1
$\begingroup$

I have to show that the derivative of 'the matrix exponential' $exp: \mathbb{C}^{n\times n}\mapsto\mathbb{C}^{n\times n}$ at the zero matrix $0$ is $id_{C^{n\times n}}$, i.e. $exp(0)=id$. The above map isn't given explicitly, so maybe someone can tell me what it does to a matrix $A$? I think the obvious guess would be $$A\mapsto e^A:=\sum_{k=0}^\infty \frac{A^k}{k!}$$, but I wouldn't know how to find the derivation of such a map since there is no variable.

So far I tried considering $$Ax\mapsto e^Ax:=\sum_{k=0}^\infty \frac{(Ax)^k}{k!}$$ instead, which is a map $\mathbb{C}\rightarrow\mathbb{C}$ I suppose.

As derivative I get $Ax\mapsto Ae^{Ax}$, which gives the identity for $x=0$, but $0$ for $A=0$.

Can somebody tell me what I got wrong, and how it is true that $exp(0)=id$?

$\endgroup$
  • 2
    $\begingroup$ $A$ is the variabile. $\endgroup$ – N74 Jun 15 '16 at 18:39
4
$\begingroup$

$e^A$ is indeed $\sum_{n=0}^\infty \frac{A^k}{k!}$. To find the derivative at the zero matrix, let $B$ be an arbitrary $n \times n$ matrix which is unit in some norm, let $\epsilon>0$ be a small number, and consider

$$e^{\epsilon B}-e^0=\sum_{k=0}^\infty \frac{\epsilon^k B^k}{k!}-\sum_{k=0}^\infty \frac{0^k}{k!} = \sum_{k=1}^\infty \frac{\epsilon^k B^k}{k!}.$$

(Here by definition $A^0=I$ for any $n \times n$ matrix $A$, including the zero matrix.) Any idea what to do now? (Hint: what is the relative size of the terms?)

$\endgroup$
  • $\begingroup$ Note: we're using the Frechet derivative here $\endgroup$ – Omnomnomnom Jun 15 '16 at 18:49
  • 1
    $\begingroup$ @Omnomnomnom The space is finite dimensional, the distinction between different sorts of derivatives doesn't matter. (Or perhaps I misunderstood the point of your comment?) $\endgroup$ – Ian Jun 15 '16 at 18:51
  • 1
    $\begingroup$ @see This definition is not so different from the calculus definition. The reason that there must be some difference is that you can't divide by a vector. So instead you perturb from your starting point going a small distance in some direction and divide by the distance that you went. So here that would be $\frac{e^{\epsilon B}-e^0}{\epsilon}$. Then you send $\epsilon\to 0$. This result depends on the direction $B$ that you went in, but (assuming your function is differentiable in the first place) this dependence is linear. We call the derivative this linear map that $B$ gets plugged into. $\endgroup$ – Ian Jun 15 '16 at 19:46
  • 1
    $\begingroup$ @see You probably used essentially the same idea when you looked at gradients in multivariable calculus. The difference is that the derivative of a scalar valued function $f$ of a vector variable $x$ at a point $x_0$ is not really $\nabla f(x_0)$ but rather the map $y \mapsto \nabla f(x_0) \cdot y$, where $\cdot$ is the dot product. Similarly the derivative of a vector valued function of a vector variable is $y \mapsto J(x_0) y$ where $J$ is the Jacobian and the multiplication is ordinary matrix multiplication. $\endgroup$ – Ian Jun 15 '16 at 19:46
  • 1
    $\begingroup$ @Ian the point is that this wikipedia page is the one with the nicest explanation of the derivative, as it applies to your answer. $\endgroup$ – Omnomnomnom Jun 15 '16 at 20:31

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.