MPS shape conventions mismatch between MPS_computational_state and DMRG

[thibxlv]

What is your issue?

I'm trying to use DMRG to initialize a quantum circuit with ancillas. I am trying to prepare a state $|k\rangle|\psi\rangle$ where $|k\rangle$ is a computational basis state and $|\psi\rangle$ is the DMRG approximation of the ground state.

Quimb uses the 'lpr' shape convention for the DMRG output indices, but everywhere else it seems to use the 'lrp' convention. Besides, calling quimb.tensor.MPS_computational_state with the option shape='lpr' produces a TypeError:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[271], line 1
----> 1 zero_mps = qtn.MPS_computational_state('0'*m_L, shape='lpr')

File ~/*****/lib/python3.12/site-packages/quimb/tensor/tensor_builder.py:3802, in MPS_computational_state(binary, dtype, cyclic, **mps_opts)
   3799     for s in binary:
   3800         yield array_map[str(s)]
-> 3802 return MPS_product_state(tuple(gen_arrays()), cyclic=cyclic, **mps_opts)

File ~/*****/lib/python3.12/site-packages/quimb/tensor/tensor_builder.py:3772, in MPS_product_state(arrays, cyclic, **mps_opts)
   3765     yield (*cyc_dim, 1, -1)
   3767 mps_arrays = (
   3768     asarray(array).reshape(*shape)
   3769     for array, shape in zip(arrays, gen_array_shapes())
   3770 )
-> 3772 return MatrixProductState(mps_arrays, shape="lrp", **mps_opts)

TypeError: quimb.tensor.tensor_1d.MatrixProductState() got multiple values for keyword argument 'shape'

It would be nice either to have a way to change the shape convention of the DMRG.state MPS, or to unify conventions by default. Thanks!

[thibxlv]

Creating a MPS from a tensor network that combines the two MPSs and creating a new bond works, but then it has no uniform shape convention for all tensors, which makes it more cumbersome if one wants to manipulate the data arrays afterwards

[jcmgray]

Hi @thibxlv, thanks for the issue. I think the method you're probably looking for is permute_arrays. Which enforces a particular internal layout convention, useful for extracting the arrays from quimb to something else.

As a side note, the shape kwarg to various TN constructors is (slightly confusingly) actually specifying the supplied layout. It would be nice update the docs and maybe even the variable names to clarify these things.