Gaussian splatting renders 3D scenes by projecting Gaussian kernels onto the image plane. The kernel is exponential — exp(-x²) — because Gaussians are Gaussians. The exponential never reaches zero, which means every Gaussian technically contributes to every pixel, requiring careful culling to keep rendering tractable.
Note (2026-11-15; read against the paper's full text): the essay attributes the whole gain to more aggressive culling and calls the culling provably lossless. The paper also motivates the swap by cost: evaluating the exponential function is a significant fraction of the already expensive blending step, and it states the aim as replacing the Gaussian kernel with a simpler, more computationally efficient approximation. The new kernel is itself an approximation: it fits the polynomial coefficients (first-, second- and third-order polynomials) with an L1 loss and a sampling strategy reflecting the practical distribution of quadric values, to stay compatible with existing 3DGS datasets, and it evaluates how quality is affected (the abstract says 'negligible impact on image quality'). The finite support gives the tight bounding radii the paper derives for culling, so the culling is exact with respect to the polynomial, but the kernel swap is not lossless with respect to the original Gaussian.
Mueller, Winter, and Steinberger replace the exponential with a polynomial approximation clipped by ReLU. The polynomial has compact support — it reaches exactly zero at a finite radius — which means culling is no longer an approximation. Gaussians beyond the polynomial's support contribute nothing, and the renderer can skip them without error.
The result: 4 to 15% performance improvement with negligible impact on image quality. The gain comes entirely from more aggressive culling. The exponential kernel requires a threshold-based cutoff — discard Gaussians contributing less than ε — and the threshold choice trades quality for speed. The polynomial kernel eliminates the tradeoff: the mathematically exact boundary enables provably lossless culling.
The structural insight: the Gaussian kernel was chosen for mathematical elegance, not computational efficiency. Exponentials are smooth, analytically tractable, closed under convolution. But rendering is not analysis. Rendering needs to know which primitives affect which pixels, and a function that never reaches zero makes that question unanswerable without arbitrary thresholds. Replacing mathematical elegance with computational pragmatism — a polynomial that does reach zero — turns an approximate operation into an exact one. The less beautiful function is the more useful one.