6
$\begingroup$

I have $f: U\rightarrow \mathbb{R}$, $f(X):=\operatorname{tr}(X^{-1})$, $U$ contains all matrices $X$, which are positive definite and symmetric. I want to show that $f$ is differentiable on $U$.

To do so, I have to figure out $f(X+tY)=\operatorname{tr}[(X+tY)^{-1}]=\operatorname{tr}[ (X(I-(-t)X^{-1}Y))^{-1}]$ Now it is: $(X(I-(-t)X^{-1}Y))^{-1}=\sum_{i=0}^{\infty}(-t)^{i}(X^{-1}Y)^{i}X^{-1}$. This is the point where I get stuck. The first term of the series vanishes with $\operatorname{tr}[X]$, when it comes to the derivation. But how can I fix the rest?

$\endgroup$

2 Answers 2

8
$\begingroup$

$g:\begin{array}{ll}S_n^{++}\left(\Bbb R\right)\to S_n^{++}\left(\Bbb R\right)\\ X \mapsto X^{-1}\end{array}$

$h:\begin{array}{ll}S_n^{++}\left(\Bbb R\right) \to \Bbb R\\X\mapsto \operatorname{Tr}(X)\end{array}$

$f=h\circ g$


Let $H\in M_n\left(\Bbb R\right)$ so that $I_n+H\in S_n^{++}\left(\Bbb R\right)$ and $\left\|H\right\|< 1$

$g(I_n+H)=(I_n+H)^{-1}=\sum\limits_{k=0}^{+\infty}(-H)^k=I_n-H+\sum\limits_{k=2}^{+\infty}(-H)^k$

$\left\|g(I_n+H)-(I_n-H)\right\|=\left\|\sum\limits_{k=2}^{+\infty}(-H)^k\right\|\le \sum\limits_{k=2}^{+\infty}\left\|(-H)\right\|^k=\cfrac{\|H\|^2}{1-\left\|H\right\|}$

$\cfrac{\left\|g(I_n+H)-g(I_n)-(-H))\right\|}{\|H\|}=\cfrac{\|H\|}{1-\left\|H\right\|}\underset{\left\|H\right\|\to 0}{\longrightarrow}0$

So $g$ is differentiable at $I_n$ and $D_{I_n}(H)=-H$


Let $X\in S_n^{++}\left(\Bbb R\right)$

Let $H\in M_n\left(\Bbb R\right)$ so that $X+H\in S_n^{++}\left(\Bbb R\right)$ and $\left\|H\right\|< \cfrac{1}{\|X^{-1}\|}$ (so that $\|X^{-1}H\|\le \|X^{-1}\| \|H\| <1$). Note that I am using the operator norm to get $\|X^{-1}H\|\le \|X^{-1}\| \|H\|$ (and I can do that because the space is finite dimensional so all norms are equivalent).

$\begin{array}{ll}g(X+H)&=g(X(I_n+X^{-1}H))\\&=(X(I_n+X^{-1}H))^{-1}\\&=(I_n+X^{-1}H)^{-1}X^{-1}\\&=(I_n-X^{-1}H+o(H))X^{-1}\\&=X^{-1}-X^{-1}HX^{-1}+o(H)X^{-1}\end{array}$

$\cfrac{\|g(X+H)-g(X)-(-X^{-1}HX^{-1})\|}{\|H\|}=\cfrac{\|o(H)X^{-1}\|}{\|H\|}\underset{\left\|H\right\|\to 0}{\longrightarrow}0$

So $g$ is differentiable at $X$ and $D_X(H)=-X^{-1}HX^{-1}$


$h$ is linear and hence differential, its derivative being itself.


Now just use the chain rule and you've got your result.

$\endgroup$
4
  • 1
    $\begingroup$ @ xaviermo2: Damn fine work! Great exposition! +1! More if it could be done! And now I can stop typing up my (essentially the same) answer. Way to go! $\endgroup$ Nov 28, 2013 at 20:02
  • $\begingroup$ Nice work! But is there a way to use my approach? I think there is just one little detail, i'm missing. $\endgroup$
    – zitoxas
    Nov 28, 2013 at 21:50
  • $\begingroup$ @zitoxas : $\sum_{i=0}^{\infty}(-t)^{i}(X^{-1}Y)^{i}X^{-1}=X^{-1}-tX^{-1}YX^{-1}+o(t)=f(X)+(D_Xf)(X)(Y) + o(t)$. Then you just have to use $\cfrac{\|f(X+tY)-f(X)-(D_Xf)(X)(Y)\|}{\|tY\|}=\cfrac{o(t)}{|t|\|Y\|}=\cfrac{1}{\|Y\|}o(1)\to 0$ to prove that you can actually call it $(D_Xf)(X)(Y)$. $\endgroup$
    – xavierm02
    Nov 28, 2013 at 22:09
  • $\begingroup$ @zitoxas : In the above comment, where I wrote $f$, I meant $g$. And then $(X)$ before the $(Y)$ has nothing to do here. By the way, you can compute the derivative of the composition like you were trying to. But I don't think you should. In this case, since it was linear, the result was still simple to get to. But if you have a composition of two complicated functions, trying to get the derivative of the composition right away is a very bad idea (unless of course the two functions kind of cancel each other). $\endgroup$
    – xavierm02
    Nov 28, 2013 at 22:14
1
$\begingroup$

This is a shorter argument. Less rigorous, but it's easy to follow.

$$ (X + \delta X)(X^{-1} + \delta X^{-1}) = I $$ $$ I + (\delta X) X^{-1} + X (\delta X^{-1}) + \text{(higher order terms)} = I $$ $$ \delta X^{-1} = - X^{-1} (\delta X) X^{-1} + \text{(higher order terms)}.$$

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .