acceptodds
Under review as a conference paper at ICLR 2027

Spectral Anatomy of Quantum Gaussian Process Kernels

Abstract

Two recent results have reshaped quantum Gaussian processes (QGPs). On the one hand, Lowe et al. (2025) rule out the exponential speedups claimed by HHL-based QGP regression in the typical regime; on the other, an independent line of work shows that highly expressive quantum kernels suffer posterior pathologies that break Bayesian optimization. We show that these seemingly unrelated phenomena are governed by the same quantity: the normalized spectral entropy of the kernel Gram matrix. We prove a Cauchy–Schwarz bound on the best rank- spectral truncation error, a finite-sample variance-contraction bound with a high-SNR connection to Bach's degrees of freedom , and a characterization of the target-dependent optimal entropy via the intrinsic dimension of the target in the kernel eigenbasis. Empirically, the diagnostic is kernel-agnostic: hardware-efficient, matchgate, IQP and RBF/Mat\'ern/RFF/deep-kernel families all collapse onto identical curves on dequantization, ECE, and variance-contraction panels. The NLL sweet spot lives at high entropy for smooth targets and at low entropy for band-limited quantum-data targets. The diagnostic transfers from simulator to IBM Heron hardware with median absolute error and mean in across configurations at , with matchgate and IQP within mean and a single HE configuration returning a outlier that drops to on rerun (attributed to calibration drift); the same diagnostic transfers to a second Heron backend (mean error ) and to a scale-up on the original backend (mean error ). No error mitigation is applied throughout.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
quantum Gaussian processes, quantum kernel methods, kernel spectral analysis, posterior calibration, Bayesian optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_96ZkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes the normalized spectral entropy s(K)=S(K)/log⁡ns(K)=S(K)/\log n of the training Gram matrix as a single, label-free coordinate for quantum Gaussian-process kernels. It argues that this quantity governs two phenomena. The first is the dequantization result of Lowe et al. (2025), whose condition-number bounds rule out exponential HHL-type speedups. The second is the concentration of expressive quantum kernels. The experiments sweep circuit depth and encoding scale for hardware-efficient, matchgate and IQP circuits (nq=6,8n_q=6,8; n=30n=30) and compare them with classical kernels. They also run Bayesian optimization and estimate s(K)s(K) on IBM Heron devices (nq=4n_q=4, n=8n=8).

Critical strengths
  • The paper implements a broad empirical study (three quantum ansatz families and two non-modular ablations, five classical kernel families, synthetic, quantum-data and two real regression targets, downstream Bayesian optimization).

  • Plotted against spectral entropy, classical and quantum kernels fall on the same curves for dq, ECE and contraction. For these GP properties the spectrum matters, not the quantumness of the feature map. This is a useful corrective to quantum-specific narratives.

  • For broadband targets, the NLL optimum lies near the Haar end (s≈0.9s\approx0.9–0.990.99), so highly expressive kernels are not automatically useless. The useful window depends on the target. Theorem 2 (water-filling) gives a simple intuition for this. Lemmas 2–3 state the two degenerate regimes precisely, including the shrinkage bias and the failure in the noiseless case.

  • The spectral statistics are estimated on two IBM Heron devices without error mitigation, and the errors are small. Lemma 4 (depolarizing noise adds a rank-one component to the Gram matrix), together with shot noise, explains the direction of the observed bias.

Critical weaknesses
  • The theory is elementary and does not support the main story. By the authors' own account, Theorem 1 does not give the needed direction ("high entropy is hard to compress").

  • The unification claim is not substantiated. Lowe et al. study the condition number of KK, which depends on the smallest eigenvalues. Spectral entropy is insensitive to these eigenvalues, and no result in the paper connects the two quantities.

  • The score dq is the fraction of eigenvalues needed to carry 99% of the trace. Both dq and s(K)s(K) depend only on the normalized spectrum, so much of their joint behaviour follows from the definitions. Moreover, dq does not measure classical hardness.

  • The quantum-data target has variance about 15 times below the noise variance. Its low-entropy optimum thus reflects a nearly constant target.

  • As a predictor of NLL, s(K)s(K) is weaker than existing statistics. Its absolute Spearman correlation is 0.28, against 0.89 for off-diagonal concentration and 0.79 for kernel–target alignment.

  • The experiments are small and the claims are hard to identify. The sweeps use n=30n=30 training points, the hardware runs use an 8×88\times8 Gram matrix, and most configurations have a single seed.

Questions for authors
  1. How is s(K)s(K) related to κ(K)\kappa(K), the quantity analyzed by Lowe et al.?
  2. Why call dq a measure of dequantization when classically simulable circuits lie on the same curve?
  3. Is it possible to exhibit the low-entropy optimum on a band-limited target whose signal is well above the noise level?
  4. Off-diagonal concentration and kernel–target alignment predict NLL better. Which practical decision does s(K)s(K) improve?
Reviewer_ADcWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the normalized spectral entropy S(K)/log n of the Gram matrix as one diagnostic for quantum GP kernels. The claim is that this number tells you whether the kernel can be dequantized with classical features, how well calibrated the posterior is, and whether the posterior variance contracts at all. On the theory side there is a Cauchy–Schwarz tail bound (Thm 1), a finite-sample bound on variance contraction (Prop 1), and Thm 2, which gives the NLL-optimal spectrum for a fixed eigenbasis. Experiments use three circuit families and some classical kernels, with a check on IBM Heron hardware and a small GP-BO study.

Critical strengths

The most critical strength is Figure 2. Once plotted against S/log n, the hardware-efficient, matchgate and IQP kernels fall on the same curves. That is what would make the diagnostic practical: you only need one Gram matrix, whatever the circuit family. Second, on hardware they kept the one bad HE point (|ΔS| = 0.30) and reran it instead of dropping it. Appendix A also shows that the main sweeps use a single seed and that some BO gaps are small.

Critical weaknesses
  1. The theory does not support the claim that dequantization and posterior pathologies are "governed by the same quantity" in the abstract. Theorem 1 is only a sufficient condition, and the authors note themselves that entropy alone does not give a useful tail bound. Proposition 1 is stated in terms of the cross-covariance K_* and does not reduce to the entropy of K. Theorem 2 fixes the eigenbasis and optimizes the training marginal likelihood. The experiments, however, change the eigenbasis by changing the circuit, and they compare held-out NLL. So what ties the three phenomena together is an empirical correlation across these circuit families. The body of the paper mostly says this (e.g. p. 8), but the abstract and C1 present it as a theoretical result.

  2. The target-dependent part of the story is weaker than what is described in the abstract. On the quantum-data target the signal variance (≈0.0034) is 15× below the noise σ² = 0.05 (p. 7). With almost no signal, shrinking to the prior wins, and that is mainly why low entropy is optimal there. It does not show that band-limited targets prefer low entropy. Algorithm 1 uses fixed 0.1/0.95 cutoffs, but in Appendix L the optimum moves from 0.91 to 0.79 as n goes from 30 to 100. In the BO study, Q-SWEET leads on the quantum target by about one standard error (−0.104 vs. −0.092), which the authors acknowledge.

Questions for authors
  1. Could you repeat the quantum-data experiment with a target whose signal is well above the noise floor, or with a smaller σ²? If low entropy still wins, that supports the band-limited claim. If not, I think the claim should come out of the abstract.
  2. Given the shift in Appendix L, how should the 0.1/0.95 cutoffs scale with n and the number of qubits?
Reviewer_PTcmOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies quantum Gaussian process to identify when quantum Gaussian process can provide non-trivial posterior updates. Two edge cases where quantum Gaussian process would provide trivial posterior updates are considered, followed by some theoretical results that try to identify when the quantum Gaussian process's kernel would avoid such trivial updates.

Critical strengths

It provides a lower bound on the variance contraction -- posterior-to-prior ratio averaged at test points -- which can provide a partial intuition when the posterior would become trivial. Specifically, if the lower bound is near 1, then it means that the posterior is more or less similar with the prior (on average over the test points) hence one may avoid using such kernels. However, such lower bound requires knowledge on the test inputs (as it requires cross-covariance between training and test inputs) hence its practical usage might not be so significant as I just mentioned now.

Critical weaknesses

Among others, my main concern is: Lowe et al. (2025) proved that quantum Gaussian processes does not provide any meaningful speedups whenever the corresponding kernel is square integrable and has a bounded diagonal; given this, what is the significance of considering quantum Gaussian process in this work? If the considered quantum Gaussian process does not belong to the setting studied by Lowe et al. (2025), e.g., unbounded kernel, then I can understand the significance of this work as there is still a reason to consider quantum Gaussian processes. If not, then I'm not sure what is the point of this work unfortunately. But this is not very clear in the current context and it looks to me that the authors' kernel also lies in the assumption studied by Lowe et al. (2025).

Now suppose that the authors' setting does not belong to the case that had been ruled-out by Lowe et al. (2025), i.e., the authors are studying the kernels that may provide meaningful quantum speedups. Now the authors are trying to show in which regime of kernels the posterior would provide non-trivial results. And the authors' flow is: (i) show the two edge cases where posterior becomes trivial; (ii) show that "participation ratio" is minimum in the first edge case; is maximum in the second edge case; (iii) thereby argued that the normalized Shannon entropy should lie strictly between 0 and 1 (that does not belong to the edge cases); (iv) also provided a lower bound on the variance contraction.

Although step (i) is interesting, I'm not sure how we can have (iii). That is, it is true that edge cases belong to Shannon entropy being 0 and 1, but it does not say that "all Shannon entropy being 0 and 1 should correspond to the edge cases hence will provide trivial posterior". Furthermore, (iv) provides a lower bound on the variance contraction but it does not provide any upper bounds so its usage is limited; also due to its reliance on the true cross-covariance term.

All in all, this work seems to contain multiple interesting points but their significance -- how the theoretical results answer the question of "what kind of kernel can provide meaningful quantum Gaussian process" -- is not clear in its current form.

Questions for authors

If the authors can clarify:

(i) the authors' kernel setting (Def. 1) that the authors' theory builds upon corresponds to the case where quantum Gaussian process can provide meaningful speedups (i.e., the setting that not ruled-out by Lowe et al. (2025))

(ii) provide a clear regime of kernel where the posterior becomes provably non-trivial (or provide a regime where the posterior becomes provably trivial)

(iii) how Proposition 1 can be practically useful (e.g., what the lower bound tells us?)

(iv) which part of this paper shows quantum advantage? I.e., what is the benefit of using quantum Gaussian process compared to classical kernels? Which part of analysis requires quantum assumption? (basically, where did the authors utilized the parametrized unitary in Def 1 for the theoretical analysis?) Without these, I'm not sure what is the significance of "quantum" in this work -- it is rather a work on classical gaussian process that studies the impact of generic kernels.

then I would be happy to adjust my assessments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.