additional restrictions on summary_func?

It should probably be a guaranteed property of general statistics that if you compute a branch statistic before and after simplifying then you’ll get the same answer. Unsimplified tree sequences can have dangling branches, which should have weight 0, so unpolarised statistics will assign value

  f(0) + f(total_weight)

to these branches. So, to ensure that this property holds, we need to require that

 f(0) = 0
 f(total_weight) = 0

This is the case for all statistics we’ve defined or discussed so far (e.g., for phenotypic covariance, we deifned the weights to sum to zero).

[jeromekelleher]

I think this still needs doing @petrelharp?

[jeromekelleher]

I agree - the general stats interface isn’t high-performance anyway, so checking the function here makes total sense. (We should keep in mind people that are not us also, who are developing stats that don’t end up being implemented in tskit)