Epistemic Abstract Argumentation Framework: Formal Foundations, Computation and Complexity

There is a defensible, though deliberately modest, continuous mirror here. Its strongest anchor is Theorem 4, and it should be presented as a high-multiplicity extension of EAAF rather than as a claim that the paper already studies weighted populations.

The regime is a large population of stakeholders participating in a common structured deliberation. The public EAAF \(\Delta=\langle A,\Omega,\Psi,\Phi\rangle\) is the finite catalogue of recurring claim types and their attack relations. A stakeholder type is a complete pair \((\Delta,a)\), where \(a\in A\) is the claim that stakeholder supports; all attack-relevant information is included in the type. If \(n_a\) stakeholders support claim \(a\), their mass is \(\mu_a=n_a/N\). The intended regime has \(N\gg |A|\): for example, thousands of respondents repeatedly make a few dozen recurring claims in a policy dispute.

This preserves the paper’s graph rather than replacing it by attack densities. Multiple stakeholders of type \(a\) do not become multiple interacting copies of the argument node; they are indistinguishable supporters of the same abstract argument. Thus rational masses are an exact high-multiplicity encoding: clearing denominators produces \(N\) cloned stakeholders, and every extension \(S\subseteq A\) represents exactly \(\sum_{a\in S}n_a\) accepted stakeholders. No argument is fractionally accepted.

My lead problem is EAAF-Epistemic-AcceptanceMass\(_\infty\), anchored in Theorem 4, proved in this paper. Restrict the population support to epistemic arguments, or use the paper’s Fact 1 to convert ordinary credulous or skeptical queries into epistemic ones. An instance consists of a well-formed EAAF \(\Delta\), a semantics \(\sigma\in\{\mathrm{gr},\mathrm{co},\mathrm{pr}\}\), a rational mass vector \(\mu\in\mathbb{Q}_{\ge 0}^{A}\) with \(\sum_{a\in A}\mu_a=1\), and a rational threshold \(\theta\). Define \(e_\sigma(\Delta,a)=1\) if \(a\) is epistemically accepted under \(\sigma\), according to the paper’s \(EA_\sigma\) predicate, and \(0\) otherwise. The problem asks whether the accepted population mass \(M_\sigma(\Delta,\mu)=\sum_{a\in A}\mu_a e_\sigma(\Delta,a)\) is at least \(\theta\). An optimization version asks for the exact rational value of \(M_\sigma(\Delta,\mu)\).

This is the continuous form of a natural question for the authors’ setting: what fraction of a population’s recurring claims are supported by the epistemic argumentation semantics? Theorem 1 makes the world view unique, and Proposition 2 makes epistemic arguments deterministic, so the aggregate statistic is well-defined even when ordinary arguments have multiple statuses.

I expect this problem to be in Class A for grounded semantics and Class B for complete and preferred semantics. For grounded semantics, Theorem 4 gives \(EA_{\mathrm{gr}}\in\mathrm{P}\); evaluating all argument types and summing their rational masses remains polynomial. For complete semantics, the one-type restriction \(\mu_g=1\), \(\theta=1\) recovers the paper’s \(EA_{\mathrm{co}}\) problem, so the \(\Theta_2^p\)-hardness from Theorem 4 transfers. The explicit weighted aggregate should remain in \(\Delta_2^p\), since one can make polynomially many calls to the individual acceptance procedure. Likewise, preferred semantics inherits the \(\Pi_2^p\)-hardness of Theorem 4 and should lie within \(\Delta_3^p\) under the paper’s explicit representation.

The hardness is not caused by the number of stakeholders; it lives in the EAAF’s attack topology and extension semantics. That is precisely why this is a useful Class B mirror: continuization does not falsely promise to dissolve every difficulty.

A second, independently defensible mirror is EAAF-Mass-Supported-Extension\(_\infty\), anchored in Theorem 5, also proved in this paper. An instance consists of \(\Delta\), \(\sigma\in\{\mathrm{gr},\mathrm{co},\mathrm{pr}\}\), a rational population vector \(\mu\in\mathbb{Q}_{\ge0}^{A}\), and a rational threshold \(q\). The question is whether there exists \(S\in\sigma(\Delta)\) such that \(S\neq\varnothing\) and \(\sum_{a\in S}\mu_a\ge q\). A solution is such an extension \(S\), together with its represented mass; the optimization form maximizes that mass over all \(\sigma\)-extensions.

This is recognizable as the paper’s non-empty existence problem with a population-sensitive objective: a complete or preferred extension must not merely contain an argument, but must represent a sufficiently large mass of stakeholders. Theorem 5 transfers exactly. Set \(\mu_a=1/|A|\) for every \(a\) and \(q=1/|A|\). Then every nonempty extension satisfies the threshold, while the empty extension does not. Hence the instance is yes exactly when \(Ex^{\neg\varnothing}_\sigma\) is yes. Therefore the problem is NP-hard, indeed NP-complete on the corresponding unweighted slice, for \(\sigma\in\{\mathrm{co},\mathrm{pr}\}\). For grounded semantics, Theorem 5 gives a polynomial-time baseline because the grounded extension is unique and its mass can be evaluated directly. The unrestricted weighted version may have a higher exact complexity, but its fundamental hardness is inherited from the finite EAAF rather than created by the continuum.

The scope is intentionally narrow. These mirrors cover Theorem 4’s epistemic acceptance results and Theorem 5’s non-empty existence results. Algorithm 1 and Theorem 2 provide a natural computational implementation route for evaluating each type, but I would not claim a new polynomial-time theorem for Algorithm 1 because it relies on an external incremental AAF solver. I would also leave stable semantics outside scope, as the paper itself postpones it.

The main further questions are whether the weighted preferred-extension problem has a tight complexity classification; whether exact or approximate maximum represented mass is fixed-parameter tractable in the number of epistemic attack types; and whether rational mass solutions admit useful finite-clone rounding guarantees. A more ambitious extension could let population mass influence which arguments are selected, but that would need to preserve the paper’s epistemic semantics carefully.

The weakest point is that mass does not feed back into attack or extension feasibility. An opponent can fairly say that this is weighted argumentation or batch evaluation, not a new mean-field EAAF. I would concede that objection. The positive claim is more modest: the paper’s abstract arguments can serve as complete population types, rational masses have an exact clone interpretation, and its named computational problems survive as meaningful population-level acceptance and representation questions. This is a genuine Class B continuous mirror, even though it does not produce a new Class A tractability gain.