1: \begin{abstract}

2: %We propose a simple and scalable non-convex method for low-rank PSD matrix problems with a generic (strongly) convex objective $f$, and additional matrix norm constraints. 

3: %Such criteria appear in quantum state tomography and phase retrieval applications, among others.

4: %%In all cases, the number of variables grows fast, 

5: %However, without careful design, existing methods quickly run into time and memory bottlenecks, as problem dimensions increase.

6: %

7: %To remedy these shortcomings, we propose the \emph{Projected Factored Gradient Descent} (\palgo) algorithm, that operates on the \emph{low-rank factorization} of the variable space.

8: %Such factorization imputes non-convexity in the optimization;

9: %%, followed by a projection step on the factored space, to accommodate constraints: this results into 

10: %%Current state of the art mostly handles the \emph{unconstrained} case.

11: %%This paper builds upon and extends such recent developments:

12: %%When given a suitable initialization, 

13: %nevertheless, we show that 

14: %our method favors local linear convergence rate in the non-convex factored space, for a class of convex norm-constrained problems.

15: %We build our theory on a novel \emph{descent lemma}, that extends recent results on the unconstrained version of the problem.

16: %Our findings are supported by empirical evidence on quantum state tomography and sparse phase retrieval applications.

17: %\end{abstract}

18: