Abstract:Motivated by the construction of confidence intervals in statistics, we study optimal configurations of $2^d-1$ lines in real projective space $RP^{d-1}$. For small $d$, we determine line sets that numerically minimize a wide variety of potential functions among all configurations of $2^d-1$ lines through the origin. Numerical experiments verify that our findings enable to assess efficiently the tightness of a bound arising from the statistical literature.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: constructing effective confidence intervals in statistics, especially uniformly valid post - selection confidence intervals. Specifically, the authors have studied how to find the optimal configuration of \(2d - 1\) lines passing through the origin in the real projective space \(\mathbb{RP}^{d - 1}\), and these line configurations can minimize a variety of potential energy functions. By finding these optimal configurations, the compactness of a bound from the statistical literature can be evaluated more effectively.
### Key problem decomposition:
1. **Statistical motivation**:
- Construct uniformly valid post - selection confidence intervals.
- Find the optimal configuration of \(2d - 1\) lines passing through the origin to minimize specific potential energy functions.
2. **Mathematical background**:
- The choices of potential energy functions include distance potential, Riesz - 1 potential and logarithmic potential.
- Study the optimal line configurations in different dimensions \(d\) and verify whether these configurations are consistent with the known best packing configurations.
3. **Application objective**:
- Verify whether the found optimal line configurations can efficiently evaluate the compactness of statistical bounds.
- Compare the performance differences between using uniformly distributed line configurations and the Monte Carlo optimization method.
### Specific problem description:
In statistical model selection, in order to construct effective confidence intervals, a function \(f_{d,r,\alpha}:D_{\leq N}\to\mathbb{R}^+\) needs to be considered, where \(\alpha\in(0, 1)\) and \(r\in\mathbb{N}^*\) are fixed parameters. For \(L\in D_{\leq N}\), the value of \(f_{d,r,\alpha}(L)\) is defined as the unique \(K > 0\) that satisfies the following condition:
\[
\mathbb{E}\left[V\left(F_{d,r}\left(K^2\cdot\max_{u\mathbb{R}=\ell\in L}\langle u, V\rangle^2\right)\right)\right]=1-\alpha,
\]
where \(F_{d,r}\) is the cumulative distribution function of the F - distribution with degrees of freedom \(d\) and \(r\), and \(V\) is a uniformly distributed random vector on \(S^{d - 1}\).
### Main contributions of the paper:
- **Theoretical analysis**: Studied the optimal configurations of \(2d - 1\) lines in different dimensions \(d\).
- **Numerical experiments**: Verified the effectiveness of the optimal line configurations through numerical experiments and compared their performance with the Monte Carlo optimization method.
- **Statistical applications**: Demonstrated the advantages of these optimal line configurations in evaluating the compactness of statistical bounds.
Through these studies, the authors hope to prove that for small dimensions \(d\), the optimal line configurations can significantly improve the accuracy and efficiency of statistical bound evaluation.