Benchmarking quantum dynamics against exact models

For a closed finite system, compare with an independently evaluated matrix exponential or an analytic two-level rotation. Check complex amplitudes up to an allowed common phase, selected observables, norm and any genuinely conserved quantities. A common energy offset changes only global phase in exact dynamics, while a wrong relative phase changes interference. Population-only comparisons can miss that distinction, especially in symmetric examples with equal occupations.

Spatial methods need additional references: exact box levels, free-packet dispersion and stationary scattering fluxes test different components. Compare like representations before calling a discrepancy a propagator error; a centered-stencil kinetic matrix and a Fourier kinetic matrix have different finite spectra. Open-system tests should include exact population and coherence decay, trace, Hermiticity, positivity and the sampling statistics of any trajectory solver. No single benchmark spans every one of these mechanisms.

Timestep subdivision or an embedded comparison estimates temporal error for a specified finite operator. Krylov dimension, Chebyshev order or iterative-solve tolerance can then be tightened so their errors do not dominate that comparison. The adaptation rule should target the reconstructed state or relevant observable under a declared norm; conserving probability alone cannot estimate a unitary method's phase defect. Very small adaptive steps also do not repair an incorrectly implemented Hamiltonian.

After temporal convergence, increase Hilbert dimension or spatial bandwidth, move boundaries and vary absorber or bath parameters as appropriate. These controls answer different questions. A result can converge beautifully in timestep toward a biased fixed-basis limit. Conversely, a basis-refinement comparison can be obscured by loose solver tolerances or Monte Carlo noise. Record enough independent controls to attribute an observed plateau instead of assuming that every plateau represents the physical answer.

For a quantity with a dominant second-order temporal error, halving timestep reduces that leading term by four and Richardson extrapolation removes it. The extrapolate remains the limit of the chosen finite model. Comparison with an external exact or experimentally justified reference is needed to diagnose shared representation or model bias. The synthetic phase example below makes the distinction explicit rather than equating an observed order with truth.

Physical validation additionally asks whether the chosen Hamiltonian, nonrelativistic regime, driving field and environmental approximation describe the intended system. An exactly solved phenomenological Lindblad equation is still a phenomenological model. Error reports should name the observable, duration and parameter range; good short-time occupations need not bound long-time phase, rare-transition probabilities or every derived measurement. Clear scope makes a benchmark informative instead of turning one successful test into a universal accuracy claim.