Part VIII. Applications and Strategy · Chapter 69
Quantum Machine Learning and Benchmark Discipline
Quantum machine learning papers arrive weekly; dependable results do not. This chapter turns QML evaluation into a checklist — encoding cost, tuned baselines, trainability, noise — so you can debug a claim in an hour instead of believing it for a quarter.
In this chapter 9 sections
Freeze the data split and feature-access model, compare against tuned classical models with the same information, report encoding, shots, optimization, and uncertainty, and reserve the quantum claim for held-out results that survive ablation and full-cost accounting.
A learning benchmark is first an information contract. If one model receives a compressed state prepared by an oracle while the baseline receives raw observations under a smaller tuning budget, the comparison is about access and effort, not model class. The data split, feature computation, and hyperparameter budget must be frozen before training results become evidence.
Define the learner’s information before its circuit
The benchmark card records dataset provenance and hashes, split policy, label balance, feature access, encoding circuit, shot budget, optimizer evaluations, classical comparators, seed set, ablations, and held-out metric. It reports every registered seed and a paired uncertainty interval rather than choosing the best run from either model.
- Split data before feature or circuit selection.
- Tune classical and quantum candidates under the same search budget.
- Report ablations, seed dispersion, wall time, and prediction uncertainty.
Benchmark card for data, model, and estimator
The QML review maps model and data-access regimes; the challenges perspective identifies trainability, generalization, and benchmarking risks. VQE and QAOA are cited as original variational precedents, not learning-advantage results. The official SDK references establish reproducible interfaces at the dated cutoff, while the benchmarking source supplies comparator discipline.
A tiny classifier fails the ablation test
Five paired held-out differences in accuracy are 0.02, -0.01, 0.01, 0.00, and -0.02. Their mean is zero. Selecting the first seed would produce a positive headline; retaining the registered seeds produces an inconclusive result. The ablation also matters: if replacing the quantum feature map does not reduce held-out accuracy, the claimed mechanism lacks support even when the final classifier performs well.
Artifact contract. A dated dataset/model/access ledger with executable paired bootstrap and ablation checks. Train/test hashes differ, the quantum and classical models receive the same features, all seeds are recorded, and the advantage flag remains false when the confidence interval crosses zero.
Executable reference fixture
from math import sqrt
REQUIRED = {"scenario": str, "source_ids": list, "train_hash": str, "test_hash": str,
"feature_contract": str, "seeds": list, "paired_differences": list,
"difference_unit": str, "ablation_delta": float}
def analyze(card, ablation_tolerance):
errors = [name for name, kind in REQUIRED.items()
if name not in card or type(card[name]) is not kind]
values = card.get("paired_differences", [])
if not values or any(type(value) not in (int, float) for value in values):
errors.append("paired_difference_type")
if card.get("difference_unit") != "accuracy points" or card.get("train_hash") == card.get("test_hash"):
errors.append("unit_or_split_hash")
if not card.get("source_ids") or not card.get("seeds") or any(type(seed) is not int for seed in card.get("seeds", [])):
errors.append("sources_or_seeds")
if errors:
return {"decision": "invalid", "errors": sorted(set(errors)), "interval": None}
mean = sum(values) / len(values)
variance = sum((value - mean) ** 2 for value in values) / (len(values) - 1)
radius = 1.96 * sqrt(variance / len(values))
interval = (mean - radius, mean + radius)
decision = "inconclusive" if interval[0] <= 0 <= interval[1] else "directional"
if abs(card["ablation_delta"]) > ablation_tolerance:
decision = "ablation-sensitive"
return {"decision": decision, "errors": [], "interval": interval, "mean": mean}
card = {"scenario": "equal-feature held-out classifier", "source_ids": ["benchmarking-2025"],
"train_hash": "train:42c1", "test_hash": "test:88af", "feature_contract": "same 12 features",
"seeds": [11, 19, 23, 31, 47], "paired_differences": [0.02, -0.01, 0.01, 0.00, -0.02],
"difference_unit": "accuracy points", "ablation_delta": 0.005}
base_result = analyze(card, 0.01)
bad = {key: value for key, value in card.items() if key != "feature_contract"}
bad_result = analyze(bad, 0.01)
shifted = {**card, "paired_differences": [value + 0.04 for value in card["paired_differences"]]}
shifted_result = analyze(shifted, 0.01)
assert base_result["decision"] == "inconclusive" and base_result["interval"][0] < 0 < base_result["interval"][1]
assert bad_result["decision"] == "invalid" and "feature_contract" in bad_result["errors"]
assert shifted_result["decision"] == "directional" and shifted_result["interval"][0] > 0
print(f"PASS: 69 QML card interval={base_result['interval']} invalid={bad_result['errors']} shifted={shifted_result['decision']}")| Seed | Quantum accuracy | Classical accuracy | Paired difference |
|---|---|---|---|
| 11 | 0.84 | 0.82 | +0.02 |
| 17 | 0.81 | 0.82 | -0.01 |
| 23 | 0.83 | 0.82 | +0.01 |
| 29 | 0.82 | 0.82 | 0.00 |
| 31 | 0.80 | 0.82 | -0.02 |
Exact validation command: python3 tools/validate_briefs.py --briefs data/editorial_briefs_64_87.json --from 64 --through 87 --check-rewritten-sources --execute-artifacts
Confidence intervals matter more than a best run
Call this a useful experiment with no observed advantage. Preserve the split hashes and use the next run to test one pre-specified mechanism, such as sensitivity to encoding depth or shot noise. Increasing the number of model variants while inspecting the same held-out set spends the test set and requires a new dataset or a nested selection protocol.
Reproduce the comparison under equal access
Prompt. Write and evaluate a benchmark card for a binary quantum-classification claim.
Deliverable. Frozen split hashes, feature-access statement, quantum and classical hyperparameter budgets, five seeds, ablations, and a paired confidence interval.
Pass condition. No test observation enters tuning, both learners receive equal information, and the claim is 'inconclusive' whenever the interval includes zero.
Model answer: useful experiment, unsupported advantage
Format. Reference benchmark card whose apparent best-seed gain vanishes under paired uncertainty.
The reference card records equal features and five seeds. The paired mean equals zero and the range crosses zero, so the decision is inconclusive. The model answer refuses to average independently selected best seeds, refuses to ignore encoding cost, and distinguishes predictive performance from evidence that a quantum feature caused that performance. The next gate is a frozen ablation with a new held-out sample.
Verification. The fixture reproduces the interval and confirms that removing the quantum feature map does not worsen the held-out result beyond the declared tolerance.
Companion work
Artifacts for this chapter
These entries resolve to checked-in local source. Commands are reproduced exactly from the chapter manifest, and source-embedded fixtures are exported as direct downloads.
evidence dossier
Leakage-resistant QML benchmark card
Reproduce or test
python3 tools/validate_briefs.py --briefs data/editorial_briefs_64_87.json --from 64 --through 87 --check-rewritten-sources --execute-artifacts
Provenance
Sources and review
- Jacob Biamonte et al.. Quantum machine learning. Nature. 2017peer-reviewed review
- Marco Cerezo et al.. Challenges and opportunities in quantum machine learning. Nature Computational Science. 2022peer-reviewed perspective
- Edward Farhi, Jeffrey Goldstone, and Sam Gutmann. A Quantum Approximate Optimization Algorithm. arXiv. 2014primary preprint
- Alberto Peruzzo et al.. A variational eigenvalue solver on a photonic quantum processor. Nature Communications. 2014primary paper
- OpenQASM Technical Steering Committee. OpenQASM 3 specification. Linux Foundation Joint Development Foundation. 2026official technical specification
- IBM Quantum. Qiskit documentation. IBM. 2026official documentation
- Google Quantum AI. Cirq documentation. Google. 2026official documentation
- Timothy Proctor et al.. Benchmarking quantum computers. Nature Reviews Physics. 2025peer-reviewed perspective
The load-bearing claims in the chapter are mapped inline to this registered source set. A citation supports only the bounded claim beside it.