Bayesian sparse reconstruction

This repository contains the code used for making the results and plots in "Bayesian sparse reconstruction: a brute-force approach to astronomical imaging and machine learning" (Higson et. al, 2019).

If you have any questions then feel free to email [email protected]. However, note that this is research code and is not actively maintained.

Requirements

Generating the results in the paper requires PolyChord v1.15, plus the requirements listed in setup.py. Results in the paper were run using Python 3.6.

Set up

The code for computation using Python likelihoods and for all data processing and plotting is contained in the bsr Python module. This can be installed, along with its dependencies, by running the following command from within this repo:

pip install . --user

You can check your installation is working using the test suite by running

nosetests

from within this repo. This requires nose.

Generating and plotting results from the paper

Nested sampling runs can be generated and the results plotted using compute_results.py; see its documentation for more details. This also contains instructions for using the C++ version of the likelihood contained in CC_ini_likelihood.cpp.

After nested sampling runs have been generated, results can also be examined in more detail in the results_testing.ipynb notebook (this also creates the results tables and plots). paper_diagrams.ipynb contains the code for making some of the explanatory figures in the paper.

Attribution

If the code is useful for your research then please cite the Bayesian sparse reconstruction paper. The BibTeX is:

@article{Higson2019bayesian,
author={Higson, Edward and Handley, Will and Hobson, Michael and Lasenby, Anthony},
title={Bayesian sparse reconstruction: a brute-force approach to astronomical imaging and machine learning},
journal={Monthly Notices of the Royal Astronomical Society},
volume={483},
number={4},
pages={4828-4846},
year={2019},
doi={10.1093/mnras/sty3307},
url={https://doi.org/10.1093/mnras/sty3307},
archivePrefix={arXiv},
arxivId={1809.04598}}

Note that some of bsr's dependencies have additional attribution requirements and associated papers.