Geometric Prototype Learning in Quantum Hilbert Space with Matrix Product States

Abstract: Quantum probability provides a novel framework for formulating machine-learning (ML) problems in Hilbert space. We introduce a prototype-based learning scheme where class representatives are encoded as generative matrix product states (MPS). Because these prototypes reside in the same Hilbert space as quantum-encoded data samples, various ML tasks such as classification and clustering can be performed through geometric measures of quantum states. This approach lifts prototype learning from classical feature space to quantum Hilbert space. Benchmarks on Fashion-MNIST and a real-world electrocardiogram dataset demonstrate that our method outperforms classical prototype approaches while remaining competitive with standard black-box neural networks. We also identify an ``attraction'' effect induced by the quantum-probabilistic prototypes and introduce a dimensionality-reduction scheme based on prototype distances. Our results establish quantum states as an explainable framework for prototype learning, opening new directions for designing ML algorithms in quantum Hilbert space.

• The paper introduces Quantum-Probabilistic Prototype Learning (QPPL) which maps data into Hilbert space using Matrix Product States for transparent classification.
• It demonstrates superior classification, anomaly detection, and clustering performance through the use of negative logarithmic fidelity as a geometric distance metric.
• The approach unifies representation, similarity measurement, and decision-making in a single quantum framework, paving the way for future quantum-enhanced machine learning.

Geometric Prototype Learning in Quantum Hilbert Space with Matrix Product States

Introduction and Problem Setting

The paper "Geometric Prototype Learning in Quantum Hilbert Space with Matrix Product States" (2605.17895) proposes and thoroughly develops a geometric, interpretable machine learning framework that operates natively in quantum Hilbert space. The work addresses a central challenge in ML: achieving strong predictive performance while maintaining model interpretability, which is particularly acute in high-stakes applications such as medical diagnostics. Traditional deep neural networks, despite state-of-the-art results, remain largely opaque to users; classical prototype-based methods offer more transparency but suffer from limitations in modeling data complexity. By elevating prototype learning to the quantum domain and leveraging Matrix Product States (MPS) as generative class representatives, the authors establish a framework where representation, similarity, and decision-making are unified through Hilbert-space geometry. This approach provides both a rigorous probabilistic foundation and potential for performance that is competitive with leading deep learning architectures.

Theoretical Framework: Quantum-Probabilistic Prototype Learning

The central methodological innovation is Quantum-Probabilistic Prototype Learning (QPPL), wherein:

• Each data sample is mapped via a quantum feature map into a product state in a high-dimensional Hilbert space, wherein each classical feature is encoded as a qubit or higher-dimensional "spin."
• Class prototypes are represented as generative MPS, a tensor network topology that models global correlations and entanglement efficiently.

Both data and prototypes are represented as normalized quantum states in the same Hilbert space. Crucially, geometric relationships—essential for classification, clustering, and anomaly detection—are measured via negative logarithmic fidelity (NLF): for two states $|\psi\rangle$ and $|\psi'\rangle$, $$D^{\mathrm{NLF}}(|\psi\rangle, |\psi'\rangle) = -\ln |\langle \psi|\psi' \rangle|$$. This distance metric maintains desirable properties for high-dimensional, many-body quantum systems and avoids pitfalls associated with standard fidelity decay.

MPS prototypes are trained generatively, modeling the data distribution by maximizing overlap (minimizing NLF) with samples from the same class. This positions each prototype as an entangled Hilbert-space "centroid" of its class, in a sense more structurally sophisticated than Euclidean centroids in classical space.

Empirical Results: Classification, Anomaly Detection, Clustering

Supervised Classification and Hierarchical Decision Schemes

QPPL is evaluated rigorously on real-world and benchmark datasets, notably a high-resolution, hierarchically annotated electrocardiogram (ECG) dataset and Fashion-MNIST. In all supervised classification tasks, QPPL demonstrates:

• Superior accuracy and AUROC compared to classical prototype methods (nearest centroid, k-NN, GLVQ, RBFN), particularly as task complexity increases (e.g., ternary anatomical ECG sublocalization).
• Competitive performance with modern neural network baselines (FC, CNN, LSTM), achieving the highest accuracy (92.14%) and AUROC (0.9888) on the hardest ECG task while remaining within a margin of CNN performance on Fashion-MNIST.
• Enhanced robustness and statistical stability (lower variance) compared to several baselines.

Importantly, the geometric organization permits hierarchical classification: by structuring successive MPS prototypes in a decision tree, the model aligns naturally with clinical reasoning and deep learning architectures for hierarchical tasks [ECGhiera, Texhiera]. This reduces class overlap and improves interpretability at each decision stage, further confirmed by an accuracy boost in multi-level classification.

Anomaly Detection

Anomaly detection is natively enabled in QPPL by quantifying outlierness via the NLF distance from a sample to its nearest prototype—no external scoring mechanism or regularization is required. On ECG anomaly detection tasks, QPPL consistently attains the highest AUROC across all levels, outperforming both prototype-based and deep-learning baselines. This performance is a direct consequence of the global, entanglement-capturing properties of MPS in Hilbert space, which enhance the sensitivity to distributional deviations.

Clustering and Dimensionality Reduction

The geometric properties of the QPPL framework facilitate superior clustering and dimensionality reduction:

• t-SNE visualizations based on Hilbert-space NLF distances yield clearer, more compact cluster structures and higher silhouette coefficients compared to Euclidean distances.
• By embedding both samples and prototypes in low-dimensional spaces using a distance matrix that includes all sample-prototype and prototype-prototype NLF distances, the clustering structure is further improved.
• The distances to prototypes define a data-adaptive, supervised metric space, yielding compact, discriminative representations suitable for both visualization and downstream ML tasks.

A notable assertion is that the positions of MPS prototypes in dimensionality-reduced space do not necessarily coincide with empirical centroids, reflecting the fundamentally non-classical encoding of global distributional information and the active prototype-induced reshaping of the data manifold.

Implications and Future Directions

The findings underscore several important implications:

• Interpretability and Transparency: All inference tasks (classification, anomaly detection, dimensional reduction) are performed directly through geometric distances in Hilbert space. The absence of black-box or readout layers affirms the interpretability of the method, satisfying critical demands in domains such as medicine.
• Unified Geometric Inference: The same core distance metric governs multiple ML operations, creating a consistent and mathematically principled decision framework.
• Potential for Hybrid and Quantum Computation: The use of MPS as prototypes paves the way for quantum-inspired and, possibly, quantum-enhanced ML methods as scalable quantum hardware matures.
• Generality Across Modalities: The demonstrated effectiveness on both temporal (ECG) and spatial (Fashion-MNIST) datasets indicates adaptability to a broad class of ML problems.

Future work may explore: scaling QPPL to even higher-dimensional data and larger class vocabularies; leveraging more expressive tensor network topologies (PEPS, MERA) for richer data distributions; integrating QPPL with quantum hardware for end-to-end quantum learning; and theoretical analysis of the geometric manifold structures induced by prototype learning in Hilbert space.

Conclusion

This work rigorously formulates and demonstrates a geometric, interpretable framework for prototype-based learning in quantum Hilbert space using MPS generative prototypes (2605.17895). The QPPL approach harmonizes high performance with transparent decision-making, enabling robust classification, anomaly detection, and clustering in a mathematically natural and extensible manner. The research opens new avenues for the design of inherently interpretable, quantum-native ML algorithms, and establishes foundational benchmarks for geometric approaches in quantum machine learning.