---
slug: 2026-arxiv-2211.16337-quantum-feature-maps-graph-machine-learning
title: Quantum Feature Maps for Graph Machine Learning on a Neutral Atom Quantum Processor
source: arXiv
link: https://arxiv.org/abs/2211.16337
category: app
triage_date: '2026-06-07'
triager: port_until_julien_parallel
bootstrap_date: '2026-06-07'
---

# Dossier: Quantum Feature Maps for Graph Machine Learning on a Neutral Atom Quantum Processor

| Field        | Value |
|--------------|-------|
| Slug         | `2026-arxiv-2211.16337-quantum-feature-maps-graph-machine-learning` |
| Source       | arXiv |
| Link         | [https://arxiv.org/abs/2211.16337](https://arxiv.org/abs/2211.16337) |
| Category     | app |
| Triaged      | 2026-06-07 by port_until_julien_parallel |
| Bootstrapped | 2026-06-07 |

## Sourcing (Romain)

### Domain classification

Primary: **machine learning** (quantum machine learning, specifically graph-kernel methods). Secondary: **biology** (cheminformatics; the headline experiment is a toxicity-screening binary classification on the PTC-FM molecular dataset). The computational contribution is a quantum feature map, the Quantum Evolution Kernel (QEK) for graph-structured data; the biochemistry dataset is the application vehicle rather than the focus. This matches the existing `pasqal/quantum_ml` framing in `vendor-app/edges.json`.

### Expert persona for Julien

A quantum machine learning practitioner with hands-on neutral-atom analog experience. Should be fluent in quantum kernel methods (constructing an SVM kernel from divergences between measurement distributions) and in analog Rydberg dynamics: the driven Ising / van der Waals Hamiltonian, control of Rabi frequency $\Omega$ and detuning $\delta$, the unit-disk-graph embedding via Rydberg blockade, and the $C_6/r^6$ interaction. Needs calibrated intuition for shot budgets and detection-error modelling, so as to separate the genuinely hardware-specific knobs (register geometry, blockade radius, $\Omega$ and $\delta$, pulse duration $T$, shot count) from the portable algorithmic skeleton (a Jensen-Shannon-divergence kernel fed to an SVM).

### Related prior work (brief bibliography)

- Henry, Thabet, Dalyac, Henriet, *Quantum evolution kernel: Machine learning on graphs with programmable arrays of qubits*, Phys. Rev. A 104, 032416 (2021), arXiv:2107.03247. The foundational QEK method that this paper implements experimentally. Already in the corpus as `2026-arxiv-2107.03247-quantum-evolution-kernel-machine-learning-graphs`.
- *Harnessing a 256-qubit Neutral Atom Simulator for Graph Classification*, arXiv:2605.04737. The same QEK-style graph-classification primitive on a different neutral-atom vendor (QuEra Aquila); the cleanest cross-vendor portability comparison. Already in the corpus as `2026-arxiv-2605.04737-aquila-graph-classification`.
- Havlíček, Córcoles, Temme, Harrow, Kandala, Chow, Gambetta, *Supervised learning with quantum-enhanced feature spaces*, Nature 567, 209 (2019), DOI:10.1038/s41586-019-0980-2. The same quantum-kernel / feature-map idea realised on a superconducting processor; a cross-modality comparison point for the primitive.
- Huang, Broughton, Mohseni, Babbush, Boixo, Neven, McClean, *Power of data in quantum machine learning*, Nat. Commun. 12, 2631 (2021), DOI:10.1038/s41467-021-22539-9. Source of the geometric-difference metric $g_{12}$ used in Sec. on geometric tests, and of the argument that classical ML can sometimes learn from quantum-hard data. Central to any portability or advantage claim.
- Schuld, Killoran, *Quantum machine learning in feature Hilbert spaces*, Phys. Rev. Lett. 122, 040504 (2019), DOI:10.1103/PhysRevLett.122.040504. Theory framing of quantum feature maps as kernels.
- Xu, Hu, Leskovec, Jegelka, *How powerful are graph neural networks?*, ICLR 2019, arXiv:1810.00826. The classical GNN / Weisfeiler-Lehman expressiveness benchmark against which the paper positions its sensitivity to global graph structure.

### Difficulty estimate

**medium**. The full source is in the folder (arXiv LaTeX plus supplemental text), so there is no paywall barrier and the hardware fingerprint is explicitly stated: $^{87}$Rb with $\ket{1}=\ket{60S_{1/2}}$, up to 32 qubits, $C_6/h \simeq 138$ GHz$\cdot\mu$m$^6$, $\Omega/2\pi \approx 1$ MHz, $\delta/2\pi \approx 0.7$ MHz, nearest-neighbour distance $r_{NN}=5.3\,\mu$m, optimal pulse $T \approx 0.66\,\mu$s, around 600 shots per point, detection errors $\varepsilon \approx 3\%$ and $\varepsilon' \approx 8\%$, and an effective dephasing of $2\pi \times 0.06$ MHz. The medium rating reflects that the relevant parameters are spread across the main text and two supplemental appendices (experimental setup, noise model), and that the analog Rydberg primitives (unit-disk-graph embedding, blockade) need care to extract correctly. There is no difficulty reading the paper itself.