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

This paper, published in the Monthly Notices of the Royal Astronomical Society, introduces the Bayesian Global Sky Model (B-GSM). It represents a significant step forward in how astrophysicists model the "foreground" noise that obscures our view of the early universe.

TL;DR

B-GSM is a new Bayesian framework that simultaneously calibrates messy radio survey data and separates it into distinct physical emission components. By moving beyond traditional Principal Component Analysis (PCA), it provides the robust error quantification necessary to finally detect the elusive 21cm signal from the Epoch of Reionization.

What Problem Does This Paper Address?

To study the Epoch of Reionization (EoR)—the era when the first stars turned on—astronomers look for a specific 21cm radio signal. The catch? This signal is buried under "foreground" emission (mostly from our own Galaxy) that is three to six orders of magnitude brighter.

Existing tools like the Global Sky Model (GSM) rely on PCA to describe this foreground. However, PCA-based models have two major flaws:

  1. Calibration Blindness: They assume the input sky surveys are perfectly calibrated. In reality, low-frequency maps often have massive errors in their temperature scales and zero-levels.
  2. Lack of Uncertainty: They provide a "best guess" for the sky but no "error bars," making it impossible to know if a detected signal is real or just a modeling artifact.

Methods

The authors treat sky modeling as a formal Bayesian inference problem. Instead of just fitting a curve, they use Nested Sampling (specifically the PolyChord algorithm) to explore the parameter space.

Key methodological innovations include:

  • Joint Likelihood: The model is conditioned on two independent datasets: high-resolution (but poorly calibrated) diffuse maps and low-resolution (but accurately calibrated) absolute temperature measurements (like those from the EDGES experiment).
  • Simultaneous Calibration: B-GSM treats calibration offsets and scale factors as "nuisance parameters" to be solved alongside the physical sky model.
  • Analytical Marginalization: To handle the computational nightmare of modeling millions of pixels, the authors derived a way to analytically "integrate out" the spatial amplitudes, focusing the sampler on the spectral parameters.
  • Model Selection: They use Bayesian Evidence ($Z$) to objectively decide how many emission components (e.g., synchrotron vs. free-free emission) the data actually supports, rather than guessing.

Datasets

As this is a validation paper, the authors used a synthetic dataset designed to mimic the headaches of real-world observations:

  • Diffuse Maps: Simulated versions of the Guzman 45 MHz, LWA1, Landecker 150 MHz, and Haslam 408 MHz surveys.
  • Real-world Artifacts: The maps included partial sky coverage, thermal noise, and intentional calibration errors (offsets up to 600K and scale errors up to 18%).
  • Ground Truth: A synthetic "absolute temperature" dataset representing what a perfectly calibrated radiometer would see as the Earth rotates (Local Sidereal Time).

Key Insights and Ideas

  • The "Ground Truth" Anchor: The absolute temperature data acts as a "calibration anchor." Even though it has no spatial resolution, it provides enough information to fix the temperature scales of the high-resolution maps.
  • Evidence-Based Complexity: The Bayesian evidence successfully "rejected" models that were too simple (1 component) or too complex (3 components), correctly identifying the 2-component truth.
  • Prior Sensitivity vs. Prediction Robustness: A fascinating finding was that while the internal spectral parameters (like the exact spectral index of a component) were sensitive to the researcher's "priors," the final sky prediction remained highly accurate and robust. This suggests B-GSM is excellent for foreground subtraction even if the underlying physics of individual components remains slightly ambiguous.

Key Results

  • Calibration Recovery: B-GSM accurately recovered the artificial calibration errors. For example, at 50 MHz, the Root-Mean-Square (RMS) residual between the model and the truth dropped from 426.61 K (uncalibrated) to just 11.79 K (calibrated).
  • Statistical Consistency: The posterior sky predictions agreed with the "true" synthetic sky within calculated statistical uncertainties across the entire 45–408 MHz range.
  • Residuals: The normalized residuals (the difference between the model and the truth) were shown to be "white noise," indicating that the model captured all the meaningful physical information in the data.

Limitations and Future Work

  • Physical Interpretability: The authors acknowledge that the spectral parameters are somewhat degenerate. While the model predicts the total sky well, assigning specific physical meanings (e.g., "this is exactly synchrotron radiation") to the individual components is still challenging due to prior sensitivity.
  • Computational Complexity: Even with marginalization, the process is intensive. The authors used an approximation for the marginal likelihood that introduces a few percent error, though they demonstrated this doesn't significantly hurt the results.
  • Next Steps: This paper is a "proof of concept." The next phase involves applying B-GSM to real observational data to produce a new, publicly available low-frequency sky model that the 21cm community can use for actual discoveries.

Summary updated

Abstract

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 spectra and spatial amplitudes of emission components. Bayesian model comparison is used to determine the optimal number of emission components and their spectral parametrization. Posterior sky predictions are conditioned on both diffuse emission and absolute temperature data sets, enabling simultaneous component separation and calibration. B-GSM is validated against a synthetic data set that mimics the partial sky coverage, thermal noise, and calibration uncertainties present in real observations. B-GSM successfully identifies a model parametrization with two emission components featuring curved power-law spectra. The Bayesian evidence strongly rejects incorrect models, confirming B-GSM’s capability to use the data set to determine model parametrization. The posterior sky predictions agree with the true sky within statistical uncertainty. The root-mean-square residuals between the true and posterior predictions for the sky temperature as a function of local sidereal time are significantly reduced, compared to the uncalibrated data set. This indicates that B-GSMs posterior sky predictions are accurately calibrated to the independent absolute temperature data set. We note that the spectral parameters and component amplitudes exhibit sensitivity to prior assumptions, however the posterior sky predictions remain robust across a selection of different priors. This paper is focused on validation of B-GSM; a later paper will present results for real data and introduce the low-frequency sky model.
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