Use zero() for type-stable calculation of _momentN()
#969
Conversation
src/moments.jl
Outdated
| function _moment2(v::AbstractArray{<:Real}, m::Real; corrected=false) | ||
| n = length(v) | ||
| s = 0.0 | ||
| s = zero(m * m) |
There was a problem hiding this comment.
In general, s does not have the correct/desired type here as this initialization purely depends on m and neglects v.
There was a problem hiding this comment.
I was making the assumption that m and v would have compatible/identical types, since m is calculated from v and the function is only called internally. Would an implementation involving v be something like zero(v[begin] * v[begin])?
There was a problem hiding this comment.
IMO we shouldn't make that assumption. For instance, in var(...; mean=...) users can easily provide a mean that is of different type.
Would an implementation involving v be something like zero(v[begin] * v[begin])
No, that would error if the array is empty and would neglect the type of m.
There was a problem hiding this comment.
Would the most recent commit (6a236d4) work for this?
There was a problem hiding this comment.
I'll have to say, I'm confused where it would fail 😅
There was a problem hiding this comment.
For instance if v::AbstractVector{Bool} and m::Bool:
julia> v = [true, true];
julia> m = true;
julia> zero(promote_type(eltype(v), typeof(m)))
false
julia> (v[1] - m)^2
0I also noticed that one actually has to make a few more changes to ensure that e.g. Float32 inputs don't yield Float64 outputs. I opened #973 to address this issue more generally.
|
Made obsolete by #973 :D |
2 participants
Solves #968