1: \begin{abstract}

2: Consider the problem of minimizing a convex differentiable function on the probability simplex, spectrahedron, or set of quantum density matrices. 

3: We prove that the exponentiated gradient method with Armjo line search always converges to the optimum, if the sequence of the iterates possesses a strictly positive limit point (element-wise for the vector case, and with respect to the L\"{o}wner partial ordering for the matrix case). 

4: To the best our knowledge, this is the first convergence result for a mirror descent-type method that only requires differentiability. 

5: The proof exploits self-concordant likeness of the log-partition function, which is of independent interest. 

6: \end{abstract}

7: