Empirical Study of Observable Sets in Multiclass Quantum Classification

Technical Summary: Empirical Study of Observable Sets in Multiclass Quantum Classification

1. Problem Statement

Most existing research in Quantum Machine Learning (QML) focuses on binary classification or uses ensembles of binary classifiers (e.g., One-vs-Rest) to handle multiple classes. However, these ensemble methods scale poorly with the number of classes, increasing training costs and susceptibility to noise and Barren Plateaus. While native multiclass models exist, there is a lack of systematic research into how the choice of observables (the operators used to extract information from the quantum state) affects model performance, trainability, and geometric convergence.

2. Methodology

The authors investigate two primary classification criteria for Parameterized Quantum Circuits (PQCs):

• Approach A (Maximizing Expected Value): Assigning a specific Pauli string observable to each class. The class is determined by the observable with the highest expected value.
• Approach B (Maximizing Fidelity): Defining a reference state (projector) for each class. The class is determined by the highest fidelity between the evolved quantum state and the reference state.

Experimental Setup:

• Observables: They compare sets of Pauli strings (specifically mutually commuting sets with varying locality) against sets of projectors into the computational basis.
• Loss Functions: They utilize Cross-Entropy loss (using a temperature-scaled softmax to handle the bounded range of Pauli expectations) and Fidelity loss.
• Datasets: Experiments were conducted on synthetic datasets (Blobs2, Blobs8, Tetrominoes) and a real-world dataset (Panels—solar panel images processed via a classical CNN for dimensionality reduction).
• Metrics: Performance was measured using the F1 score, while training dynamics were analyzed through the lens of Barren Plateaus (loss variance) and Neural Collapse (intra-class vs. inter-class fidelity).

3. Key Contributions

• Characterization of Observable Locality: A theoretical and empirical link between the locality of observables and the concentration of the loss landscape.
• Geometric Analysis of Neural Collapse: An investigation into whether quantum models exhibit "Neural Collapse"—a phenomenon where class means converge to an Equiangular Tight Frame (ETF).
• Comparative Framework: A systematic comparison of Pauli-based vs. Projector-based multiclass architectures.

4. Results

• Barren Plateaus & Locality:
  • Projectors exhibit a monotonic exponential concentration of loss as locality increases.
  • Pauli strings exhibit a parabolic relationship between loss variance and locality. This is explained by the Lie-algebraic structure of the Dynamical Lie Algebra (DLA), where the variance is driven by the dimension of the specific locality module.
• Neural Collapse:
  • Both commuting Pauli models and Projector models successfully reach Neural Collapse (high intra-class fidelity, low inter-class fidelity).
  • Non-commuting Pauli models fail to reach Neural Collapse because the non-zero overlap of their eigenspaces prevents the formation of a perfect ETF structure.
  • Curse of Dimensionality: As the number of qubits increases, the Hilbert space becomes sparser, making it harder to achieve high fidelity and causing class means to become naturally orthogonal.
• Model Performance:
  • In under-parameterized regimes (low re-uploading layers), Projector-based models showed a slight performance advantage in real-world and imbalanced datasets, likely due to the inherent imposition of the ETF structure.
  • As model complexity (re-uploading layers) increases, the performance gap between Pauli and Projector models vanishes.
• Softmax Temperature: The authors identified that the temperature parameter in the softmax function acts as a critical regularizer that controls the trade-off between training accuracy and test-set generalization.

5. Significance

This work provides a roadmap for designing more efficient multiclass QML models. It demonstrates that the choice of measurement is not merely a technical detail but a fundamental design decision that dictates:

1. Trainability: How easily the model avoids Barren Plateaus.
2. Geometric Structure: Whether the model can achieve the optimal feature representations characterized by Neural Collapse.
3. Robustness: How the model handles imbalanced data and high-dimensional feature spaces.

The findings suggest that for NISQ-era devices, selecting observables that enforce specific geometric structures (like ETFs) may provide a beneficial inductive bias for under-parameterized quantum classifiers.