Untangling dust emission and CIB anisotropies with the Scattering Transform Statistics

Srijita Sinha, Tuhin Ghosh, Erwan Allys, François Boulanger, J.‐m. DelouisarXiv 2026 · arXiv

TL;DR

Researchers have developed a new way to separate the "fog" of Galactic dust from the faint glow of the Cosmic Infrared Background (CIB) using a sophisticated mathematical tool called Scattering Transform (ST) statistics. By moving beyond simple linear correlations and capturing the complex, non-Gaussian textures of the sky, this method successfully identifies dust hidden in molecular gas regions that traditional techniques often miss.

What Problem Does This Paper Address?

In the quest to understand the early universe and the evolution of galaxies, astronomers face a massive "foreground" problem. Our own Milky Way is filled with thermal dust that glows at the same submillimeter frequencies as the Cosmic Infrared Background (CIB)—the collective light from all star-forming galaxies throughout cosmic history.

Traditionally, scientists separate the two by assuming dust lives exactly where atomic hydrogen (HI) gas is found. They use HI maps as a "template" to subtract the dust. However, this breaks down in complex regions where dust is also associated with molecular hydrogen ($H_2$), ionized gas, or "dark gas" that doesn't show up in HI surveys. This paper tackles the challenge of separating these components when the simple "dust follows HI" rule fails, particularly at intermediate and high Galactic latitudes.

Methods

The authors employ Scattering Covariance (SC) statistics, a subset of the Scattering Transform. Think of this as a middle ground between standard power spectrum analysis (which only looks at scales) and deep learning (which is powerful but often a "black box").

  1. Generative Modeling: They first identified 25 "clean" patches of sky where the HI-template method does work. They used these to train a generative model that understands the statistical "texture" of CIB anisotropies and instrumental noise.
  2. Component Separation: Instead of relying on a physical template, they used this statistical model as a constraint. They performed a pixel-based optimization to find a dust map that, when subtracted from the raw Planck data, leaves behind a residual that perfectly matches the known statistical properties of the CIB and noise.
  3. Validation: The algorithm was tested on synthetic simulations with varying signal-to-noise ratios (SNR 3 to 9) before being applied to real Planck 353 GHz observations.

Datasets

  • Planck 353 GHz (PR3): The primary intensity maps used to extract the dust signal.
  • HI4PI Survey: Used as the tracer for atomic hydrogen column density ($N_{HI}$).
  • CSFD Map: A standard corrected dust reddening map used for comparison.
  • FFP10 Simulations: End-to-end Planck noise simulations used to validate the statistical model.
  • CIB/LSS Maps (Chiang 2023): Used to cross-correlate and validate the recovered CIB residuals.

Key Insights and Ideas

  • Texture Matters: Standard power spectra ignore the "phases" of the signal—the actual shapes and filaments in the map. Scattering Transforms capture these non-Gaussian interactions across different scales, allowing the algorithm to distinguish between the "clumpy" CIB and the "filamentary" Galactic dust.
  • The $H_2$ Connection: The most significant "aha" moment came when comparing the recovered dust map to HI maps. The recovered map had a different power spectrum slope ($\Delta \alpha \approx 0.4$). This difference is a direct signature of dust associated with molecular hydrogen ($H_2$), proving the method can "see" dust that HI templates ignore.
  • Single-Frequency Power: Unlike many methods that require multiple frequency bands to separate components based on color, this approach works remarkably well at a single frequency by leveraging spatial statistics.

Key Results

  • Higher Fidelity: The recovered Planck dust map shows significantly more small-scale structure and "clumpiness" compared to the widely used CSFD map.
  • Statistical Separation: The algorithm successfully recovered the dust signal even at a low SNR of 3 in simulations, with a high correlation ($\rho = 0.97$) between the input and recovered maps.
  • Emissivity Ratios: By decomposing the recovered map, the authors estimated the ratio of dust emissivity between atomic and molecular phases. They found that $H_2$-associated dust at 353 GHz appears "colder" or has different properties than HI-associated dust, with an emissivity ratio of approximately 0.11.

Limitations and Future Work

  • Patch-Based: Currently, the method is applied to square patches ($222 \text{ deg}^2$). Expanding this to a full-sky map using the Foscat software package is the logical next step.
  • Stationarity Assumption: The method assumes that the statistical properties of the CIB are relatively consistent across the sky. While generally true for the CIB, local variations in noise or rare foregrounds could interfere.
  • Multi-Frequency Expansion: The authors plan to apply this to all Planck high-frequency channels ($\nu \ge 217 \text{ GHz}$). This would allow them to model the dust's "color" (modified black-body spectrum) and gain deeper insights into the physical temperature and properties of the interstellar medium.

Summary updated

Abstract

Template-fit approach is often used to separate the Galactic dust emission and the cosmic infrared background (CIB) anisotropies at low $\text{HI}$ column density regions with an underlying assumption that the gas and dust are tightly correlated. However, this method fails in regions where additional Galactic emission within the molecular hydrogen, diffuse ionized gas, and dark gas are present. We develop and test a statistical component separation to extract the dust signal from the contaminated $\textit{Planck}$ $353\,\rm GHz$ observations using the Scattering Covariance (SC) statistics. We first obtain a set CIB maps over $25$ square patches, each with a sky area of $222\,{\rm deg}^{2}$, using the linear correlation of dust and Galactic $21\,\rm cm$ $\text{HI}$ emission valid at low $\text{HI}$ column density regions using the template-fit approach. We then construct, from these $25$ maps, a generative model of CIB using the SC statistics. We finally rely on this contamination model to perform a component separation of dust and CIB in the $\textit{Planck}$ data for different sky regions. Applying our algorithm to the $\textit{Planck}$ $353\,\rm GHz$ observations, we recover a dust map for a test sky region that has more structures as compared to the corrected SFD map at $100\,μ\rm m$. The differences seen in the map level can be explained by decomposing the recovered $\textit{Planck}$ dust map into two gas phases: dust associated with $N_{\text{HI}}$ and dust associated with $N_{\text{H}_{2}}$. This work provides a clear pathway to map the Galactic interstellar reddening over intermediate and high Galactic latitudes.
BibTeX
@misc{sinha2026untangling,
  author = {Srijita Sinha and Tuhin Ghosh and Erwan Allys and François Boulanger and J.‐M. Delouis},
  title = {{Untangling dust emission and CIB anisotropies with the Scattering Transform Statistics}},
  year = {2026},
  eprint = {2603.06110},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.06110},
  howpublished = {arXiv},
}
Figure