2
$\begingroup$

I'm working on a paper that requires bounding

$$\Pr\left[|\vec x^\top Q \vec y| >= t\right]$$ where $Q$ is a matrix (happens to be symmetric) and $\vec x,\vec y$ are iid real mean-zero subgaussian random vectors. As I understand it, the Hanson-Wright inequality bounds this as $$2\exp\left(-c \cdot\min\left(\frac{t^2}{k^4 ||Q||_F^2}, \frac{t}{k^2||Q||_2}\right)\right)$$ where $k$ is the subgaussian parameter and $c$ is an absolute constant.

But I'm trying to derive an actual numerical bound. Is there a paper which gives an actual value of $c$, preferably a tight one? Possibly it would be different for the two sides of the min(). Thanks.

$\endgroup$
2
  • 1
    $\begingroup$ For $Q$ sparse, lemma 3.1 in [1] shows $c=\frac{1}{4e}$ works for the diagonal terms. The off diagonal terms are a bit messy but there are some constants appearing later in the same paper. [1]: www-personal.umich.edu/~shuhengz/research/sparseHW.pdf $\endgroup$ Nov 29, 2018 at 1:23
  • $\begingroup$ @JosiahPark the link above doesn't work anymore. Here is one that does: arxiv.org/pdf/1510.05517.pdf If in the future this one doesn't work, for reference the title is "Sparse Hanson-Wright inequalities for subgaussian quadratic forms" by Shuheng Zhou. $\endgroup$
    – LSK21
    Jul 26, 2023 at 9:23

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.