acceptodds
Desk rejected at ICLR 2027

Fourier-Geometric Circuit Design for Gate and Entanglement Placement in Quantum Neural Networks

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

The output of a parameterized quantum circuit (PQC) can be expressed as a finite Fourier series whose accessible frequencies are fixed by the data-encoding gates. While the encoder determines which frequencies can appear, the corresponding Fourier coefficients depend on how the trainable and entangling gates are arranged. Although existing studies provide metrics for characterizing how gate structure affects Fourier coefficients, they do not translate these analyses into an explicit design criterion specifying how gates should be arranged to make a target coefficient reachable. In this paper, we provide such a criterion. Using the adjoint action of the encoding generator, we decompose operator space into two-dimensional invariant planes indexed by frequency and show that each Fourier coefficient is exactly a sum of bilinear projections of the effective state and observable onto the planes at that frequency. Because, for Pauli encodings, high-frequency planes are spanned by mixed multi-qubit Pauli strings, a target coefficient can contribute to the output only when local rotations and entangling layers are arranged so that both the effective state and observable acquire support on one of its planes. For Pauli readouts and commuting two-qubit entanglers, this yields a circuit design rule that specifies, for a given interaction graph, a placement of local rotations and entangling layers that makes a target Fourier coefficient reachable. Using a single encoding layer, we validate the proposed design through a placement ablation, regression tasks from PDEBench and physics-informed Maxwell field modeling, and further demonstrate its robustness to moderate simulated gate noise and damping.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum Circuit Design, Parameterized Quantum Circuits, Entanglement, Quantum Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_wneeOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper use Fourier analysis to direct the circuit design of PQC, especially the arrangement of single-qubit rotation and entangling gates. While data encoding determines the possible Fourier frequencies of the circuit, the paper prove that the Fourier coefficients is determined by the projection of effective quantum state and observable. Based on this theory, the author proposed a circuit architecture composed of local rotations and entangling gates, as well as a circuit construction algorithm based on qubit interaction graph. In experiments, authors verified their circuit deigns on Fourier function fitting, advection equation, Burgers equation and Physics-informed wave equation. Results showed their circuit can use less two-qubit gates to get lower prediction error. Placement ablation study shows that even the number of sage stays the same, the arrangement of gate would affect reachable Fourier frequencies.

Critical strengths
  1. Good combination of theory and algorithm design. The paper provided the expression of Fourier coefficients, derived gate placement condition, and proved the correlation of reachable frequencies and measured qubit degree. The paper establishes the connection between Fourier coefficients and circuit structure, and further derived gate placement rules.
  2. Comprehensive experiments. The experiments include Fourier function approximation, supervised PDE regression and physics-informed learning, while also comrade with existing PQC architectures. The placement ablation study validated the theoretical Fourier reachability.
  3. Considered the constraints of real quantum hardware and noise robustness. Though still based on classical simulation, the consideration of qubit connectivity, two-qubit gate costs and depolarizing noise, damping and finite-shot noise contributed to the completeness of the work.
Critical weaknesses
  1. The theoretical analysis focuses on the reachability of individual Fourier coefficients. However, this does not guarantee that multiple coefficients can be simultaneously adjusted to match a target function. Therefore, the connection between Fourier reachability and practical approximation capability remains unclear. (see question 2)
Questions for authors
  1. Data re-uploading can also increase Fourier bandwidth. In the meantime, it may also activate Fourier coefficients that are structurally zero in a single-upload circuit. Could the author clarify whether increasing the number of data re-uploading layers can achieve similar effect as the entanglement placement?
  2. In Table 3, the proposed architecture has a best possible MSE of 0.013, but the trained MSE is 0.628. Though the reduced training budget might be able to explain part of the difference, this also lead to a question: Is the performance of the circuit constrained by optimization, or is it constricted by the dependence of Fourier coefficients? Is it possible the further distinguish the two factors by better training or by analysis of learned Fourier coefficients?
2 comments
AuthorsOpenReview ↗

Response Letter for Reviewer wnee

We thank the reviewer for these insightful questions. We conducted additional experiments on re-uploading, gate placement, optimization, and coefficient controllability. Due to the response limit, we summarize our evidence below.

Q1. We agree that data re-uploading can expand the Fourier bandwidth and activate previously inaccessible coefficients. We repeated each 14-qubit circuit block twice and identified reachable Fourier pairs numerically, without training.

ArchitectureOne uploadTwo uploadsMax. frequency
Proposed225841(14,14)
No Align122841(14,14)
Leaf Readout9289(8,8)
No Expose3255(8,7)

As the reviewer anticipated, No Align reaches all 841 pairs after two uploads, demonstrating that extra encoding layers can recover Fourier components absent from the single-upload circuit. The proposed placement also reaches all 841 pairs, while the other placements produce different accessible supports at the same depth. This clarifies their complementary roles: re-uploading increases bandwidth through encoding depth, whereas gate placement determines which frequencies become accessible at each depth.

We further tested four-layer re-uploading (Appendix C.2). Proposed (γ\gamma) achieves a held-out MSE of 0.104, compared with 0.169 for the closest encoding-matched baseline, using only 48 two-qubit gates versus 96–329 for the entangling templates. This supports our placement strategy for accurate, gate-efficient learning with re-uploading. We will incorporate the new numerical results and distinguish the existing single-upload analytical guarantees from the multi-upload experiments.

Q2. We appreciate the reviewer's question about the difference between the Fourier-support approximation bound of 0.013 and the trained MSE of 0.628 in Table 3. To investigate optimization and coefficient dependence, we conducted: longer training, coefficient-Jacobian analysis, and target-specific approximation experiments.

First, the original Table 3 experiment used both a smaller training subset and a reduced optimization budget. We compared it with the full Table 2 protocol and a doubled step budget:

Training settingStepsSeedsMSE ↓
Reduced (Table 3)≈2,57350.628±0.212
Full (Table 2)15,43880.103±0.043
Doubled budget30,87630.084

The marked improvement with the full training protocol demonstrates that training data coverage and optimization budget account for a substantial part of the original gap. The doubled budget further improves accuracy.

Second, we computed the Jacobian of the Fourier-coefficient vector with respect to circuit parameters. The fixed-CZ Proposed architecture reaches all 225 Fourier frequency pairs, with a coefficient-Jacobian rank of 58; allowing trainable entangling strengths increases this rank to 84. The Jacobian rank quantifies locally independently adjustable coefficient directions, rather than the number of reachable frequencies. Thus, the additional entangling parameters enlarge the set of coefficient combinations that can be tuned jointly. The Fourier-support LS bound permits independent choices of coefficients across all reachable frequencies; the trained PQC instead uses a structured coefficient parameterization.

To connect this observation to function approximation, we tested targets with different Fourier coefficient matrix ranks:

Target coefficient rankFixed-CZ successful fits
1–373/77
≥42/83

The 73/77 successful low-rank fits demonstrate strong approximation capability for structured coefficient patterns. These results motivate trainable entangling strengths. On an additional Advection instance, the fixed-CZ circuit attains 0.365 MSE, close to the best rank-3 Fourier approximation (0.33). A rank-5 Fourier approximation reaches 0.020, and Proposed (γ\gamma) reduces the trained MSE to 0.033. This shows how additional coefficient flexibility can translate into a substantial accuracy improvement.

Finally, we assessed optimization on targets generated independently from the same circuit family. The fixed-CZ model successfully recovers 19/20 realizable 8-qubit targets, providing evidence that the optimizer can learn functions represented by its architecture. This test directly examines optimization on targets known to be realizable by the circuit family.

Taken together, these experiments provide a clearer explanation of the original error gap: a fuller training protocol substantially improves accuracy, while Fourier-coefficient structure explains why trainable entangling strengths further improve approximation on demanding targets. We will clarify these distinct mechanisms in the revision.

Thank you for the detailed response and additional experiments. The new results help clarify the roles of data re-uploading, gate placement, optimization, and Fourier coefficient controllability, and address my main concerns. Given the substantial improvement from the full training protocol, I encourage the authors to update Table 3 using a sufficient training budget across the compared architectures, and incorporate the additional analysis into the revised paper. I will maintain my original rating.

Reviewer_UND2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propse a principled way of a designing a circuit based on Fourier geometry. While the data encoding defines the accessible frequencies, the coefficients are determined by the circuit, however, to-date there is limited understanding for this. They analyze different circuits and design them, s.t., they support Fourier coefficents up to a certain degree and provide experimental results.

Critical strengths

The task of characterizing PQC architectures and understanding how they impact the final model is interesting. The authors have a solid theoretical foundation, based on which they conduct additional experiments. I thus think the paper is instructive to read.

Critical weaknesses

While the results are valid, I am uncertain about the impact of the work. The last years have shown that the primary problem in PQCs is trainability and the way out for that is to insert inductive biases (while at the same time not making the models classically simulable).

I think this is exactly the related work that I am missing. The Barren Plateau aspect is heavily studied, and the perspective that particularly crystallized out is that it is undesireable to have a model be maximally expressive, as they essentially end up behaving randomly. This position is also relevant in the end of section 2, when the authors mention a study that found strongly entangled layers suboptimal for their model. This is a phenomenon that is extensively characterized theoretically and there are even works on entanglement-induced Barren Plateaus.

For these reasons I am uncertain about the motivation for this work, since (1) it is generally undesirable to have a model be expressive in that sense (see BP literature) and (2), in the end of section 3 in the paragraph "Design procedure" it is stated that we are given a largest frequency ω∗\omega^* that a task needs. In any general machine learning scenario we do not know the Fourier spectra of our underlying ML task. If the authors are aware of such scenarios, I would greatly appreciate any input and references.

Even taking that aside, it is well known that low-degree matching enables classical simulability, whereas including higher order moments (and keeping in mind that setting a largest frequency also includes all lower-order frequencies) leads to BPs. It is for these reasons and for the absence of the discussion on BP, that I am not sure about the moviation for this work. Also, I think answering the question on classical simulability is clearly out of scope for this work, however, I think it would merit a discussion and a paragraph in the related work section.

Moreover, there is a limited discussion on the interplay between the limited entanglement of the approach compared to their baseline. It is well known that entanglement makes the circuits harder to train and the authors' approach employ significantly fewer entanglement gates than their baseline, which is why I am uncertain whether the results imply that the improvements are indeed only attributable to the approach. Further, below are some questions about the experimental set-up. Generally, I think the current experimental setup appendix is not very clear, and I really do not understand why the setups differ so much for the different architectures and experiments. The number of repetitions are different for the different experiments, the number of iterations range from 6 epochs to 300 steps. I would appreciate a more unified set-up, which would make the paper a lot easier to follow and the claims could be verified properly.

I would also encourage the authors to rethink the structure of the paper. The paper now has a very extensive appendix, that is packed with further results, further explanations and the experimental set-up but there is no reading flow due to very few links between the main part and the appendix. It is simply not possible without extensive reading to pinpoint which parts of the appendix are relevant for which parts of the main paper, which is why I felt that large parts of the appendix were also irrelevant. I found the paper incredibly hard to follow due to that reason and I think a more thoughtful appendix could greatly improve it.

Questions for authors
  • Could the authors maybe clarify the notation? Usually, an nn qubit state would be denoted as ∣ψ⟩n\ket{\psi}_n or similar but the tensor product notation in section 3 makes it seem like they are using a single qubit state ∣ψ⟩\ket{\psi} that is tensorized nn times, which I assume is not what was intended.
  • Could the authors provide some information on how their approach can be extended to more general circuit architectures?
  • Could the authors clarify the naming convention for the baseline circuits? Why are they called circuit 16-19?
  • Why are the circuits for supervised PDE regression only trained for 6 epochs? I would expect something on the order of at least 10-50. Also the physics-informed wave equation is trained for 300 steps, so the question arises why they differ so much.
  • Why are the re-uploading circuits trained over 8 seeds, but the other one is trained only over 3 seeds? And why does the noise study only retain 5 out of the 8 seeds?
  • Have you considered that the improved results from your method could come from better trainability due to less entanglement? I think that there should be significant overlaps between your method (that limits entanglement) and the baselines (that often are highly entangled) acc to e.g., Table 2. I am uncertain that without an analysis and discussion on these these interplays the experimental results are really just attributable to the approach.
3 comments
AuthorsOpenReview ↗

Response Letter to Reviewer UND2

We thank the reviewer for highlighting barren-plateau and classical-simulability literature. Full change is reflected in the revised paper. We clarify that expressivity we refer to is Fourier expressivity, i.e., the set of Fourier frequencies structurally reachable by a circuit architecture. This differs from circuit expressibility, that describes how broadly a PQC explores the quantum state or unitary space. Our objective is not to maximize circuit expressibility, but to characterize how gate placement enables targeted Fourier reachability.

1. We study gate-placement spectral inductive bias, not maximal expressibility. To assess concentration and trainability, we measured random-parameter variance:

MethodVar. (N=14N=14)Per added qubit
HEA Circular6.7×10⁻⁵÷1.96
Strongly Entangling5.9×10⁻⁵÷1.97
Circuit 170.14Constant
Circuit 190.067Constant
Proposed0.10Constant

Proposed maintains constant variance up to 40 qubits, unlike exponentially concentrating baselines. Yet Circuit 19 also avoids concentration and reaches the same frequencies but performs worse on Advection. Thus reachability and avoiding concentration do not ensure learning success.

2. ω∗\omega^* is task-dependent, not assumed known. Wiedmann et al. [1] estimate data frequencies to select VQCs; we complement this by designing placements to reach them. Fourier least-squares (LS) bounds versus held-out Advection MSE:

SampleQubits(kx,kt)(k_x,k_t)Fourier LSMSE ↓
014(7,7)0.0130.103±0.043
012(7,5)0.0210.135±0.040
8014(7,7)0.0300.044±0.008
808(4,4)0.0510.078±0.004

Comparable 12-/14-qubit errors (p=0.21p=0.21) support bandwidth-guided selection, not guaranteed optimization.

3. We derived an exact classical evaluator: two CZ/controlled-phase stars yield four product-state terms and at most 16 overlap products, costing O(N)O(N). It agrees with statevector simulation to 1.3×10⁻¹⁴, taking 3.6 μs/input (N=14N=14) and 0.39 ms/input (N=1000N=1000) on CPU. We do not claim quantum advantage. For kk stars, cost is O(4kN)O(4^kN) (fixed kk).

Q1. The initial state is a general product state ∣ψin⟩=⨂q=1N∣ψq⟩|\psi_{in}\rangle=\bigotimes_{q=1}^N|\psi_q\rangle, not identical copies. Experiments use ∣0⟩⊗N|0\rangle^{\otimes N}; we will correct this.

Q2. We tested other entanglers and graphs beyond CZ stars. Full means the full target Fourier box:

Entangling gateReachability
CZFull
Controlled-phase (π/2\pi/2)Full
CNOT (both orientations)Full
XX(π/2)XX(\pi/2)Full
iSWAP\sqrt{\mathrm{iSWAP}}Full
iSWAP9 pairs

For ring, path, grid, heavy-hex, and 3-regular graphs:

Entangling roundsFrequency boundSuccessful cases
Singledeg⁡(r)+1\deg(r)+110/10
MultipleRadius-LL light cone28/30

These results support wider applicability; two grid cases cancel. Multi-round findings remain numerical.

Q3. Circuits 16–19 follow Sim et al. [2]; we will describe them in App. A.

Q4. We report training budgets in steps:

ExperimentQubitsEpochsSteps
PDEBench (Table 2)14615,438
Re-uploading730≈450
Wave Q-PINN2—300

Wave Q-PINN resamples points each step; epochs do not apply. We will add cross-references.

Q5. We expanded PDEBench to 7–8 seeds and noise tests to 8, with unchanged rankings:

MethodAdvection (seeds)Burgers (seeds)
HEA Circular0.230±0.135 (7)0.746±0.273 (7)
Strongly Entangling0.274±0.072 (7)0.933±0.037 (7)
Proposed0.103±0.043 (8)0.200±0.011 (7)
Proposed (γ\gamma)0.052±0.016 (7)0.189±0.014 (7)

Q6. To assess entanglement, we measured spatial–temporal entropy SxtS_{xt} alongside Fourier LS bounds and held-out Advection MSE:

Method2Q gatesFourier LSSxtS_{xt}MSE ↓
HEA Circular560.0135.910.230
Strongly Entangling560.0135.970.274
Circuit 19560.0133.540.900
Circuit 16520.9991.511.036
Proposed260.0131.480.108
Proposed (γ\gamma)260.0131.400.052

Our circuits have lower entropy and error, but entanglement alone cannot explain the gain: Circuit 19 has lower entropy yet worse accuracy at matched gate count and LS bound; Circuit 16 has low entropy but limited reachability. We emphasize placement, reachability, coefficient control, and trainability.

References

[1] M. Wiedmann et al., “Fourier Analysis of Variational Quantum Circuits for Supervised Learning,” IEEE QCE (2025).

[2] S. Sim et al., “Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum-Classical Algorithms,” Adv. Quantum Technol. 2, 1900070 (2019).

I would like to thank the authors for the extensive clarification. This made the experimental setup much clearer. I, however, cannot access the updated paper version (there are no revisions visible for me).

Besides the clarification, I still think that the scope is quite narrow, with only little practical relevance. In particular, with the models being fully classically simulable, I do not see a sufficiently substantial contribution to the scientific community to justify publication at this venue at this point.

AuthorsOpenReview ↗

Response to the Comment

We thank the reviewer for the follow-up. The revised PDF has been uploaded.

On classical simulability. We agree that our single-round circuits admit O(N)O(N) classical simulation (App. E.6). This does not, however, make the circuit-design question trivial. Known approaches to avoiding barren plateaus often also admit efficient classical simulation (Cerezo et al., 2025). The question we address is how to retain the Fourier frequencies needed for a task without losing the trainability benefits of structured circuits.

On Fourier support, coefficient magnitude, and trainability. Much work on barren plateaus seeks more favorable optimization through shallow or local circuits, or restricted dynamical Lie algebras (Cerezo et al., 2021; Ragone et al., 2024; Larocca et al., 2025). However, improving trainability does not establish that a circuit can represent the target function. Prior results identify tensions between expressibility and gradient magnitude (Holmes et al., 2022) and show that some Fourier coefficients become suppressed as circuits grow (Mhiri et al., 2025). Therefore, a circuit must reach the relevant frequencies, support usable coefficient magnitudes, and remain trainable. The circuits of Table 2 separate these requirements (pairs and bounds from Table 18, variance from Table 23):

CircuitsReachable pairs (of 225)Best possible MSEOutput variance as NN growsHeld-out MSE
Circuits 16, 18, YZY15–35≥ 0.99–1.04–1.06
Circuit 17250.997flat1.034
Basic Ent.225, real only0.81–1.02
HEA / Strongly Ent.2250.013halves per qubit0.230 / 0.274
Circuit 192250.013flat0.900
Proposed2250.013flat up to N=40N=400.103

Circuits with restricted Fourier support have high approximation-error lower bounds regardless of training; Circuit 17 shows that stable output variance alone is insufficient. HEA and Strongly Entangling circuits reach all 225 pairs but exhibit exponentially decreasing output variance. Circuit 19 also reaches all pairs without output concentration, yet it stays at the constant predictor on six of seven advection seeds (Sec. 4). Being able to represent a frequency does not mean the optimization can effectively use it.

What is new in our design. Rather than analyze the spectrum of a fixed template, as in Wiedmann et al. (2025), we use operator geometry to derive which gate placements enable the required Fourier frequencies. We also establish coefficient-magnitude bounds (Sec. 3, App. D.6) and examine output concentration. To our knowledge, prior work has not used such gate-placement conditions to systematically construct circuits with task-matched Fourier support. Support and magnitude bounds are available before training, allowing inadequate designs to be screened without optimization or shots. In Table 2, our circuit achieves the lowest held-out MSE using 26 two-qubit gates, versus 56 for the expressive baselines.

Beyond single-round circuits. The invariant-subspace decomposition does not depend on the number of rounds, although our guarantees are proved for one. The single-round O(N)O(N) simulation guarantee does not extend automatically: the stated simulation cost scales as 4k4^k for kk stars. In our seven-qubit, four-layer re-uploading experiment, the proposed circuit achieves the lowest error with 48 two-qubit gates, versus 96–329 for the baselines (App. C.2).

Practical evidence. Under finite-shot training, the proposed circuit learns at all tested budgets from 10210^2 to 10410^4 shots, while HEA and Strongly Entangling circuits stay near the constant predictor through 10310^3 shots (Table 4). The placement rule also extends to ring, grid, and heavy-hex connectivity (App. E.3) and to other two-qubit gates (App. E.5).

In conclusion, our contribution is a circuit-design principle that jointly addresses Fourier expressivity and trainability. Rather than prioritizing trainability at the expense of representation, we theoretically derive how gate placement determines reachable Fourier frequencies and coefficient magnitudes, while experimentally demonstrating that the resulting circuits maintain favorable trainability. A trainable circuit is of limited use if it cannot represent the target, just as an expressive circuit is of limited use if it cannot be trained. Since, to our knowledge, no prior work has built circuits whose Fourier support matches the task while keeping them trainable, we believe this is a substantial contribution to the community.

References. Cerezo et al., Nat. Commun. 12, 1791 (2021); Cerezo et al., Nat. Commun. 16, 7907 (2025); Holmes et al., PRX Quantum 3, 010313 (2022); Larocca et al., Nat. Rev. Phys. 7, 174 (2025); Mhiri et al., Quantum 9, 1847 (2025); Ragone et al., Nat. Commun. 15, 7172 (2024); Wiedmann et al., IEEE QCE 2025, 1785.

Reviewer_Zs3zOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies placement of gates and entanglement in quantum circuits to produce specific Fourier frequencies. The authors provide principled ways to do this, which has never been done in the literature before, and show experimental results to validate their theory.

Critical strengths

I do not have any background in parametrized quantum circuits, but if what the authors claim is true, the problem they are solving is important and novel.

Critical weaknesses

I strongly believe that this paper does not belong to ICLR. I do not see any of the topic area this paper would belong to. It might be better to target a quantum venue.

Questions for authors

N/A.

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

The paper develops a geometric description of Fourier coefficients in parameterized quantum circuits through projections of the effective state and observable onto invariant operator subspaces. Under specified encoding and entangler assumptions, it derives a gate-placement rule that makes target frequencies reachable. The approach is evaluated through circuit ablations, PDE regression, physics-informed wave modeling, and simulated noise studies.

Critical strengths
  1. The operator-space formulation provides an interpretable connection between gate placement and Fourier support. The sufficient reachability condition goes beyond identifying the frequencies permitted by the encoder and explains why rearranging the same gates can change the represented function class.
  2. The placement ablation preserves gate and parameter counts while testing predicted changes in spectral support. Together with the parameter-matched comparisons and additional re-uploading experiments, this provides meaningful evidence for the proposed structural mechanism.
Critical weaknesses
  1. I am not yet convinced that frequency reachability provides a sufficiently predictive design criterion for learning. Corollary 1 establishes generic nonzero coefficients, while Appendix D.6 realizes individual harmonics separately. Neither result characterizes whether the amplitudes and phases required by a target can be achieved simultaneously or learned reliably. The authors acknowledge this distinction, but the placement ablations mainly compare against variants that lose spectral components. They therefore leave open whether the proposed placement offers a learning benefit when competing circuits already support the same frequencies.
  2. The empirical evidence is too narrow to resolve this concern. The main PDEBench comparison uses one advection and one Burgers solution sharing the same initial condition, and reports error over the full grid despite training on 80% of it (Appendix B.1). This mainly demonstrates reconstruction of selected fields. I appreciate the held-out re-uploading study in Appendix C.2, but it covers only one advection example. Moreover, the reduced-budget placement ablation gives a Holm-corrected p-value of 0.082 (Appendix B.3), making its learning-performance evidence less conclusive than its agreement with the predicted spectral support.
Questions for authors
  1. Does the advantage persist against circuits with the same reachable Fourier spectrum and comparable gate, parameter, and optimization budgets? This would help establish whether the proposed placement improves learning once the required frequencies are already accessible.
  2. Can you report held-out-only errors for Table 2 and clarify whether the improvements persist across independent PDE initial conditions? Appendix B.1 currently reports full-grid errors for two solutions sharing the same initial condition, which limits the evidence for broader applicability.
  3. For targets containing several frequencies, what constraints does the proposed architecture impose on the jointly attainable coefficient amplitudes and phases? Corollary 1 establishes generic nonzero coefficients, while the unit-amplitude construction treats individual harmonics separately, leaving the connection to simultaneous target-spectrum approximation unclear.
  4. Algorithm 1 uses a state-side block of the form (W_R=RCR), whereas Appendix B.3 specifies (W_R=CR) for the evaluated circuit. Does the unit-amplitude guarantee in Appendix D.6 still hold for this reduced architecture, and which theoretical guarantees apply directly to the reported experiments?
  5. How sensitive are the reported improvements to learning rate and training budget? In particular, how do the effect sizes and uncertainty in the placement ablation change under the full Table 2 training budget, given the Holm-corrected (p=0.082) reported for the reduced-budget experiment?
  6. Table 4 indicates that more than (10^4) shots per input are needed merely to outperform the constant predictor at 14 qubits. What sampling budget is required to retain the reported accuracy, and is there evidence that the advantage persists when finite-shot noise is included during training?
1 comment
AuthorsOpenReview ↗

Response Letter to Reviewer VnGc

We thank the reviewers for their constructive feedback. Due to the length limit, responses are condensed; full analyses is added in revision. Unless noted, the new experiments use seven seeds.

Q1. All architectures reach the same 225 Fourier frequency pairs, including complex coefficients (App. B.3, D.6); MSEs use the held-out 20% (App. B.1). Despite matched reachability, our circuits achieve lower error with fewer parameters and gates, highlighting the importance of circuit design.

MethodParams2Q gatesAdvectionBurgers
HEA Circular168560.230±0.1350.746±0.273
Strongly Entangling168560.274±0.0720.933±0.037
Proposed126260.108±0.0440.200±0.011
Proposed (γ\gamma)152260.052±0.0160.189±0.014

Q2. Held-out reevaluation changes Table 2 MSEs by ≤1%, with unchanged rankings. We retrained on 8 Advection and 7 Burgers initial conditions satisfying the 14-qubit Fourier-support criterion ([−7,7]2[-7,7]^2); mean±SD is across conditions, one seed each.

MethodNew AdvectionNew Burgers
HEA Circular0.124±0.0870.348±0.108
Strongly Entangling0.181±0.1640.346±0.091
Proposed0.248±0.1510.383±0.048
Proposed (γ\gamma)0.108±0.1100.309±0.079

Proposed (γ\gamma) maintains the lowest mean error on both PDEs, extending its advantage to independent initial conditions. Each model is retrained; this is not zero-shot transfer.

Q3. Corollary 1 guarantees generic nonzero coefficients, not unrestricted joint amplitude/phase control. Jacobian ranks among 225 real coordinates are 138/152 (HEA/Strongly Entangling) versus 58/84 (Proposed/Proposed γ\gamma). Lower rank means stronger coupling among coefficients, not fewer frequencies. Our circuits nevertheless achieve lower Advection/Burgers errors (Q1). To test multi-frequency approximation, we use travelling-wave targets f(x,t)=∑k=1Kakcos⁡(kx−kt+ϕk)f(x,t)=\sum_{k=1}^K a_k\cos(kx-kt+\phi_k), with KK harmonics of independently chosen amplitude/phase. We test 10 targets per KK, three restarts, and MSE <10−310^{-3}.

MethodK=1K=1K=2K=2K=3K=3
HEA Circular10/1010/1010/10
Strongly Entangling10/1010/107/10
Proposed10/100/100/10
Proposed (γ\gamma)10/1010/1010/10

Proposed (γ\gamma) matches HEA on all multi-frequency targets despite lower Jacobian rank. This supports learning structured functions without unrestricted coefficient control.

Q4. Experiments use WR=CRW_R=CR, WL=RCRW_L=RCR, whereas Algorithm 1 uses WR=RCRW_R=RCR. App. D.6(i)–(ii) still guarantees generic nonzero coefficients and reachability: initial rotations prepare arbitrary product states preserving the spanning property. But D.6(iii) does not guarantee unit amplitude for the tested circuit. For fixed-CZ stars with hub degree ≥2, the maximum top-frequency amplitude is 1/2 for WR=CRW_R=CR, versus 1 for WR=RCRW_R=RCR (verified at 4 and 6 qubits). We will correct Algorithm 1 and clarify guarantee scope.

Q5. The 15,438-step ablation uses eight paired seeds and matched 126 parameters/26 CZ gates; 95% CIs compare to Proposed.

MethodMSEDifference [95% CI]Holm pp
Proposed0.103±0.043——
No Align0.751±0.0950.648 [0.566,0.730]1.0×10⁻⁸
Path1.032±0.0470.929 [0.880,0.977]2.4×10⁻¹⁵
Leaf Readout1.032±0.0470.929 [0.881,0.977]2.4×10⁻¹⁵
No Expose1.033±0.0500.929 [0.879,0.979]3.0×10⁻¹⁵

Unlike the reduced-budget result (Holm p=0.082p=0.082), full-budget differences are significant. Both baselines favor LR 0.015 in {0.003, 0.015, 0.03, 0.1}. At 30,876 steps (three Advection seeds), MSE ranges remain 0.036–0.084 (proposed family) versus 0.15–0.24 (baselines), supporting the placement advantage.

Q6. For 14-qubit Advection (three seeds), shots are per expectation evaluation. With exact training and finite-shot inference:

MethodShots to beat constant10410^4 shots10610^6 shots
HEA Circular1.7×10⁵20.30.358
Strongly Entangling5.6×10⁵57.90.867
Proposed2.0×10⁴2.090.127
Proposed (γ\gamma)1.4×10⁴1.310.052

Finite-shot training and inference, using shot-based parameter-shift gradients (mean±SD):

Method10210^2 shots10310^3 shots10410^4 shots
HEA Circular1.06±0.081.02±0.080.84±0.10
Strongly Entangling1.05±0.071.03±0.090.97±0.09
Proposed0.65±0.050.48±0.020.30±0.02
Proposed (γ\gamma)0.57±0.020.29±0.040.18±0.06

Both proposed architectures outperform baselines across tested budgets, with shot-based forward evaluations and gradients. Simulated-shot results support finite-shot robustness, albeit with lower accuracy than noiseless training.

Comments on the paper

1 comment

I thank the authors for the prompt response. I have reviewed all the references in the prior work section, and all of them are from physics journals, such as Physical Review, quantum-focused venues, such as PRX Quantum, or broader scientific journals, such as Nature Communications. This reinforces my view that the paper would be better suited to a specialized venue focused on quantum circuit design. While I recognize that quantum machine learning is a subfield of machine learning, I remain unconvinced that the methods developed in this paper have sufficient relevance or broader implications for the general machine learning community. I believe the work would reach a more appropriate audience in a specialized quantum computing venue.

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

Related papers

Discussion (0)

Sign in to comment.