pith. sign in

arxiv: 2605.20864 · v1 · pith:YPTSUDY6new · submitted 2026-05-20 · ⚛️ physics.optics

Scalable native signed optical computing enabled by dual-wavelength incoherent multiplexing

Yuan Ren , Yong Zheng , Ruixue Liu , Yunpeng Song , Qinfen Huang , Min Wang , Ya Cheng This is my paper

Pith reviewed 2026-05-21 02:11 UTC · model grok-4.3

classification ⚛️ physics.optics
keywords incoherent photonic neural networksdual-wavelength multiplexingsigned optical computingthin-film lithium niobateoptical multiplicationmultiply-accumulate operationphotonic computing
0
0 comments X

The pith

Dual-wavelength encoding enables signed optical computing with size-independent hardware overhead.

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

The paper shows how to perform native signed operations in incoherent photonic neural networks by encoding complementary signal parts on two separate wavelengths. These components are then weighted and combined inside one shared physical path on a lithium niobate chip. This removes the need for duplicated weighting hardware or extra channels that would otherwise grow with matrix size. The overhead for handling positive and negative values therefore stays fixed per multiply-accumulate step. Experiments confirm low error and working neural-network classification on standard datasets.

Core claim

By encoding complementary signal components onto two wavelength channels and performing computation within a shared physical path, the proposed scheme eliminates duplicated weighting units. As a result, the additional hardware overhead associated with signed computation remains constant per multiply accumulate operation, independent of matrix size.

What carries the argument

Dual-wavelength incoherent multiplexing that encodes complementary components of signed inputs and weights on two distinct wavelengths for computation in one shared physical path.

If this is right

  • Four-quadrant optical multiplication is realized with 1.27 percent standard deviation error.
  • The device operates at modulation bandwidths above 40 GHz.
  • Neural-network classification reaches 95.1 percent accuracy on the Moons dataset and 91.63 percent on MNIST.
  • The approach supplies a direct route to larger-scale incoherent photonic systems that still handle bipolar values natively.

Where Pith is reading between the lines

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

  • Larger matrix sizes become feasible because the extra hardware cost per operation does not increase.
  • The same dual-wavelength trick could be combined with existing wavelength-division multiplexing to raise parallelism further.
  • Similar encoding might be tested on other incoherent platforms to check whether the constant-overhead property holds beyond lithium niobate.

Load-bearing premise

Complementary signals on two wavelengths can be weighted and added together in the same path without crosstalk, wavelength-dependent loss differences, or nonlinear mixing that would destroy the linearity needed for correct signed multiplication.

What would settle it

Measurement of error that grows with matrix size or of strong crosstalk between the two wavelength channels in the fabricated device.

read the original abstract

Incoherent photonic neural networks (PNNs) provide a robust platform for analog optical computing, yet efficient implementation of native signed operations remains challenging. Existing incoherent PNNs approaches often require additional spatial channels or temporal encoding steps to represent bipolar input signals, resulting in hardware overhead that scales with system size. Here, we demonstrate a dual-wavelength incoherent photonic architecture that natively supports both signed inputs and signed weights on a thin-film lithium niobate platform. By encoding complementary signal components onto two wavelength channels and performing computation within a shared physical path, the proposed scheme eliminates duplicated weighting units. As a result, the additional hardware overhead associated with signed computation remains constant per multiply accumulate operation, independent of matrix size. The fabricated device exhibits a modulation bandwidth exceeding 40 GHz and achieves four-quadrant optical multiplication with a standard deviation error of 1.27%. System-level functionality is validated through neural-network classification, achieving 95.1% accuracy on the Moons dataset and 91.63% on MNIST. These results establish a practical route toward scalable incoherent photonic computing systems with native bipolar processing capability.

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 presents a dual-wavelength incoherent photonic architecture on thin-film lithium niobate for native signed optical computing in photonic neural networks. Complementary signal components are encoded on two distinct wavelengths and processed in a shared physical path, eliminating the need for duplicated weighting units per sign. This results in the claim that signed-computation hardware overhead remains constant per multiply-accumulate operation and independent of matrix size. The fabricated device demonstrates >40 GHz modulation bandwidth and 1.27% standard deviation error in four-quadrant multiplication. System-level tests achieve 95.1% accuracy on the Moons dataset and 91.63% on MNIST.

Significance. If the results hold, this provides a concrete experimental route to scalable incoherent PNNs with native bipolar capability, addressing a key limitation where signed operations previously incurred size-dependent overhead. The measured device bandwidth, single-MAC error metric, and end-to-end classification accuracies constitute a solid experimental demonstration. The architecture description directly supports constant-per-MAC overhead, which is a strength for larger-scale systems; explicit scaling verification would further strengthen the contribution.

major comments (2)
  1. [Architecture and scaling discussion] Architecture and scaling discussion (near the end of the device description): The central claim that overhead remains constant per MAC independent of matrix size is supported by the dual-wavelength shared-path description, but the manuscript lacks an explicit scaling calculation or simulation (e.g., hardware count for an N×N matrix as N increases from 4 to 64). Adding this would directly substantiate the independence assertion for the central claim.
  2. [Experimental validation section] Experimental validation section (results on four-quadrant multiplication): The 1.27% error is reported for the fabricated device, but to confirm negligible crosstalk and wavelength-dependent loss for the signed linearity assumption, additional measurements across a range of input amplitudes and wavelength separations would be needed; current data support the single-MAC case but leave the scaling robustness for larger matrices as an implicit extrapolation.
minor comments (2)
  1. [Figure captions] Figure captions for the device schematic and measurement setups could more explicitly label the two wavelength channels and the shared accumulation path to improve clarity for readers unfamiliar with the dual-wavelength approach.
  2. [System-level results] The MNIST accuracy of 91.63% is given without comparison to a baseline incoherent PNN without signed native support; adding this would help quantify the practical benefit of the proposed architecture.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and recommendation for minor revision. The comments have helped clarify the presentation of our scaling claims and experimental assumptions. We address each major comment below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [Architecture and scaling discussion] Architecture and scaling discussion (near the end of the device description): The central claim that overhead remains constant per MAC independent of matrix size is supported by the dual-wavelength shared-path description, but the manuscript lacks an explicit scaling calculation or simulation (e.g., hardware count for an N×N matrix as N increases from 4 to 64). Adding this would directly substantiate the independence assertion for the central claim.

    Authors: We agree that an explicit scaling calculation strengthens the central claim. In the revised manuscript we have added a dedicated paragraph with a table that enumerates component counts for N×N matrices with N = 4, 8, 16, 32, and 64. The analysis shows that the dual-wavelength shared-path approach requires exactly N² physical weighting units (each carrying two wavelength channels), whereas a duplicated spatial-channel implementation would require 2N² units. Consequently the signed-computation overhead remains fixed at one additional wavelength per MAC and does not grow with matrix dimension. revision: yes

  2. Referee: [Experimental validation section] Experimental validation section (results on four-quadrant multiplication): The 1.27% error is reported for the fabricated device, but to confirm negligible crosstalk and wavelength-dependent loss for the signed linearity assumption, additional measurements across a range of input amplitudes and wavelength separations would be needed; current data support the single-MAC case but leave the scaling robustness for larger matrices as an implicit extrapolation.

    Authors: We acknowledge that measurements over a wider parameter space would provide additional reassurance. The present device characterization already includes the 1.27 % error under the exact wavelength separation and drive conditions used for the system-level tests; the >40 GHz bandwidth further indicates that crosstalk remains low within the operating regime. The reported 95.1 % and 91.63 % accuracies on Moons and MNIST involve multiple simultaneous MAC operations and therefore supply indirect evidence of linearity at modest scale. In the revision we have expanded the experimental section with a short discussion of the measured crosstalk and loss figures and have explicitly stated the extrapolation assumptions. Full multi-matrix hardware scaling lies outside the scope of this demonstration. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an experimental demonstration of a dual-wavelength incoherent photonic architecture on a fabricated thin-film lithium niobate chip. The core scalability claim—that signed-computation hardware overhead remains constant per MAC and independent of matrix size—is an architectural description arising from encoding complementary signals on two fixed wavelengths into a shared physical path, which eliminates the need for duplicated weighting units. This is not derived from equations or parameters that reduce to the claim by construction. All quantitative results (modulation bandwidth >40 GHz, 1.27% standard deviation error in four-quadrant multiplication, 95.1% Moons accuracy, 91.63% MNIST accuracy) are reported as direct experimental measurements on the device rather than self-referential predictions or fitted inputs renamed as outputs. No load-bearing steps rely on self-citation chains, uniqueness theorems imported from prior author work, or ansatzes smuggled via citation; the derivation chain is self-contained against external benchmarks of device performance and system-level classification tasks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The architecture rests on standard linear optics and electro-optic modulation assumptions plus the practical ability to separate and recombine two wavelength channels with low crosstalk. No new particles or forces are postulated; the main addition is the encoding scheme itself.

free parameters (1)
  • wavelength pair separation
    Specific wavelengths chosen to minimize crosstalk while fitting modulator response; value not stated in abstract but required for the complementary encoding to function.
axioms (1)
  • domain assumption Electro-optic modulation remains linear over the operating range for both wavelengths
    Required for the signed multiplication to be accurate; invoked implicitly by the reported 1.27% error metric.

pith-pipeline@v0.9.0 · 5737 in / 1333 out tokens · 29156 ms · 2026-05-21T02:11:27.635136+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    By encoding complementary signal components onto two wavelength channels and performing computation within a shared physical path, the proposed scheme eliminates duplicated weighting units. As a result, the additional hardware overhead associated with signed computation remains constant per multiply-accumulate operation, independent of matrix size.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    the fabricated device exhibits a modulation bandwidth exceeding 40 GHz and achieves four-quadrant optical multiplication with a standard deviation error of 1.27%

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    & Hinton, G

    LeCun, Y ., Bengio, Y . & Hinton, G. Deep learning. Nature 521, 436-444 (2015)

    work page 2015

  2. [2]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, 770- 778 (2016)

    work page 2016

  3. [3]

    Kudithipudi, D. et al. Neuromorphic computing at scale. Nature 637, 801 -812 (2025)

    work page 2025

  4. [4]

    & Kenyon, A

    Mehonic, A. & Kenyon, A. J. Brain -inspired computing needs a master plan. Nature 604, 255-260 (2022)

    work page 2022

  5. [5]

    Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems 30 (2017)

    work page 2017

  6. [6]

    & Hinton, G

    Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems 25 (2012)

    work page 2012

  7. [7]

    Cai, H. et al. Brain organoid reservoir computing for artificial intelligence. Nature Electronics 6, 1032-1039 (2023)

    work page 2023

  8. [8]

    Aguirre, F. et al. Hardware implementation of memristor -based artificial neural networks. Nature communications 15, 1974 (2024)

    work page 1974

  9. [9]

    Shen, Y . et al. Deep learning with coherent nanophotonic circuits. Nature Photonics 11, 441-446, (2017)

    work page 2017

  10. [10]

    Ahmed, S. R. et al. Universal photonic artificial intelligence acceleration. Nature 640, 368-374, (2025)

    work page 2025

  11. [11]

    Hua, S. et al. An integrated large-scale photonic accelerator with ultralow latency. Nature 640, 361-367, (2025)

    work page 2025

  12. [12]

    Reck, M., Zeilinger, A., Bernstein, H. J. & Bertani, P. Experimental realization of any discrete unitary operator. Physical Review Letters 73, 58-61, (1994)

    work page 1994

  13. [13]

    Carolan, J. et al. Universal linear optics. Science 349, 711-716 (2015)

    work page 2015

  14. [14]

    R., Humphreys, P

    Clements, W. R., Humphreys, P. C., Metcalf, B. J., Kolthammer, W. S. & Walsmley, I. A. Optimal design for universal multiport interferometers. Optica 3, 1460 -1465, (2016)

    work page 2016

  15. [15]

    & Bogaerts, W

    Ribeiro, A., Ruocco, A., Vanacker, L. & Bogaerts, W. Demonstration of a 4 × 4 - port universal linear circuit. Optica 3, 1348-1357, (2016)

    work page 2016

  16. [16]

    Harris, N. C. et al. Linear programmable nanophotonic processors. Optica 5, 1623- 1631, (2018)

    work page 2018

  17. [17]

    & Liboiron-Ladouceur, O

    Shokraneh, F., Geoffroy-gagnon, S. & Liboiron-Ladouceur, O. The diamond mesh, a phase-error- and loss-tolerant field-programmable MZI-based optical processor for optical neural networks. Optics Express 28, 23495-23508, (2020)

    work page 2020

  18. [18]

    Zhang, H. et al. An optical neural chip for implementing complex -valued neural network. Nature Communications 12, 457, (2021)

    work page 2021

  19. [19]

    Tian, Y . et al. Scalable and compact photonic neural chip with low learning - capability-loss. Nanophotonics 11, 329-344, (2022)

    work page 2022

  20. [20]

    Pai, S. et al. Experimentally realized in situ backpropagation for deep learning in photonic neural networks. Science 380, 398-404 (2023)

    work page 2023

  21. [21]

    Zheng, Y . et al. Electro-optically programmable photonic circuits enabled by wafer-scale integration on thin -film lithium niobate. Physical Review Research 5, (2023)

    work page 2023

  22. [22]

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

    work page 2024

  23. [23]

    Lin, Z. et al. 120 GOPS Photonic tensor core in thin -film lithium niobate for inference and in situ training. Nature Communications 15, 9081, (2024)

    work page 2024

  24. [24]

    Zheng, Y . et al. Photonic Neural Network Fabricated on Thin Film Lithium Niobate for High‐Fidelity and Power‐Efficient Matrix Computation. Laser & Photonics Reviews 18, 2400565, (2024)

    work page 2024

  25. [25]

    Hu, Y . et al. Integrated lithium niobate photonic computing circuit based on efficient and high -speed electro-optic conversion. Nature Communications 16, 8178, (2025)

    work page 2025

  26. [26]

    & Bengio, Y

    Glorot, X., Bordes, A. & Bengio, Y . Deep sparse rectifier neural networks. Proceedings of the fourteenth international conference on artificial intelligence and statistics, 315-323 (2011)

    work page 2011

  27. [27]

    Mikolov, T., Sutskever, I., Chen, K., Corrado, G. S. & Dean, J. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems 26 (2013)

    work page 2013

  28. [28]

    Goodfellow, I. et al. Generative adversarial networks. Communications of the ACM 63, 139-144 (2020)

    work page 2020

  29. [29]

    Gu, J. et al. O2NN: Optical neural networks with differential detection -enabled optical operands. 2021 Design, Automation & Test in Europe Conference & Exhibition 1062-1067 (2021)

    work page 2021

  30. [30]

    Wu, R. et al. Long Low-Loss-Litium Niobate on Insulator Waveguides with Sub- Nanometer Surface Roughness. Nanomaterials 8 (2018)

    work page 2018

  31. [31]

    Wu, R. et al. Lithium niobate micro-disk resonators of quality factors above 10^7. Opt Lett 43, 4116-4119 (2018)

    work page 2018

  32. [32]

    Ren, Y . et al. Compact Ultra -Low Loss Optical True Delay Line on Thin Film Lithium Niobate. Chinese Physics Letters 42 (2025)

    work page 2025

  33. [33]

    Gao, R. et al. Lithium niobate microring with ultra -high Q factor above 10^8. Chinese Optics Letters 20, 011902 (2022)

    work page 2022

  34. [34]

    Ren, Y . et al. High-Q Lithium Niobate Microring Resonator with Electro-Optically Reconfigurable Coupling Strength. arXiv preprint arXiv:2512.22779 (2025)

    work page arXiv 2025

  35. [35]

    Jia, D. et al. Electrically tuned coupling of lithium niobate microresonators. Opt Lett 48, 2744-2747, doi:10.1364/OL.488974 (2023)

    work page doi:10.1364/ol.488974 2023

  36. [36]

    Wang, C. et al. Ultrahigh-efficiency wavelength conversion in nanophotonic periodically poled lithium niobate waveguides. Optica 5, 1438-1441 (2018)

    work page 2018

  37. [37]

    Zhao, G. et al. Low-loss thin-film periodically poled lithium niobate waveguides fabricated by femtosecond laser photolithography. Optics Letters 50, 4310-4313 (2025)

    work page 2025

  38. [38]

    Xie, X. et al. A 3.584 Tbps coherent receiver chip on InP -LiNbO3 wafer-level integration platform. Light: Science & Applications 14, 172 (2025)

    work page 2025