Statistical error bars on quantum computer calculations have been unreliable indicators of accuracy. Calculations for oxygen and beryllium hydride molecules show raw results can exceed reported uncertainties by a factor up to thirty-two; this highlights limitations in current methods used to quantify calculation errors. The Fraunhofer Institute has revealed that standard methods of assessing accuracy in quantum computer calculations can be sharply misleading.

Analyses on oxygen and beryllium hydride molecules showed actual calculation errors were between twelve to thirty-two times larger than initially reported; this means the statistical error bars currently provided do not reliably reflect true uncertainty. Consequently, more key techniques are needed to validate computations before applying them to complex scientific problems. Recent analysis by researchers at the Fraunhofer Institute reveals that standard error estimations in quantum computer calculations may be sharply misleading.

Examining oxygen and beryllium hydride molecules using’s Heron processors, the team found actual calculation errors were twelve to thirty-two times larger than those initially reported; this casts doubt on the reliability of current uncertainty quantification methods. This is akin to measuring a distance in millimeters but believing your margin of error represents miles, highly precise on paper, yet fundamentally inaccurate when applied to real-world scales.

The scientists employed a technique called zero-noise extrapolation (ZNE), which involves running calculations with increasing levels of simulated noise then extrapolating back to estimate what the result would be without any disturbance, similar to extending a trendline beyond existing data points. But even with mitigation techniques like ZNE and twirled readout error extinction, discrepancies remain, and some mitigated results fall outside physically plausible ranges.

Twirled Readout Extinction Surpasses Zero-Noise Extrapolation in Error Mitigation Performance

Pittsburgh processors saw twirled readout error extinction (TREX) remove up to 80 millihartree per job charge, a substantial improvement over the 50 mHa achieved by zero-noise extrapolation (ZNE) on the same hardware, representing nearly a 60% increase in mitigated error reduction for certain workloads. Previously impossible without sharply increasing computational cost or accepting potentially spurious results, such targeted error removal is now within reach. While ZNE reduced oxygen absolute residuals by 49.7% on Aachen and 83.5% on Pittsburgh, TREX offered further gains dependent upon system configuration and workload demands.

Mean residuals from both Aachen and Pittsburgh processors exceeded returned uncertainty by factors of 12 to 32 for oxygen and beryllium hydride molecules, highlighting significant unmitigated errors. BeH2 calculations on Pittsburgh exhibited a cross-job standard deviation of 120.4 mHa compared to a mean returned uncertainty of just 21.64 mHa, a factor of 5.6 difference.

However, controls exposed qubit-124 preparation or readout errors at levels five to six times higher than calibration values; thirteen mitigated energy estimates fell outside the expected physical range, something not observed with raw data alone, suggesting mitigation isn’t yet fully reliable across all system states.

Statistical Uncertainty Fails as an Error Proxy for Noisy Quantum Computations

A key distinction in striving for reliable results from today’s noisy machines is that reducing statistical uncertainty doesn’t automatically translate to more accurate quantum calculations. Thorough analysis by researchers at the Fraunhofer Institute reveals a troubling inconsistency between reported uncertainties and actual calculation errors when evaluating molecular energies on Aachen and Pittsburgh processors. This finding challenges reliance solely on minimising statistical fluctuations as a measure of computational reliability, raising questions about how best to validate results from near-term devices before applying them to complex scientific problems.

The team provides important guidance for workload checks and repeated measurements improving reliability going forward by pinpointing discrepancies in calculations of systems like beryllium hydride and lithium hydride. Simply reporting lower statistical uncertainty does not guarantee more accurate quantum computations; rigorous validation is vital. The investigation confirmed significant differences, up to thirty-two times larger than expected, between predicted error levels and observed residuals even with techniques such as zero-noise extrapolation and twirled readout error extinction. This underscores the need to carefully assess reported measures on near-term devices to ensure reliable results are obtained when tackling complex scientific problems.

Researchers found that decreasing statistical uncertainty in molecular energy calculations performed on IBM’s Aachen and Pittsburgh processors did not consistently improve accuracy. Analysis of 110 jobs involving five molecules revealed discrepancies, reaching factors of 12, 32, between returned uncertainties and actual calculation errors. These findings demonstrate that current methods for mitigating quantum noise do not always produce more reliable estimates, and highlight the importance of thorough workload checks alongside repeated measurements before drawing conclusions from noisy intermediate-scale quantum computations.

👉 More information
🗞 Execution bias, repeatability and mitigation cost in parameter-frozen molecular VQE energy evaluation
✍️ Julen Larrucea (Affiliation: Fraunhofer Institute for Manufacturing Technology and Advanced Materials IFAM)
🧠 ArXiv: https://arxiv.org/abs/2610.38885

More like this