Plurality Veto Voting Rule

• Plurality Veto is an ordinal single-winner rule where each candidate begins with a plurality score and voters sequentially veto their least-preferred candidate.
• It achieves a deterministic metric distortion of 3, which is optimal given the use of only ordinal preference information.
• Extensions of the rule include randomized interpolation, learning-augmented variants, and generalization to k-Approval veto cores for enhanced fairness and protection.

Plurality Veto is a single-winner voting rule in ordinal metric social choice in which each candidate begins with support equal to its plurality score, voters then successively veto their least-preferred currently standing candidate, and the last standing candidate is elected. Its central significance is that, despite using only ordinal information, it achieves the optimal deterministic metric distortion of $3$, matching the known lower bound for all deterministic rules. Subsequent work reinterprets it as the uniform-voter, plurality-weighted special case of a generalized $(p,q)$-veto core, develops anonymous and neutral continuous-time variants, and uses the same structure to obtain learning-augmented and minority-protection generalizations (Kizilkaya et al., 2022).

1. Metric formulation and objective

In the metric distortion framework, there is a finite set of candidates $C$ and a finite set of voters $V$, with $|C|=m$ and $|V|=n$. Voters and candidates are embedded in a common metric space $(X,d)$, where

$$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$

For a candidate $c \in C$, the social cost is

$SC(c,d) = \sum_{v \in V} d(v,c).$

The optimal candidate is

$(p,q)$0

The voting rule does not observe the distances $(p,q)$1; it only receives the ordinal profile $(p,q)$2, where each ranking is induced by distance: $(p,q)$3 A metric $(p,q)$4 is aligned with $(p,q)$5, written $(p,q)$6, when every voter’s ranking is consistent with $(p,q)$7. For a deterministic rule $(p,q)$8, distortion is

$(p,q)$9

No deterministic rule can achieve distortion better than $C$0, and there exist rules that achieve $C$1 exactly (Berger et al., 2023).

2. Rule definition and operational structure

Plurality Veto begins from plurality support and then applies veto pressure. For each candidate $C$2, let

$C$3

denote the plurality score. The rule initializes

$C$4

Fix an arbitrary order of voters $C$5. At round $C$6, let

$C$7

be the set of standing candidates. Voter $C$8 identifies her bottom candidate among $C$9,

$V$0

and decrements that candidate’s score by $V$1. After $V$2 rounds, the rule returns $V$3, the candidate vetoed in the last round. Equivalently, candidates start with as many “lives” as first-place votes, every voter removes one life from her least-preferred surviving candidate, and the last standing candidate wins (Kizilkaya et al., 2022).

This formulation is notable for its low communication overhead: it only makes two queries to each voter, namely a top query and a bottom-among query on the current active set. At the same time, the original sequential rule depends on the order in which voters are processed, so the winner can vary with that order; later work identifies this order dependence as the source of non-anonymity and motivates simultaneous variants (Kizilkaya et al., 2022).

3. Veto-core interpretation and simultaneous variants

Later work reframed Plurality Veto through a generalized veto core. Let $V$4 be a distribution over voters and $V$5 a distribution over candidates. A coalition $V$6 $V$7-blocks a candidate $V$8 if there exists $V$9 such that $|C|=m$0 and

$|C|=m$1

The $|C|=m$2-veto core is the set of candidates not $|C|=m$3-blocked by any coalition. The same work proves that a candidate is in the $|C|=m$4-veto core if and only if it is $|C|=m$5-dominant, meaning that it admits a $|C|=m$6-matching in the associated domination graph. Under uniform voter weights and candidate weights proportional to plurality scores, Plurality Veto is exactly the plurality-weighted special case of this framework (Kizilkaya et al., 2023).

This reinterpretation yields two structural consequences. First, previous procedures for selecting winners from veto cores can be viewed as matching algorithms, and different election methods realize different matchings. Second, a continuous-time rule, SimultaneousVeto, replaces sequential vetoing by a process in which every voter continuously brings down, at rate $|C|=m$7, the support of her bottom choice among not-yet-eliminated candidates. In the plurality specialization, each candidate starts with public support equal to its plurality score, and a candidate is eliminated if it is opposed by a voter after its support reaches $|C|=m$8. The resulting SimultaneousPluralityVeto is anonymous and neutral, returns a nonempty set of tied winners, and satisfies resolvability, monotonicity, majority, majority loser, mutual majority, and reversal symmetry (Kizilkaya et al., 2023).

The generalized-veto-core perspective therefore places the original rule, its simultaneous version, and the earlier matching-based distortion rules in a common cooperative and combinatorial framework.

4. Distortion guarantees, randomized interpolation, and learning augmentation

The original distortion proof proceeds by showing that the Plurality Veto winner has a perfect matching in an appropriate domination graph, which implies distortion $|C|=m$9. The same paper also defines a randomized family, RoundPluralityVeto$|V|=n$0, in which the veto process is run for only $|V|=n$1 rounds and then a candidate is chosen with probability proportional to its residual score. This interpolates between Random Dictatorship at $|V|=n$2 and Plurality Veto at $|V|=n$3, and for all $|V|=n$4 the rule has distortion at most $|V|=n$5 (Kizilkaya et al., 2022).

A later learning-augmented formulation modifies the plurality-weighted veto-core process using a prediction $|V|=n$6 of the optimal candidate and a parameter $|V|=n$7. For all $|V|=n$8,

$|V|=n$9

while for the predicted candidate

$(X,d)$0

Every voter down-votes her least-preferred active candidate at rate

$(X,d)$1

This algorithm, BoostedSV, achieves

$(X,d)$2

and this robustness–consistency trade-off is optimal among deterministic algorithms. When $(X,d)$3, the guarantees collapse to distortion $(X,d)$4, recovering the unboosted plurality-veto baseline. The error-sensitive bound is

$(X,d)$5

As $(X,d)$6, consistency tends to $(X,d)$7 while robustness tends to $(X,d)$8 (Berger et al., 2023).

Variant	Modification	Guarantee
Plurality Veto	Full veto process; last standing candidate wins	Distortion $(X,d)$9
RoundPluralityVeto$$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$0	Run only $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$1 veto rounds, then sample by residual score	Distortion at most $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$2 for all $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$3
BoostedSV	Add prediction-dependent boost and scaled veto rate	Optimal robustness–consistency trade-off

These extensions preserve the core intuition of plurality-weighted support depleted by bottom-directed vetoes, while altering how much the rule relies on veto dynamics, randomness, or side information.

5. Position inside the $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$4-Approval Veto spectrum

Recent work embeds Plurality Veto into a broader family called $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$5-Approval Veto. In that family, each candidate starts with its $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$6-approval score, each voter appears exactly $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$7 times in a veto order, and at each step the voter decrements the score of her least-preferred currently eligible candidate. The $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$8-approval veto core is the set of all candidates that can survive for some veto order. The case $$d : (V \cup C) \times (V \cup C) \to \mathbb{R}_{\ge 0}.$$9 is exactly Plurality Veto, except that the newer formulation allows ties, whereas the original rule immediately eliminates all candidates whose score reaches $c \in C$0 and insists on picking a single winner (Kizilkaya et al., 23 Jul 2025).

The utilitarian distortion theorem for the full class states that every candidate in the $c \in C$1-approval veto core has distortion at most

$c \in C$2

Specializing to $c \in C$3 yields the familiar bound $c \in C$4, which is tight. For the $c \in C$5-percentile objective, every candidate in the $c \in C$6-approval veto core has distortion at most $c \in C$7 for

$c \in C$8

and the distortion is unbounded for

$c \in C$9

Thus, for Plurality Veto, the threshold is $SC(c,d) = \sum_{v \in V} d(v,c).$0: every winner has $SC(c,d) = \sum_{v \in V} d(v,c).$1-percentile distortion at most $SC(c,d) = \sum_{v \in V} d(v,c).$2 for $SC(c,d) = \sum_{v \in V} d(v,c).$3, and no deterministic rule can guarantee less than $SC(c,d) = \sum_{v \in V} d(v,c).$4 on that interval. For the egalitarian objective, every candidate in the $SC(c,d) = \sum_{v \in V} d(v,c).$5-approval veto core has distortion at most $SC(c,d) = \sum_{v \in V} d(v,c).$6, so the $SC(c,d) = \sum_{v \in V} d(v,c).$7 case is again optimal (Kizilkaya et al., 23 Jul 2025).

The same paper studies mutual minority protection. In general, every candidate in the $SC(c,d) = \sum_{v \in V} d(v,c).$8-approval veto core has mutual minority protection at least $SC(c,d) = \sum_{v \in V} d(v,c).$9. For $(p,q)$00, the nontrivial guarantee is stronger than the generic statement: every Plurality Veto winner has mutual minority protection at least $(p,q)$01, because candidates ranked last by strictly more than $(p,q)$02 voters cannot possibly win under PluralityVeto. In that sense, Plurality Veto satisfies the majority loser criterion but remains the welfare-maximizing and minority-protection-minimizing endpoint of the $(p,q)$03-Approval Veto spectrum, while $(p,q)$04 yields VoteByVeto and the proportional veto core (Kizilkaya et al., 23 Jul 2025).

6. Connections to $(p,q)$05-plurality and proportional clustering

Plurality Veto also appears in geometric clustering and proportional-fairness work. A point $(p,q)$06 is a $(p,q)$07-plurality point if

$(p,q)$08

For $(p,q)$09, this becomes the metric Condorcet-point condition. The $(p,q)$10-plurality problem asks for a point with the largest possible $(p,q)$11 (Kellerhals et al., 14 Feb 2025).

Plurality Veto always selects a

$(p,q)$12

-plurality point. Since $(p,q)$13, this yields a $(p,q)$14-Droop proportional $(p,q)$15-center. The same paper proves that a point $(p,q)$16 is a $(p,q)$17-plurality point if and only if the clustering $(p,q)$18 satisfies $(p,q)$19-Droop proportionality, so the Plurality Veto guarantee translates directly into a proportional-clustering guarantee (Kellerhals et al., 14 Feb 2025).

A generic theorem in that setting states that every $(p,q)$20-plurality point has distortion

$(p,q)$21

Substituting $(p,q)$22 gives $(p,q)$23, which is weaker than the known distortion-$(p,q)$24 guarantee for Plurality Veto itself. This juxtaposition is informative rather than contradictory: it shows that the rule simultaneously satisfies a strong metric-distortion bound and a certified plurality/proportionality guarantee, and it enabled the proof that $(p,q)$25-proportionally fair clusterings can be found using purely ordinal information (Kellerhals et al., 14 Feb 2025).