UCLA researchers prove optical processors can learn any function
Diffractive optical processors hit universal approximation with rigorous math, not just theory
Researchers from UCLA (led by Aydogan Ozcan) have published a rigorous mathematical framework proving that diffractive optical processors (DOPs) can universally approximate functions—a foundational result for optical computing. The 46-page paper (arXiv:2608.04582) connects classical universal approximation theory to optical physics, showing DOPs implement finite Fourier-feature expansions with phase-encoded inputs. Their analysis quantifies error sources (Fourier truncation, PSF synthesis, noise) and derives scaling laws linking optical hardware resources (degrees of freedom, space-bandwidth product) to approximation complexity.
The work also establishes photon-budget and throughput limits, demonstrating how spatially incoherent light and coherent cascadability enable enhanced representation power distinct from digital depth-separation results. Critically, the paper formulates learnability bounds for phase-quantized DOPs, offering design principles for large-scale analog optical systems. This bridges a long-standing gap between optical computing’s potential and theoretical guarantees, positioning DOPs as viable accelerators for AI workloads with speed/energy advantages.
- Proves DOPs can universally approximate functions with quantified error bounds and scaling laws
- 46-page theoretical framework links Fourier analysis, optical physics, and learnability for phase-encoded DOPs
- Derives photon-budget limits and coherent cascadability advantages over digital neural networks
Why It Matters
Optical processors could replace GPUs for AI tasks by offering faster, lower-power alternatives with theoretical guarantees of computational universality.