Spike: QuantumSVM feature_map=auto (Qmes-style)

[Superagentic-AI]

Why

QuantumSVM accuracy depends on the encoding circuit. Exhaustive per-dataset kernel evaluation of every candidate is expensive. Qmes (arXiv:2609.04652) shows classical complexity meta-features can rank encodings with no quantum evaluation at inference time.

This spike adds a practical feature_map="auto" path for SuperQuantX QuantumSVM: recommend a name from our existing registry, optionally via Qmes, otherwise via a local heuristic.

What changed

• New registry: utils/feature_map_registry.py (ZZFeatureMap, PauliFeatureMap, AmplitudeMap, AngleEncoding, ZFeatureMap) plus Qmes pool mapping (unit, SRx, RY, HERx, RY_CX, ZFM, HD).
• New selector: utils/feature_map_auto.py with modes auto / qmes / heuristic (classical only; no quantum eval at selection time).
• QuantumSVM accepts feature_map="auto" or None, stores selected_feature_map_ and feature_map_recommendation_. Manual strings/objects still work.
• Backends (simulator, PennyLane, Qiskit, Cirq) recognize AngleEncoding and ZFeatureMap.
• Optional extra: superquantx[qmes] (git install of Qmes + problexity/pandas).
• Design note: docs/design/qmes-feature-map-auto.md.
• Unit tests: tests/unit/test_feature_map_auto.py (mocked / no hardware).

How to try

from sklearn.datasets import make_classification
from superquantx.algorithms import QuantumSVM

X, y = make_classification(n_samples=40, n_features=4, n_informative=3, random_state=0)
qsvm = QuantumSVM(backend="simulator", feature_map="auto", shots=100)
qsvm.fit(X, y)
print(qsvm.selected_feature_map_)
print(qsvm.feature_map_recommendation_)

Optional Qmes path:

pip install "superquantx[qmes]"
# or: pip install "git+https://github.com/tungduy1704/Qmes.git"

qsvm = QuantumSVM(backend="simulator", feature_map="auto", auto_feature_map_mode="qmes")

Known gaps

• Not a full Qmes port: no Qsun encodings, no 4-qubit PCA evaluator loop, no shipped SQX-trained OvO recommender.
• Mapping is approximate (closest registry analogue, not identical gate schedules).
• Heuristic quality is unvalidated against Qmes regret metrics.
• AmplitudeMap remains partial on several backends.
• Regression auto-selection not wired (QSVM is classification-focused).

References

• Paper: https://arxiv.org/abs/2609.04652
• Qmes repo: https://github.com/tungduy1704/Qmes
• Design note in-tree: docs/design/qmes-feature-map-auto.md