Weighted Spherical Centroidal Voronoi Tessellation

• Weighted Spherical Centroidal Voronoi Tessellation (WSCVT) is a method that partitions a unit sphere into regions where each generator is the centroid and cell areas are precisely controlled using weights.
• The algorithm iteratively refines generator positions and weights via convex hull and weighted circumcenter computations to ensure isotropic, area-proportional partitioning.
• WSCVT is applied in 3D tree visualization to allocate regions proportionally to data size, effectively overcoming limitations of methods like TriSphere.

Weighted Spherical Centroidal Voronoi Tessellation (WSCVT) is a method for partitioning the surface of a unit sphere into regions with strictly controlled area fractions, where each region's generator coincides with its centroid. WSCVT provides rigorous control over cell areas via assigned weights and delivers improved isotropy compared to mesh-refinement methods such as TriSphere, making it especially suitable for applications such as area-proportional 3D tree visualization on spherical surfaces (0912.3974).

1. Formal Mathematical Definition

Let $S$ denote the unit sphere in $\mathbb{R}^3$, defined as $S = \{ x \in \mathbb{R}^3 : \|x\|_2 = 1 \}$. Given a set of $n$ generator points $P = \{p_1, ..., p_n \} \subset S$, a Voronoi cell corresponding to $p_i$ in the unweighted case is

$$V(p_i) = \{ x \in S : \|x - p_i\|_2 \leq \|x - p_j\|_2 \ \forall j \neq i \}.$$

To introduce weights, assign each generator $p_i$ a positive weight $w_i$. The weighted squared distance from $x$ to $p_i$ is

$d_w(x, p_i) = \|x - p_i\|^2_2 - w_i.$

The Weighted Spherical Voronoi Tessellation (WSVT) is the partition

$$V_w(p_i) = \{ x \in S : d_w(x, p_i) \leq d_w(x, p_j) \ \forall j \neq i \}.$$

A tessellation is centroidal if each generator coincides with the centroid $c_i$ of its region, computed as the area-weighted mean of the centroids of the spherical polygon’s constituent spherical triangles. At convergence in WSCVT, $p_i = \text{centroid}(V_w(p_i))$ for all $i$.

2. Weight Interpretation and Area Control

Each weight $w_i$ is interpreted as prescribing a desired area fraction $d_i$ for cell $V_w(p_i)$:

$d_i = \frac{w_i}{\sum_{k=1}^n w_k}$

The actual area fraction after tessellation is

$a_i = \frac{\text{Area}(V_w(p_i))}{\text{Area}(S)}$

To enforce $a_i \to d_i$, weights are iteratively updated:

$w_i \gets w_i \cdot \frac{1 + a_i - d_i}{d_i}$

with a lower bound of $\epsilon \sim 10^{-6}$ to prevent degeneracy. Cell areas thus converge to user-prescribed fractions according to the assigned generator weights.

3. Algorithmic Procedure

The WSCVT algorithm, as formalized in Larrea et al. (0912.3974), proceeds as follows:

• Initialization:
  • Distribute generators $P = \{p_1, ..., p_n\}$ on $S$ by random sampling or projection from a polyhedron.
  • Compute target area fractions $d_i$.
• Iterative Refinement (repeat until $\max_i|a_i - d_i| < \epsilon$):
  • Compute the convex hull in $\mathbb{R}^3$; its faces provide a spherical Delaunay triangulation.
  • For each triangle $(p_i,p_j,p_k)$, obtain the weighted circumcenter $x_T$ as the solution of $d_w(x, p_i) = d_w(x, p_j) = d_w(x, p_k)$, then project $x_T$ to $S$.
  • These vertices form the weighted Voronoi diagram’s dual.
  • 2. Compute Cell Areas and Centroids:
  • Subdivide $V_w(p_i)$ into spherical triangles.
  • Compute each triangle’s area from its spherical excess $E = \alpha+\beta+\gamma-\pi$, or via the alternative arctangent formula in terms of side lengths.
  • Sum triangle areas to obtain $a_i$, and calculate the centroid $c_i$ as an area-weighted sum of triangle centroids.
  • 3. Move Generators: Set $p_i \gets c_i$ for all $i$ (renormalize to $S$).
  • 4. Update Weights: Apply the multiplicative adjustment above.
• Termination: Algorithm stops when each $|a_i - d_i| < \epsilon$.
• Complexity: Each iteration is $O(n \log n)$ (dominated by convex hull computation); empirical convergence occurs in tens to hundreds of iterations.

4. Comparison with TriSphere Algorithm

Aspect	TriSphere Algorithm	WSCVT
Area use	Only perfectly uniform for $n = 20\cdot 4^i$; otherwise significant wasted surface	Uses entire sphere, area exactly controlled by weights
Complexity	$O(M)$ ($M = 20\cdot 4^i$) but $M \gg n$ when $n$ is not a power of four refinement	$O(n \log n)$ per iteration; tractable for thousands of nodes
Mesh quality	Equal-area only at special mesh sizes; otherwise arbitrary leftover triangles	Centroidal, isotropic, generator at cell center (higher quality)
Example wastage	For $n=50$: $37.5\%$ wasted; $n=1000$: $21.875\%$; $n=1500$: $70.3125\%$	Always $0\%$ waste

WSCVT ensures full use of the spherical surface and precise area allocation for each cell, overcoming the uneven-packing problem characteristic of TriSphere. TriSphere requires $n=20\cdot 4^i$ for perfect uniformity, resulting in significant area wastage for other $n$. Table 1 in the cited work quantifies this wastage. WSCVT also delivers more isotropic, centroidal regions, whereas TriSphere’s refinement mesh yields only approximate uniformity except at special cases.

5. Computational Considerations and Optimizations

WSCVT’s most computationally intensive operation is the repeated computation of the 3D convex hull for $n$ points on $S$ ($O(n\log n)$ per iteration). Weighted circumcenter calculations and Voronoi graph assembly are linear in $n$, while area and centroid computations are $O(n + E)$, for $E \approx 3n$ edges. The algorithm’s overall cost is dominated by iteration count, which depends on the tightness of $\epsilon$ and variance in weights, but is empirically modest for practical $n$.

A practical optimization is to detect and flip Delaunay edges that violate the weighted distance order. This is only a partial solution; fully robust weighted Delaunay construction for the sphere is unresolved. Numerical stability in spherical geometry (areas, centroids, circumcenters) is critical; double precision and guards for degenerate triangles are recommended.

6. Application to Spherical 3D Tree Visualization

WSCVT is particularly suited for laying out the children of a tree’s root on the surface of the outermost sphere in a 3D tree visualization framework. If a child subtree contains $S_i$ leaf nodes, assigning $w_i = S_i$ ensures that larger subtrees are proportionally allocated larger surface regions. After computing the WSCVT, each surface cell $V_w(p_i)$ defines a pyramidal region that extends inward to subsequent sphere levels. Descendent nodes are distributed on inner spheres, confined within their parent’s pyramidal sector.

For scalability, WSCVT may be restricted to top-level tree partitions, with alternatives such as TriSphere-style subdivision at deeper levels. The computation is typically performed once per structure or weight change, with results cached for interactive use.

7. Results and Practical Impact

Key results reported in (0912.3974):

• For $n=20$, WSCVT precisely matches the icosahedron’s $20$ faces (0% surface wastage).
• For $n=10$, TriSphere refinement leads to $M=80$ faces of which $62.5\%$ are unused, while WSCVT partitions the full surface.
• For $n=100$ generators with weights $w_i=1,2,\ldots,100$, every cell’s area matches the prescribed fraction within error $\epsilon = 5\times 10^{-4}$.
• WSCVT consistently achieves $0\%$ surface waste, in contrast to significant wastage in TriSphere for non-special mesh sizes.

WSCVT’s capacity to enforce area-proportional partitions while maximizing isotropy and quality provides a robust basis for high-fidelity, geometrically balanced 3D tree visualizations. Its centroidal property improves perceptual uniformity and analytical legibility of structural data encoded on spherical domains (0912.3974).