acceptodds
Under review as a conference paper at ICLR 2027

Learning Symmetry-Preserving Sampling Circuits for Sample-Based Quantum Diagonalization

Abstract

Learning quantum sampling circuits requires searching discrete architectures under physical constraints and a downstream scientific objective. We study this problem in sample-based quantum diagonalization (SQD), which estimates molecular ground-state energies by classically diagonalizing a Hamiltonian in a subspace sampled from a quantum circuit. A widely used sampler is the local unitary cluster Jastrow (LUCJ) ansatz parameterized from classical coupled-cluster amplitudes, which can become unreliable under strong correlation. We formulate sampler construction as a sequential decision process and introduce a neuro-symbolic reinforcement-learning approach that combines an actor–critic policy with a particle-number-preserving gate vocabulary and symbolic construction rules. A terminal reward based on the SQD energy and circuit length directly guides the search toward accurate, compact samplers. Across the tested molecular geometries of N₂, H₂O, and LiH, shallow searched circuits recover active-space ground-state energies in simulation. In the larger N₂ (14-electron, 10-orbital) active space, a searched circuit achieves chemical accuracy with 48 transpiled CZ gates, compared with 31,260 in the automatically constructed LUCJ baseline. Ablations attribute the subspace-coverage advantage to the symmetry-preserving search space, while the learned policy improves search consistency. Executions on IBM Heron-class hardware retain 65–74% of measurements in the valid particle-number sector at equilibrium geometries. The studied active spaces remain small enough that recovered energies do not discriminate subspace quality after configuration recovery. These results establish circuit efficiency and hardware feasibility for learned SQD samplers, while leaving their advantage in subspace quality at larger scales an open question.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Reinforcement learning, neuro-symbolic learning, quantum circuit architecture search, sample-based quantum diagonalization, symmetry-preserving circuits, quantum chemistry

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jH8tOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the design of sampling circuits for SQDs, employing particle-number-preserving fixed-angle Givens rotations, predefined "bricks," and "stop actions." An actor-critic approach is used to search for gate sequences, with the final SQD energy and the number of operations serving as the reward.

Critical strengths

There is a sound rationale for evaluating circuit architectures using SQD results rather than simply optimizing VQE expectation values. The set of particle-number-preserving gates also facilitates the interpretation of the circuit's physical constraints.

Table 2 reports the hardware results, and Table 5 lists the corresponding job identifiers; the authors also explicitly acknowledge limitations regarding small-space saturation, classical recovery, single-spin-sector sampling, and stretched geometry. Seed-specific ablation data are available for verification.

Critical weaknesses

In the ablation study on pages 7–8, both the learning strategy and the unguided search yielded circuits with 24 logic gates and 48 CZ gates, all reaching the maximum depth limit. Consequently, this set of experiments did not demonstrate a reduction in gate count through learning; the length penalty remained constant across these runs.

Page 9 explicitly acknowledges that the maximum spatial dimension is only 14,400, comparable to the sample budget of 30,000, and that uniform sector sampling can also recover the FCI result after configuration recovery. While the authors have recognized this issue, it implies that the near-FCI energies reported in Table 1 do not prove that a more valuable subspace was learned.

Questions for authors

When both sets achieve chemical accuracy and share the same length, can learning reduce the number of reward evaluations needed to reach the target or improve the success rate on difficult instances?

How do single-brick, fixed-brick (with limited rotation), random, greedy, and beam search strategies compare under the same budget? The existing ADAPT-QSCI method, which explores the adaptive construction of QSCI input states, should also serve as a relevant baseline.

What are the training time, number of samples, number of SQD solver calls, and validation costs per molecule and per geometry? How much subsequent usage is required to amortize the search costs?

What specific cost does the length-based reward optimize? A single "brick" may involve multiple rotations yet counts as only one action, while all episodes in the large search space reach the 24-step limit. Does the compactness stem from the reward function, the depth limit, or the hand-crafted action set?

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

The paper uses reinforcement learning to search for shallow sampling circuits for sample-based quantum diagonalization (SQD). It combines a particle-number-preserving action vocabulary, symbolic construction rules, and a reward based on the energy recovered by SQD with a circuit-length penalty. Experiments on N₂, H₂O, and LiH demonstrate compact circuits and execution on IBM quantum hardware. A policy ablation shows improved search consistency over undirected search. However, the tested active spaces are small enough that uniform sampling with classical post-processing can also recover the reference energies, limiting conclusions about learned sampling quality.

Critical strengths

The method aligns circuit search with the actual SQD objective: producing useful sampled configurations rather than accurately preparing the ground state itself. Combining this objective with particle-number-preserving operations is a sensible use of domain knowledge and offers a promising direction for task-specific circuit search.

The hardware experiments demonstrate that the resulting shallow circuits can run on a real quantum processor and retain a substantial fraction of valid samples. This provides practical evidence of feasibility beyond simulation, although larger-scale benefits remain unproven.

Critical weaknesses

My main concern is that the experiments do not isolate the value of the task-specific designs. The method combines a particle-number-preserving action vocabulary, brick actions, and symbolic rules, but the policy ablation only compares learned and undirected search within the same space. It therefore does not establish which proposed designs drive the gains. Since these designs carry much of the contribution beyond standard RL, this missing evidence materially weakens the case for acceptance.

The tested active spaces also limit the significance of the results: uniform sampling with classical post-processing can recover the reference energies, and both learned and undirected search achieve errors far below chemical accuracy. The experiments demonstrate feasibility and improved search consistency, but do not establish a practically meaningful advantage in selecting useful configurations under a constrained budget.

Questions for authors

Could you provide controlled ablations isolating the action vocabulary, brick actions, and symbolic construction rules, with the molecular system and evaluation budgets held fixed? These results would materially affect my assessment by showing whether the task-specific designs provide gains beyond standard RL search.

Could you clarify which of these designs are new relative to prior RL circuit-search methods using chemically motivated excitation gates?

Do you have evidence that learned search reduces the sampling or search budget needed to reach a fixed accuracy compared with undirected search or a simple fixed circuit using the same gate vocabulary? This would help establish practical value beyond the small energy differences reported.

Reviewer_HfkfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explore the design of the sampling circuit inside SQD as a sequential decision problem and tackles it with a neurosymbolic reinforcement learning (NSRL) agent. The pipeline is standard SQD and the only learned component is the sampling circuit. A symbolic grammar enforces a depth limit, forbids repeated identical operations, and requires a minimum length before stopping. An actor critic network proposes operations, the terminal reward is the negative SQD energy gap to FCI. Experiments on qpu possess N₂, H₂O, and LiH dissociation curves and a larger all electron N₂ space for sampling budget, noise, and ablation tests.

Critical strengths
  • Scoring a candidate circuit by the quality of the subspace it samples is well motivated. This matters due to entire premise of SQD is that the classical diagonalization determines the answer. A variational energy reward would optimize the wrong quantity.

  • Section 4.5 and Figure 4 go beyond our circuit is shallower notion to quantify why symmetry preservation matters operationally. Every shot that leaves the particle number sector is wasted, and the searched valid sample fraction of circuit degrades from 0.771 to 0.603 across g_2 = 0.001 to 0.010 while truncated UCI collapses.

Critical weaknesses
  • The active spaces studied have particle number conserving dimensions that are below or comparable to the sampling budget, so that any sampler restricted to the correct sector, uniform sampling included, therefore reaches the FCI energy once configuration recovery is applied. This is a strong compromise given recovers the active space ground state at every geometry and chemical accuracy ... with more than two orders of magnitude fewer two qubit gates. If uniform sampling which has no circuit at all also recovers FCI at these budgets, then the 0.000 mHa gaps in Table 1 are not evidence that the searched circuit is a good sampler. They are evidence that the problem is small enough that subspace quality is irrelevant. The contributions such as circuit cost, constructibility, valid-sample fraction are actual but weak than the framing. The lead to confusion if one stops at the abstract or Table 1 would draw the wrong conclusion.

  • The learned policy against undirected search over the same space, but both arms terminate at the depth limit so circuit length is identical by construction, which means the length penalty is constant and the comparison reduces to which random circuit happens to land in a good spot. With ten seeds per arm and a Welch t-test giving p≈0.02, the statistical evidence is weak, and the per-seed values show the undirected arm ranging 0.000–0.104 mHa against 0.000–0.030 for the learned policy i.e., the best undirected seed matches the best learned seed exactly, and the difference is entirely in the tail. It is a much weaker notion. The primary system results in Sections 4.2, 4.6 come from an earlier, off policy version of the search while all 14e,10o results use the on-policy formulation, and hardware results and the ablation are not from the same method. The ablation does not actually validate the method used for the most prominent results. Additionally, the gate angle is fixed at π\pi/8 by empirical fiat, and the authors note the recovered energy is insensitive over the lower range but degrades badly at π\pi/2.

Questions for authors
  1. The uniform sampling within the correct sector recovers FCI at every budget tested in the 14e,10o space. The authors should provide a comparison of uniform sampling gap vs. searched circuit gap vs. auto UCJ gap, for instance, 1k and 5k samples.

  2. The primary system results in Table 1, Table 2 use the off-policy search, while the ablation and budget study use the on-policy formulation. Would the on-policy method on the primary systems reproduce the Table 1 gaps?

  3. With ten seeds and a best case tie between arms, the p≈0.02 result rests on the tail of undirected arm. What is the effect size if you compare medians for <0.01 mHa?

  4. The LiH hardware run at R=1.0 Å, it does not correspond to a reported classical point. The authors should provide the classical SQD gap at R=1.0 Å.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.