Multimode optical fibers are terrible for imaging. They're excellent for carrying light — thin, flexible, capable of transmitting high power — but the multiple spatial modes mix as the light propagates, scrambling any image information. Bend the fiber, and the mixing changes. The output bears no recognizable relationship to the input. This has blocked their use in endoscopy and minimally invasive imaging despite their obvious mechanical advantages.
Feng, Lee, and colleagues (arXiv:2602.20562) bypass the scrambling entirely. Instead of trying to undo the mode mixing (which changes every time the fiber moves), they encode spatial information into nonlinear spectral signatures. Second-harmonic generation — a nonlinear optical process where two photons combine into one at double the frequency — produces broadband spectra that depend on the input spatial pattern but are invariant to how the modes mix during propagation.
The trick: mode mixing is linear. It shuffles the spatial modes but preserves the relationships that nonlinear generation depends on. The second-harmonic spectrum encodes spatial information in a way that linear scrambling cannot destroy. The encoding is in the nonlinear correlations between modes, not in the modes themselves.
The result: 0.82 reconstruction accuracy on Fashion-MNIST images, 92.3% classification accuracy on biomedical datasets, and — remarkably — models trained on one fiber work on entirely different fibers they've never seen (0.74 correlation). The nonlinear encoding is not just robust to perturbation of one fiber; it generalizes across fibers. The invariant is deeper than any specific fiber's properties.
The general observation: information encoded in nonlinear correlations can survive linear scrambling. The scrambling destroys linear relationships but preserves higher-order structure. If the information lives in the correlations, not the individual signals, it is protected against any transformation that preserves those correlations.