| paper | On Middle Grounds for Preference Statements |
| authors | Anne-Marie George, Ana Ozaki |
| venue | IJCAI 2025 |
| filed under | voting · combinatorial |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Corollary 16
statement extracted from the paper’s text layer
Given finite variables and domains, a lexicographic preference language, complete stakeholder types \(\Gamma_1,\ldots,\Gamma_\tau\subseteq L\), rational masses \(\mu_1,\ldots,\mu_\tau\), and \(\varepsilon,\rho\in[0,1]\cap\mathbb{Q}\), decide whether the population has a middle-ground witness: if \(U_\mu=\bigcup_{\theta:\mu_\theta>0}\Gamma_\theta\) is consistent, output \(U_\mu\); otherwise output some \(b\in\mathcal{B}=\{\mathbf{1}_v\ge\mathbf{0}_v,\mathbf{1}_v>\mathbf{0}:v\in V\}\) satisfying \(\sum_{\theta:\{\psi,b\}\text{ is consistent for every }\psi\in\Gamma_\theta}\mu_\theta\ge1-\varepsilon\) and \(\sum_{\theta:\Gamma_\theta\models_{\mathcal L}b}\mu_\theta\ge\rho\), or report NO.
Types are complete preference bases \(\Gamma_\theta\)\), masses are \(\mu_\theta\)\), and the decision variable is a union base or candidate \(b\in\mathcal{B}\)\); the objective is to maximize compatible population mass subject to justified population mass.
The exact formulation depends only on positive-mass support, whereas making mass consequential requires replacing the paper's universal \(P3\) and existential \(P4\) conditions by thresholds.
fatal: False
The mirror covers hierarchical consistency, entailment support, and lexicographic middle-ground witnessing; it leaves the structural non-existence and non-uniqueness results and full maximal-middle-ground construction largely untouched.
There is a defensible positive case, but the continuous object must be the population of stakeholders, not the alternatives, variables, or preference statements themselves. I would use the following common model.
Fix the paper’s finite variables, domains, operator \( \oplus \), and hierarchical-model class. A stakeholder type \( \theta \) is a complete finite statement base \( \Gamma_\theta\subseteq L \). The society is a rational distribution \( \mu\) over finitely many such types. Thus \( \mu_\theta\) is the fraction of stakeholders whose complete expressed opinion is \( \Gamma_\theta\). Clearing denominators produces a finite society of identical clones, and every rational finite-type society has this interpretation.
This is plausible for a large consultation, moral-preference survey, or policy deliberation in which millions of respondents answer the same finite menu of comparisons and fall into a moderate number of recurring opinion profiles. It is much less plausible for a seven-person hiring committee; the scenario matters.
My lead mirror is:
\[ \textsf{Hierarchical-Consensus}_{\infty}. \]
An instance consists of the finite preference language \(L\), the hierarchical-model specification, stakeholder types \(\Gamma_1,\ldots,\Gamma_\tau\), rational masses \(\mu_1,\ldots,\mu_\tau\), and a threshold \(\eta\in[0,1]\). For a hierarchical model \(\pi\), define
\[ \operatorname{sat}_\mu(\pi) = \sum_{\theta:\,\pi\models\Gamma_\theta}\mu_\theta . \]
The problem is to find a hierarchical model \(\pi\) maximizing \(\operatorname{sat}_\mu(\pi)\), or, in decision form, to decide whether some \(\pi\) satisfies \(\operatorname{sat}_\mu(\pi)\ge\eta\). The special case \(\eta=1\) asks whether the entire positive-mass population has a common hierarchical model.
This is a faithful population version of the paper’s consistency problem. It does not fractionalize a hierarchy or weaken a preference statement. It asks for one of the paper’s own models that satisfies the largest fraction of the population’s complete preference bases. At \(\eta=1\), the rational-clone correspondence is exact.
The anchor is Theorem 11, proved in this paper. It states that deciding consistency of preference statements under hierarchical models is NP-complete, via a reduction from Subset Sum. The reduction already uses only three preference statements, so it embeds directly by taking one stakeholder type of mass \(1\), or by cloning that type arbitrarily many times. Consequently, \(\textsf{Hierarchical-Consensus}_{\infty}\) is expected to be Class B: hardness transfers because the combinatorics live in the variable subsets and the hierarchy, not in population multiplicity.
That is a worthwhile mirror even though continuization does not make this particular problem easy. It identifies a clean boundary: aggregating millions of repeated stakeholders does not remove hardness when the hard object is the agenda of variables.
A second, independent mirror is:
\[ \textsf{Mass-Justification}_{\infty}. \]
Given the same society and a query statement \(\varphi\), define the mass of stakeholders whose complete statements entail \(\varphi\) by
\[ J_\mu(\varphi) = \sum_{\theta:\,\Gamma_\theta\models_{\mathcal H}\varphi}\mu_\theta . \]
The problem is to compute \(J_\mu(\varphi)\), or decide whether \(J_\mu(\varphi)\ge\rho\) for a rational support threshold \(\rho\). This asks whether a proposed compromise statement is logically justified by at least a prescribed fraction of the population.
The anchor is Corollary 12, proved in the paper as a consequence of Theorem 11, using the entailment–consistency equivalence cited to Wilson et al. (2017). With one type of mass \(1\) and \(\rho=1\), the problem is exactly the paper’s coNP-complete entailment problem. Thus the mirror is again at least Class B: it is coNP-hard, while with an explicitly listed finite type support it is naturally in \(P^{\mathrm{coNP}}\) by querying entailment separately for each type. The mass threshold adds a genuine population question without changing the underlying logical language.
The strongest tractability story comes from the paper’s lexicographic result:
\[ \textsf{Lexicographic-Mass-Middle-Ground}_{\infty}. \]
For each variable \(v\), let
\[ \mathcal B = \{\mathbf 1_v\ge \mathbf 0_v : v\in V\} \cup \{\mathbf 1_v>\mathbf 0 : v\in V\}, \]
the \(2|V|\) candidate statements identified in Theorem 15. For \(b\in\mathcal B\), define
\[ \operatorname{comp}_\mu(b) = \sum_{\theta:\,\{\psi,b\}\text{ is consistent for every }\psi\in\Gamma_\theta} \mu_\theta \]
and
\[ \operatorname{just}_\mu(b) = \sum_{\theta:\,\Gamma_\theta\models_{\mathcal L} b}\mu_\theta . \]
The continuous problem is: given \(\varepsilon,\rho\in\mathbb Q\), decide whether there is a candidate \(b\in\mathcal B\) with
\[ \operatorname{comp}_\mu(b)\ge 1-\varepsilon \quad\text{and}\quad \operatorname{just}_\mu(b)\ge\rho. \]
The output is such a witness \(b\), or NO. In the exact-support special case \(\varepsilon=0\) and \(\rho=\min\{\mu_\theta:\mu_\theta>0\}\), this is precisely the paper’s middle-ground existence test after replacing repeated stakeholders by types. If the union of all positive-mass statement bases is consistent, return that union; otherwise Theorem 15 reduces the search to the \(2|V|\) candidates above.
The anchor is Corollary 16, proved here using Theorem 15 and the polynomial-time lexicographic consistency and inference results of Wilson et al. (2015). The witness problem is expected to be Class A: there are only \(2|V|\) candidates, each compatibility and entailment test is polynomial, and the mass comparisons are rational arithmetic.
This is the most attractive algorithmic mirror, but it is also the least direct one. The exact paper postulates treat stakeholder multiplicity only through support: cloning a stakeholder does not change whether a middle ground exists. The \(\varepsilon,\rho\) formulation makes mass consequential by turning P3 and P4 into population-support requirements, but that is an explicit extension of the paper’s definition. Full construction of a maximal middle ground remains potentially exponential, just as the paper leaves construction open even for lexicographic models.
I would therefore scope the positive claim narrowly. The paper supports three continuous questions: population consensus under hierarchical models, population-weighted logical justification, and lexicographic middle-ground witnessing. The first two are faithful Class-B continuizations; the third gives a credible Class-A witness problem. I would not claim that Theorems 6 and 7 alone establish computational mirrors, since their non-uniqueness and non-existence statements are structural rather than complexity results.
The weakest point is that the paper’s original middle-ground semantics are set-based, not weight-based. A critic can reasonably say that \(\textsf{Lexicographic-Mass-Middle-Ground}_{\infty}\) changes P3 and P4 rather than merely replacing individuals by masses. That objection does not defeat the lead mirror: \(\textsf{Hierarchical-Consensus}_{\infty}\) preserves the original satisfaction relation, passes the rational-clone test exactly, and gives the authors’ own stakeholder language a genuine large-population interpretation.
The proponent has found a population wrapper around the paper, but not a worthwhile continuous mirror. All three anchors run into the same structural problem: the paper treats stakeholders as logical sources, not as a population whose multiplicity affects the predicate. Preserving the paper’s semantics makes masses disappear; making masses matter changes the problem.
Theorem 11 is the clearest example. Its input is one set of statements \(\Gamma\), and the question is whether there exists a hierarchical model satisfying their conjunction. The proposed one-type instance with mass \(1\) is therefore not evidence of continuization: any problem can be embedded in a “continuous” model by assigning all mass to one type.
Nor does adding several types repair this. For the exact-consensus version with \(\eta=1\),
\[ \operatorname{sat}_\mu(\pi)=1 \]
holds exactly when
\[ \pi\models\bigcup_{\theta:\mu_\theta>0}\Gamma_\theta. \]
The masses and multiplicities vanish; only the union of the supports remains. This is simply Theorem 11 applied to a larger set of statements.
The weighted objective
\[ \max_\pi \sum_\theta \mu_\theta \mathbf 1[\pi\models\Gamma_\theta] \]
is more interesting, but it is a new weighted model-fitting or Max-consistency problem. The paper supplies neither this objective nor a reason that a stakeholder’s complete base should be the unit of weight. Majority satisfaction, weighted violated statements, and lexicographic aggregation are all plausible new projects, but they are not continuizations of Theorem 11. The plausible survey scenario does not remove that semantic gap.
Corollary 12 has the same defect. The paper’s entailment question is
\[ \Gamma\models\varphi, \]
a universal property of one statement base and one model class. The proposed quantity \(J_\mu(\varphi)\) is merely a weighted histogram of independently evaluated bases. With an explicitly listed type support, it consists of separate entailment tests followed by rational summation. There is no common model, compromise, or interaction between stakeholders.
If the original P4 semantics are retained, the relevant question is whether some stakeholder entails \(\varphi\). That depends only on whether the support contains such a type, not on its mass. Requiring \(J_\mu(\varphi)\ge\rho\) makes mass operative only by replacing the paper’s existential justification condition with an arbitrary support threshold. An implicit or sampled distribution might produce a harder problem, but then the difficulty comes from the distribution’s representation or access model, not from continuizing the paper’s logical population.
Corollary 16 is the strongest-looking anchor, but it exposes the problem most directly. In the paper, P3 is universal over stakeholders and P4 is existential: a candidate statement must be compatible with every stakeholder statement and entailed by at least one stakeholder base. After replacing duplicates by types, exact middle-ground existence depends only on the positive-mass support. A population with one respondent or a million clones has exactly the same answer. The continuum has degenerated.
The proposed \(\varepsilon,\rho\) version makes masses matter by changing both postulates. For \(\varepsilon>0\), the returned statement may be incompatible with some stakeholders, contrary to P3. For a positive justification threshold, it need not be entailed by any individual stakeholder in the sense required by P4. It is therefore an approximate weighted-consensus problem, not the paper’s middle-ground problem. Restricting the search to the \(2|V|\) statements from Theorem 15 may make this extension easy, but the theorem’s reduction applies to the exact logical postulates; it does not canonically define this approximate population notion.
A better repair still faces the same dichotomy. Preserve the paper’s P3 and P4, and multiplicity is irrelevant. Replace either condition by a mass threshold, and one has designed a new aggregation rule whose normative meaning and computational interest come from outside the paper. A large moral-preference survey could motivate such a project, but it would be a new paper on weighted preference consensus, not a continuous mirror of these results.
The honest weakness of this negative case is that the survey interpretation is not absurd. A separate author could reasonably study population-weighted hierarchical consensus. But none of the three named results survives as a population-sensitive computational object without changing its semantics. I would therefore reject the proposed mirror programme for this paper: it offers weighted wrappers around single-base reasoning, while the paper itself contains no computational result whose subject is an electorate or whose outcome depends on electorate mass.
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.