| paper | d4-conf-atal-2015-0219 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 4.1 is a named algorithmic result, but it already optimizes over real-valued aggregate adoption and fractional spending. Normalizing those quantities as population mass gives the same scalar state equations and \(O(T)\) stage search, leaving no distinct continuous-population question for this anchor.
fails bit b-already — the paper is already native to the continuum
The theorem already solves the same real-valued aggregate scheduling problem after normalization.
fatal: True
Theorem 4.1's idealized scalar control result only; no named result for the learning-by-doing setting or experimental comparisons is mirrored.
The strongest positive case is the paper’s homogeneous, aggregate model. A plausible regime is a utility running a standard rebate programme across tens of thousands of similar homes. Each home has the same adoption response and faces the same seeding cost, so the idealization has \(n\gg 1\) households but just \(\tau=1\) type. Mass then means the fraction of that cohort that has adopted or can be induced to adopt. This is a credible high-multiplicity setting, and it preserves the paper’s central simplification: social influence depends on aggregate adoption, not on household identities or network structure.
The named anchor is Theorem 4.1, proved in this paper. Under the paper’s assumptions that the aggregate adoption function \(f\) is strictly convex, strictly increasing, and satisfies \(f(D)>D\) for \(D>0\), it proves that an optimal policy spends the entire budget available at one selected stage and spends nothing at the other stages. Best-Stage tests the possible spending stages and chooses the one yielding the greatest adoption at the horizon. The paper describes this as an \(O(T)\) algorithm. It contains no numbered hardness result of its own that would provide a comparable anchor.
Call the mirror Continuum Dynamic Influence Maximization with Exogenous Budget Growth. An instance consists of a unit-mass cohort of one household type, initial adoption mass \(d_0\), initial mass-equivalent seeding budget \(b_0\), a finite horizon \(T\), and a gross budget multiplier \(q\). It also includes an exactly specified aggregate transition function \(F\), strictly convex and increasing on the reachable range, with \(F(d)>d\) there for \(d>0\). The instance guarantees that adoption capacity does not bind during the horizon. A finite computational encoding of \(F\) must be part of a concrete implementation; a rational formula is one option.
At stage \(t\), choose a mass \(x_t\) to seed, with \(0\le x_t\le b_t\). The state evolves as
\[
d_{t+1}=F(d_t+x_t),\qquad b_{t+1}=q(b_t-x_t).
\]
Here \(d_t\) is adopted population mass, and \(x_t\) is newly seeded mass. The question is to find a feasible schedule \((x_0,\ldots,x_{T-1})\) maximizing \(d_T\); a solution is an optimal schedule and its resulting adoption mass. The schedule is unrestricted: it may split the budget across stages. The one-stage policy is the consequence to be proved, not a restriction built into the question. I use \(q\) for the gross multiplier; the paper’s budget recurrence writes this as \(1+\delta\).
The expected direction is Class A: Theorem 4.1’s convexity argument applies to mass-valued states and actions, and reduces the search to \(T\) candidate stages. That gives \(T\) evaluations of the corresponding adoption outcomes. The paper’s \(O(T)\) claim does not settle bit complexity for every encoding of \(F\), or the cost of exact repeated evaluation; those depend on how the transition function and horizon are represented. Further questions include the exact complexity under a specified succinct encoding, how robust the result is to approximate evaluation of \(F\), and what happens when costs depend on cumulative adoption rather than time. The paper’s learning-by-doing experiments suggest that this last change can defeat the one-stage policy.
This mirror covers Theorem 4.1’s idealized dynamic problem. It does not claim to mirror the paper’s experimental findings as named results, nor does it cover its more realistic learning-by-doing setting, where Best-Stage becomes suboptimal. The authors should recognize the proposed problem: it keeps the same aggregate adoption dynamics, budget timing decision, and terminal adoption objective, while making the population whose adoption is measured continuous.
The weakest point is that the mirror changes the interpretation of the aggregate state more than the mathematics. The paper already allows fractions of budget, and its convexity proof works with real-valued quantities. Its model also suppresses individual and network structure from the outset. So this is a plausible high-multiplicity version of their problem, but it may yield little new algorithmic insight beyond making an implicit mass interpretation explicit.
The strongest case against is a collision at the only anchor the proponent raises: Theorem 4.1. The theorem is a computational result, so the case cannot rest on there being nothing to mirror. But its formal model already reduces the society to a real-valued aggregate state \(D\), uses real-valued budget fractions, and applies \(f\) to real inputs. It tracks no individual identities, network positions, or type profile. The finite-agent simulation in the experiments does not change the object Theorem 4.1 analyzes.
For a homogeneous population of size \(N\), normalize adopters, seeds, and budget by \(N\): \(d=D/N\), \(x=X/N\), and \(b=B/N\). On any reachable range where population capacity does not bind, define \(F_N(z)=f(Nz)/N\). The dynamics become
\[
d_{t+1}=F_N(d_t+x_t),\qquad b_{t+1}=(1+\delta)(b_t-x_t),
\]
which is the proposed mirror’s model. Best-Stage already optimizes this real-valued schedule over \(T\) possible stages. There is no discrete-to-continuous identity for the paper to establish: its theorem is already stated over aggregate quantities that can be interpreted as masses. Calling those quantities fractions of households makes the interpretation explicit but leaves the computational question unchanged.
The strongest alternative would give households types based on adoption propensity, home characteristics, and cost, then let a planner allocate seeding mass across types while aggregate uptake changes future costs. That is a plausible model of the paper’s experiments. If type-specific states and targeting matter, though, it changes the homogeneous, non-discriminating control problem of Theorem 4.1. The paper identifies heterogeneity as future work but proves no computational result about that problem. If types can instead be folded into a single aggregate transition function, the model returns to the scalar control problem the theorem already handles.
So the proposed mirror’s weakness is sharper than “the answer may be easy”: its sole named result already solves the mass-valued version. A typed, learning-by-doing extension could still be worthwhile, and I cannot rule it out as a research project. The case against is weaker there; the paper does not establish that extension as a computational result to continuize.
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.