Restricted Modulation Freedom Enhances Noise Robustness in Coherent Diffractive Optical Networks
Abstract
In coherent diffractive optical networks, greater modulation freedom allows more flexible optimization for clean inputs, but its effect on noise robustness and its physical origin remain unclear. We derive an analytical framework that identifies the physical mechanism linking modulation freedom to noise robustness. To establish this connection, we compare a continuous DDNN (C-DDNN) with continuous amplitude and phase modulation and a binary-mask DDNN (BM-DDNN) with binary amplitude modulation. Trained only on clean MNIST data, the seven-layer BM-DDNN has 1.83 percentage points lower clean test accuracy but 32.82 percentage points higher noisy test accuracy under pixel-wise Gaussian noise. This robustness advantage cannot be explained by lower total noise-only intensity at the imaging plane: the BM-DDNN has 3.22 times the noise-only intensity but 8.64 times the clean-signal intensity of the C-DDNN, reducing noise contamination relative to the clean signal. We show that this difference arises from a clean-signal transmission bias generated by spatial correlations between the structured clean-signal field and learned modulation patterns. We quantify this mechanism with a cumulative transmission-bias factor K, linking modulation freedom to relative noise contamination and robustness. The C-DDNN exhibits a stronger suppressive bias (more negative K), whereas the BM-DDNN exhibits a weaker bias (less negative K^{\tilde}), yielding the consistent ordering K<K^{\tilde}. Because K requires only clean-data forward passes, it serves as a robustness screening metric before noisy-input simulations.
Related
- On-Chip Shaping of Surface Acoustic Waves via Continuous Diffractive Modulation
- Coherent errors, syndrome extraction and the Robustness of Magic in stabiliser codes
- SPLIT: Loosely-coherent inference of long-inspiral gravitational-wave sources for robustness against modeling and data outliers
- Beyond the Thin-Layer Limit: Differentiable Volumetric Training for Visible-Range Diffractive Neural Networks
- Varifocal diffractive lenses for multi-depth microscope imaging
BibTeX
@misc{lim2026restricted,
title = {{Restricted Modulation Freedom Enhances Noise Robustness in Coherent Diffractive Optical Networks}},
author = {Hyuntae Lim and Kyoungsik Kim},
howpublished = {arXiv},
year = {2026},
month = aug,
doi = {10.48550/arxiv.2605.30820},
eprint = {2605.30820},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.30820},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Restricted Modulation Freedom Enhances Noise Robustness in Coherent Diffractive Optical Networks
AU - Hyuntae Lim
AU - Kyoungsik Kim
PY - 2026
DA - 2026/08/28
PB - arXiv
DO - 10.48550/arxiv.2605.30820
UR - https://arxiv.org/abs/2605.30820
AB - In coherent diffractive optical networks, greater modulation freedom allows more flexible optimization for clean inputs, but its effect on noise robustness and its physical origin remain unclear. We derive an analytical framework that identifies the physical mechanism linking modulation freedom to noise robustness. To establish this connection, we compare a continuous DDNN (C-DDNN) with continuous amplitude and phase modulation and a binary-mask DDNN (BM-DDNN) with binary amplitude modulation. Trained only on clean MNIST data, the seven-layer BM-DDNN has 1.83 percentage points lower clean test accuracy but 32.82 percentage points higher noisy test accuracy under pixel-wise Gaussian noise. This robustness advantage cannot be explained by lower total noise-only intensity at the imaging plane: the BM-DDNN has 3.22 times the noise-only intensity but 8.64 times the clean-signal intensity of the C-DDNN, reducing noise contamination relative to the clean signal. We show that this difference arises from a clean-signal transmission bias generated by spatial correlations between the structured clean-signal field and learned modulation patterns. We quantify this mechanism with a cumulative transmission-bias factor K, linking modulation freedom to relative noise contamination and robustness. The C-DDNN exhibits a stronger suppressive bias (more negative K), whereas the BM-DDNN exhibits a weaker bias (less negative K^{\tilde}), yielding the consistent ordering K<K^{\tilde}. Because K requires only clean-data forward passes, it serves as a robustness screening metric before noisy-input simulations.
ER -