indexing: require [()] to be a no-op for all ranks?

[lucascolley]

The standard says:

Providing an empty tuple or an ellipsis to an array of rank 0 must result in an array of the same rank (i.e., if A has rank 0, A == A[()] and A == A[...])

Why is this restricted to rank 0? Sure, a reason for the inclusion is probably to be explicit about the divergence from NumPy, but am I missing something about rank 0 being special from the standard's perspective? Since [()] must act like [...] in the 0D case, should it not always act like that, for consistency?

[seberg]

The reason is the later bullet point:

Except in the case of providing a single ellipsis (e.g., A[2:10, ...] or A[1:, ..., 2:5]), the number of provided single-axis indexing expressions (excluding None) should equal N.

Where "should" is "must" for users and no recommendation at all for implementors, IMO.

[lucascolley]

Right, but then my question is why does it not read:

"Except in the case of providing a single ellipsis or an empty tuple (e.g. ...) ..."

Do any of the array libraries have different behaviour for an empty tuple?

[seberg]

OK, I am not sure I understand the point. I would agree that there is nothing special at all about 0-D. I.e. to me the whole bullet point only serves to clarify the empty tuple case because intuition seems to sometimees lead astray for such cases.
In that sense, it would just as well make sense as a clarification to the tuple of ints bullet.

I suspect you are right, it something could be clearer, but I am not sure what :).

[lucascolley]

This arises from SciPy where (for BC) we still need to get NumPy scalars in array API compatible code, in both the production code and the tests. The only library that seems to complain about some_nonzero_rank_array[()] is array-api-strict. Of course one can just guard on ndim (e.g. h-vetinari/scipy#2 (comment)), but I wonder whether a bit of extra specification here would be useful to reduce complexity in user code (while NumPy scalars are still a thing).

Unless I am missing a reasonable alternative interpretation of [()] for higher dims (I suppose the argument is that it is nonsense for higher dims).

[seberg]

OK, then the question is if one should abandon disallowing partial indices, for the sake of saving you the idx = *idx, ... code line in this case. I do eexpect you have to abandon it for all cases, though. I doubt doing it only for 0-D is very helpful.
That has the downside of basically saying that the SymPy way of iterating elements becomes odder (you can still distinguish arr[0] and arr[0,] but it is less clear maybe). OTOH, I think in other places we may have shifted towards the narrow definition as well out of pragmatism in practice.

[lucascolley]

Yeah, perhaps the practical solution is to add a as_scalar_if_np_0d helper to array-api-compat and be clear that [()] is only defined for 0-D, while [...] is always okay. I was just thinking that if every library already agrees that a no-op is the sensible result for [()], it might be worth putting in the standard.

[asmeurer]

FWIW, I was never a fan of the "all axes should be indexed" rule and feel that it leads to unnecessary verbosity for common indexing operations.