Final Validation Report: Axioms for Learning from Pairwise Comparisons

Updated: 2026-08-26

1. Human Verdict

Formalized. The review surface covers every numbered definition, lemma, and theorem in the NeurIPS 2020 main paper and both named claims in its official supplement. All fourteen source presentations have transparent source-facing Lean statements. The ten result claims have checked proof endpoints; the four numbered definitions are reviewed directly as semantic declarations.

The formalization covers the finite pairwise-count likelihood model, MLE existence and uniqueness, Pareto efficiency, monotonicity, the impossibility of pairwise-majority consistency, and the impossibility of separability.

2. Closeout Status

3. Source and Scope

The source is Ritesh Noothigattu, Dominik Peters, and Ariel D. Procaccia, Axioms for Learning from Pairwise Comparisons, NeurIPS 2020.

The audit uses a combined copy of the official proceedings paper and official supplement (SHA-256 cbf69fcb0860f292056e7c8c3f1b7813b048877924ae6b6c037e84292d733163) and its byte-pinned layout-preserving transcript. The source-only inventory independently reconciles Lemmas 2.1–2.4; Definitions 3.1, 4.1, 5.1, and 6.1; Theorems 3.2, 4.2, 5.3, and 6.3; and supplementary Claims A.1 and C.1.

Exact source spans, semantic context, and proof routes are in audit/paper_statement_map.json. Source-facing Lean statements are in PaperInterface.lean; paired proof endpoints are in ProofInterface.lean.

4. Researcher Summary of Checked Results

Result Comparison with source
Lemma 2.1 Exact.
Lemma 2.2 Exact.
Lemma 2.3 Quantifier corrected: positive directed counts only between distinct alternatives; the source sup-norm bound follows. Correction.
Lemma 2.4 Exact.
Definitions 3.1, 4.1; Theorems 3.2, 4.2 Exact.
Definition 5.1; Theorem 5.3 Exact.
Definition 6.1; Theorem 6.3 Exact.
Supplementary Claims A.1 and C.1 Exact.

5. Remaining Boundaries and Gaps

None.

6. Additional Assumptions Beyond Paper

None.

7. Proof-Strategy Deviations

Theorems 5.3 and 6.3 replace continuity of the maximizer with a compact likelihood gap and an exact rational perturbation under the source assumptions. Proof details.

8. Proof Tricks Worth Reusing

9. Generalizations, Conjectures, and Extensions

The reusable library now includes finite pairwise-count datasets, comparison graphs, CDF-like pairwise links, fixed-reference MLEs, perfect-fit distances, likelihood existence and uniqueness tools, score-exchange lemmas, and exact counterexample constructions. These components are independent of the paper namespace and can support later ranking and preference-learning work.

10. Source Clarifications and Corrections

Lemma 2.3’s positive-count premise ranges over distinct alternatives because diagonal counts are zero. The memo gives this local clarification.

11. Paper Issues or Caveats

None.

12. Detailed Formalization Evidence

PaperInterface.lean presents fourteen source items: four semantic definitions and ten result claims. ProofInterface.lean supplies one exact-type endpoint for each result; the four definitions remain direct declarations rather than artificial theorem wrappers.

13. Paper Assumption Provenance

Assumptions.lean introduces no paper-local axiom. Four graph-selected paper-model prerequisites have current matching judgments in the paper-prerequisite ledger. No material reusable-library prerequisite is selected.

14. Displayed Formula Provenance

The source map binds pairwise counts, comparison graph, link/CDF likelihood, finite MLE, perfect fit, score bounds, and the Pareto, monotonicity, majority-consistency, and separability formulas to their source presentations.

15. Library Lift Pass

The reusable library contains the general pairwise-count dataset, graph, likelihood, link, MLE, perfect-fit, and comparison-axiom machinery. Paper numbering and explicit counterexample instances remain paper-local.

16. DAG Audit

The dependency DAG places the model and supplementary claims before the main results and shows the four axiom results depending on the finite MLE foundation. The retained visual inspection found readable labels, correct arrow direction, and no clipping or node overlap.

17. Validation Checks

The weak-order PMC definition and its Theorem 5.3 endpoint compile and pass bounded independent source comparison. Their current closeout evidence is being synchronized; the retained receipt describes its pinned prior tree.

Retained focused proof-interface and root builds passed. The current graph records ten matching direct result judgments and four matching paper-model prerequisites. The accepted graph and closure receipt bind the closeout surface; no new Lean or semantic review was run for this document edit.

18. Paper Definitions Checked

The checked source definitions are the pairwise-comparison/MLE model, Definition 3.1 Pareto efficiency, Definition 4.1 monotonicity, Definition 5.1 pairwise-majority consistency, and Definition 6.1 separability.

19. Named Theorem Statements Checked

20. Paper-Facing Statement Validator Ledger

The ten direct comparisons are in the source-to-Spec ledger; model premises are in the paper-prerequisite ledger, and source-proof fidelity records the counterexample proof replacement.

21. Source-Coverage Audit Ledger

The source map inventories all fourteen named source presentations. The ten result rows have current matching direct judgments, and the four definition rows are covered as semantic model prerequisites. The accepted graph binds both groups without counting the definitions as extra theorem claims.