scalars: enforced rule or just a best practice?

[janbrandejs]

I think we should enforce or at least declare best practice on scalars either by saying:

1. scalar has 0 modes
2. scalar has all modes of size 1
3. both are allowed

I vote for 3, but declaring that our best practice reccommendation is option 1.

When using TAPP scalars were the most frequent problem.

[evaleev]

I presume this is about scalars that result from tensor algebra, not scalars that are arguments to e.g. GEMM-like tensor products?

One can already represent a tensor of arbitrary order with all modes of unit extent as a tensor; such tensor has 1 element. Tensor product-like reductions (a(i,j,k) * b(i,k,j) -> c()) are supposed to produce tensors of order 0, which are also scalars. But other reductions (e.g. maxabs) may produce things that are true "scalars", however, in fact such scalars may belong to a different field that the argument tensor ( maxabs(Tensor) -> double ). As a result in TA we handle all reductions by separate API that produce scalars or other things, even those that are tensor product-like.

[pauldj]

I agree with Ed.
Also, we can already sketch an interface for a scal function.