Aiming at the problems of high computational complexity, inefficient modal redundant feature processing and lack of theoretical guarantees of traditional methods in multimodal fusion, this paper proposes the Sparse Spectrum Transformer (SST), a framework that achieves scalable multimodal feature fusion through the deep fusion of dynamic frequency domain analysis and sparse attention mechanism. At the theoretical level, the error upper bound of cross-modal dynamic sparse attention is derived for the first time, which proves the decisive role of sparsity s and modal alignment coefficient λ on the fusion accuracy. At the algorithmic level, the Dynamic Spectral Gating (DSG) module is designed to achieve adaptive selection of frequency-domain features and hardware co-optimization (hybrid accuracy computation with block sparse Flash Attention through frequency-domain importance sampling (retaining the first 10% low-frequency components), cross-modal similarity matrix modelling and differentiable sparsity (Top-k Gumbel-Softmax). Experiments show that SST significantly outperforms nnFormer, PolyFormer and other baseline methods in medical image segmentation (4.2% improvement in Dice Score), remote sensing target detection (3.5% improvement in mAP), and autonomous driving sensing (30 FPS inference speed at 1080p resolution), with a 58% reduction in memory usage, and has passed ACM code audit and validation for multi-platform compatibility. It also passed the ACM code audit and multi-platform compatibility verification. This study provides a new paradigm for multimodal fusion that combines theoretical rigor and engineering practicality.