The Normal Procrustes Problem: A Riemannian Optimization Approach

arXiv:2608.19513 2026 Architecture 1 ideas extracted · analyzed Sep 1, 2026

What the math gives to ML

The paper turns least-squares fitting of a matrix into a constrained problem over normal operators, defined by commuting with their adjoint: $A^*A=AA^*$. This suggests a spectral linear layer whose weight is parameterized as $A=U\operatorname{diag}(\lambda)U^*$, separating a unitary eigenbasis from learnable eigenvalues while guaranteeing normality and enabling direct control of the spectrum. The most practical transfer is to use this layer for stable recurrent or state-space transitions, with Procrustes fitting for initialization and Riemannian updates that preserve unitarity exactly.

Ideas from this paper

Unverified 2026

Normal Spectral Linear Layer

Replace an unconstrained dense transition or recurrent matrix with a normal matrix $A=U\operatorname{diag}(\lambda)U^*$, where $U$ is unitary and $\lambda$ contains learnable eigenvalues. The layer can be initialized by fitting a normal matrix to input-output pairs through the paper's objective, then trained with Riemannian updates that keep $U$ unitary and preserve the normal-operator structure.

Useful6/10
Difficulty6/10
Novelty6/10
Paper: The Normal Procrustes Problem: A Riemannian Optimization Approach arXiv:2608.19513