noisebridge 0.3.0

Adaptive quantum noise mitigation via Spiking Neural Networks (SNN-NR)

NoiseBridge

Post-processing quantum noise mitigation via a centering-matrix Syndrome Error Corrector (SEC).
Zero decoder overhead demonstrated on IBM Kingston real hardware — Surface Code d=3.

---

Key Result

LER_SNN = LER_raw = 0.0422 on IBM ibm_kingston (Surface Code d=3, p=0.25, 4096 shots)
First zero-overhead decoding result on real quantum hardware. SNN outperforms standard lookup decoder by 1.72× at p=0.10.

Version	Qubits	Depth	CZ gates	LER_raw	LER_SNN	SNN vs std
v1 linear	17	244	153	0.1785	0.2585	0.88×
v2 2D 17Q	17	242	153	0.2449	0.2944	1.01×
v3 Z-only p=0.10	13	26	25	0.0388	0.0457	1.72×
v3 Z-only p=0.25	13	26	25	0.0422	0.0422	1.27× (zero overhead)

---

Install

pip install noisebridge

With Qiskit integration:

pip install "noisebridge[qiskit]"

---

Quickstart

Surface Code syndrome decoding (IBM Kingston result)

from noisebridge import correct, load_params

# Load calibrated params for IBM Kingston
params = load_params("ibm_kingston")

# Raw Z-syndrome counts from hardware (4096 shots, p=0.25, 2 injected errors)
raw_counts = {
    "0111": 3382, "0011": 213, "0110": 184,
    "0101": 131,  "1111": 41,  "0100": 35,
}

# SNN Syndrome Error Corrector — soft-decodes Hamming-1 neighbours
corrected = correct(raw_counts, params, n=4)
# Dominant syndrome 0111 amplified; noise neighbours suppressed

Combined REM + SNN pipeline (recommended for general use)

from noisebridge import rem_snn_correct

corrected = rem_snn_correct(
    {"00": 122, "01": 3, "10": 3, "11": 128},
    n=2,
    device="iqm_garnet"
)

REM only

from noisebridge import rem_correct

p_clean = rem_correct(
    {"0": 480, "1": 520},
    n=1,
    device="ibm_kingston"
)

List supported devices

from noisebridge import list_devices

list_devices()
list_devices(recommended_only=True)

---

How It Works

NoiseBridge applies three composable post-processing strategies to raw measurement counts after a quantum circuit executes on hardware. No additional QPU shots required.

1. REM — Readout Error Mitigation
Inverts the readout confusion matrix to correct SPAM errors.
Validated on IQM Garnet real QPU: avg +0.030 fidelity (5/5 circuits positive).

2. SNN-SEC — Centering-Matrix Syndrome Error Corrector
The weight matrix uses a centering architecture:

W[i,j] = (δ(i,j) − 1/N) × W_scale

This subtracts the global mean from each syndrome probability, amplifying the dominant (true) syndrome and suppressing Hamming-1 measurement noise neighbours.
Validated on IQM Garnet real QPU: avg +0.057 fidelity (6/6 circuits positive).

3. REM → SNN pipeline (recommended)
Sequential application achieves best performance on noisy hardware.

Why Z-only for Surface Code d=3?

The state |0⟩^⊗9 is not an eigenstate of X-stabilizers — measuring them produces random 50/50 noise with zero information content. Removing X-stabilizer measurement:

• Reduces transpiled depth 242 → 26 (89% reduction)
• Reduces 2-qubit gate count 153 → 25 CZ (84% reduction)
• Eliminates all SWAP gates via native heavy-hex layout

This is the root cause of the zero-overhead result at p=0.25.

---

Hardware Support

Device	Provider	Validated	Notes
ibm_kingston	IBM Quantum	✅	Surface code d=3 benchmark (2026-05-13)
iqm_garnet	IQM	✅	20Q, avg +0.057 fidelity
iqm_emerald	IQM	⚠️	Params present, not yet validated on real QPU
ibm_fakemarrakesh	IBM (simulator)	✅	56% win rate vs ZNE, p=0.011
rigetti_aspen_m3	Rigetti	⚠️	Noise-model calibrated

---

IBM Kingston Layout (Surface Code d=3 v3)

Physical qubits:
  D0-D8  = [77, 56, 62, 66, 64, 85, 68, 57, 83]
  AZ0-AZ3 = [65, 63, 67, 84]

Z-stabilizer native edges: 10/12
  AZ0: 2/2  AZ1: 3/4  AZ2: 3/4  AZ3: 2/2

Transpiled circuit:  depth=26  CZ=25  SWAP=0

See examples/ibm_kingston_surface_d3.py for the full runnable benchmark.

---

Citation

If you use NoiseBridge in your research, please cite:

@software{noisebridge2026,
  author  = {{FractKit Project}},
  title   = {NoiseBridge: Zero-Overhead Syndrome Decoding via Centering-Matrix SNN},
  year    = {2026},
  url     = {https://github.com/MousMou/noisebridge},
  version = {0.3.0}
}

Preprint: Zero-Overhead Syndrome Decoding on Real Quantum Hardware via a Centering-Matrix SNN Syndrome Error Corrector — FractKit Project, 2026. Available on Zenodo (DOI pending).

---

License

MIT © 2026 FractKit Project