analiza aiqbench

This research introduces Quantum Spectral Models (QSMs) that leverage truncated Fourier representations and Hamiltonian constructions to create an input-dependent inductive bias. By directly constructing data-encoding unitaries from spectral values rather than standard coordinate-wise encoding, these models achieved superior test accuracy on benchmarks like Pendigits compared to existing quantum machine learning approaches. The findings highlight a significant methodological advance where sample-conditioned spectral characteristics are coupled through trainable mixers to enhance performance without relying solely on eigenvalue classification.