acceptodds
Under review as a conference paper at ICLR 2027

Order-Resolved Interaction Attribution in Noisy Quantum Machine Learning

Abstract

Feature-interaction explanations determine whether a model uses features jointly or independently. For a deployed quantum classifier, reliable attribution requires separating interactions learned from data from those induced by the circuit and accounting for hardware noise. We develop an interaction-attribution framework that addresses these effects through the encoding, entangler graph, and noise channel. The encoding fixes the trigonometric degree of the decision function, allowing low-order Sobol variances to be identified exactly from a fixed grid. The entangler graph determines where architecture-induced interactions can occur. For the circuits studied here, the pairwise reach is when the entangler commutes with the readout, and a hardware experiment on ibm_fez confirms this prediction for a distance-four interaction. Noise has a separate effect: its mode-wise contraction does not imply contraction of interaction variances, which can increase when noise breaks cancellations. We derive the attenuation law that holds and find that attenuation increases with interaction order for depolarizing and amplitude-damping noise. Across twelve datasets, the completeness residual is biased downward in all 60 single-encoding runs, while interaction rankings remain stable with mean Spearman correlation above . At higher noise, the two-point correction reduces the residual error from without correction to . These results provide a circuit-aware framework for attributing interaction structure in deployed quantum models. Code: https://anonymous.4open.science/r/ORIAN.

Primary area:
interpretability and explainable AI
Keywords:
Quantum Machine Learning, Explainable Quantum AI, Interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dUMgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper has a useful premise, especially the observation that noise can increase interaction variance by breaking cancellations. The hardware reach experiment is also a strength. However, several claims need tighter support.

Critical strengths

The problem is well motivated, and the separation of identification, architecture, and noise is a clean way to organize the analysis. The hardware experiment is pre-registered and reports interval-level results rather than only point estimates. The identification result is classical and that their claims of quantum advantage are limited. The grid-reconstruction checks and the global-depolarizing control are useful sanity checks.

Critical weaknesses
  1. Theorem 1 counts conditional mean evaluations, but computing these from circuit queries has a cost. Appendix A reconstructs the full coefficient tensor. The claimed polynomial cost needs an explicit procedure and access assumptions.
  2. Standardizing against an initialization ensemble does not establish that the remaining interactions are learned from data. The modest gains over sampled nulls warrant a narrower interpretation.
  3. Eq. 3 assumes a power law that is not guaranteed by Theorem 3. Cancellation can undermine this model. Report failure cases and compare correction methods at matched total shot budgets.
  4. The device experiment validates interaction reach, while attenuation and correction remain simulation results. The 30 to 35% discrepancy from the calibrated prediction further limits conclusions about correction on hardware.
  5. Tab. 3 compares methods targeting different interaction definitions, so agreement with Sobol variances alone does not establish superiority. The product law assumption also limits deployment claims for correlated inputs.
Questions for authors
  1. For the re-uploading circuits used in the main experiments, do you have an a priori criterion that predicts when the sign condition holds? If not, what fraction of the interactions that drive the completeness residual R_k satisfy it?
  2. In Tab. 2, do the fitted exponents remain monotone for m=2 and m=3 across all five seeds and all twelve datasets? If not, does the downward bias in R_2 still hold in those cases?
  3. Is there a version of Theorem 3 that covers amplitude damping, or is the amplitude-damping support purely empirical?
  4. In Tab. 9, under realistic calibration error, for example δ_q = 0.01, does the two-point estimator still outperform zero-noise extrapolation? Figure 3 suggests that it does not at δ_q above roughly 0.02 for the noise levels where it is preferred, so please report the crossover explicitly for the calibration error expected on ibm_fez.
6 comments
AuthorsOpenReview ↗

Reply to Reviewer dUMg (part 1)

We thank the reviewer for the detailed and specific review. We answer each point below.

Request 1.

"Theorem 1 counts conditional mean evaluations, but computing these from circuit queries has a cost. Appendix A reconstructs the full coefficient tensor. The claimed polynomial cost needs an explicit procedure and access assumptions."

Both the procedure and the access assumptions are now stated in Appendix A of the revised paper. We assume that we can choose the inputs and that each execution returns one shot of every ZqZ_q, combined into Y=b+∑qvqzqY = b + \sum_q v_q z_q with ∣Y∣≤B|Y| \le B. At a grid node of SS, the other coordinates are drawn from their product law, so E[Y]\mathbb{E}[Y] is the conditional mean at that node, and no coefficient tensor is reconstructed. By Hoeffding's inequality, N≥2B2log⁡(2K/δ)/ε2N \ge 2B^2 \log(2K/\delta)/\varepsilon^2 shots at each of the KK nodes of Theorem 1 give

∣D^S−DS∣≤2Bε cm∣S∣for all ∣S∣≤k, with probability at least 1−δ,|\hat D_S - D_S| \le 2B\varepsilon\, c_m^{|S|} \quad \text{for all } |S| \le k, \text{ with probability at least } 1-\delta,

where cm=0.78c_m = 0.78, 1.57 and 2.46 for m=1m = 1, 2 and 3. The total cost KNKN is therefore polynomial in nn for fixed kk and mm, and at fixed depth Theorem 1(b) also gives the residual itself exactly, at a cost at most linear in nn.

Request 2.

"Standardizing against an initialization ensemble does not establish that the remaining interactions are learned from data. The modest gains over sampled nulls warrant a narrower interpretation."

Thank you for this comment. We now describe zSz_S only as the excess over the untrained ansatz. Whether the remaining interactions are learned from data is tested separately, by retraining on shuffled labels. The null ratio, the largest interaction of a shuffled-label model divided by that of the trained model, has a mean between 0.04 and 0.70 across datasets, so on every dataset the trained models carry larger interactions (Table 7, formerly Tab. 6). Section 4 of the revised paper states this narrower reading.

Request 3.

"Eq. 3 assumes a power law that is not guaranteed by Theorem 3. Cancellation can undermine this model. Report failure cases and compare correction methods at matched total shot budgets."

Thank you for this suggestion. Eq. 3 is a modelling assumption that Theorem 3 does not imply, and Section 5.1 of the revised paper now states this. We apply it to the order sums DkD_k with an exponent fitted per model, and at that level it fits closely, with a lowest fit R2R^2 of 0.9967 for depolarizing noise and 0.9857 for dephasing. Failures appear only for individual subsets, as the first table below shows.

Subsets (m=1m = 1, order at most three) where DSD_S grows under noise or the power law fits poorly, with their share of variance in parentheses.

ChannelDSD_S grows at p=0.03p = 0.03Fit R2<0.99R^2 < 0.99
Depolarizing0.6% (0.9%)1.5% (1.6%)
Amplitude damping6.5% (3.9%)9.4% (4.6%)
Dephasing4.3% (10.0%)2.4% (2.3%)

The subsets where DSD_S grows are the cancellation case, and they carry little variance except under dephasing. The second table compares the correction methods at matched total shots.

Mean absolute error of R2R_2 at matched total shots per grid node, at device error (calibrated device model, 60 models).

Shots per nodeNo correctionTwo-level exponentialThree-level Richardson
2560.00210.00770.0350
1,0240.00120.00330.0105
4,0960.00080.00160.0039
16,3840.00080.00070.0018
65,5360.00080.00040.0009

Two-level exponential extrapolation is more accurate than three-level Richardson extrapolation at every budget, and at device error, correction pays off only above about 16,384 shots per node. The failure cases are in Appendix J, and the matched-budget comparison is in Appendix K (Table 13) of the revised paper.

AuthorsOpenReview ↗

Reply to Reviewer dUMg (part 2)

Request 4.

"The device experiment validates interaction reach, while attenuation and correction remain simulation results. The 30 to 35% discrepancy from the calibrated prediction further limits conclusions about correction on hardware."

We have now measured both attenuation and correction on ibm_fez. The test had to fit within the 10 minutes of IBM's free plan, as we do not have paid hardware access. Three models (n=4n = 4, m=1m = 1, L=3L = 3, seeds 0, 1 and 2, with no selection) ran on all 81 grid nodes of a four-qubit chain with gate twirling. Each circuit GG was interleaved with its folded copy GG†GG G^\dagger G, which raises the noise, and the hypotheses were fixed before any of these jobs ran.

R2R_2 on ibm_fez.

SeedExactMeasured, GGFolded, GG†GG G^\dagger GExtrapolated
00.17170.15660.13500.1683
10.12630.11140.10060.1173
20.14460.13500.12350.1412

Folding shrank the higher interaction orders more in all three models and lowered R2R_2 by a factor of 0.893 (95% interval [0.860,0.926][0.860, 0.926]), and extrapolating from the two noise levels moved R2R_2 towards its exact value in every model. The test is in Appendix H (Table 6) and Section 6.3 of the revised paper, and Section 4.1 corrects the gap to the calibrated prediction to 21% to 36%.

Request 5.

"Tab. 3 compares methods targeting different interaction definitions, so agreement with Sobol variances alone does not establish superiority. The product law assumption also limits deployment claims for correlated inputs."

Because real datasets have no ground truth, Table 3 measures agreement with DijD_{ij}. To compare the estimators independently of their definitions, we use top-KK recovery of a planted pair.

Planted-pair recovery (mean ±\pm standard deviation over five seeds).

ModelRaw DijD_{ij}zSz_SShapley interactionFaith-Shap
m=1m = 10.767±0.0220.767 \pm 0.0220.800±0.0270.800 \pm 0.0270.800±0.0270.800 \pm 0.0270.800±0.0270.800 \pm 0.027
m=2m = 20.622±0.0420.622 \pm 0.0420.744±0.0570.744 \pm 0.0570.600±0.0650.600 \pm 0.0650.600±0.0650.600 \pm 0.065
m=3m = 30.667±0.0350.667 \pm 0.0350.711±0.0220.711 \pm 0.0220.633±0.0570.633 \pm 0.0570.633±0.0570.633 \pm 0.057

With one encoding, zSz_S matches both interaction indices, and with re-uploading it recovers the planted pair more often. For correlated inputs, the residual under each dataset's empirical joint distribution still decreases under device noise in all 60 runs. These results are in Section 6.2, Table 3 and Appendix I (Table 8) of the revised paper.

Request 6.

"For the re-uploading circuits used in the main experiments, do you have an a priori criterion that predicts when the sign condition holds? If not, what fraction of the interactions that drive the completeness residual RkR_k satisfy it?"

Such a criterion is not needed, because parts (a), (b) and (d) of Theorem 3 and Corollary 4 do not depend on the sign condition. The condition can also be checked directly from the Pauli decomposition of a trained model, and it is rare in practice. On our 60 trained models, the subsets that satisfy it carry on average 0.3% of the variance above order two. Section 5 of the revised paper states this after Theorem 3, and Appendix D reports the fraction.

AuthorsOpenReview ↗

Reply to Reviewer dUMg (part 3)

Request 7.

"In Tab. 2, do the fitted exponents remain monotone for m=2 and m=3 across all five seeds and all twelve datasets? If not, does the downward bias in R2R_2 still hold in those cases?"

We thank the reviewer for this question. The exponents are not monotone in every model, but the downward bias still holds in most of the exceptions.

Models (out of 60) with e1<e2<e3e_1 < e_2 < e_3, and the non-monotone cases in which R2R_2 still falls at p=0.1p = 0.1.

Channelm=2m = 2m=3m = 3R2R_2 falls where not monotone
Depolarizing6060no exceptions
Amplitude damping565112 of 13
Dephasing394331 of 38

Under the calibrated device model, with noise on every gate, R2R_2 decreased in all 120 runs with re-uploading. The main exception is dephasing, which leaves ZZ unchanged, as the Limitations paragraph already notes, and these results are in Appendix J (Table 11) of the revised paper.

Request 8.

"Is there a version of Theorem 3 that covers amplitude damping, or is the amplitude-damping support purely empirical?"

Yes, the amplitude-damping case is now proved as Theorem 3(e). For amplitude damping with parameter γ\gamma (λX=λY=1−γ\lambda_X = \lambda_Y = \sqrt{1-\gamma}, λZ=1−γ\lambda_Z = 1-\gamma, and a shift γ\gamma along ZZ), the noisy Pauli coefficients are exactly m′=(R⊤)⊗nmm' = (R^{\top})^{\otimes n} m, where RR is its Pauli transfer matrix, and

DSN≤(1−γ)∣S∣∑a∈{X,Z}Sva(∑P: P∣S=a∣mP∣∏k∉SμˉPk)2,μˉ=(1,2/π,0,γ),D_S^{\mathcal N} \le (1-\gamma)^{|S|} \sum_{a \in \{X,Z\}^S} v_a \Big( \sum_{P:\,P|_S = a} |m_P| \prod_{k \notin S} \bar\mu_{P_k} \Big)^2, \qquad \bar\mu = (1, 2/\pi, 0, \gamma),

which is the bound of Theorem 3(b) with a ZZ outside SS contributing γ\gamma instead of zero.

On 60 trained models, the exact law agrees with simulation to 5×10−145 \times 10^{-14} and the bound is never violated. The ordering across interaction orders remains empirical, and it holds in 59 of 60 models at m=1m = 1. Theorem 3(e) is in Section 5 of the revised paper, and its proof is in Appendix D.

Request 9.

"In Tab. 9, under realistic calibration error, for example δq=0.01\delta_q = 0.01, does the two-point estimator still outperform zero-noise extrapolation? Figure 3 suggests that it does not at δq\delta_q above roughly 0.02 for the noise levels where it is preferred, so please report the crossover explicitly for the calibration error expected on ibm_fez."

Thank you for asking for this comparison. At δq=0.01\delta_q = 0.01, the two-point estimator outperforms zero-noise extrapolation only from p=0.10p = 0.10 (error 0.0011 against 0.0027), and only when qq is over-estimated.

Lowest pp in {0.01,0.02,0.03,0.05,0.07,0.10,0.15}\{0.01, 0.02, 0.03, 0.05, 0.07, 0.10, 0.15\} at which the two-point estimator has the lower error, for a shared over-estimate q^=q(1+δq)\hat q = q(1+\delta_q) (exact expectations, five seeds).

δq\delta_q00.0020.0050.010.020.05
Lowest pp0.070.050.070.100.15none

Judged by the bias it produces, the calibrated ibm_fez model lies below p=0.02p = 0.02, where the estimator never leads, so on this device we use no correction or two-level exponential extrapolation, neither of which needs qq. The crossover is in Appendix K (Table 14) of the revised paper.

Thank you again for your close reading. We hope these answers support a positive reassessment.

Reviewer_rf9eOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a circuit-aware framework for feature-interaction attribution in noisy quantum machine learning. Its key contributions include exact identification of low-order Sobol variances using finite-frequency structure, characterization of architecture-induced interaction reach through entangler graphs, and an analysis of noise-induced attenuation with a two-point correction for completeness residual bias. Experiments on twelve datasets demonstrate stable interaction rankings and improved completeness residual.

Critical strengths
  1. The paper establishes a rigorous, circuit-aware interaction attribution framework.
  2. The paper provides a thorough theoretical analysis.
Critical weaknesses
  1. The practical scalability of the proposed method is uncertain. Although low-order Sobol variances can be identified in polynomial time, exact evaluation of the completeness residual still requires exponential grid overhead, limiting its scalability.
  2. Limited scope of theoretical guarantees. The theoretical guarantees for noise-induced attenuation are derived under single-encoding assumptions.
Questions for authors
  1. Why are the methods used for comparison relatively outdated?
  2. How does the theoretical attenuation law extend to general non-unital noise channels?
  3. Given that computing the completeness residual requires an exponential evaluation grid, how can this framework be scaled to systems with larger qubit counts?
  4. Given that the hardware validation is limited to a single pre-registered six-qubit run on ibm_fez, how sensitive are the framework's attenuation and bias-correction methods to unmodeled coherent errors that may be present in larger-scale quantum processors?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer rf9e (part 1)

We thank the reviewer for the careful review. We answer each point below.

Request 1.

"The practical scalability of the proposed method is uncertain. Although low-order Sobol variances can be identified in polynomial time, exact evaluation of the completeness residual still requires exponential grid overhead, limiting its scalability."

For circuits of fixed depth, the residual can now be computed exactly, at a cost that grows at most linearly with the number of qubits (Theorem 1(b) of the revised paper). The reason is locality. Each readout term ⟨Zq⟩\langle Z_q \rangle depends only on the features in its light cone, and the ANOVA decomposition is linear, so every variance can be assembled from these local terms. Precisely, if f=b+∑qvq⟨Zq⟩f = b + \sum_q v_q \langle Z_q \rangle and each term depends on at most cc features, then every DSD_S, Var⁡(f)\operatorname{Var}(f) and RkR_k is exact from at most n(2m+1)cn(2m+1)^c circuit settings, or χ(2m+1)c\chi(2m+1)^c when the terms split into χ\chi groups with disjoint light cones. On a ring with IsingZZ entanglers and ZZ readout, c=2L−1c = 2L-1.

Circuit settings for exact RkR_k at L=3L = 3, m=1m = 1.

nnLight-cone settingsFull grid
101,21559,049
121,458531,441
501,2153503^{50}

At n=10n = 10 and 12, this computation matches the full grid to 2×10−142 \times 10^{-14}. For deep circuits, whose light cones cover all qubits, sampling at polynomial cost finds the most important pair but is too noisy for small residuals, so we use it only for ranking. Theorem 1(b) is in Section 3 of the revised paper, and its proof and the sampling procedure are in Appendix A.

Request 2.

"Limited scope of theoretical guarantees. The theoretical guarantees for noise-induced attenuation are derived under single-encoding assumptions."

We appreciate this point. The proof assumes a single encoding because only then does the noisy function factor over qubits. With re-uploading, trainable layers sit between the encoding gates of a feature and this closed form is lost, as the Limitations paragraph and Appendix D state. Beyond this setting, we therefore measured the attenuation exponents on the twelve datasets, with noise on the encoding gates and five seeds each.

Models (out of 60) whose fitted exponents increase with interaction order, e1<e2<e3e_1 < e_2 < e_3.

Channelm=1m = 1m=2m = 2m=3m = 3
Depolarizing606060
Amplitude damping595651
Dephasing133943

Under the calibrated device model, with noise on every gate, R2R_2 decreased in all 120 runs with re-uploading, so the downward bias holds beyond the setting of the proof. These results are in Appendix J (Table 11) of the revised paper.

Request 3.

"Why are the methods used for comparison relatively outdated?"

Our comparison already includes the quantity that recent methods estimate. SHAP-IQ and KernelSHAP-IQ are sampling estimators of the Shapley interaction index, which we compute exactly from all 64 coalitions, and the exact index of the shapiq library agrees with ours to 4×10−154 \times 10^{-15}. We have also added the recent Faith-Shap index to Table 3.

Rows of Table 3 in the revised paper (mean ±\pm standard deviation over 60 models).

EstimatorAgreement with exact DijD_{ij}CircuitsStability at 3×3\times device error
Exact grid117290.998±0.0020.998 \pm 0.002
Shapley interaction index0.72±0.130.72 \pm 0.132,0480.997±0.0040.997 \pm 0.004
Faith-Shap (new)0.72±0.130.72 \pm 0.132,0480.997±0.0040.997 \pm 0.004

Faith-Shap nearly coincides with the Shapley interaction index here, because both weight interaction terms of order two and three in the same way. This comparison is in Table 3 and Section 6.2 of the revised paper, and the shapiq check is in Appendix N.

AuthorsOpenReview ↗

Reply to Reviewer rf9e (part 2)

Request 4.

"How does the theoretical attenuation law extend to general non-unital noise channels?"

The law extends exactly at the level of the Pauli coefficients (Theorem 3(e) of the revised paper). Let a channel with Pauli transfer matrix RR act on every qubit after the encoding. For any channel, unital or not, the noisy function has Pauli coefficients m′=(R⊤)⊗nmm' = (R^{\top})^{\otimes n} m, because the Bloch vector of qubit kk becomes R r(θk)R\,r(\theta_k) with r(θ)=(1,sin⁡θ,0,cos⁡θ)r(\theta) = (1, \sin\theta, 0, \cos\theta), and the expectation of a Pauli string factors over qubits. Hence DSND_S^{\mathcal N} is the noiseless closed form evaluated at m′m'. If the linear part of RR is diagonal and RR shifts ZZ by tZt_Z, as for amplitude damping, generalized amplitude damping and thermal relaxation, we also prove

DSN≤λS2∑a∈{X,Z}Sva(∑P: P∣S=a∣mP∣∏k∉SμˉPk)2,μˉ=(1,2/π,0,∣tZ∣),D_S^{\mathcal N} \le \lambda_S^2 \sum_{a \in \{X,Z\}^S} v_a \Big( \sum_{P:\,P|_S = a} |m_P| \prod_{k \notin S} \bar\mu_{P_k} \Big)^2, \qquad \bar\mu = (1, 2/\pi, 0, |t_Z|),

with λS=max⁡a∏k∈S∣λak∣\lambda_S = \max_a \prod_{k \in S} |\lambda_{a_k}|. This is the bound of Theorem 3(b), except that a ZZ outside SS contributes ∣tZ∣|t_Z| instead of zero.

On 60 trained models, the exact law agrees with simulation to 5×10−145 \times 10^{-14}, and the bound holds in all 22,113 subset evaluations. The ordering across interaction orders is not implied by the theorem for these channels, but we observe it in most amplitude-damping models. Theorem 3(e) is in Section 5 of the revised paper, and its proof is in Appendix D.

Request 5.

"Given that the hardware validation is limited to a single pre-registered six-qubit run on ibm_fez, how sensitive are the framework's attenuation and bias-correction methods to unmodeled coherent errors that may be present in larger-scale quantum processors?"

Thank you for this important question. We added coherent errors to the calibrated device model for the 60 six-qubit models, namely over- and under-rotation of the IsingZZ entanglers and the single-qubit gates, and a fixed crosstalk angle on every edge.

Coherent error at device error (60 models).

Coherent errorChange in R2R_2Runs where R2R_2 fallsSpearman with ideal ranking
None−2.3±0.7-2.3 \pm 0.7%60/600.999±0.0020.999 \pm 0.002
Entangler, −5-5%−15.0±5.1-15.0 \pm 5.1%60/600.992±0.0090.992 \pm 0.009
Entangler, +5+5%+11.2±6.4+11.2 \pm 6.4%3/600.993±0.0070.993 \pm 0.007
Single-qubit, +5+5%−0.6±40.9-0.6 \pm 40.9%35/600.894±0.0640.894 \pm 0.064
Crosstalk angle 0.1+33.2±19.4+33.2 \pm 19.4%3/600.976±0.0220.976 \pm 0.022

Pair rankings stay close to the ideal ranking, but a 5% over-rotation of the entanglers reverses the bias, so our claim about its direction applies to stochastic noise, which gate twirling produces on hardware. The correction removes a coherent error that grows with the noise level but not a fixed one, and with twirling on ibm_fez, folding lowered R2R_2 by a factor of 0.893 (95% interval [0.860,0.926][0.860, 0.926]), as predicted. Section 5 of the revised paper states this scope, and Appendices H and J give the results.

We do not have paid access to quantum hardware, and each device experiment had to fit within the 10 minutes of IBM's free plan, so we make no measured claim about larger processors.

Thank you again for your questions. We hope these answers support a positive reassessment.

AuthorsOpenReview ↗

Discussion Reminder

We would like to kindly follow up on our rebuttal responses. We hope we have addressed your concerns and would appreciate your thoughts before the November 18 (AoE) deadline.

The authors' response has addressed most of my concerns, and I am inclined to increase my rating.

Reviewer_w66wOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This manuscript presents an interaction-attribution framework for quantum machine-learning models by combining functional ANOVA/Sobol variance decomposition with the encoding structure The main contributions include exact grid-based identification of low-order Sobol variances from the encoding degree (Theorem 1), an operator-weight attenuation law under local noise (Theorem 3), and a bias-correction strategy for interaction variances.

Critical strengths

The research motivation of this paper is clear.

Within the formal framework of the manuscript, the theoretical analysis is solid, and the experimental suite is broad.

Critical weaknesses

My main concern is not about the analysis conclusion of the manuscript, but about its underlying assumption - the product input law, which stipulates that the input random variables must be statistically independent. Although this is a general assumption of the ANOVA model, in reality, this assumption is overly restrictive for many practical tasks. In fact, real-world data (especially high-dimensional data) is often highly structured. It can be said that this assumption limits the usability of such models on real data.

Questions for authors

Has the author considered developing a corresponding analytical framework based on more general assumptions (relaxing the requirement of variable independence)? Or, have you considered how, within your current analytical framework, the influence of the interdependencies among variables on the effectiveness of your current model can be evaluated?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer w66w

We thank the reviewer for the positive assessment.

Request 1.

"Has the author considered developing a corresponding analytical framework based on more general assumptions (relaxing the requirement of variable independence)? Or, have you considered how, within your current analytical framework, the influence of the interdependencies among variables on the effectiveness of your current model can be evaluated?"

We have evaluated this within our current framework. The grid recovers the model's Fourier coefficients without using the input distribution, so the same 60 trained models can be re-evaluated under dependent inputs without new circuit evaluations. We used Gaussian copulas of increasing correlation and the empirical joint distribution of each dataset. Under a distribution QQ, the residual is the share of variance that the best order-two approximation cannot explain, R2Q=min⁡gVar⁡Q(f−g)/Var⁡Q(f)R_2^Q=\min_g \operatorname{Var}_Q(f-g)/\operatorname{Var}_Q(f), where gg contains only main and pairwise effects. The importance of a pair is the variance lost when its terms are removed from this approximation.

Residual and pair ranking under dependent laws (60 models, mean ±\pm standard deviation). Under the product law, R2=0.0347±0.0304R_2 = 0.0347 \pm 0.0304.

Law QQR2QR_2^QSpearman with product lawChange under device noise
Gaussian copula, ρ=0.2\rho = 0.20.0286±0.02860.0286 \pm 0.02860.980±0.0150.980 \pm 0.015−2.33±0.64-2.33 \pm 0.64%
Gaussian copula, ρ=0.4\rho = 0.40.0203±0.02360.0203 \pm 0.02360.963±0.0240.963 \pm 0.024−2.34±0.62-2.34 \pm 0.62%
Gaussian copula, ρ=0.6\rho = 0.60.0122±0.01670.0122 \pm 0.01670.950±0.0360.950 \pm 0.036−2.36±0.61-2.36 \pm 0.61%
Gaussian copula, ρ=0.8\rho = 0.80.0058±0.00960.0058 \pm 0.00960.931±0.0470.931 \pm 0.047−2.39±0.67-2.39 \pm 0.67%
Empirical joint law0.0083±0.00990.0083 \pm 0.00990.616±0.2910.616 \pm 0.291−2.65±0.79-2.65 \pm 0.79%

Under every distribution, the residual decreased under device noise in every run, so the downward bias does not depend on independent inputs. Pair rankings stay close to the product-law ranking under Gaussian dependence and differ more under the empirical distribution, as expected, because that distribution also reflects how features occur together in the data rather than only how the model combines them. These results are in Appendix I (Table 8) of the revised paper.

We hope this resolves your concern, and we would be grateful if it could be considered in the reassessment.

AuthorsOpenReview ↗

Discussion Reminder

Thank you again for your feedback. We would appreciate it if you could revisit our rebuttal responses and let us know if you have any further questions or concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for their careful comments. All changes in the revised paper are shown in blue, and the changes are listed below.

  • Theorem 1(b) shows that, at fixed depth, the residual is exact at a cost at most linear in the number of qubits, and Appendix A adds a sampling procedure for deep circuits.
  • Theorem 3(e) gives the exact noisy Pauli coefficients for any single-qubit channel and a bound for amplitude damping (Appendix D).
  • A device test of attenuation and correction on ibm_fez is added in Appendix H and summarized in Section 6.3.
  • Results under dependent input laws are added in Appendix I.
  • Attenuation on the real datasets, failure cases and coherent error are reported in Appendix J.
  • Correction at matched shot budgets and the calibration-error crossover are reported in Appendix K.
  • Faith-Shap is added to Table 3, and planted-pair recovery to Section 6.2.
  • The statements of zSz_S (Section 4) and of Eq. 3 as a model (Section 5.1) are made narrower.

We have also corrected a few statements and citations, all marked in blue. The repository is updated with the code for all new experiments.

Because new tables were added, the appendix tables are renumbered. Table 6 of the submission is now Table 7, Tables 7, 8 and 9 are now Tables 9, 10 and 12, and Tables 10 to 14 are now Tables 15 to 19.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.