The Distortion of Ordinal Social Choice

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2123 papers160 conceptsfrozen 2026-10-04

In a perfect world, a collective decision-making system would know exactly how much every person values every option. If we knew that Alice values a park at 100 units and Bob values it at 10, we could make decisions that maximize the total "social welfare."

In reality, we almost never have this data. Humans find it cognitively exhausting to assign precise numerical values to their preferences. Instead, we ask for ordinal rankings: "I prefer the park to the library." This creates an information bottleneck. By throwing away the intensity of preferences, we introduce distortion—a measure of how much social welfare we lose by making decisions based on rankings rather than true underlying utilities.

The Information Gap

The core problem, as surveyed by Elliot Anshelevich, Aris Filos-Ratsikas, Nisarg Shah, and Alexandros A. Voudouris in Distortion in Social Choice Problems: The First 15 Years and Beyond, is that ordinal rankings are "unitless."

If a majority of voters slightly prefer Candidate A over Candidate B, but a large minority passionately hates A and loves B, a simple majority vote (an ordinal rule) will pick A. However, the total social welfare might be much higher if B were chosen. This gap between the "utilitarian optimum" and the outcome of a voting rule is the distortion.

Théo Delemazure, Jérôme Lang, and Grzegorz Pierczynski argue in Independence of Irrelevant Alternatives under the Lens of Pairwise Distortion that even "irrelevant" alternatives in a ranking provide crucial clues about this hidden intensity. For example, if many candidates sit between your first and second choice, it suggests a large "distance" in your preference, a signal that rules like the Borda Count use to outperform simpler rules like Plurality.

Geometry as a Proxy for Utility

To bound this distortion mathematically, researchers often assume that preferences live in a metric space. Imagine voters and candidates as points on a map; a voter’s "cost" for a candidate is simply the distance between them.

In this geometric world, Elliot Anshelevich, Onkar Bhardwaj, and John Postl proved in Approximating Optimal Social Choice under Metric Preferences that the Copeland rule is a powerhouse. While rules like Plurality have distortion that grows linearly with the number of candidates, Copeland maintains a small, constant distortion. The triangle inequality of the metric space prevents any one candidate from being "too far" from the consensus if they are ranked highly by enough people.

Implicit Utilitarianism

If we view voting as a statistical problem, we can treat ordinal votes as noisy observations of latent cardinal utilities. This framework, known as Implicit Utilitarian Voting, was pioneered by Ioannis Caragiannis, Ariel D. Procaccia, and Nisarg Shah.

In Subset Selection Via Implicit Utilitarian Voting, they show that we don't need to know the exact utilities to maximize them. By choosing the outcome that minimizes "worst-case regret" across all possible utility scales consistent with the observed rankings, we can achieve remarkably high social welfare. This approach is so effective it powers RoboVote.org, a platform for making real-world group decisions.

The Power of a Single Bit

Can we break the fundamental limits of ordinal voting without returning to the nightmare of 1–100 star ratings? Recent work suggests that preference intensity is the key.

Anson Kahng, Mohamad Latifian, and Nisarg Shah demonstrated in Voting with Preference Intensities that eliciting just a single bit of intensity information—such as marking a "sharp drop" in utility between two candidates—can dramatically reduce distortion. Similarly, Mohamad Latifian and Alexandros A. Voudouris showed in The Distortion of Threshold Approval Matching that "threshold approvals" (approving all options above a certain personal bar) provide enough cardinal signal to exponentially improve resource allocation efficiency.

From Ballots to AI Alignment

This theoretical framework is now moving beyond politics into AI. In Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?, Paul Gölz, Nika Haghtalab, and Kunhe Yang apply distortion to Reinforcement Learning from Human Feedback (RLHF).

Because RLHF typically relies on pairwise comparisons (A > B), it suffers from the same information bottleneck as ordinal voting. If a reward model cannot capture how much a user prefers one response over another, it may optimize for a "mythical average" that leaves diverse subgroups with suboptimal outcomes. By measuring distortion, researchers can finally quantify the social welfare loss inherent in how we align AI with human values.

Papers in this version

Showing 60 of 123 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 210 papers
  2. Version 1 · 2026-10-04 · 123 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_ordinal_cardinal_information_gap_7b627a02,
  title = {ordinal-cardinal information gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/ordinal-cardinal-information-gap/rdr_7b627a02c1e441189f63a0f4579162a8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_distortion_of_ordinal_social_choice_7b627a02_v1,
  title = {The Distortion of Ordinal Social Choice},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-distortion-of-ordinal-social-choice/rdr_7b627a02c1e441189f63a0f4579162a8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}