Justine Devin / gammalib

…daries. Also streamlined the interface to avoid passing parameter indices. We now pass parameter references.

The CTA response was refactored. GCTAResponse was renamed GCTAResponseIrf, and GCTAResponse became an abstract base class for all possible CTA response functions. GCTAResponseIrf derives from GCTAResponse. Also GCTAResponseCube derives from GCTAResponse.

To make this work, the GObservation::response() method now returns a pointer instead of a reference. The interface definition is such that the pointer never can be NULL. The code has been adapted to comply to this new interface.

Upon reading an XML file, GCTAObservation now determines the appropriate reponse type and calls its reader. If no response is specified in the XML file, GCTAResponseIrf will be used by default.

Event cube bins with a negative number of counts are excluded from the binned likelihood computation. This allows filtering of event bins.

A precision of 1e-8 leads to an inacceptable increase of the computing time. A precision of 1e-6 should also to the job, although results are less accurate.

An offset has been observed in the fitted normalization of file function models that has been tracked down to an insufficient precision in the Npred integrator (see #1142). Numerical tests have shown that increasing the precision from 1e-5 to 1e-8 solve the issue.

If energy dispersion is available, the analytical gradients are wrong as they are not integrated over the energy dispersion PDF. We thus switch to numerical gradients in this case, which works pretty easy.

This is somehow a kluge, but not sure that integrating the analytical gradients numerically would at the end be faster.

The coding and design conventions (CDC) document was also updated.

The curvature matrix, which is equal to one-half time the Hessian matrix, is evaluated and not the covariance matrix. To make this clean, all covar terms have been renamed to curvature.

The likelihood computation has been removed from GObservations::likelihood and has been put in the virtual GObservation::likelihood method. This method implements standard likelihood computations for binned and unbinned analysis. It may be overloaded to implement observation specific likelihood computations.

The events() set method now takes a const reference instead of a const pointer as argument. Some methods have now been defined inline.

The GFunction::eval interface has been revised and all affected GammaLib classes have been adapted.

This reverts commit 813252eb589e317fd139d0cdde4f20bdd7959ef2.

The GModels model access operator[] now returns a reference instead of a pointer. This needed a couple of minor adjustments of the code.

The GTime interface has been reworked, and some adaption was needed. One of the main changes was the replacement of the met() method by the secs() method. The secs() method returns now the time in seconds in the native GammaLib time reference.

The validity of the observation identifier is now checked so that models can apply to individual observations.

The ostream and GLog functions are not defined at the GBase level.

As the GModels container now returns a pointer and not a reference anymore, we check the pointer validity before usage.

The GModels model access operator now returns a pointer instead of a reference. The code has been adapted to this new interface.

…el_grad and GObservation::npred_grad to 0.0002.

This value is equivalent to 1 arcsec for angular parameters, and improves the precision of the position determination.