We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: On the Optimality of the Oja's Algorithm for Online PCA

Authors: Xin Liang
Abstract: In this paper we analyze the behavior of the Oja's algorithm for online/streaming principal component subspace estimation. It is proved that with high probability it performs an efficient, gap-free, global convergence rate to approximate an principal component subspace for any sub-Gaussian distribution. Moreover, it is the first time to show that the convergence rate, namely the upper bound of the approximation, exactly matches the lower bound of an approximation obtained by the offline/classical PCA up to a constant factor.
Comments: 25 pages. arXiv admin note: text overlap with arXiv:1711.06644
Subjects: Machine Learning (cs.LG); Optimization and Control (math.OC)
MSC classes: 62H25, 68W27, 65F15
Cite as: arXiv:2104.00512 [cs.LG]
  (or arXiv:2104.00512v2 [cs.LG] for this version)

Submission history

From: Xin Liang [view email]
[v1] Wed, 31 Mar 2021 15:02:54 GMT (28kb)
[v2] Tue, 18 Jan 2022 14:22:56 GMT (29kb)

Link back to: arXiv, form interface, contact.