Any relationship between rank and operator norm of a Hermitian matrix?

251 Views Asked by At

In 'Topics in Random Matrix Theory' by Terence Tao, Exercise 2.4.3 and Exercise 2.4.4 deals with the stability of the empirical spectral distribution (ESD) w.r.t. some perturbations:

Let $M_{n}$ and $N_{n}$ be sequences of random Hermitian matrix ensembles. Let $\mu$ a (deterministic) probability measure.

Exercise 2.4.3. If $\mu_{\frac{1}{\sqrt{n}} M_{n}} \rightarrow \mu$ a.s. and $\frac{1}{n^2} \| N_{n} \|_{F}^2 \rightarrow 0$ a.s., then $\mu_{ \frac{1}{\sqrt{n}} ( M_{n} + N_{n})} \rightarrow \mu$ a.s..

Exercise 2.4.4. If $\mu_{\frac{1}{\sqrt{n}} M_{n}} \rightarrow \mu$ a.s. and $\frac{1}{n} \operatorname{rank} (N_{n}) \rightarrow 0$ a.s., then $\mu_{ \frac{1}{\sqrt{n}} ( M_{n} + N_{n})} \rightarrow \mu$ a.s..

I solved Exercise 2.4.3., which was just a corollary of Lemma 2.4.3. The hint for Exercise 2.4.4 was to use the Weyl inequalities instead of the Wielandt-Hoffman inequality. But Weyl inequalities contain terms of 'operator norm', not 'rank'. Some identities I recall are $$ \| A \|_{op} \le \| A \|_{F} \le \sqrt{\operatorname{rank} (A)} \| A\|_{op}$$ for Hermitian matrix $A$, and $$ \| A \|_{op} = \sigma(A)$$ for diagonalizable matrix $A$, where $\sigma(A)$ is the spectral radius of $A$.

My questions are:

(1) Any relationship between rank and operator norm of a Hermitian matrix?

(2) Does the assumption $\frac{1}{n} \operatorname{rank}(N_{n}) \rightarrow 0$ a.s. implies $\|M_{n} \|_{op} = o(n)$?