Topic: Condorcet Efficiency (electing the head-to-head winner)¶
Topic hub — a cross-method view. A Condorcet winner is a candidate who beats every other candidate one-on-one (by majority). A method is Condorcet-efficient if it elects that candidate whenever one exists. (Sometimes nobody qualifies — the pairwise results form a cycle — which is a separate, rarer problem.)
The one idea to take away: the Condorcet winner is the "majority's head-to-head choice." Some methods guarantee it, some get it almost always, and RCV-IRV (Hare) can miss it precisely because of center squeeze.
Which methods elect the Condorcet winner — and where each is treated¶
| Method | Condorcet winner? | Notes | Full page |
|---|---|---|---|
| Ranked Robin / Copeland | ✅ always | it is a Condorcet method — most pairwise wins | Ranked Robin |
| BTR / Baldwin / Nanson | ✅ always | Condorcet-safe IRV variants | Which RCV-IRV? |
| STAR | ⚠️ very often | not guaranteed; the runoff usually recovers it. Measured: 74–99% | STAR automatic runoff |
| RCV-IRV (Hare) | ❌ not guaranteed | can eliminate the Condorcet winner before the final round | Center squeeze |
| Approval / Plurality | ❌ | don't use the full pairwise picture | scoring methods |
When there's no Condorcet winner (a cycle), the principled "still in contention" list is the Smith set, and methods differ in how they pick from it — see cycle resolution.
The mirror image of this table is the Condorcet loser — the candidate who loses every head-to-head (the loser paradox) — and its ties-allowed refinement, the weak Condorcet loser, the candidate who beats nobody. That distinction is where STAR's guarantee has fine print: it can never elect a strict Condorcet loser, but a tie is not a loss, so a weak one can win on the runoff tiebreaker. Worked across five methods on five voters: The weak Condorcet loser. The repo's divergence ledger catalogs real library elections where STAR, IRV, and the Condorcet winner disagree.
"Isn't the Condorcet criterion just a guarantee for moderates?" No — that's the load-bearing error of a much-cited FairVote article. The Condorcet winner is defined by the electorate and moves with it (a majority-first-choice landslide candidate is automatically the Condorcet winner). The article, quoted and checked claim by claim against tabulated elections — including its own 40/15/40 hypothetical, where RCV-IRV eliminates the very moderate it discusses: FairVote's Condorcet article, claim-checked.
"What about the case for the criterion — does that hold up?" Mostly yes, which is the more useful finding. Wikipedia's "Condorcet winner criterion," claim-checked takes its four arguments in turn: stability / no weak spoilers holds and is a genuine advantage over STAR (concede it — but lead with cycles are rare, the evidence, not only a beat dislodges a beats-all winner, the definition, and never let it be heard as IIA); the participation paragraph is the rare case of an article being too soft on the family it describes (it's Moulin's theorem, provably unavoidable at 4+ candidates, not a "constructed example"); and the Smith paragraph reaches the right conclusion through a definition that is actually the Schwartz set. One sentence is outright wrong: the "306 election datasets found no participation failures for the ranked pairs–minimax family" line misreports the paper it cites — Ranked Pairs was never tested, and Copeland, Black's rule and STV each had a violating profile. We pulled the study.
"So is the Condorcet winner sacrosanct, then?" Also no — and the serious version of that argument deserves a serious reading. Edelman's "Myth of the Condorcet Winner," tabulated (BV2173, live): a "cancellation" profile going back to Condorcet himself where every majoritarian count elects one candidate and every positional count (Borda, score sum) elects another, 41 voters to 40 — plus the join-consistency/no-show theorems. The anti-Condorcet argument that survives tabulation.
"Is there anything that positively proves the pairwise approach right?" Yes — one theorem, and it's the strongest card the Condorcet family holds. Campbell–Kelly (2003), "May's Theorem for three or more alternatives": restricted to profiles where a Condorcet winner exists, electing that winner is resolute, anonymous, neutral and strategyproof — and for an odd number of voters it is the unique such rule. State it at full strength, then state the price, which is large and specific: the theorem holds only on that restricted domain, so it says nothing about cycles — exactly the profiles where methods actually disagree — and it can't distinguish Ranked Robin from Minimax from Schulze, since all of them agree there by definition. It also doesn't contradict Gibbard–Satterthwaite, and why is the useful lesson: G–S's bite comes from the full domain, not from having three or more candidates. Fair warning for STAR advocates — STAR is not a Condorcet extension, so this one lands on the other side of the table: the full treatment, limits included.
"Isn't Ranked Robin the same as Condorcet?" Almost — they're identical when a Condorcet winner exists, and part ways only in a cycle (Condorcet goes blank, Ranked Robin still picks the best record). Worked through with a real example in Ranked Robin vs. the Condorcet winner.
"But a paper proves Condorcet, IIA, and monotonicity aren't even desirable?" That's the most sophisticated form of the argument — an arXiv paper defining "ordered majority rule" and claiming IRV uniquely satisfies it. The catch: the property is IRV's own algorithm restated, defined circularly, so the "uniqueness" is a mirror. Taken apart (with the honest core — the case for cardinal ballots — kept) in Ordered majority rule and the "Condorcet isn't desirable" argument.
"Isn't there a slick new system that puts the head-to-head matchups right on the ballot?" Yes — Better Choices, a Top-3 Condorcet method whose final ballot is three bubble matchups instead of a ranking, completed by Minimax. Explained fairly and claim-checked (including the wrinkle that its cycle behavior is Minimax's, not Copeland/Ranked Robin's, and the intransitive-ballot cost of independent bubbles): Better Choices — the pairwise-ballot Condorcet method.
"Which Condorcet rule is best — is that even answerable?" For exactly three candidates, yes, and there's a neutral 2024 theorem for it: Brandt, Dong & Peters prove that maximin and its refinements (leximin, Nanson) are the uniquely best-defended Condorcet extensions against the no-show and reinforcement paradoxes — the tractable small case where Moulin's impossibility hasn't yet bitten. Summarized with the fine print (3-candidate-only; it's the tie-broken refinements that are characterized; STAR isn't a Condorcet extension so it's out of scope): Condorcet-Consistent Choice Among Three Candidates.
"There are dozens of Condorcet methods — do I have to learn them all?" For a three-way race, no: most of the famous ones are literally the same rule. Maximin, Ranked Pairs, Schulze, Kemeny, Dodgson, and Young all coincide at three candidates and only diverge at four-plus. What's left to choose is small and nameable: At three candidates, the famous Condorcet methods collapse into one. And for the smallest election where the collapse doesn't hold — where Copeland and the maximin family part company on five ballots — the minimal tilted cycle.
"Show me a Condorcet method actually misbehaving." Two towns, each of which elects Ada, that merge into an electorate electing Cara — the reinforcement paradox, run across every method (additive rules keep the promise, Condorcet rules can't, STAR's runoff breaks it): Reinforcement paradox — both halves pick Ada, the whole picks Cara.
"When there's no Condorcet winner, who decides — the ballots or the rule?" In a cycle the family splits, and the newest member says so out loud: Split Cycle, claim-checked discards each cycle's weakest defeat and returns every candidate left undefeated, rather than applying a convention. Includes a tabulated election where a candidate no voter ranks above the winner still flips Schulze's result — plus the four things that case doesn't show.
"Fine — but how often, actually?" Measured, not asserted: Condorcet efficiency, measured runs six methods over the same sampled electorates and reports the rate for each, with Ranked Robin's mandatory 100% as the control. The headline is that there is no single number — the electorate model swings the answer by more than the gap between methods. STAR runs 74–99%; RCV-IRV beats STAR under impartial culture and drops below 50% on a crowded single-issue spectrum. Includes the surprise in STAR's shortfall: most of it is preference the 0–5 ballot could not carry, not the top-two rule.
"Why does a bigger field make it worse?" Because three separate things happen to the election before any method touches it — the Condorcet winner's first-choice share halves, their narrowest head-to-head margin thins to a third, and the share of candidate pairs a 0–5 ballot cannot separate doubles. All three are measured in Why more candidates make every method miss, which also draws the efficiencies as bars, names which mechanism hits which method, and works a single 65-voter electorate through 3, 5 and 7 candidates — the crowded field, where four different people win and not one voter changes their mind.
"Where do I go to read about this properly?" Condorcet methods — a reading list: the books, papers, and free surveys worth your time, each with its lean marked — and, first, the one taxonomy (Fishburn's C1/C2/C3) that makes the family's names stop sliding. Start there if the nomenclature is what's blocking you.
Glossary: Condorcet.
Learn it by depth. The concept ladders across the curriculum: 101 — the Condorcet winner as "the majority's head-to-head choice" (intuition only); 201 — what a Condorcet extension is and the three-candidate collapse (§201.6); 301 — the variable-electorate paradoxes and the maximin result (§301.7 disagreement, §301.13 no-show, §301.14 reinforcement).
This is a topic hub (cross-method index). The authoritative write-ups live in the per-method folders linked above. See the topics index for the other topic hubs.