Contraction of Multiple ITensors?

+1 vote

Hi,

It seems the contraction * only works for two ITensors. If I have more than two ITensors that share one index, e.g. A(i,j)*B(i,k)*C(i,l) gives wrong results. Is there a function in this library doing this? Thanks.

Jin

commented by (52.6k points)
Hi, could you please expand your question a bit more? What is the result that you expected and what is the result you are getting? We'd like to know if there is a bug or whether it's just surprising behavior.

Thanks,
Miles
commented by (1.1k points)
Hi Miles,

I expect to sum over "i" and get a new tensor @@D(j,k,l) = \sum\_i A(i,j)B(i,k)C(i,l).@@
But A*B*C gives a four rank tensor. It contracts A and B first and then C.
commented by (52.6k points)
I see. This is for the C++ version, correct? (Though it should be the same for the Julia version.)

commented by (1.1k points)
Thanks. I saw this on Julia version.

selected

Hi Jin,
Thanks for the question. So this is the intended, and expected behavior of the ITensor * operator, namely that first A and B are contracted, then the result of that is contracted with C.

One reason for this is that this is just how operators work in C++. There is no notation in C++ that would allow simultaneous contraction of three or more ITensors in that language using operator overloading (*,+, and similar), apart from defining a function call such as multicontract(A,B,C,D,...) (which we might do in the future).

Another reason is that this behavior is quite often the desired behavior that one does want, so it is not in any way a bad thing. It just may not have been what you were expecting. The ITensor interface is based on tensor diagram notation which also uses a pairwise contraction convention. Contracting over an index shared between three or more tensors requires the notion of a "hyperedge" in a tensor network graph, which is perfectly fine to introduce but I'm just mentioning it to say it's not the standard thing that diagrams express.

We do offer something like a hyperedge in ITensor though: it is the delta(m,n,p) ITensor, which is a special ITensor with all its (multi-)diagonal elements set to 1.0. Not all routines involving delta tensors are as optimized as they could be, though some are and in general it's a useful tool.

Hope that helps -

Miles

commented by (1.1k points)