- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies how to design quantum expectation-value surrogate models around the reuse requirements of downstream tasks. CQS-R shares recurrent computation across circuit layers or drive periods, while CQS-P fits low-rank Fourier products using observable-dependent masks. Moment supervision supports reuse across different Hamiltonians through linear readouts, and the Fourier structure supports analytic Gaussian continuation during optimization.
Through TFIM energy prediction, FSPT trajectory prediction, and surrogate-assisted VQE, the authors analyze the contributions of dependency priors, product fitting, recurrent reuse, supervision, and search. The results support several task-dependent benefits, but the downstream significance of the additional representation gains, long-horizon fitting capability, and optimization value beyond simple feasible circuits remain limited.
Critical strengths
-
Relatively comprehensive controls for the sources of improvement. Correct, full, and blockwise sparsity-matched random masks, together with complete Fourier regression matched in dependencies and frequency coverage, help distinguish the effects of dependency information from those of the fitting procedure. The matched-mask comparisons favor the product model in moment MSE on all five independent datasets (Table 5).
-
Attention to downstream outcomes and interactions between components. All VQE endpoints are evaluated using exact energies, and methods share data, measurement records, and initializations. Appendix D crosses representation, supervision, and continuation, and reports the reduced representation gap after validation-based frequency selection. With continuation fixed, moment supervision lowers the CQS-P endpoint energy on the grid from to . The same moment model can also be reused across six couplings , outperforming NC refitted separately for each coupling in Table 25.
-
Candid discussion of important limitations. The authors distinguish recurrent update counts from measured runtime, feasible references from ground states, and prediction error from optimization performance. They also acknowledge the effects of regularization tuning, effective control dimensionality, and long-horizon fitting difficulties.
Critical weaknesses
-
The downstream significance of the additional representation gains remains insufficient and depends on baseline settings. The largest moment improvement comes from correct dependency masks. The full- and random-mask MSEs are approximately 0.19, close to the zero-prediction baselines I independently recomputed using the same seeds and test splits (approximately 0.186–0.188; Tables 5 and 29). Under matched masks, the product model improves moment MSE by 10.5%, but energy MAE by only about 0.84%; this comparison also changes factorization, regularization, and fitting simultaneously. In TFIM, tuning NC regularization substantially narrows the gap. The best CQS variant retains an approximately 10–26% MAE advantage, but at , both CQS-R and NC-tuned have near zero, while only CQS-P retains some predictive ability (approximately 0.24). NC still uses fixed regularization in FSPT and VQE, leaving unclear how much of the reported advantage would survive thorough tuning.
-
The efficiency advantage of shared recurrence has not extended to a stable long-horizon advantage. CQS-R reduces recorded training-stage and test-evaluation times by approximately 30% and 27%, respectively, but Query-GRU has lower error on four of the five datasets, and the two models use different period-conditioning encodings. Extending supervision to 24 periods approximately doubles the error on the same first-six-period targets at , with deterioration also present on the training set (Tables 16 and 17). Longer training only partially recovers performance, and the tested increase in network width does not improve mean error. On later periods, the improvement over a training-median predictor fitted separately for each period and observable is only approximately 6–7.5% (Table 18).
-
The practical VQE gains are limited, and the complete pipeline deteriorates substantially on the higher-dimensional chain. The shared-parameter grids have only 1–3 effective control dimensions. At , exact diagonalization under the paper’s settings gives a relative gap of only approximately 1.2–1.4% between the simple reference and the ground state. The average two-layer CQS-P+C improvements beyond this reference are only 0.009 and 0.006. On the chain, the two-layer ansatz contains the one-layer ansatz, yet all four pipelines deteriorate with depth: for example, CQS-P+C changes from to , while two-layer CQS-R reaches . I also note that applying on three disjoint nearest-neighbor bonds prepares and yields energy , better than all these mean endpoints. The Hamiltonian-reuse results in Table 25 also fail to reach this reference. The current results do not distinguish insufficient surrogate approximation from search difficulties.
-
Inconsistent reporting protocols affect the hardware comparison. Inspecting the supplementary code, I found that Figure 7 uses validation-tuned full-prefix ridge, whereas Table 4 uses the fixed version. Both use validation-tuned current-layer ridge, contrary to the fixed-regularization descriptions in the text and captions. The tuned full-prefix baseline in the figure outperforms CQS-R on some datasets and is close at the remaining sizes. Meanwhile, many features of the zero-padded baseline are constant during training and vary only during testing, introducing a depth-dependent distribution shift. I therefore do not find the current hardware experiments sufficient to establish an advantage for CQS-R.
-
Methodological novelty and evidence of practical scalability are limited. The main components have established precedents, with the contribution concentrated in task organization and empirical analysis. The current product implementation still performs dense computation, so masks do not reduce its computational cost; Pauli-branch mask construction for TFIM/VQE grows exponentially in the worst case. The current FSPT adapter with also omits necessary harmonics, while increasing it to does not substantially improve later periods. The clearest VQE gains are concentrated on low-dimensional, shallow grids, and broader practical scaling benefits remain unproven.
Questions for authors
-
Please reconcile the regularization-selection descriptions in Figure 7, Tables 4/13, and B.6, and fully report both fixed-regularization and validation-tuned results. Can you confirm the report text in the script stating that all six full-prefix selections reached the upper bound ? If so, can the search range be expanded? What scope of recurrent advantage does the tuned comparison support?
-
The supplementary code includes exact-energy optimization using the same procedure, but the paper does not systematically report these results. Using the same seeds 45–47, 20 initializations per seed, and Adam settings as the main table, I reproduced this diagnostic and obtained a mean exact-energy direct-search endpoint of approximately on the two-layer grid. This shows that finite-step search itself can fail. Can you systematically report direct exact-energy search, exact-energy search with the same continuation, and the single-candidate reference defined in C.3 that selects the training control with the lowest measured energy, distinguishing reliable attainment near the simple reference from improvement beyond it? Please also provide complete comparisons across independent seeds for the chain settings with already included in Table 25.
-
Why do all pipelines deteriorate with depth on the chain and fail to reach the explicitly constructible energy ? CQS-R encodes raw angles and clips optimization controls to . Are its positive-energy endpoints concentrated near the boundaries? How large are the differences between predicted and true energies at these endpoints?
-
Can CQS-R and Query-GRU be compared using the same period-conditioning encoding and actual runtime budget? Does the deterioration in training fit on the shared early-period targets primarily arise from optimization difficulties, interference between targets, or representation limitations?
-
Can constant-prediction baselines be added to Tables 5 and 29, and can label-precision or sample-size controls clarify when masks are beneficial? Do the FSPT and VQE conclusions hold after selecting NC regularization using validation data?
-
What classes of larger and deeper circuits are expected to yield practical resource benefits? Can you provide cost evidence covering mask construction, fitting, and inference, and explain how light-cone expansion, dense implementation, and harmonic requirements under repeated controls limit this expectation?
My main concern is whether task structure yields sufficiently meaningful downstream improvements after reasonable tuning and matching supervision and search procedures. I would consider raising my score if the authors demonstrate substantive gains over tuned baselines and simple references, explain the chain failures, and reconcile the hardware protocols.