1: \begin{abstract}

2: We propose   an alternating subgradient method with non-constant step sizes for

3: solving convex-concave saddle-point problems 

4: associated with general convex-concave functions.

5: We assume that the sequence of our step sizes is not summable but square summable. 

6: Then under the popular assumption of uniformly bounded subgradients, we prove

7: that a sequence of convex combinations of function values over our iterates

8: converges to the value of the function at a saddle-point. 

9: Additionally, 

10: based on our result regarding the boundedness of the sequence of our iterates,

11: we show that a sequence of the function evaluated at convex combinations of 

12: our iterates  also converges to the value of the function over a saddle-point.

13: 

14: We implement   our algorithms in  examples of 

15: a linear program in inequality form, 

16: a least-squares problem with $\ell_{1}$ regularization,

17: a matrix game, and a robust Markowitz portfolio construction problem.

18: To accelerate convergence, we reorder  the sequence of step sizes

19: in descending order, which turned out to work very-well  in our examples.

20: Our convergence results are confirmed by our numerical experiments.

21: Moreover, we also numerically compare our iterate scheme with iterates schemes

22: associated with constant step sizes. 

23: Our numerical results support  our choice of step sizes. 

24: Additionally,  we observe the convergence of the sequence of function values

25: over our iterates in multiple experiments, 

26: which currently lacks theoretical support.

27:  

28: \end{abstract}

29: