| paper | d4-journals-jet-2025-0068 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The theorems constructively characterize optimal allocation rules, so they have computational character. But both proposed mirrors optimize over a Gaussian fluctuation \(Z\), not over a society supplied as a mass distribution: for every fixed \(z\), \(\mu_N(z)\) converges to \((1/2,1/2)\). In an atomless society, independent report flips instead produce a deterministic mass transformation, removing the randomness and fluctuation scale on which these results rely.
fails bit b-changes — the continuum limit changes the question
For each fixed \(z\), \(\mu_N(z)\) converges to the same balanced society, so \(Z\) represents finite-sample fluctuation rather than mass in a limiting society; under atomless independent flips, the aggregate noisy mass is deterministic.
fatal: True
The proposed mirrors cover Theorems 1 and 2's revenue, noise-sensitivity, and surplus tradeoffs in the Gaussian fluctuation limit. They leave the implementation results, Fourier comparative statics, and imperfect-knowledge analysis without a continuous-population mirror.
The strongest honest case is a local high-multiplicity mirror of the paper’s revenue and robustness tradeoffs. A large municipality deciding whether to fund a public project could have \(N\) residents but only two complete types: low and high valuation, \(x_i\in\{-1,+1\}\). Residents of the same type share the paper’s utility parameter \(b\), transfer rules, and report-noise rate \(\delta\). Thus \(N\) is enormous while the type support is just \(\tau=2\).
Represent a society by the mass \(\mu_N(z)=(\tfrac12+\tfrac{z}{2\sqrt N},\,\tfrac12-\tfrac{z}{2\sqrt N})\) of high- and low-valuation residents. The paper’s uniform i.i.d. prior on finite profiles induces \(Z\sim N(0,1)\) for the centered, scaled mass imbalance in the large-population limit. Independent report flips induce a noisy imbalance \(Z_\delta\) that is jointly Gaussian with \(Z\), with correlation \(\rho=1-2\delta\). A provision rule \(F:\mathbb R\to\{0,1\}\) decides whether to provide the good from the reported imbalance. This keeps the population as the continuous object; the Gaussian noise kernel preserves the aggregate fluctuations that would disappear if one simply took an atomless population and applied independent flips to it.
My lead anchor is Theorem 1, proved in this paper in §6.1. It says that for the paper’s revenue-constrained noise-sensitivity minimization problem, both linear threshold rules with cutoffs \(+\phi^{-1}(r)+o(1)\) and \(-\phi^{-1}(r)+o(1)\) are asymptotically optimal: each has noise sensitivity between the optimum \(V_n(r)\) and \(V_n(r)+o(1)\). The claim is over implementable Boolean allocation rules, not merely over a threshold class.
The continuous question I would hand to a prover is \(\mathsf{RobustRev}(\delta,r)\). Given \(\delta\in(0,\tfrac12)\) and \(r\in(0,1/\sqrt{2\pi})\), choose a measurable binary rule \(F\) to minimize \(\Pr(F(Z)\ne F(Z_\delta))\), subject to \(\mathbb E[ZF(Z)]\ge r\). Here \(\mathbb E[ZF(Z)]\) is the limiting normalized maximum revenue: the paper’s revenue divided by \((1-2\delta)\sqrt N\), up to a vanishing lower-order term. The positive-moment constraint is the limiting implementability condition for anonymous rules, and meeting the positive target ensures it. A solution is an optimal \(F\), with finite-population BIC and IIR transfer implementations for its approximants. I expect this fixed-Gaussian problem to be tractable by the same structural and Gaussian-isoperimetric ideas: the expected solutions are the two upper-threshold rules \(F(z)=\mathbf 1\{z\ge \phi^{-1}(r)\}\) and \(F(z)=\mathbf 1\{z\ge-\phi^{-1}(r)\}\). Further questions include whether the result admits effective finite-\(N\) error bounds and what remains tractable for other composition priors or noise kernels.
Theorem 2, also proved here (§6.2), provides a separate anchor. It says that for sufficiently large \(n\), the revenue-constrained expected-surplus maximization problem has the linear threshold \(\ell_n(\cdot;r)\) from equation (3.2b) as an exactly optimal solution; its cutoff is \(-\phi^{-1}(r)+o(1)\). Its continuous mirror, \(\mathsf{SurplusRev}(b,\delta,r)\), uses the same Gaussian society and revenue constraint, but maximizes \(\mathbb E[F(Z)]\), the leading per-capita surplus up to the positive factor \(b/2\). For \(b>0\), the expected solution is \(F(z)=\mathbf 1\{z\ge-\phi^{-1}(r)\}\): it provides the good to the largest mass of societies consistent with the revenue target. I expect this fixed-regime problem to be tractable by Gaussian rearrangement. Natural extensions ask how the optimizer changes with richer type spaces, heterogeneous project biases, or non-Gaussian priors.
The authors would plausibly recognize both as mirrors of their problem: the agents’ private valuations, binary public-good decision, privacy noise, implementability, revenue, and surplus all remain. The paper’s asymptotic analysis already supplies much of the dictionary and the answer in this Gaussian regime. That is useful evidence for a meaningful mirror, not a claim that these characterizations are new.
The weakest point is also central. The paper has no named result classifying a problem as polynomial-time, NP-hard, FPT, or otherwise. Theorem 1 and Theorem 2 are named optimization characterizations, not formal complexity results; under a strict complexity-anchor requirement, the paper supplies no qualifying anchor. And the mirror’s state is the mass imbalance at the \(1/\sqrt N\) fluctuation scale. A literal atomless continuum with deterministic mass shares would erase the pivotal fluctuations and noise sensitivity that drive the paper’s results. Retaining that scale is a defensible co-modification of the population model, but it is where a skeptical author or referee is most likely to press.
The strongest case against these mirrors is that the proposed Gaussian limit keeps the paper’s randomness by shrinking the society to a point and zooming in on finite-population fluctuations. That is a useful asymptotic model, but it is not the continuous society in ChoCo. I would grant that the paper has a convincing high-multiplicity regime: many residents can share either of two valuation types. The problem is that multiplicity alone does not preserve the phenomena these theorems optimize.
For Theorem 1, the proposed \(\mathsf{RobustRev}\) treats \(Z\) as the normalized imbalance in a random finite electorate. In the case’s own notation, \(\mu_N(z)=(\tfrac12+\tfrac{z}{2\sqrt N},\tfrac12-\tfrac{z}{2\sqrt N})\) converges to the same society, \((\tfrac12,\tfrac12)\), for every fixed \(z\). The Gaussian variable records which finite electorate was drawn; it is not the mass distribution of a limiting society. Likewise, the random \(Z_\delta\) comes from finite-sample flips. In an atomless society with fixed masses, independent flips produce the noisy mass \(\mu'_+=\delta+(1-2\delta)\mu_+\) almost surely. For a fixed society, comparing its rule before and after that transformation is deterministic, not the paper’s noise-sensitivity probability.
One could retain stochasticity by using the local \(1/\sqrt N\) scaling, a prior over societies, finite sampling, or common aggregate noise. Each is a possible model, but the first two make the random society or its local fluctuations the object being studied; the latter two change the observation mechanism. None rescues the proposed mirror as a computational problem over the continuous society \(\mu\) that ChoCo takes as its input. In particular, the Gaussian formulation has only scalar parameters \(r\) and \(\delta\); its threshold characterization is an analytic limit, not an instance-dependent complexity question.
Theorem 2 inherits the same mismatch. Its proposed objective \(\mathbb E[F(Z)]\) averages surplus over local fluctuations in finite electorates. For a fixed continuous mass vector there is no such \(Z\). To recover that objective one must supply a prior over societies, and the revenue constraint \(\mathbb E[ZF(Z)]\ge r\) then describes that added prior and the finite-population pivotal scale. Without it, one has a different problem: surplus and transfers for a fixed mass distribution, with no direct counterpart of the theorem’s revenue-constrained threshold characterization. Theorem 2 is a real optimization result, but it does not by itself yield an algorithmic-complexity result for continuous societies.
So I would reject both proposed anchors as ChoCo mirrors, while recognizing that the paper’s Gaussian asymptotics make them close and mathematically natural. The negative case is not airtight as a claim about *every* scenario: a richer model with a prior over continuous compositions or finite sampling might be a worthwhile adjacent project. Its weakness is also its boundary: those additions may support new computational questions, but they do not establish a population-continuous mirror of Theorems 1 and 2 as stated.
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.