Chapter 7: Social Choice and Voting

Algorithmic Game Theory
Social choicePlurality and majorityBorda and approvalInstant-runoff votingCondorcet and CopelandScoring rulesCondorcet cyclesKemenyStrategic votingArrow and Gibbard–SatterthwaiteMay's theoremSingle-peaked preferencesComputational social choice
Sercan KÜLCÜ · Giresun University · Computer Engineering

Social Choice

Warm-up: The Condorcet Paradox

The rules

Each voter ranks the candidates. Every pair is decided by pairwise majority, and a candidate who beats every rival is the Condorcet winner.

Think about it
  • If every voter's ranking is consistent, must the group's preference be consistent too?
  • If you set the order of the votes, could you make any candidate you like win?
1ABC2BCA3CABvoterhigher candidateis preferred???ABCby pairwise majoritywho beats whom?

Social Choice

Learning Outcomes

CountingApply majority, plurality, Borda, approval, instant-runoff and Condorcet step by step to one preference profile.
ParadoxesShow a Condorcet cycle, the power of agenda order and IRV's failure of monotonicity; find a Kemeny ranking.
Strategic votingShow tactical voting under plurality and "burying" under Borda with concrete numbers.
TheoremsExplain what Arrow, Gibbard–Satterthwaite and May say; recognize single-peaked preferences as a way out.

Social Choice

What Is Social Choice?

  • Social choice studies rules that turn individual preferences into a collective decision or a social ranking.
  • The goal is not one universal definition of fairness, but to understand which criteria a rule satisfies and what trade-offs it makes.
  • Applications: elections, committee and club decisions, choosing a class event, sharing common resources, online polls.
input: preferences4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDf : profile → outcomemaps profiles to outcomes(plurality, Borda, IRV, …)Borda winner: Branking: B ≻ D ≻ A ≻ CSCF (social choice function)SWF (social welfare function)each column: one voter group's rankingballot = strategy
Mechanism design without money: no payments, only rankings.

Social Choice

A Sample Preference Profile

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD

A preference profile; every rule in this chapter is run on it.

13 voters, 4 candidatesVoters with the same ranking are grouped into one column.
Reading itColumn 1: 4 voters rank A ≻ D ≻ B ≻ C.
Majority thresholdMore than half of 13 votes: 7 votes.
Applied to the same profile, the rules give four different winners.

Voting Rules

Basic Voting Rules

RuleWhat does a voter report?Winner
MajorityOne candidateWhoever gets more than half of the votes (otherwise no winner)
PluralityOne candidateMost first-place votes; no majority needed
Borda countFull rankingWith m candidates, positions score m−1, m−2, …, 0; highest total wins
ApprovalSet of approved candidatesMost approvals
Instant-runoff (IRV)Full rankingLast place is eliminated and its votes transfer; repeat until a majority
CondorcetFull rankingThe candidate who beats every rival head to head (may not exist)

Note: "most votes wins" is plurality; majority requires more than 50% of the votes.

Voting Rules

Majority and Plurality Step by Step

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidatevotes
A?
B?
C?
D?

Under majority and plurality each voter reports only a first choice; the rest of the ranking is ignored.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidatefirst choice
A6
B0
C4
D3

Count first places: A = 4 + 2 = 6, C = 4, D = 3, B = 0.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidatefirst choice
A6
B0
C4
D3

Majority: 6 < 7 ⇒ no winner. Plurality: A has the most votes and wins (with just 6/13).

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidaterunoff
A6
B—
C7
D—

Two-round system (e.g. a presidential election): the top two, A and C, go to a runoff. A = 6, C = 7 ⇒ C wins.

Voting Rules

Borda Count Step by Step

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidatetotal
A0
B0
C0
D0

Borda with 4 candidates: 1st place gets 3, 2nd 2, 3rd 1, 4th 0 points. A group adds voters × points.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateG1total
A1212
B44
C00
D88

G1–G4: group contributions

Group 1 (4 voters, A ≻ D ≻ B ≻ C): A: 4×3 = 12, D: 4×2 = 8, B: 4×1 = 4, C: 4×0 = 0.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateG1G2total
A12012
B4812
C01212
D8412

G1–G4: group contributions

Group 2 (4 voters, C ≻ B ≻ D ≻ A): C: 4×3 = 12, B: 4×2 = 8, D: 4×1 = 4, A: 4×0 = 0.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateG1G2G3total
A120012
B48618
C012315
D84921

G1–G4: group contributions

Group 3 (3 voters, D ≻ B ≻ C ≻ A): D: 3×3 = 9, B: 3×2 = 6, C: 3×1 = 3, A: 3×0 = 0.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateG1G2G3G4total
A1200618
B486422
C0123217
D849021

G1–G4: group contributions

Group 4 (2 voters, A ≻ B ≻ C ≻ D): A: 2×3 = 6, B: 2×2 = 4, C: 2×1 = 2, D: 2×0 = 0.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateG1G2G3G4total
A1200618
B486422
C0123217
D849021

Totals: B = 22, D = 21, A = 18, C = 17 ⇒ Borda winner B, yet nobody ranked B first!

Voting Rules

Approval Voting

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateapproval
A?
B?
C?
D?

In approval voting each voter marks every candidate they accept. Assume each voter approves their top two.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateapproval
A6
B9
C4
D7

Approvals: A = 4 + 2 = 6, B = 4 + 3 + 2 = 9, C = 4, D = 4 + 3 = 7 ⇒ approval winner B (if everyone approves their top two, approval becomes 2-approval).

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateapproval (1st only)
A6
B0
C4
D3

The result depends on the threshold: if everyone approved only their first choice, approval = plurality and A would win.

Voting Rules

Positional Scoring Rules

A positional scoring rule fixes a vector s = (s1 ≥ s2 ≥ … ≥ sm) for m candidates; a candidate in position k gets sk points. Plurality, k-approval, Borda and veto all belong to this family.

s =11st02nd03rd04thpluralityA6B0C4D34 votersADBC4 votersCBDA3 votersDBCA2 votersABCD

s = (1, 0, 0, 0) (plurality): A = 6, C = 4, D = 3, B = 0 ⇒ winner A. Orange cells are the scoring positions.

s =11st12nd03rd04th2-approval (top two)A6B9C4D74 votersADBC4 votersCBDA3 votersDBCA2 votersABCD

s = (1, 1, 0, 0) (2-approval): B = 9, D = 7, A = 6, C = 4 ⇒ winner B.

s =31st22nd13rd04thBordaA18B22C17D214 votersADBC4 votersCBDA3 votersDBCA2 votersABCD

s = (3, 2, 1, 0) (Borda): B = 22, D = 21, A = 18, C = 17 ⇒ winner B.

s =11st12nd13rd04thveto (0 to last)A6B13C9D114 votersADBC4 votersCBDA3 votersDBCA2 votersABCD

s = (1, 1, 1, 0) (veto): B = 13, D = 11, C = 9, A = 6 ⇒ winner B.

s =101st52nd33rd04thanother vectorA60B57C55D624 votersADBC4 votersCBDA3 votersDBCA2 votersABCD

s = (10, 5, 3, 0) (another vector): D = 62, A = 60, B = 57, C = 55 ⇒ winner D. This vector picks the Condorcet winner D.

Voting Rules

Instant-Runoff Voting (IRV) Step by Step

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 1
A?
B?
C?
D?

Instant-runoff (ranked-choice) voting: each round counts every voter's top candidate still in the race; with no majority, the weakest is eliminated.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 1
A6
B0
C4
D3

Round 1: A = 6, B = 0, C = 4, D = 3. No majority (7 needed).

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 1
A6
B0
C4
D3

B has the fewest votes (0) and is eliminated. Nobody ranked B first, so no votes transfer.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 2
A6
B—
C4
D3

Round 2: A = 6, C = 4, D = 3. No majority (7 needed).

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 2
A6
B—
C4
D3

D has the fewest (3) and is eliminated. Its votes go to the next remaining choice: 3 votes D → C.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateround 3
A6
B—
C7
D—

Round 3: A = 6, C = 7. 7 ≥ 7 ⇒ C wins with a majority.

Voting Rules

Condorcet: Pairwise Comparisons Step by Step

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—0
B—0
C—0
D—0

Condorcet: for each pair, "how many voters prefer x to y?" Cell (x, y) = number of voters preferring x to y. 4 candidates ⇒ 6 comparisons.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—60
B7—1
C—0
D—0

A – B: 6 prefer A, 7 prefer B ⇒ B wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—660
B7—1
C7—1
D—0

A – C: 6 prefer A, 7 prefer C ⇒ C wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—6660
B7—1
C7—1
D7—1

A – D: 6 prefer A, 7 prefer D ⇒ D wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—6660
B7—92
C74—1
D7—1

B – C: 9 prefer B, 4 prefer C ⇒ B wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—6660
B7—962
C74—1
D77—2

B – D: 6 prefer B, 7 prefer D ⇒ D wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—6660
B7—962
C74—61
D777—3

C – D: 6 prefer C, 7 prefer D ⇒ D wins.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
x ╲ yABCDwins
A—6660
B7—962
C74—61
D777—3

D beats all three rivals 7–6: D is the Condorcet winner. A loses to everyone (the Condorcet loser), yet A won plurality!

Voting Rules

The Pairwise Majority Graph and Copeland

7–67–67–69–47–67–6DCopeland 3BCopeland 2CCopeland 1ACopeland 0
  • Each candidate is a node; an arrow x → y means a majority prefers x to y (label: vote counts).
  • The graph is a tournament: exactly one arrow between every pair (an odd number of voters rules out ties).
  • Condorcet winner = the node with no incoming arrows (D). Condorcet loser = the node with only incoming arrows (A).
  • Copeland rule: most outgoing arrows (wins) wins ⇒ D (3) ≻ B (2) ≻ C (1) ≻ A (0).
If every candidate has an incoming arrow, there is no Condorcet winner; a cycle through all the candidates causes exactly this.

Voting Rules

One Profile, Four Different Winners

RuleCountWinner
MajorityA 6 < 7none
PluralityA 6, C 4, D 3, B 0A
Two-round systemC 7 – A 6C
Instant-runoffC 7 – A 6 (round 3)C
BordaB 22, D 21, A 18, C 17B
2-approval (top two)B 9, D 7, A 6, C 4B
CondorcetD 3 winsD
KemenyD ≻ B ≻ C ≻ A (score 44)D
Lesson 1"The will of the people" depends on the rule: the rule shapes the outcome as much as the voters do.
Lesson 2Plurality can elect the Condorcet loser (A); IRV can eliminate the Condorcet winner (D).

Voting Rules

Code Trace: borda(ballots)

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A0B0C0D0

ballots = 4 groups (rows). The score dict starts empty and accumulates points per candidate.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A0B0C0D0

Group 1: 4 voters, ranking ADBC. m = 4 candidates: position k earns m − 1 − k.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B0C0D0

k = 0: A is 1st ⇒ p = 4 × (3 − 0) = 12; A total 12.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B0C0D8

k = 1: D is 2nd ⇒ p = 4 × (3 − 1) = 8; D total 8.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B4C0D8

k = 2: B is 3rd ⇒ p = 4 × (3 − 2) = 4; B total 4.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B4C0D8

k = 3: C is 4th ⇒ p = 4 × (3 − 3) = 0; C total 0.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B12C12D12

Group 2 (4 × CBDA): the inner loop runs 4 times; C +12, B +8, D +4, A +0.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A12B18C15D21

Group 3 (3 × DBCA): the inner loop runs 4 times; D +9, B +6, C +3, A +0.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A18B22C17D21

Group 4 (2 × ABCD): the inner loop runs 4 times; A +6, B +4, C +2, D +0.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A18B22C17D21

max: the top scorer is B (22), whom no voter ranked first. On ties max returns the first it sees; real rules need a tie-breaking rule.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDballots: row = voter group; 1st place 3 points … 4th 0A18B22C17D21

Complexity: n voter groups, m candidates ⇒ O(n·m). Counting is easy; the hard part is the rule's properties.

def borda(ballots):
  score = {}
  for count, order in ballots:
    m = len(order)
    for k, cand in enumerate(order):
      p = count * (m - 1 - k)
      score[cand] = score.get(cand, 0) + p
  winner = max(score, key=score.get)
  return winner, score

Voting Rules

Code Trace: irv(ballots, cands)

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA0B0C0D0majority: 2·tally > 13faded = eliminated

All candidates are in the race. Each round: count, check for a majority, otherwise drop the weakest.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA0B0C0D0majority: 2·tally > 13faded = eliminated

Round 1 begins: a zero counter for each remaining candidate.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA4B0C0D0majority: 2·tally > 13faded = eliminated

ADBC: top remaining candidate A ⇒ tally[A] += 4.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA4B0C4D0majority: 2·tally > 13faded = eliminated

CBDA: top remaining candidate C ⇒ tally[C] += 4.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA4B0C4D3majority: 2·tally > 13faded = eliminated

DBCA: top remaining candidate D ⇒ tally[D] += 3.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6B0C4D3majority: 2·tally > 13faded = eliminated

ABCD: top remaining candidate A ⇒ tally[A] += 2.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6B0C4D3majority: 2·tally > 13faded = eliminated

leader = A: 2 × 6 = 12 ≤ 13 ⇒ no majority.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6B0C4D3majority: 2·tally > 13faded = eliminated

B has the fewest votes (0) and is dropped; remaining = {A, C, D}. On ties min and max return the first they see; real rules need a tie-break.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA0BoutC0D0majority: 2·tally > 13faded = eliminated

Round 2 begins: a zero counter for each remaining candidate.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6BoutC4D3majority: 2·tally > 13faded = eliminated

Round 2 count: nobody ranked the eliminated B first, so the counts are unchanged.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6BoutC4D3majority: 2·tally > 13faded = eliminated

leader = A: 2 × 6 = 12 ≤ 13 ⇒ no majority.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6BoutC4D3majority: 2·tally > 13faded = eliminated

D has the fewest (3) and is dropped; remaining = {A, C}.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA0BoutC0Doutmajority: 2·tally > 13faded = eliminated

Round 3 begins: a zero counter for each remaining candidate.

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA4BoutC7Doutmajority: 2·tally > 13faded = eliminated

DBCA: top remaining candidate C ⇒ tally[C] += 3. (D is out, so these votes move to C.)

4 votersADBC4 votersCBDA3 votersDBCA2 votersABCDA6BoutC7Doutmajority: 2·tally > 13faded = eliminated

leader = C: 2 × 7 = 14 > 13 ⇒ return C.

def irv(ballots, cands):
  remaining = list(cands)
  while True:
    tally = {a: 0 for a in remaining}
    for count, order in ballots:
      top = next(a for a in order if a in remaining)
      tally[top] += count
    leader = max(tally, key=tally.get)
    if 2 * tally[leader] > sum(tally.values()):
      return leader
    loser = min(tally, key=tally.get)
    remaining.remove(loser)

Condorcet Paradox

A Condorcet Cycle Step by Step

rank4 voters3 voters2 voters
1.ABC
2.BCA
3.CAB
ABC

9 voters, 3 candidates. Each voter's ranking on its own is consistent (transitive).

rank4 voters3 voters2 voters
1.ABC
2.BCA
3.CAB
A ≻ B6 – 3ABC

A – B: 4 + 2 = 6 prefer A, 3 prefer B ⇒ majority A ≻ B.

rank4 voters3 voters2 voters
1.ABC
2.BCA
3.CAB
A ≻ B6 – 3B ≻ C7 – 2ABC

B – C: 4 + 3 = 7 to 2 ⇒ majority B ≻ C. Transitivity would suggest A ≻ C.

rank4 voters3 voters2 voters
1.ABC
2.BCA
3.CAB
A ≻ B6 – 3B ≻ C7 – 2C ≻ A5 – 4ABC

C – A: 3 + 2 = 5 to 4 ⇒ majority C ≻ A!

rank4 voters3 voters2 voters
1.ABC
2.BCA
3.CAB
A ≻ B6 – 3B ≻ C7 – 2C ≻ A5 – 4ABC

A ≻ B ≻ C ≻ A: the social preference is cyclic; every candidate is beaten by some majority, so there is no Condorcet winner.

Condorcet Paradox

The Agenda Decides the Winner

Sequential pairwise voting on the same cyclic profile: two candidates face off first, and the winner meets the next one.

Agenda 1: (A – B), then – C
  • A – B: 6 – 3 ⇒ A
  • A – C: 4 – 5 ⇒ C
Winner: C
Agenda 2: (B – C), then – A
  • B – C: 7 – 2 ⇒ B
  • B – A: 3 – 6 ⇒ A
Winner: A
Agenda 3: (A – C), then – B
  • A – C: 4 – 5 ⇒ C
  • C – B: 2 – 7 ⇒ B
Winner: B

A faces B first (A wins); A then meets C and loses: C.

Agenda 1: (A – B), then – C
  • A – B: 6 – 3 ⇒ A
  • A – C: 4 – 5 ⇒ C
Winner: C
Agenda 2: (B – C), then – A
  • B – C: 7 – 2 ⇒ B
  • B – A: 3 – 6 ⇒ A
Winner: A
Agenda 3: (A – C), then – B
  • A – C: 4 – 5 ⇒ C
  • C – B: 2 – 7 ⇒ B
Winner: B

First B – C (B wins), then B – A: A wins 6–3: A.

Agenda 1: (A – B), then – C
  • A – B: 6 – 3 ⇒ A
  • A – C: 4 – 5 ⇒ C
Winner: C
Agenda 2: (B – C), then – A
  • B – C: 7 – 2 ⇒ B
  • B – A: 3 – 6 ⇒ A
Winner: A
Agenda 3: (A – C), then – B
  • A – C: 4 – 5 ⇒ C
  • C – B: 2 – 7 ⇒ B
Winner: B

First A – C (C wins), then C – B: B wins 7–2: B.

Each of the three agendas yields a different winner: whoever sets the agenda (say, a committee chair) can pick the outcome. The candidate saved for last has the edge.

Condorcet Paradox

Back to the Warm-up: The Condorcet Paradox

The theory's answer: no

The warm-up profile (A≻B≻C, B≻C≻A, C≻A≻B): A–B 2–1, B–C 2–1, C–A 2–1. The social preference is cyclic; there is no Condorcet winner, and the agenda setter picks the outcome.

AgendaFirst voteWinner
(A–B), then CAC
(B–C), then ABA
(A–C), then BCB
How often?

With three voters, three candidates and all rankings equally likely, 12 of the 63 = 216 profiles are cycles: 5.6%. As the number of voters grows, this rises to about 8.8% (Guilbaud, 1952).

In practice: Election and survey data show cycles even less often than this "random preferences" estimate; real preferences are not random but similar to one another (Regenwetter et al., 2006).

In a cycle the agenda setter decides the winner: each of the three possible agendas elects a different candidate, and the one saved for last has the edge.

Condorcet Paradox

The Kemeny Rule: The Most Consistent Ranking

  • Kemeny (1959): choose the social ranking that agrees most with the voters' pairwise preferences.
  • The score of a ranking: for every pair where x comes before y, add the number of voters who prefer x to y.
  • If a Condorcet winner exists, it goes first: on the main profile D ≻ B ≻ C ≻ A (44).
  • It also decides cycles: it breaks the weakest majority ⇒ A ≻ B ≻ C (17).
  • Finding the best ranking is NP-hard (Bartholdi, Tovey, Trick, 1989).
A ≻ B 6–3B ≻ C 7–2C ≻ A 5–4ABCdashed: the weakest majority Kemeny overrides (5–4)
Applications: rank aggregation (Dwork et al., 2001): combining search engine results, ranking contestants and committee rankings.

Condorcet Paradox

Code Trace: kemeny(ballots, cands)

A ≻ B ≻ C·A ≻ C ≻ B·B ≻ A ≻ C·B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

ballots = [(4, "ABC"), (3, "BCA"), (2, "CAB")] (the cyclic profile). No best ranking yet; score −1.

order:A≻B≻CA before B+6A before C+4B before C+7score = 17A ≻ B ≻ C17A ≻ C ≻ B·B ≻ A ≻ C·B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

order = A ≻ B ≻ C: agreeing voters over the 3 pairs 6 + 4 + 7 = 17.

order:A≻B≻CA before B+6A before C+4B before C+7score = 17A ≻ B ≻ C17A ≻ C ≻ B·B ≻ A ≻ C·B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

17 > previous best ⇒ best = A ≻ B ≻ C.

order:A≻C≻BA before C+4A before B+6C before B+2score = 12A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C·B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

order = A ≻ C ≻ B: agreeing voters over the 3 pairs 4 + 6 + 2 = 12.

order:A≻C≻BA before C+4A before B+6C before B+2score = 12A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C·B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

12 ≤ 17: best unchanged.

order:B≻A≻CB before A+3B before C+7A before C+4score = 14A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

order = B ≻ A ≻ C: agreeing voters over the 3 pairs 3 + 7 + 4 = 14.

order:B≻A≻CB before A+3B before C+7A before C+4score = 14A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A·C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

14 ≤ 17: best unchanged.

order:B≻C≻AB before C+7B before A+3C before A+5score = 15A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

order = B ≻ C ≻ A: agreeing voters over the 3 pairs 7 + 3 + 5 = 15.

order:B≻C≻AB before C+7B before A+3C before A+5score = 15A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B·C ≻ B ≻ A·rankings tried (agreement score; orange = best)

15 ≤ 17: best unchanged.

order:C≻A≻BC before A+5C before B+2A before B+6score = 13A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B13C ≻ B ≻ A·rankings tried (agreement score; orange = best)

order = C ≻ A ≻ B: agreeing voters over the 3 pairs 5 + 2 + 6 = 13.

order:C≻A≻BC before A+5C before B+2A before B+6score = 13A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B13C ≻ B ≻ A·rankings tried (agreement score; orange = best)

13 ≤ 17: best unchanged.

order:C≻B≻AC before B+2C before A+5B before A+3score = 10A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B13C ≻ B ≻ A10rankings tried (agreement score; orange = best)

order = C ≻ B ≻ A: agreeing voters over the 3 pairs 2 + 5 + 3 = 10.

order:C≻B≻AC before B+2C before A+5B before A+3score = 10A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B13C ≻ B ≻ A10rankings tried (agreement score; orange = best)

10 ≤ 17: best unchanged.

A ≻ B ≻ C17A ≻ C ≻ B12B ≻ A ≻ C14B ≻ C ≻ A15C ≻ A ≻ B13C ≻ B ≻ A10rankings tried (agreement score; orange = best)

Result A ≻ B ≻ C (17): only the weakest majority, C ≻ A (5–4), is overridden. With m candidates, all m! rankings are tried.

from itertools import permutations
def kemeny(ballots, cands):
  best, best_score = None, -1
  for order in permutations(cands):
    score = 0
    for i, x in enumerate(order):
      for y in order[i + 1:]:
        score += sum(n for n, b in ballots
                     if b.index(x) < b.index(y))
    if score > best_score:
      best, best_score = order, score
  return best, best_score

Condorcet Paradox

Which Criteria Does Each Rule Satisfy?

RuleElects the
Condorcet winner
Never elects the
Condorcet loser
MonotoneStrategy
proof
Pluralitynono (A)yesno
Bordano (B)yesyesno
Instant-runoffno (C)yesnono
Approvalnonoyesno
Condorcet (Copeland)yesyesyesno
Kemenyyesyesyesno
  • Monotonicity: raising the winner on a ballot should never make it lose. IRV can violate this: moving a candidate up changes the elimination order.
  • In a cycle there is no Condorcet winner, so the bare rule cannot decide and needs an extension. The maximin (Simpson) rule takes each candidate's worst pairwise vote count; the highest wins: on the sample profile A 6, B 6, C 4, D 7 → D. It elects the Condorcet winner when one exists and still decides cycles.
  • Strategyproof: no voter can get a better outcome by misreporting their ranking. The "no"s in the last column come from a general theorem (Gibbard–Satterthwaite).

Condorcet Paradox

IRV Violates Monotonicity

rank6 voters6 voters5 voters
1.ABC
2.BCA
3.CAB
candidateround 1
A6
B6
C5

17 voters, 9 votes for a majority. Round 1: A = 6, B = 6, C = 5 ⇒ no majority; C has the fewest and is eliminated.

rank6 voters6 voters5 voters
1.ABC
2.BCA
3.CAB
candidateround 2
A11
B6
C—

C's 5 votes go to A: A = 11, B = 6 ⇒ A wins.

rank6+2 voters4 voters5 voters
1.ABC
2.BCA
3.CAB
candidate—
A?
B?
C?

Before the election A campaigns: 2 of the 6 B ≻ C ≻ A voters move A to the top (A ≻ B ≻ C). A's support only grew.

rank6+2 voters4 voters5 voters
1.ABC
2.BCA
3.CAB
candidateround 1
A8
B4
C5

New round 1: A = 8, B = 4, C = 5 ⇒ now B has the fewest and is eliminated (not C).

rank6+2 voters4 voters5 voters
1.ABC
2.BCA
3.CAB
candidateround 2
A8
B—
C9

B's 4 votes go to C: C = 9, A = 8 ⇒ C wins. A lost because it gained support: IRV is not monotone.

Strategic Voting

Tactical Voting under Plurality (Vote Splitting)

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateplurality
A6
B0
C4
D3

With sincere votes, A wins plurality. For group 3 (D ≻ B ≻ C ≻ A), A is the worst candidate!

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidateplurality
A6
B0
C4
D3

Group 3's favorite D (3 votes) cannot win. The candidate who can stop A is C, and this group prefers C to A.

rank4 voters4 voters3 voters2 voters
1.ACCA
2.DBDB
3.BDBC
4.CAAD
candidateplurality
A6
B0
C7
D0

Group 3 votes for C (ballot C ≻ D ≻ B ≻ A): C = 7, A = 6 ⇒ C wins, which this group likes better than A.

"Don't waste your vote": voters back a lesser evil who can win instead of their favorite (Duverger's law: plurality pushes toward two parties).

Strategic Voting

Strategic Voting under Borda Step by Step

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidatesincere
A18
B22
C17
D21

Sincere Borda count: B = 22, D = 21. Group 3 (D ≻ B ≻ C ≻ A) sees its favorite D losing by 1 point.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBCB
3.BDBC
4.CAAD
candidatesincere
A18
B22
C17
D21

Manipulation by burying: the 3 voters push the main rival B from 2nd to 3rd place; ballot D ≻ C ≻ B ≻ A.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBCB
3.BDBC
4.CAAD
candidatesincerechange
A180
B22−3
C17+3
D210

For each manipulator, B gets 1 point instead of 2 and C gets 2 instead of 1: B −3, C +3.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBCB
3.BDBC
4.CAAD
candidatesincerestrategic
A1818
B2219
C1720
D2121

New totals: D = 21, C = 20, B = 19, A = 18 ⇒ D wins. Group 3 got its favorite elected by lying.

Strategic Voting

How Much Damage Does Strategic Voting Do?

PluralityVote splitting: two similar candidates share votes and a third one wins. Tactical voting is common and simple.
BordaBurying a rival at the bottom is very effective; if everyone buries, the ranking drifts away from voters' true preferences.
Instant-runoffManipulation is harder to compute but possible, and the rule is not monotone. "IRV is immune to strategic voting" is a common myth.
Computational barrierFor some rules (e.g. IRV and a variant of Copeland) finding a profitable manipulation is NP-hard (Bartholdi et al., 1989–91).
The effect depends on the preference distribution: in close races strategic voting can change the winner. Transparent rules and clear information reduce some manipulation, but no general rule eliminates strategic behavior in every case.

Impossibility Theorems

Arrow's Impossibility Theorem: The Conditions

Reasonable-looking conditions for a rule that outputs a social ranking (a social welfare function):

Unrestricted domainEvery preference profile is allowed; the output is a complete, transitive ranking.
Pareto (unanimity)If everyone prefers x to y, so does society.
Independence of irrelevant alternatives (IIA)The social order of x and y depends only on how voters rank x versus y.
Non-dictatorshipNo single voter's preference is society's preference in every profile.
Arrow (1951): With at least 3 candidates, no rule satisfies all four conditions. (With 2 candidates, simple majority satisfies them all: May's theorem.)

Impossibility Theorems

Borda and Irrelevant Alternatives

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidate4 candidates
A18
B22
C17
D21

Borda with 4 candidates: B = 22 > D = 21; society ranks B above D.

rank4 voters4 voters3 voters2 voters
1.ACDA
2.DBBB
3.BDCC
4.CAAD
candidate4 candidates
A18
B22
C17
D21

Now let C drop out. C was neither winning nor second: an "irrelevant" candidate for the B–D order.

rank4 voters4 voters3 voters2 voters
1.ABDA
2.DDBB
3.BAAD
candidate4 candidatesno C
A1812
B2213
C17
D2114

New Borda (2, 1, 0): D = 14 > B = 13. No voter changed their B–D order (6 prefer B, 7 prefer D), yet the social order flipped: Borda violates IIA. Arrow's IIA can fail even with a fixed candidate set: if the 4-voter A ≻ D ≻ B ≻ C group only swaps C and B (A ≻ D ≻ C ≻ B), B drops from 22 to 18 while D stays at 21 (C rises to 21): the B–D order flips again.

Impossibility Theorems

What Arrow's Theorem Means in Practice

What it saysNo collective decision rule meets every reasonable criterion at once; every rule gives up some condition.
What it does not sayIt does not say voting is meaningless, nor does it rule out any particular rule.
Who gives up what?Borda: IIA. Condorcet/majority ranking: transitivity (cycles). Dictatorship: fairness.
For institutionsDecide explicitly which criterion takes priority, and document it.
The way out is to relax a condition: restrict the preference domain (single-peaked preferences), or use intensity information (scores, money) instead of rankings.

Impossibility Theorems

The Gibbard–Satterthwaite Theorem

Theorem (Gibbard 1973, Satterthwaite 1975): With at least 3 candidates and every ranking allowed (unrestricted domain), every rule that picks a single winner in every profile, lets every candidate win (onto) and is strategyproof (nobody gains by lying) is a dictatorship.
Evidence in our profileGroup 3 voting for C under plurality; burying B under Borda.
Relation to ArrowArrow is about rankings, G–S about a single winner; G–S can be derived from Arrow.
Mechanism designWithout money there is no general incentive-compatible rule. With money, VCG works.
  • The design goal is not to eliminate strategic behavior but to limit it: less information, computational hardness, a restricted preference domain.

Impossibility Theorems

May's Theorem: Simple Majority for Two Candidates

With two candidates (x, y), which rule is "fair"? May (1952) showed that only simple majority satisfies three conditions.

ConditionMeaningViolated by
AnonymityVoter names don't matter; permuting the ballots leaves the outcome unchangedDictatorship; weighted voting
NeutralityCandidate names don't matter; both candidates are treated symmetrically"The status quo stays without ⅔" (supermajority)
Positive responsivenessIf x is tied or winning and one voter switches to x, x wins outright"Always flip a coin"; "odd vote count wins"
With two candidates the debate over rules ends; the trouble (cycles, Arrow, G–S) starts with three or more candidates.

Impossibility Theorems

A Way Out: Single-Peaked Preferences

x1x2x3x4x5leftrightvoter 1voter 3voter 2 (median)
  • Preferences are single-peaked if candidates lie on an axis (left–right, budget size, location), each voter has a favorite position (peak), and utility falls as a candidate moves away from the peak.
  • Median voter theorem (Black, 1948): with an odd number of voters, the median voter's peak is the Condorcet winner; no cycles arise.
  • "Report your peak, pick the median" is strategyproof: G–S no longer applies once the domain is restricted.
The median mechanism for facility location is an application of this idea.

Impossibility Theorems

Four Results Side by Side

TheoremAssumptionsConclusionWay out
Arrow (1951)≥ 3 candidates, unrestricted domain, Pareto, IIAThe rule is a dictatorshipRestrict the domain (single-peaked); drop IIA; cardinal information (scores, intensity) instead of rankings
Gibbard–Satterthwaite (1973, 1975)≥ 3 candidates, unrestricted domain, onto (any candidate can win), strategyproofThe rule is a dictatorshipSingle-peaked preferences; randomized rules (e.g. random dictator); money (VCG)
May (1952)2 candidates; anonymity, neutrality, positive responsivenessThe rule is simple majority (and conversely)None needed: not an impossibility but a characterization
Single-peaked preferences (Black, 1948)Candidates on an axis, each voter has one peak, odd number of votersThe median voter's peak is the Condorcet winner; the median rule is strategyproofIt is itself a way out: it restricts the domain
In Arrow and G–S, dictatorship follows only when all assumptions are demanded together; relax one and positive results become possible.

Computational Social Choice

Computational Social Choice

Where computer science, economics and social choice meet: rules are judged not only by the criteria they satisfy but also by their computational cost.

Computing the winnerWith n voters and m candidates, plurality and Borda take O(nm) time, Copeland O(nm²). For Kemeny and Dodgson (the winner is the candidate needing the fewest adjacent swaps to become a Condorcet winner), finding the winner is NP-hard (Bartholdi, Tovey, Trick, 1989).
Computing manipulationsA single manipulator finds a profitable ballot under plurality and Borda with a polynomial-time greedy algorithm; under IRV the problem is NP-hard. Hardness is a “barrier”, but only in the worst case.
Preference elicitationA full ranking of m candidates carries log2(m!) bits (≈ 22 bits for m = 10). Asking voters questions step by step and collecting only enough to settle the winner cuts communication.
Fair division and judgment aggregationTwo more topics in this field.

Computational Social Choice

Judgment Aggregation: The Doctrinal Paradox

A three-judge court: the defendant pays damages only if p (the contract was valid) and q (the contract was breached).

JudgepqVerdict: p ∧ q (p and q)
Judge 1yesyesyes
Judge 2yesnono
Judge 3noyesno
Majorityyes (2–1)yes (2–1)no (1–2)
  • Each judge is consistent, but voting premise by premise the majority says “p true, q true, p ∧ q false”: an inconsistent collective judgment (Kornhauser and Sager, 1986).
  • Decide by the premises and the defendant pays; decide by the conclusion and they don't: the outcome again depends on procedure.
  • List and Pettit (2002): no aggregation rule is anonymous, systematic and always consistent; Arrow's theorem for judgments.

Summary

Voting Rules in the Real World

Two-round systemPresidential elections in Türkiye and France: without a first-round majority, the top two face a runoff.
Instant-runoffAustralian House of Representatives, some US cities; voters submit full rankings.
PluralitySingle-member districts in the UK and the US.
Borda-like scoringEurovision (12, 10, 8, …), sports awards (e.g. most valuable player votes).
ApprovalSome scientific society elections; "times that work for me" polls.
OnlineReddit and forum votes, community polls; class and club decisions.

Summary

Multiple Seats: Proportional Representation and D'Hondt

  • S seats instead of one winner; seat shares should track vote shares (proportional representation).
  • D'Hondt: divide each party's votes by 1, 2, 3, …; the S largest quotients win. Türkiye uses D'Hondt per district, plus a national threshold.
  • Example: 6 seats; A 42,000, B 31,000, C 15,000, D 12,000 votes → A 3, B 2, C 1, D 0.
  • A gets 50% of seats with 42% of votes; D gets none with 12%: D'Hondt favors large parties. Largest remainder (Hare): integer part of share × 6, leftovers to the largest fractions: A 2, B 2, C 1, D 1.
DivisorABCD
÷142,00031,00015,00012,000
÷221,00015,5007,5006,000
÷314,00010,3335,0004,000
÷410,5007,7503,7503,000
Seats3210

Highlighted cells: the six largest quotients.

Every method is a trade-off: D'Hondt favors stable governments, largest remainder small parties.

Summary

Which Voting Rule Can Be Called Fair?

  • Fairness depends on criteria such as representation, majority support, protection of minorities and strategyproofness.
  • Plurality is simple and fast, but can split votes and elect a winner without majority support (A: 6/13).
  • Borda uses more preference information and favors the compromise candidate (B), but is open to strategic ranking.
  • Instant-runoff seeks majority support, but is neither monotone nor strategyproof.
  • Condorcet methods respect pairwise majorities, but a Condorcet winner does not always exist and cycles need an extension rule.
Choose the rule by the context and the criteria that matter most: quick plurality may suit a club presidency, a ranked rule may be fairer for a class event.

Summary

Summary

RulesMajority, plurality, Borda, approval, instant-runoff, Condorcet, Kemeny: one profile, 4 different winners.
Condorcet paradoxTransitive individual preferences can produce a cyclic social preference; the agenda setter picks the outcome. Kemeny picks the ranking that disagrees least with the majorities; IRV is not monotone.
Strategic votingTactical voting under plurality, burying under Borda: honesty is not the best strategy under every rule.
ArrowWith ≥ 3 candidates, unrestricted domain + Pareto + IIA + non-dictatorship cannot all hold.
Gibbard–SatterthwaiteWith ≥ 3 candidates, a strategyproof onto rule = dictatorship.
May and single-peakednessWith 2 candidates, simple majority is the only fair rule; with single-peaked preferences the median is strategyproof.

Summary

Sample Problems

  1. 9 voters: 4 × A ≻ B ≻ C, 3 × C ≻ B ≻ A, 2 × B ≻ C ≻ A. Who wins under plurality, IRV, Borda and Condorcet?
  2. In problem 1, what do the A voters achieve under Borda by burying B (A ≻ C ≻ B)?
  3. On the cyclic profile (4 ABC, 3 BCA, 2 CAB), what agenda makes B win?
  4. What is the Kemeny ranking of the profile in problem 1? (Compare the scores of all 6 rankings.)
  5. "The status quo changes only with ⅔ of the vote": which of May's conditions does this rule violate?
  6. 11 voters: 2 × ACB, 3 × BAC, 2 × BCA, 4 × CBA. Who is the Condorcet winner? Which rule s = (1, a, 0) (0 ≤ a ≤ 1) elects them?
Show answers
  1. Plurality A (4). IRV: B (2) is eliminated, its votes go to C ⇒ C 5 – A 4, C. Borda: B 11, A 8, C 8 ⇒ B. Condorcet: B (beats A 5–4, C 6–3).
  2. Scores A 8, B 7, C 12 ⇒ C wins, which is worse than B for the A voters. Burying can backfire.
  3. First A – C (C wins), then C – B (B wins 7–2).
  4. B ≻ C ≻ A (score 16 = 6 + 5 + 5); the Condorcet winner B comes first.
  5. Neutrality: the two candidates are not treated symmetrically.
  6. C (beats A and B 6–5). None: for every a, B = 5 + 4a > C = 4 + 4a.

Summary

References

  • F. Brandt, V. Conitzer, U. Endriss, J. Lang, A. D. Procaccia (eds.), Handbook of Computational Social Choice, Cambridge University Press, 2016.
  • N. Nisan, T. Roughgarden, É. Tardos, V. Vazirani (eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (Chapter 9).
  • K. J. Arrow, Social Choice and Individual Values, Wiley, 1951.
  • A. Gibbard, "Manipulation of voting schemes: a general result", Econometrica 41(4), 1973.
  • M. A. Satterthwaite, "Strategy-proofness and Arrow's conditions", J. Economic Theory 10, 1975.
  • K. O. May, "A set of independent necessary and sufficient conditions for simple majority decision", Econometrica 20(4), 1952.
  • D. Black, "On the rationale of group decision-making", J. Political Economy 56(1), 1948.
  • J. J. Bartholdi, C. A. Tovey, M. A. Trick, "The computational difficulty of manipulating an election" and "Voting schemes for which it can be difficult to tell who won the election", Social Choice and Welfare 6, 1989 · M. Balinski, H. P. Young, Fair Representation, Yale University Press, 1982 (proportional representation).
  • J. G. Kemeny, "Mathematics without numbers", Daedalus 88(4), 1959; C. Dwork, R. Kumar, M. Naor, D. Sivakumar, "Rank aggregation methods for the web", WWW 2001.
  • A. Karlin, Y. Peres, Game Theory, Alive, AMS, 2017 (chapter on social choice and voting); D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge University Press, 2010 (Chapter 23: Voting).
  • Course pages: A. Procaccia, 15-896 (CMU) · N. Shah, CSC304 (Toronto).

The End

Chapter 7: Social Choice and Voting
Previous: Chapter 6: Online Platforms · Next: Chapter 8: Network Design and Routing
1 / 1 Sercan KÜLCÜ. All rights reserved.