Any tests?

What about _sqrtprod(1e-200, 1e-200)? Can an underflow be avoided?

Any tests?

In the previous PR, there was a script to demonstrate that the results are generally improved. Otherwise, I'm relying on the existing tests for correlation() which does not make any particular promises about accuracy.

What about _sqrtprod(1e-200, 1e-200)? Can an underflow be avoided?

This could be a future improvement. Off-hand I don't see an elegant way to do this without a bunch of conditional logic and rescaling, but I'll put some thought into it. For now, it is on the bottom of my priority list.

I see. Thank you for details.

Could you please include issue number in future PRs title. I sometimes dig through the history of the code and find changes that do not have references to issue. It's very hard to understand the context without it.