Introduction
The guiding question in this paper is the following. Given octonions λ 1 , λ 2 \lambda_1, \lambda_2 λ 1 , λ 2 , when does the equation
h ( w ) = ( w λ 1 ) λ 2 , w ≠ 0 , (1) h(w) = (w\lambda_1)\lambda_2,\qquad w \ne0, \tag*{(1)} h ( w ) = ( w λ 1 ) λ 2 , w = 0 , (1)
have a solution? Here h h h is the canonical structure operator on W = O ⊗ R V W = \mathbb{O} \otimes_{\mathbb{R}} V W = O ⊗ R V , where V V V is the standard real 7-dimensional G 2 G_2 G 2 -module, and right multiplication acts on each octonion coefficient of w ∈ W w \in W w ∈ W . In this paper, we study this two-step problem, posed in [6 p. 53] . The operator h h h is defined naturally by octonion multiplication and is both G 2 G_2 G 2 -equivariant and octonion-Hermitian. Among operators of left octonionic multiplication type, these conditions give a real two-parameter family [6 Remark 3.4, (3.6), p. 13] . Its nonzero members, modulo positive scale, form an S 1 S^1 S 1 -family. For non-scalar members, the real and one-step spectra and the full two-step coefficient loci are related by invertible affine changes of coordinates, as explained in Remark 2.8; the scalar members are degenerate.
With the associator convention [ a , b , c ] = a ( b c ) − ( a b ) c [a,b,c] = a(bc) - (ab)c [ a , b , c ] = a ( b c ) − ( ab ) c ,
( w λ 1 ) λ 2 = w ( λ 1 λ 2 ) − [ w , λ 1 , λ 2 ] . (w\lambda_1)\lambda_2 = w(\lambda_1\lambda_2) - [w,\lambda_1,\lambda_2]. ( w λ 1 ) λ 2 = w ( λ 1 λ 2 ) − [ w , λ 1 , λ 2 ] .
Over an associative coefficient algebra the equation collapses to the one-step problem with right eigenvalue λ 1 λ 2 \lambda_1\lambda_2 λ 1 λ 2 .
For octonions, we ask how the associator changes the spectral locus. By Artin’s Theorem, λ 1 \lambda_1 λ 1 and λ 2 \lambda_2 λ 2 lie in an associative quaternionic subalgebra, however, the octonion coefficients of w w w need not lie in that subalgebra [4 Chapter 6 and Sections 8.1–8.5] . On the quaternionic subalgebra, successive right multiplication agrees with multiplication by λ 1 λ 2 \lambda_1\lambda_2 λ 1 λ 2 ; on its orthogonal complement, it agrees with multiplication by the reversed product λ 2 λ 1 \lambda_2\lambda_1 λ 2 λ 1 (Proposition 3.5). The operator h h h can couple these subspaces, so this description does not reduce the problem to two independent one-step equations.
Equation (1) reduces to the problem of finding the kernel of the operator h − R λ 2 R λ 1 h-R_{\lambda_2}R_{\lambda_1} h − R λ 2 R λ 1 , where R R R denotes the octonion right-multiplication operator on W W W . For every pair of octonions ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) this operator, in an adapted orthonormal pair of imaginary units u , v u,v u , v , is identified with a member P a P_{\mathbf a} P a of the four-coefficient pencil
P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R v R u , where R v R u : w ↦ ( w u ) v , P_{\mathbf a}=h-a_0I-a_1R_u-a_2R_v-a_3R_vR_u,\qquad\text{where}\qquad R_vR_u:w\mapsto(wu)v, P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R v R u , where R v R u : w ↦ ( w u ) v ,
at a point a = ( a 0 , a 1 , a 2 , a 3 ) ∈ R 4 \mathbf a=(a_0,a_1,a_2,a_3)\in\mathbb{R}^4 a = ( a 0 , a 1 , a 2 , a 3 ) ∈ R 4 (Proposition 3.2 ), and every coefficient point a \mathbf a a is realized by a pair ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) (Proposition 3.3).
Write x = a 1 2 + a 2 2 x=a_1^2+a_2^2 x = a 1 2 + a 2 2 , y = a 3 2 y=a_3^2 y = a 3 2 and N = a 0 2 + x + y N=a_0^2+x+y N = a 0 2 + x + y . Then, for a pair, a 0 = Re ( λ 1 λ 2 ) a_0=\operatorname{Re}(\lambda_1\lambda_2) a 0 = Re ( λ 1 λ 2 ) , N = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 N=|\lambda_1|^2|\lambda_2|^2 N = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 , and y y y is the Gram determinant of the imaginary parts. The reduced point a \mathbf a a depends on the frame, while a 0 a_0 a 0 , N N N and y y y are frame-independent. Our results therefore apply to every pair at its reduced point.
The stabilizer of the quaternionic frame ( u , v , u v ) (u,v,uv) ( u , v , uv ) is K ≅ SU ( 2 ) K\cong\operatorname{SU}(2) K ≅ SU ( 2 ) . Its action decomposes W W W into three invariant sectors W 1 ≅ 1 ⊕ 16 W_1\cong1^{\oplus16} W 1 ≅ 1 ⊕ 16 , W 3 = 3 ⊕ 4 W_3=3^{\oplus4} W 3 = 3 ⊕ 4 , W 4 = 4 ⊕ 7 W_4=4^{\oplus7} W 4 = 4 ⊕ 7 , of real dimensions 16, 12 and 28, respectively (Theorem 3.6 ), and 1 1 1 , 3 3 3 , 4 4 4 refer to real irreducible representations of SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) .
Theorem 4.10 gives the polynomial identity
det 56 P a = N 6 Q 14 ( a 0 , x , y ) ( N + 4 a 0 ) 3 Q 12 ( a 0 , x , y ) 2 . \det_{56}P_{\mathbf a}=N^6Q_{14}(a_0,x,y)(N+4a_0)^3Q_{12}(a_0,x,y)^2. 56 det P a = N 6 Q 14 ( a 0 , x , y ) ( N + 4 a 0 ) 3 Q 12 ( a 0 , x , y ) 2 .
where Q 14 , Q 12 Q_{14},Q_{12} Q 14 , Q 12 are explicit polynomials defined in § 5.1 and § 5.3 . Theorem 4.11 identifies the real singular locus as Z ∪ S Z\cup S Z ∪ S , where
Z = { a ∈ R 4 : Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) = 0 } , Z=\{\mathbf a\in\mathbb{R}^4:Q_{14}(a_0,a_1^2+a_2^2,a_3^2)=0\}, Z = { a ∈ R 4 : Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) = 0 } ,
S = { a ∈ R 4 : ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 = 4 } . S=\{\mathbf a\in\mathbb{R}^4:(a_0+2)^2+a_1^2+a_2^2+a_3^2=4\}. S = { a ∈ R 4 : ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 = 4 } .
Their form for every pair is given in Corollary 4.13 . The sphere arises from the reversed-product sector W 3 W_3 W 3 . The quaternionic-type sector W 4 W_4 W 4 contributes no additional real singular points, while the fixed sector W 1 W_1 W 1 supplies the locus Z Z Z .
The real and one-step spectra were determined in [6 ] : the former is finite, and the latter is described in each complex slice by a quartic and a circle (§ 3 , Figure 1 ). To the best of the author’s knowledge, the sector determinant formulas and the resulting characterization of the real two-step locus are new.
Figure 1. The level-1 slice locus L \mathcal{L} L in the ( μ , Q ) (\mu,Q) ( μ , Q ) -plane, with equal scales: the quartic p 1 ( μ , Q ) = 0 p_1(\mu,Q)=0 p 1 ( μ , Q ) = 0 (solid) and the circle ( μ + 2 ) 2 + Q 2 = 4 (\mu+2)^2+Q^2=4 ( μ + 2 ) 2 + Q 2 = 4 (dashed). The dots on the line Q = 0 Q=0 Q = 0 are the real eigenvalues − 7 , − 4 , 0 , r ± -7,-4,0,r_{\pm} − 7 , − 4 , 0 , r ± of h h h . Since Q Q Q is signed, L \mathcal{L} L is even in Q Q Q . Compare [6 Figure 1] , where r ± r_{\pm} r ± is written μ ± \mu_{\pm} μ ± .
Sections 2 and 3 recall the lower levels and set up the pencil and sectors. Section 4 proves Theorems 4.10 and 4.11, and Section 5 states the remaining questions. The quaternionic-type coordinate and matrix calculation is given in Section 4.3. Appendix A supplies the algebraic rules and branching, and Appendix B gives the fixed-sector Schur reduction.
Acknowledgments
Codex and Claude Code were used to organize and verify supporting research materials, to assist with Python scripts that were used for computations, and for assistance with typesetting. ChatGPT was used for mathematical brainstorming and editorial assistance with the exposition. The main ideas are the author’s own and all mathematical results have been verified by the author, who takes full responsibility for the content.
Background and the one-step spectrum
Let us first introduce the conventions and the notation.
Definition 2.1 (Conventions). We use the conventions of [6, §2.1, (2.1)–(2.3), p. 6; Remark 5.1, p. 23; (5.4), p. 25]. Let O \mathbb{O} O be the octonions, with conjugation p ↦ p ˉ p \mapsto\bar{p} p ↦ p ˉ and inner product ⟨ p , r ⟩ = Re ( p r ˉ ) \langle p,r\rangle= \operatorname{Re}(p\bar{r}) ⟨ p , r ⟩ = Re ( p r ˉ ) , and let g g g be the restriction of this inner product to Im O \operatorname{Im}\mathbb{O} Im O . For a , b ∈ Im O a,b \in\operatorname{Im}\mathbb{O} a , b ∈ Im O , the cross product a × b a \times b a × b is defined by a b = − g ( a , b ) 1 + a × b ab = -g(a,b)1 + a \times b ab = − g ( a , b ) 1 + a × b . The associator [ a , b , c ] [a,b,c] [ a , b , c ] of a , b , c ∈ O a,b,c \in\mathbb{O} a , b , c ∈ O is defined by
( a b ) c = a ( b c ) − [ a , b , c ] . (2) (ab)c = a(bc) - [a,b,c]. \tag*{(2)} ( ab ) c = a ( b c ) − [ a , b , c ] . (2) Finally, G 2 = Aut ( O ) G_2 = \operatorname{Aut}(\mathbb{O}) G 2 = Aut ( O ) , and for a unit n ^ ∈ Im O \hat{n} \in\operatorname{Im}\mathbb{O} n ^ ∈ Im O the complex structure J n ^ J_{\hat{n}} J n ^ on n ^ ⊥ ⊂ Im O \hat{n}^{\perp} \subset\operatorname{Im}\mathbb{O} n ^ ⊥ ⊂ Im O is
J n ^ ( p ) = n ^ × p . J_{\hat{n}}(p) = \hat{n} \times p. J n ^ ( p ) = n ^ × p .
Definition 2.2. Following [6 §2.3] , fix an oriented Euclidean 7-space V V V and a G 2 G_2 G 2 -equivariant isometry V ≅ V ^ : = Im O V \cong\hat{V} := \operatorname{Im}\mathbb{O} V ≅ V ^ := Im O . We distinguish corresponding vectors by writing v ∈ V v \in V v ∈ V and v ^ ∈ V ^ \hat{v} \in\hat{V} v ^ ∈ V ^ . Set W : = O ⊗ R V W := \mathbb{O} \otimes_{\mathbb{R}} V W := O ⊗ R V , a real vector space of dimension 56.
The product metric and diagonal G 2 G_2 G 2 -action on W W W are
⟨ p ⊗ v , r ⊗ v ′ ⟩ W = ⟨ p , r ⟩ O g ( v , v ′ ) , γ ⋅ ( p ⊗ v ) = γ ( p ) ⊗ γ ( v ) , \begin{aligned}
\langle p \otimes v, r \otimes v'\rangle_W &= \langle p,r\rangle_{\mathbb{O}} g(v,v'),\\
\gamma\cdot(p \otimes v) &= \gamma(p) \otimes\gamma(v),
\end{aligned} ⟨ p ⊗ v , r ⊗ v ′ ⟩ W γ ⋅ ( p ⊗ v ) = ⟨ p , r ⟩ O g ( v , v ′ ) , = γ ( p ) ⊗ γ ( v ) , where G 2 G_2 G 2 acts on V V V through the fixed isometry.
All octonion products are formed in the first tensor factor. For a ∈ O a \in\mathbb{O} a ∈ O , define
L a ( p ⊗ v ) = ( a p ) ⊗ v and R a ( p ⊗ v ) = ( p a ) ⊗ v . L_a(p \otimes v) = (ap) \otimes v \qquad\text{and}\qquad R_a(p \otimes v) = (pa) \otimes v. L a ( p ⊗ v ) = ( a p ) ⊗ v and R a ( p ⊗ v ) = ( p a ) ⊗ v . We write w a : = R a w wa := R_a w w a := R a w for w ∈ W w \in W w ∈ W . Thus successive right multiplication is formed coefficient by coefficient, in the order ( w λ 1 ) λ 2 = R λ 2 R λ 1 w (w\lambda_1)\lambda_2 = R_{\lambda_2}R_{\lambda_1}w ( w λ 1 ) λ 2 = R λ 2 R λ 1 w .
Fix corresponding orthonormal bases { e i } \{e_i\} { e i } of V V V and { e ^ i } \{\hat{e}_i\} { e ^ i } of V ^ \hat{V} V ^ . Write w = ∑ j w ^ j ⊗ e j w = \sum_j \hat{w}_j \otimes e_j w = ∑ j w ^ j ⊗ e j , with w ^ j ∈ O \hat{w}_j \in\mathbb{O} w ^ j ∈ O ; identity (2) applies to each coefficient. In the expansion w ^ j = w 0 j 1 + ∑ i w i j e ^ i \hat{w}_j = w_{0j}1 + \sum_i w_{ij}\hat{e}_i w ^ j = w 0 j 1 + ∑ i w ij e ^ i , the first index of w i j w_{ij} w ij belongs to the octonion factor and the second to V V V .
The imaginary-coefficient part V ^ ⊗ V ⊂ W \hat{V} \otimes V \subset W V ^ ⊗ V ⊂ W is therefore identified with V ⊗ V V \otimes V V ⊗ V , with the octonion index first. This identifies tensor subspaces such as Λ 7 2 V \Lambda^2_7 V Λ 7 2 V and S 0 2 V S^2_0 V S 0 2 V with subspaces of W W W . As a G 2 G_2 G 2 -module,
W ≅ 1 ⊕ 7 ⊕ 7 ⊕ 14 ⊕ 27 , W \cong\mathbf{1} \oplus\mathbf{7} \oplus\mathbf{7} \oplus\mathbf{14} \oplus\mathbf{27}, W ≅ 1 ⊕ 7 ⊕ 7 ⊕ 14 ⊕ 27 , where bold numerals denote irreducible representations of G 2 G_2 G 2 of the indicated dimensions. The summands 1 \mathbf{1} 1 , Λ 7 2 V \Lambda^2_7 V Λ 7 2 V , 14 \mathbf{14} 14 , and 27 \mathbf{27} 27 lie in V ^ ⊗ V ≅ V ⊗ V = S 2 V ⊕ Λ 2 V \hat{V} \otimes V \cong V \otimes V = S^2 V \oplus\Lambda^2 V V ^ ⊗ V ≅ V ⊗ V = S 2 V ⊕ Λ 2 V and have imaginary O \mathbb{O} O -coefficients, whereas 1 ⊗ V \mathbf{1} \otimes V 1 ⊗ V has real O \mathbb{O} O -coefficients.
Definition 2.3. With the bases fixed above, let w = ∑ j w ^ j ⊗ e j w=\sum_j\widehat{w}_j\otimes e_j w = ∑ j w j ⊗ e j for w ^ j ∈ O \widehat{w}_j\in\mathbb{O} w j ∈ O , and define, as in [6 §3.1, Proposition 3.3 and Remark 3.4] , h ∈ End R ( W ) h\in\operatorname{End}_{\mathbb{R}}(W) h ∈ End R ( W ) by
h ( w ) = ∑ i , j ( ( e ^ i e ^ j ) w ^ j ) ⊗ e i . (3) h(w)=\sum_{i,j}((\widehat{e}_i\widehat{e}_j)\widehat{w}_j)\otimes e_i. \tag*{(3)} h ( w ) = i , j ∑ (( e i e j ) w j ) ⊗ e i . (3)
In (2.2) the product h i j : = e ^ i e ^ j h_{ij}:=\widehat{e}_i\widehat{e}_j h ij := e i e j is formed first and then multiplies w ^ j \widehat{w}_j w j on the left; the second index of h i j h_{ij} h ij is contracted against the V V V -index of w w w , and the first is the output V V V -index.
The operator h h h does not depend on the choice of corresponding orthonormal bases [6 §3.1] . The octonion product also determines the adjoint and equivariance properties of h h h .
Proposition 2.4 [6 Proposition 3.2] . The operator h h h is self-adjoint with respect to ⟨ ⋅ , ⋅ ⟩ W \langle\cdot,\cdot\rangle_W ⟨ ⋅ , ⋅ ⟩ W and G 2 G_2 G 2 -equivariant.
Definition 2.5. We call h ( w ) = μ w h(w)=\mu w h ( w ) = μ w (μ ∈ R \mu\in\mathbb{R} μ ∈ R ) the level-0 problem, h ( w ) = w λ h(w)=w\lambda h ( w ) = w λ (λ ∈ O \lambda\in\mathbb{O} λ ∈ O ) the level-1 or one-step problem, and h ( w ) = ( w λ 1 ) λ 2 h(w)=(w\lambda_1)\lambda_2 h ( w ) = ( w λ 1 ) λ 2 (λ 1 , λ 2 ∈ O \lambda_1,\lambda_2\in\mathbb{O} λ 1 , λ 2 ∈ O ) the level-2 or two-step problem. Hence, the level-1 right spectrum of h h h is Spec r i g h t ( h ) = { λ ∈ O : ker ( h − R λ ) ≠ 0 } \operatorname{Spec}_{\mathrm{right}}(h)=\{\lambda\in\mathbb{O}:\ker(h-R_\lambda)\ne0\} Spec right ( h ) = { λ ∈ O : ker ( h − R λ ) = 0 } . An octonion λ \lambda λ is a right eigenvalue , with eigenvector w ≠ 0 w\ne0 w = 0 , when h ( w ) = w λ h(w)=w\lambda h ( w ) = w λ .
Proposition 2.6 (Level-0 real spectrum [6 Lemma 4.2 and Theorem 4.3] ). The action of h h h on the isotypic components of W W W is summarized in the following table:
Table 2.
On the 7 ⊕ 7 7\oplus7 7 ⊕ 7 component, h h h acts through a 2 × 2 2\times2 2 × 2 matrix with characteristic polynomial μ 2 − μ − 8 \mu^2-\mu-8 μ 2 − μ − 8 . Consequently,
Spec R ( h ) = { − 7 , − 4 , 0 , r − , r + } , \operatorname{Spec}_{\mathbb{R}}(h)=\{-7,-4,0,r_-,r_+\}, Spec R ( h ) = { − 7 , − 4 , 0 , r − , r + } ,
with multiplicities 1 , 14 , 27 , 7 , 7 1,14,27,7,7 1 , 14 , 27 , 7 , 7 , respectively.
For a unit n ∈ V n\in V n ∈ V with image n ^ \widehat{n} n , the slice is C n ^ = R ⊕ R n ^ \mathbb{C}_{\widehat{n}}=\mathbb{R}\oplus\mathbb{R}\widehat{n} C n = R ⊕ R n , in which each λ \lambda λ is uniquely μ + Q n ^ \mu+Q\widehat{n} μ + Q n with μ , Q ∈ R \mu,Q\in\mathbb{R} μ , Q ∈ R , and the slice pencil is H n , Q = h − Q R n ^ H_{n,Q}=h-QR_{\widehat{n}} H n , Q = h − Q R n . We take Q Q Q to be signed : ( μ , Q , n ^ ) (\mu,Q,\widehat{n}) ( μ , Q , n ) and ( μ , − Q , − n ^ ) (\mu,-Q,-\widehat{n}) ( μ , − Q , − n ) represent the same λ \lambda λ , and every slice locus is even in Q Q Q . Since R λ R_\lambda R λ is real-linear in λ \lambda λ and μ \mu μ is central, w λ = μ w + Q R n ^ w w\lambda=\mu w+QR_{\widehat{n}}w w λ = μ w + Q R n w , so
h ( w ) = w λ if and only if H n , Q w = μ w . h(w)=w\lambda\quad\text{if and only if}\quad H_{n,Q}w=\mu w. h ( w ) = w λ if and only if H n , Q w = μ w .
Let SU ( 3 ) n = Stab G 2 ( n ) \operatorname{SU}(3)_n=\operatorname{Stab}_{G_2}(n) SU ( 3 ) n = Stab G 2 ( n ) . It follows immediately [6 Proposition 5.4] that R n ^ R_{\widehat{n}} R n , and hence the slice pencil H n , Q H_{n,Q} H n , Q , is equivariant under SU ( 3 ) n ≅ SU ( 3 ) \operatorname{SU}(3)_n\cong\operatorname{SU}(3) SU ( 3 ) n ≅ SU ( 3 ) . The pencil is therefore block-diagonal for the SU ( 3 ) \operatorname{SU}(3) SU ( 3 ) -decomposition of W W W . Here the bold numerals 1 , 6 , 8 , 12 name types of SU ( 3 ) n \operatorname{SU}(3)_n SU ( 3 ) n -modules that correspond to real irreducible representations of SU ( 3 ) \operatorname{SU}(3) SU ( 3 ) of the specified dimension.
The four real module types have the following descriptions:
Here 3 3 3 is the standard complex S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -module and 3 ‾ \overline{3} 3 its conjugate. The G 2 \mathrm{G}_2 G 2 summands restrict as
7 ≅ 1 ⊕ 6 , 7 \cong1 \oplus6, 7 ≅ 1 ⊕ 6 ,
14 ≅ 8 ⊕ 6 , 14 \cong8 \oplus6, 14 ≅ 8 ⊕ 6 ,
27 ≅ 1 ⊕ 6 ⊕ 8 ⊕ 12. 27 \cong1 \oplus6 \oplus8 \oplus12. 27 ≅ 1 ⊕ 6 ⊕ 8 ⊕ 12.
Consequently,
W ≅ 1 ⊕ 4 ⊕ 6 ⊕ 4 ⊕ 8 ⊕ 2 ⊕ 12 , (4) W \cong1^{\oplus4} \oplus6^{\oplus4} \oplus8^{\oplus2} \oplus12, \tag*{(4)} W ≅ 1 ⊕ 4 ⊕ 6 ⊕ 4 ⊕ 8 ⊕ 2 ⊕ 12 , (4)
where the two copies of 8 8 8 come from 14 14 14 and 27 27 27 . Each H n , Q H_{n,Q} H n , Q preserves every isotypic summand. The four types are irreducible over R \mathbb{R} R [9 Section 2, pp. 58–60] . Their real or complex types are given in [5 Section 26.3, Proposition 26.24] . The complex irreducible modules underlying 6 6 6 , 8 8 8 and 12 12 12 are described in [7, Section 6.1; proof of Proposition 6.17; p. 147 and Section 6.9, Exercise 8; Section 6.5].
Proposition 2.7 (The one-step spectrum [6 ] ). Fix a unit n ∈ V n \in V n ∈ V . We then have the following.
On the 1 ⊕ 4 1^{\oplus4} 1 ⊕ 4 summand, H n , Q − μ I H_{n,Q} - \mu I H n , Q − μ I is a real 4 × 4 4 \times4 4 × 4 operator whose determinant is
p 1 ( μ , Q ) = μ 4 + 6 μ 3 + ( 2 Q 2 − 15 ) μ 2 + ( 6 Q 2 − 56 ) μ + Q 4 + Q 2 , (5) p_1(\mu,Q) = \mu^4 + 6\mu^3 + (2Q^2 - 15)\mu^2 + (6Q^2 - 56)\mu+ Q^4 + Q^2, \tag*{(5)} p 1 ( μ , Q ) = μ 4 + 6 μ 3 + ( 2 Q 2 − 15 ) μ 2 + ( 6 Q 2 − 56 ) μ + Q 4 + Q 2 , (5) and the block contributes a right eigenvalue μ + Q n ^ \mu+ Q\hat{n} μ + Q n ^ exactly when p 1 ( μ , Q ) = 0 p_1(\mu,Q) = 0 p 1 ( μ , Q ) = 0 .
The 8 ⊕ 2 8^{\oplus2} 8 ⊕ 2 block contributes a right eigenvalue μ + Q n ^ \mu+ Q\hat{n} μ + Q n ^ exactly when
( μ + 2 ) 2 + Q 2 = 4. (6) (\mu+ 2)^2 + Q^2 = 4. \tag*{(6)} ( μ + 2 ) 2 + Q 2 = 4. (6) For real Q ≠ 0 Q \ne0 Q = 0 , H n , Q H_{n,Q} H n , Q has no real eigenvalue on the 12 12 12 summand or on the 6 ⊕ 4 6^{\oplus4} 6 ⊕ 4 block.
Let
L = { p 1 = 0 } ∪ { ( μ + 2 ) 2 + Q 2 = 4 } ⊂ R 2 , (7) \mathcal{L} = \{p_1 = 0\} \cup\{(\mu+ 2)^2 + Q^2 = 4\} \subset\mathbb{R}^2, \tag*{(7)} L = { p 1 = 0 } ∪ {( μ + 2 ) 2 + Q 2 = 4 } ⊂ R 2 , (7) where p 1 p_1 p 1 is the quartic (5) . For each unit n ∈ V n \in V n ∈ V , the right eigenvalues of h h h in C n ^ \mathbb{C}\hat{n} C n ^ are exactly the μ + Q n ^ \mu+ Q\hat{n} μ + Q n ^ with ( μ , Q ) ∈ L (\mu,Q) \in\mathcal{L} ( μ , Q ) ∈ L . The real points Q = 0 Q = 0 Q = 0 of L \mathcal{L} L are exactly Spec R ( h ) \operatorname{Spec}_{\mathbb{R}}(h) Spec R ( h ) .
The locus L \mathcal{L} L is plotted in Figure 1 .
Remark 2.8 (The family of structure operators). As shown in [6 Remark 3.4, (3.6), p. 13] , the G 2 \mathrm{G}_2 G 2 -equivariant, octonion-Hermitian operators of the multiplication type considered here form a real two-parameter family, with coefficients
( h α , β ) i j = − α g i j 1 + β φ i j k e ^ k . (h_{\alpha,\beta})_{ij} = -\alpha g_{ij}1 + \beta\varphi_{ijk}\hat{e}_k. ( h α , β ) ij = − α g ij 1 + β φ ij k e ^ k .
The nonzero members, modulo positive scale, form an S 1 S^1 S 1 -family; one may fix α 2 + β 2 = 1 \alpha^2+\beta^2=1 α 2 + β 2 = 1 . Since h = h 1 , 1 h=h_{1,1} h = h 1 , 1 , we have
h α , β = β h + ( β − α ) I . h_{\alpha,\beta}=\beta h+(\beta-\alpha)I. h α , β = β h + ( β − α ) I .
For β ≠ 0 \beta\ne0 β = 0 , the real and one-step right spectra are therefore obtained from those of h h h by λ ↦ β λ + ( β − α ) \lambda\mapsto\beta\lambda+(\beta-\alpha) λ ↦ β λ + ( β − α ) , with the same eigenspaces. For the two-step coefficient locus, put s = β − α s=\beta-\alpha s = β − α . The pencil of Section 3 satisfies
h α , β − a 0 I − a 1 R u − a 2 R v − a 3 R v R u = β P ( a 0 − s β , a 1 β , a 2 β , a 3 β ) . h_{\alpha,\beta}-a_0I-a_1R_u-a_2R_v-a_3R_vR_u=\beta P\left(\frac{a_0-s}{\beta},\frac{a_1}{\beta},\frac{a_2}{\beta},\frac{a_3}{\beta}\right). h α , β − a 0 I − a 1 R u − a 2 R v − a 3 R v R u = β P ( β a 0 − s , β a 1 , β a 2 , β a 3 ) .
This invertible affine change preserves kernels, and every coefficient point is realizable (Proposition 3.3). It concerns the full coefficient locus, rather than a fixed-angle slice or a map of individual factor pairs. When β = 0 \beta=0 β = 0 , the operator is scalar, h α , 0 = − α I h_{\alpha,0}=-\alpha I h α , 0 = − α I , and the affine equivalence above does not apply.
Reduction of the two-step problem
We reduce the two-step equation to a pencil in four real coefficients and show that every coefficient point occurs. The adapted frame then describes the right-multiplication part on two halves of W W W and splits W W W into three invariant sectors.
Definition 3.1. Let u , v ∈ Im O u,v\in\operatorname{Im}\mathbb{O} u , v ∈ Im O be orthonormal. For a = ( a 0 , a 1 , a 2 , a 3 ) ∈ R 4 \mathbf{a}=(a_0,a_1,a_2,a_3)\in\mathbb{R}^4 a = ( a 0 , a 1 , a 2 , a 3 ) ∈ R 4 the two-step pencil is the four-parameter affine family of operators on W W W
P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R v R u , where R v R u : w ↦ ( w u ) v . (8) P_{\mathbf{a}}=h-a_0I-a_1R_u-a_2R_v-a_3R_vR_u,\qquad\text{where}\qquad R_vR_u:w\mapsto(wu)v. \tag*{(8)} P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R v R u , where R v R u : w ↦ ( w u ) v . (8) :::`
Each operator in this pencil is defined for every a ∈ R 4 a \in\mathbb{R}^{4} a ∈ R 4 ; statements about P a P_a P a without further qualification refer to this ambient coefficient space. The order R v R u R_vR_u R v R u is the composition order of Section 2, in which right multiplications act on the octonion factor of W W W .
The frame and arbitrary pairs
The orthonormal pair ( u , v ) (u,v) ( u , v ) generates the associative quaternion algebra H = ⟨ 1 , u , v , z ⟩ \mathbb{H} = \langle1,u,v,z\rangle H = ⟨ 1 , u , v , z ⟩ , where z = u v z=uv z = uv [4 Section 6.8, Theorem 2] . Let A ^ = Im H \hat{A}=\operatorname{Im}\mathbb{H} A ^ = Im H and B ^ = H ⊥ \hat{B}=\mathbb{H}^{\perp} B ^ = H ⊥ in the octonion factor, and let A , B ⊂ V A,B\subset V A , B ⊂ V be their corresponding subspaces. Thus
O = H ⊕ B ^ and V = A ⊕ B . (9) \mathbb{O}=\mathbb{H}\oplus\hat{B}\qquad\text{and}\qquad V=A\oplus B. \tag*{(9)} O = H ⊕ B ^ and V = A ⊕ B . (9)
We use the basis ( ρ ^ 0 , ρ ^ 1 , ρ ^ 2 , ρ ^ 3 ) = ( 1 , u , v , z ) (\hat{\rho}_0,\hat{\rho}_1,\hat{\rho}_2,\hat{\rho}_3)=(1,u,v,z) ( ρ ^ 0 , ρ ^ 1 , ρ ^ 2 , ρ ^ 3 ) = ( 1 , u , v , z ) of H \mathbb{H} H . Choose a unit ℓ ∈ B ^ \ell\in\hat{B} ℓ ∈ B ^ and set ξ ^ α = ρ ^ α ℓ \hat{\xi}_{\alpha}=\hat{\rho}_{\alpha}\ell ξ ^ α = ρ ^ α ℓ . The corresponding orthonormal bases are
Table 3.
The splitting of the octonion factor gives two halves of W W W :
W + = H ⊗ V and W − = B ^ ⊗ V . (10) W_{+}=\mathbb{H}\otimes V\qquad\text{and}\qquad W_{-}=\hat{B}\otimes V. \tag*{(10)} W + = H ⊗ V and W − = B ^ ⊗ V . (10)
Let σ \sigma σ be the involution of O \mathbb{O} O equal to 1 1 1 on H \mathbb{H} H and to − 1 -1 − 1 on B ^ \hat{B} B ^ , and let Π B ^ \Pi_{\hat{B}} Π B ^ be the orthogonal projection of O \mathbb{O} O onto B ^ \hat{B} B ^ . Then σ = 1 − 2 Π B ^ \sigma=1-2\Pi_{\hat{B}} σ = 1 − 2 Π B ^ . The orthogonal projections to W ± W_{\pm} W ± are then Π ± \Pi_{\pm} Π ± , with Π − = Π B ^ ⊗ 1 \Pi_{-}=\Pi_{\hat{B}}\otimes1 Π − = Π B ^ ⊗ 1 .
The frame stabilizer will also give invariant sectors for the full pencil. Let K = SU ( 2 ) = Stab G 2 ( u , v ) K=\operatorname{SU}(2)=\operatorname{Stab}_{G_2}(u,v) K = SU ( 2 ) = Stab G 2 ( u , v ) be the residual symmetry group of the frame; it fixes H \mathbb{H} H pointwise [4 Chapter 6, Appendix 2, p. 81] . For a K K K -module U U U , write
U K = { s ∈ U : k s = s for every k ∈ K } . U^{K}=\{s\in U:ks=s\text{ for every }k\in K\}. U K = { s ∈ U : k s = s for every k ∈ K } .
The multiplication table of H \mathbb{H} H is in Appendix A, which also describes the action of K K K on B ^ \hat{B} B ^ and the Cayley–Dickson rules that use the basis ξ ^ α = ρ ^ α ℓ \hat{\xi}_{\alpha}=\hat{\rho}_{\alpha}\ell ξ ^ α = ρ ^ α ℓ .
An arbitrary pair ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) of octonions also determines such a frame by Gram–Schmidt orthogonalization. Take u = Im λ 1 / ∣ Im λ 1 ∣ u=\operatorname{Im}\lambda_1/|\operatorname{Im}\lambda_1| u = Im λ 1 /∣ Im λ 1 ∣ if Im λ 1 ≠ 0 \operatorname{Im}\lambda_1\ne0 Im λ 1 = 0 , and any unit imaginary octonion otherwise, and write Im λ 2 = α u + β v \operatorname{Im}\lambda_2=\alpha u+\beta v Im λ 2 = α u + β v with α , β ∈ R \alpha,\beta\in\mathbb{R} α , β ∈ R and v ⊥ u v\perp u v ⊥ u a unit imaginary octonion; if Im λ 2 ∈ R u \operatorname{Im}\lambda_2\in\mathbb{R}u Im λ 2 ∈ R u , any such v v v is allowed and β = 0 \beta=0 β = 0 . Then ( u , v ) (u,v) ( u , v ) is orthonormal, and
λ 1 = μ 1 + Q 1 u and λ 2 = μ 2 + α u + β v (11) \lambda_1=\mu_1+Q_1u\qquad\text{and}\qquad\lambda_2=\mu_2+\alpha u+\beta v \tag*{(11)} λ 1 = μ 1 + Q 1 u and λ 2 = μ 2 + α u + β v (11)
with μ 1 , Q 1 , μ 2 ∈ R \mu_1,Q_1,\mu_2\in\mathbb{R} μ 1 , Q 1 , μ 2 ∈ R .
Now, for a ∈ R 4 a\in\mathbb{R}^{4} a ∈ R 4 we define
x = a 1 2 + a 2 2 , y = a 3 2 and N = a 0 2 + x + y . (12) x=a_1^2+a_2^2,\qquad y=a_3^2\qquad\text{and}\qquad N=a_0^2+x+y. \tag*{(12)} x = a 1 2 + a 2 2 , y = a 3 2 and N = a 0 2 + x + y . (12)
and, in the frame ( u , v ) (u,v) ( u , v ) , define the quaternions
q ^ = a 0 + a 1 u + a 2 v + a 3 z and q ˇ = a 0 + a 1 u + a 2 v − a 3 z . (13) \hat{q}=a_0+a_1u+a_2v+a_3z\qquad\text{and}\qquad\check{q}=a_0+a_1u+a_2v-a_3z. \tag*{(13)} q ^ = a 0 + a 1 u + a 2 v + a 3 z and q ˇ = a 0 + a 1 u + a 2 v − a 3 z . (13)
The coefficients of the reduced pencil and their intrinsic meaning are as follows.
Proposition 3.2. Let ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) be a pair of octonions, written in a frame ( u , v ) (u,v) ( u , v ) as in (11) .
On O \mathbb{O} O and on W W W ,
R λ 2 R λ 1 = a 0 + a 1 R u + a 2 R v + a 3 R v R u . R_{\lambda_2}R_{\lambda_1}=a_0+a_1R_u+a_2R_v+a_3R_vR_u. R λ 2 R λ 1 = a 0 + a 1 R u + a 2 R v + a 3 R v R u . where
a 0 + i a 1 = ( μ 1 + i Q 1 ) ( μ 2 + i α ) , a 2 + i a 3 = β ( μ 1 + i Q 1 ) . (14) \begin{aligned}
a_0+ia_1&=(\mu_1+iQ_1)(\mu_2+i\alpha),\\
a_2+ia_3&=\beta(\mu_1+iQ_1). \tag*{(14)}
\end{aligned} a 0 + i a 1 a 2 + i a 3 = ( μ 1 + i Q 1 ) ( μ 2 + i α ) , = β ( μ 1 + i Q 1 ) . (14) Consequently, h − R λ 2 R λ 1 = P a h-R_{\lambda_2}R_{\lambda_1}=P_a h − R λ 2 R λ 1 = P a . We call a a a the reduced point of the pair in the frame ( u , v ) (u,v) ( u , v ) .
Let q ^ , q ˇ \hat q,\check q q ^ , q ˇ be as in (13) , and let x , y , N x,y,N x , y , N be as in (12) .
The quaternionic elements q ^ \hat q q ^ and q ˇ \check q q ˇ are the products of the original octonions in the two orders:
q ^ = λ 1 λ 2 and q ˇ = λ 2 λ 1 . \hat q=\lambda_1\lambda_2 \qquad\text{and}\qquad\check q=\lambda_2\lambda_1. q ^ = λ 1 λ 2 and q ˇ = λ 2 λ 1 . Moreover, the invariants admit the following intrinsic descriptions:
a 0 = Re ( λ 1 λ 2 ) = μ 1 μ 2 − ⟨ Im λ 1 , Im λ 2 ⟩ , N = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 , y = ∣ Im λ 1 ∣ 2 ∣ Im λ 2 ∣ 2 − ⟨ Im λ 1 , Im λ 2 ⟩ 2 = Q 1 2 β 2 , x = N − a 0 2 − y . \begin{aligned}
a_0&=\operatorname{Re}(\lambda_1\lambda_2)=\mu_1\mu_2-\langle\operatorname{Im}\lambda_1,\operatorname{Im}\lambda_2\rangle,\\
N&=|\lambda_1|^2|\lambda_2|^2,\\
y&=|\operatorname{Im}\lambda_1|^2|\operatorname{Im}\lambda_2|^2-\langle\operatorname{Im}\lambda_1,\operatorname{Im}\lambda_2\rangle^2=Q_1^2\beta^2,\\
x&=N-a_0^2-y.
\end{aligned} a 0 N y x = Re ( λ 1 λ 2 ) = μ 1 μ 2 − ⟨ Im λ 1 , Im λ 2 ⟩ , = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 , = ∣ Im λ 1 ∣ 2 ∣ Im λ 2 ∣ 2 − ⟨ Im λ 1 , Im λ 2 ⟩ 2 = Q 1 2 β 2 , = N − a 0 2 − y . Thus y y y is the Gram determinant of the imaginary parts of λ 1 \lambda_1 λ 1 and λ 2 \lambda_2 λ 2 .
In particular, y = a 3 2 = 0 y=a_3^2=0 y = a 3 2 = 0 if and only if Im λ 1 \operatorname{Im}\lambda_1 Im λ 1 and Im λ 2 \operatorname{Im}\lambda_2 Im λ 2 are linearly dependent. When these equivalent conditions hold, R λ 2 R λ 1 = R λ 1 λ 2 R_{\lambda_2}R_{\lambda_1}=R_{\lambda_1\lambda_2} R λ 2 R λ 1 = R λ 1 λ 2 .
Let
E = span R { u , v } and Γ = { g ∈ G 2 : g E = E } . \mathcal{E}=\operatorname{span}_{\mathbb{R}}\{u,v\}\qquad\text{and}\qquad\Gamma=\{g\in G_2:g\mathcal{E}=\mathcal{E}\}. E = span R { u , v } and Γ = { g ∈ G 2 : g E = E } . For g ∈ Γ g\in\Gamma g ∈ Γ , let κ ∈ O ( 2 ) \kappa\in O(2) κ ∈ O ( 2 ) be the matrix of its action on E \mathcal{E} E in the basis ( u , v ) (u,v) ( u , v ) . Define
κ ⋅ a = ( a 0 , κ ( a 1 , a 2 ) T , ( det κ ) a 3 ) . \kappa\cdot a=(a_0,\kappa(a_1,a_2)^{\mathsf T},(\det\kappa)a_3). κ ⋅ a = ( a 0 , κ ( a 1 , a 2 ) T , ( det κ ) a 3 ) . Then
g P a g − 1 = P κ ⋅ a for a ∈ R 4 . (15) gP_ag^{-1}=P_{\kappa\cdot a}\qquad\text{for }a\in\mathbb{R}^4. \tag*{(15)} g P a g − 1 = P κ ⋅ a for a ∈ R 4 . (15)
Proof. Part 1. By real linearity of right multiplication,
R λ 1 = μ 1 + Q 1 R u and R λ 2 = μ 2 + α R u + β R v . R_{\lambda_1}=\mu_1+Q_1R_u\qquad\text{and}\qquad R_{\lambda_2}=\mu_2+\alpha R_u+\beta R_v. R λ 1 = μ 1 + Q 1 R u and R λ 2 = μ 2 + α R u + β R v .
Right alternativity gives ( w u ) u = w ( u u ) = − w (wu)u=w(uu)=-w ( w u ) u = w ( uu ) = − w , hence R u 2 = − 1 R_u^2=-1 R u 2 = − 1 . Expanding the product, we obtain
R λ 2 R λ 1 = ( μ 1 μ 2 − Q 1 α ) + ( Q 1 μ 2 + μ 1 α ) R u + μ 1 β R v + Q 1 β R v R u . \begin{aligned}
R_{\lambda_2}R_{\lambda_1}
&=(\mu_1\mu_2-Q_1\alpha)+(Q_1\mu_2+\mu_1\alpha)R_u\\
&\quad+\mu_1\beta R_v+Q_1\beta R_vR_u.
\end{aligned} R λ 2 R λ 1 = ( μ 1 μ 2 − Q 1 α ) + ( Q 1 μ 2 + μ 1 α ) R u + μ 1 β R v + Q 1 β R v R u .
The coefficients agree with (14) , proving the stated identity on O \mathbb{O} O . The same identity holds on W W W because right multiplications act on the octonion factor.
Part 2. Since λ 1 , λ 2 \lambda_1,\lambda_2 λ 1 , λ 2 belong to the associative quaternion subalgebra generated by u u u and v v v , we may multiply them using
u 2 = v 2 = − 1 and u v = z = − v u . u^2=v^2=-1 \qquad\text{and}\qquad uv=z=-vu. u 2 = v 2 = − 1 and uv = z = − v u .
This gives
λ 1 λ 2 = ( μ 1 + Q 1 u ) ( μ 2 + α u + β v ) = ( μ 1 μ 2 − Q 1 α ) + ( μ 1 α + Q 1 μ 2 ) u + μ 1 β v + Q 1 β z = a 0 + a 1 u + a 2 v + a 3 z = q ^ . \begin{aligned}
\lambda_1\lambda_2&=(\mu_1+Q_1u)(\mu_2+\alpha u+\beta v)\\
&=(\mu_1\mu_2-Q_1\alpha)+(\mu_1\alpha+Q_1\mu_2)u\\
&\qquad+\mu_1\beta v+Q_1\beta z\\
&=a_0+a_1u+a_2v+a_3z=\hat{q}.
\end{aligned} λ 1 λ 2 = ( μ 1 + Q 1 u ) ( μ 2 + α u + β v ) = ( μ 1 μ 2 − Q 1 α ) + ( μ 1 α + Q 1 μ 2 ) u + μ 1 β v + Q 1 β z = a 0 + a 1 u + a 2 v + a 3 z = q ^ .
Reversing the order of multiplication changes only the sign of the z z z -component, since v u = − u v vu=-uv v u = − uv . Thus
λ 2 λ 1 = a 0 + a 1 u + a 2 v − a 3 z = q ˇ . \lambda_2\lambda_1=a_0+a_1u+a_2v-a_3z=\check{q}. λ 2 λ 1 = a 0 + a 1 u + a 2 v − a 3 z = q ˇ .
The imaginary parts are Q 1 u Q_1u Q 1 u and α u + β v \alpha u+\beta v α u + β v , with inner product Q 1 α Q_1\alpha Q 1 α . Thus a 0 = μ 1 μ 2 − Q 1 α a_0=\mu_1\mu_2-Q_1\alpha a 0 = μ 1 μ 2 − Q 1 α is Re ( λ 1 λ 2 ) \operatorname{Re}(\lambda_1\lambda_2) Re ( λ 1 λ 2 ) , and their Gram determinant is
Q 1 2 ( α 2 + β 2 ) − ( Q 1 α ) 2 = Q 1 2 β 2 = a 3 2 = y . Q_1^2(\alpha^2+\beta^2)-(Q_1\alpha)^2=Q_1^2\beta^2=a_3^2=y. Q 1 2 ( α 2 + β 2 ) − ( Q 1 α ) 2 = Q 1 2 β 2 = a 3 2 = y .
Norm multiplicativity gives N = ∣ q ^ ∣ 2 = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 N=|\hat{q}|^2=|\lambda_1|^2|\lambda_2|^2 N = ∣ q ^ ∣ 2 = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 , and the definition of N N N gives x = N − a 0 2 − y x=N-a_0^2-y x = N − a 0 2 − y .
Since the inner product on Im O \operatorname{Im}\mathbb{O} Im O is positive definite, the Gram determinant vanishes precisely when Im λ 1 \operatorname{Im}\lambda_1 Im λ 1 and Im λ 2 \operatorname{Im}\lambda_2 Im λ 2 are linearly dependent. This proves the stated characterization of y = 0 y=0 y = 0 .
Finally, if y = 0 y=0 y = 0 , then a 3 = 0 a_3=0 a 3 = 0 , so q ^ = a 0 + a 1 u + a 2 v \hat{q}=a_0+a_1u+a_2v q ^ = a 0 + a 1 u + a 2 v . By part 1 and real linearity of right multiplication,
R λ 2 R λ 1 = a 0 + a 1 R u + a 2 R v = R a 0 + a 1 u + a 2 v = R q ^ = R λ 1 λ 2 . \begin{aligned}
R_{\lambda_2}R_{\lambda_1}&=a_0+a_1R_u+a_2R_v\\
&=R_{a_0+a_1u+a_2v}\\
&=R_{\hat{q}}=R_{\lambda_1\lambda_2}.
\end{aligned} R λ 2 R λ 1 = a 0 + a 1 R u + a 2 R v = R a 0 + a 1 u + a 2 v = R q ^ = R λ 1 λ 2 .
Part 3. Let g ∈ Γ g\in\Gamma g ∈ Γ . Since h h h is G 2 G_2 G 2 -equivariant (Proposition 2.4) and g R p g − 1 = R g p gR_pg^{-1}=R_{gp} g R p g − 1 = R g p for p ∈ O p\in\mathbb{O} p ∈ O , we have
g P a g − 1 = h − a 0 − a 1 R g u − a 2 R g v − a 3 R g v R g u . gP_ag^{-1}=h-a_0-a_1R_{gu}-a_2R_{gv}-a_3R_{gv}R_{gu}. g P a g − 1 = h − a 0 − a 1 R g u − a 2 R g v − a 3 R g v R g u .
Right alternativity and its polarization give
R u 2 = R v 2 = − 1 and R u R v + R v R u = − 2 ⟨ u , v ⟩ = 0. (16) R_u^2=R_v^2=-1 \qquad\text{and}\qquad R_uR_v+R_vR_u=-2\langle u,v\rangle=0. \tag*{(16)} R u 2 = R v 2 = − 1 and R u R v + R v R u = − 2 ⟨ u , v ⟩ = 0. (16)
Since the columns of κ \kappa κ are orthonormal, these relations imply
R g v R g u = ( det κ ) R v R u . R_{gv}R_{gu}=(\det\kappa)R_vR_u. R g v R g u = ( det κ ) R v R u .
On the other hand, the linear combination a 1 R g u + a 2 R g v a_1R_{gu}+a_2R_{gv} a 1 R g u + a 2 R g v has coefficients κ ( a 1 , a 2 ) T \kappa(a_1,a_2)^{\mathsf{T}} κ ( a 1 , a 2 ) T in the basis ( R u , R v ) (R_u,R_v) ( R u , R v ) . Thus conjugation by g g g transforms the coefficients precisely according to κ ⋅ a \kappa\cdot a κ ⋅ a , giving
g P a g − 1 = P κ ⋅ a . gP_ag^{-1}=P_{\kappa\cdot a}. g P a g − 1 = P κ ⋅ a .
Proposition 3.3. For any fixed frame ( u , v ) (u,v) ( u , v ) , the map ( μ 1 , Q 1 , μ 2 , α , β ) ↦ a (\mu_1,Q_1,\mu_2,\alpha,\beta)\mapsto\mathbf{a} ( μ 1 , Q 1 , μ 2 , α , β ) ↦ a is surjective from R 5 \mathbb{R}^5 R 5 onto R 4 \mathbb{R}^4 R 4 .
Proof. Given a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 , set c 1 = a 0 + i a 1 c_1=a_0+ia_1 c 1 = a 0 + i a 1 and c 2 = a 2 + i a 3 c_2=a_2+ia_3 c 2 = a 2 + i a 3 . Then (14) is satisfied by
( β , μ 1 + i Q 1 , μ 2 + i α ) = { ( 1 , c 2 , c 1 / c 2 ) , c 2 ≠ 0 , ( 0 , 1 , c 1 ) , c 2 = 0. (\beta,\mu_1+iQ_1,\mu_2+i\alpha)=
\begin{cases}
(1,c_2,c_1/c_2), & c_2\ne0,\\
(0,1,c_1), & c_2=0.
\end{cases} ( β , μ 1 + i Q 1 , μ 2 + i α ) = { ( 1 , c 2 , c 1 / c 2 ) , ( 0 , 1 , c 1 ) , c 2 = 0 , c 2 = 0.
Thus every pair has a reduced point, and every coefficient point occurs. The fourth operator of the pencil, R v R u R_vR_u R v R u , can be evaluated on the frame.
Proposition 3.4. For all p ∈ O p\in\mathbb{O} p ∈ O ,
( p u ) v = σ ( p ) z , (pu)v=\sigma(p)z, ( p u ) v = σ ( p ) z , or equivalently, [ p , u , v ] = 2 ( Π B ^ p ) z [p,u,v]=2(\Pi_{\widehat{B}}p)z [ p , u , v ] = 2 ( Π B p ) z with the associator as defined by (2) . On W W W , we then have
R v R u = R z ∘ ( σ ⊗ 1 ) (17) R_vR_u=R_z\circ(\sigma\otimes1) \tag*{(17)} R v R u = R z ∘ ( σ ⊗ 1 ) (17) and consequently, for all a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 ,
P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R z ( σ ⊗ 1 ) . P_{\mathbf{a}}=h-a_0I-a_1R_u-a_2R_v-a_3R_z(\sigma\otimes1). P a = h − a 0 I − a 1 R u − a 2 R v − a 3 R z ( σ ⊗ 1 ) . Proof. For p ∈ H p\in\mathbb{H} p ∈ H , associativity of H \mathbb{H} H gives ( p u ) v = p z (pu)v=pz ( p u ) v = p z . For p = r ℓ ∈ B ^ p=r\ell\in\widehat{B} p = r ℓ ∈ B , the Cayley–Dickson rules (A.2) [2 Section 2.2] , also described in [4 Section 6.3 and Chapter 6, Appendix 1] , give
( ( r ℓ ) u ) v = ( ( r u ‾ ) ℓ ) v = ( ( r u ‾ ) v ‾ ) ℓ = ( r z ) ℓ and ( r ℓ ) z = ( r z ‾ ) ℓ = − ( r z ) ℓ , ((r\ell)u)v=((r\overline{u})\ell)v=((r\overline{u})\overline{v})\ell=(rz)\ell
\qquad\text{and}\qquad
(r\ell)z=(r\overline{z})\ell=-(rz)\ell, (( r ℓ ) u ) v = (( r u ) ℓ ) v = (( r u ) v ) ℓ = ( r z ) ℓ and ( r ℓ ) z = ( r z ) ℓ = − ( r z ) ℓ , because u ‾ v ‾ = v u = z \overline{u}\,\overline{v}=vu=z u v = v u = z ; so ( p u ) v = − p z (pu)v=-pz ( p u ) v = − p z on B ^ \widehat{B} B . Thus
( p u ) v = σ ( p ) z for every p ∈ O . (18) (pu)v=\sigma(p)z\qquad\text{for every }p\in\mathbb{O}. \tag*{(18)} ( p u ) v = σ ( p ) z for every p ∈ O . (18) Since right multiplications act on the octonion factor of W W W , (3.11) gives
R v R u = R z ∘ ( σ ⊗ 1 ) on W = O ⊗ V . R_vR_u=R_z\circ(\sigma\otimes1)\quad\text{on }W=\mathbb{O}\otimes V. R v R u = R z ∘ ( σ ⊗ 1 ) on W = O ⊗ V . The associator form follows from σ ( p ) z = p z − 2 ( Π B ^ p ) z \sigma(p)z=pz-2(\Pi_{\widehat{B}}p)z σ ( p ) z = p z − 2 ( Π B p ) z and (2) . Substitution into the pencil gives its stated form. In particular, R v R u = R z R_vR_u=R_z R v R u = R z on W + W_+ W + and R v R u = − R z R_vR_u=-R_z R v R u = − R z on W − W_- W − .
Proposition 3.5. Fix a frame ( u , v ) (u,v) ( u , v ) , with associated decomposition W = W + ⊕ W − W=W_+\oplus W_- W = W + ⊕ W − and projections Π ± \Pi_\pm Π ± .
(a) For every a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 , with q ^ , q ˇ \widehat{q},\check{q} q , q ˇ as in (13) ,
P a = h − R q ^ Π + − R q ˇ Π − . (19) P_{\mathbf{a}}=h-R_{\widehat{q}}\Pi_+-R_{\check{q}}\Pi_-. \tag*{(19)} P a = h − R q Π + − R q ˇ Π − . (19) (b) If a \mathbf{a} a is the reduced point of a pair ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) in this frame, then
P a = h − R λ 1 λ 2 Π + − R λ 2 λ 1 Π − . (20) P_{\mathbf{a}}=h-R_{\lambda_1\lambda_2}\Pi_+-R_{\lambda_2\lambda_1}\Pi_-. \tag*{(20)} P a = h − R λ 1 λ 2 Π + − R λ 2 λ 1 Π − . (20) Equivalently, right multiplication satisfies
( w λ 1 ) λ 2 = { w ( λ 1 λ 2 ) , w ∈ W + , w ( λ 2 λ 1 ) , w ∈ W − . (w\lambda_1)\lambda_2=
\begin{cases}
w(\lambda_1\lambda_2), & w\in W_+,\\
w(\lambda_2\lambda_1), & w\in W_-.
\end{cases} ( w λ 1 ) λ 2 = { w ( λ 1 λ 2 ) , w ( λ 2 λ 1 ) , w ∈ W + , w ∈ W − .
Proof. By Proposition 3.4, R v R u = ± R z R_vR_u=\pm R_z R v R u = ± R z on W ± W_{\pm} W ± . Thus, by real linearity of right multiplication,
a 0 I + a 1 R u + a 2 R v + a 3 R v R u = R a 0 + a 1 u + a 2 v ± a 3 z on W ± . a_0I+a_1R_u+a_2R_v+a_3R_vR_u=R_{a_0+a_1u+a_2v\pm a_3z}\quad\text{on }W_{\pm}. a 0 I + a 1 R u + a 2 R v + a 3 R v R u = R a 0 + a 1 u + a 2 v ± a 3 z on W ± .
These are precisely R q ^ R_{\hat q} R q ^ and R q ˇ R_{\check q} R q ˇ , proving (a). For (b), Proposition 3.2 gives P a = h − R λ 2 R λ 1 P_a=h-R_{\lambda_2}R_{\lambda_1} P a = h − R λ 2 R λ 1 , q ^ = λ 1 λ 2 \hat q=\lambda_1\lambda_2 q ^ = λ 1 λ 2 , and q ˇ = λ 2 λ 1 \check q=\lambda_2\lambda_1 q ˇ = λ 2 λ 1 . Substituting into (a) proves the claim. □ \square □
Each half is preserved by right multiplication by elements of H \mathbb{H} H . However, h h h may couple the two halves. To obtain invariant subspaces for the full pencil, we now decompose W W W into K K K -isotypic sectors.
The SU(2) sectors
The operator h h h and all four operators of the pencil commute with K = S U ( 2 ) K=\mathrm{SU}(2) K = SU ( 2 ) , so W W W splits according to representations of K K K . The action of K K K on B ^ \hat B B ^ and the types of its irreducible modules that occur in W W W are described in Appendix § 7 , equation (A.5) and Lemma A.1 . The real irreducible types that occur are 1 1 1 , 3 3 3 and 4 4 4 , of dimensions 1 1 1 , 3 3 3 and 4 4 4 , with 4 4 4 of quaternionic type. The subscript of W i W_i W i denotes the dimension of its irreducible type; the sector itself can have larger dimension. For γ ∈ H \gamma\in\mathbb{H} γ ∈ H and X ∈ s u ( 2 ) X\in\mathfrak{su}(2) X ∈ su ( 2 ) let
ε γ = ∑ α = 0 3 ( ξ ^ α γ ) ⊗ ξ α , w X , γ = ∑ α = 0 3 ( ( X ξ ^ α ) γ ) ⊗ ξ α , (21) \varepsilon_{\gamma}=\sum_{\alpha=0}^{3}(\hat{\xi}_{\alpha}\gamma)\otimes\xi_{\alpha},
\qquad
w_{X,\gamma}=\sum_{\alpha=0}^{3}((X\hat{\xi}_{\alpha})\gamma)\otimes\xi_{\alpha}, \tag*{(21)} ε γ = α = 0 ∑ 3 ( ξ ^ α γ ) ⊗ ξ α , w X , γ = α = 0 ∑ 3 (( X ξ ^ α ) γ ) ⊗ ξ α , (21)
where ε γ \varepsilon_{\gamma} ε γ does not depend on the orthonormal basis ( ξ ^ α ) (\hat{\xi}_{\alpha}) ( ξ ^ α ) of B ^ \hat B B ^ , and X X X acts on B ^ \hat B B ^ as an element of the Lie algebra of S U ( 2 ) ⊂ G 2 \mathrm{SU}(2)\subset G_2 SU ( 2 ) ⊂ G 2 , that is, as a derivation of O \mathbb{O} O that vanishes on H \mathbb{H} H and preserves B ^ \hat B B ^ .
Theorem 3.6. For i ∈ { 1 , 3 , 4 } i\in\{1,3,4\} i ∈ { 1 , 3 , 4 } , let W i W_i W i denote the K K K -isotypic component of W W W of type i i i . Then
W = W 1 ⊕ W 3 ⊕ W 4 (22) W=W_1\oplus W_3\oplus W_4 \tag*{(22)} W = W 1 ⊕ W 3 ⊕ W 4 (22) is an orthogonal decomposition, with each summand preserved by P a P_a P a for every a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 and by the plane stabilizer Γ \Gamma Γ . More precisely:
(a) The trivial isotypic component is
W 1 = W K = ( H ⊗ A ) ⊕ { ε γ : γ ∈ H } , and dim R W 1 = 16. W_1=W^K=(\mathbb{H}\otimes A)\oplus\{\varepsilon_{\gamma}:\gamma\in\mathbb{H}\},
\qquad\text{and}\qquad\dim_{\mathbb{R}}W_1=16. W 1 = W K = ( H ⊗ A ) ⊕ { ε γ : γ ∈ H } , and dim R W 1 = 16. (b) The component W 3 W_3 W 3 has real dimension 12 12 12 and satisfies
W 3 ⊂ B ^ ⊗ B and W 3 ⊂ W − . W_3\subset\hat B\otimes B
\qquad\text{and}\qquad
W_3\subset W_-. W 3 ⊂ B ^ ⊗ B and W 3 ⊂ W − . The map X ⊗ γ ↦ w X , γ X\otimes\gamma\mapsto w_{X,\gamma} X ⊗ γ ↦ w X , γ defines a K K K -equivariant isomorphism
s u ( 2 ) ⊗ H ≅ W 3 , \mathfrak{su}(2)\otimes\mathbb{H}\cong W_3, su ( 2 ) ⊗ H ≅ W 3 , where K ≅ S U ( 2 ) K\cong\mathrm{SU}(2) K ≅ SU ( 2 ) acts by Ad \operatorname{Ad} Ad on s u ( 2 ) \mathfrak{su}(2) su ( 2 ) and trivially on H \mathbb{H} H .
(c) The remaining component is
W 4 = ( H ⊗ B ) ⊕ ( B ^ ⊗ A ) , and dim R W 4 = 28. W_{4} = (\mathbb{H} \otimes B) \oplus(\widehat{B} \otimes A), \qquad\text{and} \qquad\dim_{\mathbb{R}} W_{4} = 28. W 4 = ( H ⊗ B ) ⊕ ( B ⊗ A ) , and dim R W 4 = 28.
(d) As a K K K -module,
W 4 ≅ B ^ ⊕ 7 and End K ( W 4 ) ≅ M 7 ( H ) , W_{4} \cong\widehat{B}^{\oplus7} \qquad\text{and} \qquad\operatorname{End}_{K}(W_{4}) \cong M_{7}(\mathbb{H}), W 4 ≅ B ⊕ 7 and End K ( W 4 ) ≅ M 7 ( H ) ,
where B ^ ≅ H \widehat{B} \cong\mathbb{H} B ≅ H is the irreducible module of type 4, of quaternionic type. Under a K K K -equivariant identification with B ^ ⊕ 7 \widehat{B}^{\oplus7} B ⊕ 7 , the matrices act on row vectors from the right.
Proof. The operator h h h is G 2 G_{2} G 2 -equivariant (Proposition 2.4), while R u R_{u} R u and R v R_{v} R v commute with K K K . Thus every P a P_{a} P a commutes with K K K and preserves its isotypic components. These components are mutually orthogonal since the K K K -action is orthogonal. Lemma A.1 gives
W ≅ 1 ⊕ 16 ⊕ 3 ⊕ 4 ⊕ 4 ⊕ 7 W \cong\mathbf{1}^{\oplus16} \oplus\mathbf{3}^{\oplus4} \oplus\mathbf{4}^{\oplus7} W ≅ 1 ⊕ 16 ⊕ 3 ⊕ 4 ⊕ 4 ⊕ 7
and locates these isotypic summands in the tensor decomposition. Their real dimensions are 16, 12 and 28. We now establish the explicit identifications in (a)–(d), using the commutant description in Appendix § 7 .
By Lemma A.1 , W 1 = W K = ( H ⊗ A ) ⊕ ( B ^ ⊗ B ) K W_{1} = W^{K} = (\mathbb{H} \otimes A) \oplus(\widehat{B} \otimes B)^{K} W 1 = W K = ( H ⊗ A ) ⊕ ( B ⊗ B ) K and dim W K = 16 \dim W^{K} = 16 dim W K = 16 . The metric and B ≅ B ^ B \cong\widehat{B} B ≅ B identify ( B ^ ⊗ B ) K (\widehat{B} \otimes B)^{K} ( B ⊗ B ) K with End K ( B ^ ) \operatorname{End}_{K}(\widehat{B}) End K ( B ) ; this identification B ^ ⊗ B = End ( B ^ ) \widehat{B} \otimes B = \operatorname{End}(\widehat{B}) B ⊗ B = End ( B ) is basis free, and it sends ϵ γ \epsilon_{\gamma} ϵ γ of (21) to η ^ ↦ ∑ α ⟨ η ^ , ξ ^ α ⟩ γ ξ ^ α = γ η ^ \widehat{\eta} \mapsto\sum_{\alpha} \langle\widehat{\eta}, \widehat{\xi}_{\alpha}\rangle\gamma\widehat{\xi}_{\alpha} = \gamma\widehat{\eta} η ↦ ∑ α ⟨ η , ξ α ⟩ γ ξ α = γ η , for every orthonormal basis ( ξ ^ α ) (\widehat{\xi}_{\alpha}) ( ξ α ) of B ^ \widehat{B} B . The commutant described in Appendix § 7 consists of the maps ξ ^ ↦ γ ξ ^ \widehat{\xi} \mapsto\gamma\widehat{\xi} ξ ↦ γ ξ , whose tensors are the ϵ γ \epsilon_{\gamma} ϵ γ , which therefore do not depend on the choice of basis.
Under the metric identification B ^ ⊗ B ≅ End ( B ^ ) \widehat{B} \otimes B \cong\operatorname{End}(\widehat{B}) B ⊗ B ≅ End ( B ) , the tensor w X , γ w_{X,\gamma} w X , γ corresponds to R γ ∘ X ∣ B ^ R_{\gamma} \circ X|_{\widehat{B}} R γ ∘ X ∣ B , so it does not depend on the choice of orthonormal basis. We prove the parametrization of W 3 W_{3} W 3 : the map X ⊗ γ ↦ w X , γ X \otimes\gamma\mapsto w_{X,\gamma} X ⊗ γ ↦ w X , γ of (21) is a K K K -equivariant isomorphism s u ( 2 ) ⊗ H → W 3 \mathfrak{su}(2) \otimes\mathbb{H} \to W_{3} su ( 2 ) ⊗ H → W 3 , with K K K acting by Ad \operatorname{Ad} Ad on s u ( 2 ) \mathfrak{su}(2) su ( 2 ) and trivially on the label γ \gamma γ .
The map is equivariant: since g ∈ K g \in K g ∈ K is an automorphism of O \mathbb{O} O that fixes γ \gamma γ and maps orthonormal bases to orthonormal bases, g ⋅ w X , γ = w Ad g X , γ g \cdot w_{X,\gamma} = w_{\operatorname{Ad}_{g}X,\gamma} g ⋅ w X , γ = w Ad g X , γ .
For injectivity, the operators R γ ∣ B ^ R_{\gamma}|_{\widehat{B}} R γ ∣ B form the commutant, while R ⊕ s u ( 2 ) \mathbb{R} \oplus\mathfrak{su}(2) R ⊕ su ( 2 ) is a commuting quaternion copy. The resulting identification H ⊗ H o p ≅ M 4 ( R ) \mathbb{H} \otimes\mathbb{H}^{\mathrm{op}} \cong M_{4}(\mathbb{R}) H ⊗ H op ≅ M 4 ( R ) gives injectivity on s u ( 2 ) ⊗ H \mathfrak{su}(2) \otimes\mathbb{H} su ( 2 ) ⊗ H . Finally, the image is 3-isotypic and has dimension 12, so it is all of W 3 W_{3} W 3 by Lemma A.1 . Its multiplicity space is therefore the label space H \mathbb{H} H .
By Lemma A.1 , W 4 = ( H ⊗ B ) ⊕ ( B ^ ⊗ A ) ≅ B ^ ⊗ R 7 = B ^ 7 ≅ H 7 W_{4} = (\mathbb{H} \otimes B) \oplus(\widehat{B} \otimes A) \cong\widehat{B} \otimes\mathbb{R}^{7} = \widehat{B}^{7} \cong\mathbb{H}^{7} W 4 = ( H ⊗ B ) ⊕ ( B ⊗ A ) ≅ B ⊗ R 7 = B 7 ≅ H 7 , with K K K acting by left multiplication in each slot, and B ^ \widehat{B} B is of quaternionic type with commutant the right quaternion multiplications; on seven copies these give M 7 ( H ) M_{7}(\mathbb{H}) M 7 ( H ) , the operators s ↦ s M \mathbf{s} \mapsto\mathbf{s}M s ↦ s M , M ∈ M 7 ( H ) M \in M_{7}(\mathbb{H}) M ∈ M 7 ( H ) , acting on rows s ∈ H 7 \mathbf{s} \in\mathbb{H}^{7} s ∈ H 7 from the right.
We next show that Γ \Gamma Γ preserves each component. Since K K K is the pointwise stabilizer of E = span R { u , v } E = \operatorname{span}_{\mathbb{R}}\{u,v\} E = span R { u , v } , every g ∈ Γ g \in\Gamma g ∈ Γ normalizes K K K . Every automorphism of K ≅ SU ( 2 ) K \cong\operatorname{SU}(2) K ≅ SU ( 2 ) is inner, so there is k 0 ∈ K k_{0} \in K k 0 ∈ K such that
g − 1 k g = k 0 k k 0 − 1 for k ∈ K . g^{-1}kg = k_{0}kk_{0}^{-1} \qquad\text{for } k \in K. g − 1 k g = k 0 k k 0 − 1 for k ∈ K .
Hence g k 0 gk_{0} g k 0 commutes with K K K and preserves its isotypic components. Since k 0 k_{0} k 0 also preserves them, so does g g g .
These sectors are common invariant subspaces for all operators in the pencil. Their symmetry also reduces the variables needed for the determinant calculations.
Proposition 3.7 (Dependence on the invariants). The restriction homomorphism Γ → O ( 2 ) \Gamma\to O(2) Γ → O ( 2 ) , g ↦ κ g \mapsto\kappa g ↦ κ , is surjective. For each i ∈ { 1 , 3 , 4 } \mathbf{i} \in\{1,3,4\} i ∈ { 1 , 3 , 4 } , the sector determinant
D i ( a ) = det ( P a ∣ W i ) D_i(\mathbf{a})=\det(P_{\mathbf{a}}|_{W_i}) D i ( a ) = det ( P a ∣ W i ) satisfies
D i ( κ ⋅ a ) = D i ( a ) for κ ∈ O ( 2 ) . D_i(\kappa\cdot\mathbf{a})=D_i(\mathbf{a}) \qquad\text{for }\kappa\in O(2). D i ( κ ⋅ a ) = D i ( a ) for κ ∈ O ( 2 ) . Consequently, D i D_i D i is a polynomial in a 0 a_0 a 0 , x x x and y y y .
Proof. Recall that G 2 G_2 G 2 acts transitively on ordered orthonormal pairs in Im O \operatorname{Im}\mathbb{O} Im O . Consequently, every κ ∈ O ( 2 ) \kappa\in O(2) κ ∈ O ( 2 ) is induced by some g ∈ Γ g \in\Gamma g ∈ Γ . By (15) and the Γ \Gamma Γ -invariance of each W i W_i W i ,
D i ( κ ⋅ a ) = D i ( a ) , D_i(\kappa\cdot\mathbf{a})=D_i(\mathbf{a}), D i ( κ ⋅ a ) = D i ( a ) ,
proving the claimed invariance.
Finally, fix i i i and set
f ( a 0 , τ , υ ) = D i ( a 0 , τ , 0 , υ ) . f(a_0,\tau,\upsilon)=D_i(a_0,\tau,0,\upsilon). f ( a 0 , τ , υ ) = D i ( a 0 , τ , 0 , υ ) .
The action of κ = − I \kappa=-I κ = − I makes f f f even in τ \tau τ , while κ = diag ( − 1 , 1 ) \kappa=\operatorname{diag}(-1,1) κ = diag ( − 1 , 1 ) then makes it even in υ \upsilon υ . Thus, f ( a 0 , τ , υ ) = F ( a 0 , τ 2 , υ 2 ) f(a_0,\tau,\upsilon)=F(a_0,\tau^2,\upsilon^2) f ( a 0 , τ , υ ) = F ( a 0 , τ 2 , υ 2 ) for some polynomial F F F . Since a rotation can take ( a 1 , a 2 ) (a_1,a_2) ( a 1 , a 2 ) to ( x , 0 ) (\sqrt{x},0) ( x , 0 ) without changing a 3 a_3 a 3 , rotational invariance gives
D i ( a ) = F ( a 0 , x , a 3 2 ) = F ( a 0 , x , y ) . D_i(\mathbf{a})=F(a_0,x,a_3^2)=F(a_0,x,y). D i ( a ) = F ( a 0 , x , a 3 2 ) = F ( a 0 , x , y ) .
We call W 1 W_1 W 1 the fixed sector since it is preserved pointwise by SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) . In this sector, the part H ⊗ A \mathbb{H}\otimes A H ⊗ A lies in W + W_+ W + and the four-dimensional space of the ϵ γ \epsilon_\gamma ϵ γ lies in W − W_- W − . A quaternionic coordinate system shows how the pencil acts on both parts at once.
Proposition 3.8. Identify W 1 W_1 W 1 with H 3 ⊕ H \mathbb{H}^3\oplus\mathbb{H} H 3 ⊕ H via the real-linear isomorphism
Φ : W 1 ⟶ H 3 ⊕ H , \Phi:W_1\longrightarrow\mathbb{H}^3\oplus\mathbb{H}, Φ : W 1 ⟶ H 3 ⊕ H , ∑ j = 1 3 ϑ j ⊗ ρ j + ϵ γ ⟼ ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) , where η = z γ z ‾ . \sum_{j=1}^{3}\vartheta_j\otimes\rho_j+\epsilon_\gamma\longmapsto(\vartheta_1,\vartheta_2,\vartheta_3;\eta), \qquad\text{where }\eta=z\gamma\overline{z}. j = 1 ∑ 3 ϑ j ⊗ ρ j + ϵ γ ⟼ ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) , where η = z γ z .
We label the first three quaternionic coordinates by 1 , 2 , 3 1,2,3 1 , 2 , 3 and the last by 0 0 0 . Then, for every a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 , the operator a 0 I + a 1 R u + a 2 R v + a 3 R v R u a_0I+a_1R_u+a_2R_v+a_3R_vR_u a 0 I + a 1 R u + a 2 R v + a 3 R v R u acts by right multiplication by q ^ \hat{q} q ^ on all four slots:
( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ⟼ ( ϑ 1 q ^ , ϑ 2 q ^ , ϑ 3 q ^ ; η q ^ ) . (23) (\vartheta_1,\vartheta_2,\vartheta_3;\eta)\longmapsto(\vartheta_1\hat{q},\vartheta_2\hat{q},\vartheta_3\hat{q};\eta\hat{q}). \tag*{(23)} ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ⟼ ( ϑ 1 q ^ , ϑ 2 q ^ , ϑ 3 q ^ ; η q ^ ) . (23)
The matrix of h h h in these coordinates is given in Appendix § 8 , equation (B.1).
Proof. By Proposition 3.4, R v R u = R z R_vR_u=R_z R v R u = R z on W + = H ⊗ V W_+=\mathbb{H}\otimes V W + = H ⊗ V , hence the operator acts by right multiplication by q ^ \hat{q} q ^ on slots 1 , 2 , 3 1,2,3 1 , 2 , 3 , which correspond to H ⊗ A ⊂ W + \mathbb{H}\otimes A\subset W_+ H ⊗ A ⊂ W + .
On slot 0 0 0 , the Cayley–Dickson rule (A.3) gives
R p ϵ γ = ϵ γ p ‾ and R v R u ϵ γ = ϵ γ z . R_p\epsilon_\gamma=\epsilon_\gamma\overline{p}
\qquad\text{and}\qquad
R_vR_u\epsilon_\gamma=\epsilon_\gamma z. R p ϵ γ = ϵ γ p and R v R u ϵ γ = ϵ γ z .
Hence the operator acts on γ \gamma γ by
γ ⟼ γ ( a 0 − a 1 u − a 2 v + a 3 z ) = γ ( z ˉ q ^ z ) . \gamma\longmapsto\gamma(a_0-a_1u-a_2v+a_3z)=\gamma(\bar{z}\hat{q}z). γ ⟼ γ ( a 0 − a 1 u − a 2 v + a 3 z ) = γ ( z ˉ q ^ z ) .
Under the change of variable η = z γ z ˉ \eta=z\gamma\bar{z} η = z γ z ˉ , this becomes
η ⟼ z γ ( z ˉ q ^ z ) z ˉ = η q ^ . \eta\longmapsto z\gamma(\bar{z}\hat{q}z)\bar{z}=\eta\hat{q}. η ⟼ z γ ( z ˉ q ^ z ) z ˉ = η q ^ .
Thus all four slots carry right multiplication by q ^ \hat{q} q ^ , as claimed.
The tensor metric in these coordinates is
∥ Φ − 1 ( ϑ ; η ) ∥ 2 = ∑ j = 1 3 ∣ ϑ j ∣ 2 + 4 ∣ η ∣ 2 , with G = diag ( I 12 , 4 I 4 ) . \lVert\Phi^{-1}(\vartheta;\eta)\rVert^2=\sum_{j=1}^{3}\lvert\vartheta_j\rvert^2+4\lvert\eta\rvert^2,\qquad\text{with}\qquad G=\operatorname{diag}(I_{12},4I_4). ∥ Φ − 1 ( ϑ ; η ) ∥ 2 = j = 1 ∑ 3 ∣ ϑ j ∣ 2 + 4 ∣ η ∣ 2 , with G = diag ( I 12 , 4 I 4 ) .
Indeed ∥ ε γ ∥ 2 = 4 ∣ γ ∣ 2 \lVert\varepsilon_\gamma\rVert^2=4\lvert\gamma\rvert^2 ∥ ε γ ∥ 2 = 4 ∣ γ ∣ 2 and ∣ z γ z ˉ ∣ = ∣ γ ∣ \lvert z\gamma\bar{z}\rvert=\lvert\gamma\rvert ∣ z γ z ˉ ∣ = ∣ γ ∣ .
These coordinates are not orthonormal: the fourth quaternionic slot has metric weight 4. This weight accounts for the transpose relation h A 0 = 4 h 0 A T h_{A0}=4h_{0A}^{\mathsf T} h A 0 = 4 h 0 A T used in Appendix § 8 .
The two-step determinant and spectral locus
The invariant sectors of Theorem 3.6 reduce the determinant calculation to three blocks. Table 1 summarizes the conclusions proved below: the determinant theorems 4.1, 4.5 and 4.7, the invertibility theorem 4.8, and the real-locus theorem 4.11. The factors Q 14 Q_{14} Q 14 , Q 12 Q_{12} Q 12 and sets Z Z Z , S S S are defined in the corresponding subsections.
Table 1. Sector contributions to the real two-step locus.
The fixed sector W 1 W_1 W 1
We use the fixed-sector coordinates ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ∈ H 3 ⊕ H (\vartheta_1,\vartheta_2,\vartheta_3;\eta)\in\mathbb{H}^3\oplus\mathbb{H} ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ∈ H 3 ⊕ H of Proposition 3.8 . The first three slots, indexed by A = ( 1 , 2 , 3 ) A=(1,2,3) A = ( 1 , 2 , 3 ) , correspond to H ⊗ A ⊂ W + \mathbb{H}\otimes A\subset W_+ H ⊗ A ⊂ W + , while slot 0 corresponds to ε γ ∈ W − \varepsilon_\gamma\in W_- ε γ ∈ W − . In these coordinates, the scalar part of P a P_a P a acts by R q ^ R_{\hat{q}} R q ^ on every slot.
Writing Φ h Φ − 1 \Phi h\Phi^{-1} Φ h Φ − 1 in blocks h A A h_{AA} h AA , h A 0 h_{A0} h A 0 , h 0 A h_{0A} h 0 A , h 00 h_{00} h 00 as in (B.1), we obtain
Φ P a Φ − 1 = ( P A A P A 0 P 0 A P 00 ) = ( h A A − R q ^ H 3 h A 0 h 0 A h 00 − R q ^ ) . \Phi P_a\Phi^{-1}
=
\begin{pmatrix}
P_{AA} & P_{A0}\\
P_{0A} & P_{00}
\end{pmatrix}
=
\begin{pmatrix}
h_{AA}-R_{\hat{q}}^{\mathbb{H}^3} & h_{A0}\\
h_{0A} & h_{00}-R_{\hat{q}}
\end{pmatrix}. Φ P a Φ − 1 = ( P AA P 0 A P A 0 P 00 ) = ( h AA − R q ^ H 3 h 0 A h A 0 h 00 − R q ^ ) .
where R q ^ H 3 R_{\hat{q}}^{\mathbb{H}^3} R q ^ H 3 denotes componentwise right multiplication on H 3 \mathbb{H}^3 H 3 .
To analyze the A A AA AA block, let r = ( u , v , z ) T \mathbf{r}=(u,v,z)^{\mathsf T} r = ( u , v , z ) T , and let r ∗ = r ˉ T \mathbf{r}^*=\bar{\mathbf{r}}^{\mathsf T} r ∗ = r ˉ T denote its quaternionic conjugate transpose. Then h A A = L M h_{AA}=L_M h AA = L M , where
M = − r r ∗ , and r ∗ r = 3. M=-\mathbf{r}\mathbf{r}^*,\qquad\text{and}\qquad\mathbf{r}^*\mathbf{r}=3. M = − r r ∗ , and r ∗ r = 3.
Thus H 3 \mathbb{H}^{3} H 3 decomposes into right quaternionic submodules
H 3 = U 3 ⊕ U 0 , where U 3 = r H and U 0 = ker r ∗ , \mathbb{H}^{3}=U_{3}\oplus U_{0},\qquad\text{where}\qquad U_{3}=r\mathbb{H}\quad\text{and}\quad U_{0}=\ker r^{*}, H 3 = U 3 ⊕ U 0 , where U 3 = r H and U 0 = ker r ∗ ,
of ranks one and two, respectively. On these submodules, P A A P_{AA} P AA acts as
P A A ∣ U 3 = − R q ^ + 3 and P A A ∣ U 0 = − R q ^ . P_{AA}|_{U_{3}}=-R_{\hat{q}+3}\qquad\text{and}\qquad P_{AA}|_{U_{0}}=-R_{\hat{q}}. P AA ∣ U 3 = − R q ^ + 3 and P AA ∣ U 0 = − R q ^ .
Consequently, using N = ∣ q ^ ∣ 2 N=|\hat{q}|^{2} N = ∣ q ^ ∣ 2 and N 3 = ∣ q ^ + 3 ∣ 2 = N + 6 a 0 + 9 N_{3}=|\hat{q}+3|^{2}=N+6a_{0}+9 N 3 = ∣ q ^ + 3 ∣ 2 = N + 6 a 0 + 9 , we obtain
det P A A = N 4 N 3 2 . \det P_{AA}=N^{4}N_{3}^{2}. det P AA = N 4 N 3 2 .
When N N 3 ≠ 0 NN_{3}\ne0 N N 3 = 0 , the inverse of P A A P_{AA} P AA is given by (B.2). The projections and the corresponding calculations are provided in Appendix B.
For N N 3 ≠ 0 NN_{3}\ne0 N N 3 = 0 , we eliminate the H 3 \mathbb{H}^{3} H 3 block using the Schur complement
S = P 00 − P 0 A P A A − 1 P A 0 . S=P_{00}-P_{0A}P_{AA}^{-1}P_{A0}. S = P 00 − P 0 A P AA − 1 P A 0 .
This is a real-linear operator on H \mathbb{H} H , represented by a real 4 × 4 4\times4 4 × 4 matrix in the basis ( 1 , u , v , z ) (1,u,v,z) ( 1 , u , v , z ) . Lemma B.3 constructs the real 4 × 4 4\times4 4 × 4 polynomial matrix M p o l M_{\mathrm{pol}} M pol , displayed in equation (B.8). Its entries belong to Q [ a 0 , a 1 , a 2 , a 3 ] \mathbb{Q}[a_{0},a_{1},a_{2},a_{3}] Q [ a 0 , a 1 , a 2 , a 3 ] , and we define
Q 14 : = det M p o l ∈ Q [ a 0 , a 1 , a 2 , a 3 ] . Q_{14}:=\det M_{\mathrm{pol}}\in\mathbb{Q}[a_{0},a_{1},a_{2},a_{3}]. Q 14 := det M pol ∈ Q [ a 0 , a 1 , a 2 , a 3 ] .
This polynomial definition applies at all coefficient points, including N N 3 = 0 NN_{3}=0 N N 3 = 0 . Appendix B supplies the determinant normalization and continuation argument used in the theorem below.
The explicit expression is shorter in the variables ( a 0 , N , t ) (a_{0},N,t) ( a 0 , N , t ) , where t = x − y t=x-y t = x − y . The change from ( a 0 , x , y ) (a_{0},x,y) ( a 0 , x , y ) is polynomial with polynomial inverse over Q \mathbb{Q} Q :
x = 1 2 ( N − a 0 2 + t ) , and y = 1 2 ( N − a 0 2 − t ) . x=\frac{1}{2}(N-a_{0}^{2}+t),\qquad\text{and}\qquad y=\frac{1}{2}(N-a_{0}^{2}-t). x = 2 1 ( N − a 0 2 + t ) , and y = 2 1 ( N − a 0 2 − t ) .
Theorem 4.1 (Determinant on the fixed sector). Using notation as above:
(i) The determinant of P a P_{a} P a on W 1 W_{1} W 1 is given by
det ( P a ∣ W 1 ) = N Q 14 . (24) \det(P_{a}|_{W_{1}})=NQ_{14}. \tag*{(24)} det ( P a ∣ W 1 ) = N Q 14 . (24) On N N 3 ≠ 0 NN_{3}\ne0 N N 3 = 0 , the Schur complement S S S satisfies
Q 14 = N 3 N 3 2 det R S . Q_{14}=N^{3}N_{3}^{2}\det_{\mathbb{R}}S. Q 14 = N 3 N 3 2 R det S . (ii) As an element of Q [ x , y ] [ a 0 ] \mathbb{Q}[x,y][a_{0}] Q [ x , y ] [ a 0 ] , the polynomial Q 14 ( a 0 , x , y ) Q_{14}(a_{0},x,y) Q 14 ( a 0 , x , y ) is monic of degree 14 in a 0 a_{0} a 0 . For its explicit expression, write N = a 0 2 + x + y N=a_{0}^{2}+x+y N = a 0 2 + x + y and t = x − y t=x-y t = x − y . Then:
Q 14 = 256 N ( 2 a 0 + 3 ) 2 t 2 + q 1 t + q 0 , where Q_{14}=256N(2a_{0}+3)^{2}t^{2}+q_{1}t+q_{0},\qquad\text{where} Q 14 = 256 N ( 2 a 0 + 3 ) 2 t 2 + q 1 t + q 0 , where q 1 = N [ 96 ( a 0 + 1 ) N 3 + 32 ( 26 a 0 2 + 37 a 0 + 23 ) N 2 + 128 ( 4 a 0 3 − 36 a 0 2 − 55 a 0 − 16 ) N − 1024 a 0 2 ( 2 a 0 + 3 ) ( 2 a 0 + 9 ) ] q_{1}=N\left[96(a_{0}+1)N^{3}+32(26a_{0}^{2}+37a_{0}+23)N^{2}+128(4a_{0}^{3}-36a_{0}^{2}-55a_{0}-16)N-1024a_{0}^{2}(2a_{0}+3)(2a_{0}+9)\right] q 1 = N [ 96 ( a 0 + 1 ) N 3 + 32 ( 26 a 0 2 + 37 a 0 + 23 ) N 2 + 128 ( 4 a 0 3 − 36 a 0 2 − 55 a 0 − 16 ) N − 1024 a 0 2 ( 2 a 0 + 3 ) ( 2 a 0 + 9 ) ] and q 0 = N 7 + 2 ( 8 a 0 + 9 ) N 6 + 3 ( 2 a 0 + 1 ) ( 6 a 0 + 11 ) N 5 − 4 ( 108 a 0 3 + 192 a 0 2 + 159 a 0 + 92 ) N 4 − 32 a 0 ( 54 a 0 3 + 29 a 0 2 − 52 a 0 − 30 ) N 3 + 256 a 0 ( 4 a 0 4 + 74 a 0 3 + 88 a 0 2 − 21 a 0 − 52 ) N 2 + 256 a 0 2 ( 7 a 0 + 8 ) ( 4 a 0 3 + 12 a 0 2 − 31 a 0 − 28 ) N − 8192 a 0 4 ( 2 a 0 + 3 ) ( 2 a 0 + 7 ) . (25) \begin{aligned}
\text{and}\qquad q_{0}={}&N^{7}+2(8a_{0}+9)N^{6}+3(2a_{0}+1)(6a_{0}+11)N^{5}\\
&-4(108a_{0}^{3}+192a_{0}^{2}+159a_{0}+92)N^{4}\\
&-32a_{0}(54a_{0}^{3}+29a_{0}^{2}-52a_{0}-30)N^{3}\\
&+256a_{0}(4a_{0}^{4}+74a_{0}^{3}+88a_{0}^{2}-21a_{0}-52)N^{2}\\
&+256a_{0}^{2}(7a_{0}+8)(4a_{0}^{3}+12a_{0}^{2}-31a_{0}-28)N\\
&-8192a_{0}^{4}(2a_{0}+3)(2a_{0}+7). \tag*{(25)}
\end{aligned} and q 0 = N 7 + 2 ( 8 a 0 + 9 ) N 6 + 3 ( 2 a 0 + 1 ) ( 6 a 0 + 11 ) N 5 − 4 ( 108 a 0 3 + 192 a 0 2 + 159 a 0 + 92 ) N 4 − 32 a 0 ( 54 a 0 3 + 29 a 0 2 − 52 a 0 − 30 ) N 3 + 256 a 0 ( 4 a 0 4 + 74 a 0 3 + 88 a 0 2 − 21 a 0 − 52 ) N 2 + 256 a 0 2 ( 7 a 0 + 8 ) ( 4 a 0 3 + 12 a 0 2 − 31 a 0 − 28 ) N − 8192 a 0 4 ( 2 a 0 + 3 ) ( 2 a 0 + 7 ) . (25)
Proof. Proposition B.4 gives the Schur normalization and global polynomial factorization in (i), including at N 3 = 0 N_3=0 N 3 = 0 . Proposition 3.7 then implies that Q 14 Q_{14} Q 14 is a polynomial in ( a 0 , x , y ) (a_0,x,y) ( a 0 , x , y ) . This proves (i) and the polynomial dependence asserted in (ii). The appendix evaluates the Schur determinant using Lemma A.3; see also [3 §8, (9), pp. 106–107] .
To check monicity, P a + a 0 I P_a+a_0I P a + a 0 I is independent of a 0 a_0 a 0 . Since dim R W 1 = 16 \dim_{\mathbb{R}}W_1=16 dim R W 1 = 16 , det ( P a ∣ W 1 ) \det(P_a|_{W_1}) det ( P a ∣ W 1 ) is monic of degree 16 in a 0 a_0 a 0 . As N = a 0 2 + x + y N=a_0^2+x+y N = a 0 2 + x + y is monic of degree 2, the factorization in (i) shows that Q 14 Q_{14} Q 14 is monic of degree 14 in a 0 a_0 a 0 .
The alternate expression (B.9) gives the same coefficients after x , y x,y x , y are written in terms of ( a 0 , N , t ) (a_0,N,t) ( a 0 , N , t ) . For the direct evaluation, substitute N = a 0 2 + a 1 2 + a 2 2 + a 3 2 N=a_0^2+a_1^2+a_2^2+a_3^2 N = a 0 2 + a 1 2 + a 2 2 + a 3 2 and t = a 1 2 + a 2 2 − a 3 2 t=a_1^2+a_2^2-a_3^2 t = a 1 2 + a 2 2 − a 3 2 into (4.2). Exact symbolic determinant calculation shows that the determinant of the polynomial matrix (B.8) minus this expression expands to zero in Q [ a 0 , a 1 , a 2 , a 3 ] \mathbb{Q}[a_0,a_1,a_2,a_3] Q [ a 0 , a 1 , a 2 , a 3 ] .
Definition 4.2. We write
Z = { a ∈ R 4 : Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) = 0 } (26) Z=\{\mathbf{a}\in\mathbb{R}^4:Q_{14}(a_0,a_1^2+a_2^2,a_3^2)=0\} \tag*{(26)} Z = { a ∈ R 4 : Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) = 0 } (26)
for the real zero set of Q 14 ( a 0 , x , y ) Q_{14}(a_0,x,y) Q 14 ( a 0 , x , y ) in the coefficient space. We also write Q 14 ( a ) Q_{14}(\mathbf{a}) Q 14 ( a ) for its pullback Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) Q_{14}(a_0,a_1^2+a_2^2,a_3^2) Q 14 ( a 0 , a 1 2 + a 2 2 , a 3 2 ) when a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 .
The W3 sector
By Theorem 3.6, the sector W 3 W_3 W 3 lies in B ^ ⊗ B \widehat{B}\otimes B B ⊗ B , hence entirely in the reversed half W − = B ^ ⊗ V W_-=\widehat{B}\otimes V W − = B ⊗ V . The right-multiplication term on this sector is R q ^ R_{\widehat{q}} R q . First we determine the action of the right product and h h h on elements of the form w X , γ w_{X,\gamma} w X , γ , as defined in (3.14).
Lemma 4.3. For X ∈ s u ( 2 ) X\in\mathfrak{su}(2) X ∈ su ( 2 ) and γ , p ∈ H \gamma,p\in\mathbb{H} γ , p ∈ H ,
R p w X , γ = w X , p γ and h w X , γ = w X , − 4 Re γ . R_p w_{X,\gamma}=w_{X,p\gamma}\qquad\text{and}\qquad h w_{X,\gamma}=w_{X,-4\operatorname{Re}\gamma}. R p w X , γ = w X , p γ and h w X , γ = w X , − 4 Re γ . Proof. Recall from (3.14) that
w X , γ = ∑ α = 0 3 ( ( X ξ ^ α ) γ ) ⊗ ξ α . w_{X,\gamma}=\sum_{\alpha=0}^{3}((X\widehat{\xi}_{\alpha})\gamma)\otimes\xi_{\alpha}. w X , γ = α = 0 ∑ 3 (( X ξ α ) γ ) ⊗ ξ α . Since X X X preserves B ^ \widehat{B} B , the Cayley–Dickson identity ( ξ ^ γ ) p = ξ ^ ( p γ ) (\widehat{\xi}\gamma)p=\widehat{\xi}(p\gamma) ( ξ γ ) p = ξ ( p γ ) from (A.3) gives
R p w X , γ = ∑ α = 0 3 ( ( ( X ξ ^ α ) γ ) p ) ⊗ ξ α R_p w_{X,\gamma}=\sum_{\alpha=0}^{3}(((X\widehat{\xi}_{\alpha})\gamma)p)\otimes\xi_{\alpha} R p w X , γ = α = 0 ∑ 3 ((( X ξ α ) γ ) p ) ⊗ ξ α = ∑ α = 0 3 ( ( X ξ ^ α ) ( p γ ) ) ⊗ ξ α = w X , p γ . =\sum_{\alpha=0}^{3}((X\widehat{\xi}_{\alpha})(p\gamma))\otimes\xi_{\alpha}=w_{X,p\gamma}. = α = 0 ∑ 3 (( X ξ α ) ( p γ )) ⊗ ξ α = w X , p γ . To compute the action of h h h , identify B ^ ⊗ B \widehat{B}\otimes B B ⊗ B with End R ( B ^ ) \operatorname{End}_{\mathbb{R}}(\widehat{B}) End R ( B ) using the paired orthonormal bases ( ξ ^ α ) (\widehat{\xi}_{\alpha}) ( ξ α ) and ( ξ α ) (\xi_{\alpha}) ( ξ α ) : a tensor a ⊗ ξ α a\otimes\xi_{\alpha} a ⊗ ξ α corresponds to the map ξ ^ β ↦ δ α β a \widehat{\xi}_{\beta}\mapsto\delta_{\alpha\beta}a ξ β ↦ δ α β a . Thus w X , γ w_{X,\gamma} w X , γ corresponds to the endomorphism
T X , γ = R γ X with T X , γ ( ξ ^ α ) = ( X ξ ^ α ) γ . T_{X,\gamma}=R_{\gamma}X\qquad\text{with}\qquad T_{X,\gamma}(\widehat{\xi}_{\alpha})=(X\widehat{\xi}_{\alpha})\gamma. T X , γ = R γ X with T X , γ ( ξ α ) = ( X ξ α ) γ .
Under the fixed identification V ^ ≅ V \widehat{V} \cong V V ≅ V , this endomorphism, extended by zero on A A A , represents the corresponding tensor in V ⊗ V V \otimes V V ⊗ V . By real linearity, it suffices to consider γ = 1 \gamma= 1 γ = 1 and γ ∈ Im H \gamma\in\operatorname{Im}\mathbb{H} γ ∈ Im H .
For γ = 1 \gamma= 1 γ = 1 , we have T X , 1 = X T_{X,1} = X T X , 1 = X . Since X ∈ s u ( 2 ) ⊂ g 2 X \in\mathfrak{su}(2) \subset\mathfrak{g}_2 X ∈ su ( 2 ) ⊂ g 2 acts as a skew-adjoint endomorphism of V V V , its corresponding alternating 2-tensor lies in Λ 14 2 V \Lambda^{2}_{14}V Λ 14 2 V . Hence, using the real spectrum of h h h from Proposition 2.6,
h w X , 1 = − 4 w X , 1 . h w_{X,1} = -4w_{X,1}. h w X , 1 = − 4 w X , 1 .
Now suppose γ ∈ Im H \gamma\in\operatorname{Im}\mathbb{H} γ ∈ Im H . Both X X X and R γ R_{\gamma} R γ are skew-adjoint on B ^ \widehat{B} B . Since X X X is a derivation vanishing on H \mathbb{H} H ,
X ( ξ ^ γ ) = ( X ξ ^ ) γ , X(\widehat{\xi}\gamma) = (X\widehat{\xi})\gamma, X ( ξ γ ) = ( X ξ ) γ ,
so X X X commutes with R γ R_{\gamma} R γ . Consequently, T X , γ = R γ X T_{X,\gamma} = R_{\gamma}X T X , γ = R γ X is self-adjoint and therefore corresponds to a symmetric 2-tensor. This tensor is also trace-free. Indeed, for fixed γ \gamma γ , the functional
X ⟼ tr ( R γ X ) X \longmapsto\operatorname{tr}(R_{\gamma}X) X ⟼ tr ( R γ X )
is Ad ( K ) \operatorname{Ad}(K) Ad ( K ) -invariant, since R γ R_{\gamma} R γ commutes with K K K . As s u ( 2 ) \mathfrak{su}(2) su ( 2 ) has no nonzero invariant linear functionals, the trace vanishes. Thus w X , γ ∈ S 0 2 V w_{X,\gamma} \in S^{2}_{0}V w X , γ ∈ S 0 2 V , where h h h acts by zero (§ 3 ).
Combining the two cases gives
h w X , γ = w X , − 4 Re γ h w_{X,\gamma} = w_{X,-4\operatorname{Re}\gamma} h w X , γ = w X , − 4 Re γ
for every γ ∈ H \gamma\in\mathbb{H} γ ∈ H .
Proposition 4.4.
(i) On W 3 W_{3} W 3 , the two-step pencil agrees with the restriction of the one-step pencil h − R q ˇ h - R_{\check{q}} h − R q ˇ . For a pair with reduced point a a a , this is the one-step comparison operator at q ˇ = λ 2 λ 1 \check{q} = \lambda_{2}\lambda_{1} q ˇ = λ 2 λ 1 .
(ii) Under the K K K -equivariant identification s u ( 2 ) ⊗ H ≅ W 3 \mathfrak{su}(2) \otimes\mathbb{H} \cong W_{3} su ( 2 ) ⊗ H ≅ W 3 given by X ⊗ γ ↦ w X , γ X \otimes\gamma\mapsto w_{X,\gamma} X ⊗ γ ↦ w X , γ in (3.14),
P a ∣ W 3 = I s u ( 2 ) ⊗ P 3 , where P 3 ( γ ) = − 4 Re γ − q ˇ γ . P_{a}|_{W_{3}} = I_{\mathfrak{su}(2)} \otimes P_{3}, \qquad\text{where} \qquad P_{3}(\gamma) = -4\operatorname{Re}\gamma- \check{q}\gamma. P a ∣ W 3 = I su ( 2 ) ⊗ P 3 , where P 3 ( γ ) = − 4 Re γ − q ˇ γ .
Proof. Since B ^ ⊗ V ⊂ W − \widehat{B} \otimes V \subset W_{-} B ⊗ V ⊂ W − , part (i) follows from Proposition 3.5, with q ˇ = λ 2 λ 1 \check{q} = \lambda_{2}\lambda_{1} q ˇ = λ 2 λ 1 for a reduced pair by Proposition 3.2 . The explicit action in part (ii) follows from Lemma 4.3 above.
We can now compute the determinant of the pencil on the sector W 3 W_{3} W 3 .
Theorem 4.5. Recall from Proposition 4.4 that P a ∣ W 3 P_{a}|_{W_{3}} P a ∣ W 3 is identified with I s u ( 2 ) ⊗ P 3 I_{\mathfrak{su}(2)} \otimes P_{3} I su ( 2 ) ⊗ P 3 , where P 3 : H → H P_{3} : \mathbb{H} \to\mathbb{H} P 3 : H → H is the real-linear map P 3 ( γ ) = − 4 Re γ − q ˇ γ P_{3}(\gamma) = -4\operatorname{Re}\gamma- \check{q}\gamma P 3 ( γ ) = − 4 Re γ − q ˇ γ . Then:
(i) The determinant satisfies the polynomial identity
det R ( P a ∣ W 3 ) = p 3 3 , where p 3 = N ( N + 4 a 0 ) = ∣ q ˇ ∣ 2 ( ∣ q ˇ + 2 ∣ 2 − 4 ) . (27) \det_{\mathbb{R}}(P_{a}|_{W_{3}}) = p_{3}^{3}, \qquad\text{where} \qquad p_{3} = N(N + 4a_{0}) = |\check{q}|^{2}(|\check{q} + 2|^{2} - 4). \tag*{(27)} R det ( P a ∣ W 3 ) = p 3 3 , where p 3 = N ( N + 4 a 0 ) = ∣ q ˇ ∣ 2 ( ∣ q ˇ + 2 ∣ 2 − 4 ) . (27) In particular, p 3 p_{3} p 3 depends on a a a only through a 0 a_{0} a 0 and x + y x+y x + y .
(ii) For a ∈ R 4 a \in\mathbb{R}^{4} a ∈ R 4 , the restriction P a ∣ W 3 P_{a}|_{W_{3}} P a ∣ W 3 is singular precisely on the round 3-sphere
S = { a ∈ R 4 : ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 = 4 } . S = \left\{a \in\mathbb{R}^{4} : (a_{0} + 2)^{2} + a_{1}^{2} + a_{2}^{2} + a_{3}^{2} = 4\right\}. S = { a ∈ R 4 : ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 = 4 } . This sphere contains the origin and lies in − 4 ≤ a 0 ≤ 0 -4 \leq a_{0} \leq0 − 4 ≤ a 0 ≤ 0 .
(iii) The kernel of P 3 P_3 P 3 is
ker P 3 = { { 0 } , a ∉ S , R q ~ , a ∈ S ∖ { 0 } , Im H , a = 0. \ker P_3 =
\begin{cases}
\{0\}, & \mathbf{a} \notin S,\\
\mathbb{R}\tilde{q}, & \mathbf{a} \in S \setminus\{0\},\\
\operatorname{Im}\mathbb{H}, & \mathbf{a}=0.
\end{cases} ker P 3 = ⎩ ⎨ ⎧ { 0 } , R q ~ , Im H , a ∈ / S , a ∈ S ∖ { 0 } , a = 0.
Under the identification W 3 ≅ s u ( 2 ) ⊗ H W_3 \cong\mathfrak{su}(2) \otimes\mathbb{H} W 3 ≅ su ( 2 ) ⊗ H , the kernel of P a ∣ W 3 P_{\mathbf{a}|W_3} P a ∣ W 3 corresponds to
s u ( 2 ) ⊗ ker P 3 , \mathfrak{su}(2) \otimes\ker P_3, su ( 2 ) ⊗ ker P 3 ,
and therefore has real dimension 0 0 0 , 3 3 3 , or 9 9 9 , respectively.
Proof. In the real basis ( 1 , u , v , z ) (1,u,v,z) ( 1 , u , v , z ) of H \mathbb{H} H , let e 0 = ( 1 , 0 , 0 , 0 ) T e_0=(1,0,0,0)^{\mathsf{T}} e 0 = ( 1 , 0 , 0 , 0 ) T . Since e 0 e 0 T e_0e_0^{\mathsf{T}} e 0 e 0 T represents the projection γ ↦ Re γ \gamma\mapsto\operatorname{Re}\gamma γ ↦ Re γ , we have
P 3 = − ( L q ~ + 4 e 0 e 0 T ) . P_3=-(L_{\tilde{q}}+4e_0e_0^{\mathsf{T}}). P 3 = − ( L q ~ + 4 e 0 e 0 T ) .
For N = ∣ q ~ ∣ 2 ≠ 0 N=|\tilde{q}|^2\ne0 N = ∣ q ~ ∣ 2 = 0 , left multiplication by q ~ \tilde{q} q ~ satisfies
det R L q ~ = N 2 and L q ~ − 1 = 1 N L q ~ ‾ . \det_{\mathbb{R}}L_{\tilde{q}}=N^2
\qquad\text{and}\qquad
L_{\tilde{q}}^{-1}=\frac{1}{N}L_{\overline{\tilde{q}}}. R det L q ~ = N 2 and L q ~ − 1 = N 1 L q ~ .
In particular, e 0 T L q ~ − 1 e 0 = a 0 / N e_0^{\mathsf{T}}L_{\tilde{q}}^{-1}e_0=a_0/N e 0 T L q ~ − 1 e 0 = a 0 / N . We can thus compute the determinant as:
det R P 3 = N 2 ( 1 + 4 e 0 T L q ~ − 1 e 0 ) = N 2 ( 1 + 4 a 0 N ) = N ( N + 4 a 0 ) . \det_{\mathbb{R}}P_3=N^2\left(1+4e_0^{\mathsf{T}}L_{\tilde{q}}^{-1}e_0\right)
=N^2\left(1+\frac{4a_0}{N}\right)=N(N+4a_0). R det P 3 = N 2 ( 1 + 4 e 0 T L q ~ − 1 e 0 ) = N 2 ( 1 + N 4 a 0 ) = N ( N + 4 a 0 ) .
Both sides are polynomials in a \mathbf{a} a , so the identity extends to N = 0 N=0 N = 0 . Since P a ∣ W 3 = I s u ( 2 ) ⊗ P 3 P_{\mathbf{a}|W_3}=I_{\mathfrak{su}(2)}\otimes P_3 P a ∣ W 3 = I su ( 2 ) ⊗ P 3 and dim R s u ( 2 ) = 3 \dim_{\mathbb{R}}\mathfrak{su}(2)=3 dim R su ( 2 ) = 3 , its determinant is p 3 3 p_3^3 p 3 3 , proving (i).
For real a ∈ R 4 \mathbf{a}\in\mathbb{R}^4 a ∈ R 4 , N = 0 N=0 N = 0 only at the origin. Moreover,
N + 4 a 0 = ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 − 4. N+4a_0=(a_0+2)^2+a_1^2+a_2^2+a_3^2-4. N + 4 a 0 = ( a 0 + 2 ) 2 + a 1 2 + a 2 2 + a 3 2 − 4.
Thus the zero set of p 3 p_3 p 3 is exactly the sphere S S S , proving (ii). Finally, the kernel equation is
q ~ γ = − 4 Re γ . \tilde{q}\gamma=-4\operatorname{Re}\gamma. q ~ γ = − 4 Re γ .
If q ~ ≠ 0 \tilde{q}\ne0 q ~ = 0 , multiplication by q ~ − 1 = q ~ ‾ / N \tilde{q}^{-1}=\overline{\tilde{q}}/N q ~ − 1 = q ~ / N gives
γ = − 4 Re γ N q ~ ‾ . \gamma=-\frac{4\operatorname{Re}\gamma}{N}\overline{\tilde{q}}. γ = − N 4 Re γ q ~ .
Hence every kernel element is a real multiple of q ~ ‾ \overline{\tilde{q}} q ~ . Conversely,
P 3 ( q ~ ‾ ) = − ( N + 4 a 0 ) , P_3(\overline{\tilde{q}})=-(N+4a_0), P 3 ( q ~ ) = − ( N + 4 a 0 ) ,
so this line is the kernel precisely when a ∈ S ∖ { 0 } \mathbf{a}\in S\setminus\{0\} a ∈ S ∖ { 0 } . At the origin, q ~ = 0 \tilde{q}=0 q ~ = 0 , and P 3 ( γ ) = − 4 Re γ P_3(\gamma)=-4\operatorname{Re}\gamma P 3 ( γ ) = − 4 Re γ , whose kernel is Im H \operatorname{Im}\mathbb{H} Im H . The corresponding kernels on W 3 W_3 W 3 follow by tensoring with s u ( 2 ) \mathfrak{su}(2) su ( 2 ) , giving the stated dimensions. □ \square □
Thus S S S is the one-step circle condition ∣ q ~ + 2 ∣ = 2 |\tilde{q}+2|=2 ∣ q ~ + 2∣ = 2 at the reversed product (Corollary 4.13 ).
The W 4 W_4 W 4 sector
The sector W 4 = ( H ⊗ B ) ⊕ ( B ^ ⊗ A ) W_4 = (\mathbb{H} \otimes B) \oplus(\widehat{B} \otimes A) W 4 = ( H ⊗ B ) ⊕ ( B ⊗ A ) is the direct sum of seven copies of the four-dimensional real K K K -module of quaternionic type. We first choose equivariant coordinates in which the multiplication part of P a P_a P a acts by the same quaternion q ^ \widehat{q} q on all seven slots. We then identify the resulting quaternionic matrix and compute its reduced norm. Finally, a sum of squares determines its real zero set.
In the original halves, the multiplication term involves q ^ \widehat{q} q on W + W_+ W + and q ˇ \check{q} q ˇ on W − W_- W − . The coordinates below conjugate the action on the second part to the common row action q ^ I 7 \widehat{q}I_7 q I 7 ; the coupling between the halves remains in the off-diagonal blocks of M h M_h M h .
For γ ∈ H \gamma\in\mathbb{H} γ ∈ H and ξ ^ ∈ B ^ \widehat{\xi} \in\widehat{B} ξ ∈ B , define
Ψ γ ( ξ ^ ) = 1 2 ∑ i = 0 3 ρ ^ i ⊗ ( ( ρ ^ i γ ) ξ ^ ) . \Psi_\gamma(\widehat{\xi}) = \frac{1}{2}\sum_{i=0}^{3} \widehat{\rho}_i \otimes((\widehat{\rho}_i\gamma)\widehat{\xi}). Ψ γ ( ξ ) = 2 1 i = 0 ∑ 3 ρ i ⊗ (( ρ i γ ) ξ ) .
where the second factor is in B ⊂ V B \subset V B ⊂ V . For f = ( f 0 , … , f 6 ) ∈ B ^ 7 f = (f_0,\ldots,f_6) \in\widehat{B}^7 f = ( f 0 , … , f 6 ) ∈ B 7 , define
J ( f ) = ∑ k = 0 3 Ψ ρ ^ k ( f k ) + ∑ j = 1 3 ( f 3 + j z ) ⊗ ρ j . (28) \mathcal{J}(f) = \sum_{k=0}^{3}\Psi_{\widehat{\rho}_k}(f_k) + \sum_{j=1}^{3}(f_{3+j}z) \otimes\rho_j. \tag*{(28)} J ( f ) = k = 0 ∑ 3 Ψ ρ k ( f k ) + j = 1 ∑ 3 ( f 3 + j z ) ⊗ ρ j . (28)
This is an isometric, K K K -equivariant isomorphism; the proof below also shows that Ψ γ \Psi_\gamma Ψ γ is independent of the orthonormal basis of H \mathbb{H} H used in its sum.
Fix a unit ℓ ∈ B ^ \ell\in\widehat{B} ℓ ∈ B and write f k = s k ℓ f_k = s_k\ell f k = s k ℓ with s = ( s 0 , … , s 6 ) ∈ H 7 \mathbf{s} = (s_0,\ldots,s_6) \in\mathbb{H}^7 s = ( s 0 , … , s 6 ) ∈ H 7 . For an equivariant operator T T T , its matrix M T \mathcal{M}_T M T is specified by T J ( f ) = J ( f ′ ) T\mathcal{J}(f)=\mathcal{J}(f') T J ( f ) = J ( f ′ ) , where f k ′ = ∑ i = 0 6 ( M T ) i k f i f'_k=\sum_{i=0}^{6}(\mathcal{M}_T)_{ik}f_i f k ′ = ∑ i = 0 6 ( M T ) ik f i . A quaternion entry acts on B ^ \widehat{B} B by left multiplication. Since γ ( s ℓ ) = ( s γ ) ℓ \gamma(s\ell)=(s\gamma)\ell γ ( s ℓ ) = ( s γ ) ℓ , the induced coordinate action is T M T ( s ) = s M T T\mathcal{M}_T(\mathbf{s})=\mathbf{s}\mathcal{M}_T T M T ( s ) = s M T . Thus T N ∘ T M = T M N T_N\circ T_M=T_{MN} T N ∘ T M = T M N : u I 7 uI_7 u I 7 followed by v I 7 vI_7 v I 7 gives z I 7 zI_7 z I 7 .
In these coordinates the full matrix of h h h is
M h = ( M + + M + − M − + M − − ) , M_h =
\begin{pmatrix}
M_{++} & M_{+-}\\
M_{-+} & M_{--}
\end{pmatrix}, M h = ( M ++ M −+ M +− M −− ) ,
M + + = diag ( 2 , − 2 , − 2 , − 2 ) and M − + = M + − ∗ , M_{++}=\operatorname{diag}(2,-2,-2,-2)
\qquad\text{and}\qquad
M_{-+}=M_{+-}^{*}, M ++ = diag ( 2 , − 2 , − 2 , − 2 ) and M −+ = M +− ∗ ,
M + − = ( − v u − 1 z − 1 − u 1 z − v u v z ) and M − − = ( − 1 − z − v z − 1 u v − u − 1 ) . (29) M_{+-}=
\begin{pmatrix}
-v & u & -1\\
z & -1 & -u\\
1 & z & -v\\
u & v & z
\end{pmatrix}
\qquad\text{and}\qquad
M_{--}=
\begin{pmatrix}
-1 & -z & -v\\
z & -1 & u\\
v & -u & -1
\end{pmatrix}. \tag*{(29)} M +− = − v z 1 u u − 1 z v − 1 − u − v z and M −− = − 1 z v − z − 1 − u − v u − 1 . (29)
where M + − ∗ M_{+-}^{*} M +− ∗ is the conjugate transpose. Equivalently, its indexed entries are ( M + − ) k , 3 + j = ρ ^ j ρ ^ k z (M_{+-})_{k,3+j}=\widehat{\rho}_j\widehat{\rho}_k z ( M +− ) k , 3 + j = ρ j ρ k z and ( M − − ) 3 + i , 3 + j = z ‾ ρ ^ j ρ ^ i z (M_{--})_{3+i,3+j}=\overline{z}\widehat{\rho}_j\widehat{\rho}_i z ( M −− ) 3 + i , 3 + j = z ρ j ρ i z , for k = 0 , … , 3 k=0,\ldots,3 k = 0 , … , 3 and i , j = 1 , 2 , 3 i,j=1,2,3 i , j = 1 , 2 , 3 . The first four slots correspond to H ⊗ B ⊂ W + \mathbb{H}\otimes B\subset W_+ H ⊗ B ⊂ W + and the last three to B ^ ⊗ A ⊂ W − \widehat{B}\otimes A\subset W_- B ⊗ A ⊂ W − , whence the subscripts.
For an equivariant operator T T T represented by M ∈ M 7 ( H ) \mathcal{M}\in M_7(\mathbb{H}) M ∈ M 7 ( H ) , its real determinant is
det R T = Nrd ( M ) 2 where Nrd ( M ) = det C χ ( M ) , \det_{\mathbb{R}}T=\operatorname{Nrd}(\mathcal{M})^2
\qquad\text{where}\qquad
\operatorname{Nrd}(\mathcal{M})=\det_{\mathbb{C}}\chi(\mathcal{M}), R det T = Nrd ( M ) 2 where Nrd ( M ) = C det χ ( M ) ,
where Nrd \operatorname{Nrd} Nrd denotes the quaternionic reduced norm and χ : M 7 ( H ) → M 14 ( C ) \chi:M_7(\mathbb{H})\to M_{14}(\mathbb{C}) χ : M 7 ( H ) → M 14 ( C ) is induced by a unital real-algebra embedding H ↪ M 2 ( C ) \mathbb{H}\hookrightarrow M_2(\mathbb{C}) H ↪ M 2 ( C ) . The reduced norm is real, nonnegative, and independent of the choice of embedding. It is the Study determinant Sdet \operatorname{Sdet} Sdet of Aslaksen, and det R = Sdet 2 \det_{\mathbb{R}}=\operatorname{Sdet}^2 det R = Sdet 2 [1 Theorems 5 and 9, and the remark on p. 63] . For definition of χ \chi χ , its multiplicativity, and the nonnegativity of det C χ \det_{\mathbb{C}}\chi det C χ , see [10 Theorems 4.2 and 8.1, Proposition 4.2 and its corollary] .
We use the explicit embedding
χ ( α + β u + γ v + ε z ) = ( α + i β γ + i ε − γ + i ε α − i β ) , for α , β , γ , ε ∈ R , (30) \chi(\alpha+\beta u+\gamma v+\varepsilon z)=\begin{pmatrix}\alpha+i\beta& \gamma+i\varepsilon\\ -\gamma+i\varepsilon& \alpha-i\beta\end{pmatrix}, \qquad\text{for }\alpha,\beta,\gamma,\varepsilon\in\mathbb{R}, \tag*{(30)} χ ( α + β u + γ v + ε z ) = ( α + i β − γ + i ε γ + i ε α − i β ) , for α , β , γ , ε ∈ R , (30)
and apply χ \chi χ entrywise to quaternionic matrices.
Proposition 4.6.
(i) Under J \mathcal{J} J , the algebra of K K K -equivariant operators on W 4 W_4 W 4 is M 7 ( H ) M_7(\mathbb{H}) M 7 ( H ) , acting on quaternionic row vectors from the right.
(ii) The operators R u R_u R u , R v R_v R v and R v R u R_vR_u R v R u are represented by u I 7 uI_7 u I 7 , v I 7 vI_7 v I 7 and z I 7 zI_7 z I 7 , respectively, and h h h by the quaternion-Hermitian matrix (4.6). Consequently P a ∣ W 4 P_a|_{W_4} P a ∣ W 4 is represented by M h − q ^ I 7 M_h-\hat{q}I_7 M h − q ^ I 7 , and
det R ( P a ∣ W 4 ) = Nrd ( M h − q ^ I 7 ) 2 . \det_{\mathbb{R}}(P_a|_{W_4})=\operatorname{Nrd}(M_h-\hat{q}I_7)^2. R det ( P a ∣ W 4 ) = Nrd ( M h − q ^ I 7 ) 2 .
Proof. Determinant convention. The equivariant identification with B ^ 7 \widehat{B}^7 B 7 gives the commutant M 7 ( H ) M_7(\mathbb{H}) M 7 ( H ) by Theorem 3.6(d), with the row action just described. This proves (i).
For a matrix M \mathcal{M} M , the row map T M : s ↦ s M T_{\mathcal{M}}:s\mapsto s\mathcal{M} T M : s ↦ s M commutes with left multiplication by C = R ⊕ R u \mathbb{C}=\mathbb{R}\oplus\mathbb{R}u C = R ⊕ R u , so it is C \mathbb{C} C -linear on H 7 ≅ C 14 \mathbb{H}^7\cong\mathbb{C}^{14} H 7 ≅ C 14 . Its complex matrix is the transpose of the entrywise image χ ′ ( M ) \chi'(\mathcal{M}) χ ′ ( M ) under a unital R \mathbb{R} R -algebra embedding χ ′ : H → M 2 ( C ) \chi':\mathbb{H}\to M_2(\mathbb{C}) χ ′ : H → M 2 ( C ) , so its complex determinant is det χ ′ ( M ) \det\chi'(\mathcal{M}) det χ ′ ( M ) .
Any two unital R \mathbb{R} R -algebra embeddings H → M 2 ( C ) \mathbb{H}\to M_2(\mathbb{C}) H → M 2 ( C ) are conjugate, by the Skolem–Noether theorem for H ⊗ C ≅ M 2 ( C ) \mathbb{H}\otimes\mathbb{C}\cong M_2(\mathbb{C}) H ⊗ C ≅ M 2 ( C ) ; hence det χ ′ ( M ) = det χ ( M ) \det\chi'(\mathcal{M})=\det\chi(\mathcal{M}) det χ ′ ( M ) = det χ ( M ) , and the choice of χ \chi χ does not matter.
This determinant is real because χ ( M ) \chi(\mathcal{M}) χ ( M ) commutes with a conjugate-linear map J \mathcal{J} J with J 2 = − 1 \mathcal{J}^2=-1 J 2 = − 1 . Finally, the real determinant of a C \mathbb{C} C -linear map is the squared modulus of its complex determinant, so det R T M = Nrd ( M ) 2 \det_{\mathbb{R}}T_{\mathcal{M}}=\operatorname{Nrd}(\mathcal{M})^2 det R T M = Nrd ( M ) 2 , since Nrd ( M ) \operatorname{Nrd}(\mathcal{M}) Nrd ( M ) is real.
Temporary coordinates. To prove (ii), we first compute in tensor-product coordinates and then make a unitary change that sends the multiplication part to right multiplication by q ^ \hat{q} q ^ in every slot. The same change gives the displayed matrix of h h h and the map J \mathcal{J} J in (4.5). Fix a unit ℓ ∈ B ^ \ell\in\widehat{B} ℓ ∈ B , so that O = H ⊕ H ℓ \mathcal{O}=\mathbb{H}\oplus\mathbb{H}\ell O = H ⊕ H ℓ with the Cayley–Dickson rules (A.2). The temporary coordinates ω = ( ω 0 , … , ω 6 ) ∈ H 7 \omega=(\omega_0,\ldots,\omega_6)\in\mathbb{H}^7 ω = ( ω 0 , … , ω 6 ) ∈ H 7 are given by
w = ∑ k = 0 3 ρ ^ k ⊗ ( ω k ℓ ) + ∑ j = 1 3 ( ω 3 + j ℓ ) ⊗ ρ j . w=\sum_{k=0}^{3}\widehat{\rho}_k\otimes(\omega_k\ell)+\sum_{j=1}^{3}(\omega_{3+j}\ell)\otimes\rho_j. w = k = 0 ∑ 3 ρ k ⊗ ( ω k ℓ ) + j = 1 ∑ 3 ( ω 3 + j ℓ ) ⊗ ρ j .
where the second factor of the first sum is read in B ⊂ V B\subset V B ⊂ V ; they are not the coordinates of J \mathcal{J} J . The group SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) acts by ω k ↦ g ω k \omega_k\mapsto g\omega_k ω k ↦ g ω k (Appendix § 7 , equation (A.5)), so equivariant operators act on these rows from the right.
Matrix before the coordinate change. By Definition 2.3 and bilinearity, h ( p ⊗ b ) = ∑ i ( ( e i ^ b ^ ) p ) ⊗ e i h(p\otimes b)=\sum_i((\widehat{e_i}\widehat{b})p)\otimes e_i h ( p ⊗ b ) = ∑ i (( e i b ) p ) ⊗ e i for p ∈ O p\in\mathcal{O} p ∈ O and b ∈ V b\in V b ∈ V , where b ^ = ∑ j b j e ^ j \widehat{b}=\sum_j b_j\widehat{e}_j b = ∑ j b j e j for b = ∑ j b j e j b=\sum_j b_je_j b = ∑ j b j e j (equation (2.2)). Expanding this with the Cayley–Dickson rules gives the following matrix entries in the temporary coordinates, for i , k = 0 , … , 3 i,k=0,\ldots,3 i , k = 0 , … , 3 and j , l = 1 , 2 , 3 j,l=1,2,3 j , l = 1 , 2 , 3 :
( M h ′ ) i k = − ρ ^ k ρ ^ i , ( M h ′ ) i , 3 + j = ρ ^ j ρ ^ i , (M'_h)_{ik}=-\widehat{\rho}_k\widehat{\rho}_i,\qquad(M'_h)_{i,3+j}=\widehat{\rho}_j\widehat{\rho}_i, ( M h ′ ) ik = − ρ k ρ i , ( M h ′ ) i , 3 + j = ρ j ρ i ,
( M h ′ ) 3 + j , k = − ρ ^ k ρ ^ j , ( M h ′ ) 3 + j , 3 + l = ρ ^ i ρ ^ j . (M'_h)_{3+j,k}=-\widehat{\rho}_k\widehat{\rho}_j,\qquad(M'_h)_{3+j,3+l}=\widehat{\rho}_i\widehat{\rho}_j. ( M h ′ ) 3 + j , k = − ρ k ρ j , ( M h ′ ) 3 + j , 3 + l = ρ i ρ j .
For the scalar part, R p R_p R p with p ∈ H p \in\mathbb{H} p ∈ H acts on the first four slots by the real matrix with entries ⟨ ρ ^ k , ρ ^ i p ⟩ \langle\hat{\rho}_k,\hat{\rho}_i p\rangle ⟨ ρ ^ k , ρ ^ i p ⟩ (i i i the row) and on the last three by ω ↦ ω p ˉ \omega\mapsto\omega\bar{p} ω ↦ ω p ˉ . By the sign-twist identity R v R u = R z ∘ ( σ ⊗ 1 ) R_vR_u=R_z\circ(\sigma\otimes1) R v R u = R z ∘ ( σ ⊗ 1 ) of Proposition 3.4, R v R u R_vR_u R v R u is R z R_z R z on the first four slots and ω ↦ ω z \omega\mapsto\omega z ω ↦ ω z on the last three.
Common scalar action. Change coordinates by
η k = 1 2 ∑ i = 0 3 ω i ρ ^ k ρ ^ i , k = 0 , … , 3 , \eta_k=\frac{1}{2}\sum_{i=0}^{3}\omega_i\hat{\rho}_k\hat{\rho}_i,\qquad k=0,\ldots,3, η k = 2 1 i = 0 ∑ 3 ω i ρ ^ k ρ ^ i , k = 0 , … , 3 ,
η 3 + j = ω 3 + j z , j = 1 , 2 , 3. \eta_{3+j}=\omega_{3+j}z,\qquad j=1,2,3. η 3 + j = ω 3 + j z , j = 1 , 2 , 3.
Equivalently,
η = ω T , where T = diag ( T , z I 3 ) , T i k = 1 2 ρ ^ k ρ ^ i . \eta=\omega\mathcal{T},\qquad\text{where}\qquad\mathcal{T}=\operatorname{diag}(T,zI_3),\qquad T_{ik}=\frac{1}{2}\hat{\rho}_k\hat{\rho}_i. η = ω T , where T = diag ( T , z I 3 ) , T ik = 2 1 ρ ^ k ρ ^ i .
The matrix T ∈ M 4 ( H ) T\in M_4(\mathbb{H}) T ∈ M 4 ( H ) is unitary. The raw coordinates are isometric, and the unitary change commutes with the left K K K -action. Hence the resulting map J \mathcal{J} J is an equivariant isometric isomorphism. On the first four slots this change intertwines R p R_p R p with η ↦ η p \eta\mapsto\eta p η ↦ η p .
On the last three slots, the action ω ↦ ω p ˉ \omega\mapsto\omega\bar{p} ω ↦ ω p ˉ becomes η ↦ η z ˉ p ˉ z \eta\mapsto\eta\bar{z}\bar{p}z η ↦ η z ˉ p ˉ z , which is η p \eta p η p for p = u , v p=u,v p = u , v , and the action ω ↦ ω z \omega\mapsto\omega z ω ↦ ω z of R v R u R_vR_u R v R u becomes η ↦ η z \eta\mapsto\eta z η ↦ η z . So the scalar part a 0 + a 1 R u + a 2 R v + a 3 R v R u a_0+a_1R_u+a_2R_v+a_3R_vR_u a 0 + a 1 R u + a 2 R v + a 3 R v R u becomes η ↦ η q ^ \eta\mapsto\eta\hat{q} η ↦ η q ^ on all seven slots.
Transformed matrix. Conjugating the blocks of M ′ M' M ′ by T \mathcal{T} T , with the identities ∑ k = 0 3 ρ ^ k γ ρ ^ k = − 2 γ ˉ \sum_{k=0}^{3}\hat{\rho}_k\gamma\hat{\rho}_k=-2\bar{\gamma} ∑ k = 0 3 ρ ^ k γ ρ ^ k = − 2 γ ˉ and ∑ k = 0 3 ρ ^ k γ ρ ^ k = 4 Re γ \sum_{k=0}^{3}\hat{\rho}_k\gamma\hat{\rho}_k=4\operatorname{Re}\gamma ∑ k = 0 3 ρ ^ k γ ρ ^ k = 4 Re γ for γ ∈ H \gamma\in\mathbb{H} γ ∈ H (Appendix § 7 , equation (A.4)), gives the upper-left block with entries 2 Re ( ρ ^ i ρ ^ k ) 2\operatorname{Re}(\hat{\rho}_i\hat{\rho}_k) 2 Re ( ρ ^ i ρ ^ k ) , which is diag ( 2 , − 2 , − 2 , − 2 ) \operatorname{diag}(2,-2,-2,-2) diag ( 2 , − 2 , − 2 , − 2 ) , the coupling ρ ^ j ρ ^ k z \hat{\rho}_j\hat{\rho}_kz ρ ^ j ρ ^ k z , and the lower-right block z ˉ ρ ^ j ρ ^ i z \bar{z}\hat{\rho}_j\hat{\rho}_i z z ˉ ρ ^ j ρ ^ i z entrywise. This is (4.6). The matrix M h M_h M h is Hermitian because h h h is symmetric and the coordinates are isometric.
Independence of choices. Finally, undoing the coordinates with f k = η k ℓ f_k=\eta_k\ell f k = η k ℓ gives (4.5), in which neither ℓ \ell ℓ nor a basis of B ^ \hat{B} B ^ appears. If r i ′ = ∑ j O i j r j r'_i=\sum_j O_{ij}r_j r i ′ = ∑ j O ij r j is another orthonormal basis of H \mathbb{H} H , with O O O real orthogonal, substitution into the defining sum for Ψ γ \Psi_\gamma Ψ γ contracts the two occurrences of O O O by ∑ i O i j O i k = δ j k \sum_i O_{ij}O_{ik}=\delta_{jk} ∑ i O ij O ik = δ j k . Thus that sum is basis independent.
The chosen coordinates make the scalar part equal to q ^ I 7 \hat{q}I_7 q ^ I 7 on all seven slots, while the off-diagonal blocks M + − M_{+-} M +− and M − + M_{-+} M −+ of M h M_h M h couple W + W_+ W + and W − W_- W − . The row-vector convention is essential: quaternionic transposition does not generally preserve the reduced norm, and in particular Nrd ( ( M h − q ^ I 7 ) T ) \operatorname{Nrd}((M_h-\hat{q}I_7)^{\mathsf{T}}) Nrd (( M h − q ^ I 7 ) T ) can differ from Nrd ( M h − q ^ I 7 ) \operatorname{Nrd}(M_h-\hat{q}I_7) Nrd ( M h − q ^ I 7 ) . See also [3 ] . The sources cited above let matrices act on column vectors, whereas M \mathcal{M} M acts here on row vectors from the right. Conjugating the coordinates, s ↦ s ˉ s\mapsto\bar{s} s ↦ s ˉ , turns the row action of M \mathcal{M} M into the column action of the conjugate transpose M ∗ \mathcal{M}^* M ∗ , and Sdet M ∗ = Sdet M \operatorname{Sdet}\mathcal{M}^*=\operatorname{Sdet}\mathcal{M} Sdet M ∗ = Sdet M [1 pp. 61–62] , so the identities above apply to M \mathcal{M} M as written.
We now evaluate the reduced norm of the displayed matrix M h − q ^ I 7 M_h-\hat{q}I_7 M h − q ^ I 7 . Let p 4 = Nrd ( M h − q ^ I 7 ) p_4=\operatorname{Nrd}(M_h-\hat{q}I_7) p 4 = Nrd ( M h − q ^ I 7 ) , so that det ( P a ∣ W 4 ) = p 4 2 \det(P_a|_{W_4})=p_4^2 det ( P a ∣ W 4 ) = p 4 2 by Proposition 4.6 . Recall that t = x − y t=x-y t = x − y .
Theorem 4.7. As an identity in Q [ a 0 , a 1 , a 2 , a 3 ] \mathbb{Q}[a_0,a_1,a_2,a_3] Q [ a 0 , a 1 , a 2 , a 3 ] , after substituting N = a 0 2 + a 1 2 + a 2 2 + a 3 2 N=a_0^2+a_1^2+a_2^2+a_3^2 N = a 0 2 + a 1 2 + a 2 2 + a 3 2
and t = a 1 2 + a 2 2 − a 3 2 t=a_{1}^{2}+a_{2}^{2}-a_{3}^{2} t = a 1 2 + a 2 2 − a 3 2 , we have p 4 = N Q 12 p_{4}=NQ_{12} p 4 = N Q 12 , where the polynomial Q 12 Q_{12} Q 12 is given by
Q 12 = F 0 ( a 0 , N ) + F 1 ( a 0 , N ) t + 64 N t 2 , where F 1 = 16 [ ( 3 a 0 + 3 ) N 3 + ( 28 a 0 2 + 40 a 0 + 8 ) N 2 + ( 64 a 0 3 + 104 a 0 2 + 64 a 0 ) N − 64 a 0 3 ] , F 0 = N 6 + ( 14 a 0 + 17 ) N 5 + ( 32 a 0 2 + 144 a 0 + 64 ) N 4 + ( − 304 a 0 3 + 160 a 0 2 + 512 a 0 + 64 ) N 3 + ( − 1600 a 0 4 − 256 a 0 3 + 3200 a 0 2 + 1024 a 0 ) N 2 + ( − 2048 a 0 5 + 3136 a 0 4 + 12288 a 0 3 + 4096 a 0 2 ) N + 7168 a 0 5 + 12288 a 0 4 . (31) \begin{aligned}
Q_{12}&=F_{0}(a_{0},N)+F_{1}(a_{0},N)t+64Nt^{2}, &&\text{where}\\
F_{1}&=16\left[(3a_{0}+3)N^{3}+(28a_{0}^{2}+40a_{0}+8)N^{2}\right.\\
&\qquad\left.+(64a_{0}^{3}+104a_{0}^{2}+64a_{0})N-64a_{0}^{3}\right],\\
F_{0}&=N^{6}+(14a_{0}+17)N^{5}+(32a_{0}^{2}+144a_{0}+64)N^{4}\\
&\qquad+(-304a_{0}^{3}+160a_{0}^{2}+512a_{0}+64)N^{3}\\
&\qquad+(-1600a_{0}^{4}-256a_{0}^{3}+3200a_{0}^{2}+1024a_{0})N^{2}\\
&\qquad+(-2048a_{0}^{5}+3136a_{0}^{4}+12288a_{0}^{3}+4096a_{0}^{2})N\\
&\qquad+7168a_{0}^{5}+12288a_{0}^{4}. \tag*{(31)}
\end{aligned} Q 12 F 1 F 0 = F 0 ( a 0 , N ) + F 1 ( a 0 , N ) t + 64 N t 2 , = 16 [ ( 3 a 0 + 3 ) N 3 + ( 28 a 0 2 + 40 a 0 + 8 ) N 2 + ( 64 a 0 3 + 104 a 0 2 + 64 a 0 ) N − 64 a 0 3 ] , = N 6 + ( 14 a 0 + 17 ) N 5 + ( 32 a 0 2 + 144 a 0 + 64 ) N 4 + ( − 304 a 0 3 + 160 a 0 2 + 512 a 0 + 64 ) N 3 + ( − 1600 a 0 4 − 256 a 0 3 + 3200 a 0 2 + 1024 a 0 ) N 2 + ( − 2048 a 0 5 + 3136 a 0 4 + 12288 a 0 3 + 4096 a 0 2 ) N + 7168 a 0 5 + 12288 a 0 4 . where (31)
In particular, p 4 p_{4} p 4 and Q 12 Q_{12} Q 12 are polynomials in ( a 0 , N , t ) (a_{0},N,t) ( a 0 , N , t ) , hence in ( a 0 , x , y ) (a_{0},x,y) ( a 0 , x , y ) , with rational coefficients, and Q 12 Q_{12} Q 12 has degree 2 in t t t .
Proof. Apply the embedding (30) to the quaternionic matrix M h − q ^ I 7 M_{h}-\hat qI_{7} M h − q ^ I 7 specified in (29) . Exact symbolic expansion of this 14 × 14 14\times14 14 × 14 complex determinant gives
det C χ ( M h − q ^ I 7 ) = N ( F 0 ( a 0 , N ) + F 1 ( a 0 , N ) t + 64 N t 2 ) . \det_{\mathbb{C}}\chi(M_{h}-\hat qI_{7})=N\left(F_{0}(a_{0},N)+F_{1}(a_{0},N)t+64Nt^{2}\right). C det χ ( M h − q ^ I 7 ) = N ( F 0 ( a 0 , N ) + F 1 ( a 0 , N ) t + 64 N t 2 ) .
By definition its left-hand side is p 4 p_{4} p 4 , proving the identity.
Since N Q 12 = p 4 = det C χ ( M h − q ^ I 7 ) ≥ 0 NQ_{12}=p_{4}=\det_{\mathbb{C}}\chi(M_{h}-\hat qI_{7})\geq0 N Q 12 = p 4 = det C χ ( M h − q ^ I 7 ) ≥ 0 , we find immediately that for N ≠ 0 N\ne0 N = 0 , Q 12 ≥ 0 Q_{12}\geq0 Q 12 ≥ 0 . The following sum of squares proves strict positivity of Q 12 Q_{12} Q 12 when y > 0 y>0 y > 0 , and hence invertibility of the block whenever a 3 ≠ 0 a_{3}\ne0 a 3 = 0 , since y = a 3 2 y=a_{3}^{2} y = a 3 2 .
Theorem 4.8 (Positivity and invertibility of the quaternionic sector). Set δ = N + 4 a 0 \delta=N+4a_{0} δ = N + 4 a 0 and
G 0 = δ [ N 2 + ( 3 a 0 + 4 ) N − 16 a 0 ( a 0 + 2 ) ] − 8 y ( N + 6 a 0 ) , G 1 = δ ( 3 N + 16 a 0 + 16 ) − 16 y , G 2 = δ ( 5 N + 8 a 0 − 16 ) + 16 y . \begin{aligned}
G_{0}&=\delta\left[N^{2}+(3a_{0}+4)N-16a_{0}(a_{0}+2)\right]-8y(N+6a_{0}),\\
G_{1}&=\delta(3N+16a_{0}+16)-16y,\\
G_{2}&=\delta(5N+8a_{0}-16)+16y.
\end{aligned} G 0 G 1 G 2 = δ [ N 2 + ( 3 a 0 + 4 ) N − 16 a 0 ( a 0 + 2 ) ] − 8 y ( N + 6 a 0 ) , = δ ( 3 N + 16 a 0 + 16 ) − 16 y , = δ ( 5 N + 8 a 0 − 16 ) + 16 y . Then, as a polynomial identity in Q [ a 0 , x , y ] \mathbb{Q}[a_{0},x,y] Q [ a 0 , x , y ] ,
Q 12 = G 0 2 + x G 1 2 + y G 2 2 + 320 x y δ 2 . (32) Q_{12}=G_{0}^{2}+xG_{1}^{2}+yG_{2}^{2}+320xy\delta^{2}. \tag*{(32)} Q 12 = G 0 2 + x G 1 2 + y G 2 2 + 320 x y δ 2 . (32) Consequently, Q 12 > 0 Q_{12}>0 Q 12 > 0 for real a 0 a_{0} a 0 , x ≥ 0 x\geq0 x ≥ 0 and y > 0 y>0 y > 0 . In particular, for real a a a with a 3 ≠ 0 a_{3}\ne0 a 3 = 0 , P a ∣ W 4 P_{a}|_{W_{4}} P a ∣ W 4 is invertible.
Proof. Substituting N = a 0 2 + x + y N=a_{0}^{2}+x+y N = a 0 2 + x + y , t = x − y t=x-y t = x − y into (31) , exact polynomial expansion gives (32) . Now suppose y > 0 y>0 y > 0 . If x > 0 x>0 x > 0 and δ ≠ 0 \delta\ne0 δ = 0 , the last term is positive. If x > 0 x>0 x > 0 and δ = 0 \delta=0 δ = 0 , then G 1 = − 16 y G_{1}=-16y G 1 = − 16 y , so x G 1 2 > 0 xG_{1}^{2}>0 x G 1 2 > 0 . It remains to consider x = 0 x=0 x = 0 , where N = a 0 2 + y N=a_{0}^{2}+y N = a 0 2 + y and Q 12 = G 0 2 + y G 2 2 Q_{12}=G_{0}^{2}+yG_{2}^{2} Q 12 = G 0 2 + y G 2 2 . A zero would therefore force G 0 = G 2 = 0 G_{0}=G_{2}=0 G 0 = G 2 = 0 . Set L 0 = N δ − 16 a 0 L_{0}=N\delta-16a_{0} L 0 = N δ − 16 a 0 . In this case the polynomial identities
G 0 + ( a 0 + 4 ) G 2 = ( N + 8 a 0 + 16 ) L 0 = ( ( a 0 + 4 ) 2 + y ) L 0 , G_{0}+(a_{0}+4)G_{2}=(N+8a_{0}+16)L_{0}=\left((a_{0}+4)^{2}+y\right)L_{0}, G 0 + ( a 0 + 4 ) G 2 = ( N + 8 a 0 + 16 ) L 0 = ( ( a 0 + 4 ) 2 + y ) L 0 ,
G 2 = 5 L 0 + 8 a 0 ( N + 2 a 0 + 2 ) = 5 L 0 + 8 a 0 ( ( a 0 + 1 ) 2 + y + 1 ) G_{2}=5L_{0}+8a_{0}(N+2a_{0}+2)=5L_{0}+8a_{0}\left((a_{0}+1)^{2}+y+1\right) G 2 = 5 L 0 + 8 a 0 ( N + 2 a 0 + 2 ) = 5 L 0 + 8 a 0 ( ( a 0 + 1 ) 2 + y + 1 )
follow by expansion. Thus the first identity forces L 0 = 0 L_0 = 0 L 0 = 0 and the second then forces a 0 = 0 a_0 = 0 a 0 = 0 . But a 0 = x = 0 a_0 = x = 0 a 0 = x = 0 gives N = δ = y N = \delta= y N = δ = y , so L 0 = y 2 > 0 L_0 = y^2 > 0 L 0 = y 2 > 0 , a contradiction. This proves strict positivity in all cases.
For real a a a with a 3 ≠ 0 a_3 \ne0 a 3 = 0 , we have x = a 1 2 + a 2 2 ≥ 0 x = a_1^2 + a_2^2 \ge0 x = a 1 2 + a 2 2 ≥ 0 , y = a 3 2 > 0 y = a_3^2 > 0 y = a 3 2 > 0 and N = a 0 2 + x + y ≥ y > 0 N = a_0^2 + x + y \ge y > 0 N = a 0 2 + x + y ≥ y > 0 . The strict positivity just proved gives Q 12 > 0 Q_{12} > 0 Q 12 > 0 . Thus p 4 = N Q 12 > 0 p_4 = NQ_{12} > 0 p 4 = N Q 12 > 0 , and det ( P a ∣ W 4 ) = ( N Q 12 ) 2 > 0 \det(P_a|_{W_4}) = (NQ_{12})^2 > 0 det ( P a ∣ W 4 ) = ( N Q 12 ) 2 > 0 by Theorem 4.7.
The invertibility proof uses a 3 ≠ 0 a_3 \ne0 a 3 = 0 . On the hyperplane a 3 = 0 a_3 = 0 a 3 = 0 we instead use the additional symmetry in the following subsection.
The one-step hyperplane a 3 = 0 a_3 = 0 a 3 = 0
On the hyperplane a 3 = 0 a_3 = 0 a 3 = 0 , the associator term vanishes, and every operator P a P_a P a is a one-step operator, with enhanced symmetry.
Let Q = x = a 1 2 + a 2 2 Q = \sqrt{x} = \sqrt{a_1^2 + a_2^2} Q = x = a 1 2 + a 2 2 . If Q ≠ 0 Q \ne0 Q = 0 , set
n ^ = a 1 u + a 2 v Q , \hat{n} = \frac{a_1u + a_2v}{Q}, n ^ = Q a 1 u + a 2 v ,
and if Q = 0 Q = 0 Q = 0 , take n ^ = u \hat{n} = u n ^ = u . Let n ∈ A n \in A n ∈ A be the corresponding unit vector in V V V . By linearity of right multiplication, for a = ( a 0 , a 1 , a 2 , 0 ) a = (a_0,a_1,a_2,0) a = ( a 0 , a 1 , a 2 , 0 ) we have
P a = h − Q R n ^ − a 0 I = H n , Q − a 0 I , P_a = h - QR_{\hat{n}} - a_0I = H_{n,Q} - a_0I, P a = h − Q R n ^ − a 0 I = H n , Q − a 0 I ,
the level-1 slice pencil of § 3 . In particular, P a P_a P a has the symmetry of the stabilizer S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n , which contains K = S U ( 2 ) = Stab G 2 ( u , v ) K = \mathrm{SU}(2) = \operatorname{Stab}_{G_2}(u,v) K = SU ( 2 ) = Stab G 2 ( u , v ) .
Lemma 4.9 (One-step kernel lemma). Let a = ( a 0 , a 1 , a 2 , 0 ) ∈ R 4 a = (a_0,a_1,a_2,0) \in\mathbb{R}^4 a = ( a 0 , a 1 , a 2 , 0 ) ∈ R 4 . Then P a P_a P a commutes with S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n , and every nonzero S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -invariant subspace U ⊂ W U \subset W U ⊂ W contains a nonzero K K K -fixed vector. Consequently,
ker P a ≠ 0 ⟹ ker P a ∩ W 1 ≠ 0. \ker P_a \ne0 \quad\Longrightarrow\quad\ker P_a \cap W_1 \ne0. ker P a = 0 ⟹ ker P a ∩ W 1 = 0.
Proof. By Proposition 2.4, h h h is G 2 G_2 G 2 -equivariant and R n ^ R_{\hat{n}} R n ^ commutes with S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n . Hence so does P a = H n , Q − a 0 I P_a = H_{n,Q} - a_0I P a = H n , Q − a 0 I . Moreover, K ⊂ S U ( 3 ) n K \subset\mathrm{SU}(3)_n K ⊂ SU ( 3 ) n , since K K K fixes H \mathbb{H} H , and therefore n n n , pointwise.
Under S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n , the real representation W W W decomposes as
W ≅ 1 ⊕ 4 ⊕ 6 ⊕ 4 ⊕ 8 ⊕ 2 ⊕ 12 W \cong1^{\oplus4} \oplus6^{\oplus4} \oplus8^{\oplus2} \oplus12 W ≅ 1 ⊕ 4 ⊕ 6 ⊕ 4 ⊕ 8 ⊕ 2 ⊕ 12
(§ 3 ). These four representation types are irreducible over R \mathbb{R} R (§ 3 ). Under the standard inclusion K ⊂ S U ( 3 ) n K \subset\mathrm{SU}(3)_n K ⊂ SU ( 3 ) n , each contains a nonzero K K K -fixed vector (Lemma A.2 of Appendix § 7 ).
Now let U ⊂ W U \subset W U ⊂ W be a nonzero S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -invariant subspace. By complete reducibility for the compact group S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n [7 Theorem 4.28] , U U U contains a nonzero irreducible S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -submodule. This submodule must be of one of the four types above, and hence contains a nonzero K K K -fixed vector. Therefore
U ∩ W 1 = U K ≠ 0. U \cap W_1 = U^K \ne0. U ∩ W 1 = U K = 0.
Finally, since P a P_a P a commutes with S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n , its kernel is S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -invariant. Applying the preceding conclusion to U = ker P a U = \ker P_a U = ker P a proves the lemma.
The full determinant and real singular locus
Multiplying the three sector determinants gives the full determinant below. Corollary 4.13 expresses it in terms of invariants of a pair.
Theorem 4.10 (The two-step determinant). For the pencil of an orthonormal pair u , v ∈ Im O u,v \in\operatorname{Im}\mathbb{O} u , v ∈ Im O ,
det 56 P a = N 6 Q 14 ( a 0 , x , y ) ( N + 4 a 0 ) 3 Q 12 ( a 0 , x , y ) 2 , (33) \det_{56} P_a = N^6 Q_{14}(a_0,x,y)(N+4a_0)^3 Q_{12}(a_0,x,y)^2, \tag*{(33)} 56 det P a = N 6 Q 14 ( a 0 , x , y ) ( N + 4 a 0 ) 3 Q 12 ( a 0 , x , y ) 2 , (33) where Q 14 Q_{14} Q 14 and Q 12 Q_{12} Q 12 are the polynomials of Theorems 4.1 and 4.7 .
We now determine the real singular locus of P a P_a P a . Recall the zero set Z Z Z of Q 14 Q_{14} Q 14 from (26) and the sphere S S S from Theorem 4.5 .
Theorem 4.11 (The real locus). Let Σ ⊂ R 4 \Sigma\subset\mathbb{R}^4 Σ ⊂ R 4 be the singular locus of P a P_a P a , and for i ∈ { 1 , 3 , 4 } i \in\{1,3,4\} i ∈ { 1 , 3 , 4 } , let Σ i \Sigma_i Σ i be the singular locus of its restriction to W i W_i W i . Then:
(i) The sector singular loci satisfy
Σ 1 = Z , Σ 3 = S and Σ 4 ⊂ Z ∩ { a 3 = 0 } . \Sigma_1 = Z,\qquad\Sigma_3 = S \qquad\text{and} \qquad\Sigma_4 \subset Z \cap\{a_3=0\}. Σ 1 = Z , Σ 3 = S and Σ 4 ⊂ Z ∩ { a 3 = 0 } . (ii) The full real singular locus is
Σ = Z ∪ S . \Sigma= Z \cup S. Σ = Z ∪ S .
Proof. The invariant sector decomposition gives
ker P a = ⨁ i ker ( P a ∣ W i ) , Σ = ⋃ i Σ i . \ker P_a = \bigoplus_i \ker(P_a|_{W_i}), \qquad\Sigma= \bigcup_i \Sigma_i. ker P a = i ⨁ ker ( P a ∣ W i ) , Σ = i ⋃ Σ i .
By Theorem 4.1 , the determinant on W 1 W_1 W 1 is N Q 14 NQ_{14} N Q 14 . Over R \mathbb{R} R , N N N vanishes only at the origin, which also belongs to Z Z Z since M p o l = 0 M_{\mathrm{pol}}=0 M pol = 0 there (Lemma B.3 ). Thus Σ 1 = Z \Sigma_1=Z Σ 1 = Z . Theorem 4.5 gives
For the remaining sector, Theorem 4.8 shows that
Σ 4 ⊂ { a 3 = 0 } . \Sigma_4\subset\{a_3=0\}. Σ 4 ⊂ { a 3 = 0 } .
On this hyperplane, Lemma 4.9 implies that P a P_a P a is singular if and only if its restriction to W 1 W_1 W 1 is singular. Hence
Σ ∩ { a 3 = 0 } = Z ∩ { a 3 = 0 } . \Sigma\cap\{a_3=0\}=Z\cap\{a_3=0\}. Σ ∩ { a 3 = 0 } = Z ∩ { a 3 = 0 } .
In particular, Σ 4 ⊂ Z \Sigma_4\subset Z Σ 4 ⊂ Z , completing (i). Finally, the sector decomposition gives
Σ = Σ 1 ∪ Σ 3 ∪ Σ 4 = Z ∪ S , \Sigma=\Sigma_1\cup\Sigma_3\cup\Sigma_4=Z\cup S, Σ = Σ 1 ∪ Σ 3 ∪ Σ 4 = Z ∪ S ,
proving (ii).
Remark 4.12. Define N 0 = a 0 2 + x N_0=a_0^2+x N 0 = a 0 2 + x and C 0 = N 0 + 4 a 0 C_0=N_0+4a_0 C 0 = N 0 + 4 a 0 , and write p ~ 1 ( a 0 , x ) = p 1 ( a 0 , x ) \widetilde{p}_1(a_0,x)=p_1(a_0,\sqrt{x}) p 1 ( a 0 , x ) = p 1 ( a 0 , x ) . This is a polynomial because the quartic (5) is even in Q Q Q . For y = 0 y=0 y = 0 , we find that
Q 14 ( a 0 , x , 0 ) = C 0 p ~ 1 E 8 and Q 12 ( a 0 , x , 0 ) = C 0 2 E 8 , Q_{14}(a_0,x,0)=C_0\widetilde{p}_1E_8 \qquad\text{and} \qquad Q_{12}(a_0,x,0)=C_0^2E_8, Q 14 ( a 0 , x , 0 ) = C 0 p 1 E 8 and Q 12 ( a 0 , x , 0 ) = C 0 2 E 8 ,
Here
E 8 ( a 0 , x ) = ∣ p 6 ( a 0 , x ) ∣ 2 , E_8(a_0,x)=\lvert p_6(a_0,\sqrt{x})\rvert^2, E 8 ( a 0 , x ) = ∣ p 6 ( a 0 , x ) ∣ 2 ,
where p 6 p_6 p 6 is the complex quartic determinant of the one-step 6 ⊕ 4 6\oplus4 6 ⊕ 4 multiplicity block [6 ] . The singlet quartic p 1 p_1 p 1 of (5) is the separate factor p ~ 1 \widetilde{p}_1 p 1 above. By Proposition 2.7 , E 8 > 0 E_8>0 E 8 > 0 for x > 0 x>0 x > 0 , so it adds no nonreal one-step spectral curve.
At every point of Z ∩ S Z \cap S Z ∩ S , the kernels in W 1 W_1 W 1 and W 3 W_3 W 3 are both nontrivial and contribute independently to ker P a \ker P_a ker P a .
Define the invariant-image map
π : R 4 ⟶ R × [ 0 , ∞ ) 2 , by π ( a ) = ( a 0 , a 1 2 + a 2 2 , a 3 2 ) . \pi:\mathbb{R}^4 \longrightarrow\mathbb{R} \times[0,\infty)^2,\qquad\text{by}\qquad\pi(a)=(a_0,a_1^2+a_2^2,a_3^2). π : R 4 ⟶ R × [ 0 , ∞ ) 2 , by π ( a ) = ( a 0 , a 1 2 + a 2 2 , a 3 2 ) .
Figure 2 plots π ( Z ) \pi(Z) π ( Z ) and π ( S ) \pi(S) π ( S ) in invariant-image coordinates. The sphere S S S is round in R 4 \mathbb{R}^4 R 4 , but its invariant image is
Figure 2. Invariant images of the real locus in ( a 0 , x , y ) (a_0,x,y) ( a 0 , x , y ) , with x , y ≥ 0 x,y \ge0 x , y ≥ 0 : blue-gray is the image of Z Z Z and pale yellow the image of S S S . On y = 0 y=0 y = 0 , the marks of Figure 1 are mapped by ( μ , Q ) ↦ ( a 0 , x ) = ( μ , Q 2 ) (\mu,Q) \mapsto(a_0,x)=(\mu,Q^2) ( μ , Q ) ↦ ( a 0 , x ) = ( μ , Q 2 ) : blue is the quartic image, dashed is the one-step circle image, and dots are the five real eigenvalues. The dashed curve is the trace of the image of S S S on the boundary plane y = 0 y=0 y = 0 .
π ( S ) = { ( a 0 , x , y ) : x , y ≥ 0 , x + y = 4 − ( a 0 + 2 ) 2 } . \pi(S)=\{(a_0,x,y):x,y\ge0,\ x+y=4-(a_0+2)^2\}. π ( S ) = {( a 0 , x , y ) : x , y ≥ 0 , x + y = 4 − ( a 0 + 2 ) 2 } .
Thus its image is a portion of a parabolic cylinder.
The boundary y = 0 y=0 y = 0 is the invariant image of the coefficient hyperplane a 3 = 0 a_3=0 a 3 = 0 . On this hyperplane, P a = H n , Q − a 0 P_a=H_{n,Q}-a_0 P a = H n , Q − a 0 (§ 5.4 ). Thus the trace on y = 0 y=0 y = 0 is the image of the level-1 slice locus under ( μ , Q ) ↦ ( a 0 , x ) = ( μ , Q 2 ) (\mu,Q)\mapsto(a_0,x)=(\mu,Q^2) ( μ , Q ) ↦ ( a 0 , x ) = ( μ , Q 2 ) . Its real-axis trace is the real spectrum of h h h given in Proposition 2.6.
The following corollary expresses the determinant and real-locus results in invariants of an arbitrary octonion pair. Recall that a 0 = Re ( λ 1 λ 2 ) a_0=\operatorname{Re}(\lambda_1\lambda_2) a 0 = Re ( λ 1 λ 2 ) , N = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 N=|\lambda_1|^2|\lambda_2|^2 N = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 , that y = a 3 2 y=a_3^2 y = a 3 2 is the Gram determinant of Im λ 1 \operatorname{Im}\lambda_1 Im λ 1 and Im λ 2 \operatorname{Im}\lambda_2 Im λ 2 , and that x = N − a 0 2 − y x=N-a_0^2-y x = N − a 0 2 − y (12) .
Corollary 4.13. Let ( λ 1 , λ 2 ) (\lambda_1,\lambda_2) ( λ 1 , λ 2 ) be any pair of octonions and let a a a be its reduced point.
The determinant det 56 ( h − R λ 2 R λ 1 ) \det_{56}(h-R_{\lambda_2}R_{\lambda_1}) det 56 ( h − R λ 2 R λ 1 ) is given by (33) at ( a 0 , x , y ) (a_0,x,y) ( a 0 , x , y ) , with N = a 0 2 + x + y = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 N=a_0^2+x+y=|\lambda_1|^2|\lambda_2|^2 N = a 0 2 + x + y = ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 .
The equation h ( w ) = ( w λ 1 ) λ 2 h(w)=(w\lambda_1)\lambda_2 h ( w ) = ( w λ 1 ) λ 2 has a nonzero solution w ∈ W w\in W w ∈ W if and only if a ∈ Z ∪ S a\in Z\cup S a ∈ Z ∪ S , that is, if and only if Q 14 ( a 0 , x , y ) = 0 Q_{14}(a_0,x,y)=0 Q 14 ( a 0 , x , y ) = 0 or
∣ λ 2 λ 1 + 2 ∣ = 2 ⟺ ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 + 4 Re ( λ 1 λ 2 ) = 0. (34) |\lambda_2\lambda_1+2|=2\quad\Longleftrightarrow\quad|\lambda_1|^2|\lambda_2|^2+4\operatorname{Re}(\lambda_1\lambda_2)=0. \tag*{(34)} ∣ λ 2 λ 1 + 2∣ = 2 ⟺ ∣ λ 1 ∣ 2 ∣ λ 2 ∣ 2 + 4 Re ( λ 1 λ 2 ) = 0. (34)
If Im λ 1 \operatorname{Im}\lambda_1 Im λ 1 and Im λ 2 \operatorname{Im}\lambda_2 Im λ 2 are linearly independent, that is, a 3 ≠ 0 a_3\ne0 a 3 = 0 , the W 4 W_4 W 4 block of h − R λ 2 R λ 1 h-R_{\lambda_2}R_{\lambda_1} h − R λ 2 R λ 1 is invertible.
The following corollary measures how far a two-step kernel vector is from satisfying any one-step equation. It also identifies exactly which kernel vectors have no octonionic right eigenvalue.
Corollary 4.14. Let a ∈ R 4 a\in\mathbb{R}^4 a ∈ R 4 and 0 ≠ w ∈ ker P a 0\ne w\in\ker P_a 0 = w ∈ ker P a , with sector decomposition w = w 1 + w 3 + w 4 w=w_1+w_3+w_4 w = w 1 + w 3 + w 4 . Set α = ∥ Π + w ∥ 2 \alpha=\|\Pi_+w\|^2 α = ∥ Π + w ∥ 2 and β = ∥ Π − w ∥ 2 \beta=\|\Pi_-w\|^2 β = ∥ Π − w ∥ 2 . Then
min λ ∈ O ∥ h ( w ) − w λ ∥ 2 = 4 a 3 2 α β α + β . (35) \min_{\lambda\in\mathbb{O}}\|h(w)-w\lambda\|^2=\frac{4a_3^2\alpha\beta}{\alpha+\beta}. \tag*{(35)} λ ∈ O min ∥ h ( w ) − w λ ∥ 2 = α + β 4 a 3 2 α β . (35) The unique minimizing octonion is
λ ∗ = α q ^ + β q ˇ α + β . \lambda_*=\frac{\alpha\hat q+\beta\check q}{\alpha+\beta}. λ ∗ = α + β α q ^ + β q ˇ . Consequently,
h ( w ) ≠ w λ for every λ ∈ O ⟺ a 3 ≠ 0 and w 1 ≠ 0. h(w)\ne w\lambda\text{ for every }\lambda\in\mathbb{O}\quad\Longleftrightarrow\quad a_3\ne0\text{ and }w_1\ne0. h ( w ) = w λ for every λ ∈ O ⟺ a 3 = 0 and w 1 = 0. This is equivalent to satisfying neither of the equations h ( w ) = w q ^ h(w)=w\hat q h ( w ) = w q ^ and h ( w ) = w q ˇ h(w)=w\check q h ( w ) = w q ˇ . Such a kernel vector exists precisely when a ∈ Z a\in Z a ∈ Z and a 3 ≠ 0 a_3\ne0 a 3 = 0 .
Proof. If a 3 = 0 a_3=0 a 3 = 0 , then P a = h − R q ^ P_a=h-R_{\hat q} P a = h − R q ^ and q ^ = q ˇ \hat q=\check q q ^ = q ˇ . Every kernel vector satisfies the common one-step equation, and both sides of (35) vanish.
Suppose a 3 ≠ 0 a_3\ne0 a 3 = 0 . Theorem 4.8 gives w 4 = 0 w_4=0 w 4 = 0 . By Theorem 3.6,
W 1 = ( H ⊗ A ) ⊕ ( B ^ ⊗ B ) K , W 3 ⊂ B ^ ⊗ B . W_1=(\mathbb{H}\otimes A)\oplus(\hat B\otimes B)^K,\qquad W_3\subset\hat B\otimes B. W 1 = ( H ⊗ A ) ⊕ ( B ^ ⊗ B ) K , W 3 ⊂ B ^ ⊗ B .
Thus Π + w \Pi_+w Π + w has its vector index in A A A , and Π − w \Pi_-w Π − w has its vector index in B B B . Proposition 3.5 gives, for every λ ∈ O \lambda\in\mathbb{O} λ ∈ O ,
h ( w ) − w λ = ( Π + w ) ( q ^ − λ ) + ( Π − w ) ( q ˇ − λ ) . h(w)-w\lambda=(\Pi_+w)(\hat q-\lambda)+(\Pi_-w)(\check q-\lambda). h ( w ) − w λ = ( Π + w ) ( q ^ − λ ) + ( Π − w ) ( q ˇ − λ ) .
The two terms are orthogonal because A ⊥ B A\perp B A ⊥ B in the second tensor factor; this does not require right multiplication by λ \lambda λ to preserve H \mathbb{H} H or B ^ \hat B B ^ . Multiplicativity of the octonion norm therefore yields
∥ h ( w ) − w λ ∥ 2 = α ∣ q ^ − λ ∣ 2 + β ∣ q ˇ − λ ∣ 2 . (36) \|h(w)-w\lambda\|^2=\alpha|\hat q-\lambda|^2+\beta|\check q-\lambda|^2. \tag*{(36)} ∥ h ( w ) − w λ ∥ 2 = α ∣ q ^ − λ ∣ 2 + β ∣ q ˇ − λ ∣ 2 . (36)
Since α + β = ∥ w ∥ 2 > 0 \alpha+\beta=\|w\|^2>0 α + β = ∥ w ∥ 2 > 0 , this quadratic is strictly convex. Completing the square gives the stated minimizer and minimum, using ∣ q ^ − q ˇ ∣ 2 = 4 a 3 2 |\hat q-\check q|^2=4a_3^2 ∣ q ^ − q ˇ ∣ 2 = 4 a 3 2 .
It remains to characterize α β > 0 \alpha\beta> 0 α β > 0 . If w 1 = 0 w_1 = 0 w 1 = 0 , then w = w 3 ∈ W − w = w_3 \in W_- w = w 3 ∈ W − and α = 0 \alpha= 0 α = 0 . If w 1 ≠ 0 w_1 \ne0 w 1 = 0 , write its fixed-sector coordinates as ( ϑ ; η ) ∈ H 3 ⊕ H (\vartheta;\eta) \in\mathbb{H}^3 \oplus\mathbb{H} ( ϑ ; η ) ∈ H 3 ⊕ H . Since N , N 3 > 0 N,N_3 > 0 N , N 3 > 0 , the block P A A P_{AA} P AA is invertible, so η = 0 \eta= 0 η = 0 forces ϑ = 0 \vartheta= 0 ϑ = 0 . Conversely, ϑ = 0 \vartheta= 0 ϑ = 0 forces h A 0 η = 0 h_{A0}\eta= 0 h A 0 η = 0 , and the injective block h A 0 h_{A0} h A 0 in (B.1) then gives η = 0 \eta= 0 η = 0 . Hence both projections of w 1 w_1 w 1 are nonzero. Moreover,
Π + w = Π + w 1 and Π − w = Π − w 1 + w 3 , \Pi_+w = \Pi_+w_1 \qquad\text{and}\qquad\Pi_-w = \Pi_-w_1 + w_3, Π + w = Π + w 1 and Π − w = Π − w 1 + w 3 ,
where the last two terms cannot cancel: the projections are K K K -equivariant and these terms belong to distinct isotypic sectors. Thus α β > 0 \alpha\beta> 0 α β > 0 exactly when w 1 ≠ 0 w_1 \ne0 w 1 = 0 . The minimum is attained, so its positivity proves the criterion. Evaluating (4.13) at q ^ \hat{q} q ^ and q ˇ \check{q} q ˇ gives 4 a 3 2 β 4a_3^2\beta 4 a 3 2 β and 4 a 3 2 α 4a_3^2\alpha 4 a 3 2 α , respectively, proving the neither-product equivalence. Finally, Theorem 4.11 gives a nonzero fixed-sector kernel exactly on Z Z Z , proving the existence assertion.
Corollary 4.14 concerns an individual kernel vector. The following example also excludes both products from the entire one-step right spectrum, although a nonzero two-step solution exists.
Example 4.15 (A genuinely two-step spectral point). Let b = 6 / 4 b = \sqrt{6}/4 b = 6 /4 and c = 22 / 4 c = \sqrt{22}/4 c = 22 /4 . The octonion pair
λ 1 = c u and λ 2 = 3 11 + 6 22 u + v \lambda_1 = cu \qquad\text{and}\qquad\lambda_2 = \sqrt{\frac{3}{11}}+\frac{6}{\sqrt{22}}u+v λ 1 = c u and λ 2 = 11 3 + 22 6 u + v has a one-dimensional two-step kernel, while both λ 1 λ 2 \lambda_1\lambda_2 λ 1 λ 2 and λ 2 λ 1 \lambda_2\lambda_1 λ 2 λ 1 lie outside the one-step right spectrum. Its imaginary parts are linearly independent and nonorthogonal, and its reduced point is a ∗ = ( − 3 / 2 , b , 0 , c ) \mathbf{a}_* = (-3/2,b,0,c) a ∗ = ( − 3/2 , b , 0 , c ) . Here N = N 3 = 4 N=N_3=4 N = N 3 = 4 , x = 3 / 8 x=3/8 x = 3/8 and y = 11 / 8 y=11/8 y = 11/8 .
Proof. The fixed-sector Schur complement, in the basis ( 1 , u , v , z ) (1,u,v,z) ( 1 , u , v , z ) , is
S ∗ = ( 2 0 0 2 c 0 − 2 0 0 0 0 − 2 − 2 b − 2 c 0 2 b − 2 ) . S_* =
\begin{pmatrix}
2 & 0 & 0 & 2c \\
0 & -2 & 0 & 0 \\
0 & 0 & -2 & -2b \\
-2c & 0 & 2b & -2
\end{pmatrix}. S ∗ = 2 0 0 − 2 c 0 − 2 0 0 0 0 − 2 2 b 2 c 0 − 2 b − 2 .
Since c 2 − b 2 = 1 c^2-b^2=1 c 2 − b 2 = 1 , its kernel contains ( − c , 0 , − b , 1 ) T (-c,0,-b,1)^{\mathsf T} ( − c , 0 , − b , 1 ) T . The upper-left 3 × 3 3\times3 3 × 3 minor is 8 8 8 , so
ker S ∗ = R ( − c , 0 , − b , 1 ) T . \ker S_* = \mathbb{R}(-c,0,-b,1)^{\mathsf T}. ker S ∗ = R ( − c , 0 , − b , 1 ) T .
Thus Q 14 ( a ∗ ) = 0 Q_{14}(\mathbf{a}_*)=0 Q 14 ( a ∗ ) = 0 , and the fixed-sector kernel is one-dimensional. Moreover, N + 4 a 0 = − 2 N+4a_0=-2 N + 4 a 0 = − 2 and a 3 = c ≠ 0 a_3=c\ne0 a 3 = c = 0 , so the 3 3 3 - and 4 4 4 -sectors are invertible. Consequently, dim R ker P a ∗ = 1 \dim_{\mathbb{R}}\ker P_{\mathbf{a}_*}=1 dim R ker P a ∗ = 1 . The associated quaternions are
q ^ = − 3 2 + b u + c z , and q ˇ = − 3 2 + b u − c z . \hat{q}=-\frac{3}{2}+bu+cz,\qquad\text{and}\qquad\check{q}=-\frac{3}{2}+bu-cz. q ^ = − 2 3 + b u + cz , and q ˇ = − 2 3 + b u − cz .
Both lie outside the level-1 spectrum. Indeed, their imaginary parts have norm 7 / 2 \sqrt{7}/2 7 /2 , and the one-step quartic and circle conditions give, respectively,
p 1 ( − 3 2 , 7 2 ) = 32 ≠ 0 and ( − 3 2 + 2 ) 2 + 7 4 = 2 ≠ 4. p_1\left(-\frac{3}{2},\frac{\sqrt{7}}{2}\right)=32\ne0
\qquad\text{and}\qquad
\left(-\frac{3}{2}+2\right)^2+\frac{7}{4}=2\ne4. p 1 ( − 2 3 , 2 7 ) = 32 = 0 and ( − 2 3 + 2 ) 2 + 4 7 = 2 = 4.
Thus both one-step operators h − R q ^ h-R_{\hat{q}} h − R q ^ and h − R q ˇ h-R_{\check{q}} h − R q ˇ are invertible, whereas P a ∗ P_{\mathbf{a}_*} P a ∗ has a nontrivial kernel.
Conclusion
The level-0 real spectrum of h h h is finite, while its level-1 right spectrum is described in each complex slice by a quartic and a circle (§ 3 ). At level 2, the determinant of the two-step pencil is given by Theorem 4.10 , and its real singular locus is exactly Z ∪ S Z \cup S Z ∪ S (Theorem 4.11 ). These results apply to every pair at its reduced point (Corollary 4.13 ).
Corollary 4.14 gives the exact minimum of ∥ h ( w ) − w λ ∥ 2 \lVert h(w)-w\lambda\rVert^2 ∥ h ( w ) − w λ ∥ 2 over all octonions λ \lambda λ . Away from a 3 = 0 a_3=0 a 3 = 0 , a kernel vector has no octonionic right eigenvalue exactly when its fixed-sector component is nonzero, and such vectors exist exactly on Z Z Z . Thus failure of both natural product equations already excludes every one-step right eigenvalue for that vector. Example 4.15 gives a one-dimensional kernel at a pair whose two products both lie outside the one-step spectrum.
The same questions can be asked of longer sequences of successive right multiplications, with the order of the factors retained. These sequences enlarge the family of comparison operators while reducing their generic symmetry with h h h . For unit factors, let
N k = { R q k ⋯ R q 1 : ∣ q i ∣ = 1 } ⊂ S O ( 8 ) , N 0 = { + I } . \mathcal{N}_k=\{R_{q_k}\cdots R_{q_1}: |q_i|=1\}\subset SO(8), \qquad\mathcal{N}_0=\{+I\}. N k = { R q k ⋯ R q 1 : ∣ q i ∣ = 1 } ⊂ S O ( 8 ) , N 0 = { + I } .
Conway and Smith give the dimensions n k = dim N k n_k=\dim\mathcal{N}_k n k = dim N k listed in Table 2 [4 Chapter 8, §8.5, Theorem 9, pp. 94–95] . Since h h h is G 2 G_2 G 2 -equivariant, the natural reduced ambient space for the level-k k k problem is
M k = R > 0 × ( N k / Ad ( G 2 ) ) . \mathcal{M}_k=\mathbb{R}_{>0}\times\left(\mathcal{N}_k/\operatorname{Ad}(G_2)\right). M k = R > 0 × ( N k / Ad ( G 2 ) ) .
where G 2 G_2 G 2 acts by conjugation on N k \mathcal{N}_k N k , leaving the positive scale factor separate. At level zero, the zero operator is included separately. If K k K_k K k denotes the generic stabilizer of a composite operator, the normalized quotient has generic dimension d k = n k − 14 + dim K k d_k=n_k-14+\dim K_k d k = n k − 14 + dim K k , and the overall positive scale gives dim M k = D k = d k + 1 \dim\mathcal{M}_k=D_k=d_k+1 dim M k = D k = d k + 1 on its principal stratum.
At level 1 a generic operator recovers an imaginary direction, with stabilizer S U ( 3 ) \mathrm{SU}(3) SU ( 3 ) ; at level two its vector and bivector parts recover a plane pointwise, with stabilizer S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) . For levels three and beyond, the key fact is that a triple of octonions that does not lie in a quaternionic subalgebra generates all of O \mathbb{O} O , so any octonion automorphism fixing it is the identity. It can then be easily shown (but is outside the scope of the current paper), that generic stabilizers of the corresponding composite operators will also be the identity for k ≥ 3 k\ge3 k ≥ 3 .
Table 2 (PDF p. 28). Dimensions of the families of successive right multiplications and their generic G 2 G_2 G 2 -reduced parameter spaces at levels 0–4. The stabilizers and reduced dimensions refer to generic configurations; special configurations may have larger stabilizers.
The spectral locus is a subset of this ambient space, defined by the vanishing of the G 2 G_2 G 2 -invariant determinant
Δ ( ρ A ) = det ( h − ρ A ⊗ I V ) , for ρ > 0 and A ∈ N k . \Delta(\rho A)=\det(h-\rho A\otimes I_V), \qquad\text{for }\rho>0\text{ and }A\in\mathcal{N}_k. Δ ( ρ A ) = det ( h − ρ A ⊗ I V ) , for ρ > 0 and A ∈ N k .
Thus the level-one quartic and circle, and the level-two loci defined by Q 14 = 0 Q_{14}=0 Q 14 = 0 and the sphere equation, describe spectral subsets of reduced ambient spaces of dimensions two and three, respectively. At three steps, generic residual symmetry disappears, but N 3 \mathcal{N}_3 N 3 still has codimension ten in SO ( 8 ) \operatorname{SO}(8) SO ( 8 ) . The higher-level problems therefore remain constrained spectral problems even without a nontrivial generic stabilizer. At seven steps, N 7 = SO ( 8 ) \mathcal{N}_7=\operatorname{SO}(8) N 7 = SO ( 8 ) [4 Chapter 8, §8.4, Theorem 7, p. 93] , so all levels can be viewed as restrictions of the same determinant to nested families in the full reduced space ( R > 0 × SO ( 8 ) ) / Ad ( G 2 ) (\mathbb{R}_{>0}\times\operatorname{SO}(8))/\operatorname{Ad}(G_2) ( R > 0 × SO ( 8 )) / Ad ( G 2 ) .
Frame, Cayley–Dickson rules and branching
We use the same notation as in Section 3.1. Multiplication on H \mathbb{H} H is given by
ρ ^ i ρ ^ j = − δ i j + ∑ k = 1 3 ϵ i j k ρ ^ k , for 1 ≤ i , j ≤ 3. (37) \hat{\rho}_i\hat{\rho}_j=-\delta_{ij}+\sum_{k=1}^{3}\epsilon_{ijk}\hat{\rho}_k,\qquad\text{for }1\le i,j\le3. \tag*{(37)} ρ ^ i ρ ^ j = − δ ij + k = 1 ∑ 3 ϵ ij k ρ ^ k , for 1 ≤ i , j ≤ 3. (37)
For every unit ℓ ∈ B ^ \ell\in\hat{B} ℓ ∈ B ^ , the Cayley–Dickson doubling [2 Section 2.2, equations (2) and (3)] (see also the right-coordinate formula in [4 Chapter 6, Appendix 1, p. 80] ) gives O = H ⊕ H ℓ \mathbb{O}=\mathbb{H}\oplus\mathbb{H}\ell O = H ⊕ H ℓ , B ^ = H ℓ \hat{B}=\mathbb{H}\ell B ^ = H ℓ and gives the following rules for any p , r , s ∈ H p,r,s\in\mathbb{H} p , r , s ∈ H :
p ( r ℓ ) = ( r p ) ℓ , ( r ℓ ) p = ( r p ˉ ) ℓ , ( p ℓ ) ( s ℓ ) = − s p . (38) \begin{aligned}
p(r\ell)&=(rp)\ell,\\
(r\ell)p&=(r\bar{p})\ell,\\
(p\ell)(s\ell)&=-sp. \tag*{(38)}
\end{aligned} p ( r ℓ ) ( r ℓ ) p ( p ℓ ) ( s ℓ ) = ( r p ) ℓ , = ( r p ˉ ) ℓ , = − s p . (38)
These rules hold for every unit ℓ ⊥ H \ell\perp\mathbb{H} ℓ ⊥ H . In particular, taking a unit ξ ^ ∈ B ^ \hat{\xi}\in\hat{B} ξ ^ ∈ B ^ , we obtain:
r ξ ^ = ξ ^ r ˉ , ( ξ ^ r ) p = ξ ^ ( p r ) , ( r ξ ^ ) p = ( r p ˉ ) ξ ^ . (39) \begin{aligned}
r\hat{\xi}&=\hat{\xi}\bar{r},\\
(\hat{\xi}r)p&=\hat{\xi}(pr),\\
(r\hat{\xi})p&=(r p\bar{})\hat{\xi}. \tag*{(39)}
\end{aligned} r ξ ^ ( ξ ^ r ) p ( r ξ ^ ) p = ξ ^ r ˉ , = ξ ^ ( p r ) , = ( r p ˉ ) ξ ^ . (39)
The quaternion sums used in the block calculations are
∑ j = 1 3 ρ ^ j p ρ ^ j = − p − 2 p ˉ , ∑ k = 0 3 ρ ^ k r ρ ^ k = 4 Re r , ∑ k = 0 3 ρ ^ k r ˉ ρ ^ k = − 2 r ˉ , ∑ k = 0 3 ρ ^ k r ρ ^ k = 4 Re r . (40) \begin{aligned}
\sum_{j=1}^{3}\hat{\rho}_jp\hat{\rho}_j&=-p-2\bar{p},&
\sum_{k=0}^{3}\hat{\rho}_kr\hat{\rho}_k&=4\operatorname{Re}r,\\
\sum_{k=0}^{3}\hat{\rho}_k\bar{r}\hat{\rho}_k&=-2\bar{r},&
\sum_{k=0}^{3}\hat{\rho}_kr\hat{\rho}_k&=4\operatorname{Re}r. \tag*{(40)}
\end{aligned} j = 1 ∑ 3 ρ ^ j p ρ ^ j k = 0 ∑ 3 ρ ^ k r ˉ ρ ^ k = − p − 2 p ˉ , = − 2 r ˉ , k = 0 ∑ 3 ρ ^ k r ρ ^ k k = 0 ∑ 3 ρ ^ k r ρ ^ k = 4 Re r , = 4 Re r . (40)
The group K K K fixes H \mathbb{H} H pointwise (it fixes u u u and v v v ) and acts by [4 Chapter 6, Appendix 2, p. 81]
p + r ℓ ⟼ p + ( g r ) ℓ , for g ∈ Sp ( 1 ) ≅ K . (41) p+r\ell\longmapsto p+(gr)\ell,\qquad\text{for }g\in\operatorname{Sp}(1)\cong K. \tag*{(41)} p + r ℓ ⟼ p + ( g r ) ℓ , for g ∈ Sp ( 1 ) ≅ K . (41)
Let 1 , 3 , 4 \mathbf{1},\mathbf{3},\mathbf{4} 1 , 3 , 4 denote the irreducible real K K K -modules of dimensions 1 , 3 , 4 1,3,4 1 , 3 , 4 (the K K K -types). Here 3 \mathbf{3} 3 is the real adjoint module, and 4 \mathbf{4} 4 is the underlying real defining module C 2 \mathbb{C}^{2} C 2 , of quaternionic type [8 Section 2.2, pp. 5–6, Lemma 2.3] ; see also [5 Section 26.3, Proposition 26.24] . In these coordinates B ^ ≅ H \hat{B}\cong\mathbb{H} B ^ ≅ H , with left quaternion multiplication, is the type 4 \mathbf{4} 4 , has no fixed vector, and satisfies B ^ C ≅ C 2 ⊕ C 2 \hat{B}_{\mathbb{C}}\cong\mathbb{C}^{2}\oplus\mathbb{C}^{2} B ^ C ≅ C 2 ⊕ C 2 , two copies of the defining (spin-1 2 \frac{1}{2} 2 1 ) module C 2 \mathbb{C}^{2} C 2 of SU ( 2 ) \operatorname{SU}(2) SU ( 2 ) . Its commutant is the four-dimensional algebra of right quaternion multiplications in these coordinates, equivalently the octonionic maps ξ ^ ↦ γ ξ ^ \hat{\xi}\mapsto\gamma\hat{\xi} ξ ^ ↦ γ ξ ^ , by (A.3).
Lemma A.1. Let W = O ⊗ V W=\mathbb{O}\otimes V W = O ⊗ V , O = H ⊕ B ^ \mathbb{O}=\mathbb{H}\oplus\widehat{B} O = H ⊕ B and V = A ⊕ B V=A\oplus B V = A ⊕ B . As real K K K -modules, H ⊗ A \mathbb{H}\otimes A H ⊗ A is trivial of dimension 12, and
( H ⊗ B ) ⊕ ( B ^ ⊗ A ) ≅ 4 ⊕ 7 and B ^ ⊗ B ≅ 1 ⊕ 4 ⊕ 3 ⊕ 4 . (\mathbb{H}\otimes B)\oplus(\widehat{B}\otimes A)\cong4^{\oplus7}
\qquad\text{and}\qquad
\widehat{B}\otimes B\cong1^{\oplus4}\oplus3^{\oplus4}. ( H ⊗ B ) ⊕ ( B ⊗ A ) ≅ 4 ⊕ 7 and B ⊗ B ≅ 1 ⊕ 4 ⊕ 3 ⊕ 4 . Consequently dim W K = 16 \dim W^K=16 dim W K = 16 , and the K K K -isotypic sectors of § 4.2 satisfy W 3 ⊂ B ^ ⊗ B W_3\subset\widehat{B}\otimes B W 3 ⊂ B ⊗ B , with dim W 3 = 12 \dim W_3=12 dim W 3 = 12 , and W 4 = ( H ⊗ B ) ⊕ ( B ^ ⊗ A ) W_4=(\mathbb{H}\otimes B)\oplus(\widehat{B}\otimes A) W 4 = ( H ⊗ B ) ⊕ ( B ⊗ A ) , with dim W 4 = 28 \dim W_4=28 dim W 4 = 28 .
Proof. The group K K K is trivial on H \mathbb{H} H and on A A A , so H ⊗ A ≅ 1 ⊕ 12 \mathbb{H}\otimes A\cong1^{\oplus12} H ⊗ A ≅ 1 ⊕ 12 . Since B ≅ B ^ B\cong\widehat{B} B ≅ B , the summands H ⊗ B \mathbb{H}\otimes B H ⊗ B and B ^ ⊗ A \widehat{B}\otimes A B ⊗ A give four and three copies of 4 4 4 , respectively. For the remaining tensor product, ( B ^ ⊗ B ) C ≅ 4 ( C 2 ⊗ C 2 ) ≅ 4 ( 1 ⊕ 3 ) (\widehat{B}\otimes B)_{\mathbb{C}}\cong4(\mathbb{C}^2\otimes\mathbb{C}^2)\cong4(1\oplus3) ( B ⊗ B ) C ≅ 4 ( C 2 ⊗ C 2 ) ≅ 4 ( 1 ⊕ 3 ) , by the Clebsch–Gordan decomposition C 2 ⊗ C 2 ≅ S 2 C 2 ⊕ Λ 2 C 2 \mathbb{C}^2\otimes\mathbb{C}^2\cong S^2\mathbb{C}^2\oplus\Lambda^2\mathbb{C}^2 C 2 ⊗ C 2 ≅ S 2 C 2 ⊕ Λ 2 C 2 into irreducible modules of dimensions 3 and 1 [7 p. 428 and Theorem C.1] . Thus B ^ ⊗ B \widehat{B}\otimes B B ⊗ B gives four trivial and four adjoint copies. Since H ⊗ A \mathbb{H}\otimes A H ⊗ A has only type 1, ( H ⊗ B ) ⊕ ( B ^ ⊗ A ) (\mathbb{H}\otimes B)\oplus(\widehat{B}\otimes A) ( H ⊗ B ) ⊕ ( B ⊗ A ) only type 4 and B ^ ⊗ B \widehat{B}\otimes B B ⊗ B only types 1, 3, the isotypic sectors are as stated, and dim W K = 12 + 4 \dim W^K=12+4 dim W K = 12 + 4 .
Recall the real S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -modules of § 3 , whose real irreducibility is standard [9 Section 2, pp. 58–60] . The group S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n acts on 8 = s u ( n ⊥ , J n ) \mathbf{8}=\mathfrak{su}(n^\perp,J_n) 8 = su ( n ⊥ , J n ) by the adjoint action, so that the complexification of 8 \mathbf{8} 8 is s l 3 ( C ) \mathfrak{sl}_3(\mathbb{C}) sl 3 ( C ) . These descriptions and the branching rules are those of [6 Definition 5.6, Propositions 5.8 and 5.11] .
Lemma A.2. Let n ∈ A ⊂ V n\in A\subset V n ∈ A ⊂ V be a unit, the copy of a unit n ^ ∈ A ^ \widehat{n}\in\widehat{A} n ∈ A , and S U ( 3 ) n = Stab G 2 ( n ) \mathrm{SU}(3)_n=\operatorname{Stab}_{G_2}(n) SU ( 3 ) n = Stab G 2 ( n ) . Then K ⊂ S U ( 3 ) n K\subset\mathrm{SU}(3)_n K ⊂ SU ( 3 ) n is the standard block embedding, and the real modules 1 , 6 , 8 , 12 1,6,8,12 1 , 6 , 8 , 12 of § 3 restrict as
1 ↓ K = 1 , and 6 ↓ K = 2 ⋅ 1 ⊕ 4 , 8 ↓ K = 1 ⊕ 3 ⊕ 4 , and 12 ↓ K = 2 ⋅ 1 ⊕ 2 ⋅ 3 ⊕ 4. (42) \begin{aligned}
1\mathord{\downarrow}_K&=1, &&\text{and}&
6\mathord{\downarrow}_K&=2\cdot1\oplus4,\\
8\mathord{\downarrow}_K&=1\oplus3\oplus4, &&\text{and}&
12\mathord{\downarrow}_K&=2\cdot1\oplus2\cdot3\oplus4. \tag*{(42)}
\end{aligned} 1 ↓ K 8 ↓ K = 1 , = 1 ⊕ 3 ⊕ 4 , and and 6 ↓ K 12 ↓ K = 2 ⋅ 1 ⊕ 4 , = 2 ⋅ 1 ⊕ 2 ⋅ 3 ⊕ 4. (42) where the types on the left are S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n -types and those on the right are K K K -types. Consequently each of the four modules has a nonzero K K K -fixed vector.
Proof. The group K K K fixes H \mathbb{H} H , hence A A A , pointwise, so it fixes n n n and lies in S U ( 3 ) n \mathrm{SU}(3)_n SU ( 3 ) n . It fixes pointwise the complex line A ∩ n ⊥ A\cap n^\perp A ∩ n ⊥ of ( n ⊥ , J n ) ≅ C 3 (n^\perp,J_n)\cong\mathbb{C}^3 ( n ⊥ , J n ) ≅ C 3 , which is J n J_n J n -stable because n ^ × A ^ ⊂ A ^ \widehat{n}\times\widehat{A}\subset\widehat{A} n × A ⊂ A , and it acts on the complementary B ≅ C 2 B\cong\mathbb{C}^2 B ≅ C 2 as S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) . This is the standard block embedding, under which C 3 ≅ C ⊕ C 2 \mathbb{C}^3\cong\mathbb{C}\oplus\mathbb{C}^2 C 3 ≅ C ⊕ C 2 as a complex K K K -module, with K K K trivial on C \mathbb{C} C . The branchings follow from this and the Clebsch–Gordan decomposition [7 p. 428 and Theorem C.1] ; for the adjoint module compare [7 Section 6.9, Exercise 4] .
The sector calculations produce diagonal-plus-skew real 4 × 4 4\times4 4 × 4 matrices.
Lemma A.3. For real d 0 d_0 d 0 , d 1 d_1 d 1 and w , k ∈ R 3 \mathbf{w},\mathbf{k}\in\mathbb{R}^3 w , k ∈ R 3 , let Λ = diag ( d 0 , d 1 , d 1 , d 1 ) \Lambda=\operatorname{diag}(d_0,d_1,d_1,d_1) Λ = diag ( d 0 , d 1 , d 1 , d 1 ) and Θ = ( 0 − w T w [ k ] × ) \Theta=\begin{pmatrix}0&-\mathbf{w}^{\mathsf{T}}\\ \mathbf{w}&[\mathbf{k}]_\times\end{pmatrix} Θ = ( 0 w − w T [ k ] × ) , where [ k ] × η = k × η [\mathbf{k}]_\times\boldsymbol{\eta}=\mathbf{k}\times\boldsymbol{\eta} [ k ] × η = k × η , a skew 4 × 4 4\times4 4 × 4 matrix. Then
det ( Λ + Θ ) = d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 . \det(\Lambda+\Theta)=d_0d_1^3+d_1^2\lvert\mathbf{w}\rvert^2+d_0d_1\lvert\mathbf{k}\rvert^2+(\mathbf{w}\cdot\mathbf{k})^2. det ( Λ + Θ ) = d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 . Proof. For a diagonal matrix Λ \Lambda Λ and a skew matrix Θ \Theta Θ , the standard Pfaffian expansion is
det ( Λ + Θ ) = ∑ I Pf ( Θ I ) 2 ∏ j ∉ I Λ j j , \det(\Lambda+\Theta)=\sum_{\mathcal{I}}\operatorname{Pf}(\Theta_{\mathcal{I}})^2\prod_{j\notin\mathcal{I}}\Lambda_{jj}, det ( Λ + Θ ) = I ∑ Pf ( Θ I ) 2 j ∈ / I ∏ Λ j j , the sum over I ⊂ { 0 , 1 , 2 , 3 } \mathcal{I}\subset\{0,1,2,3\} I ⊂ { 0 , 1 , 2 , 3 } with ∣ I ∣ \lvert\mathcal{I}\rvert ∣ I ∣ even (principal submatrices Θ I \Theta_{\mathcal{I}} Θ I , empty Pfaffian 1). For Λ = diag ( d 0 , d 1 , d 1 , d 1 ) \Lambda=\operatorname{diag}(d_0,d_1,d_1,d_1) Λ = diag ( d 0 , d 1 , d 1 , d 1 ) , the empty index set gives d 0 d 1 3 d_0d_1^3 d 0 d 1 3 , the full Pfaffian is w ⋅ k \mathbf{w}\cdot\mathbf{k} w ⋅ k , and the two-index terms give d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 d_1^2\lvert\mathbf{w}\rvert^2+d_0d_1\lvert\mathbf{k}\rvert^2 d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 .
The W1 calculation and Q14
This appendix supplies the fixed-sector blocks, Schur reduction, polynomial matrix and invariant-polynomial argument used in § 5.1 .
We use the quaternionic normal form of Proposition 3.8 , Φ : W 1 ⟶ H 3 ⊕ H \Phi: W_1 \longrightarrow\mathbb{H}^3 \oplus\mathbb{H} Φ : W 1 ⟶ H 3 ⊕ H , with Φ ( ϑ ⊗ ρ j ) = ϑ \Phi(\vartheta\otimes\rho_j)=\vartheta Φ ( ϑ ⊗ ρ j ) = ϑ in slot j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 . The slots 1 , 2 , 3 1,2,3 1 , 2 , 3 represent H ⊗ A ⊂ W + \mathbb{H}\otimes A\subset W_+ H ⊗ A ⊂ W + , and slot 0 0 0 represents { ε γ : γ ∈ H } ⊂ W − \{\varepsilon_\gamma:\gamma\in\mathbb{H}\}\subset W_- { ε γ : γ ∈ H } ⊂ W − , with Φ ( ε γ ) = z γ z ˉ \Phi(\varepsilon_\gamma)=z\gamma\bar{z} Φ ( ε γ ) = z γ z ˉ .
Write h A A , h A 0 , h 0 A , h 00 h_{AA},h_{A0},h_{0A},h_{00} h AA , h A 0 , h 0 A , h 00 for the blocks of Φ h Φ − 1 \Phi h\Phi^{-1} Φ h Φ − 1 , and use the same subscripts for Φ P a Φ − 1 \Phi P_a\Phi^{-1} Φ P a Φ − 1 . The scalar part is R q ^ R_{\hat q} R q ^ in every slot, by (23) , so that P A A = h A A − R q ^ P_{AA}=h_{AA}-R_{\hat q} P AA = h AA − R q ^ , P A 0 = h A 0 P_{A0}=h_{A0} P A 0 = h A 0 , P 0 A = h 0 A P_{0A}=h_{0A} P 0 A = h 0 A and P 00 = h 00 − R q ^ P_{00}=h_{00}-R_{\hat q} P 00 = h 00 − R q ^ . The coordinate metric G = diag ( I 12 , 4 I 4 ) G=\operatorname{diag}(I_{12},4I_4) G = diag ( I 12 , 4 I 4 ) gives h A 0 = 4 h 0 A T h_{A0}=4h_{0A}^{\mathsf T} h A 0 = 4 h 0 A T in Euclidean coordinate matrices.
Use x , y , N x,y,N x , y , N from (12) , q ^ , q ˇ \hat q,\check q q ^ , q ˇ from (13) , and t = x − y t=x-y t = x − y from § 5.1 . For the calculations below, define
q ⊥ = a 1 u + a 2 v and N 3 = ∣ q ^ + 3 ∣ 2 = N + 6 a 0 + 9. q_\perp=a_1u+a_2v \qquad\text{and}\qquad N_3=\lvert\hat q+3\rvert^2=N+6a_0+9. q ⊥ = a 1 u + a 2 v and N 3 = ∣ q ^ + 3 ∣ 2 = N + 6 a 0 + 9.
We also fix the real vectors p = Im q ^ = ( a 1 , a 2 , a 3 ) \mathbf p=\operatorname{Im}\hat q=(a_1,a_2,a_3) p = Im q ^ = ( a 1 , a 2 , a 3 ) and m = Im q ˇ = ( a 1 , a 2 , − a 3 ) \mathbf m=\operatorname{Im}\check q=(a_1,a_2,-a_3) m = Im q ˇ = ( a 1 , a 2 , − a 3 ) in the basis ( u , v , z ) (u,v,z) ( u , v , z ) , oriented by u × v = z u\times v=z u × v = z , so that ∣ p ∣ 2 = ∣ m ∣ 2 = x + y \lvert\mathbf p\rvert^2=\lvert\mathbf m\rvert^2=x+y ∣ p ∣ 2 = ∣ m ∣ 2 = x + y and p ⋅ m = x − y = t \mathbf p\cdot\mathbf m=x-y=t p ⋅ m = x − y = t . For a real 3 3 3 -vector v \mathbf v v , [ v ] × [\mathbf v]_\times [ v ] × denotes the matrix of η ↦ v × η \eta\mapsto\mathbf v\times\eta η ↦ v × η .
Lemma B.1. For ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ∈ H 3 ⊕ H (\vartheta_1,\vartheta_2,\vartheta_3;\eta)\in\mathbb{H}^3\oplus\mathbb{H} ( ϑ 1 , ϑ 2 , ϑ 3 ; η ) ∈ H 3 ⊕ H , with ϑ = ( ϑ 1 , ϑ 2 , ϑ 3 ) \boldsymbol{\vartheta}=(\vartheta_1,\vartheta_2,\vartheta_3) ϑ = ( ϑ 1 , ϑ 2 , ϑ 3 ) , the four blocks of Φ h Φ − 1 \Phi h\Phi^{-1} Φ h Φ − 1 are
( h A A ϑ ) i = ∑ j = 1 3 ( ρ ^ i ρ ^ j ) ϑ j , and ( h A 0 η ) i = − 4 z ˉ η z ρ ^ ˉ i , (h_{AA}\boldsymbol{\vartheta})_i=\sum_{j=1}^{3}(\hat\rho_i\hat\rho_j)\vartheta_j,\qquad\text{and}\qquad(h_{A0}\eta)_i=-4\bar z\eta z\bar{\hat\rho}_i, ( h AA ϑ ) i = j = 1 ∑ 3 ( ρ ^ i ρ ^ j ) ϑ j , and ( h A 0 η ) i = − 4 z ˉ η z ρ ^ ˉ i , h 0 A ϑ = z ( ∑ j = 1 3 ϑ j ρ ^ j ) z ˉ , and h 00 η = − 4 Re η . (43) h_{0A}\boldsymbol{\vartheta}=z\left(\sum_{j=1}^{3}\vartheta_j\hat\rho_j\right)\bar z,\qquad\text{and}\qquad h_{00}\eta=-4\operatorname{Re}\eta. \tag*{(43)} h 0 A ϑ = z ( j = 1 ∑ 3 ϑ j ρ ^ j ) z ˉ , and h 00 η = − 4 Re η . (43)
Proof. We evaluate the defining formula (3) for h h h , over the orthonormal basis { ρ 1 , ρ 2 , ρ 3 , ξ 0 , … , ξ 3 } \{\rho_1,\rho_2,\rho_3,\xi_0,\ldots,\xi_3\} { ρ 1 , ρ 2 , ρ 3 , ξ 0 , … , ξ 3 } of V V V . Take a unit ℓ ∈ B ^ \ell\in\hat B ℓ ∈ B ^ and ξ ^ α = ρ ^ α ℓ \hat\xi_\alpha=\hat\rho_\alpha\ell ξ ^ α = ρ ^ α ℓ . The Cayley–Dickson rules (A.2) of Appendix § 7 give ( ℓ a ) p = p ˉ a ℓ (\ell a)p=\bar p a\ell ( ℓ a ) p = p ˉ a ℓ and ( a ℓ ) ( r ℓ ) = − r ˉ a (a\ell)(r\ell)=-\bar r a ( a ℓ ) ( r ℓ ) = − r ˉ a for a , p , r ∈ H a,p,r\in\mathbb{H} a , p , r ∈ H . For a generic element ϑ ℓ \vartheta\ell ϑ ℓ of B ^ \hat B B ^ , with ϑ ∈ H \vartheta\in\mathbb{H} ϑ ∈ H , the contribution from B ^ ⊗ B \hat B\otimes B B ^ ⊗ B is computed before summing:
( ( ϑ ℓ ) ξ ^ α ) ( r ξ ^ α ) = − ( ( ρ ^ α r ρ ^ α ) ϑ ) ℓ , and ∑ k = 0 3 ρ ^ k r ρ ^ k = 4 Re r . ((\vartheta\ell)\hat\xi_\alpha)(r\hat\xi_\alpha)=-((\hat\rho_\alpha r\hat\rho_\alpha)\vartheta)\ell,\qquad\text{and}\qquad\sum_{k=0}^{3}\hat\rho_k r\hat\rho_k=4\operatorname{Re}r. (( ϑ ℓ ) ξ ^ α ) ( r ξ ^ α ) = − (( ρ ^ α r ρ ^ α ) ϑ ) ℓ , and k = 0 ∑ 3 ρ ^ k r ρ ^ k = 4 Re r .
Thus the sum over α \alpha α is − 4 ( Re r ) ϑ ℓ -4(\operatorname{Re}r)\vartheta\ell − 4 ( Re r ) ϑ ℓ . For the other blocks we use the first sum of (A.4) and the rule ρ ^ i ρ ^ j = − δ i j + ϵ i j k ρ ^ k \hat\rho_i\hat\rho_j=-\delta_{ij}+\epsilon_{ijk}\hat\rho_k ρ ^ i ρ ^ j = − δ ij + ϵ ij k ρ ^ k . Summing the four contributions, with η = z r z ˉ \eta=z r\bar z η = z r z ˉ , gives (B.1).
Let M i j = ρ ^ i ρ ^ j M_{ij}=\hat\rho_i\hat\rho_j M ij = ρ ^ i ρ ^ j , so h A A = L M h_{AA}=L_M h AA = L M . Set r = ( u , v , z ) T \mathbf r=(u,v,z)^{\mathsf T} r = ( u , v , z ) T ; then M = − r r ∗ M=-\mathbf r\mathbf r^* M = − r r ∗ and r ∗ r = 3 \mathbf r^*\mathbf r=3 r ∗ r = 3 . Thus H 3 = U 3 ⊕ U 0 \mathbb{H}^3=U_3\oplus U_0 H 3 = U 3 ⊕ U 0 , where U 3 = r H U_3=\mathbf r\mathbb{H} U 3 = r H and U 0 = ker r ∗ U_0=\ker\mathbf r^* U 0 = ker r ∗ orthogonally, with projections r r ∗ / 3 \mathbf r\mathbf r^*/3 r r ∗ /3 and I − r r ∗ / 3 I-\mathbf r\mathbf r^*/3 I − r r ∗ /3 and quaternionic ranks one and two. On these right submodules L M L_M L M is − 3 -3 − 3 and 0 0 0 , respectively, so P A A P_{AA} P AA is − R q ^ + 3 -R_{\hat q+3} − R q ^ + 3 and − R q ^ -R_{\hat q} − R q ^ . Right multiplication by a quaternion p p p on H \mathbb{H} H has real determinant ∣ p ∣ 4 |p|^4 ∣ p ∣ 4 , giving the determinant and inverse below. It follows that
det P A A = N 4 N 3 2 , \det P_{AA}=N^4N_3^2, det P AA = N 4 N 3 2 ,
P A A − 1 = − Π U 0 R q ^ − 1 − Π U 3 R ( 3 + q ^ ) − 1 , for N N 3 ≠ 0 , (44) P_{AA}^{-1}=-\Pi_{U_0}R_{\hat{q}}^{-1}-\Pi_{U_3}R_{(3+\hat{q})}^{-1},\qquad\text{for }NN_3\ne0, \tag*{(44)} P AA − 1 = − Π U 0 R q ^ − 1 − Π U 3 R ( 3 + q ^ ) − 1 , for N N 3 = 0 , (44)
Π U 0 = I − r r ∗ / 3 , and Π U 3 = r r ∗ / 3. \Pi_{U_0}=I-\mathbf{r}\mathbf{r}^{*}/3,\qquad\text{and}\qquad\Pi_{U_3}=\mathbf{r}\mathbf{r}^{*}/3. Π U 0 = I − r r ∗ /3 , and Π U 3 = r r ∗ /3.
Schur reduction and the closed form
Lemma B.2. Suppose N N 3 ≠ 0 NN_3\ne0 N N 3 = 0 . Define ζ 1 = ( z q ^ z ˉ ) − 1 \zeta_1=(z\hat{q}\bar{z})^{-1} ζ 1 = ( z q ^ z ˉ ) − 1 and ζ 2 = ( 3 + z q ^ z ˉ ) − 1 \zeta_2=(3+z\hat{q}\bar{z})^{-1} ζ 2 = ( 3 + z q ^ z ˉ ) − 1 . The Schur complement S = P 00 − P 0 A P A A − 1 P A 0 S=P_{00}-P_{0A}P_{AA}^{-1}P_{A0} S = P 00 − P 0 A P AA − 1 P A 0 is given by
S η = A η + η B + C η ˉ + η ˉ D , (45) S\eta=A\eta+\eta B+C\bar{\eta}+\bar{\eta}D, \tag*{(45)} S η = A η + η B + C η ˉ + η ˉ D , (45) where
A = − 2 + 4 3 ( 2 ζ ˉ 1 + 6 ζ 1 + ζ ˉ 2 ) , B = − q ^ − 16 3 ( ζ 1 − ζ 2 ) , A=-2+\frac{4}{3}(2\bar{\zeta}_1+6\zeta_1+\bar{\zeta}_2),\qquad B=-\hat{q}-\frac{16}{3}(\zeta_1-\zeta_2), A = − 2 + 3 4 ( 2 ζ ˉ 1 + 6 ζ 1 + ζ ˉ 2 ) , B = − q ^ − 3 16 ( ζ 1 − ζ 2 ) , C = − 2 − 8 3 ( ζ 1 − ζ ˉ 2 ) , D = − 8 3 ( ζ 1 − ζ 2 ) . (46) C=-2-\frac{8}{3}(\zeta_1-\bar{\zeta}_2),\qquad D=-\frac{8}{3}(\zeta_1-\zeta_2). \tag*{(46)} C = − 2 − 3 8 ( ζ 1 − ζ ˉ 2 ) , D = − 3 8 ( ζ 1 − ζ 2 ) . (46) Moreover, det ( P a ∣ W 1 ) = N 4 N 3 2 det R S \det(P_a|_{W_1})=N^4N_3^2\det_{\mathbb{R}}S det ( P a ∣ W 1 ) = N 4 N 3 2 det R S .
Proof. By direct computation from (B.2), the blocks (B.1) and the first sum of (A.4), we obtain P A A − 1 h A 0 η P_{AA}^{-1}h_{A0}\eta P AA − 1 h A 0 η . We apply h 0 A h_{0A} h 0 A using (B.1) and the same quaternion sum once more, and subtract from P 00 η = − 2 η − 2 η ˉ − η q ^ P_{00}\eta=-2\eta-2\bar{\eta}-\eta\hat{q} P 00 η = − 2 η − 2 η ˉ − η q ^ . The resulting coefficients are those in (B.4). The determinant identity is the standard Schur-complement formula [3 §8, (9), pp. 106–107] , together with det P A A = N 4 N 3 2 \det P_{AA}=N^4N_3^2 det P AA = N 4 N 3 2 from (B.2).
We now write the Schur map as a diagonal matrix plus a skew matrix. Since Im C = − Im D \operatorname{Im}C=-\operatorname{Im}D Im C = − Im D in (B.4), the real matrix of S η = A η + η B + C η ˉ + η ˉ D S\eta=A\eta+\eta B+C\bar{\eta}+\bar{\eta}D S η = A η + η B + C η ˉ + η ˉ D in the basis ( 1 , u , v , z ) (1,u,v,z) ( 1 , u , v , z ) is
( d 0 − w T w d 1 I 3 + [ k ] × ) , \begin{pmatrix}
d_0 & -\mathbf{w}^{\mathsf{T}}\\
\mathbf{w} & d_1I_3+[\mathbf{k}]_{\times}
\end{pmatrix}, ( d 0 w − w T d 1 I 3 + [ k ] × ) ,
where [ k ] × η = k × η [\mathbf{k}]_{\times}\eta=\mathbf{k}\times\eta [ k ] × η = k × η and
d 0 = Re ( A + B + C + D ) , d 1 = Re ( A + B − C − D ) , d_0=\operatorname{Re}(A+B+C+D),\qquad d_1=\operatorname{Re}(A+B-C-D), d 0 = Re ( A + B + C + D ) , d 1 = Re ( A + B − C − D ) ,
w = Im ( A + B ) , k = Im ( A − B − 2 C ) . \mathbf{w}=\operatorname{Im}(A+B),\qquad\mathbf{k}=\operatorname{Im}(A-B-2C). w = Im ( A + B ) , k = Im ( A − B − 2 C ) .
With Λ = diag ( d 0 , d 1 , d 1 , d 1 ) \Lambda=\operatorname{diag}(d_0,d_1,d_1,d_1) Λ = diag ( d 0 , d 1 , d 1 , d 1 ) and the skew matrix Θ \Theta Θ built from w , k \mathbf{w},\mathbf{k} w , k , the matrix above is Λ + Θ \Lambda+\Theta Λ + Θ , so Lemma A.3 of Appendix A gives
det R S = d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 . (47) \det_{\mathbb{R}}S=d_0d_1^3+d_1^2|\mathbf{w}|^2+d_0d_1|\mathbf{k}|^2+(\mathbf{w}\cdot\mathbf{k})^2. \tag*{(47)} R det S = d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 . (47)
Since z q ^ z ˉ = a 0 − q ⊥ + a 3 z z\hat{q}\bar{z}=a_0-q_{\perp}+a_3z z q ^ z ˉ = a 0 − q ⊥ + a 3 z , the real parts of ζ 1 , ζ 2 \zeta_1,\zeta_2 ζ 1 , ζ 2 are a 0 / N , ( a 0 + 3 ) / N 3 a_0/N,(a_0+3)/N_3 a 0 / N , ( a 0 + 3 ) / N 3 , and their imaginary parts are m / N , m / N 3 \mathbf{m}/N,\mathbf{m}/N_3 m / N , m / N 3 , respectively. Substitution gives, on N N 3 ≠ 0 NN_3\ne0 N N 3 = 0 ,
d 0 = − 4 − a 0 + 12 ( a 0 + 3 ) N 3 , and d 1 = − a 0 + 32 a 0 3 N + 4 ( a 0 + 3 ) 3 N 3 , d_0=-4-a_0+\frac{12(a_0+3)}{N_3},\qquad\text{and}\qquad d_1=-a_0+\frac{32a_0}{3N}+\frac{4(a_0+3)}{3N_3}, d 0 = − 4 − a 0 + N 3 12 ( a 0 + 3 ) , and d 1 = − a 0 + 3 N 32 a 0 + 3 N 3 4 ( a 0 + 3 ) ,
w = − p + 4 N 3 m , and k = p + ν m , (48) \mathbf{w}=-\mathbf{p}+\frac{4}{N_3}\mathbf{m},\qquad\text{and}\qquad\mathbf{k}=\mathbf{p}+\nu\mathbf{m}, \tag*{(48)} w = − p + N 3 4 m , and k = p + ν m , (48)
ν = 4 3 ( 4 N − 1 N 3 ) . \nu=\frac{4}{3}\left(\frac{4}{N}-\frac{1}{N_3}\right). ν = 3 4 ( N 4 − N 3 1 ) .
Since ∣ p ∣ 2 = ∣ m ∣ 2 = x + y \lvert\mathbf{p}\rvert^2=\lvert\mathbf{m}\rvert^2=x+y ∣ p ∣ 2 = ∣ m ∣ 2 = x + y and p ⋅ m = t \mathbf{p}\cdot\mathbf{m}=t p ⋅ m = t , we find
∣ w ∣ 2 = ( x + y ) ( 1 + 16 N 3 2 ) − 8 t N 3 \lvert\mathbf{w}\rvert^2=(x+y)\left(1+\frac{16}{N_3^2}\right)-\frac{8t}{N_3} ∣ w ∣ 2 = ( x + y ) ( 1 + N 3 2 16 ) − N 3 8 t
∣ k ∣ 2 = ( x + y ) ( 1 + ν 2 ) + 2 ν t , \lvert\mathbf{k}\rvert^2=(x+y)(1+\nu^2)+2\nu t, ∣ k ∣ 2 = ( x + y ) ( 1 + ν 2 ) + 2 ν t ,
w ⋅ k = − ( x + y ) + ( 4 N 3 − ν ) t + 4 ν ( x + y ) N 3 . \mathbf{w}\cdot\mathbf{k}=-(x+y)+\left(\frac{4}{N_3}-\nu\right)t+\frac{4\nu(x+y)}{N_3}. w ⋅ k = − ( x + y ) + ( N 3 4 − ν ) t + N 3 4 ν ( x + y ) .
These are rational functions of ( a 0 , x , y ) (a_0,x,y) ( a 0 , x , y ) .
Lemma B.3. Let p , m \mathbf{p},\mathbf{m} p , m be as above, let D = diag ( 3 , 1 , 1 , 1 ) D=\operatorname{diag}(3,1,1,1) D = diag ( 3 , 1 , 1 , 1 ) and Δ = diag ( 1 , N , N , N ) \Delta=\operatorname{diag}(1,N,N,N) Δ = diag ( 1 , N , N , N ) . Define the following matrices in the basis ( 1 , u , v , z ) (1,u,v,z) ( 1 , u , v , z ) :
S p o l = ( − 4 − a 0 p T − p − a 0 I + [ p ] × ) , S_{\mathrm{pol}}=
\begin{pmatrix}
-4-a_0 & \mathbf{p}^{\mathsf{T}}\\
-\mathbf{p} & -a_0I+[\mathbf{p}]_\times
\end{pmatrix}, S pol = ( − 4 − a 0 − p p T − a 0 I + [ p ] × ) , B N = ( 0 0 0 2 a 0 I + [ m ] × ) , B_N=
\begin{pmatrix}
0 & 0\\
0 & 2a_0I+[\mathbf{m}]_\times
\end{pmatrix}, B N = ( 0 0 0 2 a 0 I + [ m ] × ) , Υ = ( a 0 + 3 − m T m ( a 0 + 3 ) I − [ m ] × ) , \Upsilon=
\begin{pmatrix}
a_0+3 & -\mathbf{m}^{\mathsf{T}}\\
\mathbf{m} & (a_0+3)I-[\mathbf{m}]_\times
\end{pmatrix}, Υ = ( a 0 + 3 m − m T ( a 0 + 3 ) I − [ m ] × ) , where S p o l S_{\mathrm{pol}} S pol is the matrix of P 00 P_{00} P 00 and Υ \Upsilon Υ is the matrix of right multiplication by a 0 + 3 + m = 3 + z q ^ z ˉ a_0+3+\mathbf{m}=3+z\widehat{\mathbf{q}}\bar{z} a 0 + 3 + m = 3 + z q z ˉ . Then, on N N 3 ≠ 0 NN_3\ne0 N N 3 = 0 ,
S = S p o l + 16 3 N B N + 4 3 N 3 D Υ D , S=S_{\mathrm{pol}}+\frac{16}{3N}B_N+\frac{4}{3N_3}D\Upsilon D, S = S pol + 3 N 16 B N + 3 N 3 4 D Υ D , Υ Υ T = N 3 I and det Υ = N 3 2 . (49) \Upsilon\Upsilon^{\mathsf{T}}=N_3I
\qquad\text{and}\qquad
\det\Upsilon=N_3^2. \tag*{(49)} Υ Υ T = N 3 I and det Υ = N 3 2 . (49) Moreover the matrix
M p o l : = ( Δ S p o l + 16 3 B N ) Υ ~ + 4 3 diag ( 9 , N , N , N ) , M_{\mathrm{pol}}:=\left(\Delta S_{\mathrm{pol}}+\frac{16}{3}B_N\right)\widetilde{\Upsilon}+\frac{4}{3}\operatorname{diag}(9,N,N,N), M pol := ( Δ S pol + 3 16 B N ) Υ + 3 4 diag ( 9 , N , N , N ) , Υ ~ = D − 1 Υ D (50) \widetilde{\Upsilon}=D^{-1}\Upsilon D \tag*{(50)} Υ = D − 1 Υ D (50) has entries in Q [ a 0 , a 1 , a 2 , a 3 ] \mathbb{Q}[a_0,a_1,a_2,a_3] Q [ a 0 , a 1 , a 2 , a 3 ] , and M p o l = Δ S D − 1 Υ T D M_{\mathrm{pol}}=\Delta S D^{-1}\Upsilon^{\mathsf{T}}D M pol = Δ S D − 1 Υ T D on N N 3 ≠ 0 NN_3\ne0 N N 3 = 0 .
Proof. Split the poles. Split each entry of the data (B.6) into its polynomial part, its 1 / N 1/N 1/ N part and its 1 / N 3 1/N_3 1/ N 3 part, using ν m = 16 3 N m − 4 3 N 3 m \nu\mathbf{m}=\frac{16}{3N}\mathbf{m}-\frac{4}{3N_3}\mathbf{m} ν m = 3 N 16 m − 3 N 3 4 m . The polynomial parts − 4 − a 0 -4-a_0 − 4 − a 0 , − a 0 -a_0 − a 0 , − p -\mathbf{p} − p , p \mathbf{p} p of d 0 , d 1 , w , k d_0,d_1,\mathbf{w},\mathbf{k} d 0 , d 1 , w , k give S p o l S_{\mathrm{pol}} S pol ; only d 1 d_1 d 1 and k \mathbf{k} k have a 1 / N 1/N 1/ N part, 32 a 0 3 N \frac{32a_0}{3N} 3 N 32 a 0 and 16 3 N m \frac{16}{3N}\mathbf{m} 3 N 16 m , which give 16 3 N B N \frac{16}{3N}B_N 3 N 16 B N ; and the 1 / N 3 1/N_3 1/ N 3 parts 12 ( a 0 + 3 ) 12(a_0+3) 12 ( a 0 + 3 ) , 4 m 4\mathbf{m} 4 m , 4 3 ( a 0 + 3 ) \frac{4}{3}(a_0+3) 3 4 ( a 0 + 3 ) , − 4 3 m -\frac{4}{3}\mathbf{m} − 3 4 m of d 0 , w , d 1 , k d_0,\mathbf{w},d_1,\mathbf{k} d 0 , w , d 1 , k , each over N 3 N_3 N 3 , give 4 3 N 3 D Υ D \frac{4}{3N_3}D\Upsilon D 3 N 3 4 D Υ D .
Evaluate the quaternionic factor. Write Υ = ( a 0 + 3 ) I + Ξ \Upsilon=(a_0+3)I+\Xi Υ = ( a 0 + 3 ) I + Ξ with
Ξ = ( 0 − m T m − [ m ] × ) \Xi=
\begin{pmatrix}
0 & -\mathbf{m}^{\mathsf{T}}\\
\mathbf{m} & -[\mathbf{m}]_\times
\end{pmatrix} Ξ = ( 0 m − m T − [ m ] × )
skew. Then Ξ 2 = − ∣ m ∣ 2 I \Xi^2=-\lvert\mathbf{m}\rvert^2I Ξ 2 = − ∣ m ∣ 2 I , so Υ Υ T = ( ( a 0 + 3 ) 2 + x + y ) I = N 3 I \Upsilon\Upsilon^{\mathsf{T}}=((a_0+3)^2+x+y)I=N_3I Υ Υ T = (( a 0 + 3 ) 2 + x + y ) I = N 3 I . Hence ( det Υ ) 2 = N 3 4 (\det\Upsilon)^2=N_3^4 ( det Υ ) 2 = N 3 4 , and det Υ = N 3 2 \det\Upsilon=N_3^2 det Υ = N 3 2 , because det Υ − N 3 2 \det\Upsilon-N_3^2 det Υ − N 3 2 and det Υ + N 3 2 \det\Upsilon+N_3^2 det Υ + N 3 2 cannot both be nonzero in the domain Q [ a 0 , a 1 , a 2 , a 3 ] \mathbb{Q}[a_0,a_1,a_2,a_3] Q [ a 0 , a 1 , a 2 , a 3 ] and the value at m = 0 \mathbf{m}=0 m = 0 is ( a 0 + 3 ) 4 = N 3 2 (a_0+3)^4=N_3^2 ( a 0 + 3 ) 4 = N 3 2 .
Clear the denominators. By the pole decomposition, S D − 1 Υ T D = S p o l Υ ~ + 16 3 N B N Υ ~ + 4 3 N 3 D Υ Υ T D SD^{-1}\Upsilon^{\mathsf{T}}D=S_{\mathrm{pol}}\widetilde{\Upsilon}+\frac{16}{3N}B_N\widetilde{\Upsilon}+\frac{4}{3N_3}D\Upsilon\Upsilon^{\mathsf{T}}D S D − 1 Υ T D = S pol Υ + 3 N 16 B N Υ + 3 N 3 4 D Υ Υ T D , and D Υ Υ T D = N 3 D 2 D\Upsilon\Upsilon^{\mathsf{T}}D=N_3D^2 D Υ Υ T D = N 3 D 2 . The first row of B N B_N B N vanishes, so Δ B N = N B N \Delta B_N=NB_N Δ B N = N B N . Multiplying by Δ \Delta Δ on the left gives (B.8). ∎
Proposition B.4. Let M p o l M_{\mathrm{pol}} M pol be the polynomial matrix of Lemma B.3 , and define Q 14 : = det M p o l ∈ Q [ a 0 , a 1 , a 2 , a 3 ] Q_{14} := \det M_{\mathrm{pol}} \in\mathbb{Q}[a_0,a_1,a_2,a_3] Q 14 := det M pol ∈ Q [ a 0 , a 1 , a 2 , a 3 ] . Then, as identities of polynomials:
det ( P a ∣ W 1 ) = N det M p o l = N Q 14 \det(P_a|_{W_1}) = N \det M_{\mathrm{pol}} = N Q_{14} det ( P a ∣ W 1 ) = N det M pol = N Q 14 ;
on N N 3 ≠ 0 NN_3 \ne0 N N 3 = 0 ,
Q 14 = N 3 N 3 2 ( d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 ) (51) Q_{14} = N^3N_3^2\left(d_0d_1^3 + d_1^2|\mathbf{w}|^2 + d_0d_1|\mathbf{k}|^2 + (\mathbf{w}\cdot\mathbf{k})^2\right) \tag*{(51)} Q 14 = N 3 N 3 2 ( d 0 d 1 3 + d 1 2 ∣ w ∣ 2 + d 0 d 1 ∣ k ∣ 2 + ( w ⋅ k ) 2 ) (51) with the data of (B.6).
Proof. Since det Δ = N 3 \det\Delta= N^3 det Δ = N 3 and det ( D − 1 Υ T D ) = det Υ = N 2 \det(D^{-1}\Upsilon^{T}D) = \det\Upsilon= N^2 det ( D − 1 Υ T D ) = det Υ = N 2 (Lemma B.3 ), we have det M p o l = N 3 N 3 2 det S \det M_{\mathrm{pol}} = N^3N_3^2\det S det M pol = N 3 N 3 2 det S on N N 3 ≠ 0 NN_3 \ne0 N N 3 = 0 ; with (B.5), this is (B.9), which is item 2. By Lemma B.2 ,
det ( P a ∣ W 1 ) = N 4 N 3 2 det S = N det M p o l on N N 3 ≠ 0. \det(P_a|_{W_1}) = N^4N_3^2\det S = N\det M_{\mathrm{pol}} \qquad\text{on}\qquad NN_3 \ne0. det ( P a ∣ W 1 ) = N 4 N 3 2 det S = N det M pol on N N 3 = 0.
Since both sides are polynomials, and N N 3 ≠ 0 NN_3 \ne0 N N 3 = 0 is dense, item 1 holds identically. ∎