🔭 astrophysics

Impact of 21-cm foreground mitigation strategies on reionization power spectrum constraints

This paper evaluates two 21-cm foreground mitigation strategies—Foreground Avoidance and Gaussian Process Regression-based Removal—finding that while both introduce systematic biases of up to $\approx$ 1 $\sigma$ in astrophysical parameters, they successfully recover the global reionization history within 95% credible intervals, with the removal method offering broader parameter constraints by reclaiming contaminated Fourier modes.

Original authors: Sambit K. Giri, Florent Mertens

Published 2026-06-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sambit K. Giri, Florent Mertens

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the early universe as a giant, dark room where the first stars are just starting to flicker on. Astronomers want to take a picture of this room using a special "radio camera" that listens to a specific hum made by hydrogen gas (the 21-cm signal). This hum tells us exactly how the stars were forming and how the darkness was being cleared away.

However, there's a massive problem: The room is incredibly noisy. There are loud, static-filled radio stations (astrophysical foregrounds) from our own galaxy and beyond that are thousands of times louder than the faint hum of the early stars. It's like trying to hear a whisper in a stadium during a rock concert.

This paper is a test to see which method works best to silence the noise so we can hear the whisper. The researchers used a "blind test" (a simulated dataset where they didn't know the answer beforehand) to compare two main strategies.

The Two Strategies

1. The "Safe Zone" Strategy (Foreground Avoidance)
Imagine the noise in the stadium is concentrated in the lower seats and the aisles, while the upper balcony is relatively quiet.

How it works: This strategy says, "Let's just ignore the noisy lower seats and aisles entirely. We will only listen to the data from the quiet upper balcony."
The Result: The data is very clean because they threw away the messy parts. However, they also threw away a lot of information about the biggest structures in the room because the "quiet balcony" doesn't cover the whole space. It's like taking a photo of the stadium but only zooming in on the top row; you get a clear picture of that row, but you miss the big picture of the whole crowd.

2. The "Noise Cancelling" Strategy (Foreground Removal)
Imagine you have a super-smart AI that can listen to the rock concert, figure out exactly what the band is playing, and then play it backward to cancel out the noise, leaving only the whisper.

How it works: This strategy uses a mathematical model (Gaussian Process Regression) to predict what the noise looks like and subtracts it from the data. This allows them to keep the data from the "noisy lower seats" (the lower balcony), recovering more of the big picture.
The Result: They get more data and a wider view of the room. However, if the AI guesses the noise wrong, it might accidentally subtract part of the whisper or leave a ghost of the noise behind. It's a riskier game: higher reward, but a higher chance of distortion.

What They Found

The researchers ran these strategies through a computer simulation of the universe and checked how well they could figure out the "rules" of how the first stars formed (like how heavy the star-forming clouds were and how much light escaped them).

Both strategies had "glitches": Neither method was perfect. Both introduced small errors (biases) in their calculations.
- The "Safe Zone" team (Avoidance) was too cautious. Because they threw away so much data, their answers were very "fuzzy" (broad uncertainty). They couldn't pin down the exact size of the star-forming clouds.
- The "Noise Cancelling" team (Removal) was more aggressive. They got a sharper picture, but their answers were sometimes pushed in the wrong direction. For example, they guessed the star-forming clouds were much smaller than they actually were because the "noise cancelling" accidentally ate away some of the signal.
The "Whisper" vs. The "Volume":
- When trying to figure out how many stars were forming (the source parameters), both methods struggled a bit.
- However, when trying to figure out how much of the universe was still dark (the neutral fraction), both methods did a pretty good job. It's as if they couldn't agree on the exact shape of the whisper, but they could both agree on how loud it was overall.
- The Late Night Problem: At the very end of the simulation (when the universe was almost fully lit), both methods stumbled. The complex way the light bubbles merged made it hard to distinguish the signal from the remaining noise, leading to a persistent error in their calculations.

The Big Takeaway

The paper concludes that there is no perfect "magic button" to remove the noise.

If you want to be safe and avoid making up data, you have to throw away a lot of information (Avoidance).
If you want to keep all the information, you have to risk introducing errors because your math model might not be perfect (Removal).

The researchers suggest that in the future, we shouldn't just rely on one method. We need to be careful about which parts of the data we trust. If we can identify the "bad" parts of the data (the most contaminated bins) and throw those specific ones away, we can get much better results. It's like realizing that the lower seats in the stadium are too noisy to trust, so we only listen to the middle section, getting a balance between safety and information.

Ultimately, this study prepares astronomers for the Square Kilometre Array (SKA), a massive new radio telescope. It tells them that no matter how good their noise-cancelling software gets, they will always have to make trade-offs between keeping data and keeping it clean.

Technical Summary: Impact of 21-cm Foreground Mitigation Strategies on Reionization Power Spectrum Constraints

Problem Statement
The redshifted 21-cm signal from neutral hydrogen (HI) during the Epoch of Reionization (EoR) offers a unique probe of high-redshift structure formation and early galaxy astrophysics. However, this cosmological signal is overwhelmed by astrophysical foregrounds that are orders of magnitude stronger. While foregrounds are spectrally smooth, instrumental effects (e.g., beam chromaticity, calibration residuals) corrupt their spectral structure, leaking power into the line-of-sight Fourier modes essential for detection. This creates a "foreground wedge" in the $(k_\perp, k_\parallel)$ Fourier plane, leaving a cleaner "EoR window." The central challenge addressed in this work is quantifying how different strategies for mitigating these foregrounds—specifically Foreground Avoidance versus Foreground Removal—propagate systematic biases into the inference of astrophysical parameters and the global reionization history.

Methodology
The authors utilize data from the Square Kilometre Array Observatory (SKAO) Science Data Challenge 3a (SDC3a), a blind analysis of a simulated 1000-hour SKA-Low dataset where the ground truth signal was unknown to the analysis teams. The study evaluates two specific mitigation pipelines submitted by the DOTSS team:

Foreground Avoidance: Restricts analysis to the uncontaminated EoR window in Fourier space, sacrificing large-scale modes to ensure data fidelity without statistical modeling.
Foreground Removal: Utilizes Gaussian Process Regression (GPR) with machine-learning-trained kernels to model and subtract spectrally smooth foregrounds, attempting to reclaim modes within the wedge.

The authors perform Bayesian inference on the recovered power spectra using a Gaussian likelihood and the nautilus dynamic nested sampler. They employ the 21cmEMU emulator (trained on 21cmFAST) to map power spectra to four free astrophysical parameters: the minimum star-forming halo mass ( $M_{turn}$ ), star formation efficiency normalization ( $f_{\star,10}$ ), its mass dependence ( $\alpha_\star$ ), and the ionizing escape fraction normalization ( $f_{esc,10}$ ). Derived quantities include the neutral fraction ( $x_{HI}$ ) at specific redshifts. The analysis compares results against the "Foreground-free" ground truth across various scenarios, including full redshift ranges, selective redshift bins, and varying modeling uncertainties.

Key Results

Systematic Biases in Astrophysical Parameters: Both mitigation strategies introduce systematic biases of up to $\approx 1\sigma$ $\approx 1 σ$ in astrophysical parameters.
- Avoidance consistently yields broader posteriors due to the loss of large-scale modes ( $k \lesssim 0.2 \, h\text{Mpc}^{-1}$ ). It biases the minimum halo mass ( $M_{turn}$ ) toward higher values.
- Removal produces tighter constraints but introduces stronger biases, particularly pushing $M_{turn}$ toward the lower prior boundary ($7.0$) and biasing $f_{esc,10}$ toward lower values. These biases are driven by suppressed large-scale power in the recovered spectra, particularly at redshifts where the signal exhibits complex structure (e.g., $z \approx 8.90$ ).
Redshift-Dependent Performance: Both methods struggle most at $z \approx 8.90$ , where the simultaneous onset of X-ray heating and early reionization creates signal complexity difficult to distinguish from residuals. At $z \approx 6.54$ (late reionization), both methods exhibit a persistent bias in $x_{HI}$ toward lower values, likely due to non-linearities from ionized bubble overlaps decoupling power spectrum amplitude from the neutral fraction.
Resilience of Global History: Despite biases in source parameters, the global reionization history ( $x_{HI}$ ) is generally recovered within the 95% credible interval across all scenarios. This resilience is attributed to $x_{HI}$ being encoded primarily in the overall power spectrum amplitude rather than detailed shape.
Impact of Redshift Selection: Excluding poorly recovered bins (specifically $z=8.90$ and $z=6.54$ ) significantly reduces parameter biases in the "Best redshifts" scenario, reducing biases to $\approx 0.5\sigma$ for most parameters. However, identifying such bins in real observations requires independent quality metrics.
Convergence on Identical Scales: When the Removal strategy is restricted to the same $k$ -modes as Avoidance, the resulting posteriors converge to similar regions, confirming that the divergence between strategies is primarily driven by the additional modes reclaimed by Removal.

Significance and Claims
The paper claims to provide an explicit assessment of the information trade-offs inherent in foreground mitigation for the SKA era. Its primary contribution is demonstrating that while the global reionization history is robustly recovered, the inference of specific astrophysical source parameters is highly sensitive to the choice of mitigation strategy and the quality of the recovered power spectrum at specific redshifts.

The authors emphasize that:

Mitigation-induced biases are unavoidable in current power spectrum analyses, affecting parameters like $M_{turn}$ and $f_{esc}$ by up to $1\sigma$ .
Multi-redshift inference is susceptible to contamination from poorly mitigated bins, necessitating independent quality metrics (e.g., cross-correlations with Lyman- $\alpha$ forests or CMB lensing) to identify and exclude them in real observations.
Strategy selection is observable-dependent: While Avoidance and Removal yield similar results on identical scales, Removal is necessary for image-based statistics requiring sky-plane access, whereas Avoidance offers robustness against model misspecification at the cost of sensitivity.

The work concludes that while the 21-cm power spectrum remains the primary benchmark for initial EoR science, future strategies may need to incorporate foreground models directly into the inference forward model or utilize higher-order statistics to recover information lost to standard mitigation techniques.

The Two Strategies

What They Found

The Big Takeaway

More like this