Abstract:Several new geometric quantile-based measures for multivariate dispersion, skewness, kurtosis, and spherical asymmetry are defined. These measures differ from existing measures, which use volumes and are easy to calculate. Some theoretical justification is given, followed by experiments illustrating that they are reasonable measures of these distributional characteristics and computing confidence regions with the desired coverage.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: Existing methods for measuring multivariate distribution characteristics (such as dispersion, skewness, kurtosis, and spherical asymmetry) are insufficient when dealing with non - spherically symmetric distributions, especially in the case of extreme quantiles. Specifically, the traditional volume - based geometric quantile method is complex and inaccurate in calculation.
To solve these problems, the author proposes a new method based on geometric quantiles to define the dispersion, skewness, kurtosis, and spherical asymmetry of multivariate distributions. These new methods have the following characteristics:
1. **Avoid using volume**: The new method simplifies the calculation process by avoiding the calculation of complex volumes.
2. **Applicable to extreme quantiles**: The new method can better handle extreme quantiles and maintain good performance in non - spherically symmetric distributions.
3. **Theoretical support**: The author provides a theoretical basis to prove that these new methods are reasonable and verifies their effectiveness through experiments.
4. **Calculate confidence regions**: The author also shows how to use these new methods to calculate confidence regions for hypothesis testing.
### Specific problem description
#### 1. Limitations of existing methods
- **Complex calculation**: Traditional methods rely on volume calculation, which is very difficult in high - dimensional spaces.
- **Shape mismatch**: For non - spherically symmetric distributions, the equal - quantile contours of traditional methods may not well reflect the true shape of the distribution, especially in the case of extreme quantiles.
#### 2. Advantages of the new method
- **Simplify calculation**: The new method simplifies the calculation process by avoiding volume calculation.
- **Better adaptability**: The new method performs better when dealing with extreme quantiles, especially for non - spherically symmetric distributions.
- **Theoretical and empirical support**: The new method has not only a theoretical basis but also been verified in experiments.
### Mathematical formula representation
The specific formulas of the new method are as follows:
- **Dispersion**:
\[
\delta_X^1(\beta) := \sup_{\xi \in S^{n - 1}} \| q_X(\beta \xi) - q_X(-\beta \xi) \|
\]
\[
\delta_X^2(\beta) := \frac{1}{SA(n - 1)} \int_{S^{n - 1}} \| q_X(\beta \xi) - q_X(-\beta \xi) \| d\xi, \quad \beta \in (0,1)
\]
- **Skewness**:
\[
\gamma_X^1(\beta) := \frac{\sup_{\xi \in S^{n - 1}} \| q_X(\beta \xi) + q_X(-\beta \xi) - 2m_X \|}{\delta_X^1(\beta)}
\]
\[
\gamma_X^2(\beta) := \frac{\int_{S^{n - 1}} (q_X(\beta \xi) - m_X) d\xi / SA(n - 1)}{\delta_X^2(\beta)}, \quad \beta \in (0,1)
\]
- **Kurtosis**:
\[
\kappa_X^1(\beta, \beta') := \frac{\delta_X^1(\beta')}{\delta_X^1(\beta)}
\]
\[
\kappa_X^2(\beta, \beta') := \frac{\delta_X^2(\beta')}{\delta_X^2(\beta)}, \quad 0 < \beta < \beta' < 1
\]
- **Spherical asymmetry**:
\[
\alpha_X(\beta) := \log \left( \frac{\sup_{\xi \in S^{n - 1}} \| q_X(\beta \xi) - m_X \|}{\inf_{\xi \in S^{n - 1}}} \right)