Mitigating gain calibration errors from EoR observations with SKA1-Low AA*

This paper presents a post-calibration mitigation strategy combining Gaussian Process Regression, Principal Component Analysis, and foreground avoidance to address residual foregrounds from gain calibration errors in SKA1-Low AA* observations, demonstrating that the 21-cm signal can be recovered within 2$\sigma$ accuracy for calibration errors up to 1% across the $0.05 \leq k \leq 0.5$Mpc$^{-1}$ scale range.

The Gist

Imagine trying to listen to a whisper from a distant friend (the Cosmic Dawn) while standing in the middle of a roaring stadium crowd (the Foregrounds).

This paper is about a team of astronomers trying to build a better "noise-canceling headphone" for the Square Kilometre Array (SKA), a massive radio telescope in the making. Their goal is to hear the faint radio whispers of the very first stars in the universe, a time known as the Epoch of Reionization (EoR).

Here is the story of their challenge and their solution, broken down into simple concepts.

1. The Problem: The Whisper vs. The Roar

The signal from the first stars is incredibly faint. In fact, the "noise" from our own galaxy and other bright radio sources is about 10,000 to 100,000 times louder than the signal they are looking for.

• The Analogy: Imagine trying to hear a single firefly blinking in a stadium full of flashing strobe lights.
• The Twist: The astronomers know the "strobe lights" (foregrounds) are smooth and predictable, while the "firefly" (the cosmic signal) is jagged and unique. They usually try to mathematically subtract the smooth lights to reveal the firefly.

2. The Glitch: The "Dirty Ear"

The biggest problem isn't just the loud crowd; it's that the telescope itself is slightly imperfect. To hear clearly, the telescope needs to be perfectly calibrated (like tuning a radio). But in reality, the calibration is never 100% perfect.

• The Analogy: Imagine your noise-canceling headphones have a tiny crack in the seal. Even a tiny crack (a 0.01% error) lets a little bit of the stadium noise leak in.
• The Consequence: The paper shows that if this "crack" is too big (even just 1% or 10%), the leaked noise looks exactly like the firefly signal. The astronomers might think they found the first stars, but they are actually just hearing the echo of their own mistakes. This is called Gain Calibration Error.

3. The Solution: A Two-Step Strategy

The authors realized that using just one method to clean up the signal wasn't enough when the telescope was "out of tune." So, they developed a Hybrid Strategy (a mix-and-match approach).

Think of it like cleaning a muddy window:

1. The Squeegee (Subtraction): First, they use smart math (called Gaussian Process Regression and Principal Component Analysis) to wipe away the obvious mud (the predictable foregrounds).
2. The Masking Tape (Avoidance): However, if the window is too dirty (high calibration error), the squeegee leaves streaks. So, they put a piece of tape over the streakiest parts of the window and ignore them completely. They only look through the clean parts.

4. The Results: How Much Can We Hear?

The team simulated the telescope with different levels of "cracks" (errors) to see how well their hybrid strategy worked:

• Tiny Crack (0.1% error): The squeegee alone works fine. They can hear the whisper clearly across the whole window.
• Medium Crack (1% error): The squeegee leaves some streaks. They have to use the masking tape to cover the worst parts. They can still hear the whisper, but they have to ignore the edges of the window. They lose about 30% of the view, but the signal is real.
• Huge Crack (10% error): The window is a mess. The squeegee makes things worse, and the masking tape has to cover almost everything. They can only hear the whisper in the very center of the window, and it's very faint.

5. The Takeaway

The paper concludes that perfection is impossible, but smart compromise is possible.

• The Lesson: If the telescope isn't calibrated perfectly, you can't just rely on math to fix it. You have to be willing to throw away some data (the "masked" parts) to ensure the data you keep is trustworthy.
• The Future: This research helps the SKA team know exactly how precise their calibration needs to be. It tells them: "If you want to hear the first stars clearly, you need to keep your calibration errors below 1%, or be prepared to lose a lot of your view."

In a nutshell: The universe is whispering, but our ears (the telescope) are slightly out of tune. This paper teaches us how to use a combination of "math magic" and "strategic ignoring" to finally hear the story of how the universe woke up.

Technical Summary

Here is a detailed technical summary of the paper "Mitigating gain calibration errors from EoR observations with SKA1-Low AA"* by Beohar et al.

1. Problem Statement

The detection of the redshifted 21-cm signal from neutral hydrogen (Hi) during the Cosmic Dawn and Epoch of Reionization (EoR) is hindered by bright astrophysical foregrounds (4–5 orders of magnitude brighter than the signal) and instrumental systematics.

• The Specific Challenge: This work focuses on gain calibration errors (imperfections in amplitude and phase of antenna gains). Previous studies indicate that even minute errors (as low as 0.01%) can bias power spectrum estimation by an order of magnitude.
• The Mechanism: Calibration errors introduce "mode-mixing," where the spectrally smooth foreground power leaks into the "EoR window" (the region in Fourier space where the cosmological signal is expected to be uncontaminated). This creates a "foreground wedge" that obscures the Hi signal, particularly at large scales (low $k$).
• The Gap: While foreground removal (subtraction) and avoidance techniques exist, their efficacy under varying levels of realistic gain calibration errors (specifically for the future SKA1-Low array) requires systematic evaluation.

2. Methodology

The authors developed an end-to-end simulation pipeline (21cmE2E) to generate synthetic SKA1-Low AA* observations and test mitigation strategies.

A. Simulation Setup

• Telescope Model: Simulated the SKA1-Low AA* configuration (307 stations, restricted to a 2 km radius for computational efficiency) observing a 4-hour duration in the 138–146 MHz band ($z \sim 9$).
• Sky Model:
  • Hi Signal: Generated using the semi-numerical code 21cmFAST (redshift range 7.4–8.2).
  • Foregrounds: Modeled using compact point sources from the T-RECS catalogue (2,522 sources). Note: Galactic diffuse emission was excluded to isolate the impact of gain errors.
  • Gain Errors: Introduced complex gain errors ($g_i$) with independent amplitude and phase errors drawn from a normal distribution. Three error levels were tested: 0.1%, 1%, and 10%.
• Imaging: Visibilities were processed using OSKAR and WSCLEAN to create image cubes, followed by power spectrum estimation (1D and 2D).

B. Mitigation Strategies

The study evaluated three approaches to recover the Hi signal:

1. Foreground Avoidance: Discarding $k$-modes inside the "foreground wedge." Three scenarios were tested:
   • No avoidance (all modes).
   • Horizon limit ($k_\parallel \leq k_\perp \dots$).
   • Horizon + 0.1 $Mpc^{-1}$ buffer (conservative).
2. Foreground Subtraction (Blind):
   • Principal Component Analysis (PCA): Removes dominant spectral modes (low-rank foregrounds).
   • Gaussian Process Regression (GPR): Models foregrounds using a covariance kernel (combining smooth sky foregrounds and mode-mixing components) to subtract them statistically.
3. Hybrid Approach: Sequentially applying subtraction (PCA/GPR) followed by avoidance to clean the residual wedge.

3. Key Contributions

• Systematic Error Analysis: Provided a quantitative assessment of how gain calibration errors degrade the EoR window and limit the effectiveness of standard mitigation techniques.
• Hybrid Strategy Validation: Demonstrated that a hybrid approach (subtraction + avoidance) is superior to using either method alone, especially under moderate-to-severe calibration errors.
• GPR Kernel Optimization: Investigated the impact of including "mode-mixing" kernels in GPR models, showing that accounting for non-smooth components improves the recovery of Hi coherence scales ($l_{21}$).
• Threshold Definition: Established specific gain error thresholds for the SKA1-Low to achieve $2\sigma$ signal recovery.

4. Results

The performance was evaluated based on the recovery of the 1D and 2D power spectra within the range $0.05 \leq k \leq 0.5 , Mpc^{-1}$.

Gain Error Level	Findings
0.1%	Low Contamination: Foreground avoidance around the horizon is sufficient to recover the Hi signal within $2\sigma$ across all scales. Subtraction methods (PCA/GPR) offer negligible additional benefit.
1%	Moderate Contamination: • Subtraction alone: Fails to recover large scales; GPR outperforms PCA on large scales ($k \leq 0.1$) but both struggle with mode-mixing on small scales. • Avoidance alone: Requires a 0.1 $Mpc^{-1}$ buffer to reach $2\sigma$, resulting in a **~30$Mpc^{-1}$buffer *after* GPR/PCA subtraction recovers the signal within$2\sigma$ across accessible modes with better SNR than avoidance alone.
10%	Severe Contamination: • Subtraction alone: Fails completely. GPR grossly overestimates small-scale power and suppresses large scales; PCA fails to model strong mode-mixing. • Avoidance alone: Fails to reduce residual power to signal levels for $k \leq 0.2$. • Hybrid (Best): Combining GPR with a 0.1 $Mpc^{-1}$ buffer allows recovery within $2\sigma$for$0.1 \leq k \leq 0.5$, though at a ~30% SNR cost.

• Mode-Mixing Effects: The study found that as gain errors increase, the spectral smoothness of foregrounds degrades. Standard GPR kernels (e.g., RBF) struggle in high-error regimes unless augmented with mode-mixing kernels (e.g., Matérn 3/2 or 5/2).

5. Significance and Conclusion

• Calibration Requirements: The work reinforces that gain calibration errors must be kept below 0.01% for optimal performance, but even with errors up to 1%, the Hi signal can be statistically recovered using hybrid techniques.
• Trade-off: There is an inherent trade-off between calibration precision and data sensitivity. As calibration errors increase, the "EoR window" shrinks, forcing observers to discard more modes (avoidance), which reduces the SNR by approximately 30% for 1%–10% errors.
• Future Outlook: The authors conclude that future SKA1-Low observations will require a hybrid mitigation strategy combining model-based subtraction with strategic mode avoidance. They plan to extend this framework to include thermal noise, bandpass errors, and real data from the uGMRT.

In summary: This paper provides a critical roadmap for the SKA1-Low collaboration, demonstrating that while gain calibration errors are a severe systematic threat, a carefully tuned hybrid pipeline of Gaussian Process Regression and conservative mode avoidance can salvage the cosmological signal, provided the community accepts a moderate reduction in sensitivity.