Speaker
Description
Algorithmic shadow spectroscopy extracts molecular transition frequencies from randomized measurements of many observables over time at a fixed molecular geometry. We ask whether the underlying object, the time-dependent fermionic-shadow signal matrix of degree-two Majorana observables, can instead be learned across a geometry scan, amortizing shadow-spectroscopy observables so that the inputs to spectral post-processing at held-out geometries are predicted classically. We introduce the Fermionic Shadow Regressor (FSR), an explicit amplitude-phase network whose oscillation frequencies are conditioned on Hartree-Fock orbital energies and bounded, geometry by geometry, by an exact line-spectrum screen. On four simulated eight-qubit molecules a single recipe transfers with high held-out fidelity, outperforms generic Gaussian-process and Fourier-feature surrogates on the harder scans, and degrades in the spectral-congestion region identified by the screen; a LiH finite-shot test matches the infinite-shot reference at $S=1000$. The targets are classically simulated and the screen is an exact-diagonalization oracle: this is an amortization-and-sampling-design study, not a demonstration of quantum advantage.
| Topic | Quantum state and process tomography with learning-based approaches |
|---|---|
| Additional Topic | Quantum state reconstruction from data |