Researchers have demonstrated that large-scale nonlinear computations can be executed using exclusively linear optical components by coupling phase-based wavefront encoding with optimized diffractive processors. Reported in eLight by an engineering research team led by Md Sadman Sakib Rahman, Yuhang Li, Xilin Yang, Shiqi Chen, and Aydogan Ozcan, the framework overcomes a longstanding operational obstacle in optical computing: the reliance on weak, power-demanding optical nonlinearities. By encoding input variables into the phase profile of an optical wavefront and propagating the light through structured, passive phase-only surfaces, the system harnesses linear diffraction to perform complex-valued mathematical transformations, producing output intensities that accurately approximate extensive libraries of arbitrary nonlinear functions simultaneously at the diffraction limit of light.
Digital electronic processors face mounting performance limitations governed by the slowdown of Moore’s law alongside escalating energy demands from deep learning and artificial intelligence workloads. While analog optical computing offers potential advantages in operational speed and energy efficiency, conventional optical systems struggle to implement the nonlinear functions essential for general computing and neural network activation. Physical optical nonlinearities typically demand extreme field intensities, such as more than 10^11 watts per square centimeter for second harmonic generation or over 10^4 watts per square centimeter for Kerr effects, which far exceed ambient light levels. Slower mechanisms such as photochromic or photorefractive effects suffer from response times stretching over seconds or present strong material absorption exceeding 90 percent, leaving optical computing systems dependent on power-hungry optoelectronic interfaces.
The mathematical foundation introduced by the researchers proves that a diffractive processor acts as a universal function approximator for arbitrary sets of bandlimited nonlinear functions. In this formulation, an input variable is converted across a designated set of input pixels into specific phase values, creating a structured wavefront comprising spatial harmonics. As the light traverses consecutive surfaces featuring spatially optimized phase profiles, the coherent point-spread functions of the diffractive layers execute a weighted linear summation over these harmonics. Because this operation mirrors the calculation of a Fourier series expansion, the optical field arriving at individual output pixels tracks the values of distinct nonlinear target functions, executing mathematical mappings without requiring any physical nonlinear medium or digital post-processing steps.
To evaluate the scalability and precision of this framework, the authors conducted comprehensive numerical simulations across five distinct optical processor architectures designed for an operating illumination wavelength of 550 nanometers. The processors were modeled with the number of target nonlinear functions scaling from 100 up to one million distinct functions, using nine input encoding pixels and diffractive features distributed evenly across two passive phase-only surfaces. The total number of trainable diffractive features scaled proportionally with the problem complexity according to a derived capacity condition. Even when simultaneously computing one million distinct nonlinear functions at wavelength-scale spatial density, the root-mean-square approximation errors remained remarkably low, with the worst-case diffractive output curves remaining visually indistinguishable from the underlying mathematical target functions.
The study also analyzed the physical and architectural factors that dictate function approximation performance, focusing on structural layer depth and diffraction efficiency. In two-layer diffractive designs, unscattered ballistic photons passing through the center of the diffractive volume caused slightly elevated approximation errors at the central output pixels. Expanding the processor architecture to four diffractive layers provided the structural depth necessary to redirect and redistribute these ballistic waves across the output aperture, yielding lower and more spatially uniform error profiles. Furthermore, four-layer architectures demonstrated superior light throughput, achieving higher average diffraction efficiencies than two-layer designs across configurations with equal parameter counts.
Recognizing that practical computing hardware must balance computational fidelity against optical transmission losses, the researchers introduced an engineered training loss function that incorporates an explicit diffraction efficiency penalty. By tuning a target threshold parameter within this composite loss formulation during digital gradient descent optimization, the team evaluated processors trained on 10,000 functions across different efficiency regimes. The results revealed an explicit trade-off: setting higher throughput thresholds increased the overall realized optical power reaching the detectors while introducing slight increases in function approximation error. Under the most power-efficient design tested, the diffractive output closely matched target functions across the core input domain, exhibiting discernible deviations only near the extreme edges of the variable dynamic range.
Beyond one-dimensional mathematical mappings, the theoretical framework naturally extends to multi-dimensional, multi-wavelength, and complex-valued functions. By assigning individual input pixels to multi-dimensional Fourier frequencies, the researchers numerically validated the parallel execution of 10,000 distinct three-dimensional nonlinear functions using 64 input pixels. The team also showed that diffractive processors can approximate common artificial neural network activation functions, including sigmoid, tanh, ReLU, and softplus, over finite input domains, where high-frequency oscillations near boundaries stem from finite Fourier bandwidth limits. Furthermore, by optimizing a multi-wavelength diffractive processor across 100 discrete spectral channels spanning 451 to 550 nanometers, the system demonstrated the capacity to execute 10,000 independent functions divided evenly across different illumination wavelengths without relying on material dispersion engineering.
A critical operational property demonstrated in the study is the all-optical cascadability of diffractive function approximators, which operates independently of optoelectronic detection. Because the complex optical wavefield at the output aperture retains both amplitude and phase information, it can directly serve as the optical input for a subsequent diffractive processor. The team numerically demonstrated a cascaded system where an initial two-layer diffractive processor computed 100 complex-valued nonlinear functions, and its complex output wavefield was directly operated upon by a second diffractive processor to compute another set of 100 distinct functions. The cascaded stage maintained negligible approximation errors across both real and imaginary field components, confirming that diffractive processors can form sequential, multi-stage analog optical processing pipelines entirely in the optical domain.
To confirm the physical viability of the concept under real-world operating conditions, the researchers constructed an experimental prototype utilizing a 520-nanometer laser source, a single reflective phase-only spatial light modulator, and an image sensor. To circumvent experimental imperfections such as mechanical vibrations, pixel misalignments, and optical component aberrations that hinder direct transfer of digitally trained phase masks, the researchers devised an in situ optical training framework based on proximal policy optimization reinforcement learning. Within this setup, a central 96-by-96 pixel block of the modulator applied dynamic phase patterns corresponding to input arguments, while the surrounding 1080-by-1920 pixels served as learnable diffractive features optimized directly through real-time camera feedback.
Using a progressive optimization strategy where target functions were sequentially added in increments of five whenever experimental loss settled below 0.1, the physical system successfully achieved parallel single-shot execution of 35 distinct nonlinear functions over a four-day training cycle. The optical output intensities, monitored across 35 signal channels alongside 35 reference channels designed to normalize illumination intensity drift, demonstrated close agreement with mathematical target curves, achieving an experimental mean absolute error of approximately 0.06. Parallel in silico baseline simulations targeting the same 35-function set achieved near-perfect curve matching, illustrating that experimental discrepancies arose from physical environmental noise rather than structural design limitations.
The study also examined operational tolerances and manufacturing considerations that govern practical implementations of diffractive computing platforms. Numerical sensitivity evaluations revealed that monochromatically trained processors maintain stable approximation accuracy provided the relative deviation between training and operating wavelengths does not exceed 10^-8, a performance window well within the linewidth tolerances of standard external-cavity and stabilized laboratory lasers. While higher-dimensional inputs and larger function sets require proportionately more optimizable phase features to prevent accuracy degradation, modern high-resolution image sensors featuring hundreds of megapixels offer the spatial capacity required to scale this architecture toward hundreds of millions of parallel functions, establishing a versatile paradigm for passive, ultrafast analog computation built entirely from linear optical materials.
Subject of Research: Parallel nonlinear function approximation using linear diffractive optical processors
Article Title: Massively parallel and universal approximation of nonlinear functions using diffractive processors
Article References: Rahman, M. S. S., Li, Y., Yang, X., Chen, S., & Ozcan, A. (2025). Massively parallel and universal approximation of nonlinear functions using diffractive processors. eLight, 5(1), Article 32. https://doi.org/10.1186/s43593-025-00113-w
Image Credits: AI Generated
DOI: 10.1186/s43593-025-00113-w
Keywords: diffractive optical processors, analog optical computing, nonlinear function approximation, spatial light modulators, universal approximation theorem, linear optics, optical computing, wavefront encoding, in situ learning, diffraction efficiency, machine learning hardware, point-spread functions, multi-wavelength processing, cascadable optics, neuromorphic computing, optical neural networks, Fourier basis expansion
Cite Scienmag News
Teresa Odom. (November 21, 2026). Diffractive Optical Processors Approximate Millions of Nonlinear Functions Using Linear Optics. Scienmag. https://scienmag.com/diffractive-optical-processors-approximate-millions-of-nonlinear-functions-using-linear-optics/
Teresa Odom. "Diffractive Optical Processors Approximate Millions of Nonlinear Functions Using Linear Optics." Scienmag, 21 November 2026, https://scienmag.com/diffractive-optical-processors-approximate-millions-of-nonlinear-functions-using-linear-optics/. Accessed 21 November 2026.
Teresa Odom. "Diffractive Optical Processors Approximate Millions of Nonlinear Functions Using Linear Optics." Scienmag. November 21, 2026. https://scienmag.com/diffractive-optical-processors-approximate-millions-of-nonlinear-functions-using-linear-optics/








