Sub-quadratic scalable approximate linear converter using multi-plane light conversion with low-entropy mode mixers

Abstract: Optical computing is emerging as a promising platform for energy-efficient, high-throughput hardware in deep learning. A key challenge lies in the realization of optical matrix-vector multiplication, which often requires $O(N^2)$ phase shifters for exact synthesis of $N \times N$ matrices, limiting scalability. In this study, we propose an approximate matrix realization method using multi-plane light conversion (MPLC) that reduces both the system size and the number of phase shifters while maintaining acceptable error bounds. This approach uses low-entropy mode mixers, allowing more compact implementations compared to conventional mixers. We introduce Shannon matrix entropy as a measure of mode coupling strength in mixers and demonstrate that low-entropy mixers can preserve computational accuracy while reducing the requirements for the mixers. The approximation quality is evaluated using the maximum norm between the target and realized matrices. Numerical results show that the proposed method achieves sub-quadratic scaling of phase shifters by tolerating predefined error thresholds. To identify efficient architectures for implementing general linear matrices, we compare block-encoding (BE) and singular-value decomposition (SVD) schemes for realizing general linear matrices using unitary converters based on MPLC. Results indicate that BE exhibits superior iterative configuration properties beyond the unitary group. By characterizing the trade-offs between matrix entropy, number of phase shifter layers, and the error tolerance, this study provides a framework for designing scalable and efficient approximate optical converters.