{"id":"df763e9e-e07a-478d-806d-d93585878346","arxiv_id":"2607.04431","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Pairwise quantile regression via U-statistic pinball-loss minimization achieves nearly 1/n excess-risk rates under a local density condition, and is applied to facial-recognition similarity-score analysis.","lead":"This paper defines quantile regression for pairwise similarity scores—predicting, say, the 5th percentile of how similar two face photos look, given attributes of both photos. It proves fast statistical learning rates for this estimator and applies it to facial-recognition error analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's variance–excess control is derived via a false global quadratic lower bound; the local density assumption only supports it for small |q−Q|, so the fast-rate theorem requires a repaired constant.","rationale":"The reader's weakest_assumption is Assumption 3, and the reader's rationale explicitly mentions 'an unflagged local-to-global derivation gap' in Proposition 1. My stress-test agrees with that specific gap and identifies it as the most load-bearing concern: the proof of Proposition 1 uses a global quadratic lower bound that Assumption 3 does not supply. However, I do not think this invalidates the central rate claim: the variance–excess relation can be repaired using the quantile mass lower bound plus boundedness of q and s, yielding a modified constant and the same log(n)/n rate. Therefore the appropriate verdict remains CONDITIONAL, which is what the reader already recommended; my read does not move the verdict. The reason for 'partial' rather than 'agree' is that the reader's stated weakest assumption emphasizes the applicability of Assumption 3 to real biometric scores, while the more immediate mathematical issue is that Proposition 1's stated constant is not justified even when Assumption 3 holds. Both point to the same fragile link, and both are addressable without rejecting the paper's main contribution.","tokens_in":22300,"tokens_out":35886,"duration_ms":381723,"concrete_test":"Re-derive the lower-bound step in Appendix A.1 for a one-point covariate distribution with ν=50, δ=0.01, Q=0.5, score density 50 on [0.49,0.51] and 0 elsewhere, and q≡1. Compute E(q)=∫_0^{0.5}(F(0.5+t)−F(0.5))dt = 0.2475 and compare to (ν/2)||q−Q||² = 6.25; the quadratic lower bound fails by a factor of 25. Then test whether Proposition 1's conclusion itself fails by using a two-subpopulation model: Z∈{0,1} equiprobable, both conditional score densities have median 0.5 and satisfy Assumption 3 near 0.5, but differ in their tails; compute Var(k_q(V)) numerically and compare to (2Mτ²/ν)E(q). If the bound is violated, Proposition 1 is false; if it holds, the concern is confined to the proof's unjustified constant and Theorem 1 needs a modified C_var.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central fast-rate claim (Theorem 1) rests on Proposition 1's variance–excess inequality Var(k_q(V)) ≤ C_var E(q). The proof of Proposition 1 in Appendix A.1 uses the Knight identity and then claims that, under Assumption 3, E(q) ≥ (ν/2)||q−Q||²_{L2}. This step is only valid when |Δq| := |q(Z,Z')−Qs(τ|Z,Z')| ≤ δ. Assumption 3 bounds the conditional density only in a δ-neighborhood of the conditional quantile, not globally. For |Δ| > δ, the integral ∫_0^Δ (F(Qs+t)−F(Qs)) dt grows at best linearly in Δ (it contains the constant positive mass νδ accumulated by t=δ), so the quadratic lower bound fails. A concrete example: take τ=1/2, Q=0.5, and a score density equal to 50 on [0.49,0.51], zero elsewhere; for q≡1, E(q)=0.2475 while (ν/2)||q−Q||² = 6.25. Thus the proof of Proposition 1 is wrong as written. The conclusion can likely be repaired: because the conditional τ-quantile forces a mass τ below Q (or 1−τ above), k_q(v) ≥ τ(1−τ)|Δ|, and boundedness of q and s gives E(q) ≥ c E|Δ|, hence a variance–excess bound with a constant depending on max(B,||s||∞)/(τ(1−τ)). But that is not the stated C_var = 2Mτ²/ν, and Theorem 1's proof as presented does not go through without this repair. This is a correctness gap in the central theoretical claim, although the claimed log(n)/n rate is likely salvageable with modified constants.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pairwise extension of quantile regression, where the response is a symmetric similarity score s(X,X') and the covariates are pairs (Z,Z'). The empirical risk is a degree-2 U-statistic built from the pairwise pinball loss. The main theoretical claim is that, under a VC-type complexity condition on the function class and a local lower bound on the conditional density of the score near its conditional τ-quantile (Assumption 3), the excess risk of the empirical minimizer converges at the fast rate O(L log n / n) and the L2 error at O(sqrt(log n / n)). The authors also present synthetic experiments and a facial-recognition application with SHAP-based interpretability.","tokens_in":22746,"tokens_out":9016,"duration_ms":92339,"significance":"If established, the fast-rate result is a valuable addition to pairwise learning theory: it shows that the U-statistic structure yields a variance–excess-risk control with θ=1, leading to near-parametric rates under a mild local-density condition. The application to biometric similarity scores is timely and the empirical study, including coverage checks and model comparisons, is a useful contribution. The paper also provides code, which strengthens reproducibility. However, the proof of the main variance–excess proposition is not correct as written, and the proof of the main theorem contains an algebraic error in the Hoeffding decomposition; these issues affect the central theoretical claim and need to be repaired.","major_comments":[{"comment":"The proof claims the global quadratic lower bound E(q) ≥ (ν/2)||q−Qs||²_{L2}. Assumption 3 only gives density control on the δ-neighborhood of the conditional quantile. For |Δq| > δ, the Knight-identity integral ∫_0^Δ (F(Qs+t)−F(Qs))dt grows at most linearly, not quadratically. Concretely, take τ=1/2, Qs=0.5, and conditional density 50 on [0.49,0.51] and 0 elsewhere; for q≡1, E(q)=0.2475 while (ν/2)||q−Qs||² = 6.25. Thus Proposition 1 as stated is false. The fast rate can likely be salvaged by replacing the quadratic lower bound with a combination of the local quadratic bound and a linear bound for large deviations, yielding C_var depending on (B+||s||∞)/δ, but this repair is not in the manuscript and the stated constant 2Mτ²/ν is not correct.","section":"Appendix A.1, Eq. (18)"},{"comment":"The canonical (degenerate) kernel is defined as k̂_q(V,V') = K_q(V,V') − k_q(V) − k_q(V') − E(q). For a symmetric kernel K_q, the correct Hoeffding projection is K_q − k_q(V) − k_q(V') + E(q), where E(q)=E[K_q]. With the minus sign, k̂_q is not centered and the decomposition bE_n(q) − E(q) = 2T_n(q) + W_n(q) is algebraically wrong; the right-hand side differs by 2E(q). This is a load-bearing error because the subsequent bound E(q̂_τ) ≤ 2(sup|T_n| + sup|W_n|) follows from that decomposition. The argument can be fixed by changing the sign to +E(q), but as written the proof of Theorem 1 is incorrect.","section":"Appendix A.2, Eq. (20)"},{"comment":"Both results rely on Proposition 1 for the variance–excess control and on Eq. (18) for the L2-to-excess bound. Since Eq. (18) is not valid globally, the stated constants and the claimed range of applicability are not justified. The rate O(log n/n) is likely preserved after the repair, but the constants must be modified to include a factor depending on the ratio (B+||s||∞)/δ (or a similar term). The manuscript should state the repaired proposition and track the new constants through the proof.","section":"Theorem 1 and Corollary 1"}],"minor_comments":[{"comment":"For real-valued function classes, 'VC-dimension' should be defined precisely (e.g., VC subgraph dimension) or the relevant entropy integral should be stated. The current wording is slightly ambiguous.","section":"Assumption 2"},{"comment":"The phrase 'by a first-order Taylor expansion of the conditional CDF' is not accurate; the identity is an exact integral representation via the Knight identity. Rewording would improve clarity.","section":"Appendix A.1"},{"comment":"In Proposition 2 the bound contains (B+1)log(n/δ)/n; the appearance of '1' is unexplained. Presumably it is ||s||∞, but this should be stated explicitly.","section":"Appendix B.2"},{"comment":"The D2 values in Table 1 are useful, but the text does not report error bars or a measure of variability across data splits. Adding confidence intervals would strengthen the empirical claims.","section":"Section 4.2"},{"comment":"Several references are cited with page numbers or theorem numbers that are not checked against the actual sources (e.g., Theorem 8.3 in Massart 2007). The authors should verify these citations carefully.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a worthwhile problem and the empirical study is solid, but the central theoretical proof has two concrete errors: the global quadratic lower bound in Proposition 1 and the sign in the Hoeffding decomposition in Theorem 1. Both are fixable and likely do not change the qualitative rate, but they are load-bearing and require a careful revision rather than a rubber stamp. I would not reject the paper, but the authors must correct the proofs and state the repaired constants."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely new: it formulates pairwise quantile regression under the pinball loss as U-statistic ERM and proves fast learning rates under a local density condition. The ingredients are known—pinball loss, U-process concentration, Hoeffding decompositions—but the combination is new, and the observation that the variance–excess control holds automatically in the pairwise setting (θ=1) is a nice contrast with pointwise quantile regression. The synthetic experiments and the facial-recognition application, with Shapley-based interpretation of extreme scores, are a real plus. This is a solid within-subfield contribution, not a paradigm shift.\n\nThe main soft spot is Proposition 1. The proof in A.1 derives E(q) ≥ (ν/2)||q−Q||² by applying Assumption 3's local density lower bound uniformly over all deviations. That is not justified: the density is only bounded below in a δ-neighborhood of the conditional quantile, so for functions q far from Q the excess risk can grow linearly while the L2 distance is large. The stress-test example is correct: with density 50 on [0.49,0.51], q≡1 gives E(q)≈0.2475 but (ν/2)||q−Q||²=6.25. So the proof, as written, is wrong. The fix is plausible—boundedness of q and s gives E(q) ≥ c E|Δ|, and boundedness gives E|Δ|² ≤ C E|Δ|, so a variance–excess bound survives with a modified constant that still yields log(n)/n rates. But the constant C_var=2M²/ν is not established, and Theorem 1's proof does not go through without this repair. This is a correctness gap in the central theoretical claim, though likely fixable.\n\nThe experimental section is otherwise decent but lacks error bars, and the FR dataset is not yet released. The assumption that conditional score densities are uniformly lower bounded near the quantile is plausible for FR but strong; the paper argues for it via smoothness, which is reasonable but not a proof.\n\nThis paper deserves a serious referee. The main idea is sound, the gap is localized and repairable, and the application is relevant. I would send it to review with a request to fix Proposition 1 and to add some measure of variability to the experiments. I'd also bring it to a reading group—the fast-rate mechanism and the repair are good discussion material.","headline":"A natural pairwise extension of quantile regression with a plausible fast-rate theorem, but the proof of Proposition 1 has a local-to-global gap that needs repair before the main result is watertight.","tokens_in":23226,"tokens_out":3463,"would_cite":true,"duration_ms":37013,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08","62G20","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Pairwise quantile regression achieves fast log(n)/n learning rates","keywords":["pairwise quantile regression","pinball loss","U-statistics","fast learning rates","similarity scoring","facial recognition","VC dimension","conditional quantile"],"falsifier":"Construct a family of examples satisfying Assumptions 1-2 and realizability, with the conditional score distribution supported on only two points (so no positive density near the quantile), run the pairwise pinball-loss ERM on growing n, and measure excess risk and L2 error. If the observed rates are n^{-1/2} rather than log(n)/n, Assumption 3 is necessary for the fast rate.","tokens_in":22211,"feed_emoji":"🎯","tokens_out":4918,"duration_ms":49074,"temperature":0.7,"pith_summary":"This paper extends quantile regression to a pairwise setting: the response variable is a similarity score between two independent observations, and the explanatory variables are the pair of their covariates. The central claim is that empirically minimizing a pairwise version of the pinball loss over a bounded VC class yields fast learning rates—excess risk of order log(n)/n and L2 deviation from the conditional quantile of order sqrt(log(n)/n)—under a mild condition that the conditional density of the score stays bounded away from zero near the target quantile. The improved rate comes from a variance-excess risk control that holds automatically for U-statistics, without the restrictive noise conditions needed in pointwise quantile regression. The paper demonstrates the method on synthetic data and on facial recognition similarity scores, where it identifies interpretable drivers of extreme scores.","feed_headline":"Pairwise quantile regression achieves log(n)/n rates","feed_subtitle":"U-statistic pinball loss finds similarity-score quantiles at near-parametric speed, shown on face recognition.","key_machinery":"The central object is the U-statistic version of the empirical pinball loss, bR_n(q) = 2/(n(n-1)) sum_{i0 in a neighborhood of the quantile. Consequently, the L2 distance to the conditional quantile is O(sqrt(log(n)/n)). The key mechanism is Proposition 1: the variance of the first-order Hoeffding projection is bounded by a constant","pith_inferences":["The uniform density lower bound (Assumption 3) is plausible for smooth, continuous scores but may fail in real biometric data where conditioned covariates (like same-age, same-quality pairs) can make the score distribution nearly degenerate; if such degeneracy is present, the fast-rate guarantee is not in force. A practical check would be to estimate the conditional density near the quantile on su","Because the fast rate follows from the variance-excess risk relation for U-statistics, the same proof template should transfer to other pairwise 'margin-like' losses, such as expectile regression or pairwise logistic-type losses, whenever a similar Lipschitz and calibration inequality holds.","The paper notes but does not implement score recalibration via quantile alignment; one immediate extension is to use q_hat_tau to adjust raw similarity scores so that impostor/genuine distributions are standardized, then re-evaluate FAR/FRR on the adjusted scores.","The Shapley analysis shows feature importances vary across quantile levels (e.g., quality dominates low genuine quantiles, hair length affects high impostor quantiles); this suggests a stratified error analysis could be performed per quantile to guide targeted data augmentation."],"forward_implications":["The estimator achieves excess risk of order log(n)/n, faster than the generic 1/sqrt(n) rate, even though pairs are dependent.","The variance-excess risk control holds with exponent theta=1 automatically, so pairwise quantile regression gets fast rates under the same density condition that guarantees uniqueness of the risk minimizer.","Incomplete U-statistics with only O(n log n) pairs preserve near-optimal performance, making the method scalable to large datasets.","The approach directly estimates extreme conditional quantiles of similarity scores, which correspond to false rejections and false acceptances in biometric systems, enabling targeted analysis of rare errors.","A penalized model-selection scheme yields an oracle inequality over a countable collection of function classes."],"fun_headline_variants":["Pairwise quantile regression hits near-parametric rates","U-statistic pinball loss: fast rates for paired data","Quantile regression on pairs: log(n)/n excess risk bound","Pairwise learning: sharp quantile guarantees for similarity scores","Face-recognition scoring: pairwise quantile regression with proven rates"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire fast-rate result depends on Assumption 3: the conditional density of the similarity score given the covariate pair must be bounded below by a positive constant on a fixed interval around the target quantile, for every covariate pair; if that density is flat or degenerate, the variance-excess risk control and the log(n)/n rate disappear.","fun_headline_variants_meta":{"raw":{"variants":["Pairwise quantile regression hits near-parametric rates","U-statistic pinball loss: fast rates for paired data","Quantile regression on pairs: log(n)/n excess risk bound","Pairwise learning: sharp quantile guarantees for similarity scores","Face-recognition scoring: pairwise quantile regression with proven rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1267,"prompt_tokens":763,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":420}},"tokens_in":507,"tokens_out":504,"duration_ms":5791,"temperature":1.0,"reasoning_tokens":420,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T08:36:54.085436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a family of examples satisfying Assumptions 1-2 and realizability, with the conditional score distribution supported on only two points (so no positive density near the quantile), run the pairwise pinball-loss ERM on growing n, and measure excess risk and L2 error. If the observed rates are n^{-1/2} rather than log(n)/n, Assumption 3 is necessary for the fast rate.","supporting_citations":[],"review_version":2}