Corner cases of hvcat [...] syntax: column matrices

[toivoh]

hvcat concatenation syntax produces matrices, such as in

julia> [1 2; 3 4]
2×2 Array{Int64,2}:
 1  2
 3  4

julia> [1 2]
1×2 Array{Int64,2}:
 1  2

But there is no corresponding way to create a single column matrix:

julia> [1; 3]
2-element Array{Int64,1}:
 1
 3

julia> [1;]
1-element Array{Int64,1}:
 1

I suggest that the latter should invoke hvcat and not vcat, to produce matrices. This implies that there would be no special vcat syntax any more.

[thofma]

+1 since I tripped over this recently. It's super counterintuitive that [1 2; 3 4] and [1; 3] yield different types.

[martinholters]

IMO having easy syntax for both creating a vector and a one-column matrix would be handy. We could consider

[1; 2] # gives a one-column matrix
[1; 2;] # gives a vector

BUT

• the difference is very subtle
• the definition could equally well be the other way round, so it's probably not very intuitive
• would be breaking, of course.

JuliaLang/julia#16740 could make Vector[1; 2] and Matrix[1; 2] work (and do the expected) which would be explicit but might be too verbose. And of course, [1; 2] would still not meet everyone's expectations.

[StefanKarpinski]

Now that we have RowVectors, we could actually make [1 2 3] create a RowVector instead of a row matrix. That would make the space and semicolon syntaxes more symmetrical: you would only get a matrix when spaces and semicolons are used together – semicolons only ➝ (column) vector; spaces only ➝ row vector. There's something appealing about it, but it might be annoying.

[andreasnoack]

You can create vectors with ,, e.g. [1,2,3], so you could argue that [1;2;3] is redundant syntax for creating a vector. One option could be to make hcat/vcat 2D operations and always return a matrix, e.g. [1;2]. The main problem is to come up with a way to append vectors since [ones(2);ones(2)] would now be promoted to 2D. One obvious candidate for appending vectors is append, i.e. the non-mutating version of append!, but append(x,y) is not as short as [x;y].

[martinholters]

My current thinking is that something like this would be nice:

• [a b c]: RowVector if a, b, c are all RowVectors or scalars, Matrix if at least one is Vector or Matrix
• [a b c;]: always a Matrix, hvcat with only one row
• [a; b; c]: Vector if a, b, c are all Vectors or scalars, Matrix if at least one is RowVector or Matrix
• some other nice syntax variation: always a Matrix, hvcat with only one column

It's a bit too Friday afternoon for a good idea for the fourth case...

[StefanKarpinski]

Opening with [; could be used somehow (currently always a syntax error). One scheme that occurs to me is this:

• vertical indicated by internal or trailing ;
• horizontal indicated by internal space or leading ;

This would give the following syntaxes:

• vcat: [a;], [a; b] == [a; b;]
• hcat: [;a], [a b] == [;a b]
• hvcat: [;], [; a;], [a b;] == [; a b], [; a; b]

This creates a syntax for a zero-element hvcat since the semicolon is both leading and trailing, but leaves no zero-element syntaxes for hcat or vcat. I'm not sure how I feel about this. The rule is fairly simple, but it feels kind of visually messy. Not sure it would be an improvement.

[omus]

I think using the semicolon syntax for creating a one-column matrix is the right way to go. Appending vectors would no longer be able to use semicolon for this purpose but you could do [x..., y...] which already works.

I wrote a mini test suite just to get a feel for how things currently work and potentially could work:

@test size([1 2 3])   == (1, 3)
@test size([1; 2; 3]) == (3, 1)  # Currently fails
@test size([1, 2, 3]) == (3,)

@test size([1])  == (1,)
@test size([1;]) == (1, 1)  # Currently fails

v = [1, 2]
@test [v  v] == [1 1; 2 2]
@test [v; v] == [1 2 1 2]'         # [1; 2; 1; 2], Currently fails
@test [v, v] == [[1, 2], [1, 2]]
@test [v...  v...] == [1 2 1 2]
@test [v...; v...] == [1 2 1 2]'   # [1; 2; 1; 2]
@test [v..., v...] == [1, 2, 1, 2]

m = [1 2]
@test [m  m] == [1 2 1 2]
@test [m; m] == [1 2; 1 2]
@test [m, m] == [[1 2], [1 2]]

m = [1 2]'  # [1; 2]
@test [m  m] == [1 1; 2 2]
@test [m; m] == [1 2 1 2]'       # [1; 2; 1; 2]
@test [m, m] == [[1 2]', [1 2]'] # [[1; 2], [1; 2]]

m = [1 2; 3 4]
@test [m  m] == [1 2 1 2; 3 4 3 4]
@test [m; m] == [1 2; 3 4; 1 2; 3 4]
@test [m, m] == [[1 2; 3 4], [1 2; 3 4]]

[ngharrison]

A semicolon creating a 2-D array in all cases except for when combining vectors or scalars has been the only un-intuitive part of Julia's array syntax for me. The syntax suggested by @omus makes a lot of sense. Spaces are always used to concatenate horizontally (2-D). Semicolons are always used to concatenate vertically (2-D). Commas are always used to concatenate one-dimensionally. If a splat is used, then it unravels the iterator before doing any of the above concatenations. This is shown well in his examples. It gives a visually-similar yet distinct way to create and display the three cases.

An example that makes the current syntax even more confusing:

a = [1 2] # 1×2 Array
b = hvcat(1, [1, 2]) # 2×1 Array
c = [1, 2] # 2-element Array

[a, b, c] # forces it to print all three, gives:
3-element Array{Array{Int64,N} where N,1}:
 [1 2]
 [1; 2]
 [1, 2]

First, the syntax I used to create the 2x1 Array doesn't fit with the other two. I had to do a lot of searching and doc reading to get to it. Second, the printed output of Julia itself differentiates the 2x1 Array from the other two by using a semicolon. That is confusing.

This is a case where the language could be improved.

[stevengj]

For constructing one-column matrices, couldn't we just define a constructor:

Matrix(x::Vector) = reshape(x, length(x), 1)
Matrix(x::AbstractVector) = Matrix(Vector(x))

? These are currently a MethodError, so it would be backward-compatible, explicit, and not require any new syntax.

[StefanKarpinski]

The latest version of JuliaLang/julia#33697 provides a syntax for this: [1;;] is a 1x1 matrix and [1, 2;;] is a 2x1 matrix.

[User-764Q]

Hi,

Following StefanKarpniski's comment in Jan, is it time to close this issue?

Matt.

[judober]

For everyone who (like me) arrives here searching to generate an nx1 matrix:
[1, 2;;] does not work (at least in 1.7.1) but [[1, 2];;] does.
Or [1; 2;;]