Is it possible to specify maximum bond dimension when using CircuitToEinsum?

[GeorgePennington]

I want to convert a qiskit.QuantumCircuit object to a Matrix Product State (more specifically the state prepared by the circuit) and specify either a) the maximum bond dimension in the MPS or b) the truncation threshold used for the SVD.

Is there a way to specify a) or b) when using cuquantum.CircuitToEinsum? I.e. can I modify qiskit_basic.ipynb example (https://github.com/NVIDIA/cuQuantum/blob/main/python/samples/cutensornet/circuit_converter/qiskit_ba...) to do this?

I understand that I can contract the gates at a low level and specify an algorithm with specific max_extent or rel_cutoff (see https://github.com/NVIDIA/cuQuantum/blob/main/python/samples/cutensornet/tn_algorithms/mps_algorithm...) but is the same thing possible at the high level using CircuitToEinsum?

[yangcal]

Hello George,

Thanks for the question. Note here that CircuitToEinsum, as the name suggests, it's a converter type class rather than a simulator. Currently the outputs from CircuitToEinsum can be parsed to cuquantum.contract to compute corresponding quantities and it does not have direct support for MPS simulation capability.

In order to perform MPS simulation with bond or value truncation, you can parse the gates CircuitToEinsum.gates to the example MPS utility functions in mps_algorithm.ipynb notebook and add args to allow max_extent or rel_cutoff in the decomposition.

Additionally, we have python bindings for our state APIs which can also compute the final MPS state with or without truncations. Please refer to the documentation or examples for more details. We are also working on a more pythonic solution for python users so please stay tuned for our future releases.

Thanks

[GeorgePennington]

Thanks for response and help!

Despite CircuitToEinsum being a converter class, it is also possible to call CircuitToEinsum.reduced_density_matrix() or .amplitude(). To clarify, does that mean that those methods are using the state vector APIs? Or are they using tensor network contractions without truncating? Thank you again in advance for your help!

[yangcal]

If you call CircuitToEinsum.reduced_density_matrix()/amplitude(), you'll get expression and the operands. These expression and operands can then be directly parsed to cuquantum.contract to compute the outcome using cuTensorNet. We don't have pythonic APIs that directly handle expression and operands.