Pith. sign in

REVIEW 3 major objections 5 minor 20 references

Gradients of unitary optical neural networks using parameter-shift rule

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper derives exact parameter-shift gradient formulas for unitary optical neural networks, giving intensity and complex-field rules that require only shifted forward evaluations.

desk verdict Correct but derivative: the PSR formulas are right, yet the paper overclaims novelty and skips the chain rule from field derivative to loss gradient. read the letter →

arxiv 2506.11565 v1 pith:IROXOI76 submitted 2025-06-13 cs.ET cs.LGphysics.optics

Jinzhe Jiang , Yaqian Zhao , Xin Zhang , Chen Li , Yunlong Yu , Hailing Liu This is my paper · ORCID
classification cs.ETcs.LGphysics.optics
keywords parameter-shiftruleunitaryopticalneuralnetworkMach-Zehnderinterferometermeshgradientcomputationtrainingphaseshifterphotoniccomputingexactanalytical
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that gradients of unitary optical neural networks built from Mach–Zehnder interferometer meshes can be computed exactly by evaluating the network at shifted phase settings, without optical backpropagation or finite differences. For photodiode-measured intensity outputs, the gradient with respect to a phase $\theta$ is $\frac{1}{2}[f(\theta+\pi/2)-f(\theta-\pi/2)]$. For complex field outputs that pass unmeasured between layers, the gradient is $\frac{1-i}{2}[f(\theta+\pi/2)-f(\theta)]$. If correct, these formulas make in-hardware training of optical neural networks practical and avoid the noise of finite-difference estimates.

What carries the argument

The load-bearing object is the phase-shifter generator $G=\mathrm{diag}(1,0)$, which encodes the fact that one arm of the MZI accumulates phase $\theta$ while the other does not. Because $G$ has eigenvalues $1$ and $0$, the general parameter-shift rule for two-eigenvalue generators fixes the shift as $\pm\pi/(4r)=\pm\pi/2$. For complex fields, the same generator furnishes the operator identity $G=(I-U_j(\pi/2))/(1-i)$, which converts the derivative into one shifted evaluation. The mesh-level structures of Reck and Clements then guarantee that every phase parameter sits inside such an MZI, so the pointwise rule extends to the whole network.

What would settle it

Simulate or measure the gradient of an MZI mesh under finite insertion loss and beam-splitter imbalance; if the formula $\frac{1}{2}[f(\theta+\pi/2)-f(\theta-\pi/2)]$ no longer matches the exact derivative of the output intensity, the unitarity assumption has been violated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the phase shifter in an MZI is generated by the Hermitian matrix $G=\mathrm{diag}(1,0)$, which has exactly two eigenvalues, so the standard parameter-shift rule applies with shift $\Delta\theta=\pi/2$. Substituting this generator into the PSR yields $\frac{d|E_{o1}|^2}{d\theta_1}=\frac{1}{2}\left[|E_{o1}(\theta_1+\pi/2)|^2-|E_{o1}(\theta_1-\pi/2)|^2\right]$, and similarly for the second phase shifter and other measurement projectors. When the output is a complex electric field, the paper bypasses the real-output construction and derives $\frac{df(\theta)}{d\theta}=\frac{1-i}{2}[f(\theta+\pi/2)-f(\theta)]$ directly from $\frac{dU_j(\theta)}{d\theta}=iGU_j(\theta)$. The formulas apply to any MZI mesh obtained by Reck or Clements decompositions, which is what makes them gradient rules for whole UONNs.

Load-bearing premise

The derivation presumes that every phase shifter is exactly a lossless unitary $\mathrm{diag}(e^{i\theta},1)$ between ideal 50:50 beam splitters, and that intensities or fields can be read out without error; any loss, imbalance, or detector noise makes the shift formulas approximate.

Editorial extensions

If this is right

  • Training a UONN requires only forward evaluations at phase shifts $\theta\pm\pi/2$, so gradients can be measured directly on chip without a separate backpropagation light path.
  • The intensity rule gives exact gradients for photodiode outputs, eliminating the bias and noise sensitivity of finite-difference estimates.
  • The complex-field rule extends exact gradients through unmeasured intermediate layers, so multi-layer UONNs can be trained end to end with shifted evaluations.
  • Any MZI mesh built from Reck or Clements decompositions inherits the rule, making it compatible with standard UONN hardware.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Nothing in the derivation is special to MZIs: any linear optical element whose phase parameter is generated by a Hermitian operator with two distinct eigenvalues admits the same $\pi/2$-shift rule, so the method should transfer to other tunable interferometer families.
  • In a realistic chip, the equality holds only for the ideal unitary model; quantifying the bias from insertion loss and splitter asymmetry would determine how many averaged measurements are needed for a given training accuracy.
  • Plugging these formulas into a standard automatic-differentiation simulator would let a digital twin train the phase parameters while the physical device supplies the shifted evaluations, a hybrid loop the paper does not spell out.
  • The complex-field rule may enable fully coherent multi-layer training where gradients are obtained from field measurements made by homodyne detection, not just intensity.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes applying the parameter-shift rule (PSR) to unitary optical neural networks (UONNs) built from Mach-Zehnder interferometer meshes. For real-valued intensity outputs, it derives that the gradient of output intensity with respect to a phase parameter θ is 1/2 [f(θ+π/2)-f(θ-π/2)]. For complex-valued field outputs, it derives the identity df/dθ = (1-i)/2 [f(θ+π/2)-f(θ)] for the complex field derivative. The paper claims that these formulas provide exact analytical gradients computable directly from hardware measurements, thereby enabling backpropagation training of UONNs and avoiding the need for in-silico training or finite-difference approximations.

Significance. The real-output portion of the paper is a correct and clean application of the standard parameter-shift rule to MZI-based UONNs, and it is presented in a self-contained way. The algebraic derivation of the complex-field derivative in §2.3.2 is also correct under the stated ideal unitary model. If the paper were limited to intensity outputs, the result would be a useful theoretical note connecting PSR to classical optical hardware. However, the central training claim for complex-valued outputs is not established: the paper stops at the complex field derivative and never connects it to the real scalar loss gradient required for gradient-based optimization. As written, the contribution is a modest but sound derivation for the real-output case and an incomplete proposal for the complex-output case. No numerical experiments or robustness analysis are provided, so the practical claims of exactness from hardware measurements remain unvalidated.

major comments (3)
  1. [§2.3.2] The complex-output result computes df/dθ for the complex field f(θ), but training a neural network requires the gradient of a real scalar loss L(θ) = ℓ(f(θ), conj(f(θ))). The paper never defines such a loss, never introduces the chain rule for real losses through complex fields, and never connects df/dθ to the parameter update. For example, for L = |f - t|², the correct gradient is dL/dθ = 2 Re[(f - t)* df/dθ], which involves the adjoint/error field (f - t)* and a real-part projection. Without this step, the claim that the complex-output PSR 'enables backpropagation training' is unsupported. Please add the Wirtinger/chain-rule step or explicitly restrict the claim to the field derivative itself.
  2. [Abstract, §2.3.1, §3] The paper repeatedly claims 'exact analytical gradients directly from hardware measurements,' but this holds only under an ideal model with lossless, perfectly balanced MZIs, exact π/2 phase shifts, and ideal intensity or field detection. Real hardware has insertion loss, beam-splitter imbalance, phase-setting errors, and detector noise, all of which break the exactness of the shift formula. The manuscript provides no error or robustness analysis and no discussion of how these nonidealities affect the gradient estimate. Please add an explicit statement of the idealization and, ideally, a first-order sensitivity analysis or a numerical demonstration.
  3. [§2.3.2] For complex-valued outputs, the paper does not explain how the quantities f(θ) and f(θ+π/2) would be obtained from physical hardware. Unlike intensity, the complex field at an intermediate layer is not directly accessible without phase-resolved (e.g., interferometric) detection, and any tapping or probing scheme may itself perturb the unitary evolution. Please clarify the assumed measurement model for complex fields or restrict the method to simulated complex outputs.
minor comments (5)
  1. [§2.2] In the equation for dU_j†(θ_j)/dθ_j, the denominator is written as 'θj' instead of 'dθj'; please fix this typo.
  2. [§2.2.2] The notation for the Hermitian generator shift, Δθ = π/(4r), is correct, but the sign of r is not discussed; for the optical phase shifter r = -1/2, which leads to a negative shift that is then absorbed into the final symmetric form. A brief remark would help readers avoid sign confusion.
  3. [§2.3.1] The paper says the intensity expression 'is the same as the expression for f(θ) in the parameter-shift method,' but the connection is only valid when the observable is a projector Z on output modes. Please make this assumption explicit.
  4. [Throughout] There are minor language and formatting issues, e.g., 'beamspliter' should be 'beam splitter,' 'calculate the gradients' should be 'calculating the gradients' in §3, and several references lack full bibliographic information.
  5. [§3] The conclusion states that PSR 'circumvents the limitations of ... all-optical backpropagation' and 'avoids the approximation errors often associated with methods like finite differences,' but no comparison with finite-difference or backpropagation methods is provided. A small numerical example or error comparison would substantiate these claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PSR formulas follow from the stated generator and standard parameter-shift algebra, with no fitted inputs or self-citation chain.

full rationale

The paper's derivation is self-contained. For the real-output case, the loss is defined as f(θ)=Tr(Z W E E† W†), the phase shifter is written as U_PS(θ)=exp(-iaθG) with G=diag(1,0), eigenvalues 1 and 0, and a=-1, so r=-1/2 and the shift constant π/(4r) = -π/2. Substituting these values into the standard two-eigenvalue parameter-shift rule gives d f/dθ = 1/2[f(θ+π/2)-f(θ-π/2)]. This is a direct algebraic application of an external, well-established rule, not a quantity defined in terms of itself. For the complex-output case, the paper computes dU_j/dθ = iG U_j(θ) and uses I-U_j(π/2) = (1-i)G, yielding d f/dθ = i/(1-i)[f(θ)-f(θ+π/2)] = (1-i)/2[f(θ+π/2)-f(θ)]. Again, the result follows from the explicit operator identities and the definition of f(θ); there is no fitting, no parameter calibrated to a subset of outputs, and no load-bearing self-citation. The citations to standard PSR literature and to prior MZI mesh literature are contextual, not used to justify the derived identities. The skeptic's concern that the complex-output formula is a field derivative rather than the gradient of a real loss is a completeness or correctness risk, not a circularity: the derivation does not assume the conclusion it claims to prove. The paper makes no empirical prediction that reduces to a fitted input, and no uniqueness claim is imported from the authors' own prior work. Under the stated ideal unitary model, the gradient formulas are genuine consequences of the model equations.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The derivation is self-contained given the standard PSR for generators with two eigenvalues. No free parameters are fitted, and no new physical entities are introduced. All assumptions are physical idealizations of the optical hardware, primarily lossless unitarity and availability of shifted parameter evaluations.

assumptions (4)
  • domain assumption The MZI mesh is an ideal lossless unitary transformation composed of 50:50 beam splitters and phase shifters U(theta)=e^{i theta G}, with G=diag(1,0).
    Section 2.3 defines the component transfer matrices; no loss, fabrication errors, or detector non-idealities are modeled.
  • domain assumption Output intensity is proportional to |E_out|^2 and is expressible as an expectation value of a linear observable on the coherent optical field.
    Section 2.3.1 uses |Eo1|^2 = Tr(Z1 W E E^dag W^dag), treating detection as an ideal square-law measurement.
  • standard math The PSR for a generator with two eigenvalues, d/dtheta f = r [f(theta+pi/(4r)) - f(theta-pi/(4r))], is valid.
    Section 2.2.2 invokes this standard result without proof, relying on the known literature.
  • domain assumption The complex field f(theta) at intermediate layers is an available, evaluable quantity.
    Section 2.3.2 derives the complex-output rule using f(theta) and f(theta+pi/2), implying these values can be obtained. This is not generally true for hardware-hidden layers where intermediate optical fields are not measurable without tapping or homodyne detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Gradients of unitary optical neural networks using parameter-shift rule." pith.science (2026). https://pith.science/paper/IROXOI76

@misc{pith2026250611565,
  author       = {Pith},
  title        = {Pith review of: Gradients of unitary optical neural networks using parameter-shift rule},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IROXOI76}},
  note         = {Machine review of arXiv:2506.11565}
}
read the original abstract

This paper explores the application of the parameter-shift rule (PSR) for computing gradients in unitary optical neural networks (UONNs). While backpropagation has been fundamental to training conventional neural networks, its implementation in optical neural networks faces significant challenges due to the physical constraints of optical systems. We demonstrate how PSR, which calculates gradients by evaluating functions at shifted parameter values, can be effectively adapted for training UONNs constructed from Mach-Zehnder interferometer meshes. The method leverages the inherent Fourier series nature of optical interference in these systems to compute exact analytical gradients directly from hardware measurements. This approach offers a promising alternative to traditional in silico training methods and circumvents the limitations of both finite difference approximations and all-optical backpropagation implementations. We present the theoretical framework and practical methodology for applying PSR to optimize phase parameters in optical neural networks, potentially advancing the development of efficient hardware-based training strategies for optical computing systems.

Figures

Figures reproduced from arXiv: 2506.11565 by the authors.

Figure 1
Figure 1. Schematic diagram of unitary optical neural network. The light blue rectangular boxes represent optical [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of a optical layer with real-valued output. After detection, the intensity or amplitude of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of a optical layer with complex-valued output. All light information with phase is used as [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages

  1. [4]

    Svetlana N

    doi:10.1063/5.0169810. Svetlana N. Khonina, Nikolay L. Kazanskiy, Roman V . Skidanov, and Muhammad A. Butt. Exploring types of photonic neural networks for imaging and computing—a review.Nanomaterials, 14:697, 01

    work page doi:10.1063/5.0169810

  2. [5]

    URLhttps://www.mdpi.com/2079-4991/14/8/697

    doi:10.3390/nano14080697. URLhttps://www.mdpi.com/2079-4991/14/8/697. Kaixuan Wei, Xiao Li, Johannes Froech, Praneeth Chakravarthula, James Whitehead, Ethan Tseng, Arka Ma- jumdar, and Felix Heide. Spatially varying nanophotonic neural networks.Science Advances, 10, 11

    work page doi:10.3390/nano14080697

  3. [6]

    doi:10.1126/sciadv.adp0391. M. Matuszewski, A. Prystupiuk, and A. Opala. Role of all-optical neural networks.Physical Review Applied, 21, 01

    work page doi:10.1126/sciadv.adp0391

  4. [7]

    doi:10.1103/physrevapplied.21.014028. Fu Feng, Dewang Huo, Ziyang Zang, Yijie Lou, Shengyao Wang, Zhijuan Gu, Dong-Sheng Liu, Xinhui Duan, Daqian Wang, Xiaowei Liu, Ji Qi, Shaoliang Yu, Qingyang Du, Guangyong Chen, Cuicui Lu, Yu Yu, Xifeng Ren, and Xiaocong Yuan. Symbiotic evolution of photonics and artificial intelligence: a comprehensive review. Advance...

    work page doi:10.1103/physrevapplied.21.014028

  5. [8]

    URL https://doi.org/10.1117/1.AP

    doi:10.1117/1.AP.7.2.024001. URL https://doi.org/10.1117/1.AP. 7.2.024001. James Spall, Xianxin Guo, and Alexander I. Lvovsky. Training neural networks with end-to-end optical backpropagation. Advanced Photonics, 7(1):016004,

    doi:10.1117/1.ap.7.2.024001

  6. [9]

    URL https://doi.org/10.1117/1.AP

    doi:10.1117/1.AP.7.1.016004. URL https://doi.org/10.1117/1.AP. 7.1.016004. Tyler W Hughes, Momchil Minkov, Yu Shi, and Shanhui Fan. Training of photonic neural networks through in situ backpropagation and gradient measurement.Optica, 5:864–864, 07

    work page doi:10.1117/1.ap.7.1.016004

  7. [14]

    Giorgio Facelli

    doi:10.1103/physreva.109.062429. Giorgio Facelli. Exact gradients for linear optics with single photons,

    work page doi:10.1103/physreva.109.062429

  8. [16]

    Ryan Hamerly, Saumil Bandyopadhyay, and Dirk Englund

    doi:10.37190/oa240101. Ryan Hamerly, Saumil Bandyopadhyay, and Dirk Englund. Accurate self-configuration of rectangular multiport interferometers.Physical Review Applied, 18, 08

    work page doi:10.37190/oa240101

Show all 20 references
  1. [18]

    URLhttps://doi.org/10.1063/5.0053421

    doi:10.1063/5.0053421. URLhttps://doi.org/10.1063/5.0053421. 7 arXivTemplateA PREPRINT Michael Reck, Anton Zeilinger, Herbert J. Bernstein, and Philip Bertani. Experimental realization of any discrete unitary operator.Physical Review Letters, 73:58–61, 07

    doi

  2. [1986]

    Mingfeng Li

    doi:10.1038/323533a0. Mingfeng Li. Comprehensive review of backpropagation neural networks.Academic journal of science and technology, 9:150–154, 01

    doi

  3. [1994]

    William R Clements, Peter C Humphreys, Benjamin J Metcalf, Kolthammer W Steven, and Ian A Walmsley

    doi:10.1103/physrevlett.73.58. William R Clements, Peter C Humphreys, Benjamin J Metcalf, Kolthammer W Steven, and Ian A Walmsley. An optimal design for universal multiport interferometers,

    doi

  4. [2016]

    URLhttps://arxiv.org/abs/1603.08788. 8

    arXiv

  5. [2017]

    URL https://www.nature.com/articles/ nphoton.2017.93

    doi:10.1038/nphoton.2017.93. URL https://www.nature.com/articles/ nphoton.2017.93. Kunrun Lu and Xianxin Guo. Efficient training of unitary optical neural networks.Optics Express, 31:39616–39616, 10

    2017 doi

  6. [2018]

    Yichen Shen, Nicholas C

    doi:10.1364/optica.5.000864. Yichen Shen, Nicholas C. Harris, Scott Skirlo, Mihika Prabhu, Tom Baehr-Jones, Michael Hochberg, Xin Sun, Shijie Zhao, Hugo Larochelle, Dirk Englund, and Marin Soljaˇci´c. Deep learning with coherent nanophotonic circuits.Nature Photonics, 11:441–446, 06

    doi

  7. [2020]

    Liubov Markovich, Savvas Malikis, Stefano Polla, and Jordi Tura

    doi:10.1038/s41586-020-2973-6. Liubov Markovich, Savvas Malikis, Stefano Polla, and Jordi Tura. Parameter shift rule with optimal phase selection. Physical review. A/Physical review, A, 109, 06

    doi

  8. [2021]

    Zhang Ye, Wang Ruiting, Zhang Yejin, Su Yanmei, Wang Pengfei, Luo Guangzhen, Zhou Xuliang, and Pan Jiaoqing

    doi:10.1364/oe.432787. Zhang Ye, Wang Ruiting, Zhang Yejin, Su Yanmei, Wang Pengfei, Luo Guangzhen, Zhou Xuliang, and Pan Jiaoqing. Design of a photonic unitary neural network based on mzi arrays.Optica Applicata, 54,

    doi

  9. [2022]

    doi:10.1103/physrevapplied.18.024019. B. A. Bell and I. A. Walmsley. Further compactifying linear optical unitaries.APL Photonics, 6, 07

    doi

  10. [2023]

    Gordon Wetzstein, Aydogan Ozcan, Sylvain Gigan, Shanhui Fan, Dirk Englund, Marin Soljaˇci´c, Cornelia Denz, David A

    doi:10.1364/oe.500544. Gordon Wetzstein, Aydogan Ozcan, Sylvain Gigan, Shanhui Fan, Dirk Englund, Marin Soljaˇci´c, Cornelia Denz, David A. B. Miller, and Demetri Psaltis. Inference in artificial intelligence with deep optics and photonics.Nature, 588: 39–47, 12

    doi

  11. [2024]

    Hanieh Masoudian Saadabad, Lingraj Kumar, Reza Masoudian Saadabad, and Maja Colautti

    doi:10.54097/51y16r47. Hanieh Masoudian Saadabad, Lingraj Kumar, Reza Masoudian Saadabad, and Maja Colautti. Physics-constrained neural network for metasurface optical response prediction,

    doi

  12. [2025]

    Apostolos Tsakyridis, Miltiadis Moralis-Pegios, George Giamougiannis, Manos Kirtas, Nikolaos Passalis, Anastasios Tefas, and Nikos Pleros

    URLhttps://arxiv.org/abs/2504.06069. Apostolos Tsakyridis, Miltiadis Moralis-Pegios, George Giamougiannis, Manos Kirtas, Nikolaos Passalis, Anastasios Tefas, and Nikos Pleros. Photonic neural networks and optics-informed deep learning fundamentals.APL photonics, 9, 01

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.