Spectral Regularity and Degeneracy in the Dual-Number Algebra

Introduction

The dual-number algebra $\mathbb{C}_{\mathrm{dln},R} = R[\varepsilon]/(\varepsilon^2)$ is the simplest algebra over a field in which the characteristic polynomial of every element is a square. For $Q=a+b\varepsilon$, left multiplication has characteristic polynomial $(X-a)^2$, so the spectrum is the singleton $\{a\}$: the two eigenvalues of a quadratic have coalesced before any condition is imposed. This article defines the eigenvalue problem, proves the coalescence, and separates two notions that the coalescence puts at risk of being confused. Spectral degeneracy is the repetition of the eigenvalue; it holds for every element. Defect is the absence of an eigenbasis; it holds exactly when $b\neq0$. The first is universal, the second is a proper subset of the algebra, and the whole of the spectral theory of $\mathbb{C}_{\mathrm{dln},R}$ is the distance between them.

Notation and conventions. Throughout, $R$ is a real closed field, with the geometric case $R=\mathbb{R}$; the algebra is $\mathbb{C}_{\mathrm{dln},R}$, and a general element is

$$ Q = a + b\varepsilon, \qquad a, b \in R, $$

with $a = \operatorname{Re}Q$ and $b = \operatorname{Inf}Q$. In the prime notation of Dual-Numbers Algebra the same element is $Q = q + \varepsilon q'$, with $a = q$ and $b = q'$. The norm (the dual norm) is

$$ \langle Q\rangle = Q\bar Q = (a + b\varepsilon)(a - b\varepsilon) = a^2, $$

where $\bar Q = a - b\varepsilon$ is dual conjugation. The norm is a degenerate quadratic form of rank one, multiplicative, with radical the maximal ideal $\mathrm{M} = (\varepsilon) = \{a = 0\}$; it vanishes exactly on $\mathrm{M}$ and is non-zero exactly on the invertibles. These facts are those of Dual-Numbers Norm and Invertibility. The characteristic polynomial of $Q$ is the characteristic polynomial of $L_Q$; it is computed below. The matrix-element representation $\mathsf{M}_2$ of Dual-Numbers Matrix Element Representation is used only to compare conventions, and the eigenvalue decomposition of a general element is The Eigenvalue Decomposition of an Element. The companion articles The Defective Sector of the Dual-Number Algebra and Relations between the Dual Norm and Spectral Degeneracy take up the defect and the comparison with the norm.

The Eigenvalue Problem

An eigenvalue is a scalar for which the translate $Q-\lambda$ has a non-zero kernel, and the criterion is the invertibility criterion of the algebra.

Definition (eigenvalue). A scalar $\lambda\in R$ is an eigenvalue of $Q$ if there is a non-zero $v\in\mathbb{C}_{\mathrm{dln},R}$ with

$$ Qv = \lambda v. $$

Definition (eigenvector, eigenspace, spectrum). A non-zero $v$ satisfying $Qv = \lambda v$ is an eigenvector of $Q$ for $\lambda$. The eigenspace of $\lambda$ is

$$ E_\lambda(Q) = \ker L_{Q-\lambda} = \{v : (Q-\lambda)v = 0\}, $$

where $L_A(v) = Av$ is left multiplication. The spectrum $\sigma(Q)$ is the set of eigenvalues.

Proposition (the eigenvalue criterion). A scalar $\lambda$ is an eigenvalue of $Q$ if and only if $Q-\lambda$ is not invertible of $\mathbb{C}_{\mathrm{dln},R}$.

Proof. If $Q-\lambda$ is invertible and $(Q-\lambda)v = 0$, then $v = (Q-\lambda)^{-1}(Q-\lambda)v = 0$, so no non-zero eigenvector exists. Conversely, if $Q-\lambda$ is not invertible, then its real part $a-\lambda$ vanishes, because the invertibles of $\mathbb{C}_{\mathrm{dln},R}$ are exactly the elements with non-zero real part. Hence $Q-\lambda = b\varepsilon$, and

$$ (Q-\lambda)\varepsilon = b\varepsilon^2 = 0 $$

with $\varepsilon\neq0$, so $\varepsilon$ is an eigenvector and $\lambda$ is an eigenvalue. If $b=0$ as well, then $Q-\lambda=0$ and the kernel is the whole algebra.

Theorem (the spectrum is a singleton). For every $Q = a + b\varepsilon$,

$$ \sigma(Q) = \{a\}. $$

Proof. By the criterion, $\lambda$ is an eigenvalue exactly when $a-\lambda=0$, that is exactly when $\lambda=a$. The eigenvector equation at $\lambda=a$ is $b\varepsilon v = 0$, which has non-zero solutions for every $b$, as the proposition shows.

The Characteristic Polynomial

The operator is a scalar plus a nilpotent, and the nilpotent contributes nothing to the determinant.

Proposition (the operator form). The operator $L_\varepsilon$ satisfies $L_\varepsilon^2 = 0$ and

$$ L_Q = a\,\mathrm{id} + bL_\varepsilon. $$

Proof. For $v = x + y\varepsilon$,

$$ L_Q(v) = (a + b\varepsilon)(x + y\varepsilon) = ax + (ay + bx)\varepsilon, $$

while $a v + bL_\varepsilon(v) = a(x + y\varepsilon) + b x\varepsilon = ax + (ay + bx)\varepsilon$. For the square, $L_\varepsilon^2(v) = \varepsilon(\varepsilon v) = \varepsilon^2 v = 0$.

Theorem (the characteristic polynomial). In the basis $(1,\varepsilon)$ and with column vectors, the matrix of $L_Q$ is

$$ [L_Q]_{(1,\varepsilon)} = \begin{pmatrix} a & 0 \\ b & a \end{pmatrix}, $$

and

$$ p_Q(X) = \det\bigl(XI - [L_Q]_{(1,\varepsilon)}\bigr) = (X-a)^2. $$

Moreover

$$ \operatorname{tr}L_Q = 2a, \qquad \det L_Q = a^2 = \langle Q\rangle, \qquad \operatorname{disc}p_Q = 0. $$

Proof. The first column is the image of $1$, namely $a + b\varepsilon$, and the second is the image of $\varepsilon$, namely $a\varepsilon$; this gives the matrix. It is upper triangular with diagonal entries $a$, so its characteristic polynomial is $(X-a)^2$. The trace is the sum of the diagonal entries, the determinant is their product, and the discriminant of $(X-a)^2$ is $4a^2 - 4a^2 = 0$.

Corollary (the eigenspace). For every $Q$,

$$ \dim_R E_a(Q) = \begin{cases} 2, & b = 0, \\ 1, & b\neq0. \end{cases} $$

Proof. By the theorem, $E_a(Q) = \ker(bL_\varepsilon)$. If $b=0$ the kernel is everything. If $b\neq0$, then $bL_\varepsilon(x + y\varepsilon) = bx\varepsilon$, which vanishes exactly when $x=0$, so the kernel is the line $R\varepsilon = \mathrm{M}$.

Remark (the matrix conventions). The matrix $\begin{pmatrix}a&b\\0&a\end{pmatrix}$ often written for $L_Q$ is the matrix in the reversed basis $(\varepsilon,1)$, the matrix-element representation $\mathsf{M}_2(Q)$ of $M_2(R)$ Triangular Representation, and the matrix of left multiplication in the reversed basis. In the column convention and the basis $(1,\varepsilon)$ it is the transpose $\begin{pmatrix}a&0\\b&a\end{pmatrix}$, the regular representation of Dual-Numbers Linear Representations. The characteristic polynomial, trace, determinant and Jordan type are the same in all readings, since the matrices are similar or transposes.

The Theorem: Universal Spectral Degeneracy

Spectral regularity is the separation of the two roots of the characteristic polynomial; in $\mathbb{C}_{\mathrm{dln},R}$ there are never two roots to separate.

Definition (spectrally regular, spectrally degenerate). An element $Q$ is spectrally regular if its characteristic polynomial is squarefree, equivalently if it has two distinct eigenvalues in $R$. It is spectrally degenerate if its characteristic polynomial has a repeated root, equivalently if its discriminant vanishes.

Theorem (every element is spectrally degenerate). For every $Q\in\mathbb{C}_{\mathrm{dln},R}$, the characteristic polynomial is $(X-a)^2$, of discriminant zero. Hence every element is spectrally degenerate, and there is no spectrally regular element.

Proof. The characteristic polynomial was computed in the preceding section; a square has a repeated root and discriminant zero.

Definition (the sectors). For $i,j\in\{0,1\}$, let $\Sigma_{ij}$ be the set of elements with $\langle Q\rangle\neq0$ if $i=1$ and $\langle Q\rangle=0$ if $i=0$, and with $Q$ spectrally regular if $j=1$ and spectrally degenerate if $j=0$. Thus

$$ \Sigma_{11} = \{\langle Q\rangle\neq0,\ p_Q \text{ squarefree}\}, \qquad \Sigma_{10} = \{\langle Q\rangle\neq0,\ p_Q \text{ not squarefree}\}, $$

and similarly for $\Sigma_{01}$ and $\Sigma_{00}$.

Corollary (the norm/spectrum classification). In $\mathbb{C}_{\mathrm{dln},R}$,

$$ \Sigma_{11} = \varnothing, \qquad \Sigma_{01} = \varnothing, \qquad \Sigma_{10} = \{a\neq0\} = (\mathbb{C}_{\mathrm{dln},R})^\times, \qquad \Sigma_{00} = \mathrm{M} = \{a=0\}. $$

Proof. Both sets with $j=1$ are empty by the theorem. The condition $\langle Q\rangle\neq0$ is $a^2\neq0$, which over the real closed field $R$ is $a\neq0$; this is also the unit criterion. Its complement is $a=0$, namely $\mathrm{M}$.

Remark (degenerate is not defective). The theorem says that the eigenvalue is repeated; it does not say that the operator is not diagonalisable. The real axis $b=0$ consists of scalar operators $a\,\mathrm{id}$, which are diagonalisable, while $b\neq0$ is not. The word degenerate is fixed by the discriminant; defective is fixed by the minimal polynomial. The biquaternion cone draws the same distinction; see The Spectrum of a Biquaternion.

The Minimal Polynomial Dichotomy

The minimal polynomial is the first invariant that sees the defect, and it takes exactly two forms.

Definition (minimal polynomial). The minimal polynomial $m_Q$ of $Q$ is the monic polynomial of least degree in $R[X]$ with $m_Q(Q)=0$. Equivalently, it is the monic generator of the kernel of the evaluation map $R[X]\to\mathbb{C}_{\mathrm{dln},R}$, $X\mapsto Q$.

Theorem (the dichotomy). For $Q = a + b\varepsilon$,

$$ m_Q(X) = \begin{cases} X - a, & b = 0, \\ (X-a)^2, & b\neq0. \end{cases} $$

Proof. Since $Q-a = b\varepsilon$, one has $(Q-a)^2 = b^2\varepsilon^2 = 0$. If $b=0$ then $Q-a=0$, so $X-a$ annihilates $Q$ and is of degree one, hence is minimal. If $b\neq0$ then $Q-a\neq0$, so no polynomial of degree one annihilates $Q$; the polynomial $(X-a)^2$ does, and it is therefore minimal.

Corollary (semisimple and defective). The following are equivalent:

  1. $Q$ is semisimple, that is, $L_Q$ is diagonalisable over $R$;
  2. $m_Q$ is squarefree;
  3. $b = 0$;
  4. $Q$ lies on the real axis $R\cdot1$.

The complement, $b\neq0$, is the defective set. It is the complement of the real axis, not the maximal ideal: the maximal ideal $\mathrm{M}\setminus\{0\}$ is the subset of the defective set on which the eigenvalue is zero. The defect is carried by the nilpotent part $Q-a = b\varepsilon$, and $\mathrm{M}\setminus\{0\}$ is the set of possible non-zero nilpotent parts.

Proof. The equivalence of (2) and (3) is the theorem, since $X-a$ is squarefree and $(X-a)^2$ is not. For (1) and (2), a linear operator is diagonalisable exactly when its minimal polynomial is squarefree and splits, and both polynomials here split. For (3) and (4), $b=0$ is the real axis by definition. The final description follows because $Q-a = b\varepsilon$ lies in $\mathrm{M}$ and is non-zero exactly when $b\neq0$.

The Jordan Form

Off the real axis the operator is a single Jordan block, and the block is visible in a basis adapted to the nilpotent part.

Theorem (the Jordan form). Let $b\neq0$. In the basis $(\varepsilon,\,b^{-1})$ of $\mathbb{C}_{\mathrm{dln},R}$, the matrix of $L_Q$ is

$$ [L_Q]_{(\varepsilon,b^{-1})} = \begin{pmatrix} a & 1 \\ 0 & a \end{pmatrix} = J_2(a). $$

If $b=0$, the matrix is the scalar matrix $aI$ in every basis.

Proof. For $b\neq0$,

$$ L_Q(\varepsilon) = a\varepsilon, \qquad L_Q(b^{-1}) = \frac{a + b\varepsilon}{b} = \frac{a}{b} + \varepsilon = a\,b^{-1} + \varepsilon. $$

The first column is therefore $(a,0)$ and the second is $(1,a)$, which is the displayed matrix. For $b=0$, $L_Q = a\,\mathrm{id}$.

The One-Term Spectral Decomposition

A spectral decomposition separates distinct eigenvalues by orthogonal idempotents; with one eigenvalue and no non-trivial idempotent, it can have only one term.

Definition (spectral idempotent, spectral decomposition). A non-zero idempotent $e\in\mathbb{C}_{\mathrm{dln},R}$ is a spectral idempotent of $Q$ if $Qe = eQ = ae$ for some scalar $a$. A spectral decomposition of $Q$ is a finite sum

$$ Q = \sum_{i=1}^{r}\lambda_i e_i, $$

where the $\lambda_i$ are distinct scalars, the $e_i$ are non-zero idempotents with $e_ie_j=0$ for $i\neq j$, and $\sum_i e_i = 1$.

Proposition (the idempotents). Over a field, the only idempotents of $\mathbb{C}_{\mathrm{dln},R}$ are $0$ and $1$.

Proof. Let $e = x + y\varepsilon$ and impose $e^2=e$:

$$ x^2 + 2xy\varepsilon = x + y\varepsilon, $$

so $x^2=x$ and $2xy=y$. In a field $x=0$ or $x=1$. If $x=0$ then $y=0$; if $x=1$ then $2y=y$, so $y=0$ because $2$ is invertible. Hence $e=0$ or $e=1$.

Theorem (the one-term decomposition). A spectral decomposition of $Q$ exists if and only if $b=0$. In that case it is the one-term decomposition

$$ Q = a\cdot1, $$

and it is unique. No element of $\mathbb{C}_{\mathrm{dln},R}$ has a two-term spectral decomposition.

Proof. A spectral idempotent must be $1$, by the proposition, and then $Q = a\cdot1 = a$, so $b=0$. Conversely, if $b=0$ then $Q=a$ and $Q=a\cdot1$ is a one-term decomposition; it is unique because the only non-zero idempotent is $1$ and the scalar is then $Q$ itself. A two-term decomposition would require two distinct eigenvalues, which the spectrum does not contain.

Remark (the Jordan–Chevalley decomposition). Every element has the additive decomposition

$$ Q = a\cdot1 + b\varepsilon, $$

with $a\cdot1$ semisimple, $b\varepsilon$ nilpotent, and the two commuting. It is the Jordan–Chevalley decomposition of Nilpotent Elements and the Jordan–Chevalley Decomposition, and it is unique. It is not a spectral decomposition: the nilpotent part is not separated by an idempotent, and it is non-zero exactly on the defective set. The spectral decomposition is the special case $b=0$, where the nilpotent part vanishes.

Contrast with the Complex, Split-Complex, Quaternionic and Biquaternionic Algebras

In the neighbouring algebras the discriminant is the norm of the pure part; in the dual algebra the pure part lies in the radical of the norm, and the discriminant vanishes with it.

The spectral polynomial of an element is the monic quadratic whose roots are its eigenvalues. In each of the algebras below it has the form

$$ p(X) = (X-a)^2 + \langle\pi\rangle, $$

where $a$ is the scalar part, $\pi$ is the pure part, and $\langle\pi\rangle$ is the norm of $\pi$ in that algebra. The discriminant is then

$$ \operatorname{disc}p = -4\langle\pi\rangle. $$

The values are collected in the table; the dual row is the last.

algebra element pure part $\pi$ norm $\langle\pi\rangle$ spectral polynomial discriminant eigenvalues
$\mathbb{C}$ $a+bi$ $bi$ $b^2$ $(X-a)^2+b^2$ $-4b^2$ $a\pm ib$
$\mathbb{C}_{\mathrm{spl}}=\mathbb{R}[j]$ $a+bj$ $bj$ $-b^2$ $(X-a)^2-b^2$ $4b^2$ $a\pm b$
$\mathbb{H}$ $q_t+\vec q$ $\vec q$ $\lvert\vec q\rvert^2$ $(X-q_t)^2+\lvert\vec q\rvert^2$ $-4\lvert\vec q\rvert^2$ $q_t\pm i\lvert\vec q\rvert$
$\mathbb{B}$ $Q_t+\vec Q$ $\vec Q$ $B^2$ $(X-Q_t)^2+B^2$ $-4B^2$ $Q_t\pm iB$
$\mathbb{C}_{\mathrm{dln}}$ $a+b\varepsilon$ $b\varepsilon$ $0$ $(X-a)^2$ $0$ $a$ (double)

Here $\mathbb{C}_{\mathrm{spl}}=\mathbb{R}[j]$ with $j^2=1$ and norm $a^2-b^2$; $\mathbb{H}$ is the quaternion algebra with norm $q_t^2+\lvert\vec q\rvert^2$; $\mathbb{B}$ is the biquaternion algebra with $B^2=Q_x^2+Q_y^2+Q_z^2$ and spectral polynomial as in The Spectrum of a Biquaternion; and $\mathbb{C}_{\mathrm{dln}}$ is the dual algebra, where $\langle b\varepsilon\rangle = b^2\varepsilon^2 = 0$.

Theorem (the discriminant is the norm of the pure part). In each row of the table, the spectral polynomial is $(X-a)^2+\langle\pi\rangle$ and its discriminant is $-4\langle\pi\rangle$.

Proof. The complex and split-complex rows are the ordinary quadratic formulae for the matrices of multiplication by $a+bi$ and $a+bj$. The quaternion row is the spectral polynomial of Quaternion Spectral Theory, and the biquaternion row is that of The Spectrum of a Biquaternion, with pure part $\vec q$ or $\vec Q$ and its norm. The dual row is the characteristic polynomial of the preceding section, with pure part $b\varepsilon$ and norm $\langle b\varepsilon\rangle=b^2\varepsilon^2=0$.

Corollary (spectral regularity in the five algebras). In $\mathbb{C}$, $\mathbb{C}_{\mathrm{spl}}$, $\mathbb{H}$ and $\mathbb{B}$ the spectral-regular locus is the complement of the vanishing locus of a non-zero quadratic form in the pure part; it is non-empty and, over $\mathbb{R}$ or $\mathbb{C}$, open and dense. In $\mathbb{C}_{\mathrm{dln}}$ the spectral-regular locus is empty.

Proof. The discriminant is $-4\langle\pi\rangle$, and the pure-part norm is a non-zero quadratic form in the first four algebras. A non-zero quadratic form is a non-zero polynomial function, so it takes a non-zero value at some point, and its zero set is closed with empty interior; hence its complement is non-empty, open and dense. In $\mathbb{C}_{\mathrm{dln}}$ the pure part has norm zero, so the discriminant vanishes identically; this is the theorem of §The Theorem: Universal Spectral Degeneracy.

Summary

The dual-number algebra $\mathbb{C}_{\mathrm{dln},R}$ has a single spectral phenomenon: every $Q=a+b\varepsilon$ has $p_Q=(X-a)^2$, discriminant zero, spectrum $\{a\}$, and eigenspace of dimension two when $b=0$ and one when $b\neq0$. Every element is spectrally degenerate, so $\Sigma_{11}=\Sigma_{01}=\varnothing$, $\Sigma_{10}=\{a\neq0\}$ and $\Sigma_{00}=\mathrm{M}$.

Spectral degeneracy is not defect. The minimal polynomial is $X-a$ on the real axis and $(X-a)^2$ off it, so the real axis is the semisimple stratum and $b\neq0$ is the defective set; the defect is carried by the nilpotent part $b\varepsilon\in\mathrm{M}$, and $\mathrm{M}\setminus\{0\}$ is the subset with eigenvalue zero, not the whole defective set. Off the real axis the operator is $J_2(a)$ in the basis $(\varepsilon,b^{-1})$, the eigenspace is $R\varepsilon$, and there is no eigenbasis. The only idempotents are $0$ and $1$, so the only spectral decomposition is the one-term $Q=a\cdot1$ of the real axis; the additive Jordan–Chevalley decomposition $Q=a\cdot1+b\varepsilon$ always exists but is not spectral.

The comparison with the neighbouring algebras is the comparison of discriminants: in $\mathbb{C}$, $\mathbb{C}_{\mathrm{spl}}$, $\mathbb{H}$ and $\mathbb{B}$ the spectral polynomial is $(X-a)^2+\langle\pi\rangle$ with discriminant $-4\langle\pi\rangle$, so the spectral-regular locus is non-empty; in $\mathbb{C}_{\mathrm{dln}}$ the pure part $b\varepsilon$ has norm zero, the discriminant vanishes identically, and the spectral-regular locus is empty.

Summary of Notation

symbol meaning
$\mathbb{C}_{\mathrm{dln},R}=R[\varepsilon]/(\varepsilon^2)$ the dual-number algebra over $R$
$\mathbb{C}_{\mathrm{dln}}=\mathbb{C}_{\mathrm{dln},\mathbb{R}}$ the dual-number algebra over $\mathbb{R}$
$Q=a+b\varepsilon$ a general element, equal to $q+\varepsilon q'$ in the prime notation
$a=\operatorname{Re}Q$, $b=\operatorname{Inf}Q$ real and infinitesimal parts
$\bar Q=a-b\varepsilon$ dual conjugation
$\langle Q\rangle=Q\bar Q=a^2$ the norm, the dual norm; degenerate of rank one, radical $\mathrm{M}$
$\mathrm{M}=(\varepsilon)=\{a=0\}$ the maximal ideal, the vanishing locus of the norm
$L_Q(v)=Qv$, $R_Q(v)=vQ$ left and right multiplication; equal because the algebra is commutative
$p_Q(X)=(X-a)^2$ the characteristic polynomial of $L_Q$
$\operatorname{disc}p_Q=0$ the discriminant; spectral degeneracy is universal
$m_Q$ the minimal polynomial: $X-a$ if $b=0$, $(X-a)^2$ if $b\neq0$
$E_\lambda(Q)=\ker L_{Q-\lambda}$ the eigenspace of $\lambda$
$\sigma(Q)=\{a\}$ the spectrum
$J_2(a)=\begin{pmatrix}a&1\\0&a\end{pmatrix}$ the Jordan block off the real axis
$R\cdot1=\{b=0\}$ the real axis, the semisimple stratum
$\{b\neq0\}$ the defective set
$\Sigma_{ij}$ the norm/spectrum sectors: $i$ for the norm, $j$ for spectral regularity
$\mathsf{M}_2(Q)=\begin{pmatrix}a&b\\0&a\end{pmatrix}$ the matrix-element representation, the transpose convention of the regular representation

Further Reading