acceptodds
Under review as a conference paper at ICLR 2027

SGIQ: Sparse-Gradient-Informed Gradient Estimation for Variational Quantum Circuits

Abstract

Gradient estimation is a major measurement bottleneck in training complex variational quantum circuits (VQCs) because backpropagation scaling is unattainable for general cases with unknown quantum inputs and no quantum memory. Existing gradient estimators therefore face a cost-accuracy trade-off: parameter-shift rule (PSR) incurs a cost linear in the parameter dimension, while simultaneous perturbation stochastic approximation (SPSA) uses few evaluations but can suffer from dimension-dependent variance. To improve this trade-off, we exploit settings in which loss-gradient energy concentrates on a small subset of parameters. We formalize this property through gradient-energy sparsity and derive sufficient conditions from causal-cone, commutation, and response-decay mechanisms. Building on this structure, we propose , a parse-radient-nformed uantum gradient estimation algorithm that uses averaged SPSA estimates to rank all circuit parameters, replaces the selected coordinate estimates with PSR gradients, and retains the SPSA estimates elsewhere. Its cost scales with the circuit's gradient structure, rather than the parameter dimension. In ideal and finite-shot regression experiments, SGiQ uses fewer gradient circuits than PSR and attains lower normalized test mean absolute error (MAE) than SPSA baselines.

Primary area:
optimization
Keywords:
quantum machine learning, variational quantum algorithms, gradient estimation, sparse gradients

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_u9pAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies measurement-efficient gradient estimation for variatonal quantum circuits under gradient energy concentration. The hybrid estimator is intuitive, and the theory is more substantial than is typical for heuristic SPSA/PSR combinations. However, the evidence is too narrow to establish practical value: experiments use two small synthetic tasks, engineered circuit structures, three seeds, and one data split, while relevant hybrid baselines are omitted.

Critical strengths

SGiQ averageSPSA directions to rank d gradient coordinates, replaces the top-k estimates with parameter-shift-rule gradients, and retains SPSA values elsewhere. Its per-step cost is 2BQ(m+k), versus 2BQd for full PSR. The paper formalizes gradient-energy sparsity; gives sufficient conditions based on causal cones, commutation, and response decay and equal-budget results. Simulations on five- and eight-qubit Friedman regression models report lower error than selected SPSA configurations at lower cost than PSR.

Critical weaknesses

Evaluation is limited to two synthetic datasets, 5 to 8 qubits. There is no VQE, classification, realistic hardware noise, or hardware experiment like IBM machines.

The circuits include parameters whose gradients are identically zero by construction. Concentration is measured post hoc on one probe batch at four epochs, so its prevalence in less favorable circuits is unclear.

Hyperparameters m and k are manually screened, how shall we us adaptive rule selects k?

The theory assumes selection independent of current SPSA noise. The actual algorithm violates this; the recovery argument requires a persistent top-k gap that is not measured. Convergence is analyzed for SGD with fixed readout, whereas experiments use Adam and train the readout.

Questions for authors

Can you report top-k gaps and selection precision during training to test Lemma 3’s assumptions? How can k be selected online without full-gradient access or expensive tuning? Do gains persist across independent splits, more seeds, nonengineered (more general) anastz, and realistic noise?

4 comments
AuthorsOpenReview ↗

Response to W1.

W1: Limited experimental validation.

Our response to W1: We thank the reviewer for this comment. We extended our experimental validation along each requested dimension:

  • System size. We scaled both circuit families of the paper to 30 and 60 qubits, with up to 900 parameters.
  • Task. We added a classification task on quantum data using a QCNN.
  • Noise. We evaluated SGiQ under calibrated IBM device noise combined with finite-shot (1024-shot) measurement noise.

In all comparisons, the methods share the data split, initial parameters and batch order. SGiQ is compared with SPSA at the same or a larger budget of training gradient circuits.

Table W1-a. Larger systems. Both circuit families of the paper scaled to 3030 and 6060 qubits under ideal simulation. Each family keeps its readout from the paper: the global observable ⟨Z⊗n⟩\langle Z^{\otimes n}\rangle for the Figure 2 circuit and local single-qubit ZZ observables for the Figure 1 circuit. Entries report Test MAE (mean±SD\text{mean} \pm \text{SD}), with the number of training gradient circuits in parentheses.

Circuit (task)Qubits / paramsPSRSPSASGiQSGiQ vs. SPSA
Figure 2 (Friedman1)30 / 4500.0967 ± 0.0161 (5.40M)SPSA-300: 0.0863 ± 0.0136 (3.60M)SGiQ(240,60): 0.0799 ± 0.0041 (3.60M)7% lower error and more than 3× smaller SD at the same budget; also more accurate than PSR
Figure 2 (Friedman1)60 / 9000.0989 ± 0.0060 (10.8M)SPSA-680: 0.2704 ± 0.1500$^\ast$ (8.16M)SGiQ(480,200): 0.1588 ± 0.0532 (8.16M)trains in 3/3 seeds vs. 1/3; 41% lower error at the same budget
Figure 1 (Friedman2)30 / 1500.0531 ± 0.0004 (1.80M)SPSA-50: 0.0564 ± 0.0042 (0.60M)SGiQ(25,18): 0.0527 ± 0.0027 (0.49M)7% lower error with 19% fewer circuits
Figure 1 (Friedman2)60 / 3000.1325 ± 0.0021 (3.60M)SPSA-40: 0.1264 ± 0.0060 (0.48M)SGiQ(20,10): 0.1257 ± 0.0054 (0.34M)lower error with 28% fewer circuits

With the global observable, the 6060-qubit Figure 2 circuit lies in a barren-plateau regime, where SPSA frequently stalls at the constant predictor (test MAE 0.35700.3570).

Table W1-b. Classification and calibrated device noise. Entries report Test MAE (mean ± SD over 3 seeds) with the number of training gradient circuits is in parentheses.

SettingPSRSPSASGiQSGiQ vs. SPSA
QCNN, cluster-state classification, 8 qubits, 273 params0.3649 ± 0.0063 (3.28M)SPSA-160: 0.3792 ± 0.0130 (1.92M)SGiQ(120,14): 0.3747 ± 0.0046 (1.61M)lower error and nearly 3× smaller SD with 16% fewer circuits
Friedman1, 5 qubits, FakeMarrakesh noise model, 1024 shots0.1195 ± 0.0110 (0.96M)SPSA-60: 0.1505 ± 0.0170 (0.72M)SGiQ(50,10): 0.1340 ± 0.0089 (0.72M)11% lower error and half the SD at the same budget
Friedman1, 30 qubits, local readout, FakeMarrakesh noise model, 1024 shots0.0741$^\dagger$ (5.40M)SPSA-180: 0.0710 ± 0.0041 (2.16M)SGiQ(135,45): 0.0691 ± 0.0067 (2.16M)lowest mean error of all three methods, with 40% of the PSR budget

These results establish three conclusions:

  • The advantage of SGiQ over SPSA persists at scale. On both circuit families and at 30 and 60 qubits, SGiQ is more accurate than SPSA in all four settings, using either the same budget or 19–28% fewer circuits. In three of the four settings it also matches or exceeds the accuracy of PSR with 10–67% of the PSR budget. The gain is largest when the gradient signal is weakest: in the 60-qubit barren-plateau regime, SPSA stalls at the constant predictor in 2 of 3 seeds, whereas SGiQ trains in every seed and reduces the error of SPSA by 41%.
  • The advantage carries over to classification. On the QCNN, where every gate has its own parameter, SGiQ reaches a lower error than SPSA with 16% fewer circuits and a nearly 3× smaller seed-to-seed SD, using half of the PSR budget.
  • SGiQ is robust to hardware noise. Under calibrated device noise with 1024 shots, SGiQ reduces the error of SPSA by 11% at 5 qubits, and at 30 qubits it attains the lowest mean error of the three methods with 40% of the PSR budget. In six of the seven comparisons, SGiQ also has a smaller seed-to-seed SD than SPSA. This is consistent with Proposition 5: refining coordinates that carry a fraction ρ^\widehat\rho of the gradient energy with exact PSR values multiplies the directional-variance bound of SPSA with the same mm by 1−ρ^1-\widehat\rho.

We will add both tables to Appendices D.2 and D.3 of the revised paper. Further results are provided in our responses to Reviewer tHe4 (Tables R1 through R7) and Reviewer n1E4 (Q4).

AuthorsOpenReview ↗

Response to W2.

W2: Engineered circuit structures and gradient concentration.

Our response: We thank the reviewer for this insightful comment. In brief: (1) the structurally zero-gradient parameters are few (5 of 80 and 8 of 48) and come from circuits inherited from prior work, which serve as a controlled testbed for our theory and algorithm; more importantly, our conclusions still hold when these parts are removed. (2) The concentration is not a snapshot, as it holds at every checkpoint and probe batch we measured. (3) We do not claim that concentration is universal, only that it arises in identifiable classes of circuits, which we now confirm on a standard QCNN and on circuits with up to 60 qubits. Whether it is prevalent in practical circuits in general is an interesting and important open question, and we see it as a key direction for future work.

1. Zero gradients by construction. We agree that both circuits contain structurally zero parameters. The circuits are inherited from prior work (Skolik et al., 2021), and their final-layer RzR_z gates commute with the subsequent CZ gates and ZZ readouts (5 of 80 parameters in Friedman1 and 8 of 48 in Friedman2). We view this as a useful controlled setting: because these parameters are known exactly, our theory and algorithm can be checked against them. They do not change our conclusions. Since they carry no gradient energy, they leave ρk\rho_k and every kρk_\rho in Table 1 unchanged. When their cost is removed from full PSR, SGiQ uses 66.7% and 80% of the PSR budget for (40,10)(40,10) and (40,20)(40,20) in Friedman1, instead of 62.5% and 75%, and 42.5% instead of 35% in Friedman2.

We further evaluated an 8-qubit QCNN (Cong et al., 2019) adapted from the official TensorFlow Quantum tutorial (Broughton et al., 2020), with untied parameters, and reached the same conclusions, and SGiQ attains lower mean test error than SPSA at the same or a smaller budget (Table W1-b). In this standard architecture, the only structurally zero parameters are the final ZZ rotations on the source qubits of the pooling units, which commute with the subsequent CNOT. Such zeros therefore also arise in circuits designed by others.

2. Post hoc measurement. Table 1 is a diagnostic that SGiQ never uses: Algorithm 1 selects coordinates from SPSA probes without prior knowledge of concentration. To show that it is not a favorable snapshot, we measured concentration at 7 checkpoints (epochs 0, 5, ..., 30) on 3 non-overlapping probe batches, counting only the active parameters (dactive=75d_{\text{active}} = 75 for Friedman1 and 40 for Friedman2).

Table D. Number of active parameters kρk_\rho capturing a fraction ρ\rho of the gradient energy, with an i.i.d. Gaussian vector of the same dimension as a reference.

Task (dactived_{\text{active}})ρ=0.7\rho = 0.7ρ=0.8\rho = 0.8ρ=0.9\rho = 0.9
Friedman1 (75)16.5±1.116.5 \pm 1.122.8±1.222.8 \pm 1.232.5±1.332.5 \pm 1.3
Random reference (75)18.324.634.2
Friedman2 (40)2.9±0.72.9 \pm 0.74.1±1.34.1 \pm 1.36.3±1.56.3 \pm 1.5
Random reference (40)10.213.518.6

Concentration is stable from initialization to the end of training. In Friedman2, 4 to 6 of 40 active parameters carry 80% of the gradient energy, compared with about 13 for a random vector. Friedman1 is close to the random reference, i.e. essentially a dense-gradient case, where SGiQ still improves over the SPSA baselines (Table 2). The coverage ρ^\hat\rho of the coordinates SGiQ actually selects, which is the quantity entering our guarantees, is reported at every training step in our answer to Q1. We will add Table D to Appendix D.1.

3. Positioning and prevalence. We do not claim that gradient-energy concentration arises in every circuit. Our claim is that it arises in identifiable classes of circuits, through mechanisms we characterize theoretically (Propositions 1–3) as sufficient conditions, and that SGiQ exploits it when present. Whether concentration is prevalent across general VQC architectures is, in our view, an open question that deserves dedicated study, and we will state it as such in the revision. Our results establish that it arises in several distinct settings: at fixed d=240d = 240, 80% of the gradient energy lies in 1–31% of the parameters with a local readout (Reviewer n1E4, Q3) and in 14–31% with an all-qubit readout, a QCNN in which every parameter lies in the causal cone of the measured qubit, and 30- and 60-qubit circuits with global readouts ⟨Z⊗n⟩\langle Z^{\otimes n}\rangle, where SGiQ also attains lower mean test error than SPSA at the same or a smaller budget (Tables W1-a and W1-b). The mechanisms we identify also offer guidance for designing ansätze whose gradients can be estimated efficiently. When gradient magnitudes are nearly uniform, we expect no advantage over equal-budget SPSA, and we state this as a limitation.

AuthorsOpenReview ↗

Response to W3 & W4.

W3: Hyperparameter selection for mm and kk.

Our response: We use grid search for the hyperparameters mm and kk to study the sensitivity of hyperparameters. In practice, there is an empirical adaptive strategy for the online selection: m+k∼0.5d,k∼10%d,0.1≤k/(m+k)≤0.4m+k \sim 0.5d, k \sim 10\%d, 0.1 ≤ k/(m+k) ≤ 0.4. The performance is unless sensitive to the hyperparameters. We add one more ablation study shown in the following table.

Table R1. SGiQ configurations with 0.1≤k/(m+k)≤0.40.1 \le k/(m+k) \le 0.4 versus SPSA at equal budget. Test MAE, mean ± s.d. over 3 seeds unless noted; budget = number of training gradient circuits.

SettingddSGiQ (m,k)(m,k)k/(m+k)k/(m+k)BudgetSGiQSPSA (equal budget)PSR (budget)
30 qubits, local readout, ideal450(153, 27)0.152.16M0.0842 ± 0.01490.0868 ± 0.02190.0713 (5.4M)
(144, 36)0.202.16M0.0830 ± 0.00340.0868 ± 0.0219
(135, 45)0.252.16M0.0837 ± 0.00370.0868 ± 0.0219
(126, 54)0.302.16M0.0819 ± 0.00600.0868 ± 0.0219
(112, 68)0.382.16M0.0787 ± 0.00150.0868 ± 0.0219
(255, 45)0.153.60M0.0720 ± 0.00590.0768 ± 0.0111
30 qubits, local readout, 1024 shots450(153, 27)0.152.16M0.0783 ± 0.00450.0843 ± 0.00690.0784 (5.4M)
30 qubits, FakeMarrakesh noise, c=0.1c=0.1450(135, 45)0.252.16M0.0691 ± 0.00670.0710 ± 0.00410.0741 (5.4M, 1 seed)
30 qubits, FakeMarrakesh noise, c=0.5c=0.5450(135, 45)0.252.16M0.0755 ± 0.00760.0826 ± 0.0082
60 qubits, local readout, 1024 shots900(510, 90)0.157.20M0.0534 (1 seed)0.0622 ± 0.00390.0626 ± 0.0036 (10.8M)
(465, 135)0.237.20M0.0611 ± 0.00440.0622 ± 0.0039
QCNN (unshared parameters), ideal273(95, 14)0.131.31M0.38710.43680.3649 ± 0.0052 (3.28M)
(68, 41), EMA0.381.31M0.38190.4368
(142, 41)0.222.20M0.37600.3975
(156, 27), EMA0.152.20M0.37710.3975
(120, 14)0.101.61M0.3747 ± 0.00380.3792 ± 0.0106 (1.92M)
Figure 1 circuit, 60 qubits, 3 layers300(20, 10)0.330.35M0.1257 ± 0.00540.1375 ± 0.0065 (0.30M); 0.1264 ± 0.0060 (0.48M)0.1325 ± 0.0021 (3.6M)

W4: Selection independence, top-kk gap, and SGD with fixed readout versus Adam with trained readout.

Our response: We thank the reviewer for these observations and address them in turn.

1. Selection and Assumption 2. We agree that Algorithm 1 selects SS from the same gˉ\bar g and therefore violates Assumption 2, as we state in Section 4.2. This does not leave Algorithm 1 without guarantees. Proposition 4(i) and Theorem 1(ii) hold for any selection rule, including Algorithm 1 as implemented, and give convergence without independence or a gap condition. The gap condition of Lemma 3 only serves to transfer the sharper bounds of Theorem 1(i), and even without a gap, Eq. 12 bounds the loss in coverage by the estimation error. Empirically, the selection error is small. We report the measured gaps, selection precision, and coverage in detail in our answer to Q1.

2. Adam and trained readout. Our convergence analysis is for SGD with a fixed readout, whereas the experiments use Adam and train the readout. We use Adam to following the training protocol of previous works (Periyasamy et al. 2024) as Adam is the most popular optimizer.

We also have compare all methods under SGD. Specifically, we ran Friedman1 (55 qubits) with plain SGD (learning rate selected on the validation set from {0.03,0.1,0.3,1,3,10}\{0.03, 0.1, 0.3, 1, 3, 10\}). Entries below report Test MAE (mean±SD\text{mean} \pm \text{SD}), with training gradient circuits in parentheses:

SettingPSRSPSASGiQ
Ideal0.0884±0.01070.0884 \pm 0.0107 (0.96M0.96\text{M})SPSA-5050: 0.1150±0.01550.1150 \pm 0.0155 (0.60M0.60\text{M})SGiQ$(25,25)$: 0.1114±0.00700.1114 \pm 0.0070 (0.60M0.60\text{M})
10241024 shots0.1002±0.01360.1002 \pm 0.0136 (0.96M0.96\text{M})SPSA-5050: 0.1174±0.01200.1174 \pm 0.0120 (0.60M0.60\text{M})SGiQ$(40,10)$: 0.1087±0.01210.1087 \pm 0.0121 (0.60M0.60\text{M})

In both settings, SGiQ achieves lower test error than equal-budget SPSA-5050. In the ideal setting, every evaluated SGiQ configuration with k≤d/3k \le d/3 (0.10020.1002 to 0.11140.1114) outperforms SPSA-5050, confirming that the ordering holds under plain SGD as well as under Adam.

AuthorsOpenReview ↗

Response to Q1, Q2 & Q3.

Q1: Top-k gaps and selection precision during training.

Our response: We measured Top-k gaps and selection precision during training. Specificallhy, for SGiQ(40,10) along the training trajectories of the paper's two tasks (5 seeds, 7 checkpoints from epoch 0 to 30, 3 probe batches, 100 estimator draws per state), comparing the 10 coordinates selected before PSR refinement with the true top-10 of the PSR gradient.

Table E. Top-10 gap and selection quality during training (ranges over the 7 checkpoints).

Task10th vs. 11th largest gradientPrecision (selected indices in true top-10)Energy share of true top-10Energy share of selected 10Selected / true top-10 energy
Friedman1+3–5%0.32–0.410.52–0.550.27–0.3451–61%
Friedman2+10–18%0.59–0.820.93–0.960.81–0.9386–97%

Top-k gap. The gap is small: the 10th largest gradient magnitude exceeds the 11th by only 3%–5% (Friedman1) and 10%–18% (Friedman2). The gap condition of Lemma 3 for exact index recovery is therefore generally not met, and the selection precision is 0.32–0.41 and 0.59–0.82.

Energy of the selected set. The misranked coordinates lie near the boundary and have almost the same magnitude as the ones they replace, so little energy is lost: the selected set keeps 51%–61% (Friedman1) and 86%–97% (Friedman2) of the true top-10 energy. Its share of the total gradient energy (0.27–0.34 and 0.81–0.93) stays above k/(m+k)=0.20k/(m+k)=0.20 at every checkpoint, which is the condition under which Corollary 2 gives SGiQ an advantage over equal-budget SPSA. Hence SGiQ does not need the exact top-kk indices, only a high-energy subset.

Revision. We add Table E to the appendix and state after Lemma 3 that its gap condition concerns exact index recovery only.


Q2: Online selection of kk without full-gradient access.

Our response: Thank you for this comment. Please see rebutal to W3.


Q3: Robustness across independent splits, seeds, ansätze, and realistic noise.

Our response: We thank the reviewer for this question. We address each factor below, supported by our new experiments in W1 and W2.

  • Seeds. All new training benchmarks use 33 paired seeds sharing identical parameter initialization, data split, and mini-batch order across methods, while our gradient-concentration and selection diagnostics use 55 seeds. Across 66 of the 77 experimental settings reported in W1, SGiQ not only improves mean accuracy over equal-budget SPSA, but also achieves a smaller seed-to-seed standard deviation.
  • Non-engineered ansatze and readouts. SGiQ does not rely on specially engineered zero-gradient parameters. Its gains persist across two general circuit families scaled to 3030 and 6060 qubits under both global and local readouts, as well as on a QCNN classification benchmark where every gate has an independent parameter and no zero-gradient construction is present (see W1 and W2).
  • Realistic device noise. We evaluated SGiQ under calibrated IBM device noise models with 10241024 shots (see W1). On the 55-qubit benchmark, SGiQ reduces the test error of equal-budget SPSA by 11%11\% with half the seed-to-seed standard deviation. On the 3030-qubit benchmark, SGiQ achieves the lowest mean error among all three methods (outperforming both equal-budget SPSA and full PSR) while using only 40%40\% of the PSR circuit budget.
  • Independent data splits. In our experiments, all methods share the same data split so that the comparisons are paired and differences between methods are not caused by the split.
Reviewer_n1E4OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper concerns efficient estimation of quantum circuit gradients. The paper’s central idea is that not all gradient components are equally informative at each step, and therefore, every coordinate need not be estimated to the same accuracy. To this end, the paper defines a concept of “gradient sparseness” and uses it to define a gradient estimation protocol called SGIQ, which is positioned as a middle ground between PSR and SPSA.

Critical strengths

The idea of dynamically shifting accuracy towards coordinates with more impact on the overall gradient is an interesting direction. This is certainly a tradeoff that has not been fully explored in the literature. I'm curious to see how this framework can be extended to general PSR such as (https://arxiv.org/abs/2207.01587). I'm also curious to see if and how this intelligent gradient estimation can impact trainability (e.g. barren plateus). The paper is well-written and easy to follow which I appreciate.

Critical weaknesses
  1. Line 127 assumes that the observable OαO_\alpha is composed of local Pauli observables with small locality. This is a significant assumption, and its impact on the final results should be clearly explained and stated. It may be the case that SGIQ’s cost saving measure diminishes with observable locality.

  2. Corollary 2 and 3’s cost analysis while informative lacks clarity: it is not clear to me in what regimes SGIQ is expected to be cost effective, and to what degree. This is especially relevant as the experimental results are limited to a small number of qubits, and small-qubit studies may not predict impact in more realistic settings. The experimental results do not address this clearly: Table 1 reports different cost budgets and it is not clear how they should be interpreted. From Figure 6 it would seem that all methods are statistically indistinguishable in the studied setting. Given the experimental results presented, an expanded theoretical overhead analysis seems necessary.

The paper’s idea, while theoretically interesting and sound, is not quite clear in its experimental validation and certain technical details.

Questions for authors
  1. Can the authors clarify the context of line 127?
  2. In Table 1, why did the authors choose different coverage thresholds?
  3. Can the authors provide an analysis that compares qubit count directly with cost, and final accuracy?
  4. How should the gap in MAE and MSE accuracy between PSR and SPSA be interpreted? Have the authors considered a cost-normalized accuracy measure that removes the budget discrepancy between methods?
4 comments
AuthorsOpenReview ↗

Response to W1 & Q1.

W1 & Q1: Local Pauli observable assumption (Line 127).

Our response to W1 & Q1: We thank the reviewer for this suggestion. The local-observable setup in Line 127 was introduced solely in Section 3.2 to illustrate concrete circuit settings in which gradient-energy sparsity arises, specifically the causal-cone and decaying-response examples. It is not a general assumption required by SGiQ or its theoretical guarantees.

The three mechanisms in Section 3.2 show sufficient circuit conditions under which sparsity can arise; they do not exhaust the settings in which SGiQ can be applied. SGiQ identifies important coordinates directly from SPSA gradient estimates and does not require prior knowledge of which mechanism produces their concentration.

Observable locality is one of several factors that can influence gradient-energy concentration. In the causal-cone and decaying-response examples, a larger observable support may weaken the corresponding concentration guarantees. However, concentration can also arise through commutation (Proposition 2, which requires no locality assumption) or other factors involving the circuit architecture and data. Thus, decreasing observable locality (i.e., increasing observable support or Pauli weight) does not necessarily reduce SGiQ's cost savings; it may do so only if it weakens the gradient concentration that SGiQ exploits, thereby requiring a larger estimation budget (m+k)/d(m+k)/d to maintain comparable performance.

Our new experiments (Table R1 in our response to Reviewer tHe4, and Appendix A.1)) confirm that SGiQ does not rely on a local observable: (i) High-weight observable. On a 30-qubit circuit read out with the global observable ⟨Z⊗30⟩\langle Z^{\otimes 30}\rangle, a single Pauli string acting on all qubits, SGiQ$(240, 60)$ reaches a test MAE of 0.0799±0.00410.0799 \pm 0.0041 at 67%67\% of the PSR budget, achieving lower error than SPSA at the same budget (0.0863±0.01360.0863 \pm 0.0136) and full PSR (0.0967±0.01610.0967 \pm 0.0161). (ii) Readout supported on all qubits. With the parameter count fixed (d=240d = 240), a readout that averages ZZ over all qubits still concentrates the gradient energy: 80%80\% of the gradient energy is carried by 14%14\% to 31%31\% of the parameters across all width-depth allocations. Locality strengthens this concentration further (down to 1.2%1.2\% of the parameters when the readout is restricted to five qubits of a wide, shallow circuit; see details in Appendix A.1), which is consistent with the causal-cone mechanism, but SGiQ does not require it.

AuthorsOpenReview ↗

Response to W2 & Q2.

W2: Cost analysis in Corollaries 2 and 3.

Our response to W2: We thank the reviewer for this constructive comment. SGiQ is cost-effective when the refined coordinates capture more gradient energy than their share of the budget, ρ>k/(m+k)\rho>k/(m+k), and when m+k≪dm+k\ll d, the explicit conditions and factors are given below. Corollary 2 compares gradient estimators at the same per-update budget, whereas Corollary 3 analyzes the circuit budget sufficient to reach a fixed convergence accuracy. We will clarify this distinction and supplement Corollary 3 with an explicit comparison using the same numerical learning rate for all methods.

Corollary 2: one update at equal per-update cost. With SPSA using m′=m+km'=m+k directions, SPSA and SGiQ evaluate the same number of gradient circuits per step. Under Assumption 2, in the ideal setting, to leading order in cc and up to factors common to all methods:

MethodPer-update circuit costSquared biasDirectional-variance bound
PSR2BQd2BQd0000
SPSA, m′=m+km'=m+k2BQ(m+k)2BQ(m+k)O(dc4)O\big(d c^4\big)O(1m+k)O\Big(\frac{1}{m+k}\Big)
SGiQ2BQ(m+k)2BQ(m+k)O((d−k)c4)O\big((d-k) c^4\big)O(1−ρ^m)O\Big(\frac{1-\widehat{\rho}}{m}\Big)

At the same cost, the bias of SGiQ is never larger than that of SPSA, since it removes the bias on the kk refined coordinates. Its descent direction is more reliable if and only if ρ^>km+k\widehat{\rho}>\frac{k}{m+k}, i.e. the refined coordinates capture more energy than their budget share; PSR is exact but costs d/(m+k)d/(m+k) times more per update.

Corollary 3: total cost to reach a target accuracy. We retain the original Corollary 3 and add a comparison for the stationarity target 1T∑t<TE∥∇J(θt)∥22≤ε\frac{1}{T}\sum_{t<T}\mathbb{E}\lVert\nabla\mathcal{J}(\theta_t)\rVert_2^2\le\varepsilon, with the same learning rate for all methods. Below, O(⋅)O(\cdot) hides only a factor common to all methods and terms of order 1/m1/m and ε\varepsilon (exact expressions in Appendix B.8), so the displayed constants can be compared across rows.

MethodPer-update costTotal budget
PSRddO(d4ε2)O\Big(\frac{d^4}{\varepsilon^2}\Big)
SPSAm′m'O(65⋅m′d3ε2)O\Big(\frac{6}{5}\cdot\frac{m'd^3}{\varepsilon^2}\Big)
SGiQm+km+kO(11−(1−ρ)d−k6d⋅(m+k)d3ε2)O\Bigg(\frac{1}{1-(1-\rho)\frac{d-k}{6d}}\cdot\frac{(m+k)d^3}{\varepsilon^2}\Bigg)

Here ρ\rho is a lower bound on ρ^\widehat\rho over the run. At equal budget (m′=m+km'=m+k), the SGiQ factor never exceeds the SPSA factor 6/56/5 and decreases as ρ\rho grows, so SGiQ needs up to about 17% fewer circuits. Compared with PSR, SGiQ is cheaper when m+k<dm+k<d and d−m−kd−k>116m+1−ρ6+ε1536d2.\frac{d-m-k}{d-k}>\frac1{16m}+\frac{1-\rho}{6}+\frac{\varepsilon}{1536d^2}.

The conditions hold in our experiments. For SGiQ(40,10) on Friedman1 (d=80d=80), the measured coverage during training is ρ^=0.27\widehat\rho=0.27–0.340.34 (our response to Reviewer u9pA, Q1). Both conditions are then satisfied: 0.27>k/(m+k)=0.20.27>k/(m+k)=0.2, and d−m−kd−k=0.43\frac{d-m-k}{d-k}=0.43 exceeds the right-hand side (≈0.12\approx0.12). With this coverage, the bound of Corollary 3 gives SGiQ a 7–8% smaller total budget than SPSA. In larger circuits, the gradient remains concentrated while dd grows, so m+k≪dm+k\ll d becomes possible: on the Figure 1 circuit at 30 and 60 qubits, SGiQ reaches PSR-level accuracy with 27% and 10% of the PSR budget, respectively (Tables R2–R3 in our response to Reviewer tHe4).

Revision. We will add these tables and conditions to the discussion of Corollaries 2–3, with the complete derivation and applicability conditions in Appendix B.8, after the proof of Corollary 3.

Q2: Choice of coverage thresholds in Table 1.

Our response to Q2: We thank the reviewer for this question. The original thresholds were illustrative choices to show gradient-energy concentration within each dataset–circuit setting, rather than for ranking the two circuits. To make the presentation consistent, we now report the same three thresholds, ρ∈{0.7,0.8,0.9}\rho\in\{0.7,0.8,0.9\}, for both tasks.

Table A. Gradient-energy concentration after 30 epochs of PSR training, on the training set of the main experiments (200 samples for Friedman1, 100 for Friedman2). Each entry reports the minimum number of parameters needed to capture the specified fraction of the true loss-gradient energy. Counts are first averaged over three probe batches within each seed, then reported as mean ± SD across five seeds.

Task70% Energy (ρ=0.7\rho=0.7)80% Energy (ρ=0.8\rho=0.8)90% Energy (ρ=0.9\rho=0.9)
Friedman2 (Figure 1)2.9 ± 0.7 (7%)4.1 ± 1.3 (10%)6.3 ± 1.5 (16%)
Friedman1 (Figure 2)16.5 ± 1.1 (22%)22.8 ± 1.2 (30%)32.5 ± 1.3 (43%)

For both tasks, a small subset of parameters captures most of the gradient energy at all three thresholds, and kρk_{\rho} serves as a measure of gradient concentration.

Revision. We will report all three thresholds in Table 1, and add the results over all checkpoints and both dataset sizes to Appendix D.1.

AuthorsOpenReview ↗

Response to Q3.

Q3: Relationship between qubit count, cost, and final accuracy

Our response to Q3: [ We thank the reviewer for this question. The answer depends on how cost is measured, so we distinguish two definitions.

1. Total circuit cost over TT steps. PSR, SPSA and SGiQ use 2BQdT2BQdT, 2BQm′T2BQm'T and 2BQ(m+k)T2BQ(m+k)T gradient circuits. The qubit count nn enters only through dd: increasing nn raises the cost of PSR, while the cost of SPSA and SGiQ is fixed by m′m' and m+km+k.

2. Cost to reach a target accuracy ε\varepsilon. In Corollary 3, the qubit count enters both CPSR(ε)C_{\rm PSR}(\varepsilon) and CSGiQ(ε)C_{\rm SGiQ}(\varepsilon) only through dd. The causal-cone structure changes this.

The causal-cone case. When the readout is local, parameters outside the causal-cone set CO\mathcal C_O have zero gradient (Proposition 1); under mild conditions (Appendix A.1), dd in the costs of SPSA and SGiQ can then be replaced by ∣CO∣|\mathcal C_O|, while PSR still pays for all dd parameters. For a ring circuit with readout on five qubits, ∣CO∣≤min⁡{d−2n+5, L(3L+4)}|\mathcal C_O|\le\min\{d-2n+5,\ L(3L+4)\}, which does not grow with nn. Hence:

  • At fixed depth, the costs of SPSA and SGiQ stay bounded as nn grows, whereas that of PSR grows linearly.
  • At fixed dd, more qubits and fewer layers give a smaller cone and thus lower SPSA and SGiQ costs: for d=240d=240, ∣CO∣=235|\mathcal C_O|=235, 113113 and 1717 for 5×245\times24, 20×620\times6 and 60×260\times2 (qubits ×\times layers).

3. Final accuracy. Since ε\varepsilon is an optimization target rather than test error, we check test accuracy experimentally: on the Figure 1 circuit at 30 and 60 qubits, SGiQ matches PSR's test MAE (0.0527 vs. 0.0531; 0.1257 vs. 0.1325) with 27% and 10% of its cost (Reviewer tHe4, Tables R2–R3).

Revision. We will add the causal-cone cost analysis and its proof to Appendix A.1.]

AuthorsOpenReview ↗

Response to Q4.

Q4: MAE/MSE differences and cost-normalized accuracy.

Our response: We thank the reviewer for these questions. In short: (i) the MAE/MSE gap reflects the higher variance of SPSA updates, which occasionally produces large residuals that MSE amplifies more than MAE; (ii) yes, and we normalize cost through the comparison itself: SGiQ is compared with SPSA at exactly the same number of gradient circuits, and we additionally report results at several budgets (Tables R2, R3 and R7 in our response to Reviewer tHe4), rather than a single score that would require an arbitrary exchange rate between error and cost.

(i) The MAE/MSE gap between PSR and SPSA. MSE squares the residuals and is therefore dominated by occasional large errors, whereas MAE reflects the typical error. SPSA estimates the gradient from a few random directions, so its updates have higher variance. This produces occasional runs or samples with large residuals, which MSE amplifies more than MAE. For example, on the 30-qubit local-readout circuit, the seed-to-seed SD of the test MSE is 0.0056 for SPSA, versus 0.0024 for PSR and 0.0007 for SGiQ. SGiQ replaces the top-kk coordinates by exact PSR values, which reduces this variance, and it attains a lower MSE than SPSA in all settings of the table below.

(ii) Cost-normalized accuracy. SGiQ$(m,k)$ is compared with SPSA using m′=m+km'=m+k directions, i.e. exactly the same number of gradient circuits (Corollary 2). This avoids a combined error–cost score, which would depend on an arbitrary exchange rate. Test MAE / MSE (mean over 3 seeds) on the Friedman1 circuit:

SettingBudget / PSRPSRSPSA (same budget)SGiQ
5 qubits, ideal (Table 2)0.630.0849 / 0.0130SPSA-50: 0.1046 / 0.0209(40,10): 0.0950 / 0.0165
5 qubits, 1024 shots (Table 2)0.750.1087 / 0.0209SPSA-60: 0.1207 / 0.0256(40,20): 0.1071 / 0.0208
5 qubits, device noise + 1024 shots0.750.1195 / 0.0260SPSA-60: 0.1505 / 0.0358(50,10): 0.1340 / 0.0304
30 qubits, ideal0.670.0967 / 0.0173SPSA-300: 0.0863 / 0.0148(240,60): 0.0799 / 0.0117
60 qubits, ideal0.760.0989 / 0.0177SPSA-680: 0.2704 / 0.1319(480,200): 0.1588 / 0.0458

Device noise uses the FakeMarrakesh noise model. At equal budget, SGiQ beats SPSA in both MAE and MSE in all five settings; at 30 qubits and with 1024 shots at 5 qubits, it is also more accurate than PSR.

Revision. We will add this table to Appendix D, and the equal-budget SPSA rows to Table 2.

Reviewer_tHe4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed a novel gradient estimation for variational quantum circuits. This method uses averaged SPSA estimates to rank all circuit parameters, replaces the selected coordinate estimates with PSR gradients, and retains the SPSA estimates elsewhere.

Critical strengths

1.The idea is reasonable.

2.The presentation is logical and very easy to follow.

Critical weaknesses

1.The proposed method has substantial limitations: it is only applicable to scenarios where most parameters do not have gradients.

2.The scalability of the proposed approach is not well validated, given that the experiments are only performed on 8-qubit systems. The authors are encouraged to include experiments on systems such as 60 qubits.

3.The adaptability of the proposed approach lacks validation. The authors are strongly suggested to include additional experiments for QGANs and QCNNs, which are widely adopted quantum neural networks.

4.The experimental validation is inadequate. The authors need to include noise experiments to verify the efficacy of their algorithm.

Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to W1.

W1: Applicability of the proposed method.

Our response: We thank the reviewer for the comment and would like to clarify a misunderstanding. SGiQ is applicable to general circuits and includes two standard methods as special cases: SPSA (k=0k = 0) and PSR (k=dk = d).

We believe that this misunderstanding may come from our presentation of the two exact-sparsity mechanisms (Propositions 1 and 2). One of our contributions is to characterize a gradient-energy concentration regime (Definition 1) and establish conditions under which SGiQ provides a tighter descent-direction variance bound than equal-budget SPSA (Corollary 2) and a lower theoretical gradient-circuit complexity bound than PSR (Corollary 3). This regime only requires that a subset of coordinates captures a large fraction of the gradient energy. We illustrate it with three mechanisms. Propositions 1 and 2 (causal cones, commutation) produce exact zeros, but they are special cases showing how common circuit structures lead to concentration. Proposition 3 gives concentration without requiring any zero coordinate.

Our experiments confirm this. In the Friedman1 benchmark (d=80d = 80 with a global ⟨Z⊗5⟩\langle Z^{\otimes 5} \rangle readout), only 5 final-layer RzR_z parameters have structurally zero gradients, so 75 of 80 parameters (93.8%) are not structurally zero. SGiQ still attains lower mean test error than SPSA while using 25%-37.5% fewer gradient circuits than PSR (Table 2).

In summary, our SGiQ applies to general circuits. Gradient-energy concentration is not a condition for using the algorithm. It is the regime in which we rigorously prove its theoretical advantage. We have revised the presentation to make this distinction explicit.

AuthorsOpenReview ↗

Response to W2.

W2: Scalability to larger qubit systems.

Our Response: We thank the reviewer for this constructive suggestion. To thoroughly validate the scalability of SGiQ, we have conducted new 30-qubit and 60-qubit experiments (scaling up to 900 parameters) across both circuit families studied in the paper - the sparse-local architecture (Figure 1, Friedman2) and the hardware-efficient ring architecture with global readout (Figure 2, Friedman1). All experiments are evaluated over 3 random seeds with equal-budget SPSA controls under ideal simulation.

1. 30-Qubit Scaling Results:

Table R1: Figure 2 circuit (Friedman1), 30 qubits, 8 layers (450 active parameters).

MethodBudgetBudget / PSRTest MAETest MSE
PSR5.40M1.000.0967 ± 0.01610.0173
SPSA-3003.60M0.670.0863 ± 0.01360.0148
SGiQ (240, 60)3.60M0.670.0799 ± 0.00410.0117
SPSA-2002.40M0.440.1004 ± 0.01160.0182

Table R2: Figure 1 sparse-local circuit (Friedman2, 200 training samples), 30 qubits.

MethodBudgetBudget / PSRTest MAE
PSR1.80M1.000.0531 ± 0.0004
SPSA-380.46M0.250.0618 ± 0.0129
SGiQ (25, 18)0.49M0.270.0527 ± 0.0027
SPSA-500.60M0.330.0564 ± 0.0042
SPSA-600.72M0.400.0572 ± 0.0056
SGiQ (38, 26)0.73M0.410.0542 ± 0.0003

On the 30-qubit Figure 2 circuit (Table R1), SGiQ (240, 60) reduces the circuit budget by 33% relative to PSR while achieving the lowest Test MAE (0.0799 vs. 0.0863 for equal-budget SPSA-300 and 0.0967 for PSR) with more than 3×3\times lower seed variance. On the 30-qubit Figure 1 circuit (Table R2), SGiQ (25, 18) matches full PSR accuracy (0.0527 vs. 0.0531) using only 27% of the PSR budget, outperforming SPSA across all comparable budgets.

2. 60-Qubit Scaling Results:

Table R3: Figure 1 sparse-local circuit (Friedman2, 300 active parameters), 60 qubits.

MethodBudgetBudget / PSRTest MAE
PSR3.60M1.000.1325 ± 0.0021
SPSA-250.30M0.080.1375 ± 0.0065
SGiQ (20, 10)0.34M0.100.1257 ± 0.0054
SPSA-400.48M0.130.1264 ± 0.0060
SPSA-750.90M0.250.1363 ± 0.0048
SGiQ (50, 30)0.91M0.250.1292 ± 0.0060
SPSA-800.96M0.270.1399 ± 0.0109
SGiQ (50, 35)0.96M0.270.1311 ± 0.0034
SPSA-1001.20M0.330.1308 ± 0.0034
SGiQ (75, 35)1.27M0.350.1318 ± 0.0038

Takeaway (60-qubit sparse-local circuit): As system size scales to 60 qubits, the relative savings of SGiQ become even more pronounced: SGiQ (20, 10) reaches PSR-level accuracy (0.1257 vs. 0.1325) using only 10% of the PSR budget. Across matched budgets, SGiQ consistently outperforms or matches SPSA (e.g., 0.1292 vs. 0.1363 at 0.90M–0.91M; 0.1311 vs. 0.1399 at 0.96M; and equivalent within seed variation at 1.20M–1.27M).

Table R4: Figure 2 circuit (Friedman1), 60 qubits, 8 layers (900 parameters), global readout ⟨Z⊗N⟩\langle Z^{\otimes N}\rangle, U[0,π]U[0,\pi] initialization, ideal setting.

MethodBudgetBudget / PSRTest MAESeeds that train
PSR10.80M1.000.0989 ± 0.00603 / 3
SPSA-6808.16M0.760.2704 ± 0.15001 / 3
SGiQ (480, 200)8.16M0.760.1588 ± 0.05323 / 3

Takeaway (60-qubit global readout & barren-plateau regime): With the 60-qubit global observable ⟨Z⊗60⟩\langle Z^{\otimes 60}\rangle and 900 parameters, this circuit operates in a barren-plateau regime where gradient signals are weak and high-dimensional coordinate-mixing variance severely degrades pure SPSA. At an equal budget (76% of PSR), SPSA-680 fails to escape the constant predictor (test MAE 0.3570) in 2 out of 3 seeds (mean MAE 0.2704), whereas SGiQ (480, 200) successfully trains in all 3 seeds (mean MAE 0.1588). Full PSR remains best here (0.0989), which aligns with our theoretical analysis (Proposition 5): when the gradient signal is weak, replacing the dominant coordinates with exact PSR derivatives is critical for maintaining a reliable descent direction.

AuthorsOpenReview ↗

Response to W3.

W3: Adaptability to QGANs and QCNNs.

Our response: We thank the reviewer for the suggestion. To evaluate SGiQ on a widely adopted quantum neural network architecture, we added a QCNN classification experiment (cluster-state excitation classification, 8 qubits, convolution–pooling 8→4→2→18 \to 4 \to 2 \to 1, 320 samples, 3 seeds), in which every gate has an independent parameter (273 parameters).

MethodBudgetTest MAETest MSEAccuracy
PSR3.28M0.3649 ± 0.00630.22890.992
SPSA-1601.92M0.3792 ± 0.01300.23160.988
SGiQ (120, 14)1.61M0.3747 ± 0.00460.23280.992
SPSA-1201.44M0.3809 ± 0.01550.23880.992

On this QCNN benchmark, SGiQ (120, 14) achieves higher accuracy and lower error than SPSA-160 despite using a smaller budget, and approaches PSR performance at roughly half of its circuit cost.

An additional advantage: robustness to the SPSA perturbation size cc. While extending SGiQ to new models, we identified a practical benefit that does not depend on gradient sparsity. Since the SPSA estimate depends on cc through its finite-difference bias while the kk PSR-corrected coordinates are exact and independent of cc, SGiQ maintains a reliable descent direction even when cc is not finely tuned, whereas SPSA can fail completely. On the QCNN at c=0.3c = 0.3, every tested SGiQ configuration reaches 0.42–0.51 test MAE, while every SPSA configuration remains at 0.76–0.77 even with a larger budget. The same behavior holds on the 30- and 60-qubit regression circuits from W2 (3 seeds, using the same configuration across all values of cc):

Table R5: Effect of the SPSA perturbation size cc.

Setting (budget)ccSGiQSPSA
QCNN, 273 parameters0.2SGiQ (40, 55), 1.14M: 0.3907 ± 0.0157SPSA-160, 1.92M: 0.5376 ± 0.1371
0.3SGiQ (40, 55), 1.14M: 0.4160 ± 0.0510SPSA-160, 1.92M: 0.7747 ± 0.1535
Figure 2 circuit, 30 qubits (3.6M)0.1 (default)SGiQ (240, 60): 0.0799 ± 0.0041SPSA-300: 0.0863 ± 0.0136
0.30.1378 ± 0.01070.3570 (stalls in 3/3 seeds)
0.50.1398 ± 0.00610.3570 (stalls in 3/3 seeds)
Figure 2 circuit, 60 qubits, U[0,π/32]U[0,\pi/32] init. (7.2M)0.1 (default)SGiQ (320, 280): 0.1363 ± 0.0122SPSA-600: 0.1431 ± 0.0174
0.30.1938 ± 0.01640.3572 (stalls in 3/3 seeds)

Here, 0.3570 is the test MAE of a constant prediction (i.e., training fails to progress). This robustness also holds under realistic device noise (FakeBrisbane, c=0.5c = 0.5), where SGiQ outperforms SPSA across all 5 matched budgets (see W4, Table R7).

AuthorsOpenReview ↗

Response to W4.

W4: Noise experiments.

Our response: We thank the reviewer for this suggestion. In addition to the finite-shot (1024-shot) experiments in the main text (Table 2 and Appendix C), we added hardware-calibrated device-noise experiments. The 5-qubit, 8-layer circuit of Figure 2 (d=80d = 80 parameters, global readout ⟨Z⊗5⟩\langle Z^{\otimes 5}\rangle) is transpiled to the native gate sets and qubit connectivity of two IBM backends, which represent moderate gate noise (FakeMarrakesh, two-qubit error 0.26%) and high gate noise (FakeBrisbane, two-qubit error 1.51%), and is simulated with their calibrated noise models at 1024 shots per circuit evaluation. All runs use 3 seeds with identical initialization and mini-batch order across methods, and all SGiQ configurations satisfy k≤d/3k \le d/3.

Table R6: FakeMarrakesh (two-qubit error 0.26%), matched budget.

MethodBudgetBudget / PSRTest MAE
PSR960K1.000.1195 ± 0.0110
SPSA-60720K0.750.1505 ± 0.0170
SGiQ (50, 10)720K0.750.1340 ± 0.0089
SGiQ (40, 20)720K0.750.1362 ± 0.0076

At the same circuit budget (75% of PSR), both SGiQ configurations achieve lower test error than SPSA-60—closing over half of the accuracy gap to full PSR, while reducing the seed-to-seed standard deviation by roughly half.

Table R7: FakeBrisbane (two-qubit error 1.51%), SPSA perturbation size c=0.5c = 0.5, matched budget.

BudgetSGiQ (Test MAE)SPSA (Test MAE)Seeds where SGiQ is better
480KSGiQ (20, 20): 0.3241 ± 0.0255SPSA-40: 0.3524 ± 0.00973 / 3
540KSGiQ (40, 5): 0.3319 ± 0.0190SPSA-45: 0.3601 ± 0.00563 / 3
600KSGiQ (40, 10): 0.3485 ± 0.0081SPSA-50: 0.3582 ± 0.00263 / 3
720KSGiQ (40, 20): 0.3318 ± 0.0206SPSA-60: 0.3504 ± 0.01333 / 3
780KSGiQ (45, 20): 0.3329 ± 0.0247SPSA-65: 0.3395 ± 0.02022 / 3

Under combined shot and high gate noise, a larger perturbation radius cc is commonly used to amplify the finite-difference signal relative to measurement noise (cf. Eq. 7, σS2∝1/(Nshc2)\sigma_{\text{S}}^2 \propto 1/(N_{\text{sh}}c^2)). With c=0.5c = 0.5 on the higher-noise FakeBrisbane backend, SPSA barely moves from the constant predictor (0.3570) across budgets, whereas SGiQ continues to make training progress and outperforms SPSA in 14 out of 15 seed-matched runs. Since the PSR-corrected coordinates are unaffected by cc, SGiQ preserves a reliable descent direction, consistent with the cc-robustness observed in the noiseless setting (Table R5).

Comments on the paper

1 comment

We thank all reviewers for their time and constructive feedback. We are encouraged that the reviewers found the direction interesting and not yet fully explored (n1E4), the estimator intuitive and the paper easy to follow (tHe4, n1E4, u9pA), and the theory sound and more substantial than is typical for heuristic SPSA/PSR combinations (n1E4, u9pA).

The main concerns fall into two categories. We summarize our responses below and give details in the individual replies.

1. Scope of the experiments (tHe4, n1E4, u9pA). We have extended the experiments in three directions:

  1. Scale. Circuits with up to 60 qubits, including 30- and 60-qubit circuits with global readouts ⟨Z⊗n⟩\langle Z^{\otimes n}\rangle and a controlled width–depth study at a fixed parameter count d=240d = 240.
  2. Model and task. An 8-qubit QCNN classifier, a standard quantum machine learning architecture not designed by us.
  3. Noise. Experiments with hardware-calibrated noise models of two IBM backends.

In all settings, SGiQ attains lower mean test error than SPSA at the same or a smaller budget while using fewer gradient circuits than PSR.

2. Clarity of the claims and their scope (tHe4, n1E4, u9pA). Several concerns arose from how we presented the scope of our results: whether SGiQ requires zero gradients, the role of the local-observable form and of Assumption 2, and when SGiQ is cost-effective. Our theory and main conclusions are unchanged. In the revision we make explicit: 1) that SGiQ applies to general circuits, and that gradient-energy concentration, which does not require zero gradients, is the regime in which its advantage is proved; 2) which results depend on observable locality and on Assumption 2, and which hold for any observable and selection rule; 3) the conditions under which SGiQ reduces cost relative to PSR and improves on equal-budget SPSA; and 4) the limitations of the method.

We have uploaded a revised manuscript incorporating these changes. We hope these responses address the reviewers' concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.