`complex` errors on nested arrays

[johroj]

complex(::AbstractArray{T}) fails when zero(::Type{T}) is not defined. So it fails on for example nested arrays, i.e. complex([[1]]) throws. See

julia/base/complex.jl

Lines 1134 to 1139 in cdd4ac5

	function complex(A::AbstractArray{T}) where T
	if !isconcretetype(T)
	error("`complex` not defined on abstractly-typed arrays; please convert to a more specific type")
	end
	convert(AbstractArray{typeof(complex(zero(T)))}, A)
	end

It seems that real(::AbstractArray{T}), imag(::AbstractArray{T}) and conj(::AbstractArray{T}) uses broadcast_preserving_zero_d instead and this works as expected. See for example

julia/base/abstractarraymath.jl

Line 210 in cdd4ac5

conj(A::AbstractArray) = broadcast_preserving_zero_d(conj, A)

Is there any good why complex uses a different implementation?

[nsajko]

A workaround, in case it's not clear, is to use broadcasting yourself. For example:

julia> m = rand(ComplexF64, 2, 2)
2×2 Matrix{ComplexF64}:
   0.3918+0.301684im  0.138374+0.412443im
 0.663285+0.288682im  0.748438+0.903168im

julia> n = let n = Matrix{Matrix{ComplexF64}}(undef, 2, 2)
           for i ∈ eachindex(n)
               n[i] = m
           end
           n
       end
2×2 Matrix{Matrix{ComplexF64}}:
 [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]  [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]
 [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]  [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]

julia> complex(n)
ERROR: MethodError: no method matching zero(::Type{Matrix{ComplexF64}})
The function `zero` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  zero(::Type{Union{}}, Any...)
   @ Base number.jl:359
  zero(::Type{Missing})
   @ Base missing.jl:106
  zero(::Missing)
   @ Base missing.jl:103
  ...

Stacktrace:
 [1] complex(A::Matrix{Matrix{ComplexF64}})
   @ Base ./complex.jl:1138
 [2] top-level scope
   @ REPL[3]:1

julia> complex.(n)
2×2 Matrix{Matrix{ComplexF64}}:
 [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]  [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]
 [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]  [0.3918+0.301684im 0.138374+0.412443im; 0.663285+0.288682im 0.748438+0.903168im]

[johroj]

Yes of course. My problem was slightly more annoying since this call occurs in a library.

The question remains if there is any technical reason for implementing for using broadcast_preserving_zero_d for real , conj, and imag, but not for complex? Otherwise, the long-term fix to this problem is straightforward.