On Jun 16, 2016, at 4:32 PM, Peter F. Patel-Schneider <pfpschnei...@gmail.com>
wrote:
>
> So consider
> BGP( _:x :p ?y )
> against the active graph G = { :s :p _:x . }
>
> The RDF instance mapping is σ = { ( _:x, :s ) }.
> The solution mapping is μ = { ( y, _:x ) }.
> The pattern instance mapping is P = { ( _:x, :s ), ( y, _:x ) }.
> This is a solution because P( _:x :p ?y ) = (:s :p _:x) which is a subgraph
> of G.
>
> The point is that the _:x that came from the EXISTS substitution is a blank
> node and can be itself substituted for in the RDF instance mapping. This is
> counter to what I think is the desired meaning for EXISTS.
>
> Note that this is all just a little bit sloppy. To make it all precise
> would require an extra injection from blank node names to real blank nodes
> but this extra precision doesn't make any difference here.
That’s an interesting case. I haven’t gone through the definitions in depth in
a while, but I think you’re right that the discussion of matching BGPs in §18.3
and the evaluation semantics of EXISTS in §18.6 probably don’t do the expected
thing. The existing errata surrounding EXISTS evaluation shows that this was an
area where the spec fell a bit short.
Do you have a system that is actually giving you the results you describe?
My intuitive understanding, and one which I suspect most if not all
implementations use, is that the “blank nodes” that §18.3 references are the
syntactic blank nodes in the *query*, and shouldn’t apply to blank nodes that
come from the *data* that are used in an EXISTS filter evaluation.
Unfortunately, I don’t think there’s any way for substitute() to express a
(non-syntactic) blank node in the algebraic representation.
Does that align with what you think the expected behavior is?
thanks,
.greg
