Unexpected behaviour of cat for empty matrix entry

[andreasvarga]

The following is unexpected to me:

julia> cat(Float64[], rand(2,2), dims=Val((1,2)))
2×3 Matrix{Float64}:
 0.0  0.82864   0.222715
 0.0  0.230898  0.594456

which is the same as

julia> cat(zeros(0,1), rand(2,2), dims=Val((1,2)))
2×3 Matrix{Float64}:
 0.0  0.0953531  0.692899
 0.0  0.410014   0.0780325

which is correct. Also

julia> cat(zeros(1,0), rand(2,2), dims=Val((1,2)))
3×2 Matrix{Float64}:
 0.0       0.0
 0.60077   0.539492
 0.891619  0.0582591

is as expected.

To obtain the expected result I needed to use

julia> cat(zeros(0,0), rand(2,2), dims=Val((1,2)))
2×2 Matrix{Float64}:
 0.284389  0.356856
 0.229907  0.651545

Is this behaviour the right one, or is a bug?

[mcabbott]

I think this is expected, because the empty vector is effectively 0x1 (or if more dimensions matter, it is 0x1x1x1 etc.)

julia> size([], 1)
0

julia> size([], 2)
1

Concatenating an empty vector with a non-empty one arguably relies on it not being 0x0:

julia> vcat(Float64[], ones(2))
2-element Vector{Float64}:
 1.0
 1.0

julia> vcat(Float64[], ones(2,1))
2×1 Matrix{Float64}:
 1.0
 1.0

julia> vcat(zeros(0,0), ones(2))
ERROR: DimensionMismatch: number of columns of each array must match (got (0, 1))

[andreasvarga]

Frankly: Why should have an empty array column dimension 1? In Matlab size([ ]) = (0,0).
I realize that in Julia this is a very basic issue, but

size([],2) = 0

would still be the right choice. All concatenations could accept an empty array, without any effect (similarly as in Matlab).

My final comment: Knowing the current behaviour, it is easy to programmatically overcome possible ambiguous situations.
So, this issue can be closed from my point of view.