pith. sign in

arxiv: 2605.23403 · v1 · pith:PZIKX7GHnew · submitted 2026-05-22 · 💻 cs.LG · physics.ao-ph· quant-ph

Hybrid Quantum-Classical Corrective Diffusion Modeling for Meteorological Downscaling

Rui Wang , Edoardo Pasetto , Amer Delilbasic , Morris Riedel , Kristel Michielsen , Gabriele Cavallaro This is my paper

Pith reviewed 2026-05-25 05:04 UTC · model grok-4.3

classification 💻 cs.LG physics.ao-phquant-ph
keywords hybrid quantum-classicalcorrective diffusionstatistical downscalingvariational quantum circuitswind field generationprobabilistic downscalingUNet bottleneckmeteorological modeling
0
0 comments X

The pith

Placing variational quantum circuit layers in the UNet bottleneck improves MAE and CRPS for probabilistic wind downscaling.

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

The paper tests a hybrid quantum-classical corrective diffusion model for reconstructing high-resolution weather fields from coarse inputs. It inserts variational quantum circuit layers only into the most compressed part of the diffusion UNet while keeping the regression branch classical, to check whether these layers can serve as compact nonlinear feature maps. On the 2020 validation set the hybrid versions stay stable, keep the large-scale spatial patterns of the wind fields, and beat the classical baseline on both MAE and CRPS in several configurations. They also maintain similar kinetic-energy spectra and wind-speed distributions, though they alter tail behavior and extreme-wind localization in controlled ways. Gains observed on 2020 data do not carry over uniformly to the 2021 out-of-distribution test, exposing a generalization gap.

Core claim

The central claim is that bottleneck-level quantum hybridization makes a nontrivial contribution to weather statistical downscaling: the hybrid models remain stable, preserve large-scale spatial organization of generated wind fields, and improve both MAE and CRPS relative to a classical corrective diffusion model on the 2020 validation set, while structural diagnostics show comparable kinetic-energy spectra and windspeed distributions with controlled changes in extremes.

What carries the argument

Variational quantum circuit layers inserted into the compressed bottleneck of the diffusion UNet, functioning as compact nonlinear feature maps for latent-channel mixing while the rest of the architecture stays classical.

If this is right

  • Hybrid models improve both MAE and CRPS relative to the classical corrective diffusion model in several configurations on the 2020 validation set.
  • The hybrid variants preserve kinetic-energy spectra and windspeed distributions similar to the classical counterpart while producing controlled changes in tail behavior and extreme-windspeed localization.
  • Backend studies show negligible impact from simulated device noise at the tested circuit scale.
  • Real-hardware deployment remains limited by qubit availability and execution fidelity.
  • In-distribution gains do not transfer uniformly to the 2021 out-of-distribution test, revealing a generalization gap.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bottleneck placement strategy could be tested on other diffusion-based generative tasks that involve latent mixing in scientific domains such as fluid simulation.
  • The observed generalization gap suggests that adding explicit stabilization or regularization during training might help the hybrid models handle temporal shifts in weather data.
  • If qubit counts and fidelity improve, scaling the circuit depth or width inside the bottleneck could be checked for further metric gains without redesigning the full network.
  • The results point to a broader pattern in which quantum feature maps are most useful when confined to narrow latent stages rather than spread across an entire model.

Load-bearing premise

Inserting variational quantum circuit layers specifically into the UNet bottleneck will produce measurable improvements in downscaling metrics without any other changes to the architecture or training procedure.

What would settle it

A side-by-side evaluation on the 2020 validation set in which the hybrid models show no reduction in MAE or CRPS compared with the classical baseline, or in which they lose stability or large-scale spatial organization.

Figures

Figures reproduced from arXiv: 2605.23403 by Amer Delilbasic, Edoardo Pasetto, Gabriele Cavallaro, Kristel Michielsen, Morris Riedel, Rui Wang.

Figure 2
Figure 2. Figure 2: Processing scheme of the customized PyTorch class [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: High-level schematic of the CorrDiff workflow pro [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the HQConv ansatz. Panels (a)–(d) show the full circuit structure from [21], the encoding block, and the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative 10 m wind-speed comparisons for the classical and hybrid CorrDiff models under matched bottleneck [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Structural realism and extreme-event localization di [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Joint probability density of the horizontal 10 m wind components [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Differences in MAE and CRPS between predictions [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Wind map prediction for the datestamp 14.04.2020 (a) [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
read the original abstract

Statistical downscaling is a crucial component of the weather modeling field, where high-resolution outputs must be reconstructed from coarse-resolution inputs with the full cost of dynamical refinement. In this work, we investigate a hybrid quantum-classical corrective diffusion model for probabilistic statistical downscaling of weather fields. The proposed model inserts variational quantum circuit layers into the most compressed bottleneck of the diffusion UNet while leaving the regression branch fully classical. This placement tests whether quantum circuits can act as compact nonlinear feature maps for latent-channel mixing. We evaluate intra-channel and cross-channel ans\"atze on 10m wind components. On the 2020 validation set, the hybrid models remain stable, preserve the large-scale spatial organization of the generated wind fields, and improve both MAE and CRPS relative to a classical corrective diffusion model in several configurations. Structural diagnostics further show that the hybrid variants preserve kinetic-energy spectra and windspeed distributions similar to its classical counterpart while producing controlled changes in tail behavior, extreme-windspeed localization, and joint wind field components structure. Backend studies on the 2020 validation set show negligible impact from simulated device noise at the tested circuit scale, whereas real-hardware deployment remains limited by qubit availability and execution fidelity. The 2021 out-of-distribution test shows that these in-distribution gains do not transfer uniformly under temporal shift, revealing a generalization gap that motivates future mitigation through stabilization and regularization. These results show that bottleneck-level quantum hybridization can make a nontrivial contribution to weather statistical downscaling, while also highlighting that circuit scale and hardware deployment remain key limiting factors.

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.

Referee Report

3 major / 2 minor

Summary. The paper proposes a hybrid quantum-classical corrective diffusion model for probabilistic statistical downscaling of 10m wind fields. Variational quantum circuit (VQC) layers are inserted into the compressed bottleneck of a diffusion UNet (regression branch remains classical) to test compact nonlinear feature mapping. Intra- and cross-channel ansatze are evaluated; on the 2020 validation set the hybrid variants are reported to improve MAE and CRPS over a classical baseline while preserving large-scale spatial structure, kinetic-energy spectra, and wind-speed distributions. Simulated device noise has negligible effect at the tested scale, but real-hardware deployment is limited by qubit count and fidelity. The 2021 out-of-distribution test reveals a generalization gap that does not transfer the in-distribution gains uniformly.

Significance. If the reported metric gains can be shown to arise specifically from the variational quantum circuits rather than from added bottleneck capacity, the work would provide a concrete empirical demonstration of hybrid quantum-classical models in a high-stakes scientific application. The explicit documentation of the 2021 generalization failure and the stability under simulated noise are positive features that increase credibility. At present the practical significance remains modest because of the missing classical control, limited circuit scale, and lack of transfer to later years.

major comments (3)
  1. [Evaluation on 2020 validation set (abstract and results description)] The central empirical claim (improved MAE/CRPS on the 2020 validation set) is presented as evidence that VQC layers in the UNet bottleneck produce a nontrivial contribution. However, the manuscript compares only against an unmodified classical corrective diffusion baseline and does not report a capacity-matched classical control (e.g., an MLP or small convolutional block with comparable parameter count and identical input/output dimensions placed in the same bottleneck location). Without this ablation it is impossible to attribute the observed gains to the quantum feature map rather than to the architectural modification itself. This issue is load-bearing for any claim that the hybridization, rather than extra classical capacity, is responsible for the improvement.
  2. [Results on 2020 validation set] No statistical significance tests (paired t-tests, bootstrap confidence intervals, or multiple-comparison corrections) are mentioned for the MAE and CRPS differences between hybrid and classical models. Given that the improvements are described as occurring “in several configurations,” the absence of uncertainty quantification leaves open the possibility that the reported gains are within the variability of the training or sampling procedure.
  3. [2021 out-of-distribution test (abstract)] The 2021 out-of-distribution test is reported to show that in-distribution gains “do not transfer uniformly.” Because the paper’s motivation is meteorological downscaling, where temporal shifts are routine, the lack of any analysis or mitigation strategy for this generalization gap weakens the practical claim that the hybrid approach constitutes a viable contribution to the field.
minor comments (2)
  1. [Abstract] The abstract contains a typographic artifact (“ans”atze”) and an inconsistent pronoun (“its classical counterpart” should be “their”).
  2. [Methods (bottleneck description)] Notation for the variational quantum circuit ansatze (intra-channel vs. cross-channel) is introduced without an accompanying diagram or explicit parameter-count table, making it difficult to verify that the different ansatze are being compared on an equal footing.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, indicating where revisions will be made.

read point-by-point responses

  1. Referee: [Evaluation on 2020 validation set (abstract and results description)] The central empirical claim (improved MAE/CRPS on the 2020 validation set) is presented as evidence that VQC layers in the UNet bottleneck produce a nontrivial contribution. However, the manuscript compares only against an unmodified classical corrective diffusion baseline and does not report a capacity-matched classical control (e.g., an MLP or small convolutional block with comparable parameter count and identical input/output dimensions placed in the same bottleneck location). Without this ablation it is impossible to attribute the observed gains to the quantum feature map rather than to the architectural modification itself. This issue is load-bearing for any claim that the hybridization, rather than extra classical capacity, is responsible for the improvement.

    Authors: We agree that a capacity-matched classical control is necessary to isolate the contribution of the VQC layers from the effect of added bottleneck capacity. In the revised manuscript we will include an ablation study with a classical MLP (or small convolutional block) of comparable parameter count inserted at the same bottleneck location, allowing direct comparison of performance metrics and attribution of any gains. revision: yes

  2. Referee: [Results on 2020 validation set] No statistical significance tests (paired t-tests, bootstrap confidence intervals, or multiple-comparison corrections) are mentioned for the MAE and CRPS differences between hybrid and classical models. Given that the improvements are described as occurring “in several configurations,” the absence of uncertainty quantification leaves open the possibility that the reported gains are within the variability of the training or sampling procedure.

    Authors: We acknowledge the lack of statistical significance testing. In the revision we will add bootstrap confidence intervals (or paired t-tests where appropriate) for the MAE and CRPS differences across configurations, together with a description of the resampling procedure, to quantify uncertainty and support the reliability of the reported improvements. revision: yes

  3. Referee: [2021 out-of-distribution test (abstract)] The 2021 out-of-distribution test is reported to show that in-distribution gains “do not transfer uniformly.” Because the paper’s motivation is meteorological downscaling, where temporal shifts are routine, the lack of any analysis or mitigation strategy for this generalization gap weakens the practical claim that the hybrid approach constitutes a viable contribution to the field.

    Authors: The manuscript already explicitly documents the 2021 generalization gap and states that it motivates future mitigation via stabilization and regularization. We agree, however, that additional analysis of the distribution shift (e.g., examining changes in input statistics or feature sensitivity) and an outline of initial mitigation directions would strengthen the practical discussion. We will expand the relevant section accordingly while preserving the honest reporting of the observed limitation. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on held-out empirical comparisons

full rationale

The paper's central claims consist of empirical performance metrics (MAE, CRPS) measured on a 2020 validation set for hybrid quantum-classical diffusion models versus a classical baseline, with additional diagnostics on spectra and distributions. These are direct evaluations on held-out data after training, not quantities obtained by fitting parameters to the reported target metrics or by self-referential definitions. No equations, ansatze, or uniqueness theorems are presented that reduce the reported improvements to the inputs by construction. No load-bearing self-citations or renamings of known results appear in the derivation of the main results. The architecture description (insertion of VQC layers into the UNet bottleneck) is a modeling choice whose effects are tested empirically rather than assumed tautologically.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Central claim rests on empirical performance of a neural architecture whose training involves standard ML optimization; the abstract supplies no explicit free-parameter count beyond the implicit variational circuit angles and UNet weights. No invented physical entities are introduced.

free parameters (1)
  • variational quantum circuit parameters
    Angles in the VQC layers are optimized during end-to-end training and directly affect the reported MAE/CRPS values.
axioms (1)
  • domain assumption Variational quantum circuits placed at the UNet bottleneck can function as compact nonlinear feature maps for latent-channel mixing
    This modeling choice is invoked to justify the hybrid insertion and is required for the claimed benefit over the classical baseline.

pith-pipeline@v0.9.0 · 5829 in / 1270 out tokens · 23443 ms · 2026-05-25T05:04:32.039887+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 2 internal anchors

  1. [1]

    Empirical-statistical downscaling,

    R. Benestad, I. Hanssen-Bauer, and D. Chen, “Empirical-statistical downscaling,”World Scientific Editions, 09 2008

    work page 2008

  2. [2]

    An asynchronous regional regression model for statistical downscaling of daily climate variables,

    A. M. Stoner, K. Hayhoe, X. Yang, and D. Wuebbles, “An asynchronous regional regression model for statistical downscaling of daily climate variables,”International Journal of Climatology, vol. 33, 09 2013

    work page 2013

  3. [3]

    Nasa global daily downscaled projections, cmip6,

    B. Thrasher, W. Wang, A. Michaelis, F. Melton, T. Lee, and R. Nemani, “Nasa global daily downscaled projections, cmip6,”Scientific Data, vol. 9, June 2022

    work page 2022

  4. [4]

    Sta- tistical downscaling of precipitation using machine learning techniques,

    D. Sachindra, K. Ahmed, M. M. Rashid, S. Shahid, and B. Perera, “Sta- tistical downscaling of precipitation using machine learning techniques,” Atmospheric Research, vol. 212, pp. 240–258, 2018

    work page 2018

  5. [5]

    Statistical down- scaling of general circulation model outputs to catchment streamflows,

    D. Sachindra, F. Huang, A. Barton, and B. Perera, “Statistical down- scaling of general circulation model outputs to catchment streamflows,” 01 2011

    work page 2011

  6. [6]

    Effect of model resolution and subgrif scale physics on the simulation of precipitation in the continental united states,

    I. J.P, P. Duffy, B. Govindasamy, S. Thompson, and D. Randall, “Effect of model resolution and subgrif scale physics on the simulation of precipitation in the continental united states,”Climate Dynamics, vol. 23, pp. 243–258, 01 2004

    work page 2004

  7. [7]

    Impact of climate change on urban heat island effect and extreme temperatures: a case-study,

    D. A. Sachindra, A. W. M. Ng, S. Muthukumaran, and B. J. C. Perera, “Impact of climate change on urban heat island effect and extreme temperatures: a case-study,”Quarterly Journal of the Royal Meteorological Society, vol. 142, no. 694, p. 172–186, 2015

    work page 2015

  8. [8]

    The ongoing need for high-resolution regional climate models: Process understanding and stakeholder information,

    W. J. Gutowski, P. A. Ullrich, A. Hall, L. R. Leung, T. A. O’Brien, C. M. Patricola-DiRosario, R. W. Arritt, M. S. Bukovsky, K. V . Calvin, Z. Feng, A. D. Jones, G. J. Kooperman, E. Monier, M. S. Pritchard, S. C. Pryor, Y . Qian, A. M. Rhoades, A. F. Roberts, K. Sakaguchi, N. Urban, and C. Zarzycki, “The ongoing need for high-resolution regional climate m...

    work page 2020

  9. [9]

    Statistical downscaling of rainfall: a non-stationary and multi-resolution approach,

    M. M. Rashid, S. Beecham, and R. Chowdhury, “Statistical downscaling of rainfall: a non-stationary and multi-resolution approach,”Theoretical and Applied Climatology, vol. 124, 04 2015

    work page 2015

  10. [10]

    Adversarial super- resolution of climatological wind and solar data,

    K. Stengel, A. Glaws, D. Hettinger, and R. N. King, “Adversarial super- resolution of climatological wind and solar data,”Proceedings of the National Academy of Sciences, vol. 117, no. 29, pp. 16805–16815, 2020

    work page 2020

  11. [11]

    Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network,

    J. Leinonen, D. Nerini, and A. Berne, “Stochastic super-resolution for downscaling time-evolving atmospheric fields with a generative adversarial network,”IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 9, pp. 7211–7223, 2021

    work page 2021

  12. [12]

    Deep learning for downscaling tropical cyclone rainfall to hazard-relevant spatial scales,

    E. V osper, P. Watson, L. Harris, A. McRae, R. Santos-Rodriguez, L. Aitchison, and D. Mitchell, “Deep learning for downscaling tropical cyclone rainfall to hazard-relevant spatial scales,”Journal of Geophysi- cal Research: Atmospheres, vol. 128, no. 10, p. e2022JD038163, 2023. e2022JD038163 2022JD038163

    work page 2023

  13. [13]

    Climax: A foundation model for weather and climate.arXiv preprint arXiv:2301.10343, 2023

    T. Nguyen, J. Brandstetter, A. Kapoor, J. K. Gupta, and A. Grover, “Climax: A foundation model for weather and climate,”arXiv preprint arXiv:2301.10343, 2023

    work page arXiv 2023

  14. [14]

    Universal Diffusion-Based Probabilistic Downscaling

    R. Molinaro, N. Siegenheim, H. Martin, M. Frey, N. Poulsen, P. Seitz, and M. V . Gabler, “Universal diffusion-based probabilistic downscaling,” arXiv preprint arXiv:2602.11893, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Diffusion models for high-resolution solar forecasts,

    Y . Hatanaka, Y . Glaser, G. Galgon, G. Torri, and P. Sadowski, “Diffusion models for high-resolution solar forecasts,”arXiv preprint arXiv:2302.00170, 2023

    work page arXiv 2023

  16. [16]

    Upscale error growth in a high-resolution simulation of a summertime weather event over europe,

    T. Selz and G. C. Craig, “Upscale error growth in a high-resolution simulation of a summertime weather event over europe,”Monthly Weather Review, vol. 143, no. 3, pp. 813 – 827, 2015

    work page 2015

  17. [17]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    work page 2020

  18. [18]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y . Song, S. Hong, R. Xu, Y . Zhao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,”ACM computing surveys, vol. 56, no. 4, pp. 1–39, 2023

    work page 2023

  19. [19]

    Residual corrective diffusion modeling for km-scale atmospheric downscaling,

    M. Mardani, N. Brenowitz, Y . Cohen, J. Pathak, C.-Y . Chen, C.-C. Liu, A. Vahdat, M. A. Nabian, T. Ge, A. Subramaniam,et al., “Residual corrective diffusion modeling for km-scale atmospheric downscaling,” Communications Earth & Environment, vol. 6, no. 1, p. 124, 2025

    work page 2025

  20. [20]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention, pp. 234– 241, Springer, 2015

    work page 2015

  21. [21]

    Towards efficient quantum hybrid diffusion models,

    F. De Falco, A. Ceschini, A. Sebastianelli, B. L. Saux, and M. Panella, “Towards efficient quantum hybrid diffusion models,”arXiv preprint arXiv:2402.16147, 2024

    work page arXiv 2024

  22. [22]

    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, B. AkashNarayanan, A. Asadi,et al., “Pennylane: Automatic differentiation of hybrid quantum-classical com- putations,”arXiv preprint arXiv:1811.04968, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  23. [23]

    Eval- uating analytic gradients on quantum hardware,

    M. Schuld, V . Bergholm, C. Gogolin, J. Izaac, and N. Killoran, “Eval- uating analytic gradients on quantum hardware,”Physical Review A, vol. 99, no. 3, p. 032331, 2019

    work page 2019

  24. [24]

    Strictly proper scoring rules, predic- tion, and estimation,

    T. Gneiting and A. E. Raftery, “Strictly proper scoring rules, predic- tion, and estimation,”Journal of the American Statistical Association, vol. 102, pp. 359–378, mar 2007

    work page 2007

  25. [25]

    Atmospheric kinetic energy spectra from global high-resolution nonhydrostatic simu- lations,

    W. C. Skamarock, S.-H. Park, J. B. Klemp, and C. Snyder, “Atmospheric kinetic energy spectra from global high-resolution nonhydrostatic simu- lations,”Journal of the Atmospheric Sciences, vol. 71, no. 11, pp. 4369 – 4381, 2014

    work page 2014

  26. [26]

    Scale-selective verification of rainfall accumulations from high-resolution forecasts of convective events,

    N. M. Roberts and H. W. Lean, “Scale-selective verification of rainfall accumulations from high-resolution forecasts of convective events,” Monthly Weather Review, vol. 136, no. 1, pp. 78 – 97, 2008

    work page 2008

  27. [27]

    Multivariate bias correction of climate model output: Matching marginal distributions and intervariable dependence structure,

    A. J. Cannon, “Multivariate bias correction of climate model output: Matching marginal distributions and intervariable dependence structure,” Journal of Climate, vol. 29, no. 19, pp. 7045 – 7064, 2016

    work page 2016

  28. [28]

    Multivariate bias correction and downscaling of climate models with trend-preserving deep learning,

    F. Wang and D. Tian, “Multivariate bias correction and downscaling of climate models with trend-preserving deep learning,”Climate Dynamics, vol. 62, pp. 9651–9672, Oct 2024

    work page 2024

  29. [29]

    Barren plateaus in quantum neural network training landscapes,

    J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, “Barren plateaus in quantum neural network training landscapes,”Nature Communications, vol. 9, p. 4812, Nov 2018

    work page 2018

  30. [30]

    An initial- ization strategy for addressing barren plateaus in parametrized quantum circuits,

    E. Grant, L. Wossnig, M. Ostaszewski, and M. Benedetti, “An initial- ization strategy for addressing barren plateaus in parametrized quantum circuits,”Quantum, vol. 3, p. 214, Dec. 2019

    work page 2019

  31. [31]

    Cost function dependent barren plateaus in shallow parametrized quantum circuits,

    M. Cerezo, A. Sone, T. V olkoff, L. Cincio, and P. J. Coles, “Cost function dependent barren plateaus in shallow parametrized quantum circuits,” Nature Communications, vol. 12, p. 1791, Mar 2021

    work page 2021