qutone/lightstim/custom-noise

Custom Noise Model

LightStim separates circuit construction from noise injection. Build the noiseless circuit first (via builder.build_noisy_circuit), then inject noise as a post-processing step. The clean circuit remains available in builder.circuit.

The injection API

noisy_circuit = builder.build_noisy_circuit(
    noise_params,         # NoiseConfig instance
    noise_model,          # str — strategy name, see table below
)
# Returns a new stim.Circuit. Does NOT modify builder.circuit in place.

NoiseConfig fields

from lightstim.noise.config import NoiseConfig
noise = NoiseConfig(
    p_1q=1e-3,    # Depolarizing after single-qubit gates (H, S, ...)
    p_2q=1e-3,    # Depolarizing after two-qubit gates (CX, CZ, ...)
    p_meas=1e-3,  # Measurement bit-flip probability
    p_reset=1e-3, # State-preparation bit-flip probability
    p_idle=0,     # Placement depends on the selected noise model
    custom_params={},  # Arbitrary extra rates for XZ_biased or custom rules
)

Unused fields default to 0. Set only the fields the target noise model uses.

Noise model strategies

Decision guide

• Superconducting hardware: circuit_level with p_2q ≈ 10× p_1q, p_meas ≈ p_2q.
• Per-moment idling: circuit_level_with_idling; set all five rates

equal to p only when the target paper defines one shared probability.

• Threshold plots (fast): phenomenological first, then cross-check key points with circuit_level.
• Code distance study only: code_capacity — fastest, no gate noise.
• Biased hardware (neutral atoms η ≫ 1, trapped ions η ≲ 1): XZ_biased.

XZ-biased noise model

For hardware with asymmetric X/Z error rates, use XZ_biased with the compute_XZ_biased_params helper to convert physical rates + bias ratio η:

from lightstim.noise.injector import NoiseInjector
# eta = p_X / p_Z: >1 means X-biased, <1 means Z-biased
biased_noise = NoiseInjector.compute_XZ_biased_params(
    p_1q=1e-3, p_2q=2e-3, p_meas=1e-3, p_reset=1e-3,
    eta=0.01,   # strongly Z-biased (neutral atom regime)
)
noisy = builder.build_noisy_circuit(biased_noise, noise_model='XZ_biased')

compute_XZ_biased_params fills custom_params with: p_1q_x, p_1q_z, p_2q_x, p_2q_z — the per-axis rates derived from the total rates and the bias η.

Noiseless gate phases

Some protocol phases (e.g. state injection initialization, encoding circuits) should not receive noise. Pass noiseless=True to suppress noise on specific calls:

builder.initialize(init_dict, n, noiseless=True)            # RX/R tags 'noiseless'
builder.apply_syndrome_extraction(se.circuit, noiseless=True) # all gates tagged
builder.apply_unitary_block(gate, noiseless=True)           # same

The noise injector skips all instructions tagged 'noiseless'. This is how state-injection protocols avoid noising the unencoded region.

For circuit_level_with_idling, every TICK-delimited moment uses the complete physical-qubit set. Resets, measurements, and unitary gates count as operations; all other target qubits receive one DEPOLARIZE1(p_idle) channel. Empty moments also receive idle noise. Moments containing only noiseless physical operations remain noiseless.

Working examples

• lightstim/protocols/memory.py — standard circuit-level noise via build_noisy_circuit()
• lightstim/protocols/state_injection.py — selective noise (injection region excluded via tags)
• lightstim/noise/injector.py — NoiseInjector factory methods for all five built-in models