acceptodds
Under review as a conference paper at ICLR 2027

Sampling and Krylov-Truncation Uncertainty in Quantum Fisher Information Estimation

Abstract

Estimating quantum Fisher information (QFI) from randomized measurements with a truncated Krylov representation introduces two errors: sampling error around the fixed-order target and truncation error relative to the full-space QFI. Bootstrap intervals quantify the former but can become narrow around a biased target, making width-based stopping unreliable. We formulate adaptive stopping as selective uncertainty quantification, with separate sampling and truncation radii and abstention when their sum does not support the requested tolerance. For independent confirmation data, we prove finite-horizon false-declaration control provided that the combined radius is a valid total-error bound. We also introduce Krylov-first positivity stabilization to mitigate finite-sample plug-in bias. With frozen within-instance truncation calibration, a six-qubit experiment using \(\lvert+\rangle^\otimes6\) at \(K=48<64\) yields 19 declarations in 20 held-out runs, no observed false declarations, and 4.04% median relative error at a 10% target. Under matched random draws, full-space positivity produces 20 false declarations, whereas Krylov-first stabilization produces 16 correct declarations and four abstentions. The results demonstrate accurate selective QFI decisions at non-full Krylov resolution while identifying transferable truncation control as the central scalability challenge.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
quantum Fisher information, uncertainty quantification, selective inference, adaptive stopping, randomized measurements, Krylov subspace methods

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q7z9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies uncertainty quantification for quantum Fisher information (QFI) estimation when randomized measurements are combined with a truncated Krylov representation. The authors formulated adaptive stopping as selective uncertainty quantification, with separate sampling and truncation radii and abstention when their sum does not support the requested tolerance.

Critical strengths

The paper identifies a real failure mode in adaptive QFI estimation. A sampling confidence interval can become arbitrarily narrow around a biased truncated target.

The paired study of positivity stabilization is well designed. Comparing full-space and Krylov-first positivity using identical random draws helps isolate the effect of the order in which stabilization and Krylov projection are performed.

Critical weaknesses

The conceptual contribution is relatively limited. The central decomposition is essentially the standard separation of statistical estimation error and approximation bias.

The strongest positive experiment relies on exact within-instance truncation calibration.

The empirical approximation surrogate does not solve the truncation problem.

The central scalability challenge is explicitly left as future work.

Questions for authors

What do the authors consider the main technical novelty beyond the basic decomposition of sampling error and approximation error?

Can the authors provide any finite-sample coverage result for the bootstrap radius used for the nonlinear positivity-plus-QFI estimator?

1 comment

We appreciate your review. The question on finite-sample coverage was very useful to us, and it led to the main addition of the revision.

What do the authors consider the main technical novelty beyond the basic decomposition of sampling error and approximation error?

We agree that the decomposition itself is standard. For us its value lies in what it exposes. Proposition 1 and Figure 5 show that the adjacent-order change does not bound the approximation error even without sampling noise, so convergence cannot be read off the Krylov iterates. Once this is clear, the order of positivity and projection becomes important, and Krylov-first stabilization removes the false declarations caused by full-space positivity on the low-noise instance (Figure 4). The revision then adds finite-sample bounds on FF from shadow data alone. The variational and variance bounds are known. What is new is fitting HH on one batch and evaluating it on another, so that the guarantee survives both the data-driven choice of HH and adaptive stopping. These bounds stay close to FF up to eight qubits, where the randomized-measurement bound of Rath et al. loses much of the QFI (Table 2).

Can the authors provide any finite-sample coverage result for the bootstrap radius used for the nonlinear positivity-plus-QFI estimator?

This question pushed us in a productive direction. Rather than analyzing the bootstrap of the nonlinear estimator, we moved the guarantee to a linear quantity, where it can be proved. Once HH is fixed, L(H)=Tr⁡(ρA)\mathcal L(H)=\operatorname{Tr}(\rho A) with A=−2i[H,G]−H2A=-2i[H,G]-H^2. Proposition 2 (Section 3.3) states that if HH depends only on the exploration batch, and L^j\hat L_j is the mean of Tr⁡(AXm)\operatorname{Tr}(AX_m) over the first MjM_j independent confirmation snapshots with empirical Bernstein radius rjr_j at level δ/J\delta/J, then Pr⁡(∃j≤J: F<L^j−rj)≤δ.\Pr(\exists j\le J:\ F<\hat L_j-r_j)\le\delta. The bound therefore holds whenever the run stops. Corollary 2 adds the variance upper bound and gives a two-sided statement, which is reported in 144 of 160 runs from four to eight qubits at a 25% tolerance, with median error 8% and no false report.

The strongest positive experiment relies on exact within-instance truncation calibration. The empirical approximation surrogate does not solve the truncation problem.

We agree with both points for the submitted version, and this is exactly what the revision tries to address. The 10% result of Figure 4 uses the calibrated radius, and dK,Md_{K,M} is used to allocate resources, not to bound the error. The new bounds use neither calibration nor the exact state. The lower bound also catches failed declarations from data (General Response). On scalability, the number of shadows needed for a 5% radius stays at 2182^{18} to 2192^{19} from four to eight qubits (Table 2).

We would be glad to discuss any of this further.

Reviewer_MtNaOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper categorizes two sources of error in Krylov methods for estimating the Quantum Fisher Information (QFI) of a quantum state: The sampling error induced from having a finite number of available samples, and the Krylov truncation error resulting from the projection of a larger quantum state into a smaller Krylov subspace. The paper then formulates an uncertainty quantification protocol that adaptively determines whether the available resources are enough to estimate the QFI of a state to some desired precision.

Critical strengths

The paper’s treatment of sampling error and Krylov induced error is interesting and concrete.

Critical weaknesses

The paper’s writing is difficult to understand in places. Some examples:

  1. The AKS-QFI acronym is never fully explained (what is AKS?)
  2. At line 462 what is the precise meaning of “local surrogate?”
  3. BkB_k is never concretely defined. Is this standard practice in the literature?
Questions for authors
  1. Can the authors clarify the statement at line 045 “These resources need not become adequate at the same time?”
  2. Can the authors expand upon the “observable severity score” defined at line 165? It is not clear whether dd and ww share the same scaling, and whether a simple max⁡\max function would be appropriate. How exactly is I~K,M\tilde I_{K,M} related to resource allocation?
3 comments

Many thanks for the positive assessment and for the list of places where the text was hard to follow. All of them are fixed in the revision.

  1. The AKS-QFI acronym is never fully explained (what is AKS?)

Sorry for this omission. AKS stands for Adaptive Krylov Shadow. It is now spelled out in the abstract and at first use in Section 1.

  1. At line 462 what is the precise meaning of "local surrogate?"

We meant the adjacent-order change dK,M=∣F^K,M−F^K−1,M∣d_{K,M}=|\hat F_{K,M}-\hat F_{K-1,M}|. We called it "local" because it compares two neighboring orders, and a "surrogate" because it stands in for the truncation error, which cannot be observed. We agree the phrase was vague, and Section 6 now reads: "Scalability requires a computable upper bound on ∣BK∣|B_K| rather than the adjacent-order change dK,Md_{K,M}."

  1. B_K is never concretely defined. Is this standard practice in the literature?

Thank you for catching this. BK=FK−FB_K=F_K-F, where FKF_K is the order-KK estimator in the limit of infinitely many measurements. With the Krylov basis QKQ_K, now introduced in Section 2.1, FK=πK F(ρK,GK),πK=Tr⁡(QK†ρQK),ρK=QK†ρQK/πK,GK=QK†GQK.F_K=\pi_K\,F(\rho_K,G_K),\quad \pi_K=\operatorname{Tr}(Q_K^\dagger\rho Q_K),\quad \rho_K=Q_K^\dagger\rho Q_K/\pi_K,\quad G_K=Q_K^\dagger GQ_K. The definition now follows Eq. (3). The split is the usual bias and variance decomposition, with the bias coming from the projection, so we did not cite a specific source for it. We also renamed the weight to πK\pi_K so that it no longer clashes with the width wK,Mw_{K,M}.

Writing out this definition also helped us improve Figure 5. It now plots ∣BK∣|B_K| computed from the exact state, rather than the plug-in error at M=2048M=2048. The message is unchanged and, we think, clearer. At pϕ=0.12p_\phi=0.12 and from K=4K=4 on, ∣BK∣|B_K| stays at 31% of FF for six qubits and 45% for eight, while the median adjacent-order change is at least ten times smaller.

Can the authors clarify the statement at line 045 "These resources need not become adequate at the same time?"

MM controls the sampling error and KK controls the truncation error. A run can have enough samples for a narrow bootstrap interval while KK is still too small (Figure 1), or the other way around. We rewrote the sentence as: "A sample count that makes the sampling error small does not make the truncation error small, and conversely."

... whether d and w share the same scaling, and whether a simple max function would be appropriate. How exactly is \tilde I_{K,M} related to resource allocation?

Both are differences of QFI values, so they have the same units and are compared with the same tolerance ε\varepsilon. The rule increases KK if dK,M>εd_{K,M}>\varepsilon, otherwise doubles MM if wK,M>εw_{K,M}>\varepsilon, and moves to the declaration step when both checks pass. KK is checked first because more samples cannot reduce the truncation error. You are right that the max does not decide which resource to increase. It is at most ε\varepsilon exactly when both checks pass, so describing it as a score that routes allocation was misleading. Section 3.1 now states the rule in terms of dK,Md_{K,M} and wK,Mw_{K,M}, and Eq. (9) of the submission is removed.

The revision also adds data-only bounds on the QFI, summarized in the General Response, which we hope you will find a useful addition.

Thank you for your response.

I have looked at your updated manuscript and I find it has made the paper substantially more difficult to read. The text is dense and it looks like unpolished AI generated output. Section S1.1 is particularly concerning and I have trouble following any of the stated logic in the section.

Moreover, the added text seems to still suffer from the presentation weaknesses I have outlined in my original review. As an example, the acronym SLD is referenced 4 times throughout the text and it is never defined. This kind of error is exactly what I was pointing out in my original review.

Given the disappointing quality of the revisions and having considered the reviews of reviewer QuWG and q7z9, I have unfortunately decided to reduce my stated score for this paper to a Weak Rejection.

I think the paper needs a substantial overhaul in its presentation to communicate its ideas clearly and it also requires a more focused, streamlined approach to the empirical results.

Thank you for reading the revision so carefully and for telling us frankly how it reads. You are right about SLD. It is the symmetric logarithmic derivative LθL_{\theta} of Section 2.1, and the abbreviation is now introduced there in the updated PDF. We are sorry that the revision repeated the kind of slip you had pointed out in your review.

We also take your point about density. In trying to answer every question within the page limit, we packed too much into Section 3.3 and Supplementary S1.1, and the main line of the paper became harder to follow. Your suggestion of a thorough overhaul of the presentation and a more focused set of experiments is fair, and we will follow it in the next version.

We are grateful for your time.

Reviewer_QuWGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the problem of adaptive stopping in Krylov subspace-based QFI estimation. It points out that traditional bootstrap methods fail to control the Krylov truncation error, which can lead to premature stopping. The authors propose the AKS-QFI framework which integrates sampling uncertainty and Krylov stability monitoring with an abstention mechanism and enhances estimation accuracy through "Krylov-first" positivity stabilization.

Critical strengths
  1. The paper distinguishes between sampling uncertainty and Krylov truncation error, which shows only relying on bootstrap intervals to determine early stopping can lead to misleading conclusions.

  2. In a low-noise six-qubit pairing experiment, changing the order in which positivity constraints are applied leads to a substantial improvement in the accuracy of QFI estimates.

Critical weaknesses

Lack of sufficient support for the innovation and effectiveness of the stopping mechanism. AKS-QFI primarily integrates existing techniques, like, namely, bootstrap-based uncertainty estimation, Krylov stability monitoring, and adaptive stopping strategies, without addressing the challenge of deriving a reliable upper bound on truncation error.

The primary positive results in experiments were derived from only six-qubit physical instances without four and five qubits cases, which does not prove this approach can scale to larger quantum systems, nor has it shown that it can reliably generalize across different physical instances.

Questions for authors

With the same estimator and total measurement budget, how much improvement in accuracy or resource efficiency does AKS-QFI bring compared to fixed resources or a simple Early Stopping strategy?

The six-qubit experiment relies on precise and instance-specific calibration of the truncation error. In scenarios involving unknown quantum states, does AKS-QFI obtain a reliable upper bound on the truncation error? If so, how?

3 comments

Thank you for the detailed review. Your questions, especially the one about unknown states, are the reason the revision now has data-only bounds.

AKS-QFI primarily integrates existing techniques ... without addressing the challenge of deriving a reliable upper bound on truncation error.

We agree that bootstrap intervals, Krylov monitoring and adaptive stopping are each standard, and we do not claim them as contributions. What we hope the paper adds is an understanding of why their straightforward combination fails, and what can be done about it. First, the bootstrap interval can be valid and narrow while its target is biased, and the adjacent-order change does not reveal the bias (Proposition 1, Figure 5). As a result, stopping on width or on Krylov stability declares wrong values with high confidence. Second, the order of positivity and Krylov projection matters. On the low-noise six-qubit instance, projecting first turns 20 false declarations into 16 correct ones and four abstentions (Figure 4). Third, and new in the revision, we give bounds on FF that need neither calibration nor the exact state (Section 3.3). The variational bound itself is known. The difficulty is that fitting HH and evaluating it on the same data biases the bound upward, which is why we fit and evaluate on separate batches. This is what keeps the guarantee valid under adaptive stopping.

The closest existing method is the randomized-measurement bound F0F_0 of Rath et al., so we compared with it at the same shadow budget. Even with a radius that favors it, F0F_0 falls from 0.81F0.81F at four qubits to 0.62F0.62F at eight, while our bound stays near 0.9F0.9F (Table 2).

... only six-qubit physical instances without four and five qubits cases ...

Thank you for raising this. We now report results from four to eight qubits (Table 2), and in the supplement for a second probe family (GHZ) and two higher noise levels, all with new seeds and, apart from half of the low-noise runs, new generators.

Table R2 (Supplementary Table S3). Runs that report a lower bound / all runs, and median ℓ/F\ell/F.

Proben=4n=4n=5n=5n=6n=6
random-local, pϕ=0.03p_\phi=0.0340/40, 0.9240/40, 0.9140/40, 0.91
random-local, pϕ=0.12p_\phi=0.1220/20, 0.9020/20, 0.8920/20, 0.88
random-local, pϕ=0.24p_\phi=0.2420/20, 0.9020/20, 0.8820/20, 0.88
GHZ, pϕ=0.03p_\phi=0.0320/20, 0.9620/20, 0.9620/20, 0.96
GHZ, pϕ=0.12p_\phi=0.1220/20, 0.9520/20, 0.960/20, 0.93

No reported bound was false. For GHZ the SLD is an nn-body operator, so we fit HH inside the Krylov space. In the one cell without reports the QFI is small, and the bound reached 0.93F0.93F at our measurement cap, just short of the 5% stopping target. Because these bounds need no calibration, their guarantee does not depend on the instance.

With the same estimator and total measurement budget, how much improvement ... compared to fixed resources or a simple Early Stopping strategy?

This is a very natural question, and we have added the comparison as Table 3 of the revision.

RuleDeclarationsFalseWithin 10%Median / mean measurements
Fixed budgetn/an/a19/2032,768 / 32,768
Fixed budget, chosen in hindsightn/an/a20/20131,072 / 131,072
Width-only stopping20/201/2019/2065,536 / 65,536
AKS-QFI19/200/2020/20163,840 / 204,800

A fixed budget can be just as accurate, but only if one knows the right budget in advance. Here the budget that kept all 20 estimates within tolerance could only be identified with the exact FF, after the runs. With the calibrated truncation radius of Figure 4, AKS-QFI chooses its budget from the data and attaches a held-out radius to each of its 19 declarations, all of which are correct. Width-only early stopping is cheaper but made one false declaration. On the four-qubit control the difference is larger, with 15 false declarations in 19 for width-only stopping and none in 12 for our rule (Table 1).

In scenarios involving unknown quantum states, does AKS-QFI obtain a reliable upper bound on the truncation error? If so, how?

Our approach is to bound FF itself, which removes the need for a separate bound on ∣BK∣|B_K| in the data-only statements. The upper bound comes from F≤4Var⁡ρ(G)F\le4\operatorname{Var}_\rho(G), and together with the lower bound it gives a two-sided statement (Corollary 2). At a 25% tolerance it reports in most runs with no false report (Table 2). The variance bound is already optimal among bounds that are concave in the state (Tóth and Petz, 2013), and any sharper upper bound could be used in Corollary 2 directly.

If any of this is unclear or you would like to see another comparison, please let us know.

Reviewer_QuWGOpenReview ↗

Response to the author rebuttal

Thanks for the reviewer's feedback. It partially addressed my concerns, but I feel the answer is still a bit superficial in general and overall concerns to the paper still exist. I will maintain my score.

Thank you for reading our response and for letting us know where you stand. We understand that a reliable bound on the truncation error for unknown states remains the central question, and that our answer covers it only in part. Your review helped us see this more clearly, and it will guide how we develop the work. Thank you again for your time.

Comments on the paper

2 comments

We sincerely thank all three reviewers for their time and for the careful, constructive reviews. We were glad that the reviewers found the failure mode we study to be real (q7z9), the treatment of the two error sources interesting and concrete (MtNa), and the paired positivity experiment well designed (q7z9) and effective (QuWG). We have uploaded a revised PDF, with all changes in blue.

The main concern, raised by both QuWG and q7z9, is a fair one. Our strongest experiment (Figure 4) calibrates the truncation radius with the exact state, so what can one still say when the state is unknown? We agree that this is the key question for the paper, and most of our effort during the discussion period went into it.

Our answer is a pair of bounds on FF that use only shadow data and stay valid under adaptive stopping (new Section 3.3). The lower bound comes from the variational form of the QFI, with the operator fitted on one batch of shadows and evaluated on a separate one. An upper bound from 4Var⁡ρ(G)4\operatorname{Var}_\rho(G) completes a two-sided accuracy statement. The results are in Section 5.7, and the main table is below.

Table R1 (Table 2 of the revision). Random-local probes at pϕ=0.03p_\phi=0.03 with new seeds. For n≤6n\le6 half of the 40 runs use the original instance and half use new generators. Entries are medians over 40 runs (n≤6n\le6) or 20 runs (n=7,8n=7,8).

nnour ℓ/F\ell/FRath et al. F0F_0 bound$/F$shadows per boundtwo-sided reports at 25%, median error
40.920.812182^{18}36/40, 8.0%
50.910.752182^{18}37/40, 8.1%
60.910.712182^{18}38/40, 7.6%
70.900.662182^{18}17/20, 8.7%
80.900.622182^{18} to 2192^{19}16/20, 7.9%

Across all 400 lower-bound runs, including a GHZ family and higher noise in the supplement, no reported bound was false. The bound also stays well above the randomized-measurement bound of Rath et al. (2021) computed from the same shadows. A useful consequence is that the lower bound catches failed declarations from data alone. It flags all 57 false declarations of our transfer tests and none of the correct ones.

Other changes:

  • Table 3 compares AKS-QFI with fixed budgets and with width-only early stopping, counting every measurement (QuWG).
  • The presentation points of MtNa are addressed, and Figure 5 now plots ∣BK∣|B_K| computed from the exact state.
  • The counterexample, previously Proposition S5, is now Proposition 1 in the main text.
  • The supplementary code reproduces all new tables.

The new material uses the extra page allowed for the revision. We are grateful for the reviews, which have made the paper better.

Area_Chair_skyrOpenReview ↗

rebutal and feedback

Dear Reviewers,

Thank you for your efforts so far. The authors have responded to your reviews, please check them carefully (you may also check other reviewers' comments) and provide your feedback.

Best, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.