pith. sign in

arxiv: 2402.08475 · v5 · submitted 2024-02-13 · 🪐 quant-ph

HQNET: Harnessing Quantum Noise for Effective Training of Quantum Neural Networks in NISQ Era

Muhammad Kashif , Muhammad Shafique This is my paper

Pith reviewed 2026-05-24 03:35 UTC · model grok-4.3

classification 🪐 quant-ph
keywords quantum neural networksbarren plateausNISQ devicesmeasurement observablesquantum machine learningnoise mitigationPauli operatorsHermitian observable
0
0 comments X

The pith

Choosing the right measurement observable allows quantum neural networks to train on up to 10 noisy qubits without barren plateaus halting progress.

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

This paper examines how noise in current quantum devices creates barren plateaus that stop quantum neural networks from learning. It tests different ways to measure the qubits at the end of the circuit, including standard Pauli operators and one custom Hermitian matrix. The results indicate that a tailored observable keeps the training landscape usable even when measuring all qubits together, scaling to ten qubits where others fail earlier. For measuring one qubit at a time, the standard PauliZ works best up to ten qubits. The work shows that observable choice can be a practical way to push quantum machine learning further in noisy hardware.

Core claim

The paper demonstrates that among PauliX, PauliY, PauliZ, and a customized Hermitian observable, the customized one is most robust to noise-induced barren plateaus when using a global cost function, permitting effective training of quantum neural networks with up to 10 qubits. With a local cost function, PauliZ outperforms the others up to 10 qubits. Simulations show that PauliX and PauliY lead to flatter landscapes under noise with global costs.

What carries the argument

The selection of the Hermitian observable for measurement, combined with global or local cost functions, which determines the cost landscape's trainability under depolarizing noise.

Load-bearing premise

The noise model used in the simulations captures the main effects that produce barren plateaus on actual NISQ hardware, and performance differences come from the observable rather than other factors.

What would settle it

Running the QNN training on a real 10-qubit NISQ device with the customized observable under global cost and observing whether the loss decreases meaningfully or remains flat.

Figures

Figures reproduced from arXiv: 2402.08475 by Muhammad Kashif, Muhammad Shafique.

Figure 1
Figure 1. Figure 1: Cost Function Landscapes of 4-qubit QNN (a) Without Noise and (b) With Noise. (c) Comparison of Nosiy and Noise Free Training of 4-qubit QNN. The noise makes the cost function landscape flattened indicating the presence of BPs which eventually hinders the trainability of QNNs. of BPs leading to reduced trainability, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed methodology highlighting key steps for the analysis of noise impact on the trainability of QNNs with different qubit measurement strategies. The quantum circuits used in QNN design are constructed with 4 to 10 qubits. Each qubit has RX and RY gates applied on it, and neighboring qubits are entangled via the CZ gate. Two distinct quantum circuit configurations were explored, Global QNN (all qubits … view at source ↗
Figure 3
Figure 3. Figure 3: Optimization Landscape of 4-Qubit Global QNN under Noise-Free(upper panel) and Noisy (lower panel) Environments. For all the observables, the landscapes in case of no noise have multiple wider regions containing the solution whereas the landscapes with noise are mostly flat and truncated. For noisy environment, while using the customized Hermitian observable noise can be leveraged to advantage since the la… view at source ↗
Figure 4
Figure 4. Figure 4: Training Results of 4-Qubit Global QNN Under Noisy and Noise￾Free Environments. The noise-free training is better in all the cases, however, the customized Hermitian observable performs significantly better than other observables in noisy setting. 2) 6-Qubit Global QNN: a) Optimization Landscape in Noise-Free Setting: In the 6- qubit QNN, comprising 60 single-qubit and 25 two-qubit gates, we observe distin… view at source ↗
Figure 5
Figure 5. Figure 5: Optimization Landscapes of 6-Qubit Global QNN under Noise-Free (upper panel) Noise free) and Noisy (lower panel) Environments. In case of no noise, the landscapes with the customized Hermitian and PauliZ observable seems better as they have few regions containing the global miminum, whereas with PauliX and PauliY observables there are multiple local minima and almost no global minimum. In the presence of n… view at source ↗
Figure 6
Figure 6. Figure 6: Training of 6-Qubit Global QNN Under Noisy and Noise-Free Environments. The noise-free training is better with all the observables than noisy training. However, the customized Hermitian observable performs significantly better than other observables and almosy same as in case noise￾free training, in noisy setting. landscape for PauliZ observables, though truncated due to noise, presents some potential for … view at source ↗
Figure 7
Figure 7. Figure 7: Optimization Landscapes under Noisy Setting for 8-Qubit (upper panel) and 10 Qubit(lower panel) Global QNN. The landscapes for QNNs with more expressive circuits tends to become further flat indicating the occurence of BPs. The landscape for PauliX and PauliY and PauliZ observables are completely and majorly flat with almost no or very limited potential for optimization. The landscape with customized Hermi… view at source ↗
Figure 8
Figure 8. Figure 8: Training of 8 (left) and 10 (right) Qubit Global QNN with Different Observables Under Noise. For both 8 and 10 qubit QNNs, all the Pauli observables show limited or no training performance. Only the customized Hermitian observable performs effective training while benefitting from noise and escaping the BPs. Furthermore, our previous results indicated that in the absence of quantum noise, all observables s… view at source ↗
Figure 10
Figure 10. Figure 10: Training Results of 8 and 10 Qubit Local QNN with different observables in Noisy Environment. The landscape with PauliZ observable in local QNN even with 10 qubits contains multiple wider regions leading to the solution. PauliX and PauliY observables’ landscape are still completely flat. The 8-qubit QNN with local cost function definition with different measurement observables are then subjected to traini… view at source ↗
Figure 11
Figure 11. Figure 11: , do not differ much to that of 8-qubit QNN (with local cost function) design ( [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
read the original abstract

Effective training of Quantum Neural Networks (QNNs) is crucial in the Noisy Intermediate-Scale Quantum (NISQ) era, where noise accelerates the onset of barren plateaus (BPs) and limits scalability. This paper investigates how quantum noise impacts QNN trainability and demonstrates that careful selection of qubit measurement observables can mitigate these effects. We analyze PauliX, PauliY, PauliZ, and a customized Hermitian observable under both global (all-qubit measured) and local (single-qubit measured) cost functions. Our results show that with global cost function, PauliX and PauliY lead to flatter landscapes under noise, while PauliZ maintains training up to $8$ qubits before encountering BPs. The customized Hermitian observable proves most robust, enabling training up to $10$ qubits in noisy settings. For local cost function setting, PauliZ outperforms PauliX and PauliY, maintaining efficiency up to $10$ qubits. These findings highlight the importance of noise-aware observable selection, offering a practical strategy to improve QNN performance and advance quantum machine learning in noisy environments.

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

2 major / 2 minor

Summary. The manuscript claims that careful choice of measurement observables in QNNs can mitigate noise-induced barren plateaus. Simulations show PauliX/Y produce flatter landscapes under global costs, PauliZ sustains training to 8 qubits, and a customized Hermitian observable enables training to 10 qubits; under local costs PauliZ reaches 10 qubits while X/Y do not.

Significance. If the noise model and simulation protocol accurately capture dominant NISQ effects, the work supplies a concrete, observable-selection heuristic that could extend the practical reach of QNN training by 2–4 qubits without hardware changes. The empirical ranking of observables is falsifiable and directly actionable for circuit design.

major comments (2)
  1. [Abstract, §4] Abstract and §4 (Simulation Setup): the claim that the customized Hermitian observable is 'most robust' and enables training to 10 qubits rests on unspecified noise-channel parameters, circuit depth, shot counts, and error-bar reporting. Without these quantities it is impossible to judge whether the reported thresholds are statistically distinguishable from the Pauli cases or sensitive to unmodeled effects (readout, coherent, or correlated noise).
  2. [§3–4] §3–4 (Noise Model): the central attribution of performance differences to observable choice assumes the chosen depolarizing or amplitude-damping channels dominate barren-plateau formation identically to real hardware. No sensitivity analysis or hardware-validation experiment is described; if readout or crosstalk terms alter relative gradient variances, the ranking of observables would not hold.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'customized Hermitian observable' is used without an explicit operator definition or construction rule; a one-line mathematical expression should appear at first mention.
  2. [Figure captions] Figure captions (assumed §5): axis labels and legend entries for 'global' vs 'local' cost functions should explicitly state the measured observable and the precise cost-function definition to avoid ambiguity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to enhance clarity on simulation parameters and noise-model assumptions.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (Simulation Setup): the claim that the customized Hermitian observable is 'most robust' and enables training to 10 qubits rests on unspecified noise-channel parameters, circuit depth, shot counts, and error-bar reporting. Without these quantities it is impossible to judge whether the reported thresholds are statistically distinguishable from the Pauli cases or sensitive to unmodeled effects (readout, coherent, or correlated noise).

    Authors: We agree that explicit reporting is required for reproducibility. The revised manuscript will add a table in §4 listing all noise parameters (depolarizing probability, amplitude-damping rate), circuit depths, shot counts per expectation value, and error bars (standard deviation over repeated runs) together with a brief statistical comparison confirming that the 10-qubit threshold for the customized observable is distinguishable from the Pauli cases under the reported conditions. revision: yes

  2. Referee: [§3–4] §3–4 (Noise Model): the central attribution of performance differences to observable choice assumes the chosen depolarizing or amplitude-damping channels dominate barren-plateau formation identically to real hardware. No sensitivity analysis or hardware-validation experiment is described; if readout or crosstalk terms alter relative gradient variances, the ranking of observables would not hold.

    Authors: The depolarizing and amplitude-damping channels are the standard models employed in the barren-plateau literature to isolate observable effects. We will add a sensitivity analysis in the revision that varies noise strengths over a factor of two and shows the observable ranking remains stable. Hardware validation lies outside the scope of this simulation study; we will insert a limitations paragraph noting that unmodeled readout or crosstalk could affect the ranking and that experimental confirmation is needed. revision: partial

Circularity Check

0 steps flagged

No circularity; results are direct empirical simulation outputs with no self-referential reductions.

full rationale

The paper reports numerical simulation outcomes on how PauliX/Y/Z and a custom Hermitian observable affect QNN trainability thresholds (e.g., PauliZ up to 8 qubits globally, custom observable up to 10 qubits) under specified noise models and cost functions. These thresholds are presented as simulation results rather than quantities derived from equations, fitted parameters renamed as predictions, or self-citations. No load-bearing steps match the enumerated circularity patterns: there are no self-definitional observables, no fitted inputs called predictions, and no uniqueness theorems or ansatzes imported via citation. The derivation chain consists of independent circuit simulations, rendering the findings self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on empirical simulation comparisons; the custom Hermitian observable is introduced without independent evidence outside the reported runs. No explicit free parameters are described in the abstract.

axioms (1)
  • domain assumption Standard quantum mechanics and common NISQ noise models (depolarizing, amplitude damping, etc.) govern the simulated circuits.
    The analysis of noise impact on trainability presupposes these models produce representative barren-plateau behavior.
invented entities (1)
  • Customized Hermitian observable no independent evidence
    purpose: To serve as a more robust measurement operator that mitigates noise-induced barren plateaus better than Pauli operators.
    Introduced in the abstract as the observable that enables training to 10 qubits; no external falsifiable prediction is given.

pith-pipeline@v0.9.0 · 5720 in / 1425 out tokens · 33714 ms · 2026-05-24T03:35:18.513381+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Identifying Protein Co-regulatory Network Logic by Solving B-SAT Problems through Gate-based Quantum Computing

    quant-ph 2025-04 unverdicted novelty 7.0

    Grover's algorithm solves a B-SAT encoding of protein co-regulatory logic to recover high-likelihood Boolean models for a 5-protein neural development network from sparse data on quantum simulators and NISQ devices.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Quantum Computing in the NISQ era and beyond,

    J. Preskill, “Quantum Computing in the NISQ era and beyond,”Quan- tum, vol. 2, p. 79, Aug. 2018

    work page 2018

  2. [2]

    Towards advantages of parameterized quantum pulses,

    Z. Lianget al., “Towards advantages of parameterized quantum pulses,” arXiv, no. 2304.09253, 2023

    work page arXiv 2023

  3. [3]

    Quantum circuit matrix product state ansatz for large-scale simulations of molecules,

    Y . Fanet al., “Quantum circuit matrix product state ansatz for large-scale simulations of molecules,”arXiv, no. 2301.06376, 2023

    work page arXiv 2023

  4. [4]

    Computational advantage in hybrid quantum neural networks: Myth or reality?

    M. Kashifet al., “Computational advantage in hybrid quantum neural networks: Myth or reality?”arXiv:2412.04991, 2025

    work page arXiv 2025

  5. [5]

    Nisq computing: where are we and where do we go?

    J. W. Z. Lauet al., “Nisq computing: where are we and where do we go?”AAPPS Bulletin, vol. 32, no. 1, p. 27, 2022

    work page 2022

  6. [6]

    Investigating the effect of noise on the training performance of hybrid quantum neural networks,

    M. Kashifet al., “Investigating the effect of noise on the training performance of hybrid quantum neural networks,” in2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–10

    work page 2024

  7. [7]

    Quantum advantage in cryptography,

    R. Renner and R. Wolf, “Quantum advantage in cryptography,”AIAA Journal, vol. 61, no. 5, pp. 1895–1910, 2023

    work page 1910

  8. [8]

    Quantum computing for near-term applications in generative chemistry and drug discovery,

    A. Pyrkovet al., “Quantum computing for near-term applications in generative chemistry and drug discovery,”Drug Discovery Today, p. 103675, 2023

    work page 2023

  9. [9]

    Quantum machine learning,

    J. Biamonteet al., “Quantum machine learning,”Nature, vol. 549, no. 7671, pp. 195–202, sep 2017

    work page 2017

  10. [10]

    Parameterized quantum circuits as machine learning models,

    M. Benedettiet al., “Parameterized quantum circuits as machine learning models,”Quantum Science and Technology, vol. 4, no. 4, p. 043001, nov 2019

    work page 2019

  11. [11]

    Next- generation quantum neural networks: Enhancing efficiency, security, and privacy,

    N. Innan, M. Kashif, A. Marchisio, M. Bennai, and M. Shafique, “Next- generation quantum neural networks: Enhancing efficiency, security, and privacy,” in2025 IEEE 31st International Symposium on On-Line Testing and Robust System Design (IOLTS), 2025, pp. 1–4

    work page 2025

  12. [12]

    Variational quantum algorithms,

    M. Cerezoet al., “Variational quantum algorithms,”Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

    work page 2021

  13. [13]

    A survey on quantum machine learning: Current trends, challenges, opportunities, and the road ahead,

    K. Zamanet al., “A survey on quantum machine learning: Current trends, challenges, opportunities, and the road ahead,” 2023

    work page 2023

  14. [14]

    Classification with Quantum Neural Networks on Near Term Processors

    E. Farhi and H. Neven, “Classification with quantum neural networks on near term processors,”arXiv, no. 1802.06002, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Design space exploration of hybrid quantum–classical neural networks,

    M. Kashif and S. Al-Kuwari, “Design space exploration of hybrid quantum–classical neural networks,”Electronics, vol. 10, no. 23, p. 2980, 2021

    work page 2021

  16. [16]

    Barren plateaus in quantum neural network training landscapes,

    J. R. McCleanet al., “Barren plateaus in quantum neural network training landscapes,”Nature Communications, vol. 9, no. 1, nov 2018

    work page 2018

  17. [17]

    Mitigating barren plateaus with transfer-learning-inspired parameter initializations,

    H. Liuet al., “Mitigating barren plateaus with transfer-learning-inspired parameter initializations,”New Journal of Phys, vol. 25, no. 1, p. 013039, 2023

    work page 2023

  18. [18]

    Alleviating barren plateaus in parameterized quantum machine learning circuits: Investigating advanced parameter initialization strategies,

    M. Kashifet al., “Alleviating barren plateaus in parameterized quantum machine learning circuits: Investigating advanced parameter initialization strategies,”arXiv, no. 2311.13218, 2023

    work page arXiv 2023

  19. [19]

    Deep quanvolutional neural networks with enhanced trainability and gradient propagation,

    M. Kashif and M. Shafique, “Deep quanvolutional neural networks with enhanced trainability and gradient propagation,”Scientific Reports, vol. 15, no. 1, p. 21764, 2025

    work page 2025

  20. [20]

    Resqnets: a residual approach for mit- igating barren plateaus in quantum neural networks,

    M. Kashif and S. Al-Kuwari, “Resqnets: a residual approach for mit- igating barren plateaus in quantum neural networks,”EPJ Quantum Technology, vol. 11, no. 1, p. 4, 2024

    work page 2024

  21. [21]

    The dilemma of random parameter initialization and barren plateaus in variational quantum algorithms,

    M. Kashif and M. Shafique, “The dilemma of random parameter initialization and barren plateaus in variational quantum algorithms,” in 2024 IEEE International Conference on Rebooting Computing (ICRC). IEEE, 2024, pp. 1–8

    work page 2024

  22. [22]

    Entanglement-induced barren plateaus,

    O. Marreroet al., “Entanglement-induced barren plateaus,”PRX Quan- tum, vol. 2, p. 040316, Oct 2021

    work page 2021

  23. [23]

    The unified effect of data encoding, ansatz expressibility and entanglement on the trainability of hqnns,

    M. Kashif and S. Al-Kuwari, “The unified effect of data encoding, ansatz expressibility and entanglement on the trainability of hqnns,” International Journal of Parallel, Emergent and Distributed Systems, vol. 38, no. 5, pp. 362–400, 2023

    work page 2023

  24. [24]

    Cost function dependent barren plateaus in shallow parametrized quantum circuits,

    M. Cerezoet al., “Cost function dependent barren plateaus in shallow parametrized quantum circuits,”Nat. Comms, vol. 12, no. 1, 2021

    work page 2021

  25. [25]

    The impact of cost function globality and locality in hybrid quantum neural networks on nisq devices,

    M. Kashif and S. Al-Kuwari, “The impact of cost function globality and locality in hybrid quantum neural networks on nisq devices,”Machine Learning: Science and Technology, vol. 4, no. 1, p. 015004, jan 2023

    work page 2023

  26. [26]

    Quantum neural networks: A comparative analysis and noise robustness evaluation,

    T. Ahmedet al., “Quantum neural networks: A comparative analysis and noise robustness evaluation,”arXiv preprint arXiv:2501.14412, 2025

    work page arXiv 2025

  27. [27]

    Noise-induced barren plateaus in variational quantum algorithms,

    S. Wanget al., “Noise-induced barren plateaus in variational quantum algorithms,”Nature Communications, vol. 12, no. 1, nov 2021

    work page 2021

  28. [28]

    Noisy hqnns: A comprehensive analysis of noise ro- bustness in hybrid quantum neural networks,

    T. Ahmedet al., “Noisy hqnns: A comprehensive analysis of noise ro- bustness in hybrid quantum neural networks,”arXiv:2505.03378, 2025

    work page arXiv 2025

  29. [29]

    Nrqnn: The role of observable selection in noise-resilient quantum neural networks,

    M. Kashif and M. Shafique, “Nrqnn: The role of observable selection in noise-resilient quantum neural networks,”arXiv:2502.12637, 2025

    work page arXiv 2025

  30. [30]

    Pennylane: Automatic differentiation of hybrid quantum-classical computations,

    V . Bergholmet al., “Pennylane: Automatic differentiation of hybrid quantum-classical computations,”arXiv, 2018

    work page 2018