acceptodds
Under review as a conference paper at ICLR 2027

Overcoming the Representational Limitations of Amplitude Encoding for Quantum Neural Networks

Abstract

Quantum Neural Networks (QNNs) offer a framework for learning functions from classical data, but their expressive power depends critically on how inputs are encoded into quantum states. Angle encoding can require resources that scale with the number of features, whereas amplitude encoding represents features using only qubits. This compactness comes at a representational cost: normalization makes conventional amplitude encoding invariant to nonzero real rescaling of the input, up to an unobservable global phase. We show that this invariance induces an exact, architecture-independent restriction on the resulting hypothesis class. No downstream quantum circuit can distinguish inputs differing by nonzero real scaling. We develop a general condition under which deterministic preprocessing can remove this ambiguity and show that nonlinearity alone is insufficient. Based on this analysis, we introduce a fixed nonlinear feature map that provably distinguishes all real inputs under subsequent amplitude encoding, without adding trainable parameters and while retaining logarithmic qubit scaling. We evaluate the resulting QNN on three canonical classification benchmarks and a nonlinear synthetic dataset against conventional amplitude and angle encoding, as well as a classical neural network. Across these tasks, the proposed encoding consistently improves classification accuracy over both quantum encoding baselines by 21–44 %. Our results establish normalization-induced scale invariance as a concrete representational bottleneck of amplitude encoding and demonstrate that it can be removed while preserving a compact quantum representation, marking a step toward quantum models capable of capturing arbitrary relationships in classical data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Quantum machine learning, representation learning, amplitude encoding, Quantum neural networks, learning theory, data encoding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bdo2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper, “Overcoming the Representational Limitations of Amplitude Encoding for Quantum Neural Networks,” investigates a fundamental limitation of conventional amplitude encoding in variational quantum circuits (VQCs). Although amplitude encoding represents high-dimensional classical data using only O(log⁡2d)O(\log_2 d) qubits, its normalization procedure eliminates information about input magnitude, making inputs that differ only by a nonzero scaling factor physically indistinguishable. This restricts the representational capacity of downstream quantum neural networks, regardless of their circuit architecture. The paper's main contributions are twofold. First, it provides a theoretical analysis of normalization-induced scale invariance and establishes conditions under which deterministic preprocessing can eliminate this ambiguity. Second, it proposes a fixed, non-trainable nonlinear feature map, Φ(x)=(x,x∘2,cos⁡(x))\Phi(x)=(x,x^{\circ 2},\cos(x)), which preserves distinctions between real-valued inputs after amplitude encoding. The proposed approach retains logarithmic qubit scaling without introducing additional trainable parameters.

Critical strengths
  1. Rigorous Theoretical Characterization of the Representational Limitations of Amplitude Encoding.:The paper provides a mathematically grounded analysis of the scale invariance induced by amplitude normalization, establishing that inputs differing by nonzero real scaling factors become physically indistinguishable regardless of the subsequent quantum circuit architecture. More importantly, the authors formulate a general condition under which deterministic preprocessing can eliminate this ambiguity and demonstrate why nonlinearity alone is insufficient. This analysis is particularly valuable because it identifies an intrinsic limitation of the encoding mechanism rather than an architecture-dependent weakness, providing a theoretical foundation for understanding and improving the expressive capacity of quantum neural networks.

  2. Theoretically Motivated Feature Mapping with Preserved Qubit Efficiency: The proposed feature map combines linear, quadratic, and trigonometric components to distinguish arbitrary real-valued inputs after amplitude encoding. Its principal theoretical advantage lies in addressing the loss of radial information without introducing additional trainable parameters or sacrificing the logarithmic scaling of qubit requirements. This establishes a meaningful connection between representational fidelity and quantum resource efficiency. The construction is consequential because it offers a mathematically motivated approach to improving quantum data representations while preserving one of the fundamental advantages of amplitude encoding for high-dimensional inputs.

Critical weaknesses
  1. Limited Theoretical Novelty and Insufficient Analysis of Quantum Resource Complexity: Although the proposed feature map addresses the scale invariance of amplitude encoding, its underlying construction relies on a relatively straightforward concatenation of linear, quadratic, and trigonometric transformations. The paper establishes its injectivity under amplitude normalization but provides limited analysis of whether this construction offers substantive theoretical advantages over simpler alternatives, such as appending a constant component to the original feature vector. Furthermore, the claimed computational efficiency is primarily justified through logarithmic qubit scaling, without adequately examining the gate complexity and circuit depth required for quantum state preparation. These omissions are consequential because they leave both the theoretical novelty and the practical resource advantages of the proposed encoding insufficiently established.

  2. Insufficient Experimental Evidence to Support the Claimed Representational Improvements: The experimental evaluation does not adequately isolate the benefits of recovering scale information from those arising from increased feature dimensionality and nonlinear feature expansion. In particular, the absence of ablation studies comparing individual feature-map components and simpler injective encodings makes it difficult to determine which properties of the proposed construction are responsible for the observed improvements. Additionally, the experiments rely exclusively on noiseless state-vector simulations, leaving the method's robustness under realistic quantum noise and finite-shot measurements unexplored. These limitations are particularly important because the paper's central contribution concerns encoding effectiveness and resource efficiency, yet the current validation does not sufficiently establish whether the proposed construction offers advantages beyond those achievable through simpler preprocessing strategies.

Questions for authors

Q1: The proposed feature map is designed to eliminate the scaling ambiguity introduced by amplitude normalization. However, a simpler construction, such as augmenting the input with a constant component, (\Phi(x)=(x,1)), appears to provide the same injectivity property while requiring a smaller feature space. Could the authors clarify whether their construction offers any additional theoretical guarantees or representational advantages over such simpler alternatives? A rigorous comparison would help establish whether the proposed combination of polynomial and trigonometric features is theoretically necessary and constitutes a sufficiently substantive methodological contribution.

Q2: The paper emphasizes that the proposed encoding preserves logarithmic qubit scaling despite increasing the feature dimensionality. However, qubit efficiency does not necessarily imply efficient quantum state preparation. Could the authors specify the gate complexity and circuit depth required to prepare the proposed amplitude-encoded states and compare these costs with conventional amplitude encoding? This clarification is important for assessing whether the claimed resource efficiency extends beyond qubit count and remains meaningful as input dimensionality increases.

Q3: The proposed method simultaneously restores scale information and introduces additional nonlinear features. Consequently, the reported improvements may arise from either mechanism or their combination. Could the authors provide controlled comparisons using simpler injective mappings, such as ((x,1)), and reduced feature maps, such as ((x,x^{\circ 2})) or ((x,\cos(x)))? Such comparisons would help determine whether the observed improvements are primarily attributable to resolving scale ambiguity or to the additional nonlinear feature expansion.

1 comment
AuthorsOpenReview ↗

Response to Reviewer bdo2

We thank the reviewer for the time devoted to evaluating our work and for the thoughtful, detailed feedback. Your comments have helped us clarify the contribution and strengthen the experimental and resource analyses. We have conducted a major revision of the manuscript, changes are highlighted in red.

We also note the ethics-review flag. As no specific ethical concern could be identified after careful studying of the Code of Ethics, we have added an ethics statement as a precaution. If the flag concerned disclosure of AI use, we have also added the AI-use statement, which we inadvertently omitted from the initial submission. We apologize for this oversight and would appreciate clarification if any further concern remains.

Reviewer question 1: Does the proposed construction offer representational advantages over (x, 1)?

We agree that (x, 1) already removes the normalization ambiguity with a smaller feature space. We do not claim a stronger injectivity guarantee for the full map. We have added the (x, 1) map and a strictly radial dataset to the experiment. (x, x^{circ2}, cos(x)) performs best on the Radial benchmark, suggesting it removes the scale ambiguity most effectively, whereas (x, 1) performs best on the real-world datasets. We have revised the Results and Conclusion accordingly, recommending (x, 1) as a lightweight practical default, especially in light of the gate complexity.

Reviewer question 2: What are the gate complexity and circuit depth of state preparation compared with conventional amplitude encoding?

We have added gate-count and circuit-depth estimates to the Appendix, referenced in Section 5 (lines 500–504):

“(...) gate counts and circuit depth roughly double with each additional dimension. Hence, (x, x^{circ2}, cos(x)) and (x, x^{circ2}, 1) induce substantially higher complexity than the lower dimensional encodings. (x, 1) produces substantial gains on all five datasets at only approximately twice the gate count and circuit depth of conventional amplitude encoding.”

The revised discussion makes the preparation overhead explicit and supports the practical recommendation of the simpler constant-anchor map.

Reviewer question 3: Can controlled comparisons distinguish scale recovery from nonlinear feature expansion?

We have expanded the ablation study to include (x, 1), (x, x^{circ2}), (x, cos(x)), and (x, x^{circ2}, 1), alongside the initially proposed map. We have also added a Radial benchmark whose labels depend exclusively on scale. Because constant appending introduces no input-dependent nonlinear coordinates before normalization, the improvement (x, 1) demonstrates on the Radial dataset and the other benchmarks suggests that the gains do not arise solely from the added squared and cosine features.

Reviewer_3u7HOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

The paper studies the loss of scale information in conventional amplitude encoding. Nonzero real inputs that differ only by a nonzero scalar produce states that are indistinguishable up to global phase, regardless of the downstream quantum circuit. The authors give a preprocessing condition that removes this ambiguity and propose the fixed map Φ(x) = (x, x², cos(x)). It adds no trainable parameters and retains logarithmic qubit scaling in the original feature dimension. Experiments use noiseless statevector simulation, 32 trainable quantum rotation parameters, and six-fold stratified cross-validation on Breast Cancer, Wine, Digits, and a synthetic dataset. Balanced accuracy improves over conventional amplitude encoding on all four tasks, although Digits remains well below the classical reference and its ROC-AUC and AUPR decline. (Section 3.2, Equation 17; Section 4; Table 1.)

Critical strengths
  1. The analysis identifies a restriction introduced before the variational circuit. Once two inputs have become the same encoded state, a more expressive downstream circuit cannot recover the lost distinction. This makes the limitation independent of the chosen ansatz. (Sections 1 and 3.1–3.2.)
  2. The remedy is explicit and requires no additional trainable weights. Table 1 reports substantial balanced-accuracy gains over raw amplitude encoding on Breast Cancer (0.6991 to 0.8834), Wine (0.7600 to 0.9320), and Synthetic (0.5996 to 0.7717), under a stated cross-validation protocol.
Critical weaknesses
  1. The distinctive contribution of the three-block construction is unclear without a constant-anchor baseline such as (x, 1). That lower-dimensional, non-trainable map follows from the same normalization argument and appears sufficient to preserve input identity. The paper therefore has not established why the squared and cosine blocks are needed.
  2. The encoder comparisons change input dimensions, qubit requirements, and PCA preprocessing. A shared count of 32 trainable parameters does not control those differences. State-preparation cost is unmeasured, and all experiments are noiseless. The Digits results also show that the empirical benefit depends on the metric. (Sections 4–5; Table 1.)
Questions for authors
  1. What does Φ(x) = (x, x², cos(x)) gain over the simpler (x, 1) map, including at zero input? Please compare against this constant-anchor construction. (Section 3.2; Equation 17.)
  2. Can component ablations and a scale-sensitive task distinguish recovery of scale information from the more general benefit of nonlinear feature expansion? (Sections 3.2–5.)
  3. Angle encoding uses PCA to at most eight features, whereas amplitude-based methods retain the original dimensions. Please report qubit counts by dataset and compare encoders under controlled information or quantum-resource budgets. (Section 4.)
  4. What are the state-preparation gate counts, depth, and preprocessing or inference costs? Tripling the feature dimension can matter even when qubit count remains logarithmic. Please limit finite-shot and noisy-execution claims to what has been tested. (Sections 3.2–4.)
  5. How is the abstract’s 21–44% accuracy-gain range calculated? Digits balanced accuracy changes from 0.2751 to 0.3100, while ROC-AUC falls from 0.7480 to 0.7268 and AUPR from 0.3239 to 0.3127. Please specify the metric and whether the reported gain is relative or absolute. (Abstract; Section 5; Table 1.)
2 comments
AuthorsOpenReview ↗

Response to Reviewer 3u7H

We thank the reviewer for the time invested into evaluating our work and for the constructive feedback. We have substantially revised the manuscript to address these points. Changes are highlighted in red in the submitted revision.

Reviewer question 1: What does Φ(x) = (x, x², cos(x)) gain over the simpler (x, 1) map, including at zero input?

We have added (x, 1) to the experimental comparison. Both the full map and constant appending remove normalization ambiguity and remain well-defined at zero. The full map performs best on the Radial benchmark, whereas (x, 1) performs best overall on the real-world datasets. We have revised the Results and Conclusion accordingly: (x, 1) becomes our lightweight practical recommendation, especially due to the lower gate complexity.

Reviewer question 2: Can component ablations and a scale-sensitive task distinguish recovery of scale information from the more general benefit of nonlinear feature expansion?

We have added component ablations and a Radial benchmark whose labels depend exclusively on scale. Conventional amplitude encoding achieves balanced accuracy of 51.62%, compared with 76.88 % for (x, 1). The same map significantly improves classification performance on all remaining benchmarks. Because it adds only a constant reference, these gains cannot be attributed to the nonlinear feature expansion.

Reviewer question 3: Please report qubit counts by dataset and compare encoders under controlled information or quantum-resource budgets.

We have added dataset-specific qubit counts to the Appendix, referenced in Section 4 (lines 379–384). Conventional amplitude encoding requires 4–6 qubits, while the three-block maps require 6–8 qubits. Angle encoding uses one qubit per feature, so PCA reduces its inputs to at most eight principal components to keep quantum-resource budgets comparable.


Reviewer question 4: What are the state-preparation gate counts, depth, and preprocessing or inference costs? Tripling the feature dimension can matter even when qubit count remains logarithmic. Please limit finite-shot and noisy-execution claims to what has been tested.

We have added gate-count and circuit-depth estimates to the Appendix, referenced in Section 5 (lines 500–504):



“(...) gate counts and circuit depth roughly double with each additional dimension. Hence, (x, x^{circ2}, cos(x)) and (x, x^{circ2}, 1) induce substantially higher complexity than the lower dimensional encodings. (x, 1) produces substantial gains on all five datasets at only approximately twice the gate count and circuit depth of conventional amplitude encoding.”

We have also replaced the sentence referring to finite-shot execution in Section 4 (line 388) with:

“This setup computes expectation values without finite-shot sampling and does not model hardware noise. Accordingly, our empirical comparisons characterize performance only under ideal state-vector simulation. Finite-shot and noisy executions were not evaluated, and no conclusions about their effects are drawn here.”

Reviewer question 5: How is the abstract’s 21–44% accuracy-gain range calculated? Digits balanced accuracy changes from 0.2751 to 0.3100, while ROC-AUC falls from 0.7480 to 0.7268 and AUPR from 0.3239 to 0.3127. Please specify the metric and whether the reported gain is relative or absolute.


We have replaced the ambiguous range in the Abstract with an average improvement of 18.4 percentage points in balanced accuracy. For each dataset, we subtract the average balanced accuracy of angle encoding and conventional amplitude encoding from that of the best-performing encoding, then average across the five datasets. This clarifies that the reported gain refers specifically to balanced accuracy and is expressed as an absolute difference, rather than a relative percentage change.

Thank you for reporting the results transparently.

Some issues have been resolved through a comparison with the radial benchmark (x, 1). However, this result changed the evaluation of the contribution. The rationale for the paper's core proposal has disappeared, and the remaining contribution is limited to the observation that "normalization erases scale information" and a simple solution of adding constants. Accordingly, the score is lowered.

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

The paper studies the scale information lost when amplitude encoding normalizes classical inputs. It proposes a fixed feature map containing the inputs, their squares, and their cosines. The map remove encoding collisions while retaining logarithmic qubit scaling. Toy examples illustrate the problem. Experiments compare the method with standard encoding and a classical network on four datasets. Results improve substantially on three datasets. Digits show higher balanced accuracy but lower ranking metrics than conventional amplitude encoding.

Critical strengths

The paper explained the information-loss problem clearly. This is a useful reminder that model capacity cannot repair an inadequate input map. The proposed preprocessing is easy to implement. Gains on three datasets seem substantial. The results support the practical point that encoding choices can dominate performance in small quantum models.

Critical weaknesses

The central idea needs a much more stronger baseline. Appending one constant coordinate, (\Phi(x)=(x,1)), already removes the normalization ambiguity. It also preserves logarithmic qubit scaling. The paper failed to explain why its more elaborate feature map is needed. This is not a minor missing experiment. It directly affects the novelty & interpretation of the contribution.

The learning claim seems also unclear. Squares and cosines add nonlinear features. Better performance could comes from those features, rather than from restoring scale. The current experiments did not separate these explanations. Nor did injective encoding establish that the chosen circuit can learn useful functions of the preserved information. For a top ML conference, I would want a clear mechanism and evidence that can voice it. At present, the paper identifies a simple limitation and shows that a richer input representation helps. That is not yet a sufficiently strong contribution in my view.

Questions for authors

My questions:

  • Can the proposed map outperform constant appending? Please compare ((x,1)), ((x,x^{\circ2},1)), and the full map under matched register sizes. This is the most important experiment in my view.
  • Which claim should the experiments establish: scale information matters, nonlinear features help, or this particular map suits the circuit? These are different claims. Please identify the main one and explain which result isolates it.
  • Can you test a controlled problem where labels may depend on scale while direction is uninformative? Then compare with a problem where scale is irrelevant. This would test the proposed explanation directly. Dont need a much larger benchmark suite to resolve this question.
1 comment
AuthorsOpenReview ↗

Response to Reviewer rNPa

We thank the reviewer for the careful assessment and constructive suggestions. Your comments prompted a major revision, particularly the addition of constant-appending baselines, a controlled radial benchmark, and a clearer distinction between our primary claim and our findings about individual feature maps. Below, we address each question and explain the corresponding revisions. Changes are highlighted in red in the revised manuscript.

Reviewer question 3: Can you test a controlled problem where labels may depend on scale while direction is uninformative?

We have added a synthetic Radial benchmark whose labels depend exclusively on input scale, while direction is uninformative. Conventional amplitude encoding performs close to chance, consistent with its removal of the information needed for this task. By contrast, all evaluated scale-preserving augmentations substantially improve performance. The full map (x, x^{circ2}, cos(x)) achieves the best performance, closely followed by (x, x^{circ2}, 1). Both also outperform the classical neural network on this benchmark. We have incorporated these results into the revised Results section and Conclusion.

Reviewer question 1: Can the proposed map outperform constant appending? Please compare (x, 1), (x, x^{circ2}, 1), and the full map under matched register sizes.

We agree that constant appending is an essential baseline: (x, 1) already removes the normalization ambiguity while retaining logarithmic qubit scaling. We have therefore added both (x, 1) and (x, x^{circ2}, 1) to the experimental comparison, including on the new Radial benchmark.

The results change the practical interpretation of our contribution. The full map (x, x^{circ2}, cos(x)) performs best on the Radial benchmark, with (x, x^{circ2}, 1) close behind. On the real-world datasets, however, (x, 1) performs best overall. Consequently, the experiments do not establish that the full map is necessary or generally preferable to constant appending. We have revised the Results section and Conclusion accordingly. The conclusion now identifies (x, 1) as a lightweight practical default and presents the full map's advantage on the Radial benchmark as a task-specific empirical finding.

Reviewer question 2: Which claim should the experiments establish: scale information matters, nonlinear features help, or this particular map suits the circuit? These are different claims. Please identify the main one and explain which result isolates it.

Our primary claim is that discarding scale information can impair classification when that information is relevant to the labels. The relative usefulness of particular feature maps is a secondary empirical question. We agree that this distinction was insufficiently clear in the original manuscript. We now state this explicitly in Section 3 (lines 356–359), Section 5 (lines 429–436), and the Conclusion. The Radial benchmark provides the clearest evidence for the primary claim: labels depend only on scale, conventional amplitude encoding removes that information, and constant appending restores access to it and substantially improves classification. The (x, 1) baseline introduces no additional input-dependent nonlinear coordinates before normalization and therefore demonstrates that adding squares or cosines is not solely responsible for the substantial performance improvements. Accordingly, gains on the real-world datasets are consistent with a benefit from preserving scale.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers and Summary of Revisions

We thank all three reviewers for their time, careful assessments, and constructive suggestions. Your feedback has greatly helped us strengthen the experimental evaluation, the results analysis, and the clarity of the discussion. We have undertaken a major revision to address the concerns raised.

The main additions are constant-appending baselines, expanded component ablations, a controlled strictly Radial benchmark, and qubit-count, gate-count, and circuit-depth estimates. We have also clarified the distinction between preserving scale information and the relative performance of individual feature maps. The revised Results and Conclusion recommend (x, 1) as a lightweight practical default, while recognizing the stronger performance of (x, x^{circ2}, cos(x)) on the Radial benchmark. We have corrected the abstract’s performance summary and explicitly limited our empirical conclusions to noiseless state-vector simulation. Furthermore, we’ve added AI use and ethics statements.

Our individual replies address each reviewer’s questions. We cite the revised manuscript where suitable to keep the responses self-contained. To further ease the task of re-evaluation, we use red font in the revised manuscript to highlight those parts that have been modified.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.