| paper | Differentiable Economics for Randomized Affine Maximizer Auctions |
| authors | Michael Curry, Tuomas Sandholm, John Dickerson |
| venue | IJCAI 2023 |
| filed under | frontier · tools-data |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper contains no numbered theorem, lemma, corollary, or proposition asserting a computational, approximation, or parameterized result. Its DSIC and individual-rationality claims are structural, while its revenue results are empirical; cited results by Rochet and Roberts are not results of this paper. The proposed LP is a sensible high-multiplicity mechanism-design problem, but it is a re-modeling rather than a mirror of a computational result in the paper.
fails bit a — no named computational result to mirror
No qualifying result is covered: the paper's architecture, learning procedure, strategyproofness claims, and revenue experiments do not supply a named computational anchor.
The strongest honest conclusion is that this paper has no qualifying computational anchor for ChoCo.
The supplied text contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a complexity, approximation, parameterized, or algorithmic classification. Its formal claims are architectural: lottery AMAs are strategyproof and individually rational, while its revenue claims are experimental Tables 1–5. “Roberts’s Theorem” and Rochet’s characterization are cited structural results, not numbered computational results proved in this paper. The paper also cites complexity work such as Conitzer and Sandholm, but that is not a result of this paper.
So there is no anchor to quote, and no legitimate anchor-specific continuous problem or Class A/B/C verdict. Inventing “Theorem 1” from the empirical revenue tables would manufacture the required result.
The strongest charitable extension would be a new problem, not a mirror:
High-Multiplicity DSIC Lottery-Menu Design. An instance has item types \(j\in[n]\), a finite valuation-type set \(T\subseteq\mathbb{Q}_{\ge0}^n\), rational masses \(\mu_t\), and per-capita item supplies \(q_j\). A solution specifies an expected allocation \(x_t\in[0,1]^n\) and payment \(p_t\) for every type, maximizing
\[
\sum_{t\in T}\mu_t p_t
\]
subject to
\[
\sum_t \mu_t x_{tj}\le q_j,
\]
\[
v_t\cdot x_t-p_t\ge v_t\cdot x_{t'}-p_{t'}
\quad\text{for all }t,t'\in T,
\]
and
\[
v_t\cdot x_t-p_t\ge0.
\]
Here \(x_t\) is a lottery allocation in the paper’s ex-ante sense; the continuous object is the bidder population \(\mu\), not merely the randomized outcome. Clearing denominators gives a repeated market with \(N\mu_t\) identical bidders of type \(t\) and \(Nq_j\) copies of item \(j\).
This is plausibly Class A: for explicit finite \(T\), it is a polynomial-size LP. It preserves additive valuations, lotteries, revenue maximization, strategyproofness, and individual rationality, and is close to the paper’s discussion of MenuNet/RochetNet. But its weakness is decisive: it replaces the paper’s learned multi-bidder AMA architecture with a broader direct-mechanism LP and introduces per-capita supply scaling. The authors might recognize it as a sensible high-multiplicity extension, but not as the continuous mirror of a named computational result in this paper.
The negative case is stronger than the proponent acknowledges: this paper has no qualifying computational anchor at all. Its contributions are architectural and empirical. It shows that lottery AMAs are DSIC and individually rational, proposes a gradient-based training procedure, and reports revenues in Tables 1–5. It proves no complexity, approximation, parameterized, or exact-algorithm theorem. Rochet’s and Roberts’s results are external structural characterizations, not computational results of this paper. The learning procedure has no convergence, runtime, or approximation guarantee. Consequently, there is no named problem whose continuous complexity could be classified in ChoCo’s sense.
The proposed “High-Multiplicity DSIC Lottery-Menu Design” does not repair this omission. Its LP is a sensible mechanism-design problem, but it is not the paper’s multi-bidder AMA problem. The constraints
\[ v_t\cdot x_t-p_t\ge v_t\cdot x_{t'}-p_{t'} \]
describe a single-agent menu, or at most an interim mechanism with severe additional assumptions. They do not express the paper’s ex-post DSIC requirement for a multi-bidder mechanism whose allocation and payment depend on the entire bid profile. In the AMA, bidder \(i\)’s payment is defined through the counterfactual outcome after removing \(i\); the proposed LP has no such opponent-dependent allocation or externality.
Adding those dependencies produces the dilemma that the proposed mirror avoids. If bidders remain individually strategic, the mechanism must be a function of a whole profile of reports, with feasibility and incentive constraints across that profile. A continuum of nonatomic agents cannot materially alter the aggregate allocation, so individual strategyproofness becomes largely vacuous. If one retains finite strategic agents or positive-mass strategic groups to avoid that degeneration, one has not actually obtained the paper’s continuous-population object. If instead one aggregates them into type masses, the resulting direct-mechanism problem is a new market-design model rather than a continuization of the learned AMA architecture.
The supply scaling also changes the economics. The paper sells a fixed collection of \(n\) scarce items to \(m\) bidders. The proposed LP gives the unit-mass population per-capita supply \(q_j\), effectively scaling inventory with population. Keeping the paper’s fixed inventory would make per-capita supply vanish as multiplicity grows; scaling inventory produces a different replicated market. Either choice can be studied, but neither is forced by the paper.
A more faithful alternative would retain AMAs and replace repeated bidder types by masses. That fares no better as a ChoCo mirror. With finitely many bidders, the valuation distribution is already a prior over types, not a population being continuized. With a genuine continuum of bidders, the AMA’s maximization ranges over measurable allocation profiles and its VCG-style counterfactual payments become an infinite-dimensional object. The paper supplies no finite representation, computational question, or structural theorem to study there. Treating the valuation prior as the continuous society would also confuse continuity of the input distribution with continuity of the population, which the programme explicitly excludes.
Thus the LP is evidence that a related high-multiplicity mechanism-design problem could be formulated—not evidence that this paper contains a worthwhile continuous mirror. The universal claim that no imaginable adjacent model could ever be useful is not provable, but the stronger programme-relevant conclusion is well supported: this paper offers no computational result to mirror, and its most natural high-multiplicity replacements either discard the AMA/multi-bidder problem or become a different, non-finite mechanism-design problem.
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.