garrison
commented
1 year ago To do this, we'll need to store the weights in a bit more complicated way (https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/pull/385#discussion_r1310677757) and be also to map between subexperiment index and (sample, observable) index #387.
The QPD bases we have considered so far each have 6 unique elements in their
maps. However, if one looks at only a single qubit (either one), there are actually only 5 unique operations, as two of them are the sameQPDMeasureoperation:https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/blob/58376677cffdf9e0579b7f2ae3c2b8bd92e13ad0/circuit_knitting/cutting/qpd/qpd.py#L243-L250
For a separable circuit, when it comes time to run the sub-experiments, it's only necessary to run 5 distinct sub-experiments on each partition. However, right now, 6 subexperiments are run instead. We already have a test to make sure we can recognize these duplicate subexperiments. We should resolve this issue once we no longer send duplicate experiments to the backends.
More advanced thoughts regarding wire cutting
A lot of cases of duplicate circuits are straightforward, and we should start with them.
A tricky case, however, is in wire cutting. The
Imeasurement and theZmeasurement can be made through exactly the same circuit; the only difference is whether the corresponding bit of the qpd register is considered or ignored.https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/blob/669d9fdfa075ab31cc110be7a75940857c0a249c/circuit_knitting_toolbox/circuit_cutting/qpd/qpd.py#L337 https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/blob/669d9fdfa075ab31cc110be7a75940857c0a249c/circuit_knitting_toolbox/circuit_cutting/qpd/qpd.py#L340
Currently, we always take the parity of all available qpd measurement bits, but if we were to have a bit more control in post-processing, we could do this without running each circuit separately.