Noise Models¶

The package supports small, explicit noise-model dictionaries for circuit-backed workflows. Noise is opt-in: when noise_model=None or all probabilities are zero, workflows keep using noiseless analytic or finite-shot execution.

Supported Channels¶

Use qml.noise.build_noise_model(...) or pass a dictionary with these keys:

All values are probabilities in [0, 1]. Nonzero noise models run on PennyLane default.mixed; noiseless workflows continue to use default.qubit.

The probability parameters mean:

• depolarizing: probability of replacing a wire state with a depolarized state after supported circuit operations.
• amplitude_damping: probability of amplitude damping on each circuit wire.
• readout_error: probability of a bit-flip approximation to measurement readout error on measured wires.

Example¶

from qml import QuantumKernel, build_noise_model, run_vqc

noise_model = build_noise_model(
    depolarizing=0.01,
    amplitude_damping=0.0,
    readout_error=0.02,
)

result = run_vqc(
    n_samples=80,
    n_layers=1,
    steps=10,
    shots=256,
    noise_model=noise_model,
)

kernel = QuantumKernel(shots=256, noise_model=noise_model)

Here shots is the number of circuit samples per expectation or probability estimate. shots=None means analytic simulator execution.

CLI¶

Circuit-backed commands and benchmark commands accept:

qml-pennylane vqc --shots 256 --depolarizing 0.01 --readout-error 0.02

qml-pennylane benchmark finite-shots \
  --shots analytic 128 512 \
  --depolarizing 0.005 \
  --readout-error 0.02

Interpretation¶

Noisy results should be reported next to the matching noiseless analytic or finite-shot result using the same train/test split and seeds. Treat these runs as robustness checks, not hardware-validated performance claims.