The Theory behind UMAP?

A new academic paper titled 'The Theory behind UMAP?' by David Wegmann has identified and corrected significant foundational errors in the theoretical underpinnings of the widely-used UMAP (Uniform Manifold Approximation and Projection) algorithm. The work, submitted to arXiv and derived from a master's thesis, reveals that the original 2018 paper by McInnes et al. introduced a finite variant of a mathematical construct called the 'metric realization' functor, based on an unpublished draft by mathematician David Spivak. This foundational draft contained numerous errors that were subsequently reproduced in the UMAP publication and later literature, creating a flawed theoretical basis for the popular tool.

The paper's primary contribution is a complete, corrected, and self-contained derivation of Spivak's functors and McInnes et al.'s finite variant, providing an explicit description of the metric realization. By repairing these errors, Wegmann aims to establish a solid mathematical foundation for UMAP, which is crucial for researchers who rely on its theoretical properties for high-dimensional data visualization and analysis. The work also discusses the UMAP algorithm itself and examines claims about its properties, potentially impacting how the algorithm's reliability and theoretical guarantees are understood and applied in machine learning and statistics moving forward.