Hyperbolic Centroid Schemes Overview

• Hyperbolic centroid schemes are a collection of methods that extend classical centroid definitions to hyperbolic spaces using Minkowski space formulations and isometry invariance.
• They employ computational strategies like the hyperbolic Fréchet mean and recursive Möbius centroids to balance accuracy, efficiency, and geometric consistency in non-Euclidean settings.
• These schemes are applied across geometric measure theory, hyperbolic embeddings in NLP, and high-order finite volume methods for PDEs, enhancing shock-capturing and classification performance.

Hyperbolic centroid schemes refer to a diverse set of mathematical and computational constructions that generalize classical centroid concepts and schemes—from geometric mass-centers to high-resolution numerical schemes—within the context of hyperbolic geometry, hyperbolic conservation laws, and negatively curved manifolds. These schemes are foundational for applications ranging from geometric measure theory to modern hyperbolic embedding-based machine learning and high-order finite volume methods for hyperbolic PDEs.

1. Centroids in Hyperbolic Geometry

The hyperbolic centroid, also called the mass-center in hyperbolic space, generalizes the Euclidean center of mass to non-Euclidean settings. In the hyperboloid model of $H^n$, a “material vector” $a \in \mathbb{R}^{n+1}$ with $\langle a,a \rangle < 0$ (using the Minkowski inner product) represents the point $[a] = a/\sqrt{-\langle a,a \rangle}$ in $H^n$ with mass $m_a = \sqrt{-\langle a,a \rangle}$.

The unique mass-center system for finite point sets is characterized by six Galperin–Cho–Choi axioms, including isometry invariance, partitioning, overlap, and continuity. The centroid is constructed as follows: for material points $\{(r_i, m_i)\}_{i=1}^N$, with $r_i \in H^n$ and $m_i > 0$, the ambient-space sum $A = \sum_{i=1}^N m_i r_i$ yields the centroid $c = [A] = A/\sqrt{-\langle A, A \rangle}$, with squared “centered-mass” $$m_{\rm cen}^2 = \sum m_i^2 + 2\sum_{i<j} m_i m_j \cosh d(r_i, r_j)$$, where $d(\cdot,\cdot)$ is the hyperbolic distance. This formula is unique under the stated axioms, as shown by induction and linearity in the ambient Minkowski space. Examples include the centroid of a triangle in $H^2$ and a regular tetrahedron in $H^3$ (Cho et al., 2024).

A notable generalization is the Pappus centroid theorem for hyperbolic volumes: for a “Pappus solid” swept by sections $L_t$ along a centroid-curve $c(t)$, the hyperbolic volume satisfies

$$\mathrm{Vol}(P) = \int_I m_{\rm cen}(L_t) \cos\theta(t) dt,$$

where $\theta(t)$ is the angle between $c'(t)$ and the normal to $L_t$.

2. Computational Schemes for Hyperbolic Centroids

Algorithms for hyperbolic centroids in Riemannian manifolds primarily focus on the Fréchet mean and operator models in the Poincaré ball. The hyperbolic Fréchet mean solves

$$c^* = \arg\min_{c \in D^n} \sum_{i=1}^n d_H(c, x_i)^2,$$

for points $x_1, \dots, x_n$ in the Poincaré ball $D^n$, with the distance

$$d_H(x, y) = \operatorname{arcosh} \left( 1 + 2\frac{\|x-y\|^2}{(1 - \|x\|^2)(1 - \|y\|^2)} \right).$$

There is in general no closed form solution; one employs Riemannian gradient descent, leveraging the exponential/logarithmic maps of the Poincaré model.

Multiple fast and interpretable centroid schemes have been constructed for practical use:

• Naive Möbius Centroid (NC): Recursive Möbius addition and scalar multiplication, normalized at each step.
• Linear Forward/Backward Centroid (LFC/LBC): Ordered recursive weighted midpoints; LBC iterates in reverse, capturing non-commutativity effects.
• Linear Average Centroid (LAC): The midpoint between LFC and LBC.
• Binary-Tree Centroid (BTC): Recursive divide-and-conquer midpoints via balanced binary tree traversal.

Each scheme trades between computational complexity (generally $O(nd)$) and invariance properties. LBC and LAC perform robustly for hyperbolic embeddings in text classification and are particularly competitive for short or morphologically rich documents. The naive Euclidean mean baseline, followed by projection back to $D^n$, can suffice for large-scale, speed-critical tasks (Gerek et al., 2022).

3. Hyperbolic Centroid Schemes in Text and Embedding Spaces

With the advent of hyperbolic embeddings in natural language processing, document representations require centroid computations entirely within hyperbolic spaces. Averaging word embeddings in Euclidean space is not meaningful when each element is a point in hyperbolic geometry. Using the hyperbolic Fréchet mean or the aforementioned centroid schemes preserves geometric consistency, ensuring that the resulting representation remains in the model space and better reflects the intrinsic geometry.

Empirical studies on high-dimensional datasets (e.g., 20News, WebKB4, 1150Haber, Hurriyet6c1k) with Poincaré-embedded GloVe vectors demonstrate that LBC/LAC schemes yield classification accuracies matching or exceeding the Euclidean mean baseline; BTC schemes can be less robust in imbalanced scenarios. SVM and $k$-NN classifiers further confirm the superiority of order-sensitive centroid schemes when using hyperbolic embeddings (Gerek et al., 2022).

4. Central-Upwind and CWENO Centroid Schemes for Hyperbolic PDEs

In the context of hyperbolic conservation laws, “centroid” schemes manifest as central-upwind, CWENO (Central Weighted Essentially Non-Oscillatory), and weighted compact central (WCC) schemes achieving high-order accuracy and robust shock-capturing with minimal diffusion (Dumbser et al., 2016, Shen et al., 2021, Verma et al., 2015, Chertock et al., 2022).

Centroid-based CWENO schemes construct a high-order reconstruction polynomial in each simplex cell by least-squares fitting over a “centered” stencil (i.e., around the cell centroid), blended with sectorial (directional) low-degree polynomials over face-constrained stencils. Nonlinear smoothness indicators weight the candidates to yield the final non-oscillatory polynomial. This approach minimizes stencil size, supports arbitrary mesh movement (ALE), and is optimized for parallel implementation.

Local characteristic decomposition-based central-upwind schemes further improve resolution by projecting the reconstruction into the local characteristic variables, limiting high-frequency oscillations only in genuinely non-smooth characteristic fields, and then transforming back to conservative variables. The resulting scheme achieves lower numerical diffusion, sharper discontinuity resolution, and robustness across a diverse suite of 1D and 2D gas dynamics benchmarks (Chertock et al., 2022).

Weighted Compact Central (WCC) schemes combine staggered finite volume updates with local high-order polynomial reconstructions around each cell centroid, updating all spatial derivatives using compact central differences and Cauchy–Kovalevskaya time expansions. A compact WENO-type limiter ensures non-oscillatory behavior near discontinuities.

5. Properties, Performance, and Applications

Hyperbolic centroid schemes in the geometric sense enjoy uniqueness, isometry invariance, and explicit closed-form formulas, as established axiomatically and by construction. They enable the generalization of classical geometric theorems (e.g., Pappus' centroid theorem) to hyperbolic and spherical geometries (Cho et al., 2024).

In computational settings, centroid-based CWENO, central-upwind, and WCC schemes demonstrate:

• Second to arbitrarily high-order spatial accuracy (depending on the reconstruction polynomial degree).
• Drastically reduced numerical viscosity compared to traditional central or upwind schemes.
• Resolving of sharp shocks, rarefaction waves, contact discontinuities, and small-scale turbulence.
• Uniform high-order convergence rates in smooth regions and robust non-oscillatory behavior near discontinuities.
• Efficient parallel scalability and minimal communication in large-scale unstructured mesh computations—valid up to $10^9$ degrees of freedom (Dumbser et al., 2016).

In embedding and document representation contexts, hyperbolic centroid schemes underpin improved semantic representation and classification accuracy for datasets embedded in Poincaré geometry, particularly for low-resource languages or morphologically rich texts (Gerek et al., 2022).

6. Theoretical and Practical Considerations

For geometric centroids in $H^n$, uniqueness is ensured by axioms (A1)-(A6) and the ambient Minkowski-space linearity. Notably, the explicit centroid formula depends on hyperbolic distances via the sum of inner products between material vectors, with normalization enforcing the proper embedding in $H^n$.

For computational hyperbolic centroid schemes, order of operations (e.g., in LFC vs LBC) matters due to the non-commutativity of Möbius addition, and schemes may be sensitive to data arrangement, especially in word embedding applications. In central-upwind and CWENO-numerical schemes, stencil selection, limiting strategies, and characteristic decomposition are critical for both accuracy and stability.

Schemes for hyperbolic conservation laws must satisfy suitable CFL conditions: e.g., $\mathrm{CFL}\approx0.4$ for local characteristic-decomposition based central-upwind schemes (Chertock et al., 2022); for WCC, $\mathrm{CFL}=0.4$ for $P=1$, $0.3$ for $P=2$, $0.25$ for $P\ge3$ (Shen et al., 2021).

7. Summary Table: Representative Hyperbolic Centroid Schemes and Properties

Scheme/Context	Core Principle	Domain/Geometry
Hyperbolic Mass-Center Formula (Cho et al., 2024)	Minkowski-sum, normalized	Hyperboloid model $H^n$
Fréchet Mean (Poincaré Ball) (Gerek et al., 2022)	Variational minimizer of squared distances	Poincaré ball ($D^n$)
LFC/LBC/LAC/BTC Schemes (Gerek et al., 2022)	Recursive Möbius midpoints	Poincaré ball, hyperbolic embedding
Centroid-based CWENO (Dumbser et al., 2016, Verma et al., 2015)	Polynomial least-squares around centroid, nonlinear weights	Simplicial meshes, PDEs
LCD Central-Upwind (Chertock et al., 2022)	Reconstruction via local characteristic decomposition	Cartesian, hyperbolic PDEs
Weighted Compact Central (Shen et al., 2021)	High-order polynomial, super-compact stencil, WENO limiting	Structured mesh, PDEs

These methodologies offer a unifying perspective on “centroid” constructions across hyperbolic geometry, manifold optimization, and high-resolution numerical PDE schemes, revealing both the geometric and algorithmic richness of the topic.