3
$\begingroup$

Given a $N×M$ random complex gaussian matrix $X$ and $N×K$ random complex gaussian matrix $Y$ I'm interested in approximating the expectation expressed as: \begin{align} E[trace({(aX{X^H} + I)^{ - 1}}Y{Y^H})] \end{align} a Is a positive given variable. I know that $XX^H$ and $YY^H$ have Wishard distribution, However. I can not approximate the problem value. Thanks!

For context, this problem relates to the quality of MIMO communication link

$\endgroup$

2 Answers 2

2
$\begingroup$

I assume the matrices $X$ and $Y$ are independent. Since the trace commutes with the expectation value, and since the expectation value of the product of independent random variables is the product of expectation values, we have $$ F(a)=\mathbb{E}\bigl[{\rm tr}\,\bigl({(aX{X^H} + I)^{ - 1}}Y{Y^H}\bigr)\bigr]={\rm tr}\bigl(\,\mathbb{E}[(aX{X^H} + I)^{ - 1}]\mathbb{E}[Y{Y^H}]\bigr).$$ The second factor is simply $K$ times the unit matrix, so $$F(a)=K\,{\rm tr}\,\mathbb{E}[(aX{X^H} + I)^{ - 1}].$$ We can now again exchange trace and expectation value, to rewrite this as an integral over the eigenvalues $\mu_k$ of $XX^H$, with density $\rho(\mu)$, $$F(a)=K\int \rho(\mu)(a\mu+1)^{-1}\,d\mu.$$ The density $\rho(\mu)$ is known, for large matrix size it is the Marcenko-Pastur distribution.

For the Marcenko-Pastur distribution, so for $M\geq N\gg 1$, I find $$F(a)=\frac{K}{2a} \left(\sqrt{a^2 (M-N)^2+2 a (M+N)+1}+a (N-M)-1\right).$$

$\endgroup$
3
  • $\begingroup$ please can I employing the same approach can I used Wishart distribution?? $\endgroup$
    – hichem hb
    May 8, 2020 at 23:48
  • 1
    $\begingroup$ I am using the Wishart distribution: firstly when I use that $\mathbb{E}[YY^H]=K$ times the identity, and secondly the Marcenko-Pastur distribution is derived from the Wishart distribution. $\endgroup$ May 9, 2020 at 6:49
  • $\begingroup$ Is the answering missing a factor of min(M, N) since the trace is the summation of all eigenvalues? $\endgroup$ Nov 1, 2021 at 23:24
0
$\begingroup$

The following argument is quite similar to Carlo Beenakker's.


For simplicity, I only consider the real case. I'm also going to use different symbols for the sizes of the matrices. Let $X$ be an $n \times d$ random matrix with iid rows from $N(0,\Sigma)$ let $Y$ be a $m \times d$ random matrix independent of $X$, with iid rows from $N(0,S)$. Set $\widehat \Sigma := X^\top X/n$ and $\widehat S := Y^\top Y/m$. Let $F(a)$ be the expression you wish to evaluate and set $\lambda=1/(na)$. Note that $$ F(a) = \frac{m}{na} \mathbb E\,[\operatorname{tr} \widehat S(\widehat\Sigma + \lambda I_d)^{-1} ] $$

Then, in the limit $n,d \to \infty$ such that $d/n \to \phi \in (0,\infty)$, we have $$ \mathbb E\,[\operatorname{tr} \widehat S (\widehat\Sigma + \lambda I_d)^{-1} \mid Y] \simeq \operatorname{tr} \widehat S(\Sigma + \kappa I_d)^{-1}, $$ where $\kappa = \kappa(\lambda,n)$ is the unique nonnegative solution to the equation $\kappa - \lambda = \kappa \operatorname{df}_1(\kappa)/n$, where $$ \operatorname{df}_1(\kappa) := \operatorname{tr}\Sigma (\Sigma + \kappa I_d)^{-1}. $$ We deduce that $$ \mathbb E\,[\operatorname{tr} \widehat S(\widehat \Sigma + \lambda I_d)^{-1} ] \simeq \mathbb E\,[\operatorname{tr} \widehat S (\Sigma + \kappa I_d)^{-1}] = \mathbb E\,[\operatorname{tr}S (\Sigma + \kappa I_d)^{-1}]. $$

In your question, you have $\Sigma=S=I_d$, and so $$ \mathbb E\,[\operatorname{tr}\widehat S (\widehat \Sigma + \lambda I_d)^{-1}] \simeq \frac{d}{1+\kappa}. $$ Moreover, $\kappa$ is now given by the following well-known formula which is reminiscent of the Marchenko-Pastur distribution $$ \kappa = \frac{\lambda + \overline \phi + \sqrt{(\lambda + \overline\phi)^2 + 4\lambda}}{2}, $$ where $\overline \phi := \phi-1$.

Simplify...

$\endgroup$

Your Answer

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

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