EUVIMEDEUVIMED entdecken
Crossref

Orthogonal Catastrophe in Dynamical Quantum Kernels: Spectral Scrambling, Initial-State Ergodicity, and the Informative Time Window

Youle Wang, Hanmiao Cheng, Wenbin Yu, Lei Zhang

Quantum Science and Technology · 18.11.2026

DOI: 10.1088/2058-9565/aec3cf ↗

Kurz gefasst

Fachartikel in der Zeitschrift Quantum Science and Technology, online erschienen am 18.11.2026. Verfasst von Youle Wang und 3 weiteren Autor:innen. Gefördert von: Startup Foundation for Talent Introduction at NUIST, Basic Research Program of Jiangsu Province, Quantum Science and Technology National Science and Technology Major Project.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Ergodicity, and the Informative Time WindowQuantum kernel methods promise advantage by mapping data into high-dimensional Hilbert spaces, yet current parameterized quantum circuit (PQC) architectures suffer from the orthogonal kernel catastrophe-the exponential concentration of kernel entries that destroys trainability. This crisis arises because deep PQCs rapidly scramble quantum states into indistinguishable Haar-like distributions. To overcome this, we introduce dynamical quantum kernels generated by continuous-time Hamiltonian evolution. We rigorously prove that although dynamical evolution postpones the instantaneous concentration characteristic of deep PQCs by generating short-time signals, the asymptotic long-time catastrophe persists via distinct physical mechanisms: the global fidelity kernel concentrates due to spectral scrambling (arising from non-resonance conditions typical of chaotic systems), whereas the projected kernel collapses solely from initial-state ergodicity, where Haar-random delocalization extinguishes local signals independent of the Hamiltonian's chaoticity. Crucially, this catastrophe is not immediate. Through rigorous short-time perturbation analysis of local Hamiltonians, we prove the existence of a robust informative time window. Utilizing linearly parameterized Hamiltonians H(θ) = H0 + θV , we derive an exact Ω(n)t 2 feature variance scaling and show that localized product states guarantee an Ω(1) discriminative signal per site, escaping the PQK catastrophe. Our work connects kernel trainability to the spectral and ergodic properties of Hamiltonian dynamics, dictating that practical advantage requires leveraging localized initial states and the dynamical grace period before global scrambling dominates.

Förderung

  • Startup Foundation for Talent Introduction at NUIST
  • Basic Research Program of Jiangsu Province
  • Quantum Science and Technology National Science and Technology Major Project

Zitieren

APA

Wang, Y., Cheng, H., Yu, W., & Zhang, L. (2026). Orthogonal Catastrophe in Dynamical Quantum Kernels: Spectral Scrambling, Initial-State Ergodicity, and the Informative Time Window. Quantum Science and Technology. https://doi.org/10.1088/2058-9565/aec3cf

Vancouver

Wang Y, Cheng H, Yu W, Zhang L. Orthogonal Catastrophe in Dynamical Quantum Kernels: Spectral Scrambling, Initial-State Ergodicity, and the Informative Time Window. Quantum Science and Technology. 2026. doi:10.1088/2058-9565/aec3cf

BibTeX
@article{Wang2026,
  author = {Wang, Youle and Cheng, Hanmiao and Yu, Wenbin and Zhang, Lei},
  title = {Orthogonal Catastrophe in Dynamical Quantum Kernels: Spectral Scrambling, Initial-State Ergodicity, and the Informative Time Window},
  journal = {Quantum Science and Technology},
  year = {2026},
  doi = {10.1088/2058-9565/aec3cf}
}
Kostenlos starten →