acceptodds
Under review as a conference paper at ICLR 2027

Pessimistic Minimax Learning for Public-Private Information Games under Unilateral Coverage

Abstract

We study offline learning in two-player zero-sum contextual games with public and private information, motivated by strategic settings such as auctions and negotiations with private valuations. We introduce unilateral prescriptive concentrability and show that asymmetric information can change offline coverage through its effect on equilibrium behavior. For finite state-action spaces, we develop a pessimistic algorithm with an exploitability rate, matching the standard sample-size dependence for fully observed minimax games. We further develop a pessimistic policy mirror descent framework, PPA-PMD, for general function approximation and obtain a unified exploitability rate with no-regret actor updates. Together, these results provide the first theoretical framework for offline equilibrium learning under public-private information constraints.

Primary area:
learning theory
Keywords:
offline multi-agent learning, zero-sum games, asymmetric information, pessimism, unilateral coverage

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dUu9OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies offline learning in zero-sum Bayesian games with fully logged training data and private-information policies at execution. It introduces prescriptive coverage, proves that information restrictions can change coverage through equilibrium behavior, and develops pessimistic algorithms with statistical and optimization guarantees.

Critical strengths
  • Detailed proofs support the theoretical results.
  • The paper is clearly presented.
  • The studied offline learning under private information is meaningful.
Critical weaknesses
  • Missing experiments.
  • The absence of experiments is not itself disqualifying for a theory paper. However, the claims about scalability and overcoming the tabular method’s memory limitations have neither a computational complexity analysis nor empirical support. Exact construction of the pessimistic envelope, conditional expectations, and parameterized mirror-descent updates remain unresolved computational tasks.
  • Although some related works are cited, the deficiency is the lack of a precise comparison with their assumptions and proof mechanisms.
Questions for authors
  • Can C_F be bounded meaningfully for a useful structured function class?
  • Which implementable parameterized actor satisfies Assumption 5.3?
  • Is there a matching lower bound supporting “optimal learning”?
  • What computational evidence supports PPA-PMD’s scalability?
1 comment

Dear Reviewer,

Thank you again for your constructive feedback! We wanted to kindly follow up on our earlier response and would appreciate hearing whether our revisions have addressed your concerns.

If so, we would be grateful if you could consider updating your assessment. We are also happy to clarify any remaining questions.

Thank you for your time and consideration!

Reviewer_dA4FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies offline equilibrium learning in two-player zero-sum contextual games with public and private information, showing how information asymmetry affects data coverage through equilibrium behavior. It employs unilateral coverage condition and pessimistic algorithms for finite spaces and general function approximation, with exploitability guarantees.

Critical strengths

The formulation is clear, and the analysis shows how private information changes equilibrium behavior and how offline learning can utilize unilateral coverage assumption?

Critical weaknesses

The public–private information setting is interesting, but the learning analysis appears to adapt established pessimism and no-regret methods. Full observation of private types during training and the single-stage setting limit the technical challenges. The paper should clarify what existing offline-game analyses with general policy classes fail to cover and what new techniques it introduces to address these challenges.

Questions for authors

NA

1 comment

Dear Reviewer,

Thank you again for your constructive feedback! We wanted to kindly follow up on our earlier response and would appreciate hearing whether our revisions have addressed your concerns.

If so, we would be grateful if you could consider updating your assessment. We are also happy to clarify any remaining questions.

Thank you for your time and consideration!

Reviewer_GLAiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies offline learning in single-stage two-player zero-sum games with a public context and private types, assuming full-state logged data. It introduces a unilateral prescriptive concentrability coefficient CPres⋆C^\star_{\mathrm{Pres}} and shows that the information structure can change the coverage requirement relative to the fully observed coefficient CFull⋆C^\star_{\mathrm{Full}}. For finite settings, the proposed PPA-Exact achieves an O~(1/n)\tilde{O}(1/\sqrt{n}) exploitability bound with nn samples. For general function approximation, the proposed PPA-PMD achieves an O~(1/n+1/T)\tilde{O}(1/\sqrt{n} + 1/\sqrt{T}) bound after TT iterations of no-regret actor updates.

Critical strengths

Offline equilibrium learning under asymmetric information is a relevant extension of the offline zero-sum game literature. The paper is clearly organized.

Critical weaknesses
  1. The technical novelty is not entirely clear to me. Since the game has no transitions and the logged data contain the full state, the public belief b(x1,x2∣c)b(x^1, x^2 \mid c) can be estimated directly. The analysis then appears to largely follow the standard pessimism analysis of Cui & Du (2022a) and Zhong et al. (2022) with an additional belief-estimation error ∥b−b^∥1\|b - \hat{b}\|_1. Section 5 also seems to mainly combine existing components. It would be helpful if the authors could explain which part of the analysis requires new techniques due to the public–private structure.

  2. The paper claims "Optimal Learning in Finite Settings", but I could not find a lower bound supporting this claim. Moreover, the bound in Theorem 4.1 scales with ∣A∣=∣A1∣∣A2∣|\mathcal{A}| = |\mathcal{A}_1||\mathcal{A}_2|. For fully observed offline zero-sum games, Cui & Du (2022b) achieve ∣A1∣+∣A2∣|\mathcal{A}_1| + |\mathcal{A}_2| dependence using strategy-wise bonuses, and this work is cited in the paper. A discussion of this gap would be helpful.

  3. Assumption 5.2 seems strong. It must hold for every subset G⊆F\mathcal{G} \subseteq \mathcal{F} containing rr, with the supremum taken inside the expectation. The deviation set also includes the data-dependent empirical best response π^2\hat{\pi}_2. It is not clear to me how restrictive this assumption is in practice.

  4. There is no empirical evaluation. Even a small synthetic experiment would help illustrate the results.

Questions for authors
  1. Which step of the analysis requires new techniques compared with the fully observed offline zero-sum setting?
  2. Can the authors give an example function class for which Assumption 5.2 holds with a reasonable CFC_F?
  3. Are the motivating applications (auctions, procurement, negotiation) typically modeled as two-player zero-sum games? If so, could the authors give a concrete formulation or references? If not, could the authors clarify how the results apply to them?
3 comments

Dear Reviewer,

Thank you again for your constructive feedback! We wanted to kindly follow up on our earlier response and would appreciate hearing whether our revisions have addressed your concerns.

If so, we would be grateful if you could consider updating your assessment. We are also happy to clarify any remaining questions.

Thank you for your time and consideration!

I thank the authors for the detailed reply and additional experiments. My concern about novelty remains and I am still not convinced that this paper makes a substantial technical advance over prior work. Therefore, I maintain my original score.

Thank you for taking the time to revisit our response.

We would like to further clarify our technical novelty by highlighting the points where the standard analyses in previous work do not apply.

  1. Equilibrium-induced coverage. In existing fully observed analyses, unilateral coverage is defined around the equilibrium of the fully observed game. In our setting, the information constraint changes the equilibrium itself, so this coverage characterization cannot simply be reused with a smaller policy class. Proposition 3.3 and Appendix C.1.3 show that once each information structure is evaluated at its own induced equilibrium, the resulting coverage coefficients can differ by an arbitrarily large factor in either direction. Thus the new coverage phenomenon comes from the equilibrium induced by the information structure.

  2. The standard pessimism step fails. In a fully observed game, a pointwise pessimistic reward estimate r‾≤r\underline r\le r directly implies a pessimistic game value. This implication is false in our setting, because the learned objective is evaluated under the estimated public belief b^\widehat b, while the true game value is evaluated under bb. Therefore, even with a pointwise pessimistic reward estimate, the standard pessimism proof does not go through. We establish an approximate-pessimism inequality that controls this belief mismatch through the state-action uncertainty bonus, which then allows unilateral prescriptive coverage to be applied.

  3. Per-function error control is insufficient under function approximation. Our pessimistic critic takes the lower envelope of a data-dependent version space separately at each state-action pair. Consequently, the relevant error involves

E![sup⁡f∈G(f−r)2],\mathbb E!\left[\sup_{f\in\mathcal G}(f-r)^2\right],

rather than a collection of separate per-function errors. This distinction is, in fact, substantive. Appendix C.6 gives an example where the behavior and target distributions are identical, yet the envelope error is a factor dd larger than the largest individual-function error. We therefore introduce prescriptive uniform FA concentrability to control this entire version-space envelope and provide sufficient conditions for finite nonlinear and linear function classes.

Our goal is to understand and characterize how public-private information changes offline equilibrium learning. Rather than introducing fudamentally different machinery for its own sake, we identify where the information structure changes the equilibrium, coverage requirement, pessimism argument, and function-approximation analysis. We believe the results above and in our paper provide a thorough theoretical characterization of these effects. To our knowledge, this is the first theoretical framework for offline equilibrium learning in two-player zero-sum games under public-private information constraints.

Thank you again for your careful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.