Capturing System Drift with Time Series Calibration for Global 21-cm Cosmology Experiments

This paper introduces a novel time-series calibration method for global 21-cm cosmology experiments that utilizes noise wave parameters to model and correct for system drift and reference source impedance assumptions, thereby significantly reducing spectral residuals and parameter fit errors compared to previous techniques.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Picture: Listening to the Universe's "Baby Talk"

Imagine the universe as a giant, quiet room. For a long time, scientists have been trying to hear a very faint whisper from the very beginning of time—the "Cosmic Dawn." This whisper is a specific radio signal (21-cm) emitted by the first stars and galaxies as they turned on.

The problem? The room is incredibly noisy. There are loud radio stations, lightning storms, and the heat of our own galaxy screaming at us. The signal we want is like a mouse squeaking in a stadium full of cheering fans. To hear it, we need a super-sensitive microphone (a radio telescope) and a way to filter out the noise perfectly.

The Problem: The Microphone is "Drifting"

The paper focuses on an experiment called REACH. Think of REACH as a high-tech radio telescope. To measure the faint whisper, the telescope needs to be calibrated (tuned) perfectly.

However, the telescope isn't a static machine. It's like a musical instrument that changes its tuning slightly every time the temperature shifts or the electronics heat up.

• The Old Way: Previous methods treated the telescope like a perfect, unchanging robot. They took a few measurements of "known" reference sounds (like a standard tuning fork) at the start and end of the day, and assumed the telescope stayed exactly the same in between.
• The Reality: The telescope's electronics (specifically the Low Noise Amplifier, or LNA) are "drifting." They are slowly changing their behavior over time, like a guitar string going slightly out of tune while you are playing a long song.

If you don't account for this drift, your final picture of the universe will be blurry or distorted. In fact, a previous famous claim of hearing the "Cosmic Dawn" signal was likely just a mistake caused by this kind of drift.

The Solution: A "Time-Lapse" Calibration

The authors propose a new way to tune the telescope. Instead of taking a snapshot of the calibration and hoping it stays good, they treat the calibration like a movie.

1. The "Drift" Analogy: Imagine you are trying to measure the height of a growing plant every hour. If you only measure the plant at 9:00 AM and 5:00 PM, and you assume it grew in a straight line, you might miss the fact that it grew faster in the middle of the day.

   • Old Method: Measure the plant twice, draw a straight line, and guess the height at noon.
   • New Method: Measure the plant at 9:00, 10:00, 11:00, etc., and draw a smooth curve that connects all those dots. This curve tells you exactly how the plant grew at every single moment.

2. The "Surface" Fit: The authors created a mathematical "surface" (like a 3D landscape) that maps the telescope's behavior across Time and Frequency (pitch).

   • They measure the "known" reference sounds at specific times.
   • They use a smart math trick (polynomial surfaces) to fill in the gaps, predicting exactly how the telescope was behaving when it was listening to the actual sky, even if that happened between the reference measurements.

The "Hidden Trap": The Reflection Coefficient

There was a second, sneaky problem. The old math assumed that the cables and connectors in the telescope were perfectly smooth and didn't bounce any radio waves back (like a perfectly flat mirror). In reality, they are a bit bumpy.

• The Analogy: Imagine trying to measure the volume of a speaker, but you assume the room is perfectly silent and the walls don't echo. If the room actually has echoes, your volume measurement will be wrong.
• The Fix: The authors rewrote the math to admit that the cables do have imperfections. By removing the assumption that everything is "perfectly matched," they stopped the math from getting confused (a problem called "degeneracy," where two different errors cancel each other out, making it impossible to find the true answer).

The Results: From Blurry to Crystal Clear

The team tested their new method by creating a fake dataset where they knew the telescope was drifting.

• Old Method (1-D Polynomial): The result was a mess. The final signal had a "chromatic residual," which is a fancy way of saying the colors (frequencies) were all wrong, and the signal drifted up and down over time. It was like looking at a photo through a wavy, dirty window.
• New Method (Surface + Better Math):
  • The drift was completely removed.
  • The "wavy window" was cleaned.
  • The error in the measurement dropped by 97%.
  • They recovered the true signal with an accuracy of within 0.06%.

The Takeaway

This paper is about teaching our radio telescopes to be more honest about their own flaws. By acknowledging that the machine changes over time and that its parts aren't perfect, and by using a "time-lapse" mathematical approach to track those changes, we can finally hear the faint, ancient whispers of the first stars without the static of our own equipment getting in the way.

It's the difference between guessing the weather based on one morning's temperature and using a sophisticated model that tracks the wind, humidity, and pressure changes throughout the entire day.

Technical Summary

Here is a detailed technical summary of the paper "Capturing System Drift with Time Series Calibration for Global 21-cm Cosmology Experiments."

1. Problem Statement

Global 21-cm cosmology experiments (such as REACH, EDGES, and SARAS) aim to detect the faint radio signal from neutral hydrogen during the Cosmic Dawn and Epoch of Reionization. The signal amplitude is predicted to be extremely small ($\lesssim 200$ mK) compared to galactic foregrounds ($10^3 - 10^4$ K).

• The Challenge: Detecting this signal requires long integration times (e.g., ~6.5 hours for REACH). Over such durations, the receiver system (specifically the Low Noise Amplifier or LNA) is prone to system drift (changes in noise wave parameters over time).
• Limitations of Current Methods: Previous Bayesian calibration methods (e.g., Roque et al. 2021) treat calibration parameters as static or frequency-dependent only. They do not explicitly utilize time information, assuming the instrument remains stable between calibrator measurements. Furthermore, existing equations often rely on impedance matching assumptions (assuming reference noise sources and cold loads are perfectly matched to 50 $\Omega$), which introduces parameter degeneracies and biases the fit.
• Consequence: Ignoring time drift and impedance mismatches leads to unmodelled systematics, chromatic residuals, and biased fits, potentially mimicking or obscuring the cosmological signal.

2. Methodology

The authors propose a new calibration framework for the REACH experiment that integrates time-series analysis with a reformulated noise wave parameterization.

A. Time-Series Surface Fitting

Instead of fitting 1-dimensional polynomials to noise wave parameters ($T_{unc}, T_{cos}, T_{sin}$) as a function of frequency only, the authors fit 2D polynomial surfaces over both frequency ($\nu$) and time ($t$).

• Equation: The noise wave parameters are modeled as:
  $$T_{NWP}(\nu, t) = \sum_{i} \sum_{j} A_{ij} \cdot \nu^i t^j$$
• Interpolation: This allows the calibration solution to be interpolated to the specific times when the antenna is measured, which occur between the measurements of the 12 reference calibrators.
• Optimization: The polynomial orders are determined via gradient ascent on the Bayesian evidence surface, utilizing an "Occam penalty" to prevent overfitting.

B. Reformulation of the Calibration Equation

The authors address the degeneracies caused by assuming perfect impedance matching in reference sources.

• Previous Approach: Used effective temperatures ($T'_{NS}, T'_L$) which implicitly assumed $\Gamma_{NS} = \Gamma_L = 0$. This created dependencies where $T'_{NS}$ and $T'_L$ were functions of the other noise wave parameters, leading to biased fits.
• New Approach: The calibration equation is rewritten to explicitly include the reflection coefficients ($\Gamma_{NS}, \Gamma_L$) of the noise source and cold load.
  $$T_s(\nu) = \bar{X}_{unc}T_{unc} + \bar{X}_{cos}T_{cos} + \bar{X}_{sin}T_{sin} + \bar{X}_{NS}T_{NS} + \bar{X}_{L}T_{L}$$
• Benefit: This removes the degeneracy between the noise wave parameters and the effective temperatures, allowing $T_{NS}$ and $T_L$ to be treated as distinct physical quantities (with $T_L$ often fixed to the measured ambient temperature).

C. Simulation and Validation

• Data Generation: A synthetic dataset was created using REACH instrument reflection coefficients. The LNA noise wave parameters were modeled as drifting low-order polynomial surfaces over time.
• Validation Source: A flat 5000 K spectrum was simulated as a validation source to test calibration accuracy.
• Comparison: Three methods were compared:
  1. 1D Polynomial Fit (Roque et al. 2021 equation).
  2. 2D Surface Fit (Roque et al. 2021 equation).
  3. 2D Surface Fit (New Equation 16 from this work).

3. Key Contributions

1. Time-Dependent Calibration: Introduction of a polynomial surface fitting method that explicitly models the evolution of noise wave parameters over time, enabling accurate interpolation for antenna measurements taken between calibrator scans.
2. Degeneracy Removal: A reformulation of the linear calibration equation that eliminates assumptions about the impedance matching of reference sources, thereby breaking parameter degeneracies that previously biased the solution.
3. Improved Bayesian Framework: Integration of these changes into the conjugate priors framework, allowing for robust posterior sampling of the time-frequency surface coefficients.

4. Results

The performance of the three methods was evaluated using the Root Mean Square Error (RMSE) of the calibrated validation source spectrum and the accuracy of parameter recovery.

Method	Calibration Equation	Drift Correction?	Chromatic Residual?	Avg. RMSE (K)	Parameter Fit Error
1D Polynomial	Eq. 10 (Old)	No	Yes	5.27	High
2D Surface	Eq. 10 (Old)	Yes	Yes	0.73	Moderate
2D Surface	Eq. 16 (New)	Yes	No	0.13	Lowest

• Drift Removal: The 1D method failed to correct for drift, resulting in a large time-varying error. The surface fitting methods successfully removed the drift.
• Chromatic Residuals: While the surface fitting with the old equation removed drift, it left a significant chromatic (frequency-dependent) residual due to parameter degeneracies. The new equation (Eq. 16) eliminated this residual entirely.
• Accuracy:
  • The new method achieved a 97% reduction in RMSE compared to the original 1D method (from 5.27 K to 0.13 K).
  • Parameter recovery for $T_{unc}$ and $T_{cos}$ improved by up to six times compared to the previous surface fitting method.
  • The recovered parameters were within 0.06% of the ground truth.
  • The fit to the noise source temperature ($T_{NS}$) became a flat plane at the true value, confirming the removal of bias.

5. Significance

This work addresses a critical bottleneck in global 21-cm cosmology: the separation of the faint cosmological signal from instrumental systematics.

• Robustness: By accounting for system drift and removing impedance assumptions, the method significantly reduces the risk of false positives or biased signal detection caused by unmodelled systematics.
• Scalability: The method is particularly valuable for long-duration observations and for combining data from multiple days, where LNA characteristics may vary slowly over time.
• Future Impact: The reduction in RMSE and chromatic residuals brings the calibration precision closer to the requirements needed to definitively detect the Cosmic Dawn signal, potentially resolving controversies surrounding previous detections (e.g., EDGES).

Limitations: The method still assumes the receiver gain ($g$) and receiver noise offset ($T_0$) are constant during the Dicke switching cycle. Future work may explore non-polynomial basis functions (e.g., Fourier or Gaussian processes) to capture more complex LNA behaviors.