don't descend into our library of math functions

[shashi]

fixes #30

[shashi]

thats a lot of really slow CIs.

[shashi]

I think isleaf should be part of the API.

isleaf(f) = true means that if f is called with x<tag>, then return f(x)<tag> instead of trying f(x<tag>).

So if someone has a function like:

function g(a)
   while a < 1
       a+= rand()
    end
    a
end

for whatever reason

and they then want to sparsity detect

f(y, x) = y .= g.(x)

Then they can set SparsityDetection.isleaf(::typeof(g)) = true and it will work.

[ChrisRackauckas]

Wouldn't they need to also give you what the sparsity of that function is in the API?

[shashi]

Yes, but it can be optional in case of Jacobian. (because you can just say the output depends on all the inputs).. But you're right they would have to say something about it for Hessian.

Well there is a way to say that too SparsityDetection.linearity

[ChrisRackauckas]

oh it's only for scalar functions?

[shashi]

Yeah well that's the easiest thing to provide an API for right away. We currently don't have decent ways of converting sparsity matrix into our metadata (ProvinanceSet or TermCombination although that should be not so hard to do).

Let's hold this thought and if there is a real need for treating sub functions as blackboxes of known sparsity, then we can come back to this idea.