University of Crete (UOC)
Trustworthy non-Gaussian inference for weak-lensing cosmology
Abstract
dc:descriptionWeak gravitational lensing is one of the main cosmological probes of the dark sector and large-scale structure, used to constrain and test cosmological models by comparing data with predictions in a Bayesian framework. As the Stage-IV surveys (\textit{Euclid}, Rubin, and Roman) enter the era of high-precision lensing, the accuracy of their constraints is limited less by statistical noise than by how well the information in the data is extracted and modelled. Much of that information lies at small, non-linear scales, where the field is non-Gaussian and only partly characterised by two-point statistics. Extracting it requires reconstructing the matter maps, summarising them with higher-order or machine-learned statistics, and comparing these with forward simulations, since precise theoretical predictions and an exact likelihood are rarely available there. These scales are also the hardest to model, and the analyses that use them rest on simulators and learned models that can bias the result without any internal check detecting it, which is why they are not yet trusted for a survey's leading constraints. This thesis investigates what makes such non-Gaussian, simulation-based inference trustworthy when deep learning enters at every step, from the reconstruction to the summaries to the inference, and treats the calibration and robustness of every result as a design requirement. The first part concerns the maps.Reconstruction from the observed shear has long been treated as preprocessing, judged on the visual quality of the result. We show that it substantially affects the cosmological constraints. Comparing the established Kaiser--Squires inversion and its inpainted variant with the state-of-the-art MCALens in a single inference pipeline, we find that the reconstruction alone changes the four-parameter figure of merit by more than a factor of two, so the method should be chosen with the final cosmological analysis in mind, not on map quality alone. We then develop PnPMass, a new deep-learning mass-mapping method built for real surveys: as accurate as the best end-to-end networks but needing no retraining when the noise or footprint changes, fast at survey scale, and equipped with per-pixel uncertainties that carry a coverage guarantee. The second part extracts cosmology from the maps, assessing how much of the non-Gaussian information can be trusted once the systematics and the inference itself are accounted for. Higher-order statistics are widely thought to be especially vulnerable to baryonic feedback, whose imprint falls on the small scales that contain their signal. We show that much of the information instead lies on larger, uncontaminated scales, so a single conservative scale cut removes the bias at every survey area, and on those scales the wavelet \ell1-norm still constrains cosmology substantially better than the power spectrum.In the tomographic setting, we construct higher-order summaries that combine information across redshift bins, and investigate whether a learned neural compression is more constraining than a well-designed analytical statistic. Under a controlled comparison, a wavelet statistic evaluated jointly on pairs of bins is as constraining as the neural compressor, whose remaining advantage is that it combines the bins by construction. For the non-Gaussian, simulation-based analyses the coming surveys depend on, this thesis identifies which choices in reconstruction, summary, and inference control the accuracy of the result, and how the uncertainty of each can be calibrated and tested.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tersenov, Andreas
- Τερσένοβ, Ανδρέας
Subjects
dc:subject × 16- Ασϑενής βαρυτιϰή εστίαση
- Κοσµιϰή διάτµηση
- Χαρτογράφηση µάζας
- Στατιστικές υψηλότερης τάξης
- Συµπερασµατολογία βασισµένη σε προσοµοιώσεις
- Weak gravitational lensing
- Cosmic shear
- Mass mapping
- Higher-order statistics
- Simulation-based inference
- Φυσική
- Φυσικές Επιστήμες
- Αστρονομία και Αστροφυσική
- Physical Sciences
- Natural Sciences
- Astronomy and Astrophysics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62687