Mathematical Supplement for the gsplat Library

Ye and Kanazawa publish a self-contained mathematical reference for gsplat, the modular differentiable GaussianSplatting library hosted by the nerfstudio project. The original 3DGS paper by Kerbl et al. introduces the representation and the high-level rendering pipeline, but it does not spell out every term of the forward and backward passes at the level of detail needed for reimplementation or extension. gsplat's supplement fills that gap.

The note walks through projection of 3D Gaussians into screen space under a pinhole camera, including the local affine approximation of the perspective transform and the propagation of covariance through it. It covers tile-based rasterisation, front-to-back alpha compositing, and the exact gradient flow back through opacity, colour, position, scale, and rotation. Spherical-harmonic colour evaluation and its gradients are documented alongside the choices of basis convention used by the library. The text is intentionally implementation-aligned: each section maps to a specific kernel or Python entry point in the library, so a reader can move between equations and code without ambiguity.

The accompanying Python API exposes each component of the forward and backward passes individually, making gsplat a common substrate for research extensions of 3DGS — for example, surface-aware variants such as SuGaR or 2D Gaussian Splatting, dynamic methods such as 4D Gaussian Splatting, and SLAM systems such as Gaussian Splatting SLAM and GS-SLAM have all reused or imitated gsplat's rasteriser interface. As a result, the supplement is short but functions as the canonical low-level reference for the GaussianSplatting ecosystem.