---
title: 'WRF-GS: Weighted Residual Gaussian Splatting'
url: https://www.emergentmind.com/topics/weighted-residual-formulation-for-gaussian-splatting-wrf-gs
type: topic
---

# WRF-GS: Weighted Residual Gaussian Splatting

The Weighted Residual Formulation for Gaussian Splatting (WRF-GS) provides a principled, theoretically grounded methodology for function approximation and inverse problems, especially in low-dimensional settings. The WRF-GS unifies the Gaussian Splatting paradigm—where models are constructed as finite mixtures of multivariate Gaussians—with a weighted-residual, least-squares loss that is minimized via Wasserstein–Fisher–Rao (WFR) gradient flows. This synthesis allows the direct interpretation of splat model optimization as a particle descent in a geometric space of mixing measures, yielding a tractable, expressive, and interpretable approach to model fitting and estimation [2511.14042].

## 1. Gaussian Splat Model Definition

The core of the WRF-GS is the Gaussian Splat model, an extension of the classical mixture-of-Gaussians paradigm. A target function \(y = f^*(x)\) is approximated by

\[
f(x; \Theta) = \sum_{k=1}^K w_k\, \mathcal{N}(x; \mu_k, \Sigma_k)
\]

where

- \(w_k \in \mathbb{R}\) (or \(\mathbb{R}^p\) for vector-valued output) are mixture weights,
- \(\mu_k \in \mathbb{R}^d\) are Gaussian centers,
- \(\Sigma_k \in \mathrm{PD}(d)\) are symmetric positive-definite covariance matrices.

The model parameters are collectively denoted as \(\Theta = \{(w_k, \mu_k, \Sigma_k)\}_{k=1}^K\), and each component \(\mathcal{N}(x; \mu, \Sigma)\) is the standard multivariate Gaussian density. Heterogeneous and anisotropic splats are supported, allowing arbitrary scale and orientation for each basis function.

Mixtures of Gaussians constitute a dense subset in \(C(\Omega)\) for any compact domain \(\Omega \subset \mathbb{R}^d\), thus forms a universal approximator with established rates for the required number of components as a function of approximation error and dimension.

## 2. Weighted Residual Loss and Empirical Risk

The fitting objective in WRF-GS is a weighted least-squares loss over the data \((x_i, y_i)_{i=1}^N\):

\[
J(\Theta) = \sum_{i=1}^N w(x_i)\, [f(x_i; \Theta) - y_i]^2
\]

where the weighting function \(w(x) > 0\) can encode sampling density, importance, or region-specific emphasis. The residuals at each point are \(r_i(\Theta) = f(x_i; \Theta) - y_i\).

Alternatively, the loss is interpretable as an \(L^2_\pi\) risk under a measure \(\pi(dx)\) with density \(w(x)\):

\[
J(\Theta) = \int_{\mathbb{R}^d} [f(x; \Theta) - y(x)]^2\, \pi(dx)
\]

In the limit of \(N \to \infty\) and uniform sampling, this recovers the continuous risk functional.

## 3. Wasserstein–Fisher–Rao Gradient Flow Interpretation

Each Gaussian splat can be seen as a "particle" state \((w, \mu, \Sigma)\) and collected into an empirical mixing measure:

\[
\mu_\Theta = \frac{1}{K} \sum_{k=1}^K w_k\, \delta_{(\mu_k, \Sigma_k)}
\in P(\mathbb{R}^d \times \mathrm{PD}(d))
\]

Model evaluation is equivalent to an expectation over the measure \(\mu_\Theta\):

\[
f(x; \Theta) = \mathbb{E}_{(\mu, \Sigma) \sim \mu_\Theta} [w\, \mathcal{N}(x; \mu, \Sigma)]
\]

and the risk \(J(\Theta)\) is interpretable as \(F(f_{\mu_\Theta})\) for \(F(f) = \int (f(x) - y(x))^2\, \pi(dx)\).

Minimization of this empirical risk is cast as a gradient flow under the WFR metric using the JKO variational update:

\[
\mu^{(n+1)}
= \arg\min_\mu\, \frac{1}{2\tau} d_\mathrm{WFR}^2(\mu, \mu^{(n)}) + F(f_\mu)
\]

The associated Euler–Lagrange equation leads, as \(\tau \to 0\), to a continuity–reaction PDE on the mixing measure:

\[
\partial_t \mu_t + \nabla_\mu \cdot (\mu_t\, v_t) = \mu_t\, \alpha_t
\]

where \((v_t, \alpha_t)\) are the WFR gradients of the risk functional.

When pulled back to the finite-dimensional parameter space, this gives explicit ODEs for the model parameters. For scalar outputs, the parameter gradients are:

\[
\partial_{w_k} J = 2 \sum_{i=1}^N w_i r_i\, \mathcal{N}(x_i; \mu_k, \Sigma_k)
\]
\[
\partial_{\mu_k} J = 2 \sum_{i=1}^N w_i r_i\, w_k\, \mathcal{N}(x_i; \mu_k, \Sigma_k)\, \Sigma_k^{-1}(x_i - \mu_k)
\]
\[
\partial_{\Sigma_k} J = \sum_{i=1}^N w_i r_i\, w_k\, \mathcal{N}(x_i; \mu_k, \Sigma_k)\, \frac{1}{2} \left[
  \Sigma_k^{-1}(x_i-\mu_k)(x_i-\mu_k)^T\Sigma_k^{-1} - \Sigma_k^{-1}
\right]
\]

## 4. Discrete-Time Optimization Algorithm

Parameter updates in WRF-GS are performed via discrete-time gradient descent, optionally using momentum or Adam. One iteration involves:

- Computing residuals \(r_i\) at all data locations.
- For each Gaussian component:
  - Compute gradients for weight, mean, and covariance as above.
  - Update parameters with appropriate step sizes (\(\eta_w, \eta_\mu, \eta_\Sigma\)), optional weight regularization (\(\lambda_w\)), and isotropic covariance regularization (\(\rho\)).

The update step can be summarized as follows:

```python
# Inputs: {x_i, y_i, w_i} for i=1..N; current {w_k, μ_k, Σ_k} for k=1..K
r_i = sum_k(w_k * N(x_i; μ_k, Σ_k)) - y_i  # For each i
for k in 1..K:
  g_w = 2 * sum_i(w_i * r_i * N(x_i; μ_k, Σ_k))
  g_μ = 2 * sum_i(w_i * r_i * w_k * N(x_i; μ_k, Σ_k) * Σ_k⁻¹ * (x_i - μ_k))
  g_Σ = sum_i(w_i * r_i * w_k * N(x_i; μ_k, Σ_k) *
              0.5 * [Σ_k⁻¹(x_i-μ_k)(x_i-μ_k)^TΣ_k⁻¹ - Σ_k⁻¹])
  w_k = w_k - η_w * g_w - λ_w * w_k
  μ_k = μ_k - η_μ * g_μ
  Σ_k = Σ_k - η_Σ * g_Σ + ρ * (I - Σ_k)
```

This iterative scheme is computationally explicit and leverages the analytic structure of the objective and its gradients.

## 5. Analytical and Theoretical Properties

The WRF-GS framework admits several rigorous properties:

- **Regularity**: If the Gaussian kernel is \(C^s\), then the mixture \(f\) inherits \(s\) bounded derivatives.
- **Universal Approximation**: On compact domains in \(\mathbb{R}^d\), Gaussian mixtures are dense in the space of continuous functions; for any \(\epsilon > 0\), a number \(K = O(\epsilon^{-2(d+2)})\) of Gaussians suffice for sup-norm accuracy.
- **Energy Dissipation**: Along a WFR gradient flow \(\mu_t\), the objective satisfies a dissipation law:
  \[
  \frac{d}{dt} F(f_{\mu_t}) = -\|\nabla_{\mathrm{WFR}} F(\mu_t)\|^2 \leq 0
  \]
  ensuring non-increasing risk over time.
- **Convergence Under Convexity**: If \(F\) is geodesically convex in the WFR metric, the flow converges exponentially to the minimizer.

The translation of functional gradient flows to explicit parametric ODEs establishes both theoretical guarantees and a foundation for practical algorithm design.

## 6. Illustrative One-Dimensional Example

A concrete 1D example demonstrates the sequence of computations and updates:

- Given data \(x = (0, 0.5, 1.0)\), \(y = (1.0, 2.0, 1.0)\), \(w_i \equiv 1\).
- Model: \(K = 2\) components, initialized with \(w_1 = w_2 = 1.0\), \(\mu_1 = 0.2\), \(\mu_2 = 0.8\), \(\Sigma_1 = \Sigma_2 = 0.1\).
- Kernel matrix computation, model prediction, and residuals as per the formulas above.
- Gradients computed numerically, e.g., \(g_{w_1} \approx -1.209\), \(g_{\mu_1} \approx -1.589\).
- Parameter update (e.g., with step size 0.1): \(w_1 \leftarrow 1.121\), \(\mu_1 \leftarrow 0.359\), etc.

This iteration illustrates how WRF-GS jointly tunes the Gaussian parameters to fit multi-scale, localized structure in the data using analytically determined gradients.

## 7. Significance and Interpretation

The WRF-GS scheme enables interpretability via direct correspondence between model components (particles/splats) and localized features in the input space. The identification of the weighted-residual loss with a WFR gradient flow situates Gaussian Splatting within a rigorous geometric framework, clarifying the relationship between the model, loss, and optimization routine. This perspective distinguishes the WRF-GS approach from heuristic kernel density and mixture modeling, endowing it with a unifying theory and robust analytic tools suitable for diverse approximation, estimation, and inverse problems [2511.14042].

Source: https://www.emergentmind.com/topics/weighted-residual-formulation-for-gaussian-splatting-wrf-gs