The Holonomy of Optimal Mass Transport: The Smooth Case
arXiv:2608.15585
2026
Architecture
2 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper gives a constructive controllability principle: on a compact manifold, horizontal vector fields can be represented by a uniformly bounded number of Lie brackets of horizontal gradient fields, with bracket depth controlled by the distribution's step and immersion dimension. Via the Trotter property, compositions of flows generated only by these gradients become dense in the identity component of the diffeomorphism group. In the Riemannian case, the same result identifies diffeomorphic optimal-transport maps as a universal family of primitive transformations, suggesting neural ODE and generative architectures built from gradient or OT primitives.
Ideas from this paper
✗ Failed on benchmark
2026
Build a neural ODE or invertible transformation whose primitive layers are flows of learned gradient vector fields, then synthesize non-gradient directions using short Lie-bracket commutator products. The paper's bounded-bracket-generation result predicts that restricted gradient primitives can approximate a much larger class of diffeomorphisms than a plain stack of gradient flows.
Useful8/10
Difficulty6/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Parameterize a generative or density-evolving model as a composition of diffeomorphic optimal-mass-transport maps rather than unconstrained residual layers. Each layer transports one smooth positive density to another through a learned squared-distance OT map, while compositions provide a principled universal family for transformations connected to the identity.
Useful7/10
Difficulty7/10
Novelty6/10