SO(3), SE(3), and the Exponential Map

The rotations article stored orientation as a quaternion and advanced it by integrating angular velocity. The spatial-algebra article packed angular and linear velocity into one twist. The exponential map is the bridge between them: it turns a constant twist, an element of a Lie algebra, into the finite rigid motion it generates, an element of the Lie group. That is exactly the operation a physics engine needs to step a pose forward without drifting off the manifold of valid orientations.

Groups, algebras, and the flow of a twist

Two matrix Lie groups run the show. $SO(3)$ is the group of rotations ($R^{\top}R = I$, $\det R = +1$). $SE(3)$ is the group of rigid motions,

$$g = \begin{bmatrix} R & p \\ 0 & 1 \end{bmatrix}, \qquad R \in SO(3),\; p \in \mathbb{R}^{3}.$$

Each group has a Lie algebra, the flat tangent space at the identity. For $SO(3)$ it is the skew-symmetric matrices $\mathfrak{so}(3)$; for $SE(3)$ it is the twists $\mathfrak{se}(3)$, written with the hat map

$$\hat{\xi} = \begin{bmatrix} [\omega]_\times & v \\ 0 & 0 \end{bmatrix}, \qquad [\omega]_\times = \begin{bmatrix} 0 & -\omega_z & \omega_y \\ \omega_z & 0 & -\omega_x \\ -\omega_y & \omega_x & 0 \end{bmatrix}.$$

The exponential is the matrix power series $\exp(A) = \sum_{k\ge 0} A^{k}/k!$. Its meaning here is dynamical: $g(t) = \exp(t\,\hat{\xi})$ is the unique solution of $\dot g = \hat{\xi}\,g$ with $g(0) = I$. So $\exp$ integrates a constant body twist over unit time: hold the velocity $\hat{\xi}$ fixed, flow for $t \in [0,1]$, and land at the finite motion $\exp(\hat{\xi})$.

SO(3): Rodrigues' formula

Write the rotation vector as $\phi = \theta\,\hat n$ with $\theta = \lVert\phi\rVert$. Because $[\phi]_\times^{3} = -\theta^{2}[\phi]_\times$, the infinite series collapses to three terms,

$$\exp\!\big([\phi]_\times\big) = I + \frac{\sin\theta}{\theta}\,[\phi]_\times + \frac{1-\cos\theta}{\theta^{2}}\,[\phi]_\times^{2}.$$

This is Rodrigues' formula, the same $\exp$ that closed the rotations article. An engine integrates $\omega$ in the flat algebra, then maps back with this formula (or the equivalent quaternion update), so orientation never leaves $SO(3)$.¹

SE(3): the screw and the left Jacobian

The full rigid-motion exponential reuses the SO(3) block and adds a translation routed through the left Jacobian $V$,

$$\exp\!\begin{bmatrix} [\phi]_\times & v \\ 0 & 0 \end{bmatrix} = \begin{bmatrix} \exp([\phi]_\times) & V\,v \\ 0 & 1 \end{bmatrix}, \qquad V = I + \frac{1-\cos\theta}{\theta^{2}}[\phi]_\times + \frac{\theta-\sin\theta}{\theta^{3}}[\phi]_\times^{2}.$$

Chasles' theorem says the result is always a screw: a rotation about some axis together with a translation along it. This is the finite-motion counterpart of the instantaneous twist from the spatial-algebra article.²

Worked example: an SE(2) twist over unit time

Drop to the plane, where the algebra is $\xi = (v_x, v_y, \omega)$ and

$$\hat{\xi} = \begin{bmatrix} 0 & -\omega & v_x \\ \omega & 0 & v_y \\ 0 & 0 & 0 \end{bmatrix}, \qquad \exp(\hat{\xi}) = \begin{bmatrix} R(\omega) & V v \\ 0 & 1 \end{bmatrix}, \quad V = \frac{1}{\omega}\begin{bmatrix} \sin\omega & -(1-\cos\omega) \\ 1-\cos\omega & \sin\omega \end{bmatrix}.$$

Take $\xi = (1,\,0,\,\tfrac{\pi}{2})$ and flow for unit time. The rotation is $\theta = \omega = \tfrac{\pi}{2} = 90^{\circ}$, and the translation is

$$p = V\begin{bmatrix} 1 \\ 0 \end{bmatrix} = \frac{1}{\pi/2}\begin{bmatrix} \sin\tfrac{\pi}{2} \\ 1-\cos\tfrac{\pi}{2} \end{bmatrix} = \frac{2}{\pi}\begin{bmatrix} 1 \\ 1 \end{bmatrix} \approx \begin{bmatrix} 0.637 \\ 0.637 \end{bmatrix}.$$

The origin does not travel in a straight line to $p$. It rides a circular arc about the instantaneous center $c = (-v_y/\omega,\; v_x/\omega) = (0,\,2/\pi)$ of radius $\lVert v\rVert/\lvert\omega\rvert = 2/\pi \approx 0.637$, sweeping a quarter turn. Send $\omega \to 0$ and $V \to I$, so $\exp(\hat{\xi})$ becomes a pure translation by $(v_x, v_y)$: the arc straightens into a line.

exp of a constant SE(2) twist is a screwinteractive

vₓ = 1.00v_y = 0.00ω = 1.57 rad

θ = ∫ ω dt

90.0°

endpoint p

(0.637, 0.637)

arc radius

0.637

pause sweep

exp(ξ) is a screw: the frame rotates by θ about the center, sweeping a circular arc from the identity to the bold endpoint frame.

Why the engine cares

Every substep, the solver has a body twist and needs the pose it produces. Using $\exp$ (in practice a small-angle rotation applied to the quaternion, plus the translation) keeps $R$ orthogonal and $\lVert q\rVert = 1$ to machine precision, whereas naively adding velocity times $dt$ to the stored numbers drifts off the manifold and has to be renormalized. The exponential map is the reason manifold-aware integration is both cheap and exact for a constant twist, and it is the same object the state estimators and controllers built on these groups differentiate through.¹

Footnotes

1. Sola, Deray, Atchuthan, A micro Lie theory for state estimation in robotics (2018). ↩ ↩²

2. Featherstone, Rigid Body Dynamics Algorithms (2008). ↩