Sparsified Kolmogorov-Arnold Networks for Interpretable Quantum State Tomography

Researchers demonstrate that sparsified Kolmogorov-Arnold Networks (KANs) can perform quantum state tomography while remaining interpretable, recovering physical structure without superior performance. The method identifies relevant Pauli measurements from 63 total measurements and reveals internal pathways consistent with known quantum mechanics, validating that neural models can be audited against established physics.

This research addresses a fundamental challenge in machine learning: building models that not only predict accurately but also remain transparent enough to verify against known physical laws. The team applied sparsified KANs to quantum state tomography—the problem of reconstructing quantum states from measurement data—using a three-qubit GHZ entangled state as their benchmark. Rather than claiming superior accuracy, the authors demonstrated that the network's internal structure naturally organizes around physically meaningful patterns, recovering exactly 12 out of 63 Pauli measurements as most relevant, which aligns with the theoretical structure of GHZ states.

This work emerges from the broader intersection of machine learning and quantum physics, where neural networks increasingly assist in processing quantum data but their decision-making remains opaque. The sparsification approach—essentially pruning unnecessary connections—enables researchers to inspect which measurements matter and how they combine, creating what the authors call "pathway-level structural interpretability."

For quantum computing and AI development, this methodology matters because it establishes a framework for auditing learned models against physical ground truth. Rather than trusting a black-box neural network to reconstruct quantum states in applications like quantum error correction or state preparation, researchers can now systematically verify that the network's internal logic matches established physical principles. The negative controls and stability analyses across noise conditions strengthen confidence in these findings.

Looking ahead, this approach could inform similar interpretability techniques in other physics-informed machine learning domains, where domain knowledge should constrain model behavior. The key question is whether these methods scale efficiently to larger quantum systems where computational overhead becomes critical.

• →Sparsified KANs recover the correct 12 Pauli measurements from 63 total measurements, demonstrating automatic discovery of physical structure.
• →The network's internal pathways organize Z-type and X/Y observables in patterns consistent with analytical GHZ physics, validating learned representations.
• →Interpretability remains stable across multiple random initializations and noise levels, with collapse under random-label controls confirming physical meaningfulness.
• →The contribution is pathway-level transparency rather than superior regression performance, establishing a framework for auditing neural reconstruction rules.
• →This methodology enables systematic verification that machine learning models respect known physical constraints in quantum information processing.