arxiv: 2604.04828 · v1 · submitted 2026-04-06 · 🪐 quant-ph · cs.CE· cs.LG· physics.comp-ph

Recognition: no theorem link

Hybrid Fourier Neural Operator for Surrogate Modeling of Laser Processing with a Quantum-Circuit Mixer

Mateusz Papierz , Asel Sagingalieva , Alix Benoit , Toni Ivas , Elia Iseli , Alexey Melnikov

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:05 UTC · model grok-4.3

classification 🪐 quant-ph cs.CEcs.LGphysics.comp-ph

keywords hybrid quantum-classicalFourier Neural Operatorvariational quantum circuitsurrogate modelinglaser processingmultiphysics simulationneural operatorsquantum machine learning

0 comments

The pith

Replacing dense spectral blocks in Fourier Neural Operators with a mode-shared quantum circuit mixer reduces parameters by 15.6% and improves accuracy for laser processing simulations.

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

The paper introduces HQ-LP-FNO, a hybrid quantum-classical Fourier Neural Operator designed for surrogate modeling of three-dimensional laser processing that includes heat transfer, convection, deformation, and phase change. Standard FNOs incur parameter costs that scale linearly with the number of Fourier modes because of dense mode-wise channel mixing, which becomes prohibitive for 3D problems. The new architecture substitutes a configurable fraction of these mixing blocks with a compact variational quantum circuit that shares parameters across modes, so the added parameter count stays fixed regardless of mode budget. A parameter-matched classical control architecture is used to isolate the contribution of the quantum mixer. On the target multiphysics dataset the hybrid model records lower errors in both phase fraction and temperature predictions while using fewer total trainable weights.

Core claim

By replacing a fraction of the dense mode-wise spectral blocks in an FNO with a mode-shared variational quantum circuit mixer whose parameter count is independent of the Fourier mode budget, HQ-LP-FNO reduces trainable parameters by 15.6% relative to a classical baseline, lowers phase-fraction mean absolute error by 26%, and reduces relative temperature MAE from 2.89% to 2.56% on a three-dimensional laser-processing surrogate task.

What carries the argument

The mode-shared variational quantum circuit (VQC) mixer, which performs spectral channel mixing with a fixed parameter count independent of the number of retained Fourier modes.

Load-bearing premise

The observed reductions in parameters and errors arise from the mode-sharing structure of the variational quantum circuit rather than from the particular classical bottleneck design or from favorable hyper-parameter choices on this dataset.

What would settle it

A purely classical parameter-efficient mixer that matches the compactness of the VQC and produces equal or larger error reductions on the same laser-processing dataset would show that the quantum circuit is not required for the reported gains.

Figures

Figures reproduced from arXiv: 2604.04828 by Alexey Melnikov, Alix Benoit, Asel Sagingalieva, Elia Iseli, Mateusz Papierz, Toni Ivas.

**Figure 2.** Figure 2: FIG. 2: HQ-LP-FNO for 3D laser–material heat transfer. (a) Input features are lifted by a pointwise projection [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3: Qualitative comparison of classical and hybrid predictions at [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4: Temperature error maps across process space at [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5: (a) Average FIM showing diagonal dominance. (b) FIM rank vs. number of trainable layers. (c) Eigenvalue distribution [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6: Mean squared error between the noiseless and noisy outputs of the quantum spectral mixer as a function of the number [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

**Figure 9.** Figure 9: FIG. 9 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

**Figure 10.** Figure 10: FIG. 10 [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗

read the original abstract

Data-driven surrogates can replace expensive multiphysics solvers for parametric PDEs, yet building compact, accurate neural operators for three-dimensional problems remains challenging: in Fourier Neural Operators, dense mode-wise spectral channel mixing scales linearly with the number of retained Fourier modes, inflating parameter counts and limiting real-time deployability. We introduce HQ-LP-FNO, a hybrid quantum-classical FNO that replaces a configurable fraction of these dense spectral blocks with a compact, mode-shared variational quantum circuit mixer whose parameter count is independent of the Fourier mode budget. A parameter-matched classical bottleneck control is co-designed to provide a rigorous evaluation framework. Evaluated on three-dimensional surrogate modeling of high-energy laser processing, coupling heat transfer, melt-pool convection, free-surface deformation, and phase change, HQ-LP-FNO reduces trainable parameters by 15.6% relative to a classical baseline while lowering phase-fraction mean absolute error by 26% and relative temperature MAE from 2.89% to 2.56%. A sweep over the quantum-channel budget reveals that a moderate VQC allocation yields the best temperature metrics across all tested configurations, including the fully classical baseline, pointing toward an optimal classical-quantum partitioning. The ablation confirms that mode-shared mixing, naturally implemented by the VQC through its compact circuit structure, is the dominant contributor to these improvements. A noisy-simulator study under backend-calibrated noise from ibm-torino confirms numerical stability of the quantum mixer across the tested shot range. These results demonstrate that VQC-based parameter-efficient spectral mixing can improve neural operator surrogates for complex multiphysics problems and establish a controlled evaluation protocol for hybrid quantum operator learning in practice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The hybrid replaces some FNO spectral mixers with a mode-shared VQC to hold parameters fixed as mode count grows, and it reports modest accuracy gains on a laser-processing surrogate with a matched classical control.

read the letter

The paper replaces a fraction of the dense mode-wise mixing blocks in a Fourier Neural Operator with a variational quantum circuit that shares its parameters across retained Fourier modes. This keeps the total parameter count from scaling with the mode budget, which is the usual bottleneck for 3D FNOs. They evaluate the resulting HQ-LP-FNO on a multiphysics laser-processing surrogate that includes heat transfer, melt-pool flow, free-surface deformation, and phase change. The hybrid version uses 15.6% fewer parameters than a parameter-matched classical baseline, cuts phase-fraction mean absolute error by 26%, and lowers relative temperature error from 2.89% to 2.56%. They also run a quantum-channel-budget sweep and a noisy-simulator test on ibm-torino-calibrated noise, plus an ablation that points to the mode-shared mixing as the main driver.

Referee Report

3 major / 2 minor

Summary. The paper introduces HQ-LP-FNO, a hybrid quantum-classical Fourier Neural Operator for 3D surrogate modeling of laser processing multiphysics (heat transfer, melt-pool convection, phase change). It replaces a fraction of dense mode-wise spectral mixing blocks in FNOs with a mode-shared variational quantum circuit (VQC) mixer whose parameter count is independent of the retained Fourier mode budget. A parameter-matched classical bottleneck control is used for comparison. On the target dataset, HQ-LP-FNO reduces trainable parameters by 15.6%, phase-fraction MAE by 26%, and relative temperature MAE from 2.89% to 2.56%. An ablation attributes gains to the mode-shared mixing, a quantum-channel-budget sweep identifies an optimal hybrid allocation, and a noisy-simulator study with ibm-torino calibration shows stability.

Significance. If the central comparison holds, the work demonstrates a concrete route to parameter-efficient neural operators for high-dimensional multiphysics PDEs by leveraging VQC structure for spectral mixing. The explicit parameter-matched classical control, ablation isolating mode-shared mixing, and backend-calibrated noise study are strengths that establish a reproducible evaluation protocol for hybrid quantum operator learning. These elements could influence surrogate modeling in engineering domains where real-time deployment is limited by parameter scaling with mode count.

major comments (3)

[Evaluation protocol (as described in abstract and §4)] The headline comparison rests on the parameter-matched classical bottleneck control, yet the manuscript provides no explicit replacement rule or architecture diagram showing what classical layer (e.g., reduced-rank linear mixer or bottleneck MLP) occupies the identical parameter budget and slot as the VQC. Without this, the reported 15.6% parameter reduction and MAE improvements cannot be unambiguously attributed to the mode-shared VQC rather than incidental changes in capacity or optimization landscape.
[Results, quantum-channel-budget sweep paragraph] The quantum-channel-budget sweep is performed only on the hybrid model; the classical baseline is not re-optimized under the identical sweep protocol. This leaves open the possibility that the temperature MAE improvement (2.89% → 2.56%) and the claim of an optimal classical-quantum partitioning arise from unequal hyper-parameter effort rather than the VQC mixer itself.
[Ablation study subsection] The ablation study confirming that 'mode-shared mixing ... is the dominant contributor' is load-bearing for the central claim, but the manuscript does not report the precise classical ablation variant (e.g., whether it retains the same bottleneck width or uses a different mixing mechanism) or the quantitative delta when mode-sharing is removed while keeping parameter count fixed.

minor comments (2)

[Results tables/figures] Error bars or standard deviations across random seeds or data splits are not mentioned in the reported MAE values; adding them would strengthen the numerical claims.
[Methods, VQC description] Notation for the VQC circuit depth, number of qubits, and exact ansatz (e.g., hardware-efficient or problem-inspired) should be defined once in a dedicated subsection rather than scattered across text and figures.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review, which highlights both the strengths of our evaluation protocol and areas where additional clarity is needed. We address each major comment point by point below, providing explanations and committing to revisions that will strengthen the manuscript without altering its core claims.

read point-by-point responses

Referee: [Evaluation protocol (as described in abstract and §4)] The headline comparison rests on the parameter-matched classical bottleneck control, yet the manuscript provides no explicit replacement rule or architecture diagram showing what classical layer (e.g., reduced-rank linear mixer or bottleneck MLP) occupies the identical parameter budget and slot as the VQC. Without this, the reported 15.6% parameter reduction and MAE improvements cannot be unambiguously attributed to the mode-shared VQC rather than incidental changes in capacity or optimization landscape.

Authors: The referee is correct that the manuscript does not provide an explicit replacement rule or architecture diagram for the classical bottleneck control. While the abstract and Section 3 describe it as a parameter-matched control, the precise implementation (a bottleneck MLP whose hidden dimension is tuned to match the VQC parameter count) and its placement relative to the VQC mixer are not diagrammed. We will revise Section 3 and Figure 2 to include the explicit rule and a side-by-side architecture diagram, ensuring the 15.6% reduction and MAE gains can be unambiguously attributed to the mode-shared VQC. revision: yes
Referee: [Results, quantum-channel-budget sweep paragraph] The quantum-channel-budget sweep is performed only on the hybrid model; the classical baseline is not re-optimized under the identical sweep protocol. This leaves open the possibility that the temperature MAE improvement (2.89% → 2.56%) and the claim of an optimal classical-quantum partitioning arise from unequal hyper-parameter effort rather than the VQC mixer itself.

Authors: We agree that the sweep protocol was applied only to hybrid configurations and that the classical baseline was not re-optimized under the same conditions. This is a valid concern regarding potential differences in hyperparameter effort. We will conduct the additional classical optimizations across the channel-budget sweep and incorporate the comparative results into the revised results section to confirm that the reported MAE improvements and optimal partitioning are robust. revision: yes
Referee: [Ablation study subsection] The ablation study confirming that 'mode-shared mixing ... is the dominant contributor' is load-bearing for the central claim, but the manuscript does not report the precise classical ablation variant (e.g., whether it retains the same bottleneck width or uses a different mixing mechanism) or the quantitative delta when mode-sharing is removed while keeping parameter count fixed.

Authors: The referee accurately notes that the ablation subsection lacks precise details on the classical variant and the quantitative deltas under fixed parameter count. The current ablation compares the VQC to a non-shared classical mixer, but the exact bottleneck width and per-mode mechanism are not specified, nor are the exact MAE deltas reported. We will expand the ablation subsection to define the classical ablation variant explicitly (non-shared per-mode mixers with matched bottleneck width) and include the quantitative MAE deltas when mode-sharing is removed at fixed parameter count, thereby strengthening support for the dominance of mode-shared mixing. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain; results are empirical measurements against controlled baselines

full rationale

The paper proposes an architectural modification to the Fourier Neural Operator by substituting a subset of dense spectral mixing layers with a mode-shared variational quantum circuit whose parameter count does not scale with retained Fourier modes. This substitution produces a lower total parameter count by design, but the paper does not treat the resulting error reductions as a mathematical prediction; instead it reports measured MAE values on a held-out multiphysics dataset together with a parameter-matched classical bottleneck control and an explicit ablation isolating the mode-sharing effect. No equation equates a claimed performance gain to a fitted constant defined from the same data, no uniqueness theorem is imported via self-citation, and no ansatz is smuggled through prior work. The evaluation therefore remains an ordinary empirical comparison rather than a self-referential derivation.

Axiom & Free-Parameter Ledger

2 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical performance of the hybrid architecture; no new physical axioms or invented particles are introduced. The only free parameters are the usual neural-network weights plus the VQC angles, all fitted to the laser-processing data.

free parameters (2)

VQC rotation angles
Variational parameters of the quantum circuit mixer, fitted during training.
Classical bottleneck weights
Weights of the parameter-matched classical control network.

pith-pipeline@v0.9.0 · 5633 in / 1285 out tokens · 32889 ms · 2026-05-10T19:05:16.895414+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

69 extracted references · 11 canonical work pages · 3 internal anchors

[1]

ˆv(q) ℓ (k) ˆv(c) ℓ (k) # =

Partitioned hybrid spectral convolution We modify only the learnable spectral operatorK ℓ in Eq. (7). Let Ω denote the retained low-frequency index set in therfftnrepresentation (standard four-corner truncation in (k x, ky) and one-sided frequencies along the rFFT axis) [3, 4]. Using an integerquantum-channel widthC q ∈ {0,1, . . . , C}, we split theoutpu...
[2]

Quantum spectral mixer block The quantum branch operates on complex Fourier coefficients over the firstC q channels. For each modek∈Ω, we construct a real feature vector by concatenating real and imaginary parts, rℓ(k) = ℜˆu(q) ℓ (k)ℑˆu (q) ℓ (k) ∈R 2Cq .(11) Because direct mapping of Fourier coefficients to rotation angles can induce phase wrapping and u...

work page arXiv
[3]

Michaleris, Finite Elements in Analysis and Design86, 51 (2014)

P. Michaleris, Finite Elements in Analysis and Design86, 51 (2014)

2014
[4]

Mukherjee, H

T. Mukherjee, H. L. Wei, A. De, and T. DebRoy, Computational Materials Science150, 304 (2018)

2018
[5]

Kovachki, Z

N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar, Journal of Machine Learning Research24, 1 (2023), URLhttps://www.jmlr.org/papers/v24/21-1524.html

2023
[6]

Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar, inInternational Con- ference on Learning Representations(2021), URLhttps://openreview.net/forum?id=c8P9NQVtmnO

2021
[7]

Benoit, T

A. Benoit, T. Ivas, M. Papierz, A. Sagingalieva, A. Melnikov, and E. Iseli,A fast and generalizable Fourier neural operator- based surrogate for melt-pool prediction in laser processing, arXiv preprint arXiv:2602.06241 (2026), URLhttps://arxiv. org/abs/2602.06241

work page arXiv 2026
[8]

Melnikov, M

A. Melnikov, M. Kordzanganeh, A. Alodjants, and R.-K. Lee, Advances in Physics: X8, 2165452 (2023)

2023
[9]

Schuld and N

M. Schuld and N. Killoran, Physical Review Letters122, 040504 (2019)

2019
[10]

Havl´ ıˇ cek, A

V. Havl´ ıˇ cek, A. C´ orcoles, K. Temme, A. Harrow, A. Kandala, J. Chow, and J. Gambetta, Nature567, 209 (2019)

2019
[11]

Schuld, R

M. Schuld, R. Sweke, and J. J. Meyer, Physical Review A103, 032430 (2021)

2021
[12]

P´ erez-Salinas, A

A. P´ erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, Quantum4, 226 (2020)

2020
[13]

T. Goto, Q. H. Tran, and K. Nakajima, Physical Review Letters127, 090506 (2021)

2021
[14]

Jerbi, L

S. Jerbi, L. J. Fiderer, H. Poulsen Nautrup, et al., Nature Communications14, 517 (2023)

2023
[15]

Kordzanganeh, P

M. Kordzanganeh, P. Sekatski, L. Fedichkin, and A. Melnikov, Machine Learning: Science and Technology4, 035036 (2023)

2023
[16]

M. A. Nielsen and I. L. Chuang,Quantum Computation and Quantum Information: 10th Anniversary Edition(Cambridge University Press, Cambridge, 2010)

2010
[17]

F. Fan, Y. Shi, T. Guggemos, and X. X. Zhu, IEEE Transactions on Neural Networks and Learning Systems35, 18145 (2024)

2024
[18]

Landman, N

J. Landman, N. Mathur, Y. Y. Li, M. Strahm, S. Kazdaghli, A. Prakash, and I. Kerenidis, Quantum6, 881 (2022)

2022
[19]

Ajlouni, A

N. Ajlouni, A. ¨Ozyava¸ s, M. Takao˘ glu, F. Takao˘ glu, and F. Ajlouni, BMC Medical Imaging23, 126 (2023)

2023
[20]

Xiang, D

Q. Xiang, D. Li, Z. Hu, et al., Scientific Reports14, 24699 (2024)

2024
[21]

Rainjonneau, I

S. Rainjonneau, I. Tokarev, S. Iudin, S. Rayaprolu, K. Pinto, et al., IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing16, 7062 (2023)

2023
[22]

Sagingalieva, M

A. Sagingalieva, M. Kordzanganeh, A. Kurkin, A. Melnikov, D. Kuhmistrov, et al., Quantum Machine Intelligence5, 38 (2023)

2023
[23]

Lusnig, A

L. Lusnig, A. Sagingalieva, M. Surmach, T. Protasevich, O. Michiu, J. McLoughlin, C. Mansell, G. de’Petris, D. Bonazza, F. Zanconati, et al., Diagnostics14, 558 (2024)

2024
[24]

Anoshin, A

M. Anoshin, A. Sagingalieva, C. Mansell, D. Zhiganov, V. Shete, M. Pflitsch, and A. Melnikov, IEEE Transactions on Quantum Engineering5, 2500514 (2024)

2024
[25]

N. Lee, M. Shin, A. Sagingalieva, A. Senokosov, M. Anoshin, A. J. Tripathi, K. Pinto, and A. Melnikov,Predictive control of blast furnace temperature in steelmaking with hybrid depth-infused quantum neural networks, arXiv preprint arXiv:2504.12389 (2025), URLhttps://arxiv.org/abs/2504.12389

work page arXiv 2025
[26]

Kurkin, J

A. Kurkin, J. Hegemann, M. Kordzanganeh, and A. Melnikov, Engineering Applications of Artificial Intelligence160, 111912 (2025)

2025
[27]

Y. Xiao, L. M. Yang, C. Shu, S. C. Chew, B. C. Khoo, Y. D. Cui, and Y. Y. Liu, Physics of Fluids36, 097145 (2024)

2024
[28]

Berger, N

S. Berger, N. Hosters, and M. M¨ oller, Scientific Reports15, 18823 (2025)

2025
[29]

Sedykh, M

A. Sedykh, M. Podapaka, A. Sagingalieva, K. Pinto, M. Pflitsch, and A. Melnikov, Machine Learning: Science and Technology5, 025045 (2024)

2024
[30]

Sedykh, T

A. Sedykh, T. Protasevich, M. Surmach, A. Senokosov, M. Anoshin, A. Sagingalieva, and A. Melnikov,Multi-stream physics hybrid networks for solving Navier-Stokes equations, arXiv preprint arXiv:2504.01891 (2025), URLhttps://arxiv.org/ abs/2504.01891

work page arXiv 2025
[31]

Flow Science, Inc.,FLOW-3D WELD 2025r1,https://www.flow3d.com/whats-new-in-flow-3d-weld/(2025), accessed: 2026-01-22

2025
[32]

Svenungsson, I

J. Svenungsson, I. Choquet, and A. F. H. Kaplan, Physics Procedia78, 182 (2015)

2015
[33]

Aggarwal, Y

A. Aggarwal, Y. C. Shin, and A. Kumar, International Journal of Heat and Mass Transfer219, 124841 (2024)

2024
[34]

T. F. Flint, J. D. Robson, G. Parivendhan, and P. Cardiff, SoftwareX21, 101299 (2023)

2023
[35]

D. B. Hann, J. Iammi, and J. Folkes, Journal of Physics D: Applied Physics44, 445401 (2011)

2011
[36]

Boivineau, C

M. Boivineau, C. Cagran, D. Doytier, V. Eyraud, M. H. Nadal, B. Wilthan, and G. Pottlacher, International Journal of Thermophysics27, 507 (2006)

2006
[37]

Zhang, Y

B. Zhang, Y. Li, and Q. Bai, Materials & Design196, 109147 (2020)

2020
[38]

Gaussian Error Linear Units (GELUs)

D. Hendrycks and K. Gimpel,Gaussian error linear units (GELUs), arXiv preprint arXiv:1606.08415 (2016), URLhttps: //arxiv.org/abs/1606.08415. 16

work page internal anchor Pith review Pith/arXiv arXiv 2016
[39]

K¨ olle, T

M. K¨ olle, T. Witter, T. Rohe, G. Stenzel, P. Altmann, and T. Gabor, in2024 IEEE International Conference on Quantum Software (QSW)(2024), pp. 157–167

2024
[40]

Marcandelli, Y

P. Marcandelli, Y. He, S. Mariani, M. Siena, and S. Markidis, in2025 IEEE International Conference on Quantum Computing and Engineering (QCE)(2025), pp. 1676–1687

2025
[41]

N. Jain, J. Landman, N. Mathur, and I. Kerenidis, Quantum Science and Technology9, 035026 (2024)

2024
[42]

Kordzanganeh, D

M. Kordzanganeh, D. Kosichkina, and A. Melnikov, Intelligent Computing2, 0028 (2023)

2023
[43]

D. M. Abrams, N. Didier, B. R. Johnson, M. P. d. Silva, and C. A. Ryan, Nature Electronics3, 744 (2020)

2020
[44]

PennyLane: Automatic differentiation of hybrid quantum-classical computations

V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso-Linaje, A. Akash, A. Asadi, et al.,PennyLane: Automatic differentiation of hybrid quantum-classical computations, arXiv preprint arXiv:1811.04968 (2018), URLhttps://arxiv.org/abs/1811.04968

work page internal anchor Pith review Pith/arXiv arXiv 2018
[45]

Kuzmin, B

V. Kuzmin, B. Kyriacou, T. Protasevich, M. Papierz, M. Kordzanganeh, and A. Melnikov,TQml simulator: optimized simulation of quantum machine learning, arXiv preprint arXiv:2506.04891 (2025), URLhttps://arxiv.org/abs/2506. 04891

work page arXiv 2025
[46]

Y. Liu, S. Arunachalam, and K. Temme, Nature Physics17, 1013 (2021)

2021
[47]

Kontolati, S

K. Kontolati, S. Goswami, M. D. Shields, and G. E. Karniadakis, Journal of Computational Physics479, 112008 (2023)

2023
[48]

Z. Xiao, S. Kou, Z. Hao, B. Lin, and Z. Deng, inAdvances in Neural Information Processing Sys- tems(2024), vol. 37, pp. 115001–115020, URLhttps://proceedings.neurips.cc/paper_files/paper/2024/hash/ d06a797c436cd5136a6f45b063316278-Abstract-Conference.html

2024
[49]

A. Tran, A. Mathews, L. Xie, and C. S. Ong, inInternational Conference on Learning Representations(2023), URL https://openreview.net/forum?id=tmIiMPl4IPa

2023
[50]

Patapovich, M

V. Patapovich, M. Periyasamy, M. Kordzanganeh, and A. Melnikov,Superposed parameterised quantum circuits, arXiv preprint arXiv:2506.08749 (2025), URLhttps://arxiv.org/abs/2506.08749

work page arXiv 2025
[51]

Coecke and R

B. Coecke and R. Duncan, inAutomata, Languages and Programming(Springer, 2008), vol. 5126 ofLecture Notes in Computer Science, pp. 298–310

2008
[52]

ZX-calculus for the working quantum computer scientist

J. van de Wetering,ZX-calculus for the working quantum computer scientist, arXiv preprint arXiv:2012.13966 (2020), URL https://arxiv.org/abs/2012.13966

work page arXiv 2012
[53]

Abbas, D

A. Abbas, D. Sutter, C. Zoufal, A. Lucchi, A. Figalli, and S. Woerner, Nature Computational Science1, 403 (2021)

2021
[54]

Amari, Neural Computation10, 251 (1998)

S.-i. Amari, Neural Computation10, 251 (1998)

1998
[55]

Haboury, M

N. Haboury, M. Kordzanganeh, A. Melnikov, and P. Sekatski,Information plane and compression-gnostic feedback in quantum machine learning, arXiv preprint arXiv:2411.02313 (2024), URLhttps://arxiv.org/abs/2411.02313

work page arXiv 2024
[56]

Holmes, K

Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, PRX Quantum3, 010313 (2022)

2022
[57]

Larocca, N

M. Larocca, N. Ju, D. Garc´ ıa-Mart´ ın, P. J. Coles, and M. Cerezo, Nature Computational Science3, 542 (2023)

2023
[58]

J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Babbush, and H. Neven, Nature Communications9, 4812 (2018)

2018
[59]

IBM Quantum,IBM Quantum,https://quantum.cloud.ibm.com/(2026), accessed: 2026-03-03

2026
[60]

Preskill, Quantum2, 79 (2018)

J. Preskill, Quantum2, 79 (2018)

2018
[61]

Kuzmin, W

V. Kuzmin, W. Somogyi, E. Pankovets, and A. Melnikov, Advanced Quantum Technologies8, e00603 (2025)

2025
[62]

Quantum computing with Qiskit

A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, et al.,Quantum computing with Qiskit, arXiv preprint arXiv:2405.08810 (2024), URLhttps://arxiv.org/abs/2405.08810

work page internal anchor Pith review Pith/arXiv arXiv 2024
[63]

Z. Cai, R. Babbush, S. C. Benjamin, S. Endo, W. J. Huggins, Y. Li, J. R. McClean, T. E. O’Brien, et al., Reviews of Modern Physics95, 045005 (2023)

2023
[64]

Bischof and M

R. Bischof and M. A. Kraus, Computer Methods in Applied Mechanics and Engineering439, 117914 (2025)

2025
[65]

X. Chen, C. Liang, D. Huang, E. Real, K. Wang, H. Pham, X. Dong, T. Luong, C.-J. Hsieh, Y. Lu, et al., inAdvances in Neural Information Processing Systems(2023), vol. 36, URLhttps://proceedings.neurips.cc/paper_files/paper/ 2023/hash/9a39b4925e35cf447ccba8757137d84f-Abstract-Conference.html. 17 Appendix A: Quantum Circuit Analysis This section analyzes th...

2023
[66]

Within this formalism, a circuit is mapped to a graph comprising nodes (“spiders”) connected by edges

Redundancy analysis via ZX calculus ZX calculus is a graphical language for representing and simplifying quantum circuits [49]. Within this formalism, a circuit is mapped to a graph comprising nodes (“spiders”) connected by edges. Rewrite rules are then applied to eliminate redundant elements [50], and the reduced graph is compared with the original to qu...
[67]

Trainability assessment via Fisher information We assess trainability using the Fisher information matrix (FIM), which quantifies the sensitivity of the circuit output distribution to parameter changes [51–53]. The FIM defines a Riemannian metric on the parameter manifold: F(θ) =E {xi,yi} ∇θ logP(y|x, θ)∇ θ logP(y|x, θ) T .(A1) To analyze local sensitivit...
[68]

[9] shows that the output of a parameterized quantum circuit may be expressed as a truncated Fourier series

Expressivity assessment via Fourier series Ref. [9] shows that the output of a parameterized quantum circuit may be expressed as a truncated Fourier series. For a feature vectorxof lengthN, this representation takes the form fθ(x) = X ω1∈Ω1 · · · X ωN ∈ΩN cω1,...,ωN (θ)e −i ω·x,(A3) where each frequency component satisfiesω k ∈ {−d, . . . ,0, . . . , d}. ...
[69]

Summary The quantum circuit implemented in this paper exhibits a low degree of redundancy and a high degree of trainability, with a majority of parameters remaining active and contributing substantially to performance. Since a non-negligible fraction of near-zero eigenvalues is observed (suggesting potential susceptibility to barren-plateau effects under ...