1
$\begingroup$

Let $\mathbf{H}_{N,K}$ be a random matrix whose entries are i.i.d complex Gaussian random variables with variance $1$. Then, we know from the law of large number that if $N,K\rightarrow\infty$, we have $$\frac{1}{N}\mathbf{H}^\mathrm{H}\mathbf{H}\rightarrow\mathbf{I},$$ where $\mathbf{I}$ is the identity matrix. Now, I have some questions regarding this large dimension analysis: Let define $\mathbf{A}=\mathbf{H}^\mathrm{H}\mathbf{H}$. If $f$ be a continues function, is it true to write? $$f(\mathbf{A})\rightarrow f(\mathbf{\mathbb{E}[\mathbf{A}]})=f(N\mathbf{I}).$$ More over, can we write as follows? $$\mathbf{H}^{\mathrm{H}}f(\mathbf{A})\mathbf{H}\rightarrow \mathbf{H}^{\mathrm{H}}f(N\mathbf{I})\mathbf{H},$$ and for special case $f(\mathbf{A})=\mathbf{A}$, we have? $$\mathbf{H}^{\mathrm{H}}\mathbf{A}\mathbf{H}\rightarrow N\mathbf{H}^{\mathrm{H}}\mathbf{H}\rightarrow N^2\mathbf{I}.$$

$\endgroup$
2
  • $\begingroup$ I don't see how this would be consistent with Marchenko-Pastur $\endgroup$ Jan 21 '20 at 9:43
  • $\begingroup$ Your identity matrix $\mathbf I$ must be $K\times K$ and hence cannot be the limit of anything as $K\to\infty$. $\endgroup$ Jan 21 '20 at 14:19
1
$\begingroup$

Let me check this for real matrices and $f(A)={\rm tr}\,A^2$. The $N$ eigenvalues $\mu_n$ of $A$ have in the limit $N\rightarrow\infty$ at fixed $N/K=\lambda\leq 1$ the Marchenko-Pastur distribution $$\rho(\mu)=\mathbb{E}\left[\sum_{n=1}^N\delta(\mu-\mu_n)\right]=N\frac{\sqrt{\lambda_+-\mu/N}\sqrt{\mu/N-\lambda_-}}{2\pi\lambda\mu},\;\;\lambda_-<\mu/N<\lambda_+,$$ with $\lambda_\pm=(1\pm\sqrt\lambda)^2$. The function $f(A)$ tends in this limit to $$f(A)\rightarrow\int_{N\lambda_-}^{N\lambda_+}\mu^2\rho(\mu)\,d\mu=N^3(1+\lambda),$$ which differs from the answer $f(N I)=N^3$ conjectured in the OP.

$\endgroup$

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.