1: \begin{abstract}

2: We describe and analyze a simple algorithm for principal component analysis

3: and singular value decomposition, VR-PCA, which uses computationally cheap

4: stochastic iterations, yet converges exponentially fast to the optimal

5: solution. In contrast, existing algorithms suffer either from slow

6: convergence, or computationally intensive iterations whose runtime scales

7: with the data size. The algorithm builds on a recent variance-reduced

8: stochastic gradient technique, which was previously analyzed for strongly

9: convex optimization, whereas here we apply it to an inherently non-convex

10: problem, using a very different analysis.

11: \end{abstract}

12: