I'm running MKL 11 Update 3 and am getting a floating point divide by zero exception at _mkl_trs_dmintr_ls_lc() + 0x10a1 from a call to strnlspbc_solve. I'm using LS to fit a Gaussian, and the code works for many cases and I've validated it when data generated from a Gaussian. When it fails, it occurs after a number of loops through the RCI loop.
There must be something about the specific problem to which you are applying this solver, or something specific about how you call the solver in your code, which causes the failure of the solver.
There is an example code provided with MKL, called ex_nlsqp_bc_f.f, and that example runs without errors.
Therefore, not much can be done or said until you provide example code that displays the same problems that you reported.
Thanks for your feedback.
I think the function name and "divide by zero" error provides a tremendous amount of information, likely enough to find the bug. I'm tempted to say a robust solver (which is the whole idea behind TR methods) should be guarding against divide by zero error.
I'll do some more investigation and then see if I can pull together an example.