0
$\begingroup$

Suppose $\Sigma=\operatorname{diag}(h)$ where $h=(1^{-p},2^{-p},3^{-p},\ldots,d^{-p})$ and $p> 1$

$X$ is a matrix with $b$ rows sampled independently from $\operatorname{Normal}(0,\Sigma)$

Suppose $d$ is large. How does $\left\|\frac{1}{b}XX^T\right\|_{\operatorname{op}}$ depend on $b$?

For $p=1,d=30$, the formula below gives a nice fit to the average value observed in simulation, can it be improved/generalized?

$$\left\|\frac{1}{b}XX^T\right\|_{\operatorname{op}}\approx\frac{1}{b}\left(\|h\|_1+(1+b)\|h\|_\infty\right)$$

enter image description here Notebook

The formula comes from analogy with Frobenius norm where Wick's theorem can be applied.

$\endgroup$
2
  • $\begingroup$ What is the norm considered? Frobenius norm? Operator norm associated to canonical Euclidean norm on $\mathbb{R}^b$? It should be precised. $\endgroup$ Commented Mar 16, 2023 at 8:02
  • $\begingroup$ Operator norm, edited question to clarify $\endgroup$ Commented Mar 16, 2023 at 8:05

0

You must log in to answer this question.