Exact Algorithms and Lower Bounds for Forming Coalitions of Constrained Maximum Size

Foivos Fioravantes, Harmender Gahlawat, Nikolaos Melissinos · AAAI 2025 (aaai25-33514)

mirror found
paperExact Algorithms and Lower Bounds for Forming Coalitions of Constrained Maximum Size
authorsFoivos Fioravantes, Harmender Gahlawat, Nikolaos Melissinos
venueAAAI 2025
filed undercoalition · hedonic
judged bygpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1)
judge confidencehigh
authors would recognise ityes

The anchor — algorithmic

Theorem 3

C-CFw can be solved in time vcO(vc)nO(1). Next, we prove that both of the above algorithms are asymptotically optimal, i.e., we do not expect a drastic im- provement in their running times.

statement extracted from the paper’s text layer

Every anchor argued

The continuous mirror question

Given a finite type set \(T\), rational masses \(\mu_t\) summing to one, symmetric pair utilities \(w_{tu}\), and coalition capacity \(C\), choose nonnegative densities \(y_q\) for every integer type composition \(q\) with \(1 \le |q| \le C\), satisfying \(\sum_q q_t y_q = \mu_t\), to maximize \(\sum_q v(q)y_q\), where \(v(q) = \sum_{t<u} w_{tu} q_t q_u + \sum_t w_{tt} \binom{q_t}{2}\); decide whether the optimum is at least a rational threshold \(B\).

The model it lives in

A high-multiplicity additive hedonic coalition model: types are complete compatibility profiles, \(\mu\) gives population mass, \(q\) gives an integer team composition, \(y_q\) gives team density, and welfare is total normalized pair utility. The natural structural extension uses a vertex cover of the type-interaction graph, not the paper's individual-agent vertex cover.

The objection that survived

The fixed-C configuration LP is tractable without the paper's matching argument, and a type-interaction vertex cover is not the individual-agent vertex cover of Theorem 3, so the theorem-level parameterized transfer remains unestablished.

fatal: False

What the mirror covers

The mirror covers weighted bounded-coalition welfare maximization and a fixed-C analogue of Theorem 3; it leaves the unweighted hardness and FPT results, ETH lower bound, kernel results, and other structural-parameter theorems untreated.

Open questions for a prover

The case FOR (proponent)

There is a credible, though deliberately limited, continuous mirror here. The strongest anchor is the paper’s proved Theorem 3:

“\(C\)-\(CF_w\) can be solved in time \(vc^{O(vc)}n^{O(1)}\).”

The theorem is proved in this paper, using a vertex-cover decomposition and a reduction to maximum-weight matching.

The mirror I would propose is Mass-\(C\)-\(CF_w\). A type is a complete compatibility profile: two agents of type \(t\) have the same utility toward every type \(u\), and the same within-type utility. The instance consists of:

A coalition configuration is an integer vector \(q\in\mathbb Z_{\ge0}^{T}\) with \(1\le |q|\le C\). Its welfare is

\[ v(q)= \sum_{t<u}w_{tu}q_tq_u+ \sum_t w_{tt}\binom{q_t}{2}. \]

The solution chooses a nonnegative mass \(y_q\) of coalitions of each configuration, subject to

\[ \sum_q q_t y_q=\mu_t \qquad\text{for every type }t, \]

and maximizes

\[ \sum_q v(q)y_q. \]

The decision version asks whether the optimum is at least \(B\). The \(y_q\)'s are densities of teams per unit population; agents are not fractionally split inside a team, but the population contains a divisible mass of teams of each integer composition.

This is a genuine high-multiplicity version of the paper’s problem. If \(\mu_t=n_t/N\) is rational, clearing denominators gives \(n_t\) ordinary agents of type \(t\). If the optimal \(y_q\) is rational, clearing its denominators gives an ordinary finite partition with exactly the same normalized welfare. Thus the continuous problem is not merely “weighted coalition formation”; it is the rational-clone limit of \(C\)-\(CF_w\).

The regime is plausible in exactly the settings the paper mentions: large firms forming project teams, recurring training cohorts, or large pools of wireless/task-allocation agents. There may be \(10^4\)–\(10^6\) agents but only a few dozen role or compatibility types. For example, agents may be classified by skill, schedule, location, and collaboration profile. The high-multiplicity assumption is that agents with the same complete profile are interchangeable; it does not require every named agent to have identical preferences.

I would expect this mirror to be Class A at least for fixed \(C\). There are only

\[ \binom{\tau+C}{C} \]

coalition configurations, so the problem is an ordinary rational LP of polynomial size for every fixed team capacity. More interestingly, suppose the type-compatibility graph has a vertex cover \(K\) of size \(k\). Types outside \(K\) have no mutual utility, so once the configurations of the \(K\)-types are fixed, assigning peripheral mass to the remaining team slots is a transportation or matching LP. Enumerating the possible core compositions gives an algorithm fixed-parameter in \(k+C\), with polynomial dependence on \(\tau\) and the encoding length.

That is the continuous analogue of the structural idea behind Theorem 3: a small interaction core controls all valuable pairings, while the large independent remainder can be handled by matching. It also exposes a useful boundary. For unrestricted \(C\), the LP has exponentially many configurations, and its pricing problem becomes

\[ \max_{|q|\le C} \left(v(q)-\sum_t p_tq_t\right), \]

which is a quadratic optimization problem over team compositions. Hardness may then survive through the type-interaction graph rather than through population multiplicity. So the expected classification is:

The mirror is author-recognizable: it preserves the bounded coalition size, pairwise additive utilities, graph-derived compatibility, and utilitarian welfare objective. It does not replace coalitions by fractional bundles or change the welfare criterion. It only replaces a large multiplicity of interchangeable agents by rational masses and replaces integer counts of repeated team configurations by their high-multiplicity LP limit.

The weakest point is precisely the parameter transfer. The paper’s \(vc\) is the vertex-cover number of an individual-agent graph. In the proposed mirror, a small vertex cover of the type graph may expand to many interacting agents, so the paper’s exact \(vc^{O(vc)}n^{O(1)}\) theorem does not automatically apply. I would therefore present the configuration LP and its fixed-\(C\) tractability as the firm part of the case, and the role-level vertex-cover extension as the research question generated by Theorem 3—not as an already-established consequence.

I would not add the paper’s hardness and kernel theorems as positive anchors: their transfer under fractional team densities is genuinely uncertain. This one anchor is enough because it already gives a natural continuous population problem, a credible high-multiplicity regime, a concrete LP, and a nontrivial tractability boundary.

The case AGAINST (opponent, writing after the proponent)

The strongest negative case is that the proponent’s construction does not actually mirror Theorem 3. It replaces the paper’s parameter by a new type-level parameter and obtains a generic high-multiplicity configuration LP.

Theorem 3 exploits a small vertex cover of the *individual-agent graph*. That structure disappears under any nondegenerate cloning limit. If types \(t\) and \(u\), both of positive mass, have positive interaction, their clones form a complete bipartite graph. Every vertex cover must then contain all agents of one side, so its size is \(\Omega(N)\). A small vertex cover of the type graph therefore does not induce a small vertex cover in the paper’s graph.

Conversely, preserving \(vc=O(1)\) forces the interacting core to contain only \(O(1)\) named agents. With fixed \(C\) and fixed type-level weights, those agents can participate in only \(O(C)\) valuable pairs. After normalizing welfare by population size, the entire objective tends to zero. Thus the parameter regime that makes Theorem 3 interesting either explodes under continuization or degenerates into a zero-welfare limit.

The proposed LP is nevertheless mathematically valid. For fixed \(C\), enumerating all type compositions and optimizing coalition densities is a standard high-multiplicity relaxation, and denominator clearing gives the expected finite cloned instance. But that observation is independent of Theorem 3 and does not use its matching argument. It would apply to essentially any bounded-size additive packing problem. The fixed-\(C\) tractability is also obtained simply by making the number of configurations \(\binom{\tau+C}{C}\) polynomial; it does not reveal what the paper’s vertex-cover theorem becomes in the continuum.

The suggested repair—using a vertex cover of the type-interaction graph—is a plausible and perhaps worthwhile new problem. But it is not a parameter-preserving mirror: it changes individual vertex cover into a role-level parameter and changes the question from “does the paper’s sparse-core algorithm survive high multiplicity?” to “what can be done for a new finite-type configuration LP?” The possible \(k+C\)-parameter algorithm is consequently evidence for a related research direction, not for a continuous counterpart of the paper’s main result.

Still, this negative case cannot honestly establish the universal claim requested. Large pools of interchangeable role types—project teams, training cohorts, or recurring task-allocation populations—are plausible high-multiplicity societies, and the configuration LP preserves the paper’s capacity, pairwise utilities, and welfare objective. Nor can existing fractional hedonic-game work be invoked as a continuous collision, since it concerns outcome utilities rather than a continuous population.

So the anchor is weakened as a mirror of Theorem 3, but it is not defeated as a legitimate continuous problem. The honest verdict is that the universal negative case is weak: this paper does support at least one credible continuous mirror, even if the proponent overstates its connection to the stated vertex-cover theorem.

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.