Breaking the Cascade: Compact Nonlinear Optical Computing with Single-Layer Encoder-Decoder Co-Localization

Alexander Chen; Aydogan Ozcan; Md Sadman Sakib Rahman; Yuntian Wang

arxiv: 2606.01032 · v1 · pith:XSH5F3TUnew · submitted 2026-05-31 · ⚛️ physics.optics · cs.NE· physics.app-ph

Breaking the Cascade: Compact Nonlinear Optical Computing with Single-Layer Encoder-Decoder Co-Localization

Yuntian Wang , Alexander Chen , Md Sadman Sakib Rahman , Aydogan Ozcan This is my paper

Pith reviewed 2026-06-28 16:53 UTC · model grok-4.3

classification ⚛️ physics.optics cs.NEphysics.app-ph

keywords nonlinear optical computingdiffractive opticsencoder-decoder co-localizationuniversal approximationphase-only diffractive planecoherent illuminationintensity detectionsingle-layer computation

0 comments

The pith

A single phase-only diffractive plane with co-localized input-dependent encoder and static decoder approximates arbitrary real-valued band-limited nonlinear functions through interference and intensity detection.

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

The paper establishes that nonlinear optical computing is possible with only one linear diffractive surface under coherent light. An input-dependent dynamic encoder shares the plane with a static decoder; free-space propagation and intensity detection then produce the desired nonlinear mappings via coherent interference without nonlinear materials or cascaded layers. The architecture is proven to serve as a universal approximator for any real-valued band-limited nonlinear function, with accuracy set by decoder degrees of freedom, detector aperture, and propagation distance. A trained frozen phase bias on the encoder region is shown to boost expressivity and tolerate coarse phase quantization on modulators. Experimental results confirm parallel approximation of nine distinct nonlinear functions in a single optical forward pass after in-situ training.

Core claim

The E+D optical processor places a dynamic encoder and a static decoder in the same phase-only diffractive plane. Coherent illumination, free-space propagation, and intensity detection together generate programmable nonlinear input-output mappings. This construction is a universal approximator for arbitrary real-valued band-limited nonlinear functions, with approximation fidelity governed by the decoder degrees of freedom, detector aperture, and axial propagation distance. A trained frozen phase bias added to the encoder region increases functional expressivity and robustness to phase quantization.

What carries the argument

Encoder-decoder co-localization (E+D) in one phase-only diffractive plane, where input-dependent encoder fields interfere with a static decoder during propagation and intensity detection produces the nonlinearity.

If this is right

The E+D processor approximates any real-valued band-limited nonlinear function.
Approximation fidelity is set by decoder degrees of freedom, detector aperture, and axial propagation distance.
A trained frozen phase bias on the encoder improves expressivity and tolerates coarse phase quantization.
Diverse nonlinear functions including neural-network activations and complex-valued mappings can be synthesized.
Nine nonlinear functions can be computed in parallel within one optical forward pass.

Where Pith is reading between the lines

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

The single-plane design may reduce the physical depth required for optical neural networks that rely on nonlinear activations.
In-situ training on spatial light modulators could allow the same hardware to be reprogrammed for different nonlinear tasks without retraining the decoder.
The approach might extend to other coherent wave systems where controllable interference replaces material nonlinearity.

Load-bearing premise

An input-dependent dynamic encoder can be physically realized in the same phase-only diffractive plane as the static decoder, so that free-space propagation and intensity detection alone suffice to produce arbitrary nonlinear mappings without nonlinear materials.

What would settle it

An optical experiment in which the single-layer E+D setup, after optimization of decoder degrees of freedom and propagation distance, fails to approximate a chosen band-limited nonlinear target function such as a sine wave to within the predicted error bound.

read the original abstract

We demonstrate that nonlinear computing can be achieved with a single linear diffractive surface under coherent illumination. We introduce a compact encoder-decoder co-localization (E+D) architecture in which an input-dependent dynamic encoder and a static optimized decoder are integrated within the same phase-only diffractive plane. Following free-space propagation, coherent interference between the encoder and decoder fields, combined with intensity detection, generates programmable nonlinear input-output mappings without requiring nonlinear optical materials or multiple diffractive layers. We prove that the proposed E+D optical processor is a universal approximator for arbitrary real-valued band-limited nonlinear functions and identify the physical factors governing its approximation fidelity, including the decoder degrees-of-freedom, detector aperture, and axial propagation distance. Crucially, we demonstrate that introducing a trained, frozen phase bias to the encoder region systematically enhances functional expressivity, providing robustness against coarse phase quantization on spatial light modulators. Using this framework, we accurately synthesize diverse nonlinear functions, including commonly used neural network activation functions and complex-valued nonlinear functions. Finally, we experimentally validate the proposed approach using a visible-light optical set-up trained through in situ learning, demonstrating the parallel approximation of 9 nonlinear functions in a single optical forward pass. By collapsing nonlinear optical computation into a single diffractive surface, the E+D architecture substantially reduces hardware and alignment complexity while preserving powerful function-approximation capabilities, providing a compact and scalable framework for analog information processing.

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.

Desk Editor's Note private letter to a colleague

The paper puts a dynamic encoder and static decoder on one phase-only diffractive plane to get nonlinearity from interference alone, with a universality claim for band-limited functions and an experiment on nine activations.

read the letter

The core claim is that nonlinear mappings can be realized with a single diffractive surface by co-locating an input-dependent encoder region with a static decoder, then using free-space propagation and intensity detection. This collapses what usually requires multiple layers or nonlinear materials into one plane, and they supply a proof that the setup is a universal approximator for real-valued band-limited functions. They also run an in-situ trained visible-light experiment that approximates nine different nonlinear functions in one forward pass.

The architecture itself is the clearest advance. Reducing alignment and hardware count while keeping programmable nonlinearity is a practical step for diffractive optical processors. Naming the governing factors (decoder degrees of freedom, detector aperture, axial distance) and adding a frozen phase bias to improve robustness against SLM quantization are useful details. The experimental parallelization of multiple functions on one surface is concrete evidence that the idea works at least for the tested cases.

The main open question is whether the input-dependent dynamic encoder can actually be realized on the same phase-only plane without hidden assumptions or extra stages. The stress-test note flags this correctly: the cross-term interference must generate the full target function class under finite aperture and propagation, and the abstract does not spell out the physical mechanism that makes the encoder input-dependent while the decoder stays static. Without the derivation steps or error analysis, it is hard to judge how tight the band-limited restriction is or whether co-localization introduces unstated limits on expressivity.

This is aimed at groups working on analog photonic computing and diffractive neural networks. Readers who care about compact hardware implementations will find the architecture and the experiment worth examining. The work is coherent enough on its own terms to merit a serious referee, even if the proof and physical constraints need close checking.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces an encoder-decoder co-localization (E+D) architecture on a single phase-only diffractive plane under coherent illumination. An input-dependent dynamic encoder and static decoder share the plane; free-space propagation followed by intensity detection produces nonlinear mappings without nonlinear materials or cascaded layers. The central claims are a mathematical proof that the E+D processor is a universal approximator for arbitrary real-valued band-limited nonlinear functions, identification of governing physical factors (decoder DoF, detector aperture, axial distance), enhancement via a trained frozen phase bias, synthesis of NN activations and complex nonlinear functions, and experimental validation via in-situ learning that approximates 9 functions in one optical pass.

Significance. If the universality proof holds under the stated physical constraints and the experimental controls are robust, the result would be significant for compact analog optical computing: it collapses nonlinear function approximation into one diffractive surface, reducing hardware complexity and alignment demands while retaining expressivity. The in-situ learning demonstration and explicit identification of fidelity-limiting physical parameters are concrete strengths that could guide practical implementations.

major comments (2)

[Abstract] The universality claim rests on the physical realizability of an input-dependent dynamic encoder co-localized with the static decoder on a single phase-only plane. The abstract states that interference plus intensity detection alone generates arbitrary band-limited nonlinear maps, but does not specify how the input field modulates only the encoder region to produce the required cross term 2 Re(E_enc E_dec*) without additional nonlinear materials or separate stages; this assumption appears load-bearing for the span of achievable functions.
[Abstract] The abstract identifies decoder degrees-of-freedom, detector aperture, and axial propagation distance as the factors governing approximation fidelity, yet provides no quantitative bounds or scaling relations showing that finite values of these parameters suffice to reach the full target function class under the co-localization constraint.

minor comments (2)

[Abstract] The abstract refers to 'commonly used neural network activation functions' and 'complex-valued nonlinear functions' without naming the specific targets or reporting quantitative error metrics (e.g., MSE or max deviation) for each.
[Abstract] The experimental claim of 'parallel approximation of 9 nonlinear functions in a single optical forward pass' would benefit from a clear description of the input encoding scheme and output detection geometry.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We address the two major comments point-by-point below.

read point-by-point responses

Referee: [Abstract] The universality claim rests on the physical realizability of an input-dependent dynamic encoder co-localized with the static decoder on a single phase-only plane. The abstract states that interference plus intensity detection alone generates arbitrary band-limited nonlinear maps, but does not specify how the input field modulates only the encoder region to produce the required cross term 2 Re(E_enc E_dec*) without additional nonlinear materials or separate stages; this assumption appears load-bearing for the span of achievable functions.

Authors: The manuscript details that the input modulates the phase values exclusively in the encoder sub-region of the single phase-only diffractive plane (via SLM), while the decoder sub-region holds fixed optimized phases. Coherent illumination of the entire plane produces fields that propagate and interfere, with square-law detection extracting the cross term. This mechanism is derived in Section II and the universality proof in Section III; we have expanded the abstract to explicitly state the co-localized phase modulation and cross-term origin. revision: yes
Referee: [Abstract] The abstract identifies decoder degrees-of-freedom, detector aperture, and axial propagation distance as the factors governing approximation fidelity, yet provides no quantitative bounds or scaling relations showing that finite values of these parameters suffice to reach the full target function class under the co-localization constraint.

Authors: The paper supplies a universality proof for the ideal (infinite-DoF) case and uses numerical/experimental analysis to identify the three parameters as fidelity limits. Explicit analytical scaling bounds for arbitrary finite values under co-localization are not derived, as they depend on the target function and noise model; we have added a new paragraph in the revised discussion section summarizing observed scaling trends from our simulations and experiments, while noting that full closed-form bounds constitute future work. revision: partial

Circularity Check

0 steps flagged

No circularity: universal-approximation claim is a self-contained mathematical proof

full rationale

The paper's central result is a proof that the described E+D architecture (co-localized input-dependent encoder and static decoder on one phase-only diffractive plane, followed by free-space propagation and intensity detection) forms a universal approximator for real-valued band-limited nonlinear functions. This is derived from the physical model of coherent interference and intensity detection rather than from any fitted parameter, self-citation chain, or definitional equivalence. No load-bearing step reduces to its own inputs by construction; the governing physical factors (decoder DoF, aperture, propagation distance) are identified as limits on fidelity but do not tautologically force the universality result. Experimental in-situ validation is presented as independent confirmation. The derivation therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Review conducted on abstract only; full details on parameters, axioms, and entities unavailable.

free parameters (1)

trained frozen phase bias
Mentioned as enhancing expressivity and robustness to quantization; value not specified in abstract.

axioms (2)

domain assumption Coherent illumination allows interference between encoder and decoder fields to produce nonlinear mappings via intensity detection
Core physical mechanism stated in abstract.
ad hoc to paper The E+D processor is a universal approximator for band-limited real-valued nonlinear functions
Claimed as proven but location and assumptions of proof not detailed in abstract.

pith-pipeline@v0.9.1-grok · 5801 in / 1335 out tokens · 30507 ms · 2026-06-28T16:53:23.167171+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 2 canonical work pages · 1 internal anchor

[1]

1 T. N. Theis, H. . -S. P. Wong. The End of Moore’s Law: A New Beginning for Information Technology. Computing in Science & Engineering 2017; 19: 41–50. 2 Leiserson CE, Thompson NC, Emer JS, Kuszmaul BC, Lampson BW, Sanchez D et al. There’s plenty of room at the Top: What will drive computer performance after Moore’s law? Science 2020; 368: eaam9744. 3 Sc...

2017
[2]

Analysis of Diffractive Optical Neural Networks and Their Integration With Electronic Neural Networks

10 Mengu D, Luo Y, Rivenson Y, Ozcan A. Analysis of Diffractive Optical Neural Networks and Their Integration With Electronic Neural Networks. IEEE Journal of Selected Topics in Quantum Electronics 2020; 26: 1–14. 11 Chen H, Lou S, Wang Q, Huang P, Duan H, Hu Y. Diffractive deep neural networks: Theories, optimization, and applications. Applied Physics Re...

2020
[3]

Optimizing structured surfaces for diffractive waveguides

17 13 Wang Y, Li Y, Gan T, Liao K, Jarrahi M, Ozcan A. Optimizing structured surfaces for diffractive waveguides. Nature Communications 2025; 16:

2025
[4]

Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits

14 Gao S, Chen H, Wang Y, Duan Z, Zhang H, Sun Z et al. Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits. Light: Science & Applications 2024; 13:

2024
[5]

Ultrahigh-fidelity spatial mode quantum gates in high- dimensional space by diffractive deep neural networks

15 Wang Q, Liu J, Lyu D, Wang J. Ultrahigh-fidelity spatial mode quantum gates in high- dimensional space by diffractive deep neural networks. Light: Science & Applications 2024; 13:

2024
[6]

Diffraction casting

16 Ryosuke Mashiko, Makoto Naruse, Ryoichi Horisaki. Diffraction casting. Advanced Photonics 2024; 6: 056005. 17 Sun Y, Dong M, Yu M, Liu X, Zhu L. Review of diffractive deep neural networks. J Opt Soc Am B 2023; 40: 2951–2961. 18 Rahman MSS, Li Y, Yang X, Chen S, Ozcan A. Massively parallel and universal approximation of nonlinear functions using diffrac...

2024
[7]

Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network

19 Li J, Tianyi Gan, Bijie Bai, Yi Luo, Mona Jarrahi, Aydogan Ozcan. Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network. Advanced Photonics 2023; 5: 016003. 20 Mengu D, Ozcan A. All-Optical Phase Recovery: Diffractive Computing for Quantitative Phase Imaging. Advanced Optical Materials 2022; 10: ...

2023
[8]

Giant photothermal nonlinearity in a single silicon nanostructure

22 Duh Y-S, Nagasaki Y, Tang Y-L, Wu P-H, Cheng H-Y, Yen T-H et al. Giant photothermal nonlinearity in a single silicon nanostructure. Nature Communications 2020; 11:

2020
[9]

Large optical nonlinearity enhancement under electronic strong coupling

23 Wang K, Seidel M, Nagarajan K, Chervy T, Genet C, Ebbesen T. Large optical nonlinearity enhancement under electronic strong coupling. Nature Communications 2021; 12:

2021
[10]

Ultra- strong nonlinear optical processes and trigonal warping in MoS2 layers

24 Säynätjoki A, Karvonen L, Rostami H, Autere A, Mehravar S, Lombardo A et al. Ultra- strong nonlinear optical processes and trigonal warping in MoS2 layers. Nature Communications 2017; 8:

2017
[11]

Nonlinear metasurfaces: a paradigm shift in nonlinear optics

25 Krasnok A, Tymchenko M, Alù A. Nonlinear metasurfaces: a paradigm shift in nonlinear optics. Materials Today 2018; 21: 8–21. 26 Xia F, Kim K, Eliezer Y, Han S, Shaughnessy L, Gigan S et al. Nonlinear optical encoding enabled by recurrent linear scattering. Nature Photonics 2024; 18: 1067–1075. 27 Wanxin Shi, Zheng Huang, Tingzhao Fu, Hongwei Chen. Revi...

2018
[12]

Roadmap on holography

32 Sheridan JT, Kostuk RK, Gil AF, Wang Y, Lu W, Zhong H et al. Roadmap on holography. Journal of Optics 2020; 22: 123002. 33 Delaire JA, Nakatani K. Linear and Nonlinear Optical Properties of Photochromic Molecules and Materials. Chem Rev 2000; 100: 1817–1846. 34 Bai B, Luo Y, Gan T, Hu J, Li Y, Zhao Y et al. To image, or not to image: class-specific dif...

2020
[13]

Fully nonlinear neuromorphic computing with linear wave scattering

35 Wanjura CC, Marquardt F. Fully nonlinear neuromorphic computing with linear wave scattering. Nature Physics 2024; 20: 1434–1440. 36 Bai B, Wei H, Yang X, Gan T, Mengu D, Jarrahi M et al. Data-class-specific all-optical transformations and encryption. Advanced Materials 2023; 35: 2212091. 37 Yildirim M, Dinc NU, Oguz I, Psaltis D, Moser C. Nonlinear pro...

work page arXiv 2024
[14]

Decoupled Weight Decay Regularization

44 Loshchilov I, Hutter F. Decoupled Weight Decay Regularization. 2019.https://arxiv.org/abs/1711.05101. 20 Figures Fig. 1 . Single-layer diffractive optical processor (E+D processor) for nonlinear function approximation under coherent illumination. a Schematic of the proposed architecture. A phase- only input plane with uniform intensity illumination is ...

work page internal anchor Pith review Pith/arXiv arXiv 2019

[1] [1]

1 T. N. Theis, H. . -S. P. Wong. The End of Moore’s Law: A New Beginning for Information Technology. Computing in Science & Engineering 2017; 19: 41–50. 2 Leiserson CE, Thompson NC, Emer JS, Kuszmaul BC, Lampson BW, Sanchez D et al. There’s plenty of room at the Top: What will drive computer performance after Moore’s law? Science 2020; 368: eaam9744. 3 Sc...

2017

[2] [2]

Analysis of Diffractive Optical Neural Networks and Their Integration With Electronic Neural Networks

10 Mengu D, Luo Y, Rivenson Y, Ozcan A. Analysis of Diffractive Optical Neural Networks and Their Integration With Electronic Neural Networks. IEEE Journal of Selected Topics in Quantum Electronics 2020; 26: 1–14. 11 Chen H, Lou S, Wang Q, Huang P, Duan H, Hu Y. Diffractive deep neural networks: Theories, optimization, and applications. Applied Physics Re...

2020

[3] [3]

Optimizing structured surfaces for diffractive waveguides

17 13 Wang Y, Li Y, Gan T, Liao K, Jarrahi M, Ozcan A. Optimizing structured surfaces for diffractive waveguides. Nature Communications 2025; 16:

2025

[4] [4]

Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits

14 Gao S, Chen H, Wang Y, Duan Z, Zhang H, Sun Z et al. Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits. Light: Science & Applications 2024; 13:

2024

[5] [5]

Ultrahigh-fidelity spatial mode quantum gates in high- dimensional space by diffractive deep neural networks

15 Wang Q, Liu J, Lyu D, Wang J. Ultrahigh-fidelity spatial mode quantum gates in high- dimensional space by diffractive deep neural networks. Light: Science & Applications 2024; 13:

2024

[6] [6]

Diffraction casting

16 Ryosuke Mashiko, Makoto Naruse, Ryoichi Horisaki. Diffraction casting. Advanced Photonics 2024; 6: 056005. 17 Sun Y, Dong M, Yu M, Liu X, Zhu L. Review of diffractive deep neural networks. J Opt Soc Am B 2023; 40: 2951–2961. 18 Rahman MSS, Li Y, Yang X, Chen S, Ozcan A. Massively parallel and universal approximation of nonlinear functions using diffrac...

2024

[7] [7]

Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network

19 Li J, Tianyi Gan, Bijie Bai, Yi Luo, Mona Jarrahi, Aydogan Ozcan. Massively parallel universal linear transformations using a wavelength-multiplexed diffractive optical network. Advanced Photonics 2023; 5: 016003. 20 Mengu D, Ozcan A. All-Optical Phase Recovery: Diffractive Computing for Quantitative Phase Imaging. Advanced Optical Materials 2022; 10: ...

2023

[8] [8]

Giant photothermal nonlinearity in a single silicon nanostructure

22 Duh Y-S, Nagasaki Y, Tang Y-L, Wu P-H, Cheng H-Y, Yen T-H et al. Giant photothermal nonlinearity in a single silicon nanostructure. Nature Communications 2020; 11:

2020

[9] [9]

Large optical nonlinearity enhancement under electronic strong coupling

23 Wang K, Seidel M, Nagarajan K, Chervy T, Genet C, Ebbesen T. Large optical nonlinearity enhancement under electronic strong coupling. Nature Communications 2021; 12:

2021

[10] [10]

Ultra- strong nonlinear optical processes and trigonal warping in MoS2 layers

24 Säynätjoki A, Karvonen L, Rostami H, Autere A, Mehravar S, Lombardo A et al. Ultra- strong nonlinear optical processes and trigonal warping in MoS2 layers. Nature Communications 2017; 8:

2017

[11] [11]

Nonlinear metasurfaces: a paradigm shift in nonlinear optics

25 Krasnok A, Tymchenko M, Alù A. Nonlinear metasurfaces: a paradigm shift in nonlinear optics. Materials Today 2018; 21: 8–21. 26 Xia F, Kim K, Eliezer Y, Han S, Shaughnessy L, Gigan S et al. Nonlinear optical encoding enabled by recurrent linear scattering. Nature Photonics 2024; 18: 1067–1075. 27 Wanxin Shi, Zheng Huang, Tingzhao Fu, Hongwei Chen. Revi...

2018

[12] [12]

Roadmap on holography

32 Sheridan JT, Kostuk RK, Gil AF, Wang Y, Lu W, Zhong H et al. Roadmap on holography. Journal of Optics 2020; 22: 123002. 33 Delaire JA, Nakatani K. Linear and Nonlinear Optical Properties of Photochromic Molecules and Materials. Chem Rev 2000; 100: 1817–1846. 34 Bai B, Luo Y, Gan T, Hu J, Li Y, Zhao Y et al. To image, or not to image: class-specific dif...

2020

[13] [13]

Fully nonlinear neuromorphic computing with linear wave scattering

35 Wanjura CC, Marquardt F. Fully nonlinear neuromorphic computing with linear wave scattering. Nature Physics 2024; 20: 1434–1440. 36 Bai B, Wei H, Yang X, Gan T, Mengu D, Jarrahi M et al. Data-class-specific all-optical transformations and encryption. Advanced Materials 2023; 35: 2212091. 37 Yildirim M, Dinc NU, Oguz I, Psaltis D, Moser C. Nonlinear pro...

work page arXiv 2024

[14] [14]

Decoupled Weight Decay Regularization

44 Loshchilov I, Hutter F. Decoupled Weight Decay Regularization. 2019.https://arxiv.org/abs/1711.05101. 20 Figures Fig. 1 . Single-layer diffractive optical processor (E+D processor) for nonlinear function approximation under coherent illumination. a Schematic of the proposed architecture. A phase- only input plane with uniform intensity illumination is ...

work page internal anchor Pith review Pith/arXiv arXiv 2019