Reproduce “Density estimation on low-dimensional manifolds: an inflation-deflation approach” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Christian Horvat, Jean-Pascal Pfister
A new "inflation-deflation" approach enables Normalizing Flows to perform exact density estimation on low-dimensional Riemannian manifolds, addressing their topological limitations. The method demonstrates strong empirical performance across diverse synthetic manifolds and offers significant computational efficiency over existing techniques for real-world data applications like out-of-distribution detection.
Key insights
- Exact manifold density can be recovered through a principled inflation-deflation process if the "Q-normal reachability" condition is met, ensuring a unique mapping from the inflated space back to the manifold.
- When the manifold's normal space is unknown, isotropic Gaussian noise in the ambient space can serve as a robust approximation for inflation, especially in higher embedding dimensions, enabling broad applicability without explicit geometric knowledge.
- Theoretical bounds for the optimal noise magnitude (σ²) can be derived based on the manifold's curvature and the target density's characteristics, guiding hyperparameter selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.