Stochastic trace estimation with tensor train random vectors
Researchers have developed a new method for stochastic trace estimation using Gaussian random tensor train vectors. This approach offers a structured alternative to traditional methods, particularly for tensor-structured settings where unstructured vectors can be computationally expensive. The proposed technique, when applied with an appropriate tensor train rank, provides dimension-independent guarantees for the Girard--Hutchinson estimator and can achieve similar accuracy to classical methods. Furthermore, the study explores the integration of these sketches into the Nyström++ framework, potentially improving sample complexity under specific conditions. AI
IMPACT Introduces a more efficient method for matrix trace estimation in tensor-structured settings, potentially improving performance in ML algorithms.