---
title: Local Gaussian Representation in Plane Curves
url: https://www.emergentmind.com/topics/local-gaussian-representation-lgr
type: topic
---

# Local Gaussian Representation in Plane Curves

Searching arXiv for the cited paper and adjacent uses of the term to ground the article.
{"query":"2606.06505 \"A Geometric Gaussian Mixture Representation of Plane Curves\" OR \"Local Gaussian Representation\"","max_results":10}
{"query":"\"A Geometric Gaussian Mixture Representation of Plane Curves\"","max_results":5}
{"query":"\"Local Gaussian Representation\" arXiv","max_results":10}
Local Gaussian Representation (LGR) denotes, in the plane-curve setting, a probabilistic representation in which a curve is approximated by a polygonal chain and each segment is replaced by a local anisotropic Gaussian component whose mean is the segment midpoint and whose covariance encodes tangential spread and user-defined normal-direction uncertainty. In "A Geometric Gaussian Mixture Representation of Plane Curves" [2606.06505], the resulting Gaussian Mixture Model (GMM) is proposed as an analytically tractable, uncertainty-aware representation that preserves local geometry while extending a deterministic one-dimensional curve into a thin probabilistic tube.

## 1. Segment-based geometric construction

The construction starts from a polygonal approximation of a plane curve. Let the curve be approximated by vertices $\{v_i\}$ and consecutive segments $s_i$ joining $v_i$ to $v_{i+1}$. For each segment, the representation uses its Euclidean length, midpoint, unit tangent, unit normal, and a user-defined normal-direction uncertainty parameter $\sigma_i>0$ [2606.06505].

| Quantity | Definition | Role |
|---|---|---|
| $L_i$ | $L_i=\|v_{i+1}-v_i\|$ | Segment length |
| $m_i$ | $m_i=(v_i+v_{i+1})/2$ | Segment midpoint |
| $t_i$ | $t_i=(v_{i+1}-v_i)/L_i$ | Unit tangent |
| $n_i$ | $n_i=Rt_i$, $R=\begin{bmatrix}0&-1\\1&0\end{bmatrix}$ | Unit normal |
| $\sigma_i$ | $\sigma_i>0$ | Normal-direction uncertainty |

This parameterization is local in a precise geometric sense. Each segment supplies its own coordinate frame through $(t_i,n_i)$, its own scale in the tangential direction through $L_i$, and its own transverse tolerance through $\sigma_i$. The framework depends only on the chosen vertices and therefore applies to smooth, closed, open, non-regular, and self-intersecting plane curves [2606.06505].

A central feature is that discretization is user controlled. Adaptive discretization is allowed, including refinement near high curvature, singularities, or self-intersections. This makes the representation sensitive to how geometric fidelity and uncertainty are distributed along the curve.

## 2. Local random variable and Gaussian moment matching

For each segment $s_i$, the local probabilistic model introduces a random variable
$$
X_i = m_i + U_i t_i + Z_i n_i,
$$
with
$$
U_i \sim \mathrm{Uniform}(-L_i/2,L_i/2), \qquad Z_i \sim \mathcal{N}(0,\sigma_i^2), \qquad U_i \perp Z_i.
$$
The construction is therefore uniform along the tangent direction of the segment and Gaussian along the normal direction [2606.06505].

The first and second moments yield a Gaussian component by moment matching:
$$
\mu_i = \mathbb{E}[X_i]=m_i,
$$
and
$$
\Sigma_i
= \mathrm{Var}(X_i)
= \frac{L_i^2}{12}\, t_i t_i^\top + \sigma_i^2\, n_i n_i^\top.
$$

This decomposition has a direct geometric interpretation. The tangential variance $L_i^2/12$ reflects a uniform spread along the finite segment, whereas the normal variance $\sigma_i^2$ encodes transverse uncertainty chosen by the user. The covariance eigenvectors are $t_i$ and $n_i$, with eigenvalues $L_i^2/12$ and $\sigma_i^2$, respectively [2606.06505].

The construction is called a local Gaussian representation because each segment induces a Gaussian aligned with local position, local tangent, local normal, and local arc length. A common misconception is that curvature is explicit in $\Sigma_i$. It is not: curvature is reflected by changes in $t_i$ and $n_i$ across segments rather than by an explicit curvature term in an individual covariance matrix [2606.06505].

## 3. Global Gaussian mixture and weighting

The full LGR over the entire curve is the mixture
$$
p(x)=\sum_i w_i\, \mathcal{N}(x\mid \mu_i,\Sigma_i),
$$
with
$$
\mu_i=m_i, \qquad
\Sigma_i=\frac{L_i^2}{12}\, t_i t_i^\top + \sigma_i^2\, n_i n_i^\top.
$$
This GMM reconstructs the curve probabilistically by combining segment-wise anisotropic components [2606.06505].

The canonical weight choice is arc-length proportional:
$$
w_i=\frac{L_i}{\sum_j L_j},
$$
which matches uniform mass along the curve. The same source also notes that the weights may depend on application. Density-aware weighting can prioritize segments with denser sampling, and uncertainty-aware weighting can downweight large $\sigma_i$ when they indicate lower confidence or upweight them when the goal is a conservative envelope [2606.06505].

The global effect of this mixture is a thin anisotropic tube aligned with the curve. Its longitudinal spread is controlled by $L_i$, and its thickness is controlled by $\sigma_i$. This couples geometric discretization and uncertainty modeling: longer segments produce larger tangential variance, while shorter segments sharpen local components and reduce longitudinal blur.

Because the representation is a standard GMM, mixture density and moments can be computed with standard Gaussian-mixture machinery. This analytical tractability is one of the main reasons the construction is positioned as suitable for uncertainty-aware geometric modeling.

## 4. Expressivity, approximation behavior, and limitations

The representation supports smooth, closed, open, non-regular, and self-intersecting plane curves because it is defined segmentwise from vertices and does not require global injectivity or regularity [2606.06505]. This is significant for geometric data that contain cusps, isolated singularities, or repeated crossings.

Its fidelity depends jointly on discretization and uncertainty. Smaller $L_i$ reduce tangential variance and improve local sharpness; refining near high curvature improves global fidelity; and spatially varying $\sigma_i$ can encode local tolerance or safety margins. A plausible implication is that discretization and uncertainty cannot be selected independently, because the tangential variance is fixed by segment length through $L_i^2/12$.

The main approximation caveat is that the true segment distribution is not Gaussian. It is a Uniform$\times$Gaussian distribution, and the Gaussian component matches only its first and second moments. As stated in the source, this moment-matching approximation can underestimate tails or truncation effects at segment ends [2606.06505].

A second limitation is blur. Large $\sigma_i$ or coarse discretization increases smoothing, so choosing $L_i$ and $\sigma_i$ requires balancing fidelity and robustness. A third limitation arises at $\sigma_i=0$: then
$$
\Sigma_i=\frac{L_i^2}{12}\, t_i t_i^\top
$$
is rank one. For algorithms requiring $\Sigma_i^{-1}$, the source recommends regularization via
$$
\Sigma_i \leftarrow \Sigma_i + \epsilon I
$$
(Tikhonov), or treating the distribution as supported on the segment in a distributional view [2606.06505].

The paper also sketches extensions. For 3D space curves, one may use tangent, normal, and binormal frames and allow a Gaussian in the normal-binormal plane; anisotropic normal models can replace $\sigma_i^2 n_i n_i^\top$ with richer covariance structures; and higher-order models can use non-Gaussian mixtures or higher moments to better approximate the uniform-along-segment distribution [2606.06505].

## 5. Algorithmic realization and worked example

A practical LGR pipeline consists of six steps [2606.06505].

1. Discretize the curve by selecting vertices $\{v_i\}$, with adaptive refinement allowed near high curvature, singularities, or self-intersections.  
2. For each segment $s_i=(v_i,v_{i+1})$, compute
   $$
   L_i=\|v_{i+1}-v_i\|,\quad
   m_i=(v_i+v_{i+1})/2,\quad
   t_i=(v_{i+1}-v_i)/L_i,\quad
   n_i=Rt_i.
   $$
3. Choose $\sigma_i$ per segment to reflect local normal-direction uncertainty.  
4. Form Gaussian parameters by moment matching:
   $$
   \mu_i=m_i,\qquad
   \Sigma_i=\frac{L_i^2}{12}\, t_i t_i^\top + \sigma_i^2\, n_i n_i^\top.
   $$
5. Set weights $w_i$ such as $w_i\propto L_i$ and construct
   $$
   p(x)=\sum_i w_i\,\mathcal{N}(x\mid \mu_i,\Sigma_i).
   $$
6. Optionally refine sampling and $\sigma_i$ to balance fidelity, smoothness, and computational cost.

The worked example in the source uses a segment from $v_0=(0,0)$ to $v_1=(1,0)$. Then
$$
L=1,\qquad m=(0.5,0),\qquad t=(1,0),\qquad n=(0,1).
$$
Choosing $\sigma=0.1$ gives $\sigma^2=0.01$, and thus
$$
\Sigma
=\frac{1}{12}
\begin{bmatrix}
1&0\\
0&0
\end{bmatrix}
+0.01
\begin{bmatrix}
0&0\\
0&1
\end{bmatrix}
=
\begin{bmatrix}
1/12&0\\
0&0.01
\end{bmatrix}.
$$
The corresponding Gaussian component has anisotropic spread with approximately $0.2887$ standard deviation along the tangent and $0.1$ along the normal [2606.06505].

This example also clarifies an important implementation point: the covariance does not merely blur the segment isotropically. It is oriented and anisotropic by construction, with geometry entering through the frame $(t,n)$ and uncertainty entering through the distinct variances.

## 6. Applications and terminological landscape

The plane-curve LGR is proposed as particularly relevant for uncertainty-aware CAD and digital twins, probabilistic obstacle modeling in robotics, and probabilistic trajectory planning [2606.06505]. In CAD and digital twins, $\sigma_i$ can model manufacturing tolerances or wear; in robotics, curve boundaries can be inflated in the normal direction to form conservative probabilistic envelopes; and in trajectory planning, path templates and boundaries become probabilistic primitives against which risk-aware objectives may be optimized.

Within arXiv and adjacent literatures, however, the acronym **LGR** is not unique. "Local Gaussian Regression" uses the same acronym for a probabilistic, generative regression model with localizing functions and variational inference, explicitly noting that in that work LGR stands for Local Gaussian Regression, not Local Gaussian Representation [1402.0645]. In rendering, LocoGS uses a "locality-aware 3D Gaussian representation," for which the supplied description adopts LGR as shorthand for a hybrid explicit/implicit compression model over 3D Gaussian attributes [2501.05757]. In point-cloud upsampling, PU-Gaussian models each input point with an anisotropic 3D Gaussian and uses direct sampling from learned local Gaussians to generate dense point sets [2509.20207]. In online HD map construction, GaussianMap defines a learned, adaptive set of 2D Gaussian primitives on the BEV plane that carry geometry and features and are splatted into a BEV map for vectorized decoding [2606.31177].

A further branch of usage appears in econometrics, where the LGR framework of Chernozhukov et al. is described as a copula-like local parameterization of bivariate joint distributions, and later work extends it to unordered categorical outcomes through a Local Logistic Representation rather than a geometric curve model [2510.05551]. The supplied material also uses "Local Gaussian Representation" for a standard form under Gaussian local unitaries for $n$-mode Gaussian states [1311.0509], and for local Gaussian approximations used in cross-spectrum analysis of multivariate time series [1708.02495].

This terminological variation suggests that LGR is best understood as a family resemblance rather than a single trans-domain object. Across these uses, the recurring idea is local Gaussian structure with parameters tied to locality, but the operational meaning differs sharply: in plane curves it is a segment-wise geometric GMM [2606.06505], in rendering and 3D perception it is a learned Gaussian primitive representation [2501.05757; 2509.20207; 2606.31177], and in statistics or econometrics it is a local probabilistic parameterization of dependence or regression structure [1402.0645; 2510.05551; 1708.02495].

Source: https://www.emergentmind.com/topics/local-gaussian-representation-lgr