Pith. sign in

REVIEW 2 major objections 2 minor 23 references

Free-space 2f and 4f optical architectures lose many orders of magnitude less signal than universal multiport interferometers once matrix dimension exceeds one thousand elements.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-26 06:44 UTC pith:LQ3266SF

load-bearing objection The simulations claim 2f/4f free-space MVM loses far less signal than UMIs above N=1000 under a max-gain constraint, but the UMI baseline may not be normalized identically. the 2 major comments →

arxiv 2606.23988 v1 pith:LQ3266SF submitted 2026-06-22 physics.optics

Attenuation scaling and error analysis of 2f and 4f architectures for free-space optical matrix-vector multiplication

Dawson Lyles , Spencer LaVere Smith This is my paper
classification physics.optics
keywords free-space opticsmatrix-vector multiplicationoptical computingattenuation scaling2f architecture4f architectureuniversal multiport interferometer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper models two free-space optical setups called the 2f and 4f architectures for matrix-vector multiplication using wave optics simulations. After limiting the optical modulator to a maximum gain, the simulations track signal attenuation and computational error as matrix size increases, and they compare the results to the expected behavior of universal multiport interferometers used in integrated photonics. The 2f and 4f approaches show far slower growth in attenuation with problem size, remaining many orders of magnitude better than the interferometer baseline above roughly one thousand matrix elements. The study also varies modulator space-bandwidth product and output slit size to see how error and loss change across different statistical distributions of matrix values. Which of the two free-space setups performs better turns out to depend on those distributions, though the 4f version appears more adaptable overall.

Core claim

After constraining the optical modulator to a maximum gain limit, wave optics simulations of the 2f and 4f free-space architectures show many orders of magnitude less attenuation per matrix-vector multiplication than universal multiport interferometers for matrix dimensions above one thousand elements, with both error and attenuation also depending on modulator space-bandwidth product, output slit aperture, and the statistical distribution of matrix elements.

What carries the argument

Wave optics simulation of the 2f and 4f free-space optical matrix-vector multiplication architectures under a maximum modulator gain constraint.

Load-bearing premise

The models assume that constraining the optical modulator to a maximum gain limit is the dominant practical constraint and that the chosen statistical distributions of matrix elements are representative of real workloads.

What would settle it

A physical measurement of optical power loss through a 2f or 4f setup performing a 2000-element matrix-vector multiplication, compared against both the simulated attenuation and the universal multiport interferometer prediction, would test the scaling result.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Attenuation per matrix-vector multiplication grows much more slowly with problem size in the 2f and 4f architectures than in universal multiport interferometers.
  • Both computational error and attenuation change when modulator space-bandwidth product or output slit aperture is varied.
  • Which architecture, 2f or 4f, produces lower error and attenuation depends on the statistical distribution of the matrix elements.
  • The 4f architecture provides more flexibility than the 2f architecture across different matrix statistics.

Where Pith is reading between the lines

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

  • Free-space optical matrix-vector multiplication could support larger problem sizes than integrated-photonic approaches before signal loss becomes prohibitive.
  • Prototype experiments could usefully target matrix distributions typical of machine-learning workloads to decide between 2f and 4f designs.
  • The gain-limit assumption highlights the value of developing modulators with higher dynamic range to further improve scaling.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript presents wave-optics simulations of 2f and 4f free-space optical architectures for matrix-vector multiplication (MVM). After imposing a maximum gain limit on the optical modulator, the authors compare per-MVM attenuation and computational error scaling with matrix dimension N across different matrix-element distributions, reporting that both architectures experience many orders of magnitude less attenuation than universal multiport interferometers (UMIs) for N > 1000. They further examine the effects of modulator space-bandwidth product and output slit aperture, concluding that architecture preference depends on matrix statistics while 4f offers more flexibility.

Significance. If the reported scaling advantage is shown to rest on equivalent normalization and modeling choices, the result would indicate that free-space 2f/4f architectures can mitigate the attenuation penalty that limits integrated-photonic MVM implementations at large N, with potential implications for energy-efficient optical computing hardware.

major comments (2)
  1. [Abstract and UMI comparison] Abstract (and any UMI-comparison section): the central claim of 'many orders of magnitude less attenuation' for N > 1000 is load-bearing on the assumption that UMI attenuation is computed under the identical max-gain constraint and matrix-element statistics applied to the 2f/4f wave-optics models. The abstract's phrasing ('expected attenuation from a UMI') leaves open the possibility that the UMI figure is taken from the literature under a different normalization, which would make the reported gap an artifact of inconsistent power scaling rather than an architectural result.
  2. [Modeling and results sections] Modeling and results sections: the abstract states that simulations are performed after 'constraining the optical modulator... to have a maximum gain limit' and for 'different statistical distributions of matrix elements,' yet supplies no error bars, no sensitivity analysis on the chosen gain threshold, and no justification that the selected distributions are representative. Because these modeling choices directly determine the reported attenuation curves, the scaling comparison cannot be assessed without explicit documentation of how the limit and distributions were fixed.
minor comments (2)
  1. Define 'attenuation per MVM' explicitly (including any normalization by input power or by the number of matrix elements) so that the 2f/4f and UMI quantities can be compared on the same footing.
  2. Clarify whether the wave-optics simulations include realistic noise sources (e.g., shot noise, modulator crosstalk) or remain purely deterministic; this affects the interpretation of the computational-error results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and will revise the manuscript accordingly to improve clarity and documentation.

read point-by-point responses

  1. Referee: [Abstract and UMI comparison] Abstract (and any UMI-comparison section): the central claim of 'many orders of magnitude less attenuation' for N > 1000 is load-bearing on the assumption that UMI attenuation is computed under the identical max-gain constraint and matrix-element statistics applied to the 2f/4f wave-optics models. The abstract's phrasing ('expected attenuation from a UMI') leaves open the possibility that the UMI figure is taken from the literature under a different normalization, which would make the reported gap an artifact of inconsistent power scaling rather than an architectural result.

    Authors: The UMI attenuation is computed under the identical maximum-gain constraint and matrix-element statistics as the 2f/4f models, using the standard loss scaling for UMIs. The term 'expected' refers to this consistent application rather than an external literature value with different normalization. We will revise the abstract and add an explicit statement in the modeling section to remove any ambiguity about the shared normalization. revision: yes

  2. Referee: [Modeling and results sections] Modeling and results sections: the abstract states that simulations are performed after 'constraining the optical modulator... to have a maximum gain limit' and for 'different statistical distributions of matrix elements,' yet supplies no error bars, no sensitivity analysis on the chosen gain threshold, and no justification that the selected distributions are representative. Because these modeling choices directly determine the reported attenuation curves, the scaling comparison cannot be assessed without explicit documentation of how the limit and distributions were fixed.

    Authors: We agree that error bars, sensitivity analysis, and distribution justification are needed for full assessment. In the revised manuscript we will add error bars from multiple randomized-phase simulation runs, include a sensitivity analysis on the gain threshold, and justify the chosen distributions by their relevance to typical optical computing and machine-learning workloads. revision: yes

Circularity Check

0 steps flagged

No circularity; scaling claims rest on independent simulations and external UMI literature

full rationale

The paper models 2f/4f MVM via wave-optics simulations under an explicit max-gain modulator constraint, then compares resulting attenuation to the expected UMI attenuation drawn from the integrated-photonics literature. No equations, fitted parameters, or self-citations are shown that would make the reported attenuation numbers or scaling advantage reduce to quantities defined by the authors' own prior work or by construction. The UMI baseline is treated as an external reference rather than re-derived inside the paper, satisfying the independence criteria. The central claim therefore remains externally grounded.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated beyond the modulator gain constraint and the choice of matrix-element distributions.

pith-pipeline@v0.9.1-grok · 5775 in / 1209 out tokens · 15751 ms · 2026-06-26T06:44:59.660945+00:00 · methodology

0 comments
read the original abstract

Free-space optical computing has been suggested as a scalable, high speed, and energy efficient platform for performing matrix-vector multiplication (MVM). We present two free-space optical approaches for MVM, called the 2f and 4f architectures, and model them using wave optics simulations. After constraining the optical modulator in our models to have a maximum gain limit, we use our simulations to compare 2f and 4f MVM performance in terms of computational error and optical signal attenuation per MVM. We examine how 2f and 4f signal attenuation per MVM scales with increasing MVM problem size for different statistical distributions of matrix elements and compare to the expected attenuation from a universal multiport interferometer (UMI), commonly used in integrated photonics for MVM. We find that the 2f and 4f architectures scale more favorably to large problem sizes, experiencing many orders of magnitude less attenuation than UMIs for matrix dimension above a thousand elements. We furthermore examine how varying modulator space-bandwidth product and output slit aperture affect 2f and 4f attenuation and computational error across different distributions of matrix elements. We conclude that the preference of 2f or 4f MVM depends on the statistics of the matrix used, but that 4f may provide more flexibility than 2f.

Figures

Figures reproduced from arXiv: 2606.23988 by Dawson Lyles, Spencer LaVere Smith.

Figure 1
Figure 1. Figure 1: (A) Schematic of a 2 𝑓 MVM system. An input optical scalar field 𝑈in in the input plane has magnitude and phase profiles which together encode an input vector, 𝒗. Each column of the input field 𝑈in encodes an element of the input vector, shown as stripes of varying shades of green, creating a copy of 𝒗 for each row of a matrix 𝑨. The field encounters a modulator which applies the transmittance function 𝑡2 … view at source ↗
Figure 2
Figure 2. Figure 2: Top: Diagram of 2 𝑓 architecture performing vector dot product 𝐷 = 𝒂 · 𝒗 showing input field 𝑈in, modulator transmittance 𝑡2 𝑓 , and output field 𝑈out and relevant planes. (A) Input optical field 𝑈in immediately before the modulator encoding vector 𝒗 = (1, 2, 3, 4) 𝑇 . The comb-like appearance of the plot trace matches that of the modulator transmittance. (B) Modulator transmittance function 𝑡2 𝑓 residing … view at source ↗
Figure 3
Figure 3. Figure 3: Top: Diagram of 4 𝑓 architecture performing vector dot product 𝐷 = 𝒂 · 𝒗 showing input field 𝑈in, modulator transmittance 𝑡4 𝑓 , and output field 𝑈out and relevant planes. (A) Input optical field 𝑈in encoding vector 𝒗 = (1, 2, 3, 4) 𝑇 . (B) Normalized convolution kernel 𝐴eff/𝐹, where 𝐹 is the modulator pixel fill factor, derived from the modulator transmittance of (C) and (D). The numerically simulated ker… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of a free-space modulator capable of independently manipulating [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Top: Histograms show the three different distributions of the vector elements [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Top: Plots of 2 𝑓 and 4 𝑓 percent waveform error WE vs. magnitude of dot product |𝐷| = |𝒂 · 𝒗| for varying modulator pixel count 𝑀 at vector length 𝑁 = 50 for each distribution of 𝑎𝑛. 2 𝑓 and 4 𝑓 data are plotted as cool-colored solid and hot-colored hollow circles, respectively. 2 𝑓 error data is taken at a slit bandwidth of Δ𝜈𝑥 = 0.2 cm−1 . Waveform error data are binned according to dot product magnitud… view at source ↗
Figure 7
Figure 7. Figure 7: Top: Plots of 2 𝑓 percent waveform error WE vs. magnitude of dot product |𝐷| = |𝒂 · 𝒗| for varying slit bandwidth Δ𝜈𝑥 at vector length 𝑁 = 50 for each distribution of 𝑎𝑛. 2 𝑓 error data is taken at 𝑀 = 2000 modulator pixels. Waveform error data are binned according to dot product magnitude and averaged for each bin to obtain the data points (with error bars). Some error bars are smaller than the marker siz… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

23 extracted references

  1. [1]

    Analog optical computing for artificial intelligence,

    J. Wu, X. Lin, Y. Guo,et al., “Analog optical computing for artificial intelligence,” Engineering10, 133–145 (2022)

    2022

  2. [2]

    Optical neural networks: progress and challenges,

    T. Fu, J. Zhang, R. Sun,et al., “Optical neural networks: progress and challenges,” Light. Sci. & Appl.13, 263 (2024)

    2024

  3. [3]

    The physics of optical computing,

    P. L. McMahon, “The physics of optical computing,” Nat. Rev. Phys.5, 717–734 (2023)

    2023

  4. [4]

    An optical neural network using less than 1 photon per multiplication,

    T. Wang, S.-Y. Ma, L. Wright,et al., “An optical neural network using less than 1 photon per multiplication,” Nat. Commun.13(2022)

    2022

  5. [5]

    Programmable photonic arrays based on microelectromechanical elements with femtowatt-level standby power consumption,

    D. U. Kim, Y. J. Park, D. Y. Kim,et al., “Programmable photonic arrays based on microelectromechanical elements with femtowatt-level standby power consumption,” Nat. Photonics17, 1089–1096 (2023)

    2023

  6. [6]

    Large-scaleopticalneuralnetworksbasedonphotoelectricmultiplication,

    R.Hamerly,L.Bernstein,A.Sludds,et al.,“Large-scaleopticalneuralnetworksbasedonphotoelectricmultiplication,” Phys. Rev. X9, 021032 (2019)

    2019

  7. [7]

    Matrix multiplication by optical methods,

    R. A. Heinz, J. O. Artman, and S. H. Lee, “Matrix multiplication by optical methods,” Appl. Opt.9, 2161–2168 (1970)

    1970

  8. [8]

    Fullyreconfigurablecoherentopticalvector–matrixmultiplication,

    J.Spall,X.Guo,T.D.Barrett,andA.I.Lvovsky,“Fullyreconfigurablecoherentopticalvector–matrixmultiplication,” Opt. Lett.45, 5752–5755 (2020)

    2020

  9. [9]

    All-optical computation of an arbitrary linear transform using diffractive networks,

    O. Kulce, D. Mengu, Y. Rivenson, and A. Ozcan, “All-optical computation of an arbitrary linear transform using diffractive networks,” inImaging and Applied Optics Congress 2022 (3D, AOA, COSI, ISA, pcAOP),(Optica Publishing Group, 2022), p. CTu3F.4

    2022

  10. [10]

    A small microring array that performs large complex-valued matrix-vector multiplication,

    J. Cheng, Y. Zhao, W. Zhang,et al., “A small microring array that performs large complex-valued matrix-vector multiplication,” Front. Optoelectron.15, 15 (2022)

    2022

  11. [11]

    Deep learning with coherent nanophotonic circuits,

    Y. Shen, N. C. Harris, S. Skirlo,et al., “Deep learning with coherent nanophotonic circuits,” Nat. Photonics11, 441–446 (2017)

    2017

  12. [12]

    Experimental realization of any discrete unitary operator,

    M. Reck, A. Zeilinger, H. J. Bernstein, and P. Bertani, “Experimental realization of any discrete unitary operator,” Phys. Rev. Lett.73, 58–61 (1994)

    1994

  13. [13]

    Optimal design for universal multiport interferometers,

    W. R. Clements, P. C. Humphreys, B. J. Metcalf,et al., “Optimal design for universal multiport interferometers,” Optica3, 1460–1465 (2016)

    2016

  14. [14]

    Matrix Multiplication Using Coherent Optical Techniques,

    P. N. Tamura and J. C. Wyant, “Matrix Multiplication Using Coherent Optical Techniques,” inOptical Information Processing: Real Time Devices and Novel Techniques,vol.0083D.P.CasasentandA.A.Sawchuk,eds.,International Society for Optics and Photonics (SPIE, 1977), pp. 97 – 104

    1977

  15. [15]

    Fully parallel, high-speed incoherent optical method for performing discrete fourier transforms,

    J. W. Goodman, A. R. Dias, and L. M. Woody, “Fully parallel, high-speed incoherent optical method for performing discrete fourier transforms,” Opt. Lett.2, 1–3 (1978)

    1978

  16. [16]

    Incoherent optical implementation of 2-d complex discrete fourier transform and equivalent 4-f system,

    L. Zhang and L. Liu, “Incoherent optical implementation of 2-d complex discrete fourier transform and equivalent 4-f system,” Opt. Commun.74, 295–300 (1990)

    1990

  17. [17]

    All-optical neural network with nonlinear activation functions,

    Y. Zuo, B. Li, Y. Zhao,et al., “All-optical neural network with nonlinear activation functions,” Optica6, 1132–1137 (2019)

    2019

  18. [18]

    Hybrid optical-electronic convolutional neural networks with optimized diffractive optics for image classification,

    J. Chang, V. Sitzmann, X. Dun,et al., “Hybrid optical-electronic convolutional neural networks with optimized diffractive optics for image classification,” Sci. Reports8(2018)

    2018

  19. [19]

    Single-shot optical neural network,

    L. Bernstein, A. Sludds, C. Panuski,et al., “Single-shot optical neural network,” Sci. Adv.9, eadg7904 (2023)

    2023

  20. [20]

    Fourier-space diffractive deep neural network,

    T. Yan, J. Wu, T. Zhou,et al., “Fourier-space diffractive deep neural network,” Phys. Rev. Lett.123, 023901 (2019)

    2019

  21. [21]

    All linear optical devices are mode converters,

    D. A. B. Miller, “All linear optical devices are mode converters,” Opt. Express20, 23985–23993 (2012)

    2012

  22. [22]

    Arbitrarymanipulationofspatialamplitudeandphaseusingphase-onlyspatiallightmodulators,

    L.ZhuandJ.Wang,“Arbitrarymanipulationofspatialamplitudeandphaseusingphase-onlyspatiallightmodulators,” Sci. reports4, 7441 (2014)

    2014

  23. [23]

    J. W. Goodman,Introduction to Fourier Optics(McGraw-Hill, New York, 1996), 2nd ed

    1996