Yuhao smp pca

[heyyuhao]

Experiment on my laptop
can see smp-pca balances the time and error

What I did
Implement the algorithm of https://arxiv.org/abs/1610.06656 Single Pass PCA of Matrix Products

• [done] sketch matrix
• [done] scaled Johnson-Lindenstrauss (JL) in matrix form
• [skipped] element-wise sampling As I observe, element-wise sampling distribution takes a long time to construct, and EWS also comes with large errors as there are many zeros. As the task of the paper is to do SVD decomposition, I am not sure that's why EWS is there to help. But in terms of the AMM task, it does not perform well on my side. (but I can still add EWS C++ code if required)