| paper | The Semi-random Likelihood of Doctrinal Paradoxes |
| authors | Ao Liu, Lirong Xia |
| venue | AAAI 2022 |
| filed under | frontier · ja |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c22r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper fails bit (a): its numbered results classify likelihoods but do not state computational problems or complexity guarantees. The proposed \(\mathrm{CSPR}_\infty\) is a coherent high-multiplicity reformulation, but it is a new computational stochastic-optimization problem rather than a mirror of a computational result in the paper. Therefore no anchor can make the paper green.
fails bit a — no named computational result to mirror
The paper supplies no named computational problem or complexity assertion; solving the proposed mass-optimization and rate-classification problem would require an additional result, while \(N\to\infty\) removes the stochastic likelihood.
fatal: True
The proposed mirror covers the max/min semi-random likelihood classification of Theorem 1 and the singleton-\(\Pi\) specialization in Corollary 2; it leaves exact finite-\(N\) probabilities, broader agendas, and the unnumbered verification discussion outside its scope.
The honest starting point is that this paper contains no named computational-complexity theorem in the strict ChoCo sense: no numbered result proves membership in \(\mathrm P\), NP-hardness, FPT, or similar. Theorem 1 is nevertheless a usable substantive anchor. It is proved in this paper, not merely cited, and gives the complete four-way asymptotic classification of semi-random doctrinal-paradox likelihoods. Corollary 2 is supporting evidence for the same anchor, not an independent one.
The strongest mirror is a high-multiplicity version of that theorem, which I would call Continuous Semi-random Paradox Rate, \(\mathrm{CSPR}_\infty\).
Fix \(p\), \(V=\{0,1\}^p\), a logical connection \(f:V\to\{0,1\}\), a quota rule \(r_{\mathbf q,\mathbf d}\), and a finite set of strictly positive rational distributions \(\Pi=\{\pi^1,\ldots,\pi^\tau\}\) over \(V\). A type \(a\) is the complete latent judgment/noise class represented by \(\pi^a\): agents of that type have the same distribution over judgment vectors and the same role in the aggregation problem. A society is a mass vector \(\mu\in\Delta_\tau\), where \(\mu_a\) is the fraction of the population belonging to type \(a\).
At scale \(N\), a rational mass vector with \(N\mu_a\in\mathbb Z\) represents \(N\mu_a\) agents of type \(a\). Each agent independently draws a judgment vector from \(\pi^a\). Let \(H_{N,\mu}\) be the resulting histogram and define \(F_N(\mu)\) as the probability that \(r_{\mathbf q,\mathbf d}(H_{N,\mu}/N)\) is inconsistent with \(f\). The max version asks for the asymptotic rate of
\[ F_N^{\max}=\max_{\mu:\,N\mu\in\mathbb Z^\tau}F_N(\mu), \]
and the min version asks for the analogous minimum. A solution must return the applicable rate, allowing residue or parity effects, together with a witnessing mass sequence or a certificate that all mass vectors lie in the relevant safe region. The permitted answers are exactly the four classes in Theorem 1: \(0\), \(\exp(-\Theta(N))\), \(\Theta(N^{-1/2})\), or \(\Theta(1)\).
This is not a cosmetic rewriting. For finite \(\Pi\), an assignment \((\pi_1,\ldots,\pi_N)\in\Pi^N\) matters only through the multiplicity of each type. Thus \(\mu\) is precisely the high-multiplicity representation of the paper’s adversarial assignment. The random judgments remain binary and individual; only the society of agents becomes continuous. Nor is \(\mu\) a fractional vote or a fractional outcome.
The regime is plausible in large-scale online deliberation, crowdsourced legal review, or AI-assisted e-democracy. Millions of participants may evaluate the same fixed agenda of \(p\) propositions, while belonging to a small number of calibration, training, jurisdictional, or reliability classes. The number of agents is then much larger than the number \(\tau\) of distinct behavioral types. Conditional noise remains independent exactly as in the paper, while the mass vector permits arbitrary aggregate correlations in the latent population. This is also close to the paper’s own motivation involving future large-scale judgment aggregation.
The connection to Theorem 1 is especially strong because its proof already identifies the relevant continuous object. For a mass vector \(\mu\), the center of the population is
\[ \bar\pi_\mu=\sum_{a=1}^{\tau}\mu_a\pi^a. \]
As \(\mu\) ranges over the simplex, \(\bar\pi_\mu\) ranges over \(\operatorname{CH}(\Pi)\), exactly the convex hull used in the theorem. The paradox event is a union of polyhedra defined by quota inequalities. Hence the continuum problem asks where the population mass lies relative to those polyhedra and their boundary faces. The four likelihood regimes then have a natural population interpretation: no feasible mass produces a paradox, paradox regions are exponentially separated, the population can reach a knife-edge boundary, or it can occupy a robust paradox region.
I would expect the asymptotic-rate version to be Class A for explicitly listed type distributions, with polynomial dependence on the truth-table size and an FPT dependence on \(\tau\). The paper’s unnumbered computational-verification discussion already reduces the \(\kappa_2\) and \(\kappa_3\) checks to linear programs and the \(\kappa_1\) check to fixed-dimension integer programming. That is not itself a named complexity theorem, so it cannot be advertised as an established ChoCo result, but it is strong evidence that the mass formulation exposes a tractable geometric core. The exact finite-\(N\) probability, rather than its asymptotic class, may be substantially harder and deserves a separate complexity analysis.
The mirror covers Theorem 1’s max- and min-semi-random likelihood classification, and in the singleton case \(\Pi=\{\pi\}\) it also covers the phenomenon summarized by Corollary 2. It does not cover every possible closed infinite \(\Pi\), nor does it turn the paper into a hardness result. It also leaves useful follow-up questions: whether exact likelihood computation is \(\#\mathrm P\)-hard, whether the rate classification is FPT in the number of types, how accurately finite populations can be rounded from continuum masses, and whether succinct representations of \(f\) or \(\Pi\) create continuum-specific hardness.
The weakest point is fundamental: the paper studies probability under independent noise, whereas a literal continuum of agents obeys a law-of-large-numbers limit in which the aggregate histogram becomes deterministic. Keeping \(N\) as a sampling scale preserves the paper’s likelihood question, but produces a two-level model rather than a purely nonatomic one; removing \(N\) collapses the objective largely to a deterministic paradox predicate. Moreover, because the paper has no named computational-complexity result, this is best presented as a credible, faithful population mirror of Theorem 1—not as a completed ChoCo complexity anchor.
The central objection is decisive under ChoCo’s stated standard: this paper contains no named computational result to continuize. Theorem 1 classifies an asymptotic probability; it does not define an algorithmic input, a decision or optimization problem, or a complexity bound. Corollary 2 is the same kind of probabilistic statement. The later discussion of checking \(\kappa_1\)–\(\kappa_4\) is neither a numbered theorem nor a complexity result in the natural parameters: it treats \(\ell=|\Pi|\) as fixed and measures polynomiality in the explicit truth-table size \(m=2^p\). It gives no \(\mathrm{poly}(p,\tau,L)\) or parameterized guarantee. A continuous computational problem could certainly be invented from this material, but it would be a new problem, not a mirror of a computational theorem in the paper.
The proposed CSPR also has a deeper modelling defect. If there are \(N\mu_a\) agents of type \(a\), with independent votes drawn from \(\pi^a\), then
\[ \frac{H_{N,\mu}}{N}\longrightarrow \bar{\pi}_{\mu}=\sum_a \mu_a\pi^a \]
almost surely. A genuinely nonatomic population therefore has a deterministic aggregate histogram. Once \(\bar{\pi}_{\mu}\) is fixed, the quota-rule outcome—and hence whether it is inconsistent—is fixed, including at a threshold once the tie-breaking convention is specified. The four regimes \(0\), \(e^{-\Theta(N)}\), \(\Theta(N^{-1/2})\), and \(\Theta(1)\) are fluctuations of a finite population around its law-of-large-numbers limit. They are not likelihoods of the continuous society itself.
Retaining \(N\) does preserve those fluctuations, but it also shows that the proposed mirror is exactly a re-encoding of the paper’s existing model. For every rational \(\mu\) with denominator \(N\), \(F_N(\mu)\) is the probability associated with an assignment containing \(N\mu_a\) copies of \(\pi^a\). The mass vector merely groups interchangeable assignments. That grouping is legitimate high multiplicity—I do not object to \(\pi^a\) being a type—but no new population-level object has entered. Indeed, the paper already passes to the barycenter \(\bar{\pi}\in\operatorname{CH}(\Pi)\) and represents paradox events as unions of polyhedra. The proposed continuum is the bookkeeping quotient of that analysis, with the finite sampling scale left in place.
The strongest repair would be to study a large-deviation rate function \(R(\mu)\), or exact finite-\(N\) probability as a function of the mass vector. That could be a worthwhile new stochastic optimization problem. It is not, however, a continuous version of Theorem 1: the theorem does not provide its computational content, and the probability remains generated by finite independent sampling rather than by the continuum. Replacing the random kernels by deterministic judgment types does not help; the same law-of-large-numbers collapse occurs. Introducing common shocks or finitely many correlated clusters preserves randomness only by adding a different model or returning to a finite population.
Corollary 2 is even less capable of supporting the mirror. With \(\Pi=\{\pi\}\), there is only one type and necessarily \(\mu=(1)\). There is no population variable to continuize at all; the result is purely an i.i.d. finite-sample fluctuation theorem. A multi-type extension is not a mirror of the corollary but a new variant of the paper’s semi-random model.
Large online juries or calibration classes may make that new variant sensible. Thus the universal claim that no high-multiplicity interpretation is imaginable would be too strong. But under the programme’s stricter requirement, the paper supplies no computational anchor, and its central likelihood phenomenon disappears in the true continuum. The honest negative verdict is therefore that no worthwhile continuous mirror of this paper has been established; the apparent mirror survives only by retaining the finite-\(N\) stochastic model and renaming its assignment multiplicities as population masses.
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.