Sampler primitive in Terra does not support mid-circuit measurements

[garrison]

What should we add?

The reference Sampler (in qiskit.primitives) does not support mid-circuit measurements. However, other Samplers, including those provided by qiskit-aer and qiskit-ibm-runtime, do support mid-circuit measurements.

I understand the desire to keep the reference implementation simple, but it is hardly a reference implementation if it only implements the Sampler for a subset of possible inputs.

This was previously discussed at #8708. My goal here is to open a centralized place for discussion going forward

[garrison]

As for the question of "why don't you just use the Aer sampler instead?" ...

I'd like to calculate the exact probability of each outcome in a circuit with mid-circuit measurements. Aer cannot calculate exact probabilities of each classical register outcome, but the reference Sampler has this ability (set shots=None). EDIT: Aer has the same problem as this package when shots=None: Qiskit/qiskit-aer#1810.

[garrison]

I figured out and implemented a way to meet my current need by evolving the Statevector according to each instruction, and considering both outcomes each time there is a measurement in the circuit. If desired, I could attempt to contribute this upstream to Qiskit Terra, so that we could close this issue. This would also have the side effect of allowing unused bits in the classical registers, which is currently not supported:

https://github.com/Qiskit/qiskit-terra/blob/94e9480ddb5d1304a45ad8b463668ebe144a43d2/qiskit/primitives/sampler.py#L168-L173

[mtreinish]

I think this is reasonable to add to the Sampler class. That current restriction looks like it's only in place because of how the sampler goes from the probabilities to the expect bitstring and isn't really inherent to sampling. So if we can remove that and support midcircuit measurement that seems reasonable to me. @t-imamichi @ikkoham what do you think?

[t-imamichi]

Because the reference implementation Sampler and Estimator were developed before mid-circuit measurements were released, the reference Sampler and Estimator cannot handle mid-circuit measurements.
I recently opened #9692 to raise an error for Estimator.

Now that we have BackendEstimator and BackendSampler, it is an option to extend them to handle mid-circuit measurements and deprecate the current reference impl Sampler and Estimator. @ikkoham opened #9700 for the discussion.

[jacobwatkins1]

Hello, is there any update on this? I would like to use Sampler for an iterative phase estimation algorithm, which necessarily requires mid circuit measurements.

[garrison]

I would like to use Sampler for an iterative phase estimation algorithm, which necessarily requires mid circuit measurements.

You could use Aer's sampler with shots != None.

[jacobwatkins1]

Edit: @garrison already noticed this prior to the comment below.

I am actually running into the same issue with the qiskit_aer.primitives Sampler.

QiskitError: 'Some classical bits are not used for measurements. The number of classical bits 4, the used classical bits set().'

Here is the code generating the problem. It is pretty minimal. Using version >= 12 for qiskit_aer and 0.44.0 for qiskit:

[garrison]

I am actually running into the same issue with the qiskit_aer.primitives Sampler.

The corresponding issue for qiskit-aer (which is triggered only when shots=None) is Qiskit/qiskit-aer#1810.

[t-imamichi]

I close this because the Primitives V1 is deprecated #12575