| paper | d4-conf-aaai-2010-0416 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
other→P discrete result → its continuous mirror
Answer: \(D(m)=1\) for every positive integer \(m\), and \(V_\mu=\operatorname{argmax}_{x\in A}W_\mu(x)\) for every society. · discrete side: other · adjudication: proved
full proof & adversarial review →
Theorem 4
Given alternatives \(A\), a finite set \(T\) of rational normalized utility types \(u_t\), and a society distribution \(\mu\) over \(T\), apply the mass analogue of Theorem 4's embedding, \(p_t(x)=u_t(x)\). Characterize its Plurality winners and distortion against \(W_\mu(x)=\sum_{t\in T}\mu_t u_t(x)\) over all such societies.
Types are complete normalized utility vectors; \(\mu_t\) is the mass of type \(t\), and \(p_t(x)\) is the fraction of that mass voting for \(x\). Plurality scores are \(S_p(x;\mu)=\sum_t\mu_t p_t(x)\), compared with welfare \(W_\mu(x)=\sum_t\mu_t u_t(x)\). For \(p_t(x)=u_t(x)\), the scores equal welfare for every \(x\) and \(\mu\), giving distortion \(1\).
The mirror covers Theorem 4's randomized Plurality embedding, whose finite-population distortion bounds become an exact mass result; it leaves the deterministic Plurality results and the Approval and Veto results untouched.
The strongest mirror is the paper’s randomized Plurality embedding, anchored by Theorem 4, proved in this paper (parts 2 and 3 are in the full version). It says that for \(n\ge m\), the embedding that makes an agent with utility vector \(u\) vote for \(x\) with probability \(u(x)\) has distortion \(O(m^{2m/n})\) and \(O(\sqrt m)\). It also says that, for \(n\ge3\), if
\[
\varepsilon(n,m)=4\sqrt{\frac{m\ln n}{n}}<1,
\]
then its distortion is at most \(1/(1-\varepsilon(n,m))\).
Here is the continuous problem I would hand to a prover: Robust Mass-Embedding Design for Plurality. An instance gives a finite alternative set \(A\) and a finite set \(T\) of rational normalized utility vectors \(u_t\in\Delta(A)\). A society is any rational distribution \(\mu\) over \(T\). Choose a single type-local embedding \(p:T\to\Delta(A)\), used for every such society, to minimize its worst-case distortion. In the continuous population, \(p_t(x)\) means that fraction of type \(t\)’s mass casts its one vote for \(x\). Thus
\[
S_p(x;\mu)=\sum_{t\in T}\mu_t p_t(x),\qquad
W_\mu(x)=\sum_{t\in T}\mu_t u_t(x).
\]
The winners maximize \(S_p\), and distortion is the maximum \(W_\mu\) divided by the least \(W_\mu\) among those winners.
The solution is \(p_t(x)=u_t(x)\). Then \(S_p(x;\mu)=W_\mu(x)\) for every alternative and every society: every Plurality winner is welfare-maximizing, so the worst-case distortion is exactly \(1\), the best possible. The rule is specified in \(O(\tau m)\) entries and works without knowing \(\mu\).
The regime is a large cooperative system choosing among shared plans, with many agents but only a small catalog of utility vectors—for example, repeated classes of autonomous units evaluating the same mission plans. Here a type is the complete normalized utility vector over those plans; mass is the fraction of the population with that vector. The intended regime has \(n\gg\tau\), so each type can be represented by a large cohort. Every unit still casts one ordinary vote; only the cohort’s aggregate vote shares are fractional.
I expect this mirror to be tractable (Class A). The paper’s finite-\(n\) random votes fluctuate around their expected scores; in the continuous population those expected shares are the actual mass scores. That is a direct fluid limit of the authors’ embedding, and I think they would recognize it as their problem’s continuous-population form. It covers Theorem 4’s randomized Plurality result, not the paper’s deterministic bounds or its Approval and Veto results.
The weakest point is also the reason the mirror is so clean: continuity removes the sampling variation that makes Theorem 4’s bounds interesting. The continuous answer is an exact benchmark, not a new finite-population bound. And, strictly speaking, the paper has no named theorem classifying a problem as in P, NP-hard, or similar; Theorem 4 is an algorithmic performance result. A useful next question is how the distortion changes when mass has a maximum indivisible size, so the model moves back toward finite, randomly voting agents.
The strongest objection is a scope mismatch: Theorem 4 gives distortion bounds for a randomized voting rule; it does not classify the complexity of a computational problem. The paper supplies no search or decision problem whose high-multiplicity complexity those bounds answer. The proposed “Robust Mass-Embedding Design” adds an optimization problem over embeddings, but that is a new problem, not a continuous version of a computational problem named in the paper.
The more generous mirror still has a narrow computational target. With a nonatomic population, independently randomized votes aggregate to their probability shares. For \(p_t(x)=u_t(x)\), the mass score is \(\sum_t\mu_t u_t(x)\), so the continuous model has no empirical vote-count fluctuations to analyze. One could preserve those fluctuations by retaining finite cohort sizes, but then the quantity governing the result is the finite granularity \(n\), which is the paper’s setting. The mass model itself offers no search over permutations, hard optimization, or other computational structure for ChoCo to chart.
That is a reason to question this anchor’s value to a complexity programme, not to reject it as an illegitimate continuous population model. The mirror is coherent, and its fluid-limit interpretation is recognizable. Since the programme accepts tractable mirrors and does not reject a question merely because its answer is simple, I cannot honestly claim this defeats the anchor outright. The negative case here is weak: it shows that Theorem 4 may be a poor fit for ChoCo’s computational agenda, but not that no worthwhile continuous mirror exists.
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.