Free-space optical communication through the atmosphere is limited by turbulence — the wavefront arriving at the receiver is distorted, and the signal must be separated from the noise. The standard approach is mode-selective: choose a particular basis of spatial modes (Hermite-Gaussian, Laguerre-Gaussian, orbital angular momentum) and design a demultiplexer to sort the incoming light into those modes.
This paper finds that which modal basis you choose matters far less than how much spatial area (footprint) your basis covers. A compact two-plane refractive system, optimized via stochastic gradient descent over an ensemble of turbulence realizations, approaches the performance of ideal modal projection — and it does so regardless of the specific mode set. The footprint is the dominant variable; the basis is a detail.
Note (2026-11-14; read against the paper's full text): the essay says the modal basis hardly matters and the compact optimized system does well 'regardless of the specific mode set'. The paper's condition is that the choice of detection basis has limited impact on power collection provided the spatial support of the modes covers the collecting pupil efficiently; a dense packing of Gaussian modes after re-orthonormalization performs as well as an optimized Laguerre-Gaussian basis. The two-plane refractive system, optimized by gradient descent over an ensemble of turbulence realizations without any imposed modal decomposition, improves significantly over direct fiber coupling and approaches the performance of an ideal Karhunen-Loeve modal projection, and stays competitive across a range of turbulence conditions; this is a numerical design study for satellite-to-ground links with single-mode fiber bundles. I did not find a comparison with a design trained on a single turbulence realization (I searched for single realization, overfitting and generalization), so the essay's statement that such designs are brittle is not something the paper reports. 'Footprint' is the essay's word for the paper's spatial support.
The optimization target matters: designing for a single turbulence realization produces brittle systems that fail on the next one. Designing over the full turbulence ensemble produces robust demultiplexers that handle whatever the atmosphere throws at them. The constraint of optimizing against uncertainty — averaging over what you can't predict — enables generalization rather than limiting performance.
The field spent decades arguing about which modal basis is best. The answer is that the argument was about the wrong variable.