1
$\begingroup$

Let $M$ be a non-compact matrix Lie group and $T_e M$ its lie algebra.

Consider a point $x \in M $ and $ \triangle \in T_e M$.

To move from $x$ to a point $y \in M$ along $\triangle$, below group operation seems to be commonly used in iterative optimization on Lie groups.

$y = x \exp(\triangle)$

Is this straight forward? How can this operation be explained?

$\endgroup$

2 Answers 2

2
$\begingroup$

The answer of Theo basically says it all what the exponential is concerned, but I maybe can shed some light regarding the optimization perspective.

Let $M$ be your Lie group and suppose it is a subgroup of $\textrm{GL}_n$. Now, if you want to solve \[ \min f(x) \quad \textrm{s.t. $x \in M$} \] by an iterative method, one usually uses a so-called smooth retraction map \[ R_x: T_x M \to M. \] to replace your current approximation $x$ to $x_+ := R_x(\Delta)$. Think for example of doing a line-search $R_x(-t \textrm{ grad}_x f)$ for $t>0$ where $\textrm{ grad}_x f$ is the Riemannian gradient of $f$.

(The retraction map has to fulfill some properties, in order for this to work, likes smoothness and being a first-order approximation of the geodesics (see below).)

Due to the left (or right) action of a Lie group on itself by multiplication, the exponential mapping at the identity $\exp$ can be transported to get a retraction at $x$ as $x\exp$. As Theo already explained, $\exp$ does not need to be a global diffeomorphism, but that is not needed for optimization, since we perform updates locally.

Another typical choice for $R_x$ are the geodesics in $x$. For some metrics, the $\exp$ coincides with the geodesics (for instance, bi-invariant metrics), but not always. One can also use cheaper alternatives for $R_x$ if your are only concerned with optimization. A nice reference for this retraction-based optimization on manifolds (and so, also Lie groups) is http://press.princeton.edu/titles/8586.html .

$\endgroup$
5
$\begingroup$

If I understand your question correctly, the story is as follows. The Lie group $M$ acts on itself by (left, say) translation, and this action is transitive and faithful. From this, we can trivialize the tangent bundle to $M$, by identifying ${\rm T}_mM$ with ${\rm T}_eM$ via the differential of the "left-multiply by $m$" map. In this way, a choice of $\Delta \in {\rm T}_eM$ determines (and indeed is the same as) a right-invariant vector field $\Delta \in \Gamma({\rm T}M)$. It is complete, in the sense that even when $M$ is not compact, this vector field has arbitrary-length flow. The map $x \mapsto x\exp(\Delta)$ is "flow by time $1$" for this vector field; in particular $\exp(\Delta)$ is defined by flowing for time $1$ starting at $e$.

Fix $x \in M$. As $\Delta$ varies through ${\rm T}_eM$, the value of $x \exp(\Delta)$ smoothly varies over $M$. For small $\Delta$, the map $\Delta \mapsto x\exp(\Delta)$ is a local diffeomorphism (the derivative of this map as $\Delta = 0$ is the left-translation map ${\rm T}_e M \to {\rm T}_x M$, and so is an isomorphism). So when $y$ is close to $x$, there is a unique $\Delta$ close to $0$ with $y = x\exp \Delta$.

Note that there might be other solutions $\Delta$ to $y = x\exp \Delta$, but they necessarily have $\Delta$ "large". Note also that when $x,y$ are not close together, then there is no guarantee that a solution exists at all (even if $M$ is connected). (When $M$ is compact connected, there is necessarily a solution, but not in the noncompact case.) The standard example is $M = {\rm SL}(2,\mathbb C)$, the group of complex $(2\times 2)$-matrices with determinant $1$. The Lie algebra ${\rm T}_e M$ can be identified with the vector space $\mathfrak{sl}(2,\mathbb C)$ of complex $(2\times 2)$-matrices with trace $0$, and then the exponential map is the matrix exponential. Every traceless $(2\times 2)$ matrix with nonzero eigenvalues has distinct eigenvalues, and hence is diagonalizable; therefore its exponential is also diagonalizable. In particular, there does not exist a traceless $2\times 2$ matrix $\Delta$ with $$ \exp(\Delta) = \begin{pmatrix} -1 & 1 \\ 0 & -1 \end{pmatrix} $$ as this matrix is not diagonalizable, but its eigenvalues are not $1$ so it is not the exponential of a matrix with eigenvalues $0$.

I don't know anything about iterative optimization. Any reasonable book on Lie theory, even if it only discusses matrix groups, should contain a good explanation of exponentials.

$\endgroup$
2
  • $\begingroup$ Dear Theo Johnson-Freyd, this your answer is very interesting as usual. To test my comprehension I would ask you: In your description of the trivialization of the $TM$, 1) should not we emphasize the freeness instead of the faithfulness of the action of $M$ on itself by multiplication on the left? and should not the resulting vector field be left-invariant? $\endgroup$
    – agt
    Apr 8, 2011 at 13:45
  • $\begingroup$ @Giuseppe: I'm bad at left/right, so each time I use "left" or "right" above, flip a coin and if it comes down heads, switch the word to the other one. In any case, right, I should have said "free" rather than "faithful"; my bad. I guess I tend to conflate these words, which is an error. $\endgroup$ Apr 8, 2011 at 14:28

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.