Introduction
In 1950, Kenneth Arrow’s "Impossibility Theorem" established that no fair voting system can satisfy a specific set of reasonable axioms. Among the most debated is the Independence of Irrelevant Alternatives (IIA). IIA posits that the collective preference between two options, $x$ and $y$, should depend solely on how voters rank $x$ against $y$, regardless of their ranking of a third option, $z$.
While IIA appears logically sound, it imposes a significant limitation: it ignores the intensity of preferences. For instance, if 51% of voters slightly prefer $x$ over $y$, but 49% passionately prefer $y$ over $x$, a majority rule (which satisfies IIA) will select $x$. However, by examining how voters rank $x$ and $y$ relative to numerous other "irrelevant" alternatives, we can infer the strength of their preferences. This research quantifies the social welfare lost by adhering to IIA and identifies non-IIA rules that effectively recover this lost utility.
Methodology: Pairwise Distortion
The authors bridge the gap between ordinal rankings and cardinal utilities using the framework of pairwise distortion. This metric calculates the ratio between the social welfare (sum of utilities) of the "true" best alternative in a pair versus the one selected by a voting rule.
The study evaluates several rules, including Pairwise Majority (which satisfies IIA), and several that violate it: Borda, Copeland, Plurality, and Half-approval. These are tested across synthetic distributions (Uniform, Unit-sum, 2D-Euclidean) and the empirical Bars dataset, which contains real human ratings of 16 different bars.
Empirical Performance
A central insight of the paper is that irrelevant alternatives serve as "Information Proxies." The relative positions of other alternatives act as a ruler; if $x$ and $y$ are at opposite ends of a ranking, it suggests a high utility difference. By following IIA, the Majority rule is forced to ignore this data, resulting in constant distortion regardless of the number of alternatives ($m$).
In contrast, rules like Borda and Copeland "harvest" this information. As $m$ increases, their pairwise distortion decreases because more options provide a more precise "ruler" for preference intensity.
As shown in the figure above, Plurality is a notable exception. Because it only considers top choices, it discards too much information and suffers from "vote splitting," causing its distortion to increase as more alternatives are added.
Theoretical Bounds
The researchers also conducted a worst-case analysis, specifically within 1D-Euclidean spaces where voters and candidates are represented as points on a line. The theoretical results confirm the trends seen in the empirical data.
Key theoretical findings include:
- Pairwise Majority: Maintains a fixed worst-case distortion of 3.
- Borda Rule: Achieves a distortion of $\frac{m+1}{m-1}$. As $m$ grows, this value approaches 1 (perfect efficiency), assuming alternatives are positioned strategically.
- OddBorda: A variant using a scoring vector of $1, 3, 5, \dots$ was found to perform even better in specific metric settings.
Conclusion
This paper challenges the long-standing IIA axiom by demonstrating that "irrelevant" options provide crucial context for social welfare. While the theoretical results are currently focused on 1D-Euclidean spaces, the empirical evidence across various datasets strongly suggests that non-IIA rules like Borda and Copeland are superior for maximizing social utility in large-scale decision-making. Future work may explore these dynamics in higher-dimensional metric spaces or correlated voter models.
Summary updated