acceptodds
Under review as a conference paper at ICLR 2027

One-Step Pure-Quantum State Generation via Classical Noise Re-Uploading

Abstract

We propose a novel pure-quantum generative model paradigm that fundamentally avoids the issue of extremely small post-selection probabilities present in previous models. Unlike existing methods that require multi-step noise, addition denoising and classical neural networks, this paradigm enables direct single-step generation of quantum data, significantly improving generation efficiency while substantially reducing the complexity of training and quantum state preparation. Furthermore, by directly sampling classical noise to generate quantum states, the sampling process becomes easier to implement. Experimental results demonstrate that this paradigm outperforms existing quantum generative models in terms of generation quality.

Primary area:
generative models
Keywords:
quantum generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CHzLOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a one-step quantum state generative model based on classical noise re-uploading. Unlike existing quantum diffusion and transport-based approaches, the proposed method maps low-dimensional classical noise directly to quantum states through a parameterized quantum circuit, without requiring iterative denoising, ancillary-qubit post-selection, or a trainable classical neural network. The generator repeatedly encodes the same classical noise vector into a quantum circuit, interleaved with trainable rotation and entangling gates. The model is optimized using a Wasserstein distribution-matching objective based on quantum-state fidelity. The authors also introduce a conditional generation mechanism using disjoint latent intervals and establish a local distributional universality result under specific circuit controllability assumptions.

The method is evaluated on several quantum-state generation tasks, including single-qubit state distributions, entanglement-entropy fitting, correlated quantum noise estimation, ten-qubit transverse-field Ising model (TFIM) ground states, and conditional generation. Experimental comparisons with QuDDPM, QDT, QGAN, QFM, and LPQC suggest improvements in distribution fitting, parameter efficiency, and ideal generation success probability.

Critical strengths
  1. A simple and potentially efficient quantum state generation architecture.

The proposed framework eliminates iterative quantum denoising and measurement-based post-selection by generating quantum states directly from classical noise. This is an appealing architectural simplification, particularly because post-selection can substantially reduce the success probability of multi-step quantum generation. The ability to generate samples using a single complete circuit execution is a meaningful advantage.

  1. Nontrivial theoretical analysis of distributional expressivity.

A significant strength is the theoretical investigation of distribution-level universality rather than only pointwise function approximation. The authors establish a local Wasserstein universality result for pure-state distributions under explicit circuit controllability assumptions. This provides a stronger theoretical foundation for the proposed framework than purely empirical demonstrations of generative expressivity.

  1. Encouraging parameter efficiency compared with existing methods.

The experiments demonstrate promising parameter efficiency, particularly in comparison with LPQC. For correlated-noise estimation, the proposed method uses 143 trainable parameters compared with 5,164 for LPQC while achieving lower mean absolute estimation errors. Similar advantages are observed in the ten-qubit TFIM experiment, suggesting that classical noise re-uploading can provide a compact representation for certain quantum-state generation tasks.

  1. Evaluation across several quantum-state generation scenarios.

The paper considers multiple tasks, including distribution fitting, entanglement-entropy estimation, correlated-noise parameter recovery, TFIM ground-state learning, and conditional generation. The inclusion of recent baselines such as QuDDPM, QFM, and LPQC strengthens the empirical relevance of the study.

  1. A straightforward conditional generation mechanism.

The proposed use of disjoint classical latent intervals provides a simple way to control the generated quantum-state distribution without introducing additional conditional quantum operations or post-selection. The experiments on multiple circles and state clusters demonstrate the feasibility of this mechanism.

Critical weaknesses
  1. Insufficiently controlled comparisons make it difficult to attribute performance improvements to the proposed architecture.

The paper compares different methods under different training objectives and experimental configurations. For example, the proposed method uses a Wasserstein distribution-matching loss, whereas the QDT baseline uses a fidelity-kernel MMD objective and QGAN uses adversarial training. Importantly, the fidelity kernel used for QDT, (k(\rho,\sigma)=\mathrm{Tr}(\rho\sigma)), yields an MMD that depends only on the difference between the mean density operators. Consequently, distinct quantum-state distributions with identical mean density operators cannot be distinguished by this objective. This may partially explain QDT's poor recovery of the target circle in Figure 4. Without controlling the training loss, it is difficult to determine whether the reported improvements primarily arise from the generator architecture or from a more suitable optimization objective. This is a central concern because the paper makes strong claims about superior generative performance.

  1. The experiments provide limited quantitative evidence for learning complete quantum-state distributions in more complex systems.

Although the paper evaluates several tasks, most direct distribution-level comparisons are conducted on relatively simple single-qubit datasets. For the ten-qubit TFIM task, the evaluation primarily reports the distribution of transverse magnetization, which captures only one physical observable and does not fully characterize the generated quantum-state distribution. Similarly, accurate entanglement-entropy fitting does not necessarily imply accurate recovery of complete quantum states. The conditional generation results are mainly qualitative. Consequently, the experimental evidence is not yet sufficient to establish the claimed superiority in general quantum-state distribution learning, especially for larger or more complex quantum systems.

  1. The claimed computational efficiency is not comprehensively evaluated.

The paper emphasizes one-step generation, reduced training costs, and unit ideal success probability. However, one-step generation refers to a single complete circuit execution rather than a shallow quantum circuit. The proposed architecture can still require a substantial number of re-uploading layers, such as (P=100) in the ten-qubit experiment. The reported efficiency comparisons mainly focus on trainable parameter counts, post-selection probabilities, and numbers of pairwise overlap estimates. They do not provide comprehensive measurements of circuit depth, quantum gate counts, measurement shots, gradient-estimation costs, or actual training and generation time. Furthermore, the experiments rely on idealized simulation without a systematic evaluation of hardware noise and finite-shot measurement errors. These omissions make it difficult to assess the practical computational advantages of the proposed approach.

  1. The ablation studies are largely qualitative and do not sufficiently isolate the contributions of individual components.

The paper investigates noise width, latent dimension, re-uploading depth, and the number of parameter layers. However, the results are presented primarily through Bloch-sphere visualizations without quantitative Wasserstein distances, standard deviations, or convergence comparisons. Moreover, the latent-dimension experiment changes both input dimension and the number of re-uploadings under a fixed parameter budget, making it difficult to isolate the effect of dimensionality. The ablation studies also disable learnable noise scales, although these scales are used in several main experiments. A more systematic evaluation is needed to establish which components are essential and how they affect performance.

  1. The scope of the theoretical guarantees and several implementation details require clarification.

The main universality theorem relies on a sufficiently small noise domain and strong controllability assumptions on the parameter layers. It establishes the existence of suitable circuit parameters but does not provide quantitative bounds on the required re-uploading depth or guarantee that gradient-based training can discover such parameters. The relationship between these assumptions and the experimental configurations should be explained more explicitly. There are also minor reproducibility concerns: the main text reports five independent runs for the entanglement-entropy experiment, whereas Appendix D.1.3 states three runs. Additionally, the exact circuit configuration underlying the reported parameter counts, particularly the 5,703 parameters in the TFIM experiment, would benefit from clarification.

Questions for authors
  1. How much of the reported performance improvement comes from the generator architecture rather than the training objective?

The proposed method uses Wasserstein distribution matching, whereas several baselines employ different objectives. In particular, the fidelity-kernel MMD used by QDT does not distinguish all quantum-state distributions. Could the authors provide controlled comparisons in which alternative generator architectures are trained using the same Wasserstein objective, under comparable parameter or circuit-resource budgets? Such an experiment would materially strengthen the evidence for the proposed architectural contribution.

  1. Can the authors provide more comprehensive quantitative evaluations of complete quantum-state distribution matching?

The ten-qubit TFIM experiment mainly evaluates transverse magnetization, while the conditional generation experiments rely largely on qualitative visualization. Could the authors report quantitative distribution distances, such as empirical Wasserstein distances or other appropriate distribution-level metrics, for these experiments? It would also be useful to clarify whether the generated TFIM states accurately reproduce physical properties beyond magnetization. These results would help determine whether the model learns the full target distribution rather than only selected observables.

  1. How do the methods compare in terms of actual quantum computational resources?

Although the proposed method avoids post-selection, it can still involve deep quantum circuits due to repeated noise encoding. Could the authors report total circuit depth, single- and two-qubit gate counts, training time, generation time, and measurement-shot requirements under matched experimental settings? In particular, how do the advantages change when finite-shot estimation and realistic quantum noise are considered? This information is important for evaluating the practical significance of the claimed efficiency improvements.

  1. How restrictive are the assumptions required for the distributional universality theorem in practical implementations?

Theorem 1 requires a sufficiently small latent-noise domain and specific controllability properties of the parameter layers. Could the authors explain how these assumptions are verified for the precise circuit architectures used in the experiments? Additionally, is there any quantitative relationship between the approximation accuracy, the number of qubits, and the required re-uploading depth? Clarifying these points would help assess the practical implications of the theoretical contribution.

  1. Can the authors provide quantitative ablations and reconcile the implementation details?

The current ablation studies are primarily qualitative. Could the authors report quantitative distribution-matching performance across multiple random seeds and investigate the contributions of learnable noise scales and repeated encoding under matched parameter budgets? It would also be helpful to provide explicit circuit configurations and parameter-count calculations for the main experiments, and to resolve the discrepancy between the reported three and five independent runs in the entanglement-entropy experiment. These clarifications would improve confidence in the robustness and reproducibility of the results.

6 comments
AuthorsOpenReview ↗

Author Response (Part 1)


W1 & Q1: Insufficiently controlled comparisons make it difficult to attribute performance improvements to the proposed architecture.

We thank the reviewer for pointing this out. For QDT, we replaced its loss with the same loss used in our method and conducted five independent runs. The resulting test quantum-state data are summarized in the table below:

DistributionW2W_2 (mean ±\pm std.)
Target0.0420±0.00430.0420 \pm 0.0043
QDT0.3171±0.00480.3171 \pm 0.0048
QGAN0.7080±0.00000.7080 \pm 0.0000
Ours0.0820±0.01560.0820 \pm 0.0156

The empirical distribution of the first batch of target samples serves as the fixed reference for all distances; the Target row compares the empirical distributions of subsequent target batches with this reference.


W2.1 & Q2 The experiments provide limited quantitative evidence for learning complete quantum-state distributions in more complex systems.

We thank the reviewer for the reminder. We have added to Appendix G the Wasserstein distance metrics for noise learning and for complex distributions such as TFIM. Specifically, for the noise learning in Section 5.4, the Wasserstein distances are as follows:

MethodN=1000N = 1000N=2000N = 2000N=5000N = 5000
LPQC0.2264±0.01500.2264 \pm 0.01500.2261±0.01620.2261 \pm 0.01620.2264±0.01820.2264 \pm 0.0182
Ours0.1152±0.02910.1152 \pm 0.02910.1093±0.01630.1093 \pm 0.01630.1085±0.00750.1085 \pm 0.0075

Furthermore, for the TFIM model in the main text, the Wasserstein distances are as follows. As requested by the reviewer, we also use the ground-state energy value EE, one of the most commonly used metrics in quantum computing, to further verify the effectiveness of our method.

MethodW2W_2EE
Target--−15.6478±0.0033-15.6478 \pm 0.0033
LPQC0.1984±0.06310.1984 \pm 0.0631−15.0568±1.5860-15.0568 \pm 1.5860
Ours0.0505±0.02120.0505 \pm 0.0212−15.6375±0.0056-15.6375 \pm 0.0056

W2.2 Accurate entanglement-entropy fitting does not necessarily imply recovery of complete states, and the conditional-generation results are mainly qualitative.

We thank the reviewer. We would like to point out that QFM itself is a method that fits only a certain metric rather than the distribution, and we have clearly illustrated this in Appendix A.2: indeed, one metric can correspond to multiple different quantum state distributions. However, this is a limitation of QFM; our method is capable of fitting the distribution, and we have demonstrated this in the paper. The entanglement fitting is only intended to show that our method can also fit metrics, and that it performs better than methods specifically designed for metric fitting.


W2.3 the experimental evidence does not yet establish claimed superiority in general quantum-state distribution learning, particularly for larger or more complex systems.

We would like to emphasize that we have demonstrated the advantages of our method across multiple benchmarks and a range of system sizes. While evaluating every possible dataset is infeasible, we do not claim that our method universally outperforms all alternatives. Rather, we believe that the breadth of our experiments provides compelling empirical evidence of its comparative advantages within the settings considered.

The experiments presented in our manuscript involve systems of up to 10 qubits. In comparison, the largest system sizes investigated in prior studies are 4 qubits for QuDDPM, 2 qubits for both QDT and QGAN, 8 qubits for LPQC, and 11 qubits for QFM. Given that the dimension of the quantum Hilbert space grows exponentially with the number of qubits, our experiments already cover a substantial computational scale relative to existing studies on quantum generative modeling.

To further assess the performance of our method at larger system sizes, we have extended the TFIM experiments to 14, 16, 18, and 20 qubits with P=50P=50. The results are presented below:

QubitsW2W_2MtargetM_{\mathrm{target}}MtestM_{\mathrm{test}}
140.0832±0.00070.0832 \pm 0.00070.87140.8697±0.00140.8697 \pm 0.0014
160.0927±0.00030.0927 \pm 0.00030.86980.8636±0.00010.8636 \pm 0.0001
180.0966±0.00430.0966 \pm 0.00430.86850.8610±0.00390.8610 \pm 0.0039
200.0891±0.00310.0891 \pm 0.00310.86750.8641±0.00230.8641 \pm 0.0023
AuthorsOpenReview ↗

Author Response (Part 2)


W3 & Q3 Its claimed efficiency is unconvincing without full circuit-cost evaluation and noise-aware experiments.

  • Fairness of circuit configurations. We emphasize that, wherever possible, all other experimental settings are held fixed to ensure a fair comparison. Specifically, all methods evaluated in this work employ the circuit architecture shown in Fig. 2(b), with the same number of circuit layers used for our method and the baselines. The comparisons therefore use identical circuit depths, quantum gate counts, measurement-shot budgets, and gradient-estimation costs.
  • Training and generation time. Our method uses a single-step formulation for both training and generation, avoiding the computational overhead associated with repeated steps in iterative approaches. Under otherwise comparable conditions, this provides a computational advantage over methods requiring multiple steps. Compared with LPQC under the same circuit configuration, our method also eliminates the additional classical neural-network computations required by LPQC during both training and generation, thereby reducing the associated computational overhead.
  • Effects of noise and finite-shot measurements. Our objective is to learn distributions over quantum states, a task that can be formulated within the framework of fault-tolerant quantum computation. The primary contribution of this work is algorithmic; we do not claim to have established an advantage on noisy quantum hardware. Moreover, noise and finite-shot errors affect all methods considered and should therefore be assessed under matched experimental conditions. We expect our method’s advantages to persist in sufficiently weak noise regimes where the relevant quantum signal is preserved, although relative performance under noise ultimately requires empirical verification.

To assess robustness to noise, we have added an ablation study in which a depolarizing channel is applied after every single-qubit gate, with the noise-width parameter fixed at 0.02. The results for different depolarizing strengths are reported below.

Depolarizing strength ppW2W_2
00.0651 ± 0.0070
10−410^{-4}0.0652 ± 0.0071
10−310^{-3}0.0690 ± 0.0095
10−210^{-2}0.1338 ± 0.0056

These results show that weak depolarizing noise has only a limited effect on our method’s performance. Although stronger noise leads to a larger discrepancy, the generated quantum states still tend to concentrate near the training distribution, as illustrated in Fig. H.6.

We have also added an ablation study examining the effect of finite measurement-shot budgets. As shown below, the results obtained with 2,000 and 5,000 shots per overlap are close to the ideal-measurement baseline, whereas a budget of 100 shots leads to some degradation. These results indicate that our method can maintain performance close to the ideal baseline using a finite number of measurements.

Training shots per overlapW2W_2
1000.0765 ± 0.0134
20000.0414 ± 0.0067
50000.0455 ± 0.0067
Ideal0.0428 ± 0.0066

Because measurements in our method are primarily used to estimate the overlap matrix, Fig. H.8 further illustrates how the estimated matrix varies with the measurement-shot budget.

AuthorsOpenReview ↗

Author Response (Part 3)


W4 & Q5 The ablation studies are primarily qualitative.

We thank the reviewer for pointing this out. We have added the Wasserstein distances for all ablation experiments to the appendix.

In addition, we would like to point out that, while keeping the existing architecture unchanged, there is no way to simultaneously fix the number of repetitions P and vary the input dimension. With a fixed number of qubits, different input dimensions require different numbers of layers; if the number of repetitions is then also fixed, the number of parameters will be completely unequal. Unless the circuit structure is altered—for example, by placing parameterized gates after PP repetitions to keep the parameter count equal—which would entirely destroy the circuit structure and make a fair comparison impossible.

We have added an ablation study on with and without scaling, and it can be observed that scaling does indeed improve performance.

wwWithout scaling W2W_2With scaling W2W_2
0.020.0655 ± 0.00590.0436 ± 0.0009
0.20.0428 ± 0.00660.0296 ± 0.0011
20.3161 ± 0.03770.0315 ± 0.0028
2π2\pi0.4282 ± 0.01120.2522 ± 0.0041

W5 & Q4 Clarifying the Applicability and Practical Verifiability of the Universality Theorem's Assumptions

We thank the reviewer for this clarification. Our theorem establishes the expressive capacity of forward-only data re-uploading: under the stated assumptions, sufficiently deep circuits can approximate arbitrary pure-state distributions. It provides an existence guarantee, without quantitative depth bounds or guarantees for gradient-based training. We will clarify this scope and explain that the experiments assess finite-depth performance and trainability empirically.

Theorem 1 requires a sufficiently small latent-noise domain. Therefore, we also demonstrate in our ablation experiments that enlarging the noise width does indeed degrade performance.

Regarding the entanglement experiments, we have fully rerun them five times and updated the results in the main text.

Reviewer_L2LgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The author appears to propose a data re-uploading quantum circuit that uses a 2-Wasserstein loss function as a generative model. The experimental results also look good compared with those of previously proposed baselines.

Critical strengths
  1. The circuit structure is simple, and the numerical results are promising.
  2. The paper provides some theoretical results on expressivity.
Critical weaknesses
  1. The phrase “per complete loss” appears to be missing from Fig. 3(c).
  2. The description says this 2-Wasserstein cost can be evaluated via a SWAP test, but the test appears to require an unmentioned state-preparation input stage, which is very hard to implement and may require very deep circuit.
  3. The writing is somewhat confusing. The quantum state and loss function are introduced in Eqs. (3)–(5), but the conditional generation in Eq. (6) does not seem connected to the preceding description. What do the disjoint intervals refer to? intervals in Hilbert space?
  4. Regarding the term “one-step generator,” isn’t this simply a type of quantum circuit Born machine?
Questions for authors
  1. Is the training process scalable?
  2. What does the sentence in line 162—“One draw of x followed by one evaluation of Eq. (2) produces one sample. No measurement outcome is accepted or rejected, so the ideal success probability is one.”—mean? What does “success probability” refer to?
  3. Following weakness 3, what exactly are the training and sampling processes?
  4. In Fig. 1, where is the measurement process in the proposed method?
3 comments
AuthorsOpenReview ↗

Author Response (Part 1)


W1 The phrase “per complete loss” appears to be missing from Fig. 3(c).

We thank the reviewer for pointing this out. We have changed the y-label in Fig. 3(c) to "SWAP-test evaluations" and clarified in the caption: "Pairwise SWAP-test evaluations for one transport-loss computation per training stage, summed over all stages.”


W2 The description says this 2-Wasserstein cost can be evaluated via a SWAP test, but the test appears to require an unmentioned state-preparation input stage, which is very hard to implement and may require very deep circuit.

We would like to emphasize that the state-preparation input stage is required not only by our method but by all methods for computing the 2-Wasserstein cost except QGAN; this is not unique to our approach. Furthermore, assuming that the cost of preparing a single target state is cc and that the number of measurement shots is SS, in our method the computational cost of computing the inner products with MM target states per cost evaluation is only SMcSMc. In contrast, since QuDDPM involves intermediate states and assuming that TT steps are required, the preparation cost of QuDDPM is TSMcTSMc. Except for QGAN, the other methods have the same preparation cost as ours. Although QGAN has a lower preparation cost, it is unusable due to its poor performance.


W3 The writing is somewhat confusing

We thank the reviewer for this comment. We have restructured the Methods section to first introduce the generation of quantum-state distributions, followed by our proposed method, the training loss, and finally conditional generation.

Disjointness is imposed only on the classical noise regions. All conditions generate states in the same nn-qubit Hilbert space: their output-state distributions may overlap, and states from different conditions need not be orthogonal. Sampling noise from different regions allows the model to generate different conditional state distributions.

In addition, we have redrawn the illustration in the main text and provided illustrations of all methods in Appendix A. We hope these revisions make the presentation clearer and easier to follow.


W4 Regarding the term “one-step generator,” isn’t this simply a type of quantum circuit Born machine?

QCBM (quantum circuit Born machine) fits a classical distribution and generates only a single quantum state, whereas a classical distribution can correspond to multiple quantum states.

Our method fits a distribution of quantum states and can generate multiple quantum states. QCBM cannot fit a distribution of quantum states. We have illustrated the difference between the two in Figure A.7, which we believe will help you better understand.


Q1 Is the training process scalable?

To assess the scaling of our method at larger system sizes, we have extended the TFIM experiments to 14, 16, 18, and 20 qubits with P=50P=50. The results are presented below:

QubitsW2W_2MtargetM_{\mathrm{target}}MtestM_{\mathrm{test}}
140.0832±0.00070.0832 \pm 0.00070.87140.8697±0.00140.8697 \pm 0.0014
160.0927±0.00030.0927 \pm 0.00030.86980.8636±0.00010.8636 \pm 0.0001
180.0966±0.00430.0966 \pm 0.00430.86850.8610±0.00390.8610 \pm 0.0039
200.0891±0.00310.0891 \pm 0.00310.86750.8641±0.00230.8641 \pm 0.0023

By contrast, prior studies of other methods have been limited to small qubit counts: 4 qubits for QuDDPM, 2 qubits for both QDT and QGAN, 8 qubits for LPQC, and 11 qubits for QFM. We would like to emphasize that as the number of qubits grows, the quantum space grows exponentially. Although 20 qubits may appear to be only 12 more than 8, the quantum space is enlarged by a factor of 4096. We believe this experiment demonstrates the scaling of our method.

AuthorsOpenReview ↗

Author Response (Part 2)


Q2 What does “success probability” refer to?

We thank the reviewer for this reminder. We have revised the manuscript, mainly emphasizing that, compared with QuDDPM and other methods, our approach:

  1. requires no auxiliary qubits

  2. with the same noise, can generate exactly the same quantum state every time. This is highly important for future research on quantum state generative models, because studying quantum states requires a large number of measurements on the same quantum state, and the fact that different noise produces completely different quantum states greatly hinders such follow-up research.


Q3 what exactly are the training and sampling processes?

Training. We sample classical noise vectors and generate quantum states by alternating encodings of the same noise vector with trainable quantum layers. We then optimize the circuit parameters to match the generated and target state distributions using a Wasserstein-based transport loss.

Sampling. We fix the trained parameters, draw a fresh classical noise vector, and apply the complete generator circuit once to ∣0n⟩|0^n\rangle to obtain one quantum-state sample. Repeating this process produces the learned state distribution. For conditional generation, we simply sample noise from the region assigned to the desired condition.

We have vividly illustrated our method in Figure A.6, and we believe this will help you understand it.


Q4 where is the measurement process in the proposed method?

During sampling, our target is to directly generate quantum states, without involving any measurement process.

During training, our measurement process mainly consists of a Swap test between the quantum state generated by the circuit and the target state.

Dear Reviewer,

Thank you again for your thoughtful feedback. We would greatly appreciate your consideration of our responses and the revised manuscript, which includes additional experiments and clarifications addressing your concerns. Please let us know if any points remain unresolved or would benefit from further explanation; we would be happy to discuss them.

We understand that the review period is busy and sincerely appreciate your time and consideration.

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

This paper proposes a one step quantum generation model. A classical noise vector is repeatedly encoded into a parameterized quantum circuit and directly generates a quantum state. Unlike quantum diffusion models it avoids intermediate stages, ancilla measurements, and post-selection, theoretically ensuring a success probability of one. The basic idea is that the required nonlinearity arises from the variation of the classical input rather than from nonlinear quantum dynamics. The approach appears promising as a simpler and more efficient alternative for generating quantum states.

Critical strengths

a. The paper replaces the multi-step quantum diffusion process with direct, one step generation, without post-selection, which is a reasonable new idea. b. The elimination of ancillas, post-selection etc reduces the complexity of generation and can avoid extremely low success probabilities. c. Sampling classical noise is straightforwar, while discrete latent intervals allow for conditional generation with shared circuit parameters, making the idea simpler than its precursors. d. The authors provide clear visualizations of the methods prefrormance.
e. Conditional genearation results enhance the study's depth.

Critical weaknesses

a. The model relies substantially on classical noise sampling and classical data encoding, so it is not fully quantum at the level of the overall pipeline. b. The reported resource advantage remains primarily theoretical or limited to small-scale experiments. c. Despite the somewhat extensive study design, the quantitative comparisons are moslty limited to comparing with the original QuDDPM. d. Perhaps the methodological improvments are not substancial.

Questions for authors

How do the circuit depth, the number of parameters and the number of required shots scale as the number of qubits or the complexity of the target distribution increases?

4 comments
AuthorsOpenReview ↗

Author Response (Part 1)


W1** Its “pure-quantum” claim is questionable because the generation pipeline relies on classical noise sampling and data encoding.

We thank the reviewer for highlighting this distinction. Our use of “pure-quantum” was intended to describe the generator architecture without a classical neural network, and we recognize that this terminology can suggest a broader claim.

Specifically, the basic model samples three independent uniform random variables and reuses the same latent vector throughout the circuit. This sampling procedure does not learn the target distribution. The latent variables directly specify quantum rotation gates, En(x)=[Rz(x3)Ry(x2)Rz(x1)]⊗nE_n(\boldsymbol{x})=[R_z(x_3)R_y(x_2)R_z(x_1)]^{\otimes n}, interleaved with trainable quantum control layers. Thus, although the encoding parameters are classical, the encoding operations act on the quantum register. No classical neural network maps the latent input to the control-layer parameters, and the generation architecture does not require computing a classical amplitude representation of the output state.

Our contribution is therefore a one-step quantum-state generator with classical noise and no classical neural network. Conditioned on a noise, state preparation is unitary and requires neither ancillary measurements.


W3 Missing quantitative data.

We thank the reviewer for this helpful comment. We have added quantitative results for all experiments to Appendix G of the revised manuscript. Our initial emphasis on qualitative results was motivated by the intuitive nature of Bloch-sphere visualizations, which help readers interpret the experimental outcomes more readily.


Q1 How do the circuit depth, the number of parameters and the number of required shots scale as the number of qubits or the complexity of the target distribution increases?

In Appendix E.2, we added single-qubit teacher–student experiments, using the number of re-uploadings in the fixed teacher circuit PTP_T as an indicator of the target distribution complexity, and varying the student circuit's PSP_S to examine the resources required to reach W22≤10−3W_2^2 \le 10^{-3}.

We tested PS∈{1,10,50,200}P_S \in \{1, 10, 50, 200\}. If PS=1P_S = 1 already meets the threshold, we set P^S=1\widehat{P}_S = 1; otherwise, we perform log–log linear interpolation between the adjacent test points that first cross the threshold:

\frac{\log(c_a / 10^{-3})}{\log(c_a / c_b)}}, \qquad c_a > 10^{-3} \ge c_b,$$ where $c_a, c_b$ are the corresponding test $W_2^2$ values. | Teacher $P_T$ | Estimated required student $\widehat{P}_S$ | | --- | --- | | 10 | 1.0 | | 20 | 1.0 | | 50 | 3.8 | | 100 | 65.7 | | 200 | 183.1 | The results show that, within this circuit family, more complex targets require more re-uploadings. For $n$ qubits, the number of parameters is $O(nP)$; under the condition that the depth of each re-uploading block is fixed, the circuit depth is $O(P)$. Regarding shots, if $S$ measurements are used per overlap estimate, one loss evaluation for $M$ target states and $N$ generated states requires $SMN$ shots, this count does not explicitly depend on $n$ or $P$.
AuthorsOpenReview ↗

Author Response (Part 2)


W2 The reported resource advantage remains primarily theoretical or limited to small-scale experiments

We would like to emphasize that the resource advantages of our method are architectural facts, and the theorems in our paper concern only expressivity, not resources. Moreover, our resource advantage can be seen directly without requiring large-scale qubits—indeed, the more qubits there are, the more pronounced our resource advantage becomes, as detailed below:

  • Simplified input preparation. For the Haar-initialized baselines considered here, preparing the random quantum input is a required physical operation that consumes quantum resources. Our generator starts from a fixed computational-basis state and samples classical noise, eliminating this separate quantum-input preparation stage. This is a concrete saving in the implementation, not merely a theoretical prediction, and it holds regardless of the number of qubits. As preparing random quantum states becomes more costly at larger system sizes, the resources saved by removing this stage can increase.
  • One-step generation. The multi-stage QuDDPM and QFM constructions considered here require each sample to pass through a sequence of denoising or flow stages. Our method directly prepares the output through one trained generator, without separately learned intermediate stages. The elimination of this multi-stage procedure is an architectural fact, not merely a theoretical prediction, and it persists at every system size. As the number or complexity of the baseline’s intermediate stages increases, the potential savings from avoiding this procedure can become larger.
  • Single-stage distribution learning. QuDDPM requires intermediate reference ensembles and stage-specific distribution-matching objectives, each involving actual state preparations and loss evaluations. Our method fits only the final target distribution, eliminating those intermediate training requirements. Under the matched budgets specified in our resource analysis, this gives O(MN)O(MN) overlap estimates instead of O(TMN)O(TMN) across TT QuDDPM stages. These intermediate preparations and evaluations are removed from the actual procedure, not merely from a theoretical model, and their removal holds at any qubit count.
  • No classical neural network during generation. LPQC processes the sampled noise through a classical neural network to obtain input-dependent quantum-circuit parameters. Our method directly encodes the noise through quantum gates, eliminating this network evaluation and its associated parameter storage. The absence of this classical neural component is directly verifiable in the implementation, not merely theoretically predicted, and it holds at every system size. If larger quantum circuits require larger neural parameter maps, the computation and storage saved by omitting those maps can increase. Our experiments already demonstrate concrete parameter savings: the evaluated 10-qubit TFIM models use 5,703 trainable parameters for our method and 135,588 for LPQC.
  • No generation ancillas or intermediate measurements. The measurement-assisted baselines considered here require auxiliary qubits and intermediate readout operations. Our pure-state generator acts only on the data register and uses unitary evolution, requiring neither generation ancillas nor intermediate measurements. These physical resource requirements are eliminated by construction, not merely predicted to decrease, and this property remains valid regardless of system size. When larger-scale baseline implementations require more auxiliary qubits, ancillary gates, or readout operations, the corresponding resource savings can grow. For fixed parameters and a fixed noise realization, our generator also prepares a uniquely determined output state under ideal execution, without measurement-induced branching or post-selection.

To directly address the concern about system size, we have extended the TFIM experiments to 14, 16, 18, and 20 qubits. The results are reported in the revised appendix. The number of repetitions is 50 (i.e., P = 50) for all qubit counts. As can be seen, even at large qubit scales (where the Hilbert space grows exponentially), our method still achieves strong performance. By contrast, prior studies of other methods have been limited to small qubit counts: 4 qubits for QuDDPM, 2 qubits for both QDT and QGAN, 8 qubits for LPQC, and 11 qubits for QFM.

QubitsW2W_2MtargetM_{\mathrm{target}}MtestM_{\mathrm{test}}
140.0832±0.00070.0832 \pm 0.00070.87140.8697±0.00140.8697 \pm 0.0014
160.0927±0.00030.0927 \pm 0.00030.86980.8636±0.00010.8636 \pm 0.0001
180.0966±0.00430.0966 \pm 0.00430.86850.8610±0.00390.8610 \pm 0.0039
200.0891±0.00310.0891 \pm 0.00310.86750.8641±0.00230.8641 \pm 0.0023
AuthorsOpenReview ↗

Author Response (Part 3)


W4 Perhaps the methodological improvments are not substancial

We respectfully disagree with the suggestion that the methodological improvements are not substantial. Our contribution combines a different construction for quantum-state distribution generation, an expressivity guarantee for that architecture, and experimentally demonstrated improvements in generation quality and parameter efficiency. We clarify these contributions below.

  1. A different construction for quantum-state distribution generation. Our method learns a direct map from classical noise to quantum states: x∼μ,ρΘ(x)=UP(x;Θ)∣0n⟩⟨0n∣UP†(x;Θ)\boldsymbol{x}\sim\mu,\rho_\Theta(\boldsymbol{x}) = U_P(\boldsymbol{x};\Theta) |0^n\rangle\langle0^n| U_P^\dagger(\boldsymbol{x};\Theta) . The same noise is repeatedly encoded between trainable quantum control layers. Once trained, the control parameters remain fixed, while different noise samples select different output states. Relative to QuDDPM, this removes the construction and learning of intermediate diffusion ensembles. Relative to the QDT/QGAN ensemble baselines considered here, it removes the requirement for random quantum input states. Relative to LPQC, it removes the classical neural network that maps noise to circuit parameters. These changes alter how the generator represents and prepares a distribution of quantum states. They are architectural differences, rather than adjustments to the hyperparameters of an existing generator.
  2. A distributional expressivity result under explicit architectural restrictions. Removing components from a generative model raises a substantive question: can the resulting architecture still approximate arbitrary target distributions? Our theorem addresses this question for pure-state distributions. Under the stated control assumptions and with three-dimensional noise supported on a sufficiently small cube, we prove that, for every target distribution QQ and every ε>0\varepsilon>0, there exist a finite re-uploading number PP and parameters Θ\Theta such that W2(Q,QP,Θ)<εW_2(Q,Q_{P,\Theta})<\varepsilon. However, QuDDPM, QFM, QDT, and QGAN do not have similar proofs, and the proof for LPQC relies mainly on the capability of classical neural networks.
  3. Conditional generation through a shared circuit. The construction also supports selecting different target distributions by changing only the noise-sampling interval. Within each conditional model, all conditions share the same trained circuit parameters; generation requires no additional conditioning register, intermediate measurement, or post-selection. We demonstrate this using a shared model for two circle distributions and a separate shared model for four cluster distributions. This provides an explicit mechanism for controlling the generated distribution through the noise source while retaining the same quantum generator.
  4. Substantial empirical differences in the evaluated configurations. The architectural changes are accompanied by measurable improvements. On the real-amplitude circle task, our empirical W2W_2 is 0.08200.0820, compared with 0.3171 for QDT and 0.7080 for QGAN. In the 10-qubit TFIM comparison, our model achieves W2=0.0505W_2=0.0505, compared with 0.19840.1984 for LPQC, while using 5,703 rather than 135,588 trainable parameters—approximately a 24-fold reduction. Thus, the tested model achieves better distribution fitting with substantially fewer parameters. Furthermore, the additional TFIM experiments extend the evaluation to 14, 16, 18, and 20 qubits, with mean W2<0.1W_2<0.1 at every tested size over three independent runs.

The methodological contribution is therefore the construction and justification of a quantum-state distribution generator that jointly dispenses with random quantum input preparation, intermediate diffusion stages, classical neural parameter maps, and generation-time auxiliary measurements. Although data re-uploading itself is an established technique, establishing its distribution-generating capability under these restrictions, implementing conditional generation within the same framework, and demonstrating the resulting empirical benefits constitute concrete methodological advances. We believe that these contributions warrant a more substantial assessment than the characterization suggested in this comment.

Dear Reviewer,

Thank you again for your thoughtful feedback. We would greatly appreciate your consideration of our responses and the revised manuscript, which includes additional experiments and clarifications addressing your concerns. Please let us know if any points remain unresolved or would benefit from further explanation; we would be happy to discuss them.

We understand that the review period is busy and sincerely appreciate your time and consideration.

Comments on the paper

1 comment

Dear Reviewers,

We sincerely thank all reviewers for their constructive feedback. Below, we summarize the additional experiments and revisions made in response to their comments.

  • Figure 4 & Appendix G.1: Controlled comparison with QDT retrained using the same Wasserstein transport loss as our method, with quantitative results over five independent runs.
  • Appendix E.1: TFIM experiments extended to 14, 16, 18, and 20 qubits, reporting empirical Wasserstein distances and transverse magnetization.
  • Appendix E.2: Teacher–student experiments investigating how the required number of re-uploadings changes with target distribution complexity.
  • Appendices G.2–G.4: Quantitative distribution-level evaluations for correlated-noise learning, ten-qubit TFIM, and conditional generation, including additional energy statistics for TFIM.
  • Appendices G.5–G.9 & H: Quantitative ablations of noise width, latent dimension, re-uploading number, parameter layers, and trainable noise scales.
  • Appendices G.10–G.11 & H: Robustness experiments with depolarizing noise and finite measurement-shot budgets during training.
  • Figure 5 & Appendix F.1.3: Entanglement-entropy experiments rerun over five independent trials, with updated results and consistent reporting. Revisions
  • Reorganized the presentation to introduce the problem formulation, generator, training objective, and conditional generation more clearly.
  • Added illustrations of the generation and training procedures for the compared methods, including the distinction from quantum circuit Born machines (Appendix A).
  • Expanded the resource analysis to specify overlap-estimation counts, target-state preparation requirements, and generation depth under explicit comparison assumptions (Appendix C), and revised the resource labels and explanations in Figure 3.

The additional results show that our method retains its advantage over QDT under a shared training objective and provides accurate distribution fitting across the evaluated tasks, including systems of up to 20 qubits. The robustness experiments also show limited degradation under weak depolarizing noise and performance close to the ideal baseline with sufficient training shots. Major additions and revisions are highlighted in blue in the revised PDF.

We sincerely appreciate the reviewers’ comments, which have helped us improve the clarity, quantitative evaluation, and presentation of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.