In standard cosmology, the spatial distribution of matter is traditionally summarized using two-point statistics, such as the power spectrum or the two-point correlation function. This mathematical framework was established in foundational works like Measuring the Galaxy Power Spectrum with Future Redshift Surveys by A. Hamilton et al. and applied to early survey data in The Angular Correlation Function of Galaxies from Early Sloan Digital Sky Survey Data by Andrew J. Connolly et al.
Two-point statistics dominate because they are computationally efficient and perfectly sufficient for a specific class of data: Gaussian random fields. If a physical field is Gaussian, its statistical properties are entirely defined by its mean and variance across different scales. The early universe, as captured by the Cosmic Microwave Background (CMB), is nearly perfectly Gaussian, making the power spectrum an exceptionally powerful tool for early-universe physics.
However, as the universe evolves, gravity drives non-linear structure formation. Matter collapses into dense clusters and long, interconnected filaments, leaving behind massive, under-dense voids. This late-time universe—the cosmic web—is highly non-Gaussian. When we compress a non-Gaussian map into a standard power spectrum, we discard the spatial phase information that defines these distinct shapes, creating a severe information bottleneck.
Why Two-Point Statistics Fail in the Late Universe
To understand why information is lost, consider what a two-point correlation function actually measures: it calculates the probability of finding a pair of objects (like galaxies) separated by a given distance $r$. It does not care where those pairs are located or how they connect globally.
Consequently, two fields can have the exact same power spectrum but completely different spatial structures. One field could be a random, isotropic blur of noise, while the other contains a highly organized network of cosmic filaments and voids. As noted by Yongzhong Xu et al. in Comparing and combining the Saskatoon, QMAP, and COBE CMB maps, ignoring phase information prevents us from capturing the true morphology of spatial fields.
As argued by Eiichiro Komatsu et al. in Non-Gaussianity as a Probe of the Physics of the Primordial Universe and the Astrophysics of the Low Redshift Universe, compressing millions of pixels or galaxy coordinates into a 1D power spectrum averages out higher-order correlations. This compression leaves vital clues about primordial inflation and non-linear dynamics unexploited.
This problem is particularly acute in weak gravitational lensing surveys, which map dark matter by measuring how its gravity subtly distorts the shapes of distant galaxies. Standard tools primarily capture the average density across scales, as discussed in Cosmological constraints with deep learning from KiDS-450 weak lensing maps by Janis Fluri et al. By failing to account for non-Gaussian features like skewness generated by gravitational collapse, traditional estimators fall short of the theoretical precision limits defined by the Cramér–Rao bound (APPROACHING THE CRAMÉR–RAO BOUND IN WEAK LENSING WITH PDF SYMMETRIZATION by Pengjie Zhang et al.).
Reclaiming Information via Higher-Order Statistics and Topology
To bypass this bottleneck, researchers have developed statistics that explicitly target non-Gaussianity. One direct approach is to measure higher-order moments, such as the third moment (skewness), which quantifies the asymmetry caused by dense clusters and empty voids.
For instance, the Dark Energy Survey collaboration demonstrated in Dark Energy Survey Year 3 results: cosmology with moments of weak lensing mass maps (M. Gatti et al.) that incorporating second and third moments of weak lensing mass maps improves cosmological constraints by up to 25%. However, calculating these higher-order moments traditionally suffers from a severe computational bottleneck due to the exponential growth of tensor size—a challenge that can be mitigated using implicit tensor decomposition techniques (Estimating Higher-Order Moments Using Symmetric Tensor Decomposition by Samantha Sherman et al.).
Beyond simple moments, researchers use geometric and topological descriptors to map the cosmic web's architecture:
- Peak Counts and Minkowski Functionals: Instead of averaging the entire field, these methods count high-density peaks or measure the topology (e.g., the number of isolated components, holes, and cavities) of the matter distribution. D. Zürcher et al. demonstrated in Cosmological forecast for non-Gaussian statistics in large-scale weak lensing surveys that these statistics can significantly increase the constraining power of weak lensing surveys while remaining robust against systematic errors.
- Marked Power Spectra: By weighting galaxies based on their local environment—such as up-weighting low-density voids—the authors of Cosmological Information in the Marked Power Spectrum of the Galaxy Field (Elena Massara et al.) tightened constraints on matter fluctuations ($\sigma_8$) by a factor of six.
- Persistent Homology: Borrowed from topological data analysis, this framework tracks how clusters, loops, and voids persist across different spatial scales. As shown by Jacky H. T. Yip et al. in Cosmology with Persistent Homology: a Fisher Forecast and Matteo Biagetti et al. in Fisher Forecasts for Primordial non-Gaussianity from Persistent Homology, persistent homology extracts multi-scale global connectivity that standard $n$-point functions miss entirely.
The Wavelet Scattering Transform
While higher-order moments capture non-Gaussianity, they quickly become noisy and unstable at very high orders. Convolutional Neural Networks (CNNs) can learn these patterns automatically, but they require massive training sets and function as uninterpretable "black boxes."
The Wavelet Scattering Transform (WST) provides an elegant middle ground. It mimics the architecture of a CNN but uses fixed, mathematically defined wavelet filters instead of learned weights. The operation follows a clear sequence:
- Convolution: The image is convolved with a directional wavelet filter to isolate specific scales and orientations.
- Modulus: A non-linear modulus operation ($|\cdot|$) is applied. This acts like a pooling layer, shifting high-frequency phase information into lower frequencies where it can be captured by subsequent layers.
- Iteration: The process is repeated on the output, generating a hierarchy of coefficients.
Input Image ──> [ Wavelet Filter ψ ] ──> [ Modulus |·| ] ──> First-order Coefficients
│
└──> [ Wavelet Filter ψ ] ──> [ Modulus |·| ] ──> Second-order Coefficients
As explained in Scattering Spectra Models for Physics by Sihao Cheng et al., WST provides a low-dimensional, robust statistical "fingerprint" of complex physical fields. Because it preserves phase correlations, it can easily distinguish between filamentary structures and random noise that share the same power spectrum (Comparing the morphology of molecular clouds without supervision by Erwan Allys et al.).
In cosmological applications, WST has proven highly effective. The authors of A new approach to observational cosmology using the scattering transform (Sihao Cheng et al.) demonstrated that WST coefficients outperform traditional methods and match state-of-the-art deep learning in parameter extraction without requiring any training. It has been successfully deployed to characterize the non-Gaussian textures of the interstellar medium (Statistical description of dust polarized emission from the diffuse interstellar medium by Erwan Allys et al.) and to separate faint galactic dust from the cosmic infrared background (Separation of dust emission from the cosmic infrared background in Herschel observations with wavelet phase harmonics by Constant Auclair et al.; Untangling dust emission and CIB anisotropies with the Scattering Transform Statistics by Srijita Sinha et al.).
Field-Level Inference and Deep Learning
The ultimate realization of extracting non-Gaussian information is field-level inference, where models process the entire high-dimensional map directly rather than compressing it into manual summary statistics.
Deep learning models, particularly 3D CNNs and Graph Neural Networks, excel at this task. Early work by Siamak Ravanbakhsh et al. (Estimating Cosmological Parameters from the Dark Matter Distribution) showed that 3D CNNs can estimate parameters directly from raw dark matter distributions by capturing the complex geometric patterns of the cosmic web. Similarly, Arushi Gupta et al. (Non-Gaussian information from weak lensing data via deep learning) demonstrated that CNNs could achieve parameter constraints up to five times tighter than the standard power spectrum by "seeing" the morphological information embedded in weak lensing maps.
To make these models interpretable, the authors of Interpreting deep learning models for weak lensing (José Manuel Zorrilla Matilla et al.) used saliency maps to reveal that neural networks dynamically shift their focus based on noise levels—prioritizing under-dense voids in clean data but pivoting to over-dense peaks when noise is high.
To scale these approaches to real-world surveys, researchers combine deep learning with Simulation-Based Inference (SBI). For example, the SIMBIG framework (A forward modeling approach to analyzing galaxy clustering with S im BIG by ChangHoon Hahn et al.) leverages normalizing flows and forward modeling to extract information from highly non-linear scales, yielding statistical gains equivalent to a 60% larger galaxy survey. By bypassing the limitations of analytic models on small scales, field-level and simulation-based approaches ensure that no non-Gaussian information is left behind.