[Numpy-discussion] numpy.random.multinomial() cannot handle zero's

Christopher Barker Chris.Barker at noaa.gov
Mon Aug 27 14:07:00 EDT 2007

Stefan van der Walt wrote:
> For some reason, the kahan_sum of [0.2,0.0,0.8,0.0] is ever so
> slightly larger than 1.0 (in the order of 1e-16), but I'm not sure
> why, yet (this isn't specific to kahan summation -- normal summation
> shows the same behavior).

Just to make sure -- is the khan_sum "compensated summation"?

Is the kahan_sum closer? -- it should be, though compensated summation 
  is really for adding LOTS of numbers, for 4, it's pointless at best. 
Anyway, binary floating point has its errors, and compensated summation 
can help, but it's still not exact for numbers that can't be exactly 
represented by binary.

i.e. if your result is within 15 decimal digits of the exact result, 
that's as good as it gets.


Christopher Barker, Ph.D.

Emergency Response Division
NOAA/NOS/OR&R            (206) 526-6959   voice
7600 Sand Point Way NE   (206) 526-6329   fax
Seattle, WA  98115       (206) 526-6317   main reception

Chris.Barker at noaa.gov

More information about the NumPy-Discussion mailing list