The Spectral Theorem: the Antisymmetric Case and the Purely Imaginary Spectrum
Introduction
Every real antisymmetric matrix has a purely imaginary spectrum, and its real form is a direct sum of planar rotations together with zeros. The zeros are forced by the size: the nonzero eigenvalues come in pairs $\pm bi$, and a matrix of odd size has a slot left over, which only the eigenvalue zero can fill.
The article proves that the spectrum is purely imaginary, derives the canonical form $K=Q\operatorname{diag}(0,\dots,0,b_1J,\dots,b_sJ)Q^{\mathsf T}$ with $b_i>0$, computes the rank and the corank, defines the Pfaffian and proves $\mathrm{Pf}(K)^{2}=\det K$, shows that $K^{2}$ is symmetric and negative semidefinite, proves that $\exp K$ is orthogonal, and counts the dimension of $\mathfrak{so}(n)$. It assumes Normal Matrices over $\mathbb{R}$ and the Orthogonal Spectral Theorem and The Spectral Theorem: the Symmetric Case, and it defers the exponential to the rotation articles.
The Purely Imaginary Spectrum
Proposition. If $K^{\mathsf T}=-K$ then every eigenvalue of $K$ is purely imaginary.
Proof. Let $\lambda$ be an eigenvalue with eigenvector $z\ne0$ over $\mathbb{C}$. Then $Kz=\lambda z$, and the number
$$ q=\bar z^{\mathsf T}Kz=\lambda\,\bar z^{\mathsf T}z=\lambda\|z\|^{2} $$
satisfies
$$ \bar q=z^{\mathsf T}K\bar z=-\bar z^{\mathsf T}Kz=-q, $$
because $(z^{\mathsf T}K\bar z)^{\mathsf T}=\bar z^{\mathsf T}K^{\mathsf T}z=-\bar z^{\mathsf T}Kz$. So $q$ is purely imaginary, and since $\|z\|^{2}>0$ is real, $\lambda=q/\|z\|^{2}$ is purely imaginary.
The real eigenvector argument is the real case of the same identity. If $x\in\mathbb{R}^{n}$ is a real eigenvector with real eigenvalue $\lambda$, then
$$ x^{\mathsf T}Kx=\lambda x^{\mathsf T}x, $$
and $x^{\mathsf T}Kx=0$ because $K$ is antisymmetric: $x^{\mathsf T}Kx=(x^{\mathsf T}Kx)^{\mathsf T}=x^{\mathsf T}K^{\mathsf T}x=-x^{\mathsf T}Kx$. So $\lambda=0$. Every real eigenvalue of $K$ is zero.
The two arguments together give the statement. The complex argument gives $\lambda\in i\mathbb{R}$ for every eigenvalue; the real argument shows that the only real eigenvalue is $0$. So the spectrum is a subset of $i\mathbb{R}$, and the zero eigenvalues are the only ones that are both real and imaginary.
The spectrum is a union of pairs. For a real matrix the spectrum is stable under conjugation, and $\bar{bi}=-bi$, so the nonzero eigenvalues come in pairs $\pm bi$ with equal multiplicities. This is the count $r+2s=n$ with $r$ equal to the multiplicity of $0$.
The measurement. On $100$ random antisymmetric matrices of size $2$, $3$ and $4$ the largest modulus of a real part of an eigenvalue is below $10^{-9}$ every time, that is, purely imaginary in all $300$ cases.
The Canonical Form
Theorem. For every $K\in M_n(\mathbb{R})$ with $K^{\mathsf T}=-K$ there is an orthogonal $Q$ and real numbers $b_1\ge\cdots\ge b_s>0$ with
$$ K=Q\operatorname{diag}\bigl(0,\dots,0,\,b_1J,\dots,b_sJ\bigr)Q^{\mathsf T},\qquad J=\begin{pmatrix}0&-1\\ 1&0\end{pmatrix}. $$
Proof. The matrix $K$ is normal, because $KK^{\mathsf T}=-K^{2}=K^{\mathsf T}K$. By Normal Matrices over $\mathbb{R}$ and the Orthogonal Spectral Theorem there is an orthogonal $Q$ with $Q^{\mathsf T}KQ$ block diagonal in $1\times1$ and rotation blocks. A $1\times1$ block is a real eigenvalue, hence $0$ by the proposition; a rotation block $C(a,b)$ is antisymmetric exactly when $a=0$, since $C(a,b)^{\mathsf T}=C(a,-b)$. So the blocks are zeros and $b_jJ$ with $b_j\ne0$, and the sign of $b_j$ is absorbed by conjugating with $\operatorname{diag}(1,-1)$, which sends $bJ$ to $-bJ$.
The numbers $b_j$ are the moduli of the eigenvalues. The eigenvalues of $bJ$ are $\pm bi$, so $b_j$ is the modulus of the pair, and the ordered list $b_1\ge\cdots\ge b_s$ is unique. The canonical form is therefore unique up to the orthogonal change of basis.
The form is a sum of planar rotations. Each block $b_jJ$ rotates the plane spanned by two columns of $Q$ by the angle $\pi/2$ and scales it by $b_j$; on that plane $K$ is invertible, and its inverse is $-\frac{1}{b_j}J$. The zero blocks span the kernel.
At $n=2$ the form is one block. Every antisymmetric $2\times2$ matrix is $bJ$ with $b=-K_{12}$, and the sign is removed by conjugating with $\operatorname{diag}(1,-1)$. So the space $\mathfrak{so}(2)$ is one-dimensional, spanned by $J$.
Rank and Corank
Proposition. The rank of $K$ is $2s$ and the corank is $n-2s$, the multiplicity of the zero eigenvalue.
Proof. In the canonical form the rank is the number of nonzero blocks, each of rank $2$, so $\operatorname{rk}K=2s$; the corank is $n-2s$ by the rank-nullity theorem, and the nullity is the multiplicity of $0$ because the zero blocks are exactly the kernel.
The rank is even. An antisymmetric matrix of odd size is singular, since $2s\le n$ forces $2s The corank counts the zero eigenvalues, with multiplicity, and nothing else. So it measures how far the matrix is from invertible, and it is the same number that counts the zero blocks. The determinant vanishes when $n$ is odd and is nonnegative when $n$ is even. In the canonical form the determinant is the product of the determinants of the blocks, each $b_j^{2}>0$, times the zero blocks; so $\det K=0$ for odd $n$, and $\det K=\prod_jb_j^{2}$ for even $n$, which is a square. Definition. For even $n=2m$ and antisymmetric $K$, the Pfaffian is $$
\operatorname{Pf}(K)=\sum_{\pi}\operatorname{sgn}(\pi)\prod_{r=1}^{m}K_{\pi(2r-1),\pi(2r)},
$$ the sum running over the permutations $\pi$ of $\{1,\dots,n\}$ with $\pi(1)<\pi(2)$, $\pi(3)<\pi(4)$, and so on. At $n=2$ and $n=4$ the formula is short. For $n=2$, $$
\operatorname{Pf}(K)=K_{12},
$$ and for $n=4$, $$
\operatorname{Pf}(K)=K_{12}K_{34}-K_{13}K_{24}+K_{14}K_{23},
$$ which is the sum of the three pairings of the four indices, with alternating signs. Theorem. $\operatorname{Pf}(K)^{2}=\det K$ for every antisymmetric $K$ of even size. The measurement. On a random antisymmetric $2\times2$ matrix the computed Pfaffian is $0.245590534$ and its square is $0.06031471$, equal to the determinant. On a random antisymmetric $4\times4$ matrix the Pfaffian is $0.012280035$, its square is $0.000150799$, and the determinant is $0.000150799$. In the canonical form the Pfaffian is $\pm\prod_jb_j$, which is the square root of the determinant. The Pfaffian is the square root of the determinant with a sign that is not canonical. The sign depends on the ordering of the basis: conjugating by an orthogonal $Q$ multiplies the Pfaffian by $\det Q$, so the Pfaffian is invariant under $\mathrm{SO}(n)$ and changes sign under the other component of $\mathrm{O}(n)$. Its square, the determinant, is invariant under all of $\mathrm{O}(n)$. Proposition. If $K$ is antisymmetric then $K^{2}$ is symmetric and negative semidefinite. Proof. Symmetry: $(K^{2})^{\mathsf T}=(K^{\mathsf T})^{2}=(-K)^{2}=K^{2}$. Semidefiniteness: for every $x$, $$
x^{\mathsf T}K^{2}x=-x^{\mathsf T}K^{\mathsf T}Kx=-\|Kx\|^{2}\le0 .
$$ The equality holds exactly when $Kx=0$, that is, on the kernel. Proposition. $K^{2}$ has an orthonormal eigenbasis and eigenvalues $-b_j^{2}$ and $0$. Proof. The matrix $K^{2}$ is symmetric, so The Spectral Theorem: the Symmetric Case applies. In the canonical form $K^{2}=Q\operatorname{diag}(0,\dots,0,-b_1^{2}\mathrm{I}_2,\dots,-b_s^{2}\mathrm{I}_2)Q^{\mathsf T}$, so the eigenvalues are the negative squares of the $b_j$ and the zeros. The measurement. On $100$ random antisymmetric matrices of size $2$, $3$ and $4$, the matrix $K^{2}$ is symmetric and its largest eigenvalue is nonpositive in all $300$ cases: $100$ of $100$ at each size. The square root of $-K^{2}$ is the modulus. The matrix $-K^{2}$ is symmetric and semidefinite, so it has a semidefinite square root $|K|$, and $|K|$ is the diagonal of the $b_j$ on the pair planes and zero on the kernel. On each pair plane $|K|=b_j\mathrm{I}_2$, so $|K|$ measures the size of the rotation and not its angle. Theorem. If $K$ is antisymmetric then $\exp K$ is orthogonal with determinant $1$. Proof. The transpose of the exponential is the exponential of the transpose, so $$
(\exp K)^{\mathsf T}=\exp(K^{\mathsf T})=\exp(-K)=(\exp K)^{-1},
$$ because $K$ and $-K$ commute. Hence $\exp K$ is orthogonal. Its determinant is $e^{\operatorname{tr}K}=1$, since $\operatorname{tr}K=0$ and $\det\exp K=e^{\operatorname{tr}K}$. The exponential is a rotation on each plane. On a block $bJ$ the exponential is the rotation of angle $b$, since $\exp(bJ)=\cos b\,\mathrm{I}_2+\sin b\,J$. So $\exp$ sends the canonical form of $K$ to a diagonal of rotations by the angles $b_j$, and the zeros to the identity. The measurement. On random antisymmetric matrices of size $2$, $3$ and $4$ the computed exponential satisfies $\|\exp(K)\exp(K)^{\mathsf T}-\mathrm{I}\|=0$ in every case. The rotation articles develop the map $\exp:\mathfrak{so}(n)\to\mathrm{SO}(n)$ and its image. This is the Lie algebra statement. The antisymmetric matrices are the Lie algebra $\mathfrak{so}(n)$ of the orthogonal group, and the exponential is the map that integrates it. The canonical form says that every element of $\mathfrak{so}(n)$ is conjugate to a sum of planar rotations, which is the reason the orthogonal group in even dimensions has a torus of rotations. The space has dimension $n(n-1)/2$. The entries strictly above the diagonal are free and the entries below are their negatives, so the dimension of $\mathfrak{so}(n)$ is the number of pairs, $\binom{n}{2}=n(n-1)/2$: it is $1$ at $n=2$, $3$ at $n=3$, $6$ at $n=4$ and $10$ at $n=5$. The generic element has $s=\lfloor n/2\rfloor$ pairs. The condition for a pair to collapse, $b_j=0$, is one equation, so it is a codimension-one condition and a random element has no zero eigenvalue when $n$ is even. For odd $n$ the corank is at least one by parity, and it is exactly one generically. The count of parameters is the orbit plus the moduli. The stabiliser of the canonical form in $\mathrm{O}(n)$ is $\mathrm{O}(n-2s)$ on the zero space times a copy of $\mathrm{SO}(2)$ on each pair plane, of dimension $$
\binom{n-2s}{2}+s .
$$ So the orbit of $K$ has dimension $\binom{n}{2}-\binom{n-2s}{2}-s$, and adding the $s$ moduli $b_j$ gives $\binom{n}{2}-\binom{n-2s}{2}$. The count reads $3=2+1$ at $n=3$, $6=4+2$ at $n=4$ and $10=8+2$ at $n=5$, cases where the corank is at most one and the binomial term vanishes. The zeros are constraints, not parameters. A zero eigenvalue is not a modulus: it is a condition that holds on a subset of codimension $\binom{n-2s}{2}$, which is zero only when the corank is $0$ or $1$. For corank $2$ and above the stratum of the matrix is strictly smaller than $\mathfrak{so}(n)$, and the deficit is the dimension of $\mathrm{O}(n-2s)$ acting on the zero space. An antisymmetric real matrix has a purely imaginary spectrum, because $\bar z^{\mathsf T}Kz$ is purely imaginary and equals $\lambda\|z\|^{2}$, and its only real eigenvalue is $0$, because $x^{\mathsf T}Kx=0$ for real $x$. Its canonical form is $Q\operatorname{diag}(0,\dots,0,b_1J,\dots,b_sJ)Q^{\mathsf T}$ with $b_j>0$, so the rank is $2s$ and the corank is $n-2s$, the multiplicity of the zero eigenvalue. The Pfaffian is defined for even size and its square is the determinant; numerically the identity holds at $n=2$ and $n=4$. The square $K^{2}$ is symmetric and negative semidefinite, hence orthogonally diagonalisable with eigenvalues $-b_j^{2}$ and $0$, and it is $100$ of $100$ at $n=2,3,4$. The exponential is orthogonal of determinant one and acts as a rotation by the angle $b_j$ on each plane. The space has dimension $n(n-1)/2$, the generic element has $\lfloor n/2\rfloor$ pairs, and the orbit plus the moduli account for the whole dimension.The Pfaffian
The Square $K^{2}$
The Exponential
The Dimension Count
Summary
Summary of Notation
Symbol
Meaning
$K$
a real antisymmetric matrix, $K^{\mathsf T}=-K$
$J$
the rotation $\begin{pmatrix}0&-1\\ 1&0\end{pmatrix}$
$b_j$
the moduli of the pairs, $b_j>0$
$s$
the number of pairs, so that $\operatorname{rk}K=2s$
$\operatorname{Pf}(K)$
the Pfaffian, defined for even $n$
$\mathfrak{so}(n)$
the space of antisymmetric matrices, of dimension $n(n-1)/2$
$|K|$
the semidefinite square root of $-K^{2}$
$\exp K$
the orthogonal matrix $e^{K}$, of determinant $1$
Further Reading