Paper Overview
Field: Machine Learning Authors: Peiyong Wang, Udaya Parampalli, Casey R. Myers Published: 2026-07-24 arXiv: 2607.22516
Summary
A central design principle in modern machine learning and artificial intelligence is to align a model's inductive bias with the structure of its input data. For matrix-valued inputs, relevant matrix-level relationships can be characterized through spectral values and spectral subspaces; however, the common coordinate-wise rotation-gate data-encoding unitaries used in most quantum machine learning models do not explicitly construct such a matrix-level representation.
This paper introduces Quantum Spectral Models (QSMs), in which the generator of the data-encoding unitary is constructed directly from each input matrix. The authors study three QSM variants based on:
- Symmetric Hamiltonians
- Global block Hamiltonians
- Non-overlapping patch-local block Hamiltonians
- Two matrix representations of the Pendigits dataset
- Two controlled synthetic tasks defined by spectral statistics
- At the maximum evaluated circuit depth, QSM variants lead over the tested quantum models on all four benchmarks.
- Patch-local QSM leads on Pendigits, while the global block Hamiltonian QSM leads on the controlled spectral tasks.
- Ablation experiments reveal a task-dependent reversal: subspace retention control performs better on Pendigits, whereas spectral-value-only control wins in the ablations on synthetic tasks.
Their outputs admit truncated Fourier representations in which input-dependent spectral gaps provide candidate phase carriers, while spectral subspaces help determine their coefficients.
Evaluation
QSMs and competing quantum models were evaluated on:
Findings:
Takeaway
These results provide a new perspective on quantum machine learning model design, showing how input-conditioned spectral representations can offer analyzable inductive biases, with broader implications for structure-aware model design in machine learning and AI.
--- *Auto-collected on 2026-07-28*