pith. sign in

arxiv: 1906.09382 · v1 · pith:QH6HHPUKnew · submitted 2019-06-22 · 🌌 astro-ph.CO · astro-ph.GA· cs.LG· stat.ML

A Halo Merger Tree Generation and Evaluation Framework

Sandra Robles , Jonathan S. G\'omez , Ad\'in Ram\'irez Rivera , Jenny A. Gonz\'alez , Nelson D. Padilla , Diego Dujovne This is my paper

Pith reviewed 2026-05-25 18:27 UTC · model grok-4.3

classification 🌌 astro-ph.CO astro-ph.GAcs.LGstat.ML
keywords halo merger treesgenerative adversarial networkssemi-analytic modelscosmological simulationsgalaxy formationEAGLE simulationmatrix generation
0
0 comments X

The pith

A generative adversarial network can generate realistic halo merger trees by learning from simulation matrices.

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

The paper sets out to produce halo merger trees that display the smooth mass growth and spatial continuity found in large-volume cosmological simulations. It does so by recasting tree construction as a matrix generation task and training a generative adversarial network on outputs from the EAGLE suite. If the method works, semi-analytic models of galaxy formation could obtain many plausible trees at far lower cost than running full hydrodynamic simulations. A sympathetic reader would care because merger trees are the central input that lets these models test theories of galaxy evolution against observations.

Core claim

Halo merger tree construction can be treated as a matrix generation problem. A generative adversarial network trained on merger trees from the EAGLE simulation suite produces new trees that exhibit realistic features, specifically the absence of drastic changes in halo mass or jumps in physical locations.

What carries the argument

A generative adversarial network trained to output matrices that represent halo merger trees and to match their statistical properties from simulation data.

If this is right

  • Semi-analytic models gain access to large numbers of realistic merger trees without the expense of full simulations.
  • The generated trees maintain smooth mass evolution and spatial continuity matching the training data.
  • The framework operates at modest computational cost relative to direct simulation outputs.
  • Quality can be assessed by direct statistical comparison to EAGLE merger trees.

Where Pith is reading between the lines

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

  • If the network has learned the underlying distribution, the same approach could produce trees for cosmologies or initial conditions different from those in the training set.
  • The matrix representation might allow the method to be applied to other tree-structured data in astrophysics, such as galaxy assembly histories.
  • Testing the generated trees against hydrodynamic simulations run with different codes would reveal whether the captured features are universal or tied to EAGLE specifics.

Load-bearing premise

The statistical features of halo merger trees can be captured by a GAN trained only on simulation outputs without explicit physical constraints or validation on other simulations.

What would settle it

Generate trees with the trained network and compare them to merger trees from an independent simulation suite; the claim fails if the generated trees show frequent unphysical mass jumps or position discontinuities absent in the independent set.

Figures

Figures reproduced from arXiv: 1906.09382 by Ad\'in Ram\'irez Rivera, Diego Dujovne, Jenny A. Gonz\'alez, Jonathan S. G\'omez, Nelson D. Padilla, Sandra Robles.

Figure 1
Figure 1. Figure 1: Top: Generated merger tree, the color map denotes pro￾genitor masses and the progenitor type is represented by circles (main halos) and triangles (subhalos). Bottom: Same merger tree in the plane snapshot vs. distance. model, based on GANs, that generates merger trees closer to the training data. To that end, first, we represent merger trees as matrices. Next, we propose a learning framework to train our m… view at source ↗
Figure 2
Figure 2. Figure 2: Halo merger tree generation process. tion supplies a large enough amount of trees for training purposes. We obtained the trees by traversing the subhalo table of the EAGLE public catalog (McAlpine et al., 2016). The EAGLE simulation suite is a set of cosmological hy￾drodynamical simulations with cosmological parameters taken from Planck Collaboration et al. (2014), that track the evolution of both, dark ma… view at source ↗
Figure 4
Figure 4. Figure 4: Normalized cumulative distribution of the averaged prob￾abilities of the distances for the real merger trees and those gener￾ated with 2 and 3 variables. first row of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Top: Cumulative probability of the mass gain and loss for all the progenitors in all the merger trees in the training dataset (real), and in the samples of generated merger trees with 1, 2, and 3 variables. Bottom: Same for the progenitors that are main halos (MH) and subhalos (SH) in all the merger trees in the train￾ing dataset (real), and in the generated merger trees with three variables. the 2 variabl… view at source ↗
Figure 5
Figure 5. Figure 5: Normalized cumulative distribution of the number of snapshots that a progenitor spend as a subhalo before the merger takes place. number a snapshot a progenitor spend as subhalo before merging with another progenitor, for generated and real merger trees. Once again, we compared both distribution with the KS test (see last row of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

Semi-analytic models are best suited to compare galaxy formation and evolution theories with observations. These models rely heavily on halo merger trees, and their realistic features (i.e., no drastic changes on halo mass or jumps on physical locations). Our aim is to provide a new framework for halo merger tree generation that takes advantage of the results of large volume simulations, with a modest computational cost. We treat halo merger tree construction as a matrix generation problem, and propose a Generative Adversarial Network that learns to generate realistic halo merger trees. We evaluate our proposal on merger trees from the EAGLE simulation suite, and show the quality of the generated trees.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes treating halo merger tree construction as an unsupervised matrix generation task and introduces a Generative Adversarial Network trained on merger trees extracted from the EAGLE simulation suite; the central claim is that the resulting synthetic trees reproduce realistic features such as smooth mass evolution and spatial continuity at modest computational cost for use in semi-analytic models.

Significance. If the generated trees satisfy the necessary physical validity constraints and match independent simulation statistics, the approach would supply a fast, data-driven alternative to full N-body runs for populating merger trees in galaxy formation modeling, enabling larger parameter explorations than currently feasible.

major comments (3)
  1. [Abstract] Abstract: the claim that 'we evaluate our proposal on merger trees from the EAGLE simulation suite, and show the quality of the generated trees' is unsupported by any reported metrics, baselines, or quantitative comparison; without these the central assertion of realism cannot be assessed.
  2. [Method] Method section (matrix-generation framing): the architecture is described as learning from EAGLE matrices without explicit loss terms, architectural constraints, or post-processing to enforce tree invariants such as parent mass >= sum of child masses, consistent ancestry across time steps, or bounded position jumps; this omission leaves open the possibility that discriminator matching of aggregate statistics alone permits invalid configurations.
  3. [Evaluation] Evaluation: no cross-validation against an independent simulation suite (e.g., Illustris or Bolshoi) is described, so it remains unclear whether the GAN has learned simulation-specific artifacts rather than universal merger-tree statistics.
minor comments (2)
  1. [Abstract] The abstract contains minor grammatical issues ('changes on halo mass' should read 'changes in halo mass'; 'jumps on physical locations' should read 'jumps in physical locations').
  2. Notation for the matrix representation of trees is introduced without an explicit diagram or example showing how ancestry and mass accretion are encoded.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's comments. We address each major comment point-by-point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'we evaluate our proposal on merger trees from the EAGLE simulation suite, and show the quality of the generated trees' is unsupported by any reported metrics, baselines, or quantitative comparison; without these the central assertion of realism cannot be assessed.

    Authors: We agree with this assessment. The abstract is overly general and does not specify the evaluation approach. In the revised version, we will modify the abstract to include references to the quantitative metrics, visual comparisons, and any baselines used in the evaluation section to support the claim of quality. revision: yes

  2. Referee: [Method] Method section (matrix-generation framing): the architecture is described as learning from EAGLE matrices without explicit loss terms, architectural constraints, or post-processing to enforce tree invariants such as parent mass >= sum of child masses, consistent ancestry across time steps, or bounded position jumps; this omission leaves open the possibility that discriminator matching of aggregate statistics alone permits invalid configurations.

    Authors: This is a valid concern. The manuscript as currently written does not detail any explicit enforcement mechanisms beyond the data-driven learning. To address this, we will revise the method section to include a post-processing step that enforces the key tree invariants (mass conservation, ancestry consistency, and position continuity) after generation, and discuss how this ensures physical validity. revision: yes

  3. Referee: [Evaluation] Evaluation: no cross-validation against an independent simulation suite (e.g., Illustris or Bolshoi) is described, so it remains unclear whether the GAN has learned simulation-specific artifacts rather than universal merger-tree statistics.

    Authors: We acknowledge that testing on an independent simulation suite would provide stronger evidence of generalizability. The current evaluation is limited to EAGLE. In the revision, we will add a section discussing potential simulation-specific features and compare key statistics (such as merger rates) with published results from other simulations like Illustris where available, to partially address this. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework trains on external EAGLE data

full rationale

The paper frames halo merger tree generation as an unsupervised matrix-generation task solved by a standard GAN trained directly on merger trees extracted from the independent EAGLE simulation suite. No equations, loss terms, or evaluation metrics are shown to reduce generated outputs to quantities defined by the model's own fitted parameters; the central claim rests on the external training data and conventional adversarial training rather than any self-definitional, fitted-input-renamed-as-prediction, or self-citation load-bearing step. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities. The GAN training process implicitly involves many hyperparameters, but none are named or justified in the available text.

pith-pipeline@v0.9.0 · 5661 in / 1075 out tokens · 49129 ms · 2026-05-25T18:27:42.017264+00:00 · methodology

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

37 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    doi: 10.1109/TPAMI.2013.50

    ISSN 0162-8828. doi: 10.1109/TPAMI.2013.50. Benson, A. J. and Bower, R. Galaxy formation spanning cosmic history. Mon. Not. Roy. Astron. Soc., 405(3):1573– 1623, Jul

    work page doi:10.1109/tpami.2013.50 2013

  2. [2]

    doi: 10.1111/j.1365-2966.2010.16592.x. Bond, J. R., Cole, S., Efstathiou, G., and Kaiser, N. Excur- sion set mass functions for hierarchical gaussian fluctua- tions. Astrophys. J., 379:440,

    work page doi:10.1111/j.1365-2966.2010.16592.x 2010

  3. [3]

    Bower, R

    doi: 10.1086/170520. Bower, R. G., Benson, A. J., Malbon, R., Helly, J. C., Frenk, C. S., Baugh, C. M., Cole, S., and Lacey, C. G. Breaking the hierarchy of galaxy formation. Mon. Not. Roy. Astron. Soc. , 370:645–655, August

    work page doi:10.1086/170520

  4. [4]

    and Yang , J

    doi: 10.1111/j.1365-2966.2006.10519.x. Carlberg, R. G., Couchman, H. M. P., and Thomas, P. A. Cosmological velocity bias. Astrophys. J., 352:L29–L32, April

    work page doi:10.1111/j.1365-2966.2006.10519.x 2006

  5. [5]

    doi: 10.1086/185686. Cole, S. Modeling Galaxy Formation in Evolving Dark Matter Halos. Astrophys. J., 367:45, Jan

    work page doi:10.1086/185686

  6. [6]

    A Halo Merger Tree Generation and Evaluation Framework Cole, S., Aragon-Salamanca, A., Frenk, C

    doi: 10.1086/169600. A Halo Merger Tree Generation and Evaluation Framework Cole, S., Aragon-Salamanca, A., Frenk, C. S., Navarro, J. F., and Zepf, S. E. A recipe for galaxy formation. Mon. Not. Roy. Astron. Soc. , 271:781–806, Dec

    work page doi:10.1086/169600

  7. [7]

    Cole, S., Lacey, C

    doi: 10.1093/mnras/271.4.781. Cole, S., Lacey, C. G., Baugh, C. M., and Frenk, C. S. Hierarchical galaxy formation. Mon. Not. Roy. Astron. Soc., 319:168–204, November

    work page doi:10.1093/mnras/271.4.781

  8. [8]

    1365-8711.2000.03879.x

    doi: 10.1046/j. 1365-8711.2000.03879.x. Crain, R. A., Schaye, J., Bower, R. G., Furlong, M., Schaller, M., Theuns, T., Dalla Vecchia, C., Frenk, C. S., McCarthy, I. G., Helly, J. C., Jenkins, A., Rosas-Guevara, Y . M., White, S. D. M., and Trayford, J. W. The EAGLE simula- tions of galaxy formation: calibration of subgrid physics and model variations. Mon...

    work page doi:10.1046/j 2000

  9. [9]

    Croton, D

    doi: 10.1093/mnras/stv725. Croton, D. J., Springel, V ., White, S. D. M., De Lucia, G., Frenk, C. S., Gao, L., Jenkins, A., Kauffmann, G., Navarro, J. F., and Yoshida, N. The many lives of active galactic nuclei: cooling flows, black holes and the luminosities and colours of galaxies. Mon. Not. Roy. Astron. Soc. , 365:11–28, January

    work page doi:10.1093/mnras/stv725

  10. [10]

    A., okas E

    doi: 10.1111/j.1365-2966.2005.09675.x. De Lucia, G., Kauffmann, G., and White, S. D. M. Chemical enrichment of the intracluster and intergalactic medium in a hierarchical galaxy formation model. Mon. Not. Roy. Astron. Soc. , 349(3):1101–1116, Apr

    work page doi:10.1111/j.1365-2966.2005.09675.x 2005

  11. [11]

    M., Papadakis I

    doi: 10.1111/j.1365-2966.2004.07584.x. Diemand, J., Kuhlen, M., and Madau, P. Early Supersym- metric Cold Dark Matter Substructure. Astrophys. J., 649 (1):1–13, Sep

    work page doi:10.1111/j.1365-2966.2004.07584.x 2004

  12. [12]

    Dosovitskiy, A., Springenberg, J

    doi: 10.1086/506377. Dosovitskiy, A., Springenberg, J. T., and Brox, T. Learning to generate chairs with convolutional neural networks. In IEEE Inter . Conf. Comput. Vis., Pattern Recog. (CVPR), pp. 1538–1546, June

    work page doi:10.1086/506377

  13. [13]

    doi: 10.1109/CVPR.2015. 7298761. Elahi, P. J., Thacker, R. J., and Widrow, L. M. Peaks above the Maxwellian Sea: a new approach to finding substruc- tures in N-body haloes. Mon. Not. Roy. Astron. Soc., 418 (1):320–335, Nov

    work page doi:10.1109/cvpr.2015 2015

  14. [14]

    doi: 10.1111/j.1365-2966.2011. 19485.x. Gonzalez-Perez, V ., Lacey, C. G., Baugh, C. M., Lagos, C. D. P., Helly, J., Campbell, D. J. R., and Mitchell, P. D. How sensitive are predicted galaxy luminosities to the choice of stellar population synthesis model? Mon. Not. Roy. Astron. Soc., 439(1):264–283, Mar

    work page doi:10.1111/j.1365-2966.2011 2011

  15. [15]

    Goodfellow, I

    doi: 10.1093/mnras/stt2410. Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y . Generative adversarial networks. In Adv. Neural Inf. Process. Sys. (NeurIPS), June

    work page doi:10.1093/mnras/stt2410

  16. [16]

    J., Reeves J

    doi: 10.1111/j.1365-2966.2010.18114.x. Han, J., Jing, Y . P., Wang, H., and Wang, W. Resolving subhaloes’ lives with the Hierarchical Bound-Tracing algorithm. Mon. Not. Roy. Astron. Soc. , 427(3):2437– 2449, Dec

    work page doi:10.1111/j.1365-2966.2010.18114.x 2010

  17. [17]

    doi: 10.1111/j.1365-2966.2012.22111. x. Hatton, S., Devriendt, J. E. G., Ninin, S., Bouchet, F. R., Guiderdoni, B., and Vibert, D. GALICS- I: A hybrid n-body/semi-analytic model of hierarchical galaxy forma- tion. Mon. Not. Roy. Astron. Soc., 343:75–106, July

    work page doi:10.1111/j.1365-2966.2012.22111 2012

  18. [18]

    Jiang, F

    doi: 10.1046/j.1365-8711.2003.05589.x. Jiang, F. and van den Bosch, F. C. Generating merger trees for dark matter haloes: a comparison of methods. Mon. Not. Roy. Astron. Soc. , 440:193–207, May

    work page doi:10.1046/j.1365-8711.2003.05589.x 2003

  19. [19]

    Katz, N., Hernquist, L., and Weinberg, D

    doi: 10.1093/mnras/stu280. Katz, N., Hernquist, L., and Weinberg, D. H. Galaxies and gas in a cold dark matter universe. Astrophys. J., 399: L109–L112, November

    work page doi:10.1093/mnras/stu280

  20. [20]

    Kauffmann, G

    doi: 10.1086/186619. Kauffmann, G. and White, S. D. M. The merging history of dark matter haloes in a hierarchical universe. Mon. Not. Roy. Astron. Soc. , 261:921–928, Apr

    work page doi:10.1086/186619

  21. [21]

    Kauffmann, G., White, S

    doi: 10.1093/mnras/261.4.921. Kauffmann, G., White, S. D. M., and Guiderdoni, B. The formation and evolution of galaxies within merging dark matter haloes. Mon. Not. Roy. Astron. Soc., 264:201–218, Sep

    work page doi:10.1093/mnras/261.4.921

  22. [22]

    Kauffmann, G., Colberg, J

    doi: 10.1093/mnras/264.1.201. Kauffmann, G., Colberg, J. M., Diaferio, A., and White, S. D. M. Clustering of galaxies in a hierarchical universe—I. methods and results at z = 0 . Mon. Not. Roy. Astron. Soc. , 303:188–206, February

    work page doi:10.1093/mnras/264.1.201

  23. [23]

    T., Done C., Smith D

    doi: 10.1046/j.1365-8711.1999.02202.x. Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inf. Process. Sys. (NeurIPS), pp. 1097–1105,

    work page doi:10.1046/j.1365-8711.1999.02202.x 1999

  24. [24]

    Lacey, C

    doi: 10.1086/170625. Lacey, C. G., Baugh, C. M., Frenk, C. S., Benson, A. J., Bower, R. G., Cole, S., Gonzalez-Perez, V ., Helly, J. C., A Halo Merger Tree Generation and Evaluation Framework Lagos, C. D. P., and Mitchell, P. D. A unified mul- tiwavelength model of galaxy formation. Mon. Not. Roy. Astron. Soc. , 462(4):3854–3911, Nov

    work page doi:10.1086/170625

  25. [25]

    Lagos, C

    doi: 10.1093/mnras/stw1888. Lagos, C. d. P., Tobar, R. J., Robotham, A. S. G., Obreschkow, D., Mitchell, P. D., Power, C., and Elahi, P. J. Shark: introducing an open source, free, and flexi- ble semi-analytic model of galaxy formation. Mon. Not. Roy. Astron. Soc. , 481(3):3573–3603, Dec

    work page doi:10.1093/mnras/stw1888

  26. [26]

    Lee, J., Yi, S

    doi: 10.1093/mnras/sty2440. Lee, J., Yi, S. K., Elahi, P. J., Thomas, P. A., Pearce, F. R., Behroozi, P., Han, J., Helly, J., Jung, I., Knebe, A., Mao, Y .-Y ., Onions, J., Rodriguez-Gomez, V ., Schneider, A., Srisawat, C., and Tweed, D. Sussing merger trees : The impact of halo merger trees on galaxy properties in a semi-analytic model. Mon. Not. Roy. As...

    work page doi:10.1093/mnras/sty2440

  27. [27]

    McAlpine, S., Helly, J

    doi: 10.1093/mnras/stu2039. McAlpine, S., Helly, J. C., Schaller, M., Trayford, J. W., Qu, Y ., Furlong, M., Bower, R. G., Crain, R. A., Schaye, J., Theuns, T., Dalla Vecchia, C., Frenk, C. S., McCarthy, I. G., Jenkins, A., Rosas-Guevara, Y ., White, S. D. M., Baes, M., Camps, P., and Lemson, G. The EAGLE sim- ulations of galaxy formation: Public release ...

    work page doi:10.1093/mnras/stu2039

  28. [28]

    Muldrew, S

    doi: 10.1016/j.ascom.2016.02.004. Muldrew, S. I., Pearce, F. R., and Power, C. The accuracy of subhalo detection. Mon. Not. Roy. Astron. Soc., 410(4): 2617–2624, Feb

    work page doi:10.1016/j.ascom.2016.02.004 2016

  29. [29]

    doi: 10.1111/j.1365-2966.2010. 17636.x. Onions, J., Knebe, A., Pearce, F. R., Muldrew, S. I., Lux, H., Knollmann, S. R., Ascasibar, Y ., Behroozi, P., Elahi, P., Han, J., Maciejewski, M., Merch´an, M. E., Neyrinck, M., Ruiz, A. N., Sgr´o, M. A., Springel, V ., and Tweed, D. Subhaloes going Notts: the subhalo-finder comparison project. Mon. Not. Roy. Astron...

    work page doi:10.1111/j.1365-2966.2010 2010

  30. [30]

    and Shibahashi , H

    doi: 10.1111/j.1365-2966.2012.20947.x. Onions, J., Ascasibar, Y ., Behroozi, P., Casado, J., Elahi, P., Han, J., Knebe, A., Lux, H., Merch´an, M. E., Muldrew, S. I., Neyrinck, M., Old, L., Pearce, F. R., Potter, D., Ruiz, A. N., Sgr ´o, M. A., Tweed, D., and Yue, T. Subhaloes gone Notts: spin across subhaloes and finders. Mon. Not. Roy. Astron. Soc. , 429(...

    work page doi:10.1111/j.1365-2966.2012.20947.x 2012

  31. [31]

    Planck Collaboration, Ade, P

    doi: 10.1093/mnras/sts549. Planck Collaboration, Ade, P. A. R., Aghanim, N., Armitage-Caplan, C., Arnaud, M., Ashdown, M., Atrio- Barandela, F., Aumont, J., Baccigalupi, C., Banday, A. J., and et al. Planck 2013 results. XVI. cosmological param- eters. Astron. Astrophys., 571:A16, November

    work page doi:10.1093/mnras/sts549 2013

  32. [32]

    Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

    doi: 10.1051/0004-6361/201321591. Radford, A., Metz, L., and Chintala, S. Unsupervised rep- resentation learning with deep convolutional generative adversarial networks. arXiv, art. arXiv:1511.06434, Nov

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1051/0004-6361/201321591

  33. [33]

    Schaye, J., Crain, R

    doi: 10.1051/0004-6361/200913374. Schaye, J., Crain, R. A., Bower, R. G., Furlong, M., Schaller, M., Theuns, T., Dalla Vecchia, C., Frenk, C. S., McCarthy, I. G., Helly, J. C., Jenkins, A., Rosas-Guevara, Y . M., White, S. D. M., Baes, M., Booth, C. M., Camps, P., Navarro, J. F., Qu, Y ., Rahmati, A., Sawala, T., Thomas, P. A., and Trayford, J. The EAGLE ...

    work page doi:10.1051/0004-6361/200913374

  34. [34]

    Somerville, R

    doi: 10.1093/mnras/stu2058. Somerville, R. S. and Primack, J. R. Semi-analytic mod- elling of galaxy formation: the local universe. Mon. Not. Roy. Astron. Soc., 310:1087–1110, December

    work page doi:10.1093/mnras/stu2058

  35. [35]

    T., Done C., Smith D

    doi: 10.1046/j.1365-8711.1999.03032.x. Somerville, R. S., Hopkins, P. F., Cox, T. J., Robertson, B. E., and Hernquist, L. A semi-analytic model for the co-evolution of galaxies, black holes and active galactic nuclei. Mon. Not. Roy. Astron. Soc. , 391:481–506, De- cember

    work page doi:10.1046/j.1365-8711.1999.03032.x 1999

  36. [36]

    T., Ogilvie G

    doi: 10.1111/j.1365-2966.2008.13805.x. White, S. D. M. and Frenk, C. S. Galaxy formation through hierarchical clustering. Astrophys. J., 379:52–79, Septem- ber

    work page doi:10.1111/j.1365-2966.2008.13805.x 2008

  37. [37]

    doi: 10.1086/170483. A. Details of the GAN Architecture We selected a convolutional-based architecture, after hav- ing evaluated first a fully connected (FC) architecture with different number of layers. Even if the FC-based GAN successfully learned the merger tree structure, it failed to reproduce the correct mass range of the halos in branches different ...

    work page doi:10.1086/170483