Fig. 3: Comparing frequency and full identification errors.
From: Robustly learning the Hamiltonian dynamics of a superconducting quantum processor
a In an N = 6 subset of connected qubits, by varying b from 0 to 1, we implement 51 different Hamiltonians. The plot shows the Fourier transform of the time domain data. b The extracted eigenfrequencies (denoised peaks in (a)) are shown as colored dots, where the assigned color is indicative of the deviation between targeted eigenfrequencies (gray lines) and the identified ones from position of the peaks. c Analog implementation error \({{{\mathcal{E}}}}_{{{\rm{analog}}}}(\hat{h},{h}_{0})\) of the identified Hamiltonian (dark red) compared to the implementation error \({{{\mathcal{E}}}}_{{{\rm{analog}}}}({{\rm{eig}}}(\hat{h}),{{\rm{eig}}}({h}_{0}))\) of the identified frequencies (golden). Colored (gray) error bars quantify the statistical (systematic) error. d Layout of the six qubits on the Sycamore processor and median of the entry-wise absolute-value deviation of the Hamiltonian matrix entries from their targeted values across the ensemble of 51 different values of b ∈ [0, 1].
