| paper | d4-conf-aaai-2023-0872 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 6 is computational, but its knapsack choices rely on a single agent being able to change project funding. In an atomless population, an individual has zero mass and cannot do that; keeping a pivotal unit agent preserves the original strategic object rather than continuizing the operative population. A coordinated mass deviation could restore pivotality, but it would be a different question from the individual best response in the theorem.
fails bit b-changes — the continuum limit changes the question
The designated resident remains a pivotal individual, so the type masses and their aggregate contributions provide background data without transferring the theorem’s individual best-response problem to a continuous population. Making a positive-mass cohort pivotal instead would ask about coordinated deviation, which the paper does not study.
fatal: True
The proposed mirror targets Theorem 6’s discrete-contribution best-response hardness; it does not cover the equilibrium funding results, the impossibility results, or the simulations.
The strongest mirror is narrow: the paper’s optimal-deviation problem, not its equilibrium-welfare results.
The anchor is Theorem 6, proved in this paper. It states that with discrete contributions and refund schemes satisfying Contribution Monotonicity, computing an agent’s utility-maximizing contribution strategy is NP-hard, even when the contributions of all other agents are given. The proof reduces from KNAPSACK. For the reduction, it uses a common refund rule with the additivity property \(\sum_j R(x_j,\cdot)=R(\sum_j x_j,\cdot)\). I would base the mirror on that explicit subclass, rather than rely on the theorem’s broader “any set” wording: each project represents an item, the remaining amount needed to fund it represents its weight, and the agent’s net value from funding it represents its value.
A credible setting is a large civic-crowdfunding platform running many small local projects. Residents fall into a modest number of complete types: each type specifies the resident’s valuation for every project and their contribution budget. Such types could arise from standardized benefit zones and contribution-cap categories. The regime is many residents but relatively few types, \(n\gg\tau\); the number of projects \(p\) can grow. This is a scenario-specific case, not a claim that all crowdfunding populations have few types.
Here is the continuous problem I would hand to a prover: Continuous-Mass Civic Crowdfunding Best Response. Its input is a rational distribution \(\mu\) over \(\tau\) resident types, where a type \(t\) consists of a valuation vector \((\theta_{tj})_{j\in P}\) and budget \(\gamma_t\). The population has total headcount-equivalent mass \(n\); one resident of type \(t^\star\) is designated, so the rest of that type’s mass is \(n\mu_{t^\star}-1\). For each type and project, the input gives the fixed aggregate contribution \(A_{tj}\) of the other residents, subject to their type budgets. It also gives project targets \(T_j\), a common refund bonus and monotone additive refund rule from the proof’s subclass, and the designated resident’s contribution quantum \(\delta\).
The question is to find a budget-feasible vector \(x\in\{k\delta:k\in\mathbb Z_{\ge0}\}^{p}\), with \(\sum_jx_j\le\gamma_{t^\star}\), maximizing that resident’s total utility against the fixed background. For project \(j\), total contributions are \(C_j=\sum_tA_{tj}+x_j\): the resident gets \(\theta_{t^\star j}-x_j\) if \(C_j\ge T_j\), and the specified refund \(R(x_j,B,C_j)\) otherwise. A solution is a maximizing contribution vector and its utility. The input and the computation are measured in \(p,\tau\), and the binary encoding length \(L\), including the contribution-grid range.
This preserves the paper’s strategic question: the designated resident chooses contributions across projects under one budget, trading off funded utility against refunds. The rest of society enters through its type masses and aggregate contributions. That is especially plausible here because the paper’s own best-response formulation already summarizes everyone else by their contributions to each project; replacing their individual list with type masses and totals does not change what the deviator faces. The mass is the population share, while aggregate money remains a headcount-scaled quantity. The focal resident’s \(\delta\)-grid stays discrete because Theorem 6 concerns that version of the problem; the mirror continuizes the population.
I expect exact optimization to be Class B: hardness transfers. The KNAPSACK choices are which projects the focal resident funds, and those choices live in \(p\), not in the number of residents. The rational background contributions can encode the same residual funding gaps as in the reduction. The continuous distribution of everyone else therefore need not remove the combinatorial core. A useful next question is whether the result admits approximation or fixed-parameter algorithms in \(p\), \(\tau\), or the grid range. Another is what happens as \(\delta\) tends toward zero: Theorem 5 says an optimal strategy may fail to exist with continuous contributions, so that version may need to ask for a supremum or an \(\varepsilon\)-optimal deviation.
The mirror covers Theorem 6’s best-response hardness, not the paper’s claims about equilibrium funding, welfare optimality, or its simulations. Its weakest point is the focal resident. In a strictly atomless model, one individual has zero mass and cannot change whether a project crosses its funding threshold. This mirror keeps a designated unit-mass resident inside an otherwise continuous population so that the pivotal funding gaps in the paper’s problem remain meaningful. If the authors or a referee require every agent to be infinitesimal and the question to be purely type-mass-level, this is no longer a faithful mirror of their optimal deviation problem.
Theorem 6 is the paper’s strongest anchor, and the proponent is right to focus on it. But its computational object is one named agent choosing a discrete contribution vector against fixed contributions from everyone else. The agent’s own action can change whether a project crosses its funding threshold. A continuous population does not preserve that combination of features.
In a genuine atomless society, one resident has zero mass. Their contribution cannot change aggregate funding, so a unilateral best response takes funded status as fixed. Away from threshold knife edges, the resident no longer chooses which projects to make fundable—the cross-project choice that supports the paper’s knapsack reduction. Retaining pivotality requires either a finite atom or funding gaps that shrink with population size. The first leaves a distinguished individual outside the continuous population model; the second makes the result depend on vanishing, finely tuned margins. In the proposed formulation, the type distribution is largely bookkeeping: once the other residents’ aggregate contributions \(A_{tj}\) are given, the best-response optimization depends on those totals and the focal resident’s values and budget, not on the masses.
The strongest repair would make a positive-mass type cohort deviate together. Its aggregate contributions could cross project thresholds, and a project-selection problem might remain. But that asks about a coordinated coalition’s deviation, not an individual’s optimal strategy or a unilateral PSNE deviation. The paper’s simulations let a fraction of agents follow heuristics; they do not give that fraction a shared objective or coordinated choice. This could motivate a separate study of coalition stability in large crowds, but it is not a continuous mirror of Theorem 6.
So the negative case rests on the unit of agency, not on the possibility that hardness persists in the number of projects. The theorem gives a real computational result, but preserving its strategic question requires preserving a pivotal individual; removing that individual changes the question. The case is not airtight: if the programme accepts a finite focal agent against a continuous background as a sufficient continuization, the proponent’s mirror is coherent. I would argue that this hybrid leaves the population continuous only around the edges of the problem, while the computation of interest remains the original individual problem.
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.