Dimension-Free Lipschitz Bounds for Brenier Maps to Compactly Supported Log-Concave Targets
arXiv:2608.15906
2026
Architecture
2 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper gives an explicit, dimension-free smoothness certificate for the optimal transport map from a semi-log-concave source to a compactly supported log-concave target. The transferable asset is a directional Hessian bound controlled by the source curvature matrix and the target's directional width, avoiding the usual dimension-dependent diameter factor. This can be used to initialize or constrain neural optimal-transport layers, bounded latent-to-data maps, and transport-based generative models so their Jacobians have a principled global scale. A second practical use is a directional curvature regularizer that penalizes violation of the theorem's bound during finite neural approximation of the Brenier potential.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Build a neural transport layer by parameterizing a convex potential whose gradient maps a semi-log-concave latent distribution into a compact convex data domain. Use the paper's dimension-free Lipschitz certificate to set the layer's Jacobian scale, initialize the potential, and reject or regularize parameter updates that create excessive curvature. The goal is a bounded-output transport module that is less sensitive to latent dimension than diameter-based spectral heuristics.
Useful7/10
Difficulty6/10
Novelty7/10
Unverified
2026
Add a theorem-guided regularizer to neural optimal-transport potentials that limits curvature separately in each direction according to the target support width. Unlike an isotropic Hessian penalty, it permits larger curvature along directions where the target is wide and enforces stronger smoothing along narrow directions, preserving anisotropic structure while controlling the transport map's Lipschitz constant.
Useful6/10
Difficulty4/10
Novelty8/10