Part 5 of the Floquet-Ising series: from two qubits to prethermal oscillations
On an ideal quantum computer, every gate is exactly unitary. Real hardware adds relaxation, dephasing, coherent over- or under-rotation, leakage, crosstalk, and measurement errors. Error mitigation aims to reduce bias in an observable without the full logical error correction of a fault-tolerant computer.
A transparent toy noise model
In the two-qubit model, we place a local Pauli-depolarizing step after every gate. Schematically, for one qubit,
\[\mathcal{E}_p(\rho)=
(1-p)\rho+\frac{p}{3}
\left(X\rho X+Y\rho Y+Z\rho Z\right).
\]
The exact weights depend on the selected convention, but the idea is clear: a Pauli error is added with probability weight p. Because we know the full density matrix, we can evaluate the noise without sampling uncertainty.
This is not a model of a specific IBM calibration. Among other things, it omits temporal correlations, drift, leakage, and spatial crosstalk. Its simplicity is an advantage here: it reveals the mitigation arithmetic without a hidden fitting machine.
Zero-noise extrapolation
Suppose the expectation of an observable at scaled noise f decreases approximately exponentially:
\[M(f)=M(0)e^{-\gamma f}.
\]
Measure or calculate M(1) at the original noise level and M(2) at twice the effective noise. Then
\[\frac{M(1)^2}{M(2)}
=\frac{M(0)^2e^{-2\gamma}}{M(0)e^{-2\gamma}}
=M(0).
\]
This is the two-point exponential ZNE estimator. At cycle 6, the toy model gives:
| Route | Magnetization |
|---|---|
| Exact, noiseless | 0.8273524194 |
| Native noise f=1 | 0.6309694889 |
| Amplified noise f=2 | 0.4809840321 |
| Exponential ZNE | 0.8277249781 |
The correction is impressive, but largely a designed success: the simple channel follows the exponential ansatz very closely. On hardware, the wrong functional model can produce an attractive but biased zero-noise estimate.
Probabilistic error cancellation
PEC takes a different approach. If we know the noise channel accurately enough, we write its inverse as a quasiprobabilistic combination of physically executable operations:
\[\mathcal{E}^{-1}=\sum_j \eta_j\mathcal{B}_j.
\]
Some coefficients eta-j are negative. We sample operations according to the absolute weights and multiply the outcome by a sign and normalization factor. On average, the estimator can be unbiased, but its variance grows with the quasiprobability overhead.
PEC therefore trades bias for sampling cost. In deep circuits, this overhead can grow exponentially. A result without error bars or QPU costs tells only half the story.
Error suppression, readout mitigation, and mitigation
Different techniques should not be lumped together:
- Transpilation and dynamical decoupling change the circuit to suppress errors in advance.
- Gate twirling randomizes coherent error structure to create a more tractable effective channel.
- M3 readout mitigation calibrates measurement confusion and corrects measured marginals.
- ZNE extrapolates from multiple noise strengths.
- PEC uses a characterized quasiprobabilistic inverse.
In our IBM route, we combine a raw arm with an XY4 and twirling arm and use a shared balanced M3 calibration. For the global 51-bit distribution, 512 shots are far too sparse. M3 is therefore applied locally to the one- and two-qubit supports of magnetization and ZZ pairs.
Why shared calibration matters
When two measurement arms are corrected with different calibrations, calibration drift can resemble a physical difference. A shared M3 snapshot makes the comparison better paired. It does not remove all drift: raw, suppression, and calibration runs still occur at different times. The timestamps therefore belong to the scientific data.
The paper’s validation hierarchy
QESEM-Unbiased uses a PEC-like unbiased estimator and becomes statistically expensive at greater depth. QESEM-Extrapolated uses noise-amplified estimators and reaches later cycles with less overhead, but is more model-dependent. The paper validates the later route in the region where both methods still overlap.
Independent noise-amplification factors, hardware tests of the characterized noise model, early-cycle classical references, and selected Quantinuum results provide additional checks. No single control carries the entire claim.
The fairest conclusion from the toy model
The toy model shows exactly how noise damps an oscillation and how ZNE reconstructs the zero-noise value under the correct ansatz. It does not prove that the same ansatz is valid on a 51-qubit processor. Its scientific value lies in transparency: we know which assumption makes the correction work and can then test whether that assumption holds on hardware.
In part 6, we compare the scaling steps. This reveals why classical convergence, mitigation uncertainty, and resource accounting together determine whether the phrase quantum advantage is justified.
Sources
- Derivation and toy implementation.
- Z. Cai et al., Quantum Error Mitigation.
- E. Leviatan et al., QESEM validation hierarchy.