Sparse Hanson-Wright Inequalities with Applications

Yiyun He,Ke Wang,Yizhe Zhu
2024-12-10
Abstract:We derive Hanson-Wright inequalities for the quadratic form of a random vector with sparse independent components. Specifically, we consider cases where the components of the random vector are sparse $\alpha$-subexponential random variables with $\alpha>0$. Our proof relies on a novel combinatorial approach to estimate the moments of the random quadratic form. In addition, we obtain a new Bernstein-type inequality for the sum of independent sparse $\alpha$-subexponential random variables. We present two applications with the sparse Hanson-Wright inequality: (1) Local law and complete eigenvector delocalization for sparse $\alpha$-subexponential Hermitian random matrices; (2) Concentration of the Euclidean norm for the linear transformation of a sparse $\alpha$-subexponential random vector.
Probability,Combinatorics,Statistics Theory
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to derive the Hanson - Wright inequality for quadratic forms of sparse independent component random vectors. Specifically, the paper focuses on the case where each component is a sparse α - subexponential random variable. The author estimates the moments of the random quadratic form through a novel combinatorial method and obtains a new Bernstein - type inequality for the sum of independent sparse α - subexponential random variables. ### Main Problem Analysis 1. **Hanson - Wright Inequality under Sparse Structure** - The traditional Hanson - Wright inequality is mainly for random variables with a centered sub - Gaussian distribution. This paper extends it to α - subexponential random variables under a sparse structure. - Specifically, consider the random vector \( X=(X_1,\dots,X_n)^T \), where each \( X_i = x_i\xi_i \), \( x_i\sim\text{Ber}(p_i) \) and \( \xi_i \) is a centered α - subexponential random variable. 2. **Technical Challenges** - The sparse structure introduces new complexity, especially when dealing with dependencies. The author addresses this challenge by decomposing the quadratic form \( S_{\text{off}} \) (the off - diagonal part) and establishing its concentration inequality. - In addition, the author also derives new concentration results for the diagonal part \( S_{\text{diag}} \), generalizing the existing theory from independent random variables to the sparse setting. 3. **Applications** - **Local Law and Eigenvector Delocalization**: Applied to sparse α - subexponential Hermitian random matrices, it proves the local law and complete delocalization of eigenvectors. - **Concentration of Euclidean Norm**: Studies the concentration of the Euclidean norm of sparse α - subexponential random vectors after linear transformation. ### Key Formulas - **Sparse Hanson - Wright Inequality** \[ P(|X^T A X|\geq t)\leq C'\exp\left(-C_{\alpha}\min\left\{\frac{t^2}{K^4\sum_{i,j}a_{ij}^2p_ip_j},\frac{t}{K^2\gamma_{1,\infty}},\left(\frac{t}{K^2\|A\|_{\max}}\right)^{\min\{\alpha/2,1/2\}}\right\}\right) \] where \[ \gamma_{1,\infty}:=\max_i\left\{\sum_{j\neq i}|a_{ij}|p_j,\sum_{j\neq i}|a_{ji}|p_j\right\} \] - **Concentration Inequality for the Diagonal Part** \[ P\left(\left|\sum_{i = 1}^n a_{ii}(X_i^2 - E[X_i^2])\right|\geq t\right)\leq C'\exp\left(-C_{\alpha}\min\left\{\frac{t^2}{K^4\sum_{i}a_{ii}^2p_i},\left(\frac{t}{K^2\max_i|a_{ii}|}\right)^{\min\{\alpha/2,1\}}\right\}\right) \] These formulas show the specific form of the Hanson - Wright inequality under the sparse structure and its performance in different application scenarios.