Estimation versus computation
An estimation algorithm reconstructs a noiseless number after the chip has run. A computation returns the result of the run itself. Those are different architectures. Mixing them is how a heuristic reconstruction gets sold as a proved quantum result.
Thesis. Estimation is a reconstruction method (noise model, quasi-probabilities, extrapolation). Computation is what the quantum job returns when the architecture has run. The parade below is estimation sold as proved physics. QPC desks compute. Transjunctions are computed. The job readout is the result.
1. The parade: IBM + Qedma, Floquet, QESEM
The claim traces to arXiv:2607.24937 (Resolving Structure in Prethermal Floquet Dynamics with Precision Quantum Computation, 27 July 2026), announced with IBM’s “quantum advantage era” papers. Authorship is overwhelmingly Qedma, with IBM Quantum co-authors, RIKEN, and BlueQubit.
Hardware. ibm_boston, IBM Heron r3. 51- and 74-qubit heavy-hex patches,
not the full 156-qubit chip. Selected 51-qubit cycles were also reproduced on Quantinuum H2 / Helios.
Circuit. A hardware-native Floquet mixed-field Ising magnet: three edge-coloured layers of fractional-angle ZZ per cycle, about 30 cycles, about 90 entangling layers. The observable is a low-weight local magnetization (site-averaged ⟨Z⟩ on coordination-two sites), with a bounded active volume.
Software. Qiskit circuits into Qedma’s QESEM (a Qiskit Function). Two modes:
- QESEM-Unbiased: probabilistic error cancellation on a characterized sparse Pauli–Lindblad model. The accuracy guarantee attaches here, in the infinite-shot limit, given that the noise model describes the device. Sampling cost grows exponentially with active volume × error rate. That wall is why unbiased execution stopped.
- QESEM-Extrapolated: zero-noise extrapolation via probabilistic error amplification and an exponential ansatz. Cheaper. Heuristic. The authors state it can introduce extrapolation bias.
2. Where the guarantee stopped
Unmitigated execution loses accuracy by cycle 4. QESEM-Unbiased agrees with converged classical references to roughly cycle 11 and remains statistically feasible to cycle 16. From cycle 16 to cycle 30 the numbers are QESEM-Extrapolated: the heuristic estimator, in the regime where no classical check exists.
That is the crux. The estimator with the guarantee covers only the window where a classical answer also exists. The window with no classical answer is covered by the estimator without the guarantee. The unbiased method never produced a number outside the classically checkable range. The novel-physics stretch (cycles 16–30) is 100% extrapolated reconstruction.
Eleven months earlier, Qedma’s methods paper (arXiv:2508.10997) drew the same line: QP-based QESEM can produce unbiased expectation values; ZNE belongs to heuristic methods that generally lack accuracy guarantees. On a trapped-ion Ising case they argued that when ZNE disagrees with classical simulation, one cannot tell which result is correct, and that is why bias-controlled mitigation is necessary. Cycles 16–30 of the 2026 Floquet paper sit in that same epistemic situation. Classical methods disagree; the arbiter is the heuristic.
3. What was claimed, and what was delivered
IBM’s press line flattened this into trusted / advantage-era quantum computation. What was delivered is a reconstruction pipeline:
- A co-designed lattice instance (heavy-hex, native fractional RZZ, parameters scanned to keep signal while entanglement is high).
- A local observable cheap enough to mitigate. Applications that need many observables, chemical energies, or samples from the full output face a harsher overhead.
- A thermodynamic-limit amplitude fit that moves with two extra patch sizes, from a biased estimator, reported at 1σ, on an oscillation of about 1% of full magnetization.
- Validation of the extrapolated ansatz inside the window where the stretch is small, then use of that ansatz where the stretch is large (cycle 30, ~90 entangling layers). That is the textbook extrapolation failure mode.
Qedma’s own 2025 methods paper already said error mitigation on its own cannot yield asymptotic exponential advantage. Quek, Stilck França, Khatri, Meyer and Eisert (Nature Physics) showed that mitigating even slightly beyond constant depth requires super-polynomial samples in the worst case. The unbiased wall at cycle 16 is that physics, not a scheduling choice.
Qedma’s accuracy guarantee is real, and it terminates at cycle 16. Cycles 16–30, which carry the novel-physics claim, are produced by a heuristic extrapolation. What was demonstrated beyond the classically checkable window is a reconstruction that behaves consistently under self-consistency checks. That is evidence about an estimator. It is a different object from a computation that returns its own result.
4. What QPC computes
QPC desks are a different architecture. Several labelled logics occupy one quantum job. Transjunctions couple them while the computation runs. The Sampler returns the job’s counts. Those counts are the output of that computation: rankings, junction gaps, co-resident decisions. Transjunctions are computed on the QPU. The result at the end of the process is the result of the process.
Named public instance: ADMET Lead Desk Stage C2 on
IBM Heron r2 · ibm_marrakesh · 156 qubits · 8192 shots.
Seventeen co-resident ADMET logics. Locked junction_ryy.
Raw junction gap 0.185 · Aer retention 0.749 · separable contrast 0.184.
Jobs d9qubi7pemts73crsa2g+ / separable d9qubk8pdb6s73e45q3g.
Primary architecture numbers on raw counts.
Resilience, when used, is a labelled diagnostic channel on the same jobs.
Frozen tables, locked geometry, pre-registered bars, and public Runtime IDs make the computation auditable. Buyers can weigh cost, organizing, and time: one structured architectural job family versus a characterization-plus-reconstruction stack whose guarantee stops where classical checks stop.
5. Closing
Estimation answers: given this noise model and this reconstruction method, what noiseless number do we infer?
Computation answers: given this architecture on this chip, what did the job return?
The Floquet parade is the first. QPC desks are the second. Keep the names on the right objects.
Sources
- Lindner, Aharonov, van den Berg, Seif, Kandala et al. Resolving Structure in Prethermal Floquet Dynamics with Precision Quantum Computation. arXiv:2607.24937.
- Qedma. Reliable high-accuracy error mitigation for utility-scale quantum circuits (QESEM). arXiv:2508.10997.
- Quek, Stilck França, Khatri, Meyer, Eisert. Exponentially tighter bounds on limitations of quantum error mitigation. Nature Physics. arXiv:2210.11505.
- QPC ADMET WOW. Architecture scorecard with public job IDs.
- QPC coherent architecture. Desks, transjunctions, raw architecture metrics.