6
$\begingroup$

From Two Population We have two $p$ dimensional sampels $\mathbf{X}_1$ and $\mathbf{X}_2$ with sample size $n_1$ and $n_2$ respectively, Let We have $\mathbf{\Lambda} = \mathrm{diag}\{(\sigma_{11}^2 + \gamma\sigma_{21}^2)^{-1/2},\dots,(\sigma_{1p}^2+ \gamma\sigma^2_{2p})^{-1/2}\}$ which $\gamma$ is $\frac{n_1}{n_2}$, $\boldsymbol{\mu}_1$ and $\boldsymbol{\mu}_2$ are the $p \times 1$ mean vectors and $\Sigma_1$ and $\Sigma_2$ are the $p \times p$ Variance-Covariance Matrices of two populations. if we set $\frac{n_1}{n_1 + n_2} \rightarrow \kappa$ and $\tilde{\boldsymbol{\Sigma}} = (1-\kappa)\boldsymbol{\Sigma}_1 + \kappa\boldsymbol{\Sigma}_2$

Q: How can I show this result using the Cauchy-Schwarz inequality: $$ \frac{\mathrm{tr}(\boldsymbol{\Lambda}^2)}{p} \sqrt{\frac{\mathrm{tr}(\tilde{\boldsymbol{\Sigma}}^2)}{\mathrm{tr}((\tilde{\boldsymbol{\Lambda}\boldsymbol{\Sigma}\boldsymbol{\Lambda}})^2)}}>1$$ My Source is this article. Any Help will be appreciated.

$\endgroup$

1 Answer 1

6
+50
$\begingroup$

I am able to use Cauchy-Schwartz inequality, but I not quite getting the same result. I may have made a mistake, so here are all the steps.


Note that for positive semi-definite matrices, trace defines an inner product. That is $tr(AB) = \langle B^T, A \rangle$.

Then by Cauchy-Schwarz for $A$ and $B$ positive semi-definite symmetric matrices, $$tr(AB) = \langle B, A \rangle \leq \sqrt{\langle B,B\rangle \langle A,A \rangle } = \sqrt{tr(B^2) tr(A^2)} \,.$$

In addition, for positive semi-definite matrices, all eigenvalues are non-negative, so $tr(B^2) \leq tr(B)^2$. Also, the product of positive semi-definite matrices is also a positive semidefinite if the product is symmetric (so $\Lambda \tilde{\Sigma} \Lambda$ is PSD). Finally, for matrices $A,B,C$, $tr(ABC) = tr(CAB)$. Putting all of this together.

\begin{align*} \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{tr(\tilde{\Sigma}^2)}{tr((\Lambda \tilde{\Sigma} \Lambda)^2)}}& \geq \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{tr(\tilde{\Sigma}^2)}{tr(\Lambda \tilde{\Sigma} \Lambda)^2}}\\ &= \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{tr(\tilde{\Sigma}^2)}{tr(\Lambda^2 \tilde{\Sigma})^2}}\\ & = \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{tr(\tilde{\Sigma}^2)}{\langle \tilde{\Sigma}, \Lambda^2\rangle^2}}\\ &\geq \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{tr(\tilde{\Sigma}^2)}{ tr(\tilde{\Sigma}^2)tr(\Lambda^4)}}\\ & \geq \dfrac{tr(\Lambda^2)}{p} \sqrt{ \dfrac{1}{tr(\Lambda^2)^2}}\\ & = \dfrac{1}{p} \end{align*}

I don't get that this is greater than 1.

$\endgroup$
5
  • $\begingroup$ thanks dude! But shame on me. I just copy and paste the articles bad notation and writing. the denominator below to sqrt was written bad (in real was $tr((\Lambda\tilde{\Sigma}\Lambda)^2))$. I edited that. Is that make any difference? check that out again please. $\endgroup$
    – SirSaleh
    Feb 7, 2018 at 20:05
  • $\begingroup$ @SirSaleh That is horrible usage of notation in the paper. I show the steps now for the changes made. I still get the same result, but now I can't get the simulation to work. So I suspect their result holds, and at some point I am using a weak bound. $\endgroup$ Feb 7, 2018 at 20:32
  • 1
    $\begingroup$ Your result is the correct one: $1/p$ is the minimum, not $1$. You can establish this by computing things explicitly for identity matrices $\Sigma_i$. (You will have to work in terms of $\Lambda^{-1}$ because $\Lambda$ isn't even defined!) I suspect there are typos either in the question or the paper. $\endgroup$
    – whuber
    Feb 7, 2018 at 21:13
  • $\begingroup$ I don't get why with simulation in all different settings it's become more than 1! while p become bigger it's become more close to 1! $\endgroup$
    – SirSaleh
    Feb 7, 2018 at 21:55
  • $\begingroup$ sorry for late bounty. I've been waiting if anybody find sharper one. But looks like this is sharpest :)) $\endgroup$
    – SirSaleh
    Feb 11, 2018 at 12:43

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.