1: \begin{abstract}

2: We target the problem of finding a local minimum in non-convex finite-sum minimization. Towards this goal, we first prove that the trust region method with inexact gradient and Hessian estimation can achieve a convergence rate of order $\OM({1}/{k^{2/3}})$ as long as those differential estimations are sufficiently accurate.

3: 	Combining such result with a novel Hessian estimator, we propose the sample-efficient stochastic trust region (STR) algorithm which finds an $(\epsilon, \sqrt{\epsilon})$-approximate local minimum within $\OM({\sqrt{n}}/{\epsilon^{1.5}})$ stochastic Hessian oracle queries.

4: 	This improves state-of-the-art result by $\OM(n^{1/6})$.

5: 	Experiments verify theoretical conclusions and the efficiency of STR.

6: \end{abstract}

7: