New WT-PCA method improves log-PCA for probability measures with convergence guarantees
Xu et al. introduce a dynamical formulation for learning principal variations under Wasserstein geometry...
A team of researchers led by Peng Xu (arXiv:2606.17196) has developed a new approach to principal component analysis for probability measures, called Wasserstein Tangential PCA (WT-PCA). The work extends log-PCA, a linearized method for analyzing variations in probability distributions under the Wasserstein geometry, by introducing a dynamical variational formulation. WT-PCA computes local principal modes of geodesic variation by leveraging the covariance operator at the Wasserstein barycenter, giving a more principled way to capture the intrinsic structure of random probability measures.
Critically, the paper provides rigorous statistical convergence guarantees—showing that empirical WT-PCA estimates converge to the population version in terms of the 2-Wasserstein distance between the barycenters. This fills a gap in the theoretical understanding of log-PCA and opens the door to more reliable dimensionality reduction for functional data analysis, generative modeling, and any domain where distributions (not just points) are the primary data type. The work is published under the machine learning category and connects to optimal transport theory.
- Introduces Wasserstein Tangential PCA (WT-PCA) as a differentiable, dynamical reformulation of log-PCA for probability measures.
- Derives a statistical convergence rate for empirical WT-PCA using the 2-Wasserstein distance between population and empirical barycenters.
- Leverages parallel transport structure in optimal transport to capture local principal geodesic modes via a covariance operator.
Why It Matters
Provides a theoretically grounded method for PCA on distributions, enabling better analysis of shape, image, and other non-Euclidean data.