1: \begin{abstract}

2: \noindent \textbf{Abstract}: Sparse principal component analysis

3: (PCA) is an important technique for dimensionality reduction of high-dimensional

4: data. However, most existing sparse PCA algorithms are based on non-convex

5: optimization, which provide little guarantee on the global convergence.

6: Sparse PCA algorithms based on a convex formulation, for example the

7: Fantope projection and selection (FPS), overcome this difficulty,

8: but are computationally expensive. In this work we study sparse PCA

9: based on the convex FPS formulation, and propose a new algorithm that

10: is computationally efficient and applicable to large and high-dimensional

11: data sets. Nonasymptotic and explicit bounds are derived for both

12: the optimization error and the statistical accuracy, which can be

13: used for testing and inference problems. We also extend our algorithm

14: to online learning problems, where data are obtained in a streaming

15: fashion. The proposed algorithm is applied to high-dimensional gene

16: expression data for the detection of functional gene groups.\medskip{}

17: 

18: \noindent \textbf{Keywords}: sparse principal component analysis,

19: dimensionality reduction, convex optimization, gradient method, online

20: learning.

21: \end{abstract}