REVIEW 2 major objections 5 minor 8 references
Label and Recover Coherent Errors: Randomized Compiling Does Not Destroy Coherent-Error Information
T0 review · 2 major / 5 minor · reviewed 2026-07-30 · grok-4.5
Pith's one-line read Randomized compiling does not erase coherent-error information; it moves that information into the discarded twirl labels, which recover the errors at the quantum limit with no extra circuits.
desk verdict Clean Fisher conservation argument that RC relocates coherent-error information into the discarded twirl labels, with a zero-extra-cost estimator that works on hardware for injected phases. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Fisher-information conservation law (Theorem 1): the joint Fisher information of outcomes and labels equals the conditional Fisher information and splits as F_marg + Δ, where Δ ≻ 0 is exactly the information recoverable from the labels; at full twirl F_marg vanishes while Δ saturates the quantum bound.
What would settle it
Inject a known coherent phase on a device, run standard randomized compiling while logging the per-shot Pauli labels, and check whether the label–outcome estimator recovers the injected value to within shot noise while the label-discarding marginal estimator remains consistent with zero.
Extended reading notes
Core claim
The coherent-error information that randomized compiling removes from the averaged output distribution is relocated, not destroyed, into the per-shot twirl labels. Retaining those labels and forming a simple label–outcome correlation recovers the coherent parameters at the quantum Fisher-information limit, unbiased under standard incoherent channels, with no additional circuits.
Load-bearing premise
The random twirl labels must be chosen independently of the unknown coherent errors, and the signal shrinkage factor must be readable without bias from the same run’s ordinary averages.
Editorial extensions
If this is right
- Any group already running randomized compiling can extract a coherent-error map simply by logging the labels it currently discards.
- One RC dataset simultaneously yields stochastic Pauli rates from the marginal and coherent angles from the labels.
- Labeled Fisher information grows linearly with circuit depth and can surpass the bounded accumulation of an untwirled coherent circuit.
- The recovered angles can be fed back as compile-time virtual-Z corrections without new characterization experiments.
- Twirl-group design controls which linear combinations of coherent errors remain separately identifiable.
Reading between the lines
- Existing open RC and Pauli-learnability datasets could be re-analyzed offline to produce coherent-error maps that were never intended by the original experiments.
- Adaptive choice of the twirl ensemble could be used to isolate crosstalk subspaces that ordinary full twirling mixes together.
- If label logging becomes standard, routine RC runs double as continuous, zero-overhead coherent calibration for the whole device.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript argues that randomized compiling does not destroy coherent-error information but relocates it into the discarded twirl labels. Theorem 1 establishes an exact Fisher-information conservation law: F_cond = F_marg + Δ with Δ ⪰ 0 equal to the label-recoverable information, reducing at full twirl to F_marg = 0 and Δ = F_0. A complementarity relation (Eq. 5) makes this explicit in the tailoring parameter T. Theorem 2 gives a PTM-based noise-aware estimator θ̂ = arcsin(E[s·b]/v) that is unbiased once the visibility v = λ_Y is read from the same-run marginal. The conservation identity is checked to machine precision (bias ≤ 2×10^{-8}) across 12 circuit families, and a controlled injection experiment on ibm_marrakesh recovers θ* to within 0.0063 rad while the marginal estimator remains near zero at all depths.
Significance. If correct, the result reframes a standard premise of RC and yields a zero-extra-circuit coherent-error map from data already collected in any RC run. The derivation uses only the θ-independence of the twirl and the chain rule, so it is not model-fragile within that setting. Strengths that raise confidence include: (i) machine-checked numerical conservation across diverse circuits with reported bias six orders below shot noise; (ii) an explicit unbiased estimator under the stated PTM noise class with self-consistent visibility; (iii) a falsifiable hardware test in which the marginal null and labeled recovery are shown side-by-side; and (iv) public simulation code and oracles. The work is a natural coherent-sector complement to Pauli-noise learnability and is of immediate practical interest to groups already running RC.
major comments (2)
- [Abstract; Theorem 1; Depth Scaling; Outlook] Abstract and Main Result claim recovery 'at the quantum Fisher-information limit,' and Fig. 1(b) plots E_s[F_Q]. Theorem 1 rigorously conserves the classical joint/conditional Fisher information of (b,s). Equality with the quantum FI is justified in the pure/ideal generator calculation (Depth Scaling, Eq. 9) but is not proved for the noisy CPTP setting of Theorem 2; the Outlook itself lists a Cramér–Rao converse for Eq. (8) as future work. Please restrict the QFI-limit language to the regime where it is shown (ideal or pure-state generators), or supply the missing argument that the labeled classical FI saturates the QFI under the noise model of Theorem 2. This is a wording/scope fix, not a challenge to conservation or unbiasedness.
- [Theorem 2; Eq. (8); Abstract] Theorem 2 and consequence C assert an exactly unbiased estimator under 'all standard incoherent channels' via v = λ_Y. The derivation uses the Pauli-twirled adjoint form E^†_tw(Y) = t_Y I + λ_Y Y. Unmodeled coherent leakage outside the twirled Pauli model, strong non-unitality beyond the offset cancellation, or SPAM that correlates with labels could bias v or E[s·b]. The hardware run reports v ≈ 0.993 and sub-shot-noise recovery for injected R_z errors, which supports the claim in that instance, but the manuscript should state the precise noise class (Pauli-twirled CPTP with label-independent SPAM) under which unbiasedness is guaranteed, rather than the broader phrase in the abstract.
minor comments (5)
- [Fig. 1(a); Eq. (5)] Fig. 1(a) caption and Eq. (5) give leading-order (in θ) complementarity; the exact conservation F_cond = F_marg + Δ holds beyond that order. A sentence distinguishing exact conservation from the first-order split in T would avoid confusion.
- [Identifiability Matrix; Results] Identifiability matrix F ∘ Σ (Eq. 6) is important for multi-parameter or correlated twirls but is not exercised in the hardware experiment (single injected phase). A brief remark on how the ibm_marrakesh design sits inside that framework would help.
- [Table I] Table I reports |θ̂ − θ*| but not uncertainties or the number of shots entering each row; adding standard errors (or noting they match the stated shot-noise floor) would make the table self-contained.
- [Abstract; Introduction] Typographical inconsistencies: 'twirldestroys' / 'twirl labels—the' spacing in the abstract/intro, 'ibm marrakesh' vs 'ibm_marrakesh', and 'Cram´ er–Rao' accent encoding. Also 'The twirldestroyscoherent-error information' in the Introduction.
- [Depth Scaling; Eq. (9)] The claim that the labeled estimator 'eventually surpasses coherent accumulation on generic circuits' (Depth Scaling) is supported numerically in Fig. 1(b) for α = 1; a one-line condition on α (already partly stated) next to that sentence would clarify the analytic scope.
Circularity Check
No significant circularity: conservation law and recovery estimator are independent Fisher/PTM identities validated against injected ground truth
full rationale
Theorem 1 follows from the classical score identity (π independent of θ by experimental design) and the chain-rule decomposition of joint Fisher information; F_cond = F_marg + Δ is an identity, not a fit, and complementarity (Eq. 5) is a first-order expansion of that identity. Theorem 2’s noise-aware estimator uses visibility v = λ_Y read from the same-run marginal PTM eigenvalue—an independent observable under the stated CPTP/Pauli-twirl model—not a parameter fitted to the coherent label–outcome signal being recovered. Hardware recovery compares the labeled estimator to known injected θ* (Table I, Fig. 1c); the marginal returning ~0 is a null check, not a circular self-prediction. Citations are to external RC/GST/learnability literature (Wallman–Emerson, Hashim, Chen, Nielsen, Zhang); there is no author-overlapping uniqueness theorem or ansatz smuggled in as load-bearing support. Simulation bias ≤2e-8 across 12 families is an independent numeric check of the same identities. Nothing in the derivation chain reduces a claimed prediction to its own fitted input by construction.
Assumptions & free parameters
free parameters (2)
- injected coherent phase θ* =
0–0.3 rad
- visibility v = λ_Y =
≈0.993 on hardware
assumptions (4)
- domain assumption Twirl-label distribution π(s) is independent of coherent-error parameters θ
- standard math Classical Fisher information chain rule and score function properties
- domain assumption Pauli-twirled adjoint channel retains only non-unitality t_Y and diagonal λ_Y for the relevant observable
- domain assumption Full twirl has E[s] = 0 so that label-independent offsets cancel in E[s·b]
Cite this review
Pith. "Pith review of Label and Recover Coherent Errors: Randomized Compiling Does Not Destroy Coherent-Error Information." pith.science (2026). https://pith.science/paper/UZADFIBY
@misc{pith2026260726756,
author = {Pith},
title = {Pith review of: Label and Recover Coherent Errors: Randomized Compiling Does Not Destroy Coherent-Error Information},
year = {2026},
howpublished = {\url{https://pith.science/paper/UZADFIBY}},
note = {Machine review of arXiv:2607.26756}
}
read the original abstract
Randomized compiling (RC) is the standard technique for converting coherent (systematic) gate errors into stochastic noise. The prevailing view is that the twirl destroys coherent-error information. An exact Fisher-information conservation law shows the opposite: the coherent-error information RC removes from the averaged output is preserved in full in the twirl labels -- the per-shot random gate choices that standard RC discards. Retaining the labels and forming a label-outcome correlation recovers the coherent error parameters at the quantum Fisher-information limit, unbiased under all standard incoherent channels, at zero additional circuit cost. Two theorems are proved, the conservation law is verified to machine precision across 12 circuit families, and recovery is confirmed on a 127-qubit IBM Quantum processor (ibm_marrakesh). The labeled estimator recovers injected coherent phases to within 0.0063 rad of the true value across all depths tested, while the standard marginal estimator returns near-zero signal at every depth.
Figures
Reference graph
Works this paper leans on
-
[1]
Score identity.Since π is θ-independent, ∂i ln[π(s)P(b|s, θ)] =∂i lnP(b|s, θ), so Fjoint(θ) =F cond(θ).(2)
-
[2]
Hence Fjoint =F marg + ∆,∆ =E b[Fs|b]⪰0.(3) 3.Conservation.Combining Eqs
Chain rule.Factoring P (b, s|θ) = ¯P (b|θ)P (s|b, θ), the cross-term vanishes because Es|b[∂j lnP(s|b, θ)] = 0. Hence Fjoint =F marg + ∆,∆ =E b[Fs|b]⪰0.(3) 3.Conservation.Combining Eqs. (2) and (3): Fcond =F marg + ∆⪰F marg. (4) ∆ ⪰ 0 is the label-recoverable Fisher information. RC keeps Fmarg and discards ∆; retaining labels recovers ∆ exactly; see Eq. (...
arXiv 2026
-
[3]
Noise tailoring for scal- able quantum computation via randomized compiling,
J. J. Wallman and J. Emerson, “Noise tailoring for scal- able quantum computation via randomized compiling,” Phys. Rev. A94, 052325 (2016). DOI: 10.1103/Phys- RevA.94.052325
doi:10.1103/phys- 2016
-
[4]
Randomized compiling for scalable quantum computing on a noisy superconducting quan- tum processor,
A. Hashimet al., “Randomized compiling for scalable quantum computing on a noisy superconducting quan- tum processor,” Phys. Rev. X11, 041039 (2021). DOI: 10.1103/PhysRevX.11.041039
-
[5]
The learnability of Pauli noise,
S. Chen, Y. Liu, M. Otten, A. Seif, B. Fefferman, and L. Jiang, “The learnability of Pauli noise,” Nat. Commun. 14, 52 (2023). DOI: 10.1038/s41467-022-35759-4. Dataset: github.com/csenrui/Pauli Learnability
-
[6]
E. Nielsenet al., “Gate set tomography,” Quantum5, 557 (2021). DOI: 10.22331/q-2021-10-05-557
-
[7]
Hidden inverses: Coherent error cancella- tion at the circuit level,
B. Zhanget al., “Hidden inverses: Coherent error cancella- tion at the circuit level,” Phys. Rev. Applied17, 034074 (2022). DOI: 10.1103/PhysRevApplied.17.034074
-
[8]
Label and recover coherent errors — simula- tion code and data,
K. Vyas, “Label and recover coherent errors — simula- tion code and data,” github.com/kushcoder12/label-and- recover-coherent-errors (2026)
2026
Reviewed July 30, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.