Respecting priorities versus respecting preferences in school choice: When is there a trade-off?

The paper's numbered propositions are structural statements about allocations and market conditions, not computational guarantees, hardness results, or formulations used to compute. The proposed finite-type mass model is a plausible high-multiplicity extension, but its DA and mass-GMBP decision tasks supply new computational content rather than transporting a result of this paper. Since bit (a) fails for every anchor, the mirror is red under the stated gate.

The objection that survived

The opponent's decisive point is that the proposed mass-DA and mass-GMBP recognition problems add the computational question absent from Propositions 1 and 3.

What the mirror covers

The proposed extension targets Propositions 1 and 3; Proposition 2, the stylized examples, and the simulation findings remain outside it, and no paper result receives a valid computational mirror.

Open questions for a prover

The case FOR (proponent)

The honest positive case is a qualified one. The paper has no named computational result in the strict ChoCo sense: no numbered theorem or proposition states NP-hardness, membership in \( \mathrm{P} \), FPT, approximation, or a comparable complexity classification. Its numbered results are structural. Thus this cannot be a fully compliant computational-anchor case. Still, it supports a plausible, author-recognisable high-multiplicity mirror.

The natural regime is a large school district with millions of students but relatively few complete preference–priority signatures. A type \(t\) records a student’s full ranking over schools and the student’s priority category at every school: catchment area, sibling status, test band, or another feature used by the district. Let \( \mu_t \) be the fraction of students of type \(t\), with \( \tau \ll n \). School capacities must scale with population: if the finite district has \(Nq_s\) seats at school \(s\), the continuum instance has per-capita capacity \(q_s\).

The continuous allocation is a matrix \(x_{ts}\ge 0\), where \(x_{ts}\) is the mass of type \(t\) assigned to school \(s\). It satisfies \( \sum_s x_{ts}=\mu_t \) and \( \sum_t x_{ts}\le q_s \). Every infinitesimal student still receives one integral school assignment; only the aggregate mass is divisible. Priorities are orders over types, with agents inside one type treated as an anonymous priority block. If a capacity cuts through a block, the corresponding mass may be split at the boundary.

A mass allocation is envy-free if there is no type \(t\) with positive mass assigned to a less-preferred school \(r\) while \(t\) prefers \(s\) to \(r\) and school \(s\) either has capacity or assigns positive mass to a strictly lower-priority type. Pareto improvement is interpreted by first-order stochastic dominance within each type: for every rank cutoff, the improved allocation gives that type at least as much mass in schools above the cutoff, with a strict improvement somewhere.

My lead anchor is Proposition 3, proved in this paper: “School choice market \(E\) satisfies the GMBP condition if and only if it admits a unique envyfree allocation and that envyfree allocation is efficient.” This is the paper’s strongest result because it is an exact characterization, not merely a sufficient condition.

The corresponding problem is:

\(\mathsf{Continuous\text{-}NoTradeoff}\). Given \(S\), rational capacities \(q_s\), a finite type set \(T\), rational masses \( \mu_t \), each type’s preference order, and each school’s priority order over types, compute the student-proposing deferred-acceptance mass allocation \(x^{\mathrm{DA}}\). Decide whether \(x^{\mathrm{DA}}\) is the unique envy-free mass allocation and is Pareto-efficient. If the answer is YES, output \(x^{\mathrm{DA}}\) and a certificate of uniqueness and efficiency; if NO, output either a second envy-free allocation or a Pareto-dominating feasible allocation.

This is a direct population mirror of the paper’s question: when do priorities and preferences cease to conflict? It is not a fractional-outcome reformulation in the usual sense. The object being continuized is the population, while schools, priorities, rankings, and the assignment of each infinitesimal student remain discrete.

I would expect this problem to be Class A under explicit finite type support. Fractional DA can be run over type masses, and Pareto efficiency can be tested by a linear feasibility problem with variables \(y_{ts}\), row-sum and capacity constraints, and the rank-cutoff inequalities defining stochastic dominance. Uniqueness should be testable by comparing the student- and school-proposing mass-DA outcomes under responsive priorities. This is an expected result for the mirror, not something established by the paper.

The supporting anchor is Proposition 1, also proved here: if \(E\) satisfies GMBP, then there is a unique envy-free allocation, it is efficient, and it is produced by student-proposing DA, school-proposing DA, and IA.

Its continuous counterpart is:

\(\mathsf{GMBP}_\infty\). Given the same finite-type mass instance, iteratively remove a school \(s\) from type \(t\)’s relevant preference list whenever \(t\) has a more-preferred school \(s'\) with enough residual capacity to accommodate the relevant mass of \(t\) ahead of all higher-priority residual types. Then ask whether the resulting simplified mass market admits a sequential elimination of mutually best type-mass blocks: at every step, a remaining type mass is assigned to its most-preferred school with residual capacity, and that mass lies within the school’s highest-priority admissible mass. A YES solution consists of the simplification sequence, the sequential mass-matching sequence, and the resulting allocation.

This is also plausibly Class A, with at most polynomially many type–school deletions and rational arithmetic polynomial in \(m\), \( \tau \), and the input length. Proposition 1 would then predict that the three DA/IA procedures agree on the same efficient envy-free allocation.

The finite-to-continuous bridge is credible. If \( \mu_t=n_t/N \) and \(q_s=Q_s/N\), multiplying every mass by \(N\) recovers a finite market with \(n_t\) students of each type and \(Q_s\) seats. Under the anonymous priority-block convention, aggregate assignments, envy-freeness, and Pareto comparisons coincide. Thus this is a genuine high-multiplicity regime: the computational question is whether one can work with \( (\mu_t)_{t\in T} \) rather than with millions of named students.

The authors should recognise this as their problem. It preserves GMBP, simplified markets, DA, envy-freeness, Pareto efficiency, and the policy question of whether the choice of mechanism matters. The scenario is also close to their own catchment-area and sibling-priority examples, but replaces individually idiosyncratic students by repeated geographical and priority profiles.

I would not use Proposition 2 as a further anchor. A continuum version of TTC requires a nontrivial definition of fractional trading cycles, and adding it would make the case less precise without adding much beyond Proposition 1.

The weakest point is decisive: the paper does not itself contain a qualifying computational result. The proposed problems are new extensions of Propositions 1 and 3, not continuizations of a stated complexity theorem. There is also a genuine modelling issue: the paper assumes strict priorities over named students, whereas repeated complete types naturally create priority blocks and fractional boundary mass. The mirror is therefore author-recognisable and mathematically plausible, but it should be presented as a promising Class-A research extension—not as a fully compliant ChoCo anchor.

The case AGAINST (opponent, writing after the proponent)

The decisive objection is that this paper has no eligible computational anchor. Propositions 1 and 3 are structural matching results: they characterize when DA is efficient, when envy-free allocations are unique, and when GMBP holds. The paper contains no theorem or corollary about complexity, approximation, parameterized algorithms, or even the complexity of recognizing GMBP. Its simulations are empirical, not computational-complexity results.

That defeats the proposed \(\mathsf{Continuous\text{-}NoTradeoff}\) mirror of Proposition 3. The proposed problem adds the computational content: it asks for an algorithm, certificates, and perhaps an LP or flow formulation for a new aggregate model. That could be a worthwhile high-multiplicity school-choice paper, but it is not a continuization of a computational result in this paper. The original result says which finite markets have a structural property; it does not claim that deciding the property is hard, easy, or algorithmically interesting.

The best repair does not change this conclusion. One could define a fluid DA process over finitely many preference–priority types, quotient uniqueness by permutations within a type, and define Pareto efficiency over aggregate assignments. That is a coherent model. But the central theorem would then be a new theorem about high-multiplicity stability and fractional assignments. The paper supplies no computational statement whose answer is being transported to the continuum.

The type issue makes the proposed bridge less direct than claimed, although it is not by itself a fatal objection. The paper assumes strict priority orders over named students. If two students are genuinely the same type, their school-specific priorities must either be tied or their tie-breaking positions must be included in the type. Including the positions typically makes \(\tau\) essentially as large as the student population; omitting them replaces strict individual priorities by anonymous priority blocks. The latter is a sensible model, but it is a substantive redefinition. At a capacity boundary, a type may be partly admitted and partly rejected, so “the type is among the top \(q_s\) students” must be replaced by a cumulative-mass condition. Again, that may be the right extension, but it requires new definitions and proofs.

The same problem defeats the GMBP anchor. GMBP is not presented as an algorithmic result; it is a domain characterization proved by eliminating individual students and seats. The proposed mass version cannot simply remove a type: a positive-mass type may straddle a priority cutoff, and the sequential mutually-best-pair argument must be rebuilt around partial mass blocks. A repaired version using cumulative priority mass could be mathematically sound, but it would be a new continuum characterization, not a computational mirror of Proposition 1. Calling its recognition problem “polynomial” would be a new contribution, not a continuization of anything the paper established.

The rational scaling argument is valid only in the limited sense that repeated-type finite markets can be represented by rational masses, and mass allocations can often be implemented by sufficiently many identical students. That is evidence that the high-multiplicity regime is sensible; it is not evidence that this paper contains a computational result to mirror. Nor should existing high-multiplicity matching work be treated as a collision: it supports the regime rather than invalidating it.

There is also a broader loss of content. In the finite paper, allocations and justified envy concern named students, and strategyproofness concerns unilateral deviations. In an atomless population, one student has measure zero and cannot affect the outcome, so individual manipulation becomes vacuous. Treating a positive-mass type as the strategic unit instead creates a coalition-manipulation problem, not the paper’s mechanism-design problem. Aggregate efficiency and stability can be retained, but only after deliberately selecting which parts of the original object to discard.

Thus the negative case is strong under ChoCo’s actual gate: neither proposed anchor is a computational result, and every worthwhile continuous formulation would be a new high-multiplicity school-choice project rather than a mirror of this paper. It is not honest to claim that no continuous school-choice model could ever be worthwhile; the proponent has identified a plausible one. The honest verdict is narrower but decisive: this paper is not a suitable source for a continuous-computational mirror.

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.