ctools

• implemented new function save_covmat(const GFilename& filename) to GObservations class, which stores the covariance matrix to a given FITS file.
• new hidden parameter covmat for ctlike, per default set to “NONE”. If valid filename is passed to this parameter, the covariance matrix is written out.

A problem is where to define the output - at the moment it is done directly after calling the optimization, because doing it in ctlike::save() is to late, as the models likelihood function is changing when ts values are required to compute. The models are set back to the original ones afterwards, but this cannot be done for the likelihood function as there is no direct access to it at that point (this caused a bit of confusion to me...).

• Update documentation to include new feature.

Ziegler Alexander wrote:

A problem is where to define the output - at the moment it is done directly after calling the optimization, because doing it in ctlike::save() is to late, as the models likelihood function is changing when ts values are required to compute. The models are set back to the original ones afterwards, but this cannot be done for the likelihood function as there is no direct access to it at that point (this caused a bit of confusion to me...).

I think it’s fine to store the covariance matrix directly after the optimization, and then let the code do other stuff for the TS computation.

I added the new covmat feature to the docu of ctlike.
I tested it with different models by hand, reading the output covariance matrix fits file again in an calculating the errors on the parameters from that output and comparing the results to the errors given in the output model xml file.
Seems like the arrangement order in the FITS file is correct and everything works fine - ready for code review.

• 1863_CovMat (both in gammalib and ctools repo)

I checked the code you committed.

I recognised that you did the saving of the covariance matrix in the ctlike::run() method while all saving should be done in the ctlike::save() method. I understand that you did this because the covariance matrix may eventually change after the initial fit due to the TS computation. There is however a way to accommodate this: saving the curvature matrix after the fit and restoring the curvature matrix at the end of the run method. I modified the code accordingly.

Logically, the covariance matrix should be a member of GObservations::likelihood and not GObservations. There is in fact a way to do the saving in the GObservations::likelihood since the class holds a pointer to GObservations, hence it is able to access the models. I therefore move the saving to GObservations::likelihood and called the method simply GObservations::likelihood::save(). In principle, we could add code to also save the model values, gradients, etc. into a FITS file.

I merged all changes into devel. I put the status to Feedback and you may check that the code still behaves as you expect.

Alexander an me came to the conclusion that it would be better to store the covariance matrix a csv file rather than as a fits file to enable easy access to the user. We are also not sure if this output should be used in subsequent scripts because the values of the covariance matrix can easily change their order if the according model file is changed by the user.
I am starting to implement the export to a csv file now.

The output format of GObservations::likelihood::save() can now be chosen with a flag to either csv or fits. Default is csv.
The user of ctlike can not choose the format. We will update the documentation accordingly if it’s OK to make the default output csv. (The fits output still works as expected but is not favored.)

An example of a csv file is given here:

RA(Crab) DEC(Crab) Prefactor(Crab) Index(Crab) Cutoff(Crab) Scale(Crab) Normalization(Crab)
5.61309599084731e-06 -1.83433791867944e-07 4.62934176274374e-24 -1.52450978816292e-06 -9.26084864350935 0 0
-1.83433791867944e-07 4.58675292270187e-06 7.65058926211812e-23 2.98909451728534e-06 -15.9355531595409 0 0
4.62934176274372e-24 7.65058926211812e-23 3.53273692750675e-36 1.92196287270621e-19 -4.25589848340144e-13 0 0
-1.52450978816292e-06 2.98909451728534e-06 1.92196287270621e-19 0.0111048626510676 -22395.023023584 0 0
-9.26084864350935 -15.9355531595409 -4.25589848340144e-13 -22395.023023584 60131839781.0409 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0

The documentation is now also updated.

The corresponding branches are:
dtiziani/gammalib:1863-covmat
dtiziani/ctools:1863-covmat

I adapt the ctlike documentation to the logic mentioned above, i.e. that the file format is defined by the file type of the file name.

The change was merged into devel.