arXiv:2609.38678
Exponential separation in sensing continuous signals via squeezing
Francesco Anna Mele, Nadine Meister, Haimeng Zhao, Senrui Chen, Hsin-Yuan Huang
Abstract
Quantum sensing traditionally focuses on using quantum resources such as squeezing and entanglement to improve precision for sensing fixed signals. However, in many applications such as gravitational-wave detection and electromagnetic-field sensing, the signal evolves continuously and varies through time. Additionally, the learner is free to prepare, control, and measure the sensor at arbitrary times, possibly chosen adaptively. In this work, we establish exponential separation in sensing time for continuously evolving signals due to the available squeezing. The signals we study are characterized by a pattern size $T$, and we find that a sensor with squeezing at least $ω(\sqrt{\log T})$ offers a $\mathrm{poly}(T)$ sensing time, whereas those with squeezing at most $o(\sqrt{\log T})$ must use an exponential sensing time in $T$.
AI Summary
Can squeezing yield exponential reductions in sensing time for continuously evolving signals, even when interrogation times and controls are adaptive? This paper studies a single bosonic mode with signal Hamiltonian $H(t)=f(t)\hat p$, restricting protocols to Gaussian input states with squeezing at most $z$, squeezing-free Gaussian controls, destructive Gaussian measurements, and unrestricted classical processing. Signals comprise unit-duration pulses grouped into independent blocks of size $T$. Two tasks encode information in temporal correlations: recovering an unknown parity constraint on binary pulse amplitudes, and distinguishing two known sign distributions with identical marginals on every subset of at most $T-c$ coordinates. The latter uses independently randomized pulse magnitudes, assumes $c=O(T^{0.99})$, and requires inverse-polynomial total-variation separation for efficient sensing. For both tasks, $z=\omega(\sqrt{\log T})$ permits polynomial sensing time, whereas $z=o(\sqrt{\log T})$ requires at least $\exp(T^{1-o(1)})$ time. For parity learning, an explicit lower bound is $T[\frac13\exp(Te^{-3z^2}/90)-1]$. The upper-bound protocol measures each pulse separately; Gaussian readout errors decrease as $e^{-\Omega(z^2)}$. Parity learning also admits polynomial total computational time at sufficiently large squeezing; general hypothesis testing guarantees only polynomial sensing time. The lower bounds reduce Gaussian experiments to noisy linear observations and use a hierarchical learning-tree representation to handle adaptive timing within pulses, blocks, and across blocks. Stated limitations include deliberately stylized signals, a single-mode setting, destructive prepare-and-measure protocols, and no lower-bound extension to multimode adaptive sensing, nondestructive Gaussian measurements, or non-Gaussian readout.
ChatGPT and Gemini were used later on in 2026 to make minor refinements to our human-made results at the end of the project.
Central claim may not hold: The abstract presents the lower bound without the essential single-mode Gaussian, squeezing-free-control, and destructive-measurement restrictions, and the paper does not establish the same result beyond that protocol class. Throughout the stated low-squeezing regime, the proven lower bound is exp(T^(1-o(1))), not uniformly exp(Ω(T)), and the explicit parity upper bound rules out the latter interpretation.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 124 pages, 12 figures