acceptodds
Under review as a conference paper at ICLR 2027

Learning Structured Corrections through State Perturbations for Quantum Error Mitigation

Abstract

Noise remains a major obstacle to the reliable execution of algorithms on quantum devices, motivating the development of quantum error mitigation (QEM) methods. Existing learning-based QEM approaches often learn a direct mapping from noisy circuit observations to mitigated measurement expected values and do not explicitly exploit the underlying structure of the mitigation problem. In this work, we introduce a different QEM framework based on pseudo perturbations of the input quantum state. By expanding the perturbation in a Pauli basis, this QEM is casted as a structured prediction problem, for which we develop a neural predictor that infers perturbation coefficients rather than directly predicting the mitigated output. To address the exponential number of Pauli basis, we derive a Pauli propagation based selection procedure. This approach identifies informative perturbation directions and constructs a reduced perturbation subspace. Experiments across diverse quantum tasks and noise settings show that our proposed framework outperform against state-of-the-art baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum Error Mitigation, AI for science, Optimization, Quantum Simulation

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_fkrZOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QEM-STEP, a learning-based quantum error mitigation (QEM) method based on pseudo-quantum state perturbations. The method represents the output correction as a weighted sum of Pauli responses. It selects a limited set of perturbation directions by propagating Pauli operators backward through an ideal circuit and uses neural networks to predict the corresponding correction coefficients. The authors provide bounds on the finite-shot variance for response estimation and analyze errors introduced by propagation truncation. They also propose two classical simulation-based strategies for generating training data: iterative gate addition and near-Clifford states. Numerical experiments on small quantum circuits, swap test scenarios, and many-body dynamics with locally applied noise demonstrate that their method can improve QEM accuracy under various settings.

Critical strengths
  1. The paper introduces a novel QEM method using Pauli responses and demonstrates its advantages on small qubit experiments.
  2. The writing is clear and easy to follow.
  3. It proposes two strategies for generating training data that can be classically simulated, reducing the reliance on ideal labels from the target task.
Critical weaknesses
  1. My primary concerns are the issues of cost and accuracy. For the cost problem, the paper defines backward propagation truncation as:
A_{l-1}=U_l^\dagger\widetilde O_lU_l,\qquad \widetilde O_{l-1}=\operatorname{TopK}(A_{l-1}),$$ However, the operation $\mathcal{A}_{l-1}$ can still generate an exponential number of terms. Thus, expanding first and then applying TopK does not avoid exponential cost. Furthermore, the cost of searching for $K$ should also be taken into account. 2. The algorithm for finding $K$ does not truly find the optimal solution but instead uses a greedy algorithm—each Top-K step is locally optimal, but there is no guarantee of global optimality. Therefore, the method only finds a local solution, and even the per-step optimality can't always be guaranteed. As such, the use of $\arg\max$ in the main text is ambiguous. 3. If I understand correctly, the method aims to measure the coefficients of the Pauli basis. As mentioned in the part about Pauli strings (lines 112–113), the coefficients of the Pauli basis (Pauli strings) decrease exponentially as the number of qubits increases. Therefore, in practice, these coefficients are very hard to measure (especially with quantum noise and finite sampling noise). If the noise level is constant, the signal-to-noise ratio will decrease exponentially. As a result, even if $c_i$ is obtained by multiplying by $2^n$, what is amplified is not the signal but exponentially suppressed signal plus noise—meaning that for large qubit numbers, the noise is almost exclusively being amplified by $2^n$ times. 4. In the experiments of Table 3, the authors describe that the evolution is first performed under ideal many-body dynamics, then local noise is added, and only local observables are measured. Since the noise is applied to single qubits and only 2-qubit observables are measured at the end, although the experiment appears to involve large qubit systems, in practice, only the effect of noise on 2-qubit subsystems (where noise is independent and does not undergo further evolution) is being measured. Thus, these experiments do not demonstrate that the method is effective on large-qubit systems in a general sense. 5. In Table 3's experiments, under ideal conditions, ZNE achieves nearly zero error, while under a finite number of shots, the proposed method performs better. However, the specific number of shots is not stated. Moreover, with enough measurements, the estimation should approach the ideal value. ZNE works by extrapolating values under different noise strengths, so introducing excessive measurement noise inevitably degrades its performance. Thus, analyzing performance under finite shots is indeed necessary, but if ZNE outperforms the proposed method at medium shot counts, it would not demonstrate the superiority of the proposed method. ### Mirror 1. In Lemma 3.1, $\sigma_{c_i}$ is not defined in advance; the same issue occurs in Remark 3.2. 2. In Figure 2, the lines representing GTranQEM, ZNE, and OLS are too close to each other and are hard to distinguish.
Questions for authors

See Weaknesses above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer fkrZ

Dear Reviewer fkrZ,

Thanks for your constructive feedback and thoughtful review. Below, we provide detailed responses to your questions and concerns. We hope that our clarifications adequately address your concerns and encourage you to reconsider your evaluation of our paper.

Q1: My primary concerns are the issues of cost and accuracy [...]:

Answer: We thank the reviewer for raising this important concern regarding support selection cost. Our truncation algorithm follows the Pauli propagation framework of [1]. In general, Pauli propagation can generate exponentially many terms. However, our method considers Pauli observables and gates acting on at most two qubits, which are sufficient for universal quantum computation. Since each gate expands a Pauli string into at most 16 terms and we retain at most KK terms after each propagation step, the intermediate expansion Al−1A_{l-1} contains at most 16K16K terms. Therefore, the computational cost of each propagation step scales linearly with KK, avoiding exponential growth in the number of qubits. To provide empirical evidence, we have added Figure 4 to the revised manuscript, which illustrates the distributions of the selected K∗K^* and Pauli coefficients α\alpha. The results indicate that the selected support sizes remain below the full Pauli-space dimension 4n4^n in the evaluated circuits.

Q2: The algorithm for finding KK does not truly find the optimal solution but instead uses a greedy algorithm [...]

Answer: We thank the reviewer for this comment. We would like to clarify that the proposed algorithm involves two distinct procedures: sequential Top-KK Pauli truncation and optimization of the support size KK. We agree that sequential Top-KK truncation is a greedy procedure and does not guarantee a globally optimal final Pauli support. Nevertheless, at each propagation step, retaining the KK largest Pauli coefficients minimizes the instantaneous truncation error in the coefficient ℓ2\ell_2 norm, assuming exact knowledge of the current Pauli expansion. Our theorem explicitly accounts for the accumulated errors of these sequential truncations and does not assume global support optimality. Importantly, the selection of $K^* $ itself is not greedy. As detailed in Algorithm 1 of the appendix, we evaluate the objective 2E(K)2+K/(2M)2E(K)^2+K/(2M) over candidate values of KK and retain the best score. The two early-stopping conditions preserve global optimality with respect to this objective: when E(K)=0E(K)=0, larger values of KK cannot reduce the truncation error further; when (K+1)/(2M)(K+1)/(2M) exceeds the best score, the nonnegative truncation term ensures that no subsequent candidate can improve the objective. Therefore, the use of arg⁡min⁡\arg\min is justified for selecting $K^* $ with respect to the derived theoretical error-bound criterion. We emphasize that this does not imply global optimality of the resulting Pauli support or minimization of the actual mitigation error. We will clarify this distinction in the main text to avoid ambiguity.

Q3: [...] the method aims to measure the coefficients of the Pauli basis [...] the noise is almost exclusively being amplified by 2n2^n times.:

Answer: We thank the reviewer for raising this important concern regarding measurement scalability. We acknowledge that Pauli coefficients can become exponentially small as the number of qubits increases. However, their exact magnitudes depend on the structure of the observable, quantum evolution, and noise, rather than solely on the number of qubits. Although ci=2nαic_i=2^n\alpha_i, this normalization factor scales both the signal and its statistical uncertainty equally and therefore does not, by itself, degrade the signal-to-noise ratio. Nevertheless, we acknowledge that estimating small Pauli coefficients can become challenging under finite-shot measurements and quantum noise. Our framework addresses this challenge in two ways. First, Pauli-propagation-based support selection restricts measurements to a reduced set of informative Pauli directions, avoiding the need to estimate all 4n4^n coefficients. Second, Theorem 3.3 explicitly characterizes the trade-off between truncation error and finite-shot uncertainty, providing a principled criterion for selecting the support size KK. To provide empirical evidence, we have added Figure 4 to the revised manuscript, which illustrates the distributions of the selected $K^* $ and Pauli coefficients α\alpha. The results indicate that the selected support sizes remain below the full Pauli-space dimension 4n4^n and our method select the coefficients with large magnitude. This reduces the number of required coefficient estimates and can improve their robustness to finite-shot uncertainty.

[1] Shao, Yuguo, Song Cheng, and Zhengwei Liu. "Pauli propagation: Simulating quantum spin dynamics via operator complexity." arXiv e-prints (2025): arXiv-2510.

AuthorsOpenReview ↗

Response to Reviewer fkrZ (part 2)

Q4: In the experiments of Table 3, [...] these experiments do not demonstrate that the method is effective on large-qubit systems in a general sense.

Answer: We thank the reviewer for this insightful comment. We agree this experiment should not be interpreted as demonstrating the scalability of QEM-STEP on large quantum systems. We would like to clarify that the primary motivation of this experiment is to evaluate the applicability of QEM-STEP across different types of quantum evolution, rather than to establish scalability solely in terms of qubit count. Specifically, Table 1 consider gate-based quantum circuits, where the evolution is represented by a sequence of quantum gates with associated noise channels. In contrast, the many-body benchmark considers the real-time evolution of a 50-qubit transverse-field Ising system governed by a Hamiltonian, without requiring an explicit gate-based circuit representation for the mitigation model. As discussed in the Introduction, one limitation of existing ML-based QEM methods, such as GTranQEM, is their reliance on circuit encoding to represent quantum evolution. This dependence limits their direct applicability to quantum dynamics that are not provided in an explicit gate-based circuit representation. In contrast, QEM-STEP is formulated through observable responses under noisy quantum evolution and does not require an explicit circuit encoding for its neural predictor. Therefore, an important motivation of our work is to develop a more flexible QEM framework applicable to different representations of quantum evolution. We will revise the experimental description to clarify this motivation.

Q5: [...] if ZNE outperforms the proposed method at medium shot counts, it would not demonstrate the superiority of the proposed method:

Answer: We rerun the ZNE for experiment in Table 3 with 2000 shots for each circuit execution. This increase improves ZNE performance, reducing its MAE from 0.0341 to 0.0239 under amplitude damping and from 0.0311 to 0.0219 under dephasing. Nevertheless, QEM-STEP achieves substantially lower MAEs of 0.0140 and 0.0148, respectively, while maintaining considerably higher CIR values of 0.8381 and 0.8551. Notably, QEM-STEP requires only 12,367 shots per test instance on average, approximately half the increased ZNE budget. This experiment demonstrate the superiority of QEM-STEP over ZNE even with smaller budget.

MethodAmplitude Damping MAE ↓CIR ↑Dephasing MAE ↓CIR ↑Average shots per test instance
ZNE (original)0.03410.42220.03110.417312,000
ZNE (increase shots)0.02390.47200.02190.447224,000
QEM-STEP0.01400.83810.01480.855112,367

Minor: In the revised manuscript, we also update Figure 2 for better visibility and add the definition of σci\sigma_{c_i} and σα,i\sigma_{\alpha,i} in Lemma 3.1 and Remark 3.2

  1. This is very strange. Even if you consider two-qubit gates, doesn't the Pauli expansion ultimately still have to be expanded over the entire n-qubit circuit? Moreover, in your Figure 4, with only 6 qubits, the number of selected Pauli terms is already between 500 and 1000. Isn't that still exponential?

  2. In Figure 4, with only 6 qubits, the coefficient magnitudes are already on the order of 10−410^{-4} to 10−210^{-2}. Wouldn't they become even smaller with more qubits? Moreover, a single example alone cannot demonstrate that this holds in other scenarios as well.

In summary, I think this method looks good, but it can hardly be used in practice. Therefore, I choose to lower my score.

Dear Reviewer fkrZ,

Thank you for your thoughtful and constructive feedback. We sincerely appreciate the time and effort you have devoted to evaluating our work.

First, we would like to clarify that restricting a circuit to one- and two-qubit gates does not eliminate the possibility of Pauli expansion over an nn-qubit system. Rather, our previous response addressed your Q1 aimed to explain how sequential Top-KK truncation controls the size of the intermediate representation. Specifically, at each propagation step, we retain at most KK Pauli strings, and applying a two-qubit gate generates at most 16K16K candidate terms before merging and truncation. Consequently, the computational cost of each propagation step depends on the retained support size KK, rather than directly on the full Pauli-space dimension 4n4^n. Furthermore, we would also like to clarify that observing 500–1000 selected Pauli terms for a fixed six-qubit system does not, by itself, establish exponential scaling. For six qubits, the full Pauli space contains 46=40964^6=4096 terms, so the selected support remains smaller than the full space.

Second, we agree that individual Pauli coefficient magnitudes may decrease as the number of qubits increases. However, their behavior depends on the circuit structure and the distribution of Pauli coefficients. Importantly, our selection procedure does not rely on the assumption that individual coefficients remain large. Instead, it retains the largest-magnitude coefficients at each propagation step, while Theorem 3.3 bounds the accumulated truncation error in terms of the discarded squared Pauli mass, together with the finite-shot estimation error. Thus, the approximation quality is characterized by the collective contribution of the discarded coefficients rather than their individual magnitudes alone.

We thank the reviewer again for the valuable comments, which have helped us clarify our approach.

Best regards, The Authors

Reviewer_3mPuOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes QEM-STEP, a learning-based quantum error mitigation method. The mitigated value is written as the noisy expectation value plus the response of the noisy channel to a "pseudo" input perturbation, which is a weighted sum of Pauli operators. This makes the mitigated value equal to the noisy expectation plus a learned weighted sum of measured quantities, where each quantity is the expectation of the observable after the noisy channel acts on one Pauli operator (Eq. 5). Each of these quantities is a "Pauli response" and can be estimated on hardware as the difference between the results on the two normalized projectors onto the plus and minus eigenspaces of that Pauli operator (Eqs. 6-7, Lemma 3.1). Up to a dimension-dependent scale factor, the vector of Pauli responses is the vector of Pauli coefficients of the noisy observable in the Heisenberg picture (Remark 3.2). Since measuring every Pauli direction is infeasible (their number grows exponentially with the qubit count), the support is chosen by Top-K Pauli propagation of the ideal circuit, and the truncation level is set by minimizing a bound (Theorem 3.3) that combines the accumulated truncation error with a shot-noise term that grows with the support size and shrinks with the number of shots. A small permutation-style network predicts per-instance weights from a "PR context" and an "SC context" built by mean pooling, and is trained end to end with a squared-norm penalty on the weights (Eq. 11). For the setting without ideal labels, the paper proposes two data generation schemes: iteratively appending gates, with labels computable by Pauli propagation, and near-Clifford state pairs for a swap test. Experiments cover 6-qubit Trotterized circuits and random circuits under several noise models (Table 1, Fig. 2), out-of-distribution depth, noise amplification and coherent noise (Table 4), label-free training (Table 2, Fig. 3), and a 50-qubit task with damping noise applied after the evolution (Table 3). QEM-STEP gets the lowest MAE in most settings.

Critical strengths
  1. The input representation differs from earlier learning-based QEM work. Liao et al. (2024) and Bao et al. (2025) encode the circuit graph, while this paper feeds the model Pauli responses of the noisy channel measured on hardware. These quantities have a clear physical meaning: they are one row of the noisy Pauli transfer matrix restricted to the observable. So part of the noise is measured directly rather than inferred from circuit structure. I think this is a good direction, and it fits with the largest gain showing up under coherent noise (Table 4: MAE 0.0630 vs. 0.1143 for the best baseline), which is where structure-only encodings tend to do poorly.
  2. The analysis of the trade-off between truncation and shot count is simple, but correct as far as I checked (Lemma 3.1, Remark 3.2, Lemma D.1, Theorem 3.3). The layer-wise triangle inequality uses the fact that the Pauli 2-norm is unchanged under unitary conjugation, and the cross term in Eq. 19 vanishes both because the shot-noise error has zero mean and because the two error vectors have disjoint supports. The rule in Algorithm 1, which picks the support size that best balances the accumulated truncation error against the shot-noise term, gives a transparent way to choose the support size, and it extends Shao et al. (2025) to sequential truncation with finite-shot error.
Critical weaknesses
  1. The "structure" may just be a reparameterization, and the ablation that would test this is missing. The network outputs the weights from the noisy value, the vector of Pauli responses and its norm, and the resulting correction (the weighted sum of responses) is a single scalar, so the mitigated value is simply a learned function of the noisy value and the Pauli responses. As Eq. 12 shows, there is always some choice of weights that matches any target correction exactly, and the paper explicitly does not require the perturbation to be a valid quantum state, so the pseudo-perturbation imposes no physical constraint. Table 5 removes the context modules and Top-K, but never compares against the same network predicting the scalar correction directly from the same features. Without that comparison, the gains cannot be credited to the structured correction rather than to the extra measured information in the Pauli responses. Please add a direct-output-head ablation (same inputs and capacity, scalar output), and consider also a model that is linear in the Pauli responses with a single set of weights shared across instances.
  2. The comparison across methods is not matched in information or shots. QEM-STEP gets one extra hardware-measured feature per selected Pauli direction for each instance, and each feature costs twice the per-setting shot count, spent on the two mixed eigenspace states. OLS, MLP, GNN and GTranQEM only see the circuit encoding and the noisy value, and ZNE and CDR have their own different sampling overheads. The paper never reports the number of shots per setting, the support size chosen for each benchmark, or the total number of circuit executions per test instance. Please report the shots per setting, the chosen support size and the total shot count for every method, give the learning-based baselines the same Pauli-response features (for example concatenated to the inputs of MLP and GTranQEM), and include at least one experiment at an equal total shot budget.
  3. The method may only be useful where classical simulation is already feasible. Support selection requires Top-K Pauli propagation of the ideal circuit (Algorithm 1), and iteratively appending gates requires exact Pauli propagation. At 6 qubits there are only 4096 Pauli terms, so the propagation is effectively exact. If the chosen support already covers most of the Pauli weight, then for product input states the expectation of the truncated propagated observable is itself a classical estimate of the ideal value, and this Pauli-propagation-only estimate is never reported as a baseline. The 50-qubit experiment does not test whether support selection scales, because the support size of 15 follows directly from locality (App. C.4). The introduction also says circuit-encoding methods have limited applicability (p. 2), yet QEM-STEP itself needs the ideal circuit. Please report the classical Pauli propagation estimate at the same support size, and add a circuit experiment at a qubit count where exact propagation is not trivial (for example 12 to 20 qubits).
Questions for authors
  1. For each benchmark in Tables 1-4, what are the number of shots per setting and the selected support size? Compared with ZNE, CDR and the learning-based baselines, how many shots and how many distinct circuit executions does QEM-STEP use per test instance?
  2. How do you prepare the normalized plus and minus eigenspace states of each Pauli operator on hardware? Do you sample random product eigenstates for each shot? Is this sampling accounted for in the variance bound of Lemma 3.1?
  3. What happens if the same architecture predicts the scalar correction (the ideal value minus the noisy value) directly from the noisy value, the Pauli responses and their norm? And what happens if MLP or GTranQEM is given the Pauli responses as an additional input?
  4. What MAE does the purely classical estimate from your Top-K Pauli propagation (the expectation of the truncated propagated observable on the input state, at the selected support size) reach on the Table 1 test set?
  5. In Table 5, what exactly does "Top-K off" mean: all Pauli directions, a random subset of the same size, or the lowest-weight Pauli operators? Is the support size kept the same?
  6. Why would label-free training (Table 2, 0.0363) beat supervised training with in-distribution labels (Table 1 mixed, 0.0438)? Are the test circuits and noise parameters the same in both?
  7. In the many-body experiment, the Pauli responses are the same for all input states at a given noise strength. Is the model then just learning a one-dimensional map from the noisy value to the ideal value for each combination of observable and noise strength? How does it compare with a linear rescaling fitted separately for each noise strength, similar to a known-noise inverse?
  8. What constraints does iteratively appending gates place on circuit structure? What is the gate pool, which polynomial bound do you use for the per-layer support size, and how similar are the generated circuits to the Trotterized test circuits? This matters for the claim that the method avoids the structural constraints of inverse or fiducial circuits.
5 comments
AuthorsOpenReview ↗

Response to reviewer 3mPu (part 1)

Dear Reviewer 3mPu,

Thanks for your constructive feedback and thoughtful review. Below, we provide detailed responses to your questions and comments regarding our paper. We hope that our responses adequately address all your concerns.

Q1: [...] what are the number of shots per setting and the selected support size [...]

Answer: We thank the reviewer for raising the important question of measurement overhead and computational resource usage. To quantify this overhead, we report the average shot formulas in the table below and summarize the average number of measurement shots per test instance across the benchmarks in Tables 1–4. In this paper, we assume that each circuit execution requires 1000 shots. Importantly, QEM-STEP is not designed to minimize measurement shots. The support size $K^* $ is selected according to Theorem 3.3 to balance Pauli truncation error against finite-shot estimation error, thereby controlling the trade-off between approximation accuracy and measurement cost. Regarding the information used by competing methods, MLP and OLS rely on gate counts, rotation-angle histograms, and calibration summaries such as mean gate and readout errors. GNN uses a circuit graph with qubit- and gate-specific calibration features, including readout errors, gate errors, and available coherence times and gate durations. GTranQEM additionally uses measured local X, Y, and Z expectation values. Unlike these approaches, QEM-STEP constructs observable-specific Pauli-response features through additional quantum measurements, without relying on device-calibration metadata or auxiliary observable features. This makes QEM-STEP particularly suitable for settings where such information is unavailable, albeit at the cost of additional measurement shots. However, additional measurements alone do not necessarily translate into improved mitigation accuracy. This is further illustrated by the MLP and GTranQEM variants augmented with the Pauli-response vector c\boldsymbol c on Max-6-Random with Mixed noise, which do not achieve better results despite their increased measurement budgets. These results suggest that the effectiveness of QEM-STEP stems not merely from additional measurements, but also from its structured correction formulation.

MethodTotal shots formula
Raw Noisy1,0001{,}000
ZNENscales×1,000N_{\mathrm{scales}}\times1{,}000
CDR(Ncalibration+1)×1,000(N_{\mathrm{calibration}}+1)\times1{,}000
GTranQEM(2n+2)×1,000(2n+2)\times1{,}000
QEM-STEP1,000Ntest∑i=1Ntest(2Ki∗+1)\displaystyle\frac{1{,}000}{N_{\mathrm{test}}}\sum_{i=1}^{N_{\mathrm{test}}}(2K_i^*+1)
MethodTrotter IncoherentTrotter MixedTrotter ProviderRandom IncoherentRandom MixedRandom Provider
Raw Noisy1,0001,0001,0001,0001,0001,000
ZNE3,0003,0003,0003,0003,0003,000
CDR71,00071,00071,00071,00071,00071,000
GTranQEM14,00014,00014,00014,00014,00014,000
QEM-STEP1,348,5061,348,5061,348,506489,700489,700489,700
MetricStrategyRaw NoisyZNECDRGTranQEMQEM-STEP
ShotsInverse-Circuit1,0003,00071,00014,000—
ShotsIterative Gate Addition1,0003,00071,00014,0001,289,253
MethodFinite-shot Amplitude DampingFinite-shot Dephasing
Raw Noisy1,0001,000
ZNE12,00012,000
DAEM12,00012,000
QEM-STEP12,36712,367
MethodCircuit-Scale ExtrapolationNoise-Scale AmplificationCoherent Noise
Raw Noisy1,0001,0001,000
ZNE3,0003,0003,000
CDR71,00071,00071,000
MLP1,0001,0001,000
GTranQEM14,00010,00010,000
QEM-STEP1,574,040125,850119,626
MethodMAE ↓CIR ↑Shots
QEM-STEP0.03730.6833489,700
MLP0.04180.63331,000
MLP + c\boldsymbol c0.10090.3933489,700
GTranQEM0.03950.663314,000
GTranQEM + c\boldsymbol c0.09360.4400502,700
AuthorsOpenReview ↗

Response to reviewer 3mPu (part 2)

Q2: How do you prepare the normalized plus and minus eigenspace states of each Pauli operator on hardware? [...]

Answer: In this work, we focus on mitigating errors arising from noisy quantum evolution and assume ideal state preparation. The normalized positive- and negative-eigenspace states of each Pauli operator are treated as valid input density matrices. Our formulation operates directly on these mixed states and does not explicitly model their preparation through quantum circuits or randomized sampling of product eigenstates for individual measurement shots. Under this assumption, Lemma 3.1 establishes an upper bound on the finite-shot variance of the Pauli response estimator using MM independent measurement shots for each of ρi,+\rho_{i,+} and ρi,−\rho_{i,-}. Thank you for highlighting this, we will incorporate these assumption into the revised manuscript.

Q3: What happens if the same architecture predicts the scalar correction (the ideal value minus the noisy value) directly from the noisy value, the Pauli responses and their norm? And what happens if MLP or GTranQEM is given the Pauli responses as an additional input?

Answer: We have conducted additional experiments on Max-6-Random with Mixed noise comparing QEM-STEP against simpler regression models, including a plain MLP, GTranQEM, and ridge regression, using the same input features (ynoisy,c)(y_{\mathrm{noisy}}, \boldsymbol c). We also introduce a scalar-output variant of QEM-STEP. QEM-STEP achieves the lowest MAE, outperforming all alternative architectures. The scalar-output variant yields a higher MAE of 0.0432, suggesting that explicitly learning Pauli-specific correction coefficients is more effective than directly predicting a scalar correction. Furthermore, augmenting MLP and GTranQEM with the response vector c\boldsymbol c degrades their performance despite having equal shot budget as QEM-STEP. Ridge regression using the same augmented inputs also achieves a higher MAE of 0.0709. These findings provide empirical support for the proposed structured prediction formulation, indicating that its advantage cannot be attributed solely to the availability of the additional response features c\boldsymbol c. Rather, the results suggest that learning Pauli-specific perturbation coefficients and combining them through the structured correction rule contributes to improved mitigation accuracy.

MethodMAE ↓CIR ↑Shots
QEM-STEP0.03730.6833489,700
QEM-STEP (scalar output)0.04320.6567489,700
MLP0.04180.63331,000
MLP + c\boldsymbol c0.10090.3933489,700
GTranQEM0.03950.663314,000
GTranQEM + c\boldsymbol c0.09360.4400502,700
Ridge regression + c\boldsymbol c0.07090.5433489,700

Q4: What MAE does the purely classical estimate from your Top-K Pauli propagation [...]

Answer: Thank you for suggesting this classical baseline. We have test exact classical Pauli propagation achieves an MAE of 0.0052 on the Max-6-Random benchmark under mixed noise, outperforming the baselines reported in Table 1. This strong performance is consistent with the selection of K∗K^*, which minimizes the approximation and finite-shot error bound in Theorem 3.3. However, extending gate-based Pauli propagation to general many-body Hamiltonian dynamics, introduces additional challenges. Unlike gate-defined circuits, Hamiltonian evolution is specified by the continuous-time operator. Applying conventional gate-by-gate Pauli propagation therefore requires constructing an explicit circuit representation, typically through a Trotter–Suzuki product-formula decomposition. Such a decomposition introduces an additional approximation error and computational overhead, with the required number of evolution steps depending on the Hamiltonian structure, simulation time, and target accuracy. In contrast, QEM-STEP does not require explicitly decomposing the Hamiltonian evolution into quantum gates for its mitigation procedure. Instead, it learns structured corrections directly from measured Pauli responses, allowing it to operate on the noisy evolution without propagating the observable through a gate-level representation. This distinction highlights an important advantage of QEM-STEP in settings where an explicit circuit decomposition is unavailable.

Q5: What exactly does "Top-K off" mean [...]

Answer: The "Top-KK off" in Table 5 means that we disable the Pauli propagation support selection module which means the model will use all Pauli directions. This ablation demonstrates the necessity of the support selection. We will update the manuscript to clarify this experiment.

AuthorsOpenReview ↗

Response to reviewer 3mPu (part 3)

Q7: In the many-body experiment, the Pauli responses are the same for all input states at a given noise strength [...]

Answer: Thank you for highlighting this distinction. In the many-body experiment, the Pauli-response vector is determined by the observable and noise channel and is shared across input states at a fixed noise strength. However, the implemented predictor is not conditioned on only one noisy expectation: it receives the 12-point noisy-expectation curve, the Ising parameter J, the observable identity, and the measured pair's position as similar to [2]. We will add these feature descriptions in the revised manuscript. We also run this experiment with a linear regression fitted separately for each noise strength with similar input features as [2]. In the table below, QEM-STEP achieves lower MAE in three of the four settings and consistently higher CIR, demonstrating its advantage over the linear baseline.

MethodExact Amplitude Damping MAE ↓CIR ↑Exact Dephasing MAE ↓CIR ↑Finite-shot Amplitude Damping MAE ↓CIR ↑Finite-shot Dephasing MAE ↓CIR ↑
QEM-STEP0.00470.89560.00590.69490.01400.83810.01480.8551
Linear regression0.00870.55850.00870.44440.01550.72200.01260.7484

Q8: What constraints does iteratively appending gates place on circuit structure? [...]

Answer: When constructing the training circuits in the no-ideal experiments, we use the same set of gates as those in the training data. After each elementary gate addition, we propagate the target observable backward through the candidate circuit and require its Pauli support to satisfy Kl≤(n+∣U∣)2K_l \leq (n +|\boldsymbol{\mathcal{U}}|)^2 at each layer ll. Construction terminates before accepting a gate that violates this bound. The Iterative Gate Addition generates training circuit similar to the Trotter experiment in Table 1. Our generator checks the propagated Pauli support after each gate addition against the above polynomial cutoff. In the six-qubit benchmark, this cutoff is permissive: it allows 729 terms after the first Trotter block, while the largest observed support within that block is 486. Moreover, after 58 gates the cutoff exceeds the entire six-qubit Pauli space. Across the 400 generated training circuits, the largest intermediate support is 924 so no candidate is rejected. These results demonstrate that, although the Pauli support can grow to 4n4^n terms in the worst case, the generated circuits exhibit smaller support in practice. This supports the practical applicability of our training-data generation method.

[1] Bao, Tianyi, et al. "Beyond circuit connections: A non-message passing graph transformer approach for quantum error mitigation." International Conference on Learning Representations. Vol. 2025. 2025. [2] Liao, Manwen, et al. "Noise-agnostic quantum error mitigation with data augmented neural models." npj Quantum Information 11.1 (2025): 8.

Reviewer_ffxbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses quantum error mitigation. On a noisy device, measuring an observable OO after a circuit gives a biased value ynoisy=Tr(O E[ρ])y^{\rm noisy}=\mathrm{Tr}(O\,\mathcal{E}[\rho]) rather than the noise-free value yideal=Tr(O U[ρ])y^{\rm ideal}=\mathrm{Tr}(O\,\mathcal{U}[\rho]), and the goal is to estimate yidealy^{\rm ideal} from noisy runs and classical post-processing. The authors propose a learning-based method, QEM-STEP.

The method is based on a perturbation of the input state: the authors consider adding δ=∑iqiPi\delta=\sum_i q_iP_i to ρ\rho such that the noisy circuit applied to ρ+δ\rho+\delta returns the ideal value. By linearity, the output on ρ+δ\rho+\delta equals ynoisy+∑iqiciy^{\rm noisy}+\sum_i q_ic_i with ci=Tr(O E[Pi])c_i=\mathrm{Tr}(O\,\mathcal{E}[P_i]) (Eq. 5). The corrected value is therefore the noisy value plus a weighted sum of the responses cic_i, and the task is to learn the weights qiq_i.

The quantum device measures these quantities. It runs the noisy circuit on the input state to obtain ynoisyy^{\rm noisy}, and, for each Pauli PiP_i in a selected set SK\mathcal{S}_K, runs the same circuit on the two eigen-projector states of PiP_i; their difference gives cic_i (Eq. 7). The set SK\mathcal{S}_K is selected classically by Top-KK Pauli propagation of the ideal circuit. What remains is a standard regression problem. A neural network gϕ(ynoisy,c,p)g_\phi(y^{\rm noisy},\mathbf{c},\mathbf{p}) receives the noisy value, the vector c\mathbf{c}, and an encoding of each Pauli, and outputs one weight q^i\hat{q}_i per Pauli; the corrected value is y^=ynoisy+∑iq^ici\hat{y}=y^{\rm noisy}+\sum_i\hat{q}_ic_i. It is trained on circuits with known ideal values by minimizing (yideal−y^)2(y^{\rm ideal}-\hat{y})^2 plus a ridge penalty. Two ways to generate training data without ideal values are also proposed.

On the theoretical side, Lemma 3.1 bounds the shot noise of each cic_i, and Lemma D.1 bounds the Top-KK truncation error E(K)E(K). Theorem 3.3 combines them: the error in estimating the cic_i from MM shots is at most a KK-independent term plus 2E(K)2+K/2M2E(K)^2+K/2M. Retaining more Paulis reduces truncation error but increases shot noise, and the rule K⋆=arg⁡min⁡K{2E(K)2+K/2M}K^\star=\arg\min_K\{2E(K)^2+K/2M\} (Alg. 1) selects KK to balance the two.

All experiments are simulations: 6-qubit circuits against ZNE, CDR, OLS, MLP, GNN and GTranQEM, a swap test against DAEM, and a 50-qubit many-body system. QEM-STEP achieves the lowest MAE in all circuit settings and in the finite-shot many-body setting, but is worse than ZNE and DAEM with exact many-body values.

Critical strengths
  1. The main strength of the paper is how it reduces error mitigation to a classical regression problem. The correction is first written as a perturbation of the input state, and because the noisy channel is linear, the perturbed output is simply the noisy value plus a weighted sum of responses ci=Tr(O E[Pi])c_i=\mathrm{Tr}(O\,\mathcal{E}[P_i]) (Eq. 5), with no approximation involved. The responses are measured on the device by running the same circuit on the two normalized eigen-projectors of PiP_i and taking the difference (Eq. 7); these inputs are mixtures of product states and easy to prepare, and Lemma 3.1 bounds the shot noise. The result is a clean division of labor. The quantum device runs the original circuit on extra input states, and nothing else: no noise model and no quasi-probability sampling as in PEC, no noise scaling as in ZNE, no near-Clifford copies of the circuit as in CDR, and no circuit encoding as in the GNN and transformer baselines. What remains is a standard supervised learning problem on (ynoisy,c)(y^{\rm noisy},\mathbf{c}), solved with a small model (a few thousand parameters, Table 6); Eq. 12 even shows what the coefficients should look like. Once ynoisyy^{\rm noisy} and c\mathbf{c} are measured for a circuit, nothing else is asked of the device: training, model selection, and the correction all run offline, so the regressor can be tuned or replaced without new quantum runs. Finally, the inputs c\mathbf{c} carry information no baseline has: they are a measured piece of the actual noisy process, a row of its Pauli transfer matrix, obtained from the same kind of runs as the mitigation.

  2. The method applies to any noise model and any quantum evolution. Because c\mathbf{c} is measured rather than modeled, nothing about the noise is assumed, and because no circuit encoding is needed, the method works the same way for gate-based circuits and for general quantum evolutions (the 50-qubit Hamiltonian dynamics of Sec. 4.4). The experiments bear this out: the method is tested under bit-flip and depolarizing noise, readout errors, two emulated IBM devices (FakeHanoiV2 and FakeLima), coherent over-rotations, and phase and amplitude damping, with exact and finite-shot values. Its largest margin is under coherent noise (MAE 0.0630 vs 0.1143 for the best baseline, Table 4), where circuit-encoding methods have no handle on the error, and in the finite-shot many-body setting (Table 3) it halves the unmitigated error while ZNE and DAEM do worse than nothing.

Critical weaknesses
  1. The proxy is never analyzed, and simple cases break it. The Paulis to measure come from the ideal observable U†[O]\mathcal{U}^\dagger[O], a stand-in for the noisy E†[O]\mathcal{E}^\dagger[O] that sets the responses cic_i. No result says when this is a good proxy; it enters Theorem 3.3 only through a term that is never worked out. With dephasing between RX(θ)R_X(\theta) and RX(−θ)R_X(-\theta) and O=ZO=Z, the ideal observable is ZZ but the noisy one has a YY term the proxy cannot see; a coherent over-rotation by ϵ\epsilon moves weight sin⁡2ϵ\sin^2\epsilon onto a Pauli the ideal circuit never produces (the case of Table 4).

  2. The size of KK is not controlled, and cost grows with it. Algorithm 1 picks K⋆=arg⁡min⁡K{2E(K)2+K/2M}K^\star=\arg\min_K\{2E(K)^2+K/2M\}, and nothing bounds K⋆K^\star. The truncation term E(K)E(K) only becomes small once KK nears the number of Pauli terms in the ideal observable, which can reach 4n4^n. When the coefficients are spread evenly over NN terms, the score is linear in KK, 2L2(1−K/N)+K/2M2L^2(1-K/N)+K/2M, so the algorithm picks either K=1K=1 or the whole basis, K⋆=NK^\star=N, the latter once M>N/4L2M>N/4L^2. The paper's swap test is an example: propagating the measured ZancZ_{\rm anc} back through the five controlled-SWAP gates gives Zanc⊗SWAP⊗5Z_{\rm anc}\otimes\mathrm{SWAP}^{\otimes5}, i.e. 45=10244^5=1024 Pauli terms with equal coefficients. Top-KK has no reason to keep some and drop others, so Algorithm 1 selects all 10241024 once MM exceeds a few hundred: about 2×1062\times10^6 shots per circuit at M=1000M=1000, and a "subspace" that is the whole space. K⋆K^\star, MM, and shot counts are never discussed, and the only experiment touching the selection is the ablation in Table 5, whose rows switch the PR context, SC context, and Top-KK on and off; in three of the four row pairs that differ only in Top-KK, turning it on makes the MAE worse.

  3. The network design is not explained, and no other design is tried. The predictor gϕg_\phi is the main learned part of the method, but Sec. 3.2 only describes it without justifying it. Each part gets one sentence: one output per Pauli, a mean-pooled "PR context" over the responses, a mean-pooled "SC context" over the hidden states, and a zero-initialized output layer. The paper never says why these parts are needed or why mean pooling is the right way to combine the KK entries. No design question is settled by experiment either: there is no comparison with simpler models on the same inputs, such as ridge regression, a plain MLP, or a model with a single output that predicts the correction yideal−ynoisyy^{\rm ideal}-y^{\rm noisy} directly (Eq. 12 shows one output is enough: the optimal coefficients are a scalar multiple of ct\mathbf{c}_t). The only evidence is the ablation in Table 5, which switches the design's own parts on and off; there the SC context alone changes nothing (0.0375 vs 0.0376). So one cannot tell whether this network is needed or any regressor on the same inputs would do.

Questions for authors
  1. When does the Top-KK support of the ideal U†[O]\mathcal{U}^\dagger[O] capture most of the weight of the noisy E†[O]\mathcal{E}^\dagger[O]? Can a bound in terms of the noise strength be given?

  2. Is there a bound on K⋆K^\star in terms of nn, the depth, or the noise strength? Is there a class of circuits for which K⋆K^\star stays far below 4n4^n?

  3. Why this network? On the same (ynoisy,c)(y^{\rm noisy},\mathbf{c}) data, how do ridge regression (Eq. 12), a plain MLP, and a single-output model predicting yideal−ynoisyy^{\rm ideal}-y^{\rm noisy} compare?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ffxb (part 1)

Dear Reviewer ffxb,

Thanks for your constructive feedback and thoughtful review. Below, we provide detailed responses to your questions and comments regarding our paper. We hope that our responses adequately address all your concerns.

Q1 + W1: When does the Top-KK support of the ideal U†[O]\boldsymbol{\mathcal U^\dagger[O]} capture most of the weight of the noisy E†[O]\boldsymbol{\mathcal E^\dagger[O]}? Can a bound in terms of the noise strength be given?:

Answer: We acknowledge that Pauli directions selected through ideal Heisenberg propagation may not capture all coefficients introduced or amplified by noise. Theorem 3.3 provides a sufficient condition for the selected support to capture most of the noisy Pauli weight. Specifically, let SK\mathcal S_K denote the support obtained through sequential Top-KK ideal propagation, DD the discrepancy between the ideal and noisy Heisenberg observables in Pauli coefficient norm, and E(K)E(K) the accumulated truncation error. Then, from the proof of Theorem 3.3, the discarded noisy Pauli weight satisfies $\sum_{P_i\notin\mathcal S_K} |\alpha(\boldsymbol{\mathcal E}^\dagger[O],P_i)|^2 \leq (D+E(K))^2.$ Consequently, the selected support captures at least 1−η1-\eta of the total squared noisy Pauli weight whenever $D+E(K)\leq \sqrt{\eta} |\boldsymbol{\alpha}(\boldsymbol{\mathcal E}^\dagger[O])|_2.$ However, our method targets settings where the noisy quantum process has not been fully characterized, making DD cannot generally be evaluated directly. We therefore select KK by minimizing the KK-dependent terms in Theorem 3.3. which establishes an upper bound on the expected squared approximation error between the true noisy Pauli coefficients and their truncated, finite-shot estimates. This criterion balances truncation and estimation errors without requiring explicit knowledge of the noisy Pauli coefficients. However, our method does not guarantee the recovery of noise-induced Pauli directions. Establishing an explicit noise strength bound that guarantees preservation of the dominant Pauli support is therefore beyond the scope of this work.

W2: The size of KK is not controlled, and cost grows with it [...].

Answer: We thank the reviewer for raising this important question regarding the scaling behavior of $K^* $ . However, the uniformly distributed Pauli coefficients unit total squared coefficient mass at each step assumption do not generally hold for our sequential Top-KK propagation procedure. Specifically, since O~l−1=Top⁡-K(Al−1)\widetilde O_{l-1}=\operatorname{Top}\text{-}K(A_{l-1}) and Al−2=Ul−1†[O~l−1]A_{l-2}=\mathcal U_{l-1}^\dagger[\widetilde O_{l-1}], the orthogonality of the Pauli basis and unitary invariance imply ∥α(Al−2)∥22=∥α(Al−1)∥22−Δl−1(K).\lVert \boldsymbol{\alpha}(A_{l-2}) \rVert_2^2 = \lVert \boldsymbol{\alpha}(A_{l-1}) \rVert_2^2 - \Delta_{l-1}(K). Thus, whenever nonzero coefficient mass is discarded, the total squared coefficient mass decreases in subsequent propagation steps. Furthermore, the coefficient distribution at each step depends on the truncations performed at preceding steps. Consequently, the accumulated truncation error E(K)E(K) generally exhibits a nonlinear dependence on KK, and the linear relationship derived under the uniform-distribution assumption does not apply in general. To provide empirical evidence, we have added Figure 4 to the revised manuscript, which illustrates the distributions of the selected K∗K^* and Pauli coefficients α\alpha. The results indicate that the selected support sizes remain below the full Pauli-space dimension 4n4^n and skew left in the evaluated circuits while the Pauli coefficients are nonuniformly distributed.

Regarding the SWAP-test example, we agree that exact Heisenberg propagation through the five controlled-SWAP gates yields 1024 equally weighted Pauli terms. However, following the SWAP-test implementation [1], we decompose the controlled-SWAP gates into {H,S,T,CNOT}\{H,S,T,\mathrm{CNOT}\}. Since our method applies Top-KK truncation at each propagation step, the accumulated error E(K)E(K) depends on these intermediate distributions rather than solely on the 1024 equally weighted terms. To illustrate this distinction, we evaluated the accumulated truncation error using the decomposed SWAP-test circuit and obtained E(1)=2.414E(1)=2.414 and E(1024)=4.007E(1024)=4.007. These results demonstrate that different truncation levels generate different intermediate observables. In particular, for M=1000M=1000, the corresponding objective values 2E(K)2+K/(2M)2E(K)^2+K/(2M) are approximately 11.657 and 32.618, respectively. Hence, K=1K=1 yields a lower objective value than K=1024K=1024 in this implementation, contrary to the conclusion obtained by considering only the uniform coefficients of the exact final observable. We thank the reviewer for highlighting this point. We will revise the manuscript to clarify the controlled-SWAP gate decomposition.

[1] Liao, Manwen, et al. "Noise-agnostic quantum error mitigation with data augmented neural models." npj Quantum Information 11.1 (2025): 8.

AuthorsOpenReview ↗

Response to Reviewer ffxb (part 2)

Q2: Is there a bound on $K^ $ in terms of nn, the depth, or the noise strength? Is there a class of circuits for which K∗K^* stays far below 4n4^n?:*

Answer: For arbitrary nn-qubit circuits, the general upper bound is K∗≤4nK^* \leq 4^n. Our theoretical analysis provides a more refined characterization of $K^* $ through the accumulated truncation error E(K)E(K), which incorporates the effect of circuit depth LL. Thus, the optimal $K^* $ depends on both circuit depth and the distribution of propagated Pauli coefficients, though its dependence on depth is not necessarily monotonic.

For specific circuit classes, stronger theoretical guarantees are available. In particular, Clifford circuits preserve single Pauli support, yielding K∗=1K^* =1 for a Pauli observable. More generally, Clifford circuits containing tt non-Clifford Pauli rotations generate at most 2t2^t Pauli terms under exact propagation. Hence, the Pauli support remains polynomially bounded when t=O(log⁡n)t=O(\log n). For general non-Clifford circuits, however, the exact Pauli support may grow exponentially in the worst case. We agree that identifying broader classes of quantum circuits that guarantee both compact Pauli support and small approximation error is an important and interesting research direction. While a comprehensive investigation of such circuit classes is beyond the scope of the present study, we believe this represents a valuable avenue for future work.

Q3 + W3: Why this network? On the same (ynoisy,c)(\boldsymbol y^{\mathrm{noisy}}, \boldsymbol c) data, how do ridge regression (Eq. 12), a plain MLP, and a single-output model predicting ynoisy−yideal\boldsymbol y^{\mathrm{noisy}} - \boldsymbol y^{\mathrm{ideal}} compare?

Answer: We have conducted additional experiments comparing QEM-STEP against simpler regression models, including a plain MLP, GTranQEM, and ridge regression, using the same input features (ynoisy,c)(\boldsymbol y^{\mathrm{noisy}}, \boldsymbol c). We also introduce a scalar-output variant of QEM-STEP. QEM-STEP achieves the lowest MAE, outperforming all alternative architectures. The scalar-output variant yields a higher MAE of 0.0432, suggesting that explicitly learning Pauli-specific correction coefficients is more effective than directly predicting a scalar correction. Furthermore, augmenting MLP and GTranQEM with the response vector c\boldsymbol c degrades their performance. Ridge regression using the same augmented inputs also achieves a higher MAE of 0.0709. These findings provide empirical support for the proposed structured prediction formulation, indicating that its advantage cannot be attributed solely to the availability of the additional response features c\boldsymbol c. Rather, the results suggest that learning Pauli-specific perturbation coefficients and combining them through the structured correction rule contributes to improved mitigation accuracy.

MethodMAE ↓CIR ↑Shots
QEM-STEP0.03730.6833489,700
QEM-STEP (scalar output)0.04320.6567489,700
MLP0.04180.63331,000
MLP + c\boldsymbol c0.10090.3933489,700
GTranQEM0.03950.663314,000
GTranQEM + c\boldsymbol c0.09360.4400502,700
Ridge regression + c\boldsymbol c0.07090.5433489,700
Reviewer_ffxbOpenReview ↗

Staying at a weak accept.

Thank you for the additional results on the networks and clarification/ confirmation of my concerns on dependence on K and on selection of Pauli directions. Given these considerations, I will retain my assessment of a weak acceptance.

Dear Reviewer ffxb,

Thank you for your thoughtful and constructive review. We sincerely appreciate the time and consideration you have devoted to evaluating our work.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.