eltype of static matrix factorization is Any

[gdalle]

julia> using StaticArrays, LinearAlgebra

julia> A = rand(2, 2);

julia> eltype(lu(A))
Float64

julia> S = SMatrix{2,2}(A);

julia> eltype(lu(S))
Any

[fredrikekre]

What do you need it for? Note that neither LinearAlgebra.LU nor StaticArrays.LU are abstract arrays that you can index etc. They both iterate L, U, and p (matrix, matrix, vector) so considering this the Any eltype make more sense.

[gdalle]

What do you need it for?

The following chunk of code will land in an upcoming version of ImplicitDifferentiation.jl. Basically I wanna solve a linear system with a possibly failed LU decomposition, and replace the results with NaNs if needed:

function safe_solve(A_lu, b)
    T = float(promote_type(eltype(A_lu), eltype(b)))
    x_maybenan = Vector{T}(undef, size(A_lu, 2))
    if issuccess(A_lu)
        x_maybenan .= A \ b
    else
        x_maybenan .= convert(T, NaN)
    end
    return x_maybenan
end

But this errored twice: because of float(eltype(A_lu)) and because of size(A_lu).

Note that neither LinearAlgebra.LU nor StaticArrays.LU are abstract arrays

Basically that's what confused me. Since matrix operations like * and \ are defined, I thought every matrix operation would be. But I understand why that is not the case.

For my scenario, I used the fields of the LU decomposition instead and it worked.