{"id":"254cfea5-742c-48d2-98b5-b1dbbf065800","arxiv_id":"2412.01098","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"LSCP uses quantile regression to learn local weights for spatial conformal prediction, but its finite-sample coverage theorem rests on a flawed proof step.","lead":"This paper proposes LSCP, a conformal prediction method that builds prediction intervals for spatial data by learning local weights with quantile regression. It claims finite-sample conditional coverage guarantees under stationarity and spatial mixing, but the main proof contains an error that breaks the central theoretical claim.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 4.3 is incompatible with the paper's fixed-domain sampling model, so Theorem 4.6 does not apply to the stated setting; Lemma A.1 also misuses Assumption 4.2.","rationale":"The reader's weakest assumption identifies the same load-bearing problem: the strong-mixing decay assumption used to prove the finite-sample conditional coverage bound is not satisfied in the paper's own fixed-domain sampling model. My stress-test confirms this concern and adds a second, independent defect in Lemma A.1. The proof of Lemma 4.5/Corollary A.3 relies on E_{d∼g_n} α_1(d)^2 ≤ M/n^2; under the i.i.d. fixed-location-distribution model stated in §2, this expectation does not decay with n. The paper's own experiments use n points uniformly sampled from the unit square, so the theorem cannot be invoked there. Even if one accepted the increasing-domain reinterpretation, Lemma A.1's proof contains a concrete algebraic error: Assumption 4.2 bounds the sum of squared residuals, but the proof treats it as a bound on the weighted L1 sum, leading to an unjustified δ_n^2 and a failing Markov step. These are not disagreements with external consensus but internal inconsistencies between stated assumptions, proof steps, and the theorem's application. The empirical comparisons may still indicate that LSCP is practically useful, and I credit the paper for presenting synthetic and real-data results, but the central theoretical claim of a finite-sample conditional coverage guarantee is not established. Therefore the reader's REJECT verdict stands unchanged.","tokens_in":17337,"tokens_out":16915,"duration_ms":158220,"concrete_test":"Fix the experimental design: sample s_i i.i.d. uniformly on [0,1]^2 with n = 1000, and take a standard mixing rate such as α_1(a) = e^{-a/ρ}. Compute the empirical E_{d∼g_n} α_1(d)^2 over all pairs; it will be a positive O(1) constant, violating the required M/n^2 by orders of magnitude. Separately, re-derive Lemma A.1 under the stated Assumption 4.2 and check whether ∑_{i∈S} ω_i ≤ δ_n follows from ∑_i ω_i |ε̂_i - ε_i| ≤ δ_n; a counterexample with one weight 1/2 and error δ_n shows it does not. If these checks confirm the failures, the coverage guarantee is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee (Theorem 4.6) depends on Assumption 4.3, which requires E_{d∼g_n} α_1(d)^2 ≤ M/n^2, where g_n is the distribution of the distance between two sampled locations. In the stochastic design stated in §2, locations are i.i.d. from a fixed distribution g, so g_n = g for every n. For any non-degenerate stationary mixing field, E_{d∼g} α_1(d)^2 is a positive constant independent of n; it cannot be ≤ M/n^2 for large n unless the field is essentially independent at all distances present in the domain. The only way to satisfy the assumption is to let the sampling domain grow with n, which contradicts the fixed-unit-grid design used in the experiments and the fixed sampling distribution assumed in the problem setup. Thus the theorem is vacuous for the paper's own experimental setting. Independently, Lemma A.1's proof misuses Assumption 4.2: from ∑_i ||ε̂_i - ε_i||^2 ≤ δ_n^2/M_n and ω_i ≤ M_n, Cauchy-Schwarz gives ∑_i ω_i |ε̂_i - ε_i| ≤ δ_n, not ≤ δ_n^2. With only an O(δ_n) weighted-L1 error, the step concluding ∑_{i∈S} ω_i ≤ δ_n fails (Markov gives only ≤ 1), so the residual/noise CDF gap in Lemma A.1 is not established as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":17721,"tokens_out":6571,"duration_ms":55578,"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":[{"comment":"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":"Appendix A, Lemma A.1 proof, Eq. (6)"},{"comment":"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.","section":"Section 4.3, Assumption 4.3"},{"comment":"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.","section":"Appendix A, Lemma A.2 proof, Eq. (19)"}],"minor_comments":[{"comment":"The statement ends with an unfinished sentence: 'where alpha = .' immediately before the proof.","section":"Appendix A, Lemma A.2 statement"},{"comment":"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":"Theorem 4.6 and Corollary 4.7"},{"comment":"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":"Section 3.2"},{"comment":"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.","section":"Section 5.1"}],"recommendation":"reject","confidential_remarks":"The paper addresses a relevant problem and the empirical results are interesting, but the main theoretical guarantee is not established. The proof error in Lemma A.1 is local but load-bearing, and Assumption 4.3 appears incompatible with the paper's stated sampling model and experiments. I do not see how to repair these within the scope of the current manuscript, so I recommend rejection; a revised version with corrected proofs and a coherent sampling assumption could be considered as a new submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat 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.\n\nThe 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).\n\nThe 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.\n\nJust 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.\n\nIf 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.\n\nI 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.","headline":"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.","tokens_in":18175,"tokens_out":5157,"would_cite":false,"duration_ms":44659,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M30","62G15","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Localized Spatial Conformal Prediction (LSCP) claims finite-sample conditional coverage under stationarity and spatial mixing instead of exchangeability.","keywords":["spatial conformal prediction","quantile regression","conditional coverage","spatial mixing","stationarity","prediction intervals","uncertainty quantification","non-exchangeable data"],"falsifier":"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.","tokens_in":17164,"feed_emoji":"📍","tokens_out":5880,"duration_ms":47507,"temperature":0.7,"pith_summary":"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.","feed_headline":"Spatial conformal prediction gets conditional coverage via neighbor weights","feed_subtitle":"Uses stationarity and spatial mixing instead of i.i.d. assumptions to bound the coverage gap.","key_machinery":"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$(a,b)\\) for the stationary error field, a weight-decay condition (Assumption 4.1), and a bound on the estimation error of the base model (Assumption 4.2). Lemma 4.5 bounds the distance between this weighted CDF and the true error CDF, and Theorem 4.6 converts that into a conditional coverage gap.","core_discovery":"On the paper's own terms, the central discovery is that casting conformal prediction as a localized quantile regression problem allows conditional coverage guarantees for spatial data under dependence. Specifically, Theorem 4.6 states that for any \\(\\$\\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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Introduces GSCP and SLSCP, the spatial conformal baselines that LSCP is compared against and whose infill-based assumptions motivate the new theory.","marker":"Mao et al., 2024"},{"why":"Provides the localized conformal prediction framework that LSCP builds on and contrasts with.","marker":"Guan, 2023"},{"why":"Supplies the weighted conformal framework and coverage-gap analysis for non-exchangeable data that LSCP extends to the spatial setting.","marker":"Barber et al., 2023"},{"why":"Introduces EnbPI, a time-series conformal baseline, whose O(1/n) weight condition is relaxed by Assumption 4.1.","marker":"Xu & Xie, 2021"},{"why":"Quantile regression forests, the implementation used to learn the localized weights in LSCP.","marker":"Meinshausen & Ridgeway, 2006"},{"why":"Provides Lemma 1.1, the variance bound for sums of mixing variables used in the proof of Lemma A.2.","marker":"Rio et al., 2017"},{"why":"Supplies the strong-mixing coefficient definition for spatial random fields used in Assumption 4.3.","marker":"Lahiri, 2003"}],"fun_headline_variants":["Localized quantile regression gives spatial conformal conditional coverage","Spatial conformal prediction without i.i.d. via localized quantiles","LSCP: spatial conformal intervals with conditional coverage guarantees","Mixing beats i.i.d. for spatial conformal prediction intervals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Localized quantile regression gives spatial conformal conditional coverage","Spatial conformal prediction without i.i.d. via localized quantiles","LSCP: spatial conformal intervals with conditional coverage guarantees","Mixing beats i.i.d. for spatial conformal prediction intervals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1297,"prompt_tokens":949,"completion_tokens":348,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":565,"tokens_out":348,"duration_ms":3562,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:42:07.028973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces GSCP and SLSCP, the spatial conformal baselines that LSCP is compared against and whose infill-based assumptions motivate the new theory."},{"cited_title":"F., Candes, E","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted conformal framework and coverage-gap analysis for non-exchangeable data that LSCP extends to the spatial setting."},{"cited_title":"and Ridgeway, G","cited_arxiv_id":null,"evidence_quote":"Quantile regression forests, the implementation used to learn the localized weights in LSCP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Lemma 1.1, the variance bound for sums of mixing variables used in the proof of Lemma A.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the strong-mixing coefficient definition for spatial random fields used in Assumption 4.3."}],"review_version":1}