Approaching physical limits of latent dimensionality in optical computing

Chaojun Xu; Jinlong Xiang; Tao Lin; Xuan Hu; Xuhan Guo; Yao Zhou; Yikai Su; Youlve Chen; Yuchen Yin; Zhenyu Zhao

arxiv: 2605.23361 · v1 · pith:XD5TVTSSnew · submitted 2026-05-22 · ⚛️ physics.optics

Approaching physical limits of latent dimensionality in optical computing

Zhenyu Zhao , Zijun Qiu , Xuan Hu , Yao Zhou , Jinlong Xiang , Youlve Chen , Chaojun Xu , Yuchen Yin

show 3 more authors

Tao Lin Yikai Su Xuhan Guo

This is my paper

Pith reviewed 2026-05-25 03:45 UTC · model grok-4.3

classification ⚛️ physics.optics

keywords optical computingphotonic processorslatent dimensionalitymultimode opticsintegration densitywave physics

0 comments

The pith

Physical limits on latent dimensionality set universal metrics for maximum expressivity in bounded optical domains.

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

The paper establishes that fundamental physical limits on latent dimensionality exist for the maximum expressivity of any bounded optical domain. These limits are presented as universal metrics for assessing optical computing capacity, independent of specific device details. The authors validate the idea by fabricating ultracompact multimode photonic processors that approach the limits, reporting 86.7 percent accuracy on iris classification in a 2.2 by 8 micrometer device and 92.9 percent on digit recognition in a 20.6 by 44.8 micrometer device. The architecture is further scaled to implement a generative diffusion model for image synthesis. This supplies a wave-physics reference point to address the integration density ceiling in photonic AI hardware.

Core claim

Physical limits on the latent dimensionality for maximum expressivity of a bounded optical domain exist and serve as universal metrics for evaluating optical computing capacity. Ultracompact multimode photonic processors can be realized that approach these limits, delivering experimental accuracies of 86.7 percent on iris classification and 92.9 percent on handwritten digit recognition while also supporting a generative diffusion model.

What carries the argument

Physical limits on latent dimensionality derived from the wave physics of a bounded optical domain, proposed as universal metrics for computing capacity.

If this is right

Ultracompact multimode processors achieve high accuracy on iris classification and digit recognition tasks.
The architecture extends to complex generative tasks such as diffusion models for image synthesis.
Photonic processor design gains a theoretical reference point by grounding in the wave physics origin of latent dimensionality.

Where Pith is reading between the lines

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

If the limits are truly universal, they could standardize capacity comparisons across different photonic materials and geometries.
Approaching the limits may reduce the mismatch between optical domain potential and accessible manipulation range in future devices.
The same wave-physics grounding could be tested for applicability to other wave-based computing platforms.

Load-bearing premise

The derived physical limits on latent dimensionality are fundamental, independent of device geometry or material, and directly approachable in fabricated processors without additional unstated losses or constraints.

What would settle it

An experiment fabricating a multimode photonic processor that exceeds the predicted latent dimensionality for its physical size while maintaining classification accuracy, or one that cannot reach the predicted limit despite optimized design.

read the original abstract

The physical implementation of artificial intelligence requires mapping computational processes onto the dynamic physical processes of the underlying computing platform. The photonic processors offer an intrinsically parallel and low energy framework for this mapping, however, a mismatch between the potential computing capability of a bounded optical domain and the human accessible manipulation range sets a hard integration density ceiling on existing architectures. Here, we address this challenge by investigating the integration density limits in photonic processors through exploring the fundamental physical limits on the latent dimensionality for maximum expressivity of a bounded optical domain. These physical limits potentially serve as universal metrics for evaluating optical computing capacity. To validate these, we design and realize ultracompact multimode photonic processors approaching these limits: a 2.2 um by 8 um processor achieves 86.7 % accuracy in experiment for iris flower classification, and a 20.6 um by 44.8 um processor reaches 92.9% accuracy in handwritten digit recognition. Finally, we scale this architecture to highly complex tasks by implementing a generative diffusion model for image synthesis. By grounding photonic processor design in the wave physics origin of latent dimensionality, our results supply the missing theoretical reference point for optical computing architecture.

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 compact experimental devices hitting 87-93% accuracy on small tasks are the real takeaway, but the claimed universal physical limits on latent dimensionality lack shown independence from geometry or material.

read the letter

The main things here are the fabricated ultracompact multimode processors that reach 86.7% on iris classification in a 2.2 by 8 um footprint and 92.9% on digits in a 20.6 by 44.8 um one, plus scaling the same approach to a generative diffusion model. Those are tangible hardware results in optical computing. The paper also tries to tie device performance to wave-physics limits on how much latent dimensionality a bounded optical area can express. That connection is the part that could matter for scaling arguments. What works is the experimental side: they actually built and tested the small processors on real tasks instead of stopping at simulation. The sizes are aggressive and the accuracies are respectable for the footprint. The soft spot is the universality claim. The abstract calls the limits potential universal metrics, but nothing shown establishes that the bounds stay the same when refractive index, boundary conditions, or material dispersion change. Without that step, the limits read as tied to the specific setups rather than fundamental. It is also not clear from the summary how close the devices actually get to the theoretical maximum once losses are included. This paper is for people already working on photonic AI hardware who need concrete scaling data. A reader focused on device footprints and task performance will find the results useful even if the theory needs tightening. The experimental grounding is strong enough that it deserves a serious referee to check the derivations and the closeness-to-limit measurements.

Referee Report

2 major / 1 minor

Summary. The manuscript derives fundamental physical limits on latent dimensionality for maximum expressivity within a bounded optical domain, proposing these limits as universal metrics for optical computing capacity. It validates the approach by fabricating ultracompact multimode photonic processors that achieve 86.7% experimental accuracy on iris classification (2.2 µm × 8 µm device) and 92.9% on handwritten digit recognition (20.6 µm × 44.8 µm device), and scales the architecture to a generative diffusion model for image synthesis.

Significance. If the limits are rigorously shown to be invariant under geometry and material variations and the fabricated devices are quantitatively demonstrated to approach them, the work would supply a needed theoretical benchmark for photonic processor design, helping address integration-density ceilings in optical AI hardware. The reported experimental accuracies in highly compact footprints are notable, but their linkage to the claimed limits requires substantiation.

major comments (2)

[Theoretical derivation (abstract and introduction)] The central claim that the derived physical limits 'potentially serve as universal metrics' (abstract) rests on an unshown invariance: no derivation or parameter sweep demonstrates that the bounds remain unchanged under variations in refractive index profile, boundary conditions, or material dispersion. This invariance is load-bearing for the universality assertion and the 'approaching physical limits' framing.
[Experimental validation and results] No quantitative comparison of achieved versus theoretical latent dimensionality is reported for the experimental devices (accounting for losses, fabrication tolerances, or measurement of effective dimensionality), leaving the claim that the processors 'approach these limits' unanchored. The accuracies alone do not establish proximity to the bound.

minor comments (1)

[Abstract] The abstract states accuracies without error bars, number of trials, or comparison to electronic baselines of similar footprint; adding these would strengthen the experimental claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below, providing clarifications grounded in the existing derivation and experimental results while noting where revisions can strengthen the presentation.

read point-by-point responses

Referee: [Theoretical derivation (abstract and introduction)] The central claim that the derived physical limits 'potentially serve as universal metrics' (abstract) rests on an unshown invariance: no derivation or parameter sweep demonstrates that the bounds remain unchanged under variations in refractive index profile, boundary conditions, or material dispersion. This invariance is load-bearing for the universality assertion and the 'approaching physical limits' framing.

Authors: The limits are obtained from the maximum number of orthogonal spatial modes supportable within a bounded domain, as fixed by the wave equation and the ratio of domain size to wavelength. This counting of degrees of freedom is independent of the particular refractive-index distribution or dispersion relation inside the domain; the same bound applies for any passive linear medium filling a fixed spatial region. The qualifier 'potentially' in the abstract already signals that the result is framed at this level of generality rather than as a claim of exhaustive numerical invariance. We will add a short clarifying paragraph in the revised introduction and methods to make this independence explicit. revision: partial
Referee: [Experimental validation and results] No quantitative comparison of achieved versus theoretical latent dimensionality is reported for the experimental devices (accounting for losses, fabrication tolerances, or measurement of effective dimensionality), leaving the claim that the processors 'approach these limits' unanchored. The accuracies alone do not establish proximity to the bound.

Authors: We agree that a direct numerical comparison would strengthen the linkage. The device footprints were selected precisely so that the number of supported modes matches the derived bound for the operating wavelength; the reported classification accuracies demonstrate that the fabricated structures realize useful computation at those scales. In the revision we will insert explicit calculations of the theoretical latent dimensionality for each reported device size, together with estimates of effective dimensionality that incorporate measured insertion loss and fabrication tolerance, thereby anchoring the 'approaching' statement quantitatively. revision: yes

Circularity Check

0 steps flagged

No circularity: abstract asserts limits as universal metrics without showing derivation or fitted inputs

full rationale

The provided abstract states that physical limits on latent dimensionality 'potentially serve as universal metrics' and reports experimental accuracies on fabricated devices, but contains no equations, derivation steps, or parameter-fitting descriptions. No self-definitional, fitted-prediction, or self-citation load-bearing reductions are visible. The universality claim is asserted from 'wave physics origin' without exhibited reduction to device-specific data or prior self-citations. Per rules, absence of quotable circular steps requires score 0; the derivation chain is not inspectable here and cannot be flagged as circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents enumeration of free parameters, axioms, or invented entities; the central claim appears to rest on an unshown derivation of 'latent dimensionality' limits whose independence from fitted device parameters cannot be verified.

pith-pipeline@v0.9.0 · 5765 in / 1271 out tokens · 30316 ms · 2026-05-25T03:45:55.561808+00:00 · methodology

discussion (0)

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/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

latent dimensionality ... D = (W/λ)(n_max - n_min) ... integration density limit ... Δn_eff ... (Eq. 1-2, Supplementary Notes 1-3)
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction contradicts

?

contradicts
CONTRADICTS: the theorem conflicts with this paper passage, or marks a claim that would need revision before publication.

latent dimensionality ... scales positively with the cross-sectional width W ... fundamental physical limits ... independent of specific structural implementations

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

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

[1]

Ambrogio, S. et al. An analog-AI chip for energy-efficient speech recognition and transcription. Nature 620, 768–775 (2023). 8. Clements, W. R., Humphreys, P . C., Metcalf, B. J., Kolthammer, W. S. & Walsmley, I. A. Optimal design for universal multiport interferometers. Optica 3, 1460 (2016). 9. Shen, Y . et al. Deep learning with coherent nanophotonic c...

work page doi:10.1007/978-1-4613-2813-1 2023
[2]

Tian, Y . et al. Photonic transformer chip: interference is all you need. PhotoniX 6, 45 (2025). 33. Kingma, D. P . & Ba, J. Adam: A Method for Stochastic Optimization. (2014) doi:10.48550/ARXIV.1412.6980. 34. Alpaydin, E. & Kaynak, C. Optical Recognition of Handwritten Digits. UCI Machine Learning Repository https://doi.org/https://doi.org/10.24432/C50P4...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2025

[1] [1]

Ambrogio, S. et al. An analog-AI chip for energy-efficient speech recognition and transcription. Nature 620, 768–775 (2023). 8. Clements, W. R., Humphreys, P . C., Metcalf, B. J., Kolthammer, W. S. & Walsmley, I. A. Optimal design for universal multiport interferometers. Optica 3, 1460 (2016). 9. Shen, Y . et al. Deep learning with coherent nanophotonic c...

work page doi:10.1007/978-1-4613-2813-1 2023

[2] [2]

Tian, Y . et al. Photonic transformer chip: interference is all you need. PhotoniX 6, 45 (2025). 33. Kingma, D. P . & Ba, J. Adam: A Method for Stochastic Optimization. (2014) doi:10.48550/ARXIV.1412.6980. 34. Alpaydin, E. & Kaynak, C. Optical Recognition of Handwritten Digits. UCI Machine Learning Repository https://doi.org/https://doi.org/10.24432/C50P4...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2025