sum over a SparseAxisArray

[hdavid16]

It would be nice to be able to execute Base.sum over a variable defined with a SparseAxisArray in a more straightforward fashion (perhaps extending Base.sum in the utils.jl file?).

For example:

Given a variable defined as:

@variable(model, x[i=1:2, j=i:2])

You could sum over it doing:

sum(x[i,j] for i = 1:2 for j = i:2)

However, it would be nicer to be able to remove the second for:

sum(x[i,j] for i = 1:2, j = i:2)

(which can't be done due to a UndefVarError)

[odow]

This is base Julia, not JuMP:

julia> y = rand(2, 2)
2×2 Array{Float64,2}:
 0.408406  0.337018
 0.188717  0.273776

julia> sum(y[i,j] for i = 1:2, j = i:2)
ERROR: UndefVarError: i not defined
Stacktrace:
 [1] top-level scope at REPL[102]:1

If you have triangular indexing, you must specify multiple for.

[odow]

Closing because this is expected behavior.

[hdavid16]

@odow Yes. I know it's base Julia. It is just feature that would be nice in the context of JuMP since it is common to sum over a variable that has sparse structure in math programming. So I wasn't sure if it was best to post this here or in Julia. I posted it here since JuMP uses these sparse structures nicely when declaring variables or constraints.

[odow]

I don't think we want to depart from the Julia syntax, because this will just cause more confusion. JuMP's macros rewrite expressions for performance, not to add subtly different functionality.

You could open an issue in Julia. Here's an example which gets at what is happening:

julia> f() = sum(i + j for i=1:2, j=i:2)
f (generic function with 1 method)

julia> @code_lowered f()
CodeInfo(
1 ─      #414 = %new(Main.:(var"#414#415"))
│   %2 = #414
│   %3 = 1:2
│   %4 = Main.i:2
│   %5 = Base.product(%3, %4)
│   %6 = Base.Generator(%2, %5)
│   %7 = Main.sum(%6)
└──      return %7
)

julia> f()
ERROR: UndefVarError: i not defined
Stacktrace:
 [1] f() at ./REPL[127]:1
 [2] top-level scope at REPL[129]:1

Relevant issues (might be more, I didn't check thoroughly):
JuliaLang/julia#4867
JuliaLang/julia#15384

[hdavid16]

Sounds good. Thanks