Quantum reservoir computing for predicting and characterizing chaotic maps

Quantum reservoir computing has emerged as a promising paradigm for harnessing quantum systems to process temporal data efficiently by bypassing the costly training of gradient-based learning methods. Here, we demonstrate the capability of this approach to predict and characterize chaotic dynamics in discrete nonlinear maps, exemplified through the logistic and Hénon maps. While achieving excellent predictive accuracy, we also demonstrate the optimization of training hyperparameters of the quantum reservoir based on the properties of the underlying map, thus paving the way for efficient forecasting with other discrete and continuous time-series data. Using closed-loop prediction of distant future steps, our protocol discriminates between chaotic and nonchaotic phases without prior knowledge of the underlying map or the nature of the time series. Furthermore, the framework exhibits robustness against decoherence when trained in situ and shows insensitivity to reservoir Hamiltonian variations as well as robustness to finite-sampling error. These results highlight quantum reservoir computing as a scalable and noise-resilient tool for modeling complex dynamical systems, with immediate applicability in near-term quantum hardware.

💡 Research Summary

This paper presents a comprehensive study of quantum reservoir computing (QRC) applied to the prediction and characterization of chaotic dynamics in discrete nonlinear maps, specifically the logistic and Hénon maps. The authors construct a quantum reservoir using a linearly coupled transverse XY spin chain with fixed exchange couplings (J = 1) and randomly sampled on‑site magnetic fields. Input time‑series data are encoded into qubits via Pauli‑Y rotations; each past value is fed sequentially through d layers, and the same input may be repeated n_rep times at each layer to inject higher‑order nonlinear terms. After each layer the reservoir evolves under the XY Hamiltonian for a fixed time τ, after which the input qubits are discarded and fresh qubits are introduced for the next time step, while hidden qubits retain the accumulated memory. At the final step all qubits (input and hidden) are measured in the Pauli‑X basis, yielding a vector of expectation values m. Prediction of the next scalar value is performed by a simple linear read‑out x̂ = W·m, where the weight matrix W is obtained analytically by ridge regression (W* = Y Mᵀ(MMᵀ + εI)⁻¹) using a training set of short windows. No gradient‑based optimization of the reservoir itself is required, preserving the “training‑once” advantage of reservoir computing.

For the logistic map (xₜ = r xₜ₋₁(1 − xₜ₋₁), r∈