Symptom

The team is prioritising next quarter’s roadmap. Six engineers each rank the eight candidate projects, you average the ranks, and you ship the top three.

Someone points out that project H — ranked last by everyone, dead on arrival — was accidentally left off two of the ballots. You add it back, re-run the average, and the order of A and B flips. Nothing about anyone’s opinion of A or B changed. A universally-hated eighth option changed which of the top two wins.

You try other rules. Borda count does the same thing. Plurality gives a winner that four of the six people ranked in the bottom half. Runoff gives yet another answer, and you discover that one engineer could have got a better outcome by ranking their true favourite lower.

The natural conclusion is that you have not found the right voting rule yet. Arrow’s theorem is the statement that you will not find it, because it does not exist. The bug is not in the aggregation code. It is in the specification.

Statement

Fix a set of at least three alternatives and $n \ge 2$ voters. Each voter submits a strict ranking. A social welfare function maps every profile of individual rankings to a single social ranking.

Conditions.

Unrestricted domain. Every profile of rankings is a legal input. The rule never refuses to answer.

Weak Pareto. If every voter ranks $x$ above $y$, the social ranking puts $x$ above $y$.

Independence of irrelevant alternatives (IIA). The social ordering of $x$ and $y$ depends only on how each voter ranks $x$ against $y$ — not on where anyone puts $z$.

Non-dictatorship. There is no voter $d$ such that the social ranking always equals $d$’s ranking, regardless of everyone else.

Arrow (1951). With three or more alternatives, no social welfare function satisfies unrestricted domain, weak Pareto, IIA, and non-dictatorship simultaneously. Any rule satisfying the first three is a dictatorship.

The theorem is stated for rankings out, rankings in. That matters, and the Does not forbid section is mostly about what happens when you weaken either end.

Argument

The cleanest modern proof is Geanakoplos’s, and its shape is worth carrying even if you never reproduce the details. It builds a dictator out of nothing but the three positive conditions, in three moves.

Move 1: extremal lemma. Suppose every voter puts alternative $b$ either at the very top or the very bottom of their ranking — nobody puts it in the middle. Then society must also put $b$ at the top or the bottom.

Why: suppose not, so society ranks some $a$ above $b$ and $b$ above some $c$. Now imagine every voter who ranked $c$ above $a$ swaps them. This does not move $b$ for anyone, since $b$ is extremal for everyone, so by IIA the social ranking of $a$ vs $b$ and of $b$ vs $c$ is unchanged: still $a > b > c$, hence $a > c$ socially. But we just arranged for every voter to rank $c$ above $a$, so weak Pareto demands $c > a$. Contradiction. So $b$ is extremal socially.

Move 2: construct a pivotal voter. Line the voters up in some order. Start from the profile where everyone puts $b$ last; by Pareto, society puts $b$ last. Now walk down the list, moving $b$ from the bottom to the top of one voter’s ranking at a time. At the end everyone has $b$ first and, by Pareto, society does too. So somewhere in this walk the social position of $b$ moved, and by Move 1 it moved from bottom straight to top in a single step. Call that voter $d$ — the pivotal voter for $b$.

Move 3: show $d$ is a dictator. Take any two alternatives $x, y$ both different from $b$, and suppose $d$ ranks $x$ above $y$. Build a profile from the one just before $d$’s switch: put $b$ at the top for $d$, in the position between $x$ and $y$, so that $d$ ranks $x > b > y$; let everyone else order $x$ and $y$ however they like, keeping $b$ where it was in that profile.

Comparing $x$ and $b$: this profile matches the pre-switch one on every voter’s $x$-vs-$b$ ranking, where society had $b$ at the bottom, so $x > b$ socially. Comparing $b$ and $y$: it matches the post-switch profile on $b$-vs-$y$, where $b$ was at the top, so $b > y$ socially. Transitivity of the social ranking gives $x > y$ socially, regardless of what anyone other than $d$ said about $x$ and $y$. So $d$ dictates over every pair not involving $b$; a symmetric argument with a different alternative in $b$’s role extends this to all pairs. $\blacksquare$

Where each condition is spent. Pareto anchors the two ends of the walk. IIA lets us transplant a comparison between profiles that agree only on that comparison, which is what makes Moves 1 and 3 work at all. Unrestricted domain lets us construct the profiles we need — every one of the strange ballots above had to be a legal input. Transitivity of the output is what Move 3 exploits in its last line.

Back to the roadmap. The rank-averaging rule violates IIA, and the H incident is exactly that violation, visible. Borda count violates IIA for the same reason: a candidate’s Borda score depends on how many alternatives it beats, so adding a loser can change margins. Every rule you tried failed one of the four conditions, because every rule must. Once you know which one your rule fails, you know its characteristic pathology, and you can decide whether you can live with it.

Forbids

A ranked voting rule that is fair in all four senses at once. Not one that is undiscovered — one that cannot be constructed. Stop looking.

“Immune to irrelevant alternatives” as a property you can just add. IIA plus Pareto plus universality already forces a dictator. If your rule is non-dictatorial and always answers, adding an alternative can reorder the existing ones, somewhere.

Treating spoiler effects as implementation bugs. A third option changing the outcome between the first two is what IIA failure is. Every non-dictatorial ranked rule has instances of it.

The claim that a better aggregation formula would settle a contested prioritisation. The formula is not the source of the disagreement, and no formula launders it into an objectively correct answer.

Does not forbid

It does not say all voting systems are equally bad, which is the misreading that gets the theorem cited to justify apathy. The conditions fail in different ways with wildly different practical severity. Plurality is spoiled by a candidate with 5% support; instant-runoff needs a much rarer configuration; Condorcet methods produce a correct winner whenever one exists and only need a tiebreak in genuine cycles. “All rules are imperfect” and “all rules are equivalent” are very different statements, and only the first one is a theorem.

It does not apply when preferences are restricted. Unrestricted domain is a real hypothesis, not decoration. If voters’ preferences are single-peaked — each voter has an ideal point on a line and dislikes options in proportion to distance — then Black’s median voter theorem gives a rule that is transitive, Paretian, IIA-satisfying, and non-dictatorial: take the median. Prioritisation along a single axis, like how much to invest in reliability, often really is single-peaked, and there the impossibility simply does not bind.

It does not apply to cardinal input. Arrow’s voters submit orderings. If they submit comparable intensities — scores, budgets, dollars — range voting satisfies IIA and Pareto without a dictator. The catch is elsewhere: intensities are not verifiable and are trivially exaggerated, so the problem moves from aggregation to elicitation. Which is the right trade for a small internal team and the wrong one for a public election.

It does not forbid a well-defined winner, only a well-defined ranking. Much of the difficulty comes from demanding a full transitive social order. If you only need to choose one thing, the relevant result is different, and the honest one to cite is Gibbard–Satterthwaite: any non-dictatorial rule choosing among three or more alternatives is manipulable. That is a genuinely different failure mode, about strategy rather than coherence.

It does not say the dictator is a person in the room. “Dictatorship” is a mathematical property of the function — one voter’s ranking determines the output. A rule can satisfy it because a tiebreak rule always defers to whoever sorted first, which nobody chose and nobody notices.

Boundary

  • Fewer than three alternatives, and the theorem is empty. Majority rule on two options satisfies everything, which is May’s theorem, and is why binary decisions feel so much cleaner than roadmap prioritisation. They are.
  • The output is required to be a total order. Weakening the output to a quasi-transitive relation escapes dictatorship but yields an oligarchy (Gibbard), which is not obviously better.
  • Infinite voter sets escape the conclusion via ultrafilter constructions, a fact of model-theoretic interest and no practical use.
  • The theorem is about the rule, not the outcome. A dictatorial rule can produce an outcome everyone likes; a fair rule can produce one everyone hates. Arrow constrains procedures.
  • It says nothing about how often pathologies occur. For realistic preference distributions Condorcet cycles are rare. Rarity is an empirical claim about your electorate and not a property you get to assume.

Read this against T095. Stable matching asks who can defect and answers constructively; Arrow asks who can be aggregated and answers that you cannot have everything. The pair is the two possible outcomes of formalising fairness, and knowing which one you are in tells you whether to search for an algorithm or to start choosing which guarantee to give up.