Arrow's Impossibility Theorem

What the theorem does not say

This deserves its own treatment, because the misreadings are common enough to be the main thing standing between a reader and the actual result.

It does not say democracy is impossible. It says a specific class of preference-aggregation rule cannot have four specific properties at once. Real democratic systems have legislatures, deliberation, agenda-setting, amendment procedures, courts and repeated elections, none of which are social welfare functions in Arrow's sense.

It does not say all voting systems are equally bad. This is the most damaging misuse, because it is sometimes deployed to shut down electoral reform debates. Systems differ in how often the failure occurs, in whether the failure is catastrophic or cosmetic, and in how easy they are to game. Approval voting, instant-runoff voting, Condorcet methods and plurality all violate something — but plurality violates far more, far more often, and the comparison between them is entirely meaningful.

It does not say the outcome of an election is arbitrary. With single-peaked preferences — voters arrayed along a single dimension, each with one ideal point — Duncan Black showed in 1948 that majority rule is perfectly well behaved and the median voter's position beats every alternative. Real electorates are often close enough to this that Condorcet winners exist. Arrow's condition of unrestricted domain insists on covering every logically possible profile, including ones that never occur.

It does not apply to two options. With exactly two alternatives, May's theorem (1952) is the relevant result, and it goes the other way: simple majority rule is the unique rule satisfying anonymity, neutrality and positive responsiveness. Binary choice is entirely well behaved. The trouble starts at three.

It does not apply to cardinal methods. Range voting, score voting, approval voting and Balinski and Laraki's majority judgment aggregate ratings or grades, not rankings. They fall outside the theorem's hypotheses. Anyone who tells you "Arrow proved score voting can't work" is misstating the theorem.

The escapes, and what each costs

Restrict the domain. Assume single-peakedness, or any of the other conditions (Sen's value restriction, Inada's conditions) that guarantee a transitive majority relation. Cost: you cannot enforce this. It is a hope about the electorate, not a property of the rule, and it fails exactly on multi-dimensional questions.

Weaken the output requirement. Ask only for a set of winners, or drop full transitivity to quasi-transitivity or acyclicity. Sen showed in 1969 that quasi-transitive social relations satisfying the other conditions exist — the Pareto extension rule is one — but Gibbard proved they concentrate power in an oligarchy: a group that is decisive collectively and each of whose members has a veto. Cost: you replace the dictator with a small committee.

Use cardinal information. The main escape in practice. Cost: you need interpersonal comparability for the numbers to mean anything, and you have to trust voters to use the scale honestly.

Allow interpersonal comparisons explicitly. Sen's programme, developed in Collective Choice and Social Welfare (1970) and the Nobel lecture. If the framework can say "this person's gain is larger than that person's loss", then possibility results abound — utilitarian and Rawlsian aggregation rules both become available and are characterised by clean axioms. Cost: interpersonal comparison of welfare is philosophically contested and empirically hard, which is why Arrow excluded it in the first place; he was following Lionel Robbins's influential 1932 argument that such comparisons are not scientific.

Infinitely many voters. Peter Fishburn showed in 1970 that the theorem fails if the electorate is infinite. Kirman and Sondermann showed in 1972 what replaces the dictator: an ultrafilter of decisive coalitions, an "invisible dictator" with no members. Cost: there are no infinite electorates.

The wall behind the wall: Gibbard–Satterthwaite

The cardinal escape is real but it does not lead into open country.

Allan Gibbard (1973) and Mark Satterthwaite (1975) proved, independently, that any deterministic voting rule with at least three possible outcomes that is onto and not dictatorial is manipulable: there is some situation in which some voter does better by submitting a ballot that misstates their preferences.

Gibbard's version is the more general and the more damaging to the escape route, because it is stated for arbitrary game forms rather than for rank-order ballots. It therefore covers score voting, approval voting and majority judgment.

And in practice the manipulability of cardinal methods is not subtle. Under approval voting, a voter must decide where to draw their approval line, and the optimal place depends on their beliefs about how everyone else will vote — which means honest approval is generally not a best response. Under score voting, a strategic voter usually maximises influence by giving 0 or the maximum and nothing in between, which collapses score voting toward approval voting. The cardinal information that lets these methods escape Arrow is exactly the information voters have an incentive to distort.

So the honest summary is: Arrow's theorem constrains ranked aggregation; cardinal methods escape it; Gibbard–Satterthwaite constrains everything. There is no rule that is simultaneously non-dictatorial, responsive and strategy-proof over three or more outcomes. What varies — and it varies a great deal — is how hard manipulation is, how much it gains, and how much damage it does when it succeeds. Those are empirical, comparative questions, and they are the ones electoral reform should be arguing about.

Was the theorem a surprise?

Arrow said later that he had expected to find a system and prove it worked; the impossibility was the result of repeatedly failing and eventually asking why. He had not read Condorcet when he proved it, and learned of the 1785 cycle afterwards.

The result's reception was not uniform. Some economists treated it as a devastating verdict on welfare economics. Others, Paul Samuelson among them, argued that Arrow had shown something about voting rather than about social welfare, since a Bergson–Samuelson social welfare function starts from a given welfare judgement rather than deriving one from individual rankings. That disagreement about what the theorem is a theorem about has never been fully settled, and it is worth knowing that it exists.

1785. Condorcet published the cycle. Its significance was not generally recognised until the twentieth century.

1932. Lionel Robbins argued in An Essay on the Nature and Significance of Economic Science that interpersonal comparisons of utility are value judgements, not science. This was widely accepted and is why Arrow's framework uses purely ordinal, non-comparable preferences. The impossibility is downstream of that methodological commitment.

1938. Abram Bergson introduced the social welfare function in a form Samuelson later developed. This framework takes a welfare ordering as given rather than constructing one from votes.

1948. Duncan Black published the median voter result and, in 1958, the book that recovered Condorcet, Borda and Dodgson.

1950. Arrow published "A Difficulty in the Concept of Social Welfare" in the Journal of Political Economy, volume 58, pages 328–346. He was working at the RAND Corporation, where the question had arisen in a game-theoretic context: whether a nation could sensibly be treated as a player with a utility function.

1951. The dissertation appeared as Social Choice and Individual Values (Wiley). Arrow was 29 when it was published.

1957. Julian Blau, in Econometrica, showed that the conditions as stated in the 1951 edition did not quite yield the conclusion — the original used citizens' sovereignty and a monotonicity condition rather than the Pareto condition, and the argument had a gap.

1963. The second edition replaced the conditions with the unanimity-based formulation now standard, and the theorem in its modern form dates from here. This is a genuine and often-omitted detail: the famous 1951 statement was not quite right, and the fix was substantive.

1969–1970. Sen's Collective Choice and Social Welfare systematised the escape routes and introduced the impossibility of a Paretian liberal, a separate and equally uncomfortable result.

1972. Arrow shared the Nobel Memorial Prize with John Hicks. At 51, he remains the youngest economics laureate.

1973 and 1975. Gibbard and Satterthwaite closed the strategy-proofness door.

2005. John Geanakoplos published "Three Brief Proofs of Arrow's Impossibility Theorem" in Economic Theory, giving the pivotal-voter argument in the compact form now standard in teaching.

It made social choice a mathematical subject. Before Arrow, arguments about voting systems were arguments about examples: here is a case where your method does something silly. After Arrow, they were arguments about axioms — state the properties you want, and either construct a rule or prove none exists. Every major result since is in this form: Gibbard–Satterthwaite, May's theorem, Sen's liberal paradox, the characterisations of the Borda count and approval voting, and Balinski and Laraki's work on majority judgment.

It disciplined welfare economics. The theorem is a direct consequence of refusing to compare welfare across people. Confronting that made the exclusion visible as a choice rather than a fact, and Sen's career is largely the project of reintroducing comparability carefully enough to be defensible. Modern welfare economics, cost-benefit analysis and the capability approach all sit downstream.

It gave constitutional design an honest starting point. A designer cannot promise a perfect system, so the useful question becomes: which failure is tolerable here, and how often will it occur? That reframing is now standard in electoral system design, in the design of committee procedures, and in the argument over ranked ballots.

It is the template for impossibility results in other fields. State desirable properties as axioms; prove they are jointly unsatisfiable. Brewer's CAP theorem for distributed systems, the FLP result on consensus, and the fair division and fair machine learning impossibility theorems all follow this pattern. Kleinberg, Mullainathan and Raghavan's 2016 result that three natural fairness criteria for risk scores cannot be satisfied simultaneously is Arrow's method applied to algorithms, and it has had a comparable effect on that field's self-understanding.

Algorithmic fairness. The impossibility results on classifier fairness — calibration, equal false positive rates and equal false negative rates cannot generally hold together when base rates differ — replicate Arrow's structure exactly. The consequence in practice is the same: you must choose which criterion to sacrifice, and pretending otherwise produces systems that fail in undeclared ways.

Rank aggregation and recommendation. Combining rankings — search signals, review sites, model leaderboards — is the same machine. The literature on metasearch and rank fusion cites Arrow directly, and the standard workaround is the standard workaround: use scores rather than ranks, and accept that the scores are manipulable.

Multi-objective machine learning. Combining several objectives into one training signal is preference aggregation. The reason weighted-sum scalarisation is used almost universally, despite its known problems with non-convex Pareto frontiers, is that it converts an ordinal aggregation problem into a cardinal one — the same escape route, with the same cost of having to justify the weights.

Preference aggregation in AI alignment. Training systems from human preference comparisons runs into Arrow's territory directly. Individual annotators supply pairwise rankings; the training procedure aggregates them into one implied ordering. Which aggregation rule is being used, and what it fails at, is usually not specified. Recent work on social choice for AI alignment takes the point seriously and asks the question in Arrow's form.

Corporate and committee governance. Boards ranking strategic options, selection committees ranking candidates, standards bodies ranking proposals — all are social welfare functions, and all inherit the theorem. Practical governance usually escapes through domain restriction (options are pre-filtered to a single dimension), through scoring rubrics (cardinal information), or through a chair with agenda power (a partial dictator). Naming which escape is in use is a useful diagnostic.

Formal statement

Let $X$ be a finite set of alternatives with $|X| \ge 3$, and let $N = {1,\dots,n}$ be a finite set of individuals with $n \ge 2$.

Let $\mathcal{L}$ denote the set of weak orderings on $X$: binary relations that are complete and transitive. A preference profile is an $n$-tuple $\succsim ,= (\succsim_1,\dots,\succsim_n) \in \mathcal{L}^n$.

A social welfare function is a map

$$F : \mathcal{L}^n \to \mathcal{L}$$

writing $\succsim^F ,= F(\succsim)$ for the social ordering it returns. The conditions:

(U) Unrestricted domain. The domain of $F$ is all of $\mathcal{L}^n$.

(P) Weak Pareto / unanimity. For all profiles and all $x,y \in X$: if $x \succ_i y$ for every $i \in N$, then $x \succ^F y$.

(IIA) Independence of irrelevant alternatives. For all profiles $\succsim$ and $\succsim'$ and all $x, y \in X$: if for every $i \in N$, $x \succsim_i y \iff x \succsim'_i y$, then $x \succsim^F y \iff x \succsim'^F y$.

(D) Non-dictatorship. There is no $i \in N$ such that for every profile and all $x,y \in X$, $x \succ_i y \implies x \succ^F y$.

Arrow's Impossibility Theorem. If $|X| \ge 3$ and $N$ is finite, there is no social welfare function $F$ satisfying (U), (P), (IIA) and (D).

Equivalently: every $F$ satisfying (U), (P) and (IIA) is dictatorial.

Note what is built into the codomain and is therefore not listed as a condition: $\succsim^F$ must be complete and transitive. Transitivity is where majority rule fails, and it is as much a condition as the four named ones.

The pivotal-voter proof

Following Geanakoplos (2005). Assume $F$ satisfies (U), (P) and (IIA).

Lemma 1 (extremal lemma). Fix $b \in X$. If in some profile every individual ranks $b$ either strictly first or strictly last, then $b$ is ranked strictly first or strictly last by $\succsim^F$.

Proof. Suppose not: then there exist $a, c \in X \setminus {b}$ with $a \succ^F b$ and $b \succ^F c$. Now modify the profile so that every individual moves $c$ strictly above $a$, leaving each individual's placement of $b$ at their own extreme untouched. This is possible precisely because $b$ is extremal for everyone, so shuffling $a$ and $c$ never disturbs anyone's ${a,b}$ or ${b,c}$ comparison.

By (IIA), $a \succ^F b$ and $b \succ^F c$ still hold. By transitivity of $\succsim^F$, $a \succ^F c$. But every individual now ranks $c$ above $a$, so (P) gives $c \succ^F a$. Contradiction. $\square$

Lemma 2 (existence of a pivot). Fix $b$. Consider the sequence of profiles $\Pi_0, \Pi_1, \dots, \Pi_n$ where in $\Pi_k$ the individuals $1,\dots,k$ rank $b$ strictly first and individuals $k+1,\dots,n$ rank $b$ strictly last (their rankings of the other alternatives held fixed throughout). In $\Pi_0$, (P) gives $b$ strictly last socially; in $\Pi_n$, (P) gives $b$ strictly first. By Lemma 1, $b$ is socially extremal at every step. Hence there is a smallest index $n^{}$ such that $b$ is socially last in $\Pi_{n^{}-1}$ and socially first in $\Pi_{n^{*}}$.

Write I for $\Pi_{n^{}-1}$ and II for $\Pi_{n^{}}$. Individual $n^{*}$ is the pivotal voter for $b$.

Lemma 3 (the pivot dictates over $X \setminus {b}$). Let $a, c \in X \setminus {b}$ be distinct. Construct profile III:

• Individual $n^{*}$ ranks $a \succ b \succ c$.
• Individuals $1,\dots,n^{*}-1$ keep $b$ strictly first, and order the rest of $X$ arbitrarily.
• Individuals $n^{*}+1,\dots,n$ keep $b$ strictly last, and order the rest of $X$ arbitrarily.

Compare III with I on the pair ${a,b}$. In I, individual $n^{}$ has $b$ last, so $a \succ_{n^{}} b$; in III, $a \succ_{n^{}} b$. Individuals before $n^{}$ have $b$ first in both, so $b \succ a$ in both. Individuals after $n^{*}$ have $b$ last in both, so $a \succ b$ in both. Every individual's ${a,b}$ comparison agrees across the two profiles, so by (IIA) the social comparison agrees. In I, $b$ was socially last, hence $a \succ^F b$. Therefore $a \succ^F b$ in III.

Compare III with II on the pair ${b,c}$. In II, individual $n^{}$ has $b$ first, so $b \succ_{n^{}} c$; in III, $b \succ_{n^{}} c$. Individuals before $n^{}$: $b$ first in both. Individuals after $n^{*}$: $b$ last in both. Again all individual ${b,c}$ comparisons agree, so by (IIA) the social comparisons agree. In II, $b$ was socially first, hence $b \succ^F c$. Therefore $b \succ^F c$ in III.

Transitivity of $\succsim^F$ now forces $a \succ^F c$ in III.

But in III the individuals other than $n^{}$ ordered $a$ and $c$ arbitrarily. By (IIA), the social ranking of ${a,c}$ depends only on the individual ${a,c}$ rankings. Hence: **whenever individual $n^{}$ ranks $a$ above $c$, society ranks $a$ above $c$, regardless of everyone else. Individual $n^{*}$ is a dictator over $X \setminus {b}$. $\square$

Closing the gap. Lemma 3 leaves pairs involving $b$ untreated. Repeat the construction with a different distinguished alternative $b' \ne b$, obtaining a pivotal voter $m^{}$ who dictates over $X \setminus {b'}$. Since $|X| \ge 3$, the sets $X \setminus {b}$ and $X \setminus {b'}$ overlap in at least one pair, and two dictators cannot disagree on that pair, so $m^{} = n^{}$. Their combined authority covers all of $X$. Hence $n^{}$ dictates over every pair, and (D) fails. $\blacksquare$

The proof uses $|X| \ge 3$ in two places: Lemma 1 needs a third alternative to build the contradiction, and the closing step needs the two punctured sets to overlap.

Where each condition is used

Condition	Where it does the work
(U)	Constructing profiles I, II, III, none of which need be realistic
(P)	Endpoints of the pivot sequence; the contradiction in Lemma 1
(IIA)	Transporting social comparisons between profiles — used four times
Transitivity of $\succsim^F$	Lemma 1 and the final step of Lemma 3

Dropping any one gives a possibility result, and each corresponds to a real escape route. Dropping (U) gives Black's single-peaked domain. Dropping transitivity gives majority rule and Condorcet cycles. Dropping (IIA) gives the Borda count and every positional scoring rule. Dropping (P) gives constant rules. Dropping finiteness of $N$ gives Fishburn's invisible dictators.

The relationship to Gibbard–Satterthwaite

Let $f : \mathcal{L}^n \to X$ be a social choice function returning a single winner. Say $f$ is strategy-proof if no individual can ever obtain a strictly preferred outcome by submitting a false ranking, and onto if every alternative wins for some profile.

Gibbard–Satterthwaite theorem (1973, 1975). If $|X| \ge 3$ and $f$ is onto and strategy-proof, then $f$ is dictatorial.

The two theorems are closely linked: strategy-proofness implies a monotonicity condition which, with ontoness, yields an Arrovian social welfare function, and several published proofs derive one from the other.

The scope difference is what matters for the cardinal escape. Arrow's theorem is stated for rules taking rank-order ballots, so score and approval voting are outside it. Gibbard's 1973 formulation is stated for arbitrary game forms — any procedure mapping strategy profiles to outcomes — and therefore covers them. Gibbard extended the result to decision schemes with randomised outcomes in 1977, where the conclusion becomes that any strategy-proof scheme is a random dictatorship or a randomisation over pairs.

The practical upshot for range and approval voting: they are not dictatorial, are not covered by Arrow, and are manipulable. A voter's optimal ballot depends on their beliefs about others, and the equilibrium behaviour of strategic score voters tends toward approval-style min-max ballots. Whether that is a serious problem in practice — how much strategic voting occurs, and how much it distorts — is an empirical question, and it is a much better question than the one Arrow's theorem is usually invoked to close down.