{"id":"b07d5399-5c82-4647-abc9-2a3ac37f007e","arxiv_id":"2607.20821","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Ball-codifference is a new marginal screening utility for heavy-tailed predictors; its asymptotic and screening properties are stated conditionally and the supporting data example contains an internal contradiction.","lead":"This paper introduces Ball-codifference, a new statistic for screening predictors in high-dimensional data that may have heavy tails, where correlation and covariance are often undefined. It claims the method is moment-free and provides a sure-screening guarantee, but the paper's numerical evidence is mixed and the data example contradicts the text.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3's uniform-error condition is never proved for the BCodif utility; empty random-ball denominators make the cited exponential inequalities inapplicable. The sure-screening claim rests on an unverified premise, and Table 3's sign-reversed 'Improvement' contradicts the abstract's prediction","rationale":"The reader's weakest-assumption identification is exactly the point I find most load-bearing: Theorem 4.3's sure-screening conclusion depends on a uniform concentration condition that the paper does not establish for the proposed statistic. The statistic is not a simple bounded V-statistic because of the local random denominators; the proof in Appendix D handles denominators only under a lower-bound event that is not guaranteed by any stated assumption and is unlikely to hold uniformly in high-dimensional, sparse, or heavy-tailed designs. Without this concentration, the theorem is a conditional statement about an unverified premise rather than a proved property of BCodifCor-SIS. I also independently checked the data-example claim and found the same internal inconsistency the reader noted: Table 3's 'Improvement' column is the percentage increase in RMSE, so positive values indicate the proposed method performs worse. This is not a matter of taste or consensus; it is an internally inconsistent use of the paper's own numbers. I therefore support the REJECT verdict. I grant that the construction is genuinely moment-free and the simulation tables do show some gains in average top-d retention, but those gains do not offset the missing theory and the false prediction-accuracy claim.","tokens_in":13879,"tokens_out":6094,"duration_ms":67749,"concrete_test":"Derive the uniform concentration bound in Theorem 4.3 for \\widehat{BCodif}_s^2. Specifically, check whether the event min_{1≤i,j≤n} Δ_{ij}^{XY} ≥ π_0/2 has probability tending to one under the n=150, p=1000 Toeplitz designs in Section 5. If not, the inverse-denominator terms in (24) prevent the exponential-inequality argument and the theorem's premise is unverified. As a supplemental check, recompute Table 3 with Improvement defined as (RMSE_BCorr − RMSE_BCodif)/RMSE_BCorr × 100; if most entries are negative, the abstract's prediction-accuracy claim is refuted.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing gap is in Theorem 4.3. Its premise max_{1≤r≤p}|ω̂_r − ω_r| ≤ ε_n = o(c_n) with probability tending to one is not derived for \\widehat{BCodif}_s^2 or \\widehat{BCodifCor}^2. The paper asserts this follows from exponential inequalities, but no proof is given. The statistic (13) contains inverse powers of the empirical joint-ball count Δ_{ij}^{XY}; when a product ball is empty, Δ_{ij}^{XY}=0 and κ̂_{ij}=0, so the kernel is not a bounded V-statistic in the usual sense. Appendix D only controls inverses on the event min_{i,j} Δ_{ij}^{XY} ≥ π_0/2, but no lower bound on local ball probabilities π_{ij} appears in Assumption 4.1 or Theorem 4.3. Under trimming, the bias from replacing Δ by Δ∨a_n is not shown to be o_p(n^{-1/2}) when π_{ij} is of order smaller than a_n. Hence the uniform error condition — the load-bearing premise of P(A ⊆ Â_n) → 1 — is unverified. In addition, the paper's own Table 3 reports Diffmean = RMSE_BCodif − RMSE_BCorr; a positive value means BCodif has higher RMSE, i.e., worse performance. Yet the text labels these as 'Improvement' and the abstract claims significant improvement, which is contradicted by 26 of the 28 rows.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Ball-codifference, a marginal screening utility that combines random-ball geometry with a characteristic-function-based codifference weight, aiming to screen predictors under heavy-tailed distributions where covariance or correlation may not exist. The main theoretical results are an asymptotic normality claim for the empirical Ball-codifference statistic (Theorem 4.2) and a sure-screening property for a proposed BCodifCor-SIS procedure (Theorem 4.3). The numerical section compares BCodifCor-SIS with BCor-SIS on simulated Gaussian and sub-Gaussian stable designs, and a real-data example (Riboflavin) is used to claim improved prediction accuracy. The paper also provides R code for reproducibility.","tokens_in":14223,"tokens_out":4247,"duration_ms":44148,"significance":"If established, a moment-free screening statistic that is robust to heavy tails and still detects dependence would be a useful addition to the high-dimensional screening toolbox. The paper has a plausible heuristic and the supplied code is a positive feature. However, the central theoretical guarantees are conditional on unverified assumptions, and the main empirical claim in the data example is contradicted by the paper's own table. As it stands, the manuscript does not make a convincing case for either the sure-screening property or the practical superiority of the proposed method.","major_comments":[{"comment":"The sure-screening conclusion is derived only from the assumed uniform error condition max_{1≤r≤p}|ω̂_r − ω_r| ≤ ε_n with probability tending to one and the population separation condition. Neither condition is proved for the Ball-codifference utility. The text after the theorem states that the uniform error condition 'can be obtained from exponential inequalities for bounded empirical processes', but no such derivation is supplied. The statistic (13) contains inverse powers of the empirical joint-ball count Δ_{ij}^{XY}, so it is not a bounded kernel in the usual sense; the cited exponential inequalities are therefore not directly applicable. Appendix D only controls the inverse denominators on the event min_{i,j} Δ_{ij}^{XY} ≥ π_0/2, but no uniform lower bound on the local ball probabilities π_ij (defined in (26)) appears in Assumption 4.1 or in Theorem 4.3. Thus the load-bearing premis","section":"Section 4, Theorem 4.3 and Appendix C"},{"comment":"The claim that the statistic in (13) is a bounded V-statistic is not correct as stated. As shown in equations (24)–(25), the statistic has terms of the form (Δ_{ij}^{XY})^{-1} and (Δ_{ij}^{XY})^{-2} unless one conditions on all local denominators being bounded away from zero. Appendix D invokes a lower bound π_ij ≥ π_0 > 0, but Assumption 4.1(iii) only allows a trimming sequence a_n ↓ 0 with n a_n → ∞. The bias introduced by trimming Δ_{ij}^{XY} to Δ_{ij}^{XY} ∨ a_n is not analyzed; when π_ij is of order smaller than a_n, the trimming error need not be o_p(n^{-1/2}). Hence the asymptotic normality result is conditional on unstated additional assumptions, and the proof does not establish the theorem as stated.","section":"Appendix D / Theorem 4.2"},{"comment":"The column 'Improvement (%)' is computed as (RMSE_BCodif − RMSE_BCorr)/RMSE_BCorr × 100; for example, ncov=2 gives 0.0192/0.7322 ≈ 2.6%. Thus a positive 'Improvement' means BCodif has a higher RMSE, i.e., worse prediction accuracy. The table has positive values in 26 of 28 rows, meaning BCodifCor-SIS is worse than BCor-SIS in almost all configurations. The text, however, claims 'the BCodifCor-SIS screening method dominates BCor-SIS significantly, except for the two instances (ncov=6,7) of small better performance for BCor-SIS.' In fact, for ncov=6 and 7 the Diffmean is negative, indicating BCodif is better. The abstract's claim that 'our variable screening method significantly improves prediction accuracy' is therefore directly contradicted by the paper's own reported numbers.","section":"Section 6, Table 3"},{"comment":"The statement that 'Ball-codifference dominates Ball-Covariance' is an overstatement based only on the pa (average retention) columns. The exact-rank recovery probabilities pm often favor BCor-SIS; examples include Table 1 (α=0.9, ρ=0.95): pm(X15) is 0.22 for BCodifCor-SIS vs 0.40 for BCor-SIS; Table 2 (ρ=0.95): pm(X10) 0.68 vs 0.80, pm(X5) 0.90 vs 0.94. The text acknowledges that exact-rank recovery is 'more variable', but the broad claim of dominance is not supported. The selective emphasis on pa while de-emphasizing pm without a principled basis weakens the simulation evidence.","section":"Section 5, Tables 1 and 2"}],"minor_comments":[{"comment":"There are several typographical errors: 'Ball-codifierence' (Section 1), 'MEPE' instead of 'MSPE' (Section 6), and 'Section 4' in Section 3.1 where the numerical experiments actually appear in Section 5. The phrase 'The second form is used in the numerical experiments in Section 4' should refer to the correct section.","section":"Throughout"},{"comment":"The simulation setup states 'We generated (Y, X_1, ..., X_{p-1}) with p = 1000' and then says 'The response is the first component, Y = Z_1'. This is a notational inconsistency; the number of predictors and the index of the response should be clarified.","section":"Section 5, first paragraph"},{"comment":"The sentence 'The second condition can be obtained from exponential inequalities for bounded empirical processes when log p = o(n c_n^2)' is vague. Even if a proof were supplied, the phrase 'the exact rate depends on the metric entropy' leaves the condition unactionable. At minimum, a reference or a formal statement of the required entropy condition should be given.","section":"Section 4, after Theorem 4.3"}],"recommendation":"reject","confidential_remarks":"The paper has a plausible idea but the central theoretical result is not proved and the empirical claim is contradicted by the paper's own table. The sure-screening theorem is a conditional restatement of the standard SIS argument, and the uniform concentration of the Ball-codifference statistic is not established. The data example's 'Improvement' column is sign-reversed relative to the text, which is not a mere typo but reverses the paper's main practical conclusion. These are load-bearing issues that cannot be fixed by local edits. I would not encourage resubmission unless the authors can prove the uniform error condition under explicit lower-bound assumptions and correct the data analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper introduces a genuinely new screening statistic—Ball-codifference—that combines Pan et al.'s Ball covariance with the extended codifference from your group. That combination is not present in the earlier literature, and the empirical score (13) is a reasonable way to get a moment-free marginal utility. The simulations show consistent gains in average top-d retention for the Toeplitz designs, and they ship R code. So there is something here.\n\nBut two soft spots are load-bearing. First, Theorem 4.3 is a restatement of the standard SIS argument: if uniform empirical error and population separation hold, the conclusion follows. The paper never proves those conditions for this statistic. The claim that they follow from exponential inequalities is not backed up; the statistic contains inverse powers of empirical ball counts, and the appendix's control only works on an event where the local ball probability is bounded away from zero, which is not part of the assumptions. So the sure-screening property is not established.\n\nSecond, the data example contradicts the abstract. Table 3 reports Diffmean = RMSE_BCodif − RMSE_BCorr, so positive values mean BCodif is worse. The table shows positive Diffmean in 26 of 28 rows, yet the text labels those as 'Improvement' and the abstract claims significant improvement. The two rows with negative Diffmean (ncov=6,7) are actually the ones where BCodif wins, and the text misreads them as losses. That is not a typo in one cell; it flips the conclusion.\n\nThe independence characterization in Lemma 2.1 is correct, and the V-statistic CLT is standard. The asymptotic argument is fine as far as it goes, but it doesn't cover the implemented statistic in the regime that matters.\n\nWho is this for? Someone working on heavy-tailed screening might find the statistic useful as an alternative to BCor-SIS. But the paper in its current form does not support the stated contributions. I would not cite it yet, and I wouldn't present it as evidence of improved prediction. It deserves a serious referee only if the authors are willing to either prove the uniform error condition under explicit lower-bound assumptions on local ball probabilities or weaken the claim accordingly, and to correct the data analysis. As is, I'd recommend rejection with an invitation to resubmit after those fixes.","headline":"New statistic worth a look, but the sure-screening proof is conditional and the data example is misreported in the paper's own tables.","tokens_in":14712,"tokens_out":2661,"would_cite":false,"duration_ms":28543,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62G10","62H12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Ball-codifference, a moment-free dependency measure, is proposed as a sure-screening statistic for heavy-tailed ultra-high-dimensional predictors.","keywords":["Ball-codifference","sure independence screening","heavy-tailed data","stable distributions","codifference","V-statistics","high-dimensional screening","dependence measures"],"falsifier":"Simulate p = 1000, n = 150 with independent symmetric α-stable predictors (α = 0.8) and no response association, then track max_r |ω̂_r − ω_r| across replications. If this maximum does not converge to zero at the theorem's required rate — for instance, if heavy-tailed cosine contrasts keep the error large — the uniform-error premise fails and the sure-screening guarantee would not hold.","tokens_in":13753,"feed_emoji":"📊","tokens_out":6152,"duration_ms":64584,"temperature":0.7,"pith_summary":"The paper tries to establish that a new marginal association statistic, Ball-codifference, is a valid and useful screening tool for ultra-high-dimensional regression when predictors and response may be heavy-tailed. The statistic combines the geometry of random balls with a codifference weight computed from cosine contrasts, so it is defined even when covariance and correlation are undefined. The paper claims a sure-screening property: with high probability, all truly active predictors are retained among the top scores, provided population utilities are separated and empirical scores converge uniformly. Its simulations and a gene-expression data example are offered as evidence that this moment-free screening outperforms Ball-covariance screening in retaining strongly associated predictors and improving prediction accuracy.","feed_headline":"New statistic screens heavy-tailed predictors without moments","feed_subtitle":"Ranks predictors from random balls and cosine contrasts, so active variables survive even with undefined variance.","key_machinery":"The load-bearing object is the Ball-codifference functional, BCodif_s^2(X,Y) = E[D_{12}^2 κ_{12}], where D_{12} compares joint and marginal probabilities of the random ball A_{12} and κ_{12} is the local codifference inside that ball. The codifference weight is a ratio of characteristic-function-type expectations; like a trigonometric knot, it remains finite under infinite variance. The empirical version rewrites every piece as sums of indicator and cosine terms, making the statistic a bounded V-statistic whose first Hoeffding projection controls its normal limit. This combination is what gives the method its moment-free, model-free character.","core_discovery":"The central claim is that dependence for screening can be measured by the Ball-codifference: the squared difference between the joint probability of a random product ball and the product of its marginal probabilities, multiplied by a local codifference weight formed from conditional cosine expectations. This quantity is a bounded V-statistic, giving asymptotic normality through empirical-process and functional-delta arguments, and it induces a marginal utility that detects association without requiring finite moments. The paper's main theorem states that, under a population separation condition and a uniform empirical-error condition, the BCodifCor-SIS procedure retains every active predicto","pith_inferences":["Editorial extension: the bounded-contrast construction is portable; the same 'squared ball discrepancy times local codifference' weighting could be grafted onto other geometric dependence measures, potentially yielding moment-free versions of distance-correlation screening.","Editorial extension: the sure-screening theorem is conditional on a uniform-error bound that is asserted rather than proved; a central open step is a concentration inequality for max_r |ω̂_r − ω_r| that would make the screening guarantee unconditional.","Editorial extension: since the statistic only uses metrics and cosines, it should extend to functional or non-Euclidean predictors; a testable next step is comparing Ball-codifference screening on curve or spatial data against competitor screening utilities."],"forward_implications":["If the claims hold, practitioners can screen ultra-high-dimensional predictors without first verifying finite moments, extending valid inference to α-stable and other heavy-tailed data.","The sure-screening theorem means the top-d selection rule will, with probability tending to one, include every active predictor once the uniform-error and separation conditions are met.","The bounded V-statistic representation yields a normal limit, so standard errors and approximate inference for the screening scores are available under the stated assumptions.","In the paper's simulations, the codifference-weighted score has higher average retention of strongly associated predictors than Ball-covariance screening across Gaussian and heavy-tailed designs.","On the riboflavin benchmark, pre-screening with Ball-codifference reduces linear-regression prediction error relative to Ball-covariance pre-screening for most pre-selected model sizes, with the largest reported improvements above 30%."],"fun_headline_variants":["No moments needed: Ball-codifference screens heavy tails","Heavy-tailed screening without moment assumptions","Ball-codifference: screening that works even when variance is undefined","Detect active predictors with Ball-codifference under heavy tails","Screen heavy-tailed data using random balls and cosine contrasts"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The theorem's guarantee depends on every estimated Ball-codifference score being uniformly close to its population value (error o(c_n) with probability tending to one) while active and inactive population scores are separated by at least 2 c_n; the paper asserts this uniform-error condition can be obtained from exponential inequalities but does not prove it for the new statistic.","fun_headline_variants_meta":{"raw":{"variants":["No moments needed: Ball-codifference screens heavy tails","Heavy-tailed screening without moment assumptions","Ball-codifference: screening that works even when variance is undefined","Detect active predictors with Ball-codifference under heavy tails","Screen heavy-tailed data using random balls and cosine contrasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000966,"raw_usage":{"total_tokens":3917,"prompt_tokens":683,"completion_tokens":3234,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":3155}},"tokens_in":427,"tokens_out":3234,"duration_ms":23501,"temperature":1.0,"reasoning_tokens":3155,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:15:55.886757+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate p = 1000, n = 150 with independent symmetric α-stable predictors (α = 0.8) and no response association, then track max_r |ω̂_r − ω_r| across replications. If this maximum does not converge to zero at the theorem's required rate — for instance, if heavy-tailed cosine contrasts keep the error large — the uniform-error premise fails and the sure-screening guarantee would not hold.","supporting_citations":[],"review_version":1}