The Bayesian Global Sky Model (B-GSM): A Calibrated Low Frequency Sky Model for EoR Applications

Gregory Carter, Will Handley, M. Ashdown, N. Razavi‐GhodsarXiv 2025 · DOI · arXiv

TL;DR

The Bayesian Global Sky Model (B-GSM) is a new, absolutely calibrated model of low-frequency (≤408 MHz) Galactic radio emission designed to help astronomers detect the elusive 21cm signal from the early universe. By jointly analyzing spatially detailed sky maps with highly accurate temperature measurements from the EDGES experiment, the authors provide a model that not only predicts sky brightness but also rigorously quantifies its own uncertainty.

What Problem Does This Paper Address?

Cosmologists hunting for the "Epoch of Reionization" (EoR)—the period when the first stars turned on—look for a specific 21cm radio signal from neutral hydrogen. The challenge is that this signal is buried under Galactic foregrounds (mostly synchrotron radiation from our own Milky Way) that are 1,000 to 1,000,000 times brighter.

Existing sky models, like the original Global Sky Model (GSM), rely on Principal Component Analysis (PCA). While useful, these models have two major flaws for EoR research:

  1. Calibration Uncertainty: The underlying survey maps often have significant errors in their temperature scales and "zero-levels" (offsets), which PCA-based models ignore.
  2. Lack of Error Bars: They provide a single "best guess" for the sky brightness without telling you how uncertain that guess is. For a signal as faint as the 21cm line, knowing the error bars is just as important as the measurement itself.

Methods

The authors developed the B-GSM using a joint Bayesian framework. Instead of just "stacking" maps, they treated the sky model and the calibration of the input data as a single inference problem.

  • Component Separation: They modeled the sky as a sum of $k$ emission components. Using Bayesian model comparison, they determined that two components following curved power-law spectra are the optimal balance between simplicity and accuracy.
  • Joint Likelihood: They combined two very different types of data:
  • Spatially resolved maps: High resolution but poorly calibrated.
  • EDGES radiometer data: No spatial resolution (it looks at the whole sky) but extremely well-calibrated absolute temperatures.
  • Nested Sampling: They used the PolyChord algorithm to explore the high-dimensional parameter space, allowing them to calculate the "Bayesian evidence" to compare different model versions and generate full posterior distributions (error estimates) for every pixel and spectral parameter.

Datasets

The model incorporates two primary data sources:

  1. Diffuse Emission Maps: Ten publicly available surveys including the Guzman 45 MHz map, the Haslam 408 MHz map, the EDA2 159 MHz map, and several LWA1 maps (45–80 MHz).
  2. Absolute Temperature Data: Measurements from the EDGES (Experiment to Detect the Global EoR Signature) radiometer, covering 40–200 MHz. This acts as the "ground truth" for calibration.

Key Insights and Ideas

  • Two Components are Enough: The data is best explained by just two spectral components. The first ($\beta \approx -2.63$) represents the standard Galactic synchrotron radiation. The second ($\beta \approx -2.11$) likely accounts for spatial variations in the synchrotron spectrum and contributions from free-free emission in the Galactic plane.
  • Calibration is the Culprit: The study found that many existing low-frequency maps require massive corrections—up to 29% in scale and hundreds of Kelvin in offset. This explains why previous models often struggled to match absolute temperature measurements.
  • The "Gold Standard" Holds Up: The famous Haslam 408 MHz map, a staple of radio astronomy for decades, was found to be remarkably well-calibrated, requiring only a ~3% scale adjustment.
  • Southern Bias: Because the EDGES data (the calibration anchor) was taken in the Southern Hemisphere, the model's uncertainty is significantly higher in the northern sky.

Key Results

  • Spectral Parameters:
    • Component 1: Spectral index $\beta_1 = -2.633 \pm 0.002$, curvature $\gamma_1 = 0.014 \pm 0.001$.
  • Component 2: Spectral index $\beta_2 = -2.108 \pm 0.008$, curvature $\gamma_2 = -0.424 \pm 0.008$.
  • Calibration Corrections: The LWA1 maps required scale corrections of 16–25%. The EDA2 159 MHz map required a ~28% scale increase.
  • Accuracy: The model's predictions for absolute sky temperature agree with EDGES data to within <3.2% across the entire 45–200 MHz range.
  • Data Release: The authors have made the B-GSM model, the code, and the processed datasets available for the community to use in their own foreground removal pipelines.

Limitations and Future Work

  • Uniform Calibration Assumption: The authors assumed each input map has a single global calibration error. However, many maps are "mosaics" of different surveys, meaning the calibration error likely varies across the sky.
  • Northern Sky Uncertainty: The lack of high-quality, absolutely calibrated data for the Northern Hemisphere limits the model's precision there. Incorporating data from northern experiments like LEDA is a planned next step.
  • Point Sources: Currently, bright point sources are masked out. Future versions aim to include a third component specifically for these sources to provide a more complete sky model.
  • Resolution: The current model is smoothed to a 5° resolution. Future work will attempt to push this to the native resolution of the highest-frequency maps (like the 56 arcmin Haslam map).

Summary updated

Abstract

We present the Bayesian Global Sky Model (B-GSM), a new absolutely calibrated model of the diffuse Galactic foreground at frequencies below 408 MHz. We assemble a dataset of publicly available diffuse emission maps at frequencies between 45 MHz and 408 MHz, along with absolute temperature data from the EDGES radiometer between 40 and 200 MHz. We use nested sampling to perform a joint Bayesian analysis of these two datasets and determine posterior distributions of: spatially resolved diffuse components, spectral parameters for the diffuse emission, and calibration corrections for each observed map. Using Bayesian model comparison, we find that the low-frequency sky is optimally modelled by two emission components, each following a curved power-law spectrum. The spectrum for the first component has a spectral index of beta_1 = -2.633 plus/minus 0.002 and a curvature of gamma_1 = 0.014 plus/minus 0.001, while the second has beta_2 = -2.108 plus/minus 0.008 and gamma_2 = -0.424 plus/minus 0.008. The diffuse maps require temperature-scale corrections of 1% to 29%, and zero-level adjustments of a few kelvin to a few hundred kelvin. We find that the Haslam 408 MHz map is well calibrated, requiring a scale correction of 1.029 plus/minus 0.003 (about 3%) and a zero-level correction of 0.91 plus/minus 0.05 kelvin. Posterior predictions for the sky's absolute temperature are in excellent agreement with EDGES data, indicating accurate calibration. The posterior sky predictions agree with observations within statistical uncertainty across all frequencies. However, agreement varies by position, with the largest discrepancies in the Galactic plane. This is the second paper in the B-GSM series; the low-frequency sky model, along with all code and data, is available for download.
BibTeX
@misc{carter2025bayesian,
  author = {Gregory Carter and Will Handley and M. Ashdown and N. Razavi‐Ghods},
  title = {{The Bayesian Global Sky Model (B-GSM): A Calibrated Low Frequency Sky Model for EoR Applications}},
  year = {2025},
  eprint = {2504.04503},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.04503},
  howpublished = {arXiv},
}
Figure