1: \begin{abstract}

2: 

3: 

4: \noindent This paper deals with subsampled spectral gradient methods for minimizing finite sums.

5: Subsample function and gradient approximations are employed  in order to reduce the overall computational cost of  the classical spectral gradient methods. The global convergence is enforced by a nonmonotone line search procedure.  Global convergence is proved provided that functions and gradients are approximated with increasing accuracy.

6: R-linear convergence  and worst-case iteration complexity is investigated

7: in case of strongly convex objective function.

8: Numerical results  on well known binary classification problems are given  to show the effectiveness of

9: this framework and analyze the effect of different spectral coefficient approximations arising from the variable sample nature of this procedure.

10: 

11: \vskip 5 pt

12: \noindent{Key words}:  spectral gradient  methods, subsampling strategies, global convergence, nonmonotone line search.

13: 

14: \end{abstract}

15: