Welcome

bsixsa does Bayesian X-ray spectral fitting on top of PyXspec. You load your spectrum and define your model in PyXspec as usual, give each free parameter a prior, and choose how to fit: nested sampling, MCMC, an optimiser, or SIXSA (Simulation-based Inference for X-ray Spectral Analysis), the method of the "SBI with NPE applied to X-ray spectral fitting" paper series¹².

It started as a fork of the Bayesian X-ray Analysis toolkit (BXA) and keeps a similar syntax. It has since grown into a general-purpose wrapper around xspec for a wide range of fitting and sampling methods.

Installation

bsixsa 0.2 is not on PyPI yet (the bsixsa package there is an older, incompatible release), so install it from GitHub. First create an environment with XSPEC, from HEASARC's conda channel:

conda create -n bsixsa python=3.12 xspec xspec-data -c https://heasarc.gsfc.nasa.gov/FTP/software/conda/ -c conda-forge
conda activate bsixsa

Then install bsixsa with the samplers you want, given as extras in brackets:

pip install "bsixsa[nautilus] @ git+https://github.com/renecotyfanboy/bsixsa"

Each sampler is an extra named after its backend: nautilus, nessai, ultranest, emcee and iminuit (see Choosing a backend). Give several separated by commas, e.g. bsixsa[nautilus,emcee], or all of them with bsixsa[all]. SIXSA and LevenbergMarquardt need no extra. If the sampler of the backend you chose is missing, Solver(...) stops with an ImportError that gives the pip install command.

Current limitations

• Several spectra and backgrounds are not supported explicitly. The nested samplers, Emcee and Iminuit use XSPEC's total fit statistic, so they fit every loaded spectrum and account for a loaded background as XSPEC does (C-stat then becomes W-stat), but this is not tested systematically. LevenbergMarquardt fits the counts of the first spectrum only, without its loaded background: do not use it with several spectra or a background file (a background modelled as a second source, as in the NXB example, is fine). Priors apply to the models of data group 1: keep the parameters of other data groups linked to them (XSPEC's default).
• SIXSA fits one spectrum, with cstat, no loaded background, and independent priors with finite bounds.
• Plots show the first spectrum only, without its background. The observed data and the per-bin likelihood of build_inference_data also come from the first spectrum only.
• Cross-calibration is not supported through the standard xspec interface (a free constant per data group), but it can be built into the spectral model.
• Fluxes are not computed by bsixsa itself: export the posterior as an XSPEC chain and use XSPEC's flux command with errors, as in the simple fit.