Execution bias, repeatability and mitigation cost in parameter-frozen molecular VQE energy evaluation

Abstract

A smaller statistical error bar need not mean a more accurate molecular energy. We evaluate noiselessly optimized molecular variational quantum eigensolver states on IBM Aachen and Pittsburgh, two 156-qubit Heron processors (provider-reported revision 3). Only energy evaluation runs on hardware, with parameters, layouts and compiled circuits frozen and dynamical decoupling disabled. A five-molecule pilot, three-molecule isolated cohort in three windows and known-state controls comprise 110 jobs and 28.2 billed quantum processing unit (QPU) minutes; the isolated cohort spans 30.2 h within a 31.7 h campaign. Raw 1024 O2 and BeH2 mean residuals exceed returned uncertainty by factors 12-32. Pittsburgh BeH2 cross-job standard deviation is 5.6 times mean returned uncertainty. The zero-noise extrapolation (ZNE) package reduces O2 absolute residuals by 49.7% on Aachen and 83.5% on Pittsburgh at about fourfold charge. On Pittsburgh, twirled readout error extinction (TREX) removes 80 millihartree (mHa) per Raw 1024 job charge versus 50 mHa for ZNE. Controls reveal qubit-124 preparation or readout errors 5.4-5.6 times calibration; 13 mitigated energy point estimates fall outside the mapped spectrum, with no Raw counterpart. These results support workload checks and repeated measurements before additional sampling.

Abstract: arXiv descriptive metadata (CC0). Thank you to arXiv for use of its open access interoperability. This service was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.