TL;DR
This paper introduces a novel "neural marked statistics" approach to extract more cosmological information from galaxy surveys than traditional methods. By learning interpretable, morphology-aware transformations of the cosmic density field, they significantly improve constraints on key cosmological parameters like $\Omega_m$ and $\sigma_8$, even breaking long-standing degeneracies.
What Problem Does This Paper Address?
Cosmological surveys like DESI and Euclid are generating vast 3D maps of the universe. Extracting the maximum scientific return from these surveys requires sophisticated statistical tools. The standard approach relies on the power spectrum, which describes the two-point correlations of the matter density field. However, much of the interesting cosmological information, particularly related to non-Gaussianities and dark energy, lies in higher-order correlations or the complex morphology of the cosmic web. Traditional methods to capture this extra information, like "marked statistics," involve hand-designed functions that reweight the density field. While effective, these functions are often limited in their flexibility and interpretability. This paper aims to bridge this gap by developing a learned, interpretable, and more powerful way to extract information beyond the simple power spectrum.
Methods
The core innovation is a "neural mark" – a learned function that transforms the density field locally. This mark is designed to be:
- Interpretable: It's built from physically motivated, rotation-invariant features derived from spherical harmonic decompositions of the density field (capturing monopole, dipole, and quadrupole-like structures). These features are then processed by independent neural networks (MLPs).
- Cosmology-Agnostic: The mark itself is trained without knowledge of specific cosmological parameters.
- Contrastive Learning: To align the marked summary statistics with cosmological parameters, the authors employ a contrastive learning objective (specifically, InfoNCE). This encourages the learned representations of the marked field to be close to the true cosmological parameters while being far from incorrect parameter sets.
- Residualization: A crucial detail is that the learned mark's contribution is residualized against the standard power spectrum. This ensures the mark is rewarded only for the complementary information it provides, not for simply re-encoding what the power spectrum already tells us.
The output of this process is a set of marked two-point statistics (power spectra of the marked field), which are then used for cosmological inference.
Datasets
- Training and Evaluation: The authors utilize the Quijote N-body simulation suite. This suite provides a large number of simulations with varying cosmological parameters.
- A subset of 5,000 simulations was used for contrastive training.
- A separate set of simulations was used for held-out generalization tests and Fisher forecasting.
- Benchmark: The paper compares its results against "classical marks" from existing literature, specifically those optimized using Fisher information maximization.
Key Insights and Ideas
- Morphology Matters: The learned mark effectively captures information encoded in the shape and structure of the cosmic web, not just its density fluctuations. This is a significant step beyond the isotropic assumption of the power spectrum.
- Interpretability is Achievable: By designing the neural mark from interpretable components (spherical harmonics, MLPs), the authors can dissect which morphological features contribute most to the improved constraints. They find that an isotropic component dominates, but anisotropic features provide crucial refinements.
- Contrastive Learning for Alignment: The contrastive objective proves highly effective at aligning the complex, learned representations of the marked field with the underlying cosmological parameters in a shared latent space.
- Complementarity is Key: Explicitly residualizing the marked summary against the unmarked power spectrum forces the learned mark to focus on genuinely new information, leading to more efficient gains.
- Breaking Degeneracies: The learned mark demonstrably breaks the long-standing $\Omega_m - \sigma_8$ degeneracy, a critical achievement for precision cosmology.
Key Results
- Fisher Constraints: At $k_{\max} = 0.2\,h\mathrm{Mpc}^{-1}$, the neural mark tightens marginalized constraints by:
-
$2.9\times$ for $\sigma_8$
- $1.8\times$ for $\Omega_m$ This is compared to classical marks.
-
Parameter MSE: The learned mark reduces the mean-squared error (MSE) for cosmological parameters across the prior space by up to $1.45\times$ compared to the best classical mark.
- Latent Space Geometry: Principal component analysis of the learned summary embedding reveals that the first two components align almost perfectly with $\Omega_m$ and $\sigma_8$, directly reflecting the improved constraints and broken degeneracy.
- Generalization: The approach shows good generalization performance on unseen simulations, outperforming classical marks in both Fisher forecasting and a held-out MLP regression task.
Limitations and Future Work
- Resolution and Scale Dependence: At very small scales ($k_{\max} \gtrsim 0.3\,h\mathrm{Mpc}^{-1}$), the anti-aliasing filters used in the spherical harmonic decomposition reduce the advantage of the learned mark over classical ones. The authors suggest that morphology-based filters could restore superiority.
- Richer Invariants: While the current mark uses interpretable components, the authors suggest exploring "richer geometrically motivated invariants" for even more refined interpretability and targeting specific morphological features.
- Broader Applications: The framework is generalizable to other cosmological parameters beyond $\Omega_m$ and $\sigma_8$, such as primordial non-Gaussianity, which inherently requires higher-order statistics.
- Computational Cost: While the training is described as taking ~3 hours on a single GPU, the inference (applying the mark and computing spectra) is efficient. The paper does not explicitly discuss the computational cost of generating the large number of simulations required for training and evaluation.
Summary updated