arxiv: 2604.15014 · v1 · submitted 2026-04-16 · 🪐 quant-ph

Recognition: unknown

Runtime-efficient zero-noise extrapolation from mixed physical and logical data

D. V. Babukhin , W. V. Pogosov

Authors on Pith no claims yet

Pith reviewed 2026-05-10 11:55 UTC · model grok-4.3

classification 🪐 quant-ph

keywords zero-noise extrapolationquantum error mitigationquantum error correctionRichardson extrapolationmixed physical-logical datavariance reductionIsing modelnear-term quantum computing

0 comments

The pith

Mixing a few error-corrected points with many uncorrected ones reduces runtime for zero-noise extrapolation by orders of magnitude.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that zero-noise extrapolation can be performed more efficiently by constructing datasets that combine a small number of low-noise logical points obtained with error correction and a larger number of higher-noise physical points obtained without it. The logical points stabilize the low-noise end of the extrapolation while the physical points cheaply extend the sampled noise range. Under the model in which correction multiplies the gate error rate by a constant factor gamma, the resulting variance of the Richardson-extrapolated estimator drops enough to cut total physical runtime by several orders of magnitude when gamma is 0.1 or less. A numerical demonstration on digital simulation of a six-spin transverse-field Ising model confirms that mixed datasets produce lower-variance zero-noise estimates than pure logical datasets in the regime examined. The work therefore frames partial error correction plus error mitigation as a concrete route to resource-efficient computation before full fault tolerance.

Core claim

The central claim is that zero-noise extrapolation from mixed physical and logical data reduces variance amplification relative to pure logical or pure physical data because the logical points anchor the fit at low noise while the physical points enlarge the baseline at far lower cost; under the model where logical error rate equals gamma times physical error rate, this yields a lower-variance estimator whose required physical runtime is smaller by several orders of magnitude for gamma less than or equal to 0.1, as derived analytically for Richardson extrapolation and verified on the six-spin transverse-field Ising model.

What carries the argument

Mixed physical-logical dataset for zero-noise extrapolation, in which logical points anchor the low-noise regime and physical points extend the noise baseline at reduced runtime cost.

If this is right

Variance of the zero-noise estimator is reduced compared with extrapolation from only logical or only physical data.
Physical runtime required to reach a target precision drops by several orders of magnitude when gamma is at most 0.1.
Mixed datasets outperform pure error-corrected extrapolation in the studied six-spin Ising model regime.
Hybrid error correction combined with error mitigation becomes a practical resource-efficient strategy before full fault tolerance.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The runtime savings may apply to other mitigation techniques that rely on extrapolation or fitting if similar anchoring effects hold.
Hardware-specific optimization of the ratio of logical to physical shots could yield further gains once actual runtime costs are measured.
The method invites direct tests on superconducting or trapped-ion processors that implement even modest error correction to confirm the constant-gamma assumption.

Load-bearing premise

Error correction reduces the effective gate error rate by a fixed factor gamma that stays constant regardless of circuit depth or other parameters.

What would settle it

An experiment that records the observed variance of the zero-noise estimator on a quantum device for mixed versus pure-logical datasets at a measured gamma near 0.1 and checks whether the runtime to target precision matches the predicted reduction.

Figures

Figures reproduced from arXiv: 2604.15014 by D. V. Babukhin, W. V. Pogosov.

**Figure 2.** Figure 2: FIG. 2: The layout of spin cluster [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4: Variance of the mean magnetization estimates [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3: Distribution of deviations of mean magnetiza [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

read the original abstract

Partial quantum error correction and quantum error mitigation are expected to coexist in the pre-fault-tolerant regime, yet the resource advantage of combining them remains insufficiently quantified. We study zero-noise extrapolation constructed from mixed datasets that contain a small number of error-corrected data points together with data obtained without error correction. The low-noise logical points anchor the extrapolation, while the higher-noise physical points enlarge the noise baseline at a much smaller runtime cost. Under a simple model in which error correction suppresses the effective gate error rate from p to $\gamma$p, we derive the variance of the zero-noise estimator and compare the physical runtime required to reach a target precision. For Richardson extrapolation, the mixed-data strategy reduces variance amplification and can lower the required physical runtime by several orders of magnitude when $\gamma \leq 0.1$. As a proof of principle, we apply the method to digital quantum simulation of a six-spin transverse-field Ising model and find that mixed physical/logical datasets yield lower-variance zero-noise estimates and outperform extrapolation based only on error-corrected data in the parameter regime studied here. These results identify hybrid error correction and error mitigation as a practical route to resource-efficient quantum computation before full fault tolerance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Mixing a few logical points with many physical ones cuts ZNE runtime under the constant-gamma model, with clean variance formulas but limited testing.

read the letter

The paper's core result is that for Richardson extrapolation you can anchor the fit with a handful of error-corrected logical data points and fill in the rest with cheaper physical points, and under their model this slashes the total physical runtime needed for a given precision by orders of magnitude when gamma is 0.1 or smaller. The variance derivation is straightforward once you accept the model, and the six-spin Ising run shows the scaling holds in that narrow case. What is new is the explicit construction of the mixed dataset together with the closed-form variance and runtime comparison; prior ZNE work did not combine physical and logical data this way with the resource accounting. The paper does a clean job of laying out when the hybrid approach wins on paper and of keeping the assumptions visible. The main soft spot is the constant-gamma assumption itself. The derivation treats gamma as fixed and independent of depth or circuit details, which is a strong simplification. The numerical test is too small and too idealized to check whether that holds on anything closer to real hardware, so the big runtime numbers remain model-dependent. No hidden circularity or fitting issues appear in the math. This is aimed at groups already running partial error correction and looking for ways to stretch those resources with mitigation. A reader working on NISQ-to-early-FT hybrids would get usable expressions and a clear baseline for further work. The paper deserves a serious referee because the analytic part is solid and the idea is practical even if the assumptions need more scrutiny in review.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes constructing zero-noise extrapolation (ZNE) from mixed physical and logical datasets, where a small number of error-corrected logical points anchor the extrapolation while physical points extend the noise baseline at lower runtime cost. Under an explicit model in which error correction reduces the effective gate error rate from p to a constant factor γp (independent of depth), the authors derive the variance of the Richardson-extrapolated zero-noise estimator and compare the physical runtime needed to reach a target precision. They report that the mixed strategy reduces variance amplification and can lower required runtime by orders of magnitude for γ ≤ 0.1. A numerical demonstration on digital simulation of a six-spin transverse-field Ising model is included as proof of principle, showing lower-variance estimates than pure physical or pure logical extrapolation in the studied regime.

Significance. If the constant-γ model approximates hardware behavior, the analytic variance derivation and runtime scaling provide a concrete quantitative basis for hybrid error-correction-plus-mitigation strategies in the pre-fault-tolerant regime. The derivation follows directly from the stated error model and the standard Richardson formula, the numerical Ising test is consistent with the predicted scaling, and the work supplies explicit runtime comparisons rather than qualitative arguments. These elements constitute a useful resource-estimation tool for near-term devices.

major comments (1)

[§II and variance derivation] §II (error model) and the subsequent variance derivation: the assumption that error correction suppresses the gate error rate to a constant factor γp independent of circuit depth is load-bearing for the claimed orders-of-magnitude runtime reduction under Richardson extrapolation. The manuscript derives the variance expressions and runtime advantage directly from this model, but provides no bounds, sensitivity analysis, or hardware evidence for depth independence beyond the six-spin toy example; if γ varies with depth the quantitative runtime savings would change.

minor comments (2)

[Abstract] The abstract states that mixed data 'can lower the required physical runtime by several orders of magnitude when γ ≤ 0.1' but does not quote the explicit scaling factor or the range of γ examined; adding one concrete numerical example would improve clarity.
[Numerical Results] In the numerical Ising-model section, the specific values of p and γ used in the simulation should be stated explicitly so readers can verify consistency with the analytic variance formulas.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading, positive assessment of the work's significance, and constructive comment. We address the major comment point by point below and have prepared revisions to strengthen the manuscript.

read point-by-point responses

Referee: [§II and variance derivation] §II (error model) and the subsequent variance derivation: the assumption that error correction suppresses the gate error rate to a constant factor γp independent of circuit depth is load-bearing for the claimed orders-of-magnitude runtime reduction under Richardson extrapolation. The manuscript derives the variance expressions and runtime advantage directly from this model, but provides no bounds, sensitivity analysis, or hardware evidence for depth independence beyond the six-spin toy example; if γ varies with depth the quantitative runtime savings would change.

Authors: We agree that the constant-γ assumption is central to the quantitative runtime comparisons derived in the paper. Section II explicitly introduces this as a minimal model in which error correction yields a fixed suppression factor γ (motivated by sub-threshold operation of quantum error-correcting codes, where logical error rates can be made independent of depth to leading order). All variance expressions and runtime scalings are derived directly under this model, as stated, and the six-spin transverse-field Ising numerical demonstration is presented only as a proof-of-principle consistent with the model's predictions. We do not claim hardware universality for the constant-γ regime. To address the referee's concern, the revised manuscript will include a new appendix with a sensitivity analysis. This analysis will consider a depth-dependent generalization γ(d) = γ₀(1 + α d) for small α > 0 and quantify how the runtime advantage of the mixed physical-logical strategy degrades (or persists) as a function of α and circuit depth. This will supply explicit bounds on the parameter regime in which the reported orders-of-magnitude savings remain valid. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central derivation consists of an explicit model assumption (error correction reduces effective gate error rate from p to constant factor γp, independent of depth) combined with the standard Richardson extrapolation formula to obtain closed-form variance expressions for the zero-noise estimator. These expressions are then used to compare physical runtime requirements for a target precision, which is an external input rather than a fitted quantity. The six-spin Ising numerical example is presented strictly as a proof-of-principle consistency check and does not supply fitted parameters that are later renamed as predictions. No load-bearing steps rely on self-citations, uniqueness theorems imported from prior author work, or ansatzes smuggled via citation; the model is stated upfront as an assumption and the algebra follows directly from it without reduction to the inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on a parameterized error-suppression model and the standard Richardson extrapolation procedure; no new physical entities are introduced.

free parameters (1)

γ
Constant factor by which error correction is assumed to reduce the effective gate error rate; central to the variance formula and runtime comparison.

axioms (1)

domain assumption Error correction reduces the effective gate error rate from p to γp with γ constant and independent of circuit parameters
Invoked throughout the variance derivation and runtime analysis.

pith-pipeline@v0.9.0 · 5509 in / 1336 out tokens · 45172 ms · 2026-05-10T11:55:21.291027+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 14 canonical work pages

[1]

We obtain two data point from qubits with quan- tum error correction codes, so the overall level of noise of each of data point is low. Let us say that the target circuit consists ofn, and each gate has an error probabilityγp, wherepis an error proba- bility of gate without error correction (on physical qubits) andγ <1 is a noise suppression prefactor fro...
[2]

correction - operation - correction

We obtain two data points, but now we use different types of qubits - one data point comes from qubits with quantum error correction, while the other data point comes from physical qubits. The two kind of data points have different overall noise levels: the error-corrected point has a levelλ A =γnpand the physical-qubits point has a levelλ B =np. The stan...
[3]

Geometric linear extrapolation As a first example demonstrating how the general for- mulation works, we use example we provided in Section III A (and depicted in Fig. 1). In this case we have two points of data (N= 2), with a single point with parti- cal quantum error correction (N 1 = 1). We have derived above that variance prefactor in case of error-cor...
[4]

global unitary folding

Richardson extrapolation Here we consider the case of Richardson extrapola- tion forK-th order polynomial (see Appendix B for fur- ther details). The variance of the zero-noise estimate for Richardson extrapolation reads V ar[θ0] = KX k=0 β2 kV ar[O(λ k)], β k = Y l̸=k λl λl −λ k (19) For the case of error-corrected data points we consider noise levelsλ k...
[5]

SinceA +A− =B +B−, thenA +O=OB − and A−O=OB +
[6]

Then we obtain△A +A ′ O∼ △B −B ′ Ofrom the equal- ity of two angles∠A −A+O=∠B ′ B−Oand ∠A+A ′ O=∠B −B ′ O

Let us draw a line through point O, perpendic- ular to the ordinate axis, and drop perpendicu- lars from pointsA − andB + onto it. Then we obtain△A +A ′ O∼ △B −B ′ Ofrom the equal- ity of two angles∠A −A+O=∠B ′ B−Oand ∠A+A ′ O=∠B −B ′ O. Since triangles△A +A ′ O and△B −B ′ Oare similar andA +O=OB −, we obtain△A +A ′ O=△B −B ′ O
[7]

SinceP +P− ∥A +A−, then∠P −P+O=∠A −A+O as the corresponding angles at parallel lines; 7.△P −P+O∼A −A+Oby angles∠P −P+O= ∠A−A+Oand∠P −OP+ =∠A −OA+
[8]

ThenA ′ Ois the height of triangleA −A+O, P ′ Ois the height of triangleP −P+O

PointP ′ is the intersection point of the line passing through pointOand perpendicular to the ordinate axis. ThenA ′ Ois the height of triangleA −A+O, P ′ Ois the height of triangleP −P+O. From the similarity of triangles, we obtain P−P+ A−A+ = P ′ O A ′ O (A1) SinceP −P+ = 2std[M(0)],A −A+ = 2σ,P ′ O= λB +λA 2 ,A ′ O= λB −λA 2 , we obtain an estimate for...
[9]

Demonstrating a universal logical gate set in error-detecting surface codes on a superconducting quantum processor,

J. Zhanget al., “Demonstrating a universal logical gate set in error-detecting surface codes on a superconducting quantum processor,” (2024), arXiv:2405.09035 [quant- ph]

work page arXiv 2024
[11]

Heet al., Phys

T. Heet al., Phys. Rev. Lett.135, 260601 (2025)

2025
[12]

Wanget al., Nature Physics22, 308–314 (2026)

K. Wanget al., Nature Physics22, 308–314 (2026)

2026
[13]

S. Endo, Z. Cai, S. C. Benjamin, and X. Yuan, Journal of the Physical Society of Japan90, 032001 (2021)

2021
[14]

Quantum error mitigation,

Z. Caiet al., Reviews of Modern Physics95(2023), 10.1103/revmodphys.95.045005

work page doi:10.1103/revmodphys.95.045005 2023
[15]

Navaset al.(Particle Data Group Collaboration), Phys

S. Navaset al.(Particle Data Group Collaboration), Phys. Rev. D110, 030001 (2024)

2024
[16]

https://doi.org/10.1103/PhysRevLett.127.200505

C. Piveteau, D. Sutter, S. Bravyi, J. M. Gambetta, and K. Temme, Physical Review Letters127(2021), 10.1103/physrevlett.127.200505

work page doi:10.1103/physrevlett.127.200505 2021
[17]

Lostaglio and A

M. Lostaglio and A. Ciani, Physical Review Letters127 (2021), 10.1103/physrevlett.127.200506

work page doi:10.1103/physrevlett.127.200506 2021
[18]

A framework of partial er- ror correction for intermediate-scale quantum comput- ers,

N. Koukoulekidis, S. Wang, T. O’Leary, D. Bultrini, L. Cincio, and P. Czarnik, “A framework of partial er- ror correction for intermediate-scale quantum comput- ers,” (2023), arXiv:2306.15531 [quant-ph]

work page arXiv 2023
[19]

Suzuki, S

Y. Suzuki, S. Endo, K. Fujii, and Y. Tokunaga, PRX Quantum3, 010345 (2022)

2022
[20]

M. A. Wahl, A. Mari, N. Shammah, W. J. Zeng, and G. S. Ravi, in2023 IEEE International Conference on Quantum Computing and Engineering (QCE)(IEEE,
[21]

S. C. Smith, B. J. Brown, and S. D. Bartlett, Commu- nications Physics7(2024), 10.1038/s42005-024-01883-4

work page doi:10.1038/s42005-024-01883-4 2024
[22]

Error mitigation for partially error-corrected quantum computers.arXiv preprint arXiv:2510.10905, 2025

B. DalFavero and R. LaRose, “Error mitigation for partially error-corrected quantum computers,” (2025), arXiv:2510.10905 [quant-ph]

work page arXiv 2025
[23]

Zhanget al., Nature Communications17(2025), 10.1038/s41467-025-67768-4

A. Zhanget al., Nature Communications17(2025), 10.1038/s41467-025-67768-4

work page doi:10.1038/s41467-025-67768-4 2025
[24]

Quantum Error Correc- tion on Error-mitigated Physical Qubits,

M. Jeon and Z. Cai, “Quantum error correction on error- mitigated physical qubits,” (2026), arXiv:2601.18384 [quant-ph]

work page arXiv 2026
[25]

Giurgica-Tiron, Y

T. Giurgica-Tiron, Y. Hindy, R. LaRose, A. Mari, and W. J. Zeng, in2020 IEEE International Conference on Quantum Computing and Engineering (QCE)(IEEE,
[26]

arXiv:2510.26300 (2025)

F. Alamet al., “Fermionic dynamics on a trapped-ion quantum computer beyond exact classical simulation,” (2025), arXiv:2510.26300 [quant-ph]

work page arXiv 2025
[27]

Ransfordet al., Helios: A 98-qubit trapped-ion quantum computer (2025), arXiv:2511.05465 [quant-ph]

A. Ransfordet al., “Helios: A 98-qubit trapped-ion quan- tum computer,” (2025), arXiv:2511.05465 [quant-ph]

work page arXiv 2025
[28]

Granet, S.-H

E. Granetet al., “Superconducting pairing correla- tions on a trapped-ion quantum computer,” (2026), arXiv:2511.02125 [quant-ph]

work page arXiv 2026
[29]

I. D. Kivlichanet al., Quantum4, 296 (2020)

2020
[30]

S. J. Everedet al., Nature622, 268–272 (2023)

2023
[31]

Acharyaet al., Nature638, 920–926 (2024)

R. Acharyaet al., Nature638, 920–926 (2024)

2024
[32]

Myths around quantum computation before full fault tolerance: What no-go theorems rule out and what they don’t.arXiv preprint arXiv:2501.05694, 2025

Z. Zimbor´ aset al., “Myths around quantum computation before full fault tolerance: What no-go theorems rule out and what they don’t,” (2025), arXiv:2501.05694 [quant- 10 ph]

work page arXiv 2025
[33]

A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland, Physical Review A86(2012), 10.1103/phys- reva.86.032324

work page doi:10.1103/phys- 2012
[34]

S. J. Devitt, W. J. Munro, and K. Nemoto, Reports on Progress in Physics76, 076001 (2013)

2013