Store full set of minima

Hello, I wrote an extension of the current minimization part of Cobaya to store the full set of minima from multiple initial points. I think this is valuable for the user, since, for example, the dispersion of the minima from different initial conditions represents a feedback on the precision of the minimization procedure. An example is this (Fig.4), where this very concept is applied. A similar concept is also shown in the Planck's paper here (Fig.1), where they show the effect of changing CAMB's accuracy parameters.

In #332 I already mentioned this feature, which then I decided to leave for later development since it was causing some issues. Those are fixed now and Travis passed all tests on my fork.

Summarizing the changes in minimize.py:

Add an attribute where to store the minima (the aforementioned bug was caused by not passing correctly this through MPI) and default it to None called full_set_of_mins;
If multiple initial points are considered, store the minima in a dictionary giving both the value at the minimum and a boolean indicating whether the minimization process was successful or not (this gets also logged);
Pass this quantity to the products function;
Add a small description of full_set_of_mins in docs

For what regards tests, since the default value of best_of is 2, I added to the existing test an assert statement checking that full_set_of_mins is a dictionary. Then, I added a new test checking that if a single initial point is requested full_set_of_mins = None. Tho, I had to suppress this test for the MPI-part of tests. I don't know if this is an intended behavior, but in the (weird) case in which best_of < MPI size the code will act as if best_of == MPI size. This results in having 2 initial points even when I explicitly set best_of = 1 because one test is run with MPI size = 2. It doesn't make much sense to me asking intentionally less initial points than processes, so probably this is not something to worry about. Still, it could be a good idea to add a warning in minimize.py to make sure that the user is aware of what is going on.

As stated above, Travis passed all checks on my fork, so this time I will not bother you with the debugging. Of course, let me know if the current implementation is satisfactory or if you would make any modifications.

CobayaSampler / cobaya

Store full set of minima #334

Codecov Report