acceptodds
Under review as a conference paper at ICLR 2027

LieQAS: Differentiable Quantum Architecture Search via Continuous Lie-Algebraic Parameterization

Abstract

Quantum Architecture Search (QAS) aims to automate variational circuit design. While classical differentiable Neural Architecture Search (NAS) can continuously relax candidate operations, its standard weighted operation-mixing relaxation does not directly transfer to quantum circuits. Existing differentiable QAS methods therefore introduce differentiability over discrete circuit-operation choices through sampling, categorical relaxation, or mixed-state representations. These mechanisms make architecture optimization indirect: sampling yields stochastic gradient estimates, while surrogate or ensemble objectives may mismatch with the discrete circuit ultimately selected. We propose LieQAS, a differentiable QAS framework that instead parameterizes the searched unitary directly in continuous Lie-algebraic coordinates using Pauli strings. Gradients are taken directly with respect to these Pauli-string coefficients, while the architecture is defined by the active Pauli-string support and updated through progressive pruning. We further introduce hardware-cost-aware regularization and non-commutativity-aware search with singleton finalization, so that the finalized architecture admits exact direct extraction. Experiments on molecular VQE benchmarks show that LieQAS achieves chemical accuracy on all four VQE instances while reducing compiled gate count by up to 64% relative to QuantumDARTS, with lower circuit depth and no more CX gates. On MNIST 0/1 classification, LieQAS reduces CX count by up to 89% while maintaining competitive mean accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum, Architecture Search, Differentiable

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_E8PCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LieQAS, a differentiable QAS framework that instead parameterizes the searched unitary directly in continuous Lie-algebraic coordinates using Pauli strings. Gradients are taken directly with respect to these Pauli-string coefficients, while the architecture is defined by the active Pauli-string support and updated through progressive pruning. The authors further introduce hardware-cost-aware regularization and non-commutativity-aware search with singleton finalization, so that the finalized architecture admits exact direct extraction. Experiments on molecular VQE benchmarks show that LieQAS achieves chemical accuracy on all four VQE instances while reducing compiled gate count by up to 64% relative to QuantumDARTS, with lower circuit depth and no more CX gates. On MNIST 0/1 classification, LieQAS reduces CX count by up to 89% while maintaining competitive mean accuracy.

Critical strengths
  1. LieQAS parameterizes the searched unitary directly in Lie-algebraic coordinates and defines the architecture by the active Pauli-string support, instead of using Monte Carlo or Gumbel-Softmax sampling.

  2. Each layer is unitary by the Lie-algebraic architecture construction.

  3. A progressive search approach is designed to tackle the non-commutativity of Pauli strings.

Critical weaknesses
  1. The authors admit that "The resulting compact circuits therefore do not imply lower classical search cost". The searching cost of LieQAS is higher than existing methods, e.g., QuantumDARTS.

  2. The scope is smaller than traditional QAS: layers L is a fixed input in Algorithm 1 while all-to-all connectivity is assumed. Baselines do not contain HEA.

  3. The paper evaluates three independent seeds and report the minimum-depth finalized architecture among the runs that reach chemical accuracy in Appendix F. Maybe robustness study is needed for a better understanding of the method.

Questions for authors

See Weaknesses

3 comments
AuthorsOpenReview ↗

Response Part1

We thank the reviewer for identifying three important aspects that require further clarification: classical search cost, search-space scope and hardware connectivity, and search robustness. We address these concerns separately below.

  1. Classical search cost

LieQAS incurs higher measured runtime and peak GPU memory than QuantumDARTS on the evaluated benchmark. The measured runtime of LieQAS is primarily determined by dense matrix-exponential simulation and the dense-to-sparse search trajectory. These results characterize the computational cost across the evaluated benchmark groups.

2. Search-space scope, restricted connectivity, and the HEA baseline.

Initial search capacity and search-space scope. The initial layer count LL is a user-specified capacity. Within this capacity, LieQAS searches the active Pauli string support through progressive pruning. The framework does not currently grow the architecture beyond the initial capacity. We clarify this scope in the revision.

We further evaluate the effect of initial capacity with G=10G=10 on the 8-qubit Fashion-MNIST 3/6 task. Results show that increasing initial capacity raises runtime and memory without a sustained improvement in accuracy. These results support using a moderate initial capacity.

L2030405060
Acc85.0±2.585.0\pm2.585.3±1.185.3\pm1.184.7±2.284.7\pm2.284.7±2.384.7\pm2.384.9±2.384.9\pm2.3
Time (s)268.1±5.1268.1\pm5.1319.5±12.2319.5\pm12.2383.7±13.5383.7\pm13.5448.6±13.2448.6\pm13.2500.0±10.6500.0\pm10.6
Mem. (MiB)12921598188221922478

Hardware connectivity. The Pauli candidate pool used in the current search assumes all-to-all connectivity. To examine the impact of restricted hardware connectivity, we additionally compile the searched circuits onto three connectivity graphs: all-to-all, nearest-neighbor linear chain, and the IBM Brisbane heavy-hex coupling graph, with bidirectional connectivity.

The recompiled resources are shown below and each entry reports Gates/Depth/CX. The results show that LieQAS generally retains its gate-count and CX-count advantages over the QAS baselines. These experiments evaluate post-search compilation under different connectivity constraints. The current LieQAS is not yet explicitly constrained by the target hardware topology. We clarify this scope in the revised manuscript and identify connectivity-aware candidate generation as a future extension.

MoleculeMethodAll-to-allLinear ChainBrisbane
H2 (4q)UCCSD116/76/43124/76/41124/76/41
QDARTS33/17/636/23/936/23/9
LieQAS12/7/212/7/212/7/2
LiH (4q)UCCSD331/238/143384/259/188384/259/188
QDARTS71/35/1685/47/2285/47/22
LieQAS53/32/1267/38/1769/40/17
LiH (6q)UCCSD466/346/232490/348/233490/348/233
QDARTS128/50/40269/161/139245/155/112
LieQAS74/42/1891/51/3088/50/27
H2O (8q)UCCSD2167/1688/11692452/1703/13172415/1696/1319
QDARTS193/62/44365/169/189348/165/165
LieQAS149/59/44233/102/86206/94/73
EncodingMethodAll-to-allLinear ChainBrisbane
PCA-angleDQAS58.8/18.2/21.478.0/26.0/33.873.6/24.6/32.8
QDARTS48.0/19.0/18.089.2/51.0/53.273.0/40.4/41.0
LieQAS16.8/15.4/2.016.8/15.4/2.016.8/15.4/2.0
PCA-denseDQAS60.4/18.0/20.671.8/24.0/28.671.4/23.8/30.2
QDARTS46.0/17.2/16.087.2/43.8/42.277.6/41.4/35.0
LieQAS20.0/16.0/2.020.0/16.0/2.020.0/16.0/2.0
AE-angleDQAS60.4/19.4/24.686.0/32.0/45.676.6/25.2/42.2
QDARTS40.6/14.0/10.661.8/27.6/23.855.8/23.0/19.8
LieQAS25.2/20.6/6.429.4/23.8/9.025.0/20.6/6.4
AE-denseDQAS61.6/18.6/20.472.6/26.0/30.266.6/23.8/27.6
QDARTS47.0/19.0/17.082.4/45.0/45.069.0/42.8/39.0
LieQAS25.2/18.2/6.028.2/21.2/7.826.6/19.6/6.6
AuthorsOpenReview ↗

Response Part2

HEA comparison. We additionally compare LieQAS with an HEA, comprising complete layers of independently parameterized RYR_Y and RZR_Z rotations alternating with full ring-CX entangling layers. Since complete HEA layers have discrete resource increments, the resulting budgets are not exactly matched.

We perform two-sided Welch’s tt-tests across the four encodings and report the p-values. Across most configurations, LieQAS has comparable accuracy with HEA. The low accuracy of HEA on AE-angle appears related to the limited flexibility in combining features under this particular shallow topology, while LieQAS successfully discover suitable circuits.

The resource comparison also reveals a trade-off between the two architectures. HEA achieves a lower compiled depth, whereas LieQAS uses fewer gates. This is consistent with their different design objectives: HEA prioritizes shallow hardware-compatible topologies, while LieQAS emphasizes accuracy and Pauli-weight-dependent resource costs.

EncodingMethodAcc. (%)GatesCXDepthP
PCA-angleHEA99.52±0.0499.52\pm0.0440.08.012.00.2925
LieQAS97.20±4.2897.20\pm4.2816.82.015.4
PCA-denseHEA98.72±0.0098.72\pm0.0040.08.012.00.0851
LieQAS93.42±5.2193.42\pm5.2120.02.016.0
AE-angleHEA79.76±1.4879.76\pm1.4840.08.012.00.000020
LieQAS94.81±0.1394.81\pm0.1325.26.420.6
AE-denseHEA89.33±0.3889.33\pm0.3840.08.012.00.0910
LieQAS92.21±2.9092.21\pm2.9025.26.018.2

3. Search robustness and seed variability

To assess robustness directly, we show our evaluation with three seeds for each molecule. Across the four molecules, LieQAS achieves chemical accuracy in 7/12 runs, tying QuantumDARTS for the highest aggregate success count, compared with 6/12 for DQAS and 5/12 for ρ\rhoDARTS. Moreover, the selected feasible LieQAS circuits generally require fewer compiled resources.

MoleculeMethodΔESuccess1QCXGatesDepth
H₂-4qLieQAS(2.6±2.5)×10−7(2.6\pm2.5)\times10^{-7}3/311.0±1.711.0\pm1.72.0±0.02.0\pm0.013.0±1.713.0\pm1.77.3±0.67.3\pm0.6
QuantumDARTS(1.3±1.1)×10−2(1.3\pm1.1)\times10^{-2}1/327.0†27.0^\dagger6.0†6.0^\dagger33.0†33.0^\dagger17.0†17.0^\dagger
DQAS(3.1±0.6)×10−6(3.1\pm0.6)\times10^{-6}3/338.3±4.038.3\pm4.012.7±5.012.7\pm5.051.0±9.051.0\pm9.023.7±7.023.7\pm7.0
RhoDARTS(6.3±11)×10−3(6.3\pm11)\times10^{-3}2/327.0±2.827.0\pm2.89.0±1.49.0\pm1.436.0±4.236.0\pm4.218.0±1.418.0\pm1.4
LiH-4qLieQAS(1.5±1.6)×10−3(1.5\pm1.6)\times10^{-3}2/357.0±22.657.0\pm22.620.0±11.320.0\pm11.377.0±33.977.0\pm33.946.0±19.846.0\pm19.8
QuantumDARTS(1.6±2.3)×10−5(1.6\pm2.3)\times10^{-5}3/363.0±9.563.0\pm9.518.3±2.118.3\pm2.181.3±11.281.3\pm11.240.7±5.040.7\pm5.0
DQAS(3.7±6.0)×10−5(3.7\pm6.0)\times10^{-5}3/348.0±3.048.0\pm3.012.3±2.112.3\pm2.160.3±4.060.3\pm4.029.7±3.529.7\pm3.5
RhoDARTS(2.5±3.8)×10−4(2.5\pm3.8)\times10^{-4}3/333.0±2.633.0\pm2.629.0±6.029.0\pm6.062.0±7.062.0\pm7.041.3±6.741.3\pm6.7
LiH-6qLieQAS(1.5±1.9)×10−2(1.5\pm1.9)\times10^{-2}1/356.0†56.0^\dagger18.0†18.0^\dagger74.0†74.0^\dagger42.0†42.0^\dagger
QuantumDARTS(2.7±4.3)×10−2(2.7\pm4.3)\times10^{-2}1/388.0†88.0^\dagger40.0†40.0^\dagger128.0†128.0^\dagger40.0†40.0^\dagger
DQAS(4.7±2.7)×10−2(4.7\pm2.7)\times10^{-2}0/3————
RhoDARTS(2.9±1.4)×10−2(2.9\pm1.4)\times10^{-2}0/3————
H₂O-8qLieQAS(5.4±3.9)×10−3(5.4\pm3.9)\times10^{-3}1/3105.0†105.0^\dagger44.0†44.0^\dagger149.0†149.0^\dagger59.0†59.0^\dagger
QuantumDARTS(2.9±4.1)×10−3(2.9\pm4.1)\times10^{-3}2/3159.0±18.4159.0\pm18.452.5±0.752.5\pm0.7211.5±19.1211.5\pm19.157.0±2.857.0\pm2.8
DQAS7.6×10−3±4.4×10−67.6\times10^{-3}\pm4.4\times10^{-6}0/3————
RhoDARTS(6.2±2.6)×10−3(6.2\pm2.6)\times10^{-3}0/3————

I thank the authors for the detailed response. I will keep my score to accept the paper.

Reviewer_oBM8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce LieQAS which is a differentiable QAS framework that parameterizes the searched unitary directly in continuous Lie algebraic coordinates over Pauli strings. Under this process the gradients are evaluated with respect to the Pauli string coefficients, whereas the architecture is defined by the active Pauli string support. The support is updated using progressive pruning by a first order Taylor saliency. The method also adds hardware cost-aware regularization and noncommutativity-aware search with singleton finalization. The author frame the motivation around two problems of DQAS (1) sampling gives stochastic gradient estimates, but the surrogate objectives may mismatch the discrete circuit that is selected at the end. (2) The search space of DQAS is not tractable.

The paper show modest gain over other DQAS approaches. In the following I detail my comments and concern.

Critical strengths
  1. The motivation of the paper is quite clear. Searching directly in continuous Pauli string coefficients, with the architecture given by the active support, is a clear formulation that targets the mismatch between a continuous surrogate and the discretly selected circuit.

  2. The hardware cost-aware regularization and the non commutativity-aware search makes the evaluation relevant to practical deployment rather than only to energy error.

  3. The evaluation is broad and cover both quantum chemistry and QML classification, the QML results are reported over five seeds. I also appreciate the authors for being honest that the dominant cost is dense simulation and that the pool reduction concerns search coordinates rather than simulation cost.

Prompt 1: Above a provided a rough peer review of a paper I am reviewing for a top ML conference. Plaese correct the grammer, spelling, and typos, please do not change core meaning or technical claims. Keep the latex format and give the results in .md format friendly way..

Prompt 2 Rewrite so that it is concise and constructive. Do not add or alter any number, citation or claim. remove "I believe" "somehow" "it seems to me" and replace with better words.

Prompt 3 Can you check comparing the initial and final review if any concern left then output a corrected final version.

and the output of the Prompt 3 is used in the final review. I would like to sincerely clarify that the review I provided is human written but the details are corrected or modified by an AI model.

Critical weaknesses
  1. The headline claim of up to 64% gate reduction is relative to QuantumDARTS only, and Table 1 compares just against UCCSD and QuantumDARTS. The paper does not compare against RL-based QAS such as curriculum RL for QAS (which the authors cited in the paper) and more recently TensorRL-QAS [NeurIPS 2025]. There I believe are the most direct competing line of work, so any claim of efficiency should be supported against other QAS approaches as well not only on one differentiable baseline.

  2. The appendix Equation 14 and 15 states that one optimization step costs O(L8n+Ln24n)O(L8^n + Ln^24^n), dominated by dense matrix exponentiation of the layer Hamiltonians, with O(4ⁿ) memory per matrix. I appreciate the authors' transparency that the reduction from 4n−14^n−1 to O(n2)O(n^2) Pauli strings concerns the number of search coordinates and not the simulation cost. The practical consequence is that the weight-two restriction lowers the cost relative to the complete pool from 16n16^n to 8n8^n but leaves it exponential in the number of qubits, so the framework is confined to small systems. Which is reflected in the paper where the largest system studied is at 8 qubits. Since the authors report wall-clock time, I ask them to show how runtime and memory scale with qubit number and to state which gradient method the timings assume. Moreover, the paper also does not discuss the cost of Taylor-FO gradients under parameter-shift on hardware or under shot noise. A discussion of simulation backends that avoid dense matrix exponential would clarify whether the framework can extend beyond small systems.

  3. I have two concerns regarding the the choice weight-≤\leq 2 in the paper. (3a) Chemistry Hamiltonians are 2-body in the fermionic picture, but after Jordan-Wigner or Bravyi-Kitaev mapping the terms become O(n)- or O(logn)-local, respectively. Natural UCCSD generators have weight at least 4 doubles, growing up to O(n) under JW. A weight-≤\leq 2 pool must therefore build them from repeated commutators, as the authors noted in the paper, which means deeper circuits and a harder optimization landscape. (3b) The unconstrained pool may also break particle-number and spin symmetries that fermionic excitation generators preserve, and because it covers all qubit pairs it may add SWAP overhead on sparse hardware. I request an ablation over the different weight cutoff such as 1, 2, 3, 4, a benchmark whose optimal generators have weight greater than 2, and an evaluation under restricted connectivity.

  4. On LiH (4q), LieQAS reaches 6.0×10−46.0 × 10^{-4} Ha, compared with 2.0×10−62.0 × 10^{-6} Ha for QuantumDARTS, and on LiH (6q) it ties QuantumDARTS. So I observe that the gate savings come with an accuracy trade-off and the paper does not discuss it. Also (4a) Table 1 reports no standard deviations over seeds, and (4b) it appears to be noiseless, although the abstract emphasizes hardware-cost awareness.

  5. I found the QML benchmark is too easy to separate the methods. The considered MNIST 0/1 is among the simplest binary classification tasks. This is reflected by the fact that handcrafted QCNN already reaches 97.3–99.3% accuracy across all four encodings, and LieQAS-L within about one point of it. With so little improvement, many differences fall within seed variance for example, PCA-angle and AE-dense, and the text itself calls the results "similar."

DQAS and QuantumDARTS have standard deviations of 3–14 points, so part of LieQAS-S's advantage on AE-angle may reflect baseline instability rather than search quality. I request a harder benchmark. For example, multiclass MNIST or Fashion-MNIST, a reduced-size QCNN or QCNN-style search space as a budget-matched baseline, and more seeds with significance tests.

Questions for authors
  1. How does LieQAS compare with curriculum RL for QAS and TensorRL-QAS in energy error, gate count, depth, and CX count under the same settings?

  2. How do wall-clock time and memory scale with qubit number, which gradient method do the timings assume, and can the framework run on a backend that avoids dense matrix exponentials?

  3. How does LieQAS perform when the pool includes weight-3 and weight-4 Pauli strings, and what is the depth overhead of building UCCSD-type generators from weight-≤2 commutators?

  4. Does the pool preserve particle number and spin, and how does it behave under restricted hardware connectivity?

  5. Why does LiH (4q) lose accuracy relative to QuantumDARTS, and what are the seed variances and noisy-simulation results for Table 1?

  6. Does LieQAS retain its advantage on a harder QML task and against a budget-matched QCNN, with more seeds and significance tests?

10 comments
AuthorsOpenReview ↗

Response Part1

We thank the reviewer for the thorough evaluation and constructive suggestions. In response, we have substantially expanded our experiments and analysis. We have also clarified the scope of our design and strengthened the experimental evidence supporting our claims in revision. We address each concern in detail below.

  1. RL-based methods comparison

LieQAS is designed to address the discrete sampling in differentiable QAS by direct optimizing of continuous Lie-algebraic Pauli coefficients. The original comparison is limited in differentiable QAS and the reported gate-count reduction of up to 64% specifically refers to QuantumDARTS, which we clarify in the revision. We extend our evaluation to Curriculum-RL and TensorRL-QAS, considering two initialization settings to distinguish the effects of tensor-network warm-starting from those of architecture search.

MPS initialization. Following the TensorRL-QAS setting, all methods start from the same MPS-prepared state. LieQAS, TensorRL, and Curriculum-RL all achieve 5/5 successful runs. The following table reports the energy errors and compiled circuit resources with (w) and without (w/o) the cost of tensor-network initialization. The RL-based methods achieve smaller energy errors and more compact circuits than LieQAS.

MoleculeMethodError (Ha)w/ow
GatesCXDepthGatesCXDepth
BeH$_2$-6qTensorRL(6.0±1.9)×10−5(6.0\pm1.9)\times10^{-5}15.85.47.6111.220.031.6
Curriculum-RL(6.5±0.4)×10−5(6.5\pm0.4)\times10^{-5}17.25.08.4119.419.834.2
LieQAS(8.5±0.6)×10−5(8.5\pm0.6)\times10^{-5}50.411.620.0135.625.441.8
H$_2$O-8qTensorRL(8.7±3.1)×10−4(8.7\pm3.1)\times10^{-4}14.68.65.8154.029.630.4
Curriculum-RL(9.1±3.7)×10−4(9.1\pm3.7)\times10^{-4}16.27.65.4162.028.632.2
LieQAS1.2×10−3±1.1×10−51.2\times10^{-3}\pm1.1\times10^{-5}59.012.418.6184.032.640.6

Zero-state initialization. We further evaluate all methods on the BeH$_2$ hamiltonian from the zero state aligning with LieQAS, including a version of TensorRL without TN initialization. In this harder task than MPS initialization, LieQAS requires the fewest CX gates while TensorRL searches the shallowest circuit. Thus, LieQAS provides a favorable entangling-gate trade-off under this shared initialization setting, aligning with the usage of regularization to reduce the cost of weight-two Pauli string.

We further extend the zero-state comparison to H$_2$O (8q). Due to limited computational resources, we have not yet completed the full exploration budgets of the RL-based methods. Currently, two independent seeds are being evaluated for each RL baseline, and neither has exhausted its prescribed search budget. We therefore report the best circuits identified so far in these runs and both RL baselines converged to the same circuit in the two evaluated runs, resulting in identical compiled resource statistics. In contrast, both LieQAS runs have completed the full search and finalization procedure. At the current checkpoints, LieQAS requires fewer compiled gates, CX gates, and lower circuit depth than the best-so-far RL circuits.

MoleculeMethodError (Ha)CXGatesDepth
BeH$_2$-6qLieQAS5.8×10−5±3.5×10−125.8\times10^{-5}\pm3.5\times10^{-12}9.2±2.347.8±7.922.4±4.9
TensorRL without TN1.9×10−6±1.2×10−61.9\times10^{-6}\pm1.2\times10^{-6}17.2±9.136.6±16.117.0±8.3
Curriculum-RL1.4×10−6±9.3×10−71.4\times10^{-6}\pm9.3\times10^{-7}19.6±8.854.6±7.925.6±4.2
H$_2$O-8qLieQAS1.0×10−3±5.3×10−71.0\times10^{-3}\pm5.3\times10^{-7}29.0 ± 4.2110.5 ± 23.349.5 ± 6.4
TensorRL without TN1.1×10−3±3.5×10−51.1\times10^{-3}\pm3.5\times10^{-5}85.5 ± 14.8119.0 ± 12.754.5 ± 3.5
Curriculum-RL1.1×10−3±3.5×10−51.1\times10^{-3}\pm3.5\times10^{-5}85.5 ± 14.8119.0 ± 12.754.5 ± 3.5
AuthorsOpenReview ↗

Response Part2

  1. Search overhead

Runtime and memory scaling. We evaluated MNIST-0/1 with 4, 6, 8, and 10 qubits, using L=20L=20, G=10G=10. The table reports the measured wall-clock time, GPU memory, initial candidate count, actual number of pruning rounds, and total training epochs. The measured runtime reflects both dense matrix-exponential simulation and architecture-search overhead. In particular, the number of pruning rounds affects the total number of recovery epochs. The peak-memory overhead is also associated with the dense operator/state tensors and fixed-shape search tensors retained throughout pruning. These results quantify the scalability of our current implementation.

QubitsWall-clock time (s)Mem. (MiB)Initial CandidatesRoundsTotal Epochs
471.74448132051370
686.58502306059410
8262.091,292552065440
105,117.2117,630870069460

Gradient method and hardware estimation. The reported timings use PyTorch automatic differentiation (autograd) through the dense matrix-exponential simulator, without parameter-shift estimation or finite-shot measurement noise. Taylor-FO saliency reuses the gradients computed during parameter optimization and introduces no additional gradient-estimation pass in our current implementation.

On quantum hardware, gradients of mixed-generator noncommuting evolutions require suitable estimators, such as the stochastic parameter-shift rule [1]. Its execution cost depends on the estimator sampling budget, observable measurements, and target precision. Finite-shot estimation introduces an additional accuracy-measurement-budget trade-off, as analyzed in [2]. For Taylor-FO pruning, this motivates accounting for gradient uncertainty when comparing candidate importance scores. These considerations identify hardware gradient estimation and measurement-aware pruning as directions for extending LieQAS to hardware-in-the-loop search.

Alternatives to dense matrix exponentiation. The LieQAS parameterization does not intrinsically require explicitly constructing dense layer unitaries. Exponential-action methods, including Krylov-subspace approximations and scaling-and-truncated-Taylor methods, can evaluate eA∣ψ⟩e^A|\psi\rangle without forming eAe^A [3,4]. These methods offer potential alternatives for reducing the simulation overhead. Integrating exponential-action methods with efficient automatic differentiation, numerical error control, and Taylor-FO pruning remains an open implementation challenge. We clarify this in the revision.

[1] Leonardo Banchi and Gavin E Crooks. Measuring analytic gradients of general quantum evolution with the stochastic parameter shift rule. Quantum, 5:386, 2021

[2] Andrea Mari, Thomas R Bromley, and Nathan Killoran. Estimating the gradient and higher-order derivatives on quantum hardware. Physical Review A, 103(1):012405, 2021.

[3] Yousef Saad. Analysis of some krylov subspace approximations to the matrix exponential operator. SIAM Journal on Numerical Analysis, 29(1):209–228, 1992.

[4] Awad H Al-Mohy and Nicholas J Higham. Computing the action of the matrix exponential, with an application to exponential integrators. SIAM journal on scientific computing, 33(2):488–511, 2011.

AuthorsOpenReview ↗

Response Part3

  1. Pauli Weight

We investigate the cost of synthesizing higher-weight transformations using weight-≤2 Pauli rotations through UCCSD synthesis comparisons and controlled target-unitary experiments.

Maximum-weight ablation. We reran the ablation of the maximum Pauli weight wmax⁡∈{2,3,4}w_{\max}\in\{2,3,4\} in Appendix C.3.2. Weight-one Pauli string cannot generate entanglement from a product initial state, motivating weight two as our starting cutoff. On the evaluated task, expanding the pool to weight-three or weight-four provides minor accuracy improvements, while substantially increasing search time and peak memory. These results illustrate the computational trade-off underlying our choice of weight-two.

Max WeightAcc. (%)Time (s)Mem. (MiB)Compiled Resources
GatesDepth1QCX
285.0±2.585.0 \pm 2.5268.1±5.1268.1\pm5.11292129228.828.826.026.020.820.88.08.0
385.7±1.185.7 \pm 1.11104.9±116.71104.9\pm116.72780278042.442.429.429.426.826.815.615.6
485.5±1.285.5 \pm 1.24007.5±120.44007.5\pm 120.48454845450.650.635.835.830.030.020.620.6

UCCSD synthesis overhead. Using the same molecular configurations as Table 1, we additionally reconstruct the same UCCSD evolution using a construction composed of weight-≤2\leq2 Pauli rotations. The table shows the change, confirming that weight restrictions can substantially increase implementation depth and resource.

MoleculeGatesCXDepthDepth Overhead
H$_2$(4q)116→165116 \rightarrow 16543→4443 \rightarrow 4476→10876 \rightarrow 10842.1%
LiH(4q)331→537331 \rightarrow 537143→150143 \rightarrow 150238→367238 \rightarrow 36754.2%
LiH(6q)466→793466 \rightarrow 793232→230232 \rightarrow 230346→567346 \rightarrow 56763.9%
H$_2$O(8q)2167→42112167 \rightarrow 42111169→11941169 \rightarrow 11941688→30841688 \rightarrow 308482.7%

Higher-weight target synthesis. We additionally evaluate four controlled four-qubit unitary-synthesis tasks. We compare maximum Pauli weights of two and three at budgets G={1,3,5,8}G=\{1,3,5,8\}. Success requires full-unitary process infidelity below 10−610^{-6}. The circuit depth is averaged over successful runs.

TaskGweight-≤\leq 2weight-≤\leq 3
SuccessDepthSuccessDepth
XXXI10/5—4/59.0±0.09.0\pm0.0
31/517.0†17.0^\dagger5/59.0±0.09.0\pm0.0
53/517.0±6.217.0\pm6.25/59.6±1.39.6\pm1.3
85/525.0±3.725.0\pm3.75/510.2±2.210.2\pm2.2
XYZI10/5—5/59.0±0.09.0\pm0.0
32/514.0±4.214.0\pm4.25/59.0±0.09.0\pm0.0
54/518.3±2.918.3\pm2.95/510.0±2.210.0\pm2.2
84/523.3±6.523.3\pm6.55/527.0±14.927.0\pm14.9
XXXX10/5—0/5—
30/5—3/520.3±1.220.3\pm1.2
50/5—3/521.7±3.821.7\pm3.8
82/524.0±5.724.0\pm5.75/523.4±7.423.4\pm7.4
XYZX10/5—0/5—
30/5—1/517.0†17.0^\dagger
51/523.0†23.0^\dagger4/522.0±7.422.0\pm7.4
84/524.5±1.324.5\pm1.34/529.3±8.829.3\pm8.8

These results show that higher-weight generators can improve representational efficiency under tight Pauli string budgets, while lower-weight generators can recover higher-order transformations through compositions of multiple rotations.

Together, these results show that Pauli string weight introduces a trade-off among expressivity, hardware cost, and search efficiency.

Reviewer_em6oOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

Differentiable Quantum Architecture Search (QAS) methods like DQAS and QuantumDarts optimize over a discrete set of gate candidates using either Monte-carlo sampling or Gumbel Softmax relaxation with one hot encoding. This only evaluates a few sampled circuits at a time and can make gradients (QDARTs) more stochastic.

This work takes a very different approach. LieQAS parametrizes a layer using a single matrix exponential over a sum of all active candidates. This allows the method to test and optimize all candidates simultaneously with exact gradients. They use Taylor-FO saliency and pruning to iteratively define candidate importance and remove less important ones. Their benchmarks also show favorable results with similar or better performance with fewer gates.

Critical strengths

I find the idea overall solid, simple, and smart. Standard Neural Architecture Search or DARTS does not typically preserve unitary evolution (i.e. weighted sum of unitaries isn’t necessarily unitary) which motivated DQAS. DQAS requires the MC sampling of different circuits to update probabilities of the candidates, which can be considerably slow and also more restrictive given that a circuit can be formed by a wide combination of different candidates so many circuit combinations are not tested (in fact most aren’t). QuantumDarts does somewhat fix the sampling issue of DQAS however because only 1 candidate circuit is picked the probabilities and gradients are not exact and the method could suffer in terms of accuracy. Similarly it doesn’t allow for the testing of as many circuits as LieQAS.

Their method takes a completely different approach. By creating a combined unitary of all possible candidates they are “essentially” able to test ALL candidates at the same time with no sampling. The use of L1 regularization/Taylor-FO saliency to “punish” worse candidates and push them towards 0 is also very intuitive, and then the pruning follows directly once the parameters have been optimized for some time. I also liked the idea of having 2 regularizer terms, 1 for the single and another for the two qubit operations. This allows for searches that restrict expensive gates more. In practice this allows for the comparison of more candidates than the prior methods potentially allowing for completely different circuits to be chosen.

The method can also be adapted in several ways which is a major strength. For example the specific way pruning is done can easily be changed in their algorithm to whatever method someone might want to use. The scoring of a candidate's importance could also be changed, simpler approaches are also possible like simply looking at gradient histories. The main idea essentially then simplifies to 1) test all candidates, 2) evaluate some importance score, 3) prune away the worst candidates, 4) repeat until G is reached. In practice this framing is similar to the framing in WSBD [1] which is a lot clearer to follow (although that method is for freezing while this is used for architecture search I think a similar framing or stage approach would be better). I would summarize LieQAS in this 4 stage process and highlight the adaptiveness of this work. This makes the work clearer and also highlights a key strength of the paper.

Critical weaknesses

I found the writing of the LieQAS algorithm somewhat tedious to follow. The strength of the method is the simplicity, but many facts are obscured or hidden in Algorithm 1 which is in the appendix. This algorithm should be in the main paper and a simple rundown of the overall process/algorithm would make it much easier to follow (i.e. the 4 stage process from above [1]). Once I read algorithm 1 in the Appendix the concerns became a lot clearer. Particularly I have considerable concerns regarding the results, specifically in relation to the hyperparameter tuning and the fairness of the comparisons.

LieQAS introduces a lot of new hyperparameters, and some of them will likely give very different outcomes. The tuning of these hyperparameters and values chosen is not sufficiently addressed and I suspect different values were chosen on the different problems to give favorable results either for better accuracy or for better gate count. To the best of my knowledge I’ve made a list of hyperparameters that LieQAS introduces and highlighted how well they are reported.

While reporting the values used is absolutely critical to this work, I also think it needs to go further. It’s crucial to see that their method is robust for some range of these values and that these values transfer somewhat well across tasks. If LieQAS is hyperparameter sensitive it will be difficult to implement and likely need several runs before good values are chosen. If it is not sensitive, the method would be quite strong however I’m not confident it is. While some of these hyperparameters are also needed in comparable methods a rundown of what “new” hyperparameters LieQAS uses compared to other work would also be helpful. Overall I find the idea intuitive and smart but the lack of transparency and complicated rundown of LieQAS makes me lean reject.


G (or final M): Represents the final amount of surviving pauli strings (parameters). They do report setting G = 10 or 50 for MNIST but this is omitted on the VQE benchmark.

L: Represents the initial number layers for their method that they will begin to prune. What value this is set to is not reported. A larger L will make their method slower since more pruning is needed but also allow for a better search. This value is critical to report as it decides how large ΔM\Delta M is and how fast/accurate their algorithm is.

R: Represents the number of pruning rounds (each round will prune ΔM\Delta M pauli terms). What value this is set to is not reported. A smaller value will be faster since fewer pruning rounds happen, a larger value will give more time for a good search. This value is also critical to report for similar reasons as L.

ΔM\Delta M: Represents number of pauli terms that are pruned each round. What value/function this is set to is not reported. Specifically ΔM\Delta M is likely set to (Mstart−G)/R(M_{start} - G)/R, but it could be non-uniform (i.e. increasing/decaying over time) so it’s essential to know the specific formula and L, G and R for each benchmark that was run.

Epochs/Steps: Represents how many steps are done before a pruning round happens. What value this is set to is not reported. More steps means that we have a better view of what Pauli strings are more/less useful. Fewer steps will be faster but less accurate. This value is also crucial to report.

λ1\lambda_1 and λ2\lambda_2: Represents the L1 penalty for 1 and 2 qubit Pauli strings. A larger λ2\lambda2 is favorable if you want to reduce CX gates for example. This is only partially addressed in ablation 2 where they show an experiment on fashion-mnist with different values, however for the main VQE/MNIST experiment they don’t report what value they use. Likely the large reduction in CX gates comes directly from choosing a larger λ2\lambda_2 value.

α:\alpha: Represents penalty for layers with more candidates / active parameters. What value this is set to is not reported for VQE/MNIST problems.

(There might be more of these that are not reported that I missed so update accordingly).

Questions for authors

How many new hyperparameters does LieQAS introduce? (i.e. which hyperparameters are shared with prior methods and which come from your method specifically)?

[1] Kverne, C., Akewar, M., Huo, Y., Patel, T. & Bhimani, J.. (2026). WSBD: Freezing-Based Optimizer for Quantum Neural Networks . Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 300:4150-4158 Available from https://proceedings.mlr.press/v300/kverne26a.html.

3 comments
AuthorsOpenReview ↗

Response Part1

We thank the reviewer for the constructive suggestions regarding the presentation, hyperparameter transparency, and experimental robustness. We have revised the algorithm presentation and added systematic hyperparameter analyses.

  1. Algorithm presentation

We agree that the overall search procedure should be directly accessible in the main paper. We have added a concise Algorithm 1 in the methodology section. The progressive search loop consists of four steps: (1) jointly optimize the coefficients of all active Pauli strings under the composite unitary; (2) evaluate their importance using Taylor-FO saliency; (3) prune less important candidates according to congestion-adjusted scores; and (4) repeat the optimization and pruning process until the target number of occupied layers is reached. We also revised the beginning of Section 3 to explicitly show that the Lie-algebraic representation and finalization procedure are not tied to a particular importance measure. While our implementation uses Taylor-FO saliency with congestion-adjusted scoring, alternative criteria, including magnitude-based or gradient-based measures, can be incorporated within the same framework. This modularity provides flexibility for adapting LieQAS to different optimization objectives and applications. We clarify this in revision, Section 5.

  1. Hyperparameter identification and classification.

LieQAS introduces four method-specific hyper-parameters for pruning and regularization controls: the pruning ratio pp, the two Pauli-weight-dependent regularization coefficients λ1q\lambda_{1q} and λ2q\lambda_{2q}, and the layer-occupancy exponent α\alpha. The initial layer count LL and target architecture budget GG are general architecture-capacity choices, while the learning rate, batch size, and optimizer are standard differentiable optimization settings. The pruning procedure determines the number of rounds RR and the per-round removal budget ΔMr\Delta M_r. Thus, RR follows from the pruning trajectory and stopping condition.

Specifically, the initial number of active coordinates is M0=LK(n)M_0=LK(n), where K(n)=3n+9(n2)K(n)=3n+9\binom{n}{2} is the number of Pauli string candidates per layer. Given the pruning ratio pp, the removal budget at round rr is determined by

ΔMr=min⁡{Mr−G,max⁡(1,⌈pMr⌉)}.\Delta M_r=\min\{M_r-G,\max(1,\lceil pM_r\rceil)\}.

The progressive pruning stage terminates when Nocc=GN_{\mathrm{occ}}=G, potentially with M>GM>G. Singleton finalization then removes additional strings only from layers with more than one active string, until M=GM=G. Thus, RR is determined by the resulting pruning trajectory and its stopping condition.

  1. Hyperparameter robustness and transferability.

Architecture capacity. LL and GG are task specific. They primarily have the relationship of L/G=2∼3L/G=2\sim3. Smaller ratio is set for easier tasks for saving search time and larger ratio is set for harder tasks for providing broader search space.

MoleculeH2LiH 4qLiH 6qH2O
L32100195270
G16507890
CircuitLieQAS-SLieQAS-L
L30100
G1050

We further provide ablation study for the relationship between LL and GG on Fashion-MNIST 3/6-8q, aligning the configuration with other ablation study. Let G=10, and all experiments run 5 times with seed being 0-4. Results show that increasing initial capacity raises runtime and memory without a sustained improvement in mean accuracy over this range. This supports using a moderate initial capacity. L/G=2∼3L/G=2\sim3 is a reasonable configuration for tasks in this work.

L2030405060
Acc85.0±2.585.0\pm2.585.3±1.185.3\pm1.184.7±2.284.7\pm2.284.7±2.384.7\pm2.384.9±2.384.9\pm2.3
Time (s)268.1±5.1268.1\pm5.1319.5±12.2319.5\pm12.2383.7±13.5383.7\pm13.5448.6±13.2448.6\pm13.2500.0±10.6500.0\pm10.6
Mem. (MiB)12921598188221922478

Pruning Ratio. We further evaluate ratios of 5%, 10%, 15%, and 20% on the same task with L=20L=20 and G=10G=10. Accuracy remains broadly stable, with a slight accuracy decrease at 20%. Increasing the pruning ratio reduces the number of pruning rounds and the total runtime. This demonstrates the expected trade-off between pruning aggressiveness and search cost.

Pruning ratio5%10%15%20%
Acc85.0±2.585.0\pm2.585.0±2.585.0\pm2.584.9±2.684.9\pm2.684.8±2.584.8\pm2.5
Pruning round130654230
Time(s)394.5±16.8394.5\pm16.8268.1±5.1268.1\pm5.1209.2±7.6209.2\pm7.6192.6±9.5192.6\pm9.5
AuthorsOpenReview ↗

Response Part2

Regularization coefficient affects both parameter optimization and subsequent architecture selection: it changes the learned Pauli coefficients, which in turn influence the Taylor-FO saliency scores used for pruning. To examine its transferability across tasks, we evaluate all QML tasks using shared coefficients λ1=0.001\lambda_1=0.001, λ2=0.01\lambda_2=0.01. An asterisk (*) indicates that the original configuration already matches these settings. The shared coefficients maintain or improve mean accuracy on nine of the fourteen evaluated configurations. Five configurations show lower mean accuracy and the largest decreases occur in 8q-MNIST PCA-dense-S, and 8q-MNIST AE-angle-S, accompanied by increased seed variability.

TaskOri. Acc. (%)Shared coefficient Acc (%)
4q-Fashion-MNIST 0/1/2/379.1±0.979.1\pm0.970.9±5.370.9\pm5.3
*4q-Fashion-MNIST 3/687.2±0.487.2\pm0.487.2±0.487.2\pm0.4
4q-MNIST 0/1/2/371.4±2.771.4\pm2.771.1±2.971.1\pm2.9
4q-MNIST 1/095.9±0.195.9\pm0.196.6±0.596.6\pm0.5
4q-MNIST 3/693.3±0.993.3\pm0.987.7±5.787.7\pm5.7
*4q-Vowel47.2±7.347.2\pm7.347.2±7.347.2\pm7.3
8q-MNIST 1/0-PCA-angle,S97.2±4.397.2\pm4.399.2±0.199.2\pm0.1
8q-MNIST 1/0-PCA-dense,S93.4±5.293.4\pm5.282.3±17.982.3\pm17.9
8q-MNIST 1/0-AE-angle,S94.8±0.194.8\pm0.183.9±17.683.9\pm17.6
8q-MNIST 1/0-AE-dense,S92.2±2.992.2\pm2.995.3±1.895.3\pm1.8
*8q-MNIST 1/0-PCA-angle,L99.3±0.199.3\pm0.199.3±0.199.3\pm0.1
*8q-MNIST 1/0-PCA-dense,L99.5±0.199.5\pm0.199.5±0.199.5\pm0.1
*8q-MNIST 1/0-AE-angle,L98.9±0.598.9\pm0.598.9±0.598.9\pm0.5
*8q-MNIST 1/0-AE-dense,L96.1±0.996.1\pm0.996.1±0.996.1\pm0.9

Table 3 in the paper shows the ablation study on the selection of these two coefficients, demonstrating that performance is relatively stable across a broad range of tested coefficients. Together, these results characterize the sensitivity and transferability of the Pauli-weight-dependent regularization coefficients. The shared setting maintains comparable performance across most evaluated tasks, while the coefficient ablation demonstrates the accuracy-resource trade-off under different regularization strengths.

Training epochs/steps. We use a consistent training schedule within each task family. For QML, we perform 15 warmup epochs, 5 recovery epochs after each pruning round, and 100 final retraining epochs after pruning is complete. For VQE, the corresponding stages use 200 warmup steps, 200 recovery steps per pruning round, and 800 final retraining steps.

Layer-occupancy exponent α\alpha is fixed as 1.5 across all main experiments. We further add an ablation study for it on Fashion-MNIST 3/6-8q. The following table shows that the task remains stable with different α\alpha.

αTest Acc.(%)
0.584.97±2.5684.97 \pm 2.56
1.084.97±2.5084.97 \pm 2.50
1.584.98±2.5184.98 \pm 2.51
2.084.81±2.4884.81 \pm 2.48
2.584.84±2.4484.84 \pm2.44

These experiments provide a clearer description of the LieQAS procedure quantifying the effects of capacity, pruning ratio, and regularization coefficients. They support reasonable stability within the tested ranges, while identifying settings that increase cost or degrade accuracy.

I thank the authors for revising the manuscript, improving the algorithms presentation and run several new ablation studies to address both my and OBM8's concerns. Because I like the core idea of LieQAS, and the execution/transparency of their method is improved, I have raised my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.