arXiv:2610.14454
Quantum-Fisher-information-aware optimization yields a quantum reservoir advantage over tuned classical reservoirs
Phi Hung Nguyen, Daniel Soh
Abstract
We show that a quantum reservoir computer optimized with its quantum Fisher information (QFI) has a quantum advantage over a classical reservoir of equal readout cost, provided by a quantum-information advantage: optimizing the quantum parameters with the QFI metric. On Mackey-Glass forecasting, a driven Jaynes-Cummings reservoir with one to three atoms beats a tuned echo-state network at every size from 4 to 30 trained weights, by 25-69% in held-out error, and is ahead of 99.5% of 10,000 random networks. After grid and Bayesian searches saturate, the QFI-aware stage lowers the error by 29-51% at 12-30 weights, beyond equal-budget black-box search or classical-parameter gradients alone, and at 30 weights turns a loss into a win. We then derive a fundamental QFI bound on the first-order rate of change of the retrained loss and factorize the gap to this bound into five link efficiencies that identify the physical mechanism limiting optimization.
AI Summary
The paper asks how quantum-state sensitivity translates into improvement of a reservoir computer when its linear readout is retrained, and whether QFI-informed optimization improves forecasting relative to a tuned classical reservoir. For least-squares loss, it derives $D_v\mathcal L=-r^T X_vw^*$ and bounds the normalized slope by a feature-weighted QFI budget: $|D_v\mathcal L|/(\|r\|\|w^*\|)\le\sqrt{v^TG_Qv}$. The gap factorizes exactly into five efficiencies describing measurement, observable response, readout utilization, novelty outside the current feature span, and residual alignment. The optimal first-order decrease within one QFI unit is $\sqrt{g^TG_Q^+g}$. Small Jaynes–Cummings and four-spin Ising simulations test the chain and interventions informed by its factors. On five-step Mackey–Glass forecasting, simulated one-to-three-atom reservoirs with 4–30 readout weights obtain 25–69% lower held-out NRMSE than equal-readout-size echo-state networks, whose hyperparameters are tuned and performance evaluated over 10,000 random realizations. After grid and Bayesian search, a stage alternating QFI natural-gradient updates with classical-parameter optimization reduces held-out error by 29–51% at 12–30 weights; comparisons include continued black-box search and classical-parameter-only gradients. The comparison matches readout count, not total optimization freedom: all quantum-device parameters are optimized, while classical recurrent matrices remain random. Results use expectation values; retaining the reported margin requires estimated $7\times10^6$–$2\times10^{10}$ shots per feature per sample. QFI-based design requires density matrices, the theory is local and fixed-rank, and its bound is loose by two to six orders of magnitude. Link diagnostics do not predict finite-step improvement reliably.
A large language model (Claude, Anthropic) was used to assist with developing and reviewing analysis code
Central claim may not hold: The claimed attribution to the QFI metric is not isolated because the gradient control freezes the quantum parameters, leaving untested whether ordinary gradient-based optimization of the same parameters achieves the reported gains. At m=24 and m=30, the black-box baselines inherit controls or starting points from earlier natural-gradient optimization, so they do not independently establish that loss-only search cannot reach the reported designs.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.