Fail to extract block by specifying sector of each index for (N,0) or (0,N) tensors

[Yue-Zhengyuan]

For a TensorMap with N1 > 0 codomain indices and N2 > 0 domain indices, I can extract its blocks by specifying the sector of each index:

using TensorKit
V = U1Space(0 => 2, 1 => 2, -1 => 2)
a = TensorMap(randn, Float64, V, V)
I = sectortype(a)
a[(I(1), I(-1))]

Output:

2×2 StridedViews.StridedView{Float64, 2, Vector{Float64}, typeof(identity)}:
 -1.29546  -0.115814
  1.63631  -0.808197

However, if I permute all indices to codomain (or domain), this no longer works:

a2 = permute(a, ((1,2), ()))
a2[(I(1), I(-1))]

I get the error:

MethodError: no method matching FusionTree(::Tuple{}, ::U1Irrep, ::Tuple{})
The type `FusionTree` exists, but no method is defined for this combination of argument types when trying to construct it.

[Jutho]

Thanks for catching this. I can reproduce and will fix this asap.

[Jutho]

This is fixed on master, we will tag a new patch release asap.

[Yue-Zhengyuan]

Thanks for the fix! I'd appreciate it if the fix can also be backported to version 0.12? PEPSKit and MPSKit haven't been updated to work with version 0.13.

I need this fix mainly to convert blocks of a TensorMap to a list of ordinary arrays readable by other people's non-Julia programs. Usually people don't distinguish between domain and codomain indices, so I decided to permute all of them to the codomain (as a Tensor), which is also how I found the bug in the first place.

[lkdvos]

One possible way to get around this would be to do something like:

for (f1, f2) in fusiontrees(a2)
    if f1.uncoupled == (I(-1), I(1))
         @show a2[f1, f2]
    end
end

The trick is that this bypasses that step of constructing the fusiontrees by starting from them, and checking if the uncoupled charges are the ones you are looking for.
Would that be sufficient for your current use-case?

[Yue-Zhengyuan]

@lkdvos This seems to work for both 0.12 and 0.13. Thanks!