| paper | d4-conf-wine-2016-0021 |
| 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: \(\Delta^*(\nu)=\max\{G_{\mathrm I},G_{\mathrm{II}}\}>0\), with the explicit quantities defined above. The supremum is attained exactly when \(z_0>H\) or \(G_{\mathrm I}\geq G_{\mathrm{II}}\). Deciding \(\Delta^*(\nu)>g\) is in \(\mathrm P\). Every admitted prior admits arbitrarily near-optimal focal mechanisms with penalty \(c=1\); an exactly optimal focal mechanism exists precisely when the supremum is attained. · discrete side: other · adjudication: proved_with_minor_gaps
full proof & adversarial review →
Theorem 3.5
Given a rational finite-support prior \(\nu\) over latent qualities \(p\in(0,1)\), with \(\operatorname{Var}_{\nu}(p)>0\) and \(\mathbb{E}_{\nu}[p]\ne 1/2\), each realized atomless society has signal-type masses \(\mu_1=p\) and \(\mu_0=1-p\). Choose a normalized strictly proper peer-scoring rule and a penalty triggered by unanimous aggregate reports that maximize truth-telling's expected payoff gap over other informative atomless Bayesian Nash equilibria, and determine whether the penalty makes truth-telling strictly higher-paying than every other equilibrium, including equilibria with identity-based splits at indifference.
The society is a continuum of users whose signal types carry masses \((1-p,p)\), with its composition drawn from \(\nu\). Users report their signal and are paid by a strictly proper score against a randomly sampled peer's report; the decision variables are the score and a penalty on unanimous aggregate reports, and the objective is the truth-telling payoff gap over other informative equilibria.
The finite-player proof does not establish focality in the atomless game: its unanimity-probability threshold disappears, and aggregate equilibria with identity-based splits at indifference require fresh payoff comparisons.
fatal: False
The mirror covers Theorem 3.5's binary focal-mechanism design and Theorem 4.1's binary equilibrium characterization; it leaves the paper's broader non-binary and unknown-prior directions untouched.
The paper has no named theorem that classifies a problem as NP-hard, in P, or FPT. Its strongest computationally useful results are constructive: they characterize equilibria and design scoring mechanisms. The case for a mirror is therefore about continuous-population mechanism design, not a complexity result the authors actually prove.
A plausible regime is a large review platform asking a single subjective question about the same product. A latent quality \(p\) is drawn from a known prior \(\nu\); conditional on \(p\), a fraction \(p\) of a continuum of users receive signal \(1\), and the rest receive \(0\). Each user reports a signal and is paid using a randomly sampled peer’s report. A type is the user’s signal \(b\in\{0,1\}\): users with the same signal have the same posterior beliefs and incentives. Each realized society has masses \(\mu_1=p\) and \(\mu_0=1-p\), so there are arbitrarily many users but only two types. For example, if \(\nu\) puts equal probability on \(p=2/5\) and \(p=4/5\), then \(q(1\mid1)=2/3\) and \(q(1\mid0)=1/2\), giving positively correlated, signal-asymmetric signals.
The uncertain population composition is a necessary part of the mirror: peer prediction relies on correlated signals, which a fixed, publicly known type distribution alone would not provide. It is still the population that becomes continuous. In a realized society, the report distribution is an aggregate mass; an individual has zero mass, but can respond to the expected fraction of peers reporting \(1\).
My lead is Continuum Focal Peer-Prediction Design, mirroring Theorem 3.5 (stated informally in §3 and restated formally in §7; proved in this paper). The theorem starts with a symmetric, positively correlated, signal-asymmetric binary prior \(Q\). For \(n\) agents, let \(\varepsilon_Q\) be the greatest probability that a fixed set of \(n-1\) agents all receive the same signal. The paper constructs a normalized peer-prediction mechanism \(PPM(Q)\) in which truth-telling pays at least as much as every informative equilibrium. Its scoring scheme maximizes the truth-telling payoff gap over other informative equilibria for priors in \(R_1\) and \(R_2\), and approaches the supremum for priors in \(R_3\). It then adds a penalty when all other agents report the same bit. This does not change the equilibria, and truth-telling becomes strictly higher-paying than every other equilibrium if
\[
\varepsilon_Q < \frac{\Delta^*(Q)}{1-t+\Delta^*(Q)},
\]
where \(t\) is the truth-telling payoff in \(PPM(Q)\) and \(\Delta^*(Q)\) is the optimal informative-equilibrium gap. The theorem also states that the threshold depends only on \(q(1\mid1)\) and \(q(1\mid0)\).
The continuous problem is: given a finite rational prior \(\nu\) over rational \(p\in(0,1)\), choose a strictly proper scoring rule with peer-prediction payments in \([0,1]\) to maximize the truth-telling payoff minus the highest payoff of any other informative equilibrium. Return an optimal rule, or a sequence approaching the supremum if none attains it, and a penalty for unanimous aggregate reports that makes truth-telling focal among all equilibria. The solution must specify the score, penalty, and equilibrium-payoff comparisons.
The expected direction is tractable. Under truth-telling, each realized society with \(p\in(0,1)\) has a non-unanimous report distribution, so the continuum truth-telling profile incurs no unanimity penalty. Constant-report equilibria are unanimous and do incur it. The finite-population probability correction \(\varepsilon_Q\) therefore disappears in this regime. The informative-equilibrium gap still depends on the same posterior probabilities \(q(1\mid0)\) and \(q(1\mid1)\); the paper’s explicit three-region score construction gives a strong route to the continuous optimization. Further questions include how the exact optimum depends on the support size and bit length of \(\nu\), and how the continuum mechanism approximates large finite cohorts.
A useful second anchor is Theorem 4.1, also proved here. For a symmetric, positively correlated binary prior and a strictly proper scoring rule, define \(q^*\) as the break-even point where an agent is indifferent between reporting \(0\) and \(1\). The theorem says there are no asymmetric equilibria and lists all equilibria, determined by \(q^*\): always-report-\(0\), always-report-\(1\), truth-telling \((0,1)\), and the constant randomized strategy \((q^*,q^*)\); also \((0,q^*/q(1\mid1))\) and \(((q^*-q(1\mid0))/q(0\mid0),1)\); plus \((1,0)\) if \(q(0\mid1)\le q^*\le q(0\mid0)\), \((1,(q^*-q(0\mid1))/q(1\mid1))\) if \(q(0\mid1)\le q^*\), and \((q^*/q(0\mid0),0)\) if \(q^*\le q(0\mid0)\). There are seven to nine equilibria.
Its continuous mirror, Continuum Peer-Prediction Equilibrium Characterization, asks: given \(\nu\) and a strictly proper score, characterize every equilibrium of the atomless peer-prediction game, including its aggregate report masses and the measurable strategy profiles that generate them. The expected result is a finite list of aggregate equilibria governed by the same break-even conditions, together with a description of how agents can vary their behavior when indifferent. This is a tractable characterization problem; the mirror should not claim that Theorem 4.1’s “no asymmetric equilibria” conclusion survives literally.
That is the main soft spot. In an atomless game, a single user cannot change the aggregate, so agents can differ in their strategies whenever the aggregate makes them indifferent. The sensible continuous analogue is to classify aggregate outcomes and all their strategy-profile realizations. The paper’s full finite-agent classification makes that extension credible, but it does not prove it. The mirror covers binary peer-prediction equilibrium and focality, not the paper’s non-binary or unknown-prior directions.
The strongest case against spending ChoCo’s time here is programme fit. Theorem 3.5 optimizes a binary peer-prediction payment rule and adds a penalty; Theorem 4.1 lists equilibria through a break-even point \(q^*\). Neither gives a complexity classification or exposes a large computational problem. In the proposed finite-support latent-quality model, the relevant posteriors come from a few moments of \(\nu\), and the scoring rule follows from the paper’s formulas. The continuous versions may be useful extensions, but they appear to add little algorithmic territory.
There is a real modeling step in the Theorem 3.5 mirror. A fixed, known society distribution \(\mu\) does not provide the correlated private signals peer prediction needs. To preserve those incentives, the model must make the population composition uncertain, such as drawing \(p\) from \(\nu\) and having a fraction \(p\) of the population receive signal \(1\). That is a coherent and plausible large review platform, though it makes the instance a distribution over societies. It also covers only a subset of the symmetric priors in the theorem. This narrows the mirror; it does not rule it out.
For that model, the paper already supplies most of the score optimization. The new continuum feature is mainly that, when \(p\) stays away from \(0\) and \(1\), truthful reports have interior aggregate masses, so the finite population’s unanimity correction disappears. That yields a clean asymptotic design question, but no evident new complexity challenge. Its best case is therefore as a simple continuous limit result, rather than as a contribution to ChoCo’s computational landscape.
Theorem 4.1 has a similar limitation. At the aggregate level, the continuum equilibrium conditions still compare peer report rates with \(q^*\), so the listed fixed points provide a direct starting point. An atomless model also permits many measurable strategy profiles when a signal type is indifferent. The proponent is right to ask for those profiles as well as the aggregate outcomes. That is a legitimate refinement, though the added description seems to be how indifferent agents split across reports, not a new computational problem.
These objections support deprioritizing the paper for ChoCo; they do not defeat both anchors. The latent-composition platform is a natural high-multiplicity regime, and neither identity dependence nor continuum degeneracy blocks it. I cannot honestly defend the universal claim that no worthwhile continuous mirror exists. The negative case is weak on that burden: it questions the programme value of two likely low-dimensional extensions, while the proposed mirrors remain well-posed.
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.