Pith. sign in

REVIEW 2 major objections 6 minor 63 references

A hybrid normalizing-flow model turns a 110-parameter non-Gaussian near-detector likelihood into a portable, closed-form density that matches MCMC sampling at 98% relative effective sample size.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 07:03 UTC pith:HZJHHIT4

load-bearing objection Solid methods paper: hybrid NF turns a 110-param ND systematics likelihood into a portable, MCMC-matching density with 98% rESS; replica is synthetic but scoped honestly and code/data are public. the 2 major comments →

arxiv 2607.03477 v2 pith:HZJHHIT4 submitted 2026-07-03 hep-ex

Flow-Based Surrogates for High-Dimensional Likelihoods in Experimental Neutrino Physics

Mathias El Baz , Lorenzo Giannessi , Adrien Blanchet , Federico S\'anchez This is my paper
classification hep-ex
keywords neutrino physicslong-baseline oscillation experimentssystematic uncertaintieslikelihood-based inferencenormalizing flowsnear-detector constraintsnon-Gaussian likelihoods
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Long-baseline neutrino experiments must carry constrained systematic uncertainties from near-detector fits into far-detector oscillation analyses. When those constrained likelihoods are high-dimensional and non-Gaussian, the usual post-fit Gaussian is portable but biased, while MCMC samples are faithful but not closed-form or easily reusable. This paper shows that a hybrid normalizing flow—coupling layers for the large linear block and a conditional autoregressive spline flow for the small non-linear block—learns a continuous density that is both samplable and pointwise evaluable. On a controlled 110-parameter replica with ten deliberately non-Gaussian systematics, the trained flow reaches 98% relative effective sample size against roughly 5% for the Gaussian, matches MCMC one- and two-dimensional marginals and flux predictions, and can thereafter be used without the original data or fit machinery. The result is a practical middle ground for exporting constrained systematics into joint and multi-experiment analyses.

Core claim

A hybrid normalizing-flow surrogate, trained from an initial Gaussian approximation of a realistic 110-parameter near-detector likelihood replica, recovers the non-Gaussian structure of that likelihood with 98% relative effective sample size, matches a long MCMC reference in marginals and flux predictions, and remains closed-form, samplable, and pointwise evaluable for downstream use.

What carries the argument

Hybrid normalizing flow q_NF(η_A, η_B) = q_AR(η_A | η_B) q_CL(η_B): coupling layers on the high-dimensional approximately linear block, conditional autoregressive rational-quadratic spline flow on the low-dimensional non-linear block, with an optional linear stage initialized to the post-fit Gaussian.

Load-bearing premise

The hand-built 110-parameter replica with ten custom non-linear weight functions is close enough to real near-detector likelihoods that the same hybrid architecture and training scheme will transfer without major redesign.

What would settle it

Train the same hybrid flow on an actual experiment near-detector likelihood (or a higher-fidelity replica with many more non-linear parameters and realistic multi-variable weights) and check whether relative effective sample size stays near 98% and whether one- and two-dimensional marginals and propagated flux spectra still match a well-converged MCMC reference.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The manuscript proposes hybrid normalizing flows as portable, closed-form surrogates for high-dimensional constrained systematic likelihoods in long-baseline neutrino analyses. A coupling-layer block models the approximately linear high-dimensional sector while a conditional autoregressive rational-quadratic spline flow models a smaller non-linear block, initialized from a post-fit Gaussian. On a self-contained 110-parameter near-detector likelihood replica (100 linear flux-like parameters plus 10 non-linear interaction-like weights), the trained flow reaches a relative effective sample size of 98% versus ~5% for the Gaussian approximation, matches Metropolis–Hastings MCMC one- and two-dimensional marginals and flux predictions, and remains samplable and pointwise evaluable for downstream propagation.

Significance. The work fills a genuine methodological gap between portable but biased post-fit Gaussians, faithful but non-evaluable MCMC chains, and heavy full-likelihood implementations that are hard to share across ND–FD or multi-experiment fits. The hybrid architecture is well matched to the linear/non-linear partition typical of near-detector systematics, training and evaluation costs are quantified (including scaling with the autoregressive block size), and both code and data are released. If the approach transfers to production analyses, it would improve uncertainty propagation and joint fits without requiring re-execution of upstream fits. Strengths include multi-faceted quantitative validation against a long MCMC reference, public reproducibility artifacts, and a clear scoping of the demonstration to a controlled replica rather than an overclaim of immediate production readiness.

major comments (2)
  1. Section 2.3 and Methods: the MCMC reference (4.2e6 Metropolis–Hastings steps, 2e5 burn-in, acceptance tuned to 0.234) is the sole ground truth for marginals, flux predictions, and the rESS claim. In 110 dimensions, random-walk MH can mix slowly; the manuscript reports neither integrated autocorrelation times, effective sample size of the chain itself, nor multi-chain diagnostics (e.g., Gelman–Rubin). Without these, it is hard to judge residual reference bias. Please add standard MCMC convergence diagnostics (or a short HMC/NUTS comparison) so that the 98% rESS and flux-level agreement can be interpreted with known reference quality.
  2. Sections 2.2 and 4.3: the ten non-linear parameters use custom single-kinematic weight functions (Eqs. 6–7, Table 4) rather than production cross-section reweighting. The hybrid design and training protocol exploit a known linear/non-linear partition; the central demonstration is therefore only as transferable as that partition and the induced non-Gaussianity. A short quantitative comparison of profile asymmetry or tail weight against published T2K ND nuisance profiles (or a statement that such a comparison is left to future work with a production likelihood) would make the scope of the claim clearer without changing the controlled-replica result.
minor comments (6)
  1. Figure 6: the horizontal axis is labeled only as training progress; please state whether the unit is epochs, likelihood evaluations, or wall time, and mark the 64 000-epoch stopping point used for the final model.
  2. Equation (21) and Section 4.5.1: the importance-sampling estimator of the symmetric KL uses the estimated normalizing constant Z-hat (Eq. 23). A brief note on the stability of Z-hat across the FIFO buffer (or its variance) would help readers reproduce the training.
  3. Table 1: sampling throughput is given for increasing DA at fixed total dimension; please state whether the coupling-block depth and conditioner widths were held fixed or re-tuned when DA was increased.
  4. Abstract and Introduction: “future near-detector to far-detector fits” is mentioned as a use case but not demonstrated. A single sentence in the Discussion clarifying that ND–FD insertion is left for follow-up work would avoid any ambiguity about what is shown versus proposed.
  5. Figure 3: the corner plot shows only a subset of the ten non-linear parameters; listing which pairs are omitted (or providing a full corner in the supplement) would aid completeness.
  6. Minor typography: “S´ anchez” and similar accented names appear inconsistently in the author list and references; please normalize.

Circularity Check

0 steps flagged

No significant circularity: the hybrid NF is trained by importance sampling against the true unnormalized likelihood and validated against an independent MCMC of the same target; the post-fit Gaussian is only an initial proposal/initializer.

full rationale

The derivation chain is self-contained and non-circular. The target density is the normalized systematic likelihood p(η) ∝ L(η) of a constructed 110-parameter replica (Eq. 1, Methods 4.1–4.3). The hybrid flow q_NF is trained by minimizing a symmetric KL estimated via importance sampling (Eqs. 20–27), with proposals that begin from the post-fit Gaussian but are iteratively replaced by lagged flow checkpoints; the Gaussian therefore never defines the target. Fidelity is measured by rESS (reaching 98 % vs ~5 % for the Gaussian), 1-D/2-D marginals, and flux predictions, all compared to an independent Metropolis–Hastings MCMC chain of the identical likelihood (Figs. 2–7). Self-citations (e.g. prior NF applications by overlapping authors) supply background on related neutrino-physics uses of flows but are not load-bearing uniqueness theorems or ansätze that force the present architecture or numerical claims. The linear/non-linear partition and hybrid design are justified by computational cost and the known structure of the replica, not by circular definition. No fitted constant is renamed a prediction, and no result reduces by construction to its inputs. Score 1 reflects only the presence of ordinary self-citations that do not underwrite the central claim.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard change-of-variable density estimation, the Barlow–Beeston lite statistical model, and the authors’ engineered partition of systematics into linear and non-linear blocks. Free parameters are the usual neural-network and training hyperparameters; no new physical entities are postulated. The main domain assumptions are that the synthetic replica captures the relevant non-Gaussian structure of real near-detector fits and that the post-fit Gaussian is a sufficiently covering initial proposal.

free parameters (5)
  • Number of coupling blocks / autoregressive layers = 12 / 12
    Architecture depth (12 each) chosen by hand; affects expressivity and cost.
  • Spline bins and tail bound B = 9 bins, B=6
    RQ-NSF resolution and support; set to 9 bins, B=6.
  • Conditioner network widths = 1×256 / 2×512
    Coupling 1×256, autoregressive 2×512 units; standard capacity knobs.
  • Learning rate and FIFO buffer length = 1e-5, K=50
    Adam AMSGrad lr=1e-5, K=50 batches of 10k; control training stability.
  • Non-linear weight amplitudes α and kinematic ranges = α ∈ {0.1,0.3}
    Hand-chosen strengths (0.1–0.3) and kinematic intervals that generate the non-Gaussian features of the replica.
axioms (4)
  • standard math Change-of-variable formula for invertible maps yields exact density evaluation and sampling.
    Standard NF foundation used throughout Sec. 2.1 and 4.4.
  • domain assumption The constrained systematic likelihood can be partitioned into a high-dimensional approximately linear block and a low-dimensional non-linear block.
    Justifies the hybrid factorization q_AR(η_A|η_B) q_CL(η_B); stated in Sec. 2.1 and Methods 4.4.3.
  • domain assumption Barlow–Beeston lite plus quadratic systematic penalty adequately models the statistical and prior structure of the near-detector likelihood.
    Likelihood definition in Methods 4.2; standard in the field but still an assumption of the replica.
  • ad hoc to paper Post-fit Gaussian (MINUIT + HESSE) provides a covering initial proposal and linear-layer initialization for training.
    Used to seed the FIFO dataset and T^(1); not required in principle but load-bearing for the reported training procedure (Sec. 4.5).

pith-pipeline@v1.1.0-grok45 · 24502 in / 2841 out tokens · 32849 ms · 2026-07-13T07:03:24.385123+00:00 · methodology

0 comments
read the original abstract

Precision long-baseline neutrino experiments use near-detector data to constrain systematic uncertainties on the unoscillated neutrino flux, a prerequisite for accurate oscillation parameter measurements at the far detector. When the constrained likelihood is high-dimensional and non-Gaussian, this procedure demands advanced statistical treatment. Here we show that normalizing flows provide faithful and portable likelihood models for this problem. Leveraging an initial Gaussian approximation of the likelihood, we train a hybrid architecture combining coupling transformations and autoregressive spline flows. We demonstrate the method on a representative near-detector likelihood replica with 110 systematic uncertainty parameters, 10 of which explicitly introduce non-Gaussianities in the posterior. The trained model achieves a relative effective sample size of 98%, compared with about 5% for the Gaussian approximation, and reproduces a Markov chain Monte Carlo reference while remaining closed-form, samplable, and pointwise evaluable, making it suited to downstream uncertainty propagation and future near-detector to far-detector fits.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 1 canonical work pages

  1. [1]

    Nature580, 339–344 (2020) https://doi.org/10

    T2K Collaboration: Constraint on the matter-antimatter symmetry-violating phase in neutrino oscillations. Nature580, 339–344 (2020) https://doi.org/10. 1038/s41586-020-2177-0 arXiv:1910.03887 [hep-ex]

    Pith/arXiv arXiv 2020

  2. [2]

    Abe, K.,et al.: The T2K Experiment. Nucl. Instrum. Meth. A659, 106–135 (2011) https://doi.org/10.1016/j.nima.2011.06.067 arXiv:1106.1238 [physics.ins- det]

    doi:10.1016/j.nima.2011.06.067 2011

  3. [3]

    Abe, K.,et al.: Evidence of Electron Neutrino Appearance in a Muon Neutrino Beam. Phys. Rev. D88(3), 032002 (2013) https://doi.org/10.1103/PhysRevD. 88.032002 arXiv:1304.0841 [hep-ex]

    Pith/arXiv arXiv doi:10.1103/physrevd 2013

  4. [4]

    Journal of Physics G: Nuclear and Particle Physics44(5), 24 054001 (2017) https://doi.org/10.1088/1361-6471/aa61cc arXiv:1609.00258 [hep- ph]

    Ankowski, A.M., Mariani, C.: Systematic uncertainties in long-baseline neutrino- oscillation experiments. Journal of Physics G: Nuclear and Particle Physics44(5), 24 054001 (2017) https://doi.org/10.1088/1361-6471/aa61cc arXiv:1609.00258 [hep- ph]

    Pith/arXiv arXiv doi:10.1088/1361-6471/aa61cc 2017

  5. [5]

    DUNE Collaboration: Deep Underground Neutrino Experiment (DUNE), Far Detector Technical Design Report, Volume II: DUNE Physics (2020)

    2020

  6. [6]

    T2K Collaboration: Measurements of neutrino oscillation parameters from the T2K experiment using 3.6×10 21 protons on target (2023)

    2023

  7. [7]

    Annual Review of Nuclear and Particle Science73, 69–93 (2023) https://doi.org/10.1146/ annurev-nucl-102020-101615

    Di Lodovico, F., Patterson, R.B., Shiozawa, M., Worcester, E.: Experimen- tal considerations in long-baseline neutrino oscillation measurements. Annual Review of Nuclear and Particle Science73, 69–93 (2023) https://doi.org/10.1146/ annurev-nucl-102020-101615

    2023

  8. [8]

    T2K Collaboration: Results from the T2K experiment on neutrino mixing including a new far detectorµ-like sample (2025)

    2025

  9. [9]

    T2K Collaboration: Constraining Neutrino Interaction Uncertainties for Neutrino Oscillation Measurements at the T2K Experiment (2026)

    2026

  10. [10]

    Statistical Science14(1), 1–28 (1999) https://doi

    Berger, J.O., Liseo, B., Wolpert, R.L.: Integrated likelihood methods for elimi- nating nuisance parameters. Statistical Science14(1), 1–28 (1999) https://doi. org/10.1214/ss/1009211804

    doi:10.1214/ss/1009211804 1999

  11. [11]

    Phys- ical Review Letters134, 011801 (2025) https://doi.org/10.1103/PhysRevLett

    Super-Kamiokande Collaboration, T2K Collaboration: First joint oscillation anal- ysis of Super-Kamiokande atmospheric and T2K accelerator neutrino data. Phys- ical Review Letters134, 011801 (2025) https://doi.org/10.1103/PhysRevLett. 134.011801 arXiv:2405.12488 [hep-ex]

    doi:10.1103/physrevlett 2025

  12. [12]

    NOvA Collaboration, T2K Collaboration: Joint neutrino oscillation analysis from the T2K and NOvA experiments (2025)

    2025

  13. [13]

    Esteban, I., Gonzalez-Garcia, M.C., Maltoni, M., Martinez-Soler, I., Pinheiro, J.P., Schwetz, T.: NuFit-6.0: Updated global analysis of three-flavor neutrino oscillations (2024)

    2024

  14. [14]

    DUNE Collaboration: Experiment Simulation Configurations Approximating DUNE TDR (2021)

    2021

  15. [15]

    SciPost Physics12(1) (2022) https://doi.org/10.21468/scipostphys.12.1.037 25

    Cranmer, K., Kraml, S., Prosper, H., Bechtle, P., Bernlochner, F., Bloch, I.M., Canonero, E., Chrzaszcz, M., Coccaro, A., Conrad, J., Cowan, G., Feickert, M., Ferreiro, N., Fowlie, A., Heinrich, L.A., Held, A., Kuhr, T., Kvellestad, A., Madigan, M., Mahmoudi, F.N., Mor˚ a, K.D., Neubauer, M.S., Pierini, M., Rojo, J., Sekmen, S., Silvestrini, L., Sanz, V.,...

    doi:10.21468/scipostphys.12.1.037 2022

  16. [16]

    Com- puter Physics Communications77(2), 219–228 (1993) https://doi.org/10.1016/ 0010-4655(93)90005-W

    Barlow, R., Beeston, C.: Fitting using finite monte carlo samples. Com- puter Physics Communications77(2), 219–228 (1993) https://doi.org/10.1016/ 0010-4655(93)90005-W

    1993

  17. [17]

    The Journal of Chemical Physics21(6), 1087–1092 (1953) https://doi.org/10.1063/1.1699114

    Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E.: Equation of state calculations by fast computing machines. The Journal of Chemical Physics21(6), 1087–1092 (1953) https://doi.org/10.1063/1.1699114

    doi:10.1063/1.1699114 1953

  18. [18]

    Biometrika57(1), 97–109 (1970) https://doi.org/10.1093/biomet/ 57.1.97

    Hastings, W.K.: Monte carlo sampling methods using markov chains and their applications. Biometrika57(1), 97–109 (1970) https://doi.org/10.1093/biomet/ 57.1.97

    doi:10.1093/biomet/ 1970

  19. [19]

    Technical report, New York U., New York (2012)

    Cranmer, K., Lewis, G., Moneta, L., Shibata, A., Verkerke, W.: HistFactory: A tool for creating statistical models for use with RooFit and RooStats. Technical report, New York U., New York (2012). https://doi.org/10.17181/ CERN-OPEN-2012-016 . https://cds.cern.ch/record/1456844

    arXiv 2012

  20. [20]

    IEEE Transactions on Pattern Analysis and Machine Intelligence43(11), 3964–3979 (2021) https://doi.org/10.1109/tpami

    Kobyzev, I., Prince, S.J.D., Brubaker, M.A.: Normalizing flows: An introduc- tion and review of current methods. IEEE Transactions on Pattern Analysis and Machine Intelligence43(11), 3964–3979 (2021) https://doi.org/10.1109/tpami. 2020.2992934

    doi:10.1109/tpami 2021

  21. [21]

    Journal of Machine Learning Research22(57), 1–64 (2021)

    Papamakarios, G., Nalisnick, E., Rezende, D.J., Mohamed, S., Lakshmi- narayanan, B.: Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research22(57), 1–64 (2021)

    2021

  22. [22]

    Proceedings of the National Academy of Sciences117(48), 30055–30062 (2020) https://doi.org/10.1073/pnas.1912789117

    Cranmer, K., Brehmer, J., Louppe, G.: The frontier of simulation-based inference. Proceedings of the National Academy of Sciences117(48), 30055–30062 (2020) https://doi.org/10.1073/pnas.1912789117

    doi:10.1073/pnas.1912789117 2020

  23. [23]

    Chapter in Artificial Intelligence for Particle Physics (World Scientific)

    Brehmer, J., Cranmer, K.: Simulation-based inference methods for particle physics (2020) arXiv:2010.06439 [hep-ph]. Chapter in Artificial Intelligence for Particle Physics (World Scientific)

    Pith/arXiv arXiv 2020

  24. [24]

    Nature Reviews Physics3, 305 (2021) https://doi.org/10.1038/s42254-021-00305-6

    Brehmer, J.: Simulation-based inference in particle physics. Nature Reviews Physics3, 305 (2021) https://doi.org/10.1038/s42254-021-00305-6

    doi:10.1038/s42254-021-00305-6 2021

  25. [25]

    Journal of Open Source Software5(52), 2505 (2020) https: //doi.org/10.21105/joss.02505

    Tejero-Cantero, A., Boelts, J., Deistler, M., Lueckmann, J.-M., Durkan, C., Gon¸ calves, P.J., Greenberg, D.S., Macke, J.H.: sbi: A toolkit for simulation- based inference. Journal of Open Source Software5(52), 2505 (2020) https: //doi.org/10.21105/joss.02505

    doi:10.21105/joss.02505 2020

  26. [26]

    Green, S.R., Simpson, C., Gair, J.: Gravitational-wave parameter estimation with autoregressive neural network flows. Phys. Rev. D102(10), 104057 (2020) https: //doi.org/10.1103/PhysRevD.102.104057 arXiv:2002.07656 [astro-ph.IM] 26

    Pith/arXiv arXiv doi:10.1103/physrevd.102.104057 2020

  27. [27]

    Dax, M., Green, S.R., Gair, J., Macke, J.H., Buonanno, A., Sch¨ olkopf, B.: Real- time gravitational wave science with neural posterior estimation. Phys. Rev. Lett.127(24), 241103 (2021) https://doi.org/10.1103/PhysRevLett.127.241103 arXiv:2106.12594 [gr-qc]

    Pith/arXiv arXiv doi:10.1103/physrevlett.127.241103 2021

  28. [28]

    Dax, M., Green, S.R., Gair, J., P¨ urrer, M., Wildberger, J., Macke, J.H., Buo- nanno, A., Sch¨ olkopf, B.: Neural importance sampling for rapid and reliable gravitational-wave inference. Phys. Rev. Lett.130(17), 171403 (2023) https: //doi.org/10.1103/PhysRevLett.130.171403 arXiv:2210.05686 [gr-qc]

    Pith/arXiv arXiv doi:10.1103/physrevlett.130.171403 2023

  29. [29]

    Krause, C., Shih, D.: Fast and accurate simulations of calorimeter showers with normalizing flows. Phys. Rev. D107(11), 113003 (2023) https://doi.org/10.1103/ PhysRevD.107.113003 arXiv:2106.05285 [physics.ins-det]

    Pith/arXiv arXiv 2023

  30. [30]

    Krause, C., Shih, D.: Accelerating accurate simulations of calorimeter showers with normalizing flows and probability density distillation. Phys. Rev. D107(11), 113004 (2023) https://doi.org/10.1103/PhysRevD.107.113004 arXiv:2110.11377 [physics.ins-det]

    Pith/arXiv arXiv doi:10.1103/physrevd.107.113004 2023

  31. [31]

    Albergo, M.S., Kanwar, G., Shanahan, P.E.: Flow-based generative models for markov chain monte carlo in lattice field theory. Phys. Rev. D100(3), 034515 (2019) https://doi.org/10.1103/PhysRevD.100.034515 arXiv:1904.12072 [hep-lat]

    Pith/arXiv arXiv doi:10.1103/physrevd.100.034515 2019

  32. [32]

    Kanwar, G., Albergo, M.S., Boyda, D., Cranmer, K., Hackett, D.C., Racani` ere, S., Rezende, D.J., Shanahan, P.E.: Equivariant flow-based sampling for lattice gauge theory. Phys. Rev. Lett.125(12), 121601 (2020) https://doi.org/10.1103/ PhysRevLett.125.121601 arXiv:2003.06413 [hep-lat]

    Pith/arXiv arXiv 2020

  33. [33]

    Albergo, M.S., Kanwar, G., Racani` ere, S., Rezende, D.J., Urban, J.M., Boyda, D., Cranmer, K., Hackett, D.C., Shanahan, P.E.: Flow-based sampling for fermionic lattice field theories. Phys. Rev. D104(11), 114507 (2021) https://doi.org/10. 1103/PhysRevD.104.114507 arXiv:2106.05934 [hep-lat]

    Pith/arXiv arXiv 2021

  34. [34]

    Physical Review D101, 113001 (2020) https://doi.org/10.1103/PhysRevD.101.113001 arXiv:2002.09436 [hep-ph]

    Pina-Otey, S., S´ anchez, F., Gait´ an, V., Lux, T.: Likelihood-free inference of exper- imental neutrino oscillations using neural spline flows. Physical Review D101, 113001 (2020) https://doi.org/10.1103/PhysRevD.101.113001 arXiv:2002.09436 [hep-ph]

    Pith/arXiv arXiv doi:10.1103/physrevd.101.113001 2020

  35. [35]

    Pina-Otey, S., S´ anchez, F., Lux, T., Gaitan, V.: Exhaustive neural importance sampling applied to monte carlo event generation. Phys. Rev. D102, 013003 (2020) https://doi.org/10.1103/PhysRevD.102.013003

    doi:10.1103/physrevd.102.013003 2020

  36. [36]

    El Baz, M., S´ anchez, F.: Fast posterior probability sampling with normalizing flows and its applicability in bayesian analysis in particle physics. Phys. Rev. D 109, 032008 (2024) https://doi.org/10.1103/PhysRevD.109.032008 27

    doi:10.1103/physrevd.109.032008 2024

  37. [37]

    Communications Physics (2026) https://doi.org/10.1038/ s42005-026-02499-6 arXiv:2507.23297 [hep-ex]

    Gavrikov, A., Serafini, A., Dolzhikov, D., Garfagnini, A., Gonchar, M., Grassi, M.,et al.: Simulation-based inference for precision neutrino physics through neu- ral monte carlo tuning. Communications Physics (2026) https://doi.org/10.1038/ s42005-026-02499-6 arXiv:2507.23297 [hep-ex]

    Pith/arXiv arXiv 2026

  38. [38]

    IceCube Collaboration: Neural posterior estimation of the neutrino direction in IceCube using transformer-encoded normalizing flows on the sphere (2026)

    2026

  39. [39]

    El Baz, M., S´ anchez, F., Jachowicz, N., Niewczas, K., Jha, A.K., Nikolakopoulos, A.: Efficient monte carlo event generation for neutrino-nucleus exclusive cross sections. Phys. Rev. D111, 113001 (2025) https://doi.org/10.1103/v7sz-vn3b

    doi:10.1103/v7sz-vn3b 2025

  40. [40]

    https://arxiv.org/abs/1605.08803

    Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using Real NVP (2017). https://arxiv.org/abs/1605.08803

    Pith/arXiv arXiv 2017

  41. [41]

    In: Advances in Neural Information Processing Systems, vol

    Papamakarios, G., Pavlakou, T., Murray, I.: Masked autore- gressive flow for density estimation. In: Advances in Neural Information Processing Systems, vol. 30, pp. 2338–2347 (2017). https://proceedings.neurips.cc/paper/2017/file/6c1da886822c67822bcf3679d04369fa- Paper.pdf

    2017

  42. [42]

    https://arxiv.org/abs/1912.00042

    Winkler, C., Worrall, D., Hoogeboom, E., Welling, M.: Learning Likelihoods with Conditional Normalizing Flows (2019). https://arxiv.org/abs/1912.00042

    Pith/arXiv arXiv 2019

  43. [43]

    The Annals of Applied Probability7(1), 110–120 (1997) https://doi.org/10.1214/aoap/1034625254

    Roberts, G.O., Gelman, A., Gilks, W.R.: Weak convergence and optimal scaling of random walk Metropolis algorithms. The Annals of Applied Probability7(1), 110–120 (1997) https://doi.org/10.1214/aoap/1034625254

    doi:10.1214/aoap/1034625254 1997

  44. [44]

    Bernoulli19(5A), 1501–1534 (2013) https: //doi.org/10.3150/12-BEJ414

    Beskos, A., Pillai, N., Roberts, G., Sanz-Serna, J.M., Stuart, A.: Optimal tuning of the hybrid Monte Carlo algorithm. Bernoulli19(5A), 1501–1534 (2013) https: //doi.org/10.3150/12-BEJ414

    doi:10.3150/12-bej414 2013

  45. [45]

    The Euro- pean Physical Journal Special Topics230(24), 4469–4481 (2021) https://doi.org/ 10.1140/epjs/s11734-021-00287-7 arXiv:2106.15809 [hep-ph]

    Hayato, Y., Pickering, L.: The neut neutrino interaction simulation. The Euro- pean Physical Journal Special Topics230(24), 4469–4481 (2021) https://doi.org/ 10.1140/epjs/s11734-021-00287-7 arXiv:2106.15809 [hep-ph]

    Pith/arXiv arXiv doi:10.1140/epjs/s11734-021-00287-7 2021

  46. [46]

    JINST12(01), 01016 (2017) https://doi.org/10.1088/ 1748-0221/12/01/P01016 arXiv:1612.07393 [hep-ex]

    Stowell, P.,et al.: NUISANCE: a neutrino cross-section generator tuning and comparison framework. JINST12(01), 01016 (2017) https://doi.org/10.1088/ 1748-0221/12/01/P01016 arXiv:1612.07393 [hep-ex]

    Pith/arXiv arXiv 2017

  47. [47]

    Dolan, S., Pickering, L., Stowell, P., Wilkinson, C., Wret, C.: CP-violation or Nuclear Excitation: Reviewing the Role of Neutrino Interaction Model Uncertainties on Accelerator-Based Neutrino Oscillation Measurements (2026) arXiv:2605.28671 [hep-ex]

    Pith/arXiv arXiv 2026

  48. [48]

    Abe, K.,et al.: T2K neutrino flux prediction. Phys. Rev. D87, 012001 (2013) https://doi.org/10.1103/PhysRevD.87.012001 arXiv:1211.0469 [hep-ex] 28

    Pith/arXiv arXiv doi:10.1103/physrevd.87.012001 2013

  49. [49]

    Abgrall, N.,et al.: Time Projection Chambers for the T2K Near Detectors. Nucl. Instrum. Meth. A637, 25–46 (2011) https://doi.org/10.1016/j.nima.2011.02.036 arXiv:1012.0865 [physics.ins-det]

    Pith/arXiv arXiv doi:10.1016/j.nima.2011.02.036 2011

  50. [50]

    Amaudruz, P.A.,et al.: The T2K Fine-Grained Detectors. Nucl. Instrum. Meth. A696, 1–31 (2012) https://doi.org/10.1016/j.nima.2012.08.020 arXiv:1204.3666 [physics.ins-det]

    Pith/arXiv arXiv doi:10.1016/j.nima.2012.08.020 2012

  51. [51]

    In: PHYSTAT 2011, pp

    Conway, J.S.: Incorporating Nuisance Parameters in Likelihoods for Multi- source Spectra. In: PHYSTAT 2011, pp. 115–120 (2011). https://doi.org/10.5170/ CERN-2011-006.115

    2011

  52. [52]

    https://doi.org/10.5281/zenodo.13223262

    GUNDAM Collaboration: GUNDAM: Generalized and Unified Neutrino Data Analysis Methods. https://doi.org/10.5281/zenodo.13223262 . https://github. com/gundam-organization/gundam

    work page doi:10.5281/zenodo.13223262

  53. [53]

    Chakrani, J., Dolan, S., Avanzini, M.B., Ershova, A., Koch, L., McFarland, K., Megias, G.D., Munteanu, L., Pickering, L., Skwarczynski, K., Nguyen, V.Q., Wret, C.: Parametrized uncertainties in the spectral function model of neutrino charged- current quasielastic interactions for oscillation analyses. Phys. Rev. D109, 072006 (2024) https://doi.org/10.1103...

    doi:10.1103/physrevd.109.072006 2024

  54. [54]

    https://arxiv.org/abs/1410.8516

    Dinh, L., Krueger, D., Bengio, Y.: NICE: Non-linear Independent Components Estimation (2015). https://arxiv.org/abs/1410.8516

    Pith/arXiv arXiv 2015

  55. [55]

    https://arxiv.org/abs/1810.01367

    Grathwohl, W., Chen, R.T.Q., Bettencourt, J., Sutskever, I., Duvenaud, D.: FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models (2018). https://arxiv.org/abs/1810.01367

    Pith/arXiv arXiv 2018

  56. [56]

    https://arxiv.org/abs/2210.02747

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow Matching for Generative Modeling (2023). https://arxiv.org/abs/2210.02747

    Pith/arXiv arXiv 2023

  57. [57]

    In: Advances in Neural Informa- tion Processing Systems, vol

    Durkan, C., Bekasov, A., Murray, I., Papamakarios, G.: Neural spline flows. In: Advances in Neural Informa- tion Processing Systems, vol. 32, pp. 7511–7522 (2019). https://proceedings.neurips.cc/paper/2019/file/7ac71d433f282034e088473244df8c02- Paper.pdf

    2019

  58. [58]

    IMA Journal of Numerical Analysis3(3), 141–152 (1983) https: //doi.org/10.1093/imanum/3.3.141

    Delbourgo, R., Gregory, J.A.: Piecewise rational quadratic interpolation to mono- tonic data. IMA Journal of Numerical Analysis3(3), 141–152 (1983) https: //doi.org/10.1093/imanum/3.3.141

    doi:10.1093/imanum/3.3.141 1983

  59. [59]

    Symmetry16(8), 942 (2024) https://doi.org/10.3390/sym16080942 29

    Coccaro, A., Letizia, M., Reyes-Gonz´ alez, H., Torre, R.: Comparison of affine and rational quadratic spline coupling and autoregressive flows through robust statis- tical tests. Symmetry16(8), 942 (2024) https://doi.org/10.3390/sym16080942 29

    doi:10.3390/sym16080942 2024

  60. [60]

    Journal of Open Source Software8(86), 5361 (2023) https://doi.org/10.21105/joss.05361

    Stimper, V., Liu, D., Campbell, A., Berenz, V., Ryll, L., Sch¨ olkopf, B., Hern´ andez- Lobato, J.M.: normflows: A pytorch package for normalizing flows. Journal of Open Source Software8(86), 5361 (2023) https://doi.org/10.21105/joss.05361

    doi:10.21105/joss.05361 2023

  61. [61]

    James, F., Roos, M.: Minuit: A System for Function Minimization and Analysis of the Parameter Errors and Correlations. Comput. Phys. Commun.10, 343–367 (1975) https://doi.org/10.1016/0010-4655(75)90039-9

    doi:10.1016/0010-4655(75)90039-9 1975

  62. [62]

    CoRRabs/1808.03856(2018) 1808.03856

    M¨ uller, T., McWilliams, B., Rousselle, F., Gross, M., Nov´ ak, J.: Neural impor- tance sampling. CoRRabs/1808.03856(2018) 1808.03856

    Pith/arXiv arXiv 2018

  63. [63]

    1–14 (2021) 30

    Elvira, V., Martino, L.: Advances in Importance Sampling, pp. 1–14 (2021) 30

    2021