Floating point error in variance-combining

[olorin]

Current implementation is not associative, and I'm not sure if it can be made associative. However, error might be minimised by delaying combination until the end (memory permitting, need to work out if feasible) and then sorting the accumulators.

[olorin]

@adefazio suggested doing the combination in log space, which I think is a good idea.

[olorin]

Sidenote: this would be a lot easier without the need to combine partial results - if all else fails, consider doing numerics with a channel and some worker threads.

[olorin]

http://i.stanford.edu/pub/cstr/reports/cs/tr/79/773/CS-TR-79-773.pdf

[olorin]

https://www.jstor.org/stable/2286154

[olorin]

I'm now almost happy that this is as good as it's going to get. I've consulted a statistician who specialises in numerical analysis, assorted Ambiata data scientists, sundry papers and textbooks and Herbie and nothing's been better than my implementation as measured by relative error from the arbitrary-precision computation. Going to leave this open until I've derived the actual error bounds for the current version and its competitors, but I'm no longer assuming a better version exists.