A Little Subsidy Ensures MMS Allocation for Three Agents

The paper's numbered Theorems 3.1 and 3.2 assert an existential upper bound and an existential lower-bound instance, respectively; neither states a complexity result or an algorithmic guarantee. The proposed high-multiplicity configuration questions are plausible extensions, but they cannot supply the missing computational anchor required by the narrow criterion. The opponent therefore wins on the decisive objective point, even though its claim that no worthwhile mirror exists in any scenario is broader than necessary.

The objection that survived

The decisive surviving objection is that neither numbered anchor asserts a complexity or algorithmic result; the proposed continuum questions therefore lack the paper-level computational result required by criterion (a).

What the mirror covers

It covers the chores versions of Theorems 3.1 and 3.2 through a high-multiplicity extension, leaves the deferred goods result and informal NP-hardness/PTAS discussion outside, and does not turn either named theorem into a computational result.

Open questions for a prover

The case FOR (proponent)

The strongest honest case is a high-multiplicity mirror of the chores problem, led by Theorem 3.1 and supported by Theorem 3.2. It is a population-limit version of their MMS-with-subsidy question, not a claim that the paper already studies continuous populations.

Let \(T\) be a finite set of agent cost types and \(E\) a finite catalogue of indivisible item classes. Type \(t\) is the complete additive cost vector \(c_t=(c_t(e))_{e\in E}\), with \(c_t(e)\in[0,1]\). A society is a mass vector \(\mu\), where \(\mu_t\) is the fraction of agents of type \(t\). In a finite realization with \(N\) agents, there are \(N\mu_t\) agents of type \(t\), and \(Nb_e\) copies of item class \(e\), where \(b_e\) is the per-agent supply. Thus \(N\gg |T|\), while the number of distinct agent types remains small.

This is plausible, for example, in a large organization assigning standardized undesirable jobs to a large workforce with three stable skill or contract profiles. Agents within a profile have identical costs for every job class; the item copies remain indivisible. The continuous object is the agent population, not the contents of an individual bundle.

A bundle is an integral vector \(x\in\mathbb Z_{\ge0}^{E}\), with cost \(c_t\cdot x\). The asymptotic MMS of type \(t\) is the natural high-multiplicity limit \(M_t(b)=\inf_{\rho}\operatorname {ess\,sup}_{x\sim\rho}c_t\cdot x\), where \(\rho\) ranges over probability distributions on integral bundles satisfying \(\mathbb E_\rho[x]=b\). This is the limit of partitioning \(Nb_e\) item copies into \(N\) bundles.

A continuous allocation is a mass \(\lambda_{t,x}\) of type-\(t\) agents receiving bundle \(x\), satisfying \(\sum_x\lambda_{t,x}=\mu_t\) and \(\sum_{t,x}\lambda_{t,x}x_e=b_e\) for every item class \(e\). Each agent still receives an integral bundle. If agents receiving \(x\) get subsidy \(\sigma_{t,x}\), the MMS constraint is \(c_t\cdot x-\sigma_{t,x}\le M_t(b)\), and the objective is to minimize \(\sum_{t,x}\lambda_{t,x}\sigma_{t,x}\). For three equal-mass types, I normalize this by \(3\), so the quantity \(3\sum_{t,x}\lambda_{t,x}\sigma_{t,x}\) is comparable to the paper’s total subsidy for three agents.

The lead anchor is Theorem 3.1, proved in this paper. It states that every three-agent chore instance has an MMS outcome with total subsidy at most \(1/6\). The corresponding continuous problem is:

*Three-Type Chore MMS-Subsidy\(_\infty\).* Given three agent types \(T=\{1,2,3\}\), \(\mu=(1/3,1/3,1/3)\), a finite item catalogue \(E\), rational supplies \(b_e\), and rational additive costs \(c_t(e)\in[0,1]\), compute the minimum normalized subsidy \(S_\infty=3\sum_{t,x}\lambda_{t,x}\sigma_{t,x}\) over all feasible \((\lambda,\sigma)\) as defined above. In particular, decide whether \(S_\infty\le1/6\), and return the bundle-mass allocation and subsidies if so.

This preserves the paper’s actual question: every recipient must be brought up to her own MMS by additive monetary compensation, while the items remain indivisible. It does not replace MMS by expected cost, envy-freeness, or fractional bundles. The paper’s \(1/6\) becomes a conjectured asymptotic guarantee, not an assumed consequence of Theorem 3.1.

I would expect this problem to be Class A when the item catalogue size \(|E|\) is fixed or treated as a parameter. The allocation is a configuration LP over integral bundles, and fixed-dimensional integer optimization can provide the required pricing and MMS computations. For unrestricted \(|E|\), however, the pricing problem resembles knapsack or bin packing, so the paper’s statement that constructing MMS partitions is NP-hard predicts a genuine boundary. That hardness would live in the item catalogue and cost encoding, not in population multiplicity.

The paper’s own proof makes the mirror scientifically interesting. Theorem 3.1 relies on a three-agent MMS-feasibility graph, Hall’s theorem, and the two-bundle structural Lemma 3.3. In the continuous problem, one can ask whether those finite combinatorial arguments are replaced by configuration-LP duality, and whether \(1/6\) is sharp, improves, or fails as the number of agents tends to infinity. Further questions include finite-\(N\) rounding bounds, unequal type masses, more than three types, and the weighted MMS/APS variant mentioned in the conclusion.

A useful secondary anchor is Theorem 3.2, also proved here. It gives a three-agent, nine-item instance requiring subsidy at least \(2/49\). Its exact high-multiplicity mirror is particularly clean. Take three types \(R,C,U\), each of mass \(1/3\), nine item classes with \(b_e=1/3\), and normalized cost vectors

\(c_R=\frac{2}{49}(8,21,17,23,14,9,16,12,18)\),

\(c_C=\frac{2}{49}(8,21,18,23,14,10,15,11,18)\), and

\(c_U=\frac{2}{49}(49/6,43/2,52/3,49/2,43/3,49/6,43/3,34/3,55/3)\).

For every \(q\), this means \(q\) copies of each of the nine item classes and \(q\) agents of each type.

*Replicated Nine-Item MMS Gadget\(_\infty\).* Compute the minimum limiting normalized subsidy for this replicated instance, and determine whether the finite lower bound \(2/49\) survives: that is, whether \(S_\infty\ge2/49\), or whether time-sharing among different integral allocations makes the asymptotic subsidy smaller.

I expect this fixed nine-class problem to be Class A. It is a finite configuration problem whose exact value should be computable by a finite-dimensional LP or integer-optimization procedure. My prior is that the \(2/49\) obstruction may weaken, possibly substantially, because a continuum of agents can time-share among integral bundle patterns even though no single three-agent allocation works. But that is precisely a question to solve, not an assertion that the lower bound disappears. If the obstruction survives as an LP dual certificate, the mirror would reveal a robust high-multiplicity barrier; if not, it would show that Theorem 3.2 is a finite-indivisibility phenomenon.

I would not use the paper’s paragraph saying that MMS-partition construction is NP-hard as a third anchor: it is not presented as a numbered theorem in the supplied text. Nor would I anchor on the goods result, since “Result 2” is stated in the introduction while its proof is deferred. The proposed mirror therefore deliberately covers the proved chore results Theorems 3.1 and 3.2, not the whole paper.

The weakest point is that Theorem 3.1 is specifically a three-agent theorem. Its proof uses \(3\times3\) structure, and there is no automatic reason for its \(1/6\) constant to survive when \(N\) becomes large. A hostile referee could say that the proposed problem is a new weighted/high-multiplicity extension rather than the theorem’s literal continuous form. That criticism is real. The answer is that this is exactly what continuization should mean here: repeat the paper’s complete agent types, preserve integral bundles and MMS, normalize total subsidy per population unit, and study the resulting computational limit. If “three agents” is interpreted as an irreducibly fixed named population, then no nontrivial population mirror exists at all; the high-multiplicity interpretation is the only one that makes the continuization question meaningful.

The case AGAINST (opponent, writing after the proponent)

The strongest negative point is that the paper supplies no named computational anchor. Theorems 3.1 and 3.2 are existential statements: one gives a subsidy guarantee, the other gives a finite lower-bound instance. The only computational claim is an informal paragraph saying that MMS-partition computation is NP-hard and that PTASs therefore follow; no algorithmic theorem is stated. Thus the proposed problems are new high-multiplicity fair-division problems, not continuizations of computational results established by this paper.

Theorem 3.1 does not have a natural population limit. Its object is specifically three named agents and a partition into three bundles:
\[ \operatorname{MMS}_i(M,3)=\min_{A\in\Pi_3(M)}\max_{B\in A}c_i(B). \]
The proof is correspondingly three-agent: Hall’s theorem reduces the obstruction to the pair \(\{2,3\}\), and Lemma 3.3 exploits a \(3\times3\) structure.

In the proposed replication, \(3q\) agents receive \(q\) copies of every item class. Their benchmark is instead
\[ M_t(b)=\inf_{\rho:\,\mathbb E_\rho[x]=b}\operatorname{ess\,sup}_{x\sim\rho}c_t\cdot x, \]
which is the \(3q\)-agent, high-multiplicity limit. The number three now counts types, not agents or MMS bundles. That is not a harmless restatement: it changes the fairness benchmark and removes the finite matching structure on which Theorem 3.1 rests. The resulting question may be worthwhile, but it is a new many-agent weighted MMS theorem, not a mirror of the paper’s theorem.

The alternative does not help. If one retains only three bundles and lets each type’s mass split among them, the model is merely a convexified or randomized version of the original three-agent allocation. If one forbids such splitting so that each type receives one common bundle, the mass becomes cosmetic and the high-multiplicity limit has done no substantive work. There is no faithful middle ground in which the original three-agent proof remains the governing object while population mass is genuinely continuous.

Theorem 3.2 is weaker still as an anchor. Its claim is that every assignment of nine particular items to three particular agents makes someone pay at least \(1\). Under the proposed replication, \(\lambda_{t,x}\) permits different copies of one type to receive different integral bundles. The finite universal quantifier over three assignments is replaced by membership in a convex hull of configuration patterns. To preserve the obstruction, one must impose a common-bundle or correlation constraint tying all copies together; that reintroduces named-agent structure and defeats the population model. If splitting is allowed, the proposed question is simply whether time-sharing removes the finite gadget’s obstruction—a legitimate experiment, but not a computational result of the paper. For the displayed nine-class instance itself, it is also only a fixed finite calculation, not a complexity problem.

The goods result cannot rescue the case: it is stated as “Result 2,” but its proof is explicitly deferred, so it is neither a proved computational theorem nor an independent computational anchor.

The honest limitation is that the universal negative claim is not airtight. The standardized-jobs scenario is a credible high-multiplicity regime: complete cost types, repeated item classes, and indivisible bundles are all legitimate. Its configuration problem could reveal an interesting disappearance of the \(2/49\) obstruction. But that would be a new population-scale MMS programme motivated by this paper, not a continuous computational mirror of any computational theorem it actually proves. The strongest defensible verdict is therefore that this is a poor ChoCo target, not that no worthwhile mirror can exist in any scenario.

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.