We prove the asymptotic normality of PCA on the Grassmannian, and derive a tight non-asymptotic bound on its excess risk using self-concordance.

We show that high-dimensional neural activity can arise from low-dimensional latent dynamics, both in RNNs and in the brain.