| paper | d4-conf-ijcai-2021-0385 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 12 gives a closed-form sensitivity result used to calibrate a privacy mechanism, but it does not state an algorithm, complexity guarantee, or computational formulation for a decision problem. The district-type proposal changes the population being modeled and suggests an analytic fairness study, not a computational mirror anchored in the paper.
fails bit a — no named computational result to mirror
No result supplies a qualifying computational anchor; the remaining named results characterize fairness, bias, privacy, and post-processing.
I can’t make a positive case that meets the anchor requirement. The paper has named theorems and corollaries, but none asserts a complexity or algorithmic result for an underlying problem. They characterize fairness, bias, or differential privacy; the paper does not study worst-case complexity or algorithms of the kind ChoCo is mapping. So there are no qualifying anchors, and no per-anchor continuous problem to hand to a prover.
The closest result is Theorem 12, proved in this paper. It states that for the Title I allocation query
\[
P^F_i(\bm{x})=\frac{x_i a_i}{\sum_j x_j a_j},
\]
if \(a_{\max}=\max_i a_i\) and a known public lower bound \(L\) satisfies \(L\leq\sum_i x_i a_i\), then the query’s \(\ell_1\) global sensitivity under a one-person change is \(\Delta=2a_{\max}/L\). That is a useful closed-form sensitivity result, but it is not a complexity classification of an allocation problem. The other named results—for example, Theorem 4’s bounded-fairness condition and Theorems 10–11’s post-processing results—are likewise analytic, not computational-complexity anchors.
There is a plausible *adjacent* continuous angle: model many similar school districts by a distribution over district types, with mass representing the fraction of districts of each type, and study the type-level funding shares and their bias under a specified noise model. But that would change the paper’s finite collection of jurisdictions and its person-level privacy relation. It would ask for an analytic fairness or sensitivity characterization, not a computational question anchored by a result here. Fractional funding shares and Laplace noise do not themselves continuize the population.
So my strongest honest positive assessment is that the paper could inspire a new continuous-population fairness question, but it does not provide a named computational result that can anchor the requested mirror.
The strongest case against a mirror is that the paper contains no named computational result to mirror. Its theorems characterize bias, fairness, and post-processing effects; they do not classify the complexity of a problem or give an algorithm for one. Theorem 12, the closest candidate, gives a closed form for the \(\ell_1\) sensitivity of a query under a one-person change. That is an analytic bound, not a complexity result.
The district-type idea does not repair that gap. In the paper, people are the records protected by differential privacy, while counties and school districts are the entities whose allocations or decisions are compared. Turning residents into a distribution over types leaves the outputs indexed by the same finite set of jurisdictions; the input counts are already real-valued. Turning jurisdictions into types instead changes the population being continuized. It can also change the fairness question: averaging bias over jurisdiction mass can hide a badly affected jurisdiction, while preserving the paper’s worst-case comparison keeps each distinct jurisdiction relevant.
The strongest alternative is to formulate a new problem around the paper’s mitigation discussion: choose a privacy mechanism, temperature, or type-specific proxy to minimize disparity subject to accuracy and privacy constraints over a distribution of jurisdictions. That could be worth studying. But it adds an optimization problem that the paper does not solve or classify; the paper’s proposed temperature optimization is a future-work objective, not a named computational result. Calling this a mirror would make the paper a source of inspiration rather than the object of continuization.
So the negative case is strong for this paper as an anchor for ChoCo’s computational-complexity programme. It is weaker as the universal claim that no related continuous-population project could be worthwhile: a carefully specified mechanism-design or optimization problem might be. That would be a new computational project, though, not a continuous mirror of the results presented here.
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.