[Paper Review] Quantum Graph Neural Networks

Guillaume Verdon, Trevor McCourt+4 more|arXiv (Cornell University)|Sep 26, 2019

Quantum Computing Algorithms and ArchitectureComputer Science40 references60 citations

TL;DR

The paper introduces Quantum Graph Neural Networks (qgnn) and specialized variants (qgrnn, qgcnn, qsgcnn) to process graph-structured quantum data, and demonstrates four applications: learning quantum Hamiltonian dynamics, entanglement generation in quantum networks, unsupervised spectral clustering, and graph isomorphism classification. It provides numerical experiments showing effective learning and quantum sensing advantages on graph-structured tasks.

ABSTRACT

We introduce Quantum Graph Neural Networks (QGNN), a new class of quantum neural network ansatze which are tailored to represent quantum processes which have a graph structure, and are particularly suitable to be executed on distributed quantum systems over a quantum network. Along with this general class of ansatze, we introduce further specialized architectures, namely, Quantum Graph Recurrent Neural Networks (QGRNN) and Quantum Graph Convolutional Neural Networks (QGCNN). We provide four example applications of QGNNs: learning Hamiltonian dynamics of quantum systems, learning how to create multipartite entanglement in a quantum network, unsupervised learning for spectral clustering, and supervised learning for graph isomorphism classification.

Motivation & Objective

• Motivate the need for graph-aware quantum neural network ansatze tailored to distributed quantum systems.
• Propose a general qgnn framework and specialized architectures (qgrnn, qgcnn, qsgcnn) for graph-structured quantum data.
• Demonstrate four applications: learning Hamiltonian dynamics, creating multipartite entanglement in quantum networks, unsupervised spectral clustering, and supervised graph isomorphism classification.

Proposed method

• Define a general QGNN ansatz as a parameterized quantum circuit on a graph with Q Hamiltonians applied in sequence and P repetitions.
• Impose constraints for trainability, such as commuting terms within a given H_q and spatial or temporal parameter tying to create qgrnn and qgcnn variants.
• Develop qgrnn as temporally tied parameters to simulate effective Hamiltonian dynamics on graphs (Trotterized evolution under H_eff).
• Create qgcnn with global (per-graph) parameters to enforce permutation invariance, aligning with Quantum Alternating Operator Ansatze (QAOA-like).
• Extend to Quantum Spectral Graph Convolutional Networks (qsgcnn) using continuous-variable ideas and a four-Hamiltonian layer sequence to mimic spectral graph convolutions with nonlinearity via an anharmonic potential.
• Provide concrete implementations and training setups for each variant, including loss functions and evaluation methods.

Experimental results

Research questions

• RQ1 Can qgrnn learn effective graph-structured quantum Hamiltonian dynamics from data?
• RQ2 Can qgcnn-based architectures learn to generate/identify multipartite entanglement in quantum networks without full global structure knowledge?
• RQ3 Can qsgcnn perform unsupervised spectral clustering on graph-structured quantum data, including low-qubit precision scenarios?
• RQ4 Can the qsgcnn approach classify graph isomorphism with high accuracy using energy-statistic measurements?

Key findings

• Quantum Graph Recurrent Neural Networks (qgrnn) successfully learn target Ising Hamiltonian dynamics on a graph, recovering topology and parameters from time-evolved states.
• A Quantum Graph Convolutional Neural Network (qgcnn) can prepare and verify GHZ states on a quantum sensor network, achieving a phase-kickback enhancement consistent with multipartite entanglement.
• The Quantum Spectral Graph Convolutional Network (qsgcnn) recovers a quantum-coherent analogue of spectral graph convolutions, enabling unsupervised spectral clustering on graphs and showing viable performance at varying qubit precisions.
• Graph Isomorphism Classification experiments with qsgcnn achieve high accuracy, including 100% test/validation accuracy for 15-node graphs and strong performance for 6-node graphs at multiple sample counts.
• The paper reports observable quantum advantages such as a 7x boost in Rabi oscillation frequency on a 7-node network during a GHZ-state phase kickback test, illustrating Heisenberg-limited sensitivity in the learned network.
• Numerical experiments leverage a Cirq-TensorFlow interface to demonstrate qgnn capabilities across dynamics learning, sensing, clustering, and isomorphism tasks.