Model tension for astrophysics

Guardiani, Matteo (2026): Model tension for astrophysics. Dissertation, LMU München: Faculty of Physics

Abstract

Astrophysics is fundamentally a non-interventional science. Physical theories are tested not through controlled experiments, but through comparison with indirect observations shaped by noise, selection effects, and the response of complex instruments. Extracting reliable information from observational data therefore requires probabilistic forward models that support not only reconstruction, but also uncertainty quantification, model criticism, and model comparison. This thesis develops Bayesian methods for these tasks from the perspective of information field theory. We begin by showing how to test competing explanations in settings where interventions are impossible. To this end, we develop practical approaches to evidence-based model comparison in high-dimensional hierarchical generative models. Furthermore, we introduce novel moment-constrained models for bivariate causal inference, extending causal reasoning to purely observational settings and thereby addressing a central challenge shared by many problems in astrophysics. We then explore standardized latent representations as diagnostic tools: posterior latent fields can reveal where model assumptions are under tension with the data, indicating missing structure or model misspecification. We apply this idea to astrophysical imaging by introducing latent-space field tension for automated component detection, using it to identify point-like and extended emission in X-ray observations. These analyses are supported by J-UBIK, a JAX-accelerated framework that combines signal models, instrument response operators, and scalable variational inference in a common pipeline. We validate our methods on synthetic data and apply them to SRG/eROSITA observations, presenting multiband Bayesian reconstructions of the SN1987A region in the Large Magellanic Cloud: denoised, deconvolved, and decomposed images with posterior uncertainty estimates. We also discuss advances on multifrequency strong gravitational lensing with LensCharm, and present the first joint reconstruction of SPT 2147–50 combining ALMA and JWST data within a unified lensing model using the same framework. In this work, we go beyond the view of inference as a tool to estimate parameters or, at best, to reconstruct continuous fields from data. We address the questions: What happens when a model class is strained by the data? Where does it fail? Which competing explanations remain plausible? By combining information field theory, standardized latent parametrizations, evidence estimation, and practical software implementations, we developmethods that scale to high-dimensional inference problems and turn model tension into a diagnostic signal—one that guides model refinement and supports principled comparison across hypotheses.

• Item Type: Theses (Dissertation, LMU Munich)
• Subjects: 500 Natural sciences and mathematics; 500 Natural sciences and mathematics > 530 Physics
• Faculties: Faculty of Physics
• Language: English
• Date of oral examination: 13. May 2026
• 1. Referee: Enßlin, Torsten
• Signature of the printed copy: 0001/UMC 32045