Pith. sign in

REVIEW 4 major objections 5 minor 73 references

A straightforward linear compression of survey colors lowers globular-cluster contamination from ~45% to ~35% in LSST-like photometry, at the cost of missing most true clusters.

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

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

With ugrizY colors alone, an LSST-like catalog yields at best ~35% contamination and ~19% completeness for globular-cluster candidates; principal components help marginally, autoencoders do not.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Useful benchmark with a real label-contamination issue that must be fixed before the numbers are trusted. the 4 major comments →

arxiv 2512.17644 v2 pith:BLDFO7DV submitted 2025-12-19 astro-ph.GA

The contribution of the color space in LSST-like photometry for the selection of extragalactic globular cluster candidates

Nicholas Schweder-Souza , Ana L. Chies-Santos , Rafael S. de Souza , Kristen C. Dage , Charles J. Bonatto , Juan P. Caso , Michele Cantiello , Pedro dos Santos-Lopes
show 11 more authors
Pedro Floriano Thayse A. Pacheco Katherine L. Rhode Pauline Barmby Jennifer Sobeck Ana I. Ennis Yasna Ordenes-Brice\~no Teymoor Saifollahi Julia Gschwend Niranjana P. Rubens E. G. Machado
This is my paper
classification astro-ph.GA
keywords globular clustersLSST-like photometrycolor spaceprincipal component analysisrandom forestFornax Clusterphotometric classificationcontamination
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper asks how much separating power lives in the six-band ugrizY color space of an LSST-like catalog for finding point-like globular clusters around external galaxies. Using a labeled Fornax Cluster catalog of spectroscopically confirmed globular clusters, background galaxies, and foreground stars, the authors compare classifiers fed all 15 colors, principal components of those colors, and auto-encoder latent coordinates. They establish that feeding the linear principal components to a random forest lowers the minimum contamination of a globular-cluster candidate list from roughly 45% to about 35%, at the price of missing roughly 80% of true clusters. They also show that non-linear auto-encoders add nothing over PCA, and that two-dimensional color-color diagrams are the weakest input. The practical point for upcoming wide surveys is that ugrizY colors alone cannot deliver clean globular-cluster samples; morphological or near-infrared information has to be added.

Core claim

The paper assembles an LSST-like ugrizY photometric catalog of the central Fornax Cluster with labeled spectroscopically confirmed globular clusters, background galaxies, and foreground stars, and asks how well point-like globular clusters can be separated using colors alone. It finds that a random forest fed all 15 colors reaches a best precision of 57%, corresponding to roughly 43–45% contamination; feeding it the 15 principal components raises precision to 65%, corresponding to roughly 35% contamination, but recall drops to 19% — 81% of true globular clusters are missed. Non-linear auto-encoder latent coordinates do not improve on PCA or raw colors, and multi-layer perceptrons trade preci

What carries the argument

The central object is the 15-color space built from ugrizY photometry (u−g, u−r, u−i, u−z, u−Y, g−r, g−i, g−z, g−Y, r−i, r−z, r−Y, i−z, i−Y, z−Y). The machinery is the comparison of three feature representations — raw colors, their principal components (a linear rotation that concentrates variance into fewer dimensions), and non-linear auto-encoder latent coordinates — fed to a random forest and a multilayer perceptron, scored by precision, recall, and F1 for the globular-cluster class. The decisive result is that PCA compresses the color information without loss: 4 principal components reproduce the 15-color performance, while 2D projections (color-color diagrams, PC1-PC2, 2D latent spaces)

Load-bearing premise

The foreground-star class is assumed to be almost free of real globular clusters; the paper itself says the same magnitude cut may leave 35–40% or ~10% of the GC population in that star sample, and if many GCs are mislabeled as stars, the measured contamination numbers are not a clean property of the color space.

What would settle it

Re-run the random forest on the 15 principal components with the stellar class redefined using space-based morphology or astrometric parallax; if the globular-cluster precision moves materially, the ~35% contamination floor is driven by label contamination rather than by the color space itself.

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

If this is right

  • For LSST-like ugrizY catalogs, color-only globular-cluster selection has a contamination floor: expect roughly 35% contamination even at best, and only with a highly incomplete (≈19% recall) candidate list.
  • Principal components of the colors are a better input than the raw colors for random-forest globular-cluster selection; using 4 PCs recovers the 15-color performance, so linear compression is safe.
  • Non-linear auto-encoder features do not improve globular-cluster classification over linear PCA; the information available in ugrizY colors is linearly accessible.
  • Two-dimensional color-color diagrams are the weakest option, roughly doubling contamination relative to using the full high-dimensional color information.
  • Reducing contamination below this floor will require augmenting colors with morphology, near-infrared, UV, or astrometric data before more complex classifiers can help.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: retrain the same classifiers on a version of the star sample built with space-based morphology or astrometric parallax to exclude extragalactic interlopers; if the confusion-matrix numbers change materially, the measured ~35% floor is partly a label-purity effect rather than a property of ugrizY colors.
  • The same PCA-plus-random-forest recipe could be applied to the ~1000-dimensional color space of a 59-band survey such as the one the paper mentions; the authors point to PCA only as a starting point, and the extrapolation to much higher dimensional color spaces is a natural next step.
  • The paper's abstract and conclusions report different contamination numbers for the 15-color case (~30% versus ~45%); a reader planning a survey should treat the higher, discussion-stage figure as the operational one, since it is the value the authors repeat in their final assessment.
  • For LSST-era operations, the paper's numbers imply that photometric globular-cluster candidate lists will need to be cross-matched with space-based morphology before spectroscopic follow-up; otherwise most confirmed clusters will be missing at any acceptable purity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper assembles an LSST-like ugrizY photometric catalog of the Fornax Cluster by combining FDS u-band and DES grizY photometry, labels 1073 spectroscopically confirmed GCs plus 50,638 galaxies and 3,962 stars from DES morphology and magnitude cuts, and compares random forest and multi-layer perceptron classifiers fed with 15 colors, reduced color subsets, principal components, and auto-encoder latent coordinates. The central claim is that color-only GC selection reaches a minimum contamination of roughly 45% with raw colors, reduced to about 35% when principal components are used, while AEs give no improvement; the authors conclude that ancillary data are needed for LSST-era GC selection.

Significance. If the measurements are clean, the paper provides a useful quantitative benchmark for LSST preparation: it directly estimates the ceiling of ugrizY colors for unresolved extragalactic GC selection using a spectroscopically confirmed GC training set, and it systematically compares linear and non-linear dimensionality reductions. The qualitative conclusion that color-only selection is incomplete and contaminated, and that PCs help modestly, is plausible and broadly supported by Table 5. The paper would be strengthened by releasing code and data and by adding proper uncertainty estimates, but the empirical framing is a genuine contribution to a practical problem.

major comments (4)
  1. [§2.2, Eq. (2)] The GCLF paragraph is numerically self-contradictory and the purity of the 'star' class is not quantified. The text first says the mag_auto_i<=21.0 cut is at ~0.33 sigma from TOM_i~23 and leaves 35-40% of the Fornax GC population in the stellar sample, then says this fraction is ~10% for the same cut. With sigma_GCLF~1.5, a 21.0 cut sits at ~1.33 sigma, giving ~10%; a 22.5 cut gives ~37%. The star test set has 792 sources, so even a 10% hidden-GC fraction places ~80 true GCs in the 'star' class. Since precision/recall and hence the headline 35-45% contamination rates are computed against these labels, these numbers are not a clean property of the color space until the hidden-GC fraction is measured directly (e.g., from ACSFCS photometric GC candidates or a GCLF model) and label-noise sensitivity is reported.
  2. [§4; §3.1.2–3.2] The PCA and AE transformations appear to be fit on the entire catalog before the stratified train/test split ('We ran PCA and AEs over our LSST-like photometric catalog'). If so, the test-set features depend on test-set statistics, leaking information and biasing the comparison between raw colors and transformed inputs in favor of the PCs/LSCs. The CV protocol should fit PCA and train AEs inside each training fold only, then apply the fitted transform to the held-out fold. The paper should state explicitly where the transforms are fit and recompute Table 5 with a leakage-free protocol.
  3. [Table 5 and §5] The central numerical claims are quoted without uncertainty. The comparison '~45% vs ~35% contamination' corresponds to P=0.57 vs P=0.65 from a single best model per input set; there is no fold-to-fold variance, no repeated-CV error, and the best model is selected by F1 on the test set. The 4-PC result (P=0.56) is not formally distinguishable from the 15-color result (P=0.57). Report mean +/- standard deviation across CV folds (or repeated CV) and a test of significance for the PC-vs-color difference.
  4. [Abstract vs §5] The abstract quotes a minimum contamination of ~30% for all 15 colors, while Section 5 and the results text state ~45% for colors and ~35% for PCs. Table 5 gives P=0.57 (43% contamination) for 15 colors and P=0.65 (35%) for 15 PCs, so the abstract's 30% is not supported by the table. Align the abstract with the body, ideally after recomputing with the leakage-free protocol.
minor comments (5)
  1. [§4] The sentence 'On the other hand, the RFC that uses 4 LSCs as input shows a Notably, all 2D inputs yield...' is grammatically broken and should be rephrased.
  2. [§2.2] After correcting the GCLF arithmetic, rewrite the motivation for the 21.0 cut; the current paragraph simultaneously justifies the cut as leaving only a small fraction of GCs and as containing 35-40% of the GC population.
  3. [§3.1.1] The selection of the 4 and 2 most important colors uses RFC feature importance from a model trained on 15 colors; clarify whether this selection was performed inside each CV fold to avoid using test information in feature selection.
  4. [§5] 'Minimum contamination' should be qualified as 'minimum among the tested configurations', since no formal optimization over decision thresholds is performed.
  5. [Tables 6–7] The supplementary tables should adopt the same uncertainty conventions as Table 5 once they are added.

Circularity Check

0 steps flagged

No significant circularity: the central contamination/completeness claims are empirical classifier measurements on externally labeled data, not derivations from their own inputs.

full rationale

The paper's main claims—minimum contamination ~45% with colors and ~35% with principal components—are read directly from classifier precision in Table 5 (e.g., RFC 15 PCs: P=0.65, R=0.19; RFC 15 colors: P=0.57, R=0.20). These are measured confusion-matrix outputs from models trained with 5-fold stratified cross-validation on labels coming from external sources: spectroscopically confirmed GCs from Schuberth et al. (2010) and Pota et al. (2018), and DES EXTENDED_COADD morphology plus magnitude cuts for stars and galaxies. No equation defines the reported contamination/precision as equal to a fitted input, and no fitted parameter is renamed as a prediction. The PCA/AE motivation cites prior work by some of the same authors (Chies-Santos et al. 2022; de Souza et al. 2022), but the transformations themselves are standard SVD-based PCA (cited to Pedregosa et al. 2011 and Jolliffe & Cadima 2016) and standard autoencoders; those self-citations are background motivation, not the load-bearing justification for the results. The manuscript does contain an internal contradiction in Section 2.2 about the GCLF fraction of GCs inside the stellar cut (first 'roughly 35−40%', then '~10% if the cut is at 21 mag'), and the PCA/AE/feature-selection steps may have been applied before the train/test split, which would be data leakage. Both are validity and label-purity concerns that could bias the reported metrics, but neither makes the conclusion equivalent to its inputs by construction. There is no self-citation chain used as a uniqueness argument and no ansatz smuggled in via citation. The honest finding is therefore no significant circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central numbers are measured from a labeled catalog, not derived from a model. The free parameters are the data-selection and modeling choices (magnitude cuts, error threshold, match radius, AE dimensions, and the scope over which PCA/AE are fitted) that define the sample and the features. The main unvalidated premises are the purity of the star label and the fidelity of the FDS+DES combination as a surrogate for LSST photometry.

free parameters (7)
  • Star sample i-band magnitude cut = 19.0 ≤ mag_auto_i ≤ 21.0
    Defines the foreground-star class (Eq. 2). Its purity directly sets the measured GC-vs-star confusion; the paper's own contamination estimate for this cut is contradictory (35–40% vs ~10% of GCs).
  • Galaxy sample i-band magnitude cut = 19.0 ≤ mag_auto_i ≤ 22.5
    Defines the galaxy class (Eq. 1). A brighter limit would change how many compact galaxies are included in the contaminant pool.
  • Magnitude-error threshold = 0.5 mag
    Sources with any band error >0.5 mag are dropped (Section 2.4), removing ~45% of the matched sample; if error correlates with source class, the labeled catalog and measured contamination shift.
  • Cross-match radius = 1.0 arcsec
    DES×FDS matching radius based on median FWHM; it determines which 1129 of 1342 spectroscopically confirmed GCs are available (Section 2.3).
  • PCA/AE fit scope = full catalog (before CV split)
    The transforms are run on all data before classification (Section 4), so test-fold sources influence the principal components/latent axes — a mild information leak.
  • AE latent dimensions = 2 and 4
    Chosen for direct comparison with 2/4 PCs; the negative AE result is conditional on this shallow architecture (Section 3.1.3).
  • Classifier hyperparameters = randomized-search max for F1 (RFC), fixed widths for MLPC
    Reported metrics are point estimates for the best models; no fold-to-fold variance or hyperparameter uncertainty is given (Section 3.2).
axioms (5)
  • domain assumption The star label is clean enough: DES morphology cut plus 19.0≤i≤21.0 isolates foreground stars without major GC contamination
    Stated in Section 2.2 but contradicted internally: the text first says the stellar sample contains 35–40% of the Fornax GC population, then ~10%. If many GCs are labeled as stars, the classifier is trained partly on mislabeled data.
  • domain assumption FDS ugri + DES grizY is a faithful proxy for LSST ugrizy photometry
    Section 2.3; filters, apertures, depths differ from LSST; the quantitative 'LSST-like' claims depend on this equivalence.
  • domain assumption Complete-case removal of missing/error-flagged sources is ignorable
    Section 2.4 removes 9% missing Y-band and >0.5 mag-error sources (55% of the matched sample); if missingness correlates with source type, the class distributions and metrics are biased.
  • domain assumption The spectroscopically confirmed GC set is representative of the full GC population in color space
    Section 2.1; only 1129/1342 spectroscopically confirmed GCs cross-match to DES, and spectroscopy is biased toward brighter/more central clusters.
  • domain assumption Fornax is representative of other galaxies' GC systems
    The benchmark uses only the Fornax cluster around NGC 1399; color distributions and contamination levels may differ with host galaxy mass and environment.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of The contribution of the color space in LSST-like photometry for the selection of extragalactic globular cluster candidates." pith.science (2026). https://pith.science/paper/BLDFO7DV

@misc{pith2026251217644,
  author       = {Pith},
  title        = {Pith review of: The contribution of the color space in LSST-like photometry for the selection of extragalactic globular cluster candidates},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLDFO7DV}},
  note         = {Machine review of arXiv:2512.17644}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Globular clusters (GCs) are excellent tracers of their host galaxies' evolutionary histories. Traditional methods for identifying GCs in galaxies rely on cuts over photometric catalogs and can yield source lists with high levels of contamination from compact background galaxies and foreground stars. In an era when large-scale sky surveys produce photometry for millions of sources, it is essential to employ flexible and scalable tools to reliably identify GCs in external galaxies. To prepare for surveys like Rubin/LSST, we need to explore practical methodological improvements and quantify the limitations inherent in the datasets. This paper investigates the selection of point-like extragalactic GCs exclusively in the $ugrizY$ color space. We use archival data to assemble an LSST-like photometric catalog for the Fornax Cluster containing labeled confirmed GCs, galaxies, and stars. From this catalog, using principal component analysis and non-linear auto-encoders (AEs), we construct inputs to random forest and multi-layer perceptron classifiers. We show that selecting GCs using all the 15 available colors can lead to a minimum contamination rate of ~30%, whereas the use of color-color diagrams may double such rate. If only the first 4 principal components of the colors are used instead, the same minimum contamination rate is achieved without increasing incompleteness. The AEs did not improve GC identification. To further reduce contamination and extract the full potential of LSST for star cluster studies, we argue for the need to augment photometric information with ancillary data (morphology from space-based missions and near-infrared photometry) before attempting to leverage more complex models.

Figures

Figures reproduced from arXiv: 2512.17644 by Ana I. Ennis, Ana L. Chies-Santos, Charles J. Bonatto, Jennifer Sobeck, Juan P. Caso, Julia Gschwend, Katherine L. Rhode, Kristen C. Dage, Michele Cantiello, Nicholas Schweder-Souza, Niranjana P., Pauline Barmby, Pedro dos Santos-Lopes, Pedro Floriano, Rafael S. de Souza, Rubens E. G. Machado, Teymoor Saifollahi, Thayse A. Pacheco, Yasna Ordenes-Brice\~no.

Figure 1
Figure 1. Figure 1: Entire coverage of FDS in gray, the red circle indicates the 1-degree radius region around NGC 1399; our sources of interest. The black points represent the positions of spectroscopically confirmed globular clusters for which we have FDS ugri photometry available. We access DES data through the Astro Data Lab science platform (R. Nikutta et al. 2020) and, as explained in Subsection 2.1, we select sources t… view at source ↗
Figure 2
Figure 2. Figure 2: Magnitude errors plotted versus magnitudes for the bands in common between DES and FDS, gri. Black dots represent spectroscop￾ically confirmed GCs. The first row of plots refers to FDS PSF photometry data, the second to DES circular aperture photometry, and the third to DES automatic aperture (based on the Kron radius) photometry. For visualization purposes, the magnitude error axes were truncated at a val… view at source ↗
Figure 3
Figure 3. Figure 3: ∆magFDS−DES versus g,r, iFDS: the difference in magnitude for the same source, in the same band, but in different surveys against the FDS magnitude in the same band. The first row displays the plots where DES MAG APER 5 (2.92′′) data was used, the second DES MAG APER 4 (1.92′′), and the third DES MAG AUTO. The horizontal black line is y = 0. reliable compared with that of DAOphot (A. L. Chies-Santos et al.… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of magnitude values for each band. The colored bars represent the dataset before the pre-processing described in Subsection 2.4. The black edges indicate the subset that represents the dataset after all the filtering; it contains only the labeled sources. e.g., R. D’Abrusco et al. (2016) and A. L. Chies-Santos et al. (2022). R. D’Abrusco et al. (2016) selected GC candidates around NGC 1399 usi… view at source ↗
Figure 5
Figure 5. Figure 5: Diagram to illustrate the flux of data in our analysis procedure. PCs ∼ principal components, LSCs ∼ latent space coordinates, RFC ∼ random forest classifier, MLPC ∼ multi-layer perceptron classifier. 3.1. Preparing model inputs We aim to investigate whether more concise inputs could affect the identification of GCs. For this, we prepare input sets to the models using different numbers of colors, PCs, and … view at source ↗
Figure 6
Figure 6. Figure 6: Projections in the color space, PC space, and the non-linear AE latent space of our LSST-like filtered and labeled photometric catalog. The plots show that there are no qualitative differences in the distribution of the points in these spaces. the difference between the input data and its reconstructed version from the lower-dimensional latent space to obtain the best compression possible. The dimension of… view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices of the RFC and MLPC that received the set of all 15 PCs as input. It is also possible to demonstrate that PCA has some compression efficacy for this dataset; this is supported by the fact that, with RFCs, the precision scores obtained with the PCs input are greater than those obtained with the colors input, except for the comparison between the 2 colors input and the 2 PCs input. In part… view at source ↗
Figure 8
Figure 8. Figure 8: ∆magFDS−DES vs g,r, iFDS: the difference in magnitude for the same source, in the same band, but in different surveys against the FDS magnitude in the same band. The first row displays the plots where DES MAG APER 5 (2.92′′) data was used, the second DES MAG APER 4 (1.92′′), and the third DES MAG AUTO. Black points background galaxies selected via criteria (1) as in Subsection 2.2. The horizontal black lin… view at source ↗
Figure 9
Figure 9. Figure 9: ∆magFDS−DES vs g,r, iFDS : the difference in magnitude for the same source, in the same band, but in different surveys against the FDS magnitude in the same band. The first row displays the plots where DES MAG APER 5 (2.92′′) data was used, the second DES MAG APER 4 (1.92′′), and the third DES MAG AUTO. Black points foreground stars selected via criteria (2) as in Subsection 2.2. The horizontal black line … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

73 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Abbott, T. M. C., Adam´ow, M., Aguena, M., et al. 2021, ApJS, 255, 20, doi: 10.3847/1538-4365/ac00b3

    doi:10.3847/1538-4365/ac00b3 2021

  2. [2]

    2018, PASJ, 70, S8, doi: 10.1093/pasj/psx081

    Aihara, H., Armstrong, R., Bickerton, S., et al. 2018, PASJ, 70, S8, doi: 10.1093/pasj/psx081

    doi:10.1093/pasj/psx081 2018

  3. [3]

    2019, arXiv e-prints, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Akeson, R., Armus, L., Bachelet, E., et al. 2019, arXiv e-prints, arXiv:1902.05569, doi: 10.48550/arXiv.1902.05569

    Pith/arXiv arXiv doi:10.48550/arxiv.1902.05569 2019

  4. [4]

    S., Tully, R

    Anand, G. S., Tully, R. B., Cohen, Y ., et al. 2024, ApJ, 973, 83, doi: 10.3847/1538-4357/ad64c7

    doi:10.3847/1538-4357/ad64c7 2024

  5. [5]

    Austin, P. C. 2011, Multivariate Behav Res., 46, 399, doi: 10.1080/00273171.2011.568786

    arXiv 2011

  6. [6]

    2020, arXiv e-prints, arXiv:2003.05991, doi: 10.48550/arXiv.2003.05991

    Bank, D., Koenigstein, N., & Giryes, R. 2020, arXiv e-prints, arXiv:2003.05991, doi: 10.48550/arXiv.2003.05991

    Pith/arXiv arXiv doi:10.48550/arxiv.2003.05991 2020

  7. [7]

    C., et al

    Barbisan, E., Huang, J., Dage, K. C., et al. 2022, MNRAS, 514, 943, doi: 10.1093/mnras/stac1396

    work page doi:10.1093/mnras/stac1396 2022

  8. [8]

    Baumgardt, H., & Hilker, M. 2018, MNRAS, 478, 1520, doi: 10.1093/mnras/sty1057 16 15 20 25 4 2 0 2 4 magFDS DES DES selected galaxies g band DES MAG_APER_5 15 20 25 r band 15.0 17.5 20.0 22.5 25.0 i band 15 20 25 4 2 0 2 4 magFDS DES DES MAG_APER_4 15 20 25 15.0 17.5 20.0 22.5 25.0 15 20 25 gFDS [mag] 4 2 0 2 4 magFDS DES DES MAG_AUTO 15 20 25 rFDS [mag] ...

    doi:10.1093/mnras/sty1057 2018

  9. [9]

    Beasley, M. A. 2020, in Reviews in Frontiers of Modern Astrophysics; From Space Debris to Cosmology, ed. P. Kab´ath, D. Jones, & M. Skarka, 245–277, doi: 10.1007/978-3-030-38509-5 9

    doi:10.1007/978-3-030-38509-5 2020

  10. [10]

    M., Windhorst, R

    Berkheimer, J. M., Windhorst, R. A., Harris, W. E., et al. 2025, arXiv e-prints, arXiv:2508.03883, doi: 10.48550/arXiv.2508.03883

    work page doi:10.48550/arxiv.2508.03883 2025

  11. [11]

    1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

    Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

    doi:10.1051/aas:1996164 1996

  12. [12]

    B., Ivezi´c, ˇZ., Jones, R

    Bianco, F. B., Ivezi´c, ˇZ., Jones, R. L., et al. 2022, ApJS, 258, 1, doi: 10.3847/1538-4365/ac3e72

    doi:10.3847/1538-4365/ac3e72 2022

  13. [13]

    2015, arXiv e-prints, arXiv:1511.05741, doi: 10.48550/arXiv.1511.05741

    Biau, G., & Scornet, E. 2015, arXiv e-prints, arXiv:1511.05741, doi: 10.48550/arXiv.1511.05741

    Pith/arXiv arXiv doi:10.48550/arxiv.1511.05741 2015

  14. [14]

    2001, Machine Learning, 45, 5, doi: 10.1023/A:1010933404324

    Breiman, L. 2001, Machine Learning, 45, 5, doi: 10.1023/A:1010933404324

    doi:10.1023/a:1010933404324 2001

  15. [15]

    P., Romanowsky, A

    Brodie, J. P., Romanowsky, A. J., Strader, J., & Forbes, D. A. 2011, AJ, 142, 199, doi: 10.1088/0004-6256/142/6/199

    doi:10.1088/0004-6256/142/6/199 2011

  16. [16]

    P., Romanowsky, A

    Brodie, J. P., Romanowsky, A. J., Strader, J., et al. 2014, ApJ, 796, 52, doi: 10.1088/0004-637X/796/1/52

    doi:10.1088/0004-637x/796/1/52 2014

  17. [17]

    Burkert, A., & Forbes, D. A. 2020, AJ, 159, 56, doi: 10.3847/1538-3881/ab5b0e

    doi:10.3847/1538-3881/ab5b0e 2020

  18. [18]

    A., Chies-Santos, A

    Canossa-Gosteinski, M. A., Chies-Santos, A. L., Furlanetto, C., et al. 2024, MNRAS, 534, 1729, doi: 10.1093/mnras/stae2161 17 15 20 25 4 2 0 2 4 magFDS DES DES selected stars g band DES MAG_APER_5 15 20 25 r band 15.0 17.5 20.0 22.5 25.0 i band 15 20 25 4 2 0 2 4 magFDS DES DES MAG_APER_4 15 20 25 15.0 17.5 20.0 22.5 25.0 15 20 25 gFDS [mag] 4 2 0 2 4 mag...

    doi:10.1093/mnras/stae2161 2024

  19. [19]

    2018, A&A, 611, A21, doi: 10.1051/0004-6361/201731325

    Cantiello, M., Grado, A., Rejkuba, M., et al. 2018, A&A, 611, A21, doi: 10.1051/0004-6361/201731325

    doi:10.1051/0004-6361/201731325 2018

  20. [20]

    2020, A&A, 639, A136, doi: 10.1051/0004-6361/202038137

    Cantiello, M., Venhola, A., Grado, A., et al. 2020, A&A, 639, A136, doi: 10.1051/0004-6361/202038137

    work page doi:10.1051/0004-6361/202038137 2020

  21. [21]

    Castelli, A. V . 2014, MNRAS, 442, 891, doi: 10.1093/mnras/stu876

    work page doi:10.1093/mnras/stu876 2014

  22. [22]

    L., Larsen, S

    Chies-Santos, A. L., Larsen, S. S., Kuntschner, H., et al. 2011a, A&A, 525, A20, doi: 10.1051/0004-6361/201015683

    doi:10.1051/0004-6361/201015683

  23. [23]

    L., Larsen, S

    Chies-Santos, A. L., Larsen, S. S., Wehner, E. M., et al. 2011b, A&A, 525, A19, doi: 10.1051/0004-6361/201015681

    work page doi:10.1051/0004-6361/201015681

  24. [24]

    L., de Souza, R

    Chies-Santos, A. L., de Souza, R. S., Caso, J. P., et al. 2022, MNRAS, 516, 1320, doi: 10.1093/mnras/stac2002 Cˆot´e, P., Blakeslee, J. P., Ferrarese, L., et al. 2004, ApJS, 153, 223, doi: 10.1086/421490 CSST Collaboration, Gong, Y ., Miao, H., et al. 2025, arXiv e-prints, arXiv:2507.04618, doi: 10.48550/arXiv.2507.04618 D’Abrusco, R., Cantiello, M., Paol...

    doi:10.48550/arxiv.2507.04618 2022

  25. [25]

    C., Zepf, S

    Dage, K. C., Zepf, S. E., Thygesen, E., et al. 2020, MNRAS, 497, 596, doi: 10.1093/mnras/staa1963 de Souza, R. S., Maio, U., Biffi, V ., & Ciardi, B. 2014, MNRAS, 440, 240, doi: 10.1093/mnras/stu274 18 RFC MLPC 15 Colors (0.97, 0.99, 0.98) (0.99, 0.85, 0.92) 4 Colors (0.97, 0.98, 0.97) (0.99, 0.86, 0.92) 2 Colors (0.95, 0.97, 0.96) (0.99, 0.81, 0.89) 15 P...

    arXiv 2020

  26. [26]

    M., Pascale, M., Frye, B., et al

    Diego, J. M., Pascale, M., Frye, B., et al. 2023, A&A, 679, A159, doi: 10.1051/0004-6361/202345868

    doi:10.1051/0004-6361/202345868 2023

  27. [27]

    2022, A&A, 663, A81, doi: 10.1051/0004-6361/202243354

    Dold, D., & Fahrion, K. 2022, A&A, 663, A81, doi: 10.1051/0004-6361/202243354

    work page doi:10.1051/0004-6361/202243354 2022

  28. [28]

    Dornan, V ., & Harris, W. E. 2025, ApJ, 988, 70, doi: 10.3847/1538-4357/ade05e

    doi:10.3847/1538-4357/ade05e 2025

  29. [29]

    2025, arXiv e-prints, arXiv:2509.21975, doi: 10.48550/arXiv.2509.21975 Euclid Collaboration, Scaramella, R., Amiaux, J., et al

    Dou, H., Li, H., Zhang, H., Yu, H., & Wang, H. 2025, arXiv e-prints, arXiv:2509.21975, doi: 10.48550/arXiv.2509.21975 Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022, A&A, 662, A112, doi: 10.1051/0004-6361/202141938 Euclid Collaboration, V oggel, K., Lanc ¸on, A., et al. 2025, A&A, 693, A251, doi: 10.1051/0004-6361/202450851

    work page doi:10.48550/arxiv.2509.21975 2025

  30. [30]

    2020, A&A, 637, A27, doi: 10.1051/0004-6361/202037686

    Fahrion, K., Lyubenova, M., Hilker, M., et al. 2020, A&A, 637, A27, doi: 10.1051/0004-6361/202037686

    doi:10.1051/0004-6361/202037686 2020

  31. [31]

    2012, ApJS, 200, 4, doi: 10.1088/0067-0049/200/1/4

    Ferrarese, L., Cˆot´e, P., Cuillandre, J.-C., et al. 2012, ApJS, 200, 4, doi: 10.1088/0067-0049/200/1/4

    doi:10.1088/0067-0049/200/1/4 2012

  32. [32]

    A., Read, J

    Forbes, D. A., Read, J. I., Gieles, M., & Collins, M. L. M. 2018, MNRAS, 481, 5592, doi: 10.1093/mnras/sty2584

    doi:10.1093/mnras/sty2584 2018

  33. [33]

    Empirical comparison between autoencoders and traditional dimensionality reduction methods

    Fournier, Q., & Aloise, D. 2021, arXiv e-prints, arXiv:2103.04874, doi: 10.48550/arXiv.2103.04874 Gaia Collaboration, Prusti, T., de Bruijne, J. H. J., et al. 2016, A&A, 595, A1, doi: 10.1051/0004-6361/201629272

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2103.04874 2021

  34. [34]

    H., & Mechelli, A

    Garcia-Dias, R., Vieira, S., Lopez Pinaya, W. H., & Mechelli, A. 2020, in Machine Learning, ed. A. Mechelli & S. Vieira (Academic Press), 227–247, doi: https://doi.org/10.1016/B978-0-12-815739-8.00013-4

    work page doi:10.1016/b978-0-12-815739-8.00013-4 2020

  35. [35]

    R., & Rhode, K

    Hargis, J. R., & Rhode, K. L. 2012, AJ, 144, 164, doi: 10.1088/0004-6256/144/6/164

    work page doi:10.1088/0004-6256/144/6/164 2012

  36. [36]

    E., Harris, G

    Harris, W. E., Harris, G. L. H., & Alessi, M. 2013, ApJ, 772, 82, doi: 10.1088/0004-637X/772/2/82

    doi:10.1088/0004-637x/772/2/82 2013

  37. [37]

    E., Morningstar, W., Gnedin, O

    Harris, W. E., Morningstar, W., Gnedin, O. Y ., et al. 2014, ApJ, 797, 128, doi: 10.1088/0004-637X/797/2/128

    doi:10.1088/0004-637x/797/2/128 2014

  38. [38]

    E., Imai, K., King, G., & Stuart, E

    Ho, D. E., Imai, K., King, G., & Stuart, E. A. 2007, Political Analysis, 15, 199–236, doi: 10.1093/pan/mpl013

    doi:10.1093/pan/mpl013 2007

  39. [39]

    M., Ferguson, A

    Howell, J. M., Ferguson, A. M. N., Larsen, S. S., et al. 2025, arXiv e-prints, arXiv:2509.10440, doi: 10.48550/arXiv.2509.10440

    doi:10.48550/arxiv.2509.10440 2025

  40. [40]

    J., Harris, G

    Hudson, M. J., Harris, G. L., & Harris, W. E. 2014, ApJL, 787, L5, doi: 10.1088/2041-8205/787/1/L5

    doi:10.1088/2041-8205/787/1/l5 2014

  41. [41]

    K., Sand, D

    Hughes, A. K., Sand, D. J., Seth, A., et al. 2021, ApJ, 914, 16, doi: 10.3847/1538-4357/abf63c Ivezi´c, ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c

    doi:10.3847/1538-4357/abf63c 2021

  42. [42]

    T., & Cadima, J

    Jolliffe, I. T., & Cadima, J. 2016, Philosophical Transactions of the Royal Society of London Series A, 374, 20150202, doi: 10.1098/rsta.2015.0202 Jord´an, A., Peng, E. W., Blakeslee, J. P., et al. 2015, ApJS, 221, 13, doi: 10.1088/0067-0049/221/1/13 19 Jord´an, A., Blakeslee, J. P., Cˆot´e, P., et al. 2007, ApJS, 169, 213, doi: 10.1086/512778

    arXiv 2016

  43. [43]

    2022, Nature, 602, 585, doi: 10.1038/s41586-021-04354-w

    Kirsten, F., Marcote, B., Nimmo, K., et al. 2022, Nature, 602, 585, doi: 10.1038/s41586-021-04354-w

    doi:10.1038/s41586-021-04354-w 2022

  44. [44]

    A., de Souza, R

    Kuhn, M. A., de Souza, R. S., Krone-Martins, A., et al. 2021, ApJS, 254, 33, doi: 10.3847/1538-4365/abe465

    doi:10.3847/1538-4365/abe465 2021

  45. [45]

    S., Ferguson, A

    Larsen, S. S., Ferguson, A. M. N., Howell, J. M., et al. 2025, A&A, 703, A113, doi: 10.1051/0004-6361/202554661

    doi:10.1051/0004-6361/202554661 2025

  46. [46]

    W., Cˆot´e, P., et al

    Lim, S., Peng, E. W., Cˆot´e, P., et al. 2025, ApJS, 276, 34, doi: 10.3847/1538-4365/ad97b7

    doi:10.3847/1538-4365/ad97b7 2025

  47. [47]

    J., Kundu, A., Zepf, S

    Maccarone, T. J., Kundu, A., Zepf, S. E., & Rhode, K. L. 2007, Nature, 445, 183, doi: 10.1038/nature05434

    doi:10.1038/nature05434 2007

  48. [48]

    L., Jord´an, A., Cˆot´e, P., et al

    Masters, K. L., Jord´an, A., Cˆot´e, P., et al. 2010, ApJ, 715, 1419, doi: 10.1088/0004-637X/715/2/1419

    work page doi:10.1088/0004-637x/715/2/1419 2010

  49. [49]

    2024, A&A, 691, A104, doi: 10.1051/0004-6361/202451273

    Mirabile, M., Cantiello, M., Lonare, P., et al. 2024, A&A, 691, A104, doi: 10.1051/0004-6361/202451273

    doi:10.1051/0004-6361/202451273 2024

  50. [50]

    2022, Astronomy and Computing, 39, 100555, doi: 10.1016/j.ascom.2022.100555 Mu˜noz, R

    Mohammadi, M., Mutatiina, J., Saifollahi, T., & Bunte, K. 2022, Astronomy and Computing, 39, 100555, doi: 10.1016/j.ascom.2022.100555 Mu˜noz, R. P., Puzia, T. H., Lanc ¸on, A., et al. 2014, ApJS, 210, 4, doi: 10.1088/0067-0049/210/1/4

    arXiv 2022

  51. [51]

    1991, Neurocomputing, 2, 183, doi: https://doi.org/10.1016/0925-2312(91)90023-5

    Murtagh, F. 1991, Neurocomputing, 2, 183, doi: https://doi.org/10.1016/0925-2312(91)90023-5

    doi:10.1016/0925-2312(91)90023-5 1991

  52. [52]

    Nikutta, R., Fitzpatrick, M., Scott, A., & Weaver, B. A. 2020, Astronomy and Computing, 33, 100411, doi: 10.1016/j.ascom.2020.100411

    arXiv 2020

  53. [53]

    2000, A&AS, 143, 23, doi: 10.1051/aas:2000169

    Ochsenbein, F., Bauer, P., & Marcout, J. 2000, A&AS, 143, 23, doi: 10.1051/aas:2000169

    doi:10.1051/aas:2000169 2000

  54. [54]

    A., Coelho, P

    Pacheco, T. A., Coelho, P. R. T., Martins, L. P., et al. 2025, ApJ, 992, 151, doi: 10.3847/1538-4357/adff5f

    doi:10.3847/1538-4357/adff5f 2025

  55. [55]

    2011, Journal of Machine Learning Research, 12, 2825

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

    2011

  56. [56]

    2020, https://arxiv.org/abs/2008.12633

    Peletier, R., Iodice, E., Venhola, A., et al. 2020, https://arxiv.org/abs/2008.12633

    arXiv 2020

  57. [57]

    W., Jord´an, A., Cˆot´e, P., et al

    Peng, E. W., Jord´an, A., Cˆot´e, P., et al. 2006, ApJ, 639, 95, doi: 10.1086/498210

    doi:10.1086/498210 2006

  58. [58]

    R., Hilker, M., et al

    Pota, V ., Napolitano, N. R., Hilker, M., et al. 2018, MNRAS, 481, 1744, doi: 10.1093/mnras/sty2149

    doi:10.1093/mnras/sty2149 2018

  59. [59]

    L., et al

    Reina-Campos, M., Trujillo-Gomez, S., Pfeffer, J. L., et al. 2023, MNRAS, 521, 6368, doi: 10.1093/mnras/stad920

    doi:10.1093/mnras/stad920 2023

  60. [60]

    2012, Ap&SS, 341, 195, doi: 10.1007/s10509-012-0986-9

    Rejkuba, M. 2012, Ap&SS, 341, 195, doi: 10.1007/s10509-012-0986-9

    doi:10.1007/s10509-012-0986-9 2012

  61. [61]

    2022, MNRAS, 511, 4633, doi: 10.1093/mnras/stac328

    Saifollahi, T., Zaritsky, D., Trujillo, I., et al. 2022, MNRAS, 511, 4633, doi: 10.1093/mnras/stac328

    doi:10.1093/mnras/stac328 2022

  62. [62]

    F., et al

    Saifollahi, T., Janz, J., Peletier, R. F., et al. 2021, MNRAS, 504, 3580, doi: 10.1093/mnras/stab1118

    work page doi:10.1093/mnras/stab1118 2021

  63. [63]

    2025a, A&A, 697, A10, doi: 10.1051/0004-6361/202450784

    Saifollahi, T., V oggel, K., Lanc ¸on, A., et al. 2025a, A&A, 697, A10, doi: 10.1051/0004-6361/202450784

    doi:10.1051/0004-6361/202450784

  64. [64]

    2025b, arXiv e-prints, arXiv:2503.16367, doi: 10.48550/arXiv.2503.16367

    Saifollahi, T., Lanc ¸on, A., Cantiello, M., et al. 2025b, arXiv e-prints, arXiv:2503.16367, doi: 10.48550/arXiv.2503.16367

    doi:10.48550/arxiv.2503.16367

  65. [65]

    F., & Finkbeiner, D

    Schlafly, E. F., & Finkbeiner, D. P. 2011, ApJ, 737, 103, doi: 10.1088/0004-637X/737/2/103

    doi:10.1088/0004-637x/737/2/103 2011

  66. [66]

    J., Finkbeiner, D

    Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525, doi: 10.1086/305772

    doi:10.1086/305772 1998

  67. [67]

    2010, A&A, 513, A52, doi: 10.1051/0004-6361/200912482

    Schuberth, Y ., Richtler, T., Hilker, M., et al. 2010, A&A, 513, A52, doi: 10.1051/0004-6361/200912482

    doi:10.1051/0004-6361/200912482 2010

  68. [68]

    P., Forbes, D

    Usher, C., Brodie, J. P., Forbes, D. A., et al. 2019, MNRAS, 490, 491, doi: 10.1093/mnras/stz2596

    doi:10.1093/mnras/stz2596 2019

  69. [69]

    C., Girardi, L., et al

    Usher, C., Dage, K. C., Girardi, L., et al. 2023, PASP, 135, 074201, doi: 10.1088/1538-3873/ace3f7

    doi:10.1088/1538-3873/ace3f7 2023

  70. [70]

    2021, MNRAS, 505, 5815, doi: 10.1093/mnras/stab1701

    Burkert, A. 2021, MNRAS, 505, 5815, doi: 10.1093/mnras/stab1701

    doi:10.1093/mnras/stab1701 2021

  71. [71]

    M., Remus, R.-S., McKenzie, M., & Forbes, D

    Valenzuela, L. M., Remus, R.-S., McKenzie, M., & Forbes, D. A. 2024, A&A, 687, A104, doi: 10.1051/0004-6361/202348010 V oggel, K. T., Seth, A. C., Sand, D. J., et al. 2020, ApJ, 899, 140, doi: 10.3847/1538-4357/ab6f69

    doi:10.1051/0004-6361/202348010 2024

  72. [72]

    S., et al

    Xu, Q., Shen, S., de Souza, R. S., et al. 2023, MNRAS, 526, 6391, doi: 10.1093/mnras/stad3181

    doi:10.1093/mnras/stad3181 2023

  73. [73]

    2022, MNRAS, 513, 2609, doi: 10.1093/mnras/stac1072

    Zaritsky, D. 2022, MNRAS, 513, 2609, doi: 10.1093/mnras/stac1072

    doi:10.1093/mnras/stac1072 2022

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.