Transpilation routing causes StandardRB to cheat

[fvoichick]

Informations

• Qiskit Experiments version: 0.5.3
• Python version: 3.8.10
• Operating system: Ubuntu
• I first mentioned this issue in the #experiments channel, where @wshanks suggested opening an issue.

What is the current behavior?

StandardRB gives inaccurate benchmarking results for greater than two qubits. The Qiskit transpiler routing stage does not respect the barrier instructions between the Clifford gates, and so SWAP gates are optimized away using TranspileLayout relabeling. This makes it impossible to accurately benchmark three-qubit gates using StandardRB, because the randomly-sampled Clifford gates are skewed toward those that involve fewer SWAP gates. (Two-qubit RB has a custom transpilation that seems to avoid this problem.)

Steps to reproduce the problem

Run this script:

from itertools import takewhile

from qiskit import QuantumCircuit
from qiskit.quantum_info import Clifford
from qiskit_experiments.library import StandardRB


def circuit_until_barrier(circuit: QuantumCircuit) -> QuantumCircuit:
    result = QuantumCircuit(3)
    for instruction in takewhile(lambda i: i.operation.name != "barrier", circuit.data):
        result.append(instruction)
    return result


experiment = StandardRB(
    physical_qubits=(0, 1, 2),
    lengths=(1,),
    num_samples=1,
    seed=2023,
)
experiment.transpile_options["coupling_map"] = [[0, 1], [1, 0], [1, 2], [2, 1]]
experiment.transpile_options["basis_gates"] = ["rz", "sx", "x", "cx"]
before: QuantumCircuit = experiment.circuits()[0]
after: QuantumCircuit = experiment._transpiled_circuits()[0]
print(Clifford.from_circuit(circuit_until_barrier(before)))
print(Clifford.from_circuit(circuit_until_barrier(after)))
print(after.layout)

It produces the following output:

Clifford: Stabilizer = ['-XIZ', '+ZZY', '+YYY'], Destabilizer = ['+XIX', '-XZI', '+IZI']
Clifford: Stabilizer = ['-IXZ', '+ZZY', '+YYY'], Destabilizer = ['+IXX', '-ZXI', '+ZII']
TranspileLayout(initial_layout=Layout({
0: Qubit(QuantumRegister(3, 'q'), 0),
1: Qubit(QuantumRegister(3, 'q'), 1),
2: Qubit(QuantumRegister(3, 'q'), 2)
}), input_qubit_mapping={Qubit(QuantumRegister(3, 'q'), 0): 0, Qubit(QuantumRegister(3, 'q'), 1): 1, Qubit(QuantumRegister(3, 'q'), 2): 2}, final_layout=Layout({
0: Qubit(QuantumRegister(3, 'q'), 0),
2: Qubit(QuantumRegister(3, 'q'), 1),
1: Qubit(QuantumRegister(3, 'q'), 2)
}))

What is the expected behavior?

The Clifford gates before and after transpilation should be the same. (I would also expect final_layout to match initial_layout).

Suggested solutions

Edit the _transpiled_circuits method so that it decomposes each Clifford gate in a way that preserves the qubit layout.

(edited for clarity)

[wshanks]

One note -- relabeling does not occur for the case where there is all-to-all coupling. This is always the case for two qubit RB, so the special handling there does not matter. The smallest case where this effect occurs is three qubit RB with linearly coupled qubits (for three qubits all coupled to each other it also should not occur).

[coruscating]

Thanks for reporting this. The root of this issue is that the Qiskit transpiler tries to optimize the layout whenever possible by relabeling qubits regardless of the presence of barriers, so we need to add an option to force the final layout to stay the same as the initial. This work is already being tracked in Qiskit/qiskit#8185. In the meantime, we'll add a note to the documentation that 3Q+ experiments cannot be guaranteed to run on the exact physical qubits specified by the user when coupling is limited.