Parameter-Decoupled Superresolution with Airy-Disk-Informed Learning: Application to NGC 4486B
Abstract
Estimating the separation of two unresolved or nearly unresolved sources from recorded intensity images is difficult when source brightness, mutual coherence, and phase are unknown and the point-spread function (PSF) departs from a Gaussian. We formulate Hermite--Gaussian (HG) projections of one- and two-dimensional Airy amplitudes and prove that ratios of mode probabilities with equal longitudinal parity exactly cancel the brightness, coherence, and phase nuisance parameters under a symmetric two-source model. For a one-dimensional Airy slice we achieve an explicit multi-ratio separation inversion. For the physical two-dimensional circular-pupil PSF we obtain exact finite-pupil overlap integrals and locally invertible separation responses. Motivated by these analytical invariants, we examine Airy-informed convolutional neural networks (CNNs) on synthetic Airy images with known separations. At 150,000 training images, the approximate mean absolute errors are 0.018 in plotted separation units, 10 times better comparing to 0.18 for non-Airy Gaussian PSF approximation. Applied to an archival Hubble Space Telescope image of the double nucleus of NGC 4486B, the Airy-informed CNN gives estimates of approximately 9.6--12.0 pc for training sets of 100,000--150,000 images, broadly consistent with the historical approximately 12 pc value. We distinguish exact optical-mode nuisance cancellation from empirical robustness of an image-only learned estimator and identify the limits imposed by PSF mismatch and extended galaxy structure.
Related
- Parameter-Decoupling-Informed Learning for Superresolution: Laboratory Validation and Galaxy Imaging
- Physics-Informed Neural Networks vs U-Net: Application to thermal dust emission recovery in the microwave
- Metaheuristic physics-informed extreme-learning machine for initial orbit determination of near-Earth asteroids from too-short arcs
- CausticFlow: An Efficient Machine Learning Framework Combining Neural Differential Equations and Normalizing Flows for Binary Microlensing Parameter Inference
- CausticFlow: An Efficient Machine Learning Framework Combining Neural Differential Equations and Normalizing Flows for Binary Microlensing Parameter Inference
BibTeX
@misc{qu2026parameterdecoupled,
title = {{Parameter-Decoupled Superresolution with Airy-Disk-Informed Learning: Application to NGC 4486B}},
author = {James Qu and Isabella L. Lee and Gang Zhang and Abdelali Sajia and Yashashree Jadhav and Xiao-Feng Qian},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.36021},
eprint = {2610.36021},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.36021},
note = {CC BY-NC-SA 4.0},
}
RIS
TY - UNPB
TI - Parameter-Decoupled Superresolution with Airy-Disk-Informed Learning: Application to NGC 4486B
AU - James Qu
AU - Isabella L. Lee
AU - Gang Zhang
AU - Abdelali Sajia
AU - Yashashree Jadhav
AU - Xiao-Feng Qian
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.36021
UR - https://arxiv.org/abs/2610.36021
AB - Estimating the separation of two unresolved or nearly unresolved sources from recorded intensity images is difficult when source brightness, mutual coherence, and phase are unknown and the point-spread function (PSF) departs from a Gaussian. We formulate Hermite--Gaussian (HG) projections of one- and two-dimensional Airy amplitudes and prove that ratios of mode probabilities with equal longitudinal parity exactly cancel the brightness, coherence, and phase nuisance parameters under a symmetric two-source model. For a one-dimensional Airy slice we achieve an explicit multi-ratio separation inversion. For the physical two-dimensional circular-pupil PSF we obtain exact finite-pupil overlap integrals and locally invertible separation responses. Motivated by these analytical invariants, we examine Airy-informed convolutional neural networks (CNNs) on synthetic Airy images with known separations. At 150,000 training images, the approximate mean absolute errors are 0.018 in plotted separation units, 10 times better comparing to 0.18 for non-Airy Gaussian PSF approximation. Applied to an archival Hubble Space Telescope image of the double nucleus of NGC 4486B, the Airy-informed CNN gives estimates of approximately 9.6--12.0 pc for training sets of 100,000--150,000 images, broadly consistent with the historical approximately 12 pc value. We distinguish exact optical-mode nuisance cancellation from empirical robustness of an image-only learned estimator and identify the limits imposed by PSF mismatch and extended galaxy structure.
ER -