Not clear what goes wrong in case of following error: ValueError: f(a) and f(b) must have different signs

[fritshermans]

I'm running phi_k on a sample (1 mio rows) of my large dataset (>> 1 mio rows). Quite often, I get the following error:

ValueError: f(a) and f(b) must have different signs

This doesn't always happen; apparently it depends on the sample that is drawn. What could cause this error?

[RUrlus]

Ha @fritshermans, the error is coming from an optimisation failure in this call.

I suspect numerical issues (very large/small chi2 values). Could there be columns for which the sample is (nearly) perfectly correlated? It could be that with a large sample and high/low correlation we're hitting the boundary conditions

[fritshermans]

Hi @RUrlus, that answers my question. Would probably be good to raise an Exception that explains that this could be the case. Thanks for the good work! We use this package a lot.

[RUrlus]

I want to confirm my suspicion, but otherwise I agree that adding an explanation would be a good idea.