From data to knowledge

A measurement is only as trustworthy as the method behind it. Much of my work sits on the bridge between theory and observation: how do we know we have genuinely learned something new about a model from data — rather than fooled ourselves with a systematic or an over-flexible fit? This theme collects the statistical and machine-learning methods that make the cosmology defensible.

Selected threads

• Observational systematics in large-scale structure. In Karim et al. (2023), MNRAS — first author — I identified a new systematic: the bias introduced when measuring the imaging-systematics weights themselves distorts the galaxy power spectrum and, in turn, cosmological inference.
• Model selection for dark energy. Applying the expected log pointwise predictive density (ELPD) as a principled criterion in the $w_0 w_a$-versus-$\Lambda$CDM debate — asking which dark-energy model the data actually prefer, and by how much. Because ELPD decomposes point by point, it can also identify which measurements drive a preference: try the interactive DESI DR2 leave-one-tracer-out explorer.
• Faster inference with emulators (student-led). Improving the COSMOPOWER emulator through hyper-parameter tuning, to accelerate cosmological likelihood evaluation.
• Forecasting a redshift-evolving σ₈ (student-led). Forecasting how well an LSST-era analysis could constrain a redshift-evolving $\sigma_8$ model — a direct handle on the growth-of-structure tension.