A mathematical theory for understanding when abstract representations emerge in neural networks

Their proof of approximately orthogonal latent-variable representations at global minima of feedforward networks is cited in riva2026task as a parallel structural result for a different architectural family.