Condorcet-Consistent Choice Among Three Candidates — the maximin result (Brandt, Dong & Peters, 2024)¶
A rigorous social-choice result, read for what it does and doesn't prove. Felix Brandt, Chris Dong & Dominik Peters, "Condorcet-Consistent Choice Among Three Candidates" (arXiv:2411.19857, Nov 2024; journal version 2025) asks: if you restrict to exactly three candidates, which Condorcet extension best resists the two nastiest variable-electorate paradoxes — the no-show paradox and the reinforcement paradox? Their answer: maximin and two of its refinements (Nanson's rule and leximin) occupy a uniquely defensible position. This is the theoretical backbone under the Better Choices proposal's minimax count — and, unlike advocacy literature, a neutral academic result (the authors are theorists, not campaigners). It cuts for a specific Condorcet rule, and this repo — which leans STAR, a method the theorem doesn't even cover — reports it straight.
→ Runnable: the paper's Fig. 1 profile, with its minimality proved — the minimal tilted cycle (5 voters) · its Theorem 2 profile — the reinforcement paradox. → Related: Minimax / Simpson-Kramer (Felsenthal paradoxes) · the No-Show paradox · Participation topic hub · Better Choices — the pairwise-ballot method · cycle resolution · Condorcet reading list.
Why "exactly three candidates" is the whole point¶
Two candidates are trivial: majority rule satisfies essentially every fairness property at once. Three or more candidates is where Arrow and Gibbard–Satterthwaite bite and every rule starts trading one virtue for another. The paper's move is to ask whether the smallest hard case — three candidates — is tractable enough to pick a best Condorcet rule, even if no such rule exists in general. It is. Three candidates is special because it is exactly where Moulin's impossibility theorem stops applying: Moulin (1988) proved every Condorcet extension suffers the no-show paradox once there are ≥ 4 candidates (and enough voters). At three, there's room to escape — and this paper maps exactly who does.
The two paradoxes, and the findings¶
Reinforcement paradox (a.k.a. consistency / multiple-districts, Young–Levenglick 1978): two separate groups of voters each elect A, but the combined electorate does not. A rule that does this contradicts what every sub-group agreed on. Runnable — the paper's own Theorem 2 profile, cast as two towns whose merger flips the winner from Ada to Cara, counted across every method: Reinforcement paradox — when both halves pick Ada but the whole picks Cara.
Finding: with three candidates, the reinforcement paradox must occur for every Condorcet extension once there are ≥ 8 voters — no escape, for anyone. But certain refinements of maximin are immune when there are ≤ 7 voters.
No-show paradox (participation failure, Moulin 1988): a voter gets a better result by staying home than by voting sincerely — abstention beats participation.
Finding: among homogeneous Condorcet extensions (rules unchanged when you scale the whole electorate up proportionally), the only ones immune to the no-show paradox are refinements of maximin.
Add the trivial fact that any Condorcet extension elects the Condorcet winner when one exists, and the scorecard for three candidates is:
| Property | Maximin refinements (Nanson / leximin) | Any other Condorcet extension |
|---|---|---|
| Elects the Condorcet winner when one exists | ✅ (by definition) | ✅ (by definition) |
| Immune to the no-show paradox (homogeneous rules) | ✅ uniquely | ❌ |
| Immune to the reinforcement paradox, ≤ 7 voters | ✅ | ❌ (in general) |
| Immune to the reinforcement paradox, ≥ 8 voters | ❌ | ❌ — nobody is |
A companion fact from the same paper's Fig. 2 makes the family tractable: at three candidates, maximin = Ranked Pairs = Schulze = Kemeny = Dodgson = Young are one and the same rule — the three-candidate collapse.
The paper then gives axiomatic characterizations of maximin, Nanson's rule, and leximin — short lists of independently reasonable axioms that uniquely pin down each rule. That's the honest form of "use this rule": not "trust us," but "here are principles you'd likely accept, and this is the only rule satisfying all of them."
What maximin and its refinements are¶
- Maximin (Simpson–Kramer; Condorcet's own 1785 cycle rule): elect the candidate whose worst pairwise loss is smallest — least-strongly-beaten. Condorcet's three-candidate phrasing: in a cycle, "the adopted view results from the two [pairwise majorities] that are most probable [largest]." Bare maximin can tie (two candidates with equally bad worst losses), which is why refinements exist.
- Leximin — break maximin ties lexicographically: compare worst losses; if tied, second-worst; then third-worst. "Least bad, then next-least-bad."
- Nanson's rule — iteratively eliminate every candidate with a below-average Borda score. It's a Condorcet extension, and at three candidates it lands as a maximin refinement.
How this reconciles with the repo's other minimax page¶
This looks, at first, to contradict the repo's Minimax page, which (following Felsenthal) lists Minimax as vulnerable to the no-show, twin, and reinforcement paradoxes. Both are right — the difference is candidate count. Every one of Felsenthal's damning Minimax examples uses four candidates (Example 30's no-show, for instance, has a four-candidate cyclical order). Brandt–Dong–Peters restrict to three, where Moulin's impossibility hasn't kicked in — and there, the refinements of maximin (bare maximin's ties resolved) are uniquely well-behaved. So the two results don't collide: Minimax looks paradox-prone in the general (4+) case and uniquely well-behaved in the exactly-three case. That's a genuinely clarifying pairing, not a contradiction — and a good caution against citing a criterion result without its candidate-count fine print.
What this means for Better Choices — and the honest caveats¶
The Better Choices proposal counts its three-candidate final by minimax ("least bad loss"). This paper is the serious backing for that choice: at three candidates, minimax-family rules really are the best-defended Condorcet option against the two worst variable-electorate paradoxes. Three caveats keep it in proportion — all of which strengthen the case for reading the paper carefully rather than as a slogan:
- It's the refinements that are characterized, not bare minimax. Better Choices as described (plain "smallest-margin loss") can tie; the clean theorems attach to leximin / Nanson, which specify the tie-break. A faithful implementation should pin down that tie-break (leximin is the natural one), not leave it to a coin.
- It's a three-candidate result — full stop. With four or more candidates Moulin's theorem returns and every Condorcet extension, minimax included, fails no-show. Better Choices dodges this only by being a Top-3 system (its primary guarantees exactly three finalists) — which is precisely why the primary that feeds it matters so much.
- The reinforcement paradox is unavoidable for everyone at ≥ 8 voters. This is not "maximin escapes all paradoxes." It escapes no-show (uniquely) and reinforcement for small electorates. In any real public election (≫ 8 voters) Better Choices, Ranked Robin, and every other Condorcet rule can still exhibit reinforcement. The result is a "best available," not a "flawless."
And the repo's own lean, stated plainly: this theorem is entirely about the Condorcet family — rules that insist on the head-to-head winner. STAR is not a Condorcet extension (it's a score method; it elects the Condorcet winner very often but not by rule), so the theorem simply does not bind it. STAR makes the opposite trade: it gives up the Condorcet guarantee to buy preference-strength expression and a two-step count with no separate primary — and it, too, can fail no-show (the repo concedes this openly). So this paper is not "Condorcet beats STAR"; it's "if you commit to always electing the Condorcet winner, then at three candidates maximin-leximin is your strongest defense." Whether to make that commitment at all is the scores-vs-ranks fork, which this result doesn't settle.
Bottom line¶
For the narrow, tractable case of exactly three candidates, Brandt, Dong & Peters give the maximin family the strongest theoretical justification any Condorcet rule has: uniquely no-show-immune among homogeneous rules, reinforcement-immune for small electorates, with clean axiomatic characterizations of maximin, Nanson, and leximin. It is the rigorous foundation under Better Choices' minimax count — with the fine print that it's a three-candidate result, that it's the tie-broken refinements that are characterized, that reinforcement still bites everyone at scale, and that it says nothing about score methods like STAR, which decline the Condorcet commitment the theorem is about.
Source: Felix Brandt, Chris Dong & Dominik Peters, "Condorcet-Consistent Choice Among Three Candidates" (arXiv:2411.19857; author PDF). Neutral academic social-choice theory — no campaign affiliation on either side. Glossary: Condorcet · no-show paradox. See also Darlington's pro-Minimax case in the reading list — advocacy for the same rule family, read with its lean marked.