Causal cones in data sibling graphs #432
Replies: 5 comments 6 replies
-
Looks like validation checks we have a causal cone: |
Beta Was this translation helpful? Give feedback.
-
I originally thought this property would make convexity checking easier, but I'm not convinced now. @lmondada do you agree? Whether we have this property or not, it isn't the case that every convex subgraph has the property. And as for specification of a convex subgraph, again I think it makes no difference: we can specify the subgraph by its set of minimal nodes (nodes having no predecessors in the set) and its set of maximal nodes (nodes having no successors in the set). |
Beta Was this translation helpful? Give feedback.
-
So leaning towards not requiring this on the grounds that fewer edges is good. |
Beta Was this translation helpful? Give feedback.
-
Think this has been solved now and we decided not to require causal cone (right @cqc-alec? I can't find the PR!) |
Beta Was this translation helpful? Give feedback.
-
Might have been closed by #468 ? |
Beta Was this translation helpful? Give feedback.
-
At one point we discussed requiring Order edges from the Input node into nodes having no data inputs, and similarly for outputs.
This may make analysis of causal relations and convexity a bit easier.
It doesn't seem to be in the spec at the moment.
If we do have this requirement then either the
SimpleReplace
API needs updating so that such edges are inserted, or the user has to explicitly specify them.Another possible objection is that such edges prevent us from moving nodes out of the graph. Arguably Order edges should only exist when there is an actual order requirement.
Beta Was this translation helpful? Give feedback.
All reactions