Element type of `Diagonal{T}` is `T` even when `typeof(zero(T)) != T`

[tkoolen]

From https://discourse.julialang.org/t/help-creating-a-jump-variable-array/11515/8?u=tkoolen. There are cases where typeof(zero(T)) != T, so what should the eltype of Diagonal{T} for such T?

One option would be to define

Base.eltype(::Type{Diagonal{T}}) where {T} = promote_type(T, typeof(zero(T)))

but this could lead to bugs where the T in Diagonal{T} is assumed to be the element type. Another option is to make it so that the Diagonal constructor converts the input vector to e.g. Vector{promote_type(T, typeof(zero(T)))} when necessary.

Maybe erroring in the constructor when typeof(zero(T)) != T could be another option.

[StefanKarpinski]

Having a type where zero(T) is not of type T seems highly pathological.

[mbauman]

but this could lead to bugs where the T in Diagonal{T} is assumed to be the element type.

Even worse is that Diagonal reports that T to AbstractArray{T} — leading to a strange inconsistency between eltype and functions that grab T from the AbstractArray directly. I think we have to assert that zero(T) is T.

[mbauman]

Strongly related: https://github.com/JuliaLang/julia/issues/27015

[vtjnash]

leading to a strange inconsistency between eltype

Well, that's just wrong – but fixable :). If it needs a ComputedFieldType(:tm:), that is doable these days now. You just need the right incantation!. Here's an example:

julia/base/iterators.jl

Line 169 in 917ae8b

Pairs(data::A, itr::I) where {A, I} = new{eltype(I), eltype(A), I, A}(data, itr)

[tkoolen]

Having a type where zero(T) is not of type T seems highly pathological.

I think what JuMP is doing is very reasonable (see Discourse link in description). JuMP.Variable just can't represent zero.

Also related: JuliaLang/LinearAlgebra.jl#497, and actually also JuliaGeometry/Rotations.jl#55.

I think @vtjnash's approach would work well on master, now that the 'storage' type parameter is separated from the 'element' type parameter (unlike how it was in 0.6); something like

function Diagonal(diag::V) where {T, V<:AbstractVector{T}}
    S = promote_type(T, typeof(zero(T)))
    new{S, V}(diag)
end

[andyferris]

Sometilmes I feel that eltype could be more of a trait for AbstractArray (or iterate, I suppose?). It makes sense as a type parameter of Array but I've noticed that carrying it around is inconvenient in some of the concrete arrays I've created.

On the other hand. removing T from AbstractArray and forcing everyone to use trait dispatch patterns just to capture the eltype might be too inconvenient at this stage?

[brenhinkeller]

Is it still true that there are Base types for which typeof(zero(T)) != T?

It looks like some formerly inconsistent matrix zeros (e.g. JuliaLang/LinearAlgebra.jl#170) are much more consistent nowadays

[martinholters]

Is it still true that there are Base types for which typeof(zero(T)) != T?

Yes, e.g. for T = Irrational{:π}. (Although Diagonal{Irrational{:π}} may not be of much practical importance, of course.)