Learned Diffractive Optics for Quantum-Optimal Inference

Code and data for arXiv:2609.08014

Quantum mechanics sets the ultimate bounds on photon-limited sensing, yet practical measurements attaining these bounds are known only in special cases. This is particularly the case for visual sensing problems, where the goal is to infer features of a distant object based on the spatial structure of the light field it emits or reflects. Because of the potentially complex structure of such objects and fields, constructing optimal measurements on them is a challenging task. Here, we apply learned diffractive optics to state discrimination and parameter estimation of coherent and diffraction-limited incoherent light fields under a restricted photon budget. Optimized directly on each task's figure of merit, without prior knowledge of the optimal measurement, the physically realizable diffractive optical neural networks substantially outperform standard measurements and approach the quantum limits for a given number of photons as well as in the asymptotic limit.

Usage

uv sync

Run each experiment with:

uv run python -m experiments.<name>.bounds    # quantum limits
uv run python -m experiments.<name>.train     # diffractive optics
uv run python -m experiments.<name>.plot      # create figure

directory	figure	task
simple_example	2	two Hermite-Gaussian superpositions
coherent_discrimination	3	two phase-imprinted blood cells
incoherent_discrimination	4	two blurred handwritten digits
coherent_estimation	5	defocus (axial localization)
incoherent_estimation	6	two-source separation

core/ holds the shared code: the diffractive optics simulated with TorchOptics, the figures of merit, the optimiser, file I/O and the figure style. Experiment-specific physical settings are at the top of each states.py; optimizer and sampling settings are in train.py and bounds.py.