{"id":"arxiv/2609.25085","url":"https://arxiv.org/abs/2609.25085","tags":["stat.ML","cs.LG","math.NA"],"score":0,"title":"FREESIA: Covariance-Aware Posterior Transport for Expressive and Scalable Data Assimilation","domain":"stat.ML","source":"arxiv","authors":["Shiwei Ni","Yangwen Zhang","Hang Qi","Xiaofei Guan","Lili Ju"],"reviews":0,"subject":"stat.ML","summary":"Data assimilation aims to infer the state of complex dynamical systems based on observational data. However, accurate inference of the multimodal posteriors induced by nonlinear or non-injective observation operators remains a key challenge under high-dimensional and sparse observation conditions. Ensemble filters scale to high dimensions but are confined by restrictive distributional assumptions, while training-free generative filters (e.g., EnSF, EnFF) alleviate this limitation but may introduce structural errors and hinder information propagation under sparse observations. To address…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-09-19","reactions":{}}