acceptodds
Withdrawn from ICLR 2027

Spatial Radius and Group Coverage of Conformal Image Sets under Covariate Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstructing a high-resolution image from coarse measurements leaves uncertainty about regional averages. A prediction set can contain the true image with high probability while allowing large differences between accepted images. In this article, we study the smallest and largest averages over all accepted images that match the observed coarse measurements. These ranges give simultaneous bounds for every region whenever the set contains the true image. We derive exact ranges for nested regions and construct an accepted reconstruction with the smallest worst-case regional error. Standard calibration within observed groups preserves coverage when group proportions change and within-group distributions remain unchanged. This guarantee does not require an accurate predictor. For smoothly varying images, a fixed number of regional averages controls approximation error independently of resolution. The calibrated threshold also contributes to the image radius. In controlled chest X-ray experiments, calibrating regional averages without individual pixels reduces the largest regional range by 20% to 44% compared with calibrating both. Retinal experiments show that calibration within grades improves the smallest mean group coverage over marginal calibration.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal prediction, Image reconstruction, Regional averages, Group coverage, Covariate shift, Uncertainty quantification

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_rWXnOpenReview ↗
Rating
2
Confidence
1(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies conformal prediction sets for high-resolution image reconstruction under covariate shift. The method uses nested spatial averages for calibration and computes exact upper and lower bounds of regional averages through a linear-time tree algorithm. The paper also analyzes the relationship between spatial resolution and image-space radius and discusses extensions to more general types of shifts. Experiments on several medical image datasets and a masked autoencoder show that the method can achieve about 0.90 intra-group coverage; compared to pixel-level constraints, using region-level constraints can significantly reduce the prediction set radius.

Critical strengths
  1. The method is designed for precise range computation under nested spatial constraints, and it is simple and computationally efficient.
  2. It theoretically establishes a connection between spatial resolution, image regularity, and achievable predictive radius, which is useful for understanding when coarse-grained constraints can be applied
Critical weaknesses
  1. The method is relatively low in novelty. Group calibration and finite-dimensional covariate-shift calibration are quite similar to existing work, so the paper's main innovation seems to focus on the construction of spatial ranges. The paper needs to more clearly state the substantive differences compared to existing methods.
  2. The experiments do not sufficiently isolate the benefits brought by "exact range computation" itself. In the experiments, the widths of the exact range and separate intervals are very close, and the larger radius decrease mainly comes from replacing pixel-level constraints with 2×2 region constraints. Additionally, the comparison with related conditional conformal methods is not yet very convincing.
Questions for authors
  1. Under matched coverage and computational budgets, how does this method perform compared to the strongest existing finite-sample conditional/covariate-shift calibration methods?
  2. Can the benefits of exact range computation be evaluated separately, without mixing them with the gains from using coarser spatial constraints?
  3. Can experiments under truly continuous covariate shifts be added, instead of mainly focusing on group proportion shifts?
3 comments
AuthorsOpenReview ↗

Response to Reviewer rWXn (1/2)

We thank Reviewer rWXn very much for the thoughtful comments. This first part addresses the calibration comparison and the separate effects of exact ranges and center choice in Appendix S16 (p. S98) and Appendix S13 (p. S88).

Q1. Coverage and computation

Under matched coverage and computational budgets, how does this method perform compared to the strongest existing finite-sample conditional/covariate-shift calibration methods?

Following your suggestion, we have added independent level selection near 0.90 empirical coverage and report actual computation costs. For each method, score construction, and tilt θ∈−4,4\theta\in\\{-4,4\\}, we select a level from 0.860,0.865,…,0.9400.860,0.865,\ldots,0.940 on 100 tuning replications and evaluate 100 new replications. Each uses 400 calibration observations and 200 queries per tilt, sharing calibration draws, queries, and jitter across methods. Achieved mean coverage spans 0.89715 to 0.90450, and conditional linear-program (LP) radii are 0.9–1.7% larger than group-calibration radii (Appendix S16 and Table 35, p. S98).

Scores and methodCoverage at −4-4, 44Radius at −4-4, 44Time (s)
Original, Group0.90120, 0.901650.387360, 0.3691700.021
Original, Empirical common0.89900, 0.901900.385444, 0.3700810.021
Original, Oracle weighted0.89980, 0.901650.385587, 0.3697370.021
Original, Maximum weighted0.90255, 0.904250.387212, 0.3713020.021
Original, Conditional LP0.89865, 0.904500.393212, 0.3753981.560
Jitter, Conditional LP0.89715, 0.900550.391314, 0.3725571.303

Times cover threshold calculation and set geometry for 400 queries. In the common-nominal-level study, our 300-second single-thread cap per method and replication was an equal budget ceiling, not equal elapsed time, and did not bind: the longest run took 4.41 seconds and none failed. At near-matched empirical coverages, group calibration is about 62–74 times faster than the conditional LP with smaller radii. Because their coverage guarantees protect different shift classes, this comparison does not establish universal superiority (Appendix S14, p. S90, and Appendix S16 and Table 35, p. S98).

Q2. Isolating exact ranges

Can the benefits of exact range computation be evaluated separately, without mixing them with the gains from using coarser spatial constraints?

We appreciate this comment and apologize for the earlier lack of clarity. Yes, with the predictor, observation, functional class, and calibrated set fixed, exact leaf ranges reduce the mean maximum regional half-width over separate score intervals by only 0.056% and 0.053%, while the larger 20–44% reduction changes the calibrated class. The fixed-set calculation still gives exact feasible ranges and supports the optimal accepted certificate. Please see Table 32 and its following discussion (p. S90).

We decompose these effects on the same five-U-Net predictions, λ=0.0025\lambda=0.0025, and 32×3232\times32 observations over 100 controlled-shift and 100 official-split replications with 400 queries each.

  1. When we change calibration by removing 16,384 pixel constraints while retaining 5,461 regional averages, the mean largest 2×22\times2 half-range falls from 0.437 to 0.348 on the standardized scale and from 0.249 to 0.140 on the constant scale. This is the class-selection gain (Table 1 and the following discussion, p. 8).
  2. With the class and accepted sets fixed, exact leaf ranges reduce the mean maximum regional half-width from 0.3480 with separate score intervals to 0.3478 under controlled shift and from 0.3462 to 0.3461 on the official split. The tree calculation takes O(∣H∣)O(|\mathcal H|) time, detects empty sets, and gives feasible bounds for an accepted center. The intensity-offset analysis also distinguishes average width from the largest half-range, and we claim no material radius gain (Table 32, p. S90, Theorem S1, p. S5, and Appendix S18, p. S99).
  3. For center choice, let T(c)T(c) be the largest regional error over a fixed nonempty accepted set CC, rCr_C its minimum over accepted centers, and ΠC(μ^)\Pi_C(\widehat\mu) the pixel Euclidean projection of the ensemble mean. Then rC=T(yacc)≤T(ΠC(μ^))≤2radH(C)≤2rC.r_C=T(y_{\mathrm{acc}})\le T(\Pi_C(\widehat\mu))\le2\mathrm{rad}_{\mathcal H}(C)\le2r_C. The mean certificate falls from 0.3707 to 0.3505 under controlled shift and from 0.3697 to 0.3495 on the official split, while projection has smaller realized regional and pixel errors. Center selection changes the certificate, not coverage of the common set (Proposition S37, p. S88, and Table 32, p. S90).
AuthorsOpenReview ↗

Response to Reviewer rWXn (2/2)

Continuing our response, we address continuous within-group shifts (Appendix S22, pp. S103–S104) and differences from existing calibration methods (Section 3, p. 4).

Q3. Continuous covariate shifts

Thank you very much for this insightful question. In Appendix S22 (p. S103), we have added a continuous image population with known within-group density ratios and unchanged conditional fine-image distributions. Independently observed coarse averages are uniform on [1/4,3/4][1/4,3/4], with observed G∼Bernoulli⁡(1/2)G\sim\operatorname{Bernoulli}(1/2). For B=4(Z1−1/2)B=4(Z_1-1/2) and θ∈−4,−2,0,2,4\theta\in\\{-4,-2,0,2,4\\}, the target ratio is rθ=exp⁡(θB)/A(θ)r_\theta=\exp(\theta B)/A(\theta), where A(θ)=sinh⁡(θ)/θA(\theta)=\sinh(\theta)/\theta for nonzero θ\theta and A(0)=1A(0)=1. Fine images have continuous random variation with exact coarse averages and one conditional distribution given (G,Z)(G,Z).

In each of 100 independent replications, we use 400 calibration images and 200 queries per tilt, sharing five fixed predictors, the maximum regional score, nominal level 0.90, and randomization. For the group-specific conditional span, we include all ten functions 1G=grθ\mathbf1\\{G=g\\}r_\theta, and weighted calibration uses the same groups with query mass at infinity. At θ=4\theta=4, group calibration covers 0.5882 (radius 0.04744), randomized conditional calibration covers 0.8980 (Monte Carlo standard error 0.0035, radius 0.05219), and oracle and maximum weighted calibration both cover 0.9136 (radius 0.05400). Randomized conditional coverage is 0.8959 and 0.9001 in the two groups (please see Table 41, p. S104, and Table 42, p. S104).

The ten-feature construction protects each group under each declared tilt, averaging over calibration, target draws, and randomization, whereas a six-feature construction protects pooled tilts and source-group averages. Declared tilts and their fixed convex mixtures specify each guarantee. Across all 100,000 queries per panel, we obtain finite nonnegative conditional thresholds. Computing conditional thresholds takes 1.916 seconds (group-specific) and 3.033 seconds (pooled) per 1,000 queries, versus less than 0.001 seconds for each sorting-based rule. Here, exact ranges equal separate score intervals, which isolates calibration under shift.

We also retain the fixed-cohort brightness study (Appendix S14) and independent level-selection comparison (Appendix S16). On official medical splits, the smallest group coverage is 0.848–0.876 in four cohorts. Estimated-ratio adjustments improve these minima, though four cohorts remain below 0.90. In Proposition 3, we retain unmeasured density-ratio and conditional-image errors in the target-coverage bound.

W1. Differences from existing methods

The paper needs to more clearly state the substantive differences compared to existing methods.

We appreciate your bringing this to our attention, and we apologize for the lack of clarity in our previous manuscript. We differ from existing calibration methods in post-calibration geometry: exact feasible ranges, the minimax accepted center, and resolution-dependent radius bounds. In Algorithm 1 (p. 4), we use Mondrian calibration of Vovk et al. (2005), and Theorem 1 (p. 4) restates group-score coverage. We follow Gibbs et al. (2025) for our finite-dimensional target and conditional LP without claiming calibration novelty (Appendix S6.1, p. S28, and Appendix S14, p. S90).

  1. Exact ranges. Separate score intervals omit observed coarse block averages. On a nonempty accepted set, Proposition 2 (p. 6) gives exact one-level bounds and a radius equal to half the widest range, which we extend to nested partition trees in two linear-cost passes. These are feasible extrema over one common set, and we do not claim meaningful empirical width reduction (Theorem S1, p. S5).
  2. Accepted reconstruction. We minimize worst-case regional error over accepted centers. For four subregions per coarse block, the optimal certificate is at most 3/23/2 times the unrestricted radius (Proposition S1, p. S3). On nonempty tree-constrained sets, the projection comparison holds with sharp factor 2, though it does not imply smaller realized errors (Proposition S37, p. S88).
  3. Resolution. For fixed s,R>0s,R > 0, the regularity class satisfies ∣(I−Pj)y∣V≤R2−sj\\|(I-P_j)y\\|_V\le R2^{-sj} at each dyadic level, where PjP_j replaces pixels by block averages and ∣y∣V=V−1/2∣y∣2\\|y\\|_V=V^{-1/2}\\|y\\|_2. For 0<ϵ≤2R0 < \epsilon\le2R, we bound the approximation remainder by ϵ\epsilon using M≤min⁡V,4(2R/ϵ)2/sM\le\min\\{V,4(2R/\epsilon)^{2/s}\\} summaries, and the calibrated threshold stays in the radius bound. For unrestricted images, if the accepted set defined by M<VM < V linear measurements contains an image in [ρ,1−ρ]V[\rho,1-\rho]^V with 0<ρ≤1/20 < \rho\le1/2, its normalized radius is at least ρ1−M/V\rho\sqrt{1-M/V} (Theorem 3, p. 6).

Thank you again for your careful reading and constructive comments.

Thank you for the constructive responses and discussion. I should clarify that I have very limited familiarity with the literature in this area, and I am concerned that this assignment may not be well aligned with my expertise. Acquiring sufficient background to provide a well-informed assessment requires substantial time, and I therefore acknowledge the limitations of my review. After revisiting the paper and considering the other reviewers’ comments, I remain inclined toward a weak reject. My concerns primarily relate to the writing, figure presentation, and overall organization, as well as the nearly 100 pages of supplementary material, whose necessity and relevance to the main claims are not sufficiently clear to me. These issues detract from the clarity and quality of the presentation. Nevertheless, given my limited expertise in this area, I will continue to consider the substantive arguments raised by other reviewers and remain open to increasing my score.

Reviewer_BGYAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper considers uncertainty quantification for image super-resolution (focusing on specific setup: noiseless averaging over 4x4 pixel blocks), and studies conformal bounds that hold simultaneously for regional averages of the image.

Critical strengths

Different aspect seem to be considered (regions, groups, certain kind of covariate shift) but the presentation is not clear, and it is hard to understand what is actually novel in the paper.

Critical weaknesses
  • Poor presentation.
    It is hard to follow the paper and to understand what is the motivation, the central idea, the method and the novelty.

  • The motivation for this study is lacking.

  • The novelty seems limited. Most of the ingredients are well established (score based on scaling the residual by estimated std, simultaneous bounds based on max, handling groups with Mondrian-based approach, etc).

  • The empirical gains are not clear. The empirical study also seems limited: why don't you consider popular practical SR setups (practical downsampling, observation noise levels) and prominent models?

Questions for authors

n/a

1 comment

We thank Reviewer BGYA very much for the constructive comments. Our added overview and empirical studies are in Section 2 (p. 3), Appendix S15.5 (p. S96), and Appendix S19 (p. S100).

1. Presentation

Poor presentation. It is hard to follow the paper and to understand what is the motivation, the central idea, the method and the novelty.

We are sorry for the lack of clarity earlier. We added a four-step overview before the formulas in Section 2 (p. 3). We first observe coarse block averages and the clinical group, calibrate each group's largest standardized regional deviation (Algorithm 1, p. 4), and intersect the score set with those averages. Under within-group score exchangeability, this retains coverage at least 1−α1-\alpha in each group, averaged over calibration samples (Theorem 1, p. 4). For nonempty intersections, we compute exact regional ranges and a reconstruction minimizing worst-case regional error.

2. Motivation

The motivation for this study is lacking.

We agree that the motivation should be clearer. We target uncertainty for all regions of one reconstructed image at once, consistent with the observed coarse image. Regionwise risk control leaves 9.5% of regions uncovered and covers all regions in no target image, while our standardized sets cover 0.901 of target images with mean regional half-width 0.104 (Table 1, p. 8). Unlike separate intervals and the ensemble mean, our accepted reconstruction matches the coarse averages in this study (Table 8, p. S52). Group calibration protects changed group proportions under unchanged within-group score distributions, not arbitrary within-group changes (Theorem 1, p. 4).

3. Contributions

The novelty seems limited. Most of the ingredients are well established (score based on scaling the residual by estimated std, simultaneous bounds based on max, handling groups with Mondrian-based approach, etc).

We agree with your point that Algorithm 1 (p. 4) uses standard calibration. We do not propose a new score or group calibration method. Our contribution is the geometry of the accepted set after standard calibration. First, we derive exact ranges in two linear-time tree passes, with sharp factor 3/23/2 for the accepted-center certificate on four subregions per coarse block and 2 for projection (Proposition 2, Theorem S1, Propositions S1 and S37). Among minimax centers, we select the unique full-pixel reconstruction closest to the ensemble mean, preserving the certificate and reducing mean maximum regional error by 5.14% and 5.11% relative to the original minimax center in the two PneumoniaMNIST designs (Proposition S41, p. S102, Table 40, p. S103).

Second, for a fixed regularity class and approximation tolerance, we bound the summary budget independently of pixel resolution, while the calibrated threshold remains in the total radius bound. Our unrestricted-image lower bound assumes fewer linear measurements than pixels and an accepted image with all intensities bounded away from zero and one (Theorem 3, p. 6).

Our separate common-threshold coverage analysis retains calibration, shift-approximation, and conditional-image errors. Its sampling bound requires an independently fixed score and the measurable, bounded shift class with covering-number control in Theorem 2, giving inverse-square-root sample-size calibration error up to a logarithmic factor for fixed class parameters (Theorem 2, p. 5, Proposition 3, p. 7).

4. Empirical study

The empirical gains are not clear. The empirical study also seems limited: why don't you consider popular practical SR setups (practical downsampling, observation noise levels) and prominent models?

Following your suggestion, we added pretrained SwinIR-S and noisy and blurred observation studies with limited evaluation scope, and our general response separates the empirical effects. A fixed pretrained SwinIR-S 4×4\times model with five fixed transformations gives controlled-sampling coverage 0.9042 with the standardized score and 0.9014 with the constant score. This applies our construction to a pretrained transformer, where standardized sets are wider than U-Net's here and we claim no sharper sets (Appendix S19, Table 38, p. S100).

Across 100 resamples per setting, joint calibration covers 0.90028 and 0.89958 with uniform block-error amplitudes 0.005 and 0.020, and 0.89880 with Gaussian noise standard deviation 0.010. Under separable blur, 4×44\times4 averaging, and Gaussian noise, the four-region check covers 0.895 using two query images per resample, and measurement constraints give no material width reduction. Exact aligned averages enable the tree calculation, whereas operators whose rows overlap spatial leaves, such as blur before downsampling, use linear or conic programs without that linear-time guarantee. Bicubic downsampling was not evaluated (Appendix S15, p. S94, Appendix S15.5, p. S96, Table 34 and the blur analysis, p. S97).

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

This paper investigates uncertainty in regional average intensities when reconstructing high-resolution images from low-resolution observations. Rather than treating each region independently, the authors compute exact ranges of possible regional averages over candidate images that satisfy the observation and calibration constraints, while accounting for relationships between regions. They also construct a feasible reconstruction with a worst-case regional error guarantee. The paper uses existing groupwise conformal calibration to handle changes in group proportions, assuming that within-group distributions remain unchanged. Additionally, the authors analyze how image smoothness and the number of measurements affect uncertainty. Experiments on medical images and synthetic data evaluate coverage, uncertainty ranges, reconstruction accuracy, and the effects of image resolution.

Critical strengths
  1. The paper provides an efficient way to compute exact regional uncertainty ranges while accounting for dependencies between regions and the observed coarse measurements. It also provides simultaneous regional error guarantees, which is a useful technical contribution.

  2. The proposed reconstruction satisfies the observation constraints while maintaining regional accuracy close to the ensemble mean.

Critical weaknesses
  1. The practical improvement in uncertainty range is limited. In the neural experiment, the exact calculation reduces the range by less than 1%, raising questions about its practical impact.

  2. Coverage under broader distribution shifts remains limited. On separate source and test splits, some groups fall below the target coverage in four of five cohorts, suggesting limited robustness when within-group distributions change.

Questions for authors

Section 4.1 compares the proposed reconstruction with the original ensemble mean. Could the authors also consider a simpler baseline that projects the ensemble mean onto the same observation and calibration constraints, without optimizing the worst-case regional error? Comparing their error bounds would help clarify the additional benefit of the proposed optimization.

1 comment

We thank Reviewer roND very much for the thoughtful comments and address the projection baseline, the small exact-range gain, and within-group shift in Appendix S13 (p. S88), Appendix S20 (p. S101), and Appendix S17 (p. S98).

1. Projection baseline

Section 4.1 compares the proposed reconstruction with the original ensemble mean. Could the authors also consider a simpler baseline that projects the ensemble mean onto the same observation and calibration constraints, without optimizing the worst-case regional error? Comparing their error bounds would help clarify the additional benefit of the proposed optimization.

Following your suggestion, we added the full-pixel Euclidean projection of the ensemble mean onto the same nonempty accepted set C=CZ(x)C=C_Z(x), given by ΠC(μ^)=argmin⁡y∈C∣y−μ^∣22\Pi_C(\widehat\mu)=\operatorname{argmin}_{y\in C}\\|y-\widehat\mu\\|_2^2. Let T(c)T(c) be the largest regional error over CC, rC=min⁡c∈CT(c)r_C=\min_{c\in C}T(c), and roptr_{\mathrm{opt}} the unrestricted regional radius. Proposition S37 (p. S88) gives ropt≤rC=T(yacc)≤T(ΠC(μ^))≤2ropt.r_{\mathrm{opt}}\le r_C=T(y_{\mathrm{acc}})\le T(\Pi_C(\widehat\mu))\le2r_{\mathrm{opt}}. For four equal subregions per coarse block, rC≤(3/2)roptr_C\le(3/2)r_{\mathrm{opt}} (Proposition S1, p. S3). We now choose the full-pixel reconstruction closest to the ensemble mean among all centers with certificate rCr_C and prove its uniqueness and optimal certificate (Proposition S41, p. S102). On the same 100 controlled and 100 official splits, its mean maximum regional error is 0.08973 and 0.08851, versus 0.09459 and 0.09328 for the original minimax center and 0.08662 and 0.08574 for projection. Its certificates remain 0.35054 and 0.34951 (0.37066 and 0.36973 for projection), and mean pixel RMSE decreases in every paired split relative to the original center (Table 40, p. S103).

On our fixed numerical tolerance grid, certification is rare at small tolerances: minimax fractions are 4.4475% and 2.6900% at ε=0.20\varepsilon=0.20, and 95.65% and 94.39% at 0.500.50 (versus 84.09% and 82.16% for projection). Quantile points describe empirical certification curves, and a conditional error bound also depends on certification probability (Appendix S23, p. S105).

2. Exact-range gain

The practical improvement in uncertainty range is limited. In the neural experiment, the exact calculation reduces the range by less than 1%, raising questions about its practical impact.

We apologize for the lack of clarity earlier. We agree that, on the same accepted set with the class fixed, exact leaf ranges reduce mean maximum regional half-width relative to separate score intervals by only 0.056% and 0.053%. The 20–44% reduction changes the calibrated class and is not an exact-range gain. Still, we need the exact calculation to determine feasible extremes, detect empty sets, and support the minimax accepted reconstruction (Table 32 and discussion, p. S90).

For four fine regions k∈Bbk\in\mathcal B_b, let mk=μ^hk(x)m_k=\widehat\mu_{h_k}(x), ak=σ^hk(x)+λa_k=\widehat\sigma_{h_k}(x)+\lambda, and eb=∑k∈Bbmk−4zbe_b=\sum_{k\in\mathcal B_b}m_k-4z_b, with calibrated threshold qq. If the accepted set is nonempty and all four score intervals are unclipped, lj=mj−qajl_j=m_j-qa_j and uj=mj+qaju_j=m_j+qa_j give Lj−lj=max⁡0,−eb−q(∑k∈Bb∖jak−aj),uj−Uj=max⁡0,eb−q(∑k∈Bb∖jak−aj).L_j-l_j=\max\\{0,-e_b-q(\sum_{k\in\mathcal B_b\setminus\\{j\\}}a_k-a_j)\\},\qquad u_j-U_j=\max\\{0,e_b-q(\sum_{k\in\mathcal B_b\setminus\\{j\\}}a_k-a_j)\\}. The mean largest block-average mismatch is about 0.003, giving a mean largest ∣eb∣|e_b| near 0.012. For four equal scales, the threshold term is 2qaj≥2qλ2qa_j\ge2q\lambda, about 0.10 or larger at the reported mean controlled-shift thresholds. This equal-scale case is illustrative, and the full sums above apply to unequal scales (Proposition 2, p. 6, Appendix S7.5, p. S51, and Table 8, p. S52).

3. Broader shift

Coverage under broader distribution shifts remains limited. On separate source and test splits, some groups fall below the target coverage in four of five cohorts, suggesting limited robustness when within-group distributions change.

We agree that group calibration does not protect against arbitrary within-group changes. Under within-group exchangeability, Theorem 1 (p. 4) averages group coverage over calibration samples, and Proposition 1 (p. 5) separately bounds the shortfall for a realized calibration sample under its sampling assumptions. On official splits, the smallest mean group coverage is 0.848–0.876 in four cohorts (Appendix S8.7, p. S58).

Proposition 3 (p. 7) retains calibration shortfall, density-ratio approximation, and conditional-image distribution error in its target-coverage bound, and conditional-image change cannot be identified from unlabeled target covariates alone. Estimated-ratio group and conditional calibration improve the smallest group mean coverage in all five separate-pool cohorts, though four remain below 0.90. These results do not remove either shift term (Appendix S17, p. S98, and Table 36, p. S99).

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

Related papers

Discussion (0)

Sign in to comment.