Rotors Acting on the Bivector Space: the Adjoint Action
Introduction
A rotation acts on the bivectors by $K\mapsto QKQ^{\mathsf T}$, that action preserves the Killing form, so it is an orthogonal map of the space of bivectors, and at $n=3$ it is nothing but the rotation itself acting on $\mathbb{R}^{3}$.
The bivector space is the space of antisymmetric matrices, of dimension $n(n-1)/2$. The article proves that the action of $\mathrm{SO}(n)$ is a homomorphism into $\mathrm{SO}(N)$, that the invariant form is the Killing form and is negative definite, that at $n=3$ the action is the defining action read through $\operatorname{Skew}_3\cong\mathbb{R}^{3}$, that at $n=4$ the space splits into two invariant halves, and that the orbits are classified by the angles. It assumes The Exponential of the Antisymmetric Part: One-Parameter Rotations and The Rotor as a Square Root of a Rotation: $\mathrm{Spin}(n)\to\mathrm{SO}(n)$.
Throughout $n\ge2$, $N=n(n-1)/2$, and $\langle X,Y\rangle=\operatorname{tr}(X^{\mathsf T}Y)$ is the Frobenius form.
The Bivector Space
A bivector is an antisymmetric matrix, and an antisymmetric matrix is a 2-form. The space $\operatorname{Skew}_n$ is the Lie algebra of $\mathrm{SO}(n)$ under the commutator $[X,Y]=XY-YX$. An antisymmetric matrix is the same data as an alternating bilinear form, and the pairing
$$ K\longleftrightarrow\omega_{K}(u,v)=\langle Ku,v\rangle $$
identifies $\operatorname{Skew}_n$ with $\Lambda^{2}\mathbb{R}^{n}$ and with $\Lambda^{2}(\mathbb{R}^{n})^{*}$. So
$$ \dim\operatorname{Skew}_n=\binom{n}{2}=\frac{n(n-1)}{2}=N . $$
The elementary bivector is a plane with an area. The matrix $E_{ij}=e_{i}e_{j}^{\mathsf T}-e_{j}e_{i}^{\mathsf T}$ is anti-Hermitian, and the $E_{ij}$ for $i The commutator of two bivectors is a bivector. The space is closed under the commutator, and the identity $$
[E_{ij},E_{kl}]=\delta_{jk}E_{il}-\delta_{ik}E_{jl}-\delta_{jl}E_{ik}+\delta_{il}E_{jk}
$$ shows the closure on the basis. So the bivector space carries both an inner product and a bracket, and the adjoint action preserves both. Definition. For $Q\in\mathrm{SO}(n)$ and $K\in\operatorname{Skew}_n$, put $$
\operatorname{Ad}_{Q}(K)=QKQ^{\mathsf T}.
$$ The result is anti-Hermitian, since $(QKQ^{\mathsf T})^{\mathsf T}=QK^{\mathsf T}Q^{\mathsf T}=-QKQ^{\mathsf T}$, so the map lands in $\operatorname{Skew}_n$. The action is a homomorphism. For $Q,R\in\mathrm{SO}(n)$, $$
\operatorname{Ad}_{Q}\operatorname{Ad}_{R}(K)=QRK(QR)^{\mathsf T}=\operatorname{Ad}_{QR}(K),
$$ and $\operatorname{Ad}_{\mathrm{I}}$ is the identity. So $$
\operatorname{Ad}:\ \mathrm{SO}(n)\longrightarrow\mathrm{GL}\big(\operatorname{Skew}_n\big)=\mathrm{GL}_N(\mathbb{R})
$$ is a representation of the rotation group on the bivector space, of dimension $N=n(n-1)/2$. The action on a bivector is the action on its plane. At $n=3$ the map reads $[u]\mapsto[Qu]$: the rotated bivector is the bivector of the rotated vector. Over $100$ random pairs $(Q,u)$ the identity $$
Q\,[u]\,Q^{\mathsf T}=[Qu]
$$ holds with maximum error $1.9\cdot10^{-15}$. The rotor acts by the sandwich. A rotor $R$ acts on vectors by $X\mapsto RX\widetilde{R}$, and on bivectors by $K\mapsto RKR^{-1}$, because a bivector is a product of two vectors. The two actions are the same representation, one in the Clifford algebra and one in matrices. The adjoint action preserves the commutator. For anti-Hermitian $X,Y$ and orthogonal $Q$, $$
[QXQ^{\mathsf T},QYQ^{\mathsf T}]=QXQ^{\mathsf T}QYQ^{\mathsf T}-QYQ^{\mathsf T}QXQ^{\mathsf T}=Q[X,Y]Q^{\mathsf T}.
$$ So each $\operatorname{Ad}_{Q}$ is an automorphism of the Lie algebra, which is the defining property of the adjoint representation. The action preserves the Frobenius form. Since $Q^{\mathsf T}Q=\mathrm{I}$, $$
\langle QXQ^{\mathsf T},QYQ^{\mathsf T}\rangle=\operatorname{tr}\big(QXQ^{\mathsf T}QYQ^{\mathsf T}\big)=\operatorname{tr}\big(QXYQ^{\mathsf T}\big)=\operatorname{tr}(XY)=\langle X,Y\rangle .
$$ So the representation is orthogonal for the Frobenius form, and its image lies in $\mathrm{O}(N)$. The trace form is negative definite on the anti-Hermitian space. For anti-Hermitian $K\ne0$, $$
\operatorname{tr}(K^{\mathsf T}K)=\operatorname{tr}(-K^{2})=-\operatorname{tr}(K^{2})=\sum_{ij}K_{ij}^{2}>0,
$$ so $\operatorname{tr}(K^{\mathsf T}K)$ is positive definite and $\operatorname{tr}(XY)$ is negative definite on $\operatorname{Skew}_n$, with the sign of the Killing form. The Killing form of $\mathfrak{so}(n)$ is a multiple of the trace form. For $X,Y\in\operatorname{Skew}_n$, $$
B(X,Y)=\operatorname{tr}\big(\operatorname{ad}_{X}\operatorname{ad}_{Y}\big)=(n-2)\operatorname{tr}(XY).
$$ The factor vanishes at $n=2$: the algebra is abelian and its Killing form is zero. For $n\ge3$ the factor is positive. The form is negative definite. With $\operatorname{tr}(XY)<0$ for anti-Hermitian $X,Y$ and $X\ne0$, $$
B(X,X)=(n-2)\operatorname{tr}(X^{2})=-(n-2)\sum_{ij}X_{ij}^{2}<0\qquad(n\ge3).
$$ So $-B$ is an inner product on the bivector space, and $\mathrm{SO}(n)$ acts by isometries of it. At $n=3$ and $X=[x]$, $Y=[y]$ the form is $-2\langle x,y\rangle$, the Euclidean form up to the factor. The form is invariant under the adjoint action. Since $\operatorname{Ad}_{Q}$ is an algebra automorphism, $$
B\big(\operatorname{Ad}_{Q}X,\operatorname{Ad}_{Q}Y\big)=B(X,Y),
$$ and the same identity follows from the trace computation with the factor $n-2$ carried along. Over $100$ random pairs of bivectors and $100$ random rotations the maximum error is $1.4\cdot10^{-14}$. The image lies in $\mathrm{SO}(N)$, not only in $\mathrm{O}(N)$. Each $\operatorname{Ad}_{Q}$ is orthogonal of size $N$; its determinant is continuous in $Q$ with values in $\{\pm1\}$, the group $\mathrm{SO}(n)$ is connected, and $\operatorname{Ad}_{\mathrm{I}}$ has determinant $1$. So $$
\operatorname{Ad}_{Q}\in\mathrm{SO}(N),\qquad N=\frac{n(n-1)}{2},
$$ for every rotation $Q$. A check at $n=4$ returns determinant $1$ to $10^{-15}$. The eigenvalues of the adjoint map lie on the unit circle. If $Q$ has eigenvalues $\lambda_{1},\dots,\lambda_{n}$ on the unit circle, then $\operatorname{Ad}_{Q}$ acts on $\Lambda^{2}$ with eigenvalues $\lambda_{i}\lambda_{j}$ for $i $$
\operatorname{spec}\big(\operatorname{Ad}_{Q}\big)\subseteq\{z:|z|=1\},
$$ and the check at $n=4,5$ gives maximum deviation $1.8\cdot10^{-15}$. The bivector space is a three-dimensional Euclidean space. The map $u\mapsto[u]$ is an isomorphism $\mathbb{R}^{3}\to\operatorname{Skew}_3$, and $$
\langle[u],[v]\rangle=\operatorname{tr}\big([u]^{\mathsf T}[v]\big)=-2\langle u,v\rangle .
$$ So the Frobenius form is $-2$ times the Euclidean form, and the Killing form is the same up to the factor. The commutator is the cross product. For $u,v\in\mathbb{R}^{3}$, $$
\big[[u],[v]\big]=[u\times v],
$$ so $\operatorname{Skew}_3$ with the commutator is $\mathbb{R}^{3}$ with the cross product, and the rotation group acts by its defining action: $$
\operatorname{Ad}_{Q}\big([u]\big)=[Qu].
$$ The adjoint representation is the defining representation. The matrix of $\operatorname{Ad}_{Q}$ in the basis $([e_{1}],[e_{2}],[e_{3}])$ is the matrix of $Q$, because $[Qu]$ has the coordinates $Qu$. So at $n=3$ the adjoint representation is the rotation acting on vectors, disguised as an action on planes. The bivector space splits into a self-dual and an anti-self-dual half. On $\Lambda^{2}\mathbb{R}^{4}$ the Hodge star satisfies $*^{2}=1$, so the space is the direct sum of the eigenspaces $$
\Lambda^{2}_{+}=\{\omega:*\omega=\omega\},\qquad \Lambda^{2}_{-}=\{\omega:*\omega=-\omega\},
$$ each of dimension $3$. In matrix terms, with the basis $E_{12},E_{34},E_{13},E_{24},E_{14},E_{23}$, $$
\Lambda^{2}_{\pm}=\operatorname{span}\big\{E_{12}\pm E_{34},\ E_{13}\mp E_{24},\ E_{14}\pm E_{23}\big\}.
$$ The star commutes with the rotations, so the halves are invariant. The star is built from the metric and the orientation, and a rotation preserves both, so $*\operatorname{Ad}_{Q}=\operatorname{Ad}_{Q}*$. Hence each half is carried into itself, and $$
\operatorname{Skew}_4=\Lambda^{2}_{+}\oplus\Lambda^{2}_{-},\qquad 6=3+3,
$$ with both summands preserved. The direct check confirms it: the projection of $\operatorname{Ad}_{Q}K$ on a half depends only on the projection of $K$ on that half, to $10^{-15}$. Each half is an ideal, and each is a copy of $\mathfrak{so}(3)$. A commutator of two self-dual bivectors is again self-dual, because the star commutes with the bracket, so each half is closed under the bracket and is an ideal. Each is three-dimensional with a negative definite Killing form, hence isomorphic to $\mathfrak{so}(3)$, and $$
\mathfrak{so}(4)\cong\mathfrak{so}(3)\oplus\mathfrak{so}(3).
$$ The two halves see the two angles separately. A rotation of $\mathbb{R}^{4}$ has one angle per half, and equal angles in modulus give the isoclinic case, treated in Simple and Non-Simple Rotations: One Angle at $n\le3$, $\lfloor n/2\rfloor$ Planes at $n\ge4$. The orbit of a bivector is determined by its angles. Let $K\in\operatorname{Skew}_n$ have angles $\theta_{1},\dots,\theta_{s}$ with $s=\lfloor n/2\rfloor$, in the block form of the spectral theorem for antisymmetric matrices. Two bivectors lie in the same orbit exactly when $$
\{\theta_{1},\dots,\theta_{s}\}=\{\theta'_{1},\dots,\theta'_{s}\}
$$ as multisets, zeros included. The angles are the invariants of the action, and the action is transitive on the set of bivectors with a given multiset. The dimension of a generic orbit is $N-s$. The stabiliser of a bivector with distinct nonzero angles is the torus that rotates each invariant plane separately, of dimension $s$. So the orbit of such a bivector is the homogeneous space $$
\mathrm{SO}(n)\big/\big(\mathrm{SO}(2)^{s}\times\mathrm{SO}(n-2s)\big),\qquad \dim=\frac{n(n-1)}{2}-s .
$$ At $n=3$ this is $3-1=2$, a sphere: the orbit of a nonzero bivector is the sphere of bivectors of the same area. At $n=4$ with two distinct angles it is $6-2=4$, and with equal angles it is smaller, because the stabiliser grows. Repeated angles shrink the orbit. When two angles coincide, the stabiliser contains the rotations that mix the two planes, and the orbit loses dimension. The extreme case is the zero bivector, whose orbit is a point. So the orbit is a function of the multiset of angles, and the multiset is the complete invariant. The angles are read on the exponential. The eigenvalues of $\exp K$ are $e^{\pm i\theta_{j}}$ and $1$, so the angles are the arguments of the eigenvalues, in $[0,\pi]$. Two bivectors are in the same orbit exactly when $\exp K$ and $\exp K'$ are conjugate in $\mathrm{SO}(n)$: the adjoint orbits are the fibres of the exponential up to conjugacy. The invariants hold numerically. Over $100$ random rotations at $n=4$ and $100$ random pairs of bivectors, the following hold to the stated accuracy: The bivector space is $\operatorname{Skew}_n$, of dimension $N=n(n-1)/2$, identified with the 2-forms. A rotation acts on it by $\operatorname{Ad}_{Q}(K)=QKQ^{\mathsf T}$, an algebra automorphism that preserves the Frobenius form and the Killing form $B=(n-2)\operatorname{tr}(XY)$, and lands in $\mathrm{SO}(N)$ because the form is negative definite and the rotation group is connected. At $n=3$ the space is $\mathbb{R}^{3}$ with the cross product, and the adjoint action is the defining action, $\operatorname{Ad}_{Q}([u])=[Qu]$. At $n=4$ the space splits as $\Lambda^{2}_{+}\oplus\Lambda^{2}_{-}$, of dimensions $3+3$, and each half is an ideal isomorphic to $\mathfrak{so}(3)$, so $\mathfrak{so}(4)\cong\mathfrak{so}(3)\oplus\mathfrak{so}(3)$. The orbits are classified by the multiset of angles, a generic orbit has dimension $N-s$ with $s=\lfloor n/2\rfloor$, and the eigenvalues of every adjoint map lie on the unit circle.The Adjoint Action
The Action Is an Automorphism
The Killing Form
The Image Is Orthogonal
The Case $n=3$
The Case $n=4$: Two Ideals
The Orbits
The Numbers
Summary
Summary of Notation
Symbol
Meaning
$\operatorname{Skew}_n$
the bivector space, of dimension $N=n(n-1)/2$
$E_{ij}$
the elementary bivector $e_{i}e_{j}^{\mathsf T}-e_{j}e_{i}^{\mathsf T}$
$\operatorname{Ad}_{Q}(K)=QKQ^{\mathsf T}$
the adjoint action
$B(X,Y)=(n-2)\operatorname{tr}(XY)$
the Killing form
$\langle X,Y\rangle=\operatorname{tr}(X^{\mathsf T}Y)$
the Frobenius form
$\Lambda^{2}_{\pm}$
the self-dual and anti-self-dual halves at $n=4$
$s=\lfloor n/2\rfloor$
the rank, the number of angles
$\theta_{j}$
the rotation angles of a bivector
Further Reading