Benchmarking Non-Markovianity Measures in Digital Quantum Simulation of Open Quantum Systems

Abstract Digital quantum simulation of open quantum systems has produced a range of methods for reproducing non-Markovian dynamics, including pseudomode embeddings, the hierarchical equations of motion, and collision models. These methods are validated against populations, coherences, or spectra. They are not validated against the quantities that define memory itself, and the behavior of those quantities under the pressures of a real simulation is largely unexamined. We study three such pressures. First, a single-qubit mode encoding misrepresents the Breuer-Laine-Piilo (BLP) measure with an error whose sign flips between regimes: it fabricates memory at strong coupling, by about a factor of 2.6 over the converged value, and it erases memory at moderate temperature. The Rivas-Huelga-Plenio (RHP) measure corroborates the temperature-driven destruction, while at strong coupling it is numerically unstable and does not track the BLP result. We derive a cross-validated rule for the required Fock dimension that collapses both regimes onto the occupied number extent, d req ≈ ⌈⟨n⟩+c σ n ⌉+1, and we show that the constant must be calibrated on super-Poissonian statistics. Second, using the exact noisy measure as computable ground truth, we show that the raw finite-shot estimate of the BLP measure can nearly double the true value under realistic noise. We provide a debiased estimator that is validated near-unbiased in that regime, together with a Markovian null-floor control that ports to hardware. Third, temporally correlated device noise is conflated with physical memory by the measure. A resource-fair comparison finds that a pseudomode and an ancilla-train collision model reach parity at equal qubit cost for a single-Lorentzian bath. Every result is simulator-based, validated against an exact reference, and reproducible from recorded data. We make no claim of quantum advantage.