Sky Background Building of Multi-objective Fiber spectra Based on Mutual Information Network

Ali Luo; Haifeng Yang; Hui Zhang; Jianghui Cai; Lichan Zhou; Qin Han; Xiao Kong; Yuqing Yang; Zhichao Ding

REVIEW 3 major objections 4 minor 26 references

Sky Background Building of Multi-objective Fiber spectra Based on Mutual Information Network

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A mutual-information network estimates per-location sky backgrounds better than the survey's Super sky, especially in the blue end of the spectrum.

desk verdict New per-fiber sky estimation idea, but the evaluation can't distinguish learning from memorization. read the letter →

arxiv 2508.19875 v1 pith:4IK6IMX6 submitted 2025-08-27 cs.CV cs.LG

Hui Zhang , Jianghui Cai , Haifeng Yang , Ali Luo , Yuqing Yang , Xiao Kong , Zhichao Ding , Lichan Zhou

show 1 more author

Qin Han

This is my paper

classification cs.CVcs.LG

keywords skybackgroundsubtractionmutualinformationmulti-fiberspectroscopyLAMOSTdeeplearningemissionlineswavelengthcalibrationrepresentation

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 proposes SMI, a deep network that builds the sky background for each fiber position using spectra from all fibers in a single LAMOST plate, rather than relying only on dedicated sky fibers averaged into a Super sky. SMI decomposes the sky into a shared component (common continuum and emission lines) and a location-specific component (emission lines that vary with position), then uses a two-stage mutual-information training scheme to learn both. Tested against observed sky-fiber spectra, SMI's estimated sky backgrounds have smaller residuals, lower MAE and RMSE, and fewer outliers than LAMOST's Super sky, most clearly in the blue end and in sparse-emission regions. If correct, the method offers a practical route to per-fiber sky subtraction that uses on-source information rather than only sky fibers.

What carries the argument

The central object is SMI, a two-stage mutual-information network. Its key pieces are: (1) a decomposition of the sky background into a common part (Ssm, roughly position-independent continuum plus shared emission lines) and an exclusive part (So, position-dependent emission lines); (2) a feature extraction block that uses a small convolution kernel plus a wavelength-calibration module to align shifted emission-line features before mutual information is computed; (3) a first training stage that maximizes cross mutual information between representations of different spectra — with representations swapped to suppress unique information — yielding a 'shared' sky representation; (4) a second sta

What would settle it

Hold out a complete set of sky fibers — exclude them from both the pretraining labels and the network's input — and compare SMI's sky estimate at those exact positions against the observed spectra. If the accuracy advantage over LAMOST Super sky disappears or reverses on those held-out fibers, the claim that SMI generalizes beyond its training data is refuted. A complementary test injects synthetic sky emission lines with known strengths into simulated fiber data and checks whether SMI recovers them independently of the real sky-fiber spectra.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the sky background at an object fiber can be estimated more accurately by exploiting the mutual information shared across all fiber spectra in a plate than by averaging dedicated sky fibers. The model separates each observed spectrum into a common sky component and a unique local component: it maximizes mutual information between representations of different spectra to capture the common component, then minimizes mutual information between the shared representation and each location's full representation to isolate the unique emission-line component. The shared component plus the unique component yields an individual sky background at each obje

Load-bearing premise

The evaluation assumes the observed sky-fiber spectrum at a sky position is the true sky background, even though that spectrum is fed into the network as input and the pretraining labels came from the same spectrometers; if the network simply memorizes or interpolates its input sky fibers, the reported improvement over Super sky would not transfer to object fibers.

Editorial extensions

If this is right

Sky subtraction at object-fiber positions can be done with a position-specific background instead of one plate-wide average, potentially removing residual sky structure that Super sky leaves in target spectra.
The blue end of the spectrum benefits most, which matters because blue sky emission lines (e.g., OH and Hg features) are often the hardest to model and most contaminating for faint-object science.
Fewer outliers in the estimated sky background translate to fewer spurious features in reduced spectra and more reliable sky emission-line subtraction.
The shared representation produced by the first stage is itself a clean common-sky model for a field, preserving emission-line positions and profile widths across fibers.
The incremental pretraining-plus-mutual-information recipe is portable to other multi-fiber instruments such as SDSS, DESI, or JWST NIRSpec, as long as wavelength-calibrated spectra from a single plate are available.

Reading between the lines

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

The paper does not yet test downstream science outcomes; we infer that if SMI's sky estimates are genuinely better, they should improve stellar parameter measurements and line-ratio analyses after subtraction, an extension that would settle the practical gain.
The evaluation uses observed sky-fiber spectra as ground truth even though those same spectra are part of the model's input and pretraining labels; we infer that the reported accuracy could be inflated by the network partially reproducing its inputs, and a held-out test with entire sky fibers excluded would clarify this.
The mutual-information decomposition between 'shared across all fibers' and 'unique to a location' is a general recipe that could be applied beyond astronomy to any multi-sensor setting where a common background contaminates position-dependent signals, for example in multi-detector imaging or array spectroscopy.
The observed depression at dense emission-line peaks in Figure 13 suggests a specific limitation; we infer that a wider receptive field or a post-processing smoothing step could recover those peaks, making the method even more competitive in the red-arm dense-line bands.

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

You should know this paper has a genuinely new idea wrapped in an evaluation that does not support its central claim. The idea: instead of building one Super sky from sky fibers, learn an individual sky background at every fiber by exploiting mutual information across all fibers in a LAMOST plate. The architecture is a legitimate new application of MINE/Deep InfoMax: a first stage maximizes cross-representation MI to extract a shared sky component, a second stage minimizes MI to get the fiber-specific component, and a wavelength-calibration block addresses emission-line shifts. None of the cited PCA/NMF/B-spline baselines do per-fiber sky from all spectra, so the novelty claim is fair.

What the paper does well: it is honest. The summary admits no downstream validation was performed, and the shared-representation figures show the model does capture common emission-line structure. There is a real direction here.

The soft spots are serious. The evaluation is circular in a specific, load-bearing way: the model trains on all fibers, including the sky fibers, and the pretraining loss (Eq. 8) directly fits sky labels. In §3.2 the 'authentic sky background' benchmark is the observed sky-fiber spectrum at the same fiber the model saw. Low residuals can therefore be memorization, not estimation. The paper even shows target-fiber outputs deviate substantially (Fig. 10), and the tables labeled MAE contain negative values, which is impossible for an absolute error. That alone undermines the numbers. Several spectrograph rows also show SMI worse than the LAMOST pipeline, so the qualitative claim is not consistent.

If you review this, don't desk reject—the novelty is worth referee time. But the authors need to retrain without the evaluation sky fibers, report actual MAE, and show object-fiber performance or a downstream measurement. Without that, the central claim is unsupported.

Referee Report

3 major / 4 minor

Summary. The paper proposes SMI, a two-stage neural-network method that uses spectra from all fibers in a LAMOST plate to estimate an individual sky-background spectrum at each fiber position. A wavelength-calibration module is introduced to align emission-line features, and a mutual-information objective (based on MINE/Deep InfoMax) is used first to extract a 'shared' sky component and then a location-specific 'exclusive' component. The authors train and test on their own reduction of 2D LAMOST CCD data, compare SMI's sky-fiber estimates against the observed sky-fiber spectra and against LAMOST's Super Sky, and report lower residual statistics (MAE, RMSE) in several spectrographs, concluding that SMI gives better object sky backgrounds, especially in the blue end.

Significance. If the central claim were rigorously established, per-fiber sky-background estimation from all fibers would be a useful contribution to multi-object spectroscopy, complementing conventional Super Sky methods. The paper also ships no code or data, and the evaluation is largely internal. On the evidence presented, the claim is not supported: the benchmark is circular (sky-fiber spectra are both inputs and labels), the reported 'MAE' contains negative values, several spectrograph rows show SMI worse than LAMOST, and no independent object-fiber or downstream validation is provided. The idea is potentially interesting, but the current manuscript does not meet the evidentiary bar for its main conclusion.

major comments (3)

[§3.2, §2.3.1, §3.1] The quantitative evaluation is circular. §3.1 states that the model is trained with all fibers in a single observation, and §2.3.1 (Eq. 8) pretrains the feature extractor directly on sky labels from the same spectrometers. The evaluation protocol in §3.2 then treats observed sky-fiber spectra as the 'authentic sky background' and compares SMI's estimates at those very sky-fiber positions. Because those positions are part of the model's input and label set, low residuals could reflect copying or interpolating the input at that fiber rather than a genuinely better spatially varying sky model at object positions. The paper itself notes in Fig. 10 that target-fiber results have 'significant discrepancies from observational data,' and §4 concedes that no downstream validation was performed. No evidence supports the abstract's claim that SMI yields better object-fiber sky backgrounds. A valid
[Tables 1–3] The metric labeled 'MAE' takes negative values (e.g., Table 1, HD023750N350938V02 spec04: -0.61; HD094551N184100B02 spec04: -42.58; Table 2, GAC066N10B1 spec04: -103.07; Table 3 LAMOST row 6506: -0.66). A mean absolute error cannot be negative; the quantity appears to be a signed mean residual (bias), not MAE. The interpretation of smaller values as 'closer to observations' is therefore unreliable. In addition, many rows show SMI worse than LAMOST (e.g., Table 1, HD094551N184100B02 spec05 RMSE: 189.19 vs 115.28; testplanid-1 spec03 MAE: 190.22 vs 181.34; testplanid-2 spec04 MAE: 270.93 vs 264.77). These mixed results are acknowledged only as 'variations' in the text, but they substantially undermine the blanket improvement claim. No error bars or significance tests across spectra or plates are provided.
[Abstract, §3.2.1]

minor comments (4)

[Eqs. 1, 9, 18] Several typos and unclear notations: Eq. 1 'sym-bolizes'; Eq. 9 has a typo in the denominator 'p(x)(z)' and the marginal/joint densities are not defined cleanly; Eq. 18's phrase 'counts for the length of the data within the number of emission line features' is unclear, and the stopping criterion for the adaptive segmentation is not precisely specified.
[Fig. 11, Fig. 12, Fig. 13] Axis labels and units should be standardized. Some captions say x is wavelength and y is flux, but in the figures the x-axis is sometimes labeled 'spectral flux' and no units are given. This makes quantitative reading difficult.
[§3.2.1, Table 1 text] The text states that for spectrographs 04, 05, and 15 in HD023750N350938V02 'MAE and RMSE from both SMI and LAMOST nearly zero,' but Table 1 lists values between 0.21 and 44.70 with several much larger. 'Nearly zero' is inaccurate.
[§2.3.2–2.3.3] Equations 12–17 give the mutual-information objectives, but the decomposition into Ssm(λ) and So(i,λ) used in Eq. 6 is not formally connected to these objectives. Please clarify how the learned representations are mapped to the physical sky components, and give the actual values chosen for α and β in the experiments.

Circularity Check

1 steps flagged · score 6.0 of 10

Sky-fiber benchmark is an input/label: SMI's 'closer to observed sky' comparison reduces to reproducing the evaluated fiber, not a validated object-fiber estimate.

fitted input called prediction [§3.1 Data Reduction, §2.3.1 Pre-Train Model (Eq. 8), §3.2 Result and Discussion]
""this model is trained with all fibers in a single observation" (§3.1); "Spectral data containing information on sky emission lines were used as sky labels" (§2.3.1); "we propose that the sky fiber data represents the authentic sky background at the observation site and can serve as a reliable evaluation benchmark. The SMI estimations obtained from the sky fiber position are compared with the acquired sky fiber data" (§3.2)."

The quantitative benchmark for 'closer to observed sky' is the observed sky-fiber spectrum at the very fiber being evaluated. But SMI's inputs include all fiber spectra (§3.1), including that sky fiber, and Eq. 8 pretrains the feature extractor to minimize KL divergence between its output and sky labels derived from the same spectral data. A low residual at sky-fiber positions can therefore be achieved by reproducing or interpolating the input spectrum, not by independently recovering a spatially varying sky field at object positions. The abstract's central claim of a better object sky background is thus supported only by a benchmark that is an input/label, not by an independent target-fiber validation.

full rationale

The mutual-information objectives in Eqs. 12-17 are not themselves circular: they define an optimization problem and the model is trained to solve it. The circularity enters at the evaluation stage. Section 3.2 declares sky-fiber data to be the ground-truth benchmark and compares SMI's sky-fiber estimates against those data. However, Section 3.1 states the model is trained with all fibers in one observation, and Section 2.3.1 pretrains the network with sky labels from the same spectrometers (Eq. 8). Thus the reference sky-fiber spectrum is simultaneously a model input and the source of the pretraining labels; the reported improvement at sky-fiber positions can reflect memorization or interpolation rather than better object-fiber sky estimation. The paper's own Section 4 concedes that no downstream target-fiber validation was performed, which further weakens support for the central object-fiber claim. I also note that Tables 1-3 contain negative MAE values, impossible for mean absolute error; this is a numerical reliability issue, not circularity. The self-citations (Yang et al., Cai et al.) are background/future-work and are not load-bearing. Because the main quantitative claim reduces to fitting/input reproduction at the evaluation points, the circularity score is 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. The shared and exclusive sky representations are learned latent variables, not independently evidenced physical constructs. The entire central claim rests on domain assumptions about spectral decomposition and on a circular validation protocol.

free parameters (3)

alpha in Eq. 14 = not reported
Weights the L1 constraint on shared representations; no value or search range is given in the text.
beta in Eq. 17 = not reported
Weights the penalty that minimizes mutual information between exclusive representations; no value is reported.
Segment length stopping criterion (Eq. 18) = not quantified
Adaptive segmentation stops when 'the ratio tends to stabilize'; this is a hand-set heuristic that determines input segment lengths and affects the training data structure.

assumptions (5)

domain assumption The observed sky background can be decomposed into a common component Sm(i, lambda) and an exclusive component So(i, lambda) (Eqs. 3-4).
The mutual information objectives assume this decomposition is identifiable from spectra; no proof is offered.
domain assumption Sky fibers contain only sky background, with no target signal (Eq. 7).
Standard for calibration but can be violated by scattered light, fiber cross-talk, or misalignment; the paper relies on this to form ground-truth labels.
domain assumption Neighborhood median subtraction removes the continuum and yields a valid sky emission-line label (Section 2.3.1, Figure 5).
The pretraining loss fits this label; if the median residual retains continuum or suppresses broad lines, the learned features are biased.
domain assumption The 5577 Angstrom sky line can be used to normalize fiber efficiency differences (Section 2.1).
Assumes the intrinsic line flux is uniform across fibers; spatial variation of the line would introduce normalization errors.
domain assumption Maximizing and minimizing the specified mutual information terms separates shared from unique information (Section 2.3.3, Eqs. 14 and 17).
Borrowed from representation learning literature; no derivation or ablation establishes this holds for spectral data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sky Background Building of Multi-objective Fiber spectra Based on Mutual Information Network." pith.science (2026). https://pith.science/paper/4IK6IMX6

@misc{pith2026250819875,
  author       = {Pith},
  title        = {Pith review of: Sky Background Building of Multi-objective Fiber spectra Based on Mutual Information Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4IK6IMX6}},
  note         = {Machine review of arXiv:2508.19875}
}

read the original abstract

Sky background subtraction is a critical step in Multi-objective Fiber spectra process. However, current subtraction relies mainly on sky fiber spectra to build Super Sky. These average spectra are lacking in the modeling of the environment surrounding the objects. To address this issue, a sky background estimation model: Sky background building based on Mutual Information (SMI) is proposed. SMI based on mutual information and incremental training approach. It utilizes spectra from all fibers in the plate to estimate the sky background. SMI contains two main networks, the first network applies a wavelength calibration module to extract sky features from spectra, and can effectively solve the feature shift problem according to the corresponding emission position. The second network employs an incremental training approach to maximize mutual information between representations of different spectra to capturing the common component. Then, it minimizes the mutual information between adjoining spectra representations to obtain individual components. This network yields an individual sky background at each location of the object. To verify the effectiveness of the method in this paper, we conducted experiments on the spectra of LAMOST. Results show that SMI can obtain a better object sky background during the observation, especially in the blue end.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 25 canonical work pages

[1]

K., Ag¨ueros, M

Abazajian, K., Adelman-McCarthy, J. K., Ag¨ueros, M. A., et al. 2004, The Astronomical Journal, 128, 502 2

work page 2004
[2]

2017, Guang pu xue yu Guang pu fen xi= Guang pu, 37, 273 2

An, R., Pan, J., Yi, Z., & Wei, P. 2017, Guang pu xue yu Guang pu fen xi= Guang pu, 37, 273 2

work page 2017
[3]

D., & Buchwalter, W

Bachman, P., Hjelm, R. D., & Buchwalter, W. 2019, Learning representations by maximizing mutual in- formation across views, Proceedings of the 33rd International Conference on Neural Information Processing Systems (Red Hook, NY , USA: Curran Associates Inc.) 6

work page 2019
[4]

R., Zhang, L., & Ye, Z

Bai, Z. R., Zhang, L., & Ye, Z. F. 2007, Acta Astronomica Sinica, 48, 507 2, 6, 13

work page 2007
[5]

2017, Research in Astronomy and Astrophysics, 17, 091 1

Bai, Z.-R., Zhang, H.-T., Yuan, H.-L., et al. 2017, Research in Astronomy and Astrophysics, 17, 091 1

work page 2017
[6]

I., Baratin, A., Rajeswar, S., et al

Belghazi, M. I., Baratin, A., Rajeswar, S., et al. 2018, arXiv preprint arXiv:1801.04062 6

arXiv 2018
[7]

2019, SPECTROSCOPY AND SPECTRAL ANALYSIS, 39, 1301 13

Cai, J.-H., Yang, Y .-Q., Yang, H.-F., et al. 2019, SPECTROSCOPY AND SPECTRAL ANALYSIS, 39, 1301 13

work page 2019
[8]

2024, Research in Astronomy and Astrophysics, 24, 095023 1

Fu, H., Liu, P., Qi, X., & Mei, X. 2024, Research in Astronomy and Astrophysics, 24, 095023 1

work page 2024

Show all 26 references

[9]

2023, The Astronomical Journal, 165, 144 2

Guy, J., Bailey, S., Kremin, A., et al. 2023, The Astronomical Journal, 165, 144 2

2023
[10]

2023, Monthly Notices of the Royal Astronomical Society, 526, 5520 2, 4, 13 Sky Background Building 15

Han, B., Song, Y ., & Zhao, Y . 2023, Monthly Notices of the Royal Astronomical Society, 526, 5520 2, 4, 13 Sky Background Building 15

2023
[11]

D., Blasch, E., et al

Jia, B., Pham, K. D., Blasch, E., et al. 2018, in 2018 IEEE Aerospace Conference, 1 1

2018
[12]

2023, arXiv preprint arXiv:2312.06761 2

Kolganov, F., Chilingarian, I., & Grishin, K. 2023, arXiv preprint arXiv:2312.06761 2

2023 arXiv
[13]

J., & Mink, D

Kurtz, M. J., & Mink, D. J. 2000, The Astrophysical Journal, 533, L183 2

2000
[14]

2019, Publications of the Astronomical Society of the Pacific, 131, 024505 1

Rui, W., Luo, A.-L., Shuo, Z., et al. 2019, Publications of the Astronomical Society of the Pacific, 131, 024505 1

2019
[15]

H., Serrurier, M., & Ortner, M

Sanchez, E. H., Serrurier, M., & Ortner, M. 2020, in Computer Vision–ECCV 2020: 16th European

2020
[16]

Conference, Glasgow, UK, August 23–28, 2020,

2020
[17]

2014, Spectroscopy and Spectral Analysis, 34, 3132 2

Shi, Z.-X., Georges, C., Luo, A., Tu, L.-P., et al. 2014, Spectroscopy and Spectral Analysis, 34, 3132 2

2014
[18]

T., Lilly, S

Soto, K. T., Lilly, S. J., Bacon, R., Richard, J., & Conseil, S. 2016, Monthly Notices of the Royal Astronomical Society, 458, 3210 2

2016
[19]

E., Kaviraj, S., Collins, C

Watkins, A. E., Kaviraj, S., Collins, C. C., et al. 2024, Monthly Notices of the Royal Astronomical Society, 528, 4289 2

2024
[20]

G., Lee, J

Williams, T. G., Lee, J. C., Larson, K. L., et al. 2024, The Astrophysical Journal Supplement Series, 273, 13 2

2024
[21]

2018, in 2018 14th International Conference on Computational Intelligence and Security (CIS), IEEE, 476 2

Wu, P., Yin, Q., & Guo, P. 2018, in 2018 14th International Conference on Computational Intelligence and Security (CIS), IEEE, 476 2

2018
[22]

2023, Research in Astronomy and Astrophysics, 23, 055006 1

Yang, H.-F., Yin, X.-N., Cai, J.-H., et al. 2023, Research in Astronomy and Astrophysics, 23, 055006 1

2023
[23]

2024, The Astrophysical Journal Supplement Series, 272, 43 1

Yang, H.-F., Wang, R., Cai, J.-H., et al. 2024, The Astrophysical Journal Supplement Series, 272, 43 1

2024
[24]

2013, Research in Astronomy and Astrophysics, 1255 13

Yao, S., Haotong, Z., Hailong, Y ., et al. 2013, Research in Astronomy and Astrophysics, 1255 13

2013
[25]

2016, Publications of the Astronomical Society of Australia, 33, e058 2

Zhang, B., Zhang, L., & Ye, Z. 2016, Publications of the Astronomical Society of Australia, 33, e058 2

2016
[26]

2012, Publications of the Astronomical Society of Australia, 29, 78 2

Zhu, J., & Ye, Z.-f. 2012, Publications of the Astronomical Society of Australia, 29, 78 2

2012

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

K., Ag¨ueros, M

Abazajian, K., Adelman-McCarthy, J. K., Ag¨ueros, M. A., et al. 2004, The Astronomical Journal, 128, 502 2

work page 2004

[2] [2]

2017, Guang pu xue yu Guang pu fen xi= Guang pu, 37, 273 2

An, R., Pan, J., Yi, Z., & Wei, P. 2017, Guang pu xue yu Guang pu fen xi= Guang pu, 37, 273 2

work page 2017

[3] [3]

D., & Buchwalter, W

Bachman, P., Hjelm, R. D., & Buchwalter, W. 2019, Learning representations by maximizing mutual in- formation across views, Proceedings of the 33rd International Conference on Neural Information Processing Systems (Red Hook, NY , USA: Curran Associates Inc.) 6

work page 2019

[4] [4]

R., Zhang, L., & Ye, Z

Bai, Z. R., Zhang, L., & Ye, Z. F. 2007, Acta Astronomica Sinica, 48, 507 2, 6, 13

work page 2007

[5] [5]

2017, Research in Astronomy and Astrophysics, 17, 091 1

Bai, Z.-R., Zhang, H.-T., Yuan, H.-L., et al. 2017, Research in Astronomy and Astrophysics, 17, 091 1

work page 2017

[6] [6]

I., Baratin, A., Rajeswar, S., et al

Belghazi, M. I., Baratin, A., Rajeswar, S., et al. 2018, arXiv preprint arXiv:1801.04062 6

arXiv 2018

[7] [7]

2019, SPECTROSCOPY AND SPECTRAL ANALYSIS, 39, 1301 13

Cai, J.-H., Yang, Y .-Q., Yang, H.-F., et al. 2019, SPECTROSCOPY AND SPECTRAL ANALYSIS, 39, 1301 13

work page 2019

[8] [8]

2024, Research in Astronomy and Astrophysics, 24, 095023 1

Fu, H., Liu, P., Qi, X., & Mei, X. 2024, Research in Astronomy and Astrophysics, 24, 095023 1

work page 2024