| paper | d4-conf-atal-2015-0261 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
other→P discrete result → its continuous mirror
Answer: Computable in deterministic time polynomial in \(m,\tau,L,k\). · discrete side: other · adjudication: proved
full proof & adversarial review →
Theorem 1
statement extracted from the paper’s text layer
Given a finite target-strategy set \(F\), rational prior \(\pi\), finite set \(T\) of complete recommender, report, and context types, rational masses \(w_t\), and likelihoods \(\ell_t(f)\), compute the normalized weighted-evidence valuation \(\nu(f)=\frac{\pi(f)\prod_{t\in T}\ell_t(f)^{w_t}}{\sum_{g\in F}\pi(g)\prod_{t\in T}\ell_t(g)^{w_t}}\). Integer masses recover aggregation of discrete records; fractional masses use the power-likelihood extension.
A population of recommendation records represented by masses over complete recommender, report, and context types, with target strategies \(f\in F\); combine their likelihood contributions to compute a normalized weighted-evidence valuation.
Theorem 1 does not derive the fractional power-likelihood rule, leaving an open question about whether authors would treat weighted evidence as their aggregation problem.
fatal: False
The mirror extends Theorem 1's aggregation of recommendations; it leaves direct-observation updates, the broader argument against endogenous discounting, and the discussion of collusion untouched.
The strongest positive case is a continuous version of recommendation aggregation: a user receives recommendations from a very large, low-diversity population of sources and infers how a target will behave. That keeps the paper’s central question—how to form an opinion about a target from evidence—while making the population of recommenders the continuous object.
The closest anchor is Theorem 1, proved here (and presented as a generalization of earlier work on Beta models). It says that if \(e^x_1,\ldots,e^x_n\) are evidence about target \(x\), and \(e^x\) is shared evidence about other targets, then the valuation from all the evidence equals the aggregation of the valuations from each \(e^x_i\), each conditioned on that same \(e^x\):
\[
[[\tau_x(e^x_1,\ldots,e^x_n,e^x)]]
=
[[\tau_x(e^x_1,e^x)\triangleright\!\!\triangleleft\cdots\triangleright\!\!\triangleleft\tau_x(e^x_n,e^x)]].
\]
The theorem’s setting assumes independent, non-colluding recommenders. Its result is an exact decomposition of the opinion computation.
There is an important qualification: Theorem 1 is not a computational-complexity result in the sense required for an anchor. It asserts a semantic identity, not that a named problem is in P, NP-hard, or in another complexity class. The paper has no qualifying complexity anchor. Theorem 1 is the strongest nearby result for a positive case, but calling it a complexity result would overstate what the paper proves.
A concrete candidate mirror is Continuous Recommendation Aggregation. Fix a finite set \(F\) of possible target strategies, a rational prior \(\pi\) over \(F\), and a finite set \(T\) of complete recommender-and-report types. A type includes the source’s relevant reporting strategy, its report, and the context used by the model. For each \(t\in T\) and \(f\in F\), the instance supplies the likelihood \(\ell_t(f)\) of that report type if the target has strategy \(f\).
The recommender population has size \(N\) and distribution \(\mu\) over \(T\), with rational masses summing to one and support size \(\tau\). The task is to return the exact posterior valuation over \(F\):
\[
\nu(f)=
\frac{\pi(f)\prod_{t\in T}\ell_t(f)^{N\mu_t}}
{\sum_{g\in F}\pi(g)\prod_{t\in T}\ell_t(g)^{N\mu_t}}.
\]
The solution is this distribution; there is no optimization objective. When every \(N\mu_t\) is an integer, it is exactly the posterior for that many independent recommendation records of each type. Fractional counts extend the likelihood update to continuous mass. The probabilities in \(\pi\) remain uncertainty about target strategy; the continuous object is the recommender population itself.
A plausible regime is a large marketplace with millions of reviews, a standardized rating system, and a relatively small number of reporting-policy, rating, and context types. Then \(N\) can be large while \(\tau\) stays small. That scenario is plausible only where sources sharing a type really are interchangeable for this inference task and their reports satisfy the independence assumptions. Theorem 1 supports the aggregation step for discrete records; it does not prove the fractional-mass formula.
For an explicit finite \(F\) and likelihood table, I would expect this restricted problem to be tractable: compute the type contributions for each \(f\), then normalize. That is a prediction about the proposed mirror, not a result of this paper. Hardness for a succinct or continuous strategy space, correctness of the fractional extension, and the effect of collusion are further questions. The discrete-to-continuous bridge is exact when the masses represent integer counts; the continuous problem’s behavior with genuinely fractional counts needs its own justification.
This mirror covers the paper’s recommendation-aggregation result, not its full argument about endogenous discounting or its discussion of collusion. Its weakest point is substantial: fractional recommendation counts are generalized evidence weights, not literal observations, and the paper does not establish that authors should regard them as the same problem. Also, the paper studies one user’s inference about a target, rather than a collective choice among alternatives. If ChoCo requires that collective-choice structure, this is an edge case at best.
Theorem 1 is the only plausible anchor, and it is a semantic identity, not a computational-complexity result. Proposition 1 and the paper’s other formal results have the same character: they say how opinions relate, not how hard a named computational problem is or what algorithm solves it. The divide-and-conquer claim may motivate an implementation, but the paper does not analyze its complexity. There is no result here about voting, collective choice, manipulation, control, or winner robustness to continuize.
The proposed posterior-computation task is a sensible adjacent question, but it does not repair that gap. With an explicit finite strategy set \(F\) and explicit likelihood table, it asks for a posterior from a formula. Any interesting complexity would need to come from changing the representation—for example, making \(F\) succinct or adding latent unknowns. That may be worthwhile research, but the paper’s theorem neither poses nor supports those harder variants. The concern is not that an algorithm might be easy; it is that the computational challenge would be newly supplied rather than mirrored from the paper.
There is also a mismatch in what gets made continuous. The theorem aggregates evidence records from independent recommenders. If a recommender type includes its realized report, the proposed \(\mu\) is a histogram of evidence, not a distribution over stable agent types. If type instead means a reporting strategy and context, then \(\mu\) alone does not say which reports were observed; those records must be supplied separately. For integer counts, the proposed expression is just the ordinary posterior over a compressed finite dataset. For fractional counts, \(\prod_t \ell_t(f)^{N\mu_t}\) is a power likelihood: it can define a useful weighted-evidence model, but it is not the posterior for fractional independent observations, nor an identity established by Theorem 1. A genuine growing sample also makes posterior concentration, rather than a nontrivial limiting opinion, the generic outcome when strategies are identifiable.
Most fundamentally, this is an individual user inferring a target’s strategy, not a society choosing among alternatives. Recasting recommenders as a large population makes a plausible high-multiplicity trust-inference problem, but it does not make that problem computational social choice in ChoCo’s stated sense. Treating reports as ballots would change the model’s purpose: their likelihood product is evidence pooling, not a collective choice rule.
So the negative case is strong for this programme’s scope: the paper offers no qualifying computational result, and the strongest repaired anchor becomes a new continuous inference model outside collective choice. It is not airtight as a claim that no worthwhile mirror exists anywhere. A broader programme that includes epistemic pooling could reasonably study weighted or limiting evidence aggregation, and the marketplace scenario makes that direction plausible. That possibility is the weak point in the negative case; it does not, however, supply a ChoCo mirror of a computational result in this paper.
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.