Detection-Limited Eigenvalue Learning for Hybrid Quantum Linear System Solvers
Hybrid quantum linear system solvers of the Harrow--Hassidim--Lloyd (HHL) type learn the eigenvalues that carry weight in the input state and then run a short clock register, but the choice of the learning estimator has not been analysed. We show that the detection of weak spectral components decides this choice and give a counting model: a component of weight $w$ is found with about $(w+w_{\rm fl})/w^2$ samples, with $w_{\rm fl}$ the weight leaking into its label. Single-ancilla estimators on mixture statistics need about $1/w^2$ samples, and labelling each shot with a clock register (Projective HHL) needs about $1/w$, down to a leakage floor of a few per cent. Recycled sequential HHL keeps one copy of the input, measures it repeatedly with one ancilla and labels the copy by its own Bayesian posterior, without a register; its label error is tuned by a stopping threshold, which is not a proven bound, and measured in simulation. On four systems with up to eight eigenvalues and calibrated ancilla noise it gives a lower median solution infidelity than Projective HHL within the same evolution time in all eight settings, by factors from $2$ to $3\times10^3$, while neither is uniformly better for the compliance. On DC power-flow problems of 9-, 14- and 30-bus systems it is, without dephasing, the only stage tested that resolves the weak component at the smallest eigenvalue of the 9-bus system in most runs; with dephasing none reliably resolves the weak components of the larger systems, and the coherent clock limits every solver.
Authors
- Taiki Tanimae
- Hikaru Wakaura (ORCID: https://orcid.org/0000-0001-8381-8323)
Institutions
- Quantum Integrated Research Institute Inc. (Japan) (JP)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11766188/v1
- Primary Topic
- Quantum Computing Algorithms and Architecture
- Type
- preprint