1: \begin{abstract}

2: This paper considers the problem of solving systems of quadratic equations,

3: namely, recovering an object of interest $\bm{x}^{\natural}\in\mathbb{R}^{n}$

4: from $m$ quadratic equations\,/\,samples $y_{i}=(\bm{a}_{i}^{\top}\bm{x}^{\natural})^{2}$,

5: $1\leq i\leq m$. This problem, also dubbed as phase retrieval, spans

6: multiple domains including physical sciences and machine learning.

7: 

8: We investigate the efficacy of gradient descent (or Wirtinger flow)

9: designed for the nonconvex least squares problem. We prove that

10: under Gaussian designs, gradient descent --- when randomly initialized

11: --- yields an $\epsilon$-accurate solution in $O\big(\log n+\log(1/\epsilon)\big)$

12: iterations given nearly minimal samples, thus achieving near-optimal

13: computational and sample complexities at once. This provides the first

14: global convergence guarantee concerning vanilla gradient descent for

15: phase retrieval, without the need of (i) carefully-designed initialization,

16: (ii) sample splitting, or (iii) sophisticated saddle-point escaping

17: schemes. All of these are achieved by exploiting the statistical models

18: in analyzing optimization algorithms, via a leave-one-out approach

19: that enables the decoupling of certain statistical dependency between

20: the gradient descent iterates and the data.

21: 

22: \end{abstract}

23: