Review for NeurIPS paper: Learning Feature Sparse Principal Subspace

Strengths: The main contributions are novel and original, to my knowledge. The global optimality of alg.1 is very interesting and somehow surprising. Given the known hardness results of the FSPCA problem, the theorem (4.1) characterizes a subclass of problems that could be perfectly solvable. For high-rank setting, they provide a new iterative minorization-maximization procedure alg.2 by solving a low-rank covariance with alg.1. I found the MM construction here novel as existing results mostly use power method type procedure as the main algorithmic framework.