Linear ICA via Optimal Transport Replaces the Classical Non-Gaussianity Assumption

arXiv·low signal

Ashutosh Jha, Michel Besserve, and Simon Buchholz (arXiv 2607.14081, cs.LG/stat.ML) recast linear Independent Component Analysis as an optimal-transport problem, departing from the classical route of maximizing non-Gaussianity to recover independent sources from linear mixtures. This is a foundational-methods paper rather than an applied one — the practical payoff is in identifiability, which matters for representation-learning and causal-discovery work downstream. Low immediate applicability for most builders; genuine interest if you work on disentanglement.