Ideas for analysis improvements with low statistics

[mireianievas]

I'm currently in the Fermi school and there have been some ideas floating to improve the the sensitivity of the analysis without making the analysis too slow.

I'm currently thinking on the following possibilities:

A) Component analysis (as done by Fermipy). This would allow to analyze the different events according to their PSF tag and potentially improve the sensitivity. I see in the code some attempt of implementing a simpler version of this, analyzing back/front events separately. What is the current status of that?.

B) Switch to unbinned for higher energies (low statistics regime). One can think on something simple, such as running the low energy bins (where we have good statistics) with binned analysis and then switch to unbinned from a certain threshold Energy or in those bins where the number of predicted events is low. The analysis would be a slower, but that extra sensitivity at high energies might be worth it.

C) Switch to unbinned for LCs with low number of predicted events per bin.

PS. I have other ideas floating in my mind too, like implementing a module to automatically download (using astroquery) the data for the requested times/energies/roi or making sure that products are 'recycled' if they exist in the disk (maybe a previous execution crashed but we have some of the products already generated).

What do you think?.

[davidsanchez]

Hi

Thanks for all the nice ideas:

A) The code can use what we, in Fermi, call the "summed likelihood" and should be working (at least for P7 Irfs. Should be possible to do the same for the PSF and other

B) that something I had in mind since a while but did not find the time to implement it

C) sound like a good idea.

If you have time to make the changes, that would be great! these are very good ideas

cheers

[mireianievas]

Dear all,

I made a first implementation of the PSF/EDISP component analysis and uploaded it to my github. You can check it here: https://github.com/mireianievas/enrico .

I needed to modify a little bit more the code than I expected, because the iso file was not correctly predicted (there was a small bug there) for the Summed Likelihood and the I needed to add much more stuff to the list of auxiliary/IRF files to download.

For now, I preferred to keep the default config file untouched until the procedure is properly tested. To activate it, just add the section [AnalysisComponents] to your config file, and then either write the PSF = true variable or EDISP = variable.

I still need to think about B). What sounds better to you?:

1. User defines a threshold energy, for E<Eth binned, for E>Eth unbinned ?
2. Binned analysis, then predict the number of counts as a function of energy, then split the dataset (that would be optional of course) and rerun a Summed Likelihood.

For the LC (C) it would be probably 2. anyways.

[davidsanchez]

Hi,

thanks for the work. I would go for option 1

How do you plan to predict the number of counts?

cheers