Envy-freeness and maximum Nash welfare for mixed divisible and indivisible goods

The paper has no numbered result with the required computational character. The proposed continuum MNW/ EF1M problem is a sensible high-multiplicity extension, but it is not a computational mirror of Theorem 7 or any other numbered result. Under the strict gate, the missing computational anchor is decisive.

The objection that survived

The continuum formulation changes an every-pair finite guarantee into an almost-everywhere guarantee, and no numbered computational result is preserved.

What the mirror covers

The proposed mirror concerns MNW and EF1M, especially Theorem 7; it does not computationally mirror the structural equivalences, fairness guarantees, or existence corollaries.

Open questions for a prover

The case FOR (proponent)

The paper does not contain a qualifying computational anchor. Its numbered results are structural or existential:

Theorem 2: MNW, leximin, and \(\Phi\)-fairness coincide under binary indivisible valuations and identical divisible valuations.
Theorem 3: under binary linear valuations, every suitable \(\Phi\)-fair allocation is EFXM.
Theorem 7 (Caragiannis et al. [9], cited and reproved here): with additive valuations, every MNW allocation is EF1M.
Corollaries 1–3 establish existence of EFXM/EF1M and Pareto-optimal allocations.

None asserts NP-hardness, membership in P, parameterized complexity, or an algorithmic formulation. The computational statements are unnumbered remarks in Section 4.3 or results cited from elsewhere, notably the NP-hardness of mixed-goods MNW from [22] and APX-hardness for indivisible goods from [23]. Under the stated rules, they cannot serve as anchors.

The strongest conditional positive case would use Theorem 7. Its exact continuous analogue could be:

Continuum-MNW-EF1M. The input is a finite type set \(T\), masses \(\mu_t\), additive valuations \(v_t\) over finitely many indivisible-good classes and divisible-good classes, and per-capita supplies of both kinds. An allocation assigns every agent an integral bundle of indivisible goods and fractional quantities of divisible goods; aggregate variables record what fraction of each type receives each integral bundle. The objective is first to maximize the mass of agents with positive utility and then to maximize
\[ \int \log u(a)\,d\lambda(a). \]
The required output is an optimizing allocation that is EF1M for \(\lambda\times\lambda\)-almost every ordered pair of agents: if agent \(b\) receives indivisible goods, some one good from \(b\)'s bundle removes \(a\)'s envy; if \(b\) receives none, \(a\) must not envy \(b\).

A credible regime is a large workforce of recurring shift-preference types: millions of workers, perhaps only a few dozen valuation types, many repeated copies of each shift slot, and a divisible wage or compensation pool. Each worker still receives an integral shift bundle; only the population assignment is continuous. This is recognisably the same MNW/EF1M problem, not a cake-cutting problem disguised as one.

I would investigate the restricted binary or identical-valuation versions as possible Class A cases, using convex optimization over aggregate bundle masses and column generation. For general additive valuations, hardness may survive through the indivisible-good structure. The main weakness is serious: Theorem 7’s finite-agent proof improves welfare by transferring goods between particular agents, whereas changing one agent has zero effect in a nonatomic population. The correct limit therefore needs an almost-everywhere or positive-mass formulation, and the theorem does not establish that.

Thus there is a plausible continuous-population mirror, but it is not anchored by a named computational result from this paper. Under the programme’s strict protocol, that limitation should be reported explicitly rather than presented as a full positive case.

The case AGAINST (opponent, writing after the proponent)

The strongest negative case is procedural and decisive under the programme’s stated gate: this paper contains no qualifying computational anchor. Theorem 2, Theorems 3–7, and Corollaries 1–3 are structural, fairness, or existence results. The algorithmic remarks in Section 4.3 compose procedures from earlier work, while the cited hardness results belong to other papers. Continuizing any numbered theorem would therefore create a new computational problem rather than mirror one posed by this paper.

The proposed use of Theorem 7 does not escape that objection. Theorem 7 says that every MNW allocation has EF1M; it does not assert a complexity bound, formulate an optimization problem, or claim an algorithm. Replacing the finite Nash product by
\[ \int \log u(a)\,d\lambda(a) \]
and replacing agents by type-masses is a plausible new high-multiplicity fair-division problem, but it is not a continuous version of the theorem’s computational content, because there is no such content to preserve.

The best repair is to scale indivisible goods per capita, use finitely many valuation types, and let \(x_{t,B}\) denote the mass of type \(t\) receiving integral bundle \(B\). That avoids the superficial objection that indivisible goods disappear in a nonatomic limit. A recurring shift-workforce scenario is also genuinely credible: the paper itself mentions binary shift preferences, and many workers can share each type. Thus one must not claim that identity, indivisibility, or multiplicity make the model meaningless.

But that concession exposes the limit of the proposed anchor. The resulting configuration-LP or convex-optimization problem could be worthwhile research, especially for computing high-multiplicity MNW allocations. It is nevertheless a new problem motivated by the paper’s subject, not a computational mirror of a result in the paper.

There is also a substantive mismatch in the fairness conclusion. The finite theorem guarantees EF1M for every ordered pair of agents. The natural continuum statement guarantees it only for \(\lambda\times\lambda\)-almost every pair, so all exceptional individuals—and even exceptional type–bundle cells of zero mass—become invisible. Requiring the property for every type in the support would restore a stronger guarantee, but that is a newly chosen robust notion rather than the direct measure-theoretic limit.

So the paper should receive a negative verdict under the strict protocol: it supplies no named computational result to continuize, and Theorem 7 cannot serve as one merely because a new high-multiplicity optimization problem can be built around it. The universal substantive claim is weaker: the shift-workforce construction shows that a continuous population model could itself be sensible and potentially valuable. What fails is not the possible model, but its status as a worthwhile mirror of this particular paper.

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.