{"id":"b124ecf7-2da7-419b-add6-27681f679c5e","arxiv_id":"2509.25521","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A numerical implementation of the Rosenblatt conditional-CDF transform renders a classifier quasi-independent from an observable, with particle physics examples showing lower distance correlation than DisCo at comparable AUC.","lead":"The authors present two numerical recipes, irregular-bin interpolation and kernel density estimation, that implement the classic Rosenblatt conditional-CDF transform to make a classifier output statistically independent of another observable. The method is demonstrated on toy data, images, and an LHC-style dijet search, where it decorrelates the classifier while preserving more classification power than the DisCo penalty approach.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The independence guarantee is conditional on the defining background sample matching the true background density; the paper does not test robustness to mismodeling, so the practical claim may fail outside the same-density setting.","rationale":"The reader's weakest assumption — representativeness of the defining background sample — is indeed the most load-bearing condition for the central claim. The mathematical core is correct and the numerical implementation is tested on samples drawn from the same density that defines the transform, but this does not probe the failure mode that matters in real HEP analyses, where simulation may not match data. The proposed reweighting test would directly quantify sensitivity to this mismatch. Secondary issues such as absence of error bars and potential selection of parameters on the validation sample are real but do not change the verdict; they also support keeping the paper CONDITIONAL rather than ACCEPT or REJECT.","tokens_in":14989,"tokens_out":14377,"duration_ms":132965,"concrete_test":"Reproduce the LHC Olympics example, but construct the IRGI transform from a background sample reweighted by a smooth function of x, e.g., multiply event weights by exp(α(m−⟨m⟩)) with α chosen to shift the dijet mass spectrum by about one standard deviation. Then evaluate DCC(x,γ) on the original, unweighted 100k background test sample. If the DCC rises by more than an order of magnitude above the reported 0.00025, the method is not robust to defining-sample mismodeling and the practical claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The transform in Eq. (4) is the exact conditional CDF of y given x under the defining background density ρ. The numerical implementations (IRGI and KDE) estimate this CDF from a finite defining sample, and the reported DCC of 0.00025 is obtained when the defining sample and the evaluation sample are drawn from the same underlying density by construction (Section III.C). The paper explicitly restricts the guarantee to 'any sample from the same underlying background probability density' (Section II). In the LHC Olympics use case, the defining sample is Monte Carlo; if the real background density in the signal region differs — through miscalibrated mass spectrum, wrong background composition, or mismodeled track features — the estimated conditional CDF is wrong and γ will not be independent of x. The absence of any robustness test against a perturbed defining sample or an independent generator leaves the central practical claim, that the method renders a classifier quasi-independent of an observable in an actual analysis, unverified in the regime where it matters. This is separate from the mathematical correctness of the Rosenblatt transform, which is not in question.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a numerical method for transforming a random variable y into a new variable γ defined as the conditional cumulative distribution function of y given an observable x under a background probability density. The transformation (Eq. 4) is a Rosenblatt transform; the authors prove in Appendix A that γ is independent of x and marginally uniform on [0,1]. Two numerical estimators are presented: IRGI, based on irregular adaptive binning with linear interpolation, and KDE, based on Gaussian kernel smoothing. The method is intended as a plug-in decorrelation layer for classifiers in particle physics, replacing training-time constraints such as DisCo. The authors demonstrate the method on a Gaussian toy example, a cat/dog image classification task with a color observable, and an LHC Olympics dijet dataset, reporting lower distance correlation at comparable or higher AUC than DisCo in the LHC example.","tokens_in":15255,"tokens_out":8213,"duration_ms":69746,"significance":"The mathematical foundation is sound: the independence result follows from the classical Rosenblatt transform, and the proof in Appendix A is correct under the stated regularity conditions. The numerical implementations are clearly described and the LHC Olympics example is a relevant stress test. If the reported performance holds out-of-sample, the method offers a simple, model-agnostic alternative to DisCo with a rigorous independence guarantee and no adversarial optimization. However, the empirical validation is currently weakened by in-sample parameter selection (the DCC minima are obtained on the same data used to select d and σ_r) and by the absence of statistical uncertainties on DCC and AUC. In addition, the paper does not test robustness to mismodeling of the defining background sample, which is the key assumption for practical use. These issues are fixable and do not invalidate the core idea.","major_comments":[{"comment":"The DCC values reported for the IRGI and KDE classifiers (DCC 0.00025 and 0.00018, respectively) are the result of scanning the free parameters d and σ_r on the same 20k-event validation sample that is used to compute these DCC values. The minimum over a parameter grid is an in-sample optimum and is expected to be smaller than the performance on unseen data. The paper does not report DCC on the held-out testing sample (100k background, 10k signal) used for the ABCD demonstration. To support the central claim, the authors should either report DCC on the testing sample or use a nested cross-validation scheme where parameter selection is performed on a training slice and DCC is evaluated on an independent slice.","section":"III.C, Figs. 17-18"},{"comment":"In the Gaussian blob example, for each of the five subsamples the DCC minima for IRGI and KDE are obtained by scanning d or σ_r on that same subsample, so the reported minima are biased estimates of the method's DCC. Additionally, no statistical uncertainties (bootstrap or repeated draws) are reported for DCC or AUC, so it is unclear whether the differences between IRGI, KDE, and the baseline are statistically significant. The authors should provide confidence intervals or per-repetition variations, and ideally a separate validation set for parameter selection.","section":"III.A, Table III-A, Figs. 8-9"},{"comment":"The independence guarantee is conditional on the defining background sample being drawn from the same underlying probability density as the evaluation sample. The paper states this assumption but does not assess its robustness. In the LHC Olympics example the defining sample and the evaluation sample are both from the same Monte Carlo generator, so the demonstration does not cover the realistic scenario in which the simulation used to construct the transform does not perfectly describe the data (e.g., miscalibrated mass spectrum or mismodeled track features). A robustness study with a perturbed defining sample is necessary to support the practical claim that the method renders a classifier quasi-independent in an actual analysis.","section":"II (defining background sample) and III.C"},{"comment":"The IRGI algorithm constructs the transform on a finite rectangular defining domain, with interpolation vertices placed on the domain boundary. The paper does not specify how the transform is evaluated for points outside this domain, nor how it behaves when the evaluation sample extends beyond the support of the defining sample. In the LHC Olympics example, the observable x extends beyond the region shown in the defining sample; an explicit extrapolation rule is needed to define γ for those events.","section":"II.A (IRGI algorithm), Fig. 4"}],"minor_comments":[{"comment":"The acronym 'IRregular' should be 'Irregular' or the capitalization should be explained, as 'IRregular Grid Interpolation' is unusual.","section":"II.A"},{"comment":"In the acknowledgements, 'independetly' should be 'independently'.","section":"V"},{"comment":"The phrase 'bump-hunt is inpired' should be 'inspired'.","section":"III.C"},{"comment":"The abbreviation 'TMV A' appears throughout the text; it should be written consistently as 'TMVA'.","section":"Various"},{"comment":"The phrase 'unique pairs (xi,xj)' is ambiguous; it should specify all pairs with i<j.","section":"II.B, Eq. (8)"},{"comment":"Several subfigures lack complete captions; for example, Fig. 5(b) has no description after the label.","section":"Figures 5, 30, 31"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for physics.data-an. The mathematical core is correct and the numerical implementations are interesting, but the empirical validation needs stronger methodology before acceptance. The key issues are in-sample parameter selection, missing uncertainties, and lack of robustness testing against mismodeling. The authors should also be encouraged to carefully edit the language and figures. There is no concern about the novelty of the Rosenblatt transform itself; the contribution is the numerical method and its comparison to DisCo, which is legitimate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a post-processing implementation of Rosenblatt's 1952 conditional-CDF transform, applied to classifier decorrelation. The authors say that themselves, which I appreciate. What's new is the numerical machinery—an irregular-grid interpolation (IRGI) and a KDE variant—plus a clean demonstration that you can decorrelate a trained classifier from an observable without retraining, so AUC mostly survives. In their LHC Olympics example they get DCC 0.00025 with AUC 0.912, against DisCo at 0.00011 with AUC 0.828. If that's representative, it's a genuinely useful plug-in alternative for ABCD analyses.\n\nThe mathematics is fine. Appendix A proves independence of x and gamma under continuity and nonzero marginal density; the proof is a straightforward Jacobian argument and it holds. The paper also distinguishes itself from TMVA's rarity transform, which only uniformizes marginally. The examples span a toy Gaussian, CIFAR cats vs dogs, and the LHC Olympics; the cat/dog example is a nice stress test of the method's generality.\n\nSoft spots, in proportion. The reported DCC minima are selected by scanning d and sigma_r on the same validation samples used to report the numbers, so the headline DCCs are tuned, not fixed-prediction values. There are no statistical uncertainties on DCC or the ABCD closure errors, which matters when comparing classifiers that differ at the 10^-4 level. Second, the independence guarantee is conditional on the defining background sample being drawn from the same density as the events you later apply the transform to. The paper says this, but never tests robustness to a mismodeled defining sample—e.g., a miscalibrated mass spectrum or wrong background composition. That's not a fatal flaw; it's the same assumption underlying any MC-based decorrelation, including DisCo. But it should be stated as a limitation and ideally tested with a perturbed defining sample. Minor: Figures 5(b), 30(b), and 31(b) are missing from the preprint, and no code is released. Both are fixable.\n\nWho this is for: HEP analysts using ABCD or related data-driven estimates who want a cheap alternative to retraining with a decorrelation penalty. The paper deserves a serious referee—the core method is sound and the empirical comparison is informative even with the tuning caveat. I'd encourage the editor to send it out, with a request for error bars and a robustness check on the defining-sample assumption.","headline":"A post-processing Rosenblatt transform that beats DisCo at preserving AUC in their LHC Olympics test; mathematically sound, empirically promising, but the same-density assumption is untested.","tokens_in":15730,"tokens_out":2174,"would_cite":true,"duration_ms":21133,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A numerical Rosenblatt transform can force a classifier to be statistically independent of a chosen observable, and the transformed classifier keeps its separation power.","keywords":["Rosenblatt transform","forced quasi-independence","distance correlation","ABCD method","classifier decorrelation","kernel density estimation","irregular grid interpolation","LHC Olympics"],"falsifier":"Train the transform on one background sample and apply it to a second sample generated from a different but plausible background density, for example the same process with a shifted calorimeter energy scale; if the measured DCC between $x$ and $\\gamma$ in the second sample stays at the original $(x,y)$ level rather than dropping toward $10^{-4}$, the practical quasi-independence claim fails.","tokens_in":14765,"feed_emoji":"⚛️","tokens_out":9939,"duration_ms":81795,"temperature":0.7,"pith_summary":"This paper establishes that a single numerical integral transform can take any classifier $y$ and a nuisance observable $x$, and produce a new score $\\gamma$ that is quasi-independent of $x$ while remaining almost as discriminative as the original. The transform is the within-$x$ cumulative probability of $y$; when the background density is known, the change of variables makes the joint density factor into the product of the $x$-marginal and a uniform distribution, which is mathematical independence. The paper gives two practical estimators built from a background sample, one based on irregular-grid interpolation and one on kernel density estimation, and demonstrates on public particle-physics data that they reduce the distance correlation from $0.01658$ to about $0.0002$ while keeping the AUC at $0.911$-$0.912$, whereas a DisCo-trained classifier at comparable decorrelation falls to $0.828$. If correct, this removes the main drawback of decorrelation-by-training: analyses no longer trade separating power for the independence that the ABCD background-estimation method needs.","feed_headline":"Rosenblatt transform decouples classifier from observable, keeps AUC","feed_subtitle":"At the LHC Olympics test it hits DCC 0.00025 at AUC 0.912; DisCo pays the same decorrelation with AUC 0.828.","key_machinery":"The load-bearing object is the conditional cumulative distribution transform $\\gamma(x,y)$ of Eq. (4): the $y$-normalized primitive of the joint density $\\rho$ along the $y$-axis, $\\gamma(x,y)=\\int_{-\\infty}^{y} \\rho(x,y')\\,dy'/\\int_{-\\infty}^{\\infty} \\rho(x,y'')\\,dy''$. It is the same object as the classical Rosenblatt transform, re-derived independently here. Its role is to map, for every fixed $x$-slice, the conditional distribution of $y$ onto a uniform distribution; the change-of-variable calculation in Appendix A then makes the joint density factor as $\\rho_X(x)\\cdot 1_{(0,1)}(\\gamma)$, which is exactly independence. The paper's numerical contribution is to estimate this transform from a finite background sample by two routes: IRGI, which builds an irregular equal-count grid and uses normalized cumulative column sums, and KDE, which replaces the density by Gaussian kernels with the ratio of smoothing widths controlled by a free parameter $\\sigma_r$.","core_discovery":"The paper's central claim is that the numerical conditional-CDF transform of Eq. (4), $\\gamma(x,y)$, renders its output independent of $x$: for any continuous background density $\\rho$, the transformed joint density satisfies $\\rho(x,\\gamma)=\\rho_X(x)\\cdot 1_{(0,1)}(\\gamma)$. Because of this, applying the transform to the output of an arbitrary classifier makes the classifier quasi-independent of the observable while preserving the classifier's ordering of events; the ROC curve changes minimally. On the LHC Olympics dijet example, the IRGI version reaches DCC $0.00025$ with AUC $0.912$ and the KDE version DCC $0.00018$ with AUC $0.911$, compared with baseline $0.01658$/$0.911$ and a DisCo classifier at DCC $0.00011$ whose AUC drops to $0.828$.","pith_inferences":["Because the transform is built from background only, it should transfer to anomaly-detection settings where signal is unknown: the same conditional CDF can define a background-independent coordinate for weakly supervised searches, and a classifier trained or cut on it would inherit the independence property by construction.","The construction treats the defining background sample as ground truth; a natural robustness test would be to build $\\gamma$ on one background simulation and apply it to a second simulation of the same process with different detector response or generator settings, then measure whether the DCC stays near zero.","The independence proof requires the true density and exact integration; with finite samples the numerical estimators can only approximate $\\gamma$, so the practical limit on DCC is set by sample size, smoothing width, and grid depth rather than by the method itself.","A direct comparison with the ABCDisCo variant at matched DCC, using the authors' own ABCD relative-error metric, would clarify how much of the gain comes from avoiding retraining instability and how much from the transform's exact factorization."],"forward_implications":["Any trained classifier can be post-processed into a background-quasi-independent score, so the independence requirement no longer has to be baked into the loss function during training.","In the paper's LHC Olympics example, the transformed classifier keeps AUC $0.912$ (baseline $0.911$) while pushing DCC down by two orders of magnitude, so ABCD-style background estimation can be applied without a resolution-power penalty.","The method works on the defining background sample alone; signal events are not used to build the transform, which makes it applicable to control-region or simulation-based background definitions.","Both IRGI and KDE reduce DCC consistently across five independent subsamples in the toy test, and the ABCD relative error in signal estimation for the IRGI-transformed classifier is roughly three times smaller than for the baseline in the paper's setup."],"supporting_citations":[{"why":"Supplies the existence theorem for the Rosenblatt transform, which the paper re-derives and turns into a numerical procedure.","marker":"[4]"},{"why":"Defines the distance correlation coefficient used throughout to quantify quasi-independence.","marker":"[2]"},{"why":"Introduces the ABCDisCo method and its DisCo training constraint, the baseline against which the transform is compared.","marker":"[1]"},{"why":"Defines the AUC metric used to show the transformed classifier keeps its separation power.","marker":"[3]"},{"why":"Provides the closest existing numerical uniformization technique, used to delineate the difference of x-local uniformization.","marker":"[5]"},{"why":"Supplies the public LHC Olympics 2020 dijet dataset used in the high-energy physics demonstration.","marker":"[8]"}],"fun_headline_variants":["Rosenblatt method yields classifier independence with intact AUC","Numerical Rosenblatt: decorrelate classifiers, keep AUC high","Forced independence via Rosenblatt, AUC preserved","Rosenblatt decorrelates classifiers, preserves AUC","Quasi-independence with Rosenblatt, minimal AUC loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The defining background sample used to estimate the conditional CDF must be representative of the background events in the actual analysis; if simulation or sideband data disagrees with real background, the transform is miscalibrated and $\\gamma$ will not be independent in practice.","fun_headline_variants_meta":{"raw":{"variants":["Rosenblatt method yields classifier independence with intact AUC","Numerical Rosenblatt: decorrelate classifiers, keep AUC high","Forced independence via Rosenblatt, AUC preserved","Rosenblatt decorrelates classifiers, preserves AUC","Quasi-independence with Rosenblatt, minimal AUC loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000951,"raw_usage":{"total_tokens":3967,"prompt_tokens":764,"completion_tokens":3203,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":380,"completion_tokens_details":{"reasoning_tokens":3124}},"tokens_in":380,"tokens_out":3203,"duration_ms":19524,"temperature":1.0,"reasoning_tokens":3124,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:42:11.927367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the transform on one background sample and apply it to a second sample generated from a different but plausible background density, for example the same process with a shifted calorimeter energy scale; if the measured DCC between $x$ and $\\gamma$ in the second sample stays at the original $(x,y)$ level rather than dropping toward $10^{-4}$, the practical quasi-independence claim fails.","supporting_citations":[{"cited_title":"Measuring and testing dependence by correlation of distances","cited_arxiv_id":null,"evidence_quote":"Defines the distance correlation coefficient used throughout to quantify quasi-independence."},{"cited_title":"Albertsson.Users Guide - TMVA 4 - ROOT - CERN","cited_arxiv_id":null,"evidence_quote":"Provides the closest existing numerical uniformization technique, used to delineate the difference of x-local uniformization."},{"cited_title":"2019.URL: https://zenodo.org/records/2629073","cited_arxiv_id":null,"evidence_quote":"Supplies the public LHC Olympics 2020 dijet dataset used in the high-energy physics demonstration."}],"review_version":2}