| paper | d4-conf-ecai-2014-0336 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
Theorem 19
Given an attack graph \(G=(A,R)\), a finite support \(T\) of ballot types \(t\in\{0,1\}^{2|A|}\), rational masses \(\mu_t\) summing to \(1\), a positive rational smoothing mass \(\delta\), and tolerance \(\eta>0\), compute a rational vector \(\hat{x}\in[0,1]^A\) with maximum residual at most \(\eta\) for \(x_a=r_a\prod_{b\in\operatorname{Att}(a)}(1-x_b)\), where \(r_a=p_a/(p_a+q_a+\delta)\), \(p_a=\sum_t\mu_t b_t^+(a)\), and \(q_a=\sum_t\mu_t b_t^-(a)\).
A continuum electorate represented by a finite distribution over complete argument-voting ballots; each argument’s positive and negative vote shares determine its coefficient, and the decision variable is the approximate social-strength vector.
The fixed-\(\delta\) continuous semantics is not the literal large-population limit of the paper’s fixed-\(\epsilon\) count semantics: exact finite correspondence uses \(\delta=\epsilon/N\), which vanishes as \(N\) grows. The proponent gives the per-instance translation but does not settle which limiting smoothing convention the authors would adopt.
fatal: False
The mirror covers simple-product social-model approximation and the tree evaluation regime; it leaves the paper’s algorithm comparisons, benchmark findings, and broader semantic frameworks outside the stated question.
The strongest honest mirror is narrow: continuize the participants who vote on arguments, then ask for the same Simple Product social model the paper computes. The continuous object is the electorate, not the argument strengths, which are already real-valued in the paper.
Let \(G=(A,R)\) be an attack graph with \(n=|A|\) arguments. A voter type is a complete ballot \(t\in T\subseteq\{0,1\}^{2n}\): for each argument \(a\), the type specifies whether that voter casts a positive vote, a negative vote, both, or neither. The society is a rational distribution \(\mu\) over the \(\tau\) ballot types in its support, with \(\sum_t\mu_t=1\). This fits a large online consultation where millions of participants vote across a debate, while repeated ballot patterns—perhaps from communities or standardised campaigns—leave far fewer distinct types than people.
For each argument \(a\), let \(p_a=\sum_t\mu_t b^+_t(a)\) and \(q_a=\sum_t\mu_t b^-_t(a)\) be the positive and negative vote shares. With a positive per-capita smoothing mass \(\delta\), define \(r_a=p_a/(p_a+q_a+\delta)\), and ask for strengths \(x\) satisfying
\[
x_a=r_a\prod_{b\in\operatorname{Att}(a)}(1-x_b).
\]
Here the decision variable is the social-strength vector \(x\); there is no optimisation objective. The smoothing adjustment has a direct finite-population bridge: for \(N\) voters, set \(\delta=\epsilon/N\). If \(\mu_t\) is each ballot type’s fraction of the \(N\) voters, this gives exactly the paper’s vote aggregation formula with counts \(v^+_a=Np_a\) and \(v^-_a=Nq_a\).
My lead anchor is Theorem 19, proved in this paper. It states that ISS converges to a solution of the system \(x_i=\tau_i\prod_{j\in A_i}(1-x_j)\), independently of its initial guess \(x^{(0)}\in]0,1[^n\). For arguments with no positive votes, the paper’s Proposition 8 allows them to be removed and their zero strengths restored afterward. This anchors the following problem:
Mass-SAF Model Approximation. Given an attack graph \(G\), a finite list of rational ballot types and masses \(\mu\), a positive rational \(\delta\), and a tolerance \(\eta>0\), output a rational vector \(\hat x\in[0,1]^A\) whose maximum fixed-point residual is at most \(\eta\):
\[
\max_{a\in A}\left|\hat x_a-r_a\prod_{b\in\operatorname{Att}(a)}(1-\hat x_b)\right|\le\eta.
\]
Theorem 19 makes this a plausible Class A, approximate-numerical problem: for these rational inputs, ISS’s convergent iterates eventually reach any fixed residual tolerance. The paper’s convergence result is not a polynomial-time bit-complexity theorem, though, so I would describe tractability as the expected direction rather than a classification already established for this mirror.
A second, more restricted anchor is Corollary 25, also proved here: when the argumentation graph is a tree, ISSc is optimal; its proof gives total cost \(O(e)\), since the SCC procedure calls ISS on singleton components and there are no cycles. This anchors Tree Mass-SAF Model Evaluation: given the same rational mass-vote inputs but a tree attack graph, output the exact model. The equations can be evaluated in topological order, giving a unique rational solution with \(O(e)\) arithmetic operations. Its expected direction is Class A. The arithmetic-operation bound still leaves the bit cost of exact rational outputs to analyse.
Theorem 19 is the lead because it applies beyond tree debates; Corollary 25 adds a particularly clean tractable regime. Neither anchor is hardness. Further questions include proving or refuting uniqueness for this mass-input version, bounding the general graph’s running time in \(n,\tau,L,\log(1/\eta)\), and quantifying how much the model changes when a small mass shifts between ballot types.
The weakest point is substantial: the semantics uses only the \(2n\) vote-share marginals, not the joint distribution of ballot types. The mass distribution may therefore be mostly an input interpretation; the algorithm can reduce it to those marginals, so \(\tau\) need not matter after preprocessing. Still, the paper’s voters remain discrete counts, while its continuous variables are argument strengths. A large electorate voting on the debate is a plausible population regime, and the mass formulation asks the authors’ same model-computation question with normalized vote shares. It is a credible, modest mirror, but not a rich new complexity landscape on the evidence in this paper.
The strongest case against both anchors is that the paper has already compressed away the population. Its input is an attack graph and, for each argument, two vote counts. It does not represent voters or ballots, and its equations use only the resulting coefficients. A distribution over ballot types can be a coherent way to describe an electorate, but the solver sees it only through the \(2|A|\) positive and negative vote marginals. The correlations between types disappear. Moreover, Definition 9 already states the system for arbitrary real coefficients in \(]0,1[\).
That makes Theorem 19 a poor foundation for a new population-complexity question: its convergence result already covers the coefficients induced by rational vote masses. The ISS iteration depends on the attack graph and those coefficients, not on electorate size or ballot support. The paper’s running-time discussion likewise concerns graph structure and convergence rate, not the number of people or types. One could ask for bit-complexity bounds in \(n,\tau,L,\log(1/\eta)\), but that would be a new analysis of the same coefficient-input problem, not a complexity landscape created by making the population continuous. Adding mass-shift costs or a campaigning objective might create a richer problem, but neither is present in the paper’s model or anchored by Theorem 19.
There is a modeling choice around \(\varepsilon\), too. To recover the finite-count semantics from normalized masses, the per-capita smoothing term must be \(\delta=\varepsilon/N\). As \(N\) grows, that term vanishes for every argument with positive vote mass. Keeping \(\delta\) fixed gives a coherent alternative, but makes the regularizer scale with the population and changes the semantics. This does not rule out a mass version—the authors might accept that co-modification—but it shows that the proposed mirror needs an additional semantic choice.
Corollary 25 has the same limitation. For a tree attack graph, topological substitution works for arbitrary rational coefficients, whether they came from vote counts or voter masses. Exact rational evaluation is a sensible question, but its difficulty comes from rational arithmetic and graph size; the electorate and type support add no structure to the computation.
This is a case against the *payoff* of these mirrors, not against the coherence of a continuous electorate. A large consultation with many participants is plausible, and its normalized vote shares give a well-defined mass input. Since the paper’s computation already acts only on aggregate vote coefficients, however, the best case against is that this is a thin mirror with little population-specific computational content. I cannot honestly make the universal claim that no worthwhile mirror exists in any scenario; the proposed model is plausible, and the negative case is correspondingly limited.
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.