🔭 astrophysics

Generative Models of 21cm EoR Lightcones with 3D Scattering Transforms

This paper extends the scattering transform formalism to three-dimensional Epoch of Reionization (EoR) lightcones by introducing a 3D wavelet set and constructing maximum entropy generative models that successfully synthesize realistic EoR data, thereby laying the groundwork for advanced component separation techniques capable of recovering valuable non-Gaussian information obscured by foregrounds.

Original authors: Ian Hothi, Erwan Allys, Benoit Semelin, Romain Meriot

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ian Hothi, Erwan Allys, Benoit Semelin, Romain Meriot

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

The Big Picture: Finding a Whisper in a Hurricane

Imagine the early universe, just after the Big Bang, as a giant, dark room. Suddenly, the first stars turn on like lightbulbs. This era is called the Epoch of Reionization (EoR). Scientists want to study this "lighting up" moment by listening to a specific radio signal (the 21cm signal) emitted by the hydrogen gas in that room.

However, there is a massive problem. Trying to hear this faint cosmic whisper is like trying to hear a single person whispering in the middle of a roaring hurricane. The "hurricane" is made of bright radio noise from our own galaxy, the Earth, and the telescopes themselves. This noise is thousands of times louder than the signal we want.

The Problem: The Old Tools Are Too Simple

For a long time, scientists tried to filter out the noise by looking at the "volume" of the signal (called the power spectrum). It's like trying to separate a whisper from a roar just by looking at a volume meter. The problem is that the cosmic signal isn't just a simple hum; it has a complex, messy, and irregular shape (it is non-Gaussian). The old volume-meter tools throw away all that interesting shape information, leaving scientists with a blurry picture.

The New Tool: The "Scattering Transform"

The authors of this paper propose a new way to look at the data. They use a mathematical tool called the Scattering Transform.

Think of the Scattering Transform as a high-tech sieve or a multi-layered coffee filter.

Instead of just measuring the total volume, this sieve breaks the signal down into tiny, specific textures and patterns.
It looks at how the signal changes in different directions: side-to-side (the sky) and up-and-down (time/redshift).
It captures the "shape" of the data, not just its loudness.

The Innovation: A 3D Sieve

Previous versions of this sieve only worked on flat, 2D pictures (like a single photo of the sky). But the EoR signal is a 3D movie (a "lightcone") that changes as you look deeper into space and time.

The authors invented a 3D Scattering Transform.

The Analogy: Imagine you have a loaf of bread. A 2D sieve only looks at the crust of one slice. The new 3D sieve looks at the texture of the crumb inside the loaf, how the slices stack on top of each other, and how the whole loaf is shaped.
They built this by combining a 2D filter (for the sky) with a 1D filter (for the time/redshift dimension) to create a custom 3D filter that fits the shape of the universe perfectly.

The Solution: Building a "Fake" Universe

The paper doesn't just analyze data; it builds a generative model.

The Analogy: Imagine you have one perfect, rare painting (the real simulation of the universe). You want to create a machine that can paint new pictures that look exactly like the original, but with different details.
The authors fed their 3D sieve into a computer program. The program learned the "rules" of the universe's texture (the Scattering Transform statistics) from that one perfect painting.
Then, the program started painting 30 new, fake universes.

The Results: A Perfect Copycat?

The authors checked if their "fake" universes were good copies. They compared the fake ones to the original using several tests:

Visual Check: They looked at slices of the fake universes. The fake ones looked just like the real ones, with the same dark spots (absorbing gas) and bright spots (ionized gas).
The Histogram (The "Shape" of the Data): The real universe has a very weird, lopsided shape in its data distribution. The fake universes copied this lopsided shape perfectly, whereas a simple "Gaussian" (bell curve) model failed completely.
The Texture (Minkowski Functionals): They checked the complexity of the shapes (like how connected the bubbles of gas are). The fake universes matched the real ones very well, though they struggled slightly with the sharpest, most high-contrast edges.

The Limitations

The authors are honest about where their tool isn't perfect yet:

The "Blurry" Edges: When looking at the very smallest details (high frequencies), the fake universes got a little bit blurry or biased. This happened partly because of the complex math used to turn the data back into its original form.
The "Round" Bubbles: The model had a little trouble recreating the perfectly sharp, circular edges of the ionized bubbles seen in the real data.

Conclusion

In short, this paper proves that we can take a complex, 3D simulation of the early universe, analyze its unique "texture" using a new 3D mathematical sieve, and then use that knowledge to generate new, realistic 3D universes.

This is a crucial first step. By proving we can generate realistic "fake" universes that capture the true complexity of the signal, scientists can eventually use these models to help strip away the "hurricane" of noise and finally hear the "whisper" of the first stars.

Technical Summary: Generative Models of 21cm EoR Lightcones with 3D Scattering Transforms

Problem Statement
The Epoch of Reionization (EoR) is characterized by the redshifted 21cm signal from neutral hydrogen, observed as a three-dimensional lightcone comprising two spatial sky axes and one redshift (frequency) axis. Extracting this signal is challenging because it is obscured by foregrounds (e.g., Galactic synchrotron emission) and instrumental systematics that are orders of magnitude stronger. While current foreground separation techniques focus on recovering the EoR power spectrum, they often neglect the highly non-Gaussian nature of the EoR signal, which contains crucial astrophysical information regarding the progression of reionization. Existing component separation methods based on Scattering Transforms (ST) have been successful in two-dimensional contexts (e.g., Galactic interstellar medium) but have not yet been extended to the three-dimensional data structure of EoR lightcones.

Methodology
The authors propose a framework to construct and validate maximum entropy generative models for 3D EoR lightcones using Scattering Transform statistics. The methodology involves three primary stages:

Construction of 3D Wavelets: To address the specific structure of EoR data (where foregrounds dominate low parallel wavenumbers $k_\parallel$ ), the authors define a new 3D wavelet set. This is constructed via the tensor product of:
- A 1D isotropic wavelet set defined along the redshift ( $k_\parallel$ ) axis.
- A 2D isotropic wavelet set defined in the sky plane ( $k_\perp$ ) domain.
  This separation allows the wavelets to characterize the $k_\parallel$ vs. $k_\perp$ plane without mixing regions where foregrounds are dominant with those where they are sub-dominant.
Scattering Transform Statistics: The authors utilize scattering covariance statistics ( $C_{p1,p2}$ ) derived from convolutions of the lightcone with these wavelets, followed by non-linear modulus operations. The statistics include:
- $C_{11}$ : Covariance of linear wavelet coefficients (power spectrum constraints).
- $C_{01}$ and $C_{00}$ : Covariances involving modulus operations to capture non-Gaussian phase information.
- An additional set of $C'_{11}$ coefficients computed with a finer scale resolution to improve spectral constraints.
  These coefficients are normalized by a reference field (the target lightcone) to create scale-invariant statistics.
Generative Model and Data Conditioning:
- Quantile Transformation: To handle the highly skewed histogram of the EoR brightness temperature, the target lightcone is first transformed via a quantile function to approximate a Gaussian distribution. The generative model is trained in this transformed space.
- Maximum Entropy Sampling: New realizations are generated by minimizing a loss function (Euclidean distance between the ST statistics of the target and the synthesized field) using gradient descent (L-BFGS-B) on a white noise initialization.
- Inverse Transformation: The synthesized fields are mapped back to the original data space using the inverse quantile transform.
- Boundary Handling: To mitigate non-periodic boundary effects inherent in lightcone simulations, statistics are estimated on a cropped central volume, while generated samples are full periodic cubes. Unconstrained large and small scales are handled by substituting target modes or filtering, respectively.

Key Contributions

Extension to 3D: The paper extends the Scattering Transform formalism from 2D datasets to 3D EoR lightcones, introducing a specific tensor-product wavelet set tailored to the anisotropic nature of the data.
Generative Model Construction: It demonstrates the construction of a maximum entropy generative model capable of synthesizing realistic EoR lightcones from a single simulated target.
Validation Framework: The authors provide a rigorous statistical validation of the synthesized lightcones against the target using independent metrics: histograms, power spectra (3D, 2D, and cylindrically averaged), and Minkowski Functionals.

Results
The study validates the model using a single simulated lightcone from the LoReLi II database (redshift range $9.5 \le z \le 8.5$ ).

Visual and Statistical Agreement: The synthesized lightcones visually reproduce the target's distinct features, including localized absorbing regions and ionized bubbles.
Histograms: The model successfully recovers the highly skewed, non-Gaussian histogram of the target lightcone over four orders of magnitude.
Power Spectra: The 3D and 2D power spectra of the synthesized lightcones match the target well on large scales. However, a bias is observed at high wavenumbers (small scales), particularly along the redshift axis, which the authors attribute to the non-linear inverse quantile transform failing to perfectly reproduce non-Gaussian structures at small scales.
Minkowski Functionals: The morphological complexity (genus, area, perimeter) is generally well-reproduced. However, the model struggles to fully replicate the sharp, high-contrast circular features (ionized regions) in the sky plane, leading to deviations in the $M_2$ statistic.
Variance: The sample variance between synthesized lightcones is slightly lower than the variance estimated from octants of the target, suggesting the model may slightly underestimate the full lightcone-to-lightcone variability, though this is difficult to quantify precisely with a single target.

Significance and Claims
The paper presents this work as a "proof-of-concept" for applying 3D scattering transform-based generative models to EoR lightcones. The authors claim that the results demonstrate the feasibility of synthesizing realistic EoR data that preserves key non-Gaussian properties and morphology. They posit that this capability is a necessary prerequisite for developing future component separation algorithms that rely on 3D scattering transforms to separate the EoR signal from foregrounds and noise, moving beyond simple power spectrum recovery to utilize the full non-Gaussian information content of the data. The authors remain modest regarding current limitations, specifically noting the biases at small scales and the difficulty in reproducing sharp morphological features, suggesting these could be addressed by refined wavelet treatments or more sophisticated ST representations in future work.