Speaker
Description
Modern CMB and large-scale-structure analyses face some of the same statistical bottleneck: the data are high-dimensional, non-Gaussian, affected by complex systematics, and often have intractable likelihoods, while the end-to-end simulations needed for Monte Carlo validation or Simulation-Based Inference are prohibitively expensive. I will present a unified framework based on the Scattering Covariance: an interpretable, physics-informed analogue of a convolutional neural network, built from fixed oriented wavelets, nonlinearities, and cross-scale/cross-channel covariance statistics.
In recent work (Campeti et al., A&A, 2025), we used this representation to construct a fast map-level generative emulator for CMB instrumental systematics simulations. Even when trained on as few as ten high-fidelity simulations, the emulator generates statistically independent approximate realizations that reproduce power spectra, scattering statistics, Minkowski functionals, and pixel-covariance structure, enabling orders-of-magnitude simulation augmentation at negligible cost compared with full end-to-end campaigns.
I will then describe how the same scattering-covariance latent space can be used for transparent SBI pipelines: for CMB polarization, to infer parameters such as the optical depth to reionization and the tensor-to-scalar ratio while controlling non-stationary foreground and instrumental residuals; and for weak-lensing cosmology, to emulate high-resolution non-Gaussian convergence maps and infer parameters such as $\Omega_m$ and $\sigma_8$ beyond two-point statistics. The goal is a simulation-efficient, interpretable alternative to ``black-box'' neural inference: using a few expensive simulations as anchors for large, calibrated, statistically-controlled inference pipelines.
| Other topic / keywords: | Weak Lensing, Simulation-Based Inference |
|---|