| paper | d4-conf-atal-2022-0195 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Theorem 5.1
statement extracted from the paper’s text layer
Given finite value labels \(V\), actions \(A\), and supported types \(T\), each with rational judgement matrices \(J_t^+,J_t^- \in [-1,1]^{|V|\times|A|}\), positive rational per-person influence \(q_t\), and positive rational mass \(\mu_t\) summing to \(1\), find any pair \(X^+,X^- \in [-1,1]^{|V|\times|A|}\) minimizing the mass-weighted disagreement for fixed \(p\in[1,\infty]\). For finite \(p\), minimize \(\left[\sum_{t\in T}\mu_t q_t\sum_{v\in V,\,a\in A}\left(|J_t^+[v,a]-X^+[v,a]|^p+|J_t^-[v,a]-X^-[v,a]|^p\right)\right]^{1/p}\); for \(p=\infty\), minimize the maximum absolute coordinate disagreement over all positive-mass types, values, actions, and signs.
A society is a finite-support distribution over complete action-judgement types; each type carries its rational judgement matrices and per-person influence, and the decision variable is the consensus pair minimizing weighted ℓ_p disagreement, with influence omitted at \(p=\infty\).
The opponent's strongest point is that the weighted ℓ_p objective is already exactly representable by rows for the types, and its coordinates are separable, so the mirror adds no new optimization structure; this bounds its novelty, though the proponent acknowledges the re-expression and the model remains a finite list.
fatal: False
The mirror covers the judgement aggregation in Theorem 5.1; the same construction applies separately to the preference aggregation in Theorem 5.2. It leaves the European Values Study analysis and its comparisons of ethical parameters as empirical results.
The strongest mirror is the paper’s aggregation of citizens’ judgements about how values evaluate actions. Consider a large public consultation for an AI policy: there may be millions of participants, but each answers the same finite questionnaire using fixed response categories. A type is a complete response pattern—its judgement matrices, preferences over values, and per-person social influence. If \(n_t\) people share type \(t\) in a population of size \(N\), its mass is \(\mu_t=n_t/N\). In this regime the support size \(\tau\) can be far smaller than \(N\). That is a plausible high-multiplicity version of the paper’s stated goal of aggregating the value systems of a group of individuals or stakeholders.
My lead anchor is Theorem 5.1, proved in this paper. It says that finding the aggregate judgement matrices \(J_S\) that minimize the paper’s weighted \(p\)-distance \(U_p^{(J)}\) is equivalent to solving an \(\ell_p\)-regression problem \(\min_x\|Bx-b\|_p\). Here \(B\) has \(2|Ag||V||A|\) rows and \(2|V||A|\) columns, formed by stacking one block \(w_i^{1/p}I\) per agent; \(b\) stacks the corresponding vectors \(w_i^{1/p}\overrightarrow{J_i^+}\) and \(w_i^{1/p}\overrightarrow{J_i^-}\). The solution \(x\) is the vectorization of the positive and negative aggregate judgement matrices. This is a computational formulation of the underlying aggregation problem, and the paper uses it to solve instances.
The continuous question I would hand to a prover is Continuous Action-Judgement Consensus. An instance consists of finite value labels \(V\), actions \(A\), and a finite support \(T\). Each type \(t\in T\) has rational judgement matrices \(J_t^+,J_t^-\in[-1,1]^{|V|\times|A|}\), a positive rational per-person influence \(q_t\), and a positive rational mass \(\mu_t\), with \(\sum_t\mu_t=1\). Given a fixed \(p\in[1,\infty]\), find any pair \(X^+,X^-\) minimizing the mass-weighted disagreement. For finite \(p\), the objective is \(\sum_{t\in T}\mu_tq_t\sum_{v\in V,a\in A}\bigl(|J_t^+[v,a]-X^+[v,a]|^p+|J_t^-[v,a]-X^-[v,a]|^p\bigr)\). For \(p=\infty\), it is the maximum absolute disagreement over every positive-mass type, value, action, and sign; as in the paper, the influence weights disappear in that limit. A solution is any minimizing pair of matrices, with entries in \([-1,1]\).
This is the same judgement-aggregation question, with mass replacing the list of named participants. Repeating type \(t\) exactly \(n_t\) times with weight \(q_t\) gives the same minimizer as mass \(\mu_t=n_t/N\), since normalization only scales the finite-population objective. The type retains every feature the problem uses; the compression is precisely that identical participants can be represented by their total mass. In the continuous regression formulation, the paper’s block for type \(t\) is weighted by \((\mu_tq_t)^{1/p}\), so the variable dimension depends on \(|V||A|\), while the data depend on \(\tau\), rather than on \(N\).
A second, parallel anchor is Theorem 5.2, also proved here. It says that minimizing the paper’s weighted \(p\)-distance \(U_p^{(P)}\) to find the aggregate preference matrix \(P_S\) is equivalent to \(\min_x\|Cx-c\|_p\). The matrix \(C\) stacks blocks \(w_i^{1/p}I\), with \(I\) of size \(|V|^2\); \(c\) stacks \(w_i^{1/p}\overrightarrow{P_i}\); and the solution vector is \(\overrightarrow{P_S}\). Its continuous counterpart, Continuous Value-Preference Consensus, takes the same typed population and asks for \(Y\in[0,1]^{|V|\times|V|}\) minimizing, for finite \(p\), \(\sum_t\mu_tq_t\sum_{v,v'\in V}|P_t[v,v']-Y[v,v']|^p\), or the maximum coordinate disagreement over positive-mass types for \(p=\infty\). Any minimizing \(Y\) is a solution. Theorem 5.1 is my stronger anchor because it aggregates the action judgements that give the value labels their concrete meaning; Theorem 5.2 completes the paper’s separate aggregation of preferences.
I expect both mirrors to be Class A. For \(p=1\) and \(p=\infty\), the objectives have linear-program formulations; for \(p=2\), the paper gives the weighted-mean solution for judgement matrices in Proposition 5.3. For a general fixed rational \(p\), the finite-support objectives are convex, so I would expect an additive-\(\varepsilon\) algorithm with running time polynomial in \(\tau\), \(|V||A|\), input bit length, and \(\log(1/\varepsilon)\). That bit-complexity bound is a prediction for the mirror, not a theorem established by this paper: a precise algorithmic version should specify the approximation guarantee and arithmetic model. Further questions include whether the two consensuses remain unique for different \(p\), how sensitive they are to small changes in type masses, and how to choose separate ethical parameters for judgements and preferences.
The paper contains no named NP-hardness result to anchor a hardness-transfer question. Its named computational results are the regression formulations above, which the paper uses as the basis for computation. That makes this a positive case for a tractable continuous mirror, rather than a claim that continuization dissolves a hardness result.
The weakest point is that this is a close population re-expression of an optimization problem the paper already formulates over individual agents. The continuous version does not obviously create a new optimization structure: it groups identical value systems and replaces repeated observations with masses. But that is a faithful high-multiplicity regime, not a simplified substitute for the authors’ problem. Their model already aggregates the value systems of society members, and the paper motivates the task with large stakeholder groups. I would expect the authors to recognize this as their aggregation problem asked of a society described by type masses, with the practical question centered on computation in \(\tau\) when \(N\) is enormous.
The strongest case against is that Theorems 5.1 and 5.2 already accept the continuous mirror’s data in substance. I would not argue that there is no named computational result: both theorems state computational optimization formulations. The more serious point is that the paper’s weighted list of agents is already a finite distribution over value-system types.
For a society with masses \(\mu_t\) and per-person influence \(q_t\), give the paper one row for each supported type and weight that row by \(w_t=\mu_tq_t\). Its disagreement objective is then exactly the proposed mass-weighted objective, up to an irrelevant normalization. At \(p=\infty\), both models take the maximum over the positive-weight types, and the influence weights disappear. Conversely, any finite rational weighted instance in the paper can be read as a distribution over its rows. The population description changes; the optimization problem does not.
That answers both anchors. Theorem 5.1’s judgement matrices and Theorem 5.2’s preference matrices are separate blocks of the same weighted \(\ell_p\) fitting problem. Grouping repeated agents into types simply replaces repeated rows with one weighted row. The paper’s regression dimensions already become functions of the support size when the input has one row per type. The proponent’s consultation scenario is plausible, but it illustrates this exact equivalence rather than supplying a distinct problem for the programme to classify.
The equivalence is unusually direct because the model has no coupling between coordinates: each judgement or preference entry is fitted independently. For \(p=1\), this is weighted median fitting; for \(p=2\), weighted mean fitting; for \(p=\infty\), midpoint-of-extremes fitting. General fixed \(p\) remains coordinate-wise scalar regression. There is no additional social-choice structure here whose computational behavior changes when the rows are interpreted as a population distribution.
A richer version could add constraints linking judgments across actions or values, or mechanisms that treat people differently. Those additions might create new computational questions, but they would add structure beyond the aggregation problems anchored in this paper. If identities matter to such a mechanism, grouping people into types would also need to preserve whatever identity-dependent information it uses.
This is a narrow negative case: I cannot honestly claim that a large consultation with finitely many response types is an implausible high-multiplicity regime, and the authors would likely recognize that application. The objection is that the paper’s own weighted formulation already represents precisely those typed populations, so the proposed mirrors appear to restate its optimization problems in population language. If that equivalence does not count as a collision for ChoCo, these anchors make a strong case for a 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.