acceptodds
Withdrawn from ICLR 2027

When Geometry Meets Energy: Robust Operator Selection for Adaptive VQE via GAGrad

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Adaptive variational quantum eigensolvers (VQE) grow quantum circuits by iteratively selecting operators from a predefined pool. The standard energy-gradient criterion (ADAPT-VQE) aligns directly with the optimization objective but can mislead when the reference state deviates from Hartree-Fock. Geometry-driven alternatives based on the quantum geometric tensor (QGT) offer complementary strengths, prioritizing structurally non-redundant operators and maintaining robustness to reference-state perturbation. We propose the GAGrad operator selection framework, which combines energy-gradient magnitudes with geometric volume-expansion criteria measured via the Schur complement of the QGT. Across molecular benchmarks (LiH, BeH; 6–10 qubits), we uncover a gradient failure phenomenon: under moderate perturbation, gradient and Schur-complement rankings become anti-correlated (), and ADAPT and GAG diverge only in the final 1–2 operator selections, yet this late-stage difference produces 3–4 orders of magnitude gap in residual energy. Thus, geometric information is most consequential near convergence, when energetic differences between candidates become small and gradient rankings become unreliable. GAGrad matches ADAPT's step efficiency while capturing this geometric advantage, achieving comparable precision on most benchmarks with shallower circuits and stable performance under depolarizing noise. A GAG warmstart protocol further reduces circuit depth substantially.

Primary area:
optimization
Keywords:
Variational Quantum Eigensolver, Quantum Machine Learning, Quantum Geometric Tensor

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_iwpEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the selection of new operators in adaptive VQE. The authors employ the Schur complement (SkS_k) of the QGT to quantify the geometric novelty of candidate operators. They experimentally compare ADAPT, the purely geometric selection method GAG, GAGrad, and random selection across four qubit configurations for LiH and BeH₂. The authors argue that perturbations in the reference state can lead to "gradient failure," and that geometric information improves operator selection near convergence while enhancing robustness against noise.

Critical strengths

Energy decrease and tangent space expansion are distinct selection signals; combining them is of research value.

Critical weaknesses

There is a logical error in the key reasoning of Theorem 1. Page 5 assumes that the gradient ranking degrades and approaches a random state under perturbation, and the proof on page 14 asserts that the correlation between gradient and geometric rankings must pass through negative values. However: correlation gradually approaches zero does not mean correlation necessarily becomes negative.

The experiments have not established any meaningful failure or robustness, and there are clear discrepancies in the reported results; the absolute error associated with "gradient failure" is very small. All these values ​​are far below the chemical accuracy of 1.6x10^-3 Ha adopted in the paper.

The table on page 8 clearly extends beyond the layout boundaries; while the main text on page 8 states that the system ran on ibm_boston, the note for Table 10 on page 19 reads: Values ​​are estimated from calibration data; actual results may vary.

Estimates derived from calibration data cannot serve as direct evidence of actual device measurements. Until this is clarified, the paper's conclusion regarding "real hardware verification" cannot be accepted.

Questions for authors

Regarding Theorem 1, why is the emergence of a negative correlation interval inevitable? Please explicitly rule out the hypothesis that "correlation remains non-negative and tends toward zero," and explain why fixed utility and fixed geometric ranking cause GAG ​​to fail under strong perturbations.

How can one distinguish between gradient failure and residuals caused by optimization tolerance?

Where exactly is the noise in Table 1 applied?

Do the results in Table 10 represent empirical measurements, noise simulations, or calibrated estimates?

Why is there no comparison with directly relevant geometric baselines?

Reviewer_5VyAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GAGrad, an operator-selection method for adaptive VQE that combines energy gradients with the geometry of quantum state space. Whereas ADAPT-VQE selects operators by their energy-gradient magnitude, GAGrad uses the Schur complement of the quantum geometric tensor (QGT) to measure the size of a candidate’s new direction beyond the current circuit’s tangent space and combines this with the gradient magnitude. The authors evaluate energy errors, circuit sizes, and computational costs on 6–10-qubit LiH and BeH₂ systems, including experiments with perturbed reference states.

Critical strengths

I find it interesting to consider both whether an operator can help lower the energy and whether it provides a direction not already represented by the circuit. Comparing gradient-based selection, geometry-based selection, their combination, and random selection within the same adaptive pipeline helps clarify the role of each criterion. On some benchmarks, GAGrad reduces CNOT counts while retaining the same number of growth steps as ADAPT, suggesting a benefit from combining the two signals.

Critical weaknesses

The practical benefit over ADAPT remains unclear when the additional computational cost is taken into account. In the BeH₂ 8-qubit experiment in Table 7, GAGrad(mult) reduces the CNOT count from 94 to 86, but increases the total number of circuit evaluations from 1,623 to 4,402, approximately a 2.7-fold increase. The reported 3–4 orders of magnitude difference in residual error should also be considered in terms of absolute accuracy. For example, ADAPT’s error in the perturbed LiH experiment in Table 7 is 2.5×10−9 Eh2.5\times10^{-9}\,E_h, already far below the paper’s chemical-accuracy threshold of 1.6×10−3 Eh1.6\times10^{-3}\,E_h. It would help to explain when the smaller circuits and improvements at such high precision justify the additional measurement cost.

Questions for authors

How do ADAPT and GAGrad compare in the total measurement cost required to reach the same target accuracy?

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

This paper investigates the effect of the ordering of training task on memory. Specifically, it permutes identical task sets, measures the resulting differences in frozen-bank evaluation accuracy, and compares order selection with write-time memory governance. The authors proposed an ordering layer for this purpose and observed order sensitivity. The authors demonstrate that with the correct order, the task accuracy can be improved.

Critical strengths
  1. The experiments and ablations are well-designed. By holding task content constant and varying its permutation, the paper effectively isolates order effects from differences in training data. The authors also report both the best-worst spread and independently rerun selection gains to distinguish the performance gains. Moreover, the authors include no memory condition, the governed/ungoverned banks, writer specific selection, different backbones and the larger pool experiment.

  2. The proposed idea about training task re-ordering the novel. The experiments result also support the claims. The authors show in Table 4 that an order selected for one writer can hurt another, whereas selecting separately for each writer generally helps. This supports the practical conclusion that a beneficial order should not be assumed to transfer between memory implementations.

Critical weaknesses
  1. The proposed method selects the best order using scores on the same evaluation tasks used for final reporting and only the inference runs differ. The results show that the selected order remains effective across reruns, but not that it generalizes to new tasks. The fresh-pool experiments repeat the same selection procedure within each pool and therefore do not resolve this limitation. Selecting orders on a validation set and evaluating the resulting banks on a separate test set would clarify the broader value of the method.

  2. The pilot requires 36 complete training-and-evaluation runs. Although the authors propose amortizing it over repeated runs of the same workload, it is unclear why an operator would rebuild the memory bank each time instead of reusing a successful pilot bank or retaining verified solutions. The paper would benefit from a concrete deployment scenario and an end-to-end cost comparison showing when order selection becomes worthwhile.

Questions for authors
  1. Could you select the order using a validation set, then evaluate the frozen memory bank on a separate test set whose scores were never used for selection? Even one representative experiment would help establish whether the gains extend to unseen tasks.、

  2. In the intended deployment setting, why would the agent rebuild its bank for repeated runs of the same workload rather than reuse a successful pilot bank or cache verified solutions? Comparing these alternatives in accuracy and total cost, including the pilot will help clarify the deployment feasibility and value.

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

Related papers

Discussion (0)

Sign in to comment.