Social welfare function — the object Arrow's theorem is actually about¶

Almost every impossibility argument you'll meet turns on a distinction that debate rarely bothers to state: does a voting rule output a winner, or a ranking? Arrow's theorem is about the second kind. Get the type wrong and you'll either over-apply the theorem ("Arrow proved every method is unfair") or miss why some methods dodge it on a technicality. This page pins down the two objects, states the Pareto and IIA axioms at both levels, and shows the trap: majority rule is both Paretian and IIA, and escapes Arrow only because it is not a social welfare function at all.**

→ Related: Does Arrow apply to STAR? — the ordinal/cardinal escape · Gibbard–Satterthwaite — the other impossibility · distortion — social welfare as a number · criteria at a glance · the math behind Condorcet.

Two objects, two types¶

Fix a set of voters N = {1,…,n} and a set of candidates A. A voter's ballot is a linear order on A — a strict ranking, no ties. Write L(A) for the set of those, and R(A) for the set of weak orders (complete and transitive, ties allowed). A profile P = (≽₁,…,≽ₙ) is one ballot per voter.

	Outputs	Type	Anchoring theorem
Social welfare function (SWF)	a whole social ranking	`f : L(A)ⁿ → R(A)`	Arrow (1951)
Social choice function (SCF)	a winner (or tied set)	`f : L(A)ⁿ → A`	Gibbard–Satterthwaite (1973/75)

The SWF's output is called the social preference order. Note the asymmetry in Arrow's setup: individual ballots may not contain ties, but the social result may. That's a modeling convenience, not a law of nature — real ballots routinely permit equal ranks (weak ranks), and STAR's Equal Support bucket exists precisely because voters do want to say "these two are the same to me."

Nearly every method taught in this repo is an SCF — STAR, Approval, RCV-IRV, Plurality all name a winner. Ranked Robin is the interesting hybrid: its pairwise win-loss record is a social ranking, and the winner is read off the top of it.

The two axioms, stated at both levels¶

Because there are two types of object, each axiom has two readings. Conflating them is the single most common error in criterion tables.

Pareto / unanimity (Vilfredo Pareto, 1848–1923):

SWF (weak Pareto): if a ≻ᵢ b for every voter i, then a ≻ b in the social ranking.
SCF (Pareto criterion): if every voter prefers a to b, then b is not elected. A candidate is Pareto optimal if no rival is unanimously preferred over them.
Strong Pareto: all weakly prefer a, at least one strictly ⟹ a ≻ b. Arrow only needs the weak form — assuming less makes the theorem stronger.
Nonimposition is the weakest relative in the family: no candidate is unelectable (for every candidate there's some profile that elects them outright). Pareto implies nonimposition — a unanimously top-ranked candidate must win — so it's rarely assumed separately.

Why a criterion this weak is worth stating at all is clearest from the rule it excludes. Consider reverse Borda: elect whoever has the lowest Borda count. It is perfectly anonymous and perfectly neutral — the two axioms people reach for first — and it is transparently backwards. Nothing but Pareto rules it out. That's the honest job description: Pareto doesn't identify good winners, it excludes rules that are inverted. Which is also why passing it proves so little (see below).

Independence of Irrelevant Alternatives (IIA): the social ranking of a vs b depends only on how individuals rank a vs b — never on where anyone puts a third candidate c. Violating this is precisely the spoiler effect.

Arrow's theorem. With |A| ≥ 3, every SWF that is weakly Paretian and IIA is a dictatorship — there is one voter whose ranking simply is the output.

The proof runs on decisive coalitions. A coalition C is decisive for a over b if C unanimously preferring a forces a ≻ b socially. Weak Pareto is exactly the statement that the grand coalition N is decisive — that's the seed. A Contagion (Field Expansion) Lemma upgrades "decisive for one pair" to "decisive for all pairs"; a Splitting (Group Contraction) Lemma shows any decisive coalition of size ≥ 2 contains a smaller decisive one. Iterate down from N and you land on a singleton. Without Pareto there is no nonempty decisive set to start shrinking.

The trap: majority rule is Paretian and IIA¶

Pairwise majority rule satisfies both axioms, easily. So why isn't it a counterexample to Arrow?

Because it isn't an SWF. Its output need not be transitive — the Condorcet paradox produces a ≻ b ≻ c ≻ a, which is not a weak order, so it isn't in R(A) and the function f : L(A)ⁿ → R(A) is not well defined. Majority rule escapes Arrow by failing to have the right type, not by beating an axiom.

This is the correct frame for Condorcet methods generally, and it cuts against a sloppy claim in both directions:

Against the critics: "Condorcet methods sometimes elect nobody" is false for real methods. Bare "elect the Condorcet winner" is a partial rule; Ranked Robin, Ranked Pairs, and Schulze are completions that always return a winner. Any criterion table with a "Condorcet Method — Always a Winner: NO" row needs that row split.
Against the advocates: a completion doesn't dodge Arrow either. Once a Condorcet method always outputs a ranking, it is an SWF, and Arrow applies in full — so it must fail IIA or Pareto or be a dictatorship. Ranked Robin fails IIA, which is exactly where cycle-resolution rules live. Patching the cycle is what costs you IIA; it doesn't buy an exemption.

The asymmetry that keeps Pareto from being oversold¶

Pareto forbids; it does not require. The criterion says non-Pareto-optimal candidates must not win. It says nothing about which Pareto-optimal candidate should.

Plurality satisfies Pareto — a candidate ranked below X by everyone gets no first-place votes and can't win — yet plurality routinely elects a poor Pareto-optimal candidate. Passing Pareto is a floor, not a recommendation. Relatedly: Pareto ⟹ the unanimity criterion (a candidate holding every first-place vote wins), since every rival is then unanimously dominated. The converse fails.

And the sting in the tail: a dictatorship is Paretian. The dictator's top choice is never unanimously dominated. That is why Arrow's conclusion is devastating rather than reassuring — the axioms are so mild that dictatorship clears them.

Who actually fails Pareto¶

A short list, because "fails Pareto" sounds worse than it usually is — most methods pass:

Rule	Pareto	Why
Plurality, Borda, IRV, STAR, Ranked Robin	✓	unanimous domination survives the count
Dictatorship	✓	the dictator's favorite is never dominated
Imposed / constant rule ("X always wins"; "everything ties")	✗	ignores the ballots entirely
Sequential pairwise / agenda voting	✗	a unanimously-preferred candidate can be eliminated early — runnable at agenda_voting.md
Anti-plurality	✗	on unanimous `A>B>C` it elects A and B; B is dominated
Approval	✗	worked: Felsenthal Ex.6

Why Approval fails and STAR doesn't is the instructive pair, and it's a ballot-expressiveness point, not a tabulation one. An approval ballot cannot record a strict preference within the approved set, so "every voter prefers A to C" is a fact the ballots never carried and the count cannot honor.

STAR passes, and the argument is short. If every voter scores a strictly above b, then a's score total strictly exceeds b's — so b can only reach the runoff alongside a, and there every voter prefers a. b never wins. Note this leans on the scoring round and the runoff together: the runoff alone wouldn't do it.

Don't run this row on a multi-winner rule — it's a type error. Every table above is about a single-winner SCF. Apply the same definition candidate-by-candidate to a committee election and every bloc rule "fails" trivially: Bloc STAR, Bloc Approval, Bloc Ranked Robin and SNTV must fill N seats, so if fewer than N candidates are Pareto optimal, one of the seated candidates is necessarily dominated. Zwicker gives the crisp specimen — seat candidates in descending Copeland order until the committee is full, which is exactly what this repo's Bloc RR does — and immediately supplies the correction: for a committee election the alternatives are committees, not candidates, so a committee-level Pareto criterion is the one that applies. The apparent failure is an artifact of comparing objects of the wrong type — the same trap this page opens with, one level up.

The phrase does double duty, and the repo uses both:

Arrow's SWF — the ordinal object above. A function producing a social ranking. No utilities anywhere; Arrow deliberately avoided interpersonal comparison.
Welfarist social welfare — a cardinal quantity, W(u₁,…,uₙ), aggregating voter utilities. Utilitarian (sum), egalitarian (max-min), or Nash (product).

Sense 2 is the hidden spine of three pages that otherwise look unrelated: distortion (a candidate's social welfare = the sum of voter utilities; the optimum minimizes total cost), VSE (the 100% mark is the utilitarian optimum), and the ABC rules spectrum (AV / PAV / Chamberlin–Courant are one family differing only in the aggregator — see welfarist rule in the glossary). Same object, three aggregators.

The cleanest way to hold the two apart: Arrow's SWF asks "what order?"; welfarist social welfare asks "how much?" — and cardinal ballots are exactly what lets you ask the second question, which is why STAR sits outside Arrow's frame but squarely inside the distortion literature.

Reading a criterion table without being fooled¶

Grids of YES/NO across methods circulate widely. Three checks:

Are the rows SWFs or SCFs? A table mixing "elects a Condorcet winner" (SCF) with "produces a transitive ranking" (SWF) is comparing different objects.
Which columns are actually Arrow's? Unrestricted domain, transitive-and-complete output, weak Pareto, IIA, non-dictatorship. Monotonicity and Condorcet-efficiency are not Arrow conditions — including them implies the impossibility involves them, which it doesn't.
Does dictatorship pass everything? If so the table is correct and you've found Arrow's punchline: the axioms are mild enough that the worst rule clears them. That's the lesson, not a bug in the table.

Per reading these fairly: a criterion failure is a trade-off to weigh, never a disqualification.

Sources¶

Kenneth J. Arrow, Social Choice and Individual Values (1951) — the theorem (book note). Lean: neutral, foundational.
Peter C. Fishburn, "Condorcet Social Choice Functions," SIAM J. Appl. Math. 33(3), 1977 — the precise SCF vocabulary (reading list).
Amartya Sen, Collective Choice and Social Welfare (1970; exp. ed. 2017) — the welfarist sense, and the bridge between the two (book note).
The decisive-coalition proof (Contagion / Splitting Lemmas) is the standard modern presentation in the computational-social-choice textbooks; see the math behind Condorcet for how it sits alongside Gibbard–Satterthwaite.