incident with invalid entries

[slwu89]

After removing parts in the codom of a hom, the resulting rows in the printed acset appear as 0 entries. However, using 0 in incident cannot get the rows with invalid codom elements. I'm not sure how to do this, and it would be nice if incident could find those rows.

Very happy to work on PR to get this addressed before the next ACSets.jl update if someone can give me some pointers on where to look in the source.

MWE below, exhibiting this behavior with non indexed and indexed hom.

@present MySch(FreeSchema) begin
  (A,B)::Ob
  f::Hom(A,B)
end

@acset_type MyDataNoIdx(MySch)

@acset_type MyDataIdx(MySch, index=[:f])

data_noidx = @acset MyDataNoIdx begin
  A = 5
  B = 5
  f = [1,2,3,4,5]
end

data_idx = @acset MyDataNoIdx begin
  A = 5
  B = 5
  f = [1,2,3,4,5]
end

rem_parts!(data_noidx, :B, 1:3)
rem_parts!(data_idx, :B, 1:3)

incident(data_noidx, 0, :f)
incident(data_idx, 0, :f)

# would be nice if incident could do
findall(data_noidx[:,:f] .== 0)
findall(data_idx[:,:f] .== 0)

[slwu89]

Not sure if a possible "fix" that looks directly at the Mapping object that implements the hom to find the rows for which the hom is not defined is too janky, when the user calls incident with 0 as input. Something like using haskey and iterating over rows.

[slwu89]

@olynch and @epatters based off the zulip thread, I think we do desire for incident to retrieve indices of undefined elements, and that the computation should be done on the fly when called, for performance and consistency with the math.

Given that when the acset is pretty-printed, undefined elements are shown as 0, it's natural to want the incident(data_noidx, 0, :f) call to be the way the user asks for the indices of undef elements. But I guess there's a few ways to go from here.

• Because the relevant method of incident called here is

  ACSets.jl/src/DenseACSets.jl

  Lines 501 to 503 in 1655674

  	@inline function ACSetInterface.incident(acs::SimpleACSet, part, f::Symbol)
  	preimage(dom_parts(acs, f), acs.subparts[f], part)
  	end

  we could check here if part is 0 or not, and if so we could then send it off to preimage, maybe having a special method of preimage that dispatches on ::Val{0}, although I'm not sure if that's really advantageous for speed in this small case.
• Can ignore it until we get to preimage and do the checking if the user passed 0 at this point.
• Something else I can't think of.

What are your opinions on how to best make this work?

Note to self, check how getindex is converting invalid entries to 0s when using something like acs[:,:f] where f has invalid entries.

[kris-brown]

I think that it would be straightforward to add a new method undefined_parts(my_acset, my_column_name) rather than try to make incident have different behavior when given 0. It's kind of an implementation detail that 0 is how we visualize undefined parts. Plus, incident is a high-performance/low-level function that we don't want to make more complicated.