| paper | d4-conf-ijcai-2021-0331 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
A type-mass DRF-MT allocation is a coherent high-multiplicity problem, and the opponent concedes that it survives as a model. But the only anchor raised, Lemma 4, establishes a fairness property rather than a computational result; the LP runtime discussion and numbered Claims do not meet the required named-result gate.
fails bit a — no named computational result to mirror
The candidate mirror covers computing the progressive DRF-MT allocation and re-examining the sharing-incentive guarantee of Lemma 4; it leaves the other allocation properties, proportionality, and rounded indivisible allocations unclaimed.
The strongest honest case is limited by the anchor rule: this paper has no named computational-complexity result. Its numbered Lemmas 1–5 establish allocation properties, such as Pareto optimality and weighted envy-freeness. The closest fit is Lemma 4, proved here: when demands, weights, and supplies are rational and each agent’s priority weights are set from their accessible contributions to the resource pool, DRF-MT satisfies sharing incentive. That is a substantive result, but it does not assert that a problem is in P, NP-hard, or belongs to another complexity class. The paper says its algorithm runs in polynomial time, and Claim 1 supports its termination bound, but neither is a numbered theorem, lemma, corollary, or proposition asserting a complexity classification. So neither can serve as an anchor under the requested rules.
There is still a plausible population mirror. Consider a regional health system with many hospitals, where a relatively small set of complete profiles recurs: each profile specifies per-capita demands, accessible resource groups, priority weights, and, for sharing-incentive instances, accessible contributions. With \(n\) hospitals and only \(\tau \ll n\) profiles, type \(t\) has mass \(\mu_t\), its fraction of the hospital population. Supplies and demands are measured per capita. This makes the *population* continuous; divisible resource allocations are already part of the paper’s model and are not the proposed continuization.
A candidate problem, Type-Continuous DRF-MT Allocation, would take a finite resource set partitioned into meta-types, per-capita supplies \(\bar S_r\), and a finite support \(T\) with rational masses \(\mu_t\) summing to \(1\). Each type \(t\) has rational per-capita demands \(d_{t\ell}\), demand groups \(g_{t\ell}\), and priority weights \(q_{t\ell}\), with \(\sum_t\mu_tq_{t\ell}=1\) for each meta-type. A solution is a feasible aggregate allocation \(z_{tr}\) and the successive max-min levels produced by DRF-MT’s progressive-filling rule. Feasibility requires \(\sum_t z_{tr}\leq\bar S_r\); type \(t\)’s per-hospital utility is \(u_t=\min_{\ell:d_{t\ell}>0}\frac{\sum_{r\in g_{t\ell}}z_{tr}}{\mu_t d_{t\ell}}\). In each round, maximize \(y\) subject, for each active type and demand group, to \(\mu_t y q_{t\ell_t^*}d_{t\ell}/d_{t\ell_t^*}\leq\sum_{r\in g_{t\ell}}z_{tr}\), where \(\ell_t^*\in\arg\min_\ell q_{t\ell}/d_{t\ell}\); retain earlier guarantees for eliminated types and eliminate types and resources by the paper’s rule. For the sharing-incentive version, set priority weights from accessible contributions, using the paper’s phantom-weight convention where needed.
I would expect this candidate to be Class A: for \(\tau\) supported types, the paper’s successive LPs can be written over type-level allocations, with size polynomial in the number of resources, demand groups, \(\tau\), and input bit length. That expectation is an inference from the algorithm and its termination argument, not a conclusion of a qualifying named result. Lemma 4 is the closest property to carry over: the continuous version would ask whether each type’s per-capita allocation gives at least its per-capita utility from its accessible contribution.
The weakest point is that even this reasonable mirror has no qualifying anchor in the paper. It also appears computationally modest: DRF-MT already uses polynomial-sized LPs and is tested on large finite populations, so compressing repeated profiles may add little new complexity theory. The mirror covers the fair allocation rule and the sharing-incentive question; it does not establish a continuous-population result about strategy-proofness, which concerns individual reports, or about the paper’s rounding and indivisible-allocation discussion.
The strongest case against this proposal is the anchor rule: Lemma 4 is not a computational result. It proves a fairness guarantee—sharing incentive under a condition on the weights—not a complexity classification or algorithmic bound. Claim 1 establishes that an agent and a resource are eliminated each round; the polynomial-time claim follows from that and the LP formulation, but the paper does not state it as a numbered theorem, lemma, or corollary asserting complexity. So the proposed mirror has no qualifying anchor under the brief’s criteria.
The proposed type model is still coherent. If several agents have the same complete profile, their per-agent allocations can be averaged into one type-level allocation. The resulting LP aggregates the same supply and fairness constraints; identical types can be treated symmetrically. With rational masses, that is a high-multiplicity version of the finite-agent problem. The paper’s algorithm already solves LPs whose size is polynomial in the number of agents and resources, so this aggregation appears to yield an immediate support-size version. That makes the candidate look incremental, though “the answer is easy” is not, by itself, a valid objection under the programme’s rules.
There are plausible regimes for the model: hospitals using standardized demand and access profiles, or cloud users on repeated resource plans. The paper’s own settings do not rule those out, and high-multiplicity work supports the general regime. Nor does the fact that resources are already divisible make this a collision: the population remains finite in the paper. I would not claim that no sensible continuous-population model can be made here.
The negative case is therefore narrow but firm against the case as presented: the sharing-incentive mirror may be a worthwhile question, but it is not anchored to a named computational result, as required. The broader claim that this paper has no worthwhile continuous mirror in any scenario is weak; the type-level LP is a plausible mirror that would need to be advanced as a new computational question rather than as a continuization of Lemma 4.
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.