NS-RGS: Newton-Schulz based Riemannian gradient method for orthogonal group synchronization

This collection of recent ML/statistics papers presents new algorithms and theoretical guarantees across optimization, inference, generative modeling, and privacy. Key contributions: NS-RGS proposes a Newton–Schulz based Riemannian gradient scheme for orthogonal group synchronization replacing costly SVD/QR with matrix multiplications for GPU/TPU speedups and linear convergence guarantees; a variational REML (VREML) method yields scalable exact inference for Gaussian ICAR spatial models; identifiability results for transport maps and vector fields from finitely many pushforward densities with伴