Empty Array with nonzero size along some dimension, should they be disallowed?

Internals & Design
1

On both LTS and nightly we have:

julia> a = Array{Int}(undef, 0, 1)
0×1 Matrix{Int64}

julia> size(a)
(0, 1)

julia> length(a)
0

Is this a bug? I’d actually have expected the construction attempt to throw. Couldn’t find any issues on Github, but I can’t think of a rationale for this behavior.

Alternatively, perhaps it’d be good to normalize the shape, so the result of Array{Int}(undef, 0, 1) would be the same as for Array{Int}(undef, 0, 0)?

Thoughts?

2

Why would it throw or change the shape from what the user requested? A 0x1 matrix is perfectly valid.

For a practical example of where the difference between 0x0 and 0x1 matters, hcat(ones(0,1), ones(1,1)) works, but hcat(ones(0,0), ones(1,1)) will throw a `DimmensionMismatch.

7 Likes
3

Of possible interest: the plotting example “making waves” uses this with:

...
zs = zeros(0, 40)
...
@gif for i in range(0, stop = 2π, length = n)
    ...
    # add to and show the tracked values over time
    global zs = vcat(zs, z')
    ...
end

It’s convenient to initialise the time series data as a 0x40 matrix.

1 Like