If the background model has bins that are exactly 0 the significance explodes to infinity if a data event happen to land there. This in principle shouldn't happen, however, due to low statistics and insufficient smoothing/averaging, out background estimation still has 0s in some places.
As a workaround, we set a small but non-zero everywhere in the background. In the Crab spectral fit tutorial for example. While this stopped the likelihood calculation from breaking, it has been shown that the particular floor value we choose --which was arbitrary-- has a noticeable effect in the calculated significance value.
The correct way to fix this is to improve the algorithm that estimate the background by adding a step that smooths out the distribution and/or averages along certain dimensions.
Note: discarding the events where the background is 0 is another workaround, but I rather not go in that direction. If there is a region where the background is truly very small, therefore resulting in multiple 0s in the model, but the signal is actually present for real, we could be underestimating the significance by a lot.
If the background model has bins that are exactly 0 the significance explodes to infinity if a data event happen to land there. This in principle shouldn't happen, however, due to low statistics and insufficient smoothing/averaging, out background estimation still has 0s in some places.
As a workaround, we set a small but non-zero everywhere in the background. In the Crab spectral fit tutorial for example. While this stopped the likelihood calculation from breaking, it has been shown that the particular floor value we choose --which was arbitrary-- has a noticeable effect in the calculated significance value.
The correct way to fix this is to improve the algorithm that estimate the background by adding a step that smooths out the distribution and/or averages along certain dimensions.
Note: discarding the events where the background is 0 is another workaround, but I rather not go in that direction. If there is a region where the background is truly very small, therefore resulting in multiple 0s in the model, but the signal is actually present for real, we could be underestimating the significance by a lot.