| paper | d4-conf-ijcai-2003-0202 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 1 establishes when utility aggregation represents a preference relation; it does not assert that an object can be computed or that computing it is hard. The proposed population policy problem changes uncertainty about one user's preferences into aggregate welfare over user types, and the paper gives no computational result for that problem.
fails bit a — no named computational result to mirror
The proposed population extension draws on the utility aggregation behind Theorem 1; the paper's other representation, calibration, small-world, and elicitation discussions have no computational mirrors.
The strongest positive angle I can see is a population version of Theorem 1. Let each type \(t\) be a user’s utility function over a shared finite outcome set \(S\), normalized to the same best and worst outcomes. Let \(\mu_t\) be that type’s share of a large user population. Given a finite set of policies \(D\), each inducing an outcome lottery \(p_d\), choose a policy maximizing
\[
\sum_t \mu_t \sum_{s\in S} p_d(s)u_t(s).
\]
A plausible setting is a large service choosing one policy for many users, with many users but relatively few recurring preference types. The common normalization matters: Theorem 1, proved in this paper, says that when preferences share best and worst outcomes and their utility representations agree on fixed values for those outcomes, the distribution-weighted utility represents the aggregate preference relation. [Theorem 1 in the paper](https://www.cs.utoronto.ca/~cebly/Papers/foundations.pdf)
But this is only an adjacent model, not a valid computational anchor. In the paper, the distribution is uncertainty about one user’s utility function; it is not the fraction of a population holding each type. And Theorem 1 is a representation result, not a result about an algorithm or the complexity of a problem. The paper contains no named computational result of the kind required here, so I have no anchor on which to base a compliant continuous-problem question. The proposed population policy problem would need its own computational analysis, and the paper gives no complexity classification for it.
The strongest objection is that this paper has no computational result to continuize. Its contribution is foundational: it studies when expected expected utility is justified by decision-theoretic axioms, and what that means for preference elicitation. Theorem 1, as cited in the positive case, is a representation result, not a complexity theorem. The paper defines no computational problem whose complexity or algorithms could be classified for a continuous society.
The proposed large-service example is coherent: users could fall into recurring utility types, with shared best and worst outcomes, and a service could rank policy lotteries by mass-weighted utility. But that changes the paper’s setting from uncertainty about one user’s preferences to aggregating the welfare of many users. Common calibration gives the utilities a shared scale; it does not turn the paper’s axiomatic result into a computational one. To ask about algorithms, one would still have to specify a new policy space, its representation, and perhaps query costs or an elicitation objective. Those choices would define a new problem rather than mirror a result in this paper. A population extension of Theorem 1 would remain axiomatic, outside ChoCo’s stated scope.
So the proposed mirror does not survive as a computational mirror, even in its stronger population form. The limit of this negative case is that it cannot rule out a worthwhile high-multiplicity optimization problem inspired by the paper’s discussion of elicitation. It shows that such a problem would need an independent formulation and computational result; this paper supplies neither.
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.