Pith. sign in

REVIEW 3 major objections 4 minor 74 references

Enhancing anomaly detection with topology-aware autoencoders

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that an autoencoder whose latent space has the same global shape as the background phase-space manifold removes spurious reconstruction errors and improves anomaly detection.

desk verdict Useful latent-space constructions and clean toy tests, but the realistic AUC claim is not isolated: S2⊗S2 matches S2⊗RP2, so the 'homeomorphic latent space' conclusion overreaches. read the letter →

arxiv 2502.10163 v1 pith:TMC22QDL submitted 2025-02-14 hep-ph cs.LGhep-ex

Vishal S. Ngairangbam , B{l}ażej Rozwoda , Kazuki Sakurai , Michael Spannowsky This is my paper
classification hep-phcs.LGhep-ex
keywords topology-awareautoencodersanomalydetectionlatentspacetopologyphase-spacemanifoldsrealprojectiveplanetopquarkdecaysunsupervisednewphysicssearchesloss-versus-distance
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 is trying to establish that the global shape, or topology, of an autoencoder's bottleneck is not a neutral choice: if the background data naturally lives on a curved or compact manifold, a flat Euclidean latent space of the same dimension cannot represent it without introducing spurious reconstruction errors. The authors build autoencoders whose latent layers explicitly realize such manifolds, namely spheres, products of spheres, and the real projective plane, and test them on toy data and simulated hadronic top-quark decays. They report that a latent space matching the background's $S^2 \otimes \mathbb{RP}^2$ structure improves anomaly detection over the best flat four-dimensional latent space, raising the AUC from about 0.725 to about 0.732 on reconstructed events. The broader point matters because most autoencoders used in collider searches assume a Euclidean bottleneck, and the paper argues that this assumption quietly caps their sensitivity.

What carries the argument

The central object is a constricted latent layer built from an explicit embedding of the data manifold. For a sphere $S^n$, the preparation-layer values $y_i$ are normalized by their radius, $z_i = y_i/r$; for a product $S^{n_1} \otimes S^{n_2}$, two such normalizations are applied to separate node blocks; for the real projective plane $\mathbb{RP}^2$, four neurons realize the map $(z_1,z_2,z_3,z_4) = (\tilde y_1^2 - \tilde y_2^2, \tilde y_1 \tilde y_2, \tilde y_2 \tilde y_3, \tilde y_3 \tilde y_1)$ on a normalized triple $\tilde y$. This forces the encoder output to live on the chosen manifold, giving the decoder a global chart of the background and eliminating the spurious-error regions that a hard-dimensional Euclidean bottleneck produces when it cannot globally fit a non-trivial manifold.

What would settle it

Train the same six autoencoders on detector-level events from the same top-pair process using public experimental data, with the same selection cuts, and compare AUCs: if the $S^2 \otimes \mathbb{RP}^2$ bottleneck does not beat the $\mathbb{R}^4$ bottleneck, or if the loss-versus-distance curve of the $\mathbb{R}^4$ autoencoder shows no localized high-error peak, the topological-obstruction mechanism is not the operative cause of the improvement.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a latent space homeomorphic to the background phase-space manifold removes topological obstructions that a Euclidean latent space of equal dimension cannot avoid. When data lie on, for example, $S^2 \otimes \mathbb{RP}^2$, an $\mathbb{R}^4$ bottleneck must tear or puncture the manifold somewhere, producing localized regions of high reconstruction error that look like anomalies. By constructing the bottleneck directly as the manifold's embedding, the autoencoder achieves faithful global reconstruction and reserves high error for genuinely out-of-distribution events. In the hadronic top-decay benchmark, the $S^2 \otimes \mathbb{RP}^2$ autoencoder achieves AUC $0.7321 \pm 0.0037$ versus $0.7246 \pm 0.0097$ for the $\mathbb{R}^4$ autoencoder, while over-large flat spaces such as $\mathbb{R}^5$ and $S^5$ reconstruct the signal too well and drop to about $0.55$.

Load-bearing premise

The argument relies on the reconstructed background data still being well described by the idealized manifold $S^2 \otimes \mathbb{RP}^2$ after parton showering, hadronization, jet clustering, and the top-candidate selection; if finite particle widths, detector smearing, or jet misassignment change the effective topology, the matched latent prior is no longer describing the data and the small AUC advantage over a flat $\mathbb{R}^4$ bottleneck could disappear.

Editorial extensions

If this is right

  • An autoencoder trained on a background whose phase-space topology is known should use a bottleneck of the same topology and intrinsic dimension, not merely the same dimension.
  • Increasing the latent dimension beyond the background's intrinsic dimension is actively harmful: in the reconstructed top events, $\mathbb{R}^5$ and $S^5$ latent spaces give AUCs near $0.55$.
  • A low validation loss does not indicate good anomaly detection, because a larger latent space can reconstruct both background and signal well.
  • The topology-matched $S^2 \otimes \mathbb{RP}^2$ bottleneck outperforms the flat $\mathbb{R}^4$ bottleneck on reconstructed events, with AUC $0.7321 \pm 0.0037$ versus $0.7246 \pm 0.0097$.
  • On the toy $S^2 \otimes S^2$ dataset, a shallow topology-aware network with about one thousand parameters reconstructs the manifold better than a deep $\mathbb{R}^4$ autoencoder with 1.4 million parameters.

Reading between the lines

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

  • If the mechanism holds, the same construction can be applied to any final state with known phase-space topology, such as identical-particle pairs producing $\mathbb{RP}^2$ factors or multi-body decays producing higher spheres.
  • A natural next test is detector-level public data with the identical event selection; the reported advantage might grow if the signal manifold's topology differs more sharply from the background, or shrink if finite widths thicken the manifold so that the idealized topology is no longer a good description.
  • Because the paper's $\mathbb{RP}^2$ embedding makes the latent space antipodally identified but does not explicitly symmetrize the decoder under jet exchange, a fully symmetric decoder is a structural extension that could push the AUC further.
  • A sharper falsifiable consequence of the paper's picture is that the loss-versus-distance diagnostic for an $\mathbb{R}^4$ autoencoder on reconstructed top data should show a localized high-loss puncture, while the topology-matched autoencoder should not; measuring this directly on data would isolate the claimed mechanism.
Share X Bluesky LinkedIn Reddit HN

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 proposes autoencoders whose latent spaces are constrained to be homeomorphic to the assumed phase-space manifold of the background process, with the goal of improving unsupervised anomaly detection at colliders. The authors construct latent embeddings for S^n, products of spheres, and RP^2 via explicit normalization or polynomial maps in the bottleneck layer. Toy experiments on S^2 and S^2⊗S^2 data show that flat latent spaces of the same intrinsic dimension (R^2 and R^4) leave topological obstruction regions with high reconstruction error, while the corresponding compact latent spaces fit the data globally. They then apply the method to simulated hadronic top-quark decays, comparing R^3, R^4, R^5, S^2⊗S^2, S^2⊗RP^2, and S^5 latent topologies, and report AUC values for distinguishing the three-body signal from the two-body background. The central claim is that a latent topology homeomorphic to the background manifold improves anomaly detection by removing spurious reconstruction errors.

Significance. If the central claim were established, the paper would provide a principled way to choose latent geometry from first-principles physics and would be of considerable interest for unsupervised new-physics searches. The toy demonstrations are clean and directly exhibit the topological-obstruction mechanism, and the paper is generally well written with explicit instructions for constructing the non-trivial latent embeddings. The authors also correctly emphasize the distinction between existence of local charts and global embeddings. However, the collider experiment does not actually isolate the effect of homeomorphism: the S^2⊗S^2 control performs identically to S^2⊗RP^2, and the preprocessing and network construction together undermine the claim that the RP^2 quotient is realized. As a result, the central assertion is currently not supported by the presented evidence, although the underlying mechanism is plausible and the issue is addressable with additional controls.

major comments (3)
  1. [§V.A and §V.B] The input features for the reconstructed data are the three-momenta of the decay products in the top rest frame, and the paper states that 'the untagged jets are ordered by transverse momentum in the lab frame for reconstructed data.' This pT ordering makes the two W-decay jets distinguishable by construction, so the background data manifold after preprocessing is S^2⊗S^2, not S^2⊗RP^2. The claim in §V.B that 'the background manifold is S²⊗RP²' is therefore not applicable to the actual features fed into the autoencoders. This is not a minor technicality: it explains why S^2⊗S^2-AE and S^2⊗RP^2-AE perform identically in Tables I and II, and it means the S^2⊗RP^2 latent is not in fact homeomorphic to the data manifold being fitted.
  2. [§V.C] The paper admits that the network 'does not explicitly respect the discrete symmetry of the identification (pa,pb) ∼ (pb,pa)' and that the dataset contains equal amounts of both orderings. Because the RP^2 embedding in §III identifies antipodal points on the preparation sphere but the subsequent layers do not enforce the quotient, the effective latent space realized by S^2⊗RP^2-AE is the double cover S^2⊗S^2. The observation that S^2⊗S^2-AE and S^2⊗RP^2-AE give identical AUCs (0.9466±0.0002 each in Table I and within 0.001 in Table II) is exactly what this construction predicts. Consequently, the comparison cannot distinguish 'latent homeomorphic to the background' from 'latent being a double cover of the background, or any other compact 4-manifold.' To support the central claim, the authors should either enforce the RP^2 quotient symmetry (for example by symmetrizing the reconstruction loss over the two jets or using a canonical ordering consistent with the claimed manifold) or include a compact 4D control topology that is not homeomorphic to S^2⊗RP^2 (for example T^4) and show that performance degrades.
  3. [Tables I and II] The reported precision does not establish a significant AUC improvement of the topological latent spaces over the flat R^4 baseline. In Table II, R^4 achieves 0.7246±0.0097 while S^2⊗RP^2 achieves 0.7321±0.0037; the difference is 0.0075, which is smaller than the R^4 standard deviation. With ten training instances per topology, a paired test (for example a paired t-test or a Wilcoxon signed-rank test) should be reported for the pairwise comparisons. In the absence of such a test, the abstract's statement that topology-aware autoencoders 'significantly improve anomaly separation' is not supported by the numbers. The large differences relative to R^3 and R^5 are not evidence for the homeomorphic-prior claim, since those baselines have the wrong intrinsic dimension.
minor comments (4)
  1. [§II.B] The text says the momentum data manifold is 'holomorphic to S^{3n−4}'; the intended word is clearly 'homeomorphic'.
  2. [§IV.B] There is a typo in 'with 0c being the zero-vector in∈ R3c'; the 'in∈' should read 'in' or '∈'.
  3. [§IV] In the toy experiments the S^2-AE and S^2⊗S^2-AE use a much smaller network (a single hidden layer of 64 units, around 1k parameters) than the R^k-AEs (deep, 1.4M parameters). While the later collider experiments control architecture, the visual comparisons in Figs. 2–4 are not architecture-controlled, so the apparent success of the topological latent spaces could partly reflect the explicit normalization rather than the topology alone.
  4. [§V.C] The sentence 'with the non-trivial four-dimensional topologies having comparatively better anomaly detection performance for either datasets' is grammatically awkward and should be revised.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the latent topologies are physics-derived inputs, the AUC comparisons are empirical, and the self-citations are not load-bearing.

full rationale

The derivation chain is not circular. The latent topologies are stipulated from physical prior knowledge, not read off from the anomaly-detection score: Section II.B derives S3n-4, products, and RP2 from energy-momentum conservation and identical-particle symmetry, and Section V.B applies this to the top decay (S2⊗S2, then S2⊗RP2 for indistinguishable W jets). These choices are fixed before the autoencoders are trained; none of the AUC numbers in Tables I-II is used to pick or tune the latent manifold. All six architectures share the same training protocol and evaluation (Section V.B), and the matched topology S2⊗RP2 is not even the best in every case: on parton-level data it exactly ties S2⊗S2 (0.9466±0.0002 each), and on reconstructed data the nominal advantage over R4 (0.7321 vs 0.7246) is not a fit renamed as a prediction. The paper's own admission that the network 'does not explicitly respect the discrete symmetry of the identification (pa, pb) ∼ (pb, pa)' shows the RP2 quotient was not hard-coded into the architecture. The cited topological-obstruction statement is Ref. [47], an external paper (Batson et al.), and the self-citations [56, 58, 60] are related-work references rather than load-bearing premises or uniqueness theorems. The main weakness—that S2⊗S2 is not homeomorphic to S2⊗RP2 yet performs identically, so the experiment does not fully isolate 'homeomorphic matching'—is a limitation of the control/comparison, not a case of the conclusion being equivalent to its input by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on five assumptions, all flagged in the text: (1) the phase-space topology theorems imported from [47], (2) the S^2 and RP^2 kinematics for two-body and identical-particle decays, (3) the zero-width, on-shell idealization that makes the reconstructed background manifold exactly 4-dimensional, (4) the empirical realizability claim that trained autoencoders actually learn the global homeomorphism, and (5) the restriction to out-of-distribution signals with dimension greater than the background. No parameters are fitted to the anomaly-detection outcome; the only hand-chosen numbers are toy-embedding rotation angles and standard, shared training hyperparameters. No new physical entities are introduced: the latent manifolds (S^n, products, RP^2) are standard mathematical spaces with independent mathematical standing, and the RP^2 latent layer uses the classical Veronese embedding into R^4.

free parameters (2)
  • Givens rotation angles for the S^2 × S^2 toy embedding = θ1=2.6994, θ2=2.3480, θ3=3.0390, θ4=3.9448, θ5=4.3052, θ6=0.8107
    Hand-chosen fixed angles in Section IVB used to rotate the S^2 × S^2 toy dataset non-trivially inside R^9. They are arbitrary constants, not fitted to any target result, and do not affect the comparison between latent topologies.
  • Training hyperparameters (architecture 1024-512-256-128-64, batch 256, learning rate 0.001, patience 20)
    Chosen by hand and identical across all six compared autoencoders (Sections IV and VB), so they do not bias the topology comparison. Listed for completeness because they are part of the method that must be re-implemented.
assumptions (5)
  • domain assumption The phase space of n distinguishable particles with fixed total energy and momentum is homeomorphic to S^(3n-4)
    Stated in Section IIB with a citation to [47] and not re-derived here. It is used to identify the anomalous signal manifold as S^5 for the three-body top decay in Section V. The statement is imported from the cited literature rather than proven in this paper.
  • domain assumption A two-body decay at rest with fixed masses has an S^2 phase space; sequential two-body decays give products S^2 × S^2; a pair of identical particles gives RP^2
    Section IIB derives the background manifold of the hadronic top decay from these kinematics statements: t -> bW gives one S^2, W -> jj gives the other, and identical jets replace one sphere by RP^2. These are standard phase-space facts, asserted without proof.
  • domain assumption The background data manifold is exactly S^2 × S^2 (or S^2 × RP^2) in the zero-width, on-shell, massless-parton limit
    Section VB: 'Assuming distinguishable decay products and zero width decays in the top's rest frame, the background data manifold is S^2 ⊗ S^2'. This idealization is load-bearing for the choice of latent topology and is only approximately true after parton showering, hadronization, and reconstruction, a limitation the paper acknowledges in Section VC.
  • domain assumption A densely connected autoencoder can realize a global homeomorphism between the data manifold and a latent manifold of matching topology
    Section II argues from the universal approximation property that the encoder and decoder can implement the global maps, and Section III states the toy experiments show the conditions are 'enough'. The claim is demonstrated empirically on the toy datasets, not proven; the footnote in Section II notes the hypercube-domain caveat of the universal approximation theorems.
  • domain assumption The anomalous signal lives outside the background manifold and has intrinsic dimension greater than the background
    Section II explicitly restricts attention to signal manifolds that are 'out-of-distribution' and not contained in the background manifold, and Section V chooses a three-body decay signal with a 5-dimensional manifold against a 4-dimensional background. The benchmark therefore does not test submanifold-type signals, for which the paper concedes autoencoders would reconstruct the signal well.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing anomaly detection with topology-aware autoencoders." pith.science (2026). https://pith.science/paper/TMC22QDL

@misc{pith2026250210163,
  author       = {Pith},
  title        = {Pith review of: Enhancing anomaly detection with topology-aware autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TMC22QDL}},
  note         = {Machine review of arXiv:2502.10163}
}
abstract

Anomaly detection in high-energy physics is essential for identifying new physics beyond the Standard Model. Autoencoders provide a signal-agnostic approach but are limited by the topology of their latent space. This work explores topology-aware autoencoders, embedding phase-space distributions onto compact manifolds that reflect energy-momentum conservation. We construct autoencoders with spherical ($S^n$), product ($S^2 \otimes S^2$), and projective ($\mathbb{RP}^2$) latent spaces and compare their anomaly detection performance against conventional Euclidean embeddings. Our results show that autoencoders with topological priors significantly improve anomaly separation by preserving the global structure of the data manifold and reducing spurious reconstruction errors. Applying our approach to simulated hadronic top-quark decays, we show that latent spaces with appropriate topological constraints enhance sensitivity and robustness in detecting anomalous events. This study establishes topology-aware autoencoders as a powerful tool for unsupervised searches for new physics in particle-collision data.

Figures

Figures reproduced from arXiv: 2502.10163 by the authors.

Figure 1
Figure 1. Successful and unsuccessful global fits of data with different latent manifolds. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Input–output pairs for ten samples with highest reconstruction error for different latent spaces. The [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Loss-versus-distance plots for three latent layers. For each sample [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Input - output pairs for ten samples with highest reconstruction error for different latent spaces. The [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Invariant mass distribution of the W+ (mjj ) and top (mbjj ) decay products after reconstruction and baseline selection (solid lines) and the true distribution at parton level (dotted lines). The network analyses are carried out using PyTorch (v2.0.0) [73] on a single …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 24 canonical work pages

  1. [1]

    Belis, P

    V. Belis, P. Odagiu, and T. K. Aarrestad, Rev. Phys.12, 100091 (2024), 2312.14190

    arXiv 2024

  2. [2]

    Kasieczka et al., Rept

    G. Kasieczka et al., Rept. Prog. Phys.84, 124201 (2021), 2101.08320

    arXiv 2021

  3. [3]

    Aarrestad et al., SciPost Phys.12, 043 (2022), 2105.14027

    T. Aarrestad et al., SciPost Phys.12, 043 (2022), 2105.14027

    arXiv 2022

  4. [4]

    R. T. D’Agnolo and A. Wulzer, Phys. Rev. D99, 015014 (2019), 1806.02350

    arXiv 2019

  5. [5]

    J. H. Collins, K. Howe, and B. Nachman, Phys. Rev. Lett.121, 241803 (2018), 1805.02664

    arXiv 2018

  6. [6]

    J. H. Collins, K. Howe, and B. Nachman, Phys. Rev. D99, 014038 (2019), 1902.02634

    arXiv 2019

  7. [7]

    Novelty Detection Meets Collider Physics

    J. Hajer, Y.-Y. Li, T. Liu, and H. Wang, Phys. Rev. D101, 076015 (2020), 1807.10261

    work page Pith review arXiv 2020

  8. [8]

    Guiding New Physics Searches with Unsupervised Learning

    A. De Simone and T. Jacques, Eur. Phys. J. C79, 289 (2019), 1807.06038

    work page Pith review arXiv 2019

Show all 74 references
  1. [9]

    Andreassen, B

    A. Andreassen, B. Nachman, and D. Shih, Phys. Rev. D101, 095004 (2020), 2001.05001

    2020 arXiv

  2. [10]

    Nachman and D

    B. Nachman and D. Shih, Phys. Rev. D101, 075042 (2020), 2001.04990

    2020 arXiv

  3. [11]

    Knapp, O

    O. Knapp, O. Cerri, G. Dissertori, T. Q. Nguyen, M. Pierini, and J.-R. Vlimant, Eur. Phys. J. Plus136, 236 (2021), 2005.01598

    2021 arXiv

  4. [12]

    Aad et al

    G. Aad et al. (ATLAS), Phys. Rev. Lett.125, 131801 (2020), 2005.02983

    2020 arXiv

  5. [13]

    B. M. Dillon, D. A. Faroughy, J. F. Kamenik, and M. Szewc, JHEP10, 206 (2020), 2005.12319

    2020 arXiv

  6. [14]

    Crispim Romão, N

    M. Crispim Romão, N. F. Castro, and R. Pedro, Eur. Phys. J. C81, 27 (2021), [Erratum: Eur.Phys.J.C 81, 1020 (2021)], 2006.05432

    2021 arXiv

  7. [15]

    Cheng, J.-F

    T. Cheng, J.-F. Arguin, J. Leissner-Martin, J. Pilette, and T. Golling, Phys. Rev. D107, 016002 (2023), 2007.01850

    2023 arXiv

  8. [16]

    C. K. Khosa and V. Sanz, SciPost Phys.15, 053 (2023), 2007.14462

    2023 arXiv

  9. [17]

    Mikuni and F

    V. Mikuni and F. Canelli, Phys. Rev. D103, 092007 (2021), 2010.07106

    2021 arXiv

  10. [18]

    S. E. Park, D. Rankin, S.-M. Udrescu, M. Yunus, and P. Harris, JHEP21, 030 (2020), 2011.03550

    2020 arXiv

  11. [19]

    Blance and M

    A. Blance and M. Spannowsky, JHEP21, 170 (2020), 2103.03897

    2020 arXiv

  12. [20]

    Dorigo, M

    T. Dorigo, M. Fumanelli, C. Maccani, M. Mojsovska, G. C. Strong, and B. Scarpa, JHEP01, 008 (2023), 2106.05747

    2023 arXiv

  13. [21]

    Caron, L

    S. Caron, L. Hendriks, and R. Verheyen, SciPost Phys.12, 077 (2022), 2106.10164

    2022 arXiv

  14. [22]

    Hallin, J

    A. Hallin, J. Isaacson, G. Kasieczka, C. Krause, B. Nachman, T. Quadfasel, M. Schlaffer, D. Shih, and M. Sommerhalder, Phys. Rev. D106, 055006 (2022), 2109.00546

    2022 arXiv

  15. [23]

    Mikuni, B

    V. Mikuni, B. Nachman, and D. Shih, Phys. Rev. D105, 055006 (2022), 2111.06417

    2022 arXiv

  16. [24]

    R. T. d’Agnolo, G. Grosso, M. Pierini, A. Wulzer, and M. Zanetti, Eur. Phys. J. C82, 275 (2022), 2111.13633

    2022 arXiv

  17. [25]

    S. E. Park, P. Harris, and B. Ostdiek, JHEP07, 108 (2023), 2208.05484

    2023 arXiv

  18. [26]

    Hallin, G

    A. Hallin, G. Kasieczka, T. Quadfasel, D. Shih, and M. Sommerhalder, Phys. Rev. D107, 114012 (2023), 2210.14924

    2023 arXiv

  19. [27]

    Kasieczka, R

    G. Kasieczka, R. Mastandrea, V. Mikuni, B. Nachman, M. Pettee, and D. Shih, Phys. Rev. D 107, 015009 (2023), 2209.06225

    2023 arXiv

  20. [28]

    Z. Hao, R. Kansal, J. Duarte, and N. Chernyavskaya, Eur. Phys. J. C83, 485 (2023), 2212.07347

    2023 arXiv

  21. [29]

    Golling et al., in34th Conference on Neural Information Processing Systems (2023), 2303.14134

    T. Golling et al., in34th Conference on Neural Information Processing Systems (2023), 2303.14134

    2023 arXiv

  22. [30]

    Aad et al

    G. Aad et al. (ATLAS), Phys. Rev. D108, 052009 (2023), 2306.03637. 12

    2023 arXiv

  23. [31]

    E. M. Metodiev, J. Thaler, and R. Wynne, Phys. Rev. D110, 055012 (2024), 2312.00119

    2024 arXiv

  24. [32]

    Sengupta, M

    D. Sengupta, M. Leigh, J. A. Raine, S. Klein, and T. Golling, JHEP04, 109 (2024), 2312.10130

    2024 arXiv

  25. [33]

    C. L. Cheng, G. Singh, and B. Nachman (2024), 2405.08889

    2024 arXiv

  26. [34]

    Grosso, JHEP12, 093 (2024), 2407.01249

    G. Grosso, JHEP12, 093 (2024), 2407.01249

    2024 arXiv

  27. [35]

    Duffy, M

    C. Duffy, M. Hassanshah, M. Jastrzebski, and S. Malik (2024), 2407.07961

    2024 arXiv

  28. [36]

    Das and D

    R. Das and D. Shih (2024), 2410.20537

    2024 arXiv

  29. [37]

    Craig, J

    N. Craig, J. N. Howard, and H. Li (2024), 2401.15542

    2024 arXiv

  30. [38]

    J. Y. Araz and M. Spannowsky (2024), 2409.04519

    2024 arXiv

  31. [39]

    R. Das, T. Finke, M. Hein, G. Kasieczka, M. Krämer, A. Mück, and D. Shih (2024), 2411.00085

    2024 arXiv

  32. [40]

    Hammad, M

    A. Hammad, M. M. Nojiri, and M. Yamazaki (2024), 2411.09927

    2024 arXiv

  33. [41]

    Farina, Y

    M. Farina, Y. Nakai, and D. Shih, Phys. Rev. D101, 075021 (2020), 1808.08992

    2020 arXiv

  34. [42]

    Heimel, G

    T. Heimel, G. Kasieczka, T. Plehn, and J. M. Thompson, SciPost Phys.6, 030 (2019), 1808.08979

    2019 arXiv

  35. [43]

    T. S. Roy and A. H. Vijay (2019), 1903.02032

    2019 arXiv

  36. [44]

    Cerri, T

    O. Cerri, T. Q. Nguyen, M. Pierini, M. Spiropulu, and J.-R. Vlimant, JHEP05, 036 (2019), 1811.10276

    2019 arXiv

  37. [45]

    Blance, M

    A. Blance, M. Spannowsky, and P. Waite, JHEP10, 047 (2019), 1905.10384

    2019 arXiv

  38. [46]

    van Beekveld, S

    M. van Beekveld, S. Caron, L. Hendriks, P. Jackson, A. Leinweber, S. Otten, R. Patrick, R. Ruiz De Austri, M. Santoni, and M. White, JHEP09, 024 (2021), 2010.07940

    2021 arXiv

  39. [47]

    Batson, C

    J. Batson, C. G. Haaf, Y. Kahn, and D. A. Roberts, JHEP04, 280 (2021), 2102.08380

    2021 arXiv

  40. [48]

    B. M. Dillon, T. Plehn, C. Sauer, and P. Sorrenson, SciPost Phys.11, 061 (2021), 2104.08291

    2021 arXiv

  41. [49]

    Finke, M

    T. Finke, M. Krämer, A. Morandini, A. Mück, and I. Oleksiyuk, JHEP06, 161 (2021), 2104.09051

    2021 arXiv

  42. [50]

    Atkinson, A

    O. Atkinson, A. Bhardwaj, C. Englert, V. S. Ngairangbam, and M. Spannowsky, JHEP08, 080 (2021), 2105.07988

    2021 arXiv

  43. [51]

    Govorkova et al., Nature Mach

    E. Govorkova et al., Nature Mach. Intell.4, 154 (2022), 2108.03986

    2022 arXiv

  44. [52]

    Fraser, S

    K. Fraser, S. Homiller, R. K. Mishra, B. Ostdiek, and M. D. Schwartz, JHEP03, 066 (2022), 2110.06948

    2022 arXiv

  45. [53]

    S. Tsan, R. Kansal, A. Aportela, D. Diaz, J. Duarte, S. Krishna, F. Mokhtar, J.-R. Vlimant, and M. Pierini, in35th Conference on Neural Information Processing Systems (2021), 2111.12849

    2021 arXiv

  46. [54]

    Jawahar, T

    P. Jawahar, T. Aarrestad, N. Chernyavskaya, M. Pierini, K. A. Wozniak, J. Ngadiuba, J. Duarte, and S. Tsan, Front. Big Data 5, 803685 (2022), 2110.08508

    2022 arXiv

  47. [55]

    Canelli, A

    F. Canelli, A. de Cosa, L. L. Pottier, J. Niedziela, K. Pedro, and M. Pierini, JHEP02, 074 (2022), 2112.02864

    2022 arXiv

  48. [56]

    V. S. Ngairangbam, M. Spannowsky, and M. Takeuchi, Phys. Rev. D105, 095004 (2022), 2112.04958

    2022 arXiv

  49. [57]

    S. Alvi, C. W. Bauer, and B. Nachman, JHEP02, 220 (2023), 2206.08391

    2023 arXiv

  50. [58]

    Atkinson, A

    O. Atkinson, A. Bhardwaj, C. Englert, P. Konar, V. S. Ngairangbam, and M. Spannowsky, Front. Artif. Intell.5, 943135 (2022), 2204.12231

    2022 arXiv

  51. [59]

    B. M. Dillon, L. Favaro, T. Plehn, P. Sorrenson, and M. Krämer, SciPost Phys. Core6, 074 (2023), 2206.14225

    2023

  52. [60]

    Bhattacherjee, P

    B. Bhattacherjee, P. Konar, V. S. Ngairangbam, and P. Solanki (2023), 2308.13611

    2023 arXiv

  53. [61]

    Aad et al

    G. Aad et al. (ATLAS), Eur. Phys. J. C83, 681 (2023), 2211.16345

    2023 arXiv

  54. [62]

    Rep., CERN, Geneva (2016), URLhttps://cds.cern.ch/record/2205149

    Tech. Rep., CERN, Geneva (2016), URLhttps://cds.cern.ch/record/2205149

    2016

  55. [63]

    J. R. Munkres, Topology (Prentice Hall, Inc., 2000), 2nd ed., ISBN 0131816292, URLhttp://www.worldcat.org/isbn/ 0131816292

    2000

  56. [64]

    Whitney, Annals of Mathematics37, 645 (1936), ISSN 0003486X, 19398980, URLhttp://www.jstor.org/stable/ 1968482

    H. Whitney, Annals of Mathematics37, 645 (1936), ISSN 0003486X, 19398980, URLhttp://www.jstor.org/stable/ 1968482

    1936

  57. [65]

    Whitney, Annals of Mathematics45, 220 (1944), ISSN 0003486X, 19398980, URLhttp://www.jstor.org/stable/ 1969265

    H. Whitney, Annals of Mathematics45, 220 (1944), ISSN 0003486X, 19398980, URLhttp://www.jstor.org/stable/ 1969265

    1944

  58. [66]

    Alwall, R

    J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H. S. Shao, T. Stelzer, P. Torrielli, and M. Zaro, JHEP 07, 079 (2014), 1405.0301

    2014 arXiv

  59. [67]

    Bierlich et al., SciPost Phys

    C. Bierlich et al., SciPost Phys. Codeb.2022, 8 (2022), 2203.11601

    2022 arXiv

  60. [68]

    Brivio, Y

    I. Brivio, Y. Jiang, and M. Trott, JHEP12, 070 (2017), 1709.06492

    2017 arXiv

  61. [69]

    Brivio, JHEP04, 073 (2021), 2012.11343

    I. Brivio, JHEP04, 073 (2021), 2012.11343

    2021 arXiv

  62. [70]

    Cacciari, G

    M. Cacciari, G. P. Salam, and G. Soyez, JHEP04, 063 (2008), 0802.1189

    2008 arXiv

  63. [71]

    Cacciari, G

    M. Cacciari, G. P. Salam, and G. Soyez, Eur. Phys. J. C72, 1896 (2012), 1111.6097

    2012 arXiv

  64. [72]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Journal of Machine Learning Research12, 2825 (2011)

    2011

  65. [73]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., PyTorch: an imperative style, high-performance deep learning library (Curran Associates Inc., Red Hook, NY, USA, 2019)

    2019

  66. [74]

    D. P. Kingma and J. Ba, in3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , edited by Y. Bengio and Y. LeCun (2015), URLhttp://arxiv.org/abs/ 1412.6980

    2015 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.