| name | quantum-graph-machine-learning-models |
| description | Geometric Quantum Machine Learning (GQML) design toolbox for graph problems — comprehensive characterization of constituents for n-node-graph → n-qubit-state encoding; enables hybrid classical-quantum integration, generalizes known GQML models (extending expressivity at near-zero cost), and supports straightforward classical pre-training; validated numerically. |
| version | 1.0.0 |
| last_updated | 2026-07-03T00:00:00.000Z |
| arxiv_id | 2607.00698 |
| arxiv_url | https://arxiv.org/abs/2607.00698 |
| authors | ["Frédéric Sauvage","Pranav Kalidindi","Frederic Rapp","Martín Larocca"] |
| tags | ["geometric-quantum-machine-learning","GQML","graph-neural-network","equivariant-quantum","quantum-graph-models","quantum-machine-learning","classical-pre-training","hybrid-QML"] |
| category | quantum-machine-learning |
Quantum Machine Learning Models for Graphs
Source: arXiv:2607.00698 (submitted 1 Jul 2026) — Sauvage et al. (Los Alamos, LA-UR-26-23295).
When to Use
Trigger this skill when working on:
- Geometric Quantum Machine Learning (GQML) for graph-structured problems (subgraph isomorphism, graph classification, max-cut, graph isomorphism)
- Designing equivariant parameterized quantum circuits for graph inputs
- Integrating classical GNN features into quantum circuits (hybrid pipelines)
- Needing a classical pre-training strategy for quantum graph models
- Extending or generalizing existing GQML proposals to boost expressivity cheaply
Core Contribution
Provides a unifying design toolbox for GQML models on graphs, specifically the regime where an n-node graph is encoded into an n-qubit state. The toolbox comprehensively characterizes the constituents (encoding, ansatz, measurement, symmetry handling) and shows how to:
- Naturally integrate with classical models — classical GNN embeddings feed quantum layers and vice versa.
- Generalize known GQML models — sometimes extending their expressivity at virtually no additional cost.
- Pre-train classically — straightforward classical pre-training strategies seed quantum weights, reducing quantum training cost / barren-plateau exposure.
Methodology: The GQML-for-Graphs Toolbox
Constituent characterization
A graph quantum model M is decomposed into four constituents, each chosen to respect the graph symmetry group S_n (permutation equivariance):
- Encoding E(x) — maps a graph x (adjacency / node features) to an n-qubit state |ψ(x)⟩. Must be equivariant: relabeling nodes permutes qubits correspondingly.
- Ansatz U(θ) — parameterized circuit. Equivariant ansätze have layers that commute with the permutation representation.
- Measurement {O_k} — observables whose orbits under S_n produce covariant features.
- Readout — classical post-processing (pooling) that aggregates into a graph-level prediction.
Design principles (unifying perspective)
- Symmetry dictates ansatz structure: only circuits in the commutant of the permutation representation can be equivariant. This constrains the circuit family and prevents "symmetry-breaking" parameters.
