| paper | To Lead or to be Led: A Generalized Condorcet Jury Theorem under Dependence |
| authors | — |
| venue | AAMAS 2024 |
| filed under | voting · theory |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper contains no numbered theorem asserting hardness, tractability, approximation, or parameterized complexity, so bit (a) fails under the explicit ChoCo rule. The proponent nevertheless identifies a plausible high-multiplicity sensor-fusion interpretation and a direct continuum limit of Theorem 2. That recognizable mirror cannot compensate for the absence of a qualifying computational anchor, so the grade is red.
fails bit a — no named computational result to mirror
The direct continuum winner-certification question captures Theorem 2's population-limit content; Theorem 3's finite-\(n\) concentration bound is not preserved without adding a population intensity. Theorem 1, Lemma 1, simulations, and cited prior results remain uncovered.
On the literal ChoCo anchor rule, this paper has no qualifying named computational result. Theorem 1 is the classical CJT, cited from Condorcet; Lemma 1 is Hoeffding’s inequality, also cited; and Theorems 2 and 3 are proved here but assert probabilistic convergence and sample-size bounds, not membership in \(P\), NP-hardness, W[1]-hardness, or a related complexity classification. Algorithms 1 and 2 and the simulations do not repair that gap. So a strict reading should say this plainly.
There is nevertheless a credible positive mirror if quantitative algorithmic theorems are admitted as anchors. My lead anchor would be Theorem 3, proved in this paper.
The natural regime is a large sensor-fusion electorate. There are many deployed sensor units observing the same underlying world, but only a small number of hardware/calibration classes. A type \(t\) is a complete conditional approval model: for each true world \(w\in W\), \(q_t^w(S)\) is the probability that a unit of type \(t\) privately approves exactly the subset \(S\subseteq W\). Units of the same type have the same competence profile and the same response to the opinion leader. The society is a rational distribution \(\mu\) over these types, where \(\mu_t\) is the fraction of deployed units of type \(t\). A finite clone expansion with \(N\mu_t\) units of each type recovers the corresponding discrete electorate whenever those quantities are integral.
The opinion leader remains a single common external signal, exactly as in the paper. With probability \(\pi\), a unit follows the leader’s approval set \(O\); otherwise it uses its private signal. Conditional on the true world, private signals are independent across units, as required by Definition 1. Thus only the population is continuized; approvals remain subsets of \(W\), and the alternatives themselves are not fractionalized.
The continuous problem is:
Minimum Certified Jury Mass\(_\infty\). The input is \(W\), a finite type set \(T\), rational masses \(\mu\), competence marginals \(p_t^w\), an opinion-leader competence \(\hat p\), influence strength \(\pi\), a reliability margin \(\Delta>0\), and a target confidence \(P_{\min}<1\). Let \(\Delta_\mu\) be a certified lower bound on the population-average competence gap, so \(\sum_t\mu_t p_t^w\ge \Delta_\mu+\sum_t\mu_t p_t^v\) for every competitor \(v\neq w\). Find the least total population mass \(M\ge0\) for which the paper’s worst-case success certificate reaches \(P_{\min}\).
Using the implicit bound in Theorem 3, this asks for the least \(M\) satisfying \(\hat p e^{-\frac12M\Delta_\mu^2(1-\pi)^2}+(1-\hat p)e^{-\frac12M(\Delta_\mu(1-\pi)-\pi)^2}\le\frac{1-P_{\min}}{m-1}\). The leader is safely usable when \(\pi<\frac{\Delta_\mu}{\Delta_\mu+1}\). If \(\hat p\) is unavailable, the explicit version asks for any \(M\) satisfying \(M\ge\frac{2}{(\Delta_\mu(1-\pi)-\pi)^2}\ln\frac{m-1}{1-P_{\min}}\).
This is recognisably the same question as Theorem 3: how large must the jury be before approval voting tracks truth with prescribed confidence under heterogeneous competence and common influence? The only change is that the heterogeneous electorate is represented by type masses rather than an indexed list of agents. The continuous problem is expected to be Class A: the bound is monotone in \(M\), so the optimum is obtained by a closed-form expression or one-dimensional approximation. It also has an exact high-multiplicity bridge: \(M=N\) and \(N\mu_t\in\mathbb Z\) yields the corresponding repeated-agent finite instance.
A secondary anchor is Theorem 2, also proved here. Its mirror is Continuum Opinion-Leader Winner Certification. For a true world \(w\) and leader signal \(O\), the continuum score of alternative \(a\) is \(s_\mu^{w,O}(a)=\sum_t\mu_t\bigl(\pi\mathbf 1[a\in O]+(1-\pi)q_t^w(a)\bigr)\), where \(q_t^w(a)=\sum_{S\ni a}q_t^w(S)\). The question is whether \(w\) is the unique winner, meaning \(s_\mu^{w,O}(w)>s_\mu^{w,O}(a)\) for every \(a\neq w\), for every leader signal in the specified worst-case support. More generally, with a leader-signal distribution \(r_w\), output the limiting success probability \(\sum_O r_w(O)\mathbf 1[w\text{ uniquely wins}]\), and decide whether it is at least \(P_{\min}\).
This is the exact infinite-population limit of the paper’s process. For each fixed leader state, the empirical scores of a repeated finite population converge to these type-weighted expectations. Under the paper’s condition \(\pi<\frac{\Delta_\mu}{\Delta_\mu+1}\), the correct world has a positive limiting margin even when the leader is wrong, yielding the convergence claimed in Theorem 2. The certification problem is Class A: with explicit type tables, it requires only weighted sums and comparisons. Its useful extensions are exact finite-\(N\) tail computation, robustness to perturbations of \(\mu\), minimum mass removal needed to reverse the winner, and choosing a sensor mix \(\mu\) under procurement costs.
The authors should recognise this as their problem’s continuous analogue. Their own motivation explicitly includes sensor fusion; their mathematical object is already an electorate of abstract agents with heterogeneous competence, approval voting, and a common opinion leader. The mirror preserves all of those ingredients and changes only the population representation. It is not merely outcome-space continuity or a generic mean-field opinion model.
The weakness is substantial. The paper’s central results are analytic probability statements, and the continuous limit makes the stochastic aggregation almost trivial through concentration. Theorem 3’s mirror is largely an inversion of an exponential inequality rather than a new complexity landscape. More importantly, neither Theorem 2 nor Theorem 3 is a qualifying computational anchor under the programme’s strict rule. I would therefore present this as a plausible Class-A seed for a continuous population problem, strongest around Theorem 3, but not as a paper that already supplies a formal ChoCo complexity result. The mirror covers only Theorems 2 and 3; it does not cover the cited CJT literature, the simulation results, or the algorithms as computational-theory contributions.
The negative case is decisive under ChoCo’s stated anchor rule: this paper contains no qualifying computational result. Theorem 1 and Lemma 1 are imported classical results. Theorems 2 and 3 prove asymptotic probability statements and concentration bounds, not complexity, optimization, approximation, or parameterized-algorithm theorems. Algorithms 1 and 2 are simulation procedures for estimating a finite sample size. There is therefore no named computational problem here whose continuous complexity landscape the programme could chart.
Even granting the proponent’s broader reading, Theorem 3 does not survive as the proposed continuous problem. The sensor-class story is a perfectly legitimate high-multiplicity regime; that is not the objection. The problem is that \(M\) is simply \(n\) under another name. ChoCo’s society is a normalized distribution \(\mu\), so replacing \(N\mu_t\) cloned sensors by \(M\mu_t\) cloned sensors leaves \(\mu\) unchanged. The exponential terms in Theorem 3 arise entirely from finite-population sampling fluctuations.
In the genuine nonatomic limit, conditional on the world and the leader’s signal, each type’s vote share is its expectation, and the aggregate score is
\[ s_\mu^{w,O}(a) = \sum_t \mu_t \left( \pi \mathbf{1}[a\in O] + (1-\pi)q_t^w(a) \right). \]
There is no remaining \(M\)-dependent concentration probability. The correct world either has the required deterministic margin or it does not. To retain the exponential bound, one must retain a finite sample size, or introduce a Poisson population intensity. That is a stochastic finite-population model, not a continuous society in ChoCo’s sense. The exact clone correspondence is a valid high-multiplicity observation, but it does not create a new continuous computational problem.
Moreover, Theorem 3 compresses the entire heterogeneous population into the single average gap \(\Delta_\mu\). Two type distributions with the same relevant averages are indistinguishable to the theorem. Thus the proposed “minimum certified jury mass” does not use a type-space optimization, mass transfer, or any structural property of a continuous population; it merely inverts a scalar inequality already stated in \(n\).
Theorem 2 has the same defect. The proposed continuum winner-certification problem is the deterministic law-of-large-numbers limit of the theorem, not a computational mirror of its content. It removes precisely what Theorem 2 studies: how finite-population success probability approaches one as \(n\) grows. If the opinion leader remains random, the resulting probability is only a finite weighted sum over leader signals; conditional on a signal, the population outcome is deterministic. Reintroducing a confidence threshold through finite-\(M\) sampling simply returns to the previous finite-jury problem.
One could optimize the sensor mixture \(\mu\), remove mass from types, or impose procurement costs. Those may be sensible new continuous control problems, but they are not mirrors of Theorems 2 or 3: they replace truth-tracking asymptotics with intervention and design. Conversely, preserving the paper’s actual question forces one either to restore \(n\) or to accept a deterministic mean-field predicate. Enlarging the type space from finitely many classes to a continuum only replaces sums by integrals; it does not restore the missing population-scale computational question.
The negative case is not that sensor populations cannot have repeated types. They plainly can. It is that this paper’s contribution is entirely about finite-sample probability and its asymptotic limit, with no computational result to continuize. Under a broad “analytic quantitative theorem” reading, the sensor example is a plausible Class-A seed; under the announced ChoCo standard, however, both proposed anchors collapse into either a renamed finite-\(n\) calculation or a deterministic limit, and neither supplies a worthwhile continuous computational mirror.
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.