Open maxnoe opened 2 years ago
@maxnoe Thank you very much for this analysis.
We had not realized that the behavior of the Hillas parameters (length and width) changes so much because we had used a much simpler toy model where the effect of cleaning and noise was not appreciated.
We will continue to analyze this to try to adjust the calculation of the uncertainties and avoid such a large underestimation.
Describe the bug
When using cleaning or noise is present (aka always), it seems the uncertainties for the hillas parameters are dramatically under-estimated.
The formulas seem only to be valid for the case where all signal is used for the estimation and no noise is present.
To Reproduce
Notebook with toy simulations: https://github.com/maxnoe/notebooks/blob/main/ctapipe/hillas_uncertainties.ipynb
Expected behavior
The mean uncertainty of the hillas parameters should match the std over random samples from the same population.