1: \begin{abstract}

2: A new decomposition optimization algorithm, called \textit{path-following gradient-based decomposition}, is proposed to solve separable convex  optimization

3: problems. Unlike path-following Newton methods considered in the literature, this algorithm does not requires any smoothness assumption on the objective

4: function. This allows us to handle more general classes of problems arising in many real applications than in the path-following Newton methods. The new

5: algorithm is a combination of three techniques, namely smoothing, Lagrangian decomposition and path-following gradient framework. 

6: The algorithm  decomposes the original problem into smaller subproblems by using dual decomposition and smoothing via self-concordant barriers, updates

7: the dual variables using a path-following gradient method and allows one to solve the subproblem in parallel.

8: Moreover, the algorithmic parameters are updated automatically without any tuning strategy as in augmented Lagrangian approaches. 

9: We prove the global convergence of the new algorithm and analyze its local convergence rate. 

10: Then, we modify the proposed algorithm by applying Nesterov's accelerating scheme to get a new variant which has a better local convergence rate.

11: Finally, we present preliminary numerical tests that confirm the theory development.

12: 

13: \keywords{Path-following gradient method \and dual fast gradient algorithm \and separable convex optimization \and smoothing technique \and self-concordant

14: barrier \and parallel implementation.}

15: % \PACS{PACS code1 \and PACS code2 \and more}

16: % \subclass{MSC code1 \and MSC code2 \and more}

17: \end{abstract}