{
  "id": "arrows-impossibility-theorem",
  "title": "Arrow's Impossibility Theorem",
  "lang": "en",
  "source_lang": "en",
  "summary": "Arrow’s impossibility theorem shows that no rule can aggregate unrestricted individual rankings into a coherent social ranking while satisfying several basic conditions.",
  "category": "economics",
  "subcategory": "microeconomics",
  "type": "concept",
  "aliases": [
    "Arrow’s theorem",
    "Arrow’s general possibility theorem",
    "General impossibility theorem"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/arrows-impossibility-theorem/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/arrows-impossibility-theorem/",
      "json": "https://aiwiki.page/zh/arrows-impossibility-theorem.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Arrow’s impossibility theorem is a foundational result in social choice theory establishing a conflict among requirements for combining individual preferences into a collective ranking. With at least three alternatives and a finite electorate, no aggregation rule can always produce a complete, transitive social ordering while simultaneously allowing unrestricted individual preferences, respecting unanimous strict preferences, making each pairwise social comparison depend only on individual comparisons of that pair, and avoiding dictatorship. Here, dictatorship has a technical meaning: one fixed individual’s strict preferences always determine the corresponding strict social comparisons.\n\nFormulated by Kenneth Arrow and published in its mature form in Social Choice and Individual Values in 1951, the theorem transformed welfare economics, voting theory, and the study of collective decision-making. It demonstrates a logical incompatibility rather than a defect peculiar to one electoral system. Its significance depends on the precise information and output the model allows: ordinal individual rankings are combined into a social ranking, without interpersonal comparisons of welfare. Restricted preferences, richer welfare information, different consistency requirements, and procedures that select winners rather than complete rankings lead to different possibilities."
    },
    {
      "heading": "Historical origins and intellectual setting",
      "text": "The problem of reconciling individual judgments with collective decisions long predates Arrow. In eighteenth-century France, Jean-Charles de Borda and the Marquis de Condorcet developed contrasting approaches to voting. Borda advocated assigning points according to candidates’ positions in voters’ rankings. Condorcet examined pairwise majority comparisons and identified the possibility of a cycle: a majority might prefer one candidate to a second, another majority the second to a third, and another majority the third to the first. These investigations established that individually consistent preferences need not produce a collectively consistent majority relation.\n\nA related problem arose in twentieth-century welfare economics. Economists sought to compare economic arrangements without assuming that one person’s satisfaction could be measured on the same scale as another’s. The development of ordinal utility theory emphasized rankings rather than measurable quantities of pleasure or satisfaction. The Pareto criterion supplied a limited basis for comparison: a change benefits society in the Pareto sense if it benefits someone without harming anyone. But this criterion generally leaves alternatives incomparable whenever different people gain and lose. A broader theory of social evaluation therefore required some account of how conflicting interests should be aggregated.\n\nAbram Bergson’s 1938 formulation of a social welfare function and Paul Samuelson’s subsequent work provided an important background. Their framework represented social evaluations through an ordering or numerical function over economic states, with ethical judgments determining how individual interests entered that evaluation. Arrow asked a distinct aggregation question: whether a rule could construct a social ordering from every admissible profile of individual orderings while satisfying specified conditions. The phrase social welfare function consequently has related but different meanings in these traditions. Arrow’s version is a mapping from preference profiles to social orderings, not simply a numerical index of aggregate welfare.\n\nArrow published “A Difficulty in the Concept of Social Welfare” in 1950 and Social Choice and Individual Values in 1951. A second edition of the book appeared in 1963, with additional discussion and developments. Later expositions often use a streamlined combination of unrestricted domain, weak Pareto, independence of irrelevant alternatives, and nondictatorship, with collective rationality built into the definition of the output. These formulations should not be treated as verbatim statements of every condition in the original presentation. Arrow shared the 1972 Nobel Memorial Prize in Economic Sciences with John Hicks for contributions to general equilibrium theory and welfare theory; the impossibility theorem was a central part of Arrow’s contribution to the latter."
    },
    {
      "heading": "The mathematical framework",
      "text": "Let the electorate be a finite set \\(N=\\{1,\\ldots,n\\}\\), normally with \\(n\\geq 2\\), and let \\(X\\) be a set of alternatives containing at least three elements. Alternatives may be candidates, public policies, economic allocations, or complete descriptions of social states. Each individual \\(i\\) has a preference relation \\(R_i\\) on \\(X\\). The expression \\(xR_i y\\) means that individual \\(i\\) regards \\(x\\) as at least as good as \\(y\\). Its strict part is defined by\n\n\\[xP_i y \\quad\\Longleftrightarrow\\quad xR_i y\\ \\text{and not}\\ yR_i x.\\]\n\nAn ordering is complete if any two alternatives can be compared and transitive if \\(xR_i y\\) and \\(yR_i z\\) imply \\(xR_i z\\). Completeness does not prohibit indifference: both \\(xR_i y\\) and \\(yR_i x\\) may hold. A complete and transitive relation is often called a weak ordering. A strict ranking, by contrast, has no ties between distinct alternatives. Arrow’s theorem has standard formulations allowing individual indifference, as well as versions restricting individuals to strict rankings. The impossibility already arises on the latter domain.\n\nA preference profile is the list \\(\\mathbf R=(R_1,\\ldots,R_n)\\). Writing \\(\\mathcal R(X)\\) for the set of all complete and transitive relations on \\(X\\), an Arrowian aggregation rule has the form\n\n\\[F:\\mathcal R(X)^n\\longrightarrow \\mathcal R(X).\\]\n\nThe social ordering at profile \\(\\mathbf R\\) is \\(R=F(\\mathbf R)\\), with strict part \\(P\\). The codomain expresses an important requirement: the rule must return a complete and transitive social ordering, not merely a candidate, an unresolved set of alternatives, or a potentially cyclic collection of pairwise judgments. Some presentations list this requirement separately as collective rationality; others incorporate it into the definition of a social welfare function.\n\nThe information entering \\(F\\) is ordinal. An individual’s ranking reveals that one alternative is preferred to another but not how much better it is. A utility function can represent such a ranking, but applying a strictly increasing transformation to that function leaves the ranking unchanged. Consequently, the aggregation problem contains no inherent measure of preference intensity and no interpersonal scale for comparing utility differences. Two people who rank \\(x\\) above \\(y\\) may differ dramatically in the importance they attach to that comparison, yet their ordinal inputs record the same direction.\n\nThis framework also abstracts from institutional details. It does not assume a particular ballot design, campaign process, legislature, or administrative procedure. Nor does it initially ask whether people truthfully report their preferences. Instead, it examines whether any mapping from the specified inputs to the specified outputs can satisfy the axioms. That generality is the source of the result’s breadth, but also the reason its assumptions must be checked before applying it to a concrete institution."
    },
    {
      "heading": "The conditions and the impossibility statement",
      "text": "Unrestricted domain, also called universal domain, requires the rule to accept every profile of individually complete and transitive preferences. Individuals may disagree in any pattern; no common ideological axis, shared ranking of some alternatives, or restriction on how preferences vary is assumed. This is a requirement about the mathematical domain of the rule, not a prediction that every conceivable profile occurs frequently. A rule fails the universal requirement if even one admissible profile causes the other conditions to break down.\n\nWeak Pareto, or unanimity, requires a strict social preference whenever every individual has the same strict preference:\n\n\\[\\bigl(\\forall i\\in N,\\ xP_i y\\bigr)\\quad\\Longrightarrow\\quad xPy.\\]\n\nIt does not prescribe an outcome when some people prefer \\(x\\) and others prefer \\(y\\), nor does it require the society to maximize a measurable sum of welfare. Its role is to prevent collective judgments from reversing unanimous strict agreement. The adjective “weak” distinguishes this condition from stronger Pareto requirements that also address profiles containing individual indifferences.\n\nIndependence of irrelevant alternatives, usually abbreviated IIA, restricts the information that may determine a social comparison. If two profiles agree on every individual’s comparison of \\(x\\) with \\(y\\), then their social comparisons of \\(x\\) with \\(y\\) must also agree:\n\n\\[\\left[ \\forall i,\\ R_i|_{\\{x,y\\}}=R_i'|_{\\{x,y\\}} \\right] \\Longrightarrow F(\\mathbf R)|_{\\{x,y\\}} = F(\\mathbf R')|_{\\{x,y\\}}.\\]\n\nIndividuals may reorder other alternatives or move them relative to \\(x\\) and \\(y\\), provided their comparison of \\(x\\) and \\(y\\) remains unchanged. Thus IIA is a cross-profile informational restriction. It is not simply a rule about deleting a losing candidate from an election, and it is not equivalent to immunity from strategic voting. Its appeal is that a social judgment between two alternatives should reflect how people compare those alternatives. Its demanding feature is that it excludes potentially relevant information encoded elsewhere in their rankings.\n\nNondictatorship requires that there be no individual \\(d\\in N\\) such that, at every admissible profile and for every pair,\n\n\\[xP_d y\\quad\\Longrightarrow\\quad xPy.\\]\n\nA dictator in this definition need not choose the agenda, hold public office, or have authority outside the aggregation rule. The definition concerns the decisive role of that individual’s strict preferences. If individual indifference is permitted, it does not specify how the society resolves pairs on which the dictator is indifferent. Nondictatorship is also weaker than equal treatment: a nondictatorial rule may still assign unequal influence to different individuals.\n\nThe theorem states that no aggregation rule satisfies all these conditions when \\(N\\) is finite and \\(X\\) contains at least three alternatives. Equivalently, any rule satisfying unrestricted domain, collective rationality, weak Pareto, and IIA must be dictatorial. This is a conditional mathematical statement. It does not rank the conditions by ethical importance or establish that each is indispensable to every practical decision procedure."
    },
    {
      "heading": "Majority cycles and concrete rule failures",
      "text": "The simplest illustration of the underlying difficulty is the Condorcet paradox. Consider three voters with strict preferences:\n\nVoter | Ranking\n1 | \\(a \\succ b \\succ c\\)\n2 | \\(b \\succ c \\succ a\\)\n3 | \\(c \\succ a \\succ b\\)\n\nVoters 1 and 3 prefer \\(a\\) to \\(b\\); voters 1 and 2 prefer \\(b\\) to \\(c\\); and voters 2 and 3 prefer \\(c\\) to \\(a\\). Pairwise majority voting therefore produces \\(aPb\\), \\(bPc\\), and \\(cPa\\). Every individual ranking is transitive, but the majority relation is not. On strict-preference profiles with an odd number of voters, pairwise majority respects unanimity, treats the comparison of each pair independently, and is nondictatorial. Its failure is collective transitivity.\n\nThe cycle also creates scope for agenda dependence. Suppose a sequential procedure first holds a contest between two alternatives and then pits the winner against the third. Comparing \\(a\\) with \\(b\\) first produces \\(a\\), which then loses to \\(c\\). Comparing \\(b\\) with \\(c\\) first produces \\(b\\), which then loses to \\(a\\). Comparing \\(c\\) with \\(a\\) first produces \\(c\\), which then loses to \\(b\\). Each alternative can therefore emerge as the final winner under a suitable sequence of pairwise votes. This example isolates the role of the agenda; it does not imply that all preference profiles are cyclic or that every sequential procedure is equally manipulable.\n\nThe Borda count avoids this particular problem by assigning points to ranking positions and ordering alternatives by total score. With three alternatives, assign two points for first place, one for second, and zero for third. Summed scores produce a complete and transitive social ordering, allowing ties. But the method violates IIA. Consider the following two profiles:\n\nVoter | Profile A | Profile B\n1 | \\(a \\succ b \\succ c\\) | \\(a \\succ c \\succ b\\)\n2 | \\(a \\succ b \\succ c\\) | \\(a \\succ c \\succ b\\)\n3 | \\(b \\succ c \\succ a\\) | \\(b \\succ a \\succ c\\)\n\nAt Profile A, \\(a\\) receives four points and \\(b\\) receives four, so society ties them. At Profile B, \\(a\\) receives five points and \\(b\\) receives two, so society strictly prefers \\(a\\). Yet each voter’s comparison of \\(a\\) with \\(b\\) is unchanged: voters 1 and 2 prefer \\(a\\), while voter 3 prefers \\(b\\). The altered placement of \\(c\\) changes the social comparison of the original pair.\n\nOther simple rules make the trade-offs equally explicit. Copying one fixed voter’s ranking satisfies unrestricted domain, collective rationality, unanimity, and IIA, but violates nondictatorship. Returning the same fixed social ranking for every profile satisfies collective rationality and IIA, but violates unanimity whenever everyone strictly prefers the reverse of one of its comparisons. These examples show why no single familiar failure should be identified with the whole theorem. Majority cycles demonstrate that majority aggregation can fail; Arrow’s result establishes that no alternative aggregation rule preserves the entire package."
    },
    {
      "heading": "How the proof forces a dictator",
      "text": "A standard proof strategy studies decisive coalitions. A coalition \\(S\\subseteq N\\) is decisive for an ordered pair \\((x,y)\\) if, whenever every member of \\(S\\) strictly prefers \\(x\\) to \\(y\\), society strictly prefers \\(x\\) to \\(y\\), regardless of the preferences of people outside \\(S\\). A coalition is globally decisive if this holds for every ordered pair of distinct alternatives. Unanimity makes the entire electorate globally decisive. A singleton globally decisive coalition is precisely a dictator in the theorem’s technical sense.\n\nAn intermediate concept is decisiveness under polarization: everyone in \\(S\\) ranks \\(x\\) above \\(y\\), while everyone outside \\(S\\) ranks \\(y\\) above \\(x\\). In proof terminology this is sometimes called almost decisiveness. IIA makes such polarized profiles useful because their pairwise social outcomes cannot depend on how a third alternative is placed. Unrestricted domain permits the proof to construct those placements freely, and social transitivity connects the resulting pairwise conclusions. Standard field-expansion arguments show that a coalition’s power over one pair in such a configuration extends to other pairs and ultimately to decisiveness without restrictions on outsiders.\n\nThe central contraction argument can be seen by supposing that a decisive coalition \\(S\\) contains at least two people and dividing it into two nonempty parts \\(A\\) and \\(B\\). Consider three alternatives and assign rankings as follows:\n\nGroup | Ranking\n\\(A\\) | \\(x \\succ y \\succ z\\)\n\\(B\\) | \\(y \\succ z \\succ x\\)\n\\(N\\setminus S\\) | \\(z \\succ x \\succ y\\)\n\nAll members of \\(S\\) prefer \\(y\\) to \\(z\\), so decisiveness gives \\(yPz\\). Completeness of the social ordering implies either \\(xR y\\) or \\(yPx\\). In the first case, transitivity together with \\(yPz\\) gives \\(xPz\\). Only members of \\(A\\) rank \\(x\\) above \\(z\\), so \\(A\\) has prevailed in a polarized comparison. In the second case, only members of \\(B\\) rank \\(y\\) above \\(x\\), so \\(B\\) has prevailed in a polarized comparison. Applying the field-expansion result makes one of these proper subcoalitions decisive.\n\nBecause the electorate is finite, the argument can be repeated only finitely many times before a singleton decisive coalition is reached. Alternatively, start with a decisive coalition of minimum size; contraction shows that its size cannot exceed one. The resulting individual dictates all strict social comparisons. This contradicts nondictatorship and completes the proof strategy. The steps involving field expansion are substantive lemmas, not merely assumptions about how coalition power behaves: they derive that behavior from IIA, unanimity, unrestricted domain, and transitivity.\n\nThis structure explains several otherwise surprising features of the theorem. The decisive person is fixed across profiles, not selected anew for each election. Three alternatives matter because the proof uses one comparison to transmit restrictions to another through transitivity. Finiteness matters because a decisive coalition can then be reduced to an individual. The proof does not assume majority rule, voter symmetry, numerical scores, or preference intensity. Nor does it require anyone to misreport preferences: the contradiction concerns the aggregation of the preferences themselves."
    },
    {
      "heading": "Scope, boundary cases, and common misinterpretations",
      "text": "The restriction to at least three alternatives is essential. With two alternatives, a society can use majority rule and treat a tie in the vote as social indifference. This produces a complete and transitive relation, respects unanimity, satisfies IIA, and is nondictatorial for electorates of at least two people. Transitivity creates no three-alternative constraint because there is no third distinct alternative. If a strict winner rather than a weak social ordering is required, tied votes introduce an additional design question, but they do not recreate Arrow’s three-alternative impossibility.\n\nThe theorem is also about a rule defined over a domain of profiles, not about a single observed profile. An election may produce a clear, transitive majority ordering without contradicting Arrow. Even a long sequence of unproblematic elections does not establish that a rule satisfies the conditions universally. Conversely, the theorem gives no probability that a randomly selected electorate will generate a cycle or another failure. Such probabilities depend on assumptions about the distribution and correlation of preferences, the number of alternatives, and the voting rule.\n\nA complete social ordering is a stronger output than a single winner. A social choice function selects an alternative, while a choice correspondence may return a nonempty set of alternatives. A procedure that finds an acceptable winner without ranking every pair does not have exactly the codomain in Arrow’s model. It may face other impossibility theorems or consistency problems, but these require separate definitions and proofs. Similarly, allowing a social relation to be incomplete or to contain some cycles changes the original problem rather than disproving the result.\n\nArrow’s theorem does not show that democracy is impossible, meaningless, or necessarily inferior to dictatorship. Democracy includes representation, rights, deliberation, public accountability, agenda formation, and constitutional constraints, none of which is exhausted by the aggregation map. The theorem’s “dictatorship” is a formal property arising under a specified package of assumptions, not a recommendation for government. It also does not show that interpersonal judgments of well-being are impossible. Those judgments are excluded from the theorem’s ordinal informational framework, rather than refuted by it.\n\nFinally, the standard finite-electorate qualification is mathematically important. For infinite electorates, decisive coalitions can have the structure of a nonprincipal ultrafilter: no single person is decisive even though the coalition structure supports an Arrow-type aggregation rule. Such constructions can depend on set-theoretic existence assumptions and may lack attractive anonymity or computational properties. They reveal the precise role of finiteness in the proof; they do not supply ordinary electoral procedures for finite populations."
    },
    {
      "heading": "What changes when an assumption is relaxed",
      "text": "A major response is to restrict the domain of preferences. In single-peaked preferences, alternatives lie on a common linear axis, and each voter has a most-preferred position, with alternatives becoming less preferred as one moves away from that peak on either side. Duncan Black’s work, including his 1948 analysis of committee decisions, established the importance of this restriction. For a finite set of alternatives and an odd number of voters with strict single-peaked rankings, pairwise majority voting is transitive. The median voter theorem further identifies the median peak as a majority winner in the standard one-dimensional setting.\n\nThe restriction does real work. Preferences concerning a single tax rate may sometimes admit a common ordering, whereas preferences over packages combining taxation, environmental regulation, defense, and regional expenditure need not be single-peaked on any common axis. Even on a one-dimensional policy variable, individuals can have preferences that violate single-peakedness. Domain restrictions therefore produce conditional possibility results, not a universal solution. Their relevance depends on the structure of the alternatives and on the actual or institutionally admissible preferences.\n\nAnother approach accepts violations of IIA. Positional scoring rules use where alternatives stand in full rankings, not just the direction of each pairwise comparison. This information can distinguish a first-versus-last comparison from a first-versus-second comparison, although ordinal ranks do not themselves measure welfare intensity. Such methods produce coherent score orderings, but the placement or introduction of other alternatives can change the relation between an existing pair. Abandoning IIA makes that dependence permissible; it does not establish that every form of dependence is desirable.\n\nA different approach enriches the welfare information. If individual utilities have appropriately comparable units, a utilitarian evaluation can rank alternatives using\n\n\\[W(x)=\\sum_{i=1}^{n}u_i(x).\\]\n\nThe expression is numerically coherent, but it requires more structure than ordinal rankings provide. Independently rescaling one person’s utility can change the resulting social ordering. Under purely ordinal information, such rescaling cannot change any individual preference and therefore should not alter what is supposedly inferred from those preferences alone. Distribution-sensitive welfare functions likewise require explicit judgments about interpersonal comparisons, weights, or priority. Arrow’s theorem clarifies why these judgments cannot simply be extracted from unrestricted ordinal rankings.\n\nCollective rationality can also be weakened. An incomplete social relation can leave some conflicts unresolved; an acyclic relation can permit a maximal alternative without satisfying full transitivity. Alternatively, a procedure can output a lottery rather than a deterministic ranking. These changes open different mathematical possibilities, but their implications depend on how preferences over sets or lotteries are defined. A lottery requires a probability distribution over outcomes, and evaluating it may require assumptions about risk attitudes or expected utility. Each relaxation changes the information, output, or consistency demanded of collective choice."
    },
    {
      "heading": "Related theorems and institutional applications",
      "text": "Arrow’s theorem concerns preference aggregation, whereas the Gibbard–Satterthwaite theorem concerns strategic reporting in deterministic winner selection. In its standard form, the latter states that a social choice function on unrestricted strict rankings, with at least three alternatives in its range, cannot be both strategy-proof and nondictatorial. Allan Gibbard published his result in 1973 and Mark Satterthwaite in 1975. Strategy-proofness means that no individual can obtain a more-preferred outcome by falsely reporting a ranking, whatever others report. The connection is close, but Arrowian IIA and strategy-proofness are different properties.\n\nThis distinction matters in mechanism design. A voting rule can violate IIA yet be analyzed for incentives, and an aggregation rule can satisfy IIA without solving the problem of strategic behavior. Domain restrictions may alter both problems: single-peaked preferences, for example, support useful strategy-proof mechanisms based on median positions. Transfers, richer messages, and restrictions on feasible outcomes also change the design environment. Results from incentive compatibility therefore complement Arrow’s theorem rather than follow from a simple identification of independence with truthfulness.\n\nArrow’s result also has a coalition-theoretic relationship with Nakamura’s theorem. For a voting game represented by winning coalitions, the Nakamura number is the smallest number of winning coalitions whose intersection is empty, when such a collection exists. It determines a threshold for guaranteeing a nonempty core—the alternatives not defeated by an admissible winning coalition—across preference profiles. This framework connects the size of the agenda with the possibility of collective cycles. It asks a different question from Arrow’s full-ranking problem but makes similarly explicit how institutional structure and the number of alternatives constrain collective consistency.\n\nIn practical evaluation, the theorem supplies a vocabulary for identifying trade-offs. Ranked electoral methods can be examined for their use of pairwise information, score information, tie-breaking conventions, and agenda restrictions. Expert panels and recommendation systems can be modeled as aggregating rankings when their inputs genuinely have that form. If instead a system aggregates calibrated measurements, monetary quantities, or probabilities, its informational setting is different. Calling every aggregation difficulty “Arrow’s theorem” obscures the specific assumptions needed for the result.\n\nThe same distinction is important in economic policy analysis. Cost–benefit analysis typically uses monetary valuations rather than ordinal rankings alone; distributional welfare analysis may use income weights or explicit interpersonal welfare comparisons. Constitutional rules and rights can remove alternatives from the feasible set, while deliberation can alter preferences before an aggregation procedure is applied. These practices do not refute the theorem. They locate particular decisions in frameworks with additional information, restricted domains, or different objectives, where the relevant consistency and incentive properties must be established separately."
    }
  ],
  "links": [
    "social-choice-theory",
    "kenneth-arrow",
    "welfare-economics",
    "utility-economics",
    "pareto-efficiency",
    "nobel-memorial-prize-in-economic-sciences",
    "utility-function",
    "democracy",
    "probability-distribution",
    "mechanism-design",
    "incentive-compatibility",
    "cost-benefit-analysis"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}