1: \begin{abstract}

2: The paper first recalls the Blahut Arimoto algorithm for computing

3: the capacity of arbitrary discrete memoryless channels, as an

4: example of an iterative algorithm working with probability density

5: estimates. Then, a geometrical interpretation of this algorithm

6: based on projections onto linear and exponential families of

7: probabilities is provided. Finally, this understanding allows also

8: to propose to write the Blahut-Arimoto algorithm, as a true proximal

9: point algorithm. it is shown  that the corresponding version has an

10: improved convergence rate, compared to the initial algorithm, as

11: well as in  comparison with other improved versions.

12: \end{abstract}

13: