acceptodds
Under review as a conference paper at ICLR 2027

A Sharp Finite-Window No-Free-Lunch for Coverage-Sufficient Dependence: Exact Block-Conformal Coverage under Structural Envelopes

Abstract

A finite score window cannot identify the dependence functional needed by the block-conformal coverage bound, even over stationary ergodic binary processes. For every window length n, two processes can have the same observed n-window law while their two-block joint-to-product total variation differs by 1/2. Over the stationary ergodic binary class with known uniform marginal, the resulting exact minimax absolute estimation risk is (1-2^-B)/2. A second construction gives identical complete calibration laws but rank-miss probabilities on opposite sides of the nominal release level, forcing any pointwise-valid rule to refuse on a benign member; the construction extends to every conformal rank. The complete score law can identify score dependence but still need not identify full-process mixing. We therefore target the weaker score-block discrepancy that is sufficient for coverage: miscoverage is bounded by the exact conformal rank term plus this discrepancy, itself at most B beta_S(h) for a stationary score process. Because one finite trajectory cannot supply that envelope without additional assumptions, we give a conservative structural route based on an observed finite-state Markov state, fixed emissions, independent transition evidence, and Dobrushin contraction. Controlled experiments test that premise-matched route; named-data diagnostics are not dependence certificates, and all estimated methods still refuse at h=27.

Primary area:
learning theory
Keywords:
conformal prediction, dependent data, mixing conditions, distribution-free coverage, impossibility result

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_WubVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether a finite window of conformal scores can certify the dependence quantity used in block-conformal coverage bounds. It gives two-point constructions of stationary ergodic processes whose calibration windows have identical laws, but whose dependence functionals or rank-miss probabilities differ (Theorems 2.4–2.7, A.6.8). It also gives a conservative positive route under observed finite-state Markov structure (Theorems A.1–A.2). The constructions are explicit, but the paper has serious problems.

Critical strengths

None

Critical weaknesses
  1. The authors seem to have used AI irresponsibly. The manuscript contains revision history (“a previous theorem … is withdrawn,” “an earlier version of this statement … is corrected here,” “the retained V3 simulation”). Internal file paths appear in theorem statements and proofs. It uses software-test language (“Nine regressions cover…,” “48 cells … with zero failures”) and includes build notes about the LaTeX style. This suggests the text was generated and revised with little careful human editing. It reduces confidence that the authors have checked the arguments themselves. There was no AI use declaration.
  2. Very poor clarity. The central results are stated or proved only in the appendix, including the coverage bound (A.4) and the mixing-robust obstruction (A.6.8). The main text is mostly positioning tables, disclaimers and undefined jargon (“information contract,” “premise-matched,” “fail-closed,” “named-data diagnostics certify no premise”). Notation is inconsistent (O_B vs O_n; blocks vs raw scores), and cross-references are broken (two sections numbered A.2; “Proposition A.6.7 in Appendix A.5”).
Questions for authors

None

Reviewer_MShmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper examines whether limited history of observations can influence prediction intervals. The authors constructed two processes that are indistinguishable from the observed history, yet yielded different miscoverage rates under a same interval rule. This seems explain that the historical data alone may be insufficient to verify a coverage claim. Experiments covered public FrozenLake kernel, a synthetic two-state model, geometry selection, traffic data, and ETT/Electricity; the authors reported refusals, over-coverage, and a failed prespecified criterioperiments on real-world time series examples.

Critical strengths

The paper seems established a useful limitation: two processes can generate histories from a same distribution, yet yield different miscoverage rates under a same interval rule. The authors demonstrated this through explicit constructions, and it attempts to explain the reason that historical data alone may be insufficient to verify a coverage claim.

The paper also identified additional information and assumptions of coverage guarantee. It reported cases in which the method cannot produce a finite interval, and experiments that failed to meet the stated success criteria.

Critical weaknesses

The presentation and logical chain are difficult to follow.. For example, there was no briding example between Theorem 2.6 and Theorem A.1, the latter one required observed states and independent transitions. It is unclear what A.1 actually resolved. The geometry, notation, and cross-references also added to the difficulty when reading.

The practical value remains unclear to me. The assumptions seem pretty strong, extra data are substantial, 79–87% of observations were skipped in the experiments? - also the intervals were wide, and refusal rates were high. The comparisons were neither budget-matched nor target-matched (pointwise Split CP vs. block coverage), so it was difficult to understand the balancing point between the benefits and costs.

Questions for authors

Does the proposed method outperform external baselines in the current experiments when evaluated under comparable total data budgets? - and with all methods targeting coverage of the entire future block?

Reviewer_xhDLOpenReview ↗
Rating
3
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents theorems regarding conformal predictions.

Critical strengths

The paper appears to present a rigorous theory.

Critical weaknesses

While the paper presents itself as only theoretical, it is also honest (from what I can tell) that this doesn't automatically apply to any application area.

Questions for authors

I do not have any questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.