acceptodds
Withdrawn from ICLR 2027

Geometry Matters: Kernel Concentration in Quantum Feature Maps

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantum machine learning (QML) is motivated by the premise that quantum circuits can embed classical data into feature Hilbert spaces in which learning may become more effective. We evaluate this premise empirically by comparing two feature maps, the entangling ZZFeatureMap and simple angle encoding, on multiclass MNIST classification. Using four kernel geometry metrics (kernel target alignment, intra/inter similarity ratio, effective rank, and off-diagonal variance), we show that ZZFeatureMap suffers severe kernel concentration as task complexity increases: its similarity matrix becomes nearly uniform, its effective rank exceeds 300, and its alignment with the labels collapses. In contrast, angle encoding preserves a structured, low-rank kernel and achieves 83% accuracy on 8-class, 8-qubit MNIST, outperforming ZZFeatureMap by 42 percentage points. We find that kernel target alignment and off-diagonal variance are reliable predictors of classification performance, while effective rank reveals when the kernel becomes diffusely concentrated. These metrics offer practical diagnostics for selecting feature maps in near-term quantum kernels. Overall, the results indicate that practical QML performance depends less on generic circuit expressivity than on whether the chosen feature map induces a useful geometry for the data.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Quantum Machine Learning, Quantum Kernels, Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_WpV6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper compares angle encoding and ZZFeatureMap as QSVC kernels on PCA-reduced MNIST. It uses four kernel diagnostics to argue that ZZFeatureMap concentrates and that the geometry explains the accuracy gap. However, the evidence supports a narrower conclusion than the manuscript claims.

Critical strengths
  1. Holding qubit count and single-qubit structure fixed is a sensible design for isolating the entangling step.
  2. Reporting kernel geometry alongside accuracy is useful, and the 42-point gap is arithmetically correct.
Critical weaknesses
  1. The entanglement claim is not established. Comparing Ry encoding with ZZFeatureMap does not isolate entanglement. An ablation removing ZZ interactions while retaining the remaining circuit is needed.
  2. Geometry is measured for one seed, with no correlation analysis or independent validation. Table A1 even shows lower KTA despite higher accuracy for 4-class angle encoding when moving from four to eight qubits.
  3. Near-zero off-diagonal similarities imply an approximately identity like kernel, rather than a uniformly constant matrix. High effective rank or high off-diagonal variance alone does not establish poor or useful class structure.
  4. The evaluation is too limited. Add tuned classical SVM baselines, variability across seeds, and bandwidth/depth ablations. One dataset and two qubit counts provide limited novelty beyond the concentration phenomenon discussed in the paper.
Questions for authors
  1. Which exact ZZFeatureMap circuit is used in the work? Also, what scaling range is used to map PCA features to rotation angles? Since concentration in ZZ-type maps depends strongly on these choices, the answer determines whether the central finding holds for the map itself or only for this particular configuration.
  2. Do the authors test whether the ZZ kernel’s concentration persists across a range of feature scaling factors? If it disappears at a smaller scale, the comparison with angle encoding would need to be redone with matched scaling.
  3. The ZZ heatmaps show a bright diagonal over a dark background. Do the KTA and off-diagonal variance values change when the diagonal is excluded? Since the diagonal is always 1, it can inflate KTA and affect the effective rank calculation.
  4. How does an RBF or polynomial SVM on the same PCA features perform on these tasks? This comparison is needed to judge whether either quantum kernel offers any practical value, and it bears directly on the paper’s claims about usefulness.
  5. The conclusion says the results match Huang et al.’s prediction that adding qubits hurts performance. As written, angle encoding improves with more qubits, so the attribution needs support.
Reviewer_RZbsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper provides a comparison of two data-embedding techniques (namely angle encoding and ZZFeatureMap) for quantum kernel methods and empirically shows four metrics (namely kernel target alignment, intra/inter-class similarity ratio, effective rank, and off-diagonal variance) to predict the classification accuracy of the 8-class MNIST dataset, classically compressed from 784 to four or eight features using PCA.

Critical strengths

The paper focuses on the important topic of finding predictors that allow for reliably assessing the potential advantages of using different quantum routines and embeddings to improve current machine learning approaches. It goes beyond typical performance measures and looks for indicative metrics, analyzing the resulting kernel matrices themselves to provide insights into the implications of different data-embedding approaches.

Critical weaknesses

Most crucially, I am very skeptical of the authors' choice to use a hybrid setup that hinges on PCA to compress high-dimensional image data into low-dimensional feature representations. With this setup, it's hard to assess what in the results is attributable to the quantum embedding itself and what is driven by the preprocessing. Using lower-dimensional data, or embedding the non-preprocessed data to allow a non-hybrid setup, would preserve the meaningfulness of the presented results.

My next concern regards the scope of the paper. As the authors note, the potential issue of concentration for quantum kernels is not new. Neither are embedding mechanisms analyzed. Yet, both the paper and the references connect to a wide body of related prior work only very limitedly. While the paper provides four metrics and empirically shows their insightfulness for high-dimensional image data (though classically compressed to low-dimensional features), neither metric is novel. Overall, the study is limited to a single dataset and two embedding approaches and provides no classical baselines for performance comparison. Extending the study along any axis (either to different types of data, broadening the range of embedding approaches analyzed, or the comparison to classical) would have improved the paper's overall scope.

Finally, even though acknowledged by the authors, it should be noted that the paper only considers noise-free simulated results and does not provide any discussion or results on how these translate to current noisy hardware.

Questions for authors
  • Why did the authors opt for a high-dimensional image dataset when this introduces the need for advanced classical preprocessing that could impact the results in a meaningful way, rather than using lower-dimensional data that would have allowed for direct embedding?
  • How do the authors expect their results to generalize to other embedding mechanisms, data domains and -dimensionalities, and current noisy quantum hardware?
Reviewer_oiXuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper compares two quantum feature maps (ZZFeatureMap and angle encoding) to assess how their induced geometry affects training performance on quantum learning algorithms. To this end, the paper presents some empirical results pertaining to these embeddings.

Critical strengths

None

Critical weaknesses

Overall, the paper offers no new technical results and the empirical results have been general knowledge in the field for some time. Therefore, I don't think this paper contributes meaningfully to the literature.

In detail

The paper offers no new insight or direction for research. The ideas discussed in the paper have been known for quite a while and there is a rich body of literature concerning this topic. As per the Reviewer Guide, here are some examples from peer-reviewed literature from on or before July 17th 2026:

I stress that there are quite a few other works, some of which are of course only on arxiv to my knowledge, such as

This next paper is from Oct 2026 so the authors are not expected to know about it but I want to include it as a reference to illustrate the current research being done in this area. It also cites a good handful of relevant literature.

To add to this, there is a complete lack of proper citations throughout the text. Some examples just from the first few paragraphs:

  • "Quantum machine learning (QML) is motivated by the possibility that quantum circuits can embed classical data into high-dimensional Hilbert spaces in which learning may become more effective. [CITATION NEEDED]"

  • "The usefulness of a QML model therefore depends not only on the size of the Hilbert space but also on whether the chosen quantum feature map induces a geometry that is meaningful for the task.[CITATION NEEDED]"

  • "Kernel concentration is a known failure mode for expressive quantum feature maps: as circuit depth or qubit count grows, pairwise similarities between data points collapse towards a single value, and the kernel loses the ability to distinguish between classes.[CITATION NEEDED]"

The introduction and related works sections fail to properly acknowledge the body of relevant literature.

Based on these significant shortcomings I recommend the paper be rejected.

Also, Figure 3 has a clipped title

Questions for authors

None

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

In this paper, the authors empirically compared two feature maps: the entangling ZZFeatureMap and simple angle encoding on multiclass MNIST classification, using four kernel geometry metrics (kernel target alignment, intra/inter similarity ratio, effective rank, and off-diagonal variance). They observed that ZZFeatureMap suffers severe kernel concentration as task complexity increases, but angle encoding preserves a structured, low-rank kernel and achieves high accuracy, outperforming ZZFeatureMap. They also found that kernel target alignment and off-diagonal variance are reliable predictors of classification performance, while effective rank reveals when the kernel becomes diffusely concentrated. Finally, they concliuded that practical QML performance depends less on generic circuit expressivity than on whether the chosen feature map induces a useful geometry for the data.

Critical strengths
  1. The paper is addressing an important problem.
  2. The observed results are interesting.
  3. The paper is easy to follow.
Critical weaknesses

My first concern is about the novelty of this work. There are existing work discussing the relationship between quantum kernel properties and learning performance. For example, Kübler et al. (NeurIPS 2021) studied the spectral properties and inductive bias of quantum kernels, and showed that the ability of a quantum kernel to learn and generalise depends on how well its function space matches the target problem. They also investigated kernel target alignment and showed that highly expressive quantum kernels can have poor generalisation under certain conditions. Moremore, Thanasilp et al. (Nature Communications 2024) investigated exponential concentration in quantum kernel methods. They identified several factors that can cause kernel concentration, including the expressivity of data embedding, entanglement, global measurements, and noise. Their work also included numerical experiments using MNIST data and investigated how kernel concentration changes with the number of qubits and circuit depth.

Kübler, J. M., Buchholz, S., & Schölkopf, B. (2021). The inductive bias of quantum kernels. Advances in Neural Information Processing Systems (NeurIPS 2021).

Thanasilp, S., Wang, S., Cerezo, M., & Holmes, Z. (2024). Exponential concentration in quantum kernel methods. Nature Communications

In this case, I think that the main contribution of this paper is the empirical comparison between ZZFeatureMap and angle encoding using several established kernel metrics. However, while the experimental results are interesting, I am not fully convinced that they provide sufficiently new insights beyond the existing literature. I suggest the authors to explain more clearly what new understanding can be obtained from their investigation, especially when comparing to related studies.

Another concern is about the experimental design. Between ZZFeatureMap and angle encoding, there are lots of differences, which include the entangling structure of ZZFeatureMap for sure, but also include other factors, e.g. their single-qubit operations, feature encoding functions, circuit depth, and entangling gates. So, it is hard to fully conclude that the observed difference between ZZFeatureMap and angle encoding is because of the entangling structure of ZZFeatureMap. A more controlled comparison should be done.

In addition, it is not clear how PCA features are scaled before being encoded into the quantum circuits. This might be important as this might produce different kernel geometries. Does this influence the following ZZFeatureMap experiments?

Finally, I think some classical machine learning baseline should be included, which would be used to show the relative performance of these quantum methods.

Questions for authors

Please see above in Critical weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.