Sample size

CvxSPCA

[9]

[12]

[11]


n=50 k

20.9

207.1

48.7

3002

n=100k

26.2

466.9

78.3

3237.4

n=500 k

35.6

2737.06

2661.7

5276.93

n=1m

35.8

3408.59

3568

5274.26

 Since proposed CvxSPCA does not depend on eigenvalue decomposition or semidefinite programming, it is more scalable in terms of the sample size. It also requires less iterations to reach a desired sparsity