Benchmarking Quantum Convolutional Neural Networks and their classical surrogates on long-range strongly correlated models

Introduction

Understanding phase transitions in strongly correlated quantum systems remains a central challenge in condensed‑matter physics. Recent advances in quantum machine learning have introduced quantum convolutional neural networks (QCNNs) as a promising tool for classifying quantum phases directly from many‑body wavefunctions. The authors of Benchmarking Quantum Convolutional Neural Networks and their classical surrogates on long‑range strongly correlated models systematically assess the performance of QCNNs and their classical surrogates—models that approximate the quantum circuit output using weighted sums of Pauli‑string expectation values—on two representative one‑dimensional systems with beyond‑nearest‑neighbour interactions.

Testbeds: Spin and Fermionic Chains

The study focuses on two benchmark Hamiltonians:

1. Anisotropic Heisenberg spin chain – a spin‑½ model with nearest‑ and next‑nearest‑neighbour exchange couplings that exhibits rich magnetic ordering and quantum critical points.
2. Extended $t$‑$t'$ Fermi‑Hubbard model – a fermionic lattice model that adds a next‑nearest‑neighbour hopping term $t'$ to the standard Hubbard Hamiltonian, generating competing charge‑density‑wave and superconducting tendencies.

Both models are simulated on modest lattice sizes, and ground‑state wavefunctions are generated across a grid of coupling parameters. These labeled datasets serve as the training and validation sets for the QCNN classifiers.

QCNN Architecture and Hybrid Training

The QCNN employed follows the standard hierarchical design: alternating layers of local two‑qubit unitaries (parameterized by rotation angles) and pooling operations that discard qubits while preserving relevant entanglement patterns. After the final layer, a single measurement yields a binary label indicating the phase.

A key contribution of the paper is the hybrid training scheme based on Pauli backpropagation. Instead of back‑propagating through a full quantum circuit, the authors express the final measurement operator as a linear combination of Pauli strings. The coefficients are obtained analytically from the circuit parameters, allowing the loss gradient to be computed from expectation values of these strings on the input states. This approach enables classical evaluation of the QCNN output without running the quantum circuit, at the cost of handling a potentially large Pauli expansion.

Truncation Strategies and Performance Trade‑offs

To keep the Pauli expansion tractable, the authors introduce a truncation threshold that discards strings whose coefficients fall below a chosen magnitude. They explore how varying this threshold influences two competing metrics:

• Approximation accuracy – measured by the classification fidelity on a held‑out test set.
• Evaluation cost – quantified by the number of retained Pauli strings and the corresponding number of expectation‑value measurements.

Heisenberg Chain

For the spin model, aggressive truncation (thresholds as high as $10^{-2}$) still yields robust convergence: the classifier retains only a few hundred Pauli strings while achieving near‑optimal accuracy. This reflects the relatively simple entanglement structure of the ground states and the locality of the QCNN filters.

Extended Hubbard Model

In contrast, the fermionic system demands looser thresholds (down to $10^{-4}$) to maintain comparable accuracy. The required Pauli expansion grows sharply, reaching several thousand strings, which markedly increases measurement overhead. The authors attribute this to the more intricate fermionic correlations and the presence of longer‑range hopping terms that spread information across the lattice.

Practical Guidance for Hybrid Quantum‑Classical Training

The benchmarking results lead to several actionable recommendations for researchers seeking to deploy QCNNs on near‑term quantum hardware:

1. Model‑dependent truncation – Evaluate the entanglement complexity of the target system before fixing a truncation level; spin models often tolerate aggressive pruning, whereas fermionic or long‑range models may not.
2. Hybrid evaluation – Use Pauli backpropagation to pre‑screen circuit parameters classically, reserving full quantum execution for final validation or for regimes where the Pauli expansion becomes prohibitive.
3. Resource budgeting – Estimate the measurement budget by counting retained Pauli strings; this provides a clear metric for comparing different truncation regimes.
4. Iterative refinement – Start with a coarse truncation to locate promising parameter regions, then gradually lower the threshold to fine‑tune performance.

Conclusion

By juxtaposing QCNNs with their classical Pauli‑backpropagation surrogates across two distinct long‑range correlated models, the paper delineates where hybrid quantum‑classical training is feasible and where it becomes resource‑intensive. The findings underscore that the viability of QCNN‑based phase classification hinges on the underlying physics of the system and on judicious selection of truncation parameters. These insights will help guide future experimental implementations of quantum machine‑learning classifiers on noisy intermediate‑scale quantum devices.