Pith. sign in

REVIEW 3 major objections 7 minor 2 cited by

Nonlinear Computation with Linear Optics via Source-Position Encoding

T0 review · 3 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims that encoding data in the spatial position of a light source makes a fully linear optical medium compute nonlinearly, and that topology-optimized material designs built on this encoding outperform linear models and rival…

desk verdict Source-position encoding is a genuinely new idea with a sound modal derivation, but the experimental section never isolates it from the quadratic intensity readout, so the central attribution is not yet supported. read the letter →

arxiv 2504.20401 v2 pith:GZB7AGAV submitted 2025-04-29 physics.optics cs.ARcs.LG

N. Richardson , C. Bosch , R. P. Adams This is my paper · ORCID
classification physics.opticscs.ARcs.LG
keywords source-positionencodingnonlinearopticslinearmediatopologyoptimizationopticalneuralnetworksdifferentiablesimulationinversedesignmodalexpansion
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 proposes a way to perform nonlinear computation in an optical system made entirely of linear materials. The idea is to encode the input data in the spatial position of a light source rather than in its amplitude: the field measured at a fixed detector is a nonlinear function of the source's position, because it is a superposition of the medium's eigenmodes evaluated at that position. The paper shows this relationship explicitly, and introduces a topology-optimization framework that designs the material distribution so that the source-position-to-field map approximates a target classifier. On quantized MNIST and Fashion-MNIST, the resulting simulated optical network beats a linear model by a wide margin and matches a standard neural network, though it falls behind at high input precision.

What carries the argument

The load-bearing mechanism is the source-position-to-field modal expansion, Equation (8): $E(z_R) = \sum_{i} \frac{-j\omega_s}{\omega_i^2-\omega_s^2} E_i^*(z_s)\,E_i(z_R)$. The term $E_i^*(z_s)$ is the source-position-dependent mode amplitude, and in a heterogeneous engineered medium the eigenmodes $E_i$ can be complex and shaped by the material distribution, which makes the transfer function from source position to measured field highly nonlinear and tunable. The companion mechanism is an end-to-end differentiable design pipeline: a finite-difference frequency-domain (FDFD) discretization of Maxwell's equations, a level-set topology optimization parameterization (B-spline, triangular mesh, or neural implicit field), and Fiber Monte Carlo to differentiate through the level-set boundary. This allows co-optimization of the material distribution and the linear decoder.

What would settle it

Simulate the fully optimized optical classifier on a two-input binary XOR task using the same p=2, q=2 source-position encoding, and compare accuracies of the optical design, a linear model (chance), and a single-hidden-layer MLP (near 100%). If the optical design does not exceed chance, the source-position nonlinearity cannot produce the nonlinear feature interactions that the paper acknowledges are missing; this directly tests whether the elementwise ceiling is real or softened by the intensity readout's cross terms.

Watch

Extended reading notes

Core claim

The central claim is that the map from source position to measured field is nonlinear even when the optical medium is entirely linear, because the field at a fixed readout point is a weighted sum of the medium's eigenmodes, and the weights depend on the source position through the mode functions. Equation (8) expresses this as $E(z_R) = \sum_{i} \frac{-j\omega_s}{\omega_i^2-\omega_s^2} E_i^*(z_s)\,E_i(z_R)$. Since the eigenmodes $E_i$ are generally not affine functions of position, varying the source position $z_s$ produces a nonlinear response at a fixed detector. The paper claims that by optimizing the spatial distribution of permittivity and permeability through topology optimization, one can shape these mode functions so that the source-position-to-field map approximates a desired transfer function, such as a classifier. The full system encodes each input coordinate by activating one source in a row of a source array, propagates the field through the engineered medium, reads intensities at fixed locations, and applies a linear decoder before a softmax. The paper reports that on the evaluated tasks this optical classifier significantly improves over a linear model and is competitive with a standard multilayer perceptron.

Load-bearing premise

The method assumes that an elementwise nonlinearity applied independently to each input coordinate, followed by a linear readout, is expressive enough for the target classification tasks.

Editorial extensions

If this is right

  • Optical neural networks could be built from passive, linear materials with no optoelectronic conversion, because the nonlinearity arises from position encoding rather than material response.
  • The design methodology produces extremely task-specific hardware: the material distribution is optimized for a single function and is not reused across tasks.
  • The approach requires only the ability to drive a source at a data-dependent position, so it works at low power and should generalize to other linear wave systems such as microwaves and acoustics.
  • The performance gap between the optical network and a standard neural network widens as input precision increases (from +15.2% over linear at 2-bit to +2.2% at 64-bit on MNIST), indicating that the encoding precision is a limiting factor.

Reading between the lines

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

  • The intensity readout's squared-magnitude operation creates pairwise cross terms between input coordinates, so the paper's stated elementwise limitation is likely softened: the effective function class is quadratic in the per-coordinate features, which can represent interactions such as XOR that a purely elementwise nonlinearity cannot.
  • Because Equation (8) also depends on source frequency through $\omega_s$, frequency could be used as a second encoding axis alongside position, potentially increasing capacity without additional spatial hardware; the paper flags this as future work.
  • The topology-optimization pipeline built on Fiber Monte Carlo is not optics-specific and could be applied to design other physical computing elements, such as acoustic or microwave classifiers, wherever the forward physics is linear and inputs can be position-coded.
  • A sharper baseline than the numerical linear model would be an optical linear network using the same source array but a random or freespace material distribution, which would isolate the contribution of the optimized topology to the classification gain.
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 / 7 minor

Summary. The manuscript proposes a method for performing nonlinear computation in fully linear optical media by encoding input data in the spatial position of optical sources. The authors derive a modal expansion showing that the field at a fixed measurement point is a nonlinear function of source position (Eq. 8), and use topology optimization with a differentiable finite-difference frequency-domain solver and Fiber Monte Carlo shape derivatives to design material distributions that map the source position to a desired classifier output. The system is evaluated on MNIST and Fashion-MNIST with PCA-projected, quantized inputs at several bit precisions; the optical network is reported to outperform a linear baseline and to approach an MLP at low precision. The paper also discusses limitations, including the claim that the nonlinearity is elementwise.

Significance. If the attribution to source-position encoding is upheld, the work is a significant contribution to the field of optical neural networks: it offers a low-power nonlinearity mechanism that does not rely on materials-based nonlinearities, and it integrates a full inverse-design pipeline (differentiable full-wave simulation, topology optimization, and decoder co-design) that is of independent interest. The modal derivation leading to Eq. (8) is a clear strength, and the use of Fiber Monte Carlo for exact-Heaviside topology differentiation is technically interesting. However, the experimental validation as reported does not isolate the proposed mechanism from the square-law intensity readout, and the stated elementwise limitation is inconsistent with the implemented readout. The central idea is plausible, but the evidence base needs strengthening before the claims can be accepted.

major comments (3)
  1. [§2.2.3–2.2.4, Eq. (11)] The intensity readout R(E) = |E|^2 defined in §2.2.3 is quadratic in the field and therefore in the source vector b = s(x); since the fields from the p active sources (one per row) superpose before detection, R contains cross terms between different rows. This means the aggregate map f in Eq. (11) is not an elementwise nonlinearity of the input coordinates, and the certification in §2.2.4 that the decoder is linear does not isolate the source-position encoding as the source of the nonlinearity. The comparison against the linear baseline conflates at least three mechanisms: the one-hot quantized position encoding, the square-law readout, and the optimized material distribution. A system with fixed sources and amplitude-modulated inputs would also be nonlinear through the same readout, so the reported gains cannot be attributed to source-position encoding without a control experiment (e.g., amplitude encoding with fixed positions, or a field-amplitude readout). Please add such a control or otherwise analyze the separate contributions.
  2. [§4.1] The stated limitation that 'the nonlinearities hold only elementwise with respect to the data' is contradicted by the actual system because the intensity readout creates products of contributions from different rows of the source array, i.e., nonlinear mixtures of variables from different input dimensions. Consequently, the explanation of the high-precision gap in Table 1 (MNIST 64-bit: 86.7% vs 96.9%) as a capacity limit of elementwise nonlinearities is not supported by the implemented model. The authors should either correct the limitation statement to reflect the readout-induced cross terms, or analyze the functional form of the map and re-interpret the accuracy gap.
  3. [§3, Table 1] The experimental results are reported as single runs without error bars, multiple random initializations, or statistical tests. At the smaller margins (e.g., MNIST 64-bit, +2.2% relative improvement), the claimed 'significant improvements' over the linear baseline are not statistically substantiated. Please report the variance across multiple optimization runs and seeds and, if appropriate, a significance test.
minor comments (7)
  1. [§2.1, Eq. (8)] The inner product in Eq. (4) includes the weight ε_r(z), so the modal coefficient for a point source should read c_i = -j ω_s/(ω_i^2 - ω_s^2) ε_r(z_s) E_i^*(z_s); the ε_r(z_s) factor is missing in Eq. (8). If the source array is placed in a region of constant ε_r the factor is a global constant, but the general formula should be corrected or the constant-ε_r assumption stated.
  2. [§2.1, Eqs. (1)–(2)] Dirichlet boundary conditions correspond to a perfect electric conductor enclosing the domain; this is a strong idealization for an open scattering device and should be mentioned as an approximation in the text, since it may introduce cavity resonances that are not present in a free-space device.
  3. [§2.2.1 and §3.1] Section 2.2.1 states 'log(k) bits of precision' while Section 3.1 correctly writes 'log2 q'; please use consistent notation for the number of quantization bits.
  4. [Table 1 caption] The parenthetical values are labeled as relative percentage improvement in the caption, but the text does not define how the percentages are computed; please provide the definition in the caption or text.
  5. [Figure 2] The decoder box labels 'ω_d → R^d → (k→k)' are unclear; clarifying the map from the d-dimensional input to the K readout channels would improve readability.
  6. [Appendix A.2] In the notation list, the units for the electric current density J are given as W/m^2, which is a power density; the units should be A/m^2.
  7. [§2.3.1 and Figure 3 caption] The cross-reference for the neural implicit field parameterization is missing a target: the text says '(cref appendix)' without an actual reference; please fix the cross-reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity; the derivation is self-contained and the only self-citation is not load-bearing.

full rationale

The paper's central derivation, Eq. (8), follows from the spectral expansion of the wave operator and an explicit point-source substitution; it does not assume the classifier result it later reports. The inverse-design objective (14) fits topology and decoder parameters to training data, but the reported Table 1 accuracies are evaluated on out-of-sample test data, so they are not forced by construction. The only overlapping-author citation is Fiber Monte Carlo (Richardson et al., 2024), used as a differentiable area-estimation tool for topology optimization; this supports a numerical method rather than supplying the predicted nonlinearity, and the experimental outcomes remain externally testable. The concern that intensity readout introduces an additional nonlinear mechanism not isolated by the ablations is an experimental attribution issue, not an instance of a prediction reducing to its inputs by definition or by self-citation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central physics derivation introduces no free parameters or invented entities: it relies on the spectral theorem and standard material assumptions. The machine learning pipeline adds learned parameters (PCA projection, linear decoder) and chosen hyperparameters (p, q), which are not part of the nonlinearity claim. The method's expressivity rests on the unstated assumption that elementwise nonlinearities on PCA features suffice for the evaluated tasks.

free parameters (3)
  • PCA projection matrix = Learned from training set for each (p, q) configuration.
    Reduces 784-dimensional input to p dimensions before quantization; the paper does not treat this as part of the optical nonlinearity, but the evaluation depends on it.
  • Linear decoder weights theta_d = Learned by Adam during training.
    The output classifier weights are fitted to the training data; the paper keeps them linear to isolate the nonlinearity, but they are still required for the classification result.
  • Encoding hyperparameters (p, q) = p in {4,8,16}, q in {4,8,16}.
    Chosen by hand to set the bit budget (8, 24, 64 bits); they define the source array size and precision, and performance varies with them.
assumptions (6)
  • standard math Spectral theorem and normal mode expansion for the self-adjoint wave operator.
    Used in Section 2.1 to express the field as a sum of eigenmodes; no independent verification needed for this classical result.
  • domain assumption Linear, isotropic, lossless, non-dispersive media with real scalar mu_r and epsilon_r.
    Assumed at the start of Section 2.1 and in Appendix A.2; validates the mode expansion and the topology optimization model, but is an idealization.
  • domain assumption Dirichlet boundary conditions on the computational domain.
    Eq. (2) sets E=0 on the boundary; this is a computational convenience that changes the mode structure relative to open or periodic boundaries.
  • domain assumption Monochromatic point dipole source model.
    Used to derive Eq. (8) and to model the source array; the field of a physical finite-size source would differ.
  • ad hoc to paper Elementwise encoding with one active source per row and uniform quantization.
    Section 2.2.1 defines this encoding; it is the chosen input interface and is not derived from the physics.
  • ad hoc to paper Sufficiency of elementwise nonlinearity plus linear readout for the evaluated tasks.
    The architecture assumes that a per-coordinate nonlinear map followed by a linear classifier is expressive enough; Section 4.1 admits this is a limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Nonlinear Computation with Linear Optics via Source-Position Encoding." pith.science (2026). https://pith.science/paper/GZB7AGAV

@misc{pith2026250420401,
  author       = {Pith},
  title        = {Pith review of: Nonlinear Computation with Linear Optics via Source-Position Encoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZB7AGAV}},
  note         = {Machine review of arXiv:2504.20401}
}
read the original abstract

Optical computing systems provide an alternate hardware model which appears to be aligned with the demands of neural network workloads. However, the challenge of implementing energy efficient nonlinearities in optics -- a key requirement for realizing neural networks -- is a conspicuous missing link. In this work we introduce a novel method to achieve nonlinear computation in fully linear media. Our method can operate at low power and requires only the ability to drive the optical system at a data-dependent spatial position. Leveraging this positional encoding, we formulate a fully automated, topology-optimization-based hardware design framework for extremely specialized optical neural networks, drawing on modern advancements in optimization and machine learning. We evaluate our optical designs on machine learning classification tasks: demonstrating significant improvements over linear methods, and competitive performance when compared to standard artificial neural networks.

Figures

Figures reproduced from arXiv: 2504.20401 by the authors.

Figure 1
Figure 1. Three material distributions (vacuum with ϵr = 1, and two arbitrarily chosen distributions, with ϵr = 3 in blue regions and ϵr = 1 in white regions; µr = 1 for all of them) and the 1st, 3rd and 6th corresponding eigenmodes. Clearly, the mode profiles can be quite distinct from the basic freespace modes (and one another) by varying the material distribution. Complex mate￾rial distributions give rise to complex modes,… view at source ↗
Figure 2
Figure 2. The input data x is used to determine which subset of the array of sources are active (illustrated in red). This collection of active sources is assembled into an aggregate electromagnetic source b := s(x). Below the aggregate source field we illustrate the geometry (hardware) A(θg), which does not depend on the data but is optimized via a gradient method. The source and geometry are simulated in a differentiable fi… view at source ↗
Figure 3
Figure 3. Hardware parameterization instances. On the left, we show B spline patches whose interior determine freespace pores amidst a homogeneous dielectric (e.g., SiO2). The parameters θH modulate the shape of the pores through the spline control points p1, . . . , pn. In the center, a triangular mesh is parameterized by a collection of n vertices v1, . . . , vn, and permittivities ϵ1, . . . , ϵm within the interior of each… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Evolution of design topology over the course of opti￾mization. The left-most column shows the initialization of the topology (before optimization), the right-most column shows the final topology (after optimization) and the center column illustrates the topology at an …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Unifying Physical Backpropagation

    cond-mat.dis-nn 2026-08 accept novelty 7.0 of 10

    A single adjoint framework identifies when the same physical hardware that computes forward can also compute exact gradients, with distinct conditions for linear and nonlinear systems.

  2. Local Learning Rules for Out-of-Equilibrium Physical Generative Models

    cs.LG 2025-06 conditional novelty 6.0 of 10

    Force matching, a local learning rule derived from score matching, can train nonlinear oscillator networks as out-of-equilibrium score-based generative models.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages · cited by 2 Pith papers

  1. [5]

    S., Zimmer, B., Nedovic, N., Liu, N., Venkatesan, R., Wang, M., Khailany, B., Dally, W

    Rekhi, A. S., Zimmer, B., Nedovic, N., Liu, N., Venkatesan, R., Wang, M., Khailany, B., Dally, W. J., and Gray, C. T. Analog/mixed-signal hardware error modeling for deep learning inference. In Proceedings of the 56th Annual Design Automation Conference 2019, pp. 1–6,

    work page 2019

  2. [7]

    Appendix A.1

    9 Nonlinear Computation with Linear Optics via Source-Position Encoding A. Appendix A.1. Notation R⪰0 The set of nonnegative real numbers{x∈ R|x≥ 0}. R≻0 The set of positive real numbers{x∈ R|x> 0}. T⊆ R⪰0 An alias for the nonnegative reals used to suggest a temporal coordinate. ∂X Boundary of the setX, i.e., the set difference of its closure against its ...

    work page 2012

  3. [9]

    study what could be called the general case of the work presented in (Xia et al., 2024), providing an analysis of the generic technique of data repitition-based strategies. The core argument is that the repitition within diffractive volumes (i.e., the optical cavity used in (Xia et al., 2024)) cannot strictly serve as optical implementations of most conve...

    work page 2024

  4. [10]

    11 Nonlinear Computation with Linear Optics via Source-Position Encoding A.3.3

    but utilizes multiple modulating planes where data is repeatedly embedded. 11 Nonlinear Computation with Linear Optics via Source-Position Encoding A.3.3. I NVERSE DESIGN IN PHOTONICS Inverse design of electromagnetic devices is widely studied (Molesky et al., 2018; Christiansen & Sigmund, 2021; Georgieva et al., 2002), ranging from fairly application spe...

    work page 2018

  5. [12]

    Our method obviates material-derived nonlinearities and an exact topology optimization method, which we can differentiate through using Fiber Monte Carlo (Richardson et al., 2024)

    depend on a material- derived nonlinearity (i.e., the electrical response of a hypothetical optical saturable absorber), and a discrete smoothed level set topology optimization method drawn from the computer vision literature (Li et al., 2010). Our method obviates material-derived nonlinearities and an exact topology optimization method, which we can diff...

    work page 2010

  6. [1855]

    W., Minkov, M., Shi, Y ., and Fan, S

    7 Nonlinear Computation with Linear Optics via Source-Position Encoding Hughes, T. W., Minkov, M., Shi, Y ., and Fan, S. Training of Photonic Neural Networks through in Situ Backpropaga- tion and Gradient Measurement. Optica, 5(7):864–871, 2018a. Hughes, T. W., Minkov, M., Williamson, I. A. D., and Fan, S. Adjoint Method and Inverse Design for Non- linear...

    doi:10.1021/acsphotonics.8b01522

  7. [1930]

    Subsequently, a wide body of work has explored nonlinear interactions with respect to intensity, frequency, phase, and polarization (Boyd et al., 2008)

    and then observed in the early 1960s at Bell Labs (Bayer & Schaack, 1970). Subsequently, a wide body of work has explored nonlinear interactions with respect to intensity, frequency, phase, and polarization (Boyd et al., 2008). Nonlinearity is observed in all media (including the vacuum) above some finite field intensity (Sauter, 1931). Unfortunately, it ...

    work page 1970

  8. [1931]

    URL https://doi.org/10

    doi: 10.1007/BF01339461. URL https://doi.org/10. 1007/BF01339461. Schuman, C. D., Kulkarni, S. R., Parsa, M., Mitchell, J. P., Date, P., and Kay, B. Opportunities for Neuromorphic Computing Algorithms and Applications. Nature Com- putational Science, 2(1):10–19,

    doi:10.1007/bf01339461

Show all 12 references
  1. [2002]

    to flexible structural optimization with full-wave field solvers (Hughes et al., 2018b; 2019). We use the adjoint method, which amounts to implicitly differentiating through our full-wave field solver to compute derivatives with respect to the geometry and material properties ...

    2019

  2. [2018]

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J

    URL http://github.com/jax-ml/jax. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language Models are Few-shot Learners. Advances in neural information processing systems, 33: 1877–1901,

    1901

  3. [2019]

    Kingma, D. P. and Ba, J. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980,

    arXiv

  4. [2024]

    Maclaurin, D., Duvenaud, D., and Adams, R. P. Autograd: Effortless Gradients in NumPy. In ICML 2015 AutoML workshop, volume 238,

    2015

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.