| paper | d4-conf-atal-2021-0213 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | no |
Theorem 3.3 is a computational result, but the proposed continuous model changes both the winner being predicted and the dependence between match outcomes. Those changes are not forced by representing a population with mass, and the hardness embedding assigns one type to each team rather than establishing a repeated-type regime. I do not find a continuous question here that preserves the paper’s evaluation problem in a form its authors would recognize.
fails bit b-changes — the continuum limit changes the question
The proposed model predicts a winning type under shared shocks rather than a unique winning team under match-level uncertainty; its one-type-per-team reduction also does not establish hardness in a repeated-type regime.
fatal: True
The proposal addresses only Theorem 3.3; it leaves the scheduling results, Theorem 3.4 and its algorithm, and the experiments alone.
The strongest positive case I can make is a Mass-Type Round-Robin Evaluation problem. It mirrors the paper’s exact winner-probability question in a large league whose participants fall into a small number of interchangeable classes.
My anchor is Theorem 3.3, proved in this paper. For every fixed symmetric outcome set \(O\) and fixed \(r\ge 3\), \(O\)-RRTS-Evaluation is #P-hard even when the input includes an optimal schedule with just \(r\) open matchdays and every match-outcome probability is in \(\{0,\tfrac12,1\}\). The problem asks for the probability that a specified team is the unique tournament winner. The proof reduces counting perfect matchings in a fixed-degree bipartite graph to this evaluation problem.
Here is a continuous version of that question. A type \(t\) describes a participant’s current score, schedule class, and outcome probabilities against each other type. The society is a distribution \(\mu\) over a finite set \(T\) of such types, with positive rational masses summing to one. For each of the \(r\) open matchdays, an input matrix \(q^d\) specifies the fraction of each type’s mass paired against each other type; it is symmetric and satisfies \(\sum_u q^d_{tu}=\mu_t\). Thus each unit of population plays once that day. The input also gives each type’s current score and a rational outcome distribution for each type-pair matchup.
To preserve meaningful uncertainty at the type level, all matches between the same pair of types on the same day share one random outcome, drawn independently across type pairs and days. A type’s final score is its current score plus its average points over the mass of opponents it meets on the open days. The question is: what is the exact probability that a designated type \(p\) has a strictly higher final average score than every other type? A solution is that rational probability.
This is a plausible high-multiplicity regime when a large competition has many participants but relatively few complete profiles—for example, participants in a large online pairwise-ranking arena grouped by current score, schedule pattern, and matchup forecast. In a finite approximation, there can be \(K\) participants per type, so the population size is \(K|T|\) while the number of distinct types is \(|T|\). The mass \(q^d_{tu}/\mu_t\) is the share of type \(t\) meeting type \(u\) that day; scores are therefore averaged per participant, rather than represented as counts over named individuals.
Theorem 3.3 gives this mirror a real complexity anchor. At the type level, its reduction can represent each team of the hard instance as a type of mass \(1/|T|\), with the same schedule and outcome probabilities. The resulting question is still exact winner evaluation, and the combinatorics lie in the number of types and their matchup graph. I would expect #P-hardness to persist for this mass-type problem: it is a case where the discrete hardness embeds in the continuous type model, rather than hardness arising from population size alone.
The paper gives some support for authors recognizing the underlying question. It explicitly connects round-robin tournaments to Copeland voting and discusses uncertain pairwise comparisons. The mirror keeps the winner-probability objective and uses population mass for interchangeable participants. It covers Theorem 3.3’s evaluation result only; it does not try to mirror the paper’s scheduling results or its experiments.
The weak point is substantial: the paper asks whether one named team is the unique winner, while this mirror asks whether one type is the unique winning class. It also assumes shared type-pair shocks; the paper’s match-level probability model treats outcomes at the individual-match level. Those choices make the continuum question well-defined and retain uncertainty, but they change the event being predicted and the dependence structure. The proposed mirror stands only if the authors and a referee accept archetype-level winner evaluation as their problem in a high-multiplicity regime.
That limitation suggests useful follow-up questions: does the independent-match model have a meaningful high-multiplicity limit, perhaps through a nontrivial extreme-value formulation for individual winners? And, for the type-level version, can the exact probability be computed efficiently when the critical matchup graph has bounded feedback-set size?
Theorem 3.3 is the right anchor: it is a named #P-hardness result for an exact computational question, and the paper explicitly connects round-robin tournaments to Copeland. The case against cannot dismiss it as lacking a result to mirror.
The difficulty is finding a high-multiplicity version that preserves what the theorem counts. A type must include everything the problem uses about a team: its score, schedule, and remaining matchups. In the reduction, the open matchups encode edges of an arbitrary regular bipartite graph. Teams with different neighborhoods are different types; in general, preserving those neighborhoods means the support grows with the number of teams. Giving each one mass \(1/n\) then re-encodes the finite instance without creating a repeated-type regime. Merging teams into larger classes can reduce the support, but it discards the matchup incidence needed to identify the designated team and its chance of being the unique winner.
The proposed class-level model is the best repair, but it changes both the event and the uncertainty. With independent match outcomes as in the paper, averaging across \(K\) interchangeable teams makes each class’s average score concentrate around its expected score as \(K\) grows. The paper asks about one team’s realized score and whether it alone finishes first. A type’s average score cannot answer that question: it hides which member wins. Keeping a marked individual preserves its identity, but in the continuum it has zero mass; its finite-population chance of winning depends on individual outcome fluctuations and the scale of the population. Those details are lost in the mass distribution.
Sharing one random outcome across every match between a type pair avoids that concentration, but it replaces many independent match outcomes with one common shock. It also makes the predicted winner a type class, rather than a team. A common-shock model could be a sensible macro-tournament in its own right, but it does not continue the paper’s exact winner-probability problem. The proposed mirror stands only after changing the dependence structure and the object whose victory is counted.
The Copeland interpretation offers a stronger alternative: make the voters the continuous society while keeping the candidates finite. But a fixed distribution over rankings determines each pairwise majority outcome, and hence the Copeland winner, deterministically except at ties. To recover a probability of winning, one must add finite-sample ballot noise or uncertainty over the distribution itself. The former restores population size and correlated comparisons; the latter asks a Bayesian uncertainty question about a random society. Either may be worth studying, but neither follows naturally from the paper’s match-level evaluation model.
So the negative case is strongest against a faithful mirror of Theorem 3.3: its hardness depends on individual matchup structure and realized outcomes, while aggregation either loses that structure or changes the winner event. It is not airtight. A community interested in large, repeated classes of competitors under shared shocks might reasonably study the proposed class-level problem. I would treat that as a separate model with a weaker claim to being a continuous mirror of this paper.
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.