Nadarasa · reference note · G19
G19 — Prethermal Floquet dynamics and cross-platform error mitigation
Leviatan et al., arXiv:2607.24937 (July 2026). A Floquet mixed-field Ising magnet on the IBM heavy-hex lattice is driven up to 74 qubits, mitigated with QESEM, and cross-checked on Quantinuum H2 and Helios. The result is a worked example of how error-mitigated quantum processors can become quantitative instruments for non-equilibrium physics beyond the current classical frontier.
The Floquet circuit
One Floquet cycle is a hardware-native sequence of single-qubit rotations interleaved with three edge-coloured layers of nearest-neighbour ZZ rotations:
U_F = ∏_{r=1}^{3} e^{-i θ_{zz} C_r / 2} · e^{-i θ_z Z_Σ / 2} · e^{-i θ_x X_Σ / 2}
where C_r sums Z_i Z_j over the edges of colour r, and the chosen angles (θ_x ≈ π/6, θ_z ≈ π/27, θ_{zz} ≈ 1.05) are picked so the system rapidly relaxes and then enters a long-lived prethermal window. The heavy-hex connectivity lets every ZZ layer run in parallel within one native gate layer.
What prethermal means here
After a short transient, the magnetization settles into oscillations with a period roughly four times the Floquet drive, around a non-trivial prethermal value. This is not an integrable or many-body-localized mechanism; the system is generic, but thermalization is delayed long enough for structured local observables to remain visible while entanglement and operator complexity grow. Eventually the drive heats the system to an effectively infinite-temperature state.
The interesting physics is in the amplitude decay of these oscillations with system size. Classically simulable systems are too small to tell whether the oscillations survive in the thermodynamic limit; the quantum data extends finite-size scaling to 74 qubits and suggests the response persists.
QESEM mitigation stack
QESEM is a software framework that uses the same quasiprobability noise characterization to run two different estimators:
- QESEM-Unbiased — probabilistic error cancellation (PEC), giving an unbiased ideal expectation value at the cost of higher sampling overhead.
- QESEM-Extrapolated — zero-noise extrapolation (ZNE) from noise-amplified circuits, often lower overhead but heuristic.
The two estimators agree over their common time window, which is a strong internal consistency check. The paper reports percent-level precision on the magnetization at depths where unmitigated execution loses signal by cycle 4.
Validation hierarchy
The paper does not rely on a single mitigation trick. The reliability argument is layered:
- Noise-model validation on the IBM superconducting QPU itself — the characterized noise model is checked against independent calibration data.
- Estimator agreement — PEC and ZNE, two conceptually different mitigations, give consistent results where they overlap.
- Classical comparison — exact state-vector and small-scale tensor-network checks are used where converged.
- Cross-platform corroboration — selected Floquet cycles are re-run on Quantinuum System Model H2 and Quantinuum Helios without using the IBM noise model, and the extrapolated results agree.
Mapping to Nadarasa methodology
| Paper layer | Leviatan et al. tool | Nadarasa equivalent |
|---|---|---|
| Ideal benchmark | Exact state-vector / small TN | G16 noiseless QPDE gap fit |
| Noise model | QESEM characterized noise model | Depolarizing ladder anchored on H2 p_2q |
| Extrapolation | QESEM-Extrapolated (ZNE) | Richardson ZNE on G17 ladder |
| Independent estimator | QESEM-Unbiased (PEC) | Model-free curve χ² vs Taylor-fit moments |
| Cross-platform | IBM Heron ↔ Quantinuum H2/Helios | Selene emulator only; hardware cross-check is open |
The G17/G18 work already uses a smaller version of the same logic: an exact ideal baseline, a depolarizing noise ladder, Richardson ZNE, and a model-free curve check that independently guards the Taylor-fit moments. Leviatan et al. show what that same ladder looks like at 74-qubit scale and with two independent mitigation estimators.
Why it matters for the frontier
This is a quantitative-science result, not just a hardware demo. The paper's classical comparison is explicit: leading tensor-network methods (PEPS-BP on H100) and sparse Pauli-path simulations (on Fugaku) do not converge for the large-system, late-cycle dynamics that the quantum data resolves. The quantum-accessible regime is therefore not merely a benchmark of qubit count; it is a region where the error-mitigated processor becomes the controlled probe.
For Nadarasa, the takeaway is methodological: the validation hierarchy (noise model → estimator agreement → classical comparison → cross-platform) is portable. Future Selene experiments that want to claim a result beyond classical reach should budget for at least two of these layers and treat cross-platform corroboration as the strongest one.
Caveats
- This is a reading note, not a reproduction. No Python kernel or Selene data is generated here.
- Heavy-hex is IBM-native geometry; translating the same Floquet circuit to Quantinuum H2 would require a routing step, but the paper shows that the physics survives the translation for selected cycles.
- The subharmonic oscillation is not a discrete time crystal; it is a prethermal transient with a finite lifetime set by the heating rate.