What mitigation does in our hardware measurement
NISQ means that the processor remains susceptible to noise and does not provide full fault-tolerant protection for this calculation. We therefore combine several measures. IBM explains these techniques and their limitations.
- Measurement mitigation with TREX: randomized readout and calibration help correct systematic measurement bias.
- Pauli twirling: equivalent randomized gate variants make coherent errors easier to handle. They do not remove all noise.
- XY4 dynamical decoupling: X and Y pulses during suitable idle periods suppress certain disturbances. This is error suppression, not error correction.
- Zero-noise extrapolation, ZNE: we measure at noise factors 1.0, 1.2 and 1.4, then extrapolate linearly to the hypothetical zero-noise limit.
I(\lambda)\approx a+b\lambda,\qquad I_{\mathrm{ZNE}}\approx a.
\]
This formula illustrates the extrapolation principle; Runtime processing may operate on individual observable terms. The zero-noise value is not measured directly. Departures from the linear model can introduce bias, and extrapolation can amplify statistical noise. More shots do not automatically solve that modelling problem.
Our successful run used 2,048 shots per noise factor, distributed over 32 randomizations of 64 shots each. Mitigation and suppression are part of the result, not an optional footnote. The saved factor-1 results have an RMSE of approximately 0.0645 against the TN consensus; the full ZNE result has an RMSE of 0.03694. The other measures remain active at factor 1, so this is not a comparison with completely unmitigated hardware.