Sampler and Estimator with noise model doesn't work consistently for parametrized circuits

[HwajungKang]

Informations

• Qiskit Aer version: 0.12.0
• Python version: 3.11
• Operating system: osx

What is the current behavior?

Running parametrized circuits using Sampler or Estimator with noise model fails. Although it works when the noise model is built using certain fakebackends like FakeCambridge.

Steps to reproduce the problem

The following code works when backend2 is passed to the noise model but fails with backend1.
Same for Estimator as well.

from qiskit import *
from qiskit.circuit import ParameterVector
from qiskit.providers.fake_provider import FakeCambridge, FakeKolkata
from qiskit_aer.noise import NoiseModel
from qiskit_aer.primitives import Sampler

backend1, backend2 = FakeCambridge(), FakeKolkata()
noise = NoiseModel.from_backend(backend2)

run_opt = {'shots':10000}
backend_opt = {'noise_model':noise}

samp = Sampler(backend_options=backend_opt, run_options=run_opt)

ang = ParameterVector('x', 3)
qc = QuantumCircuit(1)
qc.u(*ang, [0])
qc.measure_all()

res = samp.run(qc, parameter_values=[1.3, 2.1, 0.7]).result()
print(res.quasi_dists)

What is the expected behavior?

It should work consistently regardless of what fakebackend is used for noise model.

Suggested solutions

[hhorii]

Thank you for your reporting. This is from the same reason of #1795.

Your parameterized circuit is transpiled with basis gates of noise. The transpiled circuit is as follows:

global phase: x[1]/2 + x[2]/2 - π/2
                 ┌──────────┐┌────┐┌──────────────┐┌────┐┌───────────────┐
        q_0 -> 0 ┤ Rz(x[2]) ├┤ √X ├┤ Rz(x[0] + π) ├┤ √X ├┤ Rz(x[1] + 3π) ├
                 └──────────┘└────┘└──────────────┘└────┘└───────────────┘

Its global phase includes parameter expressions, which are not correctly handled by the current implementation as reported #1795.