Abstract
Photonic computing promises energy-efficient acceleration for optimizationand learning, yet discrete combinatorial search and continuous functionapproximation have largely required distinct devices and control stacks. Herewe unify k-local Ising optimization and optical Kolmogorov-Arnold network (KAN)learning on a single photonic platform, establishing a critical convergencepoint in optical computing. We introduce an SLM-centric primitive thatrealizes, in one stroke, all-optical k-local Ising interactions and fullyoptical KAN layers. The key idea is to convert the structural nonlinearity of anominally linear scatterer into a per-window computational resource by adding asingle relay pass through the same spatial light modulator: a folded 4f relayre-images the first Fourier plane onto the SLM so that each selected clique orchannel occupies a disjoint window with its own second pass phase patch.Propagation remains linear in the optical field, yet the measured intensity ineach window becomes a freely programmable polynomial of the clique sum orprojection amplitude. This yields native, per clique k-local couplings withoutnonlinear media and, in parallel, the many independent univariatenonlinearities required by KAN layers, all trainable with in-situ physicalgradients using two frames (forward and adjoint). We outline implementations onspatial photonic Ising machines, injection-locked vertical cavity surfaceemitting laser (VCSEL) arrays, and Microsoft analog optical computers; in allcases the hardware change is one extra lens and a fold (or an on-chip 4f loop),enabling a minimal overhead, massively parallel route to high-order Isingoptimization and trainable, all-optical KAN processing on one platform.