PCA-based background models

The advantages of using background spectral models is that more information can be extracted from low-count data, as correlations between bins and instrument behaviours are known.

This page describes the machine-learning approach to derive empirical background models published in Simmonds et al., 2018, A&A, 618A, 66. For Chandra/ACIS, XMM/EPIC, Swift/XRT, NuSTAR/FPMA, RXTE, Large archives of background spectra were used to derive principal components (PCA) that empirically describe the background and its variations.

BXA includes these PCA models, which can be fitted to a specific background spectrum.

These PCA models are trained in log10(counts + 1) space to avoid negative counts. The PCA models operate on detector channels and thus should never pass through the response.

The PCA models are limited in how well they can describe additive components such as gaussian emission lines. For this reason, the fitters also try adding Gaussian lines at the location of strongest fit mismatch.

The fits keep increasing complexity (first, the number of PCA components and then, the gaussians) as long as the AIC (Akaike information criterion) improves.

In Sherpa

After setting your source model (with set_model), use:

from bxa.sherpa.background.pca import auto_background
convmodel = get_model(id)
bkg_model = auto_background(id)
set_full_model(id, get_response(id)(model) + bkg_model * get_bkg_scale(id))

Automatically fits background id based on PCA-based templates, and additional gaussian lines as needed by AIC.

A full example for fitting obscured Active Galactic Nuclei is available: https://github.com/JohannesBuchner/BXA/blob/master/examples/sherpa/xagnfitter.py

In Xspec

In xspec, there are two steps.

First, fit the background spectrum outside xspec using the autobackgroundmodel/fitbkg.py script.

You can find it here: https://github.com/JohannesBuchner/BXA/tree/master/autobackgroundmodel

It will give you instructions how to load the PCA model in xspec.

If the fit is bad

This should be judged from the q-q plot that the background fitting produces.

If it is off, it shows at what energies the problem is.

It may be that your spectrum is somehow different than typical spectra.

You can try another method (e.g., empirical background models, or fall back to Wstat statistics).