Biquaternion Fourier Inversion on the Degenerate Set: Peirce Coordinates and the Exponential
Introduction
This article inverts the biquaternion Fourier transform at the frequencies where the symbol is not an invertible element, and it does so inside the algebra. The work is done inside the algebra and nowhere else: no two-by-two array appears, no array of operators is written, and no statement is carried by a determinant or an adjugate of an array. The objects are elements of $\mathbb{B}$, the operator of left multiplication, and subspaces of the algebra read as a module over itself. The word coordinates is used in one sense only, the four Peirce components of an element, and those components stay in their four subspaces and are never assembled into an array.
Status. §The Algebra, the Conjugate and the Norm and §The Degenerate Set and the Rank of Left Multiplication are proved from the eight-coordinate product; the determinant of the regular representation and the three values of the rank are quoted from The Fourier Transform and the Symbol of a Differential Operator, where they are proved. §Inversion in the Algebra is proved, quoting Plancherel with its constant from The Inverse Fourier Transform of a Symbol. §Two Generalised Inverses is proved, the existence of the Moore–Penrose inverse in a finite-dimensional inner-product space being standard and quoted as such. §The Exponential on the Degenerate Set quotes the three closed forms of Biquaternion Elementary Functions with their hypotheses and proves everything else. §Inversion When the Symbol Is Nowhere Invertible is proved for the two families whose null element is fixed up to a scalar, and the remaining case is labelled open with its obstacle named. No part of the article is a programme.
The equation. Let $P(D)=\sum_{\lvert\alpha\rvert\le m}a_\alpha D^\alpha$ have constant coefficients in $\mathbb{B}$ and symbol $\sigma_P(\zeta)=\sum a_\alpha(i\zeta)^\alpha$, with $\zeta\in\mathbb{R}^n$ and $n$ the number of variables, as in The Fourier Transform and the Symbol of a Differential Operator. The equation to invert is
$$ P(D)f=g , $$
understood in $L^2(\mathbb{R}^n,\mathbb{B})$, with $P(D)$ the multiplier by $\sigma_P$. The inversion is pointwise in $\zeta$, so the algebra and the number of variables are separate questions.
Boundaries. The zero set, the determinant of the regular representation and the rank on the null set are The Fourier Transform and the Symbol of a Differential Operator; the placement of the kernel is The Three Kernel Placements of the Biquaternion Fourier Transform, and this article uses the left placement throughout, so the symbol multiplies on the left; the Peirce vocabulary is Special Element and Peirce Conventions, which fixes no pair of idempotents and prefers none; the forms that measure an element are Norm and Unitary Conventions and Product and Form Conventions; the exponential is Biquaternion Elementary Functions.
The Algebra, the Conjugate and the Norm
The Product, the Conjugate and the Norm
Write $Q=Q_t+Q_x\vec e_x+Q_y\vec e_y+Q_z\vec e_z$ with complex coefficients, the product being fixed by $\vec e_x\vec e_y=\vec e_z$, $\vec e_y\vec e_z=\vec e_x$, $\vec e_z\vec e_x=\vec e_y$ and $\vec e_k^{2}=-e_t$ for the three units, with the scalar $i$ central. The conjugate is
$$ \tilde Q=Q_t-Q_x\vec e_x-Q_y\vec e_y-Q_z\vec e_z , $$
and the algebraic norm is $N(Q)=Q\tilde Q$. Both are elements of the commutative algebra generated by $Q_t$ and the vector part $Q_v$, which is why the next statement has no order condition.
Theorem (the conjugate and the norm). For all $Q,R\in\mathbb{B}$ the conjugate reverses products, $\widetilde{QR}=\tilde R\tilde Q$, and
$$ Q\tilde Q=\tilde QQ=N(Q)e_t,\qquad N(QR)=N(Q)N(R),\qquad N(e_t)=1 . $$
Proof. Both $Q$ and $\tilde Q$ lie in the commutative subalgebra generated by $Q_t$ and $Q_v$, so the two products agree and equal $Q_t^{2}-Q_v^{2}$. Now $Q_v=Q_x\vec e_x+Q_y\vec e_y+Q_z\vec e_z$ has $Q_v^{2}=-(Q_x^{2}+Q_y^{2}+Q_z^{2})e_t$, because the three units anticommute and each has second power $-e_t$; hence $Q_t^{2}-Q_v^{2}=(Q_t^{2}+Q_x^{2}+Q_y^{2}+Q_z^{2})e_t=N(Q)e_t$. The reversal is read off the three product rules, which change sign when the order is exchanged. For multiplicativity, $N(QR)e_t=(QR)\widetilde{(QR)}=QR\tilde R\tilde Q=QN(R)e_t\tilde Q=N(R)N(Q)e_t$, and $N(e_t)=1$ is the definition. The norm is central and $\mathbb{C}$-valued: it is a complex number times $e_t$, and it is not the coefficient norm of §The Projection Carried by a Null Symbol.
The Invertibility Criterion
Theorem (an invertible element is exactly an element of nonzero norm). $\sigma\in\mathbb{B}$ is invertible if and only if $N(\sigma)\neq0$, and then
$$ \sigma^{-1}=\frac{\tilde\sigma}{N(\sigma)} . $$
Proof. If $N(\sigma)\neq0$ the displayed element is an inverse, by the theorem above. If $\sigma$ has an inverse $\tau$, then $N(\sigma)N(\tau)=N(\sigma\tau)=N(e_t)=1$, so $N(\sigma)\neq0$. The inverse is the conjugate divided by the norm, and no determinant of any array enters: the numerator is an element of the algebra and the denominator is a complex number.
The Quadratic Identity and the Two Halves
Theorem (the identity of an element with its two coefficients). For every $Q\in\mathbb{B}$,
$$ Q^{2}-2Q_tQ+N(Q)e_t=0 . $$
Proof. Write $Q=Q_t+Q_v$ and expand: $Q^{2}=Q_t^{2}+2Q_tQ_v+Q_v^{2}$. The last term is $-(Q_x^{2}+Q_y^{2}+Q_z^{2})e_t=-(N(Q)-Q_t^{2})e_t$, and substituting $Q_v=Q-Q_t$ in the middle term gives the identity.
Corollary (the two halves of the degenerate set). Let $\sigma\neq0$ with $N(\sigma)=0$. Then exactly one of the following holds: either $\sigma_t\neq0$ and $\sigma=2\sigma_t\Pi$ with $\Pi$ a nonzero idempotent, or $\sigma_t=0$ and $\sigma^{2}=0$.
Proof. The identity reads $\sigma^{2}=2\sigma_t\sigma$ when $N(\sigma)=0$. If $\sigma_t\neq0$, put $\Pi=\sigma/(2\sigma_t)$; then $\Pi^{2}=\sigma^{2}/(4\sigma_t^{2})=2\sigma_t\sigma/(4\sigma_t^{2})=\Pi$, and $\Pi\neq0$ because $\sigma\neq0$. If $\sigma_t=0$ the same display reads $\sigma^{2}=0$.
Remark (where the cut is made, and how the identity is read). The identity is the minimal polynomial of $\sigma$ over the commutative algebra it generates, of degree at most two, and the cut between the two halves is the scalar part, which is half the trace of left multiplication. On the first half the element is a complex multiple of an idempotent and the two eigenvalues of left multiplication are $2\sigma_t$ and $0$; on the second both eigenvalues vanish. No central idempotent makes that cut: a split into two algebras belongs to the split-biquaternion layer and not here.
The Degenerate Set and the Rank of Left Multiplication
The Set, and What Is Quoted From the Sibling Article
Definition (the degenerate set and the rank). For $\sigma\in\mathbb{B}$ let $L_\sigma$ be left multiplication, $X\mapsto\sigma X$, on the eight real dimensions of $\mathbb{B}$, and let $\operatorname{rank}L_\sigma$ be its rank as a real linear map. The degenerate set of a symbol is
$$ Z=\{\zeta\in\mathbb{R}^n : N(\sigma_P(\zeta))=0\} . $$
Proposition (the set is the set of non-invertibles). $Z=\{\zeta:\sigma_P(\zeta)\ \text{is not invertible}\}=\{\zeta:\det L_{\sigma_P(\zeta)}=0\}$.
Proof. The first equality is the unit criterion of §The Invertibility Criterion. The second is the identity $\det L_Q=\lvert N(Q)\rvert^{4}$ of The Fourier Transform and the Symbol of a Differential Operator, proved there from the regular representation; a determinant vanishes exactly when the map is singular.
Quoted facts, with their source. From the same article, and used below without recomputation: $\det L_Q=\lvert N(Q)\rvert^{4}$, so that the degree of $\det L_{\sigma_P}$ in $\zeta$ is $dm$ for a symbol of order $m$ in $d=8$ real dimensions; and
$$ \operatorname{rank}L_\sigma= \begin{cases} 8, & N(\sigma)\neq0,\\ 4, & N(\sigma)=0,\ \sigma\neq0,\\ 0, & \sigma=0 . \end{cases} $$
Remark (the modulus, and the reason it is a modulus). The determinant is a real determinant and the fourth power removes the sign of $N$, so the two agree for a real symbol and differ by a sign for a complex one. The rank statement is the one used below, and the determinant enters only through the degree count.
The Peirce Corners
Definition (the Peirce decomposition of one idempotent). Let $\Pi$ be an idempotent, $\Pi^{2}=\Pi$, $\Pi\neq0$, and put $f=e_t-\Pi$, which is an idempotent with $\Pi f=f\Pi=0$ and $\Pi+f=e_t$. Every $X\in\mathbb{B}$ has one expansion
$$ X=\Pi X\Pi+\Pi Xf+fX\Pi+fXf , $$
whose four terms lie in the four corners $\Pi\mathbb{B}\Pi$, $\Pi\mathbb{B}f$, $f\mathbb{B}\Pi$ and $f\mathbb{B}f$.
Proposition (each corner is a complex line). The diagonal corners are $\Pi\mathbb{B}\Pi=\mathbb{C}\Pi$ and $f\mathbb{B}f=\mathbb{C}f$, and the two mixed corners are complex lines, so that the Peirce coordinates of $X$ are four complex numbers.
Proof. A nonzero idempotent of $\mathbb{B}$ is primitive, so $\Pi\mathbb{B}\Pi$ is a division algebra over $\mathbb{C}$ by Primitive Idempotents and Minimal Left Ideals, and being finite-dimensional over the algebraically closed field $\mathbb{C}$ it is $\mathbb{C}\Pi$; the same argument gives the other diagonal corner. The corner $\Pi\mathbb{B}f$ is a module over $\Pi\mathbb{B}\Pi=\mathbb{C}\Pi$, hence a complex vector space; it is nonzero, since $\Pi\mathbb{B}f=0$ would give $\Pi\mathbb{B}=\Pi\mathbb{B}\Pi=\mathbb{C}\Pi$ of real dimension two, against $\operatorname{rank}L_\Pi=4$ for the null idempotent $\Pi$; and the real dimension of the two mixed corners together is $8-2-2=4$, so each is one complex line.
Remark (how far the word coordinates reaches). The four numbers are the components in the four lines, they are what the title of the article calls Peirce coordinates, and they are never written as an array and never multiplied as one: the product of two elements is read from the Peirce rule, which sends a product of corner elements to the corner named by the two indices, and everything below is stated in the algebra. No pair of idempotents is preferred, by Special Element and Peirce Conventions, and the one idempotent used in an example is declared where it is used.
The Image and the Kernel on Each Half
Theorem (the idempotent half). Let $\sigma=2\sigma_t\Pi$ with $\sigma_t\neq0$ and $\Pi$ a nonzero idempotent. Then
$$ \operatorname{im}L_\sigma=2\sigma_t\,\Pi\mathbb{B},\qquad \ker L_\sigma=f\mathbb{B},\qquad \mathbb{B}=\operatorname{im}L_\sigma\oplus\ker L_\sigma , $$
both summands of real dimension four, and $L_\sigma/(2\sigma_t)$ is the idempotent operator $L_\Pi$.
Proof. $\sigma X=2\sigma_t\Pi X$ gives the image, and $\Pi X=0$ is equivalent to $X=fX$, which gives the kernel. The two are complementary because $X=\Pi X+fX$ and the two terms lie in the two subspaces; the dimensions are four each by the quoted rank. Finally $L_\Pi^{2}=L_{\Pi^{2}}=L_\Pi$.
Theorem (the nilpotent half). Let $\sigma\neq0$ with $\sigma^{2}=0$. Then
$$ \operatorname{im}L_\sigma=\ker L_\sigma=\sigma\mathbb{B}, $$
both of real dimension four, and $L_\sigma$ is not diagonalisable.
Proof. $\sigma(\sigma X)=\sigma^{2}X=0$ gives $\sigma\mathbb{B}\subseteq\ker L_\sigma$. The image has real dimension four by the quoted rank, and the kernel has real dimension $8-4=4$ by rank and nullity, so the two subspaces coincide. For the last claim, $L_\sigma^{2}=0$ while $L_\sigma\neq0$, so the minimal polynomial is $t^{2}$ and the operator is not diagonalisable.
Remark (the whole difference between the halves, in one line). On the idempotent half the image and the kernel are complementary, and the operator behaves as a projection after a scalar; on the nilpotent half they coincide, and the operator carries the image into the kernel. That is the precise sense in which the second half is the degenerate one, and it is read off the product alone.
The Two-Quaternion Form, and the Dimension of the Set
Proposition (one real quaternion multiplied by an invertible element). Write $Q_\mu=a_\mu+ib_\mu$ with $a_\mu,b_\mu$ real and put $q_0=a_\mu e_\mu$, $q_1=b_\mu e_\mu$, so that $Q=q_0+iq_1$ with real quaternions $q_0,q_1$. Then
$$ N(Q)=\lvert q_0\rvert^{2}-\lvert q_1\rvert^{2}+2i\langle q_0,q_1\rangle , $$
so $Q$ is null exactly when $\lvert q_0\rvert=\lvert q_1\rvert$ and $\langle q_0,q_1\rangle=0$; and every nonzero null element has the form $Q=q_0(e_t+ir)$ with $r$ a real pure unit, that is $Q=2q_0\Pi_r$ with $\Pi_r=\tfrac12(e_t+ir)$ an idempotent.
Proof. Expanding $N(Q)=\sum_\mu(a_\mu+ib_\mu)^{2}$ gives the two parts as stated. If $q_0\neq0$ the two conditions give that $r=q_0^{-1}q_1$ is real, pure and of unit norm; and $q_0=0$ forces $q_1=0$. Finally $(e_t+ir)^{2}=e_t+2ir-r^{2}=2(e_t+ir)$, since $r^{2}=-e_t$.
Remark (the dimension, and the one thing the reading does not say). The parameters are four in $q_0$ and two in $r$, so the null set has real dimension six in eight, the codimension two of The Fourier Transform and the Symbol of a Differential Operator. The reading is a factorisation and not a product decomposition: $q_0$ and $r$ do not commute in general, and $Q=2q_0\Pi_r$ is a left multiple of an idempotent, not a scalar multiple of one.
Inversion in the Algebra
The Inverse Is the Conjugate Over the Norm
Corollary (the inversion formula off the set). Off $Z$ the symbol is invertible and the solution is
$$ f=\mathcal{F}^{-1}\bigl[\sigma^{-1}\mathcal{F}[g]\bigr],\qquad \sigma^{-1}(\zeta)=\frac{\tilde\sigma_P(\zeta)}{N(\sigma_P(\zeta))} , $$
where $\mathcal{F}$ is the transform of The Inverse Fourier Transform of a Symbol and the product is the product in $\mathbb{B}$.
Proof. The unit criterion gives the second formula; the first is the inversion formula of The Fourier Transform and the Symbol of a Differential Operator applied to the multiplier $\sigma^{-1}$, valid under the integrability of $\sigma^{-1}\mathcal{F}[g]$; and $\mathcal{F}[P(D)f]=\sigma_P\mathcal{F}[f]$ is the intertwining of that article.
Remark (the numerator is an element, and the denominator is a complex number). In a coordinate model the same statement reads as an adjugate over a determinant, and the adjugate is an array. Here the numerator is $\tilde\sigma$, the conjugate of the element, which is a polynomial in $\zeta$ of degree at most $m$ with coefficients in $\mathbb{B}$, and the denominator is $N(\sigma_P)$, of degree $dm$. The order of the pole is the multiplicity of the factor in $N(\sigma_P)$ and not the degree of the determinant; that distinction is the subject of Inversion of the Biquaternion Fourier Transform: Poles, Tails and the Degenerate Set, and the numerator contributes no pole at all, since $\tilde\sigma$ does not vanish where $\sigma$ does not.
What the Set Costs
Theorem (existence holds, uniqueness and boundedness fail). Suppose $N(\sigma_P)$ is not identically zero. Then $Z$ has Lebesgue measure zero and for every $g\in L^2$ there is $f\in L^2$ with $P(D)f=g$ almost everywhere. The solution is not unique: at each point of $Z$ the ambiguity is the kernel of $L_{\sigma_P(\zeta)}$, of real dimension four. The solution of least norm is
$$ \mathcal{F}[f](\zeta)=\sigma_P(\zeta)^{+}\mathcal{F}[g](\zeta) , $$
with $\sigma^{+}$ the Moore–Penrose inverse of §The Moore–Penrose Inverse; and the inverse is unbounded whenever $\sigma$ is not bounded away from zero.
Proof. $N(\sigma_P)$ is a polynomial, so if it is not identically zero its zero set has empty interior and measure zero. Off $Z$ set $\mathcal{F}[f]=\sigma^{-1}\mathcal{F}[g]$, and on $Z$ set it to zero: the result is in $L^2$ and solves the equation, so a solution exists for every $g$. At a point of $Z$ the solutions differ by the kernel, of real dimension four by the theorems above, and the least-norm choice is the one in the orthogonal complement of the kernel, which is the Moore–Penrose solution. Boundedness fails because $\sigma^{+}=\sigma^{*}/(2\lVert\sigma\rVert^{2})$ grows like $1/\lVert\sigma\rVert$ as $\sigma\to0$.
Remark (the phrase inversion on the set is usually read wrongly). A set of measure zero cannot obstruct the solvability of an $L^2$ equation, since the values of $\mathcal{F}[f]$ on $Z$ are irrelevant to the norm and to the transform. What the set obstructs is uniqueness and boundedness. The genuine obstruction is the case $N(\sigma_P)\equiv0$, treated in §Inversion When the Symbol Is Nowhere Invertible.
Two Generalised Inverses
The Projection Carried by a Null Symbol
Definition (the coefficient norm). For $X\in\mathbb{B}$ put $\lVert X\rVert^{2}=\sum_\mu\lvert X_\mu\rvert^{2}$, the sum of the eight second powers of the moduli, and write $X^{*}=\overline{\tilde X}$ for the star, quaternion conjugation followed by complex conjugation.
Theorem (a null symbol carries an Hermitian idempotent). Let $\sigma\neq0$ with $N(\sigma)=0$, written $\sigma=2q_0\Pi_r$ as in §The Two-Quaternion Form, and the Dimension of the Set, and put $s=q_0r\tilde q_0/\lvert q_0\rvert^{2}$. Then $s$ is a real pure unit, $\Pi_\sigma=\tfrac12(e_t+is)$ is an Hermitian idempotent, and
$$ \sigma\sigma^{*}=2\lVert\sigma\rVert^{2}\Pi_\sigma,\qquad \Pi_\sigma\sigma=\sigma,\qquad \operatorname{im}L_{\Pi_\sigma}=\operatorname{im}L_\sigma . $$
Proof. The element $s$ is the conjugate of the pure unit $r$ by the unit $q_0/\lvert q_0\rvert$, so it is a real pure unit; hence $(e_t+is)^{2}=2(e_t+is)$ and $\Pi_\sigma$ is an idempotent, and it is Hermitian because $i^{*}=-i$ and $s^{*}=-s$ give $\Pi_\sigma^{*}=\Pi_\sigma$. For the product, $\sigma^{*}=\tilde q_0+ir\tilde q_0$, since the conjugate reverses the order and $r$ is real; then
$$ \sigma\sigma^{*}=2\lvert q_0\rvert^{2}+2iq_0r\tilde q_0=2\lvert q_0\rvert^{2}(e_t+is)=4\lvert q_0\rvert^{2}\Pi_\sigma , $$
and $\lVert\sigma\rVert^{2}=\lvert q_0\rvert^{2}+\lvert q_1\rvert^{2}=2\lvert q_0\rvert^{2}$ gives the first display. Multiplying by $\sigma$ on the right of it and using $\lVert\sigma\rVert^{2}\sigma=2\lvert q_0\rvert^{2}\sigma$ gives $\Pi_\sigma\sigma=\sigma$, and then $\operatorname{im}L_\sigma\subseteq\operatorname{im}L_{\Pi_\sigma}$, with equality of the real dimensions four and four by the quoted rank.
Corollary (the three products that carry the inverse). With $\sigma$ null and nonzero,
$$ \sigma\sigma^{*}\sigma=2\lVert\sigma\rVert^{2}\sigma,\qquad \sigma^{*}\sigma\sigma^{*}=2\lVert\sigma\rVert^{2}\sigma^{*},\qquad (\sigma\sigma^{*})^{*}=\sigma\sigma^{*} . $$
Proof. The first is the theorem multiplied by $\sigma$ on the right; the second is the star of the first, using that $\lVert\sigma\rVert^{2}$ is real; the third is immediate from the definition of the star.
The Moore–Penrose Inverse
Definition (the Moore–Penrose inverse). For a linear map between finite-dimensional inner-product spaces the Moore–Penrose inverse is the unique map $T^{+}$ with
$$ TT^{+}T=T,\qquad T^{+}TT^{+}=T^{+},\qquad (TT^{+})^{*}=TT^{+},\qquad (T^{+}T)^{*}=T^{+}T , $$
and its existence and uniqueness in that setting are standard and quoted as such.
Theorem (the inverse on the degenerate set). Let $\sigma\neq0$ with $N(\sigma)=0$. Then the Moore–Penrose inverse of $L_\sigma$ is left multiplication by
$$ \sigma^{+}=\frac{\sigma^{*}}{2\lVert\sigma\rVert^{2}} . $$
Proof. Four equations, each read in the algebra. The first is the corollary above; the second is its star; the third holds because $\sigma\sigma^{*}$ is Hermitian and $2\lVert\sigma\rVert^{2}$ is real; the fourth holds because $(\sigma^{+}\sigma)^{*}=\sigma^{*}\sigma/(2\lVert\sigma\rVert^{2})$ and $\sigma^{*}\sigma$ is Hermitian. The four hold as identities in $\mathbb{B}$ and not merely on the operator, since left multiplication is injective.
Corollary (the two products are the two projections). On the degenerate set, $\sigma\sigma^{+}=\Pi_\sigma$ is the orthogonal projection onto $\operatorname{im}L_\sigma$, and $\sigma^{+}\sigma$ is the orthogonal projection onto the orthogonal complement of $\ker L_\sigma$.
Proof. The first is the theorem above. For the second, $\sigma^{+}\sigma=\sigma^{*}\sigma/(2\lVert\sigma\rVert^{2})$ is Hermitian and idempotent by the same two computations, and its kernel is $\ker L_\sigma$.
Remark (the inverse is not the algebraic inverse off the set, and the two must not be confused). For the unit $\sigma=2e_t+i\vec e_x$ the algebraic inverse is $\tilde\sigma/3=(2e_t-i\vec e_x)/3$ while the element $\sigma^{*}/(2\lVert\sigma\rVert^{2})$ is $\sigma/10$. The second is the Moore–Penrose inverse of the operator and not an inverse of the element, and on the degenerate set the distinction disappears, since there the first does not exist and the second does.
The Drazin Inverse
Definition (the Drazin inverse). For $T$ with index $k$, the least $k$ with $\operatorname{rank}T^{k+1}=\operatorname{rank}T^{k}$, the Drazin inverse $T^{D}$ is the unique map commuting with $T$ with $T^{D}TT^{D}=T^{D}$ and $T^{k+1}T^{D}=T^{k}$.
Theorem (the Drazin inverse on the two halves). On the idempotent half, with $\sigma=c\Pi$, $\Pi^{2}=\Pi$ and $c\neq0$, $\sigma^{D}=\Pi/c$, of index one. On the nilpotent half $\sigma^{D}=0$, of index two.
Proof. For the first, $\sigma^{2}\sigma^{D}=c^{2}\Pi\cdot\Pi/c=c\Pi=\sigma$, so the index is one and the defining equations hold with $\sigma^{D}$ commuting with $\sigma$. For the second, $\sigma^{2}=0$ gives index two, since $\operatorname{rank}L_{\sigma}=4$ and $\operatorname{rank}L_{\sigma}^{2}=0$; the element $\sigma^{D}=0$ satisfies the three equations at $k=2$, and uniqueness of the Drazin inverse gives the claim.
Remark (which inverse answers which question). The Moore–Penrose inverse uses the star, so it answers the least-norm question and its multiplier is $\sigma^{*}/2\lVert\sigma\rVert^{2}$, which reflects the frequency whenever the star does. The Drazin inverse uses only the product, so it answers the algebraic question and its multiplier is again a function of the symbol alone; the price is that it kills the nilpotent direction. The star of the symbol at $\zeta$ is the symbol of the operator with starred coefficients at $-\zeta$, so a Moore–Penrose inversion reflects the argument and a Drazin inversion does not; the choice is a choice of purpose and not a correction of one by the other.
The Exponential on the Degenerate Set
The Three Closed Forms
The three closed forms are quoted, not proved here. They are those of Biquaternion Elementary Functions, §The Closed Form, in the form of its table; the quadratic identity of §The Quadratic Identity and the Two Halves is what makes the spectral reduction apply to every $S\in\mathbb{B}$, since $\mathbb{C}[S]$ is then spanned by $e_t$ and $S$ itself. With $B^{2}=S_x^{2}+S_y^{2}+S_z^{2}$ and $\vec U=\vec S/B$ they read
$$ \exp S=\begin{cases}\exp(S_t)\bigl(\cos B\,e_t+\sin B\,\vec U\bigr), & B^{2}\neq0,\\[2pt] \exp(S_t)\bigl(e_t+\vec S\bigr), & B^{2}=0,\ \vec S\neq0,\\[2pt] \exp(S_t)\,e_t, & \vec S=0.\end{cases} $$
Proposition (each closed form is invertible). In all three cases $N(\exp S)=\exp(2S_t)$, which is never zero.
Proof. The norm is multiplicative, so $N(\exp S)=\exp(2S_t)N(\exp\vec S)$ and it remains to compute the second factor. In the first case $N(\vec U)=-1$ for a pure unit and the closed form gives $\cos^{2}B+\sin^{2}B=1$. In the second case $\vec S^{2}=0$, so $N(e_t+\vec S)=(e_t+\vec S)(e_t-\vec S)=e_t$, the conjugate of a pure element being its negative. In the third case the factor is $N(e_t)=1$.
No Nonzero Null Element Is an Exponential
Theorem (the degenerate set is outside the image of the exponential). Let $\sigma\neq0$ with $N(\sigma)=0$. Then there is no $S\in\mathbb{B}$ with $\exp S=\sigma$.
Proof. The image of the exponential is the group of invertibles, a statement of Biquaternion Elementary Functions whose hypothesis is that the element be one biquaternion and whose proof is the surjectivity of the logarithm onto the invertibles. A nonzero null element is not invertible by the unit criterion, so it is not in the image. The three closed forms give the same conclusion a second way, each of them producing an invertible element.
Remark (why the inversion is done in the algebra). The inverse kernel of the transform is an exponential because the frequency enters it through an invertible element, and the inversion of Inversion of the Biquaternion Fourier Transform: Poles, Tails and the Degenerate Set is a residue calculus in the variable that carries that exponential. At a point where the symbol is not invertible there is no exponential to take a residue of, so the calculus has no object to act on; that is not a defect of that article but the reason this one works with the element and its conjugate.
The Three Cases Against the Set
Remark (how the three cases sit, and why none of them reaches the set). The cases are cut by $B^{2}$ and not by the norm, and their intersections with $Z\setminus\{0\}$ are as follows.
| Case of the closed form | Condition | Intersection with the null set | Why the closed form misses it |
|---|---|---|---|
| $B^{2}\neq0$ | the vector part is not isotropic | $\sigma=2\sigma_t\Pi$, one eigenvalue $0$ | $\exp$ of it has eigenvalues $\exp(2\sigma_t)$ and $1$, neither zero |
| $B^{2}=0$, $\vec S\neq0$ | the vector part is isotropic and nonzero | $\sigma_t=0$, so $\sigma^{2}=0$ | $\exp$ of a nilpotent is unipotent, hence an invertible element |
| $\vec S=0$ | the element is central | $\sigma=0$ only | the scalar exponential never vanishes |
Corollary (the nilpotent ray is carried to a line of invertibles). For $\vec S\neq0$ with $\vec S^{2}=0$ and $S_t=0$ the element $\vec S$ is nilpotent and $\exp\vec S=e_t+\vec S$ is unipotent: An invertible element, the identity plus the nilpotent. The exponential therefore never returns to the nilpotent half, which is the exact sense in which its parametrisation cannot reach the half on which the inversion has to be done.
Inversion When the Symbol Is Nowhere Invertible
The Two Model Families
The case that is genuinely obstructed is $N(\sigma_P)\equiv0$. The symbol is then null at every frequency, the pointwise system is singular everywhere, and the range is a proper subspace at every point.
Theorem (a fixed null coefficient). Let $S\neq0$ be null and $\sigma_P(\zeta)=\mu(\zeta)S$ with $\mu$ a scalar polynomial vanishing only at $\zeta=0$. Then $P(D)f=g$ has a solution exactly when
$$ SS^{+}g=g , $$
that is, exactly when $g(x)$ lies in the image of left multiplication by $S$ for almost every $x$; and the solution of least norm is $f=\mathcal{F}^{-1}\bigl[\mu^{+}S^{+}\mathcal{F}[g]\bigr]$.
Proof. The pointwise system reads $\mu(\zeta)L_S\mathcal{F}[f]=\mathcal{F}[g]$. For $\zeta\neq0$ the scalar is nonzero, so solvability is $\mathcal{F}[g]\in\operatorname{im}L_S$, which is equivalent to $SS^{+}\mathcal{F}[g]=\mathcal{F}[g]$ because $SS^{+}$ is the orthogonal projection onto that image by §The Moore–Penrose Inverse. Both sides are constant in $\zeta$ except through $\mathcal{F}[g]$, so the condition is a condition on $g$ alone, and it is exactly $SS^{+}g=g$. The least-norm solution at each frequency is $\mu^{+}S^{+}\mathcal{F}[g]$, and $\mu^{+}$ is $1/\mu$ off the origin and $0$ at the origin.
Example (the nilpotent family). Let $N_0=\vec e_x+i\vec e_y$, which is nilpotent, and $P(D)=N_0\Delta$. Its symbol is $\sigma_P(\zeta)=-\lvert\zeta\rvert^{2}N_0$, null at every $\zeta$, so $\operatorname{im}L_{N_0}$ is a real four-dimensional subspace at every frequency and the condition is that $g$ take values in it. The projection is computed from the formulas above:
$$ N_0^{+}=-\frac{N_0^{*}}{4},\qquad N_0N_0^{+}=\tfrac12\bigl(e_t+i\vec e_z\bigr). $$
Example (the idempotent family). Let $P(D)=\Pi\Delta$ with $\Pi=\tfrac12(e_t+i\vec e_z)$, an Hermitian idempotent. Its symbol is $-\lvert\zeta\rvert^{2}\Pi$, null everywhere, and by §The Moore–Penrose Inverse the generalised inverse of the element is $\Pi$ itself, so the condition is $\Pi g=g$: the data must take values in the image of left multiplication by $\Pi$ at every point. The solution of least norm is $f=-\Delta^{-1}\Pi g$, the ordinary inversion followed by a projection.
Remark (the two examples are one computation). In both the condition is $SS^{+}g=g$, and the two differ only in which four-dimensional subspace the projection selects: the nilpotent one reaches its subspace through the second power of the element, the idempotent one is already the projection. That is the same difference as the one between image-equals-kernel and image-complementary-kernel in §The Image and the Kernel on Each Half.
What Remains Open
The open case, stated precisely. Let $N(\sigma_P)\equiv0$ with the null element $\sigma_P(\zeta)$ not of the form $\mu(\zeta)S$ for a fixed null $S$. The image of $L_{\sigma_P(\zeta)}$ is then a four-dimensional subspace depending on $\zeta$, and solvability requires $\mathcal{F}[g](\zeta)$ to lie in it for almost every $\zeta$. The following are open here.
- Whether the pointwise conditions are simultaneously satisfiable by the transform of one function, that is, whether the range of $P(D)$ is closed in $L^2$.
- Whether the projection $\sigma_P(\zeta)\sigma_P(\zeta)^{+}$ onto that image varies measurably in $\zeta$, and whether the resulting family defines a bounded projection on $L^2$.
- The kernel of $P(D)$ when the null element varies, which is infinite-dimensional in the two model families and undescribed in general.
Remark (why the constant-coefficient case is not a formality). In the two families the projection is constant, so the pointwise conditions assemble into one condition on $g$. When the null element varies, so does the projection, and a family of projections that does not commute with the multiplier need not integrate to a projection on $L^2$; that is the obstacle, not a gap in technique.
Remark (what the algebra does and does not settle). Every statement about a single frequency is a statement about one element and its conjugate, settled completely above. It does not turn a family of frequencies into one problem in the algebra: the elements $\sigma_P(\zeta)$ vary with $\zeta$, and the projections at the null frequencies need not be the idempotents met above.
Summary
The inversion is carried out in the algebra, on elements and their conjugates. The conjugate gives the inverse and the norm gives the denominator: for $\sigma$ an invertible element, $\sigma^{-1}=\tilde\sigma/N(\sigma)$, and the quadratic identity $\sigma^{2}-2\sigma_t\sigma+N(\sigma)e_t=0$ cuts the degenerate set into two halves, a nonzero-trace half where the element is twice its scalar part times an idempotent and a zero-trace half where the second power vanishes.
On the idempotent half the image and the kernel of left multiplication are complementary, of four real dimensions each; on the nilpotent half they coincide, and the operator is not diagonalisable. That coincidence is the whole of the degeneracy, and it is read off the product. In Peirce coordinates the four components of an element are four complex numbers in four complex lines, and the corners are never assembled into an array.
The generalised inverse on the set is a single formula, $\sigma^{+}=\sigma^{*}/(2\lVert\sigma\rVert^{2})$, valid on both halves, and it is the Moore–Penrose inverse of left multiplication: its two products $\sigma\sigma^{+}$ and $\sigma^{+}\sigma$ are the orthogonal projections onto the image and onto the orthogonal complement of the kernel. The Drazin inverse answers the other question and kills the nilpotent direction. When the norm is not identically zero the set costs uniqueness and boundedness but not existence, because it has measure zero.
The route through the exponential closes before it reaches the set: each of the three closed forms produces an invertible element, so no nonzero null element is an exponential, and a residue calculus has no object at a null frequency.
When the norm vanishes identically the obstruction is real: the range is proper at every frequency, the condition is $SS^{+}g=g$, and the inversion is a scalar inversion after a fixed projection. That is proved for the two families whose null element is fixed up to a scalar; a null element varying with the frequency is open, with the failure of the pointwise projections to integrate named as the obstacle.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $\mathbb{B}$ | The biquaternion algebra, $\mathbb{C}\otimes_{\mathbb{R}}\mathbb{H}$, of real dimension $8$ |
| $e_t$ | The scalar unit |
| $\vec e_x,\vec e_y,\vec e_z$ | The three vector units, each with second power $-e_t$ |
| $i$ | The central scalar imaginary |
| $d$ | The real dimension of the algebra, $8$ |
| $Q_t$ | The scalar part of $Q$; $Q_v$ its vector part |
| $\tilde Q$ | The conjugate, $Q_t-Q_v$, an anti-automorphism |
| $Q^{*}$ | The star, $\overline{\tilde Q}$, quaternion conjugation followed by complex conjugation |
| $N$ | The algebraic norm $Q\tilde Q$, central and $\mathbb{C}$-valued, equal to $\det L_Q$ up to the fourth power |
| $\lVert X\rVert$ | The coefficient norm, the root of the sum of the eight second powers of the moduli |
| $L_\sigma$ | Left multiplication, $X\mapsto\sigma X$ |
| $\det L_\sigma$ | The determinant of the regular representation, equal to $\lvert N(\sigma)\rvert^{4}$ |
| $\Pi$ | A nonzero idempotent, and $f=e_t-\Pi$ its complement |
| $\Pi_\sigma$ | The Hermitian idempotent carried by a null $\sigma$, $\tfrac12(e_t+is)$ |
| $\Pi_r$ | The idempotent $\tfrac12(e_t+ir)$ carried by a pure unit $r$ |
| $s$ | The real pure unit $q_0r\tilde q_0/\lvert q_0\rvert^{2}$ |
| $q_0,q_1$ | The two real quaternions of $Q=q_0+iq_1$ |
| $r$ | A real pure unit quaternion, $r^{2}=-e_t$ |
| $Z$ | The degenerate set, $\{\zeta:N(\sigma_P(\zeta))=0\}$ |
| $P(D)$ | A constant-coefficient operator with coefficients in $\mathbb{B}$ |
| $\sigma_P$, $\sigma$ | Its symbol, $\sum a_\alpha(i\zeta)^\alpha$ |
| $m$ | The total order of $P(D)$ |
| $\zeta$ | The frequency, in $\mathbb{R}^n$ |
| $n$ | The number of variables |
| $\mathcal{F}$, $\mathcal{F}^{-1}$ | The biquaternion Fourier transform and its inverse, left placement |
| $f,g$ | The unknown and the data, $L^2$ functions with values in $\mathbb{B}$ |
| $\sigma^{+}$ | The Moore–Penrose inverse, $\sigma^{*}/(2\lVert\sigma\rVert^{2})$ on the set |
| $\sigma^{D}$ | The Drazin inverse |
| $T^{+}$, $T^{D}$ | The Moore–Penrose and Drazin inverses of a linear map |
| $k$ | The index of a linear map, the least $k$ with $\operatorname{rank}T^{k+1}=\operatorname{rank}T^{k}$ |
| $S$, $N_0$ | Fixed nonzero null elements used as model coefficients |
| $\mu$ | A scalar polynomial multiplier in the model families |
| $\Delta$ | The Laplacian, the scalar factor of the two families |
| $B$ | The value with $B^{2}=S_x^{2}+S_y^{2}+S_z^{2}$ in the closed forms |
| $\vec U$ | The normalised vector part $\vec S/B$ when $B^{2}\neq0$ |
| $\mathbb{H}$ | The real quaternion algebra |
| $\lvert\cdot\rvert$ | The quaternion norm on $q_0,q_1$; the modulus on $\mathbb{C}$ |
| $\langle q_0,q_1\rangle$ | The real inner product of the four coefficients |
Further Reading
- The Fourier Transform and the Symbol of a Differential Operator, for the operator and its symbol, the determinant $\lvert N\rvert^{4}$ of the regular representation, the three values of the rank, and the codimension of the null cone.
- The Inverse Fourier Transform of a Symbol, for the transform and its glyph, for Plancherel with its constant, and for the distinction between the pole order and the degree.
- Inversion of the Biquaternion Fourier Transform: Poles, Tails and the Degenerate Set, for the residue inversion in a plane of the root, which is the route that closes before the degenerate set.
- Biquaternion Elementary Functions, for the three closed forms of the exponential and for the image of the exponential, both quoted with hypotheses.
- The Three Kernel Placements of the Biquaternion Fourier Transform, for the left placement, which transposes every formula of §Inversion in the Algebra.
- Special Element and Peirce Conventions, for the Peirce vocabulary and for the statement that no pair of idempotents is preferred.
- Primitive Idempotents and Minimal Left Ideals and Off-Diagonal Peirce Components and Noncentral Idempotents, for the division property of a diagonal corner and the product rule of the corners.
- Norm and Unitary Conventions and Product and Form Conventions, for the forms of the algebra and their letters, and for why the positive definite form and the isotropic norm must not be identified.