acceptodds
Withdrawn from ICLR 2027

When Does a Quantum Interaction Matter? Observability and Symmetry in Hybrid Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A quantum circuit can change its state under an input transformation while a chosen readout remains unchanged. We study this distinction in hybrid quantum neural networks for surrogate AC optimal power flow. For a real two-layer, non-reuploading ansatz with data encoding, we prove that single-qubit and readouts depend only on the coordinate-wise cosines of the encoded angles, whereas carries an odd component. We then use the matched readout pair versus to separate parity information from feature-count and encoder effects. With a frozen encoder, improves test MAE over on all three grid systems. With end-to-end training, the advantage is not consistent and depends on the grid and architecture. A target-aware audit further shows that the learned latent geometry changes around sign-mirror neighborhoods, but does not identify a unique optimization mechanism. The results establish a practical attribution protocol: state dependence, observable invariance, and predictive utility must be tested separately. We make no claim of quantum advantage.

Primary area:
learning theory
Keywords:
Quantum Machine Learning, Data Re-uploading, Quantum Circuit Observability, Representation Symmetry, Fourier Expressivity, Hybrid Quantum-Classical Learning, AC Optimal Power Flow

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_kJPROpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents a careful attribution study of measurement-induced information loss in hybrid quantum models. Its central distinction is to state dependence, and predictive utility is valuable, and the matched Z+X versus Z+Y intervention is well designed. The authors appropriately report a conditional rather than universal result: Y helps when the encoder is frozen but not consistently during end-to-end training.

Critical strengths

The paper clearly separates quantum-state sensitivity from readout accessibility and downstream predictive value. Then, Z+X versus Z+Y holds feature count and decoder parameterization fixed, giving a cleaner attribution test than simply enlarging the readout.

Unfortunately, I cannot say more positive encouragement.

Critical weaknesses

The manuscript is incorrectly titled “Formatting Instructions for ICLR 2027 Conference Submissions,” which must be corrected.

The theoretical result is architecture-specific and elementary: coordinate-wise parity is proved only for one two-layer, non-reuploading gate ordering. Deeper and reuploading cases lack theory.

The only consistent YY benefit occurs in the frozen-encoder control, whose MAE is four to five times worse than end-to-end training. Thus, the demonstrated mechanism is not a practical bottleneck in the primary setting.

Again, in the current form, the paper is very hard to parse.

Questions for authors

Can Proposition 2 be generalized to deeper or reuploading circuits, or can counterexamples delimit its scope?

Do results persist across resampled splits and larger AC-OPF datasets?

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

The authors consider a very specific instance of a quantum circuit and attempt to characterize certain invariances in the readout for AC optimal power flow. They provide some theoretical results and conduct experiments to support their statements.

Critical strengths

The authors tackle the task of understanding the geoemtry of variational circuits, which is an interesting endavour.

Critical weaknesses

The scope of the study is incredibly narrow. The circuit that the theoretical results cover is limited to a very specific instance, that, to the best of my knowledge, is not commonly used in practice (otherwise, I would ask the authors to provide references). It is unclear how it can be expanded to more general or more complex circuits that are of practical relevance. It would be helpful if the authors commented on this to understand the broader relevance or impact of the work, since, as the manuscript is now, I do not think it is provides sufficiently novel and interesting insights for the community to be published at this venue.

For the experimental setup, it is unclear what systems the paper refers to and there is also no appendix to clarify it. It is unclear to me which experiments are conducted, which systems are referenced and even what the different models are. I am afraid I am not able to provide further or more concrete questions/criticism/guidance, as I am simply not able to follow the manuscript and experimental setup.

Beyond that, the manuscript has several structural and formatting issues. Firstly, the authors did not insert the title of the paper into the template. Secondly, it is missing a proper background section, which is standard for an academic paper. As a minor comment, also the placement of the proofs of the proposition is very uncommon and do not help in improving the reading flow. I found the manuscript as-is quite difficult to follow.

Besides being mandatory, the paper does not include an AI statement.

Questions for authors
  • It is evident that a hybrid ML model does not have access to the full quantum state, as well as there being certain operations that the expectation value is simply invariant to. In the introduction you state that this distinction matters in quantum surrogates for AC optimal power flow. Why would that be the case for exactly this type of application?
  • Can the authors clarify if their results are applicable to more commonly used circuit architectures and feature encoding methods or how their results could be extended? A common hardware efficient ansatz usually employs different rotation gates.
  • I would also assume that the invariances get more subtle once quantum circuits that produce more "interesting" states are employed. Could the authors comment on this?
Reviewer_NHYUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work shows that X/Z measurements discard encoded sign information in a specific hybrid QNN, while Y measurements preserve some of it. Authors compare equally sized Z+X and Z+Y readouts for power grid prediction and test both frozen and trainable encoders. Adding Y helps with frozen encoders but gives no consistent end-to-end improvement.

Critical strengths

I don't see any strengths in this manuscript

Critical weaknesses
  • The proof of proposition 2 is wrong where it claims CNOT preserves all Pauli strings without Y. That is wrong. Proposition 2 holds only because the specified circuit propagate through a narrower family of pure X or pure Z strings, which avoids the counterexample. So the proof needs narrower argument.
  • Theoretical contribution is too narrow to be impactful. The results establish a sign symmetry for a restricted circuit and the straightforward fact that a decoder cannot recover information absent from its inputs. They provide no general explanation of when a learned encoder compensates for this restriction or how much prediction error the restriction causes.
  • The proposed mechanism is also not convincingly isolated. Replacing X with Y changes feature correlations, conditioning, and gradients as well as sign sensitivity. Matching feature counts and freezing the encoder therefore do not establish that recovering signs specifically causes the observed improvement.
  • Table 1 and 2 show no improvement under normal end-to-end training. The Z+Y readout is no better than the Z+X in the end-to-end comparison. Gain exists only for deliberately frozen encoder, hence the practicality of this method is not shown.
  • The experiments are trivial. Too small, only four simulated qubits and three datasets with 30 test examples each, all on fixed splits. That there are no baselines is very concerning as well.
  • This method is only shown for very restricted specific QNN. I can't see any practicality.
  • What is "target-aware audit"? No definition nor explanation.
  • Also, there is a big leap. Failure to observe a consistent benefit from YY does not prove that sign loss is no longer a prediction bottleneck. Optimization difficulties or limited statistical power could conceal a real representational disadvantage.
  • Numerical reporting in the tables are also inconsistent.
Questions for authors

Do the authors think this is a complete research? I feel the results and contributions are nowhere near impactful.

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

Related papers

Discussion (0)

Sign in to comment.