confusing types of Complex arrays

[catawbasam]

This seems confusing:

julia> ca=[0+0im, 0+0im, 0+0im]
3-element Array{Complex{Int64},1}:
 0+0im
 0+0im
 0+0im

julia> ca2=Complex[0+0im, 0+0im, 0+0im]
3-element Array{Complex{T<:Real},1}:
 0+0im
 0+0im
 0+0im

Commit 4e68cb4* (0 days old master)

[jiahao]

This is not a bug, but the result of parametric invariance. ca2 is explicitly defined to have elements of type Complex .

If this issue is about the name Complex being confusing, I had proposed renaming Complex in #8142 , but didn't seem to garner more support.

[catawbasam]

But why should ca get typed as Array{Complex{Int64},1} ?
I didn't (intentionally) do anything to narrow the type in the assignment.

[jiahao]

Why shouldn't it? How is it different from inferring [1, 2] to be of type Array{Int} or [1, "2"] to be of type Array{Any}?

[JeffBezanson]

If you don't specify an element type, we pick for you. We prefer a narrow type by default because homogeneous arrays are extremely common and essential for performance.

[catawbasam]

Ok, then why not prefer a narrow type for ca2=Complex[0+0im, 0+0im, 0+0im] ?
Aren't those elements also homogenous?

[nalimilan]

I guess that's because Complex is equivalent to Complex{T<:Real}, so it's equivalent to asking explicitly for an array that can hold any Complex{T<:Real} value. Sometimes you may need this, if you don't plan to store only Complex{Int64} values in the array.

[JeffBezanson]

Yes, that syntax is for requesting an element type. You're right that we could have decided to have T[...] convert its arguments to type T, followed by doing what [ ... ] does. However then it would be very hard to get a heterogeneous array, and wanting Any[ ... ] is common enough.