Citations & References#

The bibtex entries for PyAutoGalaxy and its affiliated software packages can be found here, with example text for citing PyAutoGalaxy in .tex format here format here and .md format here.

PyAutoGalaxy is published in the Journal of Open Source Software and its entry in the above .bib file is under the citation key pyautogalaxy.

As shown in the examples, we would greatly appreciate it if you mention PyAutoGalaxy by name and include a link to our GitHub page!

You should also specify the non-linear search(es) you use in your analysis (e.g. Dynesty, Emcee, PySwarms, etc) in the main body of text, and delete as appropriate any packages your analysis did not use. The citations.bib file includes the citation key for all of these projects.

JAX#

PyAutoGalaxy runs on a NumPy backend by default and an optional JAX backend for just-in-time compilation, automatic differentiation, and GPU/TPU execution. If you run any analysis on the JAX path, please cite JAX under the citation key jax. JAX-specific components that are also cited under their own keys when used are optax (gradient-based optimizers, key optax), the interferometer non-uniform FFT (nufftax and its FINUFFT algorithm, keys nufftax and finufft, see below), and the critical-curve/caustic solver (Jax-Zero-Contour, key jax_zero_contour, see below).

If you use the learning-rate-free af.MultiStartProdigy search, also cite the Prodigy method under the key prodigy. It is a published algorithm in its own right, run through Optax’s implementation (optax.contrib.prodigy), so cite it alongside optax rather than in place of it:

@inproceedings{prodigy,
author = {Mishchenko, Konstantin and Defazio, Aaron},
title = {{Prodigy: An Expeditiously Adaptive Parameter-Free Learner}},
booktitle = {Proceedings of the 41st International Conference on Machine Learning},
series = {Proceedings of Machine Learning Research},
volume = {235},
pages = {35779--35804},
publisher = {PMLR},
url = {https://proceedings.mlr.press/v235/mishchenko24a.html},
year = {2024}
}

The reference implementation is at https://github.com/konstmish/prodigy and the preprint at https://arxiv.org/abs/2306.06101.

Jax-Zero-Contour#

If you use the zero-contour method for critical curve and caustic computation (the default in visualize/general.yaml via critical_curves_method: zero_contour), please cite the Jax-Zero-Contour package by Coleman Krawczyk:

@software{coleman_krawczyk_2025_15730415,
  author       = {Coleman Krawczyk},
  title        = {CKrawczyk/Jax-Zero-Contour: Version 2.0.0},
  month        = jun,
  year         = 2025,
  publisher    = {Zenodo},
  version      = {v2.0.0},
  doi          = {10.5281/zenodo.15730415},
  url          = {https://doi.org/10.5281/zenodo.15730415},
}

The package is available at https://github.com/CKrawczyk/Jax-Zero-Contour and archived at https://doi.org/10.5281/zenodo.15730415.

NUFFTax#

If you fit interferometer datasets on the JAX path, the non-uniform FFT is performed by nufftax, a pure-JAX NUFFT implementation by the GragasLab team. Please cite the package:

@software{nufftax,
  author = {Gragas and Oudoumanessah, Geoffroy and Iollo, Jacopo},
  title  = {nufftax: Pure JAX implementation of the Non-Uniform Fast Fourier Transform},
  url    = {https://github.com/GragasLab/nufftax},
  year   = {2026},
}

nufftax’s algorithm is based on FINUFFT (Flatiron Institute); the upstream paper should also be cited:

@article{finufft,
  author  = {Barnett, Alexander H. and Magland, Jeremy F. and af Klinteberg, Ludvig},
  title   = {A parallel non-uniform fast Fourier transform library based on an
             'exponential of semicircle' kernel},
  journal = {SIAM J. Sci. Comput.},
  volume  = {41},
  number  = {5},
  pages   = {C479--C504},
  year    = {2019},
}

The package is available at https://github.com/GragasLab/nufftax.

Rectangular Mesh (Pixelized Reconstructions)#

If you reconstruct a galaxy using the adaptive rectangular meshes (RectangularAdaptDensity or RectangularAdaptImage), you must cite the following paper https://arxiv.org/abs/2606.30620 under citation key Enzi2026. The mesh’s adaptive coordinate transform implements the ray-guided transformed uniform (RTU) grid formulation this paper introduces, which is what makes the pixelization adaptive and fully auto-differentiable:

@article{Enzi2026,
  author        = {Enzi, Wolfgang J. R. and Krawczyk, Coleman M. and Li, Tian and Collett, Thomas E.},
  title         = {Gaussian processes on ray-guided transformed uniform grids for fast, flexible, and auto-differentiable adaptive source reconstruction in lens modelling},
  journal       = {MNRAS, submitted},
  eprint        = {2606.30620},
  archivePrefix = {arXiv},
  primaryClass  = {astro-ph.GA},
  year          = {2026},
  url           = {https://arxiv.org/abs/2606.30620},
}

Note that PyAutoGalaxy pairs the RTU grid with its own regularization schemes (e.g. reg.Constant, reg.Adapt) rather than the Gaussian-process prior used in the paper, so quantitative results are not directly comparable between the two implementations. The RectangularUniform mesh performs no RTU transform, so this citation is not required when only the uniform mesh is used.

Dynesty#

If you used the nested sampling algorithm Dynesty, please follow the citation instructions on the dynesty readthedocs.