Abstract
While it seems common sense that adding a vote for a winning candidate shouldn't make them lose, many popular voting methods—including Borda count and Copeland—frequently violate this "Positive Involvement" (PI) principle. This paper introduces a new framework to measure these violations, revealing that for certain methods, a significant portion of a voter's power is actually "turned against" their favorite candidate.
Introduction
In the field of social choice theory, the Positive Involvement (PI) axiom states a simple rule: if candidate $x$ is a winner, adding a new voter who ranks $x$ first should not cause $x$ to lose. It is a "variable-electorate" cousin to the more famous monotonicity axiom.
The challenge is that while many sophisticated voting methods (especially those that are "Condorcet consistent") violate PI in theory, the absolute frequency of these violations in random simulations often appears negligible (e.g., < 1%). This leads some to dismiss PI violations as a theoretical curiosity rather than a practical concern. The authors argue that absolute frequency is a misleading metric and set out to provide more meaningful ways to measure how often a voting method's logic is "fundamentally misconceived."
Methods and Measurement Strategies
The authors use large-scale computer simulations to test various voting methods, including Borda, Instant Runoff (IRV), Coombs, Baldwin, Copeland, Top Cycle, Uncovered Set, Ranked Pairs, and Split Cycle. They introduce two novel measurement strategies to better capture the impact of PI violations:
- Conditional Disagreement: Instead of asking how often a method violates PI across all possible elections, they ask: "In elections where Method A and Method B choose different winners, how often does Method A violate PI?"
- Voter Potency Relativization: They define "voter potency" as the probability that a single voter's action causes any candidate to lose. They then measure the ratio of PI violations to this potency. This answers the question: "If my vote is actually going to change the outcome, how likely is it to change it in a way that hurts my favorite candidate?"
Datasets
The study relies on synthetic preference profiles generated from four standard probability models in social choice theory:
- Impartial Culture (IC): Every possible ranking is equally likely.
- Impartial Anonymous Culture (IAC): Focuses on the distribution of vote counts.
- Urn Model (URN): Simulates correlation between voters.
- Mallows Model: Simulates a "noisy" version of a ground-truth ranking.
The authors sampled 50,000 profiles for each data point, covering electorates from 4 to 5,000 voters and 3 to 10 candidates.
Key Insights
- The "Insult to Injury" Effect: The paper highlights that PI violations are uniquely perverse. In a large election, your vote rarely changes the outcome. If it does change the outcome by making your favorite candidate lose, your "potency" has been weaponized against your own preferences.
- Conditioning Matters: When looking only at elections where two voting methods disagree, the probability of a PI violation jumps from near-zero to significant levels (sometimes over 20-50% for certain methods).
- The Condorcet Conflict: Most methods that always pick a Condorcet winner violate PI. However, the authors highlight Split Cycle as a notable exception that satisfies both Condorcet consistency and Positive Involvement.
- Potency Ratios: For methods like Coombs and Top Cycle, simulations show that as the number of candidates increases, a shockingly high percentage of a voter's potential influence is "negative."
Key Results
- Absolute vs. Relative: While absolute PI violations for the Baldwin method might be under 1% in a standard 5-candidate IC model, the probability of violation jumps to nearly 15% when conditioning on disagreement with Borda.
- Voter Potency: In 10-candidate elections using the Top Cycle method, the ratio of PI violations to voter potency is remarkably high, suggesting that a voter's ability to "help" their favorite is frequently overshadowed by the risk of "hurting" them.
- Consistency across Models: The "perverse" behavior of methods like Coombs and Baldwin persisted across the Urn and Mallows models.
Conclusion
While these axioms are mathematically interesting for large elections, their practical impact is most dramatic in small committees or clubs where a single vote has a high probability of being the tie-breaker. Future work remains to explore how PI violations interact with complex strategic voting behaviors.
Summary updated