1: \begin{abstract}

2: In this paper, we study the generalized phase retrieval problem: to

3: recover a signal $\bm{x}\in\mathbb{C}^n$ from the measurements

4: $y_r=\lvert \langle\bm{a}_r,\bm{x}\rangle\rvert^2$,

5: $r=1,2,\ldots,m$. The problem can be reformulated as a least-squares

6: minimization problem.  Although the cost function is nonconvex, the

7: global convergence of gradient descent algorithm from a random

8: initialization is studied, when $m$ is large enough. We improve the

9: known result of the local convergence from a spectral

10: initialization. When the signal $\bm{x}$ is real-valued, we prove that

11: the cost function is local convex near the solution

12: $\{\pm\bm{x}\}$. To accelerate the gradient descent, we review and

13: apply several efficient line search methods. We also perform a

14: comparative numerical study of the line search methods and the

15: alternative projection method. Numerical simulations demonstrate the

16: superior ability of LBFGS algorithm than other algorithms.

17: \end{abstract}

18: