| paper | d4-conf-ecai-2023-0440 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Theorem 5 constructively specifies a finite-player attack, but its equilibrium comparison depends on one agent’s defection triggering the leader’s fallback. In a nonatomic population, that deviation leaves the distribution unchanged; the proposed quantile check therefore does not certify the paper’s attack, and positive-mass cohort deviations would define a different game.
fails bit b-changes — the continuum limit changes the question
A zero-mass individual deviation cannot trigger the fallback auction, so the theorem’s comparison between lottery utility and defection utility does not survive the limit.
fatal: True
The proposed threshold check concerns only Theorem 5’s attack condition; it does not preserve that theorem’s equilibrium, and it supplies no separate continuous computational mirror for the paper’s other attack and distributional results.
The paper has no named result classifying a problem as polynomial time, hard, or fixed parameter tractable. The closest positive anchor is Theorem 5, proved here: it is a constructive strategic result, not a complexity theorem.
Theorem 5 says that for \(n=(1+\alpha)m\) buyers with \(\varepsilon\ll v_1<\cdots<v_n\), if some \(k<m\) satisfies
\[
\frac{v_{n-k+1}-B}{v_{n-m}}<\frac{n-k}{n-m},
\]
then EIP-1559 is not Stackelberg attack resilient. The proof constructs contracts: a leading agent is guaranteed inclusion for a negligible tip; a set \(C\) of \(k\) agents, including the leader, tips \(2\varepsilon\); and the others tip \(\varepsilon\), entering a lottery for the remaining slots. If a later contract fails, the contracts revert to bids based on the ordinary auction cutoff. The theorem also covers first-price auctions by setting \(B=0\). Its content is an explicit attack under a valuation condition, not a claim about the complexity of finding attacks.
A plausible high-multiplicity regime is a congested block containing many transactions from recurring bots or applications, with cohorts sharing the same value of inclusion, base-fee exposure, and ability to commit. The type would be that full transaction profile; mass would be the fraction of transactions in each type; and \(q=m/n\) would be the block capacity as a share of demand. Here \(n\) could be enormous while the number \(\tau\) of valuation and capability types is modest. The paper’s uniform- and Pareto-sampling results support large-population analysis, but continuously sampled valuations are almost surely all distinct, so those results do not themselves provide a small-\(\tau\) regime.
The continuous problem I would put to a prover is Quantile Attack Certification. Its input is a finite-support rational distribution \(\mu\) over transaction values \(v>B\), a rational capacity share \(q\in(0,1)\), and a designated leader whose value is represented in the society. Let \(Q_\mu\) be the quantile function, with boundary types fractionally split at atoms. For a proposed protected mass \(\kappa\in[0,q)\), set \(a_\mu(\kappa)=Q_\mu(1-\kappa)\) and \(b_\mu(q)=Q_\mu(1-q)\). Decide whether some \(\kappa\) satisfies
\[
\frac{a_\mu(\kappa)-B}{b_\mu(q)}
<
\frac{1-\kappa}{1-q}.
\]
A solution must return such a \(\kappa\), the protected upper-tail mass, and the corresponding attack certificate: the protected mass gets guaranteed inclusion at negligible tip; the remaining \(1-\kappa\) mass enters a lottery over \(q-\kappa\) slots, so each has allocation probability \((q-\kappa)/(1-\kappa)\). The inequality is the limit of Theorem 5’s condition and ensures the highest-valued unprotected types prefer the lottery to defying the threat; lower-valued types cannot profitably outbid the ordinary cutoff. The designated leader is retained as a tagged, negligible-mass agent so its individual payoff survives the limit.
For this theorem-specific attack family, I expect the certification problem to be tractable: sort the \(\tau\) support values and check the quantile intervals for \(\kappa\), using rational arithmetic. That would be a Class A result, but it would be a new algorithmic question built around Theorem 5, not a complexity result proved in the paper. Theorem 5 is the only anchor I would use; Theorems 12 and 14 support the paper’s large-population motivation but do not establish a different computational problem to mirror.
The weakest point is substantial: Theorem 5 assumes distinct valuations and its contracts are individualized and ordered. Turning those contracts into a type-level continuum problem may require extending the paper to valuation ties and may lose structure essential to its universal commitments. The leader also remains a tagged individual rather than a mass type. So this is a narrow, conditional positive case: it is plausible if the authors accept computing certificates for their explicit attack as their problem; it is not evidence that the paper already studies high-multiplicity computational complexity.
The only anchor is Theorem 5, and it is an equilibrium existence result, not a computational result: it gives a condition under which a particular Stackelberg attack works. Theorem 3 and the later distributional results establish when that condition is likely to hold; they do not add an algorithmic problem. The proposed quantile certificate is therefore a new problem built from the theorem, rather than a continuous version of a computational problem the paper studies.
The deeper difficulty is that the attack depends on atomic players. Contract \(A_u^C\) is indexed by a named player’s place in the contract order, and its fallback depends on whether later named players comply. The proof compares each person’s lottery payoff with their payoff if they personally defy the threat and trigger a return to the auction.
In a population continuum, an individual defector has zero mass. Their departure does not change the aggregate bids, and the leader’s fallback bid cannot make the rest of the population face the finite-player auction outcome. The threat that supports Theorem 5’s equilibrium therefore disappears. Normalizing the market makes an individual transaction’s share of block capacity vanish; the theorem’s key comparison relies on the effect of one person’s deviation. A tagged, zero-mass leader preserves that person’s payoff by stipulation, but cannot preserve the attack’s effect on the population. Keeping one indivisible slot and one atomic leader would instead make the model a continuum population with an exceptional finite player.
The stronger proposed mirror would be to find or certify an equilibrium attack in a market with a finite-support distribution of transaction types. Recurring transaction cohorts could make that distribution plausible, so I would not reject the mirror on the grounds that transactions lack multiplicity. The obstacle is the strategic unit. If individuals within a type choose contracts separately, the order and contingencies that drive the attack remain individualized, so grouping by type does not compress the game. If a whole type or cohort chooses one contract and can deviate as a unit, the threat can affect positive mass—but the players are then cohorts or coalitions, and the result depends on a new collective-deviation model. The paper supplies no natural continuum version of its universal-commitment game to select among these choices.
The proposed quantile check avoids that issue by testing the theorem’s inequality over \(\kappa\). It may be a reasonable calculation for a chosen mass-allocation model, with sensible conventions for ties. It does not establish that the attack remains an equilibrium in a continuum society. A more complete resilience question would need a continuum game and equilibrium concept first; different choices about individual deviations, cohort commitments, and lottery allocation can change the object being analyzed.
So the negative case is strongest against this anchor’s specific strategic mechanism, not against every possible continuous auction model. A worthwhile mean-field or cohort-level attack problem can be imagined, but its central modeling choices would define a new game rather than continue the paper’s individual universal-commitment attack.
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.