---
title: Spherical Gaussians in 3D Rendering & Analysis
url: https://www.emergentmind.com/topics/spherical-gaussians
type: topic
---

# Spherical Gaussians in 3D Rendering & Analysis

Spherical Gaussians (SGs) are a family of directionally concentrated, exponentially decaying functions defined on the unit sphere. They encode localized, orientation-dependent phenomena in a highly compact and computationally efficient manner, and have emerged as critical primitives in 3D rendering, geometric learning, harmonic analysis, spatial statistics, and scientific computing. Modern research leverages spherical Gaussians for view-dependent color modeling, BRDFs, illumination compression, statistical modeling, and as localized angular basis functions combining sharpness, locality, and flexible parameterization.

## 1. Mathematical Definition and Parameterization

A canonical (isotropic) spherical Gaussian is defined as a radially symmetric lobe centered on the unit sphere $\mathbb{S}^2$:

\[
G(\mathbf{d}; \alpha, \lambda, \boldsymbol{\mu}) = \alpha\, \exp\left[\lambda(\mathbf{d} \cdot \boldsymbol{\mu} - 1)\right]
\]

where:

- $\mathbf{d} \in \mathbb{R}^3$, $\|\mathbf{d}\|=1$: query or viewing direction,
- $\boldsymbol{\mu} \in \mathbb{R}^3$, $\|\boldsymbol{\mu}\|=1$: mean or “lobe” direction,
- $\lambda \in \mathbb{R}_+$: sharpness (larger $\lambda$ results in a narrower lobe),
- $\alpha \in \mathbb{R}^3$ or $\mathbb{R}$: amplitude (e.g., RGB or scalar).

The exponential term decays rapidly as the angle between $\mathbf{d}$ and $\boldsymbol{\mu}$ increases, providing angular localization. The unnormalized peak at alignment ensures that multiple SGs can be summed without over-smoothing high-frequency details [2501.00342, 2509.07021]. Anisotropic and higher-rank generalizations add further directionality via additional axis-aligned sharpness parameters [2303.08064, 2412.06511].

## 2. Efficient Representation in 3D Gaussian Splatting

Spherical Gaussians have become the dominant color representation in 3D Gaussian Splatting (3DGS), supplanting traditional spherical harmonics (SH). In SG-Splatting [2501.00342]:

- Each splat stores a view-independent diffuse color $D$ and one or more SG lobes.
- For single-lobe,
  \[
  C(\mathbf{d}) = D + \alpha\, \exp\left[\lambda (\mathbf{d} \cdot \boldsymbol{\mu} - 1)\right]
  \]
- For enhanced fidelity, multiple orthogonally-arranged axes ($\boldsymbol{\mu}_i$) may be employed.
- Parameter storage is drastically reduced: for third-degree SH, $48$ floats per splat (three channels $\times$ sixteen basis functions); for single-lobe SG, only $10$ floats ($D$: $3$, $\alpha$: $3$, $\boldsymbol{\mu}$: $3$, $\lambda$: $1$). The three-lobe orthogonal SG model uses $15$ floats.

Practical impact includes a $4$–$5\times$ reduction in memory for color representation, $40$–$50\%$ overall memory reduction, and $\sim1.5\times$ faster rendering due to arithmetic and bandwidth reduction, with SIMDintrinsic-friendly kernels [2501.00342, 2509.07021].

## 3. Multi-lobe and Mixed Spherical Gaussian Models

Complex view-dependent effects (e.g., specular highlights, reflective lobes) cannot be captured by a single SG. Multi-lobe schemes:

\[
C(\mathbf{d}) = D + \sum_{i=1}^n \alpha_i \exp\left[\lambda_i(\mathbf{d} \cdot \boldsymbol{\mu}_i - 1)\right]
\]
- For $n=3$, axes may be fixed to $(1,0,0)$, $(0,1,0)$, $(0,0,1)$ (orthogonal arrangement).
- Parameters can be shared or fixed per primitive to minimize memory overhead.

To improve the representation of both low- and high-frequency components, mixed models combine low-degree SH with SG lobes:

\[
C(\mathbf{d}) = \mathrm{SH}_{\text{low}}(\mathbf{d}) + \sum_i \alpha_i \exp\left[\lambda_i(\mathbf{d}\cdot \boldsymbol{\mu}_i - 1)\right]
\]
SH degree is adapted per primitive size (degree $0$ for small, $1$ for medium, $2$ for large), leveraging SH for smooth backgrounds and SGs for sharp, view-dependent highlights [2501.00342].

## 4. Compression, Pruning, and Parameterization Strategies

Memory and compute efficiency are further advanced via joint pruning of primitives and lobes. For instance, MEGS$^2$ [2509.07021] replaces SH with arbitrarily-oriented SG lobes (each with $7$ floats: $3$ for $\boldsymbol{\mu}$, $1$ for sharpness, $3$ for amplitude) and formulates a unified constrained optimization that prunes both primitives and individual lobes, with parameter budgets expressed as $\rho_o \|\mathbf{o}\|_0 + \rho_s \|\mathbf{s}\|_0 \leq \kappa$ ($\rho_o$ and $\rho_s$ are cost per primitive and per lobe).

Lobes with low sharpness or amplitude are pruned, and their “energy” is merged into the diffuse term via:

\[
\Delta c_0 = a_i\, \frac{1-e^{-2s_i}}{2s_i}
\]
MEGS$^2$ demonstrates 16$\times$ static and 8$\times$ rendering VRAM reduction compared to naive 3DGS, with negligible visual quality loss [2509.07021].

## 5. Comparison with Spherical Harmonics and Related Bases

Spherical Gaussians differ

Source: https://www.emergentmind.com/topics/spherical-gaussians