Line shape analysis of $\Lambda(1405)$ in $\gamma p \rightarrow K^+\Sigma^-\pi^+$ reaction using convolutional neural network

Denny Lane B. Sombillo; Vince Angelo A. Chavez

REVIEW 4 major objections 5 minor 14 references

Line shape analysis of $\Lambda(1405)$ in $\gamma p \rightarrow K^+\Sigma^-\pi^+$ reaction using convolutional neural network

T0 review · 4 major / 5 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 a convolutional neural network classifies the measured $\Sigma^-\pi^+$ line shape of $\Lambda(1405)$ as a two-pole structure on the second Riemann sheet, and takes this as agreement with the current consensus.

desk verdict A competent CNN validation on synthetic line shapes that overreaches in its experimental inference, but the method deserves a careful referee. read the letter →

arxiv 2506.04622 v1 pith:IJGKU2NH submitted 2025-06-05 hep-ph hep-ex

Vince Angelo A. Chavez , Denny Lane B. Sombillo This is my paper

classification hep-phhep-ex

keywords Lambda(1405)lineshapeanalysisconvolutionalneuralnetworktwo-polestructureRiemannsheetuniformizedS-matrixinvariantmassdistribution

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 asks whether the $\Lambda(1405)$ enhancement, whose nature has been debated since its prediction in 1960, really consists of two poles, as most recent analyses conclude. It builds a convolutional neural network that classifies line shapes by their pole structure, training it on 19 synthetic classes generated from a two-channel uniformized $S$-matrix that lets the pole positions and Riemann sheets be set by hand. When the measured $\Sigma^-\pi^+$ invariant mass distribution from the $\gamma p\to K^+\Sigma\pi$ reaction is fed through the network, the inference lands on the class with two poles both on the second Riemann sheet. The approach matters because peaks near a two-hadron threshold are ambiguous between kinematical effects and genuine dynamical states, and a classifier gives a way to test the pole interpretation without committing to a specific dynamical model.

What carries the argument

The line-shape generator $F(\sqrt{s}) = |T_{11}(\sqrt{s}) + r\,e^{-i\theta} T_{21}(\sqrt{s})|^2$ constructed from independent $S$-matrix poles is the object that carries the argument. The poles are placed through the uniformization variable $\omega = (q_1+q_2)/\sqrt{\epsilon_2^2-\epsilon_1^2}$, which gives control over the position and Riemann sheet of each pole and lets the authors generate 10,000 training line shapes for each of the 19 labels. The CNN itself has two convolutional layers followed by linear layers with ReLU activations; it is trained in six stages of four labels each, and its task is to map features such as the unitarity below the second threshold and the number and shape of peaks to the correct pole label.

What would settle it

Generate test spectra from the same 19 pole classes but add a smooth background or a third channel; if the CNN's confidence in Label 04 drops below its reported 95--100\% level, the inference on the experimental spectrum is not robust to model misspecification. Alternatively, refit the measured spectrum with a one-pole and a two-pole generator and compare fit quality; a comparable one-pole fit would undercut the two-pole conclusion.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the experimental $\Sigma^-\pi^+$ spectrum selects Label 04 -- two independent poles on the second Riemann sheet -- among the 19 pole configurations in Table 1. The network reaches this answer with 95--100\% confidence at each of the six curriculum-training stages, with the only notable confusions occurring between Label 04 and other two-pole labels in Stages 4 and 5. The paper presents this as agreement with the present consensus that $\Lambda(1405)$ is a two-pole structure, one narrow pole near the $\bar{K}N$ threshold and one broad pole near the $\Sigma\pi$ threshold.

Load-bearing premise

The measured spectrum is assumed to be fully representable by the two-channel uniformized line-shape generator $F(\sqrt{s})$ with independent poles; if backgrounds, three-body dynamics, or omitted channels shape the data, the CNN cannot detect them, and its choice of Label 04 would not be a valid inference about $\Lambda(1405)$.

Editorial extensions

If this is right

If the classification is correct, the $\Lambda(1405)$ is a two-pole structure on the second Riemann sheet, consistent with the current consensus of a narrow pole near the $\bar{K}N$ threshold and a broad pole near the $\Sigma\pi$ threshold.
The trained CNN distinguishes all 19 pole classes with precision, recall, and F1-scores mostly above 0.9, showing that line-shape features carry enough information to separate one-, two-, and three-pole configurations.
The curriculum label groupings let the model extend to more complex pole structures without retraining earlier stages, so the same approach can be scaled beyond the 19 classes considered here.
The paper's program continues with the $\Sigma^+\pi^-$ and $\Sigma^0\pi^0$ invariant mass spectra, which would provide a cross-check of the two-pole inference in other final states.

Reading between the lines

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

A reader might infer that the network is effectively a non-parametric classifier for whether a spectrum shows a threshold-cusp pattern consistent with independent poles; its 'two-pole' verdict says nothing about the dynamical origin (molecular vs. quark) of those poles.
The same architecture could be pointed at other ambiguous states, such as the $P_{\psi}^N(4312)^+$, to see whether a single training recipe separates kinematical cusps from genuine poles across different channels.
Because the training generator contains no background or three-body terms, the paper's result implicitly assumes those terms do not distort the measured spectrum; adding such terms to the generator is a direct test of whether Label 04 survives model misspecification.

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.

Desk Editor's Note

read the letter

Colleague,

Read the HADRON proceedings by Chavez and Sombillo. Short version: they train a CNN to classify the CLAS Sigma-pi invariant mass line shape into one of 19 pole structures from a two-channel uniformized S-matrix, and the network picks Label 04 (two poles on sheet II), which they read as confirmation of the Lambda(1405) two-pole consensus.

What is actually new: this is the first application of a CNN (rather than a fully connected DNN) to this reaction, and the modified curriculum training with six four-label models is a reasonable way to limit resources while keeping accuracy. The synthetic validation is decent—most f1 scores above 0.9, and the confusion matrices are honestly reported.

The soft spot is the experimental inference. The training data are generated from F(s) = |T11 + r e^{-i theta} T21|^2, with pole terms only. The classifier is closed-set: it must pick one of the 19 labels. The paper never checks whether the CLAS data actually fall inside the distribution of any generated class—no goodness-of-fit, no energy test, no out-of-distribution rejection. So the 99% confidence in Label 04 is a statement about the 19-class family, not about the physical Lambda(1405). That is not a minor quibble; stages 4 and 5 show Labels 13 and 14 (three-pole structures) with f1 ~0.94-0.95 against Label 04, so the line-shape differences among competing hypotheses are comparable to the distortions a background or resolution effect could introduce.

Also: Label 04 appears as an anchor in every curriculum stage, which could bias the inference; the authors acknowledge this and propose randomized groupings for future work. And the free parameters (r, theta, pole positions, regulators) are not specified in the proceedings text, so the training set is not reproducible from the paper alone.

Credit where due: the authors are transparent that they only consider 19 pole structures and call the results preliminary in the abstract. The final sentence 'indeed a two-pole structure' is stronger than the evidence supports, but that is typical of a conference write-up.

Who this is for: someone working on ML applications in hadron spectroscopy, or on line-shape analysis near thresholds. The paper is a useful example of both the promise and the pitfalls of using classifiers for physics inference.

Peer review: I would send it to a referee rather than desk reject, because the synthetic validation is meaningful and the method could transfer to other threshold enhancements. But the referee should require a representability check and the generation parameters before the physics claim is accepted. As a proceedings, it is fine as a preliminary report; as a journal paper, it needs that additional work.

Referee Report

4 major / 5 minor

Summary. The paper trains a convolutional neural network to classify line shapes of the Lambda(1405) in the Sigma-minus-pi-plus invariant mass spectrum into one of 19 pole-structure classes. The training and validation data are generated from a two-channel uniformized S-matrix with independent poles, using the line-shape generator F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2. After a six-stage curriculum training, the authors feed 10,000 randomized variants of the CLAS Sigma-minus-pi-plus data into the network and obtain label 04 (two poles on the second Riemann sheet) with 95--100% confidence in each stage. They conclude that Lambda(1405) is indeed a two-pole structure, consistent with the present consensus.

Significance. An ML classifier that can distinguish pole structures from line-shape data would be a useful tool for hadron spectroscopy, and the paper's synthetic-data validation is a genuine strength: most labels achieve precision, recall, and F1 scores above 0.9, and the curriculum strategy is a resource-efficient way to handle 19 classes. If the inference on the CLAS data were robust, it would support the two-pole picture with an independent line-shape analysis. However, the experimental inference is only as strong as the assumption that the CLAS spectrum lies within the 19-class generator family, and the paper does not test that assumption. The conclusion is therefore a model-based classification rather than an independent derivation.

major comments (4)

[Section 2.2 and Section 2.4] The central inference depends on the assumption that the CLAS Sigma-minus-pi-plus invariant mass spectrum is representable by F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2 with the 19 pole structures of Table 1, but no goodness-of-fit or out-of-distribution test is reported. A softmax classifier is closed-set: it must return one of the 19 labels even when the input contains backgrounds, production-vertex effects, three-body phase space, or detector resolution that the generator excludes. The claim in Section 4 that Lambda(1405) is 'indeed' a two-pole structure is therefore not justified by the CNN output alone; please add a rejection class or an explicit comparison such as an energy distance or likelihood ratio between the CLAS binned distribution and the generated Label 04 samples.
[Section 2.3, Table 2] In the curriculum training, Label 04 is an anchor in every stage: Stage 1 outputs Label 04 from labels 01--04, and each subsequent stage re-trains with Label 04 plus three new labels. If Stage 1 makes a wrong inference, the error is propagated and later stages cannot recover because Label 04 is always present and the other labels are only ever compared with it. The paper acknowledges randomized label groupings only as future work, but without a sensitivity study the 95--100% confidence reported in Table 3 does not address this bias. Please rerun the curriculum with different anchor choices or without an anchor, and report how often the final inference changes.
[Section 3, Table 3] Stages 4 and 5 show near-degenerate classification between Label 04 and Labels 13 and 14, with F1 scores of 0.94--0.95 and off-diagonal confusion in the confusion matrices. The line-shape differences among these hypotheses are thus comparable to the distortions that a small background or resolution effect would introduce, so the statement that the confusion is 'so minimal that it does not significantly affect the findings' is unsupported. Please report the full distribution of prediction counts for the 10,000 randomized CLAS inputs rather than only the maximum, and quantify how far the experimental inputs sit from the Label 04 versus Label 13/14 decision boundary.
[Section 2.2] The generator F(sqrt(s)) depends on r, theta, the pole positions omega_m, and the regulators omega_m', but the paper does not state the ranges, priors, or sampling distributions for these parameters. Without this information the training set is not reproducible, and one cannot tell whether the physical line shape could fall outside the support of the generated classes. Please specify the parameter ranges and show, if possible, that the CLAS data are covered by the generated Label 04 distribution.

minor comments (5)

[Section 2.1 and Figure 1] The Figure 1 caption says 'energy and intensity as inputs' but the figure appears to show the CLAS invariant mass distribution, not the network architecture or input format; please clarify the figure and describe how inputs are preprocessed and normalized.
[Section 2.3] Please describe how the inferred label from the previous stage is used in the next stage: is the next model trained from scratch or fine-tuned, and how are the 10,000 randomized CLAS inputs combined with the synthetic data in that retraining?
[Section 2.2] Check the expression omega = (q1+q2)/sqrt(epsilon_2^2 - epsilon_1^2); the subscripts on the thresholds are not defined consistently with q1 and q2, and the square root may be a typo.
[General] There are several typographical and formatting errors, such as 'In Stages 2 and 3„' and 'pure lineshape analysis'; please proofread the text.
[Section 2.3] The hyperparameters of the CNN (kernel sizes, number of channels, optimizer, learning rate, batch size) and the ranges used to randomize the CLAS energy and intensity values are not reported; please include them for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNN is trained on simulated line shapes from an explicit S-matrix model and applied to external CLAS data, so the classification is not equivalent to its own inputs.

full rationale

The paper's derivation chain is linear and self-contained: (1) synthetic line shapes are generated from the explicit formula F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2 using a two-channel uniformized S-matrix with pole positions chosen as labels; (2) a CNN is trained and validated exclusively on these synthetic data; (3) the trained network is applied to the experimental CLAS Sigma-pi invariant mass distribution. The labels are not defined in terms of the CLAS data, and no parameter of the generator is fitted to the experimental spectrum, so the inference is a model-based classification rather than a reduction of the conclusion to the input. The closed-set representability assumption (only 19 pole structures, no goodness-of-fit to CLAS, no rejection option) is a genuine external-validity limitation, but it is not circular: it does not make Eq. (F) equal to the CLAS input by construction. The self-citations to prior work [7]-[10] describe the DNN and independent-pole methodology, but the relevant equations and training details are provided in this paper, so the argument does not rest on an unverified self-citation. The repeated inclusion of Label 04 in the curriculum is a possible bias but does not force the outcome, since the models at later stages could select other labels and the confusion matrices show nonzero alternatives. Overall, the central claim is an independent empirical classification with stated assumptions, not a circular derivation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central inference is conditioned on the forward model used to generate training data. The free parameters r, theta, and pole positions are not specified, tying the trained classifier to unstated sampling choices. The 19-pole label set is an explicit restriction, and the uniformized S-matrix product is a domain model assumption.

free parameters (3)

r (relative strength of T21 in line-shape generator) = not specified
Appears in F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2, used to generate all training line shapes. The sampling distribution is not stated, so the classifier's decision boundary depends on it.
theta (relative phase of T21) = not specified
Same formula as above. The phase and strength together shape every synthetic line shape, and their unspecified range directly influences what the CNN learns.
pole positions omega_m and regulators omega_m' = not specified
The independent-pole S-matrix controls pole locations on Riemann sheets to generate 19 labeled datasets. Values or ranges are not given, so the trained classifier is defined relative to these choices.

assumptions (3)

domain assumption The measured Sigma-pi invariant mass distribution is proportional to F(sqrt(s)) = |T11 + r e^{-i theta} T21|^2
Section 2.2 introduces this as the dataset generating function. It assumes the CLAS spectrum is a coherent combination of the two T-matrix elements with constant relative strength and phase.
domain assumption The two-channel S-matrix can be uniformized by omega = (q1+q2)/sqrt(eps2^2 - eps1^2) and written as a finite product D_m(omega) with independent poles
Section 2.2 relies on Kato's uniformization (Ref. [14]) to build training data. This imposes a specific analytic structure and a finite set of poles.
ad hoc to paper The 19 pole structures (one to three poles on Riemann sheets II, III, IV) span the plausible interpretations of Lambda(1405)
The authors state 'we only considered 19 pole structures. This is not enough to capture all possible interpretations', so the label set is a restriction introduced for this analysis that the classifier cannot escape.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Line shape analysis of $\Lambda(1405)$ in $\gamma p \rightarrow K^+\Sigma^-\pi^+$ reaction using convolutional neural network." pith.science (2026). https://pith.science/paper/IJGKU2NH

@misc{pith2026250604622,
  author       = {Pith},
  title        = {Pith review of: Line shape analysis of $\Lambda(1405)$ in $\gamma p \rightarrow K^+\Sigma^-\pi^+$ reaction using convolutional neural network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJGKU2NH}},
  note         = {Machine review of arXiv:2506.04622}
}

abstract

Interpreting peaks or dips that appear in an invariant mass distribution is a recurring challenge in hadron physics. These enhancements can be ambiguous, especially near a two-hadron threshold since kinematical and dynamical effects play an important role in their nature. One such enhancement is an exotic baryon $\Lambda(1405)$ which was first observed in 1973. Despite the few available experimental data, the statistics of the measurements of $\Lambda(1405)$ have improved for line shape analysis. The present consensus is that it is a structure of two poles both on the second Riemann sheet. However, there are still investigations of other pole structures corresponding to $\Lambda(1405)$. Lately, the use of a deep neural network in analyzing these line shapes has been proven to be effective, especially in distinguishing pole structures. Thus, in this study, we develop a convolutional neural network, a type of DNN, to determine the general pole structure that corresponds to $\Lambda(1405)$ found in the $\Sigma^-\pi^+$ invariant mass distribution measured by CLAS in their experiment involving the $\gamma p \rightarrow K^+\Sigma\pi$ reaction. The CNN is trained using a two-channel uniformized $S$-matrix allowing us to control the position and the corresponding Riemann sheet of the poles. Our preliminary results show that the trained CNN can accurately distinguish pole structures in the $\Sigma^-\pi^+$ invariant mass distribution and agrees with the present consensus of a two-pole structure. This supports the preceding works on the $\Lambda(1405)$ and requires a thorough analysis of $\Sigma^+\pi^-$ and $\Sigma^0\pi^0$ invariant mass spectra.

Figures

Figures reproduced from arXiv: 2506.04622 by the authors.

**Figure 1.** CLAS measurement [11] of Λ(1405) in Σ −𝜋 + invariant mass with energy and intensity as inputs for the convolutional neural network model convolutional layers will then pass through a series of linear layers that aim to map these features to their corresponding pole structure having a linear function of 𝑧𝑖 = 𝑊𝑖 𝑗𝑥 𝑗 + 𝑏𝑖 . To introduce non-linearity between these linear functions, we included activation layers called… view at source ↗

**Figure 2.** Confusion matrices from the validation phase of Stages 1, 3 and 5. The labels are in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages

[1]

Dalitz and S.F

R.H. Dalitz and S.F. Tuan,Ann. Phys. 10(1960) 307

work page 1960
[2]

Thomas, A

D.W. Thomas, A. Engler, H.E. Fisk and R.W. Kraemer,Nucl. Phys. B 56 (1973) 15

work page 1973
[3]

Hemingway,Nucl

R.J. Hemingway,Nucl. Phys. B 253 (1985) 742

work page 1985
[4]

J-PARC E31collaboration, Phys. Lett. B 837 (2023) 137637 [2209.08254]

arXiv 2023
[5]

Mai and U.-G

M. Mai and U.-G. Meißner,Eur. Phys. J. A 51(2015) 30 [1411.7884]

arXiv 2015
[6]

Meson-baryon reactions with strangeness -1 within a chiral framework

Z.-H. Guo and J.A. Oller,Phys. Rev. C 87(2013) 035202 [1210.3485]

work page Pith review arXiv 2013
[7]

Sombillo, Y

D.L.B. Sombillo, Y. Ikeda, T. Sato and A. Hosaka,Phys. Rev. D 102 (2020) 016024

work page 2020
[8]

Pole structure of $P_\psi^N(4312)^+$ via machine learning and uniformized S-matrix

L.M. Santos, V.A.A. Chavez and D.L.B. Sombillo,J. Phys. G 52(2025) 015104 [2405.11906]

work page Pith review arXiv 2025

Show all 14 references

[9]

Co, V.A.A

D.A.O. Co, V.A.A. Chavez and D.L.B. Sombillo,Phys. Rev. D 110 (2024) 114034 [2403.18265]

2024 arXiv
[10]

Santos and D.L.B

L.M. Santos and D.L.B. Sombillo,Phys. Rev. C 108 (2023) 045204 [2308.03325]

2023 arXiv
[11]

CLAS collaboration, Phys. Rev. Lett. 112 (2014) 082004 [1402.2296]

2014 arXiv
[12]

Hornik, M

K. Hornik, M. Stinchcombe and H. White,Neural Networks 2 (1989) 359

1989
[13]

Indolia, A.K

S. Indolia, A.K. Goswami, S. Mishra and P. Asopa,Procedia Comp. Sci. 132 (2018) 679

2018
[14]

Kato,Ann

M. Kato,Ann. Phys. 31 (1965) 130. 6

1965

Pith tools

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

[1] [1]

Dalitz and S.F

R.H. Dalitz and S.F. Tuan,Ann. Phys. 10(1960) 307

work page 1960

[2] [2]

Thomas, A

D.W. Thomas, A. Engler, H.E. Fisk and R.W. Kraemer,Nucl. Phys. B 56 (1973) 15

work page 1973

[3] [3]

Hemingway,Nucl

R.J. Hemingway,Nucl. Phys. B 253 (1985) 742

work page 1985

[4] [4]

J-PARC E31collaboration, Phys. Lett. B 837 (2023) 137637 [2209.08254]

arXiv 2023

[5] [5]

Mai and U.-G

M. Mai and U.-G. Meißner,Eur. Phys. J. A 51(2015) 30 [1411.7884]

arXiv 2015

[6] [6]

Meson-baryon reactions with strangeness -1 within a chiral framework

Z.-H. Guo and J.A. Oller,Phys. Rev. C 87(2013) 035202 [1210.3485]

work page Pith review arXiv 2013

[7] [7]

Sombillo, Y

D.L.B. Sombillo, Y. Ikeda, T. Sato and A. Hosaka,Phys. Rev. D 102 (2020) 016024

work page 2020

[8] [8]

Pole structure of $P_\psi^N(4312)^+$ via machine learning and uniformized S-matrix

L.M. Santos, V.A.A. Chavez and D.L.B. Sombillo,J. Phys. G 52(2025) 015104 [2405.11906]

work page Pith review arXiv 2025