Reliable Adaptive Stopping for Krylov-Shadow Quantum Fisher Information Estimation

Erjie Liu; Yangshuai Wang

Reliable Adaptive Stopping for Krylov-Shadow Quantum Fisher Information Estimation

T0 review · 1 major / 0 minor · reviewed 2026-06-30 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read AKS-QFI separates Krylov truncation from sampling error to avoid false stopping decisions in shadow QFI estimation.

desk verdict The paper frames adaptive stopping for Krylov-shadow QFI as two separate reliability components and shows lower false-stop rates than width-only rules on an n=4 benchmark, but the recalibration step lacks a clear ground-truth-free procedure. read the letter →

arxiv 2605.14338 v2 pith:MFQGOT6M submitted 2026-05-14 quant-ph

Erjie Liu , Yangshuai Wang This is my paper · ORCID

classification quant-ph

keywords quantumFisherinformationKrylov-shadowestimationadaptivestoppingshadowtomographyQFImixed-statebenchmarkerrorseparationfinite-sampleuncertainty

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Krylov-shadow QFI estimation combines two resource choices: Krylov order sets how finely the state populations are resolved, while sample count sets the statistical precision at that resolution. Treating the combined width as a single stopping signal can produce false stops, in which the rule reports a tight interval around a biased low-order estimate. The paper reframes adaptive stopping as a two-component reliability task that isolates truncation error from sampling uncertainty and introduces AKS-QFI to monitor and recalibrate each component separately. On a noisy four-qubit mixed-state benchmark, width-only rules produce false success declarations between 16 and 68 percent of the time; AKS-QFI produces none under identical resource limits and, after independent recalibration, correctly declares success at the true 5 percent relative tolerance.

What carries the argument

AKS-QFI, the component-aware adaptive stopping interface that decouples Krylov truncation error from finite-sample uncertainty for Krylov-shadow QFI estimators.

What would settle it

On a system whose exact QFI is known by other means, run AKS-QFI until it declares success at the claimed tolerance and verify whether the actual relative error is below 5 percent while width-only stopping declares success on estimates whose true error exceeds that threshold.

Watch

Extended reading notes

Core claim

Krylov-shadow QFI estimation has two independent resource directions: Krylov order controls population resolution while sample count controls statistical uncertainty. Treating them as a single width produces false stops on biased estimates. AKS-QFI treats adaptive stopping as a reliability problem that separates truncation error from sampling error and recalibrates each component independently, yielding zero false success declarations on the benchmark and accurate success at 5% tolerance after recalibration.

Load-bearing premise

The two error sources of Krylov truncation and finite-sample uncertainty can be separated and recalibrated independently without the recalibration step introducing bias or requiring knowledge of the true QFI value.

Editorial extensions

If this is right

Width-only stopping rules can report narrow intervals around biased low-order QFI estimates.
AKS-QFI returns no false success declarations under the same resource limits where width-only rules fail 16 to 68 percent of the time.
Independent recalibration of Krylov resolution and sample counts produces accurate success declarations at true 5 percent relative tolerance.
Adaptive stopping functions as a reliability layer for shadow-based QFI estimation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The separation of truncation and sampling errors may extend to other shadow protocols that combine approximation order with Monte Carlo sampling.
Composite width metrics are likely insufficient for reliable stopping whenever multiple distinct error sources are present in quantum estimation tasks.
The recalibration procedure could be tested on systems larger than four qubits to check whether the independence assumption continues to hold.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The paper's core claim is that treating Krylov order and sample count as a single resource in shadow QFI estimation can produce false stops on biased low-order estimates. AKS-QFI instead separates Krylov truncation error from finite-sample uncertainty and supplies a component-aware stopping interface.

This separation is the actual new element. It directly addresses a practical failure mode where a narrow width around a truncated estimate gets accepted as success. The n=4 noisy mixed-state benchmark reports width-only false-stop rates between 0.16 and 0.68, while AKS-QFI reports zero false successes under the same budget and reaches accurate 5% relative tolerance after recalibration.

The framing is useful. Anyone running shadow-based QFI on near-term hardware has to decide when to stop, and lumping the two error sources together is an easy mistake. Naming the two directions explicitly gives a cleaner reliability layer.

The soft spot is the recalibration step. The abstract states that Krylov resolution and sample counts are adjusted to achieve the reported tolerance, yet supplies no observable-only procedure. If that adjustment needs the true QFI value for validation, the zero false-success result becomes non-reproducible in deployment. The numbers also come without error bars, run counts, or derivation details.

The work is aimed at people already using classical shadows for QFI in metrology or variational settings. Readers who need concrete stopping rules will find the two-component view worth testing, even if the current evidence is preliminary.

The central idea is coherent and the benchmark comparison is presented as external evidence rather than circular. It deserves a serious referee to examine whether the recalibration can be made practical and to see larger-system results.

Referee Report

1 major / 0 minor

Summary. The paper introduces AKS-QFI, a two-component adaptive stopping interface for Krylov-shadow QFI estimation that separates Krylov truncation error from finite-sample statistical uncertainty. It claims that width-only stopping produces false stops with rates 0.16-0.68 on an n=4 noisy mixed-state benchmark, while AKS-QFI yields zero false success declarations under the same resource limit and, after recalibrating Krylov resolution and sample counts, produces accurate success declarations at true 5% relative tolerance.

Significance. If the recalibration step can be performed using only observable quantities, the separation of error sources would supply a practical reliability layer for shadow-based QFI estimators, addressing a recognized weakness in adaptive stopping for quantum metrology. The n=4 benchmark provides initial evidence of reduced false-stop risk relative to width-only rules.

major comments (1)

[Abstract] Abstract: the concrete false-stop rates (0.16-0.68) and the claim of zero false-success declarations plus accurate 5% tolerance after recalibration are stated without any derivation, error-bar information, number of independent runs, or description of the recalibration procedure itself. Because the recalibration step is load-bearing for the reliability claim, its absence prevents evaluation of whether the reported performance can be reproduced without ground-truth QFI knowledge.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and for identifying the need for greater transparency in the abstract regarding our numerical claims. We address the comment point by point below.

read point-by-point responses

Referee: [Abstract] Abstract: the concrete false-stop rates (0.16-0.68) and the claim of zero false-success declarations plus accurate 5% tolerance after recalibration are stated without any derivation, error-bar information, number of independent runs, or description of the recalibration procedure itself. Because the recalibration step is load-bearing for the reliability claim, its absence prevents evaluation of whether the reported performance can be reproduced without ground-truth QFI knowledge.

Authors: The false-stop rates (0.16-0.68) and zero false-success declarations are computed from the n=4 noisy mixed-state benchmark experiments reported in Section 4, which consist of 100 independent runs; one-standard-deviation error bars on these rates appear in Figure 3 and Table II. The recalibration procedure itself is described in Section 5.2: at each step we use only observable quantities (the empirical shadow variance for the statistical component and the difference between successive Krylov-order estimates for the truncation component) to decide whether to increase Krylov order or sample count, without access to ground-truth QFI. We agree that the abstract is too terse on these points and will revise it to (i) state the number of runs, (ii) note that error bars are reported in the main text, and (iii) briefly indicate that recalibration uses only observable quantities, with a pointer to Section 5.2. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; benchmark claims remain external to method definition

full rationale

The paper introduces AKS-QFI as a two-component adaptive stopping rule separating Krylov truncation error from finite-sample uncertainty. The reported false-stop rates (0.16-0.68 for width-only) and zero false-success declarations under AKS-QFI are presented as outcomes of an external n=4 noisy mixed-state benchmark. No equations, fitted parameters, or self-citations are exhibited that reduce the success declarations, 5% tolerance claims, or recalibration step to a definition or input internal to the stopping rule itself. The derivation chain for the reliability layer is therefore self-contained against the stated benchmark evidence.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated in the provided text.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Reliable Adaptive Stopping for Krylov-Shadow Quantum Fisher Information Estimation." pith.science (2026). https://pith.science/paper/MFQGOT6M

@misc{pith2026260514338,
  author       = {Pith},
  title        = {Pith review of: Reliable Adaptive Stopping for Krylov-Shadow Quantum Fisher Information Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFQGOT6M}},
  note         = {Machine review of arXiv:2605.14338}
}

abstract

Scalable quantum Fisher information (QFI) estimation becomes actionable when the numerical estimate is paired with a trustworthy stopping decision. Krylov-shadow QFI estimation has two resource directions: the Krylov order sets the population resolution, whereas the sample count controls statistical uncertainty at that resolution. We show that treating these directions as one can produce false stops, where a width-based rule reports a narrow interval around a biased low-order estimate. We turn adaptive stopping into a two-component reliability problem, separating Krylov truncation from finite-sample uncertainty, and introduce AKS-QFI, a component-aware stopping interface for Krylov-shadow estimators. On a noisy mixed-state benchmark at $n=4$ qubits, width-only stopping has false-stop rates from $0.16$ to $0.68$. Under the same resource limit, AKS-QFI returns no false success declarations; after recalibrating Krylov resolution and sample counts, it returns accurate success declarations at true 5% relative tolerance. These results make adaptive stopping a reliability layer for shadow-based QFI estimation.

Figures

Figures reproduced from arXiv: 2605.14338 by the authors.

**Figure 1.** Narrow empirical intervals do not establish accuracy under width-only stopping. Terminationtime empirical interval width wM versus post hoc absolute error E for all runs (n = 4, five dephasing levels, 50 independent replicates per noise level). Width-only runs form a cluster in the upper-left region (wM ≤ ε, E > ε): these are false stops in the sense of Eq. (18). The empirical gate can pass at the terminal low-K p… view at source ↗

**Figure 2.** Component-aware stopping suppresses false success declarations under the default resource limit. False-stop rate (FSR, filled markers) and stop rate (SR, open markers) versus dephasing probability pϕ (n = 4, 50 independent replicates per point), with 95% Wilson confidence intervals. Width-only FSR increases with pϕ and reaches 0.68 at pϕ = 0.24, while componentaware FSR remains at zero throughout. Componentaware S… view at source ↗

**Figure 3.** Width-only stopping’s low effective cost reflects premature termination at a biased value. Median post hoc absolute error versus pϕ (n = 4, 50 independent replicates per point). The two rules diverge at pϕ ≥ 0.18: component-aware stopping reduces median error by up to a factor of 4.8 relative to width-only stopping despite using up to 16× larger effective terminal sample counts at these noise levels, indicating th… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: shows the decision geometry behind this control. All component-aware terminal points lie below the true-error threshold, while runs without a success label are precisely the cases where the bootstrap width remains above threshold. The contrast with the default grid is …

**Figure 5.** Figure 5: Threshold sensitivity under a fixed-resource ablation. Each cell is one stopping configuration (K stop min , Mstop min , P) aggregated over replicates at fixed noise (pϕ = 0.12, pdep = 0.03), ε = 0.2, resource limits (Kmax, Mmax) = (8, 512). Color and cell text give th…

**Figure 6.** Figure 6: Empirical Krylov convergence and observable stability. For each system size, markers show median inter-order change dK and post hoc truncation bias |BK| versus Krylov order at pϕ = 0.12; dashed curves show fitted exponential trends. For n = 4, both quantities decay wit…

**Figure 8.** Figure 8: Early termination under width-only stopping can appear efficient while remaining inaccurate. Post hoc absolute error versus effective terminal sample count for the benchmark (logarithmic vertical axis). Width-only runs that terminate at low effective sample counts fr…

**Figure 9.** Figure 9: Component-aware stopping uses more terminal samples under the current resource limits. Median effective terminal sample count versus dephasing probability p in the n = 4 benchmark (50 independent replicates per point). In the tested configuration, component-aware st…

**Figure 10.** Figure 10: Stop-precision summary for success declarations. Each cell reports the post hoc stop precision for a rule and dephasing probability in the n = 4 benchmark (50 independent replicates per point), defined as the fraction of success declarations satisfying E ≤ ε. Width-…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Quantum sens- ing

Christian L Degen, Friedemann Reinhard, and Paola Cappellaro. “Quantum sens- ing”. Reviews of modern physics 89, 035002 (2017)

work page 2017
[2]

Advances in quan- tum metrology

Vittorio Giovannetti, Seth Lloyd, and Lorenzo Maccone. “Advances in quan- tum metrology”. Nature photonics5, 222– 229 (2011)

work page 2011
[3]

Quantum metrology and its application in biology

Michael A Taylor and Warwick P Bowen. “Quantum metrology and its application in biology”. Physics Reports615, 1–59 (2016)

work page 2016
[4]

Quantum fisher informa- tion matrix and multiparameter estimation

Jing Liu, Haidong Yuan, Xiao-Ming Lu, and Xiaoguang Wang. “Quantum fisher informa- tion matrix and multiparameter estimation”. JournalofPhysicsA:MathematicalandThe- oretical53, 023001 (2020)

work page 2020
[5]

Introduc- tion to quantum fisher information

Dénes Petz and Catalin Ghinea. “Introduc- tion to quantum fisher information”. In Quantum probability and related topics. Pages 261–281. World Scientific (2011)

work page 2011
[6]

Quantum fisher information for states in exponential form

Zhang Jiang. “Quantum fisher information for states in exponential form”. Physical Re- view A89, 032128 (2014)

work page 2014
[7]

Quantum fisher information from randomized mea- surements

Aniket Rath, Cyril Branciard, Anna Min- guzzi, and Benoît Vermersch. “Quantum fisher information from randomized mea- surements”. Physical Review Letters127, 260501 (2021)

work page 2021
[8]

Fisher information in noisy intermediate-scale quantum applica- tions

Johannes Jakob Meyer. “Fisher information in noisy intermediate-scale quantum applica- tions”. Quantum5, 539 (2021)

work page 2021

Show all 22 references

[9]

Variational quantum al- gorithm for estimating the quantum fisher information

Jacob L Beckey, M Cerezo, Akira Sone, and Patrick J Coles. “Variational quantum al- gorithm for estimating the quantum fisher information”. Physical Review Research4, 013083 (2022)

2022
[10]

Robust estimation of the quan- tum fisher information on a quantum proces- sor

Vittorio Vitale, Aniket Rath, Petar Jurcevic, Andreas Elben, Cyril Branciard, and Benoît Vermersch. “Robust estimation of the quan- tum fisher information on a quantum proces- sor”. PRX Quantum5, 030338 (2024)

2024
[11]

Predicting many proper- ties of a quantum system from very few measurements

Hsin-Yuan Huang, Richard Kueng, and John Preskill. “Predicting many proper- ties of a quantum system from very few measurements”. Nature Physics16, 1050– 1057 (2020). arXiv:2002.08953

2020
[12]

The randomized measurement toolbox

Andreas Elben, Steven T. Flammia, Hsin- Yuan Huang, Richard Kueng, John Preskill, Benoît Vermersch, and Peter Zoller. “The randomized measurement toolbox”. Na- ture Reviews Physics 5, 9–24 (2023). arXiv:2203.11374

2023
[13]

Krylov shadow tomography: Efficient estimation of quantum fisher information

Da-Jian Zhang and D. M. Tong. “Krylov shadow tomography: Efficient estimation of quantum fisher information”. Phys. Rev. Lett.134, 110802 (2025). arXiv:2503.01697

2025
[14]

Shadow tomography of quantum states

Scott Aaronson. “Shadow tomography of quantum states”. SIAM Journal on Comput- ing 49, STOC18–368–STOC18–394 (2020). arXiv:1711.01053

2020 arXiv
[15]

On the classical shadow nonparametric boot- strap

Eric Ghysels and Jack Morgan. “On the classical shadow nonparametric boot- strap” (2025). arXiv:2511.09793

2025
[16]

Sequential tests of statis- tical hypotheses

Abraham Wald. “Sequential tests of statis- tical hypotheses”. Ann. Math. Statist.16, 117–186 (1945)

1945
[17]

Time- uniform, nonparametric, nonasymptotic con- fidence sequences

Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon. “Time- uniform, nonparametric, nonasymptotic con- fidence sequences”. Ann. Statist.49, 1055– 1080 (2021)

2021
[18]

Self-guided quantum to- mography

Christopher Ferrie. “Self-guided quantum to- mography”. Physical Review Letters 113, 190404 (2014)

2014
[19]

Robust on- line Hamiltonian learning

Christopher E. Granade, Christopher Ferrie, Nathan Wiebe, and D. G. Cory. “Robust on- line Hamiltonian learning”. New Journal of Physics 14, 103013 (2012). arXiv:1207.1655

2012 arXiv
[20]

Ban- dit algorithms

Tor Lattimore and Csaba Szepesvári. “Ban- dit algorithms”. Cambridge University Press. (2020)

2020
[21]

PAC bounds for multi-armed bandit and Markov decision processes

Eyal Even-Dar, Shie Mannor, and Yishay Mansour. “PAC bounds for multi-armed bandit and Markov decision processes”. In Proceedings of the 15th Annual Conference on Computational Learning Theory (COLT). Volume 2375 of Lecture Notes in Computer Science, pages 255–270. Springer (2002)

2002
[22]

Asymptotic statistics

A. W. van der Vaart. “Asymptotic statistics”. Number 3 in Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press. Cambridge (1998). 25

1998

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.

[1] [1]

Quantum sens- ing

Christian L Degen, Friedemann Reinhard, and Paola Cappellaro. “Quantum sens- ing”. Reviews of modern physics 89, 035002 (2017)

work page 2017

[2] [2]

Advances in quan- tum metrology

Vittorio Giovannetti, Seth Lloyd, and Lorenzo Maccone. “Advances in quan- tum metrology”. Nature photonics5, 222– 229 (2011)

work page 2011

[3] [3]

Quantum metrology and its application in biology

Michael A Taylor and Warwick P Bowen. “Quantum metrology and its application in biology”. Physics Reports615, 1–59 (2016)

work page 2016

[4] [4]

Quantum fisher informa- tion matrix and multiparameter estimation

Jing Liu, Haidong Yuan, Xiao-Ming Lu, and Xiaoguang Wang. “Quantum fisher informa- tion matrix and multiparameter estimation”. JournalofPhysicsA:MathematicalandThe- oretical53, 023001 (2020)

work page 2020

[5] [5]

Introduc- tion to quantum fisher information

Dénes Petz and Catalin Ghinea. “Introduc- tion to quantum fisher information”. In Quantum probability and related topics. Pages 261–281. World Scientific (2011)

work page 2011

[6] [6]

Quantum fisher information for states in exponential form

Zhang Jiang. “Quantum fisher information for states in exponential form”. Physical Re- view A89, 032128 (2014)

work page 2014

[7] [7]

Quantum fisher information from randomized mea- surements

Aniket Rath, Cyril Branciard, Anna Min- guzzi, and Benoît Vermersch. “Quantum fisher information from randomized mea- surements”. Physical Review Letters127, 260501 (2021)

work page 2021

[8] [8]

Fisher information in noisy intermediate-scale quantum applica- tions

Johannes Jakob Meyer. “Fisher information in noisy intermediate-scale quantum applica- tions”. Quantum5, 539 (2021)

work page 2021