[Paper Review] EDGES and JWST with 21cm global signal emulator

Shintaro Yoshiura, Teppei Minoda+1 more|arXiv (Cornell University)|May 19, 2023

Advanced Measurement and Metrology Techniques4 citations

TL;DR

This paper develops an artificial neural network-based emulator to efficiently predict the 21cm global signal with sub-mK accuracy, enabling fast Bayesian parameter estimation. Applied to EDGES low-band data, the emulator reveals that results are highly sensitive to foreground modeling, noise assumptions, and frequency range—Bayesian evidence favors higher-order polynomial foregrounds and enhanced noise, with EDGES-derived models consistent with recent JWST galaxy luminosity function measurements at z≈16.

ABSTRACT

The 21cm global signal is an important probe to reveal the properties of the first astrophysical objects and the processes of the structure formation from which one can constrain astrophysical and cosmological parameters. To extract the information of such parameters, one needs to efficiently evaluate the 21cm global signal for statistical analysis. First we developed an artificial neural network-based emulator to predict the 21cm global signal, which works with significantly less computational cost and high precision. Then we apply our emulator to demonstrate the parameter estimation based on the Bayesian analysis by using the publicly available EDGES low-band data. We find that the result is sensitive to the foreground model, the assumption of noise, and the frequency range used in the analysis. The Bayesian evidence suggests the models with higher order polynomial function and enhanced noise are preferred. We also compare models suggested from the EDGES low-band data and the ones from recent JWST measurements of the galaxy luminosity function at $z=16$. We find that the model which produces the 21cm absorption line at $z\approx15$ is well consistent with the central value of the observed luminosity function at $z=16$.

Motivation & Objective

• To develop a fast, accurate emulator for the 21cm global signal to enable efficient statistical inference in cosmology.
• To perform Bayesian parameter estimation using publicly available EDGES low-band data, accounting for uncertainties in foregrounds, noise, and frequency range.
• To test the consistency between EDGES-derived astrophysical models and recent JWST observations of the galaxy luminosity function at z≈16.
• To assess the robustness of 21cm signal detection in light of systematic effects such as foregrounds and noise modeling.
• To explore the implications of the EDGES absorption signal for primordial power spectrum parameters and early star formation.

Proposed method

• An artificial neural network (ANN) is trained on 21cm global signal simulations spanning 8 astrophysical and 3 primordial power spectrum parameters.
• The emulator predicts the 21cm global signal with a root mean square error (RMSE) of a few mK, drastically reducing computational cost compared to full simulations.
• Bayesian inference is performed using the Polychord sampler, with likelihoods based on EDGES low-band data and modeled foregrounds.
• Foregrounds are modeled using polynomial functions of varying order, and noise is assumed to be either white or enhanced (redshift-dependent).
• The analysis includes marginalization over model parameters and uses Bayesian evidence to compare competing models.
• The emulator is validated against full 21cmFAST simulations and used to derive posterior distributions for UV luminosity functions at z=12 and z=16.

Experimental results

Research questions

• RQ1 How does the choice of foreground model affect the inferred 21cm global signal and parameter constraints from EDGES data?
• RQ2 What is the impact of noise modeling assumptions (white vs. enhanced) on Bayesian evidence and signal detection significance?
• RQ3 To what extent are the astrophysical parameters derived from EDGES consistent with recent JWST measurements of the galaxy UV luminosity function at z≈16?
• RQ4 Can the 21cm global signal emulator reliably reproduce the signal shape and amplitude for high-redshift star formation models?
• RQ5 How sensitive are the posterior constraints on primordial power spectrum parameters to the assumed signal model and data range?

Key findings

• The ANN-based emulator achieves sub-mK accuracy (RMSE of a few mK) in predicting the 21cm global signal across a wide range of astrophysical and primordial parameters.
• Bayesian evidence favors higher-order polynomial foreground models and enhanced noise, indicating that systematic modeling is critical for reliable inference.
• The posterior distribution of the 21cm global signal amplitude is consistent with zero, suggesting the EDGES detection may be sensitive to modeling assumptions.
• The model producing a 21cm absorption line at z≈15 is consistent with the central value of the JWST-measured UV luminosity function at z=16.
• Constraints on the running of the primordial power spectrum are weak, but β_s can be constrained if the signal shape and position are well-determined.
• The posterior distributions of the UV luminosity function at z=12 and z=16 derived from EDGES data are in good agreement with recent JWST observations.