Harlan Grove wrote:

Martin Brown wrote...
...

Excel appears to flake out whenever the product of the numbers in
GEOMEAN overflows a basic IEEE floating point representation. 1.7E+308

Odd really since in other algorithms they hold intermediate results in
temporary reals which would give the naieve algorithm a lot more
headroom 1.1E+4932

...

Odder still since the geometric mean is just e raised to power of the
average of the logarithms or the product of the Nth roots of N values.
80-bit reals aren't necessary for better results. Better algorithms
would have provided ample robustness.

Yes. What I was commenting on was that they have implemented a very bad
algorithm in an absolutely disastrous fashion. Two mistakes compounded!

Another one to add to the Mickeysoft hall of statistical infamy.

Regards,
Martin Brown