pith. sign in

arxiv: 2606.23159 · v1 · pith:ZHULTPMQnew · submitted 2026-06-22 · 💻 cs.ET · physics.optics

General-Purpose Nonlinear Function Approximation via Linear Integrated Photonics

Ayana Mizuno , Isamu Takai , Makoto Nakai , Atsutaka Miyamichi , Ryuichi Konishi , Satoshi Sunada This is my paper

Pith reviewed 2026-06-26 06:10 UTC · model grok-4.3

classification 💻 cs.ET physics.optics
keywords photonic computingrandom Fourier featuresnonlinear function approximationsilicon photonicsoptical computingfunction approximationneural network hardwareoptoelectronic computing
0
0 comments X

The pith

Optical random Fourier feature mapping turns high-order nonlinear functions into linear computations in silicon photonic circuits.

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

The paper shows that nonlinear function approximation becomes possible in photonic hardware by first mapping the target function to random Fourier features, then performing the resulting linear operations optically. This sidesteps the usual requirement for nonlinear optical materials or active devices. A reader would care because current photonic accelerators excel at linear matrix multiplies but struggle with the nonlinear stages needed for neural networks and other workloads; the method promises to add those stages while keeping the hardware simple and scalable. Experiments cover tenth-order Legendre polynomials, Voigt, Fermi-Dirac and Fresnel functions, common activation functions, two-dimensional nonlinear maps, and a ten-dimensional softmax layer, all realized in a single silicon photonic circuit.

Core claim

By encoding nonlinear function evaluation as an equivalent linear computation through optical random Fourier feature mapping, high-order and high-dimensional nonlinear functions can be approximated accurately using only linear photonic operations in a simple silicon circuit, without complex nonlinear or active materials.

What carries the argument

Optical random Fourier feature mapping, which converts nonlinear function evaluation into an equivalent linear vector-matrix multiplication executed in the photonic domain.

If this is right

  • High-order polynomials and computationally intensive special functions become feasible on linear photonic hardware.
  • Neural-network activation functions and multi-dimensional operations such as softmax can be realized without additional nonlinear elements.
  • The same circuit architecture supports both linear and nonlinear stages, preserving overall throughput and scalability.
  • A pathway opens toward fully functional optical accelerators that handle complete AI workloads in integrated silicon photonics.

Where Pith is reading between the lines

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

  • Existing linear photonic processors could be extended with this mapping layer to perform end-to-end inference without hybrid electronic nonlinear stages.
  • The approach might generalize to other kernel approximations if analogous linear embeddings can be implemented optically.
  • Reduced material complexity could lower fabrication barriers for photonic neural-network chips.

Load-bearing premise

That random Fourier feature mapping realized optically will deliver sufficiently accurate approximations for the target functions even under realistic photonic noise and fabrication variation.

What would settle it

A direct measurement of the mean-squared error between the photonic circuit output and the analytic value of a tenth-order Legendre polynomial over its domain, compared against the theoretical error bound of the random Fourier feature method.

read the original abstract

Photonic computing has emerged as a promising platform for accelerating artificial intelligence workloads by enabling low-latency and energy-efficient linear operations such as vector-matrix multiplication. However, scalable on-chip high-order nonlinear processing remains challenging, limiting the functional versatility of current photonic hardware. Here, we present an optoelectronic approach for approximating high-order and high-dimensional nonlinear functions. The key to this approach lies in optical random Fourier feature mapping, which transforms nonlinear function evaluation into an equivalent linear computation. This approach enables nonlinear computing within a linear photonic framework, eliminating the need for complex optical nonlinear or active materials while preserving scalability and computational throughput in a simple silicon photonic circuit. We experimentally demonstrate a broad class of nonlinear functions, including tenth-order Legendre polynomials, computationally demanding special functions (Voigt, Fermi-Dirac, and Fresnel), neural-network activation functions, two-dimensional nonlinear functions, and a 10-dimensional softmax layer. This work establishes a general and scalable strategy for nonlinear computing in photonic integrated hardware and opens a pathway toward fully functional optical accelerators for next-generation computing 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

2 major / 2 minor

Summary. The manuscript claims that optical random Fourier feature mapping implemented via linear silicon photonic circuits enables accurate approximation of high-order and high-dimensional nonlinear functions (tenth-order Legendre polynomials, Voigt/Fermi-Dirac/Fresnel functions, neural activations, 2D nonlinear maps, and 10D softmax) without requiring nonlinear optical materials, supported by experimental demonstrations in a simple silicon photonic circuit.

Significance. If the experimental results prove robust to photonic imperfections, the work would be significant for photonic computing by providing a scalable route to general-purpose nonlinear operations in linear hardware, potentially improving versatility and energy efficiency of optical AI accelerators.

major comments (2)
  1. [Experimental demonstration] Experimental demonstration (likely §4 or equivalent): the manuscript supplies no quantitative propagation analysis showing how fabrication-induced phase/amplitude errors (typically 1–5 %), waveguide loss, and thermal crosstalk perturb the random weights w and biases b of the optical RFF map and thereby affect final approximation error for the claimed tenth-order and 10-D functions. RFF convergence holds only for ideal mappings; this omission is load-bearing for the central experimental claim.
  2. [Results for high-dimensional cases] Results for high-dimensional cases (10-D softmax and 2-D maps): no comparison is provided between achieved approximation error and the theoretical RFF error bounds as a function of the number of random features, leaving unclear whether the optical implementation meets the precision asserted under realistic noise.
minor comments (2)
  1. [Abstract] Abstract: quantitative metrics (e.g., mean-squared error, number of features, throughput) are absent, making it difficult to assess the strength of the demonstrations relative to baselines.
  2. [Theory/Methods] Notation: the mapping from optical circuit parameters to the RFF frequencies w and biases b should be stated explicitly with a diagram or equation to clarify how the linear photonic mesh realizes the feature map.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the presentation of our experimental results. We address each major comment below and will revise the manuscript to incorporate the requested analyses.

read point-by-point responses

  1. Referee: [Experimental demonstration] Experimental demonstration (likely §4 or equivalent): the manuscript supplies no quantitative propagation analysis showing how fabrication-induced phase/amplitude errors (typically 1–5 %), waveguide loss, and thermal crosstalk perturb the random weights w and biases b of the optical RFF map and thereby affect final approximation error for the claimed tenth-order and 10-D functions. RFF convergence holds only for ideal mappings; this omission is load-bearing for the central experimental claim.

    Authors: We agree that an explicit quantitative error-propagation analysis is necessary to substantiate robustness. In the revised manuscript we will add a dedicated subsection that propagates typical fabrication errors (phase/amplitude deviations of 1–5 %, waveguide loss, and thermal crosstalk) through the optical RFF mapping via Monte-Carlo simulation and, where possible, direct measurement on the same device. The resulting impact on approximation error for the tenth-order polynomials and 10-D softmax will be reported explicitly. revision: yes

  2. Referee: [Results for high-dimensional cases] Results for high-dimensional cases (10-D softmax and 2-D maps): no comparison is provided between achieved approximation error and the theoretical RFF error bounds as a function of the number of random features, leaving unclear whether the optical implementation meets the precision asserted under realistic noise.

    Authors: We will add, for both the 2-D nonlinear maps and the 10-D softmax, plots that compare the experimentally obtained approximation error against the theoretical RFF convergence bound (O(1/√D)) as a function of the number of random features D. This comparison will be performed both for the ideal mapping and under the measured noise levels of the fabricated circuit, thereby demonstrating that the optical implementation achieves the expected scaling. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central claim rests on external RFF theory plus experimental implementation

full rationale

The paper's derivation applies the established random Fourier feature (RFF) mapping—imported from machine-learning literature—to linear photonic hardware for nonlinear approximation. The abstract and described approach cite no self-referential definitions, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the result to its own inputs. Experimental demonstrations of Legendre polynomials, special functions, and high-dimensional maps are presented as validation rather than tautological outputs. The chain is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities can be extracted. The mapping technique itself is treated as a standard mathematical tool imported from machine learning.

pith-pipeline@v0.9.1-grok · 5729 in / 961 out tokens · 27065 ms · 2026-06-26T06:10:16.967332+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

60 extracted references · 4 canonical work pages

  1. [1]

    Artificial neural networks for photonic applications —from algorithms to implementation: tutorial,

    P. Freire et al. , “Artificial neural networks for photonic applications —from algorithms to implementation: tutorial,” Adv. Opt. Photonics 15, 3, 739-834 (2023)

    2023

  2. [2]

    The physics of optical computing,

    P. L. McMahon, “The physics of optical computing,” Nat. Rev. Phys. 5, 717 (2023)

    2023

  3. [3]

    Advances in photonic reservoir computing,

    G. Van der Sande, D. Brunner, and M. C. Soriano, “Advances in photonic reservoir computing,” Nanophotonics 6, 3, 561–576 (2017)

    2017

  4. [4]

    Optical generative models,

    S. Chen et al., “Optical generative models,” Nature 644, 903–911 (2025)

    2025

  5. [5]

    Roadmap on neuromorphic photonics,

    D. Brunner et al., “Roadmap on neuromorphic photonics,” arXiv:2501.07917 (2025)

    work page arXiv 2025

  6. [6]

    Experimental realization of any discrete unitary operator,

    M. Reck et al., “Experimental realization of any discrete unitary operator,” Phys. Rev. Lett. 73, 1, 58-61 (1994)

    1994

  7. [7]

    Optimal design for universal multiport interferometers,

    W. R. Clements et al., “Optimal design for universal multiport interferometers,” Optica 3, 12, 1460–1465 (2016)

    2016

  8. [8]

    Self-configuring universal linear optical component,

    D. A. B. Miller, “Self-configuring universal linear optical component,” Photonics Res. 1, 1, 1–15 (2013)

    2013

  9. [9]

    All-optical machine learning using diffractive deep neural networks,

    X. Lin et al., “All-optical machine learning using diffractive deep neural networks,” Science 361, 6406, 1004– 1008 (2018)

    2018

  10. [10]

    Photonics for artificial intelligence and neuromorphic computing,

    B. J. Shastri et al. , “Photonics for artificial intelligence and neuromorphic computing,” Nat. Photonics 15, 2, 102–114 (2021)

    2021

  11. [11]

    Deep learning with coherent nanophotonic circuits,

    Y . Shen et al., “Deep learning with coherent nanophotonic circuits,” Nat. Photonics 11, 7, 441–446 (2017)

    2017

  12. [12]

    Parallel convolutional processing using an integrated photonic tensor core,

    J. Feldmann et al., “Parallel convolutional processing using an integrated photonic tensor core,” Nature 589, 7840, 52–58 (2021)

    2021

  13. [13]

    11 TOPS photonic convolutional accelerator for optical neural networks,

    X. Xu et al., “11 TOPS photonic convolutional accelerator for optical neural networks,” Nature 589, 7840, 44– 51 (2021)

    2021

  14. [14]

    Large- scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit,

    T. Zhou et al. , “Large- scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit,” Nat. Photonics 15, 5, 367–373 (2021)

    2021

  15. [15]

    Higher-dimensional processing using a photonic tensor core with continuous-time data,

    B. Dong et al., “Higher-dimensional processing using a photonic tensor core with continuous-time data,” Nat. Photonics 17, 12, 1080–1088 (2023)

    2023

  16. [16]

    A system-on-chip microwave photonic processor solves dynamic RF interference in real time with picosecond latency,

    W. Zhang et al., “A system-on-chip microwave photonic processor solves dynamic RF interference in real time with picosecond latency,” Light Sci. Appl. 13, 14 (2024)

    2024

  17. [17]

    Silicon photonics for high- speed communications and photonic signal processing,

    X. Zhou et al. , “Silicon photonics for high- speed communications and photonic signal processing,” npj Nanophotonics 1, 27 (2024)

    2024

  18. [18]

    Versatile parallel signal processing with a scalable silicon photonic chip,

    S. Hong et al., “Versatile parallel signal processing with a scalable silicon photonic chip,” Nat. Commun. 16, 288 (2025)

    2025

  19. [19]

    Optical computing and optical signal processing for recovery of spatiotemporally coupled optical communications channels,

    J. Liu et al., “Optical computing and optical signal processing for recovery of spatiotemporally coupled optical communications channels,” Adv. Photonics 7, 6, 064003 (2025)

    2025

  20. [20]

    Quantization-aware photonic homodyne computing for accelerated artificial intelligence and scientific simulation,

    L. Zhou et al., “Quantization-aware photonic homodyne computing for accelerated artificial intelligence and scientific simulation,” arXiv:2602.08269 (2026)

    work page arXiv 2026

  21. [21]

    Optical neural engine for solving scientific partial differential equations,

    Y . Tang et al., “Optical neural engine for solving scientific partial differential equations,” Nat. Commun. 16, 4603 (2025). 25

    2025

  22. [22]

    Reconfigurable application -specific photonic integrated circuit for solving partial differential equations,

    J. Ye et al., “Reconfigurable application -specific photonic integrated circuit for solving partial differential equations,” Nanophotonics 13, 12, 2231–2239 (2024)

    2024

  23. [23]

    Physics-informed neural networks for PDE problems: A comprehensive review,

    K. Luo et al., “Physics-informed neural networks for PDE problems: A comprehensive review,” Artif. Intell. Rev. 58, 10, 323 (2025)

    2025

  24. [24]

    Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in physical systems,

    K. Kontolati et al., “Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in physical systems,” Nat. Commun. 15, 5101 (2024)

    2024

  25. [25]

    Photonic tensor cores for machine learning,

    M. Miscuglio et al., “Photonic tensor cores for machine learning,” Appl. Phys. Rev. 7, 3, 031404 (2020)

    2020

  26. [26]

    Review of nonlinear activation functions in optical neural networks,

    W. Shi et al., “Review of nonlinear activation functions in optical neural networks,” Adv. Photonics 7, 6, 064004 (2025)

    2025

  27. [27]

    Ultra-broadband all-optical nonlinear activation function enabled by MoTe₂/optical waveguide integrated devices,

    C. Chen et al., “Ultra-broadband all-optical nonlinear activation function enabled by MoTe₂/optical waveguide integrated devices,” Nat. Commun. 15, 9047 (2024)

    2024

  28. [28]

    A complete photonic integrated neuron for nonlinear all-optical computing,

    T. Yan et al., “A complete photonic integrated neuron for nonlinear all-optical computing,” Nat. Comput. Sci. 5, 1202–1213 (2025)

    2025

  29. [29]

    Field-programmable photonic nonlinearity,

    T. Wu et al., “Field-programmable photonic nonlinearity,” Nat. Photonics 19, 725–732 (2025)

    2025

  30. [30]

    Programmable on-chip nonlinear photonics

    R. Yanagimoto et al., “Programmable on-chip nonlinear photonics” Nature 649, 330–337 (2026)

    2026

  31. [31]

    All-optical neural network with nonlinear activation functions,

    Y . Zuo et al., “All-optical neural network with nonlinear activation functions,” Optica 6, 9, 1132–1137 (2019)

    2019

  32. [32]

    Image sensing with multilayer nonlinear optical neural networks,

    T. Wang et al., “Image sensing with multilayer nonlinear optical neural networks,” Nat. Photonics 17, 5, 408– 415 (2023)

    2023

  33. [33]

    All-optical ultrafast ReLU function for energy -efficient nanophotonic deep learning,

    G. H. Y . Li et al. , “All-optical ultrafast ReLU function for energy -efficient nanophotonic deep learning,” Nanophotonics 12, 5, 847–855 (2023)

    2023

  34. [34]

    Large-scale photonic computing with nonlinear disordered media,

    H. Wang et al., “Large-scale photonic computing with nonlinear disordered media,” Nat. Comput. Sci. 4, 429– 439 (2024)

    2024

  35. [35]

    Programmable photonic circuits,

    W. Bogaerts et al., “Programmable photonic circuits,” Nature 586, 207–216 (2020)

    2020

  36. [36]

    An on-chip photonic deep neural network for image classification,

    F. Ashtiani, A. J. Geers, and F. Aflatouni, “An on-chip photonic deep neural network for image classification,” Nature 606, 501–506 (2022)

    2022

  37. [37]

    Single -chip photonic deep neural network with forward- only training,

    S. Bandyopadhyay et al. , “Single -chip photonic deep neural network with forward- only training,” Nat. Photonics 18, 1335–1343 (2024)

    2024

  38. [38]

    Optoelectronic nonlinear Softmax operator based on diffractive neural networks,

    Z. Zhan et al., “Optoelectronic nonlinear Softmax operator based on diffractive neural networks,” Opt. Express 32, 15, 26458–26469 (2024)

    2024

  39. [39]

    Nonlinear optical encoding enabled by recurrent linear scattering,

    F. Xia et al., “Nonlinear optical encoding enabled by recurrent linear scattering,” Nat. Photonics 18, 1067– 1075 (2024)

    2024

  40. [40]

    Nonlinear processing with linear optics,

    M. Yildirim et al., “Nonlinear processing with linear optics,” Nat. Photonics 18, 1076–1082 (2024)

    2024

  41. [41]

    Nonlinear encoding in diffractive information processing using linear optical materials,

    Y . Li, J. Li, and A. Ozcan, “Nonlinear encoding in diffractive information processing using linear optical materials,” Light Sci. Appl. 13, 173 (2024)

    2024

  42. [42]

    Unwrapping photonic reservoirs: Enhanced expressivity via random Fourier encoding over stretched domains,

    G. McCaul et al., “Unwrapping photonic reservoirs: Enhanced expressivity via random Fourier encoding over stretched domains,” Chaos 35, 9, 093129 (2025)

    2025

  43. [43]

    Scalable optical learning operator,

    U. Teğin et al., “Scalable optical learning operator,” Nat. Comput. Sci. 1, 542-549 (2021)

    2021

  44. [44]

    Random features for large-scale kernel machines,

    A. Rahimi et. al., “Random features for large-scale kernel machines,” Adv. Neural Inf. Process. Syst. 20, (2007). 26

    2007

  45. [45]

    Random feature attention,

    P. Hao et al., “Random feature attention,” arXiv:2103.02143 (2021)

    work page arXiv 2021

  46. [46]

    Fourier features let networks learn high frequency functions in low dimensional domains,

    M. Tancik et al., “Fourier features let networks learn high frequency functions in low dimensional domains,” arXiv:2006.10739 (2020)

    work page arXiv 2006

  47. [47]

    Algorithm 916: Computing the Faddeyeva and V oigt functions,

    M. R. Zaghloul et al. , “Algorithm 916: Computing the Faddeyeva and V oigt functions,” ACM Trans. Math. Softw. 38, 2, 15 (2012)

    2012

  48. [48]

    The V oigt and complex error function: Humlíček’s rational approximation generalized,

    F. Schreier, “The V oigt and complex error function: Humlíček’s rational approximation generalized,” Mon. Not. R. Astron. Soc. 479, 3, 3068–3075 (2018)

    2018

  49. [49]

    Optimized computation of the V oigt and complex probability functions,

    J. Humlíček, “Optimized computation of the V oigt and complex probability functions,” J. Quant. Spectrosc. Radiat. Transfer 27, 4, 437–444 (1982)

    1982

  50. [50]

    Minimax rational approximation of the Fermi–Dirac distribution,

    J. E. Moussa, “Minimax rational approximation of the Fermi–Dirac distribution,” J. Chem. Phys. 145, 164108 (2016)

    2016

  51. [51]

    J. W. Goodman, Introduction to Fourier Optics 3rd edn. Roberts & Company, Englewood (2005)

    2005

  52. [52]

    Low-power thermo-optic silicon modulator for large-scale photonic integrated systems,

    S. Chung et al., “Low-power thermo-optic silicon modulator for large-scale photonic integrated systems,” Opt. Express 27, 9, 13430–13459 (2019)

    2019

  53. [53]

    Thermo-optic phase shifters based on silicon-on-insulator platform: State-of-the-art and a review,

    S. Liu et al., “Thermo-optic phase shifters based on silicon-on-insulator platform: State-of-the-art and a review,” Front. Optoelectron. 15, 1, 9 (2022)

    2022

  54. [54]

    Silicon photonic microelectromechanical phase shifters for scalable programmable photonics,

    P. Edinger et al. , “Silicon photonic microelectromechanical phase shifters for scalable programmable photonics,” Opt. Lett. 46, 22, 5671-5674 (2021)

    2021

  55. [55]

    Integrated silicon photonic MEMS,

    N. Quack et al., “Integrated silicon photonic MEMS,” Microsyst. Nanoeng. 9, 27 (2023)

    2023

  56. [56]

    High-speed optical modulation based on carrier depletion in a silicon waveguide,

    A. Liu et al., “High-speed optical modulation based on carrier depletion in a silicon waveguide,” Opt. Express 15, 2, 660-668 (2007)

    2007

  57. [57]

    High-speed carrier-depletion silicon Mach–Zehnder optical modulators with lateral PN junctions,

    G. T. Reed et al., “High-speed carrier-depletion silicon Mach–Zehnder optical modulators with lateral PN junctions,” Front. Phys. 2, 77 (2014)

    2014

  58. [58]

    High-performance coherent optical modulators based on thin- film lithium niobate platform,

    M. Xu et al., “High-performance coherent optical modulators based on thin- film lithium niobate platform,” Nat. Commun. 11, 3911 (2020)

    2020

  59. [59]

    Attojoule/bit folded thin film lithium niobate coherent modulators using air-bridge structures,

    M. Xu et al., “Attojoule/bit folded thin film lithium niobate coherent modulators using air-bridge structures,” APL Photonics 8, 6, 066104 (2023)

    2023

  60. [60]

    Waveguide -multiplexed photonic matrix –vector multiplication processor using multiport photodetectors,

    R.Tang et al., “Waveguide -multiplexed photonic matrix –vector multiplication processor using multiport photodetectors,” Optica 12, 6, 812-820 (2025)

    2025