I have access to samples of some distribution with second-moment matrix $\Sigma=E[xx^T]$ and need an estimate of $\|\Sigma\|_F^2$ (which can be used to set optimal size for LMS)
We can use Frobenius norm of sample second-moment matrix $\hat{\Sigma}$ below, but this gives a biased estimator for small $B$
$$\left\|\hat{\Sigma}\right\|_F^2=\left\|\frac{1}{B}\sum_i^B x_i x_i^T\right\|_F^2$$ For instance, take standard Gaussian in $d$ dimensions, sample estimator for $B=1$ gives $d^2$ in expectation, whereas correct value is $d$. How do I unbias it for general $B$?
Alternative way to phrase the question is this.
Another idea was to use following identity and replace expectation with sample means, but it has the same bias issue.
$$\|\Sigma\|_F^2=\frac{1}{2}\left(E[\|x\|^4]-E[\|x\|^2]+2\|E[x]\|^4\right)$$
