The noisy simulation of dynamic circuits gives to incorrect measurement outcomes

[xlelephant]

Informations

• Qiskit Aer version: 1.0.2
• Python version: 3.9.7
• Operating system: Windows

What is the current behavior?

Mid-circuit measurement and conditional branching are important features in Quantum computing.
Unfortunately, the Aer simulator is presently yielding wrong results when used for noisy simulation of dynamic circuits.

Steps to reproduce the problem

qregs = qiskit.QuantumRegister(2, name="Q")
final_reg = qiskit.ClassicalRegister(1, name="M")
cregs = qiskit.ClassicalRegister(1, name="R")
circ = qiskit.QuantumCircuit(qregs[:-1], cregs, qregs[-1:], final_reg)

circ.unitary(np.array([[0, 1], [1, 0]]), 0, label="excite")
circ.measure(qregs[0], final_reg)

circ.measure(qregs[1], cregs[0])
####### insertion of dynamic circuit #######
with circ.if_test((cregs[0], 1)):
    circ.y(qregs[-1])
####### insertion of dynamic circuit #######

run_options = {"shots": 10000, "seed_simulator": 123}
error = qiskit_aer.noise.depolarizing_error(0.1, 1)
noise_model = qiskit_aer.noise.NoiseModel()
noise_model.add_quantum_error(error, ["excite"], [0])
backend = qiskit_aer.AerSimulator(method="automatic", noise_model=noise_model)
job = backend.run(circ, noise_model=noise_model, **run_options)
result = job.result()
fid = (
    sum([count for bitstring, count in result.get_counts().items() if bitstring[0] == "1"])
    / run_options["shots"]
)
print("fidelity: ", fid)
circ.draw(output="mpl")

> fidelity: 1.0 (no noise effect)
> fidelity: 0.9509 (noisy)

no dynamic circuit

with dynamic circuit

What is the expected behavior?

The fidelity should be less than 1.0 since depolarizing error is added to the qubit(qregs[0]). However, it goes to 1.0 if I add one more qubit (qregs[1]) and insert some dynamic branches in the circuit, which I think should not effect qregs[0] in any way.

Suggested solutions

It appears that the inclusion of the conditional gate has interfered with the simulation noise model.
Please have a check~

[xlelephant]

noise_model.add_quantum_error(error, ["excite"], [0])

And if I add errors to other instructions, with names like 'rx', 'u', 'u3', the outcomes seem noisy again.

[doichanj]

This is bug in aer_compiler that the labeled gate is transplied into other gates when the circuit is dynamic circuit.

[xlelephant]

This is bug in aer_compiler that the labeled gate is transplied into other gates when the circuit is dynamic circuit.

Thanks for the answer,

I am wondering if a circuit composed solely of native gates, such as ['u', 'cz', 'ecr', ...], can circumvent this issue. Can I instruct the aer_compiler not to perform transpilation?