Eigenvalue Decomposition and Eigenbasis of a Quaternion
Introduction
The preceding article, Spectral Regularity and Degeneracy in the Quaternion Algebra, found the complex eigenvalues of a quaternion: for $\tilde q=q_t+\vec q$ they are $\lambda_\pm=q_t\pm i|\vec q|$, each of multiplicity two when $\vec q\ne0$, and the single value $q_t$ when $\vec q=0$. This article constructs the eigenvectors. For a non-real quaternion the centraliser $\mathbb R[\tilde q]$ is a copy of $\mathbb C$, the vector part defines a complex structure $J_{\tilde q}=\vec q/|\vec q|$, and $\mathbb H$ becomes a two-dimensional complex vector space. The complexified regular representation splits into two eigenspaces of complex dimension two, with bases $\pi_+,\pi_+j$ and $\pi_-,\pi_-j$; in that basis $L_{\tilde q}$ is $\operatorname{diag}(\lambda_+,\lambda_+,\lambda_-,\lambda_-)$. Over $\mathbb R$ the operator is two copies of a rotation-scaling block, and the scalar case is the limit in which the two eigenlines coalesce. The article closes with the polar decomposition and with the left, right and S-spectra of the quaternion spectrum literature.
The algebra is Quaternion Algebra; the norm is Norm and Invertibility in the Quaternion Algebra; the spectral criterion is Spectral Regularity and Degeneracy in the Quaternion Algebra; the quaternionic matrix theory is Quaternion Spectral Theory; the sphere of imaginary units is Square Roots of a General Element in the Quaternion Algebra.
Notation. $\tilde q=q_t+\vec q$, $\vec q^2=-|\vec q|^2e_t$, $s=|\vec q|$, $\tilde q^{\natural}=q_t-\vec q$, $\langle\tilde q\rangle_{\natural}=q_t^2+s^2$, $|\tilde q|=\sqrt{\langle\tilde q\rangle_{\natural}}$. The complexification is $\mathbb H_{\mathbb C}=\mathbb H\otimes_{\mathbb R}\mathbb C$.
The Centraliser and the Complex Structure
Definition. The centraliser of $\tilde q$ is $C(\tilde q)=\{x\in\mathbb H:x\tilde q=\tilde q x\}$.
Theorem. If $\tilde q\notin\mathbb R$, then $$C(\tilde q)=\mathbb R[\tilde q]=\operatorname{span}_{\mathbb R}\{1,J_{\tilde q}\}\cong\mathbb C,\qquad J_{\tilde q}=\frac{\vec q}{|\vec q|}.$$ If $\tilde q\in\mathbb R$, then $C(\tilde q)=\mathbb H$.
Proof. Write $x=x_t+\vec x$. The scalar parts commute, and $$x\tilde q-\tilde q x=[\vec x,\vec q]=2\,\vec x\times\vec q.$$ If $\vec q\ne0$, the cross product vanishes exactly when $\vec x$ is parallel to $\vec q$, so $x=x_t+\lambda\vec q=x_t+\lambda sJ_{\tilde q}$; conversely every such $x$ commutes with $\tilde q$. If $\vec q=0$, the commutator vanishes for every $x$.
Corollary. For $\vec q\ne0$, $J_{\tilde q}^2=-1$, $\tilde q=q_t+sJ_{\tilde q}$, and the map $\mathbb R[\tilde q]\to\mathbb C$ sending $J_{\tilde q}$ to $i$ is an isomorphism of fields.
Theorem (the complex structure). For $\vec q\ne0$, left multiplication by $J_{\tilde q}$ is a complex structure on $\mathbb H$: $$L_{J_{\tilde q}}^2=-\mathrm{id},\qquad L_{J_{\tilde q}}L_{\tilde q}=L_{\tilde q}L_{J_{\tilde q}}.$$ Hence $\mathbb H$ is a complex vector space of complex dimension two under the scalar action $(a+bJ_{\tilde q})x=ax+bJ_{\tilde q}x$, and $L_{\tilde q}$ is complex-linear.
Proof. $J_{\tilde q}^2=-1$ gives $L_{J_{\tilde q}}^2=L_{-e_t}=-\mathrm{id}$; the commutation is the centrality of $J_{\tilde q}$ in $\mathbb R[\tilde q]$.
Proposition (a complex basis). Let $j$ be a unit pure quaternion orthogonal to $\vec q$. Then $jJ_{\tilde q}=-J_{\tilde q}j$, the four elements $1,J_{\tilde q},j,J_{\tilde q}j$ are an orthonormal real basis of $\mathbb H$, and $\{1,j\}$ is a basis of $\mathbb H$ over $\mathbb C_{\tilde q}=\mathbb R[\tilde q]$.
Proof. The product $J_{\tilde q}j$ is pure and has square $(J_{\tilde q}j)^2=J_{\tilde q}jJ_{\tilde q}j=-J_{\tilde q}^2j^2=-1$; the orthogonality follows from $\langle J_{\tilde q}j,J_{\tilde q}\rangle=\langle J_{\tilde q}j,j\rangle=0$. The four unit vectors are therefore a real basis, and $a+bj$ with $a,b\in\operatorname{span}\{1,J_{\tilde q}\}$ exhausts their span. Uniqueness follows from the independence of the real basis.
The Two Eigenlines
Definition (the idempotents). For $\vec q\ne0$ set $$\pi_+=\tfrac12(1-iJ_{\tilde q}),\qquad \pi_-=\tfrac12(1+iJ_{\tilde q}),$$ where $i$ is the scalar imaginary of $\mathbb C_{\tilde q}\otimes\mathbb C$.
Proposition. The elements $\pi_\pm$ are idempotent and orthogonal, $$\pi_+^2=\pi_+,\qquad \pi_-^2=\pi_-,\qquad \pi_+\pi_-=\pi_-\pi_+=0,\qquad \pi_++\pi_-=1,$$ and $$J_{\tilde q}\pi_+=i\pi_+,\qquad J_{\tilde q}\pi_-=-i\pi_-.$$ Consequently $$\tilde q\pi_+=(q_t+is)\pi_+=\lambda_+\pi_+,\qquad \tilde q\pi_-=(q_t-is)\pi_-=\lambda_-\pi_-.$$
Proof. The identities follow by expanding and using $J_{\tilde q}^2=-1$. For example, $\pi_+^2=\tfrac14(1-2iJ_{\tilde q}-J_{\tilde q}^2)=\tfrac12(1-iJ_{\tilde q})=\pi_+$, and $\pi_+\pi_-=\tfrac14(1+J_{\tilde q}^2)=0$. The action of $J_{\tilde q}$ is immediate from $J_{\tilde q}\pi_+=\tfrac12(J_{\tilde q}-iJ_{\tilde q}^2)=\tfrac12(J_{\tilde q}+i)=i\pi_+$.
Definition (the eigenlines). The eigenlines of $\tilde q$ are the two lines $$\ell_+=\mathbb C\pi_+,\qquad \ell_-=\mathbb C\pi_-$$ in $\mathbb C_{\tilde q}\otimes\mathbb C$. They are the two lines cut out by the points of the spectrum of $\mathbb C_{\tilde q}\otimes\mathbb C$.
Theorem. The algebra $\mathbb C_{\tilde q}\otimes\mathbb C$ is isomorphic to $\mathbb C\times\mathbb C$ by $J_{\tilde q}\mapsto(i,-i)$, and $\pi_+,\pi_-$ are the corresponding idempotents. The eigenlines are the two lines corresponding to the two points of the spectrum; they are distinct and one-dimensional over $\mathbb C$.
Proof. The isomorphism sends $\pi_+$ to $(1,0)$ and $\pi_-$ to $(0,1)$; the nontrivial idempotents of $\mathbb C\times\mathbb C$ are exactly these two. Since $J_{\tilde q}$ has the two distinct eigenvalues $\pm i$ on $\mathbb C_{\tilde q}\otimes\mathbb C$, the two lines are distinct.
The Eigenbasis and the Diagonal Form
Theorem (the eigenbasis). Let $j$ be a unit pure quaternion orthogonal to $\vec q$, and let $$E_+=\pi_+(\mathbb H_{\mathbb C}),\qquad E_-=\pi_-(\mathbb H_{\mathbb C}).$$ Then $$E_+=\operatorname{span}_{\mathbb C}\{\pi_+,\pi_+j\},\qquad E_-=\operatorname{span}_{\mathbb C}\{\pi_-,\pi_-j\},$$ the sum is direct, $\mathbb H_{\mathbb C}=E_+\oplus E_-$, and in the basis $(\pi_+,\pi_+j,\pi_-,\pi_-j)$ the operator $L_{\tilde q}$ has the matrix $$\operatorname{diag}(\lambda_+,\lambda_+,\lambda_-,\lambda_-).$$
Proof. For $x=\pi_+y$, $\tilde qx=\tilde q\pi_+y=\lambda_+\pi_+y=\lambda_+x$; hence $E_+\subseteq\ker(L_{\tilde q}-\lambda_+)$. The four vectors are a basis: $\{\pi_+,\pi_-\}$ is a $\mathbb C$-basis of $\mathbb C_{\tilde q}\otimes\mathbb C$, and $\{\pi_+j,\pi_-j\}$ is the corresponding basis. Hence $\dim_{\mathbb C}(E_++E_-)=4$, and the eigenspaces for the distinct eigenvalues $\lambda_+\ne\lambda_-$ are independent; the sum is direct. The matrix is diagonal because each basis vector is an eigenvector.
Corollary (the multiplicities). For $\vec q\ne0$ each eigenvalue has geometric multiplicity two, equal to its algebraic multiplicity. The operator $L_{\tilde q}$ is diagonalisable over $\mathbb C$.
Corollary (the real canonical form). In the real basis $(1,J_{\tilde q},j,J_{\tilde q}j)$ the operator $L_{\tilde q}$ is block-diagonal with two copies of $$B_{\tilde q}=\begin{pmatrix}q_t&-s\\ s&q_t\end{pmatrix},\qquad s=|\vec q|.$$ Hence its real canonical form is two identical rotation-scaling blocks; the eigenvalues are $q_t\pm is$.
Proof. $L_{\tilde q}(1)=q_t+sJ_{\tilde q}$ and $L_{\tilde q}(J_{\tilde q})=-s+q_tJ_{\tilde q}$ give the first block; the same computation on $j,J_{\tilde q}j$ gives the second.
Remark (complex eigenvectors may be nilpotent). The vector $\pi_+j$ is not in $\mathbb H$, and it is square-zero: $$(\pi_+j)^2=\pi_+j\pi_+j=\pi_+\pi_-j^2=0.$$ A complex eigenbasis of a real operator may therefore contain nilpotent vectors; this is not a defect: the minimal polynomial is squarefree.
Remark (non-uniqueness). The eigenspaces $E_\pm$ are canonical, but the basis is not: any unit pure $j$ orthogonal to $\vec q$ gives one, and any invertible change of basis inside each $E_\pm$ gives another.
The Scalar Case
Theorem. If $\tilde q=q_t\in\mathbb R$, then $L_{\tilde q}=q_t\,\mathrm{id}$, the spectrum is $\{q_t\}$, every nonzero vector of $\mathbb H_{\mathbb C}$ is an eigenvector, every basis is an eigenbasis, and the diagonal form is $q_t I_4$. The centraliser is $\mathbb H$, and no complex structure $J_{\tilde q}$ is distinguished.
Proof. The matrix is scalar; the assertions are immediate.
Remark (the limit $s\to0$). The non-scalar formulas have a well-defined limit: $\lambda_\pm\to q_t$, the two eigenlines coalesce, and the basis $(\pi_+,\pi_+j,\pi_-,\pi_-j)$ tends to a basis in which $L_{\tilde q}=q_tI$. The idempotents $\pi_\pm$ themselves depend on a chosen unit pure $J$, and both act as $q_t$; the decomposition is not canonical there.
The Polar Decomposition
Theorem. Every nonzero quaternion has a polar decomposition $$\tilde q=|\tilde q|\bigl(\cos\theta+\sin\theta\,J_{\tilde q}\bigr),\qquad \theta\in[0,\pi].$$ For $\tilde q\notin\mathbb R$, $J_{\tilde q}=\vec q/|\vec q|$ and $\theta\in(0,\pi)$ are unique, with $$\cos\theta=\frac{q_t}{|\tilde q|},\qquad \sin\theta=\frac{s}{|\tilde q|}.$$ For $\tilde q>0$, $\theta=0$ and $J$ is arbitrary; for $\tilde q<0$, $\theta=\pi$ and $J$ is arbitrary.
Proof. The modulus is $|\tilde q|=\sqrt{q_t^2+s^2}$, and the displayed cosine and sine have squares summing to one. For $s\ne0$, $J_{\tilde q}$ is determined by $\vec q$ and $\theta$ by $q_t/|\tilde q|$; for $s=0$, the vector term vanishes and $J$ drops out.
Corollary (the eigenvalues as modulus and argument). For $\tilde q\ne0$, $$\lambda_\pm=|\tilde q|e^{\pm i\theta}.$$ Both eigenvalues have modulus $|\tilde q|$; the spectral radius of $L_{\tilde q}$ is $|\tilde q|$.
Proof. $q_t\pm is=|\tilde q|(\cos\theta\pm i\sin\theta)$.
Remark. The polar decomposition is the geometric content of the eigenbasis: the axis $J_{\tilde q}$ is the complex structure, the two eigenlines are its two orientations, $|\tilde q|$ is the common modulus of the eigenvalues, and $\theta$ is their argument.
The Quaternion Spectrum
Definition (left and right spectra). For a quaternionic matrix $M$, a quaternion $\lambda$ is a left eigenvalue if $Mv=\lambda v$ for some nonzero $v$, and a right eigenvalue if $Mv=v\lambda$. The sets are the left spectrum $\sigma_L(M)$ and the right spectrum $\sigma_R(M)$.
Theorem (the 1×1 case). $$\sigma_L([\tilde q])=\{\tilde q\},\qquad \sigma_R([\tilde q])=\{v^{-1}\tilde qv:v\in\mathbb H^\times\}.$$ For a non-real quaternion the right spectrum is the two-sphere $$\sigma_R([\tilde q])=q_t+\{\vec w:|\vec w|=s\};$$ for a real quaternion it is the point $\{q_t\}$.
Proof. $L_{\tilde q}v=\lambda v$ with $v\ne0$ forces $\lambda=\tilde q$ by division. $R_{\tilde q}v=v\lambda$ gives $\lambda=v^{-1}\tilde qv$. Conjugation fixes $q_t$ and rotates $\vec q$ without changing $|\vec q|$, and the rotation group acts transitively on the sphere of radius $s$.
Definition (the S-spectrum). The S-spectrum of $\tilde q$ is $$\sigma_S(\tilde q)=\{\lambda\in\mathbb H:\tilde q^2-2\operatorname{Re}(\lambda)\tilde q+|\lambda|^2\text{ is not invertible}\}.$$ For a matrix the same definition uses $M^2-2\operatorname{Re}(\lambda)M+|\lambda|^2I$.
Theorem. For a single quaternion the S-spectrum equals the right spectrum.
Proof. Complete the square: $$\tilde q^2-2t\tilde q+|\lambda|^2=(\tilde q-t)^2+|\vec\lambda|^2,\qquad \lambda=t+\vec\lambda.$$ If $\tilde q\notin\mathbb R$, the right side has vector part $2(q_t-t)\vec q$ and scalar part $(q_t-t)^2-s^2+|\vec\lambda|^2$; it vanishes exactly when $t=q_t$ and $|\vec\lambda|=s$. If $\tilde q\in\mathbb R$, it vanishes exactly at $\lambda=q_t$. In both cases this is the right spectrum.
Corollary (the complex eigenvalues on the sphere). Under the isomorphism $\mathbb C_{\tilde q}\cong\mathbb C$ sending $J_{\tilde q}$ to $i$, the complex eigenvalues correspond to the two quaternions $q_t\pm sJ_{\tilde q}$. These are the two points where the S-sphere meets the real plane $\mathbb R+\mathbb R J_{\tilde q}$, and they are antipodal on it.
Remark (which spectrum is which). The left spectrum is the single point $\{\tilde q\}$, the right and S-spectra are the two-sphere, and the complex spectrum corresponds to a pair of points on the sphere. The multiplicity two of each complex eigenvalue is the complex dimension of its eigenspace in $\mathbb H_{\mathbb C}$, not a multiplicity of points of the right spectrum. The right eigenspace attached to the representative $q_t+sJ_{\tilde q}$ is the complex line $\mathbb C_{\tilde q}$ in $\mathbb H$, of real dimension two; the complexified eigenspace $E_+$ has complex dimension two, of real dimension four. They answer different problems.
Summary
For a non-real quaternion the centraliser is $\mathbb C_{\tilde q}=\mathbb R[\tilde q]=\operatorname{span}\{1,J_{\tilde q}\}\cong\mathbb C$, with $J_{\tilde q}=\vec q/|\vec q|$ a complex structure commuting with $L_{\tilde q}$. The idempotents $\pi_\pm=\tfrac12(1\mp iJ_{\tilde q})$ satisfy $\tilde q\pi_\pm=\lambda_\pm\pi_\pm$. Choosing a unit pure $j\perp\vec q$ gives the eigenspaces $E_\pm=\operatorname{span}_{\mathbb C}\{\pi_\pm,\pi_\pm j\}$, the direct sum $\mathbb H_{\mathbb C}=E_+\oplus E_-$, and the diagonal form $\operatorname{diag}(\lambda_+,\lambda_+,\lambda_-,\lambda_-)$. Over $\mathbb R$ the operator is two copies of the rotation-scaling block. The scalar case is the limit. The polar decomposition $\tilde q=|\tilde q|(\cos\theta+\sin\theta J_{\tilde q})$ gives $\lambda_\pm=|\tilde q|e^{\pm i\theta}$, and the left, right and S-spectra are respectively a point, the two-sphere and the same two-sphere.
Summary of Notation
| symbol | meaning |
|---|---|
| $\mathbb H_{\mathbb C}=\mathbb H\otimes_{\mathbb R}\mathbb C$ | the complexification |
| $C(\tilde q)$ | the centraliser $\{x:x\tilde q=\tilde qx\}$ |
| $J_{\tilde q}=\vec q/|\vec q|$ | the complex structure of a non-real quaternion |
| $\mathbb C_{\tilde q}=\mathbb R[\tilde q]$ | the centraliser, isomorphic to $\mathbb C$ |
| $\pi_\pm=\tfrac12(1\mp iJ_{\tilde q})$ | the spectral idempotents |
| $\ell_\pm=\mathbb C\pi_\pm$ | the two eigenlines |
| $E_\pm=\pi_\pm(\mathbb H_{\mathbb C})$ | the complexified eigenspaces |
| $j$ | a unit pure quaternion orthogonal to $\vec q$ |
| $\lambda_\pm=q_t\pm i|\vec q|$ | the complex eigenvalues |
| $\sigma_L,\sigma_R,\sigma_S$ | left, right and S-spectra |
| $\theta$ | the polar angle, $\cos\theta=q_t/|\tilde q|$ |
Further Reading
- Quaternion Algebra, for the multiplication, conjugation and scalar–vector decomposition.
- Norm and Invertibility in the Quaternion Algebra, for the norm, the modulus and the polar decomposition.
- Spectral Regularity and Degeneracy in the Quaternion Algebra, for the characteristic and minimal polynomials and the eigenvalues.
- Quaternion Spectral Theory, for the left, right and S-spectra of quaternionic matrices.
- Square Roots of a General Element in the Quaternion Algebra, for the sphere of imaginary units.
- Quaternion Automorphisms and Derivations, for conjugation and similarity.
- The Eigenvalue Decomposition of an Element, for the general spectral decomposition.
- Fuzhen Zhang, Quaternions and Matrices of Quaternions, Linear Algebra and its Applications 251 (1997), 21–57, for the standard form and the right spectrum.