generalfactneutralPushing a prior and posterior to the quotient space preserves empirical and population Gibbs risks while removing KL divergence artifacts from parameter symmetriesMachine Learning (Statistics)28 Jul 2026http://arxiv.org/abs/2607.18422v1