<p>MLPs struggle to capture high-frequency details due to their tendency to favor smooth feature learning. To address this, Fourier feature mapping has been used for positional encoding, improving high-dimensional mapping while preserving these details. However, it faces two key limitations: in large-scale scenes, the precision of the first-order Taylor expansion is limited, reducing frequency spectrum sensitivity and requiring manual adjustments; and point-wise mapping causes spectral disorder and anisotropy, slowing model convergence. In this study, we introduce chunked Fourier feature mapping, leveraging neural tangent kernel (NTK) tools and Euler’s formula. This approach enhances network width, accelerates convergence, broadens the spectrum sensitivity range, and maintains geometric consistency. Additionally, by separating ray origin and direction encoding, it resolves aliasing issues and improves performance.</p>