acceptodds
Under review as a conference paper at ICLR 2027

Exact Policy Certificate Loss Under Normalization Coarsening

Abstract

Robust policy certificates lower-bound policy improvement under a sensitivity model for unobserved confounding. Group-level inverse-weight normalization can weaken these certificates, even when the policy and full-history propensity reference remain fixed. In a one-step binary-action model, we characterize this loss for arbitrary measurable normalization interfaces: it equals the excess weighted pinball risk incurred by requiring histories in the same group to share a threshold. For sensitivity radii greater than one, grouping is lossless if and only if a measurable group-level threshold belongs to the conditional treated-reward quantile set at almost every history, including with discrete rewards and nonunique quantiles. To certify fitted thresholds with estimated nuisance quantities, we develop rectangular augmented certification (\RAC), combining nuisance correction, confidence-aware fitting and independent betting. For binary rewards and an always-treat target, simultaneous full-history nuisance envelopes yield a finite-sample-valid lower confidence bound on policy improvement after fitting. Experiments show that augmentation strengthens the matched protected weak dual, while confidence-aware fitting improves average certificate strength over empirical-mean fitting across tested positive-margin conditions; fixed thresholds remain competitive. Grouping trades structural loss against statistical uncertainty. Under correctly specified low-dimensional nuisance models, structured calibration reduces protection costs at larger history supports.

Primary area:
causal reasoning
Keywords:
causal inference, off-policy evaluation, sensitivity analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MXvEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors study how coarsening the information used for inverse-weight normalization affects robust policy certificates under unobserved confounding. The main theoretical result gives an exact expression for the loss caused by grouping histories: the loss equals the increase in weighted pinball risk that comes from forcing histories in the same group to share a threshold. Then, the authors provide a characterization of when grouping is harmless (i.e., the grouped histories share a valid conditional reward quantile). The paper also introduces rectangular augmented certification (RAC), which combines several factors (e.g., simultaneous nuisance envelopes, augmented scores, threshold fitting, independent betting) to obtain finite-sample lower confidence bounds for binary rewards and an always-treat policy. Finally, the experiments are conducted to examine the trade-off between structural loss from grouping and the statistical benefit of reducing threshold complexity.

Critical strengths
  • The research question is clear and meaningful.
  • The population result is somewhat clean and easy to interpret. The authors also spent time on the finite-sample guarantee, which I appreciate.
  • The experiments are with sufficient elaboration and insights. They do more than report final performance.
  • The authors keep it transparent where the proposed choices do not help. They chose to report where fixed thresholds can be competitive, and show cases where coarsening helps only slightly.
Critical weaknesses
  • My major concern with this paper is that its theory feels somewhat incremental relative to the existing results. The authors did note that the excess-risk identity when interfaces are finite follows by subtracting existing representations from Tan (2024). The paper provides several extensions, but I am not yet convinced that they can lead to something significant enough to justify the weight placed on Theorem 1. I think the paper would be stronger if it showed one concrete consequence that genuinely depends on the new general formulation.
  • As far as I can understand, RAC heavily relies on strong assumptions of the nuisance envelopes. The authors did note that learned/misspecified features require further work. This leaves a gap between the motivation and the strongest finite-sample result. The population study with HC-MNIST is useful, but the paper does not demonstrate the full RAC scheme with realistically estimated high-dimensional nuisance functions.
  • The benefit of confidence-aware threshold fitting is not fully convincing to me. Clear improvement over empirical-mean fitting is a good result, but I do not think it is enough. A more relevant comparison may be against a strong fixed threshold. Experiments show that fixed zero can be slightly better, and this makes me unsure how important the adaptive fitting stage is to the overall contribution.
  • The theory does not yet lead to a practical rule for selecting the normalization interface. Theorem 1 characterizes when grouping loses information, but the experiments mostly evaluate handpicked interfaces. Thus, the paper explains how to measure the cost of a given grouping but leaves open the more practical issue of how to choose the grouping from data (the authors did note this as future work). However, for me, the characterization would have more practical impact if it were used to build a data-driven method to select interfaces.
  • The writing prose of the LLM the authors use as writing assistant actually hinders people from reading the paper. Some sections feel like piles of jargon (the Conclusions section for example). The paper would have been much better if the authors had actually put effort into writing it by themselves.
Questions for authors
  • How realistic are the simultaneous nuisance envelopes in applications with learned representations?
  • Can the shared-quantile characterization be used directly to construct interfaces?
  • What should the threshold fitter optimize in practice?
  • Can we have the comparison between methods in a setting where their inference assumptions hold?
Reviewer_XBsPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper systematically characterizes the policy certificate loss incurred by normalization coarsening in offline policy evaluation under unobserved confounding. Theoretically, it establishes that this loss strictly equals the excess weighted pinball risk resulting from requiring histories in the same group to share a threshold, where grouping is lossless if and only if a measurable group-level threshold belongs to the conditional treated-reward quantile set at almost every history. Methodologically, it develops the rectangular augmented certification framework, which combines simultaneous full-history nuisance envelope corrections, confidence-aware fitting, and independent betting evaluation to provide a finite-sample-valid lower confidence bound on policy improvement. Empirically, the findings reveal that coarsening fundamentally trades structural loss against statistical uncertainty, while structured nuisance calibration substantially reduces protection costs over large history supports.

Critical strengths
  1. Under standard-Borel spaces, it proves that the exact normalization coarsening loss strictly equals the excess weighted pinball risk, and establishes a necessary and sufficient condition for losslessness that accommodates discrete atomic distributions and nonunique quantiles.

  2. The proposed RAC framework dispenses with continuous-density and asymptotic normality assumptions, combining simultaneous full-history nuisance envelope corrections with independent betting inference to provide a finite-sample-valid lower confidence bound on policy improvement that strictly prevents false certification.

Critical weaknesses
  1. The paper presents confidence-aware threshold fitting as a key contribution, but experiments show that a simple fixed threshold outperforms this method in terms of both average LCB and certification rate (Appendix D.3–D.4). For example, under six positive-margin conditions, the fixed threshold achieved 5,901 out of 6,000 successful certifications, while RAC achieved only 5,898 out of 6,000. This makes the benefits of introducing the additional complexity of threshold optimization less convincing.

  2. The effectiveness of RAC depends on simultaneous nuisance confidence envelopes that cover the entire history space. However, nonparametric cellwise calibration becomes extremely conservative as the history support increases. When K=256K = 256 and N=40,000N = 40,000, no positive certificates were generated (0/1,000)(0/1,000) even when the robust margin was positive. Although structured calibration can mitigate this issue, it requires that both the propensity and outcome models be accurately captured by pre-specified low-dimensional linear models. Therefore, the paper has not yet demonstrated that this method can be effectively applied to scenarios more relevant to real-world applications.

Questions for authors
  1. Could the author further explain under what conditions the proposed confidence-aware fitting offers substantial advantages over a simple fixed threshold? Is it possible to provide theoretical conditions or more discriminating experiments to demonstrate this advantage?
  2. RAC optimizes the variance-aware empirical score, with the ultimate goal of achieving a positive betting-based LCB. Are there optimization methods that more directly target the final LCB or certification rate? The relationship between this surrogate objective and the final certification probability could perhaps be explained more clearly.
  3. Cellwise RAC fails to produce any positive certificates when K=256K=256 and N=40,000N=40,000, whereas the advantage of structured RAC depends on correctly specified low-dimensional linear nuisance models. Does the paper explain whether RAC can still produce meaningful certificates when the history space is large and the nuisance models are misspecified?
Reviewer_DFAQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the loss in robust policy-improvement certificates caused by coarsening inverse-weight normalization while retaining the full-history propensity reference and sensitivity bounds. Its main population result expresses this loss as the excess weighted pinball risk induced by requiring histories within a group to share a threshold. It also characterizes losslessness through a shared conditional quantile, allowing for discrete outcomes and nonunique quantiles.For binary outcomes and an always-treat target, the paper proposes rectangular augmented certification (RAC), combining simultaneous nuisance confidence envelopes, bias correction, variance-aware threshold fitting, and independent betting-based evaluation. Synthetic and semi-synthetic experiments examine structural loss, certificate strength, coverage, and nuisance-calibration costs.

Critical strengths
  1. The exact loss identity and shared-quantile criterion clearly explain when normalization coarsening is lossless, including with discrete outcomes.

  2. RAC explicitly accounts for nuisance uncertainty and data-dependent threshold fitting. The ablations help distinguish structural loss from statistical uncertainty.

Critical weaknesses
  1. Limited theoretical novelty. The finite-interface identity follows from Tan (2024), as acknowledged in Section 4.1. The practical significance of the general measurable-interface extension remains unclear. Unclear motivation for coarsening. RAC still requires full-history nuisance estimates and simultaneous confidence envelopes, leaving unclear which major estimation burden grouping removes. Limited evidence for adaptive fitting. Fixed-zero thresholds slightly outperform confidence-aware fitting in aggregate in both Appendices D.3 and D.4. These results do not establish a consistent practical benefit from learning thresholds.

  2. Restricted practical validation. HC–MNIST evaluates population endpoints rather than end-to-end finite-sample certification. Structured calibration relies on correctly specified, fixed-feature nuisance models.Presentation and scope. The abstract and introduction are terminology-heavy and do not clearly prioritize the main contribution. The finite-sample method is limited to binary outcomes and an always-treat target.

Questions for authors
  1. What substantive consequence of Theorem 1 goes beyond existing quantile-duality results?

  2. When does adaptive threshold fitting reliably outperform simple fixed rules?

  3. What statistical or computational savings does coarsening provide when full-history nuisance estimation is still required?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you for your feedback

We thank the reviewer for the time and care taken in reviewing our paper and discussing the revisions. We appreciate your helpful feedback and positive assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.