arXiv:2609.31027
When Can Quantum Extreme Learning Machines Replace Quantum Reservoirs?
Markus Baumann, Gerhard Stenzel, Jonas Stein, Claudia Linnhoff-Popien
Abstract
Quantum reservoir computers (QRCs) retain information from earlier inputs, whereas reset-window quantum extreme learning machines (QELMs) start afresh and receive only a recent input window. We establish when this window can replace the reservoir's memory. Under our stated assumptions and with exact expectation values, sufficiently expressive QELMs can reproduce reservoir outputs arbitrarily accurately as their windows grow precisely when the influence of the omitted past becomes uniformly negligible. However, longer windows do not always solve the problem. We construct a simple two-qubit reservoir that performs temporal recall exactly, while every fixed-window predictor faces the same positive worst-case error over unbounded delays, regardless of model capacity. Analytical benchmarks and simulations distinguish missing history from limited features and finite measurements. The central conclusion is simple: better representations can improve how available information is used, but cannot recover a past the model never receives.
AI Summary
The paper asks when a reset-window quantum extreme learning machine (QELM), receiving only the latest $d$ inputs, can reproduce a quantum reservoir computer’s outputs. It separates unavailable history from limited representation capacity and measurement noise. Define $D_d$ as the largest output difference between histories sharing their last $d$ inputs. For scalar outputs, the optimal worst-case window-only error is exactly $D_d/2$, attained by the midpoint of compatible outputs. Under finite-dimensional, compact-input, continuity and freely concatenable-input assumptions, a uniformly universal QELM family approaches this bound with exact expectations. Arbitrarily accurate emulation is therefore possible iff $D_d\to0$. A reachable–observable reduction connects this condition to a joint spectral radius $r_{\mathrm{vis}}<1$, with $\lim_d D_d^{1/d}=r_{\mathrm{vis}}$. For $0<r_{\mathrm{vis}}<1$, the required context scales asymptotically as $\log(1/\varepsilon)/[-\log r_{\mathrm{vis}}]$. A two-qubit counterexample stores an old pulse in a protected singlet population: a persistent gap $\delta$ forces error at least $\delta/2$ for every fixed window over unbounded delays. A one-radian pulse gives $\delta\approx0.1149$. Proofs, Bernstein-based QELM constructions, analytical one-qubit benchmarks and small-system simulations establish or illustrate these statements. The guarantees concern representability, not efficient learning or resource savings. One construction uses $rd$ qubits and up to $(r+1)^d$ readout coefficients, whose computation is excluded from resource counts. Finite-shot bounds are pointwise, not simultaneous over histories. Simulations do not estimate worst-case memory decay, omit hardware errors and measurement backaction, and include classical controls outperforming the tested quantum models. The separation concerns retained memory versus finite context, not quantum versus classical computation.
OpenAI Codex assisted with ideation, manuscript wording, and code and figure preparation, including debugging.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 12 pages, 3 figures