Social choice theory is the systematic study of how a group can make decisions or evaluate alternatives when its members have different preferences, interests, or judgments. Its subjects include voting rules, the measurement of social welfare, the distribution of resources, and the aggregation of judgments. A central question is whether collective decisions can satisfy several reasonable requirements at once—for example, equal treatment of individuals, respect for unanimous agreement, consistency, and resistance to strategic manipulation.
The field lies at the intersection of economics, political science, philosophy, and mathematics. Its modern foundations were established by Kenneth Arrow, whose work showed that apparently modest conditions on preference aggregation can be mutually incompatible. Subsequent research has identified both further impossibilities and constructive possibilities: restrictions on preferences, richer information about well-being, randomized procedures, and alternative conceptions of fairness can support different forms of collective decision-making.
Social choice theory does not identify a single universally best decision rule. It clarifies what rules accomplish, which values they embody, and what trade-offs their use requires. These questions arise not only in elections but also in committee decisions, public spending, resource allocation, and automated systems that combine the evaluations of many people.
Historical development and intellectual foundations
The formal study of collective choice predates modern economics. In the late thirteenth century, Ramon Llull described procedures involving comparisons among candidates. In the eighteenth century, Jean-Charles de Borda and the Marquis de Condorcet developed contrasting approaches to elections. Borda presented a rank-based scoring method to the French Academy of Sciences in 1770 and published an account in 1781. Condorcet’s 1785 essay examined majority comparisons and the conditions under which collective decisions might be reliable.
Their work exposed a distinction that remains fundamental. A scoring procedure evaluates a candidate through positions in voters’ complete rankings; a pairwise procedure asks how that candidate fares against each opponent separately. These approaches can select different winners even when they use exactly the same ballots. Condorcet also demonstrated that majority preferences can cycle, although each voter has a consistent ranking. Charles Lutwidge Dodgson, better known as Lewis Carroll, later proposed a method based on the minimum number of adjacent changes in individual rankings needed to turn a candidate into a pairwise-majority winner.
Another intellectual strand developed through welfare economics. Utilitarian approaches assessed arrangements by the welfare they produced, while ordinal economics emphasized that observed choices generally reveal rankings rather than directly measurable quantities of satisfaction. Abram Bergson’s 1938 formulation of a social welfare function made explicit that evaluating economic arrangements requires ethical judgments about individual interests. Paul Samuelson subsequently incorporated such functions into welfare analysis. This tradition asked how social evaluations should be constructed, rather than treating market outcomes as sufficient measures of social desirability.
Duncan Black’s 1948 analysis of committee decisions established an important positive result for preferences organized along a single dimension. Arrow’s Social Choice and Individual Values, published in 1951 and revised in 1963, then placed preference aggregation in a general axiomatic framework. Amartya Sen’s Collective Choice and Social Welfare appeared in 1970 and broadened the field’s treatment of information, rights, and welfare comparisons. Arrow shared the 1972 Nobel Memorial Prize in Economic Sciences with John Hicks; Sen received the prize in 1998 for contributions to welfare economics.
The field subsequently expanded beyond electoral rankings. Research on strategic behavior connected it with mechanism design, while work on allocation, judgments, and computation examined collective decisions whose outcomes cannot be represented simply as the selection of one candidate.
Preferences, information, and collective rules
A standard model begins with a set of individuals and a set of alternatives . Alternatives may be candidates, policy packages, allocations of goods, or complete descriptions of social states. Individual has a preference relation , where means that regards as at least as good as . Its strict component, , records strict preference; its indifference component records equal ranking.
Preferences are commonly assumed to be complete and transitive. Completeness means that any two alternatives can be compared. Transitivity means that and imply . Together these properties define a weak ordering. They do not require selfishness: an individual may rank alternatives according to personal consumption, other people’s welfare, moral commitments, or a combination of concerns.
A preference profile is the list . An Arrovian social welfare function maps each admissible profile to a social ordering:
where is the domain of permissible profiles and is the set of weak orderings over . A social choice function instead selects an alternative, while a social choice correspondence can select a nonempty set. These distinctions matter because producing a complete ranking is more demanding than identifying an acceptable winner.
The information supplied to a rule is equally important. Ordinal information identifies rankings but not the intensity of preference. A utility function represents ordinal preferences if exactly when . Any strictly increasing transformation preserves the same ranking. Consequently, summing arbitrary ordinal utility representations has no invariant meaning: changing the numerical representation can change the apparent social optimum without changing anyone’s preferences.
Cardinal utility supplies additional structure, but cardinality does not automatically establish interpersonal comparability. A scale may meaningfully represent one individual’s utility differences while leaving comparisons with another individual’s differences undefined. Social choice theory therefore distinguishes comparability of utility levels, comparability of differences, and the transformations under which a social evaluation must remain unchanged.
Axioms specify how rules should use their information. Anonymity requires equal treatment of individuals; neutrality requires equal treatment of alternatives. Unanimity respects agreement, while consistency conditions regulate how choices respond to changes in rankings or available options. These are separate requirements: treating voters equally does not guarantee equal treatment of candidates, and a decisive rule need not produce a transitive social ranking.
Majority rule, voting cycles, and electoral procedures
For two alternatives and an odd number of voters with strict preferences, simple majority rule selects the alternative supported by more voters. Kenneth May’s 1952 characterization showed that, in a binary setting allowing individual indifference, majority rule is uniquely identified by conditions including anonymity, neutrality, and positive responsiveness. Positive responsiveness requires that movement in favor of an alternative improve its collective position in the appropriate way. The result gives majority rule a precise justification without extending that justification automatically to choices among three or more alternatives.
The difficulty with larger choice sets is illustrated by the Condorcet paradox. Suppose three voters have the rankings
A majority prefers to , a majority prefers to , and a majority prefers to . Every individual ranking is transitive, but the majority relation is not. No alternative is a Condorcet winner, meaning an alternative that defeats every other alternative in a pairwise majority contest.
Cycles make procedures consequential. If alternatives face sequential elimination votes, the order of comparisons can determine the final winner. In the example, comparing with first leaves , which then loses to . Comparing with first instead produces as the eventual winner. This is agenda dependence, not an inconsistency in any individual’s reasoning.
Other electoral rules avoid some difficulties while introducing others. Plurality counts first choices. The Borda count assigns points by rank, conventionally with candidates. Instant-runoff voting repeatedly eliminates the candidate with the fewest first-place votes and transfers ballots to remaining candidates. Condorcet methods select the Condorcet winner when one exists and use additional principles when none exists. Approval voting lets voters support multiple candidates, but an ordinal ranking alone does not specify where each voter places the approval threshold.
No method satisfies every commonly proposed criterion. Plurality can reject a Condorcet winner, Borda rankings can depend on candidates who do not win, and instant-runoff voting can violate monotonicity: in some profiles, raising a candidate on certain ballots can cause that candidate to lose. These examples concern specified mathematical properties, not a universal ranking of electoral systems. Their practical significance also depends on nomination rules, ballot information, turnout, and whether the task is to choose one representative or a representative assembly.
Arrow’s impossibility theorem
Arrow’s impossibility theorem concerns the construction of a collective ordering from individual orderings. With at least three alternatives and a finite electorate, no aggregation rule can simultaneously satisfy unrestricted domain, the Pareto condition, independence of irrelevant alternatives, and non-dictatorship while always producing a complete and transitive social ordering.
Unrestricted domain requires the rule to accept every profile of individual weak orderings. The Pareto condition requires society to rank strictly above whenever every individual does so. Independence of irrelevant alternatives, usually abbreviated IIA, requires the social comparison of and to depend only on individuals’ comparisons of those two alternatives. If those pairwise comparisons remain unchanged between two profiles, the social comparison must also remain unchanged, regardless of how individuals rank other alternatives.
Non-dictatorship excludes an individual whose strict preference between any two alternatives always determines the strict social preference between them. A dictator in this technical sense need not hold political office or exercise coercion. The definition concerns the aggregation rule: one person’s strict ordering systematically overrides everyone else’s, while additional conventions may resolve that person’s indifferences.
Proofs commonly study decisive coalitions: groups whose preferences can force a social comparison. Pareto agreement makes the full electorate decisive. Independence and transitivity allow decisiveness to propagate across comparisons. Under the theorem’s other assumptions, this structure implies that some individual is decisive. Non-dictatorship rules out precisely that implication.
The theorem does not establish that democracy is impossible, that every election is arbitrary, or that every voting system is equally defensible. It establishes the incompatibility of a particular collection of requirements. Some procedures sacrifice independence; some restrict the domain; some return a winner rather than a complete ordering; others incorporate information beyond rankings. Each response changes the problem and must be assessed on its own terms.
Arrow’s IIA also requires careful interpretation. It compares social rankings across preference profiles; it is not simply the demand that an election winner remain unchanged whenever a losing candidate withdraws. Candidate-removal properties are related concerns but belong to a different formal setting. Likewise, the theorem does not prohibit ethical evaluation using interpersonal welfare information. Such evaluation lies outside the theorem’s purely ordinal, noncomparable informational framework.
Restricted domains and possibilities for coherent choice
Impossibility results depend on which preferences a rule must accommodate. If admissible profiles have additional structure, collective decisions may satisfy requirements that are incompatible on an unrestricted domain. The most influential example is single-peaked preferences on a common ordering of alternatives.
Consider policies ordered from low to high on one dimension, such as a spending level. A voter has a most-preferred point, or peak, and regards movement away from that peak on either side as worsening the outcome. Comparisons across opposite sides need not depend on distance in any particular way. The critical assumption is that all voters’ preferences are single-peaked relative to the same ordering of policies.
For an odd electorate with strict single-peaked preferences, the alternative at the median voter’s peak is a Condorcet winner. If the median peak is , any proposal to its left is opposed in favor of by the median voter and everyone with peaks farther right. The corresponding argument applies to proposals on the right. More generally, single-peakedness prevents strict majority cycles in the usual odd-electorate setting.
This result underlies the median voter theorem, but its implications are conditional. It does not establish that all elections are won by a politically moderate candidate. Real choices can involve several policy dimensions, voters may disagree about how alternatives should be ordered, and some preferences may have multiple peaks. Models of party competition also require assumptions about turnout, candidate objectives, and how voters respond to platforms.
Domain restrictions can improve incentives as well as consistency. On a line, selecting the median of reported peaks is strategy-proof for single-peaked voters. Holding others’ reports fixed, a voter cannot move the outcome closer to the voter’s true peak by reporting a false one. Generalized median rules extend this logic using fixed additional points, sometimes called phantom voters, to characterize broader families of procedures.
Other restrictions include single-crossing preferences, where voters can be ordered so that support for one alternative over another changes at most once along that order. Restrictions on the available alternatives can also matter. In spatial models with several dimensions, majority stability becomes much harder to obtain because coalitions can support movements in different directions. Domain assumptions are therefore substantive claims about the decision environment, not merely technical devices for avoiding a theorem.
Strategic behavior and truthful preference revelation
Preference aggregation faces a second problem beyond consistency: people may misrepresent their preferences to influence the result. A voter who favors may support because appears unable to win and is preferable to the likely winner . Social choice theory distinguishes such strategic reporting from a genuine change in the voter’s evaluation of outcomes.
A direct rule is strategy-proof if truthful reporting is a weakly dominant strategy for every participant. Holding everyone else’s reports fixed, no individual can obtain a strictly preferred outcome by submitting a false ranking. This is a demanding guarantee: it does not rely on accurate beliefs about other participants or on a favorable equilibrium.
The Gibbard–Satterthwaite theorem, established independently by Allan Gibbard in 1973 and Mark Satterthwaite in 1975, states that a deterministic, strategy-proof, onto social choice function on an unrestricted domain of strict rankings over at least three alternatives must be dictatorial. Onto means that every alternative can be selected at some profile. The theorem concerns a single selected outcome, unlike Arrow’s theorem about a complete social ordering.
There are several ways to change this setting. With only two alternatives, majority rule is strategy-proof. Restricted preferences can support non-dictatorial strategy-proof rules, as with the median rule on a single-peaked domain. Rules using payments introduce further possibilities. In auction theory and public-project mechanisms, appropriately designed transfers can align individual incentives with a social objective under assumptions such as quasilinear utility. Quasilinearity means that money enters utility additively, without changing the ranking of nonmonetary outcomes through income effects. These mechanisms face separate constraints involving budgets, distribution, and participation.
Randomization offers another approach. Under strict rankings, random dictatorship selects a voter according to fixed probabilities and implements that voter’s top choice. Equal probabilities treat voters symmetrically. The procedure is strategy-proof and ex post Pareto efficient: no selected alternative is unanimously inferior to another. The designation “random dictatorship” describes how each realized choice is determined, not a permanent assignment of political authority.
These theoretical guarantees should be distinguished from practical manipulability. A rule that permits profitable misrepresentation in some profile may still be difficult to manipulate when preferences are uncertain. Conversely, computational complexity alone does not guarantee honest behavior in ordinary cases. The link with incentive compatibility makes the field attentive both to which outcomes a procedure prescribes and to what information participants have incentives to supply.
Welfare aggregation and interpersonal comparisons
Voting models often treat ordinal rankings as the only admissible information. Social evaluation can instead incorporate the magnitude and distribution of benefits. This changes the analytical problem: rather than asking only how rankings should be combined, it asks which aspects of individual well-being can be compared and which comparisons are ethically relevant.
A familiar representation assigns a welfare number to each individual and evaluates a social state through
where are welfare weights. With equal weights and appropriately comparable utilities, this is a utilitarian objective. Its interpretation requires more than a numerical label for each person’s preferences. If each person’s utility scale can be multiplied independently by an arbitrary positive number, the social ranking generated by the sum can change. The representation therefore depends on restrictions that give utility differences an interpersonal meaning.
Alternative objectives embody different distributive priorities. A maximin rule evaluates an allocation by the well-being of its worst-off member:
Leximin first compares the welfare of the worst-off individuals; if those levels tie, it compares the second-worst-off, and continues in order. Prioritarian approaches give greater social weight to benefits received at lower levels of well-being, often using an increasing, concave transformation of comparable welfare. These principles can disagree sharply with maximizing total welfare.
Pareto efficiency provides a less demanding criterion. An allocation is Pareto efficient if no feasible change makes someone better off without making someone else worse off. It leaves many distributional questions unresolved: a highly unequal allocation may be efficient. Similarly, unanimous preference for one state over another does not specify how to compare states involving gains for some and losses for others. The distinction between efficiency and distributive evaluation is central to income inequality analysis and public policy.
Sen’s work emphasized that the informational basis of evaluation is itself a major choice. Utility can reflect adaptation to deprivation, while income does not fully capture differences in what people can achieve. The capability approach therefore considers substantive opportunities to achieve valued ways of living. Social choice can also use incomplete rankings when available information or accepted principles do not settle every comparison. An incomplete evaluation may identify clear improvements without claiming that all remaining alternatives are equally good or that every policy must receive a unique position.
Rights, liberties, and aggregation of judgments
Collective choice is not always understood as unrestricted aggregation of preferences. Legal and moral rights may reserve some decisions to individuals or exclude certain outcomes from collective authority. This creates a tension between evaluating outcomes through everyone’s preferences and protecting a person’s control over particular matters.
Sen’s 1970 liberal paradox formalized one version of this tension. Minimal liberalism requires that at least two individuals each be decisive over some pair of alternatives: their preferences determine the social comparison for those designated pairs. On an unrestricted preference domain, this requirement can conflict with the Pareto principle and a coherent social ordering. The result does not show that every conception of liberty is inconsistent with social welfare. It shows that a specific account of individual decisiveness becomes problematic when combined with unrestricted concern about other people’s conduct and the other conditions.
The mechanism involves preferences that extend beyond a person’s own actions. Someone may dislike undertaking an activity but care even more about whether another person undertakes it. Assigning each person authority over one comparison can then combine with unanimous preferences over a third comparison to generate a cycle. Responses include restricting which preferences count, representing rights as constraints on feasible decisions, or analyzing the strategic exercise of rights rather than inserting them directly into a social ranking.
A related branch, judgment aggregation, studies the combination of beliefs or judgments about logically connected propositions. Suppose a group considers propositions , , and . One member accepts all three; a second accepts but rejects and the conjunction; a third accepts but rejects and the conjunction. Every individual holds a logically consistent set of judgments. Majority voting nevertheless accepts , accepts , and rejects .
This pattern is known as the discursive dilemma. In legal settings, related examples arise when judges agree on a majority of the premises required for liability but a different majority rejects liability itself. Premise-based aggregation and conclusion-based aggregation can therefore yield different outcomes. Neither procedure can be evaluated without considering which propositions the institution must decide and how its reasons should relate to its conclusions.
Judgment aggregation extends social choice from preferences to logical structures. Collective consistency can be retained by privileging certain propositions, restricting admissible judgments, allowing incomplete collective judgments, or abandoning independent proposition-by-proposition aggregation. The relevant trade-offs differ from electoral ones because relationships of logical implication, rather than only comparisons among candidates, constrain the outcome.
Fair allocation and decisions with multiple winners
Many collective problems involve dividing resources or selecting several items rather than electing a single winner. Fair division studies allocations of divisible goods, indivisible objects, chores, and mixed resources. Its central criteria include efficiency, equal treatment, proportionality, and envy-freeness. An allocation is envy-free if no individual prefers another individual’s bundle to their own.
For a homogeneous divisible resource with additive valuations, proportionality gives each of participants at least of the total value according to that participant’s own evaluation. The familiar cut-and-choose procedure for two people illustrates the distinction between physical equality and subjective fairness. One person divides the resource into two equally valued portions according to that person’s valuation; the other chooses. Under the standard model, each obtains at least half the total value by their own assessment. The same simple procedure does not directly extend to arbitrary numbers of participants or to every additional fairness requirement.
Indivisible goods create sharper constraints. If two people both want a single object, allocating it to either person cannot be envy-free without additional devices such as payments or lotteries. Weaker notions are therefore important. Envy-freeness up to one good, or EF1, requires that an individual’s envy of another bundle disappear after removal of some good from the envied bundle. For additive, nonnegative valuations over finitely many goods, EF1 allocations exist. This does not make exact envy-freeness unnecessary; it identifies a feasible relaxation when indivisibility prevents the stronger condition.
Selecting a representative committee introduces a different form of collective fairness. Choosing the individually highest-scoring candidates may produce a homogeneous committee even when substantial groups support other candidates. Approval-based committee voting studies representation requirements using ballots on which voters approve sets of candidates. Proportional approval voting, for example, scores a voter’s approved winners using the harmonic sequence . The diminishing marginal contribution of additional approved winners rewards representation of groups that would otherwise remain unrepresented.
Participatory budgeting adds unequal project costs and a budget constraint. A city might ask residents to choose among playgrounds, transport improvements, and neighborhood facilities. Unlike a fixed-size committee, feasible outcomes must satisfy
where is project cost and is the budget. Maximizing total approvals may favor inexpensive projects or densely supported neighborhoods, while proportional rules seek other balances among groups. Complementarities between projects, geographic effects, and differing benefit intensities can require richer models than approval counts alone.
Computation, applications, and limits of formal models
Computational social choice examines the computational demands of collective decisions and applies algorithmic methods to their design. It asks how winners can be computed, how much preference information is needed, whether manipulation is tractable, and how to represent preferences over very large outcome spaces. These questions connect axiomatic analysis with computer science.
Some voting rules have straightforward counting procedures. Plurality and Borda scores can be computed by scanning ballots and adding totals. Other rules require optimization. The Kemeny rule seeks a ranking that maximizes agreement with voters’ pairwise comparisons, equivalently minimizing the total number of pairwise disagreements. Computing an exact optimal Kemeny ranking is NP-hard in general. This motivates exact methods for manageable instances, approximation algorithms, and analysis under restrictions on candidate numbers or preference structure.
Representing preferences may itself be costly. If a decision selects whether to undertake each of projects, there are possible project combinations before feasibility restrictions are imposed. Asking every participant to rank every combination quickly becomes impractical. Compact preference languages, limited queries, and iterative elicitation reduce reporting demands, but their use can exclude interactions among alternatives. A resident’s valuation of a transport interchange, for example, may depend on whether a connecting route is also funded.
Applications include combining rankings from search systems, selecting panels, organizing shared resources, and aggregating assessments produced by humans or software. In artificial intelligence, social choice concepts help analyze whose evaluations count, how conflicting objectives are combined, and what consistency or fairness guarantees an aggregation procedure offers. Rankings of products or models can display the same tensions between scoring, pairwise comparison, and strategic reporting as rankings of electoral candidates. However, mathematically similar inputs do not make the institutional stakes or appropriate safeguards identical.
Formal models have identifiable boundaries. Preferences may change through discussion, participants may possess unequal information, and a ballot may record a strategic choice rather than a stable underlying ranking. Political institutions also distribute authority, determine agendas, and enforce rights; none of these functions is exhausted by a preference-aggregation formula. Empirical work can investigate how often cycles or manipulation opportunities arise in specified environments, whereas an impossibility theorem identifies what cannot be guaranteed across its entire domain.
The design of a real procedure consequently requires explicit choices about the electorate, available alternatives, admissible information, and permitted outcomes. A rule suitable for choosing one technical standard may be unsuitable for allocating a public budget or representing a diverse population. Computational feasibility, incentive properties, and axiomatic fairness remain distinct dimensions: improving one need not improve the others.