations

A few ways to implement/model approval voting

Assign utilities to voters over three candidates {D,R,I}

●​ 2 voter: u(D) = 6, u(I)=2, U(R) = 1
●​ 1 voters: u(D) = 6, u(I) = 4, u(R) = 1
●​ 3 voters: u(I) = 5, u(D) = 4, u(R) =1
●​ 3 voters: u(R) = 5, u(I) = 2, u(D) = 0
●​ 1 voters: u(R) = 5, u(I) =4, u(D) = 0

“Approve if utility is above a threshold, e.g 3”

D: 2+1+3=6
R: 3+1=4
I:3+1+1=5

Score Voting
Voters ates each alternative on a scale of 1-10 for example. Sum up the ratings for each
alternative and the alternative with the highest score wins.

Budget Systems
Each person gets some number of votes they can allocate across the alternatives, including
voting for the same option multiple time.

Quadratic voting: Similar, but voting for the same option multiple times costs an increasing
amount of one’s budget.

Intensity of Preferences
Systems use cardinal information about preferences; Leverage more information than
ordinal systems ; Can circumvent Arrow, but clearly open to manipulation; Also requires
interpersonal utility comparisons
Universal Domain
Arrow’s theorem assumed that preferences were unrestricted
●​ Universal domain property

Sometimes reasonable to constrain what preferences can look like

Assume that alternatives lie along on dimension
We can order alternatives a<b<c<….

Single-peaked Preferences
Def: Let X be a set of ordered alternatives. We say an individual i has single-peaked
preferences if there exists xi { X such that
1.​ For all y=/ Xi, we have xi > y
2.​ For all y, z if xi<y<z then yi>z
3.​ Fir all y, z if y<z<xi then z>y
This restriction on preferences buys us leverage

Median Voter Theorem

Theorem (Median Voter Theorem)
If preferences are single-peaked then a Condorcet winner exists.
The Condorcet winner is the median voter’s most preferred candidate.

With single-peaked preference majority rule now satisfies rationality

We avoid Arrow if we are willing to restrict domain.

The MVT has strong implications for elections

Median voter is important

●​ Political candidates should be drawn to the center?
Return to this with game-theoretic models of elections

Chaos theorem
Often times many policy issues or dimensions
●​ But maybe one-dimension a good simplification
Unfortunately, the MVT does not translate to multiple dimensions even with single-peaked
preferences.

Radial Symmetry
Definition: Preference satisfy radial symmetry if there is a voter with ideal point xi, such that
each other ideal point can be paired with an ideal point on the opposite side of xi, by drawing
a line connecting them and this line intersects with xi.
Radial-symmetry is a “knife-edge” condition, any small change to any voter’s preference
violate the condition. In a way it is exceedingly unlikely to be true.

Theorem (Mckelvey-Schofield-Plott)
If there is more than one dimension and radial symmetry does not hold, then under majority
rule there does not exist a point that is unbeaten by any other point.

The theorem essentially always true

●​ Radial symmetry almost never holds
The theorem implies that starting from any policy we can reach any other policy by majority
rule votes
●​ i.e Chaos
This implies a lot of power for the agenda setter
No matter the initial status quo I can keep proposing policies until we reach my preferred
policy
Intro to Game Theory
Strategic Interactions
●​ A voter may want to misreport preferences to manipulate outcome of election
○​ But if other voters realise this then they may change how they vote too
●​ When there are multiple policy issues whoever proposes policy can get any outcome
they want under majority rule
○​ But other legislators anticipate that approving a policy today may lead to a
worse outcome in the future

Want to think about strategic interactions

Think about how my payoff depends on my action and the actions of other people

Simultaneous Move games

Start with simultaneous move games

●​ also called one-shot, strategic, or normal form agmes

There is no time in the model

We each take out action simultaneously

A strategic form game consists of

1.​ A set of players
2.​ For each player, a set of actions
3.​ For each player, a utility function over the set of action profile

Players

First ingredient of a strategic form game is the set of players

Often we denote the set of players by N, e.g.,

●​ N={1,2}
●​ N={1,2,…,n}
●​ N={Politician, Bureaucrat}
●​ N={Peter, brad, Brenton}

We typically use i or j to represent an arbitrary player, e.g i belongs to N

Actions

The second ingredient are the actions available to each player

For each player i belongs to N, we denote their actions Ai, e.g.

 ●​ In the presoners’ dilemma, each player has two actions: cooperate © and defect (D),
Ai = {C,D} for i ~ {1,2}
●​ Election model: three available candidates Ai = {R,D,I}

We represent an arbitrary choice of action by player i as ai ~ Ai

An action profile is an ordered list containing an action choice for every player in the game, a
= (a1,a2,,,,,an)

Denote collection of all possible action profiles as the set A

●​ A is all of the possible “outcomes” of the game

e.g
N={1,2}; Ai={R,P,S}

Utility Functions
Finally, each player i has an associated utility function (us : A —- R)

ui assigns a payoff to i for each possible action profile a ~ A

●​ A utility function specifies a player’s payoff from each outcome

Often write player i’s utility as a function of player i’s action, given the actions of the other
players
●​ ui(a) = ui(ai,…,ai,…,an)
●​ More compact: a = (ai,a-i) where ai is player i’s action and a-i is the profile of actions
taken by all players other than player i then ui(a) = (ai,a-i)

Prisoner’s Dilemma
Two people captured by the police held in separate cells

Police need a confession to convict and offer following deal:

●​ One confesses and other stays quiet = > 1 year for the confessor and 5 years for
quiet
●​ If both stay quiet police cannot convict = > 0 years for both
●​ If both confess = > 3 years for both

Prefer a lower number of years in prison

●​ Players: N = {1,2}
●​ Actions: Ai = {C,D} for each i ~ N
○​ I can cooperate or defect
 ●​ Utilities:
○​ u1(C,C) = 3, u1(C,D) = 0, u1(D,C) = 5, u1(D,D) = 1
○​ u2(C,C) = 3, u2(C,D) = 5, u2(D,C) = 0, u2(D,D) = 1

Public goods contribution

A group of n people can choose to contribute money to a project

●​ e.g. Built a park

If enough money is donated then the project is completed

●​ if more than T dollars then project finished
●​ otherwise money is wasted

We all benefit form the park if it is built but costly to donate

●​ Players: N = {1,2,….,n}
●​ Actions: Ai = [0,Bi] where Bi is player’s budget
●​ utilities: Let T >0 represent the total cost that has to be paid for the project to
success. With this, the utilit to each player i ~ N given action profile is :

ui(a) =

●​ 1- ai, if a1+a2+…..an >= T

●​ -ai if a1+a2+…..+an<T

Representing Two-Player Games

Only two players

Only a small number of actions for each player

=> useful to represent the game with a matrix

●​ e.g., prisoner’s dilemma

Strictly Dominated Actions

Definition:
In a strategic game player i’s action ai’’ strictly dominates action ai’ if

ui(a’’i, a-i) > ui(a’i, ai)

for every list a-i of the other players’ actions

An action strictly dominates another if it gives a higher payoff no matter what the other
players choose to do
We would expect a player to never use an action that is strictly dominated.

Practice Games
●​ Two friends trying to meet for lunch
●​ Three voters casting ballots
●​ N-player prioners’ dilemma

Which actions are strictly dominated?

Game 1:

Peter/Brad C S P

C 6,3 4,4 4,3

S 3,1 5,6 3,3

P 1,1 1,4 3,5

C is strictly dominating P for peter

Game 2:
●​ 3 voters
○​ 2 Republicans
○​ 1 Democrats
●​ Actions:
Ai = {D,R}
●​ Majority Rule
●​ u0(D) = 1; u0(R) = 0
●​ u1(D)=0; u1(R)=1

Information and Strict Dominance

strict dominance is not very demanding

It doesn’t even require me to know anything about the other player’s preferences, or make
conjectures about how they will behave

Iterated Elimination of Strictly Dominated Strategies (IESDS)

If an actor is strictly dominated then a player should not use it

Then if a player has a stictly dominated action this should be clear to other players, so they
should reason that the first player will not use the action

But now the other players may have strictly dominated actions if they assume the first player
will not use a sticky dominated action

We can see this is the lunch meetup game

Very restricted can’t tell result

Game of Chicken
●​ N={1,2}
●​ Si={swerve, straight} For i~{1,2}
●​ u1(swerve,swerve) = 0; u1 (swerve,straight) = -1
●​ u1(straight,swerve) = 1, u1(straight, straight) =-10

●​ u2(swerve,swerve). = 0; u2(swerve,straight) = 1
●​ u2(straight, swerve) = -1; u2(straight,straight)=-10
Expected Utility & Mixed Strategy Nash Equilibrium

Game 1: Matching Pennies Game

●​ Players: N = {1,2}
●​ Actions: Ai = {H,T}
●​ Utlities:
○​ u1(H,H)=1,u1(H,T)=-1,u1(T,H)=-1,u1(T,T) = 1
○​ u2(H,H)=-1,u2(H,T)=1,u2(T,H)=1,u2(T,T)=-1

Neither boxes have player playing both best responses

Expected Utility - depending on circumstances

Probability:
●​ Say an event happened many times, sometimes it has one outcome, other times
another
○​ rain or not
●​ Refer to the proportion of times an outcome occurs as the probability of that
outcomes
 ●​ Often denote the probability of outcome A as Pr(S) or use ps and qs for
probabilities

Probability distribution:
The assignment of probabilities to outcomes is the probability distribution

E.g., the weather today has 3 outcomes: {Sunny, Cloudy, Rainy}

1.​ An example of a probability a distribution here assigns probability ½ to sunny, ¼ to
cloudy, and ¼ to rainy
2.​ Another probability distribution assigns ⅓ to sunny, ⅓ to cloudy, and ⅓ to rainy

For any probability distribution:

1.​ The sum of probabilities for all possible outcomes is 1
2.​ The probability of each outcomes is >=0, <=1

Decisionmaking under Uncertainty

Consider a decision maker with a set of actions

Each action leads to a probability distribution over outcomes
Examples:
●​ If I buy a stock its value could double with probability ¼, stay the same with
probability ½, or go to 0 with probability ¼
●​ If I run for an election I win with probability ½ and lose with probability ½

Expected Utility

I am going to swin and the weather could be sunny, rainy, or cloudy

There is a probability fo ½ that is sunny, ¼ is rainy, and ¼ to be cloudy

My utility for swimming depends on the weather:

●​ u(swim|sunny) = 1; u(swim|cloudy)=0, and u(swim|rainy)=-1
●​ I prefer to swim in the sun over swim when it is cloudy over swim when it is rainy

Now we want know what is my expected utility from swimming：

Eu(Swim) = Pr(Sunny)*u(sunny)+Pr(Cloudy)*u(Cloudy)+Pr(rainy)*u(rainy)

Mixed Strategies:

A mixed strategy of a player in s strategic form game is a probability distribution over the
player’s actions

e.g Matching pennies, rock-paper-scissors, security dilemma

Suppose player i has actions Ai={a1,a2,….,ak} then a mixed strategy for i is a probability
distribution pi={pi1,pi2,….,pik}
Notice that we can get back pure strategies by setting choosing same action with probability
1.

IN a strategic form game, the profile of mixed strategies is a Nash equilibrium if and only if
player’s mixed strategy is a best response to the other player’s mixed strategy that is,

EU(p*i,p*-i) >= EU(p’9,p’-i)

for all i belongs to N, where p’i is any other possible mixed strategy for player i.

Finding MSNE
●​ If a player uses actions A and B with >0 probability in a MSNE then must be
indifferent between A and B, given the strategies of the other players
○​ i.e., EU(A|p-i) = EU(B|p-i)
○​ If not, just use action that gives the higher expected utility
○​ This gives us a way to find MSNE(mixed strategy Nash equilibrium)

E.g

Player1/Player 2 H (q prob) T (1-q prob)

H 1,-1 -1,1

T -1,1 1,-1

Eu(H|q) = q*u(H,H) +(1-q)*u(H,T) = q*1+(1-q)*(-1)=q-1+q=2q-1

Eu(T|q)=q*u(T,H)+(1-q)*u(T,T)=q(-1)+(1-q)1=1-2q

IN order to be indifferent, q = ½

E.g. 2.

Seahawks/Patriots G N

Run .4,.6 .8,.2

Pass .6,.4 .2,.8

Eus(Run|q) = q*.4+(1-q)*.8=-.4q+.8
Eus(Pass|q)= q*.6+(1-q)*.2=.4q+.2

IN order to be indifferent: q* = ¾
Lecture 7： Strict Dominance Continued & Nash Equilibrium
Game examples on notability

Information and strict dominance

-​ strict dominance is not very demanding
-​ it does not even require me to know anything about the other players’ preferences, or
make conjectures about how they will behave

Iterated elimination of strictly dominated strategies (IESDS)

-​ if an action is strictly dominated, then a player should not use it
-​ then, if a player has a strictly dominated action, this should be clear to other players
-​ so they should reason that the first player will not use the action
-​ but now the other players may have strictly dominated actions if they assume the first
player will not use a strictly dominated action
Limitations
-​ strict dominance is limited
-​ often cannot make a prediction
-​ even IESDS often fails to predict an outcome

Lecture 11: Expected Utility and Mixed Strategy NE

Probability of outcome A as Pr(A)

Probability distribution is the assignment of probabilities to outcomes

Expected utility
-​ cardinality(numerical) properties matter, not just the ordering

Randomization
-​ there might not always be a NE
-​ we need to extend our concept of how players can behave
-​ allow players to randomize over actions

Mixed strategies

-​
-​ Suppose player i has actions Ai = {a1, a2,..., ak}, then a mixed strategy for i is a
probability distribution pi = {p1, p2,..., pk}
-​ we can get back to pure strategies by setting choosing some action with probability 1

Mixed strategy Nash Equilibria

 -​
-​ if a player uses actions A and B with >0 probability in a MSNE, then must be
indifferent between A and B, given the strategies of the other players
-​ if not, just use action that gives the higher expected utility

Lecture 8: Nash Equilibrium

Stability
-​ the outcome we predict should be stable
-​ no player should be able to do better by changing actions
-​ this is the idea of Nash equilibrium: an outcome a* = (a*1, ..., a*n) with the property that
no player i can do better by choosing an action different from a*i, given that every
other player j adheres to a*j.

Nash Equilibrium definition

-​ no one has a profitable deviation

-​ check if there is a profitable deviation holding fixed the actions of every other player

Lecture 9: Nash Equilibrium & Best Response

Functions
Strict dominance does not require me to form any conjectures about how you will behave or
to even know your payoffs
-​ nash equilibrium is more demanding
-​ i need to know your actions and payoffs, you need to know my actions and payoffs, I
need to know that you know that I know your actions and payoffs, etc
-​ additionally, I need to have beliefs about how you will play the game
-​ chicken game
-​ could come from experience

Best responses
-​ in simple games, analyzing the payoff matrix to find Nash Equilibria works well
-​ in more complicated games, it is useful to think about Nash Equilibria in terms of best
response
-​ given the other player is taking some action, which of my actions gives me the
greatest payoff?
Definition of Best response

-​
-​ So fixing the actions of every other player (a−i), player i’s best response is the
action(s) that gives the highest payoff
-​ Note: may have multiple best responses

Nash Equilibria as Best Reponses

-​

Lecture 10: Best Response Functions & Expected

Utility
Best Response function definition
 -​
-​ the best response function gives a set
-​ player i’s best response is a function of what the other players do
-​ must be the case since we can have many best responses to the actions of the other
players
Nash equilibrium and BR Functions

-​
-​ A nash equilibrium is an action profile where players use mutual best responses

Downsian Model of Elections