Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Coverage Guarantees for Pseudo-Calibrated Conformal Prediction under Distribution Shift

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that pseudo-calibrated conformal prediction under label-conditional covariate shift retains target coverage of at least 1 − α − L_r(f,P) − L_γ ρ_mix, tying coverage loss to classifier error and shift size.

desk verdict Theorem 1 is a genuinely new and sound bound, but the paper's main practical claim—label-free slack tuning—is not supported; the experiments use oracle target losses. read the letter →

arxiv 2602.14913 v3 pith:7RJTJFS3 submitted 2026-02-16 cs.LG eess.IV

Farbod Siahkali , Ashwin Verma , Vijay Gupta This is my paper
classification cs.LGeess.IV
keywords conformalpredictionpseudo-calibrationdistributionshiftcoverageguaranteesWassersteindistanceramplosslabel-conditionalcovariatesets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Conformal prediction normally guarantees coverage only when calibration and test data are exchangeable; once the target distribution shifts, the guarantee breaks. The paper considers the practical setting where target labels are unavailable and one calibrates using pseudo-labels from a trained classifier. Its central result is an explicit lower bound on target coverage: the coverage gap relative to the nominal level $1-\alpha$ is at most the classifier's source ramp loss plus a Lipschitz-weighted Wasserstein measure of the per-class covariate shift. In other words, if a classifier is accurate on the source and the input shift is small, pseudo-calibrated sets remain nearly valid without any target labels. The paper also shows how to inflate the conformal threshold by a slack parameter to restore a prescribed coverage, and proposes a source-tuned rule that randomizes labels on uncertain inputs and never hurts coverage relative to hard pseudo-labeling.

What carries the argument

The load-bearing objects are the multiclass margin $\gamma_f(x,y)=M_f(x)_y - \max_{k \neq y} M_f(x)_k$, the ramp loss $\ell_r = \min\{\max(1-\gamma_f,0),1\}$, and the score $s(x,y)=-\gamma_f(x,y)$. The key identity is score dominance: $s(x,f(x)) \leq s(x,y)$ for all y, which makes pseudo-calibrated conformal thresholds never larger than true-label thresholds. The transfer step is a Kantorovich-Rubinstein bound: $|L_r(f,Q)-L_r(f,P)| \leq L_\gamma \rho_{\text{mix}}$, where $\rho_{\text{mix}}$ is the label-average $W_1$ distance between source and target class conditionals. Theorem 1 combines these to bound the coverage gap by the source ramp loss plus $L_\gamma \rho_{\text{mix}}$.

What would settle it

Simulate split conformal prediction with pseudo-labels on a synthetic problem with known $L_\gamma$, $\rho$, and $L_r(f,P)$ satisfying Assumptions 1 and 2; compute empirical target coverage over many independent calibration/test draws at a fixed $\alpha$. If for any such instance the empirical coverage is systematically below $1-\alpha - L_r(f,P) - L_\gamma \rho_{\text{mix}}$, the theorem's bound is false. A minimal case: two classes on the line, a linear classifier with known margin slope $L_\gamma$, and Gaussian class conditionals shifted by exactly $\rho$.

Watch

Extended reading notes

Core claim

The paper's main theorem (Theorem 1) states: under identical label marginals and a label-conditional covariate shift bounded by $\rho$ in Wasserstein distance, split conformal prediction calibrated on pseudo-labels $\tilde{Y}=f(X)$ from target inputs achieves $Q(Y_{n+1} \in C_{\tilde{Q}}^{1-\alpha}(X_{n+1})) \geq 1-\alpha - L_r(f,P) - L_\gamma \rho_{\text{mix}}$. Here $L_r(f,P)$ is the ramp loss of the classifier on the source distribution and $L_\gamma$ is the Lipschitz constant of the multiclass margin in the input, while $\rho_{\text{mix}}$ is the label-averaged $W_1$ distance between class conditionals. The proof uses the fact that pseudo-scores $s(x,f(x))$ are always $\leq$ true scores $s(x,y)$, so pseudo-calibrated thresholds are conservative, and transfers the clas

Load-bearing premise

The guarantee rests on knowing (or tightly estimating) the shift radius $\rho$ and the margin Lipschitz constant $L_\gamma$, and on the source and target having identical label marginals; if $\rho$ or $L_\gamma$ are misspecified, or label proportions shift, the promised coverage floor may not hold.

Editorial extensions

If this is right

  • Hard pseudo-calibrated split conformal prediction on the target keeps coverage at least 1−α−L_r(f,P)−L_γ ρ_mix, so a low-error source classifier under a small bounded shift preserves near-nominal coverage without target labels.
  • The coverage loss is explicitly attributable to classifier mistakes and shift size, giving a quantitative tradeoff: improving source ramp loss or reducing the Lipschitz sensitivity of the margin directly tightens the coverage guarantee.
  • For any desired target coverage, Corollary 1 provides a rule to choose threshold slack τ from the hinge loss ratio, converting coverage risk into a tolerable increase in prediction set size.
  • Randomizing pseudo-labels on inputs above an uncertainty threshold yields conformal thresholds that are no smaller than hard pseudo-calibration's, and hence target coverage never decreases for a fixed level.
  • Because the pseudo-vs-oracle threshold gap is bounded by the ramp loss, pseudo-calibration mistakes show up as conservative small sets; the analysis explains why hard pseudo-calibration undercovers and how to counteract it by inflation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct design corollary the paper does not spell out: train classifiers for uncertainty-quantification use by minimizing source ramp loss plus L_γ ρ_mix, i.e., penalize margin sensitivity; the bound then becomes a differentiable training objective for coverage under shift.
  • The label-conditional shift model excludes label shift; the authors note an extra total-variation term would be needed if P_Y≠Q_Y. Extending the bound to combined label+covariate shift with an estimated ratio is a natural next step and would make the method applicable to more realistic shifts.
  • The slack-selection procedure needs L_h(f,Q) on the target, which is unavailable without labels; in practice one would plug in a transfer estimate, making the realized guarantee depend on the quality of that proxy. Testing the sensitivity of the τ rule to misspecified hinge loss would clarify when the design rule is safe.
  • The source-tuned randomization threshold is chosen by sweeping u on labeled source data; one could also select u by minimizing target set size subject to a source-coverage constraint, a concrete extension that would trade off the conservatism observed in the experiments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The manuscript studies conformal prediction under distribution shift when target-domain labels are unavailable, focusing on pseudo-calibrated prediction sets. The main theoretical contribution is Theorem 1, which lower-bounds target coverage by 1 - α - L_r(f,P) - L_γ ρ_mix under assumptions of identical label marginals and bounded label-conditional Wasserstein shift. Corollary 1 extends this to a relaxed threshold with a hinge-loss term. The paper also proposes a source-tuned pseudo-calibration algorithm (Algorithm 1) that interpolates between hard pseudo-labels and random labels based on uncertainty, and reports experiments on MNIST, CIFAR-10, and CIFAR-100. The proofs in the appendix are short and mostly direct, and the core derivation of Theorem 1 appears mathematically sound. The main gap is that the practical slack-selection method advertised in the abstract and introduction is evaluated using oracle target labels, not a label-free procedure.

Significance. If the main bound is taken as stated, Theorem 1 provides a useful integration of domain-adaptation-style arguments with conformal prediction: it gives a transparent, classifier-dependent coverage guarantee under bounded label-conditional covariate shift. The proof is elementary and checkable, and the split-conformal validity step is used correctly. However, the paper's second claimed contribution—a design rule for the slack parameter τ that keeps target coverage above a prescribed level—is not actually delivered in a label-free way. The rule requires L_h(f,Q), the target-domain hinge loss, which is not available from unlabeled target data. The source-tuned algorithm is heuristic: Theorem 2 only establishes a monotonicity property for fixed u, while the selected u⋆ is not covered by any target coverage guarantee. Thus the paper contains a sound conditional theorem but the headline practical method is unsupported. This is a significant gap for a paper whose stated problem is the unlabeled-target setting.

major comments (4)
  1. [Section IV, Eq. (7), and the τ(σ) formula] The proposed design rule for τ(σ) is τ(σ) = 2( L_h(f,Q_σ)/(L_h(f,P)-Δ_P) - 1 ). This requires L_h(f,Q_σ), the target-domain hinge loss of the classifier, which is not observable without target labels. The parenthetical 'without revealing target labels, e.g. a bound via DA, or a proxy estimator' is not supported by any derivation or experiment in the paper. Figure 2 uses exactly this oracle quantity. Since the abstract claims 'we provide a method to design pseudo-calibrated sets that inflate the conformal threshold by a slack parameter,' this is a load-bearing gap. The authors should either provide a concrete label-free upper bound on L_h(f,Q_σ) or explicitly reframe Corollary 1 and the associated experiments as an oracle-based sensitivity analysis.
  2. [Theorem 1, Assumption 2, and the unlabeled-target claim] The main guarantee depends critically on Assumption 2(i), P_Y = Q_Y, and Assumption 2(ii), sup_y W_∞(P_{X|y}, Q_{X|y}) < ρ. The text states that 'when P_Y ≠ Q_Y, the result can be relaxed with an additional total variation distance term between the marginals,' but no such statement is proved or quantified. Because label shift is a very common form of distribution shift, this limitation is central to the claimed scope. The paper should either prove the TV relaxation or clearly state that the results apply only to label-conditional covariate shift with identical label marginals.
  3. [Section IV, Figs. 1 and 2] The figures that are said to show that the bounds 'track' empirical behavior are evaluated using oracle target labels. The text says 'we estimate the ramp loss L_r(f,Q_σ) using the oracle labels of samples from Q_σ,' and the same is true for L_h(f,Q_σ) in the τ-adjusted experiment. These figures therefore validate the bounds as oracle inequalities, not as a label-free procedure. This distinction should be stated explicitly in the text and figure captions, and the claim that the bounds track pseudo-calibration behavior should be tempered accordingly.
  4. [Section III-A, L_γ and ρ estimation] Theorem 1 and Corollary 1 depend on both L_γ and ρ. The paper says L_γ can be bounded via spectral norm methods or estimated, and ρ is an a priori parameter, but the experiments never report how L_γ was obtained for the plotted bound. Without this information, it is impossible to tell whether the dashed curves in Fig. 1 are genuine a priori bounds or were fitted after the fact. Please report the values of L_γ and ρ (or their estimates) used in each dataset and the procedure for obtaining them.
minor comments (4)
  1. [Section III-A, Eq. (5)] Equation (5) invokes the PDF of s#P and the condition 'when it exists.' In multiclass classification with a fixed score function, the score distribution may be discrete or have atoms. Please state the regularity conditions under which (5) holds, or note that it is used only as a heuristic illustration.
  2. [Appendix C, proof of Theorem 1] The proof invokes split-conformal validity without stating the finite-sample quantile convention. The main text uses the empirical quantile with the (1-α)(n+1)/n adjustment. For completeness, the proof should explicitly say that q_{\tilde Q,α} is the \lceil(1-α)(n+1)\rceil/n empirical quantile, so that the result holds exactly in finite samples rather than asymptotically.
  3. [Theorem 2 and Algorithm 1] Theorem 2 is stated for any fixed u. The algorithm selects u⋆ based on source data, and the proof does not directly cover the random u⋆. While one can condition on the source data to extend the argument, this step is not written. Also, the selected u⋆ is not guaranteed to maintain target coverage at 1-α; the experimental gains are heuristic. Please state this limitation clearly.
  4. [Fig. 2 caption and black curve] The black curve in Fig. 2 is described as 'the hinge-loss lower bound' but the exact expression is not given. Please specify whether this is 1-α - L_h(f,Q_σ)/(1+τ/2) and state which value of τ and which loss estimate are used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 1 is derived from first principles; the oracle-label use in the τ-slack experiment is a correctness gap, not a circular step.

full rationale

The central derivation chain is self-contained. Theorem 1 follows from split-conformal validity on pseudo-scores (giving P(S̃≤q)≥1−α), the pointwise inequality s(x,f(x))≤s(x,y), the event inclusions {S−S̃>τ}⊆{γ_f≤−τ/2}, and the DA bound |L_r(f,Q)−L_r(f,P)|≤Lγρmix via label-conditional Wasserstein. Lemma 1 and Corollary 1 are algebraic consequences of these ingredients plus the Kantorovich-Rubinstein inequality. No fitted parameter is renamed as a prediction: the source ramp loss L_r(f,P) and shift magnitude ρ are explicit assumptions/inputs, and the experiments evaluate the bounds using oracle target losses only to plot the conditional inequalities, not to generate the coverage curves. The one legitimate concern is Section IV's τ design rule, which chooses τ from L_h(f,Qσ) (the target hinge loss) and only parenthetically suggests a label-free proxy; because no label-free estimator or DA upper bound for L_h(f,Qσ) is supplied, the 'without revealing target labels' adjustment is unsupported. But this is a missing-support/correctness gap, not circularity: Corollary 1 is a conditional bound, and the adjusted coverage in Fig. 2 is measured empirically rather than being forced equal to the bound's right-hand side. There are no self-citations or imported uniqueness/ansatz results that carry the argument.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the label-conditional shift model and the Lipschitz margin assumption; both are stated explicitly in Assumptions 1-2. No new physical or mathematical entities are introduced. The free parameters are the practical calibration constants (τ, u⋆) and the assumed shift radius ρ. Note that the bound itself is not fitted; the looseness comes from worst-case DA-style arguments.

free parameters (3)
  • slack parameters τ(σ) = 2( L_h(f,Qσ)/(L_h(f,P)-Δ_P) - 1 )
    The relaxed set design uses a τ chosen by matching Corollary 1 to a source-calibrated undercoverage gap, and in the experiments τ is computed using L_h(f,Qσ) estimated from oracle target labels or a proxy; this is a fitted slack that determines the coverage guarantee.
  • uncertainty threshold u⋆ = max over grid U where source empirical coverage ≥ 1-α
    Algorithm 1 selects u⋆ on labeled source data by sweeping a grid; the choice is data-dependent and not covered by a theoretical guarantee.
  • ρ (shift bound) = assumed known / a priori
    Assumption 2(ii) treats ρ as a known specification parameter; in practice it must be guessed or estimated, and the bound's value depends on it.
assumptions (4)
  • domain assumption Label-conditional covariate shift with identical label marginals P_Y = Q_Y and sup_y W∞(P_{X|y}, Q_{X|y}) < ρ
    Assumption 2; the whole bound is built on this model of the shift, which excludes label shift and other shift classes.
  • domain assumption The margin γf is Lγ-Lipschitz in the input for every y
    Assumption 1; needed for Lemma 1 and the loss-transfer step using Kantorovich-Rubinstein.
  • standard math Kantorovich-Rubinstein inequality (Lemma 2) and the standard split-conformal quantile validity
    Appendices B and C; standard tools, but relied upon without reproof.
  • ad hoc to paper The 'ideal' quantile definitions and continuous CDFs; finite-sample quantile estimation error is ignored in the bound
    The paper uses the split-conformal threshold with empirical quantile (2) but the coverage guarantee (6) is stated for the population/idealized score distributions, without tracking the O(1/√n) finite-sample correction to the (1-α)(n+1)/n quantile.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coverage Guarantees for Pseudo-Calibrated Conformal Prediction under Distribution Shift." pith.science (2026). https://pith.science/paper/7RJTJFS3

@misc{pith2026260214913,
  author       = {Pith},
  title        = {Pith review of: Coverage Guarantees for Pseudo-Calibrated Conformal Prediction under Distribution Shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7RJTJFS3}},
  note         = {Machine review of arXiv:2602.14913}
}
read the original abstract

Conformal prediction (CP) offers distribution-free marginal coverage guarantees under an exchangeability assumption, but these guarantees can fail if the data distribution shifts. We analyze the use of pseudo-calibration as a tool to counter this performance loss under a bounded label-conditional covariate shift model. Using tools from domain adaptation, we derive a lower bound on target coverage in terms of the source-domain loss of the classifier and a Wasserstein measure of the shift. Using this result, we provide a method to design pseudo-calibrated sets that inflate the conformal threshold by a slack parameter to keep target coverage above a prescribed level. Finally, we propose a source-tuned pseudo-calibration algorithm that interpolates between hard pseudo-labels and randomized labels as a function of classifier uncertainty. Numerical experiments show that our bounds qualitatively track pseudo-calibration behavior and that the source-tuned scheme mitigates coverage degradation under distribution shift while maintaining nontrivial prediction set sizes.

Figures

Figures reproduced from arXiv: 2602.14913 by the authors.

Figure 1
Figure 1. Coverage on Qσ vs. shift σ for source calibration, hard pseudo-calibration, and source-tuned pseudo-calibration. Dashed curves show the coverage lower bounds from Theo￾rem 1. 0.0 0.5 1.0 1.5 Shift level ¾ 0.0 0.2 0.4 0.6 0.8 1.0 C o v e r a g e o n Q ¾ qQ~ ¾; ® qQ~ ¾; ® + ¿(¾) qQ¾; ® 0 2 4 6 8 10 E x p e cte d s et siz e o n Q ¾ (a) MNIST 0.0 0.5 1.0 1.5 Shift level ¾ 0.0 0.2 0.4 0.6 0.8 1.0 C o v e r a g e o n Q ¾ … view at source ↗
Figure 2
Figure 2. Coverage (solid, left axis) and ESS (dashed, right [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Label Complexity of Class-Conditional Coverage under Distribution Shift

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Under joint covariate-label shift, class-conditional quantile recovery for per-class conformal coverage costs Θ(ε^{-2} log K) target labels per class for classwise threshold procedures, and pseudo-labels buy at most a...

Reference graph

Works this paper leans on

20 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Theoretical foundations of conformal prediction,

    A. N. Angelopoulos, R. F. Barber, and S. Bates, “Theoretical foundations of conformal prediction,”arXiv preprint arXiv:2411.11824, 2024

    arXiv 2024

  2. [2]

    Con- formal prediction under covariate shift,

    R. J. Tibshirani, R. Foygel Barber, E. Candes, and A. Ramdas, “Con- formal prediction under covariate shift,”Advances in neural information processing systems, vol. 32, 2019

    2019

  3. [3]

    Optimizing in-context learning for efficient full conformal prediction,

    W. Deng, S. Park, M. Li, and O. Simeone, “Optimizing in-context learning for efficient full conformal prediction,”IEEE Signal Processing Letters, pp. 1–5, 2025

    2025

  4. [4]

    Mirror online conformal prediction with intermittent feedback,

    B. Wang, M. Zecchin, and O. Simeone, “Mirror online conformal prediction with intermittent feedback,”IEEE Signal Processing Letters, vol. 32, pp. 2888–2892, 2025

    2025

  5. [5]

    Learning minimal volume uncertainty ellipsoids,

    I. Alon, D. Arnon, and A. Wiesel, “Learning minimal volume uncertainty ellipsoids,”IEEE Signal Processing Letters, vol. 31, pp. 1655–1659, 2024

    2024

  6. [6]

    Conformal prediction beyond exchangeability,

    R. F. Barber, E. J. Candes, A. Ramdas, and R. J. Tibshirani, “Conformal prediction beyond exchangeability,”The Annals of Statistics, vol. 51, no. 2, pp. 816–845, 2023

    2023

  7. [7]

    A unifying view on dataset shift in classification,

    J. G. Moreno-Torres, T. Raeder, R. Alaiz-Rodr ´ıguez, N. V . Chawla, and F. Herrera, “A unifying view on dataset shift in classification,”Pattern recognition, vol. 45, no. 1, pp. 521–530, 2012

    2012

  8. [8]

    Distribution-free uncertainty quantification for classification under label shift,

    A. Podkopaev and A. Ramdas, “Distribution-free uncertainty quantification for classification under label shift,” inProceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, ser. Proceedings of Machine Learning Research, C. de Campos and M. H. Maathuis, Eds., vol. 161. PMLR, 27–30 Jul 2021, pp. 844–853. [Online]. Available: ht...

    2021

Show all 20 references
  1. [9]

    Wasserstein-regularized conformal prediction under general distribution shift,

    R. Xu, C. Chen, Y . Sun, P. Venkitasubramaniam, and S. Xie, “Wasserstein-regularized conformal prediction under general distribution shift,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/ forum?id=aJ3tiX1Tu4

    2025

  2. [10]

    Non-exchangeable conformal prediction with optimal transport: Tackling distribution shifts with unlabeled data,

    A. H. C. Correia and C. Louizos, “Non-exchangeable conformal prediction with optimal transport: Tackling distribution shifts with unlabeled data,” 2025. [Online]. Available: https://arxiv.org/abs/2507. 10425

    2025

  3. [11]

    Confor- mal prediction under levy-prokhorov distribution shifts: Robustness to local and global perturbations,

    L. Aolaritei, Z. O. Wang, J. Zhu, M. I. Jordan, and Y . Marzouk, “Confor- mal prediction under levy-prokhorov distribution shifts: Robustness to local and global perturbations,”arXiv preprint arXiv:2502.14105, 2025

    2025 arXiv

  4. [12]

    Adapting prediction sets to distribution shifts without labels,

    K. Kasa, Z. Zhang, H. Yang, and G. W. Taylor, “Adapting prediction sets to distribution shifts without labels,” 2025. [Online]. Available: https://arxiv.org/abs/2406.01416

    2025 arXiv

  5. [13]

    Wqlcp: Weighted adaptive conformal prediction for robust uncertainty quantification under distribution shifts,

    S. Alijani and H. Najjaran, “Wqlcp: Weighted adaptive conformal prediction for robust uncertainty quantification under distribution shifts,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2025, pp. 1732–1741

    2025

  6. [14]

    Calibrating without labels: Source-free conformal prediction using pseudo-labels,

    S. Angelman, R. Nizhar, and J. Goldberger, “Calibrating without labels: Source-free conformal prediction using pseudo-labels,” in Proceedings of the Fourteenth Symposium on Conformal and Probabilistic Prediction with Applications, ser. Proceedings of Machine Learning Research,...

    2025

  7. [15]

    Understanding self-training for gradual domain adaptation,

    A. Kumar, T. Ma, and P. Liang, “Understanding self-training for gradual domain adaptation,” inInternational conference on machine learning. PMLR, 2020, pp. 5468–5479

    2020

  8. [16]

    Gradual domain adaptation: Theory and algorithms,

    Y . He, H. Wang, B. Li, and H. Zhao, “Gradual domain adaptation: Theory and algorithms,”Journal of Machine Learning Research, vol. 25, no. 361, pp. 1–40, 2024

    2024

  9. [17]

    Probably approximately correct labels,

    E. J. Cand `es, A. Ilyas, and T. Zrnic, “Probably approximately correct labels,”arXiv preprint arXiv:2506.10908, 2025

    2025

  10. [18]

    A tutorial on conformal prediction

    G. Shafer and V . V ovk, “A tutorial on conformal prediction.”Journal of Machine Learning Research, vol. 9, no. 3, 2008

    2008

  11. [19]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification,

    A. N. Angelopoulos and S. Bates, “A gentle introduction to conformal prediction and distribution-free uncertainty quantification,”

  12. [2022]

    Available: https://arxiv.org/abs/2107.07511 6 SUPPLEMENTARYMATERIAL(APPENDIX) A

    [Online]. Available: https://arxiv.org/abs/2107.07511 6 SUPPLEMENTARYMATERIAL(APPENDIX) A. Proof of Lemma 1 Fixϵ >0. For eachy, pickπ y ∈Π(P X|y , QX|y )with ess sup (X,X ′)∼πy ∥X−X ′∥2 ≤ρ+ϵ.(9) LetY∼P Y , and conditional onY=ysample(X, X ′)∼π y. Then(X, Y)∼P XY and(X ′, Y)∼Q ...

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.