d4-conf-atal-2013-0315

The strongest mirrorable contribution is Algorithm 1, MaxTrust, which the authors introduce. There is no numbered Theorem, Lemma, Corollary, or Proposition asserting a complexity classification, so this is not a formal anchor of the kind your criteria request. The closest computational result is the algorithm and its accompanying runtime claim: it initializes trust scores, repeatedly updates one user’s score at a time by a likelihood-maximizing line search, and stops when the scores change by less than an error bound or the epoch limit is reached. It then returns the trust scores and the fused estimate’s mean and precision. The authors state a runtime of \(O(\text{epochs}\times k|S|)\), where \(k\) is the number of users and \(|S|\) the number of line-search samples; they describe the resulting solutions as suboptimal, compared with an exhaustive search of order \(O(|S|^k)\). This is their own algorithm and claim, not a result cited from elsewhere.

A concrete continuous version is Continuous MaxTrust. For one target and the paper’s single-report-per-user setting, an instance gives a total report mass \(N\), a finite support of \(\tau\) report types \(t\), and rational masses \(\mu_t\) summing to \(1\). Each type is a complete report record, including its reported location \(x_t\in\mathbb{Q}^2\), precision \(\theta_t>0\), and any other input feature used by the model. A type’s mass is the fraction of the crowd producing that report record. The decision variables are trust scores \(q_t\in[\epsilon,1]\), with \(\epsilon>0\) as in the paper.

For a trust vector \(q\), define the fused mean and precision by
\[ x_\mu(q)= \frac{\sum_t\mu_tq_t\theta_tx_t}{\sum_t\mu_tq_t\theta_t}, \qquad \Theta_\mu(q)=N\sum_t\mu_tq_t\theta_t. \]
Let \(p_t(z\mid q_t)\) be the report’s trust-scaled Gaussian density and \(f_\mu(z\mid q)\) the fused Gaussian with mean \(x_\mu(q)\) and precision \(\Theta_\mu(q)\). The problem is to maximize
\[ \sum_t N\mu_t\log\!\left( \int_{\mathbb{R}^2}p_t(z\mid q_t)f_\mu(z\mid q)\,dz \right) \]
over \(q\), and return an optimizing trust vector and its fused estimate. This replaces the paper’s sums over individual reports by mass-weighted sums; when \(N\mu_t\) is an integer for each type, it recovers the corresponding repeated-report instance. The target remains a point to estimate. It is the population of observers that becomes the continuous object.

A plausible regime is a large cell-mapping service receiving many detections of the same tower. If reports are rounded to GPS and accuracy bins, large cohorts of handsets can produce identical records, with \(N\) much larger than the number \(\tau\) of occupied report types. In that regime, replacing counts by report mass is faithful to the model: the fused mean, uncertainty scaling, and trust-based likelihood all use the reports. I expect the finite-epoch, sampled-line-search analogue to be tractable in this regime, with the algorithmic work scaling in \(\tau\) instead of \(k\). That is a practical, Class A-like expectation, not a proof that the global maximum-likelihood problem is in P. The paper gives no global-optimality or approximation guarantee; exact optimization could be continuum-specifically hard. There is no discrete hardness result here from which to expect Class B.

This mirror covers the paper’s trust inference and fusion problem, especially its single-report cell-localization setting. It does not cover the empirical claims that MaxTrust improves accuracy or consistency, nor does it make the target’s continuous location an argument that the *society* is continuous. Further questions include whether mass aggregation preserves the optimizer, whether coordinate updates converge reliably, what approximation guarantees are possible, and how results change as the total report mass \(N\) grows.

The weakest point is the homogeneity assumption the mirror needs. The paper estimates trustworthiness for individual users, while the proposed types pool users with identical report records and assign them a shared trust score. That is defensible within the paper’s model when those records make the users indistinguishable to the inference procedure, but it does not capture unobserved individual differences in reliability. The authors’ own evaluation—1,563 reports across 129 towers, or 6–46 reports per tower—also does not demonstrate a high-multiplicity regime. The mirror is strongest as a plausible extension of their model, not as a regime their experiments establish.