1: \begin{abstract}

2:   We consider the minimization of non-convex quadratic forms regularized by 

3:   a cubic term, which exhibit multiple saddle points and poor local minima. 

4:   Nonetheless, we prove that, under mild assumptions, gradient descent 

5:   approximates the \emph{global minimum} to within $\eps$ accuracy in 

6:   $O(\eps^{-1}\log(1/\eps))$ steps for large $\eps$ and $O(\log(1/\eps))$ 

7:   steps for small $\eps$ (compared to a condition number we define), with at 

8:   most logarithmic dependence on the problem dimension. When we use 

9:   gradient descent to approximate the Nesterov-Polyak cubic-regularized 

10:   Newton step, our result implies a rate of convergence to second-order 

11:   stationary points of general smooth non-convex functions.

12: \end{abstract}

13: