Random Quadratic Form with random forcing: Metastable synchronization by noise

What the math gives to ML

The paper provides a concrete projected stochastic flow on the sphere in which multiplicative random quadratic dynamics preserve antipodal symmetries, while arbitrarily small independent tangent forcing eventually breaks those symmetries and synchronizes trajectories. This suggests a principled noise injection mechanism for normalized token states or continuous-depth transformer dynamics: use shared symmetric-matrix noise for structured exploration and weak token-specific tangent noise to prevent persistent antipodal or cluster degeneracies. The most testable transfer is a train-time stochastic residual block with explicit sphere retraction, comparing symmetry breaking, representation collapse, robustness, and optimization behavior against ordinary Gaussian residual noise.