Spatial Conformal Inference through Localized Quantile Regression

Hanyang Jiang; Yao Xie

REVIEW 3 major objections 4 minor 28 references

Spatial Conformal Inference through Localized Quantile Regression

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Localized Spatial Conformal Prediction (LSCP) claims finite-sample conditional coverage under stationarity and spatial mixing instead of exchangeability.

desk verdict The algorithm is a reasonable extension of localized conformal prediction and the empirical study is solid, but the main theorem's proof and a key assumption both fail under the paper's own setup, so the advertised finite-sample guarantee is not established. read the letter →

arxiv 2412.01098 v2 pith:J3FFNW3J submitted 2024-12-02 stat.ML cs.LG

Hanyang Jiang , Yao Xie This is my paper · ORCID

classification stat.MLcs.LG MSC 62M3062G1562G08

keywords spatialconformalpredictionquantileregressionconditionalcoveragemixingstationarityintervalsuncertaintyquantificationnon-exchangeabledata

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

This paper proposes Localized Spatial Conformal Prediction (LSCP), a method for building prediction intervals at unobserved spatial locations that does not require the data to be exchangeable or i.i.d. The authors aim to prove that, under stationarity and spatial mixing of the error process, LSCP achieves conditional coverage with a finite-sample gap that shrinks to zero. This matters because classical geostatistical tools like Kriging need strong parametric assumptions, and existing conformal methods for spatial data either rely on i.i.d. assumptions or on infill sampling that is unrealistic for many applications. The paper's central result is a bound on the conditional coverage gap, together with experiments showing tighter and more consistent intervals than existing methods on synthetic and real-world mobile-signal data.

What carries the argument

The load-bearing object is the weighted empirical cumulative distribution function of the non-conformity scores, $\widehat{F}_{n+1}(y) = \sum_{i=1}^n \omega_n(X(s_i)) \mathbf{1}\{\widehat{\varepsilon}(s_i) \le y\}$, with weights learned by quantile regression (implemented via quantile random forests) over the k-nearest neighbors of the test location. The proof machinery combines this with a strong-mixing coefficient $\$\alpha$(a,b)$ for the stationary error field, a weight-decay condition (Assumption 4.1), and a bound on the estimation error of the base model (Assumption 4.2). Lemma 4.5 bounds the distance between this weighted CDF and the true error CDF, and Theorem 4.6 converts that into a conditional coverage gap.

What would settle it

Run LSCP on a stationary strongly mixing Gaussian random field observed on an increasingly dense grid over the unit square, estimating $\mathbb{E}_{d\sim g_n}\alpha_1(d)^2$ from the actual locations; if that average decays slower than $1/$n^{2}$$, the experimental setting violates Assumption 4.3, so any observed coverage is not covered by Theorem 4.6.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that casting conformal prediction as a localized quantile regression problem allows conditional coverage guarantees for spatial data under dependence. Specifically, Theorem 4.6 states that for any $\$\alpha$ \in (0,1)$ and sample size $n$, the conditional coverage gap is bounded by $4L_{n+1}\delta_n + 6M_n $n^{{(1+\gamma)/2}}$ + (4+2M+4\sqrt{M}g(b)) $n^{{-2\gamma/3}}$(\log_2 n+2)^{4/3}$, where the first term is the estimation error of the base model, the second reflects how fast the conformal weights decay, and the third reflects the strength of spatial mixing. Under the paper's assumptions the right-hand side can be driven to zero, so the prediction interval gets conditionally valid coverage asymptotically without exchangeability.

Load-bearing premise

The finite-sample bound requires the average squared strong-mixing coefficient to decay as $1/$n^{2}$$, meaning dependence between calibration locations must disappear unusually fast as the sample grows; if real spatial dependence decays more slowly, the theorem's coverage guarantee may not apply.

Editorial extensions

If this is right

LSCP provides conditional, not just marginal, coverage asymptotically for spatial data without exchangeability, a strict improvement over standard split conformal methods.
The guarantee holds for any base prediction model $\hat{f}$, so the method can wrap around black-box machine-learning predictors.
Because the theory does not rely on infill sampling, LSCP extends naturally to spatio-temporal settings where the sampling locations are not dense.
Empirically, LSCP meets the target coverage with narrower and more spatially consistent intervals than GSCP, SLSCP, LCP, and EnbPI on synthetic and real mobile-signal datasets.

Reading between the lines

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

If the mixing-rate assumption were relaxed to allow slower decay (for instance, average squared coefficient of order $n^{-1}$), the theorem would directly cover fixed-domain sampling on a bounded region, which is the setting used in the paper's own experiments.
The data-driven quantile-regression weights can be seen as a learned kernel weight function; a natural extension is to study how these weights relate to optimal kernel bandwidth selection in localized conformal inference.
A direct testable extension is to apply LSCP to spatio-temporal data, such as air-quality or traffic monitoring, checking whether conditional coverage persists when locations are fixed over time.

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.

Desk Editor's Note

read the letter

Colleague,

What you should know about arXiv:2412.01098: the method is a sensible extension of localized conformal prediction, but the paper's central finite-sample coverage guarantee does not hold as stated. Two load-bearing issues: Lemma A.1's proof misuses Assumption 4.2, and Assumption 4.3 is incompatible with the paper's own stochastic design.

The genuine novelty is learning conformal weights through quantile regression rather than fixing a kernel a priori. That is a meaningful step beyond SLSCP and LCP, and the experiments back it up: LSCP produces narrower intervals than the baselines while keeping coverage near the target in most settings. The empirical study is reasonably thorough (three synthetic scenarios, two real datasets).

The problems are in the theory. Lemma A.1 claims Σ_i ω_i |ε_i − ε̂_i| ≤ δ_n^2 from Assumption 4.2, which only bounds the sum of squared errors by δ_n^2/M_n. Applying Cauchy–Schwarz gives at best Σ_i ω_i |ε_i − ε̂_i| ≤ δ_n (when M_n = 1/n). The squared bound is needed to conclude that the weight outside the "large error" set S is o(1) via Σ_{i∈S} ω_i ≤ δ_n. Without that step, the CDF gap is not established. This is not a cosmetic issue; it is central to Lemma A.1, which feeds directly into Theorem 4.6.

Just as serious, Assumption 4.3 requires E_{d∼g_n} α_1(d)^2 ≤ M/n^2. Under the stochastic design stated in Section 2, locations are i.i.d. from a fixed distribution g, so g_n = g. Then the expectation is a constant—it cannot vanish as n grows unless the field is independent at almost all distances. The paper's own justification ("a bigger n implies sampling from a larger area") is at odds with a fixed sampling distribution. The theorem therefore does not apply to the paper's own experimental protocol, which uses the unit square.

If the authors switch to increasing-domain asymptotics, the assumption becomes plausible, but then the experiments and the stated stochastic-design framework need to change. As it stands, the theoretical contribution is not established. The method may still work in practice—plenty of conformal methods are heuristic—but the abstract's claim of "finite-sample bounds on the conditional coverage gap" is not supported.

I would send it to a serious referee: the algorithmic idea is worth engaging with, and the flaws are identifiable and possibly fixable. But I wouldn't cite the theorem as a guarantee in my own work.

Referee Report

3 major / 4 minor

Summary. The paper proposes Localized Spatial Conformal Prediction (LSCP), a conformal prediction method for spatial data that constructs prediction intervals by applying localized quantile regression to nonconformity scores. The authors claim finite-sample bounds on the conditional coverage gap under stationarity and spatial mixing rather than exchangeability, and they provide asymptotic guarantees as a corollary. The empirical section compares LSCP with existing methods on synthetic and real-world spatial datasets, reporting tighter intervals with maintained coverage.

Significance. The problem addressed is important: extending conformal prediction to spatial settings without exchangeability is a natural and timely goal. The proposed method, which learns localization weights through quantile regression, is a sensible practical contribution, and the numerical results show consistent improvements over the baselines considered. If the theoretical guarantee were valid, it would be a valuable addition to the spatial conformal prediction literature. However, the central theorem is not established due to a proof error and an assumption that appears incompatible with the paper's own experimental design; the theoretical contribution therefore cannot be credited as it stands.

major comments (3)

[Appendix A, Lemma A.1 proof, Eq. (6)] The proof asserts that sum_i omega_i |epsilon_i - hat(epsilon)_i| <= M_n sum_i |epsilon_i - hat(epsilon)_i| <= delta_n^2, citing Assumption 4.2. Assumption 4.2 only bounds the sum of squared errors by delta_n^2 / M_n. By Cauchy-Schwarz, sum_i |epsilon_i - hat(epsilon)_i| <= sqrt(n) delta_n / sqrt(M_n), so the weighted L1 sum is at most sqrt(M_n n) delta_n, not delta_n^2. Consequently, the bound sum_{i in S} omega_i <= delta_n in Eq. (7) does not follow (Markov gives at most sqrt(M_n n), which need not be small). This invalidates Lemma 4.4, and since Theorem 4.6 relies on Lemma 4.4, the main coverage guarantee is unsupported.
[Section 4.3, Assumption 4.3] Assumption 4.3 requires E_{d~g_n} alpha_1(d)^2 <= M/n^2, where g_n is the distribution of the distance between two sampled locations. In the stochastic design described in Section 2, locations are i.i.d. from a fixed distribution g, so g_n = g for every n and the expectation is a positive constant for any non-degenerate stationary mixing field; it cannot be O(n^{-2}). The paper's own experiments in Section 5.1 sample from the fixed unit grid [0,1]^2. Thus Theorem 4.6 is vacuous for the claimed setting unless the sampling domain grows with n, which is not stated in Section 2 or Section 5. The statement in Section 4.3 that 'a bigger n implies sampling from a larger area' effectively introduces a different sampling model than the rest of the paper.
[Appendix A, Lemma A.2 proof, Eq. (19)] The proof of Lemma A.2 contains an arithmetically inconsistent chain of inequalities. The displayed derivation of E sum_{i<j} alpha_ij goes through a factor n sqrt(...) then jumps to n^2 sqrt(...), and the final '<= n sqrt(M) g(b)' does not match the preceding factor. Because Lemma A.2 is the basis for Lemma 4.5 and Corollary A.3, the probability bound used in Theorem 4.6 is not established. Additionally, the constant M is used in Lemma A.2 without being defined in that lemma; it appears to come from Assumption 4.3 but this connection is not stated.

minor comments (4)

[Appendix A, Lemma A.2 statement] The statement ends with an unfinished sentence: 'where alpha = .' immediately before the proof.
[Theorem 4.6 and Corollary 4.7] These results use the notation T and bC_{t-1}, whereas the setting and proofs use n and bC_n; the notation should be made consistent for readability.
[Section 3.2] The definition beta* = argmin_{beta in [0,alpha]} (bQ_n(1-alpha+beta) - bQ_n(beta)) implicitly assumes that the estimated quantile function is monotone in beta; for quantile random forests this monotonicity is not guaranteed and should be discussed.
[Section 5.1] The experiments are described as 'beyond the assumptions of our theoretical framework,' but the mismatch between the fixed-domain experimental setup and Assumption 4.3 is not acknowledged; this connection should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LSCP's coverage bound is derived from stated assumptions rather than assumed or fitted.

full rationale

The paper's main theorem is not circular: Theorem 4.6 is a stated consequence of Assumptions 4.1-4.3 via Lemmas A.1, A.2, Corollary A.3, and Theorem A.4; none of these assumes the target conditional-coverage statement. The weights are learned rather than fixed, so the analyzed object (weights satisfying external decay/mixing conditions) may differ from the implemented object (quantile-random-forest weights), but this is a theory-method gap, not a definitional reduction. Self-citations (Xu & Xie 2021/2023; Xu et al. 2024; Jiang et al. 2024) are background comparisons and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. The proof does contain a non-circular technical issue: in Lemma A.1, Assumption 4.2 bounds sum of squared residuals, but the proof immediately uses sum omega_i |epsilon_i - epsiloni_hat| <= M_n sum |epsilon_i - epsiloni_hat| <= delta_n^2, which does not follow from the squared-error bound. This affects correctness, not circularity. Similarly, Assumption 4.3's E alpha_1(d)^2 <= M/n^2 is hard to satisfy under the fixed [0,1]^2 uniform design used in Section 5.1, making Theorem 4.6 inapplicable to the paper's own simulation setting; again this is an applicability concern, not a circular one. Therefore the derivation chain is self-contained and no step reduces to its own input.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central coverage guarantee rests on three assumptions and an additive model. Assumption 4.3's mixing decay condition is especially strong and likely violated in the experimental design; Assumption 4.2 is misused in the proof of Lemma A.1.

free parameters (2)

k (number of nearest neighbors)
Chosen by 5-fold cross-validation; controls the localization neighborhood and affects interval width and coverage.
Quantile Random Forest hyperparameters
Not specified in the paper; the learned weights depend on them, so they are free components of the method.

assumptions (5)

domain assumption Additive noise model Y(s) = f(X(s)) + ε(s) with ε stationary
Section 4.1. The whole analysis of residual CDFs rests on this decomposition.
ad hoc to paper Assumption 4.1: weights satisfy ω_n = o(n^{-(1+γ)/2})
A tailor-made condition to make the error terms vanish; not verified for the learned quantile regression weights.
domain assumption Assumption 4.2: prediction error bounds with δ_n
Requires the fitted model to approximate the truth well; the proof misapplies this as an L1 bound when it is stated as L2.
domain assumption Assumption 4.3: strong mixing with E_d α1(d)^2 ≤ M/n^2
A very fast decay condition that likely fails for fixed-domain sampling; the paper's own simulations sample on [0,1]^2.
domain assumption Fε Lipschitz continuous with constant L_{n+1}
Used to bound the probability of ε being near a threshold; typical but not universal.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spatial Conformal Inference through Localized Quantile Regression." pith.science (2026). https://pith.science/paper/J3FFNW3J

@misc{pith2026241201098,
  author       = {Pith},
  title        = {Pith review of: Spatial Conformal Inference through Localized Quantile Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J3FFNW3J}},
  note         = {Machine review of arXiv:2412.01098}
}

read the original abstract

Reliable uncertainty quantification at unobserved spatial locations, especially in the presence of complex and heterogeneous datasets, remains a core challenge in spatial statistics. Traditional approaches like Kriging rely heavily on assumptions such as normality, which often break down in large-scale, diverse datasets, leading to unreliable prediction intervals. While machine learning methods have emerged as powerful alternatives, they primarily focus on point predictions and provide limited mechanisms for uncertainty quantification. Conformal prediction, a distribution-free framework, offers valid prediction intervals without relying on parametric assumptions. However, existing conformal prediction methods are either not tailored for spatial settings, or existing ones for spatial data have relied on rather restrictive i.i.d. assumptions. In this paper, we propose Localized Spatial Conformal Prediction (LSCP), a conformal prediction method designed specifically for spatial data. LSCP leverages localized quantile regression to construct prediction intervals. Instead of i.i.d. assumptions, our theoretical analysis builds on weaker conditions of stationarity and spatial mixing, which is natural for spatial data, providing finite-sample bounds on the conditional coverage gap and establishing asymptotic guarantees for conditional coverage. We present experiments on both synthetic and real-world datasets to demonstrate that LSCP achieves accurate coverage with significantly tighter and more consistent prediction intervals across the spatial domain compared to existing methods.

Figures

Figures reproduced from arXiv: 2412.01098 by the authors.

**Figure 2.** Illustration of Localized Spatial Conformal Prediction (LSCP). Left: [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The violin plots on the left show the distribution of coverage across different areas, while [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The heatmaps illustrate the width of the prediction intervals for each method across [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages

[1]

Angelopoulos, A., Candes, E., and Tibshirani, R. J. Conformal pid control for time series prediction. Advances in neural information processing systems, 36, 2024

work page 2024
[2]

N., Bates, S., et al

Angelopoulos, A. N., Bates, S., et al. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning , 16 0 (4): 0 494--591, 2023

2023
[3]

F., Candes, E

Barber, R. F., Candes, E. J., Ramdas, A., and Tibshirani, R. J. Conformal prediction beyond exchangeability. The Annals of Statistics, 51 0 (2): 0 816--845, 2023

work page 2023
[4]

Barron, A. R. Approximation and estimation bounds for artificial neural networks. Machine learning, 14: 0 115--133, 1994

work page 1994
[5]

J., and Sun, Y

Chen, W., Li, Y., Reich, B. J., and Sun, Y. Deepkriging: Spatially dependent deep neural networks for spatial prediction. arXiv preprint arXiv:2007.11972, 2020

arXiv 2007
[6]

Statistics for spatial data

Cressie, N. Statistics for spatial data. John Wiley & Sons, 2015

2015
[7]

A., Guindani, M., and Gelfand, A

Duan, J. A., Guindani, M., and Gelfand, A. E. Generalized spatial dirichlet process models. Biometrika, 94 0 (4): 0 809--825, 2007

work page 2007
[8]

J., Ramdas, A., and Tibshirani, R

Foygel Barber, R., Candes, E. J., Ramdas, A., and Tibshirani, R. J. The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA, 10 0 (2): 0 455--482, 2021

work page 2021

Show all 28 references

[9]

Does non-stationary spatial data always require non-stationary random fields? Spatial Statistics, 14: 0 505--531, 2015

Fuglstad, G.-A., Simpson, D., Lindgren, F., and Rue, H. Does non-stationary spatial data always require non-stationary random fields? Spatial Statistics, 14: 0 505--531, 2015

2015
[10]

E., Kottas, A., and MacEachern, S

Gelfand, A. E., Kottas, A., and MacEachern, S. N. Bayesian nonparametric spatial modeling with dirichlet process mixing. Journal of the American Statistical Association, 100 0 (471): 0 1021--1035, 2005

2005
[11]

and Candes, E

Gibbs, I. and Candes, E. Adaptive conformal inference under distribution shift. Advances in Neural Information Processing Systems, 34: 0 1660--1672, 2021

2021
[12]

Localized conformal prediction: A generalized inference framework for conformal prediction

Guan, L. Localized conformal prediction: A generalized inference framework for conformal prediction. Biometrika, 110 0 (1): 0 33--50, 2023

2023
[13]

J., Datta, A., Finley, A

Heaton, M. J., Datta, A., Finley, A. O., Furrer, R., Guinness, J., Guhaniyogi, R., Gerber, F., Gramacy, R. B., Hammerling, D., Katzfuss, M., et al. A case study competition among methods for analyzing large spatial data. Journal of Agricultural, Biological and Environmental St...

2019
[14]

N., Heuvelink, G

Hengl, T., Nussbaum, M., Wright, M. N., Heuvelink, G. B., and Gr \"a ler, B. Random forest as a generic framework for predictive modeling of spatial and spatio-temporal variables. PeerJ, 6: 0 e5518, 2018

2018
[15]

Learning cellular network connection quality with conformal

Jiang, H., Belding, E., Zegure, E., and Xie, Y. Learning cellular network connection quality with conformal. arXiv preprint arXiv:2407.10976, 2024

2024 arXiv
[16]

Lahiri, S. N. Central limit theorems for weighted sums of a spatial process under a class of stochastic and fixed designs. Sankhy \=a : The Indian Journal of Statistics , pp.\ 356--388, 2003

2003
[17]

and Wasserman, L

Lei, J. and Wasserman, L. Distribution-free prediction bands for non-parametric regression. Journal of the Royal Statistical Society Series B: Statistical Methodology, 76 0 (1): 0 71--96, 2014

2014
[18]

Mao, H., Martin, R., and Reich, B. J. Valid model-free spatial prediction. Journal of the American Statistical Association, 119 0 (546): 0 904--914, 2024

2024
[19]

and Ridgeway, G

Meinshausen, N. and Ridgeway, G. Quantile regression forests. Journal of machine learning research, 7 0 (6), 2006

2006
[20]

Rio, E. et al. Asymptotic theory of weakly dependent random processes, volume 80. Springer, 2017

2017
[21]

S., Keesee, A

Siddique, T., Mahmud, M. S., Keesee, A. M., Ngwira, C. M., and Connor, H. A survey of uncertainty quantification in machine learning for space weather prediction. Geosciences, 12 0 (1): 0 27, 2022

2022
[22]

J., Foygel Barber, R., Candes, E., and Ramdas, A

Tibshirani, R. J., Foygel Barber, R., Candes, E., and Ramdas, A. Conformal prediction under covariate shift. Advances in neural information processing systems, 32, 2019

2019
[23]

Algorithmic learning in a random world, volume 29

Vovk, V., Gammerman, A., and Shafer, G. Algorithmic learning in a random world, volume 29. Springer, 2005

2005
[24]

and Xie, Y

Xu, C. and Xie, Y. Conformal prediction interval for dynamic time-series. In International Conference on Machine Learning, pp.\ 11559--11569. PMLR, 2021

2021
[25]

and Xie, Y

Xu, C. and Xie, Y. Sequential predictive conformal inference for time series. In International Conference on Machine Learning, pp.\ 38707--38727. PMLR, 2023

2023
[26]

Conformal prediction for multi-dimensional time series by ellipsoidal sets

Xu, C., Jiang, H., and Xie, Y. Conformal prediction for multi-dimensional time series by ellipsoidal sets. arXiv preprint arXiv:2403.03850, 2024

2024 arXiv
[27]

Adaptive conformal predictions for time series

Zaffran, M., F \'e ron, O., Goude, Y., Josse, J., and Dieuleveut, A. Adaptive conformal predictions for time series. In International Conference on Machine Learning, pp.\ 25834--25866. PMLR, 2022

2022
[28]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Angelopoulos, A., Candes, E., and Tibshirani, R. J. Conformal pid control for time series prediction. Advances in neural information processing systems, 36, 2024

work page 2024

[2] [2]

N., Bates, S., et al

Angelopoulos, A. N., Bates, S., et al. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning , 16 0 (4): 0 494--591, 2023

2023

[3] [3]

F., Candes, E

Barber, R. F., Candes, E. J., Ramdas, A., and Tibshirani, R. J. Conformal prediction beyond exchangeability. The Annals of Statistics, 51 0 (2): 0 816--845, 2023

work page 2023

[4] [4]

Barron, A. R. Approximation and estimation bounds for artificial neural networks. Machine learning, 14: 0 115--133, 1994

work page 1994

[5] [5]

J., and Sun, Y

Chen, W., Li, Y., Reich, B. J., and Sun, Y. Deepkriging: Spatially dependent deep neural networks for spatial prediction. arXiv preprint arXiv:2007.11972, 2020

arXiv 2007

[6] [6]

Statistics for spatial data

Cressie, N. Statistics for spatial data. John Wiley & Sons, 2015

2015

[7] [7]

A., Guindani, M., and Gelfand, A

Duan, J. A., Guindani, M., and Gelfand, A. E. Generalized spatial dirichlet process models. Biometrika, 94 0 (4): 0 809--825, 2007

work page 2007

[8] [8]

J., Ramdas, A., and Tibshirani, R

Foygel Barber, R., Candes, E. J., Ramdas, A., and Tibshirani, R. J. The limits of distribution-free conditional predictive inference. Information and Inference: A Journal of the IMA, 10 0 (2): 0 455--482, 2021

work page 2021