Refining Concentration Bounds for Gaussian Quadratic Chaos with Applications

Abstract

This paper studies concentration of measure for Gaussian quadratic chaos in the non-asymptotic regime. We begin by tightening the classic Hanson-Wright inequality, where we increase its absolute constant $κ$ from 0.125 to at least 0.1457 in the symmetric case. We present a sharper version of Laurent-Massart inequality (LMI), which increases $κ$ from the largest available value of $1-\frac{\sqrt{3}}{2}\approx 0.134$ to $\frac{9-\sqrt{17}}{32}\approx 0.152$ in the positive-semidefinite case. We develop a sequence of inequalities indexed by $m\ge1$ that involve Schatten norms of the underlying matrix. The case $m=1$ recovers HWI. The bounds undergo a phase transition in the sense that if the tail parameter is smaller than a critical threshold $τ_c$, then $m=1$ is the tightest, whereas if it is larger than $τ_c$, then $m=\infty$ is the tightest. This leads to a novel bound called the $m_\infty$ bound. All HWI, LMI and the $m_\infty$ inequality rely on Markov's inequality. By avoiding Markov's inequality, we introduce the strong and weak $χ^2$ inequalities. We provide a list of concentration bounds that depend only on the operator norm in the positive-definite case, namely, the $m_\infty$ bound, the weak $χ^2$ bound, relaxed versions of HWI and LMI, and the so-called large-deviations bound. The sharpest among these is shown to be always either the $m_\infty$ bound or the weak $χ^2$ bound. By utilizing conditioning, modified versions of $m_\infty$, $χ^2$, and HW inequalities of various orders are proposed, all of which can be significantly tighter than the original inequalities. At the expense of higher computational cost, the sharpness of these modified bounds improves dramatically as the order increases. The paper concludes with two applications in statistical signal processing and wireless communications where the modified bounds prove highly effective.