{"id":"17ca54ae-fe43-40b6-ab96-caf896a7573a","arxiv_id":"2507.15768","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LOITS is a differentiable sampling method, demonstrated in a GAN closure test, that maps sampled events back to the parameters of a target density for event-level inference.","lead":"This preprint introduces LOITS, a sampling algorithm that keeps random samples differentiable with respect to model parameters so machine learning pipelines can be trained end to end. The authors validate it on a two-dimensional toy density using a GAN, aiming at future event-level analyses of hadron structure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GAN closure test can pass by compensating for LOITS's product-density bias, so the exactness of the gradients and the unbiasedness of the inferred density are not established.","rationale":"The reader's conditional verdict is appropriate, and the weakest_assumption identifies the same gap that I consider most load-bearing: the no-MH GAN validation cannot certify that LOITS samples come from the target density. The product-density approximation is explicitly acknowledged in Section III, and the MH correction is introduced as the way to make sampling asymptotically exact, yet the main closure test omits it. In that test the discriminator only sees LOITS-generated samples, so the generator can learn a distorted image whose LOITS-biased samples match the training data; comparing that image to the ground truth then does not establish unbiased inference. The gradient derivation problems in Eqs. (4) and (6) reinforce the concern that the paper's exactness claim is not rigorously supported, but they are secondary to the sampling-bias issue because an autodiff implementation through the interpolation could still produce consistent gradients for the approximate map. No code is shipped, so the implemented gradient behavior cannot be checked independently. The proposed concrete test directly separates sampling bias from inference quality by comparing LOITS-only and LOITS+MH training, and it stresses the factorization with a strongly correlated density. This does not change the reader's conditional verdict: the LOITS idea remains plausible, but the central claim needs an unbiased or demonstrably negligible-bias demonstration before the method is used for QCF extraction.","tokens_in":19815,"tokens_out":8492,"duration_ms":103308,"concrete_test":"Re-run the Section IV closure test twice on a density from Eq. (15) in which the correlation term is made large enough that the product-density approximation is poor, once with LOITS-only and once with LOITS+MH, using identical architectures, seeds, and sample sizes. If the converged 50×50 generator images differ beyond pixel noise, the original no-MH test was compensating for LOITS sampling bias. To make the check quantitative, report the pixel-wise mean absolute deviation between each converged image and the ground truth; the LOITS+MH run should be substantially closer if the concern is valid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV validates LOITS without the MH correction defined in Eq. (14), even though Section III states that LOITS sampling is only approximate because p(x,y|θ) is replaced by p(x,y0|θ)p(x0,y|θ). The generator is trained so that the distribution of LOITS-generated samples matches the training samples; convergence therefore constrains the composition LOITS∘G, not the generator image G itself. A biased or non-normalized sampler can be compensated by a distorted generator image, so agreement between the final image and the ground truth in Figs. 5-6 does not imply that event samples are drawn from the inferred QCF density, nor that gradients of the true expected loss are exact. The analytic gradient derivation in Eqs. (4) and (6) is also not a valid derivation: Eq. (4) differentiates the CDF through a nonexistent chain rule, and Eq. (6) omits the implicit θ-dependence of xθ and has the wrong sign for inverse-function differentiation. Because the validation target is the image rather than the sampling distribution, the central claim that LOITS enables exact-gradient, event-level QCF inference is not supported by the presented evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces LOITS (local orthogonal inverse transform sampling), a differentiable sampling algorithm intended to enable gradient-based optimization of quantum correlation functions (QCFs) directly from event-level data. The method segments phase space and approximates a multi-dimensional density as a product of one-dimensional densities, then applies inverse transform sampling with local linear interpolation so that samples inherit analytic dependence on model parameters. The authors present a GAN-based closure test in which the generator produces a pixelated density image, LOITS converts that image into samples, and a discriminator compares those samples with training data drawn from a known two-dimensional density. The paper claims that LOITS computes exact gradients of the loss with respect to the parameters and that the closure test demonstrates accurate reconstruction of the test density.","tokens_in":20095,"tokens_out":5641,"duration_ms":60124,"significance":"The problem addressed is important: end-to-end, event-level inference of QCFs without surrogate models is a genuine gap in the field, and a differentiable sampling step would be a valuable building block for simulation-based inference. The LOITS idea, combining local orthogonal segmentation with inverse transform sampling and a possible Metropolis-Hastings correction, is interesting and potentially useful. The paper also explicitly connects the method to GPD/TMD imaging and the resolution question, which is a strength. However, the manuscript does not currently establish the central claims: the analytic gradient derivation contains mathematical errors, the sampling approximation is not quantitatively controlled, and the validation is only visual and in-sample. As a result, the contribution is at the stage of a promising method with a toy demonstration rather than a demonstrated exact-gradient, unbiased inference tool.","major_comments":[{"comment":"The chain rule written in Eq. (4) is not valid: CDF(x,θ) is already the integral of p(z|θ), so ∂CDF/∂θ_i = ∫_0^x (∂p(z|θ)/∂θ_i) dz; there is no functional derivative ∂CDF/∂p(z|θ) appearing in the integrand. Consequently the presented derivation of the exact gradient is incorrect as written. The paper should either state the correct expression or explain that LOITS relies on automatic differentiation through the interpolation in Eq. (7), which would compute gradients of the approximate sampler rather than of the exact inverse CDF.","section":"Section III, Eq. (4)"},{"comment":"Differentiating the identity CDF(xθ(u,θ),θ)=u with respect to θ_i gives ∂xθ/∂θ_i = −(∂CDF/∂θ_i)/(∂CDF/∂x), with the derivatives evaluated at x=xθ(u,θ). The expression in Eq. (6) omits the division by ∂CDF/∂x and the minus sign, and it does not indicate the evaluation point. This is a load-bearing error because the claim that LOITS computes exact gradients rests on this formula; the authors should provide a correct derivation or empirically verify gradients (e.g., against finite differences).","section":"Section III, Eq. (6)"},{"comment":"The local separability approximation p(x,y|θ)≈p(x,y0|θ)p(x0,y|θ) is introduced without normalization or a quantitative accuracy bound; for a general correlated density the approximation can be arbitrarily poor. The GAN closure test in Section IV uses LOITS without the MH correction of Eq. (14), so the generator is optimized against samples from the approximate product density. The discriminator therefore constrains the composition LOITS∘G, not the generator image pG itself; a biased sampler can be compensated by a distorted image. The comparison of the generator image to the ground truth in Figs. 5-6 does not establish that event samples are drawn from the inferred density.","section":"Section III, local separability approximation"},{"comment":"The closure test is evaluated only by visual comparison and pixel-ratio maps (Figs. 5-6). There is no quantitative metric (e.g., two-sample test, KL divergence, χ² per bin), no uncertainty estimate, and no baseline comparison (for instance, training the same GAN with a non-differentiable but unbiased sampler, or comparing to unbinned maximum likelihood on the parametric form). Without these, the claim that LOITS yields accurate reconstruction and informative gradients is not demonstrated.","section":"Section IV, closure test evaluation"},{"comment":"The claim that LOITS+MH provides a differentiable sampler is not justified. The acceptance step in Eq. (14) involves a discrete accept/reject decision; standard Metropolis–Hastings does not define a pathwise-differentiable map from θ to the chain state, because rejected proposals cause the state to be carried over with zero derivative through the random accept/reject step. The paper should provide a rigorous argument or a reparameterized MH formulation if this claim is retained.","section":"Section III, LOITS+MH differentiability"}],"minor_comments":[{"comment":"The statement 'CDF(0,θ) = CDF(1,θ) = 1' should read CDF(0,θ)=0 and CDF(1,θ)=1; otherwise the boundary values contradict the definition in Eq. (3).","section":"Section III, text near Eq. (3)"},{"comment":"The algorithm is called 'Local Orthogonal' in Section III but 'Longitudinal Orthogonal' in Section V; please make the name consistent throughout the paper.","section":"Section V, Conclusions"},{"comment":"Figure 3 would benefit from colorbars and normalized histograms so that the sampling fidelity can be judged quantitatively rather than by eye.","section":"Figure 3"},{"comment":"The procedure of averaging the generator output over 1,000 latent noise samples before passing it to LOITS is unusual and should be justified; as written, it suggests the generator is trained on an ensemble-averaged image rather than on individual draws, which may affect mode coverage.","section":"Section IV, GAN setup"},{"comment":"Reference [95] is cited for the distributed training strategy, but the connection to the asynchronous generative inverse problem solver is not obvious; please clarify what is taken from that work.","section":"Section IV, training details"}],"recommendation":"major_revision","confidential_remarks":"The paper is a methods contribution with a physics motivation; the toy closure test is reasonable as a first step, but the analytic gradient errors and the lack of quantitative validation make the central claim unsupported in the current form. The authors should also reconsider whether the 'exactness' language can be retained given that the sampling is approximate; a more precise statement about gradients of the approximate sampler would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kevin,\n\nLOITS is a plausible idea, but the paper oversells it. The core innovation—using local orthogonal inverse transform sampling with local interpolation to make sampling differentiable—is a genuine combination, even if each ingredient is standard. The physics motivation is clear and the paper is well-positioned in the simulation-based inference literature. The toy demonstrations show the algorithm can drive GAN training on a simple 2D density.\n\nThe soft spots are real. The analytic gradient derivation in Eqs. (4) and (6) is wrong as written: Eq. (4) has an invalid chain rule, and Eq. (6) misses the implicit θ-dependence of xθ and has the wrong sign. If the implementation uses automatic differentiation, these errors may not change the numerics, but the paper claims 'exact gradients' and nothing in the derivation supports that. Fixing this is straightforward, but it needs to be done.\n\nMore importantly, the validation has a gap. The GAN closure test uses LOITS without the MH correction defined in Eq. (14). The training loss constrains the composition LOITS∘G, not the generator image G itself. A biased or non-normalized sampler can be compensated by a distorted generator image, so the visual agreement between the generator output and the ground truth in Figs. 5 and 6 does not establish that the image is the density being sampled. To make the case, they need a quantitative comparison between the generated samples and the true samples (e.g., a distance metric), or they should run the test with the MH-corrected sampler. There is also no baseline, no error bars, and no code, which makes the closure test hard to interpret.\n\nI don't think the central idea is wrong, but the exact-gradient claim and the unbiasedness of the inferred density are not supported by the presented evidence. The paper would benefit from a corrected derivation, a quantitative validation, and ideally a public implementation.\n\nThis is for someone working on differentiable simulation-based inference for QCD, or on unbinned unfolding. It deserves a serious referee—the problem is important and the approach is promising—but it needs major revision before it can be accepted as a reliable building block.","headline":"LOITS is a plausible but oversold method for differentiable sampling; the gradient derivation is flawed and the closure test doesn't validate the sampling distribution.","tokens_in":20560,"tokens_out":4653,"would_cite":false,"duration_ms":49833,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces a differentiable sampling algorithm, LOITS, that lets gradient-based optimization flow from collision events back to the parameters of hadron-structure distributions, and validates it by training a GAN to recover a…","keywords":["differential programming","inverse transform sampling","LOITS","quantum correlation functions","generalized parton distributions","generative adversarial networks","event-level inference","simulation-based inference"],"falsifier":"Run the closure test on a two-dimensional Gaussian with correlation coefficient near 0.9 and known parameter dependence, then compare the histogram of LOITS samples against the exact density and compare gradients computed through LOITS against finite-difference gradients for a small parameter step; visible histogram bias or gradient disagreement beyond numerical precision would refute the claims of accurate sampling and exact differentiability.","tokens_in":19653,"feed_emoji":"⚛️","tokens_out":6493,"duration_ms":68813,"temperature":0.7,"pith_summary":"Reconstructing the quark and gluon structure of hadrons from collision data is an inverse problem: the data are individual events, while the quantities of interest are quantum correlation functions (QCFs) such as parton distribution functions, transverse-momentum distributions, and generalized parton distributions. The paper's central claim is that this inference can be performed directly on unbinned events by making the sampling step differentiable. Standard Monte Carlo sampling breaks the gradient chain from the model parameters to the loss, while the introduced LOITS algorithm builds a differentiable path from phase-space samples back to the parameters. A closure test with a generative adversarial network recovers a known two-dimensional density from 10,000 to 1,000,000 sampled events, with reconstruction quality improving as data grow. If the claim holds, QCF parameters could be optimized end-to-end from raw event-level data without binned summary statistics or surrogate models.","feed_headline":"A new sampling algorithm makes hadron-structure fits differentiable","feed_subtitle":"LOITS opens the path to gradient-based, event-level inference of quark-gluon distributions.","key_machinery":"The central object is the LOITS sampler. For a density $p(x,y|\\theta)$ it fixes orthogonal reference slices $(x_0,y_0)$, treats $p(x,y_0|\\theta)\\,p(x_0,y|\\theta)$ as the local sampling density, builds one-dimensional CDFs along $x$ and $y$ in each segment, and inverts them by local linear interpolation so that a uniform random input $u$ maps to a sample whose position depends analytically on $\\theta$. A subsequent accept-reject correction step can restore asymptotic exactness when the product approximation is inaccurate. The work this machinery does is turning a stochastic sampling operation into a differentiable layer that can sit between a neural-network density image and a loss function defined on samples.","core_discovery":"The paper proposes that the non-differentiability of sampling is the missing link in event-level QCF inference, and LOITS is designed to remove it. LOITS partitions phase space into orthogonal segments, approximates the target density inside each segment as a product of one-dimensional densities, computes local cumulative distribution functions, and generates samples by inverting those local CDFs with a differentiable interpolation. Because the CDF values depend on the model parameters, the generated samples inherit computable gradients through the chain rule, and the stated goal is to enable exact gradients of the loss with respect to the model parameters. The authors demonstrate in a closure test that a GAN trained through this differentiable sampler reconstructs a beta-like test density with resolution that improves with event count, and they argue the algorithm generalizes to arbitrary dimensions and to other generative or likelihood-based inference settings.","pith_inferences":["When LOITS is used without its accept-reject correction, the generator is optimized against an approximate product density; practitioners should treat the uncorrected sampler as a biased surrogate and calibrate the bias before trusting the gradients.","Strongly correlated, high-dimensional phase spaces will likely require finer segmentation than the smooth closure-test density, so an adaptive segmentation scheme or a dimension-reduction step would be a natural stress test of the method.","Combining LOITS with a differentiable detector surrogate would complete the end-to-end pipeline the paper sketches, allowing gradients to flow through detector effects as well as through the phase-space sampling.","The product-density construction suggests LOITS could also serve as a proposal generator for other exact Markov-chain samplers, trading some sampling efficiency for reliable unbiasedness."],"forward_implications":["QCF model parameters (PDFs, TMDs, GPDs) could be optimized directly from unbinned event-level data by gradient descent, skipping the traditional binned cross-section extraction and unfolding steps.","The differentiable sampler can be inserted into any simulation-based inference pipeline in which the theoretical density depends smoothly on the model parameters, including likelihood-based fits and generative models.","Neural-network representations of multidimensional hadron structure, such as pixelated GPD images, can be trained without an explicit likelihood function.","The resolution of the reconstructed density grows with available event statistics, giving a concrete handle on how much data a planned experiment needs to image hadron structure at a given scale."],"supporting_citations":[{"why":"Supplies the GAN architecture used to turn a latent vector into a pixelated density image in the closure test.","marker":"[94]"},{"why":"Gives the QCD factorization theorems that connect QCFs to differential cross sections, the target input for LOITS-based event generation.","marker":"[2]"},{"why":"Defines the simulation-based inference paradigm that motivates reconstructing parameters directly from sampled events.","marker":"[33]"},{"why":"Provides the generative-model-plus-accept-reject strategy that LOITS adapts to make its approximate sampling asymptotically exact.","marker":"[91]"},{"why":"Represents the iterative Bayesian unfolding baseline that the event-level, differentiable approach is intended to supersede.","marker":"[1]"},{"why":"Shows how exclusive observables sensitive to GPDs can be converted into event-level samples, a pipeline LOITS aims to make differentiable.","marker":"[41]"}],"fun_headline_variants":["Differentiable sampling paves way to event-level hadron fits","LOITS algorithm makes hadron-structure sampling differentiable","Gradient-based event inference for quark-gluon distributions","New sampling method enables end-to-end hadron-structure learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands on the assumption that inside each small phase-space patch the distribution factorizes into separate x and y pieces; when x and y are strongly correlated this product approximation can be badly wrong, and the paper gives no quantitative bound on the resulting bias.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable sampling paves way to event-level hadron fits","LOITS algorithm makes hadron-structure sampling differentiable","Gradient-based event inference for quark-gluon distributions","New sampling method enables end-to-end hadron-structure learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000164,"raw_usage":{"total_tokens":1231,"prompt_tokens":913,"completion_tokens":318,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":250}},"tokens_in":529,"tokens_out":318,"duration_ms":3522,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:23:56.876296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closure test on a two-dimensional Gaussian with correlation coefficient near 0.9 and known parameter dependence, then compare the histogram of LOITS samples against the exact density and compare gradients computed through LOITS against finite-difference gradients for a small parameter step; visible histogram bias or gradient disagreement beyond numerical precision would refute the claims of accurate sampling and exact differentiability.","supporting_citations":[{"cited_title":"Generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"Supplies the GAN architecture used to turn a latent vector into a pixelated density image in the closure test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the QCD factorization theorems that connect QCFs to differential cross sections, the target input for LOITS-based event generation."},{"cited_title":", K}.(8) 6 0.2 0.4 0.6 0.8 x 0.2 0.4 0.6 0.8 y LOITS(4×4) 0.2 0.4 0.6 0.8 x 0.2 0.4 0.6 0.8 y LOITS(4×4) + MH 0.2 0.4 0.6 0.8 x 0.2 0.4 0.6 0.8 y LOITS(50×50) FIG","cited_arxiv_id":null,"evidence_quote":"Represents the iterative Bayesian unfolding baseline that the event-level, differentiable approach is intended to supersede."}],"review_version":1}