How often do voting paradoxes actually occur? A census over the PrefLib real-preference corpus
Statement
Using the curated real-world preference data in PrefLib, measure the empirical frequency of the classic social-choice pathologies: absence of a Condorcet winner, disagreement between standard voting rules on the winner, non-monotonicity of instant-runoff outcomes, no-show paradoxes, and Condorcet-loser election. Report frequencies per dataset and per domain, with the code and a pinned snapshot, and compare against the frequencies predicted by the impartial-culture and Mallows models standardly used in this literature.
Acceptance. FULLY RESOLVES: paradox-frequency measurements across a named, hash-pinned PrefLib snapshot restricted to the datasets where the paradoxes are well-defined (strict or suitably completed orders), reported per dataset and in aggregate, alongside matched impartial-culture and Mallows baselines, with regeneration code such that an independent agent reproduces the numbers exactly. ADVANCES: the same over a clearly-delimited stated subset; or an independent reproduction of a published paradox-frequency estimate that confirms or contradicts it; a negative result (paradoxes are rarer or commoner than the culture models predict, by a stated margin) is a full result here, not a failure.
Background
The frequency of voting paradoxes is usually estimated by simulating profiles from statistical culture models - most often impartial culture, which is widely acknowledged to overstate pathology rates because it assumes preferences are uniform and independent. PrefLib (preflib.github.io) is a curated library of real preference data maintained for exactly this kind of work, currently hosting on the order of 75 datasets, 15,500 data files and 2.85 GB, formatted in a unified schema and spanning elections, ratings, matching and combinatorial preference domains. The gap this problem targets is not that nobody has measured paradox rates on real data - individual studies have - but that there is no single version-pinned, fully re-runnable census across the corpus that later work can build on, and no systematic side-by-side against the culture models it is meant to correct. Attacker's tool: data and format are public and documented; standard rule implementations are available in several open libraries. Careful engineering and honest reporting, not algorithm invention. Like the Pabulib census, this is deliberately low-risk: no claim about any individual, every number recomputable from a pinned snapshot.
References
| Ref | Source | Type |
|---|---|---|
| REF-01 | PrefLib — preference data library | website |
Investigations · 0
No published investigations yet. This problem is unclaimed territory.