The Spectrum of a Biquaternion

Introduction

The biquaternion algebra carries many quadratic conditions; this article studies one of them, the square of the vector part,

$$ B^2=Q_x^2+Q_y^2+Q_z^2, $$

which involves no scalar term. This condition decides whether the two eigenvalues of an element can be told apart: the characteristic polynomial is $X^2-2Q_tX+Q_t^2+B^2$, and its discriminant is $-4B^2$. This article defines the eigenvalue problem, proves the spectral criterion, and separates the elements of $\mathbb{B}$ into the two cases of the value of $B^2$: the regular case $B^2\neq0$, where the two eigenvalues are distinct and the element has a unique spectral decomposition, and the degeneracy case $B^2=0$, the spectral-degenerate cone, on which the two eigenvalues coalesce. This article is the study of this one condition and of its two cases.

Notation. An element is $\tilde Q=Q_te_t+\vec Q$, its quaternion conjugation is $\tilde Q^{\natural}=Q_t-\vec Q$, and the product is the plain product of the algebra, associative with the two-sided unit $e_t$. The vector part is $\vec Q=Q_x\vec e_x+Q_y\vec e_y+Q_z\vec e_z$, and

$$ B^2=Q_x^2+Q_y^2+Q_z^2,\qquad \vec Q^2=-B^2e_t. $$

The number $B$ is a fixed complex square root of $B^2$; the other root only exchanges the two eigenvalues. The companion article N-Regularity and N-Degeneracy in the Biquaternion Algebra studies the zero-divisor set; this article studies $B^2$ and the spectral cases it defines. The matrix realization $\mathsf{M}_2$ of $M_2(\mathbb{C})$ Pauli Representation appears only as a check, through the identity $\det\mathsf{M}_2(\tilde Q)=Q_t^2+B^2$; the criterion is proved without it. The algebra and its conventions are Introduction to the Algebra of Biquaternions; the eigenvalue decomposition, the normality criterion and the spectral theorem are The Biquaternion Eigenvalue Decomposition.

The Spectral Criterion

An eigenvalue is a scalar for which a left ideal is proper, and the criterion turns that definition into a single polynomial equation.

Definition (eigenvalue). A scalar $\lambda\in\mathbb{C}$ is an eigenvalue of $\tilde Q$ if the left ideal $\mathbb{B}(\tilde Q-\lambda e_t)$ is proper, that is $\mathbb{B}(\tilde Q-\lambda e_t)\neq\mathbb{B}$.

Definition (eigenvector). A non-zero $v\in\mathbb{B}$ is an eigenvector of $\tilde Q$ for $\lambda$ if $(\tilde Q-\lambda e_t)v=0$. The eigenvectors for $\lambda$, with the origin, are the left annihilator

$$ \operatorname{Ann}_{\ell}(\tilde Q-\lambda e_t)=\{v:\ (\tilde Q-\lambda e_t)v=0\}, $$

which is a right ideal of $\mathbb{B}$.

Definition (spectral idempotent). A non-zero idempotent $e$ is a spectral idempotent of $\tilde Q$ if $e\tilde Q=\tilde Qe=\lambda e$ for some scalar $\lambda$.

Theorem (the spectral criterion). For $\tilde Q\in\mathbb{B}$ and $\lambda\in\mathbb{C}$, the following are equivalent:

  1. $\lambda$ is an eigenvalue of $\tilde Q$;
  2. $\tilde Q-\lambda e_t$ is not invertible of the plain product;
  3. $(\lambda-Q_t)^2+B^2=0$.

Proof. Write $a=\tilde Q-\lambda e_t=a_te_t+\vec a$, with $a_t=Q_t-\lambda$ and $\vec a=\vec Q$.

(1)$\iff$(2). If $a$ is invertible then $\mathbb{B}a=\mathbb{B}$, so the left ideal is not proper. Conversely, if $\mathbb{B}a=\mathbb{B}$ then $e_t=xa$ for some $x$. Left multiplication by $a$ is injective, since $ay=0$ gives $y=xay=0$. In finite dimension an injective endomorphism is bijective, so there is $b$ with $ab=e_t$. Then $a(ba-e_t)=0$, hence $ba=e_t$ by injectivity, and $b$ is a two-sided inverse.

(2)$\iff$(3). The conjugate product is

$$ aa^{\natural}=a_t^2-\vec a^2=(a_t^2+B^2)e_t=((\lambda-Q_t)^2+B^2)e_t . $$

If the scalar is non-zero, then $a^{-1}=a^{\natural}/(a_t^2+B^2)$, so $a$ is invertible. If the scalar is zero, then $aa^{\natural}=0$; for $a\neq0$ this exhibits $a$ as a zero divisor, since $a^{\natural}\neq0$, and $a=0$ is not invertible. So $a$ is invertible exactly when the scalar is non-zero.

Remark (the determinant in the model). Under $\mathsf{M}_2$ the element $a=a_te_t+\vec a$ is

$$ \mathsf{M}_2(a)=\begin{pmatrix} a_t-ia_z & -ia_x-a_y\\ -ia_x+a_y & a_t+ia_z\end{pmatrix}, $$

whose determinant is $a_t^2+a_x^2+a_y^2+a_z^2$, the same scalar as above. The determinant is therefore a reading of $(\lambda-Q_t)^2+B^2$ in the model, not the proof of the criterion; the criterion holds without the model.

Corollary (the characteristic polynomial). For every scalar $X$,

$$ (X-\tilde Q)(X-\tilde Q^{\natural})=((X-Q_t)^2+B^2)e_t, $$

so the characteristic polynomial is

$$ p_{\tilde Q}(X)=X^2-2Q_tX+Q_t^2+B^2, $$

of discriminant

$$ \operatorname{disc}p_{\tilde Q}=4Q_t^2-4(Q_t^2+B^2)=-4B^2. $$

The eigenvalues are

$$ \lambda_\pm=Q_t\pm iB, $$

and they coincide exactly when $B=0$, that is exactly when $B^2=0$. In the model the same polynomial is $\det(\mathsf{M}_2(\tilde Q)-X\mathrm I_2)$.

Corollary (the eigenspaces). For a non-scalar $\tilde Q$ and each eigenvalue $\lambda$, the eigenspace $\operatorname{Ann}_{\ell}(\tilde Q-\lambda e_t)$ is a right ideal of complex dimension $2$. For $\tilde Q=Q_te_t$ the only eigenvalue is $Q_t$, and its eigenspace is all of $\mathbb{B}$.

Proof. For $a=\tilde Q-\lambda e_t\neq0$, the eigenspace is the left annihilator of $a$, a right ideal. It is non-zero because $a$ is a zero divisor, and it is proper because $a\neq0$. Every non-zero proper right ideal of $\mathbb{B}$ is minimal of complex dimension $2$, by the ideal theory of The Polarized Module of the Biquaternion Algebra, §The Two Minimal Left Ideals. At the scalar the difference $\tilde Q-Q_te_t$ is zero, and its annihilator is all of $\mathbb{B}$.

Corollary (Cayley–Hamilton and the inverse). Every $\tilde Q$ satisfies

$$ \tilde Q^2-2Q_t\tilde Q+(Q_t^2+B^2)e_t=0, $$

and if $Q_t^2+B^2\neq0$ then $\tilde Q^{-1}=(2Q_te_t-\tilde Q)/(Q_t^2+B^2)=\tilde Q^{\natural}/(Q_t^2+B^2)$.

Proof. With $\tilde Q=Q_te_t+\vec Q$ and $\vec Q^2=-B^2e_t$, $\tilde Q^2=Q_t^2e_t+2Q_t\vec Q-B^2e_t=2Q_t\tilde Q-(Q_t^2+B^2)e_t$. Multiplying by $2Q_te_t-\tilde Q=\tilde Q^{\natural}$ on either side gives $(Q_t^2+B^2)e_t$.

The Operator Form of the Criterion

The discriminant also reads on left multiplication: it decides whether the directional part of the operator is a complex structure or a nilpotent.

Proposition (the directional part). Write $\tilde Q=Q_te_t+\vec Q$. Then

$$ L_{\tilde Q}=Q_t\,\mathrm{id}+L_{\vec Q},\qquad R_{\tilde Q}=Q_t\,\mathrm{id}+R_{\vec Q}, $$

and

$$ L_{\vec Q}^{\,2}=R_{\vec Q}^{\,2}=-B^{2}\,\mathrm{id}. $$

Proof. For $X\in\mathbb{B}$, $L_{\vec Q}^{\,2}(X)=\vec Q(\vec QX)=\vec Q^{2}X=-B^{2}X$, because $\vec Q^{2}=-B^{2}e_t$, and $R_{\vec Q}^{\,2}(X)=(X\vec Q)\vec Q=X\vec Q^{2}=-B^{2}X$. The scalar part is central.

Corollary (the minimal polynomial of the directional part). The two directional operators have the same minimal polynomial,

$$ m_{L_{\vec Q}}=m_{R_{\vec Q}}= \begin{cases} X^{2}+B^{2}, & B^{2}\neq0,\\ X^{2}, & B^{2}=0,\ \vec Q\neq0,\\ X, & \vec Q=0. \end{cases} $$

Proof. If $B^{2}\neq0$, the polynomial $X^{2}+B^{2}=(X-iB)(X+iB)$ annihilates the operator and is squarefree, because $B\neq0$; a degree-one annihilator would make the operator scalar, which it is not when $\vec Q\neq0$. If $B^{2}=0$ and $\vec Q\neq0$, the square annihilates and the operator is not zero. If $\vec Q=0$, the operator is zero.

Corollary (the complex structure and the defect). If $B^{2}\neq0$, then $J=L_{\vec Q}/B$ satisfies $J^{2}=-\mathrm{id}$, and the eigenvalues of $L_{\vec Q}$ are $\pm iB$: the directional part is a complex structure. If $B^{2}=0$ and $\vec Q\neq0$, $L_{\vec Q}$ is nilpotent of index two. Consequently $L_{\tilde Q}$ is semisimple except on $\mathcal S\setminus\{\vec Q=0\}$, where it is defective with minimal polynomial $(X-Q_t)^{2}$ and Jordan type $(2,2)$.

Proof. The square of $J$ is $-B^{2}\mathrm{id}/B^{2}=-\mathrm{id}$; the eigenvalues of $L_{\vec Q}$ are the roots $\pm iB$ of $X^{2}+B^{2}$. Translation by $Q_t\mathrm{id}$ preserves semisimplicity and defectivity, and it preserves the Jordan type. In the defective case the minimal polynomial becomes $(X-Q_t)^{2}$; on the scalar stratum it is $X-Q_t$.

Remark (the commutator form). The same computation on commutators, $\mathrm{ad}_{\vec Q}^{\,3}=-4B^{2}\mathrm{ad}_{\vec Q}$, is the inner derivation of Relations between the N Norm and Spectral Regularity and Degeneracy, §The Two Operator Loci, where it is nilpotent exactly on $\mathcal S$.

Remark (left and right eigenvalues). The centre of $\mathbb{B}$ is $\mathbb{C}$, so complex scalars commute with every element. A left eigenvalue, $\tilde Qv=\lambda v$, and a right eigenvalue, $\tilde Qv=v\lambda$, therefore define the same set when the scalar is required to lie in $\mathbb{C}$, and that set is $\sigma(\tilde Q)$. The distinction between the two eigenvalue problems, which is genuine over a noncommutative division ring (Quaternion Spectral Theory, §The Spectrum of a Quaternion), collapses here.

The Two Cases

The discriminant of the characteristic polynomial is $-4B^2$, and its vanishing, and nothing else, separates the algebra into two spectral cases.

Definition (spectral decomposition). A spectral decomposition of $\tilde Q$ is either the one-term writing $\tilde Q=Q_te_t$ when $\tilde Q$ is scalar, or a two-term writing

$$ \tilde Q=\lambda_+e_++\lambda_-e_-, $$

where $\lambda_+\neq\lambda_-$ and $e_+,e_-$ are non-zero idempotents with

$$ e_+^2=e_+,\qquad e_-^2=e_-,\qquad e_+e_-=e_-e_+=0,\qquad e_++e_-=e_t. $$

The scalars $\lambda_\pm$ are the eigenvalues and the idempotents $e_\pm$ are the spectral idempotents.

Criterion (four descriptions). For a non-scalar $\tilde Q$ the following are equivalent:

  1. $B^2\neq0$;
  2. the eigenvalues are distinct;
  3. $\tilde Q$ is semisimple;
  4. $L_{\tilde Q}$ is diagonalisable.

Proof. (1)$\iff$(2): the eigenvalues are $Q_t\pm iB$, and they are distinct exactly when $B\neq0$. (2)$\iff$(3): for a non-scalar element the minimal polynomial is the characteristic polynomial $(X-Q_t)^2+B^2$, of degree two, since a degree-one minimal polynomial would make the element scalar; it is squarefree exactly when its two roots are distinct. (3)$\iff$(4): the left regular representation is faithful and carries $\mathbb{C}[\tilde Q]$ isomorphically onto $\mathbb{C}[L_{\tilde Q}]$, so $\tilde Q$ and $L_{\tilde Q}$ have the same minimal polynomial, and an operator is diagonalisable exactly when that polynomial is squarefree.

The two cases are collected in the table; the regular row is proved in §The Spectral Decomposition and §The Eigenbasis of Left Multiplication, the degenerate row in The Jordan Basis of a Spectrally Degenerate Biquaternion, §The Three Strata.

case condition spectrum spectral decomposition eigenbasis
regular $B^2\neq0$ $\{\lambda_+,\lambda_-\}$, distinct two terms, unique exists
degenerate $B^2=0$ $\{Q_t\}$ one term iff $\vec Q=0$, none otherwise every basis iff $\vec Q=0$, none otherwise

The Four Minimal Polynomials

The characteristic polynomial has two coefficients, and the minimal polynomial separates what the characteristic polynomial does not.

type condition $p_{\tilde Q}$ $m_{\tilde Q}$
semisimple, non-scalar $B^2\neq0$ $(X-Q_t)^2+B^2$ $(X-Q_t)^2+B^2$
scalar $B^2=0$, $\vec Q=0$ $(X-Q_t)^2$ $X-Q_t$
Jordan $B^2=0$, $\vec Q\neq0$, $Q_t\neq0$ $(X-Q_t)^2$ $(X-Q_t)^2$
nilpotent $B^2=0$, $\vec Q\neq0$, $Q_t=0$ $X^2$ $X^2$

The first row is the spectrally regular case. The second row is the central part of the spectral-degenerate cone, $\vec Q=0$: it is semisimple too, but its semisimplicity comes from the degree-one minimal polynomial, not from distinct roots. The last two rows form the non-central part of the cone, split by the scalar part. The table says what the two-case split does not. The characteristic polynomial cannot separate the scalar case from the Jordan case: both have $(X-Q_t)^2$, and only the minimal polynomial distinguishes them. It also cannot separate the nilpotent case from the origin: both have $X^2$. The two-case split of this article is the coarser reading by $B^2$: the regular case is the non-scalar semisimple row, and the degenerate case contains the scalar, Jordan and nilpotent rows. The classification is the one used in The Semisimple–Nilpotent Split of a Biquaternion.

Remark (the words). Unqualified regular means spectrally regular, $B^2\neq0$, and degenerate means spectrally degenerate, $B^2=0$; both words are fixed by the discriminant. Degenerate here does not mean defective: the vertex is semisimple, while the other two strata of the spectral-degenerate cone are defective.

Remark (the condition is not multiplicative). Neither case is stable under the product. Two elements of $B^2=0$ can multiply to an element of $B^2\neq0$: both factors of

$$ (\vec e_x+i\vec e_y)(\vec e_x-i\vec e_y)=-2i\vec e_z $$

have $B^2=0$, and the product has $B^2=-4$. Two elements of $B^2\neq0$ can multiply to an element of $B^2=0$: both factors of

$$ (e_t+\vec e_x)(e_t-\vec e_x)=2e_t $$

have $B^2=1$, and the product has $B^2=0$. So the regular case is not a subgroup, and the spectral condition is not preserved by multiplication in either direction.

The Regular Case: Distinct Eigenvalues

When $B^2\neq0$ the two eigenvalues are distinct, the spectral idempotents are the Lagrange idempotents, and the decomposition is unique.

The Spectral Decomposition

Theorem (existence). Let $B^2\neq0$ and put

$$ \vec{Ζ}=-\frac{i\vec Q}{B},\qquad \tilde\Pi_\pm=\frac{e_t\pm\vec{Ζ}}{2},\qquad \lambda_\pm=Q_t\pm iB. $$

Then $\tilde\Pi_\pm$ are non-zero idempotents with $\tilde\Pi_+\tilde\Pi_-=\tilde\Pi_-\tilde\Pi_+=0$ and $\tilde\Pi_++\tilde\Pi_-=e_t$, and

$$ \tilde Q=\lambda_+\tilde\Pi_++\lambda_-\tilde\Pi_-,\qquad \tilde Q\tilde\Pi_\pm=\lambda_\pm\tilde\Pi_\pm. $$

Proof. Since $\vec Q^2=-B^2e_t$, one has $\vec{Ζ}^2=e_t$ and $\vec Q=iB\vec{Ζ}$. Hence $\tilde\Pi_\pm^2=(e_t\pm2\vec{Ζ}+e_t)/4=\tilde\Pi_\pm$ and $\tilde\Pi_+\tilde\Pi_-=(e_t-\vec{Ζ}^2)/4=0$, and the sum is $e_t$. The idempotents are non-zero because $\vec{Ζ}$ has zero scalar part, so $\vec{Ζ}\neq\pm e_t$. For the action, $\vec Q\vec{Ζ}=(-i/B)\vec Q^2=iBe_t$, hence

$$ \tilde Q\tilde\Pi_+=\frac{(Q_t e_t+\vec Q)(e_t+\vec{Ζ})}{2} =\frac{(Q_t+iB)(e_t+\vec{Ζ})}{2}=\lambda_+\tilde\Pi_+, $$

and $\tilde Q\tilde\Pi_-=\lambda_-\tilde\Pi_-$ the same way. Adding the two identities multiplied by the eigenvalues gives the decomposition.

Corollary (Lagrange form). The idempotents are the Lagrange idempotents,

$$ \tilde\Pi_\pm=\frac{\tilde Q-\lambda_\mp e_t}{\lambda_\pm-\lambda_\mp}, $$

and the spectral decomposition is unique.

Proof. The first identity is the computation $\tilde Q-\lambda_-=iB(e_t+\vec{Ζ})$, and its analogue with the signs exchanged. For uniqueness, the minimal polynomial $(X-\lambda_+)(X-\lambda_-)$ has distinct roots, so the Chinese remainder theorem gives $\mathbb{C}[\tilde Q]\cong\mathbb{C}[X]/((X-\lambda_+)(X-\lambda_-))\cong\mathbb{C}\times\mathbb{C}$; the idempotents of that product are $0,e_t,\tilde\Pi_+,\tilde\Pi_-$, and every spectral idempotent lies in the centraliser $\mathbb{C}[\tilde Q]$.

Corollary (spectral projectors). $L_{\tilde\Pi_+}$ is the projector onto the $\lambda_+$-eigenspace: it sends $X$ to $\tilde\Pi_+X$, with image $\tilde\Pi_+\mathbb{B}$ and kernel $\tilde\Pi_-\mathbb{B}$. The same statement holds with the two signs exchanged.

The Eigenbasis of Left Multiplication

Theorem (the eigenspaces). For $B^2\neq0$ and each sign,

$$ \ker(L_{\tilde Q}-\lambda_\pm)=\tilde\Pi_\pm\mathbb{B}, $$

a right ideal of complex dimension $2$. Each eigenvalue therefore has multiplicity two.

Proof. Since $\tilde Q-\lambda_+=iB(\vec{Ζ}-e_t)=-2iB\tilde\Pi_-$, the equation $(\tilde Q-\lambda_+)X=0$ is $\tilde\Pi_-X=0$, that is $X=\tilde\Pi_+X$, that is $X\in\tilde\Pi_+\mathbb{B}$. In a basis in which $\tilde\Pi_+=\operatorname{diag}(1,0)$, the ideal $\tilde\Pi_+\mathbb{B}$ is the set of matrices whose second row vanishes, of complex dimension $2$.

Corollary (eigenbasis). Choose non-zero $\vec\Upsilon_-\in\tilde\Pi_+\mathbb{B}\tilde\Pi_-$ and $\vec\Upsilon_+\in\tilde\Pi_-\mathbb{B}\tilde\Pi_+$. Then

$$ (\tilde\Pi_+,\ \vec\Upsilon_-,\ \vec\Upsilon_+,\ \tilde\Pi_-) $$

is an eigenbasis of $L_{\tilde Q}$, with diagonal values $(\lambda_+,\lambda_+,\lambda_-,\lambda_-)$. An eigenbasis is any pair of bases of the two eigenspaces, so it is determined only up to an invertible change of basis on each.

Proof. The mixed corners are one-dimensional, and $\vec\Upsilon_-\in\tilde\Pi_+\mathbb{B}$ while $\vec\Upsilon_+\in\tilde\Pi_-\mathbb{B}$, so the four vectors form a basis of the two eigenspaces and hence of $\mathbb{B}$.

Remark (the two sides). The minimal left ideals $S_\pm=\mathbb{B}\tilde\Pi_\pm$ are the eigenspaces of the right multiplication $R_{\tilde Q}$, with eigenvalues $\lambda_\pm$. They are not eigenspaces of $L_{\tilde Q}$: $S_+$ contains $\tilde\Pi_+$ and $\vec\Upsilon_+$, with eigenvalues $\lambda_+$ and $\lambda_-$, and $S_-$ contains $\tilde\Pi_-$ and $\vec\Upsilon_-$, with eigenvalues $\lambda_-$ and $\lambda_+$. The eigenbasis above is an eigenbasis of left multiplication, not of right multiplication.

Remark (the degenerate side). If $B^2=0$ and $\vec Q\neq0$, the minimal polynomial is $(X-Q_t)^2$, the eigenspace has dimension $2$, and no eigenbasis exists. If $\vec Q=0$ the element is scalar and $L_{\tilde Q}=Q_t\,\mathrm{id}$; every basis is then an eigenbasis.

The Degenerate Case: The Spectral-Degenerate Cone

On the spectral-degenerate cone $B^2=0$ the eigenvalues have coalesced, and two questions decide the spectral behaviour of a point: whether the vector part vanishes, and whether the scalar part vanishes.

The Spectral-Degenerate Cone

Definition. The spectral-degenerate cone is

$$ \mathcal S=\{B^2=0\}=\{\vec Q^2=0\}=\{\operatorname{disc}p_{\tilde Q}=0\}. $$

Theorem (three readings of one equation). The equation $B^2=0$ is simultaneously:

  • geometric: the vector part is isotropic, $Q_x^2+Q_y^2+Q_z^2=0$;
  • algebraic: the vector part is square-zero, $\vec Q^2=0$, hence nilpotent of index two when it is non-zero;
  • spectral: the discriminant vanishes, and the two eigenvalues coalesce at $Q_t$.

Theorem (the geometry). The spectral-degenerate cone is the product of the scalar line and the isotropic cone of the vector subspace,

$$ \mathcal S=\mathbb{C}e_t\times\{\vec Q\in\mathbb{B}_{V}:\vec Q^2=0\}. $$

It is a complex cone of complex dimension $3$ and real dimension $6$. Its singular locus is the vertex $\{\vec Q=0\}$, the whole scalar line; off the vertex it is a smooth hypersurface.

Proof. The equation $B^2=0$ is one complex equation in the four complex coefficients, homogeneous of degree two. Its gradient with respect to $(Q_t,Q_x,Q_y,Q_z)$ is

$$ (0,\,2Q_x,\,2Q_y,\,2Q_z), $$

which vanishes exactly when $\vec Q=0$. Off the vertex the gradient is non-zero and the equation cuts a smooth hypersurface, of complex dimension $4-1=3$ and real dimension $8-2=6$. On the vertex the spectral-degenerate cone is $\mathbb{C}e_t\times C$, with $C$ the isotropic cone of $\mathbb{C}^3$, which is singular at its apex; the product is singular along the whole line.

The strata of the spectral-degenerate cone. The spectral-degenerate cone has three pieces: the scalar stratum $\mathcal S^{\mathrm{s}}=\{\vec Q=0\}$, the vertex; the nilpotent stratum $\mathcal S^{\mathrm{n}}=\{\vec Q\neq0,\ Q_t=0\}$; and the unipotent stratum $\mathcal S^{\mathrm{u}}=\{\vec Q\neq0,\ Q_t\neq0\}$, with $\mathcal S=\mathcal S^{\mathrm{s}}\sqcup\mathcal S^{\mathrm{n}}\sqcup\mathcal S^{\mathrm{u}}$. The last two form the Jordan part, the defective part of the spectral-degenerate cone. Their normal forms, minimal polynomials, ranks, centralisers and spectral data are in The Jordan Basis of a Spectrally Degenerate Biquaternion, §The Three Strata; how the strata sit inside the four norm–spectrum sectors is in Relations between the N Norm and Spectral Regularity and Degeneracy, §The Four Sectors Compared and §The Two Decompositions.

Warning (what the spectral-degenerate cone is not). The equation involves only $\vec Q$: the spectral-degenerate cone is blind to the scalar part, and the whole scalar line lies on it. It is not a subspace, since $\vec e_x+i\vec e_y$ and $\vec e_x-i\vec e_y$ lie on it while their sum $2\vec e_x$ does not. It is not the set of defective elements: a point of the vertex is semisimple with minimal polynomial $X-Q_t$. The spectral-degenerate cone is exactly the set of elements with a repeated eigenvalue.

Remark (a zero divisor with a simple spectrum). The spectral condition is independent of the zero-divisor condition, and the witness is the Hermitian idempotent. Let $\vec u$ be a vector with $\vec u\cdot\vec u=1$, so that $\vec u^{2}=-e_t$, and put

$$ \tilde\Pi=\frac12(e_t+i\vec u). $$

Then $\tilde\Pi^{2}=\tilde\Pi$, and

$$ \tilde\Pi\tilde\Pi^{\natural}=\Bigl(\frac14+B^{2}(\tilde\Pi)\Bigr)e_t=\Bigl(\frac14-\frac14\Bigr)e_t=0,\qquad B^{2}(\tilde\Pi)=\frac{i^{2}}{4}(\vec u\cdot\vec u)=-\frac14. $$

The spectrum is $\{1,0\}$ and the minimal polynomial is $X(X-1)$, which is squarefree: $\tilde\Pi$ is a zero divisor and is spectrally regular. The comparison of the null cone and the spectral-degenerate cone is Relations between the N Norm and Spectral Regularity and Degeneracy, §Independence of the Two Cones.

The Isotropic Line and the Flag

A non-isotropic line splits the space; an isotropic line only flags it.

Let $(\vec X,\vec Y)=X_xY_x+X_yY_y+X_zY_z$ be the symmetric form on the vector subspace, so that $(\vec Q,\vec Q)=B^{2}$. If $B^{2}\neq0$, the line $\mathbb{C}\vec Q$ is non-isotropic and

$$ \mathbb{C}^{3}=\mathbb{C}\vec Q\oplus\vec Q^{\perp} $$

is a direct sum of non-degenerate summands, canonical in $\vec Q$. If $B^{2}=0$ and $\vec Q\neq0$, the line is isotropic and $\vec Q\in\vec Q^{\perp}$, so $\mathbb{C}\vec Q+\vec Q^{\perp}=\vec Q^{\perp}$ and the sum is not direct. No complement is determined by $\vec Q$: the canonical objects are the line, the plane $\vec Q^{\perp}$ that contains it, and the flag

$$ 0\subset\mathbb{C}\vec Q\subset\vec Q^{\perp}\subset\mathbb{C}^{3}. $$

On the algebra, $\vec Q^{2}=0$ gives $\operatorname{im}L_{\vec Q}\subseteq\ker L_{\vec Q}$, and both spaces have complex dimension $2$, so they coincide. The canonical flag is

$$ 0\subset\mathbb{C}\vec Q\subset\ker L_{\vec Q}=\operatorname{im}L_{\vec Q}\subset\mathbb{B}. $$

A Jordan basis of $L_{\vec Q}$ needs two vectors outside $\ker L_{\vec Q}$ whose images span the kernel. The quotient $\mathbb{B}/\ker L_{\vec Q}$ is two-dimensional, but no splitting of it is determined by $\vec Q$ alone. The eigenbasis fails on the spectral-degenerate cone for this reason: the isotropic line admits no canonical complement. A canonical complement requires extra data, and the resulting Jordan basis is constructed in The Jordan Basis of a Spectrally Degenerate Biquaternion.

The Similarity Classification

Definition. Two biquaternions $\tilde Q_1,\tilde Q_2$ are similar if $\tilde Q_2=S\tilde Q_1S^{-1}$ for some invertible $S\in\mathbb{B}$.

Theorem (the classification).

  1. If $B^2\neq0$, the similarity class is determined by the unordered pair of eigenvalues $\{\lambda_+,\lambda_-\}$.
  2. If $B^2=0$, the spectrum is the singleton $\{Q_t\}$, and exactly two similarity classes share it: the central class $\tilde Q=Q_te_t$, and the non-semisimple class $\vec Q\neq0$ with $\vec Q^2=0$, which is similar to the Jordan block $J_2(Q_t)$. At $Q_t=0$ the second class is that of the non-zero nilpotents, $J_2(0)$; at $Q_t\neq0$ it is that of the unipotent scalings $Q_tu$, $J_2(Q_t)$.

Proof. (1) The characteristic polynomial has two distinct roots, so $\tilde Q$ is diagonalisable with those roots; two diagonalisable matrices with the same spectrum are conjugate in $M_2(\mathbb{C})\cong\mathbb{B}$, and the spectrum is unordered because the two eigenvalues are exchanged by the choice of $B$. (2) On the spectral-degenerate cone $\vec Q^2=0$. If $\vec Q=0$ the element is central. If $\vec Q\neq0$, then $\tilde Q-Q_te_t$ is non-zero with square $0$, so the minimal polynomial is $(X-Q_t)^2$ and the Jordan form is a single block of size $2$; all non-zero square-zero elements are rank-one nilpotents, hence conjugate.

The unordered spectrum therefore determines the similarity class off the spectral-degenerate cone, and carries no information on it beyond $Q_t$: the discriminant is what separates the central class from the non-semisimple one.

The Two Cases on the Remarkable Subspaces

The condition $B^2$ depends only on the vector part, so its restriction to a remarkable subspace is read from the vector coordinates of that subspace.

remarkable subspace real basis restriction of $B^2$ elements with $B^2=0$
$\mathbb{B}_{C}$, the centre $e_t,\,ie_t$ $0$ the whole centre
$\mathbb{B}_{V}$, the vector subspace $\vec e_k,\,i\vec e_k$ $\sum_kQ_k^2$ the isotropic cone, complex dimension $2$
$\mathbb{B}_{H}$, the quaternion subspace $e_t,\vec e_x,\vec e_y,\vec e_z$ $q_x^2+q_y^2+q_z^2$ $\mathbb{R}e_t$
$\mathbb{B}_{iH}$, the anti-quaternion subspace $ie_t,i\vec e_x,i\vec e_y,i\vec e_z$ $-(q_x^2+q_y^2+q_z^2)$ $\mathbb{R}ie_t$
$\mathbb{B}_{\Psi}$, the Hermitian subspace $e_t,i\vec e_x,i\vec e_y,i\vec e_z$ $-(a_x^2+a_y^2+a_z^2)$ $\mathbb{R}e_t$
$\mathbb{B}_{\Phi}$, the anti-Hermitian subspace $ie_t,\vec e_x,\vec e_y,\vec e_z$ $a_x^2+a_y^2+a_z^2$ $\mathbb{R}ie_t$

Corollary. Every element of the centre is spectrally degenerate, and the centre is the only remarkable subspace contained in $\mathcal S$. On the quaternion and anti-Hermitian subspaces $B^2$ is a sum of squares, on the anti-quaternion and Hermitian subspaces it is the negative of a sum of squares; in each case the zero set is the real or the imaginary scalar line. Only on the vector subspace is the condition a genuine cone, the isotropic cone of complex dimension $2$ and real dimension $4$.

Proof. The values are read from the displayed real bases: on the centre all three vector coefficients vanish; on $\mathbb{B}_{H}$ and $\mathbb{B}_{iH}$ the coefficients $Q_k$ are real or purely imaginary; on $\mathbb{B}_{\Psi}$ and $\mathbb{B}_{\Phi}$ the scalar coefficient is real or purely imaginary and the vector coefficients are purely imaginary or real. Each non-central restriction is a sum of three squares with a sign, so its zero set is the scalar line of the subspace.

Corollary (the strata on the remarkable subspaces). The centre and the four real-type subspaces meet only the scalar stratum; the vector subspace meets the scalar stratum at the origin and the nilpotent stratum elsewhere; the unipotent stratum meets none of the remarkable subspaces.

Proof. On the centre and on the four real-type subspaces, $B^2=0$ forces $\vec Q=0$, so the point is scalar. On the vector subspace $Q_t=0$ by construction, so a point of $\mathcal S$ there is the origin or a nilpotent. The unipotent stratum requires $Q_t\neq0$ and $\vec Q\neq0$ together, and no remarkable subspace carries both coordinates.

Summary

The biquaternion algebra carries many quadratic conditions; the one studied here is the square of the vector part, $B^2=Q_x^2+Q_y^2+Q_z^2$, which involves no scalar term. It decides whether the two eigenvalues of an element can be told apart: the characteristic polynomial is $X^2-2Q_tX+Q_t^2+B^2$, of discriminant $-4B^2$, with eigenvalues $Q_t\pm iB$.

The eigenvalue problem is stated without a module and without an acting operator: an eigenvalue is a scalar for which the left ideal $\mathbb{B}(\tilde Q-\lambda e_t)$ is proper, an eigenvector lies in the left annihilator, and a spectral idempotent is an idempotent commuting with the element. The criterion proves that $\lambda$ is an eigenvalue exactly when $\tilde Q-\lambda e_t$ is a non-invertible, exactly when $(\lambda-Q_t)^2+B^2=0$. The determinant of the matrix model is the same scalar, not the proof; the proof is the conjugate product $a a^{\natural}=(a_t^2+B^2)e_t$. The eigenspace of a non-scalar element has complex dimension $2$; Cayley–Hamilton gives the inverse whenever the determinant does not vanish. The same condition reads on the operators: $L_{\tilde Q}=Q_t\,\mathrm{id}+L_{\vec Q}$ and $R_{\tilde Q}=Q_t\,\mathrm{id}+R_{\vec Q}$, with $L_{\vec Q}^{\,2}=R_{\vec Q}^{\,2}=-B^2\,\mathrm{id}$. Off the spectral-degenerate cone the directional part is a complex structure; on the spectral-degenerate cone it is nilpotent.

The regular case is $B^2\neq0$. The eigenvalues are distinct, the element is semisimple, and its left multiplication is diagonalisable. The spectral idempotents are the Lagrange idempotents $\tilde\Pi_\pm=(\tilde Q-\lambda_\mp e_t)/(\lambda_\pm-\lambda_\mp)$, which coincide with $(e_t\pm\vec{Ζ})/2$ for $\vec{Ζ}=-i\vec Q/B$; the decomposition is unique, and an eigenbasis exists with diagonal values $(\lambda_+,\lambda_+,\lambda_-,\lambda_-)$.

The degeneracy case is the spectral-degenerate cone $B^2=0$, of complex dimension $3$ and real dimension $6$, the product of the scalar line and the isotropic cone of the vector subspace, smooth off the scalar vertex. On it the two eigenvalues have coalesced at $Q_t$. The spectral-degenerate cone splits into the scalar vertex and the Jordan part, the latter carrying the nilpotent and unipotent strata; their normal forms, minimal polynomials, ranks, centralisers and spectral data are in The Jordan Basis of a Spectrally Degenerate Biquaternion, §The Three Strata. The condition is blind to the scalar part, so the whole centre is degenerate, and no case is stable under multiplication. The spectral condition and the zero-divisor condition are independent: a Hermitian idempotent has $B^2=-\tfrac14$ and is a zero divisor.

Summary of Notation

symbol meaning
$\mathbb{B}=\mathbb{C}\otimes_{\mathbb{R}}\mathbb{H}$ the biquaternion algebra
$\tilde Q=Q_te_t+\vec Q$ a general element
$\vec Q=Q_x\vec e_x+Q_y\vec e_y+Q_z\vec e_z$ the vector part
$B^2=Q_x^2+Q_y^2+Q_z^2$ the square of the vector part, the discriminant up to $-4$
$p_{\tilde Q}(X)=X^2-2Q_tX+Q_t^2+B^2$ the characteristic polynomial, equal to $(X-Q_t)^2+B^2$
$m_{\tilde Q}$ the minimal polynomial of $\tilde Q$
$a a^{\natural}=(a_t^2+B_a^2)e_t$ the conjugate product, the intrinsic invertibility test
$\det\mathsf{M}_2(a)=a_t^2+a_x^2+a_y^2+a_z^2$ the same scalar in the matrix model
$\operatorname{disc}p_{\tilde Q}=-4B^2$ the discriminant
$\lambda_\pm=Q_t\pm iB$ the eigenvalues
$\mathcal S=\{B^2=0\}$ the spectral-degenerate cone
$\mathcal S^{\mathrm{s}},\ \mathcal S^{\mathrm{n}},\ \mathcal S^{\mathrm{u}}$ the scalar, nilpotent and unipotent strata of $\mathcal S$
$L_{\tilde Q}(X)=\tilde QX,\ R_{\tilde Q}(X)=X\tilde Q$ left and right multiplication
$\operatorname{Ann}_{\ell}(a)=\{v:av=0\}$ the left annihilator, a right ideal
$\vec{Ζ}=-i\vec Q/B$ the pure root, with $\vec{Ζ}^2=e_t$
$\tilde\Pi_\pm=(e_t\pm\vec{Ζ})/2$ the spectral idempotents, equal to $(\tilde Q-\lambda_\mp e_t)/(\lambda_\pm-\lambda_\mp)$

Further Reading