handle empty bins gracefully

[lukasheinrich]

Description

Poissons are only defined for rate parameters > 0. But it might well be that the histograms contain bins in which the total expected rate is 0 (perhaps the binning was chosen to include bins that never get filled).

Arguably those should be filtered out -- or we leave it up to the user, but then we should print a warning.

[kratsg]

Can you add an example? I'm not sure I fully understand the issue here.

[matthewfeickert]

@lukasheinrich So if I understand the idea, you're saying that as the p.mf. for the Poisson distribution is

then for bins with the p.m.f. is 0 and so they should be filtered out of the model as they don't contribute. I would think that this is the correct behavior, but can you or @kratsg think of instances in which they should not be? If so I guess I would think the default behavior should still be to filter them out but allow the user to pass some option that overrides this with a warning.

[kratsg]

For Poissons, we have to filter them out. What's not clear to me is how the code is affected. I guess this is introducing some of the "divide by zero encountered" errors that we see from time to time.

[matthewfeickert]

Closing this in favor of Issue #293.