---
title: "CohAlign: Predicting first-order noise susceptibility of gradients in equivariant quantum neural n"
id: "plos-one-3-cohalign-predicting-the-first-order-noise-susceptibility-of-gradients-in"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-3-cohalign-predicting-the-first-order-noise-susceptibility-of-gradients-in"
content_type: "clinical_feed_article"
specialty: "General"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139"
published_at: "2026-10-22T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# CohAlign: Predicting first-order noise susceptibility of gradients in equivariant quantum neural n
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-3-cohalign-predicting-the-first-order-noise-susceptibility-of-gradients-in
- **Specialty:** [General](https://medichelpline.com/clinical-feed/general.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139)
- **Published At:** 2026-10-22T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The authors introduce **CohAlign**, an open-source toolkit that computes the first-order noise susceptibility of the ensemble second moment of a specified parameter **gradient** in charge-conserving brickwork quantum neural networks. - The calculation requires only a classical description: circuit, noise model, prepared state, readout and initialization prior; no finite-noise circuit sweep is needed, though noiseless classical simulation at design time is required. - The paper distinguishes three sector-coherence quantities: the worst-case sector coherence rate, the operator mode contraction rate for the gradient-carrying operator, and the response susceptibility that predicts measured gradient decay. - Investigators reported that a channel’s overall contraction rate can differ dramatically from the measured-gradient decay because only coherences visible to the readout matter; a constructed control demonstrates near-maximal operator contraction with no measured gradient decay. - Validation was performed with exact simulations at eight and ten qubits, across two circuit topologies, through a ladder of tests including analytic recovery, finite-difference slope comparisons, nine noise models, and a two-excitation sector study. - The toolkit uses a finite probe approximation and reports sensitivity of rates to probe strength; isotropic channels converge linearly in the probe and a working probe of 10^-3 is shipped. - The response susceptibility is computed from cross-moments of the noiseless derivative and its first-order response; estimator and measurement ensembles with fixed draw counts are used to form stable predictions. - Limitations include the need for classical noiseless simulation at studied system sizes and finite-probe biases; the paper documents these and supplies full reproducible code, data, and seeds in the public repository.
## Clinical Analysis & Structured Key Points
[ Skip to main content ](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#main-content) Advertisement * [plos.org](https://plos.org/) * [Create account](https://community.plos.org/registration/new) * [Sign in](https://journals.plos.org/user/secure/login?page=%2Fplosone%2Farticle%3Fid%3D10.1371%2Fjournal.pone.0361139) * * About * Browse * Publish * [](https://journals.plos.org/plosone/ "PLOS One") * Search [advanced search](https://journals.plos.org/plosone/search) * [Browse Topics](https://journals.plos.org/plosone/subjectAreaBrowse) Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click [here](https://github.com/PLOS/plos-thesaurus/blob/master/README.md "Link opens in new window"). [](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139) [](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139) * 0 [Save](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#savedHeader) [Total Mendeley and Citeulike bookmarks.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#savedHeader) * 0 [Citation](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#citedHeader) [Paper's citation count computed by Dimensions.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#citedHeader) * 0 [View](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#viewedHeader) [PLOS views and downloads.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#viewedHeader) * 0 [Share](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#discussedHeader) [Sum of Facebook, Twitter, Reddit and Wikipedia activity.](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139#discussedHeader) Open Access Peer-reviewed Research Article # CohAlign: Predicting the first-order noise susceptibility of gradients in equivariant quantum neural networks * Hassan Ugail , Roles Conceptualization, Formal analysis, Investigation, Methodology, Software, Validation, Writing – original draft, Writing – review & editing * E-mail: h.ugail@bradford.ac.uk Affiliation Centre for Visual Computing and Intelligent Systems, University of Bradford, Bradford, United Kingdom [ ![ORCID logo](https://journals.plos.org/resource/img/orcid_16x16.png) https://orcid.org/0000-0002-3084-1797 ](https://orcid.org/0000-0002-3084-1797 "ORCID Registry") ⨯ * Newton Howard Roles Conceptualization, Formal analysis, Investigation, Methodology, Software, Validation, Writing – original draft, Writing – review & editing Affiliation School of Individualized Study, Rochester Institute of Technology, Rochester, New York, United States of America ⨯ # CohAlign: Predicting the first-order noise susceptibility of gradients in equivariant quantum neural networks * Hassan Ugail, * Newton Howard ![PLOS](https://journals.plos.org/resource/img/logo-plos-full-color.svg) x * Published: October 22, 2026 * * [Article](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139) * [Authors](https://journals.plos.org/plosone/article/authors?id=10.1371/journal.pone.0361139) * [Metrics](https://journals.plos.org/plosone/article/metrics?id=10.1371/journal.pone.0361139) * [Comments](https://journals.plos.org/plosone/article/comments?id=10.1371/journal.pone.0361139) * [Media Coverage](http://plos.altmetric.com/details/doi/10.1371/journal.pone.0361139) * [Abstract](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#abstract0) * [1. Introduction](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec001) * [2. Materials and methods](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec002) * [3. Results](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec009) * [4. Discussion](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec016) * [5. Conclusions](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec017) * [Acknowledgments](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#ack) * [References](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#references) * [Reader Comments](https://journals.plos.org/plosone/article/comments?id=10.1371/journal.pone.0361139) * [Figures](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139) ![Accessible Data Icon](https://journals.plos.org/resource/img/accessible_data.svg)Accessible Data [ See the data ![Link Icon](https://journals.plos.org/resource/img/data_link_icon.svg) ](https://github.com/ugail/CohAlign) This article includes the Accessible Data icon, an experimental feature to encourage data sharing and reuse. [Find out how research articles qualify for this feature.](https://theplosblog.plos.org/2023/07/accessible-data/) ## Abstract Quantum neural networks learn by following gradients, and on real hardware those gradients are washed out by noise. Predicting how quickly this happens, before any finite noise experiment or circuit sweep is run, remains an open problem. We present CohAlign, an open source toolkit that computes the first order noise susceptibility of the ensemble second moment of a specified parameter gradient in a charge conserving brickwork quantum neural network. The computation needs only a classical description of the circuit, the noise model, the prepared state, the readout and the initialisation prior. The predicted quantity is distinct from training convergence or final accuracy. The key observation is that only the part of the noise that a measurement can actually see matters. A purposefully constructed control channel makes this clear, damaging the operator carrying a gradient at nearly the maximal rate while the measured gradient remains exactly as predicted and untouched. We validate the predictions against separately evaluated finite difference slopes across the benchmark suite, including a signed coherent case, in exact simulations at eight and ten qubits and on two circuit topologies. ## Figures ![Table 5](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t005) ![Table 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t001) ![Table 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t002) ![Fig 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g001) ![Fig 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g002) ![Fig 3](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g003) ![Fig 4](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g004) ![Table 3](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t003) ![Fig 5](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g005) ![Fig 6](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g006) ![Fig 7](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g007) ![Fig 8](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g008) ![Table 4](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t004) ![Table 5](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t005) ![Table 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t001) ![Table 2](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.t002) ![Fig 1](https://journals.plos.org/plosone/article/figure/image?size=inline&id=10.1371/journal.pone.0361139.g001) **Citation:** Ugail H, Howard N (2026) CohAlign: Predicting the first-order noise susceptibility of gradients in equivariant quantum neural networks. PLoS One 21(10): e0361139. https://doi.org/10.1371/journal.pone.0361139 **Editor:** Zeheng Wang, Commonwealth Scientific and Industrial Research Organisation, AUSTRALIA **Received:** August 20, 2026; **Accepted:** October 6, 2026; **Published:** October 22, 2026 **Copyright:** © 2026 Ugail, Howard. This is an open access article distributed under the terms of the [Creative Commons Attribution License](http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. **Data Availability:** All data underlying this study are generated deterministically by the released pipeline notebook. The notebook, the three module sources, all per phase outputs, worked applications, figures and the run metadata, including every seed and package version, are available in the project repository at under the MIT licence. **Funding:** The author(s) received no specific funding for this work. **Competing interests:** The authors have declared that no competing interests exist. ## 1. Introduction Variational quantum algorithms are among the leading candidates for useful computation on near term quantum hardware. The idea is simple. A circuit carries trainable parameters, a cost function is measured at its output, and the parameters are adjusted until the measured cost falls [[1](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref001),[2](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref002)], with gradients obtained through parameter-shift rules [[3](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref003)–[5](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref005)]. The central obstacle to scaling this approach is trainability. Deep or highly expressive circuits suffer from barren plateaus, regions of the landscape where gradients vanish exponentially with system size [[6](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref006)–[8](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref008)], and a growing body of theory identifies architectures and initialisations that avoid them [[9](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref009)–[11](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref011)]. Hardware noise, however, introduces a second and quite distinct mechanism. Even when the noiseless landscape is benign, noise accumulating through the layers suppresses the measured gradients, in the worst case producing noise induced plateaus that no initialisation can repair [[12](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref012)–[14](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref014)]. This second mechanism, and the problem of predicting it before any experiment is run, is what the present paper addresses. Symmetry offers a promising route through both difficulties. Equivariant architectures restrict every gate to commute with a conserved charge and are among the most attractive routes to trainable models [[15](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref015)–[20](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref020)], and for our purposes the symmetry brings a further and less widely appreciated benefit. Because the dynamics is confined to a fixed charge sector, the cost function and its gradients live in a known part of the Hilbert space, and only a specific portion of the coherence within that sector is visible to a local readout. A single scalar summary of the noise therefore paints too pessimistic a picture, because it says nothing about where the damage falls. The central contrast of this paper is that channel contraction is not the measured gradient degradation. Two channels with identical worst case rates can degrade gradients at very different speeds, and a channel that rapidly contracts almost every coherence direction can leave the gradients entirely untouched, provided the directions it contracts are the ones the readout cannot see. This paper makes that observation quantitative and turns it into a working software instrument. The setting is the family of brickwork quantum neural networks conserving a global charge under phase rotations. The central mathematical object is the in-sector off-diagonal block of the noise generator, the occupation basis coherences within a degenerate charge eigenspace, which constitute the sector relative form of quantum coherence [[21](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref021)–[23](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref023)]. In the mode decomposition of Marvian and Spekkens [[24](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref024),[25](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref025)] these coherences belong to the zero frequency mode of the global phase action, generated by the degeneracy of the sector. They are therefore invariant under the symmetry rather than resources of asymmetry, and this is precisely why the covariance of a channel with the symmetry offers them no protection. On this block it is helpful to distinguish three quantities, each answering a progressively sharper question. The worst case sector coherence rate ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e001) measures the fastest speed at which the channel contracts any coherence direction, and so tells us how bad the noise could possibly be. The operator mode contraction rate ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e002) measures how fast the channel contracts the particular operator mode carrying the derivative of the readout with respect to a chosen parameter, and so tells us how badly the noise damages the object that carries the gradient. The response susceptibility ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e003) measures how fast the initialisation ensemble second moment of the measured gradient decays, and depends jointly on the channel, the architecture, the parameter, the readout, the prepared state and the initialisation ensemble. The progression from ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e004) through ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e005) to ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e006) is the organising spine of the paper, each step conditioning the question on more of the problem, and a central lesson is that the last two are genuinely different objects. We exhibit a case in which the channel contracts the gradient mode at almost the full worst case rate while the measured gradient does not decay at all, simply because the contraction is invisible to the readout. This is the direct answer to the natural question of why an operator contraction rate is not already sufficient. The practical contribution is a toolkit, called CohAlign, that computes all three quantities ahead of time from a classical description of the problem, with no finite noise circuit sweep and no measured degradation data. It does require a classical simulation of the noiseless circuit at the system sizes we study, so we describe the method throughout as a classical design time calculation. The toolkit is validated on a ladder of tests of increasing difficulty. The ladder begins with an exact analytic recovery and an engineered control whose gradient is provably immune to its noise, continues through a direct finite difference validation of the susceptibility itself, extended across parameter locations, a second circuit topology, a larger system and varied settings. The ladder ends with a family wide comparison across nine noise models and a study in the two excitation sector, where scalar descriptions are least likely to suffice. This paper is standalone in the sense that every definition, derivation, dataset and line of code needed to reproduce its results is contained in the paper and its released pipeline. No result here depends on our earlier noiseless symmetry audits [[26](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref026),[27](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref027)] or channel level diagnostics [[28](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#pone.0361139.ref028)], which do not address the parameter resolved, readout aware prediction problem treated here. The remainder of the paper is organised as follows. [Section 2](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec002) defines the model, the three rates and the validation protocol, [Section 3](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec009) presents the validation ladder, [Section 4](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec016) interprets the findings and states the limitations, and [Section 5](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0361139#sec017) concludes, with two appendices collecting the ensemble derivation, the test battery and a worked example. ## 2. Materials and methods ### 2.1. Ethics statement This study did not involve human participants, animal experimentation or personal data. All experiments are numerical simulations implemented in open source scientific Python libraries, and no institutional ethics approval was required. ### 2.2. The model, the readout and the noise channels The architecture we study is a brickwork quantum neural network on a ring of _n_ qubits, and the implementation also provides an open chain variant, which drops the cycle closing edge, together with a selectable readout site, both used in the validation reported below. Each layer applies a trainable phase rotation ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e007) on every site, followed by a fixed layer of hopping gates ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e008) on alternating edge sets, even numbered edges in even layers and odd numbered edges in odd layers, with the ring closing edge included. All indices are zero based. Every gate commutes with the total charge ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e009), so the dynamics never mixes charge sectors. The readout is the sector projected observable ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e010), where ![](https://journals.plos.org/plosone/article/file?type=thumbnail&id=10.1371/journal.pone.0361139.e011) projects onto the sector with _r_ excitations, and the
## Related Clinical Research

- [Tobacco use among Lebanese university students: prevalence, drivers, and campus policy implications](https://medichelpline.com/clinical-feed/plos-one-4-prevalence-and-determinants-of-tobacco-use-among-university-students-in-lebanon.md)
- [Immersive virtual reality produces comparable willingness-to-pay estimates for ready-to-eat cereals](https://medichelpline.com/clinical-feed/plos-one-7-can-immersive-virtual-reality-be-an-alternative-world-for-eliciting-consumer.md)
- [How to define outcomes completely in clinical trials: six-element guidance](https://medichelpline.com/clinical-feed/plos-medicine-1-guidance-for-defining-outcomes-in-clinical-trials.md)
- [Molina Q3 2026: Mixed Quarter as Medicaid Cost Pressures Meet Rising State Rates](https://medichelpline.com/clinical-feed/healthcare-dive-1-molina-joins-medicaid-recovery-bandwagon-despite-mixed-q3.md)
- [Healthcare Providers Advocate for Developer Accountability in AI Medical Devices](https://medichelpline.com/clinical-feed/healthcare-dive-2-developers-not-doctors-should-be-held-accountable-for-generative-ai-medical.md)

## Navigation
- [← Back to General Feed](https://medichelpline.com/clinical-feed/general.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.