Abstract:We prove a generalization of the Hardy-Littlewood-Pólya rearrangement inequality to positive definite matrices. The inequality can be seen as a commutation principle in the sense of Iusem and Seeger. An important instrument in the proof is a first-order perturbation formula for a certain spectral function, which could be of independent interests. The inequality is then extended to rectangular matrices. Using our main results, we derive new inequalities for several distance-like functions encountered in various signal processing or machine learning applications.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to generalize the Hardy - Littlewood - Pólya rearrangement inequality to matrix form and explore its applications in fields such as signal processing and machine learning. Specifically, the author has proven a new class of matrix rearrangement inequalities. These inequalities not only cover the vector - form rearrangement inequalities but also extend to the cases of positive definite matrices and rectangular matrices. The following is an overview of the main contributions and content of the paper:
### Main Problems
1. **Generalize Vector Rearrangement Inequality**
- Researchers have proven that the classical Hardy - Littlewood - Pólya rearrangement inequality holds for vectors \( u, v\in\mathbb{R}^n \):
\[
\sum_{i = 1}^n u_i^{\downarrow}v_i^{\uparrow}\leq\sum_{i = 1}^n u_i v_i\leq\sum_{i = 1}^n u_i^{\downarrow}v_i^{\downarrow}
\]
where \( u^{\downarrow} \) and \( v^{\downarrow} \) denote the descending order arrangements of \( u \) and \( v \), while \( u^{\uparrow} \) and \( v^{\uparrow} \) denote the ascending order arrangements.
2. **Matrix - form Rearrangement Inequality**
- The author proposes a new class of matrix rearrangement inequalities, which are applicable to positive definite matrices \( A, B\in\mathbb{R}^{n\times n} \):
\[
\sum_{i = 1}^n f(\lambda_i(A)\lambda_{n - i + 1}(B))\leq\sum_{i = 1}^n f(\lambda_i(B^{1/2}AB^{1/2}))\leq\sum_{i = 1}^n f(\lambda_i(A)\lambda_i(B))
\]
where \( \lambda_i(\cdot) \) denotes the \( i \)-th largest eigenvalue, \( f \) is a differentiable function, and \( s\mapsto sf'(s) \) is monotonically increasing or decreasing on \( \mathbb{R}^+ \).
3. **Extension to Rectangular Matrices**
- The author further extends the above inequalities to rectangular matrices \( X, Y\in\mathbb{R}^{m\times n} \):
\[
\sum_{i = 1}^{\min\{m,n\}} f(\sigma_i(X)\sigma_{n - i + 1}(Y))\leq\sum_{i = 1}^{\min\{m,n\}} f(\sigma_i(X^{\top}Y))\leq\sum_{i = 1}^{\min\{m,n\}} f(\sigma_i(X)\sigma_i(Y))
\]
where \( \sigma_i(\cdot) \) denotes the \( i \)-th largest singular value.
### Applications
1. **Schatten Quasi - norm**
- Through the matrix rearrangement inequality, the author derives inequalities regarding the Schatten quasi - norm:
\[
\sum_{i = 1}^{\min\{m,n\}}\sigma_i^q(X)\sigma_{n - i + 1}^q(Y)\leq\|X^{\top}Y\|_q^q\leq\sum_{i = 1}^{\min\{m,n\}}\sigma_i^q(X)\sigma_i^q(Y)
\]
These inequalities have potential application values in statistical analysis and numerical algorithms such as low - rank matrix recovery.
2. **Positive Definite Matrix**