| paper | d4-conf-wine-2015-0019 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | unclear |
Theorem 17 clearly has computational character, and the proposed role-mass problem is precise. But its exact correspondence depends on whole cohorts flipping together, while the paper updates one agent at a time; the more faithful split-mass version needs a new reachability definition. Whether the cohort restriction is a model the authors would recognize is undecided, so neither green nor red is warranted.
fails bit none — no continuous question survives
The exact correspondence requires each role cohort to flip as one actor, whereas the paper permits individual updates; the theorem does not establish that reachability is preserved when a cohort can split into different opinions.
fatal: False
The unresolved point is whether unanimous role-level flips are a co-modification the authors would accept, given their individual-update dynamics. Their acceptance of block-level campaigns would validate this mirror; requiring agents within each role to split would require a new reachability definition and would not inherit Theorem 17.
The mirror covers Theorem 17's small-minority reachability hardness for role-structured networks; it leaves the graph characterization in Theorems 1 and 2 and the algorithmic results, including Theorem 20, untouched.
The strongest mirror I see is the paper’s small-minority hardness result, Theorem 17. It is proved in this paper; the NP-hardness of its source problem, 2P2N-3SAT, is cited from reference [17]. Theorem 17 says that for every fixed \(0<\varepsilon<1/8\), it is NP-hard, given an \(n\)-node graph, to decide whether some initial profile with at most \(n(1/4-\varepsilon)\) agents preferring \(1\) can reach a stable profile with at least \(n/2\) agents preferring \(1\). Equivalently, an initial \(0\)-majority can be subverted by a very small \(1\)-minority. The update sequence and its stable endpoint are part of the question.
A concrete continuous version is Cohort Minority-to-Majority Reachability. An instance has a finite set \(T\) of network-role types, an undirected graph \(H=(T,F)\), and positive rational masses \(\mu_t\) summing to \(1\). Interpret each type as a continuum cohort: every member of type \(t\) is connected to all members of each neighboring type and to no members of other types. Cohorts share one opinion at each step. An initial profile is \(x^0:T\to\{0,1\}\). A legal step chooses a type \(t\) for which more than half of its neighboring mass currently has the opposite opinion, then flips that cohort’s opinion. A profile is stable when no type can legally flip.
For fixed \(0<\varepsilon<1/8\), ask whether there is an initial profile with \(\sum_{t:x^0_t=1}\mu_t\le 1/4-\varepsilon\) and a finite legal update sequence ending at a stable profile \(x\) with \(\sum_{t:x_t=1}\mu_t\ge 1/2\). A solution is that initial type assignment together with the update sequence. The decision problem is measured in \(\tau=|T|\) and the bit length \(L\) of the masses. With \(\mu_t=1/\tau\), this is exactly the paper’s graph problem on its type graph, including its thresholds. Thus Theorem 17 gives NP-hardness for this continuous problem already on the equal-mass subfamily.
A plausible regime is a large platform or organization with many users in each of a smaller number of communication roles. Users in a role share the same initial signal and see the same neighboring roles; a role’s opinion changes as a cohort. For instance, the population might have millions of users and hundreds or thousands of roles. The graph \(H\) captures the interaction structure between roles, while mass records population shares. This is a block-structured version of the paper’s network, not a claim that all social networks have such a structure.
I would expect hardness to transfer: the reduction’s combinatorics live in the interaction graph and its gadgets, not in the number of individually represented agents. In the equal-mass case the continuum is a high-multiplicity lift of the finite role graph, so changing counts into masses does not remove that encoding. The further question is whether hardness persists when role masses are unequal or when only part of a cohort may update.
The soft spot is the cohort-wide update rule. The paper selects one unhappy agent at a time; its theorem does not establish hardness when each network role contains a continuum of independently updating agents who can split into different opinions. A referee might regard that split-capable model as the more faithful mirror. I still think the cohort version is worth asking: it preserves the paper’s strict-majority rule and subversion objective in a credible large-population regime, and makes explicit where the known hardness sits—among network roles rather than individual agents.
Theorem 17 is a real computational result, so the case against cannot rest on a lack of something to mirror. The weakness is that the proposed cohort model gets its exact correspondence by treating each vertex of the reduction as a role. Those roles are the graph’s distinct positions: their neighborhoods carry the SAT gadgets. Giving each role a large mass does not make its members behave like the paper’s agents. The paper updates one agent at a time; the proposal flips an entire cohort as one actor. That is a substantive change to the campaign dynamics.
The natural repair is to let members of a role update individually and track the mass of each opinion within each role. With an interaction kernel \(W\), one could ask whether an initial minority of mass at most \(1/4-\varepsilon\) can reach a stable state with at least half the population supporting \(1\). But an individual update moves zero mass in the continuum. To get nontrivial reachability, the model must specify how positive mass moves, or how finite update sequences are rescaled into continuous time. Those choices can change which cascades and stable states are reachable. So this is a coherent continuous control problem, but the paper does not supply a canonical limit of its asynchronous, one-agent-at-a-time process.
There is also a tension in the type space. If a type includes an agent’s exact neighborhood, the arbitrary graphs used in Theorem 17 generally require a distinct type per vertex: \(\tau\) grows with the graph, and each type’s mass can vanish as the population grows. If types instead describe shared roles, the model imposes block-structured interactions that the paper’s graphs do not have. The platform scenario makes that restriction plausible, but it studies populations connected through common role-level exposure, not repeated agents with the paper’s individual neighborhoods.
That is the strongest negative case: the exact cohort mirror changes who can update, while the more faithful split-mass version needs a new and consequential definition of reachability. It weakens the case for this paper as a clean continuization target, but it does not establish that no worthwhile mirror exists. The split-mass network question remains plausible, and that is where this case against is least convincing.
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.