The former implementation tests if the norm of the correction of the trust region sub problem is smaller than innerTolerance_ which is usually a small number. If the correction dX has a large magnitude, e.g. ||dX|| = 2. The for loop with innerIterations_ runs till the end because ||dX|| will always stay larger than innerTolerance_. And the full set of innerIterations_ is performed even if corr_global is already converged within machine precision. Therefore, to terminate the loop at a better time comparing innerTolerance_ to diff which measures the change of corr_global within the current iteration.
I discussed this with @jaap and @lnebel . Maybe you two can check if i missed something?
What do you think about this? @osander