I encounter the following problem when I study ridge regression.
Problem. Let $\{d_j\}_{j=1}^\infty$ be a sequence of positive integers. Let $\{\psi_{i,j}\}_{i,j=1}^\infty$ be a collection of vectors where $\psi_{i,j}\in\mathbb{R}^{d_j}$ and $\|\psi_{i,j}\|_2\le 1$. For positive integers $n_0, n_1$ and $n_2$ ($n_1\ge n_2$), we define \begin{equation} \begin{aligned} \Lambda_1=\sum_{i\in[n_0]}\left(\bigotimes_{j\in[n_1]}\psi_{i,j}\right)\left(\bigotimes_{j\in[n_1]}\psi_{i,j}\right)^\top+\lambda I_1,\\ \Lambda_2=\sum_{i\in[n_0]}\left(\bigotimes_{j\in[n_2]}\psi_{i,j}\right)\left(\bigotimes_{j\in[n_2]}\psi_{i,j}\right)^\top+\lambda I_2, \end{aligned} \end{equation} where $\otimes$ is the Kronecker product, $I_1$ and $I_2$ are identity matrices and $\lambda\in\mathbb{R}_+$. Let $\{\varphi_k\}_{k=1}^\infty$ be a sequence of vectors where $\varphi_k\in\mathbb{R}^{d_k}$ and $\|\varphi_k\|_2\ge 1$, and we define \begin{equation} \phi_1=\bigotimes_{k\in[n_1]} \varphi_k, \quad \phi_2=\bigotimes_{k\in[n_2]} \varphi_k. \end{equation} Then, show that the following holds \begin{equation} \phi_1^\top\Lambda_1^{-1}\phi_1\ge \phi_2^\top\Lambda_2^{-1}\phi_2. \end{equation}
I have verified it via programming, and no counterexample was found. Therefore, I believe it is probably true. However, I can only prove the case where $n_0=1$. The main idea of my proof for $n_0=1$ is the following:
- Clearly, it suffices to prove that it holds when $n_1=2$ and $n_2=1$.
- We diagonalize $\Lambda_1^{-1}$ and $\Lambda_2^{-1}$.
- By diagonalization, we can translate $\phi_1^\top\Lambda_1^{-1}\phi_1$ and $\phi_2^\top\Lambda_2^{-1}\phi_2$ into a combination of eigenvalues, and finally obtain the desired inequality.
I do not how to handle the case where $n_0>1$ since there are fundamental differences--we can not easily diagonalize the matrices. Hence, I am stuck... Any help or hint would be appreciated.
Thanks in advance.