1
$\begingroup$

Suppose that $G$ is an $n\times n$ Gaussian random matrix of i.i.d. entries $N(0,1/n)$ and $D$ is an $n\times n$ deterministic diagonal elements. I'd like to know if there have been results on the singular values of the matrix product $GD$.

There is classical result on the operator norm of $GD$, which says that $$ \|GD\|_{op} \approx \|D\|_{op} \pm \frac{\|D\|_F}{\sqrt{n}}, $$ where $\|D\|_F$ is the Frobenius norm of $D$.

I'd like to know if there are similar results on other singular values, or on the trace norm $\|GD\|_\ast$, preferably a lower bound.

$\endgroup$
2
$\begingroup$

You can rephrase your question as follows: Let $U,V$ be random (=Haar distributed) independent unitaries. You can write $G=UD_1 V$ where $D_1$ is diagonal (entries are the singular values of $G$, independent of of $U,V$, and follow the Wishart distribution). Now you ask about the eigenvalues of $(D_1VD^2V^*D_1)^{1/2}$. The limit of the empirical measure can be computed by free probability methods. It is the free multiplicative convolution of the law of $D^2$ and that of $D_1^2$, pushed forward by square-root.

$\endgroup$
  • $\begingroup$ sorry I am new to free probability. What does 'free multiplicative product' mean? Can you point me to some reading? Do you mean the S-transform of the product is the product of the S-transform of the distributions $D^2$ and $D_1^2$? $\endgroup$ – user58955 Jan 7 '17 at 13:59
  • $\begingroup$ Yes. There are many sources to red about it, for example en.wikipedia.org/wiki/… You can also read more details in wisdom.weizmann.ac.il/~zeitouni/cupbook.pdf $\endgroup$ – ofer zeitouni Jan 7 '17 at 19:12
  • $\begingroup$ Thanks! btw, do you know any non-asymptotic analysis along this? By 'non-asymptotic' I mean not considering the limiting distribution but giving a quantitative bound for large $n$. The bounds on the operator norm I had in the problem statement, which holds with high probability, is a non-asymptotic bound. Do we expect something like that for singular values or trace norms of $GD$? $\endgroup$ – user58955 Jan 8 '17 at 2:17
  • $\begingroup$ There are concentration inequalities that quantify the convergence. See again our book for an introduction and references. $\endgroup$ – ofer zeitouni Jan 8 '17 at 5:14
  • $\begingroup$ Er, I mean $D$ is a deterministic $n\times n$ matrix, where $n$ is a 'fixed' large number. There is no limiting distribution for $D$ as $n\to\infty$. $\endgroup$ – user58955 Jan 13 '17 at 1:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.