- 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
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 gates commute with the subsequent CZ gates and 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 and every in Table 1 unchanged. When their cost is removed from full PSR, SGiQ uses 66.7% and 80% of the PSR budget for and 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 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 ( for Friedman1 and 40 for Friedman2).
Table D. Number of active parameters capturing a fraction of the gradient energy, with an i.i.d. Gaussian vector of the same dimension as a reference.
| Task () | |||
|---|---|---|---|
| Friedman1 (75) | |||
| Random reference (75) | 18.3 | 24.6 | 34.2 |
| Friedman2 (40) | |||
| Random reference (40) | 10.2 | 13.5 | 18.6 |
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 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 , 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 , 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.
Response to W3 & W4.
W3: Hyperparameter selection for and .
Our response: We use grid search for the hyperparameters and to study the sensitivity of hyperparameters. In practice, there is an empirical adaptive strategy for the online selection: . The performance is unless sensitive to the hyperparameters. We add one more ablation study shown in the following table.
Table R1. SGiQ configurations with versus SPSA at equal budget. Test MAE, mean ± s.d. over 3 seeds unless noted; budget = number of training gradient circuits.
| Setting | SGiQ | Budget | SGiQ | SPSA (equal budget) | PSR (budget) | ||
|---|---|---|---|---|---|---|---|
| 30 qubits, local readout, ideal | 450 | (153, 27) | 0.15 | 2.16M | 0.0842 ± 0.0149 | 0.0868 ± 0.0219 | 0.0713 (5.4M) |
| (144, 36) | 0.20 | 2.16M | 0.0830 ± 0.0034 | 0.0868 ± 0.0219 | |||
| (135, 45) | 0.25 | 2.16M | 0.0837 ± 0.0037 | 0.0868 ± 0.0219 | |||
| (126, 54) | 0.30 | 2.16M | 0.0819 ± 0.0060 | 0.0868 ± 0.0219 | |||
| (112, 68) | 0.38 | 2.16M | 0.0787 ± 0.0015 | 0.0868 ± 0.0219 | |||
| (255, 45) | 0.15 | 3.60M | 0.0720 ± 0.0059 | 0.0768 ± 0.0111 | |||
| 30 qubits, local readout, 1024 shots | 450 | (153, 27) | 0.15 | 2.16M | 0.0783 ± 0.0045 | 0.0843 ± 0.0069 | 0.0784 (5.4M) |
| 30 qubits, FakeMarrakesh noise, | 450 | (135, 45) | 0.25 | 2.16M | 0.0691 ± 0.0067 | 0.0710 ± 0.0041 | 0.0741 (5.4M, 1 seed) |
| 30 qubits, FakeMarrakesh noise, | 450 | (135, 45) | 0.25 | 2.16M | 0.0755 ± 0.0076 | 0.0826 ± 0.0082 | |
| 60 qubits, local readout, 1024 shots | 900 | (510, 90) | 0.15 | 7.20M | 0.0534 (1 seed) | 0.0622 ± 0.0039 | 0.0626 ± 0.0036 (10.8M) |
| (465, 135) | 0.23 | 7.20M | 0.0611 ± 0.0044 | 0.0622 ± 0.0039 | |||
| QCNN (unshared parameters), ideal | 273 | (95, 14) | 0.13 | 1.31M | 0.3871 | 0.4368 | 0.3649 ± 0.0052 (3.28M) |
| (68, 41), EMA | 0.38 | 1.31M | 0.3819 | 0.4368 | |||
| (142, 41) | 0.22 | 2.20M | 0.3760 | 0.3975 | |||
| (156, 27), EMA | 0.15 | 2.20M | 0.3771 | 0.3975 | |||
| (120, 14) | 0.10 | 1.61M | 0.3747 ± 0.0038 | 0.3792 ± 0.0106 (1.92M) | |||
| Figure 1 circuit, 60 qubits, 3 layers | 300 | (20, 10) | 0.33 | 0.35M | 0.1257 ± 0.0054 | 0.1375 ± 0.0065 (0.30M); 0.1264 ± 0.0060 (0.48M) | 0.1325 ± 0.0021 (3.6M) |
W4: Selection independence, top- 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 from the same 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 ( qubits) with plain SGD (learning rate selected on the validation set from ). Entries below report Test MAE (), with training gradient circuits in parentheses:
| Setting | PSR | SPSA | SGiQ |
|---|---|---|---|
| Ideal | () | SPSA-: () | SGiQ$(25,25)$: () |
| shots | () | SPSA-: () | SGiQ$(40,10)$: () |
In both settings, SGiQ achieves lower test error than equal-budget SPSA-. In the ideal setting, every evaluated SGiQ configuration with ( to ) outperforms SPSA-, confirming that the ordering holds under plain SGD as well as under Adam.
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).
| Task | 10th vs. 11th largest gradient | Precision (selected indices in true top-10) | Energy share of true top-10 | Energy share of selected 10 | Selected / true top-10 energy |
|---|---|---|---|---|---|
| Friedman1 | +3–5% | 0.32–0.41 | 0.52–0.55 | 0.27–0.34 | 51–61% |
| Friedman2 | +10–18% | 0.59–0.82 | 0.93–0.96 | 0.81–0.93 | 86–97% |
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 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- 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 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 paired seeds sharing identical parameter initialization, data split, and mini-batch order across methods, while our gradient-concentration and selection diagnostics use seeds. Across of the 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 and 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 shots (see W1). On the -qubit benchmark, SGiQ reduces the test error of equal-budget SPSA by with half the seed-to-seed standard deviation. On the -qubit benchmark, SGiQ achieves the lowest mean error among all three methods (outperforming both equal-budget SPSA and full PSR) while using only 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.
Response to W1.
Our response to W1: We thank the reviewer for this comment. We extended our experimental validation along each requested dimension:
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 30 and 60 qubits under ideal simulation. Each family keeps its readout from the paper: the global observable ⟨Z⊗n⟩ for the Figure 2 circuit and local single-qubit Z observables for the Figure 1 circuit. Entries report Test MAE (mean±SD), with the number of training gradient circuits in parentheses.
With the global observable, the 60-qubit Figure 2 circuit lies in a barren-plateau regime, where SPSA frequently stalls at the constant predictor (test MAE 0.3570).
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.
These results establish three conclusions:
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).