Published October 16, 2026 | Version v3

Artifact of Live Noise Fingerprints as an Empirical Semantics for Noisy Quantum Channels

Authors/Creators

Description

Abstract

Quantum computations execute on noisy hardware whose effective noise is opaque and varies across platforms and over time. Complete noise-channel descriptions are often too detailed for reasoning, comparison, or analysis. This work studies the semantic problem of constructing tractable abstractions of noisy quantum channels: given a finite set of observations, how can channel behaviour be summarised while preserving properties of interest?

We characterise the effective noise of a quantum channel as an abstract domain of noise fingerprints. A noise fingerprint is an observational abstraction of a quantum channel over a finite experiment interface. We relate channels and fingerprints through a probabilistic Galois connection, in which concretisation returns a probability distribution over channels and is compared with them through its expected value. This construction reduces to the classical Galois connection when every concretisation is a point mass. Measured fingerprints carry a separate (ε, δ) estimation guarantee, separating tolerance ε from failure probability δ over measurement sessions. We state four abstraction obligations: soundness, conciseness, efficiency, and adaptiveness.

The package provides an executable instantiation of the abstraction function using the classical-shadow estimator, together with a partial concretisation function that reconstructs channels from fingerprints via semidefinite programming. Using the information orders of Coecke and Martin on classical and quantum states, it also provides a probabilistic concretisation of Pauli channels as random mixtures of unitary operations. A worked example shows how fingerprints determine whether two realisations of the same program are observationally equivalent across environments and compiler passes, and localise the gate at which they diverge.

The evaluation reports that the Galois law holds on all tested pairs of Pauli channels; measured fingerprints meet their (ε, δ) coverage guarantee; the soundness bound is tight; compression relative to the concrete channel domain is exponential while nearly all information about noise strength is retained on a synthetic benchmark; sample complexity grows sublinearly in the interface size, consistent with the logarithmic bound; and incremental updates track channel drift at the optimal rate. Additional experiments examine physically meaningful noise signatures and distinguish cross-platform behaviours.

Together, this work brings abstract interpretation to the empirical semantics of noisy quantum channels, offering a lightweight, formally grounded abstraction for characterising, comparing, and monitoring noisy quantum executions.

Context and Method

Characterising noisy quantum channels remains challenging: full process tomography scales exponentially with system size, while common noise metrics offer limited semantic guarantees.

A noise fingerprint is the matrix of deviations between observed and ideal Pauli expectation values over a fixed experiment interface. Channels and fingerprints are ordered by information, using the Coecke–Martin orders on classical and quantum states.

The abstraction maps each completely positive trace-preserving (CPTP) map to its fingerprint. The concretisation returns a probability distribution over channels, and the two satisfy the Galois-connection law with the concretisation compared through its expected value.

Separately, a fingerprint measured with T rounds lies within ε of the exact fingerprint with probability at least 1 − δ.

The framework is evaluated against four obligations:

  • Soundness: the probabilistic Galois connection, an exact observational-soundness theorem, and a Lipschitz bound (L ≤ 1) relating abstract distance to diamond-norm distance.

  • Conciseness: compression relative to the full channel representation.

  • Estimation efficiency: the (ε, δ) guarantee from finite measurements, with logarithmic growth in the interface size.

  • Adaptiveness: incremental updates under channel drift.

The implementation uses the Huang–Kueng–Preskill classical-shadow estimator with median-of-means aggregation. The package includes exact SDP-based diamond norms for small systems, two independently validated large-system ground truths (completely bounded norm multiplicativity for single-site channel pairs and the exact Pauli-channel ℓ1 formula), and a product-factorised classical-shadow sampler validated against the full sampler.

Formal properties are checked by 59 automated tests. Every number reported in the paper is regenerated by a script; CLAIMS.md maps each number to its generating script and output.

What's New in Version v3

  • Probabilistic Galois connection (new definition): src/infoorder.py implements the Coecke–Martin information orders and the probabilistic concretisation of Pauli channels. experiments/galois_check.py (make galois) checks the Galois law.

  • (ε, δ) estimation guarantee: src/pgc.py and experiments/pgc_coverage.py test it over 48 configurations, with a held-out calibration check and a negative control. Version v3 called this the probabilistic Galois connection.

  • Consistent confidence criterion: every "rounds needed to reach a tolerance ε" result is the smallest budget at which the error is at most ε in at least 95 of 100 independent sessions (δ = 0.05). The median (typical case) is recorded alongside it.

  • Generated paper numbers and figures: experiments/export_numbers.py, the coverage experiment, and the Galois check write LaTeX macros (results/paper_numbers.tex, results/pgc_numbers.tex, and results/galois_numbers.tex). README.md identifies the figure file behind each figure in the paper.

  • Worked example: examples/worked_example.ipynb, with a Jupyter-free runner.

  • Expanded tests: 59 tests (51 in v3).

Artifact Evaluation Summary

The package provides the implementation, experiments, and evaluation pipeline for the paper: a pure Python/NumPy implementation of noisy quantum channels, classical-shadow fingerprint estimation, the information orders and the probabilistic Galois connection, the (ε, δ) estimation guarantee, SDP-based concretisation, 59 tests, a worked example, and scripts that reproduce every figure and quantitative claim.

Evaluators can begin with the worked example, verify correctness with one command, and regenerate the results. Precomputed data and figures allow immediate inspection; after dependencies are installed, the full pipeline runs from scratch without GPUs, quantum hardware, or network access. New programs, noise profiles, qubit counts, and properties can be added with minimal changes.

Package Contents

  • examples/ — executed worked_example.ipynb (with outputs stored) and run_worked_example.py, which runs the example without Jupyter.

  • src/ — states and channels; noise fingerprints; literal classical-shadow estimators (full and product-factorised); the Coecke–Martin information orders and probabilistic concretisation (infoorder.py); the (ε, δ) estimation guarantee (pgc.py); optional readout-error models; exact diamond norms via Watrous SDPs; validated large-system diamond-norm ground truths; SDP-based concretisation; configurable noise models; and a registry of 12 supplementary configurations (profiles.py).

  • experiments/ — scripts for each result, including audit_seed_sweep.py, compliance.py, galois_check.py, pgc_coverage.py, scale_rq.py, realistic_limits.py, shadow_validation.py, expert_verification.py, fingerprints_main.py, figure scripts, export_numbers.py, and supplementary sweeps rq1_sweep.py through rq4_sweep.py.

  • results/ — JSON data for every result and the LaTeX number macros.

  • figures/ — all figures, including the nine used in the paper.

  • tests/ — 59 tests, including tests for exact diamond-norm ground truths, the product-factorised sampler, the median-of-means estimator, the information orders and probabilistic Galois connection, and the (ε, δ) guarantee.

  • Root-level support files: verify.py, CLAIMS.md (claim-by-claim map), NOTES.md (scope, platform notes, and seed-sensitivity check), README.md, Makefile, Dockerfile, requirements.txt, and LICENSE.

Installation

Python 3.10 or later is required.

Install dependencies:

pip install -r requirements.txt

Dependencies: NumPy, SciPy, CVXPY (SCS), scikit-learn, and Matplotlib. No quantum SDK is required; all simulations use a pure NumPy density-matrix implementation. Jupyter is optional.

A Dockerfile is provided:

docker build .

Getting Started: Worked Example

Run the worked example:

python3 examples/run_worked_example.py

Alternatively, open examples/worked_example.ipynb.

The runner takes seconds and uses NumPy only. It accepts a program in OpenQASM, shows what a backend does to it, derives the observation interface from the program's test, computes the fingerprints, and decides whether two backends are interchangeable for that test.

Expected results:

  • Two backends implementing the same noise by different means are judged interchangeable on all five programs (deviations ≈ 10⁻¹⁶).

  • A leftover coherent phase rotation is flagged on four of the five programs and correctly cleared on the fifth, whose test reads only populations.

  • On a population-only interface, the phase rotation is invisible; refining the interface to the observables read by the test restores it.

  • The prefix sweep finds the first divergence at the first CX gate (distance 0.0729, rising to 0.1007 after the second CX).

  • Numerical checks: fingerprint distance equals observational discrepancy to machine precision; concretisation within tolerance is an interval; the median-of-means example has expectation 7/9.

Verification

Run:

make kick-the-tires

This takes under two minutes and runs the 59 tests, a diamond-norm SDP self-check, and a quick coverage check of the (ε, δ) guarantee.

Expected output:

59 passed
ALL CHECKS PASSED

Measurement Protocol

Every "rounds needed to reach ε" result is the smallest budget on a geometric grid at which the error is at most ε in at least 95 of 100 independent sessions (38 of 40 for the detectability study), the empirical form of the (ε, δ) guarantee with δ = 0.05. The median value is stored alongside it.

Stochastic results are reported as mean ± standard deviation over seeds. All randomness is seeded; rerunning reproduces the shipped data (SDP-based values agree to solver tolerance, approximately 10⁻⁶).

 

Expert physics oracle

The oracle lists expected deviations for each noise channel. Each entry is a `(state, observable)` pair. `large` marks pairs expected to show a substantial deviation, and `zero` marks pairs expected to show zero deviation. `Phi+` denotes the Bell state.

ORACLE = {
    "depolarizing": {
        "large": [("11", "ZZ"), ("00", "ZZ"), ("++", "XX"), ("Phi+", "XX")],
        "zero": [("00", "XX"), ("00", "YY"), ("11", "XX"), ("++", "ZZ")],
    },
    "amplitude_damping": {
        "large": [
            ("11", "ZZ"), ("11", "ZI"), ("11", "IZ"), ("10", "IZ"),
            ("01", "ZI"), ("++", "XX"), ("+1", "ZX"),
        ],
        "zero": [("00", "ZZ"), ("00", "IZ"), ("00", "ZI")],
    },
    "phase_damping": {
        "large": [
            ("++", "XX"), ("++", "IX"), ("++", "XI"),
            ("Phi+", "XX"), ("Phi+", "YY"),
        ],
        "zero": [("00", "ZZ"), ("00", "IZ"), ("11", "ZZ"), ("01", "ZI")],
    },
}

Reproducing the Paper's Results

1. Probabilistic Galois connection — make galois (under a minute)

  • 82 single-qubit Pauli channels (random rates, depolarising and dephasing families), 6,724 pairs: the Galois law holds in every case (0 violations); 1,691 pairs are ordered, so it is not satisfied vacuously; the expected concretisation reconstructs every channel exactly.

  • 43 other channels (amplitude damping, IBM-calibrated, random), 1,849 pairs: the abstract and concrete orders agree, but only trivially, because no two distinct channels are ordered: their outputs on some probe do not commute.

2. (ε, δ) estimation guarantee — make pgc (about 15 minutes)

  • 48 configurations: 2 interfaces, 4 channels, δ ∈ {0.05, 0.01}, and T ∈ {2000, 8000, 32000}; 400–1000 sessions per configuration.

  • Certified tolerance: every measured fingerprint lies within it in every session; the Clopper–Pearson lower bound is ≥ 1 − δ in all 48 configurations. The certified tolerance is 4.8–6.7 times the smallest valid one.

  • Held-out coverage of the quantile-calibrated tolerance: 0.957 (δ = 0.05) and 0.990 (δ = 0.01).

  • Negative control (per-entry interval, without a union over the interface): fails in 24 of 24 two-qubit configurations (coverage 0.00–0.03).

  • SDP concretisation never lies farther from the true fingerprint than the measurement (worst ratio 0.896 over 20 sessions).

3. Soundness, conciseness, continuity — python3 experiments/compliance.py (a few minutes)

  • Correlation between abstract and concrete distance ρ = 0.98; realised Lipschitz constant L = 1.00; observational-soundness gap 1.1 × 10⁻¹⁶.

  • Compression: the fingerprint is 81,049 times smaller than the channel at n = 6; parameter-regression R² = 0.96; continuity ratio 2.0.

4. Seed sweep — make audit (about 30 minutes)

  • Sample-complexity exponent β = 0.33 ± 0.05 at 95% confidence (median criterion 0.42 ± 0.08): as |X| grows 6.7 times, required rounds grow 1.6–2.0 times.

  • Convergence slope −0.500 ± 0.006, matching the T⁻¹/² rate.

  • Warm start after drift: 1,539 ± 167 fresh rounds versus 3,469 for a cold start (about 2.3 times) at 95% confidence; 54 versus 1,480 (about 27 times) in the typical case.

  • Reconstruction error of SDP concretisation: 0.173 ± 0.016 (T = 200) to 0.025 ± 0.001 (T = 12,800).

  • Information retention about noise strength: R_info = 0.99 ± 0.01.

5. Scaled evaluation to n = 8 — make scale (about 20 minutes)

  • Exact channel pairs up to n = 8: L = 1.000 at every size; correlation between 0.83 and 0.92.

  • Required rounds range from 3,153 to 8,290 as the interface grows 38 times (|X| = 8 to 304); pooled β = 0.28 at 95% confidence (median criterion 0.48).

6. Stress tests — make limits (about 30 minutes)

  • Detectability follows the variance law: exponent −2.06 (theory −2); 5.4 × 10⁴ rounds at depth 100 and 9.0 × 10⁵ at depth 25; a single hardware-rate layer extrapolates to about 7 × 10⁸ rounds.

  • Calibrated readout error (4.15 × 10⁻³) induces a bias floor near 10⁻².

  • The standard two-qubit probe suite has a 176-dimensional blind subspace: an explicit channel pair has ‖ΔF‖∞ < 10⁻¹⁴ but diamond distance 0.0396. Informationally complete probes remove this blindness.

  • Warm starting pays 2.0 times at drift 0.025 and 1.4 times at drift 0.10, crosses over near drift 0.16, and costs more beyond that (0.3 times at drift 0.7).

7. Supporting checks — make supporting (about 5 minutes)

  • Single-round estimator unbiased: maximum bias 0.039 over 162 cases at T = 2 × 10⁴.

  • Shared snapshots: for m ∈ {2, 4, 8, 12} observables, T ∈ {800, 800, 800, 1600} rounds at 95% confidence (median {200, 400, 400, 400}).

  • Expert signatures confirmed for depolarising, phase damping, and amplitude damping (two label-flipped and two secondary pairs flagged for amplitude damping).

  • Fingerprints separate channel families 876 times; cross-platform distances are 0.008–0.012, at least 11 times their bootstrap uncertainty.

8. Figures and numbers — make all-figures, make numbers

These commands regenerate every figure and the LaTeX macros read by the paper. Full reproduction takes about two hours on one laptop core.

Supplementary Sweeps (Not Reported in the Paper)

make rq1 through make rq4 (about 40 minutes total) run harmonised sweeps over 12 configurations — three channel families and hardware-informed IBM Boston and Quantinuum H2 profiles, each under Qiskit- and Cirq-style variants — for n = 1–8. These sweeps use the median criterion.

  • RQ1: correlation 0.97 (n = 1) to 0.83 (n = 8); realised ratio 1.0 for amplitude and phase damping, and 2/3 for depolarising noise.

  • RQ2: |X| = 228 at n = 8 against 16⁸ ≈ 4.3 × 10⁹ channel parameters (1.88 × 10⁷ times compression); per-qubit recovery R² ≈ 0.91–0.92 for n ≥ 2; off-qubit leakage R² < 0.

  • RQ3: β = 0.71 ± 0.08.

  • RQ4: convergence slopes −0.49 to −0.53; warm-start ratios 31 times to 418 times for small drift.

Scope and Limitations

All experiments are simulations. Hardware profiles are noise models built from published calibration data, not live connections to quantum devices. The Galois law is established for Pauli channels on the single-qubit interface; for other channels it holds only trivially.

Files

worked_example.ipynb

Files (9.4 MB)

Name Size
md5:e2aef72c9f89434fe37d938b0702ba4d
9.3 MB Preview Download
md5:9aaa200d2f106533da7785a6a3096b65
68.3 kB Preview Download
md5:c0f1e5849455b1036f288ecae870ff01
21.3 kB Preview Download