Justine Devin / gammalib

…for the case that the input map is not normalized (fixes #1584).

…l seem to be biased in the ellipse size.

…runcate ellipse at 3 sigma (#1561).

…ve radius that is consistent with the Gaussian normalization (#1561).

…semimajor and semiminor axis.

Before the limitation was 5 times that values and this led to convergence problems when fitting the model. Limiting the model to 2.5 times the larger axis seems to avoid the problem. Note that this is certainly not a perfect fix of the problem, but it helps to make the working model for now. Solves at least #1561.

…d. This is needed to provide correct results for cubes that are larger than the ROI.

These methods allow saving a spatial diffuse cube into a FITS file or FITS object.

…tial model component.

The eval() and eval_gradients() methods are expected to return an intensity in units of sr^-1, hence the model value needs to be divided by 4pi.

…lization in set_mc_cone() method as the fraction of the flux that is comprised within the simulation cone (#1558).

…ialDiffuseMap::set_mc_cone() method.

…s only used for Monte Carlo simulation and has a quite specific behaviour.

…for Monte Carlo simulations.

Remove the Monte Carlo cash computation from GModelSpatialDiffuseMap::prepare_map() method.

Rename GModelSpatialDiffuseMap::norm() method to  GModelSpatialDiffuseMap::mc_norm() and call GModelSpatialDiffuseMap::set_mc_cone() in that method.

set_mc_cone() called fetch_cube() which itself called load() which itself called set_mc_cone() with a larger cone again. This led to a neverending recursion.

I now introduced a load_cube() method that does the proper loading and that does not call set_mc_cone() again. This method is now used in fetch_cube().

…/s (conversion factor was missing).

The formula for drawing random numbers on a sphere was not correct. This has now been corrected, and verify using a dedicated test script (test_model_spatial_const.py).

Switching to a fixed number of iterations led to an inaccuracy of the flux determination that affected considerably the pull distributions. In short, too many events have been simulated. I thus switched back to using the integration precision instead of using a fixed number of iterations, which should be okay for flux computation (flux computation is not involved in likelihood computations).

…y intergation is done over all energy ranges in an energy boundary object (see #1450).

Updated at the same time the Doxygen information.