arXiv:2610.26228
The Fundamental Memory Ceiling of Noisy Quantum Reservoirs
Markus Baumann, Jonas Stein, Claudia Linnhoff-Popien
Abstract
Fresh qubits, not size, set how long a noisy quantum system can remember. With $k$ input qubits reinitialized per time step, the longest memory any reservoir can attain under depolarizing noise is $\exp[Θ(k)]$ time steps, whatever its size, dynamics, and input encoding. A universal upper bound and an explicit reservoir that attains the exponential scaling establish this optimal resource law. The bound depends on the dynamics only through the entropy that fresh qubits remove per time step, so it holds for any Markovian noisy dynamics without feedback, including measurement-free error correction.
AI Summary
How long can a noisy quantum reservoir retain an input if its size is unrestricted but only $k$ qubits are refreshed per time step? The paper considers arbitrary input encodings and global unitaries followed by independent single-qubit depolarizing noise $\mathcal D_q$, with no classical feedback or additional entropy sink. Memory depth is defined through the trace-distance distinguishability of echo states whose input histories agree on their last $L$ steps, excluding information retained solely from initialization. For fixed $0<q<1$ and $0<\varepsilon<q$, the maximum memory depth over all reservoir sizes scales as $W_{q,k}(\varepsilon)=\exp[\Theta_q(k)]$. The universal upper bound is $W\le\lceil\delta^{-1}\ln(1/\varepsilon)\rceil$, with $\delta^{-1}\le2^{k/(1-q)^2}$ for $k\ge4$. A classical reversible repetition-code reservoir using at most $2k$ qubits and majority-vote repair achieves $W\ge\frac14e^{\alpha_q(k-3)}\ln(q/\varepsilon)$ for sufficiently large $k$, where $\alpha_q=\frac12\ln[1/(1-q^2)]$. This establishes the converse to the known logarithmic fresh-qubit cost of retaining a bit for $T$ rounds. The proof combines hypercontractivity, which bounds echo-state purity independently of size, with reverse hypercontractivity and a contracting combination of trace and Schatten-2 norms. The framework extends to channels satisfying $\mathcal E(I)\le2^bI$, treating $b$ as an entropy-removal budget. The upper and lower exponents remain far apart: at $q=0.9$, $k=5$, and $\varepsilon=0.01$, the upper bound exceeds $2^{470}$ steps, versus 316 for the construction. Feedback, non-Markovian noise, and noise without a depolarizing component are outside scope; amplitude damping supplies an additional size-dependent budget. Memory depth alone does not establish temporal-task performance or quantum advantage.
Claude Opus 5.5 assisted with code development.
Error found: The abstract incorrectly extends the bound to any feedback-free Markovian noisy dynamics, although the theorem requires a fixed local depolarizing component and pure dephasing permits indefinite classical-bit retention. The abstract omits the requirement to count additional entropy sinks in the budget, although its own amplitude-damping construction achieves memory exponential in system size with only two fresh input qubits per step.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 25 pages, 11 figures, including Supplemental Material. Reproducibility code and data included