acceptodds
Under review as a conference paper at ICLR 2027

Compositional Quantum Surrogates: Circuit-Guided Representations for Quantum Surrogate Learning

Abstract

Quantum expectation surrogates amortize measurements across related circuit controls, but downstream tasks require different forms of reuse. Existing predictivesurrogate methods learn mean-value functions; how the fitted representation serves ordered predictions, observable reuse, and surrogate-based optimization remains open. We introduce Compositional Quantum Surrogates (CQS) through two circuit-guided architectures: CQS-R, a recurrent model for ordered layers or drive periods, and CQS-P, a masked Fourier product model for observable-dependent interactions. With matched masks and frequency coverage, product composition reduces moment prediction error by 10.5% relative to complete Fourier regression. A query-wise recurrent control reaches similar accuracy on 12-period trajectories but requires 78 updates, compared with 12 shared updates for CQS-R. In VQE, linear moment readouts enable Hamiltonian reuse, and Fourier structure enables analytic continuation; on two-layer grids, continuation mainly improves attainment near a feasible reference. The results favor dependency-aware products for moments, shared recurrence for trajectories, and continuation-stabilized search for VQE pre-training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum expectation prediction, Compositional representation learning, Quantum surrogate modeling, Variational quantum algorithms, Finite-measurement learning

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_JJuzOpenReview ↗
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
  1. 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).

  2. 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 3×33\times3 grid from −3.80-3.80 to −4.50-4.50. The same moment model can also be reused across six couplings JJ, outperforming NC refitted separately for each coupling in Table 25.

  3. 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
  1. 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 N=14N=14, both CQS-R and NC-tuned have R2R^2 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.

  2. 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 N=8,10N=8,10, 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).

  3. 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 J/h=0.2J/h=0.2, exact diagonalization under the paper’s settings gives a relative gap of only approximately 1.2–1.4% between the simple reference −N/2-N/2 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 −2.631-2.631 to −1.286-1.286, while two-layer CQS-R reaches +1.671+1.671. I also note that applying RZZ(π)R_{ZZ}(\pi) on three disjoint nearest-neighbor bonds prepares ∣−⟩⊗6|-\rangle^{\otimes6} and yields energy −3-3, 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.

  4. 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 α=1\alpha=1 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.

  5. 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 Qϕ=1Q_\phi=1 also omits necessary harmonics, while increasing it to Qϕ=8Q_\phi=8 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
  1. 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 α=100\alpha=100? If so, can the search range be expanded? What scope of recurrent advantage does the tuned comparison support?

  2. 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 −3.023-3.023 on the two-layer 3×33\times3 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 ∣J∣=0.5|J|=0.5 already included in Table 25.

  3. Why do all pipelines deteriorate with depth on the chain and fail to reach the explicitly constructible energy −3-3? CQS-R encodes raw angles and clips optimization controls to [−π,π][-\pi,\pi]. Are its positive-energy endpoints concentrated near the boundaries? How large are the differences between predicted and true energies at these endpoints?

  4. 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?

  5. 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?

  6. 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.

Reviewer_BLgjOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Compositional Quantum Surrogates (CQS), a framework for learning quantum expectation values using two architectures that exploit different structural properties of quantum circuits. CQS-R uses recurrent networks to share computation across sequential circuit layers or drive periods, while CQS-P uses observable-dependent masks and low-rank Fourier products to model interactions between circuit controls.

The authors evaluate both approaches on transverse-field Ising model prediction, Floquet trajectory prediction, and surrogate-assisted VQE. The results demonstrate improvements in prediction accuracy and computational reuse over existing surrogate methods. The paper also includes extensive ablations separating the contributions of circuit-dependent masking, Fourier factorization, moment supervision, and Gaussian continuation.

Critical strengths
  1. Well-motivated framework. The paper identifies different forms of structure that can be exploited in quantum surrogate learning. The distinction between sequential computation reuse and observable-dependent Fourier representations is sensible and potentially useful across quantum computing applications. This is a promising direction towards more scalable analysis, representation and design of quantum circuits.

  2. Encouraging empirical results. Both architectures achieve competitive results against existing methods. In particular, CQS-P improves moment prediction accuracy, while CQS-R provides computational reuse across trajectory queries without sacrificing predictive performance.

  3. Thorough experimental ablations. I particularly appreciate the effort to isolate the different sources of improvement. The comparisons with complete Fourier regression under matched masks, depth-matched recurrent models, and the factorial experiments separating representation, supervision, and continuation strengthen the empirical conclusions.

  4. Practical relevance. The ability to reuse observable predictions across Hamiltonian coefficients and the inclusion of experiments using real quantum hardware make the proposed framework potentially useful beyond the specific benchmarks studied.

Critical weaknesses
  1. Limited difficulty of the optimization benchmarks. The VQE experiments are relatively simple, particularly the two-dimensional systems, which have only one or three effective control parameters. Moreover, the strongest results largely recover an analytically accessible feasible configuration, with limited improvement beyond it. Consequently, these experiments demonstrate optimization reliability but provide limited evidence of effectiveness on challenging VQE problems.

  2. Limited methodological novelty. The individual components, including recurrent representations, low-rank Fourier products, and Gaussian continuation, are established techniques. The ablations also suggest that much of the improvement can be attributed to circuit-dependent masking and continuation, with smaller gains attributable specifically to the proposed architectures. While their combination is valuable, the fundamental algorithmic novelty is somewhat limited.

  3. Uncertain computational scalability. The construction of observable-dependent masks can require exponentially many intermediate Pauli strings, and the current implementation does not exploit mask sparsity to reduce dense computation. Similarly, the measured computational advantage of CQS-R is more modest than the reduction in recurrent update counts suggests. Further investigation is needed to establish whether these benefits persist for larger and deeper circuits.

I am willing to reconsider my score if some of these concerns are addressed convingcinly during the rebuttal.

Questions for authors
  1. How does the computational cost of constructing the CQS-P dependency masks scale in practice with circuit depth and system size? Does the increasing density of dependencies eventually limit their usefulness?

  2. Given that CQS-R and CQS-P exploit complementary forms of circuit structure, have the authors considered combining both architectures into a single model? For instance, could observable-dependent Fourier representations be incorporated into sequential recurrent updates to exploit both spatial dependencies and temporal composition?

Reviewer_2fUxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Compositional Quantum Surrogates (CQS), a framework for learning expectation-value predictors for parameterized quantum circuits. The authors propose the following surrogate architectures: 1) CQS-R, a GRU-based recurrent model to predict quantum observables across drive periods sequentially; and 2) CQS-P, a low-rank Fourier product model using circuit-dependency masks to represent the interactions among control parameters. The methods are then evaluated on TFIM energy prediction, FSPT trajectory prediction, and VQE optimization. The main findings are that: 1) once dependency masks are fixed, CQS-P achieves a 10.5% reduction in moment MSE compared with a complete Fourier regression; 2) CQS-R essentially matches query-wise GRU in accuracy with fewer recurrent updates; and 3) CQS-P + C pipeline yields the best endpoint after VQE optimization.

Critical strengths
  1. There are thorough ablations regarding dependency masks, representations, and optimization methods.
  2. The observable-dependent masks improve prediction accuracy by a lot. In addition, CQS-R demonstrates computational reuse across trajectory predictions.
Critical weaknesses
  1. Novelty could be limited, as CQS-R is essentially based on a standard GRU, while CQS-P adapts low-rank Fourier factorization techniques which already exist.
  2. Most of CQS-P's improvement comes from dependency masks rather than product factorization itself. Furthermore, the actual wall-clock speedup is considerably smaller than the reduction in recurrent updates.It would make more sense to compare it with other models, like attention, which would be cheaper than either and are not compared. Furthermore, there might be issues related to scalability and more baselines are needed. The grid experiments carry the paper’s strongest VQE claims but they have only 1-3 effective controls. The endpoints are also close to a simple analytic reference (−N/2). For the 1D chain there lacks a similarly strong complete-Fourier baseline and a reference energy for evaluating the quality of the optimization endpoints.
Questions for authors
  1. How does CQS-R compare with parallel trajectory-prediction architectures, such as a Transformer or a MLP (multiple outputs), under fixed inference constraint?
  2. On the 1D chain, does CQS-P+C still win against a complete Fourier baseline with continuation? Can the authors provide a reference for the best attainable energy within the variational ansatz?
  3. Can the authors demonstrate VQE improvements on a higher-dimensional task where the optimized energies fall meaningfully below a simple feasible reference?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.