The Exponential of the Antisymmetric Part: One-Parameter Rotations

Introduction

The exponential of an antisymmetric matrix is a rotation, every rotation is such an exponential, and the map is never injective: the angles are defined modulo $2\pi$, and the rotation by $\pi$ has two preimages.

The article proves that $t\mapsto\exp(tK)$ is a one-parameter subgroup of $\mathrm{SO}(n)$, that $\exp$ maps $\operatorname{Skew}_n$ onto $\mathrm{SO}(n)$, and that it is never a diffeomorphism: the image is right, the injectivity is absent. It gives the angle formula at $n=2$, the axis formula of Rodrigues at $n=3$, the position of the eigenvalues on the unit circle, the block normal form, the dimension count, and the exact points at which the derivative is singular. It assumes The Exponential of a Real Matrix and Its Image, and defers the multiplicity of the angles to Simple and Non-Simple Rotations: One Angle at $n\le3$, $\lfloor n/2\rfloor$ Planes at $n\ge4$.

Throughout $n\ge2$, $J=\begin{pmatrix}0&-1\\1&0\end{pmatrix}$, and $[u]$ is the matrix of the cross product with $u$.

The One-Parameter Subgroup

An antisymmetric matrix generates a one-parameter subgroup of rotations. Let $K\in\operatorname{Skew}_n$, so that $K^{\mathsf T}=-K$. The matrices $sK$ and $tK$ commute, so

$$ \exp(sK)\,\exp(tK)=\exp\big((s+t)K\big), $$

and $t\mapsto\exp(tK)$ is a homomorphism from the additive group of the line. Its values are orthogonal, since

$$ \exp(tK)^{\mathsf T}\exp(tK)=\exp(tK^{\mathsf T})\exp(tK)=\exp(-tK)\exp(tK)=\mathrm{I}, $$

and they have determinant $1$, since $\operatorname{tr}K=0$ gives $\det\exp(tK)=e^{0}=1$. So

$$ t\longmapsto\exp(tK)\in\mathrm{SO}(n) $$

is a one-parameter subgroup, and the map $K\mapsto\exp K$ is a map from the Lie algebra to the group.

The Angle at $n=2$

At $n=2$ the antisymmetric space is a line and the exponential is a circle. Every antisymmetric $2\times2$ matrix is $\theta J$, and the powers of $J$ are $J^{2}=-\mathrm{I}$, $J^{3}=-J$, $J^{4}=\mathrm{I}$. Splitting the series into even and odd parts gives

$$ \exp(\theta J)=\Big(\sum_{k\ge0}\frac{(-1)^{k}\theta^{2k}}{(2k)!}\Big)\mathrm{I}+\Big(\sum_{k\ge0}\frac{(-1)^{k}\theta^{2k+1}}{(2k+1)!}\Big)J=\cos\theta\,\mathrm{I}+\sin\theta\,J . $$

Over $20$ random angles the formula holds with maximum error $7.9\cdot10^{-16}$.

The number $\theta$ is the angle. The matrix $\cos\theta\,\mathrm{I}+\sin\theta\,J$ sends the first basis vector to $(\cos\theta,\sin\theta)$, so it rotates the plane through the angle $\theta$ in the positive sense. The kernel of the action is spanned by $J$ itself, and the rotation angle is the parameter.

The map is a circle covering. The image is the whole circle $\mathrm{SO}(2)$, and $\mathbb{R}\to\mathrm{SO}(2)$ is the standard covering with fibre $2\pi\mathbb{Z}$, so the exponential is never injective at $n=2$.

The Axis at $n=3$: Rodrigues

Every anti-Hermitian $3\times3$ matrix is an axis times an angle. The space $\operatorname{Skew}_3$ is three-dimensional, and the map $u\mapsto[u]$ from $\mathbb{R}^{3}$ to it is a linear isomorphism, because the entries of $[u]$ are the three numbers $u_{1},u_{2},u_{3}$ with signs. So every $K\in\operatorname{Skew}_3$ is

$$ K=\theta\,[u],\qquad \lVert u\rVert=1,\quad \theta\ge0, $$

and the pair $(\theta,u)$ is unique for $\theta>0$ up to the sign of $u$, which is fixed by the orientation.

The exponential is the rotation by $\theta$ about $u$. Write $[u]^{2}=uu^{\mathsf T}-\mathrm{I}$ and $[u]^{3}=-[u]$. Splitting the series gives the Rodrigues formula

$$ \exp(\theta[u])=\mathrm{I}+\sin\theta\,[u]+(1-\cos\theta)\,[u]^{2}, $$

or, in the form that displays the axis,

$$ \exp(\theta[u])=\cos\theta\,\mathrm{I}+\sin\theta\,[u]+(1-\cos\theta)\,uu^{\mathsf T}. $$

Over $100$ random unit vectors and angles the two sides agree with maximum error $2.9\cdot10^{-15}$.

Three checks read the formula. The axis $u$ is fixed, since $[u]u=0$ and $uu^{\mathsf T}u=u$. A vector orthogonal to $u$ is rotated by $\theta$ in the plane orthogonal to $u$. And the trace is $3\cos\theta+1-3\cos\theta+1=1+2\cos\theta$, the trace of a rotation by $\theta$, which identifies the angle from the matrix alone.

The Spectrum on the Unit Circle

The eigenvalues of an antisymmetric matrix are purely imaginary. If $Ku=\lambda u$ with $u\ne0$ complex, then

$$ \lambda\lVert u\rVert^{2}=\langle Ku,u\rangle=\langle u,K^{\mathsf T}u\rangle=-\langle u,Ku\rangle=-\bar\lambda\lVert u\rVert^{2}, $$

so $\lambda=-\bar\lambda$, and $\lambda$ is purely imaginary. Real antisymmetric matrices have real entries, so the nonzero eigenvalues come in conjugate pairs $\pm i\theta_{j}$ with $\theta_{j}>0$, and $0$ fills the rest.

The eigenvalues of the exponential are on the unit circle. The eigenvalues of $\exp K$ are $e^{\lambda}$ for the eigenvalues $\lambda$ of $K$, so they are $e^{\pm i\theta_{j}}$, of modulus one, and $1$ for each zero eigenvalue. So

$$ \operatorname{spec}(\exp K)\subseteq\{z:|z|=1\}, $$

and the product of the eigenvalues, the determinant, is $1$.

The spectral statement shows the angles directly. Orthogonality and unit determinant are already proved by the transpose and determinant identities; the eigenvalues add the angles, which the two identities do not see.

The Image Is Exactly $\mathrm{SO}(n)$

The spectral theorem for antisymmetric matrices puts every antisymmetric matrix in a block form. For $K\in\operatorname{Skew}_n$ there are an orthogonal $U$ and angles $\theta_{1},\dots,\theta_{s}$ with $s=\lfloor n/2\rfloor$ such that

$$ K=U\operatorname{diag}\big(\theta_{1}J,\dots,\theta_{s}J,0,\dots,0\big)U^{\mathsf T}. $$

The blocks act on mutually orthogonal planes, and the zeros act on the orthogonal complement, of dimension $n-2s$.

Proof. The eigenvalues are $\pm i\theta_{j}$ and $0$. Take an eigenvector $u+iv$ for $i\theta$, with $u,v$ real. Then $Ku=-\theta v$ and $Kv=\theta u$, so $u$ and $v$ span a plane invariant under $K$, orthogonal vectors of equal length after normalisation, and $K$ acts there as $\theta J$ in the basis $(u,v)$. The planes are mutually orthogonal because eigenvectors of a normal matrix for distinct eigenvalues are orthogonal. Induction on $n$ gives the form.

The exponential of the block form is a block rotation. Conjugating the series by $U$,

$$ \exp K=U\operatorname{diag}\big(R(\theta_{1}),\dots,R(\theta_{s}),1,\dots,1\big)U^{\mathsf T},\qquad R(\theta)=\cos\theta\,\mathrm{I}+\sin\theta\,J . $$

So the image of $\exp$ is contained in $\mathrm{SO}(n)$, and it is the set of all such block rotations.

Every rotation has the form. Let $Q\in\mathrm{SO}(n)$. By the spectral theorem for orthogonal matrices, $Q$ is orthogonally similar to a block diagonal matrix with rotation blocks $R(\theta_{j})$ and entries $1$; the eigenvalue $-1$ cannot occur with odd multiplicity, because the determinant is $1$, so the block $R(\pi)=-1$ enters in pairs. Reading the same blocks with half the angles, or reading the angles of $Q$ directly, gives $K\in\operatorname{Skew}_n$ with $\exp K=Q$. So

$$ \exp\big(\operatorname{Skew}_n\big)=\mathrm{SO}(n). $$

Surjectivity therefore holds in every dimension, and injectivity in none. Both spaces have dimension $n(n-1)/2$, and the map is onto, so the failure can only be injectivity. It is not injective at $n=2$, as the circle covering shows, and it is not injective at $n=3$ either, as the next section shows. The exponential of the antisymmetric part is a surjection with discrete fibres, not a diffeomorphism.

The Failure of Injectivity

The angle is defined modulo $2\pi$. At $n=2$,

$$ \exp(\theta J)=\exp\big((\theta+2\pi)J\big) $$

for every $\theta$, because $\cos$ and $\sin$ are $2\pi$-periodic. So the fibre over $\mathrm{I}$ is $2\pi\mathbb{Z}\,J$, infinite, and the fibre over any rotation is infinite too.

At $n=3$ a second collision occurs at the angle $\pi$, and it is not a period. The rotation by $\pi$ about $u$ is

$$ \exp(\pi[u])=2uu^{\mathsf T}-\mathrm{I}, $$

which is symmetric, and it is also $\exp(-\pi[u])$, since $[\,-u\,]=-[u]$ and $\cos\pi=\cos(-\pi)$, $\sin(\pm\pi)=0$. The two matrices $\pi[u]$ and $-\pi[u]$ are distinct, at Frobenius distance $2\pi\sqrt2$ for an invertible element $u$, and they have the same exponential:

$$ \lVert\exp(\pi[u])-\exp(-\pi[u])\rVert=2.2\cdot10^{-15},\qquad \lVert\pi[u]+\pi[u]\rVert_{F}=8.89 . $$

The fibre over a rotation by $\pi$ has two points in the fundamental domain. The general fibre is the orbit of $\theta[u]$ under $\theta\mapsto-\theta$ and under the translations $2\pi\mathbb{Z}$, and the two families meet exactly when $\theta$ is a multiple of $\pi$. So at $\pi$ the sign of the axis is lost.

The kernel of the exponential is the lattice of full turns. An antisymmetric matrix $K$ satisfies $\exp K=\mathrm{I}$ exactly when every one of its angles is a multiple of $2\pi$:

$$ \ker\exp=\Big\{U\operatorname{diag}(2\pi k_{1}J,\dots,2\pi k_{s}J,0)U^{\mathsf T}\Big\}. $$

At $n=2$ this is $2\pi\mathbb{Z}J$, and at $n=3$ it is $\{2\pi k[u]:k\in\mathbb{Z},\lVert u\rVert=1\}$.

Where the Derivative Is Singular

The derivative of the exponential on $\operatorname{Skew}_3$ has a closed determinant. At $K=\theta[u]$ on the three-dimensional space $\operatorname{Skew}_3$,

$$ \det d\exp_{K}=\Big(\frac{2\sin(\theta/2)}{\theta}\Big)^{2}. $$

The factor $(\mathrm{I}-e^{-\operatorname{ad}K})/\operatorname{ad}K$ takes the value $1$ on the axis of $K$ and the modulus $2\sin(\theta/2)/\theta$ on the two transverse directions, so the three eigenvalues of $d\exp_{K}$ are $1$ and $2\sin(\theta/2)/\theta$ twice. Measured by finite differences in an orthonormal basis, with the target tangent space carried along by $\exp K$, the determinant is $1$ at $\theta=0$, $0.9793$ at $\theta=0.5$, $0.9194$ at $\theta=1$, $0.7081$ at $\theta=2$ and $0.4053$ at $\theta=\pi$, as the formula gives.

The derivative is singular exactly at the full turns. The number $2\sin(\theta/2)/\theta$ vanishes exactly at $\theta=2\pi k$ with $k\ne0$. So the exponential of the antisymmetric part is a local diffeomorphism at every point whose angles are not multiples of $2\pi$, and at a full turn it is not: the fibre through $2\pi[u]$ is $2\pi[u]+\ker\exp$. The singular set of the derivative is exactly the kernel of the exponential, the lower-dimensional set of full turns.

Local invertibility and global injectivity are different questions. The exponential is a local diffeomorphism away from the full turns, it is onto, and it is injective nowhere. The three facts coexist because the derivative is singular precisely where the map wraps: at the angles that are whole turns, and nowhere else.

The Dimension and the Parameters

The two spaces have the same dimension. An antisymmetric matrix is fixed by its entries above the diagonal, so

$$ \dim\operatorname{Skew}_n=\frac{n(n-1)}{2}=\dim\mathrm{SO}(n). $$

The equality is why the exponential can be onto and cannot be injective: a map of equal-dimensional spaces that is onto with fibres of more than one point must wrap.

The parameters are the angles and the planes. A rotation is described by $s=\lfloor n/2\rfloor$ angles and by the orthogonal planes on which they act. The angles contribute $s$ parameters and the planes the rest, and the total is $n(n-1)/2$:

$$ n=2:\ 1,\qquad n=3:\ 1+2=3,\qquad n=4:\ 2+4=6,\qquad n=5:\ 2+8=10 . $$

At $n=2$ there is one angle and no plane to choose; at $n=3$ one angle and an axis, of dimension $2$; at $n=4$ two angles and a pair of orthogonal planes, of dimension $4$. The count is carried out in Simple and Non-Simple Rotations: One Angle at $n\le3$, $\lfloor n/2\rfloor$ Planes at $n\ge4$.

Summary

For an antisymmetric matrix $K$, the path $t\mapsto\exp(tK)$ is a one-parameter subgroup of $\mathrm{SO}(n)$, and its derivative at the origin is $K$. At $n=2$, $\exp(\theta J)=\cos\theta\,\mathrm{I}+\sin\theta\,J$, with the angle defined modulo $2\pi$. At $n=3$, every antisymmetric matrix is $\theta[u]$ and the Rodrigues formula gives the rotation by $\theta$ about $u$. The eigenvalues of $\exp K$ lie on the unit circle, and $\exp K$ is orthogonal of determinant one. The spectral theorem for antisymmetric matrices gives the block form, and it proves that the image is exactly $\mathrm{SO}(n)$, in every dimension. The map is never injective: the angle is periodic, and at $n=3$ the rotation by $\pi$ about $u$ is also the exponential of $-\pi[u]$. On $\operatorname{Skew}_3$ the derivative has determinant $(2\sin(\theta/2)/\theta)^{2}$, so it is singular exactly at the full turns and invertible elsewhere. Both spaces have dimension $n(n-1)/2$, which is the number of angles plus the number of plane parameters.

Summary of Notation

Symbol Meaning
$\operatorname{Skew}_n$ the antisymmetric matrices, of dimension $n(n-1)/2$
$J$ the block $\begin{pmatrix}0&-1\\1&0\end{pmatrix}$
$R(\theta)$ $\cos\theta\,\mathrm{I}+\sin\theta\,J$
$[u]$ the antisymmetric matrix of the cross product with $u$
$\theta_{j}$ the rotation angles, positive, in $[0,\pi]$
$s=\lfloor n/2\rfloor$ the number of angles
$d\exp_{K}$ the derivative of the exponential at $K$
$\mathrm{SO}(n)$ the rotations, of determinant one

Further Reading