Engineering Quantum Reservoirs through Krylov Complexity, Expressivity and Observability

This study employs Krylov-based information measures to understand task performance in quantum reservoir computing, a sub-field of quantum machine learning. In our study we show that fidelity and spread complexity can only explain the task performance for short time evolutions of the quantum systems. We then discuss two measures, Krylov expressivity and Krylov observability, and compare them to task performance and the information processing capacity. Our results show that Krylov observability exhibits almost identical behavior to information processing capacity, while being three orders of times faster to compute. In the case when the system is undersampled Krylov observability best captures the behavior of the task performance.

💡 Research Summary

This paper investigates how to quantify and predict the performance of quantum reservoir computing (QRC) systems using Krylov‑based information measures. Traditional metrics such as fidelity and spread (operator) complexity have been shown to correlate with task performance only during the early stages of the quantum dynamics; they quickly become oscillatory and fail to explain the long‑term saturation observed in chaotic time‑series prediction tasks. To address these shortcomings, the authors introduce two novel Krylov‑space quantities: Krylov expressivity and Krylov observability.

Krylov expressivity quantifies the effective dimensionality of the Krylov subspace generated by the initial encoded state under the system Hamiltonian. It is linked to a “Krylov grade” that can be inferred directly from the Hamiltonian’s spectral properties, providing a physically interpretable bound on how richly the reservoir can embed input data. Krylov observability, on the other hand, measures how many linearly independent expectation values can be obtained when a set of observables is measured repeatedly at different times within the same Krylov subspace. Because it is defined in terms of experimentally accessible expectation values, observability can be evaluated on real quantum hardware without full state tomography.

The authors benchmark four transverse‑field Ising reservoirs with distinct inter‑spin couplings and eigenvalue multiplicities. They train each reservoir on two chaotic prediction tasks derived from the Lorenz‑63 system: (i) a five‑step ahead prediction of the x‑coordinate (LXX) and (ii) a cross‑prediction of the z‑coordinate from the current x (LXZ). Performance is assessed using normalized root‑mean‑square error (NRMSE) and Pearson correlation, while the information processing capacity (IPC) is computed using Legendre polynomial targets.

Key findings include:

1. Fidelity and spread complexity capture only the initial rise in performance; they oscillate and cannot explain the eventual plateau.
2. Krylov expressivity grows initially but reaches a saturation point that varies across Hamiltonians; this saturation does not reliably predict task accuracy.
3. Krylov observability increases monotonically and mirrors the behavior of IPC across all reservoirs. When the read‑out dimension (NR) is large, observability and IPC are virtually indistinguishable.
4. For small NR, IPC is capped by the read‑out dimension, whereas Krylov observability continues to reflect the reservoir’s ability to retain and transform information, thus providing a more faithful predictor of task performance under undersampling conditions.
5. Computationally, evaluating observability requires roughly three orders of magnitude fewer matrix operations than constructing the full state matrix for IPC (Nobs ≈ 10⁻³ Nu). This makes observability attractive for real‑time diagnostics on quantum devices.

The paper also discusses how the Krylov grade can be derived from Hamiltonian spectra, allowing practitioners to set physically motivated cut‑offs that avoid artificial inflation of the Krylov dimension due to numerical orthogonalization errors. This insight explains why certain Hamiltonians may exhibit a smaller operator Krylov space despite having a larger state‑space Krylov dimension.

In conclusion, Krylov observability emerges as a fast, experimentally accessible, and theoretically grounded metric that captures the same information as the traditional IPC while being far more computationally efficient. It therefore offers a practical tool for designing, benchmarking, and optimizing quantum reservoirs, especially in regimes where measurement resources are limited. Krylov expressivity, while less directly predictive of task performance, provides valuable insight into encoding strategies and the intrinsic expressive power of different quantum Hamiltonians. The authors suggest future work extending observability analysis to variational quantum algorithms and quantum extreme learning machines, aiming to deepen the explainability of quantum machine‑learning models.