I was looking for a simple way to find the number of samples $n$ needed to get a decent approximation to the covariance matrix $\boldsymbol{\Sigma}$. Given a random sample $\{ \mathbf{X}_1,\mathbf{X}_2, \dots ,\mathbf{X}_n\}$, the sample covariance matrix is

$$ \boldsymbol{\Sigma}_{n} = \frac{1}{n-1} \sum^{n}_{k=1} \left( \boldsymbol{\mathbf{X}}_k - \overline{\boldsymbol{\mathbf{X}}} \right)\left(\boldsymbol{\mathbf{X}}_k - \overline{\boldsymbol{\mathbf{X}}} \right)^T $$

I want to plot the number of samples $n$ (on the x-axis) vs some scalar value computed from the estimate $\boldsymbol{\Sigma}_{n}$ (on the y-axis). What quantity can I compute in a simple way to plot such a convergence graph?. What about the matrix norm $\left \| \cdot \right \|_2$, i.e. the maximum singular value of $\boldsymbol{\Sigma}_{n}$?.

  • $\begingroup$ The near-duplicate at stats.stackexchange.com/questions/59478/… focuses on the multivariate Normal distribution, but its answers can be generalized. $\endgroup$ – whuber May 8 at 14:51
  • 1
    $\begingroup$ Did you look at Wishart distribution already? $\endgroup$ – Aksakal May 8 at 15:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.