Three-Dimensional Popular Matching with Cyclic Preferences

· AAMAS 2022 (aamas22-00038)

mirror found
paperThree-Dimensional Popular Matching with Cyclic Preferences
authors
venueAAMAS 2022
filed undercoalition · matching
judged bygpt-5.6-luna / xhigh (triple__luna__xhigh__c22r1)
judge confidencehigh
authors would recognise ityes

The anchor — hardness

Theorem 4.2

It is NP-complete to decide whether a given 3dpmvi instance with a matching 𝑀 admits a matching that is more popular than 𝑀. This holds even if the preferences are derived from a 1-master list.

statement extracted from the paper’s text layer

Every anchor argued

The continuous mirror question

Given finite type sets \(\mathcal A,\mathcal B,\mathcal C\), rational masses \(\mu^A,\mu^B,\mu^C\), cyclic acceptability and preference rankings over partner types, and a finite cell-decomposed reference mass matching \(M\), does there exist an acceptable fractional mass matching \(M'\) satisfying the type-mass capacity constraints and \(\Delta(M',M)>0\), where \(\Delta\) is the total mass-weighted head-to-head vote margin?

The model it lives in

A finite three-way transportation LP with variables \(y_{pqr}\ge0\) assigning mass from reference cells \(p,q,r\) to acceptable alternative triples, capacity constraints for every cell, and objective maximizing the linear vote margin \(\Delta(M',M)\).

The objection that survived

The mirror replaces strict rankings over named agents with rankings over partner types, effectively introducing indifference among copies and changing the literal preference representation.

fatal: False

What the mirror covers

The mirror covers only Theorem 4.2; it leaves Theorems 4.1 and 4.3–4.6, the structural lemmas, and the master-list results unaddressed.

Open questions for a prover

The case FOR (proponent)

The positive case I would actually defend is a mirror of the paper’s verification problem, anchored on Theorem 4.2, proved in this paper:

It is NP-complete to decide whether a given \(3dpmvi\) instance with a matching \(M\) admits a matching more popular than \(M\), even when preferences are derived from a \(1\)-master list.

I would use no second anchor. Theorem 4.2 is especially valuable because its continuous version has a clear interpretation and a credible algorithmic fate.

Consider a large three-sided service market: users, data sources, and servers, as in the paper’s networking applications. There may be millions of agents in each class but only finitely many operational types: service profile, compatibility requirements, priority template, and so on. A type is a complete preference and acceptability profile over partner types. Let the three type sets be \(\mathcal A,\mathcal B,\mathcal C\), with rational masses \(\mu^A,\mu^B,\mu^C\), each summing to \(1\). Preferences remain cyclic: \(A\)-types rank \(B\)-types, \(B\)-types rank \(C\)-types, and \(C\)-types rank \(A\)-types.

A mass matching is a measurable assignment of population mass to acceptable triples. A matching may split a type across several partner types, since a nonatomic population can be partitioned into subcohorts. This is not a lottery: every infinitesimal agent receives one definite triple.

The precise continuous problem is:

Mass-\(3d\)-Cyclic Popularity Verification. The input consists of the type sets, masses, cyclic preference lists, and a reference mass matching \(M\). The matching \(M\) is represented by finitely many current-partner cells. For example, an \(A\)-cell \(p\) has mass \(r_p\), preference type \(a(p)\), and current \(B\)-partner type \(b^0(p)\), or \(\bot\) if unmatched. Analogous cells \(q\) and \(r\) are supplied for \(B\) and \(C\).

The question is whether there exists another mass matching \(M'\) such that the total mass preferring \(M'\) exceeds the total mass preferring \(M\).

For a type \(x\), define

\[ v_x(z,z_0)= \begin{cases} 1 & \text{if } z\succ_x z_0,\\ 0 & \text{if } z=z_0,\\ -1 & \text{if } z_0\succ_x z. \end{cases} \]

Let \(y_{pqr}\) be the mass of alternative triples formed from \(A\)-cell \(p\), \(B\)-cell \(q\), and \(C\)-cell \(r\). The variables satisfy the mass constraints

\[ \sum_{q,r}y_{pqr}\le r_p,\qquad \sum_{p,r}y_{pqr}\le r_q,\qquad \sum_{p,q}y_{pqr}\le r_r, \]

and \(y_{pqr}=0\) whenever the corresponding type triple is unacceptable.

The vote margin is the linear expression

\[ \Delta(M',M) = K(M)+ \sum_{p,q,r} y_{pqr} \left[ \begin{aligned} &v_{a(p)}(b(q),b^0(p))-v_{a(p)}(\bot,b^0(p))\\ +{}&v_{b(q)}(c(r),c^0(q))-v_{b(q)}(\bot,c^0(q))\\ +{}&v_{c(r)}(a(p),a^0(r))-v_{c(r)}(\bot,a^0(r)) \end{aligned} \right], \]

where \(K(M)\) accounts for agents left unmatched in \(M'\). The answer is YES precisely when the maximum of \(\Delta(M',M)\) is positive.

This is recognisably the paper’s question: same three populations, same cyclic acceptability, same partner-based votes, and the same head-to-head notion of popularity. The only change is that unit agents become divisible mass. In a discrete election with \(N\) agents, a winning margin of \(k\) becomes a mass margin \(k/N>0\).

I expect this continuous problem to be Class A. Once the reference matching is fixed, every agent’s current partner is known. The best defeating alternative is therefore a finite packing/transportation LP with linear objective and linear mass constraints. The number of variables is polynomial in the number of supplied type cells, and exact rational LP methods decide whether the optimum is positive. Thus the NP-complete discrete search for a defeating integral 3-dimensional matching plausibly dissolves into fractional mass transport.

The high-multiplicity regime is credible rather than ornamental. A recurring batch of network requests, data sources, and server slots can contain millions of agents with a small catalogue of repeated service profiles. The paper itself points to networking and resource-allocation applications, and its master-list restrictions strengthen this interpretation: common priority templates are natural in pooled service systems. The relevant object is a fraction of a batch receiving each service pattern, not the identity of the \(437{,}291\)-st request.

This mirror deliberately covers only Theorem 4.2. I would not yet claim that the paper’s existence result, Theorem 4.1, has the same fate. There the reference matching is itself variable, so one must choose \(M\) while ensuring that no alternative mass matching defeats it. That introduces a nontrivial \(\exists M\,\forall M'\) condition, and the resulting problem could be tractable, continuum-specifically hard, or retain discrete hardness. It is an excellent follow-up question, but not needed to establish a credible mirror.

The weakest point is that popularity compares agents with their own current partners. If one retained only aggregate type-to-type mass and discarded the reference-partner cells, popularity would become underdefined once a type is split across several partners. The cell refinement is therefore essential. An opponent could argue that it enriches the state beyond the paper’s bare population profile. My answer is that it records the supplied matching \(M\), exactly as Theorem 4.2 does; it does not introduce idiosyncratic prices or named-agent structure. Agents in the same preference type and current-partner cell remain completely interchangeable.

The natural next questions are whether Theorem 4.1’s existence problem admits a polynomial formulation, whether discrete popularity can be recovered from the LP with useful additive guarantees, and whether hardness returns when mass splitting is forbidden or when preference types are given only through a succinct pricing oracle.

The case AGAINST (opponent, writing after the proponent)

I cannot honestly defeat the proponent’s anchor. The negative case is weak here.

Theorem 4.2 is a genuine computational result, and its mass analogue is well defined. The apparent identity-dependence of popularity can be retained by refining the population into cells indexed by preference type and current partner type. Those cells are part of the given matching \(M\), not an illicit introduction of named-agent data. An alternative mass matching is then a fractional three-way matching, and the question whether it achieves positive vote margin is exactly a finite LP.

The strongest objection is that the paper’s original preferences are strict rankings over named agents. If one insists on preserving that literally, then repeated agents cannot generally be indistinguishable: their rankings of counterpart identities encode different types. The proposed repair instead ranks partner types, effectively introducing indifference among copies. That is a genuine modelling change.

It does not, however, defeat the better mirror. In a recurring service market, users, data sources, and servers can naturally be grouped by operational profiles, compatibility classes, and priority templates. Their preferences can be over partner types, with mass split across several partner types. The paper’s networking applications and master-list restrictions make this scenario credible. Current-partner cells also remain finite and meaningful in such a market.

Nor can one object that the paper’s particular hardness gadgets use many individualized agents. The programme explicitly allows a different high-multiplicity scenario for the same mathematical problem. The fact that the repaired version becomes a straightforward LP is an algorithmic outcome, not a defect in posing it.

Thus the only defensible negative claim is narrow: the paper itself does not establish a naturally repeated-type regime, and a literal strict-ranking mirror is problematic. It cannot support the requested universal conclusion. Theorem 4.2 supplies a worthwhile continuous mirror unless the programme excludes type-level preferences and fractional reassignment by fiat.

The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.