2
$\begingroup$

I need to evaluate the moment $$\mathbb{E} (AX)^n,$$ where A is an NxN Hermitian square matrix, and X is $$X=ZZ^{\ast},$$ where $Z=\mu+Y$, where $\mu$ is mean of $Z$ and $Y$ is a zero-mean complex valued matrix with independent elements, but they do not have equal variance.

Z is a NxK matrix where $K>N$.

Now, when are the moments finite as $n\to \infty$?

If it simplifies, let Y be complex Gaussian. If it simplifies, let Y be IID (equal variances).

(I will use this to find the expected inverse of $AX$)

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

For any $N\times N$ matrix $A$ we set $\DeclareMathOperator{\tr}{tr}$ $\newcommand{\bE}{\mathbb{E}}$

$$ |A|^2:=\sum_{i,j}|a_{ij}|^2 =\tr(A A^*). $$

There exists a constant $C=C(N)>0$ such that for any complex $N\times N$ matrices $A, B$ we have

$$ |AB|\leq C |A|\cdot |B|. $$

In particular, if $A_1,\dotsc, A_n$ are complex $N\times N$ matrices we have

$$ |A_1\cdots A_n|\leq C^{n-1}|A_1|\cdots |A_n|. $$

We have

$$ \Bigl\vert \bE(AX)^n\Bigr\vert\leq \bE\bigl(\;|(A(X)^n|\;\bigr) \leq C^{2n-1}|A|^n \bE\bigl(\;|X|^n\;\bigr) =C^{2n-1}|A|^n \bE\bigl(\;|Z|^{2n}\;\bigr) $$

$$ = C^{2n-1}|A|^n\bE\Bigl( \;\Bigl(\; \sum_{i,j}|z_{ij}|^2\;\Bigr)^n\;\Bigr). $$

Now the question reduces to the following. Suppose we are given independent, complex Gaussian variables $\zeta_1,\dotsc, \zeta_m$. Are the moments of $|\zeta_1|^2+\cdots +|\zeta_m|^2$ finite? The answer is yes, because the probablity

$$ p(R)=P(|\zeta_1|^2+\cdots +|\zeta_m|^2 > R^2) $$

goes to zero exponentially fast since the random vector $(\zeta_1,\dotsc,\zeta_m)$ is Gaussian.

Next use the fact that for any nonnegative random variable $X$ we have the equality

$$ \bE(X^k) =\int_0^\infty kx^{k-1} P(X>x) dx,\;\;\forall k\geq 1. $$

Improve this answer
$\endgroup$
5
  • $\begingroup$ It is not entirely clear to me... From your answer it appears as if your conclusion is: They are always finite, no matter what $A$, the variance of $Z$, the mean $\mu$ and $n$ are...But this cannot be true. Also, the constant $C$ is important. If C>1, it is not clear at all to me that the moments are finite as $n$ grows large. I can agree that for finite $n$, the moments are finite, but the question concerns the case whether $$\lim_{n\to \infty} \mathbb{E}(AX)^n=0$$ $\endgroup$
    – pierre robert
    Jan 30, 2013 at 21:24
  • $\begingroup$ The moments are finite as $n\to\infty$. They may not be bounded as $n\to\infty$. $\endgroup$
    – Liviu Nicolaescu
    Jan 30, 2013 at 22:27
  • $\begingroup$ You need to assume something about $A$. Even the case $N=1$ you can see that if $|A|>1$ the moments go to $\infty$ like $|A|^n$. $\endgroup$
    – Liviu Nicolaescu
    Jan 30, 2013 at 22:31
  • $\begingroup$ This is essentially the question. What is needed? I need to evaluate the expected inverse of $I+AX$, which I will do through a Neumann series. But the series is not always convergent, and in that case I will use a preconditioned in order to guarantee convergence of the series expansion. The problem is now that I don't really know how to design the preconditioned. $\endgroup$
    – pierre robert
    Jan 31, 2013 at 6:34
  • $\begingroup$ The problem is tricky. Just look at the case when $A$ is diagonal and you se you need to impose some conditions on $A$. My argument shows that a condition of the type $C^2|A| <1$ will do. $\endgroup$
    – Liviu Nicolaescu
    Jan 31, 2013 at 9:50

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.