Implement sensitivity function in Cython about gammapy HOT 5 CLOSED

From my experience with TS maps, I think it's a good choice to use brentq in combination with multiprocessing. Sensitivity computation for TS maps is missing as well, I'll open a separate issue for that.

from gammapy.

The sensitivity_on_off method does not exist anymore, I think it's functionality was mostly replaced by the SensitivityEstimator class. Closing this issue.

from gammapy.

@adonath - The issue / task really is still the there, only sensitivity_on_off was renamed excess_matching_significance in the meantime.

Currently we use numpy.vectorize at

It's not a trivial task because currently the implementation is based on scipy.optimize.fsolve - I think that is callable from Cython, but only from the latest Scipy version, i.e. not available for us yet.

@adonath - Up to you if you want to re-open or keep this issue closed. It's certainly not urgent - but some day (can be next year) - this should be implemented in a fast way, and probably code duplication with the TS map code removed (by moving the stats part of that code to gammapy.stats and sharing as much as possible for these 1-dim amplitude fit tasks.

from gammapy.

Thanks @cdeil, you are right, I didn't notice the re-name of the function.

I just checked what the SensitivityEstimator does and I doubt a highly optimised version of this code is needed. Computing the minimal excess needed for a given significance for O(~30) energy bins should always be less than one second. The TS map computation does this for O(~1e4-1e6) pixels, which is a completely different magnitude. In the case of the SensitivityEstimator even a "brute-force" grid minimisation (30 energy bins and ~100 bins in flux / excess) with Numpy should be efficient and precise enough.

For me it's rather the question, whether the code should be shared e.g. with the FluxPointsEstimator or be re-written based on SED likelihood profiles, such that we can compute sensitivity curves for 1D and 3D datasets as well. I can make a reminder issue for this.

from gammapy.

Yes, a faster version would only be needed for sensitivity maps with ~ 1e6 pixels, not sensitivity spectra with @ 10 pixels. I'm not sure how to best avoid code (and test effort) duplication among the various use cases for 1D optimisation within Gammapy. Not high priority, thus the label here was "wishlist".

from gammapy.