Rename issparse?

[garrison]

... aka Bikeshed and/or improve documentation for issparse. The docstring for issparse says

issparse(S)

Returns true if S is sparse, and false otherwise.

but this is ambiguous. What, precisely, does it mean for S (presumably an array) to be sparse? Here are two (very different) possible definitions:

1. The array has a density of nonzero elements less than some parameter x (x could have some default value, but this would be rather arbitrary).
2. The array is represented by a some "sparse" data type.

It is the second option that is implemented by issparse. In a way this makes sense, as the function does not take such a parameter x (representing the threshold density of nonzero elements). But at a minimum this should be more clearly explained in the documentation. (I'm also unable to explain why issparse(::Diagonal) and related types return false -- perhaps it's a bug, or perhaps I do not understand fully what this function is meant to do.)

On the other hand, the current behavior makes issparse different from many other functions I can think of, such as istril, istriu, issymmetric, ishermitian, and isreal. Each of these considers the value itself rather than the datatype; for instance, isreal(0im) == true.

Perhaps a better name for this function, then, would be issparsetype? I can think of downsides to this name (you do not pass a type to it). Another idea would be issparsestorage, but I don't know if I like this either.

So I ask:

1. Do people agree that issparse is a poor name? If it is, can we come up with anything better?
2. What should issparse(::Diagonal), issparse(::Tridiagonal), etc. return?
3. Is this function even useful? I just searched discourse, and it's never been mentioned there. In Base,
   it is used only in the concatenation functions in sparsevector.jl. I have not grepped the package ecosystem, but I wonder how frequently it is used there.

[andreasnoack]

The methods are very old and most likely inherited from Matlab. In Julia, I think it makes more sense to check if the type is a SparseMatrixCSC or AbstractSparseMatrix but there is the complication that a this would be false for e.g. LowerTriangular{SparseMatrixCSC} for which issparse is true.

[oxinabox]

issparse should be a property of the type only.
Like a trait.

as @andreasnoack said it can't be just replaced with:

x isa AbstractSparseMatrix as it also depends on type parameters for e.g. LowerTriangular.

Its currently defined in

julia/base/sparse/abstractsparse.jl

Lines 13 to 23 in ea34968

	issparse(A::AbstractArray) = false
	issparse(S::AbstractSparseArray) = true
	issparse(S::Symmetric{<:Any,<:AbstractSparseMatrix}) = true
	issparse(S::Hermitian{<:Any,<:AbstractSparseMatrix}) = true
	issparse(S::LowerTriangular{<:Any,<:AbstractSparseMatrix}) = true
	issparse(S::LinAlg.UnitLowerTriangular{<:Any,<:AbstractSparseMatrix}) = true
	issparse(S::UpperTriangular{<:Any,<:AbstractSparseMatrix}) = true
	issparse(S::LinAlg.UnitUpperTriangular{<:Any,<:AbstractSparseMatrix}) = true
	indtype(S::AbstractSparseArray{<:Any,Ti}) where {Ti} = Ti

It would be marginally cleaner if it were:

issparse(x:T) where {T} = issparse(T)

issparse(::Type{AbstractArray}) = false
issparse(::Type{AbstractSparseArray}) = true
issparse(::Type{Symmetric{<:Any, T}}) where {T} = issparse(T)

But is it an useful trait is a good question.

Having it as a trait might be marginally cleaner in

julia/base/sparse/sparsevector.jl

Line 1015 in 82ae10d

function cat(catdims, Xin::_SparseConcatGroup...)

[garrison]

Here is a list of uses of issparse in the julia package ecosystem, created using cloneallpackages and grep.

[ViralBShah]

I like the isa solution offered above, and if people are generally onboard, it shouldn't be difficult to get this changed in all the packages.

Perhaps FemtoCleaner can automate it all.

[User-764Q]

HI,

It looks like this was merged with #34308 in Dec 18. Does that mean this issue can be closed?

Matt.