Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Flow Annealed Importance Sampling Bootstrap meets Differentiable Particle Physics

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper shows that Flow Annealed Importance Sampling Bootstrap, using a differentiable matrix element as target, trains normalizing flows for particle-physics event generation with higher sampling efficiency and fewer target…

desk verdict A solid FAB-for-matrix-elements application with an honest 2D comparison, but the 8D efficiency claim is statistically fragile and needs more seeds. read the letter →

arxiv 2411.16234 v2 pith:CKXIVZTP submitted 2024-11-25 hep-ph cs.LGphysics.comp-phphysics.data-an

Annalena Kofler , Vincent Stimper , Mikhail Mikhasenko , Michael Kagan , Lukas Heinrich This is my paper · ORCID
classification hep-phcs.LGphysics.comp-phphysics.data-an
keywords normalizingflowsannealedimportancesamplingHamiltonianMonteCarloeventgenerationmatrixelementsdifferentiableprogrammingefficiencyreplaybuffer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a training method called Flow Annealed Importance Sampling Bootstrap (FAB) is well suited to generating particle-physics events from matrix elements. FAB evaluates the differentiable target density during training, so it avoids pre-generating a large training dataset, and it uses annealed importance sampling with Hamiltonian Monte Carlo steps to guide flow samples toward the target. The authors compare FAB against reverse-KL and forward-KL flow training on two processes: a two-dimensional charm-baryon decay and an eight-dimensional top-pair production process. They report that FAB with a replay buffer reaches the highest importance sampling efficiency in the eight-dimensional case (90.59% versus 56.51% for reverse-KL) and reaches about 40% efficiency with roughly an order of magnitude fewer matrix-element evaluations. If true, this makes flow-based event generation cheaper at the high-luminosity LHC and for future colliders.

What carries the argument

The load-bearing object is the FAB surrogate loss $S(\theta) = -\mathbb{E}_{\mathrm{AIS}}\left[\bar{w}_{\mathrm{AIS}} \log q_\theta(\bar{x}_{\mathrm{AIS}})\right]$, where $q_\theta$ is a rational-quadratic spline normalizing flow, the target is a differentiable matrix element, and annealed importance sampling with Hamiltonian Monte Carlo transitions interpolates from the flow to $p^2/q_\theta$. The gradient is stopped through the AIS samples, so the flow is trained only through the log-density term. The replay buffer stores the stopped-gradient pairs $(\bar{x}_{\mathrm{AIS}}, \bar{w}_{\mathrm{AIS}})$ and samples them according to their importance weights, allowing several gradient updates per iteration and reducing the number of target evaluations. This machinery is what lets FAB train directly on the target density without pre-generated training data, and the paper argues it is what produces the high sampling efficiency in eight dimensions.

What would settle it

Run FAB with replay buffer on the eight-dimensional $e^+e^-\to t\bar t$ matrix element for many additional random seeds and record both the divergence rate and the distribution of final efficiencies; if a substantial fraction of runs diverge, or the average over surviving runs falls below the reverse-KL comparison, the central claim fails. A cheaper precursor is monitoring the AIS importance-weight variance, which should spike before a run diverges.

Watch

Extended reading notes

Core claim

The central claim is that FAB with a prioritized replay buffer, applied for the first time to matrix elements in high-energy physics, outperforms both reverse-KL and forward-KL normalizing-flow training. On the eight-dimensional $e^+e^-\to t\bar t$ process, the paper reports a final importance sampling efficiency of 90.59% for FAB with replay buffer, compared with 56.51% for rKLD and 84.25% for FAB without the buffer. The paper also finds that FAB with buffer reaches about 40% efficiency with roughly an order of magnitude fewer matrix-element evaluations than rKLD and FAB without buffer. On the two-dimensional $\Lambda_c^+\to pK^-\pi^+$ decay, FAB with buffer reaches 99.56% sampling efficiency, comparable to rKLD and better than fKLD. The authors argue the advantage comes from the annealed importance sampling chain with Hamiltonian Monte Carlo transitions guiding samples toward the target early in training, and from reusing high-weight samples from the replay buffer for multiple gradient updates.

Load-bearing premise

The central claim rests on the assumption that the single FAB-with-replay-buffer run that diverged during the eight-dimensional training was a random accident, not evidence that the method is prone to blowing up in high dimensions.

Editorial extensions

If this is right

  • If correct, FAB with a replay buffer is the most target-evaluation-efficient training method among the density-evaluation-based approaches tested for flow-based event generation.
  • Reaching about 40% efficiency with an order of magnitude fewer target evaluations means expensive matrix-element evaluations can be saved during training on high-dimensional processes.
  • Training directly on the differentiable target removes the need for the large pre-generated datasets required by forward-KL training, at the cost of requiring a differentiable matrix element.
  • The approach is portable to any process with a differentiable matrix-element implementation, and the authors explicitly expect that adding multi-channeling will lead to further performance improvements.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the eight-dimensional result depends on the single diverged FAB-with-replay-buffer run being atypical; if replay-buffer FAB is instead prone to divergences in higher dimensions, the reported 90.59% efficiency is an optimistic summary of the method's behavior.
  • Editorial inference: the order-of-magnitude saving in target evaluations suggests a scaling advantage that should grow with dimensionality, because each Hamiltonian Monte Carlo transition provides gradient information precisely where the flow is a poor approximation; this is untested beyond eight dimensions.
  • Editorial inference: the same replay-buffer-weighted surrogate loss could be applied to other expensive unnormalized targets in physics, such as cross sections with differentiable parton densities, although the paper does not demonstrate this.
  • Editorial inference: since the comparison deliberately excludes physics-aware multi-channeling, the relative gain of FAB in a production event generator may differ from what is reported here once such prior information is included.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper adapts Flow Annealed Importance Sampling Bootstrap (FAB) to high-energy-physics matrix-element sampling. FAB trains a normalizing flow by evaluating the differentiable target density, using AIS with HMC to produce weighted samples and a prioritized replay buffer for multiple gradient updates. The authors compare FAB (with and without the buffer) against reverse-KL (rKLD), forward-KL (fKLD), and VEGAS+ on a 2D Λc+→pK−π+ decay and an 8D e+e−→tt̄ process. They report importance sampling efficiency, test-set forward KL, integral estimates, and efficiency versus number of target evaluations. The paper's central claim is that FAB with a replay buffer reaches higher sampling efficiency and requires fewer target evaluations in high dimensions than the other methods.

Significance. If the high-dimensional result is robust, the paper demonstrates a practical training scheme for flow-based event generation that avoids precomputed training data and exploits differentiable matrix elements. The manuscript is clearly written, includes code and data availability statements, provides a reproducible hyperparameter description, and compares several baselines including VEGAS+. The 2D comparison is clean and shows rKLD and FAB essentially tied. However, the headline 8D claim currently rests on two FAB-with-buffer seeds after excluding one diverged run, and it is evaluated primarily with a metric that is closely aligned with FAB's own loss; these issues make the claimed advantage fragile rather than established.

major comments (3)
  1. [Section 4, first paragraph on e+e−→tt̄] The paper states that 'one of the three training runs of FAB w/ buffer diverged and exclude it from figure 5b and the performance evaluation in table 1.' This self-reported exclusion is load-bearing for the central claim: the 8D FAB-with-buffer efficiency of 90.59±0.01% in Table 1 and the 'order of magnitude fewer target evaluations' statement in Figure 5b are based on the two surviving seeds only. A 1/3 divergence rate is itself evidence about the stability of the replay-buffer mechanism in high dimensions. Please report the diverged run's trajectory, the divergence criterion, whether similar instabilities occur for the other methods, and ideally additional seeds; without this, the mean and uncertainty in Table 1 are conditional on discarding a failure and cannot support the abstract's claim.
  2. [Table 1 and Figure 5b] The reported gap between FAB with buffer (90.59±0.01%, n=2) and rKLD (56.51±40.14%, n=3) is not statistically significant under standard assumptions: the standard error of the difference is approximately 23 percentage points, giving a two-sided p-value of roughly 0.14. The phrase 'significantly higher importance sampling efficiency' therefore overstates the evidence. Please provide per-seed values, report the comparison at matched target-evaluation budgets with uncertainties, and either add seeds or qualify the conclusion accordingly.
  3. [Section 2.1 and Section 3, performance metrics] The primary quantitative metric, importance sampling efficiency ε in Eq. (2), is the normalized effective sample size based on Var_q[w] with w = p(x)/qθ(x). FAB's D_{α=2} loss in Eq. (1) explicitly minimizes the variance of exactly these weights, whereas rKLD and fKLD optimize different objectives. Comparing methods on ε therefore favors FAB by construction. The test-set forward KL in Table 1 and the unweighting efficiency in Table C1 provide some counterbalance, but the paper should make this metric-alignment explicit and justify why ε is the decisive comparison for event generation, or add an evaluation metric that is not directly optimized by FAB.
minor comments (4)
  1. [Section 2.2] There is a typo: 'irrgeular phase space boundaries' should read 'irregular phase space boundaries.'
  2. [Appendix B] The notation 'linit' is used without a definition; it should be introduced as the initial HMC step size.
  3. [Appendix B, FAB hyperparameters] The text says FAB uses 'two (linearly spaced) intermediate distributions M' but Section 2.1 defines the AIS chain as q1,...,q_{M-1}; for M=2 this would be one intermediate distribution. Please clarify the relationship between M and the number of intermediate distributions.
  4. [Appendix B, training settings] The iteration counts for rKLD and FAB appear inconsistent with the claim of equal target evaluations: rKLD uses 3×10^4 iterations in 2D and 10^8 in 8D, while FAB uses 3×10^3 and 10^7, respectively; please clarify how the per-iteration target evaluations balance these numbers.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial metric alignment makes the efficiency claim partly tautological, but the central 'fewer target evaluations' comparison is independent; no load-bearing self-citation chain.

  1. fitted input called prediction [Section 2.1 (FAB training objective) vs Section 3 (Performance Metrics, Eq. 2); abstract]
    "In FAB [23], D_α=2 is chosen as a loss function since it minimizes the variance of the importance weights w = p(x)/q_θ(x)—a desirable property of a well-performing density estimator q_θ(x) [27]. ... The importance sampling efficiency [50, 51] can be computed with samples from the normalizing flow x_i ∼ q_θ(x_i) and their importance weights w_i = p(x_i)/q_θ(x_i) as ε = 1/N [Σ_i w_i]^2 [Σ_i w_i^2]^{-1}."

    With E_q[w] = ∫p dx = const, the expectation of Eq. (2) is ε ≈ (E_q[w])^2/E_q[w^2], so maximizing ε is equivalent to minimizing E_q[w^2]. The FAB loss D_{α=2} is (1/2)E_q[w^2] (up to an additive/overall constant). Hence the flow is trained to minimize exactly the statistic that the reported head-line metric measures. The abstract's claim that 'FAB reaches higher sampling efficiency' is therefore in part a restatement of its own loss, not an independent first-principles result. The 'fewer target evaluations' component (Fig. 5b) is an empirical training-dynamics result and is not circular.

full rationale

The paper is an application/benchmarking study, not a derivation. FAB [23] is adopted from published, code-released prior work (FAB-JAX), and the differentiable matrix elements come from MadJAX/ComPWA; the overlapping-author citations are to public code and are not load-bearing in the sense of an unverified uniqueness claim. No 'uniqueness theorem' is imported. The only concrete reduction is the metric alignment described above: importance-sampling efficiency is a monotone function of E_q[w^2], and D_{α=2} is exactly that expectation. This makes the efficiency ranking partially tautological, but the 'fewer target evaluations' claim and the 8D application are independent empirical contributions. Section 4's exclusion of one diverged FAB-with-buffer run is a statistical robustness limitation (n=2 vs n=3) and not a circularity; it should be weighed under correctness risk, not circularity. Overall score 3 reflects one partial, non-load-bearing metric-alignment issue and no self-citation chain forcing the result.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical and does not rest on a mathematical derivation, so the ledger records tuned hyperparameters and domain assumptions. No new physical entities are introduced. The most consequential items are the HMC step size and the replay-buffer design, which together determine FAB's target-evaluation-count advantage.

free parameters (7)
  • HMC step size l_init = 0.05 (2D), 0.005 (8D)
    Chosen by running one FAB run with an arbitrary step size and observing the adjusted values during training (Appendix B). Load-bearing because HMC acceptance and AIS quality depend on it.
  • Number of intermediate AIS distributions M = 2
    Hand-picked (Appendix B); affects bias and variance of AIS weights and the total number of target evaluations.
  • HMC leapfrog steps per iteration = 3
    Hand-picked (Appendix B); determines how well HMC mixes between intermediate distributions.
  • Gradient updates per iteration L with replay buffer = 4 (2D), 2 (8D)
    Chosen so that buffered FAB and rKLD use the same number of target evaluations (Section 3, Appendix B). This is the mechanism behind the fewer-target-evaluations claim, so the central result depends on this choice.
  • Initial HMC acceptance rate = 0.65
    Target acceptance rate, tuned during training (Appendix B).
  • Flow architecture hyperparameters = nt=nb=10, nn=100 (2D); nt=nb=14, nn=400 (8D)
    Tuned only on rKLD runs, then reused for all methods (Appendix B); chosen based on final loss, validation loss, and efficiency.
  • Replay buffer size range = 10^4 to 10^5
    Set to 10 times and 100 times the batch size, not optimized (Appendix B).
assumptions (5)
  • standard math AIS with HMC transition operator yields asymptotically unbiased importance weights and samples for the current flow.
    Invoked in Section 2.1 for FAB; taken from [23, 28].
  • domain assumption The matrix element amplitudes implemented in ComPWA and MadJAX evaluate the unnormalized density and its gradients correctly.
    Section 2.2 relies on these differentiable implementations; errors would bias all target-evaluation-based methods.
  • domain assumption The transformations mapping physical phase space to the unit hypercube (square Dalitz, RAMBO) are invertible and account for the correct Jacobian, so a normalizing flow on the hypercube targets the true physical distribution.
    Appendix A describes the mapping but does not explicitly derive the Jacobian; the flow density is trained on transformed coordinates.
  • domain assumption Importance sampling efficiency computed on fresh flow samples is a valid proxy for event-generation quality.
    Section 3 defines epsilon from equation (2); unweighting efficiency is related but not identical (Appendix C).
  • domain assumption Normalizing flows with rational-quadratic splines can represent the target matrix element densities on the unit hypercube.
    Section 2.2 and Appendix B; expressiveness is assumed sufficient for both matrix elements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flow Annealed Importance Sampling Bootstrap meets Differentiable Particle Physics." pith.science (2026). https://pith.science/paper/CKXIVZTP

@misc{pith2026241116234,
  author       = {Pith},
  title        = {Pith review of: Flow Annealed Importance Sampling Bootstrap meets Differentiable Particle Physics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKXIVZTP}},
  note         = {Machine review of arXiv:2411.16234}
}
read the original abstract

High-energy physics requires the generation of large numbers of simulated data samples from complex but analytically tractable distributions called matrix elements. Surrogate models, such as normalizing flows, are gaining popularity for this task due to their computational efficiency. We adopt an approach based on Flow Annealed importance sampling Bootstrap (FAB) that evaluates the differentiable target density during training and helps avoid the costly generation of training data in advance. We show that FAB reaches higher sampling efficiency with fewer target evaluations in high dimensions in comparison to other methods.

Figures

Figures reproduced from arXiv: 2411.16234 by the authors.

Figure 1
Figure 1. Visualization of compared methods. The black arro [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The Λ + c → pK−π + decay (left-most diagram) consists of three decay channels characterized by their resonances Λ0 (center-left), ∆++ (center-right), and K∗ (right), each visualized in Feynman-like diagrams. an intermediate particle that can have resonances which are shown in a Dalitz plot in figure 3. A Dalitz plot is a physics-specific visualization of 2D amplitudes where the axes are chosen such that the histogra… view at source ↗
Figure 3
Figure 3. Comparison of the target density for the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Normalized marginal histograms of the invariant m [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Importance sampling efficiency depending on the num [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Corner plot with samples from MadGraph and from the best normalizing flows for each method. need to apply the inverse RAMBO transformation [48] to obtain physical information for each outgoing particle. We observe that the normalizing flows especially deviate from the …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Communicating Likelihoods with Normalising Flows

    hep-ph 2025-02 conditional novelty 4.0 of 10

    A normalizing-flow workflow compresses sample-based likelihoods into small files, validated with a radial Kolmogorov-Smirnov test on three high-energy physics examples.

Reference graph

Works this paper leans on

65 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    2020 High-Luminosity Large Hadron Collider (HL-LHC): Technica l design report CERN Yellow Reports: Monographs (Geneva: CERN) URL https:// cds.cern.ch/record/2749422

    Aberle O et al. 2020 High-Luminosity Large Hadron Collider (HL-LHC): Technica l design report CERN Yellow Reports: Monographs (Geneva: CERN) URL https:// cds.cern.ch/record/2749422

    arXiv 2020

  2. [2]

    Software C O and Computing 2022 CMS phase-2 computing mod el: Update document Tech. rep. CERN Geneva URL https://cds.cern.ch/record/2815292

    arXiv 2022

  3. [3]

    Collaboration A (ATLAS) 2022 ATLAS software and computi ng HL-LHC roadmap Tech. rep. CERN Geneva URL https://cds.cern.ch/record/2802918

    arXiv 2022

  4. [4]

    REFERENCES 20 1002/9783527618460.ch6) URL https://onlinelibrary.wiley.com/doi/abs/ 10.1002/9783527618460.ch6

    Griffiths D 1987 The Feynman Calculus (John Wiley & Sons, Ltd) chap 6, pp 189–212 ( Preprint https://onlinelibrary.wiley.com/doi/pdf/10. REFERENCES 20 1002/9783527618460.ch6) URL https://onlinelibrary.wiley.com/doi/abs/ 10.1002/9783527618460.ch6

    work page doi:10.1002/9783527618460.ch6 1987

  5. [5]

    Kleiss R and Pittau R 1994 Comput. Phys. Commun. 83 141–146 ( Preprint hep-ph/ 9405257)

    work page 1994

  6. [6]

    Alwall J, Herquet M, Maltoni F, Mattelaer O and Stelzer T 2 011 Journal of High Energy Physics 2011 128 URL https://doi.org/10.1007/JHEP06(2011)128

    doi:10.1007/jhep06(2011)128 2011

  7. [8]

    (Sherpa) 2019 SciPost Phys

    Bothmann E et al. (Sherpa) 2019 SciPost Phys. 7 034 ( Preprint 1905.09127) URL https://scipost.org/10.21468/SciPostPhys.7.3.034

    arXiv 2019

  8. [9]

    2022 SciPost Phys

    Bierlich C et al. 2022 SciPost Phys. Codebases 8 URL https://scipost.org/10. 21468/SciPostPhysCodeb.8

    work page 2022

Show all 65 references
  1. [10]

    2022 SciPost Phys

    Bierlich C et al. 2022 SciPost Phys. Codebases 8–r8.3 URL https://scipost.org/ 10.21468/SciPostPhysCodeb.8-r8.3

    2022 doi

  2. [11]

    Lepage P G 1978 Journal of Computational Physics 27 192–203 URL https:// www.sciencedirect.com/science/article/pii/0021999178900049

    1978

  3. [12]

    10 038 URL https://scipost.org/ 10.21468/SciPostPhys.10.2.038

    Stienen B and Verheyen R 2021 SciPost Phys. 10 038 URL https://scipost.org/ 10.21468/SciPostPhys.10.2.038

    2021 doi

  4. [13]

    Müller T, Mcwilliams B, Rousselle F, Gross M and Novák J 2 019 ACM Trans. Graph. 38 URL https://doi.org/10.1145/3341156

    doi

  5. [14]

    8 069 URL https://scipost.org/10.21468/SciPostPhys.8.4.069

    Bothmann E, Janßen T, Knobbe M, Schmale T and Schumann S 2 020 SciPost Phys. 8 069 URL https://scipost.org/10.21468/SciPostPhys.8.4.069

    doi

  6. [15]

    Gao C, Isaacson J and Krause C 2020 Machine Learning: Science and Technology 1 045023 URL https://dx.doi.org/10.1088/2632-2153/abab62

    2020 doi

  7. [16]

    Gao C, Höche S, Isaacson J, Krause C and Schulz H 2020 Phys. Rev. D 101(7) 076002 URL https://link.aps.org/doi/10.1103/PhysRevD.101.076002

    2020 doi

  8. [17]

    15 141 URL https://scipost.org/10.21468/ SciPostPhys.15.4.141

    Heimel T, Winterhalder R, Butter A, Isaacson J, Krause C , Maltoni F, Mattelaer O and Plehn T 2023 SciPost Phys. 15 141 URL https://scipost.org/10.21468/ SciPostPhys.15.4.141

    2023

  9. [18]

    17 023 URL https://scipost.org/10.21468/SciPostPhys

    Heimel T, Huetsch N, Maltoni F, Mattelaer O, Plehn T and W interhalder R 2024 SciPost Phys. 17 023 URL https://scipost.org/10.21468/SciPostPhys. 17.1.023

    2024 doi

  10. [19]

    Pina-Otey S, Sánchez F, Lux T and Gaitan V 2020 Phys. Rev. D 102(1) 013003 URL https://link.aps.org/doi/10.1103/PhysRevD.102.013003

    2020 doi

  11. [20]

    Deutschmann N and Götz N 2024 Journal of High Energy Physics 2024 83 URL https://doi.org/10.1007/JHEP03(2024)083

    2024 doi

  12. [21]

    Heinrich L and Kagan M 2023 J. Phys. Conf. Ser. 2438 012137 ( Preprint 2203. 00057) URL https://dx.doi.org/10.1088/1742-6596/2438/1/012137 REFERENCES 21

    2023 doi

  13. [22]

    18 017 URL https://scipost.org/10.21468/SciPostPhys.18.1.017

    Heimel T, Mattelaer O, Plehn T and Winterhalder R 2025 SciPost Phys. 18 017 URL https://scipost.org/10.21468/SciPostPhys.18.1.017

    2025 doi

  14. [23]

    Midgley L I, Stimper V, N C Simm G, Schölkopf B and M Hernan dez-Lobato J 2023 Flow annealed importance sampling bootstrap The Eleventh International Conference on Learning Representations URL https://openreview.net/forum? id=XCTVFJwS9LJ

    2023

  15. [24]

    Papamakarios G, Nalisnick E, J Rezende D, Mohamed S and L akshminarayanan B 2021 Journal of Machine Learning Research 22 1–64 URL http://jmlr.org/ papers/v22/19-1028.html

    2021

  16. [25]

    G Tabak E and Vanden-Eijnden E 2010 Communications in Mathematical Sciences 8 217 – 233 URL https://dx.doi.org/10.4310/CMS.2010.v8.n1.a11

    2010 doi

  17. [26]

    Rezende D and Mohamed S 2015 Variational inference with normalizing flows Proceedings of the 32nd International Conference on Machin e Learning (Proceedings of Machine Learning Research vol 37) pp 1530–1538 URL https:// proceedings.mlr.press/v37/rezende15.html

    2015

  18. [27]

    com/en-us/research/wp-content/uploads/2016/02/tr-20 05-173.pdf

    Minka T 2005 Technical report, Microsoft Research URL https://www.microsoft. com/en-us/research/wp-content/uploads/2016/02/tr-20 05-173.pdf

    2005

  19. [28]

    1023/A:1008923215028

    Neal R M 2001 Statistics and Computing 11 125–139 URL https://doi.org/10. 1023/A:1008923215028

    2001

  20. [29]

    Michel M, de Boer R and Pflueger S 2023 comPW A: Common part ial wave analysis - making amplitude analysis transparant, understandable, and easy to start with URL https://github.com/ComPWA

    2023

  21. [30]

    (COMPASS Collaboration) 2017 Phys

    Adolph C et al. (COMPASS Collaboration) 2017 Phys. Rev. D 95(3) 032004 URL https://link.aps.org/doi/10.1103/PhysRevD.95.032004

    2017 doi

  22. [31]

    Kaspar F M and Gerassimov S (COMPASS) 2022 Rev. Mex. Fis. Suppl. 3 0308020

    2022

  23. [32]

    High Energy Phys

    Marangotto D 2020 Adv. High Energy Phys. 2020 7463073 ( Preprint 2004.12318) URL https://onlinelibrary.wiley.com/doi/abs/10.1155/2020/7463073

    2020 arXiv

  24. [33]

    (LHCb Collaboration) 2023 Phys

    Aaij R et al. (LHCb Collaboration) 2023 Phys. Rev. D 108(1) 012023 URL https://link.aps.org/doi/10.1103/PhysRevD.108.012023

    2023 doi

  25. [34]

    (LHCb Collaboration) 2023 JHEP 07 228 ( Preprint 2301.07010)

    Aaij R et al. (LHCb Collaboration) 2023 JHEP 07 228 ( Preprint 2301.07010)

    2023 arXiv

  26. [35]

    Penalva N, Hernández E and Nieves J 2019 Phys. Rev. D 100(11) 113007 URL https://link.aps.org/doi/10.1103/PhysRevD.100.113007

    2019 doi

  27. [36]

    Hu Q Y, Li X Q, Yang Y D and Zheng D H 2021 Journal of High Energy Physics 2021 183 URL https://doi.org/10.1007/JHEP02(2021)183

    2021 doi

  28. [37]

    Byckling E and Kajantie K 1973 Particle Kinematics (Wiley)

    1973

  29. [38]

    2017 PeerJ Computer Science 3 e103 URL https://doi.org/10

    Meurer A et al. 2017 PeerJ Computer Science 3 e103 URL https://doi.org/10. 7717/peerj-cs.103

    2017

  30. [39]

    2018 JAX: composable transformations of Python+NumPy programs URL http://github.com/google/jax REFERENCES 22

    Bradbury J et al. 2018 JAX: composable transformations of Python+NumPy programs URL http://github.com/google/jax REFERENCES 22

    2018

  31. [40]

    Alwall J, Frederix R, Frixione S, Hirschi V, Maltoni F, M attelaer O, Shao H S, Stelzer T, Torrielli P and Zaro M 2014 JHEP 07 079 ( Preprint 1405.0301) URL https://doi.org/10.1007/JHEP07(2014)079

    2014 arXiv

  32. [41]

    (D0) 2010 Phys

    Abazov V M et al. (D0) 2010 Phys. Rev. D 82 071102 ( Preprint 1008.4284)

    2010 arXiv

  33. [42]

    (CMS) 2011 Phys

    Chatrchyan S et al. (CMS) 2011 Phys. Rev. D 84 092004 ( Preprint 1108.3773)

    2011 arXiv

  34. [43]

    (ATLAS) 2023 JHEP 06 138 ( Preprint 2207.01354)

    Aad G et al. (ATLAS) 2023 JHEP 06 138 ( Preprint 2207.01354)

    2023 arXiv

  35. [44]

    (ATLAS) 2023 JHEP 07 141 ( Preprint 2303.15340)

    Aad G et al. (ATLAS) 2023 JHEP 07 141 ( Preprint 2303.15340)

    2023 arXiv

  36. [45]

    André K D J, Benedikt M, Oide K and Zimmermann F (FCC-ee) 2 025 PoS ICHEP2024 830

  37. [46]

    (Linear Collider Vision) 2025 ( Preprint 2503.19983)

    Balazs C et al. (Linear Collider Vision) 2025 ( Preprint 2503.19983)

    2025

  38. [47]

    Carrazza S, Cruz-Martinez J M and Rossi M 2021 Computer Physics Communications 264 107995 URL https://www.sciencedirect.com/science/ article/pii/S0010465521001077

    2021

  39. [48]

    Plätzer S 2013 arXiv e-prints arXiv:1308.2922 (Preprint 1308.2922) URL https:// arxiv.org/abs/1308.2922

    2013 arXiv

  40. [49]

    Lepage P G 2021 Journal of Computational Physics 439 110386 URL https:// www.sciencedirect.com/science/article/pii/S0021999121002813

    2021

  41. [50]

    Martino L, Elvira V and Louzada F 2017 Signal Processing 131 386–401 URL https://www.sciencedirect.com/science/article/pii/S0165168416302110

    2017

  42. [51]

    Elvira V, Martino L and Robert C P 2022 International Statistical Review 90 525–550 URL https://onlinelibrary.wiley.com/doi/abs/10.1111/insr.12500

    2022 doi

  43. [52]

    Flow Annealed Importance Sampling Bootstrap meets Differentiab le Particle Physics

    Kofler A, Stimper V, Mikhasenko M, Kagan M and Heinrich L 2 024 Data for “Flow Annealed Importance Sampling Bootstrap meets Differentiab le Particle Physics” URL https://doi.org/10.17617/3.UZ786R

    doi

  44. [53]

    Cabezas A, Corenflos A, Lao J and Louf R 2024 BlackJAX: Com posable Bayesian inference in JAX ( Preprint 2402.10797)

    2024 arXiv

  45. [54]

    Foreman-Mackey D 2016 The Journal of Open Source Software 1 24 URL https:// doi.org/10.21105/joss.00024

    2016 doi

  46. [55]

    2020 The DeepMind JAX Ecosystem URL http://github.com/ deepmind

    Babuschkin I et al. 2020 The DeepMind JAX Ecosystem URL http://github.com/ deepmind

    2020

  47. [56]

    Hennigan T, Cai T, Norman T, Martens L and Babuschkin T 20 20 Haiku: Sonnet for JAX URL http://github.com/deepmind/dm-haiku

  48. [57]

    Hunter J D 2007 Computing in Science & Engineering 9 90–95

    2007

  49. [58]

    2020 Nature 585 357–362 URL https://doi.org/10.1038/ s41586-020-2649-2

    Harris C R et al. 2020 Nature 585 357–362 URL https://doi.org/10.1038/ s41586-020-2649-2

    2020

  50. [59]

    Heinrich L, Feickert M, Rodrigues E and P Neuwirth A 2024 pylhe: v0.9.0 URL https://github.com/scikit-hep/pylhe REFERENCES 23

    2024

  51. [60]

    Fritsch M, Pflüger S, de Boer R E, Gradl W and Peters K 2024 C om- PW A/tensorwaves: Python fitter package for multiple comput ational back-ends URL https://github.com/ComPWA/tensorwaves

    2024

  52. [61]

    Petroff M A 2021 arXiv preprints (Preprint 2107.02270)

    2021 arXiv

  53. [62]

    Durkan C, Bekasov A, Murray I and Papamakarios G 2019 Neu ral spline flows Advances in Neural Information Processing Systems vol 32 ed Wallach H, Larochelle H, Beygelzimer A, d 'Alché-Buc F, Fox E and Garnett R (Curran Associates, Inc.) URL https://proceedings.neurips.cc/paper/...

    2019

  54. [63]

    Kingma D and Ba J 2015 Adam: A method for stochastic optim ization International Conference on Learning Representations (IC LR) (San Diega, CA, USA)

    2015

  55. [64]

    9 053 URL https://scipost

    Klimek M D and Perelstein M 2020 SciPost Phys. 9 053 URL https://scipost. org/10.21468/SciPostPhys.9.4.053

    2020 doi

  56. [65]

    12 164 URL https://scipost.org/10.21468/SciPostPhys.12.5.164

    Danziger K, Janßen T, Schumann S and Siegert F 2022 SciPost Phys. 12 164 URL https://scipost.org/10.21468/SciPostPhys.12.5.164

    2022 doi

  57. [66]

    10 089 URL https://scipost.org/10.21468/SciPostPhys.10.4.089

    Backes M, Butter A, Plehn T and Winterhalder R 2021 SciPost Phys. 10 089 URL https://scipost.org/10.21468/SciPostPhys.10.4.089

    2021 doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.