Adaptive Non-Local Observable on Quantum Neural Networks

Hsin-Yi Lin, H. Eric Tseng, Samuel Yen-Chi Chen, Shinjae Yoo2025 · arXiv

TL;DR

This paper introduces a novel "Adaptive Non-Local Observable" (ANO) framework for Quantum Neural Networks (QNNs). By making the measurement observable itself trainable and allowing it to act on multiple qubits, ANO significantly enhances QNN expressivity and performance compared to traditional QNNs that use fixed, single-qubit observables. This approach offers a more powerful and resource-efficient path for quantum machine learning.

What Problem Does This Paper Address?

Quantum Machine Learning (QML) aims to leverage quantum phenomena for computational advantages in machine learning. Variational Quantum Circuits (VQCs), a core component of many QML models like Quantum Neural Networks (QNNs), typically rely on a fixed set of measurement observables, usually built from simple Pauli operators acting on individual qubits. This limitation restricts the "function space" of VQCs, hindering their ability to capture complex data distributions and leading to lower model capacity. The paper addresses this by proposing a method to make these measurement observables more flexible and powerful, thereby increasing the expressivity of QNNs.

Methods

The core innovation is the Adaptive Non-Local Observable (ANO) framework. Instead of fixed Pauli observables, ANO uses trainable, multi-qubit Hermitian matrices as observables. These observables are optimized jointly with the VQC's rotation parameters. This approach is inspired by the Heisenberg picture in quantum mechanics, where observables evolve while states remain static.

Key aspects of the method include:

  • Dynamical Hermitian Observables: The paper introduces the concept of dynamical Hermitian observables with evolving parameters. This allows the measurement strategy to adapt during training, mirroring how quantum states evolve in the Schrödinger picture.
  • Non-Local Measurements: Unlike standard VQCs that measure individual qubits, ANO allows observables to act on multiple qubits simultaneously. This enables the capture of richer, non-local quantum correlations.
  • Two Measurement Schemes:
    1. Sliding k-Local Measurements: This scheme uses a sliding window to measure overlapping groups of k qubits, forming a sequence of local measurements that can build up a global representation.
  • Pairwise Combinatorial Measurement: This scheme selects a subset of qubits and measures all possible pairs within that subset using 2-local ANO observables. This is designed to be parameter-efficient while still capturing feature interactions.
  • Theoretical Framework: The paper establishes a theoretical connection between VQCs and ANO through unitary similarity, showing that conventional VQCs are a special case of ANO operating within a single equivalence class of observables.

Datasets

The paper evaluates its proposed framework on two standard machine learning datasets:

  • Banknote Authentication Dataset: A binary classification task with 4-dimensional features.
  • MNIST Dataset: A subset of the MNIST dataset (resized to 4x4 images) for a 10-class classification task.

Key Insights and Ideas

  • Observables are as Important as Gates: The paper highlights that the measurement observable plays a crucial role in QNN expressivity, not just the quantum gates. Making observables trainable is a powerful lever for improving QNN performance.
  • Heisenberg Picture Inspiration: Drawing inspiration from the Heisenberg picture provides a new perspective on VQCs, framing them as a trajectory in the observable space rather than just state evolution.
  • VQCs as a Special Case: The theoretical analysis reveals that conventional VQCs, with their fixed observables, are a restricted form of the more general ANO framework. ANO can potentially eliminate the need for variational rotations altogether by directly optimizing the observable.
  • Trade-off between k-locality and Rotations: For smaller k-local observables, variational rotations are crucial for enhancing feature mixing and model expressivity. However, as k increases (more non-local interactions), the benefit of rotations diminishes, as the observable itself captures more complex correlations.
  • Parameter Efficiency: The pairwise combinatorial measurement scheme demonstrates that non-local interactions can be captured efficiently with fewer parameters than using higher k-local observables or extensive variational rotations.

Key Results

  • Banknote Authentication: ANO models, particularly 2-local ANO with variational rotations, significantly outperformed the fixed Pauli baseline. The performance improved with increasing k-locality, with 2-local and 3-local models showing fast convergence.
  • MNIST:
    • Sliding k-local ANO showed consistent accuracy improvements with increasing k, though with diminishing returns.
  • The pairwise combinatorial measurement scheme achieved the highest accuracy (82.0% with 3,130 parameters for 16 qubits) and was more parameter-efficient than higher k-local sliding schemes. Notably, this scheme achieved competitive results even without variational rotations.
  • Overall: The proposed ANO framework consistently outperformed conventional VQCs across both datasets, demonstrating enhanced model expressivity and, in some configurations, improved parameter efficiency.

Limitations and Future Work

  • Computational Cost: While ANO offers advantages, the number of parameters can grow rapidly with k, especially for full ANO (k=n). The paper acknowledges this and investigates smaller k-local ANOs with rotations as a compromise.
  • Overfitting: For the Banknote dataset, the 3-local ANO showed slight signs of overfitting, suggesting that careful parameter tuning or regularization might be necessary for certain tasks.
  • Future Directions: The paper suggests exploring smaller k-local ANOs combined with variational rotations as a parameter-efficient alternative to full ANO. It also hints at the potential for ANO to eliminate variational rotation gates entirely, which could lead to more streamlined and efficient quantum learning circuits. The investigation into the trade-off between k-locality and rotation gates is also a promising avenue for future research.

Summary updated

Abstract

Conventional Variational Quantum Circuits (VQCs) for Quantum Machine Learning typically rely on a fixed Hermitian observable, often built from Pauli operators. Inspired by the Heisenberg picture, we propose an adaptive non-local measurement framework that substantially increases the model complexity of the quantum circuits. Our introduction of dynamical Hermitian observables with evolving parameters shows that optimizing VQC rotations corresponds to tracing a trajectory in the observable space. This viewpoint reveals that standard VQCs are merely a special case of the Heisenberg representation. Furthermore, we show that properly incorporating variational rotations with non-local observables enhances qubit interaction and information mixture, admitting flexible circuit designs. Two non-local measurement schemes are introduced, and numerical simulations on classification tasks confirm that our approach outperforms conventional VQCs, yielding a more powerful and resource-efficient approach as a Quantum Neural Network.
BibTeX
@misc{lin2025adaptive,
  author = {Hsin-Yi Lin and H. Eric Tseng and Samuel Yen-Chi Chen and Shinjae Yoo},
  title = {{Adaptive Non-Local Observable on Quantum Neural Networks}},
  year = {2025},
  eprint = {2504.13414},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.13414},
}
Figure