Are the same of Fréchet mean and Möbius gyrocentroid in the Poincaré ball model?

67 Views Asked by At

There are several definitions of "mean" in the Poincaré ball, an isometric hyperbolic space model. The Poincaré ball $\left( \mathbb{D}^n, \mathfrak{g}^{\mathbb{D}} \right)$ is defined as a manifold $$\mathbb{D}^n_{\kappa} = \left\{ \mathbf{x} \in \mathbb{R}^n : \left\Vert \mathbf{x} \right\Vert_2 < -\frac{1}{\kappa} \right\} ~~ (\kappa < 0)$$ with the Riemannian metric $$\mathfrak{g}^{\mathbb{D}_{\kappa}}_{\mathbf{x}} = (\lambda^{\kappa}_{\mathbf{x}})^2 \mathfrak{g}^{\mathbb{E}} $$ where $$\lambda^{\kappa}_{\mathbf{x}} := \frac{2}{1 + \kappa \left\Vert \mathbf{x} \right\Vert_2^2} $$ Here, $\kappa < 0$ is the curvature, and $\mathfrak{g}^{\mathbb{E}} = \mathbf{I}_n$ is the Euclidean metric tensor.

First type of mean is the Möbius gyrocentroid, which is transformed from Einstein gyrocentroid defined in Klein model:

Given a set of points $\left\{ \mathbf{x}_i \in \mathbb{D}_{\kappa}^n \right\}_{i=1}^{N}$ with positive weights $\left\{ w_i \in \mathbb{R}_+ \right\}_{i=1}^{N}$, the Möbius gyrocentroid of them is defined as $$\bar{\mathbf{x}} = \frac{1}{2} \otimes_{\kappa} \left( \frac{\sum_{i=1}^{N} w_i \lambda_{\mathbf{x}_i}^{\kappa} \mathbf{x}_i}{\sum_{i=1}^{N} w_i (\lambda_{\mathbf{x}_i}^{\kappa} - 1) } \right) $$

Here, "$\otimes_{\kappa}$" is the Möbius scalar-vector multiplication: $$r \otimes_{\kappa} \mathbf{x} = \frac{1}{\sqrt{|\kappa|}}\tanh \left( r \tanh^{-1} \left( \sqrt{|\kappa|} \left\Vert \mathbf{x} \right\Vert_2 \right) \right) \frac{\mathbf{x}}{\left\Vert \mathbf{x} \right\Vert_2} $$

Another kind of mean if the Fréchet mean, which is defined as the point that minimizes the Fréchet variance, i.e.

Given a set of points $\left\{ \mathbf{x}_i \in \mathbb{D}_{\kappa}^n \right\}_{i=1}^{N}$ with positive weights $\left\{ w_i \in \mathbb{R}_+ \right\}_{i=1}^{N}$, the Fréchet mean of them is defined as the solution of the following optimization problem: $$\bar{\mathbf{x}} = \arg\min_{\mathbf{z} \in \mathbb{D}_{\kappa}^n} \frac{1}{N} \sum_{i=1}^N w_i \cdot d_{\mathbb{D}}^{\kappa}(\mathbf{z}, \mathbf{x}_i)^2 $$

Here, $d_{\mathbb{D}}^{\kappa}(\cdot, \cdot)$ refers to the geodesic distance in Poincaré ball, which can be computed as $$ d_{\mathbb{D}}^{\kappa} (\mathbf{x}, \mathbf{y}) = \frac{1}{\sqrt{|\kappa|}}\cosh^{-1} \left( 1- \frac{2 \kappa \left\Vert \mathbf{x} - \mathbf{y} \right\Vert_2^2}{(1 + \kappa \left\Vert \mathbf{x} \right\Vert_2^2)(1 + \kappa \left\Vert \mathbf{y} \right\Vert_2^2)} \right) $$


I am confused about the differences between above two means. Are they the same in the Poincaré ball whenever the values of weights are ($w_i \equiv 1$ for all $i$s or $w_i$ can be arbitrary positive real number)? What are the purposes of designing two types of means?