1: \begin{abstract}

2: We consider the problem of minimizing a sum of $n$ functions 

3: via projected iterations onto a convex parameter set $\C \subset \reals^p$ where $n\gg p\gg 1$. 

4: In this regime, algorithms which utilize sub-sampling techniques are known to be effective.

5: In this paper, we use sub-sampling techniques together with 

6: eigenvalue thresholding to design a new randomized batch algorithm 

7: which possesses comparable convergence rate to Newton's method, 

8: yet has much smaller per-iteration cost. 

9: The proposed algorithm is robust in terms of starting point and 

10: step size, and enjoys a composite convergence rate, namely, 

11: quadratic convergence at start and linear convergence 

12: when the iterate is close to the minimizer. 

13: We develop its theoretical analysis which also allows us to 

14: select near-optimal algorithm parameters. 

15: Our theoretical results can be used to obtain 

16: convergence rates of previously proposed sub-sampling based algorithms as well. 

17: We demonstrate how our results apply to well-known machine learning problems.

18: Lastly, we evaluate the performance of our algorithm on several datasets 

19: under various scenarios. 

20: \end{abstract}

21: