The Bayesian Global Sky Model (B-GSM): Validation of a Data Driven Bayesian Simultaneous Component Separation and Calibration Algorithm for EoR Foreground Modelling

Gregory Carter, Will Handley, M. Ashdown, N. Razavi‐Ghods

TL;DR

The Bayesian Global Sky Model (B-GSM) is a new framework that simultaneously separates radio foreground components and calibrates observational data for low-frequency astronomy. By moving beyond traditional Principal Component Analysis (PCA) and providing robust uncertainty quantification, B-GSM offers a more reliable way to model the "noise" that obscures our view of the early universe.

What Problem Does This Paper Address?

To study the Epoch of Reionization (EoR)—the period when the first stars and galaxies formed—astronomers look for the "21cm signal" from neutral hydrogen. The challenge is that this signal is incredibly faint, buried under "foreground" radio emission from our own Galaxy and other sources that are 1,000 to 1,000,000 times brighter.

Existing tools like the Global Sky Model (GSM) rely on PCA to separate these foregrounds. However, PCA has two major flaws: it assumes the input data is perfectly calibrated (which it rarely is at low frequencies), and it provides no estimate of how uncertain its predictions are. Without knowing the "error bars" on our foreground models, it is nearly impossible to claim a definitive detection of the elusive 21cm signal.

Methods

The authors introduce a fully Bayesian approach. Instead of just finding the "best fit," B-GSM explores the entire probability space of possible sky models using a technique called nested sampling (specifically the PolyChord algorithm).

The core innovation is the Joint Likelihood function. It combines two very different types of data:

  1. Diffuse Sky Maps: These have high spatial resolution (they show the "shape" of the sky) but are often poorly calibrated, containing unknown offsets and scaling errors.
  2. Absolute Temperature Measurements: These come from specialized radiometers (like EDGES). They have no spatial resolution (they are just 1D curves over time) but are extremely well-calibrated "ground truths."

By processing these together, B-GSM uses the "ground truth" data to fix the calibration errors in the high-resolution maps while simultaneously figuring out the physical components (like synchrotron radiation and free-free emission) that make up the sky.

Datasets

Since this is a validation paper (the first of two), the authors used a synthetic dataset to test the algorithm's limits:

  • Components: Two "true" components with curved power-law spectra.
  • Frequencies: Nine channels between 45 and 408 MHz.
  • Real-world imperfections: They added 5% thermal noise, masked out parts of the sky to mimic partial survey coverage, and introduced intentional calibration errors (scaling factors and temperature offsets).

Key Insights and Ideas

  • Simultaneous is Key: Calibration and component separation are usually treated as separate steps. B-GSM shows that doing them at the same time prevents errors in one from "leaking" into the other.
  • Bayesian Evidence as a Filter: The framework uses "Bayesian evidence" to determine the model's complexity. It successfully rejected models with too few or too many components, correctly identifying the 2-component structure used to create the fake data.
  • Robustness to Priors: One of the biggest "aha" moments was finding that even if the researchers' initial guesses (priors) about the sky's spectral index were slightly off, the final sky prediction remained accurate. The data effectively "overruled" the priors.
  • Uncertainty Mapping: Unlike previous models, B-GSM produces "standard deviation maps." This tells researchers exactly where the model is less reliable—typically in the southern sky where observational data is thinner.

Key Results

  • Dramatic Error Reduction: The algorithm significantly reduced residuals. For example, at 60 MHz, the root-mean-square (RMS) error dropped from ~563 K in the uncalibrated data to just ~7.8 K in the B-GSM prediction.
  • Parameter Recovery: The model recovered the "hidden" calibration offsets and scaling factors with high precision, often within $2\sigma$ of the true values.
  • Spectral Accuracy: The recovered spectra for the individual components (synchrotron and free-free) were nearly indistinguishable from the "true" spectra used to generate the data.

Limitations and Future Work

The main limitation is that this study used synthetic data. While the authors tried to make it realistic, real-world radio data contains additional headaches like ionospheric interference and complex telescope beam patterns that weren't fully modeled here.

What’s next? This paper is the "proof of concept." The second paper in the series will apply B-GSM to real, messy observational data. Once complete, the authors plan to release the low-frequency sky model for public download, providing a new standard tool for the 21cm cosmology community.

Summary updated

Abstract

We introduce the Bayesian Global Sky Model (B-GSM), a novel data-driven Bayesian approach to modelling radio foregrounds at frequencies <400~MHz. B-GSM aims to address the limitations of previous models by incorporating robust error quantification and calibration. Using nested sampling, we compute Bayesian evidence and posterior distributions for the spectral behaviour and spatial amplitudes of diffuse emission components. Bayesian model comparison is used to determine the optimal number of emission components and their spectral parametrisation. Posterior sky predictions are conditioned on both diffuse emission and absolute temperature datasets, enabling simultaneous component separation and calibration. B-GSM is validated against a synthetic dataset designed to mimic the partial sky coverage, thermal noise, and calibration uncertainties present in real observations of the diffuse sky at low frequencies. B-GSM correctly identifies a model parametrisation with two emission components featuring curved power-law spectra. The posterior sky predictions agree with the true synthetic sky within statistical uncertainty. We find that the root-mean-square (RMS) residuals between the true and posterior predictions for the sky temperature as a function of LST are significantly reduced, when compared to the uncalibrated dataset. This indicates that B-GSM is able to correctly calibrate its posterior sky prediction to the independent absolute temperature dataset. We find that while the spectral parameters and component amplitudes exhibit some sensitivity to prior assumptions, the posterior sky predictions remain robust across a selection of different priors. This is the first of two papers, and is focused on validation of B-GSMs Bayesian framework, the second paper will present results of deployment on real data and introduce the low-frequency sky model which will be available for public download.
BibTeX
@inproceedings{carter2025bayesian,
  author = {Gregory Carter and Will Handley and M. Ashdown and N. Razavi‐Ghods},
  title = {{The Bayesian Global Sky Model (B-GSM): Validation of a Data Driven Bayesian Simultaneous Component Separation and Calibration Algorithm for EoR Foreground Modelling}},
  booktitle = {Monthly},
  year = {2025},
  doi = {10.1093/mnras/staf1743},
}
Figure