Twirlator: A Pipeline for Analyzing Subgroup Symmetry Effects in Quantum Machine Learning Ansatzes

Bo Zhao; Ilmo Salmenper\"a; V\"ain\"o Mehtola; Valter Uotila

REVIEW 3 major objections 4 minor 28 references

Twirlator: A Pipeline for Analyzing Subgroup Symmetry Effects in Quantum Machine Learning Ansatzes

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Twirlator claims that symmetrizing quantum ML ansatzes with larger subgroups of the permutation group raises circuit overhead, lowers expressibility, and usually increases entangling capability, across 19 common ansatz patterns.

desk verdict Useful open-source pipeline and benchmark for partial symmetry in QML; the main caveat is that subgroup size is an aggregate knob and within-order variance is unexamined. read the letter →

arxiv 2511.04243 v2 pith:LCO3TD4W submitted 2025-11-06 quant-ph cs.LG

Valter Uotila , Väinö Mehtola , Ilmo Salmenperä , Bo Zhao This is my paper

classification quant-phcs.LG PACS 03.67.Lx

keywords geometricquantummachinelearningsubgroupsymmetrizationtwirlingformulaexpressibilityentanglingcapabilityansatzcircuitspartialsymmetry

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Symmetry is often treated as a binary choice, but this paper introduces a pipeline, Twirlator, that realizes partial symmetry by twirling ansatz generators over subgroups of a symmetric group, using subgroup size as a continuous knob. It measures generator drift, circuit depth and size, expressibility, and entangling capability for 19 standard ansatzes under every subgroup of S4 and sampled subgroups of S5. The central empirical claim is that larger subgroups consistently increase circuit overhead, reduce expressibility, and usually increase entangling capability, while depth leaves the relative trends unchanged. The work matters because it reframes symmetry as a tunable trade-off in quantum machine learning, giving practitioners a concrete way to select partial symmetries that balance hardware cost and model capacity.

What carries the argument

The central mechanism is the twirling formula over a subgroup, which replaces a gate generator G with the average T[G]=(1/|S'|) Σ_{s in S'} U_s G U_s^† over the subgroup's permutation-unitaries. Subgroup order is the tunable symmetry knob; larger order means more averaging and a larger departure from the original generator. For the angle-encoded circuits used here, the induced unitary representations are simply permutation matrices, and the twirl is applied to generators rather than full gates, so the overhead of symmetrization can be measured directly via generator drift and transpiled circuit size.

What would settle it

Run Twirlator on a single ansatz with two order-2 subgroups of S4—one generated by a single transposition and the other by a double transposition—and compare their circuit size, expressibility, and entangling capability; if these differ enough to reverse the reported trend within the same order, then subgroup order alone is not the operative knob.

Watch

Extended reading notes

Core claim

Across 19 ansatz patterns, subgroup-based symmetrization reveals a consistent trade-off: as the subgroup used for twirling grows, generator drift increases, synthesized circuit size grows, expressibility drops (the KL divergence from the uniform pure-state distribution rises), and entangling capability often rises. The effect holds across circuit depths, so the symmetry level, not depth, dominates the relative behavior. Ansatzes with non-parametrized entangling gates behave differently: their entangling capability can decrease under symmetrization. The paper argues that partial symmetries (smaller subgroups) can preserve expressibility and keep circuits small while still encoding some of the

Load-bearing premise

The load-bearing premise is that subgroup size (order) is a sufficient one-dimensional measure of symmetry strength for predicting these trade-offs, even though subgroups of the same order can act on the qubits in structurally different ways.

Editorial extensions

If this is right

Practitioners can choose a partial symmetry level: smaller subgroups give lower circuit overhead and higher expressibility than full-group symmetrization while still providing some inductive bias.
Full-group symmetrization of S4 or S5 is often too expensive, and adding circuit depth does not restore the lost expressibility, so the symmetry level must be set early in ansatz design.
The entangling capability usually increases with subgroup size, but for ansatzes containing non-parametrized two-qubit gates it can decrease, so the gate structure of the base ansatz matters as much as the subgroup size.
Generator drift and circuit size trends are consistent across depths, making depth-1 analysis a reliable proxy for the relative effects of symmetry levels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A likely implication not drawn in the paper: since subgroups of the same order can be non-isomorphic and act on qubits differently, subgroup order alone may not predict behavior for a specific subgroup; within-order variance could be large.
The tendency of larger subgroups to boost entangling capability suggests the twirl effectively introduces entangling interactions; this could be exploited deliberately, but it may conflict with expressibility in data that lacks full symmetry.
The paper's state-level relaxation for amplitude encoding could be tested separately to see whether the same monotonic trade-offs persist when the operator-level symmetry condition is relaxed.
The observed non-monotonicities around subgroup sizes 6 and 12 in S4, attributed to subgroup containment, hint that the lattice of subgroups, not just their sizes, shapes the results—a finer-grained analysis by subgroup isomorphism class would be a natural next step.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a solid software-engineering contribution: it ships an open-source pipeline, reproduces the Sim et al. baselines, and systematically benchmarks 19 ansatz patterns across all subgroups of S4 and a sampled subset of S5. The new content is the empirical mapping of generator drift, circuit overhead, expressibility, and entangling capability against subgroup order. That is a legitimate and useful contribution, especially the generator-drift metric and the observation that larger subgroups usually mean more overhead and less expressibility. The authors are also honest about scope: angle encoding only, no trainability study, no complex encodings. There is no fitting-to-data circularity; the twirl is deterministic and the trends are empirical.

The soft spots are real but not fatal. The stress-test concern is on target: the paper aggregates all subgroups of a given order into a single point without reporting within-order variance. For S4, order-2 subgroups split into transposition and double-transposition classes, and order-4 includes C4 and V4, which act differently on qubits. The paper's own observations about containment (size-3 inside size-6, size-12 inside size-24) show that structure beyond cardinality matters. So "subgroup size as a tunable knob" is an average statement, not a per-subgroup guarantee. A practitioner choosing a specific subgroup may see behavior that deviates from the plotted trend. This is the main load-bearing caveat.

A second, milder issue is statistical reporting: expressibility and entanglement are point estimates from 10,000 samples with no error bars or repeated-seed variance. Given the known sensitivity of KL divergence to binning and sampling, confidence intervals would strengthen the claims. For S5, the random subgroup sampling lacks coverage analysis, so some trend points may be driven by unrepresentative subgroups. The depth-1 default is fine, but the paper could say more about whether depth interactions are fully captured.

None of this undermines the central contribution. The tool works, the baseline check is credible, and the qualitative trends are supported by the plots. The paper is honest about its limitations and does not overclaim beyond its scope.

This deserves a serious referee. I would send it to peer review, asking for either error bars or at least an explicit discussion of within-order subgroup variance and a rationale for why the aggregate trend is useful despite it. For a QML audience, the reading-group value is real, but I would not cite it as authoritative on per-subgroup effects — only as evidence for the average trade-off.

Referee Report

3 major / 4 minor

Summary. The paper presents Twirlator, an automated pipeline that symmetrizes parameterized QML ansatzes with respect to subgroups of the symmetric group S_n using the Pauli twirling formula on generators. The pipeline is evaluated on 19 common ansatz patterns from [26], with four-qubit circuits under all subgroups of S_4 and five-qubit circuits under a random sample of subgroups of S_5. For each symmetrized circuit, the paper reports generator drift, synthesized circuit size/depth after Qiskit opt-level 3, expressibility (KL divergence from the Haar distribution), and Meyer-Wallach entangling capability. The central empirical claim is that larger subgroups typically increase generator drift and circuit overhead, reduce expressibility, and often increase entangling capability, while depth leaves relative subgroup trends unchanged. The authors reproduce the baseline expressibility/entanglement values of [26] for the unsymmetrized ansatzes and provide open-source code.

Significance. If the empirical trends hold, Twirlator is a useful, reusable tool that reframes symmetry as a tunable continuum rather than a binary choice, and its benchmark results give practitioners a concrete starting point for balancing symmetry-induced overhead against model capacity. The pipeline is deterministic in its core (twirling is a linear projection), it ships open-source code, and it reproduces an external baseline, all of which are strengths. However, the practical guidance in §6 rests entirely on the assumption that subgroup order is a sufficient one-dimensional proxy for the level of symmetry. The manuscript is transparent about its scope (angle encoding, permutation symmetries, no end-to-end trainability), so the missing pieces are not scope claims but statistical and aggregation details that are required to make the 'size-as-knob' recommendation trustworthy.

major comments (3)

[§3.1, §4, Figs. 3–6] Within-order variance is never reported. For S_4 the text says 'all subgroups' are considered, yet each figure plots a single scalar per subgroup order; for S_5 it says 'at most 30 random subgroups' are sampled per order, with no aggregation rule (mean? median? representative?) and no spread. Same-order subgroups can be non-isomorphic and act very differently on angle-encoded qubits (e.g., order-2 subgroups generated by a transposition vs. a double transposition; order-4 cyclic C4 vs. Klein V4). The paper's own explanations in §5.1–5.3 rely on containment structure (size 3 inside 6, size 12 inside 24) rather than order alone. Without per-subgroup variance or a justified selection rule, the central 'subgroup size as a knob' claim is under-supported; a practitioner might pick a subgroup whose behavior deviates from the average trend.
[§5.1, Eq. (3)] The generator drift definition is ambiguous and likely inconsistent. The text states that G_twirl is the symmetrized version of G, which implies a per-gate pairing, but Eq. (3) averages ||G - G_twirl|| over all combinations of G∈G(A_i) and G_twirl∈T[G(A_i)]. Unless the two sets have equal size and a canonical ordering is imposed, this double sum measures the average cross-distance between two sets, not the per-generator projection distance. Since generator drift is the RQ1 metric and the values in Fig. 3 are quantitative, the formula needs to be corrected or clarified to match the stated intent.
[§3.3, Figs. 5–6] Expressibility and entangling capability are reported as point estimates with no sampling error. The KL divergence is estimated from 10,000 sampled fidelities, but no confidence intervals, bootstrap, or multi-seed repetitions are provided. Consequently, statements such as 'subgroups of sizes 3 and 6 yield similar KL-divergence values' (§5.3) are not statistically grounded, and it is unclear whether the differences between adjacent subgroup orders are significant. Given that the main new quantitative findings are these trends, the paper should report error bars or at least a variance analysis over independent parameter samples or subgroup choices.

minor comments (4)

[Abstract] Typo: 'Twirlatormodels' is missing a space; also 'ansatzes'/'ansatze' spelling is inconsistent throughout.
[§3.3] The sentence 'which we updated 2' has an unresolved superscript/footnote; the footnote target is missing.
[§5.3] 'increased reduced expressibility' should be 'reduced expressibility'.
[Fig. 2] The circuit rendering is dense and hard to read; a higher-resolution or symbolic gate list would improve verifiability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the twirl pipeline is deterministic and metrics are measured against external baselines.

full rationale

The paper's central claims are empirical outputs of a deterministic pipeline: generators are twirled by Eq. (2), synthesized by Qiskit, and then compared on generator drift, circuit size, expressibility, and entangling capability. These quantities are not fitted parameters renamed as predictions; the twirl is a fixed projection and the metrics are external (KL divergence to Haar, Meyer-Wallach entanglement). The baseline reproduction against Sim et al. [26] confirms independence. The only author-overlapping reference is [9], used to supply PennyLane ansatz implementations; this is tooling, not a load-bearing theorem, and the ansatz patterns themselves originate from [26]. Generator drift's dependence on subgroup size is partly a mathematical property of twirling as a projection onto the commutant, but the paper reports non-monotonicities (e.g., size-6 subgroups) that show the measurements are not fully forced by construction. Within-order subgroup variance is unreported, which is a methodological limitation rather than circularity. Overall, no circular step can be exhibited; score 1 reflects only the minor self-citation in ansatz implementations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central contribution is a measurement pipeline; no new physical entities or fitted constants are introduced. The free parameters listed are inherited hyperparameters and sampling choices, not numbers tuned to make the central claim true.

free parameters (3)

Sampling budget for expressibility/entanglement = 10,000 states; 75 histogram bins
Inherited from [26], not fitted; affects absolute KL/MW values but the claimed trends are relative.
Depth-1 default for main comparisons = 1 layer
Chosen baseline; Fig. 7 shows depth does not reverse trends, so this choice is not load-bearing.
S5 subgroup sample cap = ≤30 random subgroups per order
Compute-budget choice; may under-represent diverse subgroup structures of the same order, affecting the averaged figures.

assumptions (4)

standard math Twirling generators is sufficient to symmetrize parametrized gates (Proposition 1 of [18]).
The pipeline replaces each generator G with T[G] and re-synthesizes; the correctness of this replacement is imported from [18].
domain assumption Equation (1): induced unitary representations as permutation matrices satisfy U_init(φ(s,x)) = U_s U_init(x) U_s† for angle encoding.
The implementation only constructs symmetrized circuits for diagonal angle encoding; this property is asserted, not proven for all x in the paper.
ad hoc to paper Subgroup order is a meaningful one-dimensional ordering of symmetry strength.
The central comparisons aggregate subgroups by size, but same-order subgroups can be non-isomorphic and act differently on qubits.
domain assumption Expressibility (KL divergence to Haar) and Meyer-Wallach entanglement are valid proxies for QML model quality.
The paper adopts these metrics from [26] and explicitly notes they are proxies, not task performance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Twirlator: A Pipeline for Analyzing Subgroup Symmetry Effects in Quantum Machine Learning Ansatzes." pith.science (2026). https://pith.science/paper/LCO3TD4W

@misc{pith2026251104243,
  author       = {Pith},
  title        = {Pith review of: Twirlator: A Pipeline for Analyzing Subgroup Symmetry Effects in Quantum Machine Learning Ansatzes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCO3TD4W}},
  note         = {Machine review of arXiv:2511.04243}
}

abstract

Symmetry is a strong inductive bias in geometric deep learning and its quantum counterpart, and has attracted increasing attention for improving the trainability of QML models. Yet incorporating symmetries into quantum machine learning (QML) ansatzes is not free: symmetrization often adds gates and constrains the circuits. To understand these effects, we present Twirlator, which is an automated pipeline that symmetrizes parameterized QML ansatzes and quantifies the trade-offs as the amount of symmetry increases. Twirlator models partial symmetries by the size of a subgroup of the symmetric group, enabling analysis between the ``no symmetry'' and ``full symmetry'' extremes. Across 19 common ansatz patterns, Twirlator symmetrizes circuits with respect to any subgroup of $S_n$ and measures (1) generator drift, (2) circuit overhead (depth and size), and (3) expressibility and entangling capability. The experimental evaluation focuses on subgroups of $S_4$ and $S_5$. Twirlator reveals that larger subgroups typically increase circuit overhead, reduce expressibility, and often increase entangling capability. The pipeline and results provide practical guidance for selecting ansatz patterns and symmetry levels that balance hardware cost and model performance in symmetry-aware QML applications.

Figures

Figures reproduced from arXiv: 2511.04243 by the authors.

**Figure 1.** Ansatz 3 from [26] with four qubits the results partially comparable to [26], we focus on four-qubit versions of these ansatzes. Because of selecting four qubits, we construct the symmetric group 𝑆4 and consider its possible subgroups, which have orders as 1 (trivial one-element subgroup), 2, 3, 4, 6, 8, 12, and 24 (complete 𝑆4 has 4! = 24 elements). We have considered all of the subgroups of 𝑆4. 3.2 Subgroup-based… view at source ↗

**Figure 2.** Symmetrized circuit corresponding the circuit in Fig. 1 using a four element subgroup of symmetric group [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Generator drift under subgroup-based symmetrization. We report the operator difference norm (Equation 3) between original gate [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Circuit overheads under subgroup-based symmetrization.We report the synthesized circuit size (total operations after Qiskit transpilation, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Expressibility under subgroup-based symmetrization (depth [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Entanglement capability of the ansatzes when depth is [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Depth sensitivity under subgroup-based symmetrization: depth shifts absolute values, while relative subgroup trends remain stable. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 canonical work pages

[26]

Johnson, and Alán Aspuru-Guzik

Sukin Sim, Peter D. Johnson, and Alán Aspuru-Guzik. 2019. Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum- Classical Algorithms.Advanced Quantum Technologies2, 12 (2019), 1900070. arXiv:https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/qute.201900070 doi:10.1002/qute.201900070

doi:10.1002/qute.201900070 2019
[1]

Matthew Ashman, Cristiana Diaconu, Adrian Weller, Wessel Bruinsma, and Richard E. Turner. 2025. Approximately equivariant neural processes. InProceed- ings of the 38th International Conference on Neural Information Processing Systems (Vancouver, BC, Canada)(NIPS ’24). Curran Associates Inc., Red Hook, NY, USA, Article 3078, 36 pages

2025
[2]

Utkarsh Azad and Animesh Sinha. 2023. qLEET: visualizing loss landscapes, expressibility, entangling power and training trajectories for parameterized quan- tum circuits.Quantum Information Processing22, 6 (6 2023), 256. doi:10.1007/ s11128-023-03998-z

2023
[4]

Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković

Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. 2021. Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. doi:10.48550/arXiv.2104.13478 arXiv:2104.13478 [cs]

arXiv doi:10.48550/arxiv.2104.13478 2021
[5]

Taco Cohen and Max Welling. 2016. Group Equivariant Convolutional Net- works. InProceedings of The 33rd International Conference on Machine Learn- ing (Proceedings of Machine Learning Research, Vol. 48), Maria Florina Balcan and Kilian Q. Weinberger (Eds.). PMLR, New York, New York, USA, 2990–2999. https://proceedings.mlr.press/v48/cohenc16.html

2016
[6]

Marc Finzi, Greg Benton, and Andrew Gordon Wilson. 2021. Residual pathway priors for soft equivariance constraints. InProceedings of the 35th International Conference on Neural Information Processing Systems (NIPS ’21). Curran Associates Inc., Red Hook, NY, USA, Article 2299, 13 pages

2021
[7]

2019.Quantum information in the real world

Jonas Helsen. 2019.Quantum information in the real world. Ph. D. Dissertation. TU Delft. doi:10.4233/uuid:312b719d-32bc-4219-82bb-8e6febc2abcc

work page doi:10.4233/uuid:312b719d-32bc-4219-82bb-8e6febc2abcc 2019
[8]

Cerezo, and Patrick J

Zoë Holmes, Kunal Sharma, M. Cerezo, and Patrick J. Coles. 2022. Connecting Ansatz Expressibility to Gradient Magnitudes and Barren Plateaus.PRX Quantum 3, 1 (Jan. 2022), 010313. doi:10.1103/PRXQuantum.3.010313

doi:10.1103/prxquantum.3.010313 2022

Show all 28 references

[9]

Linus Jern, Valter Uotila, Cong Yu, and Bo Zhao. 2025. Agent-Q: Fine-Tuning Large Language Models for Quantum Circuit Generation and Optimization. In 2025 IEEE International Conference on Quantum Computing and Engineering (QCE), Vol. 01. 1621–1632. doi:10.1109/QCE65121.2025.00179

2025
[10]

Kipf and Max Welling

Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. doi:10.48550/arXiv.16...

2017 doi arXiv
[11]

Kullback and R

S. Kullback and R. A. Leibler. 1951. On Information and Sufficiency.The Annals of Mathematical Statistics22, 1 (March 1951), 79–86. doi:10.1214/aoms/1177729694

1951
[12]

LeCun, B

Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, and L. D. Jackel. 1989. Backpropagation Applied to Handwritten Zip Code Recognition. Neural Computation1, 4 (Dec. 1989), 541–551. doi:10.1162/neco.1989.1.4.541

1989 doi
[13]

Lecun, L

Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. 1998. Gradient-based learning applied to document recognition.Proc. IEEE86, 11 (Nov. 1998), 2278–2324. doi:10.1109/5.726791

1998 doi
[14]

McClean, Sergio Boixo, Vadim N

Jarrod R. McClean, Sergio Boixo, Vadim N. Smelyanskiy, Ryan Babbush, and Hartmut Neven. 2018. Barren plateaus in quantum neural network training landscapes.Nature Communications9, 1 (Nov. 2018), 4812. doi:10.1038/s41467- 018-07090-4

2018 doi
[15]

2025.Group Invariance in Quantum Kernels for Vector Boson Scattering Identification at the LHC

Väinö Mehtola. 2025.Group Invariance in Quantum Kernels for Vector Boson Scattering Identification at the LHC. Master’s thesis. Aalto University. https: //aaltodoc.aalto.fi/handle/123456789/136648

2025
[16]

Antonio Anna Mele. 2024. Introduction to Haar Measure Tools in Quantum Information: A Beginner’s Tutorial.Quantum8 (May 2024), 1340. doi:10.22331/q- 2024-05-08-1340

2024 doi
[17]

Meyer and Nolan R

David A. Meyer and Nolan R. Wallach. 2002. Global entanglement in multiparticle systems.J. Math. Phys.43, 9 (9 2002), 4273–4278. doi:10.1063/1.1497700

2002 doi
[18]

Johannes Jakob Meyer, Marian Mularski, Elies Gil-Fuster, Antonio Anna Mele, Francesco Arzani, Alissa Wilms, and Jens Eisert. 2023. Exploiting symmetry in variational quantum machine learning.PRX Quantum4, 1 (March 2023), 010328. doi:10.1103/PRXQuantum.4.010328 arXiv:2205.06217...

2023 arXiv
[19]

Nguyen, Louis Schatzki, Paolo Braccia, Michael Ragone, Patrick J

Quynh T. Nguyen, Louis Schatzki, Paolo Braccia, Michael Ragone, Patrick J. Coles, Frédéric Sauvage, Martín Larocca, and M. Cerezo. 2024. Theory for Equivariant Quantum Neural Networks.PRX Quantum5, 2 (May 2024), 020328. doi:10.1103/ PRXQuantum.5.020328

2024
[20]

Isabel Nha Minh Le, Oriel Kiss, Julian Schuhmacher, Ivano Tavernelli, and Francesco Tacchino. 2025. Symmetry-invariant quantum machine learning force fields.New Journal of Physics27, 2 (Feb. 2025), 023015. doi:10.1088/1367- 2630/adad0c

2025 doi
[21]

Nguyen, Louis Schatzki, Patrick J

Michael Ragone, Paolo Braccia, Quynh T. Nguyen, Louis Schatzki, Patrick J. Coles, Frederic Sauvage, Martin Larocca, and M. Cerezo. 2023. Representation Theory for Geometric Quantum Machine Learning. doi:10.48550/arXiv.2210.07980 arXiv:2210.07980 [quant-ph]

2023 doi arXiv
[22]

Romero and Suhas Lohit

David W. Romero and Suhas Lohit. 2022. Learning partial equivariances from data (NIPS ’22). Curran Associates Inc., Red Hook, NY, USA, Article 2642, 13 pages

2022
[23]

Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. 2009. The Graph Neural Network Model.IEEE Transactions on Neural Networks20, 1 (Jan. 2009), 61–80. doi:10.1109/TNN.2008.2005605

2009
[24]

Nguyen, Frédéric Sauvage, and M

Louis Schatzki, Martín Larocca, Quynh T. Nguyen, Frédéric Sauvage, and M. Cerezo. 2024. Theoretical guarantees for permutation-equivariant quantum neural networks.npj Quantum Information10, 1 (Jan. 2024), 12. doi:10.1038/ s41534-024-00804-1

2024
[25]

2014.Understanding machine learning: from theory to algorithms

Shai Shalev-Shwartz and Shai Ben-David. 2014.Understanding machine learning: from theory to algorithms. Cambridge university press, New York

2014
[27]

Descoteaux, Mit Kotak, Gabriel de Miranda Nascimento, Seán R

Chuin Wei Tan, Marc L. Descoteaux, Mit Kotak, Gabriel de Miranda Nascimento, Seán R. Kavanagh, Laura Zichi, Menghang Wang, Aadit Saluja, Yizhong R. Hu, Tess Smidt, Anders Johansson, William C. Witt, Boris Kozinsky, and Albert Musaelian. 2025. High-performance training and infe...

2025 arXiv
[28]

Nurminen

Valter Uotila, Ilmo Salmenpera, Leo Becker, Arianne Meijer van de Griend, Aakash Ravindra Shinde, and Jukka K. Nurminen. 2025. Perspectives on Uti- lization of Measurements in Quantum Algorithms . In2025 IEEE International Conference on Quantum Software (QSW). IEEE Computer So...

2025
[29]

West, Jamie Heredge, Martin Sevior, and Muhammad Usman

Maxwell T. West, Jamie Heredge, Martin Sevior, and Muhammad Usman. 2024. Provably Trainable Rotationally Equivariant Quantum Machine Learning.PRX Quantum5, 3 (July 2024), 030320. doi:10.1103/PRXQuantum.5.030320

2024 doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [26]

Johnson, and Alán Aspuru-Guzik

Sukin Sim, Peter D. Johnson, and Alán Aspuru-Guzik. 2019. Expressibility and Entangling Capability of Parameterized Quantum Circuits for Hybrid Quantum- Classical Algorithms.Advanced Quantum Technologies2, 12 (2019), 1900070. arXiv:https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/qute.201900070 doi:10.1002/qute.201900070

doi:10.1002/qute.201900070 2019

[2] [1]

Matthew Ashman, Cristiana Diaconu, Adrian Weller, Wessel Bruinsma, and Richard E. Turner. 2025. Approximately equivariant neural processes. InProceed- ings of the 38th International Conference on Neural Information Processing Systems (Vancouver, BC, Canada)(NIPS ’24). Curran Associates Inc., Red Hook, NY, USA, Article 3078, 36 pages

2025

[3] [2]

Utkarsh Azad and Animesh Sinha. 2023. qLEET: visualizing loss landscapes, expressibility, entangling power and training trajectories for parameterized quan- tum circuits.Quantum Information Processing22, 6 (6 2023), 256. doi:10.1007/ s11128-023-03998-z

2023

[4] [4]

Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković

Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. 2021. Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. doi:10.48550/arXiv.2104.13478 arXiv:2104.13478 [cs]

arXiv doi:10.48550/arxiv.2104.13478 2021

[5] [5]

Taco Cohen and Max Welling. 2016. Group Equivariant Convolutional Net- works. InProceedings of The 33rd International Conference on Machine Learn- ing (Proceedings of Machine Learning Research, Vol. 48), Maria Florina Balcan and Kilian Q. Weinberger (Eds.). PMLR, New York, New York, USA, 2990–2999. https://proceedings.mlr.press/v48/cohenc16.html

2016

[6] [6]

Marc Finzi, Greg Benton, and Andrew Gordon Wilson. 2021. Residual pathway priors for soft equivariance constraints. InProceedings of the 35th International Conference on Neural Information Processing Systems (NIPS ’21). Curran Associates Inc., Red Hook, NY, USA, Article 2299, 13 pages

2021

[7] [7]

2019.Quantum information in the real world

Jonas Helsen. 2019.Quantum information in the real world. Ph. D. Dissertation. TU Delft. doi:10.4233/uuid:312b719d-32bc-4219-82bb-8e6febc2abcc

work page doi:10.4233/uuid:312b719d-32bc-4219-82bb-8e6febc2abcc 2019

[8] [8]

Cerezo, and Patrick J

Zoë Holmes, Kunal Sharma, M. Cerezo, and Patrick J. Coles. 2022. Connecting Ansatz Expressibility to Gradient Magnitudes and Barren Plateaus.PRX Quantum 3, 1 (Jan. 2022), 010313. doi:10.1103/PRXQuantum.3.010313

doi:10.1103/prxquantum.3.010313 2022