benchmarksfactneutralFisher width and inverse-Fisher width play complementary roles in statistical learning and recovery on statistical manifoldsMachine Learning (Statistics)28 Jul 2026http://arxiv.org/abs/2607.20578v1