I think a user warning is a good idea

NSE raises an error. Its not whether or not the computation is wrong - a generic tool like numpy can return NaNs when interacting with NaNs (i.e. when the input itself is invalid), but a score is not a generic tool - here the input argument itself is invalid for a score. (for this particular scenario, scores that need to have at least one dimension reduced, must have at least one dimension reduced - this is the case in NSE, and allowing for this computation - where this criteria (>=1 reduced dim) fails - makes the entire score meaningless and may obfuscate genuine computational errors).

NSE does raise a warning however for partially correct data - in the event that the parts that are correct can be used and the rest can be double checked.

In NSE in particular if I'm allowed to preserve all dimensions, I can't compute obs variance - it means I fundamentally don't understand how NSE works or something in my data pipeline is busted - this isn't a warning its a fundamental user input error. A warning is more acceptable if I did specify reduce dims, but I still had insufficient data to do the computation.

I think warnings should be used sparingly - e.g. for partially valid inputs like described earlier where the user may still be interested in proceeding with a computation. For data that is clearly or completely invalid, raising an error should be the default - they aren't road blocks, they just mean the result is incorrect or invalid and needs to be handled appropriately either by handling the exception itself or fixing the source of the error, because any further operation with this result without appropriate handling, is meaningless.