I am trying to understand a piece of the proof that the exponential map is locally one-to-one and onto in Brian Hall's Lie Groups, Lie Algebras, and Representations.

Theorem 3.42. For $\epsilon\in(0,\log 2)$, let $U_\epsilon=\{X\in M_n(\mathbb{C})|\|X\|<\epsilon\}$ and let $V_\epsilon=\exp(U_\epsilon)$. Suppose $G\subset GL(n;\mathbb{C})$ is a matrix Lie group with Lie algebra $\mathfrak{g}$. Then there exists $\epsilon\in(0,\log 2)$ such that for all $A\in V_\epsilon$, $A\in G$ if and only if $\log A\in\mathfrak{g}$.

The first step of the proof in the book is defining the map $$ \Phi(X+Y):=e^Xe^Y,\quad X\in \mathfrak{g},\ Y\in \mathfrak{g}^\perp $$ where $\mathfrak{g}^\perp$ is the orthogonal complement of $\mathfrak{g}$ in $M_n(\mathbb{C})\cong R^{2n^2}$.

The book claims that the following calculation shows the derivative of $\Phi$ at the point $0\in\mathbb{R}^{2n^2}$ is the identity:

$$\frac{d}{dt}\Phi(tX,0)\mid_{t=0}=X,\quad \frac{d}{dt}\Phi(0,tY)\mid_{t=0}=Y. \tag{0} $$

I do not understand why:

  • By definition of $\Phi$, it has only one variable in $M_n(\mathbb{C})$. What does the expression $\Phi(tX,0)$ mean?
  • Can one show by definition that the derivative of $\Phi$ at the point $0\in\mathbb{R}^{2n^2}$ is indeed the identity: $$ \lim_{t\to 0}\frac{1}{t}\bigg(\Phi(0+tZ)-\Phi(0)\bigg)=Z\,?\tag{1} $$

  • How does (0) imply (1)?

Following the definition of $\Phi$, (1) is equivalent to $$ \lim_{t\to 0}\frac{1}{t}\bigg(e^{tX}e^{tY}-I\bigg)=X+Y,\quad X\in\mathfrak{g},\ Y\in\mathfrak{g}^\perp.\tag{2} $$ But I fail to see how this is true.

  • $\begingroup$ By definition of the exponential functions, $$ e^{tX}e^{tY}= \big(I+tX+\frac{1}{2}t^2X^2+\frac{1}{3!}t^3X^3+\cdots\big) \big(I+tY+\frac{1}{2}t^2Y^2+\frac{1}{3!}t^3Y^3+\cdots\big) =I+t(X+Y)+O(t^2)\,, $$ which implies (2). But still I can't tell why (0) implies (1). $\endgroup$ – user6 Mar 2 at 23:33
  • 1
    $\begingroup$ It looks like they meant to write $\Phi(X,Y)$, not $\Phi(X+Y)$. Since one has a direct-sum decomposition $M_n(\Bbb C) = \mathfrak g\oplus \mathfrak g^\perp$, there's not much difference, as $(X,Y)$ uniquely determines $X+Y$ and vice versa. $\endgroup$ – Ted Shifrin Mar 3 at 0:31
  • $\begingroup$ @TedShifrin: That makes sense. Thanks! Do you have an idea how (0) implies that the derivative of $\Phi$ at $0$ is the identity map? It seems that (0) only proves two "directions". $\endgroup$ – user6 Mar 3 at 0:44
  • 1
    $\begingroup$ The matrix of the derivative (using the map as I said it) will be the identity if you choose any basis for $\mathfrak g$ and any basis for $\mathfrak g^\perp$. Or just use the direct sum decomposition ... $\endgroup$ – Ted Shifrin Mar 3 at 0:46
  • $\begingroup$ The exponential is not locally injective in general, by an argument of Dixmier, see math.stackexchange.com/a/1592257/35400. It's only locally injective around 0. $\endgroup$ – YCor Mar 3 at 20:51

Thanks to Ted's comment, I think I have an answer to my questions.

  • By context, $(tX,0)$ should be understood as an element in $\mathfrak{g}$ while $M_n(\mathbb{C})=\mathfrak{g}\oplus \mathfrak{g}^\perp$.
  • The function $\Phi$ is smooth by its definition and thus $$ \Phi'(0)(Z)=\lim_{t\to 0}\frac{\Phi(0+tZ)-\Phi(0)}{t}\,, $$ where the right hand size is the directional derivative of $\Phi$ at $0$ in the direction $Z$. On the other hand, if one writes $Z$ as $X+Y$ where $X\in\mathfrak{g}$ and $Y\in\mathfrak{g}^\perp$, then (0) implies that $$ \Phi'(0)(Z)=\Phi'(0)(X)+\Phi'(0)(Y)=X+Y=Z. $$

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.