pith. machine review for the scientific record. sign in

arxiv: 2604.14229 · v2 · submitted 2026-04-14 · 🪐 quant-ph · cs.AI· cs.LG· eess.IV

Recognition: unknown

Magnitude Is All You Need? Rethinking Phase in Quantum Encoding of Complex SAR Data

Sakthi Prabhu Gunasekar , Prasanna Kumar Rangarajan
Authors on Pith no claims yet

Pith reviewed 2026-05-10 14:55 UTC · model grok-4.3

classification 🪐 quant-ph cs.AIcs.LGeess.IV
keywords SARquantum machine learningphase encodinghybrid modelsMSTARcomplex datatarget recognitionquantum encoding
0
0 comments X

The pith

Hybrid quantum-classical models for SAR classification perform better with magnitude-only encoding than with phase-inclusive encodings.

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

The paper questions the expectation that quantum models for SAR automatic target recognition should incorporate phase information from complex-valued data. It evaluates five encoding methods in a consistent experimental framework on the MSTAR dataset. Magnitude-only encoding emerges as superior in hybrid quantum-classical setups, delivering 99.57 percent accuracy on three-class classification and 71.19 percent on eight-class tasks. Purely quantum models, however, show marked gains from phase data, with improvements reaching 21.65 percentage points. Overall, the work reveals that phase information's value hinges on the model architecture, with hybrids able to bypass it through classical processing.

Core claim

Magnitude-only encoding performs better than phase-inclusive encodings in hybrid quantum-classical models. It achieves 99.57 percent accuracy on the 3-class task and 71.19 percent accuracy on the 8-class task, outperforming complex-valued alternatives under the same framework. In purely quantum models with only 184 to 224 trainable parameters and no classical neural-network layers, phase information becomes much more important, improving accuracy by up to 21.65 percentage points. These results demonstrate that the usefulness of phase depends on the architecture used to process the SAR data.

What carries the argument

Comparison of five encoding strategies for complex SAR data (magnitude-only, joint magnitude-phase, in-phase and quadrature, preprocessed phase, and purely quantum) evaluated in both hybrid and pure quantum models on the MSTAR dataset.

If this is right

  • Hybrid models compensate for absent phase data via classical components.
  • Pure quantum models depend more on phase for distinguishing classes.
  • Encoding strategies must be jointly designed with model architecture.
  • Magnitude-only encoding offers practical advantages for hybrid quantum-classical SAR processing.

Where Pith is reading between the lines

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

  • Researchers should test magnitude-only baselines first when working with complex-valued data in hybrid quantum setups.
  • The architecture dependence may generalize to other complex inputs such as communications signals or medical imaging.
  • As quantum resources increase, the relative value of phase information could shift even within hybrid models.

Load-bearing premise

The unified experimental setup implements and optimizes all five encodings without systematic bias from hyperparameter selection, circuit depth, or classical-layer capacity that could favor magnitude-only in the hybrid case.

What would settle it

Reproducing the hybrid-model experiments with exhaustive hyperparameter sweeps or alternative circuit designs that allow any phase-inclusive encoding to match or exceed the reported magnitude-only accuracies would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.14229 by Prasanna Kumar Rangarajan, Sakthi Prabhu Gunasekar.

Figure 1
Figure 1. Figure 1: Quantum circuit diagrams for key encoding strategies. (a) S1: [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Unified architecture overview. (a) Hybrid quantum-classical architecture (MagQT) supporting S1–S3 via swappable encoding. (b) Dual-path architecture [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MSTAR SAR samples for all 8 classes (ROI 64 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Phase contribution across architectures and encoding strategies. In [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

Synthetic Aperture Radar (SAR) data is inherently complex-valued, while quantum machine learning (QML) models operate in complex Hilbert spaces. This similarity suggests that using both the magnitude and phase of SAR data in quantum encoding should help automatic target recognition in SAR images. In this study, we test this assumption by comparing five encoding strategies for quantum models: magnitude-only encoding, joint magnitude-phase encoding, in-phase and quadrature encoding, preprocessed phase encoding, and a purely quantum architecture. All approaches are evaluated under a unified experimental setup on the MSTAR benchmark dataset. Surprisingly, we find that magnitude-only encoding performs better than phase-inclusive encodings in hybrid quantum-classical models. It achieves 99.57 percent accuracy on the 3-class task and 71.19 percent accuracy on the 8-class task, outperforming complex-valued alternatives under the same framework. Adding phase information provides little or no improvement and can sometimes degrade performance. However, in purely quantum models with only 184 to 224 trainable parameters and no classical neural-network layers, phase information becomes much more important, improving accuracy by up to 21.65 percentage points. These findings show that the usefulness of phase information depends not only on the data, but also on the architecture used to process it. Hybrid models can compensate for missing phase information through their classical components, while pure quantum models rely more strongly on phase information for class discrimination. The results provide practical guidance for encoding complex-valued data in quantum machine learning and highlight the importance of jointly designing encoding strategies and model architectures for current quantum systems.

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 compares five quantum encoding strategies for complex-valued SAR data on the MSTAR dataset: magnitude-only, joint magnitude-phase, in-phase/quadrature, preprocessed phase, and a purely quantum architecture. It reports that magnitude-only encoding outperforms phase-inclusive variants in hybrid quantum-classical models (99.57% accuracy on 3-class, 71.19% on 8-class tasks), while phase information improves accuracy by up to 21.65 percentage points in purely quantum models with 184–224 parameters and no classical layers. The central conclusion is that phase utility is architecture-dependent, with hybrid models compensating for absent phase via classical components.

Significance. If the empirical comparisons hold under fair conditions, the work supplies concrete guidance for encoding complex data in near-term QML and illustrates that hybrid vs. pure-quantum architectures can reverse the value of phase information. The unified-setup benchmark on a standard SAR dataset is a useful contribution, though its strength depends on verification that no encoding was inadvertently favored by fixed classical capacity or hyperparameter choices.

major comments (3)
  1. [§3] §3 (Experimental Setup): The manuscript states that all five encodings were evaluated 'under a unified experimental setup' but does not specify whether circuit depth, classical post-processing layers, optimizer schedule, or hyperparameter search were performed independently for each encoding. If a single hyperparameter configuration was applied uniformly, the reported superiority of magnitude-only encoding in hybrid models (Table 2 / §4.1) cannot be isolated from possible architectural favoritism toward simpler inputs.
  2. [§4.2] §4.2 (Purely Quantum Results): The purely quantum models are reported to contain only 184–224 trainable parameters, yet the text provides no explicit mapping of how each encoding populates the limited qubit register or variational circuit; without this, the claimed 21.65 pp gain from phase inclusion cannot be assessed for consistency with the hybrid-model findings.
  3. [§4.1] §4.1 and abstract: Point estimates (99.57 %, 71.19 %) are given without error bars, standard deviations across random seeds, or details of train/validation/test splits and hyperparameter-search protocol. This omission makes it impossible to judge whether the observed gaps exceed statistical variability.
minor comments (2)
  1. [§2] Notation for the five encodings is introduced in §2 but the precise mathematical definitions (e.g., how phase is normalized or preprocessed) are not restated when results are discussed, forcing the reader to cross-reference.
  2. [§4] Figure captions and axis labels in §4 would benefit from explicit mention of the number of shots, qubit count, and classical-layer widths used in each experiment.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful and constructive review. The comments highlight important aspects of experimental rigor and clarity that we address point by point below. We will incorporate clarifications and additional details in the revised manuscript.

read point-by-point responses

  1. Referee: [§3] §3 (Experimental Setup): The manuscript states that all five encodings were evaluated 'under a unified experimental setup' but does not specify whether circuit depth, classical post-processing layers, optimizer schedule, or hyperparameter search were performed independently for each encoding. If a single hyperparameter configuration was applied uniformly, the reported superiority of magnitude-only encoding in hybrid models (Table 2 / §4.1) cannot be isolated from possible architectural favoritism toward simpler inputs.

    Authors: We deliberately employed a single, fixed hyperparameter configuration (including circuit depth, classical post-processing, optimizer schedule, and learning rate) across all five encodings. This unified setup was chosen to ensure that observed performance differences arise from the encoding strategy rather than from encoding-specific architectural tuning, which would confound the central comparison. The configuration was selected after limited preliminary runs to achieve reasonable performance for the more complex encodings while remaining practical for the simpler ones. We acknowledge that this approach does not constitute an exhaustive per-encoding hyperparameter search. In the revision we will expand §3 to explicitly list the shared hyperparameter values, the selection rationale, and a brief discussion of why independent optimization was not performed. revision: yes

  2. Referee: [§4.2] §4.2 (Purely Quantum Results): The purely quantum models are reported to contain only 184–224 trainable parameters, yet the text provides no explicit mapping of how each encoding populates the limited qubit register or variational circuit; without this, the claimed 21.65 pp gain from phase inclusion cannot be assessed for consistency with the hybrid-model findings.

    Authors: We agree that an explicit parameter mapping is required for reproducibility and for evaluating the reported accuracy gains. The 184–224 parameter range corresponds to a fixed 4-qubit variational circuit whose rotation and entanglement gates are populated differently depending on the encoding (e.g., magnitude-only uses two real-valued features per pixel while phase-inclusive encodings supply additional rotation angles). In the revised manuscript we will add a dedicated paragraph or table in §4.2 that details, for each encoding, the number of qubits, the feature-to-rotation mapping, the ansatz structure, and the exact count of trainable parameters. This will allow direct comparison with the hybrid-model results. revision: yes

  3. Referee: [§4.1] §4.1 and abstract: Point estimates (99.57 %, 71.19 %) are given without error bars, standard deviations across random seeds, or details of train/validation/test splits and hyperparameter-search protocol. This omission makes it impossible to judge whether the observed gaps exceed statistical variability.

    Authors: We recognize that reporting only point estimates limits statistical interpretation. In the revised version we will augment §4.1 and the abstract with standard deviations obtained from five independent runs using different random seeds for weight initialization and data shuffling. We will also specify the exact train/validation/test split ratios used on the MSTAR dataset and describe the hyperparameter search protocol (grid search over a modest range of learning rates and circuit depths, with the final values fixed for the unified comparison). These additions will be presented alongside the existing accuracy figures. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark comparison with no derivation chain

full rationale

The paper presents an experimental comparison of five quantum encoding strategies for complex SAR data on the MSTAR dataset, reporting accuracies for hybrid and pure quantum models. No first-principles derivation, prediction, or uniqueness theorem is claimed; results are obtained by direct training and evaluation under a unified setup. No equations reduce a claimed output to a fitted input by construction, and no self-citation chain bears the central claim. The work is self-contained as an empirical study.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on empirical performance comparisons on the MSTAR benchmark. No new mathematical axioms, physical assumptions, or postulated entities are introduced; standard quantum-circuit simulation and classical optimization are presupposed.

pith-pipeline@v0.9.0 · 5595 in / 1146 out tokens · 63618 ms · 2026-05-10T14:55:04.440818+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

23 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Automatic target recognition on synthetic aperture radar imagery: A survey,

    O. Kechagias-Stamatis and N. Aouf, “Automatic target recognition on synthetic aperture radar imagery: A survey,”IEEE Access, vol. 9, pp. 40562–40596, 2021

    2021

  2. [2]

    The auto- matic target-recognition system in SAIP,

    L. M. Novak, G. J. Owirka, W. S. Brower, and A. L. Weaver, “The auto- matic target-recognition system in SAIP,”Lincoln Laboratory Journal, vol. 10, no. 2, pp. 187–202, 1997

    1997

  3. [3]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,”Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  4. [4]

    Quantum machine learning in feature Hilbert spaces,

    M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,”Physical Review Letters, vol. 122, no. 4, p. 040504, 2019

    2019

  5. [5]

    Quantum vision transformers,

    E. A. Cherratet al., “Quantum vision transformers,”Quantum, vol. 8, p. 1265, 2024

    2024

  6. [6]

    Benchmarking MedMNIST dataset on real quantum hardware,

    G. Singh, H. Jin, and K. M. Merz Jr., “Benchmarking MedMNIST dataset on real quantum hardware,”Scientific Reports, vol. 16, Article no. 9017, 2026

    2026

  7. [7]

    HQViT: Hybrid quantum vision transformer for image classification,

    H. Zhang, Q. Zhao, M. Zhou, and L. Feng, “HQViT: Hybrid quantum vision transformer for image classification,”arXiv preprint arXiv:2504.02730, 2025

    work page arXiv 2025

  8. [8]

    Quantum machine learning for image classifica- tion,

    A. Senokosovet al., “Quantum machine learning for image classifica- tion,”arXiv preprint arXiv:2304.09224, 2024

    work page arXiv 2024

  9. [9]

    Standard SAR ATR evaluation experiments using the MSTAR public release data set,

    T. D. Ross, S. W. Worrell, V . J. Velten, J. C. Mossing, and M. L. Bryant, “Standard SAR ATR evaluation experiments using the MSTAR public release data set,” inProc. SPIE, vol. 3370, pp. 566–573, 1998

    1998

  10. [10]

    PennyLane: Automatic differentiation of hybrid quantum-classical computations

    V . Bergholmet al., “PennyLane: Automatic differentiation of hy- brid quantum-classical computations,”arXiv preprint arXiv:1811.04968, 2018

    work page internal anchor Pith review arXiv 2018

  11. [11]

    Effect of data encoding on the expressive power of variational quantum-machine-learning models,

    M. Schuld, R. Sweke, and J. J. Meyer, “Effect of data encoding on the expressive power of variational quantum-machine-learning models,” Physical Review A, vol. 103, no. 3, p. 032430, 2021

    2021

  12. [12]

    Data re-uploading for a universal quantum classifier,

    A. P ´erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, “Data re-uploading for a universal quantum classifier,”Quantum, vol. 4, p. 226, 2020

    2020

  13. [13]

    The power of quantum neural networks,

    A. Abbaset al., “The power of quantum neural networks,”Nature Computational Science, vol. 1, no. 6, pp. 403–409, 2021

    2021

  14. [14]

    Supervised learning with quantum-enhanced feature spaces,

    V . Havl´ıˇceket al., “Supervised learning with quantum-enhanced feature spaces,”Nature, vol. 567, no. 7747, pp. 209–212, 2019

    2019

  15. [15]

    Target classification using the deep convolutional networks for SAR images,

    S. Chen, H. Wang, F. Xu, and Y .-Q. Jin, “Target classification using the deep convolutional networks for SAR images,”IEEE Trans. Geosci. Remote Sens., vol. 54, no. 8, pp. 4806–4817, 2016

    2016

  16. [16]

    Quantum convolutional neural networks,

    I. Cong, S. Choi, and M. D. Lukin, “Quantum convolutional neural networks,”Nature Physics, vol. 15, no. 12, pp. 1273–1278, 2019

    2019

  17. [17]

    Classification with Quantum Neural Networks on Near Term Processors

    E. Farhi and H. Neven, “Classification with quantum neural networks on near term processors,”arXiv preprint arXiv:1802.06002, 2018

    work page Pith review arXiv 2018

  18. [18]

    Parameterized quantum circuits as machine learning models,

    M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, “Parameterized quantum circuits as machine learning models,”Quantum Science and Technology, vol. 4, no. 4, p. 043001, 2019

    2019

  19. [19]

    SAR automatic target recog- nition method based on multi-stream complex-valued networks,

    Z. Zeng, J. Sun, Z. Han, and W. Hong, “SAR automatic target recog- nition method based on multi-stream complex-valued networks,”IEEE Trans. Geosci. Remote Sens., vol. 60, p. 5228618, 2022

    2022

  20. [20]

    Deep Complex Networks

    C. Trabelsiet al., “Deep complex networks,”arXiv preprint arXiv:1705.09792, 2018

    work page Pith review arXiv 2018

  21. [21]

    Variational quantum algorithms,

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

    2021

  22. [22]

    A performance analysis of convolutional neural network models in SAR target recognition,

    J. Shao, C. Qu, and J. Li, “A performance analysis of convolutional neural network models in SAR target recognition,” inProc. IGARSS, pp. 1142–1145, 2017

    2017

  23. [23]

    SAR ATR based on ResNet with center-softmax loss,

    K. Fu, X. Zhang, and G. Wang, “SAR ATR based on ResNet with center-softmax loss,” inProc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), 2020

    2020