mean(::UnitRange) overflows

[jariji]

using Statistics
julia> mean(1:typemax(Int64))
-4.611686018427388e18

  mean(itr)

  Compute the mean of all elements in a collection.

From the documentation one might expect a number between the two endpoints, but I got a number outside that range. I think this is a bug.

[oscardssmith]

Good catch. Right now this is implemented as (first(r) + last(r)) / 2. It should probably be first(r)/2 + last(r)/2.

[petvana]

Seems to be already reported in JuliaStats/Statistics.jl#120 and there is even an unmerged PR for that JuliaStats/Statistics.jl#115.

[chunjiw]

Good catch. Right now this is implemented as (first(r) + last(r)) / 2. It should probably be first(r)/2 + last(r)/2.

Or perhaps first(r) + (last(r) - first(r)) / 2

[jariji]

I like @chunjiw 's proposal because it works better on affine spaces like dates or Fahrenheits where you can't divide positions by numbers (you can only divide displacements).

[chunjiw]

I'm wondering is the performance gain of "one less float point division" significant enough? When I started programming I was taught x + (y-x) / 2 is the best way to calculate mean.

[oscardssmith]

llvm should be able to get rid of the floating point division since it's by a constant power of 2