Suppose I've got a tensor network (TN) representing some bipartite quantum state, $|\Psi\rangle$. Using the Schmidt decomposition, I can write $|\Psi\rangle = \sum_{k=1}^r \sqrt{\lambda_k}|\chi_k\rangle_A \otimes |\phi_k\rangle_B$, for some upper bond dimension $r$. My question is: can the mutual information $I(A:B)$ be efficiently and accurately computed for this system in the case where the Hilbert space of $A$ is small but $B$ is large? (For example, consider a 1D chain, where $A$ corresponds to two spins in the centre, and $B$ is the rest of the chain).
Here's my best understanding of the problem: I know that $I(A:B) = S(A) + S(B) - S(AB)$. Because $|\Psi\rangle$ is a pure state, we know that $S(A) = S(B) = -\sum_{k=1}^r \lambda_k \log \lambda_k$. For a reasonable bond dimension, I feel like this calculation should be straightforward to implement. However, I have no good idea for how difficult it is to estimate $S(AB) \equiv S(\rho_{AB})$, where $\rho_{AB}$ is the density matrix corresponding to the whole system.
Also--does this problem become easier if we assume some structure in the TN? For example, if we assume that it's an MPS, then can we more easily calculate the mutual information?
Thanks for all the help!