On the Unique Recovery of Transport Maps and Vector Fields from Finite Measure-Valued Data

Researchers Jonah Botvinick-Greenhouse and Yunan Yang have published a significant theoretical paper titled "On the Unique Recovery of Transport Maps and Vector Fields from Finite Measure-Valued Data" that establishes mathematical guarantees for uniquely identifying generative models from limited observations. Their work addresses a fundamental question in machine learning: when can we uniquely determine a transport map (the function that transforms one probability distribution into another) from observing its effects on just a finite number of input densities? This is precisely what happens when training generative AI models like diffusion models or normalizing flows.

The paper proves that under general conditions, a diffeomorphism (a smooth, invertible transformation) can be uniquely identified from its pushforward action on finitely many densities. Using the Whitney and Takens embedding theorems, the researchers provide estimates for how many density measurements are needed, depending only on the intrinsic dimension of the problem. They introduce a new metric that compares diffeomorphisms by measuring discrepancies between their pushforward densities, offering a principled way to evaluate generative model performance.

Beyond static maps, the researchers also prove analogous results for vector fields in infinitesimal settings, where derivatives of densities along smooth vector fields are observed. This connects to dynamical systems and PDE inverse problems, with applications to continuity equations, advection-diffusion-reaction equations, and Fokker-Planck equations. The work provides theoretical foundations for understanding when AI models are uniquely determined by their training data and offers new tools for analyzing model identifiability in practical applications.