Problem-Dependent Power of Quantum Neural Networks on Multiclass Classification.

Area of Science:

• Quantum computing
• Machine learning
• Computational complexity

Background:

• Quantum neural networks (QNNs) are increasingly used for physical systems analysis.
• Understanding the specific advantages and limitations of QNNs, particularly quantum classifiers (QCs), remains an active research area.
• Classical simulation capabilities and quantum memory effects influence QC performance compared to classical models.

Purpose of the Study:

• To systematically investigate the problem-dependent power of quantum classifiers (QCs) in multiclass classification tasks.
• To analyze the expected risk, balancing training loss and generalization error, to understand QC performance.
• To identify key factors influencing QC efficacy and compare them with classical neural networks.

Main Methods:

• Analysis of expected risk, a metric combining training loss and generalization error.
• Investigation of the risk curve behavior in QCs.
• Exploration of connections between optimal QCs, the Helstrom bound, and equiangular tight frames.
• Development of a method to estimate optimal hyperparameters using QC loss dynamics.

Main Results:

• Training loss was found to be the dominant factor in QC power, rather than generalization ability.
• Quantum classifiers exhibit a U-shaped risk curve, distinct from the double-descent curve seen in deep neural classifiers.
• An intrinsic link was established between optimal QCs, the Helstrom bound, and equiangular tight frames.
• A novel method for estimating optimal hyperparameters based on loss dynamics was proposed and validated.

Conclusions:

• The study clarifies the problem-dependent power of QNNs and QCs.
• The proposed method effectively estimates optimal hyperparameters for QCs, minimizing risk.
• Numerical results confirm QC superiority over MLPs on parity tasks and highlight limitations compared to CNNs on image tasks.
• This research provides a practical framework for evaluating the potential of QNNs.