betaFeedback

WorksEconomics, Econometrics and Finance

Classification Aggregation with Multiple Memberships

Abstract

People face the problem of classifying objects into categories while aggregating their opinions. We axiomatically study such a classification aggregation problem with multiple memberships, meaning that each individual assigns one category to every object, whereas the resulting classification may assign more than one category. Under the natural settings and axioms (i.e., Independence and Supportiveness), each of our new axioms – Responsibility, Split Stability and Rejection Stability – characterises what we call the junta rules. The rules assign a category to an object exactly when some member of a fixed coalition does so. The coalitions that can secure an assignment form a grill, a topological structure implied by an ultrafilter. Hence, the study parallels the literature on Arrow's theorem characterised by the decisive coalition structures. An infinite population setting is also investigated to clarify the differences among the axioms.

Fioravanti acknowledges the financial support from the French National Research Agency within the project ANR-24-EXMA-0001 PEPR MathsVivES CONDORCET. Kurata acknowledges support from KAKENHI (No. 25KJ2146) and Global Research Center (Waseda University). The authors’ email addresses are: Federico Fioravanti, [email protected]; Leo Kurata, [email protected]; Vasilis Stamatis, [email protected]. During the preparation of this work, the authors used Claude (Anthropic) and ChatGPT (OpenAI) in order to assist with language editing and to identify possible algebraic or expository issues. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Introduction

We study the setting where a set of individuals assigns each object to exactly one of the categories, using every category at least once, to obtain a collective classification. This problem, called classification aggregation, has been considered by Maniquet and Mongin [38], who show that an aggregator satisfying the properties of Unanimity, which requires society to respect agreement on an object’s category, and Independence, which requires the collective assignment for that object to depend only on how voters classify it, must adopt the classification of one fixed voter (the dictator) at every profile. In the preference aggregation setting (Arrow, [2]), where a finite society must turn any profile of rankings over at least three alternatives into a complete and transitive ranking of its own, unanimity and independence of irrelevant alternatives also force the existence of a dictator. But non-dictatorial rules become possible when some assumptions are relaxed, such as the social ranking being quasi-transitive (Gibbard, [31]) or being incomplete (Weymark, [49]). In the same spirit, we study classification aggregation correspondences (CACs).1 Each voter still assigns one category to each object, but society may recognize several categories as final memberships of the same object, provided every object receives at least one category and every category remains represented. Which non-dictatorial rules become possible under this relaxation?

Collective classification with multiple memberships can arise in a variety of settings. For example, consider a participatory-budgeting platform that asks residents to classify project proposals. Each resident chooses one primary category per proposal from a list that includes Arts, Biodiversity, and Climate.2 Classifications help determine how public projects are compared and presented to citizens, and may also affect which sources of funding they can access. An urban-greening project may aim both to restore habitats and to support climate adaptation. One resident may therefore identify Biodiversity as its primary category, while another identifies Climate. Listing the project under both headings recognizes both contributions. If eligibility for funding follows the assigned categories, the project could still be considered under Climate when the Biodiversity budget is exhausted. It would retain both memberships even if only the Climate budget funded it.

In another application, a platform combines the predictions of several trained classifiers. Each classifier assigns one thematic label per project description, while the combined classification may retain several labels for the same project, a form of output also studied in multi-label learning (e.g., Gibaja and Ventura, [30]).

To isolate the effect of allowing multiple memberships, we admit every profile of individual classifications in the domain of Maniquet and Mongin [38].3 The requirement that every category remain represented collectively also has a practical interpretation. Some participatory-budgeting schemes allocate separate budgets to thematic categories.4 When these budgets cannot be transferred between categories, leaving a category empty would leave its budget with no eligible proposal; this motivates requiring the collective classification to represent every category.

We also retain Independence, while strengthening Unanimity to Supportiveness, which requires every category collectively assigned to an object to have been proposed for that object by at least one individual. Supportiveness implies Unanimity in our setting and is itself a consequence of Independence and Unanimity in Maniquet and Mongin’s original setting (Maniquet and Mongin, [38], Lemma 1). In this finite-population benchmark with at least three categories, Independence and Supportiveness therefore describe exactly the same class as Independence and Unanimity, namely the dictatorial rules (Proposition 1). In this sense, our approach preserves their individual domain and requirement that every category be represented, while changing the admissible form of the collective outcome. The requirement that no category be left empty concerns the collective classification as a whole, while Independence requires considering each object separately. On the other hand, once multiple memberships are allowed, even under Independence, Unanimity no longer delivers Supportiveness when there are three or more categories. The clearest example is the classification counterpart of an oligarchy. It assigns an object to the category on which the members of a coalition agree and, when they disagree, to every category. This rule satisfies both Independence and Unanimity, and it is non-dictatorial once the coalition has two members (Proposition 2). Under the membership reading, disagreement within the coalition can lead society to assert that the object belongs to categories that nobody, inside or outside the coalition, has proposed. When the designated voters disagree between Biodiversity and Climate, for example, the urban-greening proposal would also be listed under Arts. Supportiveness excludes such a deicision. In § 4.2, we further justify this strengthening of Unanimity from both the interpretational and the technical perspectives.

Accepting every proposed category gives an independent and supportive rule that is non-dictatorial when there are at least two individuals. We ask how decision power over the acceptance of assignments can be distributed while allowing some supported assignments to be rejected. In particular, can a rule treat individuals equally and still reject some proposed memberships if each accepted membership must have a supporter whose report suffices on its own? In the participatory-budgeting example, listing the urban-greening proposal under Climate expresses the panel’s judgement that the project contributes to climate adaptation. The platform may give designated panel members the authority to secure the inclusion of a thematic category. Our Responsibility axiom requires that, for each accepted assignment, some supporting voter’s complete classification be sufficient to sustain it, whatever the other voters report. This requirement is stronger than the mere existence of a supporter. Because thematic categories can overlap, the collective classification can retain both memberships when voters identify different primary categories. Different voters can therefore sustain different memberships, including memberships of the same proposal, and share decision power over the collective classification. For aggregation of learned classifications, the same requirement gives designated classifiers the authority to secure the inclusion of a label.

Our main results characterise the distribution of decision power over collective memberships. Under Independence and Supportiveness, Responsibility and Rejection Stability each characterise the junta rules on finite populations, while Split Stability characterises the subclass in which the junta depends only on the object (Theorems 1 to 3).5 A junta rule fixes, for each object and category, a coalition – which we call a junta – and assigns the object to the category exactly when at least one member of the junta does. To ensure that every object receive a category and every category be used, the juntas for each object must have a common member, as must the juntas for each category (Proposition 3). Under Category Neutrality, the junta depends only on the object; further adding Object Neutrality gives a single junta for all objects and categories. The three axioms reach these characterisations from different directions. Responsibility asks that every assigned category be secured by the complete report of one of its supporters, in the spirit of the expertise conditions of Fioravanti [24]; Split Stability asks that, when support for a category is divided between it and a category previously unused at that object, at least one of the two is accepted exactly when the original was, while the object’s other memberships remain unchanged; and Rejection Stability asks that combining supporters’ unchanged complete reports from two rejecting profiles should not secure the rejected assignment. Every member of a junta can include a category on her own, and different members can include different categories at the same object; this is possible only because multiple memberships are allowed. Note that when outcomes are required to be singletons, every junta shrinks to the same individual and the dictatorship of Maniquet and Mongin [38] is recovered.

We also examine whether individual responsibility and equal treatment of individuals allow some supported assignments to be rejected. Under Independence, Supportiveness and Responsibility, adding Anonymity singles out the full-population junta JNJ^{N} (Corollary 3). Thus, under the maintained axioms, equal treatment and individual responsibility require the acceptance of every supported assignment. The rule is also neutral for all sizes of the problem, so that multiple memberships escape the non-existence of anonymous and neutral rules found by Cailloux et al. [7].

It turns out that the coalitional structure behind these results is a grill [14]. Grills are related to the filters, familiar from aggregation theory. The decisive coalitions in the theorem of Maniquet and Mongin form an ultrafilter, and an ultrafilter is exactly a family that is both a filter and a grill (Fact 3). Much of the literature on Arrow’s theorem weakens collective rationality while keeping the Pareto principle and (binary) independence, and it finds that the decisive coalitions retain the filter property. Quasi-transitive social preferences lead to oligarchies [31], transitive but incomplete social preferences lead to oligarchies [49], and acyclic social preferences lead to prefilters [5, 3]. Under our axioms, relaxing single-valuedness in classification aggregation corresponds to keeping the other half of the ultrafilter. Grills are the duals of filters (Fact 4), and the coalitions that win under the junta rule with junta, say MM, are exactly the blocking coalitions of the oligarchy with oligarchs MM. In the Arrovian literature, each member of a decisive group can block a strict social preference. In our setting, each member of a junta can secure the inclusion of a category. A rule that lets each member of a group block every category she does not propose leaves divided objects with no category at all, violating the requirement that every object receive a category. In this sense, the passage from the theorem of Maniquet and Mongin to junta rules parallels the passage from Arrow’s theorem to oligarchies, with the roles of filters and grills exchanged. Refer to § 5.1 and Table 2 for technical details.

Social outcomeAxiomsFinite NN: Structure; rulesFinite NN: SourceInfinite NN: Structure; rulesInfinite NN: Source
Preference aggregation
Complete, transitiveWP, IIAUltrafilter (= Filter ∩\cap Grill); dictatorship[35]Ultrafilter, possibly free; non-dictatorial rules possible under AC[28], [35]
Complete, quasi-transitiveWP, IIAPrincipal filter; oligarchy[31]Filter, possibly nonprincipal; non-oligarchic rules possible[32]
Transitive, possibly incompleteWP, IIAFilter (decisive)/ Grill (quasi-decisive); oligarchy[49], [8]Filter (decisive)/ grill (quasi-decisive); not necessarily principal[4], [8]
Classification aggregation
Single-valued & surjectiveI, UnanimityUltrafilter (= Filter ∩\cap Grill); dictatorship[38]Ultrafilter, possibly free; non-dictatorial rules possible under ACCorollary 4
Set-valued & coveringI, Supportiveness, SSPrincipal grill; J(Mx)J(M_x)Theorem 2Grill; L(Mx)L(M_x)Theorem 4
dittoI, Supportiveness, Rej. Stab., CNPrincipal grill; J(Mx)J(M_x)Theorem 3Grill; L(Mx)L(M_x)Theorem 4
dittoI, Supportiveness, Resp., CNPrincipal grill; J(Mx)J(M_x)Corollary 1Principal grill; J(Mx)J(M_x)Corollary 5

Table 2. Coalitional structures and rules in preference aggregation and in classification aggregation.

Under Independence, Supportiveness and Category Neutrality, our three new axioms are equivalent on finite populations when m>p≥3m > p \ge3. By contrast, on infinite populations, Responsibility is strictly stronger than the two stability axioms under these conditions. Arrow’s theorem fails in infinite societies, because non-dictatorial rules are obtained whenever an ultrafilter is free [28, 35], and the same holds for the theorem of Maniquet and Mongin (Corollary 4). With multiple memberships, Split Stability, or Rejection Stability together with Category Neutrality, characterises the limiting junta rules, a class that extends the object-dependent junta rules and whose winning coalitions form grills that need not be principal (Theorem 4). The rule that assigns an object to every category with infinitely many supporters belongs to this class and is defined explicitly. Responsibility yields junta rules on every population by directly identifying the individuals whose classifications can sustain an assignment (Corollary 5).

The axiomatic study of classification began with partitions, in which the categories carry no labels and their number is not fixed in advance. Mirkin [42] was the first to analyse the aggregation of equivalence relations axiomatically, and Fishburn and Rubinstein [29] show that independence and unanimity leave exactly the conjunctive rules, under which two objects are equivalent in the aggregate partition if and only if they are equivalent for every member of a fixed nonempty coalition (see also Dimitrov et al., [19]). These rules are oligarchies, and they are available because the intersection of equivalence relations is again an equivalence relation: agreement within a coalition always defines an admissible outcome. With designated categories this is no longer the case. Maniquet and Mongin [38] extend to three or more categories a result that Kasher and Rubinstein [34] obtained for two, showing that Independence and Unanimity force dictatorship when individual and collective classifications must leave no category empty, and Dokow and Holzman [21] obtain the counterpart for classifications that place at most one object in each category. Subsequent work retains single-valued collective classifications and varies the remaining requirements. Cailloux et al. [6] weaken Unanimity to citizen sovereignty; Craven [15] shows that domain restrictions permitting majority-based aggregation have very limited scope in this setting, and identifies a hybrid structure in which such aggregation is possible exactly when each object is eligible for at most two categories; Fioravanti [24] examines forms of individual authority over objects or categories; Cailloux et al. [7] drop Independence and study anonymity and neutrality; and Craven [16] considers ranked categories. Filmus [23] characterises all independent rules whose outcome is always surjective, without further assumptions on the objectwise aggregation functions, and derives from it a quantitative version of the dictatorship conclusion. Hervouin [33] also considers classification aggregation correspondences. Relative to this literature, we keep the individual domain of Maniquet and Mongin [38], Independence and the requirement that every category be used, and change only the form of the collective outcome.

Group identification Kasher and Rubinstein [34] is the case in which the objects to be classified are the individuals themselves. Because a group may be left empty there, rules that give individuals authority over their own membership remain available, and the literature characterises the liberal rule together with the consent rules, which balance self-determination against the opinions of others [43, 46, 41]. Cho and Ju [12] determine memberships in three or more groups simultaneously and show that independence of irrelevant opinions, together with symmetry and non-degeneracy, leads to the liberal rule. In these models possibility comes from allowing categories to remain empty; we keep the requirement that every category be used and obtain possibility from multiple memberships instead. Several papers associate an object with several categories through degrees of membership. Cho and Park [13] allow fractional memberships in group identification and characterise weighted-average and fractional consent rules, and fuzzy versions of classification and group identification are studied by Alcantud et al. [1], Fioravanti and Tohmé [27] and Fioravanti [25], the last of which obtains weighted arithmetic means. Our memberships are not graded. An object belongs fully to each category assigned to it, and assigning it to one category does not weaken its membership in another; accordingly, the rules we obtain are coalitional rather than averaging.

Reading each statement “xx belongs to rr” as a binary issue turns a CAC into a binary aggregator whose admissible collective evaluations form a strictly larger set than the admissible individual ones: an individual assigns each object exactly one category, whereas society assigns it at least one. Our problem therefore belongs to the family in which society is held to a weaker consistency requirement than individuals. In judgment aggregation, allowing incomplete collective judgments while retaining deductive closure yields oligarchies [18], and Dokow and Holzman [20] show that the conditions on the domain that lead to dictatorship without abstentions lead to oligarchy once abstentions are allowed, with Gibbard’s oligarchy theorem among the corollaries. In these models society may take a position on fewer issues than individuals do, so agreement within a coalition remains an admissible outcome. In ours, society may accept more membership statements than any individual does, disagreement is resolved by inclusion rather than abstention, and oligarchies give way to juntas. Graph aggregation Endriss and Grandi [22] provides a close formal analogue. Its groundedness axiom requires every collective edge to belong to some individual graph, as Supportiveness does for memberships, and Endriss and Grandi [22] show that, for graph properties that are contagious and implicative, unanimous, grounded and independent rules that preserve the property are oligarchic. Applied to each membership statement, the quota rules of Dietrich and List [17] accept a category when it has enough supporters; a quota of one for every pair gives JNJ^N, and pair-specific quotas give procedures such as F4F^4 in Example 4. Finally, when individual and collective evaluations are drawn from the same set, Dokow and Holzman [21] and Szegedy and Xu [47] study which domains of non-binary evaluations force supportive or unanimous aggregation to be dictatorial. The setting of Maniquet and Mongin [38] is such a domain, with objects as issues and categories as positions, and CACs fall outside this framework precisely because the collective outcome need not belong to the individual domain.

As recalled earlier, the ultrafilter approach identifies the decisive coalitions of an Arrowian rule with an ultrafilter [35, 32], and weakenings of collective rationality correspond to weakenings of this structure, from filters and oligarchies to pref­­ilters (see also Bossert and Suzumura [4]). Cato et al. [11] develop this approach in an abstract framework in which an aggregation rule may remain silent. Under a stability condition, the decisive coalitions form a filter, and when the set of signals is finite they characterise the rules that approve a signal exactly when some coalition in a given filter or ultrafilter agrees on it. These rules are agreement-based. Junta rules are inclusion-based, and their winning coalitions form grills [14], the duals of filters. To our knowledge, grills have not been used explicitly as the decisive structure of aggregation rules, although they appear under another name: Cato [8] shows that, for transitive but possibly incomplete social preferences, the quasi-decisive coalitions form a quasi-ultrafilter, which by Fact 2 is exactly a grill. In his setting these are the coalitions that can prevent society from ranking an alternative at least as high as one that all their members prefer to it; here they are the coalitions that can secure a membership. Fishburn [28] and Kirman and Sondermann [35] show that Arrow’s conditions can be met by non-dictatorial rules on infinite populations, through free ultrafilters that cannot be explicitly constructed. Mihara [40] shows that in a countably infinite society every rule satisfying Arrow’s conditions that is pairwise computable is dictatorial, and Mihara [39] shows that in infinite societies anonymity is inconsistent with Arrow’s unanimity and independence conditions on sufficiently rich domains. Torres [48] introduces limiting dictatorial rules, of which our limiting junta rules are the counterparts, and Cato [9, 10] study further aspects of preference aggregation with infinite populations. In group identification, Fioravanti and Tohmé [26] show that, with infinitely many voters, the characterisation of the liberal rule survives whereas that of the oligarchic rule does not. A similar divergence appears here. The axiom that identifies individual supporters, Responsibility, characterises junta rules on every population, whereas the axioms that only relate coalitions, Split Stability and Rejection Stability, admit non-principal grills. The resulting rules include the Fréchet rule, which, unlike the rules built on free ultrafilters, is defined without the axiom of choice.

Paper organisation

The paper is organized as follows. § 3 introduces the model and the coalitional families we use. § 4 presents the basic axioms and the benchmark. § 5 discusses the two translations of oligarchy and introduces junta rules. § 6 contains the three characterisations, examines the role of Supportiveness, and concludes by relating Anonymity to responsibility. § 8 turns to infinite populations. § 9 concludes. Longer proofs are collected in Section A, and Section B shows that the axioms in the main characterisations are independent.

Preliminaries

Notation

Let NN be a set of individuals with at least two members. Unless stated otherwise, N={1,…,n}N = \{1,\ldots,n\} is finite; in § 8 and wherever we say so explicitly, NN may be infinite, for instance N=N={1,2,…}N = \mathbb{N} = \{1,2,\ldots\}. Let X={x1,…,xm}X = \{x_1,\ldots,x_m\} be a finite set of mm objects and PP a finite set of pp categories, with m≥p≥2m \ge p \ge2.

An individual classification is a surjective mapping c:X→Pc:X \to P. We denote the set of individual classifications by C={c:X→P∣c is surjective}\mathcal{C} = \{c:X \to P \mid c\text{ is surjective}\}. A social classification is a mapping c:X→2P∖{∅}c:X \to2^P \setminus\{\varnothing\} satisfying ⋃x∈Xc(x)=P\bigcup_{x\in X}c(x)=P. We denote the set of social classifications by C‾\overline{\mathcal{C}}.

A profile is a vector c=(ci)i∈N∈CNc=(c_i)_{i\in N}\in\mathcal{C}^N of individual classifications. A classification aggregation correspondence (CAC) is a mapping F:CN→C‾F:\mathcal{C}^N\to\overline{\mathcal{C}} that assigns a social classification to every profile. A CAC is resolute if F(c)(x)F(c)(x) is a singleton for all cc and xx; resolute CACs are the classification aggregation functions studied by Maniquet and Mongin [38]. The rule they specify is a dictatorial rule, defined by the existence of d∈Nd\in N such that F(c)(x)={cd(x)}F(c)(x)=\{c_d(x)\} for all c∈CNc\in\mathcal{C}^N and x∈Xx\in X.

Given a profile cc and an object x∈Xx\in X, we write

Sr(x,c)={i∈N∣ci(x)=r}S_r(x,c)=\{i\in N\mid c_i(x)=r\}

for the support set of category rr at object xx. Given c∈CNc\in\mathcal{C}^N, we write cx=(ci(x))i∈N∈PNc_x=(c_i(x))_{i\in N}\in P^N for the opinion vector at xx. Given an opinion vector π∈PN\pi\in P^N, Sr(π)={i∈N∣π(i)=r}S_r(\pi)=\{i\in N\mid\pi(i)=r\} denotes the support set of category rr.

Any opinion vector at a fixed object can be obtained from a profile of individual classifications. Note that both lemmas below hold for finite and infinite populations.

Lemma 1. Assume m≥p≥2m\ge p\ge2. For every π∈PN\pi\in P^N and every x∈Xx\in X, there exists a profile c∈CNc\in\mathcal{C}^N such that cx=πc_x=\pi.

Proof. Fix π∈PN\pi\in P^N and x∈Xx\in X. For each i∈Ni\in N, set ci(x)=π(i)c_i(x)=\pi(i). It remains to define cic_i on X∖{x}X\setminus\{x\} so that cic_i is surjective. The category π(i)\pi(i) is already attained at xx. Since ∣X∖{x}∣=m−1≥p−1=∣P∖{π(i)}∣|X\setminus\{x\}|=m-1\ge p-1=|P\setminus\{\pi(i)\}|, we can assign every category in P∖{π(i)}P \setminus\{\pi(i)\} to at least one object in X∖{x}X \setminus\{x\} and assign any remaining objects arbitrarily. ■

When m>pm > p, the profile can be chosen so that changing the opinion vector at xx never affects surjectivity.

Lemma 2. Assume m>p≥2m > p \ge2. For every π∈PN\pi\in P^N and every x∈Xx \in X, there exists a profile c∈CNc \in C^N such that cx=πc_x = \pi and, for every i∈Ni \in N, the restriction ci∣X∖{x}c_i|_{X\setminus\{x\}} is surjective.

Proof. Fix π∈PN\pi\in P^N and x∈Xx \in X, and write P={r1,…,rp}P = \{r_1,\ldots,r_p\}. For each i∈Ni \in N, set ci(x)=π(i)c_i(x) = \pi(i). Since ∣X∖{x}∣=m−1≥p|X \setminus\{x\}| = m - 1 \ge p, we may choose pairwise distinct objects y1,…,yp∈X∖{x}y_1,\ldots,y_p \in X \setminus\{x\} and set ci(yj)=rjc_i(y_j) = r_j for j∈{1,…,p}j \in\{1,\ldots,p\}, and ci(y)=π(i)c_i(y) = \pi(i) for every remaining y∈X∖{x}y \in X \setminus\{x\}. Then ci∣X∖{x}c_i|_{X\setminus\{x\}} is surjective, hence so is cic_i. ■

A set-valued outcome of CACs admits two readings: membership or tie. Under the membership reading, r∈F(c)(x)r \in F(c)(x) means that society assigns xx to rr, so that xx may belong to several categories at once. Under the tie reading, F(c)(x)F(c)(x) lists the categories that remain admissible for xx, and a single category is to be selected later. We adopt the membership reading throughout. The covering condition ⋃x∈Xcˉ(x)=P\bigcup_{x\in X}\bar{c}(x)=P in the definition of a social classification is formulated for this reading. It says that every category has at least one member, which is how the surjectivity requirement translates when an object may belong to several categories. Under the tie reading, we would instead require the existence of a surjective selection: a surjective mapping c:X→Pc:X\to P with c(x)∈cˉ(x)c(x)\in\bar{c}(x) for every x∈Xx\in X. Under this reading a single category must eventually be selected for each object, and if this selection respects the two postulates of Maniquet and Mongin [38], their impossibility applies at that final stage. Under the membership reading, the set-valued outcome is the final classification.

Examples of CACs

We introduce four running examples involving five rules at a common profile cc. The examples distinguish whether accepted assignments have supporters, whether an object’s classification depends on reports about other objects, and whether an individual’s report can secure an assignment. They also compare individual decision power with equal treatment of voters. We return to these examples when introducing the axioms and summarise their properties in Table 1.

F1F^1F2F^2F2′F^{2'}F3F^3F4F^4F5F^5
Independence✓✓✓×✓✓
Unanimity✓✓✓✓✓✓
Supportiveness×✓✓✓✓✓
Split Stability×✓×✓×✓6
Responsibility✓✓✓✓××
Rejection Stability×✓✓✓××
Object Neutrality✓✓×✓×✓
Category Neutrality✓✓×✓×✓
Anonymity✓✓××✓✓

Table 1. Axioms satisfied (✓) and violated (×) by the rules in Examples 1 to 5.

A participatory-budgeting platform groups project proposals into thematic categories for comparison. Two voters, N={1,2}N=\{1,2\}, classify four proposals, X={x1,x2,x3,x4}X=\{x_1,x_2,x_3,x_4\}, into three categories, P={a,b,c}P=\{a,b,c\}, which we read as Arts, Biodiversity and Climate, respectively. Each voter assigns every proposal to one category and uses every category, whereas the platform may place a proposal in several categories, for instance a project combining public art and habitat restoration in Arts and in Biodiversity. The list (c(x1),c(x2),c(x3),c(x4))∈PX(c(x_1),c(x_2),c(x_3),c(x_4))\in P^X denotes the classification cc, and a social classification cˉ\bar{c} is written in the same way. In each of the examples below, the profile c=(c1,c2)c=(c_1,c_2) is given by

c1=(a,b,c,a),c_1=(a,b,c,a),
c2=(a,a,b,c).c_2=(a,a,b,c).

The voters agree on x1x_1 and disagree on the other three proposals. Observe that both classifications are surjective.

Example 1. Let F1(c)(x)={c1(x)}F^1(c)(x)=\{c_1(x)\} if c1(x)=c2(x)c_1(x)=c_2(x), and F1(c)(x)=PF^1(c)(x)=P otherwise. At the profile above,

F1(c)=({a},P,P,P).F^1(c)=(\{a\},P,P,P).

Every category is used, so F1(c)F^1(c) is a social classification, and the same holds at every profile. The outcome for a proposal depends only on how the two voters classify that proposal, and a proposal on which they agree is placed exactly in the category they chose. At x2x_2, however, voter 1 proposes Biodiversity and voter 2 Arts, and F1F^1 places the proposal in Climate as well, although neither of them proposed it. In this case, the rule uses every category only because it places each disputed proposal in all of them. ◇

Example 2. Let F2(c)(x)={c1(x),c2(x)}F^{2}(\mathbf{c})(x)=\{c_{1}(x),c_{2}(x)\}, so that each proposal is placed in every category that some voter proposes for it. At the profile above,

F2(c)=({a},{a,b},{b,c},{a,c}).F^{2}(\mathbf{c})=(\{a\},\{a,b\},\{b,c\},\{a,c\}).

Like F1F^{1}, this rule depends only on the assignments for each proposal and respects agreement, but it never places a proposal in a category that nobody proposed. The voters whose assignments count may also vary across proposals and categories. Let F2′F^{2'} coincide with F2F^{2}, except that Biodiversity is assigned to x2x_{2} only when voter 2 proposes it, and Climate is assigned to x4x_{4} only when voter 1 proposes it, for instance because voter 2 has expertise in biodiversity conservation and voter 1 in climate adaptation. At the profile above,

F2′(c)=({a},{a},{b,c},{a}).F^{2'}(\mathbf{c})=(\{a\},\{a\},\{b,c\},\{a\}).

Every category is still used, and F2′F^{2'} produces a social classification at every profile (see also the discussion in Example 4). ⋄

Note that, under the tie reading of set-valued outcomes, no selection from F2′(c)F^{2'}(\mathbf{c}) in Example 2 is surjective, whereas the membership reading is still applicable.

Example 3. Let F3(c)(x)F^{3}(\mathbf{c})(x) consist of c1(x)c_{1}(x) together with c2(x)c_{2}(x) whenever voter 2 uses c2(x)c_{2}(x) for at least one other proposal. Voter 2’s assignment for a proposal thus counts only if she uses that category more than once; at the profile above, this is the case only for Arts. Since F3(c)F^{3}(\mathbf{c}) always contains voter 1’s classification, it is a social classification at every profile, and every category it assigns is proposed by someone. At the profile above,

F3(c)=({a},{a,b},{c},{a}).F^{3}(\mathbf{c})=(\{a\},\{a,b\},\{c\},\{a\}).

The rule also uses every category. ⋄

Lastly, the platform may instead use a quota rule, requiring a specified minimum number of supporting voters for each proposal–category pair. If one supporter suffices for every pair, this procedure coincides with F2F^{2}. With two voters, a quota of 2 requires both of them to propose the category. Raising every quota to 2 gives ({a},∅,∅,∅)(\{a\},\varnothing,\varnothing,\varnothing) at the profile above: three proposals are left without a category, and Biodiversity and Climate are not used.

Example 4. Let F4F^{4} require two supporters for Art at x1x_{1}, for Biodiversity at x2x_{2}, and for Climate at x4x_{4}, and one supporter for every other pair. At the profile above, this rule yields

F4(c)=({a},{a},{b,c},{a}).F^{4}(\mathbf{c})=(\{a\},\{a\},\{b,c\},\{a\}).

This rule produces a social classification at every profile. Its outcome at the profile above coincides with that of F2′F^{2'} in Example 2, but the reasons behind them are different. Under F2′F^{2'}, voter 1’s assignment of Biodiversity to x2x_{2} does not count; under F4F^{4}, it counts but is not enough on its own. If the voters exchange their reports, so that voter 1 submits c2c_{2} and voter 2 submits c1c_{1}, F2′F^{2'} gives ({a},{a,b},{b,c},{a,c})(\{a\},\{a,b\},\{b,c\},\{a,c\}), whereas F4F^{4} gives the same outcome as before, since the quotas count supporters without regard to who they are. Yet neither voter can secure Biodiversity for x2x_{2} by herself. In these cases, all the categories are used because all quotas are reached (which might not always be the case). ⋄

Filters, ultrafilters and grills

We use families of coalitions to describe which groups can secure an assignment. Filters and related families also describe decisive coalitions in preference aggregation ([44, 45]). We recall the definitions used below.

Definition 1. A family W\mathcal{W} of subsets of NN may have the following properties:

(p.1) ∅∉W\varnothing\notin\mathcal{W}.

(p.2) N∈WN \in\mathcal{W}.

(p.3) For all A,B⊆NA, B \subseteq N, if A∈WA \in\mathcal{W} and A⊆BA \subseteq B, then B∈WB \in\mathcal{W}.

(p.4) For all A,B∈WA, B \in\mathcal{W}, A∩B∈WA \cap B \in\mathcal{W}.

(p.5) For all A⊆NA \subseteq N, A∉WA \notin\mathcal{W} implies N∖A∈WN \setminus A \in\mathcal{W}.

(p.6) For all A,B⊆NA, B \subseteq N, A∪B∈WA \cup B \in\mathcal{W} implies A∈WA \in\mathcal{W} or B∈WB \in\mathcal{W}.

A family W\mathcal{W} of subsets of NN is called

  • a filter, denoted by F\mathcal{F}, if it satisfies (p.1)–(p.4);

  • an ultrafilter, denoted by U\mathcal{U}, if it satisfies (p.1)–(p.5); and

  • a grill, denoted by G\mathcal{G}, if it satisfies (p.1)–(p.3) and (p.6).

Property (p.6) is called the grill property. Equivalently, letting I=2N∖G\mathcal{I} = 2^N \setminus\mathcal{G}, the grill property states that I\mathcal{I} is closed under union: if A∉GA \notin\mathcal{G} and B∉GB \notin\mathcal{G}, then A∪B∉GA \cup B \notin\mathcal{G}. By induction, if A1∪⋯∪Ak∈GA_1 \cup\cdots\cup A_k \in\mathcal{G}, then Aj∈GA_j \in\mathcal{G} for some jj.

A filter is principal if there is a nonempty coalition S∈FS \in\mathcal{F} such that T⊇ST \supseteq S for all T∈FT \in\mathcal{F}. In that case F=FS\mathcal{F} = \mathcal{F}_S, where FS={T⊆N∣T⊇S}\mathcal{F}_S = \{T \subseteq N \mid T \supseteq S\}. An ultrafilter is principal if and only if it equals F{i}\mathcal{F}_{\{i\}} for some i∈Ni \in N, and it is free otherwise. A grill G\mathcal{G} on NN is principal if G=GM\mathcal{G} = \mathcal{G}_M for some nonempty M⊆NM \subseteq N, where

GM={A⊆N∣A∩M≠∅},\mathcal{G}_M = \{A \subseteq N \mid A \cap M \ne\varnothing\},

and it is free otherwise. Letting MG={i∈N∣{i}∈G}M_{\mathcal{G}} = \{i \in N \mid\{i\} \in\mathcal{G}\}, every coalition meeting MGM_{\mathcal{G}} belongs to G\mathcal{G} by (p.3). Hence, G\mathcal{G} is principal if and only if every member of G\mathcal{G} meets MGM_{\mathcal{G}}, in which case G=GMG\mathcal{G} = \mathcal{G}_{M_{\mathcal{G}}}. When NN is finite, all filters, ultrafilters and grills on NN are principal; for grills, this is the content of the next fact.

Fact 1. Let NN be finite. A family G⊆2N\mathcal{G} \subseteq2^N is a grill on NN if and only if there exists a unique nonempty M⊆NM \subseteq N such that G=GM\mathcal{G} = \mathcal{G}_M.

Proof. Let G\mathcal{G} be a grill and I=2N∖G\mathcal{I} = 2^N \setminus\mathcal{G}. By the contrapositive of (p.3), I\mathcal{I} is downward closed, and by the contrapositive of (p.6), it is closed under union. Since ∅∈I\varnothing\in\mathcal{I} and NN is finite, S=⋃A∈IAS = \bigcup_{A \in\mathcal{I}} A belongs to I\mathcal{I} and I=2S\mathcal{I} = 2^S. Let M=N∖SM = N \setminus S. Then M≠∅M \ne\varnothing because N∈GN \in\mathcal{G}, and A∈GA \in\mathcal{G} if and only if A⊈SA \not\subseteq S, that is, if and only if A∩M≠∅A \cap M \ne\varnothing. Uniqueness holds because M=MGM = M_{\mathcal{G}}. Conversely, GM\mathcal{G}_M satisfies (p.1)–(p.3) when M≠∅M \ne\varnothing, and (p.6) holds because (A∪B)∩M=(A∩M)∪(B∩M)(A \cup B) \cap M = (A \cap M) \cup(B \cap M). ■

We record three further properties of grills.

Fact 2. Let W\mathcal{W} be a family of subsets of NN satisfying (p.1)–(p.3). If W\mathcal{W} satisfies (p.6), then it satisfies (p.5).

Proof. Take any A⊆NA \subseteq N. Since A∪(N∖A)=N∈WA \cup(N \setminus A) = N \in\mathcal{W} by (p.2), property (p.6) gives A∈WA \in\mathcal{W} or N∖A∈WN \setminus A \in\mathcal{W}. ■

Cato [8] introduces the notion of a quasi-ultrafilter, defined as a family satisfying (p.1)–(p.3), (p.5), and (p.6). By Fact 2, condition (p.5) is redundant in this definition, so a quasi-ultrafilter is precisely a grill. FACT 2 also leads to the following characterisation of ultrafilters.

Fact 3. A family of subsets of NN is both a filter and a grill if and only if it is an ultrafilter.

Proof. Let U\mathcal{U} be an ultrafilter. It is a filter by definition, so it remains to verify (p.6). Suppose, to the contrary, that A∪B∈UA \cup B \in\mathcal{U} but A∉UA \notin\mathcal{U} and B∉UB \notin\mathcal{U} for some A,B⊆NA, B \subseteq N. By (p.5), N∖A∈UN \setminus A \in\mathcal{U} and N∖B∈UN \setminus B \in\mathcal{U}, and so (p.4) gives N∖(A∪B)∈UN \setminus(A \cup B) \in\mathcal{U}. Another application of (p.4) yields ∅=(A∪B)∩(N∖(A∪B))∈U\varnothing= (A \cup B) \cap(N \setminus(A \cup B)) \in\mathcal{U}, which contradicts (p.1).

Conversely, let W\mathcal{W} be both a filter and a grill. Then W\mathcal{W} satisfies (p.1)–(p.4), and it satisfies (p.5) by Fact 2. Hence, W\mathcal{W} is an ultrafilter.

For instance, consider ∣N∣≥2|N| \ge2 and W={A⊆N:A≠∅}\mathcal{W} = \{A \subseteq N : A \ne\varnothing\}. This is a grill but not an ultrafilter because {i}∩{j}=∅\{i\} \cap\{j\} = \varnothing for any distinct i,j∈Ni,j \in N.

In fact, grills and filters are dual to each other [14]. For a family G⊆2N\mathcal{G} \subseteq2^N, let G∗={N∖A∣A⊆N, A∉G}\mathcal{G}^* = \{N \setminus A \mid A \subseteq N,\ A \notin\mathcal{G}\} be the family of complements of the sets outside G\mathcal{G}. We call the members of G∗\mathcal{G}^* the blocking coalitions of G\mathcal{G}.

Fact 4. Let G⊆2N\mathcal{G} \subseteq2^N.

  1. G\mathcal{G} is a grill if and only if G∗\mathcal{G}^* is a filter.

  2. If G\mathcal{G} is a grill, then S∈GS \in\mathcal{G} if and only if S∩T≠∅S \cap T \ne\varnothing for every T∈G∗T \in\mathcal{G}^*.

  3. For every nonempty M⊆NM \subseteq N, (GM)∗=FM(\mathcal{G}_M)^* = \mathcal{F}_M and (FM)∗=GM(\mathcal{F}_M)^* = \mathcal{G}_M. In particular, a grill is principal if and only if its family of blocking coalitions is a principal filter.

Proof. (i) The map A↦N∖AA \mapsto N \setminus A is a bijection from 2N∖G2^N \setminus\mathcal{G} onto G∗\mathcal{G}^*. Under it, (p.1) for G\mathcal{G} corresponds to (p.2) for G∗\mathcal{G}^*, (p.2) for G\mathcal{G} to (p.1) for G∗\mathcal{G}^*, and upward closure of G\mathcal{G} to upward closure of G∗\mathcal{G}^*. Finally, the grill property of G\mathcal{G} says that A,B∉GA,B \notin\mathcal{G} implies A∪B∉GA \cup B \notin\mathcal{G}, which is equivalent to (N∖A)∩(N∖B)∈G∗(N \setminus A) \cap(N \setminus B) \in\mathcal{G}^* whenever N∖A,N∖B∈G∗N \setminus A,N \setminus B \in\mathcal{G}^*, that is, to (p.4) for G∗\mathcal{G}^*.

(ii) If S∈GS \in\mathcal{G} and S∩T=∅S \cap T = \varnothing for some T∈G∗T \in\mathcal{G}^*, then S⊆N∖T∉GS \subseteq N \setminus T \notin\mathcal{G}, contradicting upward closure. If S∉GS \notin\mathcal{G}, then T=N∖S∈G∗T = N \setminus S \in\mathcal{G}^* and S∩T=∅S \cap T = \varnothing.

(iii) A∉GMA \notin\mathcal{G}_M if and only if A∩M=∅A \cap M = \varnothing, that is, if and only if M⊆N∖AM \subseteq N \setminus A. Likewise, B∈(FM)∗B \in(\mathcal{F}_M)^* if and only if M⊈N∖BM \nsubseteq N \setminus B if and only if B∩M≠∅B \cap M \ne\varnothing.

Basic axioms and the Maniquet–Mongin benchmark

Independence and Unanimity

The first axiom states that the social classification of an object depends only on how the individuals classify that object.

Axiom (Independence). A CAC FF satisfies Independence if, for all x∈Xx \in X and all c,c′∈CN\mathbf{c},\mathbf{c}' \in\mathcal{C}^N, ci(x)=ci′(x)c_i(x) = c'_i(x) for every i∈Ni \in N implies F(c)(x)=F(c′)(x)F(\mathbf{c})(x) = F(\mathbf{c}')(x).

Under Independence, FF is determined by the family of mappings (fx)x∈X(f_x)_{x \in X} with fx:PN→2P∖{∅}f_x : P^N \to2^P \setminus\{\varnothing\} and fx(cx)=F(c)(x)f_x(\mathbf{c}_x) = F(\mathbf{c})(x); by Lemma 1, every fxf_x is defined on all of PNP^N.

Maniquet and Mongin [38] use the following requirement together with Independence.

Axiom (Unanimity). A CAC FF satisfies Unanimity if, for all x∈Xx \in X and all c∈CN\mathbf{c} \in\mathcal{C}^N such that ci(x)=cj(x)c_i(x) = c_j(x) for every i,j∈Ni,j \in N, it is the case that F(c)(x)={ci(x)}F(\mathbf{c})(x) = \{c_i(x)\}.

These are the two requirements in Maniquet and Mongin [38]. Allowing multiple memberships makes it necessary to examine whether Unanimity alone provides an adequate connection between individual reports and collectively accepted assignments.

Supportiveness beyond unanimity

Unanimity constrains the outcome when all individuals agree, but it does not in general prevent an unsupported assignment when they disagree. Because our setting allows multimembership, we propose the use of Supportiveness, which strengthens Unanimity. Supportiveness requires every socially assigned category to have at least one individual supporter.

Axiom (Supportiveness). A CAC FF satisfies Supportiveness if, for all x∈Xx \in X and all c∈CN\mathbf{c} \in\mathcal{C}^N,

F(c)(x)⊆{ci(x)∣i∈N}.F(\mathbf{c})(x) \subseteq\{c_i(x) \mid i \in N\}.

Supportiveness can be read as the unanimity principle applied to each membership statement “xx belongs to rr”. If every individual rejects the statement, so does society. If every individual accepts it, the outcome at xx is a nonempty subset of {r}\{r\}, so society accepts it as well. The requirement that society reject unsupported memberships has several precedents. In the single-valued setting, Maniquet and Mongin [38] derive it from their axioms, and it is part (b) of the unary unanimity condition of Craven [15]. It is the supportiveness condition of Dokow and Holzman [21] for non-binary evaluations, it corresponds to the zero unanimity of Fioravanti [25] for fuzzy classification aggregation, the groundedness axiom of Endriss and Grandi [22] for graph aggregation, and the Pareto★ principle of Kurata and Nakamura [36, 37] for Harsanyian aggregation of incomplete preferences.

Supportiveness implies Unanimity in our framework. For let FF be a CAC and suppose that every individual assigns xx to rr; by Supportiveness, F(c)(x)⊆{r}F(\mathbf{c})(x) \subseteq\{r\}, and F(c)(x)≠∅F(\mathbf{c})(x) \ne\varnothing follows from F(c)∈C‾F(\mathbf{c}) \in\overline{\mathcal{C}}, meaning F(c)(x)={r}F(\mathbf{c})(x) = \{r\}. The converse holds only when p=2p = 2 and fails when p≥3p \ge3.

Lemma 3. Every CAC FF that satisfies Supportiveness satisfies Unanimity.

Among the rules of Examples 1 to 4, only F3F^3 violates Independence: at x2x_2 and x4x_4 its outcome changes although nobody changes her classification of these proposals. Only F1F^1 violates Supportiveness: at x2x_2 it adds Climate, which neither voter proposes.

Sections 4 and 5.4 show why Unanimity cannot replace Supportiveness in the Responsibility characterisation when p≥3p \ge3. Although the two axioms are interchangeable in some of our results, Unanimity alone is insufficient under the membership reading. We adopt Supportiveness throughout so that every accepted assignment has an individual supporter. The two axioms differ when individuals disagree. Suppose that some voters classify a proposal as Arts and the others as Biodiversity. Listing the proposal under both categories is a reasonable way of reflecting this disagreement, since each category records the judgement of some members and can be defended by pointing to them. Listing it under Climate as well is different in kind. Nobody proposed Climate for the proposal, so no voter can be asked to justify that assignment, and a resident comparing proposals in Climate would find one that no voter placed there. Unanimity is silent about this profile, since the members do not agree, and so it does not prevent Climate from being added. Supportiveness excludes Climate and allows Arts, Biodiversity or both to be accepted. The rule F1F^1 of Example 1 behaves in exactly this way at x2x_2: it satisfies Unanimity, yet it lists the proposal under Climate, whereas F2F^2 lists it under Arts and Biodiversity only.

For resolute rules, the distinction does not arise, because there Unanimity, Independence and the requirement that every category be used already imply Supportiveness.

Proposition 1. Let m≥p≥2m \ge p \ge2 and m≥3m \ge3. For a resolute CAC FF, the following statements are equivalent.

(i) FF satisfies Independence and Unanimity.

(ii) FF satisfies Independence and Supportiveness.

(iii) FF is dictatorial.

Proof. A dictatorial rule satisfies Independence and Supportiveness, so (iii) implies (ii). By Lemma 3, (ii) implies (i). That (i) implies (iii) is the theorem of Maniquet and Mongin [38] when p≥3p \ge3 and Corollary 2 of Cailloux et al. [6] when p=2p = 2.

Proposition 1 shows that replacing Unanimity with Supportiveness preserves the benchmark characterisation of Maniquet and Mongin [38]. In their setting it is already implied by Independence, Unanimity and the requirement that the social classification be surjective. In our setting the requirement on the social classification is weaker, because a category can be used by assigning objects to several categories at once, and this derivation of Supportiveness is no longer available. Independence and Unanimity therefore do not suffice to exclude unsupported memberships.

Neutrality

We also consider symmetry requirements on objects and categories. The next axiom requires that the names of the objects be irrelevant for the aggregation.

Axiom (Object Neutrality). A CAC FF satisfies Object Neutrality if, for all c∈CN\boldsymbol{c} \in\mathcal{C}^{N}, all permutations μ:X→X\mu: X \to X and all x∈Xx \in X, F(c∘μ)(x)=F(c)(μ(x))F(\boldsymbol{c} \circ\mu)(x) = F(\boldsymbol{c})(\mu(x)), where c∘μ=(ci∘μ)i∈N\boldsymbol{c} \circ\mu= (c_i \circ\mu)_{i \in N}.

If c∈CN\boldsymbol{c} \in\mathcal{C}^{N} and μ\mu is a permutation of XX, then c∘μ∈CN\boldsymbol{c} \circ\mu\in\mathcal{C}^{N}, since the composition of a surjection with a bijection is a surjection. Category Neutrality requires that the aggregation rule treat all categories symmetrically.

Axiom (Category Neutrality). A CAC FF satisfies Category Neutrality if, for every permutation σ\sigma of PP and every c∈CN\boldsymbol{c} \in\mathcal{C}^{N}, F(σ∘c)(x)=σ(F(c)(x))F(\sigma\circ\boldsymbol{c})(x) = \sigma(F(\boldsymbol{c})(x)) for all x∈Xx \in X, where σ∘c=(σ∘ci)i∈N\sigma\circ\boldsymbol{c} = (\sigma\circ c_i)_{i \in N} and σ(A)={σ(r)∣r∈A}\sigma(A) = \{\sigma(r) \mid r \in A\} for A⊆PA \subseteq P.

Notice that in the running examples, all rules satisfy both neutrality axioms except for F2F^{2} and F4F^{4}: the former makes whose assignment counts depend on the proposal and category, and the latter makes the required quota depend on them.

Oligarchy and junta rules

Translating oligarchy into a CAC

In preference aggregation, an oligarchy is a coalition MM whose unanimous strict preference determines the social strict preference, while opposing strict preferences among its members result in social indifference (Gibbard, [31]). There are two ways to carry this idea over to classification: the agreement-based and inclusion-based translations.

The agreement-based translation retains a category only when every member of MM assigns xx to it, that is, it selects {r∈P∣M⊆Sr(x,c)}=⋂i∈M{ci(x)}\{r \in P \mid M \subseteq S_r(x,\boldsymbol{c})\} = \bigcap_{i \in M}\{c_i(x)\}. This set becomes empty as soon as two members of MM classify xx differently, so it defines a CAC only when MM is a singleton.7 An agreement-based CAC must therefore specify what happens when the members of MM disagree. We consider that the counterpart of social indifference in our setting is to assign xx to every category. For a nonempty coalition MM, the oligarchy rule (with all-category assignment) is defined by

O+M(c)(x)={{r}if M⊆Sr(x,c),Potherwise.O_{+}^{M}(\boldsymbol{c})(x) = \begin{cases} \{r\} & \text{if } M \subseteq S_r(x,\boldsymbol{c}),\\ P & \text{otherwise.} \end{cases}

The inclusion-based translation retains every category to which at least one member of MM assigns xx:

JM(c)(x)={r∈P∣Sr(x,c)∩M≠∅}={ci(x)∣i∈M}.J^{M}(\boldsymbol{c})(x) = \{r \in P \mid S_r(x,\boldsymbol{c}) \cap M \ne\varnothing\} = \{c_i(x) \mid i \in M\}.

We call MM a junta and JMJ^{M} the uniform junta rule. No restriction is imposed on the size of MM. With M=NM = N, the rule O+NO_{+}^{N} is F1F^{1} of Example 1, and JNJ^{N} is F2F^{2} of Example 2. At the profile of these examples, the two rules agree at x1x_1, where the voters agree, and differ at the three disputed proposals, where F1F^{1} returns PP and F2F^{2} returns only the proposed categories.

Both rules give the members of MM the same power over exclusive assignments: under either rule, the outcome at xx is {r}\{r\} if and only if Sr(x,c)∈FMS_r(x,\boldsymbol{c}) \in\mathcal{F}_M. They differ in how categories are included. Under JMJ^{M}, category rr is included if and only if Sr(x,c)∈GMS_r(x,\boldsymbol{c}) \in\mathcal{G}_M, whereas under O+MO_{+}^{M} whether rr is included is not determined by the support of rr alone, since disagreement among the other members of MM puts every category into the outcome. By Fact 4, GM\mathcal{G}_M consists exactly of the sets whose complements are not in FM\mathcal{F}_M.

Proposition 2. Let m≥p≥2m \ge p \ge2 and let M⊆NM \subseteq N be nonempty.

(i) O+MO_+^M is a CAC that satisfies Independence, Unanimity, Category Neutrality and Object Neutrality.

(ii) O+MO_+^M is dictatorial if and only if ∣M∣=1|M| = 1.

(iii) If p=2p = 2, then O+M=JMO_+^M = J^M.

(iv) If p≥3p \ge3 and ∣M∣≥2|M| \ge2, then O+MO_+^M violates Supportiveness.

The proof is in Section A. It follows from (i) and (iii) of Proposition 2 and Lemma 3 that when m≥p≥2m \ge p \ge2, every CAC that satisfies Unanimity satisfies Supportiveness if and only if p=2p = 2.

The oligarchy rule is thus the most direct translation of an oligarchy, and it satisfies the axioms of Maniquet and Mongin [38] without being dictatorial. We nevertheless regard it as unsatisfactory. Under the membership reading of § 3.1, the outcome PP asserts that the object belongs to every category, including categories to which nobody, inside or outside MM, has assigned it. In the participatory-budgeting example, a proposal on which two members of the oligarchy disagree would be placed in every category, as x2x_2, x3x_3 and x4x_4 are under F1F^1 in Example 1. The rule also satisfies the requirement that every category be used in an uninformative way: whenever two oligarchs disagree at some object, that object alone covers all categories. The inclusion-based translation satisfies Supportiveness. Under the tie reading, by contrast, the outcome PP would mean that the question is left open, and O+MO_+^M would be a reasonable counterpart of an oligarchy; the choice between the two readings is therefore substantive.

The two translations also differ in their coalitional structure. Under JMJ^M, the coalitions that secure the inclusion of a category form the grill GM\mathcal{G}_M, which satisfies (p.6) but is not closed under intersection when ∣M∣≥2|M| \ge2, and each member of MM can include a category on her own. Since GM\mathcal{G}_M consists of the sets whose complements are not in FM\mathcal{F}_M, these are exactly the coalitions that can prevent a strict social preference under the oligarchy with oligarchs MM: a coalition that can block a strict preference there can force an assignment here. Agreement, by contrast, cannot serve as the criterion for inclusion. If the coalitions that can secure the inclusion of a category formed a filter FM\mathcal{F}_M with ∣M∣≥2|M| \ge2, every member of MM could veto the inclusion of each category she does not propose, which is the counterpart of an oligarch’s veto,8 and a profile at which two members of MM disagree would leave the object without any category. Such a filter must therefore be an ultrafilter, which on a finite population means a dictatorship. Indeed, an individual with veto power is a dictator, since the outcome at every object is then a nonempty subset of {ci(x)}\{c_i(x)\}.

The class of junta rules

The uniform junta rule is a special case of the general class in which the junta may depend on both the object and the category. For a family of coalitions (Mx,r)x∈X,r∈P(M_{x,r})_{x \in X,r \in P}, the general junta rule is defined, for all c∈CNc \in C^N and all x∈Xx \in X, by

J(Mx,r)(c)(x)={r∈P∣Sr(x,c)∩Mx,r≠∅}.J^{(M_{x,r})}(c)(x)=\{r \in P \mid S_r(x,c) \cap M_{x,r} \ne\varnothing\}.

Thus, category rr is assigned to object xx if and only if at least one member of Mx,rM_{x,r} assigns xx to rr. When Mx,r=MxM_{x,r}=M_x for every x∈Xx \in X and r∈Pr \in P, the junta depends only on the object, and we obtain the object-dependent junta rule

J(Mx)(c)(x)={ci(x)∣i∈Mx}.J^{(M_x)}(c)(x)=\{c_i(x) \mid i \in M_x\}.

Likewise, Mx,r=MrM_{x,r}=M_r gives a rule J(Mr)J^{(M_r)} whose junta depends only on the category, and Mx,r=MM_{x,r}=M gives the uniform junta rule J(M)=JMJ^{(M)}=J^M.

A social classification must assign at least one category to each object and leave no category empty. The following proposition characterises when a junta rule is a CAC.

Proposition 3. Let NN be finite or infinite, m≥p≥2m \ge p \ge2, and let (Mx,r)x∈X,r∈P(M_{x,r})_{x \in X,r \in P} be a family of coalitions. Then J(Mx,r)J^{(M_{x,r})} is a CAC if and only if the following conditions hold:

⋂r∈PMx,r≠∅for every x∈X,(1)\bigcap_{r \in P} M_{x,r} \ne\varnothing\quad\text{for every } x \in X, \tag*{(1)}
⋂x∈XMx,r≠∅for every r∈P.(2)\bigcap_{x \in X} M_{x,r} \ne\varnothing\quad\text{for every } r \in P. \tag*{(2)}

In particular, J(Mx)J^{(M_x)} is a CAC if and only if

⋂x∈XMx≠∅,(3)\bigcap_{x \in X} M_x \ne\varnothing, \tag*{(3)}

J(Mr)J^{(M_r)} is a CAC if and only if ⋂r∈PMr≠∅\bigcap_{r \in P} M_r \ne\varnothing, and JMJ^M is a CAC if and only if M≠∅M \ne\varnothing.

The two conditions above correspond to the two requirements on a social classification. Condition eq. (1) ensures that, at each object, some individual belongs to the junta for every category. Whatever category she assigns to that object is therefore included in the social classification. Without such an individual, everyone can assign the object to a category whose junta does not contain her, leaving the object without any category.

Condition eq. (2) ensures that, for each category, some individual belongs to its junta at every object. Since her individual classification uses every category, she assigns at least one object to that category, and this assignment is included in the social classification. Without such an individual, everyone can use the category only at an object whose corresponding junta does not contain her, leaving the category empty. The individuals ensuring these requirements may differ across objects and categories.

By the definition of GM\mathcal{G}_M, the rule can be written in terms of its coalition families:

r∈J(Mx,r)(c)(x)⟺Sr(x,c)∈GMx,r.r \in J^{(M_{x,r})}(\mathbf{c})(x) \Longleftrightarrow S_r(x,\mathbf{c}) \in\mathcal{G}_{M_{x,r}}.

For J(Mx)J^{(M_x)}, fix an object xx. The support sets (Sr(x,c))r∈P(S_r(x,\mathbf{c}))_{r \in P} have union N∈GMxN \in\mathcal{G}_{M_x}. By the grill property applied to this finite union, at least one support set belongs to GMx\mathcal{G}_{M_x}, so the outcome at xx is nonempty whenever Mx≠∅M_x \ne\varnothing. Ensuring that no category is left empty additionally requires the common-member condition eq. (3). Every individual in ⋂x∈XMx\bigcap_{x \in X} M_x has her individual classification included in the social classification, and the assignments of the other individuals are included at the objects to whose juntas they belong.

The rule F2′F^{2'} of Example 2 is J(Mx,r)J^{(M_{x,r})} with Mx2,b={2}M_{x_2,b} = \{2\}, Mx4,c={1}M_{x_4,c} = \{1\} and Mx,r=NM_{x,r} = N for every other pair. Its juntas satisfy eqs. (1) and (2), since voter 2 belongs to every junta at x2x_2 and to every junta for Biodiversity, and voter 1 belongs to every junta at x4x_4 and to every junta for Climate. No voter, however, belongs to all juntas, and neither individual classification is included in F2′(c)F^{2'}(\mathbf{c}): voter 1’s Biodiversity at x2x_2 and voter 2’s Climate at x4x_4 are dropped.

Among the general junta rules that are CACs, only dictatorial rules are resolute. For each xx, choose ix∈⋂r∈PMx,ri_x \in\bigcap_{r \in P} M_{x,r}, which is nonempty by eq. (1). If some Mx,rM_{x,r} contained j≠ixj \ne i_x, Lemma 1 would give a profile with cj(x)=rc_j(x) = r and cix(x)=s≠rc_{i_x}(x) = s \ne r, so both rr and ss would be accepted at xx. Thus Mx,r={ix}M_{x,r} = \{i_x\} for every x,rx,r, and eq. (2) implies that all ixi_x coincide. This recovers the dictatorship of Maniquet and Mongin [38] within the class of junta rules, and allowing multiple memberships is exactly what permits juntas with more than one member.

The next section characterises these rules and then returns to Example 4 to ask whether individual responsibility and equal treatment allow any supported assignments to be rejected.

Characterisations of junta rules

This section characterises junta rules by means of three axioms, each added to Independence and Supportiveness. The three axioms – Responsibility, Split Stability, and Rejection Stability – express different requirements. Responsibility looks for an individual whose report secures an assignment, Split Stability compares profiles before and after the support of a category is split, and Rejection Stability concerns coalitions that fail to secure one. Split Stability and Rejection Stability imply the grill property, the former by splitting a winning coalition and the latter by combining losing ones, while Responsibility identifies the generating junta directly.

Characterisation via Responsibility

We begin with a characterisation of junta rules based on the role of individual reports in ensuring acceptance. Responsibility requires that every accepted category with supporters be secured by one supporter’s complete classification, regardless of the other individuals’ reports.

Consider again the voters classifying proposals into categories for the platform. Each voter assigns every proposal to one category and uses every category, while the platform’s classification may assign several categories to the same proposal. These categories provide different thematic groups for comparing proposals. One may require that every accepted assignment with individual support be retained on the basis of one supporter’s complete classification, whatever the other individuals report. This excludes acceptance that depends on agreement among reports that are separately insufficient. It is an additional normative requirement and does not follow from allowing multiple memberships. Fioravanti [24] motivates granting individuals decisive power by their expertise or their direct connection to the objects or categories involved. Responsibility makes no assumption about individuals’ expertise and does not require decision power to be equally distributed among them.

For i∈Ni \in N, denote by (ci,d−i)(c_i,d_{-i}) the profile in which individual ii reports cic_i and the other individuals report d−i∈CN∖{i}d_{-i} \in C^{N\setminus\{i\}}.

Axiom (Responsibility). A CAC FF satisfies Responsibility (Resp.) if, for every c∈CNc \in C^N, x∈Xx \in X and r∈Pr \in P such that r∈F(c)(x)r \in F(c)(x) and Sr(x,c)≠∅S_r(x,c) \ne\varnothing, there exists i∈Sr(x,c)i \in S_r(x,c) such that

r∈F(ci,d−i)(x)for every d−i∈CN∖{i}.r \in F(c_i,d_{-i})(x) \qquad\text{for every } d_{-i} \in C^{N\setminus\{i\}}.

The individual may depend on the profile, the object and the category, but her report must ensure acceptance for every choice of the other reports. Her report is sufficient for inclusion; it need not be necessary, and more than one individual may satisfy the requirement. Under Supportiveness, every assignment can be traced to at least one such supporter, who in the participatory-budgeting setting could be asked to justify it. This provides a basis for accountability, although Responsibility does not single out a unique supporter.

Example 5. Let N={1,2,3}N = \{1,2,3\} be three trained classifiers assigning two proposal descriptions, X={x1,x2}X = \{x_1,x_2\}, to two thematic labels, P={a,b}P = \{a,b\}, read as Arts and Biodiversity. The only surjective classifications are u=(a,b)u = (a,b) and v=(b,a)v = (b,a). Let F5F^5 be majority voting at each object. This rule always uses both categories and therefore defines a CAC on this domain. At c=(u,u,v)c = (u,u,v), it assigns {a}\{a\} to x1x_1. The profiles (u,v,v)(u,v,v) and (v,u,v)(v,u,v) retain the reports of individuals 1 and 2, respectively, but both assign {b}\{b\} to x1x_1. Neither supporter’s report ensures acceptance, so majority voting violates Responsibility, although each supporter is pivotal at cc. JNJ^N accepts aa whenever either supporter retains uu, whatever the others report.

Responsibility does not imply Supportiveness, since it imposes no condition on accepted categories without supporters; the constant CAC F(c)(x)=PF(c)(x) = P satisfies Responsibility. When Independence also holds, the individual’s assignment at the object suffices, whatever her assignments to the other objects.

Lemma 4. Let FF satisfy Independence and Responsibility. If r∈F(c)(x)r \in F(c)(x) and Sr(x,c)≠∅S_r(x,c) \ne\varnothing, there exists i∈Sr(x,c)i \in S_r(x,c) such that

ci′(x)=r⟹r∈F(c′)(x)for every c′∈CN.(4)c'_i(x) = r \quad\Longrightarrow\quad r \in F(c')(x) \qquad\text{for every } c' \in C^N. \tag*{(4)}

Proof. Take ii as in Responsibility and any c′c' with ci′(x)=rc'_i(x) = r. The profile (ci,c−i′)(c_i,c'_{-i}) belongs to CNC^N, and Responsibility gives r∈F(ci,c−i′)(x)r \in F(c_i,c'_{-i})(x). Since ci(x)=ci′(x)c_i(x) = c'_i(x), this profile and c′c' agree at xx, and Independence gives r∈F(c′)(x)r \in F(c')(x). ▪

Without Independence the assignment alone need not suffice. Under F3F^3 in Example 3, Arts at x2x_2 is secured by voter 2’s individual classification c2c_2, which uses Arts twice, but not by her assignment at x2x_2. If she instead reports c2′=(c,a,b,c)c'_2=(c,a,b,c), she still assigns x2x_2 to Arts, yet the proposal leaves Arts.

In the single-valued setting, Fioravanti [24] calls an individual minimally semi-decisive for a pair (x,r)(x,r) if her assignment of xx to rr suffices for society to assign xx to rr at every profile. Condition eq. (4) extends this notion to our setting. The power concerns the inclusion of a category, which need not be the only category assigned to the object. One voter may ensure that a proposal is placed in Biodiversity while another ensures that it is placed in Arts, and when they make these assignments both categories appear in the platform’s classification. Multiple memberships allow these powers to coexist at the same object. For each pair (x,r)(x,r), let Mx,rM_{x,r} be the coalition of all individuals satisfying eq. (4). Under Supportiveness, Lemma 4 guarantees that every assigned category has a supporter in the corresponding coalition, and conversely the support of any member suffices. For F2′F^{2'} in Example 2, these coalitions are the ones that define the rule: only voter 2 satisfies eq. (4) for x2x_2 and Biodiversity, only voter 1 for x4x_4 and Climate, and both voters for every other pair. This gives the following representation.

Theorem 1. Let m≥p≥2m \ge p \ge2. A CAC FF satisfies Independence, Supportiveness and Responsibility if and only if there exists a family (Mx,r)x∈X,r∈P(M_{x,r})_{x\in X,r\in P} satisfying eqs. (1) and (2) such that

F(c)(x)={r∈P∣Sr(x,c)∩Mx,r≠∅}for all c∈CN and x∈X.(5)F(\mathbf{c})(x)=\{r\in P\mid S_r(x,\mathbf{c})\cap M_{x,r}\ne\varnothing\}\quad\text{for all }\mathbf{c}\in\mathcal{C}^{N}\text{ and }x\in X. \tag*{(5)}

This family is unique.

Under Category Neutrality the same junta applies to all categories at a given object, and under Object Neutrality as well it applies to all objects.

Corollary 1. Let m≥p≥2m \ge p \ge2.

(i) A CAC FF satisfies Independence, Supportiveness, Responsibility and Category Neutrality if and only if F=J(Mx)F=J^{(M_x)} for a family of coalitions satisfying eq. (3).

(ii) A CAC FF satisfies Independence, Supportiveness, Responsibility, Category Neutrality and Object Neutrality if and only if F=JMF=J^M for a nonempty coalition M⊆NM\subseteq N.

The proof of Theorem 1 identifies a responsible individual directly, so the characterisation also holds when m=pm=p or p=2p=2.

Characterisation via Split Stability

Our second characterisation concerns changes in the support sets. It asks how the social classification should respond when support for one category is divided between that category and a previously unused one.

Given c∈CN\mathbf{c}\in\mathcal{C}^{N}, x∈Xx\in X, distinct categories k,k′∈Pk,k'\in P and a nonempty coalition B⊆Sk(x,c)B\subseteq S_k(x,\mathbf{c}), let cx,B→k′=(ci′)i∈N\mathbf{c}^{x,B\to k'}=(c'_i)_{i\in N} denote the profile obtained from c\mathbf{c} by setting ci′(x)=k′c'_i(x)=k' for every i∈Bi\in B and leaving all other entries unchanged. We call the passage from c\mathbf{c} to cx,B→k′\mathbf{c}^{x,B\to k'} a defection of BB to k′k' at xx. The defection is admissible if cx,B→k′∈CN\mathbf{c}^{x,B\to k'}\in\mathcal{C}^{N}, that is, if ci−1(k)∖{x}≠∅c_i^{-1}(k)\setminus\{x\}\ne\varnothing for every i∈Bi\in B. Not every defection from an arbitrary profile is admissible. When m>pm>p, however, Lemma 2 implies that every opinion vector at xx is realised by a profile at which every defection at xx is admissible.

Axiom (Split Stability). A CAC FF satisfies Split Stability if, for all x∈Xx\in X, all c∈CN\mathbf{c}\in\mathcal{C}^{N}, all k,k′∈Pk,k'\in P such that Sk(x,c)≠∅S_k(x,\mathbf{c})\ne\varnothing and Sk′(x,c)=∅S_{k'}(x,\mathbf{c})=\varnothing, and every nonempty B⊆Sk(x,c)B\subseteq S_k(x,\mathbf{c}) such that cx,B→k′∈CN\mathbf{c}^{x,B\to k'}\in\mathcal{C}^{N}, the following conditions hold:

(SS.1) F(cx,B→k′)(x)∖{k,k′}=F(c)(x)∖{k,k′}F(\mathbf{c}^{x,B\to k'})(x)\setminus\{k,k'\}=F(\mathbf{c})(x)\setminus\{k,k'\}.

(SS.2) F(cx,B→k′)(x)∩{k,k′}≠∅⟺k∈F(c)(x)F(\mathbf{c}^{x,B\to k'})(x)\cap\{k,k'\}\ne\varnothing\Longleftrightarrow k\in F(\mathbf{c})(x).

To understand the normative aspect of the condition, consider a profile at which some individuals assign xx to a category kk and nobody assigns it to k′k'. Suppose that part of the supporters of kk switch to k′k', for reasons unrelated to manipulation: the two categories may be similar, or new information about xx may lead some individuals to revise their view. Note that xx may also be classified into other categories under either profile. Then, condition (SS.1) requires that the categories supported by others be unaffected by this change. Condition (SS.2) requires that at least one of kk and k′k' be accepted after the split exactly when kk was accepted before it. If the original support was strong enough to place xx in kk, then after the split xx should be placed in kk or k′k'. Conversely, if xx was not placed in kk, the defectors should not be able to obtain a membership in kk or k′k' by splitting.

We return to Example 3.

Example 3 (Continued). Suppose now that voter 2 moves x1x_1 from Arts to Climate, a category that nobody had proposed for x1x_1. Since she still uses Arts for x2x_2, her new report c2′=(c,a,b,c)c'_2=(c,a,b,c) is a classification, and the profile c′=(c1,c2′)c'=(c_1,c'_2) gives

F3(c′)=({a,c},{b},{c},{a,c}).F^3(c')=\bigl(\{a,c\},\{b\},\{c\},\{a,c\}\bigr).

At x1x_1, the support for Arts has been split between Arts and Climate. The proposal stays in Arts, enters Climate and remains outside Biodiversity. Under F1F^1, by contrast, the same change turns {a}\{a\} into PP at x1x_1, so Biodiversity appears although nobody proposed it. At x2x_2 and x4x_4, nobody changed her classification, yet the outcome changes from {a,b}\{a,b\} to {b}\{b\} and from {a}\{a\} to {a,c}\{a,c\}, because voter 2 now uses Arts once and Climate twice. ⋄

When voter 2 moves x1x_1 from Arts to the previously unused category Climate, F3F^3 keeps x1x_1 under Arts, adds Climate and leaves Biodiversity out, as (SS.1) and (SS.2) require. The same move makes F1F^1 add Biodiversity at x1x_1, contrary to (SS.1). The changes that F3F^3 makes at x2x_2 and x4x_4 violate Independence. Split Stability constrains the outcome only at the object where the split occurs.

When m=pm=p, every individual classification is a bijection, no admissible split exists, and the axiom is vacuous. When p=2p=2, no category is vacant unless the profile is unanimous at xx; under Supportiveness, Split Stability then holds automatically. We therefore state the result under m>p≥3m>p\ge3.

Theorem 2. Let m>p≥3m>p\ge3. A CAC FF satisfies Independence, Supportiveness and Split Stability if and only if there exists a family (Mx)x∈X(M_x)_{x\in X} of nonempty coalitions satisfying eq. (3) such that F=J(Mx)F=J^{(M_x)}. The family is unique.

The proof, in Section A, first shows that whether kk is assigned to xx depends only on the support set of kk at xx, then that the family of winning support sets Wx\mathcal{W}_x does not depend on kk, and finally that Wx\mathcal{W}_x is a grill. Finiteness of NN is used only at the last step, to conclude that the grill is principal (Fact 1). In particular, the junta at each object must be the same for all categories. The rule F2′F^{2'} of Example 2 shows what goes wrong otherwise: if voter 1 alone proposes Biodiversity for x2x_2, it is dropped, but if she moves x2x_2 to the previously unused category Climate instead, which is admissible when she uses Biodiversity for another proposal, the proposal enters Climate, contrary to (SS.2). Adding Object Neutrality removes the dependence of the junta on the object.

Corollary 2. Let m>p≥3m>p\ge3. A CAC satisfies Independence, Supportiveness, Split Stability and Object Neutrality if and only if it is a uniform junta rule JMJ^M for some nonempty coalition MM.

Characterisation via Rejection Stability

Our third characterisation starts from the idea that coalitions whose submitted classifications are not enough to secure an assignment should not become enough merely by being combined. Consider again the platform’s classification of proposals, and suppose that a proposal is not placed in Biodiversity at either of two profiles. In each profile, the reports of the individuals who assign the proposal to Biodiversity are insufficient for the assignment. Rejection Stability requires the assignment to remain rejected whenever each supporter in a profile retains her own complete classification from one of the two rejecting profiles. The requirement concerns complete classifications because, without Independence, an individual’s assignments to other objects may affect the classification of the proposal.

Axiom (Rejection Stability). A CAC FF satisfies Rejection Stability (Rej. Stab.) if, for every x∈Xx \in X, r∈Pr \in P and c,d,e∈CN\mathbf{c}, \mathbf{d}, \mathbf{e} \in C^N such that r∉F(c)(x)r \notin F(\mathbf{c})(x) and r∉F(d)(x)r \notin F(\mathbf{d})(x), we have r∉F(e)(x)r \notin F(\mathbf{e})(x) whenever ei∈{ci,di}e_i \in\{c_i,d_i\} for every i∈Sr(x,e)i \in S_r(x,\mathbf{e}).

The two rejecting profiles may coincide. Each supporting report in e\mathbf{e} must be the same individual’s report in c\mathbf{c} or in d\mathbf{d}, but different supporters may take their reports from different profiles, and only some of the original supporters need retain their support. Individuals who do not support the assignment in e\mathbf{e} may submit arbitrary reports.

Example 6. Take the majority rule of Example 5, and consider c=(u,v,v)\mathbf{c}=(u,v,v), d=(v,u,v)\mathbf{d}=(v,u,v) and e=(u,u,v)\mathbf{e}=(u,u,v). The outcome at x1x_1 is {b}\{b\} under both source profiles and {a}\{a\} under the target. Both supporters of aa in e\mathbf{e} retain their reports, since e1=c1e_1=c_1 and e2=d2e_2=d_2. Majority voting therefore violates Rejection Stability.

Rejection Stability does not imply Supportiveness either: the constant CAC F(c)(x)=PF(\mathbf{c})(x)=P satisfies it vacuously. The closure in Rejection Stability extends to any finite number of rejecting profiles.

Lemma 5. Let FF satisfy Rejection Stability, and let c1,…,cK∈CN\mathbf{c}^1,\ldots,\mathbf{c}^K \in C^N, with K≥1K \ge1, be profiles such that r∉F(ck)(x)r \notin F(\mathbf{c}^k)(x) for every kk. If e∈CN\mathbf{e} \in C^N satisfies ei∈{ci1,…,ciK}e_i \in\{c_i^1,\ldots,c_i^K\} for every i∈Sr(x,e)i \in S_r(x,\mathbf{e}), then r∉F(e)(x)r \notin F(\mathbf{e})(x).

Proof. We argue by induction on KK. For K≤2K \le2, the claim is Rejection Stability. Let K≥3K \ge3, and let UU be the set of supporters i∈Sr(x,e)i \in S_r(x,\mathbf{e}) with ei∉{ci1,…,ciK−1}e_i \notin\{c_i^1,\ldots,c_i^{K-1}\}, so that ei=ciKe_i=c_i^K for i∈Ui \in U. By Lemma 1, there is c0∈C\mathbf{c}^0 \in C with c0(x)≠rc^0(x) \ne r. Let e′\mathbf{e}' be obtained from e\mathbf{e} by replacing eie_i with c0c^0 for every i∈Ui \in U. Every supporter of rr at xx in e′\mathbf{e}' retains a report from one of c1,…,cK−1\mathbf{c}^1,\ldots,\mathbf{c}^{K-1}, so r∉F(e′)(x)r \notin F(\mathbf{e}')(x) by the induction hypothesis. Every supporter in e\mathbf{e} retains her report from e′\mathbf{e}' or from cK\mathbf{c}^K, and Rejection Stability gives r∉F(e)(x)r \notin F(\mathbf{e})(x). ■

Lemma 6. Let m≥p≥2m \ge p \ge2.

(i) On every population, Supportiveness and Responsibility imply Rejection Stability.

(ii) If NN is finite, Rejection Stability implies Responsibility.

Proof. (i) Let FF satisfy Supportiveness and Responsibility. Consider profiles c,d\mathbf{c},\mathbf{d} that reject rr at xx and a profile e\mathbf{e} satisfying the report condition of Rejection Stability. Suppose that r∈F(e)(x)r \in F(\mathbf{e})(x). By Supportiveness, rr has a supporter in e\mathbf{e}, and Responsibility provides such a supporter ii whose report eie_i ensures acceptance for every choice of the other reports. Since ei=cie_i=c_i or ei=die_i=d_i, one of the source profiles must accept rr at xx, a contradiction. Hence Rejection Stability holds.

(ii) Let NN be finite and suppose that FF satisfies Rejection Stability. Take e\mathbf{e}, xx and rr such that r∈F(e)(x)r \in F(\mathbf{e})(x) and Sr(x,e)≠∅S_r(x,\mathbf{e}) \ne\varnothing. Suppose that Responsibility fails at this profile, object and category. Then, for every i∈Sr(x,e)i \in S_r(x,\mathbf{e}), there is a profile di∈CN\mathbf{d}^i \in C^N such that

dii=eiandr∉F(di)(x).d_i^i=e_i \qquad\text{and}\qquad r \notin F(\mathbf{d}^i)(x).

These finitely many profiles reject rr at xx, and every supporter ii in e\mathbf{e} retains her report from di\mathbf{d}^i. By Lemma 5, r∉F(e)(x)r \notin F(\mathbf{e})(x), a contradiction. Thus Responsibility holds. ■

Thus, on finite populations, Responsibility and Rejection Stability are equivalent under Supportiveness, without Independence or either neutrality axiom. Combining Lemma 6 with Theorem 1 and Corollary 1 gives the third characterisation.

Theorem 3. Let m≥p≥2m \ge p \ge2.

  1. A CAC FF satisfies Independence, Supportiveness and Rejection Stability if and only if F=J(Mx,r)F = J^{(M_{x,r})} for a unique family of coalitions satisfying eqs. (4.1a) and (4.1b).

  1. Adding Category Neutrality characterises the object-dependent junta rules J(Mx)J^{(M_x)} satisfying eq. (3), and adding Object Neutrality as well characterises the uniform junta rules JMJ^M with M≠∅M \ne\varnothing.

Under Independence and Supportiveness and when m>p≥3m > p \ge3, Split Stability and Rejection Stability both yield the grill property. Split Stability requires that splitting a winning coalition leaves at least one winning part, while Rejection Stability requires that combining two losing coalitions never produces a winning one. The difference is that Split Stability also forces the winning coalitions to be the same for all categories at a given object, whereas Rejection Stability allows them to depend on the category. Thus, on finite populations, Theorems 2 and 3 show that Split Stability is equivalent to Rejection Stability together with Category Neutrality.

Independence cannot be omitted. The rule F3F^3 of Example 3 satisfies Supportiveness and both neutrality axioms. Every category it accepts is secured by the complete report of voter 1 or of voter 2, so Responsibility holds, and Lemma 6 gives Rejection Stability. Yet it is not a junta rule, since its outcomes at x2x_2 and x4x_4 change although nobody changes her classification of these proposals. Part (ii) of Lemma 6 fails on infinite populations, as § 8 shows.

Relations between the axioms

On finite populations, under Independence and Supportiveness and for m>p≥3m > p \ge3, Split Stability is equivalent to Responsibility together with Category Neutrality, and to Rejection Stability together with Category Neutrality (Theorem 2, Corollary 1, and Lemma 6). Rejection Stability implies Responsibility for all m≥p≥2m \ge p \ge2, even without Independence or Supportiveness (Lemma 6(ii)). We now examine whether Unanimity can replace Supportiveness in the characterisations and what happens when we impose Anonymity.

Replacing Supportiveness with Unanimity

Our characterisations impose Supportiveness in place of the Unanimity axiom of Maniquet and Mongin [38]. The following proposition identifies when Unanimity and the other axioms already imply Supportiveness.

Proposition 4. Let m≥p≥2m \ge p \ge2.

(i) If m>pm > p, every CAC satisfying Unanimity, Independence and Split Stability satisfies Supportiveness.

(ii) Every CAC satisfying Unanimity and Rejection Stability satisfies Supportiveness.

Thus Theorems 2 and 3 and Corollary 2 remain valid when Supportiveness is replaced by Unanimity. For the characterisations through Responsibility in Theorem 1 and Corollary 1, this replacement is valid if and only if p=2p = 2 (cf. Proposition 2).

For p≥3p \ge3, Supportiveness is therefore needed as a separate assumption in the characterisation through Responsibility. Responsibility, as stated, imposes no condition on accepted categories without supporters. In Example 1, Climate at x2x_2 is such a category: Responsibility imposes nothing on it, and Unanimity does not exclude it because the voters disagree. One could strengthen Responsibility by requiring every accepted category to be secured by some individual’s complete classification, whether or not that individual supports the assignment. Combined with Unanimity, this condition forces the responsible individual to be a supporter: if she assigned xx to some s≠rs \ne r, letting every other individual do the same would produce a profile that is unanimous on ss at xx. Unanimity together with this condition is therefore equivalent to Supportiveness together with Responsibility, and we keep the more transparent second combination.

Anonymity and Responsibility

The preceding characterisations allow individuals to differ in their authority to secure an assignment. We now consider the effect of requiring equal treatment.

Notes: F1F^1 and F2F^2 correspond to O+NO^N_+ and JNJ^N, respectively; F2′F^{2'} is a general junta rule; F4F^4 is a quota rule; and F5F^5 is the majority rule on the domain n=3n = 3, m=p=2m = p = 2 used in Example 5.

Axiom (Anonymity). A CAC FF satisfies Anonymity if, for every bijection τ:N→N\tau: N \to N and every c∈CN\mathbf{c} \in C^N, F((cτ(i))i∈N)=F(c)F((c_{\tau(i)})_{i \in N}) = F(\mathbf{c}).

Note that in the running example, only F2′F^{2'} and F3F^3 violate Anonymity. In Example 2, if the voters exchange their reports, the outcome at x2x_2 becomes {a,b}\{a,b\} instead of {a}\{a\}; and in Example 3, only voter 2’s categories are subject to the repetition requirement.

Among general junta rules that are CACs, Anonymity holds if and only if Mx,r=NM_{x,r} = N for every object-category pair. Indeed, if one individual’s sole support secures an assignment, Anonymity gives every individual the same authority. The preceding characterisations therefore yield the following result.

Corollary 3. Let NN be finite and m≥p≥2m \ge p \ge2. A CAC satisfies Independence, Supportiveness, Anonymity and either Responsibility or Rejection Stability (or Split Stability, when m>p≥3m > p \ge3) if and only if it is JNJ^N.

Under the maintained axioms, Anonymity and Responsibility are thus compatible only through the whole-population junta rule JNJ^N, which accepts every supported assignment. To retain the possibility of rejecting some supported assignments while preserving Responsibility, individuals must be allowed to differ in whose support suffices. The quota procedure in Example 4 treats individuals equally but violates Responsibility, since neither voter can secure an assignment for which both voters’ support is required. Allowing different individuals to have different decision power therefore permits junta rules to reject some supported assignments.9

Table 1 collects the properties of the running examples. In particular, F2′F^{2'} and F4F^4 illustrate the distinction above: the former satisfies Responsibility and the latter satisfies Anonymity, although they give the same outcome at the original profile. The table also includes the majority rule F5F^5 of Example 5.

Infinite populations

On finite populations, when m>p≥3m > p \ge3, the three axioms of § 6 are interchangeable under Independence, Supportiveness and Category Neutrality. This section shows that they come apart when the population is infinite. The question is natural because the theorem of Maniquet and Mongin [38], our benchmark, fails on an infinite population (§ 8.1). Split Stability and Rejection Stability make the winning coalitions a grill that need not ેહ be principal and characterise a strictly larger class, the limiting junta rules (§ 8.2). Responsibility keeps the grill principal and characterises the junta rules as before (§ 8.3). Throughout, NN is an arbitrary set with at least two members, finite or infinite, while XX and PP remain finite with m≥p≥2m \ge p \ge2.

Benchmark: the Maniquet–Mongin theorem with an infinite population

Maniquet and Mongin [38] show that the decisive coalitions of a resolute CAC satisfying Independence and Unanimity form an ultrafilter U\mathcal{U} on NN, with F(c)(x)={r}F(\mathbf{c})(x)=\{r\} if and only if Sr(x,c)∈US_r(x,\mathbf{c})\in\mathcal{U}. Their argument uses the finiteness of NN only through the fact that every ultrafilter on a finite set is principal.

Corollary 4. Assume the axiom of choice. If NN is infinite, there exists a resolute CAC satisfying Independence, Supportiveness and Unanimity that is not dictatorial.

Proof. Let U\mathcal{U} be a free ultrafilter on NN and let F(c)(x)={r}F(\mathbf{c})(x)=\{r\} for the unique r∈Pr\in P with Sr(x,c)∈US_r(x,\mathbf{c})\in\mathcal{U}; uniqueness and existence follow from (p.4), (p.1) and the grill property of U\mathcal{U} applied to the finite partition (Sr(x,c))r∈P(S_r(x,\mathbf{c}))_{r\in P}. Since XX is finite, ⋂x∈XSrx(x,c)∈U\bigcap_{x\in X}S_{r_x}(x,\mathbf{c})\in\mathcal{U}, where F(c)(x)={rx}F(\mathbf{c})(x)=\{r_x\}, and any individual ii in this intersection satisfies ci(x)=rxc_i(x)=r_x for every xx, so the outcome is surjective. Independence and Supportiveness are immediate, and a dictator dd would give {d}∈U\{d\}\in\mathcal{U}, contradicting freeness. ■

Hence the theorem of Maniquet and Mongin holds if and only if NN is finite. The rule in Corollary 4 relies on a free ultrafilter, whose existence requires a weak form of the axiom of choice, so it cannot be defined explicitly. This is the classification counterpart of the observation of Fishburn [28] and Kirman and Sondermann [35].

Split Stability and Rejection Stability: limiting junta rules

Fix x∈Xx\in X and r∈Pr\in P. A coalition A⊆NA\subseteq N is winning for rr at xx if r∈F(c)(x)r\in F(\mathbf{c})(x) for every c∈CN\mathbf{c}\in C^N with Sr(x,c)=AS_r(x,\mathbf{c})=A, and null for rr at xx if r∉F(c)(x)r\notin F(\mathbf{c})(x) for every such c\mathbf{c}. We write Wx,r\mathcal{W}_{x,r} for the family of winning coalitions.

Under Independence and Supportiveness, Split Stability and Rejection Stability each make every coalition either winning or null for rr at xx, and make Wx,r\mathcal{W}_{x,r} a grill, whether NN is finite or infinite (Lemma 7 in Section A). Under Split Stability, or under Rejection Stability and Category Neutrality, Wx,r=Wx\mathcal{W}_{x,r}=\mathcal{W}_x does not depend on rr, and under Object Neutrality as well it does not depend on xx. On a finite population this grill is principal, which yields the junta rules of § 6. On an infinite population it need not be. To describe the resulting rules, call a nonempty family M\mathcal{M} of nonempty coalitions directed if for all M,M′∈MM,M'\in\mathcal{M} there is M′′∈MM''\in\mathcal{M} with M′′⊆M∩M′M''\subseteq M\cap M'.

Definition 2. Given a family (Mx)x∈X(\mathcal{M}_x)_{x\in X} of directed families, the object-dependent limiting junta rule L(Mx)L^{(\mathcal{M}_x)} is defined, for all c∈CN\mathbf{c}\in C^N and all x∈Xx\in X, by

L(Mx)(c)(x)=⋂M∈MxJM(c)(x)={r∈P∣Sr(x,c)∩M≠∅ for all M∈Mx}.L^{(\mathcal{M}_x)}(\mathbf{c})(x)=\bigcap_{M\in\mathcal{M}_x}J^M(\mathbf{c})(x)=\{r\in P\mid S_r(x,\mathbf{c})\cap M\ne\varnothing\text{ for all }M\in\mathcal{M}_x\}.

It is a uniform limiting junta rule, written LML^{\mathcal{M}}, if Mx=M\mathcal{M}_x=\mathcal{M} for every x∈Xx\in X.

Reading each M∈MxM\in\mathcal{M}_x as a committee, a category is attached to xx exactly when every committee contains someone who attaches it. The rule is the limit of junta rules as the committees shrink along Mx\mathcal{M}_x, in the same way as the limiting dictatorial rules of Torres [48] are limits of dictatorships. An object-dependent junta rule J(Mx)J^{(\mathcal{M}_x)} is the case Mx={Mx}\mathcal{M}_x=\{M_x\}.

For each xx, the family G(Mx)={S⊆N∣S∩M≠∅ for all M∈Mx}\mathcal{G}(\mathcal{M}_x)=\{S\subseteq N\mid S\cap M\ne\varnothing\text{ for all }M\in\mathcal{M}_x\} is a grill, every grill arises in this way from the directed family of its blocking coalitions, and r∈L(Mx)(c)(x)r\in L^{(\mathcal{M}_x)}(\mathbf{c})(x) if and only if Sr(x,c)∈G(Mx)S_r(x,\mathbf{c})\in\mathcal{G}(\mathcal{M}_x) (Lemma 8 in Section A). The limiting junta rules are therefore exactly the rules that accept a category when its support set belongs to a grill. Such a rule is a CAC if and only if

⋂x∈XMx≠∅for every family (Mx)x∈X with Mx∈Mx for each x,(6)\bigcap_{x\in X}M_x\ne\varnothing\quad\text{for every family }(M_x)_{x\in X}\text{ with }M_x\in\mathcal{M}_x\text{ for each }x, \tag*{(6)}

a condition that holds automatically for uniform limiting junta rules. This yields the characterisation on an arbitrary population.

Theorem 4. Let NN be finite or infinite.

(a) Let m>p≥3m > p \ge3. A CAC FF satisfies Independence, Supportiveness and Split Stability if and only if F=L(Mx)F = L^{(\mathcal{M}_x)} for a family of directed families satisfying eq. (6).

(b) Let m≥p≥2m \ge p \ge2. A CAC FF satisfies Independence, Supportiveness, Rejection Stability and Category Neutrality if and only if F=L(Mx)F = L^{(\mathcal{M}_x)} for a family of directed families satisfying eq. (6).

In both cases, adding Object Neutrality characterises the uniform limiting junta rules.

Thus, when NN is finite or infinite and m>p≥3m > p \ge3, a CAC satisfying Independence and Supportiveness satisfies Split Stability if and only if it satisfies Rejection Stability and Category Neutrality. The next proposition compares limiting junta rules with junta rules.

Proposition 5. (i) L(Mx)=J(Mx)L^{(\mathcal{M}_x)} = J^{(M_x)} for some family (Mx)x∈X(M_x)_{x\in X} of coalitions if and only if every Mx\mathcal{M}_x has a smallest member, in which case MxM_x is the smallest member of Mx\mathcal{M}_x.

(ii) If NN is infinite and M\mathcal{M} is the family of cofinite coalitions, then LM(c)(x)={r∈P∣Sr(x,c) is infinite}L^{\mathcal{M}}(\mathbf{c})(x) = \{r \in P \mid S_r(x,\mathbf{c})\text{ is infinite}\}, and LML^{\mathcal{M}} is not a junta rule: LM≠J(Mx,r)L^{\mathcal{M}} \ne J^{(\mathcal{M}_{x,r})} for every family (Mx,r)x∈X,r∈P(\mathcal{M}_{x,r})_{x\in X,r\in P} of coalitions.

By Fact 4, part (i) says that a limiting junta rule is an object-dependent junta rule exactly when every winning family WxW_x is principal. On a finite population this is always the case: by directedness, a member of Mx\mathcal{M}_x of minimal cardinality is contained in every other member, so every limiting junta rule is an object-dependent junta rule. This is the counterpart of the fact that every ultrafilter on a finite set is principal, and it is the only point at which finiteness matters. Under the respective assumptions of Theorem 4, Split Stability and Rejection Stability therefore characterise a class strictly larger than the object-dependent junta rules when NN is infinite. We call the rule in (ii) the Fréchet rule, since its winning family consists of the infinite sets, which are exactly the sets that meet every member of the Fréchet filter of cofinite sets.10 It satisfies Independence, Supportiveness, Split Stability, Rejection Stability and both neutrality axioms, yet it is not a junta rule. Its committees are defined without any appeal to the axiom of choice, so its existence is provable in ZF, whereas that of the rules in Corollary 4 is not. The difference lies in the form of the output. A set-valued rule needs only a directed family of committees, whereas a single-valued rule needs its decisive coalitions to form an ultrafilter. The Fréchet rule also violates Responsibility: starting from unanimous support for rr at xx, retain any one individual’s report and let all others assign xx to another category, which is possible by Lemma 1; the category then has a single supporter and is rejected.

Responsibility: junta rules on every population

Responsibility continues to characterise junta rules on infinite populations.

Corollary 5. Let NN be infinite and m≥p≥2m \ge p \ge2. Then Theorem 1 and Corollary 1 remain valid. In particular, an object-dependent limiting junta rule satisfies Responsibility if and only if every winning family WxW_x is principal, that is, if and only if it is an object-dependent junta rule.

In fact, the proofs of Theorem 1 and Corollary 1 do not use the finiteness of NN. An object-dependent limiting junta rule satisfies Independence, Supportiveness and Category Neutrality, so by Corollary 1(i) it satisfies Responsibility if and only if it is an object-dependent junta rule. By Proposition 5(i) and Fact 4, this holds if and only if every WxW_x is principal.

The contrast with Theorem 4 lies in the way each axiom identifies the members of a junta. Split Stability and Rejection Stability relate a coalition to two of its parts at a time: if the union of two coalitions is winning, so is one of them (Lemma 7). On a finite population, splitting a winning coalition again and again ends with a single winning individual, and these individuals form the junta. On an infinite population the process need not end. At every accepted category, Responsibility names a supporter whose report secures the category on her own (Lemma 4), so each member of the junta is identified directly, whatever the size of the population. Put differently, under Supportiveness Responsibility requires rejection to survive the combination of any number of rejecting profiles, one for each supporter, whereas Rejection Stability requires this only for finitely many (Lemma 5), and the two coincide when NN is finite (Lemma 6).

Concluding remarks

This paper asks how little of the framework of Maniquet and Mongin [38] must be given up to escape their impossibility. We keep their individual domain, in which every individual reports a single category for each object and uses every category, together with Independence and the requirement that the social classification classify every object and leave no category empty. We allow an object to belong to several categories in the social classification and strengthen Unanimity to Supportiveness. In the resolute benchmark, Independence and Unanimity already imply Supportiveness (Proposition 1). With at least three categories, Independence and Unanimity permit a set-valued oligarchy rule that places every disputed object in all categories, including unsupported ones (Proposition 2).

On finite populations, Independence and Supportiveness together with either Responsibility or Rejection Stability characterise the general junta rules J(Mx,r)J^{(M_x,r)} (Theorems 1 and 3). Together with Split Stability they characterise the object-dependent junta rules J(Mx)J^{(M_x)} when m>p≥3m > p \ge3 (Theorem 2), and the neutrality axioms reduce the juntas to a single coalition (Corollaries 1 and 2 and Theorem 3). In each case the juntas are uniquely determined by the rule. Supportiveness can be replaced by Unanimity in the characterisations through Split Stability and Rejection Stability but, when p≥3p \ge3, not in the one through Responsibility (Proposition 4). When outcomes are required to be singletons, every junta shrinks to the same individual, and the dictatorship of Maniquet and Mongin reappears as the resolute special case.

The characterisations describe what multiple memberships make possible; they do not single out a recommended junta. Adding Anonymity leaves only JNJ^N, which accepts every proposed membership (Corollary 3). Under the maintained axioms, a rule that rejects some supported memberships while satisfying Responsibility must therefore give some individuals more authority than others, whereas a rule that is selective and treats individuals equally, such as the quota procedure of Example 4, must give up Responsibility.

Table 2 places these results alongside the decisive structures of preference aggregation. An ultrafilter is exactly a family that is both a filter and a grill (Fact 3). The Arrowian literature on weakened collective rationality retains the filter half: under quasi-transitive social preferences, the decisive coalitions form the principal filter FM\mathcal{F}_M of an oligarchy MM, and each oligarch can veto a strict social preference (Gibbard, [31]). Multiple memberships retain the grill half: for each object and category, the coalitions that secure the membership form the principal grill GMx,r\mathcal{G}_{M_x,r}, and each member of the junta can secure the inclusion of the category she proposes. The two structures are dual (Fact 4), since the coalitions that can block a strict social preference under the oligarchy with oligarchs MM are exactly those that secure a membership under JMJ^M. If the coalitions able to secure a membership formed a filter FM\mathcal{F}_M with ∣M∣≥2|M| \ge2, an object on which two members of MM disagree would receive no category (§ 5.1). Multiple memberships let the powers of different junta members coexist at the same object, and Supportiveness ensures that every membership they create rests on an individual report.

On infinite populations the three axioms come apart. The theorem of Maniquet and Mongin fails there, as Arrow’s theorem does (Fishburn, [28]; Kirman and Sondermann [35]), but only through free ultrafilters, whose existence requires a weak form of the axiom of choice (Corollary 4). Split Stability, and Rejection Stability together with Category Neutrality, characterise the limiting junta rules, whose winning coalitions form grills that need not be principal (Theorem 4). This class contains the Fréchet rule, which assigns a category to an object exactly when infinitely many individuals propose it and is defined without any appeal to choice (Proposition 5).

Notes. In preference aggregation, individual preferences are complete and transitive, the domain is unrestricted, and the alternative set XX contains at least three elements. The filter and grill statements concern families of decisive and quasi-decisive coalitions, respectively, and do not by themselves uniquely determine the aggregation rule. The structures called quasi-ultrafilters by Cato [8] are termed grills here. WP denotes weak Pareto, IIA independence of irrelevant alternatives, and AC the axiom of choice. In classification aggregation, I denotes Independence, SS Split Stability, and CN Category Neutrality. The single-valued classification row is stated for m≥p≥3m \ge p \ge3. The Split Stability characterisations require m>p≥3m > p \ge3, while the Responsibility and Rejection Stability characterisations hold for m≥p≥2m \ge p \ge2. The junta and limiting junta rules satisfy the covering conditions stated in the corresponding results.

Responsibility, by contrast, keeps the grills principal on every population (Corollary 5), because it identifies each member of a junta directly instead of reaching her by splitting coalitions. The axioms thus differ in whether a membership must be secured by an individual supporter or may rest on infinite groups, none of whose finite subgroups secures it on its own.

Equal treatment also fares differently than in Arrowian aggregation. On infinite populations, anonymity is inconsistent with Arrow’s unanimity and independence conditions on sufficiently rich domains ([39]). The same is true of resolute CACs satisfying Independence and Unanimity, whose decisive coalitions form an ultrafilter (§ 8.1), since no ultrafilter on NN is invariant under all permutations of NN. With multiple memberships, by contrast, JNJ^N satisfies Anonymity on every population, and the Fréchet rule does so on every infinite population, because permutations of NN preserve whether a support set is nonempty or infinite.

We close with three directions for future research.

In our model individuals report one category for each object, so multiple memberships can arise only in the social classification. In many applications, however, individuals may themselves wish to place an object in several categories; a voter may judge that a project belongs to Biodiversity and Climate. A natural extension lets individual reports take the same form as social classifications, so that the input and output domains coincide. Supportiveness then requires every collective membership to be proposed by some individual, and an independence condition applied separately to each membership statement, combined with Unanimity, already delivers it. The junta rules remain well defined and are CACs under the same conditions on their juntas. Moreover, since single-valued reports remain admissible, any rule satisfying extensions of our axioms that reduce to them on single-valued profiles must coincide with a junta rule on those profiles. An individual may now support several categories at the same object, so the support sets at an object overlap instead of partitioning the population. Responsibility and Rejection Stability are stated in terms of individual reports and extend almost verbatim, whereas Split Stability, which moves supporters from one category to another, admits several non-equivalent reformulations once an individual can add a category without abandoning another. Whether these axioms still deliver grills on the richer domain, and how the rules they characterise behave at profiles with set-valued reports, remain open.

We have not studied incentives. Under a junta rule, an individual’s report at an object matters only through the category she proposes there, and only if she belongs to the junta for that object and category. She can add that category to the social classification, but she cannot remove a category proposed by another member of the corresponding junta. How this limited influence translates into incentives depends on how individuals evaluate social classifications with multiple memberships, for instance by their proximity to their own reports, so a strategic analysis must first specify such evaluations. Under the Fréchet rule, changing the reports of finitely many individuals never changes whether a support set is infinite, so no finite coalition can affect the outcome, while infinite coalitions can; on infinite populations, the relevant notion of manipulation is therefore coalitional.

We have adopted the membership reading throughout. Under the tie reading, the set assigned to an object lists the categories that remain admissible for it, and a single category is selected later. The natural requirement on outcomes is then that some selection be surjective, which means that, for every set TT of categories, at least ∣T∣|T| objects are assigned a set that meets TT. The oligarchy rule O+MO^{M}_{+} satisfies this requirement, since the classification of any member of MM is a surjective selection, and under this reading it becomes a reasonable counterpart of an oligarchy; junta rules may fail it, as F2F^{2} in Example 2 does. As noted in § 3.1, if the final selection satisfies the requirements of Maniquet and Mongin [38], their impossibility returns at that stage. The question is therefore what the set-valued stage can achieve by itself, and in particular whether the filter-based structures that are unavailable under the membership reading reappear under the tie reading, as they do in preference aggregation. A related question concerns populations modelled by measure spaces (Cato, [9]), in which individuals may be negligible, and in which the contrast between Responsibility, which relies on individual supporters, and the two stability axioms, which only compare coalitions, can be expected to take a different form.

References

Proofs

Proof of Proposition 2

(i) Fix i∈Mi \in M. By definition, ci(x)∈O+M(c)(x)c_i(x) \in O^M_+(c)(x) at every object. Thus outcomes are nonempty, and every category is used because cic_i is surjective. The outcome at xx depends only on the reports at xx, and unanimous support for rr gives {r}\{r\}. Hence Independence and Unanimity hold. The definition is unchanged by relabelling objects or categories, so both neutrality axioms hold.

(ii) If M={d}M = \{d\}, the rule returns {cd(x)}\{c_d(x)\} at every object. If ∣M∣≥2|M| \ge2, Lemma 1 gives a profile where two members disagree at some object. The outcome there is PP, so the rule is not dictatorial.

(iii) If the members of MM agree on rr, both rules return {r}\{r\}. If they disagree and p=2p = 2, both categories have a supporter in MM, so both rules return PP.

(iv) Choose distinct ii, j∈Mj \in M and distinct rr, ss, t∈Pt \in P. By Lemma 1, choose a profile where ii assigns xx to rr, jj assigns it to ss, and everyone assigns it to either rr or ss. The outcome is PP, so it includes the unsupported category tt.

Proof of Proposition 3

(a) Suppose that eqs. (4.1a) and (4.1b) hold. Fix a profile c∈CNc \in C^N. For each xx, choose i∈⋂r∈PMx,ri \in\bigcap_{r \in P} M_{x,r}. Then i∈Mx,ci(x)i \in M_{x,c_i(x)}, so xx receives ci(x)c_i(x). For each rr, choose i∈⋂x∈XMx,ri \in\bigcap_{x \in X} M_{x,r}. Surjectivity gives an object xx with ci(x)=rc_i(x) = r, and i∈Mx,ri \in M_{x,r} ensures that xx receives rr. Thus every object receives a category and every category is used.

(b) Conversely, suppose the rule is a CAC. If eq. (4.1a) fails at xx, each ii can choose rir_i with i∉Mx,rii \notin M_{x,r_i}. By Lemma 1, realise the assignments ci(x)=ric_i(x) = r_i. No category has a supporter in its junta at xx, leaving xx without a category.

(c) Likewise, if eq. (4.1b) fails for rr, each ii can choose yiy_i with i∉Myi,ri \notin M_{y_i,r}. Set ci(yi)=rc_i(y_i) = r and map the remaining objects onto P∖{r}P \setminus\{r\}, possible because m−1≥p−1m - 1 \ge p - 1. Then cic_i is surjective and uses rr only at yiy_i. Every supporter of rr is therefore outside the relevant junta, leaving rr unused.

The special cases follow by substitution.

Proof of Theorem 1

(a) Suppose FF satisfies Independence, Supportiveness and Responsibility. For each xx, rr, let Mx,rM_{x,r} consist of the individuals whose assignment of xx to rr guarantees its acceptance at every profile. By Supportiveness and Lemma 4, every accepted assignment has a supporter in Mx,rM_{x,r}. Conversely, the support of any member of Mx,rM_{x,r} ensures acceptance by definition. Thus r∈F(c)(x)⟺Sr(x,c)∩Mx,r≠∅r \in F(c)(x) \Longleftrightarrow S_r(x,c) \cap M_{x,r} \ne\varnothing. This gives eq. (5), and Proposition 3 gives eqs. (4.1a) and (4.1b).

(b) Uniqueness. By Lemma 1, any individual ii can be the sole supporter of rr at xx. At such a profile, rr is accepted exactly when i∈Mx,ri \in M_{x,r}, which determines every Mx,rM_{x,r} uniquely.

(c) Conversely, let FF have the stated junta representation. The rule is a CAC by Proposition 3 and satisfies Independence and Supportiveness by its definition. For every accepted assignment, choose i∈Sr(x,c)∩Mx,ri \in S_r(x,c) \cap M_{x,r}. Keeping cic_i unchanged preserves this support whatever the others report, so Responsibility holds.

Proof of Corollary 1

(i) By Theorem 1, write F=J(Mx,r)F = J^{(M_{x,r})}. Fix xx, distinct rr, ss, and ii. By Lemma 1, take a profile where ii alone supports rr at xx. Swapping rr and ss makes ii the sole supporter of ss at xx. Category Neutrality therefore gives i∈Mx,r⟺i∈Mx,si \in M_{x,r} \Longleftrightarrow i \in M_{x,s}. Hence Mx,r=MxM_{x,r} = M_x for every rr, and Proposition 3 gives eq. (3). Conversely, these rules satisfy Category Neutrality and the axioms of Theorem 1.

(ii) By (i), write F=J(Mx)F = J^{(M_x)}. Take a profile where ii alone supports rr at xx, and swap objects xx and yy in every report. Now ii alone supports rr at yy, so Object Neutrality gives i∈Mxi \in M_x if and only if i∈Myi \in M_y. Thus all MxM_x equal a common nonempty coalition MM. Conversely, a uniform junta rule also satisfies Object Neutrality.

■\blacksquare

Proof of Theorem 2

(a) Suppose FF satisfies Independence, Supportiveness and Split Stability, and fix xx. By Independence, write fx(cx)=F(c)(x)f_x(c_x) = F(c)(x). By Lemma 2, each opinion vector can be realised with every individual’s report surjective on X∖{x}X \setminus\{x\}. All changes at xx below therefore preserve surjectivity.

Claim 1. Whether k∈fx(π)k \in f_x(\pi) depends only on Sk(π)S_k(\pi).

Proof. Fix S⊆NS \subseteq N and ℓ≠k\ell\ne k. Start with the opinion vector assigning kk to members of SS and ℓ\ell to everyone else. To reach any π\pi with Sk(π)=SS_k(\pi) = S, move each nonempty group Sr(π)S_r(\pi), r∉{k,ℓ}r \notin\{k,\ell\}, from ℓ\ell to rr in turn. Each destination is vacant, so (SS.1) preserves whether kk is accepted throughout.

□\square

Let Wx(k)W_x(k) be the family of support sets that lead to acceptance of kk. Supportiveness and Lemma 3 give ∅∉Wx(k)\varnothing\notin W_x(k) and N∈Wx(k)N \in W_x(k).

Claim 2. The family Wx(k)W_x(k) is the same for every category kk.

Proof. Choose distinct kk, k′k', ℓ\ell and a nonempty S⊆NS \subseteq N. Let members of SS support kk and everyone else support ℓ\ell. Move all of SS from kk to the vacant category k′k'. By (SS.2), kk was accepted exactly when kk or k′k' is now accepted. Supportiveness excludes the now unsupported kk, so S∈Wx(k)S \in W_x(k) if and only if S∈Wx(k′)S \in W_x(k'). The empty set belongs to neither family.

□\square

Writing WxW_x for this common family, we have

F(c)(x)={k∈P∣Sk(x,c)∈Wx}.(7)F(c)(x) = \{k \in P \mid S_k(x,c) \in W_x\}. \tag*{(7)}

Claim 3. For disjoint A,B⊆NA, B \subseteq N, A∪B∈WxA \cup B \in W_x if and only if A∈WxA \in W_x or B∈WxB \in W_x.

Proof. If either set is empty, use ∅∉Wx\varnothing\notin W_x. Otherwise, choose distinct kk, k′k', ℓ\ell. Let A∪BA \cup B support kk and everyone else support ℓ\ell, then move BB to the vacant category k′k'. The two support sets become AA and BB, so (SS.2) gives the claim.

□\square

The claim implies upward closure: if A∈WxA \in W_x and A⊆TA \subseteq T, apply it to T=A∪(T∖A)T = A \cup(T \setminus A). It also gives the grill property: if A∪B∈WxA \cup B \in W_x, apply it to A∪(B∖A)A \cup(B \setminus A) and use upward closure. Thus WxW_x is a grill. Since NN is finite, Fact 1 gives a unique nonempty MxM_x with Wx=GMxW_x = \mathcal{G}_{M_x}. Equation (A.1) then gives F=J(Mx)F = J^{(M_x)}, and Proposition 3 gives eq. (3).

(b) Conversely, let F=J(Mx)F = J^{(M_x)} satisfy eq. (3). It is a CAC by Proposition 3; its definition gives Independence and Supportiveness. Consider an admissible defection of B⊆Sk(x,c)B \subseteq S_k(x,c) to a vacant category k′k'. Other support sets are unchanged, giving (SS.1). The support of kk splits into Sk(x,c)∖BS_k(x,c) \setminus B and BB. At least one part meets MxM_x exactly when the original set does, giving (SS.2).

■\blacksquare

Proof of Corollary 2. By Theorem 2, write F=J(Mx)F = J(M_x). Swapping objects as in the proof of Corollary 1(ii) shows that Object Neutrality makes all MxM_x equal. Conversely, every uniform junta rule satisfies Object Neutrality and the axioms of Theorem 2.

Proof of Theorem 3. On finite populations, Responsibility and Rejection Stability are equivalent under Supportiveness by Lemma 6. Parts (i) and (ii) therefore follow from Theorem 1 and Corollary 1, respectively.

Proof of Proposition 4. In each case, we show that a category with no supporters is rejected. Fix c,x,r\mathbf{c},x,r with Sr(x,c)=∅S_r(x,\mathbf{c}) = \varnothing.

(i) Let k1,…,kqk_1,\ldots,k_q be the categories used at xx in c\mathbf{c}. By Lemma 2, start from a profile where everyone assigns xx to k1k_1 and every report is surjective on X∖{x}X \setminus\{x\}. Unanimity rejects rr. For j=2,…,qj = 2,\ldots,q, move Skj(x,c)S_{k_j}(x,\mathbf{c}) from k1k_1 to the vacant category kjk_j. Every move is admissible, and (SS.1) preserves rejection of rr because r∉{k1,kj}r \notin\{k_1,k_j\}. The resulting opinion vector is cx\mathbf{c}_x, so Independence gives r∉F(c)(x)r \notin F(\mathbf{c})(x).

(ii) By Lemma 1, take a profile d\mathbf{d} where everyone assigns xx to some s≠rs \ne r. Unanimity rejects rr there. Apply Rejection Stability with both source profiles equal to d\mathbf{d} and target c\mathbf{c}. Its report condition is vacuous because c\mathbf{c} has no supporter of rr at xx, so r∉F(c)(x)r \notin F(\mathbf{c})(x).

Proof of Theorem 4.

Lemma 7. Let NN be finite or infinite, and let FF satisfy Independence, Supportiveness, and either Split Stability (with m>p≥3m > p \ge3) or Rejection Stability (with m≥p≥2m \ge p \ge2). Then, for every x∈Xx \in X and r∈Pr \in P:

(i) every A⊆NA \subseteq N is either winning or null for rr at xx;

(ii) Wx,rW_{x,r} is a grill.

Under Split Stability, or under Rejection Stability and Category Neutrality, Wx,r=WxW_{x,r} = W_x does not depend on rr; under Object Neutrality as well, it does not depend on xx.

Proof.

(a) Split Stability. The three claims in the proof of Theorem 2 and the argument following them give the result. Those steps apply to infinite populations as well: the number of defections is finite because PP is finite.

(b) Rejection Stability. Fix x,rx,r. Suppose d\mathbf{d} rejects rr with support set AA, and take any profile e\mathbf{e} with the same support set. Form c\mathbf{c} by taking eie_i for i∈Ai \in A and did_i for everyone else. Since cx=dx\mathbf{c}_x = \mathbf{d}_x, Independence makes c\mathbf{c} reject rr. Every supporter in e\mathbf{e} retains her report from c\mathbf{c}, so Rejection Stability, with both source profiles equal to c\mathbf{c}, makes e\mathbf{e} reject rr. Thus every coalition that is not winning is null.

Let Ix,r\mathcal{I}_{x,r} be the family of null coalitions. Take A,B∈Ix,rA,B \in\mathcal{I}_{x,r} and any profile e\mathbf{e} with support set S⊆A∪BS \subseteq A \cup B. Choose profiles c,d\mathbf{c},\mathbf{d} with support sets A,BA,B, retaining eie_i in c\mathbf{c} for i∈S∩Ai \in S \cap A and in d\mathbf{d} for i∈S∩Bi \in S \cap B. Complete the other reports using Lemma 1. Both profiles reject rr, and each supporter in e\mathbf{e} retains her report from at least one of them. Rejection Stability therefore makes e\mathbf{e} reject rr, so SS is null. Hence Ix,r\mathcal{I}_{x,r} is downward closed and closed under union. Supportiveness and Lemma 3 give ∅∈Ix,r\varnothing\in\mathcal{I}_{x,r} and N∉Ix,rN \notin\mathcal{I}_{x,r}. Its complement Wx,rW_{x,r} is therefore a grill.

(c) Neutrality. Swapping rr and ss turns each support set for rr into the same support set for ss, so Category Neutrality gives Wx,r=Wx,sW_{x,r} = W_{x,s}. Likewise, swapping xx and yy and applying Object Neutrality gives Wx,r=Wy,rW_{x,r} = W_{y,r}.

Lemma 8. Let (Mx)x∈X(\mathcal{M}_x)_{x\in X} be a family of directed families, and write G(Mx)={S⊆N∣S∩M≠∅\mathcal{G}(\mathcal{M}_x)=\{S\subseteq N\mid S\cap M\ne\varnothing for every M∈Mx}M\in\mathcal{M}_x\}.

(i) Each G(Mx)\mathcal{G}(\mathcal{M}_x) is a grill, and r∈L(Mx)(c)(x)r\in L^{(\mathcal{M}_x)}(\mathbf{c})(x) if and only if Sr(x,c)∈G(Mx)S_r(x,\mathbf{c})\in\mathcal{G}(\mathcal{M}_x).

(ii) Conversely, for every grill G\mathcal{G}, its family of blocking coalitions G∗\mathcal{G}^{*} is directed and G=G(G∗)\mathcal{G}=\mathcal{G}(\mathcal{G}^{*}).

(iii) L(Mx)L^{(\mathcal{M}_x)} is a CAC if and only if eq. (6) holds. This condition holds automatically for uniform limiting junta rules.

Proof.

(i) Since Mx\mathcal{M}_x is a nonempty family of nonempty coalitions, N∈G(Mx)N\in\mathcal{G}(\mathcal{M}_x) and ∅∉G(Mx)\varnothing\notin\mathcal{G}(\mathcal{M}_x). Upward closure follows from the definition. If A,B∉G(Mx)A,B\notin\mathcal{G}(\mathcal{M}_x), choose M1,M2∈MxM_1,M_2\in\mathcal{M}_x disjoint from A,BA,B, respectively. Directedness gives M3∈MxM_3\in\mathcal{M}_x with M3⊆M1∩M2M_3\subseteq M_1\cap M_2. Then M3M_3 is disjoint from A∪BA\cup B, proving the grill property. The representation of L(Mx)L^{(\mathcal{M}_x)} is its definition.

(ii) By Fact 4, G∗\mathcal{G}^{*} is a filter, hence directed, and G=G(G∗)\mathcal{G}=\mathcal{G}(\mathcal{G}^{*}).

(iii) At each object, the finitely many support sets partition NN. The grill property therefore makes at least one of them winning, so the outcome is nonempty. It remains to check that every category is used. Suppose eq. (6) holds and rr is unused. For each xx, choose Mx∈MxM_x\in\mathcal{M}_x disjoint from Sr(x,c)S_r(x,\mathbf{c}). Any ⋂xMx\bigcap_x M_x would then assign no object to rr, contradicting surjectivity.

Conversely, suppose Mx∈MxM_x\in\mathcal{M}_x for each xx and ⋂xMx=∅\bigcap_x M_x=\varnothing. Fix rr and, for each ii, choose yiy_i with i∉Myii\notin M_{y_i}. Set ci(yi)=rc_i(y_i)=r and map the remaining objects onto P∖{r}P\setminus\{r\}, possible because m−1≥p−1m-1\ge p-1. Thus Sr(x,c)∩Mx=∅S_r(x,\mathbf{c})\cap M_x=\varnothing at every xx, leaving rr unused.

For a uniform rule, directedness and finiteness of XX ensure that every intersection ⋂xMx\bigcap_x M_x contains a member of M\mathcal{M}. That member is nonempty, so eq. (6) holds.

Proof of Theorem 4.

(a) Suppose FF satisfies either set of axioms. By Lemma 7, acceptance at xx is determined by a grill WxW_x common to all categories. Set Mx=Wx∗\mathcal{M}_x=W_x^{*}. By Lemma 8(ii), Mx\mathcal{M}_x is directed and Wx=G(Mx)W_x=\mathcal{G}(\mathcal{M}_x), so F=L(Mx)F=L^{(\mathcal{M}_x)}. Part (iii) gives eq. (6) because FF is a CAC. Under Object Neutrality, Lemma 7 makes WxW_x, and hence Mx\mathcal{M}_x, the same for every object.

(b) Conversely, let F=L(Mx)F=L^{(\mathcal{M}_x)} satisfy eq. (6) and write Gx=G(Mx)\mathcal{G}_x=\mathcal{G}(\mathcal{M}_x). By Lemma 8, FF is a CAC with r∈F(c)(x)⟺Sr(x,c)∈Gxr\in F(\mathbf{c})(x)\Longleftrightarrow S_r(x,\mathbf{c})\in\mathcal{G}_x. This gives Independence and Category Neutrality; ∅∉Gx\varnothing\notin\mathcal{G}_x gives Supportiveness.

For Split Stability, an admissible defection to a vacant category leaves all other support sets unchanged, giving (SS.1). It splits the original support set into two parts. The grill property and upward closure say that the original set is winning exactly when at least one part is winning, giving (SS.2).

For Rejection Stability, let c,dc,d reject rr at xx and let ee satisfy its report condition. Then Sr(x,e)⊆Sr(x,c)∪Sr(x,d)S_r(x,e)\subseteq S_r(x,c)\cup S_r(x,d). The union is losing by the grill property, and every subset of a losing set is losing by upward closure. Thus ee rejects rr at xx. Uniform limiting junta rules also satisfy Object Neutrality.

Proof of Proposition 5

(i) If MxM_x is the smallest member of Mx\mathcal{M}_x, a coalition meets every member exactly when it meets MxM_x. Hence L(Mx)=J(Mx)L^{(\mathcal{M}_x)}=J^{(M_x)}.

Conversely, suppose L(Mx)=J(Mx′)L^{(\mathcal{M}_x)}=J^{(M'_x)}. By Lemma 1, every coalition can occur as a support set at xx, so G(Mx)=GMx′\mathcal{G}(\mathcal{M}_x)=\mathcal{G}_{M'_x}. For each i∈Mx′i\in M'_x, the singleton {i}\{i\} is winning and therefore meets every member of Mx\mathcal{M}_x. Thus Mx′M'_x is contained in every member. Since N∖Mx′N\setminus M'_x is losing, some M∈MxM\in\mathcal{M}_x is disjoint from it, so M⊆Mx′M\subseteq M'_x. Hence M=Mx′M=M'_x is the smallest member. (ii) Since NN is infinite, cofinite coalitions are nonempty and closed under finite intersections, hence directed. A finite set SS misses the cofinite set N∖SN \setminus S, while an infinite set meets every cofinite set. Thus a category is accepted exactly when its support set is infinite. Suppose this rule were J(Mx,r)J^{(M_{x,r})}. Fix x,rx,r. Unanimous support is accepted, so Mx,r≠∅M_{x,r} \ne\varnothing. For i∈Mx,ri \in M_{x,r}, Lemma 1 gives a profile where ii alone supports rr at xx. The junta rule accepts rr there, whereas the limiting rule rejects it because its support set is finite. This is a contradiction.

Independence of the axioms

The following CACs show that no axiom can be dropped from the main characterisations. We assume that NN has at least two members, and fix distinct categories a,b∈Pa,b \in P and an object x0∈Xx_0 \in X where needed.

Characterisationw/o Independencew/o Supportivenessw/o third axiomw/o neutrality
Theorem 1 (Resp.)E2E_2E3E_3E4E_4–
Corollary 1(i) (Resp., CN)E1E_1E3E_3E5E_5E6E_6
Theorem 2 (SS)E1E_1E3E_3E4E_4–
Corollary 2 (SS, ON)E1E_1E3E_3E5E_5E7E_7
Theorem 3(i) (Rej. Stab.)E2E_2E3E_3E4E_4–
Theorem 3(ii) (Rej. Stab., CN)E1E_1E3E_3E5E_5E6E_6

Table 3. Examples showing the independence of the axioms. The examples E1E_1, E2E_2, and E5E_5 require m>pm > p, m≥3m \ge3 or p≥3p \ge3, and n≥3n \ge3, respectively.

E1E_1. (Requires m>pm > p.) E1(c)(x)={c1(x)}∪{c2(x)∣c2(y)=c2(x) for some y≠x}E_1(\boldsymbol{c})(x) = \{c_1(x)\} \cup\{c_2(x) \mid c_2(y) = c_2(x)\text{ for some }y \ne x\}. It satisfies Supportiveness, Responsibility, Rejection Stability, Split Stability and both neutrality axioms, but not Independence (§ 6.3). This is the rule F3F^3 of Example 3. For Split Stability, note that in an admissible defection individual 2 can defect from kk only if kk appears elsewhere in c2c_2, and the vacant category k′k' appears elsewhere in c2c_2 because c2c_2 is surjective; hence individual 2 is included before the defection if and only if she is included after it, and the outcome changes only within {k,k′}\{k,k'\}.

E2E_2. (Requires m≥3m \ge3 or p≥3p \ge3.) For fixed y0∈Xy_0 \in X and r0∈Pr_0 \in P, E2(c)(x)={c1(x)}∪{c2(x)∣c2(y0)=r0}E_2(\boldsymbol{c})(x) = \{c_1(x)\} \cup\{c_2(x) \mid c_2(y_0) = r_0\}. It satisfies Supportiveness, Responsibility and Rejection Stability, because each accepted category is secured by the individual classification of individual 1 or 2, but not Independence.

E3E_3. The constant CAC E3(c)(x)=PE_3(\boldsymbol{c})(x) = P satisfies Independence, Split Stability, Responsibility, Rejection Stability and both neutrality axioms, but not Supportiveness.

E4E_4. E4(c)(x)={ci(x)∣i∈N}E_4(\boldsymbol{c})(x) = \{c_i(x) \mid i \in N\} for x≠x0x \ne x_0, and E4(c)(x0)={b}E_4(\boldsymbol{c})(x_0) = \{b\} if every individual assigns x0x_0 to bb and E4(c)(x0)={ci(x0)∣i∈N}∖{b}E_4(\boldsymbol{c})(x_0) = \{c_i(x_0) \mid i \in N\} \setminus\{b\} otherwise. It is a CAC: category bb is used at some object other than x0x_0 unless every individual assigns bb only to x0x_0, in which case the profile is unanimous on bb at x0x_0. It satisfies Independence and Supportiveness. It violates Responsibility at a profile that is unanimous on bb at x0x_0, hence also Rejection Stability by Lemma 6, and it violates Split Stability when p≥3p \ge3: if some individuals assign x0x_0 to aa and the others to bb, letting all supporters of bb defect to a vacant category k′k' changes the outcome from {a}\{a\} to {a,k′}\{a,k'\}, contrary to (SS.2).

E5E_5. (Requires n≥3n \ge3.) E5(c)(x)={c1(x)}∪{r∈P∣∣Sr(x,c)∣≥2}E_5(\boldsymbol{c})(x) = \{c_1(x)\} \cup\{r \in P \mid|S_r(x,\boldsymbol{c})| \ge2\}. It satisfies Independence, Supportiveness and both neutrality axioms, but not Responsibility, and hence not Rejection Stability.

  • E6E_6. J(Mx,r)J(M_{x,r}) with Mx,a={1,2}M_{x,a} = \{1, 2\} and Mx,r={1}M_{x,r} = \{1\} for r≠ar \ne a satisfies Independence, Supportiveness, Responsibility and Rejection Stability, but not Category Neutrality.

  • E7E_7. J(Mx)J(M_x) with Mx0={1,2}M_{x_0} = \{1, 2\} and Mx={1}M_x = \{1\} for x≠x0x \ne x_0 satisfies all axioms of Theorem 2 and Corollary 1 except Object Neutrality.

References

  1. [1]Alcantud, J. C. R., Díaz, S., and Montes, S. (2019). Liberalism and dictatorship in the problem of fuzzy classification. International Journal of Approximate Reasoning, 110:82–95. [5]DOI
  2. [2]Arrow, K. J. (1963). Social Choice and Individual Values. John Wiley & Sons, New York, 2nd edition. [2]
  3. [3]Bossert, W. and Cato, S. (2020). Acyclicity, anonymity, and prefilters. Journal of Mathematical Economics, 87:134–141. [4]DOI
  4. [4]Bossert, W. and Suzumura, K. (2009). Decisive coalitions and coherence properties. PIE/CIS Discussion Paper 427, Center for Intergenerational Studies, Institute of Economic Research, Hitotsubashi University. [5, 24]
  5. [5]Brown, D. J. (1975). Aggregation of preferences. Quarterly Journal of Economics, 89(3):456–469. [4]DOI
  6. [6]Cailloux, O., Hervouin, M., Özkes, A. I., and Sanver, M. R. (2024). Classification aggregation without unanimity. Mathematical Social Sciences, 128:6–9. [4, 11]arxiv.org/abs/2310.11841
  7. [7]Cailloux, O., Hervouin, M., Özkes, A. I., and Sanver, M. R. (2025). Anonymous and neutral classification aggregation. Working paper, Université Paris Dauphine–PSL. [4, 20]DOI
  8. [8]Cato, S. (2013). Quasi-decisiveness, quasi-ultrafilter, and social quasi-orderings. Social Choice and Welfare, 41(1):169–202. [5, 9, 24]DOI
  9. [9]Cato, S. (2021). Preference aggregation and atoms in measures. Journal of Mathematical Economics, 94:102446. [6, 25]DOI
  10. [10]Cato, S. (2022). Stable preference aggregation with infinite population. Social Choice and Welfare, 59(2):287–304. [6]DOI
  11. [11]Cato, S., Gonzalez, S., Réмila, E., and Solal, P. (2026). Filters in abstract aggregation with infinite sequences of signals. Mathematical Social Sciences, 142:102551. [5]DOI
  12. [12]Cho, W. J. and Ju, B.-G. (2017). Multinary group identification. Theoretical Economics, 12(2):513–531. [5]DOI
  13. [13]Cho, W. J. and Park, C. W. (2018). Fractional group identification. Journal of Mathematical Economics, 77:66–75. [5]DOI
  14. [14]Choquet, G. (1947). Sur les notions de filtre et de grille. Comptes rendus de l’Académie des sciences, Paris, 224:171–173. [4, 5, 10]
  15. [15]Craven, J. (2023). Domain restrictions in the aggregation of classifications. Global Philosophy, 33:12. [4, 11]DOI
  16. [16]Craven, J. (2024). Aggregation of ranked categories. Mathematical Social Sciences, 129:27–33. [4]DOI
  17. [17]Dietrich, F. and List, C. (2007). Judgment aggregation by quota rules: Majority voting generalized. Journal of Theoretical Politics, 19(4):391–424. [5]
  18. [18]Dietrich, F. and List, C. (2008). Judgment aggregation without full rationality. Social Choice and Welfare, 31(1):15–39. [5, 12]DOI
  19. [19]Dimitrov, D., Marchant, T., and Mishra, D. (2012). Separability and aggregation of equivalence relations. Economic Theory, 51(1):191–212. [4]DOI
  20. [20]Dokow, E. and Holzman, R. (2010a). Aggregation of binary evaluations with abstentions. Journal of Economic Theory, 145(2):544–561. [5]DOI
  21. [21]Dokow, E. and Holzman, R. (2010b). Aggregation of non-binary evaluations. Advances in Applied Mathematics, 45(4):487–504. [4, 5, 11]DOI
  22. [22]Endriss, U. and Grandi, U. (2017). Graph aggregation. Artificial Intelligence, 245:86–114. [5, 11]
  23. [23]Filmus, Y. (2026). Classification aggregation: A quantitative impossibility theorem. arXiv preprint arXiv:2605.17136. [4]arxiv.org/abs/2605.17136
  24. [24]Fioravanti, F. (2025a). Expert classification aggregation. Mathematical Social Sciences, 137:102406. [3, 4, 15, 16]
  25. [25]Fioravanti, F. (2025b). Fuzzy classification aggregation. Mathematical Social Sciences, 135:102422. [5, 11]
  26. [26]Fioravanti, F. and Tohmé, F. (2020). Asking infinite voters ‘Who is a J?’: Group identification problems in ℕ. Journal of Classification, 37(1):58–65. [6]
  27. [27]Fioravanti, F. and Tohmé, F. (2022). Fuzzy group identification problems. Fuzzy Sets and Systems, 434:159–171. [5]DOI
  28. [28]Fishburn, P. C. (1970). Arrow’s impossibility theorem: Concise proof and infinite voters. Journal of Economic Theory, 2(1):103–106. [4, 6, 21, 23, 24]DOI
  29. [29]Fishburn, P. C. and Rubinstein, A. (1986). Aggregation of equivalence relations. Journal of Classification, 3:61–65. [4]DOI
  30. [30]Gibaja, E. and Ventura, S. (2015). A tutorial on multilabel learning. ACM Computing Surveys (CSUR), 47(3):1–38. [2]DOI
  31. [31]Gibbard, A. F. (2014). Intransitive social indifference and the Arrow dilemma. Review of Economic Design, 18(1):3–10. [2, 4, 12, 23, 24]DOI
  32. [32]Hansson, B. (1976). The existence of group preference functions. Public Choice, 28(1):89–98. [5, 24]DOI
  33. [33]Hervouin, M. (2025). Advances in Classification Aggregation and Proportionality. PhD thesis, Université Paris Dauphine–PSL. [2, 4]
  34. [34]Kasher, A. and Rubinstein, A. (1997). On the question “Who is a J?”: A social choice approach. Logique et Analyse, 40(160):385–395. [4, 5]
  35. [35]Kirman, A. P. and Sondermann, D. (1972). Arrow’s theorem, many agents, and invisible dictators. Journal of Economic Theory, 5(2):267–277. [4, 5, 6, 21, 23, 24]
  36. [36]Kurata, L. and Nakamura, K. (2025). Reservation of judgment and robust collective decisions. arXiv preprint arXiv:2512.16240. [11]arxiv.org/abs/2512.16240
  37. [37]Kurata, L. and Nakamura, K. (2026). Agreement with reservation of judgment under risk. arXiv preprint arXiv:2601.01334. [11]arxiv.org/abs/2601.01334
  38. [38]Maniquet, F. and Mongin, P. (2016). A theorem on aggregating classifications. Mathematical Social Sciences, 79:6–10. [2, 3, 4, 5, 6, 7, 10, 11, 13, 14, 19, 20, 21, 23, 24, 25]DOI
  39. [39]Mihara, H. R. (1997a). Anonymity and neutrality in Arrow’s theorem with restricted coalition algebras. Social Choice and Welfare, 14(4):503–512. [6, 24]DOI
  40. [40]Mihara, H. R. (1997b). Arrow’s theorem and Turing computability. Economic Theory, 10(2):257–276. [6]DOI
  41. [41]Miller, A. D. (2008). Group identification. Games and Economic Behavior, 63(1):188–202. [5]DOI
  42. [42]Mirkin, B. G. (1975). On the problem of reconciling partitions. In Blalock, H. M., Aganbegian, A., Borodkin, F. M., Boudon, R., and Capecchi, V., editors, Quantitative Sociology: International Perspectives on Mathematical and Statistical Modeling, pages 441–449. Academic Press, New York. [4]DOI
  43. [43]Samet, D. and Schmeidler, D. (2003). Between liberalism and democracy. Journal of Economic Theory, 110(2):213–233. [5]DOI
  44. [44]Sen, A. K. (1970). Collective Choice and Social Welfare, volume 5 of Mathematical Economics Texts. Holden-Day, San Francisco. [8]
  45. [45]Sen, A. K. (1995). Rationality and social choice. The American Economic Review, 85(1):1–24. [8]
  46. [46]Sung, S. C. and Dimitrov, D. (2005). On the axiomatic characterization of “Who is a J?”. Logique et Analyse, 48(189–192):101–112. [5]
  47. [47]Szegedy, M. and Xu, Y. (2015). Impossibility theorems and the universal algebraic toolkit. arXiv preprint arXiv:1506.01315. [5]arxiv.org/abs/1506.01315
  48. [48]Torres, R. (2005). Limiting dictatorial rules. Journal of Mathematical Economics, 41(7):913–935. [6, 21]DOI
  49. [49]Weymark, J. A. (1984). Arrow’s theorem with social quasi-orderings. Public Choice, 42(3):235–246. [2, 4, 24]DOI

Paper details

vlnc:W_qrjb4ctk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Claude, ChatGPT

View statement
Looking for related papers…
Contents