How to import noise models from Qiskit¶

Noise models in Qiskit and PennyLane¶

The noise models in Qiskit are built using the tools available in the noise module of the Qiskit-Aer package. Each model is represented by a NoiseModel object that contains QuantumError to describe the errors encountered in gate operations. Optionally, it may also have a ReadoutError that describes the classical readout errors.

Let’s build a Qiskit noise model that inserts depolarization errors for single-qubit gates, bit-flip errors for the target qubit of the two-qubit gates, and amplitude damping errors for each measurement:

import numpy as np
from qiskit_aer.noise import (
    amplitude_damping_error, depolarizing_error, pauli_error, NoiseModel
)

# Building the Qiskit noise model
model_qk = NoiseModel()

# Depolarization error for single-qubit gates
prob_depol = 0.2
error_gate1 = depolarizing_error(prob_depol, 1)
model_qk.add_all_qubit_quantum_error(error_gate1, ["u1", "u2", "u3"])

# Bit flip errors for two-qubit gate
prob_bit_flip = 0.1
error_gate2 = pauli_error([('X', prob_bit_flip), ('I', 1 - prob_bit_flip)]).tensor(
    pauli_error([('I', 1)])
)
model_qk.add_all_qubit_quantum_error(error_gate2, ["cx"])

# Amplitude damping error for measurements
n_qubits = 3
exc_population = 0.2
prob_ampl_damp = np.random.default_rng(42).uniform(0, 0.2, n_qubits)
for qubit in range(n_qubits):
    error_meas = amplitude_damping_error(prob_ampl_damp[qubit], exc_population)
    model_qk.add_quantum_error(error_meas, "measure", [qubit])

print(model_qk)

NoiseModel:
  Basis gates: ['cx', 'id', 'rz', 'sx', 'u1', 'u2', 'u3']
  Instructions with noise: ['cx', 'u3', 'measure', 'u1', 'u2']
  Qubits with noise: [0, 1, 2]
  All-qubits errors: ['u1', 'u2', 'u3', 'cx']
  Specific qubit errors: [('measure', (0,)), ('measure', (1,)), ('measure', (2,))]

In contrast, the noise models in PennyLane are NoiseModel objects with Boolean conditions that select the operation for which we want to apply noise. These conditions are mapped to noise functions that apply (or queue) the corresponding noise for the selected operation or measurement process based on user-provided metadata. This allows for a more functional construction, as we can see by recreating the above noise model as shown below. For more information on this, check out our how-to for noise models in PennyLane. 🧑‍🏫

import pennylane as qml

# Depolarization error for single-qubit gates
gate1_fcond = qml.noise.op_in(["U1", "U2", "U3"]) & qml.noise.wires_in(range(n_qubits))
gate1_noise = qml.noise.partial_wires(qml.DepolarizingChannel, prob_depol)

# Bit flip errors for two-qubit gate
gate2_fcond = qml.noise.op_eq("CNOT")
def gate2_noise(op, **metadata):
    qml.BitFlip(prob_bit_flip, op.wires[1])

# Readout errors for measurements
rmeas_fcond = qml.noise.meas_eq(qml.counts)
def rmeas_noise(op, **metadata):
    for wire in op.wires:
        qml.GeneralizedAmplitudeDamping(prob_ampl_damp[wire], 1 - exc_population, wire)

# Building the PennyLane noise model
model_pl = qml.NoiseModel(
    {gate1_fcond: gate1_noise, gate2_fcond: gate2_noise}, {rmeas_fcond: rmeas_noise},
)

print(model_pl)

NoiseModel({
    OpIn(['U1', 'U2', 'U3']) & WiresIn([0, 1, 2]): DepolarizingChannel(p=0.2)
    OpEq(CNOT): gate2_noise
},
meas_map = {
    MeasEq('CountsMP'): rmeas_noise
})

It is important to verify whether these noise models work the intended way. For this purpose, we will use them while simulating a GHZ state using the default.mixed and qiskit.aer devices. Note that we require add_noise() transform for adding the PennyLane noise model but the Qiskit noise model is provided in the device definition itself:

# Preparing the devices
n_shots = int(2e6)
dev_pl_ideal = qml.device("default.mixed", wires=n_qubits, shots=n_shots)
dev_qk_noisy = qml.device("qiskit.aer", wires=n_qubits, shots=n_shots, noise_model=model_qk)

def GHZcircuit():
    qml.U2(0, np.pi, wires=[0])
    for wire in range(n_qubits-1):
        qml.CNOT([wire, wire + 1])
    return qml.counts(wires=range(n_qubits), all_outcomes=True)

# Preparing the circuits
pl_ideal_circ = qml.QNode(GHZcircuit, dev_pl_ideal)
pl_noisy_circ = qml.add_noise(pl_ideal_circ, noise_model=model_pl)
qk_noisy_circ = qml.QNode(GHZcircuit, dev_qk_noisy)

# Preparing the results
pl_noisy_res, qk_noisy_res = pl_noisy_circ(), qk_noisy_circ()

Now let’s look at the results to compare the two noise models:

pl_probs = np.array(list(pl_noisy_res.values())) / n_shots
qk_probs = np.array(list(qk_noisy_res.values())) / n_shots

print("PennyLane Results: ", np.round(pl_probs, 3))
print("Qiskit Results:    ", np.round(qk_probs, 3))
print("Are results equal? ", np.allclose(pl_probs, qk_probs, atol=1e-2))

PennyLane Results:  [0.385 0.056 0.028 0.076 0.056 0.028 0.082 0.287]
Qiskit Results:     [0.385 0.057 0.028 0.076 0.056 0.028 0.081 0.287]
Are results equal?  True

As the results are equal within a targeted tolerance, we can confirm that the two noise models are equivalent. Note that this tolerance can be further suppressed by increasing the number of shots (n_shots) in the simulation.

Importing Qiskit noise models¶

PennyLane provides the from_qiskit_noise() function to easily convert a Qiskit noise model into an equivalent PennyLane noise model. Let’s look at an example of a noise model based on the GenericBackendV2 backend that gets instantiated with the error data generated and sampled randomly from historical IBM backend data.

from qiskit.providers.fake_provider import GenericBackendV2

backend = GenericBackendV2(num_qubits=2, seed=42)
qk_noise_model = NoiseModel.from_backend(backend)
print(qk_noise_model)

NoiseModel:
  Basis gates: ['cx', 'delay', 'id', 'measure', 'reset', 'rz', 'sx', 'x']
  Instructions with noise: ['cx', 'x', 'measure', 'sx', 'id']
  Qubits with noise: [0, 1]
  Specific qubit errors: [('cx', (0, 1)), ('cx', (1, 0)), ('id', (0,)), ('id', (1,)), ('sx', (0,)), ('sx', (1,)), ('x', (0,)), ('x', (1,)), ('measure', (0,)), ('measure', (1,))]

To import this noise model as a PennyLane one, we simply do:

pl_noise_model = qml.from_qiskit_noise(qk_noise_model)
print(pl_noise_model)

/home/runner/work/qml/qml/venv/lib/python3.10/site-packages/pennylane_qiskit/noise_models.py:131: UserWarning: Readout errors are not supported currently and will be skipped.
  warn("Readout errors are not supported currently and will be skipped.")
NoiseModel({
    OpIn(['CNOT']) & WiresIn([0, 1]): QubitChannel(num_kraus=9, num_wires=2)
    OpIn(['CNOT']) & WiresIn([1, 0]): QubitChannel(num_kraus=16, num_wires=2)
    OpIn(['Identity']) & WiresIn([0]): QubitChannel(num_kraus=3, num_wires=1)
    OpIn(['Identity']) & WiresIn([1]): QubitChannel(num_kraus=3, num_wires=1)
    OpIn(['SX']) & WiresIn([0]): QubitChannel(num_kraus=3, num_wires=1)
    OpIn(['SX']) & WiresIn([1]): QubitChannel(num_kraus=3, num_wires=1)
    OpIn(['PauliX']) & WiresIn([0]): QubitChannel(num_kraus=3, num_wires=1)
    OpIn(['PauliX']) & WiresIn([1]): QubitChannel(num_kraus=3, num_wires=1)
})

This conversion leverages the standard Kraus representation of the errors stored in the qk_noise_model. Internally, this is done in a smart three-step process:

1. First, all the basis gates from the noise model are mapped to the corresponding PennyLane gate operations.

2. Next, the operations with noise are mapped to the corresponding error channels defined via QubitChannel.

3. Finally, the Boolean conditionals are constructed and combined based on their associated errors.

This can be done for any noise model defined in Qiskit with a minor catch that the classical readout errors are not supported yet in PennyLane. However, we can easily re-insert quantum readout errors into our converted noise model. Here’s an example that adds rmeas_fcond and rmeas_noise (defined earlier) to pl_noise_model:

pl_noise_model += {"meas_map": {rmeas_fcond: rmeas_noise}}
print(pl_noise_model.meas_map)

{MeasEq('CountsMP'): <function rmeas_noise at 0x7f9b78636ef0>}

Conclusion¶

Qiskit provides noise models and tools that could be used to mirror the behaviour of quantum devices. Integrating them into PennyLane is a powerful way to enable users to perform differentiable noisy simulations that help them study the effects of noise on quantum circuits and develop noise-robust quantum algorithms. In this how-to guide, we learned how to construct PennyLane noise models from Qiskit ones by manually building one-to-one mappings for each kind of error and also by using the from_qiskit_noise() function to convert the Qiskit noise model automatically. 💪

Should you have any questions about using noise models in PennyLane, you can consult the noise module documentation, the PennyLane Codebook module on Noisy Quantum Theory, or create a post on the PennyLane Discussion Forum.

About the author¶

Total running time of the script: (0 minutes 34.717 seconds)